diff mbox series

[v9,01/10] namei: obey trailing magic-link DAC permissions

Message ID 20190706145737.5299-2-cyphar@cyphar.com (mailing list archive)
State New, archived
Headers show
Series namei: openat2(2) path resolution restrictions | expand

Commit Message

Aleksa Sarai July 6, 2019, 2:57 p.m. UTC
The ability for userspace to "re-open" file descriptors through
/proc/self/fd has been a very useful tool for all sorts of usecases
(container runtimes are one common example). However, the current
interface for doing this has resulted in some pretty subtle security
holes. Userspace can re-open a file descriptor with more permissions
than the original, which can result in cases such as /proc/$pid/exe
being re-opened O_RDWR at a later date even though (by definition)
/proc/$pid/exe cannot be opened for writing. When combined with O_PATH
the results can get even more confusing.

We cannot block this outright. Aside from userspace already depending on
it, it's a useful feature which can actually increase the security of
userspace. For instance, LXC keeps an O_PATH of the container's
/dev/pts/ptmx that gets re-opened to create new ptys and then uses
TIOCGPTPEER to get the slave end. This allows for pty allocation without
resolving paths inside an (untrusted) container's rootfs. There isn't a
trivial way of doing this that is as straight-forward and safe as O_PATH
re-opening.

Instead we have to restrict it in such a way that it doesn't break
(good) users but does block potential attackers. The solution applied in
this patch is to restrict *re-opening* (not resolution through)
magic-links by requiring that mode of the link be obeyed. Normal
symlinks have modes of a+rwx but magic-links have other modes. These
magic-link modes were historically ignored during path resolution, but
they've now been re-purposed for more useful ends.

It is also necessary to define semantics for the mode of an O_PATH
descriptor, since re-opening a magic-link through an O_PATH needs to be
just as restricted as the corresponding magic-link otherwise the above
protection can be bypassed. There are two distinct cases:

 1. The target is a regular file (not a magic-link). Userspace depends
    on being able to re-open the O_PATH of a regular file, so we must
    define the mode to be a+rwx.

 2. The target is a magic-link. In this case, we simply copy the mode of
    the magic-link. This results in an O_PATH of a magic-link
    effectively acting as a no-op in terms of how much re-opening
    privileges a process has.

CAP_DAC_OVERRIDE can be used to override all of these restrictions, but
we only permit &init_userns's capabilities to affect these semantics.
The reason for this is that there isn't a clear way to track what
user_ns is the original owner of a given O_PATH chain -- thus an
unprivileged user could create a new userns and O_PATH the file
descriptor, owning it. All signs would indicate that the user really
does have CAP_DAC_OVERRIDE over the new descriptor and the protection
would be bypassed. We thus opt for the more conservative approach.

I have run this patch on several machines for several days. So far, the
only processes which have hit this case ("loadkeys" and "kbd_mode" from
the kbd package[1]) gracefully handle the permission error and do not
cause any user-visible problems. In order to give users a heads-up, a
warning is given whenever may_open_magiclink() refuses access.

[1]: http://git.altlinux.org/people/legion/packages/kbd.git

Co-developed-by: Andy Lutomirski <luto@kernel.org>
Co-developed-by: Christian Brauner <christian@brauner.io>
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
---
 fs/internal.h      |   1 +
 fs/namei.c         | 103 +++++++++++++++++++++++++++++++++++++++++++--
 fs/open.c          |   3 +-
 fs/proc/fd.c       |  23 +++++++++-
 include/linux/fs.h |   4 ++
 5 files changed, 127 insertions(+), 7 deletions(-)

Comments

Al Viro July 12, 2019, 4:14 a.m. UTC | #1
On Sun, Jul 07, 2019 at 12:57:28AM +1000, Aleksa Sarai wrote:

> @@ -514,7 +516,14 @@ static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
>  	p->stack = p->internal;
>  	p->dfd = dfd;
>  	p->name = name;
> -	p->total_link_count = old ? old->total_link_count : 0;
> +	p->total_link_count = 0;
> +	p->acc_mode = 0;
> +	p->opath_mask = FMODE_PATH_READ | FMODE_PATH_WRITE;
> +	if (old) {
> +		p->total_link_count = old->total_link_count;
> +		p->acc_mode = old->acc_mode;
> +		p->opath_mask = old->opath_mask;
> +	}

Huh?  Could somebody explain why traversals of NFS4 referrals should inherit
->acc_mode and ->opath_mask?

>  static __always_inline
> -const char *get_link(struct nameidata *nd)
> +const char *get_link(struct nameidata *nd, bool trailing)
>  {
>  	struct saved *last = nd->stack + nd->depth - 1;
>  	struct dentry *dentry = last->link.dentry;
> @@ -1081,6 +1134,44 @@ const char *get_link(struct nameidata *nd)
>  		} else {
>  			res = get(dentry, inode, &last->done);
>  		}
> +		/* If we just jumped it was because of a magic-link. */
> +		if (unlikely(nd->flags & LOOKUP_JUMPED)) {

That's not quite guaranteed (it is possible to bind a symlink on top
of a regular file, and you will get LOOKUP_JUMPED on the entry into
trailing_symlink() when looking the result up).  Moreover, why bother
with LOOKUP_JUMPED here?  See that
	nd->last_type = LAST_BIND;
several lines prior?  That's precisely to be able to recognize those
suckers.

And _that_ would've avoided another piece of ugliness - your LOOKUP_JUMPED
kludge forces you to handle that cra^Wsclero^Wvaluable security hardening
in get_link(), instead of trailing_symlink() where you apparently want
it to be.  Simply because nd_jump_root() done later in get_link() will set
LOOKUP_JUMPED for absolute symlinks, confusing your test.

Moreover, I'm not sure that trailing_symlink() is the right place for
that either - I would be rather tempted to fold do_o_path() into
path_openat(), inline path_lookupat() there (as in
        s = path_init(nd, flags);

        while (!(error = link_path_walk(s, nd))
                && ((error = lookup_last(nd)) > 0)) {
                s = trailing_symlink(nd);
        }
        if (!error)
                error = complete_walk(nd);
        if (!error && nd->flags & LOOKUP_DIRECTORY)
                if (!d_can_lookup(nd->path.dentry))
                        error = -ENOTDIR;
        if (!error) {
                audit_inode(nd->name, nd->path.dentry, 0);
                error = vfs_open(&nd->path, file);
        }
        terminate_walk(nd);
- we don't need LOOKUP_DOWN there) and then we only care about the
two callers of trailing_symlink() that are in path_openat().  Which
is where you have your ->acc_mode and ->opath_mask without the need
to dump them into nameidata.  Or to bring that mess into the
things like stat(2) et.al. - it simply doesn't belong there.

In any case, this "bool trailing" is completely wrong; whether that
check belongs in trailing_symlink() or (some of) its callers, putting
it into get_link() is a mistake, forced by kludgy check for procfs-style
symlinks.
Al Viro July 12, 2019, 6:36 a.m. UTC | #2
On Fri, Jul 12, 2019 at 05:14:54AM +0100, Al Viro wrote:

> That's not quite guaranteed (it is possible to bind a symlink on top
> of a regular file, and you will get LOOKUP_JUMPED on the entry into
> trailing_symlink() when looking the result up).  Moreover, why bother
> with LOOKUP_JUMPED here?  See that
> 	nd->last_type = LAST_BIND;
> several lines prior?  That's precisely to be able to recognize those
> suckers.

... except that this won't work these days (once upon a time it used
to, but that had been a long time ago).  However, that doesn't change
the fact that the test is really wrong.  So let's do it right:

* set a new flag along with LOOKUP_JUMPED in nd_jump_link()
* clear it in get_link() right before
	res = READ_ONCE(inode->i_link);
* check it in trailing_symlink() (or callers thereof)

The rest of comments stand, AFAICS.
Aleksa Sarai July 12, 2019, 12:20 p.m. UTC | #3
On 2019-07-12, Al Viro <viro@zeniv.linux.org.uk> wrote:
> On Sun, Jul 07, 2019 at 12:57:28AM +1000, Aleksa Sarai wrote:
> > @@ -514,7 +516,14 @@ static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
> >  	p->stack = p->internal;
> >  	p->dfd = dfd;
> >  	p->name = name;
> > -	p->total_link_count = old ? old->total_link_count : 0;
> > +	p->total_link_count = 0;
> > +	p->acc_mode = 0;
> > +	p->opath_mask = FMODE_PATH_READ | FMODE_PATH_WRITE;
> > +	if (old) {
> > +		p->total_link_count = old->total_link_count;
> > +		p->acc_mode = old->acc_mode;
> > +		p->opath_mask = old->opath_mask;
> > +	}
> 
> Huh?  Could somebody explain why traversals of NFS4 referrals should inherit
> ->acc_mode and ->opath_mask?

I'll be honest -- I don't understand what set_nameidata() did so I just
did what I thought would be an obvious change (to just copy the
contents). I thought it was related to some aspect of the symlink stack
handling.

In that case, should they both be set to 0 on set_nameidata()? This will
mean that fd re-opening (or magic-link opening) through a
set_nameidata() would always fail.

> >  static __always_inline
> > -const char *get_link(struct nameidata *nd)
> > +const char *get_link(struct nameidata *nd, bool trailing)
> >  {
> >  	struct saved *last = nd->stack + nd->depth - 1;
> >  	struct dentry *dentry = last->link.dentry;
> > @@ -1081,6 +1134,44 @@ const char *get_link(struct nameidata *nd)
> >  		} else {
> >  			res = get(dentry, inode, &last->done);
> >  		}
> > +		/* If we just jumped it was because of a magic-link. */
> > +		if (unlikely(nd->flags & LOOKUP_JUMPED)) {
> [...]
> In any case, this "bool trailing" is completely wrong; whether that
> check belongs in trailing_symlink() or (some of) its callers, putting
> it into get_link() is a mistake, forced by kludgy check for procfs-style
> symlinks.

The error path for LOOKUP_JUMPED comes from the old O_BENEATH patchset,
but all of the "bool trailing" logic is definitely my gaff (I was
quietly hoping you'd have a much better solution than the whole
get_link() thing -- it definitely felt very kludgey to write).

I will work on the suggestion in your follow-up email. Thanks!
Al Viro July 12, 2019, 1:10 p.m. UTC | #4
On Fri, Jul 12, 2019 at 10:20:17PM +1000, Aleksa Sarai wrote:
> On 2019-07-12, Al Viro <viro@zeniv.linux.org.uk> wrote:
> > On Sun, Jul 07, 2019 at 12:57:28AM +1000, Aleksa Sarai wrote:
> > > @@ -514,7 +516,14 @@ static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
> > >  	p->stack = p->internal;
> > >  	p->dfd = dfd;
> > >  	p->name = name;
> > > -	p->total_link_count = old ? old->total_link_count : 0;
> > > +	p->total_link_count = 0;
> > > +	p->acc_mode = 0;
> > > +	p->opath_mask = FMODE_PATH_READ | FMODE_PATH_WRITE;
> > > +	if (old) {
> > > +		p->total_link_count = old->total_link_count;
> > > +		p->acc_mode = old->acc_mode;
> > > +		p->opath_mask = old->opath_mask;
> > > +	}
> > 
> > Huh?  Could somebody explain why traversals of NFS4 referrals should inherit
> > ->acc_mode and ->opath_mask?
> 
> I'll be honest -- I don't understand what set_nameidata() did so I just
> did what I thought would be an obvious change (to just copy the
> contents). I thought it was related to some aspect of the symlink stack
> handling.

No.  It's handling of (very rare) nested pathwalk.  The only case I can think
of is handling of NFS4 referrals - they are triggered by ->d_automount()
and include NFS4 mount.  Which does internal pathwalk of its own, to get
to the root of subtree being automounted.

NFS has its own recursion protection on that path (no deeper nesting than
one level of referral traversals), but there some nesting is inevitable;
we do get another nameidata instance on stack.  And for nd_jump_link() we
need to keep track of the innermost one.

For symlinks nothing of that sort happens - they are dealt with on the same
struct nameidata.  ->total_link_count copying is there for one reason only -
we want the total amount of symlinks traversed during the pathwalk (including
the referral processing, etc.) to count towards MAXSYMLINKS check.  It could've
been moved from nameidata to task_struct, but it's cheaper to handle it that
way.

Again, nesting is *rare*.

> In that case, should they both be set to 0 on set_nameidata()? This will
> mean that fd re-opening (or magic-link opening) through a
> set_nameidata() would always fail.

Huh?  set_nameidata() is done for *all* instances - it's pretty much the
constructor of that object (and restore_nameidata() - a destructor).
Everything goes through it.

And again, I'm not sure we want these fields in nameidata - IMO they belong
in open_flags.  Things like e.g. stat() don't need them at all.

Incidentally, O_PATH opening of symlinks combined with subsequent procfs
symlink traversals is worth testing - that's where the things get subtle
and that's where it's easy to get in trouble on modifications.
Aleksa Sarai July 14, 2019, 7:11 a.m. UTC | #5
On 2019-07-12, Al Viro <viro@zeniv.linux.org.uk> wrote:
> On Fri, Jul 12, 2019 at 10:20:17PM +1000, Aleksa Sarai wrote:
> > On 2019-07-12, Al Viro <viro@zeniv.linux.org.uk> wrote:
> > > On Sun, Jul 07, 2019 at 12:57:28AM +1000, Aleksa Sarai wrote:
> > > > @@ -514,7 +516,14 @@ static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
> > > >  	p->stack = p->internal;
> > > >  	p->dfd = dfd;
> > > >  	p->name = name;
> > > > -	p->total_link_count = old ? old->total_link_count : 0;
> > > > +	p->total_link_count = 0;
> > > > +	p->acc_mode = 0;
> > > > +	p->opath_mask = FMODE_PATH_READ | FMODE_PATH_WRITE;
> > > > +	if (old) {
> > > > +		p->total_link_count = old->total_link_count;
> > > > +		p->acc_mode = old->acc_mode;
> > > > +		p->opath_mask = old->opath_mask;
> > > > +	}
> > > 
> > > Huh?  Could somebody explain why traversals of NFS4 referrals should inherit
> > > ->acc_mode and ->opath_mask?
> > 
> > I'll be honest -- I don't understand what set_nameidata() did so I just
> > did what I thought would be an obvious change (to just copy the
> > contents). I thought it was related to some aspect of the symlink stack
> > handling.
> 
> No.  It's handling of (very rare) nested pathwalk.  The only case I can think
> of is handling of NFS4 referrals - they are triggered by ->d_automount()
> and include NFS4 mount.  Which does internal pathwalk of its own, to get
> to the root of subtree being automounted.
> 
> NFS has its own recursion protection on that path (no deeper nesting than
> one level of referral traversals), but there some nesting is inevitable;
> we do get another nameidata instance on stack.  And for nd_jump_link() we
> need to keep track of the innermost one.
> 
> For symlinks nothing of that sort happens - they are dealt with on the same
> struct nameidata.  ->total_link_count copying is there for one reason only -
> we want the total amount of symlinks traversed during the pathwalk (including
> the referral processing, etc.) to count towards MAXSYMLINKS check.  It could've
> been moved from nameidata to task_struct, but it's cheaper to handle it that
> way.
> 
> Again, nesting is *rare*.

Thanks for the explanation, much appreciated. I will drop the old->...
copying hunk.

> > In that case, should they both be set to 0 on set_nameidata()? This will
> > mean that fd re-opening (or magic-link opening) through a
> > set_nameidata() would always fail.
> 
> Huh?  set_nameidata() is done for *all* instances - it's pretty much the
> constructor of that object (and restore_nameidata() - a destructor).
> Everything goes through it.

Sorry, I meant to drop the copy-from-old logic -- not set it to zero
explicitly in set_nameidata().

> And again, I'm not sure we want these fields in nameidata - IMO they belong
> in open_flags.  Things like e.g. stat() don't need them at all.

Yup, I'll work up a version that does the consolidation you mentioned
in your other mail.

> Incidentally, O_PATH opening of symlinks combined with subsequent procfs
> symlink traversals is worth testing - that's where the things get subtle
> and that's where it's easy to get in trouble on modifications.

I have some self-tests of a symlink-to-a-magic-link in the last patch of
the series. Did you mean something even more chained like a symlink to a
/proc/self/fd/$n of an O_NOFOLLOW|O_PATH of a symlink?
diff mbox series

Patch

diff --git a/fs/internal.h b/fs/internal.h
index a48ef81be37d..12847f502f49 100644
--- a/fs/internal.h
+++ b/fs/internal.h
@@ -119,6 +119,7 @@  struct open_flags {
 	int acc_mode;
 	int intent;
 	int lookup_flags;
+	fmode_t opath_mask;
 };
 extern struct file *do_filp_open(int dfd, struct filename *pathname,
 		const struct open_flags *op);
diff --git a/fs/namei.c b/fs/namei.c
index 20831c2fbb34..4ec6168762db 100644
--- a/fs/namei.c
+++ b/fs/namei.c
@@ -506,6 +506,8 @@  struct nameidata {
 	struct inode	*link_inode;
 	unsigned	root_seq;
 	int		dfd;
+	fmode_t 	opath_mask;
+	int		acc_mode; /* op.acc_mode */
 } __randomize_layout;
 
 static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
@@ -514,7 +516,14 @@  static void set_nameidata(struct nameidata *p, int dfd, struct filename *name)
 	p->stack = p->internal;
 	p->dfd = dfd;
 	p->name = name;
-	p->total_link_count = old ? old->total_link_count : 0;
+	p->total_link_count = 0;
+	p->acc_mode = 0;
+	p->opath_mask = FMODE_PATH_READ | FMODE_PATH_WRITE;
+	if (old) {
+		p->total_link_count = old->total_link_count;
+		p->acc_mode = old->acc_mode;
+		p->opath_mask = old->opath_mask;
+	}
 	p->saved = old;
 	current->nameidata = p;
 }
@@ -1042,8 +1051,52 @@  static int may_create_in_sticky(struct dentry * const dir,
 	return 0;
 }
 
+/**
+ * may_reopen_magiclink - Check permissions for opening a trailing magic-link
+ * @opath_mask: the O_PATH mask of the magic-link
+ * @acc_mode: ACC_MODE which the user is attempting
+ *
+ * We block magic-link re-opening if the @opath_mask is more strict than the
+ * @acc_mode being requested, unless the user is capable(CAP_DAC_OVERRIDE).
+ *
+ * Returns 0 if successful, -ve on error.
+ */
+static int may_open_magiclink(fmode_t opath_mask, int acc_mode)
+{
+	/*
+	 * We only allow for init_userns to be able to override magic-links.
+	 * This is done to avoid cases where an unprivileged userns could take
+	 * an O_PATH of the fd, resulting in it being very unclear whether
+	 * CAP_DAC_OVERRIDE should work on the new O_PATH fd (given that it
+	 * pipes through to the underlying file).
+	 */
+	if (capable(CAP_DAC_OVERRIDE))
+		return 0;
+
+	if ((acc_mode & MAY_READ) &&
+	    !(opath_mask & (FMODE_READ | FMODE_PATH_READ)))
+		goto err;
+	if ((acc_mode & MAY_WRITE) &&
+	    !(opath_mask & (FMODE_WRITE | FMODE_PATH_WRITE)))
+		goto err;
+
+	return 0;
+
+err:
+	pr_warn_ratelimited("%s[%d]: magic-link re-open blocked (acc_mode=%s%s%s, opath_mask=%s%s%s%s)",
+		current->comm, task_pid_nr(current),
+		(acc_mode & MAY_READ) ? "r": "",
+		(acc_mode & MAY_WRITE) ? "w": "",
+		(acc_mode & MAY_EXEC) ? "x": "",
+		(opath_mask & FMODE_READ) ? "R" : "",
+		(opath_mask & FMODE_PATH_READ) ? "r" : "",
+		(opath_mask & FMODE_WRITE) ? "W" : "",
+		(opath_mask & FMODE_PATH_WRITE) ? "w" : "");
+	return -EACCES;
+}
+
 static __always_inline
-const char *get_link(struct nameidata *nd)
+const char *get_link(struct nameidata *nd, bool trailing)
 {
 	struct saved *last = nd->stack + nd->depth - 1;
 	struct dentry *dentry = last->link.dentry;
@@ -1081,6 +1134,44 @@  const char *get_link(struct nameidata *nd)
 		} else {
 			res = get(dentry, inode, &last->done);
 		}
+		/* If we just jumped it was because of a magic-link. */
+		if (unlikely(nd->flags & LOOKUP_JUMPED)) {
+			/*
+			 * For trailing_symlink we check whether the symlink's
+			 * mode allows us to do what we want through acc_mode.
+			 * In addition, we need to stash away what the link
+			 * mode is in case we are about to O_PATH this
+			 * magic-link.
+			 *
+			 * This is only done for magic-links, as a security
+			 * measure to prevent users from being able to re-open
+			 * files with additional permissions or similar tricks
+			 * through procfs. This is not strictly POSIX-friendly,
+			 * but technically neither are magic-links.
+			 */
+			if (trailing) {
+				fmode_t opath_mask = 0;
+
+				/*
+				 * Figure out the O_PATH mask. Rather than
+				 * using acl_permission_check, we check whether
+				 * any of the rw bits are set in the mode.
+				 */
+				if (inode->i_mode & S_IRUGO)
+					opath_mask |= FMODE_PATH_READ;
+				if (inode->i_mode & S_IWUGO)
+					opath_mask |= FMODE_PATH_WRITE;
+
+				/*
+				 * Is the new opath_mask more restrictive than
+				 * the acc_mode being requested?
+				 */
+				error = may_open_magiclink(opath_mask, nd->acc_mode);
+				if (error)
+					return ERR_PTR(error);
+				nd->opath_mask &= opath_mask;
+			}
+		}
 		if (IS_ERR_OR_NULL(res))
 			return res;
 	}
@@ -2142,7 +2233,7 @@  static int link_path_walk(const char *name, struct nameidata *nd)
 			return err;
 
 		if (err) {
-			const char *s = get_link(nd);
+			const char *s = get_link(nd, false);
 
 			if (IS_ERR(s))
 				return PTR_ERR(s);
@@ -2258,7 +2349,7 @@  static const char *trailing_symlink(struct nameidata *nd)
 		return ERR_PTR(error);
 	nd->flags |= LOOKUP_PARENT;
 	nd->stack[0].name = NULL;
-	s = get_link(nd);
+	s = get_link(nd, true);
 	return s ? s : "";
 }
 
@@ -3508,6 +3599,7 @@  static int do_o_path(struct nameidata *nd, unsigned flags, struct file *file)
 	if (!error) {
 		audit_inode(nd->name, path.dentry, 0);
 		error = vfs_open(&path, file);
+		file->f_mode |= nd->opath_mask;
 		path_put(&path);
 	}
 	return error;
@@ -3519,6 +3611,9 @@  static struct file *path_openat(struct nameidata *nd,
 	struct file *file;
 	int error;
 
+	nd->acc_mode = op->acc_mode;
+	nd->opath_mask = op->opath_mask;
+
 	file = alloc_empty_file(op->open_flag, current_cred());
 	if (IS_ERR(file))
 		return file;
diff --git a/fs/open.c b/fs/open.c
index b5b80469b93d..ab20eae39df7 100644
--- a/fs/open.c
+++ b/fs/open.c
@@ -982,8 +982,9 @@  static inline int build_open_flags(int flags, umode_t mode, struct open_flags *o
 		acc_mode |= MAY_APPEND;
 
 	op->acc_mode = acc_mode;
-
 	op->intent = flags & O_PATH ? 0 : LOOKUP_OPEN;
+	/* For O_PATH backwards-compatibility we default to an all-set mask. */
+	op->opath_mask = FMODE_PATH_READ | FMODE_PATH_WRITE;
 
 	if (flags & O_CREAT) {
 		op->intent |= LOOKUP_CREATE;
diff --git a/fs/proc/fd.c b/fs/proc/fd.c
index 81882a13212d..9b7d8becb002 100644
--- a/fs/proc/fd.c
+++ b/fs/proc/fd.c
@@ -104,11 +104,30 @@  static void tid_fd_update_inode(struct task_struct *task, struct inode *inode,
 	task_dump_owner(task, 0, &inode->i_uid, &inode->i_gid);
 
 	if (S_ISLNK(inode->i_mode)) {
+		/*
+		 * Always set +x (depending on the fmode type), since there
+		 * currently aren't FMODE_PATH_EXEC restrictions and there is
+		 * no O_MAYEXEC yet. This might change in the future, in which
+		 * case we will restrict +x.
+		 */
 		unsigned i_mode = S_IFLNK;
+		if (f_mode & FMODE_PATH)
+			i_mode |= S_IXGRP;
+		else
+			i_mode |= S_IXUSR;
+		/*
+		 * Construct the mode bits based on the open-mode. The u+rwx
+		 * bits are for "ordinary" open modes while g+rwx are for
+		 * O_PATH modes.
+		 */
 		if (f_mode & FMODE_READ)
-			i_mode |= S_IRUSR | S_IXUSR;
+			i_mode |= S_IRUSR;
 		if (f_mode & FMODE_WRITE)
-			i_mode |= S_IWUSR | S_IXUSR;
+			i_mode |= S_IWUSR;
+		if (f_mode & FMODE_PATH_READ)
+			i_mode |= S_IRGRP;
+		if (f_mode & FMODE_PATH_WRITE)
+			i_mode |= S_IWGRP;
 		inode->i_mode = i_mode;
 	}
 	security_task_to_inode(task, inode);
diff --git a/include/linux/fs.h b/include/linux/fs.h
index f7fdfe93e25d..f7df213405ea 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -173,6 +173,10 @@  typedef int (dio_iodone_t)(struct kiocb *iocb, loff_t offset,
 /* File does not contribute to nr_files count */
 #define FMODE_NOACCOUNT		((__force fmode_t)0x20000000)
 
+/* File is an O_PATH descriptor which can be upgraded to (read, write). */
+#define FMODE_PATH_READ		((__force fmode_t)0x40000000)
+#define FMODE_PATH_WRITE	((__force fmode_t)0x80000000)
+
 /*
  * Flag for rw_copy_check_uvector and compat_rw_copy_check_uvector
  * that indicates that they should check the contents of the iovec are