| Commit message (Collapse) | Author | Age |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
fixes #326 consider nni_taskq_exec_synch()
fixes #410 kqueue implementation could be smarter
fixes #411 epoll_implementation could be smarter
fixes #426 synchronous completion can lead to panic
fixes #421 pipe close race condition/duplicate destroy
This is a major refactoring of two significant parts of the code base,
which are closely interrelated.
First the aio and taskq framework have undergone a number of simplifications,
and improvements. We have ditched a few parts of the internal API (for
example tasks no longer support cancellation) that weren't terribly useful
but added a lot of complexity, and we've made aio_schedule something that
now checks for cancellation or other "premature" completions. The
aio framework now uses the tasks more tightly, so that aio wait can
devolve into just nni_task_wait(). We did have to add a "task_prep()"
step to prevent race conditions.
Second, the entire POSIX poller framework has been simplified, and made
more robust, and more scalable. There were some fairly inherent race
conditions around the shutdown/close code, where we *thought* we were
synchronizing against the other thread, but weren't doing so adequately.
With a cleaner design, we've been able to tighten up the implementation
to remove these race conditions, while substantially reducing the chance
for lock contention, thereby improving scalability. The illumos poller
also got a performance boost by polling for multiple events.
In highly "busy" systems, we expect to see vast reductions in lock
contention, and therefore greater scalability, in addition to overall
improved reliability.
One area where we currently can do better is that there is still only
a single poller thread run. Scaling this out is a task that has to be done
differently for each poller, and carefuly to ensure that close conditions
are safe on all pollers, and that no chance for deadlock/livelock waiting
for pfd finalizers can occur.
|
| | |
|
| |
|
|
|
|
|
|
|
|
| |
While here we added a test for the aio stuff, and cleaned up some dead
code for the old fd notifications. There were a few improvements to
shorten & clean code elsewhere, such as short-circuiting task wait
when the task has no callback.
The legacy sendmsg() and recvmsg() APIs are still in the socket core
until we convert the device code to use the aios.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
If the underlying platform fails (FreeBSD is the only one I'm aware
of that does this!), we use a global lock or condition variable instead.
This means that our lock initializers never ever fail.
Probably we could eliminate most of this for Linux and Darwin, since
on those platforms, mutex and condvar initialization reasonably never
fails. Initial benchmarks show little difference either way -- so we
can revisit (optimize) later.
This removes a lot of otherwise untested code in error cases and so forth,
improving coverage and resilience in the face of allocation failures.
Platforms other than POSIX should follow a similar pattern if they need
this. (VxWorks, I'm thinking of you.) Most sane platforms won't have
an issue here, since normally these initializations do not need to allocate
memory. (Reportedly, even FreeBSD has plans to "fix" this in libthr2.)
While here, some bugs were fixed in initialization & teardown.
The fallback code is properly tested with dedicated test cases.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
A little benchmarking showed that we were encountering far too many
wakeups, leading to severe performance degradation; we had a bunch
of threads all sleeping on the same condition variable (taskqs)
and this woke them all up, resulting in heavy mutex contention.
Since we only need one of the threads to wake, and we don't care which
one, let's just wake only one. This reduced RTT latency from about
240 us down to about 30 s. (1/8 of the former cost.)
There's still a bunch of tuning to do; performance remains worse than
we would like.
|
| | |
|
| |
|
|
|
|
|
|
|
| |
This passes valgrind 100% clean for both helgrind and deep leak
checks. This represents a complete rethink of how the AIOs work,
and much simpler synchronization; the provider API is a bit simpler
to boot, as a number of failure modes have been simply eliminated.
While here a few other minor bugs were squashed.
|
| | |
|
| |
|
|
| |
block for any AIO completion.
|
| |
|
|
|
|
|
| |
The queue is bound at initialization time of the task, and we call
entries just tasks, so we don't have to pass around a taskq pointer
across all the calls. Further, nni_task_dispatch is now guaranteed
to succeed.
|
| |
|
|
|
|
|
|
|
| |
We need to remember that protocol stops can run synchronously, and
therefore we need to wait for the aio to complete. Further, we need
to break apart shutting down aio activity from deallocation, as we need
to shut down *all* async activity before deallocating *anything*.
Noticed that we had a pipe race in the surveyor pattern too.
|
| |
|
|
|
|
|
|
| |
We have seen some yet another weird situation where we had an orphaned
pipe, which was caused by not completing the callback. If we are going
to run nni_aio_fini, we should still run the callback (albeit with a
return value of NNG_ECANCELED or somesuch) to be sure that we can't
orphan stuff.
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
