xref: /openbmc/linux/net/ipv4/tcp_metrics.c (revision 7211ec63)
1 #include <linux/rcupdate.h>
2 #include <linux/spinlock.h>
3 #include <linux/jiffies.h>
4 #include <linux/module.h>
5 #include <linux/cache.h>
6 #include <linux/slab.h>
7 #include <linux/init.h>
8 #include <linux/tcp.h>
9 #include <linux/hash.h>
10 #include <linux/tcp_metrics.h>
11 #include <linux/vmalloc.h>
12 
13 #include <net/inet_connection_sock.h>
14 #include <net/net_namespace.h>
15 #include <net/request_sock.h>
16 #include <net/inetpeer.h>
17 #include <net/sock.h>
18 #include <net/ipv6.h>
19 #include <net/dst.h>
20 #include <net/tcp.h>
21 #include <net/genetlink.h>
22 
23 int sysctl_tcp_nometrics_save __read_mostly;
24 
25 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
26 						   const struct inetpeer_addr *daddr,
27 						   struct net *net, unsigned int hash);
28 
29 struct tcp_fastopen_metrics {
30 	u16	mss;
31 	u16	syn_loss:10,		/* Recurring Fast Open SYN losses */
32 		try_exp:2;		/* Request w/ exp. option (once) */
33 	unsigned long	last_syn_loss;	/* Last Fast Open SYN loss */
34 	struct	tcp_fastopen_cookie	cookie;
35 };
36 
37 /* TCP_METRIC_MAX includes 2 extra fields for userspace compatibility
38  * Kernel only stores RTT and RTTVAR in usec resolution
39  */
40 #define TCP_METRIC_MAX_KERNEL (TCP_METRIC_MAX - 2)
41 
42 struct tcp_metrics_block {
43 	struct tcp_metrics_block __rcu	*tcpm_next;
44 	possible_net_t			tcpm_net;
45 	struct inetpeer_addr		tcpm_saddr;
46 	struct inetpeer_addr		tcpm_daddr;
47 	unsigned long			tcpm_stamp;
48 	u32				tcpm_lock;
49 	u32				tcpm_vals[TCP_METRIC_MAX_KERNEL + 1];
50 	struct tcp_fastopen_metrics	tcpm_fastopen;
51 
52 	struct rcu_head			rcu_head;
53 };
54 
55 static inline struct net *tm_net(struct tcp_metrics_block *tm)
56 {
57 	return read_pnet(&tm->tcpm_net);
58 }
59 
60 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
61 			      enum tcp_metric_index idx)
62 {
63 	return tm->tcpm_lock & (1 << idx);
64 }
65 
66 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
67 			  enum tcp_metric_index idx)
68 {
69 	return tm->tcpm_vals[idx];
70 }
71 
72 static void tcp_metric_set(struct tcp_metrics_block *tm,
73 			   enum tcp_metric_index idx,
74 			   u32 val)
75 {
76 	tm->tcpm_vals[idx] = val;
77 }
78 
79 static bool addr_same(const struct inetpeer_addr *a,
80 		      const struct inetpeer_addr *b)
81 {
82 	return inetpeer_addr_cmp(a, b) == 0;
83 }
84 
85 struct tcpm_hash_bucket {
86 	struct tcp_metrics_block __rcu	*chain;
87 };
88 
89 static struct tcpm_hash_bucket	*tcp_metrics_hash __read_mostly;
90 static unsigned int		tcp_metrics_hash_log __read_mostly;
91 
92 static DEFINE_SPINLOCK(tcp_metrics_lock);
93 
94 static void tcpm_suck_dst(struct tcp_metrics_block *tm,
95 			  const struct dst_entry *dst,
96 			  bool fastopen_clear)
97 {
98 	u32 msval;
99 	u32 val;
100 
101 	tm->tcpm_stamp = jiffies;
102 
103 	val = 0;
104 	if (dst_metric_locked(dst, RTAX_RTT))
105 		val |= 1 << TCP_METRIC_RTT;
106 	if (dst_metric_locked(dst, RTAX_RTTVAR))
107 		val |= 1 << TCP_METRIC_RTTVAR;
108 	if (dst_metric_locked(dst, RTAX_SSTHRESH))
109 		val |= 1 << TCP_METRIC_SSTHRESH;
110 	if (dst_metric_locked(dst, RTAX_CWND))
111 		val |= 1 << TCP_METRIC_CWND;
112 	if (dst_metric_locked(dst, RTAX_REORDERING))
113 		val |= 1 << TCP_METRIC_REORDERING;
114 	tm->tcpm_lock = val;
115 
116 	msval = dst_metric_raw(dst, RTAX_RTT);
117 	tm->tcpm_vals[TCP_METRIC_RTT] = msval * USEC_PER_MSEC;
118 
119 	msval = dst_metric_raw(dst, RTAX_RTTVAR);
120 	tm->tcpm_vals[TCP_METRIC_RTTVAR] = msval * USEC_PER_MSEC;
121 	tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
122 	tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
123 	tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
124 	if (fastopen_clear) {
125 		tm->tcpm_fastopen.mss = 0;
126 		tm->tcpm_fastopen.syn_loss = 0;
127 		tm->tcpm_fastopen.try_exp = 0;
128 		tm->tcpm_fastopen.cookie.exp = false;
129 		tm->tcpm_fastopen.cookie.len = 0;
130 	}
131 }
132 
133 #define TCP_METRICS_TIMEOUT		(60 * 60 * HZ)
134 
135 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
136 {
137 	if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
138 		tcpm_suck_dst(tm, dst, false);
139 }
140 
141 #define TCP_METRICS_RECLAIM_DEPTH	5
142 #define TCP_METRICS_RECLAIM_PTR		(struct tcp_metrics_block *) 0x1UL
143 
144 #define deref_locked(p)	\
145 	rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
146 
147 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
148 					  struct inetpeer_addr *saddr,
149 					  struct inetpeer_addr *daddr,
150 					  unsigned int hash)
151 {
152 	struct tcp_metrics_block *tm;
153 	struct net *net;
154 	bool reclaim = false;
155 
156 	spin_lock_bh(&tcp_metrics_lock);
157 	net = dev_net(dst->dev);
158 
159 	/* While waiting for the spin-lock the cache might have been populated
160 	 * with this entry and so we have to check again.
161 	 */
162 	tm = __tcp_get_metrics(saddr, daddr, net, hash);
163 	if (tm == TCP_METRICS_RECLAIM_PTR) {
164 		reclaim = true;
165 		tm = NULL;
166 	}
167 	if (tm) {
168 		tcpm_check_stamp(tm, dst);
169 		goto out_unlock;
170 	}
171 
172 	if (unlikely(reclaim)) {
173 		struct tcp_metrics_block *oldest;
174 
175 		oldest = deref_locked(tcp_metrics_hash[hash].chain);
176 		for (tm = deref_locked(oldest->tcpm_next); tm;
177 		     tm = deref_locked(tm->tcpm_next)) {
178 			if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
179 				oldest = tm;
180 		}
181 		tm = oldest;
182 	} else {
183 		tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
184 		if (!tm)
185 			goto out_unlock;
186 	}
187 	write_pnet(&tm->tcpm_net, net);
188 	tm->tcpm_saddr = *saddr;
189 	tm->tcpm_daddr = *daddr;
190 
191 	tcpm_suck_dst(tm, dst, true);
192 
193 	if (likely(!reclaim)) {
194 		tm->tcpm_next = tcp_metrics_hash[hash].chain;
195 		rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
196 	}
197 
198 out_unlock:
199 	spin_unlock_bh(&tcp_metrics_lock);
200 	return tm;
201 }
202 
203 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
204 {
205 	if (tm)
206 		return tm;
207 	if (depth > TCP_METRICS_RECLAIM_DEPTH)
208 		return TCP_METRICS_RECLAIM_PTR;
209 	return NULL;
210 }
211 
212 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
213 						   const struct inetpeer_addr *daddr,
214 						   struct net *net, unsigned int hash)
215 {
216 	struct tcp_metrics_block *tm;
217 	int depth = 0;
218 
219 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
220 	     tm = rcu_dereference(tm->tcpm_next)) {
221 		if (addr_same(&tm->tcpm_saddr, saddr) &&
222 		    addr_same(&tm->tcpm_daddr, daddr) &&
223 		    net_eq(tm_net(tm), net))
224 			break;
225 		depth++;
226 	}
227 	return tcp_get_encode(tm, depth);
228 }
229 
230 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
231 						       struct dst_entry *dst)
232 {
233 	struct tcp_metrics_block *tm;
234 	struct inetpeer_addr saddr, daddr;
235 	unsigned int hash;
236 	struct net *net;
237 
238 	saddr.family = req->rsk_ops->family;
239 	daddr.family = req->rsk_ops->family;
240 	switch (daddr.family) {
241 	case AF_INET:
242 		inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
243 		inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
244 		hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
245 		break;
246 #if IS_ENABLED(CONFIG_IPV6)
247 	case AF_INET6:
248 		inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
249 		inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr);
250 		hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
251 		break;
252 #endif
253 	default:
254 		return NULL;
255 	}
256 
257 	net = dev_net(dst->dev);
258 	hash ^= net_hash_mix(net);
259 	hash = hash_32(hash, tcp_metrics_hash_log);
260 
261 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
262 	     tm = rcu_dereference(tm->tcpm_next)) {
263 		if (addr_same(&tm->tcpm_saddr, &saddr) &&
264 		    addr_same(&tm->tcpm_daddr, &daddr) &&
265 		    net_eq(tm_net(tm), net))
266 			break;
267 	}
268 	tcpm_check_stamp(tm, dst);
269 	return tm;
270 }
271 
272 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
273 						 struct dst_entry *dst,
274 						 bool create)
275 {
276 	struct tcp_metrics_block *tm;
277 	struct inetpeer_addr saddr, daddr;
278 	unsigned int hash;
279 	struct net *net;
280 
281 	if (sk->sk_family == AF_INET) {
282 		inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
283 		inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
284 		hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
285 	}
286 #if IS_ENABLED(CONFIG_IPV6)
287 	else if (sk->sk_family == AF_INET6) {
288 		if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
289 			inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
290 			inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
291 			hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
292 		} else {
293 			inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
294 			inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
295 			hash = ipv6_addr_hash(&sk->sk_v6_daddr);
296 		}
297 	}
298 #endif
299 	else
300 		return NULL;
301 
302 	net = dev_net(dst->dev);
303 	hash ^= net_hash_mix(net);
304 	hash = hash_32(hash, tcp_metrics_hash_log);
305 
306 	tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
307 	if (tm == TCP_METRICS_RECLAIM_PTR)
308 		tm = NULL;
309 	if (!tm && create)
310 		tm = tcpm_new(dst, &saddr, &daddr, hash);
311 	else
312 		tcpm_check_stamp(tm, dst);
313 
314 	return tm;
315 }
316 
317 /* Save metrics learned by this TCP session.  This function is called
318  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
319  * or goes from LAST-ACK to CLOSE.
320  */
321 void tcp_update_metrics(struct sock *sk)
322 {
323 	const struct inet_connection_sock *icsk = inet_csk(sk);
324 	struct dst_entry *dst = __sk_dst_get(sk);
325 	struct tcp_sock *tp = tcp_sk(sk);
326 	struct net *net = sock_net(sk);
327 	struct tcp_metrics_block *tm;
328 	unsigned long rtt;
329 	u32 val;
330 	int m;
331 
332 	sk_dst_confirm(sk);
333 	if (sysctl_tcp_nometrics_save || !dst)
334 		return;
335 
336 	rcu_read_lock();
337 	if (icsk->icsk_backoff || !tp->srtt_us) {
338 		/* This session failed to estimate rtt. Why?
339 		 * Probably, no packets returned in time.  Reset our
340 		 * results.
341 		 */
342 		tm = tcp_get_metrics(sk, dst, false);
343 		if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
344 			tcp_metric_set(tm, TCP_METRIC_RTT, 0);
345 		goto out_unlock;
346 	} else
347 		tm = tcp_get_metrics(sk, dst, true);
348 
349 	if (!tm)
350 		goto out_unlock;
351 
352 	rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
353 	m = rtt - tp->srtt_us;
354 
355 	/* If newly calculated rtt larger than stored one, store new
356 	 * one. Otherwise, use EWMA. Remember, rtt overestimation is
357 	 * always better than underestimation.
358 	 */
359 	if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
360 		if (m <= 0)
361 			rtt = tp->srtt_us;
362 		else
363 			rtt -= (m >> 3);
364 		tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
365 	}
366 
367 	if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
368 		unsigned long var;
369 
370 		if (m < 0)
371 			m = -m;
372 
373 		/* Scale deviation to rttvar fixed point */
374 		m >>= 1;
375 		if (m < tp->mdev_us)
376 			m = tp->mdev_us;
377 
378 		var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
379 		if (m >= var)
380 			var = m;
381 		else
382 			var -= (var - m) >> 2;
383 
384 		tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
385 	}
386 
387 	if (tcp_in_initial_slowstart(tp)) {
388 		/* Slow start still did not finish. */
389 		if (!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 (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
405 			tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
406 				       max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
407 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
408 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
409 			tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
410 		}
411 	} else {
412 		/* Else slow start did not finish, cwnd is non-sense,
413 		 * ssthresh may be also invalid.
414 		 */
415 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
416 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
417 			tcp_metric_set(tm, TCP_METRIC_CWND,
418 				       (val + tp->snd_ssthresh) >> 1);
419 		}
420 		if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
421 			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
422 			if (val && tp->snd_ssthresh > val)
423 				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
424 					       tp->snd_ssthresh);
425 		}
426 		if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
427 			val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
428 			if (val < tp->reordering &&
429 			    tp->reordering != net->ipv4.sysctl_tcp_reordering)
430 				tcp_metric_set(tm, TCP_METRIC_REORDERING,
431 					       tp->reordering);
432 		}
433 	}
434 	tm->tcpm_stamp = jiffies;
435 out_unlock:
436 	rcu_read_unlock();
437 }
438 
439 /* Initialize metrics on socket. */
440 
441 void tcp_init_metrics(struct sock *sk)
442 {
443 	struct dst_entry *dst = __sk_dst_get(sk);
444 	struct tcp_sock *tp = tcp_sk(sk);
445 	struct tcp_metrics_block *tm;
446 	u32 val, crtt = 0; /* cached RTT scaled by 8 */
447 
448 	sk_dst_confirm(sk);
449 	if (!dst)
450 		goto reset;
451 
452 	rcu_read_lock();
453 	tm = tcp_get_metrics(sk, dst, true);
454 	if (!tm) {
455 		rcu_read_unlock();
456 		goto reset;
457 	}
458 
459 	if (tcp_metric_locked(tm, TCP_METRIC_CWND))
460 		tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
461 
462 	val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
463 	if (val) {
464 		tp->snd_ssthresh = val;
465 		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
466 			tp->snd_ssthresh = tp->snd_cwnd_clamp;
467 	} else {
468 		/* ssthresh may have been reduced unnecessarily during.
469 		 * 3WHS. Restore it back to its initial default.
470 		 */
471 		tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
472 	}
473 	val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
474 	if (val && tp->reordering != val) {
475 		tcp_disable_fack(tp);
476 		tp->reordering = val;
477 	}
478 
479 	crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
480 	rcu_read_unlock();
481 reset:
482 	/* The initial RTT measurement from the SYN/SYN-ACK is not ideal
483 	 * to seed the RTO for later data packets because SYN packets are
484 	 * small. Use the per-dst cached values to seed the RTO but keep
485 	 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
486 	 * Later the RTO will be updated immediately upon obtaining the first
487 	 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
488 	 * influences the first RTO but not later RTT estimation.
489 	 *
490 	 * But if RTT is not available from the SYN (due to retransmits or
491 	 * syn cookies) or the cache, force a conservative 3secs timeout.
492 	 *
493 	 * A bit of theory. RTT is time passed after "normal" sized packet
494 	 * is sent until it is ACKed. In normal circumstances sending small
495 	 * packets force peer to delay ACKs and calculation is correct too.
496 	 * The algorithm is adaptive and, provided we follow specs, it
497 	 * NEVER underestimate RTT. BUT! If peer tries to make some clever
498 	 * tricks sort of "quick acks" for time long enough to decrease RTT
499 	 * to low value, and then abruptly stops to do it and starts to delay
500 	 * ACKs, wait for troubles.
501 	 */
502 	if (crtt > tp->srtt_us) {
503 		/* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
504 		crtt /= 8 * USEC_PER_SEC / HZ;
505 		inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
506 	} else if (tp->srtt_us == 0) {
507 		/* RFC6298: 5.7 We've failed to get a valid RTT sample from
508 		 * 3WHS. This is most likely due to retransmission,
509 		 * including spurious one. Reset the RTO back to 3secs
510 		 * from the more aggressive 1sec to avoid more spurious
511 		 * retransmission.
512 		 */
513 		tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
514 		tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
515 
516 		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
517 	}
518 	/* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
519 	 * retransmitted. In light of RFC6298 more aggressive 1sec
520 	 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
521 	 * retransmission has occurred.
522 	 */
523 	if (tp->total_retrans > 1)
524 		tp->snd_cwnd = 1;
525 	else
526 		tp->snd_cwnd = tcp_init_cwnd(tp, dst);
527 	tp->snd_cwnd_stamp = tcp_jiffies32;
528 }
529 
530 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
531 {
532 	struct tcp_metrics_block *tm;
533 	bool ret;
534 
535 	if (!dst)
536 		return false;
537 
538 	rcu_read_lock();
539 	tm = __tcp_get_metrics_req(req, dst);
540 	if (tm && tcp_metric_get(tm, TCP_METRIC_RTT))
541 		ret = true;
542 	else
543 		ret = false;
544 	rcu_read_unlock();
545 
546 	return ret;
547 }
548 
549 static DEFINE_SEQLOCK(fastopen_seqlock);
550 
551 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
552 			    struct tcp_fastopen_cookie *cookie,
553 			    int *syn_loss, unsigned long *last_syn_loss)
554 {
555 	struct tcp_metrics_block *tm;
556 
557 	rcu_read_lock();
558 	tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
559 	if (tm) {
560 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
561 		unsigned int seq;
562 
563 		do {
564 			seq = read_seqbegin(&fastopen_seqlock);
565 			if (tfom->mss)
566 				*mss = tfom->mss;
567 			*cookie = tfom->cookie;
568 			if (cookie->len <= 0 && tfom->try_exp == 1)
569 				cookie->exp = true;
570 			*syn_loss = tfom->syn_loss;
571 			*last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
572 		} while (read_seqretry(&fastopen_seqlock, seq));
573 	}
574 	rcu_read_unlock();
575 }
576 
577 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
578 			    struct tcp_fastopen_cookie *cookie, bool syn_lost,
579 			    u16 try_exp)
580 {
581 	struct dst_entry *dst = __sk_dst_get(sk);
582 	struct tcp_metrics_block *tm;
583 
584 	if (!dst)
585 		return;
586 	rcu_read_lock();
587 	tm = tcp_get_metrics(sk, dst, true);
588 	if (tm) {
589 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
590 
591 		write_seqlock_bh(&fastopen_seqlock);
592 		if (mss)
593 			tfom->mss = mss;
594 		if (cookie && cookie->len > 0)
595 			tfom->cookie = *cookie;
596 		else if (try_exp > tfom->try_exp &&
597 			 tfom->cookie.len <= 0 && !tfom->cookie.exp)
598 			tfom->try_exp = try_exp;
599 		if (syn_lost) {
600 			++tfom->syn_loss;
601 			tfom->last_syn_loss = jiffies;
602 		} else
603 			tfom->syn_loss = 0;
604 		write_sequnlock_bh(&fastopen_seqlock);
605 	}
606 	rcu_read_unlock();
607 }
608 
609 static struct genl_family tcp_metrics_nl_family;
610 
611 static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
612 	[TCP_METRICS_ATTR_ADDR_IPV4]	= { .type = NLA_U32, },
613 	[TCP_METRICS_ATTR_ADDR_IPV6]	= { .type = NLA_BINARY,
614 					    .len = sizeof(struct in6_addr), },
615 	/* Following attributes are not received for GET/DEL,
616 	 * we keep them for reference
617 	 */
618 #if 0
619 	[TCP_METRICS_ATTR_AGE]		= { .type = NLA_MSECS, },
620 	[TCP_METRICS_ATTR_TW_TSVAL]	= { .type = NLA_U32, },
621 	[TCP_METRICS_ATTR_TW_TS_STAMP]	= { .type = NLA_S32, },
622 	[TCP_METRICS_ATTR_VALS]		= { .type = NLA_NESTED, },
623 	[TCP_METRICS_ATTR_FOPEN_MSS]	= { .type = NLA_U16, },
624 	[TCP_METRICS_ATTR_FOPEN_SYN_DROPS]	= { .type = NLA_U16, },
625 	[TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]	= { .type = NLA_MSECS, },
626 	[TCP_METRICS_ATTR_FOPEN_COOKIE]	= { .type = NLA_BINARY,
627 					    .len = TCP_FASTOPEN_COOKIE_MAX, },
628 #endif
629 };
630 
631 /* Add attributes, caller cancels its header on failure */
632 static int tcp_metrics_fill_info(struct sk_buff *msg,
633 				 struct tcp_metrics_block *tm)
634 {
635 	struct nlattr *nest;
636 	int i;
637 
638 	switch (tm->tcpm_daddr.family) {
639 	case AF_INET:
640 		if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
641 				    inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
642 			goto nla_put_failure;
643 		if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
644 				    inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
645 			goto nla_put_failure;
646 		break;
647 	case AF_INET6:
648 		if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
649 				     inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
650 			goto nla_put_failure;
651 		if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
652 				     inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
653 			goto nla_put_failure;
654 		break;
655 	default:
656 		return -EAFNOSUPPORT;
657 	}
658 
659 	if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
660 			  jiffies - tm->tcpm_stamp,
661 			  TCP_METRICS_ATTR_PAD) < 0)
662 		goto nla_put_failure;
663 
664 	{
665 		int n = 0;
666 
667 		nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
668 		if (!nest)
669 			goto nla_put_failure;
670 		for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
671 			u32 val = tm->tcpm_vals[i];
672 
673 			if (!val)
674 				continue;
675 			if (i == TCP_METRIC_RTT) {
676 				if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
677 						val) < 0)
678 					goto nla_put_failure;
679 				n++;
680 				val = max(val / 1000, 1U);
681 			}
682 			if (i == TCP_METRIC_RTTVAR) {
683 				if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
684 						val) < 0)
685 					goto nla_put_failure;
686 				n++;
687 				val = max(val / 1000, 1U);
688 			}
689 			if (nla_put_u32(msg, i + 1, val) < 0)
690 				goto nla_put_failure;
691 			n++;
692 		}
693 		if (n)
694 			nla_nest_end(msg, nest);
695 		else
696 			nla_nest_cancel(msg, nest);
697 	}
698 
699 	{
700 		struct tcp_fastopen_metrics tfom_copy[1], *tfom;
701 		unsigned int seq;
702 
703 		do {
704 			seq = read_seqbegin(&fastopen_seqlock);
705 			tfom_copy[0] = tm->tcpm_fastopen;
706 		} while (read_seqretry(&fastopen_seqlock, seq));
707 
708 		tfom = tfom_copy;
709 		if (tfom->mss &&
710 		    nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
711 				tfom->mss) < 0)
712 			goto nla_put_failure;
713 		if (tfom->syn_loss &&
714 		    (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
715 				tfom->syn_loss) < 0 ||
716 		     nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
717 				jiffies - tfom->last_syn_loss,
718 				TCP_METRICS_ATTR_PAD) < 0))
719 			goto nla_put_failure;
720 		if (tfom->cookie.len > 0 &&
721 		    nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
722 			    tfom->cookie.len, tfom->cookie.val) < 0)
723 			goto nla_put_failure;
724 	}
725 
726 	return 0;
727 
728 nla_put_failure:
729 	return -EMSGSIZE;
730 }
731 
732 static int tcp_metrics_dump_info(struct sk_buff *skb,
733 				 struct netlink_callback *cb,
734 				 struct tcp_metrics_block *tm)
735 {
736 	void *hdr;
737 
738 	hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
739 			  &tcp_metrics_nl_family, NLM_F_MULTI,
740 			  TCP_METRICS_CMD_GET);
741 	if (!hdr)
742 		return -EMSGSIZE;
743 
744 	if (tcp_metrics_fill_info(skb, tm) < 0)
745 		goto nla_put_failure;
746 
747 	genlmsg_end(skb, hdr);
748 	return 0;
749 
750 nla_put_failure:
751 	genlmsg_cancel(skb, hdr);
752 	return -EMSGSIZE;
753 }
754 
755 static int tcp_metrics_nl_dump(struct sk_buff *skb,
756 			       struct netlink_callback *cb)
757 {
758 	struct net *net = sock_net(skb->sk);
759 	unsigned int max_rows = 1U << tcp_metrics_hash_log;
760 	unsigned int row, s_row = cb->args[0];
761 	int s_col = cb->args[1], col = s_col;
762 
763 	for (row = s_row; row < max_rows; row++, s_col = 0) {
764 		struct tcp_metrics_block *tm;
765 		struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
766 
767 		rcu_read_lock();
768 		for (col = 0, tm = rcu_dereference(hb->chain); tm;
769 		     tm = rcu_dereference(tm->tcpm_next), col++) {
770 			if (!net_eq(tm_net(tm), net))
771 				continue;
772 			if (col < s_col)
773 				continue;
774 			if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
775 				rcu_read_unlock();
776 				goto done;
777 			}
778 		}
779 		rcu_read_unlock();
780 	}
781 
782 done:
783 	cb->args[0] = row;
784 	cb->args[1] = col;
785 	return skb->len;
786 }
787 
788 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
789 			   unsigned int *hash, int optional, int v4, int v6)
790 {
791 	struct nlattr *a;
792 
793 	a = info->attrs[v4];
794 	if (a) {
795 		inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
796 		if (hash)
797 			*hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
798 		return 0;
799 	}
800 	a = info->attrs[v6];
801 	if (a) {
802 		struct in6_addr in6;
803 
804 		if (nla_len(a) != sizeof(struct in6_addr))
805 			return -EINVAL;
806 		in6 = nla_get_in6_addr(a);
807 		inetpeer_set_addr_v6(addr, &in6);
808 		if (hash)
809 			*hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
810 		return 0;
811 	}
812 	return optional ? 1 : -EAFNOSUPPORT;
813 }
814 
815 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
816 			 unsigned int *hash, int optional)
817 {
818 	return __parse_nl_addr(info, addr, hash, optional,
819 			       TCP_METRICS_ATTR_ADDR_IPV4,
820 			       TCP_METRICS_ATTR_ADDR_IPV6);
821 }
822 
823 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
824 {
825 	return __parse_nl_addr(info, addr, NULL, 0,
826 			       TCP_METRICS_ATTR_SADDR_IPV4,
827 			       TCP_METRICS_ATTR_SADDR_IPV6);
828 }
829 
830 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
831 {
832 	struct tcp_metrics_block *tm;
833 	struct inetpeer_addr saddr, daddr;
834 	unsigned int hash;
835 	struct sk_buff *msg;
836 	struct net *net = genl_info_net(info);
837 	void *reply;
838 	int ret;
839 	bool src = true;
840 
841 	ret = parse_nl_addr(info, &daddr, &hash, 0);
842 	if (ret < 0)
843 		return ret;
844 
845 	ret = parse_nl_saddr(info, &saddr);
846 	if (ret < 0)
847 		src = false;
848 
849 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
850 	if (!msg)
851 		return -ENOMEM;
852 
853 	reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
854 				  info->genlhdr->cmd);
855 	if (!reply)
856 		goto nla_put_failure;
857 
858 	hash ^= net_hash_mix(net);
859 	hash = hash_32(hash, tcp_metrics_hash_log);
860 	ret = -ESRCH;
861 	rcu_read_lock();
862 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
863 	     tm = rcu_dereference(tm->tcpm_next)) {
864 		if (addr_same(&tm->tcpm_daddr, &daddr) &&
865 		    (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
866 		    net_eq(tm_net(tm), net)) {
867 			ret = tcp_metrics_fill_info(msg, tm);
868 			break;
869 		}
870 	}
871 	rcu_read_unlock();
872 	if (ret < 0)
873 		goto out_free;
874 
875 	genlmsg_end(msg, reply);
876 	return genlmsg_reply(msg, info);
877 
878 nla_put_failure:
879 	ret = -EMSGSIZE;
880 
881 out_free:
882 	nlmsg_free(msg);
883 	return ret;
884 }
885 
886 static void tcp_metrics_flush_all(struct net *net)
887 {
888 	unsigned int max_rows = 1U << tcp_metrics_hash_log;
889 	struct tcpm_hash_bucket *hb = tcp_metrics_hash;
890 	struct tcp_metrics_block *tm;
891 	unsigned int row;
892 
893 	for (row = 0; row < max_rows; row++, hb++) {
894 		struct tcp_metrics_block __rcu **pp;
895 		spin_lock_bh(&tcp_metrics_lock);
896 		pp = &hb->chain;
897 		for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
898 			if (net_eq(tm_net(tm), net)) {
899 				*pp = tm->tcpm_next;
900 				kfree_rcu(tm, rcu_head);
901 			} else {
902 				pp = &tm->tcpm_next;
903 			}
904 		}
905 		spin_unlock_bh(&tcp_metrics_lock);
906 	}
907 }
908 
909 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
910 {
911 	struct tcpm_hash_bucket *hb;
912 	struct tcp_metrics_block *tm;
913 	struct tcp_metrics_block __rcu **pp;
914 	struct inetpeer_addr saddr, daddr;
915 	unsigned int hash;
916 	struct net *net = genl_info_net(info);
917 	int ret;
918 	bool src = true, found = false;
919 
920 	ret = parse_nl_addr(info, &daddr, &hash, 1);
921 	if (ret < 0)
922 		return ret;
923 	if (ret > 0) {
924 		tcp_metrics_flush_all(net);
925 		return 0;
926 	}
927 	ret = parse_nl_saddr(info, &saddr);
928 	if (ret < 0)
929 		src = false;
930 
931 	hash ^= net_hash_mix(net);
932 	hash = hash_32(hash, tcp_metrics_hash_log);
933 	hb = tcp_metrics_hash + hash;
934 	pp = &hb->chain;
935 	spin_lock_bh(&tcp_metrics_lock);
936 	for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
937 		if (addr_same(&tm->tcpm_daddr, &daddr) &&
938 		    (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
939 		    net_eq(tm_net(tm), net)) {
940 			*pp = tm->tcpm_next;
941 			kfree_rcu(tm, rcu_head);
942 			found = true;
943 		} else {
944 			pp = &tm->tcpm_next;
945 		}
946 	}
947 	spin_unlock_bh(&tcp_metrics_lock);
948 	if (!found)
949 		return -ESRCH;
950 	return 0;
951 }
952 
953 static const struct genl_ops tcp_metrics_nl_ops[] = {
954 	{
955 		.cmd = TCP_METRICS_CMD_GET,
956 		.doit = tcp_metrics_nl_cmd_get,
957 		.dumpit = tcp_metrics_nl_dump,
958 		.policy = tcp_metrics_nl_policy,
959 	},
960 	{
961 		.cmd = TCP_METRICS_CMD_DEL,
962 		.doit = tcp_metrics_nl_cmd_del,
963 		.policy = tcp_metrics_nl_policy,
964 		.flags = GENL_ADMIN_PERM,
965 	},
966 };
967 
968 static struct genl_family tcp_metrics_nl_family __ro_after_init = {
969 	.hdrsize	= 0,
970 	.name		= TCP_METRICS_GENL_NAME,
971 	.version	= TCP_METRICS_GENL_VERSION,
972 	.maxattr	= TCP_METRICS_ATTR_MAX,
973 	.netnsok	= true,
974 	.module		= THIS_MODULE,
975 	.ops		= tcp_metrics_nl_ops,
976 	.n_ops		= ARRAY_SIZE(tcp_metrics_nl_ops),
977 };
978 
979 static unsigned int tcpmhash_entries;
980 static int __init set_tcpmhash_entries(char *str)
981 {
982 	ssize_t ret;
983 
984 	if (!str)
985 		return 0;
986 
987 	ret = kstrtouint(str, 0, &tcpmhash_entries);
988 	if (ret)
989 		return 0;
990 
991 	return 1;
992 }
993 __setup("tcpmhash_entries=", set_tcpmhash_entries);
994 
995 static int __net_init tcp_net_metrics_init(struct net *net)
996 {
997 	size_t size;
998 	unsigned int slots;
999 
1000 	if (!net_eq(net, &init_net))
1001 		return 0;
1002 
1003 	slots = tcpmhash_entries;
1004 	if (!slots) {
1005 		if (totalram_pages >= 128 * 1024)
1006 			slots = 16 * 1024;
1007 		else
1008 			slots = 8 * 1024;
1009 	}
1010 
1011 	tcp_metrics_hash_log = order_base_2(slots);
1012 	size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1013 
1014 	tcp_metrics_hash = kvzalloc(size, GFP_KERNEL);
1015 	if (!tcp_metrics_hash)
1016 		return -ENOMEM;
1017 
1018 	return 0;
1019 }
1020 
1021 static void __net_exit tcp_net_metrics_exit(struct net *net)
1022 {
1023 	tcp_metrics_flush_all(net);
1024 }
1025 
1026 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1027 	.init	=	tcp_net_metrics_init,
1028 	.exit	=	tcp_net_metrics_exit,
1029 };
1030 
1031 void __init tcp_metrics_init(void)
1032 {
1033 	int ret;
1034 
1035 	ret = register_pernet_subsys(&tcp_net_metrics_ops);
1036 	if (ret < 0)
1037 		panic("Could not allocate the tcp_metrics hash table\n");
1038 
1039 	ret = genl_register_family(&tcp_metrics_nl_family);
1040 	if (ret < 0)
1041 		panic("Could not register tcp_metrics generic netlink\n");
1042 }
1043