xref: /openbmc/linux/net/ipv4/tcp_metrics.c (revision f0168042)
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 	struct net			*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(const struct tcp_metrics_block *tm)
55 {
56 	/* Paired with the WRITE_ONCE() in tcpm_new() */
57 	return READ_ONCE(tm->tcpm_net);
58 }
59 
60 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
61 			      enum tcp_metric_index idx)
62 {
63 	/* Paired with WRITE_ONCE() in tcpm_suck_dst() */
64 	return READ_ONCE(tm->tcpm_lock) & (1 << idx);
65 }
66 
67 static u32 tcp_metric_get(const struct tcp_metrics_block *tm,
68 			  enum tcp_metric_index idx)
69 {
70 	/* Paired with WRITE_ONCE() in tcp_metric_set() */
71 	return READ_ONCE(tm->tcpm_vals[idx]);
72 }
73 
74 static void tcp_metric_set(struct tcp_metrics_block *tm,
75 			   enum tcp_metric_index idx,
76 			   u32 val)
77 {
78 	/* Paired with READ_ONCE() in tcp_metric_get() */
79 	WRITE_ONCE(tm->tcpm_vals[idx], val);
80 }
81 
82 static bool addr_same(const struct inetpeer_addr *a,
83 		      const struct inetpeer_addr *b)
84 {
85 	return (a->family == b->family) && !inetpeer_addr_cmp(a, b);
86 }
87 
88 struct tcpm_hash_bucket {
89 	struct tcp_metrics_block __rcu	*chain;
90 };
91 
92 static struct tcpm_hash_bucket	*tcp_metrics_hash __read_mostly;
93 static unsigned int		tcp_metrics_hash_log __read_mostly;
94 
95 static DEFINE_SPINLOCK(tcp_metrics_lock);
96 static DEFINE_SEQLOCK(fastopen_seqlock);
97 
98 static void tcpm_suck_dst(struct tcp_metrics_block *tm,
99 			  const struct dst_entry *dst,
100 			  bool fastopen_clear)
101 {
102 	u32 msval;
103 	u32 val;
104 
105 	WRITE_ONCE(tm->tcpm_stamp, jiffies);
106 
107 	val = 0;
108 	if (dst_metric_locked(dst, RTAX_RTT))
109 		val |= 1 << TCP_METRIC_RTT;
110 	if (dst_metric_locked(dst, RTAX_RTTVAR))
111 		val |= 1 << TCP_METRIC_RTTVAR;
112 	if (dst_metric_locked(dst, RTAX_SSTHRESH))
113 		val |= 1 << TCP_METRIC_SSTHRESH;
114 	if (dst_metric_locked(dst, RTAX_CWND))
115 		val |= 1 << TCP_METRIC_CWND;
116 	if (dst_metric_locked(dst, RTAX_REORDERING))
117 		val |= 1 << TCP_METRIC_REORDERING;
118 	/* Paired with READ_ONCE() in tcp_metric_locked() */
119 	WRITE_ONCE(tm->tcpm_lock, val);
120 
121 	msval = dst_metric_raw(dst, RTAX_RTT);
122 	tcp_metric_set(tm, TCP_METRIC_RTT, msval * USEC_PER_MSEC);
123 
124 	msval = dst_metric_raw(dst, RTAX_RTTVAR);
125 	tcp_metric_set(tm, TCP_METRIC_RTTVAR, msval * USEC_PER_MSEC);
126 	tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
127 		       dst_metric_raw(dst, RTAX_SSTHRESH));
128 	tcp_metric_set(tm, TCP_METRIC_CWND,
129 		       dst_metric_raw(dst, RTAX_CWND));
130 	tcp_metric_set(tm, TCP_METRIC_REORDERING,
131 		       dst_metric_raw(dst, RTAX_REORDERING));
132 	if (fastopen_clear) {
133 		write_seqlock(&fastopen_seqlock);
134 		tm->tcpm_fastopen.mss = 0;
135 		tm->tcpm_fastopen.syn_loss = 0;
136 		tm->tcpm_fastopen.try_exp = 0;
137 		tm->tcpm_fastopen.cookie.exp = false;
138 		tm->tcpm_fastopen.cookie.len = 0;
139 		write_sequnlock(&fastopen_seqlock);
140 	}
141 }
142 
143 #define TCP_METRICS_TIMEOUT		(60 * 60 * HZ)
144 
145 static void tcpm_check_stamp(struct tcp_metrics_block *tm,
146 			     const struct dst_entry *dst)
147 {
148 	unsigned long limit;
149 
150 	if (!tm)
151 		return;
152 	limit = READ_ONCE(tm->tcpm_stamp) + TCP_METRICS_TIMEOUT;
153 	if (unlikely(time_after(jiffies, limit)))
154 		tcpm_suck_dst(tm, dst, false);
155 }
156 
157 #define TCP_METRICS_RECLAIM_DEPTH	5
158 #define TCP_METRICS_RECLAIM_PTR		(struct tcp_metrics_block *) 0x1UL
159 
160 #define deref_locked(p)	\
161 	rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
162 
163 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
164 					  struct inetpeer_addr *saddr,
165 					  struct inetpeer_addr *daddr,
166 					  unsigned int hash)
167 {
168 	struct tcp_metrics_block *tm;
169 	struct net *net;
170 	bool reclaim = false;
171 
172 	spin_lock_bh(&tcp_metrics_lock);
173 	net = dev_net(dst->dev);
174 
175 	/* While waiting for the spin-lock the cache might have been populated
176 	 * with this entry and so we have to check again.
177 	 */
178 	tm = __tcp_get_metrics(saddr, daddr, net, hash);
179 	if (tm == TCP_METRICS_RECLAIM_PTR) {
180 		reclaim = true;
181 		tm = NULL;
182 	}
183 	if (tm) {
184 		tcpm_check_stamp(tm, dst);
185 		goto out_unlock;
186 	}
187 
188 	if (unlikely(reclaim)) {
189 		struct tcp_metrics_block *oldest;
190 
191 		oldest = deref_locked(tcp_metrics_hash[hash].chain);
192 		for (tm = deref_locked(oldest->tcpm_next); tm;
193 		     tm = deref_locked(tm->tcpm_next)) {
194 			if (time_before(READ_ONCE(tm->tcpm_stamp),
195 					READ_ONCE(oldest->tcpm_stamp)))
196 				oldest = tm;
197 		}
198 		tm = oldest;
199 	} else {
200 		tm = kzalloc(sizeof(*tm), GFP_ATOMIC);
201 		if (!tm)
202 			goto out_unlock;
203 	}
204 	/* Paired with the READ_ONCE() in tm_net() */
205 	WRITE_ONCE(tm->tcpm_net, net);
206 
207 	tm->tcpm_saddr = *saddr;
208 	tm->tcpm_daddr = *daddr;
209 
210 	tcpm_suck_dst(tm, dst, reclaim);
211 
212 	if (likely(!reclaim)) {
213 		tm->tcpm_next = tcp_metrics_hash[hash].chain;
214 		rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
215 	}
216 
217 out_unlock:
218 	spin_unlock_bh(&tcp_metrics_lock);
219 	return tm;
220 }
221 
222 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
223 {
224 	if (tm)
225 		return tm;
226 	if (depth > TCP_METRICS_RECLAIM_DEPTH)
227 		return TCP_METRICS_RECLAIM_PTR;
228 	return NULL;
229 }
230 
231 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
232 						   const struct inetpeer_addr *daddr,
233 						   struct net *net, unsigned int hash)
234 {
235 	struct tcp_metrics_block *tm;
236 	int depth = 0;
237 
238 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
239 	     tm = rcu_dereference(tm->tcpm_next)) {
240 		if (addr_same(&tm->tcpm_saddr, saddr) &&
241 		    addr_same(&tm->tcpm_daddr, daddr) &&
242 		    net_eq(tm_net(tm), net))
243 			break;
244 		depth++;
245 	}
246 	return tcp_get_encode(tm, depth);
247 }
248 
249 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
250 						       struct dst_entry *dst)
251 {
252 	struct tcp_metrics_block *tm;
253 	struct inetpeer_addr saddr, daddr;
254 	unsigned int hash;
255 	struct net *net;
256 
257 	saddr.family = req->rsk_ops->family;
258 	daddr.family = req->rsk_ops->family;
259 	switch (daddr.family) {
260 	case AF_INET:
261 		inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
262 		inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
263 		hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
264 		break;
265 #if IS_ENABLED(CONFIG_IPV6)
266 	case AF_INET6:
267 		inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
268 		inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr);
269 		hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
270 		break;
271 #endif
272 	default:
273 		return NULL;
274 	}
275 
276 	net = dev_net(dst->dev);
277 	hash ^= net_hash_mix(net);
278 	hash = hash_32(hash, tcp_metrics_hash_log);
279 
280 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
281 	     tm = rcu_dereference(tm->tcpm_next)) {
282 		if (addr_same(&tm->tcpm_saddr, &saddr) &&
283 		    addr_same(&tm->tcpm_daddr, &daddr) &&
284 		    net_eq(tm_net(tm), net))
285 			break;
286 	}
287 	tcpm_check_stamp(tm, dst);
288 	return tm;
289 }
290 
291 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
292 						 struct dst_entry *dst,
293 						 bool create)
294 {
295 	struct tcp_metrics_block *tm;
296 	struct inetpeer_addr saddr, daddr;
297 	unsigned int hash;
298 	struct net *net;
299 
300 	if (sk->sk_family == AF_INET) {
301 		inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
302 		inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
303 		hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
304 	}
305 #if IS_ENABLED(CONFIG_IPV6)
306 	else if (sk->sk_family == AF_INET6) {
307 		if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
308 			inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
309 			inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
310 			hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
311 		} else {
312 			inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
313 			inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
314 			hash = ipv6_addr_hash(&sk->sk_v6_daddr);
315 		}
316 	}
317 #endif
318 	else
319 		return NULL;
320 
321 	net = dev_net(dst->dev);
322 	hash ^= net_hash_mix(net);
323 	hash = hash_32(hash, tcp_metrics_hash_log);
324 
325 	tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
326 	if (tm == TCP_METRICS_RECLAIM_PTR)
327 		tm = NULL;
328 	if (!tm && create)
329 		tm = tcpm_new(dst, &saddr, &daddr, hash);
330 	else
331 		tcpm_check_stamp(tm, dst);
332 
333 	return tm;
334 }
335 
336 /* Save metrics learned by this TCP session.  This function is called
337  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
338  * or goes from LAST-ACK to CLOSE.
339  */
340 void tcp_update_metrics(struct sock *sk)
341 {
342 	const struct inet_connection_sock *icsk = inet_csk(sk);
343 	struct dst_entry *dst = __sk_dst_get(sk);
344 	struct tcp_sock *tp = tcp_sk(sk);
345 	struct net *net = sock_net(sk);
346 	struct tcp_metrics_block *tm;
347 	unsigned long rtt;
348 	u32 val;
349 	int m;
350 
351 	sk_dst_confirm(sk);
352 	if (READ_ONCE(net->ipv4.sysctl_tcp_nometrics_save) || !dst)
353 		return;
354 
355 	rcu_read_lock();
356 	if (icsk->icsk_backoff || !tp->srtt_us) {
357 		/* This session failed to estimate rtt. Why?
358 		 * Probably, no packets returned in time.  Reset our
359 		 * results.
360 		 */
361 		tm = tcp_get_metrics(sk, dst, false);
362 		if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
363 			tcp_metric_set(tm, TCP_METRIC_RTT, 0);
364 		goto out_unlock;
365 	} else
366 		tm = tcp_get_metrics(sk, dst, true);
367 
368 	if (!tm)
369 		goto out_unlock;
370 
371 	rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
372 	m = rtt - tp->srtt_us;
373 
374 	/* If newly calculated rtt larger than stored one, store new
375 	 * one. Otherwise, use EWMA. Remember, rtt overestimation is
376 	 * always better than underestimation.
377 	 */
378 	if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
379 		if (m <= 0)
380 			rtt = tp->srtt_us;
381 		else
382 			rtt -= (m >> 3);
383 		tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
384 	}
385 
386 	if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
387 		unsigned long var;
388 
389 		if (m < 0)
390 			m = -m;
391 
392 		/* Scale deviation to rttvar fixed point */
393 		m >>= 1;
394 		if (m < tp->mdev_us)
395 			m = tp->mdev_us;
396 
397 		var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
398 		if (m >= var)
399 			var = m;
400 		else
401 			var -= (var - m) >> 2;
402 
403 		tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
404 	}
405 
406 	if (tcp_in_initial_slowstart(tp)) {
407 		/* Slow start still did not finish. */
408 		if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
409 		    !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
410 			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
411 			if (val && (tcp_snd_cwnd(tp) >> 1) > val)
412 				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
413 					       tcp_snd_cwnd(tp) >> 1);
414 		}
415 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
416 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
417 			if (tcp_snd_cwnd(tp) > val)
418 				tcp_metric_set(tm, TCP_METRIC_CWND,
419 					       tcp_snd_cwnd(tp));
420 		}
421 	} else if (!tcp_in_slow_start(tp) &&
422 		   icsk->icsk_ca_state == TCP_CA_Open) {
423 		/* Cong. avoidance phase, cwnd is reliable. */
424 		if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
425 		    !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
426 			tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
427 				       max(tcp_snd_cwnd(tp) >> 1, tp->snd_ssthresh));
428 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
429 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
430 			tcp_metric_set(tm, TCP_METRIC_CWND, (val + tcp_snd_cwnd(tp)) >> 1);
431 		}
432 	} else {
433 		/* Else slow start did not finish, cwnd is non-sense,
434 		 * ssthresh may be also invalid.
435 		 */
436 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
437 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
438 			tcp_metric_set(tm, TCP_METRIC_CWND,
439 				       (val + tp->snd_ssthresh) >> 1);
440 		}
441 		if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
442 		    !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
443 			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
444 			if (val && tp->snd_ssthresh > val)
445 				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
446 					       tp->snd_ssthresh);
447 		}
448 		if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
449 			val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
450 			if (val < tp->reordering &&
451 			    tp->reordering !=
452 			    READ_ONCE(net->ipv4.sysctl_tcp_reordering))
453 				tcp_metric_set(tm, TCP_METRIC_REORDERING,
454 					       tp->reordering);
455 		}
456 	}
457 	WRITE_ONCE(tm->tcpm_stamp, jiffies);
458 out_unlock:
459 	rcu_read_unlock();
460 }
461 
462 /* Initialize metrics on socket. */
463 
464 void tcp_init_metrics(struct sock *sk)
465 {
466 	struct dst_entry *dst = __sk_dst_get(sk);
467 	struct tcp_sock *tp = tcp_sk(sk);
468 	struct net *net = sock_net(sk);
469 	struct tcp_metrics_block *tm;
470 	u32 val, crtt = 0; /* cached RTT scaled by 8 */
471 
472 	sk_dst_confirm(sk);
473 	if (!dst)
474 		goto reset;
475 
476 	rcu_read_lock();
477 	tm = tcp_get_metrics(sk, dst, true);
478 	if (!tm) {
479 		rcu_read_unlock();
480 		goto reset;
481 	}
482 
483 	if (tcp_metric_locked(tm, TCP_METRIC_CWND))
484 		tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
485 
486 	val = READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) ?
487 	      0 : tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
488 	if (val) {
489 		tp->snd_ssthresh = val;
490 		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
491 			tp->snd_ssthresh = tp->snd_cwnd_clamp;
492 	} else {
493 		/* ssthresh may have been reduced unnecessarily during.
494 		 * 3WHS. Restore it back to its initial default.
495 		 */
496 		tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
497 	}
498 	val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
499 	if (val && tp->reordering != val)
500 		tp->reordering = val;
501 
502 	crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
503 	rcu_read_unlock();
504 reset:
505 	/* The initial RTT measurement from the SYN/SYN-ACK is not ideal
506 	 * to seed the RTO for later data packets because SYN packets are
507 	 * small. Use the per-dst cached values to seed the RTO but keep
508 	 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
509 	 * Later the RTO will be updated immediately upon obtaining the first
510 	 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
511 	 * influences the first RTO but not later RTT estimation.
512 	 *
513 	 * But if RTT is not available from the SYN (due to retransmits or
514 	 * syn cookies) or the cache, force a conservative 3secs timeout.
515 	 *
516 	 * A bit of theory. RTT is time passed after "normal" sized packet
517 	 * is sent until it is ACKed. In normal circumstances sending small
518 	 * packets force peer to delay ACKs and calculation is correct too.
519 	 * The algorithm is adaptive and, provided we follow specs, it
520 	 * NEVER underestimate RTT. BUT! If peer tries to make some clever
521 	 * tricks sort of "quick acks" for time long enough to decrease RTT
522 	 * to low value, and then abruptly stops to do it and starts to delay
523 	 * ACKs, wait for troubles.
524 	 */
525 	if (crtt > tp->srtt_us) {
526 		/* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
527 		crtt /= 8 * USEC_PER_SEC / HZ;
528 		inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
529 	} else if (tp->srtt_us == 0) {
530 		/* RFC6298: 5.7 We've failed to get a valid RTT sample from
531 		 * 3WHS. This is most likely due to retransmission,
532 		 * including spurious one. Reset the RTO back to 3secs
533 		 * from the more aggressive 1sec to avoid more spurious
534 		 * retransmission.
535 		 */
536 		tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
537 		tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
538 
539 		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
540 	}
541 }
542 
543 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
544 {
545 	struct tcp_metrics_block *tm;
546 	bool ret;
547 
548 	if (!dst)
549 		return false;
550 
551 	rcu_read_lock();
552 	tm = __tcp_get_metrics_req(req, dst);
553 	if (tm && tcp_metric_get(tm, TCP_METRIC_RTT))
554 		ret = true;
555 	else
556 		ret = false;
557 	rcu_read_unlock();
558 
559 	return ret;
560 }
561 
562 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
563 			    struct tcp_fastopen_cookie *cookie)
564 {
565 	struct tcp_metrics_block *tm;
566 
567 	rcu_read_lock();
568 	tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
569 	if (tm) {
570 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
571 		unsigned int seq;
572 
573 		do {
574 			seq = read_seqbegin(&fastopen_seqlock);
575 			if (tfom->mss)
576 				*mss = tfom->mss;
577 			*cookie = tfom->cookie;
578 			if (cookie->len <= 0 && tfom->try_exp == 1)
579 				cookie->exp = true;
580 		} while (read_seqretry(&fastopen_seqlock, seq));
581 	}
582 	rcu_read_unlock();
583 }
584 
585 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
586 			    struct tcp_fastopen_cookie *cookie, bool syn_lost,
587 			    u16 try_exp)
588 {
589 	struct dst_entry *dst = __sk_dst_get(sk);
590 	struct tcp_metrics_block *tm;
591 
592 	if (!dst)
593 		return;
594 	rcu_read_lock();
595 	tm = tcp_get_metrics(sk, dst, true);
596 	if (tm) {
597 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
598 
599 		write_seqlock_bh(&fastopen_seqlock);
600 		if (mss)
601 			tfom->mss = mss;
602 		if (cookie && cookie->len > 0)
603 			tfom->cookie = *cookie;
604 		else if (try_exp > tfom->try_exp &&
605 			 tfom->cookie.len <= 0 && !tfom->cookie.exp)
606 			tfom->try_exp = try_exp;
607 		if (syn_lost) {
608 			++tfom->syn_loss;
609 			tfom->last_syn_loss = jiffies;
610 		} else
611 			tfom->syn_loss = 0;
612 		write_sequnlock_bh(&fastopen_seqlock);
613 	}
614 	rcu_read_unlock();
615 }
616 
617 static struct genl_family tcp_metrics_nl_family;
618 
619 static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
620 	[TCP_METRICS_ATTR_ADDR_IPV4]	= { .type = NLA_U32, },
621 	[TCP_METRICS_ATTR_ADDR_IPV6]	= { .type = NLA_BINARY,
622 					    .len = sizeof(struct in6_addr), },
623 	/* Following attributes are not received for GET/DEL,
624 	 * we keep them for reference
625 	 */
626 #if 0
627 	[TCP_METRICS_ATTR_AGE]		= { .type = NLA_MSECS, },
628 	[TCP_METRICS_ATTR_TW_TSVAL]	= { .type = NLA_U32, },
629 	[TCP_METRICS_ATTR_TW_TS_STAMP]	= { .type = NLA_S32, },
630 	[TCP_METRICS_ATTR_VALS]		= { .type = NLA_NESTED, },
631 	[TCP_METRICS_ATTR_FOPEN_MSS]	= { .type = NLA_U16, },
632 	[TCP_METRICS_ATTR_FOPEN_SYN_DROPS]	= { .type = NLA_U16, },
633 	[TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]	= { .type = NLA_MSECS, },
634 	[TCP_METRICS_ATTR_FOPEN_COOKIE]	= { .type = NLA_BINARY,
635 					    .len = TCP_FASTOPEN_COOKIE_MAX, },
636 #endif
637 };
638 
639 /* Add attributes, caller cancels its header on failure */
640 static int tcp_metrics_fill_info(struct sk_buff *msg,
641 				 struct tcp_metrics_block *tm)
642 {
643 	struct nlattr *nest;
644 	int i;
645 
646 	switch (tm->tcpm_daddr.family) {
647 	case AF_INET:
648 		if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
649 				    inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
650 			goto nla_put_failure;
651 		if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
652 				    inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
653 			goto nla_put_failure;
654 		break;
655 	case AF_INET6:
656 		if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
657 				     inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
658 			goto nla_put_failure;
659 		if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
660 				     inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
661 			goto nla_put_failure;
662 		break;
663 	default:
664 		return -EAFNOSUPPORT;
665 	}
666 
667 	if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
668 			  jiffies - READ_ONCE(tm->tcpm_stamp),
669 			  TCP_METRICS_ATTR_PAD) < 0)
670 		goto nla_put_failure;
671 
672 	{
673 		int n = 0;
674 
675 		nest = nla_nest_start_noflag(msg, TCP_METRICS_ATTR_VALS);
676 		if (!nest)
677 			goto nla_put_failure;
678 		for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
679 			u32 val = tcp_metric_get(tm, i);
680 
681 			if (!val)
682 				continue;
683 			if (i == TCP_METRIC_RTT) {
684 				if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
685 						val) < 0)
686 					goto nla_put_failure;
687 				n++;
688 				val = max(val / 1000, 1U);
689 			}
690 			if (i == TCP_METRIC_RTTVAR) {
691 				if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
692 						val) < 0)
693 					goto nla_put_failure;
694 				n++;
695 				val = max(val / 1000, 1U);
696 			}
697 			if (nla_put_u32(msg, i + 1, val) < 0)
698 				goto nla_put_failure;
699 			n++;
700 		}
701 		if (n)
702 			nla_nest_end(msg, nest);
703 		else
704 			nla_nest_cancel(msg, nest);
705 	}
706 
707 	{
708 		struct tcp_fastopen_metrics tfom_copy[1], *tfom;
709 		unsigned int seq;
710 
711 		do {
712 			seq = read_seqbegin(&fastopen_seqlock);
713 			tfom_copy[0] = tm->tcpm_fastopen;
714 		} while (read_seqretry(&fastopen_seqlock, seq));
715 
716 		tfom = tfom_copy;
717 		if (tfom->mss &&
718 		    nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
719 				tfom->mss) < 0)
720 			goto nla_put_failure;
721 		if (tfom->syn_loss &&
722 		    (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
723 				tfom->syn_loss) < 0 ||
724 		     nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
725 				jiffies - tfom->last_syn_loss,
726 				TCP_METRICS_ATTR_PAD) < 0))
727 			goto nla_put_failure;
728 		if (tfom->cookie.len > 0 &&
729 		    nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
730 			    tfom->cookie.len, tfom->cookie.val) < 0)
731 			goto nla_put_failure;
732 	}
733 
734 	return 0;
735 
736 nla_put_failure:
737 	return -EMSGSIZE;
738 }
739 
740 static int tcp_metrics_dump_info(struct sk_buff *skb,
741 				 struct netlink_callback *cb,
742 				 struct tcp_metrics_block *tm)
743 {
744 	void *hdr;
745 
746 	hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
747 			  &tcp_metrics_nl_family, NLM_F_MULTI,
748 			  TCP_METRICS_CMD_GET);
749 	if (!hdr)
750 		return -EMSGSIZE;
751 
752 	if (tcp_metrics_fill_info(skb, tm) < 0)
753 		goto nla_put_failure;
754 
755 	genlmsg_end(skb, hdr);
756 	return 0;
757 
758 nla_put_failure:
759 	genlmsg_cancel(skb, hdr);
760 	return -EMSGSIZE;
761 }
762 
763 static int tcp_metrics_nl_dump(struct sk_buff *skb,
764 			       struct netlink_callback *cb)
765 {
766 	struct net *net = sock_net(skb->sk);
767 	unsigned int max_rows = 1U << tcp_metrics_hash_log;
768 	unsigned int row, s_row = cb->args[0];
769 	int s_col = cb->args[1], col = s_col;
770 
771 	for (row = s_row; row < max_rows; row++, s_col = 0) {
772 		struct tcp_metrics_block *tm;
773 		struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
774 
775 		rcu_read_lock();
776 		for (col = 0, tm = rcu_dereference(hb->chain); tm;
777 		     tm = rcu_dereference(tm->tcpm_next), col++) {
778 			if (!net_eq(tm_net(tm), net))
779 				continue;
780 			if (col < s_col)
781 				continue;
782 			if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
783 				rcu_read_unlock();
784 				goto done;
785 			}
786 		}
787 		rcu_read_unlock();
788 	}
789 
790 done:
791 	cb->args[0] = row;
792 	cb->args[1] = col;
793 	return skb->len;
794 }
795 
796 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
797 			   unsigned int *hash, int optional, int v4, int v6)
798 {
799 	struct nlattr *a;
800 
801 	a = info->attrs[v4];
802 	if (a) {
803 		inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
804 		if (hash)
805 			*hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
806 		return 0;
807 	}
808 	a = info->attrs[v6];
809 	if (a) {
810 		struct in6_addr in6;
811 
812 		if (nla_len(a) != sizeof(struct in6_addr))
813 			return -EINVAL;
814 		in6 = nla_get_in6_addr(a);
815 		inetpeer_set_addr_v6(addr, &in6);
816 		if (hash)
817 			*hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
818 		return 0;
819 	}
820 	return optional ? 1 : -EAFNOSUPPORT;
821 }
822 
823 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
824 			 unsigned int *hash, int optional)
825 {
826 	return __parse_nl_addr(info, addr, hash, optional,
827 			       TCP_METRICS_ATTR_ADDR_IPV4,
828 			       TCP_METRICS_ATTR_ADDR_IPV6);
829 }
830 
831 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
832 {
833 	return __parse_nl_addr(info, addr, NULL, 0,
834 			       TCP_METRICS_ATTR_SADDR_IPV4,
835 			       TCP_METRICS_ATTR_SADDR_IPV6);
836 }
837 
838 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
839 {
840 	struct tcp_metrics_block *tm;
841 	struct inetpeer_addr saddr, daddr;
842 	unsigned int hash;
843 	struct sk_buff *msg;
844 	struct net *net = genl_info_net(info);
845 	void *reply;
846 	int ret;
847 	bool src = true;
848 
849 	ret = parse_nl_addr(info, &daddr, &hash, 0);
850 	if (ret < 0)
851 		return ret;
852 
853 	ret = parse_nl_saddr(info, &saddr);
854 	if (ret < 0)
855 		src = false;
856 
857 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
858 	if (!msg)
859 		return -ENOMEM;
860 
861 	reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
862 				  info->genlhdr->cmd);
863 	if (!reply)
864 		goto nla_put_failure;
865 
866 	hash ^= net_hash_mix(net);
867 	hash = hash_32(hash, tcp_metrics_hash_log);
868 	ret = -ESRCH;
869 	rcu_read_lock();
870 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
871 	     tm = rcu_dereference(tm->tcpm_next)) {
872 		if (addr_same(&tm->tcpm_daddr, &daddr) &&
873 		    (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
874 		    net_eq(tm_net(tm), net)) {
875 			ret = tcp_metrics_fill_info(msg, tm);
876 			break;
877 		}
878 	}
879 	rcu_read_unlock();
880 	if (ret < 0)
881 		goto out_free;
882 
883 	genlmsg_end(msg, reply);
884 	return genlmsg_reply(msg, info);
885 
886 nla_put_failure:
887 	ret = -EMSGSIZE;
888 
889 out_free:
890 	nlmsg_free(msg);
891 	return ret;
892 }
893 
894 static void tcp_metrics_flush_all(struct net *net)
895 {
896 	unsigned int max_rows = 1U << tcp_metrics_hash_log;
897 	struct tcpm_hash_bucket *hb = tcp_metrics_hash;
898 	struct tcp_metrics_block *tm;
899 	unsigned int row;
900 
901 	for (row = 0; row < max_rows; row++, hb++) {
902 		struct tcp_metrics_block __rcu **pp;
903 		bool match;
904 
905 		spin_lock_bh(&tcp_metrics_lock);
906 		pp = &hb->chain;
907 		for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
908 			match = net ? net_eq(tm_net(tm), net) :
909 				!refcount_read(&tm_net(tm)->ns.count);
910 			if (match) {
911 				*pp = tm->tcpm_next;
912 				kfree_rcu(tm, rcu_head);
913 			} else {
914 				pp = &tm->tcpm_next;
915 			}
916 		}
917 		spin_unlock_bh(&tcp_metrics_lock);
918 	}
919 }
920 
921 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
922 {
923 	struct tcpm_hash_bucket *hb;
924 	struct tcp_metrics_block *tm;
925 	struct tcp_metrics_block __rcu **pp;
926 	struct inetpeer_addr saddr, daddr;
927 	unsigned int hash;
928 	struct net *net = genl_info_net(info);
929 	int ret;
930 	bool src = true, found = false;
931 
932 	ret = parse_nl_addr(info, &daddr, &hash, 1);
933 	if (ret < 0)
934 		return ret;
935 	if (ret > 0) {
936 		tcp_metrics_flush_all(net);
937 		return 0;
938 	}
939 	ret = parse_nl_saddr(info, &saddr);
940 	if (ret < 0)
941 		src = false;
942 
943 	hash ^= net_hash_mix(net);
944 	hash = hash_32(hash, tcp_metrics_hash_log);
945 	hb = tcp_metrics_hash + hash;
946 	pp = &hb->chain;
947 	spin_lock_bh(&tcp_metrics_lock);
948 	for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
949 		if (addr_same(&tm->tcpm_daddr, &daddr) &&
950 		    (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
951 		    net_eq(tm_net(tm), net)) {
952 			*pp = tm->tcpm_next;
953 			kfree_rcu(tm, rcu_head);
954 			found = true;
955 		} else {
956 			pp = &tm->tcpm_next;
957 		}
958 	}
959 	spin_unlock_bh(&tcp_metrics_lock);
960 	if (!found)
961 		return -ESRCH;
962 	return 0;
963 }
964 
965 static const struct genl_small_ops tcp_metrics_nl_ops[] = {
966 	{
967 		.cmd = TCP_METRICS_CMD_GET,
968 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
969 		.doit = tcp_metrics_nl_cmd_get,
970 		.dumpit = tcp_metrics_nl_dump,
971 	},
972 	{
973 		.cmd = TCP_METRICS_CMD_DEL,
974 		.validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
975 		.doit = tcp_metrics_nl_cmd_del,
976 		.flags = GENL_ADMIN_PERM,
977 	},
978 };
979 
980 static struct genl_family tcp_metrics_nl_family __ro_after_init = {
981 	.hdrsize	= 0,
982 	.name		= TCP_METRICS_GENL_NAME,
983 	.version	= TCP_METRICS_GENL_VERSION,
984 	.maxattr	= TCP_METRICS_ATTR_MAX,
985 	.policy = tcp_metrics_nl_policy,
986 	.netnsok	= true,
987 	.module		= THIS_MODULE,
988 	.small_ops	= tcp_metrics_nl_ops,
989 	.n_small_ops	= ARRAY_SIZE(tcp_metrics_nl_ops),
990 	.resv_start_op	= TCP_METRICS_CMD_DEL + 1,
991 };
992 
993 static unsigned int tcpmhash_entries;
994 static int __init set_tcpmhash_entries(char *str)
995 {
996 	ssize_t ret;
997 
998 	if (!str)
999 		return 0;
1000 
1001 	ret = kstrtouint(str, 0, &tcpmhash_entries);
1002 	if (ret)
1003 		return 0;
1004 
1005 	return 1;
1006 }
1007 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1008 
1009 static int __net_init tcp_net_metrics_init(struct net *net)
1010 {
1011 	size_t size;
1012 	unsigned int slots;
1013 
1014 	if (!net_eq(net, &init_net))
1015 		return 0;
1016 
1017 	slots = tcpmhash_entries;
1018 	if (!slots) {
1019 		if (totalram_pages() >= 128 * 1024)
1020 			slots = 16 * 1024;
1021 		else
1022 			slots = 8 * 1024;
1023 	}
1024 
1025 	tcp_metrics_hash_log = order_base_2(slots);
1026 	size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1027 
1028 	tcp_metrics_hash = kvzalloc(size, GFP_KERNEL);
1029 	if (!tcp_metrics_hash)
1030 		return -ENOMEM;
1031 
1032 	return 0;
1033 }
1034 
1035 static void __net_exit tcp_net_metrics_exit_batch(struct list_head *net_exit_list)
1036 {
1037 	tcp_metrics_flush_all(NULL);
1038 }
1039 
1040 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1041 	.init		=	tcp_net_metrics_init,
1042 	.exit_batch	=	tcp_net_metrics_exit_batch,
1043 };
1044 
1045 void __init tcp_metrics_init(void)
1046 {
1047 	int ret;
1048 
1049 	ret = register_pernet_subsys(&tcp_net_metrics_ops);
1050 	if (ret < 0)
1051 		panic("Could not allocate the tcp_metrics hash table\n");
1052 
1053 	ret = genl_register_family(&tcp_metrics_nl_family);
1054 	if (ret < 0)
1055 		panic("Could not register tcp_metrics generic netlink\n");
1056 }
1057