Softing firing on wrong cpu

[Martin Husemann <martin%duskware.de@localhost>]

Hi folks,

I see a (more or less reproducable) problem that looks like a softint
is scheduled on cpu0 but fires on cpu1. It then hits a KASSERT since on
that cpu the sofint is not marked si_active.

I added some instrumentation and couldn't explain things unless in 
softint_overlay the l_cpu pointer of the passed lwp is not curcpu(), so
I added a KASSERT to verify.

Unfortunately with that change I have never been able to reproduce the problem
again.

Index: kern_softint.c
===================================================================
RCS file: /cvsroot/src/sys/kern/kern_softint.c,v
retrieving revision 1.29
diff -u -p -r1.29 kern_softint.c
--- kern_softint.c      19 Jul 2009 10:11:55 -0000      1.29
+++ kern_softint.c      29 Oct 2009 13:41:28 -0000
@@ -720,6 +730,7 @@ softint_overlay(void)
        si = ((softcpu_t *)ci->ci_data.cpu_softcpu)->sc_int;
 
        KASSERT((l->l_pflag & LP_INTR) == 0);
+       KASSERT(ci == curcpu());
 
        /* Arrange to elevate priority if the LWP blocks. */
        s = splhigh();


Now if I can trust my eyes I did write ==, not =, so the problem going
away suprises me a bit.

Anybody got an idea that does not boild down to "cache coherency problem
at CPUINFO_VA"?

Martin