Fix: error handling in kernel add_channel
[lttng-tools.git] / src / common / consumer.c
index 25f891af9abf17b3f9e7beb4ed2e75bec671296e..5ca2e7bad68187f7b61b628a28ae0932658a3ccb 100644 (file)
@@ -28,6 +28,7 @@
 #include <sys/types.h>
 #include <unistd.h>
 #include <inttypes.h>
+#include <signal.h>
 
 #include <common/common.h>
 #include <common/utils.h>
@@ -47,6 +48,18 @@ struct lttng_consumer_global_data consumer_data = {
        .type = LTTNG_CONSUMER_UNKNOWN,
 };
 
+enum consumer_channel_action {
+       CONSUMER_CHANNEL_ADD,
+       CONSUMER_CHANNEL_DEL,
+       CONSUMER_CHANNEL_QUIT,
+};
+
+struct consumer_channel_msg {
+       enum consumer_channel_action action;
+       struct lttng_consumer_channel *chan;    /* add */
+       uint64_t key;                           /* del */
+};
+
 /*
  * Flag to inform the polling thread to quit when all fd hung up. Updated by
  * the consumer_thread_receive_fds when it notices that all fds has hung up.
@@ -64,18 +77,59 @@ static struct lttng_ht *metadata_ht;
 static struct lttng_ht *data_ht;
 
 /*
- * Notify a thread pipe to poll back again. This usually means that some global
- * state has changed so we just send back the thread in a poll wait call.
+ * Notify a thread lttng pipe to poll back again. This usually means that some
+ * global state has changed so we just send back the thread in a poll wait
+ * call.
  */
-static void notify_thread_pipe(int wpipe)
+static void notify_thread_lttng_pipe(struct lttng_pipe *pipe)
+{
+       struct lttng_consumer_stream *null_stream = NULL;
+
+       assert(pipe);
+
+       (void) lttng_pipe_write(pipe, &null_stream, sizeof(null_stream));
+}
+
+static void notify_channel_pipe(struct lttng_consumer_local_data *ctx,
+               struct lttng_consumer_channel *chan,
+               uint64_t key,
+               enum consumer_channel_action action)
 {
+       struct consumer_channel_msg msg;
        int ret;
 
+       memset(&msg, 0, sizeof(msg));
+
+       msg.action = action;
+       msg.chan = chan;
        do {
-               struct lttng_consumer_stream *null_stream = NULL;
+               ret = write(ctx->consumer_channel_pipe[1], &msg, sizeof(msg));
+       } while (ret < 0 && errno == EINTR);
+}
 
-               ret = write(wpipe, &null_stream, sizeof(null_stream));
+void notify_thread_del_channel(struct lttng_consumer_local_data *ctx,
+               uint64_t key)
+{
+       notify_channel_pipe(ctx, NULL, key, CONSUMER_CHANNEL_DEL);
+}
+
+static int read_channel_pipe(struct lttng_consumer_local_data *ctx,
+               struct lttng_consumer_channel **chan,
+               uint64_t *key,
+               enum consumer_channel_action *action)
+{
+       struct consumer_channel_msg msg;
+       int ret;
+
+       do {
+               ret = read(ctx->consumer_channel_pipe[0], &msg, sizeof(msg));
        } while (ret < 0 && errno == EINTR);
+       if (ret > 0) {
+               *action = msg.action;
+               *chan = msg.chan;
+               *key = msg.key;
+       }
+       return ret;
 }
 
 /*
@@ -277,9 +331,9 @@ static void cleanup_relayd_ht(void)
                destroy_relayd(relayd);
        }
 
-       lttng_ht_destroy(consumer_data.relayd_ht);
-
        rcu_read_unlock();
+
+       lttng_ht_destroy(consumer_data.relayd_ht);
 }
 
 /*
@@ -353,8 +407,8 @@ static void cleanup_relayd(struct consumer_relayd_sock_pair *relayd,
         * read of this status which happens AFTER receiving this notify.
         */
        if (ctx) {
-               notify_thread_pipe(ctx->consumer_data_pipe[1]);
-               notify_thread_pipe(ctx->consumer_metadata_pipe[1]);
+               notify_thread_lttng_pipe(ctx->consumer_data_pipe);
+               notify_thread_lttng_pipe(ctx->consumer_metadata_pipe);
        }
 }
 
@@ -409,6 +463,13 @@ void consumer_del_stream(struct lttng_consumer_stream *stream,
                                PERROR("munmap");
                        }
                }
+
+               if (stream->wait_fd >= 0) {
+                       ret = close(stream->wait_fd);
+                       if (ret) {
+                               PERROR("close");
+                       }
+               }
                break;
        case LTTNG_CONSUMER32_UST:
        case LTTNG_CONSUMER64_UST:
@@ -425,7 +486,10 @@ void consumer_del_stream(struct lttng_consumer_stream *stream,
        ret = lttng_ht_del(ht, &iter);
        assert(!ret);
 
-       /* Remove node session id from the consumer_data stream ht */
+       iter.iter.node = &stream->node_channel_id.node;
+       ret = lttng_ht_del(consumer_data.stream_per_chan_id_ht, &iter);
+       assert(!ret);
+
        iter.iter.node = &stream->node_session_id.node;
        ret = lttng_ht_del(consumer_data.stream_list_ht, &iter);
        assert(!ret);
@@ -471,8 +535,7 @@ void consumer_del_stream(struct lttng_consumer_stream *stream,
        }
        rcu_read_unlock();
 
-       uatomic_dec(&stream->chan->refcount);
-       if (!uatomic_read(&stream->chan->refcount)
+       if (!uatomic_sub_return(&stream->chan->refcount, 1)
                        && !uatomic_read(&stream->chan->nb_init_stream_left)) {
                free_chan = stream->chan;
        }
@@ -542,11 +605,14 @@ struct lttng_consumer_stream *consumer_allocate_stream(uint64_t channel_key,
        /* Key is always the wait_fd for streams. */
        lttng_ht_node_init_u64(&stream->node, stream->key);
 
+       /* Init node per channel id key */
+       lttng_ht_node_init_u64(&stream->node_channel_id, channel_key);
+
        /* Init session id node with the stream session id */
        lttng_ht_node_init_u64(&stream->node_session_id, stream->session_id);
 
-       DBG3("Allocated stream %s (key %" PRIu64 ", relayd_id %" PRIu64 ", session_id %" PRIu64,
-                       stream->name, stream->key, stream->net_seq_idx, stream->session_id);
+       DBG3("Allocated stream %s (key %" PRIu64 ", chan_key %" PRIu64 " relayd_id %" PRIu64 ", session_id %" PRIu64,
+                       stream->name, stream->key, channel_key, stream->net_seq_idx, stream->session_id);
 
        rcu_read_unlock();
        return stream;
@@ -584,6 +650,9 @@ static int add_stream(struct lttng_consumer_stream *stream,
 
        lttng_ht_add_unique_u64(ht, &stream->node);
 
+       lttng_ht_add_u64(consumer_data.stream_per_chan_id_ht,
+                       &stream->node_channel_id);
+
        /*
         * 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
@@ -597,9 +666,6 @@ static int add_stream(struct lttng_consumer_stream *stream,
                uatomic_inc(&relayd->refcount);
        }
 
-       /* Update channel refcount once added without error(s). */
-       uatomic_inc(&stream->chan->refcount);
-
        /*
         * When nb_init_stream_left reaches 0, we don't need to trigger any action
         * in terms of destroying the associated channel, because the action that
@@ -608,6 +674,8 @@ static int add_stream(struct lttng_consumer_stream *stream,
         * stream.
         */
        if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
+               /* Increment refcount before decrementing nb_init_stream_left */
+               cmm_smp_wmb();
                uatomic_dec(&stream->chan->nb_init_stream_left);
        }
 
@@ -668,6 +736,8 @@ struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
        obj->net_seq_idx = net_seq_idx;
        obj->refcount = 0;
        obj->destroy_flag = 0;
+       obj->control_sock.sock.fd = -1;
+       obj->data_sock.sock.fd = -1;
        lttng_ht_node_init_u64(&obj->node, obj->net_seq_idx);
        pthread_mutex_init(&obj->ctrl_sock_mutex, NULL);
 
@@ -732,7 +802,7 @@ static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
                }
 
                /* Metadata are always sent on the control socket. */
-               outfd = relayd->control_sock.fd;
+               outfd = relayd->control_sock.sock.fd;
        } else {
                /* Set header with stream information */
                data_hdr.stream_id = htobe64(stream->relayd_stream_id);
@@ -757,7 +827,7 @@ static int write_relayd_stream_header(struct lttng_consumer_stream *stream,
                ++stream->next_net_seq_num;
 
                /* Set to go on data socket */
-               outfd = relayd->data_sock.fd;
+               outfd = relayd->data_sock.sock.fd;
        }
 
 error:
@@ -770,14 +840,16 @@ error:
  *
  * On error, return NULL.
  */
-struct lttng_consumer_channel *consumer_allocate_channel(unsigned long key,
+struct lttng_consumer_channel *consumer_allocate_channel(uint64_t key,
                uint64_t session_id,
                const char *pathname,
                const char *name,
                uid_t uid,
                gid_t gid,
                int relayd_id,
-               enum lttng_event_output output)
+               enum lttng_event_output output,
+               uint64_t tracefile_size,
+               uint64_t tracefile_count)
 {
        struct lttng_consumer_channel *channel;
 
@@ -794,6 +866,8 @@ struct lttng_consumer_channel *consumer_allocate_channel(unsigned long key,
        channel->gid = gid;
        channel->relayd_id = relayd_id;
        channel->output = output;
+       channel->tracefile_size = tracefile_size;
+       channel->tracefile_count = tracefile_count;
 
        strncpy(channel->pathname, pathname, sizeof(channel->pathname));
        channel->pathname[sizeof(channel->pathname) - 1] = '\0';
@@ -802,6 +876,9 @@ struct lttng_consumer_channel *consumer_allocate_channel(unsigned long key,
        channel->name[sizeof(channel->name) - 1] = '\0';
 
        lttng_ht_node_init_u64(&channel->node, channel->key);
+
+       channel->wait_fd = -1;
+
        CDS_INIT_LIST_HEAD(&channel->streams.head);
 
        DBG("Allocated channel (key %" PRIu64 ")", channel->key)
@@ -813,7 +890,8 @@ end:
 /*
  * Add a channel to the global list protected by a mutex.
  */
-int consumer_add_channel(struct lttng_consumer_channel *channel)
+int consumer_add_channel(struct lttng_consumer_channel *channel,
+               struct lttng_consumer_local_data *ctx)
 {
        int ret = 0;
        struct lttng_ht_node_u64 *node;
@@ -822,14 +900,13 @@ int consumer_add_channel(struct lttng_consumer_channel *channel)
        pthread_mutex_lock(&consumer_data.lock);
        rcu_read_lock();
 
-       lttng_ht_lookup(consumer_data.channel_ht,
-                       &channel->key, &iter);
+       lttng_ht_lookup(consumer_data.channel_ht, &channel->key, &iter);
        node = lttng_ht_iter_get_node_u64(&iter);
        if (node != NULL) {
                /* Channel already exist. Ignore the insertion */
                ERR("Consumer add channel key %" PRIu64 " already exists!",
                        channel->key);
-               ret = -1;
+               ret = LTTNG_ERR_KERN_CHAN_EXIST;
                goto end;
        }
 
@@ -839,6 +916,10 @@ end:
        rcu_read_unlock();
        pthread_mutex_unlock(&consumer_data.lock);
 
+       if (!ret && channel->wait_fd != -1 &&
+                       channel->metadata_stream == NULL) {
+               notify_channel_pipe(ctx, channel, -1, CONSUMER_CHANNEL_ADD);
+       }
        return ret;
 }
 
@@ -878,7 +959,12 @@ static int update_poll_array(struct lttng_consumer_local_data *ctx,
                                stream->endpoint_status == CONSUMER_ENDPOINT_INACTIVE) {
                        continue;
                }
-               DBG("Active FD %d", stream->wait_fd);
+               /*
+                * This clobbers way too much the debug output. Uncomment that if you
+                * need it for debugging purposes.
+                *
+                * DBG("Active FD %d", stream->wait_fd);
+                */
                (*pollfd)[i].fd = stream->wait_fd;
                (*pollfd)[i].events = POLLIN | POLLPRI;
                local_stream[i] = stream;
@@ -890,7 +976,7 @@ static int update_poll_array(struct lttng_consumer_local_data *ctx,
         * 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_data_pipe[0];
+       (*pollfd)[i].fd = lttng_pipe_get_readfd(ctx->consumer_data_pipe);
        (*pollfd)[i].events = POLLIN | POLLPRI;
        return i;
 }
@@ -979,6 +1065,8 @@ void lttng_consumer_cleanup(void)
 
        cleanup_relayd_ht();
 
+       lttng_ht_destroy(consumer_data.stream_per_chan_id_ht);
+
        /*
         * This HT contains streams that are freed by either the metadata thread or
         * the data thread so we do *nothing* on the hash table and simply destroy
@@ -1063,7 +1151,7 @@ struct lttng_consumer_local_data *lttng_consumer_create(
                int (*recv_stream)(struct lttng_consumer_stream *stream),
                int (*update_stream)(int stream_key, uint32_t state))
 {
-       int ret, i;
+       int ret;
        struct lttng_consumer_local_data *ctx;
 
        assert(consumer_data.type == LTTNG_CONSUMER_UNKNOWN ||
@@ -1077,32 +1165,18 @@ struct lttng_consumer_local_data *lttng_consumer_create(
        }
 
        ctx->consumer_error_socket = -1;
+       ctx->consumer_metadata_socket = -1;
        /* assign the callbacks */
        ctx->on_buffer_ready = buffer_ready;
        ctx->on_recv_channel = recv_channel;
        ctx->on_recv_stream = recv_stream;
        ctx->on_update_stream = update_stream;
 
-       ret = pipe(ctx->consumer_data_pipe);
-       if (ret < 0) {
-               PERROR("Error creating poll pipe");
+       ctx->consumer_data_pipe = lttng_pipe_open(0);
+       if (!ctx->consumer_data_pipe) {
                goto error_poll_pipe;
        }
 
-       /* set read end of the pipe to non-blocking */
-       ret = fcntl(ctx->consumer_data_pipe[0], F_SETFL, O_NONBLOCK);
-       if (ret < 0) {
-               PERROR("fcntl O_NONBLOCK");
-               goto error_poll_fcntl;
-       }
-
-       /* set write end of the pipe to non-blocking */
-       ret = fcntl(ctx->consumer_data_pipe[1], F_SETFL, O_NONBLOCK);
-       if (ret < 0) {
-               PERROR("fcntl O_NONBLOCK");
-               goto error_poll_fcntl;
-       }
-
        ret = pipe(ctx->consumer_should_quit);
        if (ret < 0) {
                PERROR("Error creating recv pipe");
@@ -1115,8 +1189,14 @@ struct lttng_consumer_local_data *lttng_consumer_create(
                goto error_thread_pipe;
        }
 
-       ret = utils_create_pipe(ctx->consumer_metadata_pipe);
+       ret = pipe(ctx->consumer_channel_pipe);
        if (ret < 0) {
+               PERROR("Error creating channel pipe");
+               goto error_channel_pipe;
+       }
+
+       ctx->consumer_metadata_pipe = lttng_pipe_open(0);
+       if (!ctx->consumer_metadata_pipe) {
                goto error_metadata_pipe;
        }
 
@@ -1128,28 +1208,15 @@ struct lttng_consumer_local_data *lttng_consumer_create(
        return ctx;
 
 error_splice_pipe:
-       utils_close_pipe(ctx->consumer_metadata_pipe);
+       lttng_pipe_destroy(ctx->consumer_metadata_pipe);
 error_metadata_pipe:
+       utils_close_pipe(ctx->consumer_channel_pipe);
+error_channel_pipe:
        utils_close_pipe(ctx->consumer_thread_pipe);
 error_thread_pipe:
-       for (i = 0; i < 2; i++) {
-               int err;
-
-               err = close(ctx->consumer_should_quit[i]);
-               if (err) {
-                       PERROR("close");
-               }
-       }
-error_poll_fcntl:
+       utils_close_pipe(ctx->consumer_should_quit);
 error_quit_pipe:
-       for (i = 0; i < 2; i++) {
-               int err;
-
-               err = close(ctx->consumer_data_pipe[i]);
-               if (err) {
-                       PERROR("close");
-               }
-       }
+       lttng_pipe_destroy(ctx->consumer_data_pipe);
 error_poll_pipe:
        free(ctx);
 error:
@@ -1169,30 +1236,15 @@ void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx)
        if (ret) {
                PERROR("close");
        }
-       ret = close(ctx->consumer_thread_pipe[0]);
-       if (ret) {
-               PERROR("close");
-       }
-       ret = close(ctx->consumer_thread_pipe[1]);
-       if (ret) {
-               PERROR("close");
-       }
-       ret = close(ctx->consumer_data_pipe[0]);
-       if (ret) {
-               PERROR("close");
-       }
-       ret = close(ctx->consumer_data_pipe[1]);
-       if (ret) {
-               PERROR("close");
-       }
-       ret = close(ctx->consumer_should_quit[0]);
-       if (ret) {
-               PERROR("close");
-       }
-       ret = close(ctx->consumer_should_quit[1]);
+       ret = close(ctx->consumer_metadata_socket);
        if (ret) {
                PERROR("close");
        }
+       utils_close_pipe(ctx->consumer_thread_pipe);
+       utils_close_pipe(ctx->consumer_channel_pipe);
+       lttng_pipe_destroy(ctx->consumer_data_pipe);
+       lttng_pipe_destroy(ctx->consumer_metadata_pipe);
+       utils_close_pipe(ctx->consumer_should_quit);
        utils_close_pipe(ctx->consumer_splice_metadata_pipe);
 
        unlink(ctx->consumer_command_sock_path);
@@ -1290,6 +1342,7 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
                        goto end;
                }
                ret = lttng_ustctl_get_mmap_read_offset(stream, &mmap_offset);
+
                break;
        default:
                ERR("Unknown consumer_data type");
@@ -1345,6 +1398,26 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap(
        } else {
                /* No streaming, we have to set the len with the full padding */
                len += padding;
+
+               /*
+                * Check if we need to change the tracefile before writing the packet.
+                */
+               if (stream->chan->tracefile_size > 0 &&
+                               (stream->tracefile_size_current + len) >
+                               stream->chan->tracefile_size) {
+                       ret = utils_rotate_stream_file(stream->chan->pathname,
+                                       stream->name, stream->chan->tracefile_size,
+                                       stream->chan->tracefile_count, stream->uid, stream->gid,
+                                       stream->out_fd, &(stream->tracefile_count_current));
+                       if (ret < 0) {
+                               ERR("Rotating output file");
+                               goto end;
+                       }
+                       outfd = stream->out_fd = ret;
+                       /* Reset current size because we just perform a rotation. */
+                       stream->tracefile_size_current = 0;
+               }
+               stream->tracefile_size_current += len;
        }
 
        while (len > 0) {
@@ -1507,6 +1580,26 @@ ssize_t lttng_consumer_on_read_subbuffer_splice(
        } else {
                /* No streaming, we have to set the len with the full padding */
                len += padding;
+
+               /*
+                * Check if we need to change the tracefile before writing the packet.
+                */
+               if (stream->chan->tracefile_size > 0 &&
+                               (stream->tracefile_size_current + len) >
+                               stream->chan->tracefile_size) {
+                       ret = utils_rotate_stream_file(stream->chan->pathname,
+                                       stream->name, stream->chan->tracefile_size,
+                                       stream->chan->tracefile_count, stream->uid, stream->gid,
+                                       stream->out_fd, &(stream->tracefile_count_current));
+                       if (ret < 0) {
+                               ERR("Rotating output file");
+                               goto end;
+                       }
+                       outfd = stream->out_fd = ret;
+                       /* Reset current size because we just perform a rotation. */
+                       stream->tracefile_size_current = 0;
+               }
+               stream->tracefile_size_current += len;
        }
 
        while (len > 0) {
@@ -1790,6 +1883,13 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
                                PERROR("munmap metadata stream");
                        }
                }
+
+               if (stream->wait_fd >= 0) {
+                       ret = close(stream->wait_fd);
+                       if (ret < 0) {
+                               PERROR("close kernel metadata wait_fd");
+                       }
+               }
                break;
        case LTTNG_CONSUMER32_UST:
        case LTTNG_CONSUMER64_UST:
@@ -1806,7 +1906,10 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
        ret = lttng_ht_del(ht, &iter);
        assert(!ret);
 
-       /* Remove node session id from the consumer_data stream ht */
+       iter.iter.node = &stream->node_channel_id.node;
+       ret = lttng_ht_del(consumer_data.stream_per_chan_id_ht, &iter);
+       assert(!ret);
+
        iter.iter.node = &stream->node_session_id.node;
        ret = lttng_ht_del(consumer_data.stream_list_ht, &iter);
        assert(!ret);
@@ -1849,8 +1952,7 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
        rcu_read_unlock();
 
        /* Atomically decrement channel refcount since other threads can use it. */
-       uatomic_dec(&stream->chan->refcount);
-       if (!uatomic_read(&stream->chan->refcount)
+       if (!uatomic_sub_return(&stream->chan->refcount, 1)
                        && !uatomic_read(&stream->chan->nb_init_stream_left)) {
                /* Go for channel deletion! */
                free_chan = stream->chan;
@@ -1920,11 +2022,16 @@ static int add_metadata_stream(struct lttng_consumer_stream *stream,
         * stream.
         */
        if (uatomic_read(&stream->chan->nb_init_stream_left) > 0) {
+               /* Increment refcount before decrementing nb_init_stream_left */
+               cmm_smp_wmb();
                uatomic_dec(&stream->chan->nb_init_stream_left);
        }
 
        lttng_ht_add_unique_u64(ht, &stream->node);
 
+       lttng_ht_add_unique_u64(consumer_data.stream_per_chan_id_ht,
+               &stream->node_channel_id);
+
        /*
         * 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
@@ -2012,7 +2119,7 @@ void *consumer_thread_metadata_poll(void *data)
        metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
        if (!metadata_ht) {
                /* ENOMEM at this point. Better to bail out. */
-               goto error;
+               goto end_ht;
        }
 
        DBG("Thread metadata poll started");
@@ -2021,10 +2128,11 @@ void *consumer_thread_metadata_poll(void *data)
        ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
        if (ret < 0) {
                ERR("Poll set creation failed");
-               goto end;
+               goto end_poll;
        }
 
-       ret = lttng_poll_add(&events, ctx->consumer_metadata_pipe[0], LPOLLIN);
+       ret = lttng_poll_add(&events,
+                       lttng_pipe_get_readfd(ctx->consumer_metadata_pipe), LPOLLIN);
        if (ret < 0) {
                goto end;
        }
@@ -2062,30 +2170,26 @@ restart:
                                continue;
                        }
 
-                       if (pollfd == ctx->consumer_metadata_pipe[0]) {
+                       if (pollfd == lttng_pipe_get_readfd(ctx->consumer_metadata_pipe)) {
                                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]);
-                                       ret = close(ctx->consumer_metadata_pipe[0]);
-                                       if (ret < 0) {
-                                               PERROR("close metadata pipe");
-                                       }
+                                       lttng_poll_del(&events,
+                                                       lttng_pipe_get_readfd(ctx->consumer_metadata_pipe));
+                                       lttng_pipe_read_close(ctx->consumer_metadata_pipe);
                                        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");
+                                       ssize_t pipe_len;
+
+                                       pipe_len = lttng_pipe_read(ctx->consumer_metadata_pipe,
+                                                       &stream, sizeof(stream));
+                                       if (pipe_len < 0) {
+                                               ERR("read metadata stream, ret: %ld", pipe_len);
                                                /*
-                                                * Let's continue here and hope we can still work
-                                                * without stopping the consumer. XXX: Should we?
+                                                * Continue here to handle the rest of the streams.
                                                 */
                                                continue;
                                        }
@@ -2180,10 +2284,11 @@ restart:
 error:
 end:
        DBG("Metadata poll thread exiting");
-       lttng_poll_clean(&events);
 
+       lttng_poll_clean(&events);
+end_poll:
        destroy_stream_ht(metadata_ht);
-
+end_ht:
        rcu_unregister_thread();
        return NULL;
 }
@@ -2291,13 +2396,10 @@ void *consumer_thread_data_poll(void *data)
                        ssize_t pipe_readlen;
 
                        DBG("consumer_data_pipe wake up");
-                       /* Consume 1 byte of pipe data */
-                       do {
-                               pipe_readlen = read(ctx->consumer_data_pipe[0], &new_stream,
-                                               sizeof(new_stream));
-                       } while (pipe_readlen == -1 && errno == EINTR);
+                       pipe_readlen = lttng_pipe_read(ctx->consumer_data_pipe,
+                                       &new_stream, sizeof(new_stream));
                        if (pipe_readlen < 0) {
-                               PERROR("read consumer data pipe");
+                               ERR("Consumer data pipe ret %ld", pipe_readlen);
                                /* Continue so we can at least handle the current stream(s). */
                                continue;
                        }
@@ -2435,17 +2537,294 @@ end:
         * only tracked fd in the poll set. The thread will take care of closing
         * the read side.
         */
-       ret = close(ctx->consumer_metadata_pipe[1]);
+       (void) lttng_pipe_write_close(ctx->consumer_metadata_pipe);
+
+       destroy_data_stream_ht(data_ht);
+
+       rcu_unregister_thread();
+       return NULL;
+}
+
+/*
+ * Close wake-up end of each stream belonging to the channel. This will
+ * allow the poll() on the stream read-side to detect when the
+ * write-side (application) finally closes them.
+ */
+static
+void consumer_close_channel_streams(struct lttng_consumer_channel *channel)
+{
+       struct lttng_ht *ht;
+       struct lttng_consumer_stream *stream;
+       struct lttng_ht_iter iter;
+
+       ht = consumer_data.stream_per_chan_id_ht;
+
+       rcu_read_lock();
+       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) {
+               /*
+                * Protect against teardown with mutex.
+                */
+               pthread_mutex_lock(&stream->lock);
+               if (cds_lfht_is_node_deleted(&stream->node.node)) {
+                       goto next;
+               }
+               switch (consumer_data.type) {
+               case LTTNG_CONSUMER_KERNEL:
+                       break;
+               case LTTNG_CONSUMER32_UST:
+               case LTTNG_CONSUMER64_UST:
+                       /*
+                        * Note: a mutex is taken internally within
+                        * liblttng-ust-ctl to protect timer wakeup_fd
+                        * use from concurrent close.
+                        */
+                       lttng_ustconsumer_close_stream_wakeup(stream);
+                       break;
+               default:
+                       ERR("Unknown consumer_data type");
+                       assert(0);
+               }
+       next:
+               pthread_mutex_unlock(&stream->lock);
+       }
+       rcu_read_unlock();
+}
+
+static void destroy_channel_ht(struct lttng_ht *ht)
+{
+       struct lttng_ht_iter iter;
+       struct lttng_consumer_channel *channel;
+       int ret;
+
+       if (ht == NULL) {
+               return;
+       }
+
+       rcu_read_lock();
+       cds_lfht_for_each_entry(ht->ht, &iter.iter, channel, wait_fd_node.node) {
+               ret = lttng_ht_del(ht, &iter);
+               assert(ret != 0);
+       }
+       rcu_read_unlock();
+
+       lttng_ht_destroy(ht);
+}
+
+/*
+ * This thread polls the channel fds to detect when they are being
+ * closed. It closes all related streams if the channel is detected as
+ * closed. It is currently only used as a shim layer for UST because the
+ * consumerd needs to keep the per-stream wakeup end of pipes open for
+ * periodical flush.
+ */
+void *consumer_thread_channel_poll(void *data)
+{
+       int ret, i, pollfd;
+       uint32_t revents, nb_fd;
+       struct lttng_consumer_channel *chan = NULL;
+       struct lttng_ht_iter iter;
+       struct lttng_ht_node_u64 *node;
+       struct lttng_poll_event events;
+       struct lttng_consumer_local_data *ctx = data;
+       struct lttng_ht *channel_ht;
+
+       rcu_register_thread();
+
+       channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
+       if (!channel_ht) {
+               /* ENOMEM at this point. Better to bail out. */
+               goto end_ht;
+       }
+
+       DBG("Thread channel poll started");
+
+       /* Size is set to 1 for the consumer_channel pipe */
+       ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
        if (ret < 0) {
-               PERROR("close data pipe");
+               ERR("Poll set creation failed");
+               goto end_poll;
        }
 
-       destroy_data_stream_ht(data_ht);
+       ret = lttng_poll_add(&events, ctx->consumer_channel_pipe[0], LPOLLIN);
+       if (ret < 0) {
+               goto end;
+       }
+
+       /* Main loop */
+       DBG("Channel main loop started");
+
+       while (1) {
+               /* Only the channel pipe is set */
+               if (LTTNG_POLL_GETNB(&events) == 0 && consumer_quit == 1) {
+                       goto end;
+               }
+
+restart:
+               DBG("Channel poll wait with %d fd(s)", LTTNG_POLL_GETNB(&events));
+               ret = lttng_poll_wait(&events, -1);
+               DBG("Channel event catched in thread");
+               if (ret < 0) {
+                       if (errno == EINTR) {
+                               ERR("Poll EINTR catched");
+                               goto restart;
+                       }
+                       goto end;
+               }
+
+               nb_fd = ret;
+
+               /* From here, the event is a channel 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_channel_pipe[0]) {
+                               if (revents & (LPOLLERR | LPOLLHUP)) {
+                                       DBG("Channel 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_channel_pipe[0]);
+                                       continue;
+                               } else if (revents & LPOLLIN) {
+                                       enum consumer_channel_action action;
+                                       uint64_t key;
+
+                                       ret = read_channel_pipe(ctx, &chan, &key, &action);
+                                       if (ret <= 0) {
+                                               ERR("Error reading channel pipe");
+                                               continue;
+                                       }
+
+                                       switch (action) {
+                                       case CONSUMER_CHANNEL_ADD:
+                                               DBG("Adding channel %d to poll set",
+                                                       chan->wait_fd);
+
+                                               lttng_ht_node_init_u64(&chan->wait_fd_node,
+                                                       chan->wait_fd);
+                                               lttng_ht_add_unique_u64(channel_ht,
+                                                               &chan->wait_fd_node);
+                                               /* Add channel to the global poll events list */
+                                               lttng_poll_add(&events, chan->wait_fd,
+                                                               LPOLLIN | LPOLLPRI);
+                                               break;
+                                       case CONSUMER_CHANNEL_DEL:
+                                       {
+                                               chan = consumer_find_channel(key);
+                                               if (!chan) {
+                                                       ERR("UST consumer get channel key %" PRIu64 " not found for del channel", key);
+                                                       break;
+                                               }
+                                               lttng_poll_del(&events, chan->wait_fd);
+                                               ret = lttng_ht_del(channel_ht, &iter);
+                                               assert(ret == 0);
+                                               consumer_close_channel_streams(chan);
+
+                                               /*
+                                                * Release our own refcount. Force channel deletion even if
+                                                * streams were not initialized.
+                                                */
+                                               if (!uatomic_sub_return(&chan->refcount, 1)) {
+                                                       consumer_del_channel(chan);
+                                               }
+                                               goto restart;
+                                       }
+                                       case CONSUMER_CHANNEL_QUIT:
+                                               /*
+                                                * Remove the pipe from the poll set and continue the loop
+                                                * since their might be data to consume.
+                                                */
+                                               lttng_poll_del(&events, ctx->consumer_channel_pipe[0]);
+                                               continue;
+                                       default:
+                                               ERR("Unknown action");
+                                               break;
+                                       }
+                               }
+
+                               /* Handle other stream */
+                               continue;
+                       }
+
+                       rcu_read_lock();
+                       {
+                               uint64_t tmp_id = (uint64_t) pollfd;
+
+                               lttng_ht_lookup(channel_ht, &tmp_id, &iter);
+                       }
+                       node = lttng_ht_iter_get_node_u64(&iter);
+                       assert(node);
+
+                       chan = caa_container_of(node, struct lttng_consumer_channel,
+                                       wait_fd_node);
+
+                       /* Check for error event */
+                       if (revents & (LPOLLERR | LPOLLHUP)) {
+                               DBG("Channel fd %d is hup|err.", pollfd);
+
+                               lttng_poll_del(&events, chan->wait_fd);
+                               ret = lttng_ht_del(channel_ht, &iter);
+                               assert(ret == 0);
+                               consumer_close_channel_streams(chan);
+
+                               /* Release our own refcount */
+                               if (!uatomic_sub_return(&chan->refcount, 1)
+                                               && !uatomic_read(&chan->nb_init_stream_left)) {
+                                       consumer_del_channel(chan);
+                               }
+                       }
 
+                       /* Release RCU lock for the channel looked up */
+                       rcu_read_unlock();
+               }
+       }
+
+end:
+       lttng_poll_clean(&events);
+end_poll:
+       destroy_channel_ht(channel_ht);
+end_ht:
+       DBG("Channel poll thread exiting");
        rcu_unregister_thread();
        return NULL;
 }
 
+static int set_metadata_socket(struct lttng_consumer_local_data *ctx,
+               struct pollfd *sockpoll, int client_socket)
+{
+       int ret;
+
+       assert(ctx);
+       assert(sockpoll);
+
+       if (lttng_consumer_poll_socket(sockpoll) < 0) {
+               ret = -1;
+               goto error;
+       }
+       DBG("Metadata connection on client_socket");
+
+       /* Blocking call, waiting for transmission */
+       ctx->consumer_metadata_socket = lttcomm_accept_unix_sock(client_socket);
+       if (ctx->consumer_metadata_socket < 0) {
+               WARN("On accept metadata");
+               ret = -1;
+               goto error;
+       }
+       ret = 0;
+
+error:
+       return ret;
+}
+
 /*
  * This thread listens on the consumerd socket and receives the file
  * descriptors from the session daemon.
@@ -2512,6 +2891,15 @@ void *consumer_thread_sessiond_poll(void *data)
                goto end;
        }
 
+       /*
+        * Setup metadata socket which is the second socket connection on the
+        * command unix socket.
+        */
+       ret = set_metadata_socket(ctx, consumer_sockpoll, client_socket);
+       if (ret < 0) {
+               goto end;
+       }
+
        /* This socket is not useful anymore. */
        ret = close(client_socket);
        if (ret < 0) {
@@ -2568,7 +2956,9 @@ end:
         * Notify the data poll thread to poll back again and test the
         * consumer_quit state that we just set so to quit gracefully.
         */
-       notify_thread_pipe(ctx->consumer_data_pipe[1]);
+       notify_thread_lttng_pipe(ctx->consumer_data_pipe);
+
+       notify_channel_pipe(ctx, NULL, -1, CONSUMER_CHANNEL_QUIT);
 
        /* Cleaning up possibly open sockets. */
        if (sock >= 0) {
@@ -2637,6 +3027,7 @@ void lttng_consumer_init(void)
        consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
        consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
        consumer_data.stream_list_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
+       consumer_data.stream_per_chan_id_ht = lttng_ht_new(0, LTTNG_HT_TYPE_U64);
 }
 
 /*
@@ -2647,12 +3038,15 @@ void lttng_consumer_init(void)
  */
 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,
-               unsigned int sessiond_id)
+               struct pollfd *consumer_sockpoll,
+               struct lttcomm_relayd_sock *relayd_sock, unsigned int sessiond_id)
 {
        int fd = -1, ret = -1, relayd_created = 0;
        enum lttng_error_code ret_code = LTTNG_OK;
-       struct consumer_relayd_sock_pair *relayd;
+       struct consumer_relayd_sock_pair *relayd = NULL;
+
+       assert(ctx);
+       assert(relayd_sock);
 
        DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx);
 
@@ -2703,11 +3097,11 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type,
        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);
+               lttcomm_copy_sock(&relayd->control_sock.sock, &relayd_sock->sock);
+               ret = lttcomm_create_sock(&relayd->control_sock.sock);
                /* Immediately try to close the created socket if valid. */
-               if (relayd->control_sock.fd >= 0) {
-                       if (close(relayd->control_sock.fd)) {
+               if (relayd->control_sock.sock.fd >= 0) {
+                       if (close(relayd->control_sock.sock.fd)) {
                                PERROR("close relayd control socket");
                        }
                }
@@ -2717,7 +3111,10 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type,
                }
 
                /* Assign new file descriptor */
-               relayd->control_sock.fd = fd;
+               relayd->control_sock.sock.fd = fd;
+               /* Assign version values. */
+               relayd->control_sock.major = relayd_sock->major;
+               relayd->control_sock.minor = relayd_sock->minor;
 
                /*
                 * Create a session on the relayd and store the returned id. Lock the
@@ -2745,11 +3142,11 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type,
                break;
        case LTTNG_STREAM_DATA:
                /* Copy received lttcomm socket */
-               lttcomm_copy_sock(&relayd->data_sock, relayd_sock);
-               ret = lttcomm_create_sock(&relayd->data_sock);
+               lttcomm_copy_sock(&relayd->data_sock.sock, &relayd_sock->sock);
+               ret = lttcomm_create_sock(&relayd->data_sock.sock);
                /* Immediately try to close the created socket if valid. */
-               if (relayd->data_sock.fd >= 0) {
-                       if (close(relayd->data_sock.fd)) {
+               if (relayd->data_sock.sock.fd >= 0) {
+                       if (close(relayd->data_sock.sock.fd)) {
                                PERROR("close relayd data socket");
                        }
                }
@@ -2759,7 +3156,10 @@ int consumer_add_relayd_socket(int net_seq_idx, int sock_type,
                }
 
                /* Assign new file descriptor */
-               relayd->data_sock.fd = fd;
+               relayd->data_sock.sock.fd = fd;
+               /* Assign version values. */
+               relayd->data_sock.major = relayd_sock->major;
+               relayd->data_sock.minor = relayd_sock->minor;
                break;
        default:
                ERR("Unknown relayd socket type (%d)", sock_type);
This page took 0.037022 seconds and 4 git commands to generate.