Re: Xen PVM: Strange lockups when running PostgreSQL load

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 17/10/12 15:51, Jan Beulich wrote:
>>>> On 17.10.12 at 15:28, Andrew Cooper <andrew.cooper3@citrix.com> wrote:
>> In our case, certain processes were locking up, and it turned out that
>> the kernel was issuing SCHOP_poll hypercalls (same the stack trace on
>> your launchpad ticket) on its own spinlock IPI event channel
>> (understandable, as its a spinlock), but with the event channel masked,
>> so it would never wake up from the poll.
> Probably some misunderstanding: The event channel used
> here will always be masked, and whether an event channel is
> masked doesn't matter for the purposes of polling it.
>
> Jan
>

Sorry - I was not very clear.

In our case, we saw every single VCPU stuck in a SCHEDOP_poll on its own
spinlock event channel, which renderes the guest completely deadlocked.

My understanding was that while some VCPUs were blocked in the
hypervisor, the VCPU with the lock should raise an event when the lock
was released, waking the other VCPUs so one could grab the lock.  This
implication of this being that there is a bug in the spinlock code.

-- 
Andrew Cooper - Dom0 Kernel Engineer, Citrix XenServer
T: +44 (0)1223 225 900, http://www.citrix.com