Fix: cleanup inactive FDs in the consumer polling thread
[lttng-tools.git] / src / common / consumer / consumer.c
index 7ce294e815397a9da9f641b18f1dee3019b7916f..814d93e76887c64d1ba8e54ad813d8f909f2850b 100644 (file)
@@ -47,6 +47,7 @@
 #include <common/consumer/consumer-stream.h>
 #include <common/consumer/consumer-testpoint.h>
 #include <common/align.h>
+#include <common/consumer/consumer-metadata-cache.h>
 
 struct lttng_consumer_global_data consumer_data = {
        .stream_count = 0,
@@ -570,7 +571,7 @@ struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
        stream->session_id = session_id;
        stream->monitor = monitor;
        stream->endpoint_status = CONSUMER_ENDPOINT_ACTIVE;
-       stream->index_fd = -1;
+       stream->index_file = NULL;
        stream->last_sequence_number = -1ULL;
        pthread_mutex_init(&stream->lock, NULL);
        pthread_mutex_init(&stream->metadata_timer_lock, NULL);
@@ -1070,7 +1071,7 @@ int consumer_add_channel(struct lttng_consumer_channel *channel,
  */
 static int update_poll_array(struct lttng_consumer_local_data *ctx,
                struct pollfd **pollfd, struct lttng_consumer_stream **local_stream,
-               struct lttng_ht *ht)
+               struct lttng_ht *ht, int *nb_inactive_fd)
 {
        int i = 0;
        struct lttng_ht_iter iter;
@@ -1082,6 +1083,7 @@ static int update_poll_array(struct lttng_consumer_local_data *ctx,
        assert(local_stream);
 
        DBG("Updating poll fd array");
+       *nb_inactive_fd = 0;
        rcu_read_lock();
        cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) {
                /*
@@ -1092,9 +1094,14 @@ static int update_poll_array(struct lttng_consumer_local_data *ctx,
                 * just after the check. However, this is OK since the stream(s) will
                 * be deleted once the thread is notified that the end point state has
                 * changed where this function will be called back again.
+                *
+                * We track the number of inactive FDs because they still need to be
+                * closed by the polling thread after a wakeup on the data_pipe or
+                * metadata_pipe.
                 */
                if (stream->state != LTTNG_CONSUMER_ACTIVE_STREAM ||
                                stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) {
+                       (*nb_inactive_fd)++;
                        continue;
                }
                /*
@@ -1229,6 +1236,11 @@ void lttng_consumer_should_exit(struct lttng_consumer_local_data *ctx)
        DBG("Consumer flag that it should quit");
 }
 
+
+/*
+ * Flush pending writes to trace output disk file.
+ */
+static
 void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream,
                off_t orig_offset)
 {
@@ -1266,8 +1278,8 @@ void lttng_consumer_sync_trace_file(struct lttng_consumer_stream *stream,
        ret = posix_fadvise(outfd, orig_offset - stream->max_sb_size,
                        stream->max_sb_size, POSIX_FADV_DONTNEED);
        if (ret && ret != -ENOSYS) {
-               errno = -ret;
-               PERROR("posix_fadvise");
+               errno = ret;
+               PERROR("posix_fadvise on fd %i", outfd);
        }
 }
 
@@ -1525,7 +1537,6 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
                mmap_base = stream->mmap_base;
                ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset);
                if (ret < 0) {
-                       ret = -errno;
                        PERROR("tracer ctl get_mmap_read_offset");
                        goto end;
                }
@@ -1620,21 +1631,16 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
                        }
                        outfd = stream->out_fd;
 
-                       if (stream->index_fd >= 0) {
-                               ret = close(stream->index_fd);
-                               if (ret < 0) {
-                                       PERROR("Closing index");
-                                       goto end;
-                               }
-                               stream->index_fd = -1;
-                               ret = index_create_file(stream->chan->pathname,
+                       if (stream->index_file) {
+                               lttng_index_file_put(stream->index_file);
+                               stream->index_file = lttng_index_file_create(stream->chan->pathname,
                                                stream->name, stream->uid, stream->gid,
                                                stream->chan->tracefile_size,
-                                               stream->tracefile_count_current);
-                               if (ret < 0) {
+                                               stream->tracefile_count_current,
+                                               CTF_INDEX_MAJOR, CTF_INDEX_MINOR);
+                               if (!stream->index_file) {
                                        goto end;
                                }
-                               stream->index_fd = ret;
                        }
 
                        /* Reset current size because we just perform a rotation. */
@@ -1687,8 +1693,8 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
                lttng_sync_file_range(outfd, stream->out_fd_offset, len,
                                SYNC_FILE_RANGE_WRITE);
                stream->out_fd_offset += len;
+               lttng_consumer_sync_trace_file(stream, orig_offset);
        }
-       lttng_consumer_sync_trace_file(stream, orig_offset);
 
 write_error:
        /*
@@ -1827,22 +1833,16 @@ ssize_t lttng_consumer_on_read_subbuffer_splice(
                        }
                        outfd = stream->out_fd;
 
-                       if (stream->index_fd >= 0) {
-                               ret = close(stream->index_fd);
-                               if (ret < 0) {
-                                       PERROR("Closing index");
-                                       goto end;
-                               }
-                               stream->index_fd = -1;
-                               ret = index_create_file(stream->chan->pathname,
+                       if (stream->index_file) {
+                               lttng_index_file_put(stream->index_file);
+                               stream->index_file = lttng_index_file_create(stream->chan->pathname,
                                                stream->name, stream->uid, stream->gid,
                                                stream->chan->tracefile_size,
-                                               stream->tracefile_count_current);
-                               if (ret < 0) {
-                                       written = ret;
+                                               stream->tracefile_count_current,
+                                               CTF_INDEX_MAJOR, CTF_INDEX_MINOR);
+                               if (!stream->index_file) {
                                        goto end;
                                }
-                               stream->index_fd = ret;
                        }
 
                        /* Reset current size because we just perform a rotation. */
@@ -1917,7 +1917,9 @@ ssize_t lttng_consumer_on_read_subbuffer_splice(
                stream->output_written += ret_splice;
                written += ret_splice;
        }
-       lttng_consumer_sync_trace_file(stream, orig_offset);
+       if (!relayd) {
+               lttng_consumer_sync_trace_file(stream, orig_offset);
+       }
        goto end;
 
 write_error:
@@ -2057,6 +2059,10 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
        pthread_mutex_lock(&consumer_data.lock);
        pthread_mutex_lock(&stream->chan->lock);
        pthread_mutex_lock(&stream->lock);
+       if (stream->chan->metadata_cache) {
+               /* Only applicable to userspace consumers. */
+               pthread_mutex_lock(&stream->chan->metadata_cache->lock);
+       }
 
        /* Remove any reference to that stream. */
        consumer_stream_delete(stream, ht);
@@ -2080,6 +2086,9 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
         */
        stream->chan->metadata_stream = NULL;
 
+       if (stream->chan->metadata_cache) {
+               pthread_mutex_unlock(&stream->chan->metadata_cache->lock);
+       }
        pthread_mutex_unlock(&stream->lock);
        pthread_mutex_unlock(&stream->chan->lock);
        pthread_mutex_unlock(&consumer_data.lock);
@@ -2367,7 +2376,7 @@ restart:
                                        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
+                                        * a negative len, it means an error occurred thus we
                                         * simply remove it from the poll set and free the
                                         * stream.
                                         */
@@ -2394,7 +2403,7 @@ restart:
                                                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
+                                                * a negative len, it means an error occurred thus we
                                                 * simply remove it from the poll set and free the
                                                 * stream.
                                                 */
@@ -2446,6 +2455,8 @@ void *consumer_thread_data_poll(void *data)
        struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
        /* local view of consumer_data.fds_count */
        int nb_fd = 0;
+       /* Number of FDs with CONSUMER_ENDPOINT_INACTIVE but still open. */
+       int nb_inactive_fd = 0;
        struct lttng_consumer_local_data *ctx = data;
        ssize_t len;
 
@@ -2502,7 +2513,7 @@ void *consumer_thread_data_poll(void *data)
                                goto end;
                        }
                        ret = update_poll_array(ctx, &pollfd, local_stream,
-                                       data_ht);
+                                       data_ht, &nb_inactive_fd);
                        if (ret < 0) {
                                ERR("Error in allocating pollfd or local_outfds");
                                lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR);
@@ -2515,7 +2526,7 @@ void *consumer_thread_data_poll(void *data)
                pthread_mutex_unlock(&consumer_data.lock);
 
                /* No FDs and consumer_quit, consumer_cleanup the thread */
-               if (nb_fd == 0 && consumer_quit == 1) {
+               if (nb_fd == 0 && consumer_quit == 1 && nb_inactive_fd == 0) {
                        err = 0;        /* All is OK */
                        goto end;
                }
This page took 0.028568 seconds and 4 git commands to generate.