Merge tag 'vfs-6.14-rc1.statx.dio' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs

Pull vfs direct-io updates from Christian Brauner:
 "File systems that write out of place usually require different
  alignment for direct I/O writes than what they can do for reads.

  Add a separate dio read align field to statx, as many out of place
  write file systems can easily do reads aligned to the device sector
  size, but require bigger alignment for writes.

  This is usually papered over by falling back to buffered I/O for
  smaller writes and doing read-modify-write cycles, but performance for
  this sucks, so applications benefit from knowing the actual write
  alignment"

* tag 'vfs-6.14-rc1.statx.dio' of git://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs:
  xfs: report larger dio alignment for COW inodes
  xfs: report the correct read/write dio alignment for reflinked inodes
  xfs: cleanup xfs_vn_getattr
  fs: add STATX_DIO_READ_ALIGN
  fs: reformat the statx definition

3 files changed, 49 insertions, 25 deletions

diff --git a/fs/stat.c b/fs/stat.c
index 0870e969a8a0..2c0e111a098a 100644
--- a/fs/stat.c
+++ b/fs/stat.c
@@ -725,6 +725,7 @@ cp_statx(const struct kstat *stat, struct statx __user *buffer)
 	tmp.stx_mnt_id = stat->mnt_id;
 	tmp.stx_dio_mem_align = stat->dio_mem_align;
 	tmp.stx_dio_offset_align = stat->dio_offset_align;
+	tmp.stx_dio_read_offset_align = stat->dio_read_offset_align;
 	tmp.stx_subvol = stat->subvol;
 	tmp.stx_atomic_write_unit_min = stat->atomic_write_unit_min;
 	tmp.stx_atomic_write_unit_max = stat->atomic_write_unit_max;
diff --git a/fs/xfs/xfs_ioctl.c b/fs/xfs/xfs_ioctl.c
index 0789c18aaa18..f95103325318 100644
--- a/fs/xfs/xfs_ioctl.c
+++ b/fs/xfs/xfs_ioctl.c
@@ -1204,7 +1204,16 @@ xfs_file_ioctl(
 		struct xfs_buftarg	*target = xfs_inode_buftarg(ip);
 		struct dioattr		da;
 
-		da.d_mem =  da.d_miniosz = target->bt_logical_sectorsize;
+		da.d_mem = target->bt_logical_sectorsize;
+
+		/*
+		 * See xfs_report_dioalign() for an explanation about why this
+		 * reports a value larger than the sector size for COW inodes.
+		 */
+		if (xfs_is_cow_inode(ip))
+			da.d_miniosz = xfs_inode_alloc_unitsize(ip);
+		else
+			da.d_miniosz = target->bt_logical_sectorsize;
 		da.d_maxiosz = INT_MAX & ~(da.d_miniosz - 1);
 
 		if (copy_to_user(arg, &da, sizeof(da)))
diff --git a/fs/xfs/xfs_iops.c b/fs/xfs/xfs_iops.c
index 207e0dadffc3..40289fe6f5b2 100644
--- a/fs/xfs/xfs_iops.c
+++ b/fs/xfs/xfs_iops.c
@@ -573,17 +573,43 @@ xfs_stat_blksize(
 }
 
 static void
-xfs_get_atomic_write_attr(
+xfs_report_dioalign(
 	struct xfs_inode	*ip,
-	unsigned int		*unit_min,
-	unsigned int		*unit_max)
+	struct kstat		*stat)
 {
-	if (!xfs_inode_can_atomicwrite(ip)) {
-		*unit_min = *unit_max = 0;
-		return;
-	}
+	struct xfs_buftarg	*target = xfs_inode_buftarg(ip);
+	struct block_device	*bdev = target->bt_bdev;
+
+	stat->result_mask |= STATX_DIOALIGN | STATX_DIO_READ_ALIGN;
+	stat->dio_mem_align = bdev_dma_alignment(bdev) + 1;
 
-	*unit_min = *unit_max = ip->i_mount->m_sb.sb_blocksize;
+	/*
+	 * For COW inodes, we can only perform out of place writes of entire
+	 * allocation units (blocks or RT extents).
+	 * For writes smaller than the allocation unit, we must fall back to
+	 * buffered I/O to perform read-modify-write cycles.  At best this is
+	 * highly inefficient; at worst it leads to page cache invalidation
+	 * races.  Tell applications to avoid this by reporting the larger write
+	 * alignment in dio_offset_align, and the smaller read alignment in
+	 * dio_read_offset_align.
+	 */
+	stat->dio_read_offset_align = bdev_logical_block_size(bdev);
+	if (xfs_is_cow_inode(ip))
+		stat->dio_offset_align = xfs_inode_alloc_unitsize(ip);
+	else
+		stat->dio_offset_align = stat->dio_read_offset_align;
+}
+
+static void
+xfs_report_atomic_write(
+	struct xfs_inode	*ip,
+	struct kstat		*stat)
+{
+	unsigned int		unit_min = 0, unit_max = 0;
+
+	if (xfs_inode_can_atomicwrite(ip))
+		unit_min = unit_max = ip->i_mount->m_sb.sb_blocksize;
+	generic_fill_statx_atomic_writes(stat, unit_min, unit_max);
 }
 
 STATIC int
@@ -647,22 +673,10 @@ xfs_vn_getattr(
 		stat->rdev = inode->i_rdev;
 		break;
 	case S_IFREG:
-		if (request_mask & STATX_DIOALIGN) {
-			struct xfs_buftarg	*target = xfs_inode_buftarg(ip);
-			struct block_device	*bdev = target->bt_bdev;
-
-			stat->result_mask |= STATX_DIOALIGN;
-			stat->dio_mem_align = bdev_dma_alignment(bdev) + 1;
-			stat->dio_offset_align = bdev_logical_block_size(bdev);
-		}
-		if (request_mask & STATX_WRITE_ATOMIC) {
-			unsigned int	unit_min, unit_max;
-
-			xfs_get_atomic_write_attr(ip, &unit_min,
-					&unit_max);
-			generic_fill_statx_atomic_writes(stat,
-					unit_min, unit_max);
-		}
+		if (request_mask & (STATX_DIOALIGN | STATX_DIO_READ_ALIGN))
+			xfs_report_dioalign(ip, stat);
+		if (request_mask & STATX_WRITE_ATOMIC)
+			xfs_report_atomic_write(ip, stat);
 		fallthrough;
 	default:
 		stat->blksize = xfs_stat_blksize(ip);