consumerd: move rotation logic to domain-agnostic read path
[lttng-tools.git] / src / common / consumer / consumer.c
index 1ea682f028e76ebe546b3dc892981bc10d2009a5..b9290976186ea0fe232cd2e6a447307066463be8 100644 (file)
@@ -564,7 +564,9 @@ void consumer_stream_update_channel_attributes(
                        channel->tracefile_size;
 }
 
-struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
+struct lttng_consumer_stream *consumer_allocate_stream(
+               struct lttng_consumer_channel *channel,
+               uint64_t channel_key,
                uint64_t stream_key,
                const char *channel_name,
                uint64_t relayd_id,
@@ -592,6 +594,7 @@ struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
        }
 
        rcu_read_lock();
+       stream->chan = channel;
        stream->key = stream_key;
        stream->trace_chunk = trace_chunk;
        stream->out_fd = -1;
@@ -603,6 +606,7 @@ struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
        stream->endpoint_status = CONSUMER_ENDPOINT_ACTIVE;
        stream->index_file = NULL;
        stream->last_sequence_number = -1ULL;
+       stream->rotate_position = -1ULL;
        pthread_mutex_init(&stream->lock, NULL);
        pthread_mutex_init(&stream->metadata_timer_lock, NULL);
 
@@ -1063,6 +1067,7 @@ struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key,
                uint64_t session_id_per_pid,
                unsigned int monitor,
                unsigned int live_timer_interval,
+               bool is_in_live_session,
                const char *root_shm_path,
                const char *shm_path)
 {
@@ -1094,6 +1099,7 @@ struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key,
        channel->tracefile_count = tracefile_count;
        channel->monitor = monitor;
        channel->live_timer_interval = live_timer_interval;
+       channel->is_live = is_in_live_session;
        pthread_mutex_init(&channel->lock, NULL);
        pthread_mutex_init(&channel->timer_lock, NULL);
 
@@ -1670,18 +1676,19 @@ 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,
+               const struct lttng_buffer_view *buffer,
                unsigned long padding,
                struct ctf_packet_index *index)
 {
-       unsigned long mmap_offset;
-       void *mmap_base;
        ssize_t ret = 0;
        off_t orig_offset = stream->out_fd_offset;
        /* Default is on the disk */
        int outfd = stream->out_fd;
        struct consumer_relayd_sock_pair *relayd = NULL;
        unsigned int relayd_hang_up = 0;
+       const size_t subbuf_content_size = buffer->size - padding;
+       size_t write_len;
 
        /* RCU lock for the relayd pointer */
        rcu_read_lock();
@@ -1697,39 +1704,9 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
                }
        }
 
-       /* get the offset inside the fd to mmap */
-       switch (consumer_data.type) {
-       case LTTNG_CONSUMER_KERNEL:
-               mmap_base = stream->mmap_base;
-               ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
-               if (ret < 0) {
-                       PERROR("tracer ctl get_mmap_read_offset");
-                       goto end;
-               }
-               break;
-       case LTTNG_CONSUMER32_UST:
-       case LTTNG_CONSUMER64_UST:
-               mmap_base = lttng_ustctl_get_mmap_base(stream);
-               if (!mmap_base) {
-                       ERR("read mmap get mmap base for stream %s", stream->name);
-                       ret = -EPERM;
-                       goto end;
-               }
-               ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset);
-               if (ret != 0) {
-                       PERROR("tracer ctl get_mmap_read_offset");
-                       ret = -EINVAL;
-                       goto end;
-               }
-               break;
-       default:
-               ERR("Unknown consumer_data type");
-               assert(0);
-       }
-
        /* Handle stream on the relayd if the output is on the network */
        if (relayd) {
-               unsigned long netlen = len;
+               unsigned long netlen = subbuf_content_size;
 
                /*
                 * Lock the control socket for the complete duration of the function
@@ -1767,10 +1744,10 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
                                goto write_error;
                        }
                }
-       } else {
-               /* No streaming, we have to set the len with the full padding */
-               len += padding;
 
+               write_len = subbuf_content_size;
+       } else {
+               /* No streaming; we have to write the full padding. */
                if (stream->metadata_flag && stream->reset_metadata_flag) {
                        ret = utils_truncate_stream_file(stream->out_fd, 0);
                        if (ret < 0) {
@@ -1784,7 +1761,7 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
                 * Check if we need to change the tracefile before writing the packet.
                 */
                if (stream->chan->tracefile_size > 0 &&
-                               (stream->tracefile_size_current + len) >
+                               (stream->tracefile_size_current + buffer->size) >
                                stream->chan->tracefile_size) {
                        ret = consumer_stream_rotate_output_files(stream);
                        if (ret) {
@@ -1793,19 +1770,21 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
                        outfd = stream->out_fd;
                        orig_offset = 0;
                }
-               stream->tracefile_size_current += len;
+               stream->tracefile_size_current += buffer->size;
                if (index) {
                        index->offset = htobe64(stream->out_fd_offset);
                }
+
+               write_len = buffer->size;
        }
 
        /*
         * This call guarantee that len or less is returned. It's impossible to
         * receive a ret value that is bigger than len.
         */
-       ret = lttng_write(outfd, mmap_base + mmap_offset, len);
-       DBG("Consumer mmap write() ret %zd (len %lu)", ret, len);
-       if (ret < 0 || ((size_t) ret != len)) {
+       ret = lttng_write(outfd, buffer->data, write_len);
+       DBG("Consumer mmap write() ret %zd (len %lu)", ret, write_len);
+       if (ret < 0 || ((size_t) ret != write_len)) {
                /*
                 * Report error to caller if nothing was written else at least send the
                 * amount written.
@@ -1826,7 +1805,8 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
                        DBG("Consumer mmap write detected relayd hang up");
                } else {
                        /* Unhandled error, print it and stop function right now. */
-                       PERROR("Error in write mmap (ret %zd != len %lu)", ret, len);
+                       PERROR("Error in write mmap (ret %zd != write_len %zu)", ret,
+                                       write_len);
                }
                goto write_error;
        }
@@ -1835,9 +1815,9 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
        /* This call is useless on a socket so better save a syscall. */
        if (!relayd) {
                /* This won't block, but will start writeout asynchronously */
-               lttng_sync_file_range(outfd, stream->out_fd_offset, len,
+               lttng_sync_file_range(outfd, stream->out_fd_offset, write_len,
                                SYNC_FILE_RANGE_WRITE);
-               stream->out_fd_offset += len;
+               stream->out_fd_offset += write_len;
                lttng_consumer_sync_trace_file(stream, orig_offset);
        }
 
@@ -3433,6 +3413,7 @@ ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream,
                struct lttng_consumer_local_data *ctx)
 {
        ssize_t ret;
+       int rotation_ret;
 
        pthread_mutex_lock(&stream->chan->lock);
        pthread_mutex_lock(&stream->lock);
@@ -3440,6 +3421,19 @@ ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream,
                pthread_mutex_lock(&stream->metadata_rdv_lock);
        }
 
+       /*
+        * If the stream was flagged to be ready for rotation before we extract
+        * the next packet, rotate it now.
+        */
+       if (stream->rotate_ready) {
+               DBG("Rotate stream before consuming data");
+               ret = lttng_consumer_rotate_stream(ctx, stream);
+               if (ret < 0) {
+                       ERR("Stream rotation error before consuming data");
+                       goto end;
+               }
+       }
+
        switch (consumer_data.type) {
        case LTTNG_CONSUMER_KERNEL:
                ret = lttng_kconsumer_read_subbuffer(stream, ctx);
@@ -3455,13 +3449,38 @@ ssize_t lttng_consumer_read_subbuffer(struct lttng_consumer_stream *stream,
                break;
        }
 
+       if (ret < 0) {
+               goto end;
+       }
+
+       /*
+        * After extracting the packet, we check if the stream is now ready to
+        * be rotated and perform the action immediately.
+        *
+        * Don't overwrite `ret` as callers expect the number of bytes
+        * consumed to be returned on success.
+        */
+       rotation_ret = lttng_consumer_stream_is_rotate_ready(stream);
+       if (rotation_ret == 1) {
+               rotation_ret = lttng_consumer_rotate_stream(ctx, stream);
+               if (rotation_ret < 0) {
+                       ret = rotation_ret;
+                       ERR("Stream rotation error after consuming data");
+                       goto end;
+               }
+       } else if (rotation_ret < 0) {
+               ret = rotation_ret;
+               ERR("Failed to check if stream was ready to rotate after consuming data");
+               goto end;
+       }
+
+end:
        if (stream->metadata_flag) {
                pthread_cond_broadcast(&stream->metadata_rdv);
                pthread_mutex_unlock(&stream->metadata_rdv_lock);
        }
        pthread_mutex_unlock(&stream->lock);
        pthread_mutex_unlock(&stream->chan->lock);
-
        return ret;
 }
 
@@ -3943,15 +3962,23 @@ int consumer_flush_buffer(struct lttng_consumer_stream *stream, int producer_act
 
        switch (consumer_data.type) {
        case LTTNG_CONSUMER_KERNEL:
-               ret = kernctl_buffer_flush(stream->wait_fd);
-               if (ret < 0) {
-                       ERR("Failed to flush kernel stream");
-                       goto end;
+               if (producer_active) {
+                       ret = kernctl_buffer_flush(stream->wait_fd);
+                       if (ret < 0) {
+                               ERR("Failed to flush kernel stream");
+                               goto end;
+                       }
+               } else {
+                       ret = kernctl_buffer_flush_empty(stream->wait_fd);
+                       if (ret < 0) {
+                               ERR("Failed to flush kernel stream");
+                               goto end;
+                       }
                }
                break;
        case LTTNG_CONSUMER32_UST:
        case LTTNG_CONSUMER64_UST:
-               lttng_ustctl_flush_buffer(stream, producer_active);
+               lttng_ustconsumer_flush_buffer(stream, producer_active);
                break;
        default:
                ERR("Unknown consumer_data type");
@@ -4006,7 +4033,7 @@ int lttng_consumer_rotate_channel(struct lttng_consumer_channel *channel,
                        ht->hash_fct(&channel->key, lttng_ht_seed),
                        ht->match_fct, &channel->key, &iter.iter,
                        stream, node_channel_id.node) {
-               unsigned long consumed_pos;
+               unsigned long produced_pos = 0, consumed_pos = 0;
 
                health_code_update();
 
@@ -4019,65 +4046,87 @@ int lttng_consumer_rotate_channel(struct lttng_consumer_channel *channel,
                        rotating_to_new_chunk = false;
                }
 
-               ret = lttng_consumer_sample_snapshot_positions(stream);
-               if (ret < 0) {
-                       ERR("Failed to sample snapshot position during channel rotation");
-                       goto end_unlock_stream;
+               /*
+                * Do not flush an empty packet when rotating from a NULL trace
+                * chunk. The stream has no means to output data, and the prior
+                * rotation which rotated to NULL performed that side-effect already.
+                */
+               if (stream->trace_chunk) {
+                       /*
+                        * For metadata stream, do an active flush, which does not
+                        * produce empty packets. For data streams, empty-flush;
+                        * ensures we have at least one packet in each stream per trace
+                        * chunk, even if no data was produced.
+                        */
+                       ret = consumer_flush_buffer(stream, stream->metadata_flag ? 1 : 0);
+                       if (ret < 0) {
+                               ERR("Failed to flush stream %" PRIu64 " during channel rotation",
+                                               stream->key);
+                               goto end_unlock_stream;
+                       }
                }
 
-               ret = lttng_consumer_get_produced_snapshot(stream,
-                               &stream->rotate_position);
-               if (ret < 0) {
-                       ERR("Failed to sample produced position during channel rotation");
+               ret = lttng_consumer_take_snapshot(stream);
+               if (ret < 0 && ret != -ENODATA && ret != -EAGAIN) {
+                       ERR("Failed to sample snapshot position during channel rotation");
                        goto end_unlock_stream;
                }
+               if (!ret) {
+                       ret = lttng_consumer_get_produced_snapshot(stream,
+                                       &produced_pos);
+                       if (ret < 0) {
+                               ERR("Failed to sample produced position during channel rotation");
+                               goto end_unlock_stream;
+                       }
 
-               lttng_consumer_get_consumed_snapshot(stream,
-                               &consumed_pos);
-               if (consumed_pos == stream->rotate_position) {
+                       ret = lttng_consumer_get_consumed_snapshot(stream,
+                                       &consumed_pos);
+                       if (ret < 0) {
+                               ERR("Failed to sample consumed position during channel rotation");
+                               goto end_unlock_stream;
+                       }
+               }
+               /*
+                * Align produced position on the start-of-packet boundary of the first
+                * packet going into the next trace chunk.
+                */
+               produced_pos = ALIGN_FLOOR(produced_pos, stream->max_sb_size);
+               if (consumed_pos == produced_pos) {
                        stream->rotate_ready = true;
                }
-
                /*
-                * Active flush; has no effect if the production position
-                * is at a packet boundary.
+                * The rotation position is based on the packet_seq_num of the
+                * packet following the last packet that was consumed for this
+                * stream, incremented by the offset between produced and
+                * consumed positions. This rotation position is a lower bound
+                * (inclusive) at which the next trace chunk starts. Since it
+                * is a lower bound, it is OK if the packet_seq_num does not
+                * correspond exactly to the same packet identified by the
+                * consumed_pos, which can happen in overwrite mode.
                 */
-               ret = consumer_flush_buffer(stream, 1);
-               if (ret < 0) {
-                       ERR("Failed to flush stream %" PRIu64 " during channel rotation",
+               if (stream->sequence_number_unavailable) {
+                       /*
+                        * Rotation should never be performed on a session which
+                        * interacts with a pre-2.8 lttng-modules, which does
+                        * not implement packet sequence number.
+                        */
+                       ERR("Failure to rotate stream %" PRIu64 ": sequence number unavailable",
                                        stream->key);
+                       ret = -1;
                        goto end_unlock_stream;
                }
+               stream->rotate_position = stream->last_sequence_number + 1 +
+                               ((produced_pos - consumed_pos) / stream->max_sb_size);
 
                if (!is_local_trace) {
                        /*
                         * The relay daemon control protocol expects a rotation
                         * position as "the sequence number of the first packet
-                        * _after_ the current trace chunk.
-                        *
-                        * At the moment when the positions of the buffers are
-                        * sampled, the production position does not necessarily
-                        * sit at a packet boundary. The 'active' flush
-                        * operation above will push the production position to
-                        * the next packet boundary _if_ it is not already
-                        * sitting at such a boundary.
-                        *
-                        * Assuming a current production position that is not
-                        * on the bound of a packet, the 'target' sequence
-                        * number is
-                        *   (consumed_pos / subbuffer_size) + 1
-                        * Note the '+ 1' to ensure the current packet is
-                        * part of the current trace chunk.
-                        *
-                        * However, if the production position is already at
-                        * a packet boundary, the '+ 1' is not necessary as the
-                        * last packet of the current chunk is already
-                        * 'complete'.
+                        * _after_ the current trace chunk".
                         */
                        const struct relayd_stream_rotation_position position = {
                                .stream_id = stream->relayd_stream_id,
-                               .rotate_at_seq_num = (stream->rotate_position / stream->max_sb_size) +
-                                       !!(stream->rotate_position % stream->max_sb_size),
+                               .rotate_at_seq_num = stream->rotate_position,
                        };
 
                        ret = lttng_dynamic_array_add_element(
@@ -4140,44 +4189,39 @@ end:
  */
 int lttng_consumer_stream_is_rotate_ready(struct lttng_consumer_stream *stream)
 {
-       int ret;
-       unsigned long consumed_pos;
-
-       if (!stream->rotate_position && !stream->rotate_ready) {
-               ret = 0;
-               goto end;
-       }
-
        if (stream->rotate_ready) {
-               ret = 1;
-               goto end;
+               return 1;
        }
 
        /*
-        * If we don't have the rotate_ready flag, check the consumed position
-        * to determine if we need to rotate.
+        * If packet seq num is unavailable, it means we are interacting
+        * with a pre-2.8 lttng-modules which does not implement the
+        * sequence number. Rotation should never be used by sessiond in this
+        * scenario.
         */
-       ret = lttng_consumer_sample_snapshot_positions(stream);
-       if (ret < 0) {
-               ERR("Taking snapshot positions");
-               goto end;
+       if (stream->sequence_number_unavailable) {
+               ERR("Internal error: rotation used on stream %" PRIu64
+                               " with unavailable sequence number",
+                               stream->key);
+               return -1;
        }
 
-       ret = lttng_consumer_get_consumed_snapshot(stream, &consumed_pos);
-       if (ret < 0) {
-               ERR("Consumed snapshot position");
-               goto end;
+       if (stream->rotate_position == -1ULL ||
+                       stream->last_sequence_number == -1ULL) {
+               return 0;
        }
 
-       /* Rotate position not reached yet (with check for overflow). */
-       if ((long) (consumed_pos - stream->rotate_position) < 0) {
-               ret = 0;
-               goto end;
+       /*
+        * Rotate position not reached yet. The stream rotate position is
+        * the position of the next packet belonging to the next trace chunk,
+        * but consumerd considers rotation ready when reaching the last
+        * packet of the current chunk, hence the "rotate_position - 1".
+        */
+       if (stream->last_sequence_number >= stream->rotate_position - 1) {
+               return 1;
        }
-       ret = 1;
 
-end:
-       return ret;
+       return 0;
 }
 
 /*
@@ -4185,7 +4229,7 @@ end:
  */
 void lttng_consumer_reset_stream_rotate_state(struct lttng_consumer_stream *stream)
 {
-       stream->rotate_position = 0;
+       stream->rotate_position = -1ULL;
        stream->rotate_ready = false;
 }
 
This page took 0.046346 seconds and 4 git commands to generate.