[Intel-gfx] [PATCH] dma-buf: Precheck for a valid dma_fence before acquiring the reference

[Chris Wilson]

Quoting Daniel Vetter (2020-02-21 15:17:24)
> On Fri, Feb 21, 2020 at 3:38 PM Chris Wilson <chris at chris-wilson.co.uk> wrote:
> > dma_fence_get_rcu() is used to acquire a reference to under a dma-fence
> > under racey conditions -- a perfect recipe for a disaster. As we know
> > the caller may be handling stale memory, use kasan to confirm the
> > dma-fence, or rather its memory block, is valid before attempting to
> > acquire a reference. This should help us to more quickly and clearly
> > identify lost races.
> 
> Hm ... I'm a bit lost on the purpose, and what this does. Fences need
> to be rcu-freed, and I have honestly no idea how kasan treats those.
> Are we throwing false positives, because kasan thinks the stuff is
> freed, but we're still accessing it (while the grace period hasn't
> passed, so anything freed is still guaranteed to be at least in the
> slab cache somewhere).
> 
> I'm not seeing how this catches lost races quicker, since the refcount
> should get to 0 way before we get to the kfree. So the refcount check
> on the next line should catch strictly more races than the kasan
> check.

It's not about the fence itself, but those pointing to the fence. That's
where we may find garbage, and by returning NULL the kernel keeps
working for a bit longer as you try to piece together the race.
-Chris