Received-SPF: softfail (google.com: domain of transitioning linux-kernel-owner@vger.kernel.org does not designate 23.128.96.19 as permitted sender) client-ip=23.128.96.19;
From:   =?UTF-8?q?Micka=C3=ABl=20Sala=C3=BCn?= <mic@digikod.net>
To:     James Morris <jmorris@namei.org>,
        "Serge E . Hallyn" <serge@hallyn.com>
Cc:     =?UTF-8?q?Micka=C3=ABl=20Sala=C3=BCn?= <mic@digikod.net>,
        Al Viro <viro@zeniv.linux.org.uk>,
        Jann Horn <jannh@google.com>,
        John Johansen <john.johansen@canonical.com>,
        Kees Cook <keescook@chromium.org>,
        Konstantin Meskhidze <konstantin.meskhidze@huawei.com>,
        Paul Moore <paul@paul-moore.com>,
        Shuah Khan <shuah@kernel.org>,
        Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>,
        linux-doc@vger.kernel.org, linux-fsdevel@vger.kernel.org,
        linux-kernel@vger.kernel.org, linux-security-module@vger.kernel.org
Subject: [PATCH v3 07/12] landlock: Add support for file reparenting with LANDLOCK_ACCESS_FS_REFER
Date:   Fri,  6 May 2022 18:10:57 +0200
Message-Id: <20220506161102.525323-8-mic@digikod.net>
In-Reply-To: <20220506161102.525323-1-mic@digikod.net>
References: <20220506161102.525323-1-mic@digikod.net>
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
Precedence: bulk

Add a new LANDLOCK_ACCESS_FS_REFER access right to enable policy writers
to allow sandboxed processes to link and rename files from and to a
specific set of file hierarchies.  This access right should be composed
with LANDLOCK_ACCESS_FS_MAKE_* for the destination of a link or rename,
and with LANDLOCK_ACCESS_FS_REMOVE_* for a source of a rename.  This
lift a Landlock limitation that always denied changing the parent of an
inode.

Renaming or linking to the same directory is still always allowed,
whatever LANDLOCK_ACCESS_FS_REFER is used or not, because it is not
considered a threat to user data.

However, creating multiple links or renaming to a different parent
directory may lead to privilege escalations if not handled properly.
Indeed, we must be sure that the source doesn't gain more privileges by
being accessible from the destination.  This is handled by making sure
that the source hierarchy (including the referenced file or directory
itself) restricts at least as much the destination hierarchy.  If it is
not the case, an EXDEV error is returned, making it potentially possible
for user space to copy the file hierarchy instead of moving or linking
it.

Instead of creating different access rights for the source and the
destination, we choose to make it simple and consistent for users.
Indeed, considering the previous constraint, it would be weird to
require such destination access right to be also granted to the source
(to make it a superset).  Moreover, RENAME_EXCHANGE would also add to
the confusion because of paths being both a source and a destination.

See the provided documentation for additional details.

New tests are provided with a following commit.

Reviewed-by: Paul Moore <paul@paul-moore.com>
Signed-off-by: Mickaël Salaün <mic@digikod.net>
Link: https://lore.kernel.org/r/20220506161102.525323-8-mic@digikod.net
---

Changes since v2:
* Add Reviewed-by: Paul Moore.
* Update collect_domain_accesses() documentation for the RENAME_EXCHANGE
  case and fix spelling.
* Add unlikely() branches in check_access_path_dual() to favor single
  access request (i.e. non-refer actions).
* Remove useless ret assignment in collect_domain_accesses().
* Format with clang-format and rebase.

Changes since v1:
* Update current_check_access_path() to efficiently handle
  RENAME_EXCHANGE thanks to the updated LSM hook (see previous patch).
  Only one path walk is performed per rename arguments until their
  common mount point is reached.  Superset of access rights is correctly
  checked, including when exchanging a file with a directory.  This
  requires to store another matrix of layer masks.
* Reorder and rename check_access_path_dual() arguments in a more
  generic way: switch from src/dst to 1/2.  This makes it easier to
  understand the RENAME_EXCHANGE cases alongs with the others.  Update
  and improve check_access_path_dual() documentation accordingly.
* Clean up the check_access_path_dual() loop: set both allowed_parent*
  when reaching internal filesystems and remove a useless one.  This
  allows potential renames in internal filesystems (like for other
  operations).
* Move the function arguments checks from BUILD_BUG_ON() to
  WARN_ON_ONCE() to avoid clang build error.
* Rename is_superset() to no_more_access() and make it handle superset
  checks of source and destination for simple and exchange cases.
* Move the layer_masks_child* creation from current_check_refer_path()
  to check_access_path_dual(): this is simpler and less error-prone,
  especially with RENAME_EXCHANGE.
* Remove one optimization in current_check_refer_path() to make the code
  simpler, especially with the RENAME_EXCHANGE handling.
* Remove overzealous WARN_ON_ONCE() for !access_request check in
  init_layer_masks().
---
 include/uapi/linux/landlock.h                |  27 +-
 security/landlock/fs.c                       | 600 ++++++++++++++++---
 security/landlock/limits.h                   |   2 +-
 security/landlock/syscalls.c                 |   2 +-
 tools/testing/selftests/landlock/base_test.c |   2 +-
 tools/testing/selftests/landlock/fs_test.c   |   3 +-
 6 files changed, 556 insertions(+), 80 deletions(-)

diff --git a/include/uapi/linux/landlock.h b/include/uapi/linux/landlock.h
index 21c8d58283c9..23df4e0e8ace 100644
--- a/include/uapi/linux/landlock.h
+++ b/include/uapi/linux/landlock.h
@@ -21,8 +21,14 @@ struct landlock_ruleset_attr {
 	/**
 	 * @handled_access_fs: Bitmask of actions (cf. `Filesystem flags`_)
 	 * that is handled by this ruleset and should then be forbidden if no
-	 * rule explicitly allow them.  This is needed for backward
-	 * compatibility reasons.
+	 * rule explicitly allow them: it is a deny-by-default list that should
+	 * contain as much Landlock access rights as possible. Indeed, all
+	 * Landlock filesystem access rights that are not part of
+	 * handled_access_fs are allowed.  This is needed for backward
+	 * compatibility reasons.  One exception is the
+	 * LANDLOCK_ACCESS_FS_REFER access right, which is always implicitly
+	 * handled, but must still be explicitly handled to add new rules with
+	 * this access right.
 	 */
 	__u64 handled_access_fs;
 };
@@ -112,6 +118,22 @@ struct landlock_path_beneath_attr {
  * - %LANDLOCK_ACCESS_FS_MAKE_FIFO: Create (or rename or link) a named pipe.
  * - %LANDLOCK_ACCESS_FS_MAKE_BLOCK: Create (or rename or link) a block device.
  * - %LANDLOCK_ACCESS_FS_MAKE_SYM: Create (or rename or link) a symbolic link.
+ * - %LANDLOCK_ACCESS_FS_REFER: Link or rename a file from or to a different
+ *   directory (i.e. reparent a file hierarchy).  This access right is
+ *   available since the second version of the Landlock ABI.  This is also the
+ *   only access right which is always considered handled by any ruleset in
+ *   such a way that reparenting a file hierarchy is always denied by default.
+ *   To avoid privilege escalation, it is not enough to add a rule with this
+ *   access right.  When linking or renaming a file, the destination directory
+ *   hierarchy must also always have the same or a superset of restrictions of
+ *   the source hierarchy.  If it is not the case, or if the domain doesn't
+ *   handle this access right, such actions are denied by default with errno
+ *   set to EXDEV.  Linking also requires a LANDLOCK_ACCESS_FS_MAKE_* access
+ *   right on the destination directory, and renaming also requires a
+ *   LANDLOCK_ACCESS_FS_REMOVE_* access right on the source's (file or
+ *   directory) parent.  Otherwise, such actions are denied with errno set to
+ *   EACCES.  The EACCES errno prevails over EXDEV to let user space
+ *   efficiently deal with an unrecoverable error.
  *
  * .. warning::
  *
@@ -137,6 +159,7 @@ struct landlock_path_beneath_attr {
 #define LANDLOCK_ACCESS_FS_MAKE_FIFO			(1ULL << 10)
 #define LANDLOCK_ACCESS_FS_MAKE_BLOCK			(1ULL << 11)
 #define LANDLOCK_ACCESS_FS_MAKE_SYM			(1ULL << 12)
+#define LANDLOCK_ACCESS_FS_REFER			(1ULL << 13)
 /* clang-format on */
 
 #endif /* _UAPI_LINUX_LANDLOCK_H */
diff --git a/security/landlock/fs.c b/security/landlock/fs.c
index 30b42cdee52e..ec5a6247cd3e 100644
--- a/security/landlock/fs.c
+++ b/security/landlock/fs.c
@@ -4,6 +4,7 @@
  *
  * Copyright © 2016-2020 Mickaël Salaün <mic@digikod.net>
  * Copyright © 2018-2020 ANSSI
+ * Copyright © 2021-2022 Microsoft Corporation
  */
 
 #include <linux/atomic.h>
@@ -273,40 +274,262 @@ static inline bool is_nouser_or_private(const struct dentry *dentry)
 		unlikely(IS_PRIVATE(d_backing_inode(dentry))));
 }
 
-static int check_access_path(const struct landlock_ruleset *const domain,
-			     const struct path *const path,
-			     const access_mask_t access_request)
+static inline access_mask_t
+get_handled_accesses(const struct landlock_ruleset *const domain)
 {
-	layer_mask_t layer_masks[LANDLOCK_NUM_ACCESS_FS] = {};
-	bool allowed = false, has_access = false;
-	struct path walker_path;
-	size_t i;
+	access_mask_t access_dom = 0;
+	unsigned long access_bit;
+
+	for (access_bit = 0; access_bit < LANDLOCK_NUM_ACCESS_FS;
+	     access_bit++) {
+		size_t layer_level;
+
+		for (layer_level = 0; layer_level < domain->num_layers;
+		     layer_level++) {
+			if (domain->fs_access_masks[layer_level] &
+			    BIT_ULL(access_bit)) {
+				access_dom |= BIT_ULL(access_bit);
+				break;
+			}
+		}
+	}
+	return access_dom;
+}
+
+static inline access_mask_t
+init_layer_masks(const struct landlock_ruleset *const domain,
+		 const access_mask_t access_request,
+		 layer_mask_t (*const layer_masks)[LANDLOCK_NUM_ACCESS_FS])
+{
+	access_mask_t handled_accesses = 0;
+	size_t layer_level;
 
+	memset(layer_masks, 0, sizeof(*layer_masks));
+	/* An empty access request can happen because of O_WRONLY | O_RDWR. */
 	if (!access_request)
 		return 0;
-	if (WARN_ON_ONCE(!domain || !path))
-		return 0;
-	if (is_nouser_or_private(path->dentry))
-		return 0;
-	if (WARN_ON_ONCE(domain->num_layers < 1))
-		return -EACCES;
 
-	/* Saves all layers handling a subset of requested accesses. */
-	for (i = 0; i < domain->num_layers; i++) {
+	/* Saves all handled accesses per layer. */
+	for (layer_level = 0; layer_level < domain->num_layers; layer_level++) {
 		const unsigned long access_req = access_request;
 		unsigned long access_bit;
 
 		for_each_set_bit(access_bit, &access_req,
-				 ARRAY_SIZE(layer_masks)) {
-			if (domain->fs_access_masks[i] & BIT_ULL(access_bit)) {
-				layer_masks[access_bit] |= BIT_ULL(i);
-				has_access = true;
+				 ARRAY_SIZE(*layer_masks)) {
+			if (domain->fs_access_masks[layer_level] &
+			    BIT_ULL(access_bit)) {
+				(*layer_masks)[access_bit] |=
+					BIT_ULL(layer_level);
+				handled_accesses |= BIT_ULL(access_bit);
 			}
 		}
 	}
-	/* An access request not handled by the domain is allowed. */
-	if (!has_access)
+	return handled_accesses;
+}
+
+/*
+ * Check that a destination file hierarchy has more restrictions than a source
+ * file hierarchy.  This is only used for link and rename actions.
+ *
+ * @layer_masks_child2: Optional child masks.
+ */
+static inline bool no_more_access(
+	const layer_mask_t (*const layer_masks_parent1)[LANDLOCK_NUM_ACCESS_FS],
+	const layer_mask_t (*const layer_masks_child1)[LANDLOCK_NUM_ACCESS_FS],
+	const bool child1_is_directory,
+	const layer_mask_t (*const layer_masks_parent2)[LANDLOCK_NUM_ACCESS_FS],
+	const layer_mask_t (*const layer_masks_child2)[LANDLOCK_NUM_ACCESS_FS],
+	const bool child2_is_directory)
+{
+	unsigned long access_bit;
+
+	for (access_bit = 0; access_bit < ARRAY_SIZE(*layer_masks_parent2);
+	     access_bit++) {
+		/* Ignores accesses that only make sense for directories. */
+		const bool is_file_access =
+			!!(BIT_ULL(access_bit) & ACCESS_FILE);
+
+		if (child1_is_directory || is_file_access) {
+			/*
+			 * Checks if the destination restrictions are a
+			 * superset of the source ones (i.e. inherited access
+			 * rights without child exceptions):
+			 * restrictions(parent2) >= restrictions(child1)
+			 */
+			if ((((*layer_masks_parent1)[access_bit] &
+			      (*layer_masks_child1)[access_bit]) |
+			     (*layer_masks_parent2)[access_bit]) !=
+			    (*layer_masks_parent2)[access_bit])
+				return false;
+		}
+
+		if (!layer_masks_child2)
+			continue;
+		if (child2_is_directory || is_file_access) {
+			/*
+			 * Checks inverted restrictions for RENAME_EXCHANGE:
+			 * restrictions(parent1) >= restrictions(child2)
+			 */
+			if ((((*layer_masks_parent2)[access_bit] &
+			      (*layer_masks_child2)[access_bit]) |
+			     (*layer_masks_parent1)[access_bit]) !=
+			    (*layer_masks_parent1)[access_bit])
+				return false;
+		}
+	}
+	return true;
+}
+
+/*
+ * Removes @layer_masks accesses that are not requested.
+ *
+ * Returns true if the request is allowed, false otherwise.
+ */
+static inline bool
+scope_to_request(const access_mask_t access_request,
+		 layer_mask_t (*const layer_masks)[LANDLOCK_NUM_ACCESS_FS])
+{
+	const unsigned long access_req = access_request;
+	unsigned long access_bit;
+
+	if (WARN_ON_ONCE(!layer_masks))
+		return true;
+
+	for_each_clear_bit(access_bit, &access_req, ARRAY_SIZE(*layer_masks))
+		(*layer_masks)[access_bit] = 0;
+	return !memchr_inv(layer_masks, 0, sizeof(*layer_masks));
+}
+
+/*
+ * Returns true if there is at least one access right different than
+ * LANDLOCK_ACCESS_FS_REFER.
+ */
+static inline bool
+is_eacces(const layer_mask_t (*const layer_masks)[LANDLOCK_NUM_ACCESS_FS],
+	  const access_mask_t access_request)
+{
+	unsigned long access_bit;
+	/* LANDLOCK_ACCESS_FS_REFER alone must return -EXDEV. */
+	const unsigned long access_check = access_request &
+					   ~LANDLOCK_ACCESS_FS_REFER;
+
+	if (!layer_masks)
+		return false;
+
+	for_each_set_bit(access_bit, &access_check, ARRAY_SIZE(*layer_masks)) {
+		if ((*layer_masks)[access_bit])
+			return true;
+	}
+	return false;
+}
+
+/**
+ * check_access_path_dual - Check accesses for requests with a common path
+ *
+ * @domain: Domain to check against.
+ * @path: File hierarchy to walk through.
+ * @access_request_parent1: Accesses to check, once @layer_masks_parent1 is
+ *     equal to @layer_masks_parent2 (if any).  This is tied to the unique
+ *     requested path for most actions, or the source in case of a refer action
+ *     (i.e. rename or link), or the source and destination in case of
+ *     RENAME_EXCHANGE.
+ * @layer_masks_parent1: Pointer to a matrix of layer masks per access
+ *     masks, identifying the layers that forbid a specific access.  Bits from
+ *     this matrix can be unset according to the @path walk.  An empty matrix
+ *     means that @domain allows all possible Landlock accesses (i.e. not only
+ *     those identified by @access_request_parent1).  This matrix can
+ *     initially refer to domain layer masks and, when the accesses for the
+ *     destination and source are the same, to requested layer masks.
+ * @dentry_child1: Dentry to the initial child of the parent1 path.  This
+ *     pointer must be NULL for non-refer actions (i.e. not link nor rename).
+ * @access_request_parent2: Similar to @access_request_parent1 but for a
+ *     request involving a source and a destination.  This refers to the
+ *     destination, except in case of RENAME_EXCHANGE where it also refers to
+ *     the source.  Must be set to 0 when using a simple path request.
+ * @layer_masks_parent2: Similar to @layer_masks_parent1 but for a refer
+ *     action.  This must be NULL otherwise.
+ * @dentry_child2: Dentry to the initial child of the parent2 path.  This
+ *     pointer is only set for RENAME_EXCHANGE actions and must be NULL
+ *     otherwise.
+ *
+ * This helper first checks that the destination has a superset of restrictions
+ * compared to the source (if any) for a common path.  Because of
+ * RENAME_EXCHANGE actions, source and destinations may be swapped.  It then
+ * checks that the collected accesses and the remaining ones are enough to
+ * allow the request.
+ *
+ * Returns:
+ * - 0 if the access request is granted;
+ * - -EACCES if it is denied because of access right other than
+ *   LANDLOCK_ACCESS_FS_REFER;
+ * - -EXDEV if the renaming or linking would be a privileged escalation
+ *   (according to each layered policies), or if LANDLOCK_ACCESS_FS_REFER is
+ *   not allowed by the source or the destination.
+ */
+static int check_access_path_dual(
+	const struct landlock_ruleset *const domain,
+	const struct path *const path,
+	const access_mask_t access_request_parent1,
+	layer_mask_t (*const layer_masks_parent1)[LANDLOCK_NUM_ACCESS_FS],
+	const struct dentry *const dentry_child1,
+	const access_mask_t access_request_parent2,
+	layer_mask_t (*const layer_masks_parent2)[LANDLOCK_NUM_ACCESS_FS],
+	const struct dentry *const dentry_child2)
+{
+	bool allowed_parent1 = false, allowed_parent2 = false, is_dom_check,
+	     child1_is_directory = true, child2_is_directory = true;
+	struct path walker_path;
+	access_mask_t access_masked_parent1, access_masked_parent2;
+	layer_mask_t _layer_masks_child1[LANDLOCK_NUM_ACCESS_FS],
+		_layer_masks_child2[LANDLOCK_NUM_ACCESS_FS];
+	layer_mask_t(*layer_masks_child1)[LANDLOCK_NUM_ACCESS_FS] = NULL,
+	(*layer_masks_child2)[LANDLOCK_NUM_ACCESS_FS] = NULL;
+
+	if (!access_request_parent1 && !access_request_parent2)
 		return 0;
+	if (WARN_ON_ONCE(!domain || !path))
+		return 0;
+	if (is_nouser_or_private(path->dentry))
+		return 0;
+	if (WARN_ON_ONCE(domain->num_layers < 1 || !layer_masks_parent1))
+		return -EACCES;
+
+	if (unlikely(layer_masks_parent2)) {
+		if (WARN_ON_ONCE(!dentry_child1))
+			return -EACCES;
+		/*
+		 * For a double request, first check for potential privilege
+		 * escalation by looking at domain handled accesses (which are
+		 * a superset of the meaningful requested accesses).
+		 */
+		access_masked_parent1 = access_masked_parent2 =
+			get_handled_accesses(domain);
+		is_dom_check = true;
+	} else {
+		if (WARN_ON_ONCE(dentry_child1 || dentry_child2))
+			return -EACCES;
+		/* For a simple request, only check for requested accesses. */
+		access_masked_parent1 = access_request_parent1;
+		access_masked_parent2 = access_request_parent2;
+		is_dom_check = false;
+	}
+
+	if (unlikely(dentry_child1)) {
+		unmask_layers(find_rule(domain, dentry_child1),
+			      init_layer_masks(domain, LANDLOCK_MASK_ACCESS_FS,
+					       &_layer_masks_child1),
+			      &_layer_masks_child1);
+		layer_masks_child1 = &_layer_masks_child1;
+		child1_is_directory = d_is_dir(dentry_child1);
+	}
+	if (unlikely(dentry_child2)) {
+		unmask_layers(find_rule(domain, dentry_child2),
+			      init_layer_masks(domain, LANDLOCK_MASK_ACCESS_FS,
+					       &_layer_masks_child2),
+			      &_layer_masks_child2);
+		layer_masks_child2 = &_layer_masks_child2;
+		child2_is_directory = d_is_dir(dentry_child2);
+	}
 
 	walker_path = *path;
 	path_get(&walker_path);
@@ -316,11 +539,52 @@ static int check_access_path(const struct landlock_ruleset *const domain,
 	 */
 	while (true) {
 		struct dentry *parent_dentry;
+		const struct landlock_rule *rule;
+
+		/*
+		 * If at least all accesses allowed on the destination are
+		 * already allowed on the source, respectively if there is at
+		 * least as much as restrictions on the destination than on the
+		 * source, then we can safely refer files from the source to
+		 * the destination without risking a privilege escalation.
+		 * This also applies in the case of RENAME_EXCHANGE, which
+		 * implies checks on both direction.  This is crucial for
+		 * standalone multilayered security policies.  Furthermore,
+		 * this helps avoid policy writers to shoot themselves in the
+		 * foot.
+		 */
+		if (unlikely(is_dom_check &&
+			     no_more_access(
+				     layer_masks_parent1, layer_masks_child1,
+				     child1_is_directory, layer_masks_parent2,
+				     layer_masks_child2,
+				     child2_is_directory))) {
+			allowed_parent1 = scope_to_request(
+				access_request_parent1, layer_masks_parent1);
+			allowed_parent2 = scope_to_request(
+				access_request_parent2, layer_masks_parent2);
+
+			/* Stops when all accesses are granted. */
+			if (allowed_parent1 && allowed_parent2)
+				break;
 
-		allowed = unmask_layers(find_rule(domain, walker_path.dentry),
-					access_request, &layer_masks);
-		if (allowed)
-			/* Stops when a rule from each layer grants access. */
+			/*
+			 * Now, downgrades the remaining checks from domain
+			 * handled accesses to requested accesses.
+			 */
+			is_dom_check = false;
+			access_masked_parent1 = access_request_parent1;
+			access_masked_parent2 = access_request_parent2;
+		}
+
+		rule = find_rule(domain, walker_path.dentry);
+		allowed_parent1 = unmask_layers(rule, access_masked_parent1,
+						layer_masks_parent1);
+		allowed_parent2 = unmask_layers(rule, access_masked_parent2,
+						layer_masks_parent2);
+
+		/* Stops when a rule from each layer grants access. */
+		if (allowed_parent1 && allowed_parent2)
 			break;
 
 jump_up:
@@ -333,7 +597,6 @@ static int check_access_path(const struct landlock_ruleset *const domain,
 				 * Stops at the real root.  Denies access
 				 * because not all layers have granted access.
 				 */
-				allowed = false;
 				break;
 			}
 		}
@@ -343,7 +606,8 @@ static int check_access_path(const struct landlock_ruleset *const domain,
 			 * access to internal filesystems (e.g. nsfs, which is
 			 * reachable through /proc/<pid>/ns/<namespace>).
 			 */
-			allowed = !!(walker_path.mnt->mnt_flags & MNT_INTERNAL);
+			allowed_parent1 = allowed_parent2 =
+				!!(walker_path.mnt->mnt_flags & MNT_INTERNAL);
 			break;
 		}
 		parent_dentry = dget_parent(walker_path.dentry);
@@ -351,7 +615,36 @@ static int check_access_path(const struct landlock_ruleset *const domain,
 		walker_path.dentry = parent_dentry;
 	}
 	path_put(&walker_path);
-	return allowed ? 0 : -EACCES;
+
+	if (allowed_parent1 && allowed_parent2)
+		return 0;
+
+	/*
+	 * This prioritizes EACCES over EXDEV for all actions, including
+	 * renames with RENAME_EXCHANGE.
+	 */
+	if (likely(is_eacces(layer_masks_parent1, access_request_parent1) ||
+		   is_eacces(layer_masks_parent2, access_request_parent2)))
+		return -EACCES;
+
+	/*
+	 * Gracefully forbids reparenting if the destination directory
+	 * hierarchy is not a superset of restrictions of the source directory
+	 * hierarchy, or if LANDLOCK_ACCESS_FS_REFER is not allowed by the
+	 * source or the destination.
+	 */
+	return -EXDEV;
+}
+
+static inline int check_access_path(const struct landlock_ruleset *const domain,
+				    const struct path *const path,
+				    access_mask_t access_request)
+{
+	layer_mask_t layer_masks[LANDLOCK_NUM_ACCESS_FS] = {};
+
+	access_request = init_layer_masks(domain, access_request, &layer_masks);
+	return check_access_path_dual(domain, path, access_request,
+				      &layer_masks, NULL, 0, NULL, NULL);
 }
 
 static inline int current_check_access_path(const struct path *const path,
@@ -398,6 +691,206 @@ static inline access_mask_t maybe_remove(const struct dentry *const dentry)
 				  LANDLOCK_ACCESS_FS_REMOVE_FILE;
 }
 
+/**
+ * collect_domain_accesses - Walk through a file path and collect accesses
+ *
+ * @domain: Domain to check against.
+ * @mnt_root: Last directory to check.
+ * @dir: Directory to start the walk from.
+ * @layer_masks_dom: Where to store the collected accesses.
+ *
+ * This helper is useful to begin a path walk from the @dir directory to a
+ * @mnt_root directory used as a mount point.  This mount point is the common
+ * ancestor between the source and the destination of a renamed and linked
+ * file.  While walking from @dir to @mnt_root, we record all the domain's
+ * allowed accesses in @layer_masks_dom.
+ *
+ * This is similar to check_access_path_dual() but much simpler because it only
+ * handles walking on the same mount point and only check one set of accesses.
+ *
+ * Returns:
+ * - true if all the domain access rights are allowed for @dir;
+ * - false if the walk reached @mnt_root.
+ */
+static bool collect_domain_accesses(
+	const struct landlock_ruleset *const domain,
+	const struct dentry *const mnt_root, struct dentry *dir,
+	layer_mask_t (*const layer_masks_dom)[LANDLOCK_NUM_ACCESS_FS])
+{
+	unsigned long access_dom;
+	bool ret = false;
+
+	if (WARN_ON_ONCE(!domain || !mnt_root || !dir || !layer_masks_dom))
+		return true;
+	if (is_nouser_or_private(dir))
+		return true;
+
+	access_dom = init_layer_masks(domain, LANDLOCK_MASK_ACCESS_FS,
+				      layer_masks_dom);
+
+	dget(dir);
+	while (true) {
+		struct dentry *parent_dentry;
+
+		/* Gets all layers allowing all domain accesses. */
+		if (unmask_layers(find_rule(domain, dir), access_dom,
+				  layer_masks_dom)) {
+			/*
+			 * Stops when all handled accesses are allowed by at
+			 * least one rule in each layer.
+			 */
+			ret = true;
+			break;
+		}
+
+		/* We should not reach a root other than @mnt_root. */
+		if (dir == mnt_root || WARN_ON_ONCE(IS_ROOT(dir)))
+			break;
+
+		parent_dentry = dget_parent(dir);
+		dput(dir);
+		dir = parent_dentry;
+	}
+	dput(dir);
+	return ret;
+}
+
+/**
+ * current_check_refer_path - Check if a rename or link action is allowed
+ *
+ * @old_dentry: File or directory requested to be moved or linked.
+ * @new_dir: Destination parent directory.
+ * @new_dentry: Destination file or directory.
+ * @removable: Sets to true if it is a rename operation.
+ * @exchange: Sets to true if it is a rename operation with RENAME_EXCHANGE.
+ *
+ * Because of its unprivileged constraints, Landlock relies on file hierarchies
+ * (and not only inodes) to tie access rights to files.  Being able to link or
+ * rename a file hierarchy brings some challenges.  Indeed, moving or linking a
+ * file (i.e. creating a new reference to an inode) can have an impact on the
+ * actions allowed for a set of files if it would change its parent directory
+ * (i.e. reparenting).
+ *
+ * To avoid trivial access right bypasses, Landlock first checks if the file or
+ * directory requested to be moved would gain new access rights inherited from
+ * its new hierarchy.  Before returning any error, Landlock then checks that
+ * the parent source hierarchy and the destination hierarchy would allow the
+ * link or rename action.  If it is not the case, an error with EACCES is
+ * returned to inform user space that there is no way to remove or create the
+ * requested source file type.  If it should be allowed but the new inherited
+ * access rights would be greater than the source access rights, then the
+ * kernel returns an error with EXDEV.  Prioritizing EACCES over EXDEV enables
+ * user space to abort the whole operation if there is no way to do it, or to
+ * manually copy the source to the destination if this remains allowed, e.g.
+ * because file creation is allowed on the destination directory but not direct
+ * linking.
+ *
+ * To achieve this goal, the kernel needs to compare two file hierarchies: the
+ * one identifying the source file or directory (including itself), and the
+ * destination one.  This can be seen as a multilayer partial ordering problem.
+ * The kernel walks through these paths and collects in a matrix the access
+ * rights that are denied per layer.  These matrices are then compared to see
+ * if the destination one has more (or the same) restrictions as the source
+ * one.  If this is the case, the requested action will not return EXDEV, which
+ * doesn't mean the action is allowed.  The parent hierarchy of the source
+ * (i.e. parent directory), and the destination hierarchy must also be checked
+ * to verify that they explicitly allow such action (i.e.  referencing,
+ * creation and potentially removal rights).  The kernel implementation is then
+ * required to rely on potentially four matrices of access rights: one for the
+ * source file or directory (i.e. the child), a potentially other one for the
+ * other source/destination (in case of RENAME_EXCHANGE), one for the source
+ * parent hierarchy and a last one for the destination hierarchy.  These
+ * ephemeral matrices take some space on the stack, which limits the number of
+ * layers to a deemed reasonable number: 16.
+ *
+ * Returns:
+ * - 0 if access is allowed;
+ * - -EXDEV if @old_dentry would inherit new access rights from @new_dir;
+ * - -EACCES if file removal or creation is denied.
+ */
+static int current_check_refer_path(struct dentry *const old_dentry,
+				    const struct path *const new_dir,
+				    struct dentry *const new_dentry,
+				    const bool removable, const bool exchange)
+{
+	const struct landlock_ruleset *const dom =
+		landlock_get_current_domain();
+	bool allow_parent1, allow_parent2;
+	access_mask_t access_request_parent1, access_request_parent2;
+	struct path mnt_dir;
+	layer_mask_t layer_masks_parent1[LANDLOCK_NUM_ACCESS_FS],
+		layer_masks_parent2[LANDLOCK_NUM_ACCESS_FS];
+
+	if (!dom)
+		return 0;
+	if (WARN_ON_ONCE(dom->num_layers < 1))
+		return -EACCES;
+	if (unlikely(d_is_negative(old_dentry)))
+		return -ENOENT;
+	if (exchange) {
+		if (unlikely(d_is_negative(new_dentry)))
+			return -ENOENT;
+		access_request_parent1 =
+			get_mode_access(d_backing_inode(new_dentry)->i_mode);
+	} else {
+		access_request_parent1 = 0;
+	}
+	access_request_parent2 =
+		get_mode_access(d_backing_inode(old_dentry)->i_mode);
+	if (removable) {
+		access_request_parent1 |= maybe_remove(old_dentry);
+		access_request_parent2 |= maybe_remove(new_dentry);
+	}
+
+	/* The mount points are the same for old and new paths, cf. EXDEV. */
+	if (old_dentry->d_parent == new_dir->dentry) {
+		/*
+		 * The LANDLOCK_ACCESS_FS_REFER access right is not required
+		 * for same-directory referer (i.e. no reparenting).
+		 */
+		access_request_parent1 = init_layer_masks(
+			dom, access_request_parent1 | access_request_parent2,
+			&layer_masks_parent1);
+		return check_access_path_dual(dom, new_dir,
+					      access_request_parent1,
+					      &layer_masks_parent1, NULL, 0,
+					      NULL, NULL);
+	}
+
+	/* Backward compatibility: no reparenting support. */
+	if (!(get_handled_accesses(dom) & LANDLOCK_ACCESS_FS_REFER))
+		return -EXDEV;
+
+	access_request_parent1 |= LANDLOCK_ACCESS_FS_REFER;
+	access_request_parent2 |= LANDLOCK_ACCESS_FS_REFER;
+
+	/* Saves the common mount point. */
+	mnt_dir.mnt = new_dir->mnt;
+	mnt_dir.dentry = new_dir->mnt->mnt_root;
+
+	/* new_dir->dentry is equal to new_dentry->d_parent */
+	allow_parent1 = collect_domain_accesses(dom, mnt_dir.dentry,
+						old_dentry->d_parent,
+						&layer_masks_parent1);
+	allow_parent2 = collect_domain_accesses(
+		dom, mnt_dir.dentry, new_dir->dentry, &layer_masks_parent2);
+
+	if (allow_parent1 && allow_parent2)
+		return 0;
+
+	/*
+	 * To be able to compare source and destination domain access rights,
+	 * take into account the @old_dentry access rights aggregated with its
+	 * parent access rights.  This will be useful to compare with the
+	 * destination parent access rights.
+	 */
+	return check_access_path_dual(dom, &mnt_dir, access_request_parent1,
+				      &layer_masks_parent1, old_dentry,
+				      access_request_parent2,
+				      &layer_masks_parent2,
+				      exchange ? new_dentry : NULL);
+}
+
 /* Inode hooks */
 
 static void hook_inode_free_security(struct inode *const inode)
@@ -591,32 +1084,12 @@ static int hook_sb_pivotroot(const struct path *const old_path,
 
 /* Path hooks */
 
-/*
- * Creating multiple links or renaming may lead to privilege escalations if not
- * handled properly.  Indeed, we must be sure that the source doesn't gain more
- * privileges by being accessible from the destination.  This is getting more
- * complex when dealing with multiple layers.  The whole picture can be seen as
- * a multilayer partial ordering problem.  A future version of Landlock will
- * deal with that.
- */
 static int hook_path_link(struct dentry *const old_dentry,
 			  const struct path *const new_dir,
 			  struct dentry *const new_dentry)
 {
-	const struct landlock_ruleset *const dom =
-		landlock_get_current_domain();
-
-	if (!dom)
-		return 0;
-	/* The mount points are the same for old and new paths, cf. EXDEV. */
-	if (old_dentry->d_parent != new_dir->dentry)
-		/* Gracefully forbids reparenting. */
-		return -EXDEV;
-	if (unlikely(d_is_negative(old_dentry)))
-		return -ENOENT;
-	return check_access_path(
-		dom, new_dir,
-		get_mode_access(d_backing_inode(old_dentry)->i_mode));
+	return current_check_refer_path(old_dentry, new_dir, new_dentry, false,
+					false);
 }
 
 static int hook_path_rename(const struct path *const old_dir,
@@ -625,30 +1098,9 @@ static int hook_path_rename(const struct path *const old_dir,
 			    struct dentry *const new_dentry,
 			    const unsigned int flags)
 {
-	const struct landlock_ruleset *const dom =
-		landlock_get_current_domain();
-	u32 exchange_access = 0;
-
-	if (!dom)
-		return 0;
-	/* The mount points are the same for old and new paths, cf. EXDEV. */
-	if (old_dir->dentry != new_dir->dentry)
-		/* Gracefully forbids reparenting. */
-		return -EXDEV;
-	if (flags & RENAME_EXCHANGE) {
-		if (unlikely(d_is_negative(new_dentry)))
-			return -ENOENT;
-		exchange_access =
-			get_mode_access(d_backing_inode(new_dentry)->i_mode);
-	}
-	if (unlikely(d_is_negative(old_dentry)))
-		return -ENOENT;
-	/* RENAME_EXCHANGE is handled because directories are the same. */
-	return check_access_path(
-		dom, old_dir,
-		maybe_remove(old_dentry) | maybe_remove(new_dentry) |
-			exchange_access |
-			get_mode_access(d_backing_inode(old_dentry)->i_mode));
+	/* old_dir refers to old_dentry->d_parent and new_dir->mnt */
+	return current_check_refer_path(old_dentry, new_dir, new_dentry, true,
+					!!(flags & RENAME_EXCHANGE));
 }
 
 static int hook_path_mkdir(const struct path *const dir,
diff --git a/security/landlock/limits.h b/security/landlock/limits.h
index 17c2a2e7fe1e..b54184ab9439 100644
--- a/security/landlock/limits.h
+++ b/security/landlock/limits.h
@@ -18,7 +18,7 @@
 #define LANDLOCK_MAX_NUM_LAYERS		16
 #define LANDLOCK_MAX_NUM_RULES		U32_MAX
 
-#define LANDLOCK_LAST_ACCESS_FS		LANDLOCK_ACCESS_FS_MAKE_SYM
+#define LANDLOCK_LAST_ACCESS_FS		LANDLOCK_ACCESS_FS_REFER
 #define LANDLOCK_MASK_ACCESS_FS		((LANDLOCK_LAST_ACCESS_FS << 1) - 1)
 #define LANDLOCK_NUM_ACCESS_FS		__const_hweight64(LANDLOCK_MASK_ACCESS_FS)
 
diff --git a/security/landlock/syscalls.c b/security/landlock/syscalls.c
index 507d43827afe..735a0865ea11 100644
--- a/security/landlock/syscalls.c
+++ b/security/landlock/syscalls.c
@@ -129,7 +129,7 @@ static const struct file_operations ruleset_fops = {
 	.write = fop_dummy_write,
 };
 
-#define LANDLOCK_ABI_VERSION 1
+#define LANDLOCK_ABI_VERSION 2
 
 /**
  * sys_landlock_create_ruleset - Create a new ruleset
diff --git a/tools/testing/selftests/landlock/base_test.c b/tools/testing/selftests/landlock/base_test.c
index 35f64832b869..da9290817866 100644
--- a/tools/testing/selftests/landlock/base_test.c
+++ b/tools/testing/selftests/landlock/base_test.c
@@ -75,7 +75,7 @@ TEST(abi_version)
 	const struct landlock_ruleset_attr ruleset_attr = {
 		.handled_access_fs = LANDLOCK_ACCESS_FS_READ_FILE,
 	};
-	ASSERT_EQ(1, landlock_create_ruleset(NULL, 0,
+	ASSERT_EQ(2, landlock_create_ruleset(NULL, 0,
 					     LANDLOCK_CREATE_RULESET_VERSION));
 
 	ASSERT_EQ(-1, landlock_create_ruleset(&ruleset_attr, 0,
diff --git a/tools/testing/selftests/landlock/fs_test.c b/tools/testing/selftests/landlock/fs_test.c
index a4fdcda62bde..69f9c7409198 100644
--- a/tools/testing/selftests/landlock/fs_test.c
+++ b/tools/testing/selftests/landlock/fs_test.c
@@ -401,7 +401,7 @@ TEST_F_FORK(layout1, inval)
 	LANDLOCK_ACCESS_FS_WRITE_FILE | \
 	LANDLOCK_ACCESS_FS_READ_FILE)
 
-#define ACCESS_LAST LANDLOCK_ACCESS_FS_MAKE_SYM
+#define ACCESS_LAST LANDLOCK_ACCESS_FS_REFER
 
 #define ACCESS_ALL ( \
 	ACCESS_FILE | \
@@ -414,6 +414,7 @@ TEST_F_FORK(layout1, inval)
 	LANDLOCK_ACCESS_FS_MAKE_SOCK | \
 	LANDLOCK_ACCESS_FS_MAKE_FIFO | \
 	LANDLOCK_ACCESS_FS_MAKE_BLOCK | \
+	LANDLOCK_ACCESS_FS_MAKE_SYM | \
 	ACCESS_LAST)
 
 /* clang-format on */
-- 
2.35.1