xfs: fix corruptions in the directory tree

Repair corruptions in the directory tree itself.  Cycles are broken by
removing an incoming parent->child link.  Multiply-owned directories are
fixed by pruning the extra parent -> child links  Disconnected subtrees
are reconnected to the lost and found.

Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Christoph Hellwig <hch@lst.de>

11 files changed, 927 insertions, 8 deletions

diff --git a/fs/xfs/Makefile b/fs/xfs/Makefile
index 8ec0dd257a98..d1ce1213797b 100644
--- a/fs/xfs/Makefile
+++ b/fs/xfs/Makefile
@@ -204,6 +204,7 @@ xfs-y				+= $(addprefix scrub/, \
 				   bmap_repair.o \
 				   cow_repair.o \
 				   dir_repair.o \
+				   dirtree_repair.o \
 				   findparent.o \
 				   fscounters_repair.o \
 				   ialloc_repair.o \
diff --git a/fs/xfs/scrub/dirtree.c b/fs/xfs/scrub/dirtree.c
index ecc56eb5ed27..bde58fb561ea 100644
--- a/fs/xfs/scrub/dirtree.c
+++ b/fs/xfs/scrub/dirtree.c
@@ -26,6 +26,8 @@
 #include "scrub/xfblob.h"
 #include "scrub/listxattr.h"
 #include "scrub/trace.h"
+#include "scrub/repair.h"
+#include "scrub/orphanage.h"
 #include "scrub/dirtree.h"
 
 /*
@@ -95,6 +97,12 @@ xchk_setup_dirtree(
 
 	xchk_fsgates_enable(sc, XCHK_FSGATES_DIRENTS);
 
+	if (xchk_could_repair(sc)) {
+		error = xrep_setup_dirtree(sc);
+		if (error)
+			return error;
+	}
+
 	dl = kvzalloc(sizeof(struct xchk_dirtree), XCHK_GFP_FLAGS);
 	if (!dl)
 		return -ENOMEM;
@@ -104,6 +112,7 @@ xchk_setup_dirtree(
 	INIT_LIST_HEAD(&dl->path_list);
 	dl->root_ino = NULLFSINO;
 	dl->scan_ino = NULLFSINO;
+	dl->parent_ino = NULLFSINO;
 
 	mutex_init(&dl->lock);
 
@@ -142,7 +151,7 @@ out_dl:
  * Add the parent pointer described by @dl->pptr to the given path as a new
  * step.  Returns -ELNRNG if the path is too deep.
  */
-STATIC int
+int
 xchk_dirpath_append(
 	struct xchk_dirtree		*dl,
 	struct xfs_inode		*ip,
@@ -609,6 +618,22 @@ xchk_dirpath_step_is_stale(
 	if (memcmp(dl->hook_xname.name, p->name->name, p->name->len) != 0)
 		return 0;
 
+	/*
+	 * If the update comes from the repair code itself, walk the state
+	 * machine forward.
+	 */
+	if (p->ip->i_ino == dl->scan_ino &&
+	    path->outcome == XREP_DIRPATH_ADOPTING) {
+		xchk_dirpath_set_outcome(dl, path, XREP_DIRPATH_ADOPTED);
+		return 0;
+	}
+
+	if (p->ip->i_ino == dl->scan_ino &&
+	    path->outcome == XREP_DIRPATH_DELETING) {
+		xchk_dirpath_set_outcome(dl, path, XREP_DIRPATH_DELETED);
+		return 0;
+	}
+
 	/* Exact match, scan data is out of date. */
 	trace_xchk_dirpath_changed(dl->sc, path->path_nr, step_nr, p->dp,
 			p->ip, p->name);
@@ -747,7 +772,7 @@ xchk_dirtree_load_path(
  * path was too deep; -ENOSR if there were too many parent pointers; or
  * a negative errno.
  */
-STATIC int
+int
 xchk_dirtree_find_paths_to_root(
 	struct xchk_dirtree	*dl)
 {
@@ -819,7 +844,7 @@ xchk_dirtree_find_paths_to_root(
  * Figure out what to do with the paths we tried to find.  Do not call this
  * if the scan results are stale.
  */
-STATIC void
+void
 xchk_dirtree_evaluate(
 	struct xchk_dirtree		*dl,
 	struct xchk_dirtree_outcomes	*oc)
@@ -856,6 +881,13 @@ xchk_dirtree_evaluate(
 			/* This path got all the way to the root. */
 			oc->good++;
 			break;
+		case XREP_DIRPATH_DELETING:
+		case XREP_DIRPATH_DELETED:
+		case XREP_DIRPATH_ADOPTING:
+		case XREP_DIRPATH_ADOPTED:
+			/* These should not be in progress! */
+			ASSERT(0);
+			break;
 		}
 	}
 
diff --git a/fs/xfs/scrub/dirtree.h b/fs/xfs/scrub/dirtree.h
index 2ddbcf43c291..1e1686365c61 100644
--- a/fs/xfs/scrub/dirtree.h
+++ b/fs/xfs/scrub/dirtree.h
@@ -26,6 +26,11 @@ enum xchk_dirpath_outcome {
 	XCHK_DIRPATH_LOOP,		/* cycle detected further up */
 	XCHK_DIRPATH_STALE,		/* path is stale */
 	XCHK_DIRPATH_OK,		/* path reaches the root */
+
+	XREP_DIRPATH_DELETING,		/* path is being deleted */
+	XREP_DIRPATH_DELETED,		/* path has been deleted */
+	XREP_DIRPATH_ADOPTING,		/* path is being adopted */
+	XREP_DIRPATH_ADOPTED,		/* path has been adopted */
 };
 
 /*
@@ -64,6 +69,9 @@ struct xchk_dirtree_outcomes {
 
 	/* Number of XCHK_DIRPATH_OK */
 	unsigned int		good;
+
+	/* Directory needs to be added to lost+found */
+	bool			needs_adoption;
 };
 
 struct xchk_dirtree {
@@ -79,6 +87,14 @@ struct xchk_dirtree {
 	 */
 	xfs_ino_t		scan_ino;
 
+	/*
+	 * If we start deleting redundant paths to this subdirectory, this is
+	 * the inode number of the surviving parent and the dotdot entry will
+	 * be set to this value.  If the value is NULLFSINO, then use @root_ino
+	 * as a stand-in until the orphanage can adopt the subdirectory.
+	 */
+	xfs_ino_t		parent_ino;
+
 	/* Scratch buffer for scanning pptr xattrs */
 	struct xfs_parent_rec	pptr_rec;
 	struct xfs_da_args	pptr_args;
@@ -87,12 +103,18 @@ struct xchk_dirtree {
 	struct xfs_name		xname;
 	char			namebuf[MAXNAMELEN];
 
+	/* Information for reparenting this directory. */
+	struct xrep_adoption	adoption;
+
 	/*
 	 * Hook into directory updates so that we can receive live updates
 	 * from other writer threads.
 	 */
 	struct xfs_dir_hook	dhook;
 
+	/* Parent pointer update arguments. */
+	struct xfs_parent_args	ppargs;
+
 	/* lock for everything below here */
 	struct mutex		lock;
 
@@ -146,4 +168,11 @@ xchk_dirtree_parentless(const struct xchk_dirtree *dl)
 	return false;
 }
 
+int xchk_dirtree_find_paths_to_root(struct xchk_dirtree *dl);
+int xchk_dirpath_append(struct xchk_dirtree *dl, struct xfs_inode *ip,
+		struct xchk_dirpath *path, const struct xfs_name *name,
+		const struct xfs_parent_rec *pptr);
+void xchk_dirtree_evaluate(struct xchk_dirtree *dl,
+		struct xchk_dirtree_outcomes *oc);
+
 #endif /* __XFS_SCRUB_DIRTREE_H__ */
diff --git a/fs/xfs/scrub/dirtree_repair.c b/fs/xfs/scrub/dirtree_repair.c
new file mode 100644
index 000000000000..5c04e70ba951
--- /dev/null
+++ b/fs/xfs/scrub/dirtree_repair.c
@@ -0,0 +1,821 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * Copyright (c) 2023-2024 Oracle.  All Rights Reserved.
+ * Author: Darrick J. Wong <djwong@kernel.org>
+ */
+#include "xfs.h"
+#include "xfs_fs.h"
+#include "xfs_shared.h"
+#include "xfs_format.h"
+#include "xfs_trans_resv.h"
+#include "xfs_trans_space.h"
+#include "xfs_mount.h"
+#include "xfs_log_format.h"
+#include "xfs_trans.h"
+#include "xfs_inode.h"
+#include "xfs_icache.h"
+#include "xfs_dir2.h"
+#include "xfs_dir2_priv.h"
+#include "xfs_attr.h"
+#include "xfs_parent.h"
+#include "scrub/scrub.h"
+#include "scrub/common.h"
+#include "scrub/bitmap.h"
+#include "scrub/ino_bitmap.h"
+#include "scrub/xfile.h"
+#include "scrub/xfarray.h"
+#include "scrub/xfblob.h"
+#include "scrub/listxattr.h"
+#include "scrub/trace.h"
+#include "scrub/repair.h"
+#include "scrub/orphanage.h"
+#include "scrub/dirtree.h"
+#include "scrub/readdir.h"
+
+/*
+ * Directory Tree Structure Repairs
+ * ================================
+ *
+ * If we decide that the directory being scanned is participating in a
+ * directory loop, the only change we can make is to remove directory entries
+ * pointing down to @sc->ip.  If that leaves it with no parents, the directory
+ * should be adopted by the orphanage.
+ */
+
+/* Set up to repair directory loops. */
+int
+xrep_setup_dirtree(
+	struct xfs_scrub	*sc)
+{
+	return xrep_orphanage_try_create(sc);
+}
+
+/* Change the outcome of this path. */
+static inline void
+xrep_dirpath_set_outcome(
+	struct xchk_dirtree		*dl,
+	struct xchk_dirpath		*path,
+	enum xchk_dirpath_outcome	outcome)
+{
+	trace_xrep_dirpath_set_outcome(dl->sc, path->path_nr, path->nr_steps,
+			outcome);
+
+	path->outcome = outcome;
+}
+
+/* Delete all paths. */
+STATIC void
+xrep_dirtree_delete_all_paths(
+	struct xchk_dirtree		*dl,
+	struct xchk_dirtree_outcomes	*oc)
+{
+	struct xchk_dirpath		*path;
+
+	xchk_dirtree_for_each_path(dl, path) {
+		switch (path->outcome) {
+		case XCHK_DIRPATH_CORRUPT:
+		case XCHK_DIRPATH_LOOP:
+			oc->suspect--;
+			oc->bad++;
+			xrep_dirpath_set_outcome(dl, path, XCHK_DIRPATH_DELETE);
+			break;
+		case XCHK_DIRPATH_OK:
+			oc->good--;
+			oc->bad++;
+			xrep_dirpath_set_outcome(dl, path, XCHK_DIRPATH_DELETE);
+			break;
+		default:
+			break;
+		}
+	}
+
+	ASSERT(oc->suspect == 0);
+	ASSERT(oc->good == 0);
+}
+
+/* Since this is the surviving path, set the dotdot entry to this value. */
+STATIC void
+xrep_dirpath_retain_parent(
+	struct xchk_dirtree		*dl,
+	struct xchk_dirpath		*path)
+{
+	struct xchk_dirpath_step	step;
+	int				error;
+
+	error = xfarray_load(dl->path_steps, path->first_step, &step);
+	if (error)
+		return;
+
+	dl->parent_ino = be64_to_cpu(step.pptr_rec.p_ino);
+}
+
+/* Find the one surviving path so we know how to set dotdot. */
+STATIC void
+xrep_dirtree_find_surviving_path(
+	struct xchk_dirtree		*dl,
+	struct xchk_dirtree_outcomes	*oc)
+{
+	struct xchk_dirpath		*path;
+	bool				foundit = false;
+
+	xchk_dirtree_for_each_path(dl, path) {
+		switch (path->outcome) {
+		case XCHK_DIRPATH_CORRUPT:
+		case XCHK_DIRPATH_LOOP:
+		case XCHK_DIRPATH_OK:
+			if (!foundit) {
+				xrep_dirpath_retain_parent(dl, path);
+				foundit = true;
+				continue;
+			}
+			ASSERT(foundit == false);
+			break;
+		default:
+			break;
+		}
+	}
+
+	ASSERT(oc->suspect + oc->good == 1);
+}
+
+/* Delete all paths except for the one good one. */
+STATIC void
+xrep_dirtree_keep_one_good_path(
+	struct xchk_dirtree		*dl,
+	struct xchk_dirtree_outcomes	*oc)
+{
+	struct xchk_dirpath		*path;
+	bool				foundit = false;
+
+	xchk_dirtree_for_each_path(dl, path) {
+		switch (path->outcome) {
+		case XCHK_DIRPATH_CORRUPT:
+		case XCHK_DIRPATH_LOOP:
+			oc->suspect--;
+			oc->bad++;
+			xrep_dirpath_set_outcome(dl, path, XCHK_DIRPATH_DELETE);
+			break;
+		case XCHK_DIRPATH_OK:
+			if (!foundit) {
+				xrep_dirpath_retain_parent(dl, path);
+				foundit = true;
+				continue;
+			}
+			oc->good--;
+			oc->bad++;
+			xrep_dirpath_set_outcome(dl, path, XCHK_DIRPATH_DELETE);
+			break;
+		default:
+			break;
+		}
+	}
+
+	ASSERT(oc->suspect == 0);
+	ASSERT(oc->good < 2);
+}
+
+/* Delete all paths except for one suspect one. */
+STATIC void
+xrep_dirtree_keep_one_suspect_path(
+	struct xchk_dirtree		*dl,
+	struct xchk_dirtree_outcomes	*oc)
+{
+	struct xchk_dirpath		*path;
+	bool				foundit = false;
+
+	xchk_dirtree_for_each_path(dl, path) {
+		switch (path->outcome) {
+		case XCHK_DIRPATH_CORRUPT:
+		case XCHK_DIRPATH_LOOP:
+			if (!foundit) {
+				xrep_dirpath_retain_parent(dl, path);
+				foundit = true;
+				continue;
+			}
+			oc->suspect--;
+			oc->bad++;
+			xrep_dirpath_set_outcome(dl, path, XCHK_DIRPATH_DELETE);
+			break;
+		case XCHK_DIRPATH_OK:
+			ASSERT(0);
+			break;
+		default:
+			break;
+		}
+	}
+
+	ASSERT(oc->suspect == 1);
+	ASSERT(oc->good == 0);
+}
+
+/*
+ * Figure out what to do with the paths we tried to find.  Returns -EDEADLOCK
+ * if the scan results have become stale.
+ */
+STATIC void
+xrep_dirtree_decide_fate(
+	struct xchk_dirtree		*dl,
+	struct xchk_dirtree_outcomes	*oc)
+{
+	xchk_dirtree_evaluate(dl, oc);
+
+	/* Parentless directories should not have any paths at all. */
+	if (xchk_dirtree_parentless(dl)) {
+		xrep_dirtree_delete_all_paths(dl, oc);
+		return;
+	}
+
+	/* One path is exactly the number of paths we want. */
+	if (oc->good + oc->suspect == 1) {
+		xrep_dirtree_find_surviving_path(dl, oc);
+		return;
+	}
+
+	/* Zero paths means we should reattach the subdir to the orphanage. */
+	if (oc->good + oc->suspect == 0) {
+		if (dl->sc->orphanage)
+			oc->needs_adoption = true;
+		return;
+	}
+
+	/*
+	 * Otherwise, this subdirectory has too many parents.  If there's at
+	 * least one good path, keep it and delete the others.
+	 */
+	if (oc->good > 0) {
+		xrep_dirtree_keep_one_good_path(dl, oc);
+		return;
+	}
+
+	/*
+	 * There are no good paths and there are too many suspect paths.
+	 * Keep the first suspect path and delete the rest.
+	 */
+	xrep_dirtree_keep_one_suspect_path(dl, oc);
+}
+
+/*
+ * Load the first step of this path into @step and @dl->xname/pptr
+ * for later repair work.
+ */
+STATIC int
+xrep_dirtree_prep_path(
+	struct xchk_dirtree		*dl,
+	struct xchk_dirpath		*path,
+	struct xchk_dirpath_step	*step)
+{
+	int				error;
+
+	error = xfarray_load(dl->path_steps, path->first_step, step);
+	if (error)
+		return error;
+
+	error = xfblob_loadname(dl->path_names, step->name_cookie, &dl->xname,
+			step->name_len);
+	if (error)
+		return error;
+
+	dl->pptr_rec = step->pptr_rec; /* struct copy */
+	return 0;
+}
+
+/* Delete the VFS dentry for a removed child. */
+STATIC int
+xrep_dirtree_purge_dentry(
+	struct xchk_dirtree	*dl,
+	struct xfs_inode	*dp,
+	const struct xfs_name	*name)
+{
+	struct qstr		qname = QSTR_INIT(name->name, name->len);
+	struct dentry		*parent_dentry, *child_dentry;
+	int			error = 0;
+
+	/*
+	 * Find the dentry for the parent directory.  If there isn't one, we're
+	 * done.  Caller already holds i_rwsem for parent and child.
+	 */
+	parent_dentry = d_find_alias(VFS_I(dp));
+	if (!parent_dentry)
+		return 0;
+
+	/* The VFS thinks the parent is a directory, right? */
+	if (!d_is_dir(parent_dentry)) {
+		ASSERT(d_is_dir(parent_dentry));
+		error = -EFSCORRUPTED;
+		goto out_dput_parent;
+	}
+
+	/*
+	 * Try to find the dirent pointing to the child.  If there isn't one,
+	 * we're done.
+	 */
+	qname.hash = full_name_hash(parent_dentry, name->name, name->len);
+	child_dentry = d_lookup(parent_dentry, &qname);
+	if (!child_dentry) {
+		error = 0;
+		goto out_dput_parent;
+	}
+
+	trace_xrep_dirtree_delete_child(dp->i_mount, child_dentry);
+
+	/* Child is not a directory?  We're screwed. */
+	if (!d_is_dir(child_dentry)) {
+		ASSERT(d_is_dir(child_dentry));
+		error = -EFSCORRUPTED;
+		goto out_dput_child;
+	}
+
+	/* Replace the child dentry with a negative one. */
+	d_delete(child_dentry);
+
+out_dput_child:
+	dput(child_dentry);
+out_dput_parent:
+	dput(parent_dentry);
+	return error;
+}
+
+/*
+ * Prepare to delete a link by taking the IOLOCK of the parent and the child
+ * (scrub target).  Caller must hold IOLOCK_EXCL on @sc->ip.  Returns 0 if we
+ * took both locks, or a negative errno if we couldn't lock the parent in time.
+ */
+static inline int
+xrep_dirtree_unlink_iolock(
+	struct xfs_scrub	*sc,
+	struct xfs_inode	*dp)
+{
+	int			error;
+
+	ASSERT(sc->ilock_flags & XFS_IOLOCK_EXCL);
+
+	if (xfs_ilock_nowait(dp, XFS_IOLOCK_EXCL))
+		return 0;
+
+	xchk_iunlock(sc, XFS_IOLOCK_EXCL);
+	do {
+		xfs_ilock(dp, XFS_IOLOCK_EXCL);
+		if (xchk_ilock_nowait(sc, XFS_IOLOCK_EXCL))
+			break;
+		xfs_iunlock(dp, XFS_IOLOCK_EXCL);
+
+		if (xchk_should_terminate(sc, &error)) {
+			xchk_ilock(sc, XFS_IOLOCK_EXCL);
+			return error;
+		}
+
+		delay(1);
+	} while (1);
+
+	return 0;
+}
+
+/*
+ * Remove a link from the directory tree and update the dcache.  Returns
+ * -ESTALE if the scan data are now out of date.
+ */
+STATIC int
+xrep_dirtree_unlink(
+	struct xchk_dirtree		*dl,
+	struct xfs_inode		*dp,
+	struct xchk_dirpath		*path,
+	struct xchk_dirpath_step	*step)
+{
+	struct xfs_scrub		*sc = dl->sc;
+	struct xfs_mount		*mp = sc->mp;
+	xfs_ino_t			dotdot_ino;
+	xfs_ino_t			parent_ino = dl->parent_ino;
+	unsigned int			resblks;
+	int				dontcare;
+	int				error;
+
+	/* Take IOLOCK_EXCL of the parent and child. */
+	error = xrep_dirtree_unlink_iolock(sc, dp);
+	if (error)
+		return error;
+
+	/*
+	 * Create the transaction that we need to sever the path.  Ignore
+	 * EDQUOT and ENOSPC being returned via nospace_error because the
+	 * directory code can handle a reservationless update.
+	 */
+	resblks = xfs_remove_space_res(mp, step->name_len);
+	error = xfs_trans_alloc_dir(dp, &M_RES(mp)->tr_remove, sc->ip,
+			&resblks, &sc->tp, &dontcare);
+	if (error)
+		goto out_iolock;
+
+	/*
+	 * Cancel if someone invalidate the paths while we were trying to get
+	 * the ILOCK.
+	 */
+	mutex_lock(&dl->lock);
+	if (dl->stale) {
+		mutex_unlock(&dl->lock);
+		error = -ESTALE;
+		goto out_trans_cancel;
+	}
+	xrep_dirpath_set_outcome(dl, path, XREP_DIRPATH_DELETING);
+	mutex_unlock(&dl->lock);
+
+	trace_xrep_dirtree_delete_path(dl->sc, sc->ip, path->path_nr,
+			&dl->xname, &dl->pptr_rec);
+
+	/*
+	 * Decide if we need to reset the dotdot entry.  Rules:
+	 *
+	 * - If there's a surviving parent, we want dotdot to point there.
+	 * - If we don't have any surviving parents, then point dotdot at the
+	 *   root dir.
+	 * - If dotdot is already set to the value we want, pass in NULLFSINO
+	 *   for no change necessary.
+	 *
+	 * Do this /before/ we dirty anything, in case the dotdot lookup
+	 * fails.
+	 */
+	error = xchk_dir_lookup(sc, sc->ip, &xfs_name_dotdot, &dotdot_ino);
+	if (error)
+		goto out_trans_cancel;
+	if (parent_ino == NULLFSINO)
+		parent_ino = dl->root_ino;
+	if (dotdot_ino == parent_ino)
+		parent_ino = NULLFSINO;
+
+	/* Drop the link from sc->ip's dotdot entry.  */
+	error = xfs_droplink(sc->tp, dp);
+	if (error)
+		goto out_trans_cancel;
+
+	/* Reset the dotdot entry to a surviving parent. */
+	if (parent_ino != NULLFSINO) {
+		error = xfs_dir_replace(sc->tp, sc->ip, &xfs_name_dotdot,
+				parent_ino, 0);
+		if (error)
+			goto out_trans_cancel;
+	}
+
+	/* Drop the link from dp to sc->ip. */
+	error = xfs_droplink(sc->tp, sc->ip);
+	if (error)
+		goto out_trans_cancel;
+
+	error = xfs_dir_removename(sc->tp, dp, &dl->xname, sc->ip->i_ino,
+			resblks);
+	if (error) {
+		ASSERT(error != -ENOENT);
+		goto out_trans_cancel;
+	}
+
+	if (xfs_has_parent(sc->mp)) {
+		error = xfs_parent_removename(sc->tp, &dl->ppargs, dp,
+				&dl->xname, sc->ip);
+		if (error)
+			goto out_trans_cancel;
+	}
+
+	/*
+	 * Notify dirent hooks that we removed the bad link, invalidate the
+	 * dcache, and commit the repair.
+	 */
+	xfs_dir_update_hook(dp, sc->ip, -1, &dl->xname);
+	error = xrep_dirtree_purge_dentry(dl, dp, &dl->xname);
+	if (error)
+		goto out_trans_cancel;
+
+	error = xrep_trans_commit(sc);
+	goto out_ilock;
+
+out_trans_cancel:
+	xchk_trans_cancel(sc);
+out_ilock:
+	xfs_iunlock(sc->ip, XFS_ILOCK_EXCL);
+	xfs_iunlock(dp, XFS_ILOCK_EXCL);
+out_iolock:
+	xfs_iunlock(dp, XFS_IOLOCK_EXCL);
+	return error;
+}
+
+/*
+ * Delete a directory entry that points to this directory.  Returns -ESTALE
+ * if the scan data are now out of date.
+ */
+STATIC int
+xrep_dirtree_delete_path(
+	struct xchk_dirtree		*dl,
+	struct xchk_dirpath		*path)
+{
+	struct xchk_dirpath_step	step;
+	struct xfs_scrub		*sc = dl->sc;
+	struct xfs_inode		*dp;
+	int				error;
+
+	/*
+	 * Load the parent pointer and directory inode for this path, then
+	 * drop the scan lock, the ILOCK, and the transaction so that
+	 * _delete_path can reserve the proper transaction.  This sets up
+	 * @dl->xname for the deletion.
+	 */
+	error = xrep_dirtree_prep_path(dl, path, &step);
+	if (error)
+		return error;
+
+	error = xchk_iget(sc, be64_to_cpu(step.pptr_rec.p_ino), &dp);
+	if (error)
+		return error;
+
+	mutex_unlock(&dl->lock);
+	xchk_trans_cancel(sc);
+	xchk_iunlock(sc, XFS_ILOCK_EXCL);
+
+	/* Delete the directory link and release the parent. */
+	error = xrep_dirtree_unlink(dl, dp, path, &step);
+	xchk_irele(sc, dp);
+
+	/*
+	 * Retake all the resources we had at the beginning even if the repair
+	 * failed or the scan data are now stale.  This keeps things simple for
+	 * the caller.
+	 */
+	xchk_trans_alloc_empty(sc);
+	xchk_ilock(sc, XFS_ILOCK_EXCL);
+	mutex_lock(&dl->lock);
+
+	if (!error && dl->stale)
+		error = -ESTALE;
+	return error;
+}
+
+/* Add a new path to represent our in-progress adoption. */
+STATIC int
+xrep_dirtree_create_adoption_path(
+	struct xchk_dirtree		*dl)
+{
+	struct xfs_scrub		*sc = dl->sc;
+	struct xchk_dirpath		*path;
+	int				error;
+
+	/*
+	 * We should have capped the number of paths at XFS_MAXLINK-1 in the
+	 * scanner.
+	 */
+	if (dl->nr_paths > XFS_MAXLINK) {
+		ASSERT(dl->nr_paths <= XFS_MAXLINK);
+		return -EFSCORRUPTED;
+	}
+
+	/*
+	 * Create a new xchk_path structure to remember this parent pointer
+	 * and record the first name step.
+	 */
+	path = kmalloc(sizeof(struct xchk_dirpath), XCHK_GFP_FLAGS);
+	if (!path)
+		return -ENOMEM;
+
+	INIT_LIST_HEAD(&path->list);
+	xino_bitmap_init(&path->seen_inodes);
+	path->nr_steps = 0;
+	path->outcome = XREP_DIRPATH_ADOPTING;
+
+	/*
+	 * Record the new link that we just created in the orphanage.  Because
+	 * adoption is the last repair that we perform, we don't bother filling
+	 * in the path all the way back to the root.
+	 */
+	xfs_inode_to_parent_rec(&dl->pptr_rec, sc->orphanage);
+
+	error = xino_bitmap_set(&path->seen_inodes, sc->orphanage->i_ino);
+	if (error)
+		goto out_path;
+
+	trace_xrep_dirtree_create_adoption(sc, sc->ip, dl->nr_paths,
+			&dl->xname, &dl->pptr_rec);
+
+	error = xchk_dirpath_append(dl, sc->ip, path, &dl->xname,
+			&dl->pptr_rec);
+	if (error)
+		goto out_path;
+
+	path->first_step = xfarray_length(dl->path_steps) - 1;
+	path->second_step = XFARRAY_NULLIDX;
+	path->path_nr = dl->nr_paths;
+
+	list_add_tail(&path->list, &dl->path_list);
+	dl->nr_paths++;
+	return 0;
+
+out_path:
+	kfree(path);
+	return error;
+}
+
+/*
+ * Prepare to move a file to the orphanage by taking the IOLOCK of the
+ * orphanage and the child (scrub target).  Caller must hold IOLOCK_EXCL on
+ * @sc->ip.  Returns 0 if we took both locks, or a negative errno if we
+ * couldn't lock the orphanage in time.
+ */
+static inline int
+xrep_dirtree_adopt_iolock(
+	struct xfs_scrub	*sc)
+{
+	int			error;
+
+	ASSERT(sc->ilock_flags & XFS_IOLOCK_EXCL);
+
+	if (xrep_orphanage_ilock_nowait(sc, XFS_IOLOCK_EXCL))
+		return 0;
+
+	xchk_iunlock(sc, XFS_IOLOCK_EXCL);
+	do {
+		xrep_orphanage_ilock(sc, XFS_IOLOCK_EXCL);
+		if (xchk_ilock_nowait(sc, XFS_IOLOCK_EXCL))
+			break;
+		xrep_orphanage_iunlock(sc, XFS_IOLOCK_EXCL);
+
+		if (xchk_should_terminate(sc, &error)) {
+			xchk_ilock(sc, XFS_IOLOCK_EXCL);
+			return error;
+		}
+
+		delay(1);
+	} while (1);
+
+	return 0;
+}
+
+/*
+ * Reattach this orphaned directory to the orphanage.  Do not call this with
+ * any resources held.  Returns -ESTALE if the scan data have become out of
+ * date.
+ */
+STATIC int
+xrep_dirtree_adopt(
+	struct xchk_dirtree		*dl)
+{
+	struct xfs_scrub		*sc = dl->sc;
+	int				error;
+
+	/* Take the IOLOCK of the orphanage and the scrub target. */
+	error = xrep_dirtree_adopt_iolock(sc);
+	if (error)
+		return error;
+
+	/*
+	 * Set up for an adoption.  The directory tree fixer runs after the
+	 * link counts have been corrected.  Therefore, we must bump the
+	 * child's link count since there will be no further opportunity to fix
+	 * errors.
+	 */
+	error = xrep_adoption_trans_alloc(sc, &dl->adoption);
+	if (error)
+		goto out_iolock;
+	dl->adoption.bump_child_nlink = true;
+
+	/* Figure out what name we're going to use here. */
+	error = xrep_adoption_compute_name(&dl->adoption, &dl->xname);
+	if (error)
+		goto out_trans;
+
+	/*
+	 * Now that we have a proposed name for the orphanage entry, create
+	 * a faux path so that the live update hook will see it.
+	 */
+	mutex_lock(&dl->lock);
+	if (dl->stale) {
+		mutex_unlock(&dl->lock);
+		error = -ESTALE;
+		goto out_trans;
+	}
+	error = xrep_dirtree_create_adoption_path(dl);
+	mutex_unlock(&dl->lock);
+	if (error)
+		goto out_trans;
+
+	/* Reparent the directory. */
+	error = xrep_adoption_move(&dl->adoption);
+	if (error)
+		goto out_trans;
+
+	/*
+	 * Commit the name and release all inode locks except for the scrub
+	 * target's IOLOCK.
+	 */
+	error = xrep_trans_commit(sc);
+	goto out_ilock;
+
+out_trans:
+	xchk_trans_cancel(sc);
+out_ilock:
+	xchk_iunlock(sc, XFS_ILOCK_EXCL);
+	xrep_orphanage_iunlock(sc, XFS_ILOCK_EXCL);
+out_iolock:
+	xrep_orphanage_iunlock(sc, XFS_IOLOCK_EXCL);
+	return error;
+}
+
+/*
+ * This newly orphaned directory needs to be adopted by the orphanage.
+ * Make this happen.
+ */
+STATIC int
+xrep_dirtree_move_to_orphanage(
+	struct xchk_dirtree		*dl)
+{
+	struct xfs_scrub		*sc = dl->sc;
+	int				error;
+
+	/*
+	 * Start by dropping all the resources that we hold so that we can grab
+	 * all the resources that we need for the adoption.
+	 */
+	mutex_unlock(&dl->lock);
+	xchk_trans_cancel(sc);
+	xchk_iunlock(sc, XFS_ILOCK_EXCL);
+
+	/* Perform the adoption. */
+	error = xrep_dirtree_adopt(dl);
+
+	/*
+	 * Retake all the resources we had at the beginning even if the repair
+	 * failed or the scan data are now stale.  This keeps things simple for
+	 * the caller.
+	 */
+	xchk_trans_alloc_empty(sc);
+	xchk_ilock(sc, XFS_ILOCK_EXCL);
+	mutex_lock(&dl->lock);
+
+	if (!error && dl->stale)
+		error = -ESTALE;
+	return error;
+}
+
+/*
+ * Try to fix all the problems.  Returns -ESTALE if the scan data have become
+ * out of date.
+ */
+STATIC int
+xrep_dirtree_fix_problems(
+	struct xchk_dirtree		*dl,
+	struct xchk_dirtree_outcomes	*oc)
+{
+	struct xchk_dirpath		*path;
+	int				error;
+
+	/* Delete all the paths we don't want. */
+	xchk_dirtree_for_each_path(dl, path) {
+		if (path->outcome != XCHK_DIRPATH_DELETE)
+			continue;
+
+		error = xrep_dirtree_delete_path(dl, path);
+		if (error)
+			return error;
+	}
+
+	/* Reparent this directory to the orphanage. */
+	if (oc->needs_adoption) {
+		if (xrep_orphanage_can_adopt(dl->sc))
+			return xrep_dirtree_move_to_orphanage(dl);
+		return -EFSCORRUPTED;
+	}
+
+	return 0;
+}
+
+/* Fix directory loops involving this directory. */
+int
+xrep_dirtree(
+	struct xfs_scrub		*sc)
+{
+	struct xchk_dirtree		*dl = sc->buf;
+	struct xchk_dirtree_outcomes	oc;
+	int				error;
+
+	/*
+	 * Prepare to fix the directory tree by retaking the scan lock.  The
+	 * order of resource acquisition is still IOLOCK -> transaction ->
+	 * ILOCK -> scan lock.
+	 */
+	mutex_lock(&dl->lock);
+	do {
+		/*
+		 * Decide what we're going to do, then do it.  An -ESTALE
+		 * return here means the scan results are invalid and we have
+		 * to walk again.
+		 */
+		if (!dl->stale) {
+			xrep_dirtree_decide_fate(dl, &oc);
+
+			trace_xrep_dirtree_decided_fate(dl, &oc);
+
+			error = xrep_dirtree_fix_problems(dl, &oc);
+			if (!error || error != -ESTALE)
+				break;
+		}
+		error = xchk_dirtree_find_paths_to_root(dl);
+		if (error == -ELNRNG || error == -ENOSR)
+			error = -EFSCORRUPTED;
+	} while (!error);
+	mutex_unlock(&dl->lock);
+
+	return error;
+}
diff --git a/fs/xfs/scrub/orphanage.c b/fs/xfs/scrub/orphanage.c
index b2f905924d0d..b1c6c60ee1da 100644
--- a/fs/xfs/scrub/orphanage.c
+++ b/fs/xfs/scrub/orphanage.c
@@ -570,6 +570,12 @@ xrep_adoption_move(
 		xfs_bumplink(sc->tp, sc->orphanage);
 	xfs_trans_log_inode(sc->tp, sc->orphanage, XFS_ILOG_CORE);
 
+	/* Bump the link count of the child. */
+	if (adopt->bump_child_nlink) {
+		xfs_bumplink(sc->tp, sc->ip);
+		xfs_trans_log_inode(sc->tp, sc->ip, XFS_ILOG_CORE);
+	}
+
 	/* Replace the dotdot entry if the child is a subdirectory. */
 	if (isdir) {
 		error = xfs_dir_replace(sc->tp, sc->ip, &xfs_name_dotdot,
diff --git a/fs/xfs/scrub/orphanage.h b/fs/xfs/scrub/orphanage.h
index beb6b686784e..7c7a2e7d81db 100644
--- a/fs/xfs/scrub/orphanage.h
+++ b/fs/xfs/scrub/orphanage.h
@@ -60,6 +60,14 @@ struct xrep_adoption {
 	/* Block reservations for orphanage and child (if directory). */
 	unsigned int		orphanage_blkres;
 	unsigned int		child_blkres;
+
+	/*
+	 * Does the caller want us to bump the child link count?  This is not
+	 * needed when reattaching files that have become disconnected but have
+	 * nlink > 1.  It is necessary when changing the directory tree
+	 * structure.
+	 */
+	bool			bump_child_nlink:1;
 };
 
 bool xrep_orphanage_can_adopt(struct xfs_scrub *sc);
diff --git a/fs/xfs/scrub/repair.h b/fs/xfs/scrub/repair.h
index 622eb486a16f..0e0dc2bf985c 100644
--- a/fs/xfs/scrub/repair.h
+++ b/fs/xfs/scrub/repair.h
@@ -95,6 +95,7 @@ int xrep_setup_directory(struct xfs_scrub *sc);
 int xrep_setup_parent(struct xfs_scrub *sc);
 int xrep_setup_nlinks(struct xfs_scrub *sc);
 int xrep_setup_symlink(struct xfs_scrub *sc, unsigned int *resblks);
+int xrep_setup_dirtree(struct xfs_scrub *sc);
 
 /* Repair setup functions */
 int xrep_setup_ag_allocbt(struct xfs_scrub *sc);
@@ -132,6 +133,7 @@ int xrep_xattr(struct xfs_scrub *sc);
 int xrep_directory(struct xfs_scrub *sc);
 int xrep_parent(struct xfs_scrub *sc);
 int xrep_symlink(struct xfs_scrub *sc);
+int xrep_dirtree(struct xfs_scrub *sc);
 
 #ifdef CONFIG_XFS_RT
 int xrep_rtbitmap(struct xfs_scrub *sc);
@@ -205,6 +207,7 @@ xrep_setup_nothing(
 #define xrep_setup_directory		xrep_setup_nothing
 #define xrep_setup_parent		xrep_setup_nothing
 #define xrep_setup_nlinks		xrep_setup_nothing
+#define xrep_setup_dirtree		xrep_setup_nothing
 
 #define xrep_setup_inode(sc, imap)	((void)0)
 
@@ -239,6 +242,7 @@ static inline int xrep_setup_symlink(struct xfs_scrub *sc, unsigned int *x)
 #define xrep_directory			xrep_notsupported
 #define xrep_parent			xrep_notsupported
 #define xrep_symlink			xrep_notsupported
+#define xrep_dirtree			xrep_notsupported
 
 #endif /* CONFIG_XFS_ONLINE_REPAIR */
 
diff --git a/fs/xfs/scrub/scrub.c b/fs/xfs/scrub/scrub.c
index 8f1431db7739..e813b66b603a 100644
--- a/fs/xfs/scrub/scrub.c
+++ b/fs/xfs/scrub/scrub.c
@@ -441,7 +441,7 @@ static const struct xchk_meta_ops meta_scrub_ops[] = {
 		.setup	= xchk_setup_dirtree,
 		.scrub	= xchk_dirtree,
 		.has	= xfs_has_parent,
-		.repair	= xrep_notsupported,
+		.repair	= xrep_dirtree,
 	},
 };
 
diff --git a/fs/xfs/scrub/trace.h b/fs/xfs/scrub/trace.h
index 509b6f4fd0cd..b3756722bee1 100644
--- a/fs/xfs/scrub/trace.h
+++ b/fs/xfs/scrub/trace.h
@@ -1685,6 +1685,10 @@ TRACE_DEFINE_ENUM(XCHK_DIRPATH_CORRUPT);
 TRACE_DEFINE_ENUM(XCHK_DIRPATH_LOOP);
 TRACE_DEFINE_ENUM(XCHK_DIRPATH_STALE);
 TRACE_DEFINE_ENUM(XCHK_DIRPATH_OK);
+TRACE_DEFINE_ENUM(XREP_DIRPATH_DELETING);
+TRACE_DEFINE_ENUM(XREP_DIRPATH_DELETED);
+TRACE_DEFINE_ENUM(XREP_DIRPATH_ADOPTING);
+TRACE_DEFINE_ENUM(XREP_DIRPATH_ADOPTED);
 
 #define XCHK_DIRPATH_OUTCOME_STRINGS \
 	{ XCHK_DIRPATH_SCANNING,	"scanning" }, \
@@ -1692,7 +1696,11 @@ TRACE_DEFINE_ENUM(XCHK_DIRPATH_OK);
 	{ XCHK_DIRPATH_CORRUPT,		"corrupt" }, \
 	{ XCHK_DIRPATH_LOOP,		"loop" }, \
 	{ XCHK_DIRPATH_STALE,		"stale" }, \
-	{ XCHK_DIRPATH_OK,		"ok" }
+	{ XCHK_DIRPATH_OK,		"ok" }, \
+	{ XREP_DIRPATH_DELETING,	"deleting" }, \
+	{ XREP_DIRPATH_DELETED,		"deleted" }, \
+	{ XREP_DIRPATH_ADOPTING,	"adopting" }, \
+	{ XREP_DIRPATH_ADOPTED,		"adopted" }
 
 DECLARE_EVENT_CLASS(xchk_dirpath_outcome_class,
 	TP_PROTO(struct xfs_scrub *sc, unsigned long long path_nr,
@@ -1738,6 +1746,7 @@ DECLARE_EVENT_CLASS(xchk_dirtree_evaluate_class,
 		__field(unsigned int, bad)
 		__field(unsigned int, suspect)
 		__field(unsigned int, good)
+		__field(bool, needs_adoption)
 	),
 	TP_fast_assign(
 		__entry->dev = dl->sc->mp->m_super->s_dev;
@@ -1747,15 +1756,17 @@ DECLARE_EVENT_CLASS(xchk_dirtree_evaluate_class,
 		__entry->bad = oc->bad;
 		__entry->suspect = oc->suspect;
 		__entry->good = oc->good;
+		__entry->needs_adoption = oc->needs_adoption ? 1 : 0;
 	),
-	TP_printk("dev %d:%d ino 0x%llx rootino 0x%llx nr_paths %u bad %u suspect %u good %u",
+	TP_printk("dev %d:%d ino 0x%llx rootino 0x%llx nr_paths %u bad %u suspect %u good %u adopt? %d",
 		  MAJOR(__entry->dev), MINOR(__entry->dev),
 		  __entry->ino,
 		  __entry->rootino,
 		  __entry->nr_paths,
 		  __entry->bad,
 		  __entry->suspect,
-		  __entry->good)
+		  __entry->good,
+		  __entry->needs_adoption)
 );
 #define DEFINE_XCHK_DIRTREE_EVALUATE_EVENT(name) \
 DEFINE_EVENT(xchk_dirtree_evaluate_class, name, \
@@ -3181,6 +3192,7 @@ DEFINE_REPAIR_DENTRY_EVENT(xrep_adoption_check_child);
 DEFINE_REPAIR_DENTRY_EVENT(xrep_adoption_check_alias);
 DEFINE_REPAIR_DENTRY_EVENT(xrep_adoption_check_dentry);
 DEFINE_REPAIR_DENTRY_EVENT(xrep_adoption_invalidate_child);
+DEFINE_REPAIR_DENTRY_EVENT(xrep_dirtree_delete_child);
 
 TRACE_EVENT(xrep_symlink_salvage_target,
 	TP_PROTO(struct xfs_inode *ip, char *target, unsigned int targetlen),
@@ -3483,6 +3495,11 @@ TRACE_EVENT(xrep_iunlink_commit_bucket,
 		  __entry->agino)
 );
 
+DEFINE_XCHK_DIRPATH_OUTCOME_EVENT(xrep_dirpath_set_outcome);
+DEFINE_XCHK_DIRTREE_EVENT(xrep_dirtree_delete_path);
+DEFINE_XCHK_DIRTREE_EVENT(xrep_dirtree_create_adoption);
+DEFINE_XCHK_DIRTREE_EVALUATE_EVENT(xrep_dirtree_decided_fate);
+
 #endif /* IS_ENABLED(CONFIG_XFS_ONLINE_REPAIR) */
 
 #endif /* _TRACE_XFS_SCRUB_TRACE_H */
diff --git a/fs/xfs/xfs_inode.c b/fs/xfs/xfs_inode.c
index ebe2ce9bd9ee..58fb7a5062e1 100644
--- a/fs/xfs/xfs_inode.c
+++ b/fs/xfs/xfs_inode.c
@@ -893,7 +893,7 @@ xfs_init_new_inode(
  * link count to go to zero, move the inode to AGI unlinked list so that it can
  * be freed when the last active reference goes away via xfs_inactive().
  */
-static int			/* error */
+int
 xfs_droplink(
 	struct xfs_trans	*tp,
 	struct xfs_inode	*ip)
diff --git a/fs/xfs/xfs_inode.h b/fs/xfs/xfs_inode.h
index 04a91e312993..9fd4d29a5713 100644
--- a/fs/xfs/xfs_inode.h
+++ b/fs/xfs/xfs_inode.h
@@ -626,6 +626,7 @@ void xfs_end_io(struct work_struct *work);
 int xfs_ilock2_io_mmap(struct xfs_inode *ip1, struct xfs_inode *ip2);
 void xfs_iunlock2_io_mmap(struct xfs_inode *ip1, struct xfs_inode *ip2);
 void xfs_iunlock2_remapping(struct xfs_inode *ip1, struct xfs_inode *ip2);
+int xfs_droplink(struct xfs_trans *tp, struct xfs_inode *ip);
 void xfs_bumplink(struct xfs_trans *tp, struct xfs_inode *ip);
 void xfs_lock_inodes(struct xfs_inode **ips, int inodes, uint lock_mode);
 void xfs_sort_inodes(struct xfs_inode **i_tab, unsigned int num_inodes);