Message-ID: <CALYGNiPkfYTPnyy646hTpYed4A=Xr_ee9pa-vMDDmA_MLoMAGA@mail.gmail.com>
Date:	Tue, 23 Sep 2014 00:46:10 +0400
From:	Konstantin Khlebnikov <koct9i@...il.com>
To:	Rafael Aquini <aquini@...hat.com>
Cc:	Andrew Morton <akpm@...ux-foundation.org>,
	Konstantin Khlebnikov <k.khlebnikov@...sung.com>,
	Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
	"linux-mm@...ck.org" <linux-mm@...ck.org>,
	Andrey Ryabinin <ryabinin.a.a@...il.com>,
	Sasha Levin <sasha.levin@...cle.com>
Subject: Re: [PATCH v2 4/6] mm: introduce common page state for ballooned memory

On Tue, Sep 23, 2014 at 12:22 AM, Rafael Aquini <aquini@...hat.com> wrote:
> On Tue, Sep 23, 2014 at 12:06:11AM +0400, Konstantin Khlebnikov wrote:
>> On Mon, Sep 22, 2014 at 11:22 PM, Rafael Aquini <aquini@...hat.com> wrote:
>> > On Mon, Sep 22, 2014 at 10:40:34PM +0400, Konstantin Khlebnikov wrote:
>> >> On Sat, Sep 20, 2014 at 10:23 AM, Andrew Morton
>> >> <akpm@...ux-foundation.org> wrote:
>> >> > On Sat, 20 Sep 2014 09:25:01 +0400 Konstantin Khlebnikov <koct9i@...il.com> wrote:
>> >> >
>> >> >> >
>> >> >> > So I'm going to send "fix for
>> >> >> > mm-balloon_compaction-use-common-page-ballooning-v2" to Linus
>> >> >> > separately, but it has no changelog at all.
>> >> >>
>> >> >> Probably it would be better if you drop everything except actually
>> >> >> fixes and stresstest. This is gone too far, now balloon won't compile
>> >> >> in the middle of patchset. Just tell me and I'll redo the rest.
>> >> >
>> >> > I think it's best if I drop everything:
>> >> >
>> >> > mm-balloon_compaction-ignore-anonymous-pages.patch
>> >> > mm-balloon_compaction-keep-ballooned-pages-away-from-normal-migration-path.patch
>> >> > mm-balloon_compaction-isolate-balloon-pages-without-lru_lock.patch
>> >> > selftests-vm-transhuge-stress-stress-test-for-memory-compaction.patch
>> >> > mm-introduce-common-page-state-for-ballooned-memory.patch
>> >> > mm-balloon_compaction-use-common-page-ballooning.patch
>> >> > mm-balloon_compaction-general-cleanup.patch
>> >> > mm-balloon_compaction-use-common-page-ballooning-v2-fix-1.patch
>> >> >
>> >> > Please go through it and send out a new version?
>> >> >
>> >> >
>> >>
>> >> I've found yet another bug in this code. It seems here is a nest.
>> >> balloon_page_dequeue can race with  balloon_page_isolate:
>> >> balloon_page_isolate can remove page from list between
>> >> llist_for_each_entry_safe and trylock_page in balloon_page_dequeue.
>> >> balloon_page_dequeue runs under mutex_lock(&vb->balloon_lock);
>> >> both of them lock page using trylock_page so race is tight but it is
>> >> not impossible.
>> > Plausible to happen if stress testing compaction simultaneously with
>> > freezing/unloading the balloon driver. As you noted, it's quite tight
>> > despite not impossible. Nice catch.
>> >
>> >
>> >> Probably it's really easier to rewrite it than to fix bugs one by one =/
>> > I'm not against a rewrite, but I don't think that rewriting the code to get rid
>> > of such bugs changes the fact we still have to address them in the actual placed
>> > code as we go on finding them. That's why I thought your inital changeset fine,
>> > with patches for stable going first and code overhaul for next following them up.
>> >
>> > For this race you spotted, I think a simple change like the following
>> > might be enough (not-tested)
>>
>> This locking scheme is too fragile and uncommon.
>>
>
> page_lock and refcounting was what I had at my disposal to sort these
> races out since I haven't thoutgh on a special page->_mapcount when
> designing this feature. It's the way other page races are sorted out.
> Not arguing it's not fragile, but it's the way code is layed out since
> some time, so we must check the feasibility of a total overhaul for
> stable branches.
>
>
>> What about this:
>>
>> * special page->_mapcount marks ballooned pages
>> * page->private points to balloon (directly, without intermediate mapping)
>> * flag PagePrivate means page currently in balloon page list (i.e. not
>> isolated, like PageLRU for normal pages)
>> * lock_page protects all of them
>>
>> balloon_page_dequeue() will delete page from balloon list only if it's
>> not isolated, also it always clears page->private and balloon mark.
>> put-back rechecks mark after locking the page and releases it as
>> normal page if mark is gone.
>>
>
> I have already agreed with you here, since the changes above are mostly from
> your original overhaul proposal. It's a much better approach for that
> balloon code, no doubts. Thanks for doing it. Only thing we need to take
> care here is about its requirement on changing the semantics for those
> interfaces might turn the changes unfeasible for old stable branches. If we
> can ignore this mentioned fact entirely, I don't see why not going with
> your idea all branches across, otherwise I think we should overhaul the
> code for -next, and send pontual fixes for stable.
>

Ok, I'll try to implement this approach in suitable for stable branches way.
Keeping several different versions is overkill for this code.

>
> Cheers,
> -- Rafael
>
>> >
>> > diff --git a/mm/balloon_compaction.c b/mm/balloon_compaction.c
>> > index 6e45a50..fd3a497 100644
>> > --- a/mm/balloon_compaction.c
>> > +++ b/mm/balloon_compaction.c
>> > @@ -93,6 +93,16 @@ struct page *balloon_page_dequeue(struct
>> > balloon_dev_info *b_dev_info)
>> >                  * to be released by the balloon driver.
>> >                  */
>> >                 if (trylock_page(page)) {
>> > +                       /*
>> > +                        * Skip dequeue attempt for this page to a later round
>> > +                        * if balloon_page_isolate() has sucessfully isolated
>> > +                        * it just before we got the page lock here.
>> > +                        */
>> > +                       if (page_count(page) != 1) {
>> > +                               unlock_page(page);
>> > +                               continue
>> > +                       }
>> > +
>> >                         spin_lock_irqsave(&b_dev_info->pages_lock, flags);
>> >                         /*
>> >                          * Raise the page refcount here to prevent any
>> >                          * wrong
>> >
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists