X-Git-Url: https://git.lttng.org/?p=lttng-tools.git;a=blobdiff_plain;f=src%2Fcommon%2Fconsumer.c;h=efd9e7eb3d368552685c3e0a7bd3f54362943318;hp=0b2f07391009f4a317b8f7c876443addad39d660;hb=4e9a468645939ea62541fa802893b928b01888b7;hpb=43c34bc328e6970b298c9f5cd661e2ca648ebf16 diff --git a/src/common/consumer.c b/src/common/consumer.c index 0b2f07391..efd9e7eb3 100644 --- a/src/common/consumer.c +++ b/src/common/consumer.c @@ -56,7 +56,7 @@ int consumer_poll_timeout = -1; * Also updated by the signal handler (consumer_should_exit()). Read by the * polling threads. */ -volatile int consumer_quit = 0; +volatile int consumer_quit; /* * The following two hash tables are visible by all threads which are separated @@ -66,8 +66,23 @@ volatile int consumer_quit = 0; * stream element in this ht should only be updated by the metadata poll thread * for the metadata and the data poll thread for the data. */ -struct lttng_ht *metadata_ht = NULL; -struct lttng_ht *data_ht = NULL; +struct lttng_ht *metadata_ht; +struct lttng_ht *data_ht; + +/* + * Notify a thread pipe to poll back again. This usually means that some global + * state has changed so we just send back the thread in a poll wait call. + */ +static void notify_thread_pipe(int wpipe) +{ + int ret; + + do { + struct lttng_consumer_stream *null_stream = NULL; + + ret = write(wpipe, &null_stream, sizeof(null_stream)); + } while (ret < 0 && errno == EINTR); +} /* * Find a stream. The consumer_data.lock must be locked during this @@ -172,17 +187,6 @@ void consumer_free_stream(struct rcu_head *head) free(stream); } -static -void consumer_free_metadata_stream(struct rcu_head *head) -{ - struct lttng_ht_node_ulong *node = - caa_container_of(head, struct lttng_ht_node_ulong, head); - struct lttng_consumer_stream *stream = - caa_container_of(node, struct lttng_consumer_stream, waitfd_node); - - free(stream); -} - /* * RCU protected relayd socket pair free. */ @@ -193,6 +197,17 @@ static void consumer_rcu_free_relayd(struct rcu_head *head) struct consumer_relayd_sock_pair *relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node); + /* + * Close all sockets. This is done in the call RCU since we don't want the + * socket fds to be reassigned thus potentially creating bad state of the + * relayd object. + * + * We do not have to lock the control socket mutex here since at this stage + * there is no one referencing to this relayd object. + */ + (void) relayd_close(&relayd->control_sock); + (void) relayd_close(&relayd->data_sock); + free(relayd); } @@ -215,20 +230,88 @@ static void destroy_relayd(struct consumer_relayd_sock_pair *relayd) iter.iter.node = &relayd->node.node; ret = lttng_ht_del(consumer_data.relayd_ht, &iter); if (ret != 0) { - /* We assume the relayd was already destroyed */ + /* We assume the relayd is being or is destroyed */ return; } - /* Close all sockets */ - pthread_mutex_lock(&relayd->ctrl_sock_mutex); - (void) relayd_close(&relayd->control_sock); - pthread_mutex_unlock(&relayd->ctrl_sock_mutex); - (void) relayd_close(&relayd->data_sock); - /* RCU free() call */ call_rcu(&relayd->node.head, consumer_rcu_free_relayd); } +/* + * Update the end point status of all streams having the given network sequence + * index (relayd index). + * + * It's atomically set without having the stream mutex locked which is fine + * because we handle the write/read race with a pipe wakeup for each thread. + */ +static void update_endpoint_status_by_netidx(int net_seq_idx, + enum consumer_endpoint_status status) +{ + struct lttng_ht_iter iter; + struct lttng_consumer_stream *stream; + + DBG("Consumer set delete flag on stream by idx %d", net_seq_idx); + + rcu_read_lock(); + + /* Let's begin with metadata */ + cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) { + if (stream->net_seq_idx == net_seq_idx) { + uatomic_set(&stream->endpoint_status, status); + DBG("Delete flag set to metadata stream %d", stream->wait_fd); + } + } + + /* Follow up by the data streams */ + cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) { + if (stream->net_seq_idx == net_seq_idx) { + uatomic_set(&stream->endpoint_status, status); + DBG("Delete flag set to data stream %d", stream->wait_fd); + } + } + rcu_read_unlock(); +} + +/* + * Cleanup a relayd object by flagging every associated streams for deletion, + * destroying the object meaning removing it from the relayd hash table, + * closing the sockets and freeing the memory in a RCU call. + * + * If a local data context is available, notify the threads that the streams' + * state have changed. + */ +static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd, + struct lttng_consumer_local_data *ctx) +{ + int netidx; + + assert(relayd); + + /* Save the net sequence index before destroying the object */ + netidx = relayd->net_seq_idx; + + /* + * Delete the relayd from the relayd hash table, close the sockets and free + * the object in a RCU call. + */ + destroy_relayd(relayd); + + /* Set inactive endpoint to all streams */ + update_endpoint_status_by_netidx(netidx, CONSUMER_ENDPOINT_INACTIVE); + + /* + * With a local data context, notify the threads that the streams' state + * have changed. The write() action on the pipe acts as an "implicit" + * memory barrier ordering the updates of the end point status from the + * read of this status which happens AFTER receiving this notify. + */ + if (ctx) { + notify_thread_pipe(ctx->consumer_data_pipe[1]); + notify_thread_pipe(ctx->consumer_metadata_pipe[1]); + } +} + /* * Flag a relayd socket pair for destruction. Destroy it if the refcount * reaches zero. @@ -262,11 +345,14 @@ void consumer_del_stream(struct lttng_consumer_stream *stream, assert(stream); + DBG("Consumer del stream %d", stream->wait_fd); + if (ht == NULL) { /* Means the stream was allocated but not successfully added */ goto free_stream; } + pthread_mutex_lock(&stream->lock); pthread_mutex_lock(&consumer_data.lock); switch (consumer_data.type) { @@ -293,15 +379,15 @@ void consumer_del_stream(struct lttng_consumer_stream *stream, ret = lttng_ht_del(ht, &iter); assert(!ret); + /* Remove node session id from the consumer_data stream ht */ + iter.iter.node = &stream->node_session_id.node; + ret = lttng_ht_del(consumer_data.stream_list_ht, &iter); + assert(!ret); rcu_read_unlock(); - if (consumer_data.stream_count <= 0) { - goto end; - } + assert(consumer_data.stream_count > 0); consumer_data.stream_count--; - if (!stream) { - goto end; - } + if (stream->out_fd >= 0) { ret = close(stream->out_fd); if (ret) { @@ -360,6 +446,7 @@ void consumer_del_stream(struct lttng_consumer_stream *stream, end: consumer_data.need_update = 1; pthread_mutex_unlock(&consumer_data.lock); + pthread_mutex_unlock(&stream->lock); if (free_chan) { consumer_del_channel(free_chan); @@ -380,6 +467,7 @@ struct lttng_consumer_stream *consumer_allocate_stream( gid_t gid, int net_index, int metadata_flag, + uint64_t session_id, int *alloc_ret) { struct lttng_consumer_stream *stream; @@ -415,10 +503,24 @@ struct lttng_consumer_stream *consumer_allocate_stream( stream->gid = gid; stream->net_seq_idx = net_index; stream->metadata_flag = metadata_flag; + stream->session_id = session_id; strncpy(stream->path_name, path_name, sizeof(stream->path_name)); stream->path_name[sizeof(stream->path_name) - 1] = '\0'; - lttng_ht_node_init_ulong(&stream->waitfd_node, stream->wait_fd); - lttng_ht_node_init_ulong(&stream->node, stream->key); + pthread_mutex_init(&stream->lock, NULL); + + /* + * Index differently the metadata node because the thread is using an + * internal hash table to match streams in the metadata_ht to the epoll set + * file descriptor. + */ + if (metadata_flag) { + lttng_ht_node_init_ulong(&stream->node, stream->wait_fd); + } else { + lttng_ht_node_init_ulong(&stream->node, stream->key); + } + + /* Init session id node with the stream session id */ + lttng_ht_node_init_ulong(&stream->node_session_id, stream->session_id); /* * The cpu number is needed before using any ustctl_* actions. Ignored for @@ -429,10 +531,10 @@ struct lttng_consumer_stream *consumer_allocate_stream( pthread_mutex_unlock(&consumer_data.lock); DBG3("Allocated stream %s (key %d, shm_fd %d, wait_fd %d, mmap_len %llu," - " out_fd %d, net_seq_idx %d)", stream->path_name, stream->key, - stream->shm_fd, stream->wait_fd, + " out_fd %d, net_seq_idx %d, session_id %" PRIu64, + stream->path_name, stream->key, stream->shm_fd, stream->wait_fd, (unsigned long long) stream->mmap_len, stream->out_fd, - stream->net_seq_idx); + stream->net_seq_idx, stream->session_id); return stream; error: @@ -463,6 +565,13 @@ static int consumer_add_stream(struct lttng_consumer_stream *stream, lttng_ht_add_unique_ulong(ht, &stream->node); + /* + * Add stream to the stream_list_ht of the consumer data. No need to steal + * the key since the HT does not use it and we allow to add redundant keys + * into this table. + */ + lttng_ht_add_ulong(consumer_data.stream_list_ht, &stream->node_session_id); + /* Check and cleanup relayd */ relayd = consumer_find_relayd(stream->net_seq_idx); if (relayd != NULL) { @@ -630,23 +739,6 @@ error: return outfd; } -/* - * Update a stream according to what we just received. - */ -void consumer_change_stream_state(int stream_key, - enum lttng_consumer_stream_state state) -{ - struct lttng_consumer_stream *stream; - - pthread_mutex_lock(&consumer_data.lock); - stream = consumer_find_stream(stream_key, consumer_data.stream_ht); - if (stream) { - stream->state = state; - } - consumer_data.need_update = 1; - pthread_mutex_unlock(&consumer_data.lock); -} - static void consumer_free_channel(struct rcu_head *head) { @@ -810,7 +902,17 @@ static int consumer_update_poll_array( DBG("Updating poll fd array"); rcu_read_lock(); cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { - if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM) { + /* + * Only active streams with an active end point can be added to the + * poll set and local stream storage of the thread. + * + * There is a potential race here for endpoint_status to be updated + * just after the check. However, this is OK since the stream(s) will + * be deleted once the thread is notified that the end point state has + * changed where this function will be called back again. + */ + if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM || + stream->endpoint_status) { continue; } DBG("Active FD %d", stream->wait_fd); @@ -822,10 +924,10 @@ static int consumer_update_poll_array( rcu_read_unlock(); /* - * Insert the consumer_poll_pipe at the end of the array and don't + * Insert the consumer_data_pipe at the end of the array and don't * increment i so nb_fd is the number of real FD. */ - (*pollfd)[i].fd = ctx->consumer_poll_pipe[0]; + (*pollfd)[i].fd = ctx->consumer_data_pipe[0]; (*pollfd)[i].events = POLLIN | POLLPRI; return i; } @@ -904,17 +1006,6 @@ void lttng_consumer_cleanup(void) rcu_read_lock(); - /* - * close all outfd. Called when there are no more threads running (after - * joining on the threads), no need to protect list iteration with mutex. - */ - cds_lfht_for_each_entry(consumer_data.stream_ht->ht, &iter.iter, node, - node) { - struct lttng_consumer_stream *stream = - caa_container_of(node, struct lttng_consumer_stream, node); - consumer_del_stream(stream, consumer_data.stream_ht); - } - cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, node, node) { struct lttng_consumer_channel *channel = @@ -924,7 +1015,6 @@ void lttng_consumer_cleanup(void) rcu_read_unlock(); - lttng_ht_destroy(consumer_data.stream_ht); lttng_ht_destroy(consumer_data.channel_ht); } @@ -941,6 +1031,8 @@ void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx) if (ret < 0) { PERROR("write consumer quit"); } + + DBG("Consumer flag that it should quit"); } void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream, @@ -1022,21 +1114,21 @@ struct lttng_consumer_local_data *lttng_consumer_create( ctx->on_recv_stream = recv_stream; ctx->on_update_stream = update_stream; - ret = pipe(ctx->consumer_poll_pipe); + ret = pipe(ctx->consumer_data_pipe); if (ret < 0) { PERROR("Error creating poll pipe"); goto error_poll_pipe; } /* set read end of the pipe to non-blocking */ - ret = fcntl(ctx->consumer_poll_pipe[0], F_SETFL, O_NONBLOCK); + ret = fcntl(ctx->consumer_data_pipe[0], F_SETFL, O_NONBLOCK); if (ret < 0) { PERROR("fcntl O_NONBLOCK"); goto error_poll_fcntl; } /* set write end of the pipe to non-blocking */ - ret = fcntl(ctx->consumer_poll_pipe[1], F_SETFL, O_NONBLOCK); + ret = fcntl(ctx->consumer_data_pipe[1], F_SETFL, O_NONBLOCK); if (ret < 0) { PERROR("fcntl O_NONBLOCK"); goto error_poll_fcntl; @@ -1084,7 +1176,7 @@ error_quit_pipe: for (i = 0; i < 2; i++) { int err; - err = close(ctx->consumer_poll_pipe[i]); + err = close(ctx->consumer_data_pipe[i]); if (err) { PERROR("close"); } @@ -1102,6 +1194,8 @@ void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx) { int ret; + DBG("Consumer destroying it. Closing everything."); + ret = close(ctx->consumer_error_socket); if (ret) { PERROR("close"); @@ -1114,11 +1208,11 @@ void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx) if (ret) { PERROR("close"); } - ret = close(ctx->consumer_poll_pipe[0]); + ret = close(ctx->consumer_data_pipe[0]); if (ret) { PERROR("close"); } - ret = close(ctx->consumer_poll_pipe[1]); + ret = close(ctx->consumer_data_pipe[1]); if (ret) { PERROR("close"); } @@ -1183,10 +1277,13 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( /* Default is on the disk */ int outfd = stream->out_fd; struct consumer_relayd_sock_pair *relayd = NULL; + unsigned int relayd_hang_up = 0; /* RCU lock for the relayd pointer */ rcu_read_lock(); + pthread_mutex_lock(&stream->lock); + /* Flag that the current stream if set for network streaming. */ if (stream->net_seq_idx != -1) { relayd = consumer_find_relayd(stream->net_seq_idx); @@ -1240,11 +1337,22 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( ret = write_relayd_metadata_id(outfd, stream, relayd, padding); if (ret < 0) { written = ret; + /* Socket operation failed. We consider the relayd dead */ + if (ret == -EPIPE || ret == -EINVAL) { + relayd_hang_up = 1; + goto write_error; + } goto end; } } + } else { + /* Socket operation failed. We consider the relayd dead */ + if (ret == -EPIPE || ret == -EINVAL) { + relayd_hang_up = 1; + goto write_error; + } + /* Else, use the default set before which is the filesystem. */ } - /* Else, use the default set before which is the filesystem. */ } else { /* No streaming, we have to set the len with the full padding */ len += padding; @@ -1260,6 +1368,11 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( if (written == 0) { written = ret; } + /* Socket operation failed. We consider the relayd dead */ + if (errno == EPIPE || errno == EINVAL) { + relayd_hang_up = 1; + goto write_error; + } goto end; } else if (ret > len) { PERROR("Error in file write (ret %zd > len %lu)", ret, len); @@ -1281,11 +1394,21 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( } lttng_consumer_sync_trace_file(stream, orig_offset); +write_error: + /* + * This is a special case that the relayd has closed its socket. Let's + * cleanup the relayd object and all associated streams. + */ + if (relayd && relayd_hang_up) { + cleanup_relayd(relayd, ctx); + } + end: /* Unlock only if ctrl socket used */ if (relayd && stream->metadata_flag) { pthread_mutex_unlock(&relayd->ctrl_sock_mutex); } + pthread_mutex_unlock(&stream->lock); rcu_read_unlock(); return written; @@ -1309,6 +1432,7 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( int outfd = stream->out_fd; struct consumer_relayd_sock_pair *relayd = NULL; int *splice_pipe; + unsigned int relayd_hang_up = 0; switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: @@ -1325,6 +1449,8 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( /* RCU lock for the relayd pointer */ rcu_read_lock(); + pthread_mutex_lock(&stream->lock); + /* Flag that the current stream if set for network streaming. */ if (stream->net_seq_idx != -1) { relayd = consumer_find_relayd(stream->net_seq_idx); @@ -1359,6 +1485,12 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( padding); if (ret < 0) { written = ret; + /* Socket operation failed. We consider the relayd dead */ + if (ret == -EBADF) { + WARN("Remote relayd disconnected. Stopping"); + relayd_hang_up = 1; + goto write_error; + } goto end; } @@ -1370,7 +1502,12 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( /* Use the returned socket. */ outfd = ret; } else { - ERR("Remote relayd disconnected. Stopping"); + /* Socket operation failed. We consider the relayd dead */ + if (ret == -EBADF) { + WARN("Remote relayd disconnected. Stopping"); + relayd_hang_up = 1; + goto write_error; + } goto end; } } else { @@ -1419,6 +1556,12 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( if (written == 0) { written = ret_splice; } + /* Socket operation failed. We consider the relayd dead */ + if (errno == EBADF) { + WARN("Remote relayd disconnected. Stopping"); + relayd_hang_up = 1; + goto write_error; + } ret = errno; goto splice_error; } else if (ret_splice > len) { @@ -1446,12 +1589,20 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( goto end; +write_error: + /* + * This is a special case that the relayd has closed its socket. Let's + * cleanup the relayd object and all associated streams. + */ + if (relayd && relayd_hang_up) { + cleanup_relayd(relayd, ctx); + /* Skip splice error so the consumer does not fail */ + goto end; + } + splice_error: /* send the appropriate error description to sessiond */ switch (ret) { - case EBADF: - lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EBADF); - break; case EINVAL: lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL); break; @@ -1467,6 +1618,7 @@ end: if (relayd && stream->metadata_flag) { pthread_mutex_unlock(&relayd->ctrl_sock_mutex); } + pthread_mutex_unlock(&stream->lock); rcu_read_unlock(); return written; @@ -1576,11 +1728,11 @@ static void destroy_stream_ht(struct lttng_ht *ht) } rcu_read_lock(); - cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, waitfd_node.node) { + cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { ret = lttng_ht_del(ht, &iter); assert(!ret); - call_rcu(&stream->waitfd_node.head, consumer_free_metadata_stream); + call_rcu(&stream->node.head, consumer_free_stream); } rcu_read_unlock(); @@ -1612,6 +1764,8 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream, goto free_stream; } + pthread_mutex_lock(&stream->lock); + pthread_mutex_lock(&consumer_data.lock); switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: @@ -1633,9 +1787,14 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream, } rcu_read_lock(); - iter.iter.node = &stream->waitfd_node.node; + iter.iter.node = &stream->node.node; ret = lttng_ht_del(ht, &iter); assert(!ret); + + /* Remove node session id from the consumer_data stream ht */ + iter.iter.node = &stream->node_session_id.node; + ret = lttng_ht_del(consumer_data.stream_list_ht, &iter); + assert(!ret); rcu_read_unlock(); if (stream->out_fd >= 0) { @@ -1698,13 +1857,14 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream, end: pthread_mutex_unlock(&consumer_data.lock); + pthread_mutex_unlock(&stream->lock); if (free_chan) { consumer_del_channel(free_chan); } free_stream: - call_rcu(&stream->waitfd_node.head, consumer_free_metadata_stream); + call_rcu(&stream->node.head, consumer_free_stream); } /* @@ -1753,13 +1913,74 @@ static int consumer_add_metadata_stream(struct lttng_consumer_stream *stream, /* Steal stream identifier to avoid having streams with the same key */ consumer_steal_stream_key(stream->key, ht); - lttng_ht_add_unique_ulong(ht, &stream->waitfd_node); + lttng_ht_add_unique_ulong(ht, &stream->node); + + /* + * Add stream to the stream_list_ht of the consumer data. No need to steal + * the key since the HT does not use it and we allow to add redundant keys + * into this table. + */ + lttng_ht_add_ulong(consumer_data.stream_list_ht, &stream->node_session_id); + rcu_read_unlock(); pthread_mutex_unlock(&consumer_data.lock); return ret; } +/* + * Delete data stream that are flagged for deletion (endpoint_status). + */ +static void validate_endpoint_status_data_stream(void) +{ + struct lttng_ht_iter iter; + struct lttng_consumer_stream *stream; + + DBG("Consumer delete flagged data stream"); + + rcu_read_lock(); + cds_lfht_for_each_entry(data_ht->ht, &iter.iter, stream, node.node) { + /* Validate delete flag of the stream */ + if (!stream->endpoint_status) { + continue; + } + /* Delete it right now */ + consumer_del_stream(stream, data_ht); + } + rcu_read_unlock(); +} + +/* + * Delete metadata stream that are flagged for deletion (endpoint_status). + */ +static void validate_endpoint_status_metadata_stream( + struct lttng_poll_event *pollset) +{ + struct lttng_ht_iter iter; + struct lttng_consumer_stream *stream; + + DBG("Consumer delete flagged metadata stream"); + + assert(pollset); + + rcu_read_lock(); + cds_lfht_for_each_entry(metadata_ht->ht, &iter.iter, stream, node.node) { + /* Validate delete flag of the stream */ + if (!stream->endpoint_status) { + continue; + } + /* + * Remove from pollset so the metadata thread can continue without + * blocking on a deleted stream. + */ + lttng_poll_del(pollset, stream->wait_fd); + + /* Delete it right now */ + consumer_del_metadata_stream(stream, metadata_ht); + } + rcu_read_unlock(); +} + /* * Thread polls on metadata file descriptor and write them on disk or on the * network. @@ -1851,6 +2072,13 @@ restart: continue; } + /* A NULL stream means that the state has changed. */ + if (stream == NULL) { + /* Check for deleted streams. */ + validate_endpoint_status_metadata_stream(&events); + continue; + } + DBG("Adding metadata stream %d to poll set", stream->wait_fd); @@ -1878,7 +2106,7 @@ restart: assert(node); stream = caa_container_of(node, struct lttng_consumer_stream, - waitfd_node); + node); /* Check for error event */ if (revents & (LPOLLERR | LPOLLHUP)) { @@ -1915,8 +2143,9 @@ restart: len = ctx->on_buffer_ready(stream, ctx); /* It's ok to have an unavailable sub-buffer */ if (len < 0 && len != -EAGAIN && len != -ENODATA) { - rcu_read_unlock(); - goto end; + /* Clean up stream from consumer and free it. */ + lttng_poll_del(&events, stream->wait_fd); + consumer_del_metadata_stream(stream, metadata_ht); } else if (len > 0) { stream->data_read = 1; } @@ -1983,7 +2212,7 @@ void *consumer_thread_data_poll(void *data) local_stream = NULL; } - /* allocate for all fds + 1 for the consumer_poll_pipe */ + /* allocate for all fds + 1 for the consumer_data_pipe */ pollfd = zmalloc((consumer_data.stream_count + 1) * sizeof(struct pollfd)); if (pollfd == NULL) { PERROR("pollfd malloc"); @@ -1991,7 +2220,7 @@ void *consumer_thread_data_poll(void *data) goto end; } - /* allocate for all fds + 1 for the consumer_poll_pipe */ + /* allocate for all fds + 1 for the consumer_data_pipe */ local_stream = zmalloc((consumer_data.stream_count + 1) * sizeof(struct lttng_consumer_stream)); if (local_stream == NULL) { @@ -2037,17 +2266,17 @@ void *consumer_thread_data_poll(void *data) } /* - * If the consumer_poll_pipe triggered poll go directly to the + * If the consumer_data_pipe triggered poll go directly to the * beginning of the loop to update the array. We want to prioritize * array update over low-priority reads. */ if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) { size_t pipe_readlen; - DBG("consumer_poll_pipe wake up"); + DBG("consumer_data_pipe wake up"); /* Consume 1 byte of pipe data */ do { - pipe_readlen = read(ctx->consumer_poll_pipe[0], &new_stream, + pipe_readlen = read(ctx->consumer_data_pipe[0], &new_stream, sizeof(new_stream)); } while (pipe_readlen == -1 && errno == EINTR); @@ -2057,6 +2286,7 @@ void *consumer_thread_data_poll(void *data) * waking us up to test it. */ if (new_stream == NULL) { + validate_endpoint_status_data_stream(); continue; } @@ -2083,7 +2313,8 @@ void *consumer_thread_data_poll(void *data) len = ctx->on_buffer_ready(local_stream[i], ctx); /* it's ok to have an unavailable sub-buffer */ if (len < 0 && len != -EAGAIN && len != -ENODATA) { - goto end; + /* Clean the stream and free it. */ + consumer_del_stream(local_stream[i], data_ht); } else if (len > 0) { local_stream[i]->data_read = 1; } @@ -2106,7 +2337,8 @@ void *consumer_thread_data_poll(void *data) len = ctx->on_buffer_ready(local_stream[i], ctx); /* it's ok to have an unavailable sub-buffer */ if (len < 0 && len != -EAGAIN && len != -ENODATA) { - goto end; + /* Clean the stream and free it. */ + consumer_del_stream(local_stream[i], data_ht); } else if (len > 0) { local_stream[i]->data_read = 1; } @@ -2293,14 +2525,9 @@ end: /* * Notify the data poll thread to poll back again and test the - * consumer_quit state to quit gracefully. + * consumer_quit state that we just set so to quit gracefully. */ - do { - struct lttng_consumer_stream *null_stream = NULL; - - ret = write(ctx->consumer_poll_pipe[1], &null_stream, - sizeof(null_stream)); - } while (ret < 0 && errno == EINTR); + notify_thread_pipe(ctx->consumer_data_pipe[1]); rcu_unregister_thread(); return NULL; @@ -2342,9 +2569,9 @@ int lttng_consumer_on_recv_stream(struct lttng_consumer_stream *stream) */ void lttng_consumer_init(void) { - consumer_data.stream_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); + consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); assert(metadata_ht); @@ -2443,3 +2670,142 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, error: return ret; } + +/* + * Try to lock the stream mutex. + * + * On success, 1 is returned else 0 indicating that the mutex is NOT lock. + */ +static int stream_try_lock(struct lttng_consumer_stream *stream) +{ + int ret; + + assert(stream); + + /* + * Try to lock the stream mutex. On failure, we know that the stream is + * being used else where hence there is data still being extracted. + */ + ret = pthread_mutex_trylock(&stream->lock); + if (ret) { + /* For both EBUSY and EINVAL error, the mutex is NOT locked. */ + ret = 0; + goto end; + } + + ret = 1; + +end: + return ret; +} + +/* + * Check if for a given session id there is still data needed to be extract + * from the buffers. + * + * Return 1 if data is in fact available to be read or else 0. + */ +int consumer_data_available(uint64_t id) +{ + int ret; + struct lttng_ht_iter iter; + struct lttng_ht *ht; + struct lttng_consumer_stream *stream; + struct consumer_relayd_sock_pair *relayd; + int (*data_available)(struct lttng_consumer_stream *); + + DBG("Consumer data available command on session id %" PRIu64, id); + + rcu_read_lock(); + pthread_mutex_lock(&consumer_data.lock); + + switch (consumer_data.type) { + case LTTNG_CONSUMER_KERNEL: + data_available = lttng_kconsumer_data_available; + break; + case LTTNG_CONSUMER32_UST: + case LTTNG_CONSUMER64_UST: + data_available = lttng_ustconsumer_data_available; + break; + default: + ERR("Unknown consumer data type"); + assert(0); + } + + /* Ease our life a bit */ + ht = consumer_data.stream_list_ht; + + cds_lfht_for_each_entry_duplicate(ht->ht, + ht->hash_fct((void *)((unsigned long) id), 0x42UL), + ht->match_fct, (void *)((unsigned long) id), + &iter.iter, stream, node_session_id.node) { + /* If this call fails, the stream is being used hence data pending. */ + ret = stream_try_lock(stream); + if (!ret) { + goto data_not_available; + } + + /* + * A removed node from the hash table indicates that the stream has + * been deleted thus having a guarantee that the buffers are closed + * on the consumer side. However, data can still be transmitted + * over the network so don't skip the relayd check. + */ + ret = cds_lfht_is_node_deleted(&stream->node.node); + if (!ret) { + /* Check the stream if there is data in the buffers. */ + ret = data_available(stream); + if (ret == 0) { + pthread_mutex_unlock(&stream->lock); + goto data_not_available; + } + } + + /* Relayd check */ + if (stream->net_seq_idx != -1) { + relayd = consumer_find_relayd(stream->net_seq_idx); + if (!relayd) { + /* + * At this point, if the relayd object is not available for the + * given stream, it is because the relayd is being cleaned up + * so every stream associated with it (for a session id value) + * are or will be marked for deletion hence no data pending. + */ + pthread_mutex_unlock(&stream->lock); + goto data_not_available; + } + + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + if (stream->metadata_flag) { + ret = relayd_quiescent_control(&relayd->control_sock); + } else { + ret = relayd_data_available(&relayd->control_sock, + stream->relayd_stream_id, stream->next_net_seq_num); + } + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); + if (ret == 0) { + pthread_mutex_unlock(&stream->lock); + goto data_not_available; + } + } + pthread_mutex_unlock(&stream->lock); + } + + /* + * Finding _no_ node in the hash table means that the stream(s) have been + * removed thus data is guaranteed to be available for analysis from the + * trace files. This is *only* true for local consumer and not network + * streaming. + */ + + /* Data is available to be read by a viewer. */ + pthread_mutex_unlock(&consumer_data.lock); + rcu_read_unlock(); + return 1; + +data_not_available: + /* Data is still being extracted from buffers. */ + pthread_mutex_unlock(&consumer_data.lock); + rcu_read_unlock(); + return 0; +}