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