1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
4 #include <linux/rcupdate.h>
7 void tcp_fastopen_init_key_once(struct net *net)
9 u8 key[TCP_FASTOPEN_KEY_LENGTH];
10 struct tcp_fastopen_context *ctxt;
13 ctxt = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
20 /* tcp_fastopen_reset_cipher publishes the new context
21 * atomically, so we allow this race happening here.
23 * All call sites of tcp_fastopen_cookie_gen also check
24 * for a valid cookie, so this is an acceptable risk.
26 get_random_bytes(key, sizeof(key));
27 tcp_fastopen_reset_cipher(net, NULL, key, NULL);
30 static void tcp_fastopen_ctx_free(struct rcu_head *head)
32 struct tcp_fastopen_context *ctx =
33 container_of(head, struct tcp_fastopen_context, rcu);
38 void tcp_fastopen_destroy_cipher(struct sock *sk)
40 struct tcp_fastopen_context *ctx;
42 ctx = rcu_dereference_protected(
43 inet_csk(sk)->icsk_accept_queue.fastopenq.ctx, 1);
45 call_rcu(&ctx->rcu, tcp_fastopen_ctx_free);
48 void tcp_fastopen_ctx_destroy(struct net *net)
50 struct tcp_fastopen_context *ctxt;
52 ctxt = xchg((__force struct tcp_fastopen_context **)&net->ipv4.tcp_fastopen_ctx, NULL);
55 call_rcu(&ctxt->rcu, tcp_fastopen_ctx_free);
58 int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
59 void *primary_key, void *backup_key)
61 struct tcp_fastopen_context *ctx, *octx;
62 struct fastopen_queue *q;
65 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
71 ctx->key[0].key[0] = get_unaligned_le64(primary_key);
72 ctx->key[0].key[1] = get_unaligned_le64(primary_key + 8);
74 ctx->key[1].key[0] = get_unaligned_le64(backup_key);
75 ctx->key[1].key[1] = get_unaligned_le64(backup_key + 8);
82 q = &inet_csk(sk)->icsk_accept_queue.fastopenq;
83 octx = xchg((__force struct tcp_fastopen_context **)&q->ctx, ctx);
85 octx = xchg((__force struct tcp_fastopen_context **)&net->ipv4.tcp_fastopen_ctx, ctx);
89 call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
94 int tcp_fastopen_get_cipher(struct net *net, struct inet_connection_sock *icsk,
97 struct tcp_fastopen_context *ctx;
102 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
104 ctx = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
106 n_keys = tcp_fastopen_context_len(ctx);
107 for (i = 0; i < n_keys; i++) {
108 put_unaligned_le64(ctx->key[i].key[0], key + (i * 2));
109 put_unaligned_le64(ctx->key[i].key[1], key + (i * 2) + 1);
117 static bool __tcp_fastopen_cookie_gen_cipher(struct request_sock *req,
119 const siphash_key_t *key,
120 struct tcp_fastopen_cookie *foc)
122 BUILD_BUG_ON(TCP_FASTOPEN_COOKIE_SIZE != sizeof(u64));
124 if (req->rsk_ops->family == AF_INET) {
125 const struct iphdr *iph = ip_hdr(syn);
127 foc->val[0] = cpu_to_le64(siphash(&iph->saddr,
131 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
134 #if IS_ENABLED(CONFIG_IPV6)
135 if (req->rsk_ops->family == AF_INET6) {
136 const struct ipv6hdr *ip6h = ipv6_hdr(syn);
138 foc->val[0] = cpu_to_le64(siphash(&ip6h->saddr,
139 sizeof(ip6h->saddr) +
142 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
149 /* Generate the fastopen cookie by applying SipHash to both the source and
150 * destination addresses.
152 static void tcp_fastopen_cookie_gen(struct sock *sk,
153 struct request_sock *req,
155 struct tcp_fastopen_cookie *foc)
157 struct tcp_fastopen_context *ctx;
160 ctx = tcp_fastopen_get_ctx(sk);
162 __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[0], foc);
166 /* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
167 * queue this additional data / FIN.
169 void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
171 struct tcp_sock *tp = tcp_sk(sk);
173 if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
176 skb = skb_clone(skb, GFP_ATOMIC);
181 /* segs_in has been initialized to 1 in tcp_create_openreq_child().
182 * Hence, reset segs_in to 0 before calling tcp_segs_in()
183 * to avoid double counting. Also, tcp_segs_in() expects
184 * skb->len to include the tcp_hdrlen. Hence, it should
185 * be called before __skb_pull().
188 tcp_segs_in(tp, skb);
189 __skb_pull(skb, tcp_hdrlen(skb));
190 sk_forced_mem_schedule(sk, skb->truesize);
191 skb_set_owner_r(skb, sk);
193 TCP_SKB_CB(skb)->seq++;
194 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
196 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
197 __skb_queue_tail(&sk->sk_receive_queue, skb);
198 tp->syn_data_acked = 1;
200 /* u64_stats_update_begin(&tp->syncp) not needed here,
201 * as we certainly are not changing upper 32bit value (0)
203 tp->bytes_received = skb->len;
205 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
209 /* returns 0 - no key match, 1 for primary, 2 for backup */
210 static int tcp_fastopen_cookie_gen_check(struct sock *sk,
211 struct request_sock *req,
213 struct tcp_fastopen_cookie *orig,
214 struct tcp_fastopen_cookie *valid_foc)
216 struct tcp_fastopen_cookie search_foc = { .len = -1 };
217 struct tcp_fastopen_cookie *foc = valid_foc;
218 struct tcp_fastopen_context *ctx;
222 ctx = tcp_fastopen_get_ctx(sk);
225 for (i = 0; i < tcp_fastopen_context_len(ctx); i++) {
226 __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[i], foc);
227 if (tcp_fastopen_cookie_match(foc, orig)) {
238 static struct sock *tcp_fastopen_create_child(struct sock *sk,
240 struct request_sock *req)
243 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
247 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
252 spin_lock(&queue->fastopenq.lock);
253 queue->fastopenq.qlen++;
254 spin_unlock(&queue->fastopenq.lock);
256 /* Initialize the child socket. Have to fix some values to take
257 * into account the child is a Fast Open socket and is created
258 * only out of the bits carried in the SYN packet.
262 rcu_assign_pointer(tp->fastopen_rsk, req);
263 tcp_rsk(req)->tfo_listener = true;
265 /* RFC1323: The window in SYN & SYN/ACK segments is never
266 * scaled. So correct it appropriately.
268 tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
269 tp->max_window = tp->snd_wnd;
271 /* Activate the retrans timer so that SYNACK can be retransmitted.
272 * The request socket is not added to the ehash
273 * because it's been added to the accept queue directly.
275 req->timeout = tcp_timeout_init(child);
276 inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
277 req->timeout, TCP_RTO_MAX);
279 refcount_set(&req->rsk_refcnt, 2);
281 /* Now finish processing the fastopen child socket. */
282 tcp_init_transfer(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, skb);
284 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
286 tcp_fastopen_add_skb(child, skb);
288 tcp_rsk(req)->rcv_nxt = tp->rcv_nxt;
289 tp->rcv_wup = tp->rcv_nxt;
290 /* tcp_conn_request() is sending the SYNACK,
291 * and queues the child into listener accept queue.
296 static bool tcp_fastopen_queue_check(struct sock *sk)
298 struct fastopen_queue *fastopenq;
300 /* Make sure the listener has enabled fastopen, and we don't
301 * exceed the max # of pending TFO requests allowed before trying
302 * to validating the cookie in order to avoid burning CPU cycles
305 * XXX (TFO) - The implication of checking the max_qlen before
306 * processing a cookie request is that clients can't differentiate
307 * between qlen overflow causing Fast Open to be disabled
308 * temporarily vs a server not supporting Fast Open at all.
310 fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
311 if (fastopenq->max_qlen == 0)
314 if (fastopenq->qlen >= fastopenq->max_qlen) {
315 struct request_sock *req1;
316 spin_lock(&fastopenq->lock);
317 req1 = fastopenq->rskq_rst_head;
318 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
319 __NET_INC_STATS(sock_net(sk),
320 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
321 spin_unlock(&fastopenq->lock);
324 fastopenq->rskq_rst_head = req1->dl_next;
326 spin_unlock(&fastopenq->lock);
332 static bool tcp_fastopen_no_cookie(const struct sock *sk,
333 const struct dst_entry *dst,
336 return (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) & flag) ||
337 tcp_sk(sk)->fastopen_no_cookie ||
338 (dst && dst_metric(dst, RTAX_FASTOPEN_NO_COOKIE));
341 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
342 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
343 * cookie request (foc->len == 0).
345 struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
346 struct request_sock *req,
347 struct tcp_fastopen_cookie *foc,
348 const struct dst_entry *dst)
350 bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
351 int tcp_fastopen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen);
352 struct tcp_fastopen_cookie valid_foc = { .len = -1 };
356 if (foc->len == 0) /* Client requests a cookie */
357 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
359 if (!((tcp_fastopen & TFO_SERVER_ENABLE) &&
360 (syn_data || foc->len >= 0) &&
361 tcp_fastopen_queue_check(sk))) {
366 if (tcp_fastopen_no_cookie(sk, dst, TFO_SERVER_COOKIE_NOT_REQD))
370 /* Client requests a cookie. */
371 tcp_fastopen_cookie_gen(sk, req, skb, &valid_foc);
372 } else if (foc->len > 0) {
373 ret = tcp_fastopen_cookie_gen_check(sk, req, skb, foc,
376 NET_INC_STATS(sock_net(sk),
377 LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
379 /* Cookie is valid. Create a (full) child socket to
380 * accept the data in SYN before returning a SYN-ACK to
381 * ack the data. If we fail to create the socket, fall
382 * back and ack the ISN only but includes the same
385 * Note: Data-less SYN with valid cookie is allowed to
386 * send data in SYN_RECV state.
389 child = tcp_fastopen_create_child(sk, skb, req);
392 valid_foc.exp = foc->exp;
394 NET_INC_STATS(sock_net(sk),
395 LINUX_MIB_TCPFASTOPENPASSIVEALTKEY);
399 NET_INC_STATS(sock_net(sk),
400 LINUX_MIB_TCPFASTOPENPASSIVE);
403 NET_INC_STATS(sock_net(sk),
404 LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
407 valid_foc.exp = foc->exp;
412 bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
413 struct tcp_fastopen_cookie *cookie)
415 const struct dst_entry *dst;
417 tcp_fastopen_cache_get(sk, mss, cookie);
419 /* Firewall blackhole issue check */
420 if (tcp_fastopen_active_should_disable(sk)) {
425 dst = __sk_dst_get(sk);
427 if (tcp_fastopen_no_cookie(sk, dst, TFO_CLIENT_NO_COOKIE)) {
433 tcp_sk(sk)->fastopen_client_fail = TFO_COOKIE_UNAVAILABLE;
437 /* This function checks if we want to defer sending SYN until the first
438 * write(). We defer under the following conditions:
439 * 1. fastopen_connect sockopt is set
440 * 2. we have a valid cookie
441 * Return value: return true if we want to defer until application writes data
442 * return false if we want to send out SYN immediately
444 bool tcp_fastopen_defer_connect(struct sock *sk, int *err)
446 struct tcp_fastopen_cookie cookie = { .len = 0 };
447 struct tcp_sock *tp = tcp_sk(sk);
450 if (tp->fastopen_connect && !tp->fastopen_req) {
451 if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) {
452 inet_sk(sk)->defer_connect = 1;
456 /* Alloc fastopen_req in order for FO option to be included
459 tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req),
461 if (tp->fastopen_req)
462 tp->fastopen_req->cookie = cookie;
468 EXPORT_SYMBOL(tcp_fastopen_defer_connect);
471 * The following code block is to deal with middle box issues with TFO:
472 * Middlebox firewall issues can potentially cause server's data being
473 * blackholed after a successful 3WHS using TFO.
474 * The proposed solution is to disable active TFO globally under the
475 * following circumstances:
476 * 1. client side TFO socket receives out of order FIN
477 * 2. client side TFO socket receives out of order RST
478 * 3. client side TFO socket has timed out three times consecutively during
480 * We disable active side TFO globally for 1hr at first. Then if it
481 * happens again, we disable it for 2h, then 4h, 8h, ...
482 * And we reset the timeout back to 1hr when we see a successful active
483 * TFO connection with data exchanges.
486 /* Disable active TFO and record current jiffies and
487 * tfo_active_disable_times
489 void tcp_fastopen_active_disable(struct sock *sk)
491 struct net *net = sock_net(sk);
493 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout))
496 /* Paired with READ_ONCE() in tcp_fastopen_active_should_disable() */
497 WRITE_ONCE(net->ipv4.tfo_active_disable_stamp, jiffies);
499 /* Paired with smp_rmb() in tcp_fastopen_active_should_disable().
500 * We want net->ipv4.tfo_active_disable_stamp to be updated first.
502 smp_mb__before_atomic();
503 atomic_inc(&net->ipv4.tfo_active_disable_times);
505 NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE);
508 /* Calculate timeout for tfo active disable
509 * Return true if we are still in the active TFO disable period
510 * Return false if timeout already expired and we should use active TFO
512 bool tcp_fastopen_active_should_disable(struct sock *sk)
514 unsigned int tfo_bh_timeout =
515 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout);
516 unsigned long timeout;
523 tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times);
527 /* Paired with smp_mb__before_atomic() in tcp_fastopen_active_disable() */
530 /* Limit timeout to max: 2^6 * initial timeout */
531 multiplier = 1 << min(tfo_da_times - 1, 6);
533 /* Paired with the WRITE_ONCE() in tcp_fastopen_active_disable(). */
534 timeout = READ_ONCE(sock_net(sk)->ipv4.tfo_active_disable_stamp) +
535 multiplier * tfo_bh_timeout * HZ;
536 if (time_before(jiffies, timeout))
539 /* Mark check bit so we can check for successful active TFO
540 * condition and reset tfo_active_disable_times
542 tcp_sk(sk)->syn_fastopen_ch = 1;
546 /* Disable active TFO if FIN is the only packet in the ofo queue
547 * and no data is received.
548 * Also check if we can reset tfo_active_disable_times if data is
549 * received successfully on a marked active TFO sockets opened on
550 * a non-loopback interface
552 void tcp_fastopen_active_disable_ofo_check(struct sock *sk)
554 struct tcp_sock *tp = tcp_sk(sk);
555 struct dst_entry *dst;
558 if (!tp->syn_fastopen)
561 if (!tp->data_segs_in) {
562 skb = skb_rb_first(&tp->out_of_order_queue);
563 if (skb && !skb_rb_next(skb)) {
564 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
565 tcp_fastopen_active_disable(sk);
569 } else if (tp->syn_fastopen_ch &&
570 atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) {
571 dst = sk_dst_get(sk);
572 if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK)))
573 atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0);
578 void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired)
580 u32 timeouts = inet_csk(sk)->icsk_retransmits;
581 struct tcp_sock *tp = tcp_sk(sk);
583 /* Broken middle-boxes may black-hole Fast Open connection during or
584 * even after the handshake. Be extremely conservative and pause
585 * Fast Open globally after hitting the third consecutive timeout or
586 * exceeding the configured timeout limit.
588 if ((tp->syn_fastopen || tp->syn_data || tp->syn_data_acked) &&
589 (timeouts == 2 || (timeouts < 2 && expired))) {
590 tcp_fastopen_active_disable(sk);
591 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);