Skip to contentSkip to conversation

drm/amdkfd: Block per-queue reset when halt_if_hws_hang=1

PATCH In review Jay Cornwall 2 participants 2 messages Jan 16, 2025

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/amdkfd: Block per-queue reset when halt_if_hws_hang=1

The purpose of halt_if_hws_hang is to preserve GPU state for driver debugging when queue preemption fails. Issuing per-queue reset may kill wavefronts which caused the preemption failure.

---
 drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -2
index f157494bfdb1..195085079eb2 100644
@@ -2327,9 +2327,9 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm,
 	 */
 	mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
 	if (mqd_mgr->check_preemption_failed(mqd_mgr, dqm->packet_mgr.priv_queue->queue->mqd)) {
+		while (halt_if_hws_hang)
+			schedule();
 		if (reset_queues_on_hws_hang(dqm)) {
-			while (halt_if_hws_hang)
-				schedule();
 			dqm->is_hws_hang = true;
 			kfd_hws_hang(dqm);
 			retval = -ETIME;
Signature · 1 line
2.34.1
Reviewed-by on v1 1/1
#1 lore raw

[Public]

Quoted 8 lines

-----Original Message-----
From: Cornwall, Jay <[email protected]>
Sent: Thursday, January 16, 2025 3:41 PM
To: [email protected]
Cc: Cornwall, Jay <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH] drm/amdkfd: Block per-queue reset when halt_if_hws_hang=1

Quoted 6 lines

The purpose of halt_if_hws_hang is to preserve GPU state for driver debugging when queue preemption fails. Issuing per-queue reset may kill wavefronts which caused the preemption failure.

Quoted 25 lines
---
 drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -2
index f157494bfdb1..195085079eb2 100644
@@ -2327,9 +2327,9 @@ static int unmap_queues_cpsch(struct
 device_queue_manager *dqm,
       */
      mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
      if (mqd_mgr->check_preemption_failed(mqd_mgr, dqm-
 >packet_mgr.priv_queue->queue->mqd)) {
+             while (halt_if_hws_hang)
+                     schedule();
              if (reset_queues_on_hws_hang(dqm)) {
-                     while (halt_if_hws_hang)
-                             schedule();
                      dqm->is_hws_hang = true;
                      kfd_hws_hang(dqm);
                      retval = -ETIME;

--
2.34.1