Re: [PATCH V2] drm/amdgpu: limit DMA size to PAGE_SIZE for scatter-gather buffers

From: Robin Murphy
Date: Wed Apr 11 2018 - 11:32:26 EST


On 11/04/18 15:33, Sinan Kaya wrote:
On 4/11/2018 8:03 AM, Robin Murphy wrote:
On 10/04/18 21:59, Sinan Kaya wrote:
Code is expecing to observe the same number of buffers returned
from dma_map_sg() function compared to
sg_alloc_table_from_pages(). This doesn't hold true universally
especially for systems with IOMMU.

So why not fix said code? It's clearly not a real hardware
limitation, and the map_sg() APIs have potentially returned fewer
than nents since forever, so there's really no excuse.

Sure, I'll take a better fix if there is one.


IOMMU driver tries to combine buffers into a single DMA address
as much as it can. The right thing is to tell the DMA layer how
much combining IOMMU can do.

Disagree; this is a dodgy hack, since you'll now end up passing
scatterlists into dma_map_sg() which already violate max_seg_size
to begin with, and I think a conscientious DMA API implementation
would be at rights to fail the mapping for that reason (I know
arm64 happens not to, but that was a deliberate design decision to
make my life easier at the time).

As a short-term fix, at least do something like what i915 does and
constrain the table allocation to the desired segment size as well,
so things remain self-consistent. But still never claim that faking
a hardware constraint as a workaround for a driver shortcoming is
"the right thing to do" ;)

You are asking for something like this from here, right?

https://elixir.bootlin.com/linux/v4.16.1/source/drivers/gpu/drm/i915/i915_gem_dmabuf.c#L58

I just looked for callers of __sg_alloc_table_from_pages() with a limit other than SCATTERLIST_MAX_SEGMENT, and found __i915_gem_userptr_alloc_pages(), which at first glance appears to be doing something vaguely similar to amdgpu_ttm_tt_pin_userptr().

Robin.