Message-Id: <20231123125121.4064694-1-yi.zhang@huaweicloud.com>
Date: Thu, 23 Nov 2023 20:51:02 +0800
From: Zhang Yi <yi.zhang@...weicloud.com>
To: linux-ext4@...r.kernel.org
Cc: linux-fsdevel@...r.kernel.org,
	tytso@....edu,
	adilger.kernel@...ger.ca,
	jack@...e.cz,
	ritesh.list@...il.com,
	hch@...radead.org,
	djwong@...nel.org,
	yi.zhang@...wei.com,
	yi.zhang@...weicloud.com,
	chengzhihao1@...wei.com,
	yukuai3@...wei.com
Subject: [RFC PATCH 00/18] ext4: use iomap for regular file's buffered IO path and enable large foilo

From: Zhang Yi <yi.zhang@...wei.com>

Hello,

This is a RFC patch set based on 6.6 that partial switch to use iomap
for regular file's buffered IO path in ext4. Now this only support ext4
filesystem with the default features and mount options, didn't support
inline_data, bigalloc, dax, fs_verity, fs_crypt, and data=journal mode
yet. I have test it through fstests -g quick with 4K block size and
some simple performance tests, all the apparent issues have been fixed
right now. This is just for discussion and check the overall plan is
feasible or not, I haven't done other tests, there must be some other
bugs need to fix later. This is the first time I've developed such a
large feature for ext4, so I hope you would like it, I will keep on
testing and improving these patches, any comments are helpful. For the
convenience of review, I split the implements into small patches.

Patch 1-6: this is a preparation that changes ext4_map_blocks() and
ext4_set_iomap() that could recognize delayed only extents, I've send it
out separately[1] because I've done full tests and suppose this can be
reviewed and merged firstly.

Patch 7: this is also a preparation that let ext4_insert_delayed_block()
can add multi-blocks once a time.

Patch 8-16: Introduce a new aops names ext4_iomap_aops for iomap,
implement buffer read, buffer write, writeback, mmap and 
zero_range/truncate path. Also inculdes two minor modifications in
iomap, Please look at the following patch for details.

Patch 17-18: Switch to iomap for regular file's buffered IO path besides
inline_data, bigalloc, dax, fs_verity, fs_crypt, and data=journal mode,
and enable large folio.

About Tests:
 - I've test it through fstests -g quick with 4k block size.
 - I've done a few quick performance tests below.
   Fio tests with psync and 1 thread on my virt machine's ramdisk.

                    bs  buffer_head              iomap with large foilo
   ---------------------------------------------------------------
   write+fsync      1M  IOPS=651, BW=652MiB/s    IOPS=1442, BW=1442MiB/s
   overwrite+fsync  1M  IOPS=832, BW=832MiB/s    IOPS=2064, BW=2065MiB/s
   read hole        1M  IOPS=1769, BW=1769MiB/s  IOPS=3556, BW=3557MiB/s
   read data        1M  IOPS=1464, BW=1465MiB/s  IOPS=2376, BW=2377MiB/s

TODO
 - Do 'kvm-xfstests -g auto' tests and keep on doing stress tests and
   fixing.
 - Do further performance analyse.
 - I noticed the discussion from Ritesh's ext2 conversion[2], I will
   also check the common part in it.
 - Support more filesystem features and mount options and maybe could
   totally remove ext4_da_aops and ext4_aops in the future.

[1] https://lore.kernel.org/linux-ext4/20231121093429.1827390-1-yi.zhang@huaweicloud.com/T/#t
[2] https://lore.kernel.org/linux-ext4/ZV2k9pR13SbXitRT@infradead.org/T/#t

Thanks,
Yi.

Zhang Yi (18):
  ext4: introduce ext4_es_skip_hole_extent() to skip hole extents
  ext4: make ext4_es_lookup_extent() return the next extent if not found
  ext4: correct the hole length returned by ext4_map_blocks()
  ext4: add a hole extent entry in cache after punch
  ext4: make ext4_map_blocks() distinguish delayed only mapping
  ext4: make ext4_set_iomap() recognize IOMAP_DELALLOC mapping type
  ext4: allow reserving multi-delayed blocks
  ext4: add a new iomap aops for regular file's buffered IO path
  ext4: implement buffered read iomap path
  ext4: implement buffered write iomap path
  iomap: add a fs private parameter to iomap_ioend
  iomap: don't increase i_size if it's not a write operation
  ext4: impliment writeback iomap path
  ext4: impliment zero_range iomap path
  ext4: writeback partial blocks before zero range
  ext4: impliment mmap iomap path
  ext4: partial enable iomap for regular file's buffered IO path
  ext4: enable large folio for regular file which has been switched to
    use iomap

 fs/ext4/ext4.h              |  15 +-
 fs/ext4/ext4_jbd2.c         |   3 +-
 fs/ext4/extents.c           |  14 +-
 fs/ext4/extents_status.c    |  73 ++--
 fs/ext4/extents_status.h    |   6 +-
 fs/ext4/file.c              |  14 +-
 fs/ext4/ialloc.c            |   5 +
 fs/ext4/inode.c             | 737 ++++++++++++++++++++++++++++++++----
 fs/ext4/move_extent.c       |   8 +
 fs/ext4/page-io.c           |  74 ++++
 fs/ext4/super.c             |   2 +
 fs/iomap/buffered-io.c      |   5 +-
 include/linux/iomap.h       |   1 +
 include/trace/events/ext4.h |  40 +-
 14 files changed, 877 insertions(+), 120 deletions(-)

-- 
2.39.2

Powered by blists - more mailing lists