Re: [PATCH 6.18.y 0/5] drm/vkms: Backport generic vblank timer to fix ABBA deadlock

[Maarten Lankhorst]

Den 2026-05-26 kl. 14:50, skrev Sasha Levin:
> On Tue, May 26, 2026 at 02:48:55PM +0200, Maarten Lankhorst wrote:
>> Hello,
>>
>> Den 2026-05-26 kl. 14:06, skrev w15303746062:
>>>
>>> Hi Sasha,
>>>
>>>> Looking at the five commits:
>>>>
>>>>  - 1/5 (74afeb812850) is the one that actually fixes the ABBA
>>>>    deadlock you observed under Syzkaller; it adds the generic vblank
>>>>    timer that replaces the open-coded vkms hrtimer path.
>>>>
>>>>  - 2/5 (d54dbb5963bd) adds new CRTC helpers for "simple use cases".
>>>>    No Fixes:/Cc:stable, no described bug.
>>>>
>>>>  - 3/5 (02e2681ffe1a) is a refactor that converts vkms to the new
>>>>    helpers. No Fixes:/Cc:stable, no described bug.
>>>>
>>>>  - 4/5 (79ae8510b5b8) is a v7.1-rc1 timeout bump that depends on 1/5.
>>>>    It is not yet in any released stable, so applying it to 6.18.y
>>>>    would put it on an LTS before any LTS contains it.
>>>>
>>>>  - 5/5 (3946d3ba9934) is a doc fix for 1/5.
>>>>
>>>> Per stable-kernel-rules, what I need to queue is the minimum set that
>>>> fixes the bug. Could you explain, per patch, why 2/5..5/5 are required
>>>> to make 1/5 work / are required to actually fix the deadlock? If only
>>>> 1/5 is needed, please resend just that one with your Signed-off-by
>>>> added (the carried patches today only have Thomas's S-o-b, which
>>>> breaks the chain of custody on a stable submission).
>>>
>>> Thanks for the quick review and for pointing out the missing Signed-off-by. I apologize for that omission; it was my mistake during the cherry-pick process.
>>>
>>> Regarding the dependency chain, I would like to clarify why commit 1/5 alone cannot fix the issue:
>>>
>>> Commits 1/5 and 2/5 introduce the new generic vblank timer infrastructure to the DRM core but do *not* touch the vkms driver at all.
>>> Commit 3/5 (02e2681ffe1a) is the actual fix that modifies `drivers/gpu/drm/vkms/vkms_crtc.c`. It removes the buggy open-coded hrtimer that causes the ABBA deadlock and switches vkms to use the new infrastructure introduced in 1/5 and 2/5.
>>>
>>> Therefore, 1/5, 2/5, and 3/5 form an indivisible set. Applying only 1/5 would leave the deadlock in vkms completely unpatched.
>>>
>>> As for 4/5 and 5/5 (the timeout bump and doc fix), Maarten Lankhorst (DRM maintainer) explicitly recommended pulling in this exact 5-commit list as the proper upstream fix for this specific vkms issue (see the mailing list link in this thread).
>>>
>>> However, if you feel 4/5 and 5/5 introduce unnecessary risk for the 6.18.y stable tree, I can absolutely drop them and only submit 1/5, 2/5, and 3/5.
>>>
>>> I am preparing a v2 patch series now with my Signed-off-by added to the chain of custody. Could you let me know if you prefer the full 5-patch series as recommended by DRM maintainers, or just the minimal 3-patch series?
>>>
>>> Best regards,
>>> Mingyu
>>
>> 5/5 might strictly speaking not be needed as it's a documentation fix and I have no idea of the policy about those.
>>
>> The reporter made a bug report of an ABBA deadlock that was fixed in upstream by the first 4 patches, perhaps it's good to those attach here to this discussion.
> 
> I have no objection to taking all 5 if you're okay with it.
> 

Mingyu made an effort to reproduce, check if commits fixed it and then checked if other
drivers needed the same bug fixed. I'm ok with the series.

Kind regards,
~Maarten Lankhorst