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 23 int sysctl_tcp_nometrics_save __read_mostly; 24 25 static 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 29 struct tcp_fastopen_metrics { 30 u16 mss; 31 u16 syn_loss:10, /* Recurring Fast Open SYN losses */ 32 try_exp:2; /* Request w/ exp. option (once) */ 33 unsigned long last_syn_loss; /* Last Fast Open SYN loss */ 34 struct tcp_fastopen_cookie cookie; 35 }; 36 37 /* TCP_METRIC_MAX includes 2 extra fields for userspace compatibility 38 * Kernel only stores RTT and RTTVAR in usec resolution 39 */ 40 #define TCP_METRIC_MAX_KERNEL (TCP_METRIC_MAX - 2) 41 42 struct tcp_metrics_block { 43 struct tcp_metrics_block __rcu *tcpm_next; 44 possible_net_t tcpm_net; 45 struct inetpeer_addr tcpm_saddr; 46 struct inetpeer_addr tcpm_daddr; 47 unsigned long tcpm_stamp; 48 u32 tcpm_ts; 49 u32 tcpm_ts_stamp; 50 u32 tcpm_lock; 51 u32 tcpm_vals[TCP_METRIC_MAX_KERNEL + 1]; 52 struct tcp_fastopen_metrics tcpm_fastopen; 53 54 struct rcu_head rcu_head; 55 }; 56 57 static inline struct net *tm_net(struct tcp_metrics_block *tm) 58 { 59 return read_pnet(&tm->tcpm_net); 60 } 61 62 static bool tcp_metric_locked(struct tcp_metrics_block *tm, 63 enum tcp_metric_index idx) 64 { 65 return tm->tcpm_lock & (1 << idx); 66 } 67 68 static u32 tcp_metric_get(struct tcp_metrics_block *tm, 69 enum tcp_metric_index idx) 70 { 71 return tm->tcpm_vals[idx]; 72 } 73 74 static void tcp_metric_set(struct tcp_metrics_block *tm, 75 enum tcp_metric_index idx, 76 u32 val) 77 { 78 tm->tcpm_vals[idx] = val; 79 } 80 81 static bool addr_same(const struct inetpeer_addr *a, 82 const struct inetpeer_addr *b) 83 { 84 return inetpeer_addr_cmp(a, b) == 0; 85 } 86 87 struct tcpm_hash_bucket { 88 struct tcp_metrics_block __rcu *chain; 89 }; 90 91 static struct tcpm_hash_bucket *tcp_metrics_hash __read_mostly; 92 static unsigned int tcp_metrics_hash_log __read_mostly; 93 94 static DEFINE_SPINLOCK(tcp_metrics_lock); 95 96 static void tcpm_suck_dst(struct tcp_metrics_block *tm, 97 const struct dst_entry *dst, 98 bool fastopen_clear) 99 { 100 u32 msval; 101 u32 val; 102 103 tm->tcpm_stamp = jiffies; 104 105 val = 0; 106 if (dst_metric_locked(dst, RTAX_RTT)) 107 val |= 1 << TCP_METRIC_RTT; 108 if (dst_metric_locked(dst, RTAX_RTTVAR)) 109 val |= 1 << TCP_METRIC_RTTVAR; 110 if (dst_metric_locked(dst, RTAX_SSTHRESH)) 111 val |= 1 << TCP_METRIC_SSTHRESH; 112 if (dst_metric_locked(dst, RTAX_CWND)) 113 val |= 1 << TCP_METRIC_CWND; 114 if (dst_metric_locked(dst, RTAX_REORDERING)) 115 val |= 1 << TCP_METRIC_REORDERING; 116 tm->tcpm_lock = val; 117 118 msval = dst_metric_raw(dst, RTAX_RTT); 119 tm->tcpm_vals[TCP_METRIC_RTT] = msval * USEC_PER_MSEC; 120 121 msval = dst_metric_raw(dst, RTAX_RTTVAR); 122 tm->tcpm_vals[TCP_METRIC_RTTVAR] = msval * USEC_PER_MSEC; 123 tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH); 124 tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND); 125 tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING); 126 tm->tcpm_ts = 0; 127 tm->tcpm_ts_stamp = 0; 128 if (fastopen_clear) { 129 tm->tcpm_fastopen.mss = 0; 130 tm->tcpm_fastopen.syn_loss = 0; 131 tm->tcpm_fastopen.try_exp = 0; 132 tm->tcpm_fastopen.cookie.exp = false; 133 tm->tcpm_fastopen.cookie.len = 0; 134 } 135 } 136 137 #define TCP_METRICS_TIMEOUT (60 * 60 * HZ) 138 139 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst) 140 { 141 if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT))) 142 tcpm_suck_dst(tm, dst, false); 143 } 144 145 #define TCP_METRICS_RECLAIM_DEPTH 5 146 #define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL 147 148 #define deref_locked(p) \ 149 rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock)) 150 151 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst, 152 struct inetpeer_addr *saddr, 153 struct inetpeer_addr *daddr, 154 unsigned int hash) 155 { 156 struct tcp_metrics_block *tm; 157 struct net *net; 158 bool reclaim = false; 159 160 spin_lock_bh(&tcp_metrics_lock); 161 net = dev_net(dst->dev); 162 163 /* While waiting for the spin-lock the cache might have been populated 164 * with this entry and so we have to check again. 165 */ 166 tm = __tcp_get_metrics(saddr, daddr, net, hash); 167 if (tm == TCP_METRICS_RECLAIM_PTR) { 168 reclaim = true; 169 tm = NULL; 170 } 171 if (tm) { 172 tcpm_check_stamp(tm, dst); 173 goto out_unlock; 174 } 175 176 if (unlikely(reclaim)) { 177 struct tcp_metrics_block *oldest; 178 179 oldest = deref_locked(tcp_metrics_hash[hash].chain); 180 for (tm = deref_locked(oldest->tcpm_next); tm; 181 tm = deref_locked(tm->tcpm_next)) { 182 if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp)) 183 oldest = tm; 184 } 185 tm = oldest; 186 } else { 187 tm = kmalloc(sizeof(*tm), GFP_ATOMIC); 188 if (!tm) 189 goto out_unlock; 190 } 191 write_pnet(&tm->tcpm_net, net); 192 tm->tcpm_saddr = *saddr; 193 tm->tcpm_daddr = *daddr; 194 195 tcpm_suck_dst(tm, dst, true); 196 197 if (likely(!reclaim)) { 198 tm->tcpm_next = tcp_metrics_hash[hash].chain; 199 rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm); 200 } 201 202 out_unlock: 203 spin_unlock_bh(&tcp_metrics_lock); 204 return tm; 205 } 206 207 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth) 208 { 209 if (tm) 210 return tm; 211 if (depth > TCP_METRICS_RECLAIM_DEPTH) 212 return TCP_METRICS_RECLAIM_PTR; 213 return NULL; 214 } 215 216 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr, 217 const struct inetpeer_addr *daddr, 218 struct net *net, unsigned int hash) 219 { 220 struct tcp_metrics_block *tm; 221 int depth = 0; 222 223 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm; 224 tm = rcu_dereference(tm->tcpm_next)) { 225 if (addr_same(&tm->tcpm_saddr, saddr) && 226 addr_same(&tm->tcpm_daddr, daddr) && 227 net_eq(tm_net(tm), net)) 228 break; 229 depth++; 230 } 231 return tcp_get_encode(tm, depth); 232 } 233 234 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req, 235 struct dst_entry *dst) 236 { 237 struct tcp_metrics_block *tm; 238 struct inetpeer_addr saddr, daddr; 239 unsigned int hash; 240 struct net *net; 241 242 saddr.family = req->rsk_ops->family; 243 daddr.family = req->rsk_ops->family; 244 switch (daddr.family) { 245 case AF_INET: 246 inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr); 247 inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr); 248 hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr); 249 break; 250 #if IS_ENABLED(CONFIG_IPV6) 251 case AF_INET6: 252 inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr); 253 inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr); 254 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr); 255 break; 256 #endif 257 default: 258 return NULL; 259 } 260 261 net = dev_net(dst->dev); 262 hash ^= net_hash_mix(net); 263 hash = hash_32(hash, tcp_metrics_hash_log); 264 265 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm; 266 tm = rcu_dereference(tm->tcpm_next)) { 267 if (addr_same(&tm->tcpm_saddr, &saddr) && 268 addr_same(&tm->tcpm_daddr, &daddr) && 269 net_eq(tm_net(tm), net)) 270 break; 271 } 272 tcpm_check_stamp(tm, dst); 273 return tm; 274 } 275 276 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw) 277 { 278 struct tcp_metrics_block *tm; 279 struct inetpeer_addr saddr, daddr; 280 unsigned int hash; 281 struct net *net; 282 283 if (tw->tw_family == AF_INET) { 284 inetpeer_set_addr_v4(&saddr, tw->tw_rcv_saddr); 285 inetpeer_set_addr_v4(&daddr, tw->tw_daddr); 286 hash = ipv4_addr_hash(tw->tw_daddr); 287 } 288 #if IS_ENABLED(CONFIG_IPV6) 289 else if (tw->tw_family == AF_INET6) { 290 if (ipv6_addr_v4mapped(&tw->tw_v6_daddr)) { 291 inetpeer_set_addr_v4(&saddr, tw->tw_rcv_saddr); 292 inetpeer_set_addr_v4(&daddr, tw->tw_daddr); 293 hash = ipv4_addr_hash(tw->tw_daddr); 294 } else { 295 inetpeer_set_addr_v6(&saddr, &tw->tw_v6_rcv_saddr); 296 inetpeer_set_addr_v6(&daddr, &tw->tw_v6_daddr); 297 hash = ipv6_addr_hash(&tw->tw_v6_daddr); 298 } 299 } 300 #endif 301 else 302 return NULL; 303 304 net = twsk_net(tw); 305 hash ^= net_hash_mix(net); 306 hash = hash_32(hash, tcp_metrics_hash_log); 307 308 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm; 309 tm = rcu_dereference(tm->tcpm_next)) { 310 if (addr_same(&tm->tcpm_saddr, &saddr) && 311 addr_same(&tm->tcpm_daddr, &daddr) && 312 net_eq(tm_net(tm), net)) 313 break; 314 } 315 return tm; 316 } 317 318 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk, 319 struct dst_entry *dst, 320 bool create) 321 { 322 struct tcp_metrics_block *tm; 323 struct inetpeer_addr saddr, daddr; 324 unsigned int hash; 325 struct net *net; 326 327 if (sk->sk_family == AF_INET) { 328 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr); 329 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr); 330 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr); 331 } 332 #if IS_ENABLED(CONFIG_IPV6) 333 else if (sk->sk_family == AF_INET6) { 334 if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) { 335 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr); 336 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr); 337 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr); 338 } else { 339 inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr); 340 inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr); 341 hash = ipv6_addr_hash(&sk->sk_v6_daddr); 342 } 343 } 344 #endif 345 else 346 return NULL; 347 348 net = dev_net(dst->dev); 349 hash ^= net_hash_mix(net); 350 hash = hash_32(hash, tcp_metrics_hash_log); 351 352 tm = __tcp_get_metrics(&saddr, &daddr, net, hash); 353 if (tm == TCP_METRICS_RECLAIM_PTR) 354 tm = NULL; 355 if (!tm && create) 356 tm = tcpm_new(dst, &saddr, &daddr, hash); 357 else 358 tcpm_check_stamp(tm, dst); 359 360 return tm; 361 } 362 363 /* Save metrics learned by this TCP session. This function is called 364 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT 365 * or goes from LAST-ACK to CLOSE. 366 */ 367 void tcp_update_metrics(struct sock *sk) 368 { 369 const struct inet_connection_sock *icsk = inet_csk(sk); 370 struct dst_entry *dst = __sk_dst_get(sk); 371 struct tcp_sock *tp = tcp_sk(sk); 372 struct tcp_metrics_block *tm; 373 unsigned long rtt; 374 u32 val; 375 int m; 376 377 if (sysctl_tcp_nometrics_save || !dst) 378 return; 379 380 if (dst->flags & DST_HOST) 381 dst_confirm(dst); 382 383 rcu_read_lock(); 384 if (icsk->icsk_backoff || !tp->srtt_us) { 385 /* This session failed to estimate rtt. Why? 386 * Probably, no packets returned in time. Reset our 387 * results. 388 */ 389 tm = tcp_get_metrics(sk, dst, false); 390 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT)) 391 tcp_metric_set(tm, TCP_METRIC_RTT, 0); 392 goto out_unlock; 393 } else 394 tm = tcp_get_metrics(sk, dst, true); 395 396 if (!tm) 397 goto out_unlock; 398 399 rtt = tcp_metric_get(tm, TCP_METRIC_RTT); 400 m = rtt - tp->srtt_us; 401 402 /* If newly calculated rtt larger than stored one, store new 403 * one. Otherwise, use EWMA. Remember, rtt overestimation is 404 * always better than underestimation. 405 */ 406 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) { 407 if (m <= 0) 408 rtt = tp->srtt_us; 409 else 410 rtt -= (m >> 3); 411 tcp_metric_set(tm, TCP_METRIC_RTT, rtt); 412 } 413 414 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) { 415 unsigned long var; 416 417 if (m < 0) 418 m = -m; 419 420 /* Scale deviation to rttvar fixed point */ 421 m >>= 1; 422 if (m < tp->mdev_us) 423 m = tp->mdev_us; 424 425 var = tcp_metric_get(tm, TCP_METRIC_RTTVAR); 426 if (m >= var) 427 var = m; 428 else 429 var -= (var - m) >> 2; 430 431 tcp_metric_set(tm, TCP_METRIC_RTTVAR, var); 432 } 433 434 if (tcp_in_initial_slowstart(tp)) { 435 /* Slow start still did not finish. */ 436 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) { 437 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); 438 if (val && (tp->snd_cwnd >> 1) > val) 439 tcp_metric_set(tm, TCP_METRIC_SSTHRESH, 440 tp->snd_cwnd >> 1); 441 } 442 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { 443 val = tcp_metric_get(tm, TCP_METRIC_CWND); 444 if (tp->snd_cwnd > val) 445 tcp_metric_set(tm, TCP_METRIC_CWND, 446 tp->snd_cwnd); 447 } 448 } else if (!tcp_in_slow_start(tp) && 449 icsk->icsk_ca_state == TCP_CA_Open) { 450 /* Cong. avoidance phase, cwnd is reliable. */ 451 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) 452 tcp_metric_set(tm, TCP_METRIC_SSTHRESH, 453 max(tp->snd_cwnd >> 1, tp->snd_ssthresh)); 454 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { 455 val = tcp_metric_get(tm, TCP_METRIC_CWND); 456 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1); 457 } 458 } else { 459 /* Else slow start did not finish, cwnd is non-sense, 460 * ssthresh may be also invalid. 461 */ 462 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { 463 val = tcp_metric_get(tm, TCP_METRIC_CWND); 464 tcp_metric_set(tm, TCP_METRIC_CWND, 465 (val + tp->snd_ssthresh) >> 1); 466 } 467 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) { 468 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); 469 if (val && tp->snd_ssthresh > val) 470 tcp_metric_set(tm, TCP_METRIC_SSTHRESH, 471 tp->snd_ssthresh); 472 } 473 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) { 474 val = tcp_metric_get(tm, TCP_METRIC_REORDERING); 475 if (val < tp->reordering && 476 tp->reordering != sysctl_tcp_reordering) 477 tcp_metric_set(tm, TCP_METRIC_REORDERING, 478 tp->reordering); 479 } 480 } 481 tm->tcpm_stamp = jiffies; 482 out_unlock: 483 rcu_read_unlock(); 484 } 485 486 /* Initialize metrics on socket. */ 487 488 void tcp_init_metrics(struct sock *sk) 489 { 490 struct dst_entry *dst = __sk_dst_get(sk); 491 struct tcp_sock *tp = tcp_sk(sk); 492 struct tcp_metrics_block *tm; 493 u32 val, crtt = 0; /* cached RTT scaled by 8 */ 494 495 if (!dst) 496 goto reset; 497 498 dst_confirm(dst); 499 500 rcu_read_lock(); 501 tm = tcp_get_metrics(sk, dst, true); 502 if (!tm) { 503 rcu_read_unlock(); 504 goto reset; 505 } 506 507 if (tcp_metric_locked(tm, TCP_METRIC_CWND)) 508 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND); 509 510 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); 511 if (val) { 512 tp->snd_ssthresh = val; 513 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 514 tp->snd_ssthresh = tp->snd_cwnd_clamp; 515 } else { 516 /* ssthresh may have been reduced unnecessarily during. 517 * 3WHS. Restore it back to its initial default. 518 */ 519 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 520 } 521 val = tcp_metric_get(tm, TCP_METRIC_REORDERING); 522 if (val && tp->reordering != val) { 523 tcp_disable_fack(tp); 524 tcp_disable_early_retrans(tp); 525 tp->reordering = val; 526 } 527 528 crtt = tcp_metric_get(tm, TCP_METRIC_RTT); 529 rcu_read_unlock(); 530 reset: 531 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal 532 * to seed the RTO for later data packets because SYN packets are 533 * small. Use the per-dst cached values to seed the RTO but keep 534 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar). 535 * Later the RTO will be updated immediately upon obtaining the first 536 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only 537 * influences the first RTO but not later RTT estimation. 538 * 539 * But if RTT is not available from the SYN (due to retransmits or 540 * syn cookies) or the cache, force a conservative 3secs timeout. 541 * 542 * A bit of theory. RTT is time passed after "normal" sized packet 543 * is sent until it is ACKed. In normal circumstances sending small 544 * packets force peer to delay ACKs and calculation is correct too. 545 * The algorithm is adaptive and, provided we follow specs, it 546 * NEVER underestimate RTT. BUT! If peer tries to make some clever 547 * tricks sort of "quick acks" for time long enough to decrease RTT 548 * to low value, and then abruptly stops to do it and starts to delay 549 * ACKs, wait for troubles. 550 */ 551 if (crtt > tp->srtt_us) { 552 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */ 553 crtt /= 8 * USEC_PER_MSEC; 554 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk)); 555 } else if (tp->srtt_us == 0) { 556 /* RFC6298: 5.7 We've failed to get a valid RTT sample from 557 * 3WHS. This is most likely due to retransmission, 558 * including spurious one. Reset the RTO back to 3secs 559 * from the more aggressive 1sec to avoid more spurious 560 * retransmission. 561 */ 562 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK); 563 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us; 564 565 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK; 566 } 567 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 568 * retransmitted. In light of RFC6298 more aggressive 1sec 569 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 570 * retransmission has occurred. 571 */ 572 if (tp->total_retrans > 1) 573 tp->snd_cwnd = 1; 574 else 575 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 576 tp->snd_cwnd_stamp = tcp_time_stamp; 577 } 578 579 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, 580 bool paws_check, bool timestamps) 581 { 582 struct tcp_metrics_block *tm; 583 bool ret; 584 585 if (!dst) 586 return false; 587 588 rcu_read_lock(); 589 tm = __tcp_get_metrics_req(req, dst); 590 if (paws_check) { 591 if (tm && 592 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL && 593 ((s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW || 594 !timestamps)) 595 ret = false; 596 else 597 ret = true; 598 } else { 599 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp) 600 ret = true; 601 else 602 ret = false; 603 } 604 rcu_read_unlock(); 605 606 return ret; 607 } 608 EXPORT_SYMBOL_GPL(tcp_peer_is_proven); 609 610 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst) 611 { 612 struct tcp_metrics_block *tm; 613 614 rcu_read_lock(); 615 tm = tcp_get_metrics(sk, dst, true); 616 if (tm) { 617 struct tcp_sock *tp = tcp_sk(sk); 618 619 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) { 620 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp; 621 tp->rx_opt.ts_recent = tm->tcpm_ts; 622 } 623 } 624 rcu_read_unlock(); 625 } 626 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp); 627 628 /* VJ's idea. Save last timestamp seen from this destination and hold 629 * it at least for normal timewait interval to use for duplicate 630 * segment detection in subsequent connections, before they enter 631 * synchronized state. 632 */ 633 bool tcp_remember_stamp(struct sock *sk) 634 { 635 struct dst_entry *dst = __sk_dst_get(sk); 636 bool ret = false; 637 638 if (dst) { 639 struct tcp_metrics_block *tm; 640 641 rcu_read_lock(); 642 tm = tcp_get_metrics(sk, dst, true); 643 if (tm) { 644 struct tcp_sock *tp = tcp_sk(sk); 645 646 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 || 647 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL && 648 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) { 649 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp; 650 tm->tcpm_ts = tp->rx_opt.ts_recent; 651 } 652 ret = true; 653 } 654 rcu_read_unlock(); 655 } 656 return ret; 657 } 658 659 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw) 660 { 661 struct tcp_metrics_block *tm; 662 bool ret = false; 663 664 rcu_read_lock(); 665 tm = __tcp_get_metrics_tw(tw); 666 if (tm) { 667 const struct tcp_timewait_sock *tcptw; 668 struct sock *sk = (struct sock *) tw; 669 670 tcptw = tcp_twsk(sk); 671 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 || 672 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL && 673 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) { 674 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp; 675 tm->tcpm_ts = tcptw->tw_ts_recent; 676 } 677 ret = true; 678 } 679 rcu_read_unlock(); 680 681 return ret; 682 } 683 684 static DEFINE_SEQLOCK(fastopen_seqlock); 685 686 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss, 687 struct tcp_fastopen_cookie *cookie, 688 int *syn_loss, unsigned long *last_syn_loss) 689 { 690 struct tcp_metrics_block *tm; 691 692 rcu_read_lock(); 693 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false); 694 if (tm) { 695 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen; 696 unsigned int seq; 697 698 do { 699 seq = read_seqbegin(&fastopen_seqlock); 700 if (tfom->mss) 701 *mss = tfom->mss; 702 *cookie = tfom->cookie; 703 if (cookie->len <= 0 && tfom->try_exp == 1) 704 cookie->exp = true; 705 *syn_loss = tfom->syn_loss; 706 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0; 707 } while (read_seqretry(&fastopen_seqlock, seq)); 708 } 709 rcu_read_unlock(); 710 } 711 712 void tcp_fastopen_cache_set(struct sock *sk, u16 mss, 713 struct tcp_fastopen_cookie *cookie, bool syn_lost, 714 u16 try_exp) 715 { 716 struct dst_entry *dst = __sk_dst_get(sk); 717 struct tcp_metrics_block *tm; 718 719 if (!dst) 720 return; 721 rcu_read_lock(); 722 tm = tcp_get_metrics(sk, dst, true); 723 if (tm) { 724 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen; 725 726 write_seqlock_bh(&fastopen_seqlock); 727 if (mss) 728 tfom->mss = mss; 729 if (cookie && cookie->len > 0) 730 tfom->cookie = *cookie; 731 else if (try_exp > tfom->try_exp && 732 tfom->cookie.len <= 0 && !tfom->cookie.exp) 733 tfom->try_exp = try_exp; 734 if (syn_lost) { 735 ++tfom->syn_loss; 736 tfom->last_syn_loss = jiffies; 737 } else 738 tfom->syn_loss = 0; 739 write_sequnlock_bh(&fastopen_seqlock); 740 } 741 rcu_read_unlock(); 742 } 743 744 static struct genl_family tcp_metrics_nl_family = { 745 .id = GENL_ID_GENERATE, 746 .hdrsize = 0, 747 .name = TCP_METRICS_GENL_NAME, 748 .version = TCP_METRICS_GENL_VERSION, 749 .maxattr = TCP_METRICS_ATTR_MAX, 750 .netnsok = true, 751 }; 752 753 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = { 754 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, }, 755 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY, 756 .len = sizeof(struct in6_addr), }, 757 /* Following attributes are not received for GET/DEL, 758 * we keep them for reference 759 */ 760 #if 0 761 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, }, 762 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, }, 763 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, }, 764 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, }, 765 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, }, 766 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, }, 767 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, }, 768 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY, 769 .len = TCP_FASTOPEN_COOKIE_MAX, }, 770 #endif 771 }; 772 773 /* Add attributes, caller cancels its header on failure */ 774 static int tcp_metrics_fill_info(struct sk_buff *msg, 775 struct tcp_metrics_block *tm) 776 { 777 struct nlattr *nest; 778 int i; 779 780 switch (tm->tcpm_daddr.family) { 781 case AF_INET: 782 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4, 783 inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0) 784 goto nla_put_failure; 785 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4, 786 inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0) 787 goto nla_put_failure; 788 break; 789 case AF_INET6: 790 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6, 791 inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0) 792 goto nla_put_failure; 793 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6, 794 inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0) 795 goto nla_put_failure; 796 break; 797 default: 798 return -EAFNOSUPPORT; 799 } 800 801 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE, 802 jiffies - tm->tcpm_stamp) < 0) 803 goto nla_put_failure; 804 if (tm->tcpm_ts_stamp) { 805 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP, 806 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0) 807 goto nla_put_failure; 808 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL, 809 tm->tcpm_ts) < 0) 810 goto nla_put_failure; 811 } 812 813 { 814 int n = 0; 815 816 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS); 817 if (!nest) 818 goto nla_put_failure; 819 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) { 820 u32 val = tm->tcpm_vals[i]; 821 822 if (!val) 823 continue; 824 if (i == TCP_METRIC_RTT) { 825 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1, 826 val) < 0) 827 goto nla_put_failure; 828 n++; 829 val = max(val / 1000, 1U); 830 } 831 if (i == TCP_METRIC_RTTVAR) { 832 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1, 833 val) < 0) 834 goto nla_put_failure; 835 n++; 836 val = max(val / 1000, 1U); 837 } 838 if (nla_put_u32(msg, i + 1, val) < 0) 839 goto nla_put_failure; 840 n++; 841 } 842 if (n) 843 nla_nest_end(msg, nest); 844 else 845 nla_nest_cancel(msg, nest); 846 } 847 848 { 849 struct tcp_fastopen_metrics tfom_copy[1], *tfom; 850 unsigned int seq; 851 852 do { 853 seq = read_seqbegin(&fastopen_seqlock); 854 tfom_copy[0] = tm->tcpm_fastopen; 855 } while (read_seqretry(&fastopen_seqlock, seq)); 856 857 tfom = tfom_copy; 858 if (tfom->mss && 859 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS, 860 tfom->mss) < 0) 861 goto nla_put_failure; 862 if (tfom->syn_loss && 863 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS, 864 tfom->syn_loss) < 0 || 865 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS, 866 jiffies - tfom->last_syn_loss) < 0)) 867 goto nla_put_failure; 868 if (tfom->cookie.len > 0 && 869 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE, 870 tfom->cookie.len, tfom->cookie.val) < 0) 871 goto nla_put_failure; 872 } 873 874 return 0; 875 876 nla_put_failure: 877 return -EMSGSIZE; 878 } 879 880 static int tcp_metrics_dump_info(struct sk_buff *skb, 881 struct netlink_callback *cb, 882 struct tcp_metrics_block *tm) 883 { 884 void *hdr; 885 886 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq, 887 &tcp_metrics_nl_family, NLM_F_MULTI, 888 TCP_METRICS_CMD_GET); 889 if (!hdr) 890 return -EMSGSIZE; 891 892 if (tcp_metrics_fill_info(skb, tm) < 0) 893 goto nla_put_failure; 894 895 genlmsg_end(skb, hdr); 896 return 0; 897 898 nla_put_failure: 899 genlmsg_cancel(skb, hdr); 900 return -EMSGSIZE; 901 } 902 903 static int tcp_metrics_nl_dump(struct sk_buff *skb, 904 struct netlink_callback *cb) 905 { 906 struct net *net = sock_net(skb->sk); 907 unsigned int max_rows = 1U << tcp_metrics_hash_log; 908 unsigned int row, s_row = cb->args[0]; 909 int s_col = cb->args[1], col = s_col; 910 911 for (row = s_row; row < max_rows; row++, s_col = 0) { 912 struct tcp_metrics_block *tm; 913 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row; 914 915 rcu_read_lock(); 916 for (col = 0, tm = rcu_dereference(hb->chain); tm; 917 tm = rcu_dereference(tm->tcpm_next), col++) { 918 if (!net_eq(tm_net(tm), net)) 919 continue; 920 if (col < s_col) 921 continue; 922 if (tcp_metrics_dump_info(skb, cb, tm) < 0) { 923 rcu_read_unlock(); 924 goto done; 925 } 926 } 927 rcu_read_unlock(); 928 } 929 930 done: 931 cb->args[0] = row; 932 cb->args[1] = col; 933 return skb->len; 934 } 935 936 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr, 937 unsigned int *hash, int optional, int v4, int v6) 938 { 939 struct nlattr *a; 940 941 a = info->attrs[v4]; 942 if (a) { 943 inetpeer_set_addr_v4(addr, nla_get_in_addr(a)); 944 if (hash) 945 *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr)); 946 return 0; 947 } 948 a = info->attrs[v6]; 949 if (a) { 950 struct in6_addr in6; 951 952 if (nla_len(a) != sizeof(struct in6_addr)) 953 return -EINVAL; 954 in6 = nla_get_in6_addr(a); 955 inetpeer_set_addr_v6(addr, &in6); 956 if (hash) 957 *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr)); 958 return 0; 959 } 960 return optional ? 1 : -EAFNOSUPPORT; 961 } 962 963 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr, 964 unsigned int *hash, int optional) 965 { 966 return __parse_nl_addr(info, addr, hash, optional, 967 TCP_METRICS_ATTR_ADDR_IPV4, 968 TCP_METRICS_ATTR_ADDR_IPV6); 969 } 970 971 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr) 972 { 973 return __parse_nl_addr(info, addr, NULL, 0, 974 TCP_METRICS_ATTR_SADDR_IPV4, 975 TCP_METRICS_ATTR_SADDR_IPV6); 976 } 977 978 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info) 979 { 980 struct tcp_metrics_block *tm; 981 struct inetpeer_addr saddr, daddr; 982 unsigned int hash; 983 struct sk_buff *msg; 984 struct net *net = genl_info_net(info); 985 void *reply; 986 int ret; 987 bool src = true; 988 989 ret = parse_nl_addr(info, &daddr, &hash, 0); 990 if (ret < 0) 991 return ret; 992 993 ret = parse_nl_saddr(info, &saddr); 994 if (ret < 0) 995 src = false; 996 997 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL); 998 if (!msg) 999 return -ENOMEM; 1000 1001 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0, 1002 info->genlhdr->cmd); 1003 if (!reply) 1004 goto nla_put_failure; 1005 1006 hash ^= net_hash_mix(net); 1007 hash = hash_32(hash, tcp_metrics_hash_log); 1008 ret = -ESRCH; 1009 rcu_read_lock(); 1010 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm; 1011 tm = rcu_dereference(tm->tcpm_next)) { 1012 if (addr_same(&tm->tcpm_daddr, &daddr) && 1013 (!src || addr_same(&tm->tcpm_saddr, &saddr)) && 1014 net_eq(tm_net(tm), net)) { 1015 ret = tcp_metrics_fill_info(msg, tm); 1016 break; 1017 } 1018 } 1019 rcu_read_unlock(); 1020 if (ret < 0) 1021 goto out_free; 1022 1023 genlmsg_end(msg, reply); 1024 return genlmsg_reply(msg, info); 1025 1026 nla_put_failure: 1027 ret = -EMSGSIZE; 1028 1029 out_free: 1030 nlmsg_free(msg); 1031 return ret; 1032 } 1033 1034 static void tcp_metrics_flush_all(struct net *net) 1035 { 1036 unsigned int max_rows = 1U << tcp_metrics_hash_log; 1037 struct tcpm_hash_bucket *hb = tcp_metrics_hash; 1038 struct tcp_metrics_block *tm; 1039 unsigned int row; 1040 1041 for (row = 0; row < max_rows; row++, hb++) { 1042 struct tcp_metrics_block __rcu **pp; 1043 spin_lock_bh(&tcp_metrics_lock); 1044 pp = &hb->chain; 1045 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) { 1046 if (net_eq(tm_net(tm), net)) { 1047 *pp = tm->tcpm_next; 1048 kfree_rcu(tm, rcu_head); 1049 } else { 1050 pp = &tm->tcpm_next; 1051 } 1052 } 1053 spin_unlock_bh(&tcp_metrics_lock); 1054 } 1055 } 1056 1057 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info) 1058 { 1059 struct tcpm_hash_bucket *hb; 1060 struct tcp_metrics_block *tm; 1061 struct tcp_metrics_block __rcu **pp; 1062 struct inetpeer_addr saddr, daddr; 1063 unsigned int hash; 1064 struct net *net = genl_info_net(info); 1065 int ret; 1066 bool src = true, found = false; 1067 1068 ret = parse_nl_addr(info, &daddr, &hash, 1); 1069 if (ret < 0) 1070 return ret; 1071 if (ret > 0) { 1072 tcp_metrics_flush_all(net); 1073 return 0; 1074 } 1075 ret = parse_nl_saddr(info, &saddr); 1076 if (ret < 0) 1077 src = false; 1078 1079 hash ^= net_hash_mix(net); 1080 hash = hash_32(hash, tcp_metrics_hash_log); 1081 hb = tcp_metrics_hash + hash; 1082 pp = &hb->chain; 1083 spin_lock_bh(&tcp_metrics_lock); 1084 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) { 1085 if (addr_same(&tm->tcpm_daddr, &daddr) && 1086 (!src || addr_same(&tm->tcpm_saddr, &saddr)) && 1087 net_eq(tm_net(tm), net)) { 1088 *pp = tm->tcpm_next; 1089 kfree_rcu(tm, rcu_head); 1090 found = true; 1091 } else { 1092 pp = &tm->tcpm_next; 1093 } 1094 } 1095 spin_unlock_bh(&tcp_metrics_lock); 1096 if (!found) 1097 return -ESRCH; 1098 return 0; 1099 } 1100 1101 static const struct genl_ops tcp_metrics_nl_ops[] = { 1102 { 1103 .cmd = TCP_METRICS_CMD_GET, 1104 .doit = tcp_metrics_nl_cmd_get, 1105 .dumpit = tcp_metrics_nl_dump, 1106 .policy = tcp_metrics_nl_policy, 1107 }, 1108 { 1109 .cmd = TCP_METRICS_CMD_DEL, 1110 .doit = tcp_metrics_nl_cmd_del, 1111 .policy = tcp_metrics_nl_policy, 1112 .flags = GENL_ADMIN_PERM, 1113 }, 1114 }; 1115 1116 static unsigned int tcpmhash_entries; 1117 static int __init set_tcpmhash_entries(char *str) 1118 { 1119 ssize_t ret; 1120 1121 if (!str) 1122 return 0; 1123 1124 ret = kstrtouint(str, 0, &tcpmhash_entries); 1125 if (ret) 1126 return 0; 1127 1128 return 1; 1129 } 1130 __setup("tcpmhash_entries=", set_tcpmhash_entries); 1131 1132 static int __net_init tcp_net_metrics_init(struct net *net) 1133 { 1134 size_t size; 1135 unsigned int slots; 1136 1137 if (!net_eq(net, &init_net)) 1138 return 0; 1139 1140 slots = tcpmhash_entries; 1141 if (!slots) { 1142 if (totalram_pages >= 128 * 1024) 1143 slots = 16 * 1024; 1144 else 1145 slots = 8 * 1024; 1146 } 1147 1148 tcp_metrics_hash_log = order_base_2(slots); 1149 size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log; 1150 1151 tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); 1152 if (!tcp_metrics_hash) 1153 tcp_metrics_hash = vzalloc(size); 1154 1155 if (!tcp_metrics_hash) 1156 return -ENOMEM; 1157 1158 return 0; 1159 } 1160 1161 static void __net_exit tcp_net_metrics_exit(struct net *net) 1162 { 1163 tcp_metrics_flush_all(net); 1164 } 1165 1166 static __net_initdata struct pernet_operations tcp_net_metrics_ops = { 1167 .init = tcp_net_metrics_init, 1168 .exit = tcp_net_metrics_exit, 1169 }; 1170 1171 void __init tcp_metrics_init(void) 1172 { 1173 int ret; 1174 1175 ret = register_pernet_subsys(&tcp_net_metrics_ops); 1176 if (ret < 0) 1177 panic("Could not allocate the tcp_metrics hash table\n"); 1178 1179 ret = genl_register_family_with_ops(&tcp_metrics_nl_family, 1180 tcp_metrics_nl_ops); 1181 if (ret < 0) 1182 panic("Could not register tcp_metrics generic netlink\n"); 1183 } 1184