Hi Logan,
On 29.05.2020 21:05, Logan Gunthorpe wrote:
On 2020-05-29 6:45 a.m., Christoph Hellwig wrote:
On Thu, May 28, 2020 at 06:00:44PM -0600, Logan Gunthorpe wrote:
This issue is most likely in the i915 driver and is most likely caused by the driver not respecting the return value of the dma_map_ops::map_sg function. You can see the driver ignoring the return value here: https://protect2.fireeye.com/url?k=ca25a34b-97f7b813-ca242804-0cc47a31c8b4-0...
Previously this didn’t cause issues because the intel map_sg always returned the same number of elements as the input scatter gather list but with the change to this dma-iommu api this is no longer the case. I wasn’t able to track the bug down to a specific line of code unfortunately.
Mark did a big audit into the map_sg API abuse and initially had some i915 patches, but then gave up on them with this comment:
"The biggest TODO is DRM/i915 driver and I don't feel brave enough to fix it fully. The driver creatively uses sg_table->orig_nents to store the size of the allocate scatterlist and ignores the number of the entries returned by dma_map_sg function. In this patchset I only fixed the sg_table objects exported by dmabuf related functions. I hope that I didn't break anything there."
it would be really nice if the i915 maintainers could help with sorting that API abuse out.
I agree completely that the API abuse should be sorted out, but I think that's much larger than just the i915 driver. Pretty much every dma-buf map_dma_buf implementation I looked at ignores the returned nents of sg_attrs. This sucks, but I don't think it's the bug Tom ran into. See:
amdgpu_dma_buf_map armada_gem_prime_map_dma_buf drm_gem_map_dma_buf i915_gem_map_dma_buf tegra_gem_prime_map_dma_buf
So this should probably be addressed by the whole GPU community.
Patches are pending: https://lore.kernel.org/linux-iommu/20200513132114.6046-1-m.szyprowski@samsu...
However, as Robin pointed out, there are other ugly tricks like stopping iterating through the SGL when sg_dma_len() is zero. For example, the AMD driver appears to use drm_prime_sg_to_page_addr_arrays() which does this trick and thus likely isn't buggy (otherwise, I'd expect someone to have complained by now seeing AMD has already switched to IOMMU-DMA.
I'm not sure that this is a trick. Stopping at zero sg_dma_len() was somewhere documented.
As I tried to point out in my previous email, i915 does not do this trick. In fact, it completely ignores sg_dma_len() and is thus completely broken. See i915_scatterlist.h and the __sgt_iter() function. So it doesn't sound to me like Mark's fix would address the issue at all. Per my previous email, I'd guess that it can be fixed simply by adjusting the __sgt_iter() function to do something more sensible. (Better yet, if possible, ditch __sgt_iter() and use the common DRM function that AMD uses).
This would at least allow us to make progress with Tom's IOMMU-DMA patch set and once that gets in, it will be harder for other drivers to make the same mistake.
Best regards