xref: /linux/net/ipv4/tcp_metrics.c (revision c4ee0af3fa0dc65f690fc908f02b8355f9576ea0)
1 #include <linux/rcupdate.h>
2 #include <linux/spinlock.h>
3 #include <linux/jiffies.h>
4 #include <linux/module.h>
5 #include <linux/cache.h>
6 #include <linux/slab.h>
7 #include <linux/init.h>
8 #include <linux/tcp.h>
9 #include <linux/hash.h>
10 #include <linux/tcp_metrics.h>
11 #include <linux/vmalloc.h>
12 
13 #include <net/inet_connection_sock.h>
14 #include <net/net_namespace.h>
15 #include <net/request_sock.h>
16 #include <net/inetpeer.h>
17 #include <net/sock.h>
18 #include <net/ipv6.h>
19 #include <net/dst.h>
20 #include <net/tcp.h>
21 #include <net/genetlink.h>
22 
23 int sysctl_tcp_nometrics_save __read_mostly;
24 
25 struct tcp_fastopen_metrics {
26 	u16	mss;
27 	u16	syn_loss:10;		/* Recurring Fast Open SYN losses */
28 	unsigned long	last_syn_loss;	/* Last Fast Open SYN loss */
29 	struct	tcp_fastopen_cookie	cookie;
30 };
31 
32 struct tcp_metrics_block {
33 	struct tcp_metrics_block __rcu	*tcpm_next;
34 	struct inetpeer_addr		tcpm_addr;
35 	unsigned long			tcpm_stamp;
36 	u32				tcpm_ts;
37 	u32				tcpm_ts_stamp;
38 	u32				tcpm_lock;
39 	u32				tcpm_vals[TCP_METRIC_MAX + 1];
40 	struct tcp_fastopen_metrics	tcpm_fastopen;
41 
42 	struct rcu_head			rcu_head;
43 };
44 
45 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
46 			      enum tcp_metric_index idx)
47 {
48 	return tm->tcpm_lock & (1 << idx);
49 }
50 
51 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
52 			  enum tcp_metric_index idx)
53 {
54 	return tm->tcpm_vals[idx];
55 }
56 
57 static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm,
58 				  enum tcp_metric_index idx)
59 {
60 	return msecs_to_jiffies(tm->tcpm_vals[idx]);
61 }
62 
63 static void tcp_metric_set(struct tcp_metrics_block *tm,
64 			   enum tcp_metric_index idx,
65 			   u32 val)
66 {
67 	tm->tcpm_vals[idx] = val;
68 }
69 
70 static void tcp_metric_set_msecs(struct tcp_metrics_block *tm,
71 				 enum tcp_metric_index idx,
72 				 u32 val)
73 {
74 	tm->tcpm_vals[idx] = jiffies_to_msecs(val);
75 }
76 
77 static bool addr_same(const struct inetpeer_addr *a,
78 		      const struct inetpeer_addr *b)
79 {
80 	const struct in6_addr *a6, *b6;
81 
82 	if (a->family != b->family)
83 		return false;
84 	if (a->family == AF_INET)
85 		return a->addr.a4 == b->addr.a4;
86 
87 	a6 = (const struct in6_addr *) &a->addr.a6[0];
88 	b6 = (const struct in6_addr *) &b->addr.a6[0];
89 
90 	return ipv6_addr_equal(a6, b6);
91 }
92 
93 struct tcpm_hash_bucket {
94 	struct tcp_metrics_block __rcu	*chain;
95 };
96 
97 static DEFINE_SPINLOCK(tcp_metrics_lock);
98 
99 static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst,
100 			  bool fastopen_clear)
101 {
102 	u32 val;
103 
104 	tm->tcpm_stamp = jiffies;
105 
106 	val = 0;
107 	if (dst_metric_locked(dst, RTAX_RTT))
108 		val |= 1 << TCP_METRIC_RTT;
109 	if (dst_metric_locked(dst, RTAX_RTTVAR))
110 		val |= 1 << TCP_METRIC_RTTVAR;
111 	if (dst_metric_locked(dst, RTAX_SSTHRESH))
112 		val |= 1 << TCP_METRIC_SSTHRESH;
113 	if (dst_metric_locked(dst, RTAX_CWND))
114 		val |= 1 << TCP_METRIC_CWND;
115 	if (dst_metric_locked(dst, RTAX_REORDERING))
116 		val |= 1 << TCP_METRIC_REORDERING;
117 	tm->tcpm_lock = val;
118 
119 	tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT);
120 	tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR);
121 	tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
122 	tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
123 	tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
124 	tm->tcpm_ts = 0;
125 	tm->tcpm_ts_stamp = 0;
126 	if (fastopen_clear) {
127 		tm->tcpm_fastopen.mss = 0;
128 		tm->tcpm_fastopen.syn_loss = 0;
129 		tm->tcpm_fastopen.cookie.len = 0;
130 	}
131 }
132 
133 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
134 					  struct inetpeer_addr *addr,
135 					  unsigned int hash,
136 					  bool reclaim)
137 {
138 	struct tcp_metrics_block *tm;
139 	struct net *net;
140 
141 	spin_lock_bh(&tcp_metrics_lock);
142 	net = dev_net(dst->dev);
143 	if (unlikely(reclaim)) {
144 		struct tcp_metrics_block *oldest;
145 
146 		oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
147 		for (tm = rcu_dereference(oldest->tcpm_next); tm;
148 		     tm = rcu_dereference(tm->tcpm_next)) {
149 			if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
150 				oldest = tm;
151 		}
152 		tm = oldest;
153 	} else {
154 		tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
155 		if (!tm)
156 			goto out_unlock;
157 	}
158 	tm->tcpm_addr = *addr;
159 
160 	tcpm_suck_dst(tm, dst, true);
161 
162 	if (likely(!reclaim)) {
163 		tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
164 		rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
165 	}
166 
167 out_unlock:
168 	spin_unlock_bh(&tcp_metrics_lock);
169 	return tm;
170 }
171 
172 #define TCP_METRICS_TIMEOUT		(60 * 60 * HZ)
173 
174 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
175 {
176 	if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
177 		tcpm_suck_dst(tm, dst, false);
178 }
179 
180 #define TCP_METRICS_RECLAIM_DEPTH	5
181 #define TCP_METRICS_RECLAIM_PTR		(struct tcp_metrics_block *) 0x1UL
182 
183 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
184 {
185 	if (tm)
186 		return tm;
187 	if (depth > TCP_METRICS_RECLAIM_DEPTH)
188 		return TCP_METRICS_RECLAIM_PTR;
189 	return NULL;
190 }
191 
192 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *addr,
193 						   struct net *net, unsigned int hash)
194 {
195 	struct tcp_metrics_block *tm;
196 	int depth = 0;
197 
198 	for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
199 	     tm = rcu_dereference(tm->tcpm_next)) {
200 		if (addr_same(&tm->tcpm_addr, addr))
201 			break;
202 		depth++;
203 	}
204 	return tcp_get_encode(tm, depth);
205 }
206 
207 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
208 						       struct dst_entry *dst)
209 {
210 	struct tcp_metrics_block *tm;
211 	struct inetpeer_addr addr;
212 	unsigned int hash;
213 	struct net *net;
214 
215 	addr.family = req->rsk_ops->family;
216 	switch (addr.family) {
217 	case AF_INET:
218 		addr.addr.a4 = inet_rsk(req)->ir_rmt_addr;
219 		hash = (__force unsigned int) addr.addr.a4;
220 		break;
221 #if IS_ENABLED(CONFIG_IPV6)
222 	case AF_INET6:
223 		*(struct in6_addr *)addr.addr.a6 = inet_rsk(req)->ir_v6_rmt_addr;
224 		hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
225 		break;
226 #endif
227 	default:
228 		return NULL;
229 	}
230 
231 	net = dev_net(dst->dev);
232 	hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
233 
234 	for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
235 	     tm = rcu_dereference(tm->tcpm_next)) {
236 		if (addr_same(&tm->tcpm_addr, &addr))
237 			break;
238 	}
239 	tcpm_check_stamp(tm, dst);
240 	return tm;
241 }
242 
243 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
244 {
245 	struct tcp_metrics_block *tm;
246 	struct inetpeer_addr addr;
247 	unsigned int hash;
248 	struct net *net;
249 
250 	addr.family = tw->tw_family;
251 	switch (addr.family) {
252 	case AF_INET:
253 		addr.addr.a4 = tw->tw_daddr;
254 		hash = (__force unsigned int) addr.addr.a4;
255 		break;
256 #if IS_ENABLED(CONFIG_IPV6)
257 	case AF_INET6:
258 		*(struct in6_addr *)addr.addr.a6 = tw->tw_v6_daddr;
259 		hash = ipv6_addr_hash(&tw->tw_v6_daddr);
260 		break;
261 #endif
262 	default:
263 		return NULL;
264 	}
265 
266 	net = twsk_net(tw);
267 	hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
268 
269 	for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
270 	     tm = rcu_dereference(tm->tcpm_next)) {
271 		if (addr_same(&tm->tcpm_addr, &addr))
272 			break;
273 	}
274 	return tm;
275 }
276 
277 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
278 						 struct dst_entry *dst,
279 						 bool create)
280 {
281 	struct tcp_metrics_block *tm;
282 	struct inetpeer_addr addr;
283 	unsigned int hash;
284 	struct net *net;
285 	bool reclaim;
286 
287 	addr.family = sk->sk_family;
288 	switch (addr.family) {
289 	case AF_INET:
290 		addr.addr.a4 = inet_sk(sk)->inet_daddr;
291 		hash = (__force unsigned int) addr.addr.a4;
292 		break;
293 #if IS_ENABLED(CONFIG_IPV6)
294 	case AF_INET6:
295 		*(struct in6_addr *)addr.addr.a6 = sk->sk_v6_daddr;
296 		hash = ipv6_addr_hash(&sk->sk_v6_daddr);
297 		break;
298 #endif
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, crtt = 0; /* cached RTT scaled by 8 */
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 	crtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
486 	rcu_read_unlock();
487 reset:
488 	/* The initial RTT measurement from the SYN/SYN-ACK is not ideal
489 	 * to seed the RTO for later data packets because SYN packets are
490 	 * small. Use the per-dst cached values to seed the RTO but keep
491 	 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
492 	 * Later the RTO will be updated immediately upon obtaining the first
493 	 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
494 	 * influences the first RTO but not later RTT estimation.
495 	 *
496 	 * But if RTT is not available from the SYN (due to retransmits or
497 	 * syn cookies) or the cache, force a conservative 3secs timeout.
498 	 *
499 	 * A bit of theory. RTT is time passed after "normal" sized packet
500 	 * is sent until it is ACKed. In normal circumstances sending small
501 	 * packets force peer to delay ACKs and calculation is correct too.
502 	 * The algorithm is adaptive and, provided we follow specs, it
503 	 * NEVER underestimate RTT. BUT! If peer tries to make some clever
504 	 * tricks sort of "quick acks" for time long enough to decrease RTT
505 	 * to low value, and then abruptly stops to do it and starts to delay
506 	 * ACKs, wait for troubles.
507 	 */
508 	if (crtt > tp->srtt) {
509 		/* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
510 		crtt >>= 3;
511 		inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
512 	} else if (tp->srtt == 0) {
513 		/* RFC6298: 5.7 We've failed to get a valid RTT sample from
514 		 * 3WHS. This is most likely due to retransmission,
515 		 * including spurious one. Reset the RTO back to 3secs
516 		 * from the more aggressive 1sec to avoid more spurious
517 		 * retransmission.
518 		 */
519 		tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
520 		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
521 	}
522 	/* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
523 	 * retransmitted. In light of RFC6298 more aggressive 1sec
524 	 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
525 	 * retransmission has occurred.
526 	 */
527 	if (tp->total_retrans > 1)
528 		tp->snd_cwnd = 1;
529 	else
530 		tp->snd_cwnd = tcp_init_cwnd(tp, dst);
531 	tp->snd_cwnd_stamp = tcp_time_stamp;
532 }
533 
534 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
535 {
536 	struct tcp_metrics_block *tm;
537 	bool ret;
538 
539 	if (!dst)
540 		return false;
541 
542 	rcu_read_lock();
543 	tm = __tcp_get_metrics_req(req, dst);
544 	if (paws_check) {
545 		if (tm &&
546 		    (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
547 		    (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
548 			ret = false;
549 		else
550 			ret = true;
551 	} else {
552 		if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
553 			ret = true;
554 		else
555 			ret = false;
556 	}
557 	rcu_read_unlock();
558 
559 	return ret;
560 }
561 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
562 
563 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
564 {
565 	struct tcp_metrics_block *tm;
566 
567 	rcu_read_lock();
568 	tm = tcp_get_metrics(sk, dst, true);
569 	if (tm) {
570 		struct tcp_sock *tp = tcp_sk(sk);
571 
572 		if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
573 			tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
574 			tp->rx_opt.ts_recent = tm->tcpm_ts;
575 		}
576 	}
577 	rcu_read_unlock();
578 }
579 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
580 
581 /* VJ's idea. Save last timestamp seen from this destination and hold
582  * it at least for normal timewait interval to use for duplicate
583  * segment detection in subsequent connections, before they enter
584  * synchronized state.
585  */
586 bool tcp_remember_stamp(struct sock *sk)
587 {
588 	struct dst_entry *dst = __sk_dst_get(sk);
589 	bool ret = false;
590 
591 	if (dst) {
592 		struct tcp_metrics_block *tm;
593 
594 		rcu_read_lock();
595 		tm = tcp_get_metrics(sk, dst, true);
596 		if (tm) {
597 			struct tcp_sock *tp = tcp_sk(sk);
598 
599 			if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
600 			    ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
601 			     tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
602 				tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
603 				tm->tcpm_ts = tp->rx_opt.ts_recent;
604 			}
605 			ret = true;
606 		}
607 		rcu_read_unlock();
608 	}
609 	return ret;
610 }
611 
612 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
613 {
614 	struct tcp_metrics_block *tm;
615 	bool ret = false;
616 
617 	rcu_read_lock();
618 	tm = __tcp_get_metrics_tw(tw);
619 	if (tm) {
620 		const struct tcp_timewait_sock *tcptw;
621 		struct sock *sk = (struct sock *) tw;
622 
623 		tcptw = tcp_twsk(sk);
624 		if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
625 		    ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
626 		     tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
627 			tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
628 			tm->tcpm_ts	   = tcptw->tw_ts_recent;
629 		}
630 		ret = true;
631 	}
632 	rcu_read_unlock();
633 
634 	return ret;
635 }
636 
637 static DEFINE_SEQLOCK(fastopen_seqlock);
638 
639 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
640 			    struct tcp_fastopen_cookie *cookie,
641 			    int *syn_loss, unsigned long *last_syn_loss)
642 {
643 	struct tcp_metrics_block *tm;
644 
645 	rcu_read_lock();
646 	tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
647 	if (tm) {
648 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
649 		unsigned int seq;
650 
651 		do {
652 			seq = read_seqbegin(&fastopen_seqlock);
653 			if (tfom->mss)
654 				*mss = tfom->mss;
655 			*cookie = tfom->cookie;
656 			*syn_loss = tfom->syn_loss;
657 			*last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
658 		} while (read_seqretry(&fastopen_seqlock, seq));
659 	}
660 	rcu_read_unlock();
661 }
662 
663 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
664 			    struct tcp_fastopen_cookie *cookie, bool syn_lost)
665 {
666 	struct dst_entry *dst = __sk_dst_get(sk);
667 	struct tcp_metrics_block *tm;
668 
669 	if (!dst)
670 		return;
671 	rcu_read_lock();
672 	tm = tcp_get_metrics(sk, dst, true);
673 	if (tm) {
674 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
675 
676 		write_seqlock_bh(&fastopen_seqlock);
677 		if (mss)
678 			tfom->mss = mss;
679 		if (cookie && cookie->len > 0)
680 			tfom->cookie = *cookie;
681 		if (syn_lost) {
682 			++tfom->syn_loss;
683 			tfom->last_syn_loss = jiffies;
684 		} else
685 			tfom->syn_loss = 0;
686 		write_sequnlock_bh(&fastopen_seqlock);
687 	}
688 	rcu_read_unlock();
689 }
690 
691 static struct genl_family tcp_metrics_nl_family = {
692 	.id		= GENL_ID_GENERATE,
693 	.hdrsize	= 0,
694 	.name		= TCP_METRICS_GENL_NAME,
695 	.version	= TCP_METRICS_GENL_VERSION,
696 	.maxattr	= TCP_METRICS_ATTR_MAX,
697 	.netnsok	= true,
698 };
699 
700 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
701 	[TCP_METRICS_ATTR_ADDR_IPV4]	= { .type = NLA_U32, },
702 	[TCP_METRICS_ATTR_ADDR_IPV6]	= { .type = NLA_BINARY,
703 					    .len = sizeof(struct in6_addr), },
704 	/* Following attributes are not received for GET/DEL,
705 	 * we keep them for reference
706 	 */
707 #if 0
708 	[TCP_METRICS_ATTR_AGE]		= { .type = NLA_MSECS, },
709 	[TCP_METRICS_ATTR_TW_TSVAL]	= { .type = NLA_U32, },
710 	[TCP_METRICS_ATTR_TW_TS_STAMP]	= { .type = NLA_S32, },
711 	[TCP_METRICS_ATTR_VALS]		= { .type = NLA_NESTED, },
712 	[TCP_METRICS_ATTR_FOPEN_MSS]	= { .type = NLA_U16, },
713 	[TCP_METRICS_ATTR_FOPEN_SYN_DROPS]	= { .type = NLA_U16, },
714 	[TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]	= { .type = NLA_MSECS, },
715 	[TCP_METRICS_ATTR_FOPEN_COOKIE]	= { .type = NLA_BINARY,
716 					    .len = TCP_FASTOPEN_COOKIE_MAX, },
717 #endif
718 };
719 
720 /* Add attributes, caller cancels its header on failure */
721 static int tcp_metrics_fill_info(struct sk_buff *msg,
722 				 struct tcp_metrics_block *tm)
723 {
724 	struct nlattr *nest;
725 	int i;
726 
727 	switch (tm->tcpm_addr.family) {
728 	case AF_INET:
729 		if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
730 				tm->tcpm_addr.addr.a4) < 0)
731 			goto nla_put_failure;
732 		break;
733 	case AF_INET6:
734 		if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
735 			    tm->tcpm_addr.addr.a6) < 0)
736 			goto nla_put_failure;
737 		break;
738 	default:
739 		return -EAFNOSUPPORT;
740 	}
741 
742 	if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
743 			  jiffies - tm->tcpm_stamp) < 0)
744 		goto nla_put_failure;
745 	if (tm->tcpm_ts_stamp) {
746 		if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
747 				(s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
748 			goto nla_put_failure;
749 		if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
750 				tm->tcpm_ts) < 0)
751 			goto nla_put_failure;
752 	}
753 
754 	{
755 		int n = 0;
756 
757 		nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
758 		if (!nest)
759 			goto nla_put_failure;
760 		for (i = 0; i < TCP_METRIC_MAX + 1; i++) {
761 			if (!tm->tcpm_vals[i])
762 				continue;
763 			if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0)
764 				goto nla_put_failure;
765 			n++;
766 		}
767 		if (n)
768 			nla_nest_end(msg, nest);
769 		else
770 			nla_nest_cancel(msg, nest);
771 	}
772 
773 	{
774 		struct tcp_fastopen_metrics tfom_copy[1], *tfom;
775 		unsigned int seq;
776 
777 		do {
778 			seq = read_seqbegin(&fastopen_seqlock);
779 			tfom_copy[0] = tm->tcpm_fastopen;
780 		} while (read_seqretry(&fastopen_seqlock, seq));
781 
782 		tfom = tfom_copy;
783 		if (tfom->mss &&
784 		    nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
785 				tfom->mss) < 0)
786 			goto nla_put_failure;
787 		if (tfom->syn_loss &&
788 		    (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
789 				tfom->syn_loss) < 0 ||
790 		     nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
791 				jiffies - tfom->last_syn_loss) < 0))
792 			goto nla_put_failure;
793 		if (tfom->cookie.len > 0 &&
794 		    nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
795 			    tfom->cookie.len, tfom->cookie.val) < 0)
796 			goto nla_put_failure;
797 	}
798 
799 	return 0;
800 
801 nla_put_failure:
802 	return -EMSGSIZE;
803 }
804 
805 static int tcp_metrics_dump_info(struct sk_buff *skb,
806 				 struct netlink_callback *cb,
807 				 struct tcp_metrics_block *tm)
808 {
809 	void *hdr;
810 
811 	hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
812 			  &tcp_metrics_nl_family, NLM_F_MULTI,
813 			  TCP_METRICS_CMD_GET);
814 	if (!hdr)
815 		return -EMSGSIZE;
816 
817 	if (tcp_metrics_fill_info(skb, tm) < 0)
818 		goto nla_put_failure;
819 
820 	return genlmsg_end(skb, hdr);
821 
822 nla_put_failure:
823 	genlmsg_cancel(skb, hdr);
824 	return -EMSGSIZE;
825 }
826 
827 static int tcp_metrics_nl_dump(struct sk_buff *skb,
828 			       struct netlink_callback *cb)
829 {
830 	struct net *net = sock_net(skb->sk);
831 	unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
832 	unsigned int row, s_row = cb->args[0];
833 	int s_col = cb->args[1], col = s_col;
834 
835 	for (row = s_row; row < max_rows; row++, s_col = 0) {
836 		struct tcp_metrics_block *tm;
837 		struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
838 
839 		rcu_read_lock();
840 		for (col = 0, tm = rcu_dereference(hb->chain); tm;
841 		     tm = rcu_dereference(tm->tcpm_next), col++) {
842 			if (col < s_col)
843 				continue;
844 			if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
845 				rcu_read_unlock();
846 				goto done;
847 			}
848 		}
849 		rcu_read_unlock();
850 	}
851 
852 done:
853 	cb->args[0] = row;
854 	cb->args[1] = col;
855 	return skb->len;
856 }
857 
858 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
859 			 unsigned int *hash, int optional)
860 {
861 	struct nlattr *a;
862 
863 	a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV4];
864 	if (a) {
865 		addr->family = AF_INET;
866 		addr->addr.a4 = nla_get_be32(a);
867 		*hash = (__force unsigned int) addr->addr.a4;
868 		return 0;
869 	}
870 	a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV6];
871 	if (a) {
872 		if (nla_len(a) != sizeof(struct in6_addr))
873 			return -EINVAL;
874 		addr->family = AF_INET6;
875 		memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
876 		*hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
877 		return 0;
878 	}
879 	return optional ? 1 : -EAFNOSUPPORT;
880 }
881 
882 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
883 {
884 	struct tcp_metrics_block *tm;
885 	struct inetpeer_addr addr;
886 	unsigned int hash;
887 	struct sk_buff *msg;
888 	struct net *net = genl_info_net(info);
889 	void *reply;
890 	int ret;
891 
892 	ret = parse_nl_addr(info, &addr, &hash, 0);
893 	if (ret < 0)
894 		return ret;
895 
896 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
897 	if (!msg)
898 		return -ENOMEM;
899 
900 	reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
901 				  info->genlhdr->cmd);
902 	if (!reply)
903 		goto nla_put_failure;
904 
905 	hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
906 	ret = -ESRCH;
907 	rcu_read_lock();
908 	for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
909 	     tm = rcu_dereference(tm->tcpm_next)) {
910 		if (addr_same(&tm->tcpm_addr, &addr)) {
911 			ret = tcp_metrics_fill_info(msg, tm);
912 			break;
913 		}
914 	}
915 	rcu_read_unlock();
916 	if (ret < 0)
917 		goto out_free;
918 
919 	genlmsg_end(msg, reply);
920 	return genlmsg_reply(msg, info);
921 
922 nla_put_failure:
923 	ret = -EMSGSIZE;
924 
925 out_free:
926 	nlmsg_free(msg);
927 	return ret;
928 }
929 
930 #define deref_locked_genl(p)	\
931 	rcu_dereference_protected(p, lockdep_genl_is_held() && \
932 				     lockdep_is_held(&tcp_metrics_lock))
933 
934 #define deref_genl(p)	rcu_dereference_protected(p, lockdep_genl_is_held())
935 
936 static int tcp_metrics_flush_all(struct net *net)
937 {
938 	unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
939 	struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
940 	struct tcp_metrics_block *tm;
941 	unsigned int row;
942 
943 	for (row = 0; row < max_rows; row++, hb++) {
944 		spin_lock_bh(&tcp_metrics_lock);
945 		tm = deref_locked_genl(hb->chain);
946 		if (tm)
947 			hb->chain = NULL;
948 		spin_unlock_bh(&tcp_metrics_lock);
949 		while (tm) {
950 			struct tcp_metrics_block *next;
951 
952 			next = deref_genl(tm->tcpm_next);
953 			kfree_rcu(tm, rcu_head);
954 			tm = next;
955 		}
956 	}
957 	return 0;
958 }
959 
960 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
961 {
962 	struct tcpm_hash_bucket *hb;
963 	struct tcp_metrics_block *tm;
964 	struct tcp_metrics_block __rcu **pp;
965 	struct inetpeer_addr addr;
966 	unsigned int hash;
967 	struct net *net = genl_info_net(info);
968 	int ret;
969 
970 	ret = parse_nl_addr(info, &addr, &hash, 1);
971 	if (ret < 0)
972 		return ret;
973 	if (ret > 0)
974 		return tcp_metrics_flush_all(net);
975 
976 	hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
977 	hb = net->ipv4.tcp_metrics_hash + hash;
978 	pp = &hb->chain;
979 	spin_lock_bh(&tcp_metrics_lock);
980 	for (tm = deref_locked_genl(*pp); tm;
981 	     pp = &tm->tcpm_next, tm = deref_locked_genl(*pp)) {
982 		if (addr_same(&tm->tcpm_addr, &addr)) {
983 			*pp = tm->tcpm_next;
984 			break;
985 		}
986 	}
987 	spin_unlock_bh(&tcp_metrics_lock);
988 	if (!tm)
989 		return -ESRCH;
990 	kfree_rcu(tm, rcu_head);
991 	return 0;
992 }
993 
994 static const struct genl_ops tcp_metrics_nl_ops[] = {
995 	{
996 		.cmd = TCP_METRICS_CMD_GET,
997 		.doit = tcp_metrics_nl_cmd_get,
998 		.dumpit = tcp_metrics_nl_dump,
999 		.policy = tcp_metrics_nl_policy,
1000 		.flags = GENL_ADMIN_PERM,
1001 	},
1002 	{
1003 		.cmd = TCP_METRICS_CMD_DEL,
1004 		.doit = tcp_metrics_nl_cmd_del,
1005 		.policy = tcp_metrics_nl_policy,
1006 		.flags = GENL_ADMIN_PERM,
1007 	},
1008 };
1009 
1010 static unsigned int tcpmhash_entries;
1011 static int __init set_tcpmhash_entries(char *str)
1012 {
1013 	ssize_t ret;
1014 
1015 	if (!str)
1016 		return 0;
1017 
1018 	ret = kstrtouint(str, 0, &tcpmhash_entries);
1019 	if (ret)
1020 		return 0;
1021 
1022 	return 1;
1023 }
1024 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1025 
1026 static int __net_init tcp_net_metrics_init(struct net *net)
1027 {
1028 	size_t size;
1029 	unsigned int slots;
1030 
1031 	slots = tcpmhash_entries;
1032 	if (!slots) {
1033 		if (totalram_pages >= 128 * 1024)
1034 			slots = 16 * 1024;
1035 		else
1036 			slots = 8 * 1024;
1037 	}
1038 
1039 	net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1040 	size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
1041 
1042 	net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1043 	if (!net->ipv4.tcp_metrics_hash)
1044 		net->ipv4.tcp_metrics_hash = vzalloc(size);
1045 
1046 	if (!net->ipv4.tcp_metrics_hash)
1047 		return -ENOMEM;
1048 
1049 	return 0;
1050 }
1051 
1052 static void __net_exit tcp_net_metrics_exit(struct net *net)
1053 {
1054 	unsigned int i;
1055 
1056 	for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1057 		struct tcp_metrics_block *tm, *next;
1058 
1059 		tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1060 		while (tm) {
1061 			next = rcu_dereference_protected(tm->tcpm_next, 1);
1062 			kfree(tm);
1063 			tm = next;
1064 		}
1065 	}
1066 	if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1067 		vfree(net->ipv4.tcp_metrics_hash);
1068 	else
1069 		kfree(net->ipv4.tcp_metrics_hash);
1070 }
1071 
1072 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1073 	.init	=	tcp_net_metrics_init,
1074 	.exit	=	tcp_net_metrics_exit,
1075 };
1076 
1077 void __init tcp_metrics_init(void)
1078 {
1079 	int ret;
1080 
1081 	ret = register_pernet_subsys(&tcp_net_metrics_ops);
1082 	if (ret < 0)
1083 		goto cleanup;
1084 	ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1085 					    tcp_metrics_nl_ops);
1086 	if (ret < 0)
1087 		goto cleanup_subsys;
1088 	return;
1089 
1090 cleanup_subsys:
1091 	unregister_pernet_subsys(&tcp_net_metrics_ops);
1092 
1093 cleanup:
1094 	return;
1095 }
1096