----- Original Message -----
From: Michal Hocko <mhocko@xxxxxxx>
To: Felipe Contreras <felipe.contreras@xxxxxxxxx>
Cc: Linux Kernel Mailing List <linux-kernel@xxxxxxxxxxxxxxx>, "linux-mm@xxxxxxxxx" <linux-mm@xxxxxxxxx>, Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>, Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>, Mel Gorman <mgorman@xxxxxxx>, KAMEZAWA Hiroyuki <kamezawa.hiro
Subject: Re: Interactivity regression since v3.11 in mm/vmscan.c
Date: 2014-06-06 17:16
On Thu 05-06-14 09:00:10, Felipe Contreras wrote:
> On Thu, Jun 5, 2014 at 8:37 AM, Michal Hocko <mhocko@xxxxxxx> wrote:
> > On Thu 05-06-14 06:33:40, Felipe Contreras wrote:
>
> >> For a while I've noticed that my machine bogs down in certain
> >> situations, usually while doing heavy I/O operations, it is not just the
> >> I/O operations, but everything, including the graphical interface, even
> >> the mouse pointer.
> >>
> >> As far as I can recall this did not happen in the past.
> >>
> >> I noticed this specially on certain operations, for example updating a
> >> a game on Steam (to an exteranl USB 3.0 device), or copying TV episodes
> >> to a USB memory stick (probably flash-based).
> >
> > We had a similar report for opensuse. The common part was that there was
> > an IO to a slow USB device going on.
>
> Well, it's a USB 3.0 device, I can write at 250 MB/s, so it's not
> really that slow.
>
> And in fact, when I read and write to and from the same USB 3.0
> device, I don't see the issue.
>
> >> Then I went back to the latest stable version (v3.14.5), and commented
> >> out the line I think is causing the slow down:
> >>
> >> if (nr_unqueued_dirty == nr_taken || nr_immediate)
> >> congestion_wait(BLK_RW_ASYNC, HZ/10);
> >
> > Yes, I came to the same check. I didn't have any confirmation yet so
> > thanks for your confirmation. I've suggested to reduce this
> > congestion_wait only to kswapd:
> > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > index 32c661d66a45..ef6a1c0e788c 100644
> > --- a/mm/vmscan.c
> > +++ b/mm/vmscan.c
> > @@ -1566,7 +1566,7 @@ shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
> > * implies that pages are cycling through the LRU faster than
> > * they are written so also forcibly stall.
> > */
> > - if (nr_unqueued_dirty == nr_taken || nr_immediate)
> > + if ((nr_unqueued_dirty == nr_taken || nr_immediate) && current_is_kswapd())
> > congestion_wait(BLK_RW_ASYNC, HZ/10);
> > }
>
> Unfortunately that doesn't fix the issue for me.
>That is really interesting. So removing the test completely helps but
>reducing it to kswapd doesn't. I would expect stalls coming from direct
>reclaimers not the kswapd.
>Mel has a nice systemtap script (attached) to watch for stalls. Maybe
>you can give it a try?
>
Alternatively can we try wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10) ?
Hillf?韬{.n?壏?檩jg??a旃??迻)钋?韰骅w+h櫒璀?ｙ/i?⒏?婃⒎?湼Щ 烚m?? ?)钋?痂燐^櫒?觥櫒ザ?v夗﨣趓O璁?f?i?⒏?