Re: [PATCH 1/9] block: fix data loss and stale date exposure problems during append write

Andreas Gruenbacher <[email protected]> 于2025年11月22日周六 00:13写道：
>
> On Fri, Nov 21, 2025 at 11:38 AM Christoph Hellwig <[email protected]> wrote:
> > On Fri, Nov 21, 2025 at 04:17:40PM +0800, zhangshida wrote:
> > > From: Shida Zhang <[email protected]>
> > >
> > > Signed-off-by: Shida Zhang <[email protected]>
> > > ---
> > >  block/bio.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > >
> > > diff --git a/block/bio.c b/block/bio.c
> > > index b3a79285c27..55c2c1a0020 100644
> > > --- a/block/bio.c
> > > +++ b/block/bio.c
> > > @@ -322,7 +322,7 @@ static struct bio *__bio_chain_endio(struct bio *bio)
> > >
> > >  static void bio_chain_endio(struct bio *bio)
> > >  {
> > > -     bio_endio(__bio_chain_endio(bio));
> > > +     bio_endio(bio);
> >
> > I don't see how this can work.  bio_chain_endio is called literally
> > as the result of calling bio_endio, so you recurse into that.
>
> Hmm, I don't actually see where: bio_endio() only calls
> __bio_chain_endio(), which is fine.
>
> Once bio_chain_endio() only calls bio_endio(), it can probably be
> removed in a follow-up patch.
>
> Also, loosely related, what I find slightly odd is this code in
> __bio_chain_endio():
>
>         if (bio->bi_status && !parent->bi_status)
>                 parent->bi_status = bio->bi_status;
>
> I don't think it really matters whether or not parent->bi_status is
> already set here?
>
> Also, multiple completions can race setting bi_status, so shouldn't we
> at least have a WRITE_ONCE() here and in the other places that set
> bi_status?
>

I'm considering whether we need to add a WRITE_ONCE() in version 2
 of this series.

From my understanding, WRITE_ONCE() prevents write merging and
tearing by ensuring the write operation is performed as a single, atomic
access. For instance, it stops the compiler from splitting a 32-bit write
into multiple 8-bit writes that could be interleaved with reads from other
CPUs.

However, since we're dealing with a single-byte (u8/blk_status_t) write,
it's naturally atomic at the hardware level. The CPU won't tear a byte-sized
write into separate bit-level operations.

Therefore, we could potentially change it to::

        if (bio->bi_status && !READ_ONCE(parent->bi_status))
                parent->bi_status = bio->bi_status;

But as you mentioned, the check might not be critical here. So ultimately,
we can simplify it to:

        if (bio->bi_status)
                parent->bi_status = bio->bi_status;

Thanks,
shida

> Thanks,
> Andreas
>