drm/amdgpu: Fix crash when hot unplug in BACO.

Commit Message

Andrey Grodzovsky May 21, 2021, 8:14 p.m. UTC

Problem:
When device goes into sleep state due to prolonged
innactivity (e.g. BACO sleep) and then hot unplugged,
PCI core will try to wake up the device as part of
unplug process. Since the device is gone all HW
programming during rpm resume fails leading
to a bad SW state later during pci remove handling.

Fix:
Use a flag we use for PCIe error recovery to avoid
accessing registres. This allows to succefully complete
rpm resume sequence and finish pci remove.

P.S Must use pci_device_is_present and not drm_dev_enter/exit
here since rpm resume happens before PCI remove and so the 
unplug flag is not set yet.

Link: https://gitlab.freedesktop.org/drm/amd/-/issues/1081
Signed-off-by: Andrey Grodzovsky <[e-mail]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 5 +++++
 1 file changed, 5 insertions(+)

Comments

Alex Deucher May 21, 2021, 8:18 p.m. UTC | #1

On Fri, May 21, 2021 at 4:14 PM Andrey Grodzovsky
<[e-mail]> wrote:
>
> Problem:
> When device goes into sleep state due to prolonged
> innactivity (e.g. BACO sleep) and then hot unplugged,
> PCI core will try to wake up the device as part of
> unplug process. Since the device is gone all HW
> programming during rpm resume fails leading
> to a bad SW state later during pci remove handling.
>
> Fix:
> Use a flag we use for PCIe error recovery to avoid
> accessing registres. This allows to succefully complete
> rpm resume sequence and finish pci remove.

Might make sense to create a preliminary patch to change the name of
this flag to something like no_hw_access since it's not specific to
pci error handling.

Alex

>
> P.S Must use pci_device_is_present and not drm_dev_enter/exit
> here since rpm resume happens before PCI remove and so the
> unplug flag is not set yet.
>
> Link: https://gitlab.freedesktop.org/drm/amd/-/issues/1081
> Signed-off-by: Andrey Grodzovsky <[e-mail]>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 5 +++++
>  1 file changed, 5 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> index d8db5929cdd9..ab95ebf56636 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
> @@ -1555,6 +1555,11 @@ static int amdgpu_pmops_runtime_resume(struct device *dev)
>         if (!adev->runpm)
>                 return -EINVAL;
>
> +       /* Avoids registers access if device is physically gone */
> +       if (!pci_device_is_present(adev->pdev))
> +               adev->in_pci_err_recovery = true;
> +
> +
>         if (amdgpu_device_supports_px(drm_dev)) {
>                 drm_dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
>
> --
> 2.25.1
>

Andrey Grodzovsky May 21, 2021, 8:37 p.m. UTC | #2

Will do.

Andrey

On 2021-05-21 4:18 p.m., Alex Deucher wrote:
> On Fri, May 21, 2021 at 4:14 PM Andrey Grodzovsky
> <[e-mail]> wrote:
>> Problem:
>> When device goes into sleep state due to prolonged
>> innactivity (e.g. BACO sleep) and then hot unplugged,
>> PCI core will try to wake up the device as part of
>> unplug process. Since the device is gone all HW
>> programming during rpm resume fails leading
>> to a bad SW state later during pci remove handling.
>>
>> Fix:
>> Use a flag we use for PCIe error recovery to avoid
>> accessing registres. This allows to succefully complete
>> rpm resume sequence and finish pci remove.
> Might make sense to create a preliminary patch to change the name of
> this flag to something like no_hw_access since it's not specific to
> pci error handling.
>
> Alex
>
>> P.S Must use pci_device_is_present and not drm_dev_enter/exit
>> here since rpm resume happens before PCI remove and so the
>> unplug flag is not set yet.
>>
>> Link: https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgitlab.freedesktop.org%2Fdrm%2Famd%2F-%2Fissues%2F1081&amp;data=04%7C01%7Candrey.grodzovsky%40amd.com%7C2a0ec02245b64de0139808d91c959987%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637572251118922092%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&amp;sdata=egcYBxBU%2BkIqbEdypVueXQcWb%2Bqe%2BKCC30Mw%2FjgR6ag%3D&amp;reserved=0
>> Signed-off-by: Andrey Grodzovsky <[e-mail]>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 5 +++++
>>   1 file changed, 5 insertions(+)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>> index d8db5929cdd9..ab95ebf56636 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>> @@ -1555,6 +1555,11 @@ static int amdgpu_pmops_runtime_resume(struct device *dev)
>>          if (!adev->runpm)
>>                  return -EINVAL;
>>
>> +       /* Avoids registers access if device is physically gone */
>> +       if (!pci_device_is_present(adev->pdev))
>> +               adev->in_pci_err_recovery = true;
>> +
>> +
>>          if (amdgpu_device_supports_px(drm_dev)) {
>>                  drm_dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
>>
>> --
>> 2.25.1
>>

Patch

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
index d8db5929cdd9..ab95ebf56636 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
@@ -1555,6 +1555,11 @@  static int amdgpu_pmops_runtime_resume(struct device *dev)
 	if (!adev->runpm)
 		return -EINVAL;
 
+	/* Avoids registers access if device is physically gone */
+	if (!pci_device_is_present(adev->pdev))
+		adev->in_pci_err_recovery = true;
+
+
 	if (amdgpu_device_supports_px(drm_dev)) {
 		drm_dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;