Merge patch series "fallocate: introduce FALLOC_FL_WRITE_ZEROES flag"

Zhang Yi <[email protected]> says:

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. However, please note that this may be a
best-effort optimization rather than a mandatory requirement, some
devices may partially fall back to writing physical zeroes due to
factors such as receiving unaligned commands.

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.

2. Introduce a queue limit flag, BLK_FLAG_WRITE_ZEROES_UNMAP_DISABLED,
   along with a corresponding sysfs entry. Users can query the support
   status of the unmap write zeroes operation and disable this operation
   if the write zeroes operation is very slow.

       /sys/block/<disk>/queue/write_zeroes_unmap

3. Introduce a new flag, FALLOC_FL_WRITE_ZEROES, into the fallocate.
   Filesystems that support this operation should allocate written
   extents and issue zeroes to the specified range of the device. For
   local block device filesystems, this operation should depend on the
   write_zeroes_unmap operaion of the underlying block device. It should
   return -EOPNOTSUPP if the device doesn't enable unmap write zeroes
   operaion.

This series implements the BLK_FEAT_WRITE_ZEROES_UNMAP feature and
BLK_FLAG_WRITE_ZEROES_UNMAP_DISABLED flag for SCSI, NVMe and
device-mapper drivers, and add the FALLOC_FL_WRITE_ZEROES and
STATX_ATTR_WRITE_ZEROES_UNMAP 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

* patches from https://lore.kernel.org/[email protected]:
  ext4: add FALLOC_FL_WRITE_ZEROES support
  block: add FALLOC_FL_WRITE_ZEROES support
  block: factor out common part in blkdev_fallocate()
  fs: introduce FALLOC_FL_WRITE_ZEROES to fallocate
  dm: clear unmap write zeroes limits when disabling write zeroes
  scsi: sd: set max_hw_wzeroes_unmap_sectors if device supports SD_ZERO_*_UNMAP
  nvmet: set WZDS and DRB if device enables unmap write zeroes operation
  nvme: set max_hw_wzeroes_unmap_sectors if device supports DEAC bit
  block: introduce max_{hw|user}_wzeroes_unmap_sectors to queue limits

Link: https://lore.kernel.org/[email protected]
Signed-off-by: Christian Brauner <[email protected]>

Author: brauner

Documentation/ABI/stable/sysfs-block

@@ -778,6 +778,39 @@ Description:
 		0, write zeroes is not supported by the device.
 
 
+What:		/sys/block/<disk>/queue/write_zeroes_unmap_max_hw_bytes
+Date:		January 2025
+Contact:	Zhang Yi <[email protected]>
+Description:
+		[RO] This file indicates whether a device supports zeroing data
+		in a specified block range without incurring the cost of
+		physically writing zeroes to the media for each individual
+		block. If this parameter is set to write_zeroes_max_bytes, the
+		device implements a zeroing operation which opportunistically
+		avoids writing zeroes to media while still guaranteeing that
+		subsequent reads from the specified block range will return
+		zeroed data. This operation is a best-effort optimization, a
+		device may fall back to physically writing zeroes to the media
+		due to other factors such as misalignment or being asked to
+		clear a block range smaller than the device's internal
+		allocation unit. If this parameter is set to 0, the device may
+		have to write each logical block media during a zeroing
+		operation.
+
+
+What:		/sys/block/<disk>/queue/write_zeroes_unmap_max_bytes
+Date:		January 2025
+Contact:	Zhang Yi <[email protected]>
+Description:
+		[RW] While write_zeroes_unmap_max_hw_bytes is the hardware limit
+		for the device, this setting is the software limit. Since the
+		unmap write zeroes operation is a best-effort optimization, some
+		devices may still physically writing zeroes to media. So the
+		speed of this operation is not guaranteed. Writing a value of
+		'0' to this file disables this operation. Otherwise, this
+		parameter should be equal to write_zeroes_unmap_max_hw_bytes.
+
+
 What:		/sys/block/<disk>/queue/zone_append_max_bytes
 Date:		May 2020
 Contact:	[email protected]

block/blk-settings.c

@@ -50,6 +50,8 @@ void blk_set_stacking_limits(struct queue_limits *lim)
 	lim->max_sectors = UINT_MAX;
 	lim->max_dev_sectors = UINT_MAX;
 	lim->max_write_zeroes_sectors = UINT_MAX;
+	lim->max_hw_wzeroes_unmap_sectors = UINT_MAX;
+	lim->max_user_wzeroes_unmap_sectors = UINT_MAX;
 	lim->max_hw_zone_append_sectors = UINT_MAX;
 	lim->max_user_discard_sectors = UINT_MAX;
 }
@@ -333,6 +335,12 @@ int blk_validate_limits(struct queue_limits *lim)
 	if (!lim->max_segments)
 		lim->max_segments = BLK_MAX_SEGMENTS;
 
+	if (lim->max_hw_wzeroes_unmap_sectors &&
+	    lim->max_hw_wzeroes_unmap_sectors != lim->max_write_zeroes_sectors)
+		return -EINVAL;
+	lim->max_wzeroes_unmap_sectors = min(lim->max_hw_wzeroes_unmap_sectors,
+			lim->max_user_wzeroes_unmap_sectors);
+
 	lim->max_discard_sectors =
 		min(lim->max_hw_discard_sectors, lim->max_user_discard_sectors);
 
@@ -418,10 +426,11 @@ int blk_set_default_limits(struct queue_limits *lim)
 {
 	/*
 	 * Most defaults are set by capping the bounds in blk_validate_limits,
-	 * but max_user_discard_sectors is special and needs an explicit
-	 * initialization to the max value here.
+	 * but these limits are special and need an explicit initialization to
+	 * the max value here.
 	 */
 	lim->max_user_discard_sectors = UINT_MAX;
+	lim->max_user_wzeroes_unmap_sectors = UINT_MAX;
 	return blk_validate_limits(lim);
 }
 
@@ -708,6 +717,13 @@ int blk_stack_limits(struct queue_limits *t, struct queue_limits *b,
 	t->max_dev_sectors = min_not_zero(t->max_dev_sectors, b->max_dev_sectors);
 	t->max_write_zeroes_sectors = min(t->max_write_zeroes_sectors,
 					b->max_write_zeroes_sectors);
+	t->max_user_wzeroes_unmap_sectors =
+			min(t->max_user_wzeroes_unmap_sectors,
+			    b->max_user_wzeroes_unmap_sectors);
+	t->max_hw_wzeroes_unmap_sectors =
+			min(t->max_hw_wzeroes_unmap_sectors,
+			    b->max_hw_wzeroes_unmap_sectors);
+
 	t->max_hw_zone_append_sectors = min(t->max_hw_zone_append_sectors,
 					b->max_hw_zone_append_sectors);
 

block/blk-sysfs.c

@@ -161,6 +161,8 @@ static ssize_t queue_##_field##_show(struct gendisk *disk, char *page)	\
 QUEUE_SYSFS_LIMIT_SHOW_SECTORS_TO_BYTES(max_discard_sectors)
 QUEUE_SYSFS_LIMIT_SHOW_SECTORS_TO_BYTES(max_hw_discard_sectors)
 QUEUE_SYSFS_LIMIT_SHOW_SECTORS_TO_BYTES(max_write_zeroes_sectors)
+QUEUE_SYSFS_LIMIT_SHOW_SECTORS_TO_BYTES(max_hw_wzeroes_unmap_sectors)
+QUEUE_SYSFS_LIMIT_SHOW_SECTORS_TO_BYTES(max_wzeroes_unmap_sectors)
 QUEUE_SYSFS_LIMIT_SHOW_SECTORS_TO_BYTES(atomic_write_max_sectors)
 QUEUE_SYSFS_LIMIT_SHOW_SECTORS_TO_BYTES(atomic_write_boundary_sectors)
 QUEUE_SYSFS_LIMIT_SHOW_SECTORS_TO_BYTES(max_zone_append_sectors)
@@ -205,6 +207,24 @@ static int queue_max_discard_sectors_store(struct gendisk *disk,
 	return 0;
 }
 
+static int queue_max_wzeroes_unmap_sectors_store(struct gendisk *disk,
+		const char *page, size_t count, struct queue_limits *lim)
+{
+	unsigned long max_zeroes_bytes, max_hw_zeroes_bytes;
+	ssize_t ret;
+
+	ret = queue_var_store(&max_zeroes_bytes, page, count);
+	if (ret < 0)
+		return ret;
+
+	max_hw_zeroes_bytes = lim->max_hw_wzeroes_unmap_sectors << SECTOR_SHIFT;
+	if (max_zeroes_bytes != 0 && max_zeroes_bytes != max_hw_zeroes_bytes)
+		return -EINVAL;
+
+	lim->max_user_wzeroes_unmap_sectors = max_zeroes_bytes >> SECTOR_SHIFT;
+	return 0;
+}
+
 static int
 queue_max_sectors_store(struct gendisk *disk, const char *page, size_t count,
 		struct queue_limits *lim)
@@ -514,6 +534,10 @@ QUEUE_LIM_RO_ENTRY(queue_atomic_write_unit_min, "atomic_write_unit_min_bytes");
 
 QUEUE_RO_ENTRY(queue_write_same_max, "write_same_max_bytes");
 QUEUE_LIM_RO_ENTRY(queue_max_write_zeroes_sectors, "write_zeroes_max_bytes");
+QUEUE_LIM_RO_ENTRY(queue_max_hw_wzeroes_unmap_sectors,
+		"write_zeroes_unmap_max_hw_bytes");
+QUEUE_LIM_RW_ENTRY(queue_max_wzeroes_unmap_sectors,
+		"write_zeroes_unmap_max_bytes");
 QUEUE_LIM_RO_ENTRY(queue_max_zone_append_sectors, "zone_append_max_bytes");
 QUEUE_LIM_RO_ENTRY(queue_zone_write_granularity, "zone_write_granularity");
 
@@ -662,6 +686,8 @@ static struct attribute *queue_attrs[] = {
 	&queue_atomic_write_unit_min_entry.attr,
 	&queue_atomic_write_unit_max_entry.attr,
 	&queue_max_write_zeroes_sectors_entry.attr,
+	&queue_max_hw_wzeroes_unmap_sectors_entry.attr,
+	&queue_max_wzeroes_unmap_sectors_entry.attr,
 	&queue_max_zone_append_sectors_entry.attr,
 	&queue_zone_write_granularity_entry.attr,
 	&queue_rotational_entry.attr,

block/fops.c

@@ -841,7 +841,7 @@ static ssize_t blkdev_read_iter(struct kiocb *iocb, struct iov_iter *to)
 
 #define	BLKDEV_FALLOC_FL_SUPPORTED					\
 		(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |		\
-		 FALLOC_FL_ZERO_RANGE)
+		 FALLOC_FL_ZERO_RANGE | FALLOC_FL_WRITE_ZEROES)
 
 static long blkdev_fallocate(struct file *file, int mode, loff_t start,
 			     loff_t len)
@@ -850,11 +850,19 @@ static long blkdev_fallocate(struct file *file, int mode, loff_t start,
 	struct block_device *bdev = I_BDEV(inode);
 	loff_t end = start + len - 1;
 	loff_t isize;
+	unsigned int flags;
 	int error;
 
 	/* Fail if we don't recognize the flags. */
 	if (mode & ~BLKDEV_FALLOC_FL_SUPPORTED)
 		return -EOPNOTSUPP;
+	/*
+	 * Don't allow writing zeroes if the device does not enable the
+	 * unmap write zeroes operation.
+	 */
+	if ((mode & FALLOC_FL_WRITE_ZEROES) &&
+	    !bdev_write_zeroes_unmap_sectors(bdev))
+		return -EOPNOTSUPP;
 
 	/* Don't go off the end of the device. */
 	isize = bdev_nr_bytes(bdev);
@@ -877,34 +885,32 @@ static long blkdev_fallocate(struct file *file, int mode, loff_t start,
 	inode_lock(inode);
 	filemap_invalidate_lock(inode->i_mapping);
 
-	/*
-	 * Invalidate the page cache, including dirty pages, for valid
-	 * de-allocate mode calls to fallocate().
-	 */
 	switch (mode) {
 	case FALLOC_FL_ZERO_RANGE:
 	case FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE:
-		error = truncate_bdev_range(bdev, file_to_blk_mode(file), start, end);
-		if (error)
-			goto fail;
-
-		error = blkdev_issue_zeroout(bdev, start >> SECTOR_SHIFT,
-					     len >> SECTOR_SHIFT, GFP_KERNEL,
-					     BLKDEV_ZERO_NOUNMAP);
+		flags = BLKDEV_ZERO_NOUNMAP;
 		break;
 	case FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE:
-		error = truncate_bdev_range(bdev, file_to_blk_mode(file), start, end);
-		if (error)
-			goto fail;
-
-		error = blkdev_issue_zeroout(bdev, start >> SECTOR_SHIFT,
-					     len >> SECTOR_SHIFT, GFP_KERNEL,
-					     BLKDEV_ZERO_NOFALLBACK);
+		flags = BLKDEV_ZERO_NOFALLBACK;
+		break;
+	case FALLOC_FL_WRITE_ZEROES:
+		flags = 0;
 		break;
 	default:
 		error = -EOPNOTSUPP;
+		goto fail;
 	}
 
+	/*
+	 * Invalidate the page cache, including dirty pages, for valid
+	 * de-allocate mode calls to fallocate().
+	 */
+	error = truncate_bdev_range(bdev, file_to_blk_mode(file), start, end);
+	if (error)
+		goto fail;
+
+	error = blkdev_issue_zeroout(bdev, start >> SECTOR_SHIFT,
+				     len >> SECTOR_SHIFT, GFP_KERNEL, flags);
  fail:
 	filemap_invalidate_unlock(inode->i_mapping);
 	inode_unlock(inode);

drivers/md/dm-table.c

@@ -2065,8 +2065,10 @@ int dm_table_set_restrictions(struct dm_table *t, struct request_queue *q,
 		limits->discard_alignment = 0;
 	}
 
-	if (!dm_table_supports_write_zeroes(t))
+	if (!dm_table_supports_write_zeroes(t)) {
 		limits->max_write_zeroes_sectors = 0;
+		limits->max_hw_wzeroes_unmap_sectors = 0;
+	}
 
 	if (!dm_table_supports_secure_erase(t))
 		limits->max_secure_erase_sectors = 0;

drivers/nvme/host/core.c

@@ -2420,22 +2420,24 @@ static int nvme_update_ns_info_block(struct nvme_ns *ns,
 	else
 		lim.write_stream_granularity = 0;
 
-	ret = queue_limits_commit_update(ns->disk->queue, &lim);
-	if (ret) {
-		blk_mq_unfreeze_queue(ns->disk->queue, memflags);
-		goto out;
-	}
-
-	set_capacity_and_notify(ns->disk, capacity);
-
 	/*
 	 * Only set the DEAC bit if the device guarantees that reads from
 	 * deallocated data return zeroes.  While the DEAC bit does not
 	 * require that, it must be a no-op if reads from deallocated data
 	 * do not return zeroes.
 	 */
-	if ((id->dlfeat & 0x7) == 0x1 && (id->dlfeat & (1 << 3)))
+	if ((id->dlfeat & 0x7) == 0x1 && (id->dlfeat & (1 << 3))) {
 		ns->head->features |= NVME_NS_DEAC;
+		lim.max_hw_wzeroes_unmap_sectors = lim.max_write_zeroes_sectors;
+	}
+
+	ret = queue_limits_commit_update(ns->disk->queue, &lim);
+	if (ret) {
+		blk_mq_unfreeze_queue(ns->disk->queue, memflags);
+		goto out;
+	}
+
+	set_capacity_and_notify(ns->disk, capacity);
 	set_disk_ro(ns->disk, nvme_ns_is_readonly(ns, info));
 	set_bit(NVME_NS_READY, &ns->flags);
 	blk_mq_unfreeze_queue(ns->disk->queue, memflags);

drivers/nvme/target/io-cmd-bdev.c

@@ -46,6 +46,10 @@ void nvmet_bdev_set_limits(struct block_device *bdev, struct nvme_id_ns *id)
 	id->npda = id->npdg;
 	/* NOWS = Namespace Optimal Write Size */
 	id->nows = to0based(bdev_io_opt(bdev) / bdev_logical_block_size(bdev));
+
+	/* Set WZDS and DRB if device supports unmapped write zeroes */
+	if (bdev_write_zeroes_unmap_sectors(bdev))
+		id->dlfeat = (1 << 3) | 0x1;
 }
 
 void nvmet_bdev_ns_disable(struct nvmet_ns *ns)

drivers/scsi/sd.c

@@ -1141,6 +1141,11 @@ static void sd_config_write_same(struct scsi_disk *sdkp,
 out:
 	lim->max_write_zeroes_sectors =
 		sdkp->max_ws_blocks * (logical_block_size >> SECTOR_SHIFT);
+
+	if (sdkp->zeroing_mode == SD_ZERO_WS16_UNMAP ||
+	    sdkp->zeroing_mode == SD_ZERO_WS10_UNMAP)
+		lim->max_hw_wzeroes_unmap_sectors =
+				lim->max_write_zeroes_sectors;
 }
 
 static blk_status_t sd_setup_flush_cmnd(struct scsi_cmnd *cmd)