tcp_fastopen.c

  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#include <net/busy_poll.h>
  7
  8void tcp_fastopen_init_key_once(struct net *net)
  9{
 10	u8 key[TCP_FASTOPEN_KEY_LENGTH];
 11	struct tcp_fastopen_context *ctxt;
 12
 13	rcu_read_lock();
 14	ctxt = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
 15	if (ctxt) {
 16		rcu_read_unlock();
 17		return;
 18	}
 19	rcu_read_unlock();
 20
 21	/* tcp_fastopen_reset_cipher publishes the new context
 22	 * atomically, so we allow this race happening here.
 23	 *
 24	 * All call sites of tcp_fastopen_cookie_gen also check
 25	 * for a valid cookie, so this is an acceptable risk.
 26	 */
 27	get_random_bytes(key, sizeof(key));
 28	tcp_fastopen_reset_cipher(net, NULL, key, NULL);
 29}
 30
 31static void tcp_fastopen_ctx_free(struct rcu_head *head)
 32{
 33	struct tcp_fastopen_context *ctx =
 34	    container_of(head, struct tcp_fastopen_context, rcu);
 35
 36	kfree_sensitive(ctx);
 37}
 38
 39void tcp_fastopen_destroy_cipher(struct sock *sk)
 40{
 41	struct tcp_fastopen_context *ctx;
 42
 43	ctx = rcu_dereference_protected(
 44			inet_csk(sk)->icsk_accept_queue.fastopenq.ctx, 1);
 45	if (ctx)
 46		call_rcu(&ctx->rcu, tcp_fastopen_ctx_free);
 47}
 48
 49void tcp_fastopen_ctx_destroy(struct net *net)
 50{
 51	struct tcp_fastopen_context *ctxt;
 52
 53	ctxt = unrcu_pointer(xchg(&net->ipv4.tcp_fastopen_ctx, NULL));
 54
 55	if (ctxt)
 56		call_rcu(&ctxt->rcu, tcp_fastopen_ctx_free);
 57}
 58
 59int tcp_fastopen_reset_cipher(struct net *net, struct sock *sk,
 60			      void *primary_key, void *backup_key)
 61{
 62	struct tcp_fastopen_context *ctx, *octx;
 63	struct fastopen_queue *q;
 64	int err = 0;
 65
 66	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
 67	if (!ctx) {
 68		err = -ENOMEM;
 69		goto out;
 70	}
 71
 72	ctx->key[0].key[0] = get_unaligned_le64(primary_key);
 73	ctx->key[0].key[1] = get_unaligned_le64(primary_key + 8);
 74	if (backup_key) {
 75		ctx->key[1].key[0] = get_unaligned_le64(backup_key);
 76		ctx->key[1].key[1] = get_unaligned_le64(backup_key + 8);
 77		ctx->num = 2;
 78	} else {
 79		ctx->num = 1;
 80	}
 81
 82	if (sk) {
 83		q = &inet_csk(sk)->icsk_accept_queue.fastopenq;
 84		octx = unrcu_pointer(xchg(&q->ctx, RCU_INITIALIZER(ctx)));
 85	} else {
 86		octx = unrcu_pointer(xchg(&net->ipv4.tcp_fastopen_ctx,
 87					  RCU_INITIALIZER(ctx)));
 88	}
 89
 90	if (octx)
 91		call_rcu(&octx->rcu, tcp_fastopen_ctx_free);
 92out:
 93	return err;
 94}
 95
 96int tcp_fastopen_get_cipher(struct net *net, struct inet_connection_sock *icsk,
 97			    u64 *key)
 98{
 99	struct tcp_fastopen_context *ctx;
100	int n_keys = 0, i;
101
102	rcu_read_lock();
103	if (icsk)
104		ctx = rcu_dereference(icsk->icsk_accept_queue.fastopenq.ctx);
105	else
106		ctx = rcu_dereference(net->ipv4.tcp_fastopen_ctx);
107	if (ctx) {
108		n_keys = tcp_fastopen_context_len(ctx);
109		for (i = 0; i < n_keys; i++) {
110			put_unaligned_le64(ctx->key[i].key[0], key + (i * 2));
111			put_unaligned_le64(ctx->key[i].key[1], key + (i * 2) + 1);
112		}
113	}
114	rcu_read_unlock();
115
116	return n_keys;
117}
118
119static bool __tcp_fastopen_cookie_gen_cipher(struct request_sock *req,
120					     struct sk_buff *syn,
121					     const siphash_key_t *key,
122					     struct tcp_fastopen_cookie *foc)
123{
124	BUILD_BUG_ON(TCP_FASTOPEN_COOKIE_SIZE != sizeof(u64));
125
126	if (req->rsk_ops->family == AF_INET) {
127		const struct iphdr *iph = ip_hdr(syn);
128
129		foc->val[0] = cpu_to_le64(siphash(&iph->saddr,
130					  sizeof(iph->saddr) +
131					  sizeof(iph->daddr),
132					  key));
133		foc->len = TCP_FASTOPEN_COOKIE_SIZE;
134		return true;
135	}
136#if IS_ENABLED(CONFIG_IPV6)
137	if (req->rsk_ops->family == AF_INET6) {
138		const struct ipv6hdr *ip6h = ipv6_hdr(syn);
139
140		foc->val[0] = cpu_to_le64(siphash(&ip6h->saddr,
141					  sizeof(ip6h->saddr) +
142					  sizeof(ip6h->daddr),
143					  key));
144		foc->len = TCP_FASTOPEN_COOKIE_SIZE;
145		return true;
146	}
147#endif
148	return false;
149}
150
151/* Generate the fastopen cookie by applying SipHash to both the source and
152 * destination addresses.
153 */
154static void tcp_fastopen_cookie_gen(struct sock *sk,
155				    struct request_sock *req,
156				    struct sk_buff *syn,
157				    struct tcp_fastopen_cookie *foc)
158{
159	struct tcp_fastopen_context *ctx;
160
161	rcu_read_lock();
162	ctx = tcp_fastopen_get_ctx(sk);
163	if (ctx)
164		__tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[0], foc);
165	rcu_read_unlock();
166}
167
168/* If an incoming SYN or SYNACK frame contains a payload and/or FIN,
169 * queue this additional data / FIN.
170 */
171void tcp_fastopen_add_skb(struct sock *sk, struct sk_buff *skb)
172{
173	struct tcp_sock *tp = tcp_sk(sk);
174
175	if (TCP_SKB_CB(skb)->end_seq == tp->rcv_nxt)
176		return;
177
178	skb = skb_clone(skb, GFP_ATOMIC);
179	if (!skb)
180		return;
181
182	tcp_cleanup_skb(skb);
183	/* segs_in has been initialized to 1 in tcp_create_openreq_child().
184	 * Hence, reset segs_in to 0 before calling tcp_segs_in()
185	 * to avoid double counting.  Also, tcp_segs_in() expects
186	 * skb->len to include the tcp_hdrlen.  Hence, it should
187	 * be called before __skb_pull().
188	 */
189	tp->segs_in = 0;
190	tcp_segs_in(tp, skb);
191	__skb_pull(skb, tcp_hdrlen(skb));
192	sk_forced_mem_schedule(sk, skb->truesize);
193	skb_set_owner_r(skb, sk);
194
195	TCP_SKB_CB(skb)->seq++;
196	TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
197
198	tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq;
199	tcp_add_receive_queue(sk, skb);
200	tp->syn_data_acked = 1;
201
202	/* u64_stats_update_begin(&tp->syncp) not needed here,
203	 * as we certainly are not changing upper 32bit value (0)
204	 */
205	tp->bytes_received = skb->len;
206
207	if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
208		tcp_fin(sk);
209}
210
211/* returns 0 - no key match, 1 for primary, 2 for backup */
212static int tcp_fastopen_cookie_gen_check(struct sock *sk,
213					 struct request_sock *req,
214					 struct sk_buff *syn,
215					 struct tcp_fastopen_cookie *orig,
216					 struct tcp_fastopen_cookie *valid_foc)
217{
218	struct tcp_fastopen_cookie search_foc = { .len = -1 };
219	struct tcp_fastopen_cookie *foc = valid_foc;
220	struct tcp_fastopen_context *ctx;
221	int i, ret = 0;
222
223	rcu_read_lock();
224	ctx = tcp_fastopen_get_ctx(sk);
225	if (!ctx)
226		goto out;
227	for (i = 0; i < tcp_fastopen_context_len(ctx); i++) {
228		__tcp_fastopen_cookie_gen_cipher(req, syn, &ctx->key[i], foc);
229		if (tcp_fastopen_cookie_match(foc, orig)) {
230			ret = i + 1;
231			goto out;
232		}
233		foc = &search_foc;
234	}
235out:
236	rcu_read_unlock();
237	return ret;
238}
239
240static struct sock *tcp_fastopen_create_child(struct sock *sk,
241					      struct sk_buff *skb,
242					      struct request_sock *req)
243{
244	struct tcp_sock *tp;
245	struct request_sock_queue *queue = &inet_csk(sk)->icsk_accept_queue;
246	struct sock *child;
247	bool own_req;
248
249	child = inet_csk(sk)->icsk_af_ops->syn_recv_sock(sk, skb, req, NULL,
250							 NULL, &own_req, NULL);
251	if (!child)
252		return NULL;
253
254	spin_lock(&queue->fastopenq.lock);
255	queue->fastopenq.qlen++;
256	spin_unlock(&queue->fastopenq.lock);
257
258	/* Initialize the child socket. Have to fix some values to take
259	 * into account the child is a Fast Open socket and is created
260	 * only out of the bits carried in the SYN packet.
261	 */
262	tp = tcp_sk(child);
263
264	rcu_assign_pointer(tp->fastopen_rsk, req);
265	tcp_rsk(req)->tfo_listener = true;
266
267	/* RFC1323: The window in SYN & SYN/ACK segments is never
268	 * scaled. So correct it appropriately.
269	 */
270	tp->snd_wnd = ntohs(tcp_hdr(skb)->window);
271	tp->max_window = tp->snd_wnd;
272
273	/* Activate the retrans timer so that SYNACK can be retransmitted.
274	 * The request socket is not added to the ehash
275	 * because it's been added to the accept queue directly.
276	 */
277	req->timeout = tcp_timeout_init(child);
278	tcp_reset_xmit_timer(child, ICSK_TIME_RETRANS,
279			     req->timeout, false);
280
281	refcount_set(&req->rsk_refcnt, 2);
282
283	sk_mark_napi_id_set(child, skb);
284
285	/* Now finish processing the fastopen child socket. */
286	tcp_init_transfer(child, BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB, skb);
287
288	tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1;
289
290	tcp_fastopen_add_skb(child, skb);
291
292	tcp_rsk(req)->rcv_nxt = tp->rcv_nxt;
293	tp->rcv_wup = tp->rcv_nxt;
294	/* tcp_conn_request() is sending the SYNACK,
295	 * and queues the child into listener accept queue.
296	 */
297	return child;
298}
299
300static bool tcp_fastopen_queue_check(struct sock *sk)
301{
302	struct fastopen_queue *fastopenq;
303	int max_qlen;
304
305	/* Make sure the listener has enabled fastopen, and we don't
306	 * exceed the max # of pending TFO requests allowed before trying
307	 * to validating the cookie in order to avoid burning CPU cycles
308	 * unnecessarily.
309	 *
310	 * XXX (TFO) - The implication of checking the max_qlen before
311	 * processing a cookie request is that clients can't differentiate
312	 * between qlen overflow causing Fast Open to be disabled
313	 * temporarily vs a server not supporting Fast Open at all.
314	 */
315	fastopenq = &inet_csk(sk)->icsk_accept_queue.fastopenq;
316	max_qlen = READ_ONCE(fastopenq->max_qlen);
317	if (max_qlen == 0)
318		return false;
319
320	if (fastopenq->qlen >= max_qlen) {
321		struct request_sock *req1;
322		spin_lock(&fastopenq->lock);
323		req1 = fastopenq->rskq_rst_head;
324		if (!req1 || time_after(req1->rsk_timer.expires, jiffies)) {
325			__NET_INC_STATS(sock_net(sk),
326					LINUX_MIB_TCPFASTOPENLISTENOVERFLOW);
327			spin_unlock(&fastopenq->lock);
328			return false;
329		}
330		fastopenq->rskq_rst_head = req1->dl_next;
331		fastopenq->qlen--;
332		spin_unlock(&fastopenq->lock);
333		reqsk_put(req1);
334	}
335	return true;
336}
337
338static bool tcp_fastopen_no_cookie(const struct sock *sk,
339				   const struct dst_entry *dst,
340				   int flag)
341{
342	return (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen) & flag) ||
343	       tcp_sk(sk)->fastopen_no_cookie ||
344	       (dst && dst_metric(dst, RTAX_FASTOPEN_NO_COOKIE));
345}
346
347/* Returns true if we should perform Fast Open on the SYN. The cookie (foc)
348 * may be updated and return the client in the SYN-ACK later. E.g., Fast Open
349 * cookie request (foc->len == 0).
350 */
351struct sock *tcp_try_fastopen(struct sock *sk, struct sk_buff *skb,
352			      struct request_sock *req,
353			      struct tcp_fastopen_cookie *foc,
354			      const struct dst_entry *dst)
355{
356	bool syn_data = TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq + 1;
357	int tcp_fastopen = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen);
358	struct tcp_fastopen_cookie valid_foc = { .len = -1 };
359	struct sock *child;
360	int ret = 0;
361
362	if (foc->len == 0) /* Client requests a cookie */
363		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENCOOKIEREQD);
364
365	if (!((tcp_fastopen & TFO_SERVER_ENABLE) &&
366	      (syn_data || foc->len >= 0) &&
367	      tcp_fastopen_queue_check(sk))) {
368		foc->len = -1;
369		return NULL;
370	}
371
372	if (tcp_fastopen_no_cookie(sk, dst, TFO_SERVER_COOKIE_NOT_REQD))
373		goto fastopen;
374
375	if (foc->len == 0) {
376		/* Client requests a cookie. */
377		tcp_fastopen_cookie_gen(sk, req, skb, &valid_foc);
378	} else if (foc->len > 0) {
379		ret = tcp_fastopen_cookie_gen_check(sk, req, skb, foc,
380						    &valid_foc);
381		if (!ret) {
382			NET_INC_STATS(sock_net(sk),
383				      LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
384		} else {
385			/* Cookie is valid. Create a (full) child socket to
386			 * accept the data in SYN before returning a SYN-ACK to
387			 * ack the data. If we fail to create the socket, fall
388			 * back and ack the ISN only but includes the same
389			 * cookie.
390			 *
391			 * Note: Data-less SYN with valid cookie is allowed to
392			 * send data in SYN_RECV state.
393			 */
394fastopen:
395			child = tcp_fastopen_create_child(sk, skb, req);
396			if (child) {
397				if (ret == 2) {
398					valid_foc.exp = foc->exp;
399					*foc = valid_foc;
400					NET_INC_STATS(sock_net(sk),
401						      LINUX_MIB_TCPFASTOPENPASSIVEALTKEY);
402				} else {
403					foc->len = -1;
404				}
405				NET_INC_STATS(sock_net(sk),
406					      LINUX_MIB_TCPFASTOPENPASSIVE);
407				tcp_sk(child)->syn_fastopen_child = 1;
408				return child;
409			}
410			NET_INC_STATS(sock_net(sk),
411				      LINUX_MIB_TCPFASTOPENPASSIVEFAIL);
412		}
413	}
414	valid_foc.exp = foc->exp;
415	*foc = valid_foc;
416	return NULL;
417}
418
419bool tcp_fastopen_cookie_check(struct sock *sk, u16 *mss,
420			       struct tcp_fastopen_cookie *cookie)
421{
422	const struct dst_entry *dst;
423
424	tcp_fastopen_cache_get(sk, mss, cookie);
425
426	/* Firewall blackhole issue check */
427	if (tcp_fastopen_active_should_disable(sk)) {
428		cookie->len = -1;
429		return false;
430	}
431
432	dst = __sk_dst_get(sk);
433
434	if (tcp_fastopen_no_cookie(sk, dst, TFO_CLIENT_NO_COOKIE)) {
435		cookie->len = -1;
436		return true;
437	}
438	if (cookie->len > 0)
439		return true;
440	tcp_sk(sk)->fastopen_client_fail = TFO_COOKIE_UNAVAILABLE;
441	return false;
442}
443
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
450 */
451bool tcp_fastopen_defer_connect(struct sock *sk, int *err)
452{
453	struct tcp_fastopen_cookie cookie = { .len = 0 };
454	struct tcp_sock *tp = tcp_sk(sk);
455	u16 mss;
456
457	if (tp->fastopen_connect && !tp->fastopen_req) {
458		if (tcp_fastopen_cookie_check(sk, &mss, &cookie)) {
459			inet_set_bit(DEFER_CONNECT, sk);
460			return true;
461		}
462
463		/* Alloc fastopen_req in order for FO option to be included
464		 * in SYN
465		 */
466		tp->fastopen_req = kzalloc(sizeof(*tp->fastopen_req),
467					   sk->sk_allocation);
468		if (tp->fastopen_req)
469			tp->fastopen_req->cookie = cookie;
470		else
471			*err = -ENOBUFS;
472	}
473	return false;
474}
475EXPORT_IPV6_MOD(tcp_fastopen_defer_connect);
476
477/*
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
486 *      or after handshake
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.
491 */
492
493/* Disable active TFO and record current jiffies and
494 * tfo_active_disable_times
495 */
496void tcp_fastopen_active_disable(struct sock *sk)
497{
498	struct net *net = sock_net(sk);
499
500	if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout))
501		return;
502
503	/* Paired with READ_ONCE() in tcp_fastopen_active_should_disable() */
504	WRITE_ONCE(net->ipv4.tfo_active_disable_stamp, jiffies);
505
506	/* Paired with smp_rmb() in tcp_fastopen_active_should_disable().
507	 * We want net->ipv4.tfo_active_disable_stamp to be updated first.
508	 */
509	smp_mb__before_atomic();
510	atomic_inc(&net->ipv4.tfo_active_disable_times);
511
512	NET_INC_STATS(net, LINUX_MIB_TCPFASTOPENBLACKHOLE);
513}
514
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
518 */
519bool tcp_fastopen_active_should_disable(struct sock *sk)
520{
521	unsigned int tfo_bh_timeout =
522		READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_fastopen_blackhole_timeout);
523	unsigned long timeout;
524	int tfo_da_times;
525	int multiplier;
526
527	if (!tfo_bh_timeout)
528		return false;
529
530	tfo_da_times = atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times);
531	if (!tfo_da_times)
532		return false;
533
534	/* Paired with smp_mb__before_atomic() in tcp_fastopen_active_disable() */
535	smp_rmb();
536
537	/* Limit timeout to max: 2^6 * initial timeout */
538	multiplier = 1 << min(tfo_da_times - 1, 6);
539
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))
544		return true;
545
546	/* Mark check bit so we can check for successful active TFO
547	 * condition and reset tfo_active_disable_times
548	 */
549	tcp_sk(sk)->syn_fastopen_ch = 1;
550	return false;
551}
552
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
558 */
559void tcp_fastopen_active_disable_ofo_check(struct sock *sk)
560{
561	struct tcp_sock *tp = tcp_sk(sk);
562	struct net_device *dev;
563	struct dst_entry *dst;
564	struct sk_buff *skb;
565
566	if (!tp->syn_fastopen)
567		return;
568
569	if (!tp->data_segs_in) {
570		skb = skb_rb_first(&tp->out_of_order_queue);
571		if (skb && !skb_rb_next(skb)) {
572			if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
573				tcp_fastopen_active_disable(sk);
574				return;
575			}
576		}
577	} else if (tp->syn_fastopen_ch &&
578		   atomic_read(&sock_net(sk)->ipv4.tfo_active_disable_times)) {
579		rcu_read_lock();
580		dst = __sk_dst_get(sk);
581		dev = dst ? dst_dev_rcu(dst) : NULL;
582		if (!(dev && (dev->flags & IFF_LOOPBACK)))
583			atomic_set(&sock_net(sk)->ipv4.tfo_active_disable_times, 0);
584		rcu_read_unlock();
585	}
586}
587
588void tcp_fastopen_active_detect_blackhole(struct sock *sk, bool expired)
589{
590	u32 timeouts = inet_csk(sk)->icsk_retransmits;
591	struct tcp_sock *tp = tcp_sk(sk);
592
593	/* Broken middle-boxes may black-hole Fast Open connection during or
594	 * even after the handshake. Be extremely conservative and pause
595	 * Fast Open globally after hitting the third consecutive timeout or
596	 * exceeding the configured timeout limit.
597	 */
598	if ((tp->syn_fastopen || tp->syn_data || tp->syn_data_acked) &&
599	    (timeouts == 2 || (timeouts < 2 && expired))) {
600		tcp_fastopen_active_disable(sk);
601		NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVEFAIL);
602	}
603}