Re: kernel panic in fib_rules_lookup [2.6.27.7 vendor-patched]

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mardi 26 octobre 2010 à 09:53 -0400, Joe Buehler a écrit :
> Eric Dumazet wrote:
> 
> > Well, it would be surprising this being wrong and crash only once in a
> > while in fib_rules_lookup
> 
> > Did you tried my last patch ?
> 
> There was a patch to the kernel by David Daney back in January to
> improve performance of Octeon memory barriers.  The patch changes the
> generic MIPS barrier code to introduce optimizations for Octeon.  The
> LINUX version I am using is from the Octeon SDK and appears to have an
> early version of this patch.  It's broken however -- the Jan patch has
> proper SYNCW instructions in smp_wmb while the SDK version does not.
> 
> I have made your changes but will also fold in this change, then start
> some load testing.
> 
> The real-time scheduler is broken in the LINUX I am using -- I get
> kernel crashes -- and I would be most happy if the SYNCW fix fixed that
> also.
> 
>

Just to make sure, are you using a single syncw, or a double one ?

/*
 * We actually use two syncw instructions in a row when we need a write
 * memory barrier. This is because the CN3XXX series of Octeons have
 * errata Core-401. This can cause a single syncw to not enforce
 * ordering under very rare conditions. Even if it is rare, better safe
 * than sorry.
 */
#define CVMX_SYNCW asm volatile ("syncw\n\tsyncw" : : : "memory")