Re: Patch pre-submission to Sashiko

["Theodore Tso" <tytso@mit.edu>]

On Tue, Oct 20, 2026 at 07:20:53AM -0500, Bart Van Assche wrote:
> 
> Does the requirement to run a local LLM raise the barrier to start with
> kernel coding?

What I was proposing was an expectation that *if* you are using an LLM
to do kernel coding, best practices would include requiring the LLM
performing the agentic coding loop to do the following:

   * Validate the commit using checkpatch
   * Do a test compile
   * Run any relevant kunit or kselftests
   * Run the commit via an Agentic review system

and maybe even:

   * Add a kunit test covering any changes made by the agentically
     generated code.

If someone wants to send a 100% human-crafted code, today we strongly
suggest (but don't absolutely require) that people at least run
checkpatch, and to read, mark, learn, and inwardly digest (with
apologies to Thomas Cranmer) the rules in SubmittingPatches.  I've in
the past requested that people at *least* run "kvm-xfstests smoke",
but that hasn't been particularly successful.  Should we ask that
human code contributors do more, such as running kunit tests?
Perhaps.

Given that code contributors vastly outnumber maintainers and
reviewerrs, it seems that anything we can do to request that people
generating load on the maintainers and reviewers take at least some
steps to make our lives easier.

						- Ted

P.S.  Ideally I'd love it if we could ask people to run either
"kvm-xfstests -c ext4/4k -g auto" for all ext4 patches, but that would
take 1-2 hours on their development machine, or run "gce-xfstests -c
ext4/4k -g auto", but that would require that they have a Google
Compute Engine account and be willing to spend $0.25 USD.  That's
probably unreasonable, but a maintainer can dream.  :-)