Re: [PATCH v2] driver core: Fix userspace expectations of uevent_show() as a probe barrier

From: Dan Williams
Date: Mon Oct 28 2024 - 18:58:11 EST


Thorsten Leemhuis wrote:
> On 08.10.24 02:26, Dan Williams wrote:
> > Dan Williams wrote:
> >> Changes since v1 [1]:
> >> - Move the new "locked" infrastructure to private header files to make
> >> it clear it is not approved for general usage (Greg)
> >
> > Greg, per the 0day report and further testing I am missing something
> > subtle in using kernfs open files to pin device objects. So hold off on
> > this for now until I can get that root caused. If someone else can spot
> > what I missed feel free to chime in, but otherwise I will circle back.
> >
> > If I don't get back to this before -rc6 I think the theoretical deadlock
> > that would be re-introduced by a revert of 15fffc6a5624 would be
> > preferable to this reported regression. I am not aware of any reports of
> > that deadlock triggering in practice.
> Was there any progress? If not: given that Linus prefers to have things
> fixed by -rc6 I wonder if now would be a good time to get the revert on
> track for a merge later this week.

Revert 15fffc6a5624 ("driver core: Fix uevent_show() vs driver detach
race"), which reintroduces a theoretical lockdep splat, is my preference
at this point.

Even if I had a new version of this replacement patch in hand today I
would still want it to be v6.13 material, not v6.12-rc. It deserves a
full kernel cycle soak time to shake out issues before release.