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 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
  26                                                    const struct inetpeer_addr *daddr,
  27                                                    struct net *net, unsigned int hash);
  28
  29 struct tcp_fastopen_metrics {
  30         u16     mss;
  31         u16     syn_loss:10;            /* Recurring Fast Open SYN losses */
  32         unsigned long   last_syn_loss;  /* Last Fast Open SYN loss */
  33         struct  tcp_fastopen_cookie     cookie;
  34 };
  35
  36 struct tcp_metrics_block {
  37         struct tcp_metrics_block __rcu  *tcpm_next;
  38         struct inetpeer_addr            tcpm_saddr;
  39         struct inetpeer_addr            tcpm_daddr;
  40         unsigned long                   tcpm_stamp;
  41         u32                             tcpm_ts;
  42         u32                             tcpm_ts_stamp;
  43         u32                             tcpm_lock;
  44         u32                             tcpm_vals[TCP_METRIC_MAX + 1];
  45         struct tcp_fastopen_metrics     tcpm_fastopen;
  46
  47         struct rcu_head                 rcu_head;
  48 };
  49
  50 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
  51                               enum tcp_metric_index idx)
  52 {
  53         return tm->tcpm_lock & (1 << idx);
  54 }
  55
  56 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
  57                           enum tcp_metric_index idx)
  58 {
  59         return tm->tcpm_vals[idx];
  60 }
  61
  62 static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm,
  63                                   enum tcp_metric_index idx)
  64 {
  65         return msecs_to_jiffies(tm->tcpm_vals[idx]);
  66 }
  67
  68 static void tcp_metric_set(struct tcp_metrics_block *tm,
  69                            enum tcp_metric_index idx,
  70                            u32 val)
  71 {
  72         tm->tcpm_vals[idx] = val;
  73 }
  74
  75 static void tcp_metric_set_msecs(struct tcp_metrics_block *tm,
  76                                  enum tcp_metric_index idx,
  77                                  u32 val)
  78 {
  79         tm->tcpm_vals[idx] = jiffies_to_msecs(val);
  80 }
  81
  82 static bool addr_same(const struct inetpeer_addr *a,
  83                       const struct inetpeer_addr *b)
  84 {
  85         const struct in6_addr *a6, *b6;
  86
  87         if (a->family != b->family)
  88                 return false;
  89         if (a->family == AF_INET)
  90                 return a->addr.a4 == b->addr.a4;
  91
  92         a6 = (const struct in6_addr *) &a->addr.a6[0];
  93         b6 = (const struct in6_addr *) &b->addr.a6[0];
  94
  95         return ipv6_addr_equal(a6, b6);
  96 }
  97
  98 struct tcpm_hash_bucket {
  99         struct tcp_metrics_block __rcu  *chain;
 100 };
 101
 102 static DEFINE_SPINLOCK(tcp_metrics_lock);
 103
 104 static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst,
 105                           bool fastopen_clear)
 106 {
 107         u32 val;
 108
 109         tm->tcpm_stamp = jiffies;
 110
 111         val = 0;
 112         if (dst_metric_locked(dst, RTAX_RTT))
 113                 val |= 1 << TCP_METRIC_RTT;
 114         if (dst_metric_locked(dst, RTAX_RTTVAR))
 115                 val |= 1 << TCP_METRIC_RTTVAR;
 116         if (dst_metric_locked(dst, RTAX_SSTHRESH))
 117                 val |= 1 << TCP_METRIC_SSTHRESH;
 118         if (dst_metric_locked(dst, RTAX_CWND))
 119                 val |= 1 << TCP_METRIC_CWND;
 120         if (dst_metric_locked(dst, RTAX_REORDERING))
 121                 val |= 1 << TCP_METRIC_REORDERING;
 122         tm->tcpm_lock = val;
 123
 124         tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT);
 125         tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR);
 126         tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
 127         tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
 128         tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
 129         tm->tcpm_ts = 0;
 130         tm->tcpm_ts_stamp = 0;
 131         if (fastopen_clear) {
 132                 tm->tcpm_fastopen.mss = 0;
 133                 tm->tcpm_fastopen.syn_loss = 0;
 134                 tm->tcpm_fastopen.cookie.len = 0;
 135         }
 136 }
 137
 138 #define TCP_METRICS_TIMEOUT             (60 * 60 * HZ)
 139
 140 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
 141 {
 142         if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
 143                 tcpm_suck_dst(tm, dst, false);
 144 }
 145
 146 #define TCP_METRICS_RECLAIM_DEPTH       5
 147 #define TCP_METRICS_RECLAIM_PTR         (struct tcp_metrics_block *) 0x1UL
 148
 149 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
 150                                           struct inetpeer_addr *saddr,
 151                                           struct inetpeer_addr *daddr,
 152                                           unsigned int hash)
 153 {
 154         struct tcp_metrics_block *tm;
 155         struct net *net;
 156         bool reclaim = false;
 157
 158         spin_lock_bh(&tcp_metrics_lock);
 159         net = dev_net(dst->dev);
 160
 161         /* While waiting for the spin-lock the cache might have been populated
 162          * with this entry and so we have to check again.
 163          */
 164         tm = __tcp_get_metrics(saddr, daddr, net, hash);
 165         if (tm == TCP_METRICS_RECLAIM_PTR) {
 166                 reclaim = true;
 167                 tm = NULL;
 168         }
 169         if (tm) {
 170                 tcpm_check_stamp(tm, dst);
 171                 goto out_unlock;
 172         }
 173
 174         if (unlikely(reclaim)) {
 175                 struct tcp_metrics_block *oldest;
 176
 177                 oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
 178                 for (tm = rcu_dereference(oldest->tcpm_next); tm;
 179                      tm = rcu_dereference(tm->tcpm_next)) {
 180                         if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
 181                                 oldest = tm;
 182                 }
 183                 tm = oldest;
 184         } else {
 185                 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
 186                 if (!tm)
 187                         goto out_unlock;
 188         }
 189         tm->tcpm_saddr = *saddr;
 190         tm->tcpm_daddr = *daddr;
 191
 192         tcpm_suck_dst(tm, dst, true);
 193
 194         if (likely(!reclaim)) {
 195                 tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
 196                 rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
 197         }
 198
 199 out_unlock:
 200         spin_unlock_bh(&tcp_metrics_lock);
 201         return tm;
 202 }
 203
 204 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
 205 {
 206         if (tm)
 207                 return tm;
 208         if (depth > TCP_METRICS_RECLAIM_DEPTH)
 209                 return TCP_METRICS_RECLAIM_PTR;
 210         return NULL;
 211 }
 212
 213 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
 214                                                    const struct inetpeer_addr *daddr,
 215                                                    struct net *net, unsigned int hash)
 216 {
 217         struct tcp_metrics_block *tm;
 218         int depth = 0;
 219
 220         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 221              tm = rcu_dereference(tm->tcpm_next)) {
 222                 if (addr_same(&tm->tcpm_saddr, saddr) &&
 223                     addr_same(&tm->tcpm_daddr, daddr))
 224                         break;
 225                 depth++;
 226         }
 227         return tcp_get_encode(tm, depth);
 228 }
 229
 230 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
 231                                                        struct dst_entry *dst)
 232 {
 233         struct tcp_metrics_block *tm;
 234         struct inetpeer_addr saddr, daddr;
 235         unsigned int hash;
 236         struct net *net;
 237
 238         saddr.family = req->rsk_ops->family;
 239         daddr.family = req->rsk_ops->family;
 240         switch (daddr.family) {
 241         case AF_INET:
 242                 saddr.addr.a4 = inet_rsk(req)->ir_loc_addr;
 243                 daddr.addr.a4 = inet_rsk(req)->ir_rmt_addr;
 244                 hash = (__force unsigned int) daddr.addr.a4;
 245                 break;
 246 #if IS_ENABLED(CONFIG_IPV6)
 247         case AF_INET6:
 248                 *(struct in6_addr *)saddr.addr.a6 = inet_rsk(req)->ir_v6_loc_addr;
 249                 *(struct in6_addr *)daddr.addr.a6 = inet_rsk(req)->ir_v6_rmt_addr;
 250                 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
 251                 break;
 252 #endif
 253         default:
 254                 return NULL;
 255         }
 256
 257         net = dev_net(dst->dev);
 258         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 259
 260         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 261              tm = rcu_dereference(tm->tcpm_next)) {
 262                 if (addr_same(&tm->tcpm_saddr, &saddr) &&
 263                     addr_same(&tm->tcpm_daddr, &daddr))
 264                         break;
 265         }
 266         tcpm_check_stamp(tm, dst);
 267         return tm;
 268 }
 269
 270 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
 271 {
 272         struct tcp_metrics_block *tm;
 273         struct inetpeer_addr saddr, daddr;
 274         unsigned int hash;
 275         struct net *net;
 276
 277         if (tw->tw_family == AF_INET) {
 278                 saddr.family = AF_INET;
 279                 saddr.addr.a4 = tw->tw_rcv_saddr;
 280                 daddr.family = AF_INET;
 281                 daddr.addr.a4 = tw->tw_daddr;
 282                 hash = (__force unsigned int) daddr.addr.a4;
 283         }
 284 #if IS_ENABLED(CONFIG_IPV6)
 285         else if (tw->tw_family == AF_INET6) {
 286                 if (ipv6_addr_v4mapped(&tw->tw_v6_daddr)) {
 287                         saddr.family = AF_INET;
 288                         saddr.addr.a4 = tw->tw_rcv_saddr;
 289                         daddr.family = AF_INET;
 290                         daddr.addr.a4 = tw->tw_daddr;
 291                         hash = (__force unsigned int) daddr.addr.a4;
 292                 } else {
 293                         saddr.family = AF_INET6;
 294                         *(struct in6_addr *)saddr.addr.a6 = tw->tw_v6_rcv_saddr;
 295                         daddr.family = AF_INET6;
 296                         *(struct in6_addr *)daddr.addr.a6 = tw->tw_v6_daddr;
 297                         hash = ipv6_addr_hash(&tw->tw_v6_daddr);
 298                 }
 299         }
 300 #endif
 301         else
 302                 return NULL;
 303
 304         net = twsk_net(tw);
 305         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 306
 307         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 308              tm = rcu_dereference(tm->tcpm_next)) {
 309                 if (addr_same(&tm->tcpm_saddr, &saddr) &&
 310                     addr_same(&tm->tcpm_daddr, &daddr))
 311                         break;
 312         }
 313         return tm;
 314 }
 315
 316 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
 317                                                  struct dst_entry *dst,
 318                                                  bool create)
 319 {
 320         struct tcp_metrics_block *tm;
 321         struct inetpeer_addr saddr, daddr;
 322         unsigned int hash;
 323         struct net *net;
 324
 325         if (sk->sk_family == AF_INET) {
 326                 saddr.family = AF_INET;
 327                 saddr.addr.a4 = inet_sk(sk)->inet_saddr;
 328                 daddr.family = AF_INET;
 329                 daddr.addr.a4 = inet_sk(sk)->inet_daddr;
 330                 hash = (__force unsigned int) daddr.addr.a4;
 331         }
 332 #if IS_ENABLED(CONFIG_IPV6)
 333         else if (sk->sk_family == AF_INET6) {
 334                 if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
 335                         saddr.family = AF_INET;
 336                         saddr.addr.a4 = inet_sk(sk)->inet_saddr;
 337                         daddr.family = AF_INET;
 338                         daddr.addr.a4 = inet_sk(sk)->inet_daddr;
 339                         hash = (__force unsigned int) daddr.addr.a4;
 340                 } else {
 341                         saddr.family = AF_INET6;
 342                         *(struct in6_addr *)saddr.addr.a6 = sk->sk_v6_rcv_saddr;
 343                         daddr.family = AF_INET6;
 344                         *(struct in6_addr *)daddr.addr.a6 = sk->sk_v6_daddr;
 345                         hash = ipv6_addr_hash(&sk->sk_v6_daddr);
 346                 }
 347         }
 348 #endif
 349         else
 350                 return NULL;
 351
 352         net = dev_net(dst->dev);
 353         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 354
 355         tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
 356         if (tm == TCP_METRICS_RECLAIM_PTR)
 357                 tm = NULL;
 358         if (!tm && create)
 359                 tm = tcpm_new(dst, &saddr, &daddr, hash);
 360         else
 361                 tcpm_check_stamp(tm, dst);
 362
 363         return tm;
 364 }
 365
 366 /* Save metrics learned by this TCP session.  This function is called
 367  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
 368  * or goes from LAST-ACK to CLOSE.
 369  */
 370 void tcp_update_metrics(struct sock *sk)
 371 {
 372         const struct inet_connection_sock *icsk = inet_csk(sk);
 373         struct dst_entry *dst = __sk_dst_get(sk);
 374         struct tcp_sock *tp = tcp_sk(sk);
 375         struct tcp_metrics_block *tm;
 376         unsigned long rtt;
 377         u32 val;
 378         int m;
 379
 380         if (sysctl_tcp_nometrics_save || !dst)
 381                 return;
 382
 383         if (dst->flags & DST_HOST)
 384                 dst_confirm(dst);
 385
 386         rcu_read_lock();
 387         if (icsk->icsk_backoff || !tp->srtt) {
 388                 /* This session failed to estimate rtt. Why?
 389                  * Probably, no packets returned in time.  Reset our
 390                  * results.
 391                  */
 392                 tm = tcp_get_metrics(sk, dst, false);
 393                 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
 394                         tcp_metric_set(tm, TCP_METRIC_RTT, 0);
 395                 goto out_unlock;
 396         } else
 397                 tm = tcp_get_metrics(sk, dst, true);
 398
 399         if (!tm)
 400                 goto out_unlock;
 401
 402         rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
 403         m = rtt - tp->srtt;
 404
 405         /* If newly calculated rtt larger than stored one, store new
 406          * one. Otherwise, use EWMA. Remember, rtt overestimation is
 407          * always better than underestimation.
 408          */
 409         if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
 410                 if (m <= 0)
 411                         rtt = tp->srtt;
 412                 else
 413                         rtt -= (m >> 3);
 414                 tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt);
 415         }
 416
 417         if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
 418                 unsigned long var;
 419
 420                 if (m < 0)
 421                         m = -m;
 422
 423                 /* Scale deviation to rttvar fixed point */
 424                 m >>= 1;
 425                 if (m < tp->mdev)
 426                         m = tp->mdev;
 427
 428                 var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
 429                 if (m >= var)
 430                         var = m;
 431                 else
 432                         var -= (var - m) >> 2;
 433
 434                 tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var);
 435         }
 436
 437         if (tcp_in_initial_slowstart(tp)) {
 438                 /* Slow start still did not finish. */
 439                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 440                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 441                         if (val && (tp->snd_cwnd >> 1) > val)
 442                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 443                                                tp->snd_cwnd >> 1);
 444                 }
 445                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 446                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 447                         if (tp->snd_cwnd > val)
 448                                 tcp_metric_set(tm, TCP_METRIC_CWND,
 449                                                tp->snd_cwnd);
 450                 }
 451         } else if (tp->snd_cwnd > tp->snd_ssthresh &&
 452                    icsk->icsk_ca_state == TCP_CA_Open) {
 453                 /* Cong. avoidance phase, cwnd is reliable. */
 454                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
 455                         tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 456                                        max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
 457                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 458                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 459                         tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
 460                 }
 461         } else {
 462                 /* Else slow start did not finish, cwnd is non-sense,
 463                  * ssthresh may be also invalid.
 464                  */
 465                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 466                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 467                         tcp_metric_set(tm, TCP_METRIC_CWND,
 468                                        (val + tp->snd_ssthresh) >> 1);
 469                 }
 470                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 471                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 472                         if (val && tp->snd_ssthresh > val)
 473                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 474                                                tp->snd_ssthresh);
 475                 }
 476                 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
 477                         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 478                         if (val < tp->reordering &&
 479                             tp->reordering != sysctl_tcp_reordering)
 480                                 tcp_metric_set(tm, TCP_METRIC_REORDERING,
 481                                                tp->reordering);
 482                 }
 483         }
 484         tm->tcpm_stamp = jiffies;
 485 out_unlock:
 486         rcu_read_unlock();
 487 }
 488
 489 /* Initialize metrics on socket. */
 490
 491 void tcp_init_metrics(struct sock *sk)
 492 {
 493         struct dst_entry *dst = __sk_dst_get(sk);
 494         struct tcp_sock *tp = tcp_sk(sk);
 495         struct tcp_metrics_block *tm;
 496         u32 val, crtt = 0; /* cached RTT scaled by 8 */
 497
 498         if (dst == NULL)
 499                 goto reset;
 500
 501         dst_confirm(dst);
 502
 503         rcu_read_lock();
 504         tm = tcp_get_metrics(sk, dst, true);
 505         if (!tm) {
 506                 rcu_read_unlock();
 507                 goto reset;
 508         }
 509
 510         if (tcp_metric_locked(tm, TCP_METRIC_CWND))
 511                 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
 512
 513         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 514         if (val) {
 515                 tp->snd_ssthresh = val;
 516                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
 517                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
 518         } else {
 519                 /* ssthresh may have been reduced unnecessarily during.
 520                  * 3WHS. Restore it back to its initial default.
 521                  */
 522                 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 523         }
 524         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 525         if (val && tp->reordering != val) {
 526                 tcp_disable_fack(tp);
 527                 tcp_disable_early_retrans(tp);
 528                 tp->reordering = val;
 529         }
 530
 531         crtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
 532         rcu_read_unlock();
 533 reset:
 534         /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
 535          * to seed the RTO for later data packets because SYN packets are
 536          * small. Use the per-dst cached values to seed the RTO but keep
 537          * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
 538          * Later the RTO will be updated immediately upon obtaining the first
 539          * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
 540          * influences the first RTO but not later RTT estimation.
 541          *
 542          * But if RTT is not available from the SYN (due to retransmits or
 543          * syn cookies) or the cache, force a conservative 3secs timeout.
 544          *
 545          * A bit of theory. RTT is time passed after "normal" sized packet
 546          * is sent until it is ACKed. In normal circumstances sending small
 547          * packets force peer to delay ACKs and calculation is correct too.
 548          * The algorithm is adaptive and, provided we follow specs, it
 549          * NEVER underestimate RTT. BUT! If peer tries to make some clever
 550          * tricks sort of "quick acks" for time long enough to decrease RTT
 551          * to low value, and then abruptly stops to do it and starts to delay
 552          * ACKs, wait for troubles.
 553          */
 554         if (crtt > tp->srtt) {
 555                 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
 556                 crtt >>= 3;
 557                 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
 558         } else if (tp->srtt == 0) {
 559                 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
 560                  * 3WHS. This is most likely due to retransmission,
 561                  * including spurious one. Reset the RTO back to 3secs
 562                  * from the more aggressive 1sec to avoid more spurious
 563                  * retransmission.
 564                  */
 565                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
 566                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
 567         }
 568         /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
 569          * retransmitted. In light of RFC6298 more aggressive 1sec
 570          * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
 571          * retransmission has occurred.
 572          */
 573         if (tp->total_retrans > 1)
 574                 tp->snd_cwnd = 1;
 575         else
 576                 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
 577         tp->snd_cwnd_stamp = tcp_time_stamp;
 578 }
 579
 580 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
 581 {
 582         struct tcp_metrics_block *tm;
 583         bool ret;
 584
 585         if (!dst)
 586                 return false;
 587
 588         rcu_read_lock();
 589         tm = __tcp_get_metrics_req(req, dst);
 590         if (paws_check) {
 591                 if (tm &&
 592                     (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
 593                     (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
 594                         ret = false;
 595                 else
 596                         ret = true;
 597         } else {
 598                 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
 599                         ret = true;
 600                 else
 601                         ret = false;
 602         }
 603         rcu_read_unlock();
 604
 605         return ret;
 606 }
 607 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
 608
 609 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
 610 {
 611         struct tcp_metrics_block *tm;
 612
 613         rcu_read_lock();
 614         tm = tcp_get_metrics(sk, dst, true);
 615         if (tm) {
 616                 struct tcp_sock *tp = tcp_sk(sk);
 617
 618                 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
 619                         tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
 620                         tp->rx_opt.ts_recent = tm->tcpm_ts;
 621                 }
 622         }
 623         rcu_read_unlock();
 624 }
 625 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
 626
 627 /* VJ's idea. Save last timestamp seen from this destination and hold
 628  * it at least for normal timewait interval to use for duplicate
 629  * segment detection in subsequent connections, before they enter
 630  * synchronized state.
 631  */
 632 bool tcp_remember_stamp(struct sock *sk)
 633 {
 634         struct dst_entry *dst = __sk_dst_get(sk);
 635         bool ret = false;
 636
 637         if (dst) {
 638                 struct tcp_metrics_block *tm;
 639
 640                 rcu_read_lock();
 641                 tm = tcp_get_metrics(sk, dst, true);
 642                 if (tm) {
 643                         struct tcp_sock *tp = tcp_sk(sk);
 644
 645                         if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
 646                             ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 647                              tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
 648                                 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
 649                                 tm->tcpm_ts = tp->rx_opt.ts_recent;
 650                         }
 651                         ret = true;
 652                 }
 653                 rcu_read_unlock();
 654         }
 655         return ret;
 656 }
 657
 658 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
 659 {
 660         struct tcp_metrics_block *tm;
 661         bool ret = false;
 662
 663         rcu_read_lock();
 664         tm = __tcp_get_metrics_tw(tw);
 665         if (tm) {
 666                 const struct tcp_timewait_sock *tcptw;
 667                 struct sock *sk = (struct sock *) tw;
 668
 669                 tcptw = tcp_twsk(sk);
 670                 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
 671                     ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 672                      tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
 673                         tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
 674                         tm->tcpm_ts        = tcptw->tw_ts_recent;
 675                 }
 676                 ret = true;
 677         }
 678         rcu_read_unlock();
 679
 680         return ret;
 681 }
 682
 683 static DEFINE_SEQLOCK(fastopen_seqlock);
 684
 685 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
 686                             struct tcp_fastopen_cookie *cookie,
 687                             int *syn_loss, unsigned long *last_syn_loss)
 688 {
 689         struct tcp_metrics_block *tm;
 690
 691         rcu_read_lock();
 692         tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
 693         if (tm) {
 694                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 695                 unsigned int seq;
 696
 697                 do {
 698                         seq = read_seqbegin(&fastopen_seqlock);
 699                         if (tfom->mss)
 700                                 *mss = tfom->mss;
 701                         *cookie = tfom->cookie;
 702                         *syn_loss = tfom->syn_loss;
 703                         *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
 704                 } while (read_seqretry(&fastopen_seqlock, seq));
 705         }
 706         rcu_read_unlock();
 707 }
 708
 709 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
 710                             struct tcp_fastopen_cookie *cookie, bool syn_lost)
 711 {
 712         struct dst_entry *dst = __sk_dst_get(sk);
 713         struct tcp_metrics_block *tm;
 714
 715         if (!dst)
 716                 return;
 717         rcu_read_lock();
 718         tm = tcp_get_metrics(sk, dst, true);
 719         if (tm) {
 720                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 721
 722                 write_seqlock_bh(&fastopen_seqlock);
 723                 if (mss)
 724                         tfom->mss = mss;
 725                 if (cookie && cookie->len > 0)
 726                         tfom->cookie = *cookie;
 727                 if (syn_lost) {
 728                         ++tfom->syn_loss;
 729                         tfom->last_syn_loss = jiffies;
 730                 } else
 731                         tfom->syn_loss = 0;
 732                 write_sequnlock_bh(&fastopen_seqlock);
 733         }
 734         rcu_read_unlock();
 735 }
 736
 737 static struct genl_family tcp_metrics_nl_family = {
 738         .id             = GENL_ID_GENERATE,
 739         .hdrsize        = 0,
 740         .name           = TCP_METRICS_GENL_NAME,
 741         .version        = TCP_METRICS_GENL_VERSION,
 742         .maxattr        = TCP_METRICS_ATTR_MAX,
 743         .netnsok        = true,
 744 };
 745
 746 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
 747         [TCP_METRICS_ATTR_ADDR_IPV4]    = { .type = NLA_U32, },
 748         [TCP_METRICS_ATTR_ADDR_IPV6]    = { .type = NLA_BINARY,
 749                                             .len = sizeof(struct in6_addr), },
 750         /* Following attributes are not received for GET/DEL,
 751          * we keep them for reference
 752          */
 753 #if 0
 754         [TCP_METRICS_ATTR_AGE]          = { .type = NLA_MSECS, },
 755         [TCP_METRICS_ATTR_TW_TSVAL]     = { .type = NLA_U32, },
 756         [TCP_METRICS_ATTR_TW_TS_STAMP]  = { .type = NLA_S32, },
 757         [TCP_METRICS_ATTR_VALS]         = { .type = NLA_NESTED, },
 758         [TCP_METRICS_ATTR_FOPEN_MSS]    = { .type = NLA_U16, },
 759         [TCP_METRICS_ATTR_FOPEN_SYN_DROPS]      = { .type = NLA_U16, },
 760         [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]    = { .type = NLA_MSECS, },
 761         [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
 762                                             .len = TCP_FASTOPEN_COOKIE_MAX, },
 763 #endif
 764 };
 765
 766 /* Add attributes, caller cancels its header on failure */
 767 static int tcp_metrics_fill_info(struct sk_buff *msg,
 768                                  struct tcp_metrics_block *tm)
 769 {
 770         struct nlattr *nest;
 771         int i;
 772
 773         switch (tm->tcpm_daddr.family) {
 774         case AF_INET:
 775                 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
 776                                 tm->tcpm_daddr.addr.a4) < 0)
 777                         goto nla_put_failure;
 778                 if (nla_put_be32(msg, TCP_METRICS_ATTR_SADDR_IPV4,
 779                                 tm->tcpm_saddr.addr.a4) < 0)
 780                         goto nla_put_failure;
 781                 break;
 782         case AF_INET6:
 783                 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
 784                             tm->tcpm_daddr.addr.a6) < 0)
 785                         goto nla_put_failure;
 786                 if (nla_put(msg, TCP_METRICS_ATTR_SADDR_IPV6, 16,
 787                             tm->tcpm_saddr.addr.a6) < 0)
 788                         goto nla_put_failure;
 789                 break;
 790         default:
 791                 return -EAFNOSUPPORT;
 792         }
 793
 794         if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
 795                           jiffies - tm->tcpm_stamp) < 0)
 796                 goto nla_put_failure;
 797         if (tm->tcpm_ts_stamp) {
 798                 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
 799                                 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
 800                         goto nla_put_failure;
 801                 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
 802                                 tm->tcpm_ts) < 0)
 803                         goto nla_put_failure;
 804         }
 805
 806         {
 807                 int n = 0;
 808
 809                 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
 810                 if (!nest)
 811                         goto nla_put_failure;
 812                 for (i = 0; i < TCP_METRIC_MAX + 1; i++) {
 813                         if (!tm->tcpm_vals[i])
 814                                 continue;
 815                         if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0)
 816                                 goto nla_put_failure;
 817                         n++;
 818                 }
 819                 if (n)
 820                         nla_nest_end(msg, nest);
 821                 else
 822                         nla_nest_cancel(msg, nest);
 823         }
 824
 825         {
 826                 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
 827                 unsigned int seq;
 828
 829                 do {
 830                         seq = read_seqbegin(&fastopen_seqlock);
 831                         tfom_copy[0] = tm->tcpm_fastopen;
 832                 } while (read_seqretry(&fastopen_seqlock, seq));
 833
 834                 tfom = tfom_copy;
 835                 if (tfom->mss &&
 836                     nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
 837                                 tfom->mss) < 0)
 838                         goto nla_put_failure;
 839                 if (tfom->syn_loss &&
 840                     (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
 841                                 tfom->syn_loss) < 0 ||
 842                      nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
 843                                 jiffies - tfom->last_syn_loss) < 0))
 844                         goto nla_put_failure;
 845                 if (tfom->cookie.len > 0 &&
 846                     nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
 847                             tfom->cookie.len, tfom->cookie.val) < 0)
 848                         goto nla_put_failure;
 849         }
 850
 851         return 0;
 852
 853 nla_put_failure:
 854         return -EMSGSIZE;
 855 }
 856
 857 static int tcp_metrics_dump_info(struct sk_buff *skb,
 858                                  struct netlink_callback *cb,
 859                                  struct tcp_metrics_block *tm)
 860 {
 861         void *hdr;
 862
 863         hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 864                           &tcp_metrics_nl_family, NLM_F_MULTI,
 865                           TCP_METRICS_CMD_GET);
 866         if (!hdr)
 867                 return -EMSGSIZE;
 868
 869         if (tcp_metrics_fill_info(skb, tm) < 0)
 870                 goto nla_put_failure;
 871
 872         return genlmsg_end(skb, hdr);
 873
 874 nla_put_failure:
 875         genlmsg_cancel(skb, hdr);
 876         return -EMSGSIZE;
 877 }
 878
 879 static int tcp_metrics_nl_dump(struct sk_buff *skb,
 880                                struct netlink_callback *cb)
 881 {
 882         struct net *net = sock_net(skb->sk);
 883         unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
 884         unsigned int row, s_row = cb->args[0];
 885         int s_col = cb->args[1], col = s_col;
 886
 887         for (row = s_row; row < max_rows; row++, s_col = 0) {
 888                 struct tcp_metrics_block *tm;
 889                 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
 890
 891                 rcu_read_lock();
 892                 for (col = 0, tm = rcu_dereference(hb->chain); tm;
 893                      tm = rcu_dereference(tm->tcpm_next), col++) {
 894                         if (col < s_col)
 895                                 continue;
 896                         if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
 897                                 rcu_read_unlock();
 898                                 goto done;
 899                         }
 900                 }
 901                 rcu_read_unlock();
 902         }
 903
 904 done:
 905         cb->args[0] = row;
 906         cb->args[1] = col;
 907         return skb->len;
 908 }
 909
 910 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 911                            unsigned int *hash, int optional, int v4, int v6)
 912 {
 913         struct nlattr *a;
 914
 915         a = info->attrs[v4];
 916         if (a) {
 917                 addr->family = AF_INET;
 918                 addr->addr.a4 = nla_get_be32(a);
 919                 if (hash)
 920                         *hash = (__force unsigned int) addr->addr.a4;
 921                 return 0;
 922         }
 923         a = info->attrs[v6];
 924         if (a) {
 925                 if (nla_len(a) != sizeof(struct in6_addr))
 926                         return -EINVAL;
 927                 addr->family = AF_INET6;
 928                 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
 929                 if (hash)
 930                         *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
 931                 return 0;
 932         }
 933         return optional ? 1 : -EAFNOSUPPORT;
 934 }
 935
 936 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 937                          unsigned int *hash, int optional)
 938 {
 939         return __parse_nl_addr(info, addr, hash, optional,
 940                                TCP_METRICS_ATTR_ADDR_IPV4,
 941                                TCP_METRICS_ATTR_ADDR_IPV6);
 942 }
 943
 944 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
 945 {
 946         return __parse_nl_addr(info, addr, NULL, 0,
 947                                TCP_METRICS_ATTR_SADDR_IPV4,
 948                                TCP_METRICS_ATTR_SADDR_IPV6);
 949 }
 950
 951 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
 952 {
 953         struct tcp_metrics_block *tm;
 954         struct inetpeer_addr saddr, daddr;
 955         unsigned int hash;
 956         struct sk_buff *msg;
 957         struct net *net = genl_info_net(info);
 958         void *reply;
 959         int ret;
 960         bool src = true;
 961
 962         ret = parse_nl_addr(info, &daddr, &hash, 0);
 963         if (ret < 0)
 964                 return ret;
 965
 966         ret = parse_nl_saddr(info, &saddr);
 967         if (ret < 0)
 968                 src = false;
 969
 970         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 971         if (!msg)
 972                 return -ENOMEM;
 973
 974         reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
 975                                   info->genlhdr->cmd);
 976         if (!reply)
 977                 goto nla_put_failure;
 978
 979         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 980         ret = -ESRCH;
 981         rcu_read_lock();
 982         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 983              tm = rcu_dereference(tm->tcpm_next)) {
 984                 if (addr_same(&tm->tcpm_daddr, &daddr) &&
 985                     (!src || addr_same(&tm->tcpm_saddr, &saddr))) {
 986                         ret = tcp_metrics_fill_info(msg, tm);
 987                         break;
 988                 }
 989         }
 990         rcu_read_unlock();
 991         if (ret < 0)
 992                 goto out_free;
 993
 994         genlmsg_end(msg, reply);
 995         return genlmsg_reply(msg, info);
 996
 997 nla_put_failure:
 998         ret = -EMSGSIZE;
 999
1000 out_free:
1001         nlmsg_free(msg);
1002         return ret;
1003 }
1004
1005 #define deref_locked_genl(p)    \
1006         rcu_dereference_protected(p, lockdep_genl_is_held() && \
1007                                      lockdep_is_held(&tcp_metrics_lock))
1008
1009 #define deref_genl(p)   rcu_dereference_protected(p, lockdep_genl_is_held())
1010
1011 static int tcp_metrics_flush_all(struct net *net)
1012 {
1013         unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
1014         struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
1015         struct tcp_metrics_block *tm;
1016         unsigned int row;
1017
1018         for (row = 0; row < max_rows; row++, hb++) {
1019                 spin_lock_bh(&tcp_metrics_lock);
1020                 tm = deref_locked_genl(hb->chain);
1021                 if (tm)
1022                         hb->chain = NULL;
1023                 spin_unlock_bh(&tcp_metrics_lock);
1024                 while (tm) {
1025                         struct tcp_metrics_block *next;
1026
1027                         next = deref_genl(tm->tcpm_next);
1028                         kfree_rcu(tm, rcu_head);
1029                         tm = next;
1030                 }
1031         }
1032         return 0;
1033 }
1034
1035 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
1036 {
1037         struct tcpm_hash_bucket *hb;
1038         struct tcp_metrics_block *tm;
1039         struct tcp_metrics_block __rcu **pp;
1040         struct inetpeer_addr saddr, daddr;
1041         unsigned int hash;
1042         struct net *net = genl_info_net(info);
1043         int ret;
1044         bool src = true, found = false;
1045
1046         ret = parse_nl_addr(info, &daddr, &hash, 1);
1047         if (ret < 0)
1048                 return ret;
1049         if (ret > 0)
1050                 return tcp_metrics_flush_all(net);
1051         ret = parse_nl_saddr(info, &saddr);
1052         if (ret < 0)
1053                 src = false;
1054
1055         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
1056         hb = net->ipv4.tcp_metrics_hash + hash;
1057         pp = &hb->chain;
1058         spin_lock_bh(&tcp_metrics_lock);
1059         for (tm = deref_locked_genl(*pp); tm; tm = deref_locked_genl(*pp)) {
1060                 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1061                     (!src || addr_same(&tm->tcpm_saddr, &saddr))) {
1062                         *pp = tm->tcpm_next;
1063                         kfree_rcu(tm, rcu_head);
1064                         found = true;
1065                 } else {
1066                         pp = &tm->tcpm_next;
1067                 }
1068         }
1069         spin_unlock_bh(&tcp_metrics_lock);
1070         if (!found)
1071                 return -ESRCH;
1072         return 0;
1073 }
1074
1075 static const struct genl_ops tcp_metrics_nl_ops[] = {
1076         {
1077                 .cmd = TCP_METRICS_CMD_GET,
1078                 .doit = tcp_metrics_nl_cmd_get,
1079                 .dumpit = tcp_metrics_nl_dump,
1080                 .policy = tcp_metrics_nl_policy,
1081                 .flags = GENL_ADMIN_PERM,
1082         },
1083         {
1084                 .cmd = TCP_METRICS_CMD_DEL,
1085                 .doit = tcp_metrics_nl_cmd_del,
1086                 .policy = tcp_metrics_nl_policy,
1087                 .flags = GENL_ADMIN_PERM,
1088         },
1089 };
1090
1091 static unsigned int tcpmhash_entries;
1092 static int __init set_tcpmhash_entries(char *str)
1093 {
1094         ssize_t ret;
1095
1096         if (!str)
1097                 return 0;
1098
1099         ret = kstrtouint(str, 0, &tcpmhash_entries);
1100         if (ret)
1101                 return 0;
1102
1103         return 1;
1104 }
1105 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1106
1107 static int __net_init tcp_net_metrics_init(struct net *net)
1108 {
1109         size_t size;
1110         unsigned int slots;
1111
1112         slots = tcpmhash_entries;
1113         if (!slots) {
1114                 if (totalram_pages >= 128 * 1024)
1115                         slots = 16 * 1024;
1116                 else
1117                         slots = 8 * 1024;
1118         }
1119
1120         net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1121         size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
1122
1123         net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1124         if (!net->ipv4.tcp_metrics_hash)
1125                 net->ipv4.tcp_metrics_hash = vzalloc(size);
1126
1127         if (!net->ipv4.tcp_metrics_hash)
1128                 return -ENOMEM;
1129
1130         return 0;
1131 }
1132
1133 static void __net_exit tcp_net_metrics_exit(struct net *net)
1134 {
1135         unsigned int i;
1136
1137         for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1138                 struct tcp_metrics_block *tm, *next;
1139
1140                 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1141                 while (tm) {
1142                         next = rcu_dereference_protected(tm->tcpm_next, 1);
1143                         kfree(tm);
1144                         tm = next;
1145                 }
1146         }
1147         if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1148                 vfree(net->ipv4.tcp_metrics_hash);
1149         else
1150                 kfree(net->ipv4.tcp_metrics_hash);
1151 }
1152
1153 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1154         .init   =       tcp_net_metrics_init,
1155         .exit   =       tcp_net_metrics_exit,
1156 };
1157
1158 void __init tcp_metrics_init(void)
1159 {
1160         int ret;
1161
1162         ret = register_pernet_subsys(&tcp_net_metrics_ops);
1163         if (ret < 0)
1164                 goto cleanup;
1165         ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1166                                             tcp_metrics_nl_ops);
1167         if (ret < 0)
1168                 goto cleanup_subsys;
1169         return;
1170
1171 cleanup_subsys:
1172         unregister_pernet_subsys(&tcp_net_metrics_ops);
1173
1174 cleanup:
1175         return;
1176 }