On 29/11/2022 18:05, Matthew Auld wrote:
On Fri, 25 Nov 2022 at 11:14, Tvrtko Ursulin
<tvrtko.ursulin@xxxxxxxxxxxxxxx> 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@xxxxxxx>
---
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.
Regards,
Tvrtko
