Re: 2.6.9-rc3-mm2

[Rusty Russell <rusty@rustcorp.com.au>]

On Fri, 2004-10-15 at 12:09, Andrew Morton wrote:
> Rusty Russell <rusty@rustcorp.com.au> wrote:
> >
> > > We were getting warnings from somewhere or other due to smp_processor_id()
> >  > within preemptible code - I don't recall the callsite.
> > 
> >  That's weird, but implies bogosity in the caller.  Covering it up like
> >  this is not necessarily a win.
> 
> umm.
> 
> 	ip_conntrack_cleanup
> 	->ip_ct_selective_cleanup
> 	  ->death_by_timeout
> 	    ->CONNTRACK_STAT_INC
> 
> is one route.

Yep, called on module removal.

Cheers,
Rusty.

Name: Avoid warning on CONNTRACK_STAT_INC in death_by_timeout()
Status: Trivial
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>

Module removal can call death_by_timeout() manually, which isn't in
softirq context, so the CONNTRACK_STAT_INC() call there (which assumes
preempt disabled) can give a warning.  Of course, the warning here is
spurious, but the simplest workaround is to call CONNTRACK_STAT_INC()
inside the lock.

diff -urpN --exclude TAGS -X /home/rusty/devel/kernel/kernel-patches/current-dontdiff --minimal .29431-linux-2.6.9-rc4-bk2/net/ipv4/netfilter/ip_conntrack_core.c .29431-linux-2.6.9-rc4-bk2.updated/net/ipv4/netfilter/ip_conntrack_core.c
--- .29431-linux-2.6.9-rc4-bk2/net/ipv4/netfilter/ip_conntrack_core.c	2004-10-11 15:17:21.000000000 +1000
+++ .29431-linux-2.6.9-rc4-bk2.updated/net/ipv4/netfilter/ip_conntrack_core.c	2004-10-15 13:05:52.000000000 +1000
@@ -328,9 +328,10 @@ static void death_by_timeout(unsigned lo
 {
 	struct ip_conntrack *ct = (void *)ul_conntrack;
 
-	CONNTRACK_STAT_INC(delete_list);
-
 	WRITE_LOCK(&ip_conntrack_lock);
+	/* Inside lock so preempt is disabled on module removal path.
+	 * Otherwise we can get spurious warnings. */
+	CONNTRACK_STAT_INC(delete_list);
 	clean_from_lists(ct);
 	WRITE_UNLOCK(&ip_conntrack_lock);
 	ip_conntrack_put(ct);

-- 
Anyone who quotes me in their signature is an idiot -- Rusty Russell