Message-ID: <BANLkTi=esh86EtfQgRGWkHx6Z3fYV+42tA@mail.gmail.com>
Date:	Wed, 18 May 2011 13:19:37 +0900
From:	Minchan Kim <minchan.kim@...il.com>
To:	Colin Ian King <colin.king@...onical.com>
Cc:	Mel Gorman <mgorman@...e.de>,
	James Bottomley <James.Bottomley@...senpartnership.com>,
	KOSAKI Motohiro <kosaki.motohiro@...fujitsu.com>,
	akpm@...ux-foundation.org, raghu.prabhu13@...il.com, jack@...e.cz,
	chris.mason@...cle.com, cl@...ux.com, penberg@...nel.org,
	riel@...hat.com, hannes@...xchg.org, linux-fsdevel@...r.kernel.org,
	linux-mm@...ck.org, linux-kernel@...r.kernel.org,
	linux-ext4@...r.kernel.org
Subject: Re: [PATCH 4/4] mm: vmscan: If kswapd has been running too long,
 allow it to sleep

Hello Colin,

On Tue, May 17, 2011 at 10:50 PM, Colin Ian King
<colin.king@...onical.com> wrote:
> On Tue, 2011-05-17 at 11:38 +0100, Mel Gorman wrote:
>> On Tue, May 17, 2011 at 08:50:44AM +0900, Minchan Kim wrote:
>> > On Mon, May 16, 2011 at 7:27 PM, Mel Gorman <mgorman@...e.de> wrote:
>> > > On Mon, May 16, 2011 at 05:58:59PM +0900, Minchan Kim wrote:
>> > >> On Mon, May 16, 2011 at 5:45 PM, Mel Gorman <mgorman@...e.de> wrote:
>> > >> > On Mon, May 16, 2011 at 02:04:00PM +0900, Minchan Kim wrote:
>> > >> >> On Mon, May 16, 2011 at 1:21 PM, James Bottomley
>> > >> >> <James.Bottomley@...senpartnership.com> wrote:
>> > >> >> > On Sun, 2011-05-15 at 19:27 +0900, KOSAKI Motohiro wrote:
>> > >> >> >> (2011/05/13 23:03), Mel Gorman wrote:
>> > >> >> >> > Under constant allocation pressure, kswapd can be in the situation where
>> > >> >> >> > sleeping_prematurely() will always return true even if kswapd has been
>> > >> >> >> > running a long time. Check if kswapd needs to be scheduled.
>> > >> >> >> >
>> > >> >> >> > Signed-off-by: Mel Gorman<mgorman@...e.de>
>> > >> >> >> > ---
>> > >> >> >> >   mm/vmscan.c |    4 ++++
>> > >> >> >> >   1 files changed, 4 insertions(+), 0 deletions(-)
>> > >> >> >> >
>> > >> >> >> > diff --git a/mm/vmscan.c b/mm/vmscan.c
>> > >> >> >> > index af24d1e..4d24828 100644
>> > >> >> >> > --- a/mm/vmscan.c
>> > >> >> >> > +++ b/mm/vmscan.c
>> > >> >> >> > @@ -2251,6 +2251,10 @@ static bool sleeping_prematurely(pg_data_t *pgdat, int order, long remaining,
>> > >> >> >> >     unsigned long balanced = 0;
>> > >> >> >> >     bool all_zones_ok = true;
>> > >> >> >> >
>> > >> >> >> > +   /* If kswapd has been running too long, just sleep */
>> > >> >> >> > +   if (need_resched())
>> > >> >> >> > +           return false;
>> > >> >> >> > +
>> > >> >> >>
>> > >> >> >> Hmm... I don't like this patch so much. because this code does
>> > >> >> >>
>> > >> >> >> - don't sleep if kswapd got context switch at shrink_inactive_list
>> > >> >> >
>> > >> >> > This isn't entirely true:  need_resched() will be false, so we'll follow
>> > >> >> > the normal path for determining whether to sleep or not, in effect
>> > >> >> > leaving the current behaviour unchanged.
>> > >> >> >
>> > >> >> >> - sleep if kswapd didn't
>> > >> >> >
>> > >> >> > This also isn't entirely true: whether need_resched() is true at this
>> > >> >> > point depends on a whole lot more that whether we did a context switch
>> > >> >> > in shrink_inactive. It mostly depends on how long we've been running
>> > >> >> > without giving up the CPU.  Generally that will mean we've been round
>> > >> >> > the shrinker loop hundreds to thousands of times without sleeping.
>> > >> >> >
>> > >> >> >> It seems to be semi random behavior.
>> > >> >> >
>> > >> >> > Well, we have to do something.  Chris Mason first suspected the hang was
>> > >> >> > a kswapd rescheduling problem a while ago.  We tried putting
>> > >> >> > cond_rescheds() in several places in the vmscan code, but to no avail.
>> > >> >>
>> > >> >> Is it a result of  test with patch of Hannes(ie, !pgdat_balanced)?
>> > >> >>
>> > >> >> If it isn't, it would be nop regardless of putting cond_reshed at vmscan.c.
>> > >> >> Because, although we complete zone balancing, kswapd doesn't sleep as
>> > >> >> pgdat_balance returns wrong result. And at last VM calls
>> > >> >> balance_pgdat. In this case, balance_pgdat returns without any work as
>> > >> >> kswap couldn't find zones which have not enough free pages and goto
>> > >> >> out. kswapd could repeat this work infinitely. So you don't have a
>> > >> >> chance to call cond_resched.
>> > >> >>
>> > >> >> But if your test was with Hanne's patch, I am very curious how come
>> > >> >> kswapd consumes CPU a lot.
>> > >> >>
>> > >> >> > The need_resched() in sleeping_prematurely() seems to be about the best
>> > >> >> > option.  The other option might be just to put a cond_resched() in
>> > >> >> > kswapd_try_to_sleep(), but that will really have about the same effect.
>> > >> >>
>> > >> >> I don't oppose it but before that, I think we have to know why kswapd
>> > >> >> consumes CPU a lot although we applied Hannes' patch.
>> > >> >>
>> > >> >
>> > >> > Because it's still possible for processes to allocate pages at the same
>> > >> > rate kswapd is freeing them leading to a situation where kswapd does not
>> > >> > consider the zone balanced for prolonged periods of time.
>> > >>
>> > >> We have cond_resched in shrink_page_list, shrink_slab and balance_pgdat.
>> > >> So I think kswapd can be scheduled out although it's scheduled in
>> > >> after a short time as task scheduled also need page reclaim. Although
>> > >> all task in system need reclaim, kswapd cpu 99% consumption is a
>> > >> natural result, I think.
>> > >> Do I miss something?
>> > >>
>> > >
>> > > Lets see;
>> > >
>> > > shrink_page_list() only applies if inactive pages were isolated
>> > >        which in turn may not happen if all_unreclaimable is set in
>> > >        shrink_zones(). If for whatver reason, all_unreclaimable is
>> > >        set on all zones, we can miss calling cond_resched().
>> > >
>> > > shrink_slab only applies if we are reclaiming slab pages. If the first
>> > >        shrinker returns -1, we do not call cond_resched(). If that
>> > >        first shrinker is dcache and __GFP_FS is not set, direct
>> > >        reclaimers will not shrink at all. However, if there are
>> > >        enough of them running or if one of the other shrinkers
>> > >        is running for a very long time, kswapd could be starved
>> > >        acquiring the shrinker_rwsem and never reaching the
>> > >        cond_resched().
>> >
>> > Don't we have to move cond_resched?
>> >
>> > diff --git a/mm/vmscan.c b/mm/vmscan.c
>> > index 292582c..633e761 100644
>> > --- a/mm/vmscan.c
>> > +++ b/mm/vmscan.c
>> > @@ -231,8 +231,10 @@ unsigned long shrink_slab(struct shrink_control *shrink,
>> >         if (scanned == 0)
>> >                 scanned = SWAP_CLUSTER_MAX;
>> >
>> > -       if (!down_read_trylock(&shrinker_rwsem))
>> > -               return 1;       /* Assume we'll be able to shrink next time */
>> > +       if (!down_read_trylock(&shrinker_rwsem)) {
>> > +               ret = 1;
>> > +               goto out; /* Assume we'll be able to shrink next time */
>> > +       }
>> >
>> >         list_for_each_entry(shrinker, &shrinker_list, list) {
>> >                 unsigned long long delta;
>> > @@ -280,12 +282,14 @@ unsigned long shrink_slab(struct shrink_control *shrink,
>> >                         count_vm_events(SLABS_SCANNED, this_scan);
>> >                         total_scan -= this_scan;
>> >
>> > -                       cond_resched();
>> >                 }
>> >
>> >                 shrinker->nr += total_scan;
>> > +               cond_resched();
>> >         }
>> >         up_read(&shrinker_rwsem);
>> > +out:
>> > +       cond_resched();
>> >         return ret;
>> >  }
>> >
>>
>> This makes some sense for the exit path but if one or more of the
>> shrinkers takes a very long time without sleeping (extremely long
>> list searches for example) then kswapd will not call cond_resched()
>> between shrinkers and still consume a lot of CPU.
>>
>> > >
>> > > balance_pgdat() only calls cond_resched if the zones are not
>> > >        balanced. For a high-order allocation that is balanced, it
>> > >        checks order-0 again. During that window, order-0 might have
>> > >        become unbalanced so it loops again for order-0 and returns
>> > >        that was reclaiming for order-0 to kswapd(). It can then find
>> > >        that a caller has rewoken kswapd for a high-order and re-enters
>> > >        balance_pgdat() without ever have called cond_resched().
>> >
>> > If kswapd reclaims order-o followed by high order, it would have a
>> > chance to call cond_resched in shrink_page_list. But if all zones are
>> > all_unreclaimable is set, balance_pgdat could return any work. Okay.
>> > It does make sense.
>> > By your scenario, someone wakes up kswapd with higher order, again.
>> > So re-enters balance_pgdat without ever have called cond_resched.
>> > But if someone wakes up higher order again, we can't have a chance to
>> > call kswapd_try_to_sleep. So your patch effect would be nop, too.
>> >
>> > It would be better to put cond_resched after balance_pgdat?
>> >
>>
>> Which will leave kswapd runnable instead of going to sleep but
>> guarantees a scheduling point. Lets see if the problem is that
>> cond_resched is being missed although if this was the case then patch
>> 4 would truly be a no-op but Colin has already reported that patch 1 on
>> its own didn't fix his problem. If the problem is sandybridge-specific
>> where kswapd remains runnable and consuming large amounts of CPU in
>> turbo mode then we know that there are other cond_resched() decisions
>> that will need to be revisited.
>>
>> Colin or James, would you be willing to test with patch 1 from this
>> series and Minchan's patch below? Thanks.
>
> This works OK fine.  Ran 250 test cycles for about 2 hours.

Thanks for the testing!.
I would like to know exact patch for you to apply.
My modification of inserting cond_resched is two.

1) shrink_slab function
2) kswapd right after balance_pgdat.

1) or 2) ?
Or
Both?


Thanks
-- 
Kind regards,
Minchan Kim
--
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