Re: compiler bug gcc4.6/4.7 with ACCESS_ONCE and workarounds

From: Martin Schwidefsky
Date: Wed Nov 12 2014 - 04:30:00 EST


On Tue, 11 Nov 2014 16:36:06 -0800
Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:

> On Tue, Nov 11, 2014 at 4:33 PM, Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > I guess as a workaround it is fine, as long as we don't lose sight of
> > trying to eventually do a better job.
>
> Oh, and when it comes to the actual gcc bug - do you have any reason
> to believe that it's somehow triggered more easily by something
> particular in the arch/s390/kvm/gaccess.c code?
>
> IOW, why does this problem not hit the x86 spinlocks that also use
> volatile pointers to aggregate types? Or does it?

This looks similiar to what we had on s390:

old.tickets = ACCESS_ONCE(lock->tickets)

In theory x86 should be affected as well. On s390 we have lots of
instruction that operate on memory and the cost model of gcc makes
the compiler more inclined to access memory multiple times. My
guess would be that once the value is cached in a register the
cost model for x86 will usually make sure that the value is not
read a second time. But this is no guarantee.

--
blue skies,
Martin.

"Reality continues to ruin my life." - Calvin.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/