Re: [PATCH 6.6.y 0/2] backport to fix error propagation of split bios

[Sasha Levin <sashal@kernel.org>]

On Fri, Apr 17, 2026, Ruohan Lan wrote:
> Backport 9ca0e58cb752 ("btrfs: merge btrfs_orig_bbio_end_io() into
> btrfs_bio_end_io()") and d48e1dea3931 ("btrfs: fix error propagation
> of split bios") to 6.6.y.

I had to drop d48e1dea3931 ("btrfs: fix error propagation of split
bios") from 6.6.y as it doesn't build on riscv64.

The commit uses cmpxchg(&bbio->status, ...) where bbio->status is
blk_status_t (u8). riscv's __cmpxchg on 6.6 only supports 4- and
8-byte sizes and triggers BUILD_BUG() for any other size:

    fs/btrfs/bio.c: In function 'btrfs_bio_end_io':
    include/linux/compiler_types.h:474:45: error: call to
      '__compiletime_assert_386' declared with attribute error:
      BUILD_BUG failed
      ...
      cmpxchg(&bbio->status, BLK_STS_OK, status);

1-/2-byte cmpxchg support was added to riscv by 54280ca64626
("riscv/cmpxchg: Implement cmpxchg for variables of size 1 and 2")
in v6.10, which is why mainline builds fine but 6.6 doesn't.

The 9ca0e58cb752 prerequisite is in the queue; d48e1dea3931 itself
will need a re-spin that either avoids sub-word cmpxchg, or a
backport of the riscv cmpxchg8/16 infrastructure as a prerequisite
(which is likely more invasive than we'd want in stable).

Could you (or Naohiro) take a look at reworking the fix so it doesn't
require sub-word cmpxchg on 6.6.y? Something like a spinlock-protected
update, or storing the status in a wider type, would let us keep the
fix on 6.6.y.

--
Thanks,
Sasha