diff options
author | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-27 10:05:51 +0000 |
---|---|---|
committer | Daniel Baumann <daniel.baumann@progress-linux.org> | 2024-04-27 10:05:51 +0000 |
commit | 5d1646d90e1f2cceb9f0828f4b28318cd0ec7744 (patch) | |
tree | a94efe259b9009378be6d90eb30d2b019d95c194 /Documentation/infiniband | |
parent | Initial commit. (diff) | |
download | linux-upstream.tar.xz linux-upstream.zip |
Adding upstream version 5.10.209.upstream/5.10.209upstream
Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>
Diffstat (limited to 'Documentation/infiniband')
-rw-r--r-- | Documentation/infiniband/core_locking.rst | 116 | ||||
-rw-r--r-- | Documentation/infiniband/index.rst | 23 | ||||
-rw-r--r-- | Documentation/infiniband/ipoib.rst | 115 | ||||
-rw-r--r-- | Documentation/infiniband/opa_vnic.rst | 159 | ||||
-rw-r--r-- | Documentation/infiniband/sysfs.rst | 6 | ||||
-rw-r--r-- | Documentation/infiniband/tag_matching.rst | 69 | ||||
-rw-r--r-- | Documentation/infiniband/user_mad.rst | 166 | ||||
-rw-r--r-- | Documentation/infiniband/user_verbs.rst | 75 |
8 files changed, 729 insertions, 0 deletions
diff --git a/Documentation/infiniband/core_locking.rst b/Documentation/infiniband/core_locking.rst new file mode 100644 index 000000000..efd5e7603 --- /dev/null +++ b/Documentation/infiniband/core_locking.rst @@ -0,0 +1,116 @@ +=========================== +InfiniBand Midlayer Locking +=========================== + + This guide is an attempt to make explicit the locking assumptions + made by the InfiniBand midlayer. It describes the requirements on + both low-level drivers that sit below the midlayer and upper level + protocols that use the midlayer. + +Sleeping and interrupt context +============================== + + With the following exceptions, a low-level driver implementation of + all of the methods in struct ib_device may sleep. The exceptions + are any methods from the list: + + - create_ah + - modify_ah + - query_ah + - destroy_ah + - post_send + - post_recv + - poll_cq + - req_notify_cq + + which may not sleep and must be callable from any context. + + The corresponding functions exported to upper level protocol + consumers: + + - rdma_create_ah + - rdma_modify_ah + - rdma_query_ah + - rdma_destroy_ah + - ib_post_send + - ib_post_recv + - ib_req_notify_cq + + are therefore safe to call from any context. + + In addition, the function + + - ib_dispatch_event + + used by low-level drivers to dispatch asynchronous events through + the midlayer is also safe to call from any context. + +Reentrancy +---------- + + All of the methods in struct ib_device exported by a low-level + driver must be fully reentrant. The low-level driver is required to + perform all synchronization necessary to maintain consistency, even + if multiple function calls using the same object are run + simultaneously. + + The IB midlayer does not perform any serialization of function calls. + + Because low-level drivers are reentrant, upper level protocol + consumers are not required to perform any serialization. However, + some serialization may be required to get sensible results. For + example, a consumer may safely call ib_poll_cq() on multiple CPUs + simultaneously. However, the ordering of the work completion + information between different calls of ib_poll_cq() is not defined. + +Callbacks +--------- + + A low-level driver must not perform a callback directly from the + same callchain as an ib_device method call. For example, it is not + allowed for a low-level driver to call a consumer's completion event + handler directly from its post_send method. Instead, the low-level + driver should defer this callback by, for example, scheduling a + tasklet to perform the callback. + + The low-level driver is responsible for ensuring that multiple + completion event handlers for the same CQ are not called + simultaneously. The driver must guarantee that only one CQ event + handler for a given CQ is running at a time. In other words, the + following situation is not allowed:: + + CPU1 CPU2 + + low-level driver -> + consumer CQ event callback: + /* ... */ + ib_req_notify_cq(cq, ...); + low-level driver -> + /* ... */ consumer CQ event callback: + /* ... */ + return from CQ event handler + + The context in which completion event and asynchronous event + callbacks run is not defined. Depending on the low-level driver, it + may be process context, softirq context, or interrupt context. + Upper level protocol consumers may not sleep in a callback. + +Hot-plug +-------- + + A low-level driver announces that a device is ready for use by + consumers when it calls ib_register_device(), all initialization + must be complete before this call. The device must remain usable + until the driver's call to ib_unregister_device() has returned. + + A low-level driver must call ib_register_device() and + ib_unregister_device() from process context. It must not hold any + semaphores that could cause deadlock if a consumer calls back into + the driver across these calls. + + An upper level protocol consumer may begin using an IB device as + soon as the add method of its struct ib_client is called for that + device. A consumer must finish all cleanup and free all resources + relating to a device before returning from the remove method. + + A consumer is permitted to sleep in its add and remove methods. diff --git a/Documentation/infiniband/index.rst b/Documentation/infiniband/index.rst new file mode 100644 index 000000000..9cd761543 --- /dev/null +++ b/Documentation/infiniband/index.rst @@ -0,0 +1,23 @@ +.. SPDX-License-Identifier: GPL-2.0 + +========== +InfiniBand +========== + +.. toctree:: + :maxdepth: 1 + + core_locking + ipoib + opa_vnic + sysfs + tag_matching + user_mad + user_verbs + +.. only:: subproject and html + + Indices + ======= + + * :ref:`genindex` diff --git a/Documentation/infiniband/ipoib.rst b/Documentation/infiniband/ipoib.rst new file mode 100644 index 000000000..0dd36154c --- /dev/null +++ b/Documentation/infiniband/ipoib.rst @@ -0,0 +1,115 @@ +================== +IP over InfiniBand +================== + + The ib_ipoib driver is an implementation of the IP over InfiniBand + protocol as specified by RFC 4391 and 4392, issued by the IETF ipoib + working group. It is a "native" implementation in the sense of + setting the interface type to ARPHRD_INFINIBAND and the hardware + address length to 20 (earlier proprietary implementations + masqueraded to the kernel as ethernet interfaces). + +Partitions and P_Keys +===================== + + When the IPoIB driver is loaded, it creates one interface for each + port using the P_Key at index 0. To create an interface with a + different P_Key, write the desired P_Key into the main interface's + /sys/class/net/<intf name>/create_child file. For example:: + + echo 0x8001 > /sys/class/net/ib0/create_child + + This will create an interface named ib0.8001 with P_Key 0x8001. To + remove a subinterface, use the "delete_child" file:: + + echo 0x8001 > /sys/class/net/ib0/delete_child + + The P_Key for any interface is given by the "pkey" file, and the + main interface for a subinterface is in "parent." + + Child interface create/delete can also be done using IPoIB's + rtnl_link_ops, where children created using either way behave the same. + +Datagram vs Connected modes +=========================== + + The IPoIB driver supports two modes of operation: datagram and + connected. The mode is set and read through an interface's + /sys/class/net/<intf name>/mode file. + + In datagram mode, the IB UD (Unreliable Datagram) transport is used + and so the interface MTU has is equal to the IB L2 MTU minus the + IPoIB encapsulation header (4 bytes). For example, in a typical IB + fabric with a 2K MTU, the IPoIB MTU will be 2048 - 4 = 2044 bytes. + + In connected mode, the IB RC (Reliable Connected) transport is used. + Connected mode takes advantage of the connected nature of the IB + transport and allows an MTU up to the maximal IP packet size of 64K, + which reduces the number of IP packets needed for handling large UDP + datagrams, TCP segments, etc and increases the performance for large + messages. + + In connected mode, the interface's UD QP is still used for multicast + and communication with peers that don't support connected mode. In + this case, RX emulation of ICMP PMTU packets is used to cause the + networking stack to use the smaller UD MTU for these neighbours. + +Stateless offloads +================== + + If the IB HW supports IPoIB stateless offloads, IPoIB advertises + TCP/IP checksum and/or Large Send (LSO) offloading capability to the + network stack. + + Large Receive (LRO) offloading is also implemented and may be turned + on/off using ethtool calls. Currently LRO is supported only for + checksum offload capable devices. + + Stateless offloads are supported only in datagram mode. + +Interrupt moderation +==================== + + If the underlying IB device supports CQ event moderation, one can + use ethtool to set interrupt mitigation parameters and thus reduce + the overhead incurred by handling interrupts. The main code path of + IPoIB doesn't use events for TX completion signaling so only RX + moderation is supported. + +Debugging Information +===================== + + By compiling the IPoIB driver with CONFIG_INFINIBAND_IPOIB_DEBUG set + to 'y', tracing messages are compiled into the driver. They are + turned on by setting the module parameters debug_level and + mcast_debug_level to 1. These parameters can be controlled at + runtime through files in /sys/module/ib_ipoib/. + + CONFIG_INFINIBAND_IPOIB_DEBUG also enables files in the debugfs + virtual filesystem. By mounting this filesystem, for example with:: + + mount -t debugfs none /sys/kernel/debug + + it is possible to get statistics about multicast groups from the + files /sys/kernel/debug/ipoib/ib0_mcg and so on. + + The performance impact of this option is negligible, so it + is safe to enable this option with debug_level set to 0 for normal + operation. + + CONFIG_INFINIBAND_IPOIB_DEBUG_DATA enables even more debug output in + the data path when data_debug_level is set to 1. However, even with + the output disabled, enabling this configuration option will affect + performance, because it adds tests to the fast path. + +References +========== + + Transmission of IP over InfiniBand (IPoIB) (RFC 4391) + http://ietf.org/rfc/rfc4391.txt + + IP over InfiniBand (IPoIB) Architecture (RFC 4392) + http://ietf.org/rfc/rfc4392.txt + + IP over InfiniBand: Connected Mode (RFC 4755) + http://ietf.org/rfc/rfc4755.txt diff --git a/Documentation/infiniband/opa_vnic.rst b/Documentation/infiniband/opa_vnic.rst new file mode 100644 index 000000000..2f888d9ff --- /dev/null +++ b/Documentation/infiniband/opa_vnic.rst @@ -0,0 +1,159 @@ +================================================================= +Intel Omni-Path (OPA) Virtual Network Interface Controller (VNIC) +================================================================= + +Intel Omni-Path (OPA) Virtual Network Interface Controller (VNIC) feature +supports Ethernet functionality over Omni-Path fabric by encapsulating +the Ethernet packets between HFI nodes. + +Architecture +============= +The patterns of exchanges of Omni-Path encapsulated Ethernet packets +involves one or more virtual Ethernet switches overlaid on the Omni-Path +fabric topology. A subset of HFI nodes on the Omni-Path fabric are +permitted to exchange encapsulated Ethernet packets across a particular +virtual Ethernet switch. The virtual Ethernet switches are logical +abstractions achieved by configuring the HFI nodes on the fabric for +header generation and processing. In the simplest configuration all HFI +nodes across the fabric exchange encapsulated Ethernet packets over a +single virtual Ethernet switch. A virtual Ethernet switch, is effectively +an independent Ethernet network. The configuration is performed by an +Ethernet Manager (EM) which is part of the trusted Fabric Manager (FM) +application. HFI nodes can have multiple VNICs each connected to a +different virtual Ethernet switch. The below diagram presents a case +of two virtual Ethernet switches with two HFI nodes:: + + +-------------------+ + | Subnet/ | + | Ethernet | + | Manager | + +-------------------+ + / / + / / + / / + / / + +-----------------------------+ +------------------------------+ + | Virtual Ethernet Switch | | Virtual Ethernet Switch | + | +---------+ +---------+ | | +---------+ +---------+ | + | | VPORT | | VPORT | | | | VPORT | | VPORT | | + +--+---------+----+---------+-+ +-+---------+----+---------+---+ + | \ / | + | \ / | + | \/ | + | / \ | + | / \ | + +-----------+------------+ +-----------+------------+ + | VNIC | VNIC | | VNIC | VNIC | + +-----------+------------+ +-----------+------------+ + | HFI | | HFI | + +------------------------+ +------------------------+ + + +The Omni-Path encapsulated Ethernet packet format is as described below. + +==================== ================================ +Bits Field +==================== ================================ +Quad Word 0: +0-19 SLID (lower 20 bits) +20-30 Length (in Quad Words) +31 BECN bit +32-51 DLID (lower 20 bits) +52-56 SC (Service Class) +57-59 RC (Routing Control) +60 FECN bit +61-62 L2 (=10, 16B format) +63 LT (=1, Link Transfer Head Flit) + +Quad Word 1: +0-7 L4 type (=0x78 ETHERNET) +8-11 SLID[23:20] +12-15 DLID[23:20] +16-31 PKEY +32-47 Entropy +48-63 Reserved + +Quad Word 2: +0-15 Reserved +16-31 L4 header +32-63 Ethernet Packet + +Quad Words 3 to N-1: +0-63 Ethernet packet (pad extended) + +Quad Word N (last): +0-23 Ethernet packet (pad extended) +24-55 ICRC +56-61 Tail +62-63 LT (=01, Link Transfer Tail Flit) +==================== ================================ + +Ethernet packet is padded on the transmit side to ensure that the VNIC OPA +packet is quad word aligned. The 'Tail' field contains the number of bytes +padded. On the receive side the 'Tail' field is read and the padding is +removed (along with ICRC, Tail and OPA header) before passing packet up +the network stack. + +The L4 header field contains the virtual Ethernet switch id the VNIC port +belongs to. On the receive side, this field is used to de-multiplex the +received VNIC packets to different VNIC ports. + +Driver Design +============== +Intel OPA VNIC software design is presented in the below diagram. +OPA VNIC functionality has a HW dependent component and a HW +independent component. + +The support has been added for IB device to allocate and free the RDMA +netdev devices. The RDMA netdev supports interfacing with the network +stack thus creating standard network interfaces. OPA_VNIC is an RDMA +netdev device type. + +The HW dependent VNIC functionality is part of the HFI1 driver. It +implements the verbs to allocate and free the OPA_VNIC RDMA netdev. +It involves HW resource allocation/management for VNIC functionality. +It interfaces with the network stack and implements the required +net_device_ops functions. It expects Omni-Path encapsulated Ethernet +packets in the transmit path and provides HW access to them. It strips +the Omni-Path header from the received packets before passing them up +the network stack. It also implements the RDMA netdev control operations. + +The OPA VNIC module implements the HW independent VNIC functionality. +It consists of two parts. The VNIC Ethernet Management Agent (VEMA) +registers itself with IB core as an IB client and interfaces with the +IB MAD stack. It exchanges the management information with the Ethernet +Manager (EM) and the VNIC netdev. The VNIC netdev part allocates and frees +the OPA_VNIC RDMA netdev devices. It overrides the net_device_ops functions +set by HW dependent VNIC driver where required to accommodate any control +operation. It also handles the encapsulation of Ethernet packets with an +Omni-Path header in the transmit path. For each VNIC interface, the +information required for encapsulation is configured by the EM via VEMA MAD +interface. It also passes any control information to the HW dependent driver +by invoking the RDMA netdev control operations:: + + +-------------------+ +----------------------+ + | | | Linux | + | IB MAD | | Network | + | | | Stack | + +-------------------+ +----------------------+ + | | | + | | | + +----------------------------+ | + | | | + | OPA VNIC Module | | + | (OPA VNIC RDMA Netdev | | + | & EMA functions) | | + | | | + +----------------------------+ | + | | + | | + +------------------+ | + | IB core | | + +------------------+ | + | | + | | + +--------------------------------------------+ + | | + | HFI1 Driver with VNIC support | + | | + +--------------------------------------------+ diff --git a/Documentation/infiniband/sysfs.rst b/Documentation/infiniband/sysfs.rst new file mode 100644 index 000000000..f0abd6fa4 --- /dev/null +++ b/Documentation/infiniband/sysfs.rst @@ -0,0 +1,6 @@ +=========== +Sysfs files +=========== + +The sysfs interface has moved to +Documentation/ABI/stable/sysfs-class-infiniband. diff --git a/Documentation/infiniband/tag_matching.rst b/Documentation/infiniband/tag_matching.rst new file mode 100644 index 000000000..ef56ea585 --- /dev/null +++ b/Documentation/infiniband/tag_matching.rst @@ -0,0 +1,69 @@ +================== +Tag matching logic +================== + +The MPI standard defines a set of rules, known as tag-matching, for matching +source send operations to destination receives. The following parameters must +match the following source and destination parameters: + +* Communicator +* User tag - wild card may be specified by the receiver +* Source rank – wild car may be specified by the receiver +* Destination rank – wild + +The ordering rules require that when more than one pair of send and receive +message envelopes may match, the pair that includes the earliest posted-send +and the earliest posted-receive is the pair that must be used to satisfy the +matching operation. However, this doesn’t imply that tags are consumed in +the order they are created, e.g., a later generated tag may be consumed, if +earlier tags can’t be used to satisfy the matching rules. + +When a message is sent from the sender to the receiver, the communication +library may attempt to process the operation either after or before the +corresponding matching receive is posted. If a matching receive is posted, +this is an expected message, otherwise it is called an unexpected message. +Implementations frequently use different matching schemes for these two +different matching instances. + +To keep MPI library memory footprint down, MPI implementations typically use +two different protocols for this purpose: + +1. The Eager protocol- the complete message is sent when the send is +processed by the sender. A completion send is received in the send_cq +notifying that the buffer can be reused. + +2. The Rendezvous Protocol - the sender sends the tag-matching header, +and perhaps a portion of data when first notifying the receiver. When the +corresponding buffer is posted, the responder will use the information from +the header to initiate an RDMA READ operation directly to the matching buffer. +A fin message needs to be received in order for the buffer to be reused. + +Tag matching implementation +=========================== + +There are two types of matching objects used, the posted receive list and the +unexpected message list. The application posts receive buffers through calls +to the MPI receive routines in the posted receive list and posts send messages +using the MPI send routines. The head of the posted receive list may be +maintained by the hardware, with the software expected to shadow this list. + +When send is initiated and arrives at the receive side, if there is no +pre-posted receive for this arriving message, it is passed to the software and +placed in the unexpected message list. Otherwise the match is processed, +including rendezvous processing, if appropriate, delivering the data to the +specified receive buffer. This allows overlapping receive-side MPI tag +matching with computation. + +When a receive-message is posted, the communication library will first check +the software unexpected message list for a matching receive. If a match is +found, data is delivered to the user buffer, using a software controlled +protocol. The UCX implementation uses either an eager or rendezvous protocol, +depending on data size. If no match is found, the entire pre-posted receive +list is maintained by the hardware, and there is space to add one more +pre-posted receive to this list, this receive is passed to the hardware. +Software is expected to shadow this list, to help with processing MPI cancel +operations. In addition, because hardware and software are not expected to be +tightly synchronized with respect to the tag-matching operation, this shadow +list is used to detect the case that a pre-posted receive is passed to the +hardware, as the matching unexpected message is being passed from the hardware +to the software. diff --git a/Documentation/infiniband/user_mad.rst b/Documentation/infiniband/user_mad.rst new file mode 100644 index 000000000..d88abfc0e --- /dev/null +++ b/Documentation/infiniband/user_mad.rst @@ -0,0 +1,166 @@ +==================== +Userspace MAD access +==================== + +Device files +============ + + Each port of each InfiniBand device has a "umad" device and an + "issm" device attached. For example, a two-port HCA will have two + umad devices and two issm devices, while a switch will have one + device of each type (for switch port 0). + +Creating MAD agents +=================== + + A MAD agent can be created by filling in a struct ib_user_mad_reg_req + and then calling the IB_USER_MAD_REGISTER_AGENT ioctl on a file + descriptor for the appropriate device file. If the registration + request succeeds, a 32-bit id will be returned in the structure. + For example:: + + struct ib_user_mad_reg_req req = { /* ... */ }; + ret = ioctl(fd, IB_USER_MAD_REGISTER_AGENT, (char *) &req); + if (!ret) + my_agent = req.id; + else + perror("agent register"); + + Agents can be unregistered with the IB_USER_MAD_UNREGISTER_AGENT + ioctl. Also, all agents registered through a file descriptor will + be unregistered when the descriptor is closed. + + 2014 + a new registration ioctl is now provided which allows additional + fields to be provided during registration. + Users of this registration call are implicitly setting the use of + pkey_index (see below). + +Receiving MADs +============== + + MADs are received using read(). The receive side now supports + RMPP. The buffer passed to read() must be at least one + struct ib_user_mad + 256 bytes. For example: + + If the buffer passed is not large enough to hold the received + MAD (RMPP), the errno is set to ENOSPC and the length of the + buffer needed is set in mad.length. + + Example for normal MAD (non RMPP) reads:: + + struct ib_user_mad *mad; + mad = malloc(sizeof *mad + 256); + ret = read(fd, mad, sizeof *mad + 256); + if (ret != sizeof mad + 256) { + perror("read"); + free(mad); + } + + Example for RMPP reads:: + + struct ib_user_mad *mad; + mad = malloc(sizeof *mad + 256); + ret = read(fd, mad, sizeof *mad + 256); + if (ret == -ENOSPC)) { + length = mad.length; + free(mad); + mad = malloc(sizeof *mad + length); + ret = read(fd, mad, sizeof *mad + length); + } + if (ret < 0) { + perror("read"); + free(mad); + } + + In addition to the actual MAD contents, the other struct ib_user_mad + fields will be filled in with information on the received MAD. For + example, the remote LID will be in mad.lid. + + If a send times out, a receive will be generated with mad.status set + to ETIMEDOUT. Otherwise when a MAD has been successfully received, + mad.status will be 0. + + poll()/select() may be used to wait until a MAD can be read. + +Sending MADs +============ + + MADs are sent using write(). The agent ID for sending should be + filled into the id field of the MAD, the destination LID should be + filled into the lid field, and so on. The send side does support + RMPP so arbitrary length MAD can be sent. For example:: + + struct ib_user_mad *mad; + + mad = malloc(sizeof *mad + mad_length); + + /* fill in mad->data */ + + mad->hdr.id = my_agent; /* req.id from agent registration */ + mad->hdr.lid = my_dest; /* in network byte order... */ + /* etc. */ + + ret = write(fd, &mad, sizeof *mad + mad_length); + if (ret != sizeof *mad + mad_length) + perror("write"); + +Transaction IDs +=============== + + Users of the umad devices can use the lower 32 bits of the + transaction ID field (that is, the least significant half of the + field in network byte order) in MADs being sent to match + request/response pairs. The upper 32 bits are reserved for use by + the kernel and will be overwritten before a MAD is sent. + +P_Key Index Handling +==================== + + The old ib_umad interface did not allow setting the P_Key index for + MADs that are sent and did not provide a way for obtaining the P_Key + index of received MADs. A new layout for struct ib_user_mad_hdr + with a pkey_index member has been defined; however, to preserve binary + compatibility with older applications, this new layout will not be used + unless one of IB_USER_MAD_ENABLE_PKEY or IB_USER_MAD_REGISTER_AGENT2 ioctl's + are called before a file descriptor is used for anything else. + + In September 2008, the IB_USER_MAD_ABI_VERSION will be incremented + to 6, the new layout of struct ib_user_mad_hdr will be used by + default, and the IB_USER_MAD_ENABLE_PKEY ioctl will be removed. + +Setting IsSM Capability Bit +=========================== + + To set the IsSM capability bit for a port, simply open the + corresponding issm device file. If the IsSM bit is already set, + then the open call will block until the bit is cleared (or return + immediately with errno set to EAGAIN if the O_NONBLOCK flag is + passed to open()). The IsSM bit will be cleared when the issm file + is closed. No read, write or other operations can be performed on + the issm file. + +/dev files +========== + + To create the appropriate character device files automatically with + udev, a rule like:: + + KERNEL=="umad*", NAME="infiniband/%k" + KERNEL=="issm*", NAME="infiniband/%k" + + can be used. This will create device nodes named:: + + /dev/infiniband/umad0 + /dev/infiniband/issm0 + + for the first port, and so on. The InfiniBand device and port + associated with these devices can be determined from the files:: + + /sys/class/infiniband_mad/umad0/ibdev + /sys/class/infiniband_mad/umad0/port + + and:: + + /sys/class/infiniband_mad/issm0/ibdev + /sys/class/infiniband_mad/issm0/port diff --git a/Documentation/infiniband/user_verbs.rst b/Documentation/infiniband/user_verbs.rst new file mode 100644 index 000000000..8ddc4b1cf --- /dev/null +++ b/Documentation/infiniband/user_verbs.rst @@ -0,0 +1,75 @@ +====================== +Userspace verbs access +====================== + + The ib_uverbs module, built by enabling CONFIG_INFINIBAND_USER_VERBS, + enables direct userspace access to IB hardware via "verbs," as + described in chapter 11 of the InfiniBand Architecture Specification. + + To use the verbs, the libibverbs library, available from + https://github.com/linux-rdma/rdma-core, is required. libibverbs contains a + device-independent API for using the ib_uverbs interface. + libibverbs also requires appropriate device-dependent kernel and + userspace driver for your InfiniBand hardware. For example, to use + a Mellanox HCA, you will need the ib_mthca kernel module and the + libmthca userspace driver be installed. + +User-kernel communication +========================= + + Userspace communicates with the kernel for slow path, resource + management operations via the /dev/infiniband/uverbsN character + devices. Fast path operations are typically performed by writing + directly to hardware registers mmap()ed into userspace, with no + system call or context switch into the kernel. + + Commands are sent to the kernel via write()s on these device files. + The ABI is defined in drivers/infiniband/include/ib_user_verbs.h. + The structs for commands that require a response from the kernel + contain a 64-bit field used to pass a pointer to an output buffer. + Status is returned to userspace as the return value of the write() + system call. + +Resource management +=================== + + Since creation and destruction of all IB resources is done by + commands passed through a file descriptor, the kernel can keep track + of which resources are attached to a given userspace context. The + ib_uverbs module maintains idr tables that are used to translate + between kernel pointers and opaque userspace handles, so that kernel + pointers are never exposed to userspace and userspace cannot trick + the kernel into following a bogus pointer. + + This also allows the kernel to clean up when a process exits and + prevent one process from touching another process's resources. + +Memory pinning +============== + + Direct userspace I/O requires that memory regions that are potential + I/O targets be kept resident at the same physical address. The + ib_uverbs module manages pinning and unpinning memory regions via + get_user_pages() and put_page() calls. It also accounts for the + amount of memory pinned in the process's pinned_vm, and checks that + unprivileged processes do not exceed their RLIMIT_MEMLOCK limit. + + Pages that are pinned multiple times are counted each time they are + pinned, so the value of pinned_vm may be an overestimate of the + number of pages pinned by a process. + +/dev files +========== + + To create the appropriate character device files automatically with + udev, a rule like:: + + KERNEL=="uverbs*", NAME="infiniband/%k" + + can be used. This will create device nodes named:: + + /dev/infiniband/uverbs0 + + and so on. Since the InfiniBand userspace verbs should be safe for + use by non-privileged processes, it may be useful to add an + appropriate MODE or GROUP to the udev rule. |