On 9/1/26 12:45, Jay Cornwall wrote:
On 9/1/2026 08:04, Russell, Kent wrote:
+ .hqd_gfx_cleaner = kgd_gfx_v9_cleaner
Prefer a name like fault_cleaner. This is not related to a HQD.
will update
+void kgd_gfx_v9_cleaner(struct amdgpu_device *adev)
+{
+ gfx_v9_4_2_shader_cleaner_workarounds(adev);
Similarly, fault_cleaner.
will update
+static void gfx942_workaround_cleaner(struct kfd_node *dev,
uint16_t pasid)
Similarly, fault_cleaner
will update
+ /* Run the GFX cleaner workaround on the faulting device.
+ * Only run once per page fault (evicted == 1 after first evict).
+ */
+ pdd = kfd_get_process_device_data(dev, p);
+ if (pdd) {
+ /* Only call cleaner once per page fault. Run the
workaround
+ * only on the first eviction (evicted == 1 after
evict above).
+ */
+ dqm = pdd->dev->dqm;
+ dqm_lock(dqm);
+ if (pdd->qpd.evicted == 1 && dqm->dev->kfd2kgd-
hqd_gfx_cleaner)
+ dqm->dev->kfd2kgd->hqd_gfx_cleaner(dqm->dev-
adev);
+ dqm_unlock(dqm);
+ }
+
+ /* Restore all process devices after workaround completes */
I'm not sure that we want to do this. Once a process faults there is
no recovery path for its queues. If we restore them then other waves
may cause further faults, leading us back here.
Similarly, if the workqueue contains pipelined fault interrupts we
don't want to do this multiple times. Perhaps check pdd->qpd.evicted
before evicting queues and only run the workaround if (evicted == 0).
The eviction and restoration functions feature built-in concurrency
protection. If an eviction process is already underway, the function
automatically bypasses the operation.
The function utilizes the "evicted" value (which can be updated to
|pdd->qpd.evicted <= 1|) to prevent multiple concurrent executions of
the cleaner function.
