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

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

Le samedi 23 octobre 2010 à 11:40 -0400, Joe Buehler a écrit :
> Eric Dumazet wrote:
> >
> > David Daney added a nudge_writes(), actually doing a "syncw"
> > instruction, and this seems to be the smp_wmb() this platform should be
> > using in the first place, not a pure compiler barrier (barrier()) 
> >
> > So Joe, you might want to change the smp_wmb() call in
> > rcu_assign_pointer() by the nudge_writes() call, and see what happens...
> >
> >
> >   
> 
> I think Daney is Cavium's Octeon LINUX guru from the posts I've seen so
> he would definitely know the platform.  I'm not sure I quite understand
> what you are saying but it sounds as though you are saying that smp_wb
> is not doing a syncw and that sounds *totally* broken -- snycw is what
> the low-level Cavium SDK uses for memory barriers all over the place.
> 

Yes, I am saying exactly this : smp_wmb() is a barrier() only, at least
on the disassembly you provided to me. It might be fine (it is the same
on x86 for example)

fib_rules.old.s

.L234:
	.loc 1 338 0
	beq	$9,$0,.L235   if (last) {
.LBB911:					// last = prev->next
.LBB912:
	.loc 12 45 0
	ld	$2,0($9)   //next = prev->next
.LBB913:
.LBB914:
	.loc 12 22 0
	sd	$9,8($18)   // part of list_add_rcu  new->prev = prev;
	.loc 12 21 0
	sd	$2,0($18)   // new->next = next;
	.loc 12 23 0
<<!>>
	sd	$18,0($9)  //rcu_assign_pointer(prev->next, new); 
	.loc 12 24 0
	sd	$18,8($2)  // next->prev = new;
.L236:
.LBE914:
.LBE913:


No syncw here at least.