[net] af_unix: fix garbage collect vs. MSG_PEEK

Commit Message

Greg KH July 26, 2021, 3:36 p.m. UTC

From: Miklos Szeredi <[e-mail]>

Gc assumes that in-flight sockets that don't have an external ref can't
gain one while unix_gc_lock is held.  That is true because
unix_notinflight() will be called before detaching fds, which takes
unix_gc_lock.

Only MSG_PEEK was somehow overlooked.  That one also clones the fds, also
keeping them in the skb.  But through MSG_PEEK an external reference can
definitely be gained without ever touching unix_gc_lock.

Signed-off-by: Miklos Szeredi <[e-mail]>
Cc: <[e-mail]>
Signed-off-by: Greg Kroah-Hartman <[e-mail]>
---
 net/unix/af_unix.c | 16 ++++++++++++++--
 1 file changed, 14 insertions(+), 2 deletions(-)

Note, this is a resend of this old submission that somehow fell through
the cracks:
	https://lore.kernel.org/netdev/[e-mail]/
and was never submitted "properly" and this issue never seemed to get
resolved properly.

I've cleaned it up and made the change much smaller and localized to
only one file.  I kept Miklos's authorship as he did the hard work on
this, I just removed lines and fixed a formatting issue :)

Comments

Kees Cook July 26, 2021, 7:27 p.m. UTC | #1

On Mon, Jul 26, 2021 at 05:36:21PM +0200, Greg Kroah-Hartman wrote:
> From: Miklos Szeredi <[e-mail]>
> 
> Gc assumes that in-flight sockets that don't have an external ref can't

I think this commit log could be expanded. I had to really study things
to even beging to understand what was going on. I assume "Gc" here means
specifically unix_gc()?

> gain one while unix_gc_lock is held.  That is true because
> unix_notinflight() will be called before detaching fds, which takes
> unix_gc_lock.

In reading the code, I *think* what is being protected by unix_gc_lock is
user->unix_inflight, u->inflight, unix_tot_inflight, and gc_inflight_list?

I note that unix_tot_inflight isn't an atomic but is read outside of
locking by unix_release_sock() and wait_for_unix_gc(), which seems wrong
(or at least inefficient).

But regardless, are the "external references" the f_count (i.e. get_file()
of u->sk.sk_socket->file) being changed by scm_fp_dup() and read by
unix_gc() (i.e. file_count())? It seems the test in unix_gc() is for
the making sure f_count isn't out of sync with u->inflight (is this the
corresponding "internal" reference?):

                total_refs = file_count(u->sk.sk_socket->file);
                inflight_refs = atomic_long_read(&u->inflight);

                BUG_ON(inflight_refs < 1);
                BUG_ON(total_refs < inflight_refs);
                if (total_refs == inflight_refs) {

> Only MSG_PEEK was somehow overlooked.  That one also clones the fds, also
> keeping them in the skb.  But through MSG_PEEK an external reference can
> definitely be gained without ever touching unix_gc_lock.

The idea appears to be that all scm_fp_dup() callers need to refresh the
u->inflight counts which is what unix_attach_fds() and unix_detach_fds()
do. Why is lock/unlock sufficient for unix_peek_fds()?

I assume the rationale is because MSG_PEEK uses a temporary scm, which
only gets fput() clean-up on destroy ("inflight" is neither incremented
nor decremented at any point in the scm lifetime).

But I don't see why any of this helps.

unix_attach_fds():
	fget(), spin_lock(), inflight++, spin_unlock()
unix_detach_fds():
	spin_lock(), inflight--, spin_unlock(), fput()
unix_peek_fds():
	fget(), spin_lock(), spin_unlock()
unix_gx():
	spin_lock(), "total_refs == inflight_refs" to hitlist,
	spin_unlock(), free hitlist skbs

Doesn't this mean total_refs and inflight_refs can still get out of
sync? What keeps an skb from being "visible" to unix_peek_fds() between
the unix_gx() spin_unlock() and the unix_peek_fds() fget()?

A: unix_gx():
	spin_lock()
	find "total_refs == inflight_refs", add to hitlist
	spin_unlock()
B: unix_peek_fds():
	fget()
A: unix_gc():
	walk hitlist and free(skb)
B: unix_peek_fds():
	*use freed skb*

I feel like I must be missing something since the above race would
appear to exist even for unix_attach_fds()/unix_detach_fds():

A: unix_gx():
	spin_lock()
	find "total_refs == inflight_refs", add to hitlist
	spin_unlock()
B: unix_attach_fds():
	fget()
A: unix_gc():
	walk hitlist and free(skb)
B: unix_attach_fds():
	*use freed skb*

I'm assuming I'm missing a top-level usage count on skb that is held by
callers, which means the skb isn't actually freed by unix_gc(). But I
return to not understanding why adding the lock/unlock helps.

What are the expected locking semantics here?

-Kees

> 
> Signed-off-by: Miklos Szeredi <[e-mail]>
> Cc: <[e-mail]>
> Signed-off-by: Greg Kroah-Hartman <[e-mail]>
> ---
>  net/unix/af_unix.c | 16 ++++++++++++++--
>  1 file changed, 14 insertions(+), 2 deletions(-)
> 
> Note, this is a resend of this old submission that somehow fell through
> the cracks:
> 	https://lore.kernel.org/netdev/[e-mail]/
> and was never submitted "properly" and this issue never seemed to get
> resolved properly.
> 
> I've cleaned it up and made the change much smaller and localized to
> only one file.  I kept Miklos's authorship as he did the hard work on
> this, I just removed lines and fixed a formatting issue :)
> 
> 
> diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
> index 23c92ad15c61..cdea997aa5bf 100644
> --- a/net/unix/af_unix.c
> +++ b/net/unix/af_unix.c
> @@ -1526,6 +1526,18 @@ static int unix_getname(struct socket *sock, struct sockaddr *uaddr, int peer)
>  	return err;
>  }
>  
> +static void unix_peek_fds(struct scm_cookie *scm, struct sk_buff *skb)
> +{
> +	scm->fp = scm_fp_dup(UNIXCB(skb).fp);
> +
> +	/* During garbage collection it is assumed that in-flight sockets don't
> +	 * get a new external reference.  So we need to wait until current run
> +	 * finishes.
> +	 */
> +	spin_lock(&unix_gc_lock);
> +	spin_unlock(&unix_gc_lock);
> +}

Miklos Szeredi July 29, 2021, 2:29 p.m. UTC | #2

On Mon, Jul 26, 2021 at 9:27 PM Kees Cook <[e-mail]> wrote:
>
> On Mon, Jul 26, 2021 at 05:36:21PM +0200, Greg Kroah-Hartman wrote:
> > From: Miklos Szeredi <[e-mail]>
> >
> > Gc assumes that in-flight sockets that don't have an external ref can't
>
> I think this commit log could be expanded. I had to really study things
> to even beging to understand what was going on. I assume "Gc" here means
> specifically unix_gc()?

Yeah, the original description was not too good.  Commit cbcf01128d0a
("af_unix: fix garbage collect vs MSG_PEEK") now in Linus' tree has a
much expanded description.

> I note that unix_tot_inflight isn't an atomic but is read outside of
> locking by unix_release_sock() and wait_for_unix_gc(), which seems wrong
> (or at least inefficient).

I don't think it matters in practice.   Do you have specific worries?

> Doesn't this mean total_refs and inflight_refs can still get out of
> sync? What keeps an skb from being "visible" to unix_peek_fds() between
> the unix_gx() spin_unlock() and the unix_peek_fds() fget()?
>
> A: unix_gx():
>         spin_lock()
>         find "total_refs == inflight_refs", add to hitlist
>         spin_unlock()
> B: unix_peek_fds():
>         fget()
> A: unix_gc():
>         walk hitlist and free(skb)
> B: unix_peek_fds():
>         *use freed skb*
>
> I feel like I must be missing something since the above race would
> appear to exist even for unix_attach_fds()/unix_detach_fds():

What you are missing is that anything that could have been peeked must
not have been garbage collected.  I.e. the garbage collection
algorithm will find that there's an external in-flight reference to
the peeked socket and so it will not add it to the hitlist.

Thanks,
Miklos

Patch

diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
index 23c92ad15c61..cdea997aa5bf 100644
--- a/net/unix/af_unix.c
+++ b/net/unix/af_unix.c
@@ -1526,6 +1526,18 @@  static int unix_getname(struct socket *sock, struct sockaddr *uaddr, int peer)
 	return err;
 }
 
+static void unix_peek_fds(struct scm_cookie *scm, struct sk_buff *skb)
+{
+	scm->fp = scm_fp_dup(UNIXCB(skb).fp);
+
+	/* During garbage collection it is assumed that in-flight sockets don't
+	 * get a new external reference.  So we need to wait until current run
+	 * finishes.
+	 */
+	spin_lock(&unix_gc_lock);
+	spin_unlock(&unix_gc_lock);
+}
+
 static int unix_scm_to_skb(struct scm_cookie *scm, struct sk_buff *skb, bool send_fds)
 {
 	int err = 0;
@@ -2175,7 +2187,7 @@  static int unix_dgram_recvmsg(struct socket *sock, struct msghdr *msg,
 		sk_peek_offset_fwd(sk, size);
 
 		if (UNIXCB(skb).fp)
-			scm.fp = scm_fp_dup(UNIXCB(skb).fp);
+			unix_peek_fds(&scm, skb);
 	}
 	err = (flags & MSG_TRUNC) ? skb->len - skip : size;
 
@@ -2418,7 +2430,7 @@  static int unix_stream_read_generic(struct unix_stream_read_state *state,
 			/* It is questionable, see note in unix_dgram_recvmsg.
 			 */
 			if (UNIXCB(skb).fp)
-				scm.fp = scm_fp_dup(UNIXCB(skb).fp);
+				unix_peek_fds(&scm, skb);
 
 			sk_peek_offset_fwd(sk, chunk);