[PATCH] drm/amd/display: Restore 5s vbl offdelay for NV3x+ DGPUs

[PATCH] drm/amd/display: Restore 5s vbl offdelay for NV3x+ DGPUs

[sunpeng.li]

From: Leo Li <sunpeng.li@amd.com>

[Why]

Rapid vblank off is causing flip-done timeouts for NV3x and newer
family of GPUs that support more idle optimization features.

A proper fix requires further investigation. In lieu of it, let's
workaround it for now.

[How]

For NV3x and newer family of DGPUs, restore the old 5s vblank off timer.

Fixes: 9b47278cec98 ("drm/amd/display: temp w/a for dGPU to enter idle optimizations")
Link: https://gitlab.freedesktop.org/drm/amd/-/issues/3787
Link: https://lore.kernel.org/amd-gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Tested-by: Michele Palazzi <sysdadmin@m1k.cloud>
---
 .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c  | 18 +++++++++++++++---
 1 file changed, 15 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
index 3fa4dbda4517c..ce5063928413c 100644
--- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
+++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
@@ -9511,9 +9511,21 @@ static void manage_dm_interrupts(struct amdgpu_device *adev,
 	if (acrtc_state) {
 		timing = &acrtc_state->stream->timing;
 
-		if (amdgpu_ip_version(adev, DCE_HWIP, 0) <
-			   IP_VERSION(3, 5, 0) ||
-			   !(adev->flags & AMD_IS_APU)) {
+		if (amdgpu_ip_version(adev, DCE_HWIP, 0) >=
+		      IP_VERSION(3, 2, 0) &&
+		      !(adev->flags & AMD_IS_APU)) {
+			/*
+			 * DGPUs NV3x and newer that support idle optimizations
+			 * experience intermittent flip-done timeouts on cursor
+			 * updates. Restore 5s offdelay behavior for now.
+			 *
+			 * Discussion on the issue:
+			 * https://lore.kernel.org/amd-gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/
+			 */
+			config.offdelay_ms = 5000;
+			config.disable_immediate = false;
+		} else if (amdgpu_ip_version(adev, DCE_HWIP, 0) <
+			     IP_VERSION(3, 5, 0)) {
 			/*
 			 * Older HW and DGPU have issues with instant off;
 			 * use a 2 frame offdelay.
-- 
2.53.0

Re: [PATCH] drm/amd/display: Restore 5s vbl offdelay for NV3x+ DGPUs

[Mario Limonciello]

On 4/22/26 11:29, sunpeng.li@amd.com wrote:
> From: Leo Li <sunpeng.li@amd.com>
> 
> [Why]
> 
> Rapid vblank off is causing flip-done timeouts for NV3x and newer
> family of GPUs that support more idle optimization features.
> 
> A proper fix requires further investigation. In lieu of it, let's
> workaround it for now.
> 
> [How]
> 
> For NV3x and newer family of DGPUs, restore the old 5s vblank off timer.

Why would it be dGPU only though?  I thought we have various reports of 
flip_done timeouts for APUs too.

Is your theory they're different root causes?

> 
> Fixes: 9b47278cec98 ("drm/amd/display: temp w/a for dGPU to enter idle optimizations")
> Link: https://gitlab.freedesktop.org/drm/amd/-/issues/3787
> Link: https://lore.kernel.org/amd-gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/
> Signed-off-by: Leo Li <sunpeng.li@amd.com>
> Tested-by: Michele Palazzi <sysdadmin@m1k.cloud>
> ---
>   .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c  | 18 +++++++++++++++---
>   1 file changed, 15 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
> index 3fa4dbda4517c..ce5063928413c 100644
> --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
> +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
> @@ -9511,9 +9511,21 @@ static void manage_dm_interrupts(struct amdgpu_device *adev,
>   	if (acrtc_state) {
>   		timing = &acrtc_state->stream->timing;
>   
> -		if (amdgpu_ip_version(adev, DCE_HWIP, 0) <
> -			   IP_VERSION(3, 5, 0) ||
> -			   !(adev->flags & AMD_IS_APU)) {
> +		if (amdgpu_ip_version(adev, DCE_HWIP, 0) >=
> +		      IP_VERSION(3, 2, 0) &&
> +		      !(adev->flags & AMD_IS_APU)) {
> +			/*
> +			 * DGPUs NV3x and newer that support idle optimizations
> +			 * experience intermittent flip-done timeouts on cursor
> +			 * updates. Restore 5s offdelay behavior for now.
> +			 *
> +			 * Discussion on the issue:
> +			 * https://lore.kernel.org/amd-gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/
> +			 */
> +			config.offdelay_ms = 5000;
> +			config.disable_immediate = false;
> +		} else if (amdgpu_ip_version(adev, DCE_HWIP, 0) <
> +			     IP_VERSION(3, 5, 0)) {
>   			/*
>   			 * Older HW and DGPU have issues with instant off;
>   			 * use a 2 frame offdelay.

Re: [PATCH] drm/amd/display: Restore 5s vbl offdelay for NV3x+ DGPUs

[Alex Deucher]

On Wed, Apr 22, 2026 at 12:49 PM <sunpeng.li@amd.com> wrote:
>
> From: Leo Li <sunpeng.li@amd.com>
>
> [Why]
>
> Rapid vblank off is causing flip-done timeouts for NV3x and newer
> family of GPUs that support more idle optimization features.
>
> A proper fix requires further investigation. In lieu of it, let's
> workaround it for now.
>
> [How]
>
> For NV3x and newer family of DGPUs, restore the old 5s vblank off timer.
>
> Fixes: 9b47278cec98 ("drm/amd/display: temp w/a for dGPU to enter idle optimizations")
> Link: https://gitlab.freedesktop.org/drm/amd/-/issues/3787
> Link: https://lore.kernel.org/amd-gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/
> Signed-off-by: Leo Li <sunpeng.li@amd.com>
> Tested-by: Michele Palazzi <sysdadmin@m1k.cloud>
> ---
>  .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c  | 18 +++++++++++++++---
>  1 file changed, 15 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
> index 3fa4dbda4517c..ce5063928413c 100644
> --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
> +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
> @@ -9511,9 +9511,21 @@ static void manage_dm_interrupts(struct amdgpu_device *adev,
>         if (acrtc_state) {
>                 timing = &acrtc_state->stream->timing;
>
> -               if (amdgpu_ip_version(adev, DCE_HWIP, 0) <
> -                          IP_VERSION(3, 5, 0) ||
> -                          !(adev->flags & AMD_IS_APU)) {
> +               if (amdgpu_ip_version(adev, DCE_HWIP, 0) >=
> +                     IP_VERSION(3, 2, 0) &&
> +                     !(adev->flags & AMD_IS_APU)) {

Why only dGPUs?  Seems like this is reported as least as often on APUs
if not more.

Alex

> +                       /*
> +                        * DGPUs NV3x and newer that support idle optimizations
> +                        * experience intermittent flip-done timeouts on cursor
> +                        * updates. Restore 5s offdelay behavior for now.
> +                        *
> +                        * Discussion on the issue:
> +                        * https://lore.kernel.org/amd-gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/
> +                        */
> +                       config.offdelay_ms = 5000;
> +                       config.disable_immediate = false;
> +               } else if (amdgpu_ip_version(adev, DCE_HWIP, 0) <
> +                            IP_VERSION(3, 5, 0)) {
>                         /*
>                          * Older HW and DGPU have issues with instant off;
>                          * use a 2 frame offdelay.
> --
> 2.53.0
>

Re: [PATCH] drm/amd/display: Restore 5s vbl offdelay for NV3x+ DGPUs

[Leo Li]

On 2026-04-22 12:56, Alex Deucher wrote:
> On Wed, Apr 22, 2026 at 12:49 PM <sunpeng.li@amd.com> wrote:
>>
>> From: Leo Li <sunpeng.li@amd.com>
>>
>> [Why]
>>
>> Rapid vblank off is causing flip-done timeouts for NV3x and newer
>> family of GPUs that support more idle optimization features.
>>
>> A proper fix requires further investigation. In lieu of it, let's
>> workaround it for now.
>>
>> [How]
>>
>> For NV3x and newer family of DGPUs, restore the old 5s vblank off timer.
>>
>> Fixes: 9b47278cec98 ("drm/amd/display: temp w/a for dGPU to enter idle optimizations")
>> Link: https://gitlab.freedesktop.org/drm/amd/-/issues/3787
>> Link: https://lore.kernel.org/amd-gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/
>> Signed-off-by: Leo Li <sunpeng.li@amd.com>
>> Tested-by: Michele Palazzi <sysdadmin@m1k.cloud>
>> ---
>>  .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c  | 18 +++++++++++++++---
>>  1 file changed, 15 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
>> index 3fa4dbda4517c..ce5063928413c 100644
>> --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
>> +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
>> @@ -9511,9 +9511,21 @@ static void manage_dm_interrupts(struct amdgpu_device *adev,
>>         if (acrtc_state) {
>>                 timing = &acrtc_state->stream->timing;
>>
>> -               if (amdgpu_ip_version(adev, DCE_HWIP, 0) <
>> -                          IP_VERSION(3, 5, 0) ||
>> -                          !(adev->flags & AMD_IS_APU)) {
>> +               if (amdgpu_ip_version(adev, DCE_HWIP, 0) >=
>> +                     IP_VERSION(3, 2, 0) &&
>> +                     !(adev->flags & AMD_IS_APU)) {
> 
> Why only dGPUs?  Seems like this is reported as least as often on APUs
> if not more.
> 
> Alex

Hi Alex, Mario,

At least in the case of the few reporters I was working with, this specific
flip-done timeout was reproduced on NV3x and 4x systems running multi-display.
The reporter for the linked gitlab issue was also running a nv3.

The cause of these flip timeouts can be varied. The signature for this
particular issue was OTG failing to fire an interrupt that is expected to
deliver the flip-done event. I'm not aware of this particular signature in APUs
-- at least none on my radar. Do bring it to my attention if you're aware of
them.

Thanks,
Leo

> 
>> +                       /*
>> +                        * DGPUs NV3x and newer that support idle optimizations
>> +                        * experience intermittent flip-done timeouts on cursor
>> +                        * updates. Restore 5s offdelay behavior for now.
>> +                        *
>> +                        * Discussion on the issue:
>> +                        * https://lore.kernel.org/amd-gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/
>> +                        */
>> +                       config.offdelay_ms = 5000;
>> +                       config.disable_immediate = false;
>> +               } else if (amdgpu_ip_version(adev, DCE_HWIP, 0) <
>> +                            IP_VERSION(3, 5, 0)) {
>>                         /*
>>                          * Older HW and DGPU have issues with instant off;
>>                          * use a 2 frame offdelay.
>> --
>> 2.53.0
>>

Re: [PATCH] drm/amd/display: Restore 5s vbl offdelay for NV3x+ DGPUs

[Mario Limonciello]

On 4/22/26 12:19, Leo Li wrote:
> 
> 
> On 2026-04-22 12:56, Alex Deucher wrote:
>> On Wed, Apr 22, 2026 at 12:49 PM <sunpeng.li@amd.com> wrote:
>>>
>>> From: Leo Li <sunpeng.li@amd.com>
>>>
>>> [Why]
>>>
>>> Rapid vblank off is causing flip-done timeouts for NV3x and newer
>>> family of GPUs that support more idle optimization features.
>>>
>>> A proper fix requires further investigation. In lieu of it, let's
>>> workaround it for now.
>>>
>>> [How]
>>>
>>> For NV3x and newer family of DGPUs, restore the old 5s vblank off timer.
>>>
>>> Fixes: 9b47278cec98 ("drm/amd/display: temp w/a for dGPU to enter idle optimizations")
>>> Link: https://gitlab.freedesktop.org/drm/amd/-/issues/3787
>>> Link: https://lore.kernel.org/amd-gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/
>>> Signed-off-by: Leo Li <sunpeng.li@amd.com>
>>> Tested-by: Michele Palazzi <sysdadmin@m1k.cloud>
>>> ---
>>>   .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c  | 18 +++++++++++++++---
>>>   1 file changed, 15 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
>>> index 3fa4dbda4517c..ce5063928413c 100644
>>> --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
>>> +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
>>> @@ -9511,9 +9511,21 @@ static void manage_dm_interrupts(struct amdgpu_device *adev,
>>>          if (acrtc_state) {
>>>                  timing = &acrtc_state->stream->timing;
>>>
>>> -               if (amdgpu_ip_version(adev, DCE_HWIP, 0) <
>>> -                          IP_VERSION(3, 5, 0) ||
>>> -                          !(adev->flags & AMD_IS_APU)) {
>>> +               if (amdgpu_ip_version(adev, DCE_HWIP, 0) >=
>>> +                     IP_VERSION(3, 2, 0) &&
>>> +                     !(adev->flags & AMD_IS_APU)) {
>>
>> Why only dGPUs?  Seems like this is reported as least as often on APUs
>> if not more.
>>
>> Alex
> 
> Hi Alex, Mario,
> 
> At least in the case of the few reporters I was working with, this specific
> flip-done timeout was reproduced on NV3x and 4x systems running multi-display.
> The reporter for the linked gitlab issue was also running a nv3.
> 
> The cause of these flip timeouts can be varied. The signature for this
> particular issue was OTG failing to fire an interrupt that is expected to
> deliver the flip-done event. I'm not aware of this particular signature in APUs
> -- at least none on my radar. Do bring it to my attention if you're aware of
> them.
> 
> Thanks,
> Leo

In Michele's proposal 
(https://lore.kernel.org/amd-gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/) 
there was a mention that it was tested on DCN 3.5 too, which made me 
think that the exact same issue was on both.

Michele - can you readily reproduce the page flip timeout on DCN 3.5?

If so; could you modify Leo's patch to drop the IS_APU designation and 
see if it happens to be the same solution?

> 
>>
>>> +                       /*
>>> +                        * DGPUs NV3x and newer that support idle optimizations
>>> +                        * experience intermittent flip-done timeouts on cursor
>>> +                        * updates. Restore 5s offdelay behavior for now.
>>> +                        *
>>> +                        * Discussion on the issue:
>>> +                        * https://lore.kernel.org/amd-gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/
>>> +                        */
>>> +                       config.offdelay_ms = 5000;
>>> +                       config.disable_immediate = false;
>>> +               } else if (amdgpu_ip_version(adev, DCE_HWIP, 0) <
>>> +                            IP_VERSION(3, 5, 0)) {
>>>                          /*
>>>                           * Older HW and DGPU have issues with instant off;
>>>                           * use a 2 frame offdelay.
>>> --
>>> 2.53.0
>>>
> 

Re: [PATCH] drm/amd/display: Restore 5s vbl offdelay for NV3x+ DGPUs

[Michele Palazzi]

On 4/22/26 19:42, Mario Limonciello wrote:
> 
> In Michele's proposal (https://lore.kernel.org/amd- 
> gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/) there was a mention 
> that it was tested on DCN 3.5 too, which made me think that the exact 
> same issue was on both.
> 
> Michele - can you readily reproduce the page flip timeout on DCN 3.5?
> 
> If so; could you modify Leo's patch to drop the IS_APU designation and 
> see if it happens to be the same solution?
> 
>>
>>>
>>>> +                       /*
>>>> +                        * DGPUs NV3x and newer that support idle 
>>>> optimizations
>>>> +                        * experience intermittent flip-done 
>>>> timeouts on cursor
>>>> +                        * updates. Restore 5s offdelay behavior for 
>>>> now.
>>>> +                        *
>>>> +                        * Discussion on the issue:
>>>> +                        * https://lore.kernel.org/amd- 
>>>> gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/
>>>> +                        */
>>>> +                       config.offdelay_ms = 5000;
>>>> +                       config.disable_immediate = false;
>>>> +               } else if (amdgpu_ip_version(adev, DCE_HWIP, 0) <
>>>> +                            IP_VERSION(3, 5, 0)) {
>>>>                          /*
>>>>                           * Older HW and DGPU have issues with 
>>>> instant off;
>>>>                           * use a 2 frame offdelay.
>>>> -- 
>>>> 2.53.0
>>>>
>>
> 

Hi Mario, i had tested my proposed patch on multiple APUs in order to 
exclude regressions or side effects, but personally i only ever 
encountered this particular issue on dGPUs, specifically a 7900 GRE 
first and a 9070XT later.

Re: [PATCH] drm/amd/display: Restore 5s vbl offdelay for NV3x+ DGPUs

[Mario Limonciello]

On 4/22/26 12:50, Michele Palazzi wrote:
> On 4/22/26 19:42, Mario Limonciello wrote:
>>
>> In Michele's proposal (https://lore.kernel.org/amd- 
>> gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/) there was a mention 
>> that it was tested on DCN 3.5 too, which made me think that the exact 
>> same issue was on both.
>>
>> Michele - can you readily reproduce the page flip timeout on DCN 3.5?
>>
>> If so; could you modify Leo's patch to drop the IS_APU designation and 
>> see if it happens to be the same solution?
>>
>>>
>>>>
>>>>> +                       /*
>>>>> +                        * DGPUs NV3x and newer that support idle 
>>>>> optimizations
>>>>> +                        * experience intermittent flip-done 
>>>>> timeouts on cursor
>>>>> +                        * updates. Restore 5s offdelay behavior 
>>>>> for now.
>>>>> +                        *
>>>>> +                        * Discussion on the issue:
>>>>> +                        * https://lore.kernel.org/amd- 
>>>>> gfx/20260217191632.1243826-1-sysdadmin@m1k.cloud/
>>>>> +                        */
>>>>> +                       config.offdelay_ms = 5000;
>>>>> +                       config.disable_immediate = false;
>>>>> +               } else if (amdgpu_ip_version(adev, DCE_HWIP, 0) <
>>>>> +                            IP_VERSION(3, 5, 0)) {
>>>>>                          /*
>>>>>                           * Older HW and DGPU have issues with 
>>>>> instant off;
>>>>>                           * use a 2 frame offdelay.
>>>>> -- 
>>>>> 2.53.0
>>>>>
>>>
>>
> 
> Hi Mario, i had tested my proposed patch on multiple APUs in order to 
> exclude regressions or side effects, but personally i only ever 
> encountered this particular issue on dGPUs, specifically a 7900 GRE 
> first and a 9070XT later.
> 

Got it - thanks for clarifying and reaffirming Leo's assertion was 
correct on root cause.

Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>