This is the mail archive of the cygwin-xfree mailing list for the Cygwin XFree86 project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: /dev/windows and select() [was Re: Slow response to keypresses in xorg-server-1.8.0-1]

On 30/08/2010 12:23, Corinna Vinschen wrote:
On Aug 29 14:39, Jon TURNEY wrote:
On 08/08/2010 12:04, Andy Koppe wrote:
On 7 August 2010 23:07, Jon TURNEY wrote:
Hmmm, looking again at the implementation of select(), I don't immediately
see that when waiting on /dev/windows, it checks that the message queue has
old messages on it before waiting.  The MSDN documentation for
MsgWaitForMultipleObjects() seems to says that messages which had arrived
before the last PeekMessage() etc. aren't considered new and so don't end
the wait?

I think you're right, a call to PeekMessage is needed for proper select() semantics: it shouldn't block if data is available for reading.

Attached is a small test-case which seems to demonstrate this problem.

Run ./dev-windows-select-test and observe select() blocks for the
full timeout, despite the fact that the /dev/windows fd is ready for
reading (and it reported as such as the end of the timeout)

If you run './dev-windows-select-test -skip' to skip the
PeekMessage(), select() returns immediately, indicating the
/dev/windows fd is ready for reading.

Again, thanks for the testcase. I applied a patch to Cygwin which should make select on /dev/windows fully functional.

The blessed solution is to replace the call to MsgWaitForMultipleObjects
with a call to MsgWaitForMultipleObjectsEx with the MWMO_INPUTAVAILABLE
flag set.  This flag is defined to do exactly what we need.

The only downside is that this flag does not exist on NT4 and its usage
results in an "invalid argument" error.  So, for NT4, I added the
workaround I described in my yesterday's soliloquy.

I'm planning to release Cygwin 1.7.7 tomorrow at the latest, so please
give it a test as soon as possible.  Here's a binary DLL for testing
(build w/o optimization, so it's probably a bit slow):
   (md5sum: 7e07fd9eafd021697a0861c1ae4fa94e)

Thanks Corinna :-)

I tried that cygwin DLL with my test case, and with an X server with what I now realize is the workaround I'd applied reverted [1] and it seems to work fine.


Unsubscribe info:
Problem reports:

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]