tst-aio - new version - reveals a bug in glibc

[Andreas Jaeger]

>>>>> Ulrich Drepper writes:

 > Andreas Jaeger <aj@suse.de> writes:
>> here're new tests for aio.  The second aio_cancel call fails with:
>> Didn't expect signal from child: got `Segmentation fault'

 > That was a show.  Thanks for bringing this up that late.  So I had to
 > stay up and fix it.  Oh well, who needs more than 5 hours of sleep.
Sorry, I send it first thing in the morning ;-).

 > The AIO code contained lots of bugs.  I cannot believe nobody hit them
 > before.  The cancelation handling was completely broken.
I've remember once seeing a report that AIO was broken but never got a
real testcase or explanation - and didn't bother myself at that time :-(.

 > But also your test patch had some bugs.  Please look at the version I
 > checked in.
Thanks for fixing this, it looks fine.

 > Meanwhile I'm a little bit more confident that implementation is not
 > completely broken.  There is one strange thing though: reading the
At least we're now testing all functions - but not in all possible
situations yet.  If somebody likes to contribute another test program,
you're welcome.
 > current XPG6 draft it says for aio_cancel() that no fields but
 > aio_fildes and aio_sigevent are used.  This is strange since at least
 > aio_reqprio should be used as well.  Otherwise queuing the request is
 > difficult.  This wording is also not present in POSIX-1.

Andreas
-- 
 Andreas Jaeger
  SuSE Labs aj@suse.de
   private aj@arthur.inka.de