Re: [PATCH] sched: fix race between sched_move_domain() and vcpu_wake()

[George Dunlap <george.dunlap@eu.citrix.com>]

On 11/10/13 13:20, Jan Beulich wrote:
>>>> On 11.10.13 at 11:32, "Jan Beulich" <JBeulich@suse.com> wrote:
>> I suppose you scanned the code for other cases like this, and
>> there are none?
> Actually I did just now, and I think there's a similar issue in
> credit2's init_pcpu(): After taking pcpu_schedule_lock(cpu) it
> alters schedule_lock and hence effectively drops the locking,
> yet continues to do other stuff before in fact releasing it.
>
> What is being done prior to unlocking, however, looks to be
> unrelated to the lock being held, and rather independently
> (of the effective releasing) wanting &rqd->lock held.

I can't quite make out what you mean in the last sentence; but setting 
the cpu in rqd->idle and rqd->active should certainly be protected by 
rqd->lock, and it certainly looks like it's not being grabbed at the moment.

Hmm -- I think we may need to do some kind of fancy looping thing like 
we do in vcpu_migrate, to lock both the current schedule lock and 
rqd->lock; with the difference, I suppose, that rqd lock won't change 
(since the assignment of cpu->runqueue at the moment is static).

Let me put this on my list of things to do before the release.

  -George