Re: [PATCH v4 3/3] block: prevent race condition on bi_status in __bio_chain_endio

On Mon, Dec 1, 2025 at 12:25 PM Christoph Hellwig <[email protected]> wrote:
> On Mon, Dec 01, 2025 at 11:22:32AM +0100, Andreas Gruenbacher wrote:
> > > -       if (bio->bi_status && !parent->bi_status)
> > > -               parent->bi_status = bio->bi_status;
> > > +       if (bio->bi_status)
> > > +               cmpxchg(&parent->bi_status, 0, bio->bi_status);
> >
> > Hmm. I don't think cmpxchg() actually is of any value here: for all
> > the chained bios, bi_status is initialized to 0, and it is only set
> > again (to a non-0 value) when a failure occurs. When there are
> > multiple failures, we only need to make sure that one of those
> > failures is eventually reported, but for that, a simple assignment is
> > enough here.
>
> A simple assignment doesn't guarantee atomicy.

Well, we've already discussed that bi_status is a single byte and so
tearing won't be an issue. Otherwise, WRITE_ONCE() would still be
enough here.

>  It also overrides earlier with later status codes, which might not be desirable.

In an A -> B bio chain, we have two competing bi_status writers:

  (1) when the A fails, B->bi_status will be updated using cmpxchg(),
  (2) when B fails, bi_status will be assigned a non-0 value.

In that scenario, B failures will always win over A failures.

In addition, when we have an A -> B -> C bio chain, we still won't get
"first failure wins" semantics because a failure of A will only be
propagated to C once B completes as well. To "fix" that, we'd have to
"chain" all bios to the same parent instead. But I don't think any of
that is really needed.

Andreas