llcp_sock_connect() only rejects LLCP_CONNECTED and LLCP_CONNECTING, so a socket that is still linked in local->sockets can be linked a second time into local->connecting_sockets. That covers bound and listening sockets, and sockets that a peer moved to LLCP_CLOSED with DISC or DM. Both lists use sk->sk_node. When the peer answers with CC, nfc_llcp_recv_cc() removes the node from the connecting list and links it back into local->sockets, where the stale bind-time entry makes the node point to itself. Every later sk_for_each() over local->sockets then spins forever with the read lock held, which pins the CPU and stalls RCU, and the socket is never freed. Reject the connection while sk_node is still linked. A socket has to be closed before it can be connected. Reproduced on v7.4-rc3 as uid 1000: a soft lockup / RCU stall in nfc_llcp_sock_from_sn() while walking local->sockets. This issue was discovered by Tencent CodeBuddy Security. Link: https://lore.kernel.org/all/20261009092338.3146149-1-bsdhenrymartin@gmail.com/ Fixes: a69f32af86e3 ("NFC: Socket linked list") Cc: stable@vger.kernel.org Signed-off-by: Henry Martin --- v1 -> v2: reject any socket that is still hashed (sk_hashed) instead of only LLCP_BOUND, covering bind()+listen()+connect() and sockets left linked after DISC/DM. net/nfc/llcp_sock.c | 9 +++++++++ 1 file changed, 9 insertions(+) diff --git a/net/nfc/llcp_sock.c b/net/nfc/llcp_sock.c index 1e5ee4bcde68..01b311d5adda 100644 --- a/net/nfc/llcp_sock.c +++ b/net/nfc/llcp_sock.c @@ -714,6 +714,15 @@ static int llcp_sock_connect(struct socket *sock, struct sockaddr_unsized *_addr ret = -EINPROGRESS; goto error; } + /* A socket that is still linked in local->sockets must not be linked + * into local->connecting_sockets as well: both lists use sk_node, and + * the CC handler links the node back into local->sockets, which turns + * the stale entry into a self-loop. + */ + if (sk_hashed(sk)) { + ret = -EISCONN; + goto error; + } dev = nfc_get_device(addr->dev_idx); if (dev == NULL) { -- 2.43.7