Re: [PATCH -next -v4 02/17] LSM: pass struct path to the inode_setsecctx hook

From: Cai Xinchen

Date: Wed Nov 18 2026 - 04:00:01 EST


Sorry, this is a major issue. I checked the current state of the patchset

- fs/attr.c, fs/open.c, fs/inode.c, fs/utimes.c, fs/namei.c,
fs/coredump.c, drivers/block/loop.c: nd->path or file->f_path (in loop.c:
the f_path of the backing file, where file == NULL is intentional).

- drivers/base/devtmpfs.c: the mount point of the parent directory, which contains a directory entry that was just created or is about to be deleted.

- fs/cachefiles/xattr.c: cache->mnt, under which a directory entry is searched.

- fs/ecryptfs: ecryptfs_lower_path(), which is the mount point of the lower-level directory entry itself.

- fs/nfsd: exp->ex_path.mnt, which contains fh_dentry, a value decoded via
exportfs_decode_fh_raw(exp->ex_path.mnt,...), or exp->ex_path.dentry.

- fs/overlayfs: ovl_upper_mnt(ofs), which contains the actual upper-level directory entry.

There are two places that violate (1), both introduced by this series of patches in ksmbd:

smb2_parent_compressed() and smb_inherit_dacl() pair the directory entry returned by dget_parent()

with path->mnt, and problems arise once that directory entry becomes the root of its mount point.

I will fix these issues in v5 and ensure this invariant.

If you find any place where (1) still does not hold, please point it out, and I
will fix it.

On 11/18/2026 8:16 AM, Al Viro wrote:
On Tue, Nov 17, 2026 at 06:43:42PM +0000, Al Viro wrote:
On Tue, Nov 17, 2026 at 11:05:27PM +0800, Cai Xinchen wrote:
The inode_setsecctx hook and its security_inode_setsecctx() caller in
fs/nfsd operate on a dentry even though both have a struct path
available. Pass a const struct path so that the coming conversion of the
xattr helpers can be reflected in the implementations; for now SELinux
and Smack keep passing path->dentry and &nop_mnt_idmap to
__vfs_setxattr_locked().

Also document in lsm_hook_defs.h what a struct path argument may be used
for: it can be passed to VFS helpers taking a struct path, but nothing is
guaranteed about path->mnt.
No, with the side of Hell, No.

+ * Hooks that take a struct path argument: the VFS only sanctions passing
+ * the path to VFS helpers that take a struct path (e.g. notify_change());
+ * it makes no guarantees with regard to path->mnt, whose layout and
+ * implementation details are subject to change. Implementations must
+ * not rely on its properties.
NAKed-by: Al Viro <viro@xxxxxxxxxxxxxxxxxx>

If that kind of stuff actually does get passed to any existing hooks, we need
to fix that ASAP. If not, consider that a blanket NAK on any patch series
that would introduce such.

We are not going there. At all. "Must not rely on its properties" ==
"must make sure that no existing functions with struct path * arguments
ever have such malformed instances passed to them". Which is not going
to be feasible to enforce. And yes, that very much includes any VFS helpers.
To elaborate a bit: correctness of quite a few places does depend upon the
following invaraint:
path->mnt->mnt_sb == path->dentry->d_sb (1)
for all struct path pointers they might possibly run into. _Some_
functions might _currently_ be safe to call with that constraint violated;
there's neither a list of such functions nor any promise whatsoever that
call that is currently safe in such conditions will remain safe at any
later point.

_IF_ you can guarantee that (1) is not violated, we have something to
talk about; otherwise, don't call your object struct path and don't
pass it to anything that expects to deal with struct path.