Skip to contentSkip to conversation

xfs: punch COW fork delalloc on COW writeback failure

PATCH In review Ravi Singh 2 participants 3 messages Jul 9

Subsystems (2)
lore
Share & tools
lore

Ravi Singh

author
#0 lore raw

PATCH 1/1 xfs: punch COW fork delalloc on COW writeback failure

When a COW writeback I/O fails, xfs_end_ioend_write() cancels the COW fork extents and punches out any overlapping delalloc blocks. Currently the delalloc punch targets XFS_DATA_FORK, but on the error path xfs_reflink_end_cow() is never called, so the COW fork extents have not been moved to the data fork yet. Any overlapping delalloc that needs to be cleaned up is therefore still in the COW fork.

Punching the data fork is incorrect because it can remove delalloc reservations that belong to a concurrent buffered write. In particular, the convert_delay path in xfs_buffered_write_iomap_begin() drops the ILOCK before calling xfs_bmapi_convert_delalloc(), and a concurrent COW writeback failure punching data fork delalloc during this window causes xfs_bmapi_convert_one_delalloc() to find the extent missing and trigger a WARN_ON_ONCE(whichfork != XFS_COW_FORK).

This is the same class of wrong-fork issue fixed by commit f6f91d290c8b ("xfs: punch delalloc extents from the COW fork for COW writes") for the xfs_buffered_write_iomap_end() path.

Fix xfs_end_ioend_write() to punch delalloc from XFS_COW_FORK instead of XFS_DATA_FORK for IOMAP_IOEND_SHARED ioends.

Fixes 5ca5916b6bc9 ("xfs: punch out data fork delalloc blocks on COW writeback failure")
---
 fs/xfs/xfs_aops.c | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)
fs/xfs/xfs_aops.c +6 -6
index 51293b6f331f..8268a3f80398 100644
@@ -131,11 +131,11 @@ xfs_end_ioend_write(
 	}
 
 	/*
-	 * Clean up all COW blocks and underlying data fork delalloc blocks on
-	 * I/O error. The delalloc punch is required because this ioend was
-	 * mapped to blocks in the COW fork and the associated pages are no
-	 * longer dirty. If we don't remove delalloc blocks here, they become
-	 * stale and can corrupt free space accounting on unmount.
+	 * Clean up COW blocks and COW fork delalloc blocks on I/O error.
+	 * On the error path xfs_reflink_end_cow() is never called, so the
+	 * COW fork extents have not been moved to the data fork yet.  Any
+	 * overlapping delalloc that needs to be cleaned up is therefore
+	 * still in the COW fork.
 	 */
 	error = blk_status_to_errno(ioend->io_bio.bi_status);
 	if (unlikely(error)) {
@@ -153,7 +153,7 @@ xfs_end_ioend_write(
 		if (ioend->io_flags & IOMAP_IOEND_SHARED) {
 			ASSERT(!is_zoned);
 			xfs_reflink_cancel_cow_range(ip, offset, size, true);
-			xfs_bmap_punch_delalloc_range(ip, XFS_DATA_FORK, offset,
+			xfs_bmap_punch_delalloc_range(ip, XFS_COW_FORK, offset,
 					offset + size, NULL);
 		}
 		goto done;

--
2.39.3

#1 lore raw
Quoted 15 lines · Ravi Singh

When a COW writeback I/O fails, xfs_end_ioend_write() cancels the COW fork extents and punches out any overlapping delalloc blocks. Currently the delalloc punch targets XFS_DATA_FORK, but on the error path xfs_reflink_end_cow() is never called, so the COW fork extents have not been moved to the data fork yet. Any overlapping delalloc that needs to be cleaned up is therefore still in the COW fork.

Punching the data fork is incorrect because it can remove delalloc reservations that belong to a concurrent buffered write. In particular, the convert_delay path in xfs_buffered_write_iomap_begin() drops the ILOCK before calling xfs_bmapi_convert_delalloc(), and a concurrent COW writeback failure punching data fork delalloc during this window causes xfs_bmapi_convert_one_delalloc() to find the extent missing and trigger a WARN_ON_ONCE(whichfork != XFS_COW_FORK).

Quoted 3 lines

This is the same class of wrong-fork issue fixed by commit f6f91d290c8b ("xfs: punch delalloc extents from the COW fork for COW writes") for the xfs_buffered_write_iomap_end() path.

I know I suggested that punching the wrong fork -might- be the cause of the bug you were trying to reproduce after noticing that commit reference in your analysis, but I don't think my hypothesis was correct.

I've had reason to look at xfs_reflink_cancel_cow_range() very closely today, and I think you should, too. i.e. work out what the COW fork should contain after it finishes running on the range the IO spanned.

Then....

Quoted 2 lines

Fix xfs_end_ioend_write() to punch delalloc from XFS_COW_FORK instead of XFS_DATA_FORK for IOMAP_IOEND_SHARED ioends.

... it should become clear that COW fork doesn't actually contain any delalloc extents over the given range after the cancel operation returns.

So even if punching the data fork is wrong or triggering some other data fork extent manipulation race, replacing it with a cow fork punch is not a fix for whatever the underlying issue might be.

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
author
#2 lore raw
Quoted 36 lines · Dave Chinner
Quoted 15 lines · Ravi Singh

When a COW writeback I/O fails, xfs_end_ioend_write() cancels the COW fork extents and punches out any overlapping delalloc blocks. Currently the delalloc punch targets XFS_DATA_FORK, but on the error path xfs_reflink_end_cow() is never called, so the COW fork extents have not been moved to the data fork yet. Any overlapping delalloc that needs to be cleaned up is therefore still in the COW fork.

Punching the data fork is incorrect because it can remove delalloc reservations that belong to a concurrent buffered write. In particular, the convert_delay path in xfs_buffered_write_iomap_begin() drops the ILOCK before calling xfs_bmapi_convert_delalloc(), and a concurrent COW writeback failure punching data fork delalloc during this window causes xfs_bmapi_convert_one_delalloc() to find the extent missing and trigger a WARN_ON_ONCE(whichfork != XFS_COW_FORK).

Quoted 3 lines

This is the same class of wrong-fork issue fixed by commit f6f91d290c8b ("xfs: punch delalloc extents from the COW fork for COW writes") for the xfs_buffered_write_iomap_end() path.

I know I suggested that punching the wrong fork -might- be the cause of the bug you were trying to reproduce after noticing that commit reference in your analysis, but I don't think my hypothesis was correct.

I've had reason to look at xfs_reflink_cancel_cow_range() very closely today, and I think you should, too. i.e. work out what the COW fork should contain after it finishes running on the range the IO spanned.

Then....

Quoted 2 lines

Fix xfs_end_ioend_write() to punch delalloc from XFS_COW_FORK instead of XFS_DATA_FORK for IOMAP_IOEND_SHARED ioends.

Quoted 8 lines

... it should become clear that COW fork doesn't actually contain any delalloc extents over the given range after the cancel operation returns.

So even if punching the data fork is wrong or triggering some other data fork extent manipulation race, replacing it with a cow fork punch is not a fix for whatever the underlying issue might be.

Thanks Dave. You're right I should have looked at xfs_reflink_cancel_cow_range() more carefully before posting. cancel_cow_blocks() removes both delalloc and real extents from the COW fork when cancel_real is true, so there's nothing left in the COW fork to punch after it returns.

I'll go back and look at this more closely.

Quoted 7 lines

Cheers,

Dave.
--
Dave Chinner
[email protected]