[lkml]   [2013]   [Oct]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectAutomatic NUMA balancing patches for tip-urgent/stable
On Mon, Oct 07, 2013 at 11:28:38AM +0100, Mel Gorman wrote:
> This series has roughly the same goals as previous versions despite the
> size. It reduces overhead of automatic balancing through scan rate reduction
> and the avoidance of TLB flushes. It selects a preferred node and moves tasks
> towards their memory as well as moving memory toward their task. It handles
> shared pages and groups related tasks together. Some problems such as shared
> page interleaving and properly dealing with processes that are larger than
> a node are being deferred. This version should be ready for wider testing
> in -tip.

Hi Ingo,

Off-list we talked with Peter about the fact that automatic NUMA
balancing as merged in 3.10, 3.11 and 3.12 shortly may corrupt
userspace memory. There is one LKML report on this that I'm aware of -- which I prompt forgot to follow up
properly on . The user-visible effect is that pages get filled with zeros
with results such as null pointer exceptions in JVMs. It is fairly difficult
to trigger but it became much easier to trigger during the development of
the series "Basic scheduler support for automatic NUMA balancing" which
is how it was discovered and finally fixed.

In that series I tagged patches 2-9 for -stable as these patches addressed
the problem for me. I did not call it out as clearly as I should have
and did not realise the cc: stable tags were stripped. Worse, as it was
close to the release and the bug is relatively old I was ok with waiting
until 3.12 came out and then treat it as a -stable backport. It has been
highlighted that this is the wrong attitude and we should consider merging
the fixes now and backporting to -stable sooner rather than later.

The most important patches are

mm: Wait for THP migrations to complete during NUMA hinting fault
mm: Prevent parallel splits during THP migration
mm: Close races between THP migration and PMD numa clearing

but on their own they will cause conflicts with tricky fixups and -stable
would differ from mainline in annoying ways. Patches 2-9 have been heavily
tested in isolation so I'm reasonably confident they fix the problem and are
-stable material. While strictly speaking not all the patches are required
for the fix, the -stable kernels would then be directly comparable with
3.13 when the full NUMA balancing series is applied. If I rework them at
this point then I'll also have to retest delaying things until next week.

Please consider queueing patches 2-9 for 3.12 via -urgent if it is not
too late and preserve the cc: stable tags so Greg will pick them up


Mel Gorman

 \ /
  Last update: 2013-10-24 14:41    [W:0.698 / U:0.120 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site