Hmm. Andrew wasn't cc'd on this series, and usually things like this
go through the -mm tree.
Maybe Andrew saw it by virtue of the linux-mm list, but maybe he
didn't. So here he is cc'd directly.
The series looks reasonable to me,
Linus
On Thu, May 5, 2011 at 12:32 PM, Andi Kleen <andi@xxxxxxxxxxxxxx> wrote:
> 012f18004da33ba67 in 2.6.36 caused a significant performance regression in
> fork/exit intensive workloads with a lot of sharing. The problem is that
> fork/exit now contend heavily on the lock of the root anon_vma.
>
> This patchkit attempts to lower this a bit by batching the lock acquisions.
> Right now the lock is taken for every shared vma individually. This
> patchkit batches this and only reaquires the lock when actually needed.
>
> When multiple processes are doing this in parallel, they will now
> spend much less time bouncing the lock cache line around. In addition
> there should be also lower overhead in the uncontended case because
> locks are relatively slow (not measured)
>
> This doesn't completely fix the regression on a 4S system, but cuts
> it down somewhat. One particular workload suffering from this gets
> about 5% faster.
>
> This is essentially a micro optimization that just tries to mitigate
> the problem a bit.
>
> Better would be to switch back to more local locking like .35 had, but I
> guess then we would be back with the old deadlocks? I was thinking also of
> adding some deadlock avoidance as an alternative.
>
> -Andi
>
--
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/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>