1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/crypto.h>
4 #include <linux/init.h>
5 #include <linux/kernel.h>
6 #include <linux/list.h>
8 #include <linux/rcupdate.h>
9 #include <linux/rculist.h>
10 #include <net/inetpeer.h>
13 void tcp_fastopen_init_key_once(struct net *net)
15 u8 key[TCP_FASTOPEN_KEY_LENGTH];
16 struct tcp_fastopen_context *ctxt;
19 ctxt = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
26 /* tcp_fastopen_reset_cipher publishes the new context
27 * atomically, so we allow this race happening here.
29 * All call sites of tcp_fastopen_cookie_gen also check
30 * for a valid cookie, so this is an acceptable risk.
32 get_random_bytes(key, sizeof(key));
33 tcp_fastopen_reset_cipher(net, NULL, key, NULL);
36 static void tcp_fastopen_ctx_free(struct rcu_head *head)
38 struct tcp_fastopen_context *ctx =
39 container_of(head, struct tcp_fastopen_context, rcu);
44 void tcp_fastopen_destroy_cipher(struct sock *sk)
46 struct tcp_fastopen_context *ctx;
48 ctx = rcu_dereference_protected(
49 inet_csk(sk)->icsk_accept_queue.fastopenq.ctx, 1);
51 call_rcu(&ctx->rcu, tcp_fastopen_ctx_free);
54 void tcp_fastopen_ctx_destroy(struct net *net)
56 struct tcp_fastopen_context *ctxt;
58 spin_lock(&net->ipv4.tcp_fastopen_ctx_lock);
60 ctxt = rcu_dereference_protected(net->ipv4.tcp_fastopen_ctx,
61 lockdep_is_held(&net->ipv4.tcp_fastopen_ctx_lock));
62 rcu_assign_pointer(net->ipv4.tcp_fastopen_ctx, NULL);
63 spin_unlock(&net->ipv4.tcp_fastopen_ctx_lock);
66 call_rcu(&ctxt->rcu, tcp_fastopen_ctx_free);
69 int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
70 void *primary_key, void *backup_key)
72 struct tcp_fastopen_context *ctx, *octx;
73 struct fastopen_queue *q;
76 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
82 ctx->key[0].key[0] = get_unaligned_le64(primary_key);
83 ctx->key[0].key[1] = get_unaligned_le64(primary_key + 8);
85 ctx->key[1].key[0] = get_unaligned_le64(backup_key);
86 ctx->key[1].key[1] = get_unaligned_le64(backup_key + 8);
92 spin_lock(&net->ipv4.tcp_fastopen_ctx_lock);
94 q = &inet_csk(sk)->icsk_accept_queue.fastopenq;
95 octx = rcu_dereference_protected(q->ctx,
96 lockdep_is_held(&net->ipv4.tcp_fastopen_ctx_lock));
97 rcu_assign_pointer(q->ctx, ctx);
99 octx = rcu_dereference_protected(net->ipv4.tcp_fastopen_ctx,
100 lockdep_is_held(&net->ipv4.tcp_fastopen_ctx_lock));
101 rcu_assign_pointer(net->ipv4.tcp_fastopen_ctx, ctx);
103 spin_unlock(&net->ipv4.tcp_fastopen_ctx_lock);
106 call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
111 static bool __tcp_fastopen_cookie_gen_cipher(struct request_sock *req,
113 const siphash_key_t *key,
114 struct tcp_fastopen_cookie *foc)
116 BUILD_BUG_ON(TCP_FASTOPEN_COOKIE_SIZE != sizeof(u64));
118 if (req->rsk_ops->family == AF_INET) {
119 const struct iphdr *iph = ip_hdr(syn);
121 foc->val[0] = cpu_to_le64(siphash(&iph->saddr,
125 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
128 #if IS_ENABLED(CONFIG_IPV6)
129 if (req->rsk_ops->family == AF_INET6) {
130 const struct ipv6hdr *ip6h = ipv6_hdr(syn);
132 foc->val[0] = cpu_to_le64(siphash(&ip6h->saddr,
133 sizeof(ip6h->saddr) +
136 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
143 /* Generate the fastopen cookie by applying SipHash to both the source and
144 * destination addresses.
146 static void tcp_fastopen_cookie_gen(struct sock *sk,
147 struct request_sock *req,
149 struct tcp_fastopen_cookie *foc)
151 struct tcp_fastopen_context *ctx;
154 ctx = tcp_fastopen_get_ctx(sk);
156 __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[0], foc);
160 /* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
161 * queue this additional data / FIN.
163 void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
165 struct tcp_sock *tp = tcp_sk(sk);
167 if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
170 skb = skb_clone(skb, GFP_ATOMIC);
175 /* segs_in has been initialized to 1 in tcp_create_openreq_child().
176 * Hence, reset segs_in to 0 before calling tcp_segs_in()
177 * to avoid double counting. Also, tcp_segs_in() expects
178 * skb->len to include the tcp_hdrlen. Hence, it should
179 * be called before __skb_pull().
182 tcp_segs_in(tp, skb);
183 __skb_pull(skb, tcp_hdrlen(skb));
184 sk_forced_mem_schedule(sk, skb->truesize);
185 skb_set_owner_r(skb, sk);
187 TCP_SKB_CB(skb)->seq++;
188 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
190 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
191 __skb_queue_tail(&sk->sk_receive_queue, skb);
192 tp->syn_data_acked = 1;
194 /* u64_stats_update_begin(&tp->syncp) not needed here,
195 * as we certainly are not changing upper 32bit value (0)
197 tp->bytes_received = skb->len;
199 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
203 /* returns 0 - no key match, 1 for primary, 2 for backup */
204 static int tcp_fastopen_cookie_gen_check(struct sock *sk,
205 struct request_sock *req,
207 struct tcp_fastopen_cookie *orig,
208 struct tcp_fastopen_cookie *valid_foc)
210 struct tcp_fastopen_cookie search_foc = { .len = -1 };
211 struct tcp_fastopen_cookie *foc = valid_foc;
212 struct tcp_fastopen_context *ctx;
216 ctx = tcp_fastopen_get_ctx(sk);
219 for (i = 0; i < tcp_fastopen_context_len(ctx); i++) {
220 __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[i], foc);
221 if (tcp_fastopen_cookie_match(foc, orig)) {
232 static struct sock *tcp_fastopen_create_child(struct sock *sk,
234 struct request_sock *req)
237 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
241 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
246 spin_lock(&queue->fastopenq.lock);
247 queue->fastopenq.qlen++;
248 spin_unlock(&queue->fastopenq.lock);
250 /* Initialize the child socket. Have to fix some values to take
251 * into account the child is a Fast Open socket and is created
252 * only out of the bits carried in the SYN packet.
256 rcu_assign_pointer(tp->fastopen_rsk, req);
257 tcp_rsk(req)->tfo_listener = true;
259 /* RFC1323: The window in SYN & SYN/ACK segments is never
260 * scaled. So correct it appropriately.
262 tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
263 tp->max_window = tp->snd_wnd;
265 /* Activate the retrans timer so that SYNACK can be retransmitted.
266 * The request socket is not added to the ehash
267 * because it's been added to the accept queue directly.
269 inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
270 TCP_TIMEOUT_INIT, TCP_RTO_MAX);
272 refcount_set(&req->rsk_refcnt, 2);
274 /* Now finish processing the fastopen child socket. */
275 tcp_init_transfer(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB);
277 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
279 tcp_fastopen_add_skb(child, skb);
281 tcp_rsk(req)->rcv_nxt = tp->rcv_nxt;
282 tp->rcv_wup = tp->rcv_nxt;
283 /* tcp_conn_request() is sending the SYNACK,
284 * and queues the child into listener accept queue.
289 static bool tcp_fastopen_queue_check(struct sock *sk)
291 struct fastopen_queue *fastopenq;
293 /* Make sure the listener has enabled fastopen, and we don't
294 * exceed the max # of pending TFO requests allowed before trying
295 * to validating the cookie in order to avoid burning CPU cycles
298 * XXX (TFO) - The implication of checking the max_qlen before
299 * processing a cookie request is that clients can't differentiate
300 * between qlen overflow causing Fast Open to be disabled
301 * temporarily vs a server not supporting Fast Open at all.
303 fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
304 if (fastopenq->max_qlen == 0)
307 if (fastopenq->qlen >= fastopenq->max_qlen) {
308 struct request_sock *req1;
309 spin_lock(&fastopenq->lock);
310 req1 = fastopenq->rskq_rst_head;
311 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
312 __NET_INC_STATS(sock_net(sk),
313 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
314 spin_unlock(&fastopenq->lock);
317 fastopenq->rskq_rst_head = req1->dl_next;
319 spin_unlock(&fastopenq->lock);
325 static bool tcp_fastopen_no_cookie(const struct sock *sk,
326 const struct dst_entry *dst,
329 return (sock_net(sk)->ipv4.sysctl_tcp_fastopen & flag) ||
330 tcp_sk(sk)->fastopen_no_cookie ||
331 (dst && dst_metric(dst, RTAX_FASTOPEN_NO_COOKIE));
334 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
335 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
336 * cookie request (foc->len == 0).
338 struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
339 struct request_sock *req,
340 struct tcp_fastopen_cookie *foc,
341 const struct dst_entry *dst)
343 bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
344 int tcp_fastopen = sock_net(sk)->ipv4.sysctl_tcp_fastopen;
345 struct tcp_fastopen_cookie valid_foc = { .len = -1 };
349 if (foc->len == 0) /* Client requests a cookie */
350 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
352 if (!((tcp_fastopen & TFO_SERVER_ENABLE) &&
353 (syn_data || foc->len >= 0) &&
354 tcp_fastopen_queue_check(sk))) {
360 tcp_fastopen_no_cookie(sk, dst, TFO_SERVER_COOKIE_NOT_REQD))
364 /* Client requests a cookie. */
365 tcp_fastopen_cookie_gen(sk, req, skb, &valid_foc);
366 } else if (foc->len > 0) {
367 ret = tcp_fastopen_cookie_gen_check(sk, req, skb, foc,
370 NET_INC_STATS(sock_net(sk),
371 LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
373 /* Cookie is valid. Create a (full) child socket to
374 * accept the data in SYN before returning a SYN-ACK to
375 * ack the data. If we fail to create the socket, fall
376 * back and ack the ISN only but includes the same
379 * Note: Data-less SYN with valid cookie is allowed to
380 * send data in SYN_RECV state.
383 child = tcp_fastopen_create_child(sk, skb, req);
386 valid_foc.exp = foc->exp;
388 NET_INC_STATS(sock_net(sk),
389 LINUX_MIB_TCPFASTOPENPASSIVEALTKEY);
393 NET_INC_STATS(sock_net(sk),
394 LINUX_MIB_TCPFASTOPENPASSIVE);
397 NET_INC_STATS(sock_net(sk),
398 LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
401 valid_foc.exp = foc->exp;
406 bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
407 struct tcp_fastopen_cookie *cookie)
409 const struct dst_entry *dst;
411 tcp_fastopen_cache_get(sk, mss, cookie);
413 /* Firewall blackhole issue check */
414 if (tcp_fastopen_active_should_disable(sk)) {
419 dst = __sk_dst_get(sk);
421 if (tcp_fastopen_no_cookie(sk, dst, TFO_CLIENT_NO_COOKIE)) {
427 tcp_sk(sk)->fastopen_client_fail = TFO_COOKIE_UNAVAILABLE;
431 /* This function checks if we want to defer sending SYN until the first
432 * write(). We defer under the following conditions:
433 * 1. fastopen_connect sockopt is set
434 * 2. we have a valid cookie
435 * Return value: return true if we want to defer until application writes data
436 * return false if we want to send out SYN immediately
438 bool tcp_fastopen_defer_connect(struct sock *sk, int *err)
440 struct tcp_fastopen_cookie cookie = { .len = 0 };
441 struct tcp_sock *tp = tcp_sk(sk);
444 if (tp->fastopen_connect && !tp->fastopen_req) {
445 if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) {
446 inet_sk(sk)->defer_connect = 1;
450 /* Alloc fastopen_req in order for FO option to be included
453 tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req),
455 if (tp->fastopen_req)
456 tp->fastopen_req->cookie = cookie;
462 EXPORT_SYMBOL(tcp_fastopen_defer_connect);
465 * The following code block is to deal with middle box issues with TFO:
466 * Middlebox firewall issues can potentially cause server's data being
467 * blackholed after a successful 3WHS using TFO.
468 * The proposed solution is to disable active TFO globally under the
469 * following circumstances:
470 * 1. client side TFO socket receives out of order FIN
471 * 2. client side TFO socket receives out of order RST
472 * 3. client side TFO socket has timed out three times consecutively during
474 * We disable active side TFO globally for 1hr at first. Then if it
475 * happens again, we disable it for 2h, then 4h, 8h, ...
476 * And we reset the timeout back to 1hr when we see a successful active
477 * TFO connection with data exchanges.
480 /* Disable active TFO and record current jiffies and
481 * tfo_active_disable_times
483 void tcp_fastopen_active_disable(struct sock *sk)
485 struct net *net = sock_net(sk);
487 atomic_inc(&net->ipv4.tfo_active_disable_times);
488 net->ipv4.tfo_active_disable_stamp = jiffies;
489 NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE);
492 /* Calculate timeout for tfo active disable
493 * Return true if we are still in the active TFO disable period
494 * Return false if timeout already expired and we should use active TFO
496 bool tcp_fastopen_active_should_disable(struct sock *sk)
498 unsigned int tfo_bh_timeout = sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout;
499 int tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times);
500 unsigned long timeout;
506 /* Limit timout to max: 2^6 * initial timeout */
507 multiplier = 1 << min(tfo_da_times - 1, 6);
508 timeout = multiplier * tfo_bh_timeout * HZ;
509 if (time_before(jiffies, sock_net(sk)->ipv4.tfo_active_disable_stamp + timeout))
512 /* Mark check bit so we can check for successful active TFO
513 * condition and reset tfo_active_disable_times
515 tcp_sk(sk)->syn_fastopen_ch = 1;
519 /* Disable active TFO if FIN is the only packet in the ofo queue
520 * and no data is received.
521 * Also check if we can reset tfo_active_disable_times if data is
522 * received successfully on a marked active TFO sockets opened on
523 * a non-loopback interface
525 void tcp_fastopen_active_disable_ofo_check(struct sock *sk)
527 struct tcp_sock *tp = tcp_sk(sk);
528 struct dst_entry *dst;
531 if (!tp->syn_fastopen)
534 if (!tp->data_segs_in) {
535 skb = skb_rb_first(&tp->out_of_order_queue);
536 if (skb && !skb_rb_next(skb)) {
537 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
538 tcp_fastopen_active_disable(sk);
542 } else if (tp->syn_fastopen_ch &&
543 atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) {
544 dst = sk_dst_get(sk);
545 if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK)))
546 atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0);
551 void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired)
553 u32 timeouts = inet_csk(sk)->icsk_retransmits;
554 struct tcp_sock *tp = tcp_sk(sk);
556 /* Broken middle-boxes may black-hole Fast Open connection during or
557 * even after the handshake. Be extremely conservative and pause
558 * Fast Open globally after hitting the third consecutive timeout or
559 * exceeding the configured timeout limit.
561 if ((tp->syn_fastopen || tp->syn_data || tp->syn_data_acked) &&
562 (timeouts == 2 || (timeouts < 2 && expired))) {
563 tcp_fastopen_active_disable(sk);
564 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);