[PATCH] dma-buf: heaps: Don't track CMA dma-buf pages under RssFile

Daniel Vetter daniel at ffwll.ch
Thu Jan 18 20:57:52 UTC 2024


On Thu, Jan 18, 2024 at 08:57:16AM -0800, T.J. Mercier wrote:
> On Thu, Jan 18, 2024 at 6:49 AM Daniel Vetter <daniel at ffwll.ch> wrote:
> >
> > On Thu, Jan 18, 2024 at 11:02:22AM +0100, Christian König wrote:
> > > Am 17.01.24 um 19:11 schrieb T.J. Mercier:
> > > > DMA buffers allocated from the CMA dma-buf heap get counted under
> > > > RssFile for processes that map them and trigger page faults. In
> > > > addition to the incorrect accounting reported to userspace, reclaim
> > > > behavior was influenced by the MM_FILEPAGES counter until linux 6.8, but
> > > > this memory is not reclaimable. [1] Change the CMA dma-buf heap to set
> > > > VM_PFNMAP on the VMA so MM does not poke at the memory managed by this
> > > > dma-buf heap, and use vmf_insert_pfn to correct the RSS accounting.
> > > >
> > > > The system dma-buf heap does not suffer from this issue since
> > > > remap_pfn_range is used during the mmap of the buffer, which also sets
> > > > VM_PFNMAP on the VMA.
> > >
> > > Mhm, not an issue with this patch but Daniel wanted to add a check for
> > > VM_PFNMAP to dma_buf_mmap() which would have noted this earlier.
> > >
> > > I don't fully remember the discussion but for some reason that was never
> > > committed. We should probably try that again.
> >
> > Iirc the issue is that dma_mmap is not guaranteed to give you a VM_SPECIAL
> > mapping, at least on absolutely all architectures. That's why I defacto
> > dropped that idea, but it would indeed be really great if we could
> > resurrect it.
> 
> I actually had it in my head that it was a BUG_ON check for VM_PFNMAP
> in dma_buf_mmap and it was merged, so I was surprised to discover that
> it wasn't set for these CMA buffers.
> 
> > Maybe for x86 only? Or x86+armv8, I'm honestly not sure anymore which
> > exact cases ended up with a VM_NORMAL mapping ... Would need a pile of
> > digging.
> 
> Looking back at the patch, the CI email at the end of the thread lists
> a bunch of now-broken links to DMESG-WARN test failures I assume
> pointed at a large chunk of them.
> 
> https://lore.kernel.org/all/166919750173.15575.2864736980735346730@emeril.freedesktop.org/

I thought there was a more recent submission, where I at least fixed the
various fallout in gem code. But maybe I only dreamed ...

Also I did the code grepping again, and at least iommu_dma_mmap() in
drivers/iommu/dma-iommu.c and arm_iommu_mmap_attrs() for arm use
vm_map_pages in certain cases, which is _not_ VM_PFNMAP.

Means really no cases where I think we can assume we'll always get
VM_PFNMAP, and unfortunately we need VM_PFNMAP or VM_IO to prevent
get_user_pages and similar bad things from happening to dma-buf mmaps.

So still no luck :-/
-Sima


> 
> > >
> > > > [1]https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/mm/vmscan.c?id=fb46e22a9e3863e08aef8815df9f17d0f4b9aede
> > > >
> > > > Fixes: b61614ec318a ("dma-buf: heaps: Add CMA heap to dmabuf heaps")
> > > > Signed-off-by: T.J. Mercier<tjmercier at google.com>
> > >
> > > Acked-by: Christian König <christian.koenig at amd.com>
> 
> Thanks Christian.

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch


More information about the dri-devel mailing list