xref: /openbmc/linux/net/ipv4/tcp_fastopen.c (revision 2a9eb57e)
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 = xchg((__force struct tcp_fastopen_context **)&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 = xchg((__force struct tcp_fastopen_context **)&q->ctx, ctx);
84 	} else {
85 		octx = xchg((__force struct tcp_fastopen_context **)&net->ipv4.tcp_fastopen_ctx, ctx);
86 	}
87 
88 	if (octx)
89 		call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
90 out:
91 	return err;
92 }
93 
94 int tcp_fastopen_get_cipher(struct net *net, struct inet_connection_sock *icsk,
95 			    u64 *key)
96 {
97 	struct tcp_fastopen_context *ctx;
98 	int n_keys = 0, i;
99 
100 	rcu_read_lock();
101 	if (icsk)
102 		ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
103 	else
104 		ctx = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
105 	if (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);
110 		}
111 	}
112 	rcu_read_unlock();
113 
114 	return n_keys;
115 }
116 
117 static bool __tcp_fastopen_cookie_gen_cipher(struct request_sock *req,
118 					     struct sk_buff *syn,
119 					     const siphash_key_t *key,
120 					     struct tcp_fastopen_cookie *foc)
121 {
122 	BUILD_BUG_ON(TCP_FASTOPEN_COOKIE_SIZE != sizeof(u64));
123 
124 	if (req->rsk_ops->family == AF_INET) {
125 		const struct iphdr *iph = ip_hdr(syn);
126 
127 		foc->val[0] = cpu_to_le64(siphash(&iph->saddr,
128 					  sizeof(iph->saddr) +
129 					  sizeof(iph->daddr),
130 					  key));
131 		foc->len = TCP_FASTOPEN_COOKIE_SIZE;
132 		return true;
133 	}
134 #if IS_ENABLED(CONFIG_IPV6)
135 	if (req->rsk_ops->family == AF_INET6) {
136 		const struct ipv6hdr *ip6h = ipv6_hdr(syn);
137 
138 		foc->val[0] = cpu_to_le64(siphash(&ip6h->saddr,
139 					  sizeof(ip6h->saddr) +
140 					  sizeof(ip6h->daddr),
141 					  key));
142 		foc->len = TCP_FASTOPEN_COOKIE_SIZE;
143 		return true;
144 	}
145 #endif
146 	return false;
147 }
148 
149 /* Generate the fastopen cookie by applying SipHash to both the source and
150  * destination addresses.
151  */
152 static void tcp_fastopen_cookie_gen(struct sock *sk,
153 				    struct request_sock *req,
154 				    struct sk_buff *syn,
155 				    struct tcp_fastopen_cookie *foc)
156 {
157 	struct tcp_fastopen_context *ctx;
158 
159 	rcu_read_lock();
160 	ctx = tcp_fastopen_get_ctx(sk);
161 	if (ctx)
162 		__tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[0], foc);
163 	rcu_read_unlock();
164 }
165 
166 /* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
167  * queue this additional data / FIN.
168  */
169 void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
170 {
171 	struct tcp_sock *tp = tcp_sk(sk);
172 
173 	if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
174 		return;
175 
176 	skb = skb_clone(skb, GFP_ATOMIC);
177 	if (!skb)
178 		return;
179 
180 	skb_dst_drop(skb);
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().
186 	 */
187 	tp->segs_in = 0;
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);
192 
193 	TCP_SKB_CB(skb)->seq++;
194 	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
195 
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;
199 
200 	/* u64_stats_update_begin(&tp->syncp) not needed here,
201 	 * as we certainly are not changing upper 32bit value (0)
202 	 */
203 	tp->bytes_received = skb->len;
204 
205 	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
206 		tcp_fin(sk);
207 }
208 
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,
212 					 struct sk_buff *syn,
213 					 struct tcp_fastopen_cookie *orig,
214 					 struct tcp_fastopen_cookie *valid_foc)
215 {
216 	struct tcp_fastopen_cookie search_foc = { .len = -1 };
217 	struct tcp_fastopen_cookie *foc = valid_foc;
218 	struct tcp_fastopen_context *ctx;
219 	int i, ret = 0;
220 
221 	rcu_read_lock();
222 	ctx = tcp_fastopen_get_ctx(sk);
223 	if (!ctx)
224 		goto out;
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)) {
228 			ret = i + 1;
229 			goto out;
230 		}
231 		foc = &search_foc;
232 	}
233 out:
234 	rcu_read_unlock();
235 	return ret;
236 }
237 
238 static struct sock *tcp_fastopen_create_child(struct sock *sk,
239 					      struct sk_buff *skb,
240 					      struct request_sock *req)
241 {
242 	struct tcp_sock *tp;
243 	struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
244 	struct sock *child;
245 	bool own_req;
246 
247 	child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
248 							 NULL, &own_req);
249 	if (!child)
250 		return NULL;
251 
252 	spin_lock(&queue->fastopenq.lock);
253 	queue->fastopenq.qlen++;
254 	spin_unlock(&queue->fastopenq.lock);
255 
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.
259 	 */
260 	tp = tcp_sk(child);
261 
262 	rcu_assign_pointer(tp->fastopen_rsk, req);
263 	tcp_rsk(req)->tfo_listener = true;
264 
265 	/* RFC1323: The window in SYN & SYN/ACK segments is never
266 	 * scaled. So correct it appropriately.
267 	 */
268 	tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
269 	tp->max_window = tp->snd_wnd;
270 
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.
274 	 */
275 	inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
276 				  TCP_TIMEOUT_INIT, TCP_RTO_MAX);
277 
278 	refcount_set(&req->rsk_refcnt, 2);
279 
280 	/* Now finish processing the fastopen child socket. */
281 	tcp_init_transfer(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, skb);
282 
283 	tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
284 
285 	tcp_fastopen_add_skb(child, skb);
286 
287 	tcp_rsk(req)->rcv_nxt = tp->rcv_nxt;
288 	tp->rcv_wup = tp->rcv_nxt;
289 	/* tcp_conn_request() is sending the SYNACK,
290 	 * and queues the child into listener accept queue.
291 	 */
292 	return child;
293 }
294 
295 static bool tcp_fastopen_queue_check(struct sock *sk)
296 {
297 	struct fastopen_queue *fastopenq;
298 
299 	/* Make sure the listener has enabled fastopen, and we don't
300 	 * exceed the max # of pending TFO requests allowed before trying
301 	 * to validating the cookie in order to avoid burning CPU cycles
302 	 * unnecessarily.
303 	 *
304 	 * XXX (TFO) - The implication of checking the max_qlen before
305 	 * processing a cookie request is that clients can't differentiate
306 	 * between qlen overflow causing Fast Open to be disabled
307 	 * temporarily vs a server not supporting Fast Open at all.
308 	 */
309 	fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
310 	if (fastopenq->max_qlen == 0)
311 		return false;
312 
313 	if (fastopenq->qlen >= fastopenq->max_qlen) {
314 		struct request_sock *req1;
315 		spin_lock(&fastopenq->lock);
316 		req1 = fastopenq->rskq_rst_head;
317 		if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
318 			__NET_INC_STATS(sock_net(sk),
319 					LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
320 			spin_unlock(&fastopenq->lock);
321 			return false;
322 		}
323 		fastopenq->rskq_rst_head = req1->dl_next;
324 		fastopenq->qlen--;
325 		spin_unlock(&fastopenq->lock);
326 		reqsk_put(req1);
327 	}
328 	return true;
329 }
330 
331 static bool tcp_fastopen_no_cookie(const struct sock *sk,
332 				   const struct dst_entry *dst,
333 				   int flag)
334 {
335 	return (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) & flag) ||
336 	       tcp_sk(sk)->fastopen_no_cookie ||
337 	       (dst && dst_metric(dst, RTAX_FASTOPEN_NO_COOKIE));
338 }
339 
340 /* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
341  * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
342  * cookie request (foc->len == 0).
343  */
344 struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
345 			      struct request_sock *req,
346 			      struct tcp_fastopen_cookie *foc,
347 			      const struct dst_entry *dst)
348 {
349 	bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
350 	int tcp_fastopen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen);
351 	struct tcp_fastopen_cookie valid_foc = { .len = -1 };
352 	struct sock *child;
353 	int ret = 0;
354 
355 	if (foc->len == 0) /* Client requests a cookie */
356 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
357 
358 	if (!((tcp_fastopen & TFO_SERVER_ENABLE) &&
359 	      (syn_data || foc->len >= 0) &&
360 	      tcp_fastopen_queue_check(sk))) {
361 		foc->len = -1;
362 		return NULL;
363 	}
364 
365 	if (tcp_fastopen_no_cookie(sk, dst, TFO_SERVER_COOKIE_NOT_REQD))
366 		goto fastopen;
367 
368 	if (foc->len == 0) {
369 		/* Client requests a cookie. */
370 		tcp_fastopen_cookie_gen(sk, req, skb, &valid_foc);
371 	} else if (foc->len > 0) {
372 		ret = tcp_fastopen_cookie_gen_check(sk, req, skb, foc,
373 						    &valid_foc);
374 		if (!ret) {
375 			NET_INC_STATS(sock_net(sk),
376 				      LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
377 		} else {
378 			/* Cookie is valid. Create a (full) child socket to
379 			 * accept the data in SYN before returning a SYN-ACK to
380 			 * ack the data. If we fail to create the socket, fall
381 			 * back and ack the ISN only but includes the same
382 			 * cookie.
383 			 *
384 			 * Note: Data-less SYN with valid cookie is allowed to
385 			 * send data in SYN_RECV state.
386 			 */
387 fastopen:
388 			child = tcp_fastopen_create_child(sk, skb, req);
389 			if (child) {
390 				if (ret == 2) {
391 					valid_foc.exp = foc->exp;
392 					*foc = valid_foc;
393 					NET_INC_STATS(sock_net(sk),
394 						      LINUX_MIB_TCPFASTOPENPASSIVEALTKEY);
395 				} else {
396 					foc->len = -1;
397 				}
398 				NET_INC_STATS(sock_net(sk),
399 					      LINUX_MIB_TCPFASTOPENPASSIVE);
400 				return child;
401 			}
402 			NET_INC_STATS(sock_net(sk),
403 				      LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
404 		}
405 	}
406 	valid_foc.exp = foc->exp;
407 	*foc = valid_foc;
408 	return NULL;
409 }
410 
411 bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
412 			       struct tcp_fastopen_cookie *cookie)
413 {
414 	const struct dst_entry *dst;
415 
416 	tcp_fastopen_cache_get(sk, mss, cookie);
417 
418 	/* Firewall blackhole issue check */
419 	if (tcp_fastopen_active_should_disable(sk)) {
420 		cookie->len = -1;
421 		return false;
422 	}
423 
424 	dst = __sk_dst_get(sk);
425 
426 	if (tcp_fastopen_no_cookie(sk, dst, TFO_CLIENT_NO_COOKIE)) {
427 		cookie->len = -1;
428 		return true;
429 	}
430 	if (cookie->len > 0)
431 		return true;
432 	tcp_sk(sk)->fastopen_client_fail = TFO_COOKIE_UNAVAILABLE;
433 	return false;
434 }
435 
436 /* This function checks if we want to defer sending SYN until the first
437  * write().  We defer under the following conditions:
438  * 1. fastopen_connect sockopt is set
439  * 2. we have a valid cookie
440  * Return value: return true if we want to defer until application writes data
441  *               return false if we want to send out SYN immediately
442  */
443 bool tcp_fastopen_defer_connect(struct sock *sk, int *err)
444 {
445 	struct tcp_fastopen_cookie cookie = { .len = 0 };
446 	struct tcp_sock *tp = tcp_sk(sk);
447 	u16 mss;
448 
449 	if (tp->fastopen_connect && !tp->fastopen_req) {
450 		if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) {
451 			inet_sk(sk)->defer_connect = 1;
452 			return true;
453 		}
454 
455 		/* Alloc fastopen_req in order for FO option to be included
456 		 * in SYN
457 		 */
458 		tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req),
459 					   sk->sk_allocation);
460 		if (tp->fastopen_req)
461 			tp->fastopen_req->cookie = cookie;
462 		else
463 			*err = -ENOBUFS;
464 	}
465 	return false;
466 }
467 EXPORT_SYMBOL(tcp_fastopen_defer_connect);
468 
469 /*
470  * The following code block is to deal with middle box issues with TFO:
471  * Middlebox firewall issues can potentially cause server's data being
472  * blackholed after a successful 3WHS using TFO.
473  * The proposed solution is to disable active TFO globally under the
474  * following circumstances:
475  *   1. client side TFO socket receives out of order FIN
476  *   2. client side TFO socket receives out of order RST
477  *   3. client side TFO socket has timed out three times consecutively during
478  *      or after handshake
479  * We disable active side TFO globally for 1hr at first. Then if it
480  * happens again, we disable it for 2h, then 4h, 8h, ...
481  * And we reset the timeout back to 1hr when we see a successful active
482  * TFO connection with data exchanges.
483  */
484 
485 /* Disable active TFO and record current jiffies and
486  * tfo_active_disable_times
487  */
488 void tcp_fastopen_active_disable(struct sock *sk)
489 {
490 	struct net *net = sock_net(sk);
491 
492 	if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout))
493 		return;
494 
495 	/* Paired with READ_ONCE() in tcp_fastopen_active_should_disable() */
496 	WRITE_ONCE(net->ipv4.tfo_active_disable_stamp, jiffies);
497 
498 	/* Paired with smp_rmb() in tcp_fastopen_active_should_disable().
499 	 * We want net->ipv4.tfo_active_disable_stamp to be updated first.
500 	 */
501 	smp_mb__before_atomic();
502 	atomic_inc(&net->ipv4.tfo_active_disable_times);
503 
504 	NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE);
505 }
506 
507 /* Calculate timeout for tfo active disable
508  * Return true if we are still in the active TFO disable period
509  * Return false if timeout already expired and we should use active TFO
510  */
511 bool tcp_fastopen_active_should_disable(struct sock *sk)
512 {
513 	unsigned int tfo_bh_timeout =
514 		READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout);
515 	unsigned long timeout;
516 	int tfo_da_times;
517 	int multiplier;
518 
519 	if (!tfo_bh_timeout)
520 		return false;
521 
522 	tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times);
523 	if (!tfo_da_times)
524 		return false;
525 
526 	/* Paired with smp_mb__before_atomic() in tcp_fastopen_active_disable() */
527 	smp_rmb();
528 
529 	/* Limit timeout to max: 2^6 * initial timeout */
530 	multiplier = 1 << min(tfo_da_times - 1, 6);
531 
532 	/* Paired with the WRITE_ONCE() in tcp_fastopen_active_disable(). */
533 	timeout = READ_ONCE(sock_net(sk)->ipv4.tfo_active_disable_stamp) +
534 		  multiplier * tfo_bh_timeout * HZ;
535 	if (time_before(jiffies, timeout))
536 		return true;
537 
538 	/* Mark check bit so we can check for successful active TFO
539 	 * condition and reset tfo_active_disable_times
540 	 */
541 	tcp_sk(sk)->syn_fastopen_ch = 1;
542 	return false;
543 }
544 
545 /* Disable active TFO if FIN is the only packet in the ofo queue
546  * and no data is received.
547  * Also check if we can reset tfo_active_disable_times if data is
548  * received successfully on a marked active TFO sockets opened on
549  * a non-loopback interface
550  */
551 void tcp_fastopen_active_disable_ofo_check(struct sock *sk)
552 {
553 	struct tcp_sock *tp = tcp_sk(sk);
554 	struct dst_entry *dst;
555 	struct sk_buff *skb;
556 
557 	if (!tp->syn_fastopen)
558 		return;
559 
560 	if (!tp->data_segs_in) {
561 		skb = skb_rb_first(&tp->out_of_order_queue);
562 		if (skb && !skb_rb_next(skb)) {
563 			if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
564 				tcp_fastopen_active_disable(sk);
565 				return;
566 			}
567 		}
568 	} else if (tp->syn_fastopen_ch &&
569 		   atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) {
570 		dst = sk_dst_get(sk);
571 		if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK)))
572 			atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0);
573 		dst_release(dst);
574 	}
575 }
576 
577 void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired)
578 {
579 	u32 timeouts = inet_csk(sk)->icsk_retransmits;
580 	struct tcp_sock *tp = tcp_sk(sk);
581 
582 	/* Broken middle-boxes may black-hole Fast Open connection during or
583 	 * even after the handshake. Be extremely conservative and pause
584 	 * Fast Open globally after hitting the third consecutive timeout or
585 	 * exceeding the configured timeout limit.
586 	 */
587 	if ((tp->syn_fastopen || tp->syn_data || tp->syn_data_acked) &&
588 	    (timeouts == 2 || (timeouts < 2 && expired))) {
589 		tcp_fastopen_active_disable(sk);
590 		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
591 	}
592 }
593