Re: [COUNTERPATCH] mm: avoid overflowing preempt_count() in mmu_take_all_locks()

[Avi Kivity <avi@redhat.com>]

On 04/01/2010 06:54 PM, Andrea Arcangeli wrote:
>
>>> The only single reason I'd go for mutexes would be to accommodate
>>> XPMEM requirements once and for all, no other reason.
>>>
>>>        
>> There is also a minor benefit for kvm.  Reduced latency over large mmu
>> operations; code simplification (we now have some
>> copy_from_user_inatomic() that could be simplified).
>>      
> Where exactly is KVM taking these locks?

Not these locks, but if we go all the way and make mmu notifiers 
sleepable, we can convert mmu_lock to a mutex.

> KVM should only call into
> GUP, and GUP itself won't iterate over rmaps either. GUP just walks
> the host pagetables and trigger page faults if the pages aren't
> mapped.

We'll probably deadlock then, gup -> change_pte notifier -> mmu_lock.  
But we can probably work around it.

> I don't see how you're going to remove
> copy_from_user_inatomic() given we don't have vmas and other metadata
> to take those locks. Maybe we can stop calling GUP but even if we take
> the anon_vma mutex/semaphore I think it won't still prevent munmap to
> drop the anon pages from under us (even if it'd stop the VM to unmap
> them through rmap). To freeze the mapping one would need to take
> mmap_sem in write mode in addition to the anon_vma mutex/sem which is
> unlikely a win compared to just gup+copy_from_user_inatomic. So I
> don't see immediate benefits for KVM but maybe I'm missing something
> of course!
>    

I meant replace c_f_u_inatomic() by c_f_u() (that's why the benefit is 
minor - we only simplify the failure path).  Sorry for being unclear.


-- 
error compiling committee.c: too many arguments to function