X-Git-Url: http://git.liburcu.org/?a=blobdiff_plain;f=src%2Fcommon%2Fust-consumer%2Fust-consumer.c;h=4bb4f8d728794123dc6289e2253371e53b869ae5;hb=9c9abe17599b58b2f8348c23f1dad263e5965cf8;hp=bdabb5e49fcda0e44d6c65018efc0b4f296c6730;hpb=601262d65a9ccd90198558639ef2a73cda4230e1;p=lttng-tools.git diff --git a/src/common/ust-consumer/ust-consumer.c b/src/common/ust-consumer/ust-consumer.c index bdabb5e49..4bb4f8d72 100644 --- a/src/common/ust-consumer/ust-consumer.c +++ b/src/common/ust-consumer/ust-consumer.c @@ -37,6 +37,7 @@ #include #include #include +#include #include #include #include @@ -78,6 +79,7 @@ static void destroy_channel(struct lttng_consumer_channel *channel) */ if (channel->uchan) { lttng_ustconsumer_del_channel(channel); + lttng_ustconsumer_free_channel(channel); } free(channel); } @@ -326,8 +328,12 @@ 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)); + } } } @@ -623,49 +629,57 @@ static int flush_channel(uint64_t chan_key) health_code_update(); - ustctl_flush_buffer(stream->ustream, 1); + pthread_mutex_lock(&stream->lock); + if (!stream->quiescent) { + ustctl_flush_buffer(stream->ustream, 0); + stream->quiescent = true; + } + pthread_mutex_unlock(&stream->lock); } error: rcu_read_unlock(); return ret; } + /* - * Close metadata stream wakeup_fd using the given key to retrieve the channel. - * RCU read side lock MUST be acquired before calling this function. - * - * NOTE: This function does NOT take any channel nor stream lock. + * Clear quiescent state from channel's streams using the given key to + * retrieve the channel. * - * Return 0 on success else LTTng error code. + * Return 0 on success else an LTTng error code. */ -static int _close_metadata(struct lttng_consumer_channel *channel) +static int clear_quiescent_channel(uint64_t chan_key) { - int ret = LTTCOMM_CONSUMERD_SUCCESS; + int ret = 0; + struct lttng_consumer_channel *channel; + struct lttng_consumer_stream *stream; + struct lttng_ht *ht; + struct lttng_ht_iter iter; - assert(channel); - assert(channel->type == CONSUMER_CHANNEL_TYPE_METADATA); + DBG("UST consumer clear quiescent channel key %" PRIu64, chan_key); - if (channel->switch_timer_enabled == 1) { - DBG("Deleting timer on metadata channel"); - consumer_timer_switch_stop(channel); + rcu_read_lock(); + channel = consumer_find_channel(chan_key); + if (!channel) { + ERR("UST consumer clear quiescent channel %" PRIu64 " not found", chan_key); + ret = LTTNG_ERR_UST_CHAN_NOT_FOUND; + goto error; } - if (channel->metadata_stream) { - ret = ustctl_stream_close_wakeup_fd(channel->metadata_stream->ustream); - if (ret < 0) { - ERR("UST consumer unable to close fd of metadata (ret: %d)", ret); - ret = LTTCOMM_CONSUMERD_ERROR_METADATA; - } + ht = consumer_data.stream_per_chan_id_ht; - if (channel->monitor) { - /* Close the read-side in consumer_del_metadata_stream */ - ret = close(channel->metadata_stream->ust_metadata_poll_pipe[1]); - if (ret < 0) { - PERROR("Close UST metadata write-side poll pipe"); - ret = LTTCOMM_CONSUMERD_ERROR_METADATA; - } - } - } + /* For each stream of the channel id, clear quiescent state. */ + 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) { + + health_code_update(); + pthread_mutex_lock(&stream->lock); + stream->quiescent = false; + pthread_mutex_unlock(&stream->lock); + } +error: + rcu_read_unlock(); return ret; } @@ -702,7 +716,7 @@ static int close_metadata(uint64_t chan_key) goto error_unlock; } - ret = _close_metadata(channel); + lttng_ustconsumer_close_metadata(channel); error_unlock: pthread_mutex_unlock(&channel->lock); @@ -788,8 +802,9 @@ error: * the stream is still in the local stream list of the channel. This call * will make sure to clean that list. */ - cds_list_del(&metadata->metadata_stream->send_node); consumer_stream_destroy(metadata->metadata_stream, NULL); + cds_list_del(&metadata->metadata_stream->send_node); + metadata->metadata_stream = NULL; error_no_stream: end: return ret; @@ -881,8 +896,8 @@ error_stream: * Clean up the stream completly because the next snapshot will use a new * metadata stream. */ - cds_list_del(&metadata_stream->send_node); consumer_stream_destroy(metadata_stream, NULL); + cds_list_del(&metadata_stream->send_node); metadata_channel->metadata_stream = NULL; error: @@ -896,7 +911,7 @@ error: * Returns 0 on success, < 0 on error */ static int snapshot_channel(uint64_t key, char *path, uint64_t relayd_id, - uint64_t max_stream_size, struct lttng_consumer_local_data *ctx) + uint64_t nb_packets_per_stream, struct lttng_consumer_local_data *ctx) { int ret; unsigned use_relayd = 0; @@ -978,12 +993,13 @@ static int snapshot_channel(uint64_t key, char *path, uint64_t relayd_id, /* * The original value is sent back if max stream size is larger than - * the possible size of the snapshot. Also, we asume that the session + * the possible size of the snapshot. Also, we assume that the session * daemon should never send a maximum stream size that is lower than * subbuffer size. */ - consumed_pos = consumer_get_consumed_maxsize(consumed_pos, - produced_pos, max_stream_size); + consumed_pos = consumer_get_consume_start_pos(consumed_pos, + produced_pos, nb_packets_per_stream, + stream->max_sb_size); while (consumed_pos < produced_pos) { ssize_t read_len; @@ -1060,7 +1076,12 @@ 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, @@ -1154,17 +1175,8 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx, health_code_update(); - if (msg.cmd_type == LTTNG_CONSUMER_STOP) { - /* - * Notify the session daemon that the command is completed. - * - * On transport layer error, the function call will print an error - * message so handling the returned code is a bit useless since we - * return an error code anyway. - */ - (void) consumer_send_status_msg(sock, ret_code); - return -ENOENT; - } + /* deprecated */ + assert(msg.cmd_type != LTTNG_CONSUMER_STOP); health_code_update(); @@ -1192,7 +1204,7 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx, relayd = consumer_find_relayd(index); if (relayd == NULL) { DBG("Unable to find relayd %" PRIu64, index); - ret_code = LTTNG_ERR_NO_CONSUMER; + ret_code = LTTCOMM_CONSUMERD_RELAYD_FAIL; } /* @@ -1373,7 +1385,7 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx, channel = consumer_find_channel(key); if (!channel) { ERR("UST consumer get channel key %" PRIu64 " not found", key); - ret_code = LTTNG_ERR_UST_CHAN_NOT_FOUND; + ret_code = LTTCOMM_CONSUMERD_CHAN_NOT_FOUND; goto end_msg_sessiond; } @@ -1454,6 +1466,18 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx, goto end_msg_sessiond; } + case LTTNG_CONSUMER_CLEAR_QUIESCENT_CHANNEL: + { + int ret; + + ret = clear_quiescent_channel( + msg.u.clear_quiescent_channel.key); + if (ret != 0) { + ret_code = ret; + } + + goto end_msg_sessiond; + } case LTTNG_CONSUMER_PUSH_METADATA: { int ret; @@ -1481,6 +1505,15 @@ 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) { @@ -1494,7 +1527,7 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx, health_poll_entry(); ret = lttng_consumer_poll_socket(consumer_sockpoll); health_poll_exit(); - if (ret < 0) { + if (ret) { goto error_fatal; } @@ -1529,17 +1562,17 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx, ctx); if (ret < 0) { ERR("Snapshot metadata failed"); - ret_code = LTTNG_ERR_UST_META_FAIL; + ret_code = LTTCOMM_CONSUMERD_ERROR_METADATA; } } else { ret = snapshot_channel(msg.u.snapshot_channel.key, msg.u.snapshot_channel.pathname, msg.u.snapshot_channel.relayd_id, - msg.u.snapshot_channel.max_stream_size, + msg.u.snapshot_channel.nb_packets_per_stream, ctx); if (ret < 0) { ERR("Snapshot channel failed"); - ret_code = LTTNG_ERR_UST_CHAN_FAIL; + ret_code = LTTCOMM_CONSUMERD_CHANNEL_FAIL; } } @@ -1695,14 +1728,19 @@ int lttng_ustconsumer_get_current_timestamp( } /* - * Called when the stream signal the consumer that it has hang up. + * Called when the stream signals the consumer that it has hung up. */ void lttng_ustconsumer_on_stream_hangup(struct lttng_consumer_stream *stream) { assert(stream); assert(stream->ustream); - ustctl_flush_buffer(stream->ustream, 0); + pthread_mutex_lock(&stream->lock); + if (!stream->quiescent) { + ustctl_flush_buffer(stream->ustream, 0); + stream->quiescent = true; + } + pthread_mutex_unlock(&stream->lock); stream->hangup_flush_done = 1; } @@ -1714,6 +1752,13 @@ void lttng_ustconsumer_del_channel(struct lttng_consumer_channel *chan) if (chan->switch_timer_enabled == 1) { consumer_timer_switch_stop(chan); } +} + +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); } @@ -1729,6 +1774,22 @@ void lttng_ustconsumer_del_stream(struct lttng_consumer_stream *stream) ustctl_destroy_stream(stream->ustream); } +int lttng_ustconsumer_get_wakeup_fd(struct lttng_consumer_stream *stream) +{ + assert(stream); + assert(stream->ustream); + + return ustctl_stream_get_wakeup_fd(stream->ustream); +} + +int lttng_ustconsumer_close_wakeup_fd(struct lttng_consumer_stream *stream) +{ + assert(stream); + assert(stream->ustream); + + return ustctl_stream_close_wakeup_fd(stream->ustream); +} + /* * Populate index values of a UST stream. Values are set in big endian order. * @@ -1798,7 +1859,7 @@ 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 + if (stream->chan->metadata_cache->max_offset == stream->ust_metadata_pushed) { ret = 0; goto end; @@ -1806,7 +1867,7 @@ 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) { @@ -1816,7 +1877,7 @@ 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; @@ -1830,7 +1891,9 @@ 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. @@ -1844,6 +1907,7 @@ 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. @@ -1852,6 +1916,7 @@ 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) { @@ -1884,6 +1949,57 @@ end: return ret; } +/* + * Return 0 on success else a negative value. + */ +static int notify_if_more_data(struct lttng_consumer_stream *stream, + struct lttng_consumer_local_data *ctx) +{ + int ret; + struct ustctl_consumer_stream *ustream; + + assert(stream); + assert(ctx); + + ustream = stream->ustream; + + /* + * First, we are going to check if there is a new subbuffer available + * before reading the stream wait_fd. + */ + /* Get the next subbuffer */ + ret = ustctl_get_next_subbuf(ustream); + if (ret) { + /* No more data found, flag the stream. */ + stream->has_data = 0; + ret = 0; + goto end; + } + + ret = ustctl_put_subbuf(ustream); + assert(!ret); + + /* This stream still has data. Flag it and wake up the data thread. */ + stream->has_data = 1; + + if (stream->monitor && !stream->hangup_flush_done && !ctx->has_wakeup) { + ssize_t writelen; + + writelen = lttng_pipe_write(ctx->consumer_wakeup_pipe, "!", 1); + if (writelen < 0 && errno != EAGAIN && errno != EWOULDBLOCK) { + ret = writelen; + goto end; + } + + /* The wake up pipe has been notified. */ + ctx->has_wakeup = 1; + } + ret = 0; + +end: + return ret; +} + /* * Read subbuffer from the given stream. * @@ -1897,7 +2013,6 @@ int lttng_ustconsumer_read_subbuffer(struct lttng_consumer_stream *stream, unsigned long len, subbuf_size, padding; int err, write_index = 1; long ret = 0; - char dummy; struct ustctl_consumer_stream *ustream; struct ctf_packet_index index; @@ -1912,11 +2027,17 @@ int lttng_ustconsumer_read_subbuffer(struct lttng_consumer_stream *stream, ustream = stream->ustream; /* - * We can consume the 1 byte written into the wait_fd by UST. - * Don't trigger error if we cannot read this one byte (read - * returns 0), or if the error is EAGAIN or EWOULDBLOCK. + * We can consume the 1 byte written into the wait_fd by UST. Don't trigger + * error if we cannot read this one byte (read returns 0), or if the error + * is EAGAIN or EWOULDBLOCK. + * + * This is only done when the stream is monitored by a thread, before the + * flush is done after a hangup and if the stream is not flagged with data + * since there might be nothing to consume in the wait fd but still have + * data available flagged by the consumer wake up pipe. */ - if (stream->monitor && !stream->hangup_flush_done) { + if (stream->monitor && !stream->hangup_flush_done && !stream->has_data) { + char dummy; ssize_t readlen; readlen = lttng_read(stream->wait_fd, &dummy, 1); @@ -2002,6 +2123,17 @@ retry: err = ustctl_put_next_subbuf(ustream); assert(err == 0); + /* + * This will consumer the byte on the wait_fd if and only if there is not + * next subbuffer to be acquired. + */ + if (!stream->metadata_flag) { + ret = notify_if_more_data(stream, ctx); + if (ret < 0) { + goto end; + } + } + /* Write index if needed. */ if (!write_index) { goto end; @@ -2011,7 +2143,23 @@ 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; } @@ -2092,7 +2240,7 @@ 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; /* @@ -2107,7 +2255,7 @@ int lttng_ustconsumer_data_pending(struct lttng_consumer_stream *stream) */ DBG("UST consumer metadata pending check: contiguous %" PRIu64 " vs pushed %" PRIu64, contiguous, pushed); - assert(((int64_t) contiguous - pushed) >= 0); + assert(((int64_t) (contiguous - pushed)) >= 0); if ((contiguous != pushed) || (((int64_t) contiguous - pushed) > 0 || contiguous == 0)) { ret = 1; /* Data is pending */ @@ -2134,6 +2282,45 @@ end: return ret; } +/* + * Stop a given metadata channel timer if enabled and close the wait fd which + * is the poll pipe of the metadata stream. + * + * This MUST be called with the metadata channel acquired. + */ +void lttng_ustconsumer_close_metadata(struct lttng_consumer_channel *metadata) +{ + int ret; + + assert(metadata); + assert(metadata->type == CONSUMER_CHANNEL_TYPE_METADATA); + + DBG("Closing metadata channel key %" PRIu64, metadata->key); + + if (metadata->switch_timer_enabled == 1) { + consumer_timer_switch_stop(metadata); + } + + if (!metadata->metadata_stream) { + goto end; + } + + /* + * Closing write side so the thread monitoring the stream wakes up if any + * and clean the metadata stream. + */ + if (metadata->metadata_stream->ust_metadata_poll_pipe[1] >= 0) { + ret = close(metadata->metadata_stream->ust_metadata_poll_pipe[1]); + if (ret < 0) { + PERROR("closing metadata pipe write side"); + } + metadata->metadata_stream->ust_metadata_poll_pipe[1] = -1; + } + +end: + return; +} + /* * Close every metadata stream wait fd of the metadata hash table. This * function MUST be used very carefully so not to run into a race between the @@ -2143,7 +2330,7 @@ end: * producer so calling this is safe because we are assured that no state change * can occur in the metadata thread for the streams in the hash table. */ -void lttng_ustconsumer_close_metadata(struct lttng_ht *metadata_ht) +void lttng_ustconsumer_close_all_metadata(struct lttng_ht *metadata_ht) { struct lttng_ht_iter iter; struct lttng_consumer_stream *stream; @@ -2160,13 +2347,7 @@ void lttng_ustconsumer_close_metadata(struct lttng_ht *metadata_ht) health_code_update(); pthread_mutex_lock(&stream->chan->lock); - /* - * Whatever returned value, we must continue to try to close everything - * so ignore it. - */ - (void) _close_metadata(stream->chan); - DBG("Metadata wait fd %d and poll pipe fd %d closed", stream->wait_fd, - stream->ust_metadata_poll_pipe[1]); + lttng_ustconsumer_close_metadata(stream->chan); pthread_mutex_unlock(&stream->chan->lock); } @@ -2188,6 +2369,10 @@ 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) @@ -2201,6 +2386,8 @@ int lttng_ustconsumer_request_metadata(struct lttng_consumer_local_data *ctx, assert(channel); assert(channel->metadata_cache); + memset(&request, 0, sizeof(request)); + /* send the metadata request to sessiond */ switch (consumer_data.type) { case LTTNG_CONSUMER64_UST: @@ -2311,3 +2498,15 @@ end: pthread_mutex_unlock(&ctx->metadata_socket_lock); return ret; } + +/* + * Return the ustctl call for the get stream id. + */ +int lttng_ustconsumer_get_stream_id(struct lttng_consumer_stream *stream, + uint64_t *stream_id) +{ + assert(stream); + assert(stream_id); + + return ustctl_get_stream_id(stream->ustream, stream_id); +}