[lttng-dev] [PATCH lttng-tools 2/4] Move add data stream to the data thread
Mathieu Desnoyers
mathieu.desnoyers at efficios.com
Mon Oct 15 13:31:01 EDT 2012
* David Goulet (dgoulet at efficios.com) wrote:
>
>
> Mathieu Desnoyers:
> > * David Goulet (dgoulet at efficios.com) wrote:
> >> As a second step of refactoring, upon receiving a data stream, we send
> >> it to the data thread that is now in charge of handling it.
> >>
> >> Furthermore, in order for this to behave correctly, we have to make the
> >> ustctl actions on the stream upon before passing it to the right thread
> >> (the kernel does not need special actions.). This way, once the sessiond
> >> thread reply back to the session daemon, the stream is sure to be open
> >> and ready for data to be recorded on the application side so we avoid a
> >> race between the application thinking the stream is ready and the stream
> >> thread still scheduled out.
> >
> > Normally, as long as we have a reference on the SHM file descriptor, and
> > we have the wakeup FD, we should be good to fetch the data of buffers
> > belonging to an application that has already exited, even if it did so
> > before the ustctl calls are done.
> >
> > So I'm wondering why you do the ustctl calls in the sessiond thread ? It
> > seems to complexify the implementation needlessly: we could still do the
> > ustctl calls and output file open at the same location, the
> > data/metadata threads.
>
> Hmmm, it was my understanding that does
does -> those
> ustctl_* calls were needed
> before the trace could be recording thus making them quickly. Wrong?
Can you rephrase your question ? I don't understand.
Thanks,
Mathieu
>
> David
>
> >
> > Thanks,
> >
> > Mathieu
> >
> >>
> >> This commit should speed up the add stream process for the session
> >> daemon. There is still some actions to move out of the session daemon
> >> poll thread to gain speed significantly, especially for network
> >> streaming.
> >>
> >> Signed-off-by: David Goulet <dgoulet at efficios.com>
> >> ---
> >> src/common/consumer.c | 123 +++++++++++---------------
> >> src/common/consumer.h | 1 +
> >> src/common/kernel-consumer/kernel-consumer.c | 24 ++---
> >> src/common/ust-consumer/ust-consumer.c | 40 ++++-----
> >> 4 files changed, 78 insertions(+), 110 deletions(-)
> >>
> >> diff --git a/src/common/consumer.c b/src/common/consumer.c
> >> index 055de1b..1d2b1f7 100644
> >> --- a/src/common/consumer.c
> >> +++ b/src/common/consumer.c
> >> @@ -89,7 +89,7 @@ static struct lttng_consumer_stream *consumer_find_stream(int key,
> >> return stream;
> >> }
> >>
> >> -static void consumer_steal_stream_key(int key, struct lttng_ht *ht)
> >> +void consumer_steal_stream_key(int key, struct lttng_ht *ht)
> >> {
> >> struct lttng_consumer_stream *stream;
> >>
> >> @@ -409,6 +409,14 @@ struct lttng_consumer_stream *consumer_allocate_stream(
> >> lttng_ht_node_init_ulong(&stream->waitfd_node, stream->wait_fd);
> >> lttng_ht_node_init_ulong(&stream->node, stream->key);
> >>
> >> + /*
> >> + * The cpu number is needed before using any ustctl_* actions. Ignored for
> >> + * the kernel so the value does not matter.
> >> + */
> >> + pthread_mutex_lock(&consumer_data.lock);
> >> + stream->cpu = stream->chan->cpucount++;
> >> + pthread_mutex_unlock(&consumer_data.lock);
> >> +
> >> DBG3("Allocated stream %s (key %d, shm_fd %d, wait_fd %d, mmap_len %llu,"
> >> " out_fd %d, net_seq_idx %d)", stream->path_name, stream->key,
> >> stream->shm_fd, stream->wait_fd,
> >> @@ -437,28 +445,6 @@ int consumer_add_stream(struct lttng_consumer_stream *stream)
> >> pthread_mutex_lock(&consumer_data.lock);
> >> rcu_read_lock();
> >>
> >> - switch (consumer_data.type) {
> >> - case LTTNG_CONSUMER_KERNEL:
> >> - break;
> >> - case LTTNG_CONSUMER32_UST:
> >> - case LTTNG_CONSUMER64_UST:
> >> - stream->cpu = stream->chan->cpucount++;
> >> - ret = lttng_ustconsumer_add_stream(stream);
> >> - if (ret) {
> >> - ret = -EINVAL;
> >> - goto error;
> >> - }
> >> -
> >> - /* Steal stream identifier only for UST */
> >> - consumer_steal_stream_key(stream->key, consumer_data.stream_ht);
> >> - break;
> >> - default:
> >> - ERR("Unknown consumer_data type");
> >> - assert(0);
> >> - ret = -ENOSYS;
> >> - goto error;
> >> - }
> >> -
> >> lttng_ht_add_unique_ulong(consumer_data.stream_ht, &stream->node);
> >>
> >> /* Check and cleanup relayd */
> >> @@ -485,7 +471,6 @@ int consumer_add_stream(struct lttng_consumer_stream *stream)
> >> consumer_data.stream_count++;
> >> consumer_data.need_update = 1;
> >>
> >> -error:
> >> rcu_read_unlock();
> >> pthread_mutex_unlock(&consumer_data.lock);
> >>
> >> @@ -1582,17 +1567,6 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
> >>
> >> DBG3("Consumer delete metadata stream %d", stream->wait_fd);
> >>
> >> - if (ht == NULL) {
> >> - /* Means the stream was allocated but not successfully added */
> >> - goto free_stream;
> >> - }
> >> -
> >> - rcu_read_lock();
> >> - iter.iter.node = &stream->waitfd_node.node;
> >> - ret = lttng_ht_del(ht, &iter);
> >> - assert(!ret);
> >> - rcu_read_unlock();
> >> -
> >> pthread_mutex_lock(&consumer_data.lock);
> >> switch (consumer_data.type) {
> >> case LTTNG_CONSUMER_KERNEL:
> >> @@ -1613,6 +1587,18 @@ void consumer_del_metadata_stream(struct lttng_consumer_stream *stream,
> >> goto end;
> >> }
> >>
> >> + if (ht == NULL) {
> >> + pthread_mutex_unlock(&consumer_data.lock);
> >> + /* Means the stream was allocated but not successfully added */
> >> + goto free_stream;
> >> + }
> >> +
> >> + rcu_read_lock();
> >> + iter.iter.node = &stream->waitfd_node.node;
> >> + ret = lttng_ht_del(ht, &iter);
> >> + assert(!ret);
> >> + rcu_read_unlock();
> >> +
> >> if (stream->out_fd >= 0) {
> >> ret = close(stream->out_fd);
> >> if (ret) {
> >> @@ -1699,27 +1685,6 @@ static int consumer_add_metadata_stream(struct lttng_consumer_stream *stream,
> >>
> >> pthread_mutex_lock(&consumer_data.lock);
> >>
> >> - switch (consumer_data.type) {
> >> - case LTTNG_CONSUMER_KERNEL:
> >> - break;
> >> - case LTTNG_CONSUMER32_UST:
> >> - case LTTNG_CONSUMER64_UST:
> >> - ret = lttng_ustconsumer_add_stream(stream);
> >> - if (ret) {
> >> - ret = -EINVAL;
> >> - goto error;
> >> - }
> >> -
> >> - /* Steal stream identifier only for UST */
> >> - consumer_steal_stream_key(stream->wait_fd, ht);
> >> - break;
> >> - default:
> >> - ERR("Unknown consumer_data type");
> >> - assert(0);
> >> - ret = -ENOSYS;
> >> - goto error;
> >> - }
> >> -
> >> /*
> >> * From here, refcounts are updated so be _careful_ when returning an error
> >> * after this point.
> >> @@ -1749,7 +1714,6 @@ static int consumer_add_metadata_stream(struct lttng_consumer_stream *stream,
> >> lttng_ht_add_unique_ulong(ht, &stream->waitfd_node);
> >> rcu_read_unlock();
> >>
> >> -error:
> >> pthread_mutex_unlock(&consumer_data.lock);
> >> return ret;
> >> }
> >> @@ -1946,7 +1910,7 @@ void *consumer_thread_data_poll(void *data)
> >> int num_rdy, num_hup, high_prio, ret, i;
> >> struct pollfd *pollfd = NULL;
> >> /* local view of the streams */
> >> - struct lttng_consumer_stream **local_stream = NULL;
> >> + struct lttng_consumer_stream **local_stream = NULL, *new_stream = NULL;
> >> /* local view of consumer_data.fds_count */
> >> int nb_fd = 0;
> >> struct lttng_consumer_local_data *ctx = data;
> >> @@ -2034,13 +1998,35 @@ void *consumer_thread_data_poll(void *data)
> >> */
> >> if (pollfd[nb_fd].revents & (POLLIN | POLLPRI)) {
> >> size_t pipe_readlen;
> >> - char tmp;
> >>
> >> DBG("consumer_poll_pipe wake up");
> >> /* Consume 1 byte of pipe data */
> >> do {
> >> - pipe_readlen = read(ctx->consumer_poll_pipe[0], &tmp, 1);
> >> + pipe_readlen = read(ctx->consumer_poll_pipe[0], &new_stream,
> >> + sizeof(new_stream));
> >> } while (pipe_readlen == -1 && errno == EINTR);
> >> +
> >> + /*
> >> + * If the stream is NULL, just ignore it. It's also possible that
> >> + * the sessiond poll thread changed the consumer_quit state and is
> >> + * waking us up to test it.
> >> + */
> >> + if (new_stream == NULL) {
> >> + continue;
> >> + }
> >> +
> >> + ret = consumer_add_stream(new_stream);
> >> + if (ret) {
> >> + ERR("Consumer add stream %d failed. Continuing",
> >> + new_stream->key);
> >> + /*
> >> + * At this point, if the add_stream fails, it is not in the
> >> + * hash table thus passing the NULL value here.
> >> + */
> >> + consumer_del_stream(new_stream, NULL);
> >> + }
> >> +
> >> + /* Continue to update the local streams and handle prio ones */
> >> continue;
> >> }
> >>
> >> @@ -2260,19 +2246,16 @@ end:
> >> consumer_poll_timeout = LTTNG_CONSUMER_POLL_TIMEOUT;
> >>
> >> /*
> >> - * Wake-up the other end by writing a null byte in the pipe
> >> - * (non-blocking). Important note: Because writing into the
> >> - * pipe is non-blocking (and therefore we allow dropping wakeup
> >> - * data, as long as there is wakeup data present in the pipe
> >> - * buffer to wake up the other end), the other end should
> >> - * perform the following sequence for waiting:
> >> - * 1) empty the pipe (reads).
> >> - * 2) perform update operation.
> >> - * 3) wait on the pipe (poll).
> >> + * Notify the data poll thread to poll back again and test the
> >> + * consumer_quit state to quit gracefully.
> >> */
> >> do {
> >> - ret = write(ctx->consumer_poll_pipe[1], "", 1);
> >> + struct lttng_consumer_stream *null_stream = NULL;
> >> +
> >> + ret = write(ctx->consumer_poll_pipe[1], &null_stream,
> >> + sizeof(null_stream));
> >> } while (ret < 0 && errno == EINTR);
> >> +
> >> rcu_unregister_thread();
> >> return NULL;
> >> }
> >> diff --git a/src/common/consumer.h b/src/common/consumer.h
> >> index 4b225e4..8e5891a 100644
> >> --- a/src/common/consumer.h
> >> +++ b/src/common/consumer.h
> >> @@ -362,6 +362,7 @@ struct consumer_relayd_sock_pair *consumer_allocate_relayd_sock_pair(
> >> struct consumer_relayd_sock_pair *consumer_find_relayd(int key);
> >> int consumer_handle_stream_before_relayd(struct lttng_consumer_stream *stream,
> >> size_t data_size);
> >> +void consumer_steal_stream_key(int key, struct lttng_ht *ht);
> >>
> >> extern struct lttng_consumer_local_data *lttng_consumer_create(
> >> enum lttng_consumer_type type,
> >> diff --git a/src/common/kernel-consumer/kernel-consumer.c b/src/common/kernel-consumer/kernel-consumer.c
> >> index 13cbe21..444f5e0 100644
> >> --- a/src/common/kernel-consumer/kernel-consumer.c
> >> +++ b/src/common/kernel-consumer/kernel-consumer.c
> >> @@ -235,10 +235,12 @@ int lttng_kconsumer_recv_cmd(struct lttng_consumer_local_data *ctx,
> >> consumer_del_stream(new_stream, NULL);
> >> }
> >> } else {
> >> - ret = consumer_add_stream(new_stream);
> >> - if (ret) {
> >> - ERR("Consumer add stream %d failed. Continuing",
> >> - new_stream->key);
> >> + do {
> >> + ret = write(ctx->consumer_poll_pipe[1], &new_stream,
> >> + sizeof(new_stream));
> >> + } while (ret < 0 && errno == EINTR);
> >> + if (ret < 0) {
> >> + PERROR("write data pipe");
> >> consumer_del_stream(new_stream, NULL);
> >> goto end_nosignal;
> >> }
> >> @@ -284,20 +286,6 @@ int lttng_kconsumer_recv_cmd(struct lttng_consumer_local_data *ctx,
> >> goto end_nosignal;
> >> }
> >>
> >> - /*
> >> - * Wake-up the other end by writing a null byte in the pipe (non-blocking).
> >> - * Important note: Because writing into the pipe is non-blocking (and
> >> - * therefore we allow dropping wakeup data, as long as there is wakeup data
> >> - * present in the pipe buffer to wake up the other end), the other end
> >> - * should perform the following sequence for waiting:
> >> - *
> >> - * 1) empty the pipe (reads).
> >> - * 2) perform update operation.
> >> - * 3) wait on the pipe (poll).
> >> - */
> >> - do {
> >> - ret = write(ctx->consumer_poll_pipe[1], "", 1);
> >> - } while (ret < 0 && errno == EINTR);
> >> end_nosignal:
> >> rcu_read_unlock();
> >>
> >> diff --git a/src/common/ust-consumer/ust-consumer.c b/src/common/ust-consumer/ust-consumer.c
> >> index 1170687..4ca4b84 100644
> >> --- a/src/common/ust-consumer/ust-consumer.c
> >> +++ b/src/common/ust-consumer/ust-consumer.c
> >> @@ -224,6 +224,18 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx,
> >> goto end_nosignal;
> >> }
> >>
> >> + /*
> >> + * This needs to be done as soon as we can so we don't block the
> >> + * application too long.
> >> + */
> >> + ret = lttng_ustconsumer_add_stream(new_stream);
> >> + if (ret) {
> >> + consumer_del_stream(new_stream, NULL);
> >> + goto end_nosignal;
> >> + }
> >> + /* Steal stream identifier to avoid having streams with the same key */
> >> + consumer_steal_stream_key(new_stream->key, consumer_data.stream_ht);
> >> +
> >> /* The stream is not metadata. Get relayd reference if exists. */
> >> relayd = consumer_find_relayd(msg.u.stream.net_index);
> >> if (relayd != NULL) {
> >> @@ -265,14 +277,12 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx,
> >> goto end_nosignal;
> >> }
> >> } else {
> >> - ret = consumer_add_stream(new_stream);
> >> - if (ret) {
> >> - ERR("Consumer add stream %d failed. Continuing",
> >> - new_stream->key);
> >> - /*
> >> - * At this point, if the add_stream fails, it is not in the
> >> - * hash table thus passing the NULL value here.
> >> - */
> >> + do {
> >> + ret = write(ctx->consumer_poll_pipe[1], &new_stream,
> >> + sizeof(new_stream));
> >> + } while (ret < 0 && errno == EINTR);
> >> + if (ret < 0) {
> >> + PERROR("write data pipe");
> >> consumer_del_stream(new_stream, NULL);
> >> goto end_nosignal;
> >> }
> >> @@ -334,20 +344,6 @@ int lttng_ustconsumer_recv_cmd(struct lttng_consumer_local_data *ctx,
> >> break;
> >> }
> >>
> >> - /*
> >> - * Wake-up the other end by writing a null byte in the pipe (non-blocking).
> >> - * Important note: Because writing into the pipe is non-blocking (and
> >> - * therefore we allow dropping wakeup data, as long as there is wakeup data
> >> - * present in the pipe buffer to wake up the other end), the other end
> >> - * should perform the following sequence for waiting:
> >> - *
> >> - * 1) empty the pipe (reads).
> >> - * 2) perform update operation.
> >> - * 3) wait on the pipe (poll).
> >> - */
> >> - do {
> >> - ret = write(ctx->consumer_poll_pipe[1], "", 1);
> >> - } while (ret < 0 && errno == EINTR);
> >> end_nosignal:
> >> rcu_read_unlock();
> >>
> >> --
> >> 1.7.10.4
> >>
> >>
> >> _______________________________________________
> >> lttng-dev mailing list
> >> lttng-dev at lists.lttng.org
> >> http://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev
> >
--
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com
More information about the lttng-dev
mailing list