On 18/10/2021 04:42, Oak Zeng wrote:
Caller of function igt_gpu_fill_dw assumes when this function returns, the gpu command to fill dword is already finished. Caller usually checks the GPU filling result immediately, see the usage in function __igt_write_huge. If there is no wait, when function igt_gpu_fill_dw returns, the GPU command is not guaranteed to be finished. So wait for request to finish.
There is more than one caller and the first one I looked at already does wait so the above is not fully accurate.
I your example of __igt_write_huge you have i915_gem_object_prepare_read called before the assert which does object wait, so that too is not a missing wait.
I haven't checked the rest. So perhaps say what failure you are trying to fix with a link to a CI report or something?
Signed-off-by: Oak Zeng <oak.zeng@xxxxxxxxx> --- drivers/gpu/drm/i915/gem/selftests/igt_gem_utils.c | 1 + 1 file changed, 1 insertion(+) diff --git a/drivers/gpu/drm/i915/gem/selftests/igt_gem_utils.c b/drivers/gpu/drm/i915/gem/selftests/igt_gem_utils.c index b35c1219c852..5fa595ed30b1 100644 --- a/drivers/gpu/drm/i915/gem/selftests/igt_gem_utils.c +++ b/drivers/gpu/drm/i915/gem/selftests/igt_gem_utils.c @@ -157,6 +157,7 @@ int igt_gpu_fill_dw(struct intel_context *ce, if (err) i915_request_set_error_once(rq, err); i915_request_add(rq); + err = i915_request_wait(rq, 0, HZ/5);
For you reference this exploded in CI because as soon as i915_request_add() you are giving your rq reference to the backend and i915_request_wait can/will therefore ooops. If you used this pattern somewhere you'd need to do:
i915_request_get() i915_request_add() i915_request_wait() i915_request_put() Regards, Tvrtko
err_batch: i915_vma_unpin_and_release(&batch, 0); return err;
