2.6.37.6 2.6.38.2 tst-aio4.c

2.6.37.6 2.6.38.2 tst-aio4.c

[Klaus Dittrich]

Reverting the patch of kernel/signal.c which was
applied to 2.6.37.6 and  2.6.38.2 solved the problem.

Re: 2.6.37.6 2.6.38.2 tst-aio4.c

[Roland Dreier]

On Mon, Mar 28, 2011 at 11:05 AM, Klaus Dittrich <kladit@arcor.de> wrote:
> Reverting the patch of kernel/signal.c which was
> applied to 2.6.37.6 and  2.6.38.2 solved the problem.

Indeed, I found that the same patch was to blame.  Fix coming shortly.

Re: 2.6.37.6 2.6.38.2 tst-aio4.c

[Jiri Slaby]

On 03/28/2011 08:14 PM, Roland Dreier wrote:
> On Mon, Mar 28, 2011 at 11:05 AM, Klaus Dittrich <kladit@arcor.de> wrote:
>> Reverting the patch of kernel/signal.c which was
>> applied to 2.6.37.6 and  2.6.38.2 solved the problem.
> 
> Indeed, I found that the same patch was to blame.  Fix coming shortly.

Is there a patch for this already?

thanks,
-- 
js

Re: 2.6.37.6 2.6.38.2 tst-aio4.c

[Roland Dreier]

> Is there a patch for this already?

Yes, it is 243b422af9ea in Linus's tree.

Re: 2.6.37.6 2.6.38.2 tst-aio4.c

[Jiri Slaby]

On 03/29/2011 11:07 PM, Roland Dreier wrote:
>> Is there a patch for this already?
> 
> Yes, it is 243b422af9ea in Linus's tree.

Thanks. (Queued up for suse kernels.)

-- 
js

2.6.37.6 2.6.38.2 tst-aio4.c

[Klaus Dittrich]

[-- Attachment #1: Type: text/plain, Size: 171 bytes --]

With both kernels, 2.6.37.6 and 2.6.38.2 tst-aio4.c of glibc fails.

It successfully completes with 2.6.37.5 and 2.6.38.1 and before.

I appended the code for convinice.


[-- Attachment #2: tst-aio4.c --]
[-- Type: text/plain, Size: 3536 bytes --]

/* Test for completion signal handling.
   Copyright (C) 2000,01,02 Free Software Foundation, Inc.
   This file is part of the GNU C Library.

   The GNU C Library is free software; you can redistribute it and/or
   modify it under the terms of the GNU Lesser General Public
   License as published by the Free Software Foundation; either
   version 2.1 of the License, or (at your option) any later version.

   The GNU C Library is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
   Lesser General Public License for more details.

   You should have received a copy of the GNU Lesser General Public
   License along with the GNU C Library; if not, write to the Free
   Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
   02111-1307 USA.  */

#include <aio.h>
#include <signal.h>
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <errno.h>

/* We might need a bit longer timeout.  */
#define TIMEOUT 10 /* sec */

int my_signo;

volatile sig_atomic_t flag;


static void
sighandler (const int signo)
{
  flag = signo;
  printf ("signo = %d\n",signo);
}

static int
wait_flag (void)
{
  while (flag == 0)
    {
      puts ("Sleeping...");
      sleep (1);
    }

  if (flag != my_signo)
    {
      printf ("signal handler received wrong signal, flag is %d\n", flag);
      return 1;
    }

  return 0;
}

#ifndef SIGRTMIN
# define SIGRTMIN -1
# define SIGRTMAX -1
#endif

static int
do_test (int argc, char *argv[])
{
  char name[] = "/tmp/aio4.XXXXXX";
  int fd;
  struct aiocb *arr[1];
  struct aiocb cb;
  static const char buf[] = "Hello World\n";
  struct aioinit init = {10, 20, 0};
  struct sigaction sa;
  struct sigevent ev;

  if (SIGRTMIN == -1)
  {
      printf ("RT signals not supported.\n");
      return 0;
  }

  /* Select a signal from the middle of the available choices... */
  my_signo = (SIGRTMAX + SIGRTMIN) / 2;

  fd = mkstemp (name);
  if (fd == -1)
    {
      printf ("cannot open temp name: %m\n");
      return 1;
    }

  unlink (name);

  /* Test also aio_init.  */
  aio_init (&init);

  arr[0] = &cb;

  cb.aio_fildes = fd;
  cb.aio_lio_opcode = LIO_WRITE;
  cb.aio_reqprio = 0;
  cb.aio_buf = (void *) buf;
  cb.aio_nbytes = sizeof (buf) - 1;
  cb.aio_offset = 0;
  cb.aio_sigevent.sigev_notify = SIGEV_SIGNAL;
  cb.aio_sigevent.sigev_notify_function = NULL;
  cb.aio_sigevent.sigev_notify_attributes = NULL;
  cb.aio_sigevent.sigev_signo = my_signo;
  cb.aio_sigevent.sigev_value.sival_ptr = NULL;

  ev.sigev_notify = SIGEV_SIGNAL;
  ev.sigev_notify_function = NULL;
  ev.sigev_notify_attributes = NULL;
  ev.sigev_signo = my_signo;

  sa.sa_handler = sighandler;
  sigemptyset (&sa.sa_mask);
  sa.sa_flags = SA_RESTART;

  if (sigaction (my_signo, &sa, NULL) < 0)
    {
      printf ("sigaction failed: %m\n");
      return 1;
    }

  flag = 0;
  /* First use aio_write.  */
  if (aio_write (arr[0]) < 0)
    {
      if (errno == ENOSYS)
	{
	  puts ("no aio support in this configuration");
	  return 0;
	}
      printf ("aio_write failed: %m\n");
      return 1;
    }

  if (wait_flag ())
    return 1;

  puts ("aio_write OK");

  flag = 0;
  /* Again with lio_listio.  */
  if (lio_listio (LIO_NOWAIT, arr, 1, &ev) < 0)
    {
      printf ("lio_listio failed: %m\n");
      return 1;
    }

  if (wait_flag ())
    return 1;

  puts ("all OK");

  return 0;
}

int main(int argc, char *argv[]) {
	do_test (argc, argv);
}

Re: 2.6.37.6 2.6.38.2 tst-aio4.c

[Roland Dreier]

On Mon, Mar 28, 2011 at 10:17 AM, Klaus Dittrich <kladit@arcor.de> wrote:
> With both kernels, 2.6.37.6 and 2.6.38.2 tst-aio4.c of glibc fails.
>
> It successfully completes with 2.6.37.5 and 2.6.38.1 and before.
>
> I appended the code for convinice.

I took a quick look at this, and indeed the program succeeds on 2.6.38
but fails on the latest tip of Linus's tree.  I'm pretty sure that the obvious
place to look, my patch "aio: wake all waiters when destroying ctx" is not
to blame, both because the patch is obviously correct :) and also because
this test is using the glibc aio implementation, which doesn't use the
kernel's aio code at all.

Indeed, an strace of the program shows that on old kernels, the child thread
does

rt_sigqueueinfo(16143, SIGRT_17, {si_signo=SIGRT_17,
si_code=SI_ASYNCIO, si_pid=16143, si_uid=1000, si_value={int=0,
ptr=0}}) = 0

while on the new kernel, it gets EPERM:

rt_sigqueueinfo(17198, SIGRT_17, {si_signo=SIGRT_17,
si_code=SI_ASYNCIO, si_pid=17198, si_uid=1000, si_value={int=0,
ptr=0}}) = -1 EPERM (Operation not permitted)

I'll take a look to see which -stable patch might be to blame.

 - R.