Re: [PATCH 6/10] mm/memcg: take care over pc->mem_cgroup

[Konstantin Khlebnikov]

Hugh Dickins wrote:
> On Tue, 21 Feb 2012, KAMEZAWA Hiroyuki wrote:
> > On Mon, 20 Feb 2012 15:34:28 -0800 (PST)
> > Hugh Dickins<hughd@xxxxxxxxxx>  wrote:
> > 	return NULL;
> > > +	lruvec = page_lock_lruvec(page);
> > >   	lock_page_cgroup(pc);
> >
> > Do we need to take lrulock+irq disable per page in this very very hot path ?
>
> I'm sure we don't want to: I hope you were pleased to find it goes away
> (from most cases) a couple of patches later.
>
> I had lruvec lock nested inside page_cgroup lock in the rollup I sent in
> December, whereas you went for page_cgroup lock nested inside lruvec lock
> in your lrucare patch.
>
> I couldn't find an imperative reason why they should be one way round or
> the other, so I tried hard to stick with your ordering, and it did work
> (in this 6/10).  But then I couldn't work out how to get rid of the
> overheads added in doing it this way round, so swapped them back.
>
> > Hmm.... How about adding NR_ISOLATED counter into lruvec ?
> >
> > Then, we can delay freeing lruvec until all conunters goes down to zero.
> > as...
> >
> > 	bool we_can_free_lruvec = true;
> >
> > 	lock_lruvec(lruvec->lock);
> > 	for_each_lru_lruvec(lru)
> > 		if (!list_empty(&lruvec->lru[lru]))
> > 			we_can_free_lruvec = false;
> > 	if (lruvec->nr_isolated)
> > 		we_can_free_lruvec = false;
> > 	unlock_lruvec(lruvec)
> > 	if (we_can_free_lruvec)
> > 		kfree(lruvec);
> >
> > If compaction, lumpy reclaim free a page taken from LRU,
> > it knows what it does and can decrement lruvec->nr_isolated properly
> > (it seems zone's NR_ISOLATED is decremented at putback.)
>
> At the moment I'm thinking that what we end up with by 9/10 is
> better than adding such a refcount.  But I'm not entirely happy with
> mem_cgroup_reset_uncharged_to_root (it adds a further page_cgroup
> lookup just after I got rid of some others), and need yet to think
> about the race which Konstantin posits, so all options remain open.

This lruvec->nr_isolated seem reasonable, and its managegin not very costly.
In move_account() we anyway need to touch old_lruvec->lru_lock after recharge,
to stabilize PageLRU() before adding page to new_lruvec. (because that race)
In migration/compaction this handled automatically, because they always call putback_lru_page() at the end.
Main problem is shrink_page_list() for lumpy-reclaim, but seems like it never used if memory
compaction is enabled, so it can be slow and ineffective with tons of lru_list relocks.

> Hugh

--
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/