net/ipv4/tcp_metrics.c

   1 #include <linux/rcupdate.h>
   2 #include <linux/spinlock.h>
   3 #include <linux/jiffies.h>
   4 #include <linux/bootmem.h>
   5 #include <linux/module.h>
   6 #include <linux/cache.h>
   7 #include <linux/slab.h>
   8 #include <linux/init.h>
   9 #include <linux/tcp.h>
  10
  11 #include <net/inet_connection_sock.h>
  12 #include <net/net_namespace.h>
  13 #include <net/request_sock.h>
  14 #include <net/inetpeer.h>
  15 #include <net/sock.h>
  16 #include <net/ipv6.h>
  17 #include <net/dst.h>
  18 #include <net/tcp.h>
  19
  20 int sysctl_tcp_nometrics_save __read_mostly;
  21
  22 enum tcp_metric_index {
  23         TCP_METRIC_RTT,
  24         TCP_METRIC_RTTVAR,
  25         TCP_METRIC_SSTHRESH,
  26         TCP_METRIC_CWND,
  27         TCP_METRIC_REORDERING,
  28
  29         /* Always last.  */
  30         TCP_METRIC_MAX,
  31 };
  32
  33 struct tcp_metrics_block {
  34         struct tcp_metrics_block __rcu  *tcpm_next;
  35         struct inetpeer_addr            tcpm_addr;
  36         unsigned long                   tcpm_stamp;
  37         u32                             tcpm_ts;
  38         u32                             tcpm_ts_stamp;
  39         u32                             tcpm_lock;
  40         u32                             tcpm_vals[TCP_METRIC_MAX];
  41 };
  42
  43 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
  44                               enum tcp_metric_index idx)
  45 {
  46         return tm->tcpm_lock & (1 << idx);
  47 }
  48
  49 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
  50                           enum tcp_metric_index idx)
  51 {
  52         return tm->tcpm_vals[idx];
  53 }
  54
  55 static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm,
  56                                   enum tcp_metric_index idx)
  57 {
  58         return msecs_to_jiffies(tm->tcpm_vals[idx]);
  59 }
  60
  61 static void tcp_metric_set(struct tcp_metrics_block *tm,
  62                            enum tcp_metric_index idx,
  63                            u32 val)
  64 {
  65         tm->tcpm_vals[idx] = val;
  66 }
  67
  68 static void tcp_metric_set_msecs(struct tcp_metrics_block *tm,
  69                                  enum tcp_metric_index idx,
  70                                  u32 val)
  71 {
  72         tm->tcpm_vals[idx] = jiffies_to_msecs(val);
  73 }
  74
  75 static bool addr_same(const struct inetpeer_addr *a,
  76                       const struct inetpeer_addr *b)
  77 {
  78         const struct in6_addr *a6, *b6;
  79
  80         if (a->family != b->family)
  81                 return false;
  82         if (a->family == AF_INET)
  83                 return a->addr.a4 == b->addr.a4;
  84
  85         a6 = (const struct in6_addr *) &a->addr.a6[0];
  86         b6 = (const struct in6_addr *) &b->addr.a6[0];
  87
  88         return ipv6_addr_equal(a6, b6);
  89 }
  90
  91 struct tcpm_hash_bucket {
  92         struct tcp_metrics_block __rcu  *chain;
  93 };
  94
  95 static DEFINE_SPINLOCK(tcp_metrics_lock);
  96
  97 static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst)
  98 {
  99         u32 val;
 100
 101         val = 0;
 102         if (dst_metric_locked(dst, RTAX_RTT))
 103                 val |= 1 << TCP_METRIC_RTT;
 104         if (dst_metric_locked(dst, RTAX_RTTVAR))
 105                 val |= 1 << TCP_METRIC_RTTVAR;
 106         if (dst_metric_locked(dst, RTAX_SSTHRESH))
 107                 val |= 1 << TCP_METRIC_SSTHRESH;
 108         if (dst_metric_locked(dst, RTAX_CWND))
 109                 val |= 1 << TCP_METRIC_CWND;
 110         if (dst_metric_locked(dst, RTAX_REORDERING))
 111                 val |= 1 << TCP_METRIC_REORDERING;
 112         tm->tcpm_lock = val;
 113
 114         tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT);
 115         tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR);
 116         tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
 117         tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
 118         tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
 119         tm->tcpm_ts = 0;
 120         tm->tcpm_ts_stamp = 0;
 121 }
 122
 123 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
 124                                           struct inetpeer_addr *addr,
 125                                           unsigned int hash,
 126                                           bool reclaim)
 127 {
 128         struct tcp_metrics_block *tm;
 129         struct net *net;
 130
 131         spin_lock_bh(&tcp_metrics_lock);
 132         net = dev_net(dst->dev);
 133         if (unlikely(reclaim)) {
 134                 struct tcp_metrics_block *oldest;
 135
 136                 oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
 137                 for (tm = rcu_dereference(oldest->tcpm_next); tm;
 138                      tm = rcu_dereference(tm->tcpm_next)) {
 139                         if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
 140                                 oldest = tm;
 141                 }
 142                 tm = oldest;
 143         } else {
 144                 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
 145                 if (!tm)
 146                         goto out_unlock;
 147         }
 148         tm->tcpm_addr = *addr;
 149         tm->tcpm_stamp = jiffies;
 150
 151         tcpm_suck_dst(tm, dst);
 152
 153         if (likely(!reclaim)) {
 154                 tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
 155                 rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
 156         }
 157
 158 out_unlock:
 159         spin_unlock_bh(&tcp_metrics_lock);
 160         return tm;
 161 }
 162
 163 #define TCP_METRICS_TIMEOUT             (60 * 60 * HZ)
 164
 165 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
 166 {
 167         if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
 168                 tcpm_suck_dst(tm, dst);
 169 }
 170
 171 #define TCP_METRICS_RECLAIM_DEPTH       5
 172 #define TCP_METRICS_RECLAIM_PTR         (struct tcp_metrics_block *) 0x1UL
 173
 174 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
 175 {
 176         if (tm)
 177                 return tm;
 178         if (depth > TCP_METRICS_RECLAIM_DEPTH)
 179                 return TCP_METRICS_RECLAIM_PTR;
 180         return NULL;
 181 }
 182
 183 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *addr,
 184                                                    struct net *net, unsigned int hash)
 185 {
 186         struct tcp_metrics_block *tm;
 187         int depth = 0;
 188
 189         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 190              tm = rcu_dereference(tm->tcpm_next)) {
 191                 if (addr_same(&tm->tcpm_addr, addr))
 192                         break;
 193                 depth++;
 194         }
 195         return tcp_get_encode(tm, depth);
 196 }
 197
 198 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
 199                                                        struct dst_entry *dst)
 200 {
 201         struct tcp_metrics_block *tm;
 202         struct inetpeer_addr addr;
 203         unsigned int hash;
 204         struct net *net;
 205
 206         addr.family = req->rsk_ops->family;
 207         switch (addr.family) {
 208         case AF_INET:
 209                 addr.addr.a4 = inet_rsk(req)->rmt_addr;
 210                 hash = (__force unsigned int) addr.addr.a4;
 211                 break;
 212         case AF_INET6:
 213                 *(struct in6_addr *)addr.addr.a6 = inet6_rsk(req)->rmt_addr;
 214                 hash = ipv6_addr_hash(&inet6_rsk(req)->rmt_addr);
 215                 break;
 216         default:
 217                 return NULL;
 218         }
 219
 220         hash ^= (hash >> 24) ^ (hash >> 16) ^ (hash >> 8);
 221
 222         net = dev_net(dst->dev);
 223         hash &= net->ipv4.tcp_metrics_hash_mask;
 224
 225         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 226              tm = rcu_dereference(tm->tcpm_next)) {
 227                 if (addr_same(&tm->tcpm_addr, &addr))
 228                         break;
 229         }
 230         tcpm_check_stamp(tm, dst);
 231         return tm;
 232 }
 233
 234 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
 235 {
 236         struct inet6_timewait_sock *tw6;
 237         struct tcp_metrics_block *tm;
 238         struct inetpeer_addr addr;
 239         unsigned int hash;
 240         struct net *net;
 241
 242         addr.family = tw->tw_family;
 243         switch (addr.family) {
 244         case AF_INET:
 245                 addr.addr.a4 = tw->tw_daddr;
 246                 hash = (__force unsigned int) addr.addr.a4;
 247                 break;
 248         case AF_INET6:
 249                 tw6 = inet6_twsk((struct sock *)tw);
 250                 *(struct in6_addr *)addr.addr.a6 = tw6->tw_v6_daddr;
 251                 hash = ipv6_addr_hash(&tw6->tw_v6_daddr);
 252                 break;
 253         default:
 254                 return NULL;
 255         }
 256
 257         hash ^= (hash >> 24) ^ (hash >> 16) ^ (hash >> 8);
 258
 259         net = twsk_net(tw);
 260         hash &= net->ipv4.tcp_metrics_hash_mask;
 261
 262         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 263              tm = rcu_dereference(tm->tcpm_next)) {
 264                 if (addr_same(&tm->tcpm_addr, &addr))
 265                         break;
 266         }
 267         return tm;
 268 }
 269
 270 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
 271                                                  struct dst_entry *dst,
 272                                                  bool create)
 273 {
 274         struct tcp_metrics_block *tm;
 275         struct inetpeer_addr addr;
 276         unsigned int hash;
 277         struct net *net;
 278         bool reclaim;
 279
 280         addr.family = sk->sk_family;
 281         switch (addr.family) {
 282         case AF_INET:
 283                 addr.addr.a4 = inet_sk(sk)->inet_daddr;
 284                 hash = (__force unsigned int) addr.addr.a4;
 285                 break;
 286         case AF_INET6:
 287                 *(struct in6_addr *)addr.addr.a6 = inet6_sk(sk)->daddr;
 288                 hash = ipv6_addr_hash(&inet6_sk(sk)->daddr);
 289                 break;
 290         default:
 291                 return NULL;
 292         }
 293
 294         hash ^= (hash >> 24) ^ (hash >> 16) ^ (hash >> 8);
 295
 296         net = dev_net(dst->dev);
 297         hash &= net->ipv4.tcp_metrics_hash_mask;
 298
 299         tm = __tcp_get_metrics(&addr, net, hash);
 300         reclaim = false;
 301         if (tm == TCP_METRICS_RECLAIM_PTR) {
 302                 reclaim = true;
 303                 tm = NULL;
 304         }
 305         if (!tm && create)
 306                 tm = tcpm_new(dst, &addr, hash, reclaim);
 307         else
 308                 tcpm_check_stamp(tm, dst);
 309
 310         return tm;
 311 }
 312
 313 /* Save metrics learned by this TCP session.  This function is called
 314  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
 315  * or goes from LAST-ACK to CLOSE.
 316  */
 317 void tcp_update_metrics(struct sock *sk)
 318 {
 319         const struct inet_connection_sock *icsk = inet_csk(sk);
 320         struct dst_entry *dst = __sk_dst_get(sk);
 321         struct tcp_sock *tp = tcp_sk(sk);
 322         struct tcp_metrics_block *tm;
 323         unsigned long rtt;
 324         u32 val;
 325         int m;
 326
 327         if (sysctl_tcp_nometrics_save || !dst)
 328                 return;
 329
 330         if (dst->flags & DST_HOST)
 331                 dst_confirm(dst);
 332
 333         rcu_read_lock();
 334         if (icsk->icsk_backoff || !tp->srtt) {
 335                 /* This session failed to estimate rtt. Why?
 336                  * Probably, no packets returned in time.  Reset our
 337                  * results.
 338                  */
 339                 tm = tcp_get_metrics(sk, dst, false);
 340                 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
 341                         tcp_metric_set(tm, TCP_METRIC_RTT, 0);
 342                 goto out_unlock;
 343         } else
 344                 tm = tcp_get_metrics(sk, dst, true);
 345
 346         if (!tm)
 347                 goto out_unlock;
 348
 349         rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
 350         m = rtt - tp->srtt;
 351
 352         /* If newly calculated rtt larger than stored one, store new
 353          * one. Otherwise, use EWMA. Remember, rtt overestimation is
 354          * always better than underestimation.
 355          */
 356         if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
 357                 if (m <= 0)
 358                         rtt = tp->srtt;
 359                 else
 360                         rtt -= (m >> 3);
 361                 tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt);
 362         }
 363
 364         if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
 365                 unsigned long var;
 366
 367                 if (m < 0)
 368                         m = -m;
 369
 370                 /* Scale deviation to rttvar fixed point */
 371                 m >>= 1;
 372                 if (m < tp->mdev)
 373                         m = tp->mdev;
 374
 375                 var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
 376                 if (m >= var)
 377                         var = m;
 378                 else
 379                         var -= (var - m) >> 2;
 380
 381                 tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var);
 382         }
 383
 384         if (tcp_in_initial_slowstart(tp)) {
 385                 /* Slow start still did not finish. */
 386                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 387                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 388                         if (val && (tp->snd_cwnd >> 1) > val)
 389                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 390                                                tp->snd_cwnd >> 1);
 391                 }
 392                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 393                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 394                         if (tp->snd_cwnd > val)
 395                                 tcp_metric_set(tm, TCP_METRIC_CWND,
 396                                                tp->snd_cwnd);
 397                 }
 398         } else if (tp->snd_cwnd > tp->snd_ssthresh &&
 399                    icsk->icsk_ca_state == TCP_CA_Open) {
 400                 /* Cong. avoidance phase, cwnd is reliable. */
 401                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
 402                         tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 403                                        max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
 404                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 405                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 406                         tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
 407                 }
 408         } else {
 409                 /* Else slow start did not finish, cwnd is non-sense,
 410                  * ssthresh may be also invalid.
 411                  */
 412                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 413                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 414                         tcp_metric_set(tm, TCP_METRIC_CWND,
 415                                        (val + tp->snd_ssthresh) >> 1);
 416                 }
 417                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 418                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 419                         if (val && tp->snd_ssthresh > val)
 420                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 421                                                tp->snd_ssthresh);
 422                 }
 423                 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
 424                         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 425                         if (val < tp->reordering &&
 426                             tp->reordering != sysctl_tcp_reordering)
 427                                 tcp_metric_set(tm, TCP_METRIC_REORDERING,
 428                                                tp->reordering);
 429                 }
 430         }
 431         tm->tcpm_stamp = jiffies;
 432 out_unlock:
 433         rcu_read_unlock();
 434 }
 435
 436 /* Initialize metrics on socket. */
 437
 438 void tcp_init_metrics(struct sock *sk)
 439 {
 440         struct dst_entry *dst = __sk_dst_get(sk);
 441         struct tcp_sock *tp = tcp_sk(sk);
 442         struct tcp_metrics_block *tm;
 443         u32 val;
 444
 445         if (dst == NULL)
 446                 goto reset;
 447
 448         dst_confirm(dst);
 449
 450         rcu_read_lock();
 451         tm = tcp_get_metrics(sk, dst, true);
 452         if (!tm) {
 453                 rcu_read_unlock();
 454                 goto reset;
 455         }
 456
 457         if (tcp_metric_locked(tm, TCP_METRIC_CWND))
 458                 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
 459
 460         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 461         if (val) {
 462                 tp->snd_ssthresh = val;
 463                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
 464                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
 465         } else {
 466                 /* ssthresh may have been reduced unnecessarily during.
 467                  * 3WHS. Restore it back to its initial default.
 468                  */
 469                 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 470         }
 471         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 472         if (val && tp->reordering != val) {
 473                 tcp_disable_fack(tp);
 474                 tcp_disable_early_retrans(tp);
 475                 tp->reordering = val;
 476         }
 477
 478         val = tcp_metric_get(tm, TCP_METRIC_RTT);
 479         if (val == 0 || tp->srtt == 0) {
 480                 rcu_read_unlock();
 481                 goto reset;
 482         }
 483         /* Initial rtt is determined from SYN,SYN-ACK.
 484          * The segment is small and rtt may appear much
 485          * less than real one. Use per-dst memory
 486          * to make it more realistic.
 487          *
 488          * A bit of theory. RTT is time passed after "normal" sized packet
 489          * is sent until it is ACKed. In normal circumstances sending small
 490          * packets force peer to delay ACKs and calculation is correct too.
 491          * The algorithm is adaptive and, provided we follow specs, it
 492          * NEVER underestimate RTT. BUT! If peer tries to make some clever
 493          * tricks sort of "quick acks" for time long enough to decrease RTT
 494          * to low value, and then abruptly stops to do it and starts to delay
 495          * ACKs, wait for troubles.
 496          */
 497         val = msecs_to_jiffies(val);
 498         if (val > tp->srtt) {
 499                 tp->srtt = val;
 500                 tp->rtt_seq = tp->snd_nxt;
 501         }
 502         val = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
 503         if (val > tp->mdev) {
 504                 tp->mdev = val;
 505                 tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk));
 506         }
 507         rcu_read_unlock();
 508
 509         tcp_set_rto(sk);
 510 reset:
 511         if (tp->srtt == 0) {
 512                 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
 513                  * 3WHS. This is most likely due to retransmission,
 514                  * including spurious one. Reset the RTO back to 3secs
 515                  * from the more aggressive 1sec to avoid more spurious
 516                  * retransmission.
 517                  */
 518                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
 519                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
 520         }
 521         /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
 522          * retransmitted. In light of RFC6298 more aggressive 1sec
 523          * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
 524          * retransmission has occurred.
 525          */
 526         if (tp->total_retrans > 1)
 527                 tp->snd_cwnd = 1;
 528         else
 529                 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
 530         tp->snd_cwnd_stamp = tcp_time_stamp;
 531 }
 532
 533 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
 534 {
 535         struct tcp_metrics_block *tm;
 536         bool ret;
 537
 538         if (!dst)
 539                 return false;
 540
 541         rcu_read_lock();
 542         tm = __tcp_get_metrics_req(req, dst);
 543         if (paws_check) {
 544                 if (tm &&
 545                     (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
 546                     (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
 547                         ret = false;
 548                 else
 549                         ret = true;
 550         } else {
 551                 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
 552                         ret = true;
 553                 else
 554                         ret = false;
 555         }
 556         rcu_read_unlock();
 557
 558         return ret;
 559 }
 560 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
 561
 562 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
 563 {
 564         struct tcp_metrics_block *tm;
 565
 566         rcu_read_lock();
 567         tm = tcp_get_metrics(sk, dst, true);
 568         if (tm) {
 569                 struct tcp_sock *tp = tcp_sk(sk);
 570
 571                 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
 572                         tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
 573                         tp->rx_opt.ts_recent = tm->tcpm_ts;
 574                 }
 575         }
 576         rcu_read_unlock();
 577 }
 578 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
 579
 580 /* VJ's idea. Save last timestamp seen from this destination and hold
 581  * it at least for normal timewait interval to use for duplicate
 582  * segment detection in subsequent connections, before they enter
 583  * synchronized state.
 584  */
 585 bool tcp_remember_stamp(struct sock *sk)
 586 {
 587         struct dst_entry *dst = __sk_dst_get(sk);
 588         bool ret = false;
 589
 590         if (dst) {
 591                 struct tcp_metrics_block *tm;
 592
 593                 rcu_read_lock();
 594                 tm = tcp_get_metrics(sk, dst, true);
 595                 if (tm) {
 596                         struct tcp_sock *tp = tcp_sk(sk);
 597
 598                         if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
 599                             ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 600                              tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
 601                                 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
 602                                 tm->tcpm_ts = tp->rx_opt.ts_recent;
 603                         }
 604                         ret = true;
 605                 }
 606                 rcu_read_unlock();
 607         }
 608         return ret;
 609 }
 610
 611 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
 612 {
 613         struct tcp_metrics_block *tm;
 614         bool ret = false;
 615
 616         rcu_read_lock();
 617         tm = __tcp_get_metrics_tw(tw);
 618         if (tw) {
 619                 const struct tcp_timewait_sock *tcptw;
 620                 struct sock *sk = (struct sock *) tw;
 621
 622                 tcptw = tcp_twsk(sk);
 623                 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
 624                     ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 625                      tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
 626                         tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
 627                         tm->tcpm_ts        = tcptw->tw_ts_recent;
 628                 }
 629                 ret = true;
 630         }
 631         rcu_read_unlock();
 632
 633         return ret;
 634 }
 635
 636 static unsigned long tcpmhash_entries;
 637 static int __init set_tcpmhash_entries(char *str)
 638 {
 639         ssize_t ret;
 640
 641         if (!str)
 642                 return 0;
 643
 644         ret = kstrtoul(str, 0, &tcpmhash_entries);
 645         if (ret)
 646                 return 0;
 647
 648         return 1;
 649 }
 650 __setup("tcpmhash_entries=", set_tcpmhash_entries);
 651
 652 static int __net_init tcp_net_metrics_init(struct net *net)
 653 {
 654         int slots, size;
 655
 656         slots = tcpmhash_entries;
 657         if (!slots) {
 658                 if (totalram_pages >= 128 * 1024)
 659                         slots = 16 * 1024;
 660                 else
 661                         slots = 8 * 1024;
 662         }
 663
 664         size = slots * sizeof(struct tcpm_hash_bucket);
 665
 666         net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL);
 667         if (!net->ipv4.tcp_metrics_hash)
 668                 return -ENOMEM;
 669
 670         net->ipv4.tcp_metrics_hash_mask = (slots - 1);
 671
 672         return 0;
 673 }
 674
 675 static void __net_exit tcp_net_metrics_exit(struct net *net)
 676 {
 677         kfree(net->ipv4.tcp_metrics_hash);
 678 }
 679
 680 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
 681         .init   =       tcp_net_metrics_init,
 682         .exit   =       tcp_net_metrics_exit,
 683 };
 684
 685 void __init tcp_metrics_init(void)
 686 {
 687         register_pernet_subsys(&tcp_net_metrics_ops);
 688 }