[PATCH -next v5 00/32] ext4: use iomap for regular file's buffered I/O path

From: Zhang Yi <[email protected]>

Hi, all

This is v5 of the series converting ext4's buffered I/O path from
buffer_head to iomap. This iteration extends ext4_map_blocks() with
flexible journal handle management, reworks the buffered write path and
writeback path to avoid open-coded mapping block operations, rewrites
the ordering logic between EOF zeroing and i_disksize updates, postpones
submission of zeroed EOF blocks for better concurrency, and includes
several fixes for stale delalloc extents, partial block zeroing
failures, and ordered I/O wait issues. Rebased on next-20260810 and
collected RVB from Ojaswin and Jan.


You can get commits here:

 https://github.com/zhangyi089/linux/commits/ext4_buffered_iomap_v5/

Patch 01-03: Simplify truncate operations and drop unnecessary ordered
             I/O when zeroing EOF blocks. This prepares for the iomap
	     conversion.
Patch 04-21: Extend ext4_map_blocks() and implement core iomap buffered
             read/write, writeback, mmap, and partial block zeroing.
Patch 22-30: Handle ordered sequence for zeroing the EOF partial block
             and updating i_disksize when performing file extending on
	     unaligned EOF files.
Patch 31-32: Enable iomap buffered I/O path.


Tests and Performance:

Tested with xfstests-bld using -g auto, fast_commit, and 64k
configurations. No new test failures were observed, aside from the
known occasional failure of generic/127, which is an issue in the
MM large folio split logic and warrants a separate fix.

 https://lore.kernel.org/linux-fsdevel/[email protected]/

Performance was tested with FIO on a 4-core virtual machine with an
Intel(R) Xeon(R) Platinum 8380 CPU, backed by a 150GB RAM-backed
virtual-io block device.

Write perf tests (MiB/s)
========================

  Sync  : 1 = fsync() after each write
  OW    : 1 = overwrite existing blocks
  Dnc   : 1 = RWF_DONTCACHE used

  #1    : only write page cache, no writeback
  #2    : writeback triggered (set dirty_background_bytes = 100MB and
          dirty_bytes = 200MB)


  Config                   BS       ext4+bh   ext4+iomap    delta
  ---------------------------------------------------------------
  Sync=0 OW=0 Dnc=0 (#1)   1k          41.8         42.0    ~0%
  Sync=0 OW=0 Dnc=0 (#1)   4k           170          176    +4%
  Sync=0 OW=0 Dnc=0 (#1)   64k         1816         1981    +9%
  Sync=0 OW=0 Dnc=0 (#1)   1m          4295         5780   +35%

  Sync=0 OW=0 Dnc=0 (#2)   1k          42.6         43.2    +2%
  Sync=0 OW=0 Dnc=0 (#2)   4k           137          140    +2%
  Sync=0 OW=0 Dnc=0 (#2)   64k          734          957   +30%
  Sync=0 OW=0 Dnc=0 (#2)   1m          1460         1683   +15%

  Sync=0 OW=0 Dnc=1        1k          14.6         15.0    +3%
  Sync=0 OW=0 Dnc=1        4k          67.9         61.9    -9%
  Sync=0 OW=0 Dnc=1        64k          415          698   +68%
  Sync=0 OW=0 Dnc=1        1m          1404         2883  +105%

  Sync=1 OW=0 Dnc=0        1k          0.71         0.60   -15%
  Sync=1 OW=0 Dnc=0        4k          2.04         1.89    -7%
  Sync=1 OW=0 Dnc=0        64k         32.6         32.4    -1%
  Sync=1 OW=0 Dnc=0        1m           253          277   +10%

  Sync=0 OW=1 Dnc=0 (#1)   1k          19.1         21.0   +10%
  Sync=0 OW=1 Dnc=0 (#1)   4k           164          179    +9%
  Sync=0 OW=1 Dnc=0 (#1)   64k         1776         1867    +5%
  Sync=0 OW=1 Dnc=0 (#1)   1m          4105         4879   +19%

  Sync=0 OW=1 Dnc=0 (#2)   1k          19.9         21.6    +8%
  Sync=0 OW=1 Dnc=0 (#2)   4k           148          157    +6%
  Sync=0 OW=1 Dnc=0 (#2)   64k          843          840    ~0%
  Sync=0 OW=1 Dnc=0 (#2)   1m          1579         1806   +14%

  Sync=0 OW=1 Dnc=1        1k          17.4         17.9    +3%
  Sync=0 OW=1 Dnc=1        4k          80.4         87.7    +9%
  Sync=0 OW=1 Dnc=1        64k          806          923   +14%
  Sync=0 OW=1 Dnc=1        1m          1070         1122    +5%

  Sync=1 OW=1 Dnc=0        1k          0.82         0.82     0%
  Sync=1 OW=1 Dnc=0        4k          3.55         3.56    ~0%
  Sync=1 OW=1 Dnc=0        64k         48.4         49.8    +3%
  Sync=1 OW=1 Dnc=0        1m           325          336    +4%


Read tests (MiB/s)
==================

  READ HOLE           : read of a hole
  READ DATA           : read of file data without pre-exists page cache
  READ PRECACHED DATA : read of file data with warm page cache

  Scenario                 BS       ext4+bh   ext4+iomap    delta
  ---------------------------------------------------------------
  READ HOLE                1k          46.6         47.1    +1%
  READ HOLE                4k           189          198    +5%
  READ HOLE                64k         2011         2116    +5%
  READ HOLE                1m          4538         4684    +3%

  READ DATA                1k          44.6         50.4   +13%
  READ DATA                4k           178          182    +2%
  READ DATA                64k          741          768    +4%
  READ DATA                1m          1143         1210    +6%

  READ PRECACHED DATA      1k          50.8         50.7    ~0%
  READ PRECACHED DATA      4k           179          173    -3%
  READ PRECACHED DATA      64k         2009         1989    -1%
  READ PRECACHED DATA      1m          6296         6294    ~0%

Large I/O scenarios show significant write performance improvement,
while small I/O in synchronous and non-cached scenarios experience a
slight regression. Read performance shows no significant difference.


Changes since v4:
 - Rebase on next-20260810 and collects RVB from Ojaswin and Jan
   (exclude patches with substantial changes).
 - Add patch 04: skip ordered I/O wait when zeroing beyond the
   i_disksize block (suggested by Ojaswin).
 - Add patch 05: Set EXT4_MAP_NEW in ext4_da_map_blocks() to properly
   indicate that a new delayed allocation block has been inserted, fix
   the issue of stale delalloc extent on short iomap buffered write
   (pointed out by Ojaswin).
 - Add patch 06-10: expand ext4_map_blocks(), make it can 1) recheck
   extent status tree before allocating blocks, 2) allow starting it's
   own journal handle if the caller doesn't provide one, callers can
   query first on overwrite and start handle if necessary, 3) avoid
   unnecessary journal handle when callers query pre-existing unwritten
   extents, 4) skip allocating block for holes in data submitting path.
   Preparing for implementing the iomap writeback path (suggested by
   Jan).
 - In patch 13: update the commit message to correct the reason of
   passing out extent seq counter (pointed out by Jan).
 - In patch 14: add function graph to explain the issues encountered
   when trying to support data=ordered mode during iomap conversion
   (suggested by Ojaswin).
 - In patch 15, drop ext4_iomap_get_blocks(), which open-coded mapping
   block logic in the non-delalloc buffered write path, and use the
   extended ext4_map_blocks() directly instead.
 - In patch 16, First rework the writeback block mapping path, directly
   call extended ext4_map_block() instead of open coding. Then Ignore
   the data_error=abort case for now, and always queue the ioend worker
   in ext4_iomap_end_bio(). This simplifies the logic and makes it
   easier to switch to the new ioend infrastructure in the future. Then
   fix the double deferrals for RWF_DONTCACHE writes after the new
   generic bio complete in task framework merged. Finally, reuse the
   i_rsv_conversion_work and i_rsv_conversion_list instead of creat two
   new ones (suggested by Jan).
 - In patch 18, clarify the TODO comments in
   ext4_block_iomap_zero_range() and fix a race condition that could
   cause partial block zeroing to fail.
 - Add patch 19: Clarify and resolve the issue where endio fails to
   convert extents when punch hole is raced by writeback.
 - In patch 22-30, rework the ordering of EOF block zeroing and
   i_disksize updating. Use EXT4_STATE_DISKSIZE_GROW_PENDING state bit
   to check whether a pending EOF zeroing block is still awaiting
   submit, and query i_disksize when writeback dirty folios, only
   updating i_disksize when this block is written back or discarded
   (suggested by Jan). Besides, do not submit zeroed EOF block
   immediately after write zeroes, postpone it to improve concurrency
   (suggested by Ojaswin). Then, handle the cases of fallocate
   operations in delay updating i_disksize when the EOF zeroed block is
   still pending. Finally, add more tracepoints to track the state of
   the EOF block and I/Os.

v4:     https://lore.kernel.org/linux-ext4/[email protected]/
v3:     https://lore.kernel.org/linux-ext4/[email protected]/
v2:     https://lore.kernel.org/linux-ext4/[email protected]/
v1:     https://lore.kernel.org/linux-ext4/[email protected]/
RFC v4: https://lore.kernel.org/linux-ext4/[email protected]/
RFC v3: https://lore.kernel.org/linux-ext4/[email protected]/
RFC v2: https://lore.kernel.org/linux-ext4/[email protected]/
RFC v1: https://lore.kernel.org/linux-ext4/[email protected]/

Comments and suggestions are welcome!

Thanks,
Yi.


Original Cover-letter
=====================

This series adds the iomap buffered I/O path support for regular files.
It implements the core iomap APIs on ext4 and introduces the
'buffered_iomap' mount option to enable the iomap buffered I/O path.
This supports default features, default mount options and bigalloc
feature. However, this does not support online defragmentation, inline
data, fsverify, fscrypt, non-extent inodes, and data=journal mode, we
will fall to buffered_head I/O path automatically if these features and
options are used.

This iomap buffered I/O path is not enabled by default because the
preceding features are not supported. Users can explicitly enable or
disable it via 'buffered_iomap' and 'nobuffered_iomap' mount options.

Key considerations

1. Lock ordering difference

   The lock ordering of folio lock and transaction start in the iomap
   path is the opposite of that in the buffer_head path.

2. data=ordered mode is not used

   Two main reasons:
   a) The lock ordering of folio lock and transaction start for
      data=ordered mode is opposite to the iomap path, which would cause
      a deadlock.
   b) The iomap writeback path does not support partial folio submission
      (required by data=ordered mode when block size < folio size, and
      it is currently handled by ext4_bio_write_folio()), which would
      also cause a deadlock.

   To replace data=ordered mode functionality:

   - For append write: Always allocate unwritten extents (dioread_nolock
     behavior) to prevent stale data exposure.

   - For post-EOF partial block zeroing, mark the inode as
     DISKSIZE_GROW_PENDING after zeroing the EOF block, and postpone
     updating i_disksize to i_size until the zeroed block has been
     written to disk or discarded (Suggested by Jan).

     Tradeoffs:
     1) i_disksize must be updated from io_end processing, so this
        mechanism is restricted to the buffered I/O iomap path due to
        locking constraints. This limitation will eventually go away
	once everything is converted to iomap.
     2) After a crash, i_disksize may be updated while the file content
        shows zeros. This is acceptable for simplicity and does not
	break any existing guarantees.

   - For online defragmentation: Not supported yet, needs further
     consideration.

3. Always enable dioread_nolock

   Two main reasons:
   a) Since data=ordered mode cannot be used, allocating written blocks
      directly would expose stale data.
   b) To optimize writeback, we should allocate blocks based on writeback
      length rather than per-folio mapping. Direct written allocation
      would over-allocate blocks.

   dioread_nolock has been the default mount option for many years, and
   Jan pointed out that we may no longer need to disable it, so gradually
   remove this mount option in the future.



Zhang Yi (32):
  ext4: simplify size updating in ext4_setattr()
  ext4: factor out ext4_truncate_[up|down]()
  ext4: simplify error handling in ext4_setattr()
  ext4: skip ordered I/O wait when zeroing beyond i_disksize block
  ext4: set EXT4_MAP_NEW flag for delayed allocated blocks
  ext4: recheck extent status tree before block allocation
  ext4: fix orig_mlen initialization in ext4_map_blocks()
  ext4: allow ext4_map_blocks() to start its own transaction handle
  ext4: avoid unnecessary transaction in ext4_map_blocks() for unwritten
    extents
  ext4: skip block allocation for holes in the data submission path
  ext4: add iomap address space operations for buffered I/O
  ext4: implement buffered read path using iomap
  ext4: pass out extent seq counter when mapping da blocks
  ext4: do not use data=ordered mode for inodes using buffered iomap
    path
  ext4: implement buffered write path using iomap
  ext4: implement writeback path using iomap
  ext4: implement mmap path using iomap
  ext4: implement partial block zero range path using iomap
  ext4: tolerate unexpected holes in ext4_convert_unwritten_extents()
  ext4: add block mapping tracepoints for iomap buffered I/O path
  ext4: disable online defrag when inode using iomap buffered I/O path
  ext4: add EXT4_STATE_DISKSIZE_GROW_PENDING state bit and helpers
  ext4: submit and wait for pending disksize-grow I/O on writeback
  ext4: advance i_disksize to i_size upon disksize-grow I/O completion
  ext4: defer i_disksize update while DISKSIZE_GROW_PENDING is set
  ext4: submit and wait for disksize-grow I/O in fallocate paths
  ext4: clear DISKSIZE_GROW_PENDING on truncate or error
  ext4: set DISKSIZE_GROW_PENDING after zeroing unaligned EOF block
  ext4: add tracepoints for DISKSIZE_GROW_PENDING set, clear, and wait
  ext4: add tracepoints for EOF block zeroing and disksize-grow I/O
  ext4: partially enable iomap for the buffered I/O path of regular
    files
  ext4: introduce a mount option for iomap buffered I/O path

 fs/ext4/ext4.h              |   83 ++-
 fs/ext4/ext4_jbd2.c         |    8 +-
 fs/ext4/ext4_jbd2.h         |    7 +-
 fs/ext4/extents.c           |   71 ++-
 fs/ext4/file.c              |   20 +-
 fs/ext4/ialloc.c            |    1 +
 fs/ext4/inode.c             | 1183 +++++++++++++++++++++++++++++++----
 fs/ext4/migrate.c           |    2 +
 fs/ext4/move_extent.c       |   11 +
 fs/ext4/page-io.c           |  201 ++++++
 fs/ext4/super.c             |   49 +-
 include/trace/events/ext4.h |  187 ++++++
 12 files changed, 1651 insertions(+), 172 deletions(-)

-- 
2.52.0