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;
 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         val = tcp_metric_get(tm, TCP_METRIC_RTT);
 482         if (val == 0 || tp->srtt == 0) {
 483                 rcu_read_unlock();
 484                 goto reset;
 485         }
 486         /* Initial rtt is determined from SYN,SYN-ACK.
 487          * The segment is small and rtt may appear much
 488          * less than real one. Use per-dst memory
 489          * to make it more realistic.
 490          *
 491          * A bit of theory. RTT is time passed after "normal" sized packet
 492          * is sent until it is ACKed. In normal circumstances sending small
 493          * packets force peer to delay ACKs and calculation is correct too.
 494          * The algorithm is adaptive and, provided we follow specs, it
 495          * NEVER underestimate RTT. BUT! If peer tries to make some clever
 496          * tricks sort of "quick acks" for time long enough to decrease RTT
 497          * to low value, and then abruptly stops to do it and starts to delay
 498          * ACKs, wait for troubles.
 499          */
 500         val = msecs_to_jiffies(val);
 501         if (val > tp->srtt) {
 502                 tp->srtt = val;
 503                 tp->rtt_seq = tp->snd_nxt;
 504         }
 505         val = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
 506         if (val > tp->mdev) {
 507                 tp->mdev = val;
 508                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
 509         }
 510         rcu_read_unlock();
 511
 512         tcp_set_rto(sk);
 513 reset:
 514         if (tp->srtt == 0) {
 515                 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
 516                  * 3WHS. This is most likely due to retransmission,
 517                  * including spurious one. Reset the RTO back to 3secs
 518                  * from the more aggressive 1sec to avoid more spurious
 519                  * retransmission.
 520                  */
 521                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
 522                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
 523         }
 524         /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
 525          * retransmitted. In light of RFC6298 more aggressive 1sec
 526          * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
 527          * retransmission has occurred.
 528          */
 529         if (tp->total_retrans > 1)
 530                 tp->snd_cwnd = 1;
 531         else
 532                 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
 533         tp->snd_cwnd_stamp = tcp_time_stamp;
 534 }
 535
 536 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
 537 {
 538         struct tcp_metrics_block *tm;
 539         bool ret;
 540
 541         if (!dst)
 542                 return false;
 543
 544         rcu_read_lock();
 545         tm = __tcp_get_metrics_req(req, dst);
 546         if (paws_check) {
 547                 if (tm &&
 548                     (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
 549                     (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
 550                         ret = false;
 551                 else
 552                         ret = true;
 553         } else {
 554                 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
 555                         ret = true;
 556                 else
 557                         ret = false;
 558         }
 559         rcu_read_unlock();
 560
 561         return ret;
 562 }
 563 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
 564
 565 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
 566 {
 567         struct tcp_metrics_block *tm;
 568
 569         rcu_read_lock();
 570         tm = tcp_get_metrics(sk, dst, true);
 571         if (tm) {
 572                 struct tcp_sock *tp = tcp_sk(sk);
 573
 574                 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
 575                         tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
 576                         tp->rx_opt.ts_recent = tm->tcpm_ts;
 577                 }
 578         }
 579         rcu_read_unlock();
 580 }
 581 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
 582
 583 /* VJ's idea. Save last timestamp seen from this destination and hold
 584  * it at least for normal timewait interval to use for duplicate
 585  * segment detection in subsequent connections, before they enter
 586  * synchronized state.
 587  */
 588 bool tcp_remember_stamp(struct sock *sk)
 589 {
 590         struct dst_entry *dst = __sk_dst_get(sk);
 591         bool ret = false;
 592
 593         if (dst) {
 594                 struct tcp_metrics_block *tm;
 595
 596                 rcu_read_lock();
 597                 tm = tcp_get_metrics(sk, dst, true);
 598                 if (tm) {
 599                         struct tcp_sock *tp = tcp_sk(sk);
 600
 601                         if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
 602                             ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 603                              tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
 604                                 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
 605                                 tm->tcpm_ts = tp->rx_opt.ts_recent;
 606                         }
 607                         ret = true;
 608                 }
 609                 rcu_read_unlock();
 610         }
 611         return ret;
 612 }
 613
 614 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
 615 {
 616         struct tcp_metrics_block *tm;
 617         bool ret = false;
 618
 619         rcu_read_lock();
 620         tm = __tcp_get_metrics_tw(tw);
 621         if (tm) {
 622                 const struct tcp_timewait_sock *tcptw;
 623                 struct sock *sk = (struct sock *) tw;
 624
 625                 tcptw = tcp_twsk(sk);
 626                 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
 627                     ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 628                      tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
 629                         tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
 630                         tm->tcpm_ts        = tcptw->tw_ts_recent;
 631                 }
 632                 ret = true;
 633         }
 634         rcu_read_unlock();
 635
 636         return ret;
 637 }
 638
 639 static DEFINE_SEQLOCK(fastopen_seqlock);
 640
 641 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
 642                             struct tcp_fastopen_cookie *cookie,
 643                             int *syn_loss, unsigned long *last_syn_loss)
 644 {
 645         struct tcp_metrics_block *tm;
 646
 647         rcu_read_lock();
 648         tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
 649         if (tm) {
 650                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 651                 unsigned int seq;
 652
 653                 do {
 654                         seq = read_seqbegin(&fastopen_seqlock);
 655                         if (tfom->mss)
 656                                 *mss = tfom->mss;
 657                         *cookie = tfom->cookie;
 658                         *syn_loss = tfom->syn_loss;
 659                         *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
 660                 } while (read_seqretry(&fastopen_seqlock, seq));
 661         }
 662         rcu_read_unlock();
 663 }
 664
 665 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
 666                             struct tcp_fastopen_cookie *cookie, bool syn_lost)
 667 {
 668         struct dst_entry *dst = __sk_dst_get(sk);
 669         struct tcp_metrics_block *tm;
 670
 671         if (!dst)
 672                 return;
 673         rcu_read_lock();
 674         tm = tcp_get_metrics(sk, dst, true);
 675         if (tm) {
 676                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 677
 678                 write_seqlock_bh(&fastopen_seqlock);
 679                 tfom->mss = mss;
 680                 if (cookie->len > 0)
 681                         tfom->cookie = *cookie;
 682                 if (syn_lost) {
 683                         ++tfom->syn_loss;
 684                         tfom->last_syn_loss = jiffies;
 685                 } else
 686                         tfom->syn_loss = 0;
 687                 write_sequnlock_bh(&fastopen_seqlock);
 688         }
 689         rcu_read_unlock();
 690 }
 691
 692 static struct genl_family tcp_metrics_nl_family = {
 693         .id             = GENL_ID_GENERATE,
 694         .hdrsize        = 0,
 695         .name           = TCP_METRICS_GENL_NAME,
 696         .version        = TCP_METRICS_GENL_VERSION,
 697         .maxattr        = TCP_METRICS_ATTR_MAX,
 698         .netnsok        = true,
 699 };
 700
 701 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
 702         [TCP_METRICS_ATTR_ADDR_IPV4]    = { .type = NLA_U32, },
 703         [TCP_METRICS_ATTR_ADDR_IPV6]    = { .type = NLA_BINARY,
 704                                             .len = sizeof(struct in6_addr), },
 705         /* Following attributes are not received for GET/DEL,
 706          * we keep them for reference
 707          */
 708 #if 0
 709         [TCP_METRICS_ATTR_AGE]          = { .type = NLA_MSECS, },
 710         [TCP_METRICS_ATTR_TW_TSVAL]     = { .type = NLA_U32, },
 711         [TCP_METRICS_ATTR_TW_TS_STAMP]  = { .type = NLA_S32, },
 712         [TCP_METRICS_ATTR_VALS]         = { .type = NLA_NESTED, },
 713         [TCP_METRICS_ATTR_FOPEN_MSS]    = { .type = NLA_U16, },
 714         [TCP_METRICS_ATTR_FOPEN_SYN_DROPS]      = { .type = NLA_U16, },
 715         [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]    = { .type = NLA_MSECS, },
 716         [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
 717                                             .len = TCP_FASTOPEN_COOKIE_MAX, },
 718 #endif
 719 };
 720
 721 /* Add attributes, caller cancels its header on failure */
 722 static int tcp_metrics_fill_info(struct sk_buff *msg,
 723                                  struct tcp_metrics_block *tm)
 724 {
 725         struct nlattr *nest;
 726         int i;
 727
 728         switch (tm->tcpm_addr.family) {
 729         case AF_INET:
 730                 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
 731                                 tm->tcpm_addr.addr.a4) < 0)
 732                         goto nla_put_failure;
 733                 break;
 734         case AF_INET6:
 735                 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
 736                             tm->tcpm_addr.addr.a6) < 0)
 737                         goto nla_put_failure;
 738                 break;
 739         default:
 740                 return -EAFNOSUPPORT;
 741         }
 742
 743         if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
 744                           jiffies - tm->tcpm_stamp) < 0)
 745                 goto nla_put_failure;
 746         if (tm->tcpm_ts_stamp) {
 747                 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
 748                                 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
 749                         goto nla_put_failure;
 750                 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
 751                                 tm->tcpm_ts) < 0)
 752                         goto nla_put_failure;
 753         }
 754
 755         {
 756                 int n = 0;
 757
 758                 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
 759                 if (!nest)
 760                         goto nla_put_failure;
 761                 for (i = 0; i < TCP_METRIC_MAX + 1; i++) {
 762                         if (!tm->tcpm_vals[i])
 763                                 continue;
 764                         if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0)
 765                                 goto nla_put_failure;
 766                         n++;
 767                 }
 768                 if (n)
 769                         nla_nest_end(msg, nest);
 770                 else
 771                         nla_nest_cancel(msg, nest);
 772         }
 773
 774         {
 775                 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
 776                 unsigned int seq;
 777
 778                 do {
 779                         seq = read_seqbegin(&fastopen_seqlock);
 780                         tfom_copy[0] = tm->tcpm_fastopen;
 781                 } while (read_seqretry(&fastopen_seqlock, seq));
 782
 783                 tfom = tfom_copy;
 784                 if (tfom->mss &&
 785                     nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
 786                                 tfom->mss) < 0)
 787                         goto nla_put_failure;
 788                 if (tfom->syn_loss &&
 789                     (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
 790                                 tfom->syn_loss) < 0 ||
 791                      nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
 792                                 jiffies - tfom->last_syn_loss) < 0))
 793                         goto nla_put_failure;
 794                 if (tfom->cookie.len > 0 &&
 795                     nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
 796                             tfom->cookie.len, tfom->cookie.val) < 0)
 797                         goto nla_put_failure;
 798         }
 799
 800         return 0;
 801
 802 nla_put_failure:
 803         return -EMSGSIZE;
 804 }
 805
 806 static int tcp_metrics_dump_info(struct sk_buff *skb,
 807                                  struct netlink_callback *cb,
 808                                  struct tcp_metrics_block *tm)
 809 {
 810         void *hdr;
 811
 812         hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 813                           &tcp_metrics_nl_family, NLM_F_MULTI,
 814                           TCP_METRICS_CMD_GET);
 815         if (!hdr)
 816                 return -EMSGSIZE;
 817
 818         if (tcp_metrics_fill_info(skb, tm) < 0)
 819                 goto nla_put_failure;
 820
 821         return genlmsg_end(skb, hdr);
 822
 823 nla_put_failure:
 824         genlmsg_cancel(skb, hdr);
 825         return -EMSGSIZE;
 826 }
 827
 828 static int tcp_metrics_nl_dump(struct sk_buff *skb,
 829                                struct netlink_callback *cb)
 830 {
 831         struct net *net = sock_net(skb->sk);
 832         unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
 833         unsigned int row, s_row = cb->args[0];
 834         int s_col = cb->args[1], col = s_col;
 835
 836         for (row = s_row; row < max_rows; row++, s_col = 0) {
 837                 struct tcp_metrics_block *tm;
 838                 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
 839
 840                 rcu_read_lock();
 841                 for (col = 0, tm = rcu_dereference(hb->chain); tm;
 842                      tm = rcu_dereference(tm->tcpm_next), col++) {
 843                         if (col < s_col)
 844                                 continue;
 845                         if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
 846                                 rcu_read_unlock();
 847                                 goto done;
 848                         }
 849                 }
 850                 rcu_read_unlock();
 851         }
 852
 853 done:
 854         cb->args[0] = row;
 855         cb->args[1] = col;
 856         return skb->len;
 857 }
 858
 859 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 860                          unsigned int *hash, int optional)
 861 {
 862         struct nlattr *a;
 863
 864         a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV4];
 865         if (a) {
 866                 addr->family = AF_INET;
 867                 addr->addr.a4 = nla_get_be32(a);
 868                 *hash = (__force unsigned int) addr->addr.a4;
 869                 return 0;
 870         }
 871         a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV6];
 872         if (a) {
 873                 if (nla_len(a) != sizeof(struct in6_addr))
 874                         return -EINVAL;
 875                 addr->family = AF_INET6;
 876                 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
 877                 *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
 878                 return 0;
 879         }
 880         return optional ? 1 : -EAFNOSUPPORT;
 881 }
 882
 883 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
 884 {
 885         struct tcp_metrics_block *tm;
 886         struct inetpeer_addr addr;
 887         unsigned int hash;
 888         struct sk_buff *msg;
 889         struct net *net = genl_info_net(info);
 890         void *reply;
 891         int ret;
 892
 893         ret = parse_nl_addr(info, &addr, &hash, 0);
 894         if (ret < 0)
 895                 return ret;
 896
 897         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 898         if (!msg)
 899                 return -ENOMEM;
 900
 901         reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
 902                                   info->genlhdr->cmd);
 903         if (!reply)
 904                 goto nla_put_failure;
 905
 906         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 907         ret = -ESRCH;
 908         rcu_read_lock();
 909         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 910              tm = rcu_dereference(tm->tcpm_next)) {
 911                 if (addr_same(&tm->tcpm_addr, &addr)) {
 912                         ret = tcp_metrics_fill_info(msg, tm);
 913                         break;
 914                 }
 915         }
 916         rcu_read_unlock();
 917         if (ret < 0)
 918                 goto out_free;
 919
 920         genlmsg_end(msg, reply);
 921         return genlmsg_reply(msg, info);
 922
 923 nla_put_failure:
 924         ret = -EMSGSIZE;
 925
 926 out_free:
 927         nlmsg_free(msg);
 928         return ret;
 929 }
 930
 931 #define deref_locked_genl(p)    \
 932         rcu_dereference_protected(p, lockdep_genl_is_held() && \
 933                                      lockdep_is_held(&tcp_metrics_lock))
 934
 935 #define deref_genl(p)   rcu_dereference_protected(p, lockdep_genl_is_held())
 936
 937 static int tcp_metrics_flush_all(struct net *net)
 938 {
 939         unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
 940         struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
 941         struct tcp_metrics_block *tm;
 942         unsigned int row;
 943
 944         for (row = 0; row < max_rows; row++, hb++) {
 945                 spin_lock_bh(&tcp_metrics_lock);
 946                 tm = deref_locked_genl(hb->chain);
 947                 if (tm)
 948                         hb->chain = NULL;
 949                 spin_unlock_bh(&tcp_metrics_lock);
 950                 while (tm) {
 951                         struct tcp_metrics_block *next;
 952
 953                         next = deref_genl(tm->tcpm_next);
 954                         kfree_rcu(tm, rcu_head);
 955                         tm = next;
 956                 }
 957         }
 958         return 0;
 959 }
 960
 961 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
 962 {
 963         struct tcpm_hash_bucket *hb;
 964         struct tcp_metrics_block *tm;
 965         struct tcp_metrics_block __rcu **pp;
 966         struct inetpeer_addr addr;
 967         unsigned int hash;
 968         struct net *net = genl_info_net(info);
 969         int ret;
 970
 971         ret = parse_nl_addr(info, &addr, &hash, 1);
 972         if (ret < 0)
 973                 return ret;
 974         if (ret > 0)
 975                 return tcp_metrics_flush_all(net);
 976
 977         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 978         hb = net->ipv4.tcp_metrics_hash + hash;
 979         pp = &hb->chain;
 980         spin_lock_bh(&tcp_metrics_lock);
 981         for (tm = deref_locked_genl(*pp); tm;
 982              pp = &tm->tcpm_next, tm = deref_locked_genl(*pp)) {
 983                 if (addr_same(&tm->tcpm_addr, &addr)) {
 984                         *pp = tm->tcpm_next;
 985                         break;
 986                 }
 987         }
 988         spin_unlock_bh(&tcp_metrics_lock);
 989         if (!tm)
 990                 return -ESRCH;
 991         kfree_rcu(tm, rcu_head);
 992         return 0;
 993 }
 994
 995 static struct genl_ops tcp_metrics_nl_ops[] = {
 996         {
 997                 .cmd = TCP_METRICS_CMD_GET,
 998                 .doit = tcp_metrics_nl_cmd_get,
 999                 .dumpit = tcp_metrics_nl_dump,
1000                 .policy = tcp_metrics_nl_policy,
1001                 .flags = GENL_ADMIN_PERM,
1002         },
1003         {
1004                 .cmd = TCP_METRICS_CMD_DEL,
1005                 .doit = tcp_metrics_nl_cmd_del,
1006                 .policy = tcp_metrics_nl_policy,
1007                 .flags = GENL_ADMIN_PERM,
1008         },
1009 };
1010
1011 static unsigned int tcpmhash_entries;
1012 static int __init set_tcpmhash_entries(char *str)
1013 {
1014         ssize_t ret;
1015
1016         if (!str)
1017                 return 0;
1018
1019         ret = kstrtouint(str, 0, &tcpmhash_entries);
1020         if (ret)
1021                 return 0;
1022
1023         return 1;
1024 }
1025 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1026
1027 static int __net_init tcp_net_metrics_init(struct net *net)
1028 {
1029         size_t size;
1030         unsigned int slots;
1031
1032         slots = tcpmhash_entries;
1033         if (!slots) {
1034                 if (totalram_pages >= 128 * 1024)
1035                         slots = 16 * 1024;
1036                 else
1037                         slots = 8 * 1024;
1038         }
1039
1040         net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1041         size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
1042
1043         net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1044         if (!net->ipv4.tcp_metrics_hash)
1045                 net->ipv4.tcp_metrics_hash = vzalloc(size);
1046
1047         if (!net->ipv4.tcp_metrics_hash)
1048                 return -ENOMEM;
1049
1050         return 0;
1051 }
1052
1053 static void __net_exit tcp_net_metrics_exit(struct net *net)
1054 {
1055         unsigned int i;
1056
1057         for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1058                 struct tcp_metrics_block *tm, *next;
1059
1060                 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1061                 while (tm) {
1062                         next = rcu_dereference_protected(tm->tcpm_next, 1);
1063                         kfree(tm);
1064                         tm = next;
1065                 }
1066         }
1067         if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1068                 vfree(net->ipv4.tcp_metrics_hash);
1069         else
1070                 kfree(net->ipv4.tcp_metrics_hash);
1071 }
1072
1073 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1074         .init   =       tcp_net_metrics_init,
1075         .exit   =       tcp_net_metrics_exit,
1076 };
1077
1078 void __init tcp_metrics_init(void)
1079 {
1080         int ret;
1081
1082         ret = register_pernet_subsys(&tcp_net_metrics_ops);
1083         if (ret < 0)
1084                 goto cleanup;
1085         ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1086                                             tcp_metrics_nl_ops,
1087                                             ARRAY_SIZE(tcp_metrics_nl_ops));
1088         if (ret < 0)
1089                 goto cleanup_subsys;
1090         return;
1091
1092 cleanup_subsys:
1093         unregister_pernet_subsys(&tcp_net_metrics_ops);
1094
1095 cleanup:
1096         return;
1097 }