Re: Patch pre-submission to Sashiko

[Mark Brown <broonie@kernel.org>]

[-- Attachment #1: Type: text/plain, Size: 1490 bytes --]

On Wed, Oct 21, 2026 at 12:20:43PM -0400, Theodore Tso wrote:
> On Wed, Oct 21, 2026 at 12:07:09PM -0500, Mark Brown wrote:

> > TBH having seen some LLM patches that come with LLM generated testcases
> > the testcases are much worse than the patches themselves.  Open coding,
> > weird structures, asserting bizzare things without explanation - just
> > generally really hard to understand what the test is trying to
> > accomplish and why.

> I recently posted a patch[1] to add KUnit tests for
> fs/ext4/extents_status.c.  It was completely AI generated, and used
> these instructions which I gave to Gemini Pro[2].

> [1] https://lore.kernel.org/all/20261021031231.40107-1-tytso@mit.edu/
> [2] https://lore.kernel.org/all/atbZWze194QiSOTe@mit.edu/

> I'm obviously biased, but it seems to me these test cases were pretty
> good.  I'd appreciate criticisms and suggestions for improvements, but
> if your experience is test cases written with older models, the latest
> frontier models have gotten way better.

Yeah, it's not that you can't do it but rather that I'm nervous of how
much of a useful gate it is to require tests - my general feel is that
if the commits have problems then any associated tests tend to have more
problems.  One example I reviewed earlier today is here:

   https://lore.kernel.org/r/ed22ebf64ed00a56d5e1eff31ac639fef00dcae6.1792179433.git.ivanrwcm25@gmail.com

which will get the issue it was looking for but isn't something I'd want
to maintain long term.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]