Re: [BProc] Re: deadlock with 2.4.21/3.2.6

[<er...@he...>]

On Tue, Oct 14, 2003 at 12:30:41PM -0400, Nicholas Henke wrote:
> On Tue, 2003-10-14 at 11:56, er...@he... wrote:
> > 
> > Hrm.  I'm glad you brought this up.  I've recently seen a similar
> > problem.  I thought it had something to do with the recent network
> > upgrade we did.  Sounds like probably not.
> 
> Fun ;/ Thanks for taking a look at it.
> 
> > 
> > It's pretty mysterious to me.  BProc really doesn't do much with
> > interrupts turned off.  I've been working on reproducing it more
> > reliably here.
> > 
> > I don't know a good way to shake the kernel loose in that case.  One
> > thing we wwere going to try was to instrument it a bit with POST codes
> > or try and poke around in memory a bit with a bus analyzer.
> 
> Ok -- way over my head there, I wonder if a hardware watchdog card would
> help - or if that would give the same results as the nmi_watchdog..aka
> nothing.

I got an oops out of the NMI watchdog which was enlightening (or at
least indicated which code was at fault).  The following patch may
have fixed it for me.  I say "may have" since I've had some trouble
reproducing the problem reliably.

This patch turns off "sigbypass" which is a little optimization where
a process sending a signal to a ghost doesn't bother the ghost.
Instead it just throws a signal forwarding message right on the
message queue.  I'm not sure how the code is broken.  I haven't had
time to look into it yet.

Please give it a try and let me know if you still see the deadlock.


--- hooks.c	29 Aug 2003 21:46:57 -0000	1.53
+++ hooks.c	21 Oct 2003 15:41:44 -0000
@@ -314,6 +314,11 @@
      * t->sigmask */
     struct bproc_krequest_t *req;
     struct siginfo           tmpinfo;
+
+    return 0;			/* XXX disable sigbypass for now.
+				 * There seems to be something busted or
+				 * unsafe about this code... */
+
     if (!BPROC_ISGHOST(t) || !t->bproc.ghost->sigbypass)
 	return 0;