X-Git-Url: https://git.lttng.org/?p=lttng-tools.git;a=blobdiff_plain;f=src%2Fcommon%2Fconsumer.c;h=300fd2a2fc896108ac60c43ce920a5f633be111b;hp=af99333c8d2ae1bacc73f7f4af5fcd231e79e104;hb=d8ef542d25837bdfb960e5df2a91c5d18f5ef401;hpb=d21b0d71990ac6ec4272c1f80f0ca544103628b3 diff --git a/src/common/consumer.c b/src/common/consumer.c index af99333c8..300fd2a2f 100644 --- a/src/common/consumer.c +++ b/src/common/consumer.c @@ -47,6 +47,16 @@ struct lttng_consumer_global_data consumer_data = { .type = LTTNG_CONSUMER_UNKNOWN, }; +enum consumer_channel_action { + CONSUMER_CHANNEL_ADD, + CONSUMER_CHANNEL_QUIT, +}; + +struct consumer_channel_msg { + enum consumer_channel_action action; + struct lttng_consumer_channel *chan; +}; + /* * Flag to inform the polling thread to quit when all fd hung up. Updated by * the consumer_thread_receive_fds when it notices that all fds has hung up. @@ -63,21 +73,6 @@ volatile int consumer_quit; static struct lttng_ht *metadata_ht; static struct lttng_ht *data_ht; -/* - * This hash table contains the mapping between the session id of the sessiond - * and the relayd session id. Element of the ht are indexed by sessiond session - * id. - * - * Node can be added when a relayd communication is opened in the sessiond - * thread. - * - * Note that a session id of the session daemon is unique to a tracing session - * and not to a domain session. However, a domain session has one consumer - * which forces the 1-1 mapping between a consumer and a domain session (ex: - * UST). This means that we can't have duplicate in this ht. - */ -static struct lttng_ht *relayd_session_id_ht; - /* * Notify a thread pipe to poll back again. This usually means that some global * state has changed so we just send back the thread in a poll wait call. @@ -93,28 +88,59 @@ static void notify_thread_pipe(int wpipe) } while (ret < 0 && errno == EINTR); } +static void notify_channel_pipe(struct lttng_consumer_local_data *ctx, + struct lttng_consumer_channel *chan, + enum consumer_channel_action action) +{ + struct consumer_channel_msg msg; + int ret; + + msg.action = action; + msg.chan = chan; + do { + ret = write(ctx->consumer_channel_pipe[1], &msg, sizeof(msg)); + } while (ret < 0 && errno == EINTR); +} + +static int read_channel_pipe(struct lttng_consumer_local_data *ctx, + struct lttng_consumer_channel **chan, + 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; + } + 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); } @@ -124,20 +150,20 @@ 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(int 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 = -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 = -1ULL; } rcu_read_unlock(); } @@ -148,20 +174,19 @@ void consumer_steal_stream_key(int key, struct lttng_ht *ht) * RCU read side lock MUST be acquired before calling this function and * protects the channel ptr. */ -static struct lttng_consumer_channel *consumer_find_channel(int key) +struct lttng_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; } - 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); } @@ -169,42 +194,33 @@ static struct lttng_consumer_channel *consumer_find_channel(int key) 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); @@ -231,7 +247,6 @@ static void destroy_relayd(struct consumer_relayd_sock_pair *relayd) { int ret; struct lttng_ht_iter iter; - struct lttng_ht_node_ulong *node; if (relayd == NULL) { return; @@ -239,22 +254,6 @@ static void destroy_relayd(struct consumer_relayd_sock_pair *relayd) DBG("Consumer destroy and close relayd socket pair"); - /* Loockup for a relayd node in the session id map hash table. */ - lttng_ht_lookup(relayd_session_id_ht, - (void *)((unsigned long) relayd->sessiond_session_id), &iter); - node = lttng_ht_iter_get_node_ulong(&iter); - if (node == NULL) { - /* We assume the relayd is being or is destroyed */ - return; - } - - /* - * Try to delete it from the relayd session id ht. The return value is of - * no importance since either way we are going to try to delete the relayd - * from the global relayd_ht. - */ - lttng_ht_del(relayd_session_id_ht, &iter); - iter.iter.node = &relayd->node.node; ret = lttng_ht_del(consumer_data.relayd_ht, &iter); if (ret != 0) { @@ -263,7 +262,44 @@ 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; + + DBG("Consumer delete channel key %" PRIu64, channel->key); + + 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(); + + call_rcu(&channel->node.head, free_channel_rcu); +end: + pthread_mutex_unlock(&consumer_data.lock); } /* @@ -283,8 +319,6 @@ static void cleanup_relayd_ht(void) } lttng_ht_destroy(consumer_data.relayd_ht); - /* The destroy_relayd call makes sure that this ht is empty here. */ - lttng_ht_destroy(relayd_session_id_ht); rcu_read_unlock(); } @@ -402,7 +436,7 @@ void consumer_del_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); @@ -432,7 +466,10 @@ void consumer_del_stream(struct lttng_consumer_stream *stream, ret = lttng_ht_del(ht, &iter); assert(!ret); - /* Remove node session id from the consumer_data stream ht */ + iter.iter.node = &stream->node_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); @@ -447,18 +484,6 @@ void consumer_del_stream(struct lttng_consumer_stream *stream, 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(); @@ -492,7 +517,7 @@ void consumer_del_stream(struct lttng_consumer_stream *stream, uatomic_dec(&stream->chan->refcount); if (!uatomic_read(&stream->chan->refcount) - && !uatomic_read(&stream->chan->nb_init_streams)) { + && !uatomic_read(&stream->chan->nb_init_stream_left)) { free_chan = stream->chan; } @@ -505,91 +530,70 @@ end: 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); } -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, + int relayd_id, uint64_t session_id, - int *alloc_ret) + int cpu, + int *alloc_ret, + enum consumer_channel_type type) { + 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; } rcu_read_lock(); - /* - * 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; - } - 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'; 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); - 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, - stream->net_seq_idx, stream->session_id); + /* 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); rcu_read_unlock(); return stream; @@ -598,13 +602,16 @@ 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, +static int add_stream(struct lttng_consumer_stream *stream, struct lttng_ht *ht) { int ret = 0; @@ -613,23 +620,26 @@ static int consumer_add_stream(struct lttng_consumer_stream *stream, 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->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); + 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); @@ -641,14 +651,14 @@ static int consumer_add_stream(struct lttng_consumer_stream *stream, uatomic_inc(&stream->chan->refcount); /* - * When nb_init_streams reaches 0, we don't need to trigger any action in - * terms of destroying the associated channel, because the action that + * 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) { + uatomic_dec(&stream->chan->nb_init_stream_left); } /* Update consumer data once the node is inserted. */ @@ -669,22 +679,18 @@ 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; @@ -712,7 +718,7 @@ 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); + lttng_ht_node_init_u64(&obj->node, obj->net_seq_idx); pthread_mutex_init(&obj->ctrl_sock_mutex, NULL); error: @@ -726,20 +732,20 @@ 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); } @@ -782,7 +788,14 @@ static int write_relayd_stream_header(struct lttng_consumer_stream *stream, data_hdr.stream_id = htobe64(stream->relayd_stream_id); data_hdr.data_size = htobe32(data_size); data_hdr.padding_size = htobe32(padding); - 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, @@ -791,6 +804,8 @@ 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; } @@ -799,119 +814,51 @@ 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; - - DBG("Consumer delete channel key %d", channel->key); - - 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(unsigned long key, + uint64_t session_id, + const char *pathname, + const char *name, + uid_t uid, + gid_t gid, + int relayd_id, + enum lttng_event_output output) { 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->uid = uid; + channel->gid = gid; + channel->relayd_id = relayd_id; + channel->output = output; + + 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) - 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; - } - 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); end: return channel; } @@ -919,31 +866,38 @@ end: /* * Add a channel to the global list protected by a mutex. */ -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); rcu_read_lock(); lttng_ht_lookup(consumer_data.channel_ht, - (void *)((unsigned long) channel->key), &iter); - node = lttng_ht_iter_get_node_ulong(&iter); + &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; } - 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(&consumer_data.lock); - return 0; + if (!ret && channel->wait_fd != -1 && + channel->metadata_stream == NULL) { + notify_channel_pipe(ctx, channel, CONSUMER_CHANNEL_ADD); + } + return ret; } /* @@ -953,14 +907,19 @@ 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) { @@ -1027,8 +986,8 @@ 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; } @@ -1046,8 +1005,7 @@ 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, @@ -1064,14 +1022,12 @@ int lttng_consumer_send_error( 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); } @@ -1081,6 +1037,8 @@ void lttng_consumer_cleanup(void) 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 @@ -1099,7 +1057,7 @@ 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"); } @@ -1117,11 +1075,11 @@ 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); @@ -1139,8 +1097,8 @@ 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); } /* @@ -1165,7 +1123,7 @@ struct lttng_consumer_local_data *lttng_consumer_create( int (*recv_stream)(struct lttng_consumer_stream *stream), int (*update_stream)(int stream_key, uint32_t state)) { - int ret, i; + int ret; struct lttng_consumer_local_data *ctx; assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN || @@ -1217,6 +1175,12 @@ struct lttng_consumer_local_data *lttng_consumer_create( 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) { goto error_metadata_pipe; @@ -1232,26 +1196,14 @@ struct lttng_consumer_local_data *lttng_consumer_create( 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: - for (i = 0; i < 2; i++) { - int err; - - err = close(ctx->consumer_should_quit[i]); - if (err) { - PERROR("close"); - } - } + utils_close_pipe(ctx->consumer_should_quit); error_poll_fcntl: error_quit_pipe: - for (i = 0; i < 2; i++) { - int err; - - err = close(ctx->consumer_data_pipe[i]); - if (err) { - PERROR("close"); - } - } + utils_close_pipe(ctx->consumer_data_pipe); error_poll_pipe: free(ctx); error: @@ -1271,30 +1223,10 @@ 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]); - 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); + utils_close_pipe(ctx->consumer_should_quit); utils_close_pipe(ctx->consumer_splice_metadata_pipe); unlink(ctx->consumer_command_sock_path); @@ -1306,8 +1238,7 @@ 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; @@ -1317,8 +1248,22 @@ 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", @@ -1345,6 +1290,7 @@ 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 */ @@ -1366,12 +1312,18 @@ 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"); @@ -1431,11 +1383,17 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( 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; } @@ -1698,21 +1656,19 @@ 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; } - } /* @@ -1720,17 +1676,15 @@ 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); @@ -1808,6 +1762,32 @@ static void destroy_stream_ht(struct lttng_ht *ht) 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. */ @@ -1830,7 +1810,7 @@ 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); @@ -1860,7 +1840,10 @@ 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_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); @@ -1873,20 +1856,6 @@ 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); @@ -1919,7 +1888,7 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream, /* 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)) { + && !uatomic_read(&stream->chan->nb_init_stream_left)) { /* Go for channel deletion! */ free_chan = stream->chan; } @@ -1932,26 +1901,26 @@ end: 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, +static int add_metadata_stream(struct lttng_consumer_stream *stream, struct lttng_ht *ht) { int ret = 0; struct consumer_relayd_sock_pair *relayd; struct lttng_ht_iter iter; - struct lttng_ht_node_ulong *node; + 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->lock); @@ -1967,8 +1936,8 @@ static int consumer_add_metadata_stream(struct lttng_consumer_stream *stream, * Lookup the stream just to make sure it does not exist in our internal * state. This should NEVER happen. */ - lttng_ht_lookup(ht, (void *)((unsigned long) stream->wait_fd), &iter); - node = lttng_ht_iter_get_node_ulong(&iter); + lttng_ht_lookup(ht, &stream->key, &iter); + node = lttng_ht_iter_get_node_u64(&iter); assert(!node); /* Find relayd and, if one is found, increment refcount. */ @@ -1981,24 +1950,27 @@ static int consumer_add_metadata_stream(struct lttng_consumer_stream *stream, uatomic_inc(&stream->chan->refcount); /* - * When nb_init_streams reaches 0, we don't need to trigger any action in - * terms of destroying the associated channel, because the action that + * 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) { + uatomic_dec(&stream->chan->nb_init_stream_left); } - lttng_ht_add_unique_ulong(ht, &stream->node); + lttng_ht_add_unique_u64(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(); @@ -2070,17 +2042,17 @@ 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_ULONG); + metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64); if (!metadata_ht) { /* ENOMEM at this point. Better to bail out. */ - goto error; + goto end_ht; } DBG("Thread metadata poll started"); @@ -2089,7 +2061,7 @@ void *consumer_thread_metadata_poll(void *data) 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); @@ -2162,13 +2134,13 @@ 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); + ret = add_metadata_stream(stream, metadata_ht); if (ret) { ERR("Unable to add metadata stream"); /* Stream was not setup properly. Continuing. */ @@ -2186,9 +2158,12 @@ 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, @@ -2245,10 +2220,11 @@ restart: error: end: DBG("Metadata poll thread exiting"); - lttng_poll_clean(&events); + lttng_poll_clean(&events); +end_poll: destroy_stream_ht(metadata_ht); - +end_ht: rcu_unregister_thread(); return NULL; } @@ -2270,7 +2246,7 @@ 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; @@ -2288,14 +2264,11 @@ 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)); @@ -2313,7 +2286,7 @@ void *consumer_thread_data_poll(void *data) 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"); @@ -2356,7 +2329,7 @@ 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 */ @@ -2364,6 +2337,11 @@ void *consumer_thread_data_poll(void *data) pipe_readlen = read(ctx->consumer_data_pipe[0], &new_stream, sizeof(new_stream)); } while (pipe_readlen == -1 && errno == EINTR); + if (pipe_readlen < 0) { + PERROR("read consumer data pipe"); + /* Continue so we can at least handle the current stream(s). */ + continue; + } /* * If the stream is NULL, just ignore it. It's also possible that @@ -2375,9 +2353,9 @@ void *consumer_thread_data_poll(void *data) continue; } - ret = consumer_add_stream(new_stream, data_ht); + ret = add_stream(new_stream, data_ht); if (ret) { - ERR("Consumer add stream %d failed. Continuing", + ERR("Consumer add stream %" PRIu64 " failed. Continuing", new_stream->key); /* * At this point, if the add_stream fails, it is not in the @@ -2487,14 +2465,8 @@ 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 @@ -2515,6 +2487,222 @@ end: 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) { + 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); + } + } + 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; + } + + 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; + + ret = read_channel_pipe(ctx, &chan, &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); + lttng_ht_add_unique_u64(channel_ht, + &chan->wait_fd_node); + /* Add channel to the global poll events list */ + lttng_poll_add(&events, chan->wait_fd, + LPOLLIN | LPOLLPRI); + break; + 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 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; +} + /* * This thread listens on the consumerd socket and receives the file * descriptors from the session daemon. @@ -2571,7 +2759,7 @@ 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; } @@ -2614,10 +2802,18 @@ 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"); + + /* + * Close metadata streams since the producer is the session daemon which + * just died. + * + * NOTE: for now, this only applies to the UST tracer. + */ + lttng_consumer_close_metadata(); /* * when all fds have hung up, the polling thread @@ -2631,6 +2827,8 @@ end: */ notify_thread_pipe(ctx->consumer_data_pipe[1]); + notify_channel_pipe(ctx, NULL, CONSUMER_CHANNEL_QUIT); + /* Cleaning up possibly open sockets. */ if (sock >= 0) { ret = close(sock); @@ -2695,10 +2893,10 @@ 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); - relayd_session_id_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); + 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); } /* @@ -2715,7 +2913,6 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, int fd = -1, ret = -1, relayd_created = 0; enum lttng_error_code ret_code = LTTNG_OK; struct consumer_relayd_sock_pair *relayd; - struct consumer_relayd_session_id *relayd_id_node; DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx); @@ -2733,6 +2930,7 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, relayd = consumer_allocate_relayd_sock_pair(net_seq_idx); if (relayd == NULL) { lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_OUTFD_ERROR); + ret = -1; goto error; } relayd->sessiond_session_id = (uint64_t) sessiond_id; @@ -2751,7 +2949,7 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, 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; + goto error_close; } /* We have the fds without error. Send status back. */ @@ -2782,32 +2980,28 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, relayd->control_sock.fd = fd; /* - * Create a session on the relayd and store the returned id. No need to - * grab the socket lock since the relayd object is not yet visible. + * 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 (ret < 0) { - goto error; + if (!relayd_created) { + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); } - - /* Set up a relayd session id node. */ - relayd_id_node = zmalloc(sizeof(struct consumer_relayd_session_id)); - if (!relayd_id_node) { - PERROR("zmalloc relayd id node"); + 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_id_node->relayd_id = relayd->relayd_session_id; - relayd_id_node->sessiond_id = (uint64_t) sessiond_id; - - /* Indexed by session id of the sessiond. */ - lttng_ht_node_init_ulong(&relayd_id_node->node, - relayd_id_node->sessiond_id); - rcu_read_lock(); - lttng_ht_add_unique_ulong(relayd_session_id_ht, &relayd_id_node->node); - rcu_read_unlock(); - break; case LTTNG_STREAM_DATA: /* Copy received lttcomm socket */ @@ -2829,10 +3023,11 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type, break; default: ERR("Unknown relayd socket type (%d)", sock_type); + ret = -1; 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); @@ -2853,10 +3048,8 @@ error: } } +error_close: if (relayd_created) { - /* We just want to cleanup. Ignore ret value. */ - (void) relayd_close(&relayd->control_sock); - (void) relayd_close(&relayd->data_sock); free(relayd); } @@ -2900,30 +3093,24 @@ end: static struct consumer_relayd_sock_pair *find_relayd_by_session_id(uint64_t id) { struct lttng_ht_iter iter; - struct lttng_ht_node_ulong *node; struct consumer_relayd_sock_pair *relayd = NULL; - struct consumer_relayd_session_id *session_id_map; - - /* Get the session id map. */ - lttng_ht_lookup(relayd_session_id_ht, (void *)((unsigned long) id), &iter); - node = lttng_ht_iter_get_node_ulong(&iter); - if (node == NULL) { - goto end; - } - - session_id_map = caa_container_of(node, struct consumer_relayd_session_id, - node); /* 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) { - if (relayd->relayd_session_id == session_id_map->relayd_id) { + /* + * 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. */ - break; + goto found; } } -end: + return NULL; + +found: return relayd; } @@ -2977,8 +3164,8 @@ int consumer_data_pending(uint64_t id) } cds_lfht_for_each_entry_duplicate(ht->ht, - ht->hash_fct((void *)((unsigned long) id), lttng_ht_seed), - ht->match_fct, (void *)((unsigned long) id), + 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); @@ -3006,10 +3193,12 @@ int consumer_data_pending(uint64_t id) 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) { @@ -3068,3 +3257,26 @@ int consumer_send_status_msg(int sock, int 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)); +}