Commit | Line | Data |
---|---|---|
51c5d0c4 DM |
1 | #include <linux/rcupdate.h> |
2 | #include <linux/spinlock.h> | |
3 | #include <linux/jiffies.h> | |
4 | #include <linux/bootmem.h> | |
ab92bb2f | 5 | #include <linux/module.h> |
4aabd8ef | 6 | #include <linux/cache.h> |
51c5d0c4 DM |
7 | #include <linux/slab.h> |
8 | #include <linux/init.h> | |
4aabd8ef DM |
9 | #include <linux/tcp.h> |
10 | ||
11 | #include <net/inet_connection_sock.h> | |
51c5d0c4 | 12 | #include <net/net_namespace.h> |
ab92bb2f | 13 | #include <net/request_sock.h> |
51c5d0c4 | 14 | #include <net/inetpeer.h> |
4aabd8ef | 15 | #include <net/sock.h> |
51c5d0c4 | 16 | #include <net/ipv6.h> |
4aabd8ef DM |
17 | #include <net/dst.h> |
18 | #include <net/tcp.h> | |
19 | ||
20 | int sysctl_tcp_nometrics_save __read_mostly; | |
21 | ||
51c5d0c4 DM |
22 | enum tcp_metric_index { |
23 | TCP_METRIC_RTT, | |
24 | TCP_METRIC_RTTVAR, | |
25 | TCP_METRIC_SSTHRESH, | |
26 | TCP_METRIC_CWND, | |
27 | TCP_METRIC_REORDERING, | |
28 | ||
29 | /* Always last. */ | |
30 | TCP_METRIC_MAX, | |
31 | }; | |
32 | ||
1fe4c481 YC |
33 | struct tcp_fastopen_metrics { |
34 | u16 mss; | |
35 | struct tcp_fastopen_cookie cookie; | |
36 | }; | |
37 | ||
51c5d0c4 DM |
38 | struct tcp_metrics_block { |
39 | struct tcp_metrics_block __rcu *tcpm_next; | |
40 | struct inetpeer_addr tcpm_addr; | |
41 | unsigned long tcpm_stamp; | |
81166dd6 DM |
42 | u32 tcpm_ts; |
43 | u32 tcpm_ts_stamp; | |
51c5d0c4 DM |
44 | u32 tcpm_lock; |
45 | u32 tcpm_vals[TCP_METRIC_MAX]; | |
1fe4c481 | 46 | struct tcp_fastopen_metrics tcpm_fastopen; |
51c5d0c4 DM |
47 | }; |
48 | ||
49 | static bool tcp_metric_locked(struct tcp_metrics_block *tm, | |
50 | enum tcp_metric_index idx) | |
51 | { | |
52 | return tm->tcpm_lock & (1 << idx); | |
53 | } | |
54 | ||
55 | static u32 tcp_metric_get(struct tcp_metrics_block *tm, | |
56 | enum tcp_metric_index idx) | |
57 | { | |
58 | return tm->tcpm_vals[idx]; | |
59 | } | |
60 | ||
61 | static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm, | |
62 | enum tcp_metric_index idx) | |
63 | { | |
64 | return msecs_to_jiffies(tm->tcpm_vals[idx]); | |
65 | } | |
66 | ||
67 | static void tcp_metric_set(struct tcp_metrics_block *tm, | |
68 | enum tcp_metric_index idx, | |
69 | u32 val) | |
70 | { | |
71 | tm->tcpm_vals[idx] = val; | |
72 | } | |
73 | ||
74 | static void tcp_metric_set_msecs(struct tcp_metrics_block *tm, | |
75 | enum tcp_metric_index idx, | |
76 | u32 val) | |
77 | { | |
78 | tm->tcpm_vals[idx] = jiffies_to_msecs(val); | |
79 | } | |
80 | ||
81 | static bool addr_same(const struct inetpeer_addr *a, | |
82 | const struct inetpeer_addr *b) | |
83 | { | |
84 | const struct in6_addr *a6, *b6; | |
85 | ||
86 | if (a->family != b->family) | |
87 | return false; | |
88 | if (a->family == AF_INET) | |
89 | return a->addr.a4 == b->addr.a4; | |
90 | ||
91 | a6 = (const struct in6_addr *) &a->addr.a6[0]; | |
92 | b6 = (const struct in6_addr *) &b->addr.a6[0]; | |
93 | ||
94 | return ipv6_addr_equal(a6, b6); | |
95 | } | |
96 | ||
97 | struct tcpm_hash_bucket { | |
98 | struct tcp_metrics_block __rcu *chain; | |
99 | }; | |
100 | ||
101 | static DEFINE_SPINLOCK(tcp_metrics_lock); | |
102 | ||
103 | static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst) | |
104 | { | |
105 | u32 val; | |
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 | tm->tcpm_lock = val; | |
119 | ||
120 | tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT); | |
121 | tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR); | |
122 | tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH); | |
123 | tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND); | |
124 | tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING); | |
81166dd6 DM |
125 | tm->tcpm_ts = 0; |
126 | tm->tcpm_ts_stamp = 0; | |
1fe4c481 YC |
127 | tm->tcpm_fastopen.mss = 0; |
128 | tm->tcpm_fastopen.cookie.len = 0; | |
51c5d0c4 DM |
129 | } |
130 | ||
131 | static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst, | |
132 | struct inetpeer_addr *addr, | |
133 | unsigned int hash, | |
134 | bool reclaim) | |
135 | { | |
136 | struct tcp_metrics_block *tm; | |
137 | struct net *net; | |
138 | ||
139 | spin_lock_bh(&tcp_metrics_lock); | |
140 | net = dev_net(dst->dev); | |
141 | if (unlikely(reclaim)) { | |
142 | struct tcp_metrics_block *oldest; | |
143 | ||
144 | oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); | |
145 | for (tm = rcu_dereference(oldest->tcpm_next); tm; | |
146 | tm = rcu_dereference(tm->tcpm_next)) { | |
147 | if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp)) | |
148 | oldest = tm; | |
149 | } | |
150 | tm = oldest; | |
151 | } else { | |
152 | tm = kmalloc(sizeof(*tm), GFP_ATOMIC); | |
153 | if (!tm) | |
154 | goto out_unlock; | |
155 | } | |
156 | tm->tcpm_addr = *addr; | |
157 | tm->tcpm_stamp = jiffies; | |
158 | ||
159 | tcpm_suck_dst(tm, dst); | |
160 | ||
161 | if (likely(!reclaim)) { | |
162 | tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain; | |
163 | rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm); | |
164 | } | |
165 | ||
166 | out_unlock: | |
167 | spin_unlock_bh(&tcp_metrics_lock); | |
168 | return tm; | |
169 | } | |
170 | ||
171 | #define TCP_METRICS_TIMEOUT (60 * 60 * HZ) | |
172 | ||
173 | static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst) | |
174 | { | |
175 | if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT))) | |
176 | tcpm_suck_dst(tm, dst); | |
177 | } | |
178 | ||
179 | #define TCP_METRICS_RECLAIM_DEPTH 5 | |
180 | #define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL | |
181 | ||
182 | static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth) | |
183 | { | |
184 | if (tm) | |
185 | return tm; | |
186 | if (depth > TCP_METRICS_RECLAIM_DEPTH) | |
187 | return TCP_METRICS_RECLAIM_PTR; | |
188 | return NULL; | |
189 | } | |
190 | ||
191 | static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *addr, | |
192 | struct net *net, unsigned int hash) | |
193 | { | |
194 | struct tcp_metrics_block *tm; | |
195 | int depth = 0; | |
196 | ||
197 | for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; | |
198 | tm = rcu_dereference(tm->tcpm_next)) { | |
199 | if (addr_same(&tm->tcpm_addr, addr)) | |
200 | break; | |
201 | depth++; | |
202 | } | |
203 | return tcp_get_encode(tm, depth); | |
204 | } | |
205 | ||
206 | static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req, | |
207 | struct dst_entry *dst) | |
208 | { | |
209 | struct tcp_metrics_block *tm; | |
210 | struct inetpeer_addr addr; | |
211 | unsigned int hash; | |
212 | struct net *net; | |
213 | ||
214 | addr.family = req->rsk_ops->family; | |
215 | switch (addr.family) { | |
216 | case AF_INET: | |
217 | addr.addr.a4 = inet_rsk(req)->rmt_addr; | |
218 | hash = (__force unsigned int) addr.addr.a4; | |
219 | break; | |
220 | case AF_INET6: | |
221 | *(struct in6_addr *)addr.addr.a6 = inet6_rsk(req)->rmt_addr; | |
ddbe5032 | 222 | hash = ipv6_addr_hash(&inet6_rsk(req)->rmt_addr); |
51c5d0c4 DM |
223 | break; |
224 | default: | |
225 | return NULL; | |
226 | } | |
227 | ||
228 | hash ^= (hash >> 24) ^ (hash >> 16) ^ (hash >> 8); | |
229 | ||
230 | net = dev_net(dst->dev); | |
231 | hash &= net->ipv4.tcp_metrics_hash_mask; | |
232 | ||
233 | for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; | |
234 | tm = rcu_dereference(tm->tcpm_next)) { | |
235 | if (addr_same(&tm->tcpm_addr, &addr)) | |
236 | break; | |
237 | } | |
238 | tcpm_check_stamp(tm, dst); | |
239 | return tm; | |
240 | } | |
241 | ||
81166dd6 DM |
242 | static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw) |
243 | { | |
244 | struct inet6_timewait_sock *tw6; | |
245 | struct tcp_metrics_block *tm; | |
246 | struct inetpeer_addr addr; | |
247 | unsigned int hash; | |
248 | struct net *net; | |
249 | ||
250 | addr.family = tw->tw_family; | |
251 | switch (addr.family) { | |
252 | case AF_INET: | |
253 | addr.addr.a4 = tw->tw_daddr; | |
254 | hash = (__force unsigned int) addr.addr.a4; | |
255 | break; | |
256 | case AF_INET6: | |
257 | tw6 = inet6_twsk((struct sock *)tw); | |
258 | *(struct in6_addr *)addr.addr.a6 = tw6->tw_v6_daddr; | |
ddbe5032 | 259 | hash = ipv6_addr_hash(&tw6->tw_v6_daddr); |
81166dd6 DM |
260 | break; |
261 | default: | |
262 | return NULL; | |
263 | } | |
264 | ||
265 | hash ^= (hash >> 24) ^ (hash >> 16) ^ (hash >> 8); | |
266 | ||
267 | net = twsk_net(tw); | |
268 | hash &= net->ipv4.tcp_metrics_hash_mask; | |
269 | ||
270 | for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; | |
271 | tm = rcu_dereference(tm->tcpm_next)) { | |
272 | if (addr_same(&tm->tcpm_addr, &addr)) | |
273 | break; | |
274 | } | |
275 | return tm; | |
276 | } | |
277 | ||
51c5d0c4 DM |
278 | static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk, |
279 | struct dst_entry *dst, | |
280 | bool create) | |
281 | { | |
282 | struct tcp_metrics_block *tm; | |
283 | struct inetpeer_addr addr; | |
284 | unsigned int hash; | |
285 | struct net *net; | |
286 | bool reclaim; | |
287 | ||
288 | addr.family = sk->sk_family; | |
289 | switch (addr.family) { | |
290 | case AF_INET: | |
291 | addr.addr.a4 = inet_sk(sk)->inet_daddr; | |
292 | hash = (__force unsigned int) addr.addr.a4; | |
293 | break; | |
294 | case AF_INET6: | |
295 | *(struct in6_addr *)addr.addr.a6 = inet6_sk(sk)->daddr; | |
ddbe5032 | 296 | hash = ipv6_addr_hash(&inet6_sk(sk)->daddr); |
51c5d0c4 DM |
297 | break; |
298 | default: | |
299 | return NULL; | |
300 | } | |
301 | ||
302 | hash ^= (hash >> 24) ^ (hash >> 16) ^ (hash >> 8); | |
303 | ||
304 | net = dev_net(dst->dev); | |
305 | hash &= net->ipv4.tcp_metrics_hash_mask; | |
306 | ||
307 | tm = __tcp_get_metrics(&addr, net, hash); | |
308 | reclaim = false; | |
309 | if (tm == TCP_METRICS_RECLAIM_PTR) { | |
310 | reclaim = true; | |
311 | tm = NULL; | |
312 | } | |
313 | if (!tm && create) | |
314 | tm = tcpm_new(dst, &addr, hash, reclaim); | |
315 | else | |
316 | tcpm_check_stamp(tm, dst); | |
317 | ||
318 | return tm; | |
319 | } | |
320 | ||
4aabd8ef DM |
321 | /* Save metrics learned by this TCP session. This function is called |
322 | * only, when TCP finishes successfully i.e. when it enters TIME-WAIT | |
323 | * or goes from LAST-ACK to CLOSE. | |
324 | */ | |
325 | void tcp_update_metrics(struct sock *sk) | |
326 | { | |
51c5d0c4 | 327 | const struct inet_connection_sock *icsk = inet_csk(sk); |
4aabd8ef | 328 | struct dst_entry *dst = __sk_dst_get(sk); |
51c5d0c4 DM |
329 | struct tcp_sock *tp = tcp_sk(sk); |
330 | struct tcp_metrics_block *tm; | |
331 | unsigned long rtt; | |
332 | u32 val; | |
333 | int m; | |
4aabd8ef | 334 | |
51c5d0c4 | 335 | if (sysctl_tcp_nometrics_save || !dst) |
4aabd8ef DM |
336 | return; |
337 | ||
51c5d0c4 | 338 | if (dst->flags & DST_HOST) |
4aabd8ef DM |
339 | dst_confirm(dst); |
340 | ||
51c5d0c4 DM |
341 | rcu_read_lock(); |
342 | if (icsk->icsk_backoff || !tp->srtt) { | |
343 | /* This session failed to estimate rtt. Why? | |
344 | * Probably, no packets returned in time. Reset our | |
345 | * results. | |
346 | */ | |
347 | tm = tcp_get_metrics(sk, dst, false); | |
348 | if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT)) | |
349 | tcp_metric_set(tm, TCP_METRIC_RTT, 0); | |
350 | goto out_unlock; | |
351 | } else | |
352 | tm = tcp_get_metrics(sk, dst, true); | |
4aabd8ef | 353 | |
51c5d0c4 DM |
354 | if (!tm) |
355 | goto out_unlock; | |
4aabd8ef | 356 | |
51c5d0c4 DM |
357 | rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT); |
358 | m = rtt - tp->srtt; | |
4aabd8ef | 359 | |
51c5d0c4 DM |
360 | /* If newly calculated rtt larger than stored one, store new |
361 | * one. Otherwise, use EWMA. Remember, rtt overestimation is | |
362 | * always better than underestimation. | |
363 | */ | |
364 | if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) { | |
365 | if (m <= 0) | |
366 | rtt = tp->srtt; | |
367 | else | |
368 | rtt -= (m >> 3); | |
369 | tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt); | |
370 | } | |
4aabd8ef | 371 | |
51c5d0c4 DM |
372 | if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) { |
373 | unsigned long var; | |
4aabd8ef | 374 | |
51c5d0c4 DM |
375 | if (m < 0) |
376 | m = -m; | |
4aabd8ef | 377 | |
51c5d0c4 DM |
378 | /* Scale deviation to rttvar fixed point */ |
379 | m >>= 1; | |
380 | if (m < tp->mdev) | |
381 | m = tp->mdev; | |
4aabd8ef | 382 | |
51c5d0c4 DM |
383 | var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR); |
384 | if (m >= var) | |
385 | var = m; | |
386 | else | |
387 | var -= (var - m) >> 2; | |
4aabd8ef | 388 | |
51c5d0c4 DM |
389 | tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var); |
390 | } | |
391 | ||
392 | if (tcp_in_initial_slowstart(tp)) { | |
393 | /* Slow start still did not finish. */ | |
394 | if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) { | |
395 | val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); | |
396 | if (val && (tp->snd_cwnd >> 1) > val) | |
397 | tcp_metric_set(tm, TCP_METRIC_SSTHRESH, | |
398 | tp->snd_cwnd >> 1); | |
399 | } | |
400 | if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { | |
401 | val = tcp_metric_get(tm, TCP_METRIC_CWND); | |
402 | if (tp->snd_cwnd > val) | |
403 | tcp_metric_set(tm, TCP_METRIC_CWND, | |
404 | tp->snd_cwnd); | |
405 | } | |
406 | } else if (tp->snd_cwnd > tp->snd_ssthresh && | |
407 | icsk->icsk_ca_state == TCP_CA_Open) { | |
408 | /* Cong. avoidance phase, cwnd is reliable. */ | |
409 | if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) | |
410 | tcp_metric_set(tm, TCP_METRIC_SSTHRESH, | |
411 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh)); | |
412 | if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { | |
413 | val = tcp_metric_get(tm, TCP_METRIC_CWND); | |
2100844c | 414 | tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1); |
51c5d0c4 DM |
415 | } |
416 | } else { | |
417 | /* Else slow start did not finish, cwnd is non-sense, | |
418 | * ssthresh may be also invalid. | |
419 | */ | |
420 | if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { | |
421 | val = tcp_metric_get(tm, TCP_METRIC_CWND); | |
422 | tcp_metric_set(tm, TCP_METRIC_CWND, | |
423 | (val + tp->snd_ssthresh) >> 1); | |
424 | } | |
425 | if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) { | |
426 | val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); | |
427 | if (val && tp->snd_ssthresh > val) | |
428 | tcp_metric_set(tm, TCP_METRIC_SSTHRESH, | |
429 | tp->snd_ssthresh); | |
430 | } | |
431 | if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) { | |
432 | val = tcp_metric_get(tm, TCP_METRIC_REORDERING); | |
433 | if (val < tp->reordering && | |
4aabd8ef | 434 | tp->reordering != sysctl_tcp_reordering) |
51c5d0c4 DM |
435 | tcp_metric_set(tm, TCP_METRIC_REORDERING, |
436 | tp->reordering); | |
4aabd8ef DM |
437 | } |
438 | } | |
51c5d0c4 DM |
439 | tm->tcpm_stamp = jiffies; |
440 | out_unlock: | |
441 | rcu_read_unlock(); | |
4aabd8ef DM |
442 | } |
443 | ||
444 | /* Initialize metrics on socket. */ | |
445 | ||
446 | void tcp_init_metrics(struct sock *sk) | |
447 | { | |
4aabd8ef | 448 | struct dst_entry *dst = __sk_dst_get(sk); |
51c5d0c4 DM |
449 | struct tcp_sock *tp = tcp_sk(sk); |
450 | struct tcp_metrics_block *tm; | |
451 | u32 val; | |
4aabd8ef DM |
452 | |
453 | if (dst == NULL) | |
454 | goto reset; | |
455 | ||
456 | dst_confirm(dst); | |
457 | ||
51c5d0c4 DM |
458 | rcu_read_lock(); |
459 | tm = tcp_get_metrics(sk, dst, true); | |
460 | if (!tm) { | |
461 | rcu_read_unlock(); | |
462 | goto reset; | |
463 | } | |
464 | ||
465 | if (tcp_metric_locked(tm, TCP_METRIC_CWND)) | |
466 | tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND); | |
467 | ||
468 | val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); | |
469 | if (val) { | |
470 | tp->snd_ssthresh = val; | |
4aabd8ef DM |
471 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) |
472 | tp->snd_ssthresh = tp->snd_cwnd_clamp; | |
473 | } else { | |
474 | /* ssthresh may have been reduced unnecessarily during. | |
475 | * 3WHS. Restore it back to its initial default. | |
476 | */ | |
477 | tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; | |
478 | } | |
51c5d0c4 DM |
479 | val = tcp_metric_get(tm, TCP_METRIC_REORDERING); |
480 | if (val && tp->reordering != val) { | |
4aabd8ef DM |
481 | tcp_disable_fack(tp); |
482 | tcp_disable_early_retrans(tp); | |
51c5d0c4 | 483 | tp->reordering = val; |
4aabd8ef DM |
484 | } |
485 | ||
51c5d0c4 DM |
486 | val = tcp_metric_get(tm, TCP_METRIC_RTT); |
487 | if (val == 0 || tp->srtt == 0) { | |
488 | rcu_read_unlock(); | |
4aabd8ef | 489 | goto reset; |
51c5d0c4 | 490 | } |
4aabd8ef DM |
491 | /* Initial rtt is determined from SYN,SYN-ACK. |
492 | * The segment is small and rtt may appear much | |
493 | * less than real one. Use per-dst memory | |
494 | * to make it more realistic. | |
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 | */ | |
51c5d0c4 DM |
505 | val = msecs_to_jiffies(val); |
506 | if (val > tp->srtt) { | |
507 | tp->srtt = val; | |
4aabd8ef DM |
508 | tp->rtt_seq = tp->snd_nxt; |
509 | } | |
51c5d0c4 DM |
510 | val = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR); |
511 | if (val > tp->mdev) { | |
512 | tp->mdev = val; | |
4aabd8ef DM |
513 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
514 | } | |
51c5d0c4 DM |
515 | rcu_read_unlock(); |
516 | ||
4aabd8ef DM |
517 | tcp_set_rto(sk); |
518 | reset: | |
519 | if (tp->srtt == 0) { | |
520 | /* RFC6298: 5.7 We've failed to get a valid RTT sample from | |
521 | * 3WHS. This is most likely due to retransmission, | |
522 | * including spurious one. Reset the RTO back to 3secs | |
523 | * from the more aggressive 1sec to avoid more spurious | |
524 | * retransmission. | |
525 | */ | |
526 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK; | |
527 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK; | |
528 | } | |
529 | /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been | |
530 | * retransmitted. In light of RFC6298 more aggressive 1sec | |
531 | * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK | |
532 | * retransmission has occurred. | |
533 | */ | |
534 | if (tp->total_retrans > 1) | |
535 | tp->snd_cwnd = 1; | |
536 | else | |
537 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); | |
538 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
539 | } | |
ab92bb2f | 540 | |
81166dd6 | 541 | bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check) |
ab92bb2f | 542 | { |
51c5d0c4 DM |
543 | struct tcp_metrics_block *tm; |
544 | bool ret; | |
545 | ||
ab92bb2f DM |
546 | if (!dst) |
547 | return false; | |
51c5d0c4 DM |
548 | |
549 | rcu_read_lock(); | |
550 | tm = __tcp_get_metrics_req(req, dst); | |
81166dd6 DM |
551 | if (paws_check) { |
552 | if (tm && | |
553 | (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL && | |
554 | (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW) | |
555 | ret = false; | |
556 | else | |
557 | ret = true; | |
558 | } else { | |
559 | if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp) | |
560 | ret = true; | |
561 | else | |
562 | ret = false; | |
563 | } | |
51c5d0c4 DM |
564 | rcu_read_unlock(); |
565 | ||
566 | return ret; | |
ab92bb2f DM |
567 | } |
568 | EXPORT_SYMBOL_GPL(tcp_peer_is_proven); | |
51c5d0c4 | 569 | |
81166dd6 DM |
570 | void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst) |
571 | { | |
572 | struct tcp_metrics_block *tm; | |
573 | ||
574 | rcu_read_lock(); | |
575 | tm = tcp_get_metrics(sk, dst, true); | |
576 | if (tm) { | |
577 | struct tcp_sock *tp = tcp_sk(sk); | |
578 | ||
579 | if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) { | |
580 | tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp; | |
581 | tp->rx_opt.ts_recent = tm->tcpm_ts; | |
582 | } | |
583 | } | |
584 | rcu_read_unlock(); | |
585 | } | |
586 | EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp); | |
587 | ||
588 | /* VJ's idea. Save last timestamp seen from this destination and hold | |
589 | * it at least for normal timewait interval to use for duplicate | |
590 | * segment detection in subsequent connections, before they enter | |
591 | * synchronized state. | |
592 | */ | |
593 | bool tcp_remember_stamp(struct sock *sk) | |
594 | { | |
595 | struct dst_entry *dst = __sk_dst_get(sk); | |
596 | bool ret = false; | |
597 | ||
598 | if (dst) { | |
599 | struct tcp_metrics_block *tm; | |
600 | ||
601 | rcu_read_lock(); | |
602 | tm = tcp_get_metrics(sk, dst, true); | |
603 | if (tm) { | |
604 | struct tcp_sock *tp = tcp_sk(sk); | |
605 | ||
606 | if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 || | |
607 | ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL && | |
608 | tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) { | |
609 | tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp; | |
610 | tm->tcpm_ts = tp->rx_opt.ts_recent; | |
611 | } | |
612 | ret = true; | |
613 | } | |
614 | rcu_read_unlock(); | |
615 | } | |
616 | return ret; | |
617 | } | |
618 | ||
619 | bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw) | |
620 | { | |
621 | struct tcp_metrics_block *tm; | |
622 | bool ret = false; | |
623 | ||
624 | rcu_read_lock(); | |
625 | tm = __tcp_get_metrics_tw(tw); | |
626 | if (tw) { | |
627 | const struct tcp_timewait_sock *tcptw; | |
628 | struct sock *sk = (struct sock *) tw; | |
629 | ||
630 | tcptw = tcp_twsk(sk); | |
631 | if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 || | |
632 | ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL && | |
633 | tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) { | |
634 | tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp; | |
635 | tm->tcpm_ts = tcptw->tw_ts_recent; | |
636 | } | |
637 | ret = true; | |
638 | } | |
639 | rcu_read_unlock(); | |
640 | ||
641 | return ret; | |
642 | } | |
643 | ||
1fe4c481 YC |
644 | static DEFINE_SEQLOCK(fastopen_seqlock); |
645 | ||
646 | void tcp_fastopen_cache_get(struct sock *sk, u16 *mss, | |
647 | struct tcp_fastopen_cookie *cookie) | |
648 | { | |
649 | struct tcp_metrics_block *tm; | |
650 | ||
651 | rcu_read_lock(); | |
652 | tm = tcp_get_metrics(sk, __sk_dst_get(sk), false); | |
653 | if (tm) { | |
654 | struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen; | |
655 | unsigned int seq; | |
656 | ||
657 | do { | |
658 | seq = read_seqbegin(&fastopen_seqlock); | |
659 | if (tfom->mss) | |
660 | *mss = tfom->mss; | |
661 | *cookie = tfom->cookie; | |
662 | } while (read_seqretry(&fastopen_seqlock, seq)); | |
663 | } | |
664 | rcu_read_unlock(); | |
665 | } | |
666 | ||
667 | ||
668 | void tcp_fastopen_cache_set(struct sock *sk, u16 mss, | |
669 | struct tcp_fastopen_cookie *cookie) | |
670 | { | |
671 | struct tcp_metrics_block *tm; | |
672 | ||
673 | rcu_read_lock(); | |
674 | tm = tcp_get_metrics(sk, __sk_dst_get(sk), true); | |
675 | if (tm) { | |
676 | struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen; | |
677 | ||
678 | write_seqlock_bh(&fastopen_seqlock); | |
679 | tfom->mss = mss; | |
680 | if (cookie->len > 0) | |
681 | tfom->cookie = *cookie; | |
682 | write_sequnlock_bh(&fastopen_seqlock); | |
683 | } | |
684 | rcu_read_unlock(); | |
685 | } | |
686 | ||
51c5d0c4 DM |
687 | static unsigned long tcpmhash_entries; |
688 | static int __init set_tcpmhash_entries(char *str) | |
689 | { | |
690 | ssize_t ret; | |
691 | ||
692 | if (!str) | |
693 | return 0; | |
694 | ||
695 | ret = kstrtoul(str, 0, &tcpmhash_entries); | |
696 | if (ret) | |
697 | return 0; | |
698 | ||
699 | return 1; | |
700 | } | |
701 | __setup("tcpmhash_entries=", set_tcpmhash_entries); | |
702 | ||
703 | static int __net_init tcp_net_metrics_init(struct net *net) | |
704 | { | |
705 | int slots, size; | |
706 | ||
707 | slots = tcpmhash_entries; | |
708 | if (!slots) { | |
709 | if (totalram_pages >= 128 * 1024) | |
710 | slots = 16 * 1024; | |
711 | else | |
712 | slots = 8 * 1024; | |
713 | } | |
714 | ||
715 | size = slots * sizeof(struct tcpm_hash_bucket); | |
716 | ||
717 | net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL); | |
718 | if (!net->ipv4.tcp_metrics_hash) | |
719 | return -ENOMEM; | |
720 | ||
721 | net->ipv4.tcp_metrics_hash_mask = (slots - 1); | |
722 | ||
723 | return 0; | |
724 | } | |
725 | ||
726 | static void __net_exit tcp_net_metrics_exit(struct net *net) | |
727 | { | |
728 | kfree(net->ipv4.tcp_metrics_hash); | |
729 | } | |
730 | ||
731 | static __net_initdata struct pernet_operations tcp_net_metrics_ops = { | |
732 | .init = tcp_net_metrics_init, | |
733 | .exit = tcp_net_metrics_exit, | |
734 | }; | |
735 | ||
736 | void __init tcp_metrics_init(void) | |
737 | { | |
738 | register_pernet_subsys(&tcp_net_metrics_ops); | |
739 | } |