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
23int sysctl_tcp_nometrics_save __read_mostly;
24
25static 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
29struct 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
42struct 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_ts;
49	u32				tcpm_ts_stamp;
50	u32				tcpm_lock;
51	u32				tcpm_vals[TCP_METRIC_MAX_KERNEL + 1];
52	struct tcp_fastopen_metrics	tcpm_fastopen;
53
54	struct rcu_head			rcu_head;
55};
56
57static inline struct net *tm_net(struct tcp_metrics_block *tm)
58{
59	return read_pnet(&tm->tcpm_net);
60}
61
62static bool tcp_metric_locked(struct tcp_metrics_block *tm,
63			      enum tcp_metric_index idx)
64{
65	return tm->tcpm_lock & (1 << idx);
66}
67
68static u32 tcp_metric_get(struct tcp_metrics_block *tm,
69			  enum tcp_metric_index idx)
70{
71	return tm->tcpm_vals[idx];
72}
73
74static void tcp_metric_set(struct tcp_metrics_block *tm,
75			   enum tcp_metric_index idx,
76			   u32 val)
77{
78	tm->tcpm_vals[idx] = val;
79}
80
81static bool addr_same(const struct inetpeer_addr *a,
82		      const struct inetpeer_addr *b)
83{
84	if (a->family != b->family)
85		return false;
86	if (a->family == AF_INET)
87		return a->addr.a4 == b->addr.a4;
88	return ipv6_addr_equal(&a->addr.in6, &b->addr.in6);
89}
90
91struct tcpm_hash_bucket {
92	struct tcp_metrics_block __rcu	*chain;
93};
94
95static struct tcpm_hash_bucket	*tcp_metrics_hash __read_mostly;
96static unsigned int		tcp_metrics_hash_log __read_mostly;
97
98static DEFINE_SPINLOCK(tcp_metrics_lock);
99
100static void tcpm_suck_dst(struct tcp_metrics_block *tm,
101			  const struct dst_entry *dst,
102			  bool fastopen_clear)
103{
104	u32 msval;
105	u32 val;
106
107	tm->tcpm_stamp = jiffies;
108
109	val = 0;
110	if (dst_metric_locked(dst, RTAX_RTT))
111		val |= 1 << TCP_METRIC_RTT;
112	if (dst_metric_locked(dst, RTAX_RTTVAR))
113		val |= 1 << TCP_METRIC_RTTVAR;
114	if (dst_metric_locked(dst, RTAX_SSTHRESH))
115		val |= 1 << TCP_METRIC_SSTHRESH;
116	if (dst_metric_locked(dst, RTAX_CWND))
117		val |= 1 << TCP_METRIC_CWND;
118	if (dst_metric_locked(dst, RTAX_REORDERING))
119		val |= 1 << TCP_METRIC_REORDERING;
120	tm->tcpm_lock = val;
121
122	msval = dst_metric_raw(dst, RTAX_RTT);
123	tm->tcpm_vals[TCP_METRIC_RTT] = msval * USEC_PER_MSEC;
124
125	msval = dst_metric_raw(dst, RTAX_RTTVAR);
126	tm->tcpm_vals[TCP_METRIC_RTTVAR] = msval * USEC_PER_MSEC;
127	tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
128	tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
129	tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
130	tm->tcpm_ts = 0;
131	tm->tcpm_ts_stamp = 0;
132	if (fastopen_clear) {
133		tm->tcpm_fastopen.mss = 0;
134		tm->tcpm_fastopen.syn_loss = 0;
135		tm->tcpm_fastopen.try_exp = 0;
136		tm->tcpm_fastopen.cookie.exp = false;
137		tm->tcpm_fastopen.cookie.len = 0;
138	}
139}
140
141#define TCP_METRICS_TIMEOUT		(60 * 60 * HZ)
142
143static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
144{
145	if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
146		tcpm_suck_dst(tm, dst, false);
147}
148
149#define TCP_METRICS_RECLAIM_DEPTH	5
150#define TCP_METRICS_RECLAIM_PTR		(struct tcp_metrics_block *) 0x1UL
151
152#define deref_locked(p)	\
153	rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
154
155static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
156					  struct inetpeer_addr *saddr,
157					  struct inetpeer_addr *daddr,
158					  unsigned int hash)
159{
160	struct tcp_metrics_block *tm;
161	struct net *net;
162	bool reclaim = false;
163
164	spin_lock_bh(&tcp_metrics_lock);
165	net = dev_net(dst->dev);
166
167	/* While waiting for the spin-lock the cache might have been populated
168	 * with this entry and so we have to check again.
169	 */
170	tm = __tcp_get_metrics(saddr, daddr, net, hash);
171	if (tm == TCP_METRICS_RECLAIM_PTR) {
172		reclaim = true;
173		tm = NULL;
174	}
175	if (tm) {
176		tcpm_check_stamp(tm, dst);
177		goto out_unlock;
178	}
179
180	if (unlikely(reclaim)) {
181		struct tcp_metrics_block *oldest;
182
183		oldest = deref_locked(tcp_metrics_hash[hash].chain);
184		for (tm = deref_locked(oldest->tcpm_next); tm;
185		     tm = deref_locked(tm->tcpm_next)) {
186			if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
187				oldest = tm;
188		}
189		tm = oldest;
190	} else {
191		tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
192		if (!tm)
193			goto out_unlock;
194	}
195	write_pnet(&tm->tcpm_net, net);
196	tm->tcpm_saddr = *saddr;
197	tm->tcpm_daddr = *daddr;
198
199	tcpm_suck_dst(tm, dst, true);
200
201	if (likely(!reclaim)) {
202		tm->tcpm_next = tcp_metrics_hash[hash].chain;
203		rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
204	}
205
206out_unlock:
207	spin_unlock_bh(&tcp_metrics_lock);
208	return tm;
209}
210
211static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
212{
213	if (tm)
214		return tm;
215	if (depth > TCP_METRICS_RECLAIM_DEPTH)
216		return TCP_METRICS_RECLAIM_PTR;
217	return NULL;
218}
219
220static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
221						   const struct inetpeer_addr *daddr,
222						   struct net *net, unsigned int hash)
223{
224	struct tcp_metrics_block *tm;
225	int depth = 0;
226
227	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
228	     tm = rcu_dereference(tm->tcpm_next)) {
229		if (addr_same(&tm->tcpm_saddr, saddr) &&
230		    addr_same(&tm->tcpm_daddr, daddr) &&
231		    net_eq(tm_net(tm), net))
232			break;
233		depth++;
234	}
235	return tcp_get_encode(tm, depth);
236}
237
238static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
239						       struct dst_entry *dst)
240{
241	struct tcp_metrics_block *tm;
242	struct inetpeer_addr saddr, daddr;
243	unsigned int hash;
244	struct net *net;
245
246	saddr.family = req->rsk_ops->family;
247	daddr.family = req->rsk_ops->family;
248	switch (daddr.family) {
249	case AF_INET:
250		saddr.addr.a4 = inet_rsk(req)->ir_loc_addr;
251		daddr.addr.a4 = inet_rsk(req)->ir_rmt_addr;
252		hash = (__force unsigned int) daddr.addr.a4;
253		break;
254#if IS_ENABLED(CONFIG_IPV6)
255	case AF_INET6:
256		saddr.addr.in6 = inet_rsk(req)->ir_v6_loc_addr;
257		daddr.addr.in6 = inet_rsk(req)->ir_v6_rmt_addr;
258		hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
259		break;
260#endif
261	default:
262		return NULL;
263	}
264
265	net = dev_net(dst->dev);
266	hash ^= net_hash_mix(net);
267	hash = hash_32(hash, tcp_metrics_hash_log);
268
269	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
270	     tm = rcu_dereference(tm->tcpm_next)) {
271		if (addr_same(&tm->tcpm_saddr, &saddr) &&
272		    addr_same(&tm->tcpm_daddr, &daddr) &&
273		    net_eq(tm_net(tm), net))
274			break;
275	}
276	tcpm_check_stamp(tm, dst);
277	return tm;
278}
279
280static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
281{
282	struct tcp_metrics_block *tm;
283	struct inetpeer_addr saddr, daddr;
284	unsigned int hash;
285	struct net *net;
286
287	if (tw->tw_family == AF_INET) {
288		saddr.family = AF_INET;
289		saddr.addr.a4 = tw->tw_rcv_saddr;
290		daddr.family = AF_INET;
291		daddr.addr.a4 = tw->tw_daddr;
292		hash = (__force unsigned int) daddr.addr.a4;
293	}
294#if IS_ENABLED(CONFIG_IPV6)
295	else if (tw->tw_family == AF_INET6) {
296		if (ipv6_addr_v4mapped(&tw->tw_v6_daddr)) {
297			saddr.family = AF_INET;
298			saddr.addr.a4 = tw->tw_rcv_saddr;
299			daddr.family = AF_INET;
300			daddr.addr.a4 = tw->tw_daddr;
301			hash = (__force unsigned int) daddr.addr.a4;
302		} else {
303			saddr.family = AF_INET6;
304			saddr.addr.in6 = tw->tw_v6_rcv_saddr;
305			daddr.family = AF_INET6;
306			daddr.addr.in6 = tw->tw_v6_daddr;
307			hash = ipv6_addr_hash(&tw->tw_v6_daddr);
308		}
309	}
310#endif
311	else
312		return NULL;
313
314	net = twsk_net(tw);
315	hash ^= net_hash_mix(net);
316	hash = hash_32(hash, tcp_metrics_hash_log);
317
318	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
319	     tm = rcu_dereference(tm->tcpm_next)) {
320		if (addr_same(&tm->tcpm_saddr, &saddr) &&
321		    addr_same(&tm->tcpm_daddr, &daddr) &&
322		    net_eq(tm_net(tm), net))
323			break;
324	}
325	return tm;
326}
327
328static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
329						 struct dst_entry *dst,
330						 bool create)
331{
332	struct tcp_metrics_block *tm;
333	struct inetpeer_addr saddr, daddr;
334	unsigned int hash;
335	struct net *net;
336
337	if (sk->sk_family == AF_INET) {
338		saddr.family = AF_INET;
339		saddr.addr.a4 = inet_sk(sk)->inet_saddr;
340		daddr.family = AF_INET;
341		daddr.addr.a4 = inet_sk(sk)->inet_daddr;
342		hash = (__force unsigned int) daddr.addr.a4;
343	}
344#if IS_ENABLED(CONFIG_IPV6)
345	else if (sk->sk_family == AF_INET6) {
346		if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
347			saddr.family = AF_INET;
348			saddr.addr.a4 = inet_sk(sk)->inet_saddr;
349			daddr.family = AF_INET;
350			daddr.addr.a4 = inet_sk(sk)->inet_daddr;
351			hash = (__force unsigned int) daddr.addr.a4;
352		} else {
353			saddr.family = AF_INET6;
354			saddr.addr.in6 = sk->sk_v6_rcv_saddr;
355			daddr.family = AF_INET6;
356			daddr.addr.in6 = sk->sk_v6_daddr;
357			hash = ipv6_addr_hash(&sk->sk_v6_daddr);
358		}
359	}
360#endif
361	else
362		return NULL;
363
364	net = dev_net(dst->dev);
365	hash ^= net_hash_mix(net);
366	hash = hash_32(hash, tcp_metrics_hash_log);
367
368	tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
369	if (tm == TCP_METRICS_RECLAIM_PTR)
370		tm = NULL;
371	if (!tm && create)
372		tm = tcpm_new(dst, &saddr, &daddr, hash);
373	else
374		tcpm_check_stamp(tm, dst);
375
376	return tm;
377}
378
379/* Save metrics learned by this TCP session.  This function is called
380 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
381 * or goes from LAST-ACK to CLOSE.
382 */
383void tcp_update_metrics(struct sock *sk)
384{
385	const struct inet_connection_sock *icsk = inet_csk(sk);
386	struct dst_entry *dst = __sk_dst_get(sk);
387	struct tcp_sock *tp = tcp_sk(sk);
388	struct tcp_metrics_block *tm;
389	unsigned long rtt;
390	u32 val;
391	int m;
392
393	if (sysctl_tcp_nometrics_save || !dst)
394		return;
395
396	if (dst->flags & DST_HOST)
397		dst_confirm(dst);
398
399	rcu_read_lock();
400	if (icsk->icsk_backoff || !tp->srtt_us) {
401		/* This session failed to estimate rtt. Why?
402		 * Probably, no packets returned in time.  Reset our
403		 * results.
404		 */
405		tm = tcp_get_metrics(sk, dst, false);
406		if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
407			tcp_metric_set(tm, TCP_METRIC_RTT, 0);
408		goto out_unlock;
409	} else
410		tm = tcp_get_metrics(sk, dst, true);
411
412	if (!tm)
413		goto out_unlock;
414
415	rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
416	m = rtt - tp->srtt_us;
417
418	/* If newly calculated rtt larger than stored one, store new
419	 * one. Otherwise, use EWMA. Remember, rtt overestimation is
420	 * always better than underestimation.
421	 */
422	if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
423		if (m <= 0)
424			rtt = tp->srtt_us;
425		else
426			rtt -= (m >> 3);
427		tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
428	}
429
430	if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
431		unsigned long var;
432
433		if (m < 0)
434			m = -m;
435
436		/* Scale deviation to rttvar fixed point */
437		m >>= 1;
438		if (m < tp->mdev_us)
439			m = tp->mdev_us;
440
441		var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
442		if (m >= var)
443			var = m;
444		else
445			var -= (var - m) >> 2;
446
447		tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
448	}
449
450	if (tcp_in_initial_slowstart(tp)) {
451		/* Slow start still did not finish. */
452		if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
453			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
454			if (val && (tp->snd_cwnd >> 1) > val)
455				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
456					       tp->snd_cwnd >> 1);
457		}
458		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
459			val = tcp_metric_get(tm, TCP_METRIC_CWND);
460			if (tp->snd_cwnd > val)
461				tcp_metric_set(tm, TCP_METRIC_CWND,
462					       tp->snd_cwnd);
463		}
464	} else if (tp->snd_cwnd > tp->snd_ssthresh &&
465		   icsk->icsk_ca_state == TCP_CA_Open) {
466		/* Cong. avoidance phase, cwnd is reliable. */
467		if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
468			tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
469				       max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
470		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
471			val = tcp_metric_get(tm, TCP_METRIC_CWND);
472			tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
473		}
474	} else {
475		/* Else slow start did not finish, cwnd is non-sense,
476		 * ssthresh may be also invalid.
477		 */
478		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
479			val = tcp_metric_get(tm, TCP_METRIC_CWND);
480			tcp_metric_set(tm, TCP_METRIC_CWND,
481				       (val + tp->snd_ssthresh) >> 1);
482		}
483		if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
484			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
485			if (val && tp->snd_ssthresh > val)
486				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
487					       tp->snd_ssthresh);
488		}
489		if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
490			val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
491			if (val < tp->reordering &&
492			    tp->reordering != sysctl_tcp_reordering)
493				tcp_metric_set(tm, TCP_METRIC_REORDERING,
494					       tp->reordering);
495		}
496	}
497	tm->tcpm_stamp = jiffies;
498out_unlock:
499	rcu_read_unlock();
500}
501
502/* Initialize metrics on socket. */
503
504void tcp_init_metrics(struct sock *sk)
505{
506	struct dst_entry *dst = __sk_dst_get(sk);
507	struct tcp_sock *tp = tcp_sk(sk);
508	struct tcp_metrics_block *tm;
509	u32 val, crtt = 0; /* cached RTT scaled by 8 */
510
511	if (!dst)
512		goto reset;
513
514	dst_confirm(dst);
515
516	rcu_read_lock();
517	tm = tcp_get_metrics(sk, dst, true);
518	if (!tm) {
519		rcu_read_unlock();
520		goto reset;
521	}
522
523	if (tcp_metric_locked(tm, TCP_METRIC_CWND))
524		tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
525
526	val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
527	if (val) {
528		tp->snd_ssthresh = val;
529		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
530			tp->snd_ssthresh = tp->snd_cwnd_clamp;
531	} else {
532		/* ssthresh may have been reduced unnecessarily during.
533		 * 3WHS. Restore it back to its initial default.
534		 */
535		tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
536	}
537	val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
538	if (val && tp->reordering != val) {
539		tcp_disable_fack(tp);
540		tcp_disable_early_retrans(tp);
541		tp->reordering = val;
542	}
543
544	crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
545	rcu_read_unlock();
546reset:
547	/* The initial RTT measurement from the SYN/SYN-ACK is not ideal
548	 * to seed the RTO for later data packets because SYN packets are
549	 * small. Use the per-dst cached values to seed the RTO but keep
550	 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
551	 * Later the RTO will be updated immediately upon obtaining the first
552	 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
553	 * influences the first RTO but not later RTT estimation.
554	 *
555	 * But if RTT is not available from the SYN (due to retransmits or
556	 * syn cookies) or the cache, force a conservative 3secs timeout.
557	 *
558	 * A bit of theory. RTT is time passed after "normal" sized packet
559	 * is sent until it is ACKed. In normal circumstances sending small
560	 * packets force peer to delay ACKs and calculation is correct too.
561	 * The algorithm is adaptive and, provided we follow specs, it
562	 * NEVER underestimate RTT. BUT! If peer tries to make some clever
563	 * tricks sort of "quick acks" for time long enough to decrease RTT
564	 * to low value, and then abruptly stops to do it and starts to delay
565	 * ACKs, wait for troubles.
566	 */
567	if (crtt > tp->srtt_us) {
568		/* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
569		crtt /= 8 * USEC_PER_MSEC;
570		inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
571	} else if (tp->srtt_us == 0) {
572		/* RFC6298: 5.7 We've failed to get a valid RTT sample from
573		 * 3WHS. This is most likely due to retransmission,
574		 * including spurious one. Reset the RTO back to 3secs
575		 * from the more aggressive 1sec to avoid more spurious
576		 * retransmission.
577		 */
578		tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
579		tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
580
581		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
582	}
583	/* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
584	 * retransmitted. In light of RFC6298 more aggressive 1sec
585	 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
586	 * retransmission has occurred.
587	 */
588	if (tp->total_retrans > 1)
589		tp->snd_cwnd = 1;
590	else
591		tp->snd_cwnd = tcp_init_cwnd(tp, dst);
592	tp->snd_cwnd_stamp = tcp_time_stamp;
593}
594
595bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst,
596			bool paws_check, bool timestamps)
597{
598	struct tcp_metrics_block *tm;
599	bool ret;
600
601	if (!dst)
602		return false;
603
604	rcu_read_lock();
605	tm = __tcp_get_metrics_req(req, dst);
606	if (paws_check) {
607		if (tm &&
608		    (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
609		    ((s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW ||
610		     !timestamps))
611			ret = false;
612		else
613			ret = true;
614	} else {
615		if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
616			ret = true;
617		else
618			ret = false;
619	}
620	rcu_read_unlock();
621
622	return ret;
623}
624EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
625
626void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
627{
628	struct tcp_metrics_block *tm;
629
630	rcu_read_lock();
631	tm = tcp_get_metrics(sk, dst, true);
632	if (tm) {
633		struct tcp_sock *tp = tcp_sk(sk);
634
635		if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
636			tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
637			tp->rx_opt.ts_recent = tm->tcpm_ts;
638		}
639	}
640	rcu_read_unlock();
641}
642EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
643
644/* VJ's idea. Save last timestamp seen from this destination and hold
645 * it at least for normal timewait interval to use for duplicate
646 * segment detection in subsequent connections, before they enter
647 * synchronized state.
648 */
649bool tcp_remember_stamp(struct sock *sk)
650{
651	struct dst_entry *dst = __sk_dst_get(sk);
652	bool ret = false;
653
654	if (dst) {
655		struct tcp_metrics_block *tm;
656
657		rcu_read_lock();
658		tm = tcp_get_metrics(sk, dst, true);
659		if (tm) {
660			struct tcp_sock *tp = tcp_sk(sk);
661
662			if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
663			    ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
664			     tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
665				tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
666				tm->tcpm_ts = tp->rx_opt.ts_recent;
667			}
668			ret = true;
669		}
670		rcu_read_unlock();
671	}
672	return ret;
673}
674
675bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
676{
677	struct tcp_metrics_block *tm;
678	bool ret = false;
679
680	rcu_read_lock();
681	tm = __tcp_get_metrics_tw(tw);
682	if (tm) {
683		const struct tcp_timewait_sock *tcptw;
684		struct sock *sk = (struct sock *) tw;
685
686		tcptw = tcp_twsk(sk);
687		if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
688		    ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
689		     tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
690			tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
691			tm->tcpm_ts	   = tcptw->tw_ts_recent;
692		}
693		ret = true;
694	}
695	rcu_read_unlock();
696
697	return ret;
698}
699
700static DEFINE_SEQLOCK(fastopen_seqlock);
701
702void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
703			    struct tcp_fastopen_cookie *cookie,
704			    int *syn_loss, unsigned long *last_syn_loss)
705{
706	struct tcp_metrics_block *tm;
707
708	rcu_read_lock();
709	tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
710	if (tm) {
711		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
712		unsigned int seq;
713
714		do {
715			seq = read_seqbegin(&fastopen_seqlock);
716			if (tfom->mss)
717				*mss = tfom->mss;
718			*cookie = tfom->cookie;
719			if (cookie->len <= 0 && tfom->try_exp == 1)
720				cookie->exp = true;
721			*syn_loss = tfom->syn_loss;
722			*last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
723		} while (read_seqretry(&fastopen_seqlock, seq));
724	}
725	rcu_read_unlock();
726}
727
728void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
729			    struct tcp_fastopen_cookie *cookie, bool syn_lost,
730			    u16 try_exp)
731{
732	struct dst_entry *dst = __sk_dst_get(sk);
733	struct tcp_metrics_block *tm;
734
735	if (!dst)
736		return;
737	rcu_read_lock();
738	tm = tcp_get_metrics(sk, dst, true);
739	if (tm) {
740		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
741
742		write_seqlock_bh(&fastopen_seqlock);
743		if (mss)
744			tfom->mss = mss;
745		if (cookie && cookie->len > 0)
746			tfom->cookie = *cookie;
747		else if (try_exp > tfom->try_exp &&
748			 tfom->cookie.len <= 0 && !tfom->cookie.exp)
749			tfom->try_exp = try_exp;
750		if (syn_lost) {
751			++tfom->syn_loss;
752			tfom->last_syn_loss = jiffies;
753		} else
754			tfom->syn_loss = 0;
755		write_sequnlock_bh(&fastopen_seqlock);
756	}
757	rcu_read_unlock();
758}
759
760static struct genl_family tcp_metrics_nl_family = {
761	.id		= GENL_ID_GENERATE,
762	.hdrsize	= 0,
763	.name		= TCP_METRICS_GENL_NAME,
764	.version	= TCP_METRICS_GENL_VERSION,
765	.maxattr	= TCP_METRICS_ATTR_MAX,
766	.netnsok	= true,
767};
768
769static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
770	[TCP_METRICS_ATTR_ADDR_IPV4]	= { .type = NLA_U32, },
771	[TCP_METRICS_ATTR_ADDR_IPV6]	= { .type = NLA_BINARY,
772					    .len = sizeof(struct in6_addr), },
773	/* Following attributes are not received for GET/DEL,
774	 * we keep them for reference
775	 */
776#if 0
777	[TCP_METRICS_ATTR_AGE]		= { .type = NLA_MSECS, },
778	[TCP_METRICS_ATTR_TW_TSVAL]	= { .type = NLA_U32, },
779	[TCP_METRICS_ATTR_TW_TS_STAMP]	= { .type = NLA_S32, },
780	[TCP_METRICS_ATTR_VALS]		= { .type = NLA_NESTED, },
781	[TCP_METRICS_ATTR_FOPEN_MSS]	= { .type = NLA_U16, },
782	[TCP_METRICS_ATTR_FOPEN_SYN_DROPS]	= { .type = NLA_U16, },
783	[TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]	= { .type = NLA_MSECS, },
784	[TCP_METRICS_ATTR_FOPEN_COOKIE]	= { .type = NLA_BINARY,
785					    .len = TCP_FASTOPEN_COOKIE_MAX, },
786#endif
787};
788
789/* Add attributes, caller cancels its header on failure */
790static int tcp_metrics_fill_info(struct sk_buff *msg,
791				 struct tcp_metrics_block *tm)
792{
793	struct nlattr *nest;
794	int i;
795
796	switch (tm->tcpm_daddr.family) {
797	case AF_INET:
798		if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
799				    tm->tcpm_daddr.addr.a4) < 0)
800			goto nla_put_failure;
801		if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
802				    tm->tcpm_saddr.addr.a4) < 0)
803			goto nla_put_failure;
804		break;
805	case AF_INET6:
806		if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
807				     &tm->tcpm_daddr.addr.in6) < 0)
808			goto nla_put_failure;
809		if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
810				     &tm->tcpm_saddr.addr.in6) < 0)
811			goto nla_put_failure;
812		break;
813	default:
814		return -EAFNOSUPPORT;
815	}
816
817	if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
818			  jiffies - tm->tcpm_stamp) < 0)
819		goto nla_put_failure;
820	if (tm->tcpm_ts_stamp) {
821		if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
822				(s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
823			goto nla_put_failure;
824		if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
825				tm->tcpm_ts) < 0)
826			goto nla_put_failure;
827	}
828
829	{
830		int n = 0;
831
832		nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
833		if (!nest)
834			goto nla_put_failure;
835		for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
836			u32 val = tm->tcpm_vals[i];
837
838			if (!val)
839				continue;
840			if (i == TCP_METRIC_RTT) {
841				if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
842						val) < 0)
843					goto nla_put_failure;
844				n++;
845				val = max(val / 1000, 1U);
846			}
847			if (i == TCP_METRIC_RTTVAR) {
848				if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
849						val) < 0)
850					goto nla_put_failure;
851				n++;
852				val = max(val / 1000, 1U);
853			}
854			if (nla_put_u32(msg, i + 1, val) < 0)
855				goto nla_put_failure;
856			n++;
857		}
858		if (n)
859			nla_nest_end(msg, nest);
860		else
861			nla_nest_cancel(msg, nest);
862	}
863
864	{
865		struct tcp_fastopen_metrics tfom_copy[1], *tfom;
866		unsigned int seq;
867
868		do {
869			seq = read_seqbegin(&fastopen_seqlock);
870			tfom_copy[0] = tm->tcpm_fastopen;
871		} while (read_seqretry(&fastopen_seqlock, seq));
872
873		tfom = tfom_copy;
874		if (tfom->mss &&
875		    nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
876				tfom->mss) < 0)
877			goto nla_put_failure;
878		if (tfom->syn_loss &&
879		    (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
880				tfom->syn_loss) < 0 ||
881		     nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
882				jiffies - tfom->last_syn_loss) < 0))
883			goto nla_put_failure;
884		if (tfom->cookie.len > 0 &&
885		    nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
886			    tfom->cookie.len, tfom->cookie.val) < 0)
887			goto nla_put_failure;
888	}
889
890	return 0;
891
892nla_put_failure:
893	return -EMSGSIZE;
894}
895
896static int tcp_metrics_dump_info(struct sk_buff *skb,
897				 struct netlink_callback *cb,
898				 struct tcp_metrics_block *tm)
899{
900	void *hdr;
901
902	hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
903			  &tcp_metrics_nl_family, NLM_F_MULTI,
904			  TCP_METRICS_CMD_GET);
905	if (!hdr)
906		return -EMSGSIZE;
907
908	if (tcp_metrics_fill_info(skb, tm) < 0)
909		goto nla_put_failure;
910
911	genlmsg_end(skb, hdr);
912	return 0;
913
914nla_put_failure:
915	genlmsg_cancel(skb, hdr);
916	return -EMSGSIZE;
917}
918
919static int tcp_metrics_nl_dump(struct sk_buff *skb,
920			       struct netlink_callback *cb)
921{
922	struct net *net = sock_net(skb->sk);
923	unsigned int max_rows = 1U << tcp_metrics_hash_log;
924	unsigned int row, s_row = cb->args[0];
925	int s_col = cb->args[1], col = s_col;
926
927	for (row = s_row; row < max_rows; row++, s_col = 0) {
928		struct tcp_metrics_block *tm;
929		struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
930
931		rcu_read_lock();
932		for (col = 0, tm = rcu_dereference(hb->chain); tm;
933		     tm = rcu_dereference(tm->tcpm_next), col++) {
934			if (!net_eq(tm_net(tm), net))
935				continue;
936			if (col < s_col)
937				continue;
938			if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
939				rcu_read_unlock();
940				goto done;
941			}
942		}
943		rcu_read_unlock();
944	}
945
946done:
947	cb->args[0] = row;
948	cb->args[1] = col;
949	return skb->len;
950}
951
952static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
953			   unsigned int *hash, int optional, int v4, int v6)
954{
955	struct nlattr *a;
956
957	a = info->attrs[v4];
958	if (a) {
959		addr->family = AF_INET;
960		addr->addr.a4 = nla_get_in_addr(a);
961		if (hash)
962			*hash = (__force unsigned int) addr->addr.a4;
963		return 0;
964	}
965	a = info->attrs[v6];
966	if (a) {
967		if (nla_len(a) != sizeof(struct in6_addr))
968			return -EINVAL;
969		addr->family = AF_INET6;
970		addr->addr.in6 = nla_get_in6_addr(a);
971		if (hash)
972			*hash = ipv6_addr_hash(&addr->addr.in6);
973		return 0;
974	}
975	return optional ? 1 : -EAFNOSUPPORT;
976}
977
978static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
979			 unsigned int *hash, int optional)
980{
981	return __parse_nl_addr(info, addr, hash, optional,
982			       TCP_METRICS_ATTR_ADDR_IPV4,
983			       TCP_METRICS_ATTR_ADDR_IPV6);
984}
985
986static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
987{
988	return __parse_nl_addr(info, addr, NULL, 0,
989			       TCP_METRICS_ATTR_SADDR_IPV4,
990			       TCP_METRICS_ATTR_SADDR_IPV6);
991}
992
993static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
994{
995	struct tcp_metrics_block *tm;
996	struct inetpeer_addr saddr, daddr;
997	unsigned int hash;
998	struct sk_buff *msg;
999	struct net *net = genl_info_net(info);
1000	void *reply;
1001	int ret;
1002	bool src = true;
1003
1004	ret = parse_nl_addr(info, &daddr, &hash, 0);
1005	if (ret < 0)
1006		return ret;
1007
1008	ret = parse_nl_saddr(info, &saddr);
1009	if (ret < 0)
1010		src = false;
1011
1012	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1013	if (!msg)
1014		return -ENOMEM;
1015
1016	reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
1017				  info->genlhdr->cmd);
1018	if (!reply)
1019		goto nla_put_failure;
1020
1021	hash ^= net_hash_mix(net);
1022	hash = hash_32(hash, tcp_metrics_hash_log);
1023	ret = -ESRCH;
1024	rcu_read_lock();
1025	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
1026	     tm = rcu_dereference(tm->tcpm_next)) {
1027		if (addr_same(&tm->tcpm_daddr, &daddr) &&
1028		    (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1029		    net_eq(tm_net(tm), net)) {
1030			ret = tcp_metrics_fill_info(msg, tm);
1031			break;
1032		}
1033	}
1034	rcu_read_unlock();
1035	if (ret < 0)
1036		goto out_free;
1037
1038	genlmsg_end(msg, reply);
1039	return genlmsg_reply(msg, info);
1040
1041nla_put_failure:
1042	ret = -EMSGSIZE;
1043
1044out_free:
1045	nlmsg_free(msg);
1046	return ret;
1047}
1048
1049static void tcp_metrics_flush_all(struct net *net)
1050{
1051	unsigned int max_rows = 1U << tcp_metrics_hash_log;
1052	struct tcpm_hash_bucket *hb = tcp_metrics_hash;
1053	struct tcp_metrics_block *tm;
1054	unsigned int row;
1055
1056	for (row = 0; row < max_rows; row++, hb++) {
1057		struct tcp_metrics_block __rcu **pp;
1058		spin_lock_bh(&tcp_metrics_lock);
1059		pp = &hb->chain;
1060		for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1061			if (net_eq(tm_net(tm), net)) {
1062				*pp = tm->tcpm_next;
1063				kfree_rcu(tm, rcu_head);
1064			} else {
1065				pp = &tm->tcpm_next;
1066			}
1067		}
1068		spin_unlock_bh(&tcp_metrics_lock);
1069	}
1070}
1071
1072static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
1073{
1074	struct tcpm_hash_bucket *hb;
1075	struct tcp_metrics_block *tm;
1076	struct tcp_metrics_block __rcu **pp;
1077	struct inetpeer_addr saddr, daddr;
1078	unsigned int hash;
1079	struct net *net = genl_info_net(info);
1080	int ret;
1081	bool src = true, found = false;
1082
1083	ret = parse_nl_addr(info, &daddr, &hash, 1);
1084	if (ret < 0)
1085		return ret;
1086	if (ret > 0) {
1087		tcp_metrics_flush_all(net);
1088		return 0;
1089	}
1090	ret = parse_nl_saddr(info, &saddr);
1091	if (ret < 0)
1092		src = false;
1093
1094	hash ^= net_hash_mix(net);
1095	hash = hash_32(hash, tcp_metrics_hash_log);
1096	hb = tcp_metrics_hash + hash;
1097	pp = &hb->chain;
1098	spin_lock_bh(&tcp_metrics_lock);
1099	for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1100		if (addr_same(&tm->tcpm_daddr, &daddr) &&
1101		    (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1102		    net_eq(tm_net(tm), net)) {
1103			*pp = tm->tcpm_next;
1104			kfree_rcu(tm, rcu_head);
1105			found = true;
1106		} else {
1107			pp = &tm->tcpm_next;
1108		}
1109	}
1110	spin_unlock_bh(&tcp_metrics_lock);
1111	if (!found)
1112		return -ESRCH;
1113	return 0;
1114}
1115
1116static const struct genl_ops tcp_metrics_nl_ops[] = {
1117	{
1118		.cmd = TCP_METRICS_CMD_GET,
1119		.doit = tcp_metrics_nl_cmd_get,
1120		.dumpit = tcp_metrics_nl_dump,
1121		.policy = tcp_metrics_nl_policy,
1122	},
1123	{
1124		.cmd = TCP_METRICS_CMD_DEL,
1125		.doit = tcp_metrics_nl_cmd_del,
1126		.policy = tcp_metrics_nl_policy,
1127		.flags = GENL_ADMIN_PERM,
1128	},
1129};
1130
1131static unsigned int tcpmhash_entries;
1132static int __init set_tcpmhash_entries(char *str)
1133{
1134	ssize_t ret;
1135
1136	if (!str)
1137		return 0;
1138
1139	ret = kstrtouint(str, 0, &tcpmhash_entries);
1140	if (ret)
1141		return 0;
1142
1143	return 1;
1144}
1145__setup("tcpmhash_entries=", set_tcpmhash_entries);
1146
1147static int __net_init tcp_net_metrics_init(struct net *net)
1148{
1149	size_t size;
1150	unsigned int slots;
1151
1152	if (!net_eq(net, &init_net))
1153		return 0;
1154
1155	slots = tcpmhash_entries;
1156	if (!slots) {
1157		if (totalram_pages >= 128 * 1024)
1158			slots = 16 * 1024;
1159		else
1160			slots = 8 * 1024;
1161	}
1162
1163	tcp_metrics_hash_log = order_base_2(slots);
1164	size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1165
1166	tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1167	if (!tcp_metrics_hash)
1168		tcp_metrics_hash = vzalloc(size);
1169
1170	if (!tcp_metrics_hash)
1171		return -ENOMEM;
1172
1173	return 0;
1174}
1175
1176static void __net_exit tcp_net_metrics_exit(struct net *net)
1177{
1178	tcp_metrics_flush_all(net);
1179}
1180
1181static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1182	.init	=	tcp_net_metrics_init,
1183	.exit	=	tcp_net_metrics_exit,
1184};
1185
1186void __init tcp_metrics_init(void)
1187{
1188	int ret;
1189
1190	ret = register_pernet_subsys(&tcp_net_metrics_ops);
1191	if (ret < 0)
1192		panic("Could not allocate the tcp_metrics hash table\n");
1193
1194	ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1195					    tcp_metrics_nl_ops);
1196	if (ret < 0)
1197		panic("Could not register tcp_metrics generic netlink\n");
1198}
1199