Fix: sessiond vs consumerd push/get metadata deadlock
[lttng-tools.git] / src / bin / lttng-sessiond / ust-app.c
index d1cc39baefc4b3fba85b0f6b04dd8b25d9522e46..fc4b7085ec5770e57550c169924e112bb794dd49 100644 (file)
@@ -440,17 +440,20 @@ ssize_t ust_app_push_metadata(struct ust_registry_session *registry,
 {
        int ret;
        char *metadata_str = NULL;
-       size_t len, offset;
+       size_t len, offset, new_metadata_len_sent;
        ssize_t ret_val;
+       uint64_t metadata_key;
 
        assert(registry);
        assert(socket);
 
+       metadata_key = registry->metadata_key;
+
        /*
         * Means that no metadata was assigned to the session. This can
         * happens if no start has been done previously.
         */
-       if (!registry->metadata_key) {
+       if (!metadata_key) {
                return 0;
        }
 
@@ -468,6 +471,7 @@ ssize_t ust_app_push_metadata(struct ust_registry_session *registry,
 
        offset = registry->metadata_len_sent;
        len = registry->metadata_len - registry->metadata_len_sent;
+       new_metadata_len_sent = registry->metadata_len;
        if (len == 0) {
                DBG3("No metadata to push for metadata key %" PRIu64,
                                registry->metadata_key);
@@ -486,13 +490,26 @@ ssize_t ust_app_push_metadata(struct ust_registry_session *registry,
                ret_val = -ENOMEM;
                goto error;
        }
-       /* Copy what we haven't send out. */
+       /* Copy what we haven't sent out. */
        memcpy(metadata_str, registry->metadata + offset, len);
-       registry->metadata_len_sent += len;
 
 push_data:
-       ret = consumer_push_metadata(socket, registry->metadata_key,
+       pthread_mutex_unlock(&registry->lock);
+       /*
+        * We need to unlock the registry while we push metadata to
+        * break a circular dependency between the consumerd metadata
+        * lock and the sessiond registry lock. Indeed, pushing metadata
+        * to the consumerd awaits that it gets pushed all the way to
+        * relayd, but doing so requires grabbing the metadata lock. If
+        * a concurrent metadata request is being performed by
+        * consumerd, this can try to grab the registry lock on the
+        * sessiond while holding the metadata lock on the consumer
+        * daemon. Those push and pull schemes are performed on two
+        * different bidirectionnal communication sockets.
+        */
+       ret = consumer_push_metadata(socket, metadata_key,
                        metadata_str, len, offset);
+       pthread_mutex_lock(&registry->lock);
        if (ret < 0) {
                /*
                 * There is an acceptable race here between the registry
@@ -510,17 +527,29 @@ push_data:
                 */
                if (ret == -LTTCOMM_CONSUMERD_CHANNEL_FAIL) {
                        ret = 0;
+               } else {
+                       ERR("Error pushing metadata to consumer");
                }
-
-               /*
-                * Update back the actual metadata len sent since it
-                * failed here.
-                */
-               registry->metadata_len_sent -= len;
                ret_val = ret;
                goto error_push;
+       } else {
+               /*
+                * Metadata may have been concurrently pushed, since
+                * we're not holding the registry lock while pushing to
+                * consumer.  This is handled by the fact that we send
+                * the metadata content, size, and the offset at which
+                * that metadata belongs. This may arrive out of order
+                * on the consumer side, and the consumer is able to
+                * deal with overlapping fragments. The consumer
+                * supports overlapping fragments, which must be
+                * contiguous starting from offset 0. We keep the
+                * largest metadata_len_sent value of the concurrent
+                * send.
+                */
+               registry->metadata_len_sent =
+                       max_t(size_t, registry->metadata_len_sent,
+                               new_metadata_len_sent);
        }
-
        free(metadata_str);
        return len;
 
@@ -722,6 +751,8 @@ void delete_ust_app_session(int sock, struct ust_app_session *ua_sess,
        }
        pthread_mutex_unlock(&ua_sess->lock);
 
+       consumer_output_put(ua_sess->consumer);
+
        call_rcu(&ua_sess->rcu_head, delete_ust_app_session_rcu);
 }
 
@@ -1685,8 +1716,11 @@ static void shadow_copy_session(struct ust_app_session *ua_sess,
        ua_sess->egid = usess->gid;
        ua_sess->buffer_type = usess->buffer_type;
        ua_sess->bits_per_long = app->bits_per_long;
+
        /* There is only one consumer object per session possible. */
+       consumer_output_get(usess->consumer);
        ua_sess->consumer = usess->consumer;
+
        ua_sess->output_traces = usess->output_traces;
        ua_sess->live_timer_interval = usess->live_timer_interval;
        copy_channel_attr_to_ustctl(&ua_sess->metadata_attr,
@@ -1773,9 +1807,10 @@ static void shadow_copy_session(struct ust_app_session *ua_sess,
 
                lttng_ht_add_unique_str(ua_sess->channels, &ua_chan->node);
        }
+       return;
 
 error:
-       return;
+       consumer_output_put(ua_sess->consumer);
 }
 
 /*
@@ -4079,7 +4114,7 @@ int ust_app_start_trace(struct ltt_ust_session *usess, struct ust_app *app)
                ret = run_as_mkdir_recursive(usess->consumer->dst.trace_path,
                                S_IRWXU | S_IRWXG, ua_sess->euid, ua_sess->egid);
                if (ret < 0) {
-                       if (ret != -EEXIST) {
+                       if (errno != EEXIST) {
                                ERR("Trace directory creation error");
                                goto error_unlock;
                        }
This page took 0.025094 seconds and 4 git commands to generate.