Re: (stat(...) == -1 || faccessat(...) == -1) && errno == EINTR ?!??

[Tobias Bading <tbading@web.de>]

[-- Attachment #1: Type: text/plain, Size: 2496 bytes --]

Hello again.

I've been able to reproduce the problem with the attached program. With
SIGALRMs firing 10 times per second, I get maybe a dozen "handler
called" lines before stat() or faccessat() fails with errno EINTR. When
I increase the rate to 50 SIGALRMs per second, the very first stat()
fails in every test run.

Specifying SA_RESTART in sigaction() has no effect.

Unfortunately, I don't have any other network shares available at the
moment to test whether only CIFS through VPN is affected, or the problem
would occur with e.g. NFS or CIFS without a VPN as well.

Tobias

---

On 14.02.21 12:13, Tobias Bading wrote:
> Hi.
>
> A few days ago I encountered strange problems at home while using GNU
> Emacs on GNU/Linux with Windows SMB shares in the office automounted
> through the company's VPN.
>
> Details and my current findings on the Emacs devel mailing list:
> https://lists.gnu.org/archive/html/emacs-devel/2021-02/msg00835.html.
>
> TL;DR:
> Emacs' stat() and faccessat() calls sometimes return -1 and set errno
> to EINTR (according to strace). Most of the time the interruption
> seems to come from a SIGIO, but at least once a SIGCHLD was in the mix
> as well.
>
> The behavior of returning with errno == EINTR seems to be new and/or
> some kind of bug, because otherwise Emacs would handle that case with
> TEMP_FAILURE_RETRY() or something to that effect, right? I've used
> pretty much the same Emacs in the office (on an older GNU/Linux
> release) for ages with the same Windows SMB shares and never had this
> problem.
>
> The GNU/Linux man pages of stat() and faccessat() don't mention EINTR
> at all, and neither does POSIX
> (https://pubs.opengroup.org/onlinepubs/9699919799/).
>
> So I guess the one million dollar question is this:
> Are stat() and faccessat() permitted to return -1 with errno == EINTR,
> or does that violate the POSIX (or some other) spec?
>
> I'll probably write a small test program next to check if I can
> reproduce this problem outside of Emacs, maybe with a simple SIGALRM
> and an endless loop of stat()/faccessat() calls. If that reproduces
> the problem it would be nice to know whether using SA_RESTART in
> sigaction() has any effect.
>
> Any ideas are very welcome... :)
>
> Tobias
>
> PS: I'm running Ubuntu focal (20.04.2 LTS, amd64), which currently
> uses libc-bin 2.31-0ubuntu9.2, kernel linux-image-5.4.0-64-generic and
> openconnect 8.05-1 to access the VPN.
>


[-- Attachment #2: dont-interrupt-me.c --]
[-- Type: text/x-csrc, Size: 1104 bytes --]

#include <fcntl.h>
#include <signal.h>
#include <stdio.h>
#include <string.h>
#include <sys/stat.h>
#include <sys/time.h>
#include <sys/types.h>
#include <unistd.h>

static const char path[] = "/smb/server/share/dir/subdir";

void handler (int signum)
{
  static const char t[] = "handler called\n";
  write (1, t, sizeof t);
}

int main ()
{
  struct stat st;
  struct sigaction action;
  struct itimerval timer;

  memset (&action, 0, sizeof action);
  action.sa_handler = handler;
  action.sa_flags = SA_RESTART;
  if (sigaction (SIGALRM, &action, NULL))
  {
    perror ("sigaction() failed");
    return 1;
  }

  timer.it_value.tv_sec  = timer.it_interval.tv_sec  = 0;
  timer.it_value.tv_usec = timer.it_interval.tv_usec = 1000000 / 10;
  if (setitimer (ITIMER_REAL, &timer, NULL))
  {
    perror ("setitimer() failed");
    return 1;
  }

  for (;;)
  {
    int r = stat (path, &st);
    if (r)
    {
      perror ("stat() failed");
      return 1;
    }

    r = faccessat (AT_FDCWD, path, R_OK, 0);
    if (r)
    {
      perror ("faccessat() failed");
      return 1;
    }
  }

  return 0;
}