Re: [PATH -mm -v2] Fix a race condtion of oops_in_progress

[Chris Snook <csnook@redhat.com>]

Huang Ying wrote:
> On Mon, 2008-11-10 at 15:35 +0800, KOSAKI Motohiro wrote:
>>>>> As far as I know, barriers don't cause changes to be visible on other
>>>>> CPUs faster too. It just guarantees corresponding operations after will
>>>>> not get executed until that before have finished. And, I don't think we
>>>>> need make changes to be visible on other CPUs faster.
>>>> You're correct that barrier() has no impact on other CPUs.  wmb() and rmb() do. 
>>>>   If we don't need to make changes visible any faster, what's the point in using 
>>>> atomic_set()?  It's not any less racy.  atomic_inc() and atomic_dec() would be 
>>>> less racy, but you're not using those.
>>> In default bust_spinlocks() implementation in lib/bust_spinlocks.c,
>>> atomic_inc() and atomic_dec_and_test() is used. Which is used by x86
>>> too. In some other architecture, atomic_set() is used to replace
>>> "oops_in_progress = <xxx>". So this patch fixes architectures which use
>>> default bust_spinlocks(), other architectures can be fixed by
>>> corresponding architecture developers.
>> I think Chris is right.
>> So, I reccomend to read Documentation/memory-barriers.txt
>>
>> Almost architecture gurantee atomic_inc cause barrier implicitly.
>> but not _all_ architecture.
> 
> Yes. atomic_inc() doesn't imply barrier on all architecture. But we
> should not add barriers before all atomic_inc(), just ones needed. Can
> you figure out which ones in the patch should has barrier added?

You need barriers *after* writes, and *before* reads.  Adding barriers to the 
oops path should be extremely cheap for performance, unless oopsing is a common 
occurrence, in which case we have bigger problems.

-- Chris