[PATCH] drm/amdgpu: add missed write lock for pci detected state pci_channel_io_normal

[PATCH] drm/amdgpu: add missed write lock for pci detected state pci_channel_io_normal

[Guchun Chen]

When a PCI error state pci_channel_io_normal is detectd, it will
report PCI_ERS_RESULT_CAN_RECOVER status to PCI driver, and PCI driver
will continue the execution of PCI resume callback report_resume by
pci_walk_bridge, and the callback will go into amdgpu_pci_resume
finally, where write lock is releasd unconditionally without acquiring
such lock.

Fixes: c9a6b82f45e2("drm/amdgpu: Implement DPC recovery")
Signed-off-by: Guchun Chen <guchun.chen@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index bb5ad2b6ca13..12f822d51de2 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -5370,6 +5370,7 @@ pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_sta
 
 	switch (state) {
 	case pci_channel_io_normal:
+		amdgpu_device_lock_adev(adev, NULL);
 		return PCI_ERS_RESULT_CAN_RECOVER;
 	/* Fatal error, prepare for slot reset */
 	case pci_channel_io_frozen:
-- 
2.17.1

Re: [PATCH] drm/amdgpu: add missed write lock for pci detected state pci_channel_io_normal

[Andrey Grodzovsky]

On 2021-09-30 10:00 p.m., Guchun Chen wrote:

> When a PCI error state pci_channel_io_normal is detectd, it will
> report PCI_ERS_RESULT_CAN_RECOVER status to PCI driver, and PCI driver
> will continue the execution of PCI resume callback report_resume by
> pci_walk_bridge, and the callback will go into amdgpu_pci_resume
> finally, where write lock is releasd unconditionally without acquiring
> such lock.


Good catch but, the issue is even wider in scope, what about 
drm_sched_resubmit_jobs
and drm_sched_start called without being stopped before ? Better to put 
the entire scope
of code in this function under flag that set only in 
pci_channel_io_frozen. As far as i remember
we don't need to do anything in case of pci_channel_io_normal.

Andrey


>
> Fixes: c9a6b82f45e2("drm/amdgpu: Implement DPC recovery")
> Signed-off-by: Guchun Chen <guchun.chen@amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 1 +
>   1 file changed, 1 insertion(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index bb5ad2b6ca13..12f822d51de2 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -5370,6 +5370,7 @@ pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_sta
>   
>   	switch (state) {
>   	case pci_channel_io_normal:
> +		amdgpu_device_lock_adev(adev, NULL);
>   		return PCI_ERS_RESULT_CAN_RECOVER;
>   	/* Fatal error, prepare for slot reset */
>   	case pci_channel_io_frozen:

RE: [PATCH] drm/amdgpu: add missed write lock for pci detected state pci_channel_io_normal

[Chen, Guchun]

[Public]

Hi Andrey,

Do you mean to move the code of drm_sched_resubmit_jobs and drm_sched_start in amdgpu_pci_resume to amdgpu_pci_error_detected, under the case pci_channel_io_frozen?
Then leave amdgpu_pci_resume as a null function, and in this way, we can drop the acquire/lock write lock for case of pci_channel_io_normal as well?

Regards,
Guchun

-----Original Message-----
From: Grodzovsky, Andrey <Andrey.Grodzovsky@amd.com> 
Sent: Friday, October 1, 2021 10:22 AM
To: Chen, Guchun <Guchun.Chen@amd.com>; amd-gfx@lists.freedesktop.org; Koenig, Christian <Christian.Koenig@amd.com>; Pan, Xinhui <Xinhui.Pan@amd.com>; Deucher, Alexander <Alexander.Deucher@amd.com>
Subject: Re: [PATCH] drm/amdgpu: add missed write lock for pci detected state pci_channel_io_normal

On 2021-09-30 10:00 p.m., Guchun Chen wrote:

> When a PCI error state pci_channel_io_normal is detectd, it will 
> report PCI_ERS_RESULT_CAN_RECOVER status to PCI driver, and PCI driver 
> will continue the execution of PCI resume callback report_resume by 
> pci_walk_bridge, and the callback will go into amdgpu_pci_resume 
> finally, where write lock is releasd unconditionally without acquiring 
> such lock.


Good catch but, the issue is even wider in scope, what about drm_sched_resubmit_jobs and drm_sched_start called without being stopped before ? Better to put the entire scope of code in this function under flag that set only in pci_channel_io_frozen. As far as i remember we don't need to do anything in case of pci_channel_io_normal.

Andrey


>
> Fixes: c9a6b82f45e2("drm/amdgpu: Implement DPC recovery")
> Signed-off-by: Guchun Chen <guchun.chen@amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 1 +
>   1 file changed, 1 insertion(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index bb5ad2b6ca13..12f822d51de2 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -5370,6 +5370,7 @@ pci_ers_result_t 
> amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_sta
>   
>   	switch (state) {
>   	case pci_channel_io_normal:
> +		amdgpu_device_lock_adev(adev, NULL);
>   		return PCI_ERS_RESULT_CAN_RECOVER;
>   	/* Fatal error, prepare for slot reset */
>   	case pci_channel_io_frozen:

Re: [PATCH] drm/amdgpu: add missed write lock for pci detected state pci_channel_io_normal

[Andrey Grodzovsky]

No, scheduler restart and device unlock must take place 
inamdgpu_pci_resume (see struct pci_error_handlers for the various 
states of PCI recovery). So just add a flag (probably in amdgpu_device) 
so we can remember what pci_channel_state_t we came from (unfortunately 
it's not passed to us in  amdgpu_pci_resume) and unless it's set don't 
do anything in amdgpu_pci_resume.

Andrey

On 2021-10-01 4:21 a.m., Chen, Guchun wrote:
> [Public]
>
> Hi Andrey,
>
> Do you mean to move the code of drm_sched_resubmit_jobs and drm_sched_start in amdgpu_pci_resume to amdgpu_pci_error_detected, under the case pci_channel_io_frozen?
> Then leave amdgpu_pci_resume as a null function, and in this way, we can drop the acquire/lock write lock for case of pci_channel_io_normal as well?
>
> Regards,
> Guchun
>
> -----Original Message-----
> From: Grodzovsky, Andrey <Andrey.Grodzovsky@amd.com>
> Sent: Friday, October 1, 2021 10:22 AM
> To: Chen, Guchun <Guchun.Chen@amd.com>; amd-gfx@lists.freedesktop.org; Koenig, Christian <Christian.Koenig@amd.com>; Pan, Xinhui <Xinhui.Pan@amd.com>; Deucher, Alexander <Alexander.Deucher@amd.com>
> Subject: Re: [PATCH] drm/amdgpu: add missed write lock for pci detected state pci_channel_io_normal
>
> On 2021-09-30 10:00 p.m., Guchun Chen wrote:
>
>> When a PCI error state pci_channel_io_normal is detectd, it will
>> report PCI_ERS_RESULT_CAN_RECOVER status to PCI driver, and PCI driver
>> will continue the execution of PCI resume callback report_resume by
>> pci_walk_bridge, and the callback will go into amdgpu_pci_resume
>> finally, where write lock is releasd unconditionally without acquiring
>> such lock.
>
> Good catch but, the issue is even wider in scope, what about drm_sched_resubmit_jobs and drm_sched_start called without being stopped before ? Better to put the entire scope of code in this function under flag that set only in pci_channel_io_frozen. As far as i remember we don't need to do anything in case of pci_channel_io_normal.
>
> Andrey
>
>
>> Fixes: c9a6b82f45e2("drm/amdgpu: Implement DPC recovery")
>> Signed-off-by: Guchun Chen <guchun.chen@amd.com>
>> ---
>>    drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 1 +
>>    1 file changed, 1 insertion(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> index bb5ad2b6ca13..12f822d51de2 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> @@ -5370,6 +5370,7 @@ pci_ers_result_t
>> amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_sta
>>    
>>    	switch (state) {
>>    	case pci_channel_io_normal:
>> +		amdgpu_device_lock_adev(adev, NULL);
>>    		return PCI_ERS_RESULT_CAN_RECOVER;
>>    	/* Fatal error, prepare for slot reset */
>>    	case pci_channel_io_frozen:

RE: [PATCH] drm/amdgpu: add missed write lock for pci detected state pci_channel_io_normal

[Chen, Guchun]

[Public]

Got your point. Will send a new patch to address this.

Regards,
Guchun

-----Original Message-----
From: Grodzovsky, Andrey <Andrey.Grodzovsky@amd.com> 
Sent: Friday, October 1, 2021 10:29 PM
To: Chen, Guchun <Guchun.Chen@amd.com>; amd-gfx@lists.freedesktop.org; Koenig, Christian <Christian.Koenig@amd.com>; Pan, Xinhui <Xinhui.Pan@amd.com>; Deucher, Alexander <Alexander.Deucher@amd.com>
Subject: Re: [PATCH] drm/amdgpu: add missed write lock for pci detected state pci_channel_io_normal

No, scheduler restart and device unlock must take place inamdgpu_pci_resume (see struct pci_error_handlers for the various states of PCI recovery). So just add a flag (probably in amdgpu_device) so we can remember what pci_channel_state_t we came from (unfortunately it's not passed to us in  amdgpu_pci_resume) and unless it's set don't do anything in amdgpu_pci_resume.

Andrey

On 2021-10-01 4:21 a.m., Chen, Guchun wrote:
> [Public]
>
> Hi Andrey,
>
> Do you mean to move the code of drm_sched_resubmit_jobs and drm_sched_start in amdgpu_pci_resume to amdgpu_pci_error_detected, under the case pci_channel_io_frozen?
> Then leave amdgpu_pci_resume as a null function, and in this way, we can drop the acquire/lock write lock for case of pci_channel_io_normal as well?
>
> Regards,
> Guchun
>
> -----Original Message-----
> From: Grodzovsky, Andrey <Andrey.Grodzovsky@amd.com>
> Sent: Friday, October 1, 2021 10:22 AM
> To: Chen, Guchun <Guchun.Chen@amd.com>; amd-gfx@lists.freedesktop.org; 
> Koenig, Christian <Christian.Koenig@amd.com>; Pan, Xinhui 
> <Xinhui.Pan@amd.com>; Deucher, Alexander <Alexander.Deucher@amd.com>
> Subject: Re: [PATCH] drm/amdgpu: add missed write lock for pci 
> detected state pci_channel_io_normal
>
> On 2021-09-30 10:00 p.m., Guchun Chen wrote:
>
>> When a PCI error state pci_channel_io_normal is detectd, it will 
>> report PCI_ERS_RESULT_CAN_RECOVER status to PCI driver, and PCI 
>> driver will continue the execution of PCI resume callback 
>> report_resume by pci_walk_bridge, and the callback will go into 
>> amdgpu_pci_resume finally, where write lock is releasd 
>> unconditionally without acquiring such lock.
>
> Good catch but, the issue is even wider in scope, what about drm_sched_resubmit_jobs and drm_sched_start called without being stopped before ? Better to put the entire scope of code in this function under flag that set only in pci_channel_io_frozen. As far as i remember we don't need to do anything in case of pci_channel_io_normal.
>
> Andrey
>
>
>> Fixes: c9a6b82f45e2("drm/amdgpu: Implement DPC recovery")
>> Signed-off-by: Guchun Chen <guchun.chen@amd.com>
>> ---
>>    drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 1 +
>>    1 file changed, 1 insertion(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> index bb5ad2b6ca13..12f822d51de2 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> @@ -5370,6 +5370,7 @@ pci_ers_result_t 
>> amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_sta
>>    
>>    	switch (state) {
>>    	case pci_channel_io_normal:
>> +		amdgpu_device_lock_adev(adev, NULL);
>>    		return PCI_ERS_RESULT_CAN_RECOVER;
>>    	/* Fatal error, prepare for slot reset */
>>    	case pci_channel_io_frozen:

RE: [PATCH] drm/amdgpu: add missed write lock for pci detected state pci_channel_io_normal

[Chen, Guchun]

[Public]

Hi Andrey,

A new patch with subject "drm/amdgpu: handle the case of pci_channel_io_frozen only in amdgpu_pci_resume" has been sent, pls review it. Thanks.

Regards,
Guchun

-----Original Message-----
From: Chen, Guchun 
Sent: Friday, October 1, 2021 11:21 PM
To: Grodzovsky, Andrey <Andrey.Grodzovsky@amd.com>; amd-gfx@lists.freedesktop.org; Koenig, Christian <Christian.Koenig@amd.com>; Pan, Xinhui <Xinhui.Pan@amd.com>; Deucher, Alexander <Alexander.Deucher@amd.com>
Subject: RE: [PATCH] drm/amdgpu: add missed write lock for pci detected state pci_channel_io_normal

[Public]

Got your point. Will send a new patch to address this.

Regards,
Guchun

-----Original Message-----
From: Grodzovsky, Andrey <Andrey.Grodzovsky@amd.com>
Sent: Friday, October 1, 2021 10:29 PM
To: Chen, Guchun <Guchun.Chen@amd.com>; amd-gfx@lists.freedesktop.org; Koenig, Christian <Christian.Koenig@amd.com>; Pan, Xinhui <Xinhui.Pan@amd.com>; Deucher, Alexander <Alexander.Deucher@amd.com>
Subject: Re: [PATCH] drm/amdgpu: add missed write lock for pci detected state pci_channel_io_normal

No, scheduler restart and device unlock must take place inamdgpu_pci_resume (see struct pci_error_handlers for the various states of PCI recovery). So just add a flag (probably in amdgpu_device) so we can remember what pci_channel_state_t we came from (unfortunately it's not passed to us in  amdgpu_pci_resume) and unless it's set don't do anything in amdgpu_pci_resume.

Andrey

On 2021-10-01 4:21 a.m., Chen, Guchun wrote:
> [Public]
>
> Hi Andrey,
>
> Do you mean to move the code of drm_sched_resubmit_jobs and drm_sched_start in amdgpu_pci_resume to amdgpu_pci_error_detected, under the case pci_channel_io_frozen?
> Then leave amdgpu_pci_resume as a null function, and in this way, we can drop the acquire/lock write lock for case of pci_channel_io_normal as well?
>
> Regards,
> Guchun
>
> -----Original Message-----
> From: Grodzovsky, Andrey <Andrey.Grodzovsky@amd.com>
> Sent: Friday, October 1, 2021 10:22 AM
> To: Chen, Guchun <Guchun.Chen@amd.com>; amd-gfx@lists.freedesktop.org; 
> Koenig, Christian <Christian.Koenig@amd.com>; Pan, Xinhui 
> <Xinhui.Pan@amd.com>; Deucher, Alexander <Alexander.Deucher@amd.com>
> Subject: Re: [PATCH] drm/amdgpu: add missed write lock for pci 
> detected state pci_channel_io_normal
>
> On 2021-09-30 10:00 p.m., Guchun Chen wrote:
>
>> When a PCI error state pci_channel_io_normal is detectd, it will 
>> report PCI_ERS_RESULT_CAN_RECOVER status to PCI driver, and PCI 
>> driver will continue the execution of PCI resume callback 
>> report_resume by pci_walk_bridge, and the callback will go into 
>> amdgpu_pci_resume finally, where write lock is releasd 
>> unconditionally without acquiring such lock.
>
> Good catch but, the issue is even wider in scope, what about drm_sched_resubmit_jobs and drm_sched_start called without being stopped before ? Better to put the entire scope of code in this function under flag that set only in pci_channel_io_frozen. As far as i remember we don't need to do anything in case of pci_channel_io_normal.
>
> Andrey
>
>
>> Fixes: c9a6b82f45e2("drm/amdgpu: Implement DPC recovery")
>> Signed-off-by: Guchun Chen <guchun.chen@amd.com>
>> ---
>>    drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 1 +
>>    1 file changed, 1 insertion(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> index bb5ad2b6ca13..12f822d51de2 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> @@ -5370,6 +5370,7 @@ pci_ers_result_t 
>> amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_sta
>>    
>>    	switch (state) {
>>    	case pci_channel_io_normal:
>> +		amdgpu_device_lock_adev(adev, NULL);
>>    		return PCI_ERS_RESULT_CAN_RECOVER;
>>    	/* Fatal error, prepare for slot reset */
>>    	case pci_channel_io_frozen: