Conversation

Some random remarks of stuff that makes LLM kernel patches diffucult:

1. The form is right i.e., false positives are well engineered, and they happen often.
2. LLM often does fix over fix. I.e. another fix already addresses the problem and the second fix would be only meaningful if previous was not already applied.
3. This happens also at least 60-70% LLM assisted that I review as a kernel maintainer: fix is technically correct but it is not best possible way to fix the issue.This is important because it directly affects fix coverage.
4. This happens at least weekly either in review or in my own doings: LLMs ignore the history very often. I almost sent patch set that erased existentially critical kernel bug fix because of this behavior. It's good that I triple-check, isn't it? :-)

In one case I said to a contributor that if I need rewrite again bug fixes as the review response, next time I'll just do the fix myself and mark the contributor as reporter. I have right to do that if hard facts support this decision. I don't care the credit myself but it is morally right decision for contributors who deserve the credit from their hard work.