X-Git-Url: https://git.lttng.org/?p=lttng-tools.git;a=blobdiff_plain;f=src%2Fcommon%2Fconsumer.c;h=4b657f33249f7a4f777b7b3ff784b379b15f5878;hp=ca6aeba3ffd7a8748256bd9cae356eeff3e977b6;hb=6a00837f8cb0431a2ad90974d67fae138ea97dd5;hpb=9617607b21389440ecd1e02faf46f7ccf38dbd2b diff --git a/src/common/consumer.c b/src/common/consumer.c index ca6aeba3f..4b657f332 100644 --- a/src/common/consumer.c +++ b/src/common/consumer.c @@ -28,6 +28,7 @@ #include #include #include +#include #include #include @@ -40,6 +41,7 @@ #include #include "consumer.h" +#include "consumer-stream.h" struct lttng_consumer_global_data consumer_data = { .stream_count = 0, @@ -47,8 +49,17 @@ 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; +enum consumer_channel_action { + CONSUMER_CHANNEL_ADD, + CONSUMER_CHANNEL_DEL, + CONSUMER_CHANNEL_QUIT, +}; + +struct consumer_channel_msg { + enum consumer_channel_action action; + struct lttng_consumer_channel *chan; /* add */ + uint64_t key; /* del */ +}; /* * Flag to inform the polling thread to quit when all fd hung up. Updated by @@ -59,53 +70,92 @@ int consumer_poll_timeout = -1; volatile int consumer_quit; /* - * The following two hash tables are visible by all threads which are separated - * in different source files. - * * 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. */ -struct lttng_ht *metadata_ht; -struct lttng_ht *data_ht; +static struct lttng_ht *metadata_ht; +static 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. + * Notify a thread lttng 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) +static void notify_thread_lttng_pipe(struct lttng_pipe *pipe) +{ + struct lttng_consumer_stream *null_stream = NULL; + + assert(pipe); + + (void) lttng_pipe_write(pipe, &null_stream, sizeof(null_stream)); +} + +static void notify_channel_pipe(struct lttng_consumer_local_data *ctx, + struct lttng_consumer_channel *chan, + uint64_t key, + enum consumer_channel_action action) { + struct consumer_channel_msg msg; int ret; + memset(&msg, 0, sizeof(msg)); + + msg.action = action; + msg.chan = chan; + msg.key = key; do { - struct lttng_consumer_stream *null_stream = NULL; + ret = write(ctx->consumer_channel_pipe[1], &msg, sizeof(msg)); + } while (ret < 0 && errno == EINTR); +} + +void notify_thread_del_channel(struct lttng_consumer_local_data *ctx, + uint64_t key) +{ + notify_channel_pipe(ctx, NULL, key, CONSUMER_CHANNEL_DEL); +} - ret = write(wpipe, &null_stream, sizeof(null_stream)); +static int read_channel_pipe(struct lttng_consumer_local_data *ctx, + struct lttng_consumer_channel **chan, + uint64_t *key, + enum consumer_channel_action *action) +{ + struct consumer_channel_msg msg; + int ret; + + do { + ret = read(ctx->consumer_channel_pipe[0], &msg, sizeof(msg)); } while (ret < 0 && errno == EINTR); + if (ret > 0) { + *action = msg.action; + *chan = msg.chan; + *key = msg.key; + } + return ret; } /* * 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 *find_stream(uint64_t key, struct lttng_ht *ht) { struct lttng_ht_iter iter; - struct lttng_ht_node_ulong *node; + struct lttng_ht_node_u64 *node; struct lttng_consumer_stream *stream = NULL; assert(ht); - /* Negative keys are lookup failures */ - if (key < 0) { + /* -1ULL keys are lookup failures */ + if (key == (uint64_t) -1ULL) { return NULL; } rcu_read_lock(); - lttng_ht_lookup(ht, (void *)((unsigned long) key), &iter); - node = lttng_ht_iter_get_node_ulong(&iter); + lttng_ht_lookup(ht, &key, &iter); + node = lttng_ht_iter_get_node_u64(&iter); if (node != NULL) { stream = caa_container_of(node, struct lttng_consumer_stream, node); } @@ -115,85 +165,77 @@ static struct lttng_consumer_stream *consumer_find_stream(int key, return stream; } -void consumer_steal_stream_key(int key, struct lttng_ht *ht) +static void steal_stream_key(uint64_t key, struct lttng_ht *ht) { struct lttng_consumer_stream *stream; rcu_read_lock(); - stream = consumer_find_stream(key, ht); + stream = find_stream(key, ht); if (stream) { - stream->key = -1; + stream->key = (uint64_t) -1ULL; /* * We don't want the lookup to match, but we still need * to iterate on this stream when iterating over the hash table. Just * change the node key. */ - stream->node.key = -1; + stream->node.key = (uint64_t) -1ULL; } rcu_read_unlock(); } -static struct lttng_consumer_channel *consumer_find_channel(int key) +/* + * Return a channel object for the given key. + * + * RCU read side lock MUST be acquired before calling this function and + * protects the channel ptr. + */ +struct lttng_consumer_channel *consumer_find_channel(uint64_t key) { struct lttng_ht_iter iter; - struct lttng_ht_node_ulong *node; + struct lttng_ht_node_u64 *node; struct lttng_consumer_channel *channel = NULL; - /* Negative keys are lookup failures */ - if (key < 0) { + /* -1ULL keys are lookup failures */ + if (key == (uint64_t) -1ULL) { 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); + lttng_ht_lookup(consumer_data.channel_ht, &key, &iter); + node = lttng_ht_iter_get_node_u64(&iter); if (node != NULL) { channel = caa_container_of(node, struct lttng_consumer_channel, node); } - rcu_read_unlock(); - return channel; } -static void consumer_steal_channel_key(int key) +static void free_stream_rcu(struct rcu_head *head) { - struct lttng_consumer_channel *channel; + struct lttng_ht_node_u64 *node = + caa_container_of(head, struct lttng_ht_node_u64, head); + struct lttng_consumer_stream *stream = + caa_container_of(node, struct lttng_consumer_stream, node); - rcu_read_lock(); - channel = consumer_find_channel(key); - if (channel) { - channel->key = -1; - /* - * We don't want the lookup to match, but we still need - * to iterate on this channel when iterating over the hash table. Just - * change the node key. - */ - channel->node.key = -1; - } - rcu_read_unlock(); + free(stream); } -static -void consumer_free_stream(struct rcu_head *head) +static void free_channel_rcu(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, node); + struct lttng_ht_node_u64 *node = + caa_container_of(head, struct lttng_ht_node_u64, head); + struct lttng_consumer_channel *channel = + caa_container_of(node, struct lttng_consumer_channel, node); - free(stream); + free(channel); } /* * RCU protected relayd socket pair free. */ -static void consumer_rcu_free_relayd(struct rcu_head *head) +static void free_relayd_rcu(struct rcu_head *head) { - struct lttng_ht_node_ulong *node = - caa_container_of(head, struct lttng_ht_node_ulong, head); + struct lttng_ht_node_u64 *node = + caa_container_of(head, struct lttng_ht_node_u64, head); struct consumer_relayd_sock_pair *relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node); @@ -213,10 +255,8 @@ static void consumer_rcu_free_relayd(struct rcu_head *head) /* * Destroy and free relayd socket pair object. - * - * This function MUST be called with the consumer_data lock acquired. */ -static void destroy_relayd(struct consumer_relayd_sock_pair *relayd) +void consumer_destroy_relayd(struct consumer_relayd_sock_pair *relayd) { int ret; struct lttng_ht_iter iter; @@ -235,7 +275,79 @@ static void destroy_relayd(struct consumer_relayd_sock_pair *relayd) } /* RCU free() call */ - call_rcu(&relayd->node.head, consumer_rcu_free_relayd); + call_rcu(&relayd->node.head, free_relayd_rcu); +} + +/* + * Remove a channel from the global list protected by a mutex. This function is + * also responsible for freeing its data structures. + */ +void consumer_del_channel(struct lttng_consumer_channel *channel) +{ + int ret; + struct lttng_ht_iter iter; + struct lttng_consumer_stream *stream, *stmp; + + DBG("Consumer delete channel key %" PRIu64, channel->key); + + pthread_mutex_lock(&consumer_data.lock); + pthread_mutex_lock(&channel->lock); + + /* Delete streams that might have been left in the stream list. */ + cds_list_for_each_entry_safe(stream, stmp, &channel->streams.head, + send_node) { + cds_list_del(&stream->send_node); + /* + * Once a stream is added to this list, the buffers were created so + * we have a guarantee that this call will succeed. + */ + consumer_stream_destroy(stream, NULL); + } + + switch (consumer_data.type) { + case LTTNG_CONSUMER_KERNEL: + break; + case LTTNG_CONSUMER32_UST: + case LTTNG_CONSUMER64_UST: + lttng_ustconsumer_del_channel(channel); + break; + default: + ERR("Unknown consumer_data type"); + assert(0); + goto end; + } + + rcu_read_lock(); + iter.iter.node = &channel->node.node; + ret = lttng_ht_del(consumer_data.channel_ht, &iter); + assert(!ret); + rcu_read_unlock(); + + call_rcu(&channel->node.head, free_channel_rcu); +end: + pthread_mutex_unlock(&channel->lock); + pthread_mutex_unlock(&consumer_data.lock); +} + +/* + * 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) { + consumer_destroy_relayd(relayd); + } + + rcu_read_unlock(); + + lttng_ht_destroy(consumer_data.relayd_ht); } /* @@ -245,13 +357,13 @@ static void destroy_relayd(struct consumer_relayd_sock_pair *relayd) * 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, +static void update_endpoint_status_by_netidx(uint64_t 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); + DBG("Consumer set delete flag on stream by idx %" PRIu64, net_seq_idx); rcu_read_lock(); @@ -284,7 +396,7 @@ static void update_endpoint_status_by_netidx(int net_seq_idx, static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd, struct lttng_consumer_local_data *ctx) { - int netidx; + uint64_t netidx; assert(relayd); @@ -297,7 +409,7 @@ static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd, * Delete the relayd from the relayd hash table, close the sockets and free * the object in a RCU call. */ - destroy_relayd(relayd); + consumer_destroy_relayd(relayd); /* Set inactive endpoint to all streams */ update_endpoint_status_by_netidx(netidx, CONSUMER_ENDPOINT_INACTIVE); @@ -309,8 +421,8 @@ static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd, * 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]); + notify_thread_lttng_pipe(ctx->consumer_data_pipe); + notify_thread_lttng_pipe(ctx->consumer_metadata_pipe); } } @@ -329,269 +441,159 @@ void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd) /* Destroy the relayd if refcount is 0 */ if (uatomic_read(&relayd->refcount) == 0) { - destroy_relayd(relayd); + consumer_destroy_relayd(relayd); } } /* - * Remove a stream from the global list protected by a mutex. This - * function is also responsible for freeing its data structures. + * Completly destroy stream from every visiable data structure and the given + * hash table if one. + * + * One this call returns, the stream object is not longer usable nor visible. */ void consumer_del_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); - - 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) { - case LTTNG_CONSUMER_KERNEL: - if (stream->mmap_base != NULL) { - ret = munmap(stream->mmap_base, stream->mmap_len); - if (ret != 0) { - PERROR("munmap"); - } - } - break; - case LTTNG_CONSUMER32_UST: - case LTTNG_CONSUMER64_UST: - lttng_ustconsumer_del_stream(stream); - break; - default: - ERR("Unknown consumer_data type"); - assert(0); - goto end; - } - - 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(); - - assert(consumer_data.stream_count > 0); - consumer_data.stream_count--; - - if (stream->out_fd >= 0) { - ret = close(stream->out_fd); - if (ret) { - PERROR("close"); - } - } - if (stream->wait_fd >= 0 && !stream->wait_fd_is_copy) { - ret = close(stream->wait_fd); - if (ret) { - PERROR("close"); - } - } - if (stream->shm_fd >= 0 && stream->wait_fd != stream->shm_fd) { - ret = close(stream->shm_fd); - if (ret) { - PERROR("close"); - } - } - - /* Check and cleanup relayd */ - rcu_read_lock(); - relayd = consumer_find_relayd(stream->net_seq_idx); - if (relayd != NULL) { - uatomic_dec(&relayd->refcount); - assert(uatomic_read(&relayd->refcount) >= 0); - - /* Closing streams requires to lock the control socket. */ - pthread_mutex_lock(&relayd->ctrl_sock_mutex); - ret = relayd_send_close_stream(&relayd->control_sock, - stream->relayd_stream_id, - stream->next_net_seq_num - 1); - pthread_mutex_unlock(&relayd->ctrl_sock_mutex); - if (ret < 0) { - DBG("Unable to close stream on the relayd. Continuing"); - /* - * Continue here. There is nothing we can do for the relayd. - * Chances are that the relayd has closed the socket so we just - * continue cleaning up. - */ - } - - /* Both conditions are met, we destroy the relayd. */ - if (uatomic_read(&relayd->refcount) == 0 && - uatomic_read(&relayd->destroy_flag)) { - destroy_relayd(relayd); - } - } - rcu_read_unlock(); - - uatomic_dec(&stream->chan->refcount); - if (!uatomic_read(&stream->chan->refcount) - && !uatomic_read(&stream->chan->nb_init_streams)) { - free_chan = stream->chan; - } - -end: - consumer_data.need_update = 1; - pthread_mutex_unlock(&consumer_data.lock); - pthread_mutex_unlock(&stream->lock); + consumer_stream_destroy(stream, ht); +} - if (free_chan) { - consumer_del_channel(free_chan); - } +/* + * XXX naming of del vs destroy is all mixed up. + */ +void consumer_del_stream_for_data(struct lttng_consumer_stream *stream) +{ + consumer_stream_destroy(stream, data_ht); +} -free_stream: - call_rcu(&stream->node.head, consumer_free_stream); +void consumer_del_stream_for_metadata(struct lttng_consumer_stream *stream) +{ + consumer_stream_destroy(stream, metadata_ht); } -struct lttng_consumer_stream *consumer_allocate_stream( - int channel_key, int stream_key, - int shm_fd, int wait_fd, +struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key, + uint64_t stream_key, enum lttng_consumer_stream_state state, - uint64_t mmap_len, - enum lttng_event_output output, - const char *path_name, + const char *channel_name, uid_t uid, gid_t gid, - int net_index, - int metadata_flag, + uint64_t relayd_id, uint64_t session_id, - int *alloc_ret) + int cpu, + int *alloc_ret, + enum consumer_channel_type type, + unsigned int monitor) { + int ret; struct lttng_consumer_stream *stream; stream = zmalloc(sizeof(*stream)); if (stream == NULL) { PERROR("malloc struct lttng_consumer_stream"); - *alloc_ret = -ENOMEM; + ret = -ENOMEM; goto end; } - /* - * Get stream's channel reference. Needed when adding the stream to the - * global hash table. - */ - stream->chan = consumer_find_channel(channel_key); - if (!stream->chan) { - *alloc_ret = -ENOENT; - ERR("Unable to find channel for stream %d", stream_key); - goto error; - } + rcu_read_lock(); stream->key = stream_key; - stream->shm_fd = shm_fd; - stream->wait_fd = wait_fd; stream->out_fd = -1; stream->out_fd_offset = 0; stream->state = state; - stream->mmap_len = mmap_len; - stream->mmap_base = NULL; - stream->output = output; stream->uid = uid; stream->gid = gid; - stream->net_seq_idx = net_index; - stream->metadata_flag = metadata_flag; + stream->net_seq_idx = relayd_id; stream->session_id = session_id; - strncpy(stream->path_name, path_name, sizeof(stream->path_name)); - stream->path_name[sizeof(stream->path_name) - 1] = '\0'; + stream->monitor = monitor; + stream->endpoint_status = CONSUMER_ENDPOINT_ACTIVE; 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); + /* If channel is the metadata, flag this stream as metadata. */ + if (type == CONSUMER_CHANNEL_TYPE_METADATA) { + stream->metadata_flag = 1; + /* Metadata is flat out. */ + strncpy(stream->name, DEFAULT_METADATA_NAME, sizeof(stream->name)); } else { - lttng_ht_node_init_ulong(&stream->node, stream->key); + /* Format stream name to _ */ + ret = snprintf(stream->name, sizeof(stream->name), "%s_%d", + channel_name, cpu); + if (ret < 0) { + PERROR("snprintf stream name"); + goto error; + } } - /* Init session id node with the stream session id */ - lttng_ht_node_init_ulong(&stream->node_session_id, stream->session_id); + /* Key is always the wait_fd for streams. */ + lttng_ht_node_init_u64(&stream->node, stream->key); - /* - * The cpu number is needed before using any ustctl_* actions. Ignored for - * the kernel so the value does not matter. - */ - pthread_mutex_lock(&consumer_data.lock); - stream->cpu = stream->chan->cpucount++; - pthread_mutex_unlock(&consumer_data.lock); + /* Init node per channel id key */ + lttng_ht_node_init_u64(&stream->node_channel_id, channel_key); + + /* Init session id node with the stream session id */ + lttng_ht_node_init_u64(&stream->node_session_id, stream->session_id); - DBG3("Allocated stream %s (key %d, shm_fd %d, wait_fd %d, mmap_len %llu," - " 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, + DBG3("Allocated stream %s (key %" PRIu64 ", chan_key %" PRIu64 + " relayd_id %" PRIu64 ", session_id %" PRIu64, + stream->name, stream->key, channel_key, stream->net_seq_idx, stream->session_id); + + rcu_read_unlock(); return stream; error: + rcu_read_unlock(); free(stream); end: + if (alloc_ret) { + *alloc_ret = ret; + } return NULL; } /* * Add a stream to the global list protected by a mutex. */ -static int consumer_add_stream(struct lttng_consumer_stream *stream, - struct lttng_ht *ht) +int consumer_add_data_stream(struct lttng_consumer_stream *stream) { + struct lttng_ht *ht = data_ht; int ret = 0; - struct consumer_relayd_sock_pair *relayd; assert(stream); assert(ht); - DBG3("Adding consumer stream %d", stream->key); + DBG3("Adding consumer stream %" PRIu64, stream->key); pthread_mutex_lock(&consumer_data.lock); + pthread_mutex_lock(&stream->chan->lock); + pthread_mutex_lock(&stream->chan->timer_lock); + pthread_mutex_lock(&stream->lock); rcu_read_lock(); /* Steal stream identifier to avoid having streams with the same key */ - consumer_steal_stream_key(stream->key, ht); + steal_stream_key(stream->key, ht); + + lttng_ht_add_unique_u64(ht, &stream->node); - lttng_ht_add_unique_ulong(ht, &stream->node); + lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht, + &stream->node_channel_id); /* * 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); - } - - /* Update channel refcount once added without error(s). */ - uatomic_inc(&stream->chan->refcount); + lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_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 + * When nb_init_stream_left 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); + if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) { + /* Increment refcount before decrementing nb_init_stream_left */ + cmm_smp_wmb(); + uatomic_dec(&stream->chan->nb_init_stream_left); } /* Update consumer data once the node is inserted. */ @@ -599,11 +601,19 @@ static int consumer_add_stream(struct lttng_consumer_stream *stream, consumer_data.need_update = 1; rcu_read_unlock(); + pthread_mutex_unlock(&stream->lock); + pthread_mutex_unlock(&stream->chan->timer_lock); + pthread_mutex_unlock(&stream->chan->lock); pthread_mutex_unlock(&consumer_data.lock); return ret; } +void consumer_del_data_stream(struct lttng_consumer_stream *stream) +{ + consumer_del_stream(stream, data_ht); +} + /* * Add relayd socket to global consumer data hashtable. RCU read side lock MUST * be acquired before calling this. @@ -611,22 +621,18 @@ static int consumer_add_stream(struct lttng_consumer_stream *stream, static int add_relayd(struct consumer_relayd_sock_pair *relayd) { int ret = 0; - struct lttng_ht_node_ulong *node; + struct lttng_ht_node_u64 *node; struct lttng_ht_iter iter; - if (relayd == NULL) { - ret = -1; - goto end; - } + assert(relayd); lttng_ht_lookup(consumer_data.relayd_ht, - (void *)((unsigned long) relayd->net_seq_idx), &iter); - node = lttng_ht_iter_get_node_ulong(&iter); + &relayd->net_seq_idx, &iter); + node = lttng_ht_iter_get_node_u64(&iter); if (node != NULL) { - /* Relayd already exist. Ignore the insertion */ goto end; } - lttng_ht_add_unique_ulong(consumer_data.relayd_ht, &relayd->node); + lttng_ht_add_unique_u64(consumer_data.relayd_ht, &relayd->node); end: return ret; @@ -636,12 +642,12 @@ end: * Allocate and return a consumer relayd socket. */ struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair( - int net_seq_idx) + uint64_t net_seq_idx) { struct consumer_relayd_sock_pair *obj = NULL; - /* Negative net sequence index is a failure */ - if (net_seq_idx < 0) { + /* net sequence index of -1 is a failure */ + if (net_seq_idx == (uint64_t) -1ULL) { goto error; } @@ -654,7 +660,9 @@ struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair( obj->net_seq_idx = net_seq_idx; obj->refcount = 0; obj->destroy_flag = 0; - lttng_ht_node_init_ulong(&obj->node, obj->net_seq_idx); + obj->control_sock.sock.fd = -1; + obj->data_sock.sock.fd = -1; + lttng_ht_node_init_u64(&obj->node, obj->net_seq_idx); pthread_mutex_init(&obj->ctrl_sock_mutex, NULL); error: @@ -668,20 +676,20 @@ error: * RCU read-side lock must be held across this call and while using the * returned object. */ -struct consumer_relayd_sock_pair *consumer_find_relayd(int key) +struct consumer_relayd_sock_pair *consumer_find_relayd(uint64_t key) { struct lttng_ht_iter iter; - struct lttng_ht_node_ulong *node; + struct lttng_ht_node_u64 *node; struct consumer_relayd_sock_pair *relayd = NULL; /* Negative keys are lookup failures */ - if (key < 0) { + if (key == (uint64_t) -1ULL) { goto error; } - lttng_ht_lookup(consumer_data.relayd_ht, (void *)((unsigned long) key), + lttng_ht_lookup(consumer_data.relayd_ht, &key, &iter); - node = lttng_ht_iter_get_node_ulong(&iter); + node = lttng_ht_iter_get_node_u64(&iter); if (node != NULL) { relayd = caa_container_of(node, struct consumer_relayd_sock_pair, node); } @@ -691,16 +699,77 @@ error: } /* - * Handle stream for relayd transmission if the stream applies for network - * streaming where the net sequence index is set. + * Find a relayd and send the stream * - * Return destination file descriptor or negative value on error. + * Returns 0 on success, < 0 on error */ -static int write_relayd_stream_header(struct lttng_consumer_stream *stream, - size_t data_size, unsigned long padding, - struct consumer_relayd_sock_pair *relayd) +int consumer_send_relayd_stream(struct lttng_consumer_stream *stream, + char *path) { - int outfd = -1, ret; + int ret = 0; + struct consumer_relayd_sock_pair *relayd; + + assert(stream); + assert(stream->net_seq_idx != -1ULL); + assert(path); + + /* The stream is not metadata. Get relayd reference if exists. */ + rcu_read_lock(); + relayd = consumer_find_relayd(stream->net_seq_idx); + if (relayd != NULL) { + /* Add stream on the relayd */ + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + ret = relayd_add_stream(&relayd->control_sock, stream->name, + path, &stream->relayd_stream_id, + stream->chan->tracefile_size, stream->chan->tracefile_count); + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); + if (ret < 0) { + goto end; + } + uatomic_inc(&relayd->refcount); + stream->sent_to_relayd = 1; + } else { + ERR("Stream %" PRIu64 " relayd ID %" PRIu64 " unknown. Can't send it.", + stream->key, stream->net_seq_idx); + ret = -1; + goto end; + } + + DBG("Stream %s with key %" PRIu64 " sent to relayd id %" PRIu64, + stream->name, stream->key, stream->net_seq_idx); + +end: + rcu_read_unlock(); + return ret; +} + +/* + * Find a relayd and close the stream + */ +void close_relayd_stream(struct lttng_consumer_stream *stream) +{ + struct consumer_relayd_sock_pair *relayd; + + /* The stream is not metadata. Get relayd reference if exists. */ + rcu_read_lock(); + relayd = consumer_find_relayd(stream->net_seq_idx); + if (relayd) { + consumer_stream_relayd_close(stream, relayd); + } + rcu_read_unlock(); +} + +/* + * Handle stream for relayd transmission if the stream applies for network + * streaming where the net sequence index is set. + * + * Return destination file descriptor or negative value on error. + */ +static int write_relayd_stream_header(struct lttng_consumer_stream *stream, + size_t data_size, unsigned long padding, + struct consumer_relayd_sock_pair *relayd) +{ + int outfd = -1, ret; struct lttcomm_relayd_data_hdr data_hdr; /* Safety net */ @@ -718,13 +787,20 @@ static int write_relayd_stream_header(struct lttng_consumer_stream *stream, } /* Metadata are always sent on the control socket. */ - outfd = relayd->control_sock.fd; + outfd = relayd->control_sock.sock.fd; } else { /* Set header with stream information */ data_hdr.stream_id = htobe64(stream->relayd_stream_id); data_hdr.data_size = htobe32(data_size); data_hdr.padding_size = htobe32(padding); - data_hdr.net_seq_num = htobe64(stream->next_net_seq_num++); + /* + * 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 */ ret = relayd_send_data_hdr(&relayd->data_sock, &data_hdr, @@ -733,157 +809,128 @@ static int write_relayd_stream_header(struct lttng_consumer_stream *stream, goto error; } + ++stream->next_net_seq_num; + /* Set to go on data socket */ - outfd = relayd->data_sock.fd; + outfd = relayd->data_sock.sock.fd; } error: return outfd; } -static -void consumer_free_channel(struct rcu_head *head) -{ - struct lttng_ht_node_ulong *node = - caa_container_of(head, struct lttng_ht_node_ulong, head); - struct lttng_consumer_channel *channel = - caa_container_of(node, struct lttng_consumer_channel, node); - - free(channel); -} - /* - * Remove a channel from the global list protected by a mutex. This - * function is also responsible for freeing its data structures. + * Allocate and return a new lttng_consumer_channel object using the given key + * to initialize the hash table node. + * + * On error, return NULL. */ -void consumer_del_channel(struct lttng_consumer_channel *channel) -{ - int ret; - struct lttng_ht_iter iter; - - pthread_mutex_lock(&consumer_data.lock); - - switch (consumer_data.type) { - case LTTNG_CONSUMER_KERNEL: - break; - case LTTNG_CONSUMER32_UST: - case LTTNG_CONSUMER64_UST: - lttng_ustconsumer_del_channel(channel); - break; - default: - ERR("Unknown consumer_data type"); - assert(0); - goto end; - } - - rcu_read_lock(); - iter.iter.node = &channel->node.node; - ret = lttng_ht_del(consumer_data.channel_ht, &iter); - assert(!ret); - rcu_read_unlock(); - - if (channel->mmap_base != NULL) { - ret = munmap(channel->mmap_base, channel->mmap_len); - if (ret != 0) { - PERROR("munmap"); - } - } - if (channel->wait_fd >= 0 && !channel->wait_fd_is_copy) { - ret = close(channel->wait_fd); - if (ret) { - PERROR("close"); - } - } - if (channel->shm_fd >= 0 && channel->wait_fd != channel->shm_fd) { - ret = close(channel->shm_fd); - if (ret) { - PERROR("close"); - } - } - - call_rcu(&channel->node.head, consumer_free_channel); -end: - pthread_mutex_unlock(&consumer_data.lock); -} - -struct lttng_consumer_channel *consumer_allocate_channel( - int channel_key, - int shm_fd, int wait_fd, - uint64_t mmap_len, - uint64_t max_sb_size, - unsigned int nb_init_streams) +struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key, + uint64_t session_id, + const char *pathname, + const char *name, + uid_t uid, + gid_t gid, + uint64_t relayd_id, + enum lttng_event_output output, + uint64_t tracefile_size, + uint64_t tracefile_count, + uint64_t session_id_per_pid, + unsigned int monitor) { struct lttng_consumer_channel *channel; - int ret; channel = zmalloc(sizeof(*channel)); if (channel == NULL) { PERROR("malloc struct lttng_consumer_channel"); goto end; } - channel->key = channel_key; - channel->shm_fd = shm_fd; - channel->wait_fd = wait_fd; - channel->mmap_len = mmap_len; - channel->max_sb_size = max_sb_size; + + channel->key = key; channel->refcount = 0; - channel->nb_init_streams = nb_init_streams; - lttng_ht_node_init_ulong(&channel->node, channel->key); + channel->session_id = session_id; + channel->session_id_per_pid = session_id_per_pid; + channel->uid = uid; + channel->gid = gid; + channel->relayd_id = relayd_id; + channel->output = output; + channel->tracefile_size = tracefile_size; + channel->tracefile_count = tracefile_count; + channel->monitor = monitor; + pthread_mutex_init(&channel->lock, NULL); + pthread_mutex_init(&channel->timer_lock, NULL); - switch (consumer_data.type) { - case LTTNG_CONSUMER_KERNEL: - channel->mmap_base = NULL; - channel->mmap_len = 0; - break; - case LTTNG_CONSUMER32_UST: - case LTTNG_CONSUMER64_UST: - ret = lttng_ustconsumer_allocate_channel(channel); - if (ret) { - free(channel); - return NULL; - } - break; - default: - ERR("Unknown consumer_data type"); - assert(0); - goto end; + /* + * In monitor mode, the streams associated with the channel will be put in + * a special list ONLY owned by this channel. So, the refcount is set to 1 + * here meaning that the channel itself has streams that are referenced. + * + * On a channel deletion, once the channel is no longer visible, the + * refcount is decremented and checked for a zero value to delete it. With + * streams in no monitor mode, it will now be safe to destroy the channel. + */ + if (!channel->monitor) { + channel->refcount = 1; } - DBG("Allocated channel (key %d, shm_fd %d, wait_fd %d, mmap_len %llu, max_sb_size %llu)", - channel->key, channel->shm_fd, channel->wait_fd, - (unsigned long long) channel->mmap_len, - (unsigned long long) channel->max_sb_size); + + strncpy(channel->pathname, pathname, sizeof(channel->pathname)); + channel->pathname[sizeof(channel->pathname) - 1] = '\0'; + + strncpy(channel->name, name, sizeof(channel->name)); + channel->name[sizeof(channel->name) - 1] = '\0'; + + lttng_ht_node_init_u64(&channel->node, channel->key); + + channel->wait_fd = -1; + + CDS_INIT_LIST_HEAD(&channel->streams.head); + + DBG("Allocated channel (key %" PRIu64 ")", channel->key) + end: return channel; } /* * Add a channel to the global list protected by a mutex. + * + * On success 0 is returned else a negative value. */ -int consumer_add_channel(struct lttng_consumer_channel *channel) +int consumer_add_channel(struct lttng_consumer_channel *channel, + struct lttng_consumer_local_data *ctx) { - struct lttng_ht_node_ulong *node; + int ret = 0; + struct lttng_ht_node_u64 *node; struct lttng_ht_iter iter; pthread_mutex_lock(&consumer_data.lock); - /* Steal channel identifier, for UST */ - consumer_steal_channel_key(channel->key); + pthread_mutex_lock(&channel->lock); + pthread_mutex_lock(&channel->timer_lock); rcu_read_lock(); - lttng_ht_lookup(consumer_data.channel_ht, - (void *)((unsigned long) channel->key), &iter); - node = lttng_ht_iter_get_node_ulong(&iter); + lttng_ht_lookup(consumer_data.channel_ht, &channel->key, &iter); + node = lttng_ht_iter_get_node_u64(&iter); if (node != NULL) { /* Channel already exist. Ignore the insertion */ + ERR("Consumer add channel key %" PRIu64 " already exists!", + channel->key); + ret = -EEXIST; goto end; } - lttng_ht_add_unique_ulong(consumer_data.channel_ht, &channel->node); + lttng_ht_add_unique_u64(consumer_data.channel_ht, &channel->node); end: rcu_read_unlock(); + pthread_mutex_unlock(&channel->timer_lock); + pthread_mutex_unlock(&channel->lock); pthread_mutex_unlock(&consumer_data.lock); - return 0; + if (!ret && channel->wait_fd != -1 && + channel->type == CONSUMER_CHANNEL_TYPE_DATA) { + notify_channel_pipe(ctx, channel, -1, CONSUMER_CHANNEL_ADD); + } + return ret; } /* @@ -893,14 +940,19 @@ end: * * Returns the number of fds in the structures. */ -static int consumer_update_poll_array( - struct lttng_consumer_local_data *ctx, struct pollfd **pollfd, - struct lttng_consumer_stream **local_stream, struct lttng_ht *ht) +static int update_poll_array(struct lttng_consumer_local_data *ctx, + struct pollfd **pollfd, struct lttng_consumer_stream **local_stream, + struct lttng_ht *ht) { int i = 0; struct lttng_ht_iter iter; struct lttng_consumer_stream *stream; + assert(ctx); + assert(ht); + assert(pollfd); + assert(local_stream); + DBG("Updating poll fd array"); rcu_read_lock(); cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { @@ -914,10 +966,15 @@ static int consumer_update_poll_array( * changed where this function will be called back again. */ if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM || - stream->endpoint_status) { + stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) { continue; } - DBG("Active FD %d", stream->wait_fd); + /* + * This clobbers way too much the debug output. Uncomment that if you + * need it for debugging purposes. + * + * DBG("Active FD %d", stream->wait_fd); + */ (*pollfd)[i].fd = stream->wait_fd; (*pollfd)[i].events = POLLIN | POLLPRI; local_stream[i] = stream; @@ -929,7 +986,7 @@ static int consumer_update_poll_array( * 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_data_pipe[0]; + (*pollfd)[i].fd = lttng_pipe_get_readfd(ctx->consumer_data_pipe); (*pollfd)[i].events = POLLIN | POLLPRI; return i; } @@ -967,8 +1024,8 @@ exit: /* * Set the error socket. */ -void lttng_consumer_set_error_sock( - struct lttng_consumer_local_data *ctx, int sock) +void lttng_consumer_set_error_sock(struct lttng_consumer_local_data *ctx, + int sock) { ctx->consumer_error_socket = sock; } @@ -986,8 +1043,7 @@ void lttng_consumer_set_command_sock_path( * Send return code to the session daemon. * If the socket is not defined, we return 0, it is not a fatal error */ -int lttng_consumer_send_error( - struct lttng_consumer_local_data *ctx, int cmd) +int lttng_consumer_send_error(struct lttng_consumer_local_data *ctx, int cmd) { if (ctx->consumer_error_socket > 0) { return lttcomm_send_unix_sock(ctx->consumer_error_socket, &cmd, @@ -998,26 +1054,35 @@ 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) { struct lttng_ht_iter iter; - struct lttng_ht_node_ulong *node; + struct lttng_consumer_channel *channel; rcu_read_lock(); - cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, node, - node) { - struct lttng_consumer_channel *channel = - caa_container_of(node, struct lttng_consumer_channel, node); + cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, channel, + node.node) { consumer_del_channel(channel); } rcu_read_unlock(); lttng_ht_destroy(consumer_data.channel_ht); + + cleanup_relayd_ht(); + + lttng_ht_destroy(consumer_data.stream_per_chan_id_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); } /* @@ -1030,7 +1095,7 @@ 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"); } @@ -1048,11 +1113,11 @@ void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream, * Don't care about error values, as these are just hints and ways to * limit the amount of page cache used. */ - if (orig_offset < stream->chan->max_sb_size) { + if (orig_offset < stream->max_sb_size) { return; } - lttng_sync_file_range(outfd, orig_offset - stream->chan->max_sb_size, - stream->chan->max_sb_size, + lttng_sync_file_range(outfd, orig_offset - stream->max_sb_size, + stream->max_sb_size, SYNC_FILE_RANGE_WAIT_BEFORE | SYNC_FILE_RANGE_WRITE | SYNC_FILE_RANGE_WAIT_AFTER); @@ -1070,8 +1135,8 @@ void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream, * defined. So it can be expected to lead to lower throughput in * streaming. */ - posix_fadvise(outfd, orig_offset - stream->chan->max_sb_size, - stream->chan->max_sb_size, POSIX_FADV_DONTNEED); + posix_fadvise(outfd, orig_offset - stream->max_sb_size, + stream->max_sb_size, POSIX_FADV_DONTNEED); } /* @@ -1094,9 +1159,9 @@ struct lttng_consumer_local_data *lttng_consumer_create( struct lttng_consumer_local_data *ctx), int (*recv_channel)(struct lttng_consumer_channel *channel), int (*recv_stream)(struct lttng_consumer_stream *stream), - int (*update_stream)(int stream_key, uint32_t state)) + int (*update_stream)(uint64_t stream_key, uint32_t state)) { - int ret, i; + int ret; struct lttng_consumer_local_data *ctx; assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN || @@ -1110,32 +1175,19 @@ struct lttng_consumer_local_data *lttng_consumer_create( } ctx->consumer_error_socket = -1; + ctx->consumer_metadata_socket = -1; + pthread_mutex_init(&ctx->metadata_socket_lock, NULL); /* assign the callbacks */ ctx->on_buffer_ready = buffer_ready; ctx->on_recv_channel = recv_channel; ctx->on_recv_stream = recv_stream; ctx->on_update_stream = update_stream; - ret = pipe(ctx->consumer_data_pipe); - if (ret < 0) { - PERROR("Error creating poll pipe"); + ctx->consumer_data_pipe = lttng_pipe_open(0); + if (!ctx->consumer_data_pipe) { goto error_poll_pipe; } - /* set read end of the pipe to non-blocking */ - 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_data_pipe[1], F_SETFL, O_NONBLOCK); - if (ret < 0) { - PERROR("fcntl O_NONBLOCK"); - goto error_poll_fcntl; - } - ret = pipe(ctx->consumer_should_quit); if (ret < 0) { PERROR("Error creating recv pipe"); @@ -1148,8 +1200,14 @@ struct lttng_consumer_local_data *lttng_consumer_create( goto error_thread_pipe; } - ret = utils_create_pipe(ctx->consumer_metadata_pipe); + ret = pipe(ctx->consumer_channel_pipe); if (ret < 0) { + PERROR("Error creating channel pipe"); + goto error_channel_pipe; + } + + ctx->consumer_metadata_pipe = lttng_pipe_open(0); + if (!ctx->consumer_metadata_pipe) { goto error_metadata_pipe; } @@ -1161,28 +1219,15 @@ struct lttng_consumer_local_data *lttng_consumer_create( return ctx; error_splice_pipe: - utils_close_pipe(ctx->consumer_metadata_pipe); + lttng_pipe_destroy(ctx->consumer_metadata_pipe); error_metadata_pipe: + utils_close_pipe(ctx->consumer_channel_pipe); +error_channel_pipe: utils_close_pipe(ctx->consumer_thread_pipe); error_thread_pipe: - for (i = 0; i < 2; i++) { - int err; - - err = close(ctx->consumer_should_quit[i]); - if (err) { - PERROR("close"); - } - } -error_poll_fcntl: + utils_close_pipe(ctx->consumer_should_quit); error_quit_pipe: - for (i = 0; i < 2; i++) { - int err; - - err = close(ctx->consumer_data_pipe[i]); - if (err) { - PERROR("close"); - } - } + lttng_pipe_destroy(ctx->consumer_data_pipe); error_poll_pipe: free(ctx); error: @@ -1202,30 +1247,15 @@ void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx) if (ret) { PERROR("close"); } - ret = close(ctx->consumer_thread_pipe[0]); - if (ret) { - PERROR("close"); - } - ret = close(ctx->consumer_thread_pipe[1]); - if (ret) { - PERROR("close"); - } - ret = close(ctx->consumer_data_pipe[0]); - if (ret) { - PERROR("close"); - } - ret = close(ctx->consumer_data_pipe[1]); - if (ret) { - PERROR("close"); - } - ret = close(ctx->consumer_should_quit[0]); - if (ret) { - PERROR("close"); - } - ret = close(ctx->consumer_should_quit[1]); + ret = close(ctx->consumer_metadata_socket); if (ret) { PERROR("close"); } + utils_close_pipe(ctx->consumer_thread_pipe); + utils_close_pipe(ctx->consumer_channel_pipe); + lttng_pipe_destroy(ctx->consumer_data_pipe); + lttng_pipe_destroy(ctx->consumer_metadata_pipe); + utils_close_pipe(ctx->consumer_should_quit); utils_close_pipe(ctx->consumer_splice_metadata_pipe); unlink(ctx->consumer_command_sock_path); @@ -1237,8 +1267,7 @@ void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx) */ static int write_relayd_metadata_id(int fd, struct lttng_consumer_stream *stream, - struct consumer_relayd_sock_pair *relayd, - unsigned long padding) + struct consumer_relayd_sock_pair *relayd, unsigned long padding) { int ret; struct lttcomm_relayd_metadata_payload hdr; @@ -1248,8 +1277,22 @@ 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", @@ -1264,6 +1307,8 @@ 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 @@ -1274,6 +1319,7 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( unsigned long padding) { unsigned long mmap_offset; + void *mmap_base; ssize_t ret = 0, written = 0; off_t orig_offset = stream->out_fd_offset; /* Default is on the disk */ @@ -1284,10 +1330,8 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( /* 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) { + if (stream->net_seq_idx != (uint64_t) -1ULL) { relayd = consumer_find_relayd(stream->net_seq_idx); if (relayd == NULL) { goto end; @@ -1297,12 +1341,19 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( /* get the offset inside the fd to mmap */ switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: + mmap_base = stream->mmap_base; ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset); break; case LTTNG_CONSUMER32_UST: case LTTNG_CONSUMER64_UST: - ret = lttng_ustctl_get_mmap_read_offset(stream->chan->handle, - stream->buf, &mmap_offset); + mmap_base = lttng_ustctl_get_mmap_base(stream); + if (!mmap_base) { + ERR("read mmap get mmap base for stream %s", stream->name); + written = -1; + goto end; + } + ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset); + break; default: ERR("Unknown consumer_data type"); @@ -1358,15 +1409,41 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( } else { /* No streaming, we have to set the len with the full padding */ len += padding; + + /* + * Check if we need to change the tracefile before writing the packet. + */ + if (stream->chan->tracefile_size > 0 && + (stream->tracefile_size_current + len) > + stream->chan->tracefile_size) { + ret = utils_rotate_stream_file(stream->chan->pathname, + stream->name, stream->chan->tracefile_size, + stream->chan->tracefile_count, stream->uid, stream->gid, + stream->out_fd, &(stream->tracefile_count_current)); + if (ret < 0) { + ERR("Rotating output file"); + goto end; + } + outfd = stream->out_fd = ret; + /* Reset current size because we just perform a rotation. */ + stream->tracefile_size_current = 0; + } + stream->tracefile_size_current += len; } while (len > 0) { do { - ret = write(outfd, stream->mmap_base + mmap_offset, len); + ret = write(outfd, mmap_base + mmap_offset, len); } 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; } @@ -1410,7 +1487,6 @@ end: if (relayd && stream->metadata_flag) { pthread_mutex_unlock(&relayd->ctrl_sock_mutex); } - pthread_mutex_unlock(&stream->lock); rcu_read_unlock(); return written; @@ -1419,6 +1495,8 @@ 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( @@ -1451,10 +1529,8 @@ 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) { + if (stream->net_seq_idx != (uint64_t) -1ULL) { relayd = consumer_find_relayd(stream->net_seq_idx); if (relayd == NULL) { goto end; @@ -1515,6 +1591,26 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( } else { /* No streaming, we have to set the len with the full padding */ len += padding; + + /* + * Check if we need to change the tracefile before writing the packet. + */ + if (stream->chan->tracefile_size > 0 && + (stream->tracefile_size_current + len) > + stream->chan->tracefile_size) { + ret = utils_rotate_stream_file(stream->chan->pathname, + stream->name, stream->chan->tracefile_size, + stream->chan->tracefile_count, stream->uid, stream->gid, + stream->out_fd, &(stream->tracefile_count_current)); + if (ret < 0) { + ERR("Rotating output file"); + goto end; + } + outfd = stream->out_fd = ret; + /* Reset current size because we just perform a rotation. */ + stream->tracefile_size_current = 0; + } + stream->tracefile_size_current += len; } while (len > 0) { @@ -1559,7 +1655,7 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( written = ret_splice; } /* Socket operation failed. We consider the relayd dead */ - if (errno == EBADF) { + if (errno == EBADF || errno == EPIPE) { WARN("Remote relayd disconnected. Stopping"); relayd_hang_up = 1; goto write_error; @@ -1620,7 +1716,6 @@ end: if (relayd && stream->metadata_flag) { pthread_mutex_unlock(&relayd->ctrl_sock_mutex); } - pthread_mutex_unlock(&stream->lock); rcu_read_unlock(); return written; @@ -1631,21 +1726,19 @@ end: * * Returns 0 on success, < 0 on error */ -int lttng_consumer_take_snapshot(struct lttng_consumer_local_data *ctx, - struct lttng_consumer_stream *stream) +int lttng_consumer_take_snapshot(struct lttng_consumer_stream *stream) { switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: - return lttng_kconsumer_take_snapshot(ctx, stream); + return lttng_kconsumer_take_snapshot(stream); case LTTNG_CONSUMER32_UST: case LTTNG_CONSUMER64_UST: - return lttng_ustconsumer_take_snapshot(ctx, stream); + return lttng_ustconsumer_take_snapshot(stream); default: ERR("Unknown consumer_data type"); assert(0); return -ENOSYS; } - } /* @@ -1653,17 +1746,15 @@ int lttng_consumer_take_snapshot(struct lttng_consumer_local_data *ctx, * * Returns 0 on success, < 0 on error */ -int lttng_consumer_get_produced_snapshot( - struct lttng_consumer_local_data *ctx, - struct lttng_consumer_stream *stream, +int lttng_consumer_get_produced_snapshot(struct lttng_consumer_stream *stream, unsigned long *pos) { switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: - return lttng_kconsumer_get_produced_snapshot(ctx, stream, pos); + return lttng_kconsumer_get_produced_snapshot(stream, pos); case LTTNG_CONSUMER32_UST: case LTTNG_CONSUMER64_UST: - return lttng_ustconsumer_get_produced_snapshot(ctx, stream, pos); + return lttng_ustconsumer_get_produced_snapshot(stream, pos); default: ERR("Unknown consumer_data type"); assert(0); @@ -1694,7 +1785,6 @@ int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx, */ static void destroy_data_stream_ht(struct lttng_ht *ht) { - int ret; struct lttng_ht_iter iter; struct lttng_consumer_stream *stream; @@ -1704,10 +1794,11 @@ static void destroy_data_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(); @@ -1721,7 +1812,6 @@ static void destroy_data_stream_ht(struct lttng_ht *ht) */ static void destroy_stream_ht(struct lttng_ht *ht) { - int ret; struct lttng_ht_iter iter; struct lttng_consumer_stream *stream; @@ -1731,16 +1821,43 @@ 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_metadata_stream(stream, ht); } rcu_read_unlock(); lttng_ht_destroy(ht); } +void lttng_consumer_close_metadata(void) +{ + switch (consumer_data.type) { + case LTTNG_CONSUMER_KERNEL: + /* + * The Kernel consumer has a different metadata scheme so we don't + * close anything because the stream will be closed by the session + * daemon. + */ + break; + case LTTNG_CONSUMER32_UST: + case LTTNG_CONSUMER64_UST: + /* + * Close all metadata streams. The metadata hash table is passed and + * this call iterates over it by closing all wakeup fd. This is safe + * because at this point we are sure that the metadata producer is + * either dead or blocked. + */ + lttng_ustconsumer_close_metadata(metadata_ht); + break; + default: + ERR("Unknown consumer_data type"); + assert(0); + } +} + /* * Clean up a metadata stream and free its memory. */ @@ -1763,12 +1880,13 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream, if (ht == NULL) { /* Means the stream was allocated but not successfully added */ - goto free_stream; + goto free_stream_rcu; } + pthread_mutex_lock(&consumer_data.lock); + pthread_mutex_lock(&stream->chan->lock); pthread_mutex_lock(&stream->lock); - pthread_mutex_lock(&consumer_data.lock); switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: if (stream->mmap_base != NULL) { @@ -1777,9 +1895,22 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream, PERROR("munmap metadata stream"); } } + if (stream->wait_fd >= 0) { + ret = close(stream->wait_fd); + if (ret < 0) { + PERROR("close kernel metadata wait_fd"); + } + } break; case LTTNG_CONSUMER32_UST: case LTTNG_CONSUMER64_UST: + if (stream->monitor) { + /* close the write-side in close_metadata */ + ret = close(stream->ust_metadata_poll_pipe[0]); + if (ret < 0) { + PERROR("Close UST metadata read-side poll pipe"); + } + } lttng_ustconsumer_del_stream(stream); break; default: @@ -1793,8 +1924,11 @@ void consumer_del_metadata_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; + iter.iter.node = &stream->node_channel_id.node; + ret = lttng_ht_del(consumer_data.stream_per_chan_id_ht, &iter); + assert(!ret); + + iter.iter.node = &stream->node_session_id.node; ret = lttng_ht_del(consumer_data.stream_list_ht, &iter); assert(!ret); rcu_read_unlock(); @@ -1806,20 +1940,6 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream, } } - if (stream->wait_fd >= 0 && !stream->wait_fd_is_copy) { - ret = close(stream->wait_fd); - if (ret) { - PERROR("close"); - } - } - - if (stream->shm_fd >= 0 && stream->wait_fd != stream->shm_fd) { - ret = close(stream->shm_fd); - if (ret) { - PERROR("close"); - } - } - /* Check and cleanup relayd */ rcu_read_lock(); relayd = consumer_find_relayd(stream->net_seq_idx); @@ -1844,47 +1964,58 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream, /* Both conditions are met, we destroy the relayd. */ if (uatomic_read(&relayd->refcount) == 0 && uatomic_read(&relayd->destroy_flag)) { - destroy_relayd(relayd); + consumer_destroy_relayd(relayd); } } rcu_read_unlock(); /* Atomically decrement channel refcount since other threads can use it. */ - uatomic_dec(&stream->chan->refcount); - if (!uatomic_read(&stream->chan->refcount) - && !uatomic_read(&stream->chan->nb_init_streams)) { + if (!uatomic_sub_return(&stream->chan->refcount, 1) + && !uatomic_read(&stream->chan->nb_init_stream_left)) { /* Go for channel deletion! */ free_chan = stream->chan; } end: - pthread_mutex_unlock(&consumer_data.lock); + /* + * Nullify the stream reference so it is not used after deletion. The + * channel lock MUST be acquired before being able to check for + * a NULL pointer value. + */ + stream->chan->metadata_stream = NULL; + pthread_mutex_unlock(&stream->lock); + pthread_mutex_unlock(&stream->chan->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); +free_stream_rcu: + call_rcu(&stream->node.head, free_stream_rcu); } /* * Action done with the metadata stream when adding it to the consumer internal * data structures to handle it. */ -static int consumer_add_metadata_stream(struct lttng_consumer_stream *stream, - struct lttng_ht *ht) +int consumer_add_metadata_stream(struct lttng_consumer_stream *stream) { + struct lttng_ht *ht = metadata_ht; int ret = 0; - struct consumer_relayd_sock_pair *relayd; + struct lttng_ht_iter iter; + struct lttng_ht_node_u64 *node; assert(stream); assert(ht); - DBG3("Adding metadata stream %d to hash table", stream->wait_fd); + DBG3("Adding metadata stream %" PRIu64 " to hash table", stream->key); pthread_mutex_lock(&consumer_data.lock); + pthread_mutex_lock(&stream->chan->lock); + pthread_mutex_lock(&stream->chan->timer_lock); + pthread_mutex_lock(&stream->lock); /* * From here, refcounts are updated so be _careful_ when returning an error @@ -1892,40 +2023,45 @@ static int consumer_add_metadata_stream(struct lttng_consumer_stream *stream, */ rcu_read_lock(); - /* 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); + /* + * Lookup the stream just to make sure it does not exist in our internal + * state. This should NEVER happen. + */ + lttng_ht_lookup(ht, &stream->key, &iter); + node = lttng_ht_iter_get_node_u64(&iter); + assert(!node); /* - * 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 + * When nb_init_stream_left 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); + if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) { + /* Increment refcount before decrementing nb_init_stream_left */ + cmm_smp_wmb(); + uatomic_dec(&stream->chan->nb_init_stream_left); } - /* Steal stream identifier to avoid having streams with the same key */ - consumer_steal_stream_key(stream->key, ht); + lttng_ht_add_unique_u64(ht, &stream->node); - lttng_ht_add_unique_ulong(ht, &stream->node); + lttng_ht_add_unique_u64(consumer_data.stream_per_chan_id_ht, + &stream->node_channel_id); /* * 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); + lttng_ht_add_u64(consumer_data.stream_list_ht, &stream->node_session_id); rcu_read_unlock(); + pthread_mutex_unlock(&stream->lock); + pthread_mutex_unlock(&stream->chan->lock); + pthread_mutex_unlock(&stream->chan->timer_lock); pthread_mutex_unlock(&consumer_data.lock); return ret; } @@ -1943,7 +2079,7 @@ static void validate_endpoint_status_data_stream(void) 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_INACTIVE) { + if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) { continue; } /* Delete it right now */ @@ -1968,7 +2104,7 @@ static void validate_endpoint_status_metadata_stream( 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) { + if (stream->endpoint_status == CONSUMER_ENDPOINT_ACTIVE) { continue; } /* @@ -1993,23 +2129,30 @@ void *consumer_thread_metadata_poll(void *data) uint32_t revents, nb_fd; struct lttng_consumer_stream *stream = NULL; struct lttng_ht_iter iter; - struct lttng_ht_node_ulong *node; + struct lttng_ht_node_u64 *node; struct lttng_poll_event events; struct lttng_consumer_local_data *ctx = data; ssize_t len; rcu_register_thread(); + metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); + if (!metadata_ht) { + /* ENOMEM at this point. Better to bail out. */ + goto end_ht; + } + 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) { ERR("Poll set creation failed"); - goto end; + goto end_poll; } - ret = lttng_poll_add(&events, ctx->consumer_metadata_pipe[0], LPOLLIN); + ret = lttng_poll_add(&events, + lttng_pipe_get_readfd(ctx->consumer_metadata_pipe), LPOLLIN); if (ret < 0) { goto end; } @@ -2018,17 +2161,13 @@ void *consumer_thread_metadata_poll(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) { @@ -2039,6 +2178,8 @@ 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); @@ -2049,27 +2190,26 @@ restart: continue; } - if (pollfd == ctx->consumer_metadata_pipe[0]) { + if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) { if (revents & (LPOLLERR | LPOLLHUP )) { DBG("Metadata thread pipe hung up"); /* * Remove the pipe from the poll set and continue the loop * since their might be data to consume. */ - lttng_poll_del(&events, ctx->consumer_metadata_pipe[0]); - close(ctx->consumer_metadata_pipe[0]); + lttng_poll_del(&events, + lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)); + lttng_pipe_read_close(ctx->consumer_metadata_pipe); continue; } else if (revents & LPOLLIN) { - do { - /* Get the stream pointer received */ - ret = read(pollfd, &stream, sizeof(stream)); - } while (ret < 0 && errno == EINTR); - if (ret < 0 || - ret < sizeof(struct lttng_consumer_stream *)) { - PERROR("read metadata stream"); + ssize_t pipe_len; + + pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe, + &stream, sizeof(stream)); + if (pipe_len < 0) { + ERR("read metadata stream, ret: %zd", pipe_len); /* - * Let's continue here and hope we can still work - * without stopping the consumer. XXX: Should we? + * Continue here to handle the rest of the streams. */ continue; } @@ -2078,20 +2218,12 @@ restart: if (stream == NULL) { /* Check for deleted streams. */ validate_endpoint_status_metadata_stream(&events); - continue; + goto restart; } DBG("Adding metadata stream %d to poll set", stream->wait_fd); - 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); @@ -2102,9 +2234,12 @@ restart: } rcu_read_lock(); - lttng_ht_lookup(metadata_ht, (void *)((unsigned long) pollfd), - &iter); - node = lttng_ht_iter_get_node_ulong(&iter); + { + uint64_t tmp_id = (uint64_t) pollfd; + + lttng_ht_lookup(metadata_ht, &tmp_id, &iter); + } + node = lttng_ht_iter_get_node_u64(&iter); assert(node); stream = caa_container_of(node, struct lttng_consumer_stream, @@ -2142,14 +2277,21 @@ restart: DBG("Metadata available on fd %d", pollfd); assert(stream->wait_fd == pollfd); - len = ctx->on_buffer_ready(stream, ctx); + 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); + /* 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; } } @@ -2161,12 +2303,11 @@ restart: error: end: DBG("Metadata poll thread exiting"); - lttng_poll_clean(&events); - - if (metadata_ht) { - destroy_stream_ht(metadata_ht); - } + lttng_poll_clean(&events); +end_poll: + destroy_stream_ht(metadata_ht); +end_ht: rcu_unregister_thread(); return NULL; } @@ -2188,12 +2329,17 @@ void *consumer_thread_data_poll(void *data) rcu_register_thread(); - data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); + data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); if (data_ht == NULL) { + /* ENOMEM at this point. Better to bail out. */ goto end; } - local_stream = zmalloc(sizeof(struct lttng_consumer_stream)); + local_stream = zmalloc(sizeof(struct lttng_consumer_stream *)); + if (local_stream == NULL) { + PERROR("local_stream malloc"); + goto end; + } while (1) { high_prio = 0; @@ -2205,14 +2351,11 @@ void *consumer_thread_data_poll(void *data) */ pthread_mutex_lock(&consumer_data.lock); if (consumer_data.need_update) { - if (pollfd != NULL) { - free(pollfd); - pollfd = NULL; - } - if (local_stream != NULL) { - free(local_stream); - local_stream = NULL; - } + free(pollfd); + pollfd = NULL; + + free(local_stream); + local_stream = NULL; /* allocate for all fds + 1 for the consumer_data_pipe */ pollfd = zmalloc((consumer_data.stream_count + 1) * sizeof(struct pollfd)); @@ -2224,13 +2367,13 @@ void *consumer_thread_data_poll(void *data) /* allocate for all fds + 1 for the consumer_data_pipe */ local_stream = zmalloc((consumer_data.stream_count + 1) * - sizeof(struct lttng_consumer_stream)); + sizeof(struct lttng_consumer_stream *)); if (local_stream == NULL) { PERROR("local_stream malloc"); pthread_mutex_unlock(&consumer_data.lock); goto end; } - ret = consumer_update_poll_array(ctx, &pollfd, local_stream, + ret = update_poll_array(ctx, &pollfd, local_stream, data_ht); if (ret < 0) { ERR("Error in allocating pollfd or local_outfds"); @@ -2250,7 +2393,7 @@ void *consumer_thread_data_poll(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) { /* @@ -2273,14 +2416,16 @@ void *consumer_thread_data_poll(void *data) * array update over low-priority reads. */ if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) { - size_t pipe_readlen; + ssize_t pipe_readlen; DBG("consumer_data_pipe wake up"); - /* Consume 1 byte of pipe data */ - do { - pipe_readlen = read(ctx->consumer_data_pipe[0], &new_stream, - sizeof(new_stream)); - } while (pipe_readlen == -1 && errno == EINTR); + pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe, + &new_stream, sizeof(new_stream)); + if (pipe_readlen < 0) { + ERR("Consumer data pipe ret %zd", pipe_readlen); + /* 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 @@ -2292,17 +2437,6 @@ void *consumer_thread_data_poll(void *data) 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; } @@ -2404,14 +2538,8 @@ void *consumer_thread_data_poll(void *data) } end: DBG("polling thread exiting"); - if (pollfd != NULL) { - free(pollfd); - pollfd = NULL; - } - if (local_stream != NULL) { - free(local_stream); - local_stream = NULL; - } + free(pollfd); + free(local_stream); /* * Close the write side of the pipe so epoll_wait() in @@ -2421,23 +2549,329 @@ end: * only tracked fd in the poll set. The thread will take care of closing * the read side. */ - close(ctx->consumer_metadata_pipe[1]); + (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe); + + destroy_data_stream_ht(data_ht); + + rcu_unregister_thread(); + return NULL; +} + +/* + * Close wake-up end of each stream belonging to the channel. This will + * allow the poll() on the stream read-side to detect when the + * write-side (application) finally closes them. + */ +static +void consumer_close_channel_streams(struct lttng_consumer_channel *channel) +{ + struct lttng_ht *ht; + struct lttng_consumer_stream *stream; + struct lttng_ht_iter iter; + + ht = consumer_data.stream_per_chan_id_ht; + + rcu_read_lock(); + cds_lfht_for_each_entry_duplicate(ht->ht, + ht->hash_fct(&channel->key, lttng_ht_seed), + ht->match_fct, &channel->key, + &iter.iter, stream, node_channel_id.node) { + /* + * Protect against teardown with mutex. + */ + pthread_mutex_lock(&stream->lock); + if (cds_lfht_is_node_deleted(&stream->node.node)) { + goto next; + } + switch (consumer_data.type) { + case LTTNG_CONSUMER_KERNEL: + break; + case LTTNG_CONSUMER32_UST: + case LTTNG_CONSUMER64_UST: + /* + * Note: a mutex is taken internally within + * liblttng-ust-ctl to protect timer wakeup_fd + * use from concurrent close. + */ + lttng_ustconsumer_close_stream_wakeup(stream); + break; + default: + ERR("Unknown consumer_data type"); + assert(0); + } + next: + pthread_mutex_unlock(&stream->lock); + } + rcu_read_unlock(); +} + +static void destroy_channel_ht(struct lttng_ht *ht) +{ + struct lttng_ht_iter iter; + struct lttng_consumer_channel *channel; + int ret; + + if (ht == NULL) { + return; + } + + rcu_read_lock(); + cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) { + ret = lttng_ht_del(ht, &iter); + assert(ret != 0); + } + rcu_read_unlock(); + + lttng_ht_destroy(ht); +} + +/* + * This thread polls the channel fds to detect when they are being + * closed. It closes all related streams if the channel is detected as + * closed. It is currently only used as a shim layer for UST because the + * consumerd needs to keep the per-stream wakeup end of pipes open for + * periodical flush. + */ +void *consumer_thread_channel_poll(void *data) +{ + int ret, i, pollfd; + uint32_t revents, nb_fd; + struct lttng_consumer_channel *chan = NULL; + struct lttng_ht_iter iter; + struct lttng_ht_node_u64 *node; + struct lttng_poll_event events; + struct lttng_consumer_local_data *ctx = data; + struct lttng_ht *channel_ht; + + rcu_register_thread(); + + channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); + if (!channel_ht) { + /* ENOMEM at this point. Better to bail out. */ + goto end_ht; + } + + DBG("Thread channel poll started"); + + /* Size is set to 1 for the consumer_channel pipe */ + ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC); + if (ret < 0) { + ERR("Poll set creation failed"); + goto end_poll; + } - if (data_ht) { - destroy_data_stream_ht(data_ht); + ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN); + if (ret < 0) { + goto end; } + /* Main loop */ + DBG("Channel main loop started"); + + while (1) { + /* Only the channel pipe is set */ + if (LTTNG_POLL_GETNB(&events) == 0 && consumer_quit == 1) { + goto end; + } + +restart: + DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events)); + ret = lttng_poll_wait(&events, -1); + DBG("Channel event catched in thread"); + if (ret < 0) { + if (errno == EINTR) { + ERR("Poll EINTR catched"); + goto restart; + } + goto end; + } + + nb_fd = ret; + + /* From here, the event is a channel wait fd */ + for (i = 0; i < nb_fd; i++) { + revents = LTTNG_POLL_GETEV(&events, i); + pollfd = LTTNG_POLL_GETFD(&events, i); + + /* Just don't waste time if no returned events for the fd */ + if (!revents) { + continue; + } + if (pollfd == ctx->consumer_channel_pipe[0]) { + if (revents & (LPOLLERR | LPOLLHUP)) { + DBG("Channel thread pipe hung up"); + /* + * Remove the pipe from the poll set and continue the loop + * since their might be data to consume. + */ + lttng_poll_del(&events, ctx->consumer_channel_pipe[0]); + continue; + } else if (revents & LPOLLIN) { + enum consumer_channel_action action; + uint64_t key; + + ret = read_channel_pipe(ctx, &chan, &key, &action); + if (ret <= 0) { + ERR("Error reading channel pipe"); + continue; + } + + switch (action) { + case CONSUMER_CHANNEL_ADD: + DBG("Adding channel %d to poll set", + chan->wait_fd); + + lttng_ht_node_init_u64(&chan->wait_fd_node, + chan->wait_fd); + rcu_read_lock(); + lttng_ht_add_unique_u64(channel_ht, + &chan->wait_fd_node); + rcu_read_unlock(); + /* Add channel to the global poll events list */ + lttng_poll_add(&events, chan->wait_fd, + LPOLLIN | LPOLLPRI); + break; + case CONSUMER_CHANNEL_DEL: + { + struct lttng_consumer_stream *stream, *stmp; + + rcu_read_lock(); + chan = consumer_find_channel(key); + if (!chan) { + rcu_read_unlock(); + ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key); + break; + } + lttng_poll_del(&events, chan->wait_fd); + iter.iter.node = &chan->wait_fd_node.node; + ret = lttng_ht_del(channel_ht, &iter); + assert(ret == 0); + consumer_close_channel_streams(chan); + + switch (consumer_data.type) { + case LTTNG_CONSUMER_KERNEL: + break; + case LTTNG_CONSUMER32_UST: + case LTTNG_CONSUMER64_UST: + /* Delete streams that might have been left in the stream list. */ + cds_list_for_each_entry_safe(stream, stmp, &chan->streams.head, + send_node) { + cds_list_del(&stream->send_node); + lttng_ustconsumer_del_stream(stream); + uatomic_sub(&stream->chan->refcount, 1); + assert(&chan->refcount); + free(stream); + } + break; + default: + ERR("Unknown consumer_data type"); + assert(0); + } + + /* + * Release our own refcount. Force channel deletion even if + * streams were not initialized. + */ + if (!uatomic_sub_return(&chan->refcount, 1)) { + consumer_del_channel(chan); + } + rcu_read_unlock(); + goto restart; + } + case CONSUMER_CHANNEL_QUIT: + /* + * Remove the pipe from the poll set and continue the loop + * since their might be data to consume. + */ + lttng_poll_del(&events, ctx->consumer_channel_pipe[0]); + continue; + default: + ERR("Unknown action"); + break; + } + } + + /* Handle other stream */ + continue; + } + + rcu_read_lock(); + { + uint64_t tmp_id = (uint64_t) pollfd; + + lttng_ht_lookup(channel_ht, &tmp_id, &iter); + } + node = lttng_ht_iter_get_node_u64(&iter); + assert(node); + + chan = caa_container_of(node, struct lttng_consumer_channel, + wait_fd_node); + + /* Check for error event */ + if (revents & (LPOLLERR | LPOLLHUP)) { + DBG("Channel fd %d is hup|err.", pollfd); + + lttng_poll_del(&events, chan->wait_fd); + ret = lttng_ht_del(channel_ht, &iter); + assert(ret == 0); + consumer_close_channel_streams(chan); + + /* Release our own refcount */ + if (!uatomic_sub_return(&chan->refcount, 1) + && !uatomic_read(&chan->nb_init_stream_left)) { + consumer_del_channel(chan); + } + } + + /* Release RCU lock for the channel looked up */ + rcu_read_unlock(); + } + } + +end: + lttng_poll_clean(&events); +end_poll: + destroy_channel_ht(channel_ht); +end_ht: + DBG("Channel poll thread exiting"); rcu_unregister_thread(); return NULL; } +static int set_metadata_socket(struct lttng_consumer_local_data *ctx, + struct pollfd *sockpoll, int client_socket) +{ + int ret; + + assert(ctx); + assert(sockpoll); + + if (lttng_consumer_poll_socket(sockpoll) < 0) { + ret = -1; + goto error; + } + DBG("Metadata connection on client_socket"); + + /* Blocking call, waiting for transmission */ + ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket); + if (ctx->consumer_metadata_socket < 0) { + WARN("On accept metadata"); + ret = -1; + goto error; + } + ret = 0; + +error: + return ret; +} + /* * This thread listens on the consumerd socket and receives the file * descriptors from the session daemon. */ 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. @@ -2468,12 +2902,6 @@ void *consumer_thread_sessiond_poll(void *data) goto end; } - ret = fcntl(client_socket, F_SETFL, O_NONBLOCK); - if (ret < 0) { - PERROR("fcntl O_NONBLOCK"); - goto end; - } - /* prepare the FDs to poll : to client socket and the should_quit pipe */ consumer_sockpoll[0].fd = ctx->consumer_should_quit[0]; consumer_sockpoll[0].events = POLLIN | POLLPRI; @@ -2487,16 +2915,27 @@ void *consumer_thread_sessiond_poll(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; } - ret = fcntl(sock, F_SETFL, O_NONBLOCK); + + /* + * Setup metadata socket which is the second socket connection on the + * command unix socket. + */ + ret = set_metadata_socket(ctx, consumer_sockpoll, client_socket); if (ret < 0) { - PERROR("fcntl O_NONBLOCK"); 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; @@ -2523,29 +2962,46 @@ void *consumer_thread_sessiond_poll(void *data) DBG("consumer_thread_receive_fds received quit from signal"); goto end; } - DBG("received fds on sock"); + DBG("received command on sock"); } end: - DBG("consumer_thread_receive_fds exiting"); + DBG("Consumer thread sessiond poll exiting"); /* - * when all fds have hung up, the polling thread - * can exit cleanly + * Close metadata streams since the producer is the session daemon which + * just died. + * + * NOTE: for now, this only applies to the UST tracer. */ - consumer_quit = 1; + lttng_consumer_close_metadata(); /* - * 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). + * when all fds have hung up, the polling thread + * can exit cleanly */ - consumer_poll_timeout = LTTNG_CONSUMER_POLL_TIMEOUT; + consumer_quit = 1; /* * Notify the data poll thread to poll back again and test the * consumer_quit state that we just set so to quit gracefully. */ - notify_thread_pipe(ctx->consumer_data_pipe[1]); + notify_thread_lttng_pipe(ctx->consumer_data_pipe); + + notify_channel_pipe(ctx, NULL, -1, CONSUMER_CHANNEL_QUIT); + + /* 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(client_socket); + if (ret < 0) { + PERROR("close client_socket sessiond poll"); + } + } rcu_unregister_thread(); return NULL; @@ -2554,17 +3010,27 @@ 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) @@ -2587,14 +3053,10 @@ int lttng_consumer_on_recv_stream(struct lttng_consumer_stream *stream) */ void lttng_consumer_init(void) { - 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); - data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); - assert(data_ht); + consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); + consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); + consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); + consumer_data.stream_per_chan_id_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); } /* @@ -2603,37 +3065,79 @@ void lttng_consumer_init(void) * This will create a relayd socket pair and add it to the relayd hash table. * The caller MUST acquire a RCU read side lock before calling it. */ -int consumer_add_relayd_socket(int net_seq_idx, int sock_type, +int consumer_add_relayd_socket(uint64_t 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_relayd_sock *relayd_sock, uint64_t sessiond_id) { - int fd, ret = -1; - struct consumer_relayd_sock_pair *relayd; + int fd = -1, ret = -1, relayd_created = 0; + enum lttng_error_code ret_code = LTTNG_OK; + struct consumer_relayd_sock_pair *relayd = NULL; - DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx); + assert(ctx); + assert(relayd_sock); + + DBG("Consumer adding relayd socket (idx: %" PRIu64 ")", net_seq_idx); /* Get relayd reference if exists. */ relayd = consumer_find_relayd(net_seq_idx); if (relayd == NULL) { + assert(sock_type == LTTNG_STREAM_CONTROL); /* Not found. Allocate one. */ relayd = consumer_allocate_relayd_sock_pair(net_seq_idx); if (relayd == NULL) { - lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_OUTFD_ERROR); + ret = -ENOMEM; + ret_code = LTTCOMM_CONSUMERD_ENOMEM; goto error; + } else { + relayd->sessiond_session_id = sessiond_id; + relayd_created = 1; } + + /* + * This code path MUST continue to the consumer send status message to + * we can notify the session daemon and continue our work without + * killing everything. + */ + } else { + /* + * relayd key should never be found for control socket. + */ + assert(sock_type != LTTNG_STREAM_CONTROL); + } + + /* First send a status message before receiving the fds. */ + ret = consumer_send_status_msg(sock, LTTNG_OK); + if (ret < 0) { + /* Somehow, the session daemon is not responding anymore. */ + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL); + goto error_nosignal; } /* Poll on consumer socket. */ if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) { + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR); ret = -EINTR; - goto error; + goto error_nosignal; } /* Get relayd socket from session daemon */ ret = lttcomm_recv_fds_unix_sock(sock, &fd, 1); 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. */ + + /* + * Failing to receive FDs might indicate a major problem such as + * reaching a fd limit during the receive where the kernel returns a + * MSG_CTRUNC and fails to cleanup the fd in the queue. Any case, we + * don't take any chances and stop everything. + * + * XXX: Feature request #558 will fix that and avoid this possible + * issue when reaching the fd limit. + */ + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD); + ret_code = LTTCOMM_CONSUMERD_ERROR_RECV_FD; goto error; } @@ -2641,41 +3145,98 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, switch (sock_type) { case LTTNG_STREAM_CONTROL: /* Copy received lttcomm socket */ - lttcomm_copy_sock(&relayd->control_sock, relayd_sock); - ret = lttcomm_create_sock(&relayd->control_sock); + lttcomm_copy_sock(&relayd->control_sock.sock, &relayd_sock->sock); + ret = lttcomm_create_sock(&relayd->control_sock.sock); + /* Handle create_sock error. */ if (ret < 0) { + ret_code = LTTCOMM_CONSUMERD_ENOMEM; goto error; } - - /* Close the created socket fd which is useless */ - close(relayd->control_sock.fd); + /* + * Close the socket created internally by + * lttcomm_create_sock, so we can replace it by the one + * received from sessiond. + */ + if (close(relayd->control_sock.sock.fd)) { + PERROR("close"); + } /* Assign new file descriptor */ - relayd->control_sock.fd = fd; + relayd->control_sock.sock.fd = fd; + fd = -1; /* For error path */ + /* Assign version values. */ + relayd->control_sock.major = relayd_sock->major; + relayd->control_sock.minor = relayd_sock->minor; + + /* + * 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) { + /* + * Close all sockets of a relayd object. It will be freed if it was + * created at the error code path or else it will be garbage + * collect. + */ + (void) relayd_close(&relayd->control_sock); + (void) relayd_close(&relayd->data_sock); + ret_code = LTTCOMM_CONSUMERD_RELAYD_FAIL; + goto error; + } + break; case LTTNG_STREAM_DATA: /* Copy received lttcomm socket */ - lttcomm_copy_sock(&relayd->data_sock, relayd_sock); - ret = lttcomm_create_sock(&relayd->data_sock); + lttcomm_copy_sock(&relayd->data_sock.sock, &relayd_sock->sock); + ret = lttcomm_create_sock(&relayd->data_sock.sock); + /* Handle create_sock error. */ if (ret < 0) { + ret_code = LTTCOMM_CONSUMERD_ENOMEM; goto error; } - - /* Close the created socket fd which is useless */ - close(relayd->data_sock.fd); + /* + * Close the socket created internally by + * lttcomm_create_sock, so we can replace it by the one + * received from sessiond. + */ + if (close(relayd->data_sock.sock.fd)) { + PERROR("close"); + } /* Assign new file descriptor */ - relayd->data_sock.fd = fd; + relayd->data_sock.sock.fd = fd; + fd = -1; /* for eventual error paths */ + /* Assign version values. */ + relayd->data_sock.major = relayd_sock->major; + relayd->data_sock.minor = relayd_sock->minor; break; default: ERR("Unknown relayd socket type (%d)", sock_type); + ret = -1; + ret_code = LTTCOMM_CONSUMERD_FATAL; goto error; } - DBG("Consumer %s socket created successfully with net idx %d (fd: %d)", + DBG("Consumer %s socket created successfully with net idx %" PRIu64 " (fd: %d)", sock_type == LTTNG_STREAM_CONTROL ? "control" : "data", relayd->net_seq_idx, fd); + /* We successfully added the socket. Send status back. */ + ret = consumer_send_status_msg(sock, ret_code); + if (ret < 0) { + /* Somehow, the session daemon is not responding anymore. */ + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL); + goto error_nosignal; + } + /* * Add relayd socket pair to consumer data hashtable. If object already * exists or on error, the function gracefully returns. @@ -2683,9 +3244,25 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, add_relayd(relayd); /* All good! */ - ret = 0; + return 0; error: + if (consumer_send_status_msg(sock, ret_code) < 0) { + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_FATAL); + } + +error_nosignal: + /* Close received socket if valid. */ + if (fd >= 0) { + if (close(fd)) { + PERROR("close received socket"); + } + } + + if (relayd_created) { + free(relayd); + } + return ret; } @@ -2717,6 +3294,36 @@ 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. @@ -2729,7 +3336,7 @@ int consumer_data_pending(uint64_t id) struct lttng_ht_iter iter; struct lttng_ht *ht; struct lttng_consumer_stream *stream; - struct consumer_relayd_sock_pair *relayd; + struct consumer_relayd_sock_pair *relayd = NULL; int (*data_pending)(struct lttng_consumer_stream *); DBG("Consumer data pending command on session id %" PRIu64, id); @@ -2753,14 +3360,27 @@ int consumer_data_pending(uint64_t id) /* 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), 0x42UL), - ht->match_fct, (void *)((unsigned long) id), + ht->hash_fct(&id, lttng_ht_seed), + ht->match_fct, &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_pending; + goto data_pending; } /* @@ -2775,55 +3395,118 @@ int consumer_data_pending(uint64_t id) ret = data_pending(stream); if (ret == 1) { pthread_mutex_unlock(&stream->lock); - goto data_not_pending; + goto data_pending; } } /* 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_pending; - } - + if (relayd) { pthread_mutex_lock(&relayd->ctrl_sock_mutex); if (stream->metadata_flag) { - ret = relayd_quiescent_control(&relayd->control_sock); + 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); + 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_not_pending; + 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 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. + * 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_not_pending: +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)); +} + +/* + * Send a channel status message to the sessiond daemon. + * + * Return the sendmsg() return value. + */ +int consumer_send_status_channel(int sock, + struct lttng_consumer_channel *channel) +{ + struct lttcomm_consumer_status_channel msg; + + assert(sock >= 0); + + if (!channel) { + msg.ret_code = -LTTNG_ERR_UST_CHAN_FAIL; + } else { + msg.ret_code = LTTNG_OK; + msg.key = channel->key; + msg.stream_count = channel->streams.count; + } + + return lttcomm_send_unix_sock(sock, &msg, sizeof(msg)); +} + +/* + * Using a maximum stream size with the produced and consumed position of a + * stream, computes the new consumed position to be as close as possible to the + * maximum possible stream size. + * + * If maximum stream size is lower than the possible buffer size (produced - + * consumed), the consumed_pos given is returned untouched else the new value + * is returned. + */ +unsigned long consumer_get_consumed_maxsize(unsigned long consumed_pos, + unsigned long produced_pos, uint64_t max_stream_size) +{ + if (max_stream_size && max_stream_size < (produced_pos - consumed_pos)) { + /* Offset from the produced position to get the latest buffers. */ + return produced_pos - max_stream_size; + } + + return consumed_pos; +}