xfs: flush inodegc before swapon

Fix the brand new xfstest that tries to swapon on a recently unshared
file and use the chance to document the other bit of magic in this
function.

The big comment is taken from a mailinglist post by Dave Chinner.

Fixes: 5e672cd ("xfs: introduce xfs_inodegc_push()")
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Carlos Maiolino <cem@kernel.org>

fs/xfs/xfs_aops.c

@@ -19,6 +19,7 @@
 #include "xfs_reflink.h"
 #include "xfs_errortag.h"
 #include "xfs_error.h"
+#include "xfs_icache.h"
 
 struct xfs_writepage_ctx {
 	struct iomap_writepage_ctx ctx;
@@ -533,7 +534,39 @@ xfs_vm_swap_activate(
 	struct file			*swap_file,
 	sector_t			*span)
 {
-	sis->bdev = xfs_inode_buftarg(XFS_I(file_inode(swap_file)))->bt_bdev;
+	struct xfs_inode		*ip = XFS_I(file_inode(swap_file));
+
+	/*
+	 * Swap file activation can race against concurrent shared extent
+	 * removal in files that have been cloned.  If this happens,
+	 * iomap_swapfile_iter() can fail because it encountered a shared
+	 * extent even though an operation is in progress to remove those
+	 * shared extents.
+	 *
+	 * This race becomes problematic when we defer extent removal
+	 * operations beyond the end of a syscall (i.e. use async background
+	 * processing algorithms).  Users think the extents are no longer
+	 * shared, but iomap_swapfile_iter() still sees them as shared
+	 * because the refcountbt entries for the extents being removed have
+	 * not yet been updated.  Hence the swapon call fails unexpectedly.
+	 *
+	 * The race condition is currently most obvious from the unlink()
+	 * operation as extent removal is deferred until after the last
+	 * reference to the inode goes away.  We then process the extent
+	 * removal asynchronously, hence triggers the "syscall completed but
+	 * work not done" condition mentioned above.  To close this race
+	 * window, we need to flush any pending inodegc operations to ensure
+	 * they have updated the refcountbt records before we try to map the
+	 * swapfile.
+	 */
+	xfs_inodegc_flush(ip->i_mount);
+
+	/*
+	 * Direct the swap code to the correct block device when this file
+	 * sits on the RT device.
+	 */
+	sis->bdev = xfs_inode_buftarg(ip)->bt_bdev;
+
 	return iomap_swapfile_activate(sis, swap_file, span,
 			&xfs_read_iomap_ops);
 }