On 6/17/24 20:21, Kuniyuki Iwashima wrote:

From: Michal Luczaj <redacted>
Date: Mon, 17 Jun 2024 01:28:52 +0200

quoted

(...)
Another AF_UNIX sockmap issue is with OOB. When OOB packet is sent, skb is
added to recv queue, but also u->oob_skb is set. Here's the problem: when
this skb goes through bpf_sk_redirect_map() and is moved between socks,
oob_skb remains set on the original sock.

Good catch!

quoted

[   23.688994] WARNING: CPU: 2 PID: 993 at net/unix/garbage.c:351 unix_collect_queue+0x6c/0xb0
[   23.689019] CPU: 2 PID: 993 Comm: kworker/u32:13 Not tainted 6.10.0-rc2+ #137
[   23.689021] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Arch Linux 1.16.3-1-1 04/01/2014
[   23.689024] Workqueue: events_unbound __unix_gc
[   23.689027] RIP: 0010:unix_collect_queue+0x6c/0xb0

I wanted to write a patch, but then I realized I'm not sure what's the
expected behaviour. Should the oob_skb setting follow to the skb's new sock
or should it be dropped (similarly to what is happening today with
scm_fp_list, i.e. redirect strips inflights)?

The former will require large refactoring as we need to check if the
redirect happens for BPF_F_INGRESS and if the redirected sk is also
SOCK_STREAM etc.

So, I'd go with the latter.  Probably we can check if skb is u->oob_skb
and drop OOB data and retry next in unix_stream_read_skb(), and forbid
MSG_OOB in unix_bpf_recvmsg().
(...)

Yeah, sounds reasonable. I'm just not sure I understand the retry part. For
each skb_queue_tail() there's one ->sk_data_ready() (which does
->read_skb()). Why bother with a retry?

This is what I was thinking:

 static int unix_stream_read_skb(struct sock *sk, skb_read_actor_t recv_actor)
 {
+	struct unix_sock *u = unix_sk(sk);
+	struct sk_buff *skb;
+	int err;
+
 	if (unlikely(READ_ONCE(sk->sk_state) != TCP_ESTABLISHED))
 		return -ENOTCONN;
 
-	return unix_read_skb(sk, recv_actor);
+	mutex_lock(&u->iolock);
+	skb = skb_recv_datagram(sk, MSG_DONTWAIT, &err);
+
+#if IS_ENABLED(CONFIG_AF_UNIX_OOB)
+	if (skb) {
+		bool drop = false;
+
+		spin_lock(&sk->sk_receive_queue.lock);
+		if (skb == u->oob_skb) {
+			WRITE_ONCE(u->oob_skb, NULL);
+			drop = true;
+		}
+		spin_unlock(&sk->sk_receive_queue.lock);
+
+		if (drop) {
+			WARN_ON_ONCE(skb_unref(skb));
+			kfree_skb(skb);
+			skb = NULL;
+			err = 0;
+		}
+	}
+#endif
+
+	mutex_unlock(&u->iolock);
+	return skb ? recv_actor(sk, skb) : err;
 }