Re: [PATCH] kvm memslot read-locking with mmu_lock

[Andrea Arcangeli <andrea-atKUWr5tajBWk0Htik3J/w@public.gmane.org>]

On Tue, Jan 22, 2008 at 04:38:49PM +0200, Avi Kivity wrote:
> Andrea Arcangeli wrote:
>>
>>> This is arch independent code, I'm surprised mmu_lock is visible here?
>>>     
>>
>> The mmu_lock is arch independent as far as I can tell. Pretty much
>> like the mm->page_table_lock is also independent. All archs will have
>> some form of shadow pagetables in software or hardware, and mmu_lock
>> is the lock to take to serialize the pagetable updates and it also
>> allows to walk the memslots in readonly mode.
>>
>>   
>
> Well, s390 has everything in hardware, but I suppose they can just ignore 
> the lock.

If they don't take it in their lowlevel it's enough I think. It'll
still be useful to lookup memslots like in every other arch.

> But we lookup memslots in parallel in the guest walker and similar places, 
> relying on mmap_sem being taken for read.

That's ok. The vmas could also be looked up either with mmap_sam in
read mode or with only the page_table_lock taken (these days it
doesn't work anymore with ptlocks). The mmu_lock is only taken in the
modifications of the memslots, the lookups don't require it. The
modifications have to take both mmap_sem in write mode and the
mmu_lock. Lookups requires either mmap_sem in read mode or the
mmu_lock. The lookups with mmap_sem in read mode won't ever see a
userspace_addr =0, the mmu_lock have to skip over the userspace_addr=0 instead.

> Maybe we need a rwlock instead, and drop this overloaded usage of mmap_sem.

Adding another lock just for the memslots should be feasible. However
it can't be a semaphore or it won't work for the
mmu-notifiers. A rwlock would be ok, but then all memslots lookups
will have to be updated to check for userspace_addr != 0 like the
kvm_hva_to_rmapp does.

> Nothing guarantees a 1:1 mapping between memslots and vma.  You can have a 
> vma backing multiple memslots, or a memslot spanning multiple vmas.

Yes, that's why it would add complications to have it in the vma, one
would need to add a vma split to guarantee the 1:1 mapping, and then
it'd be a bit faster too (though the memslots can easily be rbtree-indexed).

-------------------------------------------------------------------------
This SF.net email is sponsored by: Microsoft
Defy all challenges. Microsoft(R) Visual Studio 2008.
http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/