Postfix 'local' crashes

[Albrecht Dreß]

Binary package hint: postfix

Ubuntu version: Ubuntu 11.04
Package version: postfix 2.8.2-1ubuntu1
System: PowerMac G4 "Silver"

After an update of my box from 10.10 to 11.04, the 'local' delivery agent of Postfix stopped working. I added verbose debugging for it in /etc/postfix/master.cf, and get the following results in /var/log/mail.log:

May 1 14:34:23 antares postfix/local[8296]: name_mask: ipv4
May 1 14:34:23 antares postfix/local[8296]: inet_addr_local: configured 2 IPv4 addresses
May 1 14:34:23 antares postfix/local[8296]: process generation: 20 (20)
May 1 14:34:23 antares postfix/local[8296]: Compiled against Berkeley DB: 4.8.30?
May 1 14:34:23 antares postfix/local[8296]: Run-time linked against Berkeley DB: 4.8.30?
May 1 14:34:23 antares postfix/local[8296]: dict_open: hash:/etc/aliases
May 1 14:34:23 antares postfix/local[8296]: match_string: fast_flush_domains ~? debug_peer_list
May 1 14:34:23 antares postfix/local[8296]: match_string: fast_flush_domains ~? fast_flush_domains
May 1 14:34:23 antares postfix/local[8296]: set_eugid: euid 112 egid 124
May 1 14:34:23 antares postfix/local[8296]: name_mask: alias
May 1 14:34:23 antares postfix/local[8296]: name_mask: forward
May 1 14:34:23 antares postfix/local[8296]: name_mask: alias
May 1 14:34:23 antares postfix/local[8296]: name_mask: forward
May 1 14:34:23 antares postfix/local[8296]: name_mask: canonical
May 1 14:34:23 antares postfix/local[8296]: name_mask: virtual
May 1 14:34:23 antares postfix/local[8296]: name_mask: command
May 1 14:34:23 antares postfix/local[8296]: name_mask: file
May 1 14:34:23 antares postfix/local[8296]: name_mask: forward
May 1 14:34:23 antares postfix/local[8296]: name_mask: fcntl
May 1 14:34:23 antares postfix/local[8296]: name_mask: dotlock
May 1 14:34:23 antares postfix/local[8296]: fatal: unable to determine open file limit
May 1 14:34:24 antares postfix/master[8124]: warning: process /usr/lib/postfix/local pid 8296 exit status 1
May 1 14:34:24 antares postfix/master[8124]: warning: /usr/lib/postfix/local: bad command startup -- throttling

Observed effects: see above
Expected results: 'local' delivery should work.
---
Architecture: powerpc
DistroRelease: Ubuntu 11.04
Package: postfix 2.8.2-1ubuntu2
PackageArchitecture: powerpc
ProcEnviron:
 LANGUAGE=de_DE:en
 PATH=(custom, no user)
 LANG=de_DE.UTF-8
 SHELL=/bin/bash
Tags: natty
Uname: Linux 2.6.36.2 ppc
UpgradeStatus: Upgraded to natty on 2011-04-30 (6 days ago)
UserGroups:

[Clint Byrum (clint-fewbar)]

Albrecht, I'm getting no such failure on my amd64 11.04 machines, I wonder if this is a problem for powerpc only.

[Clint Byrum (clint-fewbar)]

Albrecht, it would help us determine the status of this bug if you could run

sudo apport-collect 774852

in a terminal

Thanks!

[Albrecht Dreß (albrecht-dress)]

apport information

[Albrecht Dreß (albrecht-dress)]

Hi Clint:

Asking Google about this issue, I found some complaints from Gentoo users related to glibc 2.13 and pre-linking [1]. However, the prelink package is not installed on my box, and installing it doesn't change anything. It might have been a build issue, though.

An other source might be the fact that the PowerPC is a big-endian, but I doubt that this is the source, as I *never* had any problem with Postfix in the past, neither with Ubuntu, nor with Yellowdog or Fedora I used in the past.

Thanks, Albrecht

[1] <http://www.google.com/search?client=ubuntu&channel=fs&q=postfix+local+fatal%3A+%22unable+to+determine+open+file+limit%22&ie=utf-8&oe=utf-8>

[Albrecht Dreß (albrecht-dress)]

I re-compiled the package with some extra information (added strerror(errno) at varoius places). The reason for the failing 'local' process is in src/util/open_limit.c, where the call to getrlimit (yes, *not* setrlimit!) fails with EPERM. That's really strange!

BTW, I don't have SElinux enabled...

[Patrick van Oostrom (patrick-meukinc)]

I am having the same issues on my boxes, running i386

This is what I found in the mail.log

May 12 08:45:50 vps594 postfix/postdrop[6857]: fatal: getrlimit: Operation not permitted
May 12 08:45:51 vps594 postfix/sendmail[6856]: warning: command "/usr/sbin/postdrop -r" exited with status 1
May 12 08:45:51 vps594 postfix/sendmail[6856]: fatal: patrick(1000): unable to execute /usr/sbin/postdrop -r: Success
May 12 08:46:04 vps594 postfix/local[6861]: fatal: unable to determine open file limit
May 12 08:46:05 vps594 postfix/master[3986]: warning: process /usr/lib/postfix/local pid 6861 exit status 1
May 12 08:46:05 vps594 postfix/master[3986]: warning: /usr/lib/postfix/local: bad command startup -- throttling

[Albrecht Dreß (albrecht-dress)]

Side note: the following code runs fine on my box:

<code>
#include <sys/resource.h>
#include <stdio.h>

int main(int argc, char **argv)
{
    struct rlimit rl;
    int result;

    result = getrlimit(RLIMIT_NOFILE, &rl);
    printf("%d -> %d %d\n", result, (int) rl.rlim_cur, (int) rl.rlim_max);
    return 0;
}
</code>

Running it as user, the result is

albrecht@antares:~/ubuntu$ ./test_lim
0 -> 1024 4096

Strange!!

[Serge Hallyn (serge-hallyn)]

Fascinating. The getrlimit kernel code should never return -EPERM, nor lead to any selinux or apparmor calls fwiw. I'd be curious to see an strace of local. If in fact a call to the kernel/sys.c:getrlimit() is returning -EPERM, then the arguments passed to do_prlimit are being corrupted along the way.

More likely would seem to be a problem with glibc. We could try an instrumented kernel to find out more.

Is this 32- or 64-bit?

[Serge Hallyn (serge-hallyn)]

Marking this 'high' importance as I think it falls into 'Has a severe impact on a small portion of Ubuntu users'

[Serge Hallyn (serge-hallyn)]

Tried to reproduce this on i386, but couldn't.

Albrecht, it's a long shot, but could you try this modified getrlimit test program? (replace 112 and 105 with the gid and uid for postfix, respectively, if they are different on your system).

#include <sys/resource.h>
#include <stdio.h>
#include <stdlib.h>

int main(int argc, char **argv)
{
    struct rlimit rl;
    int result;

    result = setresgid(0,112,0);
    printf("setresgid result is %d\n", result);
    result = setresuid(0,105,0);
    printf("setresuid result is %d\n", result);
    result = getrlimit(RLIMIT_NOFILE, &rl);
    printf("%d -> %d %d\n", result, (int) rl.rlim_cur, (int) rl.rlim_max);
    return 0;
}

[Albrecht Dreß (albrecht-dress)]

Hi Serge:

Thanks a lot for your comment - the getrlimit() behaviour is indeed puzzling!

Having a closer look at my system, I believe I found the reason, though. The Ubuntu upgrade from 10.10 to 11.04 failed to change the bootloader config (yaboot on Mac's) to boot the kernel coming with the new release. As a result, I was still running 2.6.36.2 (self-compiled, as I had to remove some graphics stuff - the Mac shipped with a special Apple OEM ATI card, which didn't work with the stock kernel; different story). Sorry, but I really forgot to look at the proper kernel version!

I manually changed yaboot.conf, rebooted - and now postfix runs again, *without* the getrlimit issue! This looks like a glibc vs. kernel incompatibility. Maybe this is also the reason while Patrick van Oostrom (#6) has similar effects on i386? Interesting that *no* other application showed similar effects.

From my pov, you could close this one (Tag "user too stupid" :-/ )...

Thanks again for your help,
Albrecht.

[Scott Kitterman (kitterman)]

Marking invalid based on reporter feedback.