On 5/17/21 11:46 PM, Thomas Hellström wrote:
On 5/17/21 3:11 PM, Christoph Hellwig wrote:
On Mon, May 17, 2021 at 04:09:42PM +0300, Serge Belyshev wrote:
Christoph Hellwig hch@lst.de writes:
As an ad-hoc experiment: can you replace the call to remap_pfn_range with remap_pfn_range_notrack (and export it if you build i915 modular) in remap_io_sg and see if that makes any difference?
That worked, thanks -- no artifacts seen.
Looks like it is caused by the validation failure then. Which means the existing code is doing something wrong in its choice of the page protection bit. I really need help from the i915 maintainers here..
Hmm,
Apart from the caching aliasing Mattew brought up, doesn't the remap_pfn_range_xxx() family require the mmap_sem held in write mode since it modifies the vma structure? remap_io_sg() is called from the fault handler with the mmap_sem held in read mode only.
/Thomas
And worse, if we prefault a user-space buffer object map using remap_io_sg() and then zap some ptes using madvise(), the next time those ptes are accessed, we'd trigger a new call to remap_io_sg() which would now find already populated ptes. While the old code looks to just silently overwrite those, it looks like the new code would BUG in remap_pte_range()?
/Thomas
Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx