[RFC] arm64: swiotlb: cma_alloc error spew

Commit Message

dann frazier April 17, 2019, 8:48 p.m. UTC

hey,
  I'm seeing an issue on a couple of arm64 systems[*] where they spew
~10K "cma: cma_alloc: alloc failed" messages at boot. The errors are
non-fatal, and bumping up cma to a large enough size (~128M) gets rid
of them - but that seems suboptimal. Bisection shows that this started
after commit fafadcd16595 ("swiotlb: don't dip into swiotlb pool for
coherent allocations"). It looks like __dma_direct_alloc_pages()
is opportunistically using CMA memory but falls back to non-CMA if CMA
disabled or unavailable. I've demonstrated that this fallback is
indeed returning a valid pointer. So perhaps the issue is really just
the warning emission.

The following naive patch solves the problem for me - just silence the
cma errors, since it looks like a soft error. But is there a better
approach?

[*] APM X-Gene & HiSilicon Hi1620 w/ SMMU disabled

Comments

Robin Murphy April 23, 2019, 11:32 a.m. UTC | #1

On 17/04/2019 21:48, dann frazier wrote:
> hey,
>    I'm seeing an issue on a couple of arm64 systems[*] where they spew
> ~10K "cma: cma_alloc: alloc failed" messages at boot. The errors are
> non-fatal, and bumping up cma to a large enough size (~128M) gets rid
> of them - but that seems suboptimal. Bisection shows that this started
> after commit fafadcd16595 ("swiotlb: don't dip into swiotlb pool for
> coherent allocations"). It looks like __dma_direct_alloc_pages()
> is opportunistically using CMA memory but falls back to non-CMA if CMA
> disabled or unavailable. I've demonstrated that this fallback is
> indeed returning a valid pointer. So perhaps the issue is really just
> the warning emission.

The CMA area being full isn't necessarily an ignorable non-problem, 
since it means you won't be able to allocate the kind of large buffers 
for which CMA was intended. The question is, is it actually filling up 
with allocations that deserve to be there, or is this the same as I've 
seen on a log from a ThunderX2 system where it's getting exhausted by 
thousands upon thousands of trivial single page allocations? If it's the 
latter (CONFIG_CMA_DEBUG should help shed some light if necessary), then 
that does lean towards spending a bit more effort on this idea:

https://lore.kernel.org/lkml/[e-mail]/

Robin.

> The following naive patch solves the problem for me - just silence the
> cma errors, since it looks like a soft error. But is there a better
> approach?
> 
> [*] APM X-Gene & HiSilicon Hi1620 w/ SMMU disabled
> 
> diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
> index 6310ad01f915b..0324aa606c173 100644
> --- a/kernel/dma/direct.c
> +++ b/kernel/dma/direct.c
> @@ -112,7 +112,7 @@ struct page *__dma_direct_alloc_pages(struct device *dev, size_t size,
>          /* CMA can be used only in the context which permits sleeping */
>          if (gfpflags_allow_blocking(gfp)) {
>                  page = dma_alloc_from_contiguous(dev, count, page_order,
> -                                                gfp & __GFP_NOWARN);
> +                                                true);
>                  if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
>                          dma_release_from_contiguous(dev, page, count);
>                          page = NULL;
> 
> 
> 
>

dann frazier April 23, 2019, 6:03 p.m. UTC | #2

On Tue, Apr 23, 2019 at 5:32 AM Robin Murphy <[e-mail]> wrote:
>
> On 17/04/2019 21:48, dann frazier wrote:
> > hey,
> >    I'm seeing an issue on a couple of arm64 systems[*] where they spew
> > ~10K "cma: cma_alloc: alloc failed" messages at boot. The errors are
> > non-fatal, and bumping up cma to a large enough size (~128M) gets rid
> > of them - but that seems suboptimal. Bisection shows that this started
> > after commit fafadcd16595 ("swiotlb: don't dip into swiotlb pool for
> > coherent allocations"). It looks like __dma_direct_alloc_pages()
> > is opportunistically using CMA memory but falls back to non-CMA if CMA
> > disabled or unavailable. I've demonstrated that this fallback is
> > indeed returning a valid pointer. So perhaps the issue is really just
> > the warning emission.
>
> The CMA area being full isn't necessarily an ignorable non-problem,
> since it means you won't be able to allocate the kind of large buffers
> for which CMA was intended. The question is, is it actually filling up
> with allocations that deserve to be there, or is this the same as I've
> seen on a log from a ThunderX2 system where it's getting exhausted by
> thousands upon thousands of trivial single page allocations? If it's the
> latter (CONFIG_CMA_DEBUG should help shed some light if necessary),

Appears so. Here's a histogram of count/size w/ a cma= large enough to
avoid failures:

$ dmesg | grep "cma: cma_alloc(cma" | sed -r 's/.*count
([0-9]+)\,.*/\1/' | sort -n | uniq -c
   2062 1
     32 2
    266 8
      2 24
      4 32
    256 33
      7 64
      2 128
      2 1024

  -dann

> then
> that does lean towards spending a bit more effort on this idea:
>
> https://lore.kernel.org/lkml/[e-mail]/
>
> Robin.
>
> > The following naive patch solves the problem for me - just silence the
> > cma errors, since it looks like a soft error. But is there a better
> > approach?
> >
> > [*] APM X-Gene & HiSilicon Hi1620 w/ SMMU disabled
> >
> > diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
> > index 6310ad01f915b..0324aa606c173 100644
> > --- a/kernel/dma/direct.c
> > +++ b/kernel/dma/direct.c
> > @@ -112,7 +112,7 @@ struct page *__dma_direct_alloc_pages(struct device *dev, size_t size,
> >          /* CMA can be used only in the context which permits sleeping */
> >          if (gfpflags_allow_blocking(gfp)) {
> >                  page = dma_alloc_from_contiguous(dev, count, page_order,
> > -                                                gfp & __GFP_NOWARN);
> > +                                                true);
> >                  if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
> >                          dma_release_from_contiguous(dev, page, count);
> >                          page = NULL;
> >
> >
> >
> >

dann frazier April 24, 2019, 12:39 a.m. UTC | #3

On Tue, Apr 23, 2019 at 12:03 PM dann frazier
<[e-mail]> wrote:
>
> On Tue, Apr 23, 2019 at 5:32 AM Robin Murphy <[e-mail]> wrote:
> >
> > On 17/04/2019 21:48, dann frazier wrote:
> > > hey,
> > >    I'm seeing an issue on a couple of arm64 systems[*] where they spew
> > > ~10K "cma: cma_alloc: alloc failed" messages at boot. The errors are
> > > non-fatal, and bumping up cma to a large enough size (~128M) gets rid
> > > of them - but that seems suboptimal. Bisection shows that this started
> > > after commit fafadcd16595 ("swiotlb: don't dip into swiotlb pool for
> > > coherent allocations"). It looks like __dma_direct_alloc_pages()
> > > is opportunistically using CMA memory but falls back to non-CMA if CMA
> > > disabled or unavailable. I've demonstrated that this fallback is
> > > indeed returning a valid pointer. So perhaps the issue is really just
> > > the warning emission.
> >
> > The CMA area being full isn't necessarily an ignorable non-problem,
> > since it means you won't be able to allocate the kind of large buffers
> > for which CMA was intended. The question is, is it actually filling up
> > with allocations that deserve to be there, or is this the same as I've
> > seen on a log from a ThunderX2 system where it's getting exhausted by
> > thousands upon thousands of trivial single page allocations? If it's the
> > latter (CONFIG_CMA_DEBUG should help shed some light if necessary),
>
> Appears so. Here's a histogram of count/size w/ a cma= large enough to
> avoid failures:
>
> $ dmesg | grep "cma: cma_alloc(cma" | sed -r 's/.*count
> ([0-9]+)\,.*/\1/' | sort -n | uniq -c
>    2062 1
>      32 2
>     266 8
>       2 24
>       4 32
>     256 33

And IIUC, this is also a big culprit. The debugfs bitmap seems to show
that the alignment of each of these leaves 31 pages unused, which adds
up to 31MB!

  -dann

>       7 64
>       2 128
>       2 1024
>
>   -dann
>
> > then
> > that does lean towards spending a bit more effort on this idea:
> >
> > https://lore.kernel.org/lkml/[e-mail]/
> >
> > Robin.
> >
> > > The following naive patch solves the problem for me - just silence the
> > > cma errors, since it looks like a soft error. But is there a better
> > > approach?
> > >
> > > [*] APM X-Gene & HiSilicon Hi1620 w/ SMMU disabled
> > >
> > > diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
> > > index 6310ad01f915b..0324aa606c173 100644
> > > --- a/kernel/dma/direct.c
> > > +++ b/kernel/dma/direct.c
> > > @@ -112,7 +112,7 @@ struct page *__dma_direct_alloc_pages(struct device *dev, size_t size,
> > >          /* CMA can be used only in the context which permits sleeping */
> > >          if (gfpflags_allow_blocking(gfp)) {
> > >                  page = dma_alloc_from_contiguous(dev, count, page_order,
> > > -                                                gfp & __GFP_NOWARN);
> > > +                                                true);
> > >                  if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
> > >                          dma_release_from_contiguous(dev, page, count);
> > >                          page = NULL;
> > >
> > >
> > >
> > >

Patch

diff --git a/kernel/dma/direct.c b/kernel/dma/direct.c
index 6310ad01f915b..0324aa606c173 100644
--- a/kernel/dma/direct.c
+++ b/kernel/dma/direct.c
@@ -112,7 +112,7 @@  struct page *__dma_direct_alloc_pages(struct device *dev, size_t size,
        /* CMA can be used only in the context which permits sleeping */
        if (gfpflags_allow_blocking(gfp)) {
                page = dma_alloc_from_contiguous(dev, count, page_order,
-                                                gfp & __GFP_NOWARN);
+                                                true);
                if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) {
                        dma_release_from_contiguous(dev, page, count);
                        page = NULL;