sock_map: avoid race between sock_map_close and sk_psock_put

commit 4b4647add7 upstream.

sk_psock_get will return NULL if the refcount of psock has gone to 0, which
will happen when the last call of sk_psock_put is done. However,
sk_psock_drop may not have finished yet, so the close callback will still
point to sock_map_close despite psock being NULL.

This can be reproduced with a thread deleting an element from the sock map,
while the second one creates a socket, adds it to the map and closes it.

That will trigger the WARN_ON_ONCE:

------------[ cut here ]------------
WARNING: CPU: 1 PID: 7220 at net/core/sock_map.c:1701 sock_map_close+0x2a2/0x2d0 net/core/sock_map.c:1701
Modules linked in:
CPU: 1 PID: 7220 Comm: syz-executor380 Not tainted 6.9.0-syzkaller-07726-g3c999d1ae3c7 #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 04/02/2024
RIP: 0010:sock_map_close+0x2a2/0x2d0 net/core/sock_map.c:1701
Code: df e8 92 29 88 f8 48 8b 1b 48 89 d8 48 c1 e8 03 42 80 3c 20 00 74 08 48 89 df e8 79 29 88 f8 4c 8b 23 eb 89 e8 4f 15 23 f8 90 <0f> 0b 90 48 83 c4 08 5b 41 5c 41 5d 41 5e 41 5f 5d e9 13 26 3d 02
RSP: 0018:ffffc9000441fda8 EFLAGS: 00010293
RAX: ffffffff89731ae1 RBX: ffffffff94b87540 RCX: ffff888029470000
RDX: 0000000000000000 RSI: ffffffff8bcab5c0 RDI: ffffffff8c1faba0
RBP: 0000000000000000 R08: ffffffff92f9b61f R09: 1ffffffff25f36c3
R10: dffffc0000000000 R11: fffffbfff25f36c4 R12: ffffffff89731840
R13: ffff88804b587000 R14: ffff88804b587000 R15: ffffffff89731870
FS:  000055555e080380(0000) GS:ffff8880b9500000(0000) knlGS:0000000000000000
CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000000000000 CR3: 00000000207d4000 CR4: 0000000000350ef0
Call Trace:
 <TASK>
 unix_release+0x87/0xc0 net/unix/af_unix.c:1048
 __sock_release net/socket.c:659 [inline]
 sock_close+0xbe/0x240 net/socket.c:1421
 __fput+0x42b/0x8a0 fs/file_table.c:422
 __do_sys_close fs/open.c:1556 [inline]
 __se_sys_close fs/open.c:1541 [inline]
 __x64_sys_close+0x7f/0x110 fs/open.c:1541
 do_syscall_x64 arch/x86/entry/common.c:52 [inline]
 do_syscall_64+0xf5/0x240 arch/x86/entry/common.c:83
 entry_SYSCALL_64_after_hwframe+0x77/0x7f
RIP: 0033:0x7fb37d618070
Code: 00 00 48 c7 c2 b8 ff ff ff f7 d8 64 89 02 b8 ff ff ff ff eb d4 e8 10 2c 00 00 80 3d 31 f0 07 00 00 74 17 b8 03 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 48 c3 0f 1f 80 00 00 00 00 48 83 ec 18 89 7c
RSP: 002b:00007ffcd4a525d8 EFLAGS: 00000202 ORIG_RAX: 0000000000000003
RAX: ffffffffffffffda RBX: 0000000000000005 RCX: 00007fb37d618070
RDX: 0000000000000010 RSI: 00000000200001c0 RDI: 0000000000000004
RBP: 0000000000000000 R08: 0000000100000000 R09: 0000000100000000
R10: 0000000000000000 R11: 0000000000000202 R12: 0000000000000000
R13: 0000000000000000 R14: 0000000000000000 R15: 0000000000000000
 </TASK>

Use sk_psock, which will only check that the pointer is not been set to
NULL yet, which should only happen after the callbacks are restored. If,
then, a reference can still be gotten, we may call sk_psock_stop and cancel
psock->work.

As suggested by Paolo Abeni, reorder the condition so the control flow is
less convoluted.

After that change, the reproducer does not trigger the WARN_ON_ONCE
anymore.

Suggested-by: Paolo Abeni <pabeni@redhat.com>
Reported-by: syzbot+07a2e4a1a57118ef7355@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=07a2e4a1a57118ef7355
Fixes: aadb2bb83f ("sock_map: Fix a potential use-after-free in sock_map_close()")
Fixes: 5b4a79ba65 ("bpf, sockmap: Don't let sock_map_{close,destroy,unhash} call itself")
Cc: stable@vger.kernel.org
Signed-off-by: Thadeu Lima de Souza Cascardo <cascardo@igalia.com>
Acked-by: Jakub Sitnicki <jakub@cloudflare.com>
Link: https://lore.kernel.org/r/20240524144702.1178377-1-cascardo@igalia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

net/core/sock_map.c

@@ -1639,19 +1639,23 @@ void sock_map_close(struct sock *sk, long timeout)
 
 	lock_sock(sk);
 	rcu_read_lock();
-	psock = sk_psock_get(sk);
-	if (unlikely(!psock)) {
-		rcu_read_unlock();
-		release_sock(sk);
-		saved_close = READ_ONCE(sk->sk_prot)->close;
-	} else {
+	psock = sk_psock(sk);
+	if (likely(psock)) {
 		saved_close = psock->saved_close;
 		sock_map_remove_links(sk, psock);
+		psock = sk_psock_get(sk);
+		if (unlikely(!psock))
+			goto no_psock;
 		rcu_read_unlock();
 		sk_psock_stop(psock);
 		release_sock(sk);
 		cancel_delayed_work_sync(&psock->work);
 		sk_psock_put(sk, psock);
+	} else {
+		saved_close = READ_ONCE(sk->sk_prot)->close;
+no_psock:
+		rcu_read_unlock();
+		release_sock(sk);
 	}
 
 	/* Make sure we do not recurse. This is a bug.