[PATCH] Fix tst-aio7.

[PATCH] Fix tst-aio7.

[Martin Schwidefsky]

Hi,
the tst-aio7 testcase sometimes fails for s390-64. After some
debugging I found out why. The aio_suspend test starts
a read on aio buffer cb1. This read never completes.
The aio buffer is still enqueued when the block that
contains the aiocb structure is left. The exit processing
will cancel the i/o and the exit status will be written to the
stack which is reused for a stack frame in the meantime.
This sometimes segfaults on s390-64, sometimes it works.
The fix is simple: cancel the i/o before leaving the block
that defines cb1.

blue skies,
  Martin.

2003-03-06  Martin Schwidefsky  <schwidefsky@de.ibm.com>

	* rt/tst-aio7.c (do_test): Cancel i/o on cb1.

diff -urN libc/rt/tst-aio7.c libc-aiofix/rt/tst-aio7.c
--- libc/rt/tst-aio7.c	Mon Sep 30 09:19:21 2002
+++ libc-aiofix/rt/tst-aio7.c	Thu Mar  6 10:38:00 2003
@@ -180,6 +180,14 @@
 	puts ("aio_suspend([done,blocked],2,3) suspended thread");
 	++result;
       }
+
+    /* Cancel i/o on cb1. */
+    r = aio_cancel (piped[0], &cb1);
+    if (r != AIO_CANCELED)
+      {
+	puts ("aio_cancel did not return AIO_CANCELED");
+	++result;
+      }
   }
 
   return result;

Re: [PATCH] Fix tst-aio7.

[Roland McGrath]

Thanks a lot!  I've put that fix in.  I think we've seen occasional
tst-aio7 failures on multiple platforms, and this certainly explains it.


Thanks,
Roland

Re: [PATCH] Fix tst-aio7.

[Ulrich Drepper]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Roland McGrath wrote:
> Thanks a lot!  I've put that fix in.  I think we've seen occasional
> tst-aio7 failures on multiple platforms, and this certainly explains it.

No, this does not sound right.  This works around the symptoms.

- -- 
- --------------.                        ,-.            444 Castro Street
Ulrich Drepper \    ,-----------------'   \ Mountain View, CA 94041 USA
Red Hat         `--' drepper at redhat.com `---------------------------
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.1 (GNU/Linux)

iD8DBQE+Z6Ju2ijCOnn/RHQRAodFAKDJrUDw3nUbGt5iPgE4qn5CAdhnXACgyoj5
tMw4oeZ6pnnSWmFpUGqqo1Q=
=WMP/
-----END PGP SIGNATURE-----

Re: [PATCH] Fix tst-aio7.

[Roland McGrath]

> No, this does not sound right.  This works around the symptoms.

How so?  The test looks clearly wrong to me.  aio_suspend will return when
the CB0 request is done, and the CB1 request is still outstanding.  Unless
we synchronize with that request, CB1 is still in the aio data structures
when it is no longer live on the stack.

Re: [PATCH] Fix tst-aio7.

[Ulrich Drepper]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Roland McGrath wrote:

> How so?  The test looks clearly wrong to me.  aio_suspend will return when
> the CB0 request is done, and the CB1 request is still outstanding.  Unless
> we synchronize with that request, CB1 is still in the aio data structures
> when it is no longer live on the stack.

Then these variables have to have global scope.

The test in question was also testing that the AIO handling at exit time
is leak-free.  There is not explicit test but one easily sees system
resources being depleated.  This is important once we have a
kernel-supported AIO implementation.

- -- 
- --------------.                        ,-.            444 Castro Street
Ulrich Drepper \    ,-----------------'   \ Mountain View, CA 94041 USA
Red Hat         `--' drepper at redhat.com `---------------------------
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.1 (GNU/Linux)

iD8DBQE+Z6gj2ijCOnn/RHQRAiM/AKCMnrRh33tGLQtD4OX/25RZUY/YPACgof8+
1u4fOGGrnW8cD7ftE0vbQ14=
=FNCu
-----END PGP SIGNATURE-----

Re: [PATCH] Fix tst-aio7.

[Roland McGrath]

> Then these variables have to have global scope.
> 
> The test in question was also testing that the AIO handling at exit time
> is leak-free.  There is not explicit test but one easily sees system
> resources being depleated.  This is important once we have a
> kernel-supported AIO implementation.

Ok.  I've reverted the aio_cancel change and instead given CB1 permanent
extent so that it's valid.  I put in a comment mentioning this rationale.