[Linaro-mm-sig] [PATCHv16 0/9] Contiguous Memory Allocator

[Marek Szyprowski]

Hello,

On Tuesday, October 11, 2011 9:30 AM Maxime Coquelin wrote:

> On 10/11/2011 09:17 AM, Marek Szyprowski wrote:
> > On Monday, October 10, 2011 2:08 PM Maxime Coquelin wrote:
> >
> >       During our stress tests, we encountered some problems :
> >
> >       1) Contiguous allocation lockup:
> >           When system RAM is full of Anon pages, if we try to allocate a
> > contiguous buffer greater than the min_free value, we face a
> > dma_alloc_from_contiguous lockup.
> >           The expected result would be dma_alloc_from_contiguous() to fail.
> >           The problem is reproduced systematically on our side.
> > Thanks for the report. Do you use Android's lowmemorykiller? I haven't
> > tested CMA on Android kernel yet. I have no idea how it will interfere
> > with Android patches.
> >
> 
> The software used for this test (v16) is a generic 3.0 Kernel and a
> minimal filesystem using Busybox.

I'm really surprised. Could you elaborate a bit how to trigger this issue?
I've did several tests and I never get a lockup. Allocation failed from time
to time though.

> With v15 patchset, I also tested it with Android.
> IIRC, sometimes the lowmemorykiller succeed to get free space and the
> contiguous allocation succeed, sometimes we faced  the lockup.
> 
> >>       2) Contiguous allocation fail:
> >>           We have developed a small driver and a shell script to
> >> allocate/release contiguous buffers.
> >>           Sometimes, dma_alloc_from_contiguous() fails to allocate the
> >> contiguous buffer (about once every 30 runs).
> >>           We have 270MB Memory passed to the kernel in our configuration,
> >> and the CMA pool is 90MB large.
> >>           In this setup, the overall memory is either free or full of
> >> reclaimable pages.
> > Yeah. We also did such stress tests recently and faced this issue. I've
> > spent some time investigating it but I have no solution yet.
> >
> > The problem is caused by a page, which is put in the CMA area. This page
> > is movable, but it's address space provides no 'migratepage' method. In
> > such case mm subsystem uses fallback_migrate_page() function. Sadly this
> > function only returns -EAGAIN. The migration loops a few times over it
> > and fails causing the fail in the allocation procedure.
> >
> > We are investing now which kernel code created/allocated such problematic

s/investing/investigating

> > pages and how to add real migration support for them.
> >
> 
> Ok, thanks for pointing this out.

We found this issue very recently. I'm still surprised that we did not notice 
it during system testing.

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