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 net *net = sock_net(sk); 373 struct tcp_metrics_block *tm; 374 unsigned long rtt; 375 u32 val; 376 int m; 377 378 sk_dst_confirm(sk); 379 if (sysctl_tcp_nometrics_save || !dst) 380 return; 381 382 rcu_read_lock(); 383 if (icsk->icsk_backoff || !tp->srtt_us) { 384 /* This session failed to estimate rtt. Why? 385 * Probably, no packets returned in time. Reset our 386 * results. 387 */ 388 tm = tcp_get_metrics(sk, dst, false); 389 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT)) 390 tcp_metric_set(tm, TCP_METRIC_RTT, 0); 391 goto out_unlock; 392 } else 393 tm = tcp_get_metrics(sk, dst, true); 394 395 if (!tm) 396 goto out_unlock; 397 398 rtt = tcp_metric_get(tm, TCP_METRIC_RTT); 399 m = rtt - tp->srtt_us; 400 401 /* If newly calculated rtt larger than stored one, store new 402 * one. Otherwise, use EWMA. Remember, rtt overestimation is 403 * always better than underestimation. 404 */ 405 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) { 406 if (m <= 0) 407 rtt = tp->srtt_us; 408 else 409 rtt -= (m >> 3); 410 tcp_metric_set(tm, TCP_METRIC_RTT, rtt); 411 } 412 413 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) { 414 unsigned long var; 415 416 if (m < 0) 417 m = -m; 418 419 /* Scale deviation to rttvar fixed point */ 420 m >>= 1; 421 if (m < tp->mdev_us) 422 m = tp->mdev_us; 423 424 var = tcp_metric_get(tm, TCP_METRIC_RTTVAR); 425 if (m >= var) 426 var = m; 427 else 428 var -= (var - m) >> 2; 429 430 tcp_metric_set(tm, TCP_METRIC_RTTVAR, var); 431 } 432 433 if (tcp_in_initial_slowstart(tp)) { 434 /* Slow start still did not finish. */ 435 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) { 436 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); 437 if (val && (tp->snd_cwnd >> 1) > val) 438 tcp_metric_set(tm, TCP_METRIC_SSTHRESH, 439 tp->snd_cwnd >> 1); 440 } 441 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { 442 val = tcp_metric_get(tm, TCP_METRIC_CWND); 443 if (tp->snd_cwnd > val) 444 tcp_metric_set(tm, TCP_METRIC_CWND, 445 tp->snd_cwnd); 446 } 447 } else if (!tcp_in_slow_start(tp) && 448 icsk->icsk_ca_state == TCP_CA_Open) { 449 /* Cong. avoidance phase, cwnd is reliable. */ 450 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) 451 tcp_metric_set(tm, TCP_METRIC_SSTHRESH, 452 max(tp->snd_cwnd >> 1, tp->snd_ssthresh)); 453 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { 454 val = tcp_metric_get(tm, TCP_METRIC_CWND); 455 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1); 456 } 457 } else { 458 /* Else slow start did not finish, cwnd is non-sense, 459 * ssthresh may be also invalid. 460 */ 461 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { 462 val = tcp_metric_get(tm, TCP_METRIC_CWND); 463 tcp_metric_set(tm, TCP_METRIC_CWND, 464 (val + tp->snd_ssthresh) >> 1); 465 } 466 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) { 467 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); 468 if (val && tp->snd_ssthresh > val) 469 tcp_metric_set(tm, TCP_METRIC_SSTHRESH, 470 tp->snd_ssthresh); 471 } 472 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) { 473 val = tcp_metric_get(tm, TCP_METRIC_REORDERING); 474 if (val < tp->reordering && 475 tp->reordering != net->ipv4.sysctl_tcp_reordering) 476 tcp_metric_set(tm, TCP_METRIC_REORDERING, 477 tp->reordering); 478 } 479 } 480 tm->tcpm_stamp = jiffies; 481 out_unlock: 482 rcu_read_unlock(); 483 } 484 485 /* Initialize metrics on socket. */ 486 487 void tcp_init_metrics(struct sock *sk) 488 { 489 struct dst_entry *dst = __sk_dst_get(sk); 490 struct tcp_sock *tp = tcp_sk(sk); 491 struct tcp_metrics_block *tm; 492 u32 val, crtt = 0; /* cached RTT scaled by 8 */ 493 494 sk_dst_confirm(sk); 495 if (!dst) 496 goto reset; 497 498 rcu_read_lock(); 499 tm = tcp_get_metrics(sk, dst, true); 500 if (!tm) { 501 rcu_read_unlock(); 502 goto reset; 503 } 504 505 if (tcp_metric_locked(tm, TCP_METRIC_CWND)) 506 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND); 507 508 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); 509 if (val) { 510 tp->snd_ssthresh = val; 511 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 512 tp->snd_ssthresh = tp->snd_cwnd_clamp; 513 } else { 514 /* ssthresh may have been reduced unnecessarily during. 515 * 3WHS. Restore it back to its initial default. 516 */ 517 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 518 } 519 val = tcp_metric_get(tm, TCP_METRIC_REORDERING); 520 if (val && tp->reordering != val) { 521 tcp_disable_fack(tp); 522 tp->reordering = val; 523 } 524 525 crtt = tcp_metric_get(tm, TCP_METRIC_RTT); 526 rcu_read_unlock(); 527 reset: 528 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal 529 * to seed the RTO for later data packets because SYN packets are 530 * small. Use the per-dst cached values to seed the RTO but keep 531 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar). 532 * Later the RTO will be updated immediately upon obtaining the first 533 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only 534 * influences the first RTO but not later RTT estimation. 535 * 536 * But if RTT is not available from the SYN (due to retransmits or 537 * syn cookies) or the cache, force a conservative 3secs timeout. 538 * 539 * A bit of theory. RTT is time passed after "normal" sized packet 540 * is sent until it is ACKed. In normal circumstances sending small 541 * packets force peer to delay ACKs and calculation is correct too. 542 * The algorithm is adaptive and, provided we follow specs, it 543 * NEVER underestimate RTT. BUT! If peer tries to make some clever 544 * tricks sort of "quick acks" for time long enough to decrease RTT 545 * to low value, and then abruptly stops to do it and starts to delay 546 * ACKs, wait for troubles. 547 */ 548 if (crtt > tp->srtt_us) { 549 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */ 550 crtt /= 8 * USEC_PER_SEC / HZ; 551 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk)); 552 } else if (tp->srtt_us == 0) { 553 /* RFC6298: 5.7 We've failed to get a valid RTT sample from 554 * 3WHS. This is most likely due to retransmission, 555 * including spurious one. Reset the RTO back to 3secs 556 * from the more aggressive 1sec to avoid more spurious 557 * retransmission. 558 */ 559 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK); 560 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us; 561 562 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK; 563 } 564 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 565 * retransmitted. In light of RFC6298 more aggressive 1sec 566 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 567 * retransmission has occurred. 568 */ 569 if (tp->total_retrans > 1) 570 tp->snd_cwnd = 1; 571 else 572 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 573 tp->snd_cwnd_stamp = tcp_time_stamp; 574 } 575 576 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, 577 bool paws_check, bool timestamps) 578 { 579 struct tcp_metrics_block *tm; 580 bool ret; 581 582 if (!dst) 583 return false; 584 585 rcu_read_lock(); 586 tm = __tcp_get_metrics_req(req, dst); 587 if (paws_check) { 588 if (tm && 589 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL && 590 ((s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW || 591 !timestamps)) 592 ret = false; 593 else 594 ret = true; 595 } else { 596 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp) 597 ret = true; 598 else 599 ret = false; 600 } 601 rcu_read_unlock(); 602 603 return ret; 604 } 605 606 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst) 607 { 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 ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) { 616 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp; 617 tp->rx_opt.ts_recent = tm->tcpm_ts; 618 } 619 } 620 rcu_read_unlock(); 621 } 622 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp); 623 624 /* VJ's idea. Save last timestamp seen from this destination and hold 625 * it at least for normal timewait interval to use for duplicate 626 * segment detection in subsequent connections, before they enter 627 * synchronized state. 628 */ 629 bool tcp_remember_stamp(struct sock *sk) 630 { 631 struct dst_entry *dst = __sk_dst_get(sk); 632 bool ret = false; 633 634 if (dst) { 635 struct tcp_metrics_block *tm; 636 637 rcu_read_lock(); 638 tm = tcp_get_metrics(sk, dst, true); 639 if (tm) { 640 struct tcp_sock *tp = tcp_sk(sk); 641 642 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 || 643 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL && 644 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) { 645 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp; 646 tm->tcpm_ts = tp->rx_opt.ts_recent; 647 } 648 ret = true; 649 } 650 rcu_read_unlock(); 651 } 652 return ret; 653 } 654 655 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw) 656 { 657 struct tcp_metrics_block *tm; 658 bool ret = false; 659 660 rcu_read_lock(); 661 tm = __tcp_get_metrics_tw(tw); 662 if (tm) { 663 const struct tcp_timewait_sock *tcptw; 664 struct sock *sk = (struct sock *) tw; 665 666 tcptw = tcp_twsk(sk); 667 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 || 668 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL && 669 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) { 670 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp; 671 tm->tcpm_ts = tcptw->tw_ts_recent; 672 } 673 ret = true; 674 } 675 rcu_read_unlock(); 676 677 return ret; 678 } 679 680 static DEFINE_SEQLOCK(fastopen_seqlock); 681 682 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss, 683 struct tcp_fastopen_cookie *cookie, 684 int *syn_loss, unsigned long *last_syn_loss) 685 { 686 struct tcp_metrics_block *tm; 687 688 rcu_read_lock(); 689 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false); 690 if (tm) { 691 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen; 692 unsigned int seq; 693 694 do { 695 seq = read_seqbegin(&fastopen_seqlock); 696 if (tfom->mss) 697 *mss = tfom->mss; 698 *cookie = tfom->cookie; 699 if (cookie->len <= 0 && tfom->try_exp == 1) 700 cookie->exp = true; 701 *syn_loss = tfom->syn_loss; 702 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0; 703 } while (read_seqretry(&fastopen_seqlock, seq)); 704 } 705 rcu_read_unlock(); 706 } 707 708 void tcp_fastopen_cache_set(struct sock *sk, u16 mss, 709 struct tcp_fastopen_cookie *cookie, bool syn_lost, 710 u16 try_exp) 711 { 712 struct dst_entry *dst = __sk_dst_get(sk); 713 struct tcp_metrics_block *tm; 714 715 if (!dst) 716 return; 717 rcu_read_lock(); 718 tm = tcp_get_metrics(sk, dst, true); 719 if (tm) { 720 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen; 721 722 write_seqlock_bh(&fastopen_seqlock); 723 if (mss) 724 tfom->mss = mss; 725 if (cookie && cookie->len > 0) 726 tfom->cookie = *cookie; 727 else if (try_exp > tfom->try_exp && 728 tfom->cookie.len <= 0 && !tfom->cookie.exp) 729 tfom->try_exp = try_exp; 730 if (syn_lost) { 731 ++tfom->syn_loss; 732 tfom->last_syn_loss = jiffies; 733 } else 734 tfom->syn_loss = 0; 735 write_sequnlock_bh(&fastopen_seqlock); 736 } 737 rcu_read_unlock(); 738 } 739 740 static struct genl_family tcp_metrics_nl_family; 741 742 static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = { 743 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, }, 744 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY, 745 .len = sizeof(struct in6_addr), }, 746 /* Following attributes are not received for GET/DEL, 747 * we keep them for reference 748 */ 749 #if 0 750 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, }, 751 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, }, 752 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, }, 753 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, }, 754 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, }, 755 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, }, 756 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, }, 757 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY, 758 .len = TCP_FASTOPEN_COOKIE_MAX, }, 759 #endif 760 }; 761 762 /* Add attributes, caller cancels its header on failure */ 763 static int tcp_metrics_fill_info(struct sk_buff *msg, 764 struct tcp_metrics_block *tm) 765 { 766 struct nlattr *nest; 767 int i; 768 769 switch (tm->tcpm_daddr.family) { 770 case AF_INET: 771 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4, 772 inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0) 773 goto nla_put_failure; 774 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4, 775 inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0) 776 goto nla_put_failure; 777 break; 778 case AF_INET6: 779 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6, 780 inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0) 781 goto nla_put_failure; 782 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6, 783 inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0) 784 goto nla_put_failure; 785 break; 786 default: 787 return -EAFNOSUPPORT; 788 } 789 790 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE, 791 jiffies - tm->tcpm_stamp, 792 TCP_METRICS_ATTR_PAD) < 0) 793 goto nla_put_failure; 794 if (tm->tcpm_ts_stamp) { 795 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP, 796 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0) 797 goto nla_put_failure; 798 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL, 799 tm->tcpm_ts) < 0) 800 goto nla_put_failure; 801 } 802 803 { 804 int n = 0; 805 806 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS); 807 if (!nest) 808 goto nla_put_failure; 809 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) { 810 u32 val = tm->tcpm_vals[i]; 811 812 if (!val) 813 continue; 814 if (i == TCP_METRIC_RTT) { 815 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1, 816 val) < 0) 817 goto nla_put_failure; 818 n++; 819 val = max(val / 1000, 1U); 820 } 821 if (i == TCP_METRIC_RTTVAR) { 822 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1, 823 val) < 0) 824 goto nla_put_failure; 825 n++; 826 val = max(val / 1000, 1U); 827 } 828 if (nla_put_u32(msg, i + 1, val) < 0) 829 goto nla_put_failure; 830 n++; 831 } 832 if (n) 833 nla_nest_end(msg, nest); 834 else 835 nla_nest_cancel(msg, nest); 836 } 837 838 { 839 struct tcp_fastopen_metrics tfom_copy[1], *tfom; 840 unsigned int seq; 841 842 do { 843 seq = read_seqbegin(&fastopen_seqlock); 844 tfom_copy[0] = tm->tcpm_fastopen; 845 } while (read_seqretry(&fastopen_seqlock, seq)); 846 847 tfom = tfom_copy; 848 if (tfom->mss && 849 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS, 850 tfom->mss) < 0) 851 goto nla_put_failure; 852 if (tfom->syn_loss && 853 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS, 854 tfom->syn_loss) < 0 || 855 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS, 856 jiffies - tfom->last_syn_loss, 857 TCP_METRICS_ATTR_PAD) < 0)) 858 goto nla_put_failure; 859 if (tfom->cookie.len > 0 && 860 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE, 861 tfom->cookie.len, tfom->cookie.val) < 0) 862 goto nla_put_failure; 863 } 864 865 return 0; 866 867 nla_put_failure: 868 return -EMSGSIZE; 869 } 870 871 static int tcp_metrics_dump_info(struct sk_buff *skb, 872 struct netlink_callback *cb, 873 struct tcp_metrics_block *tm) 874 { 875 void *hdr; 876 877 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq, 878 &tcp_metrics_nl_family, NLM_F_MULTI, 879 TCP_METRICS_CMD_GET); 880 if (!hdr) 881 return -EMSGSIZE; 882 883 if (tcp_metrics_fill_info(skb, tm) < 0) 884 goto nla_put_failure; 885 886 genlmsg_end(skb, hdr); 887 return 0; 888 889 nla_put_failure: 890 genlmsg_cancel(skb, hdr); 891 return -EMSGSIZE; 892 } 893 894 static int tcp_metrics_nl_dump(struct sk_buff *skb, 895 struct netlink_callback *cb) 896 { 897 struct net *net = sock_net(skb->sk); 898 unsigned int max_rows = 1U << tcp_metrics_hash_log; 899 unsigned int row, s_row = cb->args[0]; 900 int s_col = cb->args[1], col = s_col; 901 902 for (row = s_row; row < max_rows; row++, s_col = 0) { 903 struct tcp_metrics_block *tm; 904 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row; 905 906 rcu_read_lock(); 907 for (col = 0, tm = rcu_dereference(hb->chain); tm; 908 tm = rcu_dereference(tm->tcpm_next), col++) { 909 if (!net_eq(tm_net(tm), net)) 910 continue; 911 if (col < s_col) 912 continue; 913 if (tcp_metrics_dump_info(skb, cb, tm) < 0) { 914 rcu_read_unlock(); 915 goto done; 916 } 917 } 918 rcu_read_unlock(); 919 } 920 921 done: 922 cb->args[0] = row; 923 cb->args[1] = col; 924 return skb->len; 925 } 926 927 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr, 928 unsigned int *hash, int optional, int v4, int v6) 929 { 930 struct nlattr *a; 931 932 a = info->attrs[v4]; 933 if (a) { 934 inetpeer_set_addr_v4(addr, nla_get_in_addr(a)); 935 if (hash) 936 *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr)); 937 return 0; 938 } 939 a = info->attrs[v6]; 940 if (a) { 941 struct in6_addr in6; 942 943 if (nla_len(a) != sizeof(struct in6_addr)) 944 return -EINVAL; 945 in6 = nla_get_in6_addr(a); 946 inetpeer_set_addr_v6(addr, &in6); 947 if (hash) 948 *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr)); 949 return 0; 950 } 951 return optional ? 1 : -EAFNOSUPPORT; 952 } 953 954 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr, 955 unsigned int *hash, int optional) 956 { 957 return __parse_nl_addr(info, addr, hash, optional, 958 TCP_METRICS_ATTR_ADDR_IPV4, 959 TCP_METRICS_ATTR_ADDR_IPV6); 960 } 961 962 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr) 963 { 964 return __parse_nl_addr(info, addr, NULL, 0, 965 TCP_METRICS_ATTR_SADDR_IPV4, 966 TCP_METRICS_ATTR_SADDR_IPV6); 967 } 968 969 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info) 970 { 971 struct tcp_metrics_block *tm; 972 struct inetpeer_addr saddr, daddr; 973 unsigned int hash; 974 struct sk_buff *msg; 975 struct net *net = genl_info_net(info); 976 void *reply; 977 int ret; 978 bool src = true; 979 980 ret = parse_nl_addr(info, &daddr, &hash, 0); 981 if (ret < 0) 982 return ret; 983 984 ret = parse_nl_saddr(info, &saddr); 985 if (ret < 0) 986 src = false; 987 988 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL); 989 if (!msg) 990 return -ENOMEM; 991 992 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0, 993 info->genlhdr->cmd); 994 if (!reply) 995 goto nla_put_failure; 996 997 hash ^= net_hash_mix(net); 998 hash = hash_32(hash, tcp_metrics_hash_log); 999 ret = -ESRCH; 1000 rcu_read_lock(); 1001 for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm; 1002 tm = rcu_dereference(tm->tcpm_next)) { 1003 if (addr_same(&tm->tcpm_daddr, &daddr) && 1004 (!src || addr_same(&tm->tcpm_saddr, &saddr)) && 1005 net_eq(tm_net(tm), net)) { 1006 ret = tcp_metrics_fill_info(msg, tm); 1007 break; 1008 } 1009 } 1010 rcu_read_unlock(); 1011 if (ret < 0) 1012 goto out_free; 1013 1014 genlmsg_end(msg, reply); 1015 return genlmsg_reply(msg, info); 1016 1017 nla_put_failure: 1018 ret = -EMSGSIZE; 1019 1020 out_free: 1021 nlmsg_free(msg); 1022 return ret; 1023 } 1024 1025 static void tcp_metrics_flush_all(struct net *net) 1026 { 1027 unsigned int max_rows = 1U << tcp_metrics_hash_log; 1028 struct tcpm_hash_bucket *hb = tcp_metrics_hash; 1029 struct tcp_metrics_block *tm; 1030 unsigned int row; 1031 1032 for (row = 0; row < max_rows; row++, hb++) { 1033 struct tcp_metrics_block __rcu **pp; 1034 spin_lock_bh(&tcp_metrics_lock); 1035 pp = &hb->chain; 1036 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) { 1037 if (net_eq(tm_net(tm), net)) { 1038 *pp = tm->tcpm_next; 1039 kfree_rcu(tm, rcu_head); 1040 } else { 1041 pp = &tm->tcpm_next; 1042 } 1043 } 1044 spin_unlock_bh(&tcp_metrics_lock); 1045 } 1046 } 1047 1048 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info) 1049 { 1050 struct tcpm_hash_bucket *hb; 1051 struct tcp_metrics_block *tm; 1052 struct tcp_metrics_block __rcu **pp; 1053 struct inetpeer_addr saddr, daddr; 1054 unsigned int hash; 1055 struct net *net = genl_info_net(info); 1056 int ret; 1057 bool src = true, found = false; 1058 1059 ret = parse_nl_addr(info, &daddr, &hash, 1); 1060 if (ret < 0) 1061 return ret; 1062 if (ret > 0) { 1063 tcp_metrics_flush_all(net); 1064 return 0; 1065 } 1066 ret = parse_nl_saddr(info, &saddr); 1067 if (ret < 0) 1068 src = false; 1069 1070 hash ^= net_hash_mix(net); 1071 hash = hash_32(hash, tcp_metrics_hash_log); 1072 hb = tcp_metrics_hash + hash; 1073 pp = &hb->chain; 1074 spin_lock_bh(&tcp_metrics_lock); 1075 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) { 1076 if (addr_same(&tm->tcpm_daddr, &daddr) && 1077 (!src || addr_same(&tm->tcpm_saddr, &saddr)) && 1078 net_eq(tm_net(tm), net)) { 1079 *pp = tm->tcpm_next; 1080 kfree_rcu(tm, rcu_head); 1081 found = true; 1082 } else { 1083 pp = &tm->tcpm_next; 1084 } 1085 } 1086 spin_unlock_bh(&tcp_metrics_lock); 1087 if (!found) 1088 return -ESRCH; 1089 return 0; 1090 } 1091 1092 static const struct genl_ops tcp_metrics_nl_ops[] = { 1093 { 1094 .cmd = TCP_METRICS_CMD_GET, 1095 .doit = tcp_metrics_nl_cmd_get, 1096 .dumpit = tcp_metrics_nl_dump, 1097 .policy = tcp_metrics_nl_policy, 1098 }, 1099 { 1100 .cmd = TCP_METRICS_CMD_DEL, 1101 .doit = tcp_metrics_nl_cmd_del, 1102 .policy = tcp_metrics_nl_policy, 1103 .flags = GENL_ADMIN_PERM, 1104 }, 1105 }; 1106 1107 static struct genl_family tcp_metrics_nl_family __ro_after_init = { 1108 .hdrsize = 0, 1109 .name = TCP_METRICS_GENL_NAME, 1110 .version = TCP_METRICS_GENL_VERSION, 1111 .maxattr = TCP_METRICS_ATTR_MAX, 1112 .netnsok = true, 1113 .module = THIS_MODULE, 1114 .ops = tcp_metrics_nl_ops, 1115 .n_ops = ARRAY_SIZE(tcp_metrics_nl_ops), 1116 }; 1117 1118 static unsigned int tcpmhash_entries; 1119 static int __init set_tcpmhash_entries(char *str) 1120 { 1121 ssize_t ret; 1122 1123 if (!str) 1124 return 0; 1125 1126 ret = kstrtouint(str, 0, &tcpmhash_entries); 1127 if (ret) 1128 return 0; 1129 1130 return 1; 1131 } 1132 __setup("tcpmhash_entries=", set_tcpmhash_entries); 1133 1134 static int __net_init tcp_net_metrics_init(struct net *net) 1135 { 1136 size_t size; 1137 unsigned int slots; 1138 1139 if (!net_eq(net, &init_net)) 1140 return 0; 1141 1142 slots = tcpmhash_entries; 1143 if (!slots) { 1144 if (totalram_pages >= 128 * 1024) 1145 slots = 16 * 1024; 1146 else 1147 slots = 8 * 1024; 1148 } 1149 1150 tcp_metrics_hash_log = order_base_2(slots); 1151 size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log; 1152 1153 tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); 1154 if (!tcp_metrics_hash) 1155 tcp_metrics_hash = vzalloc(size); 1156 1157 if (!tcp_metrics_hash) 1158 return -ENOMEM; 1159 1160 return 0; 1161 } 1162 1163 static void __net_exit tcp_net_metrics_exit(struct net *net) 1164 { 1165 tcp_metrics_flush_all(net); 1166 } 1167 1168 static __net_initdata struct pernet_operations tcp_net_metrics_ops = { 1169 .init = tcp_net_metrics_init, 1170 .exit = tcp_net_metrics_exit, 1171 }; 1172 1173 void __init tcp_metrics_init(void) 1174 { 1175 int ret; 1176 1177 ret = register_pernet_subsys(&tcp_net_metrics_ops); 1178 if (ret < 0) 1179 panic("Could not allocate the tcp_metrics hash table\n"); 1180 1181 ret = genl_register_family(&tcp_metrics_nl_family); 1182 if (ret < 0) 1183 panic("Could not register tcp_metrics generic netlink\n"); 1184 } 1185