net: use the full socket in sk_mc_loop()

sk_mc_loop() is called from ip_mc_output() and ip6_finish_output2()
with skb->sk, which is not always a full socket:

- Since commit ca6fb06518 ("tcp: attach SYNACK messages to request
  sockets instead of listener"), a SYNACK is attached to its request
  socket.

- Since commit 79636038d3 ("ipv4: tcp: give socket pointer to control
  skbs"), an ACK or RST sent by ip_send_unicast_reply() on behalf of a
  TIME_WAIT socket is attached to the timewait socket.

When such a packet uses a multicast route, inet_test_bit(MC_LOOP, sk)
reads inet_flags past the end of the request or timewait socket.

This happens for instance when a SYN with a multicast source address
reaches a listener, after being looped back with its dst attached, so
that ip_rcv_finish_core() skips the martian source check. A raw
IP_HDRINCL socket sending to a local address does this, as does
re-injection by nft dup or the iptables TEE target. The destination
can also come from a source route option, or from a route of type
multicast covering a unicast prefix.

BUG: KASAN: slab-out-of-bounds in sk_mc_loop+0x111/0x170
Read of size 8 at addr ffff88800f2b7f10 by task repro-raw/470
Call Trace:
 <IRQ>
 sk_mc_loop+0x111/0x170
 ip_mc_output+0x355/0x930
 ip_build_and_send_pkt+0xb87/0xc50
 tcp_v4_send_synack+0x500/0x6f0
 tcp_conn_request+0x2135/0x2d90
 tcp_v4_conn_request+0xa5/0x210
 tcp_rcv_state_process+0x136e/0x6920
 tcp_v4_do_rcv+0x339/0xb10
 tcp_v4_rcv+0x34ab/0x3ab0
 ip_protocol_deliver_rcu+0x6e/0x3e0
 ip_local_deliver_finish+0x34d/0x510
 ip_local_deliver+0x1bc/0x310
 ip_rcv+0x390/0x410
 __netif_receive_skb_one_core+0x199/0x1e0
 process_backlog+0x239/0x680
 __napi_poll+0xb5/0x650
 net_rx_action+0x980/0xd60
 handle_softirqs+0x17f/0x590
 do_softirq+0x3f/0x60
 </IRQ>

The buggy address belongs to the object at ffff88800f2b7d60
 which belongs to the cache request_sock_TCP of size 312
The buggy address is located 120 bytes to the right of
 allocated 312-byte region [ffff88800f2b7d60, ffff88800f2b7e98)

Use sk_const_to_full_sk(), which restores the behavior from before
the two commits above: a request socket uses the setting of its
listener, and a timewait socket gets the default (loop back), as the
per-cpu control socket used before had.

Fixes: ca6fb06518 ("tcp: attach SYNACK messages to request sockets instead of listener")
Fixes: 79636038d3 ("ipv4: tcp: give socket pointer to control skbs")
Signed-off-by: Theodor Arsenij Larionov Trichkine <theodorlarionov@gmail.com>
[edumazet: rewrote the changelog, added the TIME_WAIT case]
Signed-off-by: Eric Dumazet <edumazet@kernel.org>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20261012101029.927296-2-edumazet@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

net/core/sock.c

@@ -769,6 +769,7 @@ bool sk_mc_loop(const struct sock *sk)
 {
 	if (dev_recursion_level())
 		return false;
+	sk = sk_const_to_full_sk(sk);
 	if (!sk)
 		return true;
 	/* IPV6_ADDRFORM can change sk->sk_family under us. */