Re: [PATCH 2/2] drm/vgem: use normal cached mmap'ings

[Chris Wilson]

Quoting Rob Clark (2019-07-16 17:42:15)
> From: Rob Clark <robdclark@xxxxxxxxxxxx>
> 
> Since there is no real device associated with vgem, it is impossible to
> end up with appropriate dev->dma_ops, meaning that we have no way to
> invalidate the shmem pages allocated by vgem.  So, at least on platforms
> without drm_cflush_pages(), we end up with corruption when cache lines
> from previous usage of vgem bo pages get evicted to memory.
> 
> The only sane option is to use cached mappings.
> 
> Signed-off-by: Rob Clark <robdclark@xxxxxxxxxxxx>
> ---
> Possibly we could dma_sync_*_for_{device,cpu}() on dmabuf attach/detach,
> although the ->gem_prime_{pin,unpin}() API isn't quite ideal for that as
> it is.  And that doesn't really help for drivers that don't attach/
> detach for each use.
> 
> But AFAICT vgem is mainly used for dmabuf testing, so maybe we don't
> need to care too much about use of cached mmap'ings.

Sadly this regresses with i915 interop.

Starting subtest: 4KiB-tiny-vgem-blt-early-read-child
(gem_concurrent_blit:8309) CRITICAL: Test assertion failure function dmabuf_cmp_bo, file ../tests/i915/gem_concurrent_all.c:408:
(gem_concurrent_blit:8309) CRITICAL: Failed assertion: v[((y)*(b->width) + (((y) + pass)%(b->width)))] == val
(gem_concurrent_blit:8309) CRITICAL: error: 0 != 0xdeadbeef

and igt/prime_vgem

Can you please cc intel-gfx so CI can pick up these changes?
-Chris