Re: [PATCH RFC 3/3] slub: reparent memcg caches' slabs on memcg offline
From: Christoph Lameter
Date: Mon May 19 2014 - 12:03:25 EST
On Mon, 19 May 2014, Vladimir Davydov wrote:
> > I doubt that. The accounting occurs when a new cpu slab page is allocated.
> > But the individual allocations in the fastpath are not accounted to a
> > specific group. Thus allocation in a slab page can belong to various
> > cgroups.
>
> On each kmalloc, we pick the cache that belongs to the current memcg,
> and allocate objects from that cache (see memcg_kmem_get_cache()). And
> all slab pages allocated for a per memcg cache are accounted to the
> memcg the cache belongs to (see memcg_charge_slab). So currently, each
> kmem cache, i.e. each slab of it, can only have objects of one cgroup,
> namely its owner.
Ok that works for kmalloc. What about dentry/inodes and so on?
> OK, it seems we have no choice but keeping dead caches left after memcg
> offline until they have active slabs. How can we get rid of them then?
Then they are moved to a list and therefore you can move them to yours I
think.
> Simply counting slabs on cache and destroying cache when the count goes
> to 0 isn't enough, because slub may keep some free slabs by default (if
> they are frozen e.g.) Reaping them periodically doesn't look nice.
But those are only limited to one slab per cpu ( plus eventual cpu partial
ones but you can switch that feature off).
> What if we modify __slab_free so that it won't keep empty slabs for dead
> caches? That way we would only have to count slabs allocated to a cache,
> and destroy caches as soon as the counter drops to 0. No
Well that should already be in there. Se s->min_partial to zero?
> periodic/vmpressure reaping would be necessary. I attached the patch
> that does the trick below. The changes it introduces to __slab_free do
> not look very intrusive to me. Could you please take a look at it (to
> diff slub.c primarily) when you have time, and say if, in your opinion,
> the changes to __slab_free are acceptable or not?
Looking now.
> @@ -2620,14 +2651,16 @@ static void __slab_free(struct kmem_cache *s, struct page *page,
> return;
> }
>
> - if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
> + if (unlikely(!new.inuse &&
> + (n->nr_partial > s->min_partial || cache_dead)))
> goto slab_empty;
Could you set s->min_partial = 0 to avoid this?
>
> /*
> * Objects left in the slab. If it was not on the partial list before
> * then add it.
> */
> - if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
> + if ((!kmem_cache_has_cpu_partial(s) || cache_dead) &&
> + unlikely(!prior)) {
> if (kmem_cache_debug(s))
> remove_full(s, n, page);
> add_partial(n, page, DEACTIVATE_TO_TAIL);
Not sure why we need this and the other stuff.
--
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/