A queue MODIFY (e.g. re-enabling a queue after a queue_percentage == 0 disable) must not rebuild the MQD with init_mqd: init_mqd rebuilds the HQD from power-on defaults and zeroes the ring pointers, discarding the state firmware context-saved into the MQD on unmap. The re-mapped engine then comes up idle, drops the packet queued while disabled, and wedges into a GPU reset.
Add an optional update_mqd hook to struct amdgpu_mqd for IPs to patch the MQD in place instead - preserving the saved ring/HQD state and forwarding user_wptr into the MQD write pointer. KFD already handles the same disable/re-enable sequence by patching the MQD in place rather than re-initialising it. Signed-off-by: Jesse Zhang <[email protected]> --- drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h index 63670b41c26c..2f73eceea57d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h @@ -500,6 +500,12 @@ struct amdgpu_mqd { unsigned mqd_size; int (*init_mqd)(struct amdgpu_device *adev, void *mqd, struct amdgpu_mqd_prop *p); + /* Patch an MQD in place for a queue MODIFY: keep the ring/HQD state + * firmware context-saved into it (init_mqd instead rebuilds from + * defaults) and forward @user_wptr into the MQD write pointer. + */ + int (*update_mqd)(struct amdgpu_device *adev, void *mqd, + struct amdgpu_mqd_prop *p, u64 user_wptr); }; /* -- 2.49.0
