| Commit message (Collapse) | Author | Age |
|---|
| |
|
|
|
|
|
| |
This fixes one major problem, which was that if nni_fini() was called
once on Windows, it would not be further possible to call nni_init().
While here fixed a few compilation issues.
|
| |
|
|
|
| |
This creates a use-after-free bug if nni_fini() is run, then new
sockets are created.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
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.
|
| |
|
|
|
|
| |
This includes async send and recv, driven from the poller. This will
be requierd to support the underlying UDP and ZeroTier transports in
the future. (ZeroTier is getting done first.)
|
| | |
|
| |
|
|
|
|
|
| |
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.)
|
| | |
|
| |
|
|
|
|
| |
This logic leaves a race condition in the dial side, which will
be fixed with a subsequent change to convert that to fully asynchronous
as well.
|
| | |
|
| |
|
|
|
|
| |
This is only lightly tested, and I expect that there remain
some race conditions. Endpoint logic in particular needs
work.
|
| |
|
|
|
|
| |
We still have endpoint related races apparently; we need to examine
the possibility of handling endpoints much like we do pipes, which seem
to be race free.
|
| |
|
|
|
| |
We are still seeing likely errors with pipes outliving their associated
endpoints, so work is still needed here.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
The IOCP code has been refactored to improve reuse, and hopefully
will be easier to use with TCP now. Windows IPC using Named Pipes
is mostly working -- mostly because there is a gnarly close-race.
It seems that we need to take some more care to ensure that the
pipe is not released while requests may be outstanding -- so some
deeper synchronization between the IOCP callback logic and the
win_event code is needed. In short, we need to add a condvar to
the event, and notice when we have submitted work for async completion,
and make sure we flag the event "idle" after either completion or
cancellation of the event.
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
| |
Modern Windows (Vista and later) have light weight Slim Read/Write locks
which only occupy 64 bits, and don't require any memory allocation to
create.
While here clean up a few more unreferenced variables found with the
Microsoft compilers.
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
It turns out that I had to fix a number of subtle asynchronous
handling bugs, but now TCP is fully asynchronous. We need to
change the high-level dial and listen interfaces to be async
as well.
Some of the transport APIs have changed here, and I've elected
to change what we expose to consumers as endpoints into seperate
dialers and listeners. Under the hood they are the same, but
it turns out that its helpful to know the intended use of the
endpoint at initialization time.
Scalability still occasionally hangs on Linux. Investigation
pending.
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The connect & accept logic for IPC is now fully asynchronous.
This will serve as a straight-forward template for TCP. Note that
the upper logic still uses a thread to run this "synchronously", but
that will be able to be removed once the last transport (TCP) is made
fully async.
The unified ipcsock is also now separated, and we anticipate being
able to remove the posix_sock.c logic shortly. Separating out the
endpoint logic from the pipe logic helps makes things clearer, and
may faciliate a day where endpoints have multiple addresses (for
example with a connect() endpoint that uses a round-robin DNS list
and tries to run the entire list in parallel, stopping with the first
connection made.)
The platform header got a little cleanup while we were here.
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
| |
Note that we're going to refactor this again, for both TCP and
IPC, to actually push the endpoint abstraction further down
instead of using a combined "socket" abstraction. This may help
solve other problems, such as parallel outgoing connections.
Nonetheless, most of the work to make POSIX sockets fully async
is now done.
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
| | |
|
