lkml.org 
[lkml]   [2014]   [Jan]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] mm, hugetlb: gimme back my page
From
Date
On Thu, 2014-01-30 at 10:59 +0100, Michal Hocko wrote:
> On Wed 29-01-14 22:37:03, Davidlohr Bueso wrote:
> > From: Davidlohr Bueso <davidlohr@hp.com>
> >
> > While testing some changes, I noticed an issue triggered by the libhugetlbfs
> > test-suite. This is caused by commit 309381fe (mm: dump page when hitting a
> > VM_BUG_ON using VM_BUG_ON_PAGE), where an application can unexpectedly OOM due
> > to another program that using, or reserving, pool_size-1 pages later triggers
> > a VM_BUG_ON_PAGE and thus greedly leaves no memory to the rest of the hugetlb
> > aware tasks. For example, in libhugetlbfs 2.14:
> >
> > mmap-gettest 10 32783 (2M: 64): <---- hit VM_BUG_ON_PAGE
> > mmap-cow 32782 32783 (2M: 32): FAIL Failed to create shared mapping: Cannot allocate memory
> > mmap-cow 32782 32783 (2M: 64): FAIL Failed to create shared mapping: Cannot allocate memory
> >
> > While I have not looked into why 'mmap-gettest' keeps failing, it is of no
> > importance to this particular issue. This problem is similar to why we have
> > the hugetlb_instantiation_mutex, hugepages are quite finite.
> >
> > Revert the use of VM_BUG_ON_PAGE back to just VM_BUG_ON.
>
> I do not understand what VM_BUG_ON_PAGE has to do with the above
> failure. Could you be more specific.
>
> Hmm, now that I am looking into dump_page_badflags it shouldn't call
> mem_cgroup_print_bad_page for hugetlb pages because it doesn't make any
> sense. I will post a patch for that but that still doesn't explain the
> above changelog.

Yeah, I then looked closer at it and realized it doesn't make much
sense. I don't know why I thought a new page was being used. In any
case, bisection still shows the commit in question as the cause of the
regression. I will continue looking into it.



\
 
 \ /
  Last update: 2014-01-30 19:01    [W:0.080 / U:0.100 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site