This effectively reverts 0fea2ed61e7f ("drm/amdgpu: Remove call to
reservation_object_test_signaled_rcu before wait"), as the premise of
that commit is wrong. dma_resv_wait_timeout() without a timeout will
not turn into a wait-free dma_resv_test_signaled, but will wait a
jiffy for unsignaled fences, which is not at all what userspace expects
when it calls GEM_WAIT_IDLE with a timeout of 0.
Signed-off-by: Lucas Stach <l.stach@xxxxxxxxxxxxxx>
---
This is most likely the correct kernel-side solution for the unexpected
slowdown worked around with in userspace with this Mesa series:
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32877
---
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
index 1a5df8b94661..75b5d5149911 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
@@ -567,8 +567,13 @@ int amdgpu_gem_wait_idle_ioctl(struct drm_device *dev, void *data,
return -ENOENT;
robj = gem_to_amdgpu_bo(gobj);
- ret = dma_resv_wait_timeout(robj->tbo.base.resv, DMA_RESV_USAGE_READ,
- true, timeout);
+ if (timeout == 0)
+ ret = dma_resv_test_signaled(robj->tbo.base.resv,
+ DMA_RESV_USAGE_READ);
+ else
+ ret = dma_resv_wait_timeout(robj->tbo.base.resv,
+ DMA_RESV_USAGE_READ,
+ true, timeout);
/* ret == 0 means not signaled,
* ret > 0 means signaled
--
2.48.1
