1// SPDX-License-Identifier: GPL-2.0
2#include <linux/rcupdate.h>
3#include <linux/spinlock.h>
4#include <linux/jiffies.h>
5#include <linux/module.h>
6#include <linux/cache.h>
7#include <linux/slab.h>
8#include <linux/init.h>
9#include <linux/tcp.h>
10#include <linux/hash.h>
11#include <linux/tcp_metrics.h>
12#include <linux/vmalloc.h>
13
14#include <net/inet_connection_sock.h>
15#include <net/net_namespace.h>
16#include <net/request_sock.h>
17#include <net/inetpeer.h>
18#include <net/sock.h>
19#include <net/ipv6.h>
20#include <net/dst.h>
21#include <net/tcp.h>
22#include <net/genetlink.h>
23
24static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
25 const struct inetpeer_addr *daddr,
26 struct net *net, unsigned int hash);
27
28struct tcp_fastopen_metrics {
29 u16 mss;
30 u16 syn_loss:10, /* Recurring Fast Open SYN losses */
31 try_exp:2; /* Request w/ exp. option (once) */
32 unsigned long last_syn_loss; /* Last Fast Open SYN loss */
33 struct tcp_fastopen_cookie cookie;
34};
35
36/* TCP_METRIC_MAX includes 2 extra fields for userspace compatibility
37 * Kernel only stores RTT and RTTVAR in usec resolution
38 */
39#define TCP_METRIC_MAX_KERNEL (TCP_METRIC_MAX - 2)
40
41struct tcp_metrics_block {
42 struct tcp_metrics_block __rcu *tcpm_next;
43 struct net *tcpm_net;
44 struct inetpeer_addr tcpm_saddr;
45 struct inetpeer_addr tcpm_daddr;
46 unsigned long tcpm_stamp;
47 u32 tcpm_lock;
48 u32 tcpm_vals[TCP_METRIC_MAX_KERNEL + 1];
49 struct tcp_fastopen_metrics tcpm_fastopen;
50
51 struct rcu_head rcu_head;
52};
53
54static inline struct net *tm_net(const struct tcp_metrics_block *tm)
55{
56 /* Paired with the WRITE_ONCE() in tcpm_new() */
57 return READ_ONCE(tm->tcpm_net);
58}
59
60static bool tcp_metric_locked(struct tcp_metrics_block *tm,
61 enum tcp_metric_index idx)
62{
63 /* Paired with WRITE_ONCE() in tcpm_suck_dst() */
64 return READ_ONCE(tm->tcpm_lock) & (1 << idx);
65}
66
67static u32 tcp_metric_get(const struct tcp_metrics_block *tm,
68 enum tcp_metric_index idx)
69{
70 /* Paired with WRITE_ONCE() in tcp_metric_set() */
71 return READ_ONCE(tm->tcpm_vals[idx]);
72}
73
74static void tcp_metric_set(struct tcp_metrics_block *tm,
75 enum tcp_metric_index idx,
76 u32 val)
77{
78 /* Paired with READ_ONCE() in tcp_metric_get() */
79 WRITE_ONCE(tm->tcpm_vals[idx], val);
80}
81
82static bool addr_same(const struct inetpeer_addr *a,
83 const struct inetpeer_addr *b)
84{
85 return (a->family == b->family) && !inetpeer_addr_cmp(a, b);
86}
87
88struct tcpm_hash_bucket {
89 struct tcp_metrics_block __rcu *chain;
90};
91
92static struct tcpm_hash_bucket *tcp_metrics_hash __read_mostly;
93static unsigned int tcp_metrics_hash_log __read_mostly;
94
95static DEFINE_SPINLOCK(tcp_metrics_lock);
96static DEFINE_SEQLOCK(fastopen_seqlock);
97
98static void tcpm_suck_dst(struct tcp_metrics_block *tm,
99 const struct dst_entry *dst,
100 bool fastopen_clear)
101{
102 u32 msval;
103 u32 val;
104
105 WRITE_ONCE(tm->tcpm_stamp, jiffies);
106
107 val = 0;
108 if (dst_metric_locked(dst, RTAX_RTT))
109 val |= 1 << TCP_METRIC_RTT;
110 if (dst_metric_locked(dst, RTAX_RTTVAR))
111 val |= 1 << TCP_METRIC_RTTVAR;
112 if (dst_metric_locked(dst, RTAX_SSTHRESH))
113 val |= 1 << TCP_METRIC_SSTHRESH;
114 if (dst_metric_locked(dst, RTAX_CWND))
115 val |= 1 << TCP_METRIC_CWND;
116 if (dst_metric_locked(dst, RTAX_REORDERING))
117 val |= 1 << TCP_METRIC_REORDERING;
118 /* Paired with READ_ONCE() in tcp_metric_locked() */
119 WRITE_ONCE(tm->tcpm_lock, val);
120
121 msval = dst_metric_raw(dst, RTAX_RTT);
122 tcp_metric_set(tm, idx: TCP_METRIC_RTT, val: msval * USEC_PER_MSEC);
123
124 msval = dst_metric_raw(dst, RTAX_RTTVAR);
125 tcp_metric_set(tm, idx: TCP_METRIC_RTTVAR, val: msval * USEC_PER_MSEC);
126 tcp_metric_set(tm, idx: TCP_METRIC_SSTHRESH,
127 val: dst_metric_raw(dst, RTAX_SSTHRESH));
128 tcp_metric_set(tm, idx: TCP_METRIC_CWND,
129 val: dst_metric_raw(dst, RTAX_CWND));
130 tcp_metric_set(tm, idx: TCP_METRIC_REORDERING,
131 val: dst_metric_raw(dst, RTAX_REORDERING));
132 if (fastopen_clear) {
133 write_seqlock(sl: &fastopen_seqlock);
134 tm->tcpm_fastopen.mss = 0;
135 tm->tcpm_fastopen.syn_loss = 0;
136 tm->tcpm_fastopen.try_exp = 0;
137 tm->tcpm_fastopen.cookie.exp = false;
138 tm->tcpm_fastopen.cookie.len = 0;
139 write_sequnlock(sl: &fastopen_seqlock);
140 }
141}
142
143#define TCP_METRICS_TIMEOUT (60 * 60 * HZ)
144
145static void tcpm_check_stamp(struct tcp_metrics_block *tm,
146 const struct dst_entry *dst)
147{
148 unsigned long limit;
149
150 if (!tm)
151 return;
152 limit = READ_ONCE(tm->tcpm_stamp) + TCP_METRICS_TIMEOUT;
153 if (unlikely(time_after(jiffies, limit)))
154 tcpm_suck_dst(tm, dst, fastopen_clear: false);
155}
156
157#define TCP_METRICS_RECLAIM_DEPTH 5
158#define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL
159
160#define deref_locked(p) \
161 rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
162
163static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
164 struct inetpeer_addr *saddr,
165 struct inetpeer_addr *daddr,
166 unsigned int hash)
167{
168 struct tcp_metrics_block *tm;
169 struct net *net;
170 bool reclaim = false;
171
172 spin_lock_bh(lock: &tcp_metrics_lock);
173 net = dev_net(dev: dst->dev);
174
175 /* While waiting for the spin-lock the cache might have been populated
176 * with this entry and so we have to check again.
177 */
178 tm = __tcp_get_metrics(saddr, daddr, net, hash);
179 if (tm == TCP_METRICS_RECLAIM_PTR) {
180 reclaim = true;
181 tm = NULL;
182 }
183 if (tm) {
184 tcpm_check_stamp(tm, dst);
185 goto out_unlock;
186 }
187
188 if (unlikely(reclaim)) {
189 struct tcp_metrics_block *oldest;
190
191 oldest = deref_locked(tcp_metrics_hash[hash].chain);
192 for (tm = deref_locked(oldest->tcpm_next); tm;
193 tm = deref_locked(tm->tcpm_next)) {
194 if (time_before(READ_ONCE(tm->tcpm_stamp),
195 READ_ONCE(oldest->tcpm_stamp)))
196 oldest = tm;
197 }
198 tm = oldest;
199 } else {
200 tm = kzalloc(size: sizeof(*tm), GFP_ATOMIC);
201 if (!tm)
202 goto out_unlock;
203 }
204 /* Paired with the READ_ONCE() in tm_net() */
205 WRITE_ONCE(tm->tcpm_net, net);
206
207 tm->tcpm_saddr = *saddr;
208 tm->tcpm_daddr = *daddr;
209
210 tcpm_suck_dst(tm, dst, fastopen_clear: reclaim);
211
212 if (likely(!reclaim)) {
213 tm->tcpm_next = tcp_metrics_hash[hash].chain;
214 rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
215 }
216
217out_unlock:
218 spin_unlock_bh(lock: &tcp_metrics_lock);
219 return tm;
220}
221
222static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
223{
224 if (tm)
225 return tm;
226 if (depth > TCP_METRICS_RECLAIM_DEPTH)
227 return TCP_METRICS_RECLAIM_PTR;
228 return NULL;
229}
230
231static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
232 const struct inetpeer_addr *daddr,
233 struct net *net, unsigned int hash)
234{
235 struct tcp_metrics_block *tm;
236 int depth = 0;
237
238 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
239 tm = rcu_dereference(tm->tcpm_next)) {
240 if (addr_same(a: &tm->tcpm_saddr, b: saddr) &&
241 addr_same(a: &tm->tcpm_daddr, b: daddr) &&
242 net_eq(net1: tm_net(tm), net2: net))
243 break;
244 depth++;
245 }
246 return tcp_get_encode(tm, depth);
247}
248
249static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
250 struct dst_entry *dst)
251{
252 struct tcp_metrics_block *tm;
253 struct inetpeer_addr saddr, daddr;
254 unsigned int hash;
255 struct net *net;
256
257 saddr.family = req->rsk_ops->family;
258 daddr.family = req->rsk_ops->family;
259 switch (daddr.family) {
260 case AF_INET:
261 inetpeer_set_addr_v4(iaddr: &saddr, ip: inet_rsk(sk: req)->ir_loc_addr);
262 inetpeer_set_addr_v4(iaddr: &daddr, ip: inet_rsk(sk: req)->ir_rmt_addr);
263 hash = ipv4_addr_hash(ip: inet_rsk(sk: req)->ir_rmt_addr);
264 break;
265#if IS_ENABLED(CONFIG_IPV6)
266 case AF_INET6:
267 inetpeer_set_addr_v6(iaddr: &saddr, in6: &inet_rsk(sk: req)->ir_v6_loc_addr);
268 inetpeer_set_addr_v6(iaddr: &daddr, in6: &inet_rsk(sk: req)->ir_v6_rmt_addr);
269 hash = ipv6_addr_hash(a: &inet_rsk(sk: req)->ir_v6_rmt_addr);
270 break;
271#endif
272 default:
273 return NULL;
274 }
275
276 net = dev_net(dev: dst->dev);
277 hash ^= net_hash_mix(net);
278 hash = hash_32(val: hash, bits: tcp_metrics_hash_log);
279
280 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
281 tm = rcu_dereference(tm->tcpm_next)) {
282 if (addr_same(a: &tm->tcpm_saddr, b: &saddr) &&
283 addr_same(a: &tm->tcpm_daddr, b: &daddr) &&
284 net_eq(net1: tm_net(tm), net2: net))
285 break;
286 }
287 tcpm_check_stamp(tm, dst);
288 return tm;
289}
290
291static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
292 struct dst_entry *dst,
293 bool create)
294{
295 struct tcp_metrics_block *tm;
296 struct inetpeer_addr saddr, daddr;
297 unsigned int hash;
298 struct net *net;
299
300 if (sk->sk_family == AF_INET) {
301 inetpeer_set_addr_v4(iaddr: &saddr, inet_sk(sk)->inet_saddr);
302 inetpeer_set_addr_v4(iaddr: &daddr, inet_sk(sk)->inet_daddr);
303 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
304 }
305#if IS_ENABLED(CONFIG_IPV6)
306 else if (sk->sk_family == AF_INET6) {
307 if (ipv6_addr_v4mapped(a: &sk->sk_v6_daddr)) {
308 inetpeer_set_addr_v4(iaddr: &saddr, inet_sk(sk)->inet_saddr);
309 inetpeer_set_addr_v4(iaddr: &daddr, inet_sk(sk)->inet_daddr);
310 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
311 } else {
312 inetpeer_set_addr_v6(iaddr: &saddr, in6: &sk->sk_v6_rcv_saddr);
313 inetpeer_set_addr_v6(iaddr: &daddr, in6: &sk->sk_v6_daddr);
314 hash = ipv6_addr_hash(a: &sk->sk_v6_daddr);
315 }
316 }
317#endif
318 else
319 return NULL;
320
321 net = dev_net(dev: dst->dev);
322 hash ^= net_hash_mix(net);
323 hash = hash_32(val: hash, bits: tcp_metrics_hash_log);
324
325 tm = __tcp_get_metrics(saddr: &saddr, daddr: &daddr, net, hash);
326 if (tm == TCP_METRICS_RECLAIM_PTR)
327 tm = NULL;
328 if (!tm && create)
329 tm = tcpm_new(dst, saddr: &saddr, daddr: &daddr, hash);
330 else
331 tcpm_check_stamp(tm, dst);
332
333 return tm;
334}
335
336/* Save metrics learned by this TCP session. This function is called
337 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
338 * or goes from LAST-ACK to CLOSE.
339 */
340void tcp_update_metrics(struct sock *sk)
341{
342 const struct inet_connection_sock *icsk = inet_csk(sk);
343 struct dst_entry *dst = __sk_dst_get(sk);
344 struct tcp_sock *tp = tcp_sk(sk);
345 struct net *net = sock_net(sk);
346 struct tcp_metrics_block *tm;
347 unsigned long rtt;
348 u32 val;
349 int m;
350
351 sk_dst_confirm(sk);
352 if (READ_ONCE(net->ipv4.sysctl_tcp_nometrics_save) || !dst)
353 return;
354
355 rcu_read_lock();
356 if (icsk->icsk_backoff || !tp->srtt_us) {
357 /* This session failed to estimate rtt. Why?
358 * Probably, no packets returned in time. Reset our
359 * results.
360 */
361 tm = tcp_get_metrics(sk, dst, create: false);
362 if (tm && !tcp_metric_locked(tm, idx: TCP_METRIC_RTT))
363 tcp_metric_set(tm, idx: TCP_METRIC_RTT, val: 0);
364 goto out_unlock;
365 } else
366 tm = tcp_get_metrics(sk, dst, create: true);
367
368 if (!tm)
369 goto out_unlock;
370
371 rtt = tcp_metric_get(tm, idx: TCP_METRIC_RTT);
372 m = rtt - tp->srtt_us;
373
374 /* If newly calculated rtt larger than stored one, store new
375 * one. Otherwise, use EWMA. Remember, rtt overestimation is
376 * always better than underestimation.
377 */
378 if (!tcp_metric_locked(tm, idx: TCP_METRIC_RTT)) {
379 if (m <= 0)
380 rtt = tp->srtt_us;
381 else
382 rtt -= (m >> 3);
383 tcp_metric_set(tm, idx: TCP_METRIC_RTT, val: rtt);
384 }
385
386 if (!tcp_metric_locked(tm, idx: TCP_METRIC_RTTVAR)) {
387 unsigned long var;
388
389 if (m < 0)
390 m = -m;
391
392 /* Scale deviation to rttvar fixed point */
393 m >>= 1;
394 if (m < tp->mdev_us)
395 m = tp->mdev_us;
396
397 var = tcp_metric_get(tm, idx: TCP_METRIC_RTTVAR);
398 if (m >= var)
399 var = m;
400 else
401 var -= (var - m) >> 2;
402
403 tcp_metric_set(tm, idx: TCP_METRIC_RTTVAR, val: var);
404 }
405
406 if (tcp_in_initial_slowstart(tp)) {
407 /* Slow start still did not finish. */
408 if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
409 !tcp_metric_locked(tm, idx: TCP_METRIC_SSTHRESH)) {
410 val = tcp_metric_get(tm, idx: TCP_METRIC_SSTHRESH);
411 if (val && (tcp_snd_cwnd(tp) >> 1) > val)
412 tcp_metric_set(tm, idx: TCP_METRIC_SSTHRESH,
413 val: tcp_snd_cwnd(tp) >> 1);
414 }
415 if (!tcp_metric_locked(tm, idx: TCP_METRIC_CWND)) {
416 val = tcp_metric_get(tm, idx: TCP_METRIC_CWND);
417 if (tcp_snd_cwnd(tp) > val)
418 tcp_metric_set(tm, idx: TCP_METRIC_CWND,
419 val: tcp_snd_cwnd(tp));
420 }
421 } else if (!tcp_in_slow_start(tp) &&
422 icsk->icsk_ca_state == TCP_CA_Open) {
423 /* Cong. avoidance phase, cwnd is reliable. */
424 if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
425 !tcp_metric_locked(tm, idx: TCP_METRIC_SSTHRESH))
426 tcp_metric_set(tm, idx: TCP_METRIC_SSTHRESH,
427 max(tcp_snd_cwnd(tp) >> 1, tp->snd_ssthresh));
428 if (!tcp_metric_locked(tm, idx: TCP_METRIC_CWND)) {
429 val = tcp_metric_get(tm, idx: TCP_METRIC_CWND);
430 tcp_metric_set(tm, idx: TCP_METRIC_CWND, val: (val + tcp_snd_cwnd(tp)) >> 1);
431 }
432 } else {
433 /* Else slow start did not finish, cwnd is non-sense,
434 * ssthresh may be also invalid.
435 */
436 if (!tcp_metric_locked(tm, idx: TCP_METRIC_CWND)) {
437 val = tcp_metric_get(tm, idx: TCP_METRIC_CWND);
438 tcp_metric_set(tm, idx: TCP_METRIC_CWND,
439 val: (val + tp->snd_ssthresh) >> 1);
440 }
441 if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
442 !tcp_metric_locked(tm, idx: TCP_METRIC_SSTHRESH)) {
443 val = tcp_metric_get(tm, idx: TCP_METRIC_SSTHRESH);
444 if (val && tp->snd_ssthresh > val)
445 tcp_metric_set(tm, idx: TCP_METRIC_SSTHRESH,
446 val: tp->snd_ssthresh);
447 }
448 if (!tcp_metric_locked(tm, idx: TCP_METRIC_REORDERING)) {
449 val = tcp_metric_get(tm, idx: TCP_METRIC_REORDERING);
450 if (val < tp->reordering &&
451 tp->reordering !=
452 READ_ONCE(net->ipv4.sysctl_tcp_reordering))
453 tcp_metric_set(tm, idx: TCP_METRIC_REORDERING,
454 val: tp->reordering);
455 }
456 }
457 WRITE_ONCE(tm->tcpm_stamp, jiffies);
458out_unlock:
459 rcu_read_unlock();
460}
461
462/* Initialize metrics on socket. */
463
464void tcp_init_metrics(struct sock *sk)
465{
466 struct dst_entry *dst = __sk_dst_get(sk);
467 struct tcp_sock *tp = tcp_sk(sk);
468 struct net *net = sock_net(sk);
469 struct tcp_metrics_block *tm;
470 u32 val, crtt = 0; /* cached RTT scaled by 8 */
471
472 sk_dst_confirm(sk);
473 /* ssthresh may have been reduced unnecessarily during.
474 * 3WHS. Restore it back to its initial default.
475 */
476 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
477 if (!dst)
478 goto reset;
479
480 rcu_read_lock();
481 tm = tcp_get_metrics(sk, dst, create: false);
482 if (!tm) {
483 rcu_read_unlock();
484 goto reset;
485 }
486
487 if (tcp_metric_locked(tm, idx: TCP_METRIC_CWND))
488 tp->snd_cwnd_clamp = tcp_metric_get(tm, idx: TCP_METRIC_CWND);
489
490 val = READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) ?
491 0 : tcp_metric_get(tm, idx: TCP_METRIC_SSTHRESH);
492 if (val) {
493 tp->snd_ssthresh = val;
494 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
495 tp->snd_ssthresh = tp->snd_cwnd_clamp;
496 }
497 val = tcp_metric_get(tm, idx: TCP_METRIC_REORDERING);
498 if (val && tp->reordering != val)
499 tp->reordering = val;
500
501 crtt = tcp_metric_get(tm, idx: TCP_METRIC_RTT);
502 rcu_read_unlock();
503reset:
504 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
505 * to seed the RTO for later data packets because SYN packets are
506 * small. Use the per-dst cached values to seed the RTO but keep
507 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
508 * Later the RTO will be updated immediately upon obtaining the first
509 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
510 * influences the first RTO but not later RTT estimation.
511 *
512 * But if RTT is not available from the SYN (due to retransmits or
513 * syn cookies) or the cache, force a conservative 3secs timeout.
514 *
515 * A bit of theory. RTT is time passed after "normal" sized packet
516 * is sent until it is ACKed. In normal circumstances sending small
517 * packets force peer to delay ACKs and calculation is correct too.
518 * The algorithm is adaptive and, provided we follow specs, it
519 * NEVER underestimate RTT. BUT! If peer tries to make some clever
520 * tricks sort of "quick acks" for time long enough to decrease RTT
521 * to low value, and then abruptly stops to do it and starts to delay
522 * ACKs, wait for troubles.
523 */
524 if (crtt > tp->srtt_us) {
525 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
526 crtt /= 8 * USEC_PER_SEC / HZ;
527 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
528 } else if (tp->srtt_us == 0) {
529 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
530 * 3WHS. This is most likely due to retransmission,
531 * including spurious one. Reset the RTO back to 3secs
532 * from the more aggressive 1sec to avoid more spurious
533 * retransmission.
534 */
535 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
536 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
537
538 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
539 }
540}
541
542bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
543{
544 struct tcp_metrics_block *tm;
545 bool ret;
546
547 if (!dst)
548 return false;
549
550 rcu_read_lock();
551 tm = __tcp_get_metrics_req(req, dst);
552 if (tm && tcp_metric_get(tm, idx: TCP_METRIC_RTT))
553 ret = true;
554 else
555 ret = false;
556 rcu_read_unlock();
557
558 return ret;
559}
560
561void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
562 struct tcp_fastopen_cookie *cookie)
563{
564 struct tcp_metrics_block *tm;
565
566 rcu_read_lock();
567 tm = tcp_get_metrics(sk, dst: __sk_dst_get(sk), create: false);
568 if (tm) {
569 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
570 unsigned int seq;
571
572 do {
573 seq = read_seqbegin(sl: &fastopen_seqlock);
574 if (tfom->mss)
575 *mss = tfom->mss;
576 *cookie = tfom->cookie;
577 if (cookie->len <= 0 && tfom->try_exp == 1)
578 cookie->exp = true;
579 } while (read_seqretry(sl: &fastopen_seqlock, start: seq));
580 }
581 rcu_read_unlock();
582}
583
584void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
585 struct tcp_fastopen_cookie *cookie, bool syn_lost,
586 u16 try_exp)
587{
588 struct dst_entry *dst = __sk_dst_get(sk);
589 struct tcp_metrics_block *tm;
590
591 if (!dst)
592 return;
593 rcu_read_lock();
594 tm = tcp_get_metrics(sk, dst, create: true);
595 if (tm) {
596 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
597
598 write_seqlock_bh(sl: &fastopen_seqlock);
599 if (mss)
600 tfom->mss = mss;
601 if (cookie && cookie->len > 0)
602 tfom->cookie = *cookie;
603 else if (try_exp > tfom->try_exp &&
604 tfom->cookie.len <= 0 && !tfom->cookie.exp)
605 tfom->try_exp = try_exp;
606 if (syn_lost) {
607 ++tfom->syn_loss;
608 tfom->last_syn_loss = jiffies;
609 } else
610 tfom->syn_loss = 0;
611 write_sequnlock_bh(sl: &fastopen_seqlock);
612 }
613 rcu_read_unlock();
614}
615
616static struct genl_family tcp_metrics_nl_family;
617
618static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
619 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
620 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
621 .len = sizeof(struct in6_addr), },
622 /* Following attributes are not received for GET/DEL,
623 * we keep them for reference
624 */
625#if 0
626 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
627 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
628 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
629 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
630 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
631 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
632 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
633 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
634 .len = TCP_FASTOPEN_COOKIE_MAX, },
635#endif
636};
637
638/* Add attributes, caller cancels its header on failure */
639static int tcp_metrics_fill_info(struct sk_buff *msg,
640 struct tcp_metrics_block *tm)
641{
642 struct nlattr *nest;
643 int i;
644
645 switch (tm->tcpm_daddr.family) {
646 case AF_INET:
647 if (nla_put_in_addr(skb: msg, attrtype: TCP_METRICS_ATTR_ADDR_IPV4,
648 addr: inetpeer_get_addr_v4(iaddr: &tm->tcpm_daddr)) < 0)
649 goto nla_put_failure;
650 if (nla_put_in_addr(skb: msg, attrtype: TCP_METRICS_ATTR_SADDR_IPV4,
651 addr: inetpeer_get_addr_v4(iaddr: &tm->tcpm_saddr)) < 0)
652 goto nla_put_failure;
653 break;
654 case AF_INET6:
655 if (nla_put_in6_addr(skb: msg, attrtype: TCP_METRICS_ATTR_ADDR_IPV6,
656 addr: inetpeer_get_addr_v6(iaddr: &tm->tcpm_daddr)) < 0)
657 goto nla_put_failure;
658 if (nla_put_in6_addr(skb: msg, attrtype: TCP_METRICS_ATTR_SADDR_IPV6,
659 addr: inetpeer_get_addr_v6(iaddr: &tm->tcpm_saddr)) < 0)
660 goto nla_put_failure;
661 break;
662 default:
663 return -EAFNOSUPPORT;
664 }
665
666 if (nla_put_msecs(skb: msg, attrtype: TCP_METRICS_ATTR_AGE,
667 njiffies: jiffies - READ_ONCE(tm->tcpm_stamp),
668 padattr: TCP_METRICS_ATTR_PAD) < 0)
669 goto nla_put_failure;
670
671 {
672 int n = 0;
673
674 nest = nla_nest_start_noflag(skb: msg, attrtype: TCP_METRICS_ATTR_VALS);
675 if (!nest)
676 goto nla_put_failure;
677 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
678 u32 val = tcp_metric_get(tm, idx: i);
679
680 if (!val)
681 continue;
682 if (i == TCP_METRIC_RTT) {
683 if (nla_put_u32(skb: msg, attrtype: TCP_METRIC_RTT_US + 1,
684 value: val) < 0)
685 goto nla_put_failure;
686 n++;
687 val = max(val / 1000, 1U);
688 }
689 if (i == TCP_METRIC_RTTVAR) {
690 if (nla_put_u32(skb: msg, attrtype: TCP_METRIC_RTTVAR_US + 1,
691 value: val) < 0)
692 goto nla_put_failure;
693 n++;
694 val = max(val / 1000, 1U);
695 }
696 if (nla_put_u32(skb: msg, attrtype: i + 1, value: val) < 0)
697 goto nla_put_failure;
698 n++;
699 }
700 if (n)
701 nla_nest_end(skb: msg, start: nest);
702 else
703 nla_nest_cancel(skb: msg, start: nest);
704 }
705
706 {
707 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
708 unsigned int seq;
709
710 do {
711 seq = read_seqbegin(sl: &fastopen_seqlock);
712 tfom_copy[0] = tm->tcpm_fastopen;
713 } while (read_seqretry(sl: &fastopen_seqlock, start: seq));
714
715 tfom = tfom_copy;
716 if (tfom->mss &&
717 nla_put_u16(skb: msg, attrtype: TCP_METRICS_ATTR_FOPEN_MSS,
718 value: tfom->mss) < 0)
719 goto nla_put_failure;
720 if (tfom->syn_loss &&
721 (nla_put_u16(skb: msg, attrtype: TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
722 value: tfom->syn_loss) < 0 ||
723 nla_put_msecs(skb: msg, attrtype: TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
724 njiffies: jiffies - tfom->last_syn_loss,
725 padattr: TCP_METRICS_ATTR_PAD) < 0))
726 goto nla_put_failure;
727 if (tfom->cookie.len > 0 &&
728 nla_put(skb: msg, attrtype: TCP_METRICS_ATTR_FOPEN_COOKIE,
729 attrlen: tfom->cookie.len, data: tfom->cookie.val) < 0)
730 goto nla_put_failure;
731 }
732
733 return 0;
734
735nla_put_failure:
736 return -EMSGSIZE;
737}
738
739static int tcp_metrics_dump_info(struct sk_buff *skb,
740 struct netlink_callback *cb,
741 struct tcp_metrics_block *tm)
742{
743 void *hdr;
744
745 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, seq: cb->nlh->nlmsg_seq,
746 family: &tcp_metrics_nl_family, NLM_F_MULTI,
747 cmd: TCP_METRICS_CMD_GET);
748 if (!hdr)
749 return -EMSGSIZE;
750
751 if (tcp_metrics_fill_info(msg: skb, tm) < 0)
752 goto nla_put_failure;
753
754 genlmsg_end(skb, hdr);
755 return 0;
756
757nla_put_failure:
758 genlmsg_cancel(skb, hdr);
759 return -EMSGSIZE;
760}
761
762static int tcp_metrics_nl_dump(struct sk_buff *skb,
763 struct netlink_callback *cb)
764{
765 struct net *net = sock_net(sk: skb->sk);
766 unsigned int max_rows = 1U << tcp_metrics_hash_log;
767 unsigned int row, s_row = cb->args[0];
768 int s_col = cb->args[1], col = s_col;
769
770 for (row = s_row; row < max_rows; row++, s_col = 0) {
771 struct tcp_metrics_block *tm;
772 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
773
774 rcu_read_lock();
775 for (col = 0, tm = rcu_dereference(hb->chain); tm;
776 tm = rcu_dereference(tm->tcpm_next), col++) {
777 if (!net_eq(net1: tm_net(tm), net2: net))
778 continue;
779 if (col < s_col)
780 continue;
781 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
782 rcu_read_unlock();
783 goto done;
784 }
785 }
786 rcu_read_unlock();
787 }
788
789done:
790 cb->args[0] = row;
791 cb->args[1] = col;
792 return skb->len;
793}
794
795static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
796 unsigned int *hash, int optional, int v4, int v6)
797{
798 struct nlattr *a;
799
800 a = info->attrs[v4];
801 if (a) {
802 inetpeer_set_addr_v4(iaddr: addr, ip: nla_get_in_addr(nla: a));
803 if (hash)
804 *hash = ipv4_addr_hash(ip: inetpeer_get_addr_v4(iaddr: addr));
805 return 0;
806 }
807 a = info->attrs[v6];
808 if (a) {
809 struct in6_addr in6;
810
811 if (nla_len(nla: a) != sizeof(struct in6_addr))
812 return -EINVAL;
813 in6 = nla_get_in6_addr(nla: a);
814 inetpeer_set_addr_v6(iaddr: addr, in6: &in6);
815 if (hash)
816 *hash = ipv6_addr_hash(a: inetpeer_get_addr_v6(iaddr: addr));
817 return 0;
818 }
819 return optional ? 1 : -EAFNOSUPPORT;
820}
821
822static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
823 unsigned int *hash, int optional)
824{
825 return __parse_nl_addr(info, addr, hash, optional,
826 v4: TCP_METRICS_ATTR_ADDR_IPV4,
827 v6: TCP_METRICS_ATTR_ADDR_IPV6);
828}
829
830static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
831{
832 return __parse_nl_addr(info, addr, NULL, optional: 0,
833 v4: TCP_METRICS_ATTR_SADDR_IPV4,
834 v6: TCP_METRICS_ATTR_SADDR_IPV6);
835}
836
837static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
838{
839 struct tcp_metrics_block *tm;
840 struct inetpeer_addr saddr, daddr;
841 unsigned int hash;
842 struct sk_buff *msg;
843 struct net *net = genl_info_net(info);
844 void *reply;
845 int ret;
846 bool src = true;
847
848 ret = parse_nl_addr(info, addr: &daddr, hash: &hash, optional: 0);
849 if (ret < 0)
850 return ret;
851
852 ret = parse_nl_saddr(info, addr: &saddr);
853 if (ret < 0)
854 src = false;
855
856 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
857 if (!msg)
858 return -ENOMEM;
859
860 reply = genlmsg_put_reply(skb: msg, info, family: &tcp_metrics_nl_family, flags: 0,
861 cmd: info->genlhdr->cmd);
862 if (!reply)
863 goto nla_put_failure;
864
865 hash ^= net_hash_mix(net);
866 hash = hash_32(val: hash, bits: tcp_metrics_hash_log);
867 ret = -ESRCH;
868 rcu_read_lock();
869 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
870 tm = rcu_dereference(tm->tcpm_next)) {
871 if (addr_same(a: &tm->tcpm_daddr, b: &daddr) &&
872 (!src || addr_same(a: &tm->tcpm_saddr, b: &saddr)) &&
873 net_eq(net1: tm_net(tm), net2: net)) {
874 ret = tcp_metrics_fill_info(msg, tm);
875 break;
876 }
877 }
878 rcu_read_unlock();
879 if (ret < 0)
880 goto out_free;
881
882 genlmsg_end(skb: msg, hdr: reply);
883 return genlmsg_reply(skb: msg, info);
884
885nla_put_failure:
886 ret = -EMSGSIZE;
887
888out_free:
889 nlmsg_free(skb: msg);
890 return ret;
891}
892
893static void tcp_metrics_flush_all(struct net *net)
894{
895 unsigned int max_rows = 1U << tcp_metrics_hash_log;
896 struct tcpm_hash_bucket *hb = tcp_metrics_hash;
897 struct tcp_metrics_block *tm;
898 unsigned int row;
899
900 for (row = 0; row < max_rows; row++, hb++) {
901 struct tcp_metrics_block __rcu **pp = &hb->chain;
902 bool match;
903
904 if (!rcu_access_pointer(*pp))
905 continue;
906
907 spin_lock_bh(lock: &tcp_metrics_lock);
908 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
909 match = net ? net_eq(net1: tm_net(tm), net2: net) :
910 !refcount_read(r: &tm_net(tm)->ns.count);
911 if (match) {
912 rcu_assign_pointer(*pp, tm->tcpm_next);
913 kfree_rcu(tm, rcu_head);
914 } else {
915 pp = &tm->tcpm_next;
916 }
917 }
918 spin_unlock_bh(lock: &tcp_metrics_lock);
919 cond_resched();
920 }
921}
922
923static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
924{
925 struct tcpm_hash_bucket *hb;
926 struct tcp_metrics_block *tm;
927 struct tcp_metrics_block __rcu **pp;
928 struct inetpeer_addr saddr, daddr;
929 unsigned int hash;
930 struct net *net = genl_info_net(info);
931 int ret;
932 bool src = true, found = false;
933
934 ret = parse_nl_addr(info, addr: &daddr, hash: &hash, optional: 1);
935 if (ret < 0)
936 return ret;
937 if (ret > 0) {
938 tcp_metrics_flush_all(net);
939 return 0;
940 }
941 ret = parse_nl_saddr(info, addr: &saddr);
942 if (ret < 0)
943 src = false;
944
945 hash ^= net_hash_mix(net);
946 hash = hash_32(val: hash, bits: tcp_metrics_hash_log);
947 hb = tcp_metrics_hash + hash;
948 pp = &hb->chain;
949 spin_lock_bh(lock: &tcp_metrics_lock);
950 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
951 if (addr_same(a: &tm->tcpm_daddr, b: &daddr) &&
952 (!src || addr_same(a: &tm->tcpm_saddr, b: &saddr)) &&
953 net_eq(net1: tm_net(tm), net2: net)) {
954 rcu_assign_pointer(*pp, tm->tcpm_next);
955 kfree_rcu(tm, rcu_head);
956 found = true;
957 } else {
958 pp = &tm->tcpm_next;
959 }
960 }
961 spin_unlock_bh(lock: &tcp_metrics_lock);
962 if (!found)
963 return -ESRCH;
964 return 0;
965}
966
967static const struct genl_small_ops tcp_metrics_nl_ops[] = {
968 {
969 .cmd = TCP_METRICS_CMD_GET,
970 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
971 .doit = tcp_metrics_nl_cmd_get,
972 .dumpit = tcp_metrics_nl_dump,
973 },
974 {
975 .cmd = TCP_METRICS_CMD_DEL,
976 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
977 .doit = tcp_metrics_nl_cmd_del,
978 .flags = GENL_ADMIN_PERM,
979 },
980};
981
982static struct genl_family tcp_metrics_nl_family __ro_after_init = {
983 .hdrsize = 0,
984 .name = TCP_METRICS_GENL_NAME,
985 .version = TCP_METRICS_GENL_VERSION,
986 .maxattr = TCP_METRICS_ATTR_MAX,
987 .policy = tcp_metrics_nl_policy,
988 .netnsok = true,
989 .module = THIS_MODULE,
990 .small_ops = tcp_metrics_nl_ops,
991 .n_small_ops = ARRAY_SIZE(tcp_metrics_nl_ops),
992 .resv_start_op = TCP_METRICS_CMD_DEL + 1,
993};
994
995static unsigned int tcpmhash_entries __initdata;
996static int __init set_tcpmhash_entries(char *str)
997{
998 ssize_t ret;
999
1000 if (!str)
1001 return 0;
1002
1003 ret = kstrtouint(s: str, base: 0, res: &tcpmhash_entries);
1004 if (ret)
1005 return 0;
1006
1007 return 1;
1008}
1009__setup("tcpmhash_entries=", set_tcpmhash_entries);
1010
1011static void __init tcp_metrics_hash_alloc(void)
1012{
1013 unsigned int slots = tcpmhash_entries;
1014 size_t size;
1015
1016 if (!slots) {
1017 if (totalram_pages() >= 128 * 1024)
1018 slots = 16 * 1024;
1019 else
1020 slots = 8 * 1024;
1021 }
1022
1023 tcp_metrics_hash_log = order_base_2(slots);
1024 size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1025
1026 tcp_metrics_hash = kvzalloc(size, GFP_KERNEL);
1027 if (!tcp_metrics_hash)
1028 panic(fmt: "Could not allocate the tcp_metrics hash table\n");
1029}
1030
1031static void __net_exit tcp_net_metrics_exit_batch(struct list_head *net_exit_list)
1032{
1033 tcp_metrics_flush_all(NULL);
1034}
1035
1036static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1037 .exit_batch = tcp_net_metrics_exit_batch,
1038};
1039
1040void __init tcp_metrics_init(void)
1041{
1042 int ret;
1043
1044 tcp_metrics_hash_alloc();
1045
1046 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1047 if (ret < 0)
1048 panic(fmt: "Could not register tcp_net_metrics_ops\n");
1049
1050 ret = genl_register_family(family: &tcp_metrics_nl_family);
1051 if (ret < 0)
1052 panic(fmt: "Could not register tcp_metrics generic netlink\n");
1053}
1054

source code of linux/net/ipv4/tcp_metrics.c