On Thu, Oct 31, 2019 at 10:41:51AM -0400, Johannes Weiner wrote:
> On Thu, Oct 31, 2019 at 01:52:44AM +0000, Roman Gushchin wrote:
> > On Fri, Oct 25, 2019 at 03:41:18PM -0400, Johannes Weiner wrote:
> > > @@ -3117,15 +3095,24 @@ void __memcg_kmem_uncharge(struct page *page, int order)
> > > css_put_many(&memcg->css, nr_pages);
> > > }
> > >
> > > -int __memcg_kmem_charge_subpage(struct mem_cgroup *memcg, size_t size,
> > > - gfp_t gfp)
> > > +int obj_cgroup_charge(struct obj_cgroup *objcg, size_t size, gfp_t gfp)
> > > {
> > > - return try_charge(memcg, gfp, size, true);
> > > + int ret;
> > > +
> > > + if (consume_obj_stock(objcg, nr_bytes))
> > > + return 0;
> > > +
> > > + ret = try_charge(objcg->memcg, gfp, 1);
> > > + if (ret)
> > > + return ret;
>
> > The second problem is also here. If a task belonging to a different memcg
> > is scheduled on this cpu, most likely we will need to refill both stocks,
> > even if we need only a small temporarily allocation.
>
> Yes, that's a good thing. The reason we have the per-cpu caches in the
> first place is because most likely the same cgroup will perform
> several allocations. Both the slab allocator and the page allocator
> have per-cpu caches for the same reason. I don't really understand
> what the argument is.
I mean it seems strange (and most likely will show up in perf numbers)
to move a page from one stock to another. Is there a reason why do you want
to ask try_charge() and stock only a single page?
Can we do the following instead?
1) add a boolean argument to try_charge() to bypass the consume_stock() call
at the beginning and just go slow path immediately
2) use try_charge() with this argument set to true to fill the objc/subpage
stock with MEMCG_CHARGE_BATCH pages
In this case we'll reuse try_charge() and will have all corresponding benefits,
but will avoid the double stocking, which seems as a strange idea to me.
>
> > > +
> > > + refill_obj_stock(objcg, PAGE_SIZE - size);
> >
> > And the third problem is here. Percpu allocations (on which accounting I'm
> > working right now) can be larger than a page.
>
> How about this?
>
> nr_pages = round_up(size, PAGE_SIZE);
> try_charge(objcg->memcg, nr_pages);
> refill_obj_stock(objcg, size % PAGE_SIZE);
Ok, this will work.
>
> > This is fairly small issue in comparison to the first one. But it illustrates
> > well the main point: we can't simple get a page from the existing API and
> > sublease it in parts. The problem is that we need to break the main principle
> > that a page belongs to a single memcg.
>
> We can change the underlying assumptions of the existing API if they
> are no longer correct. We don't have to invent a parallel stack.
Ok, this makes sense. And thank you for the patch, I'll take it into the set.
Thanks!
