Message-ID: <cfd18e0f0804251305u4edb2bb1k290553965291c4fa@mail.gmail.com>
Date:	Fri, 25 Apr 2008 22:05:35 +0200
From:	"Michael Kerrisk" <mtk.manpages@...glemail.com>
To:	"Neil Horman" <nhorman@...driver.com>
Cc:	"Andi Kleen" <andi@...stfloor.org>, linux-man@...r.kernel.org,
	"Linux Kernel Mailing List" <linux-kernel@...r.kernel.org>,
	"Petr Gajdos" <pgajdos@...e.cz>, michael.kerrisk@...il.com,
	"Michael Kerrisk" <mtk.manpages@...il.com>
Subject: Re: core_pattern pipe documentation

On Fri, Apr 25, 2008 at 8:54 PM, Neil Horman <nhorman@...driver.com> wrote:
>
> On Fri, Apr 25, 2008 at 08:13:21PM +0200, Michael Kerrisk wrote:
> > On Fri, Apr 25, 2008 at 6:22 PM, Neil Horman <nhorman@...driver.com> wrote:
> > >
> > > On Fri, Apr 25, 2008 at 03:18:46PM +0200, Michael Kerrisk wrote:
> > > > Hi Neil,
> > > >
> > > > On Wed, Apr 23, 2008 at 4:59 PM, Neil Horman <nhorman@...driver.com> wrote:
> > > > >
> > > > > On Wed, Apr 23, 2008 at 02:09:14PM +0200, Michael Kerrisk wrote:
> > > > > > Andi -- ping!
> > > > > >
> > > > > > Adding Neil to CC, since it looks like he also did some work here, and
> > > > > > so can perhaps comment.
> > > > > >
> > > > > > On Fri, Apr 18, 2008 at 6:53 PM, Michael Kerrisk
> > > > > > <mtk.manpages@...glemail.com> wrote:
> > > > > > > Andi,
> > > > > > >
> > > > > > > I wrote the following description of the core_pattern pipe feature.  Does this
> > > > > > > seem okay?
> > > > > > >
> > > > > > >   Piping core dumps to a program
> > > > > > >       Since kernel 2.6.19, Linux supports an  alternate  syntax
> > > > > > >       for the /proc/sys/kernel/core_pattern file.  If the first
> > > > > > >       character of this file is a pipe  symbol  (|),  then  the
> > > > > > >       remainder  of  the line is interpreted as a program to be
> > > > > > >       executed.  Instead of being written to a disk  file,  the
> > > > > > >       core  dump  is  given  as  standard input to the program.
> > > > > > >       Note the following points:
> > > > > > >
> > > > > > >       *  The program must be specified using an absolute  path-
> > > > > > >          name  (or  a  pathname relative to the root directory,
> > > > > > >          /), and must immediately follow the '|' character.
> > > > > > >
> > > > > > >       *  The process created to run the program  runs  as  user
> > > > > > >          and group root.
> > > > > > >
> > > > > > >       *  Arguments can be supplied to the program, delimited by
> > > > > > >          white space (up to a total line length of 128  bytes).
> > > > > > >
> > > > > > > Cheers,
> > > > > > >
> > > > > > > Michael
> > > > > > >
> > > > > Thanks for CC'ing me.  The above all looks good.  I would add documentation
> > > > > however, about the available macros that can be used when core_pattern is
> > > > > specified as a pipe.  Adding something like the following would be good:
> > > > >
> > > > >        * Arguments can be statically declared or implied via the use of macros,
> > > > >          denoted by the use of the %sign.  The following macros are supported:
> > > > >                * %% - output a literal % sign on the command line
> > > > >                * %p - the pid of the crashing process
> > > > >                * %u - the uid of the crashing process
> > > > >                * %g - the gid of the crashing process
> > > > >                * %s - the signal that caused the crashing process to crash
> > > > >                * %t - the time the crashing process dumped
> > > > >                * %h - the hostname of the system
> > > > >                * %e - the executable name of the crashing process
> > > > >                * %c - the core limit size of the crashing process
> > > >
> > > > Thanks for pointing that out!  I'll note it in the page.
> > > >
> > > > >          Note that the core limit size macro may be a different value than what
> > > > >          is returned by getrlimit(RLIMIT_CORE,...).  This is due to the fact
> > > > >          that the core_pattern specified executible will be run as the same uid
> > > > >          as the crashing process, and to facilitate reception of the entire
> > > > >          core, the kernel will temporarily set RLIMIT_CORE to unlimited while
> > > > >          the dump is in progress.
> > > >
> > > > Actually, I can't seem to get an example of this behavior.  In my
> > > > experiments, %c always seems to give the "right" info (i.e., I don't
> > > > ever see %c showing 2^32 (unlimited) when I set a soft limit).  Can
> > > > you show a specific case where it doesn't give the "right" value?
> > > >
> > > Oops, you're right.  I had initially implemented my core pattern updates this
> > > way, but in the end wound up just ignoring the limit in do_coredump, rather than
> > > re-writing it.  Thanks for that.  You can scratch this.
> > >
> > >
> > > > >          Note also %u and %g may be different values
> > > > >          than getuid/getgid in the event that the core_pattern executable is
> > > > >          set[u|g]id root
> > > >
> > > > I'm slightly confused by that last point.  According to my
> > > > experiments, the core_pattern executable is always run as user and
> > > > group root, so making it set[ug]id root would seem to be a no-op.
> > > > (But anyway, %u and %g do give the "right" values -- the UID and GID
> > > > of the dumping process.)
> > > >
> > > Hmm, are you sure, I was under the impression that we fork the usermodehelper in
> > > do_coredump as a parent of current, which has the dumping processes uid/gid.  I
> > > do see that in do_coredump we call get_dumpable(mm) and if it returns with the
> > > appropriate value we switch current->fsuid to 0.  I wonder if thats what you're
> > > seeing?
> >
> > Have a look at the following.  It demonstrates what I'm seeing (that
> > the coredump program is run as root/root).
> >
> > ===
> > $ cat core_pattern_test.c
> > /* core_pattern_test.c */
> >
> > #define _GNU_SOURCE
> > #include <sys/stat.h>
> > #include <fcntl.h>
> > #include <stdio.h>
> > #include <stdlib.h>
> > #include <unistd.h>
> >
> > #define BUF_SIZE 1024
> >
> > int
> > main(int argc, char *argv[])
> > {
> >     int fd, tot, j;
> >     ssize_t nread;
> >     char buf[BUF_SIZE];
> >     FILE *fp;
> >
> >     fd = open(argv[1], O_CREAT | O_WRONLY | O_TRUNC, 0666);
> >     if (fd == -1)
> >         exit(EXIT_FAILURE);
> >
> >     fp = fdopen(fd, "a");
> >
> >     fprintf(fp, "PID=%ld\n", (long) getpid());
> >     fprintf(fp, "cwd=%s\n", get_current_dir_name());
> >     fprintf(fp, "UID=%ld; EUID=%ld\n", (long) getuid(), (long) geteuid());
> >     fprintf(fp, "GID=%ld; EGID=%ld\n", (long) getgid(), (long) getegid());
> >
> >     fprintf(fp, "argc=%d\n", argc);
> >     for (j = 0; j < argc; j++)
> >         fprintf(fp, "argc[%d]=<%s>\n", j, argv[j]);
> >
> >     /* Count bytes in standard input */
> >
> >     tot = 0;
> >     while ((nread = read(STDIN_FILENO, buf, BUF_SIZE)) > 0)
> >         tot += nread;
> >     fprintf(fp, "Total bytes in core dump: %d\n", tot);
> >
> >     exit(EXIT_SUCCESS);
> > }
> > $ cc core_pattern_test.c
> > $ sudo sh -c 'echo "|$PWD/core_pattern_test $PWD/c p_%p u_%u g_%g t_%t
> > c_%c" > /proc/sys/kernel/core_pattern'
> > root's password:
> > $ id
> > uid=1000(mtk) gid=100(users) groups=16(dialout),33(video),100(users)
> > $ sleep 100
> > [type ^\]
> > Quit (core dumped)
> > $ cat c
> > PID=6743
> > cwd=/
> > UID=0; EUID=0
> > GID=0; EGID=0
> > argc=7
> > argc[0]=</home/mtk/man-pages/man5/core_pattern_test>
> > argc[1]=</home/mtk/man-pages/man5/c>
> > argc[2]=<p_6742>
> > argc[3]=<u_1000>
> > argc[4]=<g_100>
> > argc[5]=<t_1209146940>
> > argc[6]=<c_4294967295>
> > Total bytes in core dump: 282624
> > $
> > ===
> >
> > Your thoughts?
> >
> > Cheers,
> >
> > Michael
>
> I certainly don't doubt your results.  The uid/gid options IIRC were there when
> I made my updates and so I left them alone, wokring under the asumption that
> thats what they were there for.  Clearly you have evidence to the contrary here.
> It seems like we should run the core collector as the uid of the dumping
> process, simply because it doesn't need to be run as root (and one could force
> the running of the commadn as root using the suid bit on the executable file).
>
> I say documented like your experiment shows it to work, and I'll try to find
> some time to investigate where we switch uids and why.

There was another aspect of this behavior that bugs me.  The current
working directory of the dumping program is /.  This means it's
impossible for the dumping program to produce an output file in the
current working directory of the program that is crashing (which of
course is where a "core" file is normally written).  Is that really
intended behavior?
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists