net/ipv4/tcp_metrics.c

   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         if (crtt > tp->srtt) {
 485                 /* Initial RTT (tp->srtt) from SYN usually don't measure
 486                  * serialization delay on low BW links well so RTO may be
 487                  * under-estimated. Stay conservative and seed RTO with
 488                  * the RTTs from past data exchanges, using the same seeding
 489                  * formula in tcp_rtt_estimator().
 490                  */
 491                 inet_csk(sk)->icsk_rto = crtt + max(crtt >> 2, tcp_rto_min(sk));
 492         } else if (tp->srtt == 0) {
 493                 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
 494                  * 3WHS. This is most likely due to retransmission,
 495                  * including spurious one. Reset the RTO back to 3secs
 496                  * from the more aggressive 1sec to avoid more spurious
 497                  * retransmission.
 498                  */
 499                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
 500                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
 501         }
 502         /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
 503          * retransmitted. In light of RFC6298 more aggressive 1sec
 504          * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
 505          * retransmission has occurred.
 506          */
 507         if (tp->total_retrans > 1)
 508                 tp->snd_cwnd = 1;
 509         else
 510                 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
 511         tp->snd_cwnd_stamp = tcp_time_stamp;
 512 }
 513
 514 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
 515 {
 516         struct tcp_metrics_block *tm;
 517         bool ret;
 518
 519         if (!dst)
 520                 return false;
 521
 522         rcu_read_lock();
 523         tm = __tcp_get_metrics_req(req, dst);
 524         if (paws_check) {
 525                 if (tm &&
 526                     (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
 527                     (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
 528                         ret = false;
 529                 else
 530                         ret = true;
 531         } else {
 532                 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
 533                         ret = true;
 534                 else
 535                         ret = false;
 536         }
 537         rcu_read_unlock();
 538
 539         return ret;
 540 }
 541 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
 542
 543 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
 544 {
 545         struct tcp_metrics_block *tm;
 546
 547         rcu_read_lock();
 548         tm = tcp_get_metrics(sk, dst, true);
 549         if (tm) {
 550                 struct tcp_sock *tp = tcp_sk(sk);
 551
 552                 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
 553                         tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
 554                         tp->rx_opt.ts_recent = tm->tcpm_ts;
 555                 }
 556         }
 557         rcu_read_unlock();
 558 }
 559 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
 560
 561 /* VJ's idea. Save last timestamp seen from this destination and hold
 562  * it at least for normal timewait interval to use for duplicate
 563  * segment detection in subsequent connections, before they enter
 564  * synchronized state.
 565  */
 566 bool tcp_remember_stamp(struct sock *sk)
 567 {
 568         struct dst_entry *dst = __sk_dst_get(sk);
 569         bool ret = false;
 570
 571         if (dst) {
 572                 struct tcp_metrics_block *tm;
 573
 574                 rcu_read_lock();
 575                 tm = tcp_get_metrics(sk, dst, true);
 576                 if (tm) {
 577                         struct tcp_sock *tp = tcp_sk(sk);
 578
 579                         if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
 580                             ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 581                              tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
 582                                 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
 583                                 tm->tcpm_ts = tp->rx_opt.ts_recent;
 584                         }
 585                         ret = true;
 586                 }
 587                 rcu_read_unlock();
 588         }
 589         return ret;
 590 }
 591
 592 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
 593 {
 594         struct tcp_metrics_block *tm;
 595         bool ret = false;
 596
 597         rcu_read_lock();
 598         tm = __tcp_get_metrics_tw(tw);
 599         if (tm) {
 600                 const struct tcp_timewait_sock *tcptw;
 601                 struct sock *sk = (struct sock *) tw;
 602
 603                 tcptw = tcp_twsk(sk);
 604                 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
 605                     ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 606                      tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
 607                         tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
 608                         tm->tcpm_ts        = tcptw->tw_ts_recent;
 609                 }
 610                 ret = true;
 611         }
 612         rcu_read_unlock();
 613
 614         return ret;
 615 }
 616
 617 static DEFINE_SEQLOCK(fastopen_seqlock);
 618
 619 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
 620                             struct tcp_fastopen_cookie *cookie,
 621                             int *syn_loss, unsigned long *last_syn_loss)
 622 {
 623         struct tcp_metrics_block *tm;
 624
 625         rcu_read_lock();
 626         tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
 627         if (tm) {
 628                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 629                 unsigned int seq;
 630
 631                 do {
 632                         seq = read_seqbegin(&fastopen_seqlock);
 633                         if (tfom->mss)
 634                                 *mss = tfom->mss;
 635                         *cookie = tfom->cookie;
 636                         *syn_loss = tfom->syn_loss;
 637                         *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
 638                 } while (read_seqretry(&fastopen_seqlock, seq));
 639         }
 640         rcu_read_unlock();
 641 }
 642
 643 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
 644                             struct tcp_fastopen_cookie *cookie, bool syn_lost)
 645 {
 646         struct tcp_metrics_block *tm;
 647
 648         rcu_read_lock();
 649         tm = tcp_get_metrics(sk, __sk_dst_get(sk), true);
 650         if (tm) {
 651                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 652
 653                 write_seqlock_bh(&fastopen_seqlock);
 654                 tfom->mss = mss;
 655                 if (cookie->len > 0)
 656                         tfom->cookie = *cookie;
 657                 if (syn_lost) {
 658                         ++tfom->syn_loss;
 659                         tfom->last_syn_loss = jiffies;
 660                 } else
 661                         tfom->syn_loss = 0;
 662                 write_sequnlock_bh(&fastopen_seqlock);
 663         }
 664         rcu_read_unlock();
 665 }
 666
 667 static struct genl_family tcp_metrics_nl_family = {
 668         .id             = GENL_ID_GENERATE,
 669         .hdrsize        = 0,
 670         .name           = TCP_METRICS_GENL_NAME,
 671         .version        = TCP_METRICS_GENL_VERSION,
 672         .maxattr        = TCP_METRICS_ATTR_MAX,
 673         .netnsok        = true,
 674 };
 675
 676 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
 677         [TCP_METRICS_ATTR_ADDR_IPV4]    = { .type = NLA_U32, },
 678         [TCP_METRICS_ATTR_ADDR_IPV6]    = { .type = NLA_BINARY,
 679                                             .len = sizeof(struct in6_addr), },
 680         /* Following attributes are not received for GET/DEL,
 681          * we keep them for reference
 682          */
 683 #if 0
 684         [TCP_METRICS_ATTR_AGE]          = { .type = NLA_MSECS, },
 685         [TCP_METRICS_ATTR_TW_TSVAL]     = { .type = NLA_U32, },
 686         [TCP_METRICS_ATTR_TW_TS_STAMP]  = { .type = NLA_S32, },
 687         [TCP_METRICS_ATTR_VALS]         = { .type = NLA_NESTED, },
 688         [TCP_METRICS_ATTR_FOPEN_MSS]    = { .type = NLA_U16, },
 689         [TCP_METRICS_ATTR_FOPEN_SYN_DROPS]      = { .type = NLA_U16, },
 690         [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]    = { .type = NLA_MSECS, },
 691         [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
 692                                             .len = TCP_FASTOPEN_COOKIE_MAX, },
 693 #endif
 694 };
 695
 696 /* Add attributes, caller cancels its header on failure */
 697 static int tcp_metrics_fill_info(struct sk_buff *msg,
 698                                  struct tcp_metrics_block *tm)
 699 {
 700         struct nlattr *nest;
 701         int i;
 702
 703         switch (tm->tcpm_addr.family) {
 704         case AF_INET:
 705                 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
 706                                 tm->tcpm_addr.addr.a4) < 0)
 707                         goto nla_put_failure;
 708                 break;
 709         case AF_INET6:
 710                 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
 711                             tm->tcpm_addr.addr.a6) < 0)
 712                         goto nla_put_failure;
 713                 break;
 714         default:
 715                 return -EAFNOSUPPORT;
 716         }
 717
 718         if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
 719                           jiffies - tm->tcpm_stamp) < 0)
 720                 goto nla_put_failure;
 721         if (tm->tcpm_ts_stamp) {
 722                 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
 723                                 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
 724                         goto nla_put_failure;
 725                 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
 726                                 tm->tcpm_ts) < 0)
 727                         goto nla_put_failure;
 728         }
 729
 730         {
 731                 int n = 0;
 732
 733                 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
 734                 if (!nest)
 735                         goto nla_put_failure;
 736                 for (i = 0; i < TCP_METRIC_MAX + 1; i++) {
 737                         if (!tm->tcpm_vals[i])
 738                                 continue;
 739                         if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0)
 740                                 goto nla_put_failure;
 741                         n++;
 742                 }
 743                 if (n)
 744                         nla_nest_end(msg, nest);
 745                 else
 746                         nla_nest_cancel(msg, nest);
 747         }
 748
 749         {
 750                 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
 751                 unsigned int seq;
 752
 753                 do {
 754                         seq = read_seqbegin(&fastopen_seqlock);
 755                         tfom_copy[0] = tm->tcpm_fastopen;
 756                 } while (read_seqretry(&fastopen_seqlock, seq));
 757
 758                 tfom = tfom_copy;
 759                 if (tfom->mss &&
 760                     nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
 761                                 tfom->mss) < 0)
 762                         goto nla_put_failure;
 763                 if (tfom->syn_loss &&
 764                     (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
 765                                 tfom->syn_loss) < 0 ||
 766                      nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
 767                                 jiffies - tfom->last_syn_loss) < 0))
 768                         goto nla_put_failure;
 769                 if (tfom->cookie.len > 0 &&
 770                     nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
 771                             tfom->cookie.len, tfom->cookie.val) < 0)
 772                         goto nla_put_failure;
 773         }
 774
 775         return 0;
 776
 777 nla_put_failure:
 778         return -EMSGSIZE;
 779 }
 780
 781 static int tcp_metrics_dump_info(struct sk_buff *skb,
 782                                  struct netlink_callback *cb,
 783                                  struct tcp_metrics_block *tm)
 784 {
 785         void *hdr;
 786
 787         hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 788                           &tcp_metrics_nl_family, NLM_F_MULTI,
 789                           TCP_METRICS_CMD_GET);
 790         if (!hdr)
 791                 return -EMSGSIZE;
 792
 793         if (tcp_metrics_fill_info(skb, tm) < 0)
 794                 goto nla_put_failure;
 795
 796         return genlmsg_end(skb, hdr);
 797
 798 nla_put_failure:
 799         genlmsg_cancel(skb, hdr);
 800         return -EMSGSIZE;
 801 }
 802
 803 static int tcp_metrics_nl_dump(struct sk_buff *skb,
 804                                struct netlink_callback *cb)
 805 {
 806         struct net *net = sock_net(skb->sk);
 807         unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
 808         unsigned int row, s_row = cb->args[0];
 809         int s_col = cb->args[1], col = s_col;
 810
 811         for (row = s_row; row < max_rows; row++, s_col = 0) {
 812                 struct tcp_metrics_block *tm;
 813                 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
 814
 815                 rcu_read_lock();
 816                 for (col = 0, tm = rcu_dereference(hb->chain); tm;
 817                      tm = rcu_dereference(tm->tcpm_next), col++) {
 818                         if (col < s_col)
 819                                 continue;
 820                         if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
 821                                 rcu_read_unlock();
 822                                 goto done;
 823                         }
 824                 }
 825                 rcu_read_unlock();
 826         }
 827
 828 done:
 829         cb->args[0] = row;
 830         cb->args[1] = col;
 831         return skb->len;
 832 }
 833
 834 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 835                          unsigned int *hash, int optional)
 836 {
 837         struct nlattr *a;
 838
 839         a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV4];
 840         if (a) {
 841                 addr->family = AF_INET;
 842                 addr->addr.a4 = nla_get_be32(a);
 843                 *hash = (__force unsigned int) addr->addr.a4;
 844                 return 0;
 845         }
 846         a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV6];
 847         if (a) {
 848                 if (nla_len(a) != sizeof(struct in6_addr))
 849                         return -EINVAL;
 850                 addr->family = AF_INET6;
 851                 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
 852                 *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
 853                 return 0;
 854         }
 855         return optional ? 1 : -EAFNOSUPPORT;
 856 }
 857
 858 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
 859 {
 860         struct tcp_metrics_block *tm;
 861         struct inetpeer_addr addr;
 862         unsigned int hash;
 863         struct sk_buff *msg;
 864         struct net *net = genl_info_net(info);
 865         void *reply;
 866         int ret;
 867
 868         ret = parse_nl_addr(info, &addr, &hash, 0);
 869         if (ret < 0)
 870                 return ret;
 871
 872         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 873         if (!msg)
 874                 return -ENOMEM;
 875
 876         reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
 877                                   info->genlhdr->cmd);
 878         if (!reply)
 879                 goto nla_put_failure;
 880
 881         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 882         ret = -ESRCH;
 883         rcu_read_lock();
 884         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 885              tm = rcu_dereference(tm->tcpm_next)) {
 886                 if (addr_same(&tm->tcpm_addr, &addr)) {
 887                         ret = tcp_metrics_fill_info(msg, tm);
 888                         break;
 889                 }
 890         }
 891         rcu_read_unlock();
 892         if (ret < 0)
 893                 goto out_free;
 894
 895         genlmsg_end(msg, reply);
 896         return genlmsg_reply(msg, info);
 897
 898 nla_put_failure:
 899         ret = -EMSGSIZE;
 900
 901 out_free:
 902         nlmsg_free(msg);
 903         return ret;
 904 }
 905
 906 #define deref_locked_genl(p)    \
 907         rcu_dereference_protected(p, lockdep_genl_is_held() && \
 908                                      lockdep_is_held(&tcp_metrics_lock))
 909
 910 #define deref_genl(p)   rcu_dereference_protected(p, lockdep_genl_is_held())
 911
 912 static int tcp_metrics_flush_all(struct net *net)
 913 {
 914         unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
 915         struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
 916         struct tcp_metrics_block *tm;
 917         unsigned int row;
 918
 919         for (row = 0; row < max_rows; row++, hb++) {
 920                 spin_lock_bh(&tcp_metrics_lock);
 921                 tm = deref_locked_genl(hb->chain);
 922                 if (tm)
 923                         hb->chain = NULL;
 924                 spin_unlock_bh(&tcp_metrics_lock);
 925                 while (tm) {
 926                         struct tcp_metrics_block *next;
 927
 928                         next = deref_genl(tm->tcpm_next);
 929                         kfree_rcu(tm, rcu_head);
 930                         tm = next;
 931                 }
 932         }
 933         return 0;
 934 }
 935
 936 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
 937 {
 938         struct tcpm_hash_bucket *hb;
 939         struct tcp_metrics_block *tm;
 940         struct tcp_metrics_block __rcu **pp;
 941         struct inetpeer_addr addr;
 942         unsigned int hash;
 943         struct net *net = genl_info_net(info);
 944         int ret;
 945
 946         ret = parse_nl_addr(info, &addr, &hash, 1);
 947         if (ret < 0)
 948                 return ret;
 949         if (ret > 0)
 950                 return tcp_metrics_flush_all(net);
 951
 952         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 953         hb = net->ipv4.tcp_metrics_hash + hash;
 954         pp = &hb->chain;
 955         spin_lock_bh(&tcp_metrics_lock);
 956         for (tm = deref_locked_genl(*pp); tm;
 957              pp = &tm->tcpm_next, tm = deref_locked_genl(*pp)) {
 958                 if (addr_same(&tm->tcpm_addr, &addr)) {
 959                         *pp = tm->tcpm_next;
 960                         break;
 961                 }
 962         }
 963         spin_unlock_bh(&tcp_metrics_lock);
 964         if (!tm)
 965                 return -ESRCH;
 966         kfree_rcu(tm, rcu_head);
 967         return 0;
 968 }
 969
 970 static struct genl_ops tcp_metrics_nl_ops[] = {
 971         {
 972                 .cmd = TCP_METRICS_CMD_GET,
 973                 .doit = tcp_metrics_nl_cmd_get,
 974                 .dumpit = tcp_metrics_nl_dump,
 975                 .policy = tcp_metrics_nl_policy,
 976                 .flags = GENL_ADMIN_PERM,
 977         },
 978         {
 979                 .cmd = TCP_METRICS_CMD_DEL,
 980                 .doit = tcp_metrics_nl_cmd_del,
 981                 .policy = tcp_metrics_nl_policy,
 982                 .flags = GENL_ADMIN_PERM,
 983         },
 984 };
 985
 986 static unsigned int tcpmhash_entries;
 987 static int __init set_tcpmhash_entries(char *str)
 988 {
 989         ssize_t ret;
 990
 991         if (!str)
 992                 return 0;
 993
 994         ret = kstrtouint(str, 0, &tcpmhash_entries);
 995         if (ret)
 996                 return 0;
 997
 998         return 1;
 999 }
1000 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1001
1002 static int __net_init tcp_net_metrics_init(struct net *net)
1003 {
1004         size_t size;
1005         unsigned int slots;
1006
1007         slots = tcpmhash_entries;
1008         if (!slots) {
1009                 if (totalram_pages >= 128 * 1024)
1010                         slots = 16 * 1024;
1011                 else
1012                         slots = 8 * 1024;
1013         }
1014
1015         net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1016         size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
1017
1018         net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1019         if (!net->ipv4.tcp_metrics_hash)
1020                 net->ipv4.tcp_metrics_hash = vzalloc(size);
1021
1022         if (!net->ipv4.tcp_metrics_hash)
1023                 return -ENOMEM;
1024
1025         return 0;
1026 }
1027
1028 static void __net_exit tcp_net_metrics_exit(struct net *net)
1029 {
1030         unsigned int i;
1031
1032         for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1033                 struct tcp_metrics_block *tm, *next;
1034
1035                 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1036                 while (tm) {
1037                         next = rcu_dereference_protected(tm->tcpm_next, 1);
1038                         kfree(tm);
1039                         tm = next;
1040                 }
1041         }
1042         if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1043                 vfree(net->ipv4.tcp_metrics_hash);
1044         else
1045                 kfree(net->ipv4.tcp_metrics_hash);
1046 }
1047
1048 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1049         .init   =       tcp_net_metrics_init,
1050         .exit   =       tcp_net_metrics_exit,
1051 };
1052
1053 void __init tcp_metrics_init(void)
1054 {
1055         int ret;
1056
1057         ret = register_pernet_subsys(&tcp_net_metrics_ops);
1058         if (ret < 0)
1059                 goto cleanup;
1060         ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1061                                             tcp_metrics_nl_ops,
1062                                             ARRAY_SIZE(tcp_metrics_nl_ops));
1063         if (ret < 0)
1064                 goto cleanup_subsys;
1065         return;
1066
1067 cleanup_subsys:
1068         unregister_pernet_subsys(&tcp_net_metrics_ops);
1069
1070 cleanup:
1071         return;
1072 }