diff options
Diffstat (limited to '')
42 files changed, 17289 insertions, 0 deletions
diff --git a/net/rds/Kconfig b/net/rds/Kconfig new file mode 100644 index 000000000..b9092111b --- /dev/null +++ b/net/rds/Kconfig @@ -0,0 +1,27 @@ + +config RDS + tristate "The Reliable Datagram Sockets Protocol" + depends on INET + ---help--- + The RDS (Reliable Datagram Sockets) protocol provides reliable, + sequenced delivery of datagrams over Infiniband or TCP. + +config RDS_RDMA + tristate "RDS over Infiniband" + depends on RDS && INFINIBAND && INFINIBAND_ADDR_TRANS + ---help--- + Allow RDS to use Infiniband as a transport. + This transport supports RDMA operations. + +config RDS_TCP + tristate "RDS over TCP" + depends on RDS + depends on IPV6 || !IPV6 + ---help--- + Allow RDS to use TCP as a transport. + This transport does not support RDMA operations. + +config RDS_DEBUG + bool "RDS debugging messages" + depends on RDS + default n diff --git a/net/rds/Makefile b/net/rds/Makefile new file mode 100644 index 000000000..e647f9de1 --- /dev/null +++ b/net/rds/Makefile @@ -0,0 +1,17 @@ +# SPDX-License-Identifier: GPL-2.0 +obj-$(CONFIG_RDS) += rds.o +rds-y := af_rds.o bind.o cong.o connection.o info.o message.o \ + recv.o send.o stats.o sysctl.o threads.o transport.o \ + loop.o page.o rdma.o + +obj-$(CONFIG_RDS_RDMA) += rds_rdma.o +rds_rdma-y := rdma_transport.o \ + ib.o ib_cm.o ib_recv.o ib_ring.o ib_send.o ib_stats.o \ + ib_sysctl.o ib_rdma.o ib_fmr.o ib_frmr.o + + +obj-$(CONFIG_RDS_TCP) += rds_tcp.o +rds_tcp-y := tcp.o tcp_connect.o tcp_listen.o tcp_recv.o \ + tcp_send.o tcp_stats.o + +ccflags-$(CONFIG_RDS_DEBUG) := -DRDS_DEBUG diff --git a/net/rds/af_rds.c b/net/rds/af_rds.c new file mode 100644 index 000000000..cd7e01ea8 --- /dev/null +++ b/net/rds/af_rds.c @@ -0,0 +1,839 @@ +/* + * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/module.h> +#include <linux/errno.h> +#include <linux/kernel.h> +#include <linux/gfp.h> +#include <linux/in.h> +#include <linux/ipv6.h> +#include <linux/poll.h> +#include <net/sock.h> + +#include "rds.h" + +/* this is just used for stats gathering :/ */ +static DEFINE_SPINLOCK(rds_sock_lock); +static unsigned long rds_sock_count; +static LIST_HEAD(rds_sock_list); +DECLARE_WAIT_QUEUE_HEAD(rds_poll_waitq); + +/* + * This is called as the final descriptor referencing this socket is closed. + * We have to unbind the socket so that another socket can be bound to the + * address it was using. + * + * We have to be careful about racing with the incoming path. sock_orphan() + * sets SOCK_DEAD and we use that as an indicator to the rx path that new + * messages shouldn't be queued. + */ +static int rds_release(struct socket *sock) +{ + struct sock *sk = sock->sk; + struct rds_sock *rs; + + if (!sk) + goto out; + + rs = rds_sk_to_rs(sk); + + sock_orphan(sk); + /* Note - rds_clear_recv_queue grabs rs_recv_lock, so + * that ensures the recv path has completed messing + * with the socket. */ + rds_clear_recv_queue(rs); + rds_cong_remove_socket(rs); + + rds_remove_bound(rs); + + rds_send_drop_to(rs, NULL); + rds_rdma_drop_keys(rs); + rds_notify_queue_get(rs, NULL); + rds_notify_msg_zcopy_purge(&rs->rs_zcookie_queue); + + spin_lock_bh(&rds_sock_lock); + list_del_init(&rs->rs_item); + rds_sock_count--; + spin_unlock_bh(&rds_sock_lock); + + rds_trans_put(rs->rs_transport); + + sock->sk = NULL; + sock_put(sk); +out: + return 0; +} + +/* + * Careful not to race with rds_release -> sock_orphan which clears sk_sleep. + * _bh() isn't OK here, we're called from interrupt handlers. It's probably OK + * to wake the waitqueue after sk_sleep is clear as we hold a sock ref, but + * this seems more conservative. + * NB - normally, one would use sk_callback_lock for this, but we can + * get here from interrupts, whereas the network code grabs sk_callback_lock + * with _lock_bh only - so relying on sk_callback_lock introduces livelocks. + */ +void rds_wake_sk_sleep(struct rds_sock *rs) +{ + unsigned long flags; + + read_lock_irqsave(&rs->rs_recv_lock, flags); + __rds_wake_sk_sleep(rds_rs_to_sk(rs)); + read_unlock_irqrestore(&rs->rs_recv_lock, flags); +} + +static int rds_getname(struct socket *sock, struct sockaddr *uaddr, + int peer) +{ + struct rds_sock *rs = rds_sk_to_rs(sock->sk); + struct sockaddr_in6 *sin6; + struct sockaddr_in *sin; + int uaddr_len; + + /* racey, don't care */ + if (peer) { + if (ipv6_addr_any(&rs->rs_conn_addr)) + return -ENOTCONN; + + if (ipv6_addr_v4mapped(&rs->rs_conn_addr)) { + sin = (struct sockaddr_in *)uaddr; + memset(sin->sin_zero, 0, sizeof(sin->sin_zero)); + sin->sin_family = AF_INET; + sin->sin_port = rs->rs_conn_port; + sin->sin_addr.s_addr = rs->rs_conn_addr_v4; + uaddr_len = sizeof(*sin); + } else { + sin6 = (struct sockaddr_in6 *)uaddr; + sin6->sin6_family = AF_INET6; + sin6->sin6_port = rs->rs_conn_port; + sin6->sin6_addr = rs->rs_conn_addr; + sin6->sin6_flowinfo = 0; + /* scope_id is the same as in the bound address. */ + sin6->sin6_scope_id = rs->rs_bound_scope_id; + uaddr_len = sizeof(*sin6); + } + } else { + /* If socket is not yet bound and the socket is connected, + * set the return address family to be the same as the + * connected address, but with 0 address value. If it is not + * connected, set the family to be AF_UNSPEC (value 0) and + * the address size to be that of an IPv4 address. + */ + if (ipv6_addr_any(&rs->rs_bound_addr)) { + if (ipv6_addr_any(&rs->rs_conn_addr)) { + sin = (struct sockaddr_in *)uaddr; + memset(sin, 0, sizeof(*sin)); + sin->sin_family = AF_UNSPEC; + return sizeof(*sin); + } + +#if IS_ENABLED(CONFIG_IPV6) + if (!(ipv6_addr_type(&rs->rs_conn_addr) & + IPV6_ADDR_MAPPED)) { + sin6 = (struct sockaddr_in6 *)uaddr; + memset(sin6, 0, sizeof(*sin6)); + sin6->sin6_family = AF_INET6; + return sizeof(*sin6); + } +#endif + + sin = (struct sockaddr_in *)uaddr; + memset(sin, 0, sizeof(*sin)); + sin->sin_family = AF_INET; + return sizeof(*sin); + } + if (ipv6_addr_v4mapped(&rs->rs_bound_addr)) { + sin = (struct sockaddr_in *)uaddr; + memset(sin->sin_zero, 0, sizeof(sin->sin_zero)); + sin->sin_family = AF_INET; + sin->sin_port = rs->rs_bound_port; + sin->sin_addr.s_addr = rs->rs_bound_addr_v4; + uaddr_len = sizeof(*sin); + } else { + sin6 = (struct sockaddr_in6 *)uaddr; + sin6->sin6_family = AF_INET6; + sin6->sin6_port = rs->rs_bound_port; + sin6->sin6_addr = rs->rs_bound_addr; + sin6->sin6_flowinfo = 0; + sin6->sin6_scope_id = rs->rs_bound_scope_id; + uaddr_len = sizeof(*sin6); + } + } + + return uaddr_len; +} + +/* + * RDS' poll is without a doubt the least intuitive part of the interface, + * as EPOLLIN and EPOLLOUT do not behave entirely as you would expect from + * a network protocol. + * + * EPOLLIN is asserted if + * - there is data on the receive queue. + * - to signal that a previously congested destination may have become + * uncongested + * - A notification has been queued to the socket (this can be a congestion + * update, or a RDMA completion, or a MSG_ZEROCOPY completion). + * + * EPOLLOUT is asserted if there is room on the send queue. This does not mean + * however, that the next sendmsg() call will succeed. If the application tries + * to send to a congested destination, the system call may still fail (and + * return ENOBUFS). + */ +static __poll_t rds_poll(struct file *file, struct socket *sock, + poll_table *wait) +{ + struct sock *sk = sock->sk; + struct rds_sock *rs = rds_sk_to_rs(sk); + __poll_t mask = 0; + unsigned long flags; + + poll_wait(file, sk_sleep(sk), wait); + + if (rs->rs_seen_congestion) + poll_wait(file, &rds_poll_waitq, wait); + + read_lock_irqsave(&rs->rs_recv_lock, flags); + if (!rs->rs_cong_monitor) { + /* When a congestion map was updated, we signal EPOLLIN for + * "historical" reasons. Applications can also poll for + * WRBAND instead. */ + if (rds_cong_updated_since(&rs->rs_cong_track)) + mask |= (EPOLLIN | EPOLLRDNORM | EPOLLWRBAND); + } else { + spin_lock(&rs->rs_lock); + if (rs->rs_cong_notify) + mask |= (EPOLLIN | EPOLLRDNORM); + spin_unlock(&rs->rs_lock); + } + if (!list_empty(&rs->rs_recv_queue) || + !list_empty(&rs->rs_notify_queue) || + !list_empty(&rs->rs_zcookie_queue.zcookie_head)) + mask |= (EPOLLIN | EPOLLRDNORM); + if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) + mask |= (EPOLLOUT | EPOLLWRNORM); + if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue)) + mask |= POLLERR; + read_unlock_irqrestore(&rs->rs_recv_lock, flags); + + /* clear state any time we wake a seen-congested socket */ + if (mask) + rs->rs_seen_congestion = 0; + + return mask; +} + +static int rds_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg) +{ + return -ENOIOCTLCMD; +} + +static int rds_cancel_sent_to(struct rds_sock *rs, char __user *optval, + int len) +{ + struct sockaddr_in6 sin6; + struct sockaddr_in sin; + int ret = 0; + + /* racing with another thread binding seems ok here */ + if (ipv6_addr_any(&rs->rs_bound_addr)) { + ret = -ENOTCONN; /* XXX not a great errno */ + goto out; + } + + if (len < sizeof(struct sockaddr_in)) { + ret = -EINVAL; + goto out; + } else if (len < sizeof(struct sockaddr_in6)) { + /* Assume IPv4 */ + if (copy_from_user(&sin, optval, sizeof(struct sockaddr_in))) { + ret = -EFAULT; + goto out; + } + ipv6_addr_set_v4mapped(sin.sin_addr.s_addr, &sin6.sin6_addr); + sin6.sin6_port = sin.sin_port; + } else { + if (copy_from_user(&sin6, optval, + sizeof(struct sockaddr_in6))) { + ret = -EFAULT; + goto out; + } + } + + rds_send_drop_to(rs, &sin6); +out: + return ret; +} + +static int rds_set_bool_option(unsigned char *optvar, char __user *optval, + int optlen) +{ + int value; + + if (optlen < sizeof(int)) + return -EINVAL; + if (get_user(value, (int __user *) optval)) + return -EFAULT; + *optvar = !!value; + return 0; +} + +static int rds_cong_monitor(struct rds_sock *rs, char __user *optval, + int optlen) +{ + int ret; + + ret = rds_set_bool_option(&rs->rs_cong_monitor, optval, optlen); + if (ret == 0) { + if (rs->rs_cong_monitor) { + rds_cong_add_socket(rs); + } else { + rds_cong_remove_socket(rs); + rs->rs_cong_mask = 0; + rs->rs_cong_notify = 0; + } + } + return ret; +} + +static int rds_set_transport(struct rds_sock *rs, char __user *optval, + int optlen) +{ + int t_type; + + if (rs->rs_transport) + return -EOPNOTSUPP; /* previously attached to transport */ + + if (optlen != sizeof(int)) + return -EINVAL; + + if (copy_from_user(&t_type, (int __user *)optval, sizeof(t_type))) + return -EFAULT; + + if (t_type < 0 || t_type >= RDS_TRANS_COUNT) + return -EINVAL; + + rs->rs_transport = rds_trans_get(t_type); + + return rs->rs_transport ? 0 : -ENOPROTOOPT; +} + +static int rds_enable_recvtstamp(struct sock *sk, char __user *optval, + int optlen) +{ + int val, valbool; + + if (optlen != sizeof(int)) + return -EFAULT; + + if (get_user(val, (int __user *)optval)) + return -EFAULT; + + valbool = val ? 1 : 0; + + if (valbool) + sock_set_flag(sk, SOCK_RCVTSTAMP); + else + sock_reset_flag(sk, SOCK_RCVTSTAMP); + + return 0; +} + +static int rds_recv_track_latency(struct rds_sock *rs, char __user *optval, + int optlen) +{ + struct rds_rx_trace_so trace; + int i; + + if (optlen != sizeof(struct rds_rx_trace_so)) + return -EFAULT; + + if (copy_from_user(&trace, optval, sizeof(trace))) + return -EFAULT; + + if (trace.rx_traces > RDS_MSG_RX_DGRAM_TRACE_MAX) + return -EFAULT; + + rs->rs_rx_traces = trace.rx_traces; + for (i = 0; i < rs->rs_rx_traces; i++) { + if (trace.rx_trace_pos[i] > RDS_MSG_RX_DGRAM_TRACE_MAX) { + rs->rs_rx_traces = 0; + return -EFAULT; + } + rs->rs_rx_trace[i] = trace.rx_trace_pos[i]; + } + + return 0; +} + +static int rds_setsockopt(struct socket *sock, int level, int optname, + char __user *optval, unsigned int optlen) +{ + struct rds_sock *rs = rds_sk_to_rs(sock->sk); + int ret; + + if (level != SOL_RDS) { + ret = -ENOPROTOOPT; + goto out; + } + + switch (optname) { + case RDS_CANCEL_SENT_TO: + ret = rds_cancel_sent_to(rs, optval, optlen); + break; + case RDS_GET_MR: + ret = rds_get_mr(rs, optval, optlen); + break; + case RDS_GET_MR_FOR_DEST: + ret = rds_get_mr_for_dest(rs, optval, optlen); + break; + case RDS_FREE_MR: + ret = rds_free_mr(rs, optval, optlen); + break; + case RDS_RECVERR: + ret = rds_set_bool_option(&rs->rs_recverr, optval, optlen); + break; + case RDS_CONG_MONITOR: + ret = rds_cong_monitor(rs, optval, optlen); + break; + case SO_RDS_TRANSPORT: + lock_sock(sock->sk); + ret = rds_set_transport(rs, optval, optlen); + release_sock(sock->sk); + break; + case SO_TIMESTAMP: + lock_sock(sock->sk); + ret = rds_enable_recvtstamp(sock->sk, optval, optlen); + release_sock(sock->sk); + break; + case SO_RDS_MSG_RXPATH_LATENCY: + ret = rds_recv_track_latency(rs, optval, optlen); + break; + default: + ret = -ENOPROTOOPT; + } +out: + return ret; +} + +static int rds_getsockopt(struct socket *sock, int level, int optname, + char __user *optval, int __user *optlen) +{ + struct rds_sock *rs = rds_sk_to_rs(sock->sk); + int ret = -ENOPROTOOPT, len; + int trans; + + if (level != SOL_RDS) + goto out; + + if (get_user(len, optlen)) { + ret = -EFAULT; + goto out; + } + + switch (optname) { + case RDS_INFO_FIRST ... RDS_INFO_LAST: + ret = rds_info_getsockopt(sock, optname, optval, + optlen); + break; + + case RDS_RECVERR: + if (len < sizeof(int)) + ret = -EINVAL; + else + if (put_user(rs->rs_recverr, (int __user *) optval) || + put_user(sizeof(int), optlen)) + ret = -EFAULT; + else + ret = 0; + break; + case SO_RDS_TRANSPORT: + if (len < sizeof(int)) { + ret = -EINVAL; + break; + } + trans = (rs->rs_transport ? rs->rs_transport->t_type : + RDS_TRANS_NONE); /* unbound */ + if (put_user(trans, (int __user *)optval) || + put_user(sizeof(int), optlen)) + ret = -EFAULT; + else + ret = 0; + break; + default: + break; + } + +out: + return ret; + +} + +static int rds_connect(struct socket *sock, struct sockaddr *uaddr, + int addr_len, int flags) +{ + struct sock *sk = sock->sk; + struct sockaddr_in *sin; + struct rds_sock *rs = rds_sk_to_rs(sk); + int ret = 0; + + if (addr_len < offsetofend(struct sockaddr, sa_family)) + return -EINVAL; + + lock_sock(sk); + + switch (uaddr->sa_family) { + case AF_INET: + sin = (struct sockaddr_in *)uaddr; + if (addr_len < sizeof(struct sockaddr_in)) { + ret = -EINVAL; + break; + } + if (sin->sin_addr.s_addr == htonl(INADDR_ANY)) { + ret = -EDESTADDRREQ; + break; + } + if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr)) || + sin->sin_addr.s_addr == htonl(INADDR_BROADCAST)) { + ret = -EINVAL; + break; + } + ipv6_addr_set_v4mapped(sin->sin_addr.s_addr, &rs->rs_conn_addr); + rs->rs_conn_port = sin->sin_port; + break; + +#if IS_ENABLED(CONFIG_IPV6) + case AF_INET6: { + struct sockaddr_in6 *sin6; + int addr_type; + + sin6 = (struct sockaddr_in6 *)uaddr; + if (addr_len < sizeof(struct sockaddr_in6)) { + ret = -EINVAL; + break; + } + addr_type = ipv6_addr_type(&sin6->sin6_addr); + if (!(addr_type & IPV6_ADDR_UNICAST)) { + __be32 addr4; + + if (!(addr_type & IPV6_ADDR_MAPPED)) { + ret = -EPROTOTYPE; + break; + } + + /* It is a mapped address. Need to do some sanity + * checks. + */ + addr4 = sin6->sin6_addr.s6_addr32[3]; + if (addr4 == htonl(INADDR_ANY) || + addr4 == htonl(INADDR_BROADCAST) || + IN_MULTICAST(ntohl(addr4))) { + ret = -EPROTOTYPE; + break; + } + } + + if (addr_type & IPV6_ADDR_LINKLOCAL) { + /* If socket is arleady bound to a link local address, + * the peer address must be on the same link. + */ + if (sin6->sin6_scope_id == 0 || + (!ipv6_addr_any(&rs->rs_bound_addr) && + rs->rs_bound_scope_id && + sin6->sin6_scope_id != rs->rs_bound_scope_id)) { + ret = -EINVAL; + break; + } + /* Remember the connected address scope ID. It will + * be checked against the binding local address when + * the socket is bound. + */ + rs->rs_bound_scope_id = sin6->sin6_scope_id; + } + rs->rs_conn_addr = sin6->sin6_addr; + rs->rs_conn_port = sin6->sin6_port; + break; + } +#endif + + default: + ret = -EAFNOSUPPORT; + break; + } + + release_sock(sk); + return ret; +} + +static struct proto rds_proto = { + .name = "RDS", + .owner = THIS_MODULE, + .obj_size = sizeof(struct rds_sock), +}; + +static const struct proto_ops rds_proto_ops = { + .family = AF_RDS, + .owner = THIS_MODULE, + .release = rds_release, + .bind = rds_bind, + .connect = rds_connect, + .socketpair = sock_no_socketpair, + .accept = sock_no_accept, + .getname = rds_getname, + .poll = rds_poll, + .ioctl = rds_ioctl, + .listen = sock_no_listen, + .shutdown = sock_no_shutdown, + .setsockopt = rds_setsockopt, + .getsockopt = rds_getsockopt, + .sendmsg = rds_sendmsg, + .recvmsg = rds_recvmsg, + .mmap = sock_no_mmap, + .sendpage = sock_no_sendpage, +}; + +static void rds_sock_destruct(struct sock *sk) +{ + struct rds_sock *rs = rds_sk_to_rs(sk); + + WARN_ON((&rs->rs_item != rs->rs_item.next || + &rs->rs_item != rs->rs_item.prev)); +} + +static int __rds_create(struct socket *sock, struct sock *sk, int protocol) +{ + struct rds_sock *rs; + + sock_init_data(sock, sk); + sock->ops = &rds_proto_ops; + sk->sk_protocol = protocol; + sk->sk_destruct = rds_sock_destruct; + + rs = rds_sk_to_rs(sk); + spin_lock_init(&rs->rs_lock); + rwlock_init(&rs->rs_recv_lock); + INIT_LIST_HEAD(&rs->rs_send_queue); + INIT_LIST_HEAD(&rs->rs_recv_queue); + INIT_LIST_HEAD(&rs->rs_notify_queue); + INIT_LIST_HEAD(&rs->rs_cong_list); + rds_message_zcopy_queue_init(&rs->rs_zcookie_queue); + spin_lock_init(&rs->rs_rdma_lock); + rs->rs_rdma_keys = RB_ROOT; + rs->rs_rx_traces = 0; + + spin_lock_bh(&rds_sock_lock); + list_add_tail(&rs->rs_item, &rds_sock_list); + rds_sock_count++; + spin_unlock_bh(&rds_sock_lock); + + return 0; +} + +static int rds_create(struct net *net, struct socket *sock, int protocol, + int kern) +{ + struct sock *sk; + + if (sock->type != SOCK_SEQPACKET || protocol) + return -ESOCKTNOSUPPORT; + + sk = sk_alloc(net, AF_RDS, GFP_ATOMIC, &rds_proto, kern); + if (!sk) + return -ENOMEM; + + return __rds_create(sock, sk, protocol); +} + +void rds_sock_addref(struct rds_sock *rs) +{ + sock_hold(rds_rs_to_sk(rs)); +} + +void rds_sock_put(struct rds_sock *rs) +{ + sock_put(rds_rs_to_sk(rs)); +} + +static const struct net_proto_family rds_family_ops = { + .family = AF_RDS, + .create = rds_create, + .owner = THIS_MODULE, +}; + +static void rds_sock_inc_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + struct rds_sock *rs; + struct rds_incoming *inc; + unsigned int total = 0; + + len /= sizeof(struct rds_info_message); + + spin_lock_bh(&rds_sock_lock); + + list_for_each_entry(rs, &rds_sock_list, rs_item) { + read_lock(&rs->rs_recv_lock); + + /* XXX too lazy to maintain counts.. */ + list_for_each_entry(inc, &rs->rs_recv_queue, i_item) { + total++; + if (total <= len) + rds_inc_info_copy(inc, iter, + inc->i_saddr.s6_addr32[3], + rs->rs_bound_addr_v4, + 1); + } + + read_unlock(&rs->rs_recv_lock); + } + + spin_unlock_bh(&rds_sock_lock); + + lens->nr = total; + lens->each = sizeof(struct rds_info_message); +} + +static void rds_sock_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + struct rds_info_socket sinfo; + struct rds_sock *rs; + + len /= sizeof(struct rds_info_socket); + + spin_lock_bh(&rds_sock_lock); + + if (len < rds_sock_count) + goto out; + + list_for_each_entry(rs, &rds_sock_list, rs_item) { + sinfo.sndbuf = rds_sk_sndbuf(rs); + sinfo.rcvbuf = rds_sk_rcvbuf(rs); + sinfo.bound_addr = rs->rs_bound_addr_v4; + sinfo.connected_addr = rs->rs_conn_addr_v4; + sinfo.bound_port = rs->rs_bound_port; + sinfo.connected_port = rs->rs_conn_port; + sinfo.inum = sock_i_ino(rds_rs_to_sk(rs)); + + rds_info_copy(iter, &sinfo, sizeof(sinfo)); + } + +out: + lens->nr = rds_sock_count; + lens->each = sizeof(struct rds_info_socket); + + spin_unlock_bh(&rds_sock_lock); +} + +static void rds_exit(void) +{ + sock_unregister(rds_family_ops.family); + proto_unregister(&rds_proto); + rds_conn_exit(); + rds_cong_exit(); + rds_sysctl_exit(); + rds_threads_exit(); + rds_stats_exit(); + rds_page_exit(); + rds_bind_lock_destroy(); + rds_info_deregister_func(RDS_INFO_SOCKETS, rds_sock_info); + rds_info_deregister_func(RDS_INFO_RECV_MESSAGES, rds_sock_inc_info); +} +module_exit(rds_exit); + +u32 rds_gen_num; + +static int rds_init(void) +{ + int ret; + + net_get_random_once(&rds_gen_num, sizeof(rds_gen_num)); + + ret = rds_bind_lock_init(); + if (ret) + goto out; + + ret = rds_conn_init(); + if (ret) + goto out_bind; + + ret = rds_threads_init(); + if (ret) + goto out_conn; + ret = rds_sysctl_init(); + if (ret) + goto out_threads; + ret = rds_stats_init(); + if (ret) + goto out_sysctl; + ret = proto_register(&rds_proto, 1); + if (ret) + goto out_stats; + ret = sock_register(&rds_family_ops); + if (ret) + goto out_proto; + + rds_info_register_func(RDS_INFO_SOCKETS, rds_sock_info); + rds_info_register_func(RDS_INFO_RECV_MESSAGES, rds_sock_inc_info); + + goto out; + +out_proto: + proto_unregister(&rds_proto); +out_stats: + rds_stats_exit(); +out_sysctl: + rds_sysctl_exit(); +out_threads: + rds_threads_exit(); +out_conn: + rds_conn_exit(); + rds_cong_exit(); + rds_page_exit(); +out_bind: + rds_bind_lock_destroy(); +out: + return ret; +} +module_init(rds_init); + +#define DRV_VERSION "4.0" +#define DRV_RELDATE "Feb 12, 2009" + +MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>"); +MODULE_DESCRIPTION("RDS: Reliable Datagram Sockets" + " v" DRV_VERSION " (" DRV_RELDATE ")"); +MODULE_VERSION(DRV_VERSION); +MODULE_LICENSE("Dual BSD/GPL"); +MODULE_ALIAS_NETPROTO(PF_RDS); diff --git a/net/rds/bind.c b/net/rds/bind.c new file mode 100644 index 000000000..93e336535 --- /dev/null +++ b/net/rds/bind.c @@ -0,0 +1,283 @@ +/* + * Copyright (c) 2006, 2019 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <net/sock.h> +#include <linux/in.h> +#include <linux/ipv6.h> +#include <linux/if_arp.h> +#include <linux/jhash.h> +#include <linux/ratelimit.h> +#include "rds.h" + +static struct rhashtable bind_hash_table; + +static const struct rhashtable_params ht_parms = { + .nelem_hint = 768, + .key_len = RDS_BOUND_KEY_LEN, + .key_offset = offsetof(struct rds_sock, rs_bound_key), + .head_offset = offsetof(struct rds_sock, rs_bound_node), + .max_size = 16384, + .min_size = 1024, +}; + +/* Create a key for the bind hash table manipulation. Port is in network byte + * order. + */ +static inline void __rds_create_bind_key(u8 *key, const struct in6_addr *addr, + __be16 port, __u32 scope_id) +{ + memcpy(key, addr, sizeof(*addr)); + key += sizeof(*addr); + memcpy(key, &port, sizeof(port)); + key += sizeof(port); + memcpy(key, &scope_id, sizeof(scope_id)); +} + +/* + * Return the rds_sock bound at the given local address. + * + * The rx path can race with rds_release. We notice if rds_release() has + * marked this socket and don't return a rs ref to the rx path. + */ +struct rds_sock *rds_find_bound(const struct in6_addr *addr, __be16 port, + __u32 scope_id) +{ + u8 key[RDS_BOUND_KEY_LEN]; + struct rds_sock *rs; + + __rds_create_bind_key(key, addr, port, scope_id); + rcu_read_lock(); + rs = rhashtable_lookup(&bind_hash_table, key, ht_parms); + if (rs && (sock_flag(rds_rs_to_sk(rs), SOCK_DEAD) || + !refcount_inc_not_zero(&rds_rs_to_sk(rs)->sk_refcnt))) + rs = NULL; + + rcu_read_unlock(); + + rdsdebug("returning rs %p for %pI6c:%u\n", rs, addr, + ntohs(port)); + + return rs; +} + +/* returns -ve errno or +ve port */ +static int rds_add_bound(struct rds_sock *rs, const struct in6_addr *addr, + __be16 *port, __u32 scope_id) +{ + int ret = -EADDRINUSE; + u16 rover, last; + u8 key[RDS_BOUND_KEY_LEN]; + + if (*port != 0) { + rover = be16_to_cpu(*port); + if (rover == RDS_FLAG_PROBE_PORT) + return -EINVAL; + last = rover; + } else { + rover = max_t(u16, prandom_u32(), 2); + last = rover - 1; + } + + do { + if (rover == 0) + rover++; + + if (rover == RDS_FLAG_PROBE_PORT) + continue; + __rds_create_bind_key(key, addr, cpu_to_be16(rover), + scope_id); + if (rhashtable_lookup_fast(&bind_hash_table, key, ht_parms)) + continue; + + memcpy(rs->rs_bound_key, key, sizeof(rs->rs_bound_key)); + rs->rs_bound_addr = *addr; + net_get_random_once(&rs->rs_hash_initval, + sizeof(rs->rs_hash_initval)); + rs->rs_bound_port = cpu_to_be16(rover); + rs->rs_bound_node.next = NULL; + rds_sock_addref(rs); + if (!rhashtable_insert_fast(&bind_hash_table, + &rs->rs_bound_node, ht_parms)) { + *port = rs->rs_bound_port; + rs->rs_bound_scope_id = scope_id; + ret = 0; + rdsdebug("rs %p binding to %pI6c:%d\n", + rs, addr, (int)ntohs(*port)); + break; + } else { + rs->rs_bound_addr = in6addr_any; + rds_sock_put(rs); + ret = -ENOMEM; + break; + } + } while (rover++ != last); + + return ret; +} + +void rds_remove_bound(struct rds_sock *rs) +{ + + if (ipv6_addr_any(&rs->rs_bound_addr)) + return; + + rdsdebug("rs %p unbinding from %pI6c:%d\n", + rs, &rs->rs_bound_addr, + ntohs(rs->rs_bound_port)); + + rhashtable_remove_fast(&bind_hash_table, &rs->rs_bound_node, ht_parms); + rds_sock_put(rs); + rs->rs_bound_addr = in6addr_any; +} + +int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len) +{ + struct sock *sk = sock->sk; + struct rds_sock *rs = rds_sk_to_rs(sk); + struct in6_addr v6addr, *binding_addr; + struct rds_transport *trans; + __u32 scope_id = 0; + int ret = 0; + __be16 port; + + /* We allow an RDS socket to be bound to either IPv4 or IPv6 + * address. + */ + if (addr_len < offsetofend(struct sockaddr, sa_family)) + return -EINVAL; + if (uaddr->sa_family == AF_INET) { + struct sockaddr_in *sin = (struct sockaddr_in *)uaddr; + + if (addr_len < sizeof(struct sockaddr_in) || + sin->sin_addr.s_addr == htonl(INADDR_ANY) || + sin->sin_addr.s_addr == htonl(INADDR_BROADCAST) || + IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) + return -EINVAL; + ipv6_addr_set_v4mapped(sin->sin_addr.s_addr, &v6addr); + binding_addr = &v6addr; + port = sin->sin_port; +#if IS_ENABLED(CONFIG_IPV6) + } else if (uaddr->sa_family == AF_INET6) { + struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)uaddr; + int addr_type; + + if (addr_len < sizeof(struct sockaddr_in6)) + return -EINVAL; + addr_type = ipv6_addr_type(&sin6->sin6_addr); + if (!(addr_type & IPV6_ADDR_UNICAST)) { + __be32 addr4; + + if (!(addr_type & IPV6_ADDR_MAPPED)) + return -EINVAL; + + /* It is a mapped address. Need to do some sanity + * checks. + */ + addr4 = sin6->sin6_addr.s6_addr32[3]; + if (addr4 == htonl(INADDR_ANY) || + addr4 == htonl(INADDR_BROADCAST) || + IN_MULTICAST(ntohl(addr4))) + return -EINVAL; + } + /* The scope ID must be specified for link local address. */ + if (addr_type & IPV6_ADDR_LINKLOCAL) { + if (sin6->sin6_scope_id == 0) + return -EINVAL; + scope_id = sin6->sin6_scope_id; + } + binding_addr = &sin6->sin6_addr; + port = sin6->sin6_port; +#endif + } else { + return -EINVAL; + } + lock_sock(sk); + + /* RDS socket does not allow re-binding. */ + if (!ipv6_addr_any(&rs->rs_bound_addr)) { + ret = -EINVAL; + goto out; + } + /* Socket is connected. The binding address should have the same + * scope ID as the connected address, except the case when one is + * non-link local address (scope_id is 0). + */ + if (!ipv6_addr_any(&rs->rs_conn_addr) && scope_id && + rs->rs_bound_scope_id && + scope_id != rs->rs_bound_scope_id) { + ret = -EINVAL; + goto out; + } + + /* The transport can be set using SO_RDS_TRANSPORT option before the + * socket is bound. + */ + if (rs->rs_transport) { + trans = rs->rs_transport; + if (!trans->laddr_check || + trans->laddr_check(sock_net(sock->sk), + binding_addr, scope_id) != 0) { + ret = -ENOPROTOOPT; + goto out; + } + } else { + trans = rds_trans_get_preferred(sock_net(sock->sk), + binding_addr, scope_id); + if (!trans) { + ret = -EADDRNOTAVAIL; + pr_info_ratelimited("RDS: %s could not find a transport for %pI6c, load rds_tcp or rds_rdma?\n", + __func__, binding_addr); + goto out; + } + rs->rs_transport = trans; + } + + sock_set_flag(sk, SOCK_RCU_FREE); + ret = rds_add_bound(rs, binding_addr, &port, scope_id); + if (ret) + rs->rs_transport = NULL; + +out: + release_sock(sk); + return ret; +} + +void rds_bind_lock_destroy(void) +{ + rhashtable_destroy(&bind_hash_table); +} + +int rds_bind_lock_init(void) +{ + return rhashtable_init(&bind_hash_table, &ht_parms); +} diff --git a/net/rds/cong.c b/net/rds/cong.c new file mode 100644 index 000000000..ccdff09a7 --- /dev/null +++ b/net/rds/cong.c @@ -0,0 +1,428 @@ +/* + * Copyright (c) 2007, 2017 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/slab.h> +#include <linux/types.h> +#include <linux/rbtree.h> +#include <linux/bitops.h> +#include <linux/export.h> + +#include "rds.h" + +/* + * This file implements the receive side of the unconventional congestion + * management in RDS. + * + * Messages waiting in the receive queue on the receiving socket are accounted + * against the sockets SO_RCVBUF option value. Only the payload bytes in the + * message are accounted for. If the number of bytes queued equals or exceeds + * rcvbuf then the socket is congested. All sends attempted to this socket's + * address should return block or return -EWOULDBLOCK. + * + * Applications are expected to be reasonably tuned such that this situation + * very rarely occurs. An application encountering this "back-pressure" is + * considered a bug. + * + * This is implemented by having each node maintain bitmaps which indicate + * which ports on bound addresses are congested. As the bitmap changes it is + * sent through all the connections which terminate in the local address of the + * bitmap which changed. + * + * The bitmaps are allocated as connections are brought up. This avoids + * allocation in the interrupt handling path which queues messages on sockets. + * The dense bitmaps let transports send the entire bitmap on any bitmap change + * reasonably efficiently. This is much easier to implement than some + * finer-grained communication of per-port congestion. The sender does a very + * inexpensive bit test to test if the port it's about to send to is congested + * or not. + */ + +/* + * Interaction with poll is a tad tricky. We want all processes stuck in + * poll to wake up and check whether a congested destination became uncongested. + * The really sad thing is we have no idea which destinations the application + * wants to send to - we don't even know which rds_connections are involved. + * So until we implement a more flexible rds poll interface, we have to make + * do with this: + * We maintain a global counter that is incremented each time a congestion map + * update is received. Each rds socket tracks this value, and if rds_poll + * finds that the saved generation number is smaller than the global generation + * number, it wakes up the process. + */ +static atomic_t rds_cong_generation = ATOMIC_INIT(0); + +/* + * Congestion monitoring + */ +static LIST_HEAD(rds_cong_monitor); +static DEFINE_RWLOCK(rds_cong_monitor_lock); + +/* + * Yes, a global lock. It's used so infrequently that it's worth keeping it + * global to simplify the locking. It's only used in the following + * circumstances: + * + * - on connection buildup to associate a conn with its maps + * - on map changes to inform conns of a new map to send + * + * It's sadly ordered under the socket callback lock and the connection lock. + * Receive paths can mark ports congested from interrupt context so the + * lock masks interrupts. + */ +static DEFINE_SPINLOCK(rds_cong_lock); +static struct rb_root rds_cong_tree = RB_ROOT; + +static struct rds_cong_map *rds_cong_tree_walk(const struct in6_addr *addr, + struct rds_cong_map *insert) +{ + struct rb_node **p = &rds_cong_tree.rb_node; + struct rb_node *parent = NULL; + struct rds_cong_map *map; + + while (*p) { + int diff; + + parent = *p; + map = rb_entry(parent, struct rds_cong_map, m_rb_node); + + diff = rds_addr_cmp(addr, &map->m_addr); + if (diff < 0) + p = &(*p)->rb_left; + else if (diff > 0) + p = &(*p)->rb_right; + else + return map; + } + + if (insert) { + rb_link_node(&insert->m_rb_node, parent, p); + rb_insert_color(&insert->m_rb_node, &rds_cong_tree); + } + return NULL; +} + +/* + * There is only ever one bitmap for any address. Connections try and allocate + * these bitmaps in the process getting pointers to them. The bitmaps are only + * ever freed as the module is removed after all connections have been freed. + */ +static struct rds_cong_map *rds_cong_from_addr(const struct in6_addr *addr) +{ + struct rds_cong_map *map; + struct rds_cong_map *ret = NULL; + unsigned long zp; + unsigned long i; + unsigned long flags; + + map = kzalloc(sizeof(struct rds_cong_map), GFP_KERNEL); + if (!map) + return NULL; + + map->m_addr = *addr; + init_waitqueue_head(&map->m_waitq); + INIT_LIST_HEAD(&map->m_conn_list); + + for (i = 0; i < RDS_CONG_MAP_PAGES; i++) { + zp = get_zeroed_page(GFP_KERNEL); + if (zp == 0) + goto out; + map->m_page_addrs[i] = zp; + } + + spin_lock_irqsave(&rds_cong_lock, flags); + ret = rds_cong_tree_walk(addr, map); + spin_unlock_irqrestore(&rds_cong_lock, flags); + + if (!ret) { + ret = map; + map = NULL; + } + +out: + if (map) { + for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++) + free_page(map->m_page_addrs[i]); + kfree(map); + } + + rdsdebug("map %p for addr %pI6c\n", ret, addr); + + return ret; +} + +/* + * Put the conn on its local map's list. This is called when the conn is + * really added to the hash. It's nested under the rds_conn_lock, sadly. + */ +void rds_cong_add_conn(struct rds_connection *conn) +{ + unsigned long flags; + + rdsdebug("conn %p now on map %p\n", conn, conn->c_lcong); + spin_lock_irqsave(&rds_cong_lock, flags); + list_add_tail(&conn->c_map_item, &conn->c_lcong->m_conn_list); + spin_unlock_irqrestore(&rds_cong_lock, flags); +} + +void rds_cong_remove_conn(struct rds_connection *conn) +{ + unsigned long flags; + + rdsdebug("removing conn %p from map %p\n", conn, conn->c_lcong); + spin_lock_irqsave(&rds_cong_lock, flags); + list_del_init(&conn->c_map_item); + spin_unlock_irqrestore(&rds_cong_lock, flags); +} + +int rds_cong_get_maps(struct rds_connection *conn) +{ + conn->c_lcong = rds_cong_from_addr(&conn->c_laddr); + conn->c_fcong = rds_cong_from_addr(&conn->c_faddr); + + if (!(conn->c_lcong && conn->c_fcong)) + return -ENOMEM; + + return 0; +} + +void rds_cong_queue_updates(struct rds_cong_map *map) +{ + struct rds_connection *conn; + unsigned long flags; + + spin_lock_irqsave(&rds_cong_lock, flags); + + list_for_each_entry(conn, &map->m_conn_list, c_map_item) { + struct rds_conn_path *cp = &conn->c_path[0]; + + rcu_read_lock(); + if (!test_and_set_bit(0, &conn->c_map_queued) && + !rds_destroy_pending(cp->cp_conn)) { + rds_stats_inc(s_cong_update_queued); + /* We cannot inline the call to rds_send_xmit() here + * for two reasons (both pertaining to a TCP transport): + * 1. When we get here from the receive path, we + * are already holding the sock_lock (held by + * tcp_v4_rcv()). So inlining calls to + * tcp_setsockopt and/or tcp_sendmsg will deadlock + * when it tries to get the sock_lock()) + * 2. Interrupts are masked so that we can mark the + * the port congested from both send and recv paths. + * (See comment around declaration of rdc_cong_lock). + * An attempt to get the sock_lock() here will + * therefore trigger warnings. + * Defer the xmit to rds_send_worker() instead. + */ + queue_delayed_work(rds_wq, &cp->cp_send_w, 0); + } + rcu_read_unlock(); + } + + spin_unlock_irqrestore(&rds_cong_lock, flags); +} + +void rds_cong_map_updated(struct rds_cong_map *map, uint64_t portmask) +{ + rdsdebug("waking map %p for %pI4\n", + map, &map->m_addr); + rds_stats_inc(s_cong_update_received); + atomic_inc(&rds_cong_generation); + if (waitqueue_active(&map->m_waitq)) + wake_up(&map->m_waitq); + if (waitqueue_active(&rds_poll_waitq)) + wake_up_all(&rds_poll_waitq); + + if (portmask && !list_empty(&rds_cong_monitor)) { + unsigned long flags; + struct rds_sock *rs; + + read_lock_irqsave(&rds_cong_monitor_lock, flags); + list_for_each_entry(rs, &rds_cong_monitor, rs_cong_list) { + spin_lock(&rs->rs_lock); + rs->rs_cong_notify |= (rs->rs_cong_mask & portmask); + rs->rs_cong_mask &= ~portmask; + spin_unlock(&rs->rs_lock); + if (rs->rs_cong_notify) + rds_wake_sk_sleep(rs); + } + read_unlock_irqrestore(&rds_cong_monitor_lock, flags); + } +} +EXPORT_SYMBOL_GPL(rds_cong_map_updated); + +int rds_cong_updated_since(unsigned long *recent) +{ + unsigned long gen = atomic_read(&rds_cong_generation); + + if (likely(*recent == gen)) + return 0; + *recent = gen; + return 1; +} + +/* + * We're called under the locking that protects the sockets receive buffer + * consumption. This makes it a lot easier for the caller to only call us + * when it knows that an existing set bit needs to be cleared, and vice versa. + * We can't block and we need to deal with concurrent sockets working against + * the same per-address map. + */ +void rds_cong_set_bit(struct rds_cong_map *map, __be16 port) +{ + unsigned long i; + unsigned long off; + + rdsdebug("setting congestion for %pI4:%u in map %p\n", + &map->m_addr, ntohs(port), map); + + i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS; + off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS; + + set_bit_le(off, (void *)map->m_page_addrs[i]); +} + +void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port) +{ + unsigned long i; + unsigned long off; + + rdsdebug("clearing congestion for %pI4:%u in map %p\n", + &map->m_addr, ntohs(port), map); + + i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS; + off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS; + + clear_bit_le(off, (void *)map->m_page_addrs[i]); +} + +static int rds_cong_test_bit(struct rds_cong_map *map, __be16 port) +{ + unsigned long i; + unsigned long off; + + i = be16_to_cpu(port) / RDS_CONG_MAP_PAGE_BITS; + off = be16_to_cpu(port) % RDS_CONG_MAP_PAGE_BITS; + + return test_bit_le(off, (void *)map->m_page_addrs[i]); +} + +void rds_cong_add_socket(struct rds_sock *rs) +{ + unsigned long flags; + + write_lock_irqsave(&rds_cong_monitor_lock, flags); + if (list_empty(&rs->rs_cong_list)) + list_add(&rs->rs_cong_list, &rds_cong_monitor); + write_unlock_irqrestore(&rds_cong_monitor_lock, flags); +} + +void rds_cong_remove_socket(struct rds_sock *rs) +{ + unsigned long flags; + struct rds_cong_map *map; + + write_lock_irqsave(&rds_cong_monitor_lock, flags); + list_del_init(&rs->rs_cong_list); + write_unlock_irqrestore(&rds_cong_monitor_lock, flags); + + /* update congestion map for now-closed port */ + spin_lock_irqsave(&rds_cong_lock, flags); + map = rds_cong_tree_walk(&rs->rs_bound_addr, NULL); + spin_unlock_irqrestore(&rds_cong_lock, flags); + + if (map && rds_cong_test_bit(map, rs->rs_bound_port)) { + rds_cong_clear_bit(map, rs->rs_bound_port); + rds_cong_queue_updates(map); + } +} + +int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, + struct rds_sock *rs) +{ + if (!rds_cong_test_bit(map, port)) + return 0; + if (nonblock) { + if (rs && rs->rs_cong_monitor) { + unsigned long flags; + + /* It would have been nice to have an atomic set_bit on + * a uint64_t. */ + spin_lock_irqsave(&rs->rs_lock, flags); + rs->rs_cong_mask |= RDS_CONG_MONITOR_MASK(ntohs(port)); + spin_unlock_irqrestore(&rs->rs_lock, flags); + + /* Test again - a congestion update may have arrived in + * the meantime. */ + if (!rds_cong_test_bit(map, port)) + return 0; + } + rds_stats_inc(s_cong_send_error); + return -ENOBUFS; + } + + rds_stats_inc(s_cong_send_blocked); + rdsdebug("waiting on map %p for port %u\n", map, be16_to_cpu(port)); + + return wait_event_interruptible(map->m_waitq, + !rds_cong_test_bit(map, port)); +} + +void rds_cong_exit(void) +{ + struct rb_node *node; + struct rds_cong_map *map; + unsigned long i; + + while ((node = rb_first(&rds_cong_tree))) { + map = rb_entry(node, struct rds_cong_map, m_rb_node); + rdsdebug("freeing map %p\n", map); + rb_erase(&map->m_rb_node, &rds_cong_tree); + for (i = 0; i < RDS_CONG_MAP_PAGES && map->m_page_addrs[i]; i++) + free_page(map->m_page_addrs[i]); + kfree(map); + } +} + +/* + * Allocate a RDS message containing a congestion update. + */ +struct rds_message *rds_cong_update_alloc(struct rds_connection *conn) +{ + struct rds_cong_map *map = conn->c_lcong; + struct rds_message *rm; + + rm = rds_message_map_pages(map->m_page_addrs, RDS_CONG_MAP_BYTES); + if (!IS_ERR(rm)) + rm->m_inc.i_hdr.h_flags = RDS_FLAG_CONG_BITMAP; + + return rm; +} diff --git a/net/rds/connection.c b/net/rds/connection.c new file mode 100644 index 000000000..af9f7d184 --- /dev/null +++ b/net/rds/connection.c @@ -0,0 +1,933 @@ +/* + * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/list.h> +#include <linux/slab.h> +#include <linux/export.h> +#include <net/ipv6.h> +#include <net/inet6_hashtables.h> +#include <net/addrconf.h> + +#include "rds.h" +#include "loop.h" + +#define RDS_CONNECTION_HASH_BITS 12 +#define RDS_CONNECTION_HASH_ENTRIES (1 << RDS_CONNECTION_HASH_BITS) +#define RDS_CONNECTION_HASH_MASK (RDS_CONNECTION_HASH_ENTRIES - 1) + +/* converting this to RCU is a chore for another day.. */ +static DEFINE_SPINLOCK(rds_conn_lock); +static unsigned long rds_conn_count; +static struct hlist_head rds_conn_hash[RDS_CONNECTION_HASH_ENTRIES]; +static struct kmem_cache *rds_conn_slab; + +static struct hlist_head *rds_conn_bucket(const struct in6_addr *laddr, + const struct in6_addr *faddr) +{ + static u32 rds6_hash_secret __read_mostly; + static u32 rds_hash_secret __read_mostly; + + u32 lhash, fhash, hash; + + net_get_random_once(&rds_hash_secret, sizeof(rds_hash_secret)); + net_get_random_once(&rds6_hash_secret, sizeof(rds6_hash_secret)); + + lhash = (__force u32)laddr->s6_addr32[3]; +#if IS_ENABLED(CONFIG_IPV6) + fhash = __ipv6_addr_jhash(faddr, rds6_hash_secret); +#else + fhash = (__force u32)faddr->s6_addr32[3]; +#endif + hash = __inet_ehashfn(lhash, 0, fhash, 0, rds_hash_secret); + + return &rds_conn_hash[hash & RDS_CONNECTION_HASH_MASK]; +} + +#define rds_conn_info_set(var, test, suffix) do { \ + if (test) \ + var |= RDS_INFO_CONNECTION_FLAG_##suffix; \ +} while (0) + +/* rcu read lock must be held or the connection spinlock */ +static struct rds_connection *rds_conn_lookup(struct net *net, + struct hlist_head *head, + const struct in6_addr *laddr, + const struct in6_addr *faddr, + struct rds_transport *trans, + int dev_if) +{ + struct rds_connection *conn, *ret = NULL; + + hlist_for_each_entry_rcu(conn, head, c_hash_node) { + if (ipv6_addr_equal(&conn->c_faddr, faddr) && + ipv6_addr_equal(&conn->c_laddr, laddr) && + conn->c_trans == trans && + net == rds_conn_net(conn) && + conn->c_dev_if == dev_if) { + ret = conn; + break; + } + } + rdsdebug("returning conn %p for %pI6c -> %pI6c\n", ret, + laddr, faddr); + return ret; +} + +/* + * This is called by transports as they're bringing down a connection. + * It clears partial message state so that the transport can start sending + * and receiving over this connection again in the future. It is up to + * the transport to have serialized this call with its send and recv. + */ +static void rds_conn_path_reset(struct rds_conn_path *cp) +{ + struct rds_connection *conn = cp->cp_conn; + + rdsdebug("connection %pI6c to %pI6c reset\n", + &conn->c_laddr, &conn->c_faddr); + + rds_stats_inc(s_conn_reset); + rds_send_path_reset(cp); + cp->cp_flags = 0; + + /* Do not clear next_rx_seq here, else we cannot distinguish + * retransmitted packets from new packets, and will hand all + * of them to the application. That is not consistent with the + * reliability guarantees of RDS. */ +} + +static void __rds_conn_path_init(struct rds_connection *conn, + struct rds_conn_path *cp, bool is_outgoing) +{ + spin_lock_init(&cp->cp_lock); + cp->cp_next_tx_seq = 1; + init_waitqueue_head(&cp->cp_waitq); + INIT_LIST_HEAD(&cp->cp_send_queue); + INIT_LIST_HEAD(&cp->cp_retrans); + + cp->cp_conn = conn; + atomic_set(&cp->cp_state, RDS_CONN_DOWN); + cp->cp_send_gen = 0; + cp->cp_reconnect_jiffies = 0; + INIT_DELAYED_WORK(&cp->cp_send_w, rds_send_worker); + INIT_DELAYED_WORK(&cp->cp_recv_w, rds_recv_worker); + INIT_DELAYED_WORK(&cp->cp_conn_w, rds_connect_worker); + INIT_WORK(&cp->cp_down_w, rds_shutdown_worker); + mutex_init(&cp->cp_cm_lock); + cp->cp_flags = 0; +} + +/* + * There is only every one 'conn' for a given pair of addresses in the + * system at a time. They contain messages to be retransmitted and so + * span the lifetime of the actual underlying transport connections. + * + * For now they are not garbage collected once they're created. They + * are torn down as the module is removed, if ever. + */ +static struct rds_connection *__rds_conn_create(struct net *net, + const struct in6_addr *laddr, + const struct in6_addr *faddr, + struct rds_transport *trans, + gfp_t gfp, + int is_outgoing, + int dev_if) +{ + struct rds_connection *conn, *parent = NULL; + struct hlist_head *head = rds_conn_bucket(laddr, faddr); + struct rds_transport *loop_trans; + unsigned long flags; + int ret, i; + int npaths = (trans->t_mp_capable ? RDS_MPATH_WORKERS : 1); + + rcu_read_lock(); + conn = rds_conn_lookup(net, head, laddr, faddr, trans, dev_if); + if (conn && + conn->c_loopback && + conn->c_trans != &rds_loop_transport && + ipv6_addr_equal(laddr, faddr) && + !is_outgoing) { + /* This is a looped back IB connection, and we're + * called by the code handling the incoming connect. + * We need a second connection object into which we + * can stick the other QP. */ + parent = conn; + conn = parent->c_passive; + } + rcu_read_unlock(); + if (conn) + goto out; + + conn = kmem_cache_zalloc(rds_conn_slab, gfp); + if (!conn) { + conn = ERR_PTR(-ENOMEM); + goto out; + } + conn->c_path = kcalloc(npaths, sizeof(struct rds_conn_path), gfp); + if (!conn->c_path) { + kmem_cache_free(rds_conn_slab, conn); + conn = ERR_PTR(-ENOMEM); + goto out; + } + + INIT_HLIST_NODE(&conn->c_hash_node); + conn->c_laddr = *laddr; + conn->c_isv6 = !ipv6_addr_v4mapped(laddr); + conn->c_faddr = *faddr; + conn->c_dev_if = dev_if; + +#if IS_ENABLED(CONFIG_IPV6) + /* If the local address is link local, set c_bound_if to be the + * index used for this connection. Otherwise, set it to 0 as + * the socket is not bound to an interface. c_bound_if is used + * to look up a socket when a packet is received + */ + if (ipv6_addr_type(laddr) & IPV6_ADDR_LINKLOCAL) + conn->c_bound_if = dev_if; + else +#endif + conn->c_bound_if = 0; + + rds_conn_net_set(conn, net); + + ret = rds_cong_get_maps(conn); + if (ret) { + kfree(conn->c_path); + kmem_cache_free(rds_conn_slab, conn); + conn = ERR_PTR(ret); + goto out; + } + + /* + * This is where a connection becomes loopback. If *any* RDS sockets + * can bind to the destination address then we'd rather the messages + * flow through loopback rather than either transport. + */ + loop_trans = rds_trans_get_preferred(net, faddr, conn->c_dev_if); + if (loop_trans) { + rds_trans_put(loop_trans); + conn->c_loopback = 1; + if (trans->t_prefer_loopback) { + if (likely(is_outgoing)) { + /* "outgoing" connection to local address. + * Protocol says it wants the connection + * handled by the loopback transport. + * This is what TCP does. + */ + trans = &rds_loop_transport; + } else { + /* No transport currently in use + * should end up here, but if it + * does, reset/destroy the connection. + */ + kfree(conn->c_path); + kmem_cache_free(rds_conn_slab, conn); + conn = ERR_PTR(-EOPNOTSUPP); + goto out; + } + } + } + + conn->c_trans = trans; + + init_waitqueue_head(&conn->c_hs_waitq); + for (i = 0; i < npaths; i++) { + __rds_conn_path_init(conn, &conn->c_path[i], + is_outgoing); + conn->c_path[i].cp_index = i; + } + rcu_read_lock(); + if (rds_destroy_pending(conn)) + ret = -ENETDOWN; + else + ret = trans->conn_alloc(conn, GFP_ATOMIC); + if (ret) { + rcu_read_unlock(); + kfree(conn->c_path); + kmem_cache_free(rds_conn_slab, conn); + conn = ERR_PTR(ret); + goto out; + } + + rdsdebug("allocated conn %p for %pI6c -> %pI6c over %s %s\n", + conn, laddr, faddr, + strnlen(trans->t_name, sizeof(trans->t_name)) ? + trans->t_name : "[unknown]", is_outgoing ? "(outgoing)" : ""); + + /* + * Since we ran without holding the conn lock, someone could + * have created the same conn (either normal or passive) in the + * interim. We check while holding the lock. If we won, we complete + * init and return our conn. If we lost, we rollback and return the + * other one. + */ + spin_lock_irqsave(&rds_conn_lock, flags); + if (parent) { + /* Creating passive conn */ + if (parent->c_passive) { + trans->conn_free(conn->c_path[0].cp_transport_data); + kfree(conn->c_path); + kmem_cache_free(rds_conn_slab, conn); + conn = parent->c_passive; + } else { + parent->c_passive = conn; + rds_cong_add_conn(conn); + rds_conn_count++; + } + } else { + /* Creating normal conn */ + struct rds_connection *found; + + found = rds_conn_lookup(net, head, laddr, faddr, trans, + dev_if); + if (found) { + struct rds_conn_path *cp; + int i; + + for (i = 0; i < npaths; i++) { + cp = &conn->c_path[i]; + /* The ->conn_alloc invocation may have + * allocated resource for all paths, so all + * of them may have to be freed here. + */ + if (cp->cp_transport_data) + trans->conn_free(cp->cp_transport_data); + } + kfree(conn->c_path); + kmem_cache_free(rds_conn_slab, conn); + conn = found; + } else { + conn->c_my_gen_num = rds_gen_num; + conn->c_peer_gen_num = 0; + hlist_add_head_rcu(&conn->c_hash_node, head); + rds_cong_add_conn(conn); + rds_conn_count++; + } + } + spin_unlock_irqrestore(&rds_conn_lock, flags); + rcu_read_unlock(); + +out: + return conn; +} + +struct rds_connection *rds_conn_create(struct net *net, + const struct in6_addr *laddr, + const struct in6_addr *faddr, + struct rds_transport *trans, gfp_t gfp, + int dev_if) +{ + return __rds_conn_create(net, laddr, faddr, trans, gfp, 0, dev_if); +} +EXPORT_SYMBOL_GPL(rds_conn_create); + +struct rds_connection *rds_conn_create_outgoing(struct net *net, + const struct in6_addr *laddr, + const struct in6_addr *faddr, + struct rds_transport *trans, + gfp_t gfp, int dev_if) +{ + return __rds_conn_create(net, laddr, faddr, trans, gfp, 1, dev_if); +} +EXPORT_SYMBOL_GPL(rds_conn_create_outgoing); + +void rds_conn_shutdown(struct rds_conn_path *cp) +{ + struct rds_connection *conn = cp->cp_conn; + + /* shut it down unless it's down already */ + if (!rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_DOWN)) { + /* + * Quiesce the connection mgmt handlers before we start tearing + * things down. We don't hold the mutex for the entire + * duration of the shutdown operation, else we may be + * deadlocking with the CM handler. Instead, the CM event + * handler is supposed to check for state DISCONNECTING + */ + mutex_lock(&cp->cp_cm_lock); + if (!rds_conn_path_transition(cp, RDS_CONN_UP, + RDS_CONN_DISCONNECTING) && + !rds_conn_path_transition(cp, RDS_CONN_ERROR, + RDS_CONN_DISCONNECTING)) { + rds_conn_path_error(cp, + "shutdown called in state %d\n", + atomic_read(&cp->cp_state)); + mutex_unlock(&cp->cp_cm_lock); + return; + } + mutex_unlock(&cp->cp_cm_lock); + + wait_event(cp->cp_waitq, + !test_bit(RDS_IN_XMIT, &cp->cp_flags)); + wait_event(cp->cp_waitq, + !test_bit(RDS_RECV_REFILL, &cp->cp_flags)); + + conn->c_trans->conn_path_shutdown(cp); + rds_conn_path_reset(cp); + + if (!rds_conn_path_transition(cp, RDS_CONN_DISCONNECTING, + RDS_CONN_DOWN) && + !rds_conn_path_transition(cp, RDS_CONN_ERROR, + RDS_CONN_DOWN)) { + /* This can happen - eg when we're in the middle of tearing + * down the connection, and someone unloads the rds module. + * Quite reproducible with loopback connections. + * Mostly harmless. + * + * Note that this also happens with rds-tcp because + * we could have triggered rds_conn_path_drop in irq + * mode from rds_tcp_state change on the receipt of + * a FIN, thus we need to recheck for RDS_CONN_ERROR + * here. + */ + rds_conn_path_error(cp, "%s: failed to transition " + "to state DOWN, current state " + "is %d\n", __func__, + atomic_read(&cp->cp_state)); + return; + } + } + + /* Then reconnect if it's still live. + * The passive side of an IB loopback connection is never added + * to the conn hash, so we never trigger a reconnect on this + * conn - the reconnect is always triggered by the active peer. */ + cancel_delayed_work_sync(&cp->cp_conn_w); + rcu_read_lock(); + if (!hlist_unhashed(&conn->c_hash_node)) { + rcu_read_unlock(); + rds_queue_reconnect(cp); + } else { + rcu_read_unlock(); + } +} + +/* destroy a single rds_conn_path. rds_conn_destroy() iterates over + * all paths using rds_conn_path_destroy() + */ +static void rds_conn_path_destroy(struct rds_conn_path *cp) +{ + struct rds_message *rm, *rtmp; + + if (!cp->cp_transport_data) + return; + + /* make sure lingering queued work won't try to ref the conn */ + cancel_delayed_work_sync(&cp->cp_send_w); + cancel_delayed_work_sync(&cp->cp_recv_w); + + rds_conn_path_drop(cp, true); + flush_work(&cp->cp_down_w); + + /* tear down queued messages */ + list_for_each_entry_safe(rm, rtmp, + &cp->cp_send_queue, + m_conn_item) { + list_del_init(&rm->m_conn_item); + BUG_ON(!list_empty(&rm->m_sock_item)); + rds_message_put(rm); + } + if (cp->cp_xmit_rm) + rds_message_put(cp->cp_xmit_rm); + + WARN_ON(delayed_work_pending(&cp->cp_send_w)); + WARN_ON(delayed_work_pending(&cp->cp_recv_w)); + WARN_ON(delayed_work_pending(&cp->cp_conn_w)); + WARN_ON(work_pending(&cp->cp_down_w)); + + cp->cp_conn->c_trans->conn_free(cp->cp_transport_data); +} + +/* + * Stop and free a connection. + * + * This can only be used in very limited circumstances. It assumes that once + * the conn has been shutdown that no one else is referencing the connection. + * We can only ensure this in the rmmod path in the current code. + */ +void rds_conn_destroy(struct rds_connection *conn) +{ + unsigned long flags; + int i; + struct rds_conn_path *cp; + int npaths = (conn->c_trans->t_mp_capable ? RDS_MPATH_WORKERS : 1); + + rdsdebug("freeing conn %p for %pI4 -> " + "%pI4\n", conn, &conn->c_laddr, + &conn->c_faddr); + + /* Ensure conn will not be scheduled for reconnect */ + spin_lock_irq(&rds_conn_lock); + hlist_del_init_rcu(&conn->c_hash_node); + spin_unlock_irq(&rds_conn_lock); + synchronize_rcu(); + + /* shut the connection down */ + for (i = 0; i < npaths; i++) { + cp = &conn->c_path[i]; + rds_conn_path_destroy(cp); + BUG_ON(!list_empty(&cp->cp_retrans)); + } + + /* + * The congestion maps aren't freed up here. They're + * freed by rds_cong_exit() after all the connections + * have been freed. + */ + rds_cong_remove_conn(conn); + + kfree(conn->c_path); + kmem_cache_free(rds_conn_slab, conn); + + spin_lock_irqsave(&rds_conn_lock, flags); + rds_conn_count--; + spin_unlock_irqrestore(&rds_conn_lock, flags); +} +EXPORT_SYMBOL_GPL(rds_conn_destroy); + +static void __rds_inc_msg_cp(struct rds_incoming *inc, + struct rds_info_iterator *iter, + void *saddr, void *daddr, int flip, bool isv6) +{ +#if IS_ENABLED(CONFIG_IPV6) + if (isv6) + rds6_inc_info_copy(inc, iter, saddr, daddr, flip); + else +#endif + rds_inc_info_copy(inc, iter, *(__be32 *)saddr, + *(__be32 *)daddr, flip); +} + +static void rds_conn_message_info_cmn(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens, + int want_send, bool isv6) +{ + struct hlist_head *head; + struct list_head *list; + struct rds_connection *conn; + struct rds_message *rm; + unsigned int total = 0; + unsigned long flags; + size_t i; + int j; + + if (isv6) + len /= sizeof(struct rds6_info_message); + else + len /= sizeof(struct rds_info_message); + + rcu_read_lock(); + + for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash); + i++, head++) { + hlist_for_each_entry_rcu(conn, head, c_hash_node) { + struct rds_conn_path *cp; + int npaths; + + if (!isv6 && conn->c_isv6) + continue; + + npaths = (conn->c_trans->t_mp_capable ? + RDS_MPATH_WORKERS : 1); + + for (j = 0; j < npaths; j++) { + cp = &conn->c_path[j]; + if (want_send) + list = &cp->cp_send_queue; + else + list = &cp->cp_retrans; + + spin_lock_irqsave(&cp->cp_lock, flags); + + /* XXX too lazy to maintain counts.. */ + list_for_each_entry(rm, list, m_conn_item) { + total++; + if (total <= len) + __rds_inc_msg_cp(&rm->m_inc, + iter, + &conn->c_laddr, + &conn->c_faddr, + 0, isv6); + } + + spin_unlock_irqrestore(&cp->cp_lock, flags); + } + } + } + rcu_read_unlock(); + + lens->nr = total; + if (isv6) + lens->each = sizeof(struct rds6_info_message); + else + lens->each = sizeof(struct rds_info_message); +} + +static void rds_conn_message_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens, + int want_send) +{ + rds_conn_message_info_cmn(sock, len, iter, lens, want_send, false); +} + +#if IS_ENABLED(CONFIG_IPV6) +static void rds6_conn_message_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens, + int want_send) +{ + rds_conn_message_info_cmn(sock, len, iter, lens, want_send, true); +} +#endif + +static void rds_conn_message_info_send(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + rds_conn_message_info(sock, len, iter, lens, 1); +} + +#if IS_ENABLED(CONFIG_IPV6) +static void rds6_conn_message_info_send(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + rds6_conn_message_info(sock, len, iter, lens, 1); +} +#endif + +static void rds_conn_message_info_retrans(struct socket *sock, + unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + rds_conn_message_info(sock, len, iter, lens, 0); +} + +#if IS_ENABLED(CONFIG_IPV6) +static void rds6_conn_message_info_retrans(struct socket *sock, + unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + rds6_conn_message_info(sock, len, iter, lens, 0); +} +#endif + +void rds_for_each_conn_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens, + int (*visitor)(struct rds_connection *, void *), + u64 *buffer, + size_t item_len) +{ + struct hlist_head *head; + struct rds_connection *conn; + size_t i; + + rcu_read_lock(); + + lens->nr = 0; + lens->each = item_len; + + for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash); + i++, head++) { + hlist_for_each_entry_rcu(conn, head, c_hash_node) { + + /* XXX no c_lock usage.. */ + if (!visitor(conn, buffer)) + continue; + + /* We copy as much as we can fit in the buffer, + * but we count all items so that the caller + * can resize the buffer. */ + if (len >= item_len) { + rds_info_copy(iter, buffer, item_len); + len -= item_len; + } + lens->nr++; + } + } + rcu_read_unlock(); +} +EXPORT_SYMBOL_GPL(rds_for_each_conn_info); + +static void rds_walk_conn_path_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens, + int (*visitor)(struct rds_conn_path *, void *), + u64 *buffer, + size_t item_len) +{ + struct hlist_head *head; + struct rds_connection *conn; + size_t i; + + rcu_read_lock(); + + lens->nr = 0; + lens->each = item_len; + + for (i = 0, head = rds_conn_hash; i < ARRAY_SIZE(rds_conn_hash); + i++, head++) { + hlist_for_each_entry_rcu(conn, head, c_hash_node) { + struct rds_conn_path *cp; + + /* XXX We only copy the information from the first + * path for now. The problem is that if there are + * more than one underlying paths, we cannot report + * information of all of them using the existing + * API. For example, there is only one next_tx_seq, + * which path's next_tx_seq should we report? It is + * a bug in the design of MPRDS. + */ + cp = conn->c_path; + + /* XXX no cp_lock usage.. */ + if (!visitor(cp, buffer)) + continue; + + /* We copy as much as we can fit in the buffer, + * but we count all items so that the caller + * can resize the buffer. + */ + if (len >= item_len) { + rds_info_copy(iter, buffer, item_len); + len -= item_len; + } + lens->nr++; + } + } + rcu_read_unlock(); +} + +static int rds_conn_info_visitor(struct rds_conn_path *cp, void *buffer) +{ + struct rds_info_connection *cinfo = buffer; + struct rds_connection *conn = cp->cp_conn; + + if (conn->c_isv6) + return 0; + + cinfo->next_tx_seq = cp->cp_next_tx_seq; + cinfo->next_rx_seq = cp->cp_next_rx_seq; + cinfo->laddr = conn->c_laddr.s6_addr32[3]; + cinfo->faddr = conn->c_faddr.s6_addr32[3]; + strncpy(cinfo->transport, conn->c_trans->t_name, + sizeof(cinfo->transport)); + cinfo->flags = 0; + + rds_conn_info_set(cinfo->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags), + SENDING); + /* XXX Future: return the state rather than these funky bits */ + rds_conn_info_set(cinfo->flags, + atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING, + CONNECTING); + rds_conn_info_set(cinfo->flags, + atomic_read(&cp->cp_state) == RDS_CONN_UP, + CONNECTED); + return 1; +} + +#if IS_ENABLED(CONFIG_IPV6) +static int rds6_conn_info_visitor(struct rds_conn_path *cp, void *buffer) +{ + struct rds6_info_connection *cinfo6 = buffer; + struct rds_connection *conn = cp->cp_conn; + + cinfo6->next_tx_seq = cp->cp_next_tx_seq; + cinfo6->next_rx_seq = cp->cp_next_rx_seq; + cinfo6->laddr = conn->c_laddr; + cinfo6->faddr = conn->c_faddr; + strncpy(cinfo6->transport, conn->c_trans->t_name, + sizeof(cinfo6->transport)); + cinfo6->flags = 0; + + rds_conn_info_set(cinfo6->flags, test_bit(RDS_IN_XMIT, &cp->cp_flags), + SENDING); + /* XXX Future: return the state rather than these funky bits */ + rds_conn_info_set(cinfo6->flags, + atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING, + CONNECTING); + rds_conn_info_set(cinfo6->flags, + atomic_read(&cp->cp_state) == RDS_CONN_UP, + CONNECTED); + /* Just return 1 as there is no error case. This is a helper function + * for rds_walk_conn_path_info() and it wants a return value. + */ + return 1; +} +#endif + +static void rds_conn_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + u64 buffer[(sizeof(struct rds_info_connection) + 7) / 8]; + + rds_walk_conn_path_info(sock, len, iter, lens, + rds_conn_info_visitor, + buffer, + sizeof(struct rds_info_connection)); +} + +#if IS_ENABLED(CONFIG_IPV6) +static void rds6_conn_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + u64 buffer[(sizeof(struct rds6_info_connection) + 7) / 8]; + + rds_walk_conn_path_info(sock, len, iter, lens, + rds6_conn_info_visitor, + buffer, + sizeof(struct rds6_info_connection)); +} +#endif + +int rds_conn_init(void) +{ + int ret; + + ret = rds_loop_net_init(); /* register pernet callback */ + if (ret) + return ret; + + rds_conn_slab = kmem_cache_create("rds_connection", + sizeof(struct rds_connection), + 0, 0, NULL); + if (!rds_conn_slab) { + rds_loop_net_exit(); + return -ENOMEM; + } + + rds_info_register_func(RDS_INFO_CONNECTIONS, rds_conn_info); + rds_info_register_func(RDS_INFO_SEND_MESSAGES, + rds_conn_message_info_send); + rds_info_register_func(RDS_INFO_RETRANS_MESSAGES, + rds_conn_message_info_retrans); +#if IS_ENABLED(CONFIG_IPV6) + rds_info_register_func(RDS6_INFO_CONNECTIONS, rds6_conn_info); + rds_info_register_func(RDS6_INFO_SEND_MESSAGES, + rds6_conn_message_info_send); + rds_info_register_func(RDS6_INFO_RETRANS_MESSAGES, + rds6_conn_message_info_retrans); +#endif + return 0; +} + +void rds_conn_exit(void) +{ + rds_loop_net_exit(); /* unregister pernet callback */ + rds_loop_exit(); + + WARN_ON(!hlist_empty(rds_conn_hash)); + + kmem_cache_destroy(rds_conn_slab); + + rds_info_deregister_func(RDS_INFO_CONNECTIONS, rds_conn_info); + rds_info_deregister_func(RDS_INFO_SEND_MESSAGES, + rds_conn_message_info_send); + rds_info_deregister_func(RDS_INFO_RETRANS_MESSAGES, + rds_conn_message_info_retrans); +#if IS_ENABLED(CONFIG_IPV6) + rds_info_deregister_func(RDS6_INFO_CONNECTIONS, rds6_conn_info); + rds_info_deregister_func(RDS6_INFO_SEND_MESSAGES, + rds6_conn_message_info_send); + rds_info_deregister_func(RDS6_INFO_RETRANS_MESSAGES, + rds6_conn_message_info_retrans); +#endif +} + +/* + * Force a disconnect + */ +void rds_conn_path_drop(struct rds_conn_path *cp, bool destroy) +{ + atomic_set(&cp->cp_state, RDS_CONN_ERROR); + + rcu_read_lock(); + if (!destroy && rds_destroy_pending(cp->cp_conn)) { + rcu_read_unlock(); + return; + } + queue_work(rds_wq, &cp->cp_down_w); + rcu_read_unlock(); +} +EXPORT_SYMBOL_GPL(rds_conn_path_drop); + +void rds_conn_drop(struct rds_connection *conn) +{ + WARN_ON(conn->c_trans->t_mp_capable); + rds_conn_path_drop(&conn->c_path[0], false); +} +EXPORT_SYMBOL_GPL(rds_conn_drop); + +/* + * If the connection is down, trigger a connect. We may have scheduled a + * delayed reconnect however - in this case we should not interfere. + */ +void rds_conn_path_connect_if_down(struct rds_conn_path *cp) +{ + rcu_read_lock(); + if (rds_destroy_pending(cp->cp_conn)) { + rcu_read_unlock(); + return; + } + if (rds_conn_path_state(cp) == RDS_CONN_DOWN && + !test_and_set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags)) + queue_delayed_work(rds_wq, &cp->cp_conn_w, 0); + rcu_read_unlock(); +} +EXPORT_SYMBOL_GPL(rds_conn_path_connect_if_down); + +void rds_conn_connect_if_down(struct rds_connection *conn) +{ + WARN_ON(conn->c_trans->t_mp_capable); + rds_conn_path_connect_if_down(&conn->c_path[0]); +} +EXPORT_SYMBOL_GPL(rds_conn_connect_if_down); + +void +__rds_conn_path_error(struct rds_conn_path *cp, const char *fmt, ...) +{ + va_list ap; + + va_start(ap, fmt); + vprintk(fmt, ap); + va_end(ap); + + rds_conn_path_drop(cp, false); +} diff --git a/net/rds/ib.c b/net/rds/ib.c new file mode 100644 index 000000000..ba3379085 --- /dev/null +++ b/net/rds/ib.c @@ -0,0 +1,586 @@ +/* + * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/in.h> +#include <linux/if.h> +#include <linux/netdevice.h> +#include <linux/inetdevice.h> +#include <linux/if_arp.h> +#include <linux/delay.h> +#include <linux/slab.h> +#include <linux/module.h> +#include <net/addrconf.h> + +#include "rds_single_path.h" +#include "rds.h" +#include "ib.h" +#include "ib_mr.h" + +static unsigned int rds_ib_mr_1m_pool_size = RDS_MR_1M_POOL_SIZE; +static unsigned int rds_ib_mr_8k_pool_size = RDS_MR_8K_POOL_SIZE; +unsigned int rds_ib_retry_count = RDS_IB_DEFAULT_RETRY_COUNT; +static atomic_t rds_ib_unloading; + +module_param(rds_ib_mr_1m_pool_size, int, 0444); +MODULE_PARM_DESC(rds_ib_mr_1m_pool_size, " Max number of 1M mr per HCA"); +module_param(rds_ib_mr_8k_pool_size, int, 0444); +MODULE_PARM_DESC(rds_ib_mr_8k_pool_size, " Max number of 8K mr per HCA"); +module_param(rds_ib_retry_count, int, 0444); +MODULE_PARM_DESC(rds_ib_retry_count, " Number of hw retries before reporting an error"); + +/* + * we have a clumsy combination of RCU and a rwsem protecting this list + * because it is used both in the get_mr fast path and while blocking in + * the FMR flushing path. + */ +DECLARE_RWSEM(rds_ib_devices_lock); +struct list_head rds_ib_devices; + +/* NOTE: if also grabbing ibdev lock, grab this first */ +DEFINE_SPINLOCK(ib_nodev_conns_lock); +LIST_HEAD(ib_nodev_conns); + +static void rds_ib_nodev_connect(void) +{ + struct rds_ib_connection *ic; + + spin_lock(&ib_nodev_conns_lock); + list_for_each_entry(ic, &ib_nodev_conns, ib_node) + rds_conn_connect_if_down(ic->conn); + spin_unlock(&ib_nodev_conns_lock); +} + +static void rds_ib_dev_shutdown(struct rds_ib_device *rds_ibdev) +{ + struct rds_ib_connection *ic; + unsigned long flags; + + spin_lock_irqsave(&rds_ibdev->spinlock, flags); + list_for_each_entry(ic, &rds_ibdev->conn_list, ib_node) + rds_conn_drop(ic->conn); + spin_unlock_irqrestore(&rds_ibdev->spinlock, flags); +} + +/* + * rds_ib_destroy_mr_pool() blocks on a few things and mrs drop references + * from interrupt context so we push freing off into a work struct in krdsd. + */ +static void rds_ib_dev_free(struct work_struct *work) +{ + struct rds_ib_ipaddr *i_ipaddr, *i_next; + struct rds_ib_device *rds_ibdev = container_of(work, + struct rds_ib_device, free_work); + + if (rds_ibdev->mr_8k_pool) + rds_ib_destroy_mr_pool(rds_ibdev->mr_8k_pool); + if (rds_ibdev->mr_1m_pool) + rds_ib_destroy_mr_pool(rds_ibdev->mr_1m_pool); + if (rds_ibdev->pd) + ib_dealloc_pd(rds_ibdev->pd); + + list_for_each_entry_safe(i_ipaddr, i_next, &rds_ibdev->ipaddr_list, list) { + list_del(&i_ipaddr->list); + kfree(i_ipaddr); + } + + kfree(rds_ibdev->vector_load); + + kfree(rds_ibdev); +} + +void rds_ib_dev_put(struct rds_ib_device *rds_ibdev) +{ + BUG_ON(refcount_read(&rds_ibdev->refcount) == 0); + if (refcount_dec_and_test(&rds_ibdev->refcount)) + queue_work(rds_wq, &rds_ibdev->free_work); +} + +static void rds_ib_add_one(struct ib_device *device) +{ + struct rds_ib_device *rds_ibdev; + bool has_fr, has_fmr; + + /* Only handle IB (no iWARP) devices */ + if (device->node_type != RDMA_NODE_IB_CA) + return; + + rds_ibdev = kzalloc_node(sizeof(struct rds_ib_device), GFP_KERNEL, + ibdev_to_node(device)); + if (!rds_ibdev) + return; + + spin_lock_init(&rds_ibdev->spinlock); + refcount_set(&rds_ibdev->refcount, 1); + INIT_WORK(&rds_ibdev->free_work, rds_ib_dev_free); + + INIT_LIST_HEAD(&rds_ibdev->ipaddr_list); + INIT_LIST_HEAD(&rds_ibdev->conn_list); + + rds_ibdev->max_wrs = device->attrs.max_qp_wr; + rds_ibdev->max_sge = min(device->attrs.max_send_sge, RDS_IB_MAX_SGE); + + has_fr = (device->attrs.device_cap_flags & + IB_DEVICE_MEM_MGT_EXTENSIONS); + has_fmr = (device->alloc_fmr && device->dealloc_fmr && + device->map_phys_fmr && device->unmap_fmr); + rds_ibdev->use_fastreg = (has_fr && !has_fmr); + + rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32; + rds_ibdev->max_1m_mrs = device->attrs.max_mr ? + min_t(unsigned int, (device->attrs.max_mr / 2), + rds_ib_mr_1m_pool_size) : rds_ib_mr_1m_pool_size; + + rds_ibdev->max_8k_mrs = device->attrs.max_mr ? + min_t(unsigned int, ((device->attrs.max_mr / 2) * RDS_MR_8K_SCALE), + rds_ib_mr_8k_pool_size) : rds_ib_mr_8k_pool_size; + + rds_ibdev->max_initiator_depth = device->attrs.max_qp_init_rd_atom; + rds_ibdev->max_responder_resources = device->attrs.max_qp_rd_atom; + + rds_ibdev->vector_load = kcalloc(device->num_comp_vectors, + sizeof(int), + GFP_KERNEL); + if (!rds_ibdev->vector_load) { + pr_err("RDS/IB: %s failed to allocate vector memory\n", + __func__); + goto put_dev; + } + + rds_ibdev->dev = device; + rds_ibdev->pd = ib_alloc_pd(device, 0); + if (IS_ERR(rds_ibdev->pd)) { + rds_ibdev->pd = NULL; + goto put_dev; + } + + rds_ibdev->mr_1m_pool = + rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_1M_POOL); + if (IS_ERR(rds_ibdev->mr_1m_pool)) { + rds_ibdev->mr_1m_pool = NULL; + goto put_dev; + } + + rds_ibdev->mr_8k_pool = + rds_ib_create_mr_pool(rds_ibdev, RDS_IB_MR_8K_POOL); + if (IS_ERR(rds_ibdev->mr_8k_pool)) { + rds_ibdev->mr_8k_pool = NULL; + goto put_dev; + } + + rdsdebug("RDS/IB: max_mr = %d, max_wrs = %d, max_sge = %d, fmr_max_remaps = %d, max_1m_mrs = %d, max_8k_mrs = %d\n", + device->attrs.max_fmr, rds_ibdev->max_wrs, rds_ibdev->max_sge, + rds_ibdev->fmr_max_remaps, rds_ibdev->max_1m_mrs, + rds_ibdev->max_8k_mrs); + + pr_info("RDS/IB: %s: %s supported and preferred\n", + device->name, + rds_ibdev->use_fastreg ? "FRMR" : "FMR"); + + down_write(&rds_ib_devices_lock); + list_add_tail_rcu(&rds_ibdev->list, &rds_ib_devices); + up_write(&rds_ib_devices_lock); + refcount_inc(&rds_ibdev->refcount); + + ib_set_client_data(device, &rds_ib_client, rds_ibdev); + refcount_inc(&rds_ibdev->refcount); + + rds_ib_nodev_connect(); + +put_dev: + rds_ib_dev_put(rds_ibdev); +} + +/* + * New connections use this to find the device to associate with the + * connection. It's not in the fast path so we're not concerned about the + * performance of the IB call. (As of this writing, it uses an interrupt + * blocking spinlock to serialize walking a per-device list of all registered + * clients.) + * + * RCU is used to handle incoming connections racing with device teardown. + * Rather than use a lock to serialize removal from the client_data and + * getting a new reference, we use an RCU grace period. The destruction + * path removes the device from client_data and then waits for all RCU + * readers to finish. + * + * A new connection can get NULL from this if its arriving on a + * device that is in the process of being removed. + */ +struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device) +{ + struct rds_ib_device *rds_ibdev; + + rcu_read_lock(); + rds_ibdev = ib_get_client_data(device, &rds_ib_client); + if (rds_ibdev) + refcount_inc(&rds_ibdev->refcount); + rcu_read_unlock(); + return rds_ibdev; +} + +/* + * The IB stack is letting us know that a device is going away. This can + * happen if the underlying HCA driver is removed or if PCI hotplug is removing + * the pci function, for example. + * + * This can be called at any time and can be racing with any other RDS path. + */ +static void rds_ib_remove_one(struct ib_device *device, void *client_data) +{ + struct rds_ib_device *rds_ibdev = client_data; + + if (!rds_ibdev) + return; + + rds_ib_dev_shutdown(rds_ibdev); + + /* stop connection attempts from getting a reference to this device. */ + ib_set_client_data(device, &rds_ib_client, NULL); + + down_write(&rds_ib_devices_lock); + list_del_rcu(&rds_ibdev->list); + up_write(&rds_ib_devices_lock); + + /* + * This synchronize rcu is waiting for readers of both the ib + * client data and the devices list to finish before we drop + * both of those references. + */ + synchronize_rcu(); + rds_ib_dev_put(rds_ibdev); + rds_ib_dev_put(rds_ibdev); +} + +struct ib_client rds_ib_client = { + .name = "rds_ib", + .add = rds_ib_add_one, + .remove = rds_ib_remove_one +}; + +static int rds_ib_conn_info_visitor(struct rds_connection *conn, + void *buffer) +{ + struct rds_info_rdma_connection *iinfo = buffer; + struct rds_ib_connection *ic; + + /* We will only ever look at IB transports */ + if (conn->c_trans != &rds_ib_transport) + return 0; + if (conn->c_isv6) + return 0; + + iinfo->src_addr = conn->c_laddr.s6_addr32[3]; + iinfo->dst_addr = conn->c_faddr.s6_addr32[3]; + + memset(&iinfo->src_gid, 0, sizeof(iinfo->src_gid)); + memset(&iinfo->dst_gid, 0, sizeof(iinfo->dst_gid)); + if (rds_conn_state(conn) == RDS_CONN_UP) { + struct rds_ib_device *rds_ibdev; + + ic = conn->c_transport_data; + + rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo->src_gid, + (union ib_gid *)&iinfo->dst_gid); + + rds_ibdev = ic->rds_ibdev; + iinfo->max_send_wr = ic->i_send_ring.w_nr; + iinfo->max_recv_wr = ic->i_recv_ring.w_nr; + iinfo->max_send_sge = rds_ibdev->max_sge; + rds_ib_get_mr_info(rds_ibdev, iinfo); + } + return 1; +} + +#if IS_ENABLED(CONFIG_IPV6) +/* IPv6 version of rds_ib_conn_info_visitor(). */ +static int rds6_ib_conn_info_visitor(struct rds_connection *conn, + void *buffer) +{ + struct rds6_info_rdma_connection *iinfo6 = buffer; + struct rds_ib_connection *ic; + + /* We will only ever look at IB transports */ + if (conn->c_trans != &rds_ib_transport) + return 0; + + iinfo6->src_addr = conn->c_laddr; + iinfo6->dst_addr = conn->c_faddr; + + memset(&iinfo6->src_gid, 0, sizeof(iinfo6->src_gid)); + memset(&iinfo6->dst_gid, 0, sizeof(iinfo6->dst_gid)); + + if (rds_conn_state(conn) == RDS_CONN_UP) { + struct rds_ib_device *rds_ibdev; + + ic = conn->c_transport_data; + rdma_read_gids(ic->i_cm_id, (union ib_gid *)&iinfo6->src_gid, + (union ib_gid *)&iinfo6->dst_gid); + rds_ibdev = ic->rds_ibdev; + iinfo6->max_send_wr = ic->i_send_ring.w_nr; + iinfo6->max_recv_wr = ic->i_recv_ring.w_nr; + iinfo6->max_send_sge = rds_ibdev->max_sge; + rds6_ib_get_mr_info(rds_ibdev, iinfo6); + } + return 1; +} +#endif + +static void rds_ib_ic_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + u64 buffer[(sizeof(struct rds_info_rdma_connection) + 7) / 8]; + + rds_for_each_conn_info(sock, len, iter, lens, + rds_ib_conn_info_visitor, + buffer, + sizeof(struct rds_info_rdma_connection)); +} + +#if IS_ENABLED(CONFIG_IPV6) +/* IPv6 version of rds_ib_ic_info(). */ +static void rds6_ib_ic_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + u64 buffer[(sizeof(struct rds6_info_rdma_connection) + 7) / 8]; + + rds_for_each_conn_info(sock, len, iter, lens, + rds6_ib_conn_info_visitor, + buffer, + sizeof(struct rds6_info_rdma_connection)); +} +#endif + +/* + * Early RDS/IB was built to only bind to an address if there is an IPoIB + * device with that address set. + * + * If it were me, I'd advocate for something more flexible. Sending and + * receiving should be device-agnostic. Transports would try and maintain + * connections between peers who have messages queued. Userspace would be + * allowed to influence which paths have priority. We could call userspace + * asserting this policy "routing". + */ +static int rds_ib_laddr_check(struct net *net, const struct in6_addr *addr, + __u32 scope_id) +{ + int ret; + struct rdma_cm_id *cm_id; +#if IS_ENABLED(CONFIG_IPV6) + struct sockaddr_in6 sin6; +#endif + struct sockaddr_in sin; + struct sockaddr *sa; + bool isv4; + + isv4 = ipv6_addr_v4mapped(addr); + /* Create a CMA ID and try to bind it. This catches both + * IB and iWARP capable NICs. + */ + cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler, + NULL, RDMA_PS_TCP, IB_QPT_RC); + if (IS_ERR(cm_id)) + return PTR_ERR(cm_id); + + if (isv4) { + memset(&sin, 0, sizeof(sin)); + sin.sin_family = AF_INET; + sin.sin_addr.s_addr = addr->s6_addr32[3]; + sa = (struct sockaddr *)&sin; + } else { +#if IS_ENABLED(CONFIG_IPV6) + memset(&sin6, 0, sizeof(sin6)); + sin6.sin6_family = AF_INET6; + sin6.sin6_addr = *addr; + sin6.sin6_scope_id = scope_id; + sa = (struct sockaddr *)&sin6; + + /* XXX Do a special IPv6 link local address check here. The + * reason is that rdma_bind_addr() always succeeds with IPv6 + * link local address regardless it is indeed configured in a + * system. + */ + if (ipv6_addr_type(addr) & IPV6_ADDR_LINKLOCAL) { + struct net_device *dev; + + if (scope_id == 0) { + ret = -EADDRNOTAVAIL; + goto out; + } + + /* Use init_net for now as RDS is not network + * name space aware. + */ + dev = dev_get_by_index(&init_net, scope_id); + if (!dev) { + ret = -EADDRNOTAVAIL; + goto out; + } + if (!ipv6_chk_addr(&init_net, addr, dev, 1)) { + dev_put(dev); + ret = -EADDRNOTAVAIL; + goto out; + } + dev_put(dev); + } +#else + ret = -EADDRNOTAVAIL; + goto out; +#endif + } + + /* rdma_bind_addr will only succeed for IB & iWARP devices */ + ret = rdma_bind_addr(cm_id, sa); + /* due to this, we will claim to support iWARP devices unless we + check node_type. */ + if (ret || !cm_id->device || + cm_id->device->node_type != RDMA_NODE_IB_CA) + ret = -EADDRNOTAVAIL; + + rdsdebug("addr %pI6c%%%u ret %d node type %d\n", + addr, scope_id, ret, + cm_id->device ? cm_id->device->node_type : -1); + +out: + rdma_destroy_id(cm_id); + + return ret; +} + +static void rds_ib_unregister_client(void) +{ + ib_unregister_client(&rds_ib_client); + /* wait for rds_ib_dev_free() to complete */ + flush_workqueue(rds_wq); +} + +static void rds_ib_set_unloading(void) +{ + atomic_set(&rds_ib_unloading, 1); +} + +static bool rds_ib_is_unloading(struct rds_connection *conn) +{ + struct rds_conn_path *cp = &conn->c_path[0]; + + return (test_bit(RDS_DESTROY_PENDING, &cp->cp_flags) || + atomic_read(&rds_ib_unloading) != 0); +} + +void rds_ib_exit(void) +{ + rds_ib_set_unloading(); + synchronize_rcu(); + rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info); +#if IS_ENABLED(CONFIG_IPV6) + rds_info_deregister_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info); +#endif + rds_ib_unregister_client(); + rds_ib_destroy_nodev_conns(); + rds_ib_sysctl_exit(); + rds_ib_recv_exit(); + rds_trans_unregister(&rds_ib_transport); + rds_ib_mr_exit(); +} + +struct rds_transport rds_ib_transport = { + .laddr_check = rds_ib_laddr_check, + .xmit_path_complete = rds_ib_xmit_path_complete, + .xmit = rds_ib_xmit, + .xmit_rdma = rds_ib_xmit_rdma, + .xmit_atomic = rds_ib_xmit_atomic, + .recv_path = rds_ib_recv_path, + .conn_alloc = rds_ib_conn_alloc, + .conn_free = rds_ib_conn_free, + .conn_path_connect = rds_ib_conn_path_connect, + .conn_path_shutdown = rds_ib_conn_path_shutdown, + .inc_copy_to_user = rds_ib_inc_copy_to_user, + .inc_free = rds_ib_inc_free, + .cm_initiate_connect = rds_ib_cm_initiate_connect, + .cm_handle_connect = rds_ib_cm_handle_connect, + .cm_connect_complete = rds_ib_cm_connect_complete, + .stats_info_copy = rds_ib_stats_info_copy, + .exit = rds_ib_exit, + .get_mr = rds_ib_get_mr, + .sync_mr = rds_ib_sync_mr, + .free_mr = rds_ib_free_mr, + .flush_mrs = rds_ib_flush_mrs, + .t_owner = THIS_MODULE, + .t_name = "infiniband", + .t_unloading = rds_ib_is_unloading, + .t_type = RDS_TRANS_IB +}; + +int rds_ib_init(void) +{ + int ret; + + INIT_LIST_HEAD(&rds_ib_devices); + + ret = rds_ib_mr_init(); + if (ret) + goto out; + + ret = ib_register_client(&rds_ib_client); + if (ret) + goto out_mr_exit; + + ret = rds_ib_sysctl_init(); + if (ret) + goto out_ibreg; + + ret = rds_ib_recv_init(); + if (ret) + goto out_sysctl; + + rds_trans_register(&rds_ib_transport); + + rds_info_register_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info); +#if IS_ENABLED(CONFIG_IPV6) + rds_info_register_func(RDS6_INFO_IB_CONNECTIONS, rds6_ib_ic_info); +#endif + + goto out; + +out_sysctl: + rds_ib_sysctl_exit(); +out_ibreg: + rds_ib_unregister_client(); +out_mr_exit: + rds_ib_mr_exit(); +out: + return ret; +} + +MODULE_LICENSE("GPL"); diff --git a/net/rds/ib.h b/net/rds/ib.h new file mode 100644 index 000000000..71ff356ee --- /dev/null +++ b/net/rds/ib.h @@ -0,0 +1,463 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _RDS_IB_H +#define _RDS_IB_H + +#include <rdma/ib_verbs.h> +#include <rdma/rdma_cm.h> +#include <linux/interrupt.h> +#include <linux/pci.h> +#include <linux/slab.h> +#include "rds.h" +#include "rdma_transport.h" + +#define RDS_IB_MAX_SGE 8 +#define RDS_IB_RECV_SGE 2 + +#define RDS_IB_DEFAULT_RECV_WR 1024 +#define RDS_IB_DEFAULT_SEND_WR 256 +#define RDS_IB_DEFAULT_FR_WR 256 +#define RDS_IB_DEFAULT_FR_INV_WR 256 + +#define RDS_IB_DEFAULT_RETRY_COUNT 1 + +#define RDS_IB_SUPPORTED_PROTOCOLS 0x00000003 /* minor versions supported */ + +#define RDS_IB_RECYCLE_BATCH_COUNT 32 + +#define RDS_IB_WC_MAX 32 + +extern struct rw_semaphore rds_ib_devices_lock; +extern struct list_head rds_ib_devices; + +/* + * IB posts RDS_FRAG_SIZE fragments of pages to the receive queues to + * try and minimize the amount of memory tied up both the device and + * socket receive queues. + */ +struct rds_page_frag { + struct list_head f_item; + struct list_head f_cache_entry; + struct scatterlist f_sg; +}; + +struct rds_ib_incoming { + struct list_head ii_frags; + struct list_head ii_cache_entry; + struct rds_incoming ii_inc; +}; + +struct rds_ib_cache_head { + struct list_head *first; + unsigned long count; +}; + +struct rds_ib_refill_cache { + struct rds_ib_cache_head __percpu *percpu; + struct list_head *xfer; + struct list_head *ready; +}; + +/* This is the common structure for the IB private data exchange in setting up + * an RDS connection. The exchange is different for IPv4 and IPv6 connections. + * The reason is that the address size is different and the addresses + * exchanged are in the beginning of the structure. Hence it is not possible + * for interoperability if same structure is used. + */ +struct rds_ib_conn_priv_cmn { + u8 ricpc_protocol_major; + u8 ricpc_protocol_minor; + __be16 ricpc_protocol_minor_mask; /* bitmask */ + __be32 ricpc_reserved1; + __be64 ricpc_ack_seq; + __be32 ricpc_credit; /* non-zero enables flow ctl */ +}; + +struct rds_ib_connect_private { + /* Add new fields at the end, and don't permute existing fields. */ + __be32 dp_saddr; + __be32 dp_daddr; + struct rds_ib_conn_priv_cmn dp_cmn; +}; + +struct rds6_ib_connect_private { + /* Add new fields at the end, and don't permute existing fields. */ + struct in6_addr dp_saddr; + struct in6_addr dp_daddr; + struct rds_ib_conn_priv_cmn dp_cmn; +}; + +#define dp_protocol_major dp_cmn.ricpc_protocol_major +#define dp_protocol_minor dp_cmn.ricpc_protocol_minor +#define dp_protocol_minor_mask dp_cmn.ricpc_protocol_minor_mask +#define dp_ack_seq dp_cmn.ricpc_ack_seq +#define dp_credit dp_cmn.ricpc_credit + +union rds_ib_conn_priv { + struct rds_ib_connect_private ricp_v4; + struct rds6_ib_connect_private ricp_v6; +}; + +struct rds_ib_send_work { + void *s_op; + union { + struct ib_send_wr s_wr; + struct ib_rdma_wr s_rdma_wr; + struct ib_atomic_wr s_atomic_wr; + }; + struct ib_sge s_sge[RDS_IB_MAX_SGE]; + unsigned long s_queued; +}; + +struct rds_ib_recv_work { + struct rds_ib_incoming *r_ibinc; + struct rds_page_frag *r_frag; + struct ib_recv_wr r_wr; + struct ib_sge r_sge[2]; +}; + +struct rds_ib_work_ring { + u32 w_nr; + u32 w_alloc_ptr; + u32 w_alloc_ctr; + u32 w_free_ptr; + atomic_t w_free_ctr; +}; + +/* Rings are posted with all the allocations they'll need to queue the + * incoming message to the receiving socket so this can't fail. + * All fragments start with a header, so we can make sure we're not receiving + * garbage, and we can tell a small 8 byte fragment from an ACK frame. + */ +struct rds_ib_ack_state { + u64 ack_next; + u64 ack_recv; + unsigned int ack_required:1; + unsigned int ack_next_valid:1; + unsigned int ack_recv_valid:1; +}; + + +struct rds_ib_device; + +struct rds_ib_connection { + + struct list_head ib_node; + struct rds_ib_device *rds_ibdev; + struct rds_connection *conn; + + /* alphabet soup, IBTA style */ + struct rdma_cm_id *i_cm_id; + struct ib_pd *i_pd; + struct ib_cq *i_send_cq; + struct ib_cq *i_recv_cq; + struct ib_wc i_send_wc[RDS_IB_WC_MAX]; + struct ib_wc i_recv_wc[RDS_IB_WC_MAX]; + + /* To control the number of wrs from fastreg */ + atomic_t i_fastreg_wrs; + atomic_t i_fastunreg_wrs; + + /* interrupt handling */ + struct tasklet_struct i_send_tasklet; + struct tasklet_struct i_recv_tasklet; + + /* tx */ + struct rds_ib_work_ring i_send_ring; + struct rm_data_op *i_data_op; + struct rds_header *i_send_hdrs; + dma_addr_t i_send_hdrs_dma; + struct rds_ib_send_work *i_sends; + atomic_t i_signaled_sends; + + /* rx */ + struct mutex i_recv_mutex; + struct rds_ib_work_ring i_recv_ring; + struct rds_ib_incoming *i_ibinc; + u32 i_recv_data_rem; + struct rds_header *i_recv_hdrs; + dma_addr_t i_recv_hdrs_dma; + struct rds_ib_recv_work *i_recvs; + u64 i_ack_recv; /* last ACK received */ + struct rds_ib_refill_cache i_cache_incs; + struct rds_ib_refill_cache i_cache_frags; + atomic_t i_cache_allocs; + + /* sending acks */ + unsigned long i_ack_flags; +#ifdef KERNEL_HAS_ATOMIC64 + atomic64_t i_ack_next; /* next ACK to send */ +#else + spinlock_t i_ack_lock; /* protect i_ack_next */ + u64 i_ack_next; /* next ACK to send */ +#endif + struct rds_header *i_ack; + struct ib_send_wr i_ack_wr; + struct ib_sge i_ack_sge; + dma_addr_t i_ack_dma; + unsigned long i_ack_queued; + + /* Flow control related information + * + * Our algorithm uses a pair variables that we need to access + * atomically - one for the send credits, and one posted + * recv credits we need to transfer to remote. + * Rather than protect them using a slow spinlock, we put both into + * a single atomic_t and update it using cmpxchg + */ + atomic_t i_credits; + + /* Protocol version specific information */ + unsigned int i_flowctl:1; /* enable/disable flow ctl */ + + /* Batched completions */ + unsigned int i_unsignaled_wrs; + + /* Endpoint role in connection */ + bool i_active_side; + atomic_t i_cq_quiesce; + + /* Send/Recv vectors */ + int i_scq_vector; + int i_rcq_vector; +}; + +/* This assumes that atomic_t is at least 32 bits */ +#define IB_GET_SEND_CREDITS(v) ((v) & 0xffff) +#define IB_GET_POST_CREDITS(v) ((v) >> 16) +#define IB_SET_SEND_CREDITS(v) ((v) & 0xffff) +#define IB_SET_POST_CREDITS(v) ((v) << 16) + +struct rds_ib_ipaddr { + struct list_head list; + __be32 ipaddr; + struct rcu_head rcu; +}; + +enum { + RDS_IB_MR_8K_POOL, + RDS_IB_MR_1M_POOL, +}; + +struct rds_ib_device { + struct list_head list; + struct list_head ipaddr_list; + struct list_head conn_list; + struct ib_device *dev; + struct ib_pd *pd; + bool use_fastreg; + + unsigned int max_mrs; + struct rds_ib_mr_pool *mr_1m_pool; + struct rds_ib_mr_pool *mr_8k_pool; + unsigned int fmr_max_remaps; + unsigned int max_8k_mrs; + unsigned int max_1m_mrs; + int max_sge; + unsigned int max_wrs; + unsigned int max_initiator_depth; + unsigned int max_responder_resources; + spinlock_t spinlock; /* protect the above */ + refcount_t refcount; + struct work_struct free_work; + int *vector_load; +}; + +#define ibdev_to_node(ibdev) dev_to_node((ibdev)->dev.parent) +#define rdsibdev_to_node(rdsibdev) ibdev_to_node(rdsibdev->dev) + +/* bits for i_ack_flags */ +#define IB_ACK_IN_FLIGHT 0 +#define IB_ACK_REQUESTED 1 + +/* Magic WR_ID for ACKs */ +#define RDS_IB_ACK_WR_ID (~(u64) 0) + +struct rds_ib_statistics { + uint64_t s_ib_connect_raced; + uint64_t s_ib_listen_closed_stale; + uint64_t s_ib_evt_handler_call; + uint64_t s_ib_tasklet_call; + uint64_t s_ib_tx_cq_event; + uint64_t s_ib_tx_ring_full; + uint64_t s_ib_tx_throttle; + uint64_t s_ib_tx_sg_mapping_failure; + uint64_t s_ib_tx_stalled; + uint64_t s_ib_tx_credit_updates; + uint64_t s_ib_rx_cq_event; + uint64_t s_ib_rx_ring_empty; + uint64_t s_ib_rx_refill_from_cq; + uint64_t s_ib_rx_refill_from_thread; + uint64_t s_ib_rx_alloc_limit; + uint64_t s_ib_rx_total_frags; + uint64_t s_ib_rx_total_incs; + uint64_t s_ib_rx_credit_updates; + uint64_t s_ib_ack_sent; + uint64_t s_ib_ack_send_failure; + uint64_t s_ib_ack_send_delayed; + uint64_t s_ib_ack_send_piggybacked; + uint64_t s_ib_ack_received; + uint64_t s_ib_rdma_mr_8k_alloc; + uint64_t s_ib_rdma_mr_8k_free; + uint64_t s_ib_rdma_mr_8k_used; + uint64_t s_ib_rdma_mr_8k_pool_flush; + uint64_t s_ib_rdma_mr_8k_pool_wait; + uint64_t s_ib_rdma_mr_8k_pool_depleted; + uint64_t s_ib_rdma_mr_1m_alloc; + uint64_t s_ib_rdma_mr_1m_free; + uint64_t s_ib_rdma_mr_1m_used; + uint64_t s_ib_rdma_mr_1m_pool_flush; + uint64_t s_ib_rdma_mr_1m_pool_wait; + uint64_t s_ib_rdma_mr_1m_pool_depleted; + uint64_t s_ib_rdma_mr_8k_reused; + uint64_t s_ib_rdma_mr_1m_reused; + uint64_t s_ib_atomic_cswp; + uint64_t s_ib_atomic_fadd; + uint64_t s_ib_recv_added_to_cache; + uint64_t s_ib_recv_removed_from_cache; +}; + +extern struct workqueue_struct *rds_ib_wq; + +/* + * Fake ib_dma_sync_sg_for_{cpu,device} as long as ib_verbs.h + * doesn't define it. + */ +static inline void rds_ib_dma_sync_sg_for_cpu(struct ib_device *dev, + struct scatterlist *sglist, + unsigned int sg_dma_len, + int direction) +{ + struct scatterlist *sg; + unsigned int i; + + for_each_sg(sglist, sg, sg_dma_len, i) { + ib_dma_sync_single_for_cpu(dev, + ib_sg_dma_address(dev, sg), + ib_sg_dma_len(dev, sg), + direction); + } +} +#define ib_dma_sync_sg_for_cpu rds_ib_dma_sync_sg_for_cpu + +static inline void rds_ib_dma_sync_sg_for_device(struct ib_device *dev, + struct scatterlist *sglist, + unsigned int sg_dma_len, + int direction) +{ + struct scatterlist *sg; + unsigned int i; + + for_each_sg(sglist, sg, sg_dma_len, i) { + ib_dma_sync_single_for_device(dev, + ib_sg_dma_address(dev, sg), + ib_sg_dma_len(dev, sg), + direction); + } +} +#define ib_dma_sync_sg_for_device rds_ib_dma_sync_sg_for_device + + +/* ib.c */ +extern struct rds_transport rds_ib_transport; +struct rds_ib_device *rds_ib_get_client_data(struct ib_device *device); +void rds_ib_dev_put(struct rds_ib_device *rds_ibdev); +extern struct ib_client rds_ib_client; + +extern unsigned int rds_ib_retry_count; + +extern spinlock_t ib_nodev_conns_lock; +extern struct list_head ib_nodev_conns; + +/* ib_cm.c */ +int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp); +void rds_ib_conn_free(void *arg); +int rds_ib_conn_path_connect(struct rds_conn_path *cp); +void rds_ib_conn_path_shutdown(struct rds_conn_path *cp); +void rds_ib_state_change(struct sock *sk); +int rds_ib_listen_init(void); +void rds_ib_listen_stop(void); +__printf(2, 3) +void __rds_ib_conn_error(struct rds_connection *conn, const char *, ...); +int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, + struct rdma_cm_event *event, bool isv6); +int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6); +void rds_ib_cm_connect_complete(struct rds_connection *conn, + struct rdma_cm_event *event); + + +#define rds_ib_conn_error(conn, fmt...) \ + __rds_ib_conn_error(conn, KERN_WARNING "RDS/IB: " fmt) + +/* ib_rdma.c */ +int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, + struct in6_addr *ipaddr); +void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); +void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn); +void rds_ib_destroy_nodev_conns(void); +void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc); + +/* ib_recv.c */ +int rds_ib_recv_init(void); +void rds_ib_recv_exit(void); +int rds_ib_recv_path(struct rds_conn_path *conn); +int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp); +void rds_ib_recv_free_caches(struct rds_ib_connection *ic); +void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp); +void rds_ib_inc_free(struct rds_incoming *inc); +int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to); +void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc, + struct rds_ib_ack_state *state); +void rds_ib_recv_tasklet_fn(unsigned long data); +void rds_ib_recv_init_ring(struct rds_ib_connection *ic); +void rds_ib_recv_clear_ring(struct rds_ib_connection *ic); +void rds_ib_recv_init_ack(struct rds_ib_connection *ic); +void rds_ib_attempt_ack(struct rds_ib_connection *ic); +void rds_ib_ack_send_complete(struct rds_ib_connection *ic); +u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic); +void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required); + +/* ib_ring.c */ +void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr); +void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr); +u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos); +void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val); +void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val); +int rds_ib_ring_empty(struct rds_ib_work_ring *ring); +int rds_ib_ring_low(struct rds_ib_work_ring *ring); +u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring); +u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest); +extern wait_queue_head_t rds_ib_ring_empty_wait; + +/* ib_send.c */ +void rds_ib_xmit_path_complete(struct rds_conn_path *cp); +int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm, + unsigned int hdr_off, unsigned int sg, unsigned int off); +void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc); +void rds_ib_send_init_ring(struct rds_ib_connection *ic); +void rds_ib_send_clear_ring(struct rds_ib_connection *ic); +int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op); +void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits); +void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted); +int rds_ib_send_grab_credits(struct rds_ib_connection *ic, u32 wanted, + u32 *adv_credits, int need_posted, int max_posted); +int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op); + +/* ib_stats.c */ +DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_ib_statistics, rds_ib_stats); +#define rds_ib_stats_inc(member) rds_stats_inc_which(rds_ib_stats, member) +#define rds_ib_stats_add(member, count) \ + rds_stats_add_which(rds_ib_stats, member, count) +unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter, + unsigned int avail); + +/* ib_sysctl.c */ +int rds_ib_sysctl_init(void); +void rds_ib_sysctl_exit(void); +extern unsigned long rds_ib_sysctl_max_send_wr; +extern unsigned long rds_ib_sysctl_max_recv_wr; +extern unsigned long rds_ib_sysctl_max_unsig_wrs; +extern unsigned long rds_ib_sysctl_max_unsig_bytes; +extern unsigned long rds_ib_sysctl_max_recv_allocation; +extern unsigned int rds_ib_sysctl_flow_control; + +#endif diff --git a/net/rds/ib_cm.c b/net/rds/ib_cm.c new file mode 100644 index 000000000..bfbb31f0c --- /dev/null +++ b/net/rds/ib_cm.c @@ -0,0 +1,1190 @@ +/* + * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/in.h> +#include <linux/slab.h> +#include <linux/vmalloc.h> +#include <linux/ratelimit.h> +#include <net/addrconf.h> + +#include "rds_single_path.h" +#include "rds.h" +#include "ib.h" + +/* + * Set the selected protocol version + */ +static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version) +{ + conn->c_version = version; +} + +/* + * Set up flow control + */ +static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + + if (rds_ib_sysctl_flow_control && credits != 0) { + /* We're doing flow control */ + ic->i_flowctl = 1; + rds_ib_send_add_credits(conn, credits); + } else { + ic->i_flowctl = 0; + } +} + +/* + * Tune RNR behavior. Without flow control, we use a rather + * low timeout, but not the absolute minimum - this should + * be tunable. + * + * We already set the RNR retry count to 7 (which is the + * smallest infinite number :-) above. + * If flow control is off, we want to change this back to 0 + * so that we learn quickly when our credit accounting is + * buggy. + * + * Caller passes in a qp_attr pointer - don't waste stack spacv + * by allocation this twice. + */ +static void +rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr) +{ + int ret; + + attr->min_rnr_timer = IB_RNR_TIMER_000_32; + ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER); + if (ret) + printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret); +} + +/* + * Connection established. + * We get here for both outgoing and incoming connection. + */ +void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + const union rds_ib_conn_priv *dp = NULL; + struct ib_qp_attr qp_attr; + __be64 ack_seq = 0; + __be32 credit = 0; + u8 major = 0; + u8 minor = 0; + int err; + + dp = event->param.conn.private_data; + if (conn->c_isv6) { + if (event->param.conn.private_data_len >= + sizeof(struct rds6_ib_connect_private)) { + major = dp->ricp_v6.dp_protocol_major; + minor = dp->ricp_v6.dp_protocol_minor; + credit = dp->ricp_v6.dp_credit; + /* dp structure start is not guaranteed to be 8 bytes + * aligned. Since dp_ack_seq is 64-bit extended load + * operations can be used so go through get_unaligned + * to avoid unaligned errors. + */ + ack_seq = get_unaligned(&dp->ricp_v6.dp_ack_seq); + } + } else if (event->param.conn.private_data_len >= + sizeof(struct rds_ib_connect_private)) { + major = dp->ricp_v4.dp_protocol_major; + minor = dp->ricp_v4.dp_protocol_minor; + credit = dp->ricp_v4.dp_credit; + ack_seq = get_unaligned(&dp->ricp_v4.dp_ack_seq); + } + + /* make sure it isn't empty data */ + if (major) { + rds_ib_set_protocol(conn, RDS_PROTOCOL(major, minor)); + rds_ib_set_flow_control(conn, be32_to_cpu(credit)); + } + + if (conn->c_version < RDS_PROTOCOL(3, 1)) { + pr_notice("RDS/IB: Connection <%pI6c,%pI6c> version %u.%u no longer supported\n", + &conn->c_laddr, &conn->c_faddr, + RDS_PROTOCOL_MAJOR(conn->c_version), + RDS_PROTOCOL_MINOR(conn->c_version)); + set_bit(RDS_DESTROY_PENDING, &conn->c_path[0].cp_flags); + rds_conn_destroy(conn); + return; + } else { + pr_notice("RDS/IB: %s conn connected <%pI6c,%pI6c> version %u.%u%s\n", + ic->i_active_side ? "Active" : "Passive", + &conn->c_laddr, &conn->c_faddr, + RDS_PROTOCOL_MAJOR(conn->c_version), + RDS_PROTOCOL_MINOR(conn->c_version), + ic->i_flowctl ? ", flow control" : ""); + } + + atomic_set(&ic->i_cq_quiesce, 0); + + /* Init rings and fill recv. this needs to wait until protocol + * negotiation is complete, since ring layout is different + * from 3.1 to 4.1. + */ + rds_ib_send_init_ring(ic); + rds_ib_recv_init_ring(ic); + /* Post receive buffers - as a side effect, this will update + * the posted credit count. */ + rds_ib_recv_refill(conn, 1, GFP_KERNEL); + + /* Tune RNR behavior */ + rds_ib_tune_rnr(ic, &qp_attr); + + qp_attr.qp_state = IB_QPS_RTS; + err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE); + if (err) + printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err); + + /* update ib_device with this local ipaddr */ + err = rds_ib_update_ipaddr(ic->rds_ibdev, &conn->c_laddr); + if (err) + printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", + err); + + /* If the peer gave us the last packet it saw, process this as if + * we had received a regular ACK. */ + if (dp) { + if (ack_seq) + rds_send_drop_acked(conn, be64_to_cpu(ack_seq), + NULL); + } + + rds_connect_complete(conn); +} + +static void rds_ib_cm_fill_conn_param(struct rds_connection *conn, + struct rdma_conn_param *conn_param, + union rds_ib_conn_priv *dp, + u32 protocol_version, + u32 max_responder_resources, + u32 max_initiator_depth, + bool isv6) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + struct rds_ib_device *rds_ibdev = ic->rds_ibdev; + + memset(conn_param, 0, sizeof(struct rdma_conn_param)); + + conn_param->responder_resources = + min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources); + conn_param->initiator_depth = + min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth); + conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7); + conn_param->rnr_retry_count = 7; + + if (dp) { + memset(dp, 0, sizeof(*dp)); + if (isv6) { + dp->ricp_v6.dp_saddr = conn->c_laddr; + dp->ricp_v6.dp_daddr = conn->c_faddr; + dp->ricp_v6.dp_protocol_major = + RDS_PROTOCOL_MAJOR(protocol_version); + dp->ricp_v6.dp_protocol_minor = + RDS_PROTOCOL_MINOR(protocol_version); + dp->ricp_v6.dp_protocol_minor_mask = + cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS); + dp->ricp_v6.dp_ack_seq = + cpu_to_be64(rds_ib_piggyb_ack(ic)); + + conn_param->private_data = &dp->ricp_v6; + conn_param->private_data_len = sizeof(dp->ricp_v6); + } else { + dp->ricp_v4.dp_saddr = conn->c_laddr.s6_addr32[3]; + dp->ricp_v4.dp_daddr = conn->c_faddr.s6_addr32[3]; + dp->ricp_v4.dp_protocol_major = + RDS_PROTOCOL_MAJOR(protocol_version); + dp->ricp_v4.dp_protocol_minor = + RDS_PROTOCOL_MINOR(protocol_version); + dp->ricp_v4.dp_protocol_minor_mask = + cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS); + dp->ricp_v4.dp_ack_seq = + cpu_to_be64(rds_ib_piggyb_ack(ic)); + + conn_param->private_data = &dp->ricp_v4; + conn_param->private_data_len = sizeof(dp->ricp_v4); + } + + /* Advertise flow control */ + if (ic->i_flowctl) { + unsigned int credits; + + credits = IB_GET_POST_CREDITS + (atomic_read(&ic->i_credits)); + if (isv6) + dp->ricp_v6.dp_credit = cpu_to_be32(credits); + else + dp->ricp_v4.dp_credit = cpu_to_be32(credits); + atomic_sub(IB_SET_POST_CREDITS(credits), + &ic->i_credits); + } + } +} + +static void rds_ib_cq_event_handler(struct ib_event *event, void *data) +{ + rdsdebug("event %u (%s) data %p\n", + event->event, ib_event_msg(event->event), data); +} + +/* Plucking the oldest entry from the ring can be done concurrently with + * the thread refilling the ring. Each ring operation is protected by + * spinlocks and the transient state of refilling doesn't change the + * recording of which entry is oldest. + * + * This relies on IB only calling one cq comp_handler for each cq so that + * there will only be one caller of rds_recv_incoming() per RDS connection. + */ +static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context) +{ + struct rds_connection *conn = context; + struct rds_ib_connection *ic = conn->c_transport_data; + + rdsdebug("conn %p cq %p\n", conn, cq); + + rds_ib_stats_inc(s_ib_evt_handler_call); + + tasklet_schedule(&ic->i_recv_tasklet); +} + +static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq, + struct ib_wc *wcs) +{ + int nr, i; + struct ib_wc *wc; + + while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) { + for (i = 0; i < nr; i++) { + wc = wcs + i; + rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n", + (unsigned long long)wc->wr_id, wc->status, + wc->byte_len, be32_to_cpu(wc->ex.imm_data)); + + if (wc->wr_id <= ic->i_send_ring.w_nr || + wc->wr_id == RDS_IB_ACK_WR_ID) + rds_ib_send_cqe_handler(ic, wc); + else + rds_ib_mr_cqe_handler(ic, wc); + + } + } +} + +static void rds_ib_tasklet_fn_send(unsigned long data) +{ + struct rds_ib_connection *ic = (struct rds_ib_connection *)data; + struct rds_connection *conn = ic->conn; + + rds_ib_stats_inc(s_ib_tasklet_call); + + /* if cq has been already reaped, ignore incoming cq event */ + if (atomic_read(&ic->i_cq_quiesce)) + return; + + poll_scq(ic, ic->i_send_cq, ic->i_send_wc); + ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP); + poll_scq(ic, ic->i_send_cq, ic->i_send_wc); + + if (rds_conn_up(conn) && + (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) || + test_bit(0, &conn->c_map_queued))) + rds_send_xmit(&ic->conn->c_path[0]); +} + +static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq, + struct ib_wc *wcs, + struct rds_ib_ack_state *ack_state) +{ + int nr, i; + struct ib_wc *wc; + + while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) { + for (i = 0; i < nr; i++) { + wc = wcs + i; + rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n", + (unsigned long long)wc->wr_id, wc->status, + wc->byte_len, be32_to_cpu(wc->ex.imm_data)); + + rds_ib_recv_cqe_handler(ic, wc, ack_state); + } + } +} + +static void rds_ib_tasklet_fn_recv(unsigned long data) +{ + struct rds_ib_connection *ic = (struct rds_ib_connection *)data; + struct rds_connection *conn = ic->conn; + struct rds_ib_device *rds_ibdev = ic->rds_ibdev; + struct rds_ib_ack_state state; + + if (!rds_ibdev) + rds_conn_drop(conn); + + rds_ib_stats_inc(s_ib_tasklet_call); + + /* if cq has been already reaped, ignore incoming cq event */ + if (atomic_read(&ic->i_cq_quiesce)) + return; + + memset(&state, 0, sizeof(state)); + poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state); + ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); + poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state); + + if (state.ack_next_valid) + rds_ib_set_ack(ic, state.ack_next, state.ack_required); + if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) { + rds_send_drop_acked(conn, state.ack_recv, NULL); + ic->i_ack_recv = state.ack_recv; + } + + if (rds_conn_up(conn)) + rds_ib_attempt_ack(ic); +} + +static void rds_ib_qp_event_handler(struct ib_event *event, void *data) +{ + struct rds_connection *conn = data; + struct rds_ib_connection *ic = conn->c_transport_data; + + rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event, + ib_event_msg(event->event)); + + switch (event->event) { + case IB_EVENT_COMM_EST: + rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST); + break; + default: + rdsdebug("Fatal QP Event %u (%s) " + "- connection %pI6c->%pI6c, reconnecting\n", + event->event, ib_event_msg(event->event), + &conn->c_laddr, &conn->c_faddr); + rds_conn_drop(conn); + break; + } +} + +static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context) +{ + struct rds_connection *conn = context; + struct rds_ib_connection *ic = conn->c_transport_data; + + rdsdebug("conn %p cq %p\n", conn, cq); + + rds_ib_stats_inc(s_ib_evt_handler_call); + + tasklet_schedule(&ic->i_send_tasklet); +} + +static inline int ibdev_get_unused_vector(struct rds_ib_device *rds_ibdev) +{ + int min = rds_ibdev->vector_load[rds_ibdev->dev->num_comp_vectors - 1]; + int index = rds_ibdev->dev->num_comp_vectors - 1; + int i; + + for (i = rds_ibdev->dev->num_comp_vectors - 1; i >= 0; i--) { + if (rds_ibdev->vector_load[i] < min) { + index = i; + min = rds_ibdev->vector_load[i]; + } + } + + rds_ibdev->vector_load[index]++; + return index; +} + +static inline void ibdev_put_vector(struct rds_ib_device *rds_ibdev, int index) +{ + rds_ibdev->vector_load[index]--; +} + +/* + * This needs to be very careful to not leave IS_ERR pointers around for + * cleanup to trip over. + */ +static int rds_ib_setup_qp(struct rds_connection *conn) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + struct ib_device *dev = ic->i_cm_id->device; + struct ib_qp_init_attr attr; + struct ib_cq_init_attr cq_attr = {}; + struct rds_ib_device *rds_ibdev; + int ret, fr_queue_space; + + /* + * It's normal to see a null device if an incoming connection races + * with device removal, so we don't print a warning. + */ + rds_ibdev = rds_ib_get_client_data(dev); + if (!rds_ibdev) + return -EOPNOTSUPP; + + /* The fr_queue_space is currently set to 512, to add extra space on + * completion queue and send queue. This extra space is used for FRMR + * registration and invalidation work requests + */ + fr_queue_space = rds_ibdev->use_fastreg ? + (RDS_IB_DEFAULT_FR_WR + 1) + + (RDS_IB_DEFAULT_FR_INV_WR + 1) + : 0; + + /* add the conn now so that connection establishment has the dev */ + rds_ib_add_conn(rds_ibdev, conn); + + if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1) + rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1); + if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1) + rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1); + + /* Protection domain and memory range */ + ic->i_pd = rds_ibdev->pd; + + ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev); + cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1; + cq_attr.comp_vector = ic->i_scq_vector; + ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send, + rds_ib_cq_event_handler, conn, + &cq_attr); + if (IS_ERR(ic->i_send_cq)) { + ret = PTR_ERR(ic->i_send_cq); + ic->i_send_cq = NULL; + ibdev_put_vector(rds_ibdev, ic->i_scq_vector); + rdsdebug("ib_create_cq send failed: %d\n", ret); + goto rds_ibdev_out; + } + + ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev); + cq_attr.cqe = ic->i_recv_ring.w_nr; + cq_attr.comp_vector = ic->i_rcq_vector; + ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv, + rds_ib_cq_event_handler, conn, + &cq_attr); + if (IS_ERR(ic->i_recv_cq)) { + ret = PTR_ERR(ic->i_recv_cq); + ic->i_recv_cq = NULL; + ibdev_put_vector(rds_ibdev, ic->i_rcq_vector); + rdsdebug("ib_create_cq recv failed: %d\n", ret); + goto send_cq_out; + } + + ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP); + if (ret) { + rdsdebug("ib_req_notify_cq send failed: %d\n", ret); + goto recv_cq_out; + } + + ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); + if (ret) { + rdsdebug("ib_req_notify_cq recv failed: %d\n", ret); + goto recv_cq_out; + } + + /* XXX negotiate max send/recv with remote? */ + memset(&attr, 0, sizeof(attr)); + attr.event_handler = rds_ib_qp_event_handler; + attr.qp_context = conn; + /* + 1 to allow for the single ack message */ + attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1; + attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1; + attr.cap.max_send_sge = rds_ibdev->max_sge; + attr.cap.max_recv_sge = RDS_IB_RECV_SGE; + attr.sq_sig_type = IB_SIGNAL_REQ_WR; + attr.qp_type = IB_QPT_RC; + attr.send_cq = ic->i_send_cq; + attr.recv_cq = ic->i_recv_cq; + atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR); + atomic_set(&ic->i_fastunreg_wrs, RDS_IB_DEFAULT_FR_INV_WR); + + /* + * XXX this can fail if max_*_wr is too large? Are we supposed + * to back off until we get a value that the hardware can support? + */ + ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr); + if (ret) { + rdsdebug("rdma_create_qp failed: %d\n", ret); + goto recv_cq_out; + } + + ic->i_send_hdrs = ib_dma_alloc_coherent(dev, + ic->i_send_ring.w_nr * + sizeof(struct rds_header), + &ic->i_send_hdrs_dma, GFP_KERNEL); + if (!ic->i_send_hdrs) { + ret = -ENOMEM; + rdsdebug("ib_dma_alloc_coherent send failed\n"); + goto qp_out; + } + + ic->i_recv_hdrs = ib_dma_alloc_coherent(dev, + ic->i_recv_ring.w_nr * + sizeof(struct rds_header), + &ic->i_recv_hdrs_dma, GFP_KERNEL); + if (!ic->i_recv_hdrs) { + ret = -ENOMEM; + rdsdebug("ib_dma_alloc_coherent recv failed\n"); + goto send_hdrs_dma_out; + } + + ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header), + &ic->i_ack_dma, GFP_KERNEL); + if (!ic->i_ack) { + ret = -ENOMEM; + rdsdebug("ib_dma_alloc_coherent ack failed\n"); + goto recv_hdrs_dma_out; + } + + ic->i_sends = vzalloc_node(array_size(sizeof(struct rds_ib_send_work), + ic->i_send_ring.w_nr), + ibdev_to_node(dev)); + if (!ic->i_sends) { + ret = -ENOMEM; + rdsdebug("send allocation failed\n"); + goto ack_dma_out; + } + + ic->i_recvs = vzalloc_node(array_size(sizeof(struct rds_ib_recv_work), + ic->i_recv_ring.w_nr), + ibdev_to_node(dev)); + if (!ic->i_recvs) { + ret = -ENOMEM; + rdsdebug("recv allocation failed\n"); + goto sends_out; + } + + rds_ib_recv_init_ack(ic); + + rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd, + ic->i_send_cq, ic->i_recv_cq); + + goto out; + +sends_out: + vfree(ic->i_sends); +ack_dma_out: + ib_dma_free_coherent(dev, sizeof(struct rds_header), + ic->i_ack, ic->i_ack_dma); +recv_hdrs_dma_out: + ib_dma_free_coherent(dev, ic->i_recv_ring.w_nr * + sizeof(struct rds_header), + ic->i_recv_hdrs, ic->i_recv_hdrs_dma); +send_hdrs_dma_out: + ib_dma_free_coherent(dev, ic->i_send_ring.w_nr * + sizeof(struct rds_header), + ic->i_send_hdrs, ic->i_send_hdrs_dma); +qp_out: + rdma_destroy_qp(ic->i_cm_id); +recv_cq_out: + if (!ib_destroy_cq(ic->i_recv_cq)) + ic->i_recv_cq = NULL; +send_cq_out: + if (!ib_destroy_cq(ic->i_send_cq)) + ic->i_send_cq = NULL; +rds_ibdev_out: + rds_ib_remove_conn(rds_ibdev, conn); +out: + rds_ib_dev_put(rds_ibdev); + + return ret; +} + +static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event, bool isv6) +{ + const union rds_ib_conn_priv *dp = event->param.conn.private_data; + u8 data_len, major, minor; + u32 version = 0; + __be16 mask; + u16 common; + + /* + * rdma_cm private data is odd - when there is any private data in the + * request, we will be given a pretty large buffer without telling us the + * original size. The only way to tell the difference is by looking at + * the contents, which are initialized to zero. + * If the protocol version fields aren't set, this is a connection attempt + * from an older version. This could could be 3.0 or 2.0 - we can't tell. + * We really should have changed this for OFED 1.3 :-( + */ + + /* Be paranoid. RDS always has privdata */ + if (!event->param.conn.private_data_len) { + printk(KERN_NOTICE "RDS incoming connection has no private data, " + "rejecting\n"); + return 0; + } + + if (isv6) { + data_len = sizeof(struct rds6_ib_connect_private); + major = dp->ricp_v6.dp_protocol_major; + minor = dp->ricp_v6.dp_protocol_minor; + mask = dp->ricp_v6.dp_protocol_minor_mask; + } else { + data_len = sizeof(struct rds_ib_connect_private); + major = dp->ricp_v4.dp_protocol_major; + minor = dp->ricp_v4.dp_protocol_minor; + mask = dp->ricp_v4.dp_protocol_minor_mask; + } + + /* Even if len is crap *now* I still want to check it. -ASG */ + if (event->param.conn.private_data_len < data_len || major == 0) + return RDS_PROTOCOL_3_0; + + common = be16_to_cpu(mask) & RDS_IB_SUPPORTED_PROTOCOLS; + if (major == 3 && common) { + version = RDS_PROTOCOL_3_0; + while ((common >>= 1) != 0) + version++; + } else { + if (isv6) + printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI6c using incompatible protocol version %u.%u\n", + &dp->ricp_v6.dp_saddr, major, minor); + else + printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n", + &dp->ricp_v4.dp_saddr, major, minor); + } + return version; +} + +#if IS_ENABLED(CONFIG_IPV6) +/* Given an IPv6 address, find the net_device which hosts that address and + * return its index. This is used by the rds_ib_cm_handle_connect() code to + * find the interface index of where an incoming request comes from when + * the request is using a link local address. + * + * Note one problem in this search. It is possible that two interfaces have + * the same link local address. Unfortunately, this cannot be solved unless + * the underlying layer gives us the interface which an incoming RDMA connect + * request comes from. + */ +static u32 __rds_find_ifindex(struct net *net, const struct in6_addr *addr) +{ + struct net_device *dev; + int idx = 0; + + rcu_read_lock(); + for_each_netdev_rcu(net, dev) { + if (ipv6_chk_addr(net, addr, dev, 1)) { + idx = dev->ifindex; + break; + } + } + rcu_read_unlock(); + + return idx; +} +#endif + +int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, + struct rdma_cm_event *event, bool isv6) +{ + __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id; + __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id; + const struct rds_ib_conn_priv_cmn *dp_cmn; + struct rds_connection *conn = NULL; + struct rds_ib_connection *ic = NULL; + struct rdma_conn_param conn_param; + const union rds_ib_conn_priv *dp; + union rds_ib_conn_priv dp_rep; + struct in6_addr s_mapped_addr; + struct in6_addr d_mapped_addr; + const struct in6_addr *saddr6; + const struct in6_addr *daddr6; + int destroy = 1; + u32 ifindex = 0; + u32 version; + int err = 1; + + /* Check whether the remote protocol version matches ours. */ + version = rds_ib_protocol_compatible(event, isv6); + if (!version) + goto out; + + dp = event->param.conn.private_data; + if (isv6) { +#if IS_ENABLED(CONFIG_IPV6) + dp_cmn = &dp->ricp_v6.dp_cmn; + saddr6 = &dp->ricp_v6.dp_saddr; + daddr6 = &dp->ricp_v6.dp_daddr; + /* If either address is link local, need to find the + * interface index in order to create a proper RDS + * connection. + */ + if (ipv6_addr_type(daddr6) & IPV6_ADDR_LINKLOCAL) { + /* Using init_net for now .. */ + ifindex = __rds_find_ifindex(&init_net, daddr6); + /* No index found... Need to bail out. */ + if (ifindex == 0) { + err = -EOPNOTSUPP; + goto out; + } + } else if (ipv6_addr_type(saddr6) & IPV6_ADDR_LINKLOCAL) { + /* Use our address to find the correct index. */ + ifindex = __rds_find_ifindex(&init_net, daddr6); + /* No index found... Need to bail out. */ + if (ifindex == 0) { + err = -EOPNOTSUPP; + goto out; + } + } +#else + err = -EOPNOTSUPP; + goto out; +#endif + } else { + dp_cmn = &dp->ricp_v4.dp_cmn; + ipv6_addr_set_v4mapped(dp->ricp_v4.dp_saddr, &s_mapped_addr); + ipv6_addr_set_v4mapped(dp->ricp_v4.dp_daddr, &d_mapped_addr); + saddr6 = &s_mapped_addr; + daddr6 = &d_mapped_addr; + } + + rdsdebug("saddr %pI6c daddr %pI6c RDSv%u.%u lguid 0x%llx fguid " + "0x%llx\n", saddr6, daddr6, + RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version), + (unsigned long long)be64_to_cpu(lguid), + (unsigned long long)be64_to_cpu(fguid)); + + /* RDS/IB is not currently netns aware, thus init_net */ + conn = rds_conn_create(&init_net, daddr6, saddr6, + &rds_ib_transport, GFP_KERNEL, ifindex); + if (IS_ERR(conn)) { + rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn)); + conn = NULL; + goto out; + } + + /* + * The connection request may occur while the + * previous connection exist, e.g. in case of failover. + * But as connections may be initiated simultaneously + * by both hosts, we have a random backoff mechanism - + * see the comment above rds_queue_reconnect() + */ + mutex_lock(&conn->c_cm_lock); + if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) { + if (rds_conn_state(conn) == RDS_CONN_UP) { + rdsdebug("incoming connect while connecting\n"); + rds_conn_drop(conn); + rds_ib_stats_inc(s_ib_listen_closed_stale); + } else + if (rds_conn_state(conn) == RDS_CONN_CONNECTING) { + /* Wait and see - our connect may still be succeeding */ + rds_ib_stats_inc(s_ib_connect_raced); + } + goto out; + } + + ic = conn->c_transport_data; + + rds_ib_set_protocol(conn, version); + rds_ib_set_flow_control(conn, be32_to_cpu(dp_cmn->ricpc_credit)); + + /* If the peer gave us the last packet it saw, process this as if + * we had received a regular ACK. */ + if (dp_cmn->ricpc_ack_seq) + rds_send_drop_acked(conn, be64_to_cpu(dp_cmn->ricpc_ack_seq), + NULL); + + BUG_ON(cm_id->context); + BUG_ON(ic->i_cm_id); + + ic->i_cm_id = cm_id; + cm_id->context = conn; + + /* We got halfway through setting up the ib_connection, if we + * fail now, we have to take the long route out of this mess. */ + destroy = 0; + + err = rds_ib_setup_qp(conn); + if (err) { + rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err); + goto out; + } + + rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version, + event->param.conn.responder_resources, + event->param.conn.initiator_depth, isv6); + + /* rdma_accept() calls rdma_reject() internally if it fails */ + if (rdma_accept(cm_id, &conn_param)) + rds_ib_conn_error(conn, "rdma_accept failed\n"); + +out: + if (conn) + mutex_unlock(&conn->c_cm_lock); + if (err) + rdma_reject(cm_id, NULL, 0); + return destroy; +} + + +int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id, bool isv6) +{ + struct rds_connection *conn = cm_id->context; + struct rds_ib_connection *ic = conn->c_transport_data; + struct rdma_conn_param conn_param; + union rds_ib_conn_priv dp; + int ret; + + /* If the peer doesn't do protocol negotiation, we must + * default to RDSv3.0 */ + rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0); + ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */ + + ret = rds_ib_setup_qp(conn); + if (ret) { + rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret); + goto out; + } + + rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION, + UINT_MAX, UINT_MAX, isv6); + ret = rdma_connect(cm_id, &conn_param); + if (ret) + rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret); + +out: + /* Beware - returning non-zero tells the rdma_cm to destroy + * the cm_id. We should certainly not do it as long as we still + * "own" the cm_id. */ + if (ret) { + if (ic->i_cm_id == cm_id) + ret = 0; + } + ic->i_active_side = true; + return ret; +} + +int rds_ib_conn_path_connect(struct rds_conn_path *cp) +{ + struct rds_connection *conn = cp->cp_conn; + struct sockaddr_storage src, dest; + rdma_cm_event_handler handler; + struct rds_ib_connection *ic; + int ret; + + ic = conn->c_transport_data; + + /* XXX I wonder what affect the port space has */ + /* delegate cm event handler to rdma_transport */ +#if IS_ENABLED(CONFIG_IPV6) + if (conn->c_isv6) + handler = rds6_rdma_cm_event_handler; + else +#endif + handler = rds_rdma_cm_event_handler; + ic->i_cm_id = rdma_create_id(&init_net, handler, conn, + RDMA_PS_TCP, IB_QPT_RC); + if (IS_ERR(ic->i_cm_id)) { + ret = PTR_ERR(ic->i_cm_id); + ic->i_cm_id = NULL; + rdsdebug("rdma_create_id() failed: %d\n", ret); + goto out; + } + + rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn); + + if (ipv6_addr_v4mapped(&conn->c_faddr)) { + struct sockaddr_in *sin; + + sin = (struct sockaddr_in *)&src; + sin->sin_family = AF_INET; + sin->sin_addr.s_addr = conn->c_laddr.s6_addr32[3]; + sin->sin_port = 0; + + sin = (struct sockaddr_in *)&dest; + sin->sin_family = AF_INET; + sin->sin_addr.s_addr = conn->c_faddr.s6_addr32[3]; + sin->sin_port = htons(RDS_PORT); + } else { + struct sockaddr_in6 *sin6; + + sin6 = (struct sockaddr_in6 *)&src; + sin6->sin6_family = AF_INET6; + sin6->sin6_addr = conn->c_laddr; + sin6->sin6_port = 0; + sin6->sin6_scope_id = conn->c_dev_if; + + sin6 = (struct sockaddr_in6 *)&dest; + sin6->sin6_family = AF_INET6; + sin6->sin6_addr = conn->c_faddr; + sin6->sin6_port = htons(RDS_CM_PORT); + sin6->sin6_scope_id = conn->c_dev_if; + } + + ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src, + (struct sockaddr *)&dest, + RDS_RDMA_RESOLVE_TIMEOUT_MS); + if (ret) { + rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id, + ret); + rdma_destroy_id(ic->i_cm_id); + ic->i_cm_id = NULL; + } + +out: + return ret; +} + +/* + * This is so careful about only cleaning up resources that were built up + * so that it can be called at any point during startup. In fact it + * can be called multiple times for a given connection. + */ +void rds_ib_conn_path_shutdown(struct rds_conn_path *cp) +{ + struct rds_connection *conn = cp->cp_conn; + struct rds_ib_connection *ic = conn->c_transport_data; + int err = 0; + + rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id, + ic->i_pd, ic->i_send_cq, ic->i_recv_cq, + ic->i_cm_id ? ic->i_cm_id->qp : NULL); + + if (ic->i_cm_id) { + struct ib_device *dev = ic->i_cm_id->device; + + rdsdebug("disconnecting cm %p\n", ic->i_cm_id); + err = rdma_disconnect(ic->i_cm_id); + if (err) { + /* Actually this may happen quite frequently, when + * an outgoing connect raced with an incoming connect. + */ + rdsdebug("failed to disconnect, cm: %p err %d\n", + ic->i_cm_id, err); + } + + /* + * We want to wait for tx and rx completion to finish + * before we tear down the connection, but we have to be + * careful not to get stuck waiting on a send ring that + * only has unsignaled sends in it. We've shutdown new + * sends before getting here so by waiting for signaled + * sends to complete we're ensured that there will be no + * more tx processing. + */ + wait_event(rds_ib_ring_empty_wait, + rds_ib_ring_empty(&ic->i_recv_ring) && + (atomic_read(&ic->i_signaled_sends) == 0) && + (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR) && + (atomic_read(&ic->i_fastunreg_wrs) == RDS_IB_DEFAULT_FR_INV_WR)); + tasklet_kill(&ic->i_send_tasklet); + tasklet_kill(&ic->i_recv_tasklet); + + atomic_set(&ic->i_cq_quiesce, 1); + + /* first destroy the ib state that generates callbacks */ + if (ic->i_cm_id->qp) + rdma_destroy_qp(ic->i_cm_id); + if (ic->i_send_cq) { + if (ic->rds_ibdev) + ibdev_put_vector(ic->rds_ibdev, ic->i_scq_vector); + ib_destroy_cq(ic->i_send_cq); + } + + if (ic->i_recv_cq) { + if (ic->rds_ibdev) + ibdev_put_vector(ic->rds_ibdev, ic->i_rcq_vector); + ib_destroy_cq(ic->i_recv_cq); + } + + /* then free the resources that ib callbacks use */ + if (ic->i_send_hdrs) + ib_dma_free_coherent(dev, + ic->i_send_ring.w_nr * + sizeof(struct rds_header), + ic->i_send_hdrs, + ic->i_send_hdrs_dma); + + if (ic->i_recv_hdrs) + ib_dma_free_coherent(dev, + ic->i_recv_ring.w_nr * + sizeof(struct rds_header), + ic->i_recv_hdrs, + ic->i_recv_hdrs_dma); + + if (ic->i_ack) + ib_dma_free_coherent(dev, sizeof(struct rds_header), + ic->i_ack, ic->i_ack_dma); + + if (ic->i_sends) + rds_ib_send_clear_ring(ic); + if (ic->i_recvs) + rds_ib_recv_clear_ring(ic); + + rdma_destroy_id(ic->i_cm_id); + + /* + * Move connection back to the nodev list. + */ + if (ic->rds_ibdev) + rds_ib_remove_conn(ic->rds_ibdev, conn); + + ic->i_cm_id = NULL; + ic->i_pd = NULL; + ic->i_send_cq = NULL; + ic->i_recv_cq = NULL; + ic->i_send_hdrs = NULL; + ic->i_recv_hdrs = NULL; + ic->i_ack = NULL; + } + BUG_ON(ic->rds_ibdev); + + /* Clear pending transmit */ + if (ic->i_data_op) { + struct rds_message *rm; + + rm = container_of(ic->i_data_op, struct rds_message, data); + rds_message_put(rm); + ic->i_data_op = NULL; + } + + /* Clear the ACK state */ + clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); +#ifdef KERNEL_HAS_ATOMIC64 + atomic64_set(&ic->i_ack_next, 0); +#else + ic->i_ack_next = 0; +#endif + ic->i_ack_recv = 0; + + /* Clear flow control state */ + ic->i_flowctl = 0; + atomic_set(&ic->i_credits, 0); + + rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); + rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); + + if (ic->i_ibinc) { + rds_inc_put(&ic->i_ibinc->ii_inc); + ic->i_ibinc = NULL; + } + + vfree(ic->i_sends); + ic->i_sends = NULL; + vfree(ic->i_recvs); + ic->i_recvs = NULL; + ic->i_active_side = false; +} + +int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp) +{ + struct rds_ib_connection *ic; + unsigned long flags; + int ret; + + /* XXX too lazy? */ + ic = kzalloc(sizeof(struct rds_ib_connection), gfp); + if (!ic) + return -ENOMEM; + + ret = rds_ib_recv_alloc_caches(ic, gfp); + if (ret) { + kfree(ic); + return ret; + } + + INIT_LIST_HEAD(&ic->ib_node); + tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send, + (unsigned long)ic); + tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv, + (unsigned long)ic); + mutex_init(&ic->i_recv_mutex); +#ifndef KERNEL_HAS_ATOMIC64 + spin_lock_init(&ic->i_ack_lock); +#endif + atomic_set(&ic->i_signaled_sends, 0); + + /* + * rds_ib_conn_shutdown() waits for these to be emptied so they + * must be initialized before it can be called. + */ + rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); + rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); + + ic->conn = conn; + conn->c_transport_data = ic; + + spin_lock_irqsave(&ib_nodev_conns_lock, flags); + list_add_tail(&ic->ib_node, &ib_nodev_conns); + spin_unlock_irqrestore(&ib_nodev_conns_lock, flags); + + + rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data); + return 0; +} + +/* + * Free a connection. Connection must be shut down and not set for reconnect. + */ +void rds_ib_conn_free(void *arg) +{ + struct rds_ib_connection *ic = arg; + spinlock_t *lock_ptr; + + rdsdebug("ic %p\n", ic); + + /* + * Conn is either on a dev's list or on the nodev list. + * A race with shutdown() or connect() would cause problems + * (since rds_ibdev would change) but that should never happen. + */ + lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock; + + spin_lock_irq(lock_ptr); + list_del(&ic->ib_node); + spin_unlock_irq(lock_ptr); + + rds_ib_recv_free_caches(ic); + + kfree(ic); +} + + +/* + * An error occurred on the connection + */ +void +__rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...) +{ + va_list ap; + + rds_conn_drop(conn); + + va_start(ap, fmt); + vprintk(fmt, ap); + va_end(ap); +} diff --git a/net/rds/ib_fmr.c b/net/rds/ib_fmr.c new file mode 100644 index 000000000..01e764f8f --- /dev/null +++ b/net/rds/ib_fmr.c @@ -0,0 +1,269 @@ +/* + * Copyright (c) 2016 Oracle. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "ib_mr.h" + +struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev, int npages) +{ + struct rds_ib_mr_pool *pool; + struct rds_ib_mr *ibmr = NULL; + struct rds_ib_fmr *fmr; + int err = 0; + + if (npages <= RDS_MR_8K_MSG_SIZE) + pool = rds_ibdev->mr_8k_pool; + else + pool = rds_ibdev->mr_1m_pool; + + if (atomic_read(&pool->dirty_count) >= pool->max_items / 10) + queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10); + + /* Switch pools if one of the pool is reaching upper limit */ + if (atomic_read(&pool->dirty_count) >= pool->max_items * 9 / 10) { + if (pool->pool_type == RDS_IB_MR_8K_POOL) + pool = rds_ibdev->mr_1m_pool; + else + pool = rds_ibdev->mr_8k_pool; + } + + ibmr = rds_ib_try_reuse_ibmr(pool); + if (ibmr) + return ibmr; + + ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, + rdsibdev_to_node(rds_ibdev)); + if (!ibmr) { + err = -ENOMEM; + goto out_no_cigar; + } + + fmr = &ibmr->u.fmr; + fmr->fmr = ib_alloc_fmr(rds_ibdev->pd, + (IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_READ | + IB_ACCESS_REMOTE_WRITE | + IB_ACCESS_REMOTE_ATOMIC), + &pool->fmr_attr); + if (IS_ERR(fmr->fmr)) { + err = PTR_ERR(fmr->fmr); + fmr->fmr = NULL; + pr_warn("RDS/IB: %s failed (err=%d)\n", __func__, err); + goto out_no_cigar; + } + + ibmr->pool = pool; + if (pool->pool_type == RDS_IB_MR_8K_POOL) + rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc); + else + rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc); + + return ibmr; + +out_no_cigar: + kfree(ibmr); + atomic_dec(&pool->item_count); + + return ERR_PTR(err); +} + +static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, + struct rds_ib_mr *ibmr, struct scatterlist *sg, + unsigned int nents) +{ + struct ib_device *dev = rds_ibdev->dev; + struct rds_ib_fmr *fmr = &ibmr->u.fmr; + struct scatterlist *scat = sg; + u64 io_addr = 0; + u64 *dma_pages; + u32 len; + int page_cnt, sg_dma_len; + int i, j; + int ret; + + sg_dma_len = ib_dma_map_sg(dev, sg, nents, DMA_BIDIRECTIONAL); + if (unlikely(!sg_dma_len)) { + pr_warn("RDS/IB: %s failed!\n", __func__); + return -EBUSY; + } + + len = 0; + page_cnt = 0; + + for (i = 0; i < sg_dma_len; ++i) { + unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]); + u64 dma_addr = ib_sg_dma_address(dev, &scat[i]); + + if (dma_addr & ~PAGE_MASK) { + if (i > 0) { + ib_dma_unmap_sg(dev, sg, nents, + DMA_BIDIRECTIONAL); + return -EINVAL; + } else { + ++page_cnt; + } + } + if ((dma_addr + dma_len) & ~PAGE_MASK) { + if (i < sg_dma_len - 1) { + ib_dma_unmap_sg(dev, sg, nents, + DMA_BIDIRECTIONAL); + return -EINVAL; + } else { + ++page_cnt; + } + } + + len += dma_len; + } + + page_cnt += len >> PAGE_SHIFT; + if (page_cnt > ibmr->pool->fmr_attr.max_pages) { + ib_dma_unmap_sg(dev, sg, nents, DMA_BIDIRECTIONAL); + return -EINVAL; + } + + dma_pages = kmalloc_array_node(sizeof(u64), page_cnt, GFP_ATOMIC, + rdsibdev_to_node(rds_ibdev)); + if (!dma_pages) { + ib_dma_unmap_sg(dev, sg, nents, DMA_BIDIRECTIONAL); + return -ENOMEM; + } + + page_cnt = 0; + for (i = 0; i < sg_dma_len; ++i) { + unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]); + u64 dma_addr = ib_sg_dma_address(dev, &scat[i]); + + for (j = 0; j < dma_len; j += PAGE_SIZE) + dma_pages[page_cnt++] = + (dma_addr & PAGE_MASK) + j; + } + + ret = ib_map_phys_fmr(fmr->fmr, dma_pages, page_cnt, io_addr); + if (ret) { + ib_dma_unmap_sg(dev, sg, nents, DMA_BIDIRECTIONAL); + goto out; + } + + /* Success - we successfully remapped the MR, so we can + * safely tear down the old mapping. + */ + rds_ib_teardown_mr(ibmr); + + ibmr->sg = scat; + ibmr->sg_len = nents; + ibmr->sg_dma_len = sg_dma_len; + ibmr->remap_count++; + + if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL) + rds_ib_stats_inc(s_ib_rdma_mr_8k_used); + else + rds_ib_stats_inc(s_ib_rdma_mr_1m_used); + ret = 0; + +out: + kfree(dma_pages); + + return ret; +} + +struct rds_ib_mr *rds_ib_reg_fmr(struct rds_ib_device *rds_ibdev, + struct scatterlist *sg, + unsigned long nents, + u32 *key) +{ + struct rds_ib_mr *ibmr = NULL; + struct rds_ib_fmr *fmr; + int ret; + + ibmr = rds_ib_alloc_fmr(rds_ibdev, nents); + if (IS_ERR(ibmr)) + return ibmr; + + ibmr->device = rds_ibdev; + fmr = &ibmr->u.fmr; + ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents); + if (ret == 0) + *key = fmr->fmr->rkey; + else + rds_ib_free_mr(ibmr, 0); + + return ibmr; +} + +void rds_ib_unreg_fmr(struct list_head *list, unsigned int *nfreed, + unsigned long *unpinned, unsigned int goal) +{ + struct rds_ib_mr *ibmr, *next; + struct rds_ib_fmr *fmr; + LIST_HEAD(fmr_list); + int ret = 0; + unsigned int freed = *nfreed; + + /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */ + list_for_each_entry(ibmr, list, unmap_list) { + fmr = &ibmr->u.fmr; + list_add(&fmr->fmr->list, &fmr_list); + } + + ret = ib_unmap_fmr(&fmr_list); + if (ret) + pr_warn("RDS/IB: FMR invalidation failed (err=%d)\n", ret); + + /* Now we can destroy the DMA mapping and unpin any pages */ + list_for_each_entry_safe(ibmr, next, list, unmap_list) { + fmr = &ibmr->u.fmr; + *unpinned += ibmr->sg_len; + __rds_ib_teardown_mr(ibmr); + if (freed < goal || + ibmr->remap_count >= ibmr->pool->fmr_attr.max_maps) { + if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL) + rds_ib_stats_inc(s_ib_rdma_mr_8k_free); + else + rds_ib_stats_inc(s_ib_rdma_mr_1m_free); + list_del(&ibmr->unmap_list); + ib_dealloc_fmr(fmr->fmr); + kfree(ibmr); + freed++; + } + } + *nfreed = freed; +} + +void rds_ib_free_fmr_list(struct rds_ib_mr *ibmr) +{ + struct rds_ib_mr_pool *pool = ibmr->pool; + + if (ibmr->remap_count >= pool->fmr_attr.max_maps) + llist_add(&ibmr->llnode, &pool->drop_list); + else + llist_add(&ibmr->llnode, &pool->free_list); +} diff --git a/net/rds/ib_frmr.c b/net/rds/ib_frmr.c new file mode 100644 index 000000000..46988c009 --- /dev/null +++ b/net/rds/ib_frmr.c @@ -0,0 +1,379 @@ +/* + * Copyright (c) 2016 Oracle. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "ib_mr.h" + +static struct rds_ib_mr *rds_ib_alloc_frmr(struct rds_ib_device *rds_ibdev, + int npages) +{ + struct rds_ib_mr_pool *pool; + struct rds_ib_mr *ibmr = NULL; + struct rds_ib_frmr *frmr; + int err = 0; + + if (npages <= RDS_MR_8K_MSG_SIZE) + pool = rds_ibdev->mr_8k_pool; + else + pool = rds_ibdev->mr_1m_pool; + + ibmr = rds_ib_try_reuse_ibmr(pool); + if (ibmr) + return ibmr; + + ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, + rdsibdev_to_node(rds_ibdev)); + if (!ibmr) { + err = -ENOMEM; + goto out_no_cigar; + } + + frmr = &ibmr->u.frmr; + frmr->mr = ib_alloc_mr(rds_ibdev->pd, IB_MR_TYPE_MEM_REG, + pool->fmr_attr.max_pages); + if (IS_ERR(frmr->mr)) { + pr_warn("RDS/IB: %s failed to allocate MR", __func__); + err = PTR_ERR(frmr->mr); + goto out_no_cigar; + } + + ibmr->pool = pool; + if (pool->pool_type == RDS_IB_MR_8K_POOL) + rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc); + else + rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc); + + if (atomic_read(&pool->item_count) > pool->max_items_soft) + pool->max_items_soft = pool->max_items; + + frmr->fr_state = FRMR_IS_FREE; + return ibmr; + +out_no_cigar: + kfree(ibmr); + atomic_dec(&pool->item_count); + return ERR_PTR(err); +} + +static void rds_ib_free_frmr(struct rds_ib_mr *ibmr, bool drop) +{ + struct rds_ib_mr_pool *pool = ibmr->pool; + + if (drop) + llist_add(&ibmr->llnode, &pool->drop_list); + else + llist_add(&ibmr->llnode, &pool->free_list); + atomic_add(ibmr->sg_len, &pool->free_pinned); + atomic_inc(&pool->dirty_count); + + /* If we've pinned too many pages, request a flush */ + if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned || + atomic_read(&pool->dirty_count) >= pool->max_items / 5) + queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10); +} + +static int rds_ib_post_reg_frmr(struct rds_ib_mr *ibmr) +{ + struct rds_ib_frmr *frmr = &ibmr->u.frmr; + struct ib_reg_wr reg_wr; + int ret, off = 0; + + while (atomic_dec_return(&ibmr->ic->i_fastreg_wrs) <= 0) { + atomic_inc(&ibmr->ic->i_fastreg_wrs); + cpu_relax(); + } + + ret = ib_map_mr_sg_zbva(frmr->mr, ibmr->sg, ibmr->sg_dma_len, + &off, PAGE_SIZE); + if (unlikely(ret != ibmr->sg_dma_len)) + return ret < 0 ? ret : -EINVAL; + + /* Perform a WR for the fast_reg_mr. Each individual page + * in the sg list is added to the fast reg page list and placed + * inside the fast_reg_mr WR. The key used is a rolling 8bit + * counter, which should guarantee uniqueness. + */ + ib_update_fast_reg_key(frmr->mr, ibmr->remap_count++); + frmr->fr_state = FRMR_IS_INUSE; + + memset(®_wr, 0, sizeof(reg_wr)); + reg_wr.wr.wr_id = (unsigned long)(void *)ibmr; + reg_wr.wr.opcode = IB_WR_REG_MR; + reg_wr.wr.num_sge = 0; + reg_wr.mr = frmr->mr; + reg_wr.key = frmr->mr->rkey; + reg_wr.access = IB_ACCESS_LOCAL_WRITE | + IB_ACCESS_REMOTE_READ | + IB_ACCESS_REMOTE_WRITE; + reg_wr.wr.send_flags = IB_SEND_SIGNALED; + + ret = ib_post_send(ibmr->ic->i_cm_id->qp, ®_wr.wr, NULL); + if (unlikely(ret)) { + /* Failure here can be because of -ENOMEM as well */ + frmr->fr_state = FRMR_IS_STALE; + atomic_inc(&ibmr->ic->i_fastreg_wrs); + if (printk_ratelimit()) + pr_warn("RDS/IB: %s returned error(%d)\n", + __func__, ret); + } + return ret; +} + +static int rds_ib_map_frmr(struct rds_ib_device *rds_ibdev, + struct rds_ib_mr_pool *pool, + struct rds_ib_mr *ibmr, + struct scatterlist *sg, unsigned int sg_len) +{ + struct ib_device *dev = rds_ibdev->dev; + struct rds_ib_frmr *frmr = &ibmr->u.frmr; + int i; + u32 len; + int ret = 0; + + /* We want to teardown old ibmr values here and fill it up with + * new sg values + */ + rds_ib_teardown_mr(ibmr); + + ibmr->sg = sg; + ibmr->sg_len = sg_len; + ibmr->sg_dma_len = 0; + frmr->sg_byte_len = 0; + WARN_ON(ibmr->sg_dma_len); + ibmr->sg_dma_len = ib_dma_map_sg(dev, ibmr->sg, ibmr->sg_len, + DMA_BIDIRECTIONAL); + if (unlikely(!ibmr->sg_dma_len)) { + pr_warn("RDS/IB: %s failed!\n", __func__); + return -EBUSY; + } + + frmr->sg_byte_len = 0; + frmr->dma_npages = 0; + len = 0; + + ret = -EINVAL; + for (i = 0; i < ibmr->sg_dma_len; ++i) { + unsigned int dma_len = ib_sg_dma_len(dev, &ibmr->sg[i]); + u64 dma_addr = ib_sg_dma_address(dev, &ibmr->sg[i]); + + frmr->sg_byte_len += dma_len; + if (dma_addr & ~PAGE_MASK) { + if (i > 0) + goto out_unmap; + else + ++frmr->dma_npages; + } + + if ((dma_addr + dma_len) & ~PAGE_MASK) { + if (i < ibmr->sg_dma_len - 1) + goto out_unmap; + else + ++frmr->dma_npages; + } + + len += dma_len; + } + frmr->dma_npages += len >> PAGE_SHIFT; + + if (frmr->dma_npages > ibmr->pool->fmr_attr.max_pages) { + ret = -EMSGSIZE; + goto out_unmap; + } + + ret = rds_ib_post_reg_frmr(ibmr); + if (ret) + goto out_unmap; + + if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL) + rds_ib_stats_inc(s_ib_rdma_mr_8k_used); + else + rds_ib_stats_inc(s_ib_rdma_mr_1m_used); + + return ret; + +out_unmap: + ib_dma_unmap_sg(rds_ibdev->dev, ibmr->sg, ibmr->sg_len, + DMA_BIDIRECTIONAL); + ibmr->sg_dma_len = 0; + return ret; +} + +static int rds_ib_post_inv(struct rds_ib_mr *ibmr) +{ + struct ib_send_wr *s_wr; + struct rds_ib_frmr *frmr = &ibmr->u.frmr; + struct rdma_cm_id *i_cm_id = ibmr->ic->i_cm_id; + int ret = -EINVAL; + + if (!i_cm_id || !i_cm_id->qp || !frmr->mr) + goto out; + + if (frmr->fr_state != FRMR_IS_INUSE) + goto out; + + while (atomic_dec_return(&ibmr->ic->i_fastunreg_wrs) <= 0) { + atomic_inc(&ibmr->ic->i_fastunreg_wrs); + cpu_relax(); + } + + frmr->fr_inv = true; + s_wr = &frmr->fr_wr; + + memset(s_wr, 0, sizeof(*s_wr)); + s_wr->wr_id = (unsigned long)(void *)ibmr; + s_wr->opcode = IB_WR_LOCAL_INV; + s_wr->ex.invalidate_rkey = frmr->mr->rkey; + s_wr->send_flags = IB_SEND_SIGNALED; + + ret = ib_post_send(i_cm_id->qp, s_wr, NULL); + if (unlikely(ret)) { + frmr->fr_state = FRMR_IS_STALE; + frmr->fr_inv = false; + atomic_inc(&ibmr->ic->i_fastunreg_wrs); + pr_err("RDS/IB: %s returned error(%d)\n", __func__, ret); + goto out; + } +out: + return ret; +} + +void rds_ib_mr_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc) +{ + struct rds_ib_mr *ibmr = (void *)(unsigned long)wc->wr_id; + struct rds_ib_frmr *frmr = &ibmr->u.frmr; + + if (wc->status != IB_WC_SUCCESS) { + frmr->fr_state = FRMR_IS_STALE; + if (rds_conn_up(ic->conn)) + rds_ib_conn_error(ic->conn, + "frmr completion <%pI4,%pI4> status %u(%s), vendor_err 0x%x, disconnecting and reconnecting\n", + &ic->conn->c_laddr, + &ic->conn->c_faddr, + wc->status, + ib_wc_status_msg(wc->status), + wc->vendor_err); + } + + if (frmr->fr_inv) { + frmr->fr_state = FRMR_IS_FREE; + frmr->fr_inv = false; + atomic_inc(&ic->i_fastreg_wrs); + } else { + atomic_inc(&ic->i_fastunreg_wrs); + } +} + +void rds_ib_unreg_frmr(struct list_head *list, unsigned int *nfreed, + unsigned long *unpinned, unsigned int goal) +{ + struct rds_ib_mr *ibmr, *next; + struct rds_ib_frmr *frmr; + int ret = 0; + unsigned int freed = *nfreed; + + /* String all ib_mr's onto one list and hand them to ib_unmap_fmr */ + list_for_each_entry(ibmr, list, unmap_list) { + if (ibmr->sg_dma_len) + ret |= rds_ib_post_inv(ibmr); + } + if (ret) + pr_warn("RDS/IB: %s failed (err=%d)\n", __func__, ret); + + /* Now we can destroy the DMA mapping and unpin any pages */ + list_for_each_entry_safe(ibmr, next, list, unmap_list) { + *unpinned += ibmr->sg_len; + frmr = &ibmr->u.frmr; + __rds_ib_teardown_mr(ibmr); + if (freed < goal || frmr->fr_state == FRMR_IS_STALE) { + /* Don't de-allocate if the MR is not free yet */ + if (frmr->fr_state == FRMR_IS_INUSE) + continue; + + if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL) + rds_ib_stats_inc(s_ib_rdma_mr_8k_free); + else + rds_ib_stats_inc(s_ib_rdma_mr_1m_free); + list_del(&ibmr->unmap_list); + if (frmr->mr) + ib_dereg_mr(frmr->mr); + kfree(ibmr); + freed++; + } + } + *nfreed = freed; +} + +struct rds_ib_mr *rds_ib_reg_frmr(struct rds_ib_device *rds_ibdev, + struct rds_ib_connection *ic, + struct scatterlist *sg, + unsigned long nents, u32 *key) +{ + struct rds_ib_mr *ibmr = NULL; + struct rds_ib_frmr *frmr; + int ret; + + if (!ic) { + /* TODO: Add FRWR support for RDS_GET_MR using proxy qp*/ + return ERR_PTR(-EOPNOTSUPP); + } + + do { + if (ibmr) + rds_ib_free_frmr(ibmr, true); + ibmr = rds_ib_alloc_frmr(rds_ibdev, nents); + if (IS_ERR(ibmr)) + return ibmr; + frmr = &ibmr->u.frmr; + } while (frmr->fr_state != FRMR_IS_FREE); + + ibmr->ic = ic; + ibmr->device = rds_ibdev; + ret = rds_ib_map_frmr(rds_ibdev, ibmr->pool, ibmr, sg, nents); + if (ret == 0) { + *key = frmr->mr->rkey; + } else { + rds_ib_free_frmr(ibmr, false); + ibmr = ERR_PTR(ret); + } + + return ibmr; +} + +void rds_ib_free_frmr_list(struct rds_ib_mr *ibmr) +{ + struct rds_ib_mr_pool *pool = ibmr->pool; + struct rds_ib_frmr *frmr = &ibmr->u.frmr; + + if (frmr->fr_state == FRMR_IS_STALE) + llist_add(&ibmr->llnode, &pool->drop_list); + else + llist_add(&ibmr->llnode, &pool->free_list); +} diff --git a/net/rds/ib_mr.h b/net/rds/ib_mr.h new file mode 100644 index 000000000..5da12c248 --- /dev/null +++ b/net/rds/ib_mr.h @@ -0,0 +1,146 @@ +/* + * Copyright (c) 2016 Oracle. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef _RDS_IB_MR_H +#define _RDS_IB_MR_H + +#include <linux/kernel.h> + +#include "rds.h" +#include "ib.h" + +#define RDS_MR_1M_POOL_SIZE (8192 / 2) +#define RDS_MR_1M_MSG_SIZE 256 +#define RDS_MR_8K_MSG_SIZE 2 +#define RDS_MR_8K_SCALE (256 / (RDS_MR_8K_MSG_SIZE + 1)) +#define RDS_MR_8K_POOL_SIZE (RDS_MR_8K_SCALE * (8192 / 2)) + +struct rds_ib_fmr { + struct ib_fmr *fmr; +}; + +enum rds_ib_fr_state { + FRMR_IS_FREE, /* mr invalidated & ready for use */ + FRMR_IS_INUSE, /* mr is in use or used & can be invalidated */ + FRMR_IS_STALE, /* Stale MR and needs to be dropped */ +}; + +struct rds_ib_frmr { + struct ib_mr *mr; + enum rds_ib_fr_state fr_state; + bool fr_inv; + struct ib_send_wr fr_wr; + unsigned int dma_npages; + unsigned int sg_byte_len; +}; + +/* This is stored as mr->r_trans_private. */ +struct rds_ib_mr { + struct rds_ib_device *device; + struct rds_ib_mr_pool *pool; + struct rds_ib_connection *ic; + + struct llist_node llnode; + + /* unmap_list is for freeing */ + struct list_head unmap_list; + unsigned int remap_count; + + struct scatterlist *sg; + unsigned int sg_len; + int sg_dma_len; + + union { + struct rds_ib_fmr fmr; + struct rds_ib_frmr frmr; + } u; +}; + +/* Our own little MR pool */ +struct rds_ib_mr_pool { + unsigned int pool_type; + struct mutex flush_lock; /* serialize fmr invalidate */ + struct delayed_work flush_worker; /* flush worker */ + + atomic_t item_count; /* total # of MRs */ + atomic_t dirty_count; /* # dirty of MRs */ + + struct llist_head drop_list; /* MRs not reached max_maps */ + struct llist_head free_list; /* unused MRs */ + struct llist_head clean_list; /* unused & unmapped MRs */ + wait_queue_head_t flush_wait; + + atomic_t free_pinned; /* memory pinned by free MRs */ + unsigned long max_items; + unsigned long max_items_soft; + unsigned long max_free_pinned; + struct ib_fmr_attr fmr_attr; + bool use_fastreg; +}; + +extern struct workqueue_struct *rds_ib_mr_wq; +extern bool prefer_frmr; + +struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_dev, + int npages); +void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, + struct rds_info_rdma_connection *iinfo); +void rds6_ib_get_mr_info(struct rds_ib_device *rds_ibdev, + struct rds6_info_rdma_connection *iinfo6); +void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *); +void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents, + struct rds_sock *rs, u32 *key_ret, + struct rds_connection *conn); +void rds_ib_sync_mr(void *trans_private, int dir); +void rds_ib_free_mr(void *trans_private, int invalidate); +void rds_ib_flush_mrs(void); +int rds_ib_mr_init(void); +void rds_ib_mr_exit(void); + +void __rds_ib_teardown_mr(struct rds_ib_mr *); +void rds_ib_teardown_mr(struct rds_ib_mr *); +struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *, int); +struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *); +int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *, int, struct rds_ib_mr **); +struct rds_ib_mr *rds_ib_reg_fmr(struct rds_ib_device *, struct scatterlist *, + unsigned long, u32 *); +struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *); +void rds_ib_unreg_fmr(struct list_head *, unsigned int *, + unsigned long *, unsigned int); +void rds_ib_free_fmr_list(struct rds_ib_mr *); +struct rds_ib_mr *rds_ib_reg_frmr(struct rds_ib_device *rds_ibdev, + struct rds_ib_connection *ic, + struct scatterlist *sg, + unsigned long nents, u32 *key); +void rds_ib_unreg_frmr(struct list_head *list, unsigned int *nfreed, + unsigned long *unpinned, unsigned int goal); +void rds_ib_free_frmr_list(struct rds_ib_mr *); +#endif diff --git a/net/rds/ib_rdma.c b/net/rds/ib_rdma.c new file mode 100644 index 000000000..0b347f46b --- /dev/null +++ b/net/rds/ib_rdma.c @@ -0,0 +1,651 @@ +/* + * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/slab.h> +#include <linux/rculist.h> +#include <linux/llist.h> + +#include "rds_single_path.h" +#include "ib_mr.h" + +struct workqueue_struct *rds_ib_mr_wq; + +static DEFINE_PER_CPU(unsigned long, clean_list_grace); +#define CLEAN_LIST_BUSY_BIT 0 + +static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr) +{ + struct rds_ib_device *rds_ibdev; + struct rds_ib_ipaddr *i_ipaddr; + + rcu_read_lock(); + list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) { + list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) { + if (i_ipaddr->ipaddr == ipaddr) { + refcount_inc(&rds_ibdev->refcount); + rcu_read_unlock(); + return rds_ibdev; + } + } + } + rcu_read_unlock(); + + return NULL; +} + +static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr) +{ + struct rds_ib_ipaddr *i_ipaddr; + + i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL); + if (!i_ipaddr) + return -ENOMEM; + + i_ipaddr->ipaddr = ipaddr; + + spin_lock_irq(&rds_ibdev->spinlock); + list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list); + spin_unlock_irq(&rds_ibdev->spinlock); + + return 0; +} + +static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr) +{ + struct rds_ib_ipaddr *i_ipaddr; + struct rds_ib_ipaddr *to_free = NULL; + + + spin_lock_irq(&rds_ibdev->spinlock); + list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) { + if (i_ipaddr->ipaddr == ipaddr) { + list_del_rcu(&i_ipaddr->list); + to_free = i_ipaddr; + break; + } + } + spin_unlock_irq(&rds_ibdev->spinlock); + + if (to_free) + kfree_rcu(to_free, rcu); +} + +int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, + struct in6_addr *ipaddr) +{ + struct rds_ib_device *rds_ibdev_old; + + rds_ibdev_old = rds_ib_get_device(ipaddr->s6_addr32[3]); + if (!rds_ibdev_old) + return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]); + + if (rds_ibdev_old != rds_ibdev) { + rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr->s6_addr32[3]); + rds_ib_dev_put(rds_ibdev_old); + return rds_ib_add_ipaddr(rds_ibdev, ipaddr->s6_addr32[3]); + } + rds_ib_dev_put(rds_ibdev_old); + + return 0; +} + +void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + + /* conn was previously on the nodev_conns_list */ + spin_lock_irq(&ib_nodev_conns_lock); + BUG_ON(list_empty(&ib_nodev_conns)); + BUG_ON(list_empty(&ic->ib_node)); + list_del(&ic->ib_node); + + spin_lock(&rds_ibdev->spinlock); + list_add_tail(&ic->ib_node, &rds_ibdev->conn_list); + spin_unlock(&rds_ibdev->spinlock); + spin_unlock_irq(&ib_nodev_conns_lock); + + ic->rds_ibdev = rds_ibdev; + refcount_inc(&rds_ibdev->refcount); +} + +void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + + /* place conn on nodev_conns_list */ + spin_lock(&ib_nodev_conns_lock); + + spin_lock_irq(&rds_ibdev->spinlock); + BUG_ON(list_empty(&ic->ib_node)); + list_del(&ic->ib_node); + spin_unlock_irq(&rds_ibdev->spinlock); + + list_add_tail(&ic->ib_node, &ib_nodev_conns); + + spin_unlock(&ib_nodev_conns_lock); + + ic->rds_ibdev = NULL; + rds_ib_dev_put(rds_ibdev); +} + +void rds_ib_destroy_nodev_conns(void) +{ + struct rds_ib_connection *ic, *_ic; + LIST_HEAD(tmp_list); + + /* avoid calling conn_destroy with irqs off */ + spin_lock_irq(&ib_nodev_conns_lock); + list_splice(&ib_nodev_conns, &tmp_list); + spin_unlock_irq(&ib_nodev_conns_lock); + + list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node) + rds_conn_destroy(ic->conn); +} + +void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo) +{ + struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool; + + iinfo->rdma_mr_max = pool_1m->max_items; + iinfo->rdma_mr_size = pool_1m->fmr_attr.max_pages; +} + +#if IS_ENABLED(CONFIG_IPV6) +void rds6_ib_get_mr_info(struct rds_ib_device *rds_ibdev, + struct rds6_info_rdma_connection *iinfo6) +{ + struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool; + + iinfo6->rdma_mr_max = pool_1m->max_items; + iinfo6->rdma_mr_size = pool_1m->fmr_attr.max_pages; +} +#endif + +struct rds_ib_mr *rds_ib_reuse_mr(struct rds_ib_mr_pool *pool) +{ + struct rds_ib_mr *ibmr = NULL; + struct llist_node *ret; + unsigned long *flag; + + preempt_disable(); + flag = this_cpu_ptr(&clean_list_grace); + set_bit(CLEAN_LIST_BUSY_BIT, flag); + ret = llist_del_first(&pool->clean_list); + if (ret) { + ibmr = llist_entry(ret, struct rds_ib_mr, llnode); + if (pool->pool_type == RDS_IB_MR_8K_POOL) + rds_ib_stats_inc(s_ib_rdma_mr_8k_reused); + else + rds_ib_stats_inc(s_ib_rdma_mr_1m_reused); + } + + clear_bit(CLEAN_LIST_BUSY_BIT, flag); + preempt_enable(); + return ibmr; +} + +static inline void wait_clean_list_grace(void) +{ + int cpu; + unsigned long *flag; + + for_each_online_cpu(cpu) { + flag = &per_cpu(clean_list_grace, cpu); + while (test_bit(CLEAN_LIST_BUSY_BIT, flag)) + cpu_relax(); + } +} + +void rds_ib_sync_mr(void *trans_private, int direction) +{ + struct rds_ib_mr *ibmr = trans_private; + struct rds_ib_device *rds_ibdev = ibmr->device; + + switch (direction) { + case DMA_FROM_DEVICE: + ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg, + ibmr->sg_dma_len, DMA_BIDIRECTIONAL); + break; + case DMA_TO_DEVICE: + ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg, + ibmr->sg_dma_len, DMA_BIDIRECTIONAL); + break; + } +} + +void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr) +{ + struct rds_ib_device *rds_ibdev = ibmr->device; + + if (ibmr->sg_dma_len) { + ib_dma_unmap_sg(rds_ibdev->dev, + ibmr->sg, ibmr->sg_len, + DMA_BIDIRECTIONAL); + ibmr->sg_dma_len = 0; + } + + /* Release the s/g list */ + if (ibmr->sg_len) { + unsigned int i; + + for (i = 0; i < ibmr->sg_len; ++i) { + struct page *page = sg_page(&ibmr->sg[i]); + + /* FIXME we need a way to tell a r/w MR + * from a r/o MR */ + WARN_ON(!page->mapping && irqs_disabled()); + set_page_dirty(page); + put_page(page); + } + kfree(ibmr->sg); + + ibmr->sg = NULL; + ibmr->sg_len = 0; + } +} + +void rds_ib_teardown_mr(struct rds_ib_mr *ibmr) +{ + unsigned int pinned = ibmr->sg_len; + + __rds_ib_teardown_mr(ibmr); + if (pinned) { + struct rds_ib_mr_pool *pool = ibmr->pool; + + atomic_sub(pinned, &pool->free_pinned); + } +} + +static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all) +{ + unsigned int item_count; + + item_count = atomic_read(&pool->item_count); + if (free_all) + return item_count; + + return 0; +} + +/* + * given an llist of mrs, put them all into the list_head for more processing + */ +static unsigned int llist_append_to_list(struct llist_head *llist, + struct list_head *list) +{ + struct rds_ib_mr *ibmr; + struct llist_node *node; + struct llist_node *next; + unsigned int count = 0; + + node = llist_del_all(llist); + while (node) { + next = node->next; + ibmr = llist_entry(node, struct rds_ib_mr, llnode); + list_add_tail(&ibmr->unmap_list, list); + node = next; + count++; + } + return count; +} + +/* + * this takes a list head of mrs and turns it into linked llist nodes + * of clusters. Each cluster has linked llist nodes of + * MR_CLUSTER_SIZE mrs that are ready for reuse. + */ +static void list_to_llist_nodes(struct rds_ib_mr_pool *pool, + struct list_head *list, + struct llist_node **nodes_head, + struct llist_node **nodes_tail) +{ + struct rds_ib_mr *ibmr; + struct llist_node *cur = NULL; + struct llist_node **next = nodes_head; + + list_for_each_entry(ibmr, list, unmap_list) { + cur = &ibmr->llnode; + *next = cur; + next = &cur->next; + } + *next = NULL; + *nodes_tail = cur; +} + +/* + * Flush our pool of MRs. + * At a minimum, all currently unused MRs are unmapped. + * If the number of MRs allocated exceeds the limit, we also try + * to free as many MRs as needed to get back to this limit. + */ +int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, + int free_all, struct rds_ib_mr **ibmr_ret) +{ + struct rds_ib_mr *ibmr; + struct llist_node *clean_nodes; + struct llist_node *clean_tail; + LIST_HEAD(unmap_list); + unsigned long unpinned = 0; + unsigned int nfreed = 0, dirty_to_clean = 0, free_goal; + + if (pool->pool_type == RDS_IB_MR_8K_POOL) + rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush); + else + rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush); + + if (ibmr_ret) { + DEFINE_WAIT(wait); + while (!mutex_trylock(&pool->flush_lock)) { + ibmr = rds_ib_reuse_mr(pool); + if (ibmr) { + *ibmr_ret = ibmr; + finish_wait(&pool->flush_wait, &wait); + goto out_nolock; + } + + prepare_to_wait(&pool->flush_wait, &wait, + TASK_UNINTERRUPTIBLE); + if (llist_empty(&pool->clean_list)) + schedule(); + + ibmr = rds_ib_reuse_mr(pool); + if (ibmr) { + *ibmr_ret = ibmr; + finish_wait(&pool->flush_wait, &wait); + goto out_nolock; + } + } + finish_wait(&pool->flush_wait, &wait); + } else + mutex_lock(&pool->flush_lock); + + if (ibmr_ret) { + ibmr = rds_ib_reuse_mr(pool); + if (ibmr) { + *ibmr_ret = ibmr; + goto out; + } + } + + /* Get the list of all MRs to be dropped. Ordering matters - + * we want to put drop_list ahead of free_list. + */ + dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list); + dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list); + if (free_all) + llist_append_to_list(&pool->clean_list, &unmap_list); + + free_goal = rds_ib_flush_goal(pool, free_all); + + if (list_empty(&unmap_list)) + goto out; + + if (pool->use_fastreg) + rds_ib_unreg_frmr(&unmap_list, &nfreed, &unpinned, free_goal); + else + rds_ib_unreg_fmr(&unmap_list, &nfreed, &unpinned, free_goal); + + if (!list_empty(&unmap_list)) { + /* we have to make sure that none of the things we're about + * to put on the clean list would race with other cpus trying + * to pull items off. The llist would explode if we managed to + * remove something from the clean list and then add it back again + * while another CPU was spinning on that same item in llist_del_first. + * + * This is pretty unlikely, but just in case wait for an llist grace period + * here before adding anything back into the clean list. + */ + wait_clean_list_grace(); + + list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail); + if (ibmr_ret) { + *ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode); + clean_nodes = clean_nodes->next; + } + /* more than one entry in llist nodes */ + if (clean_nodes) + llist_add_batch(clean_nodes, clean_tail, + &pool->clean_list); + + } + + atomic_sub(unpinned, &pool->free_pinned); + atomic_sub(dirty_to_clean, &pool->dirty_count); + atomic_sub(nfreed, &pool->item_count); + +out: + mutex_unlock(&pool->flush_lock); + if (waitqueue_active(&pool->flush_wait)) + wake_up(&pool->flush_wait); +out_nolock: + return 0; +} + +struct rds_ib_mr *rds_ib_try_reuse_ibmr(struct rds_ib_mr_pool *pool) +{ + struct rds_ib_mr *ibmr = NULL; + int iter = 0; + + while (1) { + ibmr = rds_ib_reuse_mr(pool); + if (ibmr) + return ibmr; + + if (atomic_inc_return(&pool->item_count) <= pool->max_items) + break; + + atomic_dec(&pool->item_count); + + if (++iter > 2) { + if (pool->pool_type == RDS_IB_MR_8K_POOL) + rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted); + else + rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted); + return ERR_PTR(-EAGAIN); + } + + /* We do have some empty MRs. Flush them out. */ + if (pool->pool_type == RDS_IB_MR_8K_POOL) + rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait); + else + rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait); + + rds_ib_flush_mr_pool(pool, 0, &ibmr); + if (ibmr) + return ibmr; + } + + return ibmr; +} + +static void rds_ib_mr_pool_flush_worker(struct work_struct *work) +{ + struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work); + + rds_ib_flush_mr_pool(pool, 0, NULL); +} + +void rds_ib_free_mr(void *trans_private, int invalidate) +{ + struct rds_ib_mr *ibmr = trans_private; + struct rds_ib_mr_pool *pool = ibmr->pool; + struct rds_ib_device *rds_ibdev = ibmr->device; + + rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len); + + /* Return it to the pool's free list */ + if (rds_ibdev->use_fastreg) + rds_ib_free_frmr_list(ibmr); + else + rds_ib_free_fmr_list(ibmr); + + atomic_add(ibmr->sg_len, &pool->free_pinned); + atomic_inc(&pool->dirty_count); + + /* If we've pinned too many pages, request a flush */ + if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned || + atomic_read(&pool->dirty_count) >= pool->max_items / 5) + queue_delayed_work(rds_ib_mr_wq, &pool->flush_worker, 10); + + if (invalidate) { + if (likely(!in_interrupt())) { + rds_ib_flush_mr_pool(pool, 0, NULL); + } else { + /* We get here if the user created a MR marked + * as use_once and invalidate at the same time. + */ + queue_delayed_work(rds_ib_mr_wq, + &pool->flush_worker, 10); + } + } + + rds_ib_dev_put(rds_ibdev); +} + +void rds_ib_flush_mrs(void) +{ + struct rds_ib_device *rds_ibdev; + + down_read(&rds_ib_devices_lock); + list_for_each_entry(rds_ibdev, &rds_ib_devices, list) { + if (rds_ibdev->mr_8k_pool) + rds_ib_flush_mr_pool(rds_ibdev->mr_8k_pool, 0, NULL); + + if (rds_ibdev->mr_1m_pool) + rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL); + } + up_read(&rds_ib_devices_lock); +} + +void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents, + struct rds_sock *rs, u32 *key_ret, + struct rds_connection *conn) +{ + struct rds_ib_device *rds_ibdev; + struct rds_ib_mr *ibmr = NULL; + struct rds_ib_connection *ic = NULL; + int ret; + + rds_ibdev = rds_ib_get_device(rs->rs_bound_addr.s6_addr32[3]); + if (!rds_ibdev) { + ret = -ENODEV; + goto out; + } + + if (conn) + ic = conn->c_transport_data; + + if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) { + ret = -ENODEV; + goto out; + } + + if (rds_ibdev->use_fastreg) + ibmr = rds_ib_reg_frmr(rds_ibdev, ic, sg, nents, key_ret); + else + ibmr = rds_ib_reg_fmr(rds_ibdev, sg, nents, key_ret); + if (IS_ERR(ibmr)) { + ret = PTR_ERR(ibmr); + pr_warn("RDS/IB: rds_ib_get_mr failed (errno=%d)\n", ret); + } else { + return ibmr; + } + + out: + if (rds_ibdev) + rds_ib_dev_put(rds_ibdev); + + return ERR_PTR(ret); +} + +void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool) +{ + cancel_delayed_work_sync(&pool->flush_worker); + rds_ib_flush_mr_pool(pool, 1, NULL); + WARN_ON(atomic_read(&pool->item_count)); + WARN_ON(atomic_read(&pool->free_pinned)); + kfree(pool); +} + +struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev, + int pool_type) +{ + struct rds_ib_mr_pool *pool; + + pool = kzalloc(sizeof(*pool), GFP_KERNEL); + if (!pool) + return ERR_PTR(-ENOMEM); + + pool->pool_type = pool_type; + init_llist_head(&pool->free_list); + init_llist_head(&pool->drop_list); + init_llist_head(&pool->clean_list); + mutex_init(&pool->flush_lock); + init_waitqueue_head(&pool->flush_wait); + INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker); + + if (pool_type == RDS_IB_MR_1M_POOL) { + /* +1 allows for unaligned MRs */ + pool->fmr_attr.max_pages = RDS_MR_1M_MSG_SIZE + 1; + pool->max_items = rds_ibdev->max_1m_mrs; + } else { + /* pool_type == RDS_IB_MR_8K_POOL */ + pool->fmr_attr.max_pages = RDS_MR_8K_MSG_SIZE + 1; + pool->max_items = rds_ibdev->max_8k_mrs; + } + + pool->max_free_pinned = pool->max_items * pool->fmr_attr.max_pages / 4; + pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps; + pool->fmr_attr.page_shift = PAGE_SHIFT; + pool->max_items_soft = rds_ibdev->max_mrs * 3 / 4; + pool->use_fastreg = rds_ibdev->use_fastreg; + + return pool; +} + +int rds_ib_mr_init(void) +{ + rds_ib_mr_wq = alloc_workqueue("rds_mr_flushd", WQ_MEM_RECLAIM, 0); + if (!rds_ib_mr_wq) + return -ENOMEM; + return 0; +} + +/* By the time this is called all the IB devices should have been torn down and + * had their pools freed. As each pool is freed its work struct is waited on, + * so the pool flushing work queue should be idle by the time we get here. + */ +void rds_ib_mr_exit(void) +{ + destroy_workqueue(rds_ib_mr_wq); +} diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c new file mode 100644 index 000000000..2f16146e4 --- /dev/null +++ b/net/rds/ib_recv.c @@ -0,0 +1,1066 @@ +/* + * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/slab.h> +#include <linux/pci.h> +#include <linux/dma-mapping.h> +#include <rdma/rdma_cm.h> + +#include "rds_single_path.h" +#include "rds.h" +#include "ib.h" + +static struct kmem_cache *rds_ib_incoming_slab; +static struct kmem_cache *rds_ib_frag_slab; +static atomic_t rds_ib_allocation = ATOMIC_INIT(0); + +void rds_ib_recv_init_ring(struct rds_ib_connection *ic) +{ + struct rds_ib_recv_work *recv; + u32 i; + + for (i = 0, recv = ic->i_recvs; i < ic->i_recv_ring.w_nr; i++, recv++) { + struct ib_sge *sge; + + recv->r_ibinc = NULL; + recv->r_frag = NULL; + + recv->r_wr.next = NULL; + recv->r_wr.wr_id = i; + recv->r_wr.sg_list = recv->r_sge; + recv->r_wr.num_sge = RDS_IB_RECV_SGE; + + sge = &recv->r_sge[0]; + sge->addr = ic->i_recv_hdrs_dma + (i * sizeof(struct rds_header)); + sge->length = sizeof(struct rds_header); + sge->lkey = ic->i_pd->local_dma_lkey; + + sge = &recv->r_sge[1]; + sge->addr = 0; + sge->length = RDS_FRAG_SIZE; + sge->lkey = ic->i_pd->local_dma_lkey; + } +} + +/* + * The entire 'from' list, including the from element itself, is put on + * to the tail of the 'to' list. + */ +static void list_splice_entire_tail(struct list_head *from, + struct list_head *to) +{ + struct list_head *from_last = from->prev; + + list_splice_tail(from_last, to); + list_add_tail(from_last, to); +} + +static void rds_ib_cache_xfer_to_ready(struct rds_ib_refill_cache *cache) +{ + struct list_head *tmp; + + tmp = xchg(&cache->xfer, NULL); + if (tmp) { + if (cache->ready) + list_splice_entire_tail(tmp, cache->ready); + else + cache->ready = tmp; + } +} + +static int rds_ib_recv_alloc_cache(struct rds_ib_refill_cache *cache, gfp_t gfp) +{ + struct rds_ib_cache_head *head; + int cpu; + + cache->percpu = alloc_percpu_gfp(struct rds_ib_cache_head, gfp); + if (!cache->percpu) + return -ENOMEM; + + for_each_possible_cpu(cpu) { + head = per_cpu_ptr(cache->percpu, cpu); + head->first = NULL; + head->count = 0; + } + cache->xfer = NULL; + cache->ready = NULL; + + return 0; +} + +int rds_ib_recv_alloc_caches(struct rds_ib_connection *ic, gfp_t gfp) +{ + int ret; + + ret = rds_ib_recv_alloc_cache(&ic->i_cache_incs, gfp); + if (!ret) { + ret = rds_ib_recv_alloc_cache(&ic->i_cache_frags, gfp); + if (ret) + free_percpu(ic->i_cache_incs.percpu); + } + + return ret; +} + +static void rds_ib_cache_splice_all_lists(struct rds_ib_refill_cache *cache, + struct list_head *caller_list) +{ + struct rds_ib_cache_head *head; + int cpu; + + for_each_possible_cpu(cpu) { + head = per_cpu_ptr(cache->percpu, cpu); + if (head->first) { + list_splice_entire_tail(head->first, caller_list); + head->first = NULL; + } + } + + if (cache->ready) { + list_splice_entire_tail(cache->ready, caller_list); + cache->ready = NULL; + } +} + +void rds_ib_recv_free_caches(struct rds_ib_connection *ic) +{ + struct rds_ib_incoming *inc; + struct rds_ib_incoming *inc_tmp; + struct rds_page_frag *frag; + struct rds_page_frag *frag_tmp; + LIST_HEAD(list); + + rds_ib_cache_xfer_to_ready(&ic->i_cache_incs); + rds_ib_cache_splice_all_lists(&ic->i_cache_incs, &list); + free_percpu(ic->i_cache_incs.percpu); + + list_for_each_entry_safe(inc, inc_tmp, &list, ii_cache_entry) { + list_del(&inc->ii_cache_entry); + WARN_ON(!list_empty(&inc->ii_frags)); + kmem_cache_free(rds_ib_incoming_slab, inc); + } + + rds_ib_cache_xfer_to_ready(&ic->i_cache_frags); + rds_ib_cache_splice_all_lists(&ic->i_cache_frags, &list); + free_percpu(ic->i_cache_frags.percpu); + + list_for_each_entry_safe(frag, frag_tmp, &list, f_cache_entry) { + list_del(&frag->f_cache_entry); + WARN_ON(!list_empty(&frag->f_item)); + kmem_cache_free(rds_ib_frag_slab, frag); + } +} + +/* fwd decl */ +static void rds_ib_recv_cache_put(struct list_head *new_item, + struct rds_ib_refill_cache *cache); +static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache); + + +/* Recycle frag and attached recv buffer f_sg */ +static void rds_ib_frag_free(struct rds_ib_connection *ic, + struct rds_page_frag *frag) +{ + rdsdebug("frag %p page %p\n", frag, sg_page(&frag->f_sg)); + + rds_ib_recv_cache_put(&frag->f_cache_entry, &ic->i_cache_frags); + atomic_add(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs); + rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE); +} + +/* Recycle inc after freeing attached frags */ +void rds_ib_inc_free(struct rds_incoming *inc) +{ + struct rds_ib_incoming *ibinc; + struct rds_page_frag *frag; + struct rds_page_frag *pos; + struct rds_ib_connection *ic = inc->i_conn->c_transport_data; + + ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); + + /* Free attached frags */ + list_for_each_entry_safe(frag, pos, &ibinc->ii_frags, f_item) { + list_del_init(&frag->f_item); + rds_ib_frag_free(ic, frag); + } + BUG_ON(!list_empty(&ibinc->ii_frags)); + + rdsdebug("freeing ibinc %p inc %p\n", ibinc, inc); + rds_ib_recv_cache_put(&ibinc->ii_cache_entry, &ic->i_cache_incs); +} + +static void rds_ib_recv_clear_one(struct rds_ib_connection *ic, + struct rds_ib_recv_work *recv) +{ + if (recv->r_ibinc) { + rds_inc_put(&recv->r_ibinc->ii_inc); + recv->r_ibinc = NULL; + } + if (recv->r_frag) { + ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, DMA_FROM_DEVICE); + rds_ib_frag_free(ic, recv->r_frag); + recv->r_frag = NULL; + } +} + +void rds_ib_recv_clear_ring(struct rds_ib_connection *ic) +{ + u32 i; + + for (i = 0; i < ic->i_recv_ring.w_nr; i++) + rds_ib_recv_clear_one(ic, &ic->i_recvs[i]); +} + +static struct rds_ib_incoming *rds_ib_refill_one_inc(struct rds_ib_connection *ic, + gfp_t slab_mask) +{ + struct rds_ib_incoming *ibinc; + struct list_head *cache_item; + int avail_allocs; + + cache_item = rds_ib_recv_cache_get(&ic->i_cache_incs); + if (cache_item) { + ibinc = container_of(cache_item, struct rds_ib_incoming, ii_cache_entry); + } else { + avail_allocs = atomic_add_unless(&rds_ib_allocation, + 1, rds_ib_sysctl_max_recv_allocation); + if (!avail_allocs) { + rds_ib_stats_inc(s_ib_rx_alloc_limit); + return NULL; + } + ibinc = kmem_cache_alloc(rds_ib_incoming_slab, slab_mask); + if (!ibinc) { + atomic_dec(&rds_ib_allocation); + return NULL; + } + rds_ib_stats_inc(s_ib_rx_total_incs); + } + INIT_LIST_HEAD(&ibinc->ii_frags); + rds_inc_init(&ibinc->ii_inc, ic->conn, &ic->conn->c_faddr); + + return ibinc; +} + +static struct rds_page_frag *rds_ib_refill_one_frag(struct rds_ib_connection *ic, + gfp_t slab_mask, gfp_t page_mask) +{ + struct rds_page_frag *frag; + struct list_head *cache_item; + int ret; + + cache_item = rds_ib_recv_cache_get(&ic->i_cache_frags); + if (cache_item) { + frag = container_of(cache_item, struct rds_page_frag, f_cache_entry); + atomic_sub(RDS_FRAG_SIZE / SZ_1K, &ic->i_cache_allocs); + rds_ib_stats_add(s_ib_recv_added_to_cache, RDS_FRAG_SIZE); + } else { + frag = kmem_cache_alloc(rds_ib_frag_slab, slab_mask); + if (!frag) + return NULL; + + sg_init_table(&frag->f_sg, 1); + ret = rds_page_remainder_alloc(&frag->f_sg, + RDS_FRAG_SIZE, page_mask); + if (ret) { + kmem_cache_free(rds_ib_frag_slab, frag); + return NULL; + } + rds_ib_stats_inc(s_ib_rx_total_frags); + } + + INIT_LIST_HEAD(&frag->f_item); + + return frag; +} + +static int rds_ib_recv_refill_one(struct rds_connection *conn, + struct rds_ib_recv_work *recv, gfp_t gfp) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + struct ib_sge *sge; + int ret = -ENOMEM; + gfp_t slab_mask = GFP_NOWAIT; + gfp_t page_mask = GFP_NOWAIT; + + if (gfp & __GFP_DIRECT_RECLAIM) { + slab_mask = GFP_KERNEL; + page_mask = GFP_HIGHUSER; + } + + if (!ic->i_cache_incs.ready) + rds_ib_cache_xfer_to_ready(&ic->i_cache_incs); + if (!ic->i_cache_frags.ready) + rds_ib_cache_xfer_to_ready(&ic->i_cache_frags); + + /* + * ibinc was taken from recv if recv contained the start of a message. + * recvs that were continuations will still have this allocated. + */ + if (!recv->r_ibinc) { + recv->r_ibinc = rds_ib_refill_one_inc(ic, slab_mask); + if (!recv->r_ibinc) + goto out; + } + + WARN_ON(recv->r_frag); /* leak! */ + recv->r_frag = rds_ib_refill_one_frag(ic, slab_mask, page_mask); + if (!recv->r_frag) + goto out; + + ret = ib_dma_map_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, + 1, DMA_FROM_DEVICE); + WARN_ON(ret != 1); + + sge = &recv->r_sge[0]; + sge->addr = ic->i_recv_hdrs_dma + (recv - ic->i_recvs) * sizeof(struct rds_header); + sge->length = sizeof(struct rds_header); + + sge = &recv->r_sge[1]; + sge->addr = ib_sg_dma_address(ic->i_cm_id->device, &recv->r_frag->f_sg); + sge->length = ib_sg_dma_len(ic->i_cm_id->device, &recv->r_frag->f_sg); + + ret = 0; +out: + return ret; +} + +static int acquire_refill(struct rds_connection *conn) +{ + return test_and_set_bit(RDS_RECV_REFILL, &conn->c_flags) == 0; +} + +static void release_refill(struct rds_connection *conn) +{ + clear_bit(RDS_RECV_REFILL, &conn->c_flags); + + /* We don't use wait_on_bit()/wake_up_bit() because our waking is in a + * hot path and finding waiters is very rare. We don't want to walk + * the system-wide hashed waitqueue buckets in the fast path only to + * almost never find waiters. + */ + if (waitqueue_active(&conn->c_waitq)) + wake_up_all(&conn->c_waitq); +} + +/* + * This tries to allocate and post unused work requests after making sure that + * they have all the allocations they need to queue received fragments into + * sockets. + */ +void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + struct rds_ib_recv_work *recv; + unsigned int posted = 0; + int ret = 0; + bool can_wait = !!(gfp & __GFP_DIRECT_RECLAIM); + u32 pos; + + /* the goal here is to just make sure that someone, somewhere + * is posting buffers. If we can't get the refill lock, + * let them do their thing + */ + if (!acquire_refill(conn)) + return; + + while ((prefill || rds_conn_up(conn)) && + rds_ib_ring_alloc(&ic->i_recv_ring, 1, &pos)) { + if (pos >= ic->i_recv_ring.w_nr) { + printk(KERN_NOTICE "Argh - ring alloc returned pos=%u\n", + pos); + break; + } + + recv = &ic->i_recvs[pos]; + ret = rds_ib_recv_refill_one(conn, recv, gfp); + if (ret) { + break; + } + + rdsdebug("recv %p ibinc %p page %p addr %lu\n", recv, + recv->r_ibinc, sg_page(&recv->r_frag->f_sg), + (long) ib_sg_dma_address( + ic->i_cm_id->device, + &recv->r_frag->f_sg)); + + /* XXX when can this fail? */ + ret = ib_post_recv(ic->i_cm_id->qp, &recv->r_wr, NULL); + if (ret) { + rds_ib_conn_error(conn, "recv post on " + "%pI6c returned %d, disconnecting and " + "reconnecting\n", &conn->c_faddr, + ret); + break; + } + + posted++; + } + + /* We're doing flow control - update the window. */ + if (ic->i_flowctl && posted) + rds_ib_advertise_credits(conn, posted); + + if (ret) + rds_ib_ring_unalloc(&ic->i_recv_ring, 1); + + release_refill(conn); + + /* if we're called from the softirq handler, we'll be GFP_NOWAIT. + * in this case the ring being low is going to lead to more interrupts + * and we can safely let the softirq code take care of it unless the + * ring is completely empty. + * + * if we're called from krdsd, we'll be GFP_KERNEL. In this case + * we might have raced with the softirq code while we had the refill + * lock held. Use rds_ib_ring_low() instead of ring_empty to decide + * if we should requeue. + */ + if (rds_conn_up(conn) && + ((can_wait && rds_ib_ring_low(&ic->i_recv_ring)) || + rds_ib_ring_empty(&ic->i_recv_ring))) { + queue_delayed_work(rds_wq, &conn->c_recv_w, 1); + } +} + +/* + * We want to recycle several types of recv allocations, like incs and frags. + * To use this, the *_free() function passes in the ptr to a list_head within + * the recyclee, as well as the cache to put it on. + * + * First, we put the memory on a percpu list. When this reaches a certain size, + * We move it to an intermediate non-percpu list in a lockless manner, with some + * xchg/compxchg wizardry. + * + * N.B. Instead of a list_head as the anchor, we use a single pointer, which can + * be NULL and xchg'd. The list is actually empty when the pointer is NULL, and + * list_empty() will return true with one element is actually present. + */ +static void rds_ib_recv_cache_put(struct list_head *new_item, + struct rds_ib_refill_cache *cache) +{ + unsigned long flags; + struct list_head *old, *chpfirst; + + local_irq_save(flags); + + chpfirst = __this_cpu_read(cache->percpu->first); + if (!chpfirst) + INIT_LIST_HEAD(new_item); + else /* put on front */ + list_add_tail(new_item, chpfirst); + + __this_cpu_write(cache->percpu->first, new_item); + __this_cpu_inc(cache->percpu->count); + + if (__this_cpu_read(cache->percpu->count) < RDS_IB_RECYCLE_BATCH_COUNT) + goto end; + + /* + * Return our per-cpu first list to the cache's xfer by atomically + * grabbing the current xfer list, appending it to our per-cpu list, + * and then atomically returning that entire list back to the + * cache's xfer list as long as it's still empty. + */ + do { + old = xchg(&cache->xfer, NULL); + if (old) + list_splice_entire_tail(old, chpfirst); + old = cmpxchg(&cache->xfer, NULL, chpfirst); + } while (old); + + + __this_cpu_write(cache->percpu->first, NULL); + __this_cpu_write(cache->percpu->count, 0); +end: + local_irq_restore(flags); +} + +static struct list_head *rds_ib_recv_cache_get(struct rds_ib_refill_cache *cache) +{ + struct list_head *head = cache->ready; + + if (head) { + if (!list_empty(head)) { + cache->ready = head->next; + list_del_init(head); + } else + cache->ready = NULL; + } + + return head; +} + +int rds_ib_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to) +{ + struct rds_ib_incoming *ibinc; + struct rds_page_frag *frag; + unsigned long to_copy; + unsigned long frag_off = 0; + int copied = 0; + int ret; + u32 len; + + ibinc = container_of(inc, struct rds_ib_incoming, ii_inc); + frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); + len = be32_to_cpu(inc->i_hdr.h_len); + + while (iov_iter_count(to) && copied < len) { + if (frag_off == RDS_FRAG_SIZE) { + frag = list_entry(frag->f_item.next, + struct rds_page_frag, f_item); + frag_off = 0; + } + to_copy = min_t(unsigned long, iov_iter_count(to), + RDS_FRAG_SIZE - frag_off); + to_copy = min_t(unsigned long, to_copy, len - copied); + + /* XXX needs + offset for multiple recvs per page */ + rds_stats_add(s_copy_to_user, to_copy); + ret = copy_page_to_iter(sg_page(&frag->f_sg), + frag->f_sg.offset + frag_off, + to_copy, + to); + if (ret != to_copy) + return -EFAULT; + + frag_off += to_copy; + copied += to_copy; + } + + return copied; +} + +/* ic starts out kzalloc()ed */ +void rds_ib_recv_init_ack(struct rds_ib_connection *ic) +{ + struct ib_send_wr *wr = &ic->i_ack_wr; + struct ib_sge *sge = &ic->i_ack_sge; + + sge->addr = ic->i_ack_dma; + sge->length = sizeof(struct rds_header); + sge->lkey = ic->i_pd->local_dma_lkey; + + wr->sg_list = sge; + wr->num_sge = 1; + wr->opcode = IB_WR_SEND; + wr->wr_id = RDS_IB_ACK_WR_ID; + wr->send_flags = IB_SEND_SIGNALED | IB_SEND_SOLICITED; +} + +/* + * You'd think that with reliable IB connections you wouldn't need to ack + * messages that have been received. The problem is that IB hardware generates + * an ack message before it has DMAed the message into memory. This creates a + * potential message loss if the HCA is disabled for any reason between when it + * sends the ack and before the message is DMAed and processed. This is only a + * potential issue if another HCA is available for fail-over. + * + * When the remote host receives our ack they'll free the sent message from + * their send queue. To decrease the latency of this we always send an ack + * immediately after we've received messages. + * + * For simplicity, we only have one ack in flight at a time. This puts + * pressure on senders to have deep enough send queues to absorb the latency of + * a single ack frame being in flight. This might not be good enough. + * + * This is implemented by have a long-lived send_wr and sge which point to a + * statically allocated ack frame. This ack wr does not fall under the ring + * accounting that the tx and rx wrs do. The QP attribute specifically makes + * room for it beyond the ring size. Send completion notices its special + * wr_id and avoids working with the ring in that case. + */ +#ifndef KERNEL_HAS_ATOMIC64 +void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required) +{ + unsigned long flags; + + spin_lock_irqsave(&ic->i_ack_lock, flags); + ic->i_ack_next = seq; + if (ack_required) + set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); + spin_unlock_irqrestore(&ic->i_ack_lock, flags); +} + +static u64 rds_ib_get_ack(struct rds_ib_connection *ic) +{ + unsigned long flags; + u64 seq; + + clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); + + spin_lock_irqsave(&ic->i_ack_lock, flags); + seq = ic->i_ack_next; + spin_unlock_irqrestore(&ic->i_ack_lock, flags); + + return seq; +} +#else +void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq, int ack_required) +{ + atomic64_set(&ic->i_ack_next, seq); + if (ack_required) { + smp_mb__before_atomic(); + set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); + } +} + +static u64 rds_ib_get_ack(struct rds_ib_connection *ic) +{ + clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); + smp_mb__after_atomic(); + + return atomic64_read(&ic->i_ack_next); +} +#endif + + +static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits) +{ + struct rds_header *hdr = ic->i_ack; + u64 seq; + int ret; + + seq = rds_ib_get_ack(ic); + + rdsdebug("send_ack: ic %p ack %llu\n", ic, (unsigned long long) seq); + rds_message_populate_header(hdr, 0, 0, 0); + hdr->h_ack = cpu_to_be64(seq); + hdr->h_credit = adv_credits; + rds_message_make_checksum(hdr); + ic->i_ack_queued = jiffies; + + ret = ib_post_send(ic->i_cm_id->qp, &ic->i_ack_wr, NULL); + if (unlikely(ret)) { + /* Failed to send. Release the WR, and + * force another ACK. + */ + clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); + set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); + + rds_ib_stats_inc(s_ib_ack_send_failure); + + rds_ib_conn_error(ic->conn, "sending ack failed\n"); + } else + rds_ib_stats_inc(s_ib_ack_sent); +} + +/* + * There are 3 ways of getting acknowledgements to the peer: + * 1. We call rds_ib_attempt_ack from the recv completion handler + * to send an ACK-only frame. + * However, there can be only one such frame in the send queue + * at any time, so we may have to postpone it. + * 2. When another (data) packet is transmitted while there's + * an ACK in the queue, we piggyback the ACK sequence number + * on the data packet. + * 3. If the ACK WR is done sending, we get called from the + * send queue completion handler, and check whether there's + * another ACK pending (postponed because the WR was on the + * queue). If so, we transmit it. + * + * We maintain 2 variables: + * - i_ack_flags, which keeps track of whether the ACK WR + * is currently in the send queue or not (IB_ACK_IN_FLIGHT) + * - i_ack_next, which is the last sequence number we received + * + * Potentially, send queue and receive queue handlers can run concurrently. + * It would be nice to not have to use a spinlock to synchronize things, + * but the one problem that rules this out is that 64bit updates are + * not atomic on all platforms. Things would be a lot simpler if + * we had atomic64 or maybe cmpxchg64 everywhere. + * + * Reconnecting complicates this picture just slightly. When we + * reconnect, we may be seeing duplicate packets. The peer + * is retransmitting them, because it hasn't seen an ACK for + * them. It is important that we ACK these. + * + * ACK mitigation adds a header flag "ACK_REQUIRED"; any packet with + * this flag set *MUST* be acknowledged immediately. + */ + +/* + * When we get here, we're called from the recv queue handler. + * Check whether we ought to transmit an ACK. + */ +void rds_ib_attempt_ack(struct rds_ib_connection *ic) +{ + unsigned int adv_credits; + + if (!test_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) + return; + + if (test_and_set_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags)) { + rds_ib_stats_inc(s_ib_ack_send_delayed); + return; + } + + /* Can we get a send credit? */ + if (!rds_ib_send_grab_credits(ic, 1, &adv_credits, 0, RDS_MAX_ADV_CREDIT)) { + rds_ib_stats_inc(s_ib_tx_throttle); + clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); + return; + } + + clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); + rds_ib_send_ack(ic, adv_credits); +} + +/* + * We get here from the send completion handler, when the + * adapter tells us the ACK frame was sent. + */ +void rds_ib_ack_send_complete(struct rds_ib_connection *ic) +{ + clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); + rds_ib_attempt_ack(ic); +} + +/* + * This is called by the regular xmit code when it wants to piggyback + * an ACK on an outgoing frame. + */ +u64 rds_ib_piggyb_ack(struct rds_ib_connection *ic) +{ + if (test_and_clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags)) + rds_ib_stats_inc(s_ib_ack_send_piggybacked); + return rds_ib_get_ack(ic); +} + +/* + * It's kind of lame that we're copying from the posted receive pages into + * long-lived bitmaps. We could have posted the bitmaps and rdma written into + * them. But receiving new congestion bitmaps should be a *rare* event, so + * hopefully we won't need to invest that complexity in making it more + * efficient. By copying we can share a simpler core with TCP which has to + * copy. + */ +static void rds_ib_cong_recv(struct rds_connection *conn, + struct rds_ib_incoming *ibinc) +{ + struct rds_cong_map *map; + unsigned int map_off; + unsigned int map_page; + struct rds_page_frag *frag; + unsigned long frag_off; + unsigned long to_copy; + unsigned long copied; + uint64_t uncongested = 0; + void *addr; + + /* catch completely corrupt packets */ + if (be32_to_cpu(ibinc->ii_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES) + return; + + map = conn->c_fcong; + map_page = 0; + map_off = 0; + + frag = list_entry(ibinc->ii_frags.next, struct rds_page_frag, f_item); + frag_off = 0; + + copied = 0; + + while (copied < RDS_CONG_MAP_BYTES) { + uint64_t *src, *dst; + unsigned int k; + + to_copy = min(RDS_FRAG_SIZE - frag_off, PAGE_SIZE - map_off); + BUG_ON(to_copy & 7); /* Must be 64bit aligned. */ + + addr = kmap_atomic(sg_page(&frag->f_sg)); + + src = addr + frag->f_sg.offset + frag_off; + dst = (void *)map->m_page_addrs[map_page] + map_off; + for (k = 0; k < to_copy; k += 8) { + /* Record ports that became uncongested, ie + * bits that changed from 0 to 1. */ + uncongested |= ~(*src) & *dst; + *dst++ = *src++; + } + kunmap_atomic(addr); + + copied += to_copy; + + map_off += to_copy; + if (map_off == PAGE_SIZE) { + map_off = 0; + map_page++; + } + + frag_off += to_copy; + if (frag_off == RDS_FRAG_SIZE) { + frag = list_entry(frag->f_item.next, + struct rds_page_frag, f_item); + frag_off = 0; + } + } + + /* the congestion map is in little endian order */ + uncongested = le64_to_cpu(uncongested); + + rds_cong_map_updated(map, uncongested); +} + +static void rds_ib_process_recv(struct rds_connection *conn, + struct rds_ib_recv_work *recv, u32 data_len, + struct rds_ib_ack_state *state) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + struct rds_ib_incoming *ibinc = ic->i_ibinc; + struct rds_header *ihdr, *hdr; + + /* XXX shut down the connection if port 0,0 are seen? */ + + rdsdebug("ic %p ibinc %p recv %p byte len %u\n", ic, ibinc, recv, + data_len); + + if (data_len < sizeof(struct rds_header)) { + rds_ib_conn_error(conn, "incoming message " + "from %pI6c didn't include a " + "header, disconnecting and " + "reconnecting\n", + &conn->c_faddr); + return; + } + data_len -= sizeof(struct rds_header); + + ihdr = &ic->i_recv_hdrs[recv - ic->i_recvs]; + + /* Validate the checksum. */ + if (!rds_message_verify_checksum(ihdr)) { + rds_ib_conn_error(conn, "incoming message " + "from %pI6c has corrupted header - " + "forcing a reconnect\n", + &conn->c_faddr); + rds_stats_inc(s_recv_drop_bad_checksum); + return; + } + + /* Process the ACK sequence which comes with every packet */ + state->ack_recv = be64_to_cpu(ihdr->h_ack); + state->ack_recv_valid = 1; + + /* Process the credits update if there was one */ + if (ihdr->h_credit) + rds_ib_send_add_credits(conn, ihdr->h_credit); + + if (ihdr->h_sport == 0 && ihdr->h_dport == 0 && data_len == 0) { + /* This is an ACK-only packet. The fact that it gets + * special treatment here is that historically, ACKs + * were rather special beasts. + */ + rds_ib_stats_inc(s_ib_ack_received); + + /* + * Usually the frags make their way on to incs and are then freed as + * the inc is freed. We don't go that route, so we have to drop the + * page ref ourselves. We can't just leave the page on the recv + * because that confuses the dma mapping of pages and each recv's use + * of a partial page. + * + * FIXME: Fold this into the code path below. + */ + rds_ib_frag_free(ic, recv->r_frag); + recv->r_frag = NULL; + return; + } + + /* + * If we don't already have an inc on the connection then this + * fragment has a header and starts a message.. copy its header + * into the inc and save the inc so we can hang upcoming fragments + * off its list. + */ + if (!ibinc) { + ibinc = recv->r_ibinc; + recv->r_ibinc = NULL; + ic->i_ibinc = ibinc; + + hdr = &ibinc->ii_inc.i_hdr; + ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] = + local_clock(); + memcpy(hdr, ihdr, sizeof(*hdr)); + ic->i_recv_data_rem = be32_to_cpu(hdr->h_len); + ibinc->ii_inc.i_rx_lat_trace[RDS_MSG_RX_START] = + local_clock(); + + rdsdebug("ic %p ibinc %p rem %u flag 0x%x\n", ic, ibinc, + ic->i_recv_data_rem, hdr->h_flags); + } else { + hdr = &ibinc->ii_inc.i_hdr; + /* We can't just use memcmp here; fragments of a + * single message may carry different ACKs */ + if (hdr->h_sequence != ihdr->h_sequence || + hdr->h_len != ihdr->h_len || + hdr->h_sport != ihdr->h_sport || + hdr->h_dport != ihdr->h_dport) { + rds_ib_conn_error(conn, + "fragment header mismatch; forcing reconnect\n"); + return; + } + } + + list_add_tail(&recv->r_frag->f_item, &ibinc->ii_frags); + recv->r_frag = NULL; + + if (ic->i_recv_data_rem > RDS_FRAG_SIZE) + ic->i_recv_data_rem -= RDS_FRAG_SIZE; + else { + ic->i_recv_data_rem = 0; + ic->i_ibinc = NULL; + + if (ibinc->ii_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) { + rds_ib_cong_recv(conn, ibinc); + } else { + rds_recv_incoming(conn, &conn->c_faddr, &conn->c_laddr, + &ibinc->ii_inc, GFP_ATOMIC); + state->ack_next = be64_to_cpu(hdr->h_sequence); + state->ack_next_valid = 1; + } + + /* Evaluate the ACK_REQUIRED flag *after* we received + * the complete frame, and after bumping the next_rx + * sequence. */ + if (hdr->h_flags & RDS_FLAG_ACK_REQUIRED) { + rds_stats_inc(s_recv_ack_required); + state->ack_required = 1; + } + + rds_inc_put(&ibinc->ii_inc); + } +} + +void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic, + struct ib_wc *wc, + struct rds_ib_ack_state *state) +{ + struct rds_connection *conn = ic->conn; + struct rds_ib_recv_work *recv; + + rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n", + (unsigned long long)wc->wr_id, wc->status, + ib_wc_status_msg(wc->status), wc->byte_len, + be32_to_cpu(wc->ex.imm_data)); + + rds_ib_stats_inc(s_ib_rx_cq_event); + recv = &ic->i_recvs[rds_ib_ring_oldest(&ic->i_recv_ring)]; + ib_dma_unmap_sg(ic->i_cm_id->device, &recv->r_frag->f_sg, 1, + DMA_FROM_DEVICE); + + /* Also process recvs in connecting state because it is possible + * to get a recv completion _before_ the rdmacm ESTABLISHED + * event is processed. + */ + if (wc->status == IB_WC_SUCCESS) { + rds_ib_process_recv(conn, recv, wc->byte_len, state); + } else { + /* We expect errors as the qp is drained during shutdown */ + if (rds_conn_up(conn) || rds_conn_connecting(conn)) + rds_ib_conn_error(conn, "recv completion on <%pI6c,%pI6c> had status %u (%s), disconnecting and reconnecting\n", + &conn->c_laddr, &conn->c_faddr, + wc->status, + ib_wc_status_msg(wc->status)); + } + + /* rds_ib_process_recv() doesn't always consume the frag, and + * we might not have called it at all if the wc didn't indicate + * success. We already unmapped the frag's pages, though, and + * the following rds_ib_ring_free() call tells the refill path + * that it will not find an allocated frag here. Make sure we + * keep that promise by freeing a frag that's still on the ring. + */ + if (recv->r_frag) { + rds_ib_frag_free(ic, recv->r_frag); + recv->r_frag = NULL; + } + rds_ib_ring_free(&ic->i_recv_ring, 1); + + /* If we ever end up with a really empty receive ring, we're + * in deep trouble, as the sender will definitely see RNR + * timeouts. */ + if (rds_ib_ring_empty(&ic->i_recv_ring)) + rds_ib_stats_inc(s_ib_rx_ring_empty); + + if (rds_ib_ring_low(&ic->i_recv_ring)) { + rds_ib_recv_refill(conn, 0, GFP_NOWAIT); + rds_ib_stats_inc(s_ib_rx_refill_from_cq); + } +} + +int rds_ib_recv_path(struct rds_conn_path *cp) +{ + struct rds_connection *conn = cp->cp_conn; + struct rds_ib_connection *ic = conn->c_transport_data; + + rdsdebug("conn %p\n", conn); + if (rds_conn_up(conn)) { + rds_ib_attempt_ack(ic); + rds_ib_recv_refill(conn, 0, GFP_KERNEL); + rds_ib_stats_inc(s_ib_rx_refill_from_thread); + } + + return 0; +} + +int rds_ib_recv_init(void) +{ + struct sysinfo si; + int ret = -ENOMEM; + + /* Default to 30% of all available RAM for recv memory */ + si_meminfo(&si); + rds_ib_sysctl_max_recv_allocation = si.totalram / 3 * PAGE_SIZE / RDS_FRAG_SIZE; + + rds_ib_incoming_slab = kmem_cache_create("rds_ib_incoming", + sizeof(struct rds_ib_incoming), + 0, SLAB_HWCACHE_ALIGN, NULL); + if (!rds_ib_incoming_slab) + goto out; + + rds_ib_frag_slab = kmem_cache_create("rds_ib_frag", + sizeof(struct rds_page_frag), + 0, SLAB_HWCACHE_ALIGN, NULL); + if (!rds_ib_frag_slab) { + kmem_cache_destroy(rds_ib_incoming_slab); + rds_ib_incoming_slab = NULL; + } else + ret = 0; +out: + return ret; +} + +void rds_ib_recv_exit(void) +{ + kmem_cache_destroy(rds_ib_incoming_slab); + kmem_cache_destroy(rds_ib_frag_slab); +} diff --git a/net/rds/ib_ring.c b/net/rds/ib_ring.c new file mode 100644 index 000000000..ff97e8eda --- /dev/null +++ b/net/rds/ib_ring.c @@ -0,0 +1,168 @@ +/* + * Copyright (c) 2006 Oracle. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> + +#include "rds.h" +#include "ib.h" + +/* + * Locking for IB rings. + * We assume that allocation is always protected by a mutex + * in the caller (this is a valid assumption for the current + * implementation). + * + * Freeing always happens in an interrupt, and hence only + * races with allocations, but not with other free()s. + * + * The interaction between allocation and freeing is that + * the alloc code has to determine the number of free entries. + * To this end, we maintain two counters; an allocation counter + * and a free counter. Both are allowed to run freely, and wrap + * around. + * The number of used entries is always (alloc_ctr - free_ctr) % NR. + * + * The current implementation makes free_ctr atomic. When the + * caller finds an allocation fails, it should set an "alloc fail" + * bit and retry the allocation. The "alloc fail" bit essentially tells + * the CQ completion handlers to wake it up after freeing some + * more entries. + */ + +/* + * This only happens on shutdown. + */ +DECLARE_WAIT_QUEUE_HEAD(rds_ib_ring_empty_wait); + +void rds_ib_ring_init(struct rds_ib_work_ring *ring, u32 nr) +{ + memset(ring, 0, sizeof(*ring)); + ring->w_nr = nr; + rdsdebug("ring %p nr %u\n", ring, ring->w_nr); +} + +static inline u32 __rds_ib_ring_used(struct rds_ib_work_ring *ring) +{ + u32 diff; + + /* This assumes that atomic_t has at least as many bits as u32 */ + diff = ring->w_alloc_ctr - (u32) atomic_read(&ring->w_free_ctr); + BUG_ON(diff > ring->w_nr); + + return diff; +} + +void rds_ib_ring_resize(struct rds_ib_work_ring *ring, u32 nr) +{ + /* We only ever get called from the connection setup code, + * prior to creating the QP. */ + BUG_ON(__rds_ib_ring_used(ring)); + ring->w_nr = nr; +} + +static int __rds_ib_ring_empty(struct rds_ib_work_ring *ring) +{ + return __rds_ib_ring_used(ring) == 0; +} + +u32 rds_ib_ring_alloc(struct rds_ib_work_ring *ring, u32 val, u32 *pos) +{ + u32 ret = 0, avail; + + avail = ring->w_nr - __rds_ib_ring_used(ring); + + rdsdebug("ring %p val %u next %u free %u\n", ring, val, + ring->w_alloc_ptr, avail); + + if (val && avail) { + ret = min(val, avail); + *pos = ring->w_alloc_ptr; + + ring->w_alloc_ptr = (ring->w_alloc_ptr + ret) % ring->w_nr; + ring->w_alloc_ctr += ret; + } + + return ret; +} + +void rds_ib_ring_free(struct rds_ib_work_ring *ring, u32 val) +{ + ring->w_free_ptr = (ring->w_free_ptr + val) % ring->w_nr; + atomic_add(val, &ring->w_free_ctr); + + if (__rds_ib_ring_empty(ring) && + waitqueue_active(&rds_ib_ring_empty_wait)) + wake_up(&rds_ib_ring_empty_wait); +} + +void rds_ib_ring_unalloc(struct rds_ib_work_ring *ring, u32 val) +{ + ring->w_alloc_ptr = (ring->w_alloc_ptr - val) % ring->w_nr; + ring->w_alloc_ctr -= val; +} + +int rds_ib_ring_empty(struct rds_ib_work_ring *ring) +{ + return __rds_ib_ring_empty(ring); +} + +int rds_ib_ring_low(struct rds_ib_work_ring *ring) +{ + return __rds_ib_ring_used(ring) <= (ring->w_nr >> 1); +} + +/* + * returns the oldest alloced ring entry. This will be the next one + * freed. This can't be called if there are none allocated. + */ +u32 rds_ib_ring_oldest(struct rds_ib_work_ring *ring) +{ + return ring->w_free_ptr; +} + +/* + * returns the number of completed work requests. + */ + +u32 rds_ib_ring_completed(struct rds_ib_work_ring *ring, u32 wr_id, u32 oldest) +{ + u32 ret; + + if (oldest <= (unsigned long long)wr_id) + ret = (unsigned long long)wr_id - oldest + 1; + else + ret = ring->w_nr - oldest + (unsigned long long)wr_id + 1; + + rdsdebug("ring %p ret %u wr_id %u oldest %u\n", ring, ret, + wr_id, oldest); + return ret; +} diff --git a/net/rds/ib_send.c b/net/rds/ib_send.c new file mode 100644 index 000000000..2dcb555e6 --- /dev/null +++ b/net/rds/ib_send.c @@ -0,0 +1,992 @@ +/* + * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/in.h> +#include <linux/device.h> +#include <linux/dmapool.h> +#include <linux/ratelimit.h> + +#include "rds_single_path.h" +#include "rds.h" +#include "ib.h" + +/* + * Convert IB-specific error message to RDS error message and call core + * completion handler. + */ +static void rds_ib_send_complete(struct rds_message *rm, + int wc_status, + void (*complete)(struct rds_message *rm, int status)) +{ + int notify_status; + + switch (wc_status) { + case IB_WC_WR_FLUSH_ERR: + return; + + case IB_WC_SUCCESS: + notify_status = RDS_RDMA_SUCCESS; + break; + + case IB_WC_REM_ACCESS_ERR: + notify_status = RDS_RDMA_REMOTE_ERROR; + break; + + default: + notify_status = RDS_RDMA_OTHER_ERROR; + break; + } + complete(rm, notify_status); +} + +static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic, + struct rm_rdma_op *op, + int wc_status) +{ + if (op->op_mapped) { + ib_dma_unmap_sg(ic->i_cm_id->device, + op->op_sg, op->op_nents, + op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE); + op->op_mapped = 0; + } + + /* If the user asked for a completion notification on this + * message, we can implement three different semantics: + * 1. Notify when we received the ACK on the RDS message + * that was queued with the RDMA. This provides reliable + * notification of RDMA status at the expense of a one-way + * packet delay. + * 2. Notify when the IB stack gives us the completion event for + * the RDMA operation. + * 3. Notify when the IB stack gives us the completion event for + * the accompanying RDS messages. + * Here, we implement approach #3. To implement approach #2, + * we would need to take an event for the rdma WR. To implement #1, + * don't call rds_rdma_send_complete at all, and fall back to the notify + * handling in the ACK processing code. + * + * Note: There's no need to explicitly sync any RDMA buffers using + * ib_dma_sync_sg_for_cpu - the completion for the RDMA + * operation itself unmapped the RDMA buffers, which takes care + * of synching. + */ + rds_ib_send_complete(container_of(op, struct rds_message, rdma), + wc_status, rds_rdma_send_complete); + + if (op->op_write) + rds_stats_add(s_send_rdma_bytes, op->op_bytes); + else + rds_stats_add(s_recv_rdma_bytes, op->op_bytes); +} + +static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic, + struct rm_atomic_op *op, + int wc_status) +{ + /* unmap atomic recvbuf */ + if (op->op_mapped) { + ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1, + DMA_FROM_DEVICE); + op->op_mapped = 0; + } + + rds_ib_send_complete(container_of(op, struct rds_message, atomic), + wc_status, rds_atomic_send_complete); + + if (op->op_type == RDS_ATOMIC_TYPE_CSWP) + rds_ib_stats_inc(s_ib_atomic_cswp); + else + rds_ib_stats_inc(s_ib_atomic_fadd); +} + +static void rds_ib_send_unmap_data(struct rds_ib_connection *ic, + struct rm_data_op *op, + int wc_status) +{ + struct rds_message *rm = container_of(op, struct rds_message, data); + + if (op->op_nents) + ib_dma_unmap_sg(ic->i_cm_id->device, + op->op_sg, op->op_nents, + DMA_TO_DEVICE); + + if (rm->rdma.op_active && rm->data.op_notify) + rds_ib_send_unmap_rdma(ic, &rm->rdma, wc_status); +} + +/* + * Unmap the resources associated with a struct send_work. + * + * Returns the rm for no good reason other than it is unobtainable + * other than by switching on wr.opcode, currently, and the caller, + * the event handler, needs it. + */ +static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic, + struct rds_ib_send_work *send, + int wc_status) +{ + struct rds_message *rm = NULL; + + /* In the error case, wc.opcode sometimes contains garbage */ + switch (send->s_wr.opcode) { + case IB_WR_SEND: + if (send->s_op) { + rm = container_of(send->s_op, struct rds_message, data); + rds_ib_send_unmap_data(ic, send->s_op, wc_status); + } + break; + case IB_WR_RDMA_WRITE: + case IB_WR_RDMA_READ: + if (send->s_op) { + rm = container_of(send->s_op, struct rds_message, rdma); + rds_ib_send_unmap_rdma(ic, send->s_op, wc_status); + } + break; + case IB_WR_ATOMIC_FETCH_AND_ADD: + case IB_WR_ATOMIC_CMP_AND_SWP: + if (send->s_op) { + rm = container_of(send->s_op, struct rds_message, atomic); + rds_ib_send_unmap_atomic(ic, send->s_op, wc_status); + } + break; + default: + printk_ratelimited(KERN_NOTICE + "RDS/IB: %s: unexpected opcode 0x%x in WR!\n", + __func__, send->s_wr.opcode); + break; + } + + send->s_wr.opcode = 0xdead; + + return rm; +} + +void rds_ib_send_init_ring(struct rds_ib_connection *ic) +{ + struct rds_ib_send_work *send; + u32 i; + + for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) { + struct ib_sge *sge; + + send->s_op = NULL; + + send->s_wr.wr_id = i; + send->s_wr.sg_list = send->s_sge; + send->s_wr.ex.imm_data = 0; + + sge = &send->s_sge[0]; + sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header)); + sge->length = sizeof(struct rds_header); + sge->lkey = ic->i_pd->local_dma_lkey; + + send->s_sge[1].lkey = ic->i_pd->local_dma_lkey; + } +} + +void rds_ib_send_clear_ring(struct rds_ib_connection *ic) +{ + struct rds_ib_send_work *send; + u32 i; + + for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) { + if (send->s_op && send->s_wr.opcode != 0xdead) + rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR); + } +} + +/* + * The only fast path caller always has a non-zero nr, so we don't + * bother testing nr before performing the atomic sub. + */ +static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr) +{ + if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) && + waitqueue_active(&rds_ib_ring_empty_wait)) + wake_up(&rds_ib_ring_empty_wait); + BUG_ON(atomic_read(&ic->i_signaled_sends) < 0); +} + +/* + * The _oldest/_free ring operations here race cleanly with the alloc/unalloc + * operations performed in the send path. As the sender allocs and potentially + * unallocs the next free entry in the ring it doesn't alter which is + * the next to be freed, which is what this is concerned with. + */ +void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc) +{ + struct rds_message *rm = NULL; + struct rds_connection *conn = ic->conn; + struct rds_ib_send_work *send; + u32 completed; + u32 oldest; + u32 i = 0; + int nr_sig = 0; + + + rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n", + (unsigned long long)wc->wr_id, wc->status, + ib_wc_status_msg(wc->status), wc->byte_len, + be32_to_cpu(wc->ex.imm_data)); + rds_ib_stats_inc(s_ib_tx_cq_event); + + if (wc->wr_id == RDS_IB_ACK_WR_ID) { + if (time_after(jiffies, ic->i_ack_queued + HZ / 2)) + rds_ib_stats_inc(s_ib_tx_stalled); + rds_ib_ack_send_complete(ic); + return; + } + + oldest = rds_ib_ring_oldest(&ic->i_send_ring); + + completed = rds_ib_ring_completed(&ic->i_send_ring, wc->wr_id, oldest); + + for (i = 0; i < completed; i++) { + send = &ic->i_sends[oldest]; + if (send->s_wr.send_flags & IB_SEND_SIGNALED) + nr_sig++; + + rm = rds_ib_send_unmap_op(ic, send, wc->status); + + if (time_after(jiffies, send->s_queued + HZ / 2)) + rds_ib_stats_inc(s_ib_tx_stalled); + + if (send->s_op) { + if (send->s_op == rm->m_final_op) { + /* If anyone waited for this message to get + * flushed out, wake them up now + */ + rds_message_unmapped(rm); + } + rds_message_put(rm); + send->s_op = NULL; + } + + oldest = (oldest + 1) % ic->i_send_ring.w_nr; + } + + rds_ib_ring_free(&ic->i_send_ring, completed); + rds_ib_sub_signaled(ic, nr_sig); + nr_sig = 0; + + if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) || + test_bit(0, &conn->c_map_queued)) + queue_delayed_work(rds_wq, &conn->c_send_w, 0); + + /* We expect errors as the qp is drained during shutdown */ + if (wc->status != IB_WC_SUCCESS && rds_conn_up(conn)) { + rds_ib_conn_error(conn, "send completion on <%pI6c,%pI6c> had status %u (%s), disconnecting and reconnecting\n", + &conn->c_laddr, &conn->c_faddr, wc->status, + ib_wc_status_msg(wc->status)); + } +} + +/* + * This is the main function for allocating credits when sending + * messages. + * + * Conceptually, we have two counters: + * - send credits: this tells us how many WRs we're allowed + * to submit without overruning the receiver's queue. For + * each SEND WR we post, we decrement this by one. + * + * - posted credits: this tells us how many WRs we recently + * posted to the receive queue. This value is transferred + * to the peer as a "credit update" in a RDS header field. + * Every time we transmit credits to the peer, we subtract + * the amount of transferred credits from this counter. + * + * It is essential that we avoid situations where both sides have + * exhausted their send credits, and are unable to send new credits + * to the peer. We achieve this by requiring that we send at least + * one credit update to the peer before exhausting our credits. + * When new credits arrive, we subtract one credit that is withheld + * until we've posted new buffers and are ready to transmit these + * credits (see rds_ib_send_add_credits below). + * + * The RDS send code is essentially single-threaded; rds_send_xmit + * sets RDS_IN_XMIT to ensure exclusive access to the send ring. + * However, the ACK sending code is independent and can race with + * message SENDs. + * + * In the send path, we need to update the counters for send credits + * and the counter of posted buffers atomically - when we use the + * last available credit, we cannot allow another thread to race us + * and grab the posted credits counter. Hence, we have to use a + * spinlock to protect the credit counter, or use atomics. + * + * Spinlocks shared between the send and the receive path are bad, + * because they create unnecessary delays. An early implementation + * using a spinlock showed a 5% degradation in throughput at some + * loads. + * + * This implementation avoids spinlocks completely, putting both + * counters into a single atomic, and updating that atomic using + * atomic_add (in the receive path, when receiving fresh credits), + * and using atomic_cmpxchg when updating the two counters. + */ +int rds_ib_send_grab_credits(struct rds_ib_connection *ic, + u32 wanted, u32 *adv_credits, int need_posted, int max_posted) +{ + unsigned int avail, posted, got = 0, advertise; + long oldval, newval; + + *adv_credits = 0; + if (!ic->i_flowctl) + return wanted; + +try_again: + advertise = 0; + oldval = newval = atomic_read(&ic->i_credits); + posted = IB_GET_POST_CREDITS(oldval); + avail = IB_GET_SEND_CREDITS(oldval); + + rdsdebug("wanted=%u credits=%u posted=%u\n", + wanted, avail, posted); + + /* The last credit must be used to send a credit update. */ + if (avail && !posted) + avail--; + + if (avail < wanted) { + struct rds_connection *conn = ic->i_cm_id->context; + + /* Oops, there aren't that many credits left! */ + set_bit(RDS_LL_SEND_FULL, &conn->c_flags); + got = avail; + } else { + /* Sometimes you get what you want, lalala. */ + got = wanted; + } + newval -= IB_SET_SEND_CREDITS(got); + + /* + * If need_posted is non-zero, then the caller wants + * the posted regardless of whether any send credits are + * available. + */ + if (posted && (got || need_posted)) { + advertise = min_t(unsigned int, posted, max_posted); + newval -= IB_SET_POST_CREDITS(advertise); + } + + /* Finally bill everything */ + if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval) + goto try_again; + + *adv_credits = advertise; + return got; +} + +void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + + if (credits == 0) + return; + + rdsdebug("credits=%u current=%u%s\n", + credits, + IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)), + test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : ""); + + atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits); + if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags)) + queue_delayed_work(rds_wq, &conn->c_send_w, 0); + + WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384); + + rds_ib_stats_inc(s_ib_rx_credit_updates); +} + +void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + + if (posted == 0) + return; + + atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits); + + /* Decide whether to send an update to the peer now. + * If we would send a credit update for every single buffer we + * post, we would end up with an ACK storm (ACK arrives, + * consumes buffer, we refill the ring, send ACK to remote + * advertising the newly posted buffer... ad inf) + * + * Performance pretty much depends on how often we send + * credit updates - too frequent updates mean lots of ACKs. + * Too infrequent updates, and the peer will run out of + * credits and has to throttle. + * For the time being, 16 seems to be a good compromise. + */ + if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16) + set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags); +} + +static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic, + struct rds_ib_send_work *send, + bool notify) +{ + /* + * We want to delay signaling completions just enough to get + * the batching benefits but not so much that we create dead time + * on the wire. + */ + if (ic->i_unsignaled_wrs-- == 0 || notify) { + ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs; + send->s_wr.send_flags |= IB_SEND_SIGNALED; + return 1; + } + return 0; +} + +/* + * This can be called multiple times for a given message. The first time + * we see a message we map its scatterlist into the IB device so that + * we can provide that mapped address to the IB scatter gather entries + * in the IB work requests. We translate the scatterlist into a series + * of work requests that fragment the message. These work requests complete + * in order so we pass ownership of the message to the completion handler + * once we send the final fragment. + * + * The RDS core uses the c_send_lock to only enter this function once + * per connection. This makes sure that the tx ring alloc/unalloc pairs + * don't get out of sync and confuse the ring. + */ +int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm, + unsigned int hdr_off, unsigned int sg, unsigned int off) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + struct ib_device *dev = ic->i_cm_id->device; + struct rds_ib_send_work *send = NULL; + struct rds_ib_send_work *first; + struct rds_ib_send_work *prev; + const struct ib_send_wr *failed_wr; + struct scatterlist *scat; + u32 pos; + u32 i; + u32 work_alloc; + u32 credit_alloc = 0; + u32 posted; + u32 adv_credits = 0; + int send_flags = 0; + int bytes_sent = 0; + int ret; + int flow_controlled = 0; + int nr_sig = 0; + + BUG_ON(off % RDS_FRAG_SIZE); + BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header)); + + /* Do not send cong updates to IB loopback */ + if (conn->c_loopback + && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) { + rds_cong_map_updated(conn->c_fcong, ~(u64) 0); + scat = &rm->data.op_sg[sg]; + ret = max_t(int, RDS_CONG_MAP_BYTES, scat->length); + return sizeof(struct rds_header) + ret; + } + + /* FIXME we may overallocate here */ + if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) + i = 1; + else + i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE); + + work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos); + if (work_alloc == 0) { + set_bit(RDS_LL_SEND_FULL, &conn->c_flags); + rds_ib_stats_inc(s_ib_tx_ring_full); + ret = -ENOMEM; + goto out; + } + + if (ic->i_flowctl) { + credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT); + adv_credits += posted; + if (credit_alloc < work_alloc) { + rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc); + work_alloc = credit_alloc; + flow_controlled = 1; + } + if (work_alloc == 0) { + set_bit(RDS_LL_SEND_FULL, &conn->c_flags); + rds_ib_stats_inc(s_ib_tx_throttle); + ret = -ENOMEM; + goto out; + } + } + + /* map the message the first time we see it */ + if (!ic->i_data_op) { + if (rm->data.op_nents) { + rm->data.op_count = ib_dma_map_sg(dev, + rm->data.op_sg, + rm->data.op_nents, + DMA_TO_DEVICE); + rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count); + if (rm->data.op_count == 0) { + rds_ib_stats_inc(s_ib_tx_sg_mapping_failure); + rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); + ret = -ENOMEM; /* XXX ? */ + goto out; + } + } else { + rm->data.op_count = 0; + } + + rds_message_addref(rm); + rm->data.op_dmasg = 0; + rm->data.op_dmaoff = 0; + ic->i_data_op = &rm->data; + + /* Finalize the header */ + if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags)) + rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED; + if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) + rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED; + + /* If it has a RDMA op, tell the peer we did it. This is + * used by the peer to release use-once RDMA MRs. */ + if (rm->rdma.op_active) { + struct rds_ext_header_rdma ext_hdr; + + ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey); + rds_message_add_extension(&rm->m_inc.i_hdr, + RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr)); + } + if (rm->m_rdma_cookie) { + rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr, + rds_rdma_cookie_key(rm->m_rdma_cookie), + rds_rdma_cookie_offset(rm->m_rdma_cookie)); + } + + /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so + * we should not do this unless we have a chance of at least + * sticking the header into the send ring. Which is why we + * should call rds_ib_ring_alloc first. */ + rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic)); + rds_message_make_checksum(&rm->m_inc.i_hdr); + + /* + * Update adv_credits since we reset the ACK_REQUIRED bit. + */ + if (ic->i_flowctl) { + rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits); + adv_credits += posted; + BUG_ON(adv_credits > 255); + } + } + + /* Sometimes you want to put a fence between an RDMA + * READ and the following SEND. + * We could either do this all the time + * or when requested by the user. Right now, we let + * the application choose. + */ + if (rm->rdma.op_active && rm->rdma.op_fence) + send_flags = IB_SEND_FENCE; + + /* Each frag gets a header. Msgs may be 0 bytes */ + send = &ic->i_sends[pos]; + first = send; + prev = NULL; + scat = &ic->i_data_op->op_sg[rm->data.op_dmasg]; + i = 0; + do { + unsigned int len = 0; + + /* Set up the header */ + send->s_wr.send_flags = send_flags; + send->s_wr.opcode = IB_WR_SEND; + send->s_wr.num_sge = 1; + send->s_wr.next = NULL; + send->s_queued = jiffies; + send->s_op = NULL; + + send->s_sge[0].addr = ic->i_send_hdrs_dma + + (pos * sizeof(struct rds_header)); + send->s_sge[0].length = sizeof(struct rds_header); + + memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header)); + + /* Set up the data, if present */ + if (i < work_alloc + && scat != &rm->data.op_sg[rm->data.op_count]) { + len = min(RDS_FRAG_SIZE, + ib_sg_dma_len(dev, scat) - rm->data.op_dmaoff); + send->s_wr.num_sge = 2; + + send->s_sge[1].addr = ib_sg_dma_address(dev, scat); + send->s_sge[1].addr += rm->data.op_dmaoff; + send->s_sge[1].length = len; + + bytes_sent += len; + rm->data.op_dmaoff += len; + if (rm->data.op_dmaoff == ib_sg_dma_len(dev, scat)) { + scat++; + rm->data.op_dmasg++; + rm->data.op_dmaoff = 0; + } + } + + rds_ib_set_wr_signal_state(ic, send, false); + + /* + * Always signal the last one if we're stopping due to flow control. + */ + if (ic->i_flowctl && flow_controlled && i == (work_alloc - 1)) { + rds_ib_set_wr_signal_state(ic, send, true); + send->s_wr.send_flags |= IB_SEND_SOLICITED; + } + + if (send->s_wr.send_flags & IB_SEND_SIGNALED) + nr_sig++; + + rdsdebug("send %p wr %p num_sge %u next %p\n", send, + &send->s_wr, send->s_wr.num_sge, send->s_wr.next); + + if (ic->i_flowctl && adv_credits) { + struct rds_header *hdr = &ic->i_send_hdrs[pos]; + + /* add credit and redo the header checksum */ + hdr->h_credit = adv_credits; + rds_message_make_checksum(hdr); + adv_credits = 0; + rds_ib_stats_inc(s_ib_tx_credit_updates); + } + + if (prev) + prev->s_wr.next = &send->s_wr; + prev = send; + + pos = (pos + 1) % ic->i_send_ring.w_nr; + send = &ic->i_sends[pos]; + i++; + + } while (i < work_alloc + && scat != &rm->data.op_sg[rm->data.op_count]); + + /* Account the RDS header in the number of bytes we sent, but just once. + * The caller has no concept of fragmentation. */ + if (hdr_off == 0) + bytes_sent += sizeof(struct rds_header); + + /* if we finished the message then send completion owns it */ + if (scat == &rm->data.op_sg[rm->data.op_count]) { + prev->s_op = ic->i_data_op; + prev->s_wr.send_flags |= IB_SEND_SOLICITED; + if (!(prev->s_wr.send_flags & IB_SEND_SIGNALED)) + nr_sig += rds_ib_set_wr_signal_state(ic, prev, true); + ic->i_data_op = NULL; + } + + /* Put back wrs & credits we didn't use */ + if (i < work_alloc) { + rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i); + work_alloc = i; + } + if (ic->i_flowctl && i < credit_alloc) + rds_ib_send_add_credits(conn, credit_alloc - i); + + if (nr_sig) + atomic_add(nr_sig, &ic->i_signaled_sends); + + /* XXX need to worry about failed_wr and partial sends. */ + failed_wr = &first->s_wr; + ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr); + rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic, + first, &first->s_wr, ret, failed_wr); + BUG_ON(failed_wr != &first->s_wr); + if (ret) { + printk(KERN_WARNING "RDS/IB: ib_post_send to %pI6c " + "returned %d\n", &conn->c_faddr, ret); + rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); + rds_ib_sub_signaled(ic, nr_sig); + if (prev->s_op) { + ic->i_data_op = prev->s_op; + prev->s_op = NULL; + } + + rds_ib_conn_error(ic->conn, "ib_post_send failed\n"); + goto out; + } + + ret = bytes_sent; +out: + BUG_ON(adv_credits); + return ret; +} + +/* + * Issue atomic operation. + * A simplified version of the rdma case, we always map 1 SG, and + * only 8 bytes, for the return value from the atomic operation. + */ +int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + struct rds_ib_send_work *send = NULL; + const struct ib_send_wr *failed_wr; + u32 pos; + u32 work_alloc; + int ret; + int nr_sig = 0; + + work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos); + if (work_alloc != 1) { + rds_ib_stats_inc(s_ib_tx_ring_full); + ret = -ENOMEM; + goto out; + } + + /* address of send request in ring */ + send = &ic->i_sends[pos]; + send->s_queued = jiffies; + + if (op->op_type == RDS_ATOMIC_TYPE_CSWP) { + send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP; + send->s_atomic_wr.compare_add = op->op_m_cswp.compare; + send->s_atomic_wr.swap = op->op_m_cswp.swap; + send->s_atomic_wr.compare_add_mask = op->op_m_cswp.compare_mask; + send->s_atomic_wr.swap_mask = op->op_m_cswp.swap_mask; + } else { /* FADD */ + send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD; + send->s_atomic_wr.compare_add = op->op_m_fadd.add; + send->s_atomic_wr.swap = 0; + send->s_atomic_wr.compare_add_mask = op->op_m_fadd.nocarry_mask; + send->s_atomic_wr.swap_mask = 0; + } + send->s_wr.send_flags = 0; + nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify); + send->s_atomic_wr.wr.num_sge = 1; + send->s_atomic_wr.wr.next = NULL; + send->s_atomic_wr.remote_addr = op->op_remote_addr; + send->s_atomic_wr.rkey = op->op_rkey; + send->s_op = op; + rds_message_addref(container_of(send->s_op, struct rds_message, atomic)); + + /* map 8 byte retval buffer to the device */ + ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE); + rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret); + if (ret != 1) { + rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); + rds_ib_stats_inc(s_ib_tx_sg_mapping_failure); + ret = -ENOMEM; /* XXX ? */ + goto out; + } + + /* Convert our struct scatterlist to struct ib_sge */ + send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg); + send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg); + send->s_sge[0].lkey = ic->i_pd->local_dma_lkey; + + rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr, + send->s_sge[0].addr, send->s_sge[0].length); + + if (nr_sig) + atomic_add(nr_sig, &ic->i_signaled_sends); + + failed_wr = &send->s_atomic_wr.wr; + ret = ib_post_send(ic->i_cm_id->qp, &send->s_atomic_wr.wr, &failed_wr); + rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic, + send, &send->s_atomic_wr, ret, failed_wr); + BUG_ON(failed_wr != &send->s_atomic_wr.wr); + if (ret) { + printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI6c " + "returned %d\n", &conn->c_faddr, ret); + rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); + rds_ib_sub_signaled(ic, nr_sig); + goto out; + } + + if (unlikely(failed_wr != &send->s_atomic_wr.wr)) { + printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret); + BUG_ON(failed_wr != &send->s_atomic_wr.wr); + } + +out: + return ret; +} + +int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op) +{ + struct rds_ib_connection *ic = conn->c_transport_data; + struct rds_ib_send_work *send = NULL; + struct rds_ib_send_work *first; + struct rds_ib_send_work *prev; + const struct ib_send_wr *failed_wr; + struct scatterlist *scat; + unsigned long len; + u64 remote_addr = op->op_remote_addr; + u32 max_sge = ic->rds_ibdev->max_sge; + u32 pos; + u32 work_alloc; + u32 i; + u32 j; + int sent; + int ret; + int num_sge; + int nr_sig = 0; + + /* map the op the first time we see it */ + if (!op->op_mapped) { + op->op_count = ib_dma_map_sg(ic->i_cm_id->device, + op->op_sg, op->op_nents, (op->op_write) ? + DMA_TO_DEVICE : DMA_FROM_DEVICE); + rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count); + if (op->op_count == 0) { + rds_ib_stats_inc(s_ib_tx_sg_mapping_failure); + ret = -ENOMEM; /* XXX ? */ + goto out; + } + + op->op_mapped = 1; + } + + /* + * Instead of knowing how to return a partial rdma read/write we insist that there + * be enough work requests to send the entire message. + */ + i = ceil(op->op_count, max_sge); + + work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos); + if (work_alloc != i) { + rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); + rds_ib_stats_inc(s_ib_tx_ring_full); + ret = -ENOMEM; + goto out; + } + + send = &ic->i_sends[pos]; + first = send; + prev = NULL; + scat = &op->op_sg[0]; + sent = 0; + num_sge = op->op_count; + + for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) { + send->s_wr.send_flags = 0; + send->s_queued = jiffies; + send->s_op = NULL; + + nr_sig += rds_ib_set_wr_signal_state(ic, send, op->op_notify); + + send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ; + send->s_rdma_wr.remote_addr = remote_addr; + send->s_rdma_wr.rkey = op->op_rkey; + + if (num_sge > max_sge) { + send->s_rdma_wr.wr.num_sge = max_sge; + num_sge -= max_sge; + } else { + send->s_rdma_wr.wr.num_sge = num_sge; + } + + send->s_rdma_wr.wr.next = NULL; + + if (prev) + prev->s_rdma_wr.wr.next = &send->s_rdma_wr.wr; + + for (j = 0; j < send->s_rdma_wr.wr.num_sge && + scat != &op->op_sg[op->op_count]; j++) { + len = ib_sg_dma_len(ic->i_cm_id->device, scat); + send->s_sge[j].addr = + ib_sg_dma_address(ic->i_cm_id->device, scat); + send->s_sge[j].length = len; + send->s_sge[j].lkey = ic->i_pd->local_dma_lkey; + + sent += len; + rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr); + + remote_addr += len; + scat++; + } + + rdsdebug("send %p wr %p num_sge %u next %p\n", send, + &send->s_rdma_wr.wr, + send->s_rdma_wr.wr.num_sge, + send->s_rdma_wr.wr.next); + + prev = send; + if (++send == &ic->i_sends[ic->i_send_ring.w_nr]) + send = ic->i_sends; + } + + /* give a reference to the last op */ + if (scat == &op->op_sg[op->op_count]) { + prev->s_op = op; + rds_message_addref(container_of(op, struct rds_message, rdma)); + } + + if (i < work_alloc) { + rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i); + work_alloc = i; + } + + if (nr_sig) + atomic_add(nr_sig, &ic->i_signaled_sends); + + failed_wr = &first->s_rdma_wr.wr; + ret = ib_post_send(ic->i_cm_id->qp, &first->s_rdma_wr.wr, &failed_wr); + rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic, + first, &first->s_rdma_wr.wr, ret, failed_wr); + BUG_ON(failed_wr != &first->s_rdma_wr.wr); + if (ret) { + printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI6c " + "returned %d\n", &conn->c_faddr, ret); + rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc); + rds_ib_sub_signaled(ic, nr_sig); + goto out; + } + + if (unlikely(failed_wr != &first->s_rdma_wr.wr)) { + printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret); + BUG_ON(failed_wr != &first->s_rdma_wr.wr); + } + + +out: + return ret; +} + +void rds_ib_xmit_path_complete(struct rds_conn_path *cp) +{ + struct rds_connection *conn = cp->cp_conn; + struct rds_ib_connection *ic = conn->c_transport_data; + + /* We may have a pending ACK or window update we were unable + * to send previously (due to flow control). Try again. */ + rds_ib_attempt_ack(ic); +} diff --git a/net/rds/ib_stats.c b/net/rds/ib_stats.c new file mode 100644 index 000000000..ac46d8961 --- /dev/null +++ b/net/rds/ib_stats.c @@ -0,0 +1,107 @@ +/* + * Copyright (c) 2006 Oracle. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/percpu.h> +#include <linux/seq_file.h> +#include <linux/proc_fs.h> + +#include "rds.h" +#include "ib.h" + +DEFINE_PER_CPU_SHARED_ALIGNED(struct rds_ib_statistics, rds_ib_stats); + +static const char *const rds_ib_stat_names[] = { + "ib_connect_raced", + "ib_listen_closed_stale", + "ib_evt_handler_call", + "ib_tasklet_call", + "ib_tx_cq_event", + "ib_tx_ring_full", + "ib_tx_throttle", + "ib_tx_sg_mapping_failure", + "ib_tx_stalled", + "ib_tx_credit_updates", + "ib_rx_cq_event", + "ib_rx_ring_empty", + "ib_rx_refill_from_cq", + "ib_rx_refill_from_thread", + "ib_rx_alloc_limit", + "ib_rx_total_frags", + "ib_rx_total_incs", + "ib_rx_credit_updates", + "ib_ack_sent", + "ib_ack_send_failure", + "ib_ack_send_delayed", + "ib_ack_send_piggybacked", + "ib_ack_received", + "ib_rdma_mr_8k_alloc", + "ib_rdma_mr_8k_free", + "ib_rdma_mr_8k_used", + "ib_rdma_mr_8k_pool_flush", + "ib_rdma_mr_8k_pool_wait", + "ib_rdma_mr_8k_pool_depleted", + "ib_rdma_mr_1m_alloc", + "ib_rdma_mr_1m_free", + "ib_rdma_mr_1m_used", + "ib_rdma_mr_1m_pool_flush", + "ib_rdma_mr_1m_pool_wait", + "ib_rdma_mr_1m_pool_depleted", + "ib_rdma_mr_8k_reused", + "ib_rdma_mr_1m_reused", + "ib_atomic_cswp", + "ib_atomic_fadd", +}; + +unsigned int rds_ib_stats_info_copy(struct rds_info_iterator *iter, + unsigned int avail) +{ + struct rds_ib_statistics stats = {0, }; + uint64_t *src; + uint64_t *sum; + size_t i; + int cpu; + + if (avail < ARRAY_SIZE(rds_ib_stat_names)) + goto out; + + for_each_online_cpu(cpu) { + src = (uint64_t *)&(per_cpu(rds_ib_stats, cpu)); + sum = (uint64_t *)&stats; + for (i = 0; i < sizeof(stats) / sizeof(uint64_t); i++) + *(sum++) += *(src++); + } + + rds_stats_info_copy(iter, (uint64_t *)&stats, rds_ib_stat_names, + ARRAY_SIZE(rds_ib_stat_names)); +out: + return ARRAY_SIZE(rds_ib_stat_names); +} diff --git a/net/rds/ib_sysctl.c b/net/rds/ib_sysctl.c new file mode 100644 index 000000000..e4e41b3af --- /dev/null +++ b/net/rds/ib_sysctl.c @@ -0,0 +1,121 @@ +/* + * Copyright (c) 2006 Oracle. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/sysctl.h> +#include <linux/proc_fs.h> + +#include "ib.h" + +static struct ctl_table_header *rds_ib_sysctl_hdr; + +unsigned long rds_ib_sysctl_max_send_wr = RDS_IB_DEFAULT_SEND_WR; +unsigned long rds_ib_sysctl_max_recv_wr = RDS_IB_DEFAULT_RECV_WR; +unsigned long rds_ib_sysctl_max_recv_allocation = (128 * 1024 * 1024) / RDS_FRAG_SIZE; +static unsigned long rds_ib_sysctl_max_wr_min = 1; +/* hardware will fail CQ creation long before this */ +static unsigned long rds_ib_sysctl_max_wr_max = (u32)~0; + +unsigned long rds_ib_sysctl_max_unsig_wrs = 16; +static unsigned long rds_ib_sysctl_max_unsig_wr_min = 1; +static unsigned long rds_ib_sysctl_max_unsig_wr_max = 64; + +/* + * This sysctl does nothing. + * + * Backwards compatibility with RDS 3.0 wire protocol + * disables initial FC credit exchange. + * If it's ever possible to drop 3.0 support, + * setting this to 1 and moving init/refill of send/recv + * rings from ib_cm_connect_complete() back into ib_setup_qp() + * will cause credits to be added before protocol negotiation. + */ +unsigned int rds_ib_sysctl_flow_control = 0; + +static struct ctl_table rds_ib_sysctl_table[] = { + { + .procname = "max_send_wr", + .data = &rds_ib_sysctl_max_send_wr, + .maxlen = sizeof(unsigned long), + .mode = 0644, + .proc_handler = proc_doulongvec_minmax, + .extra1 = &rds_ib_sysctl_max_wr_min, + .extra2 = &rds_ib_sysctl_max_wr_max, + }, + { + .procname = "max_recv_wr", + .data = &rds_ib_sysctl_max_recv_wr, + .maxlen = sizeof(unsigned long), + .mode = 0644, + .proc_handler = proc_doulongvec_minmax, + .extra1 = &rds_ib_sysctl_max_wr_min, + .extra2 = &rds_ib_sysctl_max_wr_max, + }, + { + .procname = "max_unsignaled_wr", + .data = &rds_ib_sysctl_max_unsig_wrs, + .maxlen = sizeof(unsigned long), + .mode = 0644, + .proc_handler = proc_doulongvec_minmax, + .extra1 = &rds_ib_sysctl_max_unsig_wr_min, + .extra2 = &rds_ib_sysctl_max_unsig_wr_max, + }, + { + .procname = "max_recv_allocation", + .data = &rds_ib_sysctl_max_recv_allocation, + .maxlen = sizeof(unsigned long), + .mode = 0644, + .proc_handler = proc_doulongvec_minmax, + }, + { + .procname = "flow_control", + .data = &rds_ib_sysctl_flow_control, + .maxlen = sizeof(rds_ib_sysctl_flow_control), + .mode = 0644, + .proc_handler = proc_dointvec, + }, + { } +}; + +void rds_ib_sysctl_exit(void) +{ + if (rds_ib_sysctl_hdr) + unregister_net_sysctl_table(rds_ib_sysctl_hdr); +} + +int rds_ib_sysctl_init(void) +{ + rds_ib_sysctl_hdr = register_net_sysctl(&init_net, "net/rds/ib", rds_ib_sysctl_table); + if (!rds_ib_sysctl_hdr) + return -ENOMEM; + return 0; +} diff --git a/net/rds/info.c b/net/rds/info.c new file mode 100644 index 000000000..e367a97a1 --- /dev/null +++ b/net/rds/info.c @@ -0,0 +1,243 @@ +/* + * Copyright (c) 2006 Oracle. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/percpu.h> +#include <linux/seq_file.h> +#include <linux/slab.h> +#include <linux/proc_fs.h> +#include <linux/export.h> + +#include "rds.h" + +/* + * This file implements a getsockopt() call which copies a set of fixed + * sized structs into a user-specified buffer as a means of providing + * read-only information about RDS. + * + * For a given information source there are a given number of fixed sized + * structs at a given time. The structs are only copied if the user-specified + * buffer is big enough. The destination pages that make up the buffer + * are pinned for the duration of the copy. + * + * This gives us the following benefits: + * + * - simple implementation, no copy "position" across multiple calls + * - consistent snapshot of an info source + * - atomic copy works well with whatever locking info source has + * - one portable tool to get rds info across implementations + * - long-lived tool can get info without allocating + * + * at the following costs: + * + * - info source copy must be pinned, may be "large" + */ + +struct rds_info_iterator { + struct page **pages; + void *addr; + unsigned long offset; +}; + +static DEFINE_SPINLOCK(rds_info_lock); +static rds_info_func rds_info_funcs[RDS_INFO_LAST - RDS_INFO_FIRST + 1]; + +void rds_info_register_func(int optname, rds_info_func func) +{ + int offset = optname - RDS_INFO_FIRST; + + BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST); + + spin_lock(&rds_info_lock); + BUG_ON(rds_info_funcs[offset]); + rds_info_funcs[offset] = func; + spin_unlock(&rds_info_lock); +} +EXPORT_SYMBOL_GPL(rds_info_register_func); + +void rds_info_deregister_func(int optname, rds_info_func func) +{ + int offset = optname - RDS_INFO_FIRST; + + BUG_ON(optname < RDS_INFO_FIRST || optname > RDS_INFO_LAST); + + spin_lock(&rds_info_lock); + BUG_ON(rds_info_funcs[offset] != func); + rds_info_funcs[offset] = NULL; + spin_unlock(&rds_info_lock); +} +EXPORT_SYMBOL_GPL(rds_info_deregister_func); + +/* + * Typically we hold an atomic kmap across multiple rds_info_copy() calls + * because the kmap is so expensive. This must be called before using blocking + * operations while holding the mapping and as the iterator is torn down. + */ +void rds_info_iter_unmap(struct rds_info_iterator *iter) +{ + if (iter->addr) { + kunmap_atomic(iter->addr); + iter->addr = NULL; + } +} + +/* + * get_user_pages() called flush_dcache_page() on the pages for us. + */ +void rds_info_copy(struct rds_info_iterator *iter, void *data, + unsigned long bytes) +{ + unsigned long this; + + while (bytes) { + if (!iter->addr) + iter->addr = kmap_atomic(*iter->pages); + + this = min(bytes, PAGE_SIZE - iter->offset); + + rdsdebug("page %p addr %p offset %lu this %lu data %p " + "bytes %lu\n", *iter->pages, iter->addr, + iter->offset, this, data, bytes); + + memcpy(iter->addr + iter->offset, data, this); + + data += this; + bytes -= this; + iter->offset += this; + + if (iter->offset == PAGE_SIZE) { + kunmap_atomic(iter->addr); + iter->addr = NULL; + iter->offset = 0; + iter->pages++; + } + } +} +EXPORT_SYMBOL_GPL(rds_info_copy); + +/* + * @optval points to the userspace buffer that the information snapshot + * will be copied into. + * + * @optlen on input is the size of the buffer in userspace. @optlen + * on output is the size of the requested snapshot in bytes. + * + * This function returns -errno if there is a failure, particularly -ENOSPC + * if the given userspace buffer was not large enough to fit the snapshot. + * On success it returns the positive number of bytes of each array element + * in the snapshot. + */ +int rds_info_getsockopt(struct socket *sock, int optname, char __user *optval, + int __user *optlen) +{ + struct rds_info_iterator iter; + struct rds_info_lengths lens; + unsigned long nr_pages = 0; + unsigned long start; + unsigned long i; + rds_info_func func; + struct page **pages = NULL; + int ret; + int len; + int total; + + if (get_user(len, optlen)) { + ret = -EFAULT; + goto out; + } + + /* check for all kinds of wrapping and the like */ + start = (unsigned long)optval; + if (len < 0 || len > INT_MAX - PAGE_SIZE + 1 || start + len < start) { + ret = -EINVAL; + goto out; + } + + /* a 0 len call is just trying to probe its length */ + if (len == 0) + goto call_func; + + nr_pages = (PAGE_ALIGN(start + len) - (start & PAGE_MASK)) + >> PAGE_SHIFT; + + pages = kmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); + if (!pages) { + ret = -ENOMEM; + goto out; + } + ret = get_user_pages_fast(start, nr_pages, 1, pages); + if (ret != nr_pages) { + if (ret > 0) + nr_pages = ret; + else + nr_pages = 0; + ret = -EAGAIN; /* XXX ? */ + goto out; + } + + rdsdebug("len %d nr_pages %lu\n", len, nr_pages); + +call_func: + func = rds_info_funcs[optname - RDS_INFO_FIRST]; + if (!func) { + ret = -ENOPROTOOPT; + goto out; + } + + iter.pages = pages; + iter.addr = NULL; + iter.offset = start & (PAGE_SIZE - 1); + + func(sock, len, &iter, &lens); + BUG_ON(lens.each == 0); + + total = lens.nr * lens.each; + + rds_info_iter_unmap(&iter); + + if (total > len) { + len = total; + ret = -ENOSPC; + } else { + len = total; + ret = lens.each; + } + + if (put_user(len, optlen)) + ret = -EFAULT; + +out: + for (i = 0; pages && i < nr_pages; i++) + put_page(pages[i]); + kfree(pages); + + return ret; +} diff --git a/net/rds/info.h b/net/rds/info.h new file mode 100644 index 000000000..a069b51c4 --- /dev/null +++ b/net/rds/info.h @@ -0,0 +1,31 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _RDS_INFO_H +#define _RDS_INFO_H + +struct rds_info_lengths { + unsigned int nr; + unsigned int each; +}; + +struct rds_info_iterator; + +/* + * These functions must fill in the fields of @lens to reflect the size + * of the available info source. If the snapshot fits in @len then it + * should be copied using @iter. The caller will deduce if it was copied + * or not by comparing the lengths. + */ +typedef void (*rds_info_func)(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens); + +void rds_info_register_func(int optname, rds_info_func func); +void rds_info_deregister_func(int optname, rds_info_func func); +int rds_info_getsockopt(struct socket *sock, int optname, char __user *optval, + int __user *optlen); +void rds_info_copy(struct rds_info_iterator *iter, void *data, + unsigned long bytes); +void rds_info_iter_unmap(struct rds_info_iterator *iter); + + +#endif diff --git a/net/rds/loop.c b/net/rds/loop.c new file mode 100644 index 000000000..1d73ad79c --- /dev/null +++ b/net/rds/loop.c @@ -0,0 +1,254 @@ +/* + * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/slab.h> +#include <linux/in.h> +#include <net/net_namespace.h> +#include <net/netns/generic.h> +#include <linux/ipv6.h> + +#include "rds_single_path.h" +#include "rds.h" +#include "loop.h" + +static DEFINE_SPINLOCK(loop_conns_lock); +static LIST_HEAD(loop_conns); +static atomic_t rds_loop_unloading = ATOMIC_INIT(0); + +static void rds_loop_set_unloading(void) +{ + atomic_set(&rds_loop_unloading, 1); +} + +static bool rds_loop_is_unloading(struct rds_connection *conn) +{ + return atomic_read(&rds_loop_unloading) != 0; +} + +/* + * This 'loopback' transport is a special case for flows that originate + * and terminate on the same machine. + * + * Connection build-up notices if the destination address is thought of + * as a local address by a transport. At that time it decides to use the + * loopback transport instead of the bound transport of the sending socket. + * + * The loopback transport's sending path just hands the sent rds_message + * straight to the receiving path via an embedded rds_incoming. + */ + +/* + * Usually a message transits both the sender and receiver's conns as it + * flows to the receiver. In the loopback case, though, the receive path + * is handed the sending conn so the sense of the addresses is reversed. + */ +static int rds_loop_xmit(struct rds_connection *conn, struct rds_message *rm, + unsigned int hdr_off, unsigned int sg, + unsigned int off) +{ + struct scatterlist *sgp = &rm->data.op_sg[sg]; + int ret = sizeof(struct rds_header) + + be32_to_cpu(rm->m_inc.i_hdr.h_len); + + /* Do not send cong updates to loopback */ + if (rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) { + rds_cong_map_updated(conn->c_fcong, ~(u64) 0); + ret = min_t(int, ret, sgp->length - conn->c_xmit_data_off); + goto out; + } + + BUG_ON(hdr_off || sg || off); + + rds_inc_init(&rm->m_inc, conn, &conn->c_laddr); + /* For the embedded inc. Matching put is in loop_inc_free() */ + rds_message_addref(rm); + + rds_recv_incoming(conn, &conn->c_laddr, &conn->c_faddr, &rm->m_inc, + GFP_KERNEL); + + rds_send_drop_acked(conn, be64_to_cpu(rm->m_inc.i_hdr.h_sequence), + NULL); + + rds_inc_put(&rm->m_inc); +out: + return ret; +} + +/* + * See rds_loop_xmit(). Since our inc is embedded in the rm, we + * make sure the rm lives at least until the inc is done. + */ +static void rds_loop_inc_free(struct rds_incoming *inc) +{ + struct rds_message *rm = container_of(inc, struct rds_message, m_inc); + + rds_message_put(rm); +} + +/* we need to at least give the thread something to succeed */ +static int rds_loop_recv_path(struct rds_conn_path *cp) +{ + return 0; +} + +struct rds_loop_connection { + struct list_head loop_node; + struct rds_connection *conn; +}; + +/* + * Even the loopback transport needs to keep track of its connections, + * so it can call rds_conn_destroy() on them on exit. N.B. there are + * 1+ loopback addresses (127.*.*.*) so it's not a bug to have + * multiple loopback conns allocated, although rather useless. + */ +static int rds_loop_conn_alloc(struct rds_connection *conn, gfp_t gfp) +{ + struct rds_loop_connection *lc; + unsigned long flags; + + lc = kzalloc(sizeof(struct rds_loop_connection), gfp); + if (!lc) + return -ENOMEM; + + INIT_LIST_HEAD(&lc->loop_node); + lc->conn = conn; + conn->c_transport_data = lc; + + spin_lock_irqsave(&loop_conns_lock, flags); + list_add_tail(&lc->loop_node, &loop_conns); + spin_unlock_irqrestore(&loop_conns_lock, flags); + + return 0; +} + +static void rds_loop_conn_free(void *arg) +{ + struct rds_loop_connection *lc = arg; + unsigned long flags; + + rdsdebug("lc %p\n", lc); + spin_lock_irqsave(&loop_conns_lock, flags); + list_del(&lc->loop_node); + spin_unlock_irqrestore(&loop_conns_lock, flags); + kfree(lc); +} + +static int rds_loop_conn_path_connect(struct rds_conn_path *cp) +{ + rds_connect_complete(cp->cp_conn); + return 0; +} + +static void rds_loop_conn_path_shutdown(struct rds_conn_path *cp) +{ +} + +void rds_loop_exit(void) +{ + struct rds_loop_connection *lc, *_lc; + LIST_HEAD(tmp_list); + + rds_loop_set_unloading(); + synchronize_rcu(); + /* avoid calling conn_destroy with irqs off */ + spin_lock_irq(&loop_conns_lock); + list_splice(&loop_conns, &tmp_list); + INIT_LIST_HEAD(&loop_conns); + spin_unlock_irq(&loop_conns_lock); + + list_for_each_entry_safe(lc, _lc, &tmp_list, loop_node) { + WARN_ON(lc->conn->c_passive); + rds_conn_destroy(lc->conn); + } +} + +static void rds_loop_kill_conns(struct net *net) +{ + struct rds_loop_connection *lc, *_lc; + LIST_HEAD(tmp_list); + + spin_lock_irq(&loop_conns_lock); + list_for_each_entry_safe(lc, _lc, &loop_conns, loop_node) { + struct net *c_net = read_pnet(&lc->conn->c_net); + + if (net != c_net) + continue; + list_move_tail(&lc->loop_node, &tmp_list); + } + spin_unlock_irq(&loop_conns_lock); + + list_for_each_entry_safe(lc, _lc, &tmp_list, loop_node) { + WARN_ON(lc->conn->c_passive); + rds_conn_destroy(lc->conn); + } +} + +static void __net_exit rds_loop_exit_net(struct net *net) +{ + rds_loop_kill_conns(net); +} + +static struct pernet_operations rds_loop_net_ops = { + .exit = rds_loop_exit_net, +}; + +int rds_loop_net_init(void) +{ + return register_pernet_device(&rds_loop_net_ops); +} + +void rds_loop_net_exit(void) +{ + unregister_pernet_device(&rds_loop_net_ops); +} + +/* + * This is missing .xmit_* because loop doesn't go through generic + * rds_send_xmit() and doesn't call rds_recv_incoming(). .listen_stop and + * .laddr_check are missing because transport.c doesn't iterate over + * rds_loop_transport. + */ +struct rds_transport rds_loop_transport = { + .xmit = rds_loop_xmit, + .recv_path = rds_loop_recv_path, + .conn_alloc = rds_loop_conn_alloc, + .conn_free = rds_loop_conn_free, + .conn_path_connect = rds_loop_conn_path_connect, + .conn_path_shutdown = rds_loop_conn_path_shutdown, + .inc_copy_to_user = rds_message_inc_copy_to_user, + .inc_free = rds_loop_inc_free, + .t_name = "loopback", + .t_type = RDS_TRANS_LOOP, + .t_unloading = rds_loop_is_unloading, +}; diff --git a/net/rds/loop.h b/net/rds/loop.h new file mode 100644 index 000000000..bbc8cdd03 --- /dev/null +++ b/net/rds/loop.h @@ -0,0 +1,12 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _RDS_LOOP_H +#define _RDS_LOOP_H + +/* loop.c */ +extern struct rds_transport rds_loop_transport; + +int rds_loop_net_init(void); +void rds_loop_net_exit(void); +void rds_loop_exit(void); + +#endif diff --git a/net/rds/message.c b/net/rds/message.c new file mode 100644 index 000000000..4b00b1152 --- /dev/null +++ b/net/rds/message.c @@ -0,0 +1,516 @@ +/* + * Copyright (c) 2006 Oracle. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/slab.h> +#include <linux/export.h> +#include <linux/skbuff.h> +#include <linux/list.h> +#include <linux/errqueue.h> + +#include "rds.h" + +static unsigned int rds_exthdr_size[__RDS_EXTHDR_MAX] = { +[RDS_EXTHDR_NONE] = 0, +[RDS_EXTHDR_VERSION] = sizeof(struct rds_ext_header_version), +[RDS_EXTHDR_RDMA] = sizeof(struct rds_ext_header_rdma), +[RDS_EXTHDR_RDMA_DEST] = sizeof(struct rds_ext_header_rdma_dest), +[RDS_EXTHDR_NPATHS] = sizeof(u16), +[RDS_EXTHDR_GEN_NUM] = sizeof(u32), +}; + +void rds_message_addref(struct rds_message *rm) +{ + rdsdebug("addref rm %p ref %d\n", rm, refcount_read(&rm->m_refcount)); + refcount_inc(&rm->m_refcount); +} +EXPORT_SYMBOL_GPL(rds_message_addref); + +static inline bool rds_zcookie_add(struct rds_msg_zcopy_info *info, u32 cookie) +{ + struct rds_zcopy_cookies *ck = &info->zcookies; + int ncookies = ck->num; + + if (ncookies == RDS_MAX_ZCOOKIES) + return false; + ck->cookies[ncookies] = cookie; + ck->num = ++ncookies; + return true; +} + +static struct rds_msg_zcopy_info *rds_info_from_znotifier(struct rds_znotifier *znotif) +{ + return container_of(znotif, struct rds_msg_zcopy_info, znotif); +} + +void rds_notify_msg_zcopy_purge(struct rds_msg_zcopy_queue *q) +{ + unsigned long flags; + LIST_HEAD(copy); + struct rds_msg_zcopy_info *info, *tmp; + + spin_lock_irqsave(&q->lock, flags); + list_splice(&q->zcookie_head, ©); + INIT_LIST_HEAD(&q->zcookie_head); + spin_unlock_irqrestore(&q->lock, flags); + + list_for_each_entry_safe(info, tmp, ©, rs_zcookie_next) { + list_del(&info->rs_zcookie_next); + kfree(info); + } +} + +static void rds_rm_zerocopy_callback(struct rds_sock *rs, + struct rds_znotifier *znotif) +{ + struct rds_msg_zcopy_info *info; + struct rds_msg_zcopy_queue *q; + u32 cookie = znotif->z_cookie; + struct rds_zcopy_cookies *ck; + struct list_head *head; + unsigned long flags; + + mm_unaccount_pinned_pages(&znotif->z_mmp); + q = &rs->rs_zcookie_queue; + spin_lock_irqsave(&q->lock, flags); + head = &q->zcookie_head; + if (!list_empty(head)) { + info = list_entry(head, struct rds_msg_zcopy_info, + rs_zcookie_next); + if (info && rds_zcookie_add(info, cookie)) { + spin_unlock_irqrestore(&q->lock, flags); + kfree(rds_info_from_znotifier(znotif)); + /* caller invokes rds_wake_sk_sleep() */ + return; + } + } + + info = rds_info_from_znotifier(znotif); + ck = &info->zcookies; + memset(ck, 0, sizeof(*ck)); + WARN_ON(!rds_zcookie_add(info, cookie)); + list_add_tail(&q->zcookie_head, &info->rs_zcookie_next); + + spin_unlock_irqrestore(&q->lock, flags); + /* caller invokes rds_wake_sk_sleep() */ +} + +/* + * This relies on dma_map_sg() not touching sg[].page during merging. + */ +static void rds_message_purge(struct rds_message *rm) +{ + unsigned long i, flags; + bool zcopy = false; + + if (unlikely(test_bit(RDS_MSG_PAGEVEC, &rm->m_flags))) + return; + + spin_lock_irqsave(&rm->m_rs_lock, flags); + if (rm->m_rs) { + struct rds_sock *rs = rm->m_rs; + + if (rm->data.op_mmp_znotifier) { + zcopy = true; + rds_rm_zerocopy_callback(rs, rm->data.op_mmp_znotifier); + rds_wake_sk_sleep(rs); + rm->data.op_mmp_znotifier = NULL; + } + sock_put(rds_rs_to_sk(rs)); + rm->m_rs = NULL; + } + spin_unlock_irqrestore(&rm->m_rs_lock, flags); + + for (i = 0; i < rm->data.op_nents; i++) { + /* XXX will have to put_page for page refs */ + if (!zcopy) + __free_page(sg_page(&rm->data.op_sg[i])); + else + put_page(sg_page(&rm->data.op_sg[i])); + } + rm->data.op_nents = 0; + + if (rm->rdma.op_active) + rds_rdma_free_op(&rm->rdma); + if (rm->rdma.op_rdma_mr) + rds_mr_put(rm->rdma.op_rdma_mr); + + if (rm->atomic.op_active) + rds_atomic_free_op(&rm->atomic); + if (rm->atomic.op_rdma_mr) + rds_mr_put(rm->atomic.op_rdma_mr); +} + +void rds_message_put(struct rds_message *rm) +{ + rdsdebug("put rm %p ref %d\n", rm, refcount_read(&rm->m_refcount)); + WARN(!refcount_read(&rm->m_refcount), "danger refcount zero on %p\n", rm); + if (refcount_dec_and_test(&rm->m_refcount)) { + BUG_ON(!list_empty(&rm->m_sock_item)); + BUG_ON(!list_empty(&rm->m_conn_item)); + rds_message_purge(rm); + + kfree(rm); + } +} +EXPORT_SYMBOL_GPL(rds_message_put); + +void rds_message_populate_header(struct rds_header *hdr, __be16 sport, + __be16 dport, u64 seq) +{ + hdr->h_flags = 0; + hdr->h_sport = sport; + hdr->h_dport = dport; + hdr->h_sequence = cpu_to_be64(seq); + hdr->h_exthdr[0] = RDS_EXTHDR_NONE; +} +EXPORT_SYMBOL_GPL(rds_message_populate_header); + +int rds_message_add_extension(struct rds_header *hdr, unsigned int type, + const void *data, unsigned int len) +{ + unsigned int ext_len = sizeof(u8) + len; + unsigned char *dst; + + /* For now, refuse to add more than one extension header */ + if (hdr->h_exthdr[0] != RDS_EXTHDR_NONE) + return 0; + + if (type >= __RDS_EXTHDR_MAX || len != rds_exthdr_size[type]) + return 0; + + if (ext_len >= RDS_HEADER_EXT_SPACE) + return 0; + dst = hdr->h_exthdr; + + *dst++ = type; + memcpy(dst, data, len); + + dst[len] = RDS_EXTHDR_NONE; + return 1; +} +EXPORT_SYMBOL_GPL(rds_message_add_extension); + +/* + * If a message has extension headers, retrieve them here. + * Call like this: + * + * unsigned int pos = 0; + * + * while (1) { + * buflen = sizeof(buffer); + * type = rds_message_next_extension(hdr, &pos, buffer, &buflen); + * if (type == RDS_EXTHDR_NONE) + * break; + * ... + * } + */ +int rds_message_next_extension(struct rds_header *hdr, + unsigned int *pos, void *buf, unsigned int *buflen) +{ + unsigned int offset, ext_type, ext_len; + u8 *src = hdr->h_exthdr; + + offset = *pos; + if (offset >= RDS_HEADER_EXT_SPACE) + goto none; + + /* Get the extension type and length. For now, the + * length is implied by the extension type. */ + ext_type = src[offset++]; + + if (ext_type == RDS_EXTHDR_NONE || ext_type >= __RDS_EXTHDR_MAX) + goto none; + ext_len = rds_exthdr_size[ext_type]; + if (offset + ext_len > RDS_HEADER_EXT_SPACE) + goto none; + + *pos = offset + ext_len; + if (ext_len < *buflen) + *buflen = ext_len; + memcpy(buf, src + offset, *buflen); + return ext_type; + +none: + *pos = RDS_HEADER_EXT_SPACE; + *buflen = 0; + return RDS_EXTHDR_NONE; +} + +int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset) +{ + struct rds_ext_header_rdma_dest ext_hdr; + + ext_hdr.h_rdma_rkey = cpu_to_be32(r_key); + ext_hdr.h_rdma_offset = cpu_to_be32(offset); + return rds_message_add_extension(hdr, RDS_EXTHDR_RDMA_DEST, &ext_hdr, sizeof(ext_hdr)); +} +EXPORT_SYMBOL_GPL(rds_message_add_rdma_dest_extension); + +/* + * Each rds_message is allocated with extra space for the scatterlist entries + * rds ops will need. This is to minimize memory allocation count. Then, each rds op + * can grab SGs when initializing its part of the rds_message. + */ +struct rds_message *rds_message_alloc(unsigned int extra_len, gfp_t gfp) +{ + struct rds_message *rm; + + if (extra_len > KMALLOC_MAX_SIZE - sizeof(struct rds_message)) + return NULL; + + rm = kzalloc(sizeof(struct rds_message) + extra_len, gfp); + if (!rm) + goto out; + + rm->m_used_sgs = 0; + rm->m_total_sgs = extra_len / sizeof(struct scatterlist); + + refcount_set(&rm->m_refcount, 1); + INIT_LIST_HEAD(&rm->m_sock_item); + INIT_LIST_HEAD(&rm->m_conn_item); + spin_lock_init(&rm->m_rs_lock); + init_waitqueue_head(&rm->m_flush_wait); + +out: + return rm; +} + +/* + * RDS ops use this to grab SG entries from the rm's sg pool. + */ +struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents) +{ + struct scatterlist *sg_first = (struct scatterlist *) &rm[1]; + struct scatterlist *sg_ret; + + WARN_ON(rm->m_used_sgs + nents > rm->m_total_sgs); + WARN_ON(!nents); + + if (rm->m_used_sgs + nents > rm->m_total_sgs) + return NULL; + + sg_ret = &sg_first[rm->m_used_sgs]; + sg_init_table(sg_ret, nents); + rm->m_used_sgs += nents; + + return sg_ret; +} + +struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len) +{ + struct rds_message *rm; + unsigned int i; + int num_sgs = ceil(total_len, PAGE_SIZE); + int extra_bytes = num_sgs * sizeof(struct scatterlist); + + rm = rds_message_alloc(extra_bytes, GFP_NOWAIT); + if (!rm) + return ERR_PTR(-ENOMEM); + + set_bit(RDS_MSG_PAGEVEC, &rm->m_flags); + rm->m_inc.i_hdr.h_len = cpu_to_be32(total_len); + rm->data.op_nents = ceil(total_len, PAGE_SIZE); + rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs); + if (!rm->data.op_sg) { + rds_message_put(rm); + return ERR_PTR(-ENOMEM); + } + + for (i = 0; i < rm->data.op_nents; ++i) { + sg_set_page(&rm->data.op_sg[i], + virt_to_page(page_addrs[i]), + PAGE_SIZE, 0); + } + + return rm; +} + +static int rds_message_zcopy_from_user(struct rds_message *rm, struct iov_iter *from) +{ + struct scatterlist *sg; + int ret = 0; + int length = iov_iter_count(from); + int total_copied = 0; + struct rds_msg_zcopy_info *info; + + rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from)); + + /* + * now allocate and copy in the data payload. + */ + sg = rm->data.op_sg; + + info = kzalloc(sizeof(*info), GFP_KERNEL); + if (!info) + return -ENOMEM; + INIT_LIST_HEAD(&info->rs_zcookie_next); + rm->data.op_mmp_znotifier = &info->znotif; + if (mm_account_pinned_pages(&rm->data.op_mmp_znotifier->z_mmp, + length)) { + ret = -ENOMEM; + goto err; + } + while (iov_iter_count(from)) { + struct page *pages; + size_t start; + ssize_t copied; + + copied = iov_iter_get_pages(from, &pages, PAGE_SIZE, + 1, &start); + if (copied < 0) { + struct mmpin *mmp; + int i; + + for (i = 0; i < rm->data.op_nents; i++) + put_page(sg_page(&rm->data.op_sg[i])); + mmp = &rm->data.op_mmp_znotifier->z_mmp; + mm_unaccount_pinned_pages(mmp); + ret = -EFAULT; + goto err; + } + total_copied += copied; + iov_iter_advance(from, copied); + length -= copied; + sg_set_page(sg, pages, copied, start); + rm->data.op_nents++; + sg++; + } + WARN_ON_ONCE(length != 0); + return ret; +err: + kfree(info); + rm->data.op_mmp_znotifier = NULL; + return ret; +} + +int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from, + bool zcopy) +{ + unsigned long to_copy, nbytes; + unsigned long sg_off; + struct scatterlist *sg; + int ret = 0; + + rm->m_inc.i_hdr.h_len = cpu_to_be32(iov_iter_count(from)); + + /* now allocate and copy in the data payload. */ + sg = rm->data.op_sg; + sg_off = 0; /* Dear gcc, sg->page will be null from kzalloc. */ + + if (zcopy) + return rds_message_zcopy_from_user(rm, from); + + while (iov_iter_count(from)) { + if (!sg_page(sg)) { + ret = rds_page_remainder_alloc(sg, iov_iter_count(from), + GFP_HIGHUSER); + if (ret) + return ret; + rm->data.op_nents++; + sg_off = 0; + } + + to_copy = min_t(unsigned long, iov_iter_count(from), + sg->length - sg_off); + + rds_stats_add(s_copy_from_user, to_copy); + nbytes = copy_page_from_iter(sg_page(sg), sg->offset + sg_off, + to_copy, from); + if (nbytes != to_copy) + return -EFAULT; + + sg_off += to_copy; + + if (sg_off == sg->length) + sg++; + } + + return ret; +} + +int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to) +{ + struct rds_message *rm; + struct scatterlist *sg; + unsigned long to_copy; + unsigned long vec_off; + int copied; + int ret; + u32 len; + + rm = container_of(inc, struct rds_message, m_inc); + len = be32_to_cpu(rm->m_inc.i_hdr.h_len); + + sg = rm->data.op_sg; + vec_off = 0; + copied = 0; + + while (iov_iter_count(to) && copied < len) { + to_copy = min_t(unsigned long, iov_iter_count(to), + sg->length - vec_off); + to_copy = min_t(unsigned long, to_copy, len - copied); + + rds_stats_add(s_copy_to_user, to_copy); + ret = copy_page_to_iter(sg_page(sg), sg->offset + vec_off, + to_copy, to); + if (ret != to_copy) + return -EFAULT; + + vec_off += to_copy; + copied += to_copy; + + if (vec_off == sg->length) { + vec_off = 0; + sg++; + } + } + + return copied; +} + +/* + * If the message is still on the send queue, wait until the transport + * is done with it. This is particularly important for RDMA operations. + */ +void rds_message_wait(struct rds_message *rm) +{ + wait_event_interruptible(rm->m_flush_wait, + !test_bit(RDS_MSG_MAPPED, &rm->m_flags)); +} + +void rds_message_unmapped(struct rds_message *rm) +{ + clear_bit(RDS_MSG_MAPPED, &rm->m_flags); + wake_up_interruptible(&rm->m_flush_wait); +} +EXPORT_SYMBOL_GPL(rds_message_unmapped); diff --git a/net/rds/page.c b/net/rds/page.c new file mode 100644 index 000000000..7cc57e098 --- /dev/null +++ b/net/rds/page.c @@ -0,0 +1,167 @@ +/* + * Copyright (c) 2006 Oracle. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/highmem.h> +#include <linux/gfp.h> +#include <linux/cpu.h> +#include <linux/export.h> + +#include "rds.h" + +struct rds_page_remainder { + struct page *r_page; + unsigned long r_offset; +}; + +static +DEFINE_PER_CPU_SHARED_ALIGNED(struct rds_page_remainder, rds_page_remainders); + +/** + * rds_page_remainder_alloc - build up regions of a message. + * + * @scat: Scatter list for message + * @bytes: the number of bytes needed. + * @gfp: the waiting behaviour of the allocation + * + * @gfp is always ored with __GFP_HIGHMEM. Callers must be prepared to + * kmap the pages, etc. + * + * If @bytes is at least a full page then this just returns a page from + * alloc_page(). + * + * If @bytes is a partial page then this stores the unused region of the + * page in a per-cpu structure. Future partial-page allocations may be + * satisfied from that cached region. This lets us waste less memory on + * small allocations with minimal complexity. It works because the transmit + * path passes read-only page regions down to devices. They hold a page + * reference until they are done with the region. + */ +int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes, + gfp_t gfp) +{ + struct rds_page_remainder *rem; + unsigned long flags; + struct page *page; + int ret; + + gfp |= __GFP_HIGHMEM; + + /* jump straight to allocation if we're trying for a huge page */ + if (bytes >= PAGE_SIZE) { + page = alloc_page(gfp); + if (!page) { + ret = -ENOMEM; + } else { + sg_set_page(scat, page, PAGE_SIZE, 0); + ret = 0; + } + goto out; + } + + rem = &per_cpu(rds_page_remainders, get_cpu()); + local_irq_save(flags); + + while (1) { + /* avoid a tiny region getting stuck by tossing it */ + if (rem->r_page && bytes > (PAGE_SIZE - rem->r_offset)) { + rds_stats_inc(s_page_remainder_miss); + __free_page(rem->r_page); + rem->r_page = NULL; + } + + /* hand out a fragment from the cached page */ + if (rem->r_page && bytes <= (PAGE_SIZE - rem->r_offset)) { + sg_set_page(scat, rem->r_page, bytes, rem->r_offset); + get_page(sg_page(scat)); + + if (rem->r_offset != 0) + rds_stats_inc(s_page_remainder_hit); + + rem->r_offset += ALIGN(bytes, 8); + if (rem->r_offset >= PAGE_SIZE) { + __free_page(rem->r_page); + rem->r_page = NULL; + } + ret = 0; + break; + } + + /* alloc if there is nothing for us to use */ + local_irq_restore(flags); + put_cpu(); + + page = alloc_page(gfp); + + rem = &per_cpu(rds_page_remainders, get_cpu()); + local_irq_save(flags); + + if (!page) { + ret = -ENOMEM; + break; + } + + /* did someone race to fill the remainder before us? */ + if (rem->r_page) { + __free_page(page); + continue; + } + + /* otherwise install our page and loop around to alloc */ + rem->r_page = page; + rem->r_offset = 0; + } + + local_irq_restore(flags); + put_cpu(); +out: + rdsdebug("bytes %lu ret %d %p %u %u\n", bytes, ret, + ret ? NULL : sg_page(scat), ret ? 0 : scat->offset, + ret ? 0 : scat->length); + return ret; +} +EXPORT_SYMBOL_GPL(rds_page_remainder_alloc); + +void rds_page_exit(void) +{ + unsigned int cpu; + + for_each_possible_cpu(cpu) { + struct rds_page_remainder *rem; + + rem = &per_cpu(rds_page_remainders, cpu); + rdsdebug("cpu %u\n", cpu); + + if (rem->r_page) + __free_page(rem->r_page); + rem->r_page = NULL; + } +} diff --git a/net/rds/rdma.c b/net/rds/rdma.c new file mode 100644 index 000000000..9882cebfc --- /dev/null +++ b/net/rds/rdma.c @@ -0,0 +1,891 @@ +/* + * Copyright (c) 2007, 2017 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/pagemap.h> +#include <linux/slab.h> +#include <linux/rbtree.h> +#include <linux/dma-mapping.h> /* for DMA_*_DEVICE */ + +#include "rds.h" + +/* + * XXX + * - build with sparse + * - should we detect duplicate keys on a socket? hmm. + * - an rdma is an mlock, apply rlimit? + */ + +/* + * get the number of pages by looking at the page indices that the start and + * end addresses fall in. + * + * Returns 0 if the vec is invalid. It is invalid if the number of bytes + * causes the address to wrap or overflows an unsigned int. This comes + * from being stored in the 'length' member of 'struct scatterlist'. + */ +static unsigned int rds_pages_in_vec(struct rds_iovec *vec) +{ + if ((vec->addr + vec->bytes <= vec->addr) || + (vec->bytes > (u64)UINT_MAX)) + return 0; + + return ((vec->addr + vec->bytes + PAGE_SIZE - 1) >> PAGE_SHIFT) - + (vec->addr >> PAGE_SHIFT); +} + +static struct rds_mr *rds_mr_tree_walk(struct rb_root *root, u64 key, + struct rds_mr *insert) +{ + struct rb_node **p = &root->rb_node; + struct rb_node *parent = NULL; + struct rds_mr *mr; + + while (*p) { + parent = *p; + mr = rb_entry(parent, struct rds_mr, r_rb_node); + + if (key < mr->r_key) + p = &(*p)->rb_left; + else if (key > mr->r_key) + p = &(*p)->rb_right; + else + return mr; + } + + if (insert) { + rb_link_node(&insert->r_rb_node, parent, p); + rb_insert_color(&insert->r_rb_node, root); + refcount_inc(&insert->r_refcount); + } + return NULL; +} + +/* + * Destroy the transport-specific part of a MR. + */ +static void rds_destroy_mr(struct rds_mr *mr) +{ + struct rds_sock *rs = mr->r_sock; + void *trans_private = NULL; + unsigned long flags; + + rdsdebug("RDS: destroy mr key is %x refcnt %u\n", + mr->r_key, refcount_read(&mr->r_refcount)); + + if (test_and_set_bit(RDS_MR_DEAD, &mr->r_state)) + return; + + spin_lock_irqsave(&rs->rs_rdma_lock, flags); + if (!RB_EMPTY_NODE(&mr->r_rb_node)) + rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys); + trans_private = mr->r_trans_private; + mr->r_trans_private = NULL; + spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); + + if (trans_private) + mr->r_trans->free_mr(trans_private, mr->r_invalidate); +} + +void __rds_put_mr_final(struct rds_mr *mr) +{ + rds_destroy_mr(mr); + kfree(mr); +} + +/* + * By the time this is called we can't have any more ioctls called on + * the socket so we don't need to worry about racing with others. + */ +void rds_rdma_drop_keys(struct rds_sock *rs) +{ + struct rds_mr *mr; + struct rb_node *node; + unsigned long flags; + + /* Release any MRs associated with this socket */ + spin_lock_irqsave(&rs->rs_rdma_lock, flags); + while ((node = rb_first(&rs->rs_rdma_keys))) { + mr = rb_entry(node, struct rds_mr, r_rb_node); + if (mr->r_trans == rs->rs_transport) + mr->r_invalidate = 0; + rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys); + RB_CLEAR_NODE(&mr->r_rb_node); + spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); + rds_destroy_mr(mr); + rds_mr_put(mr); + spin_lock_irqsave(&rs->rs_rdma_lock, flags); + } + spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); + + if (rs->rs_transport && rs->rs_transport->flush_mrs) + rs->rs_transport->flush_mrs(); +} + +/* + * Helper function to pin user pages. + */ +static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages, + struct page **pages, int write) +{ + int ret; + + ret = get_user_pages_fast(user_addr, nr_pages, write, pages); + + if (ret >= 0 && ret < nr_pages) { + while (ret--) + put_page(pages[ret]); + ret = -EFAULT; + } + + return ret; +} + +static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args, + u64 *cookie_ret, struct rds_mr **mr_ret, + struct rds_conn_path *cp) +{ + struct rds_mr *mr = NULL, *found; + unsigned int nr_pages; + struct page **pages = NULL; + struct scatterlist *sg; + void *trans_private; + unsigned long flags; + rds_rdma_cookie_t cookie; + unsigned int nents; + long i; + int ret; + + if (ipv6_addr_any(&rs->rs_bound_addr) || !rs->rs_transport) { + ret = -ENOTCONN; /* XXX not a great errno */ + goto out; + } + + if (!rs->rs_transport->get_mr) { + ret = -EOPNOTSUPP; + goto out; + } + + nr_pages = rds_pages_in_vec(&args->vec); + if (nr_pages == 0) { + ret = -EINVAL; + goto out; + } + + /* Restrict the size of mr irrespective of underlying transport + * To account for unaligned mr regions, subtract one from nr_pages + */ + if ((nr_pages - 1) > (RDS_MAX_MSG_SIZE >> PAGE_SHIFT)) { + ret = -EMSGSIZE; + goto out; + } + + rdsdebug("RDS: get_mr addr %llx len %llu nr_pages %u\n", + args->vec.addr, args->vec.bytes, nr_pages); + + /* XXX clamp nr_pages to limit the size of this alloc? */ + pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL); + if (!pages) { + ret = -ENOMEM; + goto out; + } + + mr = kzalloc(sizeof(struct rds_mr), GFP_KERNEL); + if (!mr) { + ret = -ENOMEM; + goto out; + } + + refcount_set(&mr->r_refcount, 1); + RB_CLEAR_NODE(&mr->r_rb_node); + mr->r_trans = rs->rs_transport; + mr->r_sock = rs; + + if (args->flags & RDS_RDMA_USE_ONCE) + mr->r_use_once = 1; + if (args->flags & RDS_RDMA_INVALIDATE) + mr->r_invalidate = 1; + if (args->flags & RDS_RDMA_READWRITE) + mr->r_write = 1; + + /* + * Pin the pages that make up the user buffer and transfer the page + * pointers to the mr's sg array. We check to see if we've mapped + * the whole region after transferring the partial page references + * to the sg array so that we can have one page ref cleanup path. + * + * For now we have no flag that tells us whether the mapping is + * r/o or r/w. We need to assume r/w, or we'll do a lot of RDMA to + * the zero page. + */ + ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1); + if (ret < 0) + goto out; + + nents = ret; + sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL); + if (!sg) { + ret = -ENOMEM; + goto out; + } + WARN_ON(!nents); + sg_init_table(sg, nents); + + /* Stick all pages into the scatterlist */ + for (i = 0 ; i < nents; i++) + sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0); + + rdsdebug("RDS: trans_private nents is %u\n", nents); + + /* Obtain a transport specific MR. If this succeeds, the + * s/g list is now owned by the MR. + * Note that dma_map() implies that pending writes are + * flushed to RAM, so no dma_sync is needed here. */ + trans_private = rs->rs_transport->get_mr(sg, nents, rs, + &mr->r_key, + cp ? cp->cp_conn : NULL); + + if (IS_ERR(trans_private)) { + for (i = 0 ; i < nents; i++) + put_page(sg_page(&sg[i])); + kfree(sg); + ret = PTR_ERR(trans_private); + goto out; + } + + mr->r_trans_private = trans_private; + + rdsdebug("RDS: get_mr put_user key is %x cookie_addr %p\n", + mr->r_key, (void *)(unsigned long) args->cookie_addr); + + /* The user may pass us an unaligned address, but we can only + * map page aligned regions. So we keep the offset, and build + * a 64bit cookie containing <R_Key, offset> and pass that + * around. */ + cookie = rds_rdma_make_cookie(mr->r_key, args->vec.addr & ~PAGE_MASK); + if (cookie_ret) + *cookie_ret = cookie; + + if (args->cookie_addr && put_user(cookie, (u64 __user *)(unsigned long) args->cookie_addr)) { + ret = -EFAULT; + goto out; + } + + /* Inserting the new MR into the rbtree bumps its + * reference count. */ + spin_lock_irqsave(&rs->rs_rdma_lock, flags); + found = rds_mr_tree_walk(&rs->rs_rdma_keys, mr->r_key, mr); + spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); + + BUG_ON(found && found != mr); + + rdsdebug("RDS: get_mr key is %x\n", mr->r_key); + if (mr_ret) { + refcount_inc(&mr->r_refcount); + *mr_ret = mr; + } + + ret = 0; +out: + kfree(pages); + if (mr) + rds_mr_put(mr); + return ret; +} + +int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen) +{ + struct rds_get_mr_args args; + + if (optlen != sizeof(struct rds_get_mr_args)) + return -EINVAL; + + if (copy_from_user(&args, (struct rds_get_mr_args __user *)optval, + sizeof(struct rds_get_mr_args))) + return -EFAULT; + + return __rds_rdma_map(rs, &args, NULL, NULL, NULL); +} + +int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen) +{ + struct rds_get_mr_for_dest_args args; + struct rds_get_mr_args new_args; + + if (optlen != sizeof(struct rds_get_mr_for_dest_args)) + return -EINVAL; + + if (copy_from_user(&args, (struct rds_get_mr_for_dest_args __user *)optval, + sizeof(struct rds_get_mr_for_dest_args))) + return -EFAULT; + + /* + * Initially, just behave like get_mr(). + * TODO: Implement get_mr as wrapper around this + * and deprecate it. + */ + new_args.vec = args.vec; + new_args.cookie_addr = args.cookie_addr; + new_args.flags = args.flags; + + return __rds_rdma_map(rs, &new_args, NULL, NULL, NULL); +} + +/* + * Free the MR indicated by the given R_Key + */ +int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen) +{ + struct rds_free_mr_args args; + struct rds_mr *mr; + unsigned long flags; + + if (optlen != sizeof(struct rds_free_mr_args)) + return -EINVAL; + + if (copy_from_user(&args, (struct rds_free_mr_args __user *)optval, + sizeof(struct rds_free_mr_args))) + return -EFAULT; + + /* Special case - a null cookie means flush all unused MRs */ + if (args.cookie == 0) { + if (!rs->rs_transport || !rs->rs_transport->flush_mrs) + return -EINVAL; + rs->rs_transport->flush_mrs(); + return 0; + } + + /* Look up the MR given its R_key and remove it from the rbtree + * so nobody else finds it. + * This should also prevent races with rds_rdma_unuse. + */ + spin_lock_irqsave(&rs->rs_rdma_lock, flags); + mr = rds_mr_tree_walk(&rs->rs_rdma_keys, rds_rdma_cookie_key(args.cookie), NULL); + if (mr) { + rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys); + RB_CLEAR_NODE(&mr->r_rb_node); + if (args.flags & RDS_RDMA_INVALIDATE) + mr->r_invalidate = 1; + } + spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); + + if (!mr) + return -EINVAL; + + /* + * call rds_destroy_mr() ourselves so that we're sure it's done by the time + * we return. If we let rds_mr_put() do it it might not happen until + * someone else drops their ref. + */ + rds_destroy_mr(mr); + rds_mr_put(mr); + return 0; +} + +/* + * This is called when we receive an extension header that + * tells us this MR was used. It allows us to implement + * use_once semantics + */ +void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force) +{ + struct rds_mr *mr; + unsigned long flags; + int zot_me = 0; + + spin_lock_irqsave(&rs->rs_rdma_lock, flags); + mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL); + if (!mr) { + pr_debug("rds: trying to unuse MR with unknown r_key %u!\n", + r_key); + spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); + return; + } + + if (mr->r_use_once || force) { + rb_erase(&mr->r_rb_node, &rs->rs_rdma_keys); + RB_CLEAR_NODE(&mr->r_rb_node); + zot_me = 1; + } + spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); + + /* May have to issue a dma_sync on this memory region. + * Note we could avoid this if the operation was a RDMA READ, + * but at this point we can't tell. */ + if (mr->r_trans->sync_mr) + mr->r_trans->sync_mr(mr->r_trans_private, DMA_FROM_DEVICE); + + /* If the MR was marked as invalidate, this will + * trigger an async flush. */ + if (zot_me) { + rds_destroy_mr(mr); + rds_mr_put(mr); + } +} + +void rds_rdma_free_op(struct rm_rdma_op *ro) +{ + unsigned int i; + + for (i = 0; i < ro->op_nents; i++) { + struct page *page = sg_page(&ro->op_sg[i]); + + /* Mark page dirty if it was possibly modified, which + * is the case for a RDMA_READ which copies from remote + * to local memory */ + if (!ro->op_write) { + WARN_ON(!page->mapping && irqs_disabled()); + set_page_dirty(page); + } + put_page(page); + } + + kfree(ro->op_notifier); + ro->op_notifier = NULL; + ro->op_active = 0; +} + +void rds_atomic_free_op(struct rm_atomic_op *ao) +{ + struct page *page = sg_page(ao->op_sg); + + /* Mark page dirty if it was possibly modified, which + * is the case for a RDMA_READ which copies from remote + * to local memory */ + set_page_dirty(page); + put_page(page); + + kfree(ao->op_notifier); + ao->op_notifier = NULL; + ao->op_active = 0; +} + + +/* + * Count the number of pages needed to describe an incoming iovec array. + */ +static int rds_rdma_pages(struct rds_iovec iov[], int nr_iovecs) +{ + int tot_pages = 0; + unsigned int nr_pages; + unsigned int i; + + /* figure out the number of pages in the vector */ + for (i = 0; i < nr_iovecs; i++) { + nr_pages = rds_pages_in_vec(&iov[i]); + if (nr_pages == 0) + return -EINVAL; + + tot_pages += nr_pages; + + /* + * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1, + * so tot_pages cannot overflow without first going negative. + */ + if (tot_pages < 0) + return -EINVAL; + } + + return tot_pages; +} + +int rds_rdma_extra_size(struct rds_rdma_args *args, + struct rds_iov_vector *iov) +{ + struct rds_iovec *vec; + struct rds_iovec __user *local_vec; + int tot_pages = 0; + unsigned int nr_pages; + unsigned int i; + + local_vec = (struct rds_iovec __user *)(unsigned long) args->local_vec_addr; + + if (args->nr_local == 0) + return -EINVAL; + + if (args->nr_local > UIO_MAXIOV) + return -EMSGSIZE; + + iov->iov = kcalloc(args->nr_local, + sizeof(struct rds_iovec), + GFP_KERNEL); + if (!iov->iov) + return -ENOMEM; + + vec = &iov->iov[0]; + + if (copy_from_user(vec, local_vec, args->nr_local * + sizeof(struct rds_iovec))) + return -EFAULT; + iov->len = args->nr_local; + + /* figure out the number of pages in the vector */ + for (i = 0; i < args->nr_local; i++, vec++) { + + nr_pages = rds_pages_in_vec(vec); + if (nr_pages == 0) + return -EINVAL; + + tot_pages += nr_pages; + + /* + * nr_pages for one entry is limited to (UINT_MAX>>PAGE_SHIFT)+1, + * so tot_pages cannot overflow without first going negative. + */ + if (tot_pages < 0) + return -EINVAL; + } + + return tot_pages * sizeof(struct scatterlist); +} + +/* + * The application asks for a RDMA transfer. + * Extract all arguments and set up the rdma_op + */ +int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, + struct cmsghdr *cmsg, + struct rds_iov_vector *vec) +{ + struct rds_rdma_args *args; + struct rm_rdma_op *op = &rm->rdma; + int nr_pages; + unsigned int nr_bytes; + struct page **pages = NULL; + struct rds_iovec *iovs; + unsigned int i, j; + int ret = 0; + + if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args)) + || rm->rdma.op_active) + return -EINVAL; + + args = CMSG_DATA(cmsg); + + if (ipv6_addr_any(&rs->rs_bound_addr)) { + ret = -ENOTCONN; /* XXX not a great errno */ + goto out_ret; + } + + if (args->nr_local > UIO_MAXIOV) { + ret = -EMSGSIZE; + goto out_ret; + } + + if (vec->len != args->nr_local) { + ret = -EINVAL; + goto out_ret; + } + + iovs = vec->iov; + + nr_pages = rds_rdma_pages(iovs, args->nr_local); + if (nr_pages < 0) { + ret = -EINVAL; + goto out_ret; + } + + pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL); + if (!pages) { + ret = -ENOMEM; + goto out_ret; + } + + op->op_write = !!(args->flags & RDS_RDMA_READWRITE); + op->op_fence = !!(args->flags & RDS_RDMA_FENCE); + op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME); + op->op_silent = !!(args->flags & RDS_RDMA_SILENT); + op->op_active = 1; + op->op_recverr = rs->rs_recverr; + WARN_ON(!nr_pages); + op->op_sg = rds_message_alloc_sgs(rm, nr_pages); + if (!op->op_sg) { + ret = -ENOMEM; + goto out_pages; + } + + if (op->op_notify || op->op_recverr) { + /* We allocate an uninitialized notifier here, because + * we don't want to do that in the completion handler. We + * would have to use GFP_ATOMIC there, and don't want to deal + * with failed allocations. + */ + op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL); + if (!op->op_notifier) { + ret = -ENOMEM; + goto out_pages; + } + op->op_notifier->n_user_token = args->user_token; + op->op_notifier->n_status = RDS_RDMA_SUCCESS; + + /* Enable rmda notification on data operation for composite + * rds messages and make sure notification is enabled only + * for the data operation which follows it so that application + * gets notified only after full message gets delivered. + */ + if (rm->data.op_sg) { + rm->rdma.op_notify = 0; + rm->data.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME); + } + } + + /* The cookie contains the R_Key of the remote memory region, and + * optionally an offset into it. This is how we implement RDMA into + * unaligned memory. + * When setting up the RDMA, we need to add that offset to the + * destination address (which is really an offset into the MR) + * FIXME: We may want to move this into ib_rdma.c + */ + op->op_rkey = rds_rdma_cookie_key(args->cookie); + op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie); + + nr_bytes = 0; + + rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n", + (unsigned long long)args->nr_local, + (unsigned long long)args->remote_vec.addr, + op->op_rkey); + + for (i = 0; i < args->nr_local; i++) { + struct rds_iovec *iov = &iovs[i]; + /* don't need to check, rds_rdma_pages() verified nr will be +nonzero */ + unsigned int nr = rds_pages_in_vec(iov); + + rs->rs_user_addr = iov->addr; + rs->rs_user_bytes = iov->bytes; + + /* If it's a WRITE operation, we want to pin the pages for reading. + * If it's a READ operation, we need to pin the pages for writing. + */ + ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write); + if (ret < 0) + goto out_pages; + else + ret = 0; + + rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n", + nr_bytes, nr, iov->bytes, iov->addr); + + nr_bytes += iov->bytes; + + for (j = 0; j < nr; j++) { + unsigned int offset = iov->addr & ~PAGE_MASK; + struct scatterlist *sg; + + sg = &op->op_sg[op->op_nents + j]; + sg_set_page(sg, pages[j], + min_t(unsigned int, iov->bytes, PAGE_SIZE - offset), + offset); + + rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n", + sg->offset, sg->length, iov->addr, iov->bytes); + + iov->addr += sg->length; + iov->bytes -= sg->length; + } + + op->op_nents += nr; + } + + if (nr_bytes > args->remote_vec.bytes) { + rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n", + nr_bytes, + (unsigned int) args->remote_vec.bytes); + ret = -EINVAL; + goto out_pages; + } + op->op_bytes = nr_bytes; + +out_pages: + kfree(pages); +out_ret: + if (ret) + rds_rdma_free_op(op); + else + rds_stats_inc(s_send_rdma); + + return ret; +} + +/* + * The application wants us to pass an RDMA destination (aka MR) + * to the remote + */ +int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm, + struct cmsghdr *cmsg) +{ + unsigned long flags; + struct rds_mr *mr; + u32 r_key; + int err = 0; + + if (cmsg->cmsg_len < CMSG_LEN(sizeof(rds_rdma_cookie_t)) || + rm->m_rdma_cookie != 0) + return -EINVAL; + + memcpy(&rm->m_rdma_cookie, CMSG_DATA(cmsg), sizeof(rm->m_rdma_cookie)); + + /* We are reusing a previously mapped MR here. Most likely, the + * application has written to the buffer, so we need to explicitly + * flush those writes to RAM. Otherwise the HCA may not see them + * when doing a DMA from that buffer. + */ + r_key = rds_rdma_cookie_key(rm->m_rdma_cookie); + + spin_lock_irqsave(&rs->rs_rdma_lock, flags); + mr = rds_mr_tree_walk(&rs->rs_rdma_keys, r_key, NULL); + if (!mr) + err = -EINVAL; /* invalid r_key */ + else + refcount_inc(&mr->r_refcount); + spin_unlock_irqrestore(&rs->rs_rdma_lock, flags); + + if (mr) { + mr->r_trans->sync_mr(mr->r_trans_private, DMA_TO_DEVICE); + rm->rdma.op_rdma_mr = mr; + } + return err; +} + +/* + * The application passes us an address range it wants to enable RDMA + * to/from. We map the area, and save the <R_Key,offset> pair + * in rm->m_rdma_cookie. This causes it to be sent along to the peer + * in an extension header. + */ +int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm, + struct cmsghdr *cmsg) +{ + if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_get_mr_args)) || + rm->m_rdma_cookie != 0) + return -EINVAL; + + return __rds_rdma_map(rs, CMSG_DATA(cmsg), &rm->m_rdma_cookie, + &rm->rdma.op_rdma_mr, rm->m_conn_path); +} + +/* + * Fill in rds_message for an atomic request. + */ +int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm, + struct cmsghdr *cmsg) +{ + struct page *page = NULL; + struct rds_atomic_args *args; + int ret = 0; + + if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_atomic_args)) + || rm->atomic.op_active) + return -EINVAL; + + args = CMSG_DATA(cmsg); + + /* Nonmasked & masked cmsg ops converted to masked hw ops */ + switch (cmsg->cmsg_type) { + case RDS_CMSG_ATOMIC_FADD: + rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD; + rm->atomic.op_m_fadd.add = args->fadd.add; + rm->atomic.op_m_fadd.nocarry_mask = 0; + break; + case RDS_CMSG_MASKED_ATOMIC_FADD: + rm->atomic.op_type = RDS_ATOMIC_TYPE_FADD; + rm->atomic.op_m_fadd.add = args->m_fadd.add; + rm->atomic.op_m_fadd.nocarry_mask = args->m_fadd.nocarry_mask; + break; + case RDS_CMSG_ATOMIC_CSWP: + rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP; + rm->atomic.op_m_cswp.compare = args->cswp.compare; + rm->atomic.op_m_cswp.swap = args->cswp.swap; + rm->atomic.op_m_cswp.compare_mask = ~0; + rm->atomic.op_m_cswp.swap_mask = ~0; + break; + case RDS_CMSG_MASKED_ATOMIC_CSWP: + rm->atomic.op_type = RDS_ATOMIC_TYPE_CSWP; + rm->atomic.op_m_cswp.compare = args->m_cswp.compare; + rm->atomic.op_m_cswp.swap = args->m_cswp.swap; + rm->atomic.op_m_cswp.compare_mask = args->m_cswp.compare_mask; + rm->atomic.op_m_cswp.swap_mask = args->m_cswp.swap_mask; + break; + default: + BUG(); /* should never happen */ + } + + rm->atomic.op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME); + rm->atomic.op_silent = !!(args->flags & RDS_RDMA_SILENT); + rm->atomic.op_active = 1; + rm->atomic.op_recverr = rs->rs_recverr; + rm->atomic.op_sg = rds_message_alloc_sgs(rm, 1); + if (!rm->atomic.op_sg) { + ret = -ENOMEM; + goto err; + } + + /* verify 8 byte-aligned */ + if (args->local_addr & 0x7) { + ret = -EFAULT; + goto err; + } + + ret = rds_pin_pages(args->local_addr, 1, &page, 1); + if (ret != 1) + goto err; + ret = 0; + + sg_set_page(rm->atomic.op_sg, page, 8, offset_in_page(args->local_addr)); + + if (rm->atomic.op_notify || rm->atomic.op_recverr) { + /* We allocate an uninitialized notifier here, because + * we don't want to do that in the completion handler. We + * would have to use GFP_ATOMIC there, and don't want to deal + * with failed allocations. + */ + rm->atomic.op_notifier = kmalloc(sizeof(*rm->atomic.op_notifier), GFP_KERNEL); + if (!rm->atomic.op_notifier) { + ret = -ENOMEM; + goto err; + } + + rm->atomic.op_notifier->n_user_token = args->user_token; + rm->atomic.op_notifier->n_status = RDS_RDMA_SUCCESS; + } + + rm->atomic.op_rkey = rds_rdma_cookie_key(args->cookie); + rm->atomic.op_remote_addr = args->remote_addr + rds_rdma_cookie_offset(args->cookie); + + return ret; +err: + if (page) + put_page(page); + rm->atomic.op_active = 0; + kfree(rm->atomic.op_notifier); + + return ret; +} diff --git a/net/rds/rdma_transport.c b/net/rds/rdma_transport.c new file mode 100644 index 000000000..6b0f57c83 --- /dev/null +++ b/net/rds/rdma_transport.c @@ -0,0 +1,293 @@ +/* + * Copyright (c) 2009, 2018 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/module.h> +#include <rdma/rdma_cm.h> + +#include "rds_single_path.h" +#include "rdma_transport.h" +#include "ib.h" + +/* Global IPv4 and IPv6 RDS RDMA listener cm_id */ +static struct rdma_cm_id *rds_rdma_listen_id; +#if IS_ENABLED(CONFIG_IPV6) +static struct rdma_cm_id *rds6_rdma_listen_id; +#endif + +static int rds_rdma_cm_event_handler_cmn(struct rdma_cm_id *cm_id, + struct rdma_cm_event *event, + bool isv6) +{ + /* this can be null in the listening path */ + struct rds_connection *conn = cm_id->context; + struct rds_transport *trans; + int ret = 0; + + rdsdebug("conn %p id %p handling event %u (%s)\n", conn, cm_id, + event->event, rdma_event_msg(event->event)); + + if (cm_id->device->node_type == RDMA_NODE_IB_CA) + trans = &rds_ib_transport; + + /* Prevent shutdown from tearing down the connection + * while we're executing. */ + if (conn) { + mutex_lock(&conn->c_cm_lock); + + /* If the connection is being shut down, bail out + * right away. We return 0 so cm_id doesn't get + * destroyed prematurely */ + if (rds_conn_state(conn) == RDS_CONN_DISCONNECTING) { + /* Reject incoming connections while we're tearing + * down an existing one. */ + if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) + ret = 1; + goto out; + } + } + + switch (event->event) { + case RDMA_CM_EVENT_CONNECT_REQUEST: + ret = trans->cm_handle_connect(cm_id, event, isv6); + break; + + case RDMA_CM_EVENT_ADDR_RESOLVED: + /* XXX do we need to clean up if this fails? */ + ret = rdma_resolve_route(cm_id, + RDS_RDMA_RESOLVE_TIMEOUT_MS); + break; + + case RDMA_CM_EVENT_ROUTE_RESOLVED: + /* Connection could have been dropped so make sure the + * cm_id is valid before proceeding + */ + if (conn) { + struct rds_ib_connection *ibic; + + ibic = conn->c_transport_data; + if (ibic && ibic->i_cm_id == cm_id) + ret = trans->cm_initiate_connect(cm_id, isv6); + else + rds_conn_drop(conn); + } + break; + + case RDMA_CM_EVENT_ESTABLISHED: + trans->cm_connect_complete(conn, event); + break; + + case RDMA_CM_EVENT_REJECTED: + rdsdebug("Connection rejected: %s\n", + rdma_reject_msg(cm_id, event->status)); + /* FALLTHROUGH */ + case RDMA_CM_EVENT_ADDR_ERROR: + case RDMA_CM_EVENT_ROUTE_ERROR: + case RDMA_CM_EVENT_CONNECT_ERROR: + case RDMA_CM_EVENT_UNREACHABLE: + case RDMA_CM_EVENT_DEVICE_REMOVAL: + case RDMA_CM_EVENT_ADDR_CHANGE: + if (conn) + rds_conn_drop(conn); + break; + + case RDMA_CM_EVENT_DISCONNECTED: + rdsdebug("DISCONNECT event - dropping connection " + "%pI6c->%pI6c\n", &conn->c_laddr, + &conn->c_faddr); + rds_conn_drop(conn); + break; + + case RDMA_CM_EVENT_TIMEWAIT_EXIT: + if (conn) { + pr_info("RDS: RDMA_CM_EVENT_TIMEWAIT_EXIT event: dropping connection %pI6c->%pI6c\n", + &conn->c_laddr, &conn->c_faddr); + rds_conn_drop(conn); + } + break; + + default: + /* things like device disconnect? */ + printk(KERN_ERR "RDS: unknown event %u (%s)!\n", + event->event, rdma_event_msg(event->event)); + break; + } + +out: + if (conn) + mutex_unlock(&conn->c_cm_lock); + + rdsdebug("id %p event %u (%s) handling ret %d\n", cm_id, event->event, + rdma_event_msg(event->event), ret); + + return ret; +} + +int rds_rdma_cm_event_handler(struct rdma_cm_id *cm_id, + struct rdma_cm_event *event) +{ + return rds_rdma_cm_event_handler_cmn(cm_id, event, false); +} + +#if IS_ENABLED(CONFIG_IPV6) +int rds6_rdma_cm_event_handler(struct rdma_cm_id *cm_id, + struct rdma_cm_event *event) +{ + return rds_rdma_cm_event_handler_cmn(cm_id, event, true); +} +#endif + +static int rds_rdma_listen_init_common(rdma_cm_event_handler handler, + struct sockaddr *sa, + struct rdma_cm_id **ret_cm_id) +{ + struct rdma_cm_id *cm_id; + int ret; + + cm_id = rdma_create_id(&init_net, handler, NULL, + RDMA_PS_TCP, IB_QPT_RC); + if (IS_ERR(cm_id)) { + ret = PTR_ERR(cm_id); + printk(KERN_ERR "RDS/RDMA: failed to setup listener, " + "rdma_create_id() returned %d\n", ret); + return ret; + } + + /* + * XXX I bet this binds the cm_id to a device. If we want to support + * fail-over we'll have to take this into consideration. + */ + ret = rdma_bind_addr(cm_id, sa); + if (ret) { + printk(KERN_ERR "RDS/RDMA: failed to setup listener, " + "rdma_bind_addr() returned %d\n", ret); + goto out; + } + + ret = rdma_listen(cm_id, 128); + if (ret) { + printk(KERN_ERR "RDS/RDMA: failed to setup listener, " + "rdma_listen() returned %d\n", ret); + goto out; + } + + rdsdebug("cm %p listening on port %u\n", cm_id, RDS_PORT); + + *ret_cm_id = cm_id; + cm_id = NULL; +out: + if (cm_id) + rdma_destroy_id(cm_id); + return ret; +} + +/* Initialize the RDS RDMA listeners. We create two listeners for + * compatibility reason. The one on RDS_PORT is used for IPv4 + * requests only. The one on RDS_CM_PORT is used for IPv6 requests + * only. So only IPv6 enabled RDS module will communicate using this + * port. + */ +static int rds_rdma_listen_init(void) +{ + int ret; +#if IS_ENABLED(CONFIG_IPV6) + struct sockaddr_in6 sin6; +#endif + struct sockaddr_in sin; + + sin.sin_family = PF_INET; + sin.sin_addr.s_addr = htonl(INADDR_ANY); + sin.sin_port = htons(RDS_PORT); + ret = rds_rdma_listen_init_common(rds_rdma_cm_event_handler, + (struct sockaddr *)&sin, + &rds_rdma_listen_id); + if (ret != 0) + return ret; + +#if IS_ENABLED(CONFIG_IPV6) + sin6.sin6_family = PF_INET6; + sin6.sin6_addr = in6addr_any; + sin6.sin6_port = htons(RDS_CM_PORT); + sin6.sin6_scope_id = 0; + sin6.sin6_flowinfo = 0; + ret = rds_rdma_listen_init_common(rds6_rdma_cm_event_handler, + (struct sockaddr *)&sin6, + &rds6_rdma_listen_id); + /* Keep going even when IPv6 is not enabled in the system. */ + if (ret != 0) + rdsdebug("Cannot set up IPv6 RDMA listener\n"); +#endif + return 0; +} + +static void rds_rdma_listen_stop(void) +{ + if (rds_rdma_listen_id) { + rdsdebug("cm %p\n", rds_rdma_listen_id); + rdma_destroy_id(rds_rdma_listen_id); + rds_rdma_listen_id = NULL; + } +#if IS_ENABLED(CONFIG_IPV6) + if (rds6_rdma_listen_id) { + rdsdebug("cm %p\n", rds6_rdma_listen_id); + rdma_destroy_id(rds6_rdma_listen_id); + rds6_rdma_listen_id = NULL; + } +#endif +} + +static int rds_rdma_init(void) +{ + int ret; + + ret = rds_ib_init(); + if (ret) + goto out; + + ret = rds_rdma_listen_init(); + if (ret) + rds_ib_exit(); +out: + return ret; +} +module_init(rds_rdma_init); + +static void rds_rdma_exit(void) +{ + /* stop listening first to ensure no new connections are attempted */ + rds_rdma_listen_stop(); + rds_ib_exit(); +} +module_exit(rds_rdma_exit); + +MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>"); +MODULE_DESCRIPTION("RDS: IB transport"); +MODULE_LICENSE("Dual BSD/GPL"); diff --git a/net/rds/rdma_transport.h b/net/rds/rdma_transport.h new file mode 100644 index 000000000..200d3134a --- /dev/null +++ b/net/rds/rdma_transport.h @@ -0,0 +1,25 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _RDMA_TRANSPORT_H +#define _RDMA_TRANSPORT_H + +#include <rdma/ib_verbs.h> +#include <rdma/rdma_cm.h> +#include "rds.h" + +/* RDMA_CM also uses 16385 as the listener port. */ +#define RDS_CM_PORT 16385 + +#define RDS_RDMA_RESOLVE_TIMEOUT_MS 5000 + +int rds_rdma_conn_connect(struct rds_connection *conn); +int rds_rdma_cm_event_handler(struct rdma_cm_id *cm_id, + struct rdma_cm_event *event); +int rds6_rdma_cm_event_handler(struct rdma_cm_id *cm_id, + struct rdma_cm_event *event); + +/* from ib.c */ +extern struct rds_transport rds_ib_transport; +int rds_ib_init(void); +void rds_ib_exit(void); + +#endif diff --git a/net/rds/rds.h b/net/rds/rds.h new file mode 100644 index 000000000..4234ab81b --- /dev/null +++ b/net/rds/rds.h @@ -0,0 +1,1006 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _RDS_RDS_H +#define _RDS_RDS_H + +#include <net/sock.h> +#include <linux/scatterlist.h> +#include <linux/highmem.h> +#include <rdma/rdma_cm.h> +#include <linux/mutex.h> +#include <linux/rds.h> +#include <linux/rhashtable.h> +#include <linux/refcount.h> +#include <linux/in6.h> + +#include "info.h" + +/* + * RDS Network protocol version + */ +#define RDS_PROTOCOL_3_0 0x0300 +#define RDS_PROTOCOL_3_1 0x0301 +#define RDS_PROTOCOL_VERSION RDS_PROTOCOL_3_1 +#define RDS_PROTOCOL_MAJOR(v) ((v) >> 8) +#define RDS_PROTOCOL_MINOR(v) ((v) & 255) +#define RDS_PROTOCOL(maj, min) (((maj) << 8) | min) + +/* The following ports, 16385, 18634, 18635, are registered with IANA as + * the ports to be used for RDS over TCP and UDP. Currently, only RDS over + * TCP and RDS over IB/RDMA are implemented. 18634 is the historical value + * used for the RDMA_CM listener port. RDS/TCP uses port 16385. After + * IPv6 work, RDMA_CM also uses 16385 as the listener port. 18634 is kept + * to ensure compatibility with older RDS modules. Those ports are defined + * in each transport's header file. + */ +#define RDS_PORT 18634 + +#ifdef ATOMIC64_INIT +#define KERNEL_HAS_ATOMIC64 +#endif + +#ifdef RDS_DEBUG +#define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args) +#else +/* sigh, pr_debug() causes unused variable warnings */ +static inline __printf(1, 2) +void rdsdebug(char *fmt, ...) +{ +} +#endif + +/* XXX is there one of these somewhere? */ +#define ceil(x, y) \ + ({ unsigned long __x = (x), __y = (y); (__x + __y - 1) / __y; }) + +#define RDS_FRAG_SHIFT 12 +#define RDS_FRAG_SIZE ((unsigned int)(1 << RDS_FRAG_SHIFT)) + +/* Used to limit both RDMA and non-RDMA RDS message to 1MB */ +#define RDS_MAX_MSG_SIZE ((unsigned int)(1 << 20)) + +#define RDS_CONG_MAP_BYTES (65536 / 8) +#define RDS_CONG_MAP_PAGES (PAGE_ALIGN(RDS_CONG_MAP_BYTES) / PAGE_SIZE) +#define RDS_CONG_MAP_PAGE_BITS (PAGE_SIZE * 8) + +struct rds_cong_map { + struct rb_node m_rb_node; + struct in6_addr m_addr; + wait_queue_head_t m_waitq; + struct list_head m_conn_list; + unsigned long m_page_addrs[RDS_CONG_MAP_PAGES]; +}; + + +/* + * This is how we will track the connection state: + * A connection is always in one of the following + * states. Updates to the state are atomic and imply + * a memory barrier. + */ +enum { + RDS_CONN_DOWN = 0, + RDS_CONN_CONNECTING, + RDS_CONN_DISCONNECTING, + RDS_CONN_UP, + RDS_CONN_RESETTING, + RDS_CONN_ERROR, +}; + +/* Bits for c_flags */ +#define RDS_LL_SEND_FULL 0 +#define RDS_RECONNECT_PENDING 1 +#define RDS_IN_XMIT 2 +#define RDS_RECV_REFILL 3 +#define RDS_DESTROY_PENDING 4 + +/* Max number of multipaths per RDS connection. Must be a power of 2 */ +#define RDS_MPATH_WORKERS 8 +#define RDS_MPATH_HASH(rs, n) (jhash_1word((rs)->rs_bound_port, \ + (rs)->rs_hash_initval) & ((n) - 1)) + +#define IS_CANONICAL(laddr, faddr) (htonl(laddr) < htonl(faddr)) + +/* Per mpath connection state */ +struct rds_conn_path { + struct rds_connection *cp_conn; + struct rds_message *cp_xmit_rm; + unsigned long cp_xmit_sg; + unsigned int cp_xmit_hdr_off; + unsigned int cp_xmit_data_off; + unsigned int cp_xmit_atomic_sent; + unsigned int cp_xmit_rdma_sent; + unsigned int cp_xmit_data_sent; + + spinlock_t cp_lock; /* protect msg queues */ + u64 cp_next_tx_seq; + struct list_head cp_send_queue; + struct list_head cp_retrans; + + u64 cp_next_rx_seq; + + void *cp_transport_data; + + atomic_t cp_state; + unsigned long cp_send_gen; + unsigned long cp_flags; + unsigned long cp_reconnect_jiffies; + struct delayed_work cp_send_w; + struct delayed_work cp_recv_w; + struct delayed_work cp_conn_w; + struct work_struct cp_down_w; + struct mutex cp_cm_lock; /* protect cp_state & cm */ + wait_queue_head_t cp_waitq; + + unsigned int cp_unacked_packets; + unsigned int cp_unacked_bytes; + unsigned int cp_index; +}; + +/* One rds_connection per RDS address pair */ +struct rds_connection { + struct hlist_node c_hash_node; + struct in6_addr c_laddr; + struct in6_addr c_faddr; + int c_dev_if; /* ifindex used for this conn */ + int c_bound_if; /* ifindex of c_laddr */ + unsigned int c_loopback:1, + c_isv6:1, + c_ping_triggered:1, + c_pad_to_32:29; + int c_npaths; + struct rds_connection *c_passive; + struct rds_transport *c_trans; + + struct rds_cong_map *c_lcong; + struct rds_cong_map *c_fcong; + + /* Protocol version */ + unsigned int c_version; + possible_net_t c_net; + + struct list_head c_map_item; + unsigned long c_map_queued; + + struct rds_conn_path *c_path; + wait_queue_head_t c_hs_waitq; /* handshake waitq */ + + u32 c_my_gen_num; + u32 c_peer_gen_num; +}; + +static inline +struct net *rds_conn_net(struct rds_connection *conn) +{ + return read_pnet(&conn->c_net); +} + +static inline +void rds_conn_net_set(struct rds_connection *conn, struct net *net) +{ + write_pnet(&conn->c_net, net); +} + +#define RDS_FLAG_CONG_BITMAP 0x01 +#define RDS_FLAG_ACK_REQUIRED 0x02 +#define RDS_FLAG_RETRANSMITTED 0x04 +#define RDS_MAX_ADV_CREDIT 255 + +/* RDS_FLAG_PROBE_PORT is the reserved sport used for sending a ping + * probe to exchange control information before establishing a connection. + * Currently the control information that is exchanged is the number of + * supported paths. If the peer is a legacy (older kernel revision) peer, + * it would return a pong message without additional control information + * that would then alert the sender that the peer was an older rev. + */ +#define RDS_FLAG_PROBE_PORT 1 +#define RDS_HS_PROBE(sport, dport) \ + ((sport == RDS_FLAG_PROBE_PORT && dport == 0) || \ + (sport == 0 && dport == RDS_FLAG_PROBE_PORT)) +/* + * Maximum space available for extension headers. + */ +#define RDS_HEADER_EXT_SPACE 16 + +struct rds_header { + __be64 h_sequence; + __be64 h_ack; + __be32 h_len; + __be16 h_sport; + __be16 h_dport; + u8 h_flags; + u8 h_credit; + u8 h_padding[4]; + __sum16 h_csum; + + u8 h_exthdr[RDS_HEADER_EXT_SPACE]; +}; + +/* + * Reserved - indicates end of extensions + */ +#define RDS_EXTHDR_NONE 0 + +/* + * This extension header is included in the very + * first message that is sent on a new connection, + * and identifies the protocol level. This will help + * rolling updates if a future change requires breaking + * the protocol. + * NB: This is no longer true for IB, where we do a version + * negotiation during the connection setup phase (protocol + * version information is included in the RDMA CM private data). + */ +#define RDS_EXTHDR_VERSION 1 +struct rds_ext_header_version { + __be32 h_version; +}; + +/* + * This extension header is included in the RDS message + * chasing an RDMA operation. + */ +#define RDS_EXTHDR_RDMA 2 +struct rds_ext_header_rdma { + __be32 h_rdma_rkey; +}; + +/* + * This extension header tells the peer about the + * destination <R_Key,offset> of the requested RDMA + * operation. + */ +#define RDS_EXTHDR_RDMA_DEST 3 +struct rds_ext_header_rdma_dest { + __be32 h_rdma_rkey; + __be32 h_rdma_offset; +}; + +/* Extension header announcing number of paths. + * Implicit length = 2 bytes. + */ +#define RDS_EXTHDR_NPATHS 5 +#define RDS_EXTHDR_GEN_NUM 6 + +#define __RDS_EXTHDR_MAX 16 /* for now */ +#define RDS_RX_MAX_TRACES (RDS_MSG_RX_DGRAM_TRACE_MAX + 1) +#define RDS_MSG_RX_HDR 0 +#define RDS_MSG_RX_START 1 +#define RDS_MSG_RX_END 2 +#define RDS_MSG_RX_CMSG 3 + +struct rds_incoming { + refcount_t i_refcount; + struct list_head i_item; + struct rds_connection *i_conn; + struct rds_conn_path *i_conn_path; + struct rds_header i_hdr; + unsigned long i_rx_jiffies; + struct in6_addr i_saddr; + + rds_rdma_cookie_t i_rdma_cookie; + struct timeval i_rx_tstamp; + u64 i_rx_lat_trace[RDS_RX_MAX_TRACES]; +}; + +struct rds_mr { + struct rb_node r_rb_node; + refcount_t r_refcount; + u32 r_key; + + /* A copy of the creation flags */ + unsigned int r_use_once:1; + unsigned int r_invalidate:1; + unsigned int r_write:1; + + /* This is for RDS_MR_DEAD. + * It would be nice & consistent to make this part of the above + * bit field here, but we need to use test_and_set_bit. + */ + unsigned long r_state; + struct rds_sock *r_sock; /* back pointer to the socket that owns us */ + struct rds_transport *r_trans; + void *r_trans_private; +}; + +/* Flags for mr->r_state */ +#define RDS_MR_DEAD 0 + +static inline rds_rdma_cookie_t rds_rdma_make_cookie(u32 r_key, u32 offset) +{ + return r_key | (((u64) offset) << 32); +} + +static inline u32 rds_rdma_cookie_key(rds_rdma_cookie_t cookie) +{ + return cookie; +} + +static inline u32 rds_rdma_cookie_offset(rds_rdma_cookie_t cookie) +{ + return cookie >> 32; +} + +/* atomic operation types */ +#define RDS_ATOMIC_TYPE_CSWP 0 +#define RDS_ATOMIC_TYPE_FADD 1 + +/* + * m_sock_item and m_conn_item are on lists that are serialized under + * conn->c_lock. m_sock_item has additional meaning in that once it is empty + * the message will not be put back on the retransmit list after being sent. + * messages that are canceled while being sent rely on this. + * + * m_inc is used by loopback so that it can pass an incoming message straight + * back up into the rx path. It embeds a wire header which is also used by + * the send path, which is kind of awkward. + * + * m_sock_item indicates the message's presence on a socket's send or receive + * queue. m_rs will point to that socket. + * + * m_daddr is used by cancellation to prune messages to a given destination. + * + * The RDS_MSG_ON_SOCK and RDS_MSG_ON_CONN flags are used to avoid lock + * nesting. As paths iterate over messages on a sock, or conn, they must + * also lock the conn, or sock, to remove the message from those lists too. + * Testing the flag to determine if the message is still on the lists lets + * us avoid testing the list_head directly. That means each path can use + * the message's list_head to keep it on a local list while juggling locks + * without confusing the other path. + * + * m_ack_seq is an optional field set by transports who need a different + * sequence number range to invalidate. They can use this in a callback + * that they pass to rds_send_drop_acked() to see if each message has been + * acked. The HAS_ACK_SEQ flag can be used to detect messages which haven't + * had ack_seq set yet. + */ +#define RDS_MSG_ON_SOCK 1 +#define RDS_MSG_ON_CONN 2 +#define RDS_MSG_HAS_ACK_SEQ 3 +#define RDS_MSG_ACK_REQUIRED 4 +#define RDS_MSG_RETRANSMITTED 5 +#define RDS_MSG_MAPPED 6 +#define RDS_MSG_PAGEVEC 7 +#define RDS_MSG_FLUSH 8 + +struct rds_znotifier { + struct mmpin z_mmp; + u32 z_cookie; +}; + +struct rds_msg_zcopy_info { + struct list_head rs_zcookie_next; + union { + struct rds_znotifier znotif; + struct rds_zcopy_cookies zcookies; + }; +}; + +struct rds_msg_zcopy_queue { + struct list_head zcookie_head; + spinlock_t lock; /* protects zcookie_head queue */ +}; + +static inline void rds_message_zcopy_queue_init(struct rds_msg_zcopy_queue *q) +{ + spin_lock_init(&q->lock); + INIT_LIST_HEAD(&q->zcookie_head); +} + +struct rds_iov_vector { + struct rds_iovec *iov; + int len; +}; + +struct rds_iov_vector_arr { + struct rds_iov_vector *vec; + int len; + int indx; + int incr; +}; + +struct rds_message { + refcount_t m_refcount; + struct list_head m_sock_item; + struct list_head m_conn_item; + struct rds_incoming m_inc; + u64 m_ack_seq; + struct in6_addr m_daddr; + unsigned long m_flags; + + /* Never access m_rs without holding m_rs_lock. + * Lock nesting is + * rm->m_rs_lock + * -> rs->rs_lock + */ + spinlock_t m_rs_lock; + wait_queue_head_t m_flush_wait; + + struct rds_sock *m_rs; + + /* cookie to send to remote, in rds header */ + rds_rdma_cookie_t m_rdma_cookie; + + unsigned int m_used_sgs; + unsigned int m_total_sgs; + + void *m_final_op; + + struct { + struct rm_atomic_op { + int op_type; + union { + struct { + uint64_t compare; + uint64_t swap; + uint64_t compare_mask; + uint64_t swap_mask; + } op_m_cswp; + struct { + uint64_t add; + uint64_t nocarry_mask; + } op_m_fadd; + }; + + u32 op_rkey; + u64 op_remote_addr; + unsigned int op_notify:1; + unsigned int op_recverr:1; + unsigned int op_mapped:1; + unsigned int op_silent:1; + unsigned int op_active:1; + struct scatterlist *op_sg; + struct rds_notifier *op_notifier; + + struct rds_mr *op_rdma_mr; + } atomic; + struct rm_rdma_op { + u32 op_rkey; + u64 op_remote_addr; + unsigned int op_write:1; + unsigned int op_fence:1; + unsigned int op_notify:1; + unsigned int op_recverr:1; + unsigned int op_mapped:1; + unsigned int op_silent:1; + unsigned int op_active:1; + unsigned int op_bytes; + unsigned int op_nents; + unsigned int op_count; + struct scatterlist *op_sg; + struct rds_notifier *op_notifier; + + struct rds_mr *op_rdma_mr; + } rdma; + struct rm_data_op { + unsigned int op_active:1; + unsigned int op_notify:1; + unsigned int op_nents; + unsigned int op_count; + unsigned int op_dmasg; + unsigned int op_dmaoff; + struct rds_znotifier *op_mmp_znotifier; + struct scatterlist *op_sg; + } data; + }; + + struct rds_conn_path *m_conn_path; +}; + +/* + * The RDS notifier is used (optionally) to tell the application about + * completed RDMA operations. Rather than keeping the whole rds message + * around on the queue, we allocate a small notifier that is put on the + * socket's notifier_list. Notifications are delivered to the application + * through control messages. + */ +struct rds_notifier { + struct list_head n_list; + uint64_t n_user_token; + int n_status; +}; + +/* Available as part of RDS core, so doesn't need to participate + * in get_preferred transport etc + */ +#define RDS_TRANS_LOOP 3 + +/** + * struct rds_transport - transport specific behavioural hooks + * + * @xmit: .xmit is called by rds_send_xmit() to tell the transport to send + * part of a message. The caller serializes on the send_sem so this + * doesn't need to be reentrant for a given conn. The header must be + * sent before the data payload. .xmit must be prepared to send a + * message with no data payload. .xmit should return the number of + * bytes that were sent down the connection, including header bytes. + * Returning 0 tells the caller that it doesn't need to perform any + * additional work now. This is usually the case when the transport has + * filled the sending queue for its connection and will handle + * triggering the rds thread to continue the send when space becomes + * available. Returning -EAGAIN tells the caller to retry the send + * immediately. Returning -ENOMEM tells the caller to retry the send at + * some point in the future. + * + * @conn_shutdown: conn_shutdown stops traffic on the given connection. Once + * it returns the connection can not call rds_recv_incoming(). + * This will only be called once after conn_connect returns + * non-zero success and will The caller serializes this with + * the send and connecting paths (xmit_* and conn_*). The + * transport is responsible for other serialization, including + * rds_recv_incoming(). This is called in process context but + * should try hard not to block. + */ + +struct rds_transport { + char t_name[TRANSNAMSIZ]; + struct list_head t_item; + struct module *t_owner; + unsigned int t_prefer_loopback:1, + t_mp_capable:1; + unsigned int t_type; + + int (*laddr_check)(struct net *net, const struct in6_addr *addr, + __u32 scope_id); + int (*conn_alloc)(struct rds_connection *conn, gfp_t gfp); + void (*conn_free)(void *data); + int (*conn_path_connect)(struct rds_conn_path *cp); + void (*conn_path_shutdown)(struct rds_conn_path *conn); + void (*xmit_path_prepare)(struct rds_conn_path *cp); + void (*xmit_path_complete)(struct rds_conn_path *cp); + int (*xmit)(struct rds_connection *conn, struct rds_message *rm, + unsigned int hdr_off, unsigned int sg, unsigned int off); + int (*xmit_rdma)(struct rds_connection *conn, struct rm_rdma_op *op); + int (*xmit_atomic)(struct rds_connection *conn, struct rm_atomic_op *op); + int (*recv_path)(struct rds_conn_path *cp); + int (*inc_copy_to_user)(struct rds_incoming *inc, struct iov_iter *to); + void (*inc_free)(struct rds_incoming *inc); + + int (*cm_handle_connect)(struct rdma_cm_id *cm_id, + struct rdma_cm_event *event, bool isv6); + int (*cm_initiate_connect)(struct rdma_cm_id *cm_id, bool isv6); + void (*cm_connect_complete)(struct rds_connection *conn, + struct rdma_cm_event *event); + + unsigned int (*stats_info_copy)(struct rds_info_iterator *iter, + unsigned int avail); + void (*exit)(void); + void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg, + struct rds_sock *rs, u32 *key_ret, + struct rds_connection *conn); + void (*sync_mr)(void *trans_private, int direction); + void (*free_mr)(void *trans_private, int invalidate); + void (*flush_mrs)(void); + bool (*t_unloading)(struct rds_connection *conn); +}; + +/* Bind hash table key length. It is the sum of the size of a struct + * in6_addr, a scope_id and a port. + */ +#define RDS_BOUND_KEY_LEN \ + (sizeof(struct in6_addr) + sizeof(__u32) + sizeof(__be16)) + +struct rds_sock { + struct sock rs_sk; + + u64 rs_user_addr; + u64 rs_user_bytes; + + /* + * bound_addr used for both incoming and outgoing, no INADDR_ANY + * support. + */ + struct rhash_head rs_bound_node; + u8 rs_bound_key[RDS_BOUND_KEY_LEN]; + struct sockaddr_in6 rs_bound_sin6; +#define rs_bound_addr rs_bound_sin6.sin6_addr +#define rs_bound_addr_v4 rs_bound_sin6.sin6_addr.s6_addr32[3] +#define rs_bound_port rs_bound_sin6.sin6_port +#define rs_bound_scope_id rs_bound_sin6.sin6_scope_id + struct in6_addr rs_conn_addr; +#define rs_conn_addr_v4 rs_conn_addr.s6_addr32[3] + __be16 rs_conn_port; + struct rds_transport *rs_transport; + + /* + * rds_sendmsg caches the conn it used the last time around. + * This helps avoid costly lookups. + */ + struct rds_connection *rs_conn; + + /* flag indicating we were congested or not */ + int rs_congested; + /* seen congestion (ENOBUFS) when sending? */ + int rs_seen_congestion; + + /* rs_lock protects all these adjacent members before the newline */ + spinlock_t rs_lock; + struct list_head rs_send_queue; + u32 rs_snd_bytes; + int rs_rcv_bytes; + struct list_head rs_notify_queue; /* currently used for failed RDMAs */ + + /* Congestion wake_up. If rs_cong_monitor is set, we use cong_mask + * to decide whether the application should be woken up. + * If not set, we use rs_cong_track to find out whether a cong map + * update arrived. + */ + uint64_t rs_cong_mask; + uint64_t rs_cong_notify; + struct list_head rs_cong_list; + unsigned long rs_cong_track; + + /* + * rs_recv_lock protects the receive queue, and is + * used to serialize with rds_release. + */ + rwlock_t rs_recv_lock; + struct list_head rs_recv_queue; + + /* just for stats reporting */ + struct list_head rs_item; + + /* these have their own lock */ + spinlock_t rs_rdma_lock; + struct rb_root rs_rdma_keys; + + /* Socket options - in case there will be more */ + unsigned char rs_recverr, + rs_cong_monitor; + u32 rs_hash_initval; + + /* Socket receive path trace points*/ + u8 rs_rx_traces; + u8 rs_rx_trace[RDS_MSG_RX_DGRAM_TRACE_MAX]; + struct rds_msg_zcopy_queue rs_zcookie_queue; +}; + +static inline struct rds_sock *rds_sk_to_rs(const struct sock *sk) +{ + return container_of(sk, struct rds_sock, rs_sk); +} +static inline struct sock *rds_rs_to_sk(struct rds_sock *rs) +{ + return &rs->rs_sk; +} + +/* + * The stack assigns sk_sndbuf and sk_rcvbuf to twice the specified value + * to account for overhead. We don't account for overhead, we just apply + * the number of payload bytes to the specified value. + */ +static inline int rds_sk_sndbuf(struct rds_sock *rs) +{ + return rds_rs_to_sk(rs)->sk_sndbuf / 2; +} +static inline int rds_sk_rcvbuf(struct rds_sock *rs) +{ + return rds_rs_to_sk(rs)->sk_rcvbuf / 2; +} + +struct rds_statistics { + uint64_t s_conn_reset; + uint64_t s_recv_drop_bad_checksum; + uint64_t s_recv_drop_old_seq; + uint64_t s_recv_drop_no_sock; + uint64_t s_recv_drop_dead_sock; + uint64_t s_recv_deliver_raced; + uint64_t s_recv_delivered; + uint64_t s_recv_queued; + uint64_t s_recv_immediate_retry; + uint64_t s_recv_delayed_retry; + uint64_t s_recv_ack_required; + uint64_t s_recv_rdma_bytes; + uint64_t s_recv_ping; + uint64_t s_send_queue_empty; + uint64_t s_send_queue_full; + uint64_t s_send_lock_contention; + uint64_t s_send_lock_queue_raced; + uint64_t s_send_immediate_retry; + uint64_t s_send_delayed_retry; + uint64_t s_send_drop_acked; + uint64_t s_send_ack_required; + uint64_t s_send_queued; + uint64_t s_send_rdma; + uint64_t s_send_rdma_bytes; + uint64_t s_send_pong; + uint64_t s_page_remainder_hit; + uint64_t s_page_remainder_miss; + uint64_t s_copy_to_user; + uint64_t s_copy_from_user; + uint64_t s_cong_update_queued; + uint64_t s_cong_update_received; + uint64_t s_cong_send_error; + uint64_t s_cong_send_blocked; + uint64_t s_recv_bytes_added_to_socket; + uint64_t s_recv_bytes_removed_from_socket; + +}; + +/* af_rds.c */ +void rds_sock_addref(struct rds_sock *rs); +void rds_sock_put(struct rds_sock *rs); +void rds_wake_sk_sleep(struct rds_sock *rs); +static inline void __rds_wake_sk_sleep(struct sock *sk) +{ + wait_queue_head_t *waitq = sk_sleep(sk); + + if (!sock_flag(sk, SOCK_DEAD) && waitq) + wake_up(waitq); +} +extern wait_queue_head_t rds_poll_waitq; + + +/* bind.c */ +int rds_bind(struct socket *sock, struct sockaddr *uaddr, int addr_len); +void rds_remove_bound(struct rds_sock *rs); +struct rds_sock *rds_find_bound(const struct in6_addr *addr, __be16 port, + __u32 scope_id); +int rds_bind_lock_init(void); +void rds_bind_lock_destroy(void); + +/* cong.c */ +int rds_cong_get_maps(struct rds_connection *conn); +void rds_cong_add_conn(struct rds_connection *conn); +void rds_cong_remove_conn(struct rds_connection *conn); +void rds_cong_set_bit(struct rds_cong_map *map, __be16 port); +void rds_cong_clear_bit(struct rds_cong_map *map, __be16 port); +int rds_cong_wait(struct rds_cong_map *map, __be16 port, int nonblock, struct rds_sock *rs); +void rds_cong_queue_updates(struct rds_cong_map *map); +void rds_cong_map_updated(struct rds_cong_map *map, uint64_t); +int rds_cong_updated_since(unsigned long *recent); +void rds_cong_add_socket(struct rds_sock *); +void rds_cong_remove_socket(struct rds_sock *); +void rds_cong_exit(void); +struct rds_message *rds_cong_update_alloc(struct rds_connection *conn); + +/* connection.c */ +extern u32 rds_gen_num; +int rds_conn_init(void); +void rds_conn_exit(void); +struct rds_connection *rds_conn_create(struct net *net, + const struct in6_addr *laddr, + const struct in6_addr *faddr, + struct rds_transport *trans, gfp_t gfp, + int dev_if); +struct rds_connection *rds_conn_create_outgoing(struct net *net, + const struct in6_addr *laddr, + const struct in6_addr *faddr, + struct rds_transport *trans, + gfp_t gfp, int dev_if); +void rds_conn_shutdown(struct rds_conn_path *cpath); +void rds_conn_destroy(struct rds_connection *conn); +void rds_conn_drop(struct rds_connection *conn); +void rds_conn_path_drop(struct rds_conn_path *cpath, bool destroy); +void rds_conn_connect_if_down(struct rds_connection *conn); +void rds_conn_path_connect_if_down(struct rds_conn_path *cp); +void rds_for_each_conn_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens, + int (*visitor)(struct rds_connection *, void *), + u64 *buffer, + size_t item_len); + +__printf(2, 3) +void __rds_conn_path_error(struct rds_conn_path *cp, const char *, ...); +#define rds_conn_path_error(cp, fmt...) \ + __rds_conn_path_error(cp, KERN_WARNING "RDS: " fmt) + +static inline int +rds_conn_path_transition(struct rds_conn_path *cp, int old, int new) +{ + return atomic_cmpxchg(&cp->cp_state, old, new) == old; +} + +static inline int +rds_conn_transition(struct rds_connection *conn, int old, int new) +{ + WARN_ON(conn->c_trans->t_mp_capable); + return rds_conn_path_transition(&conn->c_path[0], old, new); +} + +static inline int +rds_conn_path_state(struct rds_conn_path *cp) +{ + return atomic_read(&cp->cp_state); +} + +static inline int +rds_conn_state(struct rds_connection *conn) +{ + WARN_ON(conn->c_trans->t_mp_capable); + return rds_conn_path_state(&conn->c_path[0]); +} + +static inline int +rds_conn_path_up(struct rds_conn_path *cp) +{ + return atomic_read(&cp->cp_state) == RDS_CONN_UP; +} + +static inline int +rds_conn_up(struct rds_connection *conn) +{ + WARN_ON(conn->c_trans->t_mp_capable); + return rds_conn_path_up(&conn->c_path[0]); +} + +static inline int +rds_conn_path_connecting(struct rds_conn_path *cp) +{ + return atomic_read(&cp->cp_state) == RDS_CONN_CONNECTING; +} + +static inline int +rds_conn_connecting(struct rds_connection *conn) +{ + WARN_ON(conn->c_trans->t_mp_capable); + return rds_conn_path_connecting(&conn->c_path[0]); +} + +/* message.c */ +struct rds_message *rds_message_alloc(unsigned int nents, gfp_t gfp); +struct scatterlist *rds_message_alloc_sgs(struct rds_message *rm, int nents); +int rds_message_copy_from_user(struct rds_message *rm, struct iov_iter *from, + bool zcopy); +struct rds_message *rds_message_map_pages(unsigned long *page_addrs, unsigned int total_len); +void rds_message_populate_header(struct rds_header *hdr, __be16 sport, + __be16 dport, u64 seq); +int rds_message_add_extension(struct rds_header *hdr, + unsigned int type, const void *data, unsigned int len); +int rds_message_next_extension(struct rds_header *hdr, + unsigned int *pos, void *buf, unsigned int *buflen); +int rds_message_add_rdma_dest_extension(struct rds_header *hdr, u32 r_key, u32 offset); +int rds_message_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to); +void rds_message_inc_free(struct rds_incoming *inc); +void rds_message_addref(struct rds_message *rm); +void rds_message_put(struct rds_message *rm); +void rds_message_wait(struct rds_message *rm); +void rds_message_unmapped(struct rds_message *rm); +void rds_notify_msg_zcopy_purge(struct rds_msg_zcopy_queue *info); + +static inline void rds_message_make_checksum(struct rds_header *hdr) +{ + hdr->h_csum = 0; + hdr->h_csum = ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2); +} + +static inline int rds_message_verify_checksum(const struct rds_header *hdr) +{ + return !hdr->h_csum || ip_fast_csum((void *) hdr, sizeof(*hdr) >> 2) == 0; +} + + +/* page.c */ +int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes, + gfp_t gfp); +void rds_page_exit(void); + +/* recv.c */ +void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn, + struct in6_addr *saddr); +void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *conn, + struct in6_addr *saddr); +void rds_inc_put(struct rds_incoming *inc); +void rds_recv_incoming(struct rds_connection *conn, struct in6_addr *saddr, + struct in6_addr *daddr, + struct rds_incoming *inc, gfp_t gfp); +int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size, + int msg_flags); +void rds_clear_recv_queue(struct rds_sock *rs); +int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msg); +void rds_inc_info_copy(struct rds_incoming *inc, + struct rds_info_iterator *iter, + __be32 saddr, __be32 daddr, int flip); +void rds6_inc_info_copy(struct rds_incoming *inc, + struct rds_info_iterator *iter, + struct in6_addr *saddr, struct in6_addr *daddr, + int flip); + +/* send.c */ +int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len); +void rds_send_path_reset(struct rds_conn_path *conn); +int rds_send_xmit(struct rds_conn_path *cp); +struct sockaddr_in; +void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest); +typedef int (*is_acked_func)(struct rds_message *rm, uint64_t ack); +void rds_send_drop_acked(struct rds_connection *conn, u64 ack, + is_acked_func is_acked); +void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack, + is_acked_func is_acked); +void rds_send_ping(struct rds_connection *conn, int cp_index); +int rds_send_pong(struct rds_conn_path *cp, __be16 dport); + +/* rdma.c */ +void rds_rdma_unuse(struct rds_sock *rs, u32 r_key, int force); +int rds_get_mr(struct rds_sock *rs, char __user *optval, int optlen); +int rds_get_mr_for_dest(struct rds_sock *rs, char __user *optval, int optlen); +int rds_free_mr(struct rds_sock *rs, char __user *optval, int optlen); +void rds_rdma_drop_keys(struct rds_sock *rs); +int rds_rdma_extra_size(struct rds_rdma_args *args, + struct rds_iov_vector *iov); +int rds_cmsg_rdma_dest(struct rds_sock *rs, struct rds_message *rm, + struct cmsghdr *cmsg); +int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, + struct cmsghdr *cmsg, + struct rds_iov_vector *vec); +int rds_cmsg_rdma_map(struct rds_sock *rs, struct rds_message *rm, + struct cmsghdr *cmsg); +void rds_rdma_free_op(struct rm_rdma_op *ro); +void rds_atomic_free_op(struct rm_atomic_op *ao); +void rds_rdma_send_complete(struct rds_message *rm, int wc_status); +void rds_atomic_send_complete(struct rds_message *rm, int wc_status); +int rds_cmsg_atomic(struct rds_sock *rs, struct rds_message *rm, + struct cmsghdr *cmsg); + +void __rds_put_mr_final(struct rds_mr *mr); +static inline void rds_mr_put(struct rds_mr *mr) +{ + if (refcount_dec_and_test(&mr->r_refcount)) + __rds_put_mr_final(mr); +} + +static inline bool rds_destroy_pending(struct rds_connection *conn) +{ + return !check_net(rds_conn_net(conn)) || + (conn->c_trans->t_unloading && conn->c_trans->t_unloading(conn)); +} + +/* stats.c */ +DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats); +#define rds_stats_inc_which(which, member) do { \ + per_cpu(which, get_cpu()).member++; \ + put_cpu(); \ +} while (0) +#define rds_stats_inc(member) rds_stats_inc_which(rds_stats, member) +#define rds_stats_add_which(which, member, count) do { \ + per_cpu(which, get_cpu()).member += count; \ + put_cpu(); \ +} while (0) +#define rds_stats_add(member, count) rds_stats_add_which(rds_stats, member, count) +int rds_stats_init(void); +void rds_stats_exit(void); +void rds_stats_info_copy(struct rds_info_iterator *iter, + uint64_t *values, const char *const *names, + size_t nr); + +/* sysctl.c */ +int rds_sysctl_init(void); +void rds_sysctl_exit(void); +extern unsigned long rds_sysctl_sndbuf_min; +extern unsigned long rds_sysctl_sndbuf_default; +extern unsigned long rds_sysctl_sndbuf_max; +extern unsigned long rds_sysctl_reconnect_min_jiffies; +extern unsigned long rds_sysctl_reconnect_max_jiffies; +extern unsigned int rds_sysctl_max_unacked_packets; +extern unsigned int rds_sysctl_max_unacked_bytes; +extern unsigned int rds_sysctl_ping_enable; +extern unsigned long rds_sysctl_trace_flags; +extern unsigned int rds_sysctl_trace_level; + +/* threads.c */ +int rds_threads_init(void); +void rds_threads_exit(void); +extern struct workqueue_struct *rds_wq; +void rds_queue_reconnect(struct rds_conn_path *cp); +void rds_connect_worker(struct work_struct *); +void rds_shutdown_worker(struct work_struct *); +void rds_send_worker(struct work_struct *); +void rds_recv_worker(struct work_struct *); +void rds_connect_path_complete(struct rds_conn_path *conn, int curr); +void rds_connect_complete(struct rds_connection *conn); +int rds_addr_cmp(const struct in6_addr *a1, const struct in6_addr *a2); + +/* transport.c */ +void rds_trans_register(struct rds_transport *trans); +void rds_trans_unregister(struct rds_transport *trans); +struct rds_transport *rds_trans_get_preferred(struct net *net, + const struct in6_addr *addr, + __u32 scope_id); +void rds_trans_put(struct rds_transport *trans); +unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter, + unsigned int avail); +struct rds_transport *rds_trans_get(int t_type); +int rds_trans_init(void); +void rds_trans_exit(void); + +#endif diff --git a/net/rds/rds_single_path.h b/net/rds/rds_single_path.h new file mode 100644 index 000000000..9521f6e99 --- /dev/null +++ b/net/rds/rds_single_path.h @@ -0,0 +1,31 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _RDS_RDS_SINGLE_H +#define _RDS_RDS_SINGLE_H + +#define c_xmit_rm c_path[0].cp_xmit_rm +#define c_xmit_sg c_path[0].cp_xmit_sg +#define c_xmit_hdr_off c_path[0].cp_xmit_hdr_off +#define c_xmit_data_off c_path[0].cp_xmit_data_off +#define c_xmit_atomic_sent c_path[0].cp_xmit_atomic_sent +#define c_xmit_rdma_sent c_path[0].cp_xmit_rdma_sent +#define c_xmit_data_sent c_path[0].cp_xmit_data_sent +#define c_lock c_path[0].cp_lock +#define c_next_tx_seq c_path[0].cp_next_tx_seq +#define c_send_queue c_path[0].cp_send_queue +#define c_retrans c_path[0].cp_retrans +#define c_next_rx_seq c_path[0].cp_next_rx_seq +#define c_transport_data c_path[0].cp_transport_data +#define c_state c_path[0].cp_state +#define c_send_gen c_path[0].cp_send_gen +#define c_flags c_path[0].cp_flags +#define c_reconnect_jiffies c_path[0].cp_reconnect_jiffies +#define c_send_w c_path[0].cp_send_w +#define c_recv_w c_path[0].cp_recv_w +#define c_conn_w c_path[0].cp_conn_w +#define c_down_w c_path[0].cp_down_w +#define c_cm_lock c_path[0].cp_cm_lock +#define c_waitq c_path[0].cp_waitq +#define c_unacked_packets c_path[0].cp_unacked_packets +#define c_unacked_bytes c_path[0].cp_unacked_bytes + +#endif /* _RDS_RDS_SINGLE_H */ diff --git a/net/rds/recv.c b/net/rds/recv.c new file mode 100644 index 000000000..ccf0bf283 --- /dev/null +++ b/net/rds/recv.c @@ -0,0 +1,825 @@ +/* + * Copyright (c) 2006, 2019 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/slab.h> +#include <net/sock.h> +#include <linux/in.h> +#include <linux/export.h> +#include <linux/time.h> +#include <linux/rds.h> + +#include "rds.h" + +void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn, + struct in6_addr *saddr) +{ + int i; + + refcount_set(&inc->i_refcount, 1); + INIT_LIST_HEAD(&inc->i_item); + inc->i_conn = conn; + inc->i_saddr = *saddr; + inc->i_rdma_cookie = 0; + inc->i_rx_tstamp.tv_sec = 0; + inc->i_rx_tstamp.tv_usec = 0; + + for (i = 0; i < RDS_RX_MAX_TRACES; i++) + inc->i_rx_lat_trace[i] = 0; +} +EXPORT_SYMBOL_GPL(rds_inc_init); + +void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *cp, + struct in6_addr *saddr) +{ + refcount_set(&inc->i_refcount, 1); + INIT_LIST_HEAD(&inc->i_item); + inc->i_conn = cp->cp_conn; + inc->i_conn_path = cp; + inc->i_saddr = *saddr; + inc->i_rdma_cookie = 0; + inc->i_rx_tstamp.tv_sec = 0; + inc->i_rx_tstamp.tv_usec = 0; +} +EXPORT_SYMBOL_GPL(rds_inc_path_init); + +static void rds_inc_addref(struct rds_incoming *inc) +{ + rdsdebug("addref inc %p ref %d\n", inc, refcount_read(&inc->i_refcount)); + refcount_inc(&inc->i_refcount); +} + +void rds_inc_put(struct rds_incoming *inc) +{ + rdsdebug("put inc %p ref %d\n", inc, refcount_read(&inc->i_refcount)); + if (refcount_dec_and_test(&inc->i_refcount)) { + BUG_ON(!list_empty(&inc->i_item)); + + inc->i_conn->c_trans->inc_free(inc); + } +} +EXPORT_SYMBOL_GPL(rds_inc_put); + +static void rds_recv_rcvbuf_delta(struct rds_sock *rs, struct sock *sk, + struct rds_cong_map *map, + int delta, __be16 port) +{ + int now_congested; + + if (delta == 0) + return; + + rs->rs_rcv_bytes += delta; + if (delta > 0) + rds_stats_add(s_recv_bytes_added_to_socket, delta); + else + rds_stats_add(s_recv_bytes_removed_from_socket, -delta); + + /* loop transport doesn't send/recv congestion updates */ + if (rs->rs_transport->t_type == RDS_TRANS_LOOP) + return; + + now_congested = rs->rs_rcv_bytes > rds_sk_rcvbuf(rs); + + rdsdebug("rs %p (%pI6c:%u) recv bytes %d buf %d " + "now_cong %d delta %d\n", + rs, &rs->rs_bound_addr, + ntohs(rs->rs_bound_port), rs->rs_rcv_bytes, + rds_sk_rcvbuf(rs), now_congested, delta); + + /* wasn't -> am congested */ + if (!rs->rs_congested && now_congested) { + rs->rs_congested = 1; + rds_cong_set_bit(map, port); + rds_cong_queue_updates(map); + } + /* was -> aren't congested */ + /* Require more free space before reporting uncongested to prevent + bouncing cong/uncong state too often */ + else if (rs->rs_congested && (rs->rs_rcv_bytes < (rds_sk_rcvbuf(rs)/2))) { + rs->rs_congested = 0; + rds_cong_clear_bit(map, port); + rds_cong_queue_updates(map); + } + + /* do nothing if no change in cong state */ +} + +static void rds_conn_peer_gen_update(struct rds_connection *conn, + u32 peer_gen_num) +{ + int i; + struct rds_message *rm, *tmp; + unsigned long flags; + + WARN_ON(conn->c_trans->t_type != RDS_TRANS_TCP); + if (peer_gen_num != 0) { + if (conn->c_peer_gen_num != 0 && + peer_gen_num != conn->c_peer_gen_num) { + for (i = 0; i < RDS_MPATH_WORKERS; i++) { + struct rds_conn_path *cp; + + cp = &conn->c_path[i]; + spin_lock_irqsave(&cp->cp_lock, flags); + cp->cp_next_tx_seq = 1; + cp->cp_next_rx_seq = 0; + list_for_each_entry_safe(rm, tmp, + &cp->cp_retrans, + m_conn_item) { + set_bit(RDS_MSG_FLUSH, &rm->m_flags); + } + spin_unlock_irqrestore(&cp->cp_lock, flags); + } + } + conn->c_peer_gen_num = peer_gen_num; + } +} + +/* + * Process all extension headers that come with this message. + */ +static void rds_recv_incoming_exthdrs(struct rds_incoming *inc, struct rds_sock *rs) +{ + struct rds_header *hdr = &inc->i_hdr; + unsigned int pos = 0, type, len; + union { + struct rds_ext_header_version version; + struct rds_ext_header_rdma rdma; + struct rds_ext_header_rdma_dest rdma_dest; + } buffer; + + while (1) { + len = sizeof(buffer); + type = rds_message_next_extension(hdr, &pos, &buffer, &len); + if (type == RDS_EXTHDR_NONE) + break; + /* Process extension header here */ + switch (type) { + case RDS_EXTHDR_RDMA: + rds_rdma_unuse(rs, be32_to_cpu(buffer.rdma.h_rdma_rkey), 0); + break; + + case RDS_EXTHDR_RDMA_DEST: + /* We ignore the size for now. We could stash it + * somewhere and use it for error checking. */ + inc->i_rdma_cookie = rds_rdma_make_cookie( + be32_to_cpu(buffer.rdma_dest.h_rdma_rkey), + be32_to_cpu(buffer.rdma_dest.h_rdma_offset)); + + break; + } + } +} + +static void rds_recv_hs_exthdrs(struct rds_header *hdr, + struct rds_connection *conn) +{ + unsigned int pos = 0, type, len; + union { + struct rds_ext_header_version version; + u16 rds_npaths; + u32 rds_gen_num; + } buffer; + u32 new_peer_gen_num = 0; + + while (1) { + len = sizeof(buffer); + type = rds_message_next_extension(hdr, &pos, &buffer, &len); + if (type == RDS_EXTHDR_NONE) + break; + /* Process extension header here */ + switch (type) { + case RDS_EXTHDR_NPATHS: + conn->c_npaths = min_t(int, RDS_MPATH_WORKERS, + be16_to_cpu(buffer.rds_npaths)); + break; + case RDS_EXTHDR_GEN_NUM: + new_peer_gen_num = be32_to_cpu(buffer.rds_gen_num); + break; + default: + pr_warn_ratelimited("ignoring unknown exthdr type " + "0x%x\n", type); + } + } + /* if RDS_EXTHDR_NPATHS was not found, default to a single-path */ + conn->c_npaths = max_t(int, conn->c_npaths, 1); + conn->c_ping_triggered = 0; + rds_conn_peer_gen_update(conn, new_peer_gen_num); +} + +/* rds_start_mprds() will synchronously start multiple paths when appropriate. + * The scheme is based on the following rules: + * + * 1. rds_sendmsg on first connect attempt sends the probe ping, with the + * sender's npaths (s_npaths) + * 2. rcvr of probe-ping knows the mprds_paths = min(s_npaths, r_npaths). It + * sends back a probe-pong with r_npaths. After that, if rcvr is the + * smaller ip addr, it starts rds_conn_path_connect_if_down on all + * mprds_paths. + * 3. sender gets woken up, and can move to rds_conn_path_connect_if_down. + * If it is the smaller ipaddr, rds_conn_path_connect_if_down can be + * called after reception of the probe-pong on all mprds_paths. + * Otherwise (sender of probe-ping is not the smaller ip addr): just call + * rds_conn_path_connect_if_down on the hashed path. (see rule 4) + * 4. rds_connect_worker must only trigger a connection if laddr < faddr. + * 5. sender may end up queuing the packet on the cp. will get sent out later. + * when connection is completed. + */ +static void rds_start_mprds(struct rds_connection *conn) +{ + int i; + struct rds_conn_path *cp; + + if (conn->c_npaths > 1 && + rds_addr_cmp(&conn->c_laddr, &conn->c_faddr) < 0) { + for (i = 0; i < conn->c_npaths; i++) { + cp = &conn->c_path[i]; + rds_conn_path_connect_if_down(cp); + } + } +} + +/* + * The transport must make sure that this is serialized against other + * rx and conn reset on this specific conn. + * + * We currently assert that only one fragmented message will be sent + * down a connection at a time. This lets us reassemble in the conn + * instead of per-flow which means that we don't have to go digging through + * flows to tear down partial reassembly progress on conn failure and + * we save flow lookup and locking for each frag arrival. It does mean + * that small messages will wait behind large ones. Fragmenting at all + * is only to reduce the memory consumption of pre-posted buffers. + * + * The caller passes in saddr and daddr instead of us getting it from the + * conn. This lets loopback, who only has one conn for both directions, + * tell us which roles the addrs in the conn are playing for this message. + */ +void rds_recv_incoming(struct rds_connection *conn, struct in6_addr *saddr, + struct in6_addr *daddr, + struct rds_incoming *inc, gfp_t gfp) +{ + struct rds_sock *rs = NULL; + struct sock *sk; + unsigned long flags; + struct rds_conn_path *cp; + + inc->i_conn = conn; + inc->i_rx_jiffies = jiffies; + if (conn->c_trans->t_mp_capable) + cp = inc->i_conn_path; + else + cp = &conn->c_path[0]; + + rdsdebug("conn %p next %llu inc %p seq %llu len %u sport %u dport %u " + "flags 0x%x rx_jiffies %lu\n", conn, + (unsigned long long)cp->cp_next_rx_seq, + inc, + (unsigned long long)be64_to_cpu(inc->i_hdr.h_sequence), + be32_to_cpu(inc->i_hdr.h_len), + be16_to_cpu(inc->i_hdr.h_sport), + be16_to_cpu(inc->i_hdr.h_dport), + inc->i_hdr.h_flags, + inc->i_rx_jiffies); + + /* + * Sequence numbers should only increase. Messages get their + * sequence number as they're queued in a sending conn. They + * can be dropped, though, if the sending socket is closed before + * they hit the wire. So sequence numbers can skip forward + * under normal operation. They can also drop back in the conn + * failover case as previously sent messages are resent down the + * new instance of a conn. We drop those, otherwise we have + * to assume that the next valid seq does not come after a + * hole in the fragment stream. + * + * The headers don't give us a way to realize if fragments of + * a message have been dropped. We assume that frags that arrive + * to a flow are part of the current message on the flow that is + * being reassembled. This means that senders can't drop messages + * from the sending conn until all their frags are sent. + * + * XXX we could spend more on the wire to get more robust failure + * detection, arguably worth it to avoid data corruption. + */ + if (be64_to_cpu(inc->i_hdr.h_sequence) < cp->cp_next_rx_seq && + (inc->i_hdr.h_flags & RDS_FLAG_RETRANSMITTED)) { + rds_stats_inc(s_recv_drop_old_seq); + goto out; + } + cp->cp_next_rx_seq = be64_to_cpu(inc->i_hdr.h_sequence) + 1; + + if (rds_sysctl_ping_enable && inc->i_hdr.h_dport == 0) { + if (inc->i_hdr.h_sport == 0) { + rdsdebug("ignore ping with 0 sport from %pI6c\n", + saddr); + goto out; + } + rds_stats_inc(s_recv_ping); + rds_send_pong(cp, inc->i_hdr.h_sport); + /* if this is a handshake ping, start multipath if necessary */ + if (RDS_HS_PROBE(be16_to_cpu(inc->i_hdr.h_sport), + be16_to_cpu(inc->i_hdr.h_dport))) { + rds_recv_hs_exthdrs(&inc->i_hdr, cp->cp_conn); + rds_start_mprds(cp->cp_conn); + } + goto out; + } + + if (be16_to_cpu(inc->i_hdr.h_dport) == RDS_FLAG_PROBE_PORT && + inc->i_hdr.h_sport == 0) { + rds_recv_hs_exthdrs(&inc->i_hdr, cp->cp_conn); + /* if this is a handshake pong, start multipath if necessary */ + rds_start_mprds(cp->cp_conn); + wake_up(&cp->cp_conn->c_hs_waitq); + goto out; + } + + rs = rds_find_bound(daddr, inc->i_hdr.h_dport, conn->c_bound_if); + if (!rs) { + rds_stats_inc(s_recv_drop_no_sock); + goto out; + } + + /* Process extension headers */ + rds_recv_incoming_exthdrs(inc, rs); + + /* We can be racing with rds_release() which marks the socket dead. */ + sk = rds_rs_to_sk(rs); + + /* serialize with rds_release -> sock_orphan */ + write_lock_irqsave(&rs->rs_recv_lock, flags); + if (!sock_flag(sk, SOCK_DEAD)) { + rdsdebug("adding inc %p to rs %p's recv queue\n", inc, rs); + rds_stats_inc(s_recv_queued); + rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong, + be32_to_cpu(inc->i_hdr.h_len), + inc->i_hdr.h_dport); + if (sock_flag(sk, SOCK_RCVTSTAMP)) + do_gettimeofday(&inc->i_rx_tstamp); + rds_inc_addref(inc); + inc->i_rx_lat_trace[RDS_MSG_RX_END] = local_clock(); + list_add_tail(&inc->i_item, &rs->rs_recv_queue); + __rds_wake_sk_sleep(sk); + } else { + rds_stats_inc(s_recv_drop_dead_sock); + } + write_unlock_irqrestore(&rs->rs_recv_lock, flags); + +out: + if (rs) + rds_sock_put(rs); +} +EXPORT_SYMBOL_GPL(rds_recv_incoming); + +/* + * be very careful here. This is being called as the condition in + * wait_event_*() needs to cope with being called many times. + */ +static int rds_next_incoming(struct rds_sock *rs, struct rds_incoming **inc) +{ + unsigned long flags; + + if (!*inc) { + read_lock_irqsave(&rs->rs_recv_lock, flags); + if (!list_empty(&rs->rs_recv_queue)) { + *inc = list_entry(rs->rs_recv_queue.next, + struct rds_incoming, + i_item); + rds_inc_addref(*inc); + } + read_unlock_irqrestore(&rs->rs_recv_lock, flags); + } + + return *inc != NULL; +} + +static int rds_still_queued(struct rds_sock *rs, struct rds_incoming *inc, + int drop) +{ + struct sock *sk = rds_rs_to_sk(rs); + int ret = 0; + unsigned long flags; + + write_lock_irqsave(&rs->rs_recv_lock, flags); + if (!list_empty(&inc->i_item)) { + ret = 1; + if (drop) { + /* XXX make sure this i_conn is reliable */ + rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong, + -be32_to_cpu(inc->i_hdr.h_len), + inc->i_hdr.h_dport); + list_del_init(&inc->i_item); + rds_inc_put(inc); + } + } + write_unlock_irqrestore(&rs->rs_recv_lock, flags); + + rdsdebug("inc %p rs %p still %d dropped %d\n", inc, rs, ret, drop); + return ret; +} + +/* + * Pull errors off the error queue. + * If msghdr is NULL, we will just purge the error queue. + */ +int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msghdr) +{ + struct rds_notifier *notifier; + struct rds_rdma_notify cmsg; + unsigned int count = 0, max_messages = ~0U; + unsigned long flags; + LIST_HEAD(copy); + int err = 0; + + memset(&cmsg, 0, sizeof(cmsg)); /* fill holes with zero */ + + /* put_cmsg copies to user space and thus may sleep. We can't do this + * with rs_lock held, so first grab as many notifications as we can stuff + * in the user provided cmsg buffer. We don't try to copy more, to avoid + * losing notifications - except when the buffer is so small that it wouldn't + * even hold a single notification. Then we give him as much of this single + * msg as we can squeeze in, and set MSG_CTRUNC. + */ + if (msghdr) { + max_messages = msghdr->msg_controllen / CMSG_SPACE(sizeof(cmsg)); + if (!max_messages) + max_messages = 1; + } + + spin_lock_irqsave(&rs->rs_lock, flags); + while (!list_empty(&rs->rs_notify_queue) && count < max_messages) { + notifier = list_entry(rs->rs_notify_queue.next, + struct rds_notifier, n_list); + list_move(¬ifier->n_list, ©); + count++; + } + spin_unlock_irqrestore(&rs->rs_lock, flags); + + if (!count) + return 0; + + while (!list_empty(©)) { + notifier = list_entry(copy.next, struct rds_notifier, n_list); + + if (msghdr) { + cmsg.user_token = notifier->n_user_token; + cmsg.status = notifier->n_status; + + err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_RDMA_STATUS, + sizeof(cmsg), &cmsg); + if (err) + break; + } + + list_del_init(¬ifier->n_list); + kfree(notifier); + } + + /* If we bailed out because of an error in put_cmsg, + * we may be left with one or more notifications that we + * didn't process. Return them to the head of the list. */ + if (!list_empty(©)) { + spin_lock_irqsave(&rs->rs_lock, flags); + list_splice(©, &rs->rs_notify_queue); + spin_unlock_irqrestore(&rs->rs_lock, flags); + } + + return err; +} + +/* + * Queue a congestion notification + */ +static int rds_notify_cong(struct rds_sock *rs, struct msghdr *msghdr) +{ + uint64_t notify = rs->rs_cong_notify; + unsigned long flags; + int err; + + err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_CONG_UPDATE, + sizeof(notify), ¬ify); + if (err) + return err; + + spin_lock_irqsave(&rs->rs_lock, flags); + rs->rs_cong_notify &= ~notify; + spin_unlock_irqrestore(&rs->rs_lock, flags); + + return 0; +} + +/* + * Receive any control messages. + */ +static int rds_cmsg_recv(struct rds_incoming *inc, struct msghdr *msg, + struct rds_sock *rs) +{ + int ret = 0; + + if (inc->i_rdma_cookie) { + ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RDMA_DEST, + sizeof(inc->i_rdma_cookie), &inc->i_rdma_cookie); + if (ret) + goto out; + } + + if ((inc->i_rx_tstamp.tv_sec != 0) && + sock_flag(rds_rs_to_sk(rs), SOCK_RCVTSTAMP)) { + ret = put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP, + sizeof(struct timeval), + &inc->i_rx_tstamp); + if (ret) + goto out; + } + + if (rs->rs_rx_traces) { + struct rds_cmsg_rx_trace t; + int i, j; + + memset(&t, 0, sizeof(t)); + inc->i_rx_lat_trace[RDS_MSG_RX_CMSG] = local_clock(); + t.rx_traces = rs->rs_rx_traces; + for (i = 0; i < rs->rs_rx_traces; i++) { + j = rs->rs_rx_trace[i]; + t.rx_trace_pos[i] = j; + t.rx_trace[i] = inc->i_rx_lat_trace[j + 1] - + inc->i_rx_lat_trace[j]; + } + + ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RXPATH_LATENCY, + sizeof(t), &t); + if (ret) + goto out; + } + +out: + return ret; +} + +static bool rds_recvmsg_zcookie(struct rds_sock *rs, struct msghdr *msg) +{ + struct rds_msg_zcopy_queue *q = &rs->rs_zcookie_queue; + struct rds_msg_zcopy_info *info = NULL; + struct rds_zcopy_cookies *done; + unsigned long flags; + + if (!msg->msg_control) + return false; + + if (!sock_flag(rds_rs_to_sk(rs), SOCK_ZEROCOPY) || + msg->msg_controllen < CMSG_SPACE(sizeof(*done))) + return false; + + spin_lock_irqsave(&q->lock, flags); + if (!list_empty(&q->zcookie_head)) { + info = list_entry(q->zcookie_head.next, + struct rds_msg_zcopy_info, rs_zcookie_next); + list_del(&info->rs_zcookie_next); + } + spin_unlock_irqrestore(&q->lock, flags); + if (!info) + return false; + done = &info->zcookies; + if (put_cmsg(msg, SOL_RDS, RDS_CMSG_ZCOPY_COMPLETION, sizeof(*done), + done)) { + spin_lock_irqsave(&q->lock, flags); + list_add(&info->rs_zcookie_next, &q->zcookie_head); + spin_unlock_irqrestore(&q->lock, flags); + return false; + } + kfree(info); + return true; +} + +int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size, + int msg_flags) +{ + struct sock *sk = sock->sk; + struct rds_sock *rs = rds_sk_to_rs(sk); + long timeo; + int ret = 0, nonblock = msg_flags & MSG_DONTWAIT; + DECLARE_SOCKADDR(struct sockaddr_in6 *, sin6, msg->msg_name); + DECLARE_SOCKADDR(struct sockaddr_in *, sin, msg->msg_name); + struct rds_incoming *inc = NULL; + + /* udp_recvmsg()->sock_recvtimeo() gets away without locking too.. */ + timeo = sock_rcvtimeo(sk, nonblock); + + rdsdebug("size %zu flags 0x%x timeo %ld\n", size, msg_flags, timeo); + + if (msg_flags & MSG_OOB) + goto out; + if (msg_flags & MSG_ERRQUEUE) + return sock_recv_errqueue(sk, msg, size, SOL_IP, IP_RECVERR); + + while (1) { + /* If there are pending notifications, do those - and nothing else */ + if (!list_empty(&rs->rs_notify_queue)) { + ret = rds_notify_queue_get(rs, msg); + break; + } + + if (rs->rs_cong_notify) { + ret = rds_notify_cong(rs, msg); + break; + } + + if (!rds_next_incoming(rs, &inc)) { + if (nonblock) { + bool reaped = rds_recvmsg_zcookie(rs, msg); + + ret = reaped ? 0 : -EAGAIN; + break; + } + + timeo = wait_event_interruptible_timeout(*sk_sleep(sk), + (!list_empty(&rs->rs_notify_queue) || + rs->rs_cong_notify || + rds_next_incoming(rs, &inc)), timeo); + rdsdebug("recvmsg woke inc %p timeo %ld\n", inc, + timeo); + if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT) + continue; + + ret = timeo; + if (ret == 0) + ret = -ETIMEDOUT; + break; + } + + rdsdebug("copying inc %p from %pI6c:%u to user\n", inc, + &inc->i_conn->c_faddr, + ntohs(inc->i_hdr.h_sport)); + ret = inc->i_conn->c_trans->inc_copy_to_user(inc, &msg->msg_iter); + if (ret < 0) + break; + + /* + * if the message we just copied isn't at the head of the + * recv queue then someone else raced us to return it, try + * to get the next message. + */ + if (!rds_still_queued(rs, inc, !(msg_flags & MSG_PEEK))) { + rds_inc_put(inc); + inc = NULL; + rds_stats_inc(s_recv_deliver_raced); + iov_iter_revert(&msg->msg_iter, ret); + continue; + } + + if (ret < be32_to_cpu(inc->i_hdr.h_len)) { + if (msg_flags & MSG_TRUNC) + ret = be32_to_cpu(inc->i_hdr.h_len); + msg->msg_flags |= MSG_TRUNC; + } + + if (rds_cmsg_recv(inc, msg, rs)) { + ret = -EFAULT; + break; + } + rds_recvmsg_zcookie(rs, msg); + + rds_stats_inc(s_recv_delivered); + + if (msg->msg_name) { + if (ipv6_addr_v4mapped(&inc->i_saddr)) { + sin = (struct sockaddr_in *)msg->msg_name; + + sin->sin_family = AF_INET; + sin->sin_port = inc->i_hdr.h_sport; + sin->sin_addr.s_addr = + inc->i_saddr.s6_addr32[3]; + memset(sin->sin_zero, 0, sizeof(sin->sin_zero)); + msg->msg_namelen = sizeof(*sin); + } else { + sin6 = (struct sockaddr_in6 *)msg->msg_name; + + sin6->sin6_family = AF_INET6; + sin6->sin6_port = inc->i_hdr.h_sport; + sin6->sin6_addr = inc->i_saddr; + sin6->sin6_flowinfo = 0; + sin6->sin6_scope_id = rs->rs_bound_scope_id; + msg->msg_namelen = sizeof(*sin6); + } + } + break; + } + + if (inc) + rds_inc_put(inc); + +out: + return ret; +} + +/* + * The socket is being shut down and we're asked to drop messages that were + * queued for recvmsg. The caller has unbound the socket so the receive path + * won't queue any more incoming fragments or messages on the socket. + */ +void rds_clear_recv_queue(struct rds_sock *rs) +{ + struct sock *sk = rds_rs_to_sk(rs); + struct rds_incoming *inc, *tmp; + unsigned long flags; + + write_lock_irqsave(&rs->rs_recv_lock, flags); + list_for_each_entry_safe(inc, tmp, &rs->rs_recv_queue, i_item) { + rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong, + -be32_to_cpu(inc->i_hdr.h_len), + inc->i_hdr.h_dport); + list_del_init(&inc->i_item); + rds_inc_put(inc); + } + write_unlock_irqrestore(&rs->rs_recv_lock, flags); +} + +/* + * inc->i_saddr isn't used here because it is only set in the receive + * path. + */ +void rds_inc_info_copy(struct rds_incoming *inc, + struct rds_info_iterator *iter, + __be32 saddr, __be32 daddr, int flip) +{ + struct rds_info_message minfo; + + minfo.seq = be64_to_cpu(inc->i_hdr.h_sequence); + minfo.len = be32_to_cpu(inc->i_hdr.h_len); + + if (flip) { + minfo.laddr = daddr; + minfo.faddr = saddr; + minfo.lport = inc->i_hdr.h_dport; + minfo.fport = inc->i_hdr.h_sport; + } else { + minfo.laddr = saddr; + minfo.faddr = daddr; + minfo.lport = inc->i_hdr.h_sport; + minfo.fport = inc->i_hdr.h_dport; + } + + minfo.flags = 0; + + rds_info_copy(iter, &minfo, sizeof(minfo)); +} + +#if IS_ENABLED(CONFIG_IPV6) +void rds6_inc_info_copy(struct rds_incoming *inc, + struct rds_info_iterator *iter, + struct in6_addr *saddr, struct in6_addr *daddr, + int flip) +{ + struct rds6_info_message minfo6; + + minfo6.seq = be64_to_cpu(inc->i_hdr.h_sequence); + minfo6.len = be32_to_cpu(inc->i_hdr.h_len); + minfo6.tos = 0; + + if (flip) { + minfo6.laddr = *daddr; + minfo6.faddr = *saddr; + minfo6.lport = inc->i_hdr.h_dport; + minfo6.fport = inc->i_hdr.h_sport; + } else { + minfo6.laddr = *saddr; + minfo6.faddr = *daddr; + minfo6.lport = inc->i_hdr.h_sport; + minfo6.fport = inc->i_hdr.h_dport; + } + + minfo6.flags = 0; + + rds_info_copy(iter, &minfo6, sizeof(minfo6)); +} +#endif diff --git a/net/rds/send.c b/net/rds/send.c new file mode 100644 index 000000000..26e2c2305 --- /dev/null +++ b/net/rds/send.c @@ -0,0 +1,1500 @@ +/* + * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/moduleparam.h> +#include <linux/gfp.h> +#include <net/sock.h> +#include <linux/in.h> +#include <linux/list.h> +#include <linux/ratelimit.h> +#include <linux/export.h> +#include <linux/sizes.h> + +#include "rds.h" + +/* When transmitting messages in rds_send_xmit, we need to emerge from + * time to time and briefly release the CPU. Otherwise the softlock watchdog + * will kick our shin. + * Also, it seems fairer to not let one busy connection stall all the + * others. + * + * send_batch_count is the number of times we'll loop in send_xmit. Setting + * it to 0 will restore the old behavior (where we looped until we had + * drained the queue). + */ +static int send_batch_count = SZ_1K; +module_param(send_batch_count, int, 0444); +MODULE_PARM_DESC(send_batch_count, " batch factor when working the send queue"); + +static void rds_send_remove_from_sock(struct list_head *messages, int status); + +/* + * Reset the send state. Callers must ensure that this doesn't race with + * rds_send_xmit(). + */ +void rds_send_path_reset(struct rds_conn_path *cp) +{ + struct rds_message *rm, *tmp; + unsigned long flags; + + if (cp->cp_xmit_rm) { + rm = cp->cp_xmit_rm; + cp->cp_xmit_rm = NULL; + /* Tell the user the RDMA op is no longer mapped by the + * transport. This isn't entirely true (it's flushed out + * independently) but as the connection is down, there's + * no ongoing RDMA to/from that memory */ + rds_message_unmapped(rm); + rds_message_put(rm); + } + + cp->cp_xmit_sg = 0; + cp->cp_xmit_hdr_off = 0; + cp->cp_xmit_data_off = 0; + cp->cp_xmit_atomic_sent = 0; + cp->cp_xmit_rdma_sent = 0; + cp->cp_xmit_data_sent = 0; + + cp->cp_conn->c_map_queued = 0; + + cp->cp_unacked_packets = rds_sysctl_max_unacked_packets; + cp->cp_unacked_bytes = rds_sysctl_max_unacked_bytes; + + /* Mark messages as retransmissions, and move them to the send q */ + spin_lock_irqsave(&cp->cp_lock, flags); + list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) { + set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); + set_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags); + } + list_splice_init(&cp->cp_retrans, &cp->cp_send_queue); + spin_unlock_irqrestore(&cp->cp_lock, flags); +} +EXPORT_SYMBOL_GPL(rds_send_path_reset); + +static int acquire_in_xmit(struct rds_conn_path *cp) +{ + return test_and_set_bit(RDS_IN_XMIT, &cp->cp_flags) == 0; +} + +static void release_in_xmit(struct rds_conn_path *cp) +{ + clear_bit(RDS_IN_XMIT, &cp->cp_flags); + smp_mb__after_atomic(); + /* + * We don't use wait_on_bit()/wake_up_bit() because our waking is in a + * hot path and finding waiters is very rare. We don't want to walk + * the system-wide hashed waitqueue buckets in the fast path only to + * almost never find waiters. + */ + if (waitqueue_active(&cp->cp_waitq)) + wake_up_all(&cp->cp_waitq); +} + +/* + * We're making the conscious trade-off here to only send one message + * down the connection at a time. + * Pro: + * - tx queueing is a simple fifo list + * - reassembly is optional and easily done by transports per conn + * - no per flow rx lookup at all, straight to the socket + * - less per-frag memory and wire overhead + * Con: + * - queued acks can be delayed behind large messages + * Depends: + * - small message latency is higher behind queued large messages + * - large message latency isn't starved by intervening small sends + */ +int rds_send_xmit(struct rds_conn_path *cp) +{ + struct rds_connection *conn = cp->cp_conn; + struct rds_message *rm; + unsigned long flags; + unsigned int tmp; + struct scatterlist *sg; + int ret = 0; + LIST_HEAD(to_be_dropped); + int batch_count; + unsigned long send_gen = 0; + +restart: + batch_count = 0; + + /* + * sendmsg calls here after having queued its message on the send + * queue. We only have one task feeding the connection at a time. If + * another thread is already feeding the queue then we back off. This + * avoids blocking the caller and trading per-connection data between + * caches per message. + */ + if (!acquire_in_xmit(cp)) { + rds_stats_inc(s_send_lock_contention); + ret = -ENOMEM; + goto out; + } + + if (rds_destroy_pending(cp->cp_conn)) { + release_in_xmit(cp); + ret = -ENETUNREACH; /* dont requeue send work */ + goto out; + } + + /* + * we record the send generation after doing the xmit acquire. + * if someone else manages to jump in and do some work, we'll use + * this to avoid a goto restart farther down. + * + * The acquire_in_xmit() check above ensures that only one + * caller can increment c_send_gen at any time. + */ + send_gen = READ_ONCE(cp->cp_send_gen) + 1; + WRITE_ONCE(cp->cp_send_gen, send_gen); + + /* + * rds_conn_shutdown() sets the conn state and then tests RDS_IN_XMIT, + * we do the opposite to avoid races. + */ + if (!rds_conn_path_up(cp)) { + release_in_xmit(cp); + ret = 0; + goto out; + } + + if (conn->c_trans->xmit_path_prepare) + conn->c_trans->xmit_path_prepare(cp); + + /* + * spin trying to push headers and data down the connection until + * the connection doesn't make forward progress. + */ + while (1) { + + rm = cp->cp_xmit_rm; + + /* + * If between sending messages, we can send a pending congestion + * map update. + */ + if (!rm && test_and_clear_bit(0, &conn->c_map_queued)) { + rm = rds_cong_update_alloc(conn); + if (IS_ERR(rm)) { + ret = PTR_ERR(rm); + break; + } + rm->data.op_active = 1; + rm->m_inc.i_conn_path = cp; + rm->m_inc.i_conn = cp->cp_conn; + + cp->cp_xmit_rm = rm; + } + + /* + * If not already working on one, grab the next message. + * + * cp_xmit_rm holds a ref while we're sending this message down + * the connction. We can use this ref while holding the + * send_sem.. rds_send_reset() is serialized with it. + */ + if (!rm) { + unsigned int len; + + batch_count++; + + /* we want to process as big a batch as we can, but + * we also want to avoid softlockups. If we've been + * through a lot of messages, lets back off and see + * if anyone else jumps in + */ + if (batch_count >= send_batch_count) + goto over_batch; + + spin_lock_irqsave(&cp->cp_lock, flags); + + if (!list_empty(&cp->cp_send_queue)) { + rm = list_entry(cp->cp_send_queue.next, + struct rds_message, + m_conn_item); + rds_message_addref(rm); + + /* + * Move the message from the send queue to the retransmit + * list right away. + */ + list_move_tail(&rm->m_conn_item, + &cp->cp_retrans); + } + + spin_unlock_irqrestore(&cp->cp_lock, flags); + + if (!rm) + break; + + /* Unfortunately, the way Infiniband deals with + * RDMA to a bad MR key is by moving the entire + * queue pair to error state. We cold possibly + * recover from that, but right now we drop the + * connection. + * Therefore, we never retransmit messages with RDMA ops. + */ + if (test_bit(RDS_MSG_FLUSH, &rm->m_flags) || + (rm->rdma.op_active && + test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))) { + spin_lock_irqsave(&cp->cp_lock, flags); + if (test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) + list_move(&rm->m_conn_item, &to_be_dropped); + spin_unlock_irqrestore(&cp->cp_lock, flags); + continue; + } + + /* Require an ACK every once in a while */ + len = ntohl(rm->m_inc.i_hdr.h_len); + if (cp->cp_unacked_packets == 0 || + cp->cp_unacked_bytes < len) { + set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); + + cp->cp_unacked_packets = + rds_sysctl_max_unacked_packets; + cp->cp_unacked_bytes = + rds_sysctl_max_unacked_bytes; + rds_stats_inc(s_send_ack_required); + } else { + cp->cp_unacked_bytes -= len; + cp->cp_unacked_packets--; + } + + cp->cp_xmit_rm = rm; + } + + /* The transport either sends the whole rdma or none of it */ + if (rm->rdma.op_active && !cp->cp_xmit_rdma_sent) { + rm->m_final_op = &rm->rdma; + /* The transport owns the mapped memory for now. + * You can't unmap it while it's on the send queue + */ + set_bit(RDS_MSG_MAPPED, &rm->m_flags); + ret = conn->c_trans->xmit_rdma(conn, &rm->rdma); + if (ret) { + clear_bit(RDS_MSG_MAPPED, &rm->m_flags); + wake_up_interruptible(&rm->m_flush_wait); + break; + } + cp->cp_xmit_rdma_sent = 1; + + } + + if (rm->atomic.op_active && !cp->cp_xmit_atomic_sent) { + rm->m_final_op = &rm->atomic; + /* The transport owns the mapped memory for now. + * You can't unmap it while it's on the send queue + */ + set_bit(RDS_MSG_MAPPED, &rm->m_flags); + ret = conn->c_trans->xmit_atomic(conn, &rm->atomic); + if (ret) { + clear_bit(RDS_MSG_MAPPED, &rm->m_flags); + wake_up_interruptible(&rm->m_flush_wait); + break; + } + cp->cp_xmit_atomic_sent = 1; + + } + + /* + * A number of cases require an RDS header to be sent + * even if there is no data. + * We permit 0-byte sends; rds-ping depends on this. + * However, if there are exclusively attached silent ops, + * we skip the hdr/data send, to enable silent operation. + */ + if (rm->data.op_nents == 0) { + int ops_present; + int all_ops_are_silent = 1; + + ops_present = (rm->atomic.op_active || rm->rdma.op_active); + if (rm->atomic.op_active && !rm->atomic.op_silent) + all_ops_are_silent = 0; + if (rm->rdma.op_active && !rm->rdma.op_silent) + all_ops_are_silent = 0; + + if (ops_present && all_ops_are_silent + && !rm->m_rdma_cookie) + rm->data.op_active = 0; + } + + if (rm->data.op_active && !cp->cp_xmit_data_sent) { + rm->m_final_op = &rm->data; + + ret = conn->c_trans->xmit(conn, rm, + cp->cp_xmit_hdr_off, + cp->cp_xmit_sg, + cp->cp_xmit_data_off); + if (ret <= 0) + break; + + if (cp->cp_xmit_hdr_off < sizeof(struct rds_header)) { + tmp = min_t(int, ret, + sizeof(struct rds_header) - + cp->cp_xmit_hdr_off); + cp->cp_xmit_hdr_off += tmp; + ret -= tmp; + } + + sg = &rm->data.op_sg[cp->cp_xmit_sg]; + while (ret) { + tmp = min_t(int, ret, sg->length - + cp->cp_xmit_data_off); + cp->cp_xmit_data_off += tmp; + ret -= tmp; + if (cp->cp_xmit_data_off == sg->length) { + cp->cp_xmit_data_off = 0; + sg++; + cp->cp_xmit_sg++; + BUG_ON(ret != 0 && cp->cp_xmit_sg == + rm->data.op_nents); + } + } + + if (cp->cp_xmit_hdr_off == sizeof(struct rds_header) && + (cp->cp_xmit_sg == rm->data.op_nents)) + cp->cp_xmit_data_sent = 1; + } + + /* + * A rm will only take multiple times through this loop + * if there is a data op. Thus, if the data is sent (or there was + * none), then we're done with the rm. + */ + if (!rm->data.op_active || cp->cp_xmit_data_sent) { + cp->cp_xmit_rm = NULL; + cp->cp_xmit_sg = 0; + cp->cp_xmit_hdr_off = 0; + cp->cp_xmit_data_off = 0; + cp->cp_xmit_rdma_sent = 0; + cp->cp_xmit_atomic_sent = 0; + cp->cp_xmit_data_sent = 0; + + rds_message_put(rm); + } + } + +over_batch: + if (conn->c_trans->xmit_path_complete) + conn->c_trans->xmit_path_complete(cp); + release_in_xmit(cp); + + /* Nuke any messages we decided not to retransmit. */ + if (!list_empty(&to_be_dropped)) { + /* irqs on here, so we can put(), unlike above */ + list_for_each_entry(rm, &to_be_dropped, m_conn_item) + rds_message_put(rm); + rds_send_remove_from_sock(&to_be_dropped, RDS_RDMA_DROPPED); + } + + /* + * Other senders can queue a message after we last test the send queue + * but before we clear RDS_IN_XMIT. In that case they'd back off and + * not try and send their newly queued message. We need to check the + * send queue after having cleared RDS_IN_XMIT so that their message + * doesn't get stuck on the send queue. + * + * If the transport cannot continue (i.e ret != 0), then it must + * call us when more room is available, such as from the tx + * completion handler. + * + * We have an extra generation check here so that if someone manages + * to jump in after our release_in_xmit, we'll see that they have done + * some work and we will skip our goto + */ + if (ret == 0) { + bool raced; + + smp_mb(); + raced = send_gen != READ_ONCE(cp->cp_send_gen); + + if ((test_bit(0, &conn->c_map_queued) || + !list_empty(&cp->cp_send_queue)) && !raced) { + if (batch_count < send_batch_count) + goto restart; + rcu_read_lock(); + if (rds_destroy_pending(cp->cp_conn)) + ret = -ENETUNREACH; + else + queue_delayed_work(rds_wq, &cp->cp_send_w, 1); + rcu_read_unlock(); + } else if (raced) { + rds_stats_inc(s_send_lock_queue_raced); + } + } +out: + return ret; +} +EXPORT_SYMBOL_GPL(rds_send_xmit); + +static void rds_send_sndbuf_remove(struct rds_sock *rs, struct rds_message *rm) +{ + u32 len = be32_to_cpu(rm->m_inc.i_hdr.h_len); + + assert_spin_locked(&rs->rs_lock); + + BUG_ON(rs->rs_snd_bytes < len); + rs->rs_snd_bytes -= len; + + if (rs->rs_snd_bytes == 0) + rds_stats_inc(s_send_queue_empty); +} + +static inline int rds_send_is_acked(struct rds_message *rm, u64 ack, + is_acked_func is_acked) +{ + if (is_acked) + return is_acked(rm, ack); + return be64_to_cpu(rm->m_inc.i_hdr.h_sequence) <= ack; +} + +/* + * This is pretty similar to what happens below in the ACK + * handling code - except that we call here as soon as we get + * the IB send completion on the RDMA op and the accompanying + * message. + */ +void rds_rdma_send_complete(struct rds_message *rm, int status) +{ + struct rds_sock *rs = NULL; + struct rm_rdma_op *ro; + struct rds_notifier *notifier; + unsigned long flags; + unsigned int notify = 0; + + spin_lock_irqsave(&rm->m_rs_lock, flags); + + notify = rm->rdma.op_notify | rm->data.op_notify; + ro = &rm->rdma; + if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) && + ro->op_active && notify && ro->op_notifier) { + notifier = ro->op_notifier; + rs = rm->m_rs; + sock_hold(rds_rs_to_sk(rs)); + + notifier->n_status = status; + spin_lock(&rs->rs_lock); + list_add_tail(¬ifier->n_list, &rs->rs_notify_queue); + spin_unlock(&rs->rs_lock); + + ro->op_notifier = NULL; + } + + spin_unlock_irqrestore(&rm->m_rs_lock, flags); + + if (rs) { + rds_wake_sk_sleep(rs); + sock_put(rds_rs_to_sk(rs)); + } +} +EXPORT_SYMBOL_GPL(rds_rdma_send_complete); + +/* + * Just like above, except looks at atomic op + */ +void rds_atomic_send_complete(struct rds_message *rm, int status) +{ + struct rds_sock *rs = NULL; + struct rm_atomic_op *ao; + struct rds_notifier *notifier; + unsigned long flags; + + spin_lock_irqsave(&rm->m_rs_lock, flags); + + ao = &rm->atomic; + if (test_bit(RDS_MSG_ON_SOCK, &rm->m_flags) + && ao->op_active && ao->op_notify && ao->op_notifier) { + notifier = ao->op_notifier; + rs = rm->m_rs; + sock_hold(rds_rs_to_sk(rs)); + + notifier->n_status = status; + spin_lock(&rs->rs_lock); + list_add_tail(¬ifier->n_list, &rs->rs_notify_queue); + spin_unlock(&rs->rs_lock); + + ao->op_notifier = NULL; + } + + spin_unlock_irqrestore(&rm->m_rs_lock, flags); + + if (rs) { + rds_wake_sk_sleep(rs); + sock_put(rds_rs_to_sk(rs)); + } +} +EXPORT_SYMBOL_GPL(rds_atomic_send_complete); + +/* + * This is the same as rds_rdma_send_complete except we + * don't do any locking - we have all the ingredients (message, + * socket, socket lock) and can just move the notifier. + */ +static inline void +__rds_send_complete(struct rds_sock *rs, struct rds_message *rm, int status) +{ + struct rm_rdma_op *ro; + struct rm_atomic_op *ao; + + ro = &rm->rdma; + if (ro->op_active && ro->op_notify && ro->op_notifier) { + ro->op_notifier->n_status = status; + list_add_tail(&ro->op_notifier->n_list, &rs->rs_notify_queue); + ro->op_notifier = NULL; + } + + ao = &rm->atomic; + if (ao->op_active && ao->op_notify && ao->op_notifier) { + ao->op_notifier->n_status = status; + list_add_tail(&ao->op_notifier->n_list, &rs->rs_notify_queue); + ao->op_notifier = NULL; + } + + /* No need to wake the app - caller does this */ +} + +/* + * This removes messages from the socket's list if they're on it. The list + * argument must be private to the caller, we must be able to modify it + * without locks. The messages must have a reference held for their + * position on the list. This function will drop that reference after + * removing the messages from the 'messages' list regardless of if it found + * the messages on the socket list or not. + */ +static void rds_send_remove_from_sock(struct list_head *messages, int status) +{ + unsigned long flags; + struct rds_sock *rs = NULL; + struct rds_message *rm; + + while (!list_empty(messages)) { + int was_on_sock = 0; + + rm = list_entry(messages->next, struct rds_message, + m_conn_item); + list_del_init(&rm->m_conn_item); + + /* + * If we see this flag cleared then we're *sure* that someone + * else beat us to removing it from the sock. If we race + * with their flag update we'll get the lock and then really + * see that the flag has been cleared. + * + * The message spinlock makes sure nobody clears rm->m_rs + * while we're messing with it. It does not prevent the + * message from being removed from the socket, though. + */ + spin_lock_irqsave(&rm->m_rs_lock, flags); + if (!test_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) + goto unlock_and_drop; + + if (rs != rm->m_rs) { + if (rs) { + rds_wake_sk_sleep(rs); + sock_put(rds_rs_to_sk(rs)); + } + rs = rm->m_rs; + if (rs) + sock_hold(rds_rs_to_sk(rs)); + } + if (!rs) + goto unlock_and_drop; + spin_lock(&rs->rs_lock); + + if (test_and_clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags)) { + struct rm_rdma_op *ro = &rm->rdma; + struct rds_notifier *notifier; + + list_del_init(&rm->m_sock_item); + rds_send_sndbuf_remove(rs, rm); + + if (ro->op_active && ro->op_notifier && + (ro->op_notify || (ro->op_recverr && status))) { + notifier = ro->op_notifier; + list_add_tail(¬ifier->n_list, + &rs->rs_notify_queue); + if (!notifier->n_status) + notifier->n_status = status; + rm->rdma.op_notifier = NULL; + } + was_on_sock = 1; + } + spin_unlock(&rs->rs_lock); + +unlock_and_drop: + spin_unlock_irqrestore(&rm->m_rs_lock, flags); + rds_message_put(rm); + if (was_on_sock) + rds_message_put(rm); + } + + if (rs) { + rds_wake_sk_sleep(rs); + sock_put(rds_rs_to_sk(rs)); + } +} + +/* + * Transports call here when they've determined that the receiver queued + * messages up to, and including, the given sequence number. Messages are + * moved to the retrans queue when rds_send_xmit picks them off the send + * queue. This means that in the TCP case, the message may not have been + * assigned the m_ack_seq yet - but that's fine as long as tcp_is_acked + * checks the RDS_MSG_HAS_ACK_SEQ bit. + */ +void rds_send_path_drop_acked(struct rds_conn_path *cp, u64 ack, + is_acked_func is_acked) +{ + struct rds_message *rm, *tmp; + unsigned long flags; + LIST_HEAD(list); + + spin_lock_irqsave(&cp->cp_lock, flags); + + list_for_each_entry_safe(rm, tmp, &cp->cp_retrans, m_conn_item) { + if (!rds_send_is_acked(rm, ack, is_acked)) + break; + + list_move(&rm->m_conn_item, &list); + clear_bit(RDS_MSG_ON_CONN, &rm->m_flags); + } + + /* order flag updates with spin locks */ + if (!list_empty(&list)) + smp_mb__after_atomic(); + + spin_unlock_irqrestore(&cp->cp_lock, flags); + + /* now remove the messages from the sock list as needed */ + rds_send_remove_from_sock(&list, RDS_RDMA_SUCCESS); +} +EXPORT_SYMBOL_GPL(rds_send_path_drop_acked); + +void rds_send_drop_acked(struct rds_connection *conn, u64 ack, + is_acked_func is_acked) +{ + WARN_ON(conn->c_trans->t_mp_capable); + rds_send_path_drop_acked(&conn->c_path[0], ack, is_acked); +} +EXPORT_SYMBOL_GPL(rds_send_drop_acked); + +void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in6 *dest) +{ + struct rds_message *rm, *tmp; + struct rds_connection *conn; + struct rds_conn_path *cp; + unsigned long flags; + LIST_HEAD(list); + + /* get all the messages we're dropping under the rs lock */ + spin_lock_irqsave(&rs->rs_lock, flags); + + list_for_each_entry_safe(rm, tmp, &rs->rs_send_queue, m_sock_item) { + if (dest && + (!ipv6_addr_equal(&dest->sin6_addr, &rm->m_daddr) || + dest->sin6_port != rm->m_inc.i_hdr.h_dport)) + continue; + + list_move(&rm->m_sock_item, &list); + rds_send_sndbuf_remove(rs, rm); + clear_bit(RDS_MSG_ON_SOCK, &rm->m_flags); + } + + /* order flag updates with the rs lock */ + smp_mb__after_atomic(); + + spin_unlock_irqrestore(&rs->rs_lock, flags); + + if (list_empty(&list)) + return; + + /* Remove the messages from the conn */ + list_for_each_entry(rm, &list, m_sock_item) { + + conn = rm->m_inc.i_conn; + if (conn->c_trans->t_mp_capable) + cp = rm->m_inc.i_conn_path; + else + cp = &conn->c_path[0]; + + spin_lock_irqsave(&cp->cp_lock, flags); + /* + * Maybe someone else beat us to removing rm from the conn. + * If we race with their flag update we'll get the lock and + * then really see that the flag has been cleared. + */ + if (!test_and_clear_bit(RDS_MSG_ON_CONN, &rm->m_flags)) { + spin_unlock_irqrestore(&cp->cp_lock, flags); + continue; + } + list_del_init(&rm->m_conn_item); + spin_unlock_irqrestore(&cp->cp_lock, flags); + + /* + * Couldn't grab m_rs_lock in top loop (lock ordering), + * but we can now. + */ + spin_lock_irqsave(&rm->m_rs_lock, flags); + + spin_lock(&rs->rs_lock); + __rds_send_complete(rs, rm, RDS_RDMA_CANCELED); + spin_unlock(&rs->rs_lock); + + spin_unlock_irqrestore(&rm->m_rs_lock, flags); + + rds_message_put(rm); + } + + rds_wake_sk_sleep(rs); + + while (!list_empty(&list)) { + rm = list_entry(list.next, struct rds_message, m_sock_item); + list_del_init(&rm->m_sock_item); + rds_message_wait(rm); + + /* just in case the code above skipped this message + * because RDS_MSG_ON_CONN wasn't set, run it again here + * taking m_rs_lock is the only thing that keeps us + * from racing with ack processing. + */ + spin_lock_irqsave(&rm->m_rs_lock, flags); + + spin_lock(&rs->rs_lock); + __rds_send_complete(rs, rm, RDS_RDMA_CANCELED); + spin_unlock(&rs->rs_lock); + + spin_unlock_irqrestore(&rm->m_rs_lock, flags); + + rds_message_put(rm); + } +} + +/* + * we only want this to fire once so we use the callers 'queued'. It's + * possible that another thread can race with us and remove the + * message from the flow with RDS_CANCEL_SENT_TO. + */ +static int rds_send_queue_rm(struct rds_sock *rs, struct rds_connection *conn, + struct rds_conn_path *cp, + struct rds_message *rm, __be16 sport, + __be16 dport, int *queued) +{ + unsigned long flags; + u32 len; + + if (*queued) + goto out; + + len = be32_to_cpu(rm->m_inc.i_hdr.h_len); + + /* this is the only place which holds both the socket's rs_lock + * and the connection's c_lock */ + spin_lock_irqsave(&rs->rs_lock, flags); + + /* + * If there is a little space in sndbuf, we don't queue anything, + * and userspace gets -EAGAIN. But poll() indicates there's send + * room. This can lead to bad behavior (spinning) if snd_bytes isn't + * freed up by incoming acks. So we check the *old* value of + * rs_snd_bytes here to allow the last msg to exceed the buffer, + * and poll() now knows no more data can be sent. + */ + if (rs->rs_snd_bytes < rds_sk_sndbuf(rs)) { + rs->rs_snd_bytes += len; + + /* let recv side know we are close to send space exhaustion. + * This is probably not the optimal way to do it, as this + * means we set the flag on *all* messages as soon as our + * throughput hits a certain threshold. + */ + if (rs->rs_snd_bytes >= rds_sk_sndbuf(rs) / 2) + set_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags); + + list_add_tail(&rm->m_sock_item, &rs->rs_send_queue); + set_bit(RDS_MSG_ON_SOCK, &rm->m_flags); + rds_message_addref(rm); + sock_hold(rds_rs_to_sk(rs)); + rm->m_rs = rs; + + /* The code ordering is a little weird, but we're + trying to minimize the time we hold c_lock */ + rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, 0); + rm->m_inc.i_conn = conn; + rm->m_inc.i_conn_path = cp; + rds_message_addref(rm); + + spin_lock(&cp->cp_lock); + rm->m_inc.i_hdr.h_sequence = cpu_to_be64(cp->cp_next_tx_seq++); + list_add_tail(&rm->m_conn_item, &cp->cp_send_queue); + set_bit(RDS_MSG_ON_CONN, &rm->m_flags); + spin_unlock(&cp->cp_lock); + + rdsdebug("queued msg %p len %d, rs %p bytes %d seq %llu\n", + rm, len, rs, rs->rs_snd_bytes, + (unsigned long long)be64_to_cpu(rm->m_inc.i_hdr.h_sequence)); + + *queued = 1; + } + + spin_unlock_irqrestore(&rs->rs_lock, flags); +out: + return *queued; +} + +/* + * rds_message is getting to be quite complicated, and we'd like to allocate + * it all in one go. This figures out how big it needs to be up front. + */ +static int rds_rm_size(struct msghdr *msg, int num_sgs, + struct rds_iov_vector_arr *vct) +{ + struct cmsghdr *cmsg; + int size = 0; + int cmsg_groups = 0; + int retval; + bool zcopy_cookie = false; + struct rds_iov_vector *iov, *tmp_iov; + + for_each_cmsghdr(cmsg, msg) { + if (!CMSG_OK(msg, cmsg)) + return -EINVAL; + + if (cmsg->cmsg_level != SOL_RDS) + continue; + + switch (cmsg->cmsg_type) { + case RDS_CMSG_RDMA_ARGS: + if (vct->indx >= vct->len) { + vct->len += vct->incr; + tmp_iov = + krealloc(vct->vec, + vct->len * + sizeof(struct rds_iov_vector), + GFP_KERNEL); + if (!tmp_iov) { + vct->len -= vct->incr; + return -ENOMEM; + } + vct->vec = tmp_iov; + } + iov = &vct->vec[vct->indx]; + memset(iov, 0, sizeof(struct rds_iov_vector)); + vct->indx++; + cmsg_groups |= 1; + retval = rds_rdma_extra_size(CMSG_DATA(cmsg), iov); + if (retval < 0) + return retval; + size += retval; + + break; + + case RDS_CMSG_ZCOPY_COOKIE: + zcopy_cookie = true; + /* fall through */ + + case RDS_CMSG_RDMA_DEST: + case RDS_CMSG_RDMA_MAP: + cmsg_groups |= 2; + /* these are valid but do no add any size */ + break; + + case RDS_CMSG_ATOMIC_CSWP: + case RDS_CMSG_ATOMIC_FADD: + case RDS_CMSG_MASKED_ATOMIC_CSWP: + case RDS_CMSG_MASKED_ATOMIC_FADD: + cmsg_groups |= 1; + size += sizeof(struct scatterlist); + break; + + default: + return -EINVAL; + } + + } + + if ((msg->msg_flags & MSG_ZEROCOPY) && !zcopy_cookie) + return -EINVAL; + + size += num_sgs * sizeof(struct scatterlist); + + /* Ensure (DEST, MAP) are never used with (ARGS, ATOMIC) */ + if (cmsg_groups == 3) + return -EINVAL; + + return size; +} + +static int rds_cmsg_zcopy(struct rds_sock *rs, struct rds_message *rm, + struct cmsghdr *cmsg) +{ + u32 *cookie; + + if (cmsg->cmsg_len < CMSG_LEN(sizeof(*cookie)) || + !rm->data.op_mmp_znotifier) + return -EINVAL; + cookie = CMSG_DATA(cmsg); + rm->data.op_mmp_znotifier->z_cookie = *cookie; + return 0; +} + +static int rds_cmsg_send(struct rds_sock *rs, struct rds_message *rm, + struct msghdr *msg, int *allocated_mr, + struct rds_iov_vector_arr *vct) +{ + struct cmsghdr *cmsg; + int ret = 0, ind = 0; + + for_each_cmsghdr(cmsg, msg) { + if (!CMSG_OK(msg, cmsg)) + return -EINVAL; + + if (cmsg->cmsg_level != SOL_RDS) + continue; + + /* As a side effect, RDMA_DEST and RDMA_MAP will set + * rm->rdma.m_rdma_cookie and rm->rdma.m_rdma_mr. + */ + switch (cmsg->cmsg_type) { + case RDS_CMSG_RDMA_ARGS: + if (ind >= vct->indx) + return -ENOMEM; + ret = rds_cmsg_rdma_args(rs, rm, cmsg, &vct->vec[ind]); + ind++; + break; + + case RDS_CMSG_RDMA_DEST: + ret = rds_cmsg_rdma_dest(rs, rm, cmsg); + break; + + case RDS_CMSG_RDMA_MAP: + ret = rds_cmsg_rdma_map(rs, rm, cmsg); + if (!ret) + *allocated_mr = 1; + else if (ret == -ENODEV) + /* Accommodate the get_mr() case which can fail + * if connection isn't established yet. + */ + ret = -EAGAIN; + break; + case RDS_CMSG_ATOMIC_CSWP: + case RDS_CMSG_ATOMIC_FADD: + case RDS_CMSG_MASKED_ATOMIC_CSWP: + case RDS_CMSG_MASKED_ATOMIC_FADD: + ret = rds_cmsg_atomic(rs, rm, cmsg); + break; + + case RDS_CMSG_ZCOPY_COOKIE: + ret = rds_cmsg_zcopy(rs, rm, cmsg); + break; + + default: + return -EINVAL; + } + + if (ret) + break; + } + + return ret; +} + +static int rds_send_mprds_hash(struct rds_sock *rs, + struct rds_connection *conn, int nonblock) +{ + int hash; + + if (conn->c_npaths == 0) + hash = RDS_MPATH_HASH(rs, RDS_MPATH_WORKERS); + else + hash = RDS_MPATH_HASH(rs, conn->c_npaths); + if (conn->c_npaths == 0 && hash != 0) { + rds_send_ping(conn, 0); + + /* The underlying connection is not up yet. Need to wait + * until it is up to be sure that the non-zero c_path can be + * used. But if we are interrupted, we have to use the zero + * c_path in case the connection ends up being non-MP capable. + */ + if (conn->c_npaths == 0) { + /* Cannot wait for the connection be made, so just use + * the base c_path. + */ + if (nonblock) + return 0; + if (wait_event_interruptible(conn->c_hs_waitq, + conn->c_npaths != 0)) + hash = 0; + } + if (conn->c_npaths == 1) + hash = 0; + } + return hash; +} + +static int rds_rdma_bytes(struct msghdr *msg, size_t *rdma_bytes) +{ + struct rds_rdma_args *args; + struct cmsghdr *cmsg; + + for_each_cmsghdr(cmsg, msg) { + if (!CMSG_OK(msg, cmsg)) + return -EINVAL; + + if (cmsg->cmsg_level != SOL_RDS) + continue; + + if (cmsg->cmsg_type == RDS_CMSG_RDMA_ARGS) { + if (cmsg->cmsg_len < + CMSG_LEN(sizeof(struct rds_rdma_args))) + return -EINVAL; + args = CMSG_DATA(cmsg); + *rdma_bytes += args->remote_vec.bytes; + } + } + return 0; +} + +int rds_sendmsg(struct socket *sock, struct msghdr *msg, size_t payload_len) +{ + struct sock *sk = sock->sk; + struct rds_sock *rs = rds_sk_to_rs(sk); + DECLARE_SOCKADDR(struct sockaddr_in6 *, sin6, msg->msg_name); + DECLARE_SOCKADDR(struct sockaddr_in *, usin, msg->msg_name); + __be16 dport; + struct rds_message *rm = NULL; + struct rds_connection *conn; + int ret = 0; + int queued = 0, allocated_mr = 0; + int nonblock = msg->msg_flags & MSG_DONTWAIT; + long timeo = sock_sndtimeo(sk, nonblock); + struct rds_conn_path *cpath; + struct in6_addr daddr; + __u32 scope_id = 0; + size_t total_payload_len = payload_len, rdma_payload_len = 0; + bool zcopy = ((msg->msg_flags & MSG_ZEROCOPY) && + sock_flag(rds_rs_to_sk(rs), SOCK_ZEROCOPY)); + int num_sgs = ceil(payload_len, PAGE_SIZE); + int namelen; + struct rds_iov_vector_arr vct; + int ind; + + memset(&vct, 0, sizeof(vct)); + + /* expect 1 RDMA CMSG per rds_sendmsg. can still grow if more needed. */ + vct.incr = 1; + + /* Mirror Linux UDP mirror of BSD error message compatibility */ + /* XXX: Perhaps MSG_MORE someday */ + if (msg->msg_flags & ~(MSG_DONTWAIT | MSG_CMSG_COMPAT | MSG_ZEROCOPY)) { + ret = -EOPNOTSUPP; + goto out; + } + + namelen = msg->msg_namelen; + if (namelen != 0) { + if (namelen < sizeof(*usin)) { + ret = -EINVAL; + goto out; + } + switch (usin->sin_family) { + case AF_INET: + if (usin->sin_addr.s_addr == htonl(INADDR_ANY) || + usin->sin_addr.s_addr == htonl(INADDR_BROADCAST) || + IN_MULTICAST(ntohl(usin->sin_addr.s_addr))) { + ret = -EINVAL; + goto out; + } + ipv6_addr_set_v4mapped(usin->sin_addr.s_addr, &daddr); + dport = usin->sin_port; + break; + +#if IS_ENABLED(CONFIG_IPV6) + case AF_INET6: { + int addr_type; + + if (namelen < sizeof(*sin6)) { + ret = -EINVAL; + goto out; + } + addr_type = ipv6_addr_type(&sin6->sin6_addr); + if (!(addr_type & IPV6_ADDR_UNICAST)) { + __be32 addr4; + + if (!(addr_type & IPV6_ADDR_MAPPED)) { + ret = -EINVAL; + goto out; + } + + /* It is a mapped address. Need to do some + * sanity checks. + */ + addr4 = sin6->sin6_addr.s6_addr32[3]; + if (addr4 == htonl(INADDR_ANY) || + addr4 == htonl(INADDR_BROADCAST) || + IN_MULTICAST(ntohl(addr4))) { + ret = -EINVAL; + goto out; + } + } + if (addr_type & IPV6_ADDR_LINKLOCAL) { + if (sin6->sin6_scope_id == 0) { + ret = -EINVAL; + goto out; + } + scope_id = sin6->sin6_scope_id; + } + + daddr = sin6->sin6_addr; + dport = sin6->sin6_port; + break; + } +#endif + + default: + ret = -EINVAL; + goto out; + } + } else { + /* We only care about consistency with ->connect() */ + lock_sock(sk); + daddr = rs->rs_conn_addr; + dport = rs->rs_conn_port; + scope_id = rs->rs_bound_scope_id; + release_sock(sk); + } + + lock_sock(sk); + if (ipv6_addr_any(&rs->rs_bound_addr) || ipv6_addr_any(&daddr)) { + release_sock(sk); + ret = -ENOTCONN; + goto out; + } else if (namelen != 0) { + /* Cannot send to an IPv4 address using an IPv6 source + * address and cannot send to an IPv6 address using an + * IPv4 source address. + */ + if (ipv6_addr_v4mapped(&daddr) ^ + ipv6_addr_v4mapped(&rs->rs_bound_addr)) { + release_sock(sk); + ret = -EOPNOTSUPP; + goto out; + } + /* If the socket is already bound to a link local address, + * it can only send to peers on the same link. But allow + * communicating beween link local and non-link local address. + */ + if (scope_id != rs->rs_bound_scope_id) { + if (!scope_id) { + scope_id = rs->rs_bound_scope_id; + } else if (rs->rs_bound_scope_id) { + release_sock(sk); + ret = -EINVAL; + goto out; + } + } + } + release_sock(sk); + + ret = rds_rdma_bytes(msg, &rdma_payload_len); + if (ret) + goto out; + + total_payload_len += rdma_payload_len; + if (max_t(size_t, payload_len, rdma_payload_len) > RDS_MAX_MSG_SIZE) { + ret = -EMSGSIZE; + goto out; + } + + if (payload_len > rds_sk_sndbuf(rs)) { + ret = -EMSGSIZE; + goto out; + } + + if (zcopy) { + if (rs->rs_transport->t_type != RDS_TRANS_TCP) { + ret = -EOPNOTSUPP; + goto out; + } + num_sgs = iov_iter_npages(&msg->msg_iter, INT_MAX); + } + /* size of rm including all sgs */ + ret = rds_rm_size(msg, num_sgs, &vct); + if (ret < 0) + goto out; + + rm = rds_message_alloc(ret, GFP_KERNEL); + if (!rm) { + ret = -ENOMEM; + goto out; + } + + /* Attach data to the rm */ + if (payload_len) { + rm->data.op_sg = rds_message_alloc_sgs(rm, num_sgs); + if (!rm->data.op_sg) { + ret = -ENOMEM; + goto out; + } + ret = rds_message_copy_from_user(rm, &msg->msg_iter, zcopy); + if (ret) + goto out; + } + rm->data.op_active = 1; + + rm->m_daddr = daddr; + + /* rds_conn_create has a spinlock that runs with IRQ off. + * Caching the conn in the socket helps a lot. */ + if (rs->rs_conn && ipv6_addr_equal(&rs->rs_conn->c_faddr, &daddr)) + conn = rs->rs_conn; + else { + conn = rds_conn_create_outgoing(sock_net(sock->sk), + &rs->rs_bound_addr, &daddr, + rs->rs_transport, + sock->sk->sk_allocation, + scope_id); + if (IS_ERR(conn)) { + ret = PTR_ERR(conn); + goto out; + } + rs->rs_conn = conn; + } + + if (conn->c_trans->t_mp_capable) + cpath = &conn->c_path[rds_send_mprds_hash(rs, conn, nonblock)]; + else + cpath = &conn->c_path[0]; + + rm->m_conn_path = cpath; + + /* Parse any control messages the user may have included. */ + ret = rds_cmsg_send(rs, rm, msg, &allocated_mr, &vct); + if (ret) { + /* Trigger connection so that its ready for the next retry */ + if (ret == -EAGAIN) + rds_conn_connect_if_down(conn); + goto out; + } + + if (rm->rdma.op_active && !conn->c_trans->xmit_rdma) { + printk_ratelimited(KERN_NOTICE "rdma_op %p conn xmit_rdma %p\n", + &rm->rdma, conn->c_trans->xmit_rdma); + ret = -EOPNOTSUPP; + goto out; + } + + if (rm->atomic.op_active && !conn->c_trans->xmit_atomic) { + printk_ratelimited(KERN_NOTICE "atomic_op %p conn xmit_atomic %p\n", + &rm->atomic, conn->c_trans->xmit_atomic); + ret = -EOPNOTSUPP; + goto out; + } + + if (rds_destroy_pending(conn)) { + ret = -EAGAIN; + goto out; + } + + rds_conn_path_connect_if_down(cpath); + + ret = rds_cong_wait(conn->c_fcong, dport, nonblock, rs); + if (ret) { + rs->rs_seen_congestion = 1; + goto out; + } + while (!rds_send_queue_rm(rs, conn, cpath, rm, rs->rs_bound_port, + dport, &queued)) { + rds_stats_inc(s_send_queue_full); + + if (nonblock) { + ret = -EAGAIN; + goto out; + } + + timeo = wait_event_interruptible_timeout(*sk_sleep(sk), + rds_send_queue_rm(rs, conn, cpath, rm, + rs->rs_bound_port, + dport, + &queued), + timeo); + rdsdebug("sendmsg woke queued %d timeo %ld\n", queued, timeo); + if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT) + continue; + + ret = timeo; + if (ret == 0) + ret = -ETIMEDOUT; + goto out; + } + + /* + * By now we've committed to the send. We reuse rds_send_worker() + * to retry sends in the rds thread if the transport asks us to. + */ + rds_stats_inc(s_send_queued); + + ret = rds_send_xmit(cpath); + if (ret == -ENOMEM || ret == -EAGAIN) { + ret = 0; + rcu_read_lock(); + if (rds_destroy_pending(cpath->cp_conn)) + ret = -ENETUNREACH; + else + queue_delayed_work(rds_wq, &cpath->cp_send_w, 1); + rcu_read_unlock(); + } + if (ret) + goto out; + rds_message_put(rm); + + for (ind = 0; ind < vct.indx; ind++) + kfree(vct.vec[ind].iov); + kfree(vct.vec); + + return payload_len; + +out: + for (ind = 0; ind < vct.indx; ind++) + kfree(vct.vec[ind].iov); + kfree(vct.vec); + + /* If the user included a RDMA_MAP cmsg, we allocated a MR on the fly. + * If the sendmsg goes through, we keep the MR. If it fails with EAGAIN + * or in any other way, we need to destroy the MR again */ + if (allocated_mr) + rds_rdma_unuse(rs, rds_rdma_cookie_key(rm->m_rdma_cookie), 1); + + if (rm) + rds_message_put(rm); + return ret; +} + +/* + * send out a probe. Can be shared by rds_send_ping, + * rds_send_pong, rds_send_hb. + * rds_send_hb should use h_flags + * RDS_FLAG_HB_PING|RDS_FLAG_ACK_REQUIRED + * or + * RDS_FLAG_HB_PONG|RDS_FLAG_ACK_REQUIRED + */ +static int +rds_send_probe(struct rds_conn_path *cp, __be16 sport, + __be16 dport, u8 h_flags) +{ + struct rds_message *rm; + unsigned long flags; + int ret = 0; + + rm = rds_message_alloc(0, GFP_ATOMIC); + if (!rm) { + ret = -ENOMEM; + goto out; + } + + rm->m_daddr = cp->cp_conn->c_faddr; + rm->data.op_active = 1; + + rds_conn_path_connect_if_down(cp); + + ret = rds_cong_wait(cp->cp_conn->c_fcong, dport, 1, NULL); + if (ret) + goto out; + + spin_lock_irqsave(&cp->cp_lock, flags); + list_add_tail(&rm->m_conn_item, &cp->cp_send_queue); + set_bit(RDS_MSG_ON_CONN, &rm->m_flags); + rds_message_addref(rm); + rm->m_inc.i_conn = cp->cp_conn; + rm->m_inc.i_conn_path = cp; + + rds_message_populate_header(&rm->m_inc.i_hdr, sport, dport, + cp->cp_next_tx_seq); + rm->m_inc.i_hdr.h_flags |= h_flags; + cp->cp_next_tx_seq++; + + if (RDS_HS_PROBE(be16_to_cpu(sport), be16_to_cpu(dport)) && + cp->cp_conn->c_trans->t_mp_capable) { + u16 npaths = cpu_to_be16(RDS_MPATH_WORKERS); + u32 my_gen_num = cpu_to_be32(cp->cp_conn->c_my_gen_num); + + rds_message_add_extension(&rm->m_inc.i_hdr, + RDS_EXTHDR_NPATHS, &npaths, + sizeof(npaths)); + rds_message_add_extension(&rm->m_inc.i_hdr, + RDS_EXTHDR_GEN_NUM, + &my_gen_num, + sizeof(u32)); + } + spin_unlock_irqrestore(&cp->cp_lock, flags); + + rds_stats_inc(s_send_queued); + rds_stats_inc(s_send_pong); + + /* schedule the send work on rds_wq */ + rcu_read_lock(); + if (!rds_destroy_pending(cp->cp_conn)) + queue_delayed_work(rds_wq, &cp->cp_send_w, 1); + rcu_read_unlock(); + + rds_message_put(rm); + return 0; + +out: + if (rm) + rds_message_put(rm); + return ret; +} + +int +rds_send_pong(struct rds_conn_path *cp, __be16 dport) +{ + return rds_send_probe(cp, 0, dport, 0); +} + +void +rds_send_ping(struct rds_connection *conn, int cp_index) +{ + unsigned long flags; + struct rds_conn_path *cp = &conn->c_path[cp_index]; + + spin_lock_irqsave(&cp->cp_lock, flags); + if (conn->c_ping_triggered) { + spin_unlock_irqrestore(&cp->cp_lock, flags); + return; + } + conn->c_ping_triggered = 1; + spin_unlock_irqrestore(&cp->cp_lock, flags); + rds_send_probe(cp, cpu_to_be16(RDS_FLAG_PROBE_PORT), 0, 0); +} +EXPORT_SYMBOL_GPL(rds_send_ping); diff --git a/net/rds/stats.c b/net/rds/stats.c new file mode 100644 index 000000000..6bbab4d74 --- /dev/null +++ b/net/rds/stats.c @@ -0,0 +1,154 @@ +/* + * Copyright (c) 2006 Oracle. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/percpu.h> +#include <linux/seq_file.h> +#include <linux/proc_fs.h> +#include <linux/export.h> + +#include "rds.h" + +DEFINE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats); +EXPORT_PER_CPU_SYMBOL_GPL(rds_stats); + +/* :.,$s/unsigned long\>.*\<s_\(.*\);/"\1",/g */ + +static const char *const rds_stat_names[] = { + "conn_reset", + "recv_drop_bad_checksum", + "recv_drop_old_seq", + "recv_drop_no_sock", + "recv_drop_dead_sock", + "recv_deliver_raced", + "recv_delivered", + "recv_queued", + "recv_immediate_retry", + "recv_delayed_retry", + "recv_ack_required", + "recv_rdma_bytes", + "recv_ping", + "send_queue_empty", + "send_queue_full", + "send_lock_contention", + "send_lock_queue_raced", + "send_immediate_retry", + "send_delayed_retry", + "send_drop_acked", + "send_ack_required", + "send_queued", + "send_rdma", + "send_rdma_bytes", + "send_pong", + "page_remainder_hit", + "page_remainder_miss", + "copy_to_user", + "copy_from_user", + "cong_update_queued", + "cong_update_received", + "cong_send_error", + "cong_send_blocked", + "recv_bytes_added_to_sock", + "recv_bytes_freed_fromsock", +}; + +void rds_stats_info_copy(struct rds_info_iterator *iter, + uint64_t *values, const char *const *names, size_t nr) +{ + struct rds_info_counter ctr; + size_t i; + + for (i = 0; i < nr; i++) { + BUG_ON(strlen(names[i]) >= sizeof(ctr.name)); + strncpy(ctr.name, names[i], sizeof(ctr.name) - 1); + ctr.name[sizeof(ctr.name) - 1] = '\0'; + ctr.value = values[i]; + + rds_info_copy(iter, &ctr, sizeof(ctr)); + } +} +EXPORT_SYMBOL_GPL(rds_stats_info_copy); + +/* + * This gives global counters across all the transports. The strings + * are copied in so that the tool doesn't need knowledge of the specific + * stats that we're exporting. Some are pretty implementation dependent + * and may change over time. That doesn't stop them from being useful. + * + * This is the only function in the chain that knows about the byte granular + * length in userspace. It converts it to number of stat entries that the + * rest of the functions operate in. + */ +static void rds_stats_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + struct rds_statistics stats = {0, }; + uint64_t *src; + uint64_t *sum; + size_t i; + int cpu; + unsigned int avail; + + avail = len / sizeof(struct rds_info_counter); + + if (avail < ARRAY_SIZE(rds_stat_names)) { + avail = 0; + goto trans; + } + + for_each_online_cpu(cpu) { + src = (uint64_t *)&(per_cpu(rds_stats, cpu)); + sum = (uint64_t *)&stats; + for (i = 0; i < sizeof(stats) / sizeof(uint64_t); i++) + *(sum++) += *(src++); + } + + rds_stats_info_copy(iter, (uint64_t *)&stats, rds_stat_names, + ARRAY_SIZE(rds_stat_names)); + avail -= ARRAY_SIZE(rds_stat_names); + +trans: + lens->each = sizeof(struct rds_info_counter); + lens->nr = rds_trans_stats_info_copy(iter, avail) + + ARRAY_SIZE(rds_stat_names); +} + +void rds_stats_exit(void) +{ + rds_info_deregister_func(RDS_INFO_COUNTERS, rds_stats_info); +} + +int rds_stats_init(void) +{ + rds_info_register_func(RDS_INFO_COUNTERS, rds_stats_info); + return 0; +} diff --git a/net/rds/sysctl.c b/net/rds/sysctl.c new file mode 100644 index 000000000..e381bbcd9 --- /dev/null +++ b/net/rds/sysctl.c @@ -0,0 +1,110 @@ +/* + * Copyright (c) 2006 Oracle. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/sysctl.h> +#include <linux/proc_fs.h> + +#include "rds.h" + +static struct ctl_table_header *rds_sysctl_reg_table; + +static unsigned long rds_sysctl_reconnect_min = 1; +static unsigned long rds_sysctl_reconnect_max = ~0UL; + +unsigned long rds_sysctl_reconnect_min_jiffies; +unsigned long rds_sysctl_reconnect_max_jiffies = HZ; + +unsigned int rds_sysctl_max_unacked_packets = 8; +unsigned int rds_sysctl_max_unacked_bytes = (16 << 20); + +unsigned int rds_sysctl_ping_enable = 1; + +static struct ctl_table rds_sysctl_rds_table[] = { + { + .procname = "reconnect_min_delay_ms", + .data = &rds_sysctl_reconnect_min_jiffies, + .maxlen = sizeof(unsigned long), + .mode = 0644, + .proc_handler = proc_doulongvec_ms_jiffies_minmax, + .extra1 = &rds_sysctl_reconnect_min, + .extra2 = &rds_sysctl_reconnect_max_jiffies, + }, + { + .procname = "reconnect_max_delay_ms", + .data = &rds_sysctl_reconnect_max_jiffies, + .maxlen = sizeof(unsigned long), + .mode = 0644, + .proc_handler = proc_doulongvec_ms_jiffies_minmax, + .extra1 = &rds_sysctl_reconnect_min_jiffies, + .extra2 = &rds_sysctl_reconnect_max, + }, + { + .procname = "max_unacked_packets", + .data = &rds_sysctl_max_unacked_packets, + .maxlen = sizeof(int), + .mode = 0644, + .proc_handler = proc_dointvec, + }, + { + .procname = "max_unacked_bytes", + .data = &rds_sysctl_max_unacked_bytes, + .maxlen = sizeof(int), + .mode = 0644, + .proc_handler = proc_dointvec, + }, + { + .procname = "ping_enable", + .data = &rds_sysctl_ping_enable, + .maxlen = sizeof(int), + .mode = 0644, + .proc_handler = proc_dointvec, + }, + { } +}; + +void rds_sysctl_exit(void) +{ + unregister_net_sysctl_table(rds_sysctl_reg_table); +} + +int rds_sysctl_init(void) +{ + rds_sysctl_reconnect_min = msecs_to_jiffies(1); + rds_sysctl_reconnect_min_jiffies = rds_sysctl_reconnect_min; + + rds_sysctl_reg_table = + register_net_sysctl(&init_net, "net/rds", rds_sysctl_rds_table); + if (!rds_sysctl_reg_table) + return -ENOMEM; + return 0; +} diff --git a/net/rds/tcp.c b/net/rds/tcp.c new file mode 100644 index 000000000..d4e6466d3 --- /dev/null +++ b/net/rds/tcp.c @@ -0,0 +1,743 @@ +/* + * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/slab.h> +#include <linux/in.h> +#include <linux/module.h> +#include <net/tcp.h> +#include <net/net_namespace.h> +#include <net/netns/generic.h> +#include <net/addrconf.h> + +#include "rds.h" +#include "tcp.h" + +/* only for info exporting */ +static DEFINE_SPINLOCK(rds_tcp_tc_list_lock); +static LIST_HEAD(rds_tcp_tc_list); + +/* rds_tcp_tc_count counts only IPv4 connections. + * rds6_tcp_tc_count counts both IPv4 and IPv6 connections. + */ +static unsigned int rds_tcp_tc_count; +#if IS_ENABLED(CONFIG_IPV6) +static unsigned int rds6_tcp_tc_count; +#endif + +/* Track rds_tcp_connection structs so they can be cleaned up */ +static DEFINE_SPINLOCK(rds_tcp_conn_lock); +static LIST_HEAD(rds_tcp_conn_list); +static atomic_t rds_tcp_unloading = ATOMIC_INIT(0); + +static struct kmem_cache *rds_tcp_conn_slab; + +static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write, + void __user *buffer, size_t *lenp, + loff_t *fpos); + +static int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF; +static int rds_tcp_min_rcvbuf = SOCK_MIN_RCVBUF; + +static struct ctl_table rds_tcp_sysctl_table[] = { +#define RDS_TCP_SNDBUF 0 + { + .procname = "rds_tcp_sndbuf", + /* data is per-net pointer */ + .maxlen = sizeof(int), + .mode = 0644, + .proc_handler = rds_tcp_skbuf_handler, + .extra1 = &rds_tcp_min_sndbuf, + }, +#define RDS_TCP_RCVBUF 1 + { + .procname = "rds_tcp_rcvbuf", + /* data is per-net pointer */ + .maxlen = sizeof(int), + .mode = 0644, + .proc_handler = rds_tcp_skbuf_handler, + .extra1 = &rds_tcp_min_rcvbuf, + }, + { } +}; + +/* doing it this way avoids calling tcp_sk() */ +void rds_tcp_nonagle(struct socket *sock) +{ + int val = 1; + + kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (void *)&val, + sizeof(val)); +} + +u32 rds_tcp_write_seq(struct rds_tcp_connection *tc) +{ + /* seq# of the last byte of data in tcp send buffer */ + return tcp_sk(tc->t_sock->sk)->write_seq; +} + +u32 rds_tcp_snd_una(struct rds_tcp_connection *tc) +{ + return tcp_sk(tc->t_sock->sk)->snd_una; +} + +void rds_tcp_restore_callbacks(struct socket *sock, + struct rds_tcp_connection *tc) +{ + rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc); + write_lock_bh(&sock->sk->sk_callback_lock); + + /* done under the callback_lock to serialize with write_space */ + spin_lock(&rds_tcp_tc_list_lock); + list_del_init(&tc->t_list_item); +#if IS_ENABLED(CONFIG_IPV6) + rds6_tcp_tc_count--; +#endif + if (!tc->t_cpath->cp_conn->c_isv6) + rds_tcp_tc_count--; + spin_unlock(&rds_tcp_tc_list_lock); + + tc->t_sock = NULL; + + sock->sk->sk_write_space = tc->t_orig_write_space; + sock->sk->sk_data_ready = tc->t_orig_data_ready; + sock->sk->sk_state_change = tc->t_orig_state_change; + sock->sk->sk_user_data = NULL; + + write_unlock_bh(&sock->sk->sk_callback_lock); +} + +/* + * rds_tcp_reset_callbacks() switches the to the new sock and + * returns the existing tc->t_sock. + * + * The only functions that set tc->t_sock are rds_tcp_set_callbacks + * and rds_tcp_reset_callbacks. Send and receive trust that + * it is set. The absence of RDS_CONN_UP bit protects those paths + * from being called while it isn't set. + */ +void rds_tcp_reset_callbacks(struct socket *sock, + struct rds_conn_path *cp) +{ + struct rds_tcp_connection *tc = cp->cp_transport_data; + struct socket *osock = tc->t_sock; + + if (!osock) + goto newsock; + + /* Need to resolve a duelling SYN between peers. + * We have an outstanding SYN to this peer, which may + * potentially have transitioned to the RDS_CONN_UP state, + * so we must quiesce any send threads before resetting + * cp_transport_data. We quiesce these threads by setting + * cp_state to something other than RDS_CONN_UP, and then + * waiting for any existing threads in rds_send_xmit to + * complete release_in_xmit(). (Subsequent threads entering + * rds_send_xmit() will bail on !rds_conn_up(). + * + * However an incoming syn-ack at this point would end up + * marking the conn as RDS_CONN_UP, and would again permit + * rds_send_xmi() threads through, so ideally we would + * synchronize on RDS_CONN_UP after lock_sock(), but cannot + * do that: waiting on !RDS_IN_XMIT after lock_sock() may + * end up deadlocking with tcp_sendmsg(), and the RDS_IN_XMIT + * would not get set. As a result, we set c_state to + * RDS_CONN_RESETTTING, to ensure that rds_tcp_state_change + * cannot mark rds_conn_path_up() in the window before lock_sock() + */ + atomic_set(&cp->cp_state, RDS_CONN_RESETTING); + wait_event(cp->cp_waitq, !test_bit(RDS_IN_XMIT, &cp->cp_flags)); + lock_sock(osock->sk); + /* reset receive side state for rds_tcp_data_recv() for osock */ + cancel_delayed_work_sync(&cp->cp_send_w); + cancel_delayed_work_sync(&cp->cp_recv_w); + if (tc->t_tinc) { + rds_inc_put(&tc->t_tinc->ti_inc); + tc->t_tinc = NULL; + } + tc->t_tinc_hdr_rem = sizeof(struct rds_header); + tc->t_tinc_data_rem = 0; + rds_tcp_restore_callbacks(osock, tc); + release_sock(osock->sk); + sock_release(osock); +newsock: + rds_send_path_reset(cp); + lock_sock(sock->sk); + rds_tcp_set_callbacks(sock, cp); + release_sock(sock->sk); +} + +/* Add tc to rds_tcp_tc_list and set tc->t_sock. See comments + * above rds_tcp_reset_callbacks for notes about synchronization + * with data path + */ +void rds_tcp_set_callbacks(struct socket *sock, struct rds_conn_path *cp) +{ + struct rds_tcp_connection *tc = cp->cp_transport_data; + + rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc); + write_lock_bh(&sock->sk->sk_callback_lock); + + /* done under the callback_lock to serialize with write_space */ + spin_lock(&rds_tcp_tc_list_lock); + list_add_tail(&tc->t_list_item, &rds_tcp_tc_list); +#if IS_ENABLED(CONFIG_IPV6) + rds6_tcp_tc_count++; +#endif + if (!tc->t_cpath->cp_conn->c_isv6) + rds_tcp_tc_count++; + spin_unlock(&rds_tcp_tc_list_lock); + + /* accepted sockets need our listen data ready undone */ + if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready) + sock->sk->sk_data_ready = sock->sk->sk_user_data; + + tc->t_sock = sock; + tc->t_cpath = cp; + tc->t_orig_data_ready = sock->sk->sk_data_ready; + tc->t_orig_write_space = sock->sk->sk_write_space; + tc->t_orig_state_change = sock->sk->sk_state_change; + + sock->sk->sk_user_data = cp; + sock->sk->sk_data_ready = rds_tcp_data_ready; + sock->sk->sk_write_space = rds_tcp_write_space; + sock->sk->sk_state_change = rds_tcp_state_change; + + write_unlock_bh(&sock->sk->sk_callback_lock); +} + +/* Handle RDS_INFO_TCP_SOCKETS socket option. It only returns IPv4 + * connections for backward compatibility. + */ +static void rds_tcp_tc_info(struct socket *rds_sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + struct rds_info_tcp_socket tsinfo; + struct rds_tcp_connection *tc; + unsigned long flags; + + spin_lock_irqsave(&rds_tcp_tc_list_lock, flags); + + if (len / sizeof(tsinfo) < rds_tcp_tc_count) + goto out; + + list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) { + struct inet_sock *inet = inet_sk(tc->t_sock->sk); + + if (tc->t_cpath->cp_conn->c_isv6) + continue; + + tsinfo.local_addr = inet->inet_saddr; + tsinfo.local_port = inet->inet_sport; + tsinfo.peer_addr = inet->inet_daddr; + tsinfo.peer_port = inet->inet_dport; + + tsinfo.hdr_rem = tc->t_tinc_hdr_rem; + tsinfo.data_rem = tc->t_tinc_data_rem; + tsinfo.last_sent_nxt = tc->t_last_sent_nxt; + tsinfo.last_expected_una = tc->t_last_expected_una; + tsinfo.last_seen_una = tc->t_last_seen_una; + + rds_info_copy(iter, &tsinfo, sizeof(tsinfo)); + } + +out: + lens->nr = rds_tcp_tc_count; + lens->each = sizeof(tsinfo); + + spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags); +} + +#if IS_ENABLED(CONFIG_IPV6) +/* Handle RDS6_INFO_TCP_SOCKETS socket option. It returns both IPv4 and + * IPv6 connections. IPv4 connection address is returned in an IPv4 mapped + * address. + */ +static void rds6_tcp_tc_info(struct socket *sock, unsigned int len, + struct rds_info_iterator *iter, + struct rds_info_lengths *lens) +{ + struct rds6_info_tcp_socket tsinfo6; + struct rds_tcp_connection *tc; + unsigned long flags; + + spin_lock_irqsave(&rds_tcp_tc_list_lock, flags); + + if (len / sizeof(tsinfo6) < rds6_tcp_tc_count) + goto out; + + list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) { + struct sock *sk = tc->t_sock->sk; + struct inet_sock *inet = inet_sk(sk); + + tsinfo6.local_addr = sk->sk_v6_rcv_saddr; + tsinfo6.local_port = inet->inet_sport; + tsinfo6.peer_addr = sk->sk_v6_daddr; + tsinfo6.peer_port = inet->inet_dport; + + tsinfo6.hdr_rem = tc->t_tinc_hdr_rem; + tsinfo6.data_rem = tc->t_tinc_data_rem; + tsinfo6.last_sent_nxt = tc->t_last_sent_nxt; + tsinfo6.last_expected_una = tc->t_last_expected_una; + tsinfo6.last_seen_una = tc->t_last_seen_una; + + rds_info_copy(iter, &tsinfo6, sizeof(tsinfo6)); + } + +out: + lens->nr = rds6_tcp_tc_count; + lens->each = sizeof(tsinfo6); + + spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags); +} +#endif + +int rds_tcp_laddr_check(struct net *net, const struct in6_addr *addr, + __u32 scope_id) +{ + struct net_device *dev = NULL; +#if IS_ENABLED(CONFIG_IPV6) + int ret; +#endif + + if (ipv6_addr_v4mapped(addr)) { + if (inet_addr_type(net, addr->s6_addr32[3]) == RTN_LOCAL) + return 0; + return -EADDRNOTAVAIL; + } + + /* If the scope_id is specified, check only those addresses + * hosted on the specified interface. + */ + if (scope_id != 0) { + rcu_read_lock(); + dev = dev_get_by_index_rcu(net, scope_id); + /* scope_id is not valid... */ + if (!dev) { + rcu_read_unlock(); + return -EADDRNOTAVAIL; + } + rcu_read_unlock(); + } +#if IS_ENABLED(CONFIG_IPV6) + ret = ipv6_chk_addr(net, addr, dev, 0); + if (ret) + return 0; +#endif + return -EADDRNOTAVAIL; +} + +static void rds_tcp_conn_free(void *arg) +{ + struct rds_tcp_connection *tc = arg; + unsigned long flags; + + rdsdebug("freeing tc %p\n", tc); + + spin_lock_irqsave(&rds_tcp_conn_lock, flags); + if (!tc->t_tcp_node_detached) + list_del(&tc->t_tcp_node); + spin_unlock_irqrestore(&rds_tcp_conn_lock, flags); + + kmem_cache_free(rds_tcp_conn_slab, tc); +} + +static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp) +{ + struct rds_tcp_connection *tc; + int i, j; + int ret = 0; + + for (i = 0; i < RDS_MPATH_WORKERS; i++) { + tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp); + if (!tc) { + ret = -ENOMEM; + goto fail; + } + mutex_init(&tc->t_conn_path_lock); + tc->t_sock = NULL; + tc->t_tinc = NULL; + tc->t_tinc_hdr_rem = sizeof(struct rds_header); + tc->t_tinc_data_rem = 0; + + conn->c_path[i].cp_transport_data = tc; + tc->t_cpath = &conn->c_path[i]; + tc->t_tcp_node_detached = true; + + rdsdebug("rds_conn_path [%d] tc %p\n", i, + conn->c_path[i].cp_transport_data); + } + spin_lock_irq(&rds_tcp_conn_lock); + for (i = 0; i < RDS_MPATH_WORKERS; i++) { + tc = conn->c_path[i].cp_transport_data; + tc->t_tcp_node_detached = false; + list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list); + } + spin_unlock_irq(&rds_tcp_conn_lock); +fail: + if (ret) { + for (j = 0; j < i; j++) + rds_tcp_conn_free(conn->c_path[j].cp_transport_data); + } + return ret; +} + +static bool list_has_conn(struct list_head *list, struct rds_connection *conn) +{ + struct rds_tcp_connection *tc, *_tc; + + list_for_each_entry_safe(tc, _tc, list, t_tcp_node) { + if (tc->t_cpath->cp_conn == conn) + return true; + } + return false; +} + +static void rds_tcp_set_unloading(void) +{ + atomic_set(&rds_tcp_unloading, 1); +} + +static bool rds_tcp_is_unloading(struct rds_connection *conn) +{ + return atomic_read(&rds_tcp_unloading) != 0; +} + +static void rds_tcp_destroy_conns(void) +{ + struct rds_tcp_connection *tc, *_tc; + LIST_HEAD(tmp_list); + + /* avoid calling conn_destroy with irqs off */ + spin_lock_irq(&rds_tcp_conn_lock); + list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) { + if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn)) + list_move_tail(&tc->t_tcp_node, &tmp_list); + } + spin_unlock_irq(&rds_tcp_conn_lock); + + list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) + rds_conn_destroy(tc->t_cpath->cp_conn); +} + +static void rds_tcp_exit(void); + +struct rds_transport rds_tcp_transport = { + .laddr_check = rds_tcp_laddr_check, + .xmit_path_prepare = rds_tcp_xmit_path_prepare, + .xmit_path_complete = rds_tcp_xmit_path_complete, + .xmit = rds_tcp_xmit, + .recv_path = rds_tcp_recv_path, + .conn_alloc = rds_tcp_conn_alloc, + .conn_free = rds_tcp_conn_free, + .conn_path_connect = rds_tcp_conn_path_connect, + .conn_path_shutdown = rds_tcp_conn_path_shutdown, + .inc_copy_to_user = rds_tcp_inc_copy_to_user, + .inc_free = rds_tcp_inc_free, + .stats_info_copy = rds_tcp_stats_info_copy, + .exit = rds_tcp_exit, + .t_owner = THIS_MODULE, + .t_name = "tcp", + .t_type = RDS_TRANS_TCP, + .t_prefer_loopback = 1, + .t_mp_capable = 1, + .t_unloading = rds_tcp_is_unloading, +}; + +static unsigned int rds_tcp_netid; + +/* per-network namespace private data for this module */ +struct rds_tcp_net { + struct socket *rds_tcp_listen_sock; + struct work_struct rds_tcp_accept_w; + struct ctl_table_header *rds_tcp_sysctl; + struct ctl_table *ctl_table; + int sndbuf_size; + int rcvbuf_size; +}; + +/* All module specific customizations to the RDS-TCP socket should be done in + * rds_tcp_tune() and applied after socket creation. + */ +void rds_tcp_tune(struct socket *sock) +{ + struct sock *sk = sock->sk; + struct net *net = sock_net(sk); + struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid); + + rds_tcp_nonagle(sock); + lock_sock(sk); + if (rtn->sndbuf_size > 0) { + sk->sk_sndbuf = rtn->sndbuf_size; + sk->sk_userlocks |= SOCK_SNDBUF_LOCK; + } + if (rtn->rcvbuf_size > 0) { + sk->sk_rcvbuf = rtn->rcvbuf_size; + sk->sk_userlocks |= SOCK_RCVBUF_LOCK; + } + release_sock(sk); +} + +static void rds_tcp_accept_worker(struct work_struct *work) +{ + struct rds_tcp_net *rtn = container_of(work, + struct rds_tcp_net, + rds_tcp_accept_w); + + while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0) + cond_resched(); +} + +void rds_tcp_accept_work(struct sock *sk) +{ + struct net *net = sock_net(sk); + struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid); + + queue_work(rds_wq, &rtn->rds_tcp_accept_w); +} + +static __net_init int rds_tcp_init_net(struct net *net) +{ + struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid); + struct ctl_table *tbl; + int err = 0; + + memset(rtn, 0, sizeof(*rtn)); + + /* {snd, rcv}buf_size default to 0, which implies we let the + * stack pick the value, and permit auto-tuning of buffer size. + */ + if (net == &init_net) { + tbl = rds_tcp_sysctl_table; + } else { + tbl = kmemdup(rds_tcp_sysctl_table, + sizeof(rds_tcp_sysctl_table), GFP_KERNEL); + if (!tbl) { + pr_warn("could not set allocate syctl table\n"); + return -ENOMEM; + } + rtn->ctl_table = tbl; + } + tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size; + tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size; + rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl); + if (!rtn->rds_tcp_sysctl) { + pr_warn("could not register sysctl\n"); + err = -ENOMEM; + goto fail; + } + +#if IS_ENABLED(CONFIG_IPV6) + rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net, true); +#else + rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net, false); +#endif + if (!rtn->rds_tcp_listen_sock) { + pr_warn("could not set up IPv6 listen sock\n"); + +#if IS_ENABLED(CONFIG_IPV6) + /* Try IPv4 as some systems disable IPv6 */ + rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net, false); + if (!rtn->rds_tcp_listen_sock) { +#endif + unregister_net_sysctl_table(rtn->rds_tcp_sysctl); + rtn->rds_tcp_sysctl = NULL; + err = -EAFNOSUPPORT; + goto fail; +#if IS_ENABLED(CONFIG_IPV6) + } +#endif + } + INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker); + return 0; + +fail: + if (net != &init_net) + kfree(tbl); + return err; +} + +static void rds_tcp_kill_sock(struct net *net) +{ + struct rds_tcp_connection *tc, *_tc; + LIST_HEAD(tmp_list); + struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid); + struct socket *lsock = rtn->rds_tcp_listen_sock; + + rtn->rds_tcp_listen_sock = NULL; + rds_tcp_listen_stop(lsock, &rtn->rds_tcp_accept_w); + spin_lock_irq(&rds_tcp_conn_lock); + list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) { + struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net); + + if (net != c_net) + continue; + if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn)) { + list_move_tail(&tc->t_tcp_node, &tmp_list); + } else { + list_del(&tc->t_tcp_node); + tc->t_tcp_node_detached = true; + } + } + spin_unlock_irq(&rds_tcp_conn_lock); + list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) + rds_conn_destroy(tc->t_cpath->cp_conn); +} + +static void __net_exit rds_tcp_exit_net(struct net *net) +{ + struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid); + + rds_tcp_kill_sock(net); + + if (rtn->rds_tcp_sysctl) + unregister_net_sysctl_table(rtn->rds_tcp_sysctl); + + if (net != &init_net && rtn->ctl_table) + kfree(rtn->ctl_table); +} + +static struct pernet_operations rds_tcp_net_ops = { + .init = rds_tcp_init_net, + .exit = rds_tcp_exit_net, + .id = &rds_tcp_netid, + .size = sizeof(struct rds_tcp_net), +}; + +void *rds_tcp_listen_sock_def_readable(struct net *net) +{ + struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid); + struct socket *lsock = rtn->rds_tcp_listen_sock; + + if (!lsock) + return NULL; + + return lsock->sk->sk_user_data; +} + +/* when sysctl is used to modify some kernel socket parameters,this + * function resets the RDS connections in that netns so that we can + * restart with new parameters. The assumption is that such reset + * events are few and far-between. + */ +static void rds_tcp_sysctl_reset(struct net *net) +{ + struct rds_tcp_connection *tc, *_tc; + + spin_lock_irq(&rds_tcp_conn_lock); + list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) { + struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net); + + if (net != c_net || !tc->t_sock) + continue; + + /* reconnect with new parameters */ + rds_conn_path_drop(tc->t_cpath, false); + } + spin_unlock_irq(&rds_tcp_conn_lock); +} + +static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write, + void __user *buffer, size_t *lenp, + loff_t *fpos) +{ + struct net *net = current->nsproxy->net_ns; + int err; + + err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos); + if (err < 0) { + pr_warn("Invalid input. Must be >= %d\n", + *(int *)(ctl->extra1)); + return err; + } + if (write) + rds_tcp_sysctl_reset(net); + return 0; +} + +static void rds_tcp_exit(void) +{ + rds_tcp_set_unloading(); + synchronize_rcu(); + rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info); +#if IS_ENABLED(CONFIG_IPV6) + rds_info_deregister_func(RDS6_INFO_TCP_SOCKETS, rds6_tcp_tc_info); +#endif + unregister_pernet_device(&rds_tcp_net_ops); + rds_tcp_destroy_conns(); + rds_trans_unregister(&rds_tcp_transport); + rds_tcp_recv_exit(); + kmem_cache_destroy(rds_tcp_conn_slab); +} +module_exit(rds_tcp_exit); + +static int rds_tcp_init(void) +{ + int ret; + + rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection", + sizeof(struct rds_tcp_connection), + 0, 0, NULL); + if (!rds_tcp_conn_slab) { + ret = -ENOMEM; + goto out; + } + + ret = rds_tcp_recv_init(); + if (ret) + goto out_slab; + + ret = register_pernet_device(&rds_tcp_net_ops); + if (ret) + goto out_recv; + + rds_trans_register(&rds_tcp_transport); + + rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info); +#if IS_ENABLED(CONFIG_IPV6) + rds_info_register_func(RDS6_INFO_TCP_SOCKETS, rds6_tcp_tc_info); +#endif + + goto out; +out_recv: + rds_tcp_recv_exit(); +out_slab: + kmem_cache_destroy(rds_tcp_conn_slab); +out: + return ret; +} +module_init(rds_tcp_init); + +MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>"); +MODULE_DESCRIPTION("RDS: TCP transport"); +MODULE_LICENSE("Dual BSD/GPL"); diff --git a/net/rds/tcp.h b/net/rds/tcp.h new file mode 100644 index 000000000..4620549ec --- /dev/null +++ b/net/rds/tcp.h @@ -0,0 +1,100 @@ +/* SPDX-License-Identifier: GPL-2.0 */ +#ifndef _RDS_TCP_H +#define _RDS_TCP_H + +#define RDS_TCP_PORT 16385 + +struct rds_tcp_incoming { + struct rds_incoming ti_inc; + struct sk_buff_head ti_skb_list; +}; + +struct rds_tcp_connection { + + struct list_head t_tcp_node; + bool t_tcp_node_detached; + struct rds_conn_path *t_cpath; + /* t_conn_path_lock synchronizes the connection establishment between + * rds_tcp_accept_one and rds_tcp_conn_path_connect + */ + struct mutex t_conn_path_lock; + struct socket *t_sock; + void *t_orig_write_space; + void *t_orig_data_ready; + void *t_orig_state_change; + + struct rds_tcp_incoming *t_tinc; + size_t t_tinc_hdr_rem; + size_t t_tinc_data_rem; + + /* XXX error report? */ + struct work_struct t_conn_w; + struct work_struct t_send_w; + struct work_struct t_down_w; + struct work_struct t_recv_w; + + /* for info exporting only */ + struct list_head t_list_item; + u32 t_last_sent_nxt; + u32 t_last_expected_una; + u32 t_last_seen_una; +}; + +struct rds_tcp_statistics { + uint64_t s_tcp_data_ready_calls; + uint64_t s_tcp_write_space_calls; + uint64_t s_tcp_sndbuf_full; + uint64_t s_tcp_connect_raced; + uint64_t s_tcp_listen_closed_stale; +}; + +/* tcp.c */ +void rds_tcp_tune(struct socket *sock); +void rds_tcp_nonagle(struct socket *sock); +void rds_tcp_set_callbacks(struct socket *sock, struct rds_conn_path *cp); +void rds_tcp_reset_callbacks(struct socket *sock, struct rds_conn_path *cp); +void rds_tcp_restore_callbacks(struct socket *sock, + struct rds_tcp_connection *tc); +u32 rds_tcp_write_seq(struct rds_tcp_connection *tc); +u32 rds_tcp_snd_una(struct rds_tcp_connection *tc); +u64 rds_tcp_map_seq(struct rds_tcp_connection *tc, u32 seq); +extern struct rds_transport rds_tcp_transport; +void rds_tcp_accept_work(struct sock *sk); +int rds_tcp_laddr_check(struct net *net, const struct in6_addr *addr, + __u32 scope_id); +/* tcp_connect.c */ +int rds_tcp_conn_path_connect(struct rds_conn_path *cp); +void rds_tcp_conn_path_shutdown(struct rds_conn_path *conn); +void rds_tcp_state_change(struct sock *sk); + +/* tcp_listen.c */ +struct socket *rds_tcp_listen_init(struct net *net, bool isv6); +void rds_tcp_listen_stop(struct socket *sock, struct work_struct *acceptor); +void rds_tcp_listen_data_ready(struct sock *sk); +int rds_tcp_accept_one(struct socket *sock); +int rds_tcp_keepalive(struct socket *sock); +void *rds_tcp_listen_sock_def_readable(struct net *net); +void rds_tcp_set_linger(struct socket *sock); + +/* tcp_recv.c */ +int rds_tcp_recv_init(void); +void rds_tcp_recv_exit(void); +void rds_tcp_data_ready(struct sock *sk); +int rds_tcp_recv_path(struct rds_conn_path *cp); +void rds_tcp_inc_free(struct rds_incoming *inc); +int rds_tcp_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to); + +/* tcp_send.c */ +void rds_tcp_xmit_path_prepare(struct rds_conn_path *cp); +void rds_tcp_xmit_path_complete(struct rds_conn_path *cp); +int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm, + unsigned int hdr_off, unsigned int sg, unsigned int off); +void rds_tcp_write_space(struct sock *sk); + +/* tcp_stats.c */ +DECLARE_PER_CPU(struct rds_tcp_statistics, rds_tcp_stats); +#define rds_tcp_stats_inc(member) rds_stats_inc_which(rds_tcp_stats, member) +unsigned int rds_tcp_stats_info_copy(struct rds_info_iterator *iter, + unsigned int avail); + +#endif diff --git a/net/rds/tcp_connect.c b/net/rds/tcp_connect.c new file mode 100644 index 000000000..008f50fb2 --- /dev/null +++ b/net/rds/tcp_connect.c @@ -0,0 +1,225 @@ +/* + * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/in.h> +#include <net/tcp.h> + +#include "rds.h" +#include "tcp.h" + +void rds_tcp_state_change(struct sock *sk) +{ + void (*state_change)(struct sock *sk); + struct rds_conn_path *cp; + struct rds_tcp_connection *tc; + + read_lock_bh(&sk->sk_callback_lock); + cp = sk->sk_user_data; + if (!cp) { + state_change = sk->sk_state_change; + goto out; + } + tc = cp->cp_transport_data; + state_change = tc->t_orig_state_change; + + rdsdebug("sock %p state_change to %d\n", tc->t_sock, sk->sk_state); + + switch (sk->sk_state) { + /* ignore connecting sockets as they make progress */ + case TCP_SYN_SENT: + case TCP_SYN_RECV: + break; + case TCP_ESTABLISHED: + /* Force the peer to reconnect so that we have the + * TCP ports going from <smaller-ip>.<transient> to + * <larger-ip>.<RDS_TCP_PORT>. We avoid marking the + * RDS connection as RDS_CONN_UP until the reconnect, + * to avoid RDS datagram loss. + */ + if (rds_addr_cmp(&cp->cp_conn->c_laddr, + &cp->cp_conn->c_faddr) >= 0 && + rds_conn_path_transition(cp, RDS_CONN_CONNECTING, + RDS_CONN_ERROR)) { + rds_conn_path_drop(cp, false); + } else { + rds_connect_path_complete(cp, RDS_CONN_CONNECTING); + } + break; + case TCP_CLOSE_WAIT: + case TCP_CLOSE: + rds_conn_path_drop(cp, false); + default: + break; + } +out: + read_unlock_bh(&sk->sk_callback_lock); + state_change(sk); +} + +int rds_tcp_conn_path_connect(struct rds_conn_path *cp) +{ + struct socket *sock = NULL; + struct sockaddr_in6 sin6; + struct sockaddr_in sin; + struct sockaddr *addr; + int addrlen; + bool isv6; + int ret; + struct rds_connection *conn = cp->cp_conn; + struct rds_tcp_connection *tc = cp->cp_transport_data; + + /* for multipath rds,we only trigger the connection after + * the handshake probe has determined the number of paths. + */ + if (cp->cp_index > 0 && cp->cp_conn->c_npaths < 2) + return -EAGAIN; + + mutex_lock(&tc->t_conn_path_lock); + + if (rds_conn_path_up(cp)) { + mutex_unlock(&tc->t_conn_path_lock); + return 0; + } + if (ipv6_addr_v4mapped(&conn->c_laddr)) { + ret = sock_create_kern(rds_conn_net(conn), PF_INET, + SOCK_STREAM, IPPROTO_TCP, &sock); + isv6 = false; + } else { + ret = sock_create_kern(rds_conn_net(conn), PF_INET6, + SOCK_STREAM, IPPROTO_TCP, &sock); + isv6 = true; + } + + if (ret < 0) + goto out; + + rds_tcp_tune(sock); + + if (isv6) { + sin6.sin6_family = AF_INET6; + sin6.sin6_addr = conn->c_laddr; + sin6.sin6_port = 0; + sin6.sin6_flowinfo = 0; + sin6.sin6_scope_id = conn->c_dev_if; + addr = (struct sockaddr *)&sin6; + addrlen = sizeof(sin6); + } else { + sin.sin_family = AF_INET; + sin.sin_addr.s_addr = conn->c_laddr.s6_addr32[3]; + sin.sin_port = 0; + addr = (struct sockaddr *)&sin; + addrlen = sizeof(sin); + } + + ret = sock->ops->bind(sock, addr, addrlen); + if (ret) { + rdsdebug("bind failed with %d at address %pI6c\n", + ret, &conn->c_laddr); + goto out; + } + + if (isv6) { + sin6.sin6_family = AF_INET6; + sin6.sin6_addr = conn->c_faddr; + sin6.sin6_port = htons(RDS_TCP_PORT); + sin6.sin6_flowinfo = 0; + sin6.sin6_scope_id = conn->c_dev_if; + addr = (struct sockaddr *)&sin6; + addrlen = sizeof(sin6); + } else { + sin.sin_family = AF_INET; + sin.sin_addr.s_addr = conn->c_faddr.s6_addr32[3]; + sin.sin_port = htons(RDS_TCP_PORT); + addr = (struct sockaddr *)&sin; + addrlen = sizeof(sin); + } + + /* + * once we call connect() we can start getting callbacks and they + * own the socket + */ + rds_tcp_set_callbacks(sock, cp); + ret = sock->ops->connect(sock, addr, addrlen, O_NONBLOCK); + + rdsdebug("connect to address %pI6c returned %d\n", &conn->c_faddr, ret); + if (ret == -EINPROGRESS) + ret = 0; + if (ret == 0) { + rds_tcp_keepalive(sock); + sock = NULL; + } else { + rds_tcp_restore_callbacks(sock, cp->cp_transport_data); + } + +out: + mutex_unlock(&tc->t_conn_path_lock); + if (sock) + sock_release(sock); + return ret; +} + +/* + * Before killing the tcp socket this needs to serialize with callbacks. The + * caller has already grabbed the sending sem so we're serialized with other + * senders. + * + * TCP calls the callbacks with the sock lock so we hold it while we reset the + * callbacks to those set by TCP. Our callbacks won't execute again once we + * hold the sock lock. + */ +void rds_tcp_conn_path_shutdown(struct rds_conn_path *cp) +{ + struct rds_tcp_connection *tc = cp->cp_transport_data; + struct socket *sock = tc->t_sock; + + rdsdebug("shutting down conn %p tc %p sock %p\n", + cp->cp_conn, tc, sock); + + if (sock) { + if (rds_destroy_pending(cp->cp_conn)) + rds_tcp_set_linger(sock); + sock->ops->shutdown(sock, RCV_SHUTDOWN | SEND_SHUTDOWN); + lock_sock(sock->sk); + rds_tcp_restore_callbacks(sock, tc); /* tc->tc_sock = NULL */ + + release_sock(sock->sk); + sock_release(sock); + } + + if (tc->t_tinc) { + rds_inc_put(&tc->t_tinc->ti_inc); + tc->t_tinc = NULL; + } + tc->t_tinc_hdr_rem = sizeof(struct rds_header); + tc->t_tinc_data_rem = 0; +} diff --git a/net/rds/tcp_listen.c b/net/rds/tcp_listen.c new file mode 100644 index 000000000..0d095d3f5 --- /dev/null +++ b/net/rds/tcp_listen.c @@ -0,0 +1,376 @@ +/* + * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/gfp.h> +#include <linux/in.h> +#include <net/tcp.h> + +#include "rds.h" +#include "tcp.h" + +int rds_tcp_keepalive(struct socket *sock) +{ + /* values below based on xs_udp_default_timeout */ + int keepidle = 5; /* send a probe 'keepidle' secs after last data */ + int keepcnt = 5; /* number of unack'ed probes before declaring dead */ + int keepalive = 1; + int ret = 0; + + ret = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE, + (char *)&keepalive, sizeof(keepalive)); + if (ret < 0) + goto bail; + + ret = kernel_setsockopt(sock, IPPROTO_TCP, TCP_KEEPCNT, + (char *)&keepcnt, sizeof(keepcnt)); + if (ret < 0) + goto bail; + + ret = kernel_setsockopt(sock, IPPROTO_TCP, TCP_KEEPIDLE, + (char *)&keepidle, sizeof(keepidle)); + if (ret < 0) + goto bail; + + /* KEEPINTVL is the interval between successive probes. We follow + * the model in xs_tcp_finish_connecting() and re-use keepidle. + */ + ret = kernel_setsockopt(sock, IPPROTO_TCP, TCP_KEEPINTVL, + (char *)&keepidle, sizeof(keepidle)); +bail: + return ret; +} + +/* rds_tcp_accept_one_path(): if accepting on cp_index > 0, make sure the + * client's ipaddr < server's ipaddr. Otherwise, close the accepted + * socket and force a reconneect from smaller -> larger ip addr. The reason + * we special case cp_index 0 is to allow the rds probe ping itself to itself + * get through efficiently. + * Since reconnects are only initiated from the node with the numerically + * smaller ip address, we recycle conns in RDS_CONN_ERROR on the passive side + * by moving them to CONNECTING in this function. + */ +static +struct rds_tcp_connection *rds_tcp_accept_one_path(struct rds_connection *conn) +{ + int i; + int npaths = max_t(int, 1, conn->c_npaths); + + /* for mprds, all paths MUST be initiated by the peer + * with the smaller address. + */ + if (rds_addr_cmp(&conn->c_faddr, &conn->c_laddr) >= 0) { + /* Make sure we initiate at least one path if this + * has not already been done; rds_start_mprds() will + * take care of additional paths, if necessary. + */ + if (npaths == 1) + rds_conn_path_connect_if_down(&conn->c_path[0]); + return NULL; + } + + for (i = 0; i < npaths; i++) { + struct rds_conn_path *cp = &conn->c_path[i]; + + if (rds_conn_path_transition(cp, RDS_CONN_DOWN, + RDS_CONN_CONNECTING) || + rds_conn_path_transition(cp, RDS_CONN_ERROR, + RDS_CONN_CONNECTING)) { + return cp->cp_transport_data; + } + } + return NULL; +} + +void rds_tcp_set_linger(struct socket *sock) +{ + struct linger no_linger = { + .l_onoff = 1, + .l_linger = 0, + }; + + kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER, + (char *)&no_linger, sizeof(no_linger)); +} + +int rds_tcp_accept_one(struct socket *sock) +{ + struct socket *new_sock = NULL; + struct rds_connection *conn; + int ret; + struct inet_sock *inet; + struct rds_tcp_connection *rs_tcp = NULL; + int conn_state; + struct rds_conn_path *cp; + struct in6_addr *my_addr, *peer_addr; +#if !IS_ENABLED(CONFIG_IPV6) + struct in6_addr saddr, daddr; +#endif + int dev_if = 0; + + if (!sock) /* module unload or netns delete in progress */ + return -ENETUNREACH; + + ret = sock_create_lite(sock->sk->sk_family, + sock->sk->sk_type, sock->sk->sk_protocol, + &new_sock); + if (ret) + goto out; + + ret = sock->ops->accept(sock, new_sock, O_NONBLOCK, true); + if (ret < 0) + goto out; + + /* sock_create_lite() does not get a hold on the owner module so we + * need to do it here. Note that sock_release() uses sock->ops to + * determine if it needs to decrement the reference count. So set + * sock->ops after calling accept() in case that fails. And there's + * no need to do try_module_get() as the listener should have a hold + * already. + */ + new_sock->ops = sock->ops; + __module_get(new_sock->ops->owner); + + ret = rds_tcp_keepalive(new_sock); + if (ret < 0) + goto out; + + rds_tcp_tune(new_sock); + + inet = inet_sk(new_sock->sk); + +#if IS_ENABLED(CONFIG_IPV6) + my_addr = &new_sock->sk->sk_v6_rcv_saddr; + peer_addr = &new_sock->sk->sk_v6_daddr; +#else + ipv6_addr_set_v4mapped(inet->inet_saddr, &saddr); + ipv6_addr_set_v4mapped(inet->inet_daddr, &daddr); + my_addr = &saddr; + peer_addr = &daddr; +#endif + rdsdebug("accepted family %d tcp %pI6c:%u -> %pI6c:%u\n", + sock->sk->sk_family, + my_addr, ntohs(inet->inet_sport), + peer_addr, ntohs(inet->inet_dport)); + +#if IS_ENABLED(CONFIG_IPV6) + /* sk_bound_dev_if is not set if the peer address is not link local + * address. In this case, it happens that mcast_oif is set. So + * just use it. + */ + if ((ipv6_addr_type(my_addr) & IPV6_ADDR_LINKLOCAL) && + !(ipv6_addr_type(peer_addr) & IPV6_ADDR_LINKLOCAL)) { + struct ipv6_pinfo *inet6; + + inet6 = inet6_sk(new_sock->sk); + dev_if = inet6->mcast_oif; + } else { + dev_if = new_sock->sk->sk_bound_dev_if; + } +#endif + + if (!rds_tcp_laddr_check(sock_net(sock->sk), peer_addr, dev_if)) { + /* local address connection is only allowed via loopback */ + ret = -EOPNOTSUPP; + goto out; + } + + conn = rds_conn_create(sock_net(sock->sk), + my_addr, peer_addr, + &rds_tcp_transport, GFP_KERNEL, dev_if); + + if (IS_ERR(conn)) { + ret = PTR_ERR(conn); + goto out; + } + /* An incoming SYN request came in, and TCP just accepted it. + * + * If the client reboots, this conn will need to be cleaned up. + * rds_tcp_state_change() will do that cleanup + */ + rs_tcp = rds_tcp_accept_one_path(conn); + if (!rs_tcp) + goto rst_nsk; + mutex_lock(&rs_tcp->t_conn_path_lock); + cp = rs_tcp->t_cpath; + conn_state = rds_conn_path_state(cp); + WARN_ON(conn_state == RDS_CONN_UP); + if (conn_state != RDS_CONN_CONNECTING && conn_state != RDS_CONN_ERROR) + goto rst_nsk; + if (rs_tcp->t_sock) { + /* Duelling SYN has been handled in rds_tcp_accept_one() */ + rds_tcp_reset_callbacks(new_sock, cp); + /* rds_connect_path_complete() marks RDS_CONN_UP */ + rds_connect_path_complete(cp, RDS_CONN_RESETTING); + } else { + rds_tcp_set_callbacks(new_sock, cp); + rds_connect_path_complete(cp, RDS_CONN_CONNECTING); + } + new_sock = NULL; + ret = 0; + if (conn->c_npaths == 0) + rds_send_ping(cp->cp_conn, cp->cp_index); + goto out; +rst_nsk: + /* reset the newly returned accept sock and bail. + * It is safe to set linger on new_sock because the RDS connection + * has not been brought up on new_sock, so no RDS-level data could + * be pending on it. By setting linger, we achieve the side-effect + * of avoiding TIME_WAIT state on new_sock. + */ + rds_tcp_set_linger(new_sock); + kernel_sock_shutdown(new_sock, SHUT_RDWR); + ret = 0; +out: + if (rs_tcp) + mutex_unlock(&rs_tcp->t_conn_path_lock); + if (new_sock) + sock_release(new_sock); + return ret; +} + +void rds_tcp_listen_data_ready(struct sock *sk) +{ + void (*ready)(struct sock *sk); + + rdsdebug("listen data ready sk %p\n", sk); + + read_lock_bh(&sk->sk_callback_lock); + ready = sk->sk_user_data; + if (!ready) { /* check for teardown race */ + ready = sk->sk_data_ready; + goto out; + } + + /* + * ->sk_data_ready is also called for a newly established child socket + * before it has been accepted and the accepter has set up their + * data_ready.. we only want to queue listen work for our listening + * socket + * + * (*ready)() may be null if we are racing with netns delete, and + * the listen socket is being torn down. + */ + if (sk->sk_state == TCP_LISTEN) + rds_tcp_accept_work(sk); + else + ready = rds_tcp_listen_sock_def_readable(sock_net(sk)); + +out: + read_unlock_bh(&sk->sk_callback_lock); + if (ready) + ready(sk); +} + +struct socket *rds_tcp_listen_init(struct net *net, bool isv6) +{ + struct socket *sock = NULL; + struct sockaddr_storage ss; + struct sockaddr_in6 *sin6; + struct sockaddr_in *sin; + int addr_len; + int ret; + + ret = sock_create_kern(net, isv6 ? PF_INET6 : PF_INET, SOCK_STREAM, + IPPROTO_TCP, &sock); + if (ret < 0) { + rdsdebug("could not create %s listener socket: %d\n", + isv6 ? "IPv6" : "IPv4", ret); + goto out; + } + + sock->sk->sk_reuse = SK_CAN_REUSE; + rds_tcp_nonagle(sock); + + write_lock_bh(&sock->sk->sk_callback_lock); + sock->sk->sk_user_data = sock->sk->sk_data_ready; + sock->sk->sk_data_ready = rds_tcp_listen_data_ready; + write_unlock_bh(&sock->sk->sk_callback_lock); + + if (isv6) { + sin6 = (struct sockaddr_in6 *)&ss; + sin6->sin6_family = PF_INET6; + sin6->sin6_addr = in6addr_any; + sin6->sin6_port = (__force u16)htons(RDS_TCP_PORT); + sin6->sin6_scope_id = 0; + sin6->sin6_flowinfo = 0; + addr_len = sizeof(*sin6); + } else { + sin = (struct sockaddr_in *)&ss; + sin->sin_family = PF_INET; + sin->sin_addr.s_addr = INADDR_ANY; + sin->sin_port = (__force u16)htons(RDS_TCP_PORT); + addr_len = sizeof(*sin); + } + + ret = sock->ops->bind(sock, (struct sockaddr *)&ss, addr_len); + if (ret < 0) { + rdsdebug("could not bind %s listener socket: %d\n", + isv6 ? "IPv6" : "IPv4", ret); + goto out; + } + + ret = sock->ops->listen(sock, 64); + if (ret < 0) + goto out; + + return sock; +out: + if (sock) + sock_release(sock); + return NULL; +} + +void rds_tcp_listen_stop(struct socket *sock, struct work_struct *acceptor) +{ + struct sock *sk; + + if (!sock) + return; + + sk = sock->sk; + + /* serialize with and prevent further callbacks */ + lock_sock(sk); + write_lock_bh(&sk->sk_callback_lock); + if (sk->sk_user_data) { + sk->sk_data_ready = sk->sk_user_data; + sk->sk_user_data = NULL; + } + write_unlock_bh(&sk->sk_callback_lock); + release_sock(sk); + + /* wait for accepts to stop and close the socket */ + flush_workqueue(rds_wq); + flush_work(acceptor); + sock_release(sock); +} diff --git a/net/rds/tcp_recv.c b/net/rds/tcp_recv.c new file mode 100644 index 000000000..42c5ff1ed --- /dev/null +++ b/net/rds/tcp_recv.c @@ -0,0 +1,349 @@ +/* + * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/slab.h> +#include <net/tcp.h> + +#include "rds.h" +#include "tcp.h" + +static struct kmem_cache *rds_tcp_incoming_slab; + +static void rds_tcp_inc_purge(struct rds_incoming *inc) +{ + struct rds_tcp_incoming *tinc; + tinc = container_of(inc, struct rds_tcp_incoming, ti_inc); + rdsdebug("purging tinc %p inc %p\n", tinc, inc); + skb_queue_purge(&tinc->ti_skb_list); +} + +void rds_tcp_inc_free(struct rds_incoming *inc) +{ + struct rds_tcp_incoming *tinc; + tinc = container_of(inc, struct rds_tcp_incoming, ti_inc); + rds_tcp_inc_purge(inc); + rdsdebug("freeing tinc %p inc %p\n", tinc, inc); + kmem_cache_free(rds_tcp_incoming_slab, tinc); +} + +/* + * this is pretty lame, but, whatever. + */ +int rds_tcp_inc_copy_to_user(struct rds_incoming *inc, struct iov_iter *to) +{ + struct rds_tcp_incoming *tinc; + struct sk_buff *skb; + int ret = 0; + + if (!iov_iter_count(to)) + goto out; + + tinc = container_of(inc, struct rds_tcp_incoming, ti_inc); + + skb_queue_walk(&tinc->ti_skb_list, skb) { + unsigned long to_copy, skb_off; + for (skb_off = 0; skb_off < skb->len; skb_off += to_copy) { + to_copy = iov_iter_count(to); + to_copy = min(to_copy, skb->len - skb_off); + + if (skb_copy_datagram_iter(skb, skb_off, to, to_copy)) + return -EFAULT; + + rds_stats_add(s_copy_to_user, to_copy); + ret += to_copy; + + if (!iov_iter_count(to)) + goto out; + } + } +out: + return ret; +} + +/* + * We have a series of skbs that have fragmented pieces of the congestion + * bitmap. They must add up to the exact size of the congestion bitmap. We + * use the skb helpers to copy those into the pages that make up the in-memory + * congestion bitmap for the remote address of this connection. We then tell + * the congestion core that the bitmap has been changed so that it can wake up + * sleepers. + * + * This is racing with sending paths which are using test_bit to see if the + * bitmap indicates that their recipient is congested. + */ + +static void rds_tcp_cong_recv(struct rds_connection *conn, + struct rds_tcp_incoming *tinc) +{ + struct sk_buff *skb; + unsigned int to_copy, skb_off; + unsigned int map_off; + unsigned int map_page; + struct rds_cong_map *map; + int ret; + + /* catch completely corrupt packets */ + if (be32_to_cpu(tinc->ti_inc.i_hdr.h_len) != RDS_CONG_MAP_BYTES) + return; + + map_page = 0; + map_off = 0; + map = conn->c_fcong; + + skb_queue_walk(&tinc->ti_skb_list, skb) { + skb_off = 0; + while (skb_off < skb->len) { + to_copy = min_t(unsigned int, PAGE_SIZE - map_off, + skb->len - skb_off); + + BUG_ON(map_page >= RDS_CONG_MAP_PAGES); + + /* only returns 0 or -error */ + ret = skb_copy_bits(skb, skb_off, + (void *)map->m_page_addrs[map_page] + map_off, + to_copy); + BUG_ON(ret != 0); + + skb_off += to_copy; + map_off += to_copy; + if (map_off == PAGE_SIZE) { + map_off = 0; + map_page++; + } + } + } + + rds_cong_map_updated(map, ~(u64) 0); +} + +struct rds_tcp_desc_arg { + struct rds_conn_path *conn_path; + gfp_t gfp; +}; + +static int rds_tcp_data_recv(read_descriptor_t *desc, struct sk_buff *skb, + unsigned int offset, size_t len) +{ + struct rds_tcp_desc_arg *arg = desc->arg.data; + struct rds_conn_path *cp = arg->conn_path; + struct rds_tcp_connection *tc = cp->cp_transport_data; + struct rds_tcp_incoming *tinc = tc->t_tinc; + struct sk_buff *clone; + size_t left = len, to_copy; + + rdsdebug("tcp data tc %p skb %p offset %u len %zu\n", tc, skb, offset, + len); + + /* + * tcp_read_sock() interprets partial progress as an indication to stop + * processing. + */ + while (left) { + if (!tinc) { + tinc = kmem_cache_alloc(rds_tcp_incoming_slab, + arg->gfp); + if (!tinc) { + desc->error = -ENOMEM; + goto out; + } + tc->t_tinc = tinc; + rdsdebug("alloced tinc %p\n", tinc); + rds_inc_path_init(&tinc->ti_inc, cp, + &cp->cp_conn->c_faddr); + tinc->ti_inc.i_rx_lat_trace[RDS_MSG_RX_HDR] = + local_clock(); + + /* + * XXX * we might be able to use the __ variants when + * we've already serialized at a higher level. + */ + skb_queue_head_init(&tinc->ti_skb_list); + } + + if (left && tc->t_tinc_hdr_rem) { + to_copy = min(tc->t_tinc_hdr_rem, left); + rdsdebug("copying %zu header from skb %p\n", to_copy, + skb); + skb_copy_bits(skb, offset, + (char *)&tinc->ti_inc.i_hdr + + sizeof(struct rds_header) - + tc->t_tinc_hdr_rem, + to_copy); + tc->t_tinc_hdr_rem -= to_copy; + left -= to_copy; + offset += to_copy; + + if (tc->t_tinc_hdr_rem == 0) { + /* could be 0 for a 0 len message */ + tc->t_tinc_data_rem = + be32_to_cpu(tinc->ti_inc.i_hdr.h_len); + tinc->ti_inc.i_rx_lat_trace[RDS_MSG_RX_START] = + local_clock(); + } + } + + if (left && tc->t_tinc_data_rem) { + to_copy = min(tc->t_tinc_data_rem, left); + + clone = pskb_extract(skb, offset, to_copy, arg->gfp); + if (!clone) { + desc->error = -ENOMEM; + goto out; + } + + skb_queue_tail(&tinc->ti_skb_list, clone); + + rdsdebug("skb %p data %p len %d off %u to_copy %zu -> " + "clone %p data %p len %d\n", + skb, skb->data, skb->len, offset, to_copy, + clone, clone->data, clone->len); + + tc->t_tinc_data_rem -= to_copy; + left -= to_copy; + offset += to_copy; + } + + if (tc->t_tinc_hdr_rem == 0 && tc->t_tinc_data_rem == 0) { + struct rds_connection *conn = cp->cp_conn; + + if (tinc->ti_inc.i_hdr.h_flags == RDS_FLAG_CONG_BITMAP) + rds_tcp_cong_recv(conn, tinc); + else + rds_recv_incoming(conn, &conn->c_faddr, + &conn->c_laddr, + &tinc->ti_inc, + arg->gfp); + + tc->t_tinc_hdr_rem = sizeof(struct rds_header); + tc->t_tinc_data_rem = 0; + tc->t_tinc = NULL; + rds_inc_put(&tinc->ti_inc); + tinc = NULL; + } + } +out: + rdsdebug("returning len %zu left %zu skb len %d rx queue depth %d\n", + len, left, skb->len, + skb_queue_len(&tc->t_sock->sk->sk_receive_queue)); + return len - left; +} + +/* the caller has to hold the sock lock */ +static int rds_tcp_read_sock(struct rds_conn_path *cp, gfp_t gfp) +{ + struct rds_tcp_connection *tc = cp->cp_transport_data; + struct socket *sock = tc->t_sock; + read_descriptor_t desc; + struct rds_tcp_desc_arg arg; + + /* It's like glib in the kernel! */ + arg.conn_path = cp; + arg.gfp = gfp; + desc.arg.data = &arg; + desc.error = 0; + desc.count = 1; /* give more than one skb per call */ + + tcp_read_sock(sock->sk, &desc, rds_tcp_data_recv); + rdsdebug("tcp_read_sock for tc %p gfp 0x%x returned %d\n", tc, gfp, + desc.error); + + return desc.error; +} + +/* + * We hold the sock lock to serialize our rds_tcp_recv->tcp_read_sock from + * data_ready. + * + * if we fail to allocate we're in trouble.. blindly wait some time before + * trying again to see if the VM can free up something for us. + */ +int rds_tcp_recv_path(struct rds_conn_path *cp) +{ + struct rds_tcp_connection *tc = cp->cp_transport_data; + struct socket *sock = tc->t_sock; + int ret = 0; + + rdsdebug("recv worker path [%d] tc %p sock %p\n", + cp->cp_index, tc, sock); + + lock_sock(sock->sk); + ret = rds_tcp_read_sock(cp, GFP_KERNEL); + release_sock(sock->sk); + + return ret; +} + +void rds_tcp_data_ready(struct sock *sk) +{ + void (*ready)(struct sock *sk); + struct rds_conn_path *cp; + struct rds_tcp_connection *tc; + + rdsdebug("data ready sk %p\n", sk); + + read_lock_bh(&sk->sk_callback_lock); + cp = sk->sk_user_data; + if (!cp) { /* check for teardown race */ + ready = sk->sk_data_ready; + goto out; + } + + tc = cp->cp_transport_data; + ready = tc->t_orig_data_ready; + rds_tcp_stats_inc(s_tcp_data_ready_calls); + + if (rds_tcp_read_sock(cp, GFP_ATOMIC) == -ENOMEM) { + rcu_read_lock(); + if (!rds_destroy_pending(cp->cp_conn)) + queue_delayed_work(rds_wq, &cp->cp_recv_w, 0); + rcu_read_unlock(); + } +out: + read_unlock_bh(&sk->sk_callback_lock); + ready(sk); +} + +int rds_tcp_recv_init(void) +{ + rds_tcp_incoming_slab = kmem_cache_create("rds_tcp_incoming", + sizeof(struct rds_tcp_incoming), + 0, 0, NULL); + if (!rds_tcp_incoming_slab) + return -ENOMEM; + return 0; +} + +void rds_tcp_recv_exit(void) +{ + kmem_cache_destroy(rds_tcp_incoming_slab); +} diff --git a/net/rds/tcp_send.c b/net/rds/tcp_send.c new file mode 100644 index 000000000..78a2554a4 --- /dev/null +++ b/net/rds/tcp_send.c @@ -0,0 +1,230 @@ +/* + * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/in.h> +#include <net/tcp.h> + +#include "rds_single_path.h" +#include "rds.h" +#include "tcp.h" + +static void rds_tcp_cork(struct socket *sock, int val) +{ + kernel_setsockopt(sock, SOL_TCP, TCP_CORK, (void *)&val, sizeof(val)); +} + +void rds_tcp_xmit_path_prepare(struct rds_conn_path *cp) +{ + struct rds_tcp_connection *tc = cp->cp_transport_data; + + rds_tcp_cork(tc->t_sock, 1); +} + +void rds_tcp_xmit_path_complete(struct rds_conn_path *cp) +{ + struct rds_tcp_connection *tc = cp->cp_transport_data; + + rds_tcp_cork(tc->t_sock, 0); +} + +/* the core send_sem serializes this with other xmit and shutdown */ +static int rds_tcp_sendmsg(struct socket *sock, void *data, unsigned int len) +{ + struct kvec vec = { + .iov_base = data, + .iov_len = len, + }; + struct msghdr msg = { + .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL, + }; + + return kernel_sendmsg(sock, &msg, &vec, 1, vec.iov_len); +} + +/* the core send_sem serializes this with other xmit and shutdown */ +int rds_tcp_xmit(struct rds_connection *conn, struct rds_message *rm, + unsigned int hdr_off, unsigned int sg, unsigned int off) +{ + struct rds_conn_path *cp = rm->m_inc.i_conn_path; + struct rds_tcp_connection *tc = cp->cp_transport_data; + int done = 0; + int ret = 0; + int more; + + if (hdr_off == 0) { + /* + * m_ack_seq is set to the sequence number of the last byte of + * header and data. see rds_tcp_is_acked(). + */ + tc->t_last_sent_nxt = rds_tcp_write_seq(tc); + rm->m_ack_seq = tc->t_last_sent_nxt + + sizeof(struct rds_header) + + be32_to_cpu(rm->m_inc.i_hdr.h_len) - 1; + smp_mb__before_atomic(); + set_bit(RDS_MSG_HAS_ACK_SEQ, &rm->m_flags); + tc->t_last_expected_una = rm->m_ack_seq + 1; + + if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags)) + rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED; + + rdsdebug("rm %p tcp nxt %u ack_seq %llu\n", + rm, rds_tcp_write_seq(tc), + (unsigned long long)rm->m_ack_seq); + } + + if (hdr_off < sizeof(struct rds_header)) { + /* see rds_tcp_write_space() */ + set_bit(SOCK_NOSPACE, &tc->t_sock->sk->sk_socket->flags); + + ret = rds_tcp_sendmsg(tc->t_sock, + (void *)&rm->m_inc.i_hdr + hdr_off, + sizeof(rm->m_inc.i_hdr) - hdr_off); + if (ret < 0) + goto out; + done += ret; + if (hdr_off + done != sizeof(struct rds_header)) + goto out; + } + + more = rm->data.op_nents > 1 ? (MSG_MORE | MSG_SENDPAGE_NOTLAST) : 0; + while (sg < rm->data.op_nents) { + int flags = MSG_DONTWAIT | MSG_NOSIGNAL | more; + + ret = tc->t_sock->ops->sendpage(tc->t_sock, + sg_page(&rm->data.op_sg[sg]), + rm->data.op_sg[sg].offset + off, + rm->data.op_sg[sg].length - off, + flags); + rdsdebug("tcp sendpage %p:%u:%u ret %d\n", (void *)sg_page(&rm->data.op_sg[sg]), + rm->data.op_sg[sg].offset + off, rm->data.op_sg[sg].length - off, + ret); + if (ret <= 0) + break; + + off += ret; + done += ret; + if (off == rm->data.op_sg[sg].length) { + off = 0; + sg++; + } + if (sg == rm->data.op_nents - 1) + more = 0; + } + +out: + if (ret <= 0) { + /* write_space will hit after EAGAIN, all else fatal */ + if (ret == -EAGAIN) { + rds_tcp_stats_inc(s_tcp_sndbuf_full); + ret = 0; + } else { + /* No need to disconnect/reconnect if path_drop + * has already been triggered, because, e.g., of + * an incoming RST. + */ + if (rds_conn_path_up(cp)) { + pr_warn("RDS/tcp: send to %pI6c on cp [%d]" + "returned %d, " + "disconnecting and reconnecting\n", + &conn->c_faddr, cp->cp_index, ret); + rds_conn_path_drop(cp, false); + } + } + } + if (done == 0) + done = ret; + return done; +} + +/* + * rm->m_ack_seq is set to the tcp sequence number that corresponds to the + * last byte of the message, including the header. This means that the + * entire message has been received if rm->m_ack_seq is "before" the next + * unacked byte of the TCP sequence space. We have to do very careful + * wrapping 32bit comparisons here. + */ +static int rds_tcp_is_acked(struct rds_message *rm, uint64_t ack) +{ + if (!test_bit(RDS_MSG_HAS_ACK_SEQ, &rm->m_flags)) + return 0; + return (__s32)((u32)rm->m_ack_seq - (u32)ack) < 0; +} + +void rds_tcp_write_space(struct sock *sk) +{ + void (*write_space)(struct sock *sk); + struct rds_conn_path *cp; + struct rds_tcp_connection *tc; + + read_lock_bh(&sk->sk_callback_lock); + cp = sk->sk_user_data; + if (!cp) { + write_space = sk->sk_write_space; + goto out; + } + + tc = cp->cp_transport_data; + rdsdebug("write_space for tc %p\n", tc); + write_space = tc->t_orig_write_space; + rds_tcp_stats_inc(s_tcp_write_space_calls); + + rdsdebug("tcp una %u\n", rds_tcp_snd_una(tc)); + tc->t_last_seen_una = rds_tcp_snd_una(tc); + rds_send_path_drop_acked(cp, rds_tcp_snd_una(tc), rds_tcp_is_acked); + + rcu_read_lock(); + if ((refcount_read(&sk->sk_wmem_alloc) << 1) <= sk->sk_sndbuf && + !rds_destroy_pending(cp->cp_conn)) + queue_delayed_work(rds_wq, &cp->cp_send_w, 0); + rcu_read_unlock(); + +out: + read_unlock_bh(&sk->sk_callback_lock); + + /* + * write_space is only called when data leaves tcp's send queue if + * SOCK_NOSPACE is set. We set SOCK_NOSPACE every time we put + * data in tcp's send queue because we use write_space to parse the + * sequence numbers and notice that rds messages have been fully + * received. + * + * tcp's write_space clears SOCK_NOSPACE if the send queue has more + * than a certain amount of space. So we need to set it again *after* + * we call tcp's write_space or else we might only get called on the + * first of a series of incoming tcp acks. + */ + write_space(sk); + + if (sk->sk_socket) + set_bit(SOCK_NOSPACE, &sk->sk_socket->flags); +} diff --git a/net/rds/tcp_stats.c b/net/rds/tcp_stats.c new file mode 100644 index 000000000..f8a7954f1 --- /dev/null +++ b/net/rds/tcp_stats.c @@ -0,0 +1,74 @@ +/* + * Copyright (c) 2006 Oracle. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/percpu.h> +#include <linux/seq_file.h> +#include <linux/proc_fs.h> + +#include "rds.h" +#include "tcp.h" + +DEFINE_PER_CPU(struct rds_tcp_statistics, rds_tcp_stats) + ____cacheline_aligned; + +static const char * const rds_tcp_stat_names[] = { + "tcp_data_ready_calls", + "tcp_write_space_calls", + "tcp_sndbuf_full", + "tcp_connect_raced", + "tcp_listen_closed_stale", +}; + +unsigned int rds_tcp_stats_info_copy(struct rds_info_iterator *iter, + unsigned int avail) +{ + struct rds_tcp_statistics stats = {0, }; + uint64_t *src; + uint64_t *sum; + size_t i; + int cpu; + + if (avail < ARRAY_SIZE(rds_tcp_stat_names)) + goto out; + + for_each_online_cpu(cpu) { + src = (uint64_t *)&(per_cpu(rds_tcp_stats, cpu)); + sum = (uint64_t *)&stats; + for (i = 0; i < sizeof(stats) / sizeof(uint64_t); i++) + *(sum++) += *(src++); + } + + rds_stats_info_copy(iter, (uint64_t *)&stats, rds_tcp_stat_names, + ARRAY_SIZE(rds_tcp_stat_names)); +out: + return ARRAY_SIZE(rds_tcp_stat_names); +} diff --git a/net/rds/threads.c b/net/rds/threads.c new file mode 100644 index 000000000..e64f9e4c3 --- /dev/null +++ b/net/rds/threads.c @@ -0,0 +1,308 @@ +/* + * Copyright (c) 2006, 2018 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/random.h> +#include <linux/export.h> + +#include "rds.h" + +/* + * All of connection management is simplified by serializing it through + * work queues that execute in a connection managing thread. + * + * TCP wants to send acks through sendpage() in response to data_ready(), + * but it needs a process context to do so. + * + * The receive paths need to allocate but can't drop packets (!) so we have + * a thread around to block allocating if the receive fast path sees an + * allocation failure. + */ + +/* Grand Unified Theory of connection life cycle: + * At any point in time, the connection can be in one of these states: + * DOWN, CONNECTING, UP, DISCONNECTING, ERROR + * + * The following transitions are possible: + * ANY -> ERROR + * UP -> DISCONNECTING + * ERROR -> DISCONNECTING + * DISCONNECTING -> DOWN + * DOWN -> CONNECTING + * CONNECTING -> UP + * + * Transition to state DISCONNECTING/DOWN: + * - Inside the shutdown worker; synchronizes with xmit path + * through RDS_IN_XMIT, and with connection management callbacks + * via c_cm_lock. + * + * For receive callbacks, we rely on the underlying transport + * (TCP, IB/RDMA) to provide the necessary synchronisation. + */ +struct workqueue_struct *rds_wq; +EXPORT_SYMBOL_GPL(rds_wq); + +void rds_connect_path_complete(struct rds_conn_path *cp, int curr) +{ + if (!rds_conn_path_transition(cp, curr, RDS_CONN_UP)) { + printk(KERN_WARNING "%s: Cannot transition to state UP, " + "current state is %d\n", + __func__, + atomic_read(&cp->cp_state)); + rds_conn_path_drop(cp, false); + return; + } + + rdsdebug("conn %p for %pI6c to %pI6c complete\n", + cp->cp_conn, &cp->cp_conn->c_laddr, &cp->cp_conn->c_faddr); + + cp->cp_reconnect_jiffies = 0; + set_bit(0, &cp->cp_conn->c_map_queued); + rcu_read_lock(); + if (!rds_destroy_pending(cp->cp_conn)) { + queue_delayed_work(rds_wq, &cp->cp_send_w, 0); + queue_delayed_work(rds_wq, &cp->cp_recv_w, 0); + } + rcu_read_unlock(); +} +EXPORT_SYMBOL_GPL(rds_connect_path_complete); + +void rds_connect_complete(struct rds_connection *conn) +{ + rds_connect_path_complete(&conn->c_path[0], RDS_CONN_CONNECTING); +} +EXPORT_SYMBOL_GPL(rds_connect_complete); + +/* + * This random exponential backoff is relied on to eventually resolve racing + * connects. + * + * If connect attempts race then both parties drop both connections and come + * here to wait for a random amount of time before trying again. Eventually + * the backoff range will be so much greater than the time it takes to + * establish a connection that one of the pair will establish the connection + * before the other's random delay fires. + * + * Connection attempts that arrive while a connection is already established + * are also considered to be racing connects. This lets a connection from + * a rebooted machine replace an existing stale connection before the transport + * notices that the connection has failed. + * + * We should *always* start with a random backoff; otherwise a broken connection + * will always take several iterations to be re-established. + */ +void rds_queue_reconnect(struct rds_conn_path *cp) +{ + unsigned long rand; + struct rds_connection *conn = cp->cp_conn; + + rdsdebug("conn %p for %pI6c to %pI6c reconnect jiffies %lu\n", + conn, &conn->c_laddr, &conn->c_faddr, + cp->cp_reconnect_jiffies); + + /* let peer with smaller addr initiate reconnect, to avoid duels */ + if (conn->c_trans->t_type == RDS_TRANS_TCP && + rds_addr_cmp(&conn->c_laddr, &conn->c_faddr) >= 0) + return; + + set_bit(RDS_RECONNECT_PENDING, &cp->cp_flags); + if (cp->cp_reconnect_jiffies == 0) { + cp->cp_reconnect_jiffies = rds_sysctl_reconnect_min_jiffies; + rcu_read_lock(); + if (!rds_destroy_pending(cp->cp_conn)) + queue_delayed_work(rds_wq, &cp->cp_conn_w, 0); + rcu_read_unlock(); + return; + } + + get_random_bytes(&rand, sizeof(rand)); + rdsdebug("%lu delay %lu ceil conn %p for %pI6c -> %pI6c\n", + rand % cp->cp_reconnect_jiffies, cp->cp_reconnect_jiffies, + conn, &conn->c_laddr, &conn->c_faddr); + rcu_read_lock(); + if (!rds_destroy_pending(cp->cp_conn)) + queue_delayed_work(rds_wq, &cp->cp_conn_w, + rand % cp->cp_reconnect_jiffies); + rcu_read_unlock(); + + cp->cp_reconnect_jiffies = min(cp->cp_reconnect_jiffies * 2, + rds_sysctl_reconnect_max_jiffies); +} + +void rds_connect_worker(struct work_struct *work) +{ + struct rds_conn_path *cp = container_of(work, + struct rds_conn_path, + cp_conn_w.work); + struct rds_connection *conn = cp->cp_conn; + int ret; + + if (cp->cp_index > 0 && + rds_addr_cmp(&cp->cp_conn->c_laddr, &cp->cp_conn->c_faddr) >= 0) + return; + clear_bit(RDS_RECONNECT_PENDING, &cp->cp_flags); + ret = rds_conn_path_transition(cp, RDS_CONN_DOWN, RDS_CONN_CONNECTING); + if (ret) { + ret = conn->c_trans->conn_path_connect(cp); + rdsdebug("conn %p for %pI6c to %pI6c dispatched, ret %d\n", + conn, &conn->c_laddr, &conn->c_faddr, ret); + + if (ret) { + if (rds_conn_path_transition(cp, + RDS_CONN_CONNECTING, + RDS_CONN_DOWN)) + rds_queue_reconnect(cp); + else + rds_conn_path_error(cp, "connect failed\n"); + } + } +} + +void rds_send_worker(struct work_struct *work) +{ + struct rds_conn_path *cp = container_of(work, + struct rds_conn_path, + cp_send_w.work); + int ret; + + if (rds_conn_path_state(cp) == RDS_CONN_UP) { + clear_bit(RDS_LL_SEND_FULL, &cp->cp_flags); + ret = rds_send_xmit(cp); + cond_resched(); + rdsdebug("conn %p ret %d\n", cp->cp_conn, ret); + switch (ret) { + case -EAGAIN: + rds_stats_inc(s_send_immediate_retry); + queue_delayed_work(rds_wq, &cp->cp_send_w, 0); + break; + case -ENOMEM: + rds_stats_inc(s_send_delayed_retry); + queue_delayed_work(rds_wq, &cp->cp_send_w, 2); + default: + break; + } + } +} + +void rds_recv_worker(struct work_struct *work) +{ + struct rds_conn_path *cp = container_of(work, + struct rds_conn_path, + cp_recv_w.work); + int ret; + + if (rds_conn_path_state(cp) == RDS_CONN_UP) { + ret = cp->cp_conn->c_trans->recv_path(cp); + rdsdebug("conn %p ret %d\n", cp->cp_conn, ret); + switch (ret) { + case -EAGAIN: + rds_stats_inc(s_recv_immediate_retry); + queue_delayed_work(rds_wq, &cp->cp_recv_w, 0); + break; + case -ENOMEM: + rds_stats_inc(s_recv_delayed_retry); + queue_delayed_work(rds_wq, &cp->cp_recv_w, 2); + default: + break; + } + } +} + +void rds_shutdown_worker(struct work_struct *work) +{ + struct rds_conn_path *cp = container_of(work, + struct rds_conn_path, + cp_down_w); + + rds_conn_shutdown(cp); +} + +void rds_threads_exit(void) +{ + destroy_workqueue(rds_wq); +} + +int rds_threads_init(void) +{ + rds_wq = create_singlethread_workqueue("krdsd"); + if (!rds_wq) + return -ENOMEM; + + return 0; +} + +/* Compare two IPv6 addresses. Return 0 if the two addresses are equal. + * Return 1 if the first is greater. Return -1 if the second is greater. + */ +int rds_addr_cmp(const struct in6_addr *addr1, + const struct in6_addr *addr2) +{ +#if defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS) && BITS_PER_LONG == 64 + const __be64 *a1, *a2; + u64 x, y; + + a1 = (__be64 *)addr1; + a2 = (__be64 *)addr2; + + if (*a1 != *a2) { + if (be64_to_cpu(*a1) < be64_to_cpu(*a2)) + return -1; + else + return 1; + } else { + x = be64_to_cpu(*++a1); + y = be64_to_cpu(*++a2); + if (x < y) + return -1; + else if (x > y) + return 1; + else + return 0; + } +#else + u32 a, b; + int i; + + for (i = 0; i < 4; i++) { + if (addr1->s6_addr32[i] != addr2->s6_addr32[i]) { + a = ntohl(addr1->s6_addr32[i]); + b = ntohl(addr2->s6_addr32[i]); + if (a < b) + return -1; + else if (a > b) + return 1; + } + } + return 0; +#endif +} +EXPORT_SYMBOL_GPL(rds_addr_cmp); diff --git a/net/rds/transport.c b/net/rds/transport.c new file mode 100644 index 000000000..46f709a4b --- /dev/null +++ b/net/rds/transport.c @@ -0,0 +1,161 @@ +/* + * Copyright (c) 2006, 2017 Oracle and/or its affiliates. All rights reserved. + * + * This software is available to you under a choice of one of two + * licenses. You may choose to be licensed under the terms of the GNU + * General Public License (GPL) Version 2, available from the file + * COPYING in the main directory of this source tree, or the + * OpenIB.org BSD license below: + * + * Redistribution and use in source and binary forms, with or + * without modification, are permitted provided that the following + * conditions are met: + * + * - Redistributions of source code must retain the above + * copyright notice, this list of conditions and the following + * disclaimer. + * + * - Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimer in the documentation and/or other materials + * provided with the distribution. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF + * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS + * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN + * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN + * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + * + */ +#include <linux/kernel.h> +#include <linux/module.h> +#include <linux/in.h> +#include <linux/ipv6.h> + +#include "rds.h" +#include "loop.h" + +static struct rds_transport *transports[RDS_TRANS_COUNT]; +static DECLARE_RWSEM(rds_trans_sem); + +void rds_trans_register(struct rds_transport *trans) +{ + BUG_ON(strlen(trans->t_name) + 1 > TRANSNAMSIZ); + + down_write(&rds_trans_sem); + + if (transports[trans->t_type]) + printk(KERN_ERR "RDS Transport type %d already registered\n", + trans->t_type); + else { + transports[trans->t_type] = trans; + printk(KERN_INFO "Registered RDS/%s transport\n", trans->t_name); + } + + up_write(&rds_trans_sem); +} +EXPORT_SYMBOL_GPL(rds_trans_register); + +void rds_trans_unregister(struct rds_transport *trans) +{ + down_write(&rds_trans_sem); + + transports[trans->t_type] = NULL; + printk(KERN_INFO "Unregistered RDS/%s transport\n", trans->t_name); + + up_write(&rds_trans_sem); +} +EXPORT_SYMBOL_GPL(rds_trans_unregister); + +void rds_trans_put(struct rds_transport *trans) +{ + if (trans) + module_put(trans->t_owner); +} + +struct rds_transport *rds_trans_get_preferred(struct net *net, + const struct in6_addr *addr, + __u32 scope_id) +{ + struct rds_transport *ret = NULL; + struct rds_transport *trans; + unsigned int i; + + if (ipv6_addr_v4mapped(addr)) { + if (*(u_int8_t *)&addr->s6_addr32[3] == IN_LOOPBACKNET) + return &rds_loop_transport; + } else if (ipv6_addr_loopback(addr)) { + return &rds_loop_transport; + } + + down_read(&rds_trans_sem); + for (i = 0; i < RDS_TRANS_COUNT; i++) { + trans = transports[i]; + + if (trans && (trans->laddr_check(net, addr, scope_id) == 0) && + (!trans->t_owner || try_module_get(trans->t_owner))) { + ret = trans; + break; + } + } + up_read(&rds_trans_sem); + + return ret; +} + +struct rds_transport *rds_trans_get(int t_type) +{ + struct rds_transport *ret = NULL; + struct rds_transport *trans; + unsigned int i; + + down_read(&rds_trans_sem); + for (i = 0; i < RDS_TRANS_COUNT; i++) { + trans = transports[i]; + + if (trans && trans->t_type == t_type && + (!trans->t_owner || try_module_get(trans->t_owner))) { + ret = trans; + break; + } + } + up_read(&rds_trans_sem); + + return ret; +} + +/* + * This returns the number of stats entries in the snapshot and only + * copies them using the iter if there is enough space for them. The + * caller passes in the global stats so that we can size and copy while + * holding the lock. + */ +unsigned int rds_trans_stats_info_copy(struct rds_info_iterator *iter, + unsigned int avail) + +{ + struct rds_transport *trans; + unsigned int total = 0; + unsigned int part; + int i; + + rds_info_iter_unmap(iter); + down_read(&rds_trans_sem); + + for (i = 0; i < RDS_TRANS_COUNT; i++) { + trans = transports[i]; + if (!trans || !trans->stats_info_copy) + continue; + + part = trans->stats_info_copy(iter, avail); + avail -= min(avail, part); + total += part; + } + + up_read(&rds_trans_sem); + + return total; +} |