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 struct tcp_fastopen_metrics { 26 u16 mss; 27 u16 syn_loss:10; /* Recurring Fast Open SYN losses */ 28 unsigned long last_syn_loss; /* Last Fast Open SYN loss */ 29 struct tcp_fastopen_cookie cookie; 30 }; 31 32 struct tcp_metrics_block { 33 struct tcp_metrics_block __rcu *tcpm_next; 34 struct inetpeer_addr tcpm_addr; 35 unsigned long tcpm_stamp; 36 u32 tcpm_ts; 37 u32 tcpm_ts_stamp; 38 u32 tcpm_lock; 39 u32 tcpm_vals[TCP_METRIC_MAX + 1]; 40 struct tcp_fastopen_metrics tcpm_fastopen; 41 42 struct rcu_head rcu_head; 43 }; 44 45 static bool tcp_metric_locked(struct tcp_metrics_block *tm, 46 enum tcp_metric_index idx) 47 { 48 return tm->tcpm_lock & (1 << idx); 49 } 50 51 static u32 tcp_metric_get(struct tcp_metrics_block *tm, 52 enum tcp_metric_index idx) 53 { 54 return tm->tcpm_vals[idx]; 55 } 56 57 static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm, 58 enum tcp_metric_index idx) 59 { 60 return msecs_to_jiffies(tm->tcpm_vals[idx]); 61 } 62 63 static void tcp_metric_set(struct tcp_metrics_block *tm, 64 enum tcp_metric_index idx, 65 u32 val) 66 { 67 tm->tcpm_vals[idx] = val; 68 } 69 70 static void tcp_metric_set_msecs(struct tcp_metrics_block *tm, 71 enum tcp_metric_index idx, 72 u32 val) 73 { 74 tm->tcpm_vals[idx] = jiffies_to_msecs(val); 75 } 76 77 static bool addr_same(const struct inetpeer_addr *a, 78 const struct inetpeer_addr *b) 79 { 80 const struct in6_addr *a6, *b6; 81 82 if (a->family != b->family) 83 return false; 84 if (a->family == AF_INET) 85 return a->addr.a4 == b->addr.a4; 86 87 a6 = (const struct in6_addr *) &a->addr.a6[0]; 88 b6 = (const struct in6_addr *) &b->addr.a6[0]; 89 90 return ipv6_addr_equal(a6, b6); 91 } 92 93 struct tcpm_hash_bucket { 94 struct tcp_metrics_block __rcu *chain; 95 }; 96 97 static DEFINE_SPINLOCK(tcp_metrics_lock); 98 99 static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst, 100 bool fastopen_clear) 101 { 102 u32 val; 103 104 tm->tcpm_stamp = jiffies; 105 106 val = 0; 107 if (dst_metric_locked(dst, RTAX_RTT)) 108 val |= 1 << TCP_METRIC_RTT; 109 if (dst_metric_locked(dst, RTAX_RTTVAR)) 110 val |= 1 << TCP_METRIC_RTTVAR; 111 if (dst_metric_locked(dst, RTAX_SSTHRESH)) 112 val |= 1 << TCP_METRIC_SSTHRESH; 113 if (dst_metric_locked(dst, RTAX_CWND)) 114 val |= 1 << TCP_METRIC_CWND; 115 if (dst_metric_locked(dst, RTAX_REORDERING)) 116 val |= 1 << TCP_METRIC_REORDERING; 117 tm->tcpm_lock = val; 118 119 tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT); 120 tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR); 121 tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH); 122 tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND); 123 tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING); 124 tm->tcpm_ts = 0; 125 tm->tcpm_ts_stamp = 0; 126 if (fastopen_clear) { 127 tm->tcpm_fastopen.mss = 0; 128 tm->tcpm_fastopen.syn_loss = 0; 129 tm->tcpm_fastopen.cookie.len = 0; 130 } 131 } 132 133 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst, 134 struct inetpeer_addr *addr, 135 unsigned int hash, 136 bool reclaim) 137 { 138 struct tcp_metrics_block *tm; 139 struct net *net; 140 141 spin_lock_bh(&tcp_metrics_lock); 142 net = dev_net(dst->dev); 143 if (unlikely(reclaim)) { 144 struct tcp_metrics_block *oldest; 145 146 oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); 147 for (tm = rcu_dereference(oldest->tcpm_next); tm; 148 tm = rcu_dereference(tm->tcpm_next)) { 149 if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp)) 150 oldest = tm; 151 } 152 tm = oldest; 153 } else { 154 tm = kmalloc(sizeof(*tm), GFP_ATOMIC); 155 if (!tm) 156 goto out_unlock; 157 } 158 tm->tcpm_addr = *addr; 159 160 tcpm_suck_dst(tm, dst, true); 161 162 if (likely(!reclaim)) { 163 tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain; 164 rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm); 165 } 166 167 out_unlock: 168 spin_unlock_bh(&tcp_metrics_lock); 169 return tm; 170 } 171 172 #define TCP_METRICS_TIMEOUT (60 * 60 * HZ) 173 174 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst) 175 { 176 if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT))) 177 tcpm_suck_dst(tm, dst, false); 178 } 179 180 #define TCP_METRICS_RECLAIM_DEPTH 5 181 #define TCP_METRICS_RECLAIM_PTR (struct tcp_metrics_block *) 0x1UL 182 183 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth) 184 { 185 if (tm) 186 return tm; 187 if (depth > TCP_METRICS_RECLAIM_DEPTH) 188 return TCP_METRICS_RECLAIM_PTR; 189 return NULL; 190 } 191 192 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *addr, 193 struct net *net, unsigned int hash) 194 { 195 struct tcp_metrics_block *tm; 196 int depth = 0; 197 198 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; 199 tm = rcu_dereference(tm->tcpm_next)) { 200 if (addr_same(&tm->tcpm_addr, addr)) 201 break; 202 depth++; 203 } 204 return tcp_get_encode(tm, depth); 205 } 206 207 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req, 208 struct dst_entry *dst) 209 { 210 struct tcp_metrics_block *tm; 211 struct inetpeer_addr addr; 212 unsigned int hash; 213 struct net *net; 214 215 addr.family = req->rsk_ops->family; 216 switch (addr.family) { 217 case AF_INET: 218 addr.addr.a4 = inet_rsk(req)->rmt_addr; 219 hash = (__force unsigned int) addr.addr.a4; 220 break; 221 case AF_INET6: 222 *(struct in6_addr *)addr.addr.a6 = inet6_rsk(req)->rmt_addr; 223 hash = ipv6_addr_hash(&inet6_rsk(req)->rmt_addr); 224 break; 225 default: 226 return NULL; 227 } 228 229 net = dev_net(dst->dev); 230 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); 231 232 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; 233 tm = rcu_dereference(tm->tcpm_next)) { 234 if (addr_same(&tm->tcpm_addr, &addr)) 235 break; 236 } 237 tcpm_check_stamp(tm, dst); 238 return tm; 239 } 240 241 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw) 242 { 243 struct inet6_timewait_sock *tw6; 244 struct tcp_metrics_block *tm; 245 struct inetpeer_addr addr; 246 unsigned int hash; 247 struct net *net; 248 249 addr.family = tw->tw_family; 250 switch (addr.family) { 251 case AF_INET: 252 addr.addr.a4 = tw->tw_daddr; 253 hash = (__force unsigned int) addr.addr.a4; 254 break; 255 case AF_INET6: 256 tw6 = inet6_twsk((struct sock *)tw); 257 *(struct in6_addr *)addr.addr.a6 = tw6->tw_v6_daddr; 258 hash = ipv6_addr_hash(&tw6->tw_v6_daddr); 259 break; 260 default: 261 return NULL; 262 } 263 264 net = twsk_net(tw); 265 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); 266 267 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; 268 tm = rcu_dereference(tm->tcpm_next)) { 269 if (addr_same(&tm->tcpm_addr, &addr)) 270 break; 271 } 272 return tm; 273 } 274 275 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk, 276 struct dst_entry *dst, 277 bool create) 278 { 279 struct tcp_metrics_block *tm; 280 struct inetpeer_addr addr; 281 unsigned int hash; 282 struct net *net; 283 bool reclaim; 284 285 addr.family = sk->sk_family; 286 switch (addr.family) { 287 case AF_INET: 288 addr.addr.a4 = inet_sk(sk)->inet_daddr; 289 hash = (__force unsigned int) addr.addr.a4; 290 break; 291 case AF_INET6: 292 *(struct in6_addr *)addr.addr.a6 = inet6_sk(sk)->daddr; 293 hash = ipv6_addr_hash(&inet6_sk(sk)->daddr); 294 break; 295 default: 296 return NULL; 297 } 298 299 net = dev_net(dst->dev); 300 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); 301 302 tm = __tcp_get_metrics(&addr, net, hash); 303 reclaim = false; 304 if (tm == TCP_METRICS_RECLAIM_PTR) { 305 reclaim = true; 306 tm = NULL; 307 } 308 if (!tm && create) 309 tm = tcpm_new(dst, &addr, hash, reclaim); 310 else 311 tcpm_check_stamp(tm, dst); 312 313 return tm; 314 } 315 316 /* Save metrics learned by this TCP session. This function is called 317 * only, when TCP finishes successfully i.e. when it enters TIME-WAIT 318 * or goes from LAST-ACK to CLOSE. 319 */ 320 void tcp_update_metrics(struct sock *sk) 321 { 322 const struct inet_connection_sock *icsk = inet_csk(sk); 323 struct dst_entry *dst = __sk_dst_get(sk); 324 struct tcp_sock *tp = tcp_sk(sk); 325 struct tcp_metrics_block *tm; 326 unsigned long rtt; 327 u32 val; 328 int m; 329 330 if (sysctl_tcp_nometrics_save || !dst) 331 return; 332 333 if (dst->flags & DST_HOST) 334 dst_confirm(dst); 335 336 rcu_read_lock(); 337 if (icsk->icsk_backoff || !tp->srtt) { 338 /* This session failed to estimate rtt. Why? 339 * Probably, no packets returned in time. Reset our 340 * results. 341 */ 342 tm = tcp_get_metrics(sk, dst, false); 343 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT)) 344 tcp_metric_set(tm, TCP_METRIC_RTT, 0); 345 goto out_unlock; 346 } else 347 tm = tcp_get_metrics(sk, dst, true); 348 349 if (!tm) 350 goto out_unlock; 351 352 rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT); 353 m = rtt - tp->srtt; 354 355 /* If newly calculated rtt larger than stored one, store new 356 * one. Otherwise, use EWMA. Remember, rtt overestimation is 357 * always better than underestimation. 358 */ 359 if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) { 360 if (m <= 0) 361 rtt = tp->srtt; 362 else 363 rtt -= (m >> 3); 364 tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt); 365 } 366 367 if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) { 368 unsigned long var; 369 370 if (m < 0) 371 m = -m; 372 373 /* Scale deviation to rttvar fixed point */ 374 m >>= 1; 375 if (m < tp->mdev) 376 m = tp->mdev; 377 378 var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR); 379 if (m >= var) 380 var = m; 381 else 382 var -= (var - m) >> 2; 383 384 tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var); 385 } 386 387 if (tcp_in_initial_slowstart(tp)) { 388 /* Slow start still did not finish. */ 389 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) { 390 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); 391 if (val && (tp->snd_cwnd >> 1) > val) 392 tcp_metric_set(tm, TCP_METRIC_SSTHRESH, 393 tp->snd_cwnd >> 1); 394 } 395 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { 396 val = tcp_metric_get(tm, TCP_METRIC_CWND); 397 if (tp->snd_cwnd > val) 398 tcp_metric_set(tm, TCP_METRIC_CWND, 399 tp->snd_cwnd); 400 } 401 } else if (tp->snd_cwnd > tp->snd_ssthresh && 402 icsk->icsk_ca_state == TCP_CA_Open) { 403 /* Cong. avoidance phase, cwnd is reliable. */ 404 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) 405 tcp_metric_set(tm, TCP_METRIC_SSTHRESH, 406 max(tp->snd_cwnd >> 1, tp->snd_ssthresh)); 407 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { 408 val = tcp_metric_get(tm, TCP_METRIC_CWND); 409 tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1); 410 } 411 } else { 412 /* Else slow start did not finish, cwnd is non-sense, 413 * ssthresh may be also invalid. 414 */ 415 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) { 416 val = tcp_metric_get(tm, TCP_METRIC_CWND); 417 tcp_metric_set(tm, TCP_METRIC_CWND, 418 (val + tp->snd_ssthresh) >> 1); 419 } 420 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) { 421 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); 422 if (val && tp->snd_ssthresh > val) 423 tcp_metric_set(tm, TCP_METRIC_SSTHRESH, 424 tp->snd_ssthresh); 425 } 426 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) { 427 val = tcp_metric_get(tm, TCP_METRIC_REORDERING); 428 if (val < tp->reordering && 429 tp->reordering != sysctl_tcp_reordering) 430 tcp_metric_set(tm, TCP_METRIC_REORDERING, 431 tp->reordering); 432 } 433 } 434 tm->tcpm_stamp = jiffies; 435 out_unlock: 436 rcu_read_unlock(); 437 } 438 439 /* Initialize metrics on socket. */ 440 441 void tcp_init_metrics(struct sock *sk) 442 { 443 struct dst_entry *dst = __sk_dst_get(sk); 444 struct tcp_sock *tp = tcp_sk(sk); 445 struct tcp_metrics_block *tm; 446 u32 val, crtt = 0; /* cached RTT scaled by 8 */ 447 448 if (dst == NULL) 449 goto reset; 450 451 dst_confirm(dst); 452 453 rcu_read_lock(); 454 tm = tcp_get_metrics(sk, dst, true); 455 if (!tm) { 456 rcu_read_unlock(); 457 goto reset; 458 } 459 460 if (tcp_metric_locked(tm, TCP_METRIC_CWND)) 461 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND); 462 463 val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH); 464 if (val) { 465 tp->snd_ssthresh = val; 466 if (tp->snd_ssthresh > tp->snd_cwnd_clamp) 467 tp->snd_ssthresh = tp->snd_cwnd_clamp; 468 } else { 469 /* ssthresh may have been reduced unnecessarily during. 470 * 3WHS. Restore it back to its initial default. 471 */ 472 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; 473 } 474 val = tcp_metric_get(tm, TCP_METRIC_REORDERING); 475 if (val && tp->reordering != val) { 476 tcp_disable_fack(tp); 477 tcp_disable_early_retrans(tp); 478 tp->reordering = val; 479 } 480 481 crtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT); 482 rcu_read_unlock(); 483 reset: 484 /* The initial RTT measurement from the SYN/SYN-ACK is not ideal 485 * to seed the RTO for later data packets because SYN packets are 486 * small. Use the per-dst cached values to seed the RTO but keep 487 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar). 488 * Later the RTO will be updated immediately upon obtaining the first 489 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only 490 * influences the first RTO but not later RTT estimation. 491 * 492 * But if RTT is not available from the SYN (due to retransmits or 493 * syn cookies) or the cache, force a conservative 3secs timeout. 494 * 495 * A bit of theory. RTT is time passed after "normal" sized packet 496 * is sent until it is ACKed. In normal circumstances sending small 497 * packets force peer to delay ACKs and calculation is correct too. 498 * The algorithm is adaptive and, provided we follow specs, it 499 * NEVER underestimate RTT. BUT! If peer tries to make some clever 500 * tricks sort of "quick acks" for time long enough to decrease RTT 501 * to low value, and then abruptly stops to do it and starts to delay 502 * ACKs, wait for troubles. 503 */ 504 if (crtt > tp->srtt) { 505 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */ 506 crtt >>= 3; 507 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk)); 508 } else if (tp->srtt == 0) { 509 /* RFC6298: 5.7 We've failed to get a valid RTT sample from 510 * 3WHS. This is most likely due to retransmission, 511 * including spurious one. Reset the RTO back to 3secs 512 * from the more aggressive 1sec to avoid more spurious 513 * retransmission. 514 */ 515 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK; 516 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK; 517 } 518 /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been 519 * retransmitted. In light of RFC6298 more aggressive 1sec 520 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK 521 * retransmission has occurred. 522 */ 523 if (tp->total_retrans > 1) 524 tp->snd_cwnd = 1; 525 else 526 tp->snd_cwnd = tcp_init_cwnd(tp, dst); 527 tp->snd_cwnd_stamp = tcp_time_stamp; 528 } 529 530 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check) 531 { 532 struct tcp_metrics_block *tm; 533 bool ret; 534 535 if (!dst) 536 return false; 537 538 rcu_read_lock(); 539 tm = __tcp_get_metrics_req(req, dst); 540 if (paws_check) { 541 if (tm && 542 (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL && 543 (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW) 544 ret = false; 545 else 546 ret = true; 547 } else { 548 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp) 549 ret = true; 550 else 551 ret = false; 552 } 553 rcu_read_unlock(); 554 555 return ret; 556 } 557 EXPORT_SYMBOL_GPL(tcp_peer_is_proven); 558 559 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst) 560 { 561 struct tcp_metrics_block *tm; 562 563 rcu_read_lock(); 564 tm = tcp_get_metrics(sk, dst, true); 565 if (tm) { 566 struct tcp_sock *tp = tcp_sk(sk); 567 568 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) { 569 tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp; 570 tp->rx_opt.ts_recent = tm->tcpm_ts; 571 } 572 } 573 rcu_read_unlock(); 574 } 575 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp); 576 577 /* VJ's idea. Save last timestamp seen from this destination and hold 578 * it at least for normal timewait interval to use for duplicate 579 * segment detection in subsequent connections, before they enter 580 * synchronized state. 581 */ 582 bool tcp_remember_stamp(struct sock *sk) 583 { 584 struct dst_entry *dst = __sk_dst_get(sk); 585 bool ret = false; 586 587 if (dst) { 588 struct tcp_metrics_block *tm; 589 590 rcu_read_lock(); 591 tm = tcp_get_metrics(sk, dst, true); 592 if (tm) { 593 struct tcp_sock *tp = tcp_sk(sk); 594 595 if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 || 596 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL && 597 tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) { 598 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp; 599 tm->tcpm_ts = tp->rx_opt.ts_recent; 600 } 601 ret = true; 602 } 603 rcu_read_unlock(); 604 } 605 return ret; 606 } 607 608 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw) 609 { 610 struct tcp_metrics_block *tm; 611 bool ret = false; 612 613 rcu_read_lock(); 614 tm = __tcp_get_metrics_tw(tw); 615 if (tm) { 616 const struct tcp_timewait_sock *tcptw; 617 struct sock *sk = (struct sock *) tw; 618 619 tcptw = tcp_twsk(sk); 620 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 || 621 ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL && 622 tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) { 623 tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp; 624 tm->tcpm_ts = tcptw->tw_ts_recent; 625 } 626 ret = true; 627 } 628 rcu_read_unlock(); 629 630 return ret; 631 } 632 633 static DEFINE_SEQLOCK(fastopen_seqlock); 634 635 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss, 636 struct tcp_fastopen_cookie *cookie, 637 int *syn_loss, unsigned long *last_syn_loss) 638 { 639 struct tcp_metrics_block *tm; 640 641 rcu_read_lock(); 642 tm = tcp_get_metrics(sk, __sk_dst_get(sk), false); 643 if (tm) { 644 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen; 645 unsigned int seq; 646 647 do { 648 seq = read_seqbegin(&fastopen_seqlock); 649 if (tfom->mss) 650 *mss = tfom->mss; 651 *cookie = tfom->cookie; 652 *syn_loss = tfom->syn_loss; 653 *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0; 654 } while (read_seqretry(&fastopen_seqlock, seq)); 655 } 656 rcu_read_unlock(); 657 } 658 659 void tcp_fastopen_cache_set(struct sock *sk, u16 mss, 660 struct tcp_fastopen_cookie *cookie, bool syn_lost) 661 { 662 struct tcp_metrics_block *tm; 663 664 rcu_read_lock(); 665 tm = tcp_get_metrics(sk, __sk_dst_get(sk), true); 666 if (tm) { 667 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen; 668 669 write_seqlock_bh(&fastopen_seqlock); 670 tfom->mss = mss; 671 if (cookie->len > 0) 672 tfom->cookie = *cookie; 673 if (syn_lost) { 674 ++tfom->syn_loss; 675 tfom->last_syn_loss = jiffies; 676 } else 677 tfom->syn_loss = 0; 678 write_sequnlock_bh(&fastopen_seqlock); 679 } 680 rcu_read_unlock(); 681 } 682 683 static struct genl_family tcp_metrics_nl_family = { 684 .id = GENL_ID_GENERATE, 685 .hdrsize = 0, 686 .name = TCP_METRICS_GENL_NAME, 687 .version = TCP_METRICS_GENL_VERSION, 688 .maxattr = TCP_METRICS_ATTR_MAX, 689 .netnsok = true, 690 }; 691 692 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = { 693 [TCP_METRICS_ATTR_ADDR_IPV4] = { .type = NLA_U32, }, 694 [TCP_METRICS_ATTR_ADDR_IPV6] = { .type = NLA_BINARY, 695 .len = sizeof(struct in6_addr), }, 696 /* Following attributes are not received for GET/DEL, 697 * we keep them for reference 698 */ 699 #if 0 700 [TCP_METRICS_ATTR_AGE] = { .type = NLA_MSECS, }, 701 [TCP_METRICS_ATTR_TW_TSVAL] = { .type = NLA_U32, }, 702 [TCP_METRICS_ATTR_TW_TS_STAMP] = { .type = NLA_S32, }, 703 [TCP_METRICS_ATTR_VALS] = { .type = NLA_NESTED, }, 704 [TCP_METRICS_ATTR_FOPEN_MSS] = { .type = NLA_U16, }, 705 [TCP_METRICS_ATTR_FOPEN_SYN_DROPS] = { .type = NLA_U16, }, 706 [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS] = { .type = NLA_MSECS, }, 707 [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY, 708 .len = TCP_FASTOPEN_COOKIE_MAX, }, 709 #endif 710 }; 711 712 /* Add attributes, caller cancels its header on failure */ 713 static int tcp_metrics_fill_info(struct sk_buff *msg, 714 struct tcp_metrics_block *tm) 715 { 716 struct nlattr *nest; 717 int i; 718 719 switch (tm->tcpm_addr.family) { 720 case AF_INET: 721 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4, 722 tm->tcpm_addr.addr.a4) < 0) 723 goto nla_put_failure; 724 break; 725 case AF_INET6: 726 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16, 727 tm->tcpm_addr.addr.a6) < 0) 728 goto nla_put_failure; 729 break; 730 default: 731 return -EAFNOSUPPORT; 732 } 733 734 if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE, 735 jiffies - tm->tcpm_stamp) < 0) 736 goto nla_put_failure; 737 if (tm->tcpm_ts_stamp) { 738 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP, 739 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0) 740 goto nla_put_failure; 741 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL, 742 tm->tcpm_ts) < 0) 743 goto nla_put_failure; 744 } 745 746 { 747 int n = 0; 748 749 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS); 750 if (!nest) 751 goto nla_put_failure; 752 for (i = 0; i < TCP_METRIC_MAX + 1; i++) { 753 if (!tm->tcpm_vals[i]) 754 continue; 755 if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0) 756 goto nla_put_failure; 757 n++; 758 } 759 if (n) 760 nla_nest_end(msg, nest); 761 else 762 nla_nest_cancel(msg, nest); 763 } 764 765 { 766 struct tcp_fastopen_metrics tfom_copy[1], *tfom; 767 unsigned int seq; 768 769 do { 770 seq = read_seqbegin(&fastopen_seqlock); 771 tfom_copy[0] = tm->tcpm_fastopen; 772 } while (read_seqretry(&fastopen_seqlock, seq)); 773 774 tfom = tfom_copy; 775 if (tfom->mss && 776 nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS, 777 tfom->mss) < 0) 778 goto nla_put_failure; 779 if (tfom->syn_loss && 780 (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS, 781 tfom->syn_loss) < 0 || 782 nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS, 783 jiffies - tfom->last_syn_loss) < 0)) 784 goto nla_put_failure; 785 if (tfom->cookie.len > 0 && 786 nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE, 787 tfom->cookie.len, tfom->cookie.val) < 0) 788 goto nla_put_failure; 789 } 790 791 return 0; 792 793 nla_put_failure: 794 return -EMSGSIZE; 795 } 796 797 static int tcp_metrics_dump_info(struct sk_buff *skb, 798 struct netlink_callback *cb, 799 struct tcp_metrics_block *tm) 800 { 801 void *hdr; 802 803 hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq, 804 &tcp_metrics_nl_family, NLM_F_MULTI, 805 TCP_METRICS_CMD_GET); 806 if (!hdr) 807 return -EMSGSIZE; 808 809 if (tcp_metrics_fill_info(skb, tm) < 0) 810 goto nla_put_failure; 811 812 return genlmsg_end(skb, hdr); 813 814 nla_put_failure: 815 genlmsg_cancel(skb, hdr); 816 return -EMSGSIZE; 817 } 818 819 static int tcp_metrics_nl_dump(struct sk_buff *skb, 820 struct netlink_callback *cb) 821 { 822 struct net *net = sock_net(skb->sk); 823 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log; 824 unsigned int row, s_row = cb->args[0]; 825 int s_col = cb->args[1], col = s_col; 826 827 for (row = s_row; row < max_rows; row++, s_col = 0) { 828 struct tcp_metrics_block *tm; 829 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row; 830 831 rcu_read_lock(); 832 for (col = 0, tm = rcu_dereference(hb->chain); tm; 833 tm = rcu_dereference(tm->tcpm_next), col++) { 834 if (col < s_col) 835 continue; 836 if (tcp_metrics_dump_info(skb, cb, tm) < 0) { 837 rcu_read_unlock(); 838 goto done; 839 } 840 } 841 rcu_read_unlock(); 842 } 843 844 done: 845 cb->args[0] = row; 846 cb->args[1] = col; 847 return skb->len; 848 } 849 850 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr, 851 unsigned int *hash, int optional) 852 { 853 struct nlattr *a; 854 855 a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV4]; 856 if (a) { 857 addr->family = AF_INET; 858 addr->addr.a4 = nla_get_be32(a); 859 *hash = (__force unsigned int) addr->addr.a4; 860 return 0; 861 } 862 a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV6]; 863 if (a) { 864 if (nla_len(a) != sizeof(struct in6_addr)) 865 return -EINVAL; 866 addr->family = AF_INET6; 867 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6)); 868 *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6); 869 return 0; 870 } 871 return optional ? 1 : -EAFNOSUPPORT; 872 } 873 874 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info) 875 { 876 struct tcp_metrics_block *tm; 877 struct inetpeer_addr addr; 878 unsigned int hash; 879 struct sk_buff *msg; 880 struct net *net = genl_info_net(info); 881 void *reply; 882 int ret; 883 884 ret = parse_nl_addr(info, &addr, &hash, 0); 885 if (ret < 0) 886 return ret; 887 888 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL); 889 if (!msg) 890 return -ENOMEM; 891 892 reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0, 893 info->genlhdr->cmd); 894 if (!reply) 895 goto nla_put_failure; 896 897 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); 898 ret = -ESRCH; 899 rcu_read_lock(); 900 for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm; 901 tm = rcu_dereference(tm->tcpm_next)) { 902 if (addr_same(&tm->tcpm_addr, &addr)) { 903 ret = tcp_metrics_fill_info(msg, tm); 904 break; 905 } 906 } 907 rcu_read_unlock(); 908 if (ret < 0) 909 goto out_free; 910 911 genlmsg_end(msg, reply); 912 return genlmsg_reply(msg, info); 913 914 nla_put_failure: 915 ret = -EMSGSIZE; 916 917 out_free: 918 nlmsg_free(msg); 919 return ret; 920 } 921 922 #define deref_locked_genl(p) \ 923 rcu_dereference_protected(p, lockdep_genl_is_held() && \ 924 lockdep_is_held(&tcp_metrics_lock)) 925 926 #define deref_genl(p) rcu_dereference_protected(p, lockdep_genl_is_held()) 927 928 static int tcp_metrics_flush_all(struct net *net) 929 { 930 unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log; 931 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash; 932 struct tcp_metrics_block *tm; 933 unsigned int row; 934 935 for (row = 0; row < max_rows; row++, hb++) { 936 spin_lock_bh(&tcp_metrics_lock); 937 tm = deref_locked_genl(hb->chain); 938 if (tm) 939 hb->chain = NULL; 940 spin_unlock_bh(&tcp_metrics_lock); 941 while (tm) { 942 struct tcp_metrics_block *next; 943 944 next = deref_genl(tm->tcpm_next); 945 kfree_rcu(tm, rcu_head); 946 tm = next; 947 } 948 } 949 return 0; 950 } 951 952 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info) 953 { 954 struct tcpm_hash_bucket *hb; 955 struct tcp_metrics_block *tm; 956 struct tcp_metrics_block __rcu **pp; 957 struct inetpeer_addr addr; 958 unsigned int hash; 959 struct net *net = genl_info_net(info); 960 int ret; 961 962 ret = parse_nl_addr(info, &addr, &hash, 1); 963 if (ret < 0) 964 return ret; 965 if (ret > 0) 966 return tcp_metrics_flush_all(net); 967 968 hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log); 969 hb = net->ipv4.tcp_metrics_hash + hash; 970 pp = &hb->chain; 971 spin_lock_bh(&tcp_metrics_lock); 972 for (tm = deref_locked_genl(*pp); tm; 973 pp = &tm->tcpm_next, tm = deref_locked_genl(*pp)) { 974 if (addr_same(&tm->tcpm_addr, &addr)) { 975 *pp = tm->tcpm_next; 976 break; 977 } 978 } 979 spin_unlock_bh(&tcp_metrics_lock); 980 if (!tm) 981 return -ESRCH; 982 kfree_rcu(tm, rcu_head); 983 return 0; 984 } 985 986 static struct genl_ops tcp_metrics_nl_ops[] = { 987 { 988 .cmd = TCP_METRICS_CMD_GET, 989 .doit = tcp_metrics_nl_cmd_get, 990 .dumpit = tcp_metrics_nl_dump, 991 .policy = tcp_metrics_nl_policy, 992 .flags = GENL_ADMIN_PERM, 993 }, 994 { 995 .cmd = TCP_METRICS_CMD_DEL, 996 .doit = tcp_metrics_nl_cmd_del, 997 .policy = tcp_metrics_nl_policy, 998 .flags = GENL_ADMIN_PERM, 999 }, 1000 }; 1001 1002 static unsigned int tcpmhash_entries; 1003 static int __init set_tcpmhash_entries(char *str) 1004 { 1005 ssize_t ret; 1006 1007 if (!str) 1008 return 0; 1009 1010 ret = kstrtouint(str, 0, &tcpmhash_entries); 1011 if (ret) 1012 return 0; 1013 1014 return 1; 1015 } 1016 __setup("tcpmhash_entries=", set_tcpmhash_entries); 1017 1018 static int __net_init tcp_net_metrics_init(struct net *net) 1019 { 1020 size_t size; 1021 unsigned int slots; 1022 1023 slots = tcpmhash_entries; 1024 if (!slots) { 1025 if (totalram_pages >= 128 * 1024) 1026 slots = 16 * 1024; 1027 else 1028 slots = 8 * 1024; 1029 } 1030 1031 net->ipv4.tcp_metrics_hash_log = order_base_2(slots); 1032 size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log; 1033 1034 net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN); 1035 if (!net->ipv4.tcp_metrics_hash) 1036 net->ipv4.tcp_metrics_hash = vzalloc(size); 1037 1038 if (!net->ipv4.tcp_metrics_hash) 1039 return -ENOMEM; 1040 1041 return 0; 1042 } 1043 1044 static void __net_exit tcp_net_metrics_exit(struct net *net) 1045 { 1046 unsigned int i; 1047 1048 for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) { 1049 struct tcp_metrics_block *tm, *next; 1050 1051 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1); 1052 while (tm) { 1053 next = rcu_dereference_protected(tm->tcpm_next, 1); 1054 kfree(tm); 1055 tm = next; 1056 } 1057 } 1058 if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash)) 1059 vfree(net->ipv4.tcp_metrics_hash); 1060 else 1061 kfree(net->ipv4.tcp_metrics_hash); 1062 } 1063 1064 static __net_initdata struct pernet_operations tcp_net_metrics_ops = { 1065 .init = tcp_net_metrics_init, 1066 .exit = tcp_net_metrics_exit, 1067 }; 1068 1069 void __init tcp_metrics_init(void) 1070 { 1071 int ret; 1072 1073 ret = register_pernet_subsys(&tcp_net_metrics_ops); 1074 if (ret < 0) 1075 goto cleanup; 1076 ret = genl_register_family_with_ops(&tcp_metrics_nl_family, 1077 tcp_metrics_nl_ops, 1078 ARRAY_SIZE(tcp_metrics_nl_ops)); 1079 if (ret < 0) 1080 goto cleanup_subsys; 1081 return; 1082 1083 cleanup_subsys: 1084 unregister_pernet_subsys(&tcp_net_metrics_ops); 1085 1086 cleanup: 1087 return; 1088 } 1089