Message-ID: <4D71B0D1-B7A5-4C29-BBA2-2E055E19A4C9@nvidia.com>
Date: Tue, 30 Dec 2025 16:13:16 -0500
From: Zi Yan <ziy@...dia.com>
To: Roman Gushchin <roman.gushchin@...ux.dev>
Cc: Shakeel Butt <shakeel.butt@...ux.dev>, Qi Zheng <qi.zheng@...ux.dev>,
 hannes@...xchg.org, hughd@...gle.com, mhocko@...e.com, muchun.song@...ux.dev,
 david@...nel.org, lorenzo.stoakes@...cle.com, harry.yoo@...cle.com,
 imran.f.khan@...cle.com, kamalesh.babulal@...cle.com,
 axelrasmussen@...gle.com, yuanchu@...gle.com, weixugc@...gle.com,
 chenridong@...weicloud.com, mkoutny@...e.com, akpm@...ux-foundation.org,
 hamzamahfooz@...ux.microsoft.com, apais@...ux.microsoft.com,
 lance.yang@...ux.dev, linux-mm@...ck.org, linux-kernel@...r.kernel.org,
 cgroups@...r.kernel.org, Qi Zheng <zhengqi.arch@...edance.com>
Subject: Re: [PATCH v2 00/28] Eliminate Dying Memory Cgroup

On 30 Dec 2025, at 14:34, Roman Gushchin wrote:

> Zi Yan <ziy@...dia.com> writes:
>
>> On 29 Dec 2025, at 23:48, Shakeel Butt wrote:
>>
>>> On Tue, Dec 30, 2025 at 12:25:31PM +0800, Qi Zheng wrote:
>>>>
>>>>
>>> [...]
>>>>>>
>>>>>> Thank you for running the AI review for this patchset, but please do not
>>>>>> directly send the raw data from the AI review to the community, as this
>>>>>> is no different from automated review by a robot.
>>>>>
>>>>> Hi Qi,
>>>>>
>>>>> I don't know why you're so negative towards it. It's been great at
>>>>
>>>> No, I don't object to having a dedicated robot to do this.
>>>>
>>>>> finding pretty tricky bugs often missed by human reviewers. In no way
>>>>> it's a replacement for human reviews, but if a robot can find real
>>>>> issues and make the kernel more reliable and safe, I'm in.
>>>>
>>>> I just think you should do a preliminary review of the AI ​​review results
>>>> instead of sending them out directly. Otherwise, if everyone does this,
>>>> the community will be full of bots.
>>>>
>>>> No?
>
> The problem is that it works only when AI is obviously wrong,
> which is not a large percentage of cases with latest models.
> In my practice with Gemini 3 and Chris Mason's prompts, it almost
> never dead wrong: it's either a real issue or some gray zone.
> And you really often need a deep expertise and a significant amount
> of time to decide if it's real or not, so it's not like you can
> assign a single person who can review all ai reviews.
>
>>>>
>>>
>>> We don't want too many bots but we definitely want at least one AI
>>> review bot. Now we have precedence of BPF and networking subsystem and
>>> the results I have seen are really good. I think the MM community needs
>>> to come together and decide on the formalities of AI review process and
>>> I see Roman is doing some early experimentation and result looks great.
>>
>> Do you mind explaining why the result looks great? Does it mean you agree
>> the regressions pointed out by the AI review?
>>
>> If we want to do AI reviews, the process should be improved instead of
>> just pasting the output from AI. In the initial stage, I think some human
>> intervention is needed, at least adding some comment on AI reviews would
>> be helpful. Otherwise, it looks like you agree completely with AI reviews.
>> In addition, “50% of the reported issues are real”, is the AI tossing
>> a coin when reporting issues?
>
> I said at least 50% in my experience. If there is a 50% chance that
> someone is pointing at a real issue in my code, I'd rather look into it
> and fix or explain why it's not an issue. Btw, this is exactly how I
> learned about this stuff - sent some bpf patches (bpf oom) and got
> excited about a number of real issues discovered by ai review.
>
> I agree though that we should not pollute email threads with a number of
> AI-generated reports with a similar context.
>
>> When I am looking into the prompt part, I have the following questions:
>>
>> 1. What is “Prompts SHA: 192922ae6bf4 ("bpf.md: adjust the documentation
>> about bpf kfunc parameter validation”)”? I got the actual prompts
>> from irc: https://github.com/masoncl/review-prompts/tree/main, but it
>> should be provided along with the review for others to reproduce.
>
> It's a significant amount of text, way too much to directly include into
> emails. SHA from the prompts git should be enough, no?

I mean at least the GitHub link should be provided, otherwise, how can people
know the exact prompts?

>
>> 2. Looking at the mm prompt: https://github.com/masoncl/review-prompts/blob/main/mm.md, are you sure the patterns are all right?
>> 	a. Page/Folio States, Large folios require per-page state tracking for
>> 		Reference counts. I thought we want to get rid of per page refcount.
>>     b. Migration Invariants, NUMA balancing expects valid PTE combinations.
>> 		PROTNONE PTEs are hardware invalid to trigger fault.
>> 	c. TLB flushes required after PTE modifications. How about spurious fault
>> 		handling?
>>
>> 3. For a cgroup patchset, I was expecting some cgroup specific prompt rules,
>> 	but could not find any. What am I missing?
>
> MM and cgroups-specific prompts are definitely in a very early stage.
> But to develop/improve them we need data.

Not just data. You are a maintainer of cgroup, so at least you could add
more cgroup specific rules to improve the quality of AI reviews.


Best Regards,
Yan, Zi

Powered by blists - more mailing lists