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 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
516 * retransmitted. In light of RFC6298 more aggressive 1sec
517 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
518 * retransmission has occurred.
519 */
520 if (tp->total_retrans > 1)
521 tp->snd_cwnd = 1;
522 else
523 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
524 tp->snd_cwnd_stamp = tcp_jiffies32;
525}
526
527bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
528{
529 struct tcp_metrics_block *tm;
530 bool ret;
531
532 if (!dst)
533 return false;
534
535 rcu_read_lock();
536 tm = __tcp_get_metrics_req(req, dst);
537 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT))
538 ret = true;
539 else
540 ret = false;
541 rcu_read_unlock();
542
543 return ret;
544}
545
546static DEFINE_SEQLOCK(fastopen_seqlock);
547
548void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
549 struct tcp_fastopen_cookie *cookie)
550{
551 struct tcp_metrics_block *tm;
552
553 rcu_read_lock();
554 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
555 if (tm) {
556 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
557 unsigned int seq;
558
559 do {
560 seq = read_seqbegin(&fastopen_seqlock);
561 if (tfom->mss)
562 *mss = tfom->mss;
563 *cookie = tfom->cookie;
564 if (cookie->len <= 0 && tfom->try_exp == 1)
565 cookie->exp = true;
566 } while (read_seqretry(&fastopen_seqlock, seq));
567 }
568 rcu_read_unlock();
569}
570
571void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
572 struct tcp_fastopen_cookie *cookie, bool syn_lost,
573 u16 try_exp)
574{
575 struct dst_entry *dst = __sk_dst_get(sk);
576 struct tcp_metrics_block *tm;
577
578 if (!dst)
579 return;
580 rcu_read_lock();
581 tm = tcp_get_metrics(sk, dst, true);
582 if (tm) {
583 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
584
585 write_seqlock_bh(&fastopen_seqlock);
586 if (mss)
587 tfom->mss = mss;
588 if (cookie && cookie->len > 0)
589 tfom->cookie = *cookie;
590 else if (try_exp > tfom->try_exp &&
591 tfom->cookie.len <= 0 && !tfom->cookie.exp)
592 tfom->try_exp = try_exp;
593 if (syn_lost) {
594 ++tfom->syn_loss;
595 tfom->last_syn_loss = jiffies;
596 } else
597 tfom->syn_loss = 0;
598 write_sequnlock_bh(&fastopen_seqlock);
599 }
600 rcu_read_unlock();
601}
602
603static struct genl_family tcp_metrics_nl_family;
604
605static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
606 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
607 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
608 .len = sizeof(struct in6_addr), },
609 /* Following attributes are not received for GET/DEL,
610 * we keep them for reference
611 */
612#if 0
613 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
614 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
615 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
616 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
617 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
618 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
619 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
620 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
621 .len = TCP_FASTOPEN_COOKIE_MAX, },
622#endif
623};
624
625/* Add attributes, caller cancels its header on failure */
626static int tcp_metrics_fill_info(struct sk_buff *msg,
627 struct tcp_metrics_block *tm)
628{
629 struct nlattr *nest;
630 int i;
631
632 switch (tm->tcpm_daddr.family) {
633 case AF_INET:
634 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
635 inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
636 goto nla_put_failure;
637 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
638 inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
639 goto nla_put_failure;
640 break;
641 case AF_INET6:
642 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
643 inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
644 goto nla_put_failure;
645 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
646 inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
647 goto nla_put_failure;
648 break;
649 default:
650 return -EAFNOSUPPORT;
651 }
652
653 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
654 jiffies - tm->tcpm_stamp,
655 TCP_METRICS_ATTR_PAD) < 0)
656 goto nla_put_failure;
657
658 {
659 int n = 0;
660
661 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
662 if (!nest)
663 goto nla_put_failure;
664 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
665 u32 val = tm->tcpm_vals[i];
666
667 if (!val)
668 continue;
669 if (i == TCP_METRIC_RTT) {
670 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
671 val) < 0)
672 goto nla_put_failure;
673 n++;
674 val = max(val / 1000, 1U);
675 }
676 if (i == TCP_METRIC_RTTVAR) {
677 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
678 val) < 0)
679 goto nla_put_failure;
680 n++;
681 val = max(val / 1000, 1U);
682 }
683 if (nla_put_u32(msg, i + 1, val) < 0)
684 goto nla_put_failure;
685 n++;
686 }
687 if (n)
688 nla_nest_end(msg, nest);
689 else
690 nla_nest_cancel(msg, nest);
691 }
692
693 {
694 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
695 unsigned int seq;
696
697 do {
698 seq = read_seqbegin(&fastopen_seqlock);
699 tfom_copy[0] = tm->tcpm_fastopen;
700 } while (read_seqretry(&fastopen_seqlock, seq));
701
702 tfom = tfom_copy;
703 if (tfom->mss &&
704 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
705 tfom->mss) < 0)
706 goto nla_put_failure;
707 if (tfom->syn_loss &&
708 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
709 tfom->syn_loss) < 0 ||
710 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
711 jiffies - tfom->last_syn_loss,
712 TCP_METRICS_ATTR_PAD) < 0))
713 goto nla_put_failure;
714 if (tfom->cookie.len > 0 &&
715 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
716 tfom->cookie.len, tfom->cookie.val) < 0)
717 goto nla_put_failure;
718 }
719
720 return 0;
721
722nla_put_failure:
723 return -EMSGSIZE;
724}
725
726static int tcp_metrics_dump_info(struct sk_buff *skb,
727 struct netlink_callback *cb,
728 struct tcp_metrics_block *tm)
729{
730 void *hdr;
731
732 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
733 &tcp_metrics_nl_family, NLM_F_MULTI,
734 TCP_METRICS_CMD_GET);
735 if (!hdr)
736 return -EMSGSIZE;
737
738 if (tcp_metrics_fill_info(skb, tm) < 0)
739 goto nla_put_failure;
740
741 genlmsg_end(skb, hdr);
742 return 0;
743
744nla_put_failure:
745 genlmsg_cancel(skb, hdr);
746 return -EMSGSIZE;
747}
748
749static int tcp_metrics_nl_dump(struct sk_buff *skb,
750 struct netlink_callback *cb)
751{
752 struct net *net = sock_net(skb->sk);
753 unsigned int max_rows = 1U << tcp_metrics_hash_log;
754 unsigned int row, s_row = cb->args[0];
755 int s_col = cb->args[1], col = s_col;
756
757 for (row = s_row; row < max_rows; row++, s_col = 0) {
758 struct tcp_metrics_block *tm;
759 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
760
761 rcu_read_lock();
762 for (col = 0, tm = rcu_dereference(hb->chain); tm;
763 tm = rcu_dereference(tm->tcpm_next), col++) {
764 if (!net_eq(tm_net(tm), net))
765 continue;
766 if (col < s_col)
767 continue;
768 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
769 rcu_read_unlock();
770 goto done;
771 }
772 }
773 rcu_read_unlock();
774 }
775
776done:
777 cb->args[0] = row;
778 cb->args[1] = col;
779 return skb->len;
780}
781
782static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
783 unsigned int *hash, int optional, int v4, int v6)
784{
785 struct nlattr *a;
786
787 a = info->attrs[v4];
788 if (a) {
789 inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
790 if (hash)
791 *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
792 return 0;
793 }
794 a = info->attrs[v6];
795 if (a) {
796 struct in6_addr in6;
797
798 if (nla_len(a) != sizeof(struct in6_addr))
799 return -EINVAL;
800 in6 = nla_get_in6_addr(a);
801 inetpeer_set_addr_v6(addr, &in6);
802 if (hash)
803 *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
804 return 0;
805 }
806 return optional ? 1 : -EAFNOSUPPORT;
807}
808
809static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
810 unsigned int *hash, int optional)
811{
812 return __parse_nl_addr(info, addr, hash, optional,
813 TCP_METRICS_ATTR_ADDR_IPV4,
814 TCP_METRICS_ATTR_ADDR_IPV6);
815}
816
817static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
818{
819 return __parse_nl_addr(info, addr, NULL, 0,
820 TCP_METRICS_ATTR_SADDR_IPV4,
821 TCP_METRICS_ATTR_SADDR_IPV6);
822}
823
824static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
825{
826 struct tcp_metrics_block *tm;
827 struct inetpeer_addr saddr, daddr;
828 unsigned int hash;
829 struct sk_buff *msg;
830 struct net *net = genl_info_net(info);
831 void *reply;
832 int ret;
833 bool src = true;
834
835 ret = parse_nl_addr(info, &daddr, &hash, 0);
836 if (ret < 0)
837 return ret;
838
839 ret = parse_nl_saddr(info, &saddr);
840 if (ret < 0)
841 src = false;
842
843 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
844 if (!msg)
845 return -ENOMEM;
846
847 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
848 info->genlhdr->cmd);
849 if (!reply)
850 goto nla_put_failure;
851
852 hash ^= net_hash_mix(net);
853 hash = hash_32(hash, tcp_metrics_hash_log);
854 ret = -ESRCH;
855 rcu_read_lock();
856 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
857 tm = rcu_dereference(tm->tcpm_next)) {
858 if (addr_same(&tm->tcpm_daddr, &daddr) &&
859 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
860 net_eq(tm_net(tm), net)) {
861 ret = tcp_metrics_fill_info(msg, tm);
862 break;
863 }
864 }
865 rcu_read_unlock();
866 if (ret < 0)
867 goto out_free;
868
869 genlmsg_end(msg, reply);
870 return genlmsg_reply(msg, info);
871
872nla_put_failure:
873 ret = -EMSGSIZE;
874
875out_free:
876 nlmsg_free(msg);
877 return ret;
878}
879
880static void tcp_metrics_flush_all(struct net *net)
881{
882 unsigned int max_rows = 1U << tcp_metrics_hash_log;
883 struct tcpm_hash_bucket *hb = tcp_metrics_hash;
884 struct tcp_metrics_block *tm;
885 unsigned int row;
886
887 for (row = 0; row < max_rows; row++, hb++) {
888 struct tcp_metrics_block __rcu **pp;
889 bool match;
890
891 spin_lock_bh(&tcp_metrics_lock);
892 pp = &hb->chain;
893 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
894 match = net ? net_eq(tm_net(tm), net) :
895 !refcount_read(&tm_net(tm)->count);
896 if (match) {
897 *pp = tm->tcpm_next;
898 kfree_rcu(tm, rcu_head);
899 } else {
900 pp = &tm->tcpm_next;
901 }
902 }
903 spin_unlock_bh(&tcp_metrics_lock);
904 }
905}
906
907static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
908{
909 struct tcpm_hash_bucket *hb;
910 struct tcp_metrics_block *tm;
911 struct tcp_metrics_block __rcu **pp;
912 struct inetpeer_addr saddr, daddr;
913 unsigned int hash;
914 struct net *net = genl_info_net(info);
915 int ret;
916 bool src = true, found = false;
917
918 ret = parse_nl_addr(info, &daddr, &hash, 1);
919 if (ret < 0)
920 return ret;
921 if (ret > 0) {
922 tcp_metrics_flush_all(net);
923 return 0;
924 }
925 ret = parse_nl_saddr(info, &saddr);
926 if (ret < 0)
927 src = false;
928
929 hash ^= net_hash_mix(net);
930 hash = hash_32(hash, tcp_metrics_hash_log);
931 hb = tcp_metrics_hash + hash;
932 pp = &hb->chain;
933 spin_lock_bh(&tcp_metrics_lock);
934 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
935 if (addr_same(&tm->tcpm_daddr, &daddr) &&
936 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
937 net_eq(tm_net(tm), net)) {
938 *pp = tm->tcpm_next;
939 kfree_rcu(tm, rcu_head);
940 found = true;
941 } else {
942 pp = &tm->tcpm_next;
943 }
944 }
945 spin_unlock_bh(&tcp_metrics_lock);
946 if (!found)
947 return -ESRCH;
948 return 0;
949}
950
951static const struct genl_ops tcp_metrics_nl_ops[] = {
952 {
953 .cmd = TCP_METRICS_CMD_GET,
954 .doit = tcp_metrics_nl_cmd_get,
955 .dumpit = tcp_metrics_nl_dump,
956 .policy = tcp_metrics_nl_policy,
957 },
958 {
959 .cmd = TCP_METRICS_CMD_DEL,
960 .doit = tcp_metrics_nl_cmd_del,
961 .policy = tcp_metrics_nl_policy,
962 .flags = GENL_ADMIN_PERM,
963 },
964};
965
966static struct genl_family tcp_metrics_nl_family __ro_after_init = {
967 .hdrsize = 0,
968 .name = TCP_METRICS_GENL_NAME,
969 .version = TCP_METRICS_GENL_VERSION,
970 .maxattr = TCP_METRICS_ATTR_MAX,
971 .netnsok = true,
972 .module = THIS_MODULE,
973 .ops = tcp_metrics_nl_ops,
974 .n_ops = ARRAY_SIZE(tcp_metrics_nl_ops),
975};
976
977static unsigned int tcpmhash_entries;
978static int __init set_tcpmhash_entries(char *str)
979{
980 ssize_t ret;
981
982 if (!str)
983 return 0;
984
985 ret = kstrtouint(str, 0, &tcpmhash_entries);
986 if (ret)
987 return 0;
988
989 return 1;
990}
991__setup("tcpmhash_entries=", set_tcpmhash_entries);
992
993static int __net_init tcp_net_metrics_init(struct net *net)
994{
995 size_t size;
996 unsigned int slots;
997
998 if (!net_eq(net, &init_net))
999 return 0;
1000
1001 slots = tcpmhash_entries;
1002 if (!slots) {
1003 if (totalram_pages >= 128 * 1024)
1004 slots = 16 * 1024;
1005 else
1006 slots = 8 * 1024;
1007 }
1008
1009 tcp_metrics_hash_log = order_base_2(slots);
1010 size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1011
1012 tcp_metrics_hash = kvzalloc(size, GFP_KERNEL);
1013 if (!tcp_metrics_hash)
1014 return -ENOMEM;
1015
1016 return 0;
1017}
1018
1019static void __net_exit tcp_net_metrics_exit_batch(struct list_head *net_exit_list)
1020{
1021 tcp_metrics_flush_all(NULL);
1022}
1023
1024static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1025 .init = tcp_net_metrics_init,
1026 .exit_batch = tcp_net_metrics_exit_batch,
1027};
1028
1029void __init tcp_metrics_init(void)
1030{
1031 int ret;
1032
1033 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1034 if (ret < 0)
1035 panic("Could not allocate the tcp_metrics hash table\n");
1036
1037 ret = genl_register_family(&tcp_metrics_nl_family);
1038 if (ret < 0)
1039 panic("Could not register tcp_metrics generic netlink\n");
1040}
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}
609
610void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
611{
612 struct tcp_metrics_block *tm;
613
614 rcu_read_lock();
615 tm = tcp_get_metrics(sk, dst, true);
616 if (tm) {
617 struct tcp_sock *tp = tcp_sk(sk);
618
619 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
620 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
621 tp->rx_opt.ts_recent = tm->tcpm_ts;
622 }
623 }
624 rcu_read_unlock();
625}
626EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
627
628/* VJ's idea. Save last timestamp seen from this destination and hold
629 * it at least for normal timewait interval to use for duplicate
630 * segment detection in subsequent connections, before they enter
631 * synchronized state.
632 */
633bool tcp_remember_stamp(struct sock *sk)
634{
635 struct dst_entry *dst = __sk_dst_get(sk);
636 bool ret = false;
637
638 if (dst) {
639 struct tcp_metrics_block *tm;
640
641 rcu_read_lock();
642 tm = tcp_get_metrics(sk, dst, true);
643 if (tm) {
644 struct tcp_sock *tp = tcp_sk(sk);
645
646 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
647 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
648 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
649 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
650 tm->tcpm_ts = tp->rx_opt.ts_recent;
651 }
652 ret = true;
653 }
654 rcu_read_unlock();
655 }
656 return ret;
657}
658
659bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
660{
661 struct tcp_metrics_block *tm;
662 bool ret = false;
663
664 rcu_read_lock();
665 tm = __tcp_get_metrics_tw(tw);
666 if (tm) {
667 const struct tcp_timewait_sock *tcptw;
668 struct sock *sk = (struct sock *) tw;
669
670 tcptw = tcp_twsk(sk);
671 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
672 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
673 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
674 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
675 tm->tcpm_ts = tcptw->tw_ts_recent;
676 }
677 ret = true;
678 }
679 rcu_read_unlock();
680
681 return ret;
682}
683
684static DEFINE_SEQLOCK(fastopen_seqlock);
685
686void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
687 struct tcp_fastopen_cookie *cookie,
688 int *syn_loss, unsigned long *last_syn_loss)
689{
690 struct tcp_metrics_block *tm;
691
692 rcu_read_lock();
693 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
694 if (tm) {
695 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
696 unsigned int seq;
697
698 do {
699 seq = read_seqbegin(&fastopen_seqlock);
700 if (tfom->mss)
701 *mss = tfom->mss;
702 *cookie = tfom->cookie;
703 if (cookie->len <= 0 && tfom->try_exp == 1)
704 cookie->exp = true;
705 *syn_loss = tfom->syn_loss;
706 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
707 } while (read_seqretry(&fastopen_seqlock, seq));
708 }
709 rcu_read_unlock();
710}
711
712void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
713 struct tcp_fastopen_cookie *cookie, bool syn_lost,
714 u16 try_exp)
715{
716 struct dst_entry *dst = __sk_dst_get(sk);
717 struct tcp_metrics_block *tm;
718
719 if (!dst)
720 return;
721 rcu_read_lock();
722 tm = tcp_get_metrics(sk, dst, true);
723 if (tm) {
724 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
725
726 write_seqlock_bh(&fastopen_seqlock);
727 if (mss)
728 tfom->mss = mss;
729 if (cookie && cookie->len > 0)
730 tfom->cookie = *cookie;
731 else if (try_exp > tfom->try_exp &&
732 tfom->cookie.len <= 0 && !tfom->cookie.exp)
733 tfom->try_exp = try_exp;
734 if (syn_lost) {
735 ++tfom->syn_loss;
736 tfom->last_syn_loss = jiffies;
737 } else
738 tfom->syn_loss = 0;
739 write_sequnlock_bh(&fastopen_seqlock);
740 }
741 rcu_read_unlock();
742}
743
744static struct genl_family tcp_metrics_nl_family;
745
746static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
747 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
748 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
749 .len = sizeof(struct in6_addr), },
750 /* Following attributes are not received for GET/DEL,
751 * we keep them for reference
752 */
753#if 0
754 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
755 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
756 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
757 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
758 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
759 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
760 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
761 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
762 .len = TCP_FASTOPEN_COOKIE_MAX, },
763#endif
764};
765
766/* Add attributes, caller cancels its header on failure */
767static int tcp_metrics_fill_info(struct sk_buff *msg,
768 struct tcp_metrics_block *tm)
769{
770 struct nlattr *nest;
771 int i;
772
773 switch (tm->tcpm_daddr.family) {
774 case AF_INET:
775 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
776 inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
777 goto nla_put_failure;
778 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
779 inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
780 goto nla_put_failure;
781 break;
782 case AF_INET6:
783 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
784 inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
785 goto nla_put_failure;
786 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
787 inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
788 goto nla_put_failure;
789 break;
790 default:
791 return -EAFNOSUPPORT;
792 }
793
794 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
795 jiffies - tm->tcpm_stamp,
796 TCP_METRICS_ATTR_PAD) < 0)
797 goto nla_put_failure;
798 if (tm->tcpm_ts_stamp) {
799 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
800 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
801 goto nla_put_failure;
802 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
803 tm->tcpm_ts) < 0)
804 goto nla_put_failure;
805 }
806
807 {
808 int n = 0;
809
810 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
811 if (!nest)
812 goto nla_put_failure;
813 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
814 u32 val = tm->tcpm_vals[i];
815
816 if (!val)
817 continue;
818 if (i == TCP_METRIC_RTT) {
819 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
820 val) < 0)
821 goto nla_put_failure;
822 n++;
823 val = max(val / 1000, 1U);
824 }
825 if (i == TCP_METRIC_RTTVAR) {
826 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
827 val) < 0)
828 goto nla_put_failure;
829 n++;
830 val = max(val / 1000, 1U);
831 }
832 if (nla_put_u32(msg, i + 1, val) < 0)
833 goto nla_put_failure;
834 n++;
835 }
836 if (n)
837 nla_nest_end(msg, nest);
838 else
839 nla_nest_cancel(msg, nest);
840 }
841
842 {
843 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
844 unsigned int seq;
845
846 do {
847 seq = read_seqbegin(&fastopen_seqlock);
848 tfom_copy[0] = tm->tcpm_fastopen;
849 } while (read_seqretry(&fastopen_seqlock, seq));
850
851 tfom = tfom_copy;
852 if (tfom->mss &&
853 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
854 tfom->mss) < 0)
855 goto nla_put_failure;
856 if (tfom->syn_loss &&
857 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
858 tfom->syn_loss) < 0 ||
859 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
860 jiffies - tfom->last_syn_loss,
861 TCP_METRICS_ATTR_PAD) < 0))
862 goto nla_put_failure;
863 if (tfom->cookie.len > 0 &&
864 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
865 tfom->cookie.len, tfom->cookie.val) < 0)
866 goto nla_put_failure;
867 }
868
869 return 0;
870
871nla_put_failure:
872 return -EMSGSIZE;
873}
874
875static int tcp_metrics_dump_info(struct sk_buff *skb,
876 struct netlink_callback *cb,
877 struct tcp_metrics_block *tm)
878{
879 void *hdr;
880
881 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
882 &tcp_metrics_nl_family, NLM_F_MULTI,
883 TCP_METRICS_CMD_GET);
884 if (!hdr)
885 return -EMSGSIZE;
886
887 if (tcp_metrics_fill_info(skb, tm) < 0)
888 goto nla_put_failure;
889
890 genlmsg_end(skb, hdr);
891 return 0;
892
893nla_put_failure:
894 genlmsg_cancel(skb, hdr);
895 return -EMSGSIZE;
896}
897
898static int tcp_metrics_nl_dump(struct sk_buff *skb,
899 struct netlink_callback *cb)
900{
901 struct net *net = sock_net(skb->sk);
902 unsigned int max_rows = 1U << tcp_metrics_hash_log;
903 unsigned int row, s_row = cb->args[0];
904 int s_col = cb->args[1], col = s_col;
905
906 for (row = s_row; row < max_rows; row++, s_col = 0) {
907 struct tcp_metrics_block *tm;
908 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
909
910 rcu_read_lock();
911 for (col = 0, tm = rcu_dereference(hb->chain); tm;
912 tm = rcu_dereference(tm->tcpm_next), col++) {
913 if (!net_eq(tm_net(tm), net))
914 continue;
915 if (col < s_col)
916 continue;
917 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
918 rcu_read_unlock();
919 goto done;
920 }
921 }
922 rcu_read_unlock();
923 }
924
925done:
926 cb->args[0] = row;
927 cb->args[1] = col;
928 return skb->len;
929}
930
931static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
932 unsigned int *hash, int optional, int v4, int v6)
933{
934 struct nlattr *a;
935
936 a = info->attrs[v4];
937 if (a) {
938 inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
939 if (hash)
940 *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
941 return 0;
942 }
943 a = info->attrs[v6];
944 if (a) {
945 struct in6_addr in6;
946
947 if (nla_len(a) != sizeof(struct in6_addr))
948 return -EINVAL;
949 in6 = nla_get_in6_addr(a);
950 inetpeer_set_addr_v6(addr, &in6);
951 if (hash)
952 *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
953 return 0;
954 }
955 return optional ? 1 : -EAFNOSUPPORT;
956}
957
958static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
959 unsigned int *hash, int optional)
960{
961 return __parse_nl_addr(info, addr, hash, optional,
962 TCP_METRICS_ATTR_ADDR_IPV4,
963 TCP_METRICS_ATTR_ADDR_IPV6);
964}
965
966static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
967{
968 return __parse_nl_addr(info, addr, NULL, 0,
969 TCP_METRICS_ATTR_SADDR_IPV4,
970 TCP_METRICS_ATTR_SADDR_IPV6);
971}
972
973static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
974{
975 struct tcp_metrics_block *tm;
976 struct inetpeer_addr saddr, daddr;
977 unsigned int hash;
978 struct sk_buff *msg;
979 struct net *net = genl_info_net(info);
980 void *reply;
981 int ret;
982 bool src = true;
983
984 ret = parse_nl_addr(info, &daddr, &hash, 0);
985 if (ret < 0)
986 return ret;
987
988 ret = parse_nl_saddr(info, &saddr);
989 if (ret < 0)
990 src = false;
991
992 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
993 if (!msg)
994 return -ENOMEM;
995
996 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
997 info->genlhdr->cmd);
998 if (!reply)
999 goto nla_put_failure;
1000
1001 hash ^= net_hash_mix(net);
1002 hash = hash_32(hash, tcp_metrics_hash_log);
1003 ret = -ESRCH;
1004 rcu_read_lock();
1005 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
1006 tm = rcu_dereference(tm->tcpm_next)) {
1007 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1008 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1009 net_eq(tm_net(tm), net)) {
1010 ret = tcp_metrics_fill_info(msg, tm);
1011 break;
1012 }
1013 }
1014 rcu_read_unlock();
1015 if (ret < 0)
1016 goto out_free;
1017
1018 genlmsg_end(msg, reply);
1019 return genlmsg_reply(msg, info);
1020
1021nla_put_failure:
1022 ret = -EMSGSIZE;
1023
1024out_free:
1025 nlmsg_free(msg);
1026 return ret;
1027}
1028
1029static void tcp_metrics_flush_all(struct net *net)
1030{
1031 unsigned int max_rows = 1U << tcp_metrics_hash_log;
1032 struct tcpm_hash_bucket *hb = tcp_metrics_hash;
1033 struct tcp_metrics_block *tm;
1034 unsigned int row;
1035
1036 for (row = 0; row < max_rows; row++, hb++) {
1037 struct tcp_metrics_block __rcu **pp;
1038 spin_lock_bh(&tcp_metrics_lock);
1039 pp = &hb->chain;
1040 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1041 if (net_eq(tm_net(tm), net)) {
1042 *pp = tm->tcpm_next;
1043 kfree_rcu(tm, rcu_head);
1044 } else {
1045 pp = &tm->tcpm_next;
1046 }
1047 }
1048 spin_unlock_bh(&tcp_metrics_lock);
1049 }
1050}
1051
1052static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
1053{
1054 struct tcpm_hash_bucket *hb;
1055 struct tcp_metrics_block *tm;
1056 struct tcp_metrics_block __rcu **pp;
1057 struct inetpeer_addr saddr, daddr;
1058 unsigned int hash;
1059 struct net *net = genl_info_net(info);
1060 int ret;
1061 bool src = true, found = false;
1062
1063 ret = parse_nl_addr(info, &daddr, &hash, 1);
1064 if (ret < 0)
1065 return ret;
1066 if (ret > 0) {
1067 tcp_metrics_flush_all(net);
1068 return 0;
1069 }
1070 ret = parse_nl_saddr(info, &saddr);
1071 if (ret < 0)
1072 src = false;
1073
1074 hash ^= net_hash_mix(net);
1075 hash = hash_32(hash, tcp_metrics_hash_log);
1076 hb = tcp_metrics_hash + hash;
1077 pp = &hb->chain;
1078 spin_lock_bh(&tcp_metrics_lock);
1079 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1080 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1081 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1082 net_eq(tm_net(tm), net)) {
1083 *pp = tm->tcpm_next;
1084 kfree_rcu(tm, rcu_head);
1085 found = true;
1086 } else {
1087 pp = &tm->tcpm_next;
1088 }
1089 }
1090 spin_unlock_bh(&tcp_metrics_lock);
1091 if (!found)
1092 return -ESRCH;
1093 return 0;
1094}
1095
1096static const struct genl_ops tcp_metrics_nl_ops[] = {
1097 {
1098 .cmd = TCP_METRICS_CMD_GET,
1099 .doit = tcp_metrics_nl_cmd_get,
1100 .dumpit = tcp_metrics_nl_dump,
1101 .policy = tcp_metrics_nl_policy,
1102 },
1103 {
1104 .cmd = TCP_METRICS_CMD_DEL,
1105 .doit = tcp_metrics_nl_cmd_del,
1106 .policy = tcp_metrics_nl_policy,
1107 .flags = GENL_ADMIN_PERM,
1108 },
1109};
1110
1111static struct genl_family tcp_metrics_nl_family __ro_after_init = {
1112 .hdrsize = 0,
1113 .name = TCP_METRICS_GENL_NAME,
1114 .version = TCP_METRICS_GENL_VERSION,
1115 .maxattr = TCP_METRICS_ATTR_MAX,
1116 .netnsok = true,
1117 .module = THIS_MODULE,
1118 .ops = tcp_metrics_nl_ops,
1119 .n_ops = ARRAY_SIZE(tcp_metrics_nl_ops),
1120};
1121
1122static unsigned int tcpmhash_entries;
1123static int __init set_tcpmhash_entries(char *str)
1124{
1125 ssize_t ret;
1126
1127 if (!str)
1128 return 0;
1129
1130 ret = kstrtouint(str, 0, &tcpmhash_entries);
1131 if (ret)
1132 return 0;
1133
1134 return 1;
1135}
1136__setup("tcpmhash_entries=", set_tcpmhash_entries);
1137
1138static int __net_init tcp_net_metrics_init(struct net *net)
1139{
1140 size_t size;
1141 unsigned int slots;
1142
1143 if (!net_eq(net, &init_net))
1144 return 0;
1145
1146 slots = tcpmhash_entries;
1147 if (!slots) {
1148 if (totalram_pages >= 128 * 1024)
1149 slots = 16 * 1024;
1150 else
1151 slots = 8 * 1024;
1152 }
1153
1154 tcp_metrics_hash_log = order_base_2(slots);
1155 size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1156
1157 tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1158 if (!tcp_metrics_hash)
1159 tcp_metrics_hash = vzalloc(size);
1160
1161 if (!tcp_metrics_hash)
1162 return -ENOMEM;
1163
1164 return 0;
1165}
1166
1167static void __net_exit tcp_net_metrics_exit(struct net *net)
1168{
1169 tcp_metrics_flush_all(net);
1170}
1171
1172static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1173 .init = tcp_net_metrics_init,
1174 .exit = tcp_net_metrics_exit,
1175};
1176
1177void __init tcp_metrics_init(void)
1178{
1179 int ret;
1180
1181 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1182 if (ret < 0)
1183 panic("Could not allocate the tcp_metrics hash table\n");
1184
1185 ret = genl_register_family(&tcp_metrics_nl_family);
1186 if (ret < 0)
1187 panic("Could not register tcp_metrics generic netlink\n");
1188}