X-Git-Url: https://git.lttng.org/?p=lttng-tools.git;a=blobdiff_plain;f=src%2Fcommon%2Fconsumer.c;h=9df0c484c9bfe4c0a29371ed473b07e846f93e77;hp=b55fd153d7830e1d817ed4144dbd65fdf776b00b;hb=18261bd1f606ed31dd91cf82bbb91af7839573f5;hpb=f724d81e402a1eb319ce0392c5b6770787a28edc diff --git a/src/common/consumer.c b/src/common/consumer.c index b55fd153d..9df0c484c 100644 --- a/src/common/consumer.c +++ b/src/common/consumer.c @@ -47,27 +47,65 @@ struct lttng_consumer_global_data consumer_data = { .type = LTTNG_CONSUMER_UNKNOWN, }; -/* timeout parameter, to control the polling thread grace period. */ -int consumer_poll_timeout = -1; - /* * Flag to inform the polling thread to quit when all fd hung up. Updated by * the consumer_thread_receive_fds when it notices that all fds has hung up. * Also updated by the signal handler (consumer_should_exit()). Read by the * polling threads. */ -volatile int consumer_quit = 0; +volatile int consumer_quit; + +/* + * Global hash table containing respectively metadata and data streams. The + * 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. + */ +static struct lttng_ht *metadata_ht; +static struct lttng_ht *data_ht; + +/* + * This hash table contains the mapping between the session id of the sessiond + * and the relayd session id. Element of the ht are indexed by sessiond session + * id. + * + * Node can be added when a relayd communication is opened in the sessiond + * thread. + * + * Note that a session id of the session daemon is unique to a tracing session + * and not to a domain session. However, a domain session has one consumer + * which forces the 1-1 mapping between a consumer and a domain session (ex: + * UST). This means that we can't have duplicate in this ht. + */ +static struct lttng_ht *relayd_session_id_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 * call. */ -static struct lttng_consumer_stream *consumer_find_stream(int key) +static struct lttng_consumer_stream *consumer_find_stream(int key, + struct lttng_ht *ht) { struct lttng_ht_iter iter; struct lttng_ht_node_ulong *node; struct lttng_consumer_stream *stream = NULL; + assert(ht); + /* Negative keys are lookup failures */ if (key < 0) { return NULL; @@ -75,8 +113,7 @@ static struct lttng_consumer_stream *consumer_find_stream(int key) rcu_read_lock(); - lttng_ht_lookup(consumer_data.stream_ht, (void *)((unsigned long) key), - &iter); + lttng_ht_lookup(ht, (void *)((unsigned long) key), &iter); node = lttng_ht_iter_get_node_ulong(&iter); if (node != NULL) { stream = caa_container_of(node, struct lttng_consumer_stream, node); @@ -87,12 +124,12 @@ static struct lttng_consumer_stream *consumer_find_stream(int key) return stream; } -static void consumer_steal_stream_key(int key) +void consumer_steal_stream_key(int key, struct lttng_ht *ht) { struct lttng_consumer_stream *stream; rcu_read_lock(); - stream = consumer_find_stream(key); + stream = consumer_find_stream(key, ht); if (stream) { stream->key = -1; /* @@ -105,6 +142,12 @@ static void consumer_steal_stream_key(int key) rcu_read_unlock(); } +/* + * Return a channel object for the given key. + * + * RCU read side lock MUST be acquired before calling this function and + * protects the channel ptr. + */ static struct lttng_consumer_channel *consumer_find_channel(int key) { struct lttng_ht_iter iter; @@ -116,8 +159,6 @@ static struct lttng_consumer_channel *consumer_find_channel(int key) return NULL; } - rcu_read_lock(); - lttng_ht_lookup(consumer_data.channel_ht, (void *)((unsigned long) key), &iter); node = lttng_ht_iter_get_node_ulong(&iter); @@ -125,8 +166,6 @@ static struct lttng_consumer_channel *consumer_find_channel(int key) channel = caa_container_of(node, struct lttng_consumer_channel, node); } - rcu_read_unlock(); - return channel; } @@ -169,6 +208,17 @@ 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); } @@ -181,6 +231,7 @@ static void destroy_relayd(struct consumer_relayd_sock_pair *relayd) { int ret; struct lttng_ht_iter iter; + struct lttng_ht_node_ulong *node; if (relayd == NULL) { return; @@ -188,23 +239,132 @@ static void destroy_relayd(struct consumer_relayd_sock_pair *relayd) DBG("Consumer destroy and close relayd socket pair"); + /* Loockup for a relayd node in the session id map hash table. */ + lttng_ht_lookup(relayd_session_id_ht, + (void *)((unsigned long) relayd->sessiond_session_id), &iter); + node = lttng_ht_iter_get_node_ulong(&iter); + if (node == NULL) { + /* We assume the relayd is being or is destroyed */ + return; + } + + /* + * Try to delete it from the relayd session id ht. The return value is of + * no importance since either way we are going to try to delete the relayd + * from the global relayd_ht. + */ + lttng_ht_del(relayd_session_id_ht, &iter); + 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); } +/* + * Iterate over the relayd hash table and destroy each element. Finally, + * destroy the whole hash table. + */ +static void cleanup_relayd_ht(void) +{ + struct lttng_ht_iter iter; + struct consumer_relayd_sock_pair *relayd; + + rcu_read_lock(); + + cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd, + node.node) { + destroy_relayd(relayd); + } + + lttng_ht_destroy(consumer_data.relayd_ht); + /* The destroy_relayd call makes sure that this ht is empty here. */ + lttng_ht_destroy(relayd_session_id_ht); + + rcu_read_unlock(); +} + +/* + * 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); + + DBG("Cleaning up relayd sockets"); + + /* 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. @@ -228,7 +388,8 @@ void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd) * Remove a stream from the global list protected by a mutex. This * function is also responsible for freeing its data structures. */ -void consumer_del_stream(struct lttng_consumer_stream *stream) +void consumer_del_stream(struct lttng_consumer_stream *stream, + struct lttng_ht *ht) { int ret; struct lttng_ht_iter iter; @@ -237,7 +398,15 @@ 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(&consumer_data.lock); + pthread_mutex_lock(&stream->lock); switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: @@ -260,18 +429,18 @@ void consumer_del_stream(struct lttng_consumer_stream *stream) rcu_read_lock(); iter.iter.node = &stream->node.node; - ret = lttng_ht_del(consumer_data.stream_ht, &iter); + 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) { @@ -327,14 +496,17 @@ void consumer_del_stream(struct lttng_consumer_stream *stream) free_chan = stream->chan; } - call_rcu(&stream->node.head, consumer_free_stream); end: consumer_data.need_update = 1; + pthread_mutex_unlock(&stream->lock); pthread_mutex_unlock(&consumer_data.lock); if (free_chan) { consumer_del_channel(free_chan); } + +free_stream: + call_rcu(&stream->node.head, consumer_free_stream); } struct lttng_consumer_stream *consumer_allocate_stream( @@ -348,10 +520,10 @@ 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; - int ret; stream = zmalloc(sizeof(*stream)); if (stream == NULL) { @@ -360,6 +532,8 @@ struct lttng_consumer_stream *consumer_allocate_stream( goto end; } + rcu_read_lock(); + /* * Get stream's channel reference. Needed when adding the stream to the * global hash table. @@ -370,7 +544,7 @@ struct lttng_consumer_stream *consumer_allocate_stream( ERR("Unable to find channel for stream %d", stream_key); goto error; } - stream->chan->refcount++; + stream->key = stream_key; stream->shm_fd = shm_fd; stream->wait_fd = wait_fd; @@ -384,48 +558,44 @@ 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->node, stream->key); - lttng_ht_node_init_ulong(&stream->waitfd_node, stream->wait_fd); + pthread_mutex_init(&stream->lock, NULL); - switch (consumer_data.type) { - case LTTNG_CONSUMER_KERNEL: - break; - case LTTNG_CONSUMER32_UST: - case LTTNG_CONSUMER64_UST: - stream->cpu = stream->chan->cpucount++; - ret = lttng_ustconsumer_allocate_stream(stream); - if (ret) { - *alloc_ret = -EINVAL; - goto error; - } - break; - default: - ERR("Unknown consumer_data type"); - *alloc_ret = -EINVAL; - goto error; + /* + * 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); + /* - * When nb_init_streams reaches 0, we don't need to trigger any action in - * terms of destroying the associated channel, because the action that - * causes the count to become 0 also causes a stream to be added. The - * channel deletion will thus be triggered by the following removal of this - * stream. + * The cpu number is needed before using any ustctl_* actions. Ignored for + * the kernel so the value does not matter. */ - if (uatomic_read(&stream->chan->nb_init_streams) > 0) { - uatomic_dec(&stream->chan->nb_init_streams); - } + pthread_mutex_lock(&consumer_data.lock); + stream->cpu = stream->chan->cpucount++; + 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); + + rcu_read_unlock(); return stream; error: + rcu_read_unlock(); free(stream); end: return NULL; @@ -434,41 +604,59 @@ end: /* * Add a stream to the global list protected by a mutex. */ -int consumer_add_stream(struct lttng_consumer_stream *stream) +static int consumer_add_stream(struct lttng_consumer_stream *stream, + struct lttng_ht *ht) { int ret = 0; - struct lttng_ht_node_ulong *node; - struct lttng_ht_iter iter; struct consumer_relayd_sock_pair *relayd; - pthread_mutex_lock(&consumer_data.lock); - /* Steal stream identifier, for UST */ - consumer_steal_stream_key(stream->key); + assert(stream); + assert(ht); + DBG3("Adding consumer stream %d", stream->key); + + pthread_mutex_lock(&consumer_data.lock); + pthread_mutex_lock(&stream->lock); rcu_read_lock(); - lttng_ht_lookup(consumer_data.stream_ht, - (void *)((unsigned long) stream->key), &iter); - node = lttng_ht_iter_get_node_ulong(&iter); - if (node != NULL) { - rcu_read_unlock(); - /* Stream already exist. Ignore the insertion */ - goto end; - } - lttng_ht_add_unique_ulong(consumer_data.stream_ht, &stream->node); + /* 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->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) { uatomic_inc(&relayd->refcount); } - rcu_read_unlock(); - /* Update consumer data */ + /* Update channel refcount once added without error(s). */ + uatomic_inc(&stream->chan->refcount); + + /* + * When nb_init_streams reaches 0, we don't need to trigger any action in + * terms of destroying the associated channel, because the action that + * causes the count to become 0 also causes a stream to be added. The + * channel deletion will thus be triggered by the following removal of this + * stream. + */ + if (uatomic_read(&stream->chan->nb_init_streams) > 0) { + uatomic_dec(&stream->chan->nb_init_streams); + } + + /* Update consumer data once the node is inserted. */ consumer_data.stream_count++; consumer_data.need_update = 1; -end: + rcu_read_unlock(); + pthread_mutex_unlock(&stream->lock); pthread_mutex_unlock(&consumer_data.lock); return ret; @@ -594,6 +782,13 @@ static int write_relayd_stream_header(struct lttng_consumer_stream *stream, data_hdr.stream_id = htobe64(stream->relayd_stream_id); data_hdr.data_size = htobe32(data_size); data_hdr.padding_size = htobe32(padding); + /* + * Note that net_seq_num below is assigned with the *current* value of + * next_net_seq_num and only after that the next_net_seq_num will be + * increment. This is why when issuing a command on the relayd using + * this next value, 1 should always be substracted in order to compare + * the last seen sequence number on the relayd side to the last sent. + */ data_hdr.net_seq_num = htobe64(stream->next_net_seq_num++); /* Other fields are zeroed previously */ @@ -611,23 +806,6 @@ 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); - 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) { @@ -648,6 +826,8 @@ void consumer_del_channel(struct lttng_consumer_channel *channel) int ret; struct lttng_ht_iter iter; + DBG("Consumer delete channel key %d", channel->key); + pthread_mutex_lock(&consumer_data.lock); switch (consumer_data.type) { @@ -780,9 +960,9 @@ end: * * Returns the number of fds in the structures. */ -int consumer_update_poll_array( +static int consumer_update_poll_array( struct lttng_consumer_local_data *ctx, struct pollfd **pollfd, - struct lttng_consumer_stream **local_stream) + struct lttng_consumer_stream **local_stream, struct lttng_ht *ht) { int i = 0; struct lttng_ht_iter iter; @@ -790,9 +970,18 @@ int consumer_update_poll_array( DBG("Updating poll fd array"); rcu_read_lock(); - cds_lfht_for_each_entry(consumer_data.stream_ht->ht, &iter.iter, stream, - node.node) { - if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM) { + cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { + /* + * 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 == CONSUMER_ENDPOINT_INACTIVE) { continue; } DBG("Active FD %d", stream->wait_fd); @@ -804,10 +993,10 @@ 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; } @@ -876,8 +1065,8 @@ int lttng_consumer_send_error( } /* - * Close all the tracefiles and stream fds, should be called when all instances - * are destroyed. + * Close all the tracefiles and stream fds and MUST be called when all + * instances are destroyed i.e. when all threads were joined and are ended. */ void lttng_consumer_cleanup(void) { @@ -886,17 +1075,6 @@ 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); - } - cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, node, node) { struct lttng_consumer_channel *channel = @@ -906,8 +1084,16 @@ void lttng_consumer_cleanup(void) rcu_read_unlock(); - lttng_ht_destroy(consumer_data.stream_ht); lttng_ht_destroy(consumer_data.channel_ht); + + cleanup_relayd_ht(); + + /* + * This HT contains streams that are freed by either the metadata thread or + * the data thread so we do *nothing* on the hash table and simply destroy + * it. + */ + lttng_ht_destroy(consumer_data.stream_list_ht); } /* @@ -920,9 +1106,11 @@ void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx) do { ret = write(ctx->consumer_should_quit[1], "4", 1); } while (ret < 0 && errno == EINTR); - if (ret < 0) { + if (ret < 0 || ret != 1) { PERROR("write consumer quit"); } + + DBG("Consumer flag that it should quit"); } void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream, @@ -1004,21 +1192,21 @@ 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; @@ -1066,7 +1254,7 @@ 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"); } @@ -1084,6 +1272,8 @@ 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"); @@ -1096,11 +1286,11 @@ 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"); } @@ -1134,8 +1324,22 @@ static int write_relayd_metadata_id(int fd, do { ret = write(fd, (void *) &hdr, sizeof(hdr)); } while (ret < 0 && errno == EINTR); - if (ret < 0) { - PERROR("write metadata stream id"); + if (ret < 0 || ret != sizeof(hdr)) { + /* + * This error means that the fd's end is closed so ignore the perror + * not to clubber the error output since this can happen in a normal + * code path. + */ + if (errno != EPIPE) { + PERROR("write metadata stream id"); + } + DBG3("Consumer failed to write relayd metadata id (errno: %d)", errno); + /* + * Set ret to a negative value because if ret != sizeof(hdr), we don't + * handle writting the missing part so report that as an error and + * don't lie to the caller. + */ + ret = -1; goto end; } DBG("Metadata stream id %" PRIu64 " with padding %lu written before data", @@ -1150,6 +1354,8 @@ end: * core function for writing trace buffers to either the local filesystem or * the network. * + * It must be called with the stream lock held. + * * Careful review MUST be put if any changes occur! * * Returns the number of bytes written @@ -1165,6 +1371,7 @@ 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(); @@ -1222,11 +1429,22 @@ 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; @@ -1238,10 +1456,21 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( } while (ret < 0 && errno == EINTR); DBG("Consumer mmap write() ret %zd (len %lu)", ret, len); if (ret < 0) { - PERROR("Error in file write"); + /* + * This is possible if the fd is closed on the other side (outfd) + * or any write problem. It can be verbose a bit for a normal + * execution if for instance the relayd is stopped abruptly. This + * can happen so set this to a DBG statement. + */ + DBG("Error in file write 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); @@ -1263,6 +1492,15 @@ 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) { @@ -1276,6 +1514,8 @@ end: /* * Splice the data from the ring buffer to the tracefile. * + * It must be called with the stream lock held. + * * Returns the number of bytes spliced. */ ssize_t lttng_consumer_on_read_subbuffer_splice( @@ -1291,6 +1531,7 @@ 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: @@ -1341,6 +1582,12 @@ 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; } @@ -1352,7 +1599,12 @@ 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 { @@ -1401,6 +1653,12 @@ 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 || errno == EPIPE) { + WARN("Remote relayd disconnected. Stopping"); + relayd_hang_up = 1; + goto write_error; + } ret = errno; goto splice_error; } else if (ret_splice > len) { @@ -1428,12 +1686,20 @@ 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; @@ -1517,10 +1783,11 @@ int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx, /* * Iterate over all streams of the hashtable and free them properly. + * + * WARNING: *MUST* be used with data stream only. */ -static void destroy_stream_ht(struct lttng_ht *ht) +static void destroy_data_stream_ht(struct lttng_ht *ht) { - int ret; struct lttng_ht_iter iter; struct lttng_consumer_stream *stream; @@ -1530,10 +1797,11 @@ static void destroy_stream_ht(struct lttng_ht *ht) rcu_read_lock(); cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { - ret = lttng_ht_del(ht, &iter); - assert(!ret); - - call_rcu(&stream->node.head, consumer_free_stream); + /* + * Ignore return value since we are currently cleaning up so any error + * can't be handled. + */ + (void) consumer_del_stream(stream, ht); } rcu_read_unlock(); @@ -1541,21 +1809,60 @@ static void destroy_stream_ht(struct lttng_ht *ht) } /* - * Clean up a metadata stream and free its memory. + * Iterate over all streams of the hashtable and free them properly. + * + * XXX: Should not be only for metadata stream or else use an other name. */ -static void consumer_del_metadata_stream(struct lttng_consumer_stream *stream) +static void destroy_stream_ht(struct lttng_ht *ht) { - int ret; - struct consumer_relayd_sock_pair *relayd; + struct lttng_ht_iter iter; + struct lttng_consumer_stream *stream; - assert(stream); - /* - * This call should NEVER receive regular stream. It must always be - * metadata stream and this is crucial for data structure synchronization. - */ - assert(stream->metadata_flag); + if (ht == NULL) { + return; + } + + rcu_read_lock(); + cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { + /* + * Ignore return value since we are currently cleaning up so any error + * can't be handled. + */ + (void) consumer_del_metadata_stream(stream, ht); + } + rcu_read_unlock(); + + lttng_ht_destroy(ht); +} + +/* + * Clean up a metadata stream and free its memory. + */ +void consumer_del_metadata_stream(struct lttng_consumer_stream *stream, + struct lttng_ht *ht) +{ + int ret; + struct lttng_ht_iter iter; + struct lttng_consumer_channel *free_chan = NULL; + struct consumer_relayd_sock_pair *relayd; + + assert(stream); + /* + * This call should NEVER receive regular stream. It must always be + * metadata stream and this is crucial for data structure synchronization. + */ + assert(stream->metadata_flag); + + DBG3("Consumer delete metadata stream %d", stream->wait_fd); + + if (ht == NULL) { + /* Means the stream was allocated but not successfully added */ + goto free_stream; + } pthread_mutex_lock(&consumer_data.lock); + pthread_mutex_lock(&stream->lock); + switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: if (stream->mmap_base != NULL) { @@ -1572,8 +1879,19 @@ static void consumer_del_metadata_stream(struct lttng_consumer_stream *stream) default: ERR("Unknown consumer_data type"); assert(0); + goto end; } - pthread_mutex_unlock(&consumer_data.lock); + + rcu_read_lock(); + 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) { ret = close(stream->out_fd); @@ -1630,9 +1948,18 @@ static void consumer_del_metadata_stream(struct lttng_consumer_stream *stream) if (!uatomic_read(&stream->chan->refcount) && !uatomic_read(&stream->chan->nb_init_streams)) { /* Go for channel deletion! */ - consumer_del_channel(stream->chan); + free_chan = stream->chan; } +end: + pthread_mutex_unlock(&stream->lock); + pthread_mutex_unlock(&consumer_data.lock); + + if (free_chan) { + consumer_del_channel(free_chan); + } + +free_stream: call_rcu(&stream->node.head, consumer_free_stream); } @@ -1640,16 +1967,123 @@ static void consumer_del_metadata_stream(struct lttng_consumer_stream *stream) * Action done with the metadata stream when adding it to the consumer internal * data structures to handle it. */ -static void consumer_add_metadata_stream(struct lttng_consumer_stream *stream) +static int consumer_add_metadata_stream(struct lttng_consumer_stream *stream, + struct lttng_ht *ht) { + int ret = 0; struct consumer_relayd_sock_pair *relayd; + struct lttng_ht_iter iter; + struct lttng_ht_node_ulong *node; + + assert(stream); + assert(ht); + + DBG3("Adding metadata stream %d to hash table", stream->wait_fd); + + pthread_mutex_lock(&consumer_data.lock); + pthread_mutex_lock(&stream->lock); + + /* + * From here, refcounts are updated so be _careful_ when returning an error + * after this point. + */ - /* Find relayd and, if one is found, increment refcount. */ rcu_read_lock(); + + /* + * Lookup the stream just to make sure it does not exist in our internal + * state. This should NEVER happen. + */ + lttng_ht_lookup(ht, (void *)((unsigned long) stream->wait_fd), &iter); + node = lttng_ht_iter_get_node_ulong(&iter); + assert(!node); + + /* Find relayd and, if one is found, increment refcount. */ relayd = consumer_find_relayd(stream->net_seq_idx); if (relayd != NULL) { uatomic_inc(&relayd->refcount); } + + /* Update channel refcount once added without error(s). */ + uatomic_inc(&stream->chan->refcount); + + /* + * When nb_init_streams reaches 0, we don't need to trigger any action in + * terms of destroying the associated channel, because the action that + * causes the count to become 0 also causes a stream to be added. The + * channel deletion will thus be triggered by the following removal of this + * stream. + */ + if (uatomic_read(&stream->chan->nb_init_streams) > 0) { + uatomic_dec(&stream->chan->nb_init_streams); + } + + 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(&stream->lock); + 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 == CONSUMER_ENDPOINT_ACTIVE) { + 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 == CONSUMER_ENDPOINT_ACTIVE) { + 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(); } @@ -1657,27 +2091,27 @@ static void consumer_add_metadata_stream(struct lttng_consumer_stream *stream) * Thread polls on metadata file descriptor and write them on disk or on the * network. */ -void *lttng_consumer_thread_poll_metadata(void *data) +void *consumer_thread_metadata_poll(void *data) { int ret, i, pollfd; uint32_t revents, nb_fd; - struct lttng_consumer_stream *stream; + struct lttng_consumer_stream *stream = NULL; struct lttng_ht_iter iter; struct lttng_ht_node_ulong *node; - struct lttng_ht *metadata_ht = NULL; struct lttng_poll_event events; struct lttng_consumer_local_data *ctx = data; ssize_t len; rcu_register_thread(); - DBG("Thread metadata poll started"); - metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); - if (metadata_ht == NULL) { - goto end; + if (!metadata_ht) { + /* ENOMEM at this point. Better to bail out. */ + goto error; } + DBG("Thread metadata poll started"); + /* Size is set to 1 for the consumer_metadata pipe */ ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC); if (ret < 0) { @@ -1694,31 +2128,35 @@ void *lttng_consumer_thread_poll_metadata(void *data) DBG("Metadata main loop started"); while (1) { - lttng_poll_reset(&events); - - nb_fd = LTTNG_POLL_GETNB(&events); - /* Only the metadata pipe is set */ - if (nb_fd == 0 && consumer_quit == 1) { + if (LTTNG_POLL_GETNB(&events) == 0 && consumer_quit == 1) { goto end; } restart: - DBG("Metadata poll wait with %d fd(s)", nb_fd); + DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events)); ret = lttng_poll_wait(&events, -1); DBG("Metadata event catched in thread"); if (ret < 0) { if (errno == EINTR) { + ERR("Poll EINTR catched"); goto restart; } goto error; } + nb_fd = ret; + + /* From here, the event is a metadata wait fd */ for (i = 0; i < nb_fd; i++) { revents = LTTNG_POLL_GETEV(&events, i); pollfd = LTTNG_POLL_GETFD(&events, i); - /* Check the metadata pipe for incoming metadata. */ + /* Just don't waste time if no returned events for the fd */ + if (!revents) { + continue; + } + if (pollfd == ctx->consumer_metadata_pipe[0]) { if (revents & (LPOLLERR | LPOLLHUP )) { DBG("Metadata thread pipe hung up"); @@ -1727,7 +2165,10 @@ restart: * since their might be data to consume. */ lttng_poll_del(&events, ctx->consumer_metadata_pipe[0]); - close(ctx->consumer_metadata_pipe[0]); + ret = close(ctx->consumer_metadata_pipe[0]); + if (ret < 0) { + PERROR("close metadata pipe"); + } continue; } else if (revents & LPOLLIN) { do { @@ -1744,62 +2185,45 @@ 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); - rcu_read_lock(); - /* The node should be init at this point */ - lttng_ht_add_unique_ulong(metadata_ht, - &stream->waitfd_node); - rcu_read_unlock(); + ret = consumer_add_metadata_stream(stream, metadata_ht); + if (ret) { + ERR("Unable to add metadata stream"); + /* Stream was not setup properly. Continuing. */ + consumer_del_metadata_stream(stream, NULL); + continue; + } /* Add metadata stream to the global poll events list */ lttng_poll_add(&events, stream->wait_fd, LPOLLIN | LPOLLPRI); - - consumer_add_metadata_stream(stream); } - /* Metadata pipe handled. Continue handling the others */ + /* Handle other stream */ continue; } - /* From here, the event is a metadata wait fd */ - rcu_read_lock(); lttng_ht_lookup(metadata_ht, (void *)((unsigned long) pollfd), &iter); node = lttng_ht_iter_get_node_ulong(&iter); - if (node == NULL) { - /* FD not found, continue loop */ - rcu_read_unlock(); - continue; - } + assert(node); stream = caa_container_of(node, struct lttng_consumer_stream, - waitfd_node); - - /* Get the data out of the metadata file descriptor */ - if (revents & (LPOLLIN | LPOLLPRI)) { - DBG("Metadata available on fd %d", pollfd); - assert(stream->wait_fd == pollfd); + node); - len = ctx->on_buffer_ready(stream, ctx); - /* It's ok to have an unavailable sub-buffer */ - if (len < 0 && len != -EAGAIN) { - rcu_read_unlock(); - goto end; - } else if (len > 0) { - stream->data_read = 1; - } - } - - /* - * Remove the stream from the hash table since there is no data - * left on the fd because we previously did a read on the buffer. - */ + /* Check for error event */ if (revents & (LPOLLERR | LPOLLHUP)) { - DBG("Metadata fd %d is hup|err|nval.", pollfd); + DBG("Metadata fd %d is hup|err.", pollfd); if (!stream->hangup_flush_done && (consumer_data.type == LTTNG_CONSUMER32_UST || consumer_data.type == LTTNG_CONSUMER64_UST)) { @@ -1807,20 +2231,40 @@ restart: lttng_ustconsumer_on_stream_hangup(stream); /* We just flushed the stream now read it. */ - len = ctx->on_buffer_ready(stream, ctx); - /* It's ok to have an unavailable sub-buffer */ - if (len < 0 && len != -EAGAIN) { - rcu_read_unlock(); - goto end; - } + do { + len = ctx->on_buffer_ready(stream, ctx); + /* + * We don't check the return value here since if we get + * a negative len, it means an error occured thus we + * simply remove it from the poll set and free the + * stream. + */ + } while (len > 0); } - /* Removing it from hash table, poll set and free memory */ - lttng_ht_del(metadata_ht, &iter); - lttng_poll_del(&events, stream->wait_fd); - consumer_del_metadata_stream(stream); + /* + * This call update the channel states, closes file descriptors + * and securely free the stream. + */ + consumer_del_metadata_stream(stream, metadata_ht); + } else if (revents & (LPOLLIN | LPOLLPRI)) { + /* Get the data out of the metadata file descriptor */ + DBG("Metadata available on fd %d", pollfd); + assert(stream->wait_fd == pollfd); + + len = ctx->on_buffer_ready(stream, ctx); + /* It's ok to have an unavailable sub-buffer */ + if (len < 0 && len != -EAGAIN && len != -ENODATA) { + /* 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; + } } + + /* Release RCU lock for the stream looked up */ rcu_read_unlock(); } } @@ -1830,9 +2274,7 @@ end: DBG("Metadata poll thread exiting"); lttng_poll_clean(&events); - if (metadata_ht) { - destroy_stream_ht(metadata_ht); - } + destroy_stream_ht(metadata_ht); rcu_unregister_thread(); return NULL; @@ -1842,26 +2284,22 @@ end: * This thread polls the fds in the set to consume the data and write * it to tracefile if necessary. */ -void *lttng_consumer_thread_poll_fds(void *data) +void *consumer_thread_data_poll(void *data) { int num_rdy, num_hup, high_prio, ret, i; struct pollfd *pollfd = NULL; /* local view of the streams */ - struct lttng_consumer_stream **local_stream = NULL; + struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL; /* local view of consumer_data.fds_count */ int nb_fd = 0; struct lttng_consumer_local_data *ctx = data; ssize_t len; - pthread_t metadata_thread; - void *status; rcu_register_thread(); - /* Start metadata polling thread */ - ret = pthread_create(&metadata_thread, NULL, - lttng_consumer_thread_poll_metadata, (void *) ctx); - if (ret < 0) { - PERROR("pthread_create metadata thread"); + data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); + if (data_ht == NULL) { + /* ENOMEM at this point. Better to bail out. */ goto end; } @@ -1886,7 +2324,7 @@ void *lttng_consumer_thread_poll_fds(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"); @@ -1894,7 +2332,7 @@ void *lttng_consumer_thread_poll_fds(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) { @@ -1902,7 +2340,8 @@ void *lttng_consumer_thread_poll_fds(void *data) pthread_mutex_unlock(&consumer_data.lock); goto end; } - ret = consumer_update_poll_array(ctx, &pollfd, local_stream); + ret = consumer_update_poll_array(ctx, &pollfd, local_stream, + data_ht); if (ret < 0) { ERR("Error in allocating pollfd or local_outfds"); lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR); @@ -1921,7 +2360,7 @@ void *lttng_consumer_thread_poll_fds(void *data) /* poll on the array of fds */ restart: DBG("polling on %d fd", nb_fd + 1); - num_rdy = poll(pollfd, nb_fd + 1, consumer_poll_timeout); + num_rdy = poll(pollfd, nb_fd + 1, -1); DBG("poll num_rdy : %d", num_rdy); if (num_rdy == -1) { /* @@ -1939,31 +2378,64 @@ void *lttng_consumer_thread_poll_fds(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; - char tmp; - 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], &tmp, 1); + pipe_readlen = read(ctx->consumer_data_pipe[0], &new_stream, + sizeof(new_stream)); } while (pipe_readlen == -1 && errno == EINTR); + if (pipe_readlen < 0) { + PERROR("read consumer data pipe"); + /* Continue so we can at least handle the current stream(s). */ + continue; + } + + /* + * If the stream is NULL, just ignore it. It's also possible that + * the sessiond poll thread changed the consumer_quit state and is + * waking us up to test it. + */ + if (new_stream == NULL) { + validate_endpoint_status_data_stream(); + continue; + } + + ret = consumer_add_stream(new_stream, data_ht); + if (ret) { + ERR("Consumer add stream %d failed. Continuing", + new_stream->key); + /* + * At this point, if the add_stream fails, it is not in the + * hash table thus passing the NULL value here. + */ + consumer_del_stream(new_stream, NULL); + } + + /* Continue to update the local streams and handle prio ones */ continue; } /* Take care of high priority channels first. */ for (i = 0; i < nb_fd; i++) { + if (local_stream[i] == NULL) { + continue; + } if (pollfd[i].revents & POLLPRI) { DBG("Urgent read on fd %d", pollfd[i].fd); high_prio = 1; len = ctx->on_buffer_ready(local_stream[i], ctx); /* it's ok to have an unavailable sub-buffer */ - if (len < 0 && len != -EAGAIN) { - goto end; + if (len < 0 && len != -EAGAIN && len != -ENODATA) { + /* Clean the stream and free it. */ + consumer_del_stream(local_stream[i], data_ht); + local_stream[i] = NULL; } else if (len > 0) { local_stream[i]->data_read = 1; } @@ -1980,13 +2452,18 @@ void *lttng_consumer_thread_poll_fds(void *data) /* Take care of low priority channels. */ for (i = 0; i < nb_fd; i++) { + if (local_stream[i] == NULL) { + continue; + } if ((pollfd[i].revents & POLLIN) || local_stream[i]->hangup_flush_done) { DBG("Normal read on fd %d", pollfd[i].fd); len = ctx->on_buffer_ready(local_stream[i], ctx); /* it's ok to have an unavailable sub-buffer */ - if (len < 0 && len != -EAGAIN) { - goto end; + if (len < 0 && len != -EAGAIN && len != -ENODATA) { + /* Clean the stream and free it. */ + consumer_del_stream(local_stream[i], data_ht); + local_stream[i] = NULL; } else if (len > 0) { local_stream[i]->data_read = 1; } @@ -1995,12 +2472,15 @@ void *lttng_consumer_thread_poll_fds(void *data) /* Handle hangup and errors */ for (i = 0; i < nb_fd; i++) { + if (local_stream[i] == NULL) { + continue; + } if (!local_stream[i]->hangup_flush_done && (pollfd[i].revents & (POLLHUP | POLLERR | POLLNVAL)) && (consumer_data.type == LTTNG_CONSUMER32_UST || consumer_data.type == LTTNG_CONSUMER64_UST)) { DBG("fd %d is hup|err|nval. Attempting flush and read.", - pollfd[i].fd); + pollfd[i].fd); lttng_ustconsumer_on_stream_hangup(local_stream[i]); /* Attempt read again, for the data we just flushed. */ local_stream[i]->data_read = 1; @@ -2013,23 +2493,28 @@ void *lttng_consumer_thread_poll_fds(void *data) if ((pollfd[i].revents & POLLHUP)) { DBG("Polling fd %d tells it has hung up.", pollfd[i].fd); if (!local_stream[i]->data_read) { - consumer_del_stream(local_stream[i]); + consumer_del_stream(local_stream[i], data_ht); + local_stream[i] = NULL; num_hup++; } } else if (pollfd[i].revents & POLLERR) { ERR("Error returned in polling fd %d.", pollfd[i].fd); if (!local_stream[i]->data_read) { - consumer_del_stream(local_stream[i]); + consumer_del_stream(local_stream[i], data_ht); + local_stream[i] = NULL; num_hup++; } } else if (pollfd[i].revents & POLLNVAL) { ERR("Polling fd %d tells fd is not open.", pollfd[i].fd); if (!local_stream[i]->data_read) { - consumer_del_stream(local_stream[i]); + consumer_del_stream(local_stream[i], data_ht); + local_stream[i] = NULL; num_hup++; } } - local_stream[i]->data_read = 0; + if (local_stream[i] != NULL) { + local_stream[i]->data_read = 0; + } } } end: @@ -2045,20 +2530,19 @@ end: /* * Close the write side of the pipe so epoll_wait() in - * lttng_consumer_thread_poll_metadata can catch it. The thread is - * monitoring the read side of the pipe. If we close them both, epoll_wait - * strangely does not return and could create a endless wait period if the - * pipe is the only tracked fd in the poll set. The thread will take care - * of closing the read side. + * consumer_thread_metadata_poll can catch it. The thread is monitoring the + * read side of the pipe. If we close them both, epoll_wait strangely does + * not return and could create a endless wait period if the pipe is the + * only tracked fd in the poll set. The thread will take care of closing + * the read side. */ - close(ctx->consumer_metadata_pipe[1]); - if (ret) { - ret = pthread_join(metadata_thread, &status); - if (ret < 0) { - PERROR("pthread_join metadata thread"); - } + ret = close(ctx->consumer_metadata_pipe[1]); + if (ret < 0) { + PERROR("close data pipe"); } + destroy_data_stream_ht(data_ht); + rcu_unregister_thread(); return NULL; } @@ -2067,9 +2551,9 @@ end: * This thread listens on the consumerd socket and receives the file * descriptors from the session daemon. */ -void *lttng_consumer_thread_receive_fds(void *data) +void *consumer_thread_sessiond_poll(void *data) { - int sock, client_socket, ret; + int sock = -1, client_socket, ret; /* * structure to poll for incoming data on communication socket avoids * making blocking sockets. @@ -2119,7 +2603,7 @@ void *lttng_consumer_thread_receive_fds(void *data) /* Blocking call, waiting for transmission */ sock = lttcomm_accept_unix_sock(client_socket); - if (sock <= 0) { + if (sock < 0) { WARN("On accept"); goto end; } @@ -2129,6 +2613,13 @@ void *lttng_consumer_thread_receive_fds(void *data) goto end; } + /* This socket is not useful anymore. */ + ret = close(client_socket); + if (ret < 0) { + PERROR("close client_socket"); + } + client_socket = -1; + /* update the polling structure to poll on the established socket */ consumer_sockpoll[1].fd = sock; consumer_sockpoll[1].events = POLLIN | POLLPRI; @@ -2167,26 +2658,25 @@ end: consumer_quit = 1; /* - * 2s of grace period, if no polling events occur during - * this period, the polling thread will exit even if there - * are still open FDs (should not happen, but safety mechanism). + * Notify the data poll thread to poll back again and test the + * consumer_quit state that we just set so to quit gracefully. */ - consumer_poll_timeout = LTTNG_CONSUMER_POLL_TIMEOUT; + notify_thread_pipe(ctx->consumer_data_pipe[1]); + + /* Cleaning up possibly open sockets. */ + if (sock >= 0) { + ret = close(sock); + if (ret < 0) { + PERROR("close sock sessiond poll"); + } + } + if (client_socket >= 0) { + ret = close(sock); + if (ret < 0) { + PERROR("close client_socket sessiond poll"); + } + } - /* - * Wake-up the other end by writing a null byte in the pipe - * (non-blocking). Important note: Because writing into the - * pipe is non-blocking (and therefore we allow dropping wakeup - * data, as long as there is wakeup data present in the pipe - * buffer to wake up the other end), the other end should - * perform the following sequence for waiting: - * 1) empty the pipe (reads). - * 2) perform update operation. - * 3) wait on the pipe (poll). - */ - do { - ret = write(ctx->consumer_poll_pipe[1], "", 1); - } while (ret < 0 && errno == EINTR); rcu_unregister_thread(); return NULL; } @@ -2194,17 +2684,27 @@ end: ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream, struct lttng_consumer_local_data *ctx) { + ssize_t ret; + + pthread_mutex_lock(&stream->lock); + switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: - return lttng_kconsumer_read_subbuffer(stream, ctx); + ret = lttng_kconsumer_read_subbuffer(stream, ctx); + break; case LTTNG_CONSUMER32_UST: case LTTNG_CONSUMER64_UST: - return lttng_ustconsumer_read_subbuffer(stream, ctx); + ret = lttng_ustconsumer_read_subbuffer(stream, ctx); + break; default: ERR("Unknown consumer_data type"); assert(0); - return -ENOSYS; + ret = -ENOSYS; + break; } + + pthread_mutex_unlock(&stream->lock); + return ret; } int lttng_consumer_on_recv_stream(struct lttng_consumer_stream *stream) @@ -2227,9 +2727,10 @@ 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); + relayd_session_id_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); } /* @@ -2240,13 +2741,23 @@ void lttng_consumer_init(void) */ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, struct lttng_consumer_local_data *ctx, int sock, - struct pollfd *consumer_sockpoll, struct lttcomm_sock *relayd_sock) + struct pollfd *consumer_sockpoll, struct lttcomm_sock *relayd_sock, + unsigned int sessiond_id) { - int fd, ret = -1; + int fd = -1, ret = -1, relayd_created = 0; + enum lttng_error_code ret_code = LTTNG_OK; struct consumer_relayd_sock_pair *relayd; + struct consumer_relayd_session_id *relayd_id_node; DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx); + /* First send a status message before receiving the fds. */ + ret = consumer_send_status_msg(sock, ret_code); + if (ret < 0) { + /* Somehow, the session daemon is not responding anymore. */ + goto error; + } + /* Get relayd reference if exists. */ relayd = consumer_find_relayd(net_seq_idx); if (relayd == NULL) { @@ -2254,8 +2765,11 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, relayd = consumer_allocate_relayd_sock_pair(net_seq_idx); if (relayd == NULL) { lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_OUTFD_ERROR); + ret = -1; goto error; } + relayd->sessiond_session_id = (uint64_t) sessiond_id; + relayd_created = 1; } /* Poll on consumer socket. */ @@ -2269,6 +2783,14 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, if (ret != sizeof(fd)) { lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD); ret = -1; + fd = -1; /* Just in case it gets set with an invalid value. */ + goto error; + } + + /* We have the fds without error. Send status back. */ + ret = consumer_send_status_msg(sock, ret_code); + if (ret < 0) { + /* Somehow, the session daemon is not responding anymore. */ goto error; } @@ -2278,32 +2800,76 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, /* Copy received lttcomm socket */ lttcomm_copy_sock(&relayd->control_sock, relayd_sock); ret = lttcomm_create_sock(&relayd->control_sock); + /* Immediately try to close the created socket if valid. */ + if (relayd->control_sock.fd >= 0) { + if (close(relayd->control_sock.fd)) { + PERROR("close relayd control socket"); + } + } + /* Handle create_sock error. */ if (ret < 0) { goto error; } - /* Close the created socket fd which is useless */ - close(relayd->control_sock.fd); - /* Assign new file descriptor */ relayd->control_sock.fd = fd; + + /* + * Create a session on the relayd and store the returned id. Lock the + * control socket mutex if the relayd was NOT created before. + */ + if (!relayd_created) { + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + } + ret = relayd_create_session(&relayd->control_sock, + &relayd->relayd_session_id); + if (!relayd_created) { + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); + } + if (ret < 0) { + goto error; + } + + /* Set up a relayd session id node. */ + relayd_id_node = zmalloc(sizeof(struct consumer_relayd_session_id)); + if (!relayd_id_node) { + PERROR("zmalloc relayd id node"); + ret = -1; + goto error; + } + + relayd_id_node->relayd_id = relayd->relayd_session_id; + relayd_id_node->sessiond_id = (uint64_t) sessiond_id; + + /* Indexed by session id of the sessiond. */ + lttng_ht_node_init_ulong(&relayd_id_node->node, + relayd_id_node->sessiond_id); + rcu_read_lock(); + lttng_ht_add_unique_ulong(relayd_session_id_ht, &relayd_id_node->node); + rcu_read_unlock(); + break; case LTTNG_STREAM_DATA: /* Copy received lttcomm socket */ lttcomm_copy_sock(&relayd->data_sock, relayd_sock); ret = lttcomm_create_sock(&relayd->data_sock); + /* Immediately try to close the created socket if valid. */ + if (relayd->data_sock.fd >= 0) { + if (close(relayd->data_sock.fd)) { + PERROR("close relayd data socket"); + } + } + /* Handle create_sock error. */ if (ret < 0) { goto error; } - /* Close the created socket fd which is useless */ - close(relayd->data_sock.fd); - /* Assign new file descriptor */ relayd->data_sock.fd = fd; break; default: ERR("Unknown relayd socket type (%d)", sock_type); + ret = -1; goto error; } @@ -2318,8 +2884,224 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, add_relayd(relayd); /* All good! */ - ret = 0; + return 0; error: + /* Close received socket if valid. */ + if (fd >= 0) { + if (close(fd)) { + PERROR("close received socket"); + } + } + + if (relayd_created) { + /* We just want to cleanup. Ignore ret value. */ + (void) relayd_close(&relayd->control_sock); + (void) relayd_close(&relayd->data_sock); + free(relayd); + } + + 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; } + +/* + * Search for a relayd associated to the session id and return the reference. + * + * A rcu read side lock MUST be acquire before calling this function and locked + * until the relayd object is no longer necessary. + */ +static struct consumer_relayd_sock_pair *find_relayd_by_session_id(uint64_t id) +{ + struct lttng_ht_iter iter; + struct consumer_relayd_sock_pair *relayd = NULL; + + /* Iterate over all relayd since they are indexed by net_seq_idx. */ + cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd, + node.node) { + /* + * Check by sessiond id which is unique here where the relayd session + * id might not be when having multiple relayd. + */ + if (relayd->sessiond_session_id == id) { + /* Found the relayd. There can be only one per id. */ + goto found; + } + } + + return NULL; + +found: + return relayd; +} + +/* + * Check if for a given session id there is still data needed to be extract + * from the buffers. + * + * Return 1 if data is pending or else 0 meaning ready to be read. + */ +int consumer_data_pending(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 = NULL; + int (*data_pending)(struct lttng_consumer_stream *); + + DBG("Consumer data pending command on session id %" PRIu64, id); + + rcu_read_lock(); + pthread_mutex_lock(&consumer_data.lock); + + switch (consumer_data.type) { + case LTTNG_CONSUMER_KERNEL: + data_pending = lttng_kconsumer_data_pending; + break; + case LTTNG_CONSUMER32_UST: + case LTTNG_CONSUMER64_UST: + data_pending = lttng_ustconsumer_data_pending; + break; + default: + ERR("Unknown consumer data type"); + assert(0); + } + + /* Ease our life a bit */ + ht = consumer_data.stream_list_ht; + + relayd = find_relayd_by_session_id(id); + if (relayd) { + /* Send init command for data pending. */ + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + ret = relayd_begin_data_pending(&relayd->control_sock, + relayd->relayd_session_id); + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); + if (ret < 0) { + /* Communication error thus the relayd so no data pending. */ + goto data_not_pending; + } + } + + cds_lfht_for_each_entry_duplicate(ht->ht, + ht->hash_fct((void *)((unsigned long) id), lttng_ht_seed), + 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_pending; + } + + /* + * 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_pending(stream); + if (ret == 1) { + pthread_mutex_unlock(&stream->lock); + goto data_pending; + } + } + + /* Relayd check */ + if (relayd) { + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + if (stream->metadata_flag) { + ret = relayd_quiescent_control(&relayd->control_sock, + stream->relayd_stream_id); + } else { + ret = relayd_data_pending(&relayd->control_sock, + stream->relayd_stream_id, + stream->next_net_seq_num - 1); + } + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); + if (ret == 1) { + pthread_mutex_unlock(&stream->lock); + goto data_pending; + } + } + pthread_mutex_unlock(&stream->lock); + } + + if (relayd) { + unsigned int is_data_inflight = 0; + + /* Send init command for data pending. */ + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + ret = relayd_end_data_pending(&relayd->control_sock, + relayd->relayd_session_id, &is_data_inflight); + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); + if (ret < 0) { + goto data_not_pending; + } + if (is_data_inflight) { + goto data_pending; + } + } + + /* + * Finding _no_ node in the hash table and no inflight data means that the + * stream(s) have been removed thus data is guaranteed to be available for + * analysis from the trace files. + */ + +data_not_pending: + /* Data is available to be read by a viewer. */ + pthread_mutex_unlock(&consumer_data.lock); + rcu_read_unlock(); + return 0; + +data_pending: + /* Data is still being extracted from buffers. */ + pthread_mutex_unlock(&consumer_data.lock); + rcu_read_unlock(); + return 1; +} + +/* + * Send a ret code status message to the sessiond daemon. + * + * Return the sendmsg() return value. + */ +int consumer_send_status_msg(int sock, int ret_code) +{ + struct lttcomm_consumer_status_msg msg; + + msg.ret_code = ret_code; + + return lttcomm_send_unix_sock(sock, &msg, sizeof(msg)); +}