On Wed, Dec 20, 2023 at 10:12 PM Christoph Hellwig [off-list ref] wrote:

On Wed, Dec 20, 2023 at 05:27:14PM -0800, Song Liu wrote:

quoted

Upper layer (fs, etc.) may issue flush with something like:

  bio_reset(bio, bdev, REQ_OP_FLUSH);
  bio->bi_end_io = xxx;
  submit_bio(bio);

No, they can't.  REQ_OP_FLUSH is currently only used by request
based drivers and only generated by the flush state machine.

Hmm... Then this call trace from bcachefs is the exception here:

bch2_journal_write=>
  submit_bio=>
    ...
    md_handle_request =>
      raid5_make_request =>
        chunk_aligned_read =>
          raid5_read_one_chunk =>

(Not that I particularly like this wart, I'd much prefer file systems
to submit a REQ_OP_FLUSH than an empty write with a preflush flag,
but that's not a trivial change).

 Kent, I think we need to update how bcachefs issue flushes.

Thanks,
Song