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                 try_exp:2;              /* Request w/ exp. option (once) */
  33         unsigned long   last_syn_loss;  /* Last Fast Open SYN loss */
  34         struct  tcp_fastopen_cookie     cookie;
  35 };
  36
  37 /* TCP_METRIC_MAX includes 2 extra fields for userspace compatibility
  38  * Kernel only stores RTT and RTTVAR in usec resolution
  39  */
  40 #define TCP_METRIC_MAX_KERNEL (TCP_METRIC_MAX - 2)
  41
  42 struct tcp_metrics_block {
  43         struct tcp_metrics_block __rcu  *tcpm_next;
  44         possible_net_t                  tcpm_net;
  45         struct inetpeer_addr            tcpm_saddr;
  46         struct inetpeer_addr            tcpm_daddr;
  47         unsigned long                   tcpm_stamp;
  48         u32                             tcpm_ts;
  49         u32                             tcpm_ts_stamp;
  50         u32                             tcpm_lock;
  51         u32                             tcpm_vals[TCP_METRIC_MAX_KERNEL + 1];
  52         struct tcp_fastopen_metrics     tcpm_fastopen;
  53
  54         struct rcu_head                 rcu_head;
  55 };
  56
  57 static inline struct net *tm_net(struct tcp_metrics_block *tm)
  58 {
  59         return read_pnet(&tm->tcpm_net);
  60 }
  61
  62 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
  63                               enum tcp_metric_index idx)
  64 {
  65         return tm->tcpm_lock & (1 << idx);
  66 }
  67
  68 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
  69                           enum tcp_metric_index idx)
  70 {
  71         return tm->tcpm_vals[idx];
  72 }
  73
  74 static void tcp_metric_set(struct tcp_metrics_block *tm,
  75                            enum tcp_metric_index idx,
  76                            u32 val)
  77 {
  78         tm->tcpm_vals[idx] = val;
  79 }
  80
  81 static bool addr_same(const struct inetpeer_addr *a,
  82                       const struct inetpeer_addr *b)
  83 {
  84         return inetpeer_addr_cmp(a, b) == 0;
  85 }
  86
  87 struct tcpm_hash_bucket {
  88         struct tcp_metrics_block __rcu  *chain;
  89 };
  90
  91 static struct tcpm_hash_bucket  *tcp_metrics_hash __read_mostly;
  92 static unsigned int             tcp_metrics_hash_log __read_mostly;
  93
  94 static DEFINE_SPINLOCK(tcp_metrics_lock);
  95
  96 static void tcpm_suck_dst(struct tcp_metrics_block *tm,
  97                           const struct dst_entry *dst,
  98                           bool fastopen_clear)
  99 {
 100         u32 msval;
 101         u32 val;
 102
 103         tm->tcpm_stamp = jiffies;
 104
 105         val = 0;
 106         if (dst_metric_locked(dst, RTAX_RTT))
 107                 val |= 1 << TCP_METRIC_RTT;
 108         if (dst_metric_locked(dst, RTAX_RTTVAR))
 109                 val |= 1 << TCP_METRIC_RTTVAR;
 110         if (dst_metric_locked(dst, RTAX_SSTHRESH))
 111                 val |= 1 << TCP_METRIC_SSTHRESH;
 112         if (dst_metric_locked(dst, RTAX_CWND))
 113                 val |= 1 << TCP_METRIC_CWND;
 114         if (dst_metric_locked(dst, RTAX_REORDERING))
 115                 val |= 1 << TCP_METRIC_REORDERING;
 116         tm->tcpm_lock = val;
 117
 118         msval = dst_metric_raw(dst, RTAX_RTT);
 119         tm->tcpm_vals[TCP_METRIC_RTT] = msval * USEC_PER_MSEC;
 120
 121         msval = dst_metric_raw(dst, RTAX_RTTVAR);
 122         tm->tcpm_vals[TCP_METRIC_RTTVAR] = msval * USEC_PER_MSEC;
 123         tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
 124         tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
 125         tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
 126         tm->tcpm_ts = 0;
 127         tm->tcpm_ts_stamp = 0;
 128         if (fastopen_clear) {
 129                 tm->tcpm_fastopen.mss = 0;
 130                 tm->tcpm_fastopen.syn_loss = 0;
 131                 tm->tcpm_fastopen.try_exp = 0;
 132                 tm->tcpm_fastopen.cookie.exp = false;
 133                 tm->tcpm_fastopen.cookie.len = 0;
 134         }
 135 }
 136
 137 #define TCP_METRICS_TIMEOUT             (60 * 60 * HZ)
 138
 139 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
 140 {
 141         if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
 142                 tcpm_suck_dst(tm, dst, false);
 143 }
 144
 145 #define TCP_METRICS_RECLAIM_DEPTH       5
 146 #define TCP_METRICS_RECLAIM_PTR         (struct tcp_metrics_block *) 0x1UL
 147
 148 #define deref_locked(p) \
 149         rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
 150
 151 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
 152                                           struct inetpeer_addr *saddr,
 153                                           struct inetpeer_addr *daddr,
 154                                           unsigned int hash)
 155 {
 156         struct tcp_metrics_block *tm;
 157         struct net *net;
 158         bool reclaim = false;
 159
 160         spin_lock_bh(&tcp_metrics_lock);
 161         net = dev_net(dst->dev);
 162
 163         /* While waiting for the spin-lock the cache might have been populated
 164          * with this entry and so we have to check again.
 165          */
 166         tm = __tcp_get_metrics(saddr, daddr, net, hash);
 167         if (tm == TCP_METRICS_RECLAIM_PTR) {
 168                 reclaim = true;
 169                 tm = NULL;
 170         }
 171         if (tm) {
 172                 tcpm_check_stamp(tm, dst);
 173                 goto out_unlock;
 174         }
 175
 176         if (unlikely(reclaim)) {
 177                 struct tcp_metrics_block *oldest;
 178
 179                 oldest = deref_locked(tcp_metrics_hash[hash].chain);
 180                 for (tm = deref_locked(oldest->tcpm_next); tm;
 181                      tm = deref_locked(tm->tcpm_next)) {
 182                         if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
 183                                 oldest = tm;
 184                 }
 185                 tm = oldest;
 186         } else {
 187                 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
 188                 if (!tm)
 189                         goto out_unlock;
 190         }
 191         write_pnet(&tm->tcpm_net, net);
 192         tm->tcpm_saddr = *saddr;
 193         tm->tcpm_daddr = *daddr;
 194
 195         tcpm_suck_dst(tm, dst, true);
 196
 197         if (likely(!reclaim)) {
 198                 tm->tcpm_next = tcp_metrics_hash[hash].chain;
 199                 rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
 200         }
 201
 202 out_unlock:
 203         spin_unlock_bh(&tcp_metrics_lock);
 204         return tm;
 205 }
 206
 207 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
 208 {
 209         if (tm)
 210                 return tm;
 211         if (depth > TCP_METRICS_RECLAIM_DEPTH)
 212                 return TCP_METRICS_RECLAIM_PTR;
 213         return NULL;
 214 }
 215
 216 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
 217                                                    const struct inetpeer_addr *daddr,
 218                                                    struct net *net, unsigned int hash)
 219 {
 220         struct tcp_metrics_block *tm;
 221         int depth = 0;
 222
 223         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
 224              tm = rcu_dereference(tm->tcpm_next)) {
 225                 if (addr_same(&tm->tcpm_saddr, saddr) &&
 226                     addr_same(&tm->tcpm_daddr, daddr) &&
 227                     net_eq(tm_net(tm), net))
 228                         break;
 229                 depth++;
 230         }
 231         return tcp_get_encode(tm, depth);
 232 }
 233
 234 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
 235                                                        struct dst_entry *dst)
 236 {
 237         struct tcp_metrics_block *tm;
 238         struct inetpeer_addr saddr, daddr;
 239         unsigned int hash;
 240         struct net *net;
 241
 242         saddr.family = req->rsk_ops->family;
 243         daddr.family = req->rsk_ops->family;
 244         switch (daddr.family) {
 245         case AF_INET:
 246                 inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
 247                 inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
 248                 hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
 249                 break;
 250 #if IS_ENABLED(CONFIG_IPV6)
 251         case AF_INET6:
 252                 inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
 253                 inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr);
 254                 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
 255                 break;
 256 #endif
 257         default:
 258                 return NULL;
 259         }
 260
 261         net = dev_net(dst->dev);
 262         hash ^= net_hash_mix(net);
 263         hash = hash_32(hash, tcp_metrics_hash_log);
 264
 265         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
 266              tm = rcu_dereference(tm->tcpm_next)) {
 267                 if (addr_same(&tm->tcpm_saddr, &saddr) &&
 268                     addr_same(&tm->tcpm_daddr, &daddr) &&
 269                     net_eq(tm_net(tm), net))
 270                         break;
 271         }
 272         tcpm_check_stamp(tm, dst);
 273         return tm;
 274 }
 275
 276 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
 277 {
 278         struct tcp_metrics_block *tm;
 279         struct inetpeer_addr saddr, daddr;
 280         unsigned int hash;
 281         struct net *net;
 282
 283         if (tw->tw_family == AF_INET) {
 284                 inetpeer_set_addr_v4(&saddr, tw->tw_rcv_saddr);
 285                 inetpeer_set_addr_v4(&daddr, tw->tw_daddr);
 286                 hash = ipv4_addr_hash(tw->tw_daddr);
 287         }
 288 #if IS_ENABLED(CONFIG_IPV6)
 289         else if (tw->tw_family == AF_INET6) {
 290                 if (ipv6_addr_v4mapped(&tw->tw_v6_daddr)) {
 291                         inetpeer_set_addr_v4(&saddr, tw->tw_rcv_saddr);
 292                         inetpeer_set_addr_v4(&daddr, tw->tw_daddr);
 293                         hash = ipv4_addr_hash(tw->tw_daddr);
 294                 } else {
 295                         inetpeer_set_addr_v6(&saddr, &tw->tw_v6_rcv_saddr);
 296                         inetpeer_set_addr_v6(&daddr, &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 ^= net_hash_mix(net);
 306         hash = hash_32(hash, tcp_metrics_hash_log);
 307
 308         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
 309              tm = rcu_dereference(tm->tcpm_next)) {
 310                 if (addr_same(&tm->tcpm_saddr, &saddr) &&
 311                     addr_same(&tm->tcpm_daddr, &daddr) &&
 312                     net_eq(tm_net(tm), net))
 313                         break;
 314         }
 315         return tm;
 316 }
 317
 318 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
 319                                                  struct dst_entry *dst,
 320                                                  bool create)
 321 {
 322         struct tcp_metrics_block *tm;
 323         struct inetpeer_addr saddr, daddr;
 324         unsigned int hash;
 325         struct net *net;
 326
 327         if (sk->sk_family == AF_INET) {
 328                 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
 329                 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
 330                 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
 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                         inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
 336                         inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
 337                         hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
 338                 } else {
 339                         inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
 340                         inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
 341                         hash = ipv6_addr_hash(&sk->sk_v6_daddr);
 342                 }
 343         }
 344 #endif
 345         else
 346                 return NULL;
 347
 348         net = dev_net(dst->dev);
 349         hash ^= net_hash_mix(net);
 350         hash = hash_32(hash, tcp_metrics_hash_log);
 351
 352         tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
 353         if (tm == TCP_METRICS_RECLAIM_PTR)
 354                 tm = NULL;
 355         if (!tm && create)
 356                 tm = tcpm_new(dst, &saddr, &daddr, hash);
 357         else
 358                 tcpm_check_stamp(tm, dst);
 359
 360         return tm;
 361 }
 362
 363 /* Save metrics learned by this TCP session.  This function is called
 364  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
 365  * or goes from LAST-ACK to CLOSE.
 366  */
 367 void tcp_update_metrics(struct sock *sk)
 368 {
 369         const struct inet_connection_sock *icsk = inet_csk(sk);
 370         struct dst_entry *dst = __sk_dst_get(sk);
 371         struct tcp_sock *tp = tcp_sk(sk);
 372         struct net *net = sock_net(sk);
 373         struct tcp_metrics_block *tm;
 374         unsigned long rtt;
 375         u32 val;
 376         int m;
 377
 378         if (sysctl_tcp_nometrics_save || !dst)
 379                 return;
 380
 381         if (dst->flags & DST_HOST)
 382                 dst_confirm(dst);
 383
 384         rcu_read_lock();
 385         if (icsk->icsk_backoff || !tp->srtt_us) {
 386                 /* This session failed to estimate rtt. Why?
 387                  * Probably, no packets returned in time.  Reset our
 388                  * results.
 389                  */
 390                 tm = tcp_get_metrics(sk, dst, false);
 391                 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
 392                         tcp_metric_set(tm, TCP_METRIC_RTT, 0);
 393                 goto out_unlock;
 394         } else
 395                 tm = tcp_get_metrics(sk, dst, true);
 396
 397         if (!tm)
 398                 goto out_unlock;
 399
 400         rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
 401         m = rtt - tp->srtt_us;
 402
 403         /* If newly calculated rtt larger than stored one, store new
 404          * one. Otherwise, use EWMA. Remember, rtt overestimation is
 405          * always better than underestimation.
 406          */
 407         if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
 408                 if (m <= 0)
 409                         rtt = tp->srtt_us;
 410                 else
 411                         rtt -= (m >> 3);
 412                 tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
 413         }
 414
 415         if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
 416                 unsigned long var;
 417
 418                 if (m < 0)
 419                         m = -m;
 420
 421                 /* Scale deviation to rttvar fixed point */
 422                 m >>= 1;
 423                 if (m < tp->mdev_us)
 424                         m = tp->mdev_us;
 425
 426                 var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
 427                 if (m >= var)
 428                         var = m;
 429                 else
 430                         var -= (var - m) >> 2;
 431
 432                 tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
 433         }
 434
 435         if (tcp_in_initial_slowstart(tp)) {
 436                 /* Slow start still did not finish. */
 437                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 438                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 439                         if (val && (tp->snd_cwnd >> 1) > val)
 440                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 441                                                tp->snd_cwnd >> 1);
 442                 }
 443                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 444                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 445                         if (tp->snd_cwnd > val)
 446                                 tcp_metric_set(tm, TCP_METRIC_CWND,
 447                                                tp->snd_cwnd);
 448                 }
 449         } else if (!tcp_in_slow_start(tp) &&
 450                    icsk->icsk_ca_state == TCP_CA_Open) {
 451                 /* Cong. avoidance phase, cwnd is reliable. */
 452                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
 453                         tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 454                                        max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
 455                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 456                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 457                         tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
 458                 }
 459         } else {
 460                 /* Else slow start did not finish, cwnd is non-sense,
 461                  * ssthresh may be also invalid.
 462                  */
 463                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 464                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 465                         tcp_metric_set(tm, TCP_METRIC_CWND,
 466                                        (val + tp->snd_ssthresh) >> 1);
 467                 }
 468                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 469                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 470                         if (val && tp->snd_ssthresh > val)
 471                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 472                                                tp->snd_ssthresh);
 473                 }
 474                 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
 475                         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 476                         if (val < tp->reordering &&
 477                             tp->reordering != net->ipv4.sysctl_tcp_reordering)
 478                                 tcp_metric_set(tm, TCP_METRIC_REORDERING,
 479                                                tp->reordering);
 480                 }
 481         }
 482         tm->tcpm_stamp = jiffies;
 483 out_unlock:
 484         rcu_read_unlock();
 485 }
 486
 487 /* Initialize metrics on socket. */
 488
 489 void tcp_init_metrics(struct sock *sk)
 490 {
 491         struct dst_entry *dst = __sk_dst_get(sk);
 492         struct tcp_sock *tp = tcp_sk(sk);
 493         struct tcp_metrics_block *tm;
 494         u32 val, crtt = 0; /* cached RTT scaled by 8 */
 495
 496         if (!dst)
 497                 goto reset;
 498
 499         dst_confirm(dst);
 500
 501         rcu_read_lock();
 502         tm = tcp_get_metrics(sk, dst, true);
 503         if (!tm) {
 504                 rcu_read_unlock();
 505                 goto reset;
 506         }
 507
 508         if (tcp_metric_locked(tm, TCP_METRIC_CWND))
 509                 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
 510
 511         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 512         if (val) {
 513                 tp->snd_ssthresh = val;
 514                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
 515                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
 516         } else {
 517                 /* ssthresh may have been reduced unnecessarily during.
 518                  * 3WHS. Restore it back to its initial default.
 519                  */
 520                 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 521         }
 522         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 523         if (val && tp->reordering != val) {
 524                 tcp_disable_fack(tp);
 525                 tcp_disable_early_retrans(tp);
 526                 tp->reordering = val;
 527         }
 528
 529         crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
 530         rcu_read_unlock();
 531 reset:
 532         /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
 533          * to seed the RTO for later data packets because SYN packets are
 534          * small. Use the per-dst cached values to seed the RTO but keep
 535          * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
 536          * Later the RTO will be updated immediately upon obtaining the first
 537          * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
 538          * influences the first RTO but not later RTT estimation.
 539          *
 540          * But if RTT is not available from the SYN (due to retransmits or
 541          * syn cookies) or the cache, force a conservative 3secs timeout.
 542          *
 543          * A bit of theory. RTT is time passed after "normal" sized packet
 544          * is sent until it is ACKed. In normal circumstances sending small
 545          * packets force peer to delay ACKs and calculation is correct too.
 546          * The algorithm is adaptive and, provided we follow specs, it
 547          * NEVER underestimate RTT. BUT! If peer tries to make some clever
 548          * tricks sort of "quick acks" for time long enough to decrease RTT
 549          * to low value, and then abruptly stops to do it and starts to delay
 550          * ACKs, wait for troubles.
 551          */
 552         if (crtt > tp->srtt_us) {
 553                 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
 554                 crtt /= 8 * USEC_PER_SEC / HZ;
 555                 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
 556         } else if (tp->srtt_us == 0) {
 557                 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
 558                  * 3WHS. This is most likely due to retransmission,
 559                  * including spurious one. Reset the RTO back to 3secs
 560                  * from the more aggressive 1sec to avoid more spurious
 561                  * retransmission.
 562                  */
 563                 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
 564                 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
 565
 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,
 581                         bool paws_check, bool timestamps)
 582 {
 583         struct tcp_metrics_block *tm;
 584         bool ret;
 585
 586         if (!dst)
 587                 return false;
 588
 589         rcu_read_lock();
 590         tm = __tcp_get_metrics_req(req, dst);
 591         if (paws_check) {
 592                 if (tm &&
 593                     (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
 594                     ((s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW ||
 595                      !timestamps))
 596                         ret = false;
 597                 else
 598                         ret = true;
 599         } else {
 600                 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
 601                         ret = true;
 602                 else
 603                         ret = false;
 604         }
 605         rcu_read_unlock();
 606
 607         return ret;
 608 }
 609 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
 610
 611 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
 612 {
 613         struct tcp_metrics_block *tm;
 614
 615         rcu_read_lock();
 616         tm = tcp_get_metrics(sk, dst, true);
 617         if (tm) {
 618                 struct tcp_sock *tp = tcp_sk(sk);
 619
 620                 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
 621                         tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
 622                         tp->rx_opt.ts_recent = tm->tcpm_ts;
 623                 }
 624         }
 625         rcu_read_unlock();
 626 }
 627 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
 628
 629 /* VJ's idea. Save last timestamp seen from this destination and hold
 630  * it at least for normal timewait interval to use for duplicate
 631  * segment detection in subsequent connections, before they enter
 632  * synchronized state.
 633  */
 634 bool tcp_remember_stamp(struct sock *sk)
 635 {
 636         struct dst_entry *dst = __sk_dst_get(sk);
 637         bool ret = false;
 638
 639         if (dst) {
 640                 struct tcp_metrics_block *tm;
 641
 642                 rcu_read_lock();
 643                 tm = tcp_get_metrics(sk, dst, true);
 644                 if (tm) {
 645                         struct tcp_sock *tp = tcp_sk(sk);
 646
 647                         if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
 648                             ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 649                              tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
 650                                 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
 651                                 tm->tcpm_ts = tp->rx_opt.ts_recent;
 652                         }
 653                         ret = true;
 654                 }
 655                 rcu_read_unlock();
 656         }
 657         return ret;
 658 }
 659
 660 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
 661 {
 662         struct tcp_metrics_block *tm;
 663         bool ret = false;
 664
 665         rcu_read_lock();
 666         tm = __tcp_get_metrics_tw(tw);
 667         if (tm) {
 668                 const struct tcp_timewait_sock *tcptw;
 669                 struct sock *sk = (struct sock *) tw;
 670
 671                 tcptw = tcp_twsk(sk);
 672                 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
 673                     ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 674                      tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
 675                         tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
 676                         tm->tcpm_ts        = tcptw->tw_ts_recent;
 677                 }
 678                 ret = true;
 679         }
 680         rcu_read_unlock();
 681
 682         return ret;
 683 }
 684
 685 static DEFINE_SEQLOCK(fastopen_seqlock);
 686
 687 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
 688                             struct tcp_fastopen_cookie *cookie,
 689                             int *syn_loss, unsigned long *last_syn_loss)
 690 {
 691         struct tcp_metrics_block *tm;
 692
 693         rcu_read_lock();
 694         tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
 695         if (tm) {
 696                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 697                 unsigned int seq;
 698
 699                 do {
 700                         seq = read_seqbegin(&fastopen_seqlock);
 701                         if (tfom->mss)
 702                                 *mss = tfom->mss;
 703                         *cookie = tfom->cookie;
 704                         if (cookie->len <= 0 && tfom->try_exp == 1)
 705                                 cookie->exp = true;
 706                         *syn_loss = tfom->syn_loss;
 707                         *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
 708                 } while (read_seqretry(&fastopen_seqlock, seq));
 709         }
 710         rcu_read_unlock();
 711 }
 712
 713 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
 714                             struct tcp_fastopen_cookie *cookie, bool syn_lost,
 715                             u16 try_exp)
 716 {
 717         struct dst_entry *dst = __sk_dst_get(sk);
 718         struct tcp_metrics_block *tm;
 719
 720         if (!dst)
 721                 return;
 722         rcu_read_lock();
 723         tm = tcp_get_metrics(sk, dst, true);
 724         if (tm) {
 725                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 726
 727                 write_seqlock_bh(&fastopen_seqlock);
 728                 if (mss)
 729                         tfom->mss = mss;
 730                 if (cookie && cookie->len > 0)
 731                         tfom->cookie = *cookie;
 732                 else if (try_exp > tfom->try_exp &&
 733                          tfom->cookie.len <= 0 && !tfom->cookie.exp)
 734                         tfom->try_exp = try_exp;
 735                 if (syn_lost) {
 736                         ++tfom->syn_loss;
 737                         tfom->last_syn_loss = jiffies;
 738                 } else
 739                         tfom->syn_loss = 0;
 740                 write_sequnlock_bh(&fastopen_seqlock);
 741         }
 742         rcu_read_unlock();
 743 }
 744
 745 static struct genl_family tcp_metrics_nl_family;
 746
 747 static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
 748         [TCP_METRICS_ATTR_ADDR_IPV4]    = { .type = NLA_U32, },
 749         [TCP_METRICS_ATTR_ADDR_IPV6]    = { .type = NLA_BINARY,
 750                                             .len = sizeof(struct in6_addr), },
 751         /* Following attributes are not received for GET/DEL,
 752          * we keep them for reference
 753          */
 754 #if 0
 755         [TCP_METRICS_ATTR_AGE]          = { .type = NLA_MSECS, },
 756         [TCP_METRICS_ATTR_TW_TSVAL]     = { .type = NLA_U32, },
 757         [TCP_METRICS_ATTR_TW_TS_STAMP]  = { .type = NLA_S32, },
 758         [TCP_METRICS_ATTR_VALS]         = { .type = NLA_NESTED, },
 759         [TCP_METRICS_ATTR_FOPEN_MSS]    = { .type = NLA_U16, },
 760         [TCP_METRICS_ATTR_FOPEN_SYN_DROPS]      = { .type = NLA_U16, },
 761         [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]    = { .type = NLA_MSECS, },
 762         [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
 763                                             .len = TCP_FASTOPEN_COOKIE_MAX, },
 764 #endif
 765 };
 766
 767 /* Add attributes, caller cancels its header on failure */
 768 static int tcp_metrics_fill_info(struct sk_buff *msg,
 769                                  struct tcp_metrics_block *tm)
 770 {
 771         struct nlattr *nest;
 772         int i;
 773
 774         switch (tm->tcpm_daddr.family) {
 775         case AF_INET:
 776                 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
 777                                     inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
 778                         goto nla_put_failure;
 779                 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
 780                                     inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
 781                         goto nla_put_failure;
 782                 break;
 783         case AF_INET6:
 784                 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
 785                                      inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
 786                         goto nla_put_failure;
 787                 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
 788                                      inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
 789                         goto nla_put_failure;
 790                 break;
 791         default:
 792                 return -EAFNOSUPPORT;
 793         }
 794
 795         if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
 796                           jiffies - tm->tcpm_stamp,
 797                           TCP_METRICS_ATTR_PAD) < 0)
 798                 goto nla_put_failure;
 799         if (tm->tcpm_ts_stamp) {
 800                 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
 801                                 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
 802                         goto nla_put_failure;
 803                 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
 804                                 tm->tcpm_ts) < 0)
 805                         goto nla_put_failure;
 806         }
 807
 808         {
 809                 int n = 0;
 810
 811                 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
 812                 if (!nest)
 813                         goto nla_put_failure;
 814                 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
 815                         u32 val = tm->tcpm_vals[i];
 816
 817                         if (!val)
 818                                 continue;
 819                         if (i == TCP_METRIC_RTT) {
 820                                 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
 821                                                 val) < 0)
 822                                         goto nla_put_failure;
 823                                 n++;
 824                                 val = max(val / 1000, 1U);
 825                         }
 826                         if (i == TCP_METRIC_RTTVAR) {
 827                                 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
 828                                                 val) < 0)
 829                                         goto nla_put_failure;
 830                                 n++;
 831                                 val = max(val / 1000, 1U);
 832                         }
 833                         if (nla_put_u32(msg, i + 1, val) < 0)
 834                                 goto nla_put_failure;
 835                         n++;
 836                 }
 837                 if (n)
 838                         nla_nest_end(msg, nest);
 839                 else
 840                         nla_nest_cancel(msg, nest);
 841         }
 842
 843         {
 844                 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
 845                 unsigned int seq;
 846
 847                 do {
 848                         seq = read_seqbegin(&fastopen_seqlock);
 849                         tfom_copy[0] = tm->tcpm_fastopen;
 850                 } while (read_seqretry(&fastopen_seqlock, seq));
 851
 852                 tfom = tfom_copy;
 853                 if (tfom->mss &&
 854                     nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
 855                                 tfom->mss) < 0)
 856                         goto nla_put_failure;
 857                 if (tfom->syn_loss &&
 858                     (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
 859                                 tfom->syn_loss) < 0 ||
 860                      nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
 861                                 jiffies - tfom->last_syn_loss,
 862                                 TCP_METRICS_ATTR_PAD) < 0))
 863                         goto nla_put_failure;
 864                 if (tfom->cookie.len > 0 &&
 865                     nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
 866                             tfom->cookie.len, tfom->cookie.val) < 0)
 867                         goto nla_put_failure;
 868         }
 869
 870         return 0;
 871
 872 nla_put_failure:
 873         return -EMSGSIZE;
 874 }
 875
 876 static int tcp_metrics_dump_info(struct sk_buff *skb,
 877                                  struct netlink_callback *cb,
 878                                  struct tcp_metrics_block *tm)
 879 {
 880         void *hdr;
 881
 882         hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 883                           &tcp_metrics_nl_family, NLM_F_MULTI,
 884                           TCP_METRICS_CMD_GET);
 885         if (!hdr)
 886                 return -EMSGSIZE;
 887
 888         if (tcp_metrics_fill_info(skb, tm) < 0)
 889                 goto nla_put_failure;
 890
 891         genlmsg_end(skb, hdr);
 892         return 0;
 893
 894 nla_put_failure:
 895         genlmsg_cancel(skb, hdr);
 896         return -EMSGSIZE;
 897 }
 898
 899 static int tcp_metrics_nl_dump(struct sk_buff *skb,
 900                                struct netlink_callback *cb)
 901 {
 902         struct net *net = sock_net(skb->sk);
 903         unsigned int max_rows = 1U << tcp_metrics_hash_log;
 904         unsigned int row, s_row = cb->args[0];
 905         int s_col = cb->args[1], col = s_col;
 906
 907         for (row = s_row; row < max_rows; row++, s_col = 0) {
 908                 struct tcp_metrics_block *tm;
 909                 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
 910
 911                 rcu_read_lock();
 912                 for (col = 0, tm = rcu_dereference(hb->chain); tm;
 913                      tm = rcu_dereference(tm->tcpm_next), col++) {
 914                         if (!net_eq(tm_net(tm), net))
 915                                 continue;
 916                         if (col < s_col)
 917                                 continue;
 918                         if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
 919                                 rcu_read_unlock();
 920                                 goto done;
 921                         }
 922                 }
 923                 rcu_read_unlock();
 924         }
 925
 926 done:
 927         cb->args[0] = row;
 928         cb->args[1] = col;
 929         return skb->len;
 930 }
 931
 932 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 933                            unsigned int *hash, int optional, int v4, int v6)
 934 {
 935         struct nlattr *a;
 936
 937         a = info->attrs[v4];
 938         if (a) {
 939                 inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
 940                 if (hash)
 941                         *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
 942                 return 0;
 943         }
 944         a = info->attrs[v6];
 945         if (a) {
 946                 struct in6_addr in6;
 947
 948                 if (nla_len(a) != sizeof(struct in6_addr))
 949                         return -EINVAL;
 950                 in6 = nla_get_in6_addr(a);
 951                 inetpeer_set_addr_v6(addr, &in6);
 952                 if (hash)
 953                         *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
 954                 return 0;
 955         }
 956         return optional ? 1 : -EAFNOSUPPORT;
 957 }
 958
 959 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 960                          unsigned int *hash, int optional)
 961 {
 962         return __parse_nl_addr(info, addr, hash, optional,
 963                                TCP_METRICS_ATTR_ADDR_IPV4,
 964                                TCP_METRICS_ATTR_ADDR_IPV6);
 965 }
 966
 967 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
 968 {
 969         return __parse_nl_addr(info, addr, NULL, 0,
 970                                TCP_METRICS_ATTR_SADDR_IPV4,
 971                                TCP_METRICS_ATTR_SADDR_IPV6);
 972 }
 973
 974 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
 975 {
 976         struct tcp_metrics_block *tm;
 977         struct inetpeer_addr saddr, daddr;
 978         unsigned int hash;
 979         struct sk_buff *msg;
 980         struct net *net = genl_info_net(info);
 981         void *reply;
 982         int ret;
 983         bool src = true;
 984
 985         ret = parse_nl_addr(info, &daddr, &hash, 0);
 986         if (ret < 0)
 987                 return ret;
 988
 989         ret = parse_nl_saddr(info, &saddr);
 990         if (ret < 0)
 991                 src = false;
 992
 993         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 994         if (!msg)
 995                 return -ENOMEM;
 996
 997         reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
 998                                   info->genlhdr->cmd);
 999         if (!reply)
1000                 goto nla_put_failure;
1001
1002         hash ^= net_hash_mix(net);
1003         hash = hash_32(hash, tcp_metrics_hash_log);
1004         ret = -ESRCH;
1005         rcu_read_lock();
1006         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
1007              tm = rcu_dereference(tm->tcpm_next)) {
1008                 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1009                     (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1010                     net_eq(tm_net(tm), net)) {
1011                         ret = tcp_metrics_fill_info(msg, tm);
1012                         break;
1013                 }
1014         }
1015         rcu_read_unlock();
1016         if (ret < 0)
1017                 goto out_free;
1018
1019         genlmsg_end(msg, reply);
1020         return genlmsg_reply(msg, info);
1021
1022 nla_put_failure:
1023         ret = -EMSGSIZE;
1024
1025 out_free:
1026         nlmsg_free(msg);
1027         return ret;
1028 }
1029
1030 static void tcp_metrics_flush_all(struct net *net)
1031 {
1032         unsigned int max_rows = 1U << tcp_metrics_hash_log;
1033         struct tcpm_hash_bucket *hb = tcp_metrics_hash;
1034         struct tcp_metrics_block *tm;
1035         unsigned int row;
1036
1037         for (row = 0; row < max_rows; row++, hb++) {
1038                 struct tcp_metrics_block __rcu **pp;
1039                 spin_lock_bh(&tcp_metrics_lock);
1040                 pp = &hb->chain;
1041                 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1042                         if (net_eq(tm_net(tm), net)) {
1043                                 *pp = tm->tcpm_next;
1044                                 kfree_rcu(tm, rcu_head);
1045                         } else {
1046                                 pp = &tm->tcpm_next;
1047                         }
1048                 }
1049                 spin_unlock_bh(&tcp_metrics_lock);
1050         }
1051 }
1052
1053 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
1054 {
1055         struct tcpm_hash_bucket *hb;
1056         struct tcp_metrics_block *tm;
1057         struct tcp_metrics_block __rcu **pp;
1058         struct inetpeer_addr saddr, daddr;
1059         unsigned int hash;
1060         struct net *net = genl_info_net(info);
1061         int ret;
1062         bool src = true, found = false;
1063
1064         ret = parse_nl_addr(info, &daddr, &hash, 1);
1065         if (ret < 0)
1066                 return ret;
1067         if (ret > 0) {
1068                 tcp_metrics_flush_all(net);
1069                 return 0;
1070         }
1071         ret = parse_nl_saddr(info, &saddr);
1072         if (ret < 0)
1073                 src = false;
1074
1075         hash ^= net_hash_mix(net);
1076         hash = hash_32(hash, tcp_metrics_hash_log);
1077         hb = tcp_metrics_hash + hash;
1078         pp = &hb->chain;
1079         spin_lock_bh(&tcp_metrics_lock);
1080         for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
1081                 if (addr_same(&tm->tcpm_daddr, &daddr) &&
1082                     (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
1083                     net_eq(tm_net(tm), net)) {
1084                         *pp = tm->tcpm_next;
1085                         kfree_rcu(tm, rcu_head);
1086                         found = true;
1087                 } else {
1088                         pp = &tm->tcpm_next;
1089                 }
1090         }
1091         spin_unlock_bh(&tcp_metrics_lock);
1092         if (!found)
1093                 return -ESRCH;
1094         return 0;
1095 }
1096
1097 static const struct genl_ops tcp_metrics_nl_ops[] = {
1098         {
1099                 .cmd = TCP_METRICS_CMD_GET,
1100                 .doit = tcp_metrics_nl_cmd_get,
1101                 .dumpit = tcp_metrics_nl_dump,
1102                 .policy = tcp_metrics_nl_policy,
1103         },
1104         {
1105                 .cmd = TCP_METRICS_CMD_DEL,
1106                 .doit = tcp_metrics_nl_cmd_del,
1107                 .policy = tcp_metrics_nl_policy,
1108                 .flags = GENL_ADMIN_PERM,
1109         },
1110 };
1111
1112 static struct genl_family tcp_metrics_nl_family __ro_after_init = {
1113         .hdrsize        = 0,
1114         .name           = TCP_METRICS_GENL_NAME,
1115         .version        = TCP_METRICS_GENL_VERSION,
1116         .maxattr        = TCP_METRICS_ATTR_MAX,
1117         .netnsok        = true,
1118         .module         = THIS_MODULE,
1119         .ops            = tcp_metrics_nl_ops,
1120         .n_ops          = ARRAY_SIZE(tcp_metrics_nl_ops),
1121 };
1122
1123 static unsigned int tcpmhash_entries;
1124 static int __init set_tcpmhash_entries(char *str)
1125 {
1126         ssize_t ret;
1127
1128         if (!str)
1129                 return 0;
1130
1131         ret = kstrtouint(str, 0, &tcpmhash_entries);
1132         if (ret)
1133                 return 0;
1134
1135         return 1;
1136 }
1137 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1138
1139 static int __net_init tcp_net_metrics_init(struct net *net)
1140 {
1141         size_t size;
1142         unsigned int slots;
1143
1144         if (!net_eq(net, &init_net))
1145                 return 0;
1146
1147         slots = tcpmhash_entries;
1148         if (!slots) {
1149                 if (totalram_pages >= 128 * 1024)
1150                         slots = 16 * 1024;
1151                 else
1152                         slots = 8 * 1024;
1153         }
1154
1155         tcp_metrics_hash_log = order_base_2(slots);
1156         size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1157
1158         tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1159         if (!tcp_metrics_hash)
1160                 tcp_metrics_hash = vzalloc(size);
1161
1162         if (!tcp_metrics_hash)
1163                 return -ENOMEM;
1164
1165         return 0;
1166 }
1167
1168 static void __net_exit tcp_net_metrics_exit(struct net *net)
1169 {
1170         tcp_metrics_flush_all(net);
1171 }
1172
1173 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1174         .init   =       tcp_net_metrics_init,
1175         .exit   =       tcp_net_metrics_exit,
1176 };
1177
1178 void __init tcp_metrics_init(void)
1179 {
1180         int ret;
1181
1182         ret = register_pernet_subsys(&tcp_net_metrics_ops);
1183         if (ret < 0)
1184                 panic("Could not allocate the tcp_metrics hash table\n");
1185
1186         ret = genl_register_family(&tcp_metrics_nl_family);
1187         if (ret < 0)
1188                 panic("Could not register tcp_metrics generic netlink\n");
1189 }