On 9/22/26 11:23 AM, Matthew Schwartz wrote:
> On 9/22/26 11:10 AM, Leo Li wrote:
>>
>>
>> On 2026-09-22 14:04, Matthew Schwartz wrote:
>>> On 9/22/26 10:50 AM, Leo Li wrote:
>>>>
>>>>
>>>> On 2026-09-21 15:47, Matthew Schwartz wrote:
>>>>> Hi George,
>>>>>
>>>>> After updating to Linux 7.2.x, I noticed gamescope's frame limiter falling
>>>>> below the requested rate. The reproducer was a Navi 33 driving 4K120
>>>>> through a DP-HDMI PCON, with VRR disabled. A 40fps limit was producing
>>>>> roughly 37fps.
>>>>>
>>>>> Gamescope was pacing against the fixed 120Hz refresh interval, but we
>>>>> observed frame intervals stretching to around 9.34ms instead of 8.33ms.
>>>>> The OTG vtotal also increased from 2249 to 2521. This was enough to
>>>>> disrupt the limiter's pacing despite userspace leaving VRR_ENABLED=0.
>>>>
>>>> Hi Matt,
>>>>
>>>> I wonder if this has to do with re-introducing 2-frame vblank off for NV3+
>>>> DGPUs, which made it into 7.2:
>>>>
>>>> fba211b078d6 ("Revert "drm/amd/display: Restore 5s vbl offdelay for NV3x+
>>>> DGPUs"")
>>>>
>>>> Since gamescope is only updating every 3rd frame, it's possible vblanks
>>>> were
>>>> turned off during those two frames, signaling driver to enable idle
>>>> optimizations.
>>>>
>>>> To test this idea, does bumping this line to something like 5 frames
>>>> ((u64)50 *) help?
>>>> https://elixir.bootlin.com/linux/v7.3-rc3/source/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c#L3377
>>>
>>> Tested this by reverting my own patch in the kernel and replacing it with
>>> this, and unfortunately the issue
>>> still reproduces here. In Hades 2, a 40FPS limit is locking to 37FPS
>>> erroneously.
>>
>> Interesting... Maybe try reverting fba211b078d6 completely as well.
>
> No change, still locks to 37FPS.
>
>>
>> If that doesn't help, could you try isolating FPO by force disabling it
>> (without applying this fix, of course)?
>> Set `disable_fpo_optimizations = true` in both dcn32_resource.c and
>> dcn321_resource.c
>
> Added this change in addition to reverting fba211b078d6, and it now locks to
> 40FPS rather than 37FPS.
Actually, 40FPS works but 30FPS locks to 28FPS rather than 30FPS still, so it
appears it doesn't fully resolve the issue.
>
>>
>> - Leo
>>
>>>
>>> Matt
>>>
>>>>
>>>> Thanks,
>>>> Leo
>>>>>
>>>>> Tracing this led to allow_freesync remaining true for a VRR-capable
>>>>> sink even when the FreeSync state was inactive. That allowed FPO to
>>>>> stretch vblank around UCLK switches. The PCON whitelist change exposed
>>>>> this on our setup by making the sink eligible for that path.
>>>>>
>>>>> The intent is to keep the frame period fixed when userspace has
>>>>> disabled VRR, while retaining FPO eligibility when variable refresh is
>>>>> actually requested. With the patch, the original setup holds the
>>>>> requested 40fps again. I also tested VRR toggling on a Legion Go 2's
>>>>> internal eDP OLED panel and did not observe a regression.
>>>>>
>>>>> Thanks,
>>>>> Matt
>>>>
>>>
>>
>
