1cf80e0e4SHerbert Xu #include <linux/crypto.h> 210467163SJerry Chu #include <linux/err.h> 32100c8d2SYuchung Cheng #include <linux/init.h> 42100c8d2SYuchung Cheng #include <linux/kernel.h> 510467163SJerry Chu #include <linux/list.h> 610467163SJerry Chu #include <linux/tcp.h> 710467163SJerry Chu #include <linux/rcupdate.h> 810467163SJerry Chu #include <linux/rculist.h> 910467163SJerry Chu #include <net/inetpeer.h> 1010467163SJerry Chu #include <net/tcp.h> 112100c8d2SYuchung Cheng 120d41cca4SYuchung Cheng int sysctl_tcp_fastopen __read_mostly = TFO_CLIENT_ENABLE; 1310467163SJerry Chu 1410467163SJerry Chu struct tcp_fastopen_context __rcu *tcp_fastopen_ctx; 1510467163SJerry Chu 1610467163SJerry Chu static DEFINE_SPINLOCK(tcp_fastopen_ctx_lock); 1710467163SJerry Chu 18222e83d2SHannes Frederic Sowa void tcp_fastopen_init_key_once(bool publish) 19222e83d2SHannes Frederic Sowa { 20222e83d2SHannes Frederic Sowa static u8 key[TCP_FASTOPEN_KEY_LENGTH]; 21222e83d2SHannes Frederic Sowa 22222e83d2SHannes Frederic Sowa /* tcp_fastopen_reset_cipher publishes the new context 23222e83d2SHannes Frederic Sowa * atomically, so we allow this race happening here. 24222e83d2SHannes Frederic Sowa * 25222e83d2SHannes Frederic Sowa * All call sites of tcp_fastopen_cookie_gen also check 26222e83d2SHannes Frederic Sowa * for a valid cookie, so this is an acceptable risk. 27222e83d2SHannes Frederic Sowa */ 28222e83d2SHannes Frederic Sowa if (net_get_random_once(key, sizeof(key)) && publish) 29222e83d2SHannes Frederic Sowa tcp_fastopen_reset_cipher(key, sizeof(key)); 30222e83d2SHannes Frederic Sowa } 31222e83d2SHannes Frederic Sowa 3210467163SJerry Chu static void tcp_fastopen_ctx_free(struct rcu_head *head) 3310467163SJerry Chu { 3410467163SJerry Chu struct tcp_fastopen_context *ctx = 3510467163SJerry Chu container_of(head, struct tcp_fastopen_context, rcu); 3610467163SJerry Chu crypto_free_cipher(ctx->tfm); 3710467163SJerry Chu kfree(ctx); 3810467163SJerry Chu } 3910467163SJerry Chu 4010467163SJerry Chu int tcp_fastopen_reset_cipher(void *key, unsigned int len) 4110467163SJerry Chu { 4210467163SJerry Chu int err; 4310467163SJerry Chu struct tcp_fastopen_context *ctx, *octx; 4410467163SJerry Chu 4510467163SJerry Chu ctx = kmalloc(sizeof(*ctx), GFP_KERNEL); 4610467163SJerry Chu if (!ctx) 4710467163SJerry Chu return -ENOMEM; 4810467163SJerry Chu ctx->tfm = crypto_alloc_cipher("aes", 0, 0); 4910467163SJerry Chu 5010467163SJerry Chu if (IS_ERR(ctx->tfm)) { 5110467163SJerry Chu err = PTR_ERR(ctx->tfm); 5210467163SJerry Chu error: kfree(ctx); 5310467163SJerry Chu pr_err("TCP: TFO aes cipher alloc error: %d\n", err); 5410467163SJerry Chu return err; 5510467163SJerry Chu } 5610467163SJerry Chu err = crypto_cipher_setkey(ctx->tfm, key, len); 5710467163SJerry Chu if (err) { 5810467163SJerry Chu pr_err("TCP: TFO cipher key error: %d\n", err); 5910467163SJerry Chu crypto_free_cipher(ctx->tfm); 6010467163SJerry Chu goto error; 6110467163SJerry Chu } 6210467163SJerry Chu memcpy(ctx->key, key, len); 6310467163SJerry Chu 6410467163SJerry Chu spin_lock(&tcp_fastopen_ctx_lock); 6510467163SJerry Chu 6610467163SJerry Chu octx = rcu_dereference_protected(tcp_fastopen_ctx, 6710467163SJerry Chu lockdep_is_held(&tcp_fastopen_ctx_lock)); 6810467163SJerry Chu rcu_assign_pointer(tcp_fastopen_ctx, ctx); 6910467163SJerry Chu spin_unlock(&tcp_fastopen_ctx_lock); 7010467163SJerry Chu 7110467163SJerry Chu if (octx) 7210467163SJerry Chu call_rcu(&octx->rcu, tcp_fastopen_ctx_free); 7310467163SJerry Chu return err; 7410467163SJerry Chu } 7510467163SJerry Chu 763a19ce0eSDaniel Lee static bool __tcp_fastopen_cookie_gen(const void *path, 77149479d0SYuchung Cheng struct tcp_fastopen_cookie *foc) 7810467163SJerry Chu { 7910467163SJerry Chu struct tcp_fastopen_context *ctx; 803a19ce0eSDaniel Lee bool ok = false; 8110467163SJerry Chu 8210467163SJerry Chu rcu_read_lock(); 8310467163SJerry Chu ctx = rcu_dereference(tcp_fastopen_ctx); 8410467163SJerry Chu if (ctx) { 853a19ce0eSDaniel Lee crypto_cipher_encrypt_one(ctx->tfm, foc->val, path); 8610467163SJerry Chu foc->len = TCP_FASTOPEN_COOKIE_SIZE; 873a19ce0eSDaniel Lee ok = true; 8810467163SJerry Chu } 8910467163SJerry Chu rcu_read_unlock(); 903a19ce0eSDaniel Lee return ok; 913a19ce0eSDaniel Lee } 923a19ce0eSDaniel Lee 933a19ce0eSDaniel Lee /* Generate the fastopen cookie by doing aes128 encryption on both 943a19ce0eSDaniel Lee * the source and destination addresses. Pad 0s for IPv4 or IPv4-mapped-IPv6 953a19ce0eSDaniel Lee * addresses. For the longer IPv6 addresses use CBC-MAC. 963a19ce0eSDaniel Lee * 973a19ce0eSDaniel Lee * XXX (TFO) - refactor when TCP_FASTOPEN_COOKIE_SIZE != AES_BLOCK_SIZE. 983a19ce0eSDaniel Lee */ 993a19ce0eSDaniel Lee static bool tcp_fastopen_cookie_gen(struct request_sock *req, 1003a19ce0eSDaniel Lee struct sk_buff *syn, 1013a19ce0eSDaniel Lee struct tcp_fastopen_cookie *foc) 1023a19ce0eSDaniel Lee { 1033a19ce0eSDaniel Lee if (req->rsk_ops->family == AF_INET) { 1043a19ce0eSDaniel Lee const struct iphdr *iph = ip_hdr(syn); 1053a19ce0eSDaniel Lee 1063a19ce0eSDaniel Lee __be32 path[4] = { iph->saddr, iph->daddr, 0, 0 }; 1073a19ce0eSDaniel Lee return __tcp_fastopen_cookie_gen(path, foc); 1083a19ce0eSDaniel Lee } 1093a19ce0eSDaniel Lee 1103a19ce0eSDaniel Lee #if IS_ENABLED(CONFIG_IPV6) 1113a19ce0eSDaniel Lee if (req->rsk_ops->family == AF_INET6) { 1123a19ce0eSDaniel Lee const struct ipv6hdr *ip6h = ipv6_hdr(syn); 1133a19ce0eSDaniel Lee struct tcp_fastopen_cookie tmp; 1143a19ce0eSDaniel Lee 1153a19ce0eSDaniel Lee if (__tcp_fastopen_cookie_gen(&ip6h->saddr, &tmp)) { 1163a19ce0eSDaniel Lee struct in6_addr *buf = (struct in6_addr *) tmp.val; 11741c91996SLi RongQing int i; 1183a19ce0eSDaniel Lee 1193a19ce0eSDaniel Lee for (i = 0; i < 4; i++) 1203a19ce0eSDaniel Lee buf->s6_addr32[i] ^= ip6h->daddr.s6_addr32[i]; 1213a19ce0eSDaniel Lee return __tcp_fastopen_cookie_gen(buf, foc); 1223a19ce0eSDaniel Lee } 1233a19ce0eSDaniel Lee } 1243a19ce0eSDaniel Lee #endif 1253a19ce0eSDaniel Lee return false; 12610467163SJerry Chu } 1275b7ed089SYuchung Cheng 1287c85af88SEric Dumazet static struct sock *tcp_fastopen_create_child(struct sock *sk, 1295b7ed089SYuchung Cheng struct sk_buff *skb, 130843f4a55SYuchung Cheng struct dst_entry *dst, 1315b7ed089SYuchung Cheng struct request_sock *req) 1325b7ed089SYuchung Cheng { 13317846376SDave Jones struct tcp_sock *tp; 1345b7ed089SYuchung Cheng struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue; 1355b7ed089SYuchung Cheng struct sock *child; 136ba34e6d9SEric Dumazet u32 end_seq; 1375e0724d0SEric Dumazet bool own_req; 1385b7ed089SYuchung Cheng 1395b7ed089SYuchung Cheng req->num_retrans = 0; 1405b7ed089SYuchung Cheng req->num_timeout = 0; 1415b7ed089SYuchung Cheng req->sk = NULL; 1425b7ed089SYuchung Cheng 1435e0724d0SEric Dumazet child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL, 1445e0724d0SEric Dumazet NULL, &own_req); 14551456b29SIan Morris if (!child) 1467c85af88SEric Dumazet return NULL; 1475b7ed089SYuchung Cheng 1480536fcc0SEric Dumazet spin_lock(&queue->fastopenq.lock); 1490536fcc0SEric Dumazet queue->fastopenq.qlen++; 1500536fcc0SEric Dumazet spin_unlock(&queue->fastopenq.lock); 1515b7ed089SYuchung Cheng 1525b7ed089SYuchung Cheng /* Initialize the child socket. Have to fix some values to take 1535b7ed089SYuchung Cheng * into account the child is a Fast Open socket and is created 1545b7ed089SYuchung Cheng * only out of the bits carried in the SYN packet. 1555b7ed089SYuchung Cheng */ 1565b7ed089SYuchung Cheng tp = tcp_sk(child); 1575b7ed089SYuchung Cheng 1585b7ed089SYuchung Cheng tp->fastopen_rsk = req; 1599439ce00SEric Dumazet tcp_rsk(req)->tfo_listener = true; 1605b7ed089SYuchung Cheng 1615b7ed089SYuchung Cheng /* RFC1323: The window in SYN & SYN/ACK segments is never 1625b7ed089SYuchung Cheng * scaled. So correct it appropriately. 1635b7ed089SYuchung Cheng */ 1645b7ed089SYuchung Cheng tp->snd_wnd = ntohs(tcp_hdr(skb)->window); 1655b7ed089SYuchung Cheng 1665b7ed089SYuchung Cheng /* Activate the retrans timer so that SYNACK can be retransmitted. 167ca6fb065SEric Dumazet * The request socket is not added to the ehash 1685b7ed089SYuchung Cheng * because it's been added to the accept queue directly. 1695b7ed089SYuchung Cheng */ 1705b7ed089SYuchung Cheng inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS, 1715b7ed089SYuchung Cheng TCP_TIMEOUT_INIT, TCP_RTO_MAX); 1725b7ed089SYuchung Cheng 173ca6fb065SEric Dumazet atomic_set(&req->rsk_refcnt, 2); 1745b7ed089SYuchung Cheng 1755b7ed089SYuchung Cheng /* Now finish processing the fastopen child socket. */ 1765b7ed089SYuchung Cheng inet_csk(child)->icsk_af_ops->rebuild_header(child); 1775b7ed089SYuchung Cheng tcp_init_congestion_control(child); 1785b7ed089SYuchung Cheng tcp_mtup_init(child); 1795b7ed089SYuchung Cheng tcp_init_metrics(child); 1805b7ed089SYuchung Cheng tcp_init_buffer_space(child); 1815b7ed089SYuchung Cheng 1827656d842SEric Dumazet /* Queue the data carried in the SYN packet. 1837656d842SEric Dumazet * We used to play tricky games with skb_get(). 1847656d842SEric Dumazet * With lockless listener, it is a dead end. 1857656d842SEric Dumazet * Do not think about it. 1865b7ed089SYuchung Cheng * 187843f4a55SYuchung Cheng * XXX (TFO) - we honor a zero-payload TFO request for now, 188843f4a55SYuchung Cheng * (any reason not to?) but no need to queue the skb since 189843f4a55SYuchung Cheng * there is no data. How about SYN+FIN? 1905b7ed089SYuchung Cheng */ 191ba34e6d9SEric Dumazet end_seq = TCP_SKB_CB(skb)->end_seq; 192ba34e6d9SEric Dumazet if (end_seq != TCP_SKB_CB(skb)->seq + 1) { 1937656d842SEric Dumazet struct sk_buff *skb2 = skb_clone(skb, GFP_ATOMIC); 194ba34e6d9SEric Dumazet 195ba34e6d9SEric Dumazet if (likely(skb2)) { 196ba34e6d9SEric Dumazet skb_dst_drop(skb2); 197ba34e6d9SEric Dumazet __skb_pull(skb2, tcp_hdrlen(skb)); 198ba34e6d9SEric Dumazet skb_set_owner_r(skb2, child); 199ba34e6d9SEric Dumazet __skb_queue_tail(&child->sk_receive_queue, skb2); 2005b7ed089SYuchung Cheng tp->syn_data_acked = 1; 201d654976cSEric Dumazet 202d654976cSEric Dumazet /* u64_stats_update_begin(&tp->syncp) not needed here, 203d654976cSEric Dumazet * as we certainly are not changing upper 32bit value (0) 204d654976cSEric Dumazet */ 205bdd1f9edSEric Dumazet tp->bytes_received = end_seq - TCP_SKB_CB(skb)->seq - 1; 206ba34e6d9SEric Dumazet } else { 207ba34e6d9SEric Dumazet end_seq = TCP_SKB_CB(skb)->seq + 1; 2085b7ed089SYuchung Cheng } 209ba34e6d9SEric Dumazet } 210ba34e6d9SEric Dumazet tcp_rsk(req)->rcv_nxt = tp->rcv_nxt = end_seq; 2117656d842SEric Dumazet /* tcp_conn_request() is sending the SYNACK, 2127656d842SEric Dumazet * and queues the child into listener accept queue. 2137c85af88SEric Dumazet */ 2147c85af88SEric Dumazet return child; 2155b7ed089SYuchung Cheng } 2165b7ed089SYuchung Cheng 2175b7ed089SYuchung Cheng static bool tcp_fastopen_queue_check(struct sock *sk) 2185b7ed089SYuchung Cheng { 2195b7ed089SYuchung Cheng struct fastopen_queue *fastopenq; 2205b7ed089SYuchung Cheng 2215b7ed089SYuchung Cheng /* Make sure the listener has enabled fastopen, and we don't 2225b7ed089SYuchung Cheng * exceed the max # of pending TFO requests allowed before trying 2235b7ed089SYuchung Cheng * to validating the cookie in order to avoid burning CPU cycles 2245b7ed089SYuchung Cheng * unnecessarily. 2255b7ed089SYuchung Cheng * 2265b7ed089SYuchung Cheng * XXX (TFO) - The implication of checking the max_qlen before 2275b7ed089SYuchung Cheng * processing a cookie request is that clients can't differentiate 2285b7ed089SYuchung Cheng * between qlen overflow causing Fast Open to be disabled 2295b7ed089SYuchung Cheng * temporarily vs a server not supporting Fast Open at all. 2305b7ed089SYuchung Cheng */ 2310536fcc0SEric Dumazet fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq; 2320536fcc0SEric Dumazet if (fastopenq->max_qlen == 0) 2335b7ed089SYuchung Cheng return false; 2345b7ed089SYuchung Cheng 2355b7ed089SYuchung Cheng if (fastopenq->qlen >= fastopenq->max_qlen) { 2365b7ed089SYuchung Cheng struct request_sock *req1; 2375b7ed089SYuchung Cheng spin_lock(&fastopenq->lock); 2385b7ed089SYuchung Cheng req1 = fastopenq->rskq_rst_head; 239fa76ce73SEric Dumazet if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) { 2405b7ed089SYuchung Cheng spin_unlock(&fastopenq->lock); 2415b7ed089SYuchung Cheng NET_INC_STATS_BH(sock_net(sk), 2425b7ed089SYuchung Cheng LINUX_MIB_TCPFASTOPENLISTENOVERFLOW); 2435b7ed089SYuchung Cheng return false; 2445b7ed089SYuchung Cheng } 2455b7ed089SYuchung Cheng fastopenq->rskq_rst_head = req1->dl_next; 2465b7ed089SYuchung Cheng fastopenq->qlen--; 2475b7ed089SYuchung Cheng spin_unlock(&fastopenq->lock); 24813854e5aSEric Dumazet reqsk_put(req1); 2495b7ed089SYuchung Cheng } 2505b7ed089SYuchung Cheng return true; 2515b7ed089SYuchung Cheng } 2525b7ed089SYuchung Cheng 25389278c9dSYuchung Cheng /* Returns true if we should perform Fast Open on the SYN. The cookie (foc) 25489278c9dSYuchung Cheng * may be updated and return the client in the SYN-ACK later. E.g., Fast Open 25589278c9dSYuchung Cheng * cookie request (foc->len == 0). 25689278c9dSYuchung Cheng */ 2577c85af88SEric Dumazet struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb, 2585b7ed089SYuchung Cheng struct request_sock *req, 259843f4a55SYuchung Cheng struct tcp_fastopen_cookie *foc, 260843f4a55SYuchung Cheng struct dst_entry *dst) 2615b7ed089SYuchung Cheng { 26289278c9dSYuchung Cheng struct tcp_fastopen_cookie valid_foc = { .len = -1 }; 26389278c9dSYuchung Cheng bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1; 2647c85af88SEric Dumazet struct sock *child; 2655b7ed089SYuchung Cheng 266531c94a9SYuchung Cheng if (foc->len == 0) /* Client requests a cookie */ 267531c94a9SYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD); 268531c94a9SYuchung Cheng 26989278c9dSYuchung Cheng if (!((sysctl_tcp_fastopen & TFO_SERVER_ENABLE) && 27089278c9dSYuchung Cheng (syn_data || foc->len >= 0) && 27189278c9dSYuchung Cheng tcp_fastopen_queue_check(sk))) { 27289278c9dSYuchung Cheng foc->len = -1; 2737c85af88SEric Dumazet return NULL; 2745b7ed089SYuchung Cheng } 27589278c9dSYuchung Cheng 27689278c9dSYuchung Cheng if (syn_data && (sysctl_tcp_fastopen & TFO_SERVER_COOKIE_NOT_REQD)) 27789278c9dSYuchung Cheng goto fastopen; 27889278c9dSYuchung Cheng 279531c94a9SYuchung Cheng if (foc->len >= 0 && /* Client presents or requests a cookie */ 280531c94a9SYuchung Cheng tcp_fastopen_cookie_gen(req, skb, &valid_foc) && 2813a19ce0eSDaniel Lee foc->len == TCP_FASTOPEN_COOKIE_SIZE && 28289278c9dSYuchung Cheng foc->len == valid_foc.len && 28389278c9dSYuchung Cheng !memcmp(foc->val, valid_foc.val, foc->len)) { 284843f4a55SYuchung Cheng /* Cookie is valid. Create a (full) child socket to accept 285843f4a55SYuchung Cheng * the data in SYN before returning a SYN-ACK to ack the 286843f4a55SYuchung Cheng * data. If we fail to create the socket, fall back and 287843f4a55SYuchung Cheng * ack the ISN only but includes the same cookie. 288843f4a55SYuchung Cheng * 289843f4a55SYuchung Cheng * Note: Data-less SYN with valid cookie is allowed to send 290843f4a55SYuchung Cheng * data in SYN_RECV state. 291843f4a55SYuchung Cheng */ 29289278c9dSYuchung Cheng fastopen: 2937c85af88SEric Dumazet child = tcp_fastopen_create_child(sk, skb, dst, req); 2947c85af88SEric Dumazet if (child) { 29589278c9dSYuchung Cheng foc->len = -1; 296843f4a55SYuchung Cheng NET_INC_STATS_BH(sock_net(sk), 297843f4a55SYuchung Cheng LINUX_MIB_TCPFASTOPENPASSIVE); 2987c85af88SEric Dumazet return child; 2995b7ed089SYuchung Cheng } 300531c94a9SYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL); 301531c94a9SYuchung Cheng } else if (foc->len > 0) /* Client presents an invalid cookie */ 302531c94a9SYuchung Cheng NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPFASTOPENPASSIVEFAIL); 3035b7ed089SYuchung Cheng 3047f9b838bSDaniel Lee valid_foc.exp = foc->exp; 30589278c9dSYuchung Cheng *foc = valid_foc; 3067c85af88SEric Dumazet return NULL; 3075b7ed089SYuchung Cheng } 308