|
|
The following reply was made to PR kern/38637; it has been noted by GNATS.
From: Andrew Doran <ad@xxxxxxxxxx>
To: Manuel Bouyer <bouyer@xxxxxxxxxxxxxxx>
Cc: gnats-bugs@xxxxxxxxxx, kern-bug-people@xxxxxxxxxx,
gnats-admin@xxxxxxxxxx, netbsd-bugs@xxxxxxxxxx
Subject: Re: kern/38637: pppoe fails to reconnect sometimes
Date: Tue, 5 May 2009 20:52:45 +0000
On Tue, May 05, 2009 at 10:20:21PM +0200, Manuel Bouyer wrote:
> On Fri, Apr 03, 2009 at 08:06:00PM +0200, Manuel Bouyer wrote:
> > On Thu, Oct 02, 2008 at 10:50:05AM +0000, Martin Husemann wrote:
> > > Sorry, I don't see how this could happen besides the softint handling
> > > breaking completely (something corrupting the call wheel?):
> > > [...]
> >
> > Update on this: this time (with a 5.0_RC3 kernel), it's not the ppoeinq
> > which stopped being processed and did overflow, but ipintrq !
> > So it's really related to soft interrupt not being called anymore ...
>
> I think I found the reason for this (or at last a possible reason).
> In sys/kern/kern_softint.c, sh_flags should be declared volatile.
> Without it, on ports where splhigh() is inline, the compiler will optimise
> the second SOFTINT_PENDING test in softint_schedule(). A dissasembly
> of softint_schedule() with and without the volatile sh_flags confirm this
> on sparc.
> Because of this there is a race that could lead to the softhand_t
> being enqueued twice on si_q, leading to a corrupted queue and
> some handler being SOFTINT_PENDING but never called.
Nice diagnosis! However the softint code is correct.
splhigh/splx on sparc64 should be real functions, or if one believes there
is a performance advantage, should be inlines with __insn_barrier in the
correct spots. Consider the potential effects elsewhere. I pointed this
issue out to Martin or Matthew a couple of years ago and then promptly
forgot about it.
|
|