Skip to contentSkip to conversation

xfs: add FALLOC_FL_WRITE_ZEROES to XFS code base

RFC PATCH In review Lukas Herbolt 4 participants 7 messages Oct 2, 2025 – Oct 15, 2025 XFS filesystem

lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 xfs: add FALLOC_FL_WRITE_ZEROES to XFS code base

Add support for FALLOC_FL_WRITE_ZEROES if the underlying device enable the unmap write zeroes operation.

Inspired by the Ext4 implementation of the FALLOC_FL_WRITE_ZEROES. It can speed up some patterns on specific hardware.

time ( ./fallocate -l 360M /mnt/test.file; dd if=/dev/zero of=/mnt/test.file \
bs=1M count=360 conv=notrunc,nocreat oflag=direct,dsync)

360+0 records in
360+0 records out
377487360 bytes (377 MB, 360 MiB) copied, 22.0027 s, 17.2 MB/s

real 0m22.114s
user 0m0.006s
sys 0m3.085s

time (./fallocate -wl 360M /mnt/test.file; dd if=/dev/zero of=/mnt/test.file \
bs=1M count=360 conv=notrunc,nocreat oflag=direct,dsync );
360+0 records in
360+0 records out
377487360 bytes (377 MB, 360 MiB) copied, 2.02512 s, 186 MB/s

real 0m6.384s
user 0m0.002s
sys 0m5.823s

---
 fs/xfs/xfs_bmap_util.c |  5 +++--
 fs/xfs/xfs_bmap_util.h |  4 ++--
 fs/xfs/xfs_file.c      | 23 ++++++++++++++++++-----
 3 files changed, 23 insertions(+), 9 deletions(-)
fs/xfs/xfs_bmap_util.c +3 -2
index 06ca11731e430..a91596a280ba5 100644
@@ -645,6 +645,7 @@ xfs_free_eofblocks(
 int
 xfs_alloc_file_space(
 	struct xfs_inode	*ip,
+    uint32_t        flags,      /* XFS_BMAPI_... */
 	xfs_off_t		offset,
 	xfs_off_t		len)
 {
@@ -748,8 +749,8 @@ xfs_alloc_file_space(
 		 * will eventually reach the requested range.
 		 */
 		error = xfs_bmapi_write(tp, ip, startoffset_fsb,
-				allocatesize_fsb, XFS_BMAPI_PREALLOC, 0, imapp,
-				&nimaps);
+				allocatesize_fsb, flags, 0, imapp, &nimaps);
+
 		if (error) {
 			if (error != -ENOSR)
 				goto error;
fs/xfs/xfs_bmap_util.h +2 -2
index c477b33616304..67770830eb245 100644
@@ -55,8 +55,8 @@ int	xfs_bmap_last_extent(struct xfs_trans *tp, struct xfs_inode *ip,
 			     int *is_empty);
 
 /* preallocation and hole punch interface */
-int	xfs_alloc_file_space(struct xfs_inode *ip, xfs_off_t offset,
-		xfs_off_t len);
+int	xfs_alloc_file_space(struct xfs_inode *ip, uint32_t flags,
+		xfs_off_t offset, xfs_off_t len);
 int	xfs_free_file_space(struct xfs_inode *ip, xfs_off_t offset,
 		xfs_off_t len, struct xfs_zone_alloc_ctx *ac);
 int	xfs_collapse_file_space(struct xfs_inode *, xfs_off_t offset,
fs/xfs/xfs_file.c +18 -5
index f96fbf5c54c99..48559a011e9b4 100644
@@ -1255,7 +1255,7 @@ xfs_falloc_insert_range(
 static int
 xfs_falloc_zero_range(
 	struct file		*file,
-	int			mode,
+	int				mode,
 	loff_t			offset,
 	loff_t			len,
 	struct xfs_zone_alloc_ctx *ac)
@@ -1277,7 +1277,16 @@ xfs_falloc_zero_range(
 
 	len = round_up(offset + len, blksize) - round_down(offset, blksize);
 	offset = round_down(offset, blksize);
-	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
+	if (mode & FALLOC_FL_WRITE_ZEROES) {
+		if (!bdev_write_zeroes_unmap_sectors(inode->i_sb->s_bdev))
+	        return -EOPNOTSUPP;
+		error = xfs_alloc_file_space(XFS_I(inode), XFS_BMAPI_ZERO,
+				offset, len);
+	}
+	else
+		error = xfs_alloc_file_space(XFS_I(inode), XFS_BMAPI_PREALLOC,
+				offset, len);
+
 	if (error)
 		return error;
 	return xfs_falloc_setsize(file, new_size);
@@ -1302,7 +1311,8 @@ xfs_falloc_unshare_range(
 	if (error)
 		return error;
 
-	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
+	error = xfs_alloc_file_space(XFS_I(inode), XFS_BMAPI_PREALLOC,
+			offset, len);
 	if (error)
 		return error;
 	return xfs_falloc_setsize(file, new_size);
@@ -1330,7 +1340,9 @@ xfs_falloc_allocate_range(
 	if (error)
 		return error;
 
-	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
+	error = xfs_alloc_file_space(XFS_I(inode), XFS_BMAPI_PREALLOC,
+			offset, len);
+
 	if (error)
 		return error;
 	return xfs_falloc_setsize(file, new_size);
@@ -1340,7 +1352,7 @@ xfs_falloc_allocate_range(
 		(FALLOC_FL_ALLOCATE_RANGE | FALLOC_FL_KEEP_SIZE |	\
 		 FALLOC_FL_PUNCH_HOLE |	FALLOC_FL_COLLAPSE_RANGE |	\
 		 FALLOC_FL_ZERO_RANGE |	FALLOC_FL_INSERT_RANGE |	\
-		 FALLOC_FL_UNSHARE_RANGE)
+		 FALLOC_FL_UNSHARE_RANGE | FALLOC_FL_WRITE_ZEROES )
 
 STATIC long
 __xfs_file_fallocate(
@@ -1383,6 +1395,7 @@ __xfs_file_fallocate(
 	case FALLOC_FL_INSERT_RANGE:
 		error = xfs_falloc_insert_range(file, offset, len);
 		break;
+	case FALLOC_FL_WRITE_ZEROES:
 	case FALLOC_FL_ZERO_RANGE:
 		error = xfs_falloc_zero_range(file, mode, offset, len, ac);
 		break;
Signature · 1 line
2.51.0
#1 lore raw

fs/xfs/xfs_bmap_util.c xfs_free_eofblocks( · line 648

Quoted 4 lines · Lukas Herbolt

int
xfs_alloc_file_space(

fs/xfs/xfs_bmap_util.c +1 -0
 	struct xfs_inode	*ip,
+    uint32_t        flags,      /* XFS_BMAPI_... */

This seems to mix tabs and spaces.

fs/xfs/xfs_file.c xfs_falloc_insert_range( · line 1258

Quoted 5 lines

static int
xfs_falloc_zero_range(

fs/xfs/xfs_file.c +1 -1
 	struct file		*file,
-	int			mode,
+	int				mode,

More whitespace damage here.

Quoted 5 lines
fs/xfs/xfs_file.c +0 -0
 	loff_t			offset,
 	loff_t			len,
 	struct xfs_zone_alloc_ctx *ac)
@@ -1277,7 +1277,16 @@ xfs_falloc_zero_range(
 

fs/xfs/xfs_file.c xfs_falloc_zero_range( · line 1286

Quoted 10 lines
fs/xfs/xfs_file.c +7 -1
 	len = round_up(offset + len, blksize) - round_down(offset, blksize);
 	offset = round_down(offset, blksize);
-	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
+	if (mode & FALLOC_FL_WRITE_ZEROES) {
+		if (!bdev_write_zeroes_unmap_sectors(inode->i_sb->s_bdev))
+	        return -EOPNOTSUPP;
+		error = xfs_alloc_file_space(XFS_I(inode), XFS_BMAPI_ZERO,
+				offset, len);
+	}
+	else

The closing brace goes onto the same line as the else, and we usually add the brace in all branches.

Otherwise this looks fine.

#2 lore raw
Quoted 21 lines · Christoph Hellwig

fs/xfs/xfs_bmap_util.c xfs_free_eofblocks( · line 648

Quoted 4 lines · Lukas Herbolt

int
xfs_alloc_file_space(

fs/xfs/xfs_bmap_util.c +1 -0
 	struct xfs_inode	*ip,
+    uint32_t        flags,      /* XFS_BMAPI_... */

This seems to mix tabs and spaces.

fs/xfs/xfs_file.c xfs_falloc_insert_range( · line 1258

Quoted 5 lines

static int
xfs_falloc_zero_range(

fs/xfs/xfs_file.c +1 -1
 	struct file		*file,
-	int			mode,
+	int				mode,

More whitespace damage here.

Quoted 5 lines
fs/xfs/xfs_file.c +0 -0
 	loff_t			offset,
 	loff_t			len,
 	struct xfs_zone_alloc_ctx *ac)
@@ -1277,7 +1277,16 @@ xfs_falloc_zero_range(
 
Quoted 5 lines

fs/xfs/xfs_file.c xfs_falloc_zero_range( · line 1281

Quoted 5 lines
fs/xfs/xfs_file.c +2 -1
 	len = round_up(offset + len, blksize) - round_down(offset, blksize);
 	offset = round_down(offset, blksize);
-	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
+	if (mode & FALLOC_FL_WRITE_ZEROES) {
+		if (!bdev_write_zeroes_unmap_sectors(inode->i_sb->s_bdev))
     		                         not correct ^^^^^^^^^^^^^^^^^^^

You need to find the block device associated with the file because XFS is a multi-device filesystem.

--D

Quoted 11 lines

fs/xfs/xfs_file.c xfs_falloc_zero_range( · line 1286

Quoted 5 lines
fs/xfs/xfs_file.c +5 -0
+	        return -EOPNOTSUPP;
+		error = xfs_alloc_file_space(XFS_I(inode), XFS_BMAPI_ZERO,
+				offset, len);
+	}
+	else

The closing brace goes onto the same line as the else, and we usually add the brace in all branches.

Otherwise this looks fine.

#3 lore raw
Quoted 8 lines · Darrick J. Wong
Quoted 3 lines

fs/xfs/xfs_file.c xfs_falloc_zero_range( · line 1281

Quoted 3 lines
fs/xfs/xfs_file.c +2 -1
-	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
+	if (mode & FALLOC_FL_WRITE_ZEROES) {
+		if (!bdev_write_zeroes_unmap_sectors(inode->i_sb->s_bdev))
     		                         not correct ^^^^^^^^^^^^^^^^^^^

You need to find the block device associated with the file because XFS is a multi-device filesystem.

Indeed. xfs_inode_buftarg will do the work, but we'll need to ensure the RT bit doesn't get flipped, i.e. it needs to hold a lock between that check and allocation the blocks if there were none yet.

#4 lore raw
Quoted 11 lines · Christoph Hellwig
Quoted 8 lines · Darrick J. Wong
Quoted 3 lines

fs/xfs/xfs_file.c xfs_falloc_zero_range( · line 1281

Quoted 3 lines
fs/xfs/xfs_file.c +2 -1
-	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
+	if (mode & FALLOC_FL_WRITE_ZEROES) {
+		if (!bdev_write_zeroes_unmap_sectors(inode->i_sb->s_bdev))
     		                         not correct ^^^^^^^^^^^^^^^^^^^

You need to find the block device associated with the file because XFS is a multi-device filesystem.

Indeed. xfs_inode_buftarg will do the work, but we'll need to ensure

Thanks for xfs_inode_buftarg pointer.

Quoted 2 lines

the RT bit doesn't get flipped, i.e. it needs to hold a lock between that check and allocation the blocks if there were none yet.

I am having bit of trouble with that. If I get it right we should hold the XFS_ILOCK_EXCL but this lock is then grabbed in the xfs_trans_alloc_inode.

So I would need to release before and there would be again a small window where the RT flag can be flipped.

Looking at the xfs_alloc_file_space, there is also check for the RT bit without lock, so this also need an attention.

    rt = XFS_IS_REALTIME_INODE(ip);
    extsz = xfs_get_extsz_hint(ip);

Or the xfs_trans_alloc_inode would need to check if we are a;ready holding the lock Is there a way how to check the current thread is the owner of the xfs_ilock rw_sem?

Something like?
---
static inline bool is_current_writer(struct rw_semaphore *sem)

{
    /* The rwsem_owner() helper can be used to get the owner */
    return sem->owner == current;
}

---

#5 lore raw
Quoted 28 lines · On Tue, Oct 14, 2025 at 10:10:51AM +0200, [email protected] wrote:
Quoted 11 lines · Christoph Hellwig
Quoted 8 lines · Darrick J. Wong
Quoted 3 lines

fs/xfs/xfs_file.c xfs_falloc_zero_range( · line 1281

Quoted 3 lines
fs/xfs/xfs_file.c +2 -1
-	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
+	if (mode & FALLOC_FL_WRITE_ZEROES) {
+		if (!bdev_write_zeroes_unmap_sectors(inode->i_sb->s_bdev))
     		                         not correct ^^^^^^^^^^^^^^^^^^^

You need to find the block device associated with the file because XFS is a multi-device filesystem.

Indeed. xfs_inode_buftarg will do the work, but we'll need to ensure

Thanks for xfs_inode_buftarg pointer.

Quoted 2 lines

the RT bit doesn't get flipped, i.e. it needs to hold a lock between that check and allocation the blocks if there were none yet.

I am having bit of trouble with that. If I get it right we should hold the XFS_ILOCK_EXCL but this lock is then grabbed in the xfs_trans_alloc_inode.

So I would need to release before and there would be again a small window where the RT flag can be flipped.

Looking at the xfs_alloc_file_space, there is also check for the RT bit without
lock, so this also need an attention.

Quoted 7 lines
    rt = XFS_IS_REALTIME_INODE(ip);
    extsz = xfs_get_extsz_hint(ip);

Or the xfs_trans_alloc_inode would need to check if we are a;ready holding the
lock Is there a way how to check the current thread is the owner of the xfs_ilock rw_sem?

Hm? The VFS takes inode_lock in vfs_fileattr_set prior to changing the rt status bit, and and __xfs_file_fallocate takes the same lock (aka the IOLOCK) before calling xfs_falloc_*.

IOWs, a call to xfs_inode_buftarg would see a stable realtime flag and return the correct results.

--D

Quoted 10 lines

Something like?
---
static inline bool is_current_writer(struct rw_semaphore *sem)

{
    /* The rwsem_owner() helper can be used to get the owner */
    return sem->owner == current;
}

---

#6 lore raw
In reply to Lukas Herbolt
Quoted 41 lines · Lukas Herbolt

Add support for FALLOC_FL_WRITE_ZEROES if the underlying device enable the unmap write zeroes operation.

Inspired by the Ext4 implementation of the FALLOC_FL_WRITE_ZEROES. It can speed up some patterns on specific hardware.

time ( ./fallocate -l 360M /mnt/test.file; dd if=/dev/zero of=/mnt/test.file \
bs=1M count=360 conv=notrunc,nocreat oflag=direct,dsync)

360+0 records in
360+0 records out
377487360 bytes (377 MB, 360 MiB) copied, 22.0027 s, 17.2 MB/s

real 0m22.114s
user 0m0.006s
sys 0m3.085s

time (./fallocate -wl 360M /mnt/test.file; dd if=/dev/zero of=/mnt/test.file \
bs=1M count=360 conv=notrunc,nocreat oflag=direct,dsync );
360+0 records in
360+0 records out
377487360 bytes (377 MB, 360 MiB) copied, 2.02512 s, 186 MB/s

real 0m6.384s
user 0m0.002s
sys 0m5.823s

---
 fs/xfs/xfs_bmap_util.c |  5 +++--
 fs/xfs/xfs_bmap_util.h |  4 ++--
 fs/xfs/xfs_file.c      | 23 ++++++++++++++++++-----
 3 files changed, 23 insertions(+), 9 deletions(-)
fs/xfs/xfs_bmap_util.c +0 -0
index 06ca11731e430..a91596a280ba5 100644

fs/xfs/xfs_bmap_util.c xfs_free_eofblocks( · line 648

Quoted 4 lines
fs/xfs/xfs_bmap_util.c +1 -0
@@ -645,6 +645,7 @@ xfs_free_eofblocks(
 int
 xfs_alloc_file_space(
 	struct xfs_inode	*ip,
+    uint32_t        flags,      /* XFS_BMAPI_... */

Spacing issue

Quoted 43 lines
fs/xfs/xfs_bmap_util.c +2 -2
 	xfs_off_t		offset,
 	xfs_off_t		len)
 {
@@ -748,8 +749,8 @@ xfs_alloc_file_space(
 		 * will eventually reach the requested range.
 		 */
 		error = xfs_bmapi_write(tp, ip, startoffset_fsb,
-				allocatesize_fsb, XFS_BMAPI_PREALLOC, 0, imapp,
-				&nimaps);
+				allocatesize_fsb, flags, 0, imapp, &nimaps);
+
 		if (error) {
 			if (error != -ENOSR)
 				goto error;
fs/xfs/xfs_bmap_util.h +2 -2
index c477b33616304..67770830eb245 100644
@@ -55,8 +55,8 @@ int	xfs_bmap_last_extent(struct xfs_trans *tp, struct xfs_inode *ip,
 			     int *is_empty);
 
 /* preallocation and hole punch interface */
-int	xfs_alloc_file_space(struct xfs_inode *ip, xfs_off_t offset,
-		xfs_off_t len);
+int	xfs_alloc_file_space(struct xfs_inode *ip, uint32_t flags,
+		xfs_off_t offset, xfs_off_t len);
 int	xfs_free_file_space(struct xfs_inode *ip, xfs_off_t offset,
 		xfs_off_t len, struct xfs_zone_alloc_ctx *ac);
 int	xfs_collapse_file_space(struct xfs_inode *, xfs_off_t offset,
fs/xfs/xfs_file.c +1 -1
index f96fbf5c54c99..48559a011e9b4 100644
@@ -1255,7 +1255,7 @@ xfs_falloc_insert_range(
 static int
 xfs_falloc_zero_range(
 	struct file		*file,
-	int			mode,
+	int				mode,
 	loff_t			offset,
 	loff_t			len,
 	struct xfs_zone_alloc_ctx *ac)

fs/xfs/xfs_file.c xfs_falloc_zero_range( · line 1282

Quoted 7 lines
fs/xfs/xfs_file.c +3 -1
@@ -1277,7 +1277,16 @@ xfs_falloc_zero_range(
 
 	len = round_up(offset + len, blksize) - round_down(offset, blksize);
 	offset = round_down(offset, blksize);
-	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
+	if (mode & FALLOC_FL_WRITE_ZEROES) {
+		if (!bdev_write_zeroes_unmap_sectors(inode->i_sb->s_bdev))
+	        return -EOPNOTSUPP;

indentation issue above.
You can use ./scripts/checkpatch.pl --max-line-length=80 <path to patch files>. This should catch most of the common styling issues. Yes, there are XFS specific styles which the above script will not be able to catch.
--NR

Quoted 48 lines
fs/xfs/xfs_file.c +14 -3
+		error = xfs_alloc_file_space(XFS_I(inode), XFS_BMAPI_ZERO,
+				offset, len);
+	}
+	else
+		error = xfs_alloc_file_space(XFS_I(inode), XFS_BMAPI_PREALLOC,
+				offset, len);
+
 	if (error)
 		return error;
 	return xfs_falloc_setsize(file, new_size);
@@ -1302,7 +1311,8 @@ xfs_falloc_unshare_range(
 	if (error)
 		return error;
 
-	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
+	error = xfs_alloc_file_space(XFS_I(inode), XFS_BMAPI_PREALLOC,
+			offset, len);
 	if (error)
 		return error;
 	return xfs_falloc_setsize(file, new_size);
@@ -1330,7 +1340,9 @@ xfs_falloc_allocate_range(
 	if (error)
 		return error;
 
-	error = xfs_alloc_file_space(XFS_I(inode), offset, len);
+	error = xfs_alloc_file_space(XFS_I(inode), XFS_BMAPI_PREALLOC,
+			offset, len);
+
 	if (error)
 		return error;
 	return xfs_falloc_setsize(file, new_size);
@@ -1340,7 +1352,7 @@ xfs_falloc_allocate_range(
 		(FALLOC_FL_ALLOCATE_RANGE | FALLOC_FL_KEEP_SIZE |	\
 		 FALLOC_FL_PUNCH_HOLE |	FALLOC_FL_COLLAPSE_RANGE |	\
 		 FALLOC_FL_ZERO_RANGE |	FALLOC_FL_INSERT_RANGE |	\
-		 FALLOC_FL_UNSHARE_RANGE)
+		 FALLOC_FL_UNSHARE_RANGE | FALLOC_FL_WRITE_ZEROES )
 
 STATIC long
 __xfs_file_fallocate(
@@ -1383,6 +1395,7 @@ __xfs_file_fallocate(
 	case FALLOC_FL_INSERT_RANGE:
 		error = xfs_falloc_insert_range(file, offset, len);
 		break;
+	case FALLOC_FL_WRITE_ZEROES:
 	case FALLOC_FL_ZERO_RANGE:
 		error = xfs_falloc_zero_range(file, mode, offset, len, ac);
 		break;