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 int tcp_fastopen_get_cipher(struct net *net, struct inet_connection_sock *icsk,
114 struct tcp_fastopen_context *ctx;
119 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
121 ctx = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
123 n_keys = tcp_fastopen_context_len(ctx);
124 for (i = 0; i < n_keys; i++) {
125 put_unaligned_le64(ctx->key[i].key[0], key + (i * 2));
126 put_unaligned_le64(ctx->key[i].key[1], key + (i * 2) + 1);
134 static bool __tcp_fastopen_cookie_gen_cipher(struct request_sock *req,
136 const siphash_key_t *key,
137 struct tcp_fastopen_cookie *foc)
139 BUILD_BUG_ON(TCP_FASTOPEN_COOKIE_SIZE != sizeof(u64));
141 if (req->rsk_ops->family == AF_INET) {
142 const struct iphdr *iph = ip_hdr(syn);
144 foc->val[0] = cpu_to_le64(siphash(&iph->saddr,
148 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
151 #if IS_ENABLED(CONFIG_IPV6)
152 if (req->rsk_ops->family == AF_INET6) {
153 const struct ipv6hdr *ip6h = ipv6_hdr(syn);
155 foc->val[0] = cpu_to_le64(siphash(&ip6h->saddr,
156 sizeof(ip6h->saddr) +
159 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
166 /* Generate the fastopen cookie by applying SipHash to both the source and
167 * destination addresses.
169 static void tcp_fastopen_cookie_gen(struct sock *sk,
170 struct request_sock *req,
172 struct tcp_fastopen_cookie *foc)
174 struct tcp_fastopen_context *ctx;
177 ctx = tcp_fastopen_get_ctx(sk);
179 __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[0], foc);
183 /* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
184 * queue this additional data / FIN.
186 void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
188 struct tcp_sock *tp = tcp_sk(sk);
190 if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
193 skb = skb_clone(skb, GFP_ATOMIC);
198 /* segs_in has been initialized to 1 in tcp_create_openreq_child().
199 * Hence, reset segs_in to 0 before calling tcp_segs_in()
200 * to avoid double counting. Also, tcp_segs_in() expects
201 * skb->len to include the tcp_hdrlen. Hence, it should
202 * be called before __skb_pull().
205 tcp_segs_in(tp, skb);
206 __skb_pull(skb, tcp_hdrlen(skb));
207 sk_forced_mem_schedule(sk, skb->truesize);
208 skb_set_owner_r(skb, sk);
210 TCP_SKB_CB(skb)->seq++;
211 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
213 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
214 __skb_queue_tail(&sk->sk_receive_queue, skb);
215 tp->syn_data_acked = 1;
217 /* u64_stats_update_begin(&tp->syncp) not needed here,
218 * as we certainly are not changing upper 32bit value (0)
220 tp->bytes_received = skb->len;
222 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
226 /* returns 0 - no key match, 1 for primary, 2 for backup */
227 static int tcp_fastopen_cookie_gen_check(struct sock *sk,
228 struct request_sock *req,
230 struct tcp_fastopen_cookie *orig,
231 struct tcp_fastopen_cookie *valid_foc)
233 struct tcp_fastopen_cookie search_foc = { .len = -1 };
234 struct tcp_fastopen_cookie *foc = valid_foc;
235 struct tcp_fastopen_context *ctx;
239 ctx = tcp_fastopen_get_ctx(sk);
242 for (i = 0; i < tcp_fastopen_context_len(ctx); i++) {
243 __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[i], foc);
244 if (tcp_fastopen_cookie_match(foc, orig)) {
255 static struct sock *tcp_fastopen_create_child(struct sock *sk,
257 struct request_sock *req)
260 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
264 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
269 spin_lock(&queue->fastopenq.lock);
270 queue->fastopenq.qlen++;
271 spin_unlock(&queue->fastopenq.lock);
273 /* Initialize the child socket. Have to fix some values to take
274 * into account the child is a Fast Open socket and is created
275 * only out of the bits carried in the SYN packet.
279 rcu_assign_pointer(tp->fastopen_rsk, req);
280 tcp_rsk(req)->tfo_listener = true;
282 /* RFC1323: The window in SYN & SYN/ACK segments is never
283 * scaled. So correct it appropriately.
285 tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
286 tp->max_window = tp->snd_wnd;
288 /* Activate the retrans timer so that SYNACK can be retransmitted.
289 * The request socket is not added to the ehash
290 * because it's been added to the accept queue directly.
292 inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
293 TCP_TIMEOUT_INIT, TCP_RTO_MAX);
295 refcount_set(&req->rsk_refcnt, 2);
297 /* Now finish processing the fastopen child socket. */
298 tcp_init_transfer(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, skb);
300 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
302 tcp_fastopen_add_skb(child, skb);
304 tcp_rsk(req)->rcv_nxt = tp->rcv_nxt;
305 tp->rcv_wup = tp->rcv_nxt;
306 /* tcp_conn_request() is sending the SYNACK,
307 * and queues the child into listener accept queue.
312 static bool tcp_fastopen_queue_check(struct sock *sk)
314 struct fastopen_queue *fastopenq;
316 /* Make sure the listener has enabled fastopen, and we don't
317 * exceed the max # of pending TFO requests allowed before trying
318 * to validating the cookie in order to avoid burning CPU cycles
321 * XXX (TFO) - The implication of checking the max_qlen before
322 * processing a cookie request is that clients can't differentiate
323 * between qlen overflow causing Fast Open to be disabled
324 * temporarily vs a server not supporting Fast Open at all.
326 fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
327 if (fastopenq->max_qlen == 0)
330 if (fastopenq->qlen >= fastopenq->max_qlen) {
331 struct request_sock *req1;
332 spin_lock(&fastopenq->lock);
333 req1 = fastopenq->rskq_rst_head;
334 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
335 __NET_INC_STATS(sock_net(sk),
336 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
337 spin_unlock(&fastopenq->lock);
340 fastopenq->rskq_rst_head = req1->dl_next;
342 spin_unlock(&fastopenq->lock);
348 static bool tcp_fastopen_no_cookie(const struct sock *sk,
349 const struct dst_entry *dst,
352 return (sock_net(sk)->ipv4.sysctl_tcp_fastopen & flag) ||
353 tcp_sk(sk)->fastopen_no_cookie ||
354 (dst && dst_metric(dst, RTAX_FASTOPEN_NO_COOKIE));
357 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
358 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
359 * cookie request (foc->len == 0).
361 struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
362 struct request_sock *req,
363 struct tcp_fastopen_cookie *foc,
364 const struct dst_entry *dst)
366 bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
367 int tcp_fastopen = sock_net(sk)->ipv4.sysctl_tcp_fastopen;
368 struct tcp_fastopen_cookie valid_foc = { .len = -1 };
372 if (foc->len == 0) /* Client requests a cookie */
373 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
375 if (!((tcp_fastopen & TFO_SERVER_ENABLE) &&
376 (syn_data || foc->len >= 0) &&
377 tcp_fastopen_queue_check(sk))) {
383 tcp_fastopen_no_cookie(sk, dst, TFO_SERVER_COOKIE_NOT_REQD))
387 /* Client requests a cookie. */
388 tcp_fastopen_cookie_gen(sk, req, skb, &valid_foc);
389 } else if (foc->len > 0) {
390 ret = tcp_fastopen_cookie_gen_check(sk, req, skb, foc,
393 NET_INC_STATS(sock_net(sk),
394 LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
396 /* Cookie is valid. Create a (full) child socket to
397 * accept the data in SYN before returning a SYN-ACK to
398 * ack the data. If we fail to create the socket, fall
399 * back and ack the ISN only but includes the same
402 * Note: Data-less SYN with valid cookie is allowed to
403 * send data in SYN_RECV state.
406 child = tcp_fastopen_create_child(sk, skb, req);
409 valid_foc.exp = foc->exp;
411 NET_INC_STATS(sock_net(sk),
412 LINUX_MIB_TCPFASTOPENPASSIVEALTKEY);
416 NET_INC_STATS(sock_net(sk),
417 LINUX_MIB_TCPFASTOPENPASSIVE);
420 NET_INC_STATS(sock_net(sk),
421 LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
424 valid_foc.exp = foc->exp;
429 bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
430 struct tcp_fastopen_cookie *cookie)
432 const struct dst_entry *dst;
434 tcp_fastopen_cache_get(sk, mss, cookie);
436 /* Firewall blackhole issue check */
437 if (tcp_fastopen_active_should_disable(sk)) {
442 dst = __sk_dst_get(sk);
444 if (tcp_fastopen_no_cookie(sk, dst, TFO_CLIENT_NO_COOKIE)) {
450 tcp_sk(sk)->fastopen_client_fail = TFO_COOKIE_UNAVAILABLE;
454 /* This function checks if we want to defer sending SYN until the first
455 * write(). We defer under the following conditions:
456 * 1. fastopen_connect sockopt is set
457 * 2. we have a valid cookie
458 * Return value: return true if we want to defer until application writes data
459 * return false if we want to send out SYN immediately
461 bool tcp_fastopen_defer_connect(struct sock *sk, int *err)
463 struct tcp_fastopen_cookie cookie = { .len = 0 };
464 struct tcp_sock *tp = tcp_sk(sk);
467 if (tp->fastopen_connect && !tp->fastopen_req) {
468 if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) {
469 inet_sk(sk)->defer_connect = 1;
473 /* Alloc fastopen_req in order for FO option to be included
476 tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req),
478 if (tp->fastopen_req)
479 tp->fastopen_req->cookie = cookie;
485 EXPORT_SYMBOL(tcp_fastopen_defer_connect);
488 * The following code block is to deal with middle box issues with TFO:
489 * Middlebox firewall issues can potentially cause server's data being
490 * blackholed after a successful 3WHS using TFO.
491 * The proposed solution is to disable active TFO globally under the
492 * following circumstances:
493 * 1. client side TFO socket receives out of order FIN
494 * 2. client side TFO socket receives out of order RST
495 * 3. client side TFO socket has timed out three times consecutively during
497 * We disable active side TFO globally for 1hr at first. Then if it
498 * happens again, we disable it for 2h, then 4h, 8h, ...
499 * And we reset the timeout back to 1hr when we see a successful active
500 * TFO connection with data exchanges.
503 /* Disable active TFO and record current jiffies and
504 * tfo_active_disable_times
506 void tcp_fastopen_active_disable(struct sock *sk)
508 struct net *net = sock_net(sk);
510 if (!sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout)
513 /* Paired with READ_ONCE() in tcp_fastopen_active_should_disable() */
514 WRITE_ONCE(net->ipv4.tfo_active_disable_stamp, jiffies);
516 /* Paired with smp_rmb() in tcp_fastopen_active_should_disable().
517 * We want net->ipv4.tfo_active_disable_stamp to be updated first.
519 smp_mb__before_atomic();
520 atomic_inc(&net->ipv4.tfo_active_disable_times);
522 NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE);
525 /* Calculate timeout for tfo active disable
526 * Return true if we are still in the active TFO disable period
527 * Return false if timeout already expired and we should use active TFO
529 bool tcp_fastopen_active_should_disable(struct sock *sk)
531 unsigned int tfo_bh_timeout = sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout;
532 unsigned long timeout;
539 tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times);
543 /* Paired with smp_mb__before_atomic() in tcp_fastopen_active_disable() */
546 /* Limit timout to max: 2^6 * initial timeout */
547 multiplier = 1 << min(tfo_da_times - 1, 6);
549 /* Paired with the WRITE_ONCE() in tcp_fastopen_active_disable(). */
550 timeout = READ_ONCE(sock_net(sk)->ipv4.tfo_active_disable_stamp) +
551 multiplier * tfo_bh_timeout * HZ;
552 if (time_before(jiffies, timeout))
555 /* Mark check bit so we can check for successful active TFO
556 * condition and reset tfo_active_disable_times
558 tcp_sk(sk)->syn_fastopen_ch = 1;
562 /* Disable active TFO if FIN is the only packet in the ofo queue
563 * and no data is received.
564 * Also check if we can reset tfo_active_disable_times if data is
565 * received successfully on a marked active TFO sockets opened on
566 * a non-loopback interface
568 void tcp_fastopen_active_disable_ofo_check(struct sock *sk)
570 struct tcp_sock *tp = tcp_sk(sk);
571 struct dst_entry *dst;
574 if (!tp->syn_fastopen)
577 if (!tp->data_segs_in) {
578 skb = skb_rb_first(&tp->out_of_order_queue);
579 if (skb && !skb_rb_next(skb)) {
580 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
581 tcp_fastopen_active_disable(sk);
585 } else if (tp->syn_fastopen_ch &&
586 atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) {
587 dst = sk_dst_get(sk);
588 if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK)))
589 atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0);
594 void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired)
596 u32 timeouts = inet_csk(sk)->icsk_retransmits;
597 struct tcp_sock *tp = tcp_sk(sk);
599 /* Broken middle-boxes may black-hole Fast Open connection during or
600 * even after the handshake. Be extremely conservative and pause
601 * Fast Open globally after hitting the third consecutive timeout or
602 * exceeding the configured timeout limit.
604 if ((tp->syn_fastopen || tp->syn_data || tp->syn_data_acked) &&
605 (timeouts == 2 || (timeouts < 2 && expired))) {
606 tcp_fastopen_active_disable(sk);
607 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);