Re: xenmp testing

["Cherry G. Mathew" <cherry.g.mathew%gmail.com@localhost>]

On 19 January 2012 16:46, Manuel Bouyer <bouyer%antioche.eu.org@localhost> 
wrote:
> On Thu, Jan 19, 2012 at 10:39:59AM +0100, Manuel Bouyer wrote:
>> > There's a race where a pmap pm_pdir[] can have a lingering entry in
>> > per-cpu L4 pdirs if they have alternate mappings to that pmap. I've
>> > tried to solve it with this:
>> >
>> > http://85.88.27.205/tmp/apte.diff
>> >
>> > But yet to hear if that fixes the problem.
>> >
>> > I'm not sure if these are related. Can it be tested ?
>>
>> I tested, it doesn't help with my problem.


You might have noticed that I knocked off apte use in xen pmap.

>
> FWIW, I couldn't reproduce the problem on a UP domU.
> It looks like a race somewhere where a PTP or PTE page is reused as data page
> before being unregisterd from hypervisor.
>
> This doesn't mean the race doesn't exist with a UP domU though ...
>

I fixed a race between xen_kpm_sync() and stale TLB entries on other
cpus that have L3 entries pointing back at the original page frame by
enforcing immediate shootdown.

The domUs look more stable to me now.

Do have a look when you find time.
-- 
~Cherry