On 2024/10/22 14:59, Sedat Dilek wrote:
> On Tue, Oct 22, 2024 at 5:13 AM Zhang Yi <yi.zhang@xxxxxxxxxxxxxxx> wrote:
>>
>> From: Zhang Yi <yi.zhang@xxxxxxxxxx>
>>
>> Hello！
>>
>> This patch series is the latest version based on my previous RFC
>> series[1], which converts the buffered I/O path of ext4 regular files to
>> iomap and enables large folios. After several months of work, almost all
>> preparatory changes have been upstreamed, thanks a lot for the review
>> and comments from Jan, Dave, Christoph, Darrick and Ritesh. Now it is
>> time for the main implementation of this conversion.
>>
>> This series is the main part of iomap buffered iomap conversion, it's
>> based on 6.12-rc4, and the code context is also depend on my anohter
>> cleanup series[1] (I've put that in this seris so we can merge it
>> directly), fixed all minor bugs found in my previous RFC v4 series.
>> Additionally, I've update change logs in each patch and also includes
>> some code modifications as Dave's suggestions. This series implements
>> the core iomap APIs on ext4 and introduces a mount option called
>> "buffered_iomap" to enable the iomap buffered I/O path. We have already
>> supported the default features, default mount options and bigalloc
>> feature. However, we do not yet support online defragmentation, inline
>> data, fs_verify, fs_crypt, ext3, and data=journal mode, ext4 will fall
>> to buffered_head I/O path automatically if you use those features and
>> options. Some of these features should be supported gradually in the
>> near future.
>>
>> Most of the implementations resemble the original buffered_head path;
>> however, there are four key differences.
>>
>> 1. The first aspect is the block allocation in the writeback path. The
>> iomap frame will invoke ->map_blocks() at least once for each dirty
>> folio. To ensure optimal writeback performance, we aim to allocate a
>> range of delalloc blocks that is as long as possible within the
>> writeback length for each invocation. In certain situations, we may
>> allocate a range of blocks that exceeds the amount we will actually
>> write back. Therefore,
>> 1) we cannot allocate a written extent for those blocks because it may
>> expose stale data in such short write cases. Instead, we should
>> allocate an unwritten extent, which means we must always enable the
>> dioread_nolock option. This change could also bring many other
>> benefits.
>> 2) We should postpone updating the 'i_disksize' until the end of the I/O
>> process, based on the actual written length. This approach can also
>> prevent the exposure of zero data, which may occur if there is a
>> power failure during an append write.
>> 3) We do not need to pre-split extents during write-back, we can
>> postpone this task until the end I/O process while converting
>> unwritten extents.
>>
>> 2. The second reason is that since we always allocate unwritten space
>> for new blocks, there is no risk of exposing stale data. As a result,
>> we do not need to order the data, which allows us to disable the
>> data=ordered mode. Consequently, we also do not require the reserved
>> handle when converting the unwritten extent in the final I/O worker,
>> we can directly start with the normal handle.
>>
>> Series details:
>>
>> Patch 1-10 is just another series of mine that refactors the fallocate
>> functions[1]. This series relies on the code context of that but has no
>> logical dependencies. I put this here just for easy access and merge.
>>
>> Patch 11-21 implement the iomap buffered read/write path, dirty folio
>> write back path and mmap path for ext4 regular file.
>>
>> Patch 22-23 disable the unsupported online-defragmentation function and
>> disable the changing of the inode journal flag to data=journal mode.
>> Please look at the following patch for details.
>>
>> Patch 24-27 introduce "buffered_iomap" mount option (is not enabled by
>> default now) to partially enable the iomap buffered I/O path and also
>> enable large folio.
>>
>>
>> About performance:
>>
>> Fio tests with psync on my machine with Intel Xeon Gold 6240 CPU with
>> 400GB system ram, 200GB ramdisk and 4TB nvme ssd disk.
>>
>> fio -directory=/mnt -direct=0 -iodepth=$iodepth -fsync=$sync -rw=$rw \
>> -numjobs=${numjobs} -bs=${bs} -ioengine=psync -size=$size \
>> -runtime=60 -norandommap=0 -fallocate=none -overwrite=$overwrite \
>> -group_reportin -name=$name --output=/tmp/test_log
>>
>
> Hi Zhang Yi,
>
> can you clarify about the FIO values for the diverse parameters?
>
Hi Sedat,
Sure, the test I present here is a simple single-thread and single-I/O
depth case with psync ioengine. Most of the FIO parameters are shown
in the tables below.
For the rest, the 'iodepth' and 'numjobs' are always set to 1 and the
'size' is 40GB. During the write cache test, I also disable the write
back process through:
echo 0 > /proc/sys/vm/dirty_writeback_centisecs
echo 100 > /proc/sys/vm/dirty_background_ratio
echo 100 > /proc/sys/vm/dirty_ratio
Thanks,
Yi.
>
>> == buffer read ==
>>
>> buffer_head iomap + large folio
>> type bs IOPS BW(MiB/s) IOPS BW(MiB/s)
>> -------------------------------------------------------
>> hole 4K 576k 2253 762k 2975 +32%
>> hole 64K 48.7k 3043 77.8k 4860 +60%
>> hole 1M 2960 2960 4942 4942 +67%
>> ramdisk 4K 443k 1732 530k 2069 +19%
>> ramdisk 64K 34.5k 2156 45.6k 2850 +32%
>> ramdisk 1M 2093 2093 2841 2841 +36%
>> nvme 4K 339k 1323 364k 1425 +8%
>> nvme 64K 23.6k 1471 25.2k 1574 +7%
>> nvme 1M 2012 2012 2153 2153 +7%
>>
>>
>> == buffer write ==
>>
>> buffer_head iomap + large folio
>> type Overwrite Sync Writeback bs IOPS BW IOPS BW(MiB/s)
>> ----------------------------------------------------------------------
>> cache N N N 4K 417k 1631 440k 1719 +5%
>> cache N N N 64K 33.4k 2088 81.5k 5092 +144%
>> cache N N N 1M 2143 2143 5716 5716 +167%
>> cache Y N N 4K 449k 1755 469k 1834 +5%
>> cache Y N N 64K 36.6k 2290 82.3k 5142 +125%
>> cache Y N N 1M 2352 2352 5577 5577 +137%
>> ramdisk N N Y 4K 365k 1424 354k 1384 -3%
>> ramdisk N N Y 64K 31.2k 1950 74.2k 4640 +138%
>> ramdisk N N Y 1M 1968 1968 5201 5201 +164%
>> ramdisk N Y N 4K 9984 39 12.9k 51 +29%
>> ramdisk N Y N 64K 5936 371 8960 560 +51%
>> ramdisk N Y N 1M 1050 1050 1835 1835 +75%
>> ramdisk Y N Y 4K 411k 1609 443k 1731 +8%
>> ramdisk Y N Y 64K 34.1k 2134 77.5k 4844 +127%
>> ramdisk Y N Y 1M 2248 2248 5372 5372 +139%
>> ramdisk Y Y N 4K 182k 711 186k 730 +3%
>> ramdisk Y Y N 64K 18.7k 1170 34.7k 2171 +86%
>> ramdisk Y Y N 1M 1229 1229 2269 2269 +85%
>> nvme N N Y 4K 373k 1458 387k 1512 +4%
>> nvme N N Y 64K 29.2k 1827 70.9k 4431 +143%
>> nvme N N Y 1M 1835 1835 4919 4919 +168%
>> nvme N Y N 4K 11.7k 46 11.7k 46 0%
>> nvme N Y N 64K 6453 403 8661 541 +34%
>> nvme N Y N 1M 649 649 1351 1351 +108%
>> nvme Y N Y 4K 372k 1456 433k 1693 +16%
>> nvme Y N Y 64K 33.0k 2064 74.7k 4669 +126%
>> nvme Y N Y 1M 2131 2131 5273 5273 +147%
>> nvme Y Y N 4K 56.7k 222 56.4k 220 -1%
>> nvme Y Y N 64K 13.4k 840 19.4k 1214 +45%
>> nvme Y Y N 1M 714 714 1504 1504 +111%
>>
>> Thanks,
>> Yi.
>>
>> Major changes since RFC v4:
>> - Disable unsupported online defragmentation, do not fall back to
>> buffer_head path.
>> - Wite and wait data back while doing partial block truncate down to
>> fix a stale data problem.
>> - Disable the online changing of the inode journal flag to data=journal
>> mode.
>> - Since iomap can zero out dirty pages with unwritten extent, do not
>> write data before zeroing out in ext4_zero_range(), and also do not
>> zero partial blocks under a started journal handle.
>>
>> [1] https://lore.kernel.org/linux-ext4/20241010133333.146793-1-yi.zhang@xxxxxxxxxx/
>>
>> ---
>> RFC v4: https://lore.kernel.org/linux-ext4/20240410142948.2817554-1-yi.zhang@xxxxxxxxxxxxxxx/
>> RFC v3: https://lore.kernel.org/linux-ext4/20240127015825.1608160-1-yi.zhang@xxxxxxxxxxxxxxx/
>> RFC v2: https://lore.kernel.org/linux-ext4/20240102123918.799062-1-yi.zhang@xxxxxxxxxxxxxxx/
>> RFC v1: https://lore.kernel.org/linux-ext4/20231123125121.4064694-1-yi.zhang@xxxxxxxxxxxxxxx/
>>
>>
>> Zhang Yi (27):
>> ext4: remove writable userspace mappings before truncating page cache
>> ext4: don't explicit update times in ext4_fallocate()
>> ext4: don't write back data before punch hole in nojournal mode
>> ext4: refactor ext4_punch_hole()
>> ext4: refactor ext4_zero_range()
>> ext4: refactor ext4_collapse_range()
>> ext4: refactor ext4_insert_range()
>> ext4: factor out ext4_do_fallocate()
>> ext4: move out inode_lock into ext4_fallocate()
>> ext4: move out common parts into ext4_fallocate()
>> ext4: use reserved metadata blocks when splitting extent on endio
>> ext4: introduce seq counter for the extent status entry
>> 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
>> ext4: don't order data for inode with EXT4_STATE_BUFFERED_IOMAP
>> ext4: implement writeback iomap path
>> ext4: implement mmap iomap path
>> ext4: do not always order data when partial zeroing out a block
>> ext4: do not start handle if unnecessary while partial zeroing out a
>> block
>> ext4: implement zero_range iomap path
>> ext4: disable online defrag when inode using iomap buffered I/O path
>> ext4: disable inode journal mode when using iomap buffered I/O path
>> ext4: partially enable iomap for the buffered I/O path of regular
>> files
>> ext4: enable large folio for regular file with iomap buffered I/O path
>> ext4: change mount options code style
>> ext4: introduce a mount option for iomap buffered I/O path
>>
>> fs/ext4/ext4.h | 17 +-
>> fs/ext4/ext4_jbd2.c | 3 +-
>> fs/ext4/ext4_jbd2.h | 8 +
>> fs/ext4/extents.c | 568 +++++++++++----------------
>> fs/ext4/extents_status.c | 13 +-
>> fs/ext4/file.c | 19 +-
>> fs/ext4/ialloc.c | 5 +
>> fs/ext4/inode.c | 755 ++++++++++++++++++++++++++++++------
>> fs/ext4/move_extent.c | 7 +
>> fs/ext4/page-io.c | 105 +++++
>> fs/ext4/super.c | 185 ++++-----
>> include/trace/events/ext4.h | 57 +--
>> 12 files changed, 1153 insertions(+), 589 deletions(-)
>>
>> --
>> 2.46.1
>>
>>
