Re: CONFIG_DMA_CMA causes ttm performance problems/hangs.

[Jerome Glisse]

 On Mon, Aug 11, 2014 at 12:11:21PM +0200, Thomas Hellstrom wrote:
> On 08/10/2014 08:02 PM, Mario Kleiner wrote:
> > On 08/10/2014 01:03 PM, Thomas Hellstrom wrote:
> >> On 08/10/2014 05:11 AM, Mario Kleiner wrote:
> >>> Resent this time without HTML formatting which lkml doesn't like.
> >>> Sorry.
> >>>
> >>> On 08/09/2014 03:58 PM, Thomas Hellstrom wrote:
> >>>> On 08/09/2014 03:33 PM, Konrad Rzeszutek Wilk wrote:
> >>>>> On August 9, 2014 1:39:39 AM EDT, Thomas
> >>>>> Hellstrom<thellstrom@xxxxxxxxxx>  wrote:
> >>>>>> Hi.
> >>>>>>
> >>>>> Hey Thomas!
> >>>>>
> >>>>>> IIRC I don't think the TTM DMA pool allocates coherent pages more
> >>>>>> than
> >>>>>> one page at a time, and _if that's true_ it's pretty unnecessary for
> >>>>>> the
> >>>>>> dma subsystem to route those allocations to CMA. Maybe Konrad could
> >>>>>> shed
> >>>>>> some light over this?
> >>>>> It should allocate in batches and keep them in the TTM DMA pool for
> >>>>> some time to be reused.
> >>>>>
> >>>>> The pages that it gets are in 4kb granularity though.
> >>>> Then I feel inclined to say this is a DMA subsystem bug. Single page
> >>>> allocations shouldn't get routed to CMA.
> >>>>
> >>>> /Thomas
> >>> Yes, seems you're both right. I read through the code a bit more and
> >>> indeed the TTM DMA pool allocates only one page during each
> >>> dma_alloc_coherent() call, so it doesn't need CMA memory. The current
> >>> allocators don't check for single page CMA allocations and therefore
> >>> try to get it from the CMA area anyway, instead of skipping to the
> >>> much cheaper fallback.
> >>>
> >>> So the callers of dma_alloc_from_contiguous() could need that little
> >>> optimization of skipping it if only one page is requested. For
> >>>
> >>> dma_generic_alloc_coherent
> >>> <https://urldefense.proofpoint.com/v1/url?u=http://lxr.free-electrons.com/ident?i%3Ddma_generic_alloc_coherent&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=QQSN6uVpEiw6RuWLAfK%2FKWBFV5HspJUfDh4Y2mUz%2FH4%3D%0A&s=d1852625e2ab2ff07eb34a7f33fc1f55f7f13959912d5a6ce9316d23070ce939>
> >>>
> >>> andintel_alloc_coherent
> >>> <https://urldefense.proofpoint.com/v1/url?u=http://lxr.free-electrons.com/ident?i%3Dintel_alloc_coherent&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=QQSN6uVpEiw6RuWLAfK%2FKWBFV5HspJUfDh4Y2mUz%2FH4%3D%0A&s=82d587e9b6aeced5cf9a7caefa91bf47fba809f3522b7379d22e45a2d5d35ebd> 
> >>> this
> >>> seems easy to do. Looking at the arm arch variants, e.g.,
> >>>
> >>> https://urldefense.proofpoint.com/v1/url?u=http://lxr.free-electrons.com/source/arch/arm/mm/dma-mapping.c%23L1194&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=QQSN6uVpEiw6RuWLAfK%2FKWBFV5HspJUfDh4Y2mUz%2FH4%3D%0A&s=4c178257eab9b5d7ca650dedba76cf27abeb49ddc7aebb9433f52b6c8bb3bbac
> >>>
> >>>
> >>> and
> >>>
> >>> https://urldefense.proofpoint.com/v1/url?u=http://lxr.free-electrons.com/source/arch/arm64/mm/dma-mapping.c%23L44&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=QQSN6uVpEiw6RuWLAfK%2FKWBFV5HspJUfDh4Y2mUz%2FH4%3D%0A&s=5f62f4cbe8cee1f1dd4cbba656354efe6867bcdc664cf90e9719e2f42a85de08
> >>>
> >>>
> >>> i'm not sure if it is that easily done, as there aren't any fallbacks
> >>> for such a case and the code looks to me as if that's at least
> >>> somewhat intentional.
> >>>
> >>> As far as TTM goes, one quick one-line fix to prevent it from using
> >>> the CMA at least on SWIOTLB, NOMMU and Intel IOMMU (when using the
> >>> above methods) would be to clear the __GFP_WAIT
> >>> <https://urldefense.proofpoint.com/v1/url?u=http://lxr.free-electrons.com/ident?i%3D__GFP_WAIT&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=QQSN6uVpEiw6RuWLAfK%2FKWBFV5HspJUfDh4Y2mUz%2FH4%3D%0A&s=d56d076770d3416264be6c9ea2829ac0d6951203696fa3ad04144f13307577bc>
> >>> flag from the
> >>> passed gfp_t flags. That would trigger the well working fallback.
> >>> So, is
> >>>
> >>> __GFP_WAIT 
> >>> <https://urldefense.proofpoint.com/v1/url?u=http://lxr.free-electrons.com/ident?i%3D__GFP_WAIT&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=QQSN6uVpEiw6RuWLAfK%2FKWBFV5HspJUfDh4Y2mUz%2FH4%3D%0A&s=d56d076770d3416264be6c9ea2829ac0d6951203696fa3ad04144f13307577bc> 
> >>> needed
> >>> for those single page allocations that go through__ttm_dma_alloc_page
> >>> <https://urldefense.proofpoint.com/v1/url?u=http://lxr.free-electrons.com/ident?i%3D__ttm_dma_alloc_page&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=QQSN6uVpEiw6RuWLAfK%2FKWBFV5HspJUfDh4Y2mUz%2FH4%3D%0A&s=7898522bba274e4dcc332735fbcf0c96e48918f60c2ee8e9a3e9c73ab3487bd0>?
> >>>
> >>>
> >>> It would be nice to have such a simple, non-intrusive one-line patch
> >>> that we still could get into 3.17 and then backported to older stable
> >>> kernels to avoid the same desktop hangs there if CMA is enabled. It
> >>> would be also nice for actual users of CMA to not use up lots of CMA
> >>> space for gpu's which don't need it. I think DMA_CMA was introduced
> >>> around 3.12.
> >>>
> >> I don't think that's a good idea. Omitting __GFP_WAIT would cause
> >> unnecessary memory allocation errors on systems under stress.
> >> I think this should be filed as a DMA subsystem kernel bug / regression
> >> and an appropriate solution should be worked out together with the DMA
> >> subsystem maintainers and then backported.
> >
> > Ok, so it is needed. I'll file a bug report.
> >
> >>> The other problem is that probably TTM does not reuse pages from the
> >>> DMA pool. If i trace the __ttm_dma_alloc_page
> >>> <https://urldefense.proofpoint.com/v1/url?u=http://lxr.free-electrons.com/ident?i%3D__ttm_dma_alloc_page&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=QQSN6uVpEiw6RuWLAfK%2FKWBFV5HspJUfDh4Y2mUz%2FH4%3D%0A&s=7898522bba274e4dcc332735fbcf0c96e48918f60c2ee8e9a3e9c73ab3487bd0>
> >>> and
> >>> __ttm_dma_free_page
> >>> <https://urldefense.proofpoint.com/v1/url?u=http://lxr.free-electrons.com/ident?i%3D__ttm_dma_alloc_page&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=QQSN6uVpEiw6RuWLAfK%2FKWBFV5HspJUfDh4Y2mUz%2FH4%3D%0A&s=7898522bba274e4dcc332735fbcf0c96e48918f60c2ee8e9a3e9c73ab3487bd0>
> >>> calls for
> >>> those single page allocs/frees, then over a 20 second interval of
> >>> tracing and switching tabs in firefox, scrolling things around etc. i
> >>> find about as many alloc's as i find free's, e.g., 1607 allocs vs.
> >>> 1648 frees.
> >> This is because historically the pools have been designed to keep only
> >> pages with nonstandard caching attributes since changing page caching
> >> attributes have been very slow but the kernel page allocators have been
> >> reasonably fast.
> >>
> >> /Thomas
> >
> > Ok. A bit more ftraceing showed my hang problem case goes through the
> > "if (is_cached)" paths, so the pool doesn't recycle anything and i see
> > it bouncing up and down by 4 pages all the time.
> >
> > But for the non-cached case, which i don't hit with my problem, could
> > one of you look at line 954...
> >
> > https://urldefense.proofpoint.com/v1/url?u=http://lxr.free-electrons.com/source/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c%23L954&k=oIvRg1%2BdGAgOoM1BIlLLqw%3D%3D%0A&r=l5Ago9ekmVFZ3c4M6eauqrJWGwjf6fTb%2BP3CxbBFkVM%3D%0A&m=QQSN6uVpEiw6RuWLAfK%2FKWBFV5HspJUfDh4Y2mUz%2FH4%3D%0A&s=e15c51805d429ee6d8960d6b88035e9811a1cdbfbf13168eec2fbb2214b99c60
> >
> >
> > ... and tell me why that unconditional npages = count; assignment
> > makes sense? It seems to essentially disable all recycling for the dma
> > pool whenever the pool isn't filled up to/beyond its maximum with free
> > pages? When the pool is filled up, lots of stuff is recycled, but when
> > it is already somewhat below capacity, it gets "punished" by not
> > getting refilled? I'd just like to understand the logic behind that line.
> >
> > thanks,
> > -mario
> 
> I'll happily forward that question to Konrad who wrote the code (or it
> may even stem from the ordinary page pool code which IIRC has Dave
> Airlie / Jerome Glisse as authors)

This is effectively bogus code, i now wonder how it came to stay alive.
Attached patch will fix that.


> 
> /Thomas
>