Re: Unexpected EINVAL from pthread_join

[Corinna Vinschen <corinna-cygwin at cygwin dot com>]

Hi Lasse,

On Feb 24 20:37, Lasse Collin wrote:
> On 2015-02-24 Corinna Vinschen wrote:
> > On Feb 24 15:54, Lasse Collin wrote:
> > > Many other pthread functions are similar in sense that they must
> > > never return EINTR. A bug similar to the one in pthread::join exist
> > > in pthread_mutex::lock. If SA_RESTART isn't used, signals can make
> > > multiple threads get a lock on the same mutex at the same time. A
> > > test program is attached. Adding cw_sig_restart to the cygwait call
> > > in pthread_mutex::lock fixes this.
> > 
> > Can you collect the info which functions are affected so that lazy me
> > just has to apply the cw_sig_restart patches in bulk?
> 
> I grepped for cygwait uses earlier but I don't understand the code well
> enough to be sure about anything. The following is pretty much
> guessing, but hopefully there's something useful still.
> 
> These look the most suspicious to me, possibly needing cw_sig_restart:
> 
>     thread.h          fast_mutex::lock
>     fhandler_tape.cc  fhandler_dev_tape::_lock
> 
> When looking for things needing cw_sig_restart, I started to wonder if
> cw_sig_eintr is used properly in all cases. Often the caller of
> cygwait + cw_sig_eintr will also conditionally call
> _my_tls.call_signal_handler, but the following functions don't. The
> comment in handle_sigsuspend makes me think that this may be a false
> alarm but I mention these just in case.
> 
>     exceptions.cc     handle_sigsuspend
>     posix_ipc.cc      ipc_mutex_lock
>     signal.cc         clock_nanosleep
>     signal.cc         sigwaitinfo
>     thread.cc         pthread_cond::wait
>     thread.cc         semaphore::_timedwait
>     thread.cc         semaphore::_wait

The signal code has gone through a couple of iterations so there's a
chance that some places do useless stuff; detritus from former
iterations.

> In the following functions the situation is kind of reversed. Those
> functions call _my_tls.call_signal_handler even though cw_sig_eintr
> wasn't specified. cygwait will already have called the signal handler
> since cw_sig has been specified via cygwait(DWORD). So I guess in these
> functions the signal handler might get called twice.
> 
>     fhandler_dsp.cc   fhandler_dev_dsp::Audio_out::waitforspace
>     fhandler_dsp.cc   fhandler_dev_dsp::Audio_in::waitfordata

Cool, thanks for looking so closely.  I wasn't expecting that, rather
just a look through the pthread functions but this is very helpful.
I'll check these code snippets in the next couple of days, maybe even
tomorrow.


Thanks again,
Corinna

-- 
Corinna Vinschen                  Please, send mails regarding Cygwin to
Cygwin Maintainer                 cygwin AT cygwin DOT com
Red Hat

Attachment: pgpC2XCVo2FbS.pgp
Description: PGP signature