git: 48caab76e520 - releng/15.2 - rbeneath lookup: do not allow to step up with dotdot lookup over VV_ROOT

From: Mark Johnston <markj_at_FreeBSD.org>
Date: Tue, 17 Nov 2026 17:09:38 UTC
The branch releng/15.2 has been updated by markj:

URL: https://cgit.FreeBSD.org/src/commit/?id=48caab76e52002a88e2fc8a244f1eeda3a9c419f

commit 48caab76e52002a88e2fc8a244f1eeda3a9c419f
Author:     Konstantin Belousov <kib@FreeBSD.org>
AuthorDate: 2026-10-23 03:48:47 +0000
Commit:     Mark Johnston <markj@FreeBSD.org>
CommitDate: 2026-11-17 13:34:51 +0000

    rbeneath lookup: do not allow to step up with dotdot lookup over VV_ROOT
    
    In vfs_lookup(), the ".." loop steps from a mount root to
    its mnt_vnodecovered vnode before the name is looked up.
    nameicap_check_dotdot() runs only after VOP_LOOKUP() and compares the
    directory being left, ni_dvp, with ni_rbeneath_dpp, the start. After
    the mount crossing ni_dvp is the covered vnode, not the start, so the
    comparison never matches and ".." is allowed. [*]
    
    Before 2f60984053 the loop called nameicap_check_dotdot(ndp, dp) right
    after the mnt_vnodecovered step, and the covered vnode, which was not
    among the vnodes the lookup had visited, was refused. The change replaced
    the visited-vnode tracking with the comparison against ni_rbeneath_dpp,
    and removed that call.
    
    A process restricted to a directory descriptor by Capsicum (capability
    mode, or a descriptor limited with cap_rights_limit()), or a program that
    relies on O_RESOLVE_BENEATH / AT_RESOLVE_BENEATH to keep a lookup beneath
    a directory, is not contained when that directory is a mount root (for
    example a tmpfs, nullfs or devfs mount handed to a sandbox). Through ".."
    it can name anything reachable from the directory that holds the mount
    point, within the descriptor's rights and the usual permission checks.
    Sandboxes that are handed a mount root (a per-sandbox tmpfs is a common
    pattern) lose their file-system boundary.
    
    Similarly, if the ni_rbeneath_dpp is reclaimed and then re-instantiated,
    e.g. for nocache nullfs or just by vnlru, the startdir might be bypassed
    because the comparison would not hold.  Check for doomed state to close
    the possibility.
    
    Approved by:    re (cperciva)
    Approved by:    so
    Security:       FreeBSD-SA-26:81.vfs
    Security:       CVE-2026-101312
    Reported by:    pjd [*]
    Fixes:  2f60984053e5 ("namei dotdot tracker: take mnt_renamelock shared to prevent parallel renames")
    Sponsored by:   The FreeBSD Foundation
---
 sys/kern/vfs_lookup.c | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)

diff --git a/sys/kern/vfs_lookup.c b/sys/kern/vfs_lookup.c
index 0a6e85d4f2ee..4ddf7d5433bf 100644
--- a/sys/kern/vfs_lookup.c
+++ b/sys/kern/vfs_lookup.c
@@ -931,6 +931,18 @@ vfs_lookup_isroot(struct nameidata *ndp, struct vnode *dvp)
 	    dvp == rootvnode);
 }
 
+static bool
+vfs_lookup_invalid_boundary(struct nameidata *ndp, struct vnode *dvp)
+{
+	struct vnode *ndpp;
+
+	ndpp = ndp->ni_rbeneath_dpp;
+	/* The doomed check prevents escape through the startdir reclaim. */
+	if (dvp == ndpp || (ndpp != NULL && VN_IS_DOOMED(ndpp)))
+		return (true);
+	return (vfs_lookup_isroot(ndp, dvp));
+}
+
 /*
  * FAILIFEXISTS handling.
  *
@@ -1280,7 +1292,7 @@ dirloop:
 		for (;;) {
 			bool isroot;
 
-			isroot = vfs_lookup_isroot(ndp, dp);
+			isroot = vfs_lookup_invalid_boundary(ndp, dp);
 			if (__predict_false(isroot && (ndp->ni_lcf &
 			    (NI_LCF_STRICTREL | NI_LCF_STRICTREL_KTR)) != 0)) {
 				if ((ndp->ni_lcf & NI_LCF_STRICTREL_KTR) != 0)