?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge branch 'bnxt_en-next'
[linux/fpc-iii.git] / net / ipv4 / tcp_metrics.c
blobb617826e24770a09ba07202b3d546ae248b68a50
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 return inetpeer_addr_cmp(a, b) == 0;
85 }
86
87 struct tcpm_hash_bucket {
88 struct tcp_metrics_block __rcu *chain;
89 };
90
91 static struct tcpm_hash_bucket *tcp_metrics_hash __read_mostly;
92 static unsigned int tcp_metrics_hash_log __read_mostly;
93
94 static DEFINE_SPINLOCK(tcp_metrics_lock);
95
96 static void tcpm_suck_dst(struct tcp_metrics_block *tm,
97 const struct dst_entry *dst,
98 bool fastopen_clear)
99 {
100 u32 msval;
101 u32 val;
102
103 tm->tcpm_stamp = jiffies;
104
105 val = 0;
106 if (dst_metric_locked(dst, RTAX_RTT))
107 val |= 1 << TCP_METRIC_RTT;
108 if (dst_metric_locked(dst, RTAX_RTTVAR))
109 val |= 1 << TCP_METRIC_RTTVAR;
110 if (dst_metric_locked(dst, RTAX_SSTHRESH))
111 val |= 1 << TCP_METRIC_SSTHRESH;
112 if (dst_metric_locked(dst, RTAX_CWND))
113 val |= 1 << TCP_METRIC_CWND;
114 if (dst_metric_locked(dst, RTAX_REORDERING))
115 val |= 1 << TCP_METRIC_REORDERING;
116 tm->tcpm_lock = val;
117
118 msval = dst_metric_raw(dst, RTAX_RTT);
119 tm->tcpm_vals[TCP_METRIC_RTT] = msval * USEC_PER_MSEC;
120
121 msval = dst_metric_raw(dst, RTAX_RTTVAR);
122 tm->tcpm_vals[TCP_METRIC_RTTVAR] = msval * USEC_PER_MSEC;
123 tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
124 tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
125 tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
126 tm->tcpm_ts = 0;
127 tm->tcpm_ts_stamp = 0;
128 if (fastopen_clear) {
129 tm->tcpm_fastopen.mss = 0;
130 tm->tcpm_fastopen.syn_loss = 0;
131 tm->tcpm_fastopen.try_exp = 0;
132 tm->tcpm_fastopen.cookie.exp = false;
133 tm->tcpm_fastopen.cookie.len = 0;
134 }
135 }
136
137 #define TCP_METRICS_TIMEOUT (60 * 60 * HZ)
138
139 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
140 {
141 if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
142 tcpm_suck_dst(tm, dst, false);
143 }
144
145 #define TCP_METRICS_RECLAIM_DEPTH 5
146 #define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL
147
148 #define deref_locked(p) \
149 rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
150
151 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
152 struct inetpeer_addr *saddr,
153 struct inetpeer_addr *daddr,
154 unsigned int hash)
155 {
156 struct tcp_metrics_block *tm;
157 struct net *net;
158 bool reclaim = false;
159
160 spin_lock_bh(&tcp_metrics_lock);
161 net = dev_net(dst->dev);
162
163 /* While waiting for the spin-lock the cache might have been populated
164 * with this entry and so we have to check again.
165 */
166 tm = __tcp_get_metrics(saddr, daddr, net, hash);
167 if (tm == TCP_METRICS_RECLAIM_PTR) {
168 reclaim = true;
169 tm = NULL;
170 }
171 if (tm) {
172 tcpm_check_stamp(tm, dst);
173 goto out_unlock;
174 }
175
176 if (unlikely(reclaim)) {
177 struct tcp_metrics_block *oldest;
178
179 oldest = deref_locked(tcp_metrics_hash[hash].chain);
180 for (tm = deref_locked(oldest->tcpm_next); tm;
181 tm = deref_locked(tm->tcpm_next)) {
182 if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
183 oldest = tm;
184 }
185 tm = oldest;
186 } else {
187 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
188 if (!tm)
189 goto out_unlock;
190 }
191 write_pnet(&tm->tcpm_net, net);
192 tm->tcpm_saddr = *saddr;
193 tm->tcpm_daddr = *daddr;
194
195 tcpm_suck_dst(tm, dst, true);
196
197 if (likely(!reclaim)) {
198 tm->tcpm_next = tcp_metrics_hash[hash].chain;
199 rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
200 }
201
202 out_unlock:
203 spin_unlock_bh(&tcp_metrics_lock);
204 return tm;
205 }
206
207 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
208 {
209 if (tm)
210 return tm;
211 if (depth > TCP_METRICS_RECLAIM_DEPTH)
212 return TCP_METRICS_RECLAIM_PTR;
213 return NULL;
214 }
215
216 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
217 const struct inetpeer_addr *daddr,
218 struct net *net, unsigned int hash)
219 {
220 struct tcp_metrics_block *tm;
221 int depth = 0;
222
223 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
224 tm = rcu_dereference(tm->tcpm_next)) {
225 if (addr_same(&tm->tcpm_saddr, saddr) &&
226 addr_same(&tm->tcpm_daddr, daddr) &&
227 net_eq(tm_net(tm), net))
228 break;
229 depth++;
230 }
231 return tcp_get_encode(tm, depth);
232 }
233
234 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
235 struct dst_entry *dst)
236 {
237 struct tcp_metrics_block *tm;
238 struct inetpeer_addr saddr, daddr;
239 unsigned int hash;
240 struct net *net;
241
242 saddr.family = req->rsk_ops->family;
243 daddr.family = req->rsk_ops->family;
244 switch (daddr.family) {
245 case AF_INET:
246 inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
247 inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
248 hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
249 break;
250 #if IS_ENABLED(CONFIG_IPV6)
251 case AF_INET6:
252 inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
253 inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr);
254 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
255 break;
256 #endif
257 default:
258 return NULL;
259 }
260
261 net = dev_net(dst->dev);
262 hash ^= net_hash_mix(net);
263 hash = hash_32(hash, tcp_metrics_hash_log);
264
265 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
266 tm = rcu_dereference(tm->tcpm_next)) {
267 if (addr_same(&tm->tcpm_saddr, &saddr) &&
268 addr_same(&tm->tcpm_daddr, &daddr) &&
269 net_eq(tm_net(tm), net))
270 break;
271 }
272 tcpm_check_stamp(tm, dst);
273 return tm;
274 }
275
276 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
277 {
278 struct tcp_metrics_block *tm;
279 struct inetpeer_addr saddr, daddr;
280 unsigned int hash;
281 struct net *net;
282
283 if (tw->tw_family == AF_INET) {
284 inetpeer_set_addr_v4(&saddr, tw->tw_rcv_saddr);
285 inetpeer_set_addr_v4(&daddr, tw->tw_daddr);
286 hash = ipv4_addr_hash(tw->tw_daddr);
287 }
288 #if IS_ENABLED(CONFIG_IPV6)
289 else if (tw->tw_family == AF_INET6) {
290 if (ipv6_addr_v4mapped(&tw->tw_v6_daddr)) {
291 inetpeer_set_addr_v4(&saddr, tw->tw_rcv_saddr);
292 inetpeer_set_addr_v4(&daddr, tw->tw_daddr);
293 hash = ipv4_addr_hash(tw->tw_daddr);
294 } else {
295 inetpeer_set_addr_v6(&saddr, &tw->tw_v6_rcv_saddr);
296 inetpeer_set_addr_v6(&daddr, &tw->tw_v6_daddr);
297 hash = ipv6_addr_hash(&tw->tw_v6_daddr);
298 }
299 }
300 #endif
301 else
302 return NULL;
303
304 net = twsk_net(tw);
305 hash ^= net_hash_mix(net);
306 hash = hash_32(hash, tcp_metrics_hash_log);
307
308 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
309 tm = rcu_dereference(tm->tcpm_next)) {
310 if (addr_same(&tm->tcpm_saddr, &saddr) &&
311 addr_same(&tm->tcpm_daddr, &daddr) &&
312 net_eq(tm_net(tm), net))
313 break;
314 }
315 return tm;
316 }
317
318 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
319 struct dst_entry *dst,
320 bool create)
321 {
322 struct tcp_metrics_block *tm;
323 struct inetpeer_addr saddr, daddr;
324 unsigned int hash;
325 struct net *net;
326
327 if (sk->sk_family == AF_INET) {
328 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
329 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
330 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
331 }
332 #if IS_ENABLED(CONFIG_IPV6)
333 else if (sk->sk_family == AF_INET6) {
334 if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
335 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
336 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
337 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
338 } else {
339 inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
340 inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
341 hash = ipv6_addr_hash(&sk->sk_v6_daddr);
342 }
343 }
344 #endif
345 else
346 return NULL;
347
348 net = dev_net(dst->dev);
349 hash ^= net_hash_mix(net);
350 hash = hash_32(hash, tcp_metrics_hash_log);
351
352 tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
353 if (tm == TCP_METRICS_RECLAIM_PTR)
354 tm = NULL;
355 if (!tm && create)
356 tm = tcpm_new(dst, &saddr, &daddr, hash);
357 else
358 tcpm_check_stamp(tm, dst);
359
360 return tm;
361 }
362
363 /* Save metrics learned by this TCP session. This function is called
364 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
365 * or goes from LAST-ACK to CLOSE.
366 */
367 void tcp_update_metrics(struct sock *sk)
368 {
369 const struct inet_connection_sock *icsk = inet_csk(sk);
370 struct dst_entry *dst = __sk_dst_get(sk);
371 struct tcp_sock *tp = tcp_sk(sk);
372 struct net *net = sock_net(sk);
373 struct tcp_metrics_block *tm;
374 unsigned long rtt;
375 u32 val;
376 int m;
377
378 if (sysctl_tcp_nometrics_save || !dst)
379 return;
380
381 if (dst->flags & DST_HOST)
382 dst_confirm(dst);
383
384 rcu_read_lock();
385 if (icsk->icsk_backoff || !tp->srtt_us) {
386 /* This session failed to estimate rtt. Why?
387 * Probably, no packets returned in time. Reset our
388 * results.
389 */
390 tm = tcp_get_metrics(sk, dst, false);
391 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
392 tcp_metric_set(tm, TCP_METRIC_RTT, 0);
393 goto out_unlock;
394 } else
395 tm = tcp_get_metrics(sk, dst, true);
396
397 if (!tm)
398 goto out_unlock;
399
400 rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
401 m = rtt - tp->srtt_us;
402
403 /* If newly calculated rtt larger than stored one, store new
404 * one. Otherwise, use EWMA. Remember, rtt overestimation is
405 * always better than underestimation.
406 */
407 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
408 if (m <= 0)
409 rtt = tp->srtt_us;
410 else
411 rtt -= (m >> 3);
412 tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
413 }
414
415 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
416 unsigned long var;
417
418 if (m < 0)
419 m = -m;
420
421 /* Scale deviation to rttvar fixed point */
422 m >>= 1;
423 if (m < tp->mdev_us)
424 m = tp->mdev_us;
425
426 var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
427 if (m >= var)
428 var = m;
429 else
430 var -= (var - m) >> 2;
431
432 tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
433 }
434
435 if (tcp_in_initial_slowstart(tp)) {
436 /* Slow start still did not finish. */
437 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
438 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
439 if (val && (tp->snd_cwnd >> 1) > val)
440 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
441 tp->snd_cwnd >> 1);
442 }
443 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
444 val = tcp_metric_get(tm, TCP_METRIC_CWND);
445 if (tp->snd_cwnd > val)
446 tcp_metric_set(tm, TCP_METRIC_CWND,
447 tp->snd_cwnd);
448 }
449 } else if (!tcp_in_slow_start(tp) &&
450 icsk->icsk_ca_state == TCP_CA_Open) {
451 /* Cong. avoidance phase, cwnd is reliable. */
452 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
453 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
454 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
455 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
456 val = tcp_metric_get(tm, TCP_METRIC_CWND);
457 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
458 }
459 } else {
460 /* Else slow start did not finish, cwnd is non-sense,
461 * ssthresh may be also invalid.
462 */
463 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
464 val = tcp_metric_get(tm, TCP_METRIC_CWND);
465 tcp_metric_set(tm, TCP_METRIC_CWND,
466 (val + tp->snd_ssthresh) >> 1);
467 }
468 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
469 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
470 if (val && tp->snd_ssthresh > val)
471 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
472 tp->snd_ssthresh);
473 }
474 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
475 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
476 if (val < tp->reordering &&
477 tp->reordering != net->ipv4.sysctl_tcp_reordering)
478 tcp_metric_set(tm, TCP_METRIC_REORDERING,
479 tp->reordering);
480 }
481 }
482 tm->tcpm_stamp = jiffies;
483 out_unlock:
484 rcu_read_unlock();
485 }
486
487 /* Initialize metrics on socket. */
488
489 void tcp_init_metrics(struct sock *sk)
490 {
491 struct dst_entry *dst = __sk_dst_get(sk);
492 struct tcp_sock *tp = tcp_sk(sk);
493 struct tcp_metrics_block *tm;
494 u32 val, crtt = 0; /* cached RTT scaled by 8 */
495
496 if (!dst)
497 goto reset;
498
499 dst_confirm(dst);
500
501 rcu_read_lock();
502 tm = tcp_get_metrics(sk, dst, true);
503 if (!tm) {
504 rcu_read_unlock();
505 goto reset;
506 }
507
508 if (tcp_metric_locked(tm, TCP_METRIC_CWND))
509 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
510
511 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
512 if (val) {
513 tp->snd_ssthresh = val;
514 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
515 tp->snd_ssthresh = tp->snd_cwnd_clamp;
516 } else {
517 /* ssthresh may have been reduced unnecessarily during.
518 * 3WHS. Restore it back to its initial default.
519 */
520 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
521 }
522 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
523 if (val && tp->reordering != val) {
524 tcp_disable_fack(tp);
525 tcp_disable_early_retrans(tp);
526 tp->reordering = val;
527 }
528
529 crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
530 rcu_read_unlock();
531 reset:
532 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
533 * to seed the RTO for later data packets because SYN packets are
534 * small. Use the per-dst cached values to seed the RTO but keep
535 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
536 * Later the RTO will be updated immediately upon obtaining the first
537 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
538 * influences the first RTO but not later RTT estimation.
539 *
540 * But if RTT is not available from the SYN (due to retransmits or
541 * syn cookies) or the cache, force a conservative 3secs timeout.
542 *
543 * A bit of theory. RTT is time passed after "normal" sized packet
544 * is sent until it is ACKed. In normal circumstances sending small
545 * packets force peer to delay ACKs and calculation is correct too.
546 * The algorithm is adaptive and, provided we follow specs, it
547 * NEVER underestimate RTT. BUT! If peer tries to make some clever
548 * tricks sort of "quick acks" for time long enough to decrease RTT
549 * to low value, and then abruptly stops to do it and starts to delay
550 * ACKs, wait for troubles.
551 */
552 if (crtt > tp->srtt_us) {
553 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
554 crtt /= 8 * USEC_PER_SEC / HZ;
555 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
556 } else if (tp->srtt_us == 0) {
557 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
558 * 3WHS. This is most likely due to retransmission,
559 * including spurious one. Reset the RTO back to 3secs
560 * from the more aggressive 1sec to avoid more spurious
561 * retransmission.
562 */
563 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
564 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
565
566 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
567 }
568 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
569 * retransmitted. In light of RFC6298 more aggressive 1sec
570 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
571 * retransmission has occurred.
572 */
573 if (tp->total_retrans > 1)
574 tp->snd_cwnd = 1;
575 else
576 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
577 tp->snd_cwnd_stamp = tcp_time_stamp;
578 }
579
580 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst,
581 bool paws_check, bool timestamps)
582 {
583 struct tcp_metrics_block *tm;
584 bool ret;
585
586 if (!dst)
587 return false;
588
589 rcu_read_lock();
590 tm = __tcp_get_metrics_req(req, dst);
591 if (paws_check) {
592 if (tm &&
593 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
594 ((s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW ||
595 !timestamps))
596 ret = false;
597 else
598 ret = true;
599 } else {
600 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
601 ret = true;
602 else
603 ret = false;
604 }
605 rcu_read_unlock();
606
607 return ret;
608 }
609 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
610
611 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
612 {
613 struct tcp_metrics_block *tm;
614
615 rcu_read_lock();
616 tm = tcp_get_metrics(sk, dst, true);
617 if (tm) {
618 struct tcp_sock *tp = tcp_sk(sk);
619
620 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
621 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
622 tp->rx_opt.ts_recent = tm->tcpm_ts;
623 }
624 }
625 rcu_read_unlock();
626 }
627 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
628
629 /* VJ's idea. Save last timestamp seen from this destination and hold
630 * it at least for normal timewait interval to use for duplicate
631 * segment detection in subsequent connections, before they enter
632 * synchronized state.
633 */
634 bool tcp_remember_stamp(struct sock *sk)
635 {
636 struct dst_entry *dst = __sk_dst_get(sk);
637 bool ret = false;
638
639 if (dst) {
640 struct tcp_metrics_block *tm;
641
642 rcu_read_lock();
643 tm = tcp_get_metrics(sk, dst, true);
644 if (tm) {
645 struct tcp_sock *tp = tcp_sk(sk);
646
647 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
648 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
649 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
650 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
651 tm->tcpm_ts = tp->rx_opt.ts_recent;
652 }
653 ret = true;
654 }
655 rcu_read_unlock();
656 }
657 return ret;
658 }
659
660 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
661 {
662 struct tcp_metrics_block *tm;
663 bool ret = false;
664
665 rcu_read_lock();
666 tm = __tcp_get_metrics_tw(tw);
667 if (tm) {
668 const struct tcp_timewait_sock *tcptw;
669 struct sock *sk = (struct sock *) tw;
670
671 tcptw = tcp_twsk(sk);
672 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
673 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
674 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
675 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
676 tm->tcpm_ts = tcptw->tw_ts_recent;
677 }
678 ret = true;
679 }
680 rcu_read_unlock();
681
682 return ret;
683 }
684
685 static DEFINE_SEQLOCK(fastopen_seqlock);
686
687 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
688 struct tcp_fastopen_cookie *cookie,
689 int *syn_loss, unsigned long *last_syn_loss)
690 {
691 struct tcp_metrics_block *tm;
692
693 rcu_read_lock();
694 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
695 if (tm) {
696 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
697 unsigned int seq;
698
699 do {
700 seq = read_seqbegin(&fastopen_seqlock);
701 if (tfom->mss)
702 *mss = tfom->mss;
703 *cookie = tfom->cookie;
704 if (cookie->len <= 0 && tfom->try_exp == 1)
705 cookie->exp = true;
706 *syn_loss = tfom->syn_loss;
707 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
708 } while (read_seqretry(&fastopen_seqlock, seq));
709 }
710 rcu_read_unlock();
711 }
712
713 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
714 struct tcp_fastopen_cookie *cookie, bool syn_lost,
715 u16 try_exp)
716 {
717 struct dst_entry *dst = __sk_dst_get(sk);
718 struct tcp_metrics_block *tm;
719
720 if (!dst)
721 return;
722 rcu_read_lock();
723 tm = tcp_get_metrics(sk, dst, true);
724 if (tm) {
725 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
726
727 write_seqlock_bh(&fastopen_seqlock);
728 if (mss)
729 tfom->mss = mss;
730 if (cookie && cookie->len > 0)
731 tfom->cookie = *cookie;
732 else if (try_exp > tfom->try_exp &&
733 tfom->cookie.len <= 0 && !tfom->cookie.exp)
734 tfom->try_exp = try_exp;
735 if (syn_lost) {
736 ++tfom->syn_loss;
737 tfom->last_syn_loss = jiffies;
738 } else
739 tfom->syn_loss = 0;
740 write_sequnlock_bh(&fastopen_seqlock);
741 }
742 rcu_read_unlock();
743 }
744
745 static struct genl_family tcp_metrics_nl_family = {
746 .id = GENL_ID_GENERATE,
747 .hdrsize = 0,
748 .name = TCP_METRICS_GENL_NAME,
749 .version = TCP_METRICS_GENL_VERSION,
750 .maxattr = TCP_METRICS_ATTR_MAX,
751 .netnsok = true,
752 };
753
754 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
755 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
756 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
757 .len = sizeof(struct in6_addr), },
758 /* Following attributes are not received for GET/DEL,
759 * we keep them for reference
760 */
761 #if 0
762 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
763 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
764 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
765 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
766 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
767 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
768 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
769 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
770 .len = TCP_FASTOPEN_COOKIE_MAX, },
771 #endif
772 };
773
774 /* Add attributes, caller cancels its header on failure */
775 static int tcp_metrics_fill_info(struct sk_buff *msg,
776 struct tcp_metrics_block *tm)
777 {
778 struct nlattr *nest;
779 int i;
780
781 switch (tm->tcpm_daddr.family) {
782 case AF_INET:
783 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
784 inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
785 goto nla_put_failure;
786 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
787 inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
788 goto nla_put_failure;
789 break;
790 case AF_INET6:
791 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
792 inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
793 goto nla_put_failure;
794 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
795 inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
796 goto nla_put_failure;
797 break;
798 default:
799 return -EAFNOSUPPORT;
800 }
801
802 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
803 jiffies - tm->tcpm_stamp,
804 TCP_METRICS_ATTR_PAD) < 0)
805 goto nla_put_failure;
806 if (tm->tcpm_ts_stamp) {
807 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
808 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
809 goto nla_put_failure;
810 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
811 tm->tcpm_ts) < 0)
812 goto nla_put_failure;
813 }
814
815 {
816 int n = 0;
817
818 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
819 if (!nest)
820 goto nla_put_failure;
821 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
822 u32 val = tm->tcpm_vals[i];
823
824 if (!val)
825 continue;
826 if (i == TCP_METRIC_RTT) {
827 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
828 val) < 0)
829 goto nla_put_failure;
830 n++;
831 val = max(val / 1000, 1U);
832 }
833 if (i == TCP_METRIC_RTTVAR) {
834 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
835 val) < 0)
836 goto nla_put_failure;
837 n++;
838 val = max(val / 1000, 1U);
839 }
840 if (nla_put_u32(msg, i + 1, val) < 0)
841 goto nla_put_failure;
842 n++;
843 }
844 if (n)
845 nla_nest_end(msg, nest);
846 else
847 nla_nest_cancel(msg, nest);
848 }
849
850 {
851 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
852 unsigned int seq;
853
854 do {
855 seq = read_seqbegin(&fastopen_seqlock);
856 tfom_copy[0] = tm->tcpm_fastopen;
857 } while (read_seqretry(&fastopen_seqlock, seq));
858
859 tfom = tfom_copy;
860 if (tfom->mss &&
861 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
862 tfom->mss) < 0)
863 goto nla_put_failure;
864 if (tfom->syn_loss &&
865 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
866 tfom->syn_loss) < 0 ||
867 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
868 jiffies - tfom->last_syn_loss,
869 TCP_METRICS_ATTR_PAD) < 0))
870 goto nla_put_failure;
871 if (tfom->cookie.len > 0 &&
872 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
873 tfom->cookie.len, tfom->cookie.val) < 0)
874 goto nla_put_failure;
875 }
876
877 return 0;
878
879 nla_put_failure:
880 return -EMSGSIZE;
881 }
882
883 static int tcp_metrics_dump_info(struct sk_buff *skb,
884 struct netlink_callback *cb,
885 struct tcp_metrics_block *tm)
886 {
887 void *hdr;
888
889 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
890 &tcp_metrics_nl_family, NLM_F_MULTI,
891 TCP_METRICS_CMD_GET);
892 if (!hdr)
893 return -EMSGSIZE;
894
895 if (tcp_metrics_fill_info(skb, tm) < 0)
896 goto nla_put_failure;
897
898 genlmsg_end(skb, hdr);
899 return 0;
900
901 nla_put_failure:
902 genlmsg_cancel(skb, hdr);
903 return -EMSGSIZE;
904 }
905
906 static int tcp_metrics_nl_dump(struct sk_buff *skb,
907 struct netlink_callback *cb)
908 {
909 struct net *net = sock_net(skb->sk);
910 unsigned int max_rows = 1U << tcp_metrics_hash_log;
911 unsigned int row, s_row = cb->args[0];
912 int s_col = cb->args[1], col = s_col;
913
914 for (row = s_row; row < max_rows; row++, s_col = 0) {
915 struct tcp_metrics_block *tm;
916 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
917
918 rcu_read_lock();
919 for (col = 0, tm = rcu_dereference(hb->chain); tm;
920 tm = rcu_dereference(tm->tcpm_next), col++) {
921 if (!net_eq(tm_net(tm), net))
922 continue;
923 if (col < s_col)
924 continue;
925 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
926 rcu_read_unlock();
927 goto done;
928 }
929 }
930 rcu_read_unlock();
931 }
932
933 done:
934 cb->args[0] = row;
935 cb->args[1] = col;
936 return skb->len;
937 }
938
939 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
940 unsigned int *hash, int optional, int v4, int v6)
941 {
942 struct nlattr *a;
943
944 a = info->attrs[v4];
945 if (a) {
946 inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
947 if (hash)
948 *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
949 return 0;
950 }
951 a = info->attrs[v6];
952 if (a) {
953 struct in6_addr in6;
954
955 if (nla_len(a) != sizeof(struct in6_addr))
956 return -EINVAL;
957 in6 = nla_get_in6_addr(a);
958 inetpeer_set_addr_v6(addr, &in6);
959 if (hash)
960 *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
961 return 0;
962 }
963 return optional ? 1 : -EAFNOSUPPORT;
964 }
965
966 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
967 unsigned int *hash, int optional)
968 {
969 return __parse_nl_addr(info, addr, hash, optional,
970 TCP_METRICS_ATTR_ADDR_IPV4,
971 TCP_METRICS_ATTR_ADDR_IPV6);
972 }
973
974 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
975 {
976 return __parse_nl_addr(info, addr, NULL, 0,
977 TCP_METRICS_ATTR_SADDR_IPV4,
978 TCP_METRICS_ATTR_SADDR_IPV6);
979 }
980
981 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
982 {
983 struct tcp_metrics_block *tm;
984 struct inetpeer_addr saddr, daddr;
985 unsigned int hash;
986 struct sk_buff *msg;
987 struct net *net = genl_info_net(info);
988 void *reply;
989 int ret;
990 bool src = true;
991
992 ret = parse_nl_addr(info, &daddr, &hash, 0);
993 if (ret < 0)
994 return ret;
995
996 ret = parse_nl_saddr(info, &saddr);
997 if (ret < 0)
998 src = false;
999
1000 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1001 if (!msg)
1002 return -ENOMEM;
1003
1004 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
1005 info->genlhdr->cmd);
1006 if (!reply)
1007 goto nla_put_failure;
1008
1009 hash ^= net_hash_mix(net);
1010 hash = hash_32(hash, tcp_metrics_hash_log);
1011 ret = -ESRCH;
1012 rcu_read_lock();
1013 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
1014 tm = rcu_dereference(tm->tcpm_next)) {
1015 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1016 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1017 net_eq(tm_net(tm), net)) {
1018 ret = tcp_metrics_fill_info(msg, tm);
1019 break;
1020 }
1021 }
1022 rcu_read_unlock();
1023 if (ret < 0)
1024 goto out_free;
1025
1026 genlmsg_end(msg, reply);
1027 return genlmsg_reply(msg, info);
1028
1029 nla_put_failure:
1030 ret = -EMSGSIZE;
1031
1032 out_free:
1033 nlmsg_free(msg);
1034 return ret;
1035 }
1036
1037 static void tcp_metrics_flush_all(struct net *net)
1038 {
1039 unsigned int max_rows = 1U << tcp_metrics_hash_log;
1040 struct tcpm_hash_bucket *hb = tcp_metrics_hash;
1041 struct tcp_metrics_block *tm;
1042 unsigned int row;
1043
1044 for (row = 0; row < max_rows; row++, hb++) {
1045 struct tcp_metrics_block __rcu **pp;
1046 spin_lock_bh(&tcp_metrics_lock);
1047 pp = &hb->chain;
1048 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1049 if (net_eq(tm_net(tm), net)) {
1050 *pp = tm->tcpm_next;
1051 kfree_rcu(tm, rcu_head);
1052 } else {
1053 pp = &tm->tcpm_next;
1054 }
1055 }
1056 spin_unlock_bh(&tcp_metrics_lock);
1057 }
1058 }
1059
1060 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
1061 {
1062 struct tcpm_hash_bucket *hb;
1063 struct tcp_metrics_block *tm;
1064 struct tcp_metrics_block __rcu **pp;
1065 struct inetpeer_addr saddr, daddr;
1066 unsigned int hash;
1067 struct net *net = genl_info_net(info);
1068 int ret;
1069 bool src = true, found = false;
1070
1071 ret = parse_nl_addr(info, &daddr, &hash, 1);
1072 if (ret < 0)
1073 return ret;
1074 if (ret > 0) {
1075 tcp_metrics_flush_all(net);
1076 return 0;
1077 }
1078 ret = parse_nl_saddr(info, &saddr);
1079 if (ret < 0)
1080 src = false;
1081
1082 hash ^= net_hash_mix(net);
1083 hash = hash_32(hash, tcp_metrics_hash_log);
1084 hb = tcp_metrics_hash + hash;
1085 pp = &hb->chain;
1086 spin_lock_bh(&tcp_metrics_lock);
1087 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1088 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1089 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1090 net_eq(tm_net(tm), net)) {
1091 *pp = tm->tcpm_next;
1092 kfree_rcu(tm, rcu_head);
1093 found = true;
1094 } else {
1095 pp = &tm->tcpm_next;
1096 }
1097 }
1098 spin_unlock_bh(&tcp_metrics_lock);
1099 if (!found)
1100 return -ESRCH;
1101 return 0;
1102 }
1103
1104 static const struct genl_ops tcp_metrics_nl_ops[] = {
1105 {
1106 .cmd = TCP_METRICS_CMD_GET,
1107 .doit = tcp_metrics_nl_cmd_get,
1108 .dumpit = tcp_metrics_nl_dump,
1109 .policy = tcp_metrics_nl_policy,
1110 },
1111 {
1112 .cmd = TCP_METRICS_CMD_DEL,
1113 .doit = tcp_metrics_nl_cmd_del,
1114 .policy = tcp_metrics_nl_policy,
1115 .flags = GENL_ADMIN_PERM,
1116 },
1117 };
1118
1119 static unsigned int tcpmhash_entries;
1120 static int __init set_tcpmhash_entries(char *str)
1121 {
1122 ssize_t ret;
1123
1124 if (!str)
1125 return 0;
1126
1127 ret = kstrtouint(str, 0, &tcpmhash_entries);
1128 if (ret)
1129 return 0;
1130
1131 return 1;
1132 }
1133 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1134
1135 static int __net_init tcp_net_metrics_init(struct net *net)
1136 {
1137 size_t size;
1138 unsigned int slots;
1139
1140 if (!net_eq(net, &init_net))
1141 return 0;
1142
1143 slots = tcpmhash_entries;
1144 if (!slots) {
1145 if (totalram_pages >= 128 * 1024)
1146 slots = 16 * 1024;
1147 else
1148 slots = 8 * 1024;
1149 }
1150
1151 tcp_metrics_hash_log = order_base_2(slots);
1152 size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1153
1154 tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1155 if (!tcp_metrics_hash)
1156 tcp_metrics_hash = vzalloc(size);
1157
1158 if (!tcp_metrics_hash)
1159 return -ENOMEM;
1160
1161 return 0;
1162 }
1163
1164 static void __net_exit tcp_net_metrics_exit(struct net *net)
1165 {
1166 tcp_metrics_flush_all(net);
1167 }
1168
1169 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1170 .init = tcp_net_metrics_init,
1171 .exit = tcp_net_metrics_exit,
1172 };
1173
1174 void __init tcp_metrics_init(void)
1175 {
1176 int ret;
1177
1178 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1179 if (ret < 0)
1180 panic("Could not allocate the tcp_metrics hash table\n");
1181
1182 ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1183 tcp_metrics_nl_ops);
1184 if (ret < 0)
1185 panic("Could not register tcp_metrics generic netlink\n");
1186 }
Linux with added FPC-III support