On Wed, May 05, 2010 at 07:34:37AM -0700, Linus Torvalds wrote:
>
>
> On Wed, 5 May 2010, Mel Gorman wrote:
> >
> > With the recent anon_vma changes, there can be more than one anon_vma->lock
> > to take in a anon_vma_chain but a second lock cannot be spinned upon in case
> > of deadlock. The rmap walker tries to take locks of different anon_vma's
> > but if the attempt fails, locks are released and the operation is restarted.
>
> Btw, is this really needed?
>
I could not convince myself that it wasn't. lockdep throws a fit if you try
but it can be taught about the situation if necessary.
> Nobody else takes two anon_vma locks at the same time, so in order to
> avoid ABBA deadlocks all we need to guarantee is that rmap_walk_ksm() and
> rmap_walk_anon() always lock the anon_vma's in the same order.
>
rmap_walk() appears to be the only one that takes multiple locks but it itself
is not serialised. If there are more than one process calling rmap_walk()
on different processes sharing the same VMAs, is there a guarantee they walk
it in the same order? I didn't think so at the time the patch because the
anon_vma the walk starts from is based on the page being migrated rather
than any idea of starting from a parent or primary anon_vma.
> And they do, as far as I can tell. How could we ever get a deadlock when
> we have both cases doing the locking by walking the same_anon_vma list?
>
If we always started the list walk in the same place then it'd be fine but
if they start in different places, it could deadlock.
> list_for_each_entry(avc, &anon_vma->head, same_anon_vma) {
>
> So I think the "retry" logic looks unnecessary, and actually opens us up
> to a possible livelock bug (imagine a long chain, and heavy page fault
> activity elsewhere that ends up locking some anon_vma in the chain, and
> just the right behavior that gets us into a lockstep situation),
I imagined it and I'm not super-happy about it. It's one of the reasons Rik
called it "fragile".
> rather than fixing an ABBA deadlock.
>
> Now, if it's true that somebody else _does_ do nested anon_vma locking,
> I'm obviously wrong. But I don't see such usage.
>
> Comments?
>
Just what I have above. I couldn't convince myself that two callers to
rmap_walk from pages based on different VMAs on the same_anon_vma list would
always started in the same place.
--
Mel Gorman
Part-time Phd Student Linux Technology Center
University of Limerick IBM Dublin Software Lab
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>