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
24 static 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
28 struct 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
41 struct 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
tm_net(struct tcp_metrics_block * tm)54 static inline struct net *tm_net(struct tcp_metrics_block *tm)
55 {
56 return read_pnet(&tm->tcpm_net);
57 }
58
tcp_metric_locked(struct tcp_metrics_block * tm,enum tcp_metric_index idx)59 static 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
tcp_metric_get(struct tcp_metrics_block * tm,enum tcp_metric_index idx)65 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
66 enum tcp_metric_index idx)
67 {
68 return tm->tcpm_vals[idx];
69 }
70
tcp_metric_set(struct tcp_metrics_block * tm,enum tcp_metric_index idx,u32 val)71 static 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
addr_same(const struct inetpeer_addr * a,const struct inetpeer_addr * b)78 static 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
84 struct tcpm_hash_bucket {
85 struct tcp_metrics_block __rcu *chain;
86 };
87
88 static struct tcpm_hash_bucket *tcp_metrics_hash __read_mostly;
89 static unsigned int tcp_metrics_hash_log __read_mostly;
90
91 static DEFINE_SPINLOCK(tcp_metrics_lock);
92
tcpm_suck_dst(struct tcp_metrics_block * tm,const struct dst_entry * dst,bool fastopen_clear)93 static 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
tcpm_check_stamp(struct tcp_metrics_block * tm,struct dst_entry * dst)134 static 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
tcpm_new(struct dst_entry * dst,struct inetpeer_addr * saddr,struct inetpeer_addr * daddr,unsigned int hash)146 static 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
197 out_unlock:
198 spin_unlock_bh(&tcp_metrics_lock);
199 return tm;
200 }
201
tcp_get_encode(struct tcp_metrics_block * tm,int depth)202 static 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
__tcp_get_metrics(const struct inetpeer_addr * saddr,const struct inetpeer_addr * daddr,struct net * net,unsigned int hash)211 static 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
__tcp_get_metrics_req(struct request_sock * req,struct dst_entry * dst)229 static 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
tcp_get_metrics(struct sock * sk,struct dst_entry * dst,bool create)271 static 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 */
tcp_update_metrics(struct sock * sk)320 void 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 (READ_ONCE(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 (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
389 !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
390 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
391 if (val && (tcp_snd_cwnd(tp) >> 1) > val)
392 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
393 tcp_snd_cwnd(tp) >> 1);
394 }
395 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
396 val = tcp_metric_get(tm, TCP_METRIC_CWND);
397 if (tcp_snd_cwnd(tp) > val)
398 tcp_metric_set(tm, TCP_METRIC_CWND,
399 tcp_snd_cwnd(tp));
400 }
401 } else if (!tcp_in_slow_start(tp) &&
402 icsk->icsk_ca_state == TCP_CA_Open) {
403 /* Cong. avoidance phase, cwnd is reliable. */
404 if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
405 !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
406 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
407 max(tcp_snd_cwnd(tp) >> 1, tp->snd_ssthresh));
408 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
409 val = tcp_metric_get(tm, TCP_METRIC_CWND);
410 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tcp_snd_cwnd(tp)) >> 1);
411 }
412 } else {
413 /* Else slow start did not finish, cwnd is non-sense,
414 * ssthresh may be also invalid.
415 */
416 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
417 val = tcp_metric_get(tm, TCP_METRIC_CWND);
418 tcp_metric_set(tm, TCP_METRIC_CWND,
419 (val + tp->snd_ssthresh) >> 1);
420 }
421 if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
422 !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
423 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
424 if (val && tp->snd_ssthresh > val)
425 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
426 tp->snd_ssthresh);
427 }
428 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
429 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
430 if (val < tp->reordering &&
431 tp->reordering !=
432 READ_ONCE(net->ipv4.sysctl_tcp_reordering))
433 tcp_metric_set(tm, TCP_METRIC_REORDERING,
434 tp->reordering);
435 }
436 }
437 tm->tcpm_stamp = jiffies;
438 out_unlock:
439 rcu_read_unlock();
440 }
441
442 /* Initialize metrics on socket. */
443
tcp_init_metrics(struct sock * sk)444 void tcp_init_metrics(struct sock *sk)
445 {
446 struct dst_entry *dst = __sk_dst_get(sk);
447 struct tcp_sock *tp = tcp_sk(sk);
448 struct net *net = sock_net(sk);
449 struct tcp_metrics_block *tm;
450 u32 val, crtt = 0; /* cached RTT scaled by 8 */
451
452 sk_dst_confirm(sk);
453 if (!dst)
454 goto reset;
455
456 rcu_read_lock();
457 tm = tcp_get_metrics(sk, dst, true);
458 if (!tm) {
459 rcu_read_unlock();
460 goto reset;
461 }
462
463 if (tcp_metric_locked(tm, TCP_METRIC_CWND))
464 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
465
466 val = READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) ?
467 0 : tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
468 if (val) {
469 tp->snd_ssthresh = val;
470 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
471 tp->snd_ssthresh = tp->snd_cwnd_clamp;
472 } else {
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 }
478 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
479 if (val && tp->reordering != val)
480 tp->reordering = val;
481
482 crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
483 rcu_read_unlock();
484 reset:
485 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
486 * to seed the RTO for later data packets because SYN packets are
487 * small. Use the per-dst cached values to seed the RTO but keep
488 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
489 * Later the RTO will be updated immediately upon obtaining the first
490 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
491 * influences the first RTO but not later RTT estimation.
492 *
493 * But if RTT is not available from the SYN (due to retransmits or
494 * syn cookies) or the cache, force a conservative 3secs timeout.
495 *
496 * A bit of theory. RTT is time passed after "normal" sized packet
497 * is sent until it is ACKed. In normal circumstances sending small
498 * packets force peer to delay ACKs and calculation is correct too.
499 * The algorithm is adaptive and, provided we follow specs, it
500 * NEVER underestimate RTT. BUT! If peer tries to make some clever
501 * tricks sort of "quick acks" for time long enough to decrease RTT
502 * to low value, and then abruptly stops to do it and starts to delay
503 * ACKs, wait for troubles.
504 */
505 if (crtt > tp->srtt_us) {
506 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
507 crtt /= 8 * USEC_PER_SEC / HZ;
508 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
509 } else if (tp->srtt_us == 0) {
510 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
511 * 3WHS. This is most likely due to retransmission,
512 * including spurious one. Reset the RTO back to 3secs
513 * from the more aggressive 1sec to avoid more spurious
514 * retransmission.
515 */
516 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
517 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
518
519 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
520 }
521 }
522
tcp_peer_is_proven(struct request_sock * req,struct dst_entry * dst)523 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
524 {
525 struct tcp_metrics_block *tm;
526 bool ret;
527
528 if (!dst)
529 return false;
530
531 rcu_read_lock();
532 tm = __tcp_get_metrics_req(req, dst);
533 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT))
534 ret = true;
535 else
536 ret = false;
537 rcu_read_unlock();
538
539 return ret;
540 }
541
542 static DEFINE_SEQLOCK(fastopen_seqlock);
543
tcp_fastopen_cache_get(struct sock * sk,u16 * mss,struct tcp_fastopen_cookie * cookie)544 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
545 struct tcp_fastopen_cookie *cookie)
546 {
547 struct tcp_metrics_block *tm;
548
549 rcu_read_lock();
550 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
551 if (tm) {
552 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
553 unsigned int seq;
554
555 do {
556 seq = read_seqbegin(&fastopen_seqlock);
557 if (tfom->mss)
558 *mss = tfom->mss;
559 *cookie = tfom->cookie;
560 if (cookie->len <= 0 && tfom->try_exp == 1)
561 cookie->exp = true;
562 } while (read_seqretry(&fastopen_seqlock, seq));
563 }
564 rcu_read_unlock();
565 }
566
tcp_fastopen_cache_set(struct sock * sk,u16 mss,struct tcp_fastopen_cookie * cookie,bool syn_lost,u16 try_exp)567 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
568 struct tcp_fastopen_cookie *cookie, bool syn_lost,
569 u16 try_exp)
570 {
571 struct dst_entry *dst = __sk_dst_get(sk);
572 struct tcp_metrics_block *tm;
573
574 if (!dst)
575 return;
576 rcu_read_lock();
577 tm = tcp_get_metrics(sk, dst, true);
578 if (tm) {
579 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
580
581 write_seqlock_bh(&fastopen_seqlock);
582 if (mss)
583 tfom->mss = mss;
584 if (cookie && cookie->len > 0)
585 tfom->cookie = *cookie;
586 else if (try_exp > tfom->try_exp &&
587 tfom->cookie.len <= 0 && !tfom->cookie.exp)
588 tfom->try_exp = try_exp;
589 if (syn_lost) {
590 ++tfom->syn_loss;
591 tfom->last_syn_loss = jiffies;
592 } else
593 tfom->syn_loss = 0;
594 write_sequnlock_bh(&fastopen_seqlock);
595 }
596 rcu_read_unlock();
597 }
598
599 static struct genl_family tcp_metrics_nl_family;
600
601 static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
602 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
603 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
604 .len = sizeof(struct in6_addr), },
605 /* Following attributes are not received for GET/DEL,
606 * we keep them for reference
607 */
608 #if 0
609 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
610 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
611 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
612 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
613 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
614 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
615 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
616 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
617 .len = TCP_FASTOPEN_COOKIE_MAX, },
618 #endif
619 };
620
621 /* Add attributes, caller cancels its header on failure */
tcp_metrics_fill_info(struct sk_buff * msg,struct tcp_metrics_block * tm)622 static int tcp_metrics_fill_info(struct sk_buff *msg,
623 struct tcp_metrics_block *tm)
624 {
625 struct nlattr *nest;
626 int i;
627
628 switch (tm->tcpm_daddr.family) {
629 case AF_INET:
630 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
631 inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
632 goto nla_put_failure;
633 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
634 inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
635 goto nla_put_failure;
636 break;
637 case AF_INET6:
638 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
639 inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
640 goto nla_put_failure;
641 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
642 inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
643 goto nla_put_failure;
644 break;
645 default:
646 return -EAFNOSUPPORT;
647 }
648
649 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
650 jiffies - tm->tcpm_stamp,
651 TCP_METRICS_ATTR_PAD) < 0)
652 goto nla_put_failure;
653
654 {
655 int n = 0;
656
657 nest = nla_nest_start_noflag(msg, TCP_METRICS_ATTR_VALS);
658 if (!nest)
659 goto nla_put_failure;
660 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
661 u32 val = tm->tcpm_vals[i];
662
663 if (!val)
664 continue;
665 if (i == TCP_METRIC_RTT) {
666 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
667 val) < 0)
668 goto nla_put_failure;
669 n++;
670 val = max(val / 1000, 1U);
671 }
672 if (i == TCP_METRIC_RTTVAR) {
673 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
674 val) < 0)
675 goto nla_put_failure;
676 n++;
677 val = max(val / 1000, 1U);
678 }
679 if (nla_put_u32(msg, i + 1, val) < 0)
680 goto nla_put_failure;
681 n++;
682 }
683 if (n)
684 nla_nest_end(msg, nest);
685 else
686 nla_nest_cancel(msg, nest);
687 }
688
689 {
690 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
691 unsigned int seq;
692
693 do {
694 seq = read_seqbegin(&fastopen_seqlock);
695 tfom_copy[0] = tm->tcpm_fastopen;
696 } while (read_seqretry(&fastopen_seqlock, seq));
697
698 tfom = tfom_copy;
699 if (tfom->mss &&
700 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
701 tfom->mss) < 0)
702 goto nla_put_failure;
703 if (tfom->syn_loss &&
704 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
705 tfom->syn_loss) < 0 ||
706 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
707 jiffies - tfom->last_syn_loss,
708 TCP_METRICS_ATTR_PAD) < 0))
709 goto nla_put_failure;
710 if (tfom->cookie.len > 0 &&
711 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
712 tfom->cookie.len, tfom->cookie.val) < 0)
713 goto nla_put_failure;
714 }
715
716 return 0;
717
718 nla_put_failure:
719 return -EMSGSIZE;
720 }
721
tcp_metrics_dump_info(struct sk_buff * skb,struct netlink_callback * cb,struct tcp_metrics_block * tm)722 static int tcp_metrics_dump_info(struct sk_buff *skb,
723 struct netlink_callback *cb,
724 struct tcp_metrics_block *tm)
725 {
726 void *hdr;
727
728 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
729 &tcp_metrics_nl_family, NLM_F_MULTI,
730 TCP_METRICS_CMD_GET);
731 if (!hdr)
732 return -EMSGSIZE;
733
734 if (tcp_metrics_fill_info(skb, tm) < 0)
735 goto nla_put_failure;
736
737 genlmsg_end(skb, hdr);
738 return 0;
739
740 nla_put_failure:
741 genlmsg_cancel(skb, hdr);
742 return -EMSGSIZE;
743 }
744
tcp_metrics_nl_dump(struct sk_buff * skb,struct netlink_callback * cb)745 static int tcp_metrics_nl_dump(struct sk_buff *skb,
746 struct netlink_callback *cb)
747 {
748 struct net *net = sock_net(skb->sk);
749 unsigned int max_rows = 1U << tcp_metrics_hash_log;
750 unsigned int row, s_row = cb->args[0];
751 int s_col = cb->args[1], col = s_col;
752
753 for (row = s_row; row < max_rows; row++, s_col = 0) {
754 struct tcp_metrics_block *tm;
755 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
756
757 rcu_read_lock();
758 for (col = 0, tm = rcu_dereference(hb->chain); tm;
759 tm = rcu_dereference(tm->tcpm_next), col++) {
760 if (!net_eq(tm_net(tm), net))
761 continue;
762 if (col < s_col)
763 continue;
764 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
765 rcu_read_unlock();
766 goto done;
767 }
768 }
769 rcu_read_unlock();
770 }
771
772 done:
773 cb->args[0] = row;
774 cb->args[1] = col;
775 return skb->len;
776 }
777
__parse_nl_addr(struct genl_info * info,struct inetpeer_addr * addr,unsigned int * hash,int optional,int v4,int v6)778 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
779 unsigned int *hash, int optional, int v4, int v6)
780 {
781 struct nlattr *a;
782
783 a = info->attrs[v4];
784 if (a) {
785 inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
786 if (hash)
787 *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
788 return 0;
789 }
790 a = info->attrs[v6];
791 if (a) {
792 struct in6_addr in6;
793
794 if (nla_len(a) != sizeof(struct in6_addr))
795 return -EINVAL;
796 in6 = nla_get_in6_addr(a);
797 inetpeer_set_addr_v6(addr, &in6);
798 if (hash)
799 *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
800 return 0;
801 }
802 return optional ? 1 : -EAFNOSUPPORT;
803 }
804
parse_nl_addr(struct genl_info * info,struct inetpeer_addr * addr,unsigned int * hash,int optional)805 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
806 unsigned int *hash, int optional)
807 {
808 return __parse_nl_addr(info, addr, hash, optional,
809 TCP_METRICS_ATTR_ADDR_IPV4,
810 TCP_METRICS_ATTR_ADDR_IPV6);
811 }
812
parse_nl_saddr(struct genl_info * info,struct inetpeer_addr * addr)813 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
814 {
815 return __parse_nl_addr(info, addr, NULL, 0,
816 TCP_METRICS_ATTR_SADDR_IPV4,
817 TCP_METRICS_ATTR_SADDR_IPV6);
818 }
819
tcp_metrics_nl_cmd_get(struct sk_buff * skb,struct genl_info * info)820 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
821 {
822 struct tcp_metrics_block *tm;
823 struct inetpeer_addr saddr, daddr;
824 unsigned int hash;
825 struct sk_buff *msg;
826 struct net *net = genl_info_net(info);
827 void *reply;
828 int ret;
829 bool src = true;
830
831 ret = parse_nl_addr(info, &daddr, &hash, 0);
832 if (ret < 0)
833 return ret;
834
835 ret = parse_nl_saddr(info, &saddr);
836 if (ret < 0)
837 src = false;
838
839 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
840 if (!msg)
841 return -ENOMEM;
842
843 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
844 info->genlhdr->cmd);
845 if (!reply)
846 goto nla_put_failure;
847
848 hash ^= net_hash_mix(net);
849 hash = hash_32(hash, tcp_metrics_hash_log);
850 ret = -ESRCH;
851 rcu_read_lock();
852 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
853 tm = rcu_dereference(tm->tcpm_next)) {
854 if (addr_same(&tm->tcpm_daddr, &daddr) &&
855 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
856 net_eq(tm_net(tm), net)) {
857 ret = tcp_metrics_fill_info(msg, tm);
858 break;
859 }
860 }
861 rcu_read_unlock();
862 if (ret < 0)
863 goto out_free;
864
865 genlmsg_end(msg, reply);
866 return genlmsg_reply(msg, info);
867
868 nla_put_failure:
869 ret = -EMSGSIZE;
870
871 out_free:
872 nlmsg_free(msg);
873 return ret;
874 }
875
tcp_metrics_flush_all(struct net * net)876 static void tcp_metrics_flush_all(struct net *net)
877 {
878 unsigned int max_rows = 1U << tcp_metrics_hash_log;
879 struct tcpm_hash_bucket *hb = tcp_metrics_hash;
880 struct tcp_metrics_block *tm;
881 unsigned int row;
882
883 for (row = 0; row < max_rows; row++, hb++) {
884 struct tcp_metrics_block __rcu **pp;
885 bool match;
886
887 spin_lock_bh(&tcp_metrics_lock);
888 pp = &hb->chain;
889 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
890 match = net ? net_eq(tm_net(tm), net) :
891 !refcount_read(&tm_net(tm)->ns.count);
892 if (match) {
893 *pp = tm->tcpm_next;
894 kfree_rcu(tm, rcu_head);
895 } else {
896 pp = &tm->tcpm_next;
897 }
898 }
899 spin_unlock_bh(&tcp_metrics_lock);
900 }
901 }
902
tcp_metrics_nl_cmd_del(struct sk_buff * skb,struct genl_info * info)903 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
904 {
905 struct tcpm_hash_bucket *hb;
906 struct tcp_metrics_block *tm;
907 struct tcp_metrics_block __rcu **pp;
908 struct inetpeer_addr saddr, daddr;
909 unsigned int hash;
910 struct net *net = genl_info_net(info);
911 int ret;
912 bool src = true, found = false;
913
914 ret = parse_nl_addr(info, &daddr, &hash, 1);
915 if (ret < 0)
916 return ret;
917 if (ret > 0) {
918 tcp_metrics_flush_all(net);
919 return 0;
920 }
921 ret = parse_nl_saddr(info, &saddr);
922 if (ret < 0)
923 src = false;
924
925 hash ^= net_hash_mix(net);
926 hash = hash_32(hash, tcp_metrics_hash_log);
927 hb = tcp_metrics_hash + hash;
928 pp = &hb->chain;
929 spin_lock_bh(&tcp_metrics_lock);
930 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
931 if (addr_same(&tm->tcpm_daddr, &daddr) &&
932 (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
933 net_eq(tm_net(tm), net)) {
934 *pp = tm->tcpm_next;
935 kfree_rcu(tm, rcu_head);
936 found = true;
937 } else {
938 pp = &tm->tcpm_next;
939 }
940 }
941 spin_unlock_bh(&tcp_metrics_lock);
942 if (!found)
943 return -ESRCH;
944 return 0;
945 }
946
947 static const struct genl_small_ops tcp_metrics_nl_ops[] = {
948 {
949 .cmd = TCP_METRICS_CMD_GET,
950 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
951 .doit = tcp_metrics_nl_cmd_get,
952 .dumpit = tcp_metrics_nl_dump,
953 },
954 {
955 .cmd = TCP_METRICS_CMD_DEL,
956 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
957 .doit = tcp_metrics_nl_cmd_del,
958 .flags = GENL_ADMIN_PERM,
959 },
960 };
961
962 static struct genl_family tcp_metrics_nl_family __ro_after_init = {
963 .hdrsize = 0,
964 .name = TCP_METRICS_GENL_NAME,
965 .version = TCP_METRICS_GENL_VERSION,
966 .maxattr = TCP_METRICS_ATTR_MAX,
967 .policy = tcp_metrics_nl_policy,
968 .netnsok = true,
969 .module = THIS_MODULE,
970 .small_ops = tcp_metrics_nl_ops,
971 .n_small_ops = ARRAY_SIZE(tcp_metrics_nl_ops),
972 .resv_start_op = TCP_METRICS_CMD_DEL + 1,
973 };
974
975 static unsigned int tcpmhash_entries;
set_tcpmhash_entries(char * str)976 static int __init set_tcpmhash_entries(char *str)
977 {
978 ssize_t ret;
979
980 if (!str)
981 return 0;
982
983 ret = kstrtouint(str, 0, &tcpmhash_entries);
984 if (ret)
985 return 0;
986
987 return 1;
988 }
989 __setup("tcpmhash_entries=", set_tcpmhash_entries);
990
tcp_net_metrics_init(struct net * net)991 static int __net_init tcp_net_metrics_init(struct net *net)
992 {
993 size_t size;
994 unsigned int slots;
995
996 if (!net_eq(net, &init_net))
997 return 0;
998
999 slots = tcpmhash_entries;
1000 if (!slots) {
1001 if (totalram_pages() >= 128 * 1024)
1002 slots = 16 * 1024;
1003 else
1004 slots = 8 * 1024;
1005 }
1006
1007 tcp_metrics_hash_log = order_base_2(slots);
1008 size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1009
1010 tcp_metrics_hash = kvzalloc(size, GFP_KERNEL);
1011 if (!tcp_metrics_hash)
1012 return -ENOMEM;
1013
1014 return 0;
1015 }
1016
tcp_net_metrics_exit_batch(struct list_head * net_exit_list)1017 static void __net_exit tcp_net_metrics_exit_batch(struct list_head *net_exit_list)
1018 {
1019 tcp_metrics_flush_all(NULL);
1020 }
1021
1022 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1023 .init = tcp_net_metrics_init,
1024 .exit_batch = tcp_net_metrics_exit_batch,
1025 };
1026
tcp_metrics_init(void)1027 void __init tcp_metrics_init(void)
1028 {
1029 int ret;
1030
1031 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1032 if (ret < 0)
1033 panic("Could not allocate the tcp_metrics hash table\n");
1034
1035 ret = genl_register_family(&tcp_metrics_nl_family);
1036 if (ret < 0)
1037 panic("Could not register tcp_metrics generic netlink\n");
1038 }
1039