When unaligned truncate down a big realtime file, xfs_truncate_page()
only zeros out the tail EOF block, __xfs_bunmapi() should split the tail
written extent and convert the later one that beyond EOF block to
unwritten, but it couldn't work as expected now since the reserved block
is zero in xfs_setattr_size(), this could expose stale data just after
commit '943bc088 ("iomap: don't increase i_size if it's not a write
operation")'.

If we truncate file that contains a large enough written extent:

     |<    rxext    >|<    rtext    >|
  ...WWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWWW
        ^ (new EOF)      ^ old EOF

Since we only zeros out the tail of the EOF block, and
xfs_itruncate_extents()->..->__xfs_bunmapi() unmap the whole ailgned
extents, it becomes this state:

     |<    rxext    >|
  ...WWWzWWWWWWWWWWWWW
        ^ new EOF

Then if we do an extending write like this, the blocks in the previous
tail extent becomes stale:

     |<    rxext    >|
  ...WWWzSSSSSSSSSSSSS..........WWWWWWWWWWWWWWWWW
        ^ old EOF               ^ append start  ^ new EOF

Fix this by reserving XFS_DIOSTRAT_SPACE_RES blocks for big realtime
inode.
Signed-off-by: Zhang Yi <yi.zhang@huawei.com>
Link: https://lore.kernel.org/r/20240618142112.1315279-2-yi.zhang@huaweicloud.comReviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>