1// SPDX-License-Identifier: GPL-2.0
2#include <linux/crypto.h>
3#include <linux/err.h>
4#include <linux/init.h>
5#include <linux/kernel.h>
6#include <linux/list.h>
7#include <linux/tcp.h>
8#include <linux/rcupdate.h>
9#include <linux/rculist.h>
10#include <net/inetpeer.h>
11#include <net/tcp.h>
12
13void tcp_fastopen_init_key_once(struct net *net)
14{
15	u8 key[TCP_FASTOPEN_KEY_LENGTH];
16	struct tcp_fastopen_context *ctxt;
17
18	rcu_read_lock();
19	ctxt = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
20	if (ctxt) {
21		rcu_read_unlock();
22		return;
23	}
24	rcu_read_unlock();
25
26	/* tcp_fastopen_reset_cipher publishes the new context
27	 * atomically, so we allow this race happening here.
28	 *
29	 * All call sites of tcp_fastopen_cookie_gen also check
30	 * for a valid cookie, so this is an acceptable risk.
31	 */
32	get_random_bytes(key, sizeof(key));
33	tcp_fastopen_reset_cipher(net, NULL, key, NULL);
34}
35
36static void tcp_fastopen_ctx_free(struct rcu_head *head)
37{
38	struct tcp_fastopen_context *ctx =
39	    container_of(head, struct tcp_fastopen_context, rcu);
40
41	kfree_sensitive(ctx);
42}
43
44void tcp_fastopen_destroy_cipher(struct sock *sk)
45{
46	struct tcp_fastopen_context *ctx;
47
48	ctx = rcu_dereference_protected(
49			inet_csk(sk)->icsk_accept_queue.fastopenq.ctx, 1);
50	if (ctx)
51		call_rcu(&ctx->rcu, tcp_fastopen_ctx_free);
52}
53
54void tcp_fastopen_ctx_destroy(struct net *net)
55{
56	struct tcp_fastopen_context *ctxt;
57
58	spin_lock(&net->ipv4.tcp_fastopen_ctx_lock);
59
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);
64
65	if (ctxt)
66		call_rcu(&ctxt->rcu, tcp_fastopen_ctx_free);
67}
68
69int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
70			      void *primary_key, void *backup_key)
71{
72	struct tcp_fastopen_context *ctx, *octx;
73	struct fastopen_queue *q;
74	int err = 0;
75
76	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
77	if (!ctx) {
78		err = -ENOMEM;
79		goto out;
80	}
81
82	ctx->key[0].key[0] = get_unaligned_le64(primary_key);
83	ctx->key[0].key[1] = get_unaligned_le64(primary_key + 8);
84	if (backup_key) {
85		ctx->key[1].key[0] = get_unaligned_le64(backup_key);
86		ctx->key[1].key[1] = get_unaligned_le64(backup_key + 8);
87		ctx->num = 2;
88	} else {
89		ctx->num = 1;
90	}
91
92	spin_lock(&net->ipv4.tcp_fastopen_ctx_lock);
93	if (sk) {
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);
98	} else {
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);
102	}
103	spin_unlock(&net->ipv4.tcp_fastopen_ctx_lock);
104
105	if (octx)
106		call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
107out:
108	return err;
109}
110
111int tcp_fastopen_get_cipher(struct net *net, struct inet_connection_sock *icsk,
112			    u64 *key)
113{
114	struct tcp_fastopen_context *ctx;
115	int n_keys = 0, i;
116
117	rcu_read_lock();
118	if (icsk)
119		ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
120	else
121		ctx = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
122	if (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);
127		}
128	}
129	rcu_read_unlock();
130
131	return n_keys;
132}
133
134static bool __tcp_fastopen_cookie_gen_cipher(struct request_sock *req,
135					     struct sk_buff *syn,
136					     const siphash_key_t *key,
137					     struct tcp_fastopen_cookie *foc)
138{
139	BUILD_BUG_ON(TCP_FASTOPEN_COOKIE_SIZE != sizeof(u64));
140
141	if (req->rsk_ops->family == AF_INET) {
142		const struct iphdr *iph = ip_hdr(syn);
143
144		foc->val[0] = cpu_to_le64(siphash(&iph->saddr,
145					  sizeof(iph->saddr) +
146					  sizeof(iph->daddr),
147					  key));
148		foc->len = TCP_FASTOPEN_COOKIE_SIZE;
149		return true;
150	}
151#if IS_ENABLED(CONFIG_IPV6)
152	if (req->rsk_ops->family == AF_INET6) {
153		const struct ipv6hdr *ip6h = ipv6_hdr(syn);
154
155		foc->val[0] = cpu_to_le64(siphash(&ip6h->saddr,
156					  sizeof(ip6h->saddr) +
157					  sizeof(ip6h->daddr),
158					  key));
159		foc->len = TCP_FASTOPEN_COOKIE_SIZE;
160		return true;
161	}
162#endif
163	return false;
164}
165
166/* Generate the fastopen cookie by applying SipHash to both the source and
167 * destination addresses.
168 */
169static void tcp_fastopen_cookie_gen(struct sock *sk,
170				    struct request_sock *req,
171				    struct sk_buff *syn,
172				    struct tcp_fastopen_cookie *foc)
173{
174	struct tcp_fastopen_context *ctx;
175
176	rcu_read_lock();
177	ctx = tcp_fastopen_get_ctx(sk);
178	if (ctx)
179		__tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[0], foc);
180	rcu_read_unlock();
181}
182
183/* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
184 * queue this additional data / FIN.
185 */
186void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
187{
188	struct tcp_sock *tp = tcp_sk(sk);
189
190	if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
191		return;
192
193	skb = skb_clone(skb, GFP_ATOMIC);
194	if (!skb)
195		return;
196
197	skb_dst_drop(skb);
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().
203	 */
204	tp->segs_in = 0;
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);
209
210	TCP_SKB_CB(skb)->seq++;
211	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
212
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;
216
217	/* u64_stats_update_begin(&tp->syncp) not needed here,
218	 * as we certainly are not changing upper 32bit value (0)
219	 */
220	tp->bytes_received = skb->len;
221
222	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
223		tcp_fin(sk);
224}
225
226/* returns 0 - no key match, 1 for primary, 2 for backup */
227static int tcp_fastopen_cookie_gen_check(struct sock *sk,
228					 struct request_sock *req,
229					 struct sk_buff *syn,
230					 struct tcp_fastopen_cookie *orig,
231					 struct tcp_fastopen_cookie *valid_foc)
232{
233	struct tcp_fastopen_cookie search_foc = { .len = -1 };
234	struct tcp_fastopen_cookie *foc = valid_foc;
235	struct tcp_fastopen_context *ctx;
236	int i, ret = 0;
237
238	rcu_read_lock();
239	ctx = tcp_fastopen_get_ctx(sk);
240	if (!ctx)
241		goto out;
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)) {
245			ret = i + 1;
246			goto out;
247		}
248		foc = &search_foc;
249	}
250out:
251	rcu_read_unlock();
252	return ret;
253}
254
255static struct sock *tcp_fastopen_create_child(struct sock *sk,
256					      struct sk_buff *skb,
257					      struct request_sock *req)
258{
259	struct tcp_sock *tp;
260	struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
261	struct sock *child;
262	bool own_req;
263
264	child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
265							 NULL, &own_req);
266	if (!child)
267		return NULL;
268
269	spin_lock(&queue->fastopenq.lock);
270	queue->fastopenq.qlen++;
271	spin_unlock(&queue->fastopenq.lock);
272
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.
276	 */
277	tp = tcp_sk(child);
278
279	rcu_assign_pointer(tp->fastopen_rsk, req);
280	tcp_rsk(req)->tfo_listener = true;
281
282	/* RFC1323: The window in SYN & SYN/ACK segments is never
283	 * scaled. So correct it appropriately.
284	 */
285	tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
286	tp->max_window = tp->snd_wnd;
287
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.
291	 */
292	inet_csk_reset_xmit_timer(child, ICSK_TIME_RETRANS,
293				  TCP_TIMEOUT_INIT, TCP_RTO_MAX);
294
295	refcount_set(&req->rsk_refcnt, 2);
296
297	/* Now finish processing the fastopen child socket. */
298	tcp_init_transfer(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, skb);
299
300	tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
301
302	tcp_fastopen_add_skb(child, skb);
303
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.
308	 */
309	return child;
310}
311
312static bool tcp_fastopen_queue_check(struct sock *sk)
313{
314	struct fastopen_queue *fastopenq;
315	int max_qlen;
316
317	/* Make sure the listener has enabled fastopen, and we don't
318	 * exceed the max # of pending TFO requests allowed before trying
319	 * to validating the cookie in order to avoid burning CPU cycles
320	 * unnecessarily.
321	 *
322	 * XXX (TFO) - The implication of checking the max_qlen before
323	 * processing a cookie request is that clients can't differentiate
324	 * between qlen overflow causing Fast Open to be disabled
325	 * temporarily vs a server not supporting Fast Open at all.
326	 */
327	fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
328	max_qlen = READ_ONCE(fastopenq->max_qlen);
329	if (max_qlen == 0)
330		return false;
331
332	if (fastopenq->qlen >= max_qlen) {
333		struct request_sock *req1;
334		spin_lock(&fastopenq->lock);
335		req1 = fastopenq->rskq_rst_head;
336		if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
337			__NET_INC_STATS(sock_net(sk),
338					LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
339			spin_unlock(&fastopenq->lock);
340			return false;
341		}
342		fastopenq->rskq_rst_head = req1->dl_next;
343		fastopenq->qlen--;
344		spin_unlock(&fastopenq->lock);
345		reqsk_put(req1);
346	}
347	return true;
348}
349
350static bool tcp_fastopen_no_cookie(const struct sock *sk,
351				   const struct dst_entry *dst,
352				   int flag)
353{
354	return (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) & flag) ||
355	       tcp_sk(sk)->fastopen_no_cookie ||
356	       (dst && dst_metric(dst, RTAX_FASTOPEN_NO_COOKIE));
357}
358
359/* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
360 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
361 * cookie request (foc->len == 0).
362 */
363struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
364			      struct request_sock *req,
365			      struct tcp_fastopen_cookie *foc,
366			      const struct dst_entry *dst)
367{
368	bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
369	int tcp_fastopen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen);
370	struct tcp_fastopen_cookie valid_foc = { .len = -1 };
371	struct sock *child;
372	int ret = 0;
373
374	if (foc->len == 0) /* Client requests a cookie */
375		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
376
377	if (!((tcp_fastopen & TFO_SERVER_ENABLE) &&
378	      (syn_data || foc->len >= 0) &&
379	      tcp_fastopen_queue_check(sk))) {
380		foc->len = -1;
381		return NULL;
382	}
383
384	if (tcp_fastopen_no_cookie(sk, dst, TFO_SERVER_COOKIE_NOT_REQD))
385		goto fastopen;
386
387	if (foc->len == 0) {
388		/* Client requests a cookie. */
389		tcp_fastopen_cookie_gen(sk, req, skb, &valid_foc);
390	} else if (foc->len > 0) {
391		ret = tcp_fastopen_cookie_gen_check(sk, req, skb, foc,
392						    &valid_foc);
393		if (!ret) {
394			NET_INC_STATS(sock_net(sk),
395				      LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
396		} else {
397			/* Cookie is valid. Create a (full) child socket to
398			 * accept the data in SYN before returning a SYN-ACK to
399			 * ack the data. If we fail to create the socket, fall
400			 * back and ack the ISN only but includes the same
401			 * cookie.
402			 *
403			 * Note: Data-less SYN with valid cookie is allowed to
404			 * send data in SYN_RECV state.
405			 */
406fastopen:
407			child = tcp_fastopen_create_child(sk, skb, req);
408			if (child) {
409				if (ret == 2) {
410					valid_foc.exp = foc->exp;
411					*foc = valid_foc;
412					NET_INC_STATS(sock_net(sk),
413						      LINUX_MIB_TCPFASTOPENPASSIVEALTKEY);
414				} else {
415					foc->len = -1;
416				}
417				NET_INC_STATS(sock_net(sk),
418					      LINUX_MIB_TCPFASTOPENPASSIVE);
419				return child;
420			}
421			NET_INC_STATS(sock_net(sk),
422				      LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
423		}
424	}
425	valid_foc.exp = foc->exp;
426	*foc = valid_foc;
427	return NULL;
428}
429
430bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
431			       struct tcp_fastopen_cookie *cookie)
432{
433	const struct dst_entry *dst;
434
435	tcp_fastopen_cache_get(sk, mss, cookie);
436
437	/* Firewall blackhole issue check */
438	if (tcp_fastopen_active_should_disable(sk)) {
439		cookie->len = -1;
440		return false;
441	}
442
443	dst = __sk_dst_get(sk);
444
445	if (tcp_fastopen_no_cookie(sk, dst, TFO_CLIENT_NO_COOKIE)) {
446		cookie->len = -1;
447		return true;
448	}
449	if (cookie->len > 0)
450		return true;
451	tcp_sk(sk)->fastopen_client_fail = TFO_COOKIE_UNAVAILABLE;
452	return false;
453}
454
455/* This function checks if we want to defer sending SYN until the first
456 * write().  We defer under the following conditions:
457 * 1. fastopen_connect sockopt is set
458 * 2. we have a valid cookie
459 * Return value: return true if we want to defer until application writes data
460 *               return false if we want to send out SYN immediately
461 */
462bool tcp_fastopen_defer_connect(struct sock *sk, int *err)
463{
464	struct tcp_fastopen_cookie cookie = { .len = 0 };
465	struct tcp_sock *tp = tcp_sk(sk);
466	u16 mss;
467
468	if (tp->fastopen_connect && !tp->fastopen_req) {
469		if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) {
470			inet_sk(sk)->defer_connect = 1;
471			return true;
472		}
473
474		/* Alloc fastopen_req in order for FO option to be included
475		 * in SYN
476		 */
477		tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req),
478					   sk->sk_allocation);
479		if (tp->fastopen_req)
480			tp->fastopen_req->cookie = cookie;
481		else
482			*err = -ENOBUFS;
483	}
484	return false;
485}
486EXPORT_SYMBOL(tcp_fastopen_defer_connect);
487
488/*
489 * The following code block is to deal with middle box issues with TFO:
490 * Middlebox firewall issues can potentially cause server's data being
491 * blackholed after a successful 3WHS using TFO.
492 * The proposed solution is to disable active TFO globally under the
493 * following circumstances:
494 *   1. client side TFO socket receives out of order FIN
495 *   2. client side TFO socket receives out of order RST
496 *   3. client side TFO socket has timed out three times consecutively during
497 *      or after handshake
498 * We disable active side TFO globally for 1hr at first. Then if it
499 * happens again, we disable it for 2h, then 4h, 8h, ...
500 * And we reset the timeout back to 1hr when we see a successful active
501 * TFO connection with data exchanges.
502 */
503
504/* Disable active TFO and record current jiffies and
505 * tfo_active_disable_times
506 */
507void tcp_fastopen_active_disable(struct sock *sk)
508{
509	struct net *net = sock_net(sk);
510
511	if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout))
512		return;
513
514	/* Paired with READ_ONCE() in tcp_fastopen_active_should_disable() */
515	WRITE_ONCE(net->ipv4.tfo_active_disable_stamp, jiffies);
516
517	/* Paired with smp_rmb() in tcp_fastopen_active_should_disable().
518	 * We want net->ipv4.tfo_active_disable_stamp to be updated first.
519	 */
520	smp_mb__before_atomic();
521	atomic_inc(&net->ipv4.tfo_active_disable_times);
522
523	NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE);
524}
525
526/* Calculate timeout for tfo active disable
527 * Return true if we are still in the active TFO disable period
528 * Return false if timeout already expired and we should use active TFO
529 */
530bool tcp_fastopen_active_should_disable(struct sock *sk)
531{
532	unsigned int tfo_bh_timeout =
533		READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout);
534	unsigned long timeout;
535	int tfo_da_times;
536	int multiplier;
537
538	if (!tfo_bh_timeout)
539		return false;
540
541	tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times);
542	if (!tfo_da_times)
543		return false;
544
545	/* Paired with smp_mb__before_atomic() in tcp_fastopen_active_disable() */
546	smp_rmb();
547
548	/* Limit timout to max: 2^6 * initial timeout */
549	multiplier = 1 << min(tfo_da_times - 1, 6);
550
551	/* Paired with the WRITE_ONCE() in tcp_fastopen_active_disable(). */
552	timeout = READ_ONCE(sock_net(sk)->ipv4.tfo_active_disable_stamp) +
553		  multiplier * tfo_bh_timeout * HZ;
554	if (time_before(jiffies, timeout))
555		return true;
556
557	/* Mark check bit so we can check for successful active TFO
558	 * condition and reset tfo_active_disable_times
559	 */
560	tcp_sk(sk)->syn_fastopen_ch = 1;
561	return false;
562}
563
564/* Disable active TFO if FIN is the only packet in the ofo queue
565 * and no data is received.
566 * Also check if we can reset tfo_active_disable_times if data is
567 * received successfully on a marked active TFO sockets opened on
568 * a non-loopback interface
569 */
570void tcp_fastopen_active_disable_ofo_check(struct sock *sk)
571{
572	struct tcp_sock *tp = tcp_sk(sk);
573	struct dst_entry *dst;
574	struct sk_buff *skb;
575
576	if (!tp->syn_fastopen)
577		return;
578
579	if (!tp->data_segs_in) {
580		skb = skb_rb_first(&tp->out_of_order_queue);
581		if (skb && !skb_rb_next(skb)) {
582			if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
583				tcp_fastopen_active_disable(sk);
584				return;
585			}
586		}
587	} else if (tp->syn_fastopen_ch &&
588		   atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) {
589		dst = sk_dst_get(sk);
590		if (!(dst && dst->dev && (dst->dev->flags & IFF_LOOPBACK)))
591			atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0);
592		dst_release(dst);
593	}
594}
595
596void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired)
597{
598	u32 timeouts = inet_csk(sk)->icsk_retransmits;
599	struct tcp_sock *tp = tcp_sk(sk);
600
601	/* Broken middle-boxes may black-hole Fast Open connection during or
602	 * even after the handshake. Be extremely conservative and pause
603	 * Fast Open globally after hitting the third consecutive timeout or
604	 * exceeding the configured timeout limit.
605	 */
606	if ((tp->syn_fastopen || tp->syn_data || tp->syn_data_acked) &&
607	    (timeouts == 2 || (timeouts < 2 && expired))) {
608		tcp_fastopen_active_disable(sk);
609		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
610	}
611}
612