Fix: ring buffer: handle concurrent update in nested buffer wrap around check
[lttng-ust.git] / libringbuffer / ring_buffer_frontend.c
index 0996ecefc0313ff5ef8b955ce7dffd3a6a793849..aff9a25c7267eb02398ed38b41feb1f68ee32260 100644 (file)
@@ -1,7 +1,22 @@
 /*
  * ring_buffer_frontend.c
  *
- * (C) Copyright 2005-2010 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ * Copyright (C) 2005-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
+ *
+ * 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; only
+ * version 2.1 of the License.
+ *
+ * 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
+ *
  *
  * Ring buffer wait-free buffer synchronization. Producer-consumer and flight
  * recorder (overwrite) modes. See thesis:
@@ -34,8 +49,6 @@
  *   - splice one subbuffer worth of data to a pipe
  *   - splice the data from pipe to disk/network
  *   - put_subbuf
- *
- * Dual LGPL v2.1/GPL v2 license.
  */
 
 #define _GNU_SOURCE
@@ -53,6 +66,8 @@
 #include "backend.h"
 #include "frontend.h"
 #include "shm.h"
+#include "tlsfixup.h"
+#include "../liblttng-ust/compat.h"    /* For ENODATA */
 
 #ifndef max
 #define max(a, b)      ((a) > (b) ? (a) : (b))
@@ -981,23 +996,30 @@ void lib_ring_buffer_print_errors(struct channel *chan,
        const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
        void *priv = channel_get_private(chan);
 
-       DBG("ring buffer %s, cpu %d: %lu records written, "
-                         "%lu records overrun\n",
-                         chan->backend.name, cpu,
-                         v_read(config, &buf->records_count),
-                         v_read(config, &buf->records_overrun));
-
-       if (v_read(config, &buf->records_lost_full)
-           || v_read(config, &buf->records_lost_wrap)
-           || v_read(config, &buf->records_lost_big))
-               DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
-                      "  [ %lu buffer full, %lu nest buffer wrap-around, "
-                      "%lu event too big ]\n",
-                      chan->backend.name, cpu,
-                      v_read(config, &buf->records_lost_full),
-                      v_read(config, &buf->records_lost_wrap),
-                      v_read(config, &buf->records_lost_big));
-
+       if (!strcmp(chan->backend.name, "relay-metadata-mmap")) {
+               DBG("ring buffer %s: %lu records written, "
+                       "%lu records overrun\n",
+                       chan->backend.name,
+                       v_read(config, &buf->records_count),
+                       v_read(config, &buf->records_overrun));
+       } else {
+               DBG("ring buffer %s, cpu %d: %lu records written, "
+                       "%lu records overrun\n",
+                       chan->backend.name, cpu,
+                       v_read(config, &buf->records_count),
+                       v_read(config, &buf->records_overrun));
+
+               if (v_read(config, &buf->records_lost_full)
+                   || v_read(config, &buf->records_lost_wrap)
+                   || v_read(config, &buf->records_lost_big))
+                       DBG("ring buffer %s, cpu %d: records were lost. Caused by:\n"
+                               "  [ %lu buffer full, %lu nest buffer wrap-around, "
+                               "%lu event too big ]\n",
+                               chan->backend.name, cpu,
+                               v_read(config, &buf->records_lost_full),
+                               v_read(config, &buf->records_lost_wrap),
+                               v_read(config, &buf->records_lost_big));
+       }
        lib_ring_buffer_print_buffer_errors(buf, chan, priv, cpu, handle);
 }
 
@@ -1187,6 +1209,19 @@ int lib_ring_buffer_try_switch_slow(enum switch_mode mode,
         */
        if (mode == SWITCH_FLUSH || off > 0) {
                if (caa_unlikely(off == 0)) {
+                       /*
+                        * A final flush that encounters an empty
+                        * sub-buffer cannot switch buffer if a
+                        * reader is located within this sub-buffer.
+                        * Anyway, the purpose of final flushing of a
+                        * sub-buffer at offset 0 is to handle the case
+                        * of entirely empty stream.
+                        */
+                       if (caa_unlikely(subbuf_trunc(offsets->begin, chan)
+                                        - subbuf_trunc((unsigned long)
+                                            uatomic_read(&buf->consumed), chan)
+                                       >= chan->backend.buf_size))
+                               return -1;
                        /*
                         * The client does not save any header information.
                         * Don't switch empty subbuffer on finalize, because it
@@ -1281,9 +1316,10 @@ int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf,
 {
        const struct lttng_ust_lib_ring_buffer_config *config = &chan->backend.config;
        struct lttng_ust_shm_handle *handle = ctx->handle;
-       unsigned long reserve_commit_diff;
+       unsigned long reserve_commit_diff, offset_cmp;
 
-       offsets->begin = v_read(config, &buf->offset);
+retry:
+       offsets->begin = offset_cmp = v_read(config, &buf->offset);
        offsets->old = offsets->begin;
        offsets->switch_new_start = 0;
        offsets->switch_new_end = 0;
@@ -1315,7 +1351,7 @@ int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf,
                }
        }
        if (caa_unlikely(offsets->switch_new_start)) {
-               unsigned long sb_index;
+               unsigned long sb_index, commit_count;
 
                /*
                 * We are typically not filling the previous buffer completely.
@@ -1326,12 +1362,32 @@ int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf,
                                 + config->cb.subbuffer_header_size();
                /* Test new buffer integrity */
                sb_index = subbuf_index(offsets->begin, chan);
+               /*
+                * Read buf->offset before buf->commit_cold[sb_index].cc_sb.
+                * lib_ring_buffer_check_deliver() has the matching
+                * memory barriers required around commit_cold cc_sb
+                * updates to ensure reserve and commit counter updates
+                * are not seen reordered when updated by another CPU.
+                */
+               cmm_smp_rmb();
+               commit_count = v_read(config,
+                               &shmp_index(handle, buf->commit_cold,
+                                       sb_index)->cc_sb);
+               /* Read buf->commit_cold[sb_index].cc_sb before buf->offset. */
+               cmm_smp_rmb();
+               if (caa_unlikely(offset_cmp != v_read(config, &buf->offset))) {
+                       /*
+                        * The reserve counter have been concurrently updated
+                        * while we read the commit counter. This means the
+                        * commit counter we read might not match buf->offset
+                        * due to concurrent update. We therefore need to retry.
+                        */
+                       goto retry;
+               }
                reserve_commit_diff =
                  (buf_trunc(offsets->begin, chan)
                   >> chan->backend.num_subbuf_order)
-                 - ((unsigned long) v_read(config,
-                                           &shmp_index(handle, buf->commit_cold, sb_index)->cc_sb)
-                    & chan->commit_count_mask);
+                 - (commit_count & chan->commit_count_mask);
                if (caa_likely(reserve_commit_diff == 0)) {
                        /* Next subbuffer not being written to. */
                        if (caa_unlikely(config->mode != RING_BUFFER_OVERWRITE &&
@@ -1366,7 +1422,8 @@ int lib_ring_buffer_try_reserve_slow(struct lttng_ust_lib_ring_buffer *buf,
 
                        /*
                         * Next subbuffer reserve offset does not match the
-                        * commit offset. Drop record in producer-consumer and
+                        * commit offset, and this did not involve update to the
+                        * reserve counter. Drop record in producer-consumer and
                         * overwrite mode. Caused by either a writer OOPS or too
                         * many nested writes over a reserve/commit pair.
                         */
@@ -1507,3 +1564,11 @@ int lib_ring_buffer_reserve_slow(struct lttng_ust_lib_ring_buffer_ctx *ctx)
        ctx->buf_offset = offsets.begin + offsets.pre_header_padding;
        return 0;
 }
+
+/*
+ * Force a read (imply TLS fixup for dlopen) of TLS variables.
+ */
+void lttng_fixup_ringbuffer_tls(void)
+{
+       asm volatile ("" : : "m" (lib_ring_buffer_nesting));
+}
This page took 0.025636 seconds and 4 git commands to generate.