On Thu, Jun 17, 2021 at 01:14:41PM +0800, Qu Wenruo wrote:

There are quite some problems in compression code:

- Weird compressed_bio::pending_bios dance
  If we just don't want compressed_bio being freed halfway, we have more
  sane methods, just like btrfs_subpage::readers.

  So here we fix it by introducing compressed_bio::io_sectors to do the
  job.

- BUG_ON()s inside btrfs_submit_compressed_*()
  Even they are just ENOMEM, we should handle them.
  With io_sectors introduced, we have a way to finish compressed_bio
  all by ourselves, as long as we haven't submitted last bio.

  If we have last bio submitted, then endio will handle it well.

- Duplicated code for compressed bio allocation and submission
  Just small refactor can handle it

- Stripe boundary is checked every time one page is added
  This is overkilled.
  Just learn from extent_io.c refactor which use bio_ctrl to do the
  boundary check only once for each bio.

  Although in compression context, we don't need extra checks in
  extent_io.c, thus we don't need bio_ctrl structure, but
  can afford to do it locally.

- Dead code removal
  One dead comment and a new zombie function,
  btrfs_bio_fits_in_stripe(), can be removed now.

I went through it several times, the changes are scary, but the overall
direction is IMHO the right one, not to say it's fixing the difficult
BUG_ONs.

I'll put it to for-next once it passes a few rounds of fstests. Taking
it to 5.14 could be risky if we don't have enough review and testing,
time is almost up before the code freeze.