Re: REGRESSION: Performance regressions from switching anon_vma->lockto mutex

[Linus Torvalds]

On Tue, Jun 14, 2011 at 5:29 PM, Tim Chen <tim.c.chen@xxxxxxxxxxxxxxx> wrote:
>
> On 2.6.39, the contention of anon_vma->lock occupies 3.25% of cpu.
> However, after the switch of the lock to mutex on 3.0-rc2, the mutex
> acquisition jumps to 18.6% of cpu.  This seems to be the main cause of
> the 52% throughput regression.

Argh. That's nasty.

Even the 3.25% is horrible. We scale so well in other situations that
it's really sad how the anon_vma lock is now one of our worst issues.

Anyway, please check me if I'm wrong, but won't the "anon_vma->root"
be the same for all the anon_vma's that are associated with one
particular vma?

The reason I ask is because when I look at anon_vma_clone(), we do that

   list_for_each_entry_reverse(pavc, &src->anon_vma_chain, same_vma) {
      ...
      anon_vma_chain_link(dst, avc, pavc->anon_vma);
   }

an dthen we do that anon_vma_lock()/unlock() dance on each of those
pavc->anon_vma's. But if the anon_vma->root is always the same, then
that would mean that we could do the lock just once, and hold it over
the loop.

Because I think the real problem with that anon_vma locking is that it
gets called so _much_. We'd be better off holding the lock for a
longer time, and just not do the lock/unlock thing so often. The
contention would go down simply because we wouldn't waste our time
with those atomic lock/unlock instructions as much.

Gaah. I knew exactly how the anon_vma locking worked a few months ago,
but it's complicated enough that I've swapped out all the details. So
I'm not at all sure that the anon_vma->root will be the same for every
anon_vma on the same_vma list.

Somebody hit me over the head with a clue-bat. Anybody?

                   Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/