?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
net: Add set,get helpers for inetpeer addresses
[linux/fpc-iii.git] / net / ipv4 / tcp_metrics.c
blob4ef4dd4bf38cc98d6d2bcc78b2c008d37eea3144
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_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
57 static inline struct net *tm_net(struct tcp_metrics_block *tm)
58 {
59 return read_pnet(&tm->tcpm_net);
60 }
61
62 static 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
68 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
69 enum tcp_metric_index idx)
70 {
71 return tm->tcpm_vals[idx];
72 }
73
74 static void tcp_metric_set(struct tcp_metrics_block *tm,
75 enum tcp_metric_index idx,
76 u32 val)
77 {
78 tm->tcpm_vals[idx] = val;
79 }
80
81 static 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
91 struct tcpm_hash_bucket {
92 struct tcp_metrics_block __rcu *chain;
93 };
94
95 static struct tcpm_hash_bucket *tcp_metrics_hash __read_mostly;
96 static unsigned int tcp_metrics_hash_log __read_mostly;
97
98 static DEFINE_SPINLOCK(tcp_metrics_lock);
99
100 static 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
143 static 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
155 static 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
206 out_unlock:
207 spin_unlock_bh(&tcp_metrics_lock);
208 return tm;
209 }
210
211 static 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
220 static 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
238 static 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 inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
251 inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
252 hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
253 break;
254 #if IS_ENABLED(CONFIG_IPV6)
255 case AF_INET6:
256 inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
257 inetpeer_set_addr_v6(&daddr, &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
280 static 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 inetpeer_set_addr_v4(&saddr, tw->tw_rcv_saddr);
289 inetpeer_set_addr_v4(&daddr, tw->tw_daddr);
290 hash = ipv4_addr_hash(tw->tw_daddr);
291 }
292 #if IS_ENABLED(CONFIG_IPV6)
293 else if (tw->tw_family == AF_INET6) {
294 if (ipv6_addr_v4mapped(&tw->tw_v6_daddr)) {
295 inetpeer_set_addr_v4(&saddr, tw->tw_rcv_saddr);
296 inetpeer_set_addr_v4(&daddr, tw->tw_daddr);
297 hash = ipv4_addr_hash(tw->tw_daddr);
298 } else {
299 inetpeer_set_addr_v6(&saddr, &tw->tw_v6_rcv_saddr);
300 inetpeer_set_addr_v6(&daddr, &tw->tw_v6_daddr);
301 hash = ipv6_addr_hash(&tw->tw_v6_daddr);
302 }
303 }
304 #endif
305 else
306 return NULL;
307
308 net = twsk_net(tw);
309 hash ^= net_hash_mix(net);
310 hash = hash_32(hash, tcp_metrics_hash_log);
311
312 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
313 tm = rcu_dereference(tm->tcpm_next)) {
314 if (addr_same(&tm->tcpm_saddr, &saddr) &&
315 addr_same(&tm->tcpm_daddr, &daddr) &&
316 net_eq(tm_net(tm), net))
317 break;
318 }
319 return tm;
320 }
321
322 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
323 struct dst_entry *dst,
324 bool create)
325 {
326 struct tcp_metrics_block *tm;
327 struct inetpeer_addr saddr, daddr;
328 unsigned int hash;
329 struct net *net;
330
331 if (sk->sk_family == AF_INET) {
332 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
333 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
334 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
335 }
336 #if IS_ENABLED(CONFIG_IPV6)
337 else if (sk->sk_family == AF_INET6) {
338 if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
339 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
340 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
341 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
342 } else {
343 inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
344 inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
345 hash = ipv6_addr_hash(&sk->sk_v6_daddr);
346 }
347 }
348 #endif
349 else
350 return NULL;
351
352 net = dev_net(dst->dev);
353 hash ^= net_hash_mix(net);
354 hash = hash_32(hash, tcp_metrics_hash_log);
355
356 tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
357 if (tm == TCP_METRICS_RECLAIM_PTR)
358 tm = NULL;
359 if (!tm && create)
360 tm = tcpm_new(dst, &saddr, &daddr, hash);
361 else
362 tcpm_check_stamp(tm, dst);
363
364 return tm;
365 }
366
367 /* Save metrics learned by this TCP session. This function is called
368 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
369 * or goes from LAST-ACK to CLOSE.
370 */
371 void tcp_update_metrics(struct sock *sk)
372 {
373 const struct inet_connection_sock *icsk = inet_csk(sk);
374 struct dst_entry *dst = __sk_dst_get(sk);
375 struct tcp_sock *tp = tcp_sk(sk);
376 struct tcp_metrics_block *tm;
377 unsigned long rtt;
378 u32 val;
379 int m;
380
381 if (sysctl_tcp_nometrics_save || !dst)
382 return;
383
384 if (dst->flags & DST_HOST)
385 dst_confirm(dst);
386
387 rcu_read_lock();
388 if (icsk->icsk_backoff || !tp->srtt_us) {
389 /* This session failed to estimate rtt. Why?
390 * Probably, no packets returned in time. Reset our
391 * results.
392 */
393 tm = tcp_get_metrics(sk, dst, false);
394 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
395 tcp_metric_set(tm, TCP_METRIC_RTT, 0);
396 goto out_unlock;
397 } else
398 tm = tcp_get_metrics(sk, dst, true);
399
400 if (!tm)
401 goto out_unlock;
402
403 rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
404 m = rtt - tp->srtt_us;
405
406 /* If newly calculated rtt larger than stored one, store new
407 * one. Otherwise, use EWMA. Remember, rtt overestimation is
408 * always better than underestimation.
409 */
410 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
411 if (m <= 0)
412 rtt = tp->srtt_us;
413 else
414 rtt -= (m >> 3);
415 tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
416 }
417
418 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
419 unsigned long var;
420
421 if (m < 0)
422 m = -m;
423
424 /* Scale deviation to rttvar fixed point */
425 m >>= 1;
426 if (m < tp->mdev_us)
427 m = tp->mdev_us;
428
429 var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
430 if (m >= var)
431 var = m;
432 else
433 var -= (var - m) >> 2;
434
435 tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
436 }
437
438 if (tcp_in_initial_slowstart(tp)) {
439 /* Slow start still did not finish. */
440 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
441 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
442 if (val && (tp->snd_cwnd >> 1) > val)
443 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
444 tp->snd_cwnd >> 1);
445 }
446 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
447 val = tcp_metric_get(tm, TCP_METRIC_CWND);
448 if (tp->snd_cwnd > val)
449 tcp_metric_set(tm, TCP_METRIC_CWND,
450 tp->snd_cwnd);
451 }
452 } else if (!tcp_in_slow_start(tp) &&
453 icsk->icsk_ca_state == TCP_CA_Open) {
454 /* Cong. avoidance phase, cwnd is reliable. */
455 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
456 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
457 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
458 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
459 val = tcp_metric_get(tm, TCP_METRIC_CWND);
460 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
461 }
462 } else {
463 /* Else slow start did not finish, cwnd is non-sense,
464 * ssthresh may be also invalid.
465 */
466 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
467 val = tcp_metric_get(tm, TCP_METRIC_CWND);
468 tcp_metric_set(tm, TCP_METRIC_CWND,
469 (val + tp->snd_ssthresh) >> 1);
470 }
471 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
472 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
473 if (val && tp->snd_ssthresh > val)
474 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
475 tp->snd_ssthresh);
476 }
477 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
478 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
479 if (val < tp->reordering &&
480 tp->reordering != sysctl_tcp_reordering)
481 tcp_metric_set(tm, TCP_METRIC_REORDERING,
482 tp->reordering);
483 }
484 }
485 tm->tcpm_stamp = jiffies;
486 out_unlock:
487 rcu_read_unlock();
488 }
489
490 /* Initialize metrics on socket. */
491
492 void tcp_init_metrics(struct sock *sk)
493 {
494 struct dst_entry *dst = __sk_dst_get(sk);
495 struct tcp_sock *tp = tcp_sk(sk);
496 struct tcp_metrics_block *tm;
497 u32 val, crtt = 0; /* cached RTT scaled by 8 */
498
499 if (!dst)
500 goto reset;
501
502 dst_confirm(dst);
503
504 rcu_read_lock();
505 tm = tcp_get_metrics(sk, dst, true);
506 if (!tm) {
507 rcu_read_unlock();
508 goto reset;
509 }
510
511 if (tcp_metric_locked(tm, TCP_METRIC_CWND))
512 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
513
514 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
515 if (val) {
516 tp->snd_ssthresh = val;
517 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
518 tp->snd_ssthresh = tp->snd_cwnd_clamp;
519 } else {
520 /* ssthresh may have been reduced unnecessarily during.
521 * 3WHS. Restore it back to its initial default.
522 */
523 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
524 }
525 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
526 if (val && tp->reordering != val) {
527 tcp_disable_fack(tp);
528 tcp_disable_early_retrans(tp);
529 tp->reordering = val;
530 }
531
532 crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
533 rcu_read_unlock();
534 reset:
535 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
536 * to seed the RTO for later data packets because SYN packets are
537 * small. Use the per-dst cached values to seed the RTO but keep
538 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
539 * Later the RTO will be updated immediately upon obtaining the first
540 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
541 * influences the first RTO but not later RTT estimation.
542 *
543 * But if RTT is not available from the SYN (due to retransmits or
544 * syn cookies) or the cache, force a conservative 3secs timeout.
545 *
546 * A bit of theory. RTT is time passed after "normal" sized packet
547 * is sent until it is ACKed. In normal circumstances sending small
548 * packets force peer to delay ACKs and calculation is correct too.
549 * The algorithm is adaptive and, provided we follow specs, it
550 * NEVER underestimate RTT. BUT! If peer tries to make some clever
551 * tricks sort of "quick acks" for time long enough to decrease RTT
552 * to low value, and then abruptly stops to do it and starts to delay
553 * ACKs, wait for troubles.
554 */
555 if (crtt > tp->srtt_us) {
556 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
557 crtt /= 8 * USEC_PER_MSEC;
558 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
559 } else if (tp->srtt_us == 0) {
560 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
561 * 3WHS. This is most likely due to retransmission,
562 * including spurious one. Reset the RTO back to 3secs
563 * from the more aggressive 1sec to avoid more spurious
564 * retransmission.
565 */
566 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
567 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
568
569 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
570 }
571 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
572 * retransmitted. In light of RFC6298 more aggressive 1sec
573 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
574 * retransmission has occurred.
575 */
576 if (tp->total_retrans > 1)
577 tp->snd_cwnd = 1;
578 else
579 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
580 tp->snd_cwnd_stamp = tcp_time_stamp;
581 }
582
583 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst,
584 bool paws_check, bool timestamps)
585 {
586 struct tcp_metrics_block *tm;
587 bool ret;
588
589 if (!dst)
590 return false;
591
592 rcu_read_lock();
593 tm = __tcp_get_metrics_req(req, dst);
594 if (paws_check) {
595 if (tm &&
596 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
597 ((s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW ||
598 !timestamps))
599 ret = false;
600 else
601 ret = true;
602 } else {
603 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
604 ret = true;
605 else
606 ret = false;
607 }
608 rcu_read_unlock();
609
610 return ret;
611 }
612 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
613
614 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
615 {
616 struct tcp_metrics_block *tm;
617
618 rcu_read_lock();
619 tm = tcp_get_metrics(sk, dst, true);
620 if (tm) {
621 struct tcp_sock *tp = tcp_sk(sk);
622
623 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
624 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
625 tp->rx_opt.ts_recent = tm->tcpm_ts;
626 }
627 }
628 rcu_read_unlock();
629 }
630 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
631
632 /* VJ's idea. Save last timestamp seen from this destination and hold
633 * it at least for normal timewait interval to use for duplicate
634 * segment detection in subsequent connections, before they enter
635 * synchronized state.
636 */
637 bool tcp_remember_stamp(struct sock *sk)
638 {
639 struct dst_entry *dst = __sk_dst_get(sk);
640 bool ret = false;
641
642 if (dst) {
643 struct tcp_metrics_block *tm;
644
645 rcu_read_lock();
646 tm = tcp_get_metrics(sk, dst, true);
647 if (tm) {
648 struct tcp_sock *tp = tcp_sk(sk);
649
650 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
651 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
652 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
653 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
654 tm->tcpm_ts = tp->rx_opt.ts_recent;
655 }
656 ret = true;
657 }
658 rcu_read_unlock();
659 }
660 return ret;
661 }
662
663 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
664 {
665 struct tcp_metrics_block *tm;
666 bool ret = false;
667
668 rcu_read_lock();
669 tm = __tcp_get_metrics_tw(tw);
670 if (tm) {
671 const struct tcp_timewait_sock *tcptw;
672 struct sock *sk = (struct sock *) tw;
673
674 tcptw = tcp_twsk(sk);
675 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
676 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
677 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
678 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
679 tm->tcpm_ts = tcptw->tw_ts_recent;
680 }
681 ret = true;
682 }
683 rcu_read_unlock();
684
685 return ret;
686 }
687
688 static DEFINE_SEQLOCK(fastopen_seqlock);
689
690 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
691 struct tcp_fastopen_cookie *cookie,
692 int *syn_loss, unsigned long *last_syn_loss)
693 {
694 struct tcp_metrics_block *tm;
695
696 rcu_read_lock();
697 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
698 if (tm) {
699 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
700 unsigned int seq;
701
702 do {
703 seq = read_seqbegin(&fastopen_seqlock);
704 if (tfom->mss)
705 *mss = tfom->mss;
706 *cookie = tfom->cookie;
707 if (cookie->len <= 0 && tfom->try_exp == 1)
708 cookie->exp = true;
709 *syn_loss = tfom->syn_loss;
710 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
711 } while (read_seqretry(&fastopen_seqlock, seq));
712 }
713 rcu_read_unlock();
714 }
715
716 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
717 struct tcp_fastopen_cookie *cookie, bool syn_lost,
718 u16 try_exp)
719 {
720 struct dst_entry *dst = __sk_dst_get(sk);
721 struct tcp_metrics_block *tm;
722
723 if (!dst)
724 return;
725 rcu_read_lock();
726 tm = tcp_get_metrics(sk, dst, true);
727 if (tm) {
728 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
729
730 write_seqlock_bh(&fastopen_seqlock);
731 if (mss)
732 tfom->mss = mss;
733 if (cookie && cookie->len > 0)
734 tfom->cookie = *cookie;
735 else if (try_exp > tfom->try_exp &&
736 tfom->cookie.len <= 0 && !tfom->cookie.exp)
737 tfom->try_exp = try_exp;
738 if (syn_lost) {
739 ++tfom->syn_loss;
740 tfom->last_syn_loss = jiffies;
741 } else
742 tfom->syn_loss = 0;
743 write_sequnlock_bh(&fastopen_seqlock);
744 }
745 rcu_read_unlock();
746 }
747
748 static struct genl_family tcp_metrics_nl_family = {
749 .id = GENL_ID_GENERATE,
750 .hdrsize = 0,
751 .name = TCP_METRICS_GENL_NAME,
752 .version = TCP_METRICS_GENL_VERSION,
753 .maxattr = TCP_METRICS_ATTR_MAX,
754 .netnsok = true,
755 };
756
757 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
758 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
759 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
760 .len = sizeof(struct in6_addr), },
761 /* Following attributes are not received for GET/DEL,
762 * we keep them for reference
763 */
764 #if 0
765 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
766 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
767 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
768 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
769 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
770 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
771 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
772 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
773 .len = TCP_FASTOPEN_COOKIE_MAX, },
774 #endif
775 };
776
777 /* Add attributes, caller cancels its header on failure */
778 static int tcp_metrics_fill_info(struct sk_buff *msg,
779 struct tcp_metrics_block *tm)
780 {
781 struct nlattr *nest;
782 int i;
783
784 switch (tm->tcpm_daddr.family) {
785 case AF_INET:
786 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
787 inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
788 goto nla_put_failure;
789 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
790 inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
791 goto nla_put_failure;
792 break;
793 case AF_INET6:
794 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
795 inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
796 goto nla_put_failure;
797 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
798 inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
799 goto nla_put_failure;
800 break;
801 default:
802 return -EAFNOSUPPORT;
803 }
804
805 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
806 jiffies - tm->tcpm_stamp) < 0)
807 goto nla_put_failure;
808 if (tm->tcpm_ts_stamp) {
809 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
810 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
811 goto nla_put_failure;
812 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
813 tm->tcpm_ts) < 0)
814 goto nla_put_failure;
815 }
816
817 {
818 int n = 0;
819
820 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
821 if (!nest)
822 goto nla_put_failure;
823 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
824 u32 val = tm->tcpm_vals[i];
825
826 if (!val)
827 continue;
828 if (i == TCP_METRIC_RTT) {
829 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
830 val) < 0)
831 goto nla_put_failure;
832 n++;
833 val = max(val / 1000, 1U);
834 }
835 if (i == TCP_METRIC_RTTVAR) {
836 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
837 val) < 0)
838 goto nla_put_failure;
839 n++;
840 val = max(val / 1000, 1U);
841 }
842 if (nla_put_u32(msg, i + 1, val) < 0)
843 goto nla_put_failure;
844 n++;
845 }
846 if (n)
847 nla_nest_end(msg, nest);
848 else
849 nla_nest_cancel(msg, nest);
850 }
851
852 {
853 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
854 unsigned int seq;
855
856 do {
857 seq = read_seqbegin(&fastopen_seqlock);
858 tfom_copy[0] = tm->tcpm_fastopen;
859 } while (read_seqretry(&fastopen_seqlock, seq));
860
861 tfom = tfom_copy;
862 if (tfom->mss &&
863 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
864 tfom->mss) < 0)
865 goto nla_put_failure;
866 if (tfom->syn_loss &&
867 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
868 tfom->syn_loss) < 0 ||
869 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
870 jiffies - tfom->last_syn_loss) < 0))
871 goto nla_put_failure;
872 if (tfom->cookie.len > 0 &&
873 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
874 tfom->cookie.len, tfom->cookie.val) < 0)
875 goto nla_put_failure;
876 }
877
878 return 0;
879
880 nla_put_failure:
881 return -EMSGSIZE;
882 }
883
884 static int tcp_metrics_dump_info(struct sk_buff *skb,
885 struct netlink_callback *cb,
886 struct tcp_metrics_block *tm)
887 {
888 void *hdr;
889
890 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
891 &tcp_metrics_nl_family, NLM_F_MULTI,
892 TCP_METRICS_CMD_GET);
893 if (!hdr)
894 return -EMSGSIZE;
895
896 if (tcp_metrics_fill_info(skb, tm) < 0)
897 goto nla_put_failure;
898
899 genlmsg_end(skb, hdr);
900 return 0;
901
902 nla_put_failure:
903 genlmsg_cancel(skb, hdr);
904 return -EMSGSIZE;
905 }
906
907 static int tcp_metrics_nl_dump(struct sk_buff *skb,
908 struct netlink_callback *cb)
909 {
910 struct net *net = sock_net(skb->sk);
911 unsigned int max_rows = 1U << tcp_metrics_hash_log;
912 unsigned int row, s_row = cb->args[0];
913 int s_col = cb->args[1], col = s_col;
914
915 for (row = s_row; row < max_rows; row++, s_col = 0) {
916 struct tcp_metrics_block *tm;
917 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
918
919 rcu_read_lock();
920 for (col = 0, tm = rcu_dereference(hb->chain); tm;
921 tm = rcu_dereference(tm->tcpm_next), col++) {
922 if (!net_eq(tm_net(tm), net))
923 continue;
924 if (col < s_col)
925 continue;
926 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
927 rcu_read_unlock();
928 goto done;
929 }
930 }
931 rcu_read_unlock();
932 }
933
934 done:
935 cb->args[0] = row;
936 cb->args[1] = col;
937 return skb->len;
938 }
939
940 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
941 unsigned int *hash, int optional, int v4, int v6)
942 {
943 struct nlattr *a;
944
945 a = info->attrs[v4];
946 if (a) {
947 inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
948 if (hash)
949 *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
950 return 0;
951 }
952 a = info->attrs[v6];
953 if (a) {
954 struct in6_addr in6;
955
956 if (nla_len(a) != sizeof(struct in6_addr))
957 return -EINVAL;
958 in6 = nla_get_in6_addr(a);
959 inetpeer_set_addr_v6(addr, &in6);
960 if (hash)
961 *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
962 return 0;
963 }
964 return optional ? 1 : -EAFNOSUPPORT;
965 }
966
967 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
968 unsigned int *hash, int optional)
969 {
970 return __parse_nl_addr(info, addr, hash, optional,
971 TCP_METRICS_ATTR_ADDR_IPV4,
972 TCP_METRICS_ATTR_ADDR_IPV6);
973 }
974
975 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
976 {
977 return __parse_nl_addr(info, addr, NULL, 0,
978 TCP_METRICS_ATTR_SADDR_IPV4,
979 TCP_METRICS_ATTR_SADDR_IPV6);
980 }
981
982 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
983 {
984 struct tcp_metrics_block *tm;
985 struct inetpeer_addr saddr, daddr;
986 unsigned int hash;
987 struct sk_buff *msg;
988 struct net *net = genl_info_net(info);
989 void *reply;
990 int ret;
991 bool src = true;
992
993 ret = parse_nl_addr(info, &daddr, &hash, 0);
994 if (ret < 0)
995 return ret;
996
997 ret = parse_nl_saddr(info, &saddr);
998 if (ret < 0)
999 src = false;
1000
1001 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1002 if (!msg)
1003 return -ENOMEM;
1004
1005 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
1006 info->genlhdr->cmd);
1007 if (!reply)
1008 goto nla_put_failure;
1009
1010 hash ^= net_hash_mix(net);
1011 hash = hash_32(hash, tcp_metrics_hash_log);
1012 ret = -ESRCH;
1013 rcu_read_lock();
1014 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
1015 tm = rcu_dereference(tm->tcpm_next)) {
1016 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1017 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1018 net_eq(tm_net(tm), net)) {
1019 ret = tcp_metrics_fill_info(msg, tm);
1020 break;
1021 }
1022 }
1023 rcu_read_unlock();
1024 if (ret < 0)
1025 goto out_free;
1026
1027 genlmsg_end(msg, reply);
1028 return genlmsg_reply(msg, info);
1029
1030 nla_put_failure:
1031 ret = -EMSGSIZE;
1032
1033 out_free:
1034 nlmsg_free(msg);
1035 return ret;
1036 }
1037
1038 static void tcp_metrics_flush_all(struct net *net)
1039 {
1040 unsigned int max_rows = 1U << tcp_metrics_hash_log;
1041 struct tcpm_hash_bucket *hb = tcp_metrics_hash;
1042 struct tcp_metrics_block *tm;
1043 unsigned int row;
1044
1045 for (row = 0; row < max_rows; row++, hb++) {
1046 struct tcp_metrics_block __rcu **pp;
1047 spin_lock_bh(&tcp_metrics_lock);
1048 pp = &hb->chain;
1049 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1050 if (net_eq(tm_net(tm), net)) {
1051 *pp = tm->tcpm_next;
1052 kfree_rcu(tm, rcu_head);
1053 } else {
1054 pp = &tm->tcpm_next;
1055 }
1056 }
1057 spin_unlock_bh(&tcp_metrics_lock);
1058 }
1059 }
1060
1061 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
1062 {
1063 struct tcpm_hash_bucket *hb;
1064 struct tcp_metrics_block *tm;
1065 struct tcp_metrics_block __rcu **pp;
1066 struct inetpeer_addr saddr, daddr;
1067 unsigned int hash;
1068 struct net *net = genl_info_net(info);
1069 int ret;
1070 bool src = true, found = false;
1071
1072 ret = parse_nl_addr(info, &daddr, &hash, 1);
1073 if (ret < 0)
1074 return ret;
1075 if (ret > 0) {
1076 tcp_metrics_flush_all(net);
1077 return 0;
1078 }
1079 ret = parse_nl_saddr(info, &saddr);
1080 if (ret < 0)
1081 src = false;
1082
1083 hash ^= net_hash_mix(net);
1084 hash = hash_32(hash, tcp_metrics_hash_log);
1085 hb = tcp_metrics_hash + hash;
1086 pp = &hb->chain;
1087 spin_lock_bh(&tcp_metrics_lock);
1088 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1089 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1090 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1091 net_eq(tm_net(tm), net)) {
1092 *pp = tm->tcpm_next;
1093 kfree_rcu(tm, rcu_head);
1094 found = true;
1095 } else {
1096 pp = &tm->tcpm_next;
1097 }
1098 }
1099 spin_unlock_bh(&tcp_metrics_lock);
1100 if (!found)
1101 return -ESRCH;
1102 return 0;
1103 }
1104
1105 static const struct genl_ops tcp_metrics_nl_ops[] = {
1106 {
1107 .cmd = TCP_METRICS_CMD_GET,
1108 .doit = tcp_metrics_nl_cmd_get,
1109 .dumpit = tcp_metrics_nl_dump,
1110 .policy = tcp_metrics_nl_policy,
1111 },
1112 {
1113 .cmd = TCP_METRICS_CMD_DEL,
1114 .doit = tcp_metrics_nl_cmd_del,
1115 .policy = tcp_metrics_nl_policy,
1116 .flags = GENL_ADMIN_PERM,
1117 },
1118 };
1119
1120 static unsigned int tcpmhash_entries;
1121 static int __init set_tcpmhash_entries(char *str)
1122 {
1123 ssize_t ret;
1124
1125 if (!str)
1126 return 0;
1127
1128 ret = kstrtouint(str, 0, &tcpmhash_entries);
1129 if (ret)
1130 return 0;
1131
1132 return 1;
1133 }
1134 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1135
1136 static int __net_init tcp_net_metrics_init(struct net *net)
1137 {
1138 size_t size;
1139 unsigned int slots;
1140
1141 if (!net_eq(net, &init_net))
1142 return 0;
1143
1144 slots = tcpmhash_entries;
1145 if (!slots) {
1146 if (totalram_pages >= 128 * 1024)
1147 slots = 16 * 1024;
1148 else
1149 slots = 8 * 1024;
1150 }
1151
1152 tcp_metrics_hash_log = order_base_2(slots);
1153 size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1154
1155 tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1156 if (!tcp_metrics_hash)
1157 tcp_metrics_hash = vzalloc(size);
1158
1159 if (!tcp_metrics_hash)
1160 return -ENOMEM;
1161
1162 return 0;
1163 }
1164
1165 static void __net_exit tcp_net_metrics_exit(struct net *net)
1166 {
1167 tcp_metrics_flush_all(net);
1168 }
1169
1170 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1171 .init = tcp_net_metrics_init,
1172 .exit = tcp_net_metrics_exit,
1173 };
1174
1175 void __init tcp_metrics_init(void)
1176 {
1177 int ret;
1178
1179 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1180 if (ret < 0)
1181 panic("Could not allocate the tcp_metrics hash table\n");
1182
1183 ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1184 tcp_metrics_nl_ops);
1185 if (ret < 0)
1186 panic("Could not register tcp_metrics generic netlink\n");
1187 }
Linux with added FPC-III support