tcp_metrics.c revision efeaa5550e4bfd335396415958fe3615530e5d5c
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
25struct tcp_fastopen_metrics {
26	u16	mss;
27	u16	syn_loss:10;		/* Recurring Fast Open SYN losses */
28	unsigned long	last_syn_loss;	/* Last Fast Open SYN loss */
29	struct	tcp_fastopen_cookie	cookie;
30};
31
32struct tcp_metrics_block {
33	struct tcp_metrics_block __rcu	*tcpm_next;
34	struct inetpeer_addr		tcpm_addr;
35	unsigned long			tcpm_stamp;
36	u32				tcpm_ts;
37	u32				tcpm_ts_stamp;
38	u32				tcpm_lock;
39	u32				tcpm_vals[TCP_METRIC_MAX + 1];
40	struct tcp_fastopen_metrics	tcpm_fastopen;
41
42	struct rcu_head			rcu_head;
43};
44
45static bool tcp_metric_locked(struct tcp_metrics_block *tm,
46			      enum tcp_metric_index idx)
47{
48	return tm->tcpm_lock & (1 << idx);
49}
50
51static u32 tcp_metric_get(struct tcp_metrics_block *tm,
52			  enum tcp_metric_index idx)
53{
54	return tm->tcpm_vals[idx];
55}
56
57static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm,
58				  enum tcp_metric_index idx)
59{
60	return msecs_to_jiffies(tm->tcpm_vals[idx]);
61}
62
63static void tcp_metric_set(struct tcp_metrics_block *tm,
64			   enum tcp_metric_index idx,
65			   u32 val)
66{
67	tm->tcpm_vals[idx] = val;
68}
69
70static void tcp_metric_set_msecs(struct tcp_metrics_block *tm,
71				 enum tcp_metric_index idx,
72				 u32 val)
73{
74	tm->tcpm_vals[idx] = jiffies_to_msecs(val);
75}
76
77static bool addr_same(const struct inetpeer_addr *a,
78		      const struct inetpeer_addr *b)
79{
80	const struct in6_addr *a6, *b6;
81
82	if (a->family != b->family)
83		return false;
84	if (a->family == AF_INET)
85		return a->addr.a4 == b->addr.a4;
86
87	a6 = (const struct in6_addr *) &a->addr.a6[0];
88	b6 = (const struct in6_addr *) &b->addr.a6[0];
89
90	return ipv6_addr_equal(a6, b6);
91}
92
93struct tcpm_hash_bucket {
94	struct tcp_metrics_block __rcu	*chain;
95};
96
97static DEFINE_SPINLOCK(tcp_metrics_lock);
98
99static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst,
100			  bool fastopen_clear)
101{
102	u32 val;
103
104	tm->tcpm_stamp = jiffies;
105
106	val = 0;
107	if (dst_metric_locked(dst, RTAX_RTT))
108		val |= 1 << TCP_METRIC_RTT;
109	if (dst_metric_locked(dst, RTAX_RTTVAR))
110		val |= 1 << TCP_METRIC_RTTVAR;
111	if (dst_metric_locked(dst, RTAX_SSTHRESH))
112		val |= 1 << TCP_METRIC_SSTHRESH;
113	if (dst_metric_locked(dst, RTAX_CWND))
114		val |= 1 << TCP_METRIC_CWND;
115	if (dst_metric_locked(dst, RTAX_REORDERING))
116		val |= 1 << TCP_METRIC_REORDERING;
117	tm->tcpm_lock = val;
118
119	tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT);
120	tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR);
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	tm->tcpm_ts = 0;
125	tm->tcpm_ts_stamp = 0;
126	if (fastopen_clear) {
127		tm->tcpm_fastopen.mss = 0;
128		tm->tcpm_fastopen.syn_loss = 0;
129		tm->tcpm_fastopen.cookie.len = 0;
130	}
131}
132
133static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
134					  struct inetpeer_addr *addr,
135					  unsigned int hash,
136					  bool reclaim)
137{
138	struct tcp_metrics_block *tm;
139	struct net *net;
140
141	spin_lock_bh(&tcp_metrics_lock);
142	net = dev_net(dst->dev);
143	if (unlikely(reclaim)) {
144		struct tcp_metrics_block *oldest;
145
146		oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
147		for (tm = rcu_dereference(oldest->tcpm_next); tm;
148		     tm = rcu_dereference(tm->tcpm_next)) {
149			if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
150				oldest = tm;
151		}
152		tm = oldest;
153	} else {
154		tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
155		if (!tm)
156			goto out_unlock;
157	}
158	tm->tcpm_addr = *addr;
159
160	tcpm_suck_dst(tm, dst, true);
161
162	if (likely(!reclaim)) {
163		tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
164		rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
165	}
166
167out_unlock:
168	spin_unlock_bh(&tcp_metrics_lock);
169	return tm;
170}
171
172#define TCP_METRICS_TIMEOUT		(60 * 60 * HZ)
173
174static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
175{
176	if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
177		tcpm_suck_dst(tm, dst, false);
178}
179
180#define TCP_METRICS_RECLAIM_DEPTH	5
181#define TCP_METRICS_RECLAIM_PTR		(struct tcp_metrics_block *) 0x1UL
182
183static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
184{
185	if (tm)
186		return tm;
187	if (depth > TCP_METRICS_RECLAIM_DEPTH)
188		return TCP_METRICS_RECLAIM_PTR;
189	return NULL;
190}
191
192static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *addr,
193						   struct net *net, unsigned int hash)
194{
195	struct tcp_metrics_block *tm;
196	int depth = 0;
197
198	for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
199	     tm = rcu_dereference(tm->tcpm_next)) {
200		if (addr_same(&tm->tcpm_addr, addr))
201			break;
202		depth++;
203	}
204	return tcp_get_encode(tm, depth);
205}
206
207static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
208						       struct dst_entry *dst)
209{
210	struct tcp_metrics_block *tm;
211	struct inetpeer_addr addr;
212	unsigned int hash;
213	struct net *net;
214
215	addr.family = req->rsk_ops->family;
216	switch (addr.family) {
217	case AF_INET:
218		addr.addr.a4 = inet_rsk(req)->rmt_addr;
219		hash = (__force unsigned int) addr.addr.a4;
220		break;
221	case AF_INET6:
222		*(struct in6_addr *)addr.addr.a6 = inet6_rsk(req)->rmt_addr;
223		hash = ipv6_addr_hash(&inet6_rsk(req)->rmt_addr);
224		break;
225	default:
226		return NULL;
227	}
228
229	net = dev_net(dst->dev);
230	hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
231
232	for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
233	     tm = rcu_dereference(tm->tcpm_next)) {
234		if (addr_same(&tm->tcpm_addr, &addr))
235			break;
236	}
237	tcpm_check_stamp(tm, dst);
238	return tm;
239}
240
241static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
242{
243	struct inet6_timewait_sock *tw6;
244	struct tcp_metrics_block *tm;
245	struct inetpeer_addr addr;
246	unsigned int hash;
247	struct net *net;
248
249	addr.family = tw->tw_family;
250	switch (addr.family) {
251	case AF_INET:
252		addr.addr.a4 = tw->tw_daddr;
253		hash = (__force unsigned int) addr.addr.a4;
254		break;
255	case AF_INET6:
256		tw6 = inet6_twsk((struct sock *)tw);
257		*(struct in6_addr *)addr.addr.a6 = tw6->tw_v6_daddr;
258		hash = ipv6_addr_hash(&tw6->tw_v6_daddr);
259		break;
260	default:
261		return NULL;
262	}
263
264	net = twsk_net(tw);
265	hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
266
267	for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
268	     tm = rcu_dereference(tm->tcpm_next)) {
269		if (addr_same(&tm->tcpm_addr, &addr))
270			break;
271	}
272	return tm;
273}
274
275static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
276						 struct dst_entry *dst,
277						 bool create)
278{
279	struct tcp_metrics_block *tm;
280	struct inetpeer_addr addr;
281	unsigned int hash;
282	struct net *net;
283	bool reclaim;
284
285	addr.family = sk->sk_family;
286	switch (addr.family) {
287	case AF_INET:
288		addr.addr.a4 = inet_sk(sk)->inet_daddr;
289		hash = (__force unsigned int) addr.addr.a4;
290		break;
291	case AF_INET6:
292		*(struct in6_addr *)addr.addr.a6 = inet6_sk(sk)->daddr;
293		hash = ipv6_addr_hash(&inet6_sk(sk)->daddr);
294		break;
295	default:
296		return NULL;
297	}
298
299	net = dev_net(dst->dev);
300	hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
301
302	tm = __tcp_get_metrics(&addr, net, hash);
303	reclaim = false;
304	if (tm == TCP_METRICS_RECLAIM_PTR) {
305		reclaim = true;
306		tm = NULL;
307	}
308	if (!tm && create)
309		tm = tcpm_new(dst, &addr, hash, reclaim);
310	else
311		tcpm_check_stamp(tm, dst);
312
313	return tm;
314}
315
316/* Save metrics learned by this TCP session.  This function is called
317 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
318 * or goes from LAST-ACK to CLOSE.
319 */
320void tcp_update_metrics(struct sock *sk)
321{
322	const struct inet_connection_sock *icsk = inet_csk(sk);
323	struct dst_entry *dst = __sk_dst_get(sk);
324	struct tcp_sock *tp = tcp_sk(sk);
325	struct tcp_metrics_block *tm;
326	unsigned long rtt;
327	u32 val;
328	int m;
329
330	if (sysctl_tcp_nometrics_save || !dst)
331		return;
332
333	if (dst->flags & DST_HOST)
334		dst_confirm(dst);
335
336	rcu_read_lock();
337	if (icsk->icsk_backoff || !tp->srtt) {
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_jiffies(tm, TCP_METRIC_RTT);
353	m = rtt - tp->srtt;
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;
362		else
363			rtt -= (m >> 3);
364		tcp_metric_set_msecs(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)
376			m = tp->mdev;
377
378		var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
379		if (m >= var)
380			var = m;
381		else
382			var -= (var - m) >> 2;
383
384		tcp_metric_set_msecs(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 (tp->snd_cwnd > tp->snd_ssthresh &&
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 != sysctl_tcp_reordering)
430				tcp_metric_set(tm, TCP_METRIC_REORDERING,
431					       tp->reordering);
432		}
433	}
434	tm->tcpm_stamp = jiffies;
435out_unlock:
436	rcu_read_unlock();
437}
438
439/* Initialize metrics on socket. */
440
441void 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;
447
448	if (dst == NULL)
449		goto reset;
450
451	dst_confirm(dst);
452
453	rcu_read_lock();
454	tm = tcp_get_metrics(sk, dst, true);
455	if (!tm) {
456		rcu_read_unlock();
457		goto reset;
458	}
459
460	if (tcp_metric_locked(tm, TCP_METRIC_CWND))
461		tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
462
463	val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
464	if (val) {
465		tp->snd_ssthresh = val;
466		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
467			tp->snd_ssthresh = tp->snd_cwnd_clamp;
468	} else {
469		/* ssthresh may have been reduced unnecessarily during.
470		 * 3WHS. Restore it back to its initial default.
471		 */
472		tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
473	}
474	val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
475	if (val && tp->reordering != val) {
476		tcp_disable_fack(tp);
477		tcp_disable_early_retrans(tp);
478		tp->reordering = val;
479	}
480
481	val = tcp_metric_get(tm, TCP_METRIC_RTT);
482	if (val == 0 || tp->srtt == 0) {
483		rcu_read_unlock();
484		goto reset;
485	}
486	/* Initial rtt is determined from SYN,SYN-ACK.
487	 * The segment is small and rtt may appear much
488	 * less than real one. Use per-dst memory
489	 * to make it more realistic.
490	 *
491	 * A bit of theory. RTT is time passed after "normal" sized packet
492	 * is sent until it is ACKed. In normal circumstances sending small
493	 * packets force peer to delay ACKs and calculation is correct too.
494	 * The algorithm is adaptive and, provided we follow specs, it
495	 * NEVER underestimate RTT. BUT! If peer tries to make some clever
496	 * tricks sort of "quick acks" for time long enough to decrease RTT
497	 * to low value, and then abruptly stops to do it and starts to delay
498	 * ACKs, wait for troubles.
499	 */
500	val = msecs_to_jiffies(val);
501	if (val > tp->srtt) {
502		tp->srtt = val;
503		tp->rtt_seq = tp->snd_nxt;
504	}
505	val = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
506	if (val > tp->mdev) {
507		tp->mdev = val;
508		tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
509	}
510	rcu_read_unlock();
511
512	tcp_set_rto(sk);
513reset:
514	if (tp->srtt == 0) {
515		/* RFC6298: 5.7 We've failed to get a valid RTT sample from
516		 * 3WHS. This is most likely due to retransmission,
517		 * including spurious one. Reset the RTO back to 3secs
518		 * from the more aggressive 1sec to avoid more spurious
519		 * retransmission.
520		 */
521		tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
522		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
523	}
524	/* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
525	 * retransmitted. In light of RFC6298 more aggressive 1sec
526	 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
527	 * retransmission has occurred.
528	 */
529	if (tp->total_retrans > 1)
530		tp->snd_cwnd = 1;
531	else
532		tp->snd_cwnd = tcp_init_cwnd(tp, dst);
533	tp->snd_cwnd_stamp = tcp_time_stamp;
534}
535
536bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
537{
538	struct tcp_metrics_block *tm;
539	bool ret;
540
541	if (!dst)
542		return false;
543
544	rcu_read_lock();
545	tm = __tcp_get_metrics_req(req, dst);
546	if (paws_check) {
547		if (tm &&
548		    (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
549		    (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
550			ret = false;
551		else
552			ret = true;
553	} else {
554		if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
555			ret = true;
556		else
557			ret = false;
558	}
559	rcu_read_unlock();
560
561	return ret;
562}
563EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
564
565void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
566{
567	struct tcp_metrics_block *tm;
568
569	rcu_read_lock();
570	tm = tcp_get_metrics(sk, dst, true);
571	if (tm) {
572		struct tcp_sock *tp = tcp_sk(sk);
573
574		if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
575			tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
576			tp->rx_opt.ts_recent = tm->tcpm_ts;
577		}
578	}
579	rcu_read_unlock();
580}
581EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
582
583/* VJ's idea. Save last timestamp seen from this destination and hold
584 * it at least for normal timewait interval to use for duplicate
585 * segment detection in subsequent connections, before they enter
586 * synchronized state.
587 */
588bool tcp_remember_stamp(struct sock *sk)
589{
590	struct dst_entry *dst = __sk_dst_get(sk);
591	bool ret = false;
592
593	if (dst) {
594		struct tcp_metrics_block *tm;
595
596		rcu_read_lock();
597		tm = tcp_get_metrics(sk, dst, true);
598		if (tm) {
599			struct tcp_sock *tp = tcp_sk(sk);
600
601			if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
602			    ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
603			     tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
604				tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
605				tm->tcpm_ts = tp->rx_opt.ts_recent;
606			}
607			ret = true;
608		}
609		rcu_read_unlock();
610	}
611	return ret;
612}
613
614bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
615{
616	struct tcp_metrics_block *tm;
617	bool ret = false;
618
619	rcu_read_lock();
620	tm = __tcp_get_metrics_tw(tw);
621	if (tm) {
622		const struct tcp_timewait_sock *tcptw;
623		struct sock *sk = (struct sock *) tw;
624
625		tcptw = tcp_twsk(sk);
626		if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
627		    ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
628		     tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
629			tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
630			tm->tcpm_ts	   = tcptw->tw_ts_recent;
631		}
632		ret = true;
633	}
634	rcu_read_unlock();
635
636	return ret;
637}
638
639static DEFINE_SEQLOCK(fastopen_seqlock);
640
641void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
642			    struct tcp_fastopen_cookie *cookie,
643			    int *syn_loss, unsigned long *last_syn_loss)
644{
645	struct tcp_metrics_block *tm;
646
647	rcu_read_lock();
648	tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
649	if (tm) {
650		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
651		unsigned int seq;
652
653		do {
654			seq = read_seqbegin(&fastopen_seqlock);
655			if (tfom->mss)
656				*mss = tfom->mss;
657			*cookie = tfom->cookie;
658			*syn_loss = tfom->syn_loss;
659			*last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
660		} while (read_seqretry(&fastopen_seqlock, seq));
661	}
662	rcu_read_unlock();
663}
664
665void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
666			    struct tcp_fastopen_cookie *cookie, bool syn_lost)
667{
668	struct tcp_metrics_block *tm;
669
670	rcu_read_lock();
671	tm = tcp_get_metrics(sk, __sk_dst_get(sk), true);
672	if (tm) {
673		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
674
675		write_seqlock_bh(&fastopen_seqlock);
676		tfom->mss = mss;
677		if (cookie->len > 0)
678			tfom->cookie = *cookie;
679		if (syn_lost) {
680			++tfom->syn_loss;
681			tfom->last_syn_loss = jiffies;
682		} else
683			tfom->syn_loss = 0;
684		write_sequnlock_bh(&fastopen_seqlock);
685	}
686	rcu_read_unlock();
687}
688
689static struct genl_family tcp_metrics_nl_family = {
690	.id		= GENL_ID_GENERATE,
691	.hdrsize	= 0,
692	.name		= TCP_METRICS_GENL_NAME,
693	.version	= TCP_METRICS_GENL_VERSION,
694	.maxattr	= TCP_METRICS_ATTR_MAX,
695	.netnsok	= true,
696};
697
698static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
699	[TCP_METRICS_ATTR_ADDR_IPV4]	= { .type = NLA_U32, },
700	[TCP_METRICS_ATTR_ADDR_IPV6]	= { .type = NLA_BINARY,
701					    .len = sizeof(struct in6_addr), },
702	/* Following attributes are not received for GET/DEL,
703	 * we keep them for reference
704	 */
705#if 0
706	[TCP_METRICS_ATTR_AGE]		= { .type = NLA_MSECS, },
707	[TCP_METRICS_ATTR_TW_TSVAL]	= { .type = NLA_U32, },
708	[TCP_METRICS_ATTR_TW_TS_STAMP]	= { .type = NLA_S32, },
709	[TCP_METRICS_ATTR_VALS]		= { .type = NLA_NESTED, },
710	[TCP_METRICS_ATTR_FOPEN_MSS]	= { .type = NLA_U16, },
711	[TCP_METRICS_ATTR_FOPEN_SYN_DROPS]	= { .type = NLA_U16, },
712	[TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]	= { .type = NLA_MSECS, },
713	[TCP_METRICS_ATTR_FOPEN_COOKIE]	= { .type = NLA_BINARY,
714					    .len = TCP_FASTOPEN_COOKIE_MAX, },
715#endif
716};
717
718/* Add attributes, caller cancels its header on failure */
719static int tcp_metrics_fill_info(struct sk_buff *msg,
720				 struct tcp_metrics_block *tm)
721{
722	struct nlattr *nest;
723	int i;
724
725	switch (tm->tcpm_addr.family) {
726	case AF_INET:
727		if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
728				tm->tcpm_addr.addr.a4) < 0)
729			goto nla_put_failure;
730		break;
731	case AF_INET6:
732		if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
733			    tm->tcpm_addr.addr.a6) < 0)
734			goto nla_put_failure;
735		break;
736	default:
737		return -EAFNOSUPPORT;
738	}
739
740	if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
741			  jiffies - tm->tcpm_stamp) < 0)
742		goto nla_put_failure;
743	if (tm->tcpm_ts_stamp) {
744		if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
745				(s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
746			goto nla_put_failure;
747		if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
748				tm->tcpm_ts) < 0)
749			goto nla_put_failure;
750	}
751
752	{
753		int n = 0;
754
755		nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
756		if (!nest)
757			goto nla_put_failure;
758		for (i = 0; i < TCP_METRIC_MAX + 1; i++) {
759			if (!tm->tcpm_vals[i])
760				continue;
761			if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0)
762				goto nla_put_failure;
763			n++;
764		}
765		if (n)
766			nla_nest_end(msg, nest);
767		else
768			nla_nest_cancel(msg, nest);
769	}
770
771	{
772		struct tcp_fastopen_metrics tfom_copy[1], *tfom;
773		unsigned int seq;
774
775		do {
776			seq = read_seqbegin(&fastopen_seqlock);
777			tfom_copy[0] = tm->tcpm_fastopen;
778		} while (read_seqretry(&fastopen_seqlock, seq));
779
780		tfom = tfom_copy;
781		if (tfom->mss &&
782		    nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
783				tfom->mss) < 0)
784			goto nla_put_failure;
785		if (tfom->syn_loss &&
786		    (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
787				tfom->syn_loss) < 0 ||
788		     nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
789				jiffies - tfom->last_syn_loss) < 0))
790			goto nla_put_failure;
791		if (tfom->cookie.len > 0 &&
792		    nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
793			    tfom->cookie.len, tfom->cookie.val) < 0)
794			goto nla_put_failure;
795	}
796
797	return 0;
798
799nla_put_failure:
800	return -EMSGSIZE;
801}
802
803static int tcp_metrics_dump_info(struct sk_buff *skb,
804				 struct netlink_callback *cb,
805				 struct tcp_metrics_block *tm)
806{
807	void *hdr;
808
809	hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
810			  &tcp_metrics_nl_family, NLM_F_MULTI,
811			  TCP_METRICS_CMD_GET);
812	if (!hdr)
813		return -EMSGSIZE;
814
815	if (tcp_metrics_fill_info(skb, tm) < 0)
816		goto nla_put_failure;
817
818	return genlmsg_end(skb, hdr);
819
820nla_put_failure:
821	genlmsg_cancel(skb, hdr);
822	return -EMSGSIZE;
823}
824
825static int tcp_metrics_nl_dump(struct sk_buff *skb,
826			       struct netlink_callback *cb)
827{
828	struct net *net = sock_net(skb->sk);
829	unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
830	unsigned int row, s_row = cb->args[0];
831	int s_col = cb->args[1], col = s_col;
832
833	for (row = s_row; row < max_rows; row++, s_col = 0) {
834		struct tcp_metrics_block *tm;
835		struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
836
837		rcu_read_lock();
838		for (col = 0, tm = rcu_dereference(hb->chain); tm;
839		     tm = rcu_dereference(tm->tcpm_next), col++) {
840			if (col < s_col)
841				continue;
842			if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
843				rcu_read_unlock();
844				goto done;
845			}
846		}
847		rcu_read_unlock();
848	}
849
850done:
851	cb->args[0] = row;
852	cb->args[1] = col;
853	return skb->len;
854}
855
856static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
857			 unsigned int *hash, int optional)
858{
859	struct nlattr *a;
860
861	a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV4];
862	if (a) {
863		addr->family = AF_INET;
864		addr->addr.a4 = nla_get_be32(a);
865		*hash = (__force unsigned int) addr->addr.a4;
866		return 0;
867	}
868	a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV6];
869	if (a) {
870		if (nla_len(a) != sizeof(struct in6_addr))
871			return -EINVAL;
872		addr->family = AF_INET6;
873		memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
874		*hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
875		return 0;
876	}
877	return optional ? 1 : -EAFNOSUPPORT;
878}
879
880static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
881{
882	struct tcp_metrics_block *tm;
883	struct inetpeer_addr addr;
884	unsigned int hash;
885	struct sk_buff *msg;
886	struct net *net = genl_info_net(info);
887	void *reply;
888	int ret;
889
890	ret = parse_nl_addr(info, &addr, &hash, 0);
891	if (ret < 0)
892		return ret;
893
894	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
895	if (!msg)
896		return -ENOMEM;
897
898	reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
899				  info->genlhdr->cmd);
900	if (!reply)
901		goto nla_put_failure;
902
903	hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
904	ret = -ESRCH;
905	rcu_read_lock();
906	for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
907	     tm = rcu_dereference(tm->tcpm_next)) {
908		if (addr_same(&tm->tcpm_addr, &addr)) {
909			ret = tcp_metrics_fill_info(msg, tm);
910			break;
911		}
912	}
913	rcu_read_unlock();
914	if (ret < 0)
915		goto out_free;
916
917	genlmsg_end(msg, reply);
918	return genlmsg_reply(msg, info);
919
920nla_put_failure:
921	ret = -EMSGSIZE;
922
923out_free:
924	nlmsg_free(msg);
925	return ret;
926}
927
928#define deref_locked_genl(p)	\
929	rcu_dereference_protected(p, lockdep_genl_is_held() && \
930				     lockdep_is_held(&tcp_metrics_lock))
931
932#define deref_genl(p)	rcu_dereference_protected(p, lockdep_genl_is_held())
933
934static int tcp_metrics_flush_all(struct net *net)
935{
936	unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
937	struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
938	struct tcp_metrics_block *tm;
939	unsigned int row;
940
941	for (row = 0; row < max_rows; row++, hb++) {
942		spin_lock_bh(&tcp_metrics_lock);
943		tm = deref_locked_genl(hb->chain);
944		if (tm)
945			hb->chain = NULL;
946		spin_unlock_bh(&tcp_metrics_lock);
947		while (tm) {
948			struct tcp_metrics_block *next;
949
950			next = deref_genl(tm->tcpm_next);
951			kfree_rcu(tm, rcu_head);
952			tm = next;
953		}
954	}
955	return 0;
956}
957
958static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
959{
960	struct tcpm_hash_bucket *hb;
961	struct tcp_metrics_block *tm;
962	struct tcp_metrics_block __rcu **pp;
963	struct inetpeer_addr addr;
964	unsigned int hash;
965	struct net *net = genl_info_net(info);
966	int ret;
967
968	ret = parse_nl_addr(info, &addr, &hash, 1);
969	if (ret < 0)
970		return ret;
971	if (ret > 0)
972		return tcp_metrics_flush_all(net);
973
974	hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
975	hb = net->ipv4.tcp_metrics_hash + hash;
976	pp = &hb->chain;
977	spin_lock_bh(&tcp_metrics_lock);
978	for (tm = deref_locked_genl(*pp); tm;
979	     pp = &tm->tcpm_next, tm = deref_locked_genl(*pp)) {
980		if (addr_same(&tm->tcpm_addr, &addr)) {
981			*pp = tm->tcpm_next;
982			break;
983		}
984	}
985	spin_unlock_bh(&tcp_metrics_lock);
986	if (!tm)
987		return -ESRCH;
988	kfree_rcu(tm, rcu_head);
989	return 0;
990}
991
992static struct genl_ops tcp_metrics_nl_ops[] = {
993	{
994		.cmd = TCP_METRICS_CMD_GET,
995		.doit = tcp_metrics_nl_cmd_get,
996		.dumpit = tcp_metrics_nl_dump,
997		.policy = tcp_metrics_nl_policy,
998		.flags = GENL_ADMIN_PERM,
999	},
1000	{
1001		.cmd = TCP_METRICS_CMD_DEL,
1002		.doit = tcp_metrics_nl_cmd_del,
1003		.policy = tcp_metrics_nl_policy,
1004		.flags = GENL_ADMIN_PERM,
1005	},
1006};
1007
1008static unsigned int tcpmhash_entries;
1009static int __init set_tcpmhash_entries(char *str)
1010{
1011	ssize_t ret;
1012
1013	if (!str)
1014		return 0;
1015
1016	ret = kstrtouint(str, 0, &tcpmhash_entries);
1017	if (ret)
1018		return 0;
1019
1020	return 1;
1021}
1022__setup("tcpmhash_entries=", set_tcpmhash_entries);
1023
1024static int __net_init tcp_net_metrics_init(struct net *net)
1025{
1026	size_t size;
1027	unsigned int slots;
1028
1029	slots = tcpmhash_entries;
1030	if (!slots) {
1031		if (totalram_pages >= 128 * 1024)
1032			slots = 16 * 1024;
1033		else
1034			slots = 8 * 1024;
1035	}
1036
1037	net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1038	size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
1039
1040	net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1041	if (!net->ipv4.tcp_metrics_hash)
1042		net->ipv4.tcp_metrics_hash = vzalloc(size);
1043
1044	if (!net->ipv4.tcp_metrics_hash)
1045		return -ENOMEM;
1046
1047	return 0;
1048}
1049
1050static void __net_exit tcp_net_metrics_exit(struct net *net)
1051{
1052	unsigned int i;
1053
1054	for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1055		struct tcp_metrics_block *tm, *next;
1056
1057		tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1058		while (tm) {
1059			next = rcu_dereference_protected(tm->tcpm_next, 1);
1060			kfree(tm);
1061			tm = next;
1062		}
1063	}
1064	if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1065		vfree(net->ipv4.tcp_metrics_hash);
1066	else
1067		kfree(net->ipv4.tcp_metrics_hash);
1068}
1069
1070static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1071	.init	=	tcp_net_metrics_init,
1072	.exit	=	tcp_net_metrics_exit,
1073};
1074
1075void __init tcp_metrics_init(void)
1076{
1077	int ret;
1078
1079	ret = register_pernet_subsys(&tcp_net_metrics_ops);
1080	if (ret < 0)
1081		goto cleanup;
1082	ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1083					    tcp_metrics_nl_ops,
1084					    ARRAY_SIZE(tcp_metrics_nl_ops));
1085	if (ret < 0)
1086		goto cleanup_subsys;
1087	return;
1088
1089cleanup_subsys:
1090	unregister_pernet_subsys(&tcp_net_metrics_ops);
1091
1092cleanup:
1093	return;
1094}
1095