xref: /openbmc/linux/net/ipv4/tcp_metrics.c (revision 95e9fd10)
1 #include <linux/rcupdate.h>
2 #include <linux/spinlock.h>
3 #include <linux/jiffies.h>
4 #include <linux/bootmem.h>
5 #include <linux/module.h>
6 #include <linux/cache.h>
7 #include <linux/slab.h>
8 #include <linux/init.h>
9 #include <linux/tcp.h>
10 #include <linux/hash.h>
11 
12 #include <net/inet_connection_sock.h>
13 #include <net/net_namespace.h>
14 #include <net/request_sock.h>
15 #include <net/inetpeer.h>
16 #include <net/sock.h>
17 #include <net/ipv6.h>
18 #include <net/dst.h>
19 #include <net/tcp.h>
20 
21 int sysctl_tcp_nometrics_save __read_mostly;
22 
23 enum tcp_metric_index {
24 	TCP_METRIC_RTT,
25 	TCP_METRIC_RTTVAR,
26 	TCP_METRIC_SSTHRESH,
27 	TCP_METRIC_CWND,
28 	TCP_METRIC_REORDERING,
29 
30 	/* Always last.  */
31 	TCP_METRIC_MAX,
32 };
33 
34 struct tcp_fastopen_metrics {
35 	u16	mss;
36 	u16	syn_loss:10;		/* Recurring Fast Open SYN losses */
37 	unsigned long	last_syn_loss;	/* Last Fast Open SYN loss */
38 	struct	tcp_fastopen_cookie	cookie;
39 };
40 
41 struct tcp_metrics_block {
42 	struct tcp_metrics_block __rcu	*tcpm_next;
43 	struct inetpeer_addr		tcpm_addr;
44 	unsigned long			tcpm_stamp;
45 	u32				tcpm_ts;
46 	u32				tcpm_ts_stamp;
47 	u32				tcpm_lock;
48 	u32				tcpm_vals[TCP_METRIC_MAX];
49 	struct tcp_fastopen_metrics	tcpm_fastopen;
50 };
51 
52 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
53 			      enum tcp_metric_index idx)
54 {
55 	return tm->tcpm_lock & (1 << idx);
56 }
57 
58 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
59 			  enum tcp_metric_index idx)
60 {
61 	return tm->tcpm_vals[idx];
62 }
63 
64 static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm,
65 				  enum tcp_metric_index idx)
66 {
67 	return msecs_to_jiffies(tm->tcpm_vals[idx]);
68 }
69 
70 static void tcp_metric_set(struct tcp_metrics_block *tm,
71 			   enum tcp_metric_index idx,
72 			   u32 val)
73 {
74 	tm->tcpm_vals[idx] = val;
75 }
76 
77 static void tcp_metric_set_msecs(struct tcp_metrics_block *tm,
78 				 enum tcp_metric_index idx,
79 				 u32 val)
80 {
81 	tm->tcpm_vals[idx] = jiffies_to_msecs(val);
82 }
83 
84 static bool addr_same(const struct inetpeer_addr *a,
85 		      const struct inetpeer_addr *b)
86 {
87 	const struct in6_addr *a6, *b6;
88 
89 	if (a->family != b->family)
90 		return false;
91 	if (a->family == AF_INET)
92 		return a->addr.a4 == b->addr.a4;
93 
94 	a6 = (const struct in6_addr *) &a->addr.a6[0];
95 	b6 = (const struct in6_addr *) &b->addr.a6[0];
96 
97 	return ipv6_addr_equal(a6, b6);
98 }
99 
100 struct tcpm_hash_bucket {
101 	struct tcp_metrics_block __rcu	*chain;
102 };
103 
104 static DEFINE_SPINLOCK(tcp_metrics_lock);
105 
106 static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst)
107 {
108 	u32 val;
109 
110 	tm->tcpm_stamp = jiffies;
111 
112 	val = 0;
113 	if (dst_metric_locked(dst, RTAX_RTT))
114 		val |= 1 << TCP_METRIC_RTT;
115 	if (dst_metric_locked(dst, RTAX_RTTVAR))
116 		val |= 1 << TCP_METRIC_RTTVAR;
117 	if (dst_metric_locked(dst, RTAX_SSTHRESH))
118 		val |= 1 << TCP_METRIC_SSTHRESH;
119 	if (dst_metric_locked(dst, RTAX_CWND))
120 		val |= 1 << TCP_METRIC_CWND;
121 	if (dst_metric_locked(dst, RTAX_REORDERING))
122 		val |= 1 << TCP_METRIC_REORDERING;
123 	tm->tcpm_lock = val;
124 
125 	tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT);
126 	tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR);
127 	tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
128 	tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
129 	tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
130 	tm->tcpm_ts = 0;
131 	tm->tcpm_ts_stamp = 0;
132 	tm->tcpm_fastopen.mss = 0;
133 	tm->tcpm_fastopen.syn_loss = 0;
134 	tm->tcpm_fastopen.cookie.len = 0;
135 }
136 
137 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
138 					  struct inetpeer_addr *addr,
139 					  unsigned int hash,
140 					  bool reclaim)
141 {
142 	struct tcp_metrics_block *tm;
143 	struct net *net;
144 
145 	spin_lock_bh(&tcp_metrics_lock);
146 	net = dev_net(dst->dev);
147 	if (unlikely(reclaim)) {
148 		struct tcp_metrics_block *oldest;
149 
150 		oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
151 		for (tm = rcu_dereference(oldest->tcpm_next); tm;
152 		     tm = rcu_dereference(tm->tcpm_next)) {
153 			if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
154 				oldest = tm;
155 		}
156 		tm = oldest;
157 	} else {
158 		tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
159 		if (!tm)
160 			goto out_unlock;
161 	}
162 	tm->tcpm_addr = *addr;
163 
164 	tcpm_suck_dst(tm, dst);
165 
166 	if (likely(!reclaim)) {
167 		tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
168 		rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
169 	}
170 
171 out_unlock:
172 	spin_unlock_bh(&tcp_metrics_lock);
173 	return tm;
174 }
175 
176 #define TCP_METRICS_TIMEOUT		(60 * 60 * HZ)
177 
178 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
179 {
180 	if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
181 		tcpm_suck_dst(tm, dst);
182 }
183 
184 #define TCP_METRICS_RECLAIM_DEPTH	5
185 #define TCP_METRICS_RECLAIM_PTR		(struct tcp_metrics_block *) 0x1UL
186 
187 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
188 {
189 	if (tm)
190 		return tm;
191 	if (depth > TCP_METRICS_RECLAIM_DEPTH)
192 		return TCP_METRICS_RECLAIM_PTR;
193 	return NULL;
194 }
195 
196 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *addr,
197 						   struct net *net, unsigned int hash)
198 {
199 	struct tcp_metrics_block *tm;
200 	int depth = 0;
201 
202 	for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
203 	     tm = rcu_dereference(tm->tcpm_next)) {
204 		if (addr_same(&tm->tcpm_addr, addr))
205 			break;
206 		depth++;
207 	}
208 	return tcp_get_encode(tm, depth);
209 }
210 
211 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
212 						       struct dst_entry *dst)
213 {
214 	struct tcp_metrics_block *tm;
215 	struct inetpeer_addr addr;
216 	unsigned int hash;
217 	struct net *net;
218 
219 	addr.family = req->rsk_ops->family;
220 	switch (addr.family) {
221 	case AF_INET:
222 		addr.addr.a4 = inet_rsk(req)->rmt_addr;
223 		hash = (__force unsigned int) addr.addr.a4;
224 		break;
225 	case AF_INET6:
226 		*(struct in6_addr *)addr.addr.a6 = inet6_rsk(req)->rmt_addr;
227 		hash = ipv6_addr_hash(&inet6_rsk(req)->rmt_addr);
228 		break;
229 	default:
230 		return NULL;
231 	}
232 
233 	net = dev_net(dst->dev);
234 	hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
235 
236 	for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
237 	     tm = rcu_dereference(tm->tcpm_next)) {
238 		if (addr_same(&tm->tcpm_addr, &addr))
239 			break;
240 	}
241 	tcpm_check_stamp(tm, dst);
242 	return tm;
243 }
244 
245 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
246 {
247 	struct inet6_timewait_sock *tw6;
248 	struct tcp_metrics_block *tm;
249 	struct inetpeer_addr addr;
250 	unsigned int hash;
251 	struct net *net;
252 
253 	addr.family = tw->tw_family;
254 	switch (addr.family) {
255 	case AF_INET:
256 		addr.addr.a4 = tw->tw_daddr;
257 		hash = (__force unsigned int) addr.addr.a4;
258 		break;
259 	case AF_INET6:
260 		tw6 = inet6_twsk((struct sock *)tw);
261 		*(struct in6_addr *)addr.addr.a6 = tw6->tw_v6_daddr;
262 		hash = ipv6_addr_hash(&tw6->tw_v6_daddr);
263 		break;
264 	default:
265 		return NULL;
266 	}
267 
268 	net = twsk_net(tw);
269 	hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
270 
271 	for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
272 	     tm = rcu_dereference(tm->tcpm_next)) {
273 		if (addr_same(&tm->tcpm_addr, &addr))
274 			break;
275 	}
276 	return tm;
277 }
278 
279 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
280 						 struct dst_entry *dst,
281 						 bool create)
282 {
283 	struct tcp_metrics_block *tm;
284 	struct inetpeer_addr addr;
285 	unsigned int hash;
286 	struct net *net;
287 	bool reclaim;
288 
289 	addr.family = sk->sk_family;
290 	switch (addr.family) {
291 	case AF_INET:
292 		addr.addr.a4 = inet_sk(sk)->inet_daddr;
293 		hash = (__force unsigned int) addr.addr.a4;
294 		break;
295 	case AF_INET6:
296 		*(struct in6_addr *)addr.addr.a6 = inet6_sk(sk)->daddr;
297 		hash = ipv6_addr_hash(&inet6_sk(sk)->daddr);
298 		break;
299 	default:
300 		return NULL;
301 	}
302 
303 	net = dev_net(dst->dev);
304 	hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
305 
306 	tm = __tcp_get_metrics(&addr, net, hash);
307 	reclaim = false;
308 	if (tm == TCP_METRICS_RECLAIM_PTR) {
309 		reclaim = true;
310 		tm = NULL;
311 	}
312 	if (!tm && create)
313 		tm = tcpm_new(dst, &addr, hash, reclaim);
314 	else
315 		tcpm_check_stamp(tm, dst);
316 
317 	return tm;
318 }
319 
320 /* Save metrics learned by this TCP session.  This function is called
321  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
322  * or goes from LAST-ACK to CLOSE.
323  */
324 void tcp_update_metrics(struct sock *sk)
325 {
326 	const struct inet_connection_sock *icsk = inet_csk(sk);
327 	struct dst_entry *dst = __sk_dst_get(sk);
328 	struct tcp_sock *tp = tcp_sk(sk);
329 	struct tcp_metrics_block *tm;
330 	unsigned long rtt;
331 	u32 val;
332 	int m;
333 
334 	if (sysctl_tcp_nometrics_save || !dst)
335 		return;
336 
337 	if (dst->flags & DST_HOST)
338 		dst_confirm(dst);
339 
340 	rcu_read_lock();
341 	if (icsk->icsk_backoff || !tp->srtt) {
342 		/* This session failed to estimate rtt. Why?
343 		 * Probably, no packets returned in time.  Reset our
344 		 * results.
345 		 */
346 		tm = tcp_get_metrics(sk, dst, false);
347 		if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
348 			tcp_metric_set(tm, TCP_METRIC_RTT, 0);
349 		goto out_unlock;
350 	} else
351 		tm = tcp_get_metrics(sk, dst, true);
352 
353 	if (!tm)
354 		goto out_unlock;
355 
356 	rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
357 	m = rtt - tp->srtt;
358 
359 	/* If newly calculated rtt larger than stored one, store new
360 	 * one. Otherwise, use EWMA. Remember, rtt overestimation is
361 	 * always better than underestimation.
362 	 */
363 	if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
364 		if (m <= 0)
365 			rtt = tp->srtt;
366 		else
367 			rtt -= (m >> 3);
368 		tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt);
369 	}
370 
371 	if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
372 		unsigned long var;
373 
374 		if (m < 0)
375 			m = -m;
376 
377 		/* Scale deviation to rttvar fixed point */
378 		m >>= 1;
379 		if (m < tp->mdev)
380 			m = tp->mdev;
381 
382 		var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
383 		if (m >= var)
384 			var = m;
385 		else
386 			var -= (var - m) >> 2;
387 
388 		tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var);
389 	}
390 
391 	if (tcp_in_initial_slowstart(tp)) {
392 		/* Slow start still did not finish. */
393 		if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
394 			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
395 			if (val && (tp->snd_cwnd >> 1) > val)
396 				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
397 					       tp->snd_cwnd >> 1);
398 		}
399 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
400 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
401 			if (tp->snd_cwnd > val)
402 				tcp_metric_set(tm, TCP_METRIC_CWND,
403 					       tp->snd_cwnd);
404 		}
405 	} else if (tp->snd_cwnd > tp->snd_ssthresh &&
406 		   icsk->icsk_ca_state == TCP_CA_Open) {
407 		/* Cong. avoidance phase, cwnd is reliable. */
408 		if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
409 			tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
410 				       max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
411 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
412 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
413 			tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
414 		}
415 	} else {
416 		/* Else slow start did not finish, cwnd is non-sense,
417 		 * ssthresh may be also invalid.
418 		 */
419 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
420 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
421 			tcp_metric_set(tm, TCP_METRIC_CWND,
422 				       (val + tp->snd_ssthresh) >> 1);
423 		}
424 		if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
425 			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
426 			if (val && tp->snd_ssthresh > val)
427 				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
428 					       tp->snd_ssthresh);
429 		}
430 		if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
431 			val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
432 			if (val < tp->reordering &&
433 			    tp->reordering != sysctl_tcp_reordering)
434 				tcp_metric_set(tm, TCP_METRIC_REORDERING,
435 					       tp->reordering);
436 		}
437 	}
438 	tm->tcpm_stamp = jiffies;
439 out_unlock:
440 	rcu_read_unlock();
441 }
442 
443 /* Initialize metrics on socket. */
444 
445 void tcp_init_metrics(struct sock *sk)
446 {
447 	struct dst_entry *dst = __sk_dst_get(sk);
448 	struct tcp_sock *tp = tcp_sk(sk);
449 	struct tcp_metrics_block *tm;
450 	u32 val;
451 
452 	if (dst == NULL)
453 		goto reset;
454 
455 	dst_confirm(dst);
456 
457 	rcu_read_lock();
458 	tm = tcp_get_metrics(sk, dst, true);
459 	if (!tm) {
460 		rcu_read_unlock();
461 		goto reset;
462 	}
463 
464 	if (tcp_metric_locked(tm, TCP_METRIC_CWND))
465 		tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
466 
467 	val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
468 	if (val) {
469 		tp->snd_ssthresh = val;
470 		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
471 			tp->snd_ssthresh = tp->snd_cwnd_clamp;
472 	} else {
473 		/* ssthresh may have been reduced unnecessarily during.
474 		 * 3WHS. Restore it back to its initial default.
475 		 */
476 		tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
477 	}
478 	val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
479 	if (val && tp->reordering != val) {
480 		tcp_disable_fack(tp);
481 		tcp_disable_early_retrans(tp);
482 		tp->reordering = val;
483 	}
484 
485 	val = tcp_metric_get(tm, TCP_METRIC_RTT);
486 	if (val == 0 || tp->srtt == 0) {
487 		rcu_read_unlock();
488 		goto reset;
489 	}
490 	/* Initial rtt is determined from SYN,SYN-ACK.
491 	 * The segment is small and rtt may appear much
492 	 * less than real one. Use per-dst memory
493 	 * to make it more realistic.
494 	 *
495 	 * A bit of theory. RTT is time passed after "normal" sized packet
496 	 * is sent until it is ACKed. In normal circumstances sending small
497 	 * packets force peer to delay ACKs and calculation is correct too.
498 	 * The algorithm is adaptive and, provided we follow specs, it
499 	 * NEVER underestimate RTT. BUT! If peer tries to make some clever
500 	 * tricks sort of "quick acks" for time long enough to decrease RTT
501 	 * to low value, and then abruptly stops to do it and starts to delay
502 	 * ACKs, wait for troubles.
503 	 */
504 	val = msecs_to_jiffies(val);
505 	if (val > tp->srtt) {
506 		tp->srtt = val;
507 		tp->rtt_seq = tp->snd_nxt;
508 	}
509 	val = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
510 	if (val > tp->mdev) {
511 		tp->mdev = val;
512 		tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
513 	}
514 	rcu_read_unlock();
515 
516 	tcp_set_rto(sk);
517 reset:
518 	if (tp->srtt == 0) {
519 		/* RFC6298: 5.7 We've failed to get a valid RTT sample from
520 		 * 3WHS. This is most likely due to retransmission,
521 		 * including spurious one. Reset the RTO back to 3secs
522 		 * from the more aggressive 1sec to avoid more spurious
523 		 * retransmission.
524 		 */
525 		tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
526 		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
527 	}
528 	/* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
529 	 * retransmitted. In light of RFC6298 more aggressive 1sec
530 	 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
531 	 * retransmission has occurred.
532 	 */
533 	if (tp->total_retrans > 1)
534 		tp->snd_cwnd = 1;
535 	else
536 		tp->snd_cwnd = tcp_init_cwnd(tp, dst);
537 	tp->snd_cwnd_stamp = tcp_time_stamp;
538 }
539 
540 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
541 {
542 	struct tcp_metrics_block *tm;
543 	bool ret;
544 
545 	if (!dst)
546 		return false;
547 
548 	rcu_read_lock();
549 	tm = __tcp_get_metrics_req(req, dst);
550 	if (paws_check) {
551 		if (tm &&
552 		    (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
553 		    (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
554 			ret = false;
555 		else
556 			ret = true;
557 	} else {
558 		if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
559 			ret = true;
560 		else
561 			ret = false;
562 	}
563 	rcu_read_unlock();
564 
565 	return ret;
566 }
567 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
568 
569 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
570 {
571 	struct tcp_metrics_block *tm;
572 
573 	rcu_read_lock();
574 	tm = tcp_get_metrics(sk, dst, true);
575 	if (tm) {
576 		struct tcp_sock *tp = tcp_sk(sk);
577 
578 		if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
579 			tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
580 			tp->rx_opt.ts_recent = tm->tcpm_ts;
581 		}
582 	}
583 	rcu_read_unlock();
584 }
585 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
586 
587 /* VJ's idea. Save last timestamp seen from this destination and hold
588  * it at least for normal timewait interval to use for duplicate
589  * segment detection in subsequent connections, before they enter
590  * synchronized state.
591  */
592 bool tcp_remember_stamp(struct sock *sk)
593 {
594 	struct dst_entry *dst = __sk_dst_get(sk);
595 	bool ret = false;
596 
597 	if (dst) {
598 		struct tcp_metrics_block *tm;
599 
600 		rcu_read_lock();
601 		tm = tcp_get_metrics(sk, dst, true);
602 		if (tm) {
603 			struct tcp_sock *tp = tcp_sk(sk);
604 
605 			if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
606 			    ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
607 			     tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
608 				tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
609 				tm->tcpm_ts = tp->rx_opt.ts_recent;
610 			}
611 			ret = true;
612 		}
613 		rcu_read_unlock();
614 	}
615 	return ret;
616 }
617 
618 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
619 {
620 	struct tcp_metrics_block *tm;
621 	bool ret = false;
622 
623 	rcu_read_lock();
624 	tm = __tcp_get_metrics_tw(tw);
625 	if (tm) {
626 		const struct tcp_timewait_sock *tcptw;
627 		struct sock *sk = (struct sock *) tw;
628 
629 		tcptw = tcp_twsk(sk);
630 		if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
631 		    ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
632 		     tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
633 			tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
634 			tm->tcpm_ts	   = tcptw->tw_ts_recent;
635 		}
636 		ret = true;
637 	}
638 	rcu_read_unlock();
639 
640 	return ret;
641 }
642 
643 static DEFINE_SEQLOCK(fastopen_seqlock);
644 
645 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
646 			    struct tcp_fastopen_cookie *cookie,
647 			    int *syn_loss, unsigned long *last_syn_loss)
648 {
649 	struct tcp_metrics_block *tm;
650 
651 	rcu_read_lock();
652 	tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
653 	if (tm) {
654 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
655 		unsigned int seq;
656 
657 		do {
658 			seq = read_seqbegin(&fastopen_seqlock);
659 			if (tfom->mss)
660 				*mss = tfom->mss;
661 			*cookie = tfom->cookie;
662 			*syn_loss = tfom->syn_loss;
663 			*last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
664 		} while (read_seqretry(&fastopen_seqlock, seq));
665 	}
666 	rcu_read_unlock();
667 }
668 
669 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
670 			    struct tcp_fastopen_cookie *cookie, bool syn_lost)
671 {
672 	struct tcp_metrics_block *tm;
673 
674 	rcu_read_lock();
675 	tm = tcp_get_metrics(sk, __sk_dst_get(sk), true);
676 	if (tm) {
677 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
678 
679 		write_seqlock_bh(&fastopen_seqlock);
680 		tfom->mss = mss;
681 		if (cookie->len > 0)
682 			tfom->cookie = *cookie;
683 		if (syn_lost) {
684 			++tfom->syn_loss;
685 			tfom->last_syn_loss = jiffies;
686 		} else
687 			tfom->syn_loss = 0;
688 		write_sequnlock_bh(&fastopen_seqlock);
689 	}
690 	rcu_read_unlock();
691 }
692 
693 static unsigned int tcpmhash_entries;
694 static int __init set_tcpmhash_entries(char *str)
695 {
696 	ssize_t ret;
697 
698 	if (!str)
699 		return 0;
700 
701 	ret = kstrtouint(str, 0, &tcpmhash_entries);
702 	if (ret)
703 		return 0;
704 
705 	return 1;
706 }
707 __setup("tcpmhash_entries=", set_tcpmhash_entries);
708 
709 static int __net_init tcp_net_metrics_init(struct net *net)
710 {
711 	size_t size;
712 	unsigned int slots;
713 
714 	slots = tcpmhash_entries;
715 	if (!slots) {
716 		if (totalram_pages >= 128 * 1024)
717 			slots = 16 * 1024;
718 		else
719 			slots = 8 * 1024;
720 	}
721 
722 	net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
723 	size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
724 
725 	net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL);
726 	if (!net->ipv4.tcp_metrics_hash)
727 		return -ENOMEM;
728 
729 	return 0;
730 }
731 
732 static void __net_exit tcp_net_metrics_exit(struct net *net)
733 {
734 	unsigned int i;
735 
736 	for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
737 		struct tcp_metrics_block *tm, *next;
738 
739 		tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
740 		while (tm) {
741 			next = rcu_dereference_protected(tm->tcpm_next, 1);
742 			kfree(tm);
743 			tm = next;
744 		}
745 	}
746 	kfree(net->ipv4.tcp_metrics_hash);
747 }
748 
749 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
750 	.init	=	tcp_net_metrics_init,
751 	.exit	=	tcp_net_metrics_exit,
752 };
753 
754 void __init tcp_metrics_init(void)
755 {
756 	register_pernet_subsys(&tcp_net_metrics_ops);
757 }
758