xref: /linux/net/ipv4/tcp_metrics.c (revision 4f3c8320c78cdd11c8fdd23c33787407f719322e)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/rcupdate.h>
3 #include <linux/spinlock.h>
4 #include <linux/jiffies.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 #include <linux/tcp_metrics.h>
12 #include <linux/vmalloc.h>
13 
14 #include <net/inet_connection_sock.h>
15 #include <net/net_namespace.h>
16 #include <net/request_sock.h>
17 #include <net/inetpeer.h>
18 #include <net/sock.h>
19 #include <net/ipv6.h>
20 #include <net/dst.h>
21 #include <net/tcp.h>
22 #include <net/genetlink.h>
23 
24 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
25 						   const struct inetpeer_addr *daddr,
26 						   struct net *net, unsigned int hash);
27 
28 struct tcp_fastopen_metrics {
29 	u16	mss;
30 	u16	syn_loss:10,		/* Recurring Fast Open SYN losses */
31 		try_exp:2;		/* Request w/ exp. option (once) */
32 	unsigned long	last_syn_loss;	/* Last Fast Open SYN loss */
33 	struct	tcp_fastopen_cookie	cookie;
34 };
35 
36 /* TCP_METRIC_MAX includes 2 extra fields for userspace compatibility
37  * Kernel only stores RTT and RTTVAR in usec resolution
38  */
39 #define TCP_METRIC_MAX_KERNEL (TCP_METRIC_MAX - 2)
40 
41 struct tcp_metrics_block {
42 	struct tcp_metrics_block __rcu	*tcpm_next;
43 	possible_net_t			tcpm_net;
44 	struct inetpeer_addr		tcpm_saddr;
45 	struct inetpeer_addr		tcpm_daddr;
46 	unsigned long			tcpm_stamp;
47 	u32				tcpm_lock;
48 	u32				tcpm_vals[TCP_METRIC_MAX_KERNEL + 1];
49 	struct tcp_fastopen_metrics	tcpm_fastopen;
50 
51 	struct rcu_head			rcu_head;
52 };
53 
54 static inline struct net *tm_net(struct tcp_metrics_block *tm)
55 {
56 	return read_pnet(&tm->tcpm_net);
57 }
58 
59 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
60 			      enum tcp_metric_index idx)
61 {
62 	return tm->tcpm_lock & (1 << idx);
63 }
64 
65 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
66 			  enum tcp_metric_index idx)
67 {
68 	return tm->tcpm_vals[idx];
69 }
70 
71 static void tcp_metric_set(struct tcp_metrics_block *tm,
72 			   enum tcp_metric_index idx,
73 			   u32 val)
74 {
75 	tm->tcpm_vals[idx] = val;
76 }
77 
78 static bool addr_same(const struct inetpeer_addr *a,
79 		      const struct inetpeer_addr *b)
80 {
81 	return inetpeer_addr_cmp(a, b) == 0;
82 }
83 
84 struct tcpm_hash_bucket {
85 	struct tcp_metrics_block __rcu	*chain;
86 };
87 
88 static struct tcpm_hash_bucket	*tcp_metrics_hash __read_mostly;
89 static unsigned int		tcp_metrics_hash_log __read_mostly;
90 
91 static DEFINE_SPINLOCK(tcp_metrics_lock);
92 
93 static void tcpm_suck_dst(struct tcp_metrics_block *tm,
94 			  const struct dst_entry *dst,
95 			  bool fastopen_clear)
96 {
97 	u32 msval;
98 	u32 val;
99 
100 	tm->tcpm_stamp = jiffies;
101 
102 	val = 0;
103 	if (dst_metric_locked(dst, RTAX_RTT))
104 		val |= 1 << TCP_METRIC_RTT;
105 	if (dst_metric_locked(dst, RTAX_RTTVAR))
106 		val |= 1 << TCP_METRIC_RTTVAR;
107 	if (dst_metric_locked(dst, RTAX_SSTHRESH))
108 		val |= 1 << TCP_METRIC_SSTHRESH;
109 	if (dst_metric_locked(dst, RTAX_CWND))
110 		val |= 1 << TCP_METRIC_CWND;
111 	if (dst_metric_locked(dst, RTAX_REORDERING))
112 		val |= 1 << TCP_METRIC_REORDERING;
113 	tm->tcpm_lock = val;
114 
115 	msval = dst_metric_raw(dst, RTAX_RTT);
116 	tm->tcpm_vals[TCP_METRIC_RTT] = msval * USEC_PER_MSEC;
117 
118 	msval = dst_metric_raw(dst, RTAX_RTTVAR);
119 	tm->tcpm_vals[TCP_METRIC_RTTVAR] = msval * USEC_PER_MSEC;
120 	tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
121 	tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
122 	tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
123 	if (fastopen_clear) {
124 		tm->tcpm_fastopen.mss = 0;
125 		tm->tcpm_fastopen.syn_loss = 0;
126 		tm->tcpm_fastopen.try_exp = 0;
127 		tm->tcpm_fastopen.cookie.exp = false;
128 		tm->tcpm_fastopen.cookie.len = 0;
129 	}
130 }
131 
132 #define TCP_METRICS_TIMEOUT		(60 * 60 * HZ)
133 
134 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
135 {
136 	if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
137 		tcpm_suck_dst(tm, dst, false);
138 }
139 
140 #define TCP_METRICS_RECLAIM_DEPTH	5
141 #define TCP_METRICS_RECLAIM_PTR		(struct tcp_metrics_block *) 0x1UL
142 
143 #define deref_locked(p)	\
144 	rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
145 
146 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
147 					  struct inetpeer_addr *saddr,
148 					  struct inetpeer_addr *daddr,
149 					  unsigned int hash)
150 {
151 	struct tcp_metrics_block *tm;
152 	struct net *net;
153 	bool reclaim = false;
154 
155 	spin_lock_bh(&tcp_metrics_lock);
156 	net = dev_net(dst->dev);
157 
158 	/* While waiting for the spin-lock the cache might have been populated
159 	 * with this entry and so we have to check again.
160 	 */
161 	tm = __tcp_get_metrics(saddr, daddr, net, hash);
162 	if (tm == TCP_METRICS_RECLAIM_PTR) {
163 		reclaim = true;
164 		tm = NULL;
165 	}
166 	if (tm) {
167 		tcpm_check_stamp(tm, dst);
168 		goto out_unlock;
169 	}
170 
171 	if (unlikely(reclaim)) {
172 		struct tcp_metrics_block *oldest;
173 
174 		oldest = deref_locked(tcp_metrics_hash[hash].chain);
175 		for (tm = deref_locked(oldest->tcpm_next); tm;
176 		     tm = deref_locked(tm->tcpm_next)) {
177 			if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
178 				oldest = tm;
179 		}
180 		tm = oldest;
181 	} else {
182 		tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
183 		if (!tm)
184 			goto out_unlock;
185 	}
186 	write_pnet(&tm->tcpm_net, net);
187 	tm->tcpm_saddr = *saddr;
188 	tm->tcpm_daddr = *daddr;
189 
190 	tcpm_suck_dst(tm, dst, true);
191 
192 	if (likely(!reclaim)) {
193 		tm->tcpm_next = tcp_metrics_hash[hash].chain;
194 		rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
195 	}
196 
197 out_unlock:
198 	spin_unlock_bh(&tcp_metrics_lock);
199 	return tm;
200 }
201 
202 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
203 {
204 	if (tm)
205 		return tm;
206 	if (depth > TCP_METRICS_RECLAIM_DEPTH)
207 		return TCP_METRICS_RECLAIM_PTR;
208 	return NULL;
209 }
210 
211 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
212 						   const struct inetpeer_addr *daddr,
213 						   struct net *net, unsigned int hash)
214 {
215 	struct tcp_metrics_block *tm;
216 	int depth = 0;
217 
218 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
219 	     tm = rcu_dereference(tm->tcpm_next)) {
220 		if (addr_same(&tm->tcpm_saddr, saddr) &&
221 		    addr_same(&tm->tcpm_daddr, daddr) &&
222 		    net_eq(tm_net(tm), net))
223 			break;
224 		depth++;
225 	}
226 	return tcp_get_encode(tm, depth);
227 }
228 
229 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
230 						       struct dst_entry *dst)
231 {
232 	struct tcp_metrics_block *tm;
233 	struct inetpeer_addr saddr, daddr;
234 	unsigned int hash;
235 	struct net *net;
236 
237 	saddr.family = req->rsk_ops->family;
238 	daddr.family = req->rsk_ops->family;
239 	switch (daddr.family) {
240 	case AF_INET:
241 		inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
242 		inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
243 		hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
244 		break;
245 #if IS_ENABLED(CONFIG_IPV6)
246 	case AF_INET6:
247 		inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
248 		inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr);
249 		hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
250 		break;
251 #endif
252 	default:
253 		return NULL;
254 	}
255 
256 	net = dev_net(dst->dev);
257 	hash ^= net_hash_mix(net);
258 	hash = hash_32(hash, tcp_metrics_hash_log);
259 
260 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
261 	     tm = rcu_dereference(tm->tcpm_next)) {
262 		if (addr_same(&tm->tcpm_saddr, &saddr) &&
263 		    addr_same(&tm->tcpm_daddr, &daddr) &&
264 		    net_eq(tm_net(tm), net))
265 			break;
266 	}
267 	tcpm_check_stamp(tm, dst);
268 	return tm;
269 }
270 
271 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
272 						 struct dst_entry *dst,
273 						 bool create)
274 {
275 	struct tcp_metrics_block *tm;
276 	struct inetpeer_addr saddr, daddr;
277 	unsigned int hash;
278 	struct net *net;
279 
280 	if (sk->sk_family == AF_INET) {
281 		inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
282 		inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
283 		hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
284 	}
285 #if IS_ENABLED(CONFIG_IPV6)
286 	else if (sk->sk_family == AF_INET6) {
287 		if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
288 			inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
289 			inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
290 			hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
291 		} else {
292 			inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
293 			inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
294 			hash = ipv6_addr_hash(&sk->sk_v6_daddr);
295 		}
296 	}
297 #endif
298 	else
299 		return NULL;
300 
301 	net = dev_net(dst->dev);
302 	hash ^= net_hash_mix(net);
303 	hash = hash_32(hash, tcp_metrics_hash_log);
304 
305 	tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
306 	if (tm == TCP_METRICS_RECLAIM_PTR)
307 		tm = NULL;
308 	if (!tm && create)
309 		tm = tcpm_new(dst, &saddr, &daddr, hash);
310 	else
311 		tcpm_check_stamp(tm, dst);
312 
313 	return tm;
314 }
315 
316 /* Save metrics learned by this TCP session.  This function is called
317  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
318  * or goes from LAST-ACK to CLOSE.
319  */
320 void tcp_update_metrics(struct sock *sk)
321 {
322 	const struct inet_connection_sock *icsk = inet_csk(sk);
323 	struct dst_entry *dst = __sk_dst_get(sk);
324 	struct tcp_sock *tp = tcp_sk(sk);
325 	struct net *net = sock_net(sk);
326 	struct tcp_metrics_block *tm;
327 	unsigned long rtt;
328 	u32 val;
329 	int m;
330 
331 	sk_dst_confirm(sk);
332 	if (net->ipv4.sysctl_tcp_nometrics_save || !dst)
333 		return;
334 
335 	rcu_read_lock();
336 	if (icsk->icsk_backoff || !tp->srtt_us) {
337 		/* This session failed to estimate rtt. Why?
338 		 * Probably, no packets returned in time.  Reset our
339 		 * results.
340 		 */
341 		tm = tcp_get_metrics(sk, dst, false);
342 		if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
343 			tcp_metric_set(tm, TCP_METRIC_RTT, 0);
344 		goto out_unlock;
345 	} else
346 		tm = tcp_get_metrics(sk, dst, true);
347 
348 	if (!tm)
349 		goto out_unlock;
350 
351 	rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
352 	m = rtt - tp->srtt_us;
353 
354 	/* If newly calculated rtt larger than stored one, store new
355 	 * one. Otherwise, use EWMA. Remember, rtt overestimation is
356 	 * always better than underestimation.
357 	 */
358 	if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
359 		if (m <= 0)
360 			rtt = tp->srtt_us;
361 		else
362 			rtt -= (m >> 3);
363 		tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
364 	}
365 
366 	if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
367 		unsigned long var;
368 
369 		if (m < 0)
370 			m = -m;
371 
372 		/* Scale deviation to rttvar fixed point */
373 		m >>= 1;
374 		if (m < tp->mdev_us)
375 			m = tp->mdev_us;
376 
377 		var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
378 		if (m >= var)
379 			var = m;
380 		else
381 			var -= (var - m) >> 2;
382 
383 		tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
384 	}
385 
386 	if (tcp_in_initial_slowstart(tp)) {
387 		/* Slow start still did not finish. */
388 		if (!net->ipv4.sysctl_tcp_no_ssthresh_metrics_save &&
389 		    !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
390 			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
391 			if (val && (tp->snd_cwnd >> 1) > val)
392 				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
393 					       tp->snd_cwnd >> 1);
394 		}
395 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
396 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
397 			if (tp->snd_cwnd > val)
398 				tcp_metric_set(tm, TCP_METRIC_CWND,
399 					       tp->snd_cwnd);
400 		}
401 	} else if (!tcp_in_slow_start(tp) &&
402 		   icsk->icsk_ca_state == TCP_CA_Open) {
403 		/* Cong. avoidance phase, cwnd is reliable. */
404 		if (!net->ipv4.sysctl_tcp_no_ssthresh_metrics_save &&
405 		    !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
406 			tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
407 				       max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
408 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
409 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
410 			tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
411 		}
412 	} else {
413 		/* Else slow start did not finish, cwnd is non-sense,
414 		 * ssthresh may be also invalid.
415 		 */
416 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
417 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
418 			tcp_metric_set(tm, TCP_METRIC_CWND,
419 				       (val + tp->snd_ssthresh) >> 1);
420 		}
421 		if (!net->ipv4.sysctl_tcp_no_ssthresh_metrics_save &&
422 		    !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
423 			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
424 			if (val && tp->snd_ssthresh > val)
425 				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
426 					       tp->snd_ssthresh);
427 		}
428 		if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
429 			val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
430 			if (val < tp->reordering &&
431 			    tp->reordering != net->ipv4.sysctl_tcp_reordering)
432 				tcp_metric_set(tm, TCP_METRIC_REORDERING,
433 					       tp->reordering);
434 		}
435 	}
436 	tm->tcpm_stamp = jiffies;
437 out_unlock:
438 	rcu_read_unlock();
439 }
440 
441 /* Initialize metrics on socket. */
442 
443 void tcp_init_metrics(struct sock *sk)
444 {
445 	struct dst_entry *dst = __sk_dst_get(sk);
446 	struct tcp_sock *tp = tcp_sk(sk);
447 	struct net *net = sock_net(sk);
448 	struct tcp_metrics_block *tm;
449 	u32 val, crtt = 0; /* cached RTT scaled by 8 */
450 
451 	sk_dst_confirm(sk);
452 	if (!dst)
453 		goto reset;
454 
455 	rcu_read_lock();
456 	tm = tcp_get_metrics(sk, dst, true);
457 	if (!tm) {
458 		rcu_read_unlock();
459 		goto reset;
460 	}
461 
462 	if (tcp_metric_locked(tm, TCP_METRIC_CWND))
463 		tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
464 
465 	val = net->ipv4.sysctl_tcp_no_ssthresh_metrics_save ?
466 	      0 : tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
467 	if (val) {
468 		tp->snd_ssthresh = val;
469 		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
470 			tp->snd_ssthresh = tp->snd_cwnd_clamp;
471 	} else {
472 		/* ssthresh may have been reduced unnecessarily during.
473 		 * 3WHS. Restore it back to its initial default.
474 		 */
475 		tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
476 	}
477 	val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
478 	if (val && tp->reordering != val)
479 		tp->reordering = val;
480 
481 	crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
482 	rcu_read_unlock();
483 reset:
484 	/* The initial RTT measurement from the SYN/SYN-ACK is not ideal
485 	 * to seed the RTO for later data packets because SYN packets are
486 	 * small. Use the per-dst cached values to seed the RTO but keep
487 	 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
488 	 * Later the RTO will be updated immediately upon obtaining the first
489 	 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
490 	 * influences the first RTO but not later RTT estimation.
491 	 *
492 	 * But if RTT is not available from the SYN (due to retransmits or
493 	 * syn cookies) or the cache, force a conservative 3secs timeout.
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 	if (crtt > tp->srtt_us) {
505 		/* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
506 		crtt /= 8 * USEC_PER_SEC / HZ;
507 		inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
508 	} else if (tp->srtt_us == 0) {
509 		/* RFC6298: 5.7 We've failed to get a valid RTT sample from
510 		 * 3WHS. This is most likely due to retransmission,
511 		 * including spurious one. Reset the RTO back to 3secs
512 		 * from the more aggressive 1sec to avoid more spurious
513 		 * retransmission.
514 		 */
515 		tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
516 		tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
517 
518 		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
519 	}
520 }
521 
522 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
523 {
524 	struct tcp_metrics_block *tm;
525 	bool ret;
526 
527 	if (!dst)
528 		return false;
529 
530 	rcu_read_lock();
531 	tm = __tcp_get_metrics_req(req, dst);
532 	if (tm && tcp_metric_get(tm, TCP_METRIC_RTT))
533 		ret = true;
534 	else
535 		ret = false;
536 	rcu_read_unlock();
537 
538 	return ret;
539 }
540 
541 static DEFINE_SEQLOCK(fastopen_seqlock);
542 
543 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
544 			    struct tcp_fastopen_cookie *cookie)
545 {
546 	struct tcp_metrics_block *tm;
547 
548 	rcu_read_lock();
549 	tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
550 	if (tm) {
551 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
552 		unsigned int seq;
553 
554 		do {
555 			seq = read_seqbegin(&fastopen_seqlock);
556 			if (tfom->mss)
557 				*mss = tfom->mss;
558 			*cookie = tfom->cookie;
559 			if (cookie->len <= 0 && tfom->try_exp == 1)
560 				cookie->exp = true;
561 		} while (read_seqretry(&fastopen_seqlock, seq));
562 	}
563 	rcu_read_unlock();
564 }
565 
566 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
567 			    struct tcp_fastopen_cookie *cookie, bool syn_lost,
568 			    u16 try_exp)
569 {
570 	struct dst_entry *dst = __sk_dst_get(sk);
571 	struct tcp_metrics_block *tm;
572 
573 	if (!dst)
574 		return;
575 	rcu_read_lock();
576 	tm = tcp_get_metrics(sk, dst, true);
577 	if (tm) {
578 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
579 
580 		write_seqlock_bh(&fastopen_seqlock);
581 		if (mss)
582 			tfom->mss = mss;
583 		if (cookie && cookie->len > 0)
584 			tfom->cookie = *cookie;
585 		else if (try_exp > tfom->try_exp &&
586 			 tfom->cookie.len <= 0 && !tfom->cookie.exp)
587 			tfom->try_exp = try_exp;
588 		if (syn_lost) {
589 			++tfom->syn_loss;
590 			tfom->last_syn_loss = jiffies;
591 		} else
592 			tfom->syn_loss = 0;
593 		write_sequnlock_bh(&fastopen_seqlock);
594 	}
595 	rcu_read_unlock();
596 }
597 
598 static struct genl_family tcp_metrics_nl_family;
599 
600 static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
601 	[TCP_METRICS_ATTR_ADDR_IPV4]	= { .type = NLA_U32, },
602 	[TCP_METRICS_ATTR_ADDR_IPV6]	= { .type = NLA_BINARY,
603 					    .len = sizeof(struct in6_addr), },
604 	/* Following attributes are not received for GET/DEL,
605 	 * we keep them for reference
606 	 */
607 #if 0
608 	[TCP_METRICS_ATTR_AGE]		= { .type = NLA_MSECS, },
609 	[TCP_METRICS_ATTR_TW_TSVAL]	= { .type = NLA_U32, },
610 	[TCP_METRICS_ATTR_TW_TS_STAMP]	= { .type = NLA_S32, },
611 	[TCP_METRICS_ATTR_VALS]		= { .type = NLA_NESTED, },
612 	[TCP_METRICS_ATTR_FOPEN_MSS]	= { .type = NLA_U16, },
613 	[TCP_METRICS_ATTR_FOPEN_SYN_DROPS]	= { .type = NLA_U16, },
614 	[TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]	= { .type = NLA_MSECS, },
615 	[TCP_METRICS_ATTR_FOPEN_COOKIE]	= { .type = NLA_BINARY,
616 					    .len = TCP_FASTOPEN_COOKIE_MAX, },
617 #endif
618 };
619 
620 /* Add attributes, caller cancels its header on failure */
621 static int tcp_metrics_fill_info(struct sk_buff *msg,
622 				 struct tcp_metrics_block *tm)
623 {
624 	struct nlattr *nest;
625 	int i;
626 
627 	switch (tm->tcpm_daddr.family) {
628 	case AF_INET:
629 		if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
630 				    inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
631 			goto nla_put_failure;
632 		if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
633 				    inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
634 			goto nla_put_failure;
635 		break;
636 	case AF_INET6:
637 		if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
638 				     inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
639 			goto nla_put_failure;
640 		if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
641 				     inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
642 			goto nla_put_failure;
643 		break;
644 	default:
645 		return -EAFNOSUPPORT;
646 	}
647 
648 	if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
649 			  jiffies - tm->tcpm_stamp,
650 			  TCP_METRICS_ATTR_PAD) < 0)
651 		goto nla_put_failure;
652 
653 	{
654 		int n = 0;
655 
656 		nest = nla_nest_start_noflag(msg, TCP_METRICS_ATTR_VALS);
657 		if (!nest)
658 			goto nla_put_failure;
659 		for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
660 			u32 val = tm->tcpm_vals[i];
661 
662 			if (!val)
663 				continue;
664 			if (i == TCP_METRIC_RTT) {
665 				if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
666 						val) < 0)
667 					goto nla_put_failure;
668 				n++;
669 				val = max(val / 1000, 1U);
670 			}
671 			if (i == TCP_METRIC_RTTVAR) {
672 				if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
673 						val) < 0)
674 					goto nla_put_failure;
675 				n++;
676 				val = max(val / 1000, 1U);
677 			}
678 			if (nla_put_u32(msg, i + 1, val) < 0)
679 				goto nla_put_failure;
680 			n++;
681 		}
682 		if (n)
683 			nla_nest_end(msg, nest);
684 		else
685 			nla_nest_cancel(msg, nest);
686 	}
687 
688 	{
689 		struct tcp_fastopen_metrics tfom_copy[1], *tfom;
690 		unsigned int seq;
691 
692 		do {
693 			seq = read_seqbegin(&fastopen_seqlock);
694 			tfom_copy[0] = tm->tcpm_fastopen;
695 		} while (read_seqretry(&fastopen_seqlock, seq));
696 
697 		tfom = tfom_copy;
698 		if (tfom->mss &&
699 		    nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
700 				tfom->mss) < 0)
701 			goto nla_put_failure;
702 		if (tfom->syn_loss &&
703 		    (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
704 				tfom->syn_loss) < 0 ||
705 		     nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
706 				jiffies - tfom->last_syn_loss,
707 				TCP_METRICS_ATTR_PAD) < 0))
708 			goto nla_put_failure;
709 		if (tfom->cookie.len > 0 &&
710 		    nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
711 			    tfom->cookie.len, tfom->cookie.val) < 0)
712 			goto nla_put_failure;
713 	}
714 
715 	return 0;
716 
717 nla_put_failure:
718 	return -EMSGSIZE;
719 }
720 
721 static int tcp_metrics_dump_info(struct sk_buff *skb,
722 				 struct netlink_callback *cb,
723 				 struct tcp_metrics_block *tm)
724 {
725 	void *hdr;
726 
727 	hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
728 			  &tcp_metrics_nl_family, NLM_F_MULTI,
729 			  TCP_METRICS_CMD_GET);
730 	if (!hdr)
731 		return -EMSGSIZE;
732 
733 	if (tcp_metrics_fill_info(skb, tm) < 0)
734 		goto nla_put_failure;
735 
736 	genlmsg_end(skb, hdr);
737 	return 0;
738 
739 nla_put_failure:
740 	genlmsg_cancel(skb, hdr);
741 	return -EMSGSIZE;
742 }
743 
744 static int tcp_metrics_nl_dump(struct sk_buff *skb,
745 			       struct netlink_callback *cb)
746 {
747 	struct net *net = sock_net(skb->sk);
748 	unsigned int max_rows = 1U << tcp_metrics_hash_log;
749 	unsigned int row, s_row = cb->args[0];
750 	int s_col = cb->args[1], col = s_col;
751 
752 	for (row = s_row; row < max_rows; row++, s_col = 0) {
753 		struct tcp_metrics_block *tm;
754 		struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
755 
756 		rcu_read_lock();
757 		for (col = 0, tm = rcu_dereference(hb->chain); tm;
758 		     tm = rcu_dereference(tm->tcpm_next), col++) {
759 			if (!net_eq(tm_net(tm), net))
760 				continue;
761 			if (col < s_col)
762 				continue;
763 			if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
764 				rcu_read_unlock();
765 				goto done;
766 			}
767 		}
768 		rcu_read_unlock();
769 	}
770 
771 done:
772 	cb->args[0] = row;
773 	cb->args[1] = col;
774 	return skb->len;
775 }
776 
777 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
778 			   unsigned int *hash, int optional, int v4, int v6)
779 {
780 	struct nlattr *a;
781 
782 	a = info->attrs[v4];
783 	if (a) {
784 		inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
785 		if (hash)
786 			*hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
787 		return 0;
788 	}
789 	a = info->attrs[v6];
790 	if (a) {
791 		struct in6_addr in6;
792 
793 		if (nla_len(a) != sizeof(struct in6_addr))
794 			return -EINVAL;
795 		in6 = nla_get_in6_addr(a);
796 		inetpeer_set_addr_v6(addr, &in6);
797 		if (hash)
798 			*hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
799 		return 0;
800 	}
801 	return optional ? 1 : -EAFNOSUPPORT;
802 }
803 
804 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
805 			 unsigned int *hash, int optional)
806 {
807 	return __parse_nl_addr(info, addr, hash, optional,
808 			       TCP_METRICS_ATTR_ADDR_IPV4,
809 			       TCP_METRICS_ATTR_ADDR_IPV6);
810 }
811 
812 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
813 {
814 	return __parse_nl_addr(info, addr, NULL, 0,
815 			       TCP_METRICS_ATTR_SADDR_IPV4,
816 			       TCP_METRICS_ATTR_SADDR_IPV6);
817 }
818 
819 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
820 {
821 	struct tcp_metrics_block *tm;
822 	struct inetpeer_addr saddr, daddr;
823 	unsigned int hash;
824 	struct sk_buff *msg;
825 	struct net *net = genl_info_net(info);
826 	void *reply;
827 	int ret;
828 	bool src = true;
829 
830 	ret = parse_nl_addr(info, &daddr, &hash, 0);
831 	if (ret < 0)
832 		return ret;
833 
834 	ret = parse_nl_saddr(info, &saddr);
835 	if (ret < 0)
836 		src = false;
837 
838 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
839 	if (!msg)
840 		return -ENOMEM;
841 
842 	reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
843 				  info->genlhdr->cmd);
844 	if (!reply)
845 		goto nla_put_failure;
846 
847 	hash ^= net_hash_mix(net);
848 	hash = hash_32(hash, tcp_metrics_hash_log);
849 	ret = -ESRCH;
850 	rcu_read_lock();
851 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
852 	     tm = rcu_dereference(tm->tcpm_next)) {
853 		if (addr_same(&tm->tcpm_daddr, &daddr) &&
854 		    (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
855 		    net_eq(tm_net(tm), net)) {
856 			ret = tcp_metrics_fill_info(msg, tm);
857 			break;
858 		}
859 	}
860 	rcu_read_unlock();
861 	if (ret < 0)
862 		goto out_free;
863 
864 	genlmsg_end(msg, reply);
865 	return genlmsg_reply(msg, info);
866 
867 nla_put_failure:
868 	ret = -EMSGSIZE;
869 
870 out_free:
871 	nlmsg_free(msg);
872 	return ret;
873 }
874 
875 static void tcp_metrics_flush_all(struct net *net)
876 {
877 	unsigned int max_rows = 1U << tcp_metrics_hash_log;
878 	struct tcpm_hash_bucket *hb = tcp_metrics_hash;
879 	struct tcp_metrics_block *tm;
880 	unsigned int row;
881 
882 	for (row = 0; row < max_rows; row++, hb++) {
883 		struct tcp_metrics_block __rcu **pp;
884 		bool match;
885 
886 		spin_lock_bh(&tcp_metrics_lock);
887 		pp = &hb->chain;
888 		for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
889 			match = net ? net_eq(tm_net(tm), net) :
890 				!refcount_read(&tm_net(tm)->ns.count);
891 			if (match) {
892 				*pp = tm->tcpm_next;
893 				kfree_rcu(tm, rcu_head);
894 			} else {
895 				pp = &tm->tcpm_next;
896 			}
897 		}
898 		spin_unlock_bh(&tcp_metrics_lock);
899 	}
900 }
901 
902 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
903 {
904 	struct tcpm_hash_bucket *hb;
905 	struct tcp_metrics_block *tm;
906 	struct tcp_metrics_block __rcu **pp;
907 	struct inetpeer_addr saddr, daddr;
908 	unsigned int hash;
909 	struct net *net = genl_info_net(info);
910 	int ret;
911 	bool src = true, found = false;
912 
913 	ret = parse_nl_addr(info, &daddr, &hash, 1);
914 	if (ret < 0)
915 		return ret;
916 	if (ret > 0) {
917 		tcp_metrics_flush_all(net);
918 		return 0;
919 	}
920 	ret = parse_nl_saddr(info, &saddr);
921 	if (ret < 0)
922 		src = false;
923 
924 	hash ^= net_hash_mix(net);
925 	hash = hash_32(hash, tcp_metrics_hash_log);
926 	hb = tcp_metrics_hash + hash;
927 	pp = &hb->chain;
928 	spin_lock_bh(&tcp_metrics_lock);
929 	for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
930 		if (addr_same(&tm->tcpm_daddr, &daddr) &&
931 		    (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
932 		    net_eq(tm_net(tm), net)) {
933 			*pp = tm->tcpm_next;
934 			kfree_rcu(tm, rcu_head);
935 			found = true;
936 		} else {
937 			pp = &tm->tcpm_next;
938 		}
939 	}
940 	spin_unlock_bh(&tcp_metrics_lock);
941 	if (!found)
942 		return -ESRCH;
943 	return 0;
944 }
945 
946 static const struct genl_small_ops tcp_metrics_nl_ops[] = {
947 	{
948 		.cmd = TCP_METRICS_CMD_GET,
949 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
950 		.doit = tcp_metrics_nl_cmd_get,
951 		.dumpit = tcp_metrics_nl_dump,
952 	},
953 	{
954 		.cmd = TCP_METRICS_CMD_DEL,
955 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
956 		.doit = tcp_metrics_nl_cmd_del,
957 		.flags = GENL_ADMIN_PERM,
958 	},
959 };
960 
961 static struct genl_family tcp_metrics_nl_family __ro_after_init = {
962 	.hdrsize	= 0,
963 	.name		= TCP_METRICS_GENL_NAME,
964 	.version	= TCP_METRICS_GENL_VERSION,
965 	.maxattr	= TCP_METRICS_ATTR_MAX,
966 	.policy = tcp_metrics_nl_policy,
967 	.netnsok	= true,
968 	.module		= THIS_MODULE,
969 	.small_ops	= tcp_metrics_nl_ops,
970 	.n_small_ops	= ARRAY_SIZE(tcp_metrics_nl_ops),
971 };
972 
973 static unsigned int tcpmhash_entries;
974 static int __init set_tcpmhash_entries(char *str)
975 {
976 	ssize_t ret;
977 
978 	if (!str)
979 		return 0;
980 
981 	ret = kstrtouint(str, 0, &tcpmhash_entries);
982 	if (ret)
983 		return 0;
984 
985 	return 1;
986 }
987 __setup("tcpmhash_entries=", set_tcpmhash_entries);
988 
989 static int __net_init tcp_net_metrics_init(struct net *net)
990 {
991 	size_t size;
992 	unsigned int slots;
993 
994 	if (!net_eq(net, &init_net))
995 		return 0;
996 
997 	slots = tcpmhash_entries;
998 	if (!slots) {
999 		if (totalram_pages() >= 128 * 1024)
1000 			slots = 16 * 1024;
1001 		else
1002 			slots = 8 * 1024;
1003 	}
1004 
1005 	tcp_metrics_hash_log = order_base_2(slots);
1006 	size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1007 
1008 	tcp_metrics_hash = kvzalloc(size, GFP_KERNEL);
1009 	if (!tcp_metrics_hash)
1010 		return -ENOMEM;
1011 
1012 	return 0;
1013 }
1014 
1015 static void __net_exit tcp_net_metrics_exit_batch(struct list_head *net_exit_list)
1016 {
1017 	tcp_metrics_flush_all(NULL);
1018 }
1019 
1020 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1021 	.init		=	tcp_net_metrics_init,
1022 	.exit_batch	=	tcp_net_metrics_exit_batch,
1023 };
1024 
1025 void __init tcp_metrics_init(void)
1026 {
1027 	int ret;
1028 
1029 	ret = register_pernet_subsys(&tcp_net_metrics_ops);
1030 	if (ret < 0)
1031 		panic("Could not allocate the tcp_metrics hash table\n");
1032 
1033 	ret = genl_register_family(&tcp_metrics_nl_family);
1034 	if (ret < 0)
1035 		panic("Could not register tcp_metrics generic netlink\n");
1036 }
1037