projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
tcp: metrics: Add source-address to tcp-metrics
[linux-2.6-block.git] / net / ipv4 / tcp_metrics.c
CommitLineData
51c5d0c4
DM		 1#include <linux/rcupdate.h>
2#include <linux/spinlock.h>
3#include <linux/jiffies.h>
ab92bb2f 4#include <linux/module.h>
4aabd8ef 5#include <linux/cache.h>
51c5d0c4
DM		 6#include <linux/slab.h>
7#include <linux/init.h>
4aabd8ef 8#include <linux/tcp.h>
5815d5e7 9#include <linux/hash.h>
d23ff701 10#include <linux/tcp_metrics.h>
976a702a 11#include <linux/vmalloc.h>
4aabd8ef
DM		 12
13#include <net/inet_connection_sock.h>
51c5d0c4 14#include <net/net_namespace.h>
ab92bb2f 15#include <net/request_sock.h>
51c5d0c4 16#include <net/inetpeer.h>
4aabd8ef 17#include <net/sock.h>
51c5d0c4 18#include <net/ipv6.h>
4aabd8ef
DM		 19#include <net/dst.h>
20#include <net/tcp.h>
d23ff701 21#include <net/genetlink.h>
4aabd8ef
DM		 22
23int sysctl_tcp_nometrics_save __read_mostly;
24
1fe4c481
YC		 25struct tcp_fastopen_metrics {
26 u16 mss;
aab48743
YC		 27 u16 syn_loss:10; /* Recurring Fast Open SYN losses */
28 unsigned long last_syn_loss; /* Last Fast Open SYN loss */
1fe4c481
YC		 29 struct tcp_fastopen_cookie cookie;
30};
31
51c5d0c4
DM		 32struct tcp_metrics_block {
33 struct tcp_metrics_block __rcu *tcpm_next;
a5443028 34 struct inetpeer_addr tcpm_saddr;
324fd55a 35 struct inetpeer_addr tcpm_daddr;
51c5d0c4 36 unsigned long tcpm_stamp;
81166dd6
DM		 37 u32 tcpm_ts;
38 u32 tcpm_ts_stamp;
51c5d0c4 39 u32 tcpm_lock;
d23ff701 40 u32 tcpm_vals[TCP_METRIC_MAX + 1];
1fe4c481 41 struct tcp_fastopen_metrics tcpm_fastopen;
d23ff701
JA		 42
43 struct rcu_head rcu_head;
51c5d0c4
DM		 44};
45
46static bool tcp_metric_locked(struct tcp_metrics_block *tm,
47 enum tcp_metric_index idx)
48{
49 return tm->tcpm_lock & (1 << idx);
50}
51
52static u32 tcp_metric_get(struct tcp_metrics_block *tm,
53 enum tcp_metric_index idx)
54{
55 return tm->tcpm_vals[idx];
56}
57
58static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm,
59 enum tcp_metric_index idx)
60{
61 return msecs_to_jiffies(tm->tcpm_vals[idx]);
62}
63
64static void tcp_metric_set(struct tcp_metrics_block *tm,
65 enum tcp_metric_index idx,
66 u32 val)
67{
68 tm->tcpm_vals[idx] = val;
69}
70
71static void tcp_metric_set_msecs(struct tcp_metrics_block *tm,
72 enum tcp_metric_index idx,
73 u32 val)
74{
75 tm->tcpm_vals[idx] = jiffies_to_msecs(val);
76}
77
78static bool addr_same(const struct inetpeer_addr *a,
79 const struct inetpeer_addr *b)
80{
81 const struct in6_addr *a6, *b6;
82
83 if (a->family != b->family)
84 return false;
85 if (a->family == AF_INET)
86 return a->addr.a4 == b->addr.a4;
87
88 a6 = (const struct in6_addr *) &a->addr.a6[0];
89 b6 = (const struct in6_addr *) &b->addr.a6[0];
90
91 return ipv6_addr_equal(a6, b6);
92}
93
94struct tcpm_hash_bucket {
95 struct tcp_metrics_block __rcu *chain;
96};
97
98static DEFINE_SPINLOCK(tcp_metrics_lock);
99
efeaa555
ED		 100static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst,
101 bool fastopen_clear)
51c5d0c4
DM		 102{
103 u32 val;
104
9a0a9502
JA		 105 tm->tcpm_stamp = jiffies;
106
51c5d0c4
DM		 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;
efeaa555
ED		 127 if (fastopen_clear) {
128 tm->tcpm_fastopen.mss = 0;
129 tm->tcpm_fastopen.syn_loss = 0;
130 tm->tcpm_fastopen.cookie.len = 0;
131 }
51c5d0c4
DM		 132}
133
134static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
a5443028 135 struct inetpeer_addr *saddr,
324fd55a 136 struct inetpeer_addr *daddr,
51c5d0c4
DM		 137 unsigned int hash,
138 bool reclaim)
139{
140 struct tcp_metrics_block *tm;
141 struct net *net;
142
143 spin_lock_bh(&tcp_metrics_lock);
144 net = dev_net(dst->dev);
145 if (unlikely(reclaim)) {
146 struct tcp_metrics_block *oldest;
147
148 oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
149 for (tm = rcu_dereference(oldest->tcpm_next); tm;
150 tm = rcu_dereference(tm->tcpm_next)) {
151 if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
152 oldest = tm;
153 }
154 tm = oldest;
155 } else {
156 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
157 if (!tm)
158 goto out_unlock;
159 }
a5443028 160 tm->tcpm_saddr = *saddr;
324fd55a 161 tm->tcpm_daddr = *daddr;
51c5d0c4 162
efeaa555 163 tcpm_suck_dst(tm, dst, true);
51c5d0c4
DM		 164
165 if (likely(!reclaim)) {
166 tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
167 rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
168 }
169
170out_unlock:
171 spin_unlock_bh(&tcp_metrics_lock);
172 return tm;
173}
174
175#define TCP_METRICS_TIMEOUT (60 * 60 * HZ)
176
177static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
178{
179 if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
efeaa555 180 tcpm_suck_dst(tm, dst, false);
51c5d0c4
DM		 181}
182
183#define TCP_METRICS_RECLAIM_DEPTH 5
184#define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL
185
186static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
187{
188 if (tm)
189 return tm;
190 if (depth > TCP_METRICS_RECLAIM_DEPTH)
191 return TCP_METRICS_RECLAIM_PTR;
192 return NULL;
193}
194
a5443028
CP		 195static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
196 const struct inetpeer_addr *daddr,
51c5d0c4
DM		 197 struct net *net, unsigned int hash)
198{
199 struct tcp_metrics_block *tm;
200 int depth = 0;
201
202 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
203 tm = rcu_dereference(tm->tcpm_next)) {
a5443028
CP		 204 if (addr_same(&tm->tcpm_saddr, saddr) &&
205 addr_same(&tm->tcpm_daddr, daddr))
51c5d0c4
DM		 206 break;
207 depth++;
208 }
209 return tcp_get_encode(tm, depth);
210}
211
212static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
213 struct dst_entry *dst)
214{
215 struct tcp_metrics_block *tm;
a5443028 216 struct inetpeer_addr saddr, daddr;
51c5d0c4
DM		 217 unsigned int hash;
218 struct net *net;
219
a5443028 220 saddr.family = req->rsk_ops->family;
324fd55a
CP		 221 daddr.family = req->rsk_ops->family;
222 switch (daddr.family) {
51c5d0c4 223 case AF_INET:
a5443028 224 saddr.addr.a4 = inet_rsk(req)->ir_loc_addr;
324fd55a
CP		 225 daddr.addr.a4 = inet_rsk(req)->ir_rmt_addr;
226 hash = (__force unsigned int) daddr.addr.a4;
51c5d0c4 227 break;
634fb979 228#if IS_ENABLED(CONFIG_IPV6)
51c5d0c4 229 case AF_INET6:
a5443028 230 *(struct in6_addr *)saddr.addr.a6 = inet_rsk(req)->ir_v6_loc_addr;
324fd55a 231 *(struct in6_addr *)daddr.addr.a6 = inet_rsk(req)->ir_v6_rmt_addr;
634fb979 232 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
51c5d0c4 233 break;
634fb979 234#endif
51c5d0c4
DM		 235 default:
236 return NULL;
237 }
238
51c5d0c4 239 net = dev_net(dst->dev);
5815d5e7 240 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
51c5d0c4
DM		 241
242 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
243 tm = rcu_dereference(tm->tcpm_next)) {
a5443028
CP		 244 if (addr_same(&tm->tcpm_saddr, &saddr) &&
245 addr_same(&tm->tcpm_daddr, &daddr))
51c5d0c4
DM		 246 break;
247 }
248 tcpm_check_stamp(tm, dst);
249 return tm;
250}
251
81166dd6
DM		 252static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
253{
81166dd6 254 struct tcp_metrics_block *tm;
a5443028 255 struct inetpeer_addr saddr, daddr;
81166dd6
DM		 256 unsigned int hash;
257 struct net *net;
258
a5443028 259 saddr.family = tw->tw_family;
324fd55a
CP		 260 daddr.family = tw->tw_family;
261 switch (daddr.family) {
81166dd6 262 case AF_INET:
a5443028 263 saddr.addr.a4 = tw->tw_rcv_saddr;
324fd55a
CP		 264 daddr.addr.a4 = tw->tw_daddr;
265 hash = (__force unsigned int) daddr.addr.a4;
81166dd6 266 break;
c2bb06db 267#if IS_ENABLED(CONFIG_IPV6)
81166dd6 268 case AF_INET6:
a5443028 269 *(struct in6_addr *)saddr.addr.a6 = tw->tw_v6_rcv_saddr;
324fd55a 270 *(struct in6_addr *)daddr.addr.a6 = tw->tw_v6_daddr;
efe4208f 271 hash = ipv6_addr_hash(&tw->tw_v6_daddr);
81166dd6 272 break;
c2bb06db 273#endif
81166dd6
DM		 274 default:
275 return NULL;
276 }
277
81166dd6 278 net = twsk_net(tw);
5815d5e7 279 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
81166dd6
DM		 280
281 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
282 tm = rcu_dereference(tm->tcpm_next)) {
a5443028
CP		 283 if (addr_same(&tm->tcpm_saddr, &saddr) &&
284 addr_same(&tm->tcpm_daddr, &daddr))
81166dd6
DM		 285 break;
286 }
287 return tm;
288}
289
51c5d0c4
DM		 290static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
291 struct dst_entry *dst,
292 bool create)
293{
294 struct tcp_metrics_block *tm;
a5443028 295 struct inetpeer_addr saddr, daddr;
51c5d0c4
DM		 296 unsigned int hash;
297 struct net *net;
298 bool reclaim;
299
a5443028 300 saddr.family = sk->sk_family;
324fd55a
CP		 301 daddr.family = sk->sk_family;
302 switch (daddr.family) {
51c5d0c4 303 case AF_INET:
a5443028 304 saddr.addr.a4 = inet_sk(sk)->inet_saddr;
324fd55a
CP		 305 daddr.addr.a4 = inet_sk(sk)->inet_daddr;
306 hash = (__force unsigned int) daddr.addr.a4;
51c5d0c4 307 break;
c2bb06db 308#if IS_ENABLED(CONFIG_IPV6)
51c5d0c4 309 case AF_INET6:
a5443028 310 *(struct in6_addr *)saddr.addr.a6 = sk->sk_v6_rcv_saddr;
324fd55a 311 *(struct in6_addr *)daddr.addr.a6 = sk->sk_v6_daddr;
efe4208f 312 hash = ipv6_addr_hash(&sk->sk_v6_daddr);
51c5d0c4 313 break;
c2bb06db 314#endif
51c5d0c4
DM		 315 default:
316 return NULL;
317 }
318
51c5d0c4 319 net = dev_net(dst->dev);
5815d5e7 320 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
51c5d0c4 321
a5443028 322 tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
51c5d0c4
DM		 323 reclaim = false;
324 if (tm == TCP_METRICS_RECLAIM_PTR) {
325 reclaim = true;
326 tm = NULL;
327 }
328 if (!tm && create)
a5443028 329 tm = tcpm_new(dst, &saddr, &daddr, hash, reclaim);
51c5d0c4
DM		 330 else
331 tcpm_check_stamp(tm, dst);
332
333 return tm;
334}
335
4aabd8ef
DM		 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{
51c5d0c4 342 const struct inet_connection_sock *icsk = inet_csk(sk);
4aabd8ef 343 struct dst_entry *dst = __sk_dst_get(sk);
51c5d0c4
DM		 344 struct tcp_sock *tp = tcp_sk(sk);
345 struct tcp_metrics_block *tm;
346 unsigned long rtt;
347 u32 val;
348 int m;
4aabd8ef 349
51c5d0c4 350 if (sysctl_tcp_nometrics_save || !dst)
4aabd8ef
DM		 351 return;
352
51c5d0c4 353 if (dst->flags & DST_HOST)
4aabd8ef
DM		 354 dst_confirm(dst);
355
51c5d0c4
DM		 356 rcu_read_lock();
357 if (icsk->icsk_backoff || !tp->srtt) {
358 /* This session failed to estimate rtt. Why?
359 * Probably, no packets returned in time. Reset our
360 * results.
361 */
362 tm = tcp_get_metrics(sk, dst, false);
363 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
364 tcp_metric_set(tm, TCP_METRIC_RTT, 0);
365 goto out_unlock;
366 } else
367 tm = tcp_get_metrics(sk, dst, true);
4aabd8ef 368
51c5d0c4
DM		 369 if (!tm)
370 goto out_unlock;
4aabd8ef 371
51c5d0c4
DM		 372 rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
373 m = rtt - tp->srtt;
4aabd8ef 374
51c5d0c4
DM		 375 /* If newly calculated rtt larger than stored one, store new
376 * one. Otherwise, use EWMA. Remember, rtt overestimation is
377 * always better than underestimation.
378 */
379 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
380 if (m <= 0)
381 rtt = tp->srtt;
382 else
383 rtt -= (m >> 3);
384 tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt);
385 }
4aabd8ef 386
51c5d0c4
DM		 387 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
388 unsigned long var;
4aabd8ef 389
51c5d0c4
DM		 390 if (m < 0)
391 m = -m;
4aabd8ef 392
51c5d0c4
DM		 393 /* Scale deviation to rttvar fixed point */
394 m >>= 1;
395 if (m < tp->mdev)
396 m = tp->mdev;
4aabd8ef 397
51c5d0c4
DM		 398 var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
399 if (m >= var)
400 var = m;
401 else
402 var -= (var - m) >> 2;
4aabd8ef 403
51c5d0c4
DM		 404 tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var);
405 }
406
407 if (tcp_in_initial_slowstart(tp)) {
408 /* Slow start still did not finish. */
409 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
410 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
411 if (val && (tp->snd_cwnd >> 1) > val)
412 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
413 tp->snd_cwnd >> 1);
414 }
415 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
416 val = tcp_metric_get(tm, TCP_METRIC_CWND);
417 if (tp->snd_cwnd > val)
418 tcp_metric_set(tm, TCP_METRIC_CWND,
419 tp->snd_cwnd);
420 }
421 } else if (tp->snd_cwnd > tp->snd_ssthresh &&
422 icsk->icsk_ca_state == TCP_CA_Open) {
423 /* Cong. avoidance phase, cwnd is reliable. */
424 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
425 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
426 max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
427 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
428 val = tcp_metric_get(tm, TCP_METRIC_CWND);
2100844c 429 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
51c5d0c4
DM		 430 }
431 } else {
432 /* Else slow start did not finish, cwnd is non-sense,
433 * ssthresh may be also invalid.
434 */
435 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
436 val = tcp_metric_get(tm, TCP_METRIC_CWND);
437 tcp_metric_set(tm, TCP_METRIC_CWND,
438 (val + tp->snd_ssthresh) >> 1);
439 }
440 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
441 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
442 if (val && tp->snd_ssthresh > val)
443 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
444 tp->snd_ssthresh);
445 }
446 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
447 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
448 if (val < tp->reordering &&
4aabd8ef 449 tp->reordering != sysctl_tcp_reordering)
51c5d0c4
DM		 450 tcp_metric_set(tm, TCP_METRIC_REORDERING,
451 tp->reordering);
4aabd8ef
DM		 452 }
453 }
51c5d0c4
DM		 454 tm->tcpm_stamp = jiffies;
455out_unlock:
456 rcu_read_unlock();
4aabd8ef
DM		 457}
458
459/* Initialize metrics on socket. */
460
461void tcp_init_metrics(struct sock *sk)
462{
4aabd8ef 463 struct dst_entry *dst = __sk_dst_get(sk);
51c5d0c4
DM		 464 struct tcp_sock *tp = tcp_sk(sk);
465 struct tcp_metrics_block *tm;
1b7fdd2a 466 u32 val, crtt = 0; /* cached RTT scaled by 8 */
4aabd8ef
DM		 467
468 if (dst == NULL)
469 goto reset;
470
471 dst_confirm(dst);
472
51c5d0c4
DM		 473 rcu_read_lock();
474 tm = tcp_get_metrics(sk, dst, true);
475 if (!tm) {
476 rcu_read_unlock();
477 goto reset;
478 }
479
480 if (tcp_metric_locked(tm, TCP_METRIC_CWND))
481 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
482
483 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
484 if (val) {
485 tp->snd_ssthresh = val;
4aabd8ef
DM		 486 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
487 tp->snd_ssthresh = tp->snd_cwnd_clamp;
488 } else {
489 /* ssthresh may have been reduced unnecessarily during.
490 * 3WHS. Restore it back to its initial default.
491 */
492 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
493 }
51c5d0c4
DM		 494 val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
495 if (val && tp->reordering != val) {
4aabd8ef
DM		 496 tcp_disable_fack(tp);
497 tcp_disable_early_retrans(tp);
51c5d0c4 498 tp->reordering = val;
4aabd8ef
DM		 499 }
500
1b7fdd2a 501 crtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
51c5d0c4 502 rcu_read_unlock();
4aabd8ef 503reset:
52f20e65
YC		 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 */
1b7fdd2a 524 if (crtt > tp->srtt) {
269aa759
NC		 525 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
526 crtt >>= 3;
527 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
1b7fdd2a 528 } else if (tp->srtt == 0) {
4aabd8ef
DM		 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->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
536 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
537 }
538 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
539 * retransmitted. In light of RFC6298 more aggressive 1sec
540 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
541 * retransmission has occurred.
542 */
543 if (tp->total_retrans > 1)
544 tp->snd_cwnd = 1;
545 else
546 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
547 tp->snd_cwnd_stamp = tcp_time_stamp;
548}
ab92bb2f 549
81166dd6 550bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
ab92bb2f 551{
51c5d0c4
DM		 552 struct tcp_metrics_block *tm;
553 bool ret;
554
ab92bb2f
DM		 555 if (!dst)
556 return false;
51c5d0c4
DM		 557
558 rcu_read_lock();
559 tm = __tcp_get_metrics_req(req, dst);
81166dd6
DM		 560 if (paws_check) {
561 if (tm &&
562 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
563 (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
564 ret = false;
565 else
566 ret = true;
567 } else {
568 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
569 ret = true;
570 else
571 ret = false;
572 }
51c5d0c4
DM		 573 rcu_read_unlock();
574
575 return ret;
ab92bb2f
DM		 576}
577EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
51c5d0c4 578
81166dd6
DM		 579void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
580{
581 struct tcp_metrics_block *tm;
582
583 rcu_read_lock();
584 tm = tcp_get_metrics(sk, dst, true);
585 if (tm) {
586 struct tcp_sock *tp = tcp_sk(sk);
587
588 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
589 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
590 tp->rx_opt.ts_recent = tm->tcpm_ts;
591 }
592 }
593 rcu_read_unlock();
594}
595EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
596
597/* VJ's idea. Save last timestamp seen from this destination and hold
598 * it at least for normal timewait interval to use for duplicate
599 * segment detection in subsequent connections, before they enter
600 * synchronized state.
601 */
602bool tcp_remember_stamp(struct sock *sk)
603{
604 struct dst_entry *dst = __sk_dst_get(sk);
605 bool ret = false;
606
607 if (dst) {
608 struct tcp_metrics_block *tm;
609
610 rcu_read_lock();
611 tm = tcp_get_metrics(sk, dst, true);
612 if (tm) {
613 struct tcp_sock *tp = tcp_sk(sk);
614
615 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
616 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
617 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
618 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
619 tm->tcpm_ts = tp->rx_opt.ts_recent;
620 }
621 ret = true;
622 }
623 rcu_read_unlock();
624 }
625 return ret;
626}
627
628bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
629{
630 struct tcp_metrics_block *tm;
631 bool ret = false;
632
633 rcu_read_lock();
634 tm = __tcp_get_metrics_tw(tw);
9a0a9502 635 if (tm) {
81166dd6
DM		 636 const struct tcp_timewait_sock *tcptw;
637 struct sock *sk = (struct sock *) tw;
638
639 tcptw = tcp_twsk(sk);
640 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
641 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
642 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
643 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
644 tm->tcpm_ts = tcptw->tw_ts_recent;
645 }
646 ret = true;
647 }
648 rcu_read_unlock();
649
650 return ret;
651}
652
1fe4c481
YC		 653static DEFINE_SEQLOCK(fastopen_seqlock);
654
655void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
aab48743
YC		 656 struct tcp_fastopen_cookie *cookie,
657 int *syn_loss, unsigned long *last_syn_loss)
1fe4c481
YC		 658{
659 struct tcp_metrics_block *tm;
660
661 rcu_read_lock();
662 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
663 if (tm) {
664 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
665 unsigned int seq;
666
667 do {
668 seq = read_seqbegin(&fastopen_seqlock);
669 if (tfom->mss)
670 *mss = tfom->mss;
671 *cookie = tfom->cookie;
aab48743
YC		 672 *syn_loss = tfom->syn_loss;
673 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
1fe4c481
YC		 674 } while (read_seqretry(&fastopen_seqlock, seq));
675 }
676 rcu_read_unlock();
677}
678
1fe4c481 679void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
aab48743 680 struct tcp_fastopen_cookie *cookie, bool syn_lost)
1fe4c481 681{
dccf76ca 682 struct dst_entry *dst = __sk_dst_get(sk);
1fe4c481
YC		 683 struct tcp_metrics_block *tm;
684
dccf76ca
ED		 685 if (!dst)
686 return;
1fe4c481 687 rcu_read_lock();
dccf76ca 688 tm = tcp_get_metrics(sk, dst, true);
1fe4c481
YC		 689 if (tm) {
690 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
691
692 write_seqlock_bh(&fastopen_seqlock);
c968601d
YC		 693 if (mss)
694 tfom->mss = mss;
695 if (cookie && cookie->len > 0)
1fe4c481 696 tfom->cookie = *cookie;
aab48743
YC		 697 if (syn_lost) {
698 ++tfom->syn_loss;
699 tfom->last_syn_loss = jiffies;
700 } else
701 tfom->syn_loss = 0;
1fe4c481
YC		 702 write_sequnlock_bh(&fastopen_seqlock);
703 }
704 rcu_read_unlock();
705}
706
d23ff701
JA		 707static struct genl_family tcp_metrics_nl_family = {
708 .id = GENL_ID_GENERATE,
709 .hdrsize = 0,
710 .name = TCP_METRICS_GENL_NAME,
711 .version = TCP_METRICS_GENL_VERSION,
712 .maxattr = TCP_METRICS_ATTR_MAX,
713 .netnsok = true,
714};
715
716static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
717 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, },
718 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY,
719 .len = sizeof(struct in6_addr), },
720 /* Following attributes are not received for GET/DEL,
721 * we keep them for reference
722 */
723#if 0
724 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, },
725 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, },
726 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, },
727 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, },
728 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, },
729 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, },
730 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, },
731 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
732 .len = TCP_FASTOPEN_COOKIE_MAX, },
733#endif
734};
735
736/* Add attributes, caller cancels its header on failure */
737static int tcp_metrics_fill_info(struct sk_buff *msg,
738 struct tcp_metrics_block *tm)
739{
740 struct nlattr *nest;
741 int i;
742
324fd55a 743 switch (tm->tcpm_daddr.family) {
d23ff701
JA		 744 case AF_INET:
745 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
324fd55a 746 tm->tcpm_daddr.addr.a4) < 0)
d23ff701
JA		 747 goto nla_put_failure;
748 break;
749 case AF_INET6:
750 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
324fd55a 751 tm->tcpm_daddr.addr.a6) < 0)
d23ff701
JA		 752 goto nla_put_failure;
753 break;
754 default:
755 return -EAFNOSUPPORT;
756 }
757
758 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
759 jiffies - tm->tcpm_stamp) < 0)
760 goto nla_put_failure;
761 if (tm->tcpm_ts_stamp) {
762 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
763 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
764 goto nla_put_failure;
765 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
766 tm->tcpm_ts) < 0)
767 goto nla_put_failure;
768 }
769
770 {
771 int n = 0;
772
773 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
774 if (!nest)
775 goto nla_put_failure;
776 for (i = 0; i < TCP_METRIC_MAX + 1; i++) {
777 if (!tm->tcpm_vals[i])
778 continue;
779 if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0)
780 goto nla_put_failure;
781 n++;
782 }
783 if (n)
784 nla_nest_end(msg, nest);
785 else
786 nla_nest_cancel(msg, nest);
787 }
788
789 {
790 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
791 unsigned int seq;
792
793 do {
794 seq = read_seqbegin(&fastopen_seqlock);
795 tfom_copy[0] = tm->tcpm_fastopen;
796 } while (read_seqretry(&fastopen_seqlock, seq));
797
798 tfom = tfom_copy;
799 if (tfom->mss &&
800 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
801 tfom->mss) < 0)
802 goto nla_put_failure;
803 if (tfom->syn_loss &&
804 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
805 tfom->syn_loss) < 0 ||
806 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
807 jiffies - tfom->last_syn_loss) < 0))
808 goto nla_put_failure;
809 if (tfom->cookie.len > 0 &&
810 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
811 tfom->cookie.len, tfom->cookie.val) < 0)
812 goto nla_put_failure;
813 }
814
815 return 0;
816
817nla_put_failure:
818 return -EMSGSIZE;
819}
820
821static int tcp_metrics_dump_info(struct sk_buff *skb,
822 struct netlink_callback *cb,
823 struct tcp_metrics_block *tm)
824{
825 void *hdr;
826
15e47304 827 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
d23ff701
JA		 828 &tcp_metrics_nl_family, NLM_F_MULTI,
829 TCP_METRICS_CMD_GET);
830 if (!hdr)
831 return -EMSGSIZE;
832
833 if (tcp_metrics_fill_info(skb, tm) < 0)
834 goto nla_put_failure;
835
836 return genlmsg_end(skb, hdr);
837
838nla_put_failure:
839 genlmsg_cancel(skb, hdr);
840 return -EMSGSIZE;
841}
842
843static int tcp_metrics_nl_dump(struct sk_buff *skb,
844 struct netlink_callback *cb)
845{
846 struct net *net = sock_net(skb->sk);
847 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
848 unsigned int row, s_row = cb->args[0];
849 int s_col = cb->args[1], col = s_col;
850
851 for (row = s_row; row < max_rows; row++, s_col = 0) {
852 struct tcp_metrics_block *tm;
853 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
854
855 rcu_read_lock();
856 for (col = 0, tm = rcu_dereference(hb->chain); tm;
857 tm = rcu_dereference(tm->tcpm_next), col++) {
858 if (col < s_col)
859 continue;
860 if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
861 rcu_read_unlock();
862 goto done;
863 }
864 }
865 rcu_read_unlock();
866 }
867
868done:
869 cb->args[0] = row;
870 cb->args[1] = col;
871 return skb->len;
872}
873
874static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
875 unsigned int *hash, int optional)
876{
877 struct nlattr *a;
878
879 a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV4];
880 if (a) {
881 addr->family = AF_INET;
882 addr->addr.a4 = nla_get_be32(a);
883 *hash = (__force unsigned int) addr->addr.a4;
884 return 0;
885 }
886 a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV6];
887 if (a) {
2c42a3fb 888 if (nla_len(a) != sizeof(struct in6_addr))
d23ff701
JA		 889 return -EINVAL;
890 addr->family = AF_INET6;
891 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
892 *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
893 return 0;
894 }
895 return optional ? 1 : -EAFNOSUPPORT;
896}
897
898static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
899{
900 struct tcp_metrics_block *tm;
324fd55a 901 struct inetpeer_addr daddr;
d23ff701
JA		 902 unsigned int hash;
903 struct sk_buff *msg;
904 struct net *net = genl_info_net(info);
905 void *reply;
906 int ret;
907
324fd55a 908 ret = parse_nl_addr(info, &daddr, &hash, 0);
d23ff701
JA		 909 if (ret < 0)
910 return ret;
911
912 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
913 if (!msg)
914 return -ENOMEM;
915
916 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
917 info->genlhdr->cmd);
918 if (!reply)
919 goto nla_put_failure;
920
921 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
922 ret = -ESRCH;
923 rcu_read_lock();
924 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
925 tm = rcu_dereference(tm->tcpm_next)) {
324fd55a 926 if (addr_same(&tm->tcpm_daddr, &daddr)) {
d23ff701
JA		 927 ret = tcp_metrics_fill_info(msg, tm);
928 break;
929 }
930 }
931 rcu_read_unlock();
932 if (ret < 0)
933 goto out_free;
934
935 genlmsg_end(msg, reply);
936 return genlmsg_reply(msg, info);
937
938nla_put_failure:
939 ret = -EMSGSIZE;
940
941out_free:
942 nlmsg_free(msg);
943 return ret;
944}
945
946#define deref_locked_genl(p) \
947 rcu_dereference_protected(p, lockdep_genl_is_held() && \
948 lockdep_is_held(&tcp_metrics_lock))
949
950#define deref_genl(p) rcu_dereference_protected(p, lockdep_genl_is_held())
951
952static int tcp_metrics_flush_all(struct net *net)
953{
954 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
955 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
956 struct tcp_metrics_block *tm;
957 unsigned int row;
958
959 for (row = 0; row < max_rows; row++, hb++) {
960 spin_lock_bh(&tcp_metrics_lock);
961 tm = deref_locked_genl(hb->chain);
962 if (tm)
963 hb->chain = NULL;
964 spin_unlock_bh(&tcp_metrics_lock);
965 while (tm) {
966 struct tcp_metrics_block *next;
967
968 next = deref_genl(tm->tcpm_next);
969 kfree_rcu(tm, rcu_head);
970 tm = next;
971 }
972 }
973 return 0;
974}
975
976static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
977{
978 struct tcpm_hash_bucket *hb;
979 struct tcp_metrics_block *tm;
980 struct tcp_metrics_block __rcu **pp;
324fd55a 981 struct inetpeer_addr daddr;
d23ff701
JA		 982 unsigned int hash;
983 struct net *net = genl_info_net(info);
984 int ret;
985
324fd55a 986 ret = parse_nl_addr(info, &daddr, &hash, 1);
d23ff701
JA		 987 if (ret < 0)
988 return ret;
989 if (ret > 0)
990 return tcp_metrics_flush_all(net);
991
992 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
993 hb = net->ipv4.tcp_metrics_hash + hash;
994 pp = &hb->chain;
995 spin_lock_bh(&tcp_metrics_lock);
996 for (tm = deref_locked_genl(*pp); tm;
997 pp = &tm->tcpm_next, tm = deref_locked_genl(*pp)) {
324fd55a 998 if (addr_same(&tm->tcpm_daddr, &daddr)) {
d23ff701
JA		 999 *pp = tm->tcpm_next;
1000 break;
1001 }
1002 }
1003 spin_unlock_bh(&tcp_metrics_lock);
1004 if (!tm)
1005 return -ESRCH;
1006 kfree_rcu(tm, rcu_head);
1007 return 0;
1008}
1009
4534de83 1010static const struct genl_ops tcp_metrics_nl_ops[] = {
d23ff701
JA		 1011 {
1012 .cmd = TCP_METRICS_CMD_GET,
1013 .doit = tcp_metrics_nl_cmd_get,
1014 .dumpit = tcp_metrics_nl_dump,
1015 .policy = tcp_metrics_nl_policy,
1016 .flags = GENL_ADMIN_PERM,
1017 },
1018 {
1019 .cmd = TCP_METRICS_CMD_DEL,
1020 .doit = tcp_metrics_nl_cmd_del,
1021 .policy = tcp_metrics_nl_policy,
1022 .flags = GENL_ADMIN_PERM,
1023 },
1024};
1025
5815d5e7 1026static unsigned int tcpmhash_entries;
51c5d0c4
DM		 1027static int __init set_tcpmhash_entries(char *str)
1028{
1029 ssize_t ret;
1030
1031 if (!str)
1032 return 0;
1033
5815d5e7 1034 ret = kstrtouint(str, 0, &tcpmhash_entries);
51c5d0c4
DM		 1035 if (ret)
1036 return 0;
1037
1038 return 1;
1039}
1040__setup("tcpmhash_entries=", set_tcpmhash_entries);
1041
1042static int __net_init tcp_net_metrics_init(struct net *net)
1043{
5815d5e7
ED		 1044 size_t size;
1045 unsigned int slots;
51c5d0c4
DM		 1046
1047 slots = tcpmhash_entries;
1048 if (!slots) {
1049 if (totalram_pages >= 128 * 1024)
1050 slots = 16 * 1024;
1051 else
1052 slots = 8 * 1024;
1053 }
1054
5815d5e7
ED		 1055 net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1056 size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
51c5d0c4 1057
976a702a
ED		 1058 net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1059 if (!net->ipv4.tcp_metrics_hash)
1060 net->ipv4.tcp_metrics_hash = vzalloc(size);
1061
51c5d0c4
DM		 1062 if (!net->ipv4.tcp_metrics_hash)
1063 return -ENOMEM;
1064
51c5d0c4
DM		 1065 return 0;
1066}
1067
1068static void __net_exit tcp_net_metrics_exit(struct net *net)
1069{
36471012
ED		 1070 unsigned int i;
1071
1072 for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1073 struct tcp_metrics_block *tm, *next;
1074
1075 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1076 while (tm) {
1077 next = rcu_dereference_protected(tm->tcpm_next, 1);
1078 kfree(tm);
1079 tm = next;
1080 }
1081 }
976a702a
ED		 1082 if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1083 vfree(net->ipv4.tcp_metrics_hash);
1084 else
1085 kfree(net->ipv4.tcp_metrics_hash);
51c5d0c4
DM		 1086}
1087
1088static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1089 .init = tcp_net_metrics_init,
1090 .exit = tcp_net_metrics_exit,
1091};
1092
1093void __init tcp_metrics_init(void)
1094{
d23ff701
JA		 1095 int ret;
1096
1097 ret = register_pernet_subsys(&tcp_net_metrics_ops);
1098 if (ret < 0)
1099 goto cleanup;
1100 ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
c53ed742 1101 tcp_metrics_nl_ops);
d23ff701
JA		 1102 if (ret < 0)
1103 goto cleanup_subsys;
1104 return;
1105
1106cleanup_subsys:
1107 unregister_pernet_subsys(&tcp_net_metrics_ops);
1108
1109cleanup:
1110 return;
51c5d0c4 1111}
Linux 6.x block layer and io_uring tree(s)