Re: [MAINTAINERS SUMMIT] Time to call it quits for the maintainer summit?

[Krzysztof Kozlowski <krzk@kernel.org>]

On 12/08/2026 08:20, Dave Airlie wrote:
> On Wed, 12 Aug 2026 at 16:01, Krzysztof Kozlowski <krzk@kernel.org> wrote:
>>
>> On 11/08/2026 05:29, Dave Airlie wrote:
>>>>>> If your assumption is correct, this is internal cat herding ... MM has
>>>>>> much the same problem except that it has to deal with somewhat
>>>>>> opinionated architecture maintainers (around 22 of them) to agree on
>>>>>> the internal abstractions for MM primitives.  Perhaps rather than
>>>>>> getting into my problem is bigger than yours type arguments, we could
>>>>>> observe that MM might run a bit more smoothly because it gets an
>>>>>> additional 3 day conference (LSF/MM) plus a MC at Plumbers to sort
>>>>>> itself out?
>>>>
>>>>> I guess the question is, what exactly is being big for DRM that it needs to do
>>>>> cherry-picking for their linux-next branch whereas Networking and ARM do
>>>>> not? Is DRM bigger than either of those? From what I understand VFS doesn't
>>>>> have to do that either. And VFS has a large range of file systems to deal
>>>>> with. But it also has one of the best abstraction layers of the kernel.
>>>>
>>>> It's not even all of DRM, it's specifically the AMD and Intel driver
>>>> stacks which as far as I can tell just routinely cherry pick all their
>>>> fixes between their development and fixes branches without even
>>>> considering the possibility of either merging up the fixes branch or
>>>> just waiting and letting the fixes propagate back.  None of the rest of
>>>> the DRM trees ever seems to cause these issues.
>>>
>>> AMD and Intel are just the two largest teams with the longest
>>> pipelines of internal engineers and teams. There are between 50 and
>>> 100 people in those groups, across multiple disjoint teams feeding
>>> into a single driver. It just doesn't scale for all 50-100 people to
>>> understand the cycle of the upstream Linus tree at all times for all
>>> patches.
>>
>> Qualcomm has also between 50-100 people in different teams upstreaming
>> to similar subsystems (like SoC) and they try to understand the cycle.
>>
>> It's just AMD and Intel do not care to understand, because it is easier
>> for them and no one nags them...
> 
> Get back to me when their devices are in a place to be running
> upstream kernels as much as AMD or Intel at the same scale.

So I am back. All recent flagship SoCs since a 3-4 years have full
upstream patches posted from day 0 of hardware release. Since a year
Qualcomm upstreams most of the patches before hardware release.

And their devices is a SoC, consisting of not only GPU but few other
DSPs and multum of other IP blocks, so not even single hardware.

I am not saying that they are doing perfect, but they received clear
guideline they need to adjust to upstream process and cycle and they do
try to adjust. I know it from first hand since I have been directly
involved in this for a year now.

> 
> We've tried to keep the upstream schedule for ages, as Rodrigo points
> out it falls down, stuff goes missing in the shutdown periods, telling
> teams to stop working for 2-3 weeks isn't an option in most companies.

Well, we tell that Qualcomm and Qualcomm listens to maintainers. Are you
saying that others cannot listen to maintainers? That is a very bad
precedent, because now Qualcomm managers will get back to me and say
"why are you so harsh to us? can't we get some slack like others?"

> Intel and AMD are trying to upstream GPUs that aren't even on the
> market yet, velocity mattters a lot more because thier customers are

Same Qualcomm. Qualcomm posts patches all in public like 9 months before
hardware is going to be announced. Publicly announced, so availability
even later!

> usually on the end of the pipeline via Linus' tree,
> 
> Qualcomm is not in the same position, a lot of their pipeline delivery
> is via Android or ChromeOS and they have nowhere near the amount of
> regression finding problems that upstream GPUs have.

No true anymore. They do release to Android but that's downstream and we
do not talk about it here. I talk about upstream and their
upstream-first releases for all of their major and minor products.

That's mainline Linux distros: Debian, Ubuntu, Yocto. Their current BSP
is pure upstream based.

That's the same amount of regression handling as Intel and AMD.

Do not assume AMD and Intel are the only ones or are somehow special.


Best regards,
Krzysztof