5 files changed, 86 insertions, 18 deletions

diff --git a/fs/xfs/libxfs/xfs_ag.c b/fs/xfs/libxfs/xfs_ag.c
index 778ec52cce70..ee9ec0c50bec 100644
--- a/fs/xfs/libxfs/xfs_ag.c
+++ b/fs/xfs/libxfs/xfs_ag.c
@@ -804,6 +804,14 @@ xfs_ag_shrink_space(
 	args.fsbno = XFS_AGB_TO_FSB(mp, agno, aglen - delta);
 
 	/*
+	 * Make sure that the last inode cluster cannot overlap with the new
+	 * end of the AG, even if it's sparse.
+	 */
+	error = xfs_ialloc_check_shrink(*tpp, agno, agibp, aglen - delta);
+	if (error)
+		return error;
+
+	/*
 	 * Disable perag reservations so it doesn't cause the allocation request
 	 * to fail. We'll reestablish reservation before we return.
 	 */
diff --git a/fs/xfs/libxfs/xfs_ialloc.c b/fs/xfs/libxfs/xfs_ialloc.c
index 57d9cb632983..aaf8805a82df 100644
--- a/fs/xfs/libxfs/xfs_ialloc.c
+++ b/fs/xfs/libxfs/xfs_ialloc.c
@@ -2928,3 +2928,58 @@ xfs_ialloc_calc_rootino(
 
 	return XFS_AGINO_TO_INO(mp, 0, XFS_AGB_TO_AGINO(mp, first_bno));
 }
+
+/*
+ * Ensure there are not sparse inode clusters that cross the new EOAG.
+ *
+ * This is a no-op for non-spinode filesystems since clusters are always fully
+ * allocated and checking the bnobt suffices.  However, a spinode filesystem
+ * could have a record where the upper inodes are free blocks.  If those blocks
+ * were removed from the filesystem, the inode record would extend beyond EOAG,
+ * which will be flagged as corruption.
+ */
+int
+xfs_ialloc_check_shrink(
+	struct xfs_trans	*tp,
+	xfs_agnumber_t		agno,
+	struct xfs_buf		*agibp,
+	xfs_agblock_t		new_length)
+{
+	struct xfs_inobt_rec_incore rec;
+	struct xfs_btree_cur	*cur;
+	struct xfs_mount	*mp = tp->t_mountp;
+	struct xfs_perag	*pag;
+	xfs_agino_t		agino = XFS_AGB_TO_AGINO(mp, new_length);
+	int			has;
+	int			error;
+
+	if (!xfs_sb_version_hassparseinodes(&mp->m_sb))
+		return 0;
+
+	pag = xfs_perag_get(mp, agno);
+	cur = xfs_inobt_init_cursor(mp, tp, agibp, pag, XFS_BTNUM_INO);
+
+	/* Look up the inobt record that would correspond to the new EOFS. */
+	error = xfs_inobt_lookup(cur, agino, XFS_LOOKUP_LE, &has);
+	if (error || !has)
+		goto out;
+
+	error = xfs_inobt_get_rec(cur, &rec, &has);
+	if (error)
+		goto out;
+
+	if (!has) {
+		error = -EFSCORRUPTED;
+		goto out;
+	}
+
+	/* If the record covers inodes that would be beyond EOFS, bail out. */
+	if (rec.ir_startino + XFS_INODES_PER_CHUNK > agino) {
+		error = -ENOSPC;
+		goto out;
+	}
+out:
+	xfs_btree_del_cursor(cur, error);
+	xfs_perag_put(pag);
+	return error;
+}
diff --git a/fs/xfs/libxfs/xfs_ialloc.h b/fs/xfs/libxfs/xfs_ialloc.h
index 9df7c80408ff..9a2112b4ad5e 100644
--- a/fs/xfs/libxfs/xfs_ialloc.h
+++ b/fs/xfs/libxfs/xfs_ialloc.h
@@ -122,4 +122,7 @@ int xfs_ialloc_cluster_alignment(struct xfs_mount *mp);
 void xfs_ialloc_setup_geometry(struct xfs_mount *mp);
 xfs_ino_t xfs_ialloc_calc_rootino(struct xfs_mount *mp, int sunit);
 
+int xfs_ialloc_check_shrink(struct xfs_trans *tp, xfs_agnumber_t agno,
+		struct xfs_buf *agibp, xfs_agblock_t new_length);
+
 #endif	/* __XFS_IALLOC_H__ */
diff --git a/fs/xfs/libxfs/xfs_inode_buf.c b/fs/xfs/libxfs/xfs_inode_buf.c
index 04ce361688f7..84ea2e0af9f0 100644
--- a/fs/xfs/libxfs/xfs_inode_buf.c
+++ b/fs/xfs/libxfs/xfs_inode_buf.c
@@ -592,23 +592,27 @@ xfs_inode_validate_extsize(
 	/*
 	 * This comment describes a historic gap in this verifier function.
 	 *
-	 * On older kernels, the extent size hint verifier doesn't check that
-	 * the extent size hint is an integer multiple of the realtime extent
-	 * size on a directory with both RTINHERIT and EXTSZINHERIT flags set.
-	 * The verifier has always enforced the alignment rule for regular
-	 * files with the REALTIME flag set.
+	 * For a directory with both RTINHERIT and EXTSZINHERIT flags set, this
+	 * function has never checked that the extent size hint is an integer
+	 * multiple of the realtime extent size.  Since we allow users to set
+	 * this combination  on non-rt filesystems /and/ to change the rt
+	 * extent size when adding a rt device to a filesystem, the net effect
+	 * is that users can configure a filesystem anticipating one rt
+	 * geometry and change their minds later.  Directories do not use the
+	 * extent size hint, so this is harmless for them.
 	 *
 	 * If a directory with a misaligned extent size hint is allowed to
 	 * propagate that hint into a new regular realtime file, the result
 	 * is that the inode cluster buffer verifier will trigger a corruption
-	 * shutdown the next time it is run.
+	 * shutdown the next time it is run, because the verifier has always
+	 * enforced the alignment rule for regular files.
 	 *
-	 * Unfortunately, there could be filesystems with these misconfigured
-	 * directories in the wild, so we cannot add a check to this verifier
-	 * at this time because that will result a new source of directory
-	 * corruption errors when reading an existing filesystem.  Instead, we
-	 * permit the misconfiguration to pass through the verifiers so that
-	 * callers of this function can correct and mitigate externally.
+	 * Because we allow administrators to set a new rt extent size when
+	 * adding a rt section, we cannot add a check to this verifier because
+	 * that will result a new source of directory corruption errors when
+	 * reading an existing filesystem.  Instead, we rely on callers to
+	 * decide when alignment checks are appropriate, and fix things up as
+	 * needed.
 	 */
 
 	if (rt_flag)
diff --git a/fs/xfs/libxfs/xfs_trans_inode.c b/fs/xfs/libxfs/xfs_trans_inode.c
index 8d595a5c4abd..16f723ebe8dd 100644
--- a/fs/xfs/libxfs/xfs_trans_inode.c
+++ b/fs/xfs/libxfs/xfs_trans_inode.c
@@ -143,16 +143,14 @@ xfs_trans_log_inode(
 	}
 
 	/*
-	 * Inode verifiers on older kernels don't check that the extent size
-	 * hint is an integer multiple of the rt extent size on a directory
-	 * with both rtinherit and extszinherit flags set.  If we're logging a
-	 * directory that is misconfigured in this way, clear the hint.
+	 * Inode verifiers do not check that the extent size hint is an integer
+	 * multiple of the rt extent size on a directory with both rtinherit
+	 * and extszinherit flags set.  If we're logging a directory that is
+	 * misconfigured in this way, clear the hint.
 	 */
 	if ((ip->i_diflags & XFS_DIFLAG_RTINHERIT) &&
 	    (ip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) &&
 	    (ip->i_extsize % ip->i_mount->m_sb.sb_rextsize) > 0) {
-		xfs_info_once(ip->i_mount,
-	"Correcting misaligned extent size hint in inode 0x%llx.", ip->i_ino);
 		ip->i_diflags &= ~(XFS_DIFLAG_EXTSIZE |
 				   XFS_DIFLAG_EXTSZINHERIT);
 		ip->i_extsize = 0;