Skip to contentSkip to conversation

drm/amdgpu/userq: properly account for resets

PATCH In review Alex Deucher 6 participants 17 messages Jul 13 – Jul 15

Subsystems (2)
lore
Share & tools
lore

Alex Deucher

author maintainer
#0 lore raw

PATCH 1/1 drm/amdgpu/userq: properly account for resets

We need to increment the reset counter, force fence completion, and set the wedged event when a user queue is reset.

mes_userq_reset_queue() handles this for collateral damage, but the caller needs to handle this directly for the original guilty queue.

Cc Srinivasan Shanmugam <[email protected]>
Cc Prike Liang <[email protected]>
Cc Sunil Khatri <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +6 -1
index 6aa75da27f912..5e1262636e1e9 100644
@@ -146,8 +146,13 @@ static void amdgpu_userq_hang_detect_work(struct work_struct *work)
 							 queue, NULL, NULL);
 		else
 			r = userq_funcs->reset(queue);
-		if (r)
+		if (r) {
 			gpu_reset = true;
+		} else {
+			atomic_inc(&adev->gpu_reset_counter);
+			amdgpu_userq_fence_driver_force_completion(queue);
+			drm_dev_wedged_event(adev_to_drm(adev), DRM_WEDGE_RECOVERY_NONE, NULL);
+		}
 	} else {
 		gpu_reset = true;
 	}
Signature · 1 line
2.55.0
Reviewed-by on v1 1/1
#1 lore raw

AMD General

Quoted 48 lines

-----Original Message-----
From: amd-gfx <[email protected]> On Behalf Of Alex Deucher
Sent: Tuesday, July 14, 2026 4:39 AM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets

We need to increment the reset counter, force fence completion, and set the wedged event when a user queue is reset.

mes_userq_reset_queue() handles this for collateral damage, but the caller needs to handle this directly for the original guilty queue.

Cc Srinivasan Shanmugam <[email protected]>
Cc Prike Liang <[email protected]>
Cc Sunil Khatri <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +6 -1
index 6aa75da27f912..5e1262636e1e9 100644
@@ -146,8 +146,13 @@ static void amdgpu_userq_hang_detect_work(struct
 work_struct *work)
                                                       queue, NULL, NULL);
              else
                      r = userq_funcs->reset(queue);
-             if (r)
+             if (r) {
                      gpu_reset = true;
+             } else {
+                     atomic_inc(&adev->gpu_reset_counter);
+                     amdgpu_userq_fence_driver_force_completion(queue);
+                     drm_dev_wedged_event(adev_to_drm(adev),
 DRM_WEDGE_RECOVERY_NONE, NULL);
+             }
      } else {
              gpu_reset = true;
      }

--
2.55.0

#2 lore raw
In reply to Alex Deucher

AMD General

Regards,
Prike

Quoted 31 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 4:39 AM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets

We need to increment the reset counter, force fence completion, and set the wedged event when a user queue is reset.

mes_userq_reset_queue() handles this for collateral damage, but the caller needs to handle this directly for the original guilty queue.

Cc Srinivasan Shanmugam <[email protected]>
Cc Prike Liang <[email protected]>
Cc Sunil Khatri <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +0 -0
index 6aa75da27f912..5e1262636e1e9 100644
@@ -146,8 +146,13 @@ static void amdgpu_userq_hang_detect_work(struct
 work_struct *work)
                                                       queue, NULL, NULL);

drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c static void amdgpu_userq_hang_detect_work(struct · line 156

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +5 -1
              else
                      r = userq_funcs->reset(queue);
-             if (r)
+             if (r) {
                      gpu_reset = true;
+             } else {
+                     atomic_inc(&adev->gpu_reset_counter);
+                     amdgpu_userq_fence_driver_force_completion(queue);
+                     drm_dev_wedged_event(adev_to_drm(adev),
 DRM_WEDGE_RECOVERY_NONE, NULL);

If we do the userq reset post similar process here just like as mes_userq_reset_queue(), so do we need to clear the duplicate the code in the mes_userq_reset_queue()? Otherwise, the gpu reset counter should be incorrect. Except that, we might input the wedge task info to report the guilty process info.

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +1 -0
+             }
      } else {
              gpu_reset = true;
      }

--
2.55.0

#3 lore raw

AMD General

Regards,
Prike

Quoted 52 lines

-----Original Message-----
From: Liang, Prike
Sent: Tuesday, July 14, 2026 10:49 AM
To: 'Alex Deucher' <[email protected]>; [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Khatri, Sunil <[email protected]>
Subject: RE: [PATCH] drm/amdgpu/userq: properly account for resets

Regards,
Prike

Quoted 38 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 4:39 AM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets

We need to increment the reset counter, force fence completion, and set the wedged event when a user queue is reset.

mes_userq_reset_queue() handles this for collateral damage, but the caller needs to handle this directly for the original guilty queue.

Cc Srinivasan Shanmugam <[email protected]>
Cc Prike Liang <[email protected]>
Cc Sunil Khatri <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +3 -1
index 6aa75da27f912..5e1262636e1e9 100644
@@ -146,8 +146,13 @@ static void amdgpu_userq_hang_detect_work(struct
 work_struct *work)
                                                     queue, NULL, NULL);
            else
                    r = userq_funcs->reset(queue);
-           if (r)
+           if (r) {
                    gpu_reset = true;
+           } else {
+                   atomic_inc(&adev->gpu_reset_counter);
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c static void amdgpu_userq_hang_detect_work(struct · line 156

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +2 -0
+                   amdgpu_userq_fence_driver_force_completion(queue);
+                   drm_dev_wedged_event(adev_to_drm(adev),
 DRM_WEDGE_RECOVERY_NONE, NULL);

If we do the userq reset post similar process here just like as mes_userq_reset_queue(), so do we need to clear the duplicate the code in the mes_userq_reset_queue()? Otherwise, the gpu reset counter should be incorrect. Except that, we might input the wedge task info to report the guilty process info.

As to the first question, I recheck the code and found out the guilty user queue in the amdgpu_gfx_reset_mes_compute() was reset separately before mes_userq_reset_queue(), with this current reset sequence then this change it make sense to update reset counter and force complete fence.

But why need to reset the original guilty queue separately, should this guilty queue was found by the amdgpu_mes_detect_and_reset_hung_queues() in the MES firmware? If so, we may need to reset the guilty queue through mes_userq_reset_queue() rather than reset separately, then we don't need to add this change anymore.

Quoted 7 lines
Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +1 -0
+           }
    } else {
            gpu_reset = true;
    }

--
2.55.0

#4 lore raw
Quoted 67 lines · Prike

AMD General

Regards,
Prike

Quoted 52 lines

-----Original Message-----
From: Liang, Prike
Sent: Tuesday, July 14, 2026 10:49 AM
To: 'Alex Deucher' <[email protected]>; [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Khatri, Sunil <[email protected]>
Subject: RE: [PATCH] drm/amdgpu/userq: properly account for resets

Regards,
Prike

Quoted 38 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 4:39 AM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets

We need to increment the reset counter, force fence completion, and set the wedged event when a user queue is reset.

mes_userq_reset_queue() handles this for collateral damage, but the caller needs to handle this directly for the original guilty queue.

Cc Srinivasan Shanmugam <[email protected]>
Cc Prike Liang <[email protected]>
Cc Sunil Khatri <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +3 -1
index 6aa75da27f912..5e1262636e1e9 100644
@@ -146,8 +146,13 @@ static void amdgpu_userq_hang_detect_work(struct
 work_struct *work)
                                                     queue, NULL, NULL);
            else
                    r = userq_funcs->reset(queue);
-           if (r)
+           if (r) {
                    gpu_reset = true;
+           } else {
+                   atomic_inc(&adev->gpu_reset_counter);
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c static void amdgpu_userq_hang_detect_work(struct · line 156

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +2 -0
+                   amdgpu_userq_fence_driver_force_completion(queue);
+                   drm_dev_wedged_event(adev_to_drm(adev),
 DRM_WEDGE_RECOVERY_NONE, NULL);

If we do the userq reset post similar process here just like as mes_userq_reset_queue(), so do we need to clear the duplicate the code in the mes_userq_reset_queue()? Otherwise, the gpu reset counter should be incorrect. Except that, we might input the wedge task info to report the guilty process info.

Quoted 3 lines

As to the first question, I recheck the code and found out the guilty user queue in the amdgpu_gfx_reset_mes_compute() was reset separately before mes_userq_reset_queue(), with this current reset sequence then this change it make sense to update reset counter and force complete fence.

But why need to reset the original guilty queue separately, should this guilty queue was found by the amdgpu_mes_detect_and_reset_hung_queues() in the MES firmware? If so, we may need to reset the guilty queue through mes_userq_reset_queue() rather than reset separately, then we don't need to add this change anymore.

We still need it for gfx and sdma queues. Those don't get reset via amdgpu_gfx_reset_mes_compute(),

Alex

Quoted 9 lines
Quoted 8 lines
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +1 -0
+           }
    } else {
            gpu_reset = true;
    }

--
2.55.0

#5 lore raw

AMD General

Regards,
Prike

Quoted 85 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 9:52 PM
To: Liang, Prike <[email protected]>
Cc: Deucher, Alexander <[email protected]>; amd- [email protected]; SHANMUGAM, SRINIVASAN <[email protected]>; Khatri, Sunil <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 67 lines · Prike

AMD General

Regards,
Prike

Quoted 61 lines

-----Original Message-----
From: Liang, Prike
Sent: Tuesday, July 14, 2026 10:49 AM
To: 'Alex Deucher' <[email protected]>; [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Khatri, Sunil <[email protected]>
Subject: RE: [PATCH] drm/amdgpu/userq: properly account for resets

Regards,
Prike

Quoted 41 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 4:39 AM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets

We need to increment the reset counter, force fence completion, and set the wedged event when a user queue is reset.

mes_userq_reset_queue() handles this for collateral damage, but the caller needs to handle this directly for the original guilty queue.

Cc Srinivasan Shanmugam <[email protected]>
Cc Prike Liang <[email protected]>
Cc Sunil Khatri <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +0 -0
index 6aa75da27f912..5e1262636e1e9 100644
@@ -146,8 +146,13 @@ static void

amdgpu_userq_hang_detect_work(struct
work_struct *work)

                                                     queue, NULL, NULL);
            else
                    r = userq_funcs->reset(queue);
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +5 -1
-           if (r)
+           if (r) {
                    gpu_reset = true;
+           } else {
+                   atomic_inc(&adev->gpu_reset_counter);
+                   amdgpu_userq_fence_driver_force_completion(queue);
+                   drm_dev_wedged_event(adev_to_drm(adev),
Quoted 1 line

DRM_WEDGE_RECOVERY_NONE, NULL);

If we do the userq reset post similar process here just like as mes_userq_reset_queue(), so do we need to clear the duplicate the code in the mes_userq_reset_queue()? Otherwise, the gpu reset counter

should be incorrect.

Quoted 3 lines
Quoted 1 line

Except that, we might input the wedge task info to report the guilty process info.

As to the first question, I recheck the code and found out the guilty user queue in

the amdgpu_gfx_reset_mes_compute() was reset separately before mes_userq_reset_queue(), with this current reset sequence then this change it make sense to update reset counter and force complete fence.

Quoted 1 line
Quoted 8 lines
Quoted 1 line

But why need to reset the original guilty queue separately, should this guilty queue

was found by the amdgpu_mes_detect_and_reset_hung_queues() in the MES firmware? If so, we may need to reset the guilty queue through mes_userq_reset_queue() rather than reset separately, then we don't need to add this change anymore.

We still need it for gfx and sdma queues. Those don't get reset via amdgpu_gfx_reset_mes_compute(),

Yes, I will check whether the original guilty compute queue can be identified by amdgpu_mes_detect_and_reset_hung_queues(). If so, the separate guilty queue reset via mes_userq_reset_queue() becomes redundant and can be removed. The compute userq reset refcount and completed fence handling would then be relocated into amdgpu_userq_hang_detect_work(), consolidating the reset path into a single flow.

Thanks,
Prike

Quoted 11 lines

Alex

Quoted 9 lines
Quoted 8 lines
Quoted 8 lines

+ }

    } else {
            gpu_reset = true;
    }

--
2.55.0

#6 lore raw
Quoted 97 lines · Prike

AMD General

Regards,
Prike

Quoted 91 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 9:52 PM
To: Liang, Prike <[email protected]>
Cc: Deucher, Alexander <[email protected]>; amd- [email protected]; SHANMUGAM, SRINIVASAN <[email protected]>; Khatri, Sunil <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 67 lines · Prike

AMD General

Regards,
Prike

Quoted 61 lines

-----Original Message-----
From: Liang, Prike
Sent: Tuesday, July 14, 2026 10:49 AM
To: 'Alex Deucher' <[email protected]>; [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Khatri, Sunil <[email protected]>
Subject: RE: [PATCH] drm/amdgpu/userq: properly account for resets

Regards,
Prike

Quoted 41 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 4:39 AM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets

We need to increment the reset counter, force fence completion, and set the wedged event when a user queue is reset.

mes_userq_reset_queue() handles this for collateral damage, but the caller needs to handle this directly for the original guilty queue.

Cc Srinivasan Shanmugam <[email protected]>
Cc Prike Liang <[email protected]>
Cc Sunil Khatri <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +0 -0
index 6aa75da27f912..5e1262636e1e9 100644
@@ -146,8 +146,13 @@ static void

amdgpu_userq_hang_detect_work(struct
work_struct *work)

                                                     queue, NULL, NULL);
            else
                    r = userq_funcs->reset(queue);
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +5 -1
-           if (r)
+           if (r) {
                    gpu_reset = true;
+           } else {
+                   atomic_inc(&adev->gpu_reset_counter);
+                   amdgpu_userq_fence_driver_force_completion(queue);
+                   drm_dev_wedged_event(adev_to_drm(adev),
Quoted 1 line

DRM_WEDGE_RECOVERY_NONE, NULL);

If we do the userq reset post similar process here just like as mes_userq_reset_queue(), so do we need to clear the duplicate the code in the mes_userq_reset_queue()? Otherwise, the gpu reset counter

should be incorrect.

Quoted 3 lines
Quoted 1 line

Except that, we might input the wedge task info to report the guilty process info.

As to the first question, I recheck the code and found out the guilty user queue in

the amdgpu_gfx_reset_mes_compute() was reset separately before mes_userq_reset_queue(), with this current reset sequence then this change it make sense to update reset counter and force complete fence.

Quoted 2 lines

But why need to reset the original guilty queue separately, should this guilty queue

was found by the amdgpu_mes_detect_and_reset_hung_queues() in the MES firmware? If so, we may need to reset the guilty queue through mes_userq_reset_queue() rather than reset separately, then we don't need to add this change anymore.

Quoted 5 lines
Quoted 2 lines

We still need it for gfx and sdma queues. Those don't get reset via amdgpu_gfx_reset_mes_compute(),

Yes, I will check whether the original guilty compute queue can be identified by amdgpu_mes_detect_and_reset_hung_queues(). If so, the separate guilty queue reset via mes_userq_reset_queue() becomes redundant and can be removed. The compute userq reset refcount and completed fence handling would then be relocated into amdgpu_userq_hang_detect_work(), consolidating the reset path into a single flow.

I don't think we need that; see V3 of the patch I sent out earlier today.

Alex

Quoted 15 lines

Thanks,
Prike

Quoted 11 lines

Alex

Quoted 9 lines
Quoted 8 lines
Quoted 8 lines

+ }

    } else {
            gpu_reset = true;
    }

--
2.55.0

Reviewed-by on v1 1/1
#7 lore raw
In reply to Alex Deucher

AMD General

Quoted 43 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 2:09 AM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets

We need to increment the reset counter, force fence completion, and set the wedged event when a user queue is reset.

mes_userq_reset_queue() handles this for collateral damage, but the caller needs to handle this directly for the original guilty queue.

Cc Srinivasan Shanmugam <[email protected]>
Cc Prike Liang <[email protected]>
Cc Sunil Khatri <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +6 -1
index 6aa75da27f912..5e1262636e1e9 100644
@@ -146,8 +146,13 @@ static void amdgpu_userq_hang_detect_work(struct
 work_struct *work)
                                                       queue, NULL, NULL);
              else
                      r = userq_funcs->reset(queue);
-             if (r)
+             if (r) {
                      gpu_reset = true;
+             } else {
+                     atomic_inc(&adev->gpu_reset_counter);
+                     amdgpu_userq_fence_driver_force_completion(queue);
+                     drm_dev_wedged_event(adev_to_drm(adev),
 DRM_WEDGE_RECOVERY_NONE, NULL);
+             }
      } else {
Quoted 2 lines
              gpu_reset = true;
      }

After the original queue was reset successfully, it did not update gpu_reset_counter, complete its pending fences, or send the wedged event. mes_userq_reset_queue() already updates gpu_reset_counter, completes the pending fences, and sends the wedged event for the other affected queues, but skips the original queue because it has already been reset.
This patch adds the same steps for the original queue, making the reset handling consistent for both the original queue and the other affected queues.

Quoted 2 lines

--
2.55.0

#8 lore raw
Quoted 43 lines · SRINIVASAN

AMD General

Quoted 41 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 2:09 AM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets

We need to increment the reset counter, force fence completion, and set the wedged event when a user queue is reset.

mes_userq_reset_queue() handles this for collateral damage, but the caller needs to handle this directly for the original guilty queue.

Cc Srinivasan Shanmugam <[email protected]>
Cc Prike Liang <[email protected]>
Cc Sunil Khatri <[email protected]>
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++-
  1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +5 -1
index 6aa75da27f912..5e1262636e1e9 100644
@@ -146,8 +146,13 @@ static void amdgpu_userq_hang_detect_work(struct
 work_struct *work)
                                                        queue, NULL, NULL);
               else
                       r = userq_funcs->reset(queue);
-             if (r)
+             if (r) {
                       gpu_reset = true;
+             } else {
+                     atomic_inc(&adev->gpu_reset_counter);
+                     amdgpu_userq_fence_driver_force_completion(queue);
+                     drm_dev_wedged_event(adev_to_drm(adev),
 DRM_WEDGE_RECOVERY_NONE, NULL);
Quoted 8 lines
Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +1 -0
+             }
       } else {
               gpu_reset = true;
       }

After the original queue was reset successfully, it did not update gpu_reset_counter, complete its pending fences, or send the wedged event. mes_userq_reset_queue() already updates gpu_reset_counter, completes the pending fences, and sends the wedged event for the other affected queues, but skips the original queue because it has already been reset.

What is the rationale of sending multiple device wedged events on a per queue basis?

Thanks,
Lijo

Quoted 7 lines

This patch adds the same steps for the original queue, making the reset handling consistent for both the original queue and the other affected queues.

Quoted 2 lines

--
2.55.0

#9 lore raw

AMD General

Get Outlook for Android<https://aka.ms/AAb9ysg>

________________________________
From: Lazar, Lijo <[email protected]>
Sent: Tuesday, July 14, 2026 3:14:34 PM
To: SHANMUGAM, SRINIVASAN <[email protected]>; Deucher, Alexander <[email protected]>; [email protected] <[email protected]>
Cc: Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 43 lines · SRINIVASAN

AMD General

Quoted 41 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 2:09 AM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets

We need to increment the reset counter, force fence completion, and set the wedged event when a user queue is reset.

mes_userq_reset_queue() handles this for collateral damage, but the caller needs to handle this directly for the original guilty queue.

Cc Srinivasan Shanmugam <[email protected]>
Cc Prike Liang <[email protected]>
Cc Sunil Khatri <[email protected]>
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++-
  1 file changed, 6 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +5 -1
index 6aa75da27f912..5e1262636e1e9 100644
@@ -146,8 +146,13 @@ static void amdgpu_userq_hang_detect_work(struct
 work_struct *work)
                                                        queue, NULL, NULL);
               else
                       r = userq_funcs->reset(queue);
-             if (r)
+             if (r) {
                       gpu_reset = true;
+             } else {
+                     atomic_inc(&adev->gpu_reset_counter);
+                     amdgpu_userq_fence_driver_force_completion(queue);
+                     drm_dev_wedged_event(adev_to_drm(adev),
 DRM_WEDGE_RECOVERY_NONE, NULL);
Quoted 8 lines
Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +1 -0
+             }
       } else {
               gpu_reset = true;
       }

After the original queue was reset successfully, it did not update gpu_reset_counter, complete its pending fences, or send the wedged event. mes_userq_reset_queue() already updates gpu_reset_counter, completes the pending fences, and sends the wedged event for the other affected queues, but skips the original queue because it has already been reset.

What is the rationale of sending multiple device wedged events on a per queue basis?

The question of whether drm_dev_wedged_event() should be emitted once per queue or once per overall recovery seems like a broader design discussion.

Thanks,
Srini

#10 lore raw
Quoted 70 lines · SRINIVASAN

AMD General

Get Outlook for Android <https://aka.ms/AAb9ysg>

------------------------------------------------------------------------
*From:* Lazar, Lijo <[email protected]>
*Sent:* Tuesday, July 14, 2026 3:14:34 PM
*To:* SHANMUGAM, SRINIVASAN <[email protected]>; Deucher, Alexander <[email protected]>; [email protected] <[email protected]> *Cc:* Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
*Subject:* Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 12 lines · SRINIVASAN

AMD General

Quoted 10 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 2:09 AM
To: [email protected]
Cc: Deucher, Alexander <[email protected]>; SHANMUGAM, SRINIVASAN <[email protected]>; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets

We need to increment the reset counter, force fence completion, and

set the

Quoted 3 lines
Quoted 3 lines

wedged event when a user queue is reset.

mes_userq_reset_queue() handles this for collateral damage, but the

caller needs

Quoted 10 lines
Quoted 10 lines

to handle this directly for the original guilty queue.

Cc Srinivasan Shanmugam <[email protected]>
Cc Prike Liang <[email protected]>
Cc Sunil Khatri <[email protected]>

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++-

   1 file changed, 6 insertions(+), 1 deletion(-)
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c static void amdgpu_userq_hang_detect_work(struct · line 147

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +0 -0
index 6aa75da27f912..5e1262636e1e9 100644
@@ -146,8 +146,13 @@ static void amdgpu_userq_hang_detect_work(struct
 work_struct *work)
                                                          queue, NULL, 

NULL);

Quoted 15 lines
Quoted 6 lines

                else
                        r = userq_funcs->reset(queue); -             if (r)
+             if (r) {
                        gpu_reset = true;
+             } else {

Quoted 8 lines

+                     atomic_inc(&adev->gpu_reset_counter); +                     amdgpu_userq_fence_driver_force_completion(queue); +                     drm_dev_wedged_event(adev_to_drm(adev), DRM_WEDGE_RECOVERY_NONE, NULL);
+             }
        } else {
                gpu_reset = true;
        }

Quoted 1 line

After the original queue was reset successfully, it did not update

gpu_reset_counter, complete its pending fences, or send the wedged event.

Quoted 9 lines
Quoted 1 line

mes_userq_reset_queue() already updates gpu_reset_counter, completes

the pending fences, and sends the wedged event for the other affected queues,

Quoted 1 line

but skips the original queue because it has already been reset.

What is the rationale of sending multiple device wedged events on a per queue basis?

The question of whether drm_dev_wedged_event() should be emitted once per queue or once per overall recovery seems like a broader design discussion.

Along with that, also need to consider if device reset_counter needs to be incremented on a per queue basis or based on reset event recovery. It could get incremented multiple times inside this - mes_userq_reset_queue.

Thanks,
Lijo

Quoted 5 lines

Thanks,
Srini

#11 lore raw

AMD General

Quoted 89 lines

-----Original Message-----
From: Lazar, Lijo <[email protected]>
Sent: Tuesday, July 14, 2026 4:02 PM
To: SHANMUGAM, SRINIVASAN <[email protected]>; Deucher, Alexander <[email protected]>; amd- [email protected]
Cc: Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 5 lines · SRINIVASAN

AMD General

Quoted 7 lines

Get Outlook for Android <https://aka.ms/AAb9ysg>

----------------------------------------------------------------------
--
*From:* Lazar, Lijo <[email protected]>
*Sent:* Tuesday, July 14, 2026 3:14:34 PM
*To:* SHANMUGAM, SRINIVASAN <[email protected]>;

Deucher,

Quoted 64 lines

Alexander <[email protected]>; [email protected] <[email protected]>
*Cc:* Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
*Subject:* Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 4 lines · SRINIVASAN

AMD General

Quoted 2 lines

-----Original Message-----
From: Alex Deucher <[email protected]> >> Sent: Tuesday,

July 14, 2026 2:09 AM >> To: [email protected] >> Cc:
Deucher, Alexander <[email protected]>; SHANMUGAM, >> SRINIVASAN <[email protected]>; Liang, Prike >> <[email protected]>; Khatri, Sunil <[email protected]> >>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets >> >> We need to increment the reset counter, force fence completion, and set the >> wedged event when a user queue is reset.

Quoted 2 lines
Quoted 2 lines

mes_userq_reset_queue() handles this for collateral damage, but

the caller needs >> to handle this directly for the original guilty queue.

Quoted 2 lines
Quoted 2 lines

Srinivasan Shanmugam <[email protected]> >> Cc: Prike Liang <[email protected]> >> Cc: Sunil Khatri <[email protected]> >> --- >>
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++- >> 1 file
changed, 6 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

Quoted 2 lines
Quoted 2 lines

b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
index 6aa75da27f912..5e1262636e1e9 100644 >> ---

a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

Quoted 2 lines

Review static void

Quoted 2 lines

+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

(quoted diff) +0 -0
@@ -146,8 +146,13 @@ static void

amdgpu_userq_hang_detect_work(struct

Quoted 2 lines

Review static void · line 147

Quoted 2 lines
(quoted diff) +0 -0
 work_struct *work)
                                                        queue,

NULL, NULL); >> else >> r =
userq_funcs->reset(queue); >> - if (r) >> +
if (r) { >> gpu_reset = true; >> +
} else { >> +
atomic_inc(&adev->gpu_reset_counter);

Quoted 1 line

Review static void · line 148

Quoted 1 line
(quoted diff) +1 -0
+

amdgpu_userq_fence_driver_force_completion(queue);

Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c static void amdgpu_userq_hang_detect_work(struct work_struct *work) · line 158

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +2 -0
+                     drm_dev_wedged_event(adev_to_drm(adev),
 DRM_WEDGE_RECOVERY_NONE, NULL);
+             }
       } else {
               gpu_reset = true;
       }

After the original queue was reset successfully, it did not update

gpu_reset_counter, complete its pending fences, or send the wedged event.

Quoted 1 line

mes_userq_reset_queue() already updates gpu_reset_counter,

completes the pending fences, and sends the wedged event for the other affected queues, > but skips the original queue because it has already been reset.

What is the rationale of sending multiple device wedged events on a per queue basis?

Quoted 8 lines
Quoted 4 lines

The question of whether drm_dev_wedged_event() should be emitted once per queue or once per overall recovery seems like a broader design discussion.

Along with that, also need to consider if device reset_counter needs to be incremented on a per queue basis or based on reset event recovery. It could get incremented multiple times inside this - mes_userq_reset_queue.

Looking at the current flow, both gpu_reset_counter and drm_dev_wedged_event() are updated once for each successfully reset queue. It would be helpful to clarify whether they are intended to be updated per affected queue or once per overall recovery.

Thanks,
Srini

Quoted 9 lines

Thanks,
Lijo

Quoted 5 lines

Thanks,
Srini

#12 lore raw
Quoted 97 lines · SRINIVASAN

AMD General

Quoted 94 lines

-----Original Message-----
From: Lazar, Lijo <[email protected]>
Sent: Tuesday, July 14, 2026 4:02 PM
To: SHANMUGAM, SRINIVASAN <[email protected]>; Deucher, Alexander <[email protected]>; amd- [email protected]
Cc: Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 5 lines · SRINIVASAN

AMD General

Quoted 7 lines

Get Outlook for Android <https://aka.ms/AAb9ysg>

----------------------------------------------------------------------
--
*From:* Lazar, Lijo <[email protected]>
*Sent:* Tuesday, July 14, 2026 3:14:34 PM
*To:* SHANMUGAM, SRINIVASAN <[email protected]>;

Deucher,

Quoted 68 lines

Alexander <[email protected]>; [email protected] <[email protected]>
*Cc:* Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
*Subject:* Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 4 lines · SRINIVASAN

AMD General

Quoted 2 lines

-----Original Message-----
From: Alex Deucher <[email protected]> >> Sent: Tuesday,

July 14, 2026 2:09 AM >> To: [email protected] >> Cc:
Deucher, Alexander <[email protected]>; SHANMUGAM, >> SRINIVASAN <[email protected]>; Liang, Prike >> <[email protected]>; Khatri, Sunil <[email protected]> >>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets >> >> We need to increment the reset counter, force fence completion, and set the >> wedged event when a user queue is reset.

Quoted 2 lines
Quoted 2 lines

mes_userq_reset_queue() handles this for collateral damage, but

the caller needs >> to handle this directly for the original guilty queue.

Quoted 2 lines
Quoted 2 lines

Srinivasan Shanmugam <[email protected]> >> Cc: Prike Liang <[email protected]> >> Cc: Sunil Khatri <[email protected]> >> --- >>
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++- >> 1 file
changed, 6 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

Quoted 2 lines
Quoted 2 lines

b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
index 6aa75da27f912..5e1262636e1e9 100644 >> ---

a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

Quoted 2 lines

Review static void

Quoted 2 lines

+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

(quoted diff) +0 -0
@@ -146,8 +146,13 @@ static void

amdgpu_userq_hang_detect_work(struct

Quoted 2 lines

Review static void · line 147

Quoted 2 lines
(quoted diff) +0 -0
 work_struct *work)
                                                        queue,

NULL, NULL); >> else >> r =
userq_funcs->reset(queue); >> - if (r) >> +
if (r) { >> gpu_reset = true; >> +
} else { >> +
atomic_inc(&adev->gpu_reset_counter);

Quoted 1 line

Review static void · line 148

Quoted 1 line
(quoted diff) +1 -0
+

amdgpu_userq_fence_driver_force_completion(queue);

Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c static void amdgpu_userq_hang_detect_work(struct work_struct *work) · line 158

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +2 -0
+                     drm_dev_wedged_event(adev_to_drm(adev),
 DRM_WEDGE_RECOVERY_NONE, NULL);
+             }
       } else {
               gpu_reset = true;
       }

After the original queue was reset successfully, it did not update

gpu_reset_counter, complete its pending fences, or send the wedged event.

Quoted 1 line

mes_userq_reset_queue() already updates gpu_reset_counter,

completes the pending fences, and sends the wedged event for the other affected queues, > but skips the original queue because it has already been reset.

What is the rationale of sending multiple device wedged events on a per queue basis?

The question of whether drm_dev_wedged_event() should be emitted once per queue or once per overall recovery seems like a broader design discussion.

Quoted 6 lines
Quoted 3 lines

Along with that, also need to consider if device reset_counter needs to be incremented on a per queue basis or based on reset event recovery. It could get incremented multiple times inside this - mes_userq_reset_queue.

Looking at the current flow, both gpu_reset_counter and drm_dev_wedged_event() are updated once for each successfully reset queue. It would be helpful to clarify whether they are intended to be updated per affected queue or once per overall recovery.

What are the semantics around the reset counter and wedged events? Presumably each should be incremented for each queue that is reset? If a hang affects multiple queues shouldn't each be a separate "reset"? In the most common case, there should just be one since queue reset should be able to reset just the guilty queue.

Alex

Quoted 13 lines

Thanks,
Srini

Quoted 9 lines

Thanks,
Lijo

Quoted 5 lines

Thanks,
Srini

#13 lore raw

AMD General

Quoted 126 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 7:26 PM
To: SHANMUGAM, SRINIVASAN <[email protected]>
Cc: Lazar, Lijo <[email protected]>; Deucher, Alexander <[email protected]>; [email protected]; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 7 lines · SRINIVASAN

AMD General

Quoted 4 lines

-----Original Message-----
From: Lazar, Lijo <[email protected]>
Sent: Tuesday, July 14, 2026 4:02 PM
To: SHANMUGAM, SRINIVASAN <[email protected]>;

Deucher,

Quoted 14 lines
Quoted 14 lines

Alexander <[email protected]>; amd- [email protected]
Cc: Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 5 lines · SRINIVASAN

AMD General

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

Get Outlook for Android <https://aka.ms/AAb9ysg>

------------------------------------------------------------------
----
--
*From:* Lazar, Lijo <[email protected]>
*Sent:* Tuesday, July 14, 2026 3:14:34 PM
*To:* SHANMUGAM, SRINIVASAN

<[email protected]>;

Quoted 10 lines
Quoted 10 lines

Deucher,

Quoted 9 lines

Alexander <[email protected]>; [email protected] <[email protected]> *Cc:* Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
*Subject:* Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines · SRINIVASAN
Quoted 4 lines

AMD General

Quoted 2 lines

-----Original Message-----
From: Alex Deucher <[email protected]> >> Sent:

Tuesday, July 14, 2026 2:09 AM >> To: [email protected] >>

Cc:

Quoted 40 lines
Quoted 40 lines
Quoted 40 lines

Deucher, Alexander <[email protected]>; SHANMUGAM, >> SRINIVASAN <[email protected]>; Liang, Prike >> <[email protected]>; Khatri, Sunil <[email protected]> >>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets >>

Quoted 1 line
Quoted 1 line

We need to increment the reset counter, force fence completion,

and set the >> wedged event when a user queue is reset.

Quoted 2 lines
Quoted 2 lines

mes_userq_reset_queue() handles this for collateral damage,

but the caller needs >> to handle this directly for the original guilty queue.

Quoted 2 lines
Quoted 2 lines

Srinivasan Shanmugam <[email protected]> >> Cc: Prike Liang <[email protected]> >> Cc: Sunil Khatri <[email protected]> >> --- >>
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++- >> 1 file
changed, 6 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

Quoted 2 lines
Quoted 2 lines

b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
index 6aa75da27f912..5e1262636e1e9 100644 >> ---

a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

Quoted 2 lines

Review static void

Quoted 2 lines

+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

(quoted diff) +0 -0
@@ -146,8 +146,13 @@ static void

amdgpu_userq_hang_detect_work(struct

Quoted 2 lines

Review static void · line 147

Quoted 2 lines
(quoted diff) +0 -0
 work_struct *work)
                                                        queue,

NULL, NULL); >> else >> r =
userq_funcs->reset(queue); >> - if (r) >> +
if (r) { >> gpu_reset = true; >> +
} else { >> +
atomic_inc(&adev->gpu_reset_counter);

Quoted 1 line

Review static void · line 148

Quoted 1 line
(quoted diff) +1 -0
+

amdgpu_userq_fence_driver_force_completion(queue);

Quoted 7 lines
Quoted 6 lines
(quoted diff) +2 -0
+                     drm_dev_wedged_event(adev_to_drm(adev),
 DRM_WEDGE_RECOVERY_NONE, NULL);
+             }
       } else {
               gpu_reset = true;
       }
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 1 line

After the original queue was reset successfully, it did not

update gpu_reset_counter, complete its pending fences, or send the wedged

event.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 1 line

mes_userq_reset_queue() already updates gpu_reset_counter,

completes the pending fences, and sends the wedged event for the other affected queues, > but skips the original queue because it has already been reset.

What is the rationale of sending multiple device wedged events on a per queue basis?

Quoted 8 lines
Quoted 8 lines
Quoted 4 lines

The question of whether drm_dev_wedged_event() should be emitted once per queue or once per overall recovery seems like a broader design discussion.

Along with that, also need to consider if device reset_counter needs to be incremented on a per queue basis or based on reset event recovery. It could get incremented multiple times inside this -

mes_userq_reset_queue.

Quoted 2 lines

Looking at the current flow, both gpu_reset_counter and

drm_dev_wedged_event() are updated once for each successfully reset queue. It would be helpful to clarify whether they are intended to be updated per affected queue or once per overall recovery.

Quoted 1 line
Quoted 5 lines

What are the semantics around the reset counter and wedged events? Presumably each should be incremented for each queue that is reset? If a hang affects multiple queues shouldn't each be a separate "reset"?
In the most common case, there should just be one since queue reset should be able to reset just the guilty queue.

Thanks for the clarification, Alex. Understood that the reset counter and wedged event are intended to be updated once for each queue that is reset.

Best Regards,
Srini

Quoted 2 lines

Alex

#14 lore raw
Quoted 129 lines · SRINIVASAN

AMD General

Quoted 126 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Tuesday, July 14, 2026 7:26 PM
To: SHANMUGAM, SRINIVASAN <[email protected]>
Cc: Lazar, Lijo <[email protected]>; Deucher, Alexander <[email protected]>; [email protected]; Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 7 lines · SRINIVASAN

AMD General

Quoted 4 lines

-----Original Message-----
From: Lazar, Lijo <[email protected]>
Sent: Tuesday, July 14, 2026 4:02 PM
To: SHANMUGAM, SRINIVASAN <[email protected]>;

Deucher,

Quoted 14 lines
Quoted 14 lines

Alexander <[email protected]>; amd- [email protected]
Cc: Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 5 lines · SRINIVASAN

AMD General

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines

Get Outlook for Android <https://aka.ms/AAb9ysg>

------------------------------------------------------------------
----
--
*From:* Lazar, Lijo <[email protected]>
*Sent:* Tuesday, July 14, 2026 3:14:34 PM
*To:* SHANMUGAM, SRINIVASAN

<[email protected]>;

Quoted 10 lines
Quoted 10 lines

Deucher,

Quoted 9 lines

Alexander <[email protected]>; [email protected] <[email protected]> *Cc:* Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
*Subject:* Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines · SRINIVASAN
Quoted 4 lines

AMD General

Quoted 2 lines

-----Original Message-----
From: Alex Deucher <[email protected]> >> Sent:

Tuesday, July 14, 2026 2:09 AM >> To: [email protected] >>

Cc:

Quoted 40 lines
Quoted 40 lines
Quoted 40 lines

Deucher, Alexander <[email protected]>; SHANMUGAM, >> SRINIVASAN <[email protected]>; Liang, Prike >> <[email protected]>; Khatri, Sunil <[email protected]> >>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets >>

Quoted 1 line
Quoted 1 line

We need to increment the reset counter, force fence completion,

and set the >> wedged event when a user queue is reset.

Quoted 2 lines
Quoted 2 lines

mes_userq_reset_queue() handles this for collateral damage,

but the caller needs >> to handle this directly for the original guilty queue.

Quoted 2 lines
Quoted 2 lines

Srinivasan Shanmugam <[email protected]> >> Cc: Prike Liang <[email protected]> >> Cc: Sunil Khatri <[email protected]> >> --- >>
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++- >> 1 file
changed, 6 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

Quoted 2 lines
Quoted 2 lines

b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
index 6aa75da27f912..5e1262636e1e9 100644 >> ---

a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

Quoted 2 lines

Review static void

Quoted 2 lines

+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

(quoted diff) +0 -0
@@ -146,8 +146,13 @@ static void

amdgpu_userq_hang_detect_work(struct

Quoted 2 lines

Review static void · line 147

Quoted 2 lines
(quoted diff) +0 -0
 work_struct *work)
                                                        queue,

NULL, NULL); >> else >> r =
userq_funcs->reset(queue); >> - if (r) >> +
if (r) { >> gpu_reset = true; >> +
} else { >> +
atomic_inc(&adev->gpu_reset_counter);

Quoted 1 line

Review static void · line 148

Quoted 1 line
(quoted diff) +1 -0
+

amdgpu_userq_fence_driver_force_completion(queue);

Quoted 7 lines
Quoted 6 lines
(quoted diff) +2 -0
+                     drm_dev_wedged_event(adev_to_drm(adev),
 DRM_WEDGE_RECOVERY_NONE, NULL);
+             }
       } else {
               gpu_reset = true;
       }
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 1 line

After the original queue was reset successfully, it did not

update gpu_reset_counter, complete its pending fences, or send the wedged

event.

Quoted 8 lines
Quoted 8 lines
Quoted 8 lines
Quoted 1 line

mes_userq_reset_queue() already updates gpu_reset_counter,

completes the pending fences, and sends the wedged event for the other affected queues, > but skips the original queue because it has already been reset.

What is the rationale of sending multiple device wedged events on a per queue basis?

Quoted 8 lines
Quoted 8 lines
Quoted 4 lines

The question of whether drm_dev_wedged_event() should be emitted once per queue or once per overall recovery seems like a broader design discussion.

Along with that, also need to consider if device reset_counter needs to be incremented on a per queue basis or based on reset event recovery. It could get incremented multiple times inside this -

mes_userq_reset_queue.

Quoted 2 lines

Looking at the current flow, both gpu_reset_counter and

drm_dev_wedged_event() are updated once for each successfully reset queue. It would be helpful to clarify whether they are intended to be updated per affected queue or once per overall recovery.

Quoted 1 line
Quoted 7 lines
Quoted 5 lines

What are the semantics around the reset counter and wedged events? Presumably each should be incremented for each queue that is reset? If a hang affects multiple queues shouldn't each be a separate "reset"?
In the most common case, there should just be one since queue reset should be able to reset just the guilty queue.

Thanks for the clarification, Alex. Understood that the reset counter and wedged event are intended to be updated once for each queue that is reset.

Well, I guess that is the question. We are the semantics around these?

Alex

Quoted 6 lines

Best Regards,
Srini

Quoted 2 lines

Alex

#15 lore raw
In reply to Alex Deucher
Quoted 103 lines · Alex Deucher
Quoted 101 lines · SRINIVASAN

AMD General

Quoted 97 lines

-----Original Message-----
From: Lazar, Lijo <[email protected]>
Sent: Tuesday, July 14, 2026 4:02 PM
To: SHANMUGAM, SRINIVASAN <[email protected]>; Deucher, Alexander <[email protected]>; amd- [email protected]
Cc: Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 5 lines · SRINIVASAN

AMD General

Quoted 7 lines

Get Outlook for Android <https://aka.ms/AAb9ysg>

----------------------------------------------------------------------
--
*From:* Lazar, Lijo <[email protected]>
*Sent:* Tuesday, July 14, 2026 3:14:34 PM
*To:* SHANMUGAM, SRINIVASAN <[email protected]>;

Deucher,

Quoted 61 lines

Alexander <[email protected]>; [email protected] <[email protected]>
*Cc:* Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
*Subject:* Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 4 lines · SRINIVASAN

AMD General

Quoted 2 lines

-----Original Message-----
From: Alex Deucher <[email protected]> >> Sent: Tuesday,

July 14, 2026 2:09 AM >> To: [email protected] >> Cc:
Deucher, Alexander <[email protected]>; SHANMUGAM, >> SRINIVASAN <[email protected]>; Liang, Prike >> <[email protected]>; Khatri, Sunil <[email protected]> >>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets >> >> We need to increment the reset counter, force fence completion, and set the >> wedged event when a user queue is reset.

Quoted 2 lines
Quoted 2 lines

mes_userq_reset_queue() handles this for collateral damage, but

the caller needs >> to handle this directly for the original guilty queue.

Quoted 2 lines
Quoted 2 lines

Srinivasan Shanmugam <[email protected]> >> Cc: Prike Liang <[email protected]> >> Cc: Sunil Khatri <[email protected]> >> --- >>
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++- >> 1 file
changed, 6 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

Quoted 2 lines
Quoted 2 lines

b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
index 6aa75da27f912..5e1262636e1e9 100644 >> ---

a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

Quoted 2 lines

Review static void

Quoted 2 lines

+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c

(quoted diff) +0 -0
@@ -146,8 +146,13 @@ static void

amdgpu_userq_hang_detect_work(struct

Quoted 2 lines

Review static void · line 147

Quoted 2 lines
(quoted diff) +0 -0
 work_struct *work)
                                                        queue,

NULL, NULL); >> else >> r =
userq_funcs->reset(queue); >> - if (r) >> +
if (r) { >> gpu_reset = true; >> +
} else { >> +
atomic_inc(&adev->gpu_reset_counter);

Quoted 1 line

Review static void · line 148

Quoted 1 line
(quoted diff) +1 -0
+

amdgpu_userq_fence_driver_force_completion(queue);

Quoted 8 lines

Review static void · line 154

Quoted 6 lines
(quoted diff) +2 -0
+                     drm_dev_wedged_event(adev_to_drm(adev),
 DRM_WEDGE_RECOVERY_NONE, NULL);
+             }
       } else {
               gpu_reset = true;
       }

After the original queue was reset successfully, it did not update

gpu_reset_counter, complete its pending fences, or send the wedged event.

Quoted 1 line

mes_userq_reset_queue() already updates gpu_reset_counter,

completes the pending fences, and sends the wedged event for the other affected queues, > but skips the original queue because it has already been reset.

Quoted 7 lines

What is the rationale of sending multiple device wedged events on a per queue basis?

The question of whether drm_dev_wedged_event() should be emitted once per queue or once per overall recovery seems like a broader design discussion.

Along with that, also need to consider if device reset_counter needs to be incremented on a per queue basis or based on reset event recovery. It could get incremented multiple times inside this - mes_userq_reset_queue.

Quoted 9 lines
Quoted 2 lines

Looking at the current flow, both gpu_reset_counter and drm_dev_wedged_event() are updated once for each successfully reset queue. It would be helpful to clarify whether they are intended to be updated per affected queue or once per overall recovery.

What are the semantics around the reset counter and wedged events? Presumably each should be incremented for each queue that is reset? If a hang affects multiple queues shouldn't each be a separate "reset"? In the most common case, there should just be one since queue reset should be able to reset just the guilty queue.

Device wedged event is a udev event. It's not good to send multiple events when multiple queues are recovered as part of single recovery process. Otherwise userspace monitoring the device will see back to back events. There is no context associated so it doesn't have any clue of what sort of reset happened.

I think the same goes with device. I believe it's to indicate a recovery event associated with device and not as a per queue event.

Thanks,
Lijo

Quoted 15 lines

Alex

Quoted 13 lines

Thanks,
Srini

Quoted 9 lines

Thanks,
Lijo

Quoted 5 lines

Thanks,
Srini

#16 lore raw
Quoted 116 lines · Lijo
Quoted 103 lines · Alex Deucher
Quoted 101 lines · SRINIVASAN

AMD General

Quoted 97 lines

-----Original Message-----
From: Lazar, Lijo <[email protected]>
Sent: Tuesday, July 14, 2026 4:02 PM
To: SHANMUGAM, SRINIVASAN <[email protected]>; Deucher, Alexander <[email protected]>; amd- [email protected]
Cc: Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
Subject: Re: [PATCH] drm/amdgpu/userq: properly account for resets

Quoted 5 lines · SRINIVASAN

AMD General

Quoted 7 lines

Get Outlook for Android <https://aka.ms/AAb9ysg>

----------------------------------------------------------------------

Signature · 3 lines
*From:* Lazar, Lijo <[email protected]>
*Sent:* Tuesday, July 14, 2026 3:14:34 PM
*To:* SHANMUGAM, SRINIVASAN <[email protected]>;

Deucher,

Quoted 61 lines

Alexander <[email protected]>; [email protected] <[email protected]>
*Cc:* Liang, Prike <[email protected]>; Khatri, Sunil <[email protected]>
*Subject:* Re: [PATCH] drm/amdgpu/userq: properly account for resets

On 14-Jul-26 10:16 AM, SHANMUGAM, SRINIVASAN wrote:
  > AMD General
  >
  >> -----Original Message-----
  >> From: Alex Deucher <[email protected]>  >> Sent: Tuesday, July 14, 2026 2:09 AM  >> To: [email protected]  >> Cc:
Deucher, Alexander <[email protected]>; SHANMUGAM,  >> SRINIVASAN <[email protected]>; Liang, Prike  >> <[email protected]>; Khatri, Sunil <[email protected]>  >>
Subject: [PATCH] drm/amdgpu/userq: properly account for resets  >>  >> We need to increment the reset counter, force fence completion, and set the  >> wedged event when a user queue is reset.
  >>
  >> mes_userq_reset_queue() handles this for collateral damage, but the caller needs  >> to handle this directly for the original guilty queue.
  >>
  >> Signed-off-by: Alex Deucher <[email protected]>  >> Cc:
Srinivasan Shanmugam <[email protected]>  >> Cc: Prike Liang <[email protected]>  >> Cc: Sunil Khatri <[email protected]>  >> ---  >> drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 7 ++++++-  >>   1 file changed, 6 insertions(+), 1 deletion(-)  >>  >> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
  >> b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
  >> index 6aa75da27f912..5e1262636e1e9 100644  >> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
  >> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
  >> @@ -146,8 +146,13 @@ static void amdgpu_userq_hang_detect_work(struct
  >> work_struct *work)
  >>                                                         queue, NULL, NULL);  >>                else  >>                        r = userq_funcs->reset(queue);  >> -             if (r)  >> +
if (r) {  >>                        gpu_reset = true;  >> +
} else {  >> +
atomic_inc(&adev->gpu_reset_counter);
  >> +
amdgpu_userq_fence_driver_force_completion(queue);
  >> +                     drm_dev_wedged_event(adev_to_drm(adev), >> DRM_WEDGE_RECOVERY_NONE, NULL);
  >> +             }
  >>        } else {
  >>                gpu_reset = true;
  >>        }
  >
  > After the original queue was reset successfully, it did not update gpu_reset_counter, complete its pending fences, or send the wedged event. > mes_userq_reset_queue() already updates gpu_reset_counter, completes the pending fences, and sends the wedged event for the other affected queues,  > but skips the original queue because it has already been reset.

Quoted 7 lines

What is the rationale of sending multiple device wedged events on a per queue basis?

The question of whether drm_dev_wedged_event() should be emitted once per queue or once per overall recovery seems like a broader design discussion.

Along with that, also need to consider if device reset_counter needs to be incremented on a per queue basis or based on reset event recovery. It could get incremented multiple times inside this - mes_userq_reset_queue.

Quoted 9 lines
Quoted 2 lines

Looking at the current flow, both gpu_reset_counter and drm_dev_wedged_event() are updated once for each successfully reset queue. It would be helpful to clarify whether they are intended to be updated per affected queue or once per overall recovery.

What are the semantics around the reset counter and wedged events? Presumably each should be incremented for each queue that is reset? If a hang affects multiple queues shouldn't each be a separate "reset"? In the most common case, there should just be one since queue reset should be able to reset just the guilty queue.

Quoted 4 lines

Device wedged event is a udev event. It's not good to send  multiple events when multiple queues are recovered as part of single recovery process. Otherwise userspace monitoring the device will see back to back events. There is no context associated so it doesn't have any clue of what sort of reset happened.

I think the same goes with device. I believe it's to indicate a recovery event associated with device and not as a per queue event.

"I think the same goes with device" = device reset counter

Thanks,
Lijo

Quoted 19 lines

Thanks,
Lijo

Quoted 15 lines

Alex

Quoted 13 lines

Thanks,
Srini

Quoted 9 lines

Thanks,
Lijo

Quoted 5 lines

Thanks,
Srini