mbox series

[v9,0/9] Ext4 fast commit performance patchset

Message ID 20250508175908.1004880-1-harshadshirwadkar@gmail.com
Headers show
Series Ext4 fast commit performance patchset | expand

Message

Harshad Shirwadkar May 8, 2025, 5:58 p.m. UTC 
V9 Cover Letter
---------------

This is the V9 of the patch series. This patch series is rebased on
6.15-rc3. This patch series fixes FSTests errors introduced patchset
8. To fix these errors, I have added a few more fixes to "ext4: rework
fast commit commit path" patch. The summary of the changes is as
follows:

- The new commit path first flushes all the data from all the inodes
  before marking inodes as being committed by setting bit
  "EXT4_STATE_FC_COMMITTING". This gets rid of a deadlock between:

  - `ext4_do_writepages()` which ends up waiting for fast commit to
    finish

  - `ext4_fc_perform_commit()` which triggers writeback path while
    submitting inode data.

- The fix to the race above, introduces a new race between
  ext4_fc_del() and ext4_fc_perform_commit(). Since now we delay
  setting "EXT4_STATE_FC_COMMITTING" until after data flush has
  finished, it is now possible that the inode being flushed may get
  deleted by a competing ext4_fc_del(). To fix this race, I introduced
  a new inode state called
  "EXT4_STATE_FC_FLUSHING_DATA". ext4_fc_del() now waits for
  "EXT4_STATE_FC_FLUSHING_DATA" bit to clear.

Verified that no new regression are introduced in -g auto and -g log
groups with fast_commit config.

V8 Cover Letter
--------------

This is the V8 of the patch series. This patch series contains fixes to
review comments by Jan Kara (<jack@suse.cz>). The main changes are as
follows:

- sbi->s_fc_lock is now a mutex type, and it is not released during
  commit path. This gets rid of a number of soft lockups due to race
  with ext4_fc_del().

- I have dropped "ext4: make fast commit ineligible on
  ext4_reserve_inode_write failure" patch.

- I have split "ext4: hold s_fc_lock while during fast commit" into
  two patches: "ext4: convert s_fc_lock to mutex type" and "ext4: hold
  s_fc_lock while during fast commit" to make reviewing easy.

V7 Cover Letter
---------------

This is the V7 of the patch series. This patch series contains fixes to
review comments by Jan Kara (<jack@suse.cz>). The main changes are as
follows:

- As discussed in V6 review, I have dropped "ext4: add nolock mode to
  ext4_map_blocks()" patch given that we now ensure that i_data_sem is
  always grabbed before ext4_mark_inode_dirty() is called.

- I have also dropped "ext4: introduce selective flushing in fast commit"
  given that correctly implementing that would require more changes, and I
  think they would be best done outside of this series.

- I added "ext4: introduce selective flushing in fast commit" as the last
  patch in the series. While testing log group tests I found a few failures
  which get fixed due to this patch.

V6 Cover Letter
---------------

This is the V6 of the patch series. This patch series contains fixes to a
bunch of kernel build warnings reported by Kernel Test Robot
(lkp@intel.com) and Dan Carpenter (dan.carpenter@linaro.org). Thank you!

V5 Cover Letter
---------------

This patch series supersedes the patch series "ext4: remove journal barrier
during fast commit" sent in July 2022[1].  Except following three patches
all the other patches are newly introduced in this series.

* ext4: convert i_fc_lock to spinlock
* ext4: for committing inode, make ext4_fc_track_inode wait
* ext4: drop i_fc_updates from inode fc info

This patchset does not introduce any on-disk format and / or user visible
API change. This patchset reworks fast-commit's commit path improve overall
performance of the commit path. Following optimizations have been added in
this series:

* Avoid having to lock the journal throughout the fast commit.
* Remove tracking of open handles per inode.
* Avoid issuing cache flushes when fast commits are contained within a
  single FUA writes and there is no data that needs flushing.
* Temporarily lift committing thread's priority to match that of the
  journal thread to reduce scheduling delays.

With the changes introduced in this patch series, now the commit path for
fast commits is as follows:

1. Lock the journal by calling jbd2_journal_lock_updates_no_rsv(). This
   ensures that all the exsiting handles finish and no new handles can
   start.
2. Mark all the fast commit eligible inodes as undergoing fast commit by
   setting "EXT4_STATE_FC_COMMITTING" state.
3. Unlock the journal by calling jbd2_journal_unlock_updates. This allows
   starting of new handles. If new handles try to start an update on any of
   the inodes that are being committed, ext4_fc_track_inode() will block
   until those inodes have finished the fast commit.
4. Submit data buffers of all the committing inodes.
5. Wait for [4] to complete.
6. Commit all the directory entry updates in the fast commit space.
7. Commit all the changed inodes in the fast commit space and clear
   "EXT4_STATE_FC_COMMITTING" for all the inodes.
8. Write tail tag to ensure atomicity of commits.

(The above flow has been documented in the code as well)

I verified that the patch series introduces no regressions in "log" groups
when "fast_commit" feature is enabled.

Also, we have a paper on fast commits in USENIX ATC 2024 this year which
should become available on the website[2] in a few months.

[1] https://lwn.net/Articles/902022/
[2] https://www.usenix.org/conference/atc24

Harshad Shirwadkar (9):
  ext4: convert i_fc_lock to spinlock
  ext4: for committing inode, make ext4_fc_track_inode wait
  ext4: mark inode dirty before grabbing i_data_sem in ext4_setattr
  ext4: rework fast commit commit path
  ext4: drop i_fc_updates from inode fc info
  ext4: update code documentation
  ext4: temporarily elevate commit thread priority
  ext4: convert s_fc_lock to mutex type
  ext4: hold s_fc_lock while during fast commit

 fs/ext4/ext4.h        |  19 +-
 fs/ext4/fast_commit.c | 446 ++++++++++++++++++++++--------------------
 fs/ext4/inline.c      |   1 +
 fs/ext4/inode.c       |  12 +-
 fs/ext4/super.c       |   9 +-
 fs/jbd2/journal.c     |   2 -
 6 files changed, 260 insertions(+), 229 deletions(-)

Comments

Theodore Ts'o May 20, 2025, 2:40 p.m. UTC | #1 
On Thu, 08 May 2025 17:58:59 +0000, Harshad Shirwadkar wrote:
> V9 Cover Letter
> ---------------
> 
> This is the V9 of the patch series. This patch series is rebased on
> 6.15-rc3. This patch series fixes FSTests errors introduced patchset
> 8. To fix these errors, I have added a few more fixes to "ext4: rework
> fast commit commit path" patch. The summary of the changes is as
> follows:
> 
> [...]

Applied, thanks!

[1/9] ext4: convert i_fc_lock to spinlock
      commit: 834224e81cdc265456f73fed748a349e43e2d8ef
[2/9] ext4: for committing inode, make ext4_fc_track_inode wait
      commit: 4d3266463ed06af2916b306bdb0ebd647726ba49
[3/9] ext4: mark inode dirty before grabbing i_data_sem in ext4_setattr
      commit: 0b64fd74dd2054834c98867422bc9813e3bba8f4
[4/9] ext4: rework fast commit commit path
      commit: 857d32f2618166765ce9306a246d0745afc76859
[5/9] ext4: drop i_fc_updates from inode fc info
      commit: ed45d331135c317c7f80e8c4e0dad644ca8ca9dc
[6/9] ext4: update code documentation
      commit: 69f35ca189300ddba29a16214159beef45bbd984
[7/9] ext4: temporarily elevate commit thread priority
      commit: 86e07d4b9b0497afef78af773c74258c8f63030f
[8/9] ext4: convert s_fc_lock to mutex type
      commit: 12e64e7f859ed19c5bb497866284d0318c3194a2
[9/9] ext4: hold s_fc_lock while during fast commit
      commit: 6593714d67bab860a733d07895a94404f4ac3039

Best regards,