| Commit message (Collapse) | Author | Age |
|---|
| |
|
|
|
|
|
|
|
| |
This is a major change, and includes changes to use a polymorphic
stream API for all transports. There have been related bugs fixed
along the way. Additionally the man pages have changed.
The old non-polymorphic APIs are removed now. This is a breaking
change, but the old APIs were never part of any released public API.
|
| | |
|
| |
|
|
|
|
|
|
|
| |
This changes the signature of the aio cancellation routines
to take the argument for cancellation directly, so we do not
need to lookup the argument using the nni_aio_get_prov_data.
We should probably consider eliminating nni_aio_get_prov_data,
and co, and changing the prov_extra to reflect prov_data. Later.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* fixes #419 want to nni_aio_stop without blocking
This actually introduces an nni_aio_close() API that causes
nni_aio_begin to return NNG_ECLOSED, while scheduling a callback
on the AIO to do an NNG_ECLOSED as well. This should be called
in non-blocking close() contexts instead of nni_aio_stop(), and
the cases where we call nni_aio_fini() multiple times are updated
updated to add nni_aio_stop() calls on all "interlinked" aios before
finalizing them.
Furthermore, we call nni_aio_close() as soon as practical in the
close path. This closes an annoying race condition where the
callback from a lower subsystem could wind up rescheduling an
operation that we wanted to abort.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This closes a fundamental flaw in the way aio structures were
handled. In paticular, aio expiration could race ahead, and
fire before the aio was properly registered by the provider.
This ultimately led to the possibility of duplicate completions
on the same aio.
The solution involved breaking up nni_aio_start into two functions.
nni_aio_begin (which can be run outside of external locks) simply
validates that nni_aio_fini() has not been called, and clears certain
fields in the aio to make it ready for use by the provider.
nni_aio_schedule does the work to register the aio with the expiration
thread, and should only be called when the aio is actually scheduled
for asynchronous completion. nni_aio_schedule_verify does the same thing,
but returns NNG_ETIMEDOUT if the aio has a zero length timeout.
This change has a small negative performance impact. We have plans to
rectify that by converting nni_aio_begin to use a locklesss flag for
the aio->a_fini bit.
While we were here, we fixed some error paths in the POSIX subsystem,
which would have returned incorrect error codes, and we made some
optmizations in the message queues to reduce conditionals while holding
locks in the hot code path.
|
| |
|
|
|
|
|
|
| |
fixes #325 synchronous aio completion crash
fixes #327 move nni_clock() operations to outside the nni_aio_lk.
This work was done for the context tree, and is necessary to properly
enable that branch.
|
| | |
|
| |
|
|
|
|
|
| |
We enabled verbose compiler warnings, and found a lot of issues.
Some of these were even real bugs. As a bonus, we actually save
some initialization steps in the compat layer, and avoid passing
some variables we don't need.
|
| |
|
|
|
|
|
|
| |
This addresses the use of the pipe special field, and eliminates it.
The message APIs (recvmsg, sendmsg) need to be updated as well still,
but I want to handle that as part of a separate issue.
While here we fixed various compiler warnings, etc.
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This introduces enough of the HTTP API to support fully server
applications, including creation of websocket style protocols,
pluggable handlers, and so forth.
We have also introduced scatter/gather I/O (rudimentary) for
aios, and made other enhancements to the AIO framework. The
internals of the AIOs themselves are now fully private, and we
have eliminated the aio->a_addr member, with plans to remove the
pipe and possibly message members as well.
A few other minor issues were found and fixed as well.
The HTTP API includes request, response, and connection objects,
which can be used with both servers and clients. It also defines
the HTTP server and handler objects, which support server applications.
Support for client applications will require a client object to be
exposed, and that should be happening shortly.
None of this is "documented" yet, bug again, we will follow up shortly.
|
| | |
|
| |
|
|
|
| |
fixes #210 Want NNG_OPT_TLS_* options for TLS transport
fixes #212 Eliminate a_endpt member of aio
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is a rather large changeset -- it fundamentally adds websocket
transport, but as part of this changeset we added a generic framework
for both HTTP and websocket. We also made some supporting changes to
the core, such as changing the way timeouts work for AIOs and adding
additional state keeping for AIOs, and adding a common framework for
deferred finalization (to avoid certain kinds of circular deadlocks
during resource cleanup). We also invented a new initialization framework
so that we can avoid wiring in knowledge about them into the master
initialization framework.
The HTTP framework is not yet complete, but it is good enough for simple
static serving and building additional services on top of -- including
websocket. We expect both websocket and HTTP support to evolve
considerably, and so these are not part of the public API yet.
Property support for the websocket transport (in particular address
properties) is still missing, as is support for TLS.
The websocket transport here is a bit more robust than the original
nanomsg implementation, as it supports multiple sockets listening at
the same port sharing the same HTTP server instance, discriminating
between them based on URI (and possibly the virtual host).
Websocket is enabled by default at present, and work to conditionalize
HTTP and websocket further (to minimize bloat) is still pending.
|
| |
|
|
|
|
|
|
|
|
| |
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.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
We introduced richer, deeper tests for UDP functionality.
These tests uncovered a number of issues which this commit fixes.
The Windows IOCP code needs to support multiple aios on a single
nni_win_event. A redesign of the IOCP handling addresses that.
The POSIX UDP code also needed fixes; foremost among them is the
fact that the UDP file descriptor is not placed into non-blocking
mode, leading to potential hangs.
A number of race conditions and bugs along the implementation of
the above items were uncovered and fixed. To the best of our knowledge
the current code is bug-free.
|
| | |
|
| |
|
|
|
|
|
|
|
| |
We add a flag (auto-clearing) that can be set on an AIO to indicate
that the AIO should not processed asynchronously on a taskq. This
can be used to enhance performance in some cases, but it can also
be used to permit an AIO be destroyed from a completion callback.
(For the latter, the callback must execute the new nni_aio_fini_cb()
routine, which destroys the AIO without waiting for it to finish.)
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
We allocate AIO structures dynamically, so that we can use them
abstractly in more places without inlining them. This will be used
for the ZeroTier transport to allow us to create operations consisting
of just the AIO. Furthermore, we provide accessors for some of the
aio members, in the hopes that we will be able to wrap these for
"safe" version of the AIO capability to export to applications, and
to protocol and transport implementors.
While here we cleaned up the protocol details to use consistently
shorter names (no nni_ prefix for static symbols needed), and we
also fixed a bug in the surveyor code.
|
| |
|
|
|
|
|
|
|
| |
This moves the DNS related functionality into common code, and also
removes all the URL parsing stuff out of the platform specific code
and into the transports. Now the transports just take sockaddr's on
initialization. (We may want to move this until later.)
We also add UDP resolution as another separate API.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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.
|
| |
|
|
|
|
| |
Apparently there are circumstances when a pipedesc may get orphaned form the
pollq. This triggers an assertion failure when it occurs. I am still
trying to understand how this can occur. Stay tuned.
|
| |
|
|
|
|
|
|
|
|
|
| |
We have seen leaks of pipes causing test failures (e.g. the Windows
IPC test) due to EADDRINUSE. This was caused by a case where we
failed to pass the pipe up because the AIO had already been canceled,
and we didn't realize that we had oprhaned the pipe. The fix is to
add a return value to nni_aio_finish, and verify that we did finish
properly, or if we did not then we must free the pipe ourself. (The
zero return from nni_aio_finish indicates that it accepts ownership
of resources passed via the aio.)
|
| |
|
|
|
|
|
|
| |
We closed a few subtle races in the AIO subsystem as well, and now
we were able to eliminate the separate timer handling the MQ code.
There appear to be some opportunities to further enhance the code
for MQs as well -- eventually probably the only access to MQs will
be with AIOs.
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This eliminates the two threads per pipe that were being used to provide
basic I/O handling, replacing them with a single global thread for now,
that uses poll and nonblocking I/O. This should lead to great scalability.
The infrastructure is in place to easily expand to multiple polling worker
threads. Some thought needs to be given about how to scale this to engage
multiple CPUs. Horizontal scaling may also shorten the poll() lists easing
C10K problem.
We should look into better solutions than poll() for platforms that have
them (epoll on Linux, kqueue on BSD, and event ports on illumos).
Note that the file descriptors start out in blocking mode for now, but
then are placed into non-blocking mode. This is because the negotiation
phase is not yet callback driven, and so needs to be synchronous.
|
| |
|
|
|
|
|
| |
Transport-level pipe initialization is now sepearate and explicit.
The POSIX send/recv logic still uses threads under the hood, but
makes use of the AIO framework for send/recv. This is a key stepping
stone towards enabling poll() or similar async I/O approaches.
|
| | |
|
| | |
|
| | |
|
| | |
|
|
|
We will still need some kind of specific handling of cancellation for
msg queues, but it will be simpler to just implement that for the queues,
and not worry about cancellation in the general case around poll etc.
(The low level poll and I/O routines will get notified by their underlying
transport pipes/descriptors closing.)
|
