xref: /openbmc/linux/net/ipv4/tcp_metrics.c (revision a44e4f3a)
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 (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
389 			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
390 			if (val && (tp->snd_cwnd >> 1) > val)
391 				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
392 					       tp->snd_cwnd >> 1);
393 		}
394 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
395 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
396 			if (tp->snd_cwnd > val)
397 				tcp_metric_set(tm, TCP_METRIC_CWND,
398 					       tp->snd_cwnd);
399 		}
400 	} else if (!tcp_in_slow_start(tp) &&
401 		   icsk->icsk_ca_state == TCP_CA_Open) {
402 		/* Cong. avoidance phase, cwnd is reliable. */
403 		if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
404 			tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
405 				       max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
406 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
407 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
408 			tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
409 		}
410 	} else {
411 		/* Else slow start did not finish, cwnd is non-sense,
412 		 * ssthresh may be also invalid.
413 		 */
414 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
415 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
416 			tcp_metric_set(tm, TCP_METRIC_CWND,
417 				       (val + tp->snd_ssthresh) >> 1);
418 		}
419 		if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
420 			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
421 			if (val && tp->snd_ssthresh > val)
422 				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
423 					       tp->snd_ssthresh);
424 		}
425 		if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
426 			val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
427 			if (val < tp->reordering &&
428 			    tp->reordering != net->ipv4.sysctl_tcp_reordering)
429 				tcp_metric_set(tm, TCP_METRIC_REORDERING,
430 					       tp->reordering);
431 		}
432 	}
433 	tm->tcpm_stamp = jiffies;
434 out_unlock:
435 	rcu_read_unlock();
436 }
437 
438 /* Initialize metrics on socket. */
439 
440 void tcp_init_metrics(struct sock *sk)
441 {
442 	struct dst_entry *dst = __sk_dst_get(sk);
443 	struct tcp_sock *tp = tcp_sk(sk);
444 	struct tcp_metrics_block *tm;
445 	u32 val, crtt = 0; /* cached RTT scaled by 8 */
446 
447 	sk_dst_confirm(sk);
448 	if (!dst)
449 		goto reset;
450 
451 	rcu_read_lock();
452 	tm = tcp_get_metrics(sk, dst, true);
453 	if (!tm) {
454 		rcu_read_unlock();
455 		goto reset;
456 	}
457 
458 	if (tcp_metric_locked(tm, TCP_METRIC_CWND))
459 		tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
460 
461 	val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
462 	if (val) {
463 		tp->snd_ssthresh = val;
464 		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
465 			tp->snd_ssthresh = tp->snd_cwnd_clamp;
466 	} else {
467 		/* ssthresh may have been reduced unnecessarily during.
468 		 * 3WHS. Restore it back to its initial default.
469 		 */
470 		tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
471 	}
472 	val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
473 	if (val && tp->reordering != val)
474 		tp->reordering = val;
475 
476 	crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
477 	rcu_read_unlock();
478 reset:
479 	/* The initial RTT measurement from the SYN/SYN-ACK is not ideal
480 	 * to seed the RTO for later data packets because SYN packets are
481 	 * small. Use the per-dst cached values to seed the RTO but keep
482 	 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
483 	 * Later the RTO will be updated immediately upon obtaining the first
484 	 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
485 	 * influences the first RTO but not later RTT estimation.
486 	 *
487 	 * But if RTT is not available from the SYN (due to retransmits or
488 	 * syn cookies) or the cache, force a conservative 3secs timeout.
489 	 *
490 	 * A bit of theory. RTT is time passed after "normal" sized packet
491 	 * is sent until it is ACKed. In normal circumstances sending small
492 	 * packets force peer to delay ACKs and calculation is correct too.
493 	 * The algorithm is adaptive and, provided we follow specs, it
494 	 * NEVER underestimate RTT. BUT! If peer tries to make some clever
495 	 * tricks sort of "quick acks" for time long enough to decrease RTT
496 	 * to low value, and then abruptly stops to do it and starts to delay
497 	 * ACKs, wait for troubles.
498 	 */
499 	if (crtt > tp->srtt_us) {
500 		/* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
501 		crtt /= 8 * USEC_PER_SEC / HZ;
502 		inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
503 	} else if (tp->srtt_us == 0) {
504 		/* RFC6298: 5.7 We've failed to get a valid RTT sample from
505 		 * 3WHS. This is most likely due to retransmission,
506 		 * including spurious one. Reset the RTO back to 3secs
507 		 * from the more aggressive 1sec to avoid more spurious
508 		 * retransmission.
509 		 */
510 		tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
511 		tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
512 
513 		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
514 	}
515 }
516 
517 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
518 {
519 	struct tcp_metrics_block *tm;
520 	bool ret;
521 
522 	if (!dst)
523 		return false;
524 
525 	rcu_read_lock();
526 	tm = __tcp_get_metrics_req(req, dst);
527 	if (tm && tcp_metric_get(tm, TCP_METRIC_RTT))
528 		ret = true;
529 	else
530 		ret = false;
531 	rcu_read_unlock();
532 
533 	return ret;
534 }
535 
536 static DEFINE_SEQLOCK(fastopen_seqlock);
537 
538 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
539 			    struct tcp_fastopen_cookie *cookie)
540 {
541 	struct tcp_metrics_block *tm;
542 
543 	rcu_read_lock();
544 	tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
545 	if (tm) {
546 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
547 		unsigned int seq;
548 
549 		do {
550 			seq = read_seqbegin(&fastopen_seqlock);
551 			if (tfom->mss)
552 				*mss = tfom->mss;
553 			*cookie = tfom->cookie;
554 			if (cookie->len <= 0 && tfom->try_exp == 1)
555 				cookie->exp = true;
556 		} while (read_seqretry(&fastopen_seqlock, seq));
557 	}
558 	rcu_read_unlock();
559 }
560 
561 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
562 			    struct tcp_fastopen_cookie *cookie, bool syn_lost,
563 			    u16 try_exp)
564 {
565 	struct dst_entry *dst = __sk_dst_get(sk);
566 	struct tcp_metrics_block *tm;
567 
568 	if (!dst)
569 		return;
570 	rcu_read_lock();
571 	tm = tcp_get_metrics(sk, dst, true);
572 	if (tm) {
573 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
574 
575 		write_seqlock_bh(&fastopen_seqlock);
576 		if (mss)
577 			tfom->mss = mss;
578 		if (cookie && cookie->len > 0)
579 			tfom->cookie = *cookie;
580 		else if (try_exp > tfom->try_exp &&
581 			 tfom->cookie.len <= 0 && !tfom->cookie.exp)
582 			tfom->try_exp = try_exp;
583 		if (syn_lost) {
584 			++tfom->syn_loss;
585 			tfom->last_syn_loss = jiffies;
586 		} else
587 			tfom->syn_loss = 0;
588 		write_sequnlock_bh(&fastopen_seqlock);
589 	}
590 	rcu_read_unlock();
591 }
592 
593 static struct genl_family tcp_metrics_nl_family;
594 
595 static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
596 	[TCP_METRICS_ATTR_ADDR_IPV4]	= { .type = NLA_U32, },
597 	[TCP_METRICS_ATTR_ADDR_IPV6]	= { .type = NLA_BINARY,
598 					    .len = sizeof(struct in6_addr), },
599 	/* Following attributes are not received for GET/DEL,
600 	 * we keep them for reference
601 	 */
602 #if 0
603 	[TCP_METRICS_ATTR_AGE]		= { .type = NLA_MSECS, },
604 	[TCP_METRICS_ATTR_TW_TSVAL]	= { .type = NLA_U32, },
605 	[TCP_METRICS_ATTR_TW_TS_STAMP]	= { .type = NLA_S32, },
606 	[TCP_METRICS_ATTR_VALS]		= { .type = NLA_NESTED, },
607 	[TCP_METRICS_ATTR_FOPEN_MSS]	= { .type = NLA_U16, },
608 	[TCP_METRICS_ATTR_FOPEN_SYN_DROPS]	= { .type = NLA_U16, },
609 	[TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]	= { .type = NLA_MSECS, },
610 	[TCP_METRICS_ATTR_FOPEN_COOKIE]	= { .type = NLA_BINARY,
611 					    .len = TCP_FASTOPEN_COOKIE_MAX, },
612 #endif
613 };
614 
615 /* Add attributes, caller cancels its header on failure */
616 static int tcp_metrics_fill_info(struct sk_buff *msg,
617 				 struct tcp_metrics_block *tm)
618 {
619 	struct nlattr *nest;
620 	int i;
621 
622 	switch (tm->tcpm_daddr.family) {
623 	case AF_INET:
624 		if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
625 				    inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
626 			goto nla_put_failure;
627 		if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
628 				    inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
629 			goto nla_put_failure;
630 		break;
631 	case AF_INET6:
632 		if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
633 				     inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
634 			goto nla_put_failure;
635 		if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
636 				     inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
637 			goto nla_put_failure;
638 		break;
639 	default:
640 		return -EAFNOSUPPORT;
641 	}
642 
643 	if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
644 			  jiffies - tm->tcpm_stamp,
645 			  TCP_METRICS_ATTR_PAD) < 0)
646 		goto nla_put_failure;
647 
648 	{
649 		int n = 0;
650 
651 		nest = nla_nest_start_noflag(msg, TCP_METRICS_ATTR_VALS);
652 		if (!nest)
653 			goto nla_put_failure;
654 		for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
655 			u32 val = tm->tcpm_vals[i];
656 
657 			if (!val)
658 				continue;
659 			if (i == TCP_METRIC_RTT) {
660 				if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
661 						val) < 0)
662 					goto nla_put_failure;
663 				n++;
664 				val = max(val / 1000, 1U);
665 			}
666 			if (i == TCP_METRIC_RTTVAR) {
667 				if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
668 						val) < 0)
669 					goto nla_put_failure;
670 				n++;
671 				val = max(val / 1000, 1U);
672 			}
673 			if (nla_put_u32(msg, i + 1, val) < 0)
674 				goto nla_put_failure;
675 			n++;
676 		}
677 		if (n)
678 			nla_nest_end(msg, nest);
679 		else
680 			nla_nest_cancel(msg, nest);
681 	}
682 
683 	{
684 		struct tcp_fastopen_metrics tfom_copy[1], *tfom;
685 		unsigned int seq;
686 
687 		do {
688 			seq = read_seqbegin(&fastopen_seqlock);
689 			tfom_copy[0] = tm->tcpm_fastopen;
690 		} while (read_seqretry(&fastopen_seqlock, seq));
691 
692 		tfom = tfom_copy;
693 		if (tfom->mss &&
694 		    nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
695 				tfom->mss) < 0)
696 			goto nla_put_failure;
697 		if (tfom->syn_loss &&
698 		    (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
699 				tfom->syn_loss) < 0 ||
700 		     nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
701 				jiffies - tfom->last_syn_loss,
702 				TCP_METRICS_ATTR_PAD) < 0))
703 			goto nla_put_failure;
704 		if (tfom->cookie.len > 0 &&
705 		    nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
706 			    tfom->cookie.len, tfom->cookie.val) < 0)
707 			goto nla_put_failure;
708 	}
709 
710 	return 0;
711 
712 nla_put_failure:
713 	return -EMSGSIZE;
714 }
715 
716 static int tcp_metrics_dump_info(struct sk_buff *skb,
717 				 struct netlink_callback *cb,
718 				 struct tcp_metrics_block *tm)
719 {
720 	void *hdr;
721 
722 	hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
723 			  &tcp_metrics_nl_family, NLM_F_MULTI,
724 			  TCP_METRICS_CMD_GET);
725 	if (!hdr)
726 		return -EMSGSIZE;
727 
728 	if (tcp_metrics_fill_info(skb, tm) < 0)
729 		goto nla_put_failure;
730 
731 	genlmsg_end(skb, hdr);
732 	return 0;
733 
734 nla_put_failure:
735 	genlmsg_cancel(skb, hdr);
736 	return -EMSGSIZE;
737 }
738 
739 static int tcp_metrics_nl_dump(struct sk_buff *skb,
740 			       struct netlink_callback *cb)
741 {
742 	struct net *net = sock_net(skb->sk);
743 	unsigned int max_rows = 1U << tcp_metrics_hash_log;
744 	unsigned int row, s_row = cb->args[0];
745 	int s_col = cb->args[1], col = s_col;
746 
747 	for (row = s_row; row < max_rows; row++, s_col = 0) {
748 		struct tcp_metrics_block *tm;
749 		struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
750 
751 		rcu_read_lock();
752 		for (col = 0, tm = rcu_dereference(hb->chain); tm;
753 		     tm = rcu_dereference(tm->tcpm_next), col++) {
754 			if (!net_eq(tm_net(tm), net))
755 				continue;
756 			if (col < s_col)
757 				continue;
758 			if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
759 				rcu_read_unlock();
760 				goto done;
761 			}
762 		}
763 		rcu_read_unlock();
764 	}
765 
766 done:
767 	cb->args[0] = row;
768 	cb->args[1] = col;
769 	return skb->len;
770 }
771 
772 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
773 			   unsigned int *hash, int optional, int v4, int v6)
774 {
775 	struct nlattr *a;
776 
777 	a = info->attrs[v4];
778 	if (a) {
779 		inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
780 		if (hash)
781 			*hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
782 		return 0;
783 	}
784 	a = info->attrs[v6];
785 	if (a) {
786 		struct in6_addr in6;
787 
788 		if (nla_len(a) != sizeof(struct in6_addr))
789 			return -EINVAL;
790 		in6 = nla_get_in6_addr(a);
791 		inetpeer_set_addr_v6(addr, &in6);
792 		if (hash)
793 			*hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
794 		return 0;
795 	}
796 	return optional ? 1 : -EAFNOSUPPORT;
797 }
798 
799 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
800 			 unsigned int *hash, int optional)
801 {
802 	return __parse_nl_addr(info, addr, hash, optional,
803 			       TCP_METRICS_ATTR_ADDR_IPV4,
804 			       TCP_METRICS_ATTR_ADDR_IPV6);
805 }
806 
807 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
808 {
809 	return __parse_nl_addr(info, addr, NULL, 0,
810 			       TCP_METRICS_ATTR_SADDR_IPV4,
811 			       TCP_METRICS_ATTR_SADDR_IPV6);
812 }
813 
814 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
815 {
816 	struct tcp_metrics_block *tm;
817 	struct inetpeer_addr saddr, daddr;
818 	unsigned int hash;
819 	struct sk_buff *msg;
820 	struct net *net = genl_info_net(info);
821 	void *reply;
822 	int ret;
823 	bool src = true;
824 
825 	ret = parse_nl_addr(info, &daddr, &hash, 0);
826 	if (ret < 0)
827 		return ret;
828 
829 	ret = parse_nl_saddr(info, &saddr);
830 	if (ret < 0)
831 		src = false;
832 
833 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
834 	if (!msg)
835 		return -ENOMEM;
836 
837 	reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
838 				  info->genlhdr->cmd);
839 	if (!reply)
840 		goto nla_put_failure;
841 
842 	hash ^= net_hash_mix(net);
843 	hash = hash_32(hash, tcp_metrics_hash_log);
844 	ret = -ESRCH;
845 	rcu_read_lock();
846 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
847 	     tm = rcu_dereference(tm->tcpm_next)) {
848 		if (addr_same(&tm->tcpm_daddr, &daddr) &&
849 		    (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
850 		    net_eq(tm_net(tm), net)) {
851 			ret = tcp_metrics_fill_info(msg, tm);
852 			break;
853 		}
854 	}
855 	rcu_read_unlock();
856 	if (ret < 0)
857 		goto out_free;
858 
859 	genlmsg_end(msg, reply);
860 	return genlmsg_reply(msg, info);
861 
862 nla_put_failure:
863 	ret = -EMSGSIZE;
864 
865 out_free:
866 	nlmsg_free(msg);
867 	return ret;
868 }
869 
870 static void tcp_metrics_flush_all(struct net *net)
871 {
872 	unsigned int max_rows = 1U << tcp_metrics_hash_log;
873 	struct tcpm_hash_bucket *hb = tcp_metrics_hash;
874 	struct tcp_metrics_block *tm;
875 	unsigned int row;
876 
877 	for (row = 0; row < max_rows; row++, hb++) {
878 		struct tcp_metrics_block __rcu **pp;
879 		bool match;
880 
881 		spin_lock_bh(&tcp_metrics_lock);
882 		pp = &hb->chain;
883 		for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
884 			match = net ? net_eq(tm_net(tm), net) :
885 				!refcount_read(&tm_net(tm)->count);
886 			if (match) {
887 				*pp = tm->tcpm_next;
888 				kfree_rcu(tm, rcu_head);
889 			} else {
890 				pp = &tm->tcpm_next;
891 			}
892 		}
893 		spin_unlock_bh(&tcp_metrics_lock);
894 	}
895 }
896 
897 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
898 {
899 	struct tcpm_hash_bucket *hb;
900 	struct tcp_metrics_block *tm;
901 	struct tcp_metrics_block __rcu **pp;
902 	struct inetpeer_addr saddr, daddr;
903 	unsigned int hash;
904 	struct net *net = genl_info_net(info);
905 	int ret;
906 	bool src = true, found = false;
907 
908 	ret = parse_nl_addr(info, &daddr, &hash, 1);
909 	if (ret < 0)
910 		return ret;
911 	if (ret > 0) {
912 		tcp_metrics_flush_all(net);
913 		return 0;
914 	}
915 	ret = parse_nl_saddr(info, &saddr);
916 	if (ret < 0)
917 		src = false;
918 
919 	hash ^= net_hash_mix(net);
920 	hash = hash_32(hash, tcp_metrics_hash_log);
921 	hb = tcp_metrics_hash + hash;
922 	pp = &hb->chain;
923 	spin_lock_bh(&tcp_metrics_lock);
924 	for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
925 		if (addr_same(&tm->tcpm_daddr, &daddr) &&
926 		    (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
927 		    net_eq(tm_net(tm), net)) {
928 			*pp = tm->tcpm_next;
929 			kfree_rcu(tm, rcu_head);
930 			found = true;
931 		} else {
932 			pp = &tm->tcpm_next;
933 		}
934 	}
935 	spin_unlock_bh(&tcp_metrics_lock);
936 	if (!found)
937 		return -ESRCH;
938 	return 0;
939 }
940 
941 static const struct genl_ops tcp_metrics_nl_ops[] = {
942 	{
943 		.cmd = TCP_METRICS_CMD_GET,
944 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
945 		.doit = tcp_metrics_nl_cmd_get,
946 		.dumpit = tcp_metrics_nl_dump,
947 	},
948 	{
949 		.cmd = TCP_METRICS_CMD_DEL,
950 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
951 		.doit = tcp_metrics_nl_cmd_del,
952 		.flags = GENL_ADMIN_PERM,
953 	},
954 };
955 
956 static struct genl_family tcp_metrics_nl_family __ro_after_init = {
957 	.hdrsize	= 0,
958 	.name		= TCP_METRICS_GENL_NAME,
959 	.version	= TCP_METRICS_GENL_VERSION,
960 	.maxattr	= TCP_METRICS_ATTR_MAX,
961 	.policy = tcp_metrics_nl_policy,
962 	.netnsok	= true,
963 	.module		= THIS_MODULE,
964 	.ops		= tcp_metrics_nl_ops,
965 	.n_ops		= ARRAY_SIZE(tcp_metrics_nl_ops),
966 };
967 
968 static unsigned int tcpmhash_entries;
969 static int __init set_tcpmhash_entries(char *str)
970 {
971 	ssize_t ret;
972 
973 	if (!str)
974 		return 0;
975 
976 	ret = kstrtouint(str, 0, &tcpmhash_entries);
977 	if (ret)
978 		return 0;
979 
980 	return 1;
981 }
982 __setup("tcpmhash_entries=", set_tcpmhash_entries);
983 
984 static int __net_init tcp_net_metrics_init(struct net *net)
985 {
986 	size_t size;
987 	unsigned int slots;
988 
989 	if (!net_eq(net, &init_net))
990 		return 0;
991 
992 	slots = tcpmhash_entries;
993 	if (!slots) {
994 		if (totalram_pages() >= 128 * 1024)
995 			slots = 16 * 1024;
996 		else
997 			slots = 8 * 1024;
998 	}
999 
1000 	tcp_metrics_hash_log = order_base_2(slots);
1001 	size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1002 
1003 	tcp_metrics_hash = kvzalloc(size, GFP_KERNEL);
1004 	if (!tcp_metrics_hash)
1005 		return -ENOMEM;
1006 
1007 	return 0;
1008 }
1009 
1010 static void __net_exit tcp_net_metrics_exit_batch(struct list_head *net_exit_list)
1011 {
1012 	tcp_metrics_flush_all(NULL);
1013 }
1014 
1015 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1016 	.init		=	tcp_net_metrics_init,
1017 	.exit_batch	=	tcp_net_metrics_exit_batch,
1018 };
1019 
1020 void __init tcp_metrics_init(void)
1021 {
1022 	int ret;
1023 
1024 	ret = register_pernet_subsys(&tcp_net_metrics_ops);
1025 	if (ret < 0)
1026 		panic("Could not allocate the tcp_metrics hash table\n");
1027 
1028 	ret = genl_register_family(&tcp_metrics_nl_family);
1029 	if (ret < 0)
1030 		panic("Could not register tcp_metrics generic netlink\n");
1031 }
1032