[PATCH v9 04/14] iomap: don't punch delalloc under a dirty EOF block

[Chris Mason <mason@kernel.org>]

After a failed buffered write, writeback on XFS can write in place to
the shared block that contains EOF, and corrupt the other file that
shares it.

After a failed buffered write, iomap_write_delalloc_release() punches
the delalloc reservation of the write, except under dirty folios, which
still need it. It looks for dirty folios only up to i_size, and punches
the rest unconditionally. But XFS rounds that punch down to a block
boundary, so the block that contains EOF loses its COW fork reservation,
and writeback goes to the shared block:

    block size 4096, i_size 5000, reservation 4096 to 8192
    searched for dirty folios:  4096 to 5000
    punched unconditionally:    5000 to 8192
    rounded to blocks by XFS:   4096 to 8192

The fix used here is to look for dirty folios up to the end of the block
that contains EOF.

Fixes: f43dc4dc3eff ("iomap: buffered write failure should not truncate the page cache")
Fixes: f6f91d290c8b ("xfs: punch delalloc extents from the COW fork for COW writes")
Signed-off-by: Chris Mason <mason@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>
---

Notes:
    v6->v9: no change

 fs/iomap/buffered-io.c | 34 ++++++++++++++++++++++++++++++----
 1 file changed, 30 insertions(+), 4 deletions(-)

diff --git a/fs/iomap/buffered-io.c b/fs/iomap/buffered-io.c
index d6e2f1ded5ca..d0980a7f27d1 100644
--- a/fs/iomap/buffered-io.c
+++ b/fs/iomap/buffered-io.c
@@ -1495,9 +1495,34 @@ static void iomap_write_delalloc_scan(struct inode *inode,
  * mapping_seek_hole_data() will return it, and we will skip the punch because
  * the folio is dirty. THis is incorrect - we always need to punch out delalloc
  * beyond EOF in this case as writeback will never write back and covert that
- * delalloc block beyond EOF. Hence we limit the cached data scan range to EOF,
- * resulting in always punching out the range from the EOF to the end of the
- * range the iomap spans.
+ * delalloc block beyond EOF. Hence we limit the cached data scan range, and
+ * always punch out the range from the end of the scan to the end of the range
+ * the iomap spans.
+ *
+ * The scan cannot simply end at EOF, because EOF is a byte offset and a
+ * delalloc reservation covers whole blocks. If EOF lies inside a block, it
+ * splits that block in two:
+ *
+ *	[block start, EOF)	File data. If it is dirty it will be written
+ *				back, so the block must keep its delalloc
+ *				backing.
+ *	[EOF, block end)	Beyond EOF. By the rule above, this would always
+ *				be punched out.
+ *
+ * Both parts share one reservation, and it cannot be split: a punch of a part
+ * of a block takes away the reservation for the whole block. If the scan ended
+ * at EOF, we would skip the punch for the first part because it is dirty, and
+ * then punch the second part, which takes the reservation away from under the
+ * dirty data in the first part. On XFS, writeback that has lost a COW fork
+ * reservation like this writes in place, over a block that is shared with
+ * another file.
+ *
+ * Hence we treat the EOF block as a whole, and let the cached data over it
+ * decide: the scan range is rounded up to the end of the block that EOF lies
+ * in. If the data over the EOF block is dirty, the scan moves the punch start
+ * past the block. If it is clean, the block is punched like any other block in
+ * the range. Only blocks that start at or beyond EOF are punched
+ * unconditionally. If EOF is block aligned, the round up changes nothing.
  *
  * Intervals are of the form [start_byte, end_byte) (i.e. open ended) because it
  * matches the intervals returned by mapping_seek_hole_data(). i.e. SEEK_DATA
@@ -1511,7 +1536,8 @@ void iomap_write_delalloc_release(struct inode *inode, loff_t start_byte,
 		iomap_punch_t punch)
 {
 	loff_t punch_start_byte = start_byte;
-	loff_t scan_end_byte = min(i_size_read(inode), end_byte);
+	loff_t scan_end_byte = min_t(u64, end_byte,
+			round_up((u64)i_size_read(inode), i_blocksize(inode)));
 
 	/*
 	 * The caller must hold invalidate_lock to avoid races with page faults
-- 
2.55.0