On 06/12/2021 03:30, Qu Wenruo wrote:

This patchset be fetched from this branch:

https://github.com/adam900710/linux/tree/refactor_chunk_map

[BACKGROUND]

Currently btrfs never uses bio_split() to split its bio against RAID
stripe boundaries.

Instead inside btrfs we check our stripe boundary everytime we allocate
a new bio, and ensure the new bio never cross stripe boundaries.

[PROBLEMS]

Although this behavior works fine, it's against the common practice used in
stacked drivers, and is making the effort to convert to iomap harder.

There is also an hidden burden, every time we allocate a new bio, we uses
BIO_MAX_BVECS, but since we know the boundaries, for RAID0/RAID10 we can
only fit at most 16 pages (fixed 64K stripe size, and 4K page size),
wasting the 256 slots we allocated.

[CHALLENGES]

To change the situation, this patchset attempts to improve the situation
by moving the bio split into btrfs_map_bio() time, so upper layer should
no longer bother the bio split against RAID stripes or even chunk
boundaries.

But there are several challenges:

- Conflicts in various endio functions
  We want the existing granularity, instead of chained endio, thus we
  must make the involved endio functions to handle split bios.

  Although most endio functions are already doing their works
  independent of the bio size, they are not yet fully handling split
  bio.

  This patch will convert them to use saved bi_iter and only iterate
  the split range instead of the whole bio.
  This change involved 3 types of IOs:

  * Buffered IO
    Including both data and metadata
  * Direct IO
  * Compressed IO

  Their endio functions needs different level of updates to handle split
  bios.

  Furthermore, there is another endio, end_workqueue_bio(), it can't
  handle split bios at all, thus we change the timing so that
  btrfs_bio_wq_end_io() is only called after the bio being split.

- Checksum verification
  Currently we rely on btrfs_bio::csum to contain the checksum for the
  whole bio.
  If one bio get split, csum will no longer points to the correct
  location for the split bio.

  This can be solved by introducing btrfs_bio::offset_to_original, and
  use that new member to calculate where we should read csum from.

  For the parent bio, it still has btrfs_bio::csum for the whole bio,
  thus it can still free it correctly.

- Independent endio for each split bio
  Unlike stack drivers, for RAID10 btrfs needs to try its best effort to
  read every sectors, to handle the following case: (X means bad, either
  unable to read or failed to pass checksum verification, V means good)

  Dev 1	(missing) | D1 (X) |
  Dev 2 (OK)	  | D1 (V) |
  Dev 3 (OK)	  | D2 (V) |
  Dev 4 (OK)	  | D2 (X) |

  In the above RAID10 case, dev1 is missing, and although dev4 is fine,
  its D2 sector is corrupted (by bit rot or whatever).

  If we use bio_chain(), read bio for both D1 and D2 will be split, and
  since D1 is missing, the whole D1 and D2 read will be marked as error,
  thus we will try to read from dev2 and dev4.

  But D2 in dev4 has csum mismatch, we can only rebuild D1 and D2
  correctly from dev2:D1 and dev3:D2.

  This patchset resolve this by saving bi_iter into btrfs_bio::iter, and
  uses that at endio to iterate only the split part of an bio.
  Other than this, existing read/write page endio functions can handle
  them properly without problem.

- Bad RAID56 naming/functionality
  There are quite some RAID56 call sites relies on specific behavior on
  __btrfs_map_block(), like returning @map_length as stripe_len other
  than real mapped length.

  This is handled by some small cleanups specific for RAID56.

[CHANGELOG]
RFC->v1:
- Better patch split
  Now patch 01~06 are refactors/cleanups/preparations.
  While 07~13 are the patches that doing the conversion while can handle
  both old and new bio split timings.
  Finally patch 14~16 convert the bio split call sites one by one to
  newer facility.
  The final patch is just a small clean up.

- Various bug fixes
  During the full fstests run, various stupid bugs are exposed and
  fixed.

v2:
- Fix the error paths for allocated but never submitted bios
  There are tons of error path that we allocate a bio but it goes
  bio_endio() directly without going through btrfs_map_bio().
  New ASSERTS() in endio functions require a populated btrfs_bio::iter,
  thus for such bios we still need to call btrfs_bio_save_iter() to
  populate btrfs_bio::iter to prevent such ASSERT()s get triggered.

- Fix scrub_stripe_index_and_offset() which abuses stripe_len and
  mapped_length

FYI the patchset doesn't apply cleanly to misc-next anymore. I've
pulled your branch form github and queued it for testing on zoned
devices.

I'll report any findings.