Message-ID: <20170210075726.GA10893@dhcp22.suse.cz>
Date:   Fri, 10 Feb 2017 08:57:26 +0100
From:   Michal Hocko <mhocko@...nel.org>
To:     peter enderborg <peter.enderborg@...ymobile.com>
Cc:     devel@...verdev.osuosl.org,
        Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        linux-kernel@...r.kernel.org,
        Arve Hjønnevåg <arve@...roid.com>,
        Riley Andrews <riandrews@...roid.com>,
        Linus Torvalds <torvalds@...ux-foundation.org>,
        linux-mm@...ck.org
Subject: Re: [PATCH 3/3 staging-next] mm: Remove RCU and tasklocks from lmk

On Fri 10-02-17 08:39:11, peter enderborg wrote:
> On 02/09/2017 09:05 PM, Michal Hocko wrote:
> > On Thu 09-02-17 14:21:52, peter enderborg wrote:
> >> Fundamental changes:
> >> 1 Does NOT take any RCU lock in shrinker functions.
> >> 2 It returns same result for scan and counts, so  we dont need to do
> >>   shinker will know when it is pointless to call scan.
> >> 3 It does not lock any other process than the one that is
> >>   going to be killed.
> >>
> >> Background.
> >> The low memory killer scans for process that can be killed to free
> >> memory. This can be cpu consuming when there is a high demand for
> >> memory. This can be seen by analysing the kswapd0 task work.
> >> The stats function added in earler patch adds a counter for waste work.
> >>
> >> How it works.
> >> This patch create a structure within the lowmemory killer that caches
> >> the user spaces processes that it might kill. It is done with a
> >> sorted rbtree so we can very easy find the candidate to be killed,
> >> and knows its properies as memory usage and sorted by oom_score_adj
> >> to look up the task with highest oom_score_adj. To be able to achive
> >> this it uses oom_score_notify events.
> >>
> >> This patch also as a other effect, we are now free to do other
> >> lowmemorykiller configurations.  Without the patch there is a need
> >> for a tradeoff between freed memory and task and rcu locks. This
> >> is no longer a concern for tuning lmk. This patch is not intended
> >> to do any calculation changes other than we do use the cache for
> >> calculate the count values and that makes kswapd0 to shrink other
> >> areas.
> > I have to admit I really do not understand big part of the above
> > paragraph as well as how this all is supposed to work. A quick glance
> > over the implementation. __lmk_task_insert seems to be only called from
> > the oom_score notifier context. If nobody updates the value then no task
> > will get into the tree. Or am I missing something really obvious here?
> > Moreover oom scores tend to be mostly same for tasks. That means that
> > your sorted tree will become sorted by pids in most cases. I do not see
> > any sorting based on the rss nor any updates that would reflect updates
> > of rss. How can this possibly work?
> 
> The task tree nodes are created,updated or removed from the notifier when
> there is a relevant oom_score_adj change. If no one create a task that
> is in the range for the lowmemorykiller the tree will be empty. This is
> an android feature so the score will be updated very often. It is
> part of activity manager to prioritise tasks.  Why should we do sort of
> rss?
 
Because the current lmk selects the tasks based on rss. And the patch
doesn't explain why this is no longer suitable and a different metric
shoult be used. If you also consider that the scale of oom_score_adj is
quite small, conllisions when you simply sort based on pids which is
more than questionable. I really fail to see how this can work
reasonably and why the change of the lmk semantic is even acceptable.

-- 
Michal Hocko
SUSE Labs

Powered by blists - more mailing lists