Skip to contentSkip to conversation

amdgpu/sriov Stop data exchange for wholegpu reset

PATCH In review Jack Zhang 3 participants 5 messages Jan 7, 2021 – Jan 12, 2021

Subsystems (2)
lore
Share & tools
lore

Jack Zhang

author
#0 lore raw

PATCH 1/1 amdgpu/sriov Stop data exchange for wholegpu reset

[Why]
When host trigger a whole gpu reset, guest will keep waiting till host finish reset. But there's a work queue in guest exchanging data between vf&pf which need to access frame buffer. During whole gpu reset, frame buffer is not accessable, and this causes the call trace.

[How]
After vf get reset notification from pf, stop data exchange.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c | 1 +
 drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c    | 1 +
 drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c    | 1 +
 3 files changed, 3 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c +1 -0
index 83ca5cbffe2c..3e212862cf5d 100644
@@ -571,6 +571,7 @@ void amdgpu_virt_fini_data_exchange(struct amdgpu_device *adev)
 		DRM_INFO("clean up the vf2pf work item\n");
 		flush_delayed_work(&adev->virt.vf2pf_work);
 		cancel_delayed_work_sync(&adev->virt.vf2pf_work);
+		adev->virt.vf2pf_update_interval_ms = 0;
 	}
 }
 
drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c +1 -0
index 7767ccca526b..3ee481557fc9 100644
@@ -255,6 +255,7 @@ static void xgpu_ai_mailbox_flr_work(struct work_struct *work)
 	if (!down_read_trylock(&adev->reset_sem))
 		return;
 
+	amdgpu_virt_fini_data_exchange(adev);
 	atomic_set(&adev->in_gpu_reset, 1);
 
 	do {
drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c +1 -0
index dd5c1e6ce009..48e588d3c409 100644
@@ -276,6 +276,7 @@ static void xgpu_nv_mailbox_flr_work(struct work_struct *work)
 	if (!down_read_trylock(&adev->reset_sem))
 		return;
 
+	amdgpu_virt_fini_data_exchange(adev);
 	atomic_set(&adev->in_gpu_reset, 1);
 
 	do {
Signature · 6 lines
2.25.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#1 lore raw

Ping

-----Original Message-----
From: Jack Zhang <[email protected]>
Sent: Thursday, January 7, 2021 6:47 PM
To: [email protected]
Cc: Zhang, Jack (Jian) <[email protected]>; Zhang, Jack (Jian) <[email protected]>; Chen, JingWen <[email protected]>
Subject: [PATCH] amdgpu/sriov Stop data exchange for wholegpu reset

[Why]
When host trigger a whole gpu reset, guest will keep waiting till host finish reset. But there's a work queue in guest exchanging data between vf&pf which need to access frame buffer. During whole gpu reset, frame buffer is not accessable, and this causes the call trace.

[How]
After vf get reset notification from pf, stop data exchange.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c | 1 +
 drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c    | 1 +
 drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c    | 1 +
 3 files changed, 3 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c +1 -0
index 83ca5cbffe2c..3e212862cf5d 100644
@@ -571,6 +571,7 @@ void amdgpu_virt_fini_data_exchange(struct amdgpu_device *adev)
 		DRM_INFO("clean up the vf2pf work item\n");
 		flush_delayed_work(&adev->virt.vf2pf_work);
 		cancel_delayed_work_sync(&adev->virt.vf2pf_work);
+		adev->virt.vf2pf_update_interval_ms = 0;
 	}
 }
 
drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c +1 -0
index 7767ccca526b..3ee481557fc9 100644
@@ -255,6 +255,7 @@ static void xgpu_ai_mailbox_flr_work(struct work_struct *work)
 	if (!down_read_trylock(&adev->reset_sem))
 		return;
 
+	amdgpu_virt_fini_data_exchange(adev);
 	atomic_set(&adev->in_gpu_reset, 1);
 
 	do {
drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c +1 -0
index dd5c1e6ce009..48e588d3c409 100644
@@ -276,6 +276,7 @@ static void xgpu_nv_mailbox_flr_work(struct work_struct *work)
 	if (!down_read_trylock(&adev->reset_sem))
 		return;
 
+	amdgpu_virt_fini_data_exchange(adev);
 	atomic_set(&adev->in_gpu_reset, 1);
 
 	do {

--
2.25.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#2 lore raw

[AMD Official Use Only - Internal Distribution Only]

Ping...

-----Original Message-----
From: Zhang, Jack (Jian)
Sent: Friday, January 8, 2021 11:07 AM
To: [email protected]; Liu, Monk <[email protected]>; Chen, JingWen <[email protected]>; Deucher, Alexander <[email protected]>; Deng, Emily <[email protected]>
Subject: RE: [PATCH] amdgpu/sriov Stop data exchange for wholegpu reset

Ping

-----Original Message-----
From: Jack Zhang <[email protected]>
Sent: Thursday, January 7, 2021 6:47 PM
To: [email protected]
Cc: Zhang, Jack (Jian) <[email protected]>; Zhang, Jack (Jian) <[email protected]>; Chen, JingWen <[email protected]>
Subject: [PATCH] amdgpu/sriov Stop data exchange for wholegpu reset

[Why]
When host trigger a whole gpu reset, guest will keep waiting till host finish reset. But there's a work queue in guest exchanging data between vf&pf which need to access frame buffer. During whole gpu reset, frame buffer is not accessable, and this causes the call trace.

[How]
After vf get reset notification from pf, stop data exchange.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c | 1 +
 drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c    | 1 +
 drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c    | 1 +
 3 files changed, 3 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c +1 -0
index 83ca5cbffe2c..3e212862cf5d 100644
@@ -571,6 +571,7 @@ void amdgpu_virt_fini_data_exchange(struct amdgpu_device *adev)
 DRM_INFO("clean up the vf2pf work item\n");
 flush_delayed_work(&adev->virt.vf2pf_work);
 cancel_delayed_work_sync(&adev->virt.vf2pf_work);
+adev->virt.vf2pf_update_interval_ms = 0;
 }
 }
 
drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c +1 -0
index 7767ccca526b..3ee481557fc9 100644
@@ -255,6 +255,7 @@ static void xgpu_ai_mailbox_flr_work(struct work_struct *work)
 if (!down_read_trylock(&adev->reset_sem))
 return;
 
+amdgpu_virt_fini_data_exchange(adev);
 atomic_set(&adev->in_gpu_reset, 1);
 
 do {
drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c +1 -0
index dd5c1e6ce009..48e588d3c409 100644
@@ -276,6 +276,7 @@ static void xgpu_nv_mailbox_flr_work(struct work_struct *work)
 if (!down_read_trylock(&adev->reset_sem))
 return;
 
+amdgpu_virt_fini_data_exchange(adev);
 atomic_set(&adev->in_gpu_reset, 1);
 
 do {

--
2.25.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Reviewed-by
#3 lore raw

[AMD Official Use Only - Internal Distribution Only]

Thanks

------------------------------------------
Monk Liu | Cloud-GPU Core team
------------------------------------------

-----Original Message-----
From: Zhang, Jack (Jian) <[email protected]>
Sent: Tuesday, January 12, 2021 11:20 AM
To: Zhang, Hawking <[email protected]>; [email protected]; Liu, Monk <[email protected]>; Chen, JingWen <[email protected]>; Deucher, Alexander <[email protected]>; Deng, Emily <[email protected]>
Subject: RE: [PATCH] amdgpu/sriov Stop data exchange for wholegpu reset

[AMD Official Use Only - Internal Distribution Only]

Ping...

-----Original Message-----
From: Zhang, Jack (Jian)
Sent: Friday, January 8, 2021 11:07 AM
To: [email protected]; Liu, Monk <[email protected]>; Chen, JingWen <[email protected]>; Deucher, Alexander <[email protected]>; Deng, Emily <[email protected]>
Subject: RE: [PATCH] amdgpu/sriov Stop data exchange for wholegpu reset

Ping

-----Original Message-----
From: Jack Zhang <[email protected]>
Sent: Thursday, January 7, 2021 6:47 PM
To: [email protected]
Cc: Zhang, Jack (Jian) <[email protected]>; Zhang, Jack (Jian) <[email protected]>; Chen, JingWen <[email protected]>
Subject: [PATCH] amdgpu/sriov Stop data exchange for wholegpu reset

[Why]
When host trigger a whole gpu reset, guest will keep waiting till host finish reset. But there's a work queue in guest exchanging data between vf&pf which need to access frame buffer. During whole gpu reset, frame buffer is not accessable, and this causes the call trace.

[How]
After vf get reset notification from pf, stop data exchange.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c | 1 +
 drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c    | 1 +
 drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c    | 1 +
 3 files changed, 3 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c +1 -0
index 83ca5cbffe2c..3e212862cf5d 100644
@@ -571,6 +571,7 @@ void amdgpu_virt_fini_data_exchange(struct amdgpu_device *adev)  DRM_INFO("clean up the vf2pf work item\n");  flush_delayed_work(&adev->virt.vf2pf_work);
 cancel_delayed_work_sync(&adev->virt.vf2pf_work);
+adev->virt.vf2pf_update_interval_ms = 0;
 }
 }
 
drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c +1 -0
index 7767ccca526b..3ee481557fc9 100644
@@ -255,6 +255,7 @@ static void xgpu_ai_mailbox_flr_work(struct work_struct *work)  if (!down_read_trylock(&adev->reset_sem))
 return;
 
+amdgpu_virt_fini_data_exchange(adev);
 atomic_set(&adev->in_gpu_reset, 1);
 
 do {
drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c +1 -1
index dd5c1e6ce009..48e588d3c409 100644
@@ -276,6 +276,7 @@ static void xgpu_nv_mailbox_flr_work(struct work_struct *work)  if (!down_read_trylock(&adev->reset_sem))
 return;
 
+amdgpu_virt_fini_data_exchange(adev);
 atomic_set(&adev->in_gpu_reset, 1);
 
 do {
--

2.25.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#4 lore raw
In reply to Jack Zhang

Dear Jack,

Thank you for your patch.

Please add a colon after amdgpu/sriov in the commit message summary.

Am 07.01.21 um 11:46 schrieb Jack Zhang:

Quoted 2 lines

[Why]
When host trigger a whole gpu reset, guest will keep

*hosts trigger* or *host triggers*

Quoted 1 line

waiting till host finish reset. But there's a work

finishes

Quoted 1 line

queue in guest exchanging data between vf&pf which need

needs

Quoted 2 lines

to access frame buffer. During whole gpu reset, frame buffer is not accessable, and this causes the call trace.

accessible (a spell checker should have caught that)

Can you please paste part of the trace, so it’s easily findable by users running into this.

Quoted 2 lines

[How]
After vf get reset notification from pf, stop data exchange.

How can this be reproduced and tested?

Kind regards,

Paul

Quoted 38 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c | 1 +
  drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c    | 1 +
  drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c    | 1 +
  3 files changed, 3 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c +1 -0
index 83ca5cbffe2c..3e212862cf5d 100644
@@ -571,6 +571,7 @@ void amdgpu_virt_fini_data_exchange(struct amdgpu_device *adev)
  		DRM_INFO("clean up the vf2pf work item\n");
  		flush_delayed_work(&adev->virt.vf2pf_work);
  		cancel_delayed_work_sync(&adev->virt.vf2pf_work);
+		adev->virt.vf2pf_update_interval_ms = 0;
  	}
  }
 
drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c +1 -0
index 7767ccca526b..3ee481557fc9 100644
@@ -255,6 +255,7 @@ static void xgpu_ai_mailbox_flr_work(struct work_struct *work)
  	if (!down_read_trylock(&adev->reset_sem))
  		return;
 
+	amdgpu_virt_fini_data_exchange(adev);
  	atomic_set(&adev->in_gpu_reset, 1);
 
  	do {
drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c +0 -0
index dd5c1e6ce009..48e588d3c409 100644

drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c static void xgpu_nv_mailbox_flr_work(struct work_struct *work) · line 282

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c +1 -0
@@ -276,6 +276,7 @@ static void xgpu_nv_mailbox_flr_work(struct work_struct *work)
  	if (!down_read_trylock(&adev->reset_sem))
  		return;
 
+	amdgpu_virt_fini_data_exchange(adev);
  	atomic_set(&adev->in_gpu_reset, 1);
 
  	do {

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx