lkml.org 
[lkml]   [2014]   [Jul]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [patch] x86, perf: avoid spamming kernel log for bts buffer failure
On Wed, 2 Jul 2014, Peter Zijlstra wrote:

> > David and I discussed this. He can probably add more background
> > info, if needed.
>
> It would still be good to see why compaction etc is failing.
>

"Why compaction is failing" has been the story of my life for the past few
weeks, unfortunately. One person ran into this and here is the breakdown
at the time of the page allocation failure:

Node 0 DMA32: 12478*4kB 7595*8kB 2087*16kB 26*32kB 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 144896kB
Node 0 Normal: 55037*4kB 0*8kB 0*16kB 0*32kB 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 1*2048kB 0*4096kB = 222196kB
Node 1 Normal: 165860*4kB 18*8kB 1*16kB 0*32kB 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB = 663600kB

So we really don't have any order-4 or higher memory and if compaction and
reclaim fails, then the allocation failure is printed every time. This is
what the patch is trying to address and we can't guarantee order-4 is
always allocatable even with GFP_KERNEL. This allocation is just outside
the PAGE_ALLOC_COSTLY_ORDER == 3 threshold that would have oom killed
something and retried; the oom killer is deferred for this case because
there's no guarantee killing a process would have resulted in order-4
memory being available (and nobody wants killing when they are 902MB above
the per-zone min watermarks like this poor guy).


\
 
 \ /
  Last update: 2014-07-03 01:41    [W:0.081 / U:0.436 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site