Re: [PATCH v2] docs: submitting-patches: ask bug fixes for their user-visible impact
From: Leon Romanovsky
Date: Sun Nov 22 2026 - 03:50:34 EST
On Thu, Nov 19, 2026 at 02:01:13PM +0100, Willy Tarreau wrote:
> On Thu, Nov 19, 2026 at 02:59:47PM +0200, Leon Romanovsky wrote:
> > On Thu, Nov 19, 2026 at 01:34:45PM +0100, Willy Tarreau wrote:
> > > On Tue, Nov 17, 2026 at 05:27:21PM +0200, Leon Romanovsky wrote:
> > > > > > Since you're already making changes in that area, could you add a sentence
> > > > > > asking reporters to provide the full dmesg stack trace?
> > > > >
> > > > > In this doc, we already ask them not to [1] :)
> > > >
> > > > This is really bad. You assumed they would trim the logs correctly, and I,
> > > > as the recipient was expected to trust that they were accurate. They are not.
> > > >
> > > > They trim everything, which makes these logs useless to me. As I said, they
> > > > include AI-generated reports, cases with reproductions that are impossible
> > > > with upstream code, and reports for completely out-of-tree code. Everything
> > > > like that was visible in original dump stack and it is lost now.
> > > >
> > > > >
> > > > > > Many AI-generated patches reduce it to one or two lines, which gives no
> > > > > > indication of whether the report is genuine or fabricated. Early in the AI
> > > > > > era, I saw some completely fabricated stack traces.
> > > > >
> > > > > Is it still an issue?
> > > >
> > > > I don't know about faked, but useless yes, especially in the cases where
> > > > in-kernel API is involved.
> > >
> > > The situation completely changed over the last year. LLMs that would
> > > hallucinate bugs and dmesg in single-turn conversations have been
> > > replaced with agents/harnesses calling shell commands and able to
> > > capture contents without inventing them. I agree that we should ask
> > > them to capture such dmesg which nowadays have no reason to be faked
> > > anymore.
> >
> > The meaning of "faked" has shifted. As you said, it previously referred to a
> > completely imaginary call stack, now it refers to a custom module that abuses
> > the in-kernel API. The latter is still "faked" because it cannot be done with
> > the current code and unlikely will be possible in the future.
>
> Ah OK we were indeed not speaking about the same cases.
Over the weekend, I realized that after this wave of incorrect and deliberate
misuse of kernel APIs, we’ll see the next level - misuse of internal driver
APIs as well.
I’m referring to the various EXPORT_SYMBOL()s in large, complex drivers
such as mlx5, bnxt, and Intel’s drivers.
The idea that “let’s trim the dmesg stack” is going to be a disaster for
everyone involved.
Thanks
>
> Willy