This is the mail archive of the cygwin-developers@cygwin.com mailing list for the Cygwin 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: Many pthread failures in the test suite, one setgroup failure


On Mon, Sep 30, 2002 at 12:18:04AM +1000, Robert Collins wrote:
>On Mon, 2002-09-30 at 00:16, Christopher Faylor wrote:
>
>> I've uploaded a pthread-condvar6.exe.bz2 to the same location.  I assume
>> that this must have something to do with gcc 3.2, too, since it fails
>> so consistently for me.
>
>Ok, I'll pick it up and play with it tomorrow. It's bedtime for me now
>though.

AFAICT, the problem may be a known one:

/* FIXME: there's a potential race with PTHREAD_MUTEX_INITALIZER:
   the mutex is not actually inited until the first use.
   So two threads trying to lock/trylock may collide.
   Solution: we need a global mutex on mutex creation, or possibly simply
   on all constructors that allow INITIALIZER macros.
   the lock should be very small: only around the init routine, not
   every test, or all mutex access will be synchronised.  */

I don't know why it is being tickled now or why it is so consistent but
if I initialize the mutex in condvar6.c, I don't see the problem.  When
the program fails, it is returning 22 (EINVAL) from pthread_cond_timedwait.

  if (!pthread_mutex::isGoodObject (themutex))
    return EINVAL;

I couldn't quite catch what was setting the EINVAL, for some reason,
though.  I rebuilt the DLL with debugging and set some breakpoints on
likely "return EINVAL"s but they didn't get hit in this context.

I did see an occasional EINVAL, though, and it seemed to be happening at
__pthread_cond_dowait here:

  if (!pthread_mutex::isGoodObject (themutex))
    return EINVAL;

I couldn't tell if this was the result of a debugging artifact or if
it was an actual problem but in this context themutex == cond, which
certainly seems wrong.

This is unfortunately just speculation so far since I haven't been
able to catch what's going on.  It does seem to be timing related
since stopping and single stepping often seems to cause the program
to "just work".

Anyway, that's my current brain dump.

cgf


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