af_unix: Initialise scc_index in unix_add_edge().

Quang Le reported that the AF_UNIX GC could garbage-collect a
receive queue of an alive in-flight socket, with a nice repro.

The repro consists of three stages.

  1)
    1-a. Create a single cyclic reference with many sockets
    1-b. close() all sockets
    1-c. Trigger GC

  2)
    2-a. Pass sk-A to an embryo sk-B
    2-b. Pass sk-X to sk-X
    2-c. Trigger GC

  3)
    3-a. accept() the embryo sk-B
    3-b. Pass sk-B to sk-C
    3-c. close() the in-flight sk-A
    3-d. Trigger GC

As of 2-c, sk-A and sk-X are linked to unix_unvisited_vertices,
and unix_walk_scc() groups them into two different SCCs:

  unix_sk(sk-A)->vertex->scc_index = 2 (UNIX_VERTEX_INDEX_START)
  unix_sk(sk-X)->vertex->scc_index = 3

Once GC completes, unix_graph_grouped is set to true.
Also, unix_graph_maybe_cyclic is set to true due to sk-X's
cyclic self-reference, which makes close() trigger GC.

At 3-b, unix_add_edge() allocates unix_sk(sk-B)->vertex and
links it to unix_unvisited_vertices.

unix_update_graph() is called at 3-a. and 3-b., but neither
unix_graph_grouped nor unix_graph_maybe_cyclic is changed
because both sk-B's listener and sk-C are not in-flight.

3-c decrements sk-A's file refcnt to 1.

Since unix_graph_grouped is true at 3-d, unix_walk_scc_fast()
is finally called and iterates 3 sockets sk-A, sk-B, and sk-X:

  sk-A -> sk-B (-> sk-C)
  sk-X -> sk-X

This is totally fine.  All of them are not yet close()d and
should be grouped into different SCCs.

However, unix_vertex_dead() misjudges that sk-A and sk-B are
in the same SCC and sk-A is dead.

  unix_sk(sk-A)->scc_index == unix_sk(sk-B)->scc_index <-- Wrong!
  &&
  sk-A's file refcnt == unix_sk(sk-A)->vertex->out_degree
                                       ^-- 1 in-flight count for sk-B
  -> sk-A is dead !?

The problem is that unix_add_edge() does not initialise scc_index.

Stage 1) is used for heap spraying, making a newly allocated
vertex have vertex->scc_index == 2 (UNIX_VERTEX_INDEX_START)
set by unix_walk_scc() at 1-c.

Let's track the max SCC index from the previous unix_walk_scc()
call and assign the max + 1 to a new vertex's scc_index.

This way, we can continue to avoid Tarjan's algorithm while
preventing misjudgments.

Fixes: ad081928a8 ("af_unix: Avoid Tarjan's algorithm if unnecessary.")
Reported-by: Quang Le <quanglex97@gmail.com>
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20251109025233.3659187-1-kuniyu@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>

net/unix/garbage.c

@@ -145,6 +145,7 @@ enum unix_vertex_index {
 };
 
 static unsigned long unix_vertex_unvisited_index = UNIX_VERTEX_INDEX_MARK1;
+static unsigned long unix_vertex_max_scc_index = UNIX_VERTEX_INDEX_START;
 
 static void unix_add_edge(struct scm_fp_list *fpl, struct unix_edge *edge)
 {
@@ -153,6 +154,7 @@ static void unix_add_edge(struct scm_fp_list *fpl, struct unix_edge *edge)
 	if (!vertex) {
 		vertex = list_first_entry(&fpl->vertices, typeof(*vertex), entry);
 		vertex->index = unix_vertex_unvisited_index;
+		vertex->scc_index = ++unix_vertex_max_scc_index;
 		vertex->out_degree = 0;
 		INIT_LIST_HEAD(&vertex->edges);
 		INIT_LIST_HEAD(&vertex->scc_entry);
@@ -489,10 +491,15 @@ prev_vertex:
 				scc_dead = unix_vertex_dead(v);
 		}
 
-		if (scc_dead)
+		if (scc_dead) {
 			unix_collect_skb(&scc, hitlist);
-		else if (!unix_graph_maybe_cyclic)
+		} else {
-			unix_graph_maybe_cyclic = unix_scc_cyclic(&scc);
+			if (unix_vertex_max_scc_index < vertex->scc_index)
+				unix_vertex_max_scc_index = vertex->scc_index;
+
+			if (!unix_graph_maybe_cyclic)
+				unix_graph_maybe_cyclic = unix_scc_cyclic(&scc);
+		}
 
 		list_del(&scc);
 	}
@@ -507,6 +514,7 @@ static void unix_walk_scc(struct sk_buff_head *hitlist)
 	unsigned long last_index = UNIX_VERTEX_INDEX_START;
 
 	unix_graph_maybe_cyclic = false;
+	unix_vertex_max_scc_index = UNIX_VERTEX_INDEX_START;
 
 	/* Visit every vertex exactly once.
 	 * __unix_walk_scc() moves visited vertices to unix_visited_vertices.