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 ctxt = xchg((__force struct tcp_fastopen_context **)&net->ipv4.tcp_fastopen_ctx, NULL);
61 call_rcu(&ctxt->rcu, tcp_fastopen_ctx_free);
64 int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
65 void *primary_key, void *backup_key)
67 struct tcp_fastopen_context *ctx, *octx;
68 struct fastopen_queue *q;
71 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
77 ctx->key[0].key[0] = get_unaligned_le64(primary_key);
78 ctx->key[0].key[1] = get_unaligned_le64(primary_key + 8);
80 ctx->key[1].key[0] = get_unaligned_le64(backup_key);
81 ctx->key[1].key[1] = get_unaligned_le64(backup_key + 8);
88 q = &inet_csk(sk)->icsk_accept_queue.fastopenq;
89 octx = xchg((__force struct tcp_fastopen_context **)&q->ctx, ctx);
91 octx = xchg((__force struct tcp_fastopen_context **)&net->ipv4.tcp_fastopen_ctx, ctx);
95 call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
100 int tcp_fastopen_get_cipher(struct net *net, struct inet_connection_sock *icsk,
103 struct tcp_fastopen_context *ctx;
108 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
110 ctx = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
112 n_keys = tcp_fastopen_context_len(ctx);
113 for (i = 0; i < n_keys; i++) {
114 put_unaligned_le64(ctx->key[i].key[0], key + (i * 2));
115 put_unaligned_le64(ctx->key[i].key[1], key + (i * 2) + 1);
123 static bool __tcp_fastopen_cookie_gen_cipher(struct request_sock *req,
125 const siphash_key_t *key,
126 struct tcp_fastopen_cookie *foc)
128 BUILD_BUG_ON(TCP_FASTOPEN_COOKIE_SIZE != sizeof(u64));
130 if (req->rsk_ops->family == AF_INET) {
131 const struct iphdr *iph = ip_hdr(syn);
133 foc->val[0] = cpu_to_le64(siphash(&iph->saddr,
137 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
140 #if IS_ENABLED(CONFIG_IPV6)
141 if (req->rsk_ops->family == AF_INET6) {
142 const struct ipv6hdr *ip6h = ipv6_hdr(syn);
144 foc->val[0] = cpu_to_le64(siphash(&ip6h->saddr,
145 sizeof(ip6h->saddr) +
148 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
155 /* Generate the fastopen cookie by applying SipHash to both the source and
156 * destination addresses.
158 static void tcp_fastopen_cookie_gen(struct sock *sk,
159 struct request_sock *req,
161 struct tcp_fastopen_cookie *foc)
163 struct tcp_fastopen_context *ctx;
166 ctx = tcp_fastopen_get_ctx(sk);
168 __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[0], foc);
172 /* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
173 * queue this additional data / FIN.
175 void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
177 struct tcp_sock *tp = tcp_sk(sk);
179 if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
182 skb = skb_clone(skb, GFP_ATOMIC);
187 /* segs_in has been initialized to 1 in tcp_create_openreq_child().
188 * Hence, reset segs_in to 0 before calling tcp_segs_in()
189 * to avoid double counting. Also, tcp_segs_in() expects
190 * skb->len to include the tcp_hdrlen. Hence, it should
191 * be called before __skb_pull().
194 tcp_segs_in(tp, skb);
195 __skb_pull(skb, tcp_hdrlen(skb));
196 sk_forced_mem_schedule(sk, skb->truesize);
197 skb_set_owner_r(skb, sk);
199 TCP_SKB_CB(skb)->seq++;
200 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
202 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
203 __skb_queue_tail(&sk->sk_receive_queue, skb);
204 tp->syn_data_acked = 1;
206 /* u64_stats_update_begin(&tp->syncp) not needed here,
207 * as we certainly are not changing upper 32bit value (0)
209 tp->bytes_received = skb->len;
211 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
215 /* returns 0 - no key match, 1 for primary, 2 for backup */
216 static int tcp_fastopen_cookie_gen_check(struct sock *sk,
217 struct request_sock *req,
219 struct tcp_fastopen_cookie *orig,
220 struct tcp_fastopen_cookie *valid_foc)
222 struct tcp_fastopen_cookie search_foc = { .len = -1 };
223 struct tcp_fastopen_cookie *foc = valid_foc;
224 struct tcp_fastopen_context *ctx;
228 ctx = tcp_fastopen_get_ctx(sk);
231 for (i = 0; i < tcp_fastopen_context_len(ctx); i++) {
232 __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[i], foc);
233 if (tcp_fastopen_cookie_match(foc, orig)) {
244 static struct sock *tcp_fastopen_create_child(struct sock *sk,
246 struct request_sock *req)
249 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
253 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
258 spin_lock(&queue->fastopenq.lock);
259 queue->fastopenq.qlen++;
260 spin_unlock(&queue->fastopenq.lock);
262 /* Initialize the child socket. Have to fix some values to take
263 * into account the child is a Fast Open socket and is created
264 * only out of the bits carried in the SYN packet.
268 rcu_assign_pointer(tp->fastopen_rsk, req);
269 tcp_rsk(req)->tfo_listener = true;
271 /* RFC1323: The window in SYN & SYN/ACK segments is never
272 * scaled. So correct it appropriately.
274 tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
275 tp->max_window = tp->snd_wnd;
277 /* Activate the retrans timer so that SYNACK can be retransmitted.
278 * The request socket is not added to the ehash
279 * because it's been added to the accept queue directly.
281 inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
282 TCP_TIMEOUT_INIT, TCP_RTO_MAX);
284 refcount_set(&req->rsk_refcnt, 2);
286 /* Now finish processing the fastopen child socket. */
287 tcp_init_transfer(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, skb);
289 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
291 tcp_fastopen_add_skb(child, skb);
293 tcp_rsk(req)->rcv_nxt = tp->rcv_nxt;
294 tp->rcv_wup = tp->rcv_nxt;
295 /* tcp_conn_request() is sending the SYNACK,
296 * and queues the child into listener accept queue.
301 static bool tcp_fastopen_queue_check(struct sock *sk)
303 struct fastopen_queue *fastopenq;
306 /* Make sure the listener has enabled fastopen, and we don't
307 * exceed the max # of pending TFO requests allowed before trying
308 * to validating the cookie in order to avoid burning CPU cycles
311 * XXX (TFO) - The implication of checking the max_qlen before
312 * processing a cookie request is that clients can't differentiate
313 * between qlen overflow causing Fast Open to be disabled
314 * temporarily vs a server not supporting Fast Open at all.
316 fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
317 max_qlen = READ_ONCE(fastopenq->max_qlen);
321 if (fastopenq->qlen >= max_qlen) {
322 struct request_sock *req1;
323 spin_lock(&fastopenq->lock);
324 req1 = fastopenq->rskq_rst_head;
325 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
326 __NET_INC_STATS(sock_net(sk),
327 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
328 spin_unlock(&fastopenq->lock);
331 fastopenq->rskq_rst_head = req1->dl_next;
333 spin_unlock(&fastopenq->lock);
339 static bool tcp_fastopen_no_cookie(const struct sock *sk,
340 const struct dst_entry *dst,
343 return (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) & flag) ||
344 tcp_sk(sk)->fastopen_no_cookie ||
345 (dst && dst_metric(dst, RTAX_FASTOPEN_NO_COOKIE));
348 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
349 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
350 * cookie request (foc->len == 0).
352 struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
353 struct request_sock *req,
354 struct tcp_fastopen_cookie *foc,
355 const struct dst_entry *dst)
357 bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
358 int tcp_fastopen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen);
359 struct tcp_fastopen_cookie valid_foc = { .len = -1 };
363 if (foc->len == 0) /* Client requests a cookie */
364 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
366 if (!((tcp_fastopen & TFO_SERVER_ENABLE) &&
367 (syn_data || foc->len >= 0) &&
368 tcp_fastopen_queue_check(sk))) {
373 if (tcp_fastopen_no_cookie(sk, dst, TFO_SERVER_COOKIE_NOT_REQD))
377 /* Client requests a cookie. */
378 tcp_fastopen_cookie_gen(sk, req, skb, &valid_foc);
379 } else if (foc->len > 0) {
380 ret = tcp_fastopen_cookie_gen_check(sk, req, skb, foc,
383 NET_INC_STATS(sock_net(sk),
384 LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
386 /* Cookie is valid. Create a (full) child socket to
387 * accept the data in SYN before returning a SYN-ACK to
388 * ack the data. If we fail to create the socket, fall
389 * back and ack the ISN only but includes the same
392 * Note: Data-less SYN with valid cookie is allowed to
393 * send data in SYN_RECV state.
396 child = tcp_fastopen_create_child(sk, skb, req);
399 valid_foc.exp = foc->exp;
401 NET_INC_STATS(sock_net(sk),
402 LINUX_MIB_TCPFASTOPENPASSIVEALTKEY);
406 NET_INC_STATS(sock_net(sk),
407 LINUX_MIB_TCPFASTOPENPASSIVE);
410 NET_INC_STATS(sock_net(sk),
411 LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
414 valid_foc.exp = foc->exp;
419 bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
420 struct tcp_fastopen_cookie *cookie)
422 const struct dst_entry *dst;
424 tcp_fastopen_cache_get(sk, mss, cookie);
426 /* Firewall blackhole issue check */
427 if (tcp_fastopen_active_should_disable(sk)) {
432 dst = __sk_dst_get(sk);
434 if (tcp_fastopen_no_cookie(sk, dst, TFO_CLIENT_NO_COOKIE)) {
440 tcp_sk(sk)->fastopen_client_fail = TFO_COOKIE_UNAVAILABLE;
444 /* This function checks if we want to defer sending SYN until the first
445 * write(). We defer under the following conditions:
446 * 1. fastopen_connect sockopt is set
447 * 2. we have a valid cookie
448 * Return value: return true if we want to defer until application writes data
449 * return false if we want to send out SYN immediately
451 bool tcp_fastopen_defer_connect(struct sock *sk, int *err)
453 struct tcp_fastopen_cookie cookie = { .len = 0 };
454 struct tcp_sock *tp = tcp_sk(sk);
457 if (tp->fastopen_connect && !tp->fastopen_req) {
458 if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) {
459 inet_sk(sk)->defer_connect = 1;
463 /* Alloc fastopen_req in order for FO option to be included
466 tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req),
468 if (tp->fastopen_req)
469 tp->fastopen_req->cookie = cookie;
475 EXPORT_SYMBOL(tcp_fastopen_defer_connect);
478 * The following code block is to deal with middle box issues with TFO:
479 * Middlebox firewall issues can potentially cause server's data being
480 * blackholed after a successful 3WHS using TFO.
481 * The proposed solution is to disable active TFO globally under the
482 * following circumstances:
483 * 1. client side TFO socket receives out of order FIN
484 * 2. client side TFO socket receives out of order RST
485 * 3. client side TFO socket has timed out three times consecutively during
487 * We disable active side TFO globally for 1hr at first. Then if it
488 * happens again, we disable it for 2h, then 4h, 8h, ...
489 * And we reset the timeout back to 1hr when we see a successful active
490 * TFO connection with data exchanges.
493 /* Disable active TFO and record current jiffies and
494 * tfo_active_disable_times
496 void tcp_fastopen_active_disable(struct sock *sk)
498 struct net *net = sock_net(sk);
500 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout))
503 /* Paired with READ_ONCE() in tcp_fastopen_active_should_disable() */
504 WRITE_ONCE(net->ipv4.tfo_active_disable_stamp, jiffies);
506 /* Paired with smp_rmb() in tcp_fastopen_active_should_disable().
507 * We want net->ipv4.tfo_active_disable_stamp to be updated first.
509 smp_mb__before_atomic();
510 atomic_inc(&net->ipv4.tfo_active_disable_times);
512 NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE);
515 /* Calculate timeout for tfo active disable
516 * Return true if we are still in the active TFO disable period
517 * Return false if timeout already expired and we should use active TFO
519 bool tcp_fastopen_active_should_disable(struct sock *sk)
521 unsigned int tfo_bh_timeout =
522 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout);
523 unsigned long timeout;
530 tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times);
534 /* Paired with smp_mb__before_atomic() in tcp_fastopen_active_disable() */
537 /* Limit timeout to max: 2^6 * initial timeout */
538 multiplier = 1 << min(tfo_da_times - 1, 6);
540 /* Paired with the WRITE_ONCE() in tcp_fastopen_active_disable(). */
541 timeout = READ_ONCE(sock_net(sk)->ipv4.tfo_active_disable_stamp) +
542 multiplier * tfo_bh_timeout * HZ;
543 if (time_before(jiffies, timeout))
546 /* Mark check bit so we can check for successful active TFO
547 * condition and reset tfo_active_disable_times
549 tcp_sk(sk)->syn_fastopen_ch = 1;
553 /* Disable active TFO if FIN is the only packet in the ofo queue
554 * and no data is received.
555 * Also check if we can reset tfo_active_disable_times if data is
556 * received successfully on a marked active TFO sockets opened on
557 * a non-loopback interface
559 void tcp_fastopen_active_disable_ofo_check(struct sock *sk)
561 struct tcp_sock *tp = tcp_sk(sk);
562 struct dst_entry *dst;
565 if (!tp->syn_fastopen)
568 if (!tp->data_segs_in) {
569 skb = skb_rb_first(&tp->out_of_order_queue);
570 if (skb && !skb_rb_next(skb)) {
571 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
572 tcp_fastopen_active_disable(sk);
576 } else if (tp->syn_fastopen_ch &&
577 atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) {
578 dst = sk_dst_get(sk);
579 if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK)))
580 atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0);
585 void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired)
587 u32 timeouts = inet_csk(sk)->icsk_retransmits;
588 struct tcp_sock *tp = tcp_sk(sk);
590 /* Broken middle-boxes may black-hole Fast Open connection during or
591 * even after the handshake. Be extremely conservative and pause
592 * Fast Open globally after hitting the third consecutive timeout or
593 * exceeding the configured timeout limit.
595 if ((tp->syn_fastopen || tp->syn_data || tp->syn_data_acked) &&
596 (timeouts == 2 || (timeouts < 2 && expired))) {
597 tcp_fastopen_active_disable(sk);
598 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);