Re: [PATCH next] mm/swap.c: reduce lock contention in lru_cache_add

[Vlastimil Babka]

On 11/26/20 4:12 AM, Alex Shi wrote:
> 在 2020/11/25 下午11:38, Vlastimil Babka 写道:
> > On 11/20/20 9:27 AM, Alex Shi wrote:
> > > The current relock logical will change lru_lock when found a new
> > > lruvec, so if 2 memcgs are reading file or alloc page at same time,
> > > they could hold the lru_lock alternately, and wait for each other for
> > > fairness attribute of ticket spin lock.
> > >
> > > This patch will sort that all lru_locks and only hold them once in
> > > above scenario. That could reduce fairness waiting for lock reget.
> > > Than, vm-scalability/case-lru-file-readtwice could get ~5% performance
> > > gain on my 2P*20core*HT machine.
> >
> > Hm, once you sort the pages like this, it's a shame not to splice them instead of more list_del() + list_add() iterations. update_lru_size() could be also called once?
>
> Yes, looks it's a good idea to use splice instead of list_del/add, but pages
> may on different lru list in a same lruvec, and also may come from different
> zones. That could involve 5 cycles for different lists, and more for zones...

Hmm, zones wouldn't affect splicing (there's a per-node lru these days), but 
would affect accounting. And yeah, there are 5 lru lists, and we probably need 
to be under lru lock to stabilize where a page belongs, so pre-sorting without 
lock wouldn't be safe? Bummer.

> I give up the try.