On Wed, 3 Jun 2026, Samuel Moelius wrote:
> dm-ebs can discard a block while a dirty dm-bufio buffer for the
> same block is still cached. If that buffer is later written back,
> stale data can be written over the discarded state.
>
> That resurrects data that userspace explicitly discarded and breaks
> the expected discard semantics of the target.
>
> Invalidate or clean the matching bufio state when processing discards
> so that old dirty buffers cannot be written back after the discard.
>
> Assisted-by: Codex:gpt-5.5-cyber-preview
> Signed-off-by: Samuel Moelius <[email protected]>
> ---
> drivers/md/dm-ebs-target.c | 8 ++++++--
> 1 file changed, 6 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/md/dm-ebs-target.c b/drivers/md/dm-ebs-target.c
> index 1e52bde48b91..9fece535a0b3 100644
> --- a/drivers/md/dm-ebs-target.c
> +++ b/drivers/md/dm-ebs-target.c
> @@ -212,8 +212,12 @@ static void __ebs_process_bios(struct work_struct *ws)
> write = true;
> r = __ebs_rw_bio(ec, REQ_OP_WRITE, bio);
> } else if (bio_op(bio) == REQ_OP_DISCARD) {
> - __ebs_forget_bio(ec, bio);
> - r = __ebs_discard_bio(ec, bio);
> + if (write)
> + r = dm_bufio_write_dirty_buffers(ec->bufio);
> + if (!r) {
> + __ebs_forget_bio(ec, bio);
> + r = __ebs_discard_bio(ec, bio);
> + }
> }
>
> if (r < 0)
Hi
There's a bug in the patch - when the "write" variable is false, "r" is
equal to -EIO, thus __ebs_forget_bio and __ebs_discard_bio is skipped at
all.
If dm_bufio_write_dirty_buffers returns non-zero, the error is erroneously
reported on the discard bio (that is innocent) and not on the write bios
that triggered the error.
Mikulas
