From: Zhang Yi <yi.zhang@xxxxxxxxxx> Changes since RFC v2: - Rebase codes on next-20250314. - Add support for nvme multipath. - Add support for NVMeT with block device backing. - Clear FALLOC_FL_WRITE_ZEROES if dm clear limits->max_write_zeroes_sectors. - Complement the counterpart userspace tools(util-linux and xfs_io) and tests(blktests and xfstests), please see below for details. Changes since RFC v1: - Switch to add a new write zeroes operation, FALLOC_FL_WRITE_ZEROES, in fallocate, instead of just adding a supported flag to FALLOC_FL_ZERO_RANGE. - Introduce a new flag BLK_FEAT_WRITE_ZEROES_UNMAP to the block device's queue limit features, and implement it on SCSI sd driver, NVMe SSD driver and dm driver. - Implement FALLOC_FL_WRITE_ZEROES on both the ext4 filesystem and block device (bdev). RFC v2: https://lore.kernel.org/linux-fsdevel/20250115114637.2705887-1-yi.zhang@xxxxxxxxxxxxxxx/ RFC v1: https://lore.kernel.org/linux-fsdevel/20241228014522.2395187-1-yi.zhang@xxxxxxxxxxxxxxx/ The counterpart userspace tools changes and tests are here: - util-linux: https://lore.kernel.org/linux-fsdevel/20250318073218.3513262-1-yi.zhang@xxxxxxxxxxxxxxx/ - xfsprogs: https://lore.kernel.org/linux-fsdevel/20250318072318.3502037-1-yi.zhang@xxxxxxxxxxxxxxx/ - xfstests: https://lore.kernel.org/linux-fsdevel/20250318072615.3505873-1-yi.zhang@xxxxxxxxxxxxxxx/ - blktests: https://lore.kernel.org/linux-fsdevel/20250318072835.3508696-1-yi.zhang@xxxxxxxxxxxxxxx/ Currently, we can use the fallocate command to quickly create a pre-allocated file. However, on most filesystems, such as ext4 and XFS, fallocate create pre-allocation blocks in an unwritten state, and the FALLOC_FL_ZERO_RANGE flag also behaves similarly. The extent state must be converted to a written state when the user writes data into this range later, which can trigger numerous metadata changes and consequent journal I/O. This may leads to significant write amplification and performance degradation in synchronous write mode. Therefore, we need a method to create a pre-allocated file with written extents that can be used for pure overwriting. At the monent, the only method available is to create an empty file and write zero data into it (for example, using 'dd' with a large block size). However, this method is slow and consumes a considerable amount of disk bandwidth, we must pre-allocate files in advance but cannot add pre-allocated files while user business services are running. Fortunately, with the development and more and more widely used of flash-based storage devices, we can efficiently write zeros to SSDs using the unmap write zeroes command if the devices do not write physical zeroes to the media. For example, if SCSI SSDs support the UMMAP bit or NVMe SSDs support the DEAC bit[1], the write zeroes command does not write actual data to the device, instead, NVMe converts the zeroed range to a deallocated state, which works fast and consumes almost no disk write bandwidth. Consequently, this feature can provide us with a faster method for creating pre-allocated files with written extents and zeroed data. This series aims to implement this by: 1. Introduce a new feature BLK_FEAT_WRITE_ZEROES_UNMAP to the block device queue limit features, which indicates whether the storage is device explicitly supports the unmapped write zeroes command. This flag should be set to 1 by the driver if the attached disk supports this command. Users can check this flag by querying: /sys/block/<disk>/queue/write_zeroes_unmap 2. Introduce a new flag FALLOC_FL_WRITE_ZEROES into the fallocate, filesystems with this operaion should allocate written extents and issuing zeroes to the range of the device. If the device supports unmap write zeroes command, the zeroing can be accelerated, if not, we currently still allow to fall back to submit zeroes data. Users can verify if the device supports the unmap write zeroes command and then decide whether to use it. This series implemented the BLK_FEAT_WRITE_ZEROES_UNMAP flag for SCSI, NVMe and device-mapper drivers, and added the FALLOC_FL_WRITE_ZEROES support for ext4 and raw bdev devices. Any comments are welcome. I've tested performance with this series on ext4 filesystem on my machine with an Intel Xeon Gold 6248R CPU, a 7TB KCD61LUL7T68 NVMe SSD which supports unmap write zeroes command with the Deallocated state and the DEAC bit. Feel free to give it a try. 0. Ensure the NVMe device supports WRITE_ZERO command. $ cat /sys/block/nvme5n1/queue/write_zeroes_max_bytes 8388608 $ nvme id-ns -H /dev/nvme5n1 | grep -i -A 3 "dlfeat" dlfeat : 25 [4:4] : 0x1 Guard Field of Deallocated Logical Blocks is set to CRC of The Value Read [3:3] : 0x1 Deallocate Bit in the Write Zeroes Command is Supported [2:0] : 0x1 Bytes Read From a Deallocated Logical Block and its Metadata are 0x00 1. Compare 'dd' and fallocate with unmap write zeroes, the later one is significantly faster than 'dd'. Create a 1GB and 10GB zeroed file. $dd if=/dev/zero of=foo bs=2M count=$count oflag=direct $time fallocate -w -l $size bar #1G dd: 0.5s FALLOC_FL_WRITE_ZEROES: 0.17s #10G dd: 5.0s FALLOC_FL_WRITE_ZEROES: 1.7s 2. Run fio overwrite and fallocate with unmap write zeroes simultaneously, fallocate has little impact on write bandwidth and only slightly affects write latency. a) Test bandwidth costs. $ fio -directory=/test -direct=1 -iodepth=10 -fsync=0 -rw=write \ -numjobs=10 -bs=2M -ioengine=libaio -size=20G -runtime=20 \ -fallocate=none -overwrite=1 -group_reportin -name=bw_test Without background zero range: bw (MiB/s): min= 2068, max= 2280, per=100.00%, avg=2186.40 With background zero range: bw (MiB/s): min= 2056, max= 2308, per=100.00%, avg=2186.20 b) Test write latency costs. $ fio -filename=/test/foo -direct=1 -iodepth=1 -fsync=0 -rw=write \ -numjobs=1 -bs=4k -ioengine=psync -size=5G -runtime=20 \ -fallocate=none -overwrite=1 -group_reportin -name=lat_test Without background zero range: lat (nsec): min=9269, max=71635, avg=9840.65 With a background zero range: lat (usec): min=9, max=982, avg=11.03 3. Compare overwriting in a pre-allocated unwritten file and a written file in O_DSYNC mode. Write to a file with written extents is much faster. # First mkfs and create a test file according to below three cases, # and then run fio. $ fio -filename=/test/foo -direct=1 -iodepth=1 -fdatasync=1 \ -rw=write -numjobs=1 -bs=4k -ioengine=psync -size=5G \ -runtime=20 -fallocate=none -group_reportin -name=test unwritten file: IOPS=20.1k, BW=78.7MiB/s unwritten file + fast_commit: IOPS=42.9k, BW=167MiB/s written file: IOPS=98.8k, BW=386MiB/s Thanks, Yi. --- [1] https://nvmexpress.org/specifications/ NVM Command Set Specification, section 3.2.8 Zhang Yi (10): block: introduce BLK_FEAT_WRITE_ZEROES_UNMAP to queue limits features nvme: set BLK_FEAT_WRITE_ZEROES_UNMAP if device supports DEAC bit nvme-multipath: add BLK_FEAT_WRITE_ZEROES_UNMAP support nvmet: set WZDS and DRB if device supports BLK_FEAT_WRITE_ZEROES_UNMAP scsi: sd: set BLK_FEAT_WRITE_ZEROES_UNMAP if device supports unmap zeroing mode dm: add BLK_FEAT_WRITE_ZEROES_UNMAP support fs: introduce FALLOC_FL_WRITE_ZEROES to fallocate block: add FALLOC_FL_WRITE_ZEROES support block: factor out common part in blkdev_fallocate() ext4: add FALLOC_FL_WRITE_ZEROES support Documentation/ABI/stable/sysfs-block | 14 +++++++ block/blk-settings.c | 6 +++ block/blk-sysfs.c | 3 ++ block/fops.c | 37 +++++++++-------- drivers/md/dm-table.c | 7 +++- drivers/md/dm.c | 1 + drivers/nvme/host/core.c | 21 +++++----- drivers/nvme/host/multipath.c | 3 +- drivers/nvme/target/io-cmd-bdev.c | 4 ++ drivers/scsi/sd.c | 5 +++ fs/ext4/extents.c | 59 ++++++++++++++++++++++------ fs/open.c | 1 + include/linux/blkdev.h | 8 ++++ include/linux/falloc.h | 3 +- include/trace/events/ext4.h | 3 +- include/uapi/linux/falloc.h | 18 +++++++++ 16 files changed, 149 insertions(+), 44 deletions(-) -- 2.46.1
