X-Git-Url: https://git.lttng.org/?p=lttng-tools.git;a=blobdiff_plain;f=src%2Fcommon%2Fconsumer.c;h=8d897e5df1e20e20c67ddbc6ed8f99af7e7de2aa;hp=c1dadddb3dd08854bffd8eef8cca3e9a76e82bed;hb=c8f59ee5fc11492ef472dc5cfd2fd2c4926b1787;hpb=3f8e211fbe73cbcf69d52af5e839b14d1a951ed7 diff --git a/src/common/consumer.c b/src/common/consumer.c index c1dadddb3..8d897e5df 100644 --- a/src/common/consumer.c +++ b/src/common/consumer.c @@ -27,8 +27,11 @@ #include #include #include +#include #include +#include +#include #include #include #include @@ -55,24 +58,38 @@ int consumer_poll_timeout = -1; */ volatile int consumer_quit = 0; +/* + * The following two hash tables are visible by all threads which are separated + * in different source files. + * + * Global hash table containing respectively metadata and data streams. The + * stream element in this ht should only be updated by the metadata poll thread + * for the metadata and the data poll thread for the data. + */ +struct lttng_ht *metadata_ht = NULL; +struct lttng_ht *data_ht = NULL; + /* * Find a stream. The consumer_data.lock must be locked during this * call. */ -static struct lttng_consumer_stream *consumer_find_stream(int key) +static struct lttng_consumer_stream *consumer_find_stream(int key, + struct lttng_ht *ht) { struct lttng_ht_iter iter; struct lttng_ht_node_ulong *node; struct lttng_consumer_stream *stream = NULL; + assert(ht); + /* Negative keys are lookup failures */ - if (key < 0) + if (key < 0) { return NULL; + } rcu_read_lock(); - lttng_ht_lookup(consumer_data.stream_ht, (void *)((unsigned long) key), - &iter); + lttng_ht_lookup(ht, (void *)((unsigned long) key), &iter); node = lttng_ht_iter_get_node_ulong(&iter); if (node != NULL) { stream = caa_container_of(node, struct lttng_consumer_stream, node); @@ -83,12 +100,12 @@ static struct lttng_consumer_stream *consumer_find_stream(int key) return stream; } -static void consumer_steal_stream_key(int key) +void consumer_steal_stream_key(int key, struct lttng_ht *ht) { struct lttng_consumer_stream *stream; rcu_read_lock(); - stream = consumer_find_stream(key); + stream = consumer_find_stream(key, ht); if (stream) { stream->key = -1; /* @@ -108,8 +125,9 @@ static struct lttng_consumer_channel *consumer_find_channel(int key) struct lttng_consumer_channel *channel = NULL; /* Negative keys are lookup failures */ - if (key < 0) + if (key < 0) { return NULL; + } rcu_read_lock(); @@ -172,7 +190,7 @@ static void consumer_rcu_free_relayd(struct rcu_head *head) * * This function MUST be called with the consumer_data lock acquired. */ -void consumer_destroy_relayd(struct consumer_relayd_sock_pair *relayd) +static void destroy_relayd(struct consumer_relayd_sock_pair *relayd) { int ret; struct lttng_ht_iter iter; @@ -200,11 +218,31 @@ void consumer_destroy_relayd(struct consumer_relayd_sock_pair *relayd) call_rcu(&relayd->node.head, consumer_rcu_free_relayd); } +/* + * Flag a relayd socket pair for destruction. Destroy it if the refcount + * reaches zero. + * + * RCU read side lock MUST be aquired before calling this function. + */ +void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd) +{ + assert(relayd); + + /* Set destroy flag for this object */ + uatomic_set(&relayd->destroy_flag, 1); + + /* Destroy the relayd if refcount is 0 */ + if (uatomic_read(&relayd->refcount) == 0) { + destroy_relayd(relayd); + } +} + /* * Remove a stream from the global list protected by a mutex. This * function is also responsible for freeing its data structures. */ -void consumer_del_stream(struct lttng_consumer_stream *stream) +void consumer_del_stream(struct lttng_consumer_stream *stream, + struct lttng_ht *ht) { int ret; struct lttng_ht_iter iter; @@ -213,6 +251,11 @@ void consumer_del_stream(struct lttng_consumer_stream *stream) assert(stream); + if (ht == NULL) { + /* Means the stream was allocated but not successfully added */ + goto free_stream; + } + pthread_mutex_lock(&consumer_data.lock); switch (consumer_data.type) { @@ -220,7 +263,7 @@ void consumer_del_stream(struct lttng_consumer_stream *stream) if (stream->mmap_base != NULL) { ret = munmap(stream->mmap_base, stream->mmap_len); if (ret != 0) { - perror("munmap"); + PERROR("munmap"); } } break; @@ -236,18 +279,18 @@ void consumer_del_stream(struct lttng_consumer_stream *stream) rcu_read_lock(); iter.iter.node = &stream->node.node; - ret = lttng_ht_del(consumer_data.stream_ht, &iter); + ret = lttng_ht_del(ht, &iter); assert(!ret); + /* Remove node session id from the consumer_data stream ht */ + iter.iter.node = &stream->node_session_id.node; + ret = lttng_ht_del(consumer_data.stream_list_ht, &iter); + assert(!ret); rcu_read_unlock(); - if (consumer_data.stream_count <= 0) { - goto end; - } + assert(consumer_data.stream_count > 0); consumer_data.stream_count--; - if (!stream) { - goto end; - } + if (stream->out_fd >= 0) { ret = close(stream->out_fd); if (ret) { @@ -292,23 +335,27 @@ void consumer_del_stream(struct lttng_consumer_stream *stream) /* Both conditions are met, we destroy the relayd. */ if (uatomic_read(&relayd->refcount) == 0 && uatomic_read(&relayd->destroy_flag)) { - consumer_destroy_relayd(relayd); + destroy_relayd(relayd); } } rcu_read_unlock(); - if (!--stream->chan->refcount) { + uatomic_dec(&stream->chan->refcount); + if (!uatomic_read(&stream->chan->refcount) + && !uatomic_read(&stream->chan->nb_init_streams)) { free_chan = stream->chan; } - - call_rcu(&stream->node.head, consumer_free_stream); end: consumer_data.need_update = 1; pthread_mutex_unlock(&consumer_data.lock); - if (free_chan) + if (free_chan) { consumer_del_channel(free_chan); + } + +free_stream: + call_rcu(&stream->node.head, consumer_free_stream); } struct lttng_consumer_stream *consumer_allocate_stream( @@ -321,22 +368,30 @@ struct lttng_consumer_stream *consumer_allocate_stream( uid_t uid, gid_t gid, int net_index, - int metadata_flag) + int metadata_flag, + uint64_t session_id, + int *alloc_ret) { struct lttng_consumer_stream *stream; - int ret; stream = zmalloc(sizeof(*stream)); if (stream == NULL) { - perror("malloc struct lttng_consumer_stream"); + PERROR("malloc struct lttng_consumer_stream"); + *alloc_ret = -ENOMEM; goto end; } + + /* + * Get stream's channel reference. Needed when adding the stream to the + * global hash table. + */ stream->chan = consumer_find_channel(channel_key); if (!stream->chan) { - perror("Unable to find channel key"); - goto end; + *alloc_ret = -ENOENT; + ERR("Unable to find channel for stream %d", stream_key); + goto error; } - stream->chan->refcount++; + stream->key = stream_key; stream->shm_fd = shm_fd; stream->wait_fd = wait_fd; @@ -350,91 +405,100 @@ struct lttng_consumer_stream *consumer_allocate_stream( stream->gid = gid; stream->net_seq_idx = net_index; stream->metadata_flag = metadata_flag; + stream->session_id = session_id; strncpy(stream->path_name, path_name, sizeof(stream->path_name)); stream->path_name[sizeof(stream->path_name) - 1] = '\0'; - lttng_ht_node_init_ulong(&stream->node, stream->key); - lttng_ht_node_init_ulong(&stream->waitfd_node, stream->wait_fd); + pthread_mutex_init(&stream->lock, NULL); - switch (consumer_data.type) { - case LTTNG_CONSUMER_KERNEL: - break; - case LTTNG_CONSUMER32_UST: - case LTTNG_CONSUMER64_UST: - stream->cpu = stream->chan->cpucount++; - ret = lttng_ustconsumer_allocate_stream(stream); - if (ret) { - free(stream); - return NULL; - } - break; - default: - ERR("Unknown consumer_data type"); - assert(0); - goto end; + /* + * Index differently the metadata node because the thread is using an + * internal hash table to match streams in the metadata_ht to the epoll set + * file descriptor. + */ + if (metadata_flag) { + lttng_ht_node_init_ulong(&stream->node, stream->wait_fd); + } else { + lttng_ht_node_init_ulong(&stream->node, stream->key); } - DBG("Allocated stream %s (key %d, shm_fd %d, wait_fd %d, mmap_len %llu, out_fd %d, net_seq_idx %d)", - stream->path_name, stream->key, - stream->shm_fd, - stream->wait_fd, - (unsigned long long) stream->mmap_len, - stream->out_fd, - stream->net_seq_idx); -end: + + /* Init session id node with the stream session id */ + lttng_ht_node_init_ulong(&stream->node_session_id, stream->session_id); + + /* + * The cpu number is needed before using any ustctl_* actions. Ignored for + * 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); + + 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); return stream; + +error: + free(stream); +end: + return NULL; } /* * Add a stream to the global list protected by a mutex. */ -int consumer_add_stream(struct lttng_consumer_stream *stream) +static int consumer_add_stream(struct lttng_consumer_stream *stream, + struct lttng_ht *ht) { int ret = 0; - struct lttng_ht_node_ulong *node; - struct lttng_ht_iter iter; struct consumer_relayd_sock_pair *relayd; - pthread_mutex_lock(&consumer_data.lock); - /* Steal stream identifier, for UST */ - consumer_steal_stream_key(stream->key); + assert(stream); + assert(ht); + + DBG3("Adding consumer stream %d", stream->key); + pthread_mutex_lock(&consumer_data.lock); rcu_read_lock(); - lttng_ht_lookup(consumer_data.stream_ht, - (void *)((unsigned long) stream->key), &iter); - node = lttng_ht_iter_get_node_ulong(&iter); - if (node != NULL) { - rcu_read_unlock(); - /* Stream already exist. Ignore the insertion */ - goto end; - } - lttng_ht_add_unique_ulong(consumer_data.stream_ht, &stream->node); + /* Steal stream identifier to avoid having streams with the same key */ + consumer_steal_stream_key(stream->key, ht); + + lttng_ht_add_unique_ulong(ht, &stream->node); + + /* + * Add stream to the stream_list_ht of the consumer data. No need to steal + * the key since the HT does not use it and we allow to add redundant keys + * into this table. + */ + lttng_ht_add_ulong(consumer_data.stream_list_ht, &stream->node_session_id); /* Check and cleanup relayd */ relayd = consumer_find_relayd(stream->net_seq_idx); if (relayd != NULL) { uatomic_inc(&relayd->refcount); } - rcu_read_unlock(); - /* Update consumer data */ - consumer_data.stream_count++; - consumer_data.need_update = 1; + /* Update channel refcount once added without error(s). */ + uatomic_inc(&stream->chan->refcount); - switch (consumer_data.type) { - case LTTNG_CONSUMER_KERNEL: - break; - case LTTNG_CONSUMER32_UST: - case LTTNG_CONSUMER64_UST: - /* Streams are in CPU number order (we rely on this) */ - stream->cpu = stream->chan->nr_streams++; - break; - default: - ERR("Unknown consumer_data type"); - assert(0); - goto end; + /* + * When nb_init_streams reaches 0, we don't need to trigger any action in + * terms of destroying the associated channel, because the action that + * causes the count to become 0 also causes a stream to be added. The + * channel deletion will thus be triggered by the following removal of this + * stream. + */ + if (uatomic_read(&stream->chan->nb_init_streams) > 0) { + uatomic_dec(&stream->chan->nb_init_streams); } -end: + /* Update consumer data once the node is inserted. */ + consumer_data.stream_count++; + consumer_data.need_update = 1; + + rcu_read_unlock(); pthread_mutex_unlock(&consumer_data.lock); return ret; @@ -444,8 +508,7 @@ end: * Add relayd socket to global consumer data hashtable. RCU read side lock MUST * be acquired before calling this. */ - -int consumer_add_relayd(struct consumer_relayd_sock_pair *relayd) +static int add_relayd(struct consumer_relayd_sock_pair *relayd) { int ret = 0; struct lttng_ht_node_ulong *node; @@ -534,7 +597,8 @@ error: * Return destination file descriptor or negative value on error. */ static int write_relayd_stream_header(struct lttng_consumer_stream *stream, - size_t data_size, struct consumer_relayd_sock_pair *relayd) + size_t data_size, unsigned long padding, + struct consumer_relayd_sock_pair *relayd) { int outfd = -1, ret; struct lttcomm_relayd_data_hdr data_hdr; @@ -559,6 +623,7 @@ static int write_relayd_stream_header(struct lttng_consumer_stream *stream, /* Set header with stream information */ data_hdr.stream_id = htobe64(stream->relayd_stream_id); data_hdr.data_size = htobe32(data_size); + data_hdr.padding_size = htobe32(padding); data_hdr.net_seq_num = htobe64(stream->next_net_seq_num++); /* Other fields are zeroed previously */ @@ -576,23 +641,6 @@ error: return outfd; } -/* - * Update a stream according to what we just received. - */ -void consumer_change_stream_state(int stream_key, - enum lttng_consumer_stream_state state) -{ - struct lttng_consumer_stream *stream; - - pthread_mutex_lock(&consumer_data.lock); - stream = consumer_find_stream(stream_key); - if (stream) { - stream->state = state; - } - consumer_data.need_update = 1; - pthread_mutex_unlock(&consumer_data.lock); -} - static void consumer_free_channel(struct rcu_head *head) { @@ -637,7 +685,7 @@ void consumer_del_channel(struct lttng_consumer_channel *channel) if (channel->mmap_base != NULL) { ret = munmap(channel->mmap_base, channel->mmap_len); if (ret != 0) { - perror("munmap"); + PERROR("munmap"); } } if (channel->wait_fd >= 0 && !channel->wait_fd_is_copy) { @@ -662,14 +710,15 @@ 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) + uint64_t max_sb_size, + unsigned int nb_init_streams) { struct lttng_consumer_channel *channel; int ret; channel = zmalloc(sizeof(*channel)); if (channel == NULL) { - perror("malloc struct lttng_consumer_channel"); + PERROR("malloc struct lttng_consumer_channel"); goto end; } channel->key = channel_key; @@ -678,7 +727,7 @@ struct lttng_consumer_channel *consumer_allocate_channel( channel->mmap_len = mmap_len; channel->max_sb_size = max_sb_size; channel->refcount = 0; - channel->nr_streams = 0; + channel->nb_init_streams = nb_init_streams; lttng_ht_node_init_ulong(&channel->node, channel->key); switch (consumer_data.type) { @@ -744,10 +793,9 @@ end: * * Returns the number of fds in the structures. */ -int consumer_update_poll_array( +static int consumer_update_poll_array( struct lttng_consumer_local_data *ctx, struct pollfd **pollfd, - struct lttng_consumer_stream **local_stream, - struct lttng_ht *metadata_ht) + struct lttng_consumer_stream **local_stream, struct lttng_ht *ht) { int i = 0; struct lttng_ht_iter iter; @@ -755,28 +803,23 @@ int consumer_update_poll_array( DBG("Updating poll fd array"); rcu_read_lock(); - cds_lfht_for_each_entry(consumer_data.stream_ht->ht, &iter.iter, stream, - node.node) { + cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM) { continue; } DBG("Active FD %d", stream->wait_fd); (*pollfd)[i].fd = stream->wait_fd; (*pollfd)[i].events = POLLIN | POLLPRI; - if (stream->metadata_flag && metadata_ht) { - lttng_ht_add_unique_ulong(metadata_ht, &stream->waitfd_node); - DBG("Active FD added to metadata hash table"); - } local_stream[i] = stream; i++; } rcu_read_unlock(); /* - * Insert the consumer_poll_pipe at the end of the array and don't + * Insert the consumer_data_pipe at the end of the array and don't * increment i so nb_fd is the number of real FD. */ - (*pollfd)[i].fd = ctx->consumer_poll_pipe[0]; + (*pollfd)[i].fd = ctx->consumer_data_pipe[0]; (*pollfd)[i].events = POLLIN | POLLPRI; return i; } @@ -798,7 +841,7 @@ restart: if (errno == EINTR) { goto restart; } - perror("Poll error"); + PERROR("Poll error"); goto exit; } if (consumer_sockpoll[0].revents & (POLLIN | POLLPRI)) { @@ -855,17 +898,6 @@ void lttng_consumer_cleanup(void) rcu_read_lock(); - /* - * close all outfd. Called when there are no more threads running (after - * joining on the threads), no need to protect list iteration with mutex. - */ - cds_lfht_for_each_entry(consumer_data.stream_ht->ht, &iter.iter, node, - node) { - struct lttng_consumer_stream *stream = - caa_container_of(node, struct lttng_consumer_stream, node); - consumer_del_stream(stream); - } - cds_lfht_for_each_entry(consumer_data.channel_ht->ht, &iter.iter, node, node) { struct lttng_consumer_channel *channel = @@ -875,7 +907,6 @@ void lttng_consumer_cleanup(void) rcu_read_unlock(); - lttng_ht_destroy(consumer_data.stream_ht); lttng_ht_destroy(consumer_data.channel_ht); } @@ -890,7 +921,7 @@ void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx) ret = write(ctx->consumer_should_quit[1], "4", 1); } while (ret < 0 && errno == EINTR); if (ret < 0) { - perror("write consumer quit"); + PERROR("write consumer quit"); } } @@ -962,7 +993,7 @@ struct lttng_consumer_local_data *lttng_consumer_create( ctx = zmalloc(sizeof(struct lttng_consumer_local_data)); if (ctx == NULL) { - perror("allocating context"); + PERROR("allocating context"); goto error; } @@ -973,41 +1004,54 @@ struct lttng_consumer_local_data *lttng_consumer_create( ctx->on_recv_stream = recv_stream; ctx->on_update_stream = update_stream; - ret = pipe(ctx->consumer_poll_pipe); + ret = pipe(ctx->consumer_data_pipe); if (ret < 0) { - perror("Error creating poll pipe"); + PERROR("Error creating poll pipe"); goto error_poll_pipe; } /* set read end of the pipe to non-blocking */ - ret = fcntl(ctx->consumer_poll_pipe[0], F_SETFL, O_NONBLOCK); + ret = fcntl(ctx->consumer_data_pipe[0], F_SETFL, O_NONBLOCK); if (ret < 0) { - perror("fcntl O_NONBLOCK"); + PERROR("fcntl O_NONBLOCK"); goto error_poll_fcntl; } /* set write end of the pipe to non-blocking */ - ret = fcntl(ctx->consumer_poll_pipe[1], F_SETFL, O_NONBLOCK); + ret = fcntl(ctx->consumer_data_pipe[1], F_SETFL, O_NONBLOCK); if (ret < 0) { - perror("fcntl O_NONBLOCK"); + PERROR("fcntl O_NONBLOCK"); goto error_poll_fcntl; } ret = pipe(ctx->consumer_should_quit); if (ret < 0) { - perror("Error creating recv pipe"); + PERROR("Error creating recv pipe"); goto error_quit_pipe; } ret = pipe(ctx->consumer_thread_pipe); if (ret < 0) { - perror("Error creating thread pipe"); + PERROR("Error creating thread pipe"); goto error_thread_pipe; } - return ctx; + ret = utils_create_pipe(ctx->consumer_metadata_pipe); + if (ret < 0) { + 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_thread_pipe); error_thread_pipe: for (i = 0; i < 2; i++) { int err; @@ -1022,7 +1066,7 @@ error_quit_pipe: for (i = 0; i < 2; i++) { int err; - err = close(ctx->consumer_poll_pipe[i]); + err = close(ctx->consumer_data_pipe[i]); if (err) { PERROR("close"); } @@ -1052,11 +1096,11 @@ void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx) if (ret) { PERROR("close"); } - ret = close(ctx->consumer_poll_pipe[0]); + ret = close(ctx->consumer_data_pipe[0]); if (ret) { PERROR("close"); } - ret = close(ctx->consumer_poll_pipe[1]); + ret = close(ctx->consumer_data_pipe[1]); if (ret) { PERROR("close"); } @@ -1068,6 +1112,8 @@ void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx) if (ret) { PERROR("close"); } + utils_close_pipe(ctx->consumer_splice_metadata_pipe); + unlink(ctx->consumer_command_sock_path); free(ctx); } @@ -1077,22 +1123,23 @@ 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) + struct consumer_relayd_sock_pair *relayd, + unsigned long padding) { int ret; - uint64_t metadata_id; + struct lttcomm_relayd_metadata_payload hdr; - metadata_id = htobe64(stream->relayd_stream_id); + hdr.stream_id = htobe64(stream->relayd_stream_id); + hdr.padding_size = htobe32(padding); do { - ret = write(fd, (void *) &metadata_id, - sizeof(stream->relayd_stream_id)); + ret = write(fd, (void *) &hdr, sizeof(hdr)); } while (ret < 0 && errno == EINTR); if (ret < 0) { PERROR("write metadata stream id"); goto end; } - DBG("Metadata stream id %zu written before data", - stream->relayd_stream_id); + DBG("Metadata stream id %" PRIu64 " with padding %lu written before data", + stream->relayd_stream_id, padding); end: return ret; @@ -1109,7 +1156,8 @@ end: */ ssize_t lttng_consumer_on_read_subbuffer_mmap( struct lttng_consumer_local_data *ctx, - struct lttng_consumer_stream *stream, unsigned long len) + struct lttng_consumer_stream *stream, unsigned long len, + unsigned long padding) { unsigned long mmap_offset; ssize_t ret = 0, written = 0; @@ -1121,6 +1169,8 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( /* RCU lock for the relayd pointer */ rcu_read_lock(); + pthread_mutex_lock(&stream->lock); + /* Flag that the current stream if set for network streaming. */ if (stream->net_seq_idx != -1) { relayd = consumer_find_relayd(stream->net_seq_idx); @@ -1161,17 +1211,17 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( if (stream->metadata_flag) { /* Metadata requires the control socket. */ pthread_mutex_lock(&relayd->ctrl_sock_mutex); - netlen += sizeof(stream->relayd_stream_id); + netlen += sizeof(struct lttcomm_relayd_metadata_payload); } - ret = write_relayd_stream_header(stream, netlen, relayd); + ret = write_relayd_stream_header(stream, netlen, padding, relayd); if (ret >= 0) { /* Use the returned socket. */ outfd = ret; /* Write metadata stream id before payload */ if (stream->metadata_flag) { - ret = write_relayd_metadata_id(outfd, stream, relayd); + ret = write_relayd_metadata_id(outfd, stream, relayd, padding); if (ret < 0) { written = ret; goto end; @@ -1179,12 +1229,16 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( } } /* Else, use the default set before which is the filesystem. */ + } else { + /* No streaming, we have to set the len with the full padding */ + len += padding; } while (len > 0) { do { ret = write(outfd, stream->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"); if (written == 0) { @@ -1192,14 +1246,13 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( } goto end; } else if (ret > len) { - PERROR("Error in file write (ret %ld > len %lu)", ret, len); + PERROR("Error in file write (ret %zd > len %lu)", ret, len); written += ret; goto end; } else { len -= ret; mmap_offset += ret; } - DBG("Consumer mmap write() ret %ld (len %lu)", ret, len); /* This call is useless on a socket so better save a syscall. */ if (!relayd) { @@ -1213,6 +1266,7 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( lttng_consumer_sync_trace_file(stream, orig_offset); end: + pthread_mutex_unlock(&stream->lock); /* Unlock only if ctrl socket used */ if (relayd && stream->metadata_flag) { pthread_mutex_unlock(&relayd->ctrl_sock_mutex); @@ -1229,7 +1283,8 @@ end: */ ssize_t lttng_consumer_on_read_subbuffer_splice( struct lttng_consumer_local_data *ctx, - struct lttng_consumer_stream *stream, unsigned long len) + struct lttng_consumer_stream *stream, unsigned long len, + unsigned long padding) { ssize_t ret = 0, written = 0, ret_splice = 0; loff_t offset = 0; @@ -1238,6 +1293,7 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( /* Default is on the disk */ int outfd = stream->out_fd; struct consumer_relayd_sock_pair *relayd = NULL; + int *splice_pipe; switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: @@ -1254,6 +1310,8 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( /* RCU lock for the relayd pointer */ rcu_read_lock(); + pthread_mutex_lock(&stream->lock); + /* Flag that the current stream if set for network streaming. */ if (stream->net_seq_idx != -1) { relayd = consumer_find_relayd(stream->net_seq_idx); @@ -1262,26 +1320,55 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( } } + /* + * 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; + } + /* Write metadata stream id before payload */ - if (stream->metadata_flag && relayd) { - /* - * Lock the control socket for the complete duration of the function - * since from this point on we will use the socket. - */ - pthread_mutex_lock(&relayd->ctrl_sock_mutex); + if (relayd) { + int total_len = len; - ret = write_relayd_metadata_id(ctx->consumer_thread_pipe[1], - stream, relayd); - if (ret < 0) { - written = ret; + if (stream->metadata_flag) { + /* + * Lock the control socket for the complete duration of the function + * since from this point on we will use the socket. + */ + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + + ret = write_relayd_metadata_id(splice_pipe[1], stream, relayd, + padding); + if (ret < 0) { + written = ret; + goto end; + } + + 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 { + ERR("Remote relayd disconnected. Stopping"); goto end; } + } else { + /* No streaming, we have to set the len with the full padding */ + len += padding; } while (len > 0) { - DBG("splice chan to pipe offset %lu of len %lu (fd : %d)", - (unsigned long)offset, len, fd); - ret_splice = splice(fd, &offset, ctx->consumer_thread_pipe[1], NULL, len, + DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)", + (unsigned long)offset, len, fd, splice_pipe[1]); + ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len, SPLICE_F_MOVE | SPLICE_F_MORE); DBG("splice chan to pipe, ret %zd", ret_splice); if (ret_splice < 0) { @@ -1296,30 +1383,24 @@ ssize_t lttng_consumer_on_read_subbuffer_splice( /* 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); + /* Update counter to fit the spliced data */ - ret_splice += sizeof(stream->relayd_stream_id); - len += sizeof(stream->relayd_stream_id); + 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 -= sizeof(stream->relayd_stream_id); - } - - ret = write_relayd_stream_header(stream, ret_splice, relayd); - if (ret >= 0) { - /* Use the returned socket. */ - outfd = ret; - } else { - ERR("Remote relayd disconnected. Stopping"); - goto end; + written -= metadata_payload_size; } } /* Splice data out */ - ret_splice = splice(ctx->consumer_thread_pipe[0], NULL, outfd, NULL, + ret_splice = splice(splice_pipe[0], NULL, outfd, NULL, ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE); - DBG("Kernel consumer splice pipe to file, ret %zd", ret_splice); + DBG("Consumer splice pipe to file, ret %zd", ret_splice); if (ret_splice < 0) { PERROR("Error in file splice"); if (written == 0) { @@ -1356,20 +1437,21 @@ splice_error: /* send the appropriate error description to sessiond */ switch (ret) { case EBADF: - lttng_consumer_send_error(ctx, CONSUMERD_SPLICE_EBADF); + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EBADF); break; case EINVAL: - lttng_consumer_send_error(ctx, CONSUMERD_SPLICE_EINVAL); + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL); break; case ENOMEM: - lttng_consumer_send_error(ctx, CONSUMERD_SPLICE_ENOMEM); + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM); break; case ESPIPE: - lttng_consumer_send_error(ctx, CONSUMERD_SPLICE_ESPIPE); + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE); break; } end: + pthread_mutex_unlock(&stream->lock); if (relayd && stream->metadata_flag) { pthread_mutex_unlock(&relayd->ctrl_sock_mutex); } @@ -1440,96 +1522,515 @@ int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx, } /* - * This thread polls the fds in the set to consume the data and write - * it to tracefile if necessary. + * Iterate over all streams of the hashtable and free them properly. + * + * WARNING: *MUST* be used with data stream only. */ -void *lttng_consumer_thread_poll_fds(void *data) +static void destroy_data_stream_ht(struct lttng_ht *ht) { - int num_rdy, num_hup, high_prio, ret, i; - struct pollfd *pollfd = NULL; - /* local view of the streams */ - struct lttng_consumer_stream **local_stream = NULL; - /* local view of consumer_data.fds_count */ - int nb_fd = 0; - struct lttng_consumer_local_data *ctx = data; - struct lttng_ht *metadata_ht; + int ret; struct lttng_ht_iter iter; - struct lttng_ht_node_ulong *node; - struct lttng_consumer_stream *metadata_stream; - ssize_t len; + struct lttng_consumer_stream *stream; - metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); + if (ht == NULL) { + return; + } - rcu_register_thread(); + rcu_read_lock(); + cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { + ret = lttng_ht_del(ht, &iter); + assert(!ret); - local_stream = zmalloc(sizeof(struct lttng_consumer_stream)); + call_rcu(&stream->node.head, consumer_free_stream); + } + rcu_read_unlock(); - while (1) { - high_prio = 0; - num_hup = 0; + lttng_ht_destroy(ht); +} - /* - * the fds set has been updated, we need to update our - * local array as well - */ - 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; - } +/* + * 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) +{ + int ret; + struct lttng_ht_iter iter; + struct lttng_consumer_stream *stream; - /* allocate for all fds + 1 for the consumer_poll_pipe */ - pollfd = zmalloc((consumer_data.stream_count + 1) * sizeof(struct pollfd)); - if (pollfd == NULL) { - perror("pollfd malloc"); - pthread_mutex_unlock(&consumer_data.lock); - goto end; - } + if (ht == NULL) { + return; + } - /* allocate for all fds + 1 for the consumer_poll_pipe */ - local_stream = zmalloc((consumer_data.stream_count + 1) * - sizeof(struct lttng_consumer_stream)); - if (local_stream == NULL) { - perror("local_stream malloc"); - pthread_mutex_unlock(&consumer_data.lock); - goto end; - } - ret = consumer_update_poll_array(ctx, &pollfd, local_stream, - metadata_ht); - if (ret < 0) { - ERR("Error in allocating pollfd or local_outfds"); - lttng_consumer_send_error(ctx, CONSUMERD_POLL_ERROR); - pthread_mutex_unlock(&consumer_data.lock); - goto end; - } - nb_fd = ret; - consumer_data.need_update = 0; - } - pthread_mutex_unlock(&consumer_data.lock); + rcu_read_lock(); + cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { + ret = lttng_ht_del(ht, &iter); + assert(!ret); - /* No FDs and consumer_quit, consumer_cleanup the thread */ - if (nb_fd == 0 && consumer_quit == 1) { - goto end; - } - /* poll on the array of fds */ - restart: - DBG("polling on %d fd", nb_fd + 1); - num_rdy = poll(pollfd, nb_fd + 1, consumer_poll_timeout); - DBG("poll num_rdy : %d", num_rdy); - if (num_rdy == -1) { - /* - * Restart interrupted system call. - */ - if (errno == EINTR) { - goto restart; - } - perror("Poll error"); - lttng_consumer_send_error(ctx, CONSUMERD_POLL_ERROR); + call_rcu(&stream->node.head, consumer_free_stream); + } + rcu_read_unlock(); + + lttng_ht_destroy(ht); +} + +/* + * Clean up a metadata stream and free its memory. + */ +void consumer_del_metadata_stream(struct lttng_consumer_stream *stream, + struct lttng_ht *ht) +{ + int ret; + struct lttng_ht_iter iter; + struct lttng_consumer_channel *free_chan = NULL; + struct consumer_relayd_sock_pair *relayd; + + assert(stream); + /* + * This call should NEVER receive regular stream. It must always be + * metadata stream and this is crucial for data structure synchronization. + */ + assert(stream->metadata_flag); + + DBG3("Consumer delete metadata stream %d", stream->wait_fd); + + if (ht == NULL) { + /* Means the stream was allocated but not successfully added */ + goto free_stream; + } + + pthread_mutex_lock(&consumer_data.lock); + 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); + + /* Remove node session id from the consumer_data stream ht */ + iter.iter.node = &stream->node_session_id.node; + ret = lttng_ht_del(consumer_data.stream_list_ht, &iter); + assert(!ret); + rcu_read_unlock(); + + if (stream->out_fd >= 0) { + ret = close(stream->out_fd); + if (ret) { + PERROR("close"); + } + } + + if (stream->wait_fd >= 0 && !stream->wait_fd_is_copy) { + ret = close(stream->wait_fd); + if (ret) { + PERROR("close"); + } + } + + if (stream->shm_fd >= 0 && stream->wait_fd != stream->shm_fd) { + ret = close(stream->shm_fd); + if (ret) { + PERROR("close"); + } + } + + /* Check and cleanup relayd */ + rcu_read_lock(); + relayd = consumer_find_relayd(stream->net_seq_idx); + if (relayd != NULL) { + uatomic_dec(&relayd->refcount); + assert(uatomic_read(&relayd->refcount) >= 0); + + /* Closing streams requires to lock the control socket. */ + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + ret = relayd_send_close_stream(&relayd->control_sock, + stream->relayd_stream_id, stream->next_net_seq_num - 1); + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); + if (ret < 0) { + DBG("Unable to close stream on the relayd. Continuing"); + /* + * Continue here. There is nothing we can do for the relayd. + * Chances are that the relayd has closed the socket so we just + * continue cleaning up. + */ + } + + /* Both conditions are met, we destroy the relayd. */ + if (uatomic_read(&relayd->refcount) == 0 && + uatomic_read(&relayd->destroy_flag)) { + destroy_relayd(relayd); + } + } + rcu_read_unlock(); + + /* 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)) { + /* Go for channel deletion! */ + free_chan = stream->chan; + } + +end: + pthread_mutex_unlock(&consumer_data.lock); + + if (free_chan) { + consumer_del_channel(free_chan); + } + +free_stream: + call_rcu(&stream->node.head, consumer_free_stream); +} + +/* + * Action done with the metadata stream when adding it to the consumer internal + * data structures to handle it. + */ +static int consumer_add_metadata_stream(struct lttng_consumer_stream *stream, + struct lttng_ht *ht) +{ + int ret = 0; + struct consumer_relayd_sock_pair *relayd; + + assert(stream); + assert(ht); + + DBG3("Adding metadata stream %d to hash table", stream->wait_fd); + + pthread_mutex_lock(&consumer_data.lock); + + /* + * From here, refcounts are updated so be _careful_ when returning an error + * after this point. + */ + + rcu_read_lock(); + /* Find relayd and, if one is found, increment refcount. */ + relayd = consumer_find_relayd(stream->net_seq_idx); + if (relayd != NULL) { + uatomic_inc(&relayd->refcount); + } + + /* Update channel refcount once added without error(s). */ + uatomic_inc(&stream->chan->refcount); + + /* + * When nb_init_streams reaches 0, we don't need to trigger any action in + * terms of destroying the associated channel, because the action that + * causes the count to become 0 also causes a stream to be added. The + * channel deletion will thus be triggered by the following removal of this + * stream. + */ + if (uatomic_read(&stream->chan->nb_init_streams) > 0) { + uatomic_dec(&stream->chan->nb_init_streams); + } + + /* Steal stream identifier to avoid having streams with the same key */ + consumer_steal_stream_key(stream->key, ht); + + lttng_ht_add_unique_ulong(ht, &stream->node); + + /* + * Add stream to the stream_list_ht of the consumer data. No need to steal + * the key since the HT does not use it and we allow to add redundant keys + * into this table. + */ + lttng_ht_add_ulong(consumer_data.stream_list_ht, &stream->node_session_id); + + rcu_read_unlock(); + + pthread_mutex_unlock(&consumer_data.lock); + return ret; +} + +/* + * Thread polls on metadata file descriptor and write them on disk or on the + * network. + */ +void *consumer_thread_metadata_poll(void *data) +{ + int ret, i, pollfd; + uint32_t revents, nb_fd; + struct lttng_consumer_stream *stream = NULL; + struct lttng_ht_iter iter; + struct lttng_ht_node_ulong *node; + struct lttng_poll_event events; + struct lttng_consumer_local_data *ctx = data; + ssize_t len; + + rcu_register_thread(); + + DBG("Thread metadata poll started"); + + /* Size is set to 1 for the consumer_metadata pipe */ + ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC); + if (ret < 0) { + ERR("Poll set creation failed"); + goto end; + } + + ret = lttng_poll_add(&events, ctx->consumer_metadata_pipe[0], LPOLLIN); + if (ret < 0) { + goto end; + } + + /* Main loop */ + DBG("Metadata main loop started"); + + while (1) { + lttng_poll_reset(&events); + + nb_fd = LTTNG_POLL_GETNB(&events); + + /* Only the metadata pipe is set */ + if (nb_fd == 0 && consumer_quit == 1) { + goto end; + } + +restart: + DBG("Metadata poll wait with %d fd(s)", nb_fd); + ret = lttng_poll_wait(&events, -1); + DBG("Metadata event catched in thread"); + if (ret < 0) { + if (errno == EINTR) { + ERR("Poll EINTR catched"); + goto restart; + } + goto error; + } + + /* From here, the event is a metadata wait fd */ + for (i = 0; i < nb_fd; i++) { + revents = LTTNG_POLL_GETEV(&events, i); + pollfd = LTTNG_POLL_GETFD(&events, i); + + /* Just don't waste time if no returned events for the fd */ + if (!revents) { + continue; + } + + if (pollfd == ctx->consumer_metadata_pipe[0]) { + if (revents & (LPOLLERR | LPOLLHUP )) { + DBG("Metadata thread pipe hung up"); + /* + * Remove the pipe from the poll set and continue the loop + * since their might be data to consume. + */ + lttng_poll_del(&events, ctx->consumer_metadata_pipe[0]); + close(ctx->consumer_metadata_pipe[0]); + continue; + } else if (revents & LPOLLIN) { + do { + /* Get the stream pointer received */ + ret = read(pollfd, &stream, sizeof(stream)); + } while (ret < 0 && errno == EINTR); + if (ret < 0 || + ret < sizeof(struct lttng_consumer_stream *)) { + PERROR("read metadata stream"); + /* + * Let's continue here and hope we can still work + * without stopping the consumer. XXX: Should we? + */ + continue; + } + + DBG("Adding metadata stream %d to poll set", + stream->wait_fd); + + ret = consumer_add_metadata_stream(stream, metadata_ht); + if (ret) { + ERR("Unable to add metadata stream"); + /* Stream was not setup properly. Continuing. */ + consumer_del_metadata_stream(stream, NULL); + continue; + } + + /* Add metadata stream to the global poll events list */ + lttng_poll_add(&events, stream->wait_fd, + LPOLLIN | LPOLLPRI); + } + + /* Handle other stream */ + continue; + } + + rcu_read_lock(); + lttng_ht_lookup(metadata_ht, (void *)((unsigned long) pollfd), + &iter); + node = lttng_ht_iter_get_node_ulong(&iter); + assert(node); + + stream = caa_container_of(node, struct lttng_consumer_stream, + node); + + /* Check for error event */ + if (revents & (LPOLLERR | LPOLLHUP)) { + DBG("Metadata fd %d is hup|err.", pollfd); + if (!stream->hangup_flush_done + && (consumer_data.type == LTTNG_CONSUMER32_UST + || consumer_data.type == LTTNG_CONSUMER64_UST)) { + DBG("Attempting to flush and consume the UST buffers"); + lttng_ustconsumer_on_stream_hangup(stream); + + /* We just flushed the stream now read it. */ + do { + len = ctx->on_buffer_ready(stream, ctx); + /* + * We don't check the return value here since if we get + * a negative len, it means an error occured thus we + * simply remove it from the poll set and free the + * stream. + */ + } while (len > 0); + } + + lttng_poll_del(&events, stream->wait_fd); + /* + * This call update the channel states, closes file descriptors + * and securely free the stream. + */ + consumer_del_metadata_stream(stream, metadata_ht); + } else if (revents & (LPOLLIN | LPOLLPRI)) { + /* Get the data out of the metadata file descriptor */ + DBG("Metadata available on fd %d", pollfd); + assert(stream->wait_fd == pollfd); + + len = ctx->on_buffer_ready(stream, ctx); + /* It's ok to have an unavailable sub-buffer */ + if (len < 0 && len != -EAGAIN && len != -ENODATA) { + rcu_read_unlock(); + goto end; + } else if (len > 0) { + stream->data_read = 1; + } + } + + /* Release RCU lock for the stream looked up */ + rcu_read_unlock(); + } + } + +error: +end: + DBG("Metadata poll thread exiting"); + lttng_poll_clean(&events); + + if (metadata_ht) { + destroy_stream_ht(metadata_ht); + } + + rcu_unregister_thread(); + return NULL; +} + +/* + * This thread polls the fds in the set to consume the data and write + * it to tracefile if necessary. + */ +void *consumer_thread_data_poll(void *data) +{ + int num_rdy, num_hup, high_prio, ret, i; + struct pollfd *pollfd = NULL; + /* local view of the streams */ + struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL; + /* local view of consumer_data.fds_count */ + int nb_fd = 0; + struct lttng_consumer_local_data *ctx = data; + ssize_t len; + + rcu_register_thread(); + + data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); + if (data_ht == NULL) { + goto end; + } + + local_stream = zmalloc(sizeof(struct lttng_consumer_stream)); + + while (1) { + high_prio = 0; + num_hup = 0; + + /* + * the fds set has been updated, we need to update our + * local array as well + */ + 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; + } + + /* allocate for all fds + 1 for the consumer_data_pipe */ + pollfd = zmalloc((consumer_data.stream_count + 1) * sizeof(struct pollfd)); + if (pollfd == NULL) { + PERROR("pollfd malloc"); + 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)); + if (local_stream == NULL) { + PERROR("local_stream malloc"); + pthread_mutex_unlock(&consumer_data.lock); + goto end; + } + ret = consumer_update_poll_array(ctx, &pollfd, local_stream, + data_ht); + if (ret < 0) { + ERR("Error in allocating pollfd or local_outfds"); + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR); + pthread_mutex_unlock(&consumer_data.lock); + goto end; + } + nb_fd = ret; + consumer_data.need_update = 0; + } + pthread_mutex_unlock(&consumer_data.lock); + + /* No FDs and consumer_quit, consumer_cleanup the thread */ + if (nb_fd == 0 && consumer_quit == 1) { + goto end; + } + /* poll on the array of fds */ + restart: + DBG("polling on %d fd", nb_fd + 1); + num_rdy = poll(pollfd, nb_fd + 1, consumer_poll_timeout); + DBG("poll num_rdy : %d", num_rdy); + if (num_rdy == -1) { + /* + * Restart interrupted system call. + */ + if (errno == EINTR) { + goto restart; + } + PERROR("Poll error"); + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR); goto end; } else if (num_rdy == 0) { DBG("Polling thread timed out"); @@ -1537,47 +2038,52 @@ void *lttng_consumer_thread_poll_fds(void *data) } /* - * If the consumer_poll_pipe triggered poll go directly to the + * If the consumer_data_pipe triggered poll go directly to the * beginning of the loop to update the array. We want to prioritize * array update over low-priority reads. */ if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) { size_t pipe_readlen; - char tmp; - DBG("consumer_poll_pipe wake up"); + DBG("consumer_data_pipe wake up"); /* Consume 1 byte of pipe data */ do { - pipe_readlen = read(ctx->consumer_poll_pipe[0], &tmp, 1); + pipe_readlen = read(ctx->consumer_data_pipe[0], &new_stream, + sizeof(new_stream)); } while (pipe_readlen == -1 && errno == EINTR); + + /* + * If the stream is NULL, just ignore it. It's also possible that + * the sessiond poll thread changed the consumer_quit state and is + * waking us up to test it. + */ + if (new_stream == NULL) { + continue; + } + + ret = consumer_add_stream(new_stream, data_ht); + if (ret) { + ERR("Consumer add stream %d failed. Continuing", + new_stream->key); + /* + * At this point, if the add_stream fails, it is not in the + * hash table thus passing the NULL value here. + */ + consumer_del_stream(new_stream, NULL); + } + + /* Continue to update the local streams and handle prio ones */ continue; } /* Take care of high priority channels first. */ for (i = 0; i < nb_fd; i++) { - /* Lookup for metadata which is the highest priority */ - lttng_ht_lookup(metadata_ht, - (void *)((unsigned long) pollfd[i].fd), &iter); - node = lttng_ht_iter_get_node_ulong(&iter); - if (node != NULL && - (pollfd[i].revents & (POLLIN | POLLPRI))) { - DBG("Urgent metadata read on fd %d", pollfd[i].fd); - metadata_stream = caa_container_of(node, - struct lttng_consumer_stream, waitfd_node); - high_prio = 1; - len = ctx->on_buffer_ready(metadata_stream, ctx); - /* it's ok to have an unavailable sub-buffer */ - if (len < 0 && len != -EAGAIN) { - goto end; - } else if (len > 0) { - metadata_stream->data_read = 1; - } - } else if (pollfd[i].revents & POLLPRI) { + if (pollfd[i].revents & POLLPRI) { DBG("Urgent read on fd %d", pollfd[i].fd); high_prio = 1; len = ctx->on_buffer_ready(local_stream[i], ctx); /* it's ok to have an unavailable sub-buffer */ - if (len < 0 && len != -EAGAIN) { + if (len < 0 && len != -EAGAIN && len != -ENODATA) { goto end; } else if (len > 0) { local_stream[i]->data_read = 1; @@ -1600,7 +2106,7 @@ void *lttng_consumer_thread_poll_fds(void *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 */ - if (len < 0 && len != -EAGAIN) { + if (len < 0 && len != -EAGAIN && len != -ENODATA) { goto end; } else if (len > 0) { local_stream[i]->data_read = 1; @@ -1628,34 +2134,19 @@ void *lttng_consumer_thread_poll_fds(void *data) if ((pollfd[i].revents & POLLHUP)) { DBG("Polling fd %d tells it has hung up.", pollfd[i].fd); if (!local_stream[i]->data_read) { - if (local_stream[i]->metadata_flag) { - iter.iter.node = &local_stream[i]->waitfd_node.node; - ret = lttng_ht_del(metadata_ht, &iter); - assert(!ret); - } - consumer_del_stream(local_stream[i]); + consumer_del_stream(local_stream[i], data_ht); num_hup++; } } else if (pollfd[i].revents & POLLERR) { ERR("Error returned in polling fd %d.", pollfd[i].fd); if (!local_stream[i]->data_read) { - if (local_stream[i]->metadata_flag) { - iter.iter.node = &local_stream[i]->waitfd_node.node; - ret = lttng_ht_del(metadata_ht, &iter); - assert(!ret); - } - consumer_del_stream(local_stream[i]); + consumer_del_stream(local_stream[i], data_ht); num_hup++; } } else if (pollfd[i].revents & POLLNVAL) { ERR("Polling fd %d tells fd is not open.", pollfd[i].fd); if (!local_stream[i]->data_read) { - if (local_stream[i]->metadata_flag) { - iter.iter.node = &local_stream[i]->waitfd_node.node; - ret = lttng_ht_del(metadata_ht, &iter); - assert(!ret); - } - consumer_del_stream(local_stream[i]); + consumer_del_stream(local_stream[i], data_ht); num_hup++; } } @@ -1672,6 +2163,21 @@ end: free(local_stream); local_stream = NULL; } + + /* + * Close the write side of the pipe so epoll_wait() in + * consumer_thread_metadata_poll can catch it. The thread is monitoring the + * read side of the pipe. If we close them both, epoll_wait strangely does + * not return and could create a endless wait period if the pipe is the + * only tracked fd in the poll set. The thread will take care of closing + * the read side. + */ + close(ctx->consumer_metadata_pipe[1]); + + if (data_ht) { + destroy_data_stream_ht(data_ht); + } + rcu_unregister_thread(); return NULL; } @@ -1680,7 +2186,7 @@ end: * This thread listens on the consumerd socket and receives the file * descriptors from the session daemon. */ -void *lttng_consumer_thread_receive_fds(void *data) +void *consumer_thread_sessiond_poll(void *data) { int sock, client_socket, ret; /* @@ -1706,7 +2212,7 @@ void *lttng_consumer_thread_receive_fds(void *data) } DBG("Sending ready command to lttng-sessiond"); - ret = lttng_consumer_send_error(ctx, CONSUMERD_COMMAND_SOCK_READY); + ret = lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY); /* return < 0 on error, but == 0 is not fatal */ if (ret < 0) { ERR("Error sending ready command to lttng-sessiond"); @@ -1715,7 +2221,7 @@ void *lttng_consumer_thread_receive_fds(void *data) ret = fcntl(client_socket, F_SETFL, O_NONBLOCK); if (ret < 0) { - perror("fcntl O_NONBLOCK"); + PERROR("fcntl O_NONBLOCK"); goto end; } @@ -1738,7 +2244,7 @@ void *lttng_consumer_thread_receive_fds(void *data) } ret = fcntl(sock, F_SETFL, O_NONBLOCK); if (ret < 0) { - perror("fcntl O_NONBLOCK"); + PERROR("fcntl O_NONBLOCK"); goto end; } @@ -1756,8 +2262,12 @@ void *lttng_consumer_thread_receive_fds(void *data) DBG("Received STOP command"); goto end; } - if (ret < 0) { - ERR("Communication interrupted on command socket"); + if (ret <= 0) { + /* + * This could simply be a session daemon quitting. Don't output + * ERR() here. + */ + DBG("Communication interrupted on command socket"); goto end; } if (consumer_quit) { @@ -1783,19 +2293,16 @@ end: consumer_poll_timeout = LTTNG_CONSUMER_POLL_TIMEOUT; /* - * Wake-up the other end by writing a null byte in the pipe - * (non-blocking). Important note: Because writing into the - * pipe is non-blocking (and therefore we allow dropping wakeup - * data, as long as there is wakeup data present in the pipe - * buffer to wake up the other end), the other end should - * perform the following sequence for waiting: - * 1) empty the pipe (reads). - * 2) perform update operation. - * 3) wait on the pipe (poll). + * Notify the data poll thread to poll back again and test the + * consumer_quit state to quit gracefully. */ do { - ret = write(ctx->consumer_poll_pipe[1], "", 1); + struct lttng_consumer_stream *null_stream = NULL; + + ret = write(ctx->consumer_data_pipe[1], &null_stream, + sizeof(null_stream)); } while (ret < 0 && errno == EINTR); + rcu_unregister_thread(); return NULL; } @@ -1836,7 +2343,190 @@ int lttng_consumer_on_recv_stream(struct lttng_consumer_stream *stream) */ void lttng_consumer_init(void) { - consumer_data.stream_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); + consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); + + metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); + assert(metadata_ht); + data_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); + assert(data_ht); +} + +/* + * Process the ADD_RELAYD command receive by a consumer. + * + * 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, + struct lttng_consumer_local_data *ctx, int sock, + struct pollfd *consumer_sockpoll, struct lttcomm_sock *relayd_sock) +{ + int fd, ret = -1; + struct consumer_relayd_sock_pair *relayd; + + DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx); + + /* Get relayd reference if exists. */ + relayd = consumer_find_relayd(net_seq_idx); + if (relayd == NULL) { + /* 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); + goto error; + } + } + + /* Poll on consumer socket. */ + if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) { + ret = -EINTR; + goto error; + } + + /* 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; + goto error; + } + + /* Copy socket information and received FD */ + switch (sock_type) { + case LTTNG_STREAM_CONTROL: + /* Copy received lttcomm socket */ + lttcomm_copy_sock(&relayd->control_sock, relayd_sock); + ret = lttcomm_create_sock(&relayd->control_sock); + if (ret < 0) { + goto error; + } + + /* Close the created socket fd which is useless */ + close(relayd->control_sock.fd); + + /* Assign new file descriptor */ + relayd->control_sock.fd = fd; + break; + case LTTNG_STREAM_DATA: + /* Copy received lttcomm socket */ + lttcomm_copy_sock(&relayd->data_sock, relayd_sock); + ret = lttcomm_create_sock(&relayd->data_sock); + if (ret < 0) { + goto error; + } + + /* Close the created socket fd which is useless */ + close(relayd->data_sock.fd); + + /* Assign new file descriptor */ + relayd->data_sock.fd = fd; + break; + default: + ERR("Unknown relayd socket type (%d)", sock_type); + goto error; + } + + DBG("Consumer %s socket created successfully with net idx %d (fd: %d)", + sock_type == LTTNG_STREAM_CONTROL ? "control" : "data", + relayd->net_seq_idx, fd); + + /* + * Add relayd socket pair to consumer data hashtable. If object already + * exists or on error, the function gracefully returns. + */ + add_relayd(relayd); + + /* All good! */ + ret = 0; + +error: + return ret; +} + +/* + * Check if for a given session id there is still data needed to be extract + * from the buffers. + * + * Return 1 if data is in fact available to be read or else 0. + */ +int consumer_data_available(uint64_t id) +{ + int ret; + struct lttng_ht_iter iter; + struct lttng_ht *ht; + struct lttng_consumer_stream *stream; + struct consumer_relayd_sock_pair *relayd; + int (*data_available)(struct lttng_consumer_stream *); + + DBG("Consumer data available command on session id %" PRIu64, id); + + pthread_mutex_lock(&consumer_data.lock); + + switch (consumer_data.type) { + case LTTNG_CONSUMER_KERNEL: + data_available = lttng_kconsumer_data_available; + break; + case LTTNG_CONSUMER32_UST: + case LTTNG_CONSUMER64_UST: + data_available = lttng_ustconsumer_data_available; + break; + default: + ERR("Unknown consumer data type"); + assert(0); + } + + rcu_read_lock(); + + /* Ease our life a bit */ + ht = consumer_data.stream_list_ht; + + cds_lfht_for_each_entry_duplicate(ht->ht, + ht->hash_fct((void *)((unsigned long) id), 0x42UL), + ht->match_fct, (void *)((unsigned long) id), + &iter.iter, stream, node_session_id.node) { + /* Check the stream for data. */ + ret = data_available(stream); + if (ret == 0) { + goto data_not_available; + } + + if (stream->net_seq_idx != -1) { + relayd = consumer_find_relayd(stream->net_seq_idx); + assert(relayd); + + pthread_mutex_lock(&stream->lock); + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + if (stream->metadata_flag) { + ret = relayd_quiescent_control(&relayd->control_sock); + } else { + ret = relayd_data_available(&relayd->control_sock, + stream->relayd_stream_id, stream->next_net_seq_num); + } + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); + pthread_mutex_unlock(&stream->lock); + if (ret == 0) { + goto data_not_available; + } + } + } + + /* + * Finding _no_ node in the hash table means that the stream(s) have been + * removed thus data is guaranteed to be available for analysis from the + * trace files. This is *only* true for local consumer and not network + * streaming. + */ + + /* Data is available to be read by a viewer. */ + pthread_mutex_unlock(&consumer_data.lock); + rcu_read_unlock(); + return 1; + +data_not_available: + /* Data is still being extracted from buffers. */ + pthread_mutex_unlock(&consumer_data.lock); + rcu_read_unlock(); + return 0; }