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