tcp_metrics.c revision 00ca9c5b2b11d44eaf20a4b647efc999734323ec
1#include <linux/rcupdate.h> 2#include <linux/spinlock.h> 3#include <linux/jiffies.h> 4#include <linux/module.h> 5#include <linux/cache.h> 6#include <linux/slab.h> 7#include <linux/init.h> 8#include <linux/tcp.h> 9#include <linux/hash.h> 10#include <linux/tcp_metrics.h> 11#include <linux/vmalloc.h> 12 13#include <net/inet_connection_sock.h> 14#include <net/net_namespace.h> 15#include <net/request_sock.h> 16#include <net/inetpeer.h> 17#include <net/sock.h> 18#include <net/ipv6.h> 19#include <net/dst.h> 20#include <net/tcp.h> 21#include <net/genetlink.h> 22 23int sysctl_tcp_nometrics_save __read_mostly; 24 25static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr, 26 const struct inetpeer_addr *daddr, 27 struct net *net, unsigned int hash); 28 29struct tcp_fastopen_metrics { 30 u16 mss; 31 u16 syn_loss:10; /* Recurring Fast Open SYN losses */ 32 unsigned long last_syn_loss; /* Last Fast Open SYN loss */ 33 struct tcp_fastopen_cookie cookie; 34}; 35 36struct tcp_metrics_block { 37 struct tcp_metrics_block __rcu *tcpm_next; 38 struct inetpeer_addr tcpm_saddr; 39 struct inetpeer_addr tcpm_daddr; 40 unsigned long tcpm_stamp; 41 u32 tcpm_ts; 42 u32 tcpm_ts_stamp; 43 u32 tcpm_lock; 44 u32 tcpm_vals[TCP_METRIC_MAX + 1]; 45 struct tcp_fastopen_metrics tcpm_fastopen; 46 47 struct rcu_head rcu_head; 48}; 49 50static bool tcp_metric_locked(struct tcp_metrics_block *tm, 51 enum tcp_metric_index idx) 52{ 53 return tm->tcpm_lock & (1 << idx); 54} 55 56static u32 tcp_metric_get(struct tcp_metrics_block *tm, 57 enum tcp_metric_index idx) 58{ 59 return tm->tcpm_vals[idx]; 60} 61 62static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm, 63 enum tcp_metric_index idx) 64{ 65 return msecs_to_jiffies(tm->tcpm_vals[idx]); 66} 67 68static void tcp_metric_set(struct tcp_metrics_block *tm, 69 enum tcp_metric_index idx, 70 u32 val) 71{ 72 tm->tcpm_vals[idx] = val; 73} 74 75static void tcp_metric_set_msecs(struct tcp_metrics_block *tm, 76 enum tcp_metric_index idx, 77 u32 val) 78{ 79 tm->tcpm_vals[idx] = jiffies_to_msecs(val); 80} 81 82static bool addr_same(const struct inetpeer_addr *a, 83 const struct inetpeer_addr *b) 84{ 85 const struct in6_addr *a6, *b6; 86 87 if (a->family != b->family) 88 return false; 89 if (a->family == AF_INET) 90 return a->addr.a4 == b->addr.a4; 91 92 a6 = (const struct in6_addr *) &a->addr.a6[0]; 93 b6 = (const struct in6_addr *) &b->addr.a6[0]; 94 95 return ipv6_addr_equal(a6, b6); 96} 97 98struct tcpm_hash_bucket { 99 struct tcp_metrics_block __rcu *chain; 100}; 101 102static DEFINE_SPINLOCK(tcp_metrics_lock); 103 104static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst, 105 bool fastopen_clear) 106{ 107 u32 val; 108 109 tm->tcpm_stamp = jiffies; 110 111 val = 0; 112 if (dst_metric_locked(dst, RTAX_RTT)) 113 val |= 1 << TCP_METRIC_RTT; 114 if (dst_metric_locked(dst, RTAX_RTTVAR)) 115 val |= 1 << TCP_METRIC_RTTVAR; 116 if (dst_metric_locked(dst, RTAX_SSTHRESH)) 117 val |= 1 << TCP_METRIC_SSTHRESH; 118 if (dst_metric_locked(dst, RTAX_CWND)) 119 val |= 1 << TCP_METRIC_CWND; 120 if (dst_metric_locked(dst, RTAX_REORDERING)) 121 val |= 1 << TCP_METRIC_REORDERING; 122 tm->tcpm_lock = val; 123 124 tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT); 125 tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR); 126 tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH); 127 tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND); 128 tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING); 129 tm->tcpm_ts = 0; 130 tm->tcpm_ts_stamp = 0; 131 if (fastopen_clear) { 132 tm->tcpm_fastopen.mss = 0; 133 tm->tcpm_fastopen.syn_loss = 0; 134 tm->tcpm_fastopen.cookie.len = 0; 135 } 136} 137 138#define TCP_METRICS_TIMEOUT (60 * 60 * HZ) 139 140static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst) 141{ 142 if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT))) 143 tcpm_suck_dst(tm, dst, false); 144} 145 146#define TCP_METRICS_RECLAIM_DEPTH 5 147#define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL 148 149static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst, 150 struct inetpeer_addr *saddr, 151 struct inetpeer_addr *daddr, 152 unsigned int hash) 153{ 154 struct tcp_metrics_block *tm; 155 struct net *net; 156 bool reclaim = false; 157 158 spin_lock_bh(&tcp_metrics_lock); 159 net = dev_net(dst->dev); 160 161 /* While waiting for the spin-lock the cache might have been populated 162 * with this entry and so we have to check again. 163 */ 164 tm = __tcp_get_metrics(saddr, daddr, net, hash); 165 if (tm == TCP_METRICS_RECLAIM_PTR) { 166 reclaim = true; 167 tm = NULL; 168 } 169 if (tm) { 170 tcpm_check_stamp(tm, dst); 171 goto out_unlock; 172 } 173 174 if (unlikely(reclaim)) { 175 struct tcp_metrics_block *oldest; 176 177 oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); 178 for (tm = rcu_dereference(oldest->tcpm_next); tm; 179 tm = rcu_dereference(tm->tcpm_next)) { 180 if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp)) 181 oldest = tm; 182 } 183 tm = oldest; 184 } else { 185 tm = kmalloc(sizeof(*tm), GFP_ATOMIC); 186 if (!tm) 187 goto out_unlock; 188 } 189 tm->tcpm_saddr = *saddr; 190 tm->tcpm_daddr = *daddr; 191 192 tcpm_suck_dst(tm, dst, true); 193 194 if (likely(!reclaim)) { 195 tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain; 196 rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm); 197 } 198 199out_unlock: 200 spin_unlock_bh(&tcp_metrics_lock); 201 return tm; 202} 203 204static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth) 205{ 206 if (tm) 207 return tm; 208 if (depth > TCP_METRICS_RECLAIM_DEPTH) 209 return TCP_METRICS_RECLAIM_PTR; 210 return NULL; 211} 212 213static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr, 214 const struct inetpeer_addr *daddr, 215 struct net *net, unsigned int hash) 216{ 217 struct tcp_metrics_block *tm; 218 int depth = 0; 219 220 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; 221 tm = rcu_dereference(tm->tcpm_next)) { 222 if (addr_same(&tm->tcpm_saddr, saddr) && 223 addr_same(&tm->tcpm_daddr, daddr)) 224 break; 225 depth++; 226 } 227 return tcp_get_encode(tm, depth); 228} 229 230static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req, 231 struct dst_entry *dst) 232{ 233 struct tcp_metrics_block *tm; 234 struct inetpeer_addr saddr, daddr; 235 unsigned int hash; 236 struct net *net; 237 238 saddr.family = req->rsk_ops->family; 239 daddr.family = req->rsk_ops->family; 240 switch (daddr.family) { 241 case AF_INET: 242 saddr.addr.a4 = inet_rsk(req)->ir_loc_addr; 243 daddr.addr.a4 = inet_rsk(req)->ir_rmt_addr; 244 hash = (__force unsigned int) daddr.addr.a4; 245 break; 246#if IS_ENABLED(CONFIG_IPV6) 247 case AF_INET6: 248 *(struct in6_addr *)saddr.addr.a6 = inet_rsk(req)->ir_v6_loc_addr; 249 *(struct in6_addr *)daddr.addr.a6 = inet_rsk(req)->ir_v6_rmt_addr; 250 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr); 251 break; 252#endif 253 default: 254 return NULL; 255 } 256 257 net = dev_net(dst->dev); 258 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); 259 260 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; 261 tm = rcu_dereference(tm->tcpm_next)) { 262 if (addr_same(&tm->tcpm_saddr, &saddr) && 263 addr_same(&tm->tcpm_daddr, &daddr)) 264 break; 265 } 266 tcpm_check_stamp(tm, dst); 267 return tm; 268} 269 270static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw) 271{ 272 struct tcp_metrics_block *tm; 273 struct inetpeer_addr saddr, daddr; 274 unsigned int hash; 275 struct net *net; 276 277 saddr.family = tw->tw_family; 278 daddr.family = tw->tw_family; 279 switch (daddr.family) { 280 case AF_INET: 281 saddr.addr.a4 = tw->tw_rcv_saddr; 282 daddr.addr.a4 = tw->tw_daddr; 283 hash = (__force unsigned int) daddr.addr.a4; 284 break; 285#if IS_ENABLED(CONFIG_IPV6) 286 case AF_INET6: 287 *(struct in6_addr *)saddr.addr.a6 = tw->tw_v6_rcv_saddr; 288 *(struct in6_addr *)daddr.addr.a6 = tw->tw_v6_daddr; 289 hash = ipv6_addr_hash(&tw->tw_v6_daddr); 290 break; 291#endif 292 default: 293 return NULL; 294 } 295 296 net = twsk_net(tw); 297 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); 298 299 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; 300 tm = rcu_dereference(tm->tcpm_next)) { 301 if (addr_same(&tm->tcpm_saddr, &saddr) && 302 addr_same(&tm->tcpm_daddr, &daddr)) 303 break; 304 } 305 return tm; 306} 307 308static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk, 309 struct dst_entry *dst, 310 bool create) 311{ 312 struct tcp_metrics_block *tm; 313 struct inetpeer_addr saddr, daddr; 314 unsigned int hash; 315 struct net *net; 316 317 saddr.family = sk->sk_family; 318 daddr.family = sk->sk_family; 319 switch (daddr.family) { 320 case AF_INET: 321 saddr.addr.a4 = inet_sk(sk)->inet_saddr; 322 daddr.addr.a4 = inet_sk(sk)->inet_daddr; 323 hash = (__force unsigned int) daddr.addr.a4; 324 break; 325#if IS_ENABLED(CONFIG_IPV6) 326 case AF_INET6: 327 *(struct in6_addr *)saddr.addr.a6 = sk->sk_v6_rcv_saddr; 328 *(struct in6_addr *)daddr.addr.a6 = sk->sk_v6_daddr; 329 hash = ipv6_addr_hash(&sk->sk_v6_daddr); 330 break; 331#endif 332 default: 333 return NULL; 334 } 335 336 net = dev_net(dst->dev); 337 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); 338 339 tm = __tcp_get_metrics(&saddr, &daddr, net, hash); 340 if (tm == TCP_METRICS_RECLAIM_PTR) 341 tm = NULL; 342 if (!tm && create) 343 tm = tcpm_new(dst, &saddr, &daddr, hash); 344 else 345 tcpm_check_stamp(tm, dst); 346 347 return tm; 348} 349 350/* Save metrics learned by this TCP session. This function is called 351 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT 352 * or goes from LAST-ACK to CLOSE. 353 */ 354void tcp_update_metrics(struct sock *sk) 355{ 356 const struct inet_connection_sock *icsk = inet_csk(sk); 357 struct dst_entry *dst = __sk_dst_get(sk); 358 struct tcp_sock *tp = tcp_sk(sk); 359 struct tcp_metrics_block *tm; 360 unsigned long rtt; 361 u32 val; 362 int m; 363 364 if (sysctl_tcp_nometrics_save || !dst) 365 return; 366 367 if (dst->flags & DST_HOST) 368 dst_confirm(dst); 369 370 rcu_read_lock(); 371 if (icsk->icsk_backoff || !tp->srtt) { 372 /* This session failed to estimate rtt. Why? 373 * Probably, no packets returned in time. Reset our 374 * results. 375 */ 376 tm = tcp_get_metrics(sk, dst, false); 377 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT)) 378 tcp_metric_set(tm, TCP_METRIC_RTT, 0); 379 goto out_unlock; 380 } else 381 tm = tcp_get_metrics(sk, dst, true); 382 383 if (!tm) 384 goto out_unlock; 385 386 rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT); 387 m = rtt - tp->srtt; 388 389 /* If newly calculated rtt larger than stored one, store new 390 * one. Otherwise, use EWMA. Remember, rtt overestimation is 391 * always better than underestimation. 392 */ 393 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) { 394 if (m <= 0) 395 rtt = tp->srtt; 396 else 397 rtt -= (m >> 3); 398 tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt); 399 } 400 401 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) { 402 unsigned long var; 403 404 if (m < 0) 405 m = -m; 406 407 /* Scale deviation to rttvar fixed point */ 408 m >>= 1; 409 if (m < tp->mdev) 410 m = tp->mdev; 411 412 var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR); 413 if (m >= var) 414 var = m; 415 else 416 var -= (var - m) >> 2; 417 418 tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var); 419 } 420 421 if (tcp_in_initial_slowstart(tp)) { 422 /* Slow start still did not finish. */ 423 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) { 424 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); 425 if (val && (tp->snd_cwnd >> 1) > val) 426 tcp_metric_set(tm, TCP_METRIC_SSTHRESH, 427 tp->snd_cwnd >> 1); 428 } 429 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { 430 val = tcp_metric_get(tm, TCP_METRIC_CWND); 431 if (tp->snd_cwnd > val) 432 tcp_metric_set(tm, TCP_METRIC_CWND, 433 tp->snd_cwnd); 434 } 435 } else if (tp->snd_cwnd > tp->snd_ssthresh && 436 icsk->icsk_ca_state == TCP_CA_Open) { 437 /* Cong. avoidance phase, cwnd is reliable. */ 438 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) 439 tcp_metric_set(tm, TCP_METRIC_SSTHRESH, 440 max(tp->snd_cwnd >> 1, tp->snd_ssthresh)); 441 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { 442 val = tcp_metric_get(tm, TCP_METRIC_CWND); 443 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1); 444 } 445 } else { 446 /* Else slow start did not finish, cwnd is non-sense, 447 * ssthresh may be also invalid. 448 */ 449 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { 450 val = tcp_metric_get(tm, TCP_METRIC_CWND); 451 tcp_metric_set(tm, TCP_METRIC_CWND, 452 (val + tp->snd_ssthresh) >> 1); 453 } 454 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) { 455 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); 456 if (val && tp->snd_ssthresh > val) 457 tcp_metric_set(tm, TCP_METRIC_SSTHRESH, 458 tp->snd_ssthresh); 459 } 460 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) { 461 val = tcp_metric_get(tm, TCP_METRIC_REORDERING); 462 if (val < tp->reordering && 463 tp->reordering != sysctl_tcp_reordering) 464 tcp_metric_set(tm, TCP_METRIC_REORDERING, 465 tp->reordering); 466 } 467 } 468 tm->tcpm_stamp = jiffies; 469out_unlock: 470 rcu_read_unlock(); 471} 472 473/* Initialize metrics on socket. */ 474 475void tcp_init_metrics(struct sock *sk) 476{ 477 struct dst_entry *dst = __sk_dst_get(sk); 478 struct tcp_sock *tp = tcp_sk(sk); 479 struct tcp_metrics_block *tm; 480 u32 val, crtt = 0; /* cached RTT scaled by 8 */ 481 482 if (dst == NULL) 483 goto reset; 484 485 dst_confirm(dst); 486 487 rcu_read_lock(); 488 tm = tcp_get_metrics(sk, dst, true); 489 if (!tm) { 490 rcu_read_unlock(); 491 goto reset; 492 } 493 494 if (tcp_metric_locked(tm, TCP_METRIC_CWND)) 495 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND); 496 497 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); 498 if (val) { 499 tp->snd_ssthresh = val; 500 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 501 tp->snd_ssthresh = tp->snd_cwnd_clamp; 502 } else { 503 /* ssthresh may have been reduced unnecessarily during. 504 * 3WHS. Restore it back to its initial default. 505 */ 506 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 507 } 508 val = tcp_metric_get(tm, TCP_METRIC_REORDERING); 509 if (val && tp->reordering != val) { 510 tcp_disable_fack(tp); 511 tcp_disable_early_retrans(tp); 512 tp->reordering = val; 513 } 514 515 crtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT); 516 rcu_read_unlock(); 517reset: 518 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal 519 * to seed the RTO for later data packets because SYN packets are 520 * small. Use the per-dst cached values to seed the RTO but keep 521 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar). 522 * Later the RTO will be updated immediately upon obtaining the first 523 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only 524 * influences the first RTO but not later RTT estimation. 525 * 526 * But if RTT is not available from the SYN (due to retransmits or 527 * syn cookies) or the cache, force a conservative 3secs timeout. 528 * 529 * A bit of theory. RTT is time passed after "normal" sized packet 530 * is sent until it is ACKed. In normal circumstances sending small 531 * packets force peer to delay ACKs and calculation is correct too. 532 * The algorithm is adaptive and, provided we follow specs, it 533 * NEVER underestimate RTT. BUT! If peer tries to make some clever 534 * tricks sort of "quick acks" for time long enough to decrease RTT 535 * to low value, and then abruptly stops to do it and starts to delay 536 * ACKs, wait for troubles. 537 */ 538 if (crtt > tp->srtt) { 539 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */ 540 crtt >>= 3; 541 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk)); 542 } else if (tp->srtt == 0) { 543 /* RFC6298: 5.7 We've failed to get a valid RTT sample from 544 * 3WHS. This is most likely due to retransmission, 545 * including spurious one. Reset the RTO back to 3secs 546 * from the more aggressive 1sec to avoid more spurious 547 * retransmission. 548 */ 549 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK; 550 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK; 551 } 552 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 553 * retransmitted. In light of RFC6298 more aggressive 1sec 554 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 555 * retransmission has occurred. 556 */ 557 if (tp->total_retrans > 1) 558 tp->snd_cwnd = 1; 559 else 560 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 561 tp->snd_cwnd_stamp = tcp_time_stamp; 562} 563 564bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check) 565{ 566 struct tcp_metrics_block *tm; 567 bool ret; 568 569 if (!dst) 570 return false; 571 572 rcu_read_lock(); 573 tm = __tcp_get_metrics_req(req, dst); 574 if (paws_check) { 575 if (tm && 576 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL && 577 (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW) 578 ret = false; 579 else 580 ret = true; 581 } else { 582 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp) 583 ret = true; 584 else 585 ret = false; 586 } 587 rcu_read_unlock(); 588 589 return ret; 590} 591EXPORT_SYMBOL_GPL(tcp_peer_is_proven); 592 593void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst) 594{ 595 struct tcp_metrics_block *tm; 596 597 rcu_read_lock(); 598 tm = tcp_get_metrics(sk, dst, true); 599 if (tm) { 600 struct tcp_sock *tp = tcp_sk(sk); 601 602 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) { 603 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp; 604 tp->rx_opt.ts_recent = tm->tcpm_ts; 605 } 606 } 607 rcu_read_unlock(); 608} 609EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp); 610 611/* VJ's idea. Save last timestamp seen from this destination and hold 612 * it at least for normal timewait interval to use for duplicate 613 * segment detection in subsequent connections, before they enter 614 * synchronized state. 615 */ 616bool tcp_remember_stamp(struct sock *sk) 617{ 618 struct dst_entry *dst = __sk_dst_get(sk); 619 bool ret = false; 620 621 if (dst) { 622 struct tcp_metrics_block *tm; 623 624 rcu_read_lock(); 625 tm = tcp_get_metrics(sk, dst, true); 626 if (tm) { 627 struct tcp_sock *tp = tcp_sk(sk); 628 629 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 || 630 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL && 631 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) { 632 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp; 633 tm->tcpm_ts = tp->rx_opt.ts_recent; 634 } 635 ret = true; 636 } 637 rcu_read_unlock(); 638 } 639 return ret; 640} 641 642bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw) 643{ 644 struct tcp_metrics_block *tm; 645 bool ret = false; 646 647 rcu_read_lock(); 648 tm = __tcp_get_metrics_tw(tw); 649 if (tm) { 650 const struct tcp_timewait_sock *tcptw; 651 struct sock *sk = (struct sock *) tw; 652 653 tcptw = tcp_twsk(sk); 654 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 || 655 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL && 656 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) { 657 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp; 658 tm->tcpm_ts = tcptw->tw_ts_recent; 659 } 660 ret = true; 661 } 662 rcu_read_unlock(); 663 664 return ret; 665} 666 667static DEFINE_SEQLOCK(fastopen_seqlock); 668 669void tcp_fastopen_cache_get(struct sock *sk, u16 *mss, 670 struct tcp_fastopen_cookie *cookie, 671 int *syn_loss, unsigned long *last_syn_loss) 672{ 673 struct tcp_metrics_block *tm; 674 675 rcu_read_lock(); 676 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false); 677 if (tm) { 678 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen; 679 unsigned int seq; 680 681 do { 682 seq = read_seqbegin(&fastopen_seqlock); 683 if (tfom->mss) 684 *mss = tfom->mss; 685 *cookie = tfom->cookie; 686 *syn_loss = tfom->syn_loss; 687 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0; 688 } while (read_seqretry(&fastopen_seqlock, seq)); 689 } 690 rcu_read_unlock(); 691} 692 693void tcp_fastopen_cache_set(struct sock *sk, u16 mss, 694 struct tcp_fastopen_cookie *cookie, bool syn_lost) 695{ 696 struct dst_entry *dst = __sk_dst_get(sk); 697 struct tcp_metrics_block *tm; 698 699 if (!dst) 700 return; 701 rcu_read_lock(); 702 tm = tcp_get_metrics(sk, dst, true); 703 if (tm) { 704 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen; 705 706 write_seqlock_bh(&fastopen_seqlock); 707 if (mss) 708 tfom->mss = mss; 709 if (cookie && cookie->len > 0) 710 tfom->cookie = *cookie; 711 if (syn_lost) { 712 ++tfom->syn_loss; 713 tfom->last_syn_loss = jiffies; 714 } else 715 tfom->syn_loss = 0; 716 write_sequnlock_bh(&fastopen_seqlock); 717 } 718 rcu_read_unlock(); 719} 720 721static struct genl_family tcp_metrics_nl_family = { 722 .id = GENL_ID_GENERATE, 723 .hdrsize = 0, 724 .name = TCP_METRICS_GENL_NAME, 725 .version = TCP_METRICS_GENL_VERSION, 726 .maxattr = TCP_METRICS_ATTR_MAX, 727 .netnsok = true, 728}; 729 730static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = { 731 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, }, 732 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY, 733 .len = sizeof(struct in6_addr), }, 734 /* Following attributes are not received for GET/DEL, 735 * we keep them for reference 736 */ 737#if 0 738 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, }, 739 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, }, 740 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, }, 741 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, }, 742 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, }, 743 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, }, 744 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, }, 745 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY, 746 .len = TCP_FASTOPEN_COOKIE_MAX, }, 747#endif 748}; 749 750/* Add attributes, caller cancels its header on failure */ 751static int tcp_metrics_fill_info(struct sk_buff *msg, 752 struct tcp_metrics_block *tm) 753{ 754 struct nlattr *nest; 755 int i; 756 757 switch (tm->tcpm_daddr.family) { 758 case AF_INET: 759 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4, 760 tm->tcpm_daddr.addr.a4) < 0) 761 goto nla_put_failure; 762 if (nla_put_be32(msg, TCP_METRICS_ATTR_SADDR_IPV4, 763 tm->tcpm_saddr.addr.a4) < 0) 764 goto nla_put_failure; 765 break; 766 case AF_INET6: 767 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16, 768 tm->tcpm_daddr.addr.a6) < 0) 769 goto nla_put_failure; 770 if (nla_put(msg, TCP_METRICS_ATTR_SADDR_IPV6, 16, 771 tm->tcpm_saddr.addr.a6) < 0) 772 goto nla_put_failure; 773 break; 774 default: 775 return -EAFNOSUPPORT; 776 } 777 778 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE, 779 jiffies - tm->tcpm_stamp) < 0) 780 goto nla_put_failure; 781 if (tm->tcpm_ts_stamp) { 782 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP, 783 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0) 784 goto nla_put_failure; 785 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL, 786 tm->tcpm_ts) < 0) 787 goto nla_put_failure; 788 } 789 790 { 791 int n = 0; 792 793 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS); 794 if (!nest) 795 goto nla_put_failure; 796 for (i = 0; i < TCP_METRIC_MAX + 1; i++) { 797 if (!tm->tcpm_vals[i]) 798 continue; 799 if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0) 800 goto nla_put_failure; 801 n++; 802 } 803 if (n) 804 nla_nest_end(msg, nest); 805 else 806 nla_nest_cancel(msg, nest); 807 } 808 809 { 810 struct tcp_fastopen_metrics tfom_copy[1], *tfom; 811 unsigned int seq; 812 813 do { 814 seq = read_seqbegin(&fastopen_seqlock); 815 tfom_copy[0] = tm->tcpm_fastopen; 816 } while (read_seqretry(&fastopen_seqlock, seq)); 817 818 tfom = tfom_copy; 819 if (tfom->mss && 820 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS, 821 tfom->mss) < 0) 822 goto nla_put_failure; 823 if (tfom->syn_loss && 824 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS, 825 tfom->syn_loss) < 0 || 826 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS, 827 jiffies - tfom->last_syn_loss) < 0)) 828 goto nla_put_failure; 829 if (tfom->cookie.len > 0 && 830 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE, 831 tfom->cookie.len, tfom->cookie.val) < 0) 832 goto nla_put_failure; 833 } 834 835 return 0; 836 837nla_put_failure: 838 return -EMSGSIZE; 839} 840 841static int tcp_metrics_dump_info(struct sk_buff *skb, 842 struct netlink_callback *cb, 843 struct tcp_metrics_block *tm) 844{ 845 void *hdr; 846 847 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq, 848 &tcp_metrics_nl_family, NLM_F_MULTI, 849 TCP_METRICS_CMD_GET); 850 if (!hdr) 851 return -EMSGSIZE; 852 853 if (tcp_metrics_fill_info(skb, tm) < 0) 854 goto nla_put_failure; 855 856 return genlmsg_end(skb, hdr); 857 858nla_put_failure: 859 genlmsg_cancel(skb, hdr); 860 return -EMSGSIZE; 861} 862 863static int tcp_metrics_nl_dump(struct sk_buff *skb, 864 struct netlink_callback *cb) 865{ 866 struct net *net = sock_net(skb->sk); 867 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log; 868 unsigned int row, s_row = cb->args[0]; 869 int s_col = cb->args[1], col = s_col; 870 871 for (row = s_row; row < max_rows; row++, s_col = 0) { 872 struct tcp_metrics_block *tm; 873 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row; 874 875 rcu_read_lock(); 876 for (col = 0, tm = rcu_dereference(hb->chain); tm; 877 tm = rcu_dereference(tm->tcpm_next), col++) { 878 if (col < s_col) 879 continue; 880 if (tcp_metrics_dump_info(skb, cb, tm) < 0) { 881 rcu_read_unlock(); 882 goto done; 883 } 884 } 885 rcu_read_unlock(); 886 } 887 888done: 889 cb->args[0] = row; 890 cb->args[1] = col; 891 return skb->len; 892} 893 894static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr, 895 unsigned int *hash, int optional, int v4, int v6) 896{ 897 struct nlattr *a; 898 899 a = info->attrs[v4]; 900 if (a) { 901 addr->family = AF_INET; 902 addr->addr.a4 = nla_get_be32(a); 903 if (hash) 904 *hash = (__force unsigned int) addr->addr.a4; 905 return 0; 906 } 907 a = info->attrs[v6]; 908 if (a) { 909 if (nla_len(a) != sizeof(struct in6_addr)) 910 return -EINVAL; 911 addr->family = AF_INET6; 912 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6)); 913 if (hash) 914 *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6); 915 return 0; 916 } 917 return optional ? 1 : -EAFNOSUPPORT; 918} 919 920static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr, 921 unsigned int *hash, int optional) 922{ 923 return __parse_nl_addr(info, addr, hash, optional, 924 TCP_METRICS_ATTR_ADDR_IPV4, 925 TCP_METRICS_ATTR_ADDR_IPV6); 926} 927 928static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr) 929{ 930 return __parse_nl_addr(info, addr, NULL, 0, 931 TCP_METRICS_ATTR_SADDR_IPV4, 932 TCP_METRICS_ATTR_SADDR_IPV6); 933} 934 935static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info) 936{ 937 struct tcp_metrics_block *tm; 938 struct inetpeer_addr saddr, daddr; 939 unsigned int hash; 940 struct sk_buff *msg; 941 struct net *net = genl_info_net(info); 942 void *reply; 943 int ret; 944 bool src = true; 945 946 ret = parse_nl_addr(info, &daddr, &hash, 0); 947 if (ret < 0) 948 return ret; 949 950 ret = parse_nl_saddr(info, &saddr); 951 if (ret < 0) 952 src = false; 953 954 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL); 955 if (!msg) 956 return -ENOMEM; 957 958 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0, 959 info->genlhdr->cmd); 960 if (!reply) 961 goto nla_put_failure; 962 963 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); 964 ret = -ESRCH; 965 rcu_read_lock(); 966 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; 967 tm = rcu_dereference(tm->tcpm_next)) { 968 if (addr_same(&tm->tcpm_daddr, &daddr) && 969 (!src || addr_same(&tm->tcpm_saddr, &saddr))) { 970 ret = tcp_metrics_fill_info(msg, tm); 971 break; 972 } 973 } 974 rcu_read_unlock(); 975 if (ret < 0) 976 goto out_free; 977 978 genlmsg_end(msg, reply); 979 return genlmsg_reply(msg, info); 980 981nla_put_failure: 982 ret = -EMSGSIZE; 983 984out_free: 985 nlmsg_free(msg); 986 return ret; 987} 988 989#define deref_locked_genl(p) \ 990 rcu_dereference_protected(p, lockdep_genl_is_held() && \ 991 lockdep_is_held(&tcp_metrics_lock)) 992 993#define deref_genl(p) rcu_dereference_protected(p, lockdep_genl_is_held()) 994 995static int tcp_metrics_flush_all(struct net *net) 996{ 997 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log; 998 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash; 999 struct tcp_metrics_block *tm; 1000 unsigned int row; 1001 1002 for (row = 0; row < max_rows; row++, hb++) { 1003 spin_lock_bh(&tcp_metrics_lock); 1004 tm = deref_locked_genl(hb->chain); 1005 if (tm) 1006 hb->chain = NULL; 1007 spin_unlock_bh(&tcp_metrics_lock); 1008 while (tm) { 1009 struct tcp_metrics_block *next; 1010 1011 next = deref_genl(tm->tcpm_next); 1012 kfree_rcu(tm, rcu_head); 1013 tm = next; 1014 } 1015 } 1016 return 0; 1017} 1018 1019static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info) 1020{ 1021 struct tcpm_hash_bucket *hb; 1022 struct tcp_metrics_block *tm; 1023 struct tcp_metrics_block __rcu **pp; 1024 struct inetpeer_addr saddr, daddr; 1025 unsigned int hash; 1026 struct net *net = genl_info_net(info); 1027 int ret; 1028 bool src = true, found = false; 1029 1030 ret = parse_nl_addr(info, &daddr, &hash, 1); 1031 if (ret < 0) 1032 return ret; 1033 if (ret > 0) 1034 return tcp_metrics_flush_all(net); 1035 ret = parse_nl_saddr(info, &saddr); 1036 if (ret < 0) 1037 src = false; 1038 1039 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); 1040 hb = net->ipv4.tcp_metrics_hash + hash; 1041 pp = &hb->chain; 1042 spin_lock_bh(&tcp_metrics_lock); 1043 for (tm = deref_locked_genl(*pp); tm; tm = deref_locked_genl(*pp)) { 1044 if (addr_same(&tm->tcpm_daddr, &daddr) && 1045 (!src || addr_same(&tm->tcpm_saddr, &saddr))) { 1046 *pp = tm->tcpm_next; 1047 kfree_rcu(tm, rcu_head); 1048 found = true; 1049 } else { 1050 pp = &tm->tcpm_next; 1051 } 1052 } 1053 spin_unlock_bh(&tcp_metrics_lock); 1054 if (!found) 1055 return -ESRCH; 1056 return 0; 1057} 1058 1059static const struct genl_ops tcp_metrics_nl_ops[] = { 1060 { 1061 .cmd = TCP_METRICS_CMD_GET, 1062 .doit = tcp_metrics_nl_cmd_get, 1063 .dumpit = tcp_metrics_nl_dump, 1064 .policy = tcp_metrics_nl_policy, 1065 .flags = GENL_ADMIN_PERM, 1066 }, 1067 { 1068 .cmd = TCP_METRICS_CMD_DEL, 1069 .doit = tcp_metrics_nl_cmd_del, 1070 .policy = tcp_metrics_nl_policy, 1071 .flags = GENL_ADMIN_PERM, 1072 }, 1073}; 1074 1075static unsigned int tcpmhash_entries; 1076static int __init set_tcpmhash_entries(char *str) 1077{ 1078 ssize_t ret; 1079 1080 if (!str) 1081 return 0; 1082 1083 ret = kstrtouint(str, 0, &tcpmhash_entries); 1084 if (ret) 1085 return 0; 1086 1087 return 1; 1088} 1089__setup("tcpmhash_entries=", set_tcpmhash_entries); 1090 1091static int __net_init tcp_net_metrics_init(struct net *net) 1092{ 1093 size_t size; 1094 unsigned int slots; 1095 1096 slots = tcpmhash_entries; 1097 if (!slots) { 1098 if (totalram_pages >= 128 * 1024) 1099 slots = 16 * 1024; 1100 else 1101 slots = 8 * 1024; 1102 } 1103 1104 net->ipv4.tcp_metrics_hash_log = order_base_2(slots); 1105 size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log; 1106 1107 net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); 1108 if (!net->ipv4.tcp_metrics_hash) 1109 net->ipv4.tcp_metrics_hash = vzalloc(size); 1110 1111 if (!net->ipv4.tcp_metrics_hash) 1112 return -ENOMEM; 1113 1114 return 0; 1115} 1116 1117static void __net_exit tcp_net_metrics_exit(struct net *net) 1118{ 1119 unsigned int i; 1120 1121 for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) { 1122 struct tcp_metrics_block *tm, *next; 1123 1124 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1); 1125 while (tm) { 1126 next = rcu_dereference_protected(tm->tcpm_next, 1); 1127 kfree(tm); 1128 tm = next; 1129 } 1130 } 1131 if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash)) 1132 vfree(net->ipv4.tcp_metrics_hash); 1133 else 1134 kfree(net->ipv4.tcp_metrics_hash); 1135} 1136 1137static __net_initdata struct pernet_operations tcp_net_metrics_ops = { 1138 .init = tcp_net_metrics_init, 1139 .exit = tcp_net_metrics_exit, 1140}; 1141 1142void __init tcp_metrics_init(void) 1143{ 1144 int ret; 1145 1146 ret = register_pernet_subsys(&tcp_net_metrics_ops); 1147 if (ret < 0) 1148 goto cleanup; 1149 ret = genl_register_family_with_ops(&tcp_metrics_nl_family, 1150 tcp_metrics_nl_ops); 1151 if (ret < 0) 1152 goto cleanup_subsys; 1153 return; 1154 1155cleanup_subsys: 1156 unregister_pernet_subsys(&tcp_net_metrics_ops); 1157 1158cleanup: 1159 return; 1160} 1161