process file descriptor limit handling
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
The current kernel (and all before as far as I can see) have a problem with the file system limit handling. The behavior does not conform to the current POSIX spec.
Look at the attached example code. The programs opens descriptors up to the limit, then lowers the limit, closes a descriptor below the number given to the last setrlimit call, and tries to open a descriptor. This currently succeeds.
The problem is that RLIMIT_NOFILE is defined as the number of open descriptors. In the example case, before the final open call, there are N1-1 open descriptors, with RLIMIT_NOFILE set to N2 (which is << N1). The open call should fail. Another, more common solution, is to have the setrlimit call fail in case the new limit is larger than the largest file descriptor in use. Returning EINVAL in this case is just fine and is apparently what other platforms do.
One could also take the position that the current behavior has it's advantages. A program could open all the file descriptors it needs, and then close a few which can be used to open some files for the normal mode of operation. Possible, maybe even quite secure, but still smells a bit like a hack.
It might also be that some wording is getting in the specification which will allow the current kernel behavior to continue to exist. More through a loophole, but still.
Anyway, I think the existing behavior is probably an oversight. Whether it is worth keeping is a question somebody (= Linus) will have to answer. My recommendation is probably, to make setrlimit fail.
- --
➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖ -----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.3 (GNU/Linux)
iD8DBQE/vRvS2ijCOnn/RHQRAl4HAKDK3xp5vydo1bqfDNvVZUCxObukMQCeI/1l xFnOkWYL4iXmzqjVuIXfYWI=
=ftW/
-----END PGP SIGNATURE-----
#include <errno.h>
#include <fcntl.h>
#include <stdio.h>
#include <unistd.h>
#include <sys/resource.h>
int
main (void)
{
#define N1 100
struct rlimit r;
r.rlim_cur = N1;
r.rlim_max = N1;
setrlimit (RLIMIT_NOFILE, &r);
int i;
int fd[N1];
for (i = 0; ; ++i)
if ((fd[i] = open ("/dev/null", 0)) < 0)
{
if (i == N1 - 3)
printf ("fine, %d file descriptors open\n", N1);
else
{
puts ("*** make sure the parent doesn't open any descriptors other than 0, 1, 2");
return 1;
}
break;
}#define N2 50
r.rlim_cur = N2;
r.rlim_max = N2;
int e = setrlimit (RLIMIT_NOFILE, &r);
if (e == EINVAL)
{
puts ("good, setrlimit sees the open descriptors");
return 0;
}
getrlimit (RLIMIT_NOFILE, &r);
if (r.rlim_cur != N2 || r.rlim_max != N2)
{
puts ("getrlimit returned different values");
return 1;
}
close (fd[N2 - 4]);
fd[N2 - 4] = open ("/dev/null", 0);
if (fd[N2 - 4] != -1)
{
printf ("opening a new descriptor succeeded even though %d are open and the limit is %d\n",
N1 - 1, N2);
return 1;
} return 0;
}
Quoted 6 lines · * Ulrich Drepper ([email protected]) wrote:
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1The current kernel (and all before as far as I can see) have a problem with the file system limit handling. The behavior does not conform to the current POSIX spec.
<snip>
Quoted 3 lines
It might also be that some wording is getting in the specification which will allow the current kernel behavior to continue to exist. More through a loophole, but still.
This seems the case. SuS v3 says:
setrlimitRLIMIT_NOFILE
This is a number one greater than the maximum value that the system may assign to a newly-created descriptor. If this limit is exceeded, functions that allocate a file descriptor shall fail with errno set to [EMFILE]. This limit constrains the number of file descriptors that a process may allocate.
open
[EMFILE]
{OPEN_MAX} file descriptors are currently open in the calling process.
limits.h
{OPEN_MAX}
Maximum number of files that one process can have open at any one time. Minimum Acceptable Value: {_POSIX_OPEN_MAX}
So, one view says your test program is within the spec, since the new fd is still one less than the current rlimit.
Anyway, here's a simple patch that would fail the second setrlimit, as you suggested.
thanks,
-chris
Signature · 1 line
Linux Security Modules http://lsm.immunix.org http://lsm.bkbits.net