document metadata_switch_timer() deadlock

[lttng-tools.git] / src / common / consumer-timer.c

diff --git a/src/common/consumer-timer.c b/src/common/consumer-timer.c
index f9c41c0eb0f8f9f32def26028f743ce7df45d322..a32dbf64d2d229e730f6e87a4add490c75567a39 100644 (file)
--- a/src/common/consumer-timer.c
+++ b/src/common/consumer-timer.c
@@ -56,6 +56,10 @@ static void setmask(sigset_t *mask)
 
 /*
  * Execute action on a timer switch.
+ *
+ * Beware: metadata_switch_timer() should *never* take a mutex also held
+ * while consumer_timer_switch_stop() is called. It would result in
+ * deadlocks.
  */
 static void metadata_switch_timer(struct lttng_consumer_local_data *ctx,
                int sig, siginfo_t *si, void *uc)
@@ -74,6 +78,22 @@ static void metadata_switch_timer(struct lttng_consumer_local_data *ctx,
        switch (ctx->type) {
        case LTTNG_CONSUMER32_UST:
        case LTTNG_CONSUMER64_UST:
+               /*
+                * Locks taken by lttng_ustconsumer_request_metadata():
+                * - metadata_socket_lock
+                *   - Calling lttng_ustconsumer_recv_metadata():
+                *     - consumer_data.lock
+                *       - channel->lock
+                *         - channel->metadata_cache->lock
+                *     - Calling consumer_metadata_cache_flushed():
+                *       - consumer_data.lock
+                *         - channel->lock
+                *           - channel->metadata_cache->lock
+                *
+                * Both consumer_data.lock and channel->lock currently
+                * cause a deadlock, since they are held while
+                * consumer_timer_switch_stop() is called.
+                */
                ret = lttng_ustconsumer_request_metadata(ctx, channel);
                if (ret < 0) {
                        channel->switch_timer_error = 1;