1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/kernel.h>
3 #include <linux/tcp.h>
4 #include <linux/rcupdate.h>
5 #include <net/tcp.h>
6
7 void tcp_fastopen_init_key_once(struct net *net)
8 {
9 u8 key[TCP_FASTOPEN_KEY_LENGTH];
10 struct tcp_fastopen_context *ctxt;
11
12 rcu_read_lock();
13 ctxt = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
14 if (ctxt) {
15 rcu_read_unlock();
16 return;
17 }
18 rcu_read_unlock();
19
20 /* tcp_fastopen_reset_cipher publishes the new context
21 * atomically, so we allow this race happening here.
22 *
23 * All call sites of tcp_fastopen_cookie_gen also check
24 * for a valid cookie, so this is an acceptable risk.
25 */
26 get_random_bytes(key, sizeof(key));
27 tcp_fastopen_reset_cipher(net, NULL, key, NULL);
28 }
29
30 static void tcp_fastopen_ctx_free(struct rcu_head *head)
31 {
32 struct tcp_fastopen_context *ctx =
33 container_of(head, struct tcp_fastopen_context, rcu);
34
35 kfree_sensitive(ctx);
36 }
37
38 void tcp_fastopen_destroy_cipher(struct sock *sk)
39 {
40 struct tcp_fastopen_context *ctx;
41
42 ctx = rcu_dereference_protected(
43 inet_csk(sk)->icsk_accept_queue.fastopenq.ctx, 1);
44 if (ctx)
45 call_rcu(&ctx->rcu, tcp_fastopen_ctx_free);
46 }
47
48 void tcp_fastopen_ctx_destroy(struct net *net)
49 {
50 struct tcp_fastopen_context *ctxt;
51
52 ctxt = unrcu_pointer(xchg(&net->ipv4.tcp_fastopen_ctx, NULL));
53
54 if (ctxt)
55 call_rcu(&ctxt->rcu, tcp_fastopen_ctx_free);
56 }
57
58 int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
59 void *primary_key, void *backup_key)
60 {
61 struct tcp_fastopen_context *ctx, *octx;
62 struct fastopen_queue *q;
63 int err = 0;
64
65 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
66 if (!ctx) {
67 err = -ENOMEM;
68 goto out;
69 }
70
71 ctx->key[0].key[0] = get_unaligned_le64(primary_key);
72 ctx->key[0].key[1] = get_unaligned_le64(primary_key + 8);
73 if (backup_key) {
74 ctx->key[1].key[0] = get_unaligned_le64(backup_key);
75 ctx->key[1].key[1] = get_unaligned_le64(backup_key + 8);
76 ctx->num = 2;
77 } else {
78 ctx->num = 1;
79 }
80
81 if (sk) {
82 q = &inet_csk(sk)->icsk_accept_queue.fastopenq;
83 octx = unrcu_pointer(xchg(&q->ctx, RCU_INITIALIZER(ctx)));
84 } else {
85 octx = unrcu_pointer(xchg(&net->ipv4.tcp_fastopen_ctx,
86 RCU_INITIALIZER(ctx)));
87 }
88
89 if (octx)
90 call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
91 out:
92 return err;
93 }
94
95 int tcp_fastopen_get_cipher(struct net *net, struct inet_connection_sock *icsk,
96 u64 *key)
97 {
98 struct tcp_fastopen_context *ctx;
99 int n_keys = 0, i;
100
101 rcu_read_lock();
102 if (icsk)
103 ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
104 else
105 ctx = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
106 if (ctx) {
107 n_keys = tcp_fastopen_context_len(ctx);
108 for (i = 0; i < n_keys; i++) {
109 put_unaligned_le64(ctx->key[i].key[0], key + (i * 2));
110 put_unaligned_le64(ctx->key[i].key[1], key + (i * 2) + 1);
111 }
112 }
113 rcu_read_unlock();
114
115 return n_keys;
116 }
117
118 static bool __tcp_fastopen_cookie_gen_cipher(struct request_sock *req,
119 struct sk_buff *syn,
120 const siphash_key_t *key,
121 struct tcp_fastopen_cookie *foc)
122 {
123 BUILD_BUG_ON(TCP_FASTOPEN_COOKIE_SIZE != sizeof(u64));
124
125 if (req->rsk_ops->family == AF_INET) {
126 const struct iphdr *iph = ip_hdr(syn);
127
128 foc->val[0] = cpu_to_le64(siphash(&iph->saddr,
129 sizeof(iph->saddr) +
130 sizeof(iph->daddr),
131 key));
132 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
133 return true;
134 }
135 #if IS_ENABLED(CONFIG_IPV6)
136 if (req->rsk_ops->family == AF_INET6) {
137 const struct ipv6hdr *ip6h = ipv6_hdr(syn);
138
139 foc->val[0] = cpu_to_le64(siphash(&ip6h->saddr,
140 sizeof(ip6h->saddr) +
141 sizeof(ip6h->daddr),
142 key));
143 foc->len = TCP_FASTOPEN_COOKIE_SIZE;
144 return true;
145 }
146 #endif
147 return false;
148 }
149
150 /* Generate the fastopen cookie by applying SipHash to both the source and
151 * destination addresses.
152 */
153 static void tcp_fastopen_cookie_gen(struct sock *sk,
154 struct request_sock *req,
155 struct sk_buff *syn,
156 struct tcp_fastopen_cookie *foc)
157 {
158 struct tcp_fastopen_context *ctx;
159
160 rcu_read_lock();
161 ctx = tcp_fastopen_get_ctx(sk);
162 if (ctx)
163 __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[0], foc);
164 rcu_read_unlock();
165 }
166
167 /* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
168 * queue this additional data / FIN.
169 */
170 void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
171 {
172 struct tcp_sock *tp = tcp_sk(sk);
173
174 if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
175 return;
176
177 skb = skb_clone(skb, GFP_ATOMIC);
178 if (!skb)
179 return;
180
181 skb_dst_drop(skb);
182 /* segs_in has been initialized to 1 in tcp_create_openreq_child().
183 * Hence, reset segs_in to 0 before calling tcp_segs_in()
184 * to avoid double counting. Also, tcp_segs_in() expects
185 * skb->len to include the tcp_hdrlen. Hence, it should
186 * be called before __skb_pull().
187 */
188 tp->segs_in = 0;
189 tcp_segs_in(tp, skb);
190 __skb_pull(skb, tcp_hdrlen(skb));
191 sk_forced_mem_schedule(sk, skb->truesize);
192 skb_set_owner_r(skb, sk);
193
194 TCP_SKB_CB(skb)->seq++;
195 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
196
197 tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
198 __skb_queue_tail(&sk->sk_receive_queue, skb);
199 tp->syn_data_acked = 1;
200
201 /* u64_stats_update_begin(&tp->syncp) not needed here,
202 * as we certainly are not changing upper 32bit value (0)
203 */
204 tp->bytes_received = skb->len;
205
206 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
207 tcp_fin(sk);
208 }
209
210 /* returns 0 - no key match, 1 for primary, 2 for backup */
211 static int tcp_fastopen_cookie_gen_check(struct sock *sk,
212 struct request_sock *req,
213 struct sk_buff *syn,
214 struct tcp_fastopen_cookie *orig,
215 struct tcp_fastopen_cookie *valid_foc)
216 {
217 struct tcp_fastopen_cookie search_foc = { .len = -1 };
218 struct tcp_fastopen_cookie *foc = valid_foc;
219 struct tcp_fastopen_context *ctx;
220 int i, ret = 0;
221
222 rcu_read_lock();
223 ctx = tcp_fastopen_get_ctx(sk);
224 if (!ctx)
225 goto out;
226 for (i = 0; i < tcp_fastopen_context_len(ctx); i++) {
227 __tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[i], foc);
228 if (tcp_fastopen_cookie_match(foc, orig)) {
229 ret = i + 1;
230 goto out;
231 }
232 foc = &search_foc;
233 }
234 out:
235 rcu_read_unlock();
236 return ret;
237 }
238
239 static struct sock *tcp_fastopen_create_child(struct sock *sk,
240 struct sk_buff *skb,
241 struct request_sock *req)
242 {
243 struct tcp_sock *tp;
244 struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
245 struct sock *child;
246 bool own_req;
247
248 child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
249 NULL, &own_req);
250 if (!child)
251 return NULL;
252
253 spin_lock(&queue->fastopenq.lock);
254 queue->fastopenq.qlen++;
255 spin_unlock(&queue->fastopenq.lock);
256
257 /* Initialize the child socket. Have to fix some values to take
258 * into account the child is a Fast Open socket and is created
259 * only out of the bits carried in the SYN packet.
260 */
261 tp = tcp_sk(child);
262
263 rcu_assign_pointer(tp->fastopen_rsk, req);
264 tcp_rsk(req)->tfo_listener = true;
265
266 /* RFC1323: The window in SYN & SYN/ACK segments is never
267 * scaled. So correct it appropriately.
268 */
269 tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
270 tp->max_window = tp->snd_wnd;
271
272 /* Activate the retrans timer so that SYNACK can be retransmitted.
273 * The request socket is not added to the ehash
274 * because it's been added to the accept queue directly.
275 */
276 req->timeout = tcp_timeout_init(child);
277 inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
278 req->timeout, TCP_RTO_MAX);
279
280 refcount_set(&req->rsk_refcnt, 2);
281
282 /* Now finish processing the fastopen child socket. */
283 tcp_init_transfer(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, skb);
284
285 tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
286
287 tcp_fastopen_add_skb(child, skb);
288
289 tcp_rsk(req)->rcv_nxt = tp->rcv_nxt;
290 tp->rcv_wup = tp->rcv_nxt;
291 /* tcp_conn_request() is sending the SYNACK,
292 * and queues the child into listener accept queue.
293 */
294 return child;
295 }
296
297 static bool tcp_fastopen_queue_check(struct sock *sk)
298 {
299 struct fastopen_queue *fastopenq;
300 int max_qlen;
301
302 /* Make sure the listener has enabled fastopen, and we don't
303 * exceed the max # of pending TFO requests allowed before trying
304 * to validating the cookie in order to avoid burning CPU cycles
305 * unnecessarily.
306 *
307 * XXX (TFO) - The implication of checking the max_qlen before
308 * processing a cookie request is that clients can't differentiate
309 * between qlen overflow causing Fast Open to be disabled
310 * temporarily vs a server not supporting Fast Open at all.
311 */
312 fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
313 max_qlen = READ_ONCE(fastopenq->max_qlen);
314 if (max_qlen == 0)
315 return false;
316
317 if (fastopenq->qlen >= max_qlen) {
318 struct request_sock *req1;
319 spin_lock(&fastopenq->lock);
320 req1 = fastopenq->rskq_rst_head;
321 if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
322 __NET_INC_STATS(sock_net(sk),
323 LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
324 spin_unlock(&fastopenq->lock);
325 return false;
326 }
327 fastopenq->rskq_rst_head = req1->dl_next;
328 fastopenq->qlen--;
329 spin_unlock(&fastopenq->lock);
330 reqsk_put(req1);
331 }
332 return true;
333 }
334
335 static bool tcp_fastopen_no_cookie(const struct sock *sk,
336 const struct dst_entry *dst,
337 int flag)
338 {
339 return (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) & flag) ||
340 tcp_sk(sk)->fastopen_no_cookie ||
341 (dst && dst_metric(dst, RTAX_FASTOPEN_NO_COOKIE));
342 }
343
344 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
345 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
346 * cookie request (foc->len == 0).
347 */
348 struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
349 struct request_sock *req,
350 struct tcp_fastopen_cookie *foc,
351 const struct dst_entry *dst)
352 {
353 bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
354 int tcp_fastopen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen);
355 struct tcp_fastopen_cookie valid_foc = { .len = -1 };
356 struct sock *child;
357 int ret = 0;
358
359 if (foc->len == 0) /* Client requests a cookie */
360 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
361
362 if (!((tcp_fastopen & TFO_SERVER_ENABLE) &&
363 (syn_data || foc->len >= 0) &&
364 tcp_fastopen_queue_check(sk))) {
365 foc->len = -1;
366 return NULL;
367 }
368
369 if (tcp_fastopen_no_cookie(sk, dst, TFO_SERVER_COOKIE_NOT_REQD))
370 goto fastopen;
371
372 if (foc->len == 0) {
373 /* Client requests a cookie. */
374 tcp_fastopen_cookie_gen(sk, req, skb, &valid_foc);
375 } else if (foc->len > 0) {
376 ret = tcp_fastopen_cookie_gen_check(sk, req, skb, foc,
377 &valid_foc);
378 if (!ret) {
379 NET_INC_STATS(sock_net(sk),
380 LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
381 } else {
382 /* Cookie is valid. Create a (full) child socket to
383 * accept the data in SYN before returning a SYN-ACK to
384 * ack the data. If we fail to create the socket, fall
385 * back and ack the ISN only but includes the same
386 * cookie.
387 *
388 * Note: Data-less SYN with valid cookie is allowed to
389 * send data in SYN_RECV state.
390 */
391 fastopen:
392 child = tcp_fastopen_create_child(sk, skb, req);
393 if (child) {
394 if (ret == 2) {
395 valid_foc.exp = foc->exp;
396 *foc = valid_foc;
397 NET_INC_STATS(sock_net(sk),
398 LINUX_MIB_TCPFASTOPENPASSIVEALTKEY);
399 } else {
400 foc->len = -1;
401 }
402 NET_INC_STATS(sock_net(sk),
403 LINUX_MIB_TCPFASTOPENPASSIVE);
404 return child;
405 }
406 NET_INC_STATS(sock_net(sk),
407 LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
408 }
409 }
410 valid_foc.exp = foc->exp;
411 *foc = valid_foc;
412 return NULL;
413 }
414
415 bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
416 struct tcp_fastopen_cookie *cookie)
417 {
418 const struct dst_entry *dst;
419
420 tcp_fastopen_cache_get(sk, mss, cookie);
421
422 /* Firewall blackhole issue check */
423 if (tcp_fastopen_active_should_disable(sk)) {
424 cookie->len = -1;
425 return false;
426 }
427
428 dst = __sk_dst_get(sk);
429
430 if (tcp_fastopen_no_cookie(sk, dst, TFO_CLIENT_NO_COOKIE)) {
431 cookie->len = -1;
432 return true;
433 }
434 if (cookie->len > 0)
435 return true;
436 tcp_sk(sk)->fastopen_client_fail = TFO_COOKIE_UNAVAILABLE;
437 return false;
438 }
439
440 /* This function checks if we want to defer sending SYN until the first
441 * write(). We defer under the following conditions:
442 * 1. fastopen_connect sockopt is set
443 * 2. we have a valid cookie
444 * Return value: return true if we want to defer until application writes data
445 * return false if we want to send out SYN immediately
446 */
447 bool tcp_fastopen_defer_connect(struct sock *sk, int *err)
448 {
449 struct tcp_fastopen_cookie cookie = { .len = 0 };
450 struct tcp_sock *tp = tcp_sk(sk);
451 u16 mss;
452
453 if (tp->fastopen_connect && !tp->fastopen_req) {
454 if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) {
455 inet_set_bit(DEFER_CONNECT, sk);
456 return true;
457 }
458
459 /* Alloc fastopen_req in order for FO option to be included
460 * in SYN
461 */
462 tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req),
463 sk->sk_allocation);
464 if (tp->fastopen_req)
465 tp->fastopen_req->cookie = cookie;
466 else
467 *err = -ENOBUFS;
468 }
469 return false;
470 }
471 EXPORT_SYMBOL(tcp_fastopen_defer_connect);
472
473 /*
474 * The following code block is to deal with middle box issues with TFO:
475 * Middlebox firewall issues can potentially cause server's data being
476 * blackholed after a successful 3WHS using TFO.
477 * The proposed solution is to disable active TFO globally under the
478 * following circumstances:
479 * 1. client side TFO socket receives out of order FIN
480 * 2. client side TFO socket receives out of order RST
481 * 3. client side TFO socket has timed out three times consecutively during
482 * or after handshake
483 * We disable active side TFO globally for 1hr at first. Then if it
484 * happens again, we disable it for 2h, then 4h, 8h, ...
485 * And we reset the timeout back to 1hr when we see a successful active
486 * TFO connection with data exchanges.
487 */
488
489 /* Disable active TFO and record current jiffies and
490 * tfo_active_disable_times
491 */
492 void tcp_fastopen_active_disable(struct sock *sk)
493 {
494 struct net *net = sock_net(sk);
495
496 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout))
497 return;
498
499 /* Paired with READ_ONCE() in tcp_fastopen_active_should_disable() */
500 WRITE_ONCE(net->ipv4.tfo_active_disable_stamp, jiffies);
501
502 /* Paired with smp_rmb() in tcp_fastopen_active_should_disable().
503 * We want net->ipv4.tfo_active_disable_stamp to be updated first.
504 */
505 smp_mb__before_atomic();
506 atomic_inc(&net->ipv4.tfo_active_disable_times);
507
508 NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE);
509 }
510
511 /* Calculate timeout for tfo active disable
512 * Return true if we are still in the active TFO disable period
513 * Return false if timeout already expired and we should use active TFO
514 */
515 bool tcp_fastopen_active_should_disable(struct sock *sk)
516 {
517 unsigned int tfo_bh_timeout =
518 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout);
519 unsigned long timeout;
520 int tfo_da_times;
521 int multiplier;
522
523 if (!tfo_bh_timeout)
524 return false;
525
526 tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times);
527 if (!tfo_da_times)
528 return false;
529
530 /* Paired with smp_mb__before_atomic() in tcp_fastopen_active_disable() */
531 smp_rmb();
532
533 /* Limit timeout to max: 2^6 * initial timeout */
534 multiplier = 1 << min(tfo_da_times - 1, 6);
535
536 /* Paired with the WRITE_ONCE() in tcp_fastopen_active_disable(). */
537 timeout = READ_ONCE(sock_net(sk)->ipv4.tfo_active_disable_stamp) +
538 multiplier * tfo_bh_timeout * HZ;
539 if (time_before(jiffies, timeout))
540 return true;
541
542 /* Mark check bit so we can check for successful active TFO
543 * condition and reset tfo_active_disable_times
544 */
545 tcp_sk(sk)->syn_fastopen_ch = 1;
546 return false;
547 }
548
549 /* Disable active TFO if FIN is the only packet in the ofo queue
550 * and no data is received.
551 * Also check if we can reset tfo_active_disable_times if data is
552 * received successfully on a marked active TFO sockets opened on
553 * a non-loopback interface
554 */
555 void tcp_fastopen_active_disable_ofo_check(struct sock *sk)
556 {
557 struct tcp_sock *tp = tcp_sk(sk);
558 struct dst_entry *dst;
559 struct sk_buff *skb;
560
561 if (!tp->syn_fastopen)
562 return;
563
564 if (!tp->data_segs_in) {
565 skb = skb_rb_first(&tp->out_of_order_queue);
566 if (skb && !skb_rb_next(skb)) {
567 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
568 tcp_fastopen_active_disable(sk);
569 return;
570 }
571 }
572 } else if (tp->syn_fastopen_ch &&
573 atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) {
574 dst = sk_dst_get(sk);
575 if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK)))
576 atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0);
577 dst_release(dst);
578 }
579 }
580
581 void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired)
582 {
583 u32 timeouts = inet_csk(sk)->icsk_retransmits;
584 struct tcp_sock *tp = tcp_sk(sk);
585
586 /* Broken middle-boxes may black-hole Fast Open connection during or
587 * even after the handshake. Be extremely conservative and pause
588 * Fast Open globally after hitting the third consecutive timeout or
589 * exceeding the configured timeout limit.
590 */
591 if ((tp->syn_fastopen || tp->syn_data || tp->syn_data_acked) &&
592 (timeouts == 2 || (timeouts < 2 && expired))) {
593 tcp_fastopen_active_disable(sk);
594 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
595 }
596 }
597
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.