X-Git-Url: https://git.lttng.org/?p=lttng-tools.git;a=blobdiff_plain;f=src%2Fcommon%2Fconsumer.c;h=a2980e77d29c33b57c4a12889cb82db019956fa0;hp=63d0d65ee3157b05ef1113336185535ea23a9dbe;hb=d09e1200ec761aef77c721bd648a299eefcc8565;hpb=173af62f4804133d4a7f45e34b6f72126f3eca5f diff --git a/src/common/consumer.c b/src/common/consumer.c index 63d0d65ee..a2980e77d 100644 --- a/src/common/consumer.c +++ b/src/common/consumer.c @@ -27,8 +27,11 @@ #include #include #include +#include #include +#include +#include #include #include #include @@ -172,7 +175,7 @@ static void consumer_rcu_free_relayd(struct rcu_head *head) * * This function MUST be called with the consumer_data lock acquired. */ -void consumer_destroy_relayd(struct consumer_relayd_sock_pair *relayd) +static void destroy_relayd(struct consumer_relayd_sock_pair *relayd) { int ret; struct lttng_ht_iter iter; @@ -200,6 +203,25 @@ void consumer_destroy_relayd(struct consumer_relayd_sock_pair *relayd) call_rcu(&relayd->node.head, consumer_rcu_free_relayd); } +/* + * Flag a relayd socket pair for destruction. Destroy it if the refcount + * reaches zero. + * + * RCU read side lock MUST be aquired before calling this function. + */ +void consumer_flag_relayd_for_destroy(struct consumer_relayd_sock_pair *relayd) +{ + assert(relayd); + + /* Set destroy flag for this object */ + uatomic_set(&relayd->destroy_flag, 1); + + /* Destroy the relayd if refcount is 0 */ + if (uatomic_read(&relayd->refcount) == 0) { + destroy_relayd(relayd); + } +} + /* * Remove a stream from the global list protected by a mutex. This * function is also responsible for freeing its data structures. @@ -274,18 +296,25 @@ void consumer_del_stream(struct lttng_consumer_stream *stream) uatomic_dec(&relayd->refcount); assert(uatomic_read(&relayd->refcount) >= 0); + /* Closing streams requires to lock the control socket. */ + pthread_mutex_lock(&relayd->ctrl_sock_mutex); ret = relayd_send_close_stream(&relayd->control_sock, stream->relayd_stream_id, stream->next_net_seq_num - 1); + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); if (ret < 0) { - ERR("Unable to close stream on the relayd. Continuing"); - /* Continue here. There is nothing we can do for the relayd.*/ + DBG("Unable to close stream on the relayd. Continuing"); + /* + * Continue here. There is nothing we can do for the relayd. + * Chances are that the relayd has closed the socket so we just + * continue cleaning up. + */ } /* Both conditions are met, we destroy the relayd. */ if (uatomic_read(&relayd->refcount) == 0 && uatomic_read(&relayd->destroy_flag)) { - consumer_destroy_relayd(relayd); + destroy_relayd(relayd); } } rcu_read_unlock(); @@ -413,20 +442,6 @@ int consumer_add_stream(struct lttng_consumer_stream *stream) consumer_data.stream_count++; consumer_data.need_update = 1; - switch (consumer_data.type) { - case LTTNG_CONSUMER_KERNEL: - break; - case LTTNG_CONSUMER32_UST: - case LTTNG_CONSUMER64_UST: - /* Streams are in CPU number order (we rely on this) */ - stream->cpu = stream->chan->nr_streams++; - break; - default: - ERR("Unknown consumer_data type"); - assert(0); - goto end; - } - end: pthread_mutex_unlock(&consumer_data.lock); @@ -434,9 +449,10 @@ end: } /* - * Add relayd socket to global consumer data hashtable. + * Add relayd socket to global consumer data hashtable. RCU read side lock MUST + * be acquired before calling this. */ -int consumer_add_relayd(struct consumer_relayd_sock_pair *relayd) +static int add_relayd(struct consumer_relayd_sock_pair *relayd) { int ret = 0; struct lttng_ht_node_ulong *node; @@ -447,20 +463,15 @@ int consumer_add_relayd(struct consumer_relayd_sock_pair *relayd) goto end; } - rcu_read_lock(); - lttng_ht_lookup(consumer_data.relayd_ht, (void *)((unsigned long) relayd->net_seq_idx), &iter); node = lttng_ht_iter_get_node_ulong(&iter); if (node != NULL) { - rcu_read_unlock(); /* Relayd already exist. Ignore the insertion */ goto end; } lttng_ht_add_unique_ulong(consumer_data.relayd_ht, &relayd->node); - rcu_read_unlock(); - end: return ret; } @@ -529,28 +540,20 @@ error: * * Return destination file descriptor or negative value on error. */ -int consumer_handle_stream_before_relayd(struct lttng_consumer_stream *stream, - size_t data_size) +static int write_relayd_stream_header(struct lttng_consumer_stream *stream, + size_t data_size, unsigned long padding, + struct consumer_relayd_sock_pair *relayd) { int outfd = -1, ret; - struct consumer_relayd_sock_pair *relayd; struct lttcomm_relayd_data_hdr data_hdr; /* Safety net */ assert(stream); + assert(relayd); /* Reset data header */ memset(&data_hdr, 0, sizeof(data_hdr)); - rcu_read_lock(); - /* Get relayd reference of the stream. */ - relayd = consumer_find_relayd(stream->net_seq_idx); - if (relayd == NULL) { - /* Stream is either local or corrupted */ - goto error; - } - - DBG("Consumer found relayd socks with index %d", stream->net_seq_idx); if (stream->metadata_flag) { /* Caller MUST acquire the relayd control socket lock */ ret = relayd_send_metadata(&relayd->control_sock, data_size); @@ -564,6 +567,7 @@ int consumer_handle_stream_before_relayd(struct lttng_consumer_stream *stream, /* Set header with stream information */ data_hdr.stream_id = htobe64(stream->relayd_stream_id); data_hdr.data_size = htobe32(data_size); + data_hdr.padding_size = htobe32(padding); data_hdr.net_seq_num = htobe64(stream->next_net_seq_num++); /* Other fields are zeroed previously */ @@ -578,7 +582,6 @@ int consumer_handle_stream_before_relayd(struct lttng_consumer_stream *stream, } error: - rcu_read_unlock(); return outfd; } @@ -684,7 +687,6 @@ struct lttng_consumer_channel *consumer_allocate_channel( channel->mmap_len = mmap_len; channel->max_sb_size = max_sb_size; channel->refcount = 0; - channel->nr_streams = 0; lttng_ht_node_init_ulong(&channel->node, channel->key); switch (consumer_data.type) { @@ -752,8 +754,7 @@ end: */ int consumer_update_poll_array( struct lttng_consumer_local_data *ctx, struct pollfd **pollfd, - struct lttng_consumer_stream **local_stream, - struct lttng_ht *metadata_ht) + struct lttng_consumer_stream **local_stream) { int i = 0; struct lttng_ht_iter iter; @@ -769,10 +770,6 @@ int consumer_update_poll_array( DBG("Active FD %d", stream->wait_fd); (*pollfd)[i].fd = stream->wait_fd; (*pollfd)[i].events = POLLIN | POLLPRI; - if (stream->metadata_flag && metadata_ht) { - lttng_ht_add_unique_ulong(metadata_ht, &stream->waitfd_node); - DBG("Active FD added to metadata hash table"); - } local_stream[i] = stream; i++; } @@ -1011,9 +1008,22 @@ struct lttng_consumer_local_data *lttng_consumer_create( goto error_thread_pipe; } - return ctx; + ret = utils_create_pipe(ctx->consumer_metadata_pipe); + if (ret < 0) { + goto error_metadata_pipe; + } + ret = utils_create_pipe(ctx->consumer_splice_metadata_pipe); + if (ret < 0) { + goto error_splice_pipe; + } + + return ctx; +error_splice_pipe: + utils_close_pipe(ctx->consumer_metadata_pipe); +error_metadata_pipe: + utils_close_pipe(ctx->consumer_thread_pipe); error_thread_pipe: for (i = 0; i < 2; i++) { int err; @@ -1074,31 +1084,165 @@ void lttng_consumer_destroy(struct lttng_consumer_local_data *ctx) if (ret) { PERROR("close"); } + utils_close_pipe(ctx->consumer_splice_metadata_pipe); + unlink(ctx->consumer_command_sock_path); free(ctx); } /* - * Mmap the ring buffer, read it and write the data to the tracefile. + * Write the metadata stream id on the specified file descriptor. + */ +static int write_relayd_metadata_id(int fd, + struct lttng_consumer_stream *stream, + struct consumer_relayd_sock_pair *relayd, + unsigned long padding) +{ + int ret; + struct lttcomm_relayd_metadata_payload hdr; + + hdr.stream_id = htobe64(stream->relayd_stream_id); + hdr.padding_size = htobe32(padding); + do { + ret = write(fd, (void *) &hdr, sizeof(hdr)); + } while (ret < 0 && errno == EINTR); + if (ret < 0) { + PERROR("write metadata stream id"); + goto end; + } + DBG("Metadata stream id %" PRIu64 " with padding %lu written before data", + stream->relayd_stream_id, padding); + +end: + return ret; +} + +/* + * Mmap the ring buffer, read it and write the data to the tracefile. This is a + * core function for writing trace buffers to either the local filesystem or + * the network. + * + * Careful review MUST be put if any changes occur! * * Returns the number of bytes written */ 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, unsigned long len, + unsigned long padding) { + unsigned long mmap_offset; + ssize_t ret = 0, written = 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; + + /* RCU lock for the relayd pointer */ + rcu_read_lock(); + + /* Flag that the current stream if set for network streaming. */ + if (stream->net_seq_idx != -1) { + relayd = consumer_find_relayd(stream->net_seq_idx); + if (relayd == NULL) { + goto end; + } + } + + /* get the offset inside the fd to mmap */ switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: - return lttng_kconsumer_on_read_subbuffer_mmap(ctx, stream, len); + ret = kernctl_get_mmap_read_offset(stream->wait_fd, &mmap_offset); + break; case LTTNG_CONSUMER32_UST: case LTTNG_CONSUMER64_UST: - return lttng_ustconsumer_on_read_subbuffer_mmap(ctx, stream, len); + ret = lttng_ustctl_get_mmap_read_offset(stream->chan->handle, + stream->buf, &mmap_offset); + break; default: ERR("Unknown consumer_data type"); assert(0); } + if (ret != 0) { + errno = -ret; + PERROR("tracer ctl get_mmap_read_offset"); + written = ret; + goto end; + } - return 0; + /* Handle stream on the relayd if the output is on the network */ + if (relayd) { + unsigned long netlen = len; + + /* + * Lock the control socket for the complete duration of the function + * since from this point on we will use the socket. + */ + if (stream->metadata_flag) { + /* Metadata requires the control socket. */ + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + netlen += sizeof(struct lttcomm_relayd_metadata_payload); + } + + ret = write_relayd_stream_header(stream, netlen, padding, relayd); + if (ret >= 0) { + /* Use the returned socket. */ + outfd = ret; + + /* Write metadata stream id before payload */ + if (stream->metadata_flag) { + ret = write_relayd_metadata_id(outfd, stream, relayd, padding); + if (ret < 0) { + written = ret; + goto end; + } + } + } + /* Else, use the default set before which is the filesystem. */ + } else { + /* No streaming, we have to set the len with the full padding */ + len += padding; + } + + while (len > 0) { + do { + ret = write(outfd, stream->mmap_base + mmap_offset, len); + } while (ret < 0 && errno == EINTR); + DBG("Consumer mmap write() ret %zd (len %lu)", ret, len); + if (ret < 0) { + PERROR("Error in file write"); + if (written == 0) { + written = ret; + } + goto end; + } else if (ret > len) { + PERROR("Error in file write (ret %zd > len %lu)", ret, len); + written += ret; + goto end; + } else { + len -= ret; + mmap_offset += ret; + } + + /* 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, ret, + SYNC_FILE_RANGE_WRITE); + stream->out_fd_offset += ret; + } + written += ret; + } + lttng_consumer_sync_trace_file(stream, orig_offset); + +end: + /* Unlock only if ctrl socket used */ + if (relayd && stream->metadata_flag) { + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); + } + + rcu_read_unlock(); + return written; } /* @@ -1108,20 +1252,178 @@ ssize_t lttng_consumer_on_read_subbuffer_mmap( */ ssize_t lttng_consumer_on_read_subbuffer_splice( struct lttng_consumer_local_data *ctx, - struct lttng_consumer_stream *stream, unsigned long len) + struct lttng_consumer_stream *stream, unsigned long len, + unsigned long padding) { + ssize_t ret = 0, written = 0, ret_splice = 0; + loff_t offset = 0; + off_t orig_offset = stream->out_fd_offset; + int fd = stream->wait_fd; + /* Default is on the disk */ + int outfd = stream->out_fd; + struct consumer_relayd_sock_pair *relayd = NULL; + int *splice_pipe; + switch (consumer_data.type) { case LTTNG_CONSUMER_KERNEL: - return lttng_kconsumer_on_read_subbuffer_splice(ctx, stream, len); + break; case LTTNG_CONSUMER32_UST: case LTTNG_CONSUMER64_UST: + /* Not supported for user space tracing */ return -ENOSYS; default: ERR("Unknown consumer_data type"); assert(0); - return -ENOSYS; } + /* RCU lock for the relayd pointer */ + rcu_read_lock(); + + /* Flag that the current stream if set for network streaming. */ + if (stream->net_seq_idx != -1) { + relayd = consumer_find_relayd(stream->net_seq_idx); + if (relayd == NULL) { + goto end; + } + } + + /* + * Choose right pipe for splice. Metadata and trace data are handled by + * different threads hence the use of two pipes in order not to race or + * corrupt the written data. + */ + if (stream->metadata_flag) { + splice_pipe = ctx->consumer_splice_metadata_pipe; + } else { + splice_pipe = ctx->consumer_thread_pipe; + } + + /* Write metadata stream id before payload */ + if (relayd) { + int total_len = len; + + if (stream->metadata_flag) { + /* + * Lock the control socket for the complete duration of the function + * since from this point on we will use the socket. + */ + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + + ret = write_relayd_metadata_id(splice_pipe[1], stream, relayd, + padding); + if (ret < 0) { + written = ret; + goto end; + } + + total_len += sizeof(struct lttcomm_relayd_metadata_payload); + } + + ret = write_relayd_stream_header(stream, total_len, padding, relayd); + if (ret >= 0) { + /* Use the returned socket. */ + outfd = ret; + } else { + ERR("Remote relayd disconnected. Stopping"); + goto end; + } + } else { + /* No streaming, we have to set the len with the full padding */ + len += padding; + } + + while (len > 0) { + DBG("splice chan to pipe offset %lu of len %lu (fd : %d, pipe: %d)", + (unsigned long)offset, len, fd, splice_pipe[1]); + ret_splice = splice(fd, &offset, splice_pipe[1], NULL, len, + SPLICE_F_MOVE | SPLICE_F_MORE); + DBG("splice chan to pipe, ret %zd", ret_splice); + if (ret_splice < 0) { + PERROR("Error in relay splice"); + if (written == 0) { + written = ret_splice; + } + ret = errno; + goto splice_error; + } + + /* Handle stream on the relayd if the output is on the network */ + if (relayd) { + if (stream->metadata_flag) { + size_t metadata_payload_size = + sizeof(struct lttcomm_relayd_metadata_payload); + + /* Update counter to fit the spliced data */ + ret_splice += metadata_payload_size; + len += metadata_payload_size; + /* + * We do this so the return value can match the len passed as + * argument to this function. + */ + written -= metadata_payload_size; + } + } + + /* Splice data out */ + ret_splice = splice(splice_pipe[0], NULL, outfd, NULL, + ret_splice, SPLICE_F_MOVE | SPLICE_F_MORE); + DBG("Consumer splice pipe to file, ret %zd", ret_splice); + if (ret_splice < 0) { + PERROR("Error in file splice"); + if (written == 0) { + written = ret_splice; + } + ret = errno; + goto splice_error; + } else if (ret_splice > len) { + errno = EINVAL; + PERROR("Wrote more data than requested %zd (len: %lu)", + ret_splice, len); + written += ret_splice; + ret = errno; + goto splice_error; + } + len -= ret_splice; + + /* 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, ret_splice, + SYNC_FILE_RANGE_WRITE); + stream->out_fd_offset += ret_splice; + } + written += ret_splice; + } + lttng_consumer_sync_trace_file(stream, orig_offset); + + ret = ret_splice; + + goto end; + +splice_error: + /* send the appropriate error description to sessiond */ + switch (ret) { + case EBADF: + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EBADF); + break; + case EINVAL: + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_EINVAL); + break; + case ENOMEM: + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ENOMEM); + break; + case ESPIPE: + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_SPLICE_ESPIPE); + break; + } + +end: + if (relayd && stream->metadata_flag) { + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); + } + + rcu_read_unlock(); + return written; } /* @@ -1185,6 +1487,341 @@ int lttng_consumer_recv_cmd(struct lttng_consumer_local_data *ctx, } } +/* + * Iterate over all stream element of the hashtable and free them. This is race + * free since the hashtable received MUST be in a race free synchronization + * state. It's the caller responsability to make sure of that. + */ +static void destroy_stream_ht(struct lttng_ht *ht) +{ + int ret; + struct lttng_ht_iter iter; + struct lttng_consumer_stream *stream; + + if (ht == NULL) { + return; + } + + rcu_read_lock(); + cds_lfht_for_each_entry(ht->ht, &iter.iter, stream, node.node) { + ret = lttng_ht_del(ht, &iter); + assert(!ret); + + free(stream); + } + rcu_read_unlock(); + + lttng_ht_destroy(ht); +} + +/* + * Clean up a metadata stream and free its memory. + */ +static void consumer_del_metadata_stream(struct lttng_consumer_stream *stream) +{ + int ret; + struct lttng_consumer_channel *free_chan = NULL; + struct consumer_relayd_sock_pair *relayd; + + assert(stream); + /* + * This call should NEVER receive regular stream. It must always be + * metadata stream and this is crucial for data structure synchronization. + */ + assert(stream->metadata_flag); + + pthread_mutex_lock(&consumer_data.lock); + switch (consumer_data.type) { + case LTTNG_CONSUMER_KERNEL: + if (stream->mmap_base != NULL) { + ret = munmap(stream->mmap_base, stream->mmap_len); + if (ret != 0) { + PERROR("munmap metadata stream"); + } + } + break; + case LTTNG_CONSUMER32_UST: + case LTTNG_CONSUMER64_UST: + lttng_ustconsumer_del_stream(stream); + break; + default: + ERR("Unknown consumer_data type"); + assert(0); + } + pthread_mutex_unlock(&consumer_data.lock); + + if (stream->out_fd >= 0) { + ret = close(stream->out_fd); + if (ret) { + PERROR("close"); + } + } + + if (stream->wait_fd >= 0 && !stream->wait_fd_is_copy) { + ret = close(stream->wait_fd); + if (ret) { + PERROR("close"); + } + } + + if (stream->shm_fd >= 0 && stream->wait_fd != stream->shm_fd) { + ret = close(stream->shm_fd); + if (ret) { + PERROR("close"); + } + } + + /* Check and cleanup relayd */ + rcu_read_lock(); + relayd = consumer_find_relayd(stream->net_seq_idx); + if (relayd != NULL) { + uatomic_dec(&relayd->refcount); + assert(uatomic_read(&relayd->refcount) >= 0); + + /* Closing streams requires to lock the control socket. */ + pthread_mutex_lock(&relayd->ctrl_sock_mutex); + ret = relayd_send_close_stream(&relayd->control_sock, + stream->relayd_stream_id, stream->next_net_seq_num - 1); + pthread_mutex_unlock(&relayd->ctrl_sock_mutex); + if (ret < 0) { + DBG("Unable to close stream on the relayd. Continuing"); + /* + * Continue here. There is nothing we can do for the relayd. + * Chances are that the relayd has closed the socket so we just + * continue cleaning up. + */ + } + + /* Both conditions are met, we destroy the relayd. */ + if (uatomic_read(&relayd->refcount) == 0 && + uatomic_read(&relayd->destroy_flag)) { + destroy_relayd(relayd); + } + } + rcu_read_unlock(); + + /* Atomically decrement channel refcount since other threads can use it. */ + uatomic_dec(&stream->chan->refcount); + if (!uatomic_read(&stream->chan->refcount)) { + free_chan = stream->chan; + } + + if (free_chan) { + consumer_del_channel(free_chan); + } + + free(stream); +} + +/* + * Action done with the metadata stream when adding it to the consumer internal + * data structures to handle it. + */ +static void consumer_add_metadata_stream(struct lttng_consumer_stream *stream) +{ + struct consumer_relayd_sock_pair *relayd; + + /* Find relayd and, if one is found, increment refcount. */ + rcu_read_lock(); + relayd = consumer_find_relayd(stream->net_seq_idx); + if (relayd != NULL) { + uatomic_inc(&relayd->refcount); + } + rcu_read_unlock(); +} + +/* + * Thread polls on metadata file descriptor and write them on disk or on the + * network. + */ +void *lttng_consumer_thread_poll_metadata(void *data) +{ + int ret, i, pollfd; + uint32_t revents, nb_fd; + struct lttng_consumer_stream *stream; + struct lttng_ht_iter iter; + struct lttng_ht_node_ulong *node; + struct lttng_ht *metadata_ht = NULL; + struct lttng_poll_event events; + struct lttng_consumer_local_data *ctx = data; + ssize_t len; + + rcu_register_thread(); + + DBG("Thread metadata poll started"); + + metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); + if (metadata_ht == NULL) { + goto end; + } + + /* Size is set to 1 for the consumer_metadata pipe */ + ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC); + if (ret < 0) { + ERR("Poll set creation failed"); + goto end; + } + + ret = lttng_poll_add(&events, ctx->consumer_metadata_pipe[0], LPOLLIN); + if (ret < 0) { + goto end; + } + + /* Main loop */ + DBG("Metadata main loop started"); + + while (1) { + lttng_poll_reset(&events); + + nb_fd = LTTNG_POLL_GETNB(&events); + + /* Only the metadata pipe is set */ + if (nb_fd == 0 && consumer_quit == 1) { + goto end; + } + +restart: + DBG("Metadata poll wait with %d fd(s)", nb_fd); + ret = lttng_poll_wait(&events, -1); + DBG("Metadata event catched in thread"); + if (ret < 0) { + if (errno == EINTR) { + goto restart; + } + goto error; + } + + for (i = 0; i < nb_fd; i++) { + revents = LTTNG_POLL_GETEV(&events, i); + pollfd = LTTNG_POLL_GETFD(&events, i); + + /* Check the metadata pipe for incoming metadata. */ + if (pollfd == ctx->consumer_metadata_pipe[0]) { + 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]); + close(ctx->consumer_metadata_pipe[0]); + continue; + } else if (revents & LPOLLIN) { + stream = zmalloc(sizeof(struct lttng_consumer_stream)); + if (stream == NULL) { + PERROR("zmalloc metadata consumer stream"); + goto error; + } + + do { + /* Get the stream and add it to the local hash table */ + ret = read(pollfd, stream, + sizeof(struct lttng_consumer_stream)); + } while (ret < 0 && errno == EINTR); + if (ret < 0 || ret < sizeof(struct lttng_consumer_stream)) { + PERROR("read metadata stream"); + free(stream); + /* + * Let's continue here and hope we can still work + * without stopping the consumer. XXX: Should we? + */ + continue; + } + + DBG("Adding metadata stream %d to poll set", + stream->wait_fd); + + rcu_read_lock(); + /* The node should be init at this point */ + lttng_ht_add_unique_ulong(metadata_ht, + &stream->waitfd_node); + rcu_read_unlock(); + + /* Add metadata stream to the global poll events list */ + lttng_poll_add(&events, stream->wait_fd, + LPOLLIN | LPOLLPRI); + + consumer_add_metadata_stream(stream); + } + + /* Metadata pipe handled. Continue handling the others */ + continue; + } + + /* From here, the event is a metadata wait fd */ + + rcu_read_lock(); + lttng_ht_lookup(metadata_ht, (void *)((unsigned long) pollfd), + &iter); + node = lttng_ht_iter_get_node_ulong(&iter); + if (node == NULL) { + /* FD not found, continue loop */ + rcu_read_unlock(); + continue; + } + + stream = caa_container_of(node, struct lttng_consumer_stream, + waitfd_node); + + /* Get the data out of the metadata file descriptor */ + if (revents & (LPOLLIN | LPOLLPRI)) { + DBG("Metadata available on fd %d", pollfd); + assert(stream->wait_fd == pollfd); + + len = ctx->on_buffer_ready(stream, ctx); + /* It's ok to have an unavailable sub-buffer */ + if (len < 0 && len != -EAGAIN) { + rcu_read_unlock(); + goto end; + } else if (len > 0) { + stream->data_read = 1; + } + } + + /* + * Remove the stream from the hash table since there is no data + * left on the fd because we previously did a read on the buffer. + */ + if (revents & (LPOLLERR | LPOLLHUP)) { + DBG("Metadata fd %d is hup|err|nval.", pollfd); + if (!stream->hangup_flush_done + && (consumer_data.type == LTTNG_CONSUMER32_UST + || consumer_data.type == LTTNG_CONSUMER64_UST)) { + DBG("Attempting to flush and consume the UST buffers"); + lttng_ustconsumer_on_stream_hangup(stream); + + /* We just flushed the stream now read it. */ + len = ctx->on_buffer_ready(stream, ctx); + /* It's ok to have an unavailable sub-buffer */ + if (len < 0 && len != -EAGAIN) { + rcu_read_unlock(); + goto end; + } + } + + /* Removing it from hash table, poll set and free memory */ + lttng_ht_del(metadata_ht, &iter); + + lttng_poll_del(&events, stream->wait_fd); + consumer_del_metadata_stream(stream); + } + rcu_read_unlock(); + } + } + +error: +end: + DBG("Metadata poll thread exiting"); + lttng_poll_clean(&events); + + if (metadata_ht) { + destroy_stream_ht(metadata_ht); + } + + rcu_unregister_thread(); + return NULL; +} + /* * This thread polls the fds in the set to consume the data and write * it to tracefile if necessary. @@ -1198,16 +1835,20 @@ void *lttng_consumer_thread_poll_fds(void *data) /* local view of consumer_data.fds_count */ int nb_fd = 0; struct lttng_consumer_local_data *ctx = data; - struct lttng_ht *metadata_ht; - struct lttng_ht_iter iter; - struct lttng_ht_node_ulong *node; - struct lttng_consumer_stream *metadata_stream; ssize_t len; - - metadata_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); + pthread_t metadata_thread; + void *status; rcu_register_thread(); + /* Start metadata polling thread */ + ret = pthread_create(&metadata_thread, NULL, + lttng_consumer_thread_poll_metadata, (void *) ctx); + if (ret < 0) { + PERROR("pthread_create metadata thread"); + goto end; + } + local_stream = zmalloc(sizeof(struct lttng_consumer_stream)); while (1) { @@ -1245,11 +1886,10 @@ void *lttng_consumer_thread_poll_fds(void *data) pthread_mutex_unlock(&consumer_data.lock); goto end; } - ret = consumer_update_poll_array(ctx, &pollfd, local_stream, - metadata_ht); + ret = consumer_update_poll_array(ctx, &pollfd, local_stream); if (ret < 0) { ERR("Error in allocating pollfd or local_outfds"); - lttng_consumer_send_error(ctx, CONSUMERD_POLL_ERROR); + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR); pthread_mutex_unlock(&consumer_data.lock); goto end; } @@ -1275,7 +1915,7 @@ void *lttng_consumer_thread_poll_fds(void *data) goto restart; } perror("Poll error"); - lttng_consumer_send_error(ctx, CONSUMERD_POLL_ERROR); + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_POLL_ERROR); goto end; } else if (num_rdy == 0) { DBG("Polling thread timed out"); @@ -1301,24 +1941,7 @@ void *lttng_consumer_thread_poll_fds(void *data) /* Take care of high priority channels first. */ for (i = 0; i < nb_fd; i++) { - /* Lookup for metadata which is the highest priority */ - lttng_ht_lookup(metadata_ht, - (void *)((unsigned long) pollfd[i].fd), &iter); - node = lttng_ht_iter_get_node_ulong(&iter); - if (node != NULL && - (pollfd[i].revents & (POLLIN | POLLPRI))) { - DBG("Urgent metadata read on fd %d", pollfd[i].fd); - metadata_stream = caa_container_of(node, - struct lttng_consumer_stream, waitfd_node); - high_prio = 1; - len = ctx->on_buffer_ready(metadata_stream, ctx); - /* it's ok to have an unavailable sub-buffer */ - if (len < 0 && len != -EAGAIN) { - goto end; - } else if (len > 0) { - metadata_stream->data_read = 1; - } - } else if (pollfd[i].revents & POLLPRI) { + if (pollfd[i].revents & POLLPRI) { DBG("Urgent read on fd %d", pollfd[i].fd); high_prio = 1; len = ctx->on_buffer_ready(local_stream[i], ctx); @@ -1374,33 +1997,18 @@ void *lttng_consumer_thread_poll_fds(void *data) if ((pollfd[i].revents & POLLHUP)) { DBG("Polling fd %d tells it has hung up.", pollfd[i].fd); if (!local_stream[i]->data_read) { - if (local_stream[i]->metadata_flag) { - iter.iter.node = &local_stream[i]->waitfd_node.node; - ret = lttng_ht_del(metadata_ht, &iter); - assert(!ret); - } consumer_del_stream(local_stream[i]); num_hup++; } } else if (pollfd[i].revents & POLLERR) { ERR("Error returned in polling fd %d.", pollfd[i].fd); if (!local_stream[i]->data_read) { - if (local_stream[i]->metadata_flag) { - iter.iter.node = &local_stream[i]->waitfd_node.node; - ret = lttng_ht_del(metadata_ht, &iter); - assert(!ret); - } consumer_del_stream(local_stream[i]); num_hup++; } } else if (pollfd[i].revents & POLLNVAL) { ERR("Polling fd %d tells fd is not open.", pollfd[i].fd); if (!local_stream[i]->data_read) { - if (local_stream[i]->metadata_flag) { - iter.iter.node = &local_stream[i]->waitfd_node.node; - ret = lttng_ht_del(metadata_ht, &iter); - assert(!ret); - } consumer_del_stream(local_stream[i]); num_hup++; } @@ -1418,6 +2026,23 @@ end: free(local_stream); local_stream = NULL; } + + /* + * Close the write side of the pipe so epoll_wait() in + * lttng_consumer_thread_poll_metadata can catch it. The thread is + * monitoring the read side of the pipe. If we close them both, epoll_wait + * strangely does not return and could create a endless wait period if the + * pipe is the only tracked fd in the poll set. The thread will take care + * of closing the read side. + */ + close(ctx->consumer_metadata_pipe[1]); + if (ret) { + ret = pthread_join(metadata_thread, &status); + if (ret < 0) { + PERROR("pthread_join metadata thread"); + } + } + rcu_unregister_thread(); return NULL; } @@ -1452,7 +2077,7 @@ void *lttng_consumer_thread_receive_fds(void *data) } DBG("Sending ready command to lttng-sessiond"); - ret = lttng_consumer_send_error(ctx, CONSUMERD_COMMAND_SOCK_READY); + ret = lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_COMMAND_SOCK_READY); /* return < 0 on error, but == 0 is not fatal */ if (ret < 0) { ERR("Error sending ready command to lttng-sessiond"); @@ -1502,8 +2127,12 @@ void *lttng_consumer_thread_receive_fds(void *data) DBG("Received STOP command"); goto end; } - if (ret < 0) { - ERR("Communication interrupted on command socket"); + if (ret <= 0) { + /* + * This could simply be a session daemon quitting. Don't output + * ERR() here. + */ + DBG("Communication interrupted on command socket"); goto end; } if (consumer_quit) { @@ -1586,3 +2215,95 @@ void lttng_consumer_init(void) consumer_data.channel_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); consumer_data.relayd_ht = lttng_ht_new(0, LTTNG_HT_TYPE_ULONG); } + +/* + * Process the ADD_RELAYD command receive by a consumer. + * + * This will create a relayd socket pair and add it to the relayd hash table. + * The caller MUST acquire a RCU read side lock before calling it. + */ +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) +{ + int fd, ret = -1; + struct consumer_relayd_sock_pair *relayd; + + DBG("Consumer adding relayd socket (idx: %d)", net_seq_idx); + + /* Get relayd reference if exists. */ + relayd = consumer_find_relayd(net_seq_idx); + if (relayd == NULL) { + /* Not found. Allocate one. */ + relayd = consumer_allocate_relayd_sock_pair(net_seq_idx); + if (relayd == NULL) { + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_OUTFD_ERROR); + goto error; + } + } + + /* Poll on consumer socket. */ + if (lttng_consumer_poll_socket(consumer_sockpoll) < 0) { + ret = -EINTR; + goto error; + } + + /* Get relayd socket from session daemon */ + ret = lttcomm_recv_fds_unix_sock(sock, &fd, 1); + if (ret != sizeof(fd)) { + lttng_consumer_send_error(ctx, LTTCOMM_CONSUMERD_ERROR_RECV_FD); + ret = -1; + goto error; + } + + /* Copy socket information and received FD */ + 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); + if (ret < 0) { + goto error; + } + + /* Close the created socket fd which is useless */ + close(relayd->control_sock.fd); + + /* Assign new file descriptor */ + relayd->control_sock.fd = fd; + break; + case LTTNG_STREAM_DATA: + /* Copy received lttcomm socket */ + lttcomm_copy_sock(&relayd->data_sock, relayd_sock); + ret = lttcomm_create_sock(&relayd->data_sock); + if (ret < 0) { + goto error; + } + + /* Close the created socket fd which is useless */ + close(relayd->data_sock.fd); + + /* Assign new file descriptor */ + relayd->data_sock.fd = fd; + break; + default: + ERR("Unknown relayd socket type (%d)", sock_type); + goto error; + } + + DBG("Consumer %s socket created successfully with net idx %d (fd: %d)", + sock_type == LTTNG_STREAM_CONTROL ? "control" : "data", + relayd->net_seq_idx, fd); + + /* + * Add relayd socket pair to consumer data hashtable. If object already + * exists or on error, the function gracefully returns. + */ + add_relayd(relayd); + + /* All good! */ + ret = 0; + +error: + return ret; +}