[PATCH 2/5] kernel.h: Add non_block_start/end()

[Jerome Glisse]

On Thu, Aug 15, 2019 at 02:35:57PM -0300, Jason Gunthorpe wrote:
> On Thu, Aug 15, 2019 at 06:25:16PM +0200, Daniel Vetter wrote:
> 
> > I'm not really well versed in the details of our userptr, but both
> > amdgpu and i915 wait for the gpu to complete from
> > invalidate_range_start. Jerome has at least looked a lot at the amdgpu
> > one, so maybe he can explain what exactly it is we're doing ...
> 
> amdgpu is (wrongly) using hmm for something, I can't really tell what
> it is trying to do. The calls to dma_fence under the
> invalidate_range_start do not give me a good feeling.
> 
> However, i915 shows all the signs of trying to follow the registration
> cache model, it even has a nice comment in
> i915_gem_userptr_get_pages() explaining that the races it has don't
> matter because it is a user space bug to change the VA mapping in the
> first place. That just screams registration cache to me.
> 
> So it is fine to run HW that way, but if you do, there is no reason to
> fence inside the invalidate_range end. Just orphan the DMA buffer and
> clean it up & release the page pins when all DMA buffer refs go to
> zero. The next access to that VA should get a new DMA buffer with the
> right mapping.
> 
> In other words the invalidation should be very simple without
> complicated locking, or wait_event's. Look at hfi1 for example.

This would break the today usage model of uptr and it will
break userspace expectation ie if GPU is writting to that
memory and that memory then the userspace want to make sure
that it will see what the GPU write.

Yes i915 is broken in respect to not having a end notifier
and tracking active invalidation for a range but the GUP
side of thing kind of hide this bug and it shrinks the window
for bad to happen to something so small that i doubt anyone
could ever hit it (still a bug thought).

Cheers,
Jérôme