summaryrefslogtreecommitdiffstats
path: root/src/ring.c
diff options
context:
space:
mode:
Diffstat (limited to 'src/ring.c')
-rw-r--r--src/ring.c482
1 files changed, 482 insertions, 0 deletions
diff --git a/src/ring.c b/src/ring.c
new file mode 100644
index 0000000..849221e
--- /dev/null
+++ b/src/ring.c
@@ -0,0 +1,482 @@
+/*
+ * Ring buffer management
+ *
+ * Copyright (C) 2000-2019 Willy Tarreau - w@1wt.eu
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation, version 2.1
+ * exclusively.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include <stdlib.h>
+#include <haproxy/api.h>
+#include <haproxy/applet.h>
+#include <haproxy/buf.h>
+#include <haproxy/cli.h>
+#include <haproxy/ring.h>
+#include <haproxy/sc_strm.h>
+#include <haproxy/stconn.h>
+#include <haproxy/thread.h>
+
+/* context used to dump the contents of a ring via "show events" or "show errors" */
+struct show_ring_ctx {
+ struct ring *ring; /* ring to be dumped */
+ size_t ofs; /* storage offset to restart from; ~0=oldest */
+ uint flags; /* set of RING_WF_* */
+};
+
+/* Initialize a pre-allocated ring with the buffer area
+ * of size */
+void ring_init(struct ring *ring, void *area, size_t size)
+{
+ HA_RWLOCK_INIT(&ring->lock);
+ LIST_INIT(&ring->waiters);
+ ring->readers_count = 0;
+ ring->buf = b_make(area, size, 0, 0);
+ /* write the initial RC byte */
+ b_putchr(&ring->buf, 0);
+}
+
+/* Creates and returns a ring buffer of size <size> bytes. Returns NULL on
+ * allocation failure.
+ */
+struct ring *ring_new(size_t size)
+{
+ struct ring *ring = NULL;
+ void *area = NULL;
+
+ if (size < 2)
+ goto fail;
+
+ ring = malloc(sizeof(*ring));
+ if (!ring)
+ goto fail;
+
+ area = malloc(size);
+ if (!area)
+ goto fail;
+
+ ring_init(ring, area, size);
+ return ring;
+ fail:
+ free(area);
+ free(ring);
+ return NULL;
+}
+
+/* Creates a unified ring + storage area at address <area> for <size> bytes.
+ * If <area> is null, then it's allocated of the requested size. The ring
+ * struct is part of the area so the usable area is slightly reduced. However
+ * the ring storage is immediately adjacent to the struct. ring_free() will
+ * ignore such rings, so the caller is responsible for releasing them.
+ */
+struct ring *ring_make_from_area(void *area, size_t size)
+{
+ struct ring *ring = NULL;
+
+ if (size < sizeof(*ring))
+ return NULL;
+
+ if (!area)
+ area = malloc(size);
+ if (!area)
+ return NULL;
+
+ ring = area;
+ area += sizeof(*ring);
+ ring_init(ring, area, size - sizeof(*ring));
+ return ring;
+}
+
+/* Cast an unified ring + storage area to a ring from <area>, without
+ * reinitializing the data buffer.
+ *
+ * Reinitialize the waiters and the lock.
+ */
+struct ring *ring_cast_from_area(void *area)
+{
+ struct ring *ring = NULL;
+
+ ring = area;
+ ring->buf.area = area + sizeof(*ring);
+
+ HA_RWLOCK_INIT(&ring->lock);
+ LIST_INIT(&ring->waiters);
+ ring->readers_count = 0;
+
+ return ring;
+}
+
+/* Resizes existing ring <ring> to <size> which must be larger, without losing
+ * its contents. The new size must be at least as large as the previous one or
+ * no change will be performed. The pointer to the ring is returned on success,
+ * or NULL on allocation failure. This will lock the ring for writes.
+ */
+struct ring *ring_resize(struct ring *ring, size_t size)
+{
+ void *area;
+
+ if (b_size(&ring->buf) >= size)
+ return ring;
+
+ area = malloc(size);
+ if (!area)
+ return NULL;
+
+ HA_RWLOCK_WRLOCK(RING_LOCK, &ring->lock);
+
+ /* recheck the buffer's size, it may have changed during the malloc */
+ if (b_size(&ring->buf) < size) {
+ /* copy old contents */
+ b_getblk(&ring->buf, area, ring->buf.data, 0);
+ area = HA_ATOMIC_XCHG(&ring->buf.area, area);
+ ring->buf.size = size;
+ }
+
+ HA_RWLOCK_WRUNLOCK(RING_LOCK, &ring->lock);
+
+ free(area);
+ return ring;
+}
+
+/* destroys and frees ring <ring> */
+void ring_free(struct ring *ring)
+{
+ if (!ring)
+ return;
+
+ /* make sure it was not allocated by ring_make_from_area */
+ if (ring->buf.area == (void *)ring + sizeof(*ring))
+ return;
+
+ free(ring->buf.area);
+ free(ring);
+}
+
+/* Tries to send <npfx> parts from <prefix> followed by <nmsg> parts from <msg>
+ * to ring <ring>. The message is sent atomically. It may be truncated to
+ * <maxlen> bytes if <maxlen> is non-null. There is no distinction between the
+ * two lists, it's just a convenience to help the caller prepend some prefixes
+ * when necessary. It takes the ring's write lock to make sure no other thread
+ * will touch the buffer during the update. Returns the number of bytes sent,
+ * or <=0 on failure.
+ */
+ssize_t ring_write(struct ring *ring, size_t maxlen, const struct ist pfx[], size_t npfx, const struct ist msg[], size_t nmsg)
+{
+ struct buffer *buf = &ring->buf;
+ struct appctx *appctx;
+ size_t totlen = 0;
+ size_t lenlen;
+ uint64_t dellen;
+ int dellenlen;
+ ssize_t sent = 0;
+ int i;
+
+ /* we have to find some room to add our message (the buffer is
+ * never empty and at least contains the previous counter) and
+ * to update both the buffer contents and heads at the same
+ * time (it's doable using atomic ops but not worth the
+ * trouble, let's just lock). For this we first need to know
+ * the total message's length. We cannot measure it while
+ * copying due to the varint encoding of the length.
+ */
+ for (i = 0; i < npfx; i++)
+ totlen += pfx[i].len;
+ for (i = 0; i < nmsg; i++)
+ totlen += msg[i].len;
+
+ if (totlen > maxlen)
+ totlen = maxlen;
+
+ lenlen = varint_bytes(totlen);
+
+ HA_RWLOCK_WRLOCK(RING_LOCK, &ring->lock);
+ if (lenlen + totlen + 1 + 1 > b_size(buf))
+ goto done_buf;
+
+ while (b_room(buf) < lenlen + totlen + 1) {
+ /* we need to delete the oldest message (from the end),
+ * and we have to stop if there's a reader stuck there.
+ * Unless there's corruption in the buffer it's guaranteed
+ * that we have enough data to find 1 counter byte, a
+ * varint-encoded length (1 byte min) and the message
+ * payload (0 bytes min).
+ */
+ if (*b_head(buf))
+ goto done_buf;
+ dellenlen = b_peek_varint(buf, 1, &dellen);
+ if (!dellenlen)
+ goto done_buf;
+ BUG_ON(b_data(buf) < 1 + dellenlen + dellen);
+
+ b_del(buf, 1 + dellenlen + dellen);
+ }
+
+ /* OK now we do have room */
+ __b_put_varint(buf, totlen);
+
+ totlen = 0;
+ for (i = 0; i < npfx; i++) {
+ size_t len = pfx[i].len;
+
+ if (len + totlen > maxlen)
+ len = maxlen - totlen;
+ if (len)
+ __b_putblk(buf, pfx[i].ptr, len);
+ totlen += len;
+ }
+
+ for (i = 0; i < nmsg; i++) {
+ size_t len = msg[i].len;
+
+ if (len + totlen > maxlen)
+ len = maxlen - totlen;
+ if (len)
+ __b_putblk(buf, msg[i].ptr, len);
+ totlen += len;
+ }
+
+ *b_tail(buf) = 0; buf->data++; // new read counter
+ sent = lenlen + totlen + 1;
+
+ /* notify potential readers */
+ list_for_each_entry(appctx, &ring->waiters, wait_entry)
+ appctx_wakeup(appctx);
+
+ done_buf:
+ HA_RWLOCK_WRUNLOCK(RING_LOCK, &ring->lock);
+ return sent;
+}
+
+/* Tries to attach appctx <appctx> as a new reader on ring <ring>. This is
+ * meant to be used by low level appctx code such as CLI or ring forwarding.
+ * For higher level functions, please see the relevant parts in appctx or CLI.
+ * It returns non-zero on success or zero on failure if too many users are
+ * already attached. On success, the caller MUST call ring_detach_appctx()
+ * to detach itself, even if it was never woken up.
+ */
+int ring_attach(struct ring *ring)
+{
+ int users = ring->readers_count;
+
+ do {
+ if (users >= 255)
+ return 0;
+ } while (!_HA_ATOMIC_CAS(&ring->readers_count, &users, users + 1));
+ return 1;
+}
+
+/* detach an appctx from a ring. The appctx is expected to be waiting at offset
+ * <ofs> relative to the beginning of the storage, or ~0 if not waiting yet.
+ * Nothing is done if <ring> is NULL.
+ */
+void ring_detach_appctx(struct ring *ring, struct appctx *appctx, size_t ofs)
+{
+ if (!ring)
+ return;
+
+ HA_RWLOCK_WRLOCK(RING_LOCK, &ring->lock);
+ if (ofs != ~0) {
+ /* reader was still attached */
+ if (ofs < b_head_ofs(&ring->buf))
+ ofs += b_size(&ring->buf) - b_head_ofs(&ring->buf);
+ else
+ ofs -= b_head_ofs(&ring->buf);
+
+ BUG_ON(ofs >= b_size(&ring->buf));
+ LIST_DEL_INIT(&appctx->wait_entry);
+ HA_ATOMIC_DEC(b_peek(&ring->buf, ofs));
+ }
+ HA_ATOMIC_DEC(&ring->readers_count);
+ HA_RWLOCK_WRUNLOCK(RING_LOCK, &ring->lock);
+}
+
+/* Tries to attach CLI handler <appctx> as a new reader on ring <ring>. This is
+ * meant to be used when registering a CLI function to dump a buffer, so it
+ * returns zero on success, or non-zero on failure with a message in the appctx
+ * CLI context. It automatically sets the io_handler and io_release callbacks if
+ * they were not set. The <flags> take a combination of RING_WF_*.
+ */
+int ring_attach_cli(struct ring *ring, struct appctx *appctx, uint flags)
+{
+ struct show_ring_ctx *ctx = applet_reserve_svcctx(appctx, sizeof(*ctx));
+
+ if (!ring_attach(ring))
+ return cli_err(appctx,
+ "Sorry, too many watchers (255) on this ring buffer. "
+ "What could it have so interesting to attract so many watchers ?");
+
+ if (!appctx->io_handler)
+ appctx->io_handler = cli_io_handler_show_ring;
+ if (!appctx->io_release)
+ appctx->io_release = cli_io_release_show_ring;
+
+ memset(ctx, 0, sizeof(*ctx));
+ ctx->ring = ring;
+ ctx->ofs = ~0; // start from the oldest event
+ ctx->flags = flags;
+ return 0;
+}
+
+/* This function dumps all events from the ring whose pointer is in <p0> into
+ * the appctx's output buffer, and takes from <o0> the seek offset into the
+ * buffer's history (0 for oldest known event). It looks at <i0> for boolean
+ * options: bit0 means it must wait for new data or any key to be pressed. Bit1
+ * means it must seek directly to the end to wait for new contents. It returns
+ * 0 if the output buffer or events are missing is full and it needs to be
+ * called again, otherwise non-zero. It is meant to be used with
+ * cli_release_show_ring() to clean up.
+ */
+int cli_io_handler_show_ring(struct appctx *appctx)
+{
+ struct show_ring_ctx *ctx = appctx->svcctx;
+ struct stconn *sc = appctx_sc(appctx);
+ struct ring *ring = ctx->ring;
+ struct buffer *buf = &ring->buf;
+ size_t ofs;
+ size_t last_ofs;
+ uint64_t msg_len;
+ size_t len, cnt;
+ int ret;
+
+ /* FIXME: Don't watch the other side !*/
+ if (unlikely(sc_opposite(sc)->flags & SC_FL_SHUT_DONE))
+ return 1;
+
+ HA_RWLOCK_WRLOCK(RING_LOCK, &ring->lock);
+ LIST_DEL_INIT(&appctx->wait_entry);
+ HA_RWLOCK_WRUNLOCK(RING_LOCK, &ring->lock);
+
+ HA_RWLOCK_RDLOCK(RING_LOCK, &ring->lock);
+
+ /* explanation for the initialization below: it would be better to do
+ * this in the parsing function but this would occasionally result in
+ * dropped events because we'd take a reference on the oldest message
+ * and keep it while being scheduled. Thus instead let's take it the
+ * first time we enter here so that we have a chance to pass many
+ * existing messages before grabbing a reference to a location. This
+ * value cannot be produced after initialization.
+ */
+ if (unlikely(ctx->ofs == ~0)) {
+ /* going to the end means looking at tail-1 */
+ ctx->ofs = b_peek_ofs(buf, (ctx->flags & RING_WF_SEEK_NEW) ? b_data(buf) - 1 : 0);
+ HA_ATOMIC_INC(b_orig(buf) + ctx->ofs);
+ }
+
+ /* we were already there, adjust the offset to be relative to
+ * the buffer's head and remove us from the counter.
+ */
+ ofs = ctx->ofs - b_head_ofs(buf);
+ if (ctx->ofs < b_head_ofs(buf))
+ ofs += b_size(buf);
+
+ BUG_ON(ofs >= buf->size);
+ HA_ATOMIC_DEC(b_peek(buf, ofs));
+
+ /* in this loop, ofs always points to the counter byte that precedes
+ * the message so that we can take our reference there if we have to
+ * stop before the end (ret=0).
+ */
+ ret = 1;
+ while (ofs + 1 < b_data(buf)) {
+ cnt = 1;
+ len = b_peek_varint(buf, ofs + cnt, &msg_len);
+ if (!len)
+ break;
+ cnt += len;
+ BUG_ON(msg_len + ofs + cnt + 1 > b_data(buf));
+
+ if (unlikely(msg_len + 1 > b_size(&trash))) {
+ /* too large a message to ever fit, let's skip it */
+ ofs += cnt + msg_len;
+ continue;
+ }
+
+ chunk_reset(&trash);
+ len = b_getblk(buf, trash.area, msg_len, ofs + cnt);
+ trash.data += len;
+ trash.area[trash.data++] = '\n';
+
+ if (applet_putchk(appctx, &trash) == -1) {
+ ret = 0;
+ break;
+ }
+ ofs += cnt + msg_len;
+ }
+
+ HA_ATOMIC_INC(b_peek(buf, ofs));
+ last_ofs = b_tail_ofs(buf);
+ ctx->ofs = b_peek_ofs(buf, ofs);
+ HA_RWLOCK_RDUNLOCK(RING_LOCK, &ring->lock);
+
+ if (ret && (ctx->flags & RING_WF_WAIT_MODE)) {
+ /* we've drained everything and are configured to wait for more
+ * data or an event (keypress, close)
+ */
+ if (!sc_oc(sc)->output && !(sc->flags & SC_FL_SHUT_DONE)) {
+ /* let's be woken up once new data arrive */
+ HA_RWLOCK_WRLOCK(RING_LOCK, &ring->lock);
+ LIST_APPEND(&ring->waiters, &appctx->wait_entry);
+ ofs = b_tail_ofs(&ring->buf);
+ HA_RWLOCK_WRUNLOCK(RING_LOCK, &ring->lock);
+ if (ofs != last_ofs) {
+ /* more data was added into the ring between the
+ * unlock and the lock, and the writer might not
+ * have seen us. We need to reschedule a read.
+ */
+ applet_have_more_data(appctx);
+ } else
+ applet_have_no_more_data(appctx);
+ ret = 0;
+ }
+ /* always drain all the request */
+ co_skip(sc_oc(sc), sc_oc(sc)->output);
+ }
+
+ applet_expect_no_data(appctx);
+ return ret;
+}
+
+/* must be called after cli_io_handler_show_ring() above */
+void cli_io_release_show_ring(struct appctx *appctx)
+{
+ struct show_ring_ctx *ctx = appctx->svcctx;
+ struct ring *ring = ctx->ring;
+ size_t ofs = ctx->ofs;
+
+ ring_detach_appctx(ring, appctx, ofs);
+}
+
+/* Returns the MAXIMUM payload len that could theoretically fit into the ring
+ * based on ring buffer size.
+ *
+ * Computation logic relies on implementation details from 'ring-t.h'.
+ */
+size_t ring_max_payload(const struct ring *ring)
+{
+ size_t max;
+
+ /* initial max = bufsize - 1 (initial RC) - 1 (payload RC) */
+ max = b_size(&ring->buf) - 1 - 1;
+
+ /* subtract payload VI (varint-encoded size) */
+ max -= varint_bytes(max);
+ return max;
+}
+
+/*
+ * Local variables:
+ * c-indent-level: 8
+ * c-basic-offset: 8
+ * End:
+ */