From: sashiko-bot@kernel.org
To: "Jarkko Sakkinen" <jarkko@kernel.org>
Cc: linux-mm@kvack.org, linux-security-module@vger.kernel.org,
"Alexander Viro" <viro@zeniv.linux.org.uk>,
"Andrew Morton" <akpm@linux-foundation.org>,
"Andy Lutomirski" <luto@amacapital.net>,
"Ben Segall" <bsegall@google.com>,
"Christian Brauner" <brauner@kernel.org>,
"David Hildenbrand" <david@kernel.org>,
"David Howells" <dhowells@redhat.com>,
"Davidlohr Bueso" <dave@stgolabs.net>,
"Dietmar Eggemann" <dietmar.eggemann@arm.com>,
"Ingo Molnar" <mingo@redhat.com>, "Jan Kara" <jack@suse.cz>,
"Jonathan Corbet" <corbet@lwn.net>,
"John Johansen" <john.johansen@canonical.com>,
"Juri Lelli" <juri.lelli@redhat.com>,
"Kees Cook" <kees@kernel.org>,
"K Prateek Nayak" <kprateek.nayak@amd.com>,
"Liam R. Howlett" <liam@infradead.org>,
"Lorenzo Stoakes" <ljs@kernel.org>,
"Manfred Spraul" <manfred@colorfullife.com>,
"Mel Gorman" <mgorman@suse.de>, "Michal Hocko" <mhocko@suse.com>,
"Mike Rapoport" <rppt@kernel.org>, "Neal Gompa" <neal@gompa.dev>,
"Paul Moore" <paul@paul-moore.com>,
"Pavel Tikhomirov" <ptikhomirov@virtuozzo.com>,
"Peter Zijlstra" <peterz@infradead.org>,
"Randy Dunlap" <rdunlap@infradead.org>,
"Serge Hallyn" <sergeh@kernel.org>,
"Shuah Khan" <skhan@linuxfoundation.org>,
"Steven Rostedt" <rostedt@goodmis.org>,
"Suren Baghdasaryan" <surenb@google.com>,
"Valentin Schneider" <vschneid@redhat.com>,
"Vincent Guittot" <vincent.guittot@linaro.org>,
"Vlastimil Babka" <vbabka@kernel.org>,
"Zygmunt Krynicki" <zygmunt.krynicki@canonical.com>
Subject: Re: [RFC PATCH v2 09/13] fs: Support installing rootns mount trees via move_mount()
Date: Wed, 18 Nov 2026 23:15:21 +0000 [thread overview]
Message-ID: <sashiko-outbox-222163@kernel.org> (raw)
In-Reply-To: <20261117231352.3816603-10-jarkko@kernel.org>
Thank you for your contribution! Sashiko AI review found 5 potential
issues to consider:
- [High] fs: check_for_nsfs_mounts() checks caller mnt_ns instead of
rootns target
- [High] fs: do_move_mount() calls lock_mnt_tree() after
attach_recursive_mnt()
- [High] fs: relative target paths in move_mount_rootns() bypass is_root
checks
- [High] fs: TOCTOU race on rootns->state in move_mount_rootns()
corrupts state
- [High] fs: fsmount() alloc_mnt_ns(fc->user_ns) bypasses submount
MNT_LOCKED
--
commit f0ccb5459fd227ad26c0bf5f42a1937aaa266930
Author: Jarkko Sakkinen <jarkko@kernel.org>
fs: Support installing rootns mount trees via move_mount()
This commit adds MOVE_MOUNT_T_ROOTNS_ROOT to allow attaching detached
anonymous mount trees into a rootns mount namespace via its file descriptor,
and allocates fsmount() anonymous mount namespaces in fc->user_ns.
> diff --git a/fs/namespace.c b/fs/namespace.c
> index 010ed5d68d229..2dfb5e01d40f0 100644
> --- a/fs/namespace.c
> +++ b/fs/namespace.c
[ ... ]
> @@ -3792,7 +3796,13 @@ static int do_move_mount(const struct path *old_path,
> if (IS_ERR(mp.parent))
> return PTR_ERR(mp.parent);
>
> - if (check_mnt(old)) {
> + if (flags & MNT_TREE_ROOTNS) {
> + if (!anon_ns_root(old) || old->mnt_ns == mp.parent->mnt_ns)
> + return -EINVAL;
> + if ((flags & MNT_TREE_ROOTNS_INIT) &&
> + mp.parent != real_mount(new_path->mnt))
> + return -EBUSY;
> + } else if (check_mnt(old)) {
[Severity: High]
Can moving an nsfs mount into a rootns with MNT_TREE_ROOTNS create an
unbreakable mount namespace reference cycle?
Later in do_move_mount(), check_for_nsfs_mounts(old) calls mnt_ns_loop(),
which checks against current->nsproxy->mnt_ns rather than the destination
mp.parent->mnt_ns:
fs/namespace.c:mnt_ns_loop() {
...
mnt_ns = to_mnt_ns(get_proc_ns(d_inode(dentry)));
return current->nsproxy->mnt_ns->ns.ns_id >= mnt_ns->ns.ns_id;
}
Because the target rootns mount namespace was created after the caller's
mount namespace, current->nsproxy->mnt_ns->ns.ns_id is smaller than
rootns_mnt_ns->ns.ns_id. If the caller clones an nsfs mount referring to
rootns_mnt_ns via open_tree(OPEN_TREE_CLONE) and moves it into rootns_mnt_ns
with MOVE_MOUNT_T_ROOTNS_ROOT, mnt_ns_loop() returns false. The mounted
nsfs dentry then pins rootns_mnt_ns while rootns_mnt_ns pins the mount,
permanently leaking the mount namespace and all mounts inside it.
Passing mp.parent->mnt_ns to check_for_nsfs_mounts() and mnt_ns_loop() would
check against the actual destination namespace.
> @@ -3845,7 +3855,16 @@ static int do_move_mount(const struct path *old_path,
> if (mount_is_ancestor(old, mp.parent))
> return -ELOOP;
>
> - return attach_recursive_mnt(old, &mp);
> + lock_tree = (flags & MNT_TREE_ROOTNS) &&
> + old->mnt_ns->user_ns != mp.parent->mnt_ns->user_ns;
> + err = attach_recursive_mnt(old, &mp);
> + if (!err && lock_tree) {
> + scoped_guard(mount_writer) {
> + lock_mnt_tree(old);
> + old->mnt.mnt_flags |= MNT_LOCKED;
> + }
> + }
> + return err;
> }
[Severity: High]
Does calling lock_mnt_tree(old) after attach_recursive_mnt(old, &mp)
mishandle overmounted subtrees, shared mount propagation, and concurrent
readers?
Inside attach_recursive_mnt(), any existing mount q at the target mountpoint
(for example when MOVE_MOUNT_BENEATH or overmounting is used) is reparented
onto topmost_overmount(old) via mnt_change_mountpoint(r, mp, q), and
MNT_LOCKED is cleared on q so it remains unmountable. Running
lock_mnt_tree(old) afterwards walks into q and all of q's descendants via
next_mnt(p, old), setting MNT_LOCKED and MNT_LOCK_* on pre-existing mounts
in the rootns and preventing q from being unmounted.
In addition, if mp.parent is a shared mount in the rootns,
attach_recursive_mnt() clones old for propagation before lock_mnt_tree(old)
runs. Because attach_recursive_mnt() sets user_ns = ns->user_ns (the target
rootns user_ns rather than old->mnt_ns->user_ns), the cross-userns check
child->mnt_parent->mnt_ns->user_ns != user_ns is false for peers in the
target user_ns, leaving propagated copies unlocked. Finally, dropping
mount_lock in unlock_mount_hash() before re-acquiring mount_writer leaves a
window where old is already visible in the rootns without MNT_LOCKED or
MNT_LOCK_* set.
> @@ -4701,7 +4720,7 @@ SYSCALL_DEFINE3(fsmount, int, fs_fd, unsigned int, flags,
> return FD_ADD((flags & FSMOUNT_CLOEXEC) ? O_CLOEXEC : 0,
> open_new_namespace(&new_path, MOUNT_COPY_NEW));
>
> - ns = alloc_mnt_ns(current->nsproxy->mnt_ns->user_ns, true);
> + ns = alloc_mnt_ns(fc->user_ns, true);
> if (IS_ERR(ns))
> return PTR_ERR(ns);
> mnt = real_mount(new_path.mnt);
[Severity: High]
Does allocating the anonymous mount namespace in fc->user_ns allow
submounts from a parent user namespace to be installed into an unprivileged
rootns without MNT_LOCKED?
When a process in a parent user namespace configures a superblock for an
unprivileged rootns via FSCONFIG_SET_ROOTNS, fc->user_ns is set to the
target rootns user_ns, so fsmount() creates a root mount M_root with
M_root->mnt_ns->user_ns equal to the rootns user_ns. If the process then
attaches a host mount M_sub (for example, a read-only bind mount or masking
mount from init_user_ns) onto M_root via move_mount(), attach_recursive_mnt()
calls lock_mnt_tree(M_sub), which only sets MNT_LOCKED on descendants p !=
M_sub and leaves M_sub itself without MNT_LOCKED.
When M_root is subsequently moved into the rootns with
MOVE_MOUNT_T_ROOTNS_ROOT, do_move_mount() sees M_root->mnt_ns->user_ns ==
mp.parent->mnt_ns->user_ns and skips lock_mnt_tree(M_root). As a result,
M_sub is never marked MNT_LOCKED in the rootns, allowing an unprivileged
user in the rootns to unmount M_sub and expose the underlying directory.
Conversely, attaching another FSCONFIG_SET_ROOTNS mount onto M_root via
normal move_mount() causes attach_recursive_mnt() to compare against
current->nsproxy->mnt_ns->user_ns (because M_root->mnt_ns is anonymous) and
spuriously lock the same-userns submount.
> @@ -4741,6 +4760,129 @@ static inline int vfs_move_mount(const struct path *from_path,
[ ... ]
> + rootns_mnt_ns = rootns->ns->mnt_ns;
> + is_root = to_name->name[0] == '/' && to_name->name[1] == '\0';
> + if (is_root) {
> + if (flags & MOVE_MOUNT_BENEATH)
> + return -EINVAL;
> + if (READ_ONCE(rootns->state) != ROOTNS_EMPTY)
> + return -EBUSY;
> +
> + to_path.mnt = mntget(&rootns_mnt_ns->root->mnt);
> + to_path.dentry = dget(rootns_mnt_ns->root->mnt.mnt_root);
> + mflags |= MNT_TREE_ROOTNS_INIT;
> + } else {
> + if (to_name->name[0] == '/')
> + return -EINVAL;
> +
> + scoped_guard(spinlock, &rootns->lock) {
> + if (rootns->state == ROOTNS_EMPTY)
> + return -ENOENT;
> + }
> +
> + lflags = 0;
> + if (flags & MOVE_MOUNT_T_SYMLINKS)
> + lflags |= LOOKUP_FOLLOW;
> + if (flags & MOVE_MOUNT_T_AUTOMOUNTS)
> + lflags |= LOOKUP_AUTOMOUNT;
> + ret = filename_lookup(AT_FDCWD, to_name, lflags, &to_path,
> + &rootns->root);
> + if (ret)
> + return ret;
> + if (real_mount(to_path.mnt)->mnt_ns != rootns_mnt_ns)
> + return -EINVAL;
> + }
[Severity: High]
Can relative target paths such as "." or "subdir/.." resolve to_path to
rootns->root and bypass the is_root checks?
Because is_root only tests for the literal string "/", passing "." after
root installation enters the else branch and resolves to_path to
rootns->root:
1. If MOVE_MOUNT_BENEATH is set with to_pathname = ".", the MOVE_MOUNT_BENEATH
check in the is_root branch is bypassed. do_lock_mount() selects the
underlying nullfs root as mp.parent and mounts the new tree beneath
rootns->root, clearing MNT_LOCKED on rootns->root in
attach_recursive_mnt() so an unprivileged rootns user can unmount a
locked cross-userns root mount.
2. Without MOVE_MOUNT_BENEATH, to_pathname = "." overmounts rootns->root
without updating rootns->root (and similarly on initial "/" installation
if from_path already has overmounts stacked on its root dentry). Because
copy_fs() and preset-root lookups in path_init() do not traverse
overmounts on rootns->root, subsequent CLONE_INTO_ROOTNS processes and
move_mount_rootns() lookups stay in the shadowed lower mount.
> +
> + ret = lookup_move_mount_from(from_dfd, from_pathname, flags,
> + &from_path);
> + if (ret)
> + return ret;
> +
> + ret = vfs_move_mount(&from_path, &to_path, mflags | MNT_TREE_ROOTNS);
> + if (!ret && is_root) {
> + scoped_guard(spinlock, &rootns->lock) {
> + path_get(&from_path);
> + rootns->root = from_path;
> + WRITE_ONCE(rootns->state, ROOTNS_READY);
> + }
> + }
> +
> + return ret;
> +}
[Severity: High]
Can concurrent "/" installations race on rootns->state, leaking the previous
rootns->root and resetting a running or dead rootns back to ROOTNS_READY?
When is_root is true, rootns->state == ROOTNS_EMPTY is only checked before
lookup_move_mount_from() and before namespace_lock() is taken, while
MNT_TREE_ROOTNS_INIT in do_move_mount() only checks whether
rootns_mnt_ns->root currently has an overmount:
CPU0 CPU1
move_mount_rootns(..., "/") move_mount_rootns(..., "/")
READ_ONCE(rootns->state) ==
ROOTNS_EMPTY
READ_ONCE(rootns->state) ==
ROOTNS_EMPTY
lookup_move_mount_from()
vfs_move_mount() -> attaches M1
rootns->root = M1
rootns->state = ROOTNS_READY
...
clone3(CLONE_INTO_ROOTNS)
rootns->state = ROOTNS_RUNNING
umount2("/", MNT_DETACH)
detaches M1 from nullfs
vfs_move_mount() -> attaches M2
spin_lock(&rootns->lock)
rootns->root = M2 <-- leaks M1 path
rootns->state = ROOTNS_READY
<-- resets state from RUNNING
Overwriting rootns->root without path_put() leaks M1's vfsmount and dentry
references and races with lockless readers of rootns->root in
filename_lookup() and copy_fs(). Resetting rootns->state to ROOTNS_READY
also allows a second clone3(CLONE_INTO_ROOTNS) to run the init bootstrap path
in copy_rootns(), overwriting rootns->init_pid without put_pid() and taking
an extra get_rootns() reference that exit_rootns() never drops.
--
Sashiko AI review · https://sashiko.dev/#/patchset/20261117231352.3816603-1-jarkko@kernel.org?part=9
next prev parent reply other threads:[~2026-11-18 23:15 UTC|newest]
Thread overview: 32+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-17 23:13 [RFC PATCH v2 00/13] Namespace for namespace roots Jarkko Sakkinen
2026-11-17 23:13 ` [RFC PATCH v2 01/13] nsproxy: Pass nsproxy to create_new_namespaces() Jarkko Sakkinen
2026-11-17 23:13 ` [RFC PATCH v2 02/13] rootns: Add core infrastructure for root namespaces Jarkko Sakkinen
2026-11-17 23:13 ` [RFC PATCH v2 03/13] rootns: Track init lifecycle via copy_process() and pid_namespace Jarkko Sakkinen
2026-11-18 23:15 ` sashiko-bot
2026-11-17 23:13 ` [RFC PATCH v2 04/13] rootns: Add rootns_create() system call Jarkko Sakkinen
2026-11-18 23:15 ` sashiko-bot
2026-11-17 23:13 ` [RFC PATCH v2 05/13] rootns: Add rootns_wait() " Jarkko Sakkinen
2026-11-18 23:15 ` sashiko-bot
2026-11-17 23:13 ` [RFC PATCH v2 06/13] rootns: Add rootns_kill() " Jarkko Sakkinen
2026-11-18 23:15 ` sashiko-bot
2026-11-17 23:13 ` [RFC PATCH v2 07/13] fork: Support CLONE_INTO_ROOTNS in clone3() Jarkko Sakkinen
2026-11-18 23:15 ` sashiko-bot
2026-11-17 23:13 ` [RFC PATCH v2 08/13] fs: Support FSCONFIG_SET_ROOTNS in fsconfig() Jarkko Sakkinen
2026-11-18 23:15 ` sashiko-bot
2026-11-17 23:13 ` [RFC PATCH v2 09/13] fs: Support installing rootns mount trees via move_mount() Jarkko Sakkinen
2026-11-18 23:15 ` sashiko-bot [this message]
2026-11-17 23:13 ` [RFC PATCH v2 10/13] arch: Wire up rootns system calls Jarkko Sakkinen
2026-11-17 23:13 ` [RFC PATCH v2 11/13] docs: Document root namespace uAPI Jarkko Sakkinen
2026-11-17 23:13 ` [RFC PATCH v2 12/13] selftests/rootns: Add root namespace test suite Jarkko Sakkinen
2026-11-18 23:15 ` sashiko-bot
2026-11-17 23:13 ` [RFC PATCH v2 13/13] MAINTAINERS: Add entry for ROOT NAMESPACE Jarkko Sakkinen
2026-11-18 14:47 ` [RFC PATCH v2 00/13] Namespace for namespace roots Josef Bacik
2026-11-18 15:28 ` Jarkko Sakkinen
2026-11-18 15:39 ` Jarkko Sakkinen
2026-11-18 15:46 ` Jarkko Sakkinen
2026-11-18 16:59 ` Andy Lutomirski
2026-11-18 17:04 ` Jarkko Sakkinen
2026-11-18 17:08 ` Jarkko Sakkinen
2026-11-18 17:30 ` Jarkko Sakkinen
2026-11-18 17:46 ` Andy Lutomirski
2026-11-18 18:23 ` Jarkko Sakkinen
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=sashiko-outbox-222163@kernel.org \
--to=sashiko-bot@kernel.org \
--cc=akpm@linux-foundation.org \
--cc=brauner@kernel.org \
--cc=bsegall@google.com \
--cc=corbet@lwn.net \
--cc=dave@stgolabs.net \
--cc=david@kernel.org \
--cc=dhowells@redhat.com \
--cc=dietmar.eggemann@arm.com \
--cc=jack@suse.cz \
--cc=jarkko@kernel.org \
--cc=john.johansen@canonical.com \
--cc=juri.lelli@redhat.com \
--cc=kees@kernel.org \
--cc=kprateek.nayak@amd.com \
--cc=liam@infradead.org \
--cc=linux-mm@kvack.org \
--cc=linux-security-module@vger.kernel.org \
--cc=ljs@kernel.org \
--cc=luto@amacapital.net \
--cc=manfred@colorfullife.com \
--cc=mgorman@suse.de \
--cc=mhocko@suse.com \
--cc=mingo@redhat.com \
--cc=neal@gompa.dev \
--cc=paul@paul-moore.com \
--cc=peterz@infradead.org \
--cc=ptikhomirov@virtuozzo.com \
--cc=rdunlap@infradead.org \
--cc=rostedt@goodmis.org \
--cc=rppt@kernel.org \
--cc=sashiko-reviews@lists.linux.dev \
--cc=sergeh@kernel.org \
--cc=skhan@linuxfoundation.org \
--cc=surenb@google.com \
--cc=vbabka@kernel.org \
--cc=vincent.guittot@linaro.org \
--cc=viro@zeniv.linux.org.uk \
--cc=vschneid@redhat.com \
--cc=zygmunt.krynicki@canonical.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox