Re: Known and unfixed active data loss bug in MM + XFS with large folios since Dec 2021 (any kernel from 6.1 upwards)

[Linus Torvalds]

On Tue, 24 Sept 2024 at 12:18, Chris Mason <clm@xxxxxxxx> wrote:
>
> A few days of load later and some extra printks, it turns out that
> taking the writer lock in __filemap_add_folio() makes us dramatically
> more likely to just return EEXIST than go into the xas_split_alloc() dance.

.. and that sounds like a good thing, except for the test coverage, I guess.

Which you seem to have fixed:

> With the changes in 6.10, we only get into that xas_destroy() case above
> when the conflicting entry is a shadow entry, so I changed my repro to
> use memory pressure instead of fadvise.
>
> I also added a schedule_timeout(1) after the split alloc, and with all
> of that I'm able to consistently make the xas_destroy() case trigger
> without causing any system instability.  Kairui Song's patches do seem
> to have fixed things nicely.

<confused thumbs up / fingers crossed emoji>

              Linus