lkml.org 
[lkml]   [2017]   [Nov]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: hugetlb page migration vs. overcommit
Date
On Tue 28-11-17 11:19:07, Michal Hocko wrote:
> On Wed 22-11-17 16:28:32, Michal Hocko wrote:
> > Hi,
> > is there any reason why we enforce the overcommit limit during hugetlb
> > pages migration? It's in alloc_huge_page_node->__alloc_buddy_huge_page
> > path. I am wondering whether this is really an intentional behavior.
> > The page migration allocates a page just temporarily so we should be
> > able to go over the overcommit limit for the migration duration. The
> > reason I am asking is that hugetlb pages tend to be utilized usually
> > (otherwise the memory would be just wasted and pool shrunk) but then
> > the migration simply fails which breaks memory hotplug and other
> > migration dependent functionality which is quite suboptimal. You can
> > workaround that by increasing the overcommit limit.
> >
> > Why don't we simply migrate as long as we are able to allocate the
> > target hugetlb page? I have a half baked patch to remove this
> > restriction, would there be an opposition to do something like that?
>
> So I finally got to think about this some more and looked at how we
> actually account things more thoroughly. And it is, you both of you
> expected, quite subtle and not easy to get around. Per NUMA pools make
> things quite complicated. Why? Migration can really increase the overall
> pool size. Say we are migrating from Node1 to Node2. Node2 doesn't have
> any pre-allocated pages but assume that the overcommit allows us to move
> on. All good. Except that the original page will return to the pool
> because free_huge_page will see Node1 without any surplus pages and
> therefore moves back the page to the pool. Node2 will release the
> surplus page only after it is freed which can be an unbound amount of
> time.
>
> While we are still effectively under the overcommit limit the semantic
> is kind of strange and I am not sure the behavior is really intended.
> I see why per node surplus counter is used here. We simply want to
> maintain per node counts after regular page free. So I was thinking
> to add a temporary/migrate state to the huge page for migration pages
> (start with new page, state transfered to the old page on success) and
> free such a page to the allocator regardless of the surplus counters.
>
> This would mean that the page migration might change inter node pool
> sizes but I guess that should be acceptable. What do you guys think?
> I can send a draft patch if that helps you to understand the idea.

This is what I have currently and it seems to work (or at least it
doesn't it doesn't blow up immediately). The first patch is a cleanup
and patch2 implements the temporary page idea.

Does this make any sense to you at all?
--
Michal Hocko

\
 
 \ /
  Last update: 2017-11-28 15:14    [W:0.110 / U:21.080 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site