X-Git-Url: https://git.lttng.org/?p=lttng-tools.git;a=blobdiff_plain;f=src%2Fcommon%2Fust-consumer%2Fust-consumer.c;h=a113ef12941762bfd054541e42d0c26924043629;hp=fbc3bbb43d5e3d9164606676dc38875d05473823;hb=ea853771023cfeb9ca8ca3853a64d444855d60fb;hpb=4628484ae860b8cf506427e864c7d1933da32777 diff --git a/src/common/ust-consumer/ust-consumer.c b/src/common/ust-consumer/ust-consumer.c index fbc3bbb43..a113ef129 100644 --- a/src/common/ust-consumer/ust-consumer.c +++ b/src/common/ust-consumer/ust-consumer.c @@ -16,7 +16,6 @@ * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. */ -#define _GNU_SOURCE #define _LGPL_SOURCE #include #include @@ -39,15 +38,15 @@ #include #include #include -#include -#include -#include +#include +#include +#include #include #include #include "ust-consumer.h" -#define UINT_MAX_STR_LEN 11 /* includes \0 */ +#define INT_MAX_STR_LEN 12 /* includes \0 */ extern struct lttng_consumer_global_data consumer_data; extern int consumer_poll_timeout; @@ -82,6 +81,7 @@ static void destroy_channel(struct lttng_consumer_channel *channel) */ if (channel->uchan) { lttng_ustconsumer_del_channel(channel); + lttng_ustconsumer_free_channel(channel); } free(channel); } @@ -247,14 +247,14 @@ error: static int get_stream_shm_path(char *stream_shm_path, const char *shm_path, int cpu) { - char cpu_nr[UINT_MAX_STR_LEN]; /* unsigned int max len */ + char cpu_nr[INT_MAX_STR_LEN]; /* int max len */ int ret; strncpy(stream_shm_path, shm_path, PATH_MAX); stream_shm_path[PATH_MAX - 1] = '\0'; - ret = snprintf(cpu_nr, UINT_MAX_STR_LEN, "%u", cpu); - if (ret != 1) { - ret = -1; + ret = snprintf(cpu_nr, INT_MAX_STR_LEN, "%i", cpu); + if (ret < 0) { + PERROR("snprintf"); goto end; } strncat(stream_shm_path, cpu_nr, @@ -352,8 +352,12 @@ static int create_ust_streams(struct lttng_consumer_channel *channel, /* Keep stream reference when creating metadata. */ if (channel->type == CONSUMER_CHANNEL_TYPE_METADATA) { channel->metadata_stream = stream; - stream->ust_metadata_poll_pipe[0] = ust_metadata_pipe[0]; - stream->ust_metadata_poll_pipe[1] = ust_metadata_pipe[1]; + if (channel->monitor) { + /* Set metadata poll pipe if we created one */ + memcpy(stream->ust_metadata_poll_pipe, + ust_metadata_pipe, + sizeof(ust_metadata_pipe)); + } } } @@ -500,7 +504,6 @@ error_open: closeret = run_as_unlink(shm_path, channel->uid, channel->gid); if (closeret) { - errno = -closeret; PERROR("unlink %s", shm_path); } } @@ -1026,6 +1029,8 @@ static int snapshot_channel(uint64_t key, char *path, uint64_t relayd_id, DBG("UST consumer snapshot channel %" PRIu64, key); cds_list_for_each_entry(stream, &channel->streams.head, send_node) { + /* Are we at a position _before_ the first available packet ? */ + bool before_first_packet = true; health_code_update(); @@ -1092,6 +1097,7 @@ static int snapshot_channel(uint64_t key, char *path, uint64_t relayd_id, while (consumed_pos < produced_pos) { ssize_t read_len; unsigned long len, padded_len; + int lost_packet = 0; health_code_update(); @@ -1105,6 +1111,15 @@ static int snapshot_channel(uint64_t key, char *path, uint64_t relayd_id, } DBG("UST consumer get subbuf failed. Skipping it."); consumed_pos += stream->max_sb_size; + + /* + * Start accounting lost packets only when we + * already have extracted packets (to match the + * content of the final snapshot). + */ + if (!before_first_packet) { + lost_packet = 1; + } continue; } @@ -1140,6 +1155,16 @@ static int snapshot_channel(uint64_t key, char *path, uint64_t relayd_id, goto error_close_stream; } consumed_pos += stream->max_sb_size; + + /* + * Only account lost packets located between + * succesfully extracted packets (do not account before + * and after since they are not visible in the + * resulting snapshot). + */ + stream->chan->lost_packets += lost_packet; + lost_packet = 0; + before_first_packet = false; } /* Simply close the stream so we can use it on the next snapshot. */ @@ -1164,11 +1189,16 @@ error: } /* - * Receive the metadata updates from the sessiond. + * Receive the metadata updates from the sessiond. Supports receiving + * overlapping metadata, but is needs to always belong to a contiguous + * range starting from 0. + * Be careful about the locks held when calling this function: it needs + * the metadata cache flush to concurrently progress in order to + * complete. */ int lttng_ustconsumer_recv_metadata(int sock, uint64_t key, uint64_t offset, - uint64_t len, struct lttng_consumer_channel *channel, - int timer, int wait) + uint64_t len, uint64_t version, + struct lttng_consumer_channel *channel, int timer, int wait) { int ret, ret_code = LTTCOMM_CONSUMERD_SUCCESS; char *metadata_str; @@ -1195,7 +1225,8 @@ int lttng_ustconsumer_recv_metadata(int sock, uint64_t key, uint64_t offset, health_code_update(); pthread_mutex_lock(&channel->metadata_cache->lock); - ret = consumer_metadata_cache_write(channel, offset, len, metadata_str); + ret = consumer_metadata_cache_write(channel, offset, len, version, + metadata_str); if (ret < 0) { /* Unable to handle metadata. Notify session daemon. */ ret_code = LTTCOMM_CONSUMERD_ERROR_METADATA; @@ -1557,6 +1588,7 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx, uint64_t len = msg.u.push_metadata.len; uint64_t key = msg.u.push_metadata.key; uint64_t offset = msg.u.push_metadata.target_offset; + uint64_t version = msg.u.push_metadata.version; struct lttng_consumer_channel *channel; DBG("UST consumer push metadata key %" PRIu64 " of len %" PRIu64, key, @@ -1578,6 +1610,15 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx, health_code_update(); + if (!len) { + /* + * There is nothing to receive. We have simply + * checked whether the channel can be found. + */ + ret_code = LTTCOMM_CONSUMERD_SUCCESS; + goto end_msg_sessiond; + } + /* Tell session daemon we are ready to receive the metadata. */ ret = consumer_send_status_msg(sock, LTTCOMM_CONSUMERD_SUCCESS); if (ret < 0) { @@ -1598,7 +1639,7 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx, health_code_update(); ret = lttng_ustconsumer_recv_metadata(sock, key, offset, - len, channel, 0, 1); + len, version, channel, 0, 1); if (ret < 0) { /* error receiving from sessiond */ goto error_fatal; @@ -1649,6 +1690,107 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx, health_code_update(); break; } + case LTTNG_CONSUMER_DISCARDED_EVENTS: + { + uint64_t ret; + struct lttng_ht_iter iter; + struct lttng_ht *ht; + struct lttng_consumer_stream *stream; + uint64_t id = msg.u.discarded_events.session_id; + uint64_t key = msg.u.discarded_events.channel_key; + + DBG("UST consumer discarded events command for session id %" + PRIu64, id); + rcu_read_lock(); + pthread_mutex_lock(&consumer_data.lock); + + ht = consumer_data.stream_list_ht; + + /* + * We only need a reference to the channel, but they are not + * directly indexed, so we just use the first matching stream + * to extract the information we need, we default to 0 if not + * found (no events are dropped if the channel is not yet in + * use). + */ + ret = 0; + cds_lfht_for_each_entry_duplicate(ht->ht, + ht->hash_fct(&id, lttng_ht_seed), + ht->match_fct, &id, + &iter.iter, stream, node_session_id.node) { + if (stream->chan->key == key) { + ret = stream->chan->discarded_events; + break; + } + } + pthread_mutex_unlock(&consumer_data.lock); + rcu_read_unlock(); + + DBG("UST consumer discarded events command for session id %" + PRIu64 ", channel key %" PRIu64, id, key); + + health_code_update(); + + /* Send back returned value to session daemon */ + ret = lttcomm_send_unix_sock(sock, &ret, sizeof(ret)); + if (ret < 0) { + PERROR("send discarded events"); + goto error_fatal; + } + + break; + } + case LTTNG_CONSUMER_LOST_PACKETS: + { + int ret; + uint64_t lost_packets; + struct lttng_ht_iter iter; + struct lttng_ht *ht; + struct lttng_consumer_stream *stream; + uint64_t id = msg.u.lost_packets.session_id; + uint64_t key = msg.u.lost_packets.channel_key; + + DBG("UST consumer lost packets command for session id %" + PRIu64, id); + rcu_read_lock(); + pthread_mutex_lock(&consumer_data.lock); + + ht = consumer_data.stream_list_ht; + + /* + * We only need a reference to the channel, but they are not + * directly indexed, so we just use the first matching stream + * to extract the information we need, we default to 0 if not + * found (no packets lost if the channel is not yet in use). + */ + lost_packets = 0; + cds_lfht_for_each_entry_duplicate(ht->ht, + ht->hash_fct(&id, lttng_ht_seed), + ht->match_fct, &id, + &iter.iter, stream, node_session_id.node) { + if (stream->chan->key == key) { + lost_packets = stream->chan->lost_packets; + break; + } + } + pthread_mutex_unlock(&consumer_data.lock); + rcu_read_unlock(); + + DBG("UST consumer lost packets command for session id %" + PRIu64 ", channel key %" PRIu64, id, key); + + health_code_update(); + + /* Send back returned value to session daemon */ + ret = lttcomm_send_unix_sock(sock, &lost_packets, + sizeof(lost_packets)); + if (ret < 0) { + PERROR("send lost packets"); + goto error_fatal; + } + + break; + } default: break; } @@ -1791,6 +1933,16 @@ int lttng_ustconsumer_get_current_timestamp( return ustctl_get_current_timestamp(stream->ustream, ts); } +int lttng_ustconsumer_get_sequence_number( + struct lttng_consumer_stream *stream, uint64_t *seq) +{ + assert(stream); + assert(stream->ustream); + assert(seq); + + return ustctl_get_sequence_number(stream->ustream, seq); +} + /* * Called when the stream signal the consumer that it has hang up. */ @@ -1813,8 +1965,6 @@ void lttng_ustconsumer_del_channel(struct lttng_consumer_channel *chan) if (chan->switch_timer_enabled == 1) { consumer_timer_switch_stop(chan); } - consumer_metadata_cache_destroy(chan); - ustctl_destroy_channel(chan->uchan); for (i = 0; i < chan->nr_stream_fds; i++) { int ret; @@ -1831,17 +1981,25 @@ void lttng_ustconsumer_del_channel(struct lttng_consumer_channel *chan) } ret = run_as_unlink(shm_path, chan->uid, chan->gid); if (ret) { - errno = -ret; PERROR("unlink %s", shm_path); } } } - free(chan->stream_fds); +} + +void lttng_ustconsumer_free_channel(struct lttng_consumer_channel *chan) +{ + assert(chan); + assert(chan->uchan); + + consumer_metadata_cache_destroy(chan); + ustctl_destroy_channel(chan->uchan); /* Try to rmdir all directories under shm_path root. */ if (chan->root_shm_path[0]) { (void) run_as_recursive_rmdir(chan->root_shm_path, chan->uid, chan->gid); } + free(chan->stream_fds); } void lttng_ustconsumer_del_stream(struct lttng_consumer_stream *stream) @@ -1923,10 +2081,56 @@ static int get_index_values(struct ctf_packet_index *index, } index->stream_id = htobe64(index->stream_id); + ret = ustctl_get_instance_id(ustream, &index->stream_instance_id); + if (ret < 0) { + PERROR("ustctl_get_instance_id"); + goto error; + } + index->stream_instance_id = htobe64(index->stream_instance_id); + + ret = ustctl_get_sequence_number(ustream, &index->packet_seq_num); + if (ret < 0) { + PERROR("ustctl_get_sequence_number"); + goto error; + } + index->packet_seq_num = htobe64(index->packet_seq_num); + error: return ret; } +static +void metadata_stream_reset_cache(struct lttng_consumer_stream *stream, + struct consumer_metadata_cache *cache) +{ + DBG("Metadata stream update to version %" PRIu64, + cache->version); + stream->ust_metadata_pushed = 0; + stream->metadata_version = cache->version; + stream->reset_metadata_flag = 1; +} + +/* + * Check if the version of the metadata stream and metadata cache match. + * If the cache got updated, reset the metadata stream. + * The stream lock and metadata cache lock MUST be held. + * Return 0 on success, a negative value on error. + */ +static +int metadata_stream_check_version(struct lttng_consumer_stream *stream) +{ + int ret = 0; + struct consumer_metadata_cache *cache = stream->chan->metadata_cache; + + if (cache->version == stream->metadata_version) { + goto end; + } + metadata_stream_reset_cache(stream, cache); + +end: + return ret; +} + /* * Write up to one packet from the metadata cache to the channel. * @@ -1940,7 +2144,11 @@ int commit_one_metadata_packet(struct lttng_consumer_stream *stream) int ret; pthread_mutex_lock(&stream->chan->metadata_cache->lock); - if (stream->chan->metadata_cache->contiguous + ret = metadata_stream_check_version(stream); + if (ret < 0) { + goto end; + } + if (stream->chan->metadata_cache->max_offset == stream->ust_metadata_pushed) { ret = 0; goto end; @@ -1948,7 +2156,7 @@ int commit_one_metadata_packet(struct lttng_consumer_stream *stream) write_len = ustctl_write_one_packet_to_channel(stream->chan->uchan, &stream->chan->metadata_cache->data[stream->ust_metadata_pushed], - stream->chan->metadata_cache->contiguous + stream->chan->metadata_cache->max_offset - stream->ust_metadata_pushed); assert(write_len != 0); if (write_len < 0) { @@ -1958,7 +2166,7 @@ int commit_one_metadata_packet(struct lttng_consumer_stream *stream) } stream->ust_metadata_pushed += write_len; - assert(stream->chan->metadata_cache->contiguous >= + assert(stream->chan->metadata_cache->max_offset >= stream->ust_metadata_pushed); ret = write_len; @@ -1972,7 +2180,9 @@ end: * Sync metadata meaning request them to the session daemon and snapshot to the * metadata thread can consumer them. * - * Metadata stream lock MUST be acquired. + * Metadata stream lock is held here, but we need to release it when + * interacting with sessiond, else we cause a deadlock with live + * awaiting on metadata to be pushed out. * * Return 0 if new metadatda is available, EAGAIN if the metadata stream * is empty or a negative value on error. @@ -1986,6 +2196,7 @@ int lttng_ustconsumer_sync_metadata(struct lttng_consumer_local_data *ctx, assert(ctx); assert(metadata); + pthread_mutex_unlock(&metadata->lock); /* * Request metadata from the sessiond, but don't wait for the flush * because we locked the metadata thread. @@ -1994,6 +2205,7 @@ int lttng_ustconsumer_sync_metadata(struct lttng_consumer_local_data *ctx, if (ret < 0) { goto end; } + pthread_mutex_lock(&metadata->lock); ret = commit_one_metadata_packet(metadata); if (ret <= 0) { @@ -2077,6 +2289,61 @@ end: return ret; } +static +int update_stream_stats(struct lttng_consumer_stream *stream) +{ + int ret; + uint64_t seq, discarded; + + ret = ustctl_get_sequence_number(stream->ustream, &seq); + if (ret < 0) { + PERROR("ustctl_get_sequence_number"); + goto end; + } + /* + * Start the sequence when we extract the first packet in case we don't + * start at 0 (for example if a consumer is not connected to the + * session immediately after the beginning). + */ + if (stream->last_sequence_number == -1ULL) { + stream->last_sequence_number = seq; + } else if (seq > stream->last_sequence_number) { + stream->chan->lost_packets += seq - + stream->last_sequence_number - 1; + } else { + /* seq <= last_sequence_number */ + ERR("Sequence number inconsistent : prev = %" PRIu64 + ", current = %" PRIu64, + stream->last_sequence_number, seq); + ret = -1; + goto end; + } + stream->last_sequence_number = seq; + + ret = ustctl_get_events_discarded(stream->ustream, &discarded); + if (ret < 0) { + PERROR("kernctl_get_events_discarded"); + goto end; + } + if (discarded < stream->last_discarded_events) { + /* + * Overflow has occured. We assume only one wrap-around + * has occured. + */ + stream->chan->discarded_events += + (1ULL << (CAA_BITS_PER_LONG - 1)) - + stream->last_discarded_events + discarded; + } else { + stream->chan->discarded_events += discarded - + stream->last_discarded_events; + } + stream->last_discarded_events = discarded; + ret = 0; + +end: + return ret; +} + /* * Read subbuffer from the given stream. * @@ -2160,6 +2427,13 @@ retry: if (ret < 0) { goto end; } + + /* Update the stream's sequence and discarded events count. */ + ret = update_stream_stats(stream); + if (ret < 0) { + PERROR("kernctl_get_events_discarded"); + goto end; + } } else { write_index = 0; } @@ -2220,7 +2494,23 @@ retry: /* * In live, block until all the metadata is sent. */ + pthread_mutex_lock(&stream->metadata_timer_lock); + assert(!stream->missed_metadata_flush); + stream->waiting_on_metadata = true; + pthread_mutex_unlock(&stream->metadata_timer_lock); + err = consumer_stream_sync_metadata(ctx, stream->session_id); + + pthread_mutex_lock(&stream->metadata_timer_lock); + stream->waiting_on_metadata = false; + if (stream->missed_metadata_flush) { + stream->missed_metadata_flush = false; + pthread_mutex_unlock(&stream->metadata_timer_lock); + (void) consumer_flush_ust_index(stream); + } else { + pthread_mutex_unlock(&stream->metadata_timer_lock); + } + if (err < 0) { goto end; } @@ -2301,7 +2591,7 @@ int lttng_ustconsumer_data_pending(struct lttng_consumer_stream *stream) uint64_t contiguous, pushed; /* Ease our life a bit. */ - contiguous = stream->chan->metadata_cache->contiguous; + contiguous = stream->chan->metadata_cache->max_offset; pushed = stream->ust_metadata_pushed; /* @@ -2430,6 +2720,10 @@ void lttng_ustconsumer_close_stream_wakeup(struct lttng_consumer_stream *stream) * function or any of its callees. Timers have a very strict locking * semantic with respect to teardown. Failure to respect this semantic * introduces deadlocks. + * + * DON'T hold the metadata lock when calling this function, else this + * can cause deadlock involving consumer awaiting for metadata to be + * pushed out due to concurrent interaction with the session daemon. */ int lttng_ustconsumer_request_metadata(struct lttng_consumer_local_data *ctx, struct lttng_consumer_channel *channel, int timer, int wait) @@ -2437,7 +2731,7 @@ int lttng_ustconsumer_request_metadata(struct lttng_consumer_local_data *ctx, struct lttcomm_metadata_request_msg request; struct lttcomm_consumer_msg msg; enum lttcomm_return_code ret_code = LTTCOMM_CONSUMERD_SUCCESS; - uint64_t len, key, offset; + uint64_t len, key, offset, version; int ret; assert(channel); @@ -2517,6 +2811,7 @@ int lttng_ustconsumer_request_metadata(struct lttng_consumer_local_data *ctx, len = msg.u.push_metadata.len; key = msg.u.push_metadata.key; offset = msg.u.push_metadata.target_offset; + version = msg.u.push_metadata.version; assert(key == channel->key); if (len == 0) { @@ -2539,7 +2834,7 @@ int lttng_ustconsumer_request_metadata(struct lttng_consumer_local_data *ctx, health_code_update(); ret = lttng_ustconsumer_recv_metadata(ctx->consumer_metadata_socket, - key, offset, len, channel, timer, wait); + key, offset, len, version, channel, timer, wait); if (ret >= 0) { /* * Only send the status msg if the sessiond is alive meaning a positive