[ofa-general] Re: [PATCH 01 of 12] Core of mmu notifiers
clameter at sgi.com
Mon Apr 28 18:28:06 PDT 2008
On Tue, 29 Apr 2008, Andrea Arcangeli wrote:
> Frankly I've absolutely no idea why rcu is needed in all rmap code
> when walking the page->mapping. Definitely the PG_locked is taken so
> there's no way page->mapping could possibly go away under the rmap
> code, hence the anon_vma can't go away as it's queued in the vma, and
> the vma has to go away before the page is zapped out of the pte.
zap_pte_range can race with the rmap code and it does not take the page
lock. The page may not go away since a refcount was taken but the mapping
can go away. Without RCU you have no guarantee that the anon_vma is
existing when you take the lock.
How long were you away from VM development?
> Now the double atomic op may not be horrible when not contented, as it
> works on the same cacheline but with cacheline bouncing with
> contention it sounds doubly horrible than a single cacheline bounce
> and I don't see the point of it as you can't use rcu anyways, so you
> can't possibly take advantage of whatever microoptimization done over
> the original locking.
Cachelines are acquired for exclusive use for a mininum duration.
Multiple atomic operations can be performed after a cacheline becomes
exclusive without danger of bouncing.
More information about the general