2.6.37.6 2.6.38.2 tst-aio4.c

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.

/* 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);
}

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.