[PATCH net v2] af_unix: Fix garbage collector racing against connect()

[Michal Luczaj <mhal@rbox.co>]

Garbage collector does not take into account the risk of embryo getting
enqueued during the garbage collection. If such embryo has a peer that
carries SCM_RIGHTS, two consecutive passes of scan_children() may see a
different set of children. Leading to an incorrectly elevated inflight
count, and then a dangling pointer within the gc_inflight_list.

sockets are AF_UNIX/SOCK_STREAM
S is an unconnected socket
L is a listening in-flight socket bound to addr, not in fdtable
V's fd will be passed via sendmsg(), gets inflight count bumped

connect(S, addr)	sendmsg(S, [V]); close(V)	__unix_gc()
----------------	-------------------------	-----------

NS = unix_create1()
skb1 = sock_wmalloc(NS)
L = unix_find_other(addr)
unix_state_lock(L)
unix_peer(S) = NS
			// V count=1 inflight=0

 			NS = unix_peer(S)
 			skb2 = sock_alloc()
			skb_queue_tail(NS, skb2[V])

			// V became in-flight
			// V count=2 inflight=1

			close(V)

			// V count=1 inflight=1
			// GC candidate condition met

						for u in gc_inflight_list:
						  if (total_refs == inflight_refs)
						    add u to gc_candidates

						// gc_candidates={L, V}

						for u in gc_candidates:
						  scan_children(u, dec_inflight)

						// embryo (skb1) was not
						// reachable from L yet, so V's
						// inflight remains unchanged
__skb_queue_tail(L, skb1)
unix_state_unlock(L)
						for u in gc_candidates:
						  if (u.inflight)
						    scan_children(u, inc_inflight_move_tail)

						// V count=1 inflight=2 (!)

If there is a GC-candidate listening socket, lock/unlock its state. This
makes GC wait until the end of any ongoing connect() to that socket. After
flipping the lock, a possibly SCM-laden embryo is already enqueued. And if
there is another embryo coming, it can not possibly carry SCM_RIGHTS. At
this point, unix_inflight() can not happen because unix_gc_lock is already
taken. Inflight graph remains unaffected.

Fixes: 1fd05ba5a2f2 ("[AF_UNIX]: Rewrite garbage collector, fixes race.")
Signed-off-by: Michal Luczaj <mhal@rbox.co>
---
v2:
  - Adhere to reverse xmas tree variable ordering
  - Expand commit message
  - Drop the reproducer

v1: https://lore.kernel.org/netdev/20240408161336.612064-1-mhal@rbox.co/

 net/unix/garbage.c | 18 +++++++++++++++++-
 1 file changed, 17 insertions(+), 1 deletion(-)

diff --git a/net/unix/garbage.c b/net/unix/garbage.c
index fa39b6265238..6433a414acf8 100644
--- a/net/unix/garbage.c
+++ b/net/unix/garbage.c
@@ -274,11 +274,22 @@ static void __unix_gc(struct work_struct *work)
 	 * receive queues.  Other, non candidate sockets _can_ be
 	 * added to queue, so we must make sure only to touch
 	 * candidates.
+	 *
+	 * Embryos, though never candidates themselves, affect which
+	 * candidates are reachable by the garbage collector.  Before
+	 * being added to a listener's queue, an embryo may already
+	 * receive data carrying SCM_RIGHTS, potentially making the
+	 * passed socket a candidate that is not yet reachable by the
+	 * collector.  It becomes reachable once the embryo is
+	 * enqueued.  Therefore, we must ensure that no SCM-laden
+	 * embryo appears in a (candidate) listener's queue between
+	 * consecutive scan_children() calls.
 	 */
 	list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
+		struct sock *sk = &u->sk;
 		long total_refs;
 
-		total_refs = file_count(u->sk.sk_socket->file);
+		total_refs = file_count(sk->sk_socket->file);
 
 		WARN_ON_ONCE(!u->inflight);
 		WARN_ON_ONCE(total_refs < u->inflight);
@@ -286,6 +297,11 @@ static void __unix_gc(struct work_struct *work)
 			list_move_tail(&u->link, &gc_candidates);
 			__set_bit(UNIX_GC_CANDIDATE, &u->gc_flags);
 			__set_bit(UNIX_GC_MAYBE_CYCLE, &u->gc_flags);
+
+			if (sk->sk_state == TCP_LISTEN) {
+				unix_state_lock(sk);
+				unix_state_unlock(sk);
+			}
 		}
 	}
 
-- 
2.44.0