linux-kernel - Re: [PATCH] mm: thp: move deferred split queue to memcg's nodeinfo

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20191008144437.fr374cxtpnrnnjsv@box>
Date:   Tue, 8 Oct 2019 17:44:37 +0300
From:   "Kirill A. Shutemov" <kirill@...temov.name>
To:     Michal Hocko <mhocko@...nel.org>
Cc:     Vlastimil Babka <vbabka@...e.cz>,
        Yang Shi <yang.shi@...ux.alibaba.com>,
        kirill.shutemov@...ux.intel.com, ktkhai@...tuozzo.com,
        hannes@...xchg.org, hughd@...gle.com, shakeelb@...gle.com,
        rientjes@...gle.com, akpm@...ux-foundation.org, linux-mm@...ck.org,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH] mm: thp: move deferred split queue to memcg's nodeinfo

On Mon, Oct 07, 2019 at 04:30:30PM +0200, Michal Hocko wrote:
> On Mon 07-10-19 16:19:59, Vlastimil Babka wrote:
> > On 10/2/19 10:43 AM, Michal Hocko wrote:
> > > On Wed 02-10-19 06:16:43, Yang Shi wrote:
> > >> The commit 87eaceb3faa59b9b4d940ec9554ce251325d83fe ("mm: thp: make
> > >> deferred split shrinker memcg aware") makes deferred split queue per
> > >> memcg to resolve memcg pre-mature OOM problem.  But, all nodes end up
> > >> sharing the same queue instead of one queue per-node before the commit.
> > >> It is not a big deal for memcg limit reclaim, but it may cause global
> > >> kswapd shrink THPs from a different node.
> > >>
> > >> And, 0-day testing reported -19.6% regression of stress-ng's madvise
> > >> test [1].  I didn't see that much regression on my test box (24 threads,
> > >> 48GB memory, 2 nodes), with the same test (stress-ng --timeout 1
> > >> --metrics-brief --sequential 72  --class vm --exclude spawn,exec), I saw
> > >> average -3% (run the same test 10 times then calculate the average since
> > >> the test itself may have most 15% variation according to my test)
> > >> regression sometimes (not every time, sometimes I didn't see regression
> > >> at all).
> > >>
> > >> This might be caused by deferred split queue lock contention.  With some
> > >> configuration (i.e. just one root memcg) the lock contention my be worse
> > >> than before (given 2 nodes, two locks are reduced to one lock).
> > >>
> > >> So, moving deferred split queue to memcg's nodeinfo to make it NUMA
> > >> aware again.
> > >>
> > >> With this change stress-ng's madvise test shows average 4% improvement
> > >> sometimes and I didn't see degradation anymore.
> > > 
> > > My concern about this getting more and more complex
> > > (http://lkml.kernel.org/r/20191002084014.GH15624@dhcp22.suse.cz) holds
> > > here even more. Can we step back and reconsider the whole thing please?
> > 
> > What about freeing immediately after split via workqueue and also have a
> > synchronous version called before going oom? Maybe there would be also
> > other things that would benefit from this scheme instead of traditional
> > reclaim and shrinkers?
> 
> That is exactly what we have discussed some time ago.

Yes, I've posted the patch:

http://lkml.kernel.org/r/20190827125911.boya23eowxhqmopa@box

But I still not sure that the approach is right. I expect it to trigger
performance regressions. For system with pleanty of free memory, we will
just pay split cost for nothing in many cases.

-- 
 Kirill A. Shutemov