[PATCH] kthread: finer-grained lockdep/cross-release completion

Byungchul Park byungchul.park at lge.com
Wed Dec 20 01:14:07 UTC 2017


On 12/19/2017 6:59 PM, Daniel Vetter wrote:
> On Mon, Dec 18, 2017 at 09:42:13AM -0800, Linus Torvalds wrote:
>> On Sun, Dec 17, 2017 at 11:11 PM, Daniel Vetter <daniel at ffwll.ch> wrote:
>>>
>>> This didn't seem to have made it into -rc4. Anything needed to get it
>>> going?
>>
>> Do you actually see the problem in -rc4?
>>
>> Because we ended up removing the cross-release checking due to other
>> developers complaining. It seemed to need a lot more work before it
>> was ready.
>>
>> So I suspect the patch is simply not relevant any more (even if it's
>> not necessarily wrong either).
> 
> Awww ... I like the cross release stuff, it's catching lots of good bugs
> for us - writing a gpu memory manager that's supposed to interface with
> the core one ain't the simplest thing to get right. That's also why we're
> going around and fixing up fallout (we've worked on the worker annotations
> for 4.14 too). I guess next release, hopefully.
> 
> I think between 4.14 and 4.15 attemps cross-release spotted around 5 or so
> genuine deadlocks in i915 code. And at least right now, with the current
> set of fixups our CI runs are splat-free. So at least a Kconfig option
> would be nice, to be able to enable it easily when you want to.
> 
> Wrt us not noticing: Getting the patches in through normal means takes too
> long, so we constantly carry arounda  bunch of fixup patches to address
> serious issues that block CI (no lockdep is no good CI). That's why we
> won't immediately notice when an issue is resolved some other way.

Hello Ingo and Linus,

IMO, taking it off by default is enough. What fs folk actually
wanted is not to remove whole stuff but make it off by default.

Cross-release logic itself makes sense. Please consider it back
and apply my patch making it off by default.

I will do what I can do for the classification in layered fs.

-- 
Thanks,
Byungchul


More information about the dri-devel mailing list