Loading...
1#include <linux/rcupdate.h>
2#include <linux/spinlock.h>
3#include <linux/jiffies.h>
4#include <linux/module.h>
5#include <linux/cache.h>
6#include <linux/slab.h>
7#include <linux/init.h>
8#include <linux/tcp.h>
9#include <linux/hash.h>
10#include <linux/tcp_metrics.h>
11#include <linux/vmalloc.h>
12
13#include <net/inet_connection_sock.h>
14#include <net/net_namespace.h>
15#include <net/request_sock.h>
16#include <net/inetpeer.h>
17#include <net/sock.h>
18#include <net/ipv6.h>
19#include <net/dst.h>
20#include <net/tcp.h>
21#include <net/genetlink.h>
22
23int sysctl_tcp_nometrics_save __read_mostly;
24
25static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
26 const struct inetpeer_addr *daddr,
27 struct net *net, unsigned int hash);
28
29struct tcp_fastopen_metrics {
30 u16 mss;
31 u16 syn_loss:10; /* Recurring Fast Open SYN losses */
32 unsigned long last_syn_loss; /* Last Fast Open SYN loss */
33 struct tcp_fastopen_cookie cookie;
34};
35
36/* TCP_METRIC_MAX includes 2 extra fields for userspace compatibility
37 * Kernel only stores RTT and RTTVAR in usec resolution
38 */
39#define TCP_METRIC_MAX_KERNEL (TCP_METRIC_MAX - 2)
40
41struct tcp_metrics_block {
42 struct tcp_metrics_block __rcu *tcpm_next;
43 struct inetpeer_addr tcpm_saddr;
44 struct inetpeer_addr tcpm_daddr;
45 unsigned long tcpm_stamp;
46 u32 tcpm_ts;
47 u32 tcpm_ts_stamp;
48 u32 tcpm_lock;
49 u32 tcpm_vals[TCP_METRIC_MAX_KERNEL + 1];
50 struct tcp_fastopen_metrics tcpm_fastopen;
51
52 struct rcu_head rcu_head;
53};
54
55static bool tcp_metric_locked(struct tcp_metrics_block *tm,
56 enum tcp_metric_index idx)
57{
58 return tm->tcpm_lock & (1 << idx);
59}
60
61static u32 tcp_metric_get(struct tcp_metrics_block *tm,
62 enum tcp_metric_index idx)
63{
64 return tm->tcpm_vals[idx];
65}
66
67static void tcp_metric_set(struct tcp_metrics_block *tm,
68 enum tcp_metric_index idx,
69 u32 val)
70{
71 tm->tcpm_vals[idx] = val;
72}
73
74static bool addr_same(const struct inetpeer_addr *a,
75 const struct inetpeer_addr *b)
76{
77 const struct in6_addr *a6, *b6;
78
79 if (a->family != b->family)
80 return false;
81 if (a->family == AF_INET)
82 return a->addr.a4 == b->addr.a4;
83
84 a6 = (const struct in6_addr *) &a->addr.a6[0];
85 b6 = (const struct in6_addr *) &b->addr.a6[0];
86
87 return ipv6_addr_equal(a6, b6);
88}
89
90struct tcpm_hash_bucket {
91 struct tcp_metrics_block __rcu *chain;
92};
93
94static DEFINE_SPINLOCK(tcp_metrics_lock);
95
96static 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.cookie.len = 0;
132 }
133}
134
135#define TCP_METRICS_TIMEOUT (60 * 60 * HZ)
136
137static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
138{
139 if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
140 tcpm_suck_dst(tm, dst, false);
141}
142
143#define TCP_METRICS_RECLAIM_DEPTH 5
144#define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL
145
146static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
147 struct inetpeer_addr *saddr,
148 struct inetpeer_addr *daddr,
149 unsigned int hash)
150{
151 struct tcp_metrics_block *tm;
152 struct net *net;
153 bool reclaim = false;
154
155 spin_lock_bh(&tcp_metrics_lock);
156 net = dev_net(dst->dev);
157
158 /* While waiting for the spin-lock the cache might have been populated
159 * with this entry and so we have to check again.
160 */
161 tm = __tcp_get_metrics(saddr, daddr, net, hash);
162 if (tm == TCP_METRICS_RECLAIM_PTR) {
163 reclaim = true;
164 tm = NULL;
165 }
166 if (tm) {
167 tcpm_check_stamp(tm, dst);
168 goto out_unlock;
169 }
170
171 if (unlikely(reclaim)) {
172 struct tcp_metrics_block *oldest;
173
174 oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
175 for (tm = rcu_dereference(oldest->tcpm_next); tm;
176 tm = rcu_dereference(tm->tcpm_next)) {
177 if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
178 oldest = tm;
179 }
180 tm = oldest;
181 } else {
182 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
183 if (!tm)
184 goto out_unlock;
185 }
186 tm->tcpm_saddr = *saddr;
187 tm->tcpm_daddr = *daddr;
188
189 tcpm_suck_dst(tm, dst, true);
190
191 if (likely(!reclaim)) {
192 tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
193 rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
194 }
195
196out_unlock:
197 spin_unlock_bh(&tcp_metrics_lock);
198 return tm;
199}
200
201static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
202{
203 if (tm)
204 return tm;
205 if (depth > TCP_METRICS_RECLAIM_DEPTH)
206 return TCP_METRICS_RECLAIM_PTR;
207 return NULL;
208}
209
210static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
211 const struct inetpeer_addr *daddr,
212 struct net *net, unsigned int hash)
213{
214 struct tcp_metrics_block *tm;
215 int depth = 0;
216
217 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
218 tm = rcu_dereference(tm->tcpm_next)) {
219 if (addr_same(&tm->tcpm_saddr, saddr) &&
220 addr_same(&tm->tcpm_daddr, daddr))
221 break;
222 depth++;
223 }
224 return tcp_get_encode(tm, depth);
225}
226
227static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
228 struct dst_entry *dst)
229{
230 struct tcp_metrics_block *tm;
231 struct inetpeer_addr saddr, daddr;
232 unsigned int hash;
233 struct net *net;
234
235 saddr.family = req->rsk_ops->family;
236 daddr.family = req->rsk_ops->family;
237 switch (daddr.family) {
238 case AF_INET:
239 saddr.addr.a4 = inet_rsk(req)->ir_loc_addr;
240 daddr.addr.a4 = inet_rsk(req)->ir_rmt_addr;
241 hash = (__force unsigned int) daddr.addr.a4;
242 break;
243#if IS_ENABLED(CONFIG_IPV6)
244 case AF_INET6:
245 *(struct in6_addr *)saddr.addr.a6 = inet_rsk(req)->ir_v6_loc_addr;
246 *(struct in6_addr *)daddr.addr.a6 = inet_rsk(req)->ir_v6_rmt_addr;
247 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
248 break;
249#endif
250 default:
251 return NULL;
252 }
253
254 net = dev_net(dst->dev);
255 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
256
257 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
258 tm = rcu_dereference(tm->tcpm_next)) {
259 if (addr_same(&tm->tcpm_saddr, &saddr) &&
260 addr_same(&tm->tcpm_daddr, &daddr))
261 break;
262 }
263 tcpm_check_stamp(tm, dst);
264 return tm;
265}
266
267static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
268{
269 struct tcp_metrics_block *tm;
270 struct inetpeer_addr saddr, daddr;
271 unsigned int hash;
272 struct net *net;
273
274 if (tw->tw_family == AF_INET) {
275 saddr.family = AF_INET;
276 saddr.addr.a4 = tw->tw_rcv_saddr;
277 daddr.family = AF_INET;
278 daddr.addr.a4 = tw->tw_daddr;
279 hash = (__force unsigned int) daddr.addr.a4;
280 }
281#if IS_ENABLED(CONFIG_IPV6)
282 else if (tw->tw_family == AF_INET6) {
283 if (ipv6_addr_v4mapped(&tw->tw_v6_daddr)) {
284 saddr.family = AF_INET;
285 saddr.addr.a4 = tw->tw_rcv_saddr;
286 daddr.family = AF_INET;
287 daddr.addr.a4 = tw->tw_daddr;
288 hash = (__force unsigned int) daddr.addr.a4;
289 } else {
290 saddr.family = AF_INET6;
291 *(struct in6_addr *)saddr.addr.a6 = tw->tw_v6_rcv_saddr;
292 daddr.family = AF_INET6;
293 *(struct in6_addr *)daddr.addr.a6 = tw->tw_v6_daddr;
294 hash = ipv6_addr_hash(&tw->tw_v6_daddr);
295 }
296 }
297#endif
298 else
299 return NULL;
300
301 net = twsk_net(tw);
302 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
303
304 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
305 tm = rcu_dereference(tm->tcpm_next)) {
306 if (addr_same(&tm->tcpm_saddr, &saddr) &&
307 addr_same(&tm->tcpm_daddr, &daddr))
308 break;
309 }
310 return tm;
311}
312
313static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
314 struct dst_entry *dst,
315 bool create)
316{
317 struct tcp_metrics_block *tm;
318 struct inetpeer_addr saddr, daddr;
319 unsigned int hash;
320 struct net *net;
321
322 if (sk->sk_family == AF_INET) {
323 saddr.family = AF_INET;
324 saddr.addr.a4 = inet_sk(sk)->inet_saddr;
325 daddr.family = AF_INET;
326 daddr.addr.a4 = inet_sk(sk)->inet_daddr;
327 hash = (__force unsigned int) daddr.addr.a4;
328 }
329#if IS_ENABLED(CONFIG_IPV6)
330 else if (sk->sk_family == AF_INET6) {
331 if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
332 saddr.family = AF_INET;
333 saddr.addr.a4 = inet_sk(sk)->inet_saddr;
334 daddr.family = AF_INET;
335 daddr.addr.a4 = inet_sk(sk)->inet_daddr;
336 hash = (__force unsigned int) daddr.addr.a4;
337 } else {
338 saddr.family = AF_INET6;
339 *(struct in6_addr *)saddr.addr.a6 = sk->sk_v6_rcv_saddr;
340 daddr.family = AF_INET6;
341 *(struct in6_addr *)daddr.addr.a6 = sk->sk_v6_daddr;
342 hash = ipv6_addr_hash(&sk->sk_v6_daddr);
343 }
344 }
345#endif
346 else
347 return NULL;
348
349 net = dev_net(dst->dev);
350 hash = hash_32(hash, net->ipv4.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 */
367void 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 tcp_metrics_block *tm;
373 unsigned long rtt;
374 u32 val;
375 int m;
376
377 if (sysctl_tcp_nometrics_save || !dst)
378 return;
379
380 if (dst->flags & DST_HOST)
381 dst_confirm(dst);
382
383 rcu_read_lock();
384 if (icsk->icsk_backoff || !tp->srtt_us) {
385 /* This session failed to estimate rtt. Why?
386 * Probably, no packets returned in time. Reset our
387 * results.
388 */
389 tm = tcp_get_metrics(sk, dst, false);
390 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
391 tcp_metric_set(tm, TCP_METRIC_RTT, 0);
392 goto out_unlock;
393 } else
394 tm = tcp_get_metrics(sk, dst, true);
395
396 if (!tm)
397 goto out_unlock;
398
399 rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
400 m = rtt - tp->srtt_us;
401
402 /* If newly calculated rtt larger than stored one, store new
403 * one. Otherwise, use EWMA. Remember, rtt overestimation is
404 * always better than underestimation.
405 */
406 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
407 if (m <= 0)
408 rtt = tp->srtt_us;
409 else
410 rtt -= (m >> 3);
411 tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
412 }
413
414 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
415 unsigned long var;
416
417 if (m < 0)
418 m = -m;
419
420 /* Scale deviation to rttvar fixed point */
421 m >>= 1;
422 if (m < tp->mdev_us)
423 m = tp->mdev_us;
424
425 var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
426 if (m >= var)
427 var = m;
428 else
429 var -= (var - m) >> 2;
430
431 tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
432 }
433
434 if (tcp_in_initial_slowstart(tp)) {
435 /* Slow start still did not finish. */
436 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
437 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
438 if (val && (tp->snd_cwnd >> 1) > val)
439 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
440 tp->snd_cwnd >> 1);
441 }
442 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
443 val = tcp_metric_get(tm, TCP_METRIC_CWND);
444 if (tp->snd_cwnd > val)
445 tcp_metric_set(tm, TCP_METRIC_CWND,
446 tp->snd_cwnd);
447 }
448 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
449 icsk->icsk_ca_state == TCP_CA_Open) {
450 /* Cong. avoidance phase, cwnd is reliable. */
451 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
452 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
453 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
454 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
455 val = tcp_metric_get(tm, TCP_METRIC_CWND);
456 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
457 }
458 } else {
459 /* Else slow start did not finish, cwnd is non-sense,
460 * ssthresh may be also invalid.
461 */
462 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
463 val = tcp_metric_get(tm, TCP_METRIC_CWND);
464 tcp_metric_set(tm, TCP_METRIC_CWND,
465 (val + tp->snd_ssthresh) >> 1);
466 }
467 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
468 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
469 if (val && tp->snd_ssthresh > val)
470 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
471 tp->snd_ssthresh);
472 }
473 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
474 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
475 if (val < tp->reordering &&
476 tp->reordering != sysctl_tcp_reordering)
477 tcp_metric_set(tm, TCP_METRIC_REORDERING,
478 tp->reordering);
479 }
480 }
481 tm->tcpm_stamp = jiffies;
482out_unlock:
483 rcu_read_unlock();
484}
485
486/* Initialize metrics on socket. */
487
488void tcp_init_metrics(struct sock *sk)
489{
490 struct dst_entry *dst = __sk_dst_get(sk);
491 struct tcp_sock *tp = tcp_sk(sk);
492 struct tcp_metrics_block *tm;
493 u32 val, crtt = 0; /* cached RTT scaled by 8 */
494
495 if (dst == NULL)
496 goto reset;
497
498 dst_confirm(dst);
499
500 rcu_read_lock();
501 tm = tcp_get_metrics(sk, dst, true);
502 if (!tm) {
503 rcu_read_unlock();
504 goto reset;
505 }
506
507 if (tcp_metric_locked(tm, TCP_METRIC_CWND))
508 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
509
510 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
511 if (val) {
512 tp->snd_ssthresh = val;
513 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
514 tp->snd_ssthresh = tp->snd_cwnd_clamp;
515 } else {
516 /* ssthresh may have been reduced unnecessarily during.
517 * 3WHS. Restore it back to its initial default.
518 */
519 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
520 }
521 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
522 if (val && tp->reordering != val) {
523 tcp_disable_fack(tp);
524 tcp_disable_early_retrans(tp);
525 tp->reordering = val;
526 }
527
528 crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
529 rcu_read_unlock();
530reset:
531 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
532 * to seed the RTO for later data packets because SYN packets are
533 * small. Use the per-dst cached values to seed the RTO but keep
534 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
535 * Later the RTO will be updated immediately upon obtaining the first
536 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
537 * influences the first RTO but not later RTT estimation.
538 *
539 * But if RTT is not available from the SYN (due to retransmits or
540 * syn cookies) or the cache, force a conservative 3secs timeout.
541 *
542 * A bit of theory. RTT is time passed after "normal" sized packet
543 * is sent until it is ACKed. In normal circumstances sending small
544 * packets force peer to delay ACKs and calculation is correct too.
545 * The algorithm is adaptive and, provided we follow specs, it
546 * NEVER underestimate RTT. BUT! If peer tries to make some clever
547 * tricks sort of "quick acks" for time long enough to decrease RTT
548 * to low value, and then abruptly stops to do it and starts to delay
549 * ACKs, wait for troubles.
550 */
551 if (crtt > tp->srtt_us) {
552 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
553 crtt /= 8 * USEC_PER_MSEC;
554 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
555 } else if (tp->srtt_us == 0) {
556 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
557 * 3WHS. This is most likely due to retransmission,
558 * including spurious one. Reset the RTO back to 3secs
559 * from the more aggressive 1sec to avoid more spurious
560 * retransmission.
561 */
562 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
563 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
564
565 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
566 }
567 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
568 * retransmitted. In light of RFC6298 more aggressive 1sec
569 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
570 * retransmission has occurred.
571 */
572 if (tp->total_retrans > 1)
573 tp->snd_cwnd = 1;
574 else
575 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
576 tp->snd_cwnd_stamp = tcp_time_stamp;
577}
578
579bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
580{
581 struct tcp_metrics_block *tm;
582 bool ret;
583
584 if (!dst)
585 return false;
586
587 rcu_read_lock();
588 tm = __tcp_get_metrics_req(req, dst);
589 if (paws_check) {
590 if (tm &&
591 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
592 (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
593 ret = false;
594 else
595 ret = true;
596 } else {
597 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
598 ret = true;
599 else
600 ret = false;
601 }
602 rcu_read_unlock();
603
604 return ret;
605}
606EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
607
608void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
609{
610 struct tcp_metrics_block *tm;
611
612 rcu_read_lock();
613 tm = tcp_get_metrics(sk, dst, true);
614 if (tm) {
615 struct tcp_sock *tp = tcp_sk(sk);
616
617 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
618 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
619 tp->rx_opt.ts_recent = tm->tcpm_ts;
620 }
621 }
622 rcu_read_unlock();
623}
624EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
625
626/* VJ's idea. Save last timestamp seen from this destination and hold
627 * it at least for normal timewait interval to use for duplicate
628 * segment detection in subsequent connections, before they enter
629 * synchronized state.
630 */
631bool tcp_remember_stamp(struct sock *sk)
632{
633 struct dst_entry *dst = __sk_dst_get(sk);
634 bool ret = false;
635
636 if (dst) {
637 struct tcp_metrics_block *tm;
638
639 rcu_read_lock();
640 tm = tcp_get_metrics(sk, dst, true);
641 if (tm) {
642 struct tcp_sock *tp = tcp_sk(sk);
643
644 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
645 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
646 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
647 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
648 tm->tcpm_ts = tp->rx_opt.ts_recent;
649 }
650 ret = true;
651 }
652 rcu_read_unlock();
653 }
654 return ret;
655}
656
657bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
658{
659 struct tcp_metrics_block *tm;
660 bool ret = false;
661
662 rcu_read_lock();
663 tm = __tcp_get_metrics_tw(tw);
664 if (tm) {
665 const struct tcp_timewait_sock *tcptw;
666 struct sock *sk = (struct sock *) tw;
667
668 tcptw = tcp_twsk(sk);
669 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
670 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
671 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
672 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
673 tm->tcpm_ts = tcptw->tw_ts_recent;
674 }
675 ret = true;
676 }
677 rcu_read_unlock();
678
679 return ret;
680}
681
682static DEFINE_SEQLOCK(fastopen_seqlock);
683
684void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
685 struct tcp_fastopen_cookie *cookie,
686 int *syn_loss, unsigned long *last_syn_loss)
687{
688 struct tcp_metrics_block *tm;
689
690 rcu_read_lock();
691 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
692 if (tm) {
693 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
694 unsigned int seq;
695
696 do {
697 seq = read_seqbegin(&fastopen_seqlock);
698 if (tfom->mss)
699 *mss = tfom->mss;
700 *cookie = tfom->cookie;
701 *syn_loss = tfom->syn_loss;
702 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
703 } while (read_seqretry(&fastopen_seqlock, seq));
704 }
705 rcu_read_unlock();
706}
707
708void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
709 struct tcp_fastopen_cookie *cookie, bool syn_lost)
710{
711 struct dst_entry *dst = __sk_dst_get(sk);
712 struct tcp_metrics_block *tm;
713
714 if (!dst)
715 return;
716 rcu_read_lock();
717 tm = tcp_get_metrics(sk, dst, true);
718 if (tm) {
719 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
720
721 write_seqlock_bh(&fastopen_seqlock);
722 if (mss)
723 tfom->mss = mss;
724 if (cookie && cookie->len > 0)
725 tfom->cookie = *cookie;
726 if (syn_lost) {
727 ++tfom->syn_loss;
728 tfom->last_syn_loss = jiffies;
729 } else
730 tfom->syn_loss = 0;
731 write_sequnlock_bh(&fastopen_seqlock);
732 }
733 rcu_read_unlock();
734}
735
736static struct genl_family tcp_metrics_nl_family = {
737 .id = GENL_ID_GENERATE,
738 .hdrsize = 0,
739 .name = TCP_METRICS_GENL_NAME,
740 .version = TCP_METRICS_GENL_VERSION,
741 .maxattr = TCP_METRICS_ATTR_MAX,
742 .netnsok = true,
743};
744
745static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
746 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
747 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
748 .len = sizeof(struct in6_addr), },
749 /* Following attributes are not received for GET/DEL,
750 * we keep them for reference
751 */
752#if 0
753 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
754 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
755 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
756 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
757 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
758 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
759 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
760 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
761 .len = TCP_FASTOPEN_COOKIE_MAX, },
762#endif
763};
764
765/* Add attributes, caller cancels its header on failure */
766static int tcp_metrics_fill_info(struct sk_buff *msg,
767 struct tcp_metrics_block *tm)
768{
769 struct nlattr *nest;
770 int i;
771
772 switch (tm->tcpm_daddr.family) {
773 case AF_INET:
774 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
775 tm->tcpm_daddr.addr.a4) < 0)
776 goto nla_put_failure;
777 if (nla_put_be32(msg, TCP_METRICS_ATTR_SADDR_IPV4,
778 tm->tcpm_saddr.addr.a4) < 0)
779 goto nla_put_failure;
780 break;
781 case AF_INET6:
782 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
783 tm->tcpm_daddr.addr.a6) < 0)
784 goto nla_put_failure;
785 if (nla_put(msg, TCP_METRICS_ATTR_SADDR_IPV6, 16,
786 tm->tcpm_saddr.addr.a6) < 0)
787 goto nla_put_failure;
788 break;
789 default:
790 return -EAFNOSUPPORT;
791 }
792
793 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
794 jiffies - tm->tcpm_stamp) < 0)
795 goto nla_put_failure;
796 if (tm->tcpm_ts_stamp) {
797 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
798 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
799 goto nla_put_failure;
800 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
801 tm->tcpm_ts) < 0)
802 goto nla_put_failure;
803 }
804
805 {
806 int n = 0;
807
808 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
809 if (!nest)
810 goto nla_put_failure;
811 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
812 u32 val = tm->tcpm_vals[i];
813
814 if (!val)
815 continue;
816 if (i == TCP_METRIC_RTT) {
817 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
818 val) < 0)
819 goto nla_put_failure;
820 n++;
821 val = max(val / 1000, 1U);
822 }
823 if (i == TCP_METRIC_RTTVAR) {
824 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
825 val) < 0)
826 goto nla_put_failure;
827 n++;
828 val = max(val / 1000, 1U);
829 }
830 if (nla_put_u32(msg, i + 1, val) < 0)
831 goto nla_put_failure;
832 n++;
833 }
834 if (n)
835 nla_nest_end(msg, nest);
836 else
837 nla_nest_cancel(msg, nest);
838 }
839
840 {
841 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
842 unsigned int seq;
843
844 do {
845 seq = read_seqbegin(&fastopen_seqlock);
846 tfom_copy[0] = tm->tcpm_fastopen;
847 } while (read_seqretry(&fastopen_seqlock, seq));
848
849 tfom = tfom_copy;
850 if (tfom->mss &&
851 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
852 tfom->mss) < 0)
853 goto nla_put_failure;
854 if (tfom->syn_loss &&
855 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
856 tfom->syn_loss) < 0 ||
857 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
858 jiffies - tfom->last_syn_loss) < 0))
859 goto nla_put_failure;
860 if (tfom->cookie.len > 0 &&
861 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
862 tfom->cookie.len, tfom->cookie.val) < 0)
863 goto nla_put_failure;
864 }
865
866 return 0;
867
868nla_put_failure:
869 return -EMSGSIZE;
870}
871
872static int tcp_metrics_dump_info(struct sk_buff *skb,
873 struct netlink_callback *cb,
874 struct tcp_metrics_block *tm)
875{
876 void *hdr;
877
878 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
879 &tcp_metrics_nl_family, NLM_F_MULTI,
880 TCP_METRICS_CMD_GET);
881 if (!hdr)
882 return -EMSGSIZE;
883
884 if (tcp_metrics_fill_info(skb, tm) < 0)
885 goto nla_put_failure;
886
887 return genlmsg_end(skb, hdr);
888
889nla_put_failure:
890 genlmsg_cancel(skb, hdr);
891 return -EMSGSIZE;
892}
893
894static int tcp_metrics_nl_dump(struct sk_buff *skb,
895 struct netlink_callback *cb)
896{
897 struct net *net = sock_net(skb->sk);
898 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
899 unsigned int row, s_row = cb->args[0];
900 int s_col = cb->args[1], col = s_col;
901
902 for (row = s_row; row < max_rows; row++, s_col = 0) {
903 struct tcp_metrics_block *tm;
904 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
905
906 rcu_read_lock();
907 for (col = 0, tm = rcu_dereference(hb->chain); tm;
908 tm = rcu_dereference(tm->tcpm_next), col++) {
909 if (col < s_col)
910 continue;
911 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
912 rcu_read_unlock();
913 goto done;
914 }
915 }
916 rcu_read_unlock();
917 }
918
919done:
920 cb->args[0] = row;
921 cb->args[1] = col;
922 return skb->len;
923}
924
925static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
926 unsigned int *hash, int optional, int v4, int v6)
927{
928 struct nlattr *a;
929
930 a = info->attrs[v4];
931 if (a) {
932 addr->family = AF_INET;
933 addr->addr.a4 = nla_get_be32(a);
934 if (hash)
935 *hash = (__force unsigned int) addr->addr.a4;
936 return 0;
937 }
938 a = info->attrs[v6];
939 if (a) {
940 if (nla_len(a) != sizeof(struct in6_addr))
941 return -EINVAL;
942 addr->family = AF_INET6;
943 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
944 if (hash)
945 *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
946 return 0;
947 }
948 return optional ? 1 : -EAFNOSUPPORT;
949}
950
951static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
952 unsigned int *hash, int optional)
953{
954 return __parse_nl_addr(info, addr, hash, optional,
955 TCP_METRICS_ATTR_ADDR_IPV4,
956 TCP_METRICS_ATTR_ADDR_IPV6);
957}
958
959static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
960{
961 return __parse_nl_addr(info, addr, NULL, 0,
962 TCP_METRICS_ATTR_SADDR_IPV4,
963 TCP_METRICS_ATTR_SADDR_IPV6);
964}
965
966static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
967{
968 struct tcp_metrics_block *tm;
969 struct inetpeer_addr saddr, daddr;
970 unsigned int hash;
971 struct sk_buff *msg;
972 struct net *net = genl_info_net(info);
973 void *reply;
974 int ret;
975 bool src = true;
976
977 ret = parse_nl_addr(info, &daddr, &hash, 0);
978 if (ret < 0)
979 return ret;
980
981 ret = parse_nl_saddr(info, &saddr);
982 if (ret < 0)
983 src = false;
984
985 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
986 if (!msg)
987 return -ENOMEM;
988
989 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
990 info->genlhdr->cmd);
991 if (!reply)
992 goto nla_put_failure;
993
994 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
995 ret = -ESRCH;
996 rcu_read_lock();
997 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
998 tm = rcu_dereference(tm->tcpm_next)) {
999 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1000 (!src || addr_same(&tm->tcpm_saddr, &saddr))) {
1001 ret = tcp_metrics_fill_info(msg, tm);
1002 break;
1003 }
1004 }
1005 rcu_read_unlock();
1006 if (ret < 0)
1007 goto out_free;
1008
1009 genlmsg_end(msg, reply);
1010 return genlmsg_reply(msg, info);
1011
1012nla_put_failure:
1013 ret = -EMSGSIZE;
1014
1015out_free:
1016 nlmsg_free(msg);
1017 return ret;
1018}
1019
1020#define deref_locked_genl(p) \
1021 rcu_dereference_protected(p, lockdep_genl_is_held() && \
1022 lockdep_is_held(&tcp_metrics_lock))
1023
1024#define deref_genl(p) rcu_dereference_protected(p, lockdep_genl_is_held())
1025
1026static int tcp_metrics_flush_all(struct net *net)
1027{
1028 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
1029 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
1030 struct tcp_metrics_block *tm;
1031 unsigned int row;
1032
1033 for (row = 0; row < max_rows; row++, hb++) {
1034 spin_lock_bh(&tcp_metrics_lock);
1035 tm = deref_locked_genl(hb->chain);
1036 if (tm)
1037 hb->chain = NULL;
1038 spin_unlock_bh(&tcp_metrics_lock);
1039 while (tm) {
1040 struct tcp_metrics_block *next;
1041
1042 next = deref_genl(tm->tcpm_next);
1043 kfree_rcu(tm, rcu_head);
1044 tm = next;
1045 }
1046 }
1047 return 0;
1048}
1049
1050static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
1051{
1052 struct tcpm_hash_bucket *hb;
1053 struct tcp_metrics_block *tm;
1054 struct tcp_metrics_block __rcu **pp;
1055 struct inetpeer_addr saddr, daddr;
1056 unsigned int hash;
1057 struct net *net = genl_info_net(info);
1058 int ret;
1059 bool src = true, found = false;
1060
1061 ret = parse_nl_addr(info, &daddr, &hash, 1);
1062 if (ret < 0)
1063 return ret;
1064 if (ret > 0)
1065 return tcp_metrics_flush_all(net);
1066 ret = parse_nl_saddr(info, &saddr);
1067 if (ret < 0)
1068 src = false;
1069
1070 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
1071 hb = net->ipv4.tcp_metrics_hash + hash;
1072 pp = &hb->chain;
1073 spin_lock_bh(&tcp_metrics_lock);
1074 for (tm = deref_locked_genl(*pp); tm; tm = deref_locked_genl(*pp)) {
1075 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1076 (!src || addr_same(&tm->tcpm_saddr, &saddr))) {
1077 *pp = tm->tcpm_next;
1078 kfree_rcu(tm, rcu_head);
1079 found = true;
1080 } else {
1081 pp = &tm->tcpm_next;
1082 }
1083 }
1084 spin_unlock_bh(&tcp_metrics_lock);
1085 if (!found)
1086 return -ESRCH;
1087 return 0;
1088}
1089
1090static const struct genl_ops tcp_metrics_nl_ops[] = {
1091 {
1092 .cmd = TCP_METRICS_CMD_GET,
1093 .doit = tcp_metrics_nl_cmd_get,
1094 .dumpit = tcp_metrics_nl_dump,
1095 .policy = tcp_metrics_nl_policy,
1096 .flags = GENL_ADMIN_PERM,
1097 },
1098 {
1099 .cmd = TCP_METRICS_CMD_DEL,
1100 .doit = tcp_metrics_nl_cmd_del,
1101 .policy = tcp_metrics_nl_policy,
1102 .flags = GENL_ADMIN_PERM,
1103 },
1104};
1105
1106static unsigned int tcpmhash_entries;
1107static int __init set_tcpmhash_entries(char *str)
1108{
1109 ssize_t ret;
1110
1111 if (!str)
1112 return 0;
1113
1114 ret = kstrtouint(str, 0, &tcpmhash_entries);
1115 if (ret)
1116 return 0;
1117
1118 return 1;
1119}
1120__setup("tcpmhash_entries=", set_tcpmhash_entries);
1121
1122static int __net_init tcp_net_metrics_init(struct net *net)
1123{
1124 size_t size;
1125 unsigned int slots;
1126
1127 slots = tcpmhash_entries;
1128 if (!slots) {
1129 if (totalram_pages >= 128 * 1024)
1130 slots = 16 * 1024;
1131 else
1132 slots = 8 * 1024;
1133 }
1134
1135 net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1136 size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
1137
1138 net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1139 if (!net->ipv4.tcp_metrics_hash)
1140 net->ipv4.tcp_metrics_hash = vzalloc(size);
1141
1142 if (!net->ipv4.tcp_metrics_hash)
1143 return -ENOMEM;
1144
1145 return 0;
1146}
1147
1148static void __net_exit tcp_net_metrics_exit(struct net *net)
1149{
1150 unsigned int i;
1151
1152 for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1153 struct tcp_metrics_block *tm, *next;
1154
1155 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1156 while (tm) {
1157 next = rcu_dereference_protected(tm->tcpm_next, 1);
1158 kfree(tm);
1159 tm = next;
1160 }
1161 }
1162 if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1163 vfree(net->ipv4.tcp_metrics_hash);
1164 else
1165 kfree(net->ipv4.tcp_metrics_hash);
1166}
1167
1168static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1169 .init = tcp_net_metrics_init,
1170 .exit = tcp_net_metrics_exit,
1171};
1172
1173void __init tcp_metrics_init(void)
1174{
1175 int ret;
1176
1177 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1178 if (ret < 0)
1179 goto cleanup;
1180 ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1181 tcp_metrics_nl_ops);
1182 if (ret < 0)
1183 goto cleanup_subsys;
1184 return;
1185
1186cleanup_subsys:
1187 unregister_pernet_subsys(&tcp_net_metrics_ops);
1188
1189cleanup:
1190 return;
1191}
1// SPDX-License-Identifier: GPL-2.0
2#include <linux/rcupdate.h>
3#include <linux/spinlock.h>
4#include <linux/jiffies.h>
5#include <linux/module.h>
6#include <linux/cache.h>
7#include <linux/slab.h>
8#include <linux/init.h>
9#include <linux/tcp.h>
10#include <linux/hash.h>
11#include <linux/tcp_metrics.h>
12#include <linux/vmalloc.h>
13
14#include <net/inet_connection_sock.h>
15#include <net/net_namespace.h>
16#include <net/request_sock.h>
17#include <net/inetpeer.h>
18#include <net/sock.h>
19#include <net/ipv6.h>
20#include <net/dst.h>
21#include <net/tcp.h>
22#include <net/genetlink.h>
23
24static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
25 const struct inetpeer_addr *daddr,
26 struct net *net, unsigned int hash);
27
28struct tcp_fastopen_metrics {
29 u16 mss;
30 u16 syn_loss:10, /* Recurring Fast Open SYN losses */
31 try_exp:2; /* Request w/ exp. option (once) */
32 unsigned long last_syn_loss; /* Last Fast Open SYN loss */
33 struct tcp_fastopen_cookie cookie;
34};
35
36/* TCP_METRIC_MAX includes 2 extra fields for userspace compatibility
37 * Kernel only stores RTT and RTTVAR in usec resolution
38 */
39#define TCP_METRIC_MAX_KERNEL (TCP_METRIC_MAX - 2)
40
41struct tcp_metrics_block {
42 struct tcp_metrics_block __rcu *tcpm_next;
43 struct net *tcpm_net;
44 struct inetpeer_addr tcpm_saddr;
45 struct inetpeer_addr tcpm_daddr;
46 unsigned long tcpm_stamp;
47 u32 tcpm_lock;
48 u32 tcpm_vals[TCP_METRIC_MAX_KERNEL + 1];
49 struct tcp_fastopen_metrics tcpm_fastopen;
50
51 struct rcu_head rcu_head;
52};
53
54static inline struct net *tm_net(const struct tcp_metrics_block *tm)
55{
56 /* Paired with the WRITE_ONCE() in tcpm_new() */
57 return READ_ONCE(tm->tcpm_net);
58}
59
60static bool tcp_metric_locked(struct tcp_metrics_block *tm,
61 enum tcp_metric_index idx)
62{
63 /* Paired with WRITE_ONCE() in tcpm_suck_dst() */
64 return READ_ONCE(tm->tcpm_lock) & (1 << idx);
65}
66
67static u32 tcp_metric_get(const struct tcp_metrics_block *tm,
68 enum tcp_metric_index idx)
69{
70 /* Paired with WRITE_ONCE() in tcp_metric_set() */
71 return READ_ONCE(tm->tcpm_vals[idx]);
72}
73
74static void tcp_metric_set(struct tcp_metrics_block *tm,
75 enum tcp_metric_index idx,
76 u32 val)
77{
78 /* Paired with READ_ONCE() in tcp_metric_get() */
79 WRITE_ONCE(tm->tcpm_vals[idx], val);
80}
81
82static bool addr_same(const struct inetpeer_addr *a,
83 const struct inetpeer_addr *b)
84{
85 return (a->family == b->family) && !inetpeer_addr_cmp(a, b);
86}
87
88struct tcpm_hash_bucket {
89 struct tcp_metrics_block __rcu *chain;
90};
91
92static struct tcpm_hash_bucket *tcp_metrics_hash __read_mostly;
93static unsigned int tcp_metrics_hash_log __read_mostly;
94
95static DEFINE_SPINLOCK(tcp_metrics_lock);
96static DEFINE_SEQLOCK(fastopen_seqlock);
97
98static void tcpm_suck_dst(struct tcp_metrics_block *tm,
99 const struct dst_entry *dst,
100 bool fastopen_clear)
101{
102 u32 msval;
103 u32 val;
104
105 WRITE_ONCE(tm->tcpm_stamp, jiffies);
106
107 val = 0;
108 if (dst_metric_locked(dst, RTAX_RTT))
109 val |= 1 << TCP_METRIC_RTT;
110 if (dst_metric_locked(dst, RTAX_RTTVAR))
111 val |= 1 << TCP_METRIC_RTTVAR;
112 if (dst_metric_locked(dst, RTAX_SSTHRESH))
113 val |= 1 << TCP_METRIC_SSTHRESH;
114 if (dst_metric_locked(dst, RTAX_CWND))
115 val |= 1 << TCP_METRIC_CWND;
116 if (dst_metric_locked(dst, RTAX_REORDERING))
117 val |= 1 << TCP_METRIC_REORDERING;
118 /* Paired with READ_ONCE() in tcp_metric_locked() */
119 WRITE_ONCE(tm->tcpm_lock, val);
120
121 msval = dst_metric_raw(dst, RTAX_RTT);
122 tcp_metric_set(tm, TCP_METRIC_RTT, msval * USEC_PER_MSEC);
123
124 msval = dst_metric_raw(dst, RTAX_RTTVAR);
125 tcp_metric_set(tm, TCP_METRIC_RTTVAR, msval * USEC_PER_MSEC);
126 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
127 dst_metric_raw(dst, RTAX_SSTHRESH));
128 tcp_metric_set(tm, TCP_METRIC_CWND,
129 dst_metric_raw(dst, RTAX_CWND));
130 tcp_metric_set(tm, TCP_METRIC_REORDERING,
131 dst_metric_raw(dst, RTAX_REORDERING));
132 if (fastopen_clear) {
133 write_seqlock(&fastopen_seqlock);
134 tm->tcpm_fastopen.mss = 0;
135 tm->tcpm_fastopen.syn_loss = 0;
136 tm->tcpm_fastopen.try_exp = 0;
137 tm->tcpm_fastopen.cookie.exp = false;
138 tm->tcpm_fastopen.cookie.len = 0;
139 write_sequnlock(&fastopen_seqlock);
140 }
141}
142
143#define TCP_METRICS_TIMEOUT (60 * 60 * HZ)
144
145static void tcpm_check_stamp(struct tcp_metrics_block *tm,
146 const struct dst_entry *dst)
147{
148 unsigned long limit;
149
150 if (!tm)
151 return;
152 limit = READ_ONCE(tm->tcpm_stamp) + TCP_METRICS_TIMEOUT;
153 if (unlikely(time_after(jiffies, limit)))
154 tcpm_suck_dst(tm, dst, false);
155}
156
157#define TCP_METRICS_RECLAIM_DEPTH 5
158#define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL
159
160#define deref_locked(p) \
161 rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
162
163static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
164 struct inetpeer_addr *saddr,
165 struct inetpeer_addr *daddr,
166 unsigned int hash)
167{
168 struct tcp_metrics_block *tm;
169 struct net *net;
170 bool reclaim = false;
171
172 spin_lock_bh(&tcp_metrics_lock);
173 net = dev_net(dst->dev);
174
175 /* While waiting for the spin-lock the cache might have been populated
176 * with this entry and so we have to check again.
177 */
178 tm = __tcp_get_metrics(saddr, daddr, net, hash);
179 if (tm == TCP_METRICS_RECLAIM_PTR) {
180 reclaim = true;
181 tm = NULL;
182 }
183 if (tm) {
184 tcpm_check_stamp(tm, dst);
185 goto out_unlock;
186 }
187
188 if (unlikely(reclaim)) {
189 struct tcp_metrics_block *oldest;
190
191 oldest = deref_locked(tcp_metrics_hash[hash].chain);
192 for (tm = deref_locked(oldest->tcpm_next); tm;
193 tm = deref_locked(tm->tcpm_next)) {
194 if (time_before(READ_ONCE(tm->tcpm_stamp),
195 READ_ONCE(oldest->tcpm_stamp)))
196 oldest = tm;
197 }
198 tm = oldest;
199 } else {
200 tm = kzalloc(sizeof(*tm), GFP_ATOMIC);
201 if (!tm)
202 goto out_unlock;
203 }
204 /* Paired with the READ_ONCE() in tm_net() */
205 WRITE_ONCE(tm->tcpm_net, net);
206
207 tm->tcpm_saddr = *saddr;
208 tm->tcpm_daddr = *daddr;
209
210 tcpm_suck_dst(tm, dst, reclaim);
211
212 if (likely(!reclaim)) {
213 tm->tcpm_next = tcp_metrics_hash[hash].chain;
214 rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
215 }
216
217out_unlock:
218 spin_unlock_bh(&tcp_metrics_lock);
219 return tm;
220}
221
222static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
223{
224 if (tm)
225 return tm;
226 if (depth > TCP_METRICS_RECLAIM_DEPTH)
227 return TCP_METRICS_RECLAIM_PTR;
228 return NULL;
229}
230
231static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
232 const struct inetpeer_addr *daddr,
233 struct net *net, unsigned int hash)
234{
235 struct tcp_metrics_block *tm;
236 int depth = 0;
237
238 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
239 tm = rcu_dereference(tm->tcpm_next)) {
240 if (addr_same(&tm->tcpm_saddr, saddr) &&
241 addr_same(&tm->tcpm_daddr, daddr) &&
242 net_eq(tm_net(tm), net))
243 break;
244 depth++;
245 }
246 return tcp_get_encode(tm, depth);
247}
248
249static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
250 struct dst_entry *dst)
251{
252 struct tcp_metrics_block *tm;
253 struct inetpeer_addr saddr, daddr;
254 unsigned int hash;
255 struct net *net;
256
257 saddr.family = req->rsk_ops->family;
258 daddr.family = req->rsk_ops->family;
259 switch (daddr.family) {
260 case AF_INET:
261 inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
262 inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
263 hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
264 break;
265#if IS_ENABLED(CONFIG_IPV6)
266 case AF_INET6:
267 inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
268 inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr);
269 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
270 break;
271#endif
272 default:
273 return NULL;
274 }
275
276 net = dev_net(dst->dev);
277 hash ^= net_hash_mix(net);
278 hash = hash_32(hash, tcp_metrics_hash_log);
279
280 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
281 tm = rcu_dereference(tm->tcpm_next)) {
282 if (addr_same(&tm->tcpm_saddr, &saddr) &&
283 addr_same(&tm->tcpm_daddr, &daddr) &&
284 net_eq(tm_net(tm), net))
285 break;
286 }
287 tcpm_check_stamp(tm, dst);
288 return tm;
289}
290
291static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
292 struct dst_entry *dst,
293 bool create)
294{
295 struct tcp_metrics_block *tm;
296 struct inetpeer_addr saddr, daddr;
297 unsigned int hash;
298 struct net *net;
299
300 if (sk->sk_family == AF_INET) {
301 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
302 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
303 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
304 }
305#if IS_ENABLED(CONFIG_IPV6)
306 else if (sk->sk_family == AF_INET6) {
307 if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
308 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
309 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
310 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
311 } else {
312 inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
313 inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
314 hash = ipv6_addr_hash(&sk->sk_v6_daddr);
315 }
316 }
317#endif
318 else
319 return NULL;
320
321 net = dev_net(dst->dev);
322 hash ^= net_hash_mix(net);
323 hash = hash_32(hash, tcp_metrics_hash_log);
324
325 tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
326 if (tm == TCP_METRICS_RECLAIM_PTR)
327 tm = NULL;
328 if (!tm && create)
329 tm = tcpm_new(dst, &saddr, &daddr, hash);
330 else
331 tcpm_check_stamp(tm, dst);
332
333 return tm;
334}
335
336/* Save metrics learned by this TCP session. This function is called
337 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
338 * or goes from LAST-ACK to CLOSE.
339 */
340void tcp_update_metrics(struct sock *sk)
341{
342 const struct inet_connection_sock *icsk = inet_csk(sk);
343 struct dst_entry *dst = __sk_dst_get(sk);
344 struct tcp_sock *tp = tcp_sk(sk);
345 struct net *net = sock_net(sk);
346 struct tcp_metrics_block *tm;
347 unsigned long rtt;
348 u32 val;
349 int m;
350
351 sk_dst_confirm(sk);
352 if (READ_ONCE(net->ipv4.sysctl_tcp_nometrics_save) || !dst)
353 return;
354
355 rcu_read_lock();
356 if (icsk->icsk_backoff || !tp->srtt_us) {
357 /* This session failed to estimate rtt. Why?
358 * Probably, no packets returned in time. Reset our
359 * results.
360 */
361 tm = tcp_get_metrics(sk, dst, false);
362 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
363 tcp_metric_set(tm, TCP_METRIC_RTT, 0);
364 goto out_unlock;
365 } else
366 tm = tcp_get_metrics(sk, dst, true);
367
368 if (!tm)
369 goto out_unlock;
370
371 rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
372 m = rtt - tp->srtt_us;
373
374 /* If newly calculated rtt larger than stored one, store new
375 * one. Otherwise, use EWMA. Remember, rtt overestimation is
376 * always better than underestimation.
377 */
378 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
379 if (m <= 0)
380 rtt = tp->srtt_us;
381 else
382 rtt -= (m >> 3);
383 tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
384 }
385
386 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
387 unsigned long var;
388
389 if (m < 0)
390 m = -m;
391
392 /* Scale deviation to rttvar fixed point */
393 m >>= 1;
394 if (m < tp->mdev_us)
395 m = tp->mdev_us;
396
397 var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
398 if (m >= var)
399 var = m;
400 else
401 var -= (var - m) >> 2;
402
403 tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
404 }
405
406 if (tcp_in_initial_slowstart(tp)) {
407 /* Slow start still did not finish. */
408 if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
409 !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
410 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
411 if (val && (tcp_snd_cwnd(tp) >> 1) > val)
412 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
413 tcp_snd_cwnd(tp) >> 1);
414 }
415 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
416 val = tcp_metric_get(tm, TCP_METRIC_CWND);
417 if (tcp_snd_cwnd(tp) > val)
418 tcp_metric_set(tm, TCP_METRIC_CWND,
419 tcp_snd_cwnd(tp));
420 }
421 } else if (!tcp_in_slow_start(tp) &&
422 icsk->icsk_ca_state == TCP_CA_Open) {
423 /* Cong. avoidance phase, cwnd is reliable. */
424 if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
425 !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
426 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
427 max(tcp_snd_cwnd(tp) >> 1, tp->snd_ssthresh));
428 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
429 val = tcp_metric_get(tm, TCP_METRIC_CWND);
430 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tcp_snd_cwnd(tp)) >> 1);
431 }
432 } else {
433 /* Else slow start did not finish, cwnd is non-sense,
434 * ssthresh may be also invalid.
435 */
436 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
437 val = tcp_metric_get(tm, TCP_METRIC_CWND);
438 tcp_metric_set(tm, TCP_METRIC_CWND,
439 (val + tp->snd_ssthresh) >> 1);
440 }
441 if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
442 !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
443 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
444 if (val && tp->snd_ssthresh > val)
445 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
446 tp->snd_ssthresh);
447 }
448 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
449 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
450 if (val < tp->reordering &&
451 tp->reordering !=
452 READ_ONCE(net->ipv4.sysctl_tcp_reordering))
453 tcp_metric_set(tm, TCP_METRIC_REORDERING,
454 tp->reordering);
455 }
456 }
457 WRITE_ONCE(tm->tcpm_stamp, jiffies);
458out_unlock:
459 rcu_read_unlock();
460}
461
462/* Initialize metrics on socket. */
463
464void tcp_init_metrics(struct sock *sk)
465{
466 struct dst_entry *dst = __sk_dst_get(sk);
467 struct tcp_sock *tp = tcp_sk(sk);
468 struct net *net = sock_net(sk);
469 struct tcp_metrics_block *tm;
470 u32 val, crtt = 0; /* cached RTT scaled by 8 */
471
472 sk_dst_confirm(sk);
473 /* ssthresh may have been reduced unnecessarily during.
474 * 3WHS. Restore it back to its initial default.
475 */
476 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
477 if (!dst)
478 goto reset;
479
480 rcu_read_lock();
481 tm = tcp_get_metrics(sk, dst, false);
482 if (!tm) {
483 rcu_read_unlock();
484 goto reset;
485 }
486
487 if (tcp_metric_locked(tm, TCP_METRIC_CWND))
488 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
489
490 val = READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) ?
491 0 : tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
492 if (val) {
493 tp->snd_ssthresh = val;
494 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
495 tp->snd_ssthresh = tp->snd_cwnd_clamp;
496 }
497 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
498 if (val && tp->reordering != val)
499 tp->reordering = val;
500
501 crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
502 rcu_read_unlock();
503reset:
504 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
505 * to seed the RTO for later data packets because SYN packets are
506 * small. Use the per-dst cached values to seed the RTO but keep
507 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
508 * Later the RTO will be updated immediately upon obtaining the first
509 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
510 * influences the first RTO but not later RTT estimation.
511 *
512 * But if RTT is not available from the SYN (due to retransmits or
513 * syn cookies) or the cache, force a conservative 3secs timeout.
514 *
515 * A bit of theory. RTT is time passed after "normal" sized packet
516 * is sent until it is ACKed. In normal circumstances sending small
517 * packets force peer to delay ACKs and calculation is correct too.
518 * The algorithm is adaptive and, provided we follow specs, it
519 * NEVER underestimate RTT. BUT! If peer tries to make some clever
520 * tricks sort of "quick acks" for time long enough to decrease RTT
521 * to low value, and then abruptly stops to do it and starts to delay
522 * ACKs, wait for troubles.
523 */
524 if (crtt > tp->srtt_us) {
525 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
526 crtt /= 8 * USEC_PER_SEC / HZ;
527 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
528 } else if (tp->srtt_us == 0) {
529 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
530 * 3WHS. This is most likely due to retransmission,
531 * including spurious one. Reset the RTO back to 3secs
532 * from the more aggressive 1sec to avoid more spurious
533 * retransmission.
534 */
535 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
536 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
537
538 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
539 }
540}
541
542bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
543{
544 struct tcp_metrics_block *tm;
545 bool ret;
546
547 if (!dst)
548 return false;
549
550 rcu_read_lock();
551 tm = __tcp_get_metrics_req(req, dst);
552 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT))
553 ret = true;
554 else
555 ret = false;
556 rcu_read_unlock();
557
558 return ret;
559}
560
561void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
562 struct tcp_fastopen_cookie *cookie)
563{
564 struct tcp_metrics_block *tm;
565
566 rcu_read_lock();
567 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
568 if (tm) {
569 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
570 unsigned int seq;
571
572 do {
573 seq = read_seqbegin(&fastopen_seqlock);
574 if (tfom->mss)
575 *mss = tfom->mss;
576 *cookie = tfom->cookie;
577 if (cookie->len <= 0 && tfom->try_exp == 1)
578 cookie->exp = true;
579 } while (read_seqretry(&fastopen_seqlock, seq));
580 }
581 rcu_read_unlock();
582}
583
584void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
585 struct tcp_fastopen_cookie *cookie, bool syn_lost,
586 u16 try_exp)
587{
588 struct dst_entry *dst = __sk_dst_get(sk);
589 struct tcp_metrics_block *tm;
590
591 if (!dst)
592 return;
593 rcu_read_lock();
594 tm = tcp_get_metrics(sk, dst, true);
595 if (tm) {
596 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
597
598 write_seqlock_bh(&fastopen_seqlock);
599 if (mss)
600 tfom->mss = mss;
601 if (cookie && cookie->len > 0)
602 tfom->cookie = *cookie;
603 else if (try_exp > tfom->try_exp &&
604 tfom->cookie.len <= 0 && !tfom->cookie.exp)
605 tfom->try_exp = try_exp;
606 if (syn_lost) {
607 ++tfom->syn_loss;
608 tfom->last_syn_loss = jiffies;
609 } else
610 tfom->syn_loss = 0;
611 write_sequnlock_bh(&fastopen_seqlock);
612 }
613 rcu_read_unlock();
614}
615
616static struct genl_family tcp_metrics_nl_family;
617
618static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
619 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
620 [TCP_METRICS_ATTR_ADDR_IPV6] =
621 NLA_POLICY_EXACT_LEN(sizeof(struct in6_addr)),
622
623 [TCP_METRICS_ATTR_SADDR_IPV4] = { .type = NLA_U32, },
624 [TCP_METRICS_ATTR_SADDR_IPV6] =
625 NLA_POLICY_EXACT_LEN(sizeof(struct in6_addr)),
626
627 /* Following attributes are not received for GET/DEL,
628 * we keep them for reference
629 */
630#if 0
631 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
632 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
633 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
634 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
635 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
636 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
637 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
638 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
639 .len = TCP_FASTOPEN_COOKIE_MAX, },
640#endif
641};
642
643/* Add attributes, caller cancels its header on failure */
644static int tcp_metrics_fill_info(struct sk_buff *msg,
645 struct tcp_metrics_block *tm)
646{
647 struct nlattr *nest;
648 int i;
649
650 switch (tm->tcpm_daddr.family) {
651 case AF_INET:
652 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
653 inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
654 goto nla_put_failure;
655 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
656 inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
657 goto nla_put_failure;
658 break;
659 case AF_INET6:
660 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
661 inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
662 goto nla_put_failure;
663 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
664 inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
665 goto nla_put_failure;
666 break;
667 default:
668 return -EAFNOSUPPORT;
669 }
670
671 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
672 jiffies - READ_ONCE(tm->tcpm_stamp),
673 TCP_METRICS_ATTR_PAD) < 0)
674 goto nla_put_failure;
675
676 {
677 int n = 0;
678
679 nest = nla_nest_start_noflag(msg, TCP_METRICS_ATTR_VALS);
680 if (!nest)
681 goto nla_put_failure;
682 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
683 u32 val = tcp_metric_get(tm, i);
684
685 if (!val)
686 continue;
687 if (i == TCP_METRIC_RTT) {
688 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
689 val) < 0)
690 goto nla_put_failure;
691 n++;
692 val = max(val / 1000, 1U);
693 }
694 if (i == TCP_METRIC_RTTVAR) {
695 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
696 val) < 0)
697 goto nla_put_failure;
698 n++;
699 val = max(val / 1000, 1U);
700 }
701 if (nla_put_u32(msg, i + 1, val) < 0)
702 goto nla_put_failure;
703 n++;
704 }
705 if (n)
706 nla_nest_end(msg, nest);
707 else
708 nla_nest_cancel(msg, nest);
709 }
710
711 {
712 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
713 unsigned int seq;
714
715 do {
716 seq = read_seqbegin(&fastopen_seqlock);
717 tfom_copy[0] = tm->tcpm_fastopen;
718 } while (read_seqretry(&fastopen_seqlock, seq));
719
720 tfom = tfom_copy;
721 if (tfom->mss &&
722 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
723 tfom->mss) < 0)
724 goto nla_put_failure;
725 if (tfom->syn_loss &&
726 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
727 tfom->syn_loss) < 0 ||
728 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
729 jiffies - tfom->last_syn_loss,
730 TCP_METRICS_ATTR_PAD) < 0))
731 goto nla_put_failure;
732 if (tfom->cookie.len > 0 &&
733 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
734 tfom->cookie.len, tfom->cookie.val) < 0)
735 goto nla_put_failure;
736 }
737
738 return 0;
739
740nla_put_failure:
741 return -EMSGSIZE;
742}
743
744static int tcp_metrics_dump_info(struct sk_buff *skb,
745 struct netlink_callback *cb,
746 struct tcp_metrics_block *tm)
747{
748 void *hdr;
749
750 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
751 &tcp_metrics_nl_family, NLM_F_MULTI,
752 TCP_METRICS_CMD_GET);
753 if (!hdr)
754 return -EMSGSIZE;
755
756 if (tcp_metrics_fill_info(skb, tm) < 0)
757 goto nla_put_failure;
758
759 genlmsg_end(skb, hdr);
760 return 0;
761
762nla_put_failure:
763 genlmsg_cancel(skb, hdr);
764 return -EMSGSIZE;
765}
766
767static int tcp_metrics_nl_dump(struct sk_buff *skb,
768 struct netlink_callback *cb)
769{
770 struct net *net = sock_net(skb->sk);
771 unsigned int max_rows = 1U << tcp_metrics_hash_log;
772 unsigned int row, s_row = cb->args[0];
773 int s_col = cb->args[1], col = s_col;
774 int res = 0;
775
776 for (row = s_row; row < max_rows; row++, s_col = 0) {
777 struct tcp_metrics_block *tm;
778 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
779
780 rcu_read_lock();
781 for (col = 0, tm = rcu_dereference(hb->chain); tm;
782 tm = rcu_dereference(tm->tcpm_next), col++) {
783 if (!net_eq(tm_net(tm), net))
784 continue;
785 if (col < s_col)
786 continue;
787 res = tcp_metrics_dump_info(skb, cb, tm);
788 if (res < 0) {
789 rcu_read_unlock();
790 goto done;
791 }
792 }
793 rcu_read_unlock();
794 }
795
796done:
797 cb->args[0] = row;
798 cb->args[1] = col;
799 return res;
800}
801
802static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
803 unsigned int *hash, int optional, int v4, int v6)
804{
805 struct nlattr *a;
806
807 a = info->attrs[v4];
808 if (a) {
809 inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
810 if (hash)
811 *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
812 return 0;
813 }
814 a = info->attrs[v6];
815 if (a) {
816 struct in6_addr in6;
817
818 in6 = nla_get_in6_addr(a);
819 inetpeer_set_addr_v6(addr, &in6);
820 if (hash)
821 *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
822 return 0;
823 }
824 return optional ? 1 : -EAFNOSUPPORT;
825}
826
827static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
828 unsigned int *hash, int optional)
829{
830 return __parse_nl_addr(info, addr, hash, optional,
831 TCP_METRICS_ATTR_ADDR_IPV4,
832 TCP_METRICS_ATTR_ADDR_IPV6);
833}
834
835static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
836{
837 return __parse_nl_addr(info, addr, NULL, 0,
838 TCP_METRICS_ATTR_SADDR_IPV4,
839 TCP_METRICS_ATTR_SADDR_IPV6);
840}
841
842static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
843{
844 struct tcp_metrics_block *tm;
845 struct inetpeer_addr saddr, daddr;
846 unsigned int hash;
847 struct sk_buff *msg;
848 struct net *net = genl_info_net(info);
849 void *reply;
850 int ret;
851 bool src = true;
852
853 ret = parse_nl_addr(info, &daddr, &hash, 0);
854 if (ret < 0)
855 return ret;
856
857 ret = parse_nl_saddr(info, &saddr);
858 if (ret < 0)
859 src = false;
860
861 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
862 if (!msg)
863 return -ENOMEM;
864
865 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
866 info->genlhdr->cmd);
867 if (!reply)
868 goto nla_put_failure;
869
870 hash ^= net_hash_mix(net);
871 hash = hash_32(hash, tcp_metrics_hash_log);
872 ret = -ESRCH;
873 rcu_read_lock();
874 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
875 tm = rcu_dereference(tm->tcpm_next)) {
876 if (addr_same(&tm->tcpm_daddr, &daddr) &&
877 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
878 net_eq(tm_net(tm), net)) {
879 ret = tcp_metrics_fill_info(msg, tm);
880 break;
881 }
882 }
883 rcu_read_unlock();
884 if (ret < 0)
885 goto out_free;
886
887 genlmsg_end(msg, reply);
888 return genlmsg_reply(msg, info);
889
890nla_put_failure:
891 ret = -EMSGSIZE;
892
893out_free:
894 nlmsg_free(msg);
895 return ret;
896}
897
898static void tcp_metrics_flush_all(struct net *net)
899{
900 unsigned int max_rows = 1U << tcp_metrics_hash_log;
901 struct tcpm_hash_bucket *hb = tcp_metrics_hash;
902 struct tcp_metrics_block *tm;
903 unsigned int row;
904
905 for (row = 0; row < max_rows; row++, hb++) {
906 struct tcp_metrics_block __rcu **pp = &hb->chain;
907 bool match;
908
909 if (!rcu_access_pointer(*pp))
910 continue;
911
912 spin_lock_bh(&tcp_metrics_lock);
913 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
914 match = net ? net_eq(tm_net(tm), net) :
915 !refcount_read(&tm_net(tm)->ns.count);
916 if (match) {
917 rcu_assign_pointer(*pp, tm->tcpm_next);
918 kfree_rcu(tm, rcu_head);
919 } else {
920 pp = &tm->tcpm_next;
921 }
922 }
923 spin_unlock_bh(&tcp_metrics_lock);
924 cond_resched();
925 }
926}
927
928static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
929{
930 struct tcpm_hash_bucket *hb;
931 struct tcp_metrics_block *tm;
932 struct tcp_metrics_block __rcu **pp;
933 struct inetpeer_addr saddr, daddr;
934 unsigned int hash;
935 struct net *net = genl_info_net(info);
936 int ret;
937 bool src = true, found = false;
938
939 ret = parse_nl_addr(info, &daddr, &hash, 1);
940 if (ret < 0)
941 return ret;
942 if (ret > 0) {
943 tcp_metrics_flush_all(net);
944 return 0;
945 }
946 ret = parse_nl_saddr(info, &saddr);
947 if (ret < 0)
948 src = false;
949
950 hash ^= net_hash_mix(net);
951 hash = hash_32(hash, tcp_metrics_hash_log);
952 hb = tcp_metrics_hash + hash;
953 pp = &hb->chain;
954 spin_lock_bh(&tcp_metrics_lock);
955 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
956 if (addr_same(&tm->tcpm_daddr, &daddr) &&
957 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
958 net_eq(tm_net(tm), net)) {
959 rcu_assign_pointer(*pp, tm->tcpm_next);
960 kfree_rcu(tm, rcu_head);
961 found = true;
962 } else {
963 pp = &tm->tcpm_next;
964 }
965 }
966 spin_unlock_bh(&tcp_metrics_lock);
967 if (!found)
968 return -ESRCH;
969 return 0;
970}
971
972static const struct genl_small_ops tcp_metrics_nl_ops[] = {
973 {
974 .cmd = TCP_METRICS_CMD_GET,
975 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
976 .doit = tcp_metrics_nl_cmd_get,
977 .dumpit = tcp_metrics_nl_dump,
978 },
979 {
980 .cmd = TCP_METRICS_CMD_DEL,
981 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
982 .doit = tcp_metrics_nl_cmd_del,
983 .flags = GENL_ADMIN_PERM,
984 },
985};
986
987static struct genl_family tcp_metrics_nl_family __ro_after_init = {
988 .hdrsize = 0,
989 .name = TCP_METRICS_GENL_NAME,
990 .version = TCP_METRICS_GENL_VERSION,
991 .maxattr = TCP_METRICS_ATTR_MAX,
992 .policy = tcp_metrics_nl_policy,
993 .netnsok = true,
994 .parallel_ops = true,
995 .module = THIS_MODULE,
996 .small_ops = tcp_metrics_nl_ops,
997 .n_small_ops = ARRAY_SIZE(tcp_metrics_nl_ops),
998 .resv_start_op = TCP_METRICS_CMD_DEL + 1,
999};
1000
1001static unsigned int tcpmhash_entries __initdata;
1002static int __init set_tcpmhash_entries(char *str)
1003{
1004 ssize_t ret;
1005
1006 if (!str)
1007 return 0;
1008
1009 ret = kstrtouint(str, 0, &tcpmhash_entries);
1010 if (ret)
1011 return 0;
1012
1013 return 1;
1014}
1015__setup("tcpmhash_entries=", set_tcpmhash_entries);
1016
1017static void __init tcp_metrics_hash_alloc(void)
1018{
1019 unsigned int slots = tcpmhash_entries;
1020 size_t size;
1021
1022 if (!slots) {
1023 if (totalram_pages() >= 128 * 1024)
1024 slots = 16 * 1024;
1025 else
1026 slots = 8 * 1024;
1027 }
1028
1029 tcp_metrics_hash_log = order_base_2(slots);
1030 size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1031
1032 tcp_metrics_hash = kvzalloc(size, GFP_KERNEL);
1033 if (!tcp_metrics_hash)
1034 panic("Could not allocate the tcp_metrics hash table\n");
1035}
1036
1037static void __net_exit tcp_net_metrics_exit_batch(struct list_head *net_exit_list)
1038{
1039 tcp_metrics_flush_all(NULL);
1040}
1041
1042static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1043 .exit_batch = tcp_net_metrics_exit_batch,
1044};
1045
1046void __init tcp_metrics_init(void)
1047{
1048 int ret;
1049
1050 tcp_metrics_hash_alloc();
1051
1052 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1053 if (ret < 0)
1054 panic("Could not register tcp_net_metrics_ops\n");
1055
1056 ret = genl_register_family(&tcp_metrics_nl_family);
1057 if (ret < 0)
1058 panic("Could not register tcp_metrics generic netlink\n");
1059}