On Tue, Jun 26, 2007 at 11:49:15PM -0400, Andreas Dilger wrote:

On Jun 27, 2007  09:14 +1000, David Chinner wrote:

quoted

Someone on the XFs list had an interesting request - preallocated
swap files. You can't use unwritten extents for this because
of sys_swapon()s use of bmap() (XFS returns holes for reading
unwritten extents), so we need a method of preallocating that does
not zero or mark the extent unread. i.e. FA_MKSWAP.

Is there a reason why unwritten extents return 0 to bmap()?

It's a fallout of xfs_get_blocks not mapping unwritten extents
on read because we want do_mpage_readpage() to treat them
as a hole. i.e. zero fill them instead of doing I/O. This is
the way XFS was shoehorned into the generic read path :/

This
would seem to be the only impediment from using fallocated files
for swap files.  Maybe if FIEMAP was used by mkswap to get an
"UNWRITTEN" flag back instead of "HOLE" it wouldn't be a problem.

Probably. If we taught do_mpage_readpage() about unwritten mappings,
then would could map them on read if and then sys_swapon can remain
blissfully unaware of unwritten extents.

I think this is pretty much all I need to do to acheive that is
(untested):

---

Teach do_mpage_readpage() about unwritten extents so we can
always map them in get_blocks rather than they are are holes on
read. Allows setup_swap_extents() to use preallocated files on XFS
filesystems for swap files without ever needing to convert them.

Signed-Off-By: Dave Chinner <redacted>

---
 fs/mpage.c                  |    5 +++--
 fs/xfs/linux-2.6/xfs_aops.c |   13 +++----------
 2 files changed, 6 insertions(+), 12 deletions(-)

Index: 2.6.x-xfs-new/fs/mpage.c
===================================================================

--- 2.6.x-xfs-new.orig/fs/mpage.c	2007-05-29 16:17:59.000000000 +1000
+++ 2.6.x-xfs-new/fs/mpage.c	2007-06-27 22:39:35.568852270 +1000

@@ -207,7 +207,8 @@ do_mpage_readpage(struct bio *bio, struc
 	 * Map blocks using the result from the previous get_blocks call first.
 	 */
 	nblocks = map_bh->b_size >> blkbits;
-	if (buffer_mapped(map_bh) && block_in_file > *first_logical_block &&
+	if (buffer_mapped(map_bh) && !buffer_unwritten(map_bh) &&
+			block_in_file > *first_logical_block &&
 			block_in_file < (*first_logical_block + nblocks)) {
 		unsigned map_offset = block_in_file - *first_logical_block;
 		unsigned last = nblocks - map_offset;

@@ -242,7 +243,7 @@ do_mpage_readpage(struct bio *bio, struc
 			*first_logical_block = block_in_file;
 		}
 
-		if (!buffer_mapped(map_bh)) {
+		if (!buffer_mapped(map_bh) || buffer_unwritten(map_bh)) {
 			fully_mapped = 0;
 			if (first_hole == blocks_per_page)
 				first_hole = page_block;

Index: 2.6.x-xfs-new/fs/xfs/linux-2.6/xfs_aops.c
===================================================================

--- 2.6.x-xfs-new.orig/fs/xfs/linux-2.6/xfs_aops.c	2007-06-05 22:14:39.000000000 +1000
+++ 2.6.x-xfs-new/fs/xfs/linux-2.6/xfs_aops.c	2007-06-27 22:39:29.545636749 +1000

@@ -1340,16 +1340,9 @@ __xfs_get_blocks(
 		return 0;
 
 	if (iomap.iomap_bn != IOMAP_DADDR_NULL) {
-		/*
-		 * For unwritten extents do not report a disk address on
-		 * the read case (treat as if we're reading into a hole).
-		 */
-		if (create || !(iomap.iomap_flags & IOMAP_UNWRITTEN)) {
-			xfs_map_buffer(bh_result, &iomap, offset,
-				       inode->i_blkbits);
-		}
-		if (create && (iomap.iomap_flags & IOMAP_UNWRITTEN)) {
-			if (direct)
+		xfs_map_buffer(bh_result, &iomap, offset, inode->i_blkbits);
+		if (iomap.iomap_flags & IOMAP_UNWRITTEN) {
+			if (create && direct)
 				bh_result->b_private = inode;
 			set_buffer_unwritten(bh_result);
 		}

Cheers,

Dave.
-- 
Dave Chinner
Principal Engineer
SGI Australian Software Group