Pileus Git - ~andy/linux/blob - 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_saddr;
  35         struct inetpeer_addr            tcpm_daddr;
  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 + 1];
  41         struct tcp_fastopen_metrics     tcpm_fastopen;
  42
  43         struct rcu_head                 rcu_head;
  44 };
  45
  46 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
  47                               enum tcp_metric_index idx)
  48 {
  49         return tm->tcpm_lock & (1 << idx);
  50 }
  51
  52 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
  53                           enum tcp_metric_index idx)
  54 {
  55         return tm->tcpm_vals[idx];
  56 }
  57
  58 static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm,
  59                                   enum tcp_metric_index idx)
  60 {
  61         return msecs_to_jiffies(tm->tcpm_vals[idx]);
  62 }
  63
  64 static void tcp_metric_set(struct tcp_metrics_block *tm,
  65                            enum tcp_metric_index idx,
  66                            u32 val)
  67 {
  68         tm->tcpm_vals[idx] = val;
  69 }
  70
  71 static void tcp_metric_set_msecs(struct tcp_metrics_block *tm,
  72                                  enum tcp_metric_index idx,
  73                                  u32 val)
  74 {
  75         tm->tcpm_vals[idx] = jiffies_to_msecs(val);
  76 }
  77
  78 static bool addr_same(const struct inetpeer_addr *a,
  79                       const struct inetpeer_addr *b)
  80 {
  81         const struct in6_addr *a6, *b6;
  82
  83         if (a->family != b->family)
  84                 return false;
  85         if (a->family == AF_INET)
  86                 return a->addr.a4 == b->addr.a4;
  87
  88         a6 = (const struct in6_addr *) &a->addr.a6[0];
  89         b6 = (const struct in6_addr *) &b->addr.a6[0];
  90
  91         return ipv6_addr_equal(a6, b6);
  92 }
  93
  94 struct tcpm_hash_bucket {
  95         struct tcp_metrics_block __rcu  *chain;
  96 };
  97
  98 static DEFINE_SPINLOCK(tcp_metrics_lock);
  99
 100 static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst,
 101                           bool fastopen_clear)
 102 {
 103         u32 val;
 104
 105         tm->tcpm_stamp = jiffies;
 106
 107         val = 0;
 108         if (dst_metric_locked(dst, RTAX_RTT))
 109                 val |= 1 << TCP_METRIC_RTT;
 110         if (dst_metric_locked(dst, RTAX_RTTVAR))
 111                 val |= 1 << TCP_METRIC_RTTVAR;
 112         if (dst_metric_locked(dst, RTAX_SSTHRESH))
 113                 val |= 1 << TCP_METRIC_SSTHRESH;
 114         if (dst_metric_locked(dst, RTAX_CWND))
 115                 val |= 1 << TCP_METRIC_CWND;
 116         if (dst_metric_locked(dst, RTAX_REORDERING))
 117                 val |= 1 << TCP_METRIC_REORDERING;
 118         tm->tcpm_lock = val;
 119
 120         tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT);
 121         tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR);
 122         tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
 123         tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
 124         tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
 125         tm->tcpm_ts = 0;
 126         tm->tcpm_ts_stamp = 0;
 127         if (fastopen_clear) {
 128                 tm->tcpm_fastopen.mss = 0;
 129                 tm->tcpm_fastopen.syn_loss = 0;
 130                 tm->tcpm_fastopen.cookie.len = 0;
 131         }
 132 }
 133
 134 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
 135                                           struct inetpeer_addr *saddr,
 136                                           struct inetpeer_addr *daddr,
 137                                           unsigned int hash,
 138                                           bool reclaim)
 139 {
 140         struct tcp_metrics_block *tm;
 141         struct net *net;
 142
 143         spin_lock_bh(&tcp_metrics_lock);
 144         net = dev_net(dst->dev);
 145         if (unlikely(reclaim)) {
 146                 struct tcp_metrics_block *oldest;
 147
 148                 oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
 149                 for (tm = rcu_dereference(oldest->tcpm_next); tm;
 150                      tm = rcu_dereference(tm->tcpm_next)) {
 151                         if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
 152                                 oldest = tm;
 153                 }
 154                 tm = oldest;
 155         } else {
 156                 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
 157                 if (!tm)
 158                         goto out_unlock;
 159         }
 160         tm->tcpm_saddr = *saddr;
 161         tm->tcpm_daddr = *daddr;
 162
 163         tcpm_suck_dst(tm, dst, true);
 164
 165         if (likely(!reclaim)) {
 166                 tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
 167                 rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
 168         }
 169
 170 out_unlock:
 171         spin_unlock_bh(&tcp_metrics_lock);
 172         return tm;
 173 }
 174
 175 #define TCP_METRICS_TIMEOUT             (60 * 60 * HZ)
 176
 177 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
 178 {
 179         if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
 180                 tcpm_suck_dst(tm, dst, false);
 181 }
 182
 183 #define TCP_METRICS_RECLAIM_DEPTH       5
 184 #define TCP_METRICS_RECLAIM_PTR         (struct tcp_metrics_block *) 0x1UL
 185
 186 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
 187 {
 188         if (tm)
 189                 return tm;
 190         if (depth > TCP_METRICS_RECLAIM_DEPTH)
 191                 return TCP_METRICS_RECLAIM_PTR;
 192         return NULL;
 193 }
 194
 195 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
 196                                                    const struct inetpeer_addr *daddr,
 197                                                    struct net *net, unsigned int hash)
 198 {
 199         struct tcp_metrics_block *tm;
 200         int depth = 0;
 201
 202         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 203              tm = rcu_dereference(tm->tcpm_next)) {
 204                 if (addr_same(&tm->tcpm_saddr, saddr) &&
 205                     addr_same(&tm->tcpm_daddr, daddr))
 206                         break;
 207                 depth++;
 208         }
 209         return tcp_get_encode(tm, depth);
 210 }
 211
 212 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
 213                                                        struct dst_entry *dst)
 214 {
 215         struct tcp_metrics_block *tm;
 216         struct inetpeer_addr saddr, daddr;
 217         unsigned int hash;
 218         struct net *net;
 219
 220         saddr.family = req->rsk_ops->family;
 221         daddr.family = req->rsk_ops->family;
 222         switch (daddr.family) {
 223         case AF_INET:
 224                 saddr.addr.a4 = inet_rsk(req)->ir_loc_addr;
 225                 daddr.addr.a4 = inet_rsk(req)->ir_rmt_addr;
 226                 hash = (__force unsigned int) daddr.addr.a4;
 227                 break;
 228 #if IS_ENABLED(CONFIG_IPV6)
 229         case AF_INET6:
 230                 *(struct in6_addr *)saddr.addr.a6 = inet_rsk(req)->ir_v6_loc_addr;
 231                 *(struct in6_addr *)daddr.addr.a6 = inet_rsk(req)->ir_v6_rmt_addr;
 232                 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
 233                 break;
 234 #endif
 235         default:
 236                 return NULL;
 237         }
 238
 239         net = dev_net(dst->dev);
 240         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 241
 242         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 243              tm = rcu_dereference(tm->tcpm_next)) {
 244                 if (addr_same(&tm->tcpm_saddr, &saddr) &&
 245                     addr_same(&tm->tcpm_daddr, &daddr))
 246                         break;
 247         }
 248         tcpm_check_stamp(tm, dst);
 249         return tm;
 250 }
 251
 252 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
 253 {
 254         struct tcp_metrics_block *tm;
 255         struct inetpeer_addr saddr, daddr;
 256         unsigned int hash;
 257         struct net *net;
 258
 259         saddr.family = tw->tw_family;
 260         daddr.family = tw->tw_family;
 261         switch (daddr.family) {
 262         case AF_INET:
 263                 saddr.addr.a4 = tw->tw_rcv_saddr;
 264                 daddr.addr.a4 = tw->tw_daddr;
 265                 hash = (__force unsigned int) daddr.addr.a4;
 266                 break;
 267 #if IS_ENABLED(CONFIG_IPV6)
 268         case AF_INET6:
 269                 *(struct in6_addr *)saddr.addr.a6 = tw->tw_v6_rcv_saddr;
 270                 *(struct in6_addr *)daddr.addr.a6 = tw->tw_v6_daddr;
 271                 hash = ipv6_addr_hash(&tw->tw_v6_daddr);
 272                 break;
 273 #endif
 274         default:
 275                 return NULL;
 276         }
 277
 278         net = twsk_net(tw);
 279         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 280
 281         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 282              tm = rcu_dereference(tm->tcpm_next)) {
 283                 if (addr_same(&tm->tcpm_saddr, &saddr) &&
 284                     addr_same(&tm->tcpm_daddr, &daddr))
 285                         break;
 286         }
 287         return tm;
 288 }
 289
 290 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
 291                                                  struct dst_entry *dst,
 292                                                  bool create)
 293 {
 294         struct tcp_metrics_block *tm;
 295         struct inetpeer_addr saddr, daddr;
 296         unsigned int hash;
 297         struct net *net;
 298         bool reclaim;
 299
 300         saddr.family = sk->sk_family;
 301         daddr.family = sk->sk_family;
 302         switch (daddr.family) {
 303         case AF_INET:
 304                 saddr.addr.a4 = inet_sk(sk)->inet_saddr;
 305                 daddr.addr.a4 = inet_sk(sk)->inet_daddr;
 306                 hash = (__force unsigned int) daddr.addr.a4;
 307                 break;
 308 #if IS_ENABLED(CONFIG_IPV6)
 309         case AF_INET6:
 310                 *(struct in6_addr *)saddr.addr.a6 = sk->sk_v6_rcv_saddr;
 311                 *(struct in6_addr *)daddr.addr.a6 = sk->sk_v6_daddr;
 312                 hash = ipv6_addr_hash(&sk->sk_v6_daddr);
 313                 break;
 314 #endif
 315         default:
 316                 return NULL;
 317         }
 318
 319         net = dev_net(dst->dev);
 320         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 321
 322         tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
 323         reclaim = false;
 324         if (tm == TCP_METRICS_RECLAIM_PTR) {
 325                 reclaim = true;
 326                 tm = NULL;
 327         }
 328         if (!tm && create)
 329                 tm = tcpm_new(dst, &saddr, &daddr, hash, reclaim);
 330         else
 331                 tcpm_check_stamp(tm, dst);
 332
 333         return tm;
 334 }
 335
 336 /* Save metrics learned by this TCP session.  This function is called
 337  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
 338  * or goes from LAST-ACK to CLOSE.
 339  */
 340 void tcp_update_metrics(struct sock *sk)
 341 {
 342         const struct inet_connection_sock *icsk = inet_csk(sk);
 343         struct dst_entry *dst = __sk_dst_get(sk);
 344         struct tcp_sock *tp = tcp_sk(sk);
 345         struct tcp_metrics_block *tm;
 346         unsigned long rtt;
 347         u32 val;
 348         int m;
 349
 350         if (sysctl_tcp_nometrics_save || !dst)
 351                 return;
 352
 353         if (dst->flags & DST_HOST)
 354                 dst_confirm(dst);
 355
 356         rcu_read_lock();
 357         if (icsk->icsk_backoff || !tp->srtt) {
 358                 /* This session failed to estimate rtt. Why?
 359                  * Probably, no packets returned in time.  Reset our
 360                  * results.
 361                  */
 362                 tm = tcp_get_metrics(sk, dst, false);
 363                 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
 364                         tcp_metric_set(tm, TCP_METRIC_RTT, 0);
 365                 goto out_unlock;
 366         } else
 367                 tm = tcp_get_metrics(sk, dst, true);
 368
 369         if (!tm)
 370                 goto out_unlock;
 371
 372         rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
 373         m = rtt - tp->srtt;
 374
 375         /* If newly calculated rtt larger than stored one, store new
 376          * one. Otherwise, use EWMA. Remember, rtt overestimation is
 377          * always better than underestimation.
 378          */
 379         if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
 380                 if (m <= 0)
 381                         rtt = tp->srtt;
 382                 else
 383                         rtt -= (m >> 3);
 384                 tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt);
 385         }
 386
 387         if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
 388                 unsigned long var;
 389
 390                 if (m < 0)
 391                         m = -m;
 392
 393                 /* Scale deviation to rttvar fixed point */
 394                 m >>= 1;
 395                 if (m < tp->mdev)
 396                         m = tp->mdev;
 397
 398                 var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
 399                 if (m >= var)
 400                         var = m;
 401                 else
 402                         var -= (var - m) >> 2;
 403
 404                 tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var);
 405         }
 406
 407         if (tcp_in_initial_slowstart(tp)) {
 408                 /* Slow start still did not finish. */
 409                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 410                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 411                         if (val && (tp->snd_cwnd >> 1) > val)
 412                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 413                                                tp->snd_cwnd >> 1);
 414                 }
 415                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 416                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 417                         if (tp->snd_cwnd > val)
 418                                 tcp_metric_set(tm, TCP_METRIC_CWND,
 419                                                tp->snd_cwnd);
 420                 }
 421         } else if (tp->snd_cwnd > tp->snd_ssthresh &&
 422                    icsk->icsk_ca_state == TCP_CA_Open) {
 423                 /* Cong. avoidance phase, cwnd is reliable. */
 424                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
 425                         tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 426                                        max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
 427                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 428                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 429                         tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
 430                 }
 431         } else {
 432                 /* Else slow start did not finish, cwnd is non-sense,
 433                  * ssthresh may be also invalid.
 434                  */
 435                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 436                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 437                         tcp_metric_set(tm, TCP_METRIC_CWND,
 438                                        (val + tp->snd_ssthresh) >> 1);
 439                 }
 440                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 441                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 442                         if (val && tp->snd_ssthresh > val)
 443                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 444                                                tp->snd_ssthresh);
 445                 }
 446                 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
 447                         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 448                         if (val < tp->reordering &&
 449                             tp->reordering != sysctl_tcp_reordering)
 450                                 tcp_metric_set(tm, TCP_METRIC_REORDERING,
 451                                                tp->reordering);
 452                 }
 453         }
 454         tm->tcpm_stamp = jiffies;
 455 out_unlock:
 456         rcu_read_unlock();
 457 }
 458
 459 /* Initialize metrics on socket. */
 460
 461 void tcp_init_metrics(struct sock *sk)
 462 {
 463         struct dst_entry *dst = __sk_dst_get(sk);
 464         struct tcp_sock *tp = tcp_sk(sk);
 465         struct tcp_metrics_block *tm;
 466         u32 val, crtt = 0; /* cached RTT scaled by 8 */
 467
 468         if (dst == NULL)
 469                 goto reset;
 470
 471         dst_confirm(dst);
 472
 473         rcu_read_lock();
 474         tm = tcp_get_metrics(sk, dst, true);
 475         if (!tm) {
 476                 rcu_read_unlock();
 477                 goto reset;
 478         }
 479
 480         if (tcp_metric_locked(tm, TCP_METRIC_CWND))
 481                 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
 482
 483         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 484         if (val) {
 485                 tp->snd_ssthresh = val;
 486                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
 487                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
 488         } else {
 489                 /* ssthresh may have been reduced unnecessarily during.
 490                  * 3WHS. Restore it back to its initial default.
 491                  */
 492                 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 493         }
 494         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 495         if (val && tp->reordering != val) {
 496                 tcp_disable_fack(tp);
 497                 tcp_disable_early_retrans(tp);
 498                 tp->reordering = val;
 499         }
 500
 501         crtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
 502         rcu_read_unlock();
 503 reset:
 504         /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
 505          * to seed the RTO for later data packets because SYN packets are
 506          * small. Use the per-dst cached values to seed the RTO but keep
 507          * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
 508          * Later the RTO will be updated immediately upon obtaining the first
 509          * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
 510          * influences the first RTO but not later RTT estimation.
 511          *
 512          * But if RTT is not available from the SYN (due to retransmits or
 513          * syn cookies) or the cache, force a conservative 3secs timeout.
 514          *
 515          * A bit of theory. RTT is time passed after "normal" sized packet
 516          * is sent until it is ACKed. In normal circumstances sending small
 517          * packets force peer to delay ACKs and calculation is correct too.
 518          * The algorithm is adaptive and, provided we follow specs, it
 519          * NEVER underestimate RTT. BUT! If peer tries to make some clever
 520          * tricks sort of "quick acks" for time long enough to decrease RTT
 521          * to low value, and then abruptly stops to do it and starts to delay
 522          * ACKs, wait for troubles.
 523          */
 524         if (crtt > tp->srtt) {
 525                 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
 526                 crtt >>= 3;
 527                 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
 528         } else if (tp->srtt == 0) {
 529                 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
 530                  * 3WHS. This is most likely due to retransmission,
 531                  * including spurious one. Reset the RTO back to 3secs
 532                  * from the more aggressive 1sec to avoid more spurious
 533                  * retransmission.
 534                  */
 535                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
 536                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
 537         }
 538         /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
 539          * retransmitted. In light of RFC6298 more aggressive 1sec
 540          * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
 541          * retransmission has occurred.
 542          */
 543         if (tp->total_retrans > 1)
 544                 tp->snd_cwnd = 1;
 545         else
 546                 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
 547         tp->snd_cwnd_stamp = tcp_time_stamp;
 548 }
 549
 550 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
 551 {
 552         struct tcp_metrics_block *tm;
 553         bool ret;
 554
 555         if (!dst)
 556                 return false;
 557
 558         rcu_read_lock();
 559         tm = __tcp_get_metrics_req(req, dst);
 560         if (paws_check) {
 561                 if (tm &&
 562                     (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
 563                     (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
 564                         ret = false;
 565                 else
 566                         ret = true;
 567         } else {
 568                 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
 569                         ret = true;
 570                 else
 571                         ret = false;
 572         }
 573         rcu_read_unlock();
 574
 575         return ret;
 576 }
 577 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
 578
 579 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
 580 {
 581         struct tcp_metrics_block *tm;
 582
 583         rcu_read_lock();
 584         tm = tcp_get_metrics(sk, dst, true);
 585         if (tm) {
 586                 struct tcp_sock *tp = tcp_sk(sk);
 587
 588                 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
 589                         tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
 590                         tp->rx_opt.ts_recent = tm->tcpm_ts;
 591                 }
 592         }
 593         rcu_read_unlock();
 594 }
 595 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
 596
 597 /* VJ's idea. Save last timestamp seen from this destination and hold
 598  * it at least for normal timewait interval to use for duplicate
 599  * segment detection in subsequent connections, before they enter
 600  * synchronized state.
 601  */
 602 bool tcp_remember_stamp(struct sock *sk)
 603 {
 604         struct dst_entry *dst = __sk_dst_get(sk);
 605         bool ret = false;
 606
 607         if (dst) {
 608                 struct tcp_metrics_block *tm;
 609
 610                 rcu_read_lock();
 611                 tm = tcp_get_metrics(sk, dst, true);
 612                 if (tm) {
 613                         struct tcp_sock *tp = tcp_sk(sk);
 614
 615                         if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
 616                             ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 617                              tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
 618                                 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
 619                                 tm->tcpm_ts = tp->rx_opt.ts_recent;
 620                         }
 621                         ret = true;
 622                 }
 623                 rcu_read_unlock();
 624         }
 625         return ret;
 626 }
 627
 628 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
 629 {
 630         struct tcp_metrics_block *tm;
 631         bool ret = false;
 632
 633         rcu_read_lock();
 634         tm = __tcp_get_metrics_tw(tw);
 635         if (tm) {
 636                 const struct tcp_timewait_sock *tcptw;
 637                 struct sock *sk = (struct sock *) tw;
 638
 639                 tcptw = tcp_twsk(sk);
 640                 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
 641                     ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 642                      tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
 643                         tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
 644                         tm->tcpm_ts        = tcptw->tw_ts_recent;
 645                 }
 646                 ret = true;
 647         }
 648         rcu_read_unlock();
 649
 650         return ret;
 651 }
 652
 653 static DEFINE_SEQLOCK(fastopen_seqlock);
 654
 655 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
 656                             struct tcp_fastopen_cookie *cookie,
 657                             int *syn_loss, unsigned long *last_syn_loss)
 658 {
 659         struct tcp_metrics_block *tm;
 660
 661         rcu_read_lock();
 662         tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
 663         if (tm) {
 664                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 665                 unsigned int seq;
 666
 667                 do {
 668                         seq = read_seqbegin(&fastopen_seqlock);
 669                         if (tfom->mss)
 670                                 *mss = tfom->mss;
 671                         *cookie = tfom->cookie;
 672                         *syn_loss = tfom->syn_loss;
 673                         *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
 674                 } while (read_seqretry(&fastopen_seqlock, seq));
 675         }
 676         rcu_read_unlock();
 677 }
 678
 679 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
 680                             struct tcp_fastopen_cookie *cookie, bool syn_lost)
 681 {
 682         struct dst_entry *dst = __sk_dst_get(sk);
 683         struct tcp_metrics_block *tm;
 684
 685         if (!dst)
 686                 return;
 687         rcu_read_lock();
 688         tm = tcp_get_metrics(sk, dst, true);
 689         if (tm) {
 690                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 691
 692                 write_seqlock_bh(&fastopen_seqlock);
 693                 if (mss)
 694                         tfom->mss = mss;
 695                 if (cookie && cookie->len > 0)
 696                         tfom->cookie = *cookie;
 697                 if (syn_lost) {
 698                         ++tfom->syn_loss;
 699                         tfom->last_syn_loss = jiffies;
 700                 } else
 701                         tfom->syn_loss = 0;
 702                 write_sequnlock_bh(&fastopen_seqlock);
 703         }
 704         rcu_read_unlock();
 705 }
 706
 707 static struct genl_family tcp_metrics_nl_family = {
 708         .id             = GENL_ID_GENERATE,
 709         .hdrsize        = 0,
 710         .name           = TCP_METRICS_GENL_NAME,
 711         .version        = TCP_METRICS_GENL_VERSION,
 712         .maxattr        = TCP_METRICS_ATTR_MAX,
 713         .netnsok        = true,
 714 };
 715
 716 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
 717         [TCP_METRICS_ATTR_ADDR_IPV4]    = { .type = NLA_U32, },
 718         [TCP_METRICS_ATTR_ADDR_IPV6]    = { .type = NLA_BINARY,
 719                                             .len = sizeof(struct in6_addr), },
 720         /* Following attributes are not received for GET/DEL,
 721          * we keep them for reference
 722          */
 723 #if 0
 724         [TCP_METRICS_ATTR_AGE]          = { .type = NLA_MSECS, },
 725         [TCP_METRICS_ATTR_TW_TSVAL]     = { .type = NLA_U32, },
 726         [TCP_METRICS_ATTR_TW_TS_STAMP]  = { .type = NLA_S32, },
 727         [TCP_METRICS_ATTR_VALS]         = { .type = NLA_NESTED, },
 728         [TCP_METRICS_ATTR_FOPEN_MSS]    = { .type = NLA_U16, },
 729         [TCP_METRICS_ATTR_FOPEN_SYN_DROPS]      = { .type = NLA_U16, },
 730         [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]    = { .type = NLA_MSECS, },
 731         [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
 732                                             .len = TCP_FASTOPEN_COOKIE_MAX, },
 733 #endif
 734 };
 735
 736 /* Add attributes, caller cancels its header on failure */
 737 static int tcp_metrics_fill_info(struct sk_buff *msg,
 738                                  struct tcp_metrics_block *tm)
 739 {
 740         struct nlattr *nest;
 741         int i;
 742
 743         switch (tm->tcpm_daddr.family) {
 744         case AF_INET:
 745                 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
 746                                 tm->tcpm_daddr.addr.a4) < 0)
 747                         goto nla_put_failure;
 748                 if (nla_put_be32(msg, TCP_METRICS_ATTR_SADDR_IPV4,
 749                                 tm->tcpm_saddr.addr.a4) < 0)
 750                         goto nla_put_failure;
 751                 break;
 752         case AF_INET6:
 753                 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
 754                             tm->tcpm_daddr.addr.a6) < 0)
 755                         goto nla_put_failure;
 756                 if (nla_put(msg, TCP_METRICS_ATTR_SADDR_IPV6, 16,
 757                             tm->tcpm_saddr.addr.a6) < 0)
 758                         goto nla_put_failure;
 759                 break;
 760         default:
 761                 return -EAFNOSUPPORT;
 762         }
 763
 764         if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
 765                           jiffies - tm->tcpm_stamp) < 0)
 766                 goto nla_put_failure;
 767         if (tm->tcpm_ts_stamp) {
 768                 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
 769                                 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
 770                         goto nla_put_failure;
 771                 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
 772                                 tm->tcpm_ts) < 0)
 773                         goto nla_put_failure;
 774         }
 775
 776         {
 777                 int n = 0;
 778
 779                 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
 780                 if (!nest)
 781                         goto nla_put_failure;
 782                 for (i = 0; i < TCP_METRIC_MAX + 1; i++) {
 783                         if (!tm->tcpm_vals[i])
 784                                 continue;
 785                         if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0)
 786                                 goto nla_put_failure;
 787                         n++;
 788                 }
 789                 if (n)
 790                         nla_nest_end(msg, nest);
 791                 else
 792                         nla_nest_cancel(msg, nest);
 793         }
 794
 795         {
 796                 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
 797                 unsigned int seq;
 798
 799                 do {
 800                         seq = read_seqbegin(&fastopen_seqlock);
 801                         tfom_copy[0] = tm->tcpm_fastopen;
 802                 } while (read_seqretry(&fastopen_seqlock, seq));
 803
 804                 tfom = tfom_copy;
 805                 if (tfom->mss &&
 806                     nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
 807                                 tfom->mss) < 0)
 808                         goto nla_put_failure;
 809                 if (tfom->syn_loss &&
 810                     (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
 811                                 tfom->syn_loss) < 0 ||
 812                      nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
 813                                 jiffies - tfom->last_syn_loss) < 0))
 814                         goto nla_put_failure;
 815                 if (tfom->cookie.len > 0 &&
 816                     nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
 817                             tfom->cookie.len, tfom->cookie.val) < 0)
 818                         goto nla_put_failure;
 819         }
 820
 821         return 0;
 822
 823 nla_put_failure:
 824         return -EMSGSIZE;
 825 }
 826
 827 static int tcp_metrics_dump_info(struct sk_buff *skb,
 828                                  struct netlink_callback *cb,
 829                                  struct tcp_metrics_block *tm)
 830 {
 831         void *hdr;
 832
 833         hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 834                           &tcp_metrics_nl_family, NLM_F_MULTI,
 835                           TCP_METRICS_CMD_GET);
 836         if (!hdr)
 837                 return -EMSGSIZE;
 838
 839         if (tcp_metrics_fill_info(skb, tm) < 0)
 840                 goto nla_put_failure;
 841
 842         return genlmsg_end(skb, hdr);
 843
 844 nla_put_failure:
 845         genlmsg_cancel(skb, hdr);
 846         return -EMSGSIZE;
 847 }
 848
 849 static int tcp_metrics_nl_dump(struct sk_buff *skb,
 850                                struct netlink_callback *cb)
 851 {
 852         struct net *net = sock_net(skb->sk);
 853         unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
 854         unsigned int row, s_row = cb->args[0];
 855         int s_col = cb->args[1], col = s_col;
 856
 857         for (row = s_row; row < max_rows; row++, s_col = 0) {
 858                 struct tcp_metrics_block *tm;
 859                 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
 860
 861                 rcu_read_lock();
 862                 for (col = 0, tm = rcu_dereference(hb->chain); tm;
 863                      tm = rcu_dereference(tm->tcpm_next), col++) {
 864                         if (col < s_col)
 865                                 continue;
 866                         if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
 867                                 rcu_read_unlock();
 868                                 goto done;
 869                         }
 870                 }
 871                 rcu_read_unlock();
 872         }
 873
 874 done:
 875         cb->args[0] = row;
 876         cb->args[1] = col;
 877         return skb->len;
 878 }
 879
 880 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 881                          unsigned int *hash, int optional)
 882 {
 883         struct nlattr *a;
 884
 885         a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV4];
 886         if (a) {
 887                 addr->family = AF_INET;
 888                 addr->addr.a4 = nla_get_be32(a);
 889                 *hash = (__force unsigned int) addr->addr.a4;
 890                 return 0;
 891         }
 892         a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV6];
 893         if (a) {
 894                 if (nla_len(a) != sizeof(struct in6_addr))
 895                         return -EINVAL;
 896                 addr->family = AF_INET6;
 897                 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
 898                 *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
 899                 return 0;
 900         }
 901         return optional ? 1 : -EAFNOSUPPORT;
 902 }
 903
 904 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
 905 {
 906         struct tcp_metrics_block *tm;
 907         struct inetpeer_addr daddr;
 908         unsigned int hash;
 909         struct sk_buff *msg;
 910         struct net *net = genl_info_net(info);
 911         void *reply;
 912         int ret;
 913
 914         ret = parse_nl_addr(info, &daddr, &hash, 0);
 915         if (ret < 0)
 916                 return ret;
 917
 918         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 919         if (!msg)
 920                 return -ENOMEM;
 921
 922         reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
 923                                   info->genlhdr->cmd);
 924         if (!reply)
 925                 goto nla_put_failure;
 926
 927         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 928         ret = -ESRCH;
 929         rcu_read_lock();
 930         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 931              tm = rcu_dereference(tm->tcpm_next)) {
 932                 if (addr_same(&tm->tcpm_daddr, &daddr)) {
 933                         ret = tcp_metrics_fill_info(msg, tm);
 934                         break;
 935                 }
 936         }
 937         rcu_read_unlock();
 938         if (ret < 0)
 939                 goto out_free;
 940
 941         genlmsg_end(msg, reply);
 942         return genlmsg_reply(msg, info);
 943
 944 nla_put_failure:
 945         ret = -EMSGSIZE;
 946
 947 out_free:
 948         nlmsg_free(msg);
 949         return ret;
 950 }
 951
 952 #define deref_locked_genl(p)    \
 953         rcu_dereference_protected(p, lockdep_genl_is_held() && \
 954                                      lockdep_is_held(&tcp_metrics_lock))
 955
 956 #define deref_genl(p)   rcu_dereference_protected(p, lockdep_genl_is_held())
 957
 958 static int tcp_metrics_flush_all(struct net *net)
 959 {
 960         unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
 961         struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
 962         struct tcp_metrics_block *tm;
 963         unsigned int row;
 964
 965         for (row = 0; row < max_rows; row++, hb++) {
 966                 spin_lock_bh(&tcp_metrics_lock);
 967                 tm = deref_locked_genl(hb->chain);
 968                 if (tm)
 969                         hb->chain = NULL;
 970                 spin_unlock_bh(&tcp_metrics_lock);
 971                 while (tm) {
 972                         struct tcp_metrics_block *next;
 973
 974                         next = deref_genl(tm->tcpm_next);
 975                         kfree_rcu(tm, rcu_head);
 976                         tm = next;
 977                 }
 978         }
 979         return 0;
 980 }
 981
 982 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
 983 {
 984         struct tcpm_hash_bucket *hb;
 985         struct tcp_metrics_block *tm;
 986         struct tcp_metrics_block __rcu **pp;
 987         struct inetpeer_addr daddr;
 988         unsigned int hash;
 989         struct net *net = genl_info_net(info);
 990         int ret;
 991
 992         ret = parse_nl_addr(info, &daddr, &hash, 1);
 993         if (ret < 0)
 994                 return ret;
 995         if (ret > 0)
 996                 return tcp_metrics_flush_all(net);
 997
 998         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 999         hb = net->ipv4.tcp_metrics_hash + hash;
1000         pp = &hb->chain;
1001         spin_lock_bh(&tcp_metrics_lock);
1002         for (tm = deref_locked_genl(*pp); tm;
1003              pp = &tm->tcpm_next, tm = deref_locked_genl(*pp)) {
1004                 if (addr_same(&tm->tcpm_daddr, &daddr)) {
1005                         *pp = tm->tcpm_next;
1006                         break;
1007                 }
1008         }
1009         spin_unlock_bh(&tcp_metrics_lock);
1010         if (!tm)
1011                 return -ESRCH;
1012         kfree_rcu(tm, rcu_head);
1013         return 0;
1014 }
1015
1016 static const struct genl_ops tcp_metrics_nl_ops[] = {
1017         {
1018                 .cmd = TCP_METRICS_CMD_GET,
1019                 .doit = tcp_metrics_nl_cmd_get,
1020                 .dumpit = tcp_metrics_nl_dump,
1021                 .policy = tcp_metrics_nl_policy,
1022                 .flags = GENL_ADMIN_PERM,
1023         },
1024         {
1025                 .cmd = TCP_METRICS_CMD_DEL,
1026                 .doit = tcp_metrics_nl_cmd_del,
1027                 .policy = tcp_metrics_nl_policy,
1028                 .flags = GENL_ADMIN_PERM,
1029         },
1030 };
1031
1032 static unsigned int tcpmhash_entries;
1033 static int __init set_tcpmhash_entries(char *str)
1034 {
1035         ssize_t ret;
1036
1037         if (!str)
1038                 return 0;
1039
1040         ret = kstrtouint(str, 0, &tcpmhash_entries);
1041         if (ret)
1042                 return 0;
1043
1044         return 1;
1045 }
1046 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1047
1048 static int __net_init tcp_net_metrics_init(struct net *net)
1049 {
1050         size_t size;
1051         unsigned int slots;
1052
1053         slots = tcpmhash_entries;
1054         if (!slots) {
1055                 if (totalram_pages >= 128 * 1024)
1056                         slots = 16 * 1024;
1057                 else
1058                         slots = 8 * 1024;
1059         }
1060
1061         net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1062         size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
1063
1064         net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1065         if (!net->ipv4.tcp_metrics_hash)
1066                 net->ipv4.tcp_metrics_hash = vzalloc(size);
1067
1068         if (!net->ipv4.tcp_metrics_hash)
1069                 return -ENOMEM;
1070
1071         return 0;
1072 }
1073
1074 static void __net_exit tcp_net_metrics_exit(struct net *net)
1075 {
1076         unsigned int i;
1077
1078         for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1079                 struct tcp_metrics_block *tm, *next;
1080
1081                 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1082                 while (tm) {
1083                         next = rcu_dereference_protected(tm->tcpm_next, 1);
1084                         kfree(tm);
1085                         tm = next;
1086                 }
1087         }
1088         if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1089                 vfree(net->ipv4.tcp_metrics_hash);
1090         else
1091                 kfree(net->ipv4.tcp_metrics_hash);
1092 }
1093
1094 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1095         .init   =       tcp_net_metrics_init,
1096         .exit   =       tcp_net_metrics_exit,
1097 };
1098
1099 void __init tcp_metrics_init(void)
1100 {
1101         int ret;
1102
1103         ret = register_pernet_subsys(&tcp_net_metrics_ops);
1104         if (ret < 0)
1105                 goto cleanup;
1106         ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1107                                             tcp_metrics_nl_ops);
1108         if (ret < 0)
1109                 goto cleanup_subsys;
1110         return;
1111
1112 cleanup_subsys:
1113         unregister_pernet_subsys(&tcp_net_metrics_ops);
1114
1115 cleanup:
1116         return;
1117 }