X-Git-Url: https://git.lttng.org/?p=lttng-tools.git;a=blobdiff_plain;f=src%2Fcommon%2Fconsumer.c;h=b0b926bb01a1f65c58c9d25da47bc980b4125e83;hp=a9070b1c9bc3f976ef26a2fbbc183c7fd4b75343;hb=9d0482e7fb70c4deaebe57872cbd75a6a1c213fb;hpb=6151a90fe7fa3dea52c57771df9083e56de7a60b diff --git a/src/common/consumer.c b/src/common/consumer.c index a9070b1c9..b0b926bb0 100644 --- a/src/common/consumer.c +++ b/src/common/consumer.c @@ -18,6 +18,7 @@ */ #define _GNU_SOURCE +#define _LGPL_SOURCE #include #include #include @@ -30,17 +31,23 @@ #include #include +#include #include #include #include +#include +#include #include #include #include #include #include #include +#include #include "consumer.h" +#include "consumer-stream.h" +#include "consumer-testpoint.h" struct lttng_consumer_global_data consumer_data = { .stream_count = 0, @@ -50,12 +57,14 @@ struct lttng_consumer_global_data consumer_data = { 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; + struct lttng_consumer_channel *chan; /* add */ + uint64_t key; /* del */ }; /* @@ -75,49 +84,97 @@ 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) { - int ret; + struct lttng_consumer_stream *null_stream = NULL; + + assert(pipe); + + (void) lttng_pipe_write(pipe, &null_stream, sizeof(null_stream)); +} - do { - struct lttng_consumer_stream *null_stream = NULL; +static void notify_health_quit_pipe(int *pipe) +{ + ssize_t ret; - ret = write(wpipe, &null_stream, sizeof(null_stream)); - } while (ret < 0 && errno == EINTR); + ret = lttng_write(pipe[1], "4", 1); + if (ret < 1) { + PERROR("write consumer health quit"); + } } 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; + ssize_t ret; + + memset(&msg, 0, sizeof(msg)); msg.action = action; msg.chan = chan; - do { - ret = write(ctx->consumer_channel_pipe[1], &msg, sizeof(msg)); - } while (ret < 0 && errno == EINTR); + msg.key = key; + ret = lttng_write(ctx->consumer_channel_pipe[1], &msg, sizeof(msg)); + if (ret < sizeof(msg)) { + PERROR("notify_channel_pipe write error"); + } +} + +void notify_thread_del_channel(struct lttng_consumer_local_data *ctx, + uint64_t key) +{ + notify_channel_pipe(ctx, NULL, key, CONSUMER_CHANNEL_DEL); } 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; + ssize_t 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; + ret = lttng_read(ctx->consumer_channel_pipe[0], &msg, sizeof(msg)); + if (ret < sizeof(msg)) { + ret = -1; + goto error; + } + *action = msg.action; + *chan = msg.chan; + *key = msg.key; +error: + return (int) ret; +} + +/* + * Cleanup the stream list of a channel. Those streams are not yet globally + * visible + */ +static void clean_channel_stream_list(struct lttng_consumer_channel *channel) +{ + struct lttng_consumer_stream *stream, *stmp; + + assert(channel); + + /* 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. Setting the monitor + * mode to 0 so we don't lock nor try to delete the stream from the + * global hash table. + */ + stream->monitor = 0; + consumer_stream_destroy(stream, NULL); } - return ret; } /* @@ -151,20 +208,20 @@ static struct lttng_consumer_stream *find_stream(uint64_t key, return stream; } -static void 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 = find_stream(key, ht); if (stream) { - stream->key = -1ULL; + 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 = -1ULL; + stream->node.key = (uint64_t) -1ULL; } rcu_read_unlock(); } @@ -195,14 +252,30 @@ struct lttng_consumer_channel *consumer_find_channel(uint64_t key) return channel; } -static void free_stream_rcu(struct rcu_head *head) +/* + * There is a possibility that the consumer does not have enough time between + * the close of the channel on the session daemon and the cleanup in here thus + * once we have a channel add with an existing key, we know for sure that this + * channel will eventually get cleaned up by all streams being closed. + * + * This function just nullifies the already existing channel key. + */ +static void steal_channel_key(uint64_t key) { - 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); + struct lttng_consumer_channel *channel; - free(stream); + rcu_read_lock(); + channel = consumer_find_channel(key); + if (channel) { + channel->key = (uint64_t) -1ULL; + /* + * 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 = (uint64_t) -1ULL; + } + rcu_read_unlock(); } static void free_channel_rcu(struct rcu_head *head) @@ -241,10 +314,8 @@ static void free_relayd_rcu(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; @@ -278,6 +349,14 @@ void consumer_del_channel(struct lttng_consumer_channel *channel) DBG("Consumer delete channel key %" PRIu64, channel->key); pthread_mutex_lock(&consumer_data.lock); + pthread_mutex_lock(&channel->lock); + + /* Destroy streams that might have been left in the stream list. */ + clean_channel_stream_list(channel); + + if (channel->live_timer_enabled == 1) { + consumer_timer_live_stop(channel); + } switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: @@ -300,6 +379,7 @@ void consumer_del_channel(struct lttng_consumer_channel *channel) call_rcu(&channel->node.head, free_channel_rcu); end: + pthread_mutex_unlock(&channel->lock); pthread_mutex_unlock(&consumer_data.lock); } @@ -316,12 +396,12 @@ static void cleanup_relayd_ht(void) cds_lfht_for_each_entry(consumer_data.relayd_ht->ht, &iter.iter, relayd, node.node) { - destroy_relayd(relayd); + consumer_destroy_relayd(relayd); } - lttng_ht_destroy(consumer_data.relayd_ht); - rcu_read_unlock(); + + lttng_ht_destroy(consumer_data.relayd_ht); } /* @@ -331,13 +411,13 @@ static void cleanup_relayd_ht(void) * 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(); @@ -370,7 +450,7 @@ 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); @@ -383,7 +463,7 @@ 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); @@ -395,8 +475,8 @@ 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); } } @@ -415,124 +495,33 @@ 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_rcu; - } - - pthread_mutex_lock(&consumer_data.lock); - pthread_mutex_lock(&stream->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); - - 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(); - - assert(consumer_data.stream_count > 0); - consumer_data.stream_count--; - - if (stream->out_fd >= 0) { - ret = close(stream->out_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_stream_left)) { - free_chan = stream->chan; - } - -end: - consumer_data.need_update = 1; - pthread_mutex_unlock(&stream->lock); - pthread_mutex_unlock(&consumer_data.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_rcu: - call_rcu(&stream->node.head, free_stream_rcu); +void consumer_del_stream_for_metadata(struct lttng_consumer_stream *stream) +{ + consumer_stream_destroy(stream, metadata_ht); } struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key, @@ -541,11 +530,12 @@ struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key, const char *channel_name, uid_t uid, gid_t gid, - int relayd_id, + uint64_t relayd_id, uint64_t session_id, int cpu, int *alloc_ret, - enum consumer_channel_type type) + enum consumer_channel_type type, + unsigned int monitor) { int ret; struct lttng_consumer_stream *stream; @@ -562,11 +552,15 @@ struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key, stream->key = stream_key; stream->out_fd = -1; stream->out_fd_offset = 0; + stream->output_written = 0; stream->state = state; stream->uid = uid; stream->gid = gid; stream->net_seq_idx = relayd_id; stream->session_id = session_id; + stream->monitor = monitor; + stream->endpoint_status = CONSUMER_ENDPOINT_ACTIVE; + stream->index_fd = -1; pthread_mutex_init(&stream->lock, NULL); /* If channel is the metadata, flag this stream as metadata. */ @@ -574,6 +568,9 @@ struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key, stream->metadata_flag = 1; /* Metadata is flat out. */ strncpy(stream->name, DEFAULT_METADATA_NAME, sizeof(stream->name)); + /* Live rendez-vous point. */ + pthread_cond_init(&stream->metadata_rdv, NULL); + pthread_mutex_init(&stream->metadata_rdv_lock, NULL); } else { /* Format stream name to _ */ ret = snprintf(stream->name, sizeof(stream->name), "%s_%d", @@ -593,8 +590,10 @@ struct lttng_consumer_stream *consumer_allocate_stream(uint64_t 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 %" PRIu64 ", chan_key %" PRIu64 " relayd_id %" PRIu64 ", session_id %" PRIu64, - stream->name, stream->key, channel_key, stream->net_seq_idx, stream->session_id); + 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; @@ -612,11 +611,10 @@ end: /* * Add a stream to the global list protected by a mutex. */ -static int 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); @@ -624,6 +622,8 @@ static int add_stream(struct lttng_consumer_stream *stream, 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(); @@ -642,15 +642,6 @@ static int add_stream(struct lttng_consumer_stream *stream, */ lttng_ht_add_u64(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); - /* * 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 @@ -659,6 +650,8 @@ static int add_stream(struct lttng_consumer_stream *stream, * stream. */ 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); } @@ -668,11 +661,18 @@ static int add_stream(struct lttng_consumer_stream *stream, 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. @@ -701,12 +701,12 @@ 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; } @@ -719,6 +719,8 @@ struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair( obj->net_seq_idx = net_seq_idx; obj->refcount = 0; obj->destroy_flag = 0; + 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); @@ -755,6 +757,106 @@ error: return relayd; } +/* + * Find a relayd and send the stream + * + * Returns 0 on success, < 0 on error + */ +int consumer_send_relayd_stream(struct lttng_consumer_stream *stream, + char *path) +{ + 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 send the streams sent message + * + * Returns 0 on success, < 0 on error + */ +int consumer_send_relayd_streams_sent(uint64_t net_seq_idx) +{ + int ret = 0; + struct consumer_relayd_sock_pair *relayd; + + assert(net_seq_idx != -1ULL); + + /* The stream is not metadata. Get relayd reference if exists. */ + rcu_read_lock(); + relayd = consumer_find_relayd(net_seq_idx); + if (relayd != NULL) { + /* Add stream on the relayd */ + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + ret = relayd_streams_sent(&relayd->control_sock); + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); + if (ret < 0) { + goto end; + } + } else { + ERR("Relayd ID %" PRIu64 " unknown. Can't send streams_sent.", + net_seq_idx); + ret = -1; + goto end; + } + + ret = 0; + DBG("All streams sent relayd id %" PRIu64, 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. @@ -827,10 +929,13 @@ struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key, const char *name, uid_t uid, gid_t gid, - int relayd_id, + uint64_t relayd_id, enum lttng_event_output output, uint64_t tracefile_size, - uint64_t tracefile_count) + uint64_t tracefile_count, + uint64_t session_id_per_pid, + unsigned int monitor, + unsigned int live_timer_interval) { struct lttng_consumer_channel *channel; @@ -843,12 +948,43 @@ struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key, channel->key = key; channel->refcount = 0; 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; + channel->live_timer_interval = live_timer_interval; + pthread_mutex_init(&channel->lock, NULL); + pthread_mutex_init(&channel->timer_lock, NULL); + + switch (output) { + case LTTNG_EVENT_SPLICE: + channel->output = CONSUMER_CHANNEL_SPLICE; + break; + case LTTNG_EVENT_MMAP: + channel->output = CONSUMER_CHANNEL_MMAP; + break; + default: + assert(0); + free(channel); + channel = NULL; + 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; + } strncpy(channel->pathname, pathname, sizeof(channel->pathname)); channel->pathname[sizeof(channel->pathname) - 1] = '\0'; @@ -870,38 +1006,36 @@ end: /* * Add a channel to the global list protected by a mutex. + * + * Always return 0 indicating success. */ int consumer_add_channel(struct lttng_consumer_channel *channel, struct lttng_consumer_local_data *ctx) { - int ret = 0; - struct lttng_ht_node_u64 *node; - struct lttng_ht_iter iter; - pthread_mutex_lock(&consumer_data.lock); - rcu_read_lock(); + pthread_mutex_lock(&channel->lock); + pthread_mutex_lock(&channel->timer_lock); - 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 = -1; - goto end; - } + /* + * This gives us a guarantee that the channel we are about to add to the + * channel hash table will be unique. See this function comment on the why + * we need to steel the channel key at this stage. + */ + steal_channel_key(channel->key); + rcu_read_lock(); 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); - if (!ret && channel->wait_fd != -1 && - channel->metadata_stream == NULL) { - notify_channel_pipe(ctx, channel, CONSUMER_CHANNEL_ADD); + if (channel->wait_fd != -1 && channel->type == CONSUMER_CHANNEL_TYPE_DATA) { + notify_channel_pipe(ctx, channel, -1, CONSUMER_CHANNEL_ADD); } - return ret; + + return 0; } /* @@ -957,14 +1091,17 @@ static int update_poll_array(struct lttng_consumer_local_data *ctx, * 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; + + (*pollfd)[i + 1].fd = lttng_pipe_get_readfd(ctx->consumer_wakeup_pipe); + (*pollfd)[i + 1].events = POLLIN | POLLPRI; return i; } /* - * Poll on the should_quit pipe and the command socket return -1 on error and - * should exit, 0 if data is available on the command socket + * Poll on the should_quit pipe and the command socket return -1 on + * error, 1 if should exit, 0 if data is available on the command socket */ int lttng_consumer_poll_socket(struct pollfd *consumer_sockpoll) { @@ -980,16 +1117,13 @@ restart: goto restart; } PERROR("Poll error"); - goto exit; + return -1; } if (consumer_sockpoll[0].revents & (POLLIN | POLLPRI)) { DBG("consumer_should_quit wake up"); - goto exit; + return 1; } return 0; - -exit: - return -1; } /* @@ -1061,12 +1195,11 @@ void lttng_consumer_cleanup(void) */ void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx) { - int ret; + ssize_t ret; + consumer_quit = 1; - do { - ret = write(ctx->consumer_should_quit[1], "4", 1); - } while (ret < 0 && errno == EINTR); - if (ret < 0 || ret != 1) { + ret = lttng_write(ctx->consumer_should_quit[1], "4", 1); + if (ret < 1) { PERROR("write consumer quit"); } @@ -1130,7 +1263,7 @@ 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; struct lttng_consumer_local_data *ctx; @@ -1147,30 +1280,21 @@ 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; + ctx->consumer_wakeup_pipe = lttng_pipe_open(0); + if (!ctx->consumer_wakeup_pipe) { + goto error_wakeup_pipe; } ret = pipe(ctx->consumer_should_quit); @@ -1179,47 +1303,84 @@ struct lttng_consumer_local_data *lttng_consumer_create( goto error_quit_pipe; } - ret = pipe(ctx->consumer_thread_pipe); - if (ret < 0) { - PERROR("Error creating thread pipe"); - goto error_thread_pipe; - } - ret = pipe(ctx->consumer_channel_pipe); if (ret < 0) { PERROR("Error creating channel pipe"); goto error_channel_pipe; } - ret = utils_create_pipe(ctx->consumer_metadata_pipe); - if (ret < 0) { + ctx->consumer_metadata_pipe = lttng_pipe_open(0); + if (!ctx->consumer_metadata_pipe) { goto error_metadata_pipe; } - ret = utils_create_pipe(ctx->consumer_splice_metadata_pipe); - if (ret < 0) { - goto error_splice_pipe; - } - return ctx; -error_splice_pipe: - utils_close_pipe(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: utils_close_pipe(ctx->consumer_should_quit); -error_poll_fcntl: error_quit_pipe: - utils_close_pipe(ctx->consumer_data_pipe); + lttng_pipe_destroy(ctx->consumer_wakeup_pipe); +error_wakeup_pipe: + lttng_pipe_destroy(ctx->consumer_data_pipe); error_poll_pipe: free(ctx); error: return NULL; } +/* + * Iterate over all streams of the hashtable and free them properly. + */ +static void destroy_data_stream_ht(struct lttng_ht *ht) +{ + struct lttng_ht_iter iter; + struct lttng_consumer_stream *stream; + + if (ht == NULL) { + return; + } + + rcu_read_lock(); + cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { + /* + * Ignore return value since we are currently cleaning up so any error + * can't be handled. + */ + (void) consumer_del_stream(stream, ht); + } + rcu_read_unlock(); + + lttng_ht_destroy(ht); +} + +/* + * Iterate over all streams of the metadata hashtable and free them + * properly. + */ +static void destroy_metadata_stream_ht(struct lttng_ht *ht) +{ + struct lttng_ht_iter iter; + struct lttng_consumer_stream *stream; + + if (ht == NULL) { + return; + } + + rcu_read_lock(); + cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { + /* + * Ignore return value since we are currently cleaning up so any error + * can't be handled. + */ + (void) consumer_del_metadata_stream(stream, ht); + } + rcu_read_unlock(); + + lttng_ht_destroy(ht); +} + /* * Close all fds associated with the instance and free the context. */ @@ -1229,6 +1390,13 @@ void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx) DBG("Consumer destroying it. Closing everything."); + if (!ctx) { + return; + } + + destroy_data_stream_ht(data_ht); + destroy_metadata_stream_ht(metadata_ht); + ret = close(ctx->consumer_error_socket); if (ret) { PERROR("close"); @@ -1237,11 +1405,11 @@ void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx) if (ret) { PERROR("close"); } - utils_close_pipe(ctx->consumer_thread_pipe); utils_close_pipe(ctx->consumer_channel_pipe); - utils_close_pipe(ctx->consumer_data_pipe); + lttng_pipe_destroy(ctx->consumer_data_pipe); + lttng_pipe_destroy(ctx->consumer_metadata_pipe); + lttng_pipe_destroy(ctx->consumer_wakeup_pipe); utils_close_pipe(ctx->consumer_should_quit); - utils_close_pipe(ctx->consumer_splice_metadata_pipe); unlink(ctx->consumer_command_sock_path); free(ctx); @@ -1254,17 +1422,15 @@ static int write_relayd_metadata_id(int fd, struct lttng_consumer_stream *stream, struct consumer_relayd_sock_pair *relayd, unsigned long padding) { - int ret; + ssize_t ret; struct lttcomm_relayd_metadata_payload hdr; hdr.stream_id = htobe64(stream->relayd_stream_id); hdr.padding_size = htobe32(padding); - do { - ret = write(fd, (void *) &hdr, sizeof(hdr)); - } while (ret < 0 && errno == EINTR); - if (ret < 0 || ret != sizeof(hdr)) { + ret = lttng_write(fd, (void *) &hdr, sizeof(hdr)); + if (ret < sizeof(hdr)) { /* - * This error means that the fd's end is closed so ignore the perror + * 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. */ @@ -1284,100 +1450,7 @@ static int write_relayd_metadata_id(int fd, stream->relayd_stream_id, padding); end: - return ret; -} - -/* - * Create the tracefile on disk. - * - * Return 0 on success or else a negative value. - */ -int lttng_create_output_file(struct lttng_consumer_stream *stream) -{ - int ret; - char full_path[PATH_MAX]; - char *path_name_id = NULL; - char *path; - - assert(stream); - - /* Don't create anything if this is set for streaming. */ - if (stream->net_seq_idx != (uint64_t) -1ULL) { - ret = 0; - goto end; - } - - ret = snprintf(full_path, sizeof(full_path), "%s/%s", - stream->chan->pathname, stream->name); - if (ret < 0) { - PERROR("snprintf create output file"); - goto error; - } - - /* - * If we split the trace in multiple files, we have to add the tracefile - * current count at the end of the tracefile name - */ - if (stream->chan->tracefile_size > 0) { - ret = asprintf(&path_name_id, "%s_%" PRIu64, full_path, - stream->tracefile_count_current); - if (ret < 0) { - PERROR("Allocating path name ID"); - goto error; - } - path = path_name_id; - } else { - path = full_path; - } - - ret = run_as_open(path, O_WRONLY | O_CREAT | O_TRUNC, - S_IRWXU | S_IRWXG | S_IRWXO, stream->uid, stream->gid); - if (ret < 0) { - PERROR("open stream path %s", path); - goto error_open; - } - stream->out_fd = ret; - stream->tracefile_size_current = 0; - -error_open: - free(path_name_id); -error: -end: - return ret; -} - -/* - * Change the output tracefile according to the tracefile_size and - * tracefile_count parameters. The stream lock MUST be held before calling this - * function because we are modifying the stream status. - * - * Return 0 on success or else a negative value. - */ -static int rotate_output_file(struct lttng_consumer_stream *stream) -{ - int ret; - - assert(stream); - assert(stream->tracefile_size_current); - - ret = close(stream->out_fd); - if (ret < 0) { - PERROR("Closing tracefile"); - goto end; - } - - if (stream->chan->tracefile_count > 0) { - stream->tracefile_count_current = - (stream->tracefile_count_current + 1) % - stream->chan->tracefile_count; - } else { - stream->tracefile_count_current++; - } - - return lttng_create_output_file(stream); - -end: - return ret; + return (int) ret; } /* @@ -1394,11 +1467,12 @@ end: ssize_t lttng_consumer_on_read_subbuffer_mmap( struct lttng_consumer_local_data *ctx, struct lttng_consumer_stream *stream, unsigned long len, - unsigned long padding) + unsigned long padding, + struct ctf_packet_index *index) { unsigned long mmap_offset; void *mmap_base; - ssize_t ret = 0, written = 0; + ssize_t ret = 0; off_t orig_offset = stream->out_fd_offset; /* Default is on the disk */ int outfd = stream->out_fd; @@ -1409,9 +1483,10 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( rcu_read_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) { + ret = -EPIPE; goto end; } } @@ -1421,28 +1496,31 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( case LTTNG_CONSUMER_KERNEL: mmap_base = stream->mmap_base; ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset); + if (ret < 0) { + ret = -errno; + PERROR("tracer ctl get_mmap_read_offset"); + goto end; + } break; case LTTNG_CONSUMER32_UST: case LTTNG_CONSUMER64_UST: mmap_base = lttng_ustctl_get_mmap_base(stream); if (!mmap_base) { ERR("read mmap get mmap base for stream %s", stream->name); - written = -1; + ret = -EPERM; goto end; } ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset); - + if (ret != 0) { + PERROR("tracer ctl get_mmap_read_offset"); + ret = -EINVAL; + goto end; + } break; default: ERR("Unknown consumer_data type"); assert(0); } - if (ret != 0) { - errno = -ret; - PERROR("tracer ctl get_mmap_read_offset"); - written = ret; - goto end; - } /* Handle stream on the relayd if the output is on the network */ if (relayd) { @@ -1459,30 +1537,20 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( } ret = write_relayd_stream_header(stream, netlen, padding, relayd); - if (ret >= 0) { - /* Use the returned socket. */ - outfd = ret; + if (ret < 0) { + relayd_hang_up = 1; + goto write_error; + } + /* Use the returned socket. */ + outfd = ret; - /* Write metadata stream id before payload */ - if (stream->metadata_flag) { - ret = write_relayd_metadata_id(outfd, stream, relayd, padding); - if (ret < 0) { - written = ret; - /* Socket operation failed. We consider the relayd dead */ - if (ret == -EPIPE || ret == -EINVAL) { - relayd_hang_up = 1; - goto write_error; - } - goto end; - } - } - } else { - /* Socket operation failed. We consider the relayd dead */ - if (ret == -EPIPE || ret == -EINVAL) { + /* Write metadata stream id before payload */ + if (stream->metadata_flag) { + ret = write_relayd_metadata_id(outfd, stream, relayd, padding); + if (ret < 0) { relayd_hang_up = 1; goto write_error; } - /* Else, use the default set before which is the filesystem. */ } } else { /* No streaming, we have to set the len with the full padding */ @@ -1494,55 +1562,78 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( if (stream->chan->tracefile_size > 0 && (stream->tracefile_size_current + len) > stream->chan->tracefile_size) { - ret = rotate_output_file(stream); + 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), + &stream->out_fd); if (ret < 0) { ERR("Rotating output file"); goto end; } outfd = stream->out_fd; + + if (stream->index_fd >= 0) { + ret = index_create_file(stream->chan->pathname, + stream->name, stream->uid, stream->gid, + stream->chan->tracefile_size, + stream->tracefile_count_current); + if (ret < 0) { + goto end; + } + stream->index_fd = ret; + } + + /* Reset current size because we just perform a rotation. */ + stream->tracefile_size_current = 0; + stream->out_fd_offset = 0; + orig_offset = 0; } stream->tracefile_size_current += len; + if (index) { + index->offset = htobe64(stream->out_fd_offset); + } } - while (len > 0) { - do { - ret = write(outfd, mmap_base + mmap_offset, len); - } while (ret < 0 && errno == EINTR); - DBG("Consumer mmap write() ret %zd (len %lu)", ret, len); + /* + * This call guarantee that len or less is returned. It's impossible to + * receive a ret value that is bigger than len. + */ + ret = lttng_write(outfd, mmap_base + mmap_offset, len); + DBG("Consumer mmap write() ret %zd (len %lu)", ret, len); + if (ret < 0 || ((size_t) ret != len)) { + /* + * Report error to caller if nothing was written else at least send the + * amount written. + */ if (ret < 0) { + ret = -errno; + } + relayd_hang_up = 1; + + /* Socket operation failed. We consider the relayd dead */ + if (errno == EPIPE || errno == EINVAL || errno == EBADF) { /* - * 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. + * This is possible if the fd is closed on the other side + * (outfd) or any write problem. It can be verbose a bit for a + * normal execution if for instance the relayd is stopped + * abruptly. This can happen so set this to a DBG statement. */ - DBG("Error in file write mmap"); - if (written == 0) { - written = ret; - } - /* Socket operation failed. We consider the relayd dead */ - if (errno == EPIPE || errno == EINVAL) { - relayd_hang_up = 1; - goto write_error; - } - goto end; - } else if (ret > len) { - PERROR("Error in file write (ret %zd > len %lu)", ret, len); - written += ret; - goto end; + DBG("Consumer mmap write detected relayd hang up"); } else { - len -= ret; - mmap_offset += ret; + /* Unhandled error, print it and stop function right now. */ + PERROR("Error in write mmap (ret %zd != len %lu)", ret, len); } + goto write_error; + } + stream->output_written += ret; - /* This call is useless on a socket so better save a syscall. */ - if (!relayd) { - /* This won't block, but will start writeout asynchronously */ - lttng_sync_file_range(outfd, stream->out_fd_offset, ret, - SYNC_FILE_RANGE_WRITE); - stream->out_fd_offset += ret; - } - written += ret; + /* This call is useless on a socket so better save a syscall. */ + if (!relayd) { + /* This won't block, but will start writeout asynchronously */ + lttng_sync_file_range(outfd, stream->out_fd_offset, len, + SYNC_FILE_RANGE_WRITE); + stream->out_fd_offset += len; } lttng_consumer_sync_trace_file(stream, orig_offset); @@ -1562,7 +1653,7 @@ end: } rcu_read_unlock(); - return written; + return ret; } /* @@ -1575,7 +1666,8 @@ end: ssize_t lttng_consumer_on_read_subbuffer_splice( struct lttng_consumer_local_data *ctx, struct lttng_consumer_stream *stream, unsigned long len, - unsigned long padding) + unsigned long padding, + struct ctf_packet_index *index) { ssize_t ret = 0, written = 0, ret_splice = 0; loff_t offset = 0; @@ -1603,27 +1695,18 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( rcu_read_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) { + written = -ret; goto end; } } - - /* - * Choose right pipe for splice. Metadata and trace data are handled by - * different threads hence the use of two pipes in order not to race or - * corrupt the written data. - */ - if (stream->metadata_flag) { - splice_pipe = ctx->consumer_splice_metadata_pipe; - } else { - splice_pipe = ctx->consumer_thread_pipe; - } + splice_pipe = stream->splice_pipe; /* Write metadata stream id before payload */ if (relayd) { - int total_len = len; + unsigned long total_len = len; if (stream->metadata_flag) { /* @@ -1636,31 +1719,21 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( padding); if (ret < 0) { written = ret; - /* Socket operation failed. We consider the relayd dead */ - if (ret == -EBADF) { - WARN("Remote relayd disconnected. Stopping"); - relayd_hang_up = 1; - goto write_error; - } - goto end; + relayd_hang_up = 1; + goto write_error; } total_len += sizeof(struct lttcomm_relayd_metadata_payload); } ret = write_relayd_stream_header(stream, total_len, padding, relayd); - if (ret >= 0) { - /* Use the returned socket. */ - outfd = ret; - } else { - /* Socket operation failed. We consider the relayd dead */ - if (ret == -EBADF) { - WARN("Remote relayd disconnected. Stopping"); - relayd_hang_up = 1; - goto write_error; - } - goto end; + if (ret < 0) { + written = ret; + relayd_hang_up = 1; + goto write_error; } + /* Use the returned socket. */ + outfd = ret; } else { /* No streaming, we have to set the len with the full padding */ len += padding; @@ -1671,14 +1744,37 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( if (stream->chan->tracefile_size > 0 && (stream->tracefile_size_current + len) > stream->chan->tracefile_size) { - ret = rotate_output_file(stream); + 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), + &stream->out_fd); if (ret < 0) { + written = ret; ERR("Rotating output file"); goto end; } outfd = stream->out_fd; + + if (stream->index_fd >= 0) { + ret = index_create_file(stream->chan->pathname, + stream->name, stream->uid, stream->gid, + stream->chan->tracefile_size, + stream->tracefile_count_current); + if (ret < 0) { + written = ret; + goto end; + } + stream->index_fd = ret; + } + + /* Reset current size because we just perform a rotation. */ + stream->tracefile_size_current = 0; + stream->out_fd_offset = 0; + orig_offset = 0; } stream->tracefile_size_current += len; + index->offset = htobe64(stream->out_fd_offset); } while (len > 0) { @@ -1688,57 +1784,51 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( SPLICE_F_MOVE | SPLICE_F_MORE); DBG("splice chan to pipe, ret %zd", ret_splice); if (ret_splice < 0) { - PERROR("Error in relay splice"); - if (written == 0) { - written = ret_splice; - } ret = errno; + written = -ret; + PERROR("Error in relay splice"); goto splice_error; } /* Handle stream on the relayd if the output is on the network */ - if (relayd) { - if (stream->metadata_flag) { - size_t metadata_payload_size = - sizeof(struct lttcomm_relayd_metadata_payload); + if (relayd && stream->metadata_flag) { + size_t metadata_payload_size = + sizeof(struct lttcomm_relayd_metadata_payload); - /* Update counter to fit the spliced data */ - ret_splice += metadata_payload_size; - len += metadata_payload_size; - /* - * We do this so the return value can match the len passed as - * argument to this function. - */ - written -= metadata_payload_size; - } + /* Update counter to fit the spliced data */ + ret_splice += metadata_payload_size; + len += metadata_payload_size; + /* + * We do this so the return value can match the len passed as + * argument to this function. + */ + written -= metadata_payload_size; } /* Splice data out */ ret_splice = splice(splice_pipe[0], NULL, outfd, NULL, ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE); - DBG("Consumer splice pipe to file, ret %zd", ret_splice); + DBG("Consumer splice pipe to file (out_fd: %d), ret %zd", + outfd, ret_splice); if (ret_splice < 0) { - PERROR("Error in file splice"); - if (written == 0) { - written = ret_splice; - } - /* Socket operation failed. We consider the relayd dead */ - if (errno == EBADF || errno == EPIPE) { - WARN("Remote relayd disconnected. Stopping"); - relayd_hang_up = 1; - goto write_error; - } ret = errno; - goto splice_error; + written = -ret; + relayd_hang_up = 1; + goto write_error; } else if (ret_splice > len) { - errno = EINVAL; - PERROR("Wrote more data than requested %zd (len: %lu)", - ret_splice, len); - written += ret_splice; + /* + * We don't expect this code path to be executed but you never know + * so this is an extra protection agains a buggy splice(). + */ ret = errno; + written += ret_splice; + PERROR("Wrote more data than requested %zd (len: %lu)", ret_splice, + len); goto splice_error; + } else { + /* All good, update current len and continue. */ + len -= ret_splice; } - len -= ret_splice; /* This call is useless on a socket so better save a syscall. */ if (!relayd) { @@ -1747,12 +1837,10 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( SYNC_FILE_RANGE_WRITE); stream->out_fd_offset += ret_splice; } + stream->output_written += ret_splice; written += ret_splice; } lttng_consumer_sync_trace_file(stream, orig_offset); - - ret = ret_splice; - goto end; write_error: @@ -1846,61 +1934,7 @@ int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx, } } -/* - * Iterate over all streams of the hashtable and free them properly. - * - * WARNING: *MUST* be used with data stream only. - */ -static void destroy_data_stream_ht(struct lttng_ht *ht) -{ - struct lttng_ht_iter iter; - struct lttng_consumer_stream *stream; - - if (ht == NULL) { - return; - } - - rcu_read_lock(); - cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { - /* - * Ignore return value since we are currently cleaning up so any error - * can't be handled. - */ - (void) consumer_del_stream(stream, ht); - } - rcu_read_unlock(); - - lttng_ht_destroy(ht); -} - -/* - * Iterate over all streams of the hashtable and free them properly. - * - * XXX: Should not be only for metadata stream or else use an other name. - */ -static void destroy_stream_ht(struct lttng_ht *ht) -{ - struct lttng_ht_iter iter; - struct lttng_consumer_stream *stream; - - if (ht == NULL) { - return; - } - - rcu_read_lock(); - cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { - /* - * Ignore return value since we are currently cleaning up so any error - * can't be handled. - */ - (void) consumer_del_metadata_stream(stream, ht); - } - rcu_read_unlock(); - - lttng_ht_destroy(ht); -} - -void lttng_consumer_close_metadata(void) +void lttng_consumer_close_all_metadata(void) { switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: @@ -1918,7 +1952,7 @@ void lttng_consumer_close_metadata(void) * because at this point we are sure that the metadata producer is * either dead or blocked. */ - lttng_ustconsumer_close_metadata(metadata_ht); + lttng_ustconsumer_close_all_metadata(metadata_ht); break; default: ERR("Unknown consumer_data type"); @@ -1932,126 +1966,62 @@ void lttng_consumer_close_metadata(void) void consumer_del_metadata_stream(struct lttng_consumer_stream *stream, struct lttng_ht *ht) { - int ret; - struct lttng_ht_iter iter; struct lttng_consumer_channel *free_chan = NULL; - struct consumer_relayd_sock_pair *relayd; - - assert(stream); - /* - * This call should NEVER receive regular stream. It must always be - * metadata stream and this is crucial for data structure synchronization. - */ - assert(stream->metadata_flag); - - DBG3("Consumer delete metadata stream %d", stream->wait_fd); - - if (ht == NULL) { - /* Means the stream was allocated but not successfully added */ - goto free_stream_rcu; - } - - pthread_mutex_lock(&consumer_data.lock); - pthread_mutex_lock(&stream->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 metadata stream"); - } - } - 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); - - 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(); - if (stream->out_fd >= 0) { - ret = close(stream->out_fd); - if (ret) { - PERROR("close"); - } - } + assert(stream); + /* + * This call should NEVER receive regular stream. It must always be + * metadata stream and this is crucial for data structure synchronization. + */ + assert(stream->metadata_flag); - /* 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); + DBG3("Consumer delete metadata stream %d", stream->wait_fd); - /* 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. - */ - } + pthread_mutex_lock(&consumer_data.lock); + pthread_mutex_lock(&stream->chan->lock); + pthread_mutex_lock(&stream->lock); - /* 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(); + /* Remove any reference to that stream. */ + consumer_stream_delete(stream, ht); + + /* Close down everything including the relayd if one. */ + consumer_stream_close(stream); + /* Destroy tracer buffers of the stream. */ + consumer_stream_destroy_buffers(stream); /* Atomically decrement channel refcount since other threads can use it. */ - uatomic_dec(&stream->chan->refcount); - if (!uatomic_read(&stream->chan->refcount) + 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: + /* + * 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_rcu: - call_rcu(&stream->node.head, free_stream_rcu); + consumer_stream_free(stream); } /* * Action done with the metadata stream when adding it to the consumer internal * data structures to handle it. */ -static int 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; @@ -2061,6 +2031,8 @@ static int add_metadata_stream(struct lttng_consumer_stream *stream, 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); /* @@ -2078,15 +2050,6 @@ static int add_metadata_stream(struct lttng_consumer_stream *stream, node = lttng_ht_iter_get_node_u64(&iter); assert(!node); - /* Find relayd and, if one is found, increment refcount. */ - relayd = consumer_find_relayd(stream->net_seq_idx); - if (relayd != NULL) { - uatomic_inc(&relayd->refcount); - } - - /* Update channel refcount once added without error(s). */ - uatomic_inc(&stream->chan->refcount); - /* * When nb_init_stream_left reaches 0, we don't need to trigger any action * in terms of destroying the associated channel, because the action that @@ -2095,6 +2058,8 @@ static int add_metadata_stream(struct lttng_consumer_stream *stream, * stream. */ 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); } @@ -2113,6 +2078,8 @@ static int add_metadata_stream(struct lttng_consumer_stream *stream, 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; } @@ -2176,7 +2143,7 @@ static void validate_endpoint_status_metadata_stream( */ void *consumer_thread_metadata_poll(void *data) { - int ret, i, pollfd; + int ret, i, pollfd, err = -1; uint32_t revents, nb_fd; struct lttng_consumer_stream *stream = NULL; struct lttng_ht_iter iter; @@ -2187,12 +2154,14 @@ void *consumer_thread_metadata_poll(void *data) 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; + health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_METADATA); + + if (testpoint(consumerd_thread_metadata)) { + goto error_testpoint; } + health_code_update(); + DBG("Thread metadata poll started"); /* Size is set to 1 for the consumer_metadata pipe */ @@ -2202,7 +2171,8 @@ void *consumer_thread_metadata_poll(void *data) 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; } @@ -2211,59 +2181,60 @@ void *consumer_thread_metadata_poll(void *data) DBG("Metadata main loop started"); while (1) { - /* Only the metadata pipe is set */ - if (LTTNG_POLL_GETNB(&events) == 0 && consumer_quit == 1) { - goto end; - } - restart: - DBG("Metadata poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events)); + health_code_update(); + health_poll_entry(); + DBG("Metadata poll wait"); ret = lttng_poll_wait(&events, -1); + DBG("Metadata poll return from wait with %d fd(s)", + LTTNG_POLL_GETNB(&events)); + health_poll_exit(); DBG("Metadata event catched in thread"); if (ret < 0) { if (errno == EINTR) { ERR("Poll EINTR catched"); goto restart; } - goto error; + if (LTTNG_POLL_GETNB(&events) == 0) { + err = 0; /* All is OK */ + } + goto end; } nb_fd = ret; /* From here, the event is a metadata wait fd */ for (i = 0; i < nb_fd; i++) { + health_code_update(); + 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) { + /* No activity for this FD (poll implementation). */ 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]); - ret = close(ctx->consumer_metadata_pipe[0]); - if (ret < 0) { - PERROR("close metadata pipe"); - } + 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 *)) { + ssize_t pipe_len; + + pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe, + &stream, sizeof(stream)); + if (pipe_len < sizeof(stream)) { PERROR("read metadata stream"); /* - * 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; } @@ -2278,17 +2249,9 @@ restart: DBG("Adding metadata stream %d to poll set", stream->wait_fd); - ret = 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); + LPOLLIN | LPOLLPRI | LPOLLHUP); } /* Handle other stream */ @@ -2318,6 +2281,8 @@ restart: /* We just flushed the stream now read it. */ do { + health_code_update(); + len = ctx->on_buffer_ready(stream, ctx); /* * We don't check the return value here since if we get @@ -2339,14 +2304,23 @@ restart: DBG("Metadata available on fd %d", pollfd); assert(stream->wait_fd == pollfd); - len = ctx->on_buffer_ready(stream, ctx); + do { + health_code_update(); + + 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; } } @@ -2355,14 +2329,19 @@ restart: } } -error: + /* All is OK */ + err = 0; end: DBG("Metadata poll thread exiting"); lttng_poll_clean(&events); end_poll: - destroy_stream_ht(metadata_ht); -end_ht: +error_testpoint: + if (err) { + health_error(); + ERR("Health error occurred in %s", __func__); + } + health_unregister(health_consumerd); rcu_unregister_thread(); return NULL; } @@ -2373,7 +2352,7 @@ end_ht: */ void *consumer_thread_data_poll(void *data) { - int num_rdy, num_hup, high_prio, ret, i; + int num_rdy, num_hup, high_prio, ret, i, err = -1; struct pollfd *pollfd = NULL; /* local view of the streams */ struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL; @@ -2384,15 +2363,23 @@ void *consumer_thread_data_poll(void *data) rcu_register_thread(); - data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); - if (data_ht == NULL) { - /* ENOMEM at this point. Better to bail out. */ - goto end; + health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_DATA); + + if (testpoint(consumerd_thread_data)) { + goto error_testpoint; } - local_stream = zmalloc(sizeof(struct lttng_consumer_stream)); + health_code_update(); + + local_stream = zmalloc(sizeof(struct lttng_consumer_stream *)); + if (local_stream == NULL) { + PERROR("local_stream malloc"); + goto end; + } while (1) { + health_code_update(); + high_prio = 0; num_hup = 0; @@ -2408,17 +2395,19 @@ void *consumer_thread_data_poll(void *data) 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)); + /* + * Allocate for all fds +1 for the consumer_data_pipe and +1 for + * wake up pipe. + */ + pollfd = zmalloc((consumer_data.stream_count + 2) * sizeof(struct pollfd)); if (pollfd == NULL) { PERROR("pollfd malloc"); pthread_mutex_unlock(&consumer_data.lock); goto end; } - /* allocate for all fds + 1 for the consumer_data_pipe */ - local_stream = zmalloc((consumer_data.stream_count + 1) * - sizeof(struct lttng_consumer_stream)); + local_stream = zmalloc((consumer_data.stream_count + 2) * + sizeof(struct lttng_consumer_stream *)); if (local_stream == NULL) { PERROR("local_stream malloc"); pthread_mutex_unlock(&consumer_data.lock); @@ -2439,12 +2428,15 @@ void *consumer_thread_data_poll(void *data) /* No FDs and consumer_quit, consumer_cleanup the thread */ if (nb_fd == 0 && consumer_quit == 1) { + err = 0; /* All is OK */ goto end; } /* poll on the array of fds */ restart: - DBG("polling on %d fd", nb_fd + 1); - num_rdy = poll(pollfd, nb_fd + 1, -1); + DBG("polling on %d fd", nb_fd + 2); + health_poll_entry(); + num_rdy = poll(pollfd, nb_fd + 2, -1); + health_poll_exit(); DBG("poll num_rdy : %d", num_rdy); if (num_rdy == -1) { /* @@ -2470,13 +2462,10 @@ void *consumer_thread_data_poll(void *data) 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); - if (pipe_readlen < 0) { - PERROR("read consumer data pipe"); + pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe, + &new_stream, sizeof(new_stream)); + if (pipe_readlen < sizeof(new_stream)) { + PERROR("Consumer data pipe"); /* Continue so we can at least handle the current stream(s). */ continue; } @@ -2491,23 +2480,28 @@ void *consumer_thread_data_poll(void *data) continue; } - ret = add_stream(new_stream, data_ht); - if (ret) { - ERR("Consumer add stream %" PRIu64 " 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; } + /* Handle wakeup pipe. */ + if (pollfd[nb_fd + 1].revents & (POLLIN | POLLPRI)) { + char dummy; + ssize_t pipe_readlen; + + pipe_readlen = lttng_pipe_read(ctx->consumer_wakeup_pipe, &dummy, + sizeof(dummy)); + if (pipe_readlen < 0) { + PERROR("Consumer data wakeup pipe"); + } + /* We've been awakened to handle stream(s). */ + ctx->has_wakeup = 0; + } + /* Take care of high priority channels first. */ for (i = 0; i < nb_fd; i++) { + health_code_update(); + if (local_stream[i] == NULL) { continue; } @@ -2536,11 +2530,14 @@ void *consumer_thread_data_poll(void *data) /* Take care of low priority channels. */ for (i = 0; i < nb_fd; i++) { + health_code_update(); + if (local_stream[i] == NULL) { continue; } if ((pollfd[i].revents & POLLIN) || - local_stream[i]->hangup_flush_done) { + local_stream[i]->hangup_flush_done || + local_stream[i]->has_data) { DBG("Normal read on fd %d", pollfd[i].fd); len = ctx->on_buffer_ready(local_stream[i], ctx); /* it's ok to have an unavailable sub-buffer */ @@ -2556,6 +2553,8 @@ void *consumer_thread_data_poll(void *data) /* Handle hangup and errors */ for (i = 0; i < nb_fd; i++) { + health_code_update(); + if (local_stream[i] == NULL) { continue; } @@ -2601,6 +2600,8 @@ void *consumer_thread_data_poll(void *data) } } } + /* All is OK */ + err = 0; end: DBG("polling thread exiting"); free(pollfd); @@ -2614,12 +2615,14 @@ end: * only tracked fd in the poll set. The thread will take care of closing * the read side. */ - ret = close(ctx->consumer_metadata_pipe[1]); - if (ret < 0) { - PERROR("close data pipe"); - } + (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe); - destroy_data_stream_ht(data_ht); +error_testpoint: + if (err) { + health_error(); + ERR("Health error occurred in %s", __func__); + } + health_unregister(health_consumerd); rcu_unregister_thread(); return NULL; @@ -2644,22 +2647,36 @@ void consumer_close_channel_streams(struct lttng_consumer_channel *channel) 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); + if (stream->metadata_flag) { + /* Safe and protected by the stream lock. */ + lttng_ustconsumer_close_metadata(stream->chan); + } else { + /* + * 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(); } @@ -2693,7 +2710,7 @@ static void destroy_channel_ht(struct lttng_ht *ht) */ void *consumer_thread_channel_poll(void *data) { - int ret, i, pollfd; + int ret, i, pollfd, err = -1; uint32_t revents, nb_fd; struct lttng_consumer_channel *chan = NULL; struct lttng_ht_iter iter; @@ -2704,6 +2721,14 @@ void *consumer_thread_channel_poll(void *data) rcu_register_thread(); + health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_CHANNEL); + + if (testpoint(consumerd_thread_channel)) { + goto error_testpoint; + } + + health_code_update(); + channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); if (!channel_ht) { /* ENOMEM at this point. Better to bail out. */ @@ -2728,20 +2753,23 @@ void *consumer_thread_channel_poll(void *data) 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)); + health_code_update(); + DBG("Channel poll wait"); + health_poll_entry(); ret = lttng_poll_wait(&events, -1); + DBG("Channel poll return from wait with %d fd(s)", + LTTNG_POLL_GETNB(&events)); + health_poll_exit(); DBG("Channel event catched in thread"); if (ret < 0) { if (errno == EINTR) { ERR("Poll EINTR catched"); goto restart; } + if (LTTNG_POLL_GETNB(&events) == 0) { + err = 0; /* All is OK */ + } goto end; } @@ -2749,13 +2777,16 @@ restart: /* From here, the event is a channel wait fd */ for (i = 0; i < nb_fd; i++) { + health_code_update(); + 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) { + /* No activity for this FD (poll implementation). */ continue; } + if (pollfd == ctx->consumer_channel_pipe[0]) { if (revents & (LPOLLERR | LPOLLHUP)) { DBG("Channel thread pipe hung up"); @@ -2767,8 +2798,9 @@ restart: continue; } else if (revents & LPOLLIN) { enum consumer_channel_action action; + uint64_t key; - ret = read_channel_pipe(ctx, &chan, &action); + ret = read_channel_pipe(ctx, &chan, &key, &action); if (ret <= 0) { ERR("Error reading channel pipe"); continue; @@ -2781,12 +2813,61 @@ restart: 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: + { + /* + * This command should never be called if the channel + * has streams monitored by either the data or metadata + * thread. The consumer only notify this thread with a + * channel del. command if it receives a destroy + * channel command from the session daemon that send it + * if a command prior to the GET_CHANNEL failed. + */ + + 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); + + switch (consumer_data.type) { + case LTTNG_CONSUMER_KERNEL: + break; + case LTTNG_CONSUMER32_UST: + case LTTNG_CONSUMER64_UST: + health_code_update(); + /* Destroy streams that might have been left in the stream list. */ + clean_channel_stream_list(chan); + 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 @@ -2823,7 +2904,19 @@ restart: lttng_poll_del(&events, chan->wait_fd); ret = lttng_ht_del(channel_ht, &iter); assert(ret == 0); + + /* + * This will close the wait fd for each stream associated to + * this channel AND monitored by the data/metadata thread thus + * will be clean by the right thread. + */ 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 */ @@ -2831,12 +2924,20 @@ restart: } } + /* All is OK */ + err = 0; end: lttng_poll_clean(&events); end_poll: destroy_channel_ht(channel_ht); end_ht: +error_testpoint: DBG("Channel poll thread exiting"); + if (err) { + health_error(); + ERR("Health error occurred in %s", __func__); + } + health_unregister(health_consumerd); rcu_unregister_thread(); return NULL; } @@ -2849,8 +2950,8 @@ static int set_metadata_socket(struct lttng_consumer_local_data *ctx, assert(ctx); assert(sockpoll); - if (lttng_consumer_poll_socket(sockpoll) < 0) { - ret = -1; + ret = lttng_consumer_poll_socket(sockpoll); + if (ret) { goto error; } DBG("Metadata connection on client_socket"); @@ -2874,7 +2975,7 @@ error: */ void *consumer_thread_sessiond_poll(void *data) { - int sock = -1, client_socket, ret; + int sock = -1, client_socket, ret, err = -1; /* * structure to poll for incoming data on communication socket avoids * making blocking sockets. @@ -2884,6 +2985,14 @@ void *consumer_thread_sessiond_poll(void *data) rcu_register_thread(); + health_register(health_consumerd, HEALTH_CONSUMERD_TYPE_SESSIOND); + + if (testpoint(consumerd_thread_sessiond)) { + goto error_testpoint; + } + + health_code_update(); + DBG("Creating command socket %s", ctx->consumer_command_sock_path); unlink(ctx->consumer_command_sock_path); client_socket = lttcomm_create_unix_sock(ctx->consumer_command_sock_path); @@ -2905,19 +3014,18 @@ 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; consumer_sockpoll[1].fd = client_socket; consumer_sockpoll[1].events = POLLIN | POLLPRI; - if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) { + ret = lttng_consumer_poll_socket(consumer_sockpoll); + if (ret) { + if (ret > 0) { + /* should exit */ + err = 0; + } goto end; } DBG("Connection on client_socket"); @@ -2928,18 +3036,17 @@ void *consumer_thread_sessiond_poll(void *data) WARN("On accept"); goto end; } - ret = fcntl(sock, F_SETFL, O_NONBLOCK); - if (ret < 0) { - PERROR("fcntl O_NONBLOCK"); - goto end; - } /* * 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) { + if (ret) { + if (ret > 0) { + /* should exit */ + err = 0; + } goto end; } @@ -2955,29 +3062,39 @@ void *consumer_thread_sessiond_poll(void *data) consumer_sockpoll[1].events = POLLIN | POLLPRI; while (1) { - if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) { + health_code_update(); + + health_poll_entry(); + ret = lttng_consumer_poll_socket(consumer_sockpoll); + health_poll_exit(); + if (ret) { + if (ret > 0) { + /* should exit */ + err = 0; + } goto end; } DBG("Incoming command on sock"); ret = lttng_consumer_recv_cmd(ctx, sock, consumer_sockpoll); - if (ret == -ENOENT) { - DBG("Received STOP command"); - goto end; - } if (ret <= 0) { /* * This could simply be a session daemon quitting. Don't output * ERR() here. */ DBG("Communication interrupted on command socket"); + err = 0; goto end; } if (consumer_quit) { DBG("consumer_thread_receive_fds received quit from signal"); + err = 0; /* All is OK */ goto end; } DBG("received command on sock"); } + /* All is OK */ + err = 0; + end: DBG("Consumer thread sessiond poll exiting"); @@ -2987,7 +3104,7 @@ end: * * NOTE: for now, this only applies to the UST tracer. */ - lttng_consumer_close_metadata(); + lttng_consumer_close_all_metadata(); /* * when all fds have hung up, the polling thread @@ -2999,9 +3116,11 @@ end: * 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); - notify_channel_pipe(ctx, NULL, CONSUMER_CHANNEL_QUIT); + notify_health_quit_pipe(health_quit_pipe); /* Cleaning up possibly open sockets. */ if (sock >= 0) { @@ -3011,12 +3130,19 @@ end: } } if (client_socket >= 0) { - ret = close(sock); + ret = close(client_socket); if (ret < 0) { PERROR("close client_socket sessiond poll"); } } +error_testpoint: + if (err) { + health_error(); + ERR("Health error occurred in %s", __func__); + } + health_unregister(health_consumerd); + rcu_unregister_thread(); return NULL; } @@ -3027,6 +3153,9 @@ ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream, ssize_t ret; pthread_mutex_lock(&stream->lock); + if (stream->metadata_flag) { + pthread_mutex_lock(&stream->metadata_rdv_lock); + } switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: @@ -3043,6 +3172,10 @@ ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream, break; } + if (stream->metadata_flag) { + pthread_cond_broadcast(&stream->metadata_rdv); + pthread_mutex_unlock(&stream->metadata_rdv_lock); + } pthread_mutex_unlock(&stream->lock); return ret; } @@ -3065,12 +3198,42 @@ int lttng_consumer_on_recv_stream(struct lttng_consumer_stream *stream) /* * Allocate and set consumer data hash tables. */ -void lttng_consumer_init(void) +int lttng_consumer_init(void) { consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); + if (!consumer_data.channel_ht) { + goto error; + } + consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); + if (!consumer_data.relayd_ht) { + goto error; + } + consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); + if (!consumer_data.stream_list_ht) { + goto error; + } + consumer_data.stream_per_chan_id_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); + if (!consumer_data.stream_per_chan_id_ht) { + goto error; + } + + data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); + if (!data_ht) { + goto error; + } + + metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); + if (!metadata_ht) { + goto error; + } + + return 0; + +error: + return -1; } /* @@ -3079,60 +3242,82 @@ 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_relayd_sock *relayd_sock, unsigned int sessiond_id) + struct lttcomm_relayd_sock *relayd_sock, uint64_t sessiond_id, + uint64_t relayd_session_id) { int fd = -1, ret = -1, relayd_created = 0; - enum lttng_error_code ret_code = LTTNG_OK; - struct consumer_relayd_sock_pair *relayd; + enum lttcomm_return_code ret_code = LTTCOMM_CONSUMERD_SUCCESS; + struct consumer_relayd_sock_pair *relayd = NULL; assert(ctx); assert(relayd_sock); - DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx); - - /* First send a status message before receiving the fds. */ - ret = consumer_send_status_msg(sock, ret_code); - if (ret < 0) { - /* Somehow, the session daemon is not responding anymore. */ - goto error; - } + 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 = -1; + ret = -ENOMEM; + ret_code = LTTCOMM_CONSUMERD_ENOMEM; goto error; + } else { + relayd->sessiond_session_id = sessiond_id; + relayd_created = 1; } - relayd->sessiond_session_id = (uint64_t) 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, LTTCOMM_CONSUMERD_SUCCESS); + 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) { + ret = lttng_consumer_poll_socket(consumer_sockpoll); + if (ret) { + /* Needing to exit in the middle of a command: error. */ + 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. */ - goto error_close; - } - /* We have the fds without error. Send status back. */ - ret = consumer_send_status_msg(sock, ret_code); - if (ret < 0) { - /* Somehow, the session daemon is not responding anymore. */ + /* + * 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; } @@ -3142,64 +3327,51 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, /* Copy received lttcomm socket */ lttcomm_copy_sock(&relayd->control_sock.sock, &relayd_sock->sock); ret = lttcomm_create_sock(&relayd->control_sock.sock); - /* Immediately try to close the created socket if valid. */ - if (relayd->control_sock.sock.fd >= 0) { - if (close(relayd->control_sock.sock.fd)) { - PERROR("close relayd control socket"); - } - } /* Handle create_sock error. */ if (ret < 0) { + ret_code = LTTCOMM_CONSUMERD_ENOMEM; goto error; } + /* + * 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.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); - goto error; - } + relayd->relayd_session_id = relayd_session_id; break; case LTTNG_STREAM_DATA: /* Copy received lttcomm socket */ lttcomm_copy_sock(&relayd->data_sock.sock, &relayd_sock->sock); ret = lttcomm_create_sock(&relayd->data_sock.sock); - /* Immediately try to close the created socket if valid. */ - if (relayd->data_sock.sock.fd >= 0) { - if (close(relayd->data_sock.sock.fd)) { - PERROR("close relayd data socket"); - } - } /* Handle create_sock error. */ if (ret < 0) { + ret_code = LTTCOMM_CONSUMERD_ENOMEM; goto error; } + /* + * 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.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; @@ -3207,6 +3379,7 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, default: ERR("Unknown relayd socket type (%d)", sock_type); ret = -1; + ret_code = LTTCOMM_CONSUMERD_FATAL; goto error; } @@ -3214,6 +3387,14 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, 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. @@ -3224,6 +3405,11 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, 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)) { @@ -3231,7 +3417,6 @@ error: } } -error_close: if (relayd_created) { free(relayd); } @@ -3436,6 +3621,7 @@ int consumer_send_status_msg(int sock, int ret_code) { struct lttcomm_consumer_status_msg msg; + memset(&msg, 0, sizeof(msg)); msg.ret_code = ret_code; return lttcomm_send_unix_sock(sock, &msg, sizeof(msg)); @@ -3453,13 +3639,34 @@ int consumer_send_status_channel(int sock, assert(sock >= 0); + memset(&msg, 0, sizeof(msg)); if (!channel) { - msg.ret_code = -LTTNG_ERR_UST_CHAN_FAIL; + msg.ret_code = LTTCOMM_CONSUMERD_CHANNEL_FAIL; } else { - msg.ret_code = LTTNG_OK; + msg.ret_code = LTTCOMM_CONSUMERD_SUCCESS; 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; +}