[Intel-gfx] [PATCH 7/9] drm/i915: stop using ttm_bo_wait

[Daniel Vetter]

On Wed, 30 Nov 2022 at 14:03, Tvrtko Ursulin
<tvrtko.ursulin at linux.intel.com> wrote:
> On 29/11/2022 18:05, Matthew Auld wrote:
> > On Fri, 25 Nov 2022 at 11:14, Tvrtko Ursulin
> > <tvrtko.ursulin at linux.intel.com> wrote:
> >>
> >>
> >> + Matt
> >>
> >> On 25/11/2022 10:21, Christian König wrote:
> >>> TTM is just wrapping core DMA functionality here, remove the mid-layer.
> >>> No functional change.
> >>>
> >>> Signed-off-by: Christian König <christian.koenig at amd.com>
> >>> ---
> >>>    drivers/gpu/drm/i915/gem/i915_gem_ttm.c | 9 ++++++---
> >>>    1 file changed, 6 insertions(+), 3 deletions(-)
> >>>
> >>> diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
> >>> index 5247d88b3c13..d409a77449a3 100644
> >>> --- a/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
> >>> +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm.c
> >>> @@ -599,13 +599,16 @@ i915_ttm_resource_get_st(struct drm_i915_gem_object *obj,
> >>>    static int i915_ttm_truncate(struct drm_i915_gem_object *obj)
> >>>    {
> >>>        struct ttm_buffer_object *bo = i915_gem_to_ttm(obj);
> >>> -     int err;
> >>> +     long err;
> >>>
> >>>        WARN_ON_ONCE(obj->mm.madv == I915_MADV_WILLNEED);
> >>>
> >>> -     err = ttm_bo_wait(bo, true, false);
> >>> -     if (err)
> >>> +     err = dma_resv_wait_timeout(bo->base.resv, DMA_RESV_USAGE_BOOKKEEP,
> >>> +                                 true, 15 * HZ);
> >>
> >> This 15 second stuck out a bit for me and then on a slightly deeper look
> >> it seems this timeout will "leak" into a few of i915 code paths. If we
> >> look at the difference between the legacy shmem and ttm backend I am not
> >> sure if the legacy one is blocking or not - but if it can block I don't
> >> think it would have an arbitrary timeout like this. Matt your thoughts?
> >
> > Not sure what is meant by leak here, but the legacy shmem must also
> > wait/block when unbinding each VMA, before calling truncate. It's the
>
> By "leak" I meant if 15s timeout propagates into some code paths visible
> from userspace which with a legacy backend instead have an indefinite
> wait. If we have that it's probably not very good to have this
> inconsistency, or to apply an arbitrary timeout to those path to start with.
>
> > same story for the ttm backend, except slightly more complicated in
> > that there might be no currently bound VMA, and yet the GPU could
> > still be accessing the pages due to async unbinds, kernel moves etc,
> > which the wait here (and in i915_ttm_shrink) is meant to protect
> > against. If the wait times out it should just fail gracefully. I guess
> > we could just use MAX_SCHEDULE_TIMEOUT here? Not sure if it really
> > matters though.
>
> Right, depends if it can leak or not to userspace and diverge between
> backends.

Generally lock_timeout() is a design bug. It's either
lock_interruptible (or maybe lock_killable) or try_lock, but
lock_timeout is just duct-tape. I haven't dug in to figure out what
should be here, but it smells fishy.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch