lkml.org 
[lkml]   [2013]   [Mar]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC/PATCH 0/5] Contiguous Memory Allocator and get_user_pages()
Hello,

On 3/6/2013 9:47 AM, Minchan Kim wrote:
> Hello,
>
> On Tue, Mar 5, 2013 at 3:57 PM, Marek Szyprowski
> <m.szyprowski@samsung.com> wrote:
> > Hello,
> >
> > Contiguous Memory Allocator is very sensitive about migration failures
> > of the individual pages. A single page, which causes permanent migration
> > failure can break large conitguous allocations and cause the failure of
> > a multimedia device driver.
> >
> > One of the known issues with migration of CMA pages are the problems of
> > migrating the anonymous user pages, for which the others called
> > get_user_pages(). This takes a reference to the given user pages to let
> > kernel to operate directly on the page content. This is usually used for
> > preventing swaping out the page contents and doing direct DMA to/from
> > userspace.
> >
> > To solving this issue requires preventing locking of the pages, which
> > are placed in CMA regions, for a long time. Our idea is to migrate
> > anonymous page content before locking the page in get_user_pages(). This
> > cannot be done automatically, as get_user_pages() interface is used very
> > often for various operations, which usually last for a short period of
> > time (like for example exec syscall). We have added a new flag
> > indicating that the given get_user_space() call will grab pages for a
> > long time, thus it is suitable to use the migration workaround in such
> > cases.
> >
> > The proposed extensions is used by V4L2/VideoBuf2
> > (drivers/media/v4l2-core/videobuf2-dma-contig.c), but that is not the
> > only place which might benefit from it, like any driver which use DMA to
> > userspace with get_user_pages(). This one is provided to demonstrate the
> > use case.
> >
> > I would like to hear some comments on the presented approach. What do
> > you think about it? Is there a chance to get such workaround merged at
> > some point to mainline?
> >
>
> I discussed similar patch from memory-hotplug guys with Mel.
> Look at http://marc.info/?l=linux-mm&m=136014458829566&w=2
>
> The conern is that we ends up forcing using FOLL_DURABLE/GUP_NM for
> all drivers and subsystems for making sure CMA/memory-hotplug works
> well.
>
> You mentioned driver grab a page for a long time should use
> FOLL_DURABLE flag but "for a long time" is very ambiguous. For
> example, there is a driver
>
> get_user_pages()
> some operation.
> put_pages
>
> You can make sure some operation is really fast always?

Well, in our case (judging from the logs) we observed 2 usage patterns
for get_user_pages() calls. One group was lots of short time locks, whose
call stacks originated in various kernel places, the second group was
device drivers which used get_user_pages() to create a buffer for the
DMA. Such buffers were used for the whole lifetime of the session to
the given device, what was equivalent to infinity from the migration/CMA
point of view. This was however based on the specific use case at out
target system, that's why I wanted to start the discussion and find
some generic approach.


> For example, what if it depends on other event which is normally very
> fast but quite slow once a week or try to do dynamic memory allocation
> but memory pressure is severe?
>
> For 100% working well, at last we need to change all GUP user with
> GUP_NM or your FOLL_DURABLE whatever but the concern Mel pointed out
> is it could cause lowmem exhaustion problem.

This way we sooner or later end up without any movable pages at all.
I assume that keeping some temporary references on movable/cma pages
must be allowed, because otherwise we limit the functionality too much.

> At the moment, there is other problem against migratoin, which are not
> related with your patch. ex, zcache, zram, zswap. Their pages couldn't
> be migrated out so I think below Mel's suggestion or some generic
> infrastructure can move pinned page is more proper way to go.

zcache/zram/zswap (vsmalloc based code) can be also extended to support
migration. It requires some significant amount of work, but it is really
doable.

> "To guarantee CMA can migrate pages pinned by drivers I think you need
> migrate-related callsbacks to unpin, barrier the driver until migration
> completes and repin."

Right, this might improve the migration reliability. Are there any works
being done in this direction?

Best regards
--
Marek Szyprowski
Samsung Poland R&D Center




\
 
 \ /
  Last update: 2013-03-06 12:21    [W:0.173 / U:0.284 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site