This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
Re: RFA: close-on-exec internal file descriptors
- From: Mark Kettenis <mark dot kettenis at xs4all dot nl>
- To: tromey at redhat dot com
- Cc: gdb-patches at sourceware dot org
- Date: Sat, 6 Dec 2008 16:40:48 +0100 (CET)
- Subject: Re: RFA: close-on-exec internal file descriptors
- References: <m3skp2w32i.fsf@fleche.redhat.com>
> From: Tom Tromey <tromey@redhat.com>
> Date: Fri, 05 Dec 2008 17:38:13 -0700
>
> A few weeks ago, someone on irc pointed out that gdb leaks file
> descriptors to its child processes. One simple way to see this is to
> start gdb with an exec and then shell out:
>
> opsy. gdb /tmp/r
> [...]
> (gdb) shell ls -l /proc/$$/fd
> total 0
> lrwx------ 1 tromey tromey 64 2008-12-05 17:10 0 -> /dev/pts/1
> lrwx------ 1 tromey tromey 64 2008-12-05 17:10 1 -> /dev/pts/1
> lrwx------ 1 tromey tromey 64 2008-12-05 17:10 2 -> /dev/pts/1
> lr-x------ 1 tromey tromey 64 2008-12-05 17:10 3 -> pipe:[1100229]
> l-wx------ 1 tromey tromey 64 2008-12-05 17:10 4 -> pipe:[1100229]
> lr-x------ 1 tromey tromey 64 2008-12-05 17:10 5 -> /proc/8096/fd
>
> I believe those 'pipe' entries are from the call to pipe in
> linux-nat.c:linux_nat_set_async.
>
> This patch fixes this problem by introducing new wrapper functions
> which create close-on-exec file descriptors. Then, these functions
> are used everywhere in gdb. After this patch, these wrapper functions
> should be used in all new code as well.
>
> I chose to take advantage of the new glibc flags like O_CLOEXEC when
> they are available. This is friendlier in the Python case -- here,
> gdb might have multiple threads, and the glibc flags enable us to
> avoid a window where a file descriptor is not marked close-on-exec.
> If the new flags or functions are not available, we fall back to
> fcntl.
Sorry, but I don't see the point in having #ifdef O_CLOEXEC code when
there is a perfectly portable way to do this using fcntl. It leads to
more bits of code that can possibly go untested.
I also think it would actually be better to explicitly close file
descriptors before doing an exec instead of relying on people to use
the proper _cloexec call throughout gdb.