On 2026-03-10 12:19, Michele Palazzi wrote:
> On 3/5/26 23:21, [email protected] wrote:
>> From: Leo Li <[email protected]>
>>
>> [Why]
>>
>> It's suspected that there's a race condition where cursor-only updates
>> can race with page-flip updates when updating the amdgpu_crtc->event,
>> which is used to track the pending vblank event until it's send by the
>> interrupt handler.
>>
>> Theoretically, DRM should prevent concurrent updates on the same CRTC.
>> But evidently, according to findings[1] on a previous attemp to fix
>> this, there is a race.
>>
>> [How]
>>
>> Save pending vblank events from cursor-only updates seperately in
>> amdgpu_crtc->cursor_event. Since cursor updates won't arm the pflip_irq
>> to fire, we'll send it in crtc/vupdate_irq handlers.
>>
>> [1]
>> https://lore.kernel.org/amd-gfx/[email protected]/
>>
>
> fyi i am testing this patch, looking good so far but i'd give it another 24
> hours to be sure
There a report that this patch doesn't help on the original gitlab issue:
https://gitlab.freedesktop.org/drm/amd/-/issues/3787#note_3366574
I'm suspecting the issue isn't at the arming side, but something is preventing
delivery at the interrupt handlers.
I'll reply to the thread with your first patch with some things to try, after I
mull on it a little more.
- Leo
