lkml.org 
[lkml]   [2012]   [Feb]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] x86, mce: Fix rcu splat in drain_mce_log_buffer()
On Sun, Feb 12, 2012 at 09:41:33AM -0800, Paul E. McKenney wrote:
> On Mon, Feb 06, 2012 at 05:18:02PM +0100, Borislav Petkov wrote:
> > On Fri, Jan 13, 2012 at 01:44:35PM -0800, Paul E. McKenney wrote:
> > > Looks good to me, but I do need to defer to people who know this code
> > > better than do I. The key thing that (from what I can see) makes
> > > rcu_dereference() unnecessary is that the smp_rmb() used in conjunction
> > > with polling the .finished field takes care of ordering.
> >
> > Right, this was me trying hard not to screw up touching mcelog.next,
> > thus trying to use the rcu_dereference_index_check() primitive without
> > thinking it through too much. But you're right, I'm polling the
> > ->finished field 4 times (totally arbitrary, btw) which should suffice
> > while the mce_log() routine above writes those entries.
> >
> > Although, the question still remains, since mce_log() accesses
> > mcelog.next through the rcu_dereference_index_check() primitive,
> > shouldn't I do it the same way?
>
> I don't claim to be an mce_log() expert, but when I looked it over,
> I didn't see a need for rcu_dereference_index_check(). Unless I am
> confused (quite possible), the memory barriers are sufficient.
>
> The rcu_read_lock() and rcu_read_unlock() seem to be needed to avoid
> premature freeing, though.

Looka ere:

commit f56e8a0765cc4374e02f4e3a79e2427b5096b075
Author: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Date: Fri Mar 5 15:03:27 2010 -0800

x86/mce: Fix RCU lockdep splats

Create an rcu_dereference_check_mce() that checks for RCU-sched
read side and mce_read_mutex being held on update side. Replace
uses of rcu_dereference() in arch/x86/kernel/cpu/mcheck/mce.c
with this new macro.

...

+#define rcu_dereference_check_mce(p) \
+ rcu_dereference_check((p), \
+ rcu_read_lock_sched_held() || \
+ lockdep_is_held(&mce_read_mutex))
+
So I guess the check is to see that we're rather holding the
mce_chrdev_read_mutex when accessing this from userspace (the mcelog
userspace thing).

I can't comment on the RCU-sched thing because I don't know it that
well. It has to do with extended CPU idle times but why do we require to
be in RCU read-side critical section when accessing the index? Hmmm, I
think a guy called Paul should know :-) ...

Thanks.

--
Regards/Gruss,
Boris.

Advanced Micro Devices GmbH
Einsteinring 24, 85609 Dornach
GM: Alberto Bozzo
Reg: Dornach, Landkreis Muenchen
HRB Nr. 43632 WEEE Registernr: 129 19551


\
 
 \ /
  Last update: 2012-02-16 11:19    [from the cache]
©2003-2011 Jasper Spaans