Re: RFA: Actual support for tracing forks on GNU/Linux

[Daniel Jacobowitz <drow@mvista.com>]

On Wed, Jul 09, 2003 at 05:57:13PM -0400, Daniel Jacobowitz wrote:
> On Sat, Jun 28, 2003 at 12:34:44PM -0400, Daniel Jacobowitz wrote:
> > On Wed, Jun 18, 2003 at 07:29:42PM -0400, Daniel Jacobowitz wrote:
> > > This patch enables "catch fork" and "set follow-fork-mode child" for fork(). 
> > > 
> > > Other things from my working directory that doesn't include:
> > >  - vfork support - I'll do that separately once this in
> > >  - exec support - the code is still a mess
> > >  - exit event support - i.e. stopping the process just before it exits,
> > >    to take a look around - this confuses GDB very badly
> > >  - gdbserver support for any of the above
> > >  - enabling the testsuite checks for this - even worse of a mess
> > > I hope to get at least native vfork support into GDB 6.0.
> > > 
> > > This patch works by adding a hook every time we attach to an LWP or fork a
> > > new LWP, to set tracing flags on it.  We enable event reporting for fork()
> > > [which requires ~ 2.5.34 kernel; I heard that one of RH's backport kernels
> > > included this, but I don't know if it still does.]  Then, when we get a
> > > fork, we end up attached to both parent and child.  We remove breakpoints in
> > > whichever one we don't care about, and then we detach it.
> > > 
> > > This both lets us choose which one to trace, and also fixes the problem
> > > where breakpoints would be left in the inferior after it forked, causing the
> > > child to die with SIGTRAP.
> > > 
> > > I had to override kill_inferior, for an issue discovered in testing: when
> > > we're stopped with both processes attached, we have to make sure to kill
> > > them both.  We have to deal with this because the user could "catch fork",
> > > and when they see the fork decide which one to debug.
> > > 
> > > I think that's everything.  I'd like at least a nod from the threading
> > > maintainers, since I had to hook into lin-lwp.c.  OK?
> 
> Ping?  If neither of our thread maintainers feels comfortable
> commenting on this patch I'm willing to approve it myself, but I'd
> prefer a comment from the maintainers of lin-lwp.

Ping.

> > > 2003-06-18  Daniel Jacobowitz  <drow@mvista.com>
> > > 
> > > 	* config/i386/nm-linux.h (LINUX_CHILD_POST_STARTUP_INFERIOR): Define.
> > > 	* config/nm-linux.h (linux_enable_event_reporting)
> > > 	(linux_handle_extended_wait, linux_child_post_startup_inferior): New
> > > 	prototypes.
> > > 	(CHILD_POST_STARTUP_INFERIOR, CHILD_POST_ATTACH, CHILD_FOLLOW_FORK)
> > > 	(KILL_INFERIOR): Define.
> > > 	* i386-linux-nat.c (child_post_startup_inferior): New function.
> > > 	* i386-nat.c (child_post_startup_inferior): Wrap in #ifdef.
> > > 	* infptrace.c (kill_inferior): Wrap in #ifdef.
> > > 	* lin-lwp.c (lin_lwp_attach_lwp): Call child_post_attach after
> > > 	attaching to each LWP.
> > > 	(child_wait, lin_lwp_wait): Call linux_handle_extended_wait.
> > > 	(init_lin_lwp_ops): Fill in some more operations.
> > > 	* linux-nat.c (linux_enable_event_reporting): New function.
> > > 	(child_post_attach, linux_child_post_startup_inferior)
> > > 	(child_post_startup_inferior, child_follow_fork)
> > > 	(linux_handle_extended_wait, kill_inferior): New functions.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer