statmount: accept fd as a parameter

We would like to add support for checkpoint/restoring file descriptors open on these "unmounted" mounts to CRIU (Checkpoint/Restore in Userspace) [1].

Currently, we have no way to get mount info for these "unmounted" mounts since they do appear in /proc/<pid>/mountinfo and statmount does not work on them, since they do not belong to any mount namespace.

This patch helps us by providing a way to get mountinfo for these "unmounted" mounts by using a fd on the mount.

Changes from v6 [2] to v7:
* Add kselftests for STATMOUNT_BY_FD flag.

* Instead of renaming mnt_id_req.mnt_ns_fd to mnt_id_req.fd introduce a union so struct mnt_id_req looks like this:

    struct mnt_id_req {
            __u32 size;
            union {
                    __u32 mnt_ns_fd;
                    __u32 mnt_fd;
            };
            __u64 mnt_id;
            __u64 param;
            __u64 mnt_ns_id;
    };

* In case of STATMOUNT_BY_FD grab mnt_ns inside of do_statmount(), since we get mnt_ns from mnt, which should happen under namespace lock.

* Remove the modifications made to grab_requested_mnt_ns, those were never needed.

Changes from v5 [3] to v6:
* Instead of returning "[unmounted]" as the mount point for "unmounted" mounts, we unset the STATMOUNT_MNT_POINT flag in statmount.mask.

* Instead of returning 0 as the mnt_ns_id for "unmounted" mounts, we unset the STATMOUNT_MNT_NS_ID flag in statmount.mask.

* Added comment in do_statmount clarifying that the caller sets s->mnt in case of STATMOUNT_BY_FD.

* In do_statmount move the mnt_ns_id and mnt_ns_empty() check just before lookup_mnt_in_ns().

* We took another look at the capability checks for getting information for "unmounted" mounts using an fd and decided to remove them for the following reasons:

- All fs related information is available via fstatfs() without any capability check.

- Mount information is also available via /proc/pid/mountinfo (without any capability check).

- Given that we have access to a fd on the mount which tells us that we had access to the mount at some point (or someone that had access gave us the fd). So, we should be able to access mount info.

Changes from v4 [4] to v5:
Check only for s->root.mnt to be NULL instead of checking for both s->root.mnt and s->root.dentry (I did not find a case where only one of them would be NULL).

* Only allow system root (CAP_SYS_ADMIN in init_user_ns) to call statmount() on fd's on "unmounted" mounts. We (mostly Pavel) spent some time thinking about how our previous approach (of checking the opener's file credentials) caused problems.

Please take a look at the linked pictures they describe everything more clearly.

Case 1: A fd is on a normal mount (Link to Picture: [5]) Consider, a situation where we have two processes P1 and P2 and a file F1. F1 is opened on mount ns M1 by P1. P1 is nested inside user namespace U1 and U2. P2 is also in U1. P2 is also in a pid namespace and mount namespace separate from M1.

P1 sends F1 to P2 (using a unix socket). But, P2 is unable to call statmount() on F1 because since it is a separate pid and mount namespace. This is good and expected.

Case 2: A fd is on a "unmounted" mount (Link to Picture: [6]) Consider a similar situation as Case 1. But now F1 is on a mounted that has been "unmounted". Now, since we used openers credentials to check for permissions P2 ends up having the ability call statmount() and get mount info for this "unmounted" mount.

Hence, It is better to restrict the ability to call statmount() on fds on "unmounted" mounts to system root only (There could also be other cases than the one described above).

Changes from v3 [7] to v4:
* Change the string returned when there is no mountpoint to be "[unmounted]" instead of "[detached]".
* Remove the new DEFINE_FREE put_file and use the one already present in include/linux/file.h (fput) [8].
* Inside listmount consistently pass 0 in flags to copy_mnt_id_req and prepare_klistmount()->grab_requested_mnt_ns() and remove flags from the prepare_klistmount prototype.
* If STATMOUNT_BY_FD is set, check for mnt_ns_id == 0 && mnt_id == 0.

Changes from v2 [9] to v3:
* Rename STATMOUNT_FD flag to STATMOUNT_BY_FD.
* Fixed UAF bug caused by the reference to fd_mount being bound by scope of CLASS(fd_raw, f)(kreq.fd) by using fget_raw instead.
* Reused @spare parameter in mnt_id_req instead of adding new fields to the struct.

Changes from v1 [10] to v2:
v1 of this patchset, took a different approach and introduced a new umount_mnt_ns, to which "unmounted" mounts would be moved to (instead of their namespace being NULL) thus allowing them to be still available via statmount.

Introducing umount_mnt_ns complicated namespace locking and modified performance sensitive code [11] and it was agreed upon that fd-based statmount would be better.

This code is also available on github [12].

[1]: https://github.com/checkpoint-restore/criu/pull/2754
[2]: https://lore.kernel.org/all/[email protected]/
[3]: https://lore.kernel.org/criu/[email protected]/T/#u
[4]: https://lore.kernel.org/all/[email protected]/
[5]: https://github.com/bsach64/linux/blob/statmount-fd-v5/fd_on_normal_mount.png
[6]: https://github.com/bsach64/linux/blob/statmount-fd-v5/file_on_unmounted_mount.png
[7]: https://lore.kernel.org/all/[email protected]/
[8]: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/include/linux/file.h#n97
[9]: https://lore.kernel.org/linux-fsdevel/[email protected]/
[10]: https://lore.kernel.org/linux-fsdevel/[email protected]/
[11]: https://lore.kernel.org/linux-fsdevel/[email protected]/
[12]: https://github.com/bsach64/linux/tree/statmount-fd-v7

Bhavik Sachdev (3):
statmount: permission check should return EPERM
statmount: accept fd as a parameter
selftests: statmount: tests for STATMOUNT_BY_FD

 fs/namespace.c                                | 102 ++++---
 include/uapi/linux/mount.h                    |  10 +-
 .../filesystems/statmount/statmount.h         |  15 +-
 .../filesystems/statmount/statmount_test.c    | 261 +++++++++++++++++-
 .../filesystems/statmount/statmount_test_ns.c | 101 ++++++-
 5 files changed, 430 insertions(+), 59 deletions(-)

2.52.0

Currently, statmount() returns ENOENT when caller is not CAP_SYS_ADMIN in the user namespace owner of target mount namespace. This should be EPERM instead.

---
 fs/namespace.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

index 2bad25709b2c..ee36d67f1ac2 100644
@@ -5795,7 +5795,7 @@ SYSCALL_DEFINE4(statmount, const struct mnt_id_req __user *, req,
 
 	if (kreq.mnt_ns_id && (ns != current->nsproxy->mnt_ns) &&
 	    !ns_capable_noaudit(ns->user_ns, CAP_SYS_ADMIN))
-		return -ENOENT;
+		return -EPERM;
 
 	ks = kmalloc(sizeof(*ks), GFP_KERNEL_ACCOUNT);
 	if (!ks)

2.52.0

Extend struct mnt_id_req to take in a fd and introduce STATMOUNT_BY_FD flag. When a valid fd is provided and STATMOUNT_BY_FD is set, statmount will return mountinfo about the mount the fd is on.

This even works for "unmounted" mounts (mounts that have been umounted using umount2(mnt, MNT_DETACH)), if you have access to a file descriptor on that mount. These "umounted" mounts will have no mountpoint and no valid mount namespace. Hence, we unset the STATMOUNT_MNT_POINT and STATMOUNT_MNT_NS_ID in statmount.mask for "unmounted" mounts.

In case of STATMOUNT_BY_FD, given that we already have access to an fd on the mount, accessing mount information without a capability check seems fine because of the following reasons:

- All fs related information is available via fstatfs() without any capability check.
- Mount information is also available via /proc/pid/mountinfo (without any capability check).
- Given that we have access to a fd on the mount which tells us that we had access to the mount at some point (or someone that had access gave us the fd). So, we should be able to access mount info.

---
 fs/namespace.c             | 102 ++++++++++++++++++++++++-------------
 include/uapi/linux/mount.h |  10 +++-
 2 files changed, 76 insertions(+), 36 deletions(-)

index ee36d67f1ac2..73ffa1fbdad7 100644
@@ -5563,31 +5563,49 @@ static int grab_requested_root(struct mnt_namespace *ns, struct path *root)
 
 /* locks: namespace_shared */
 static int do_statmount(struct kstatmount *s, u64 mnt_id, u64 mnt_ns_id,
-			struct mnt_namespace *ns)
+                        struct file *mnt_file, struct mnt_namespace *ns)
 {
-	struct mount *m;
 	int err;
 
-	/* Has the namespace already been emptied? */
-	if (mnt_ns_id && mnt_ns_empty(ns))
-		return -ENOENT;
+	if (mnt_file) {
+		WARN_ON_ONCE(ns != NULL);
 
-	s->mnt = lookup_mnt_in_ns(mnt_id, ns);
-	if (!s->mnt)
-		return -ENOENT;
+		s->mnt = mnt_file->f_path.mnt;
+		ns = real_mount(s->mnt)->mnt_ns;
+		if (!ns)
+			/*
+			 * We can't set mount point and mnt_ns_id since we don't have a
+			 * ns for the mount. This can happen if the mount is unmounted
+			 * with MNT_DETACH.
+			 */
+			s->mask &= ~(STATMOUNT_MNT_POINT | STATMOUNT_MNT_NS_ID);
+	} else {
+		/* Has the namespace already been emptied? */
+		if (mnt_ns_id && mnt_ns_empty(ns))
+			return -ENOENT;
 
-	err = grab_requested_root(ns, &s->root);
-	if (err)
-		return err;
+		s->mnt = lookup_mnt_in_ns(mnt_id, ns);
+		if (!s->mnt)
+			return -ENOENT;
+	}
 
-	/*
-	 * Don't trigger audit denials. We just want to determine what
-	 * mounts to show users.
-	 */
-	m = real_mount(s->mnt);
-	if (!is_path_reachable(m, m->mnt.mnt_root, &s->root) &&
-	    !ns_capable_noaudit(ns->user_ns, CAP_SYS_ADMIN))
-		return -EPERM;
+	if (ns) {
+		err = grab_requested_root(ns, &s->root);
+		if (err)
+			return err;
+
+		if (!mnt_file) {
+			struct mount *m;
+			/*
+			 * Don't trigger audit denials. We just want to determine what
+			 * mounts to show users.
+			 */
+			m = real_mount(s->mnt);
+			if (!is_path_reachable(m, m->mnt.mnt_root, &s->root) &&
+			    !ns_capable_noaudit(ns->user_ns, CAP_SYS_ADMIN))
+				return -EPERM;
+		}
+	}
 
 	err = security_sb_statfs(s->mnt->mnt_root);
 	if (err)

@@ -5709,7 +5727,7 @@ static int prepare_kstatmount(struct kstatmount *ks, struct mnt_id_req *kreq,
 }
 
 static int copy_mnt_id_req(const struct mnt_id_req __user *req,
-			   struct mnt_id_req *kreq)
+			   struct mnt_id_req *kreq, unsigned int flags)
 {
 	int ret;
 	size_t usize;

@@ -5727,11 +5745,17 @@ static int copy_mnt_id_req(const struct mnt_id_req __user *req,
 	ret = copy_struct_from_user(kreq, sizeof(*kreq), req, usize);
 	if (ret)
 		return ret;
-	if (kreq->mnt_ns_fd != 0 && kreq->mnt_ns_id)
-		return -EINVAL;
-	/* The first valid unique mount id is MNT_UNIQUE_ID_OFFSET + 1. */
-	if (kreq->mnt_id <= MNT_UNIQUE_ID_OFFSET)
-		return -EINVAL;
+
+	if (flags & STATMOUNT_BY_FD) {
+		if (kreq->mnt_id || kreq->mnt_ns_id)
+			return -EINVAL;
+	} else {
+		if (kreq->mnt_ns_fd != 0 && kreq->mnt_ns_id)
+			return -EINVAL;
+		/* The first valid unique mount id is MNT_UNIQUE_ID_OFFSET + 1. */
+		if (kreq->mnt_id <= MNT_UNIQUE_ID_OFFSET)
+			return -EINVAL;
+	}
 	return 0;
 }
 

@@ -5777,25 +5801,33 @@ SYSCALL_DEFINE4(statmount, const struct mnt_id_req __user *, req,
 {
 	struct mnt_namespace *ns __free(mnt_ns_release) = NULL;
 	struct kstatmount *ks __free(kfree) = NULL;
+	struct file *mnt_file __free(fput) = NULL;
 	struct mnt_id_req kreq;
 	/* We currently support retrieval of 3 strings. */
 	size_t seq_size = 3 * PATH_MAX;
 	int ret;
 
-	if (flags)
+	if (flags & ~STATMOUNT_BY_FD)
 		return -EINVAL;
 
-	ret = copy_mnt_id_req(req, &kreq);
+	ret = copy_mnt_id_req(req, &kreq, flags);
 	if (ret)
 		return ret;
 
-	ns = grab_requested_mnt_ns(&kreq);
-	if (IS_ERR(ns))
-		return PTR_ERR(ns);
+	if (flags & STATMOUNT_BY_FD) {
+		mnt_file = fget_raw(kreq.mnt_fd);
+		if (!mnt_file)
+			return -EBADF;
+		/* do_statmount sets ns in case of STATMOUNT_BY_FD */
+	} else {
+		ns = grab_requested_mnt_ns(&kreq);
+		if (IS_ERR(ns))
+			return PTR_ERR(ns);
 
-	if (kreq.mnt_ns_id && (ns != current->nsproxy->mnt_ns) &&
-	    !ns_capable_noaudit(ns->user_ns, CAP_SYS_ADMIN))
-		return -EPERM;
+		if (kreq.mnt_ns_id && (ns != current->nsproxy->mnt_ns) &&
+		    !ns_capable_noaudit(ns->user_ns, CAP_SYS_ADMIN))
+			return -EPERM;
+	}
 
 	ks = kmalloc(sizeof(*ks), GFP_KERNEL_ACCOUNT);
 	if (!ks)

@@ -5807,7 +5839,7 @@ SYSCALL_DEFINE4(statmount, const struct mnt_id_req __user *, req,
 		return ret;
 
 	scoped_guard(namespace_shared)
-		ret = do_statmount(ks, kreq.mnt_id, kreq.mnt_ns_id, ns);
+		ret = do_statmount(ks, kreq.mnt_id, kreq.mnt_ns_id, mnt_file, ns);
 
 	if (!ret)
 		ret = copy_statmount_to_user(ks);

@@ -5947,7 +5979,7 @@ SYSCALL_DEFINE4(listmount, const struct mnt_id_req __user *, req,
 	if (!access_ok(mnt_ids, nr_mnt_ids * sizeof(*mnt_ids)))
 		return -EFAULT;
 
-	ret = copy_mnt_id_req(req, &kreq);
+	ret = copy_mnt_id_req(req, &kreq, 0);
 	if (ret)
 		return ret;
 

index 5d3f8c9e3a62..18c624405268 100644
@@ -197,7 +197,10 @@ struct statmount {
  */
 struct mnt_id_req {
 	__u32 size;
-	__u32 mnt_ns_fd;
+	union {
+		__u32 mnt_ns_fd;
+		__u32 mnt_fd;
+	};
 	__u64 mnt_id;
 	__u64 param;
 	__u64 mnt_ns_id;

@@ -232,4 +235,9 @@ struct mnt_id_req {
 #define LSMT_ROOT		0xffffffffffffffff	/* root mount */
 #define LISTMOUNT_REVERSE	(1 << 0) /* List later mounts first */
 
+/*
+ * @flag bits for statmount(2)
+ */
+#define STATMOUNT_BY_FD		0x00000001U	/* want mountinfo for given fd */
+
 #endif /* _UAPI_LINUX_MOUNT_H */

2.52.0

Add tests for STATMOUNT_BY_FD flag, which adds support for passing a file descriptors to statmount(). The fd can also be on a "unmounted" mount (mount unmounted with MNT_DETACH), we also include tests for that.

---
 .../filesystems/statmount/statmount.h         |  15 +-
 .../filesystems/statmount/statmount_test.c    | 261 +++++++++++++++++-
 .../filesystems/statmount/statmount_test_ns.c | 101 ++++++-
 3 files changed, 354 insertions(+), 23 deletions(-)

index 99e5ad082fb1..e1cba4bfd8d9 100644
@@ -43,19 +43,24 @@
 	#endif
 #endif
 
-static inline int statmount(uint64_t mnt_id, uint64_t mnt_ns_id, uint64_t mask,
-			    struct statmount *buf, size_t bufsize,
+static inline int statmount(uint64_t mnt_id, uint64_t mnt_ns_id, uint32_t fd,
+			    uint64_t mask, struct statmount *buf, size_t bufsize,
 			    unsigned int flags)
 {
 	struct mnt_id_req req = {
 		.size = MNT_ID_REQ_SIZE_VER0,
-		.mnt_id = mnt_id,
 		.param = mask,
 	};
 
-	if (mnt_ns_id) {
+	if (flags & STATMOUNT_BY_FD) {
 		req.size = MNT_ID_REQ_SIZE_VER1;
-		req.mnt_ns_id = mnt_ns_id;
+		req.mnt_fd = fd;
+	} else {
+		req.mnt_id = mnt_id;
+		if (mnt_ns_id) {
+			req.size = MNT_ID_REQ_SIZE_VER1;
+			req.mnt_ns_id = mnt_ns_id;
+		}
 	}
 
 	return syscall(__NR_statmount, &req, buf, bufsize, flags);

2.52.0

This seems good to me now. Maybe Miklos wants to give this a nod as well.

---

Applied to the vfs-6.20.namespace branch of the vfs/vfs.git tree. Patches in the vfs-6.20.namespace branch should appear in linux-next soon.

Please report any outstanding bugs that were missed during review in a new review to the original patch series allowing us to drop it.

It's encouraged to provide Acked-bys and Reviewed-bys even though the patch has now been applied. If possible patch trailers will be updated.

Note that commit hashes shown below are subject to change due to rebase, trailer updates or similar. If in doubt, please check the listed branch.

tree:   https://git.kernel.org/pub/scm/linux/kernel/git/vfs/vfs.git
branch: vfs-6.20.namespace

[1/3] statmount: permission check should return EPERM

      https://git.kernel.org/vfs/vfs/c/5d68cefb00d9
[2/3] statmount: accept fd as a parameter
      https://git.kernel.org/vfs/vfs/c/c5255171323f
[3/3] selftests: statmount: tests for STATMOUNT_BY_FD
      https://git.kernel.org/vfs/vfs/c/15d50368501a

This looks useful.