net/ipv4/tcp_metrics.c

   1 #include <linux/rcupdate.h>
   2 #include <linux/spinlock.h>
   3 #include <linux/jiffies.h>
   4 #include <linux/module.h>
   5 #include <linux/cache.h>
   6 #include <linux/slab.h>
   7 #include <linux/init.h>
   8 #include <linux/tcp.h>
   9 #include <linux/hash.h>
  10 #include <linux/tcp_metrics.h>
  11 #include <linux/vmalloc.h>
  12
  13 #include <net/inet_connection_sock.h>
  14 #include <net/net_namespace.h>
  15 #include <net/request_sock.h>
  16 #include <net/inetpeer.h>
  17 #include <net/sock.h>
  18 #include <net/ipv6.h>
  19 #include <net/dst.h>
  20 #include <net/tcp.h>
  21 #include <net/genetlink.h>
  22
  23 int sysctl_tcp_nometrics_save __read_mostly;
  24
  25 struct tcp_fastopen_metrics {
  26         u16     mss;
  27         u16     syn_loss:10;            /* Recurring Fast Open SYN losses */
  28         unsigned long   last_syn_loss;  /* Last Fast Open SYN loss */
  29         struct  tcp_fastopen_cookie     cookie;
  30 };
  31
  32 struct tcp_metrics_block {
  33         struct tcp_metrics_block __rcu  *tcpm_next;
  34         struct inetpeer_addr            tcpm_addr;
  35         unsigned long                   tcpm_stamp;
  36         u32                             tcpm_ts;
  37         u32                             tcpm_ts_stamp;
  38         u32                             tcpm_lock;
  39         u32                             tcpm_vals[TCP_METRIC_MAX + 1];
  40         struct tcp_fastopen_metrics     tcpm_fastopen;
  41
  42         struct rcu_head                 rcu_head;
  43 };
  44
  45 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
  46                               enum tcp_metric_index idx)
  47 {
  48         return tm->tcpm_lock & (1 << idx);
  49 }
  50
  51 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
  52                           enum tcp_metric_index idx)
  53 {
  54         return tm->tcpm_vals[idx];
  55 }
  56
  57 static u32 tcp_metric_get_jiffies(struct tcp_metrics_block *tm,
  58                                   enum tcp_metric_index idx)
  59 {
  60         return msecs_to_jiffies(tm->tcpm_vals[idx]);
  61 }
  62
  63 static void tcp_metric_set(struct tcp_metrics_block *tm,
  64                            enum tcp_metric_index idx,
  65                            u32 val)
  66 {
  67         tm->tcpm_vals[idx] = val;
  68 }
  69
  70 static void tcp_metric_set_msecs(struct tcp_metrics_block *tm,
  71                                  enum tcp_metric_index idx,
  72                                  u32 val)
  73 {
  74         tm->tcpm_vals[idx] = jiffies_to_msecs(val);
  75 }
  76
  77 static bool addr_same(const struct inetpeer_addr *a,
  78                       const struct inetpeer_addr *b)
  79 {
  80         const struct in6_addr *a6, *b6;
  81
  82         if (a->family != b->family)
  83                 return false;
  84         if (a->family == AF_INET)
  85                 return a->addr.a4 == b->addr.a4;
  86
  87         a6 = (const struct in6_addr *) &a->addr.a6[0];
  88         b6 = (const struct in6_addr *) &b->addr.a6[0];
  89
  90         return ipv6_addr_equal(a6, b6);
  91 }
  92
  93 struct tcpm_hash_bucket {
  94         struct tcp_metrics_block __rcu  *chain;
  95 };
  96
  97 static DEFINE_SPINLOCK(tcp_metrics_lock);
  98
  99 static void tcpm_suck_dst(struct tcp_metrics_block *tm, struct dst_entry *dst,
 100                           bool fastopen_clear)
 101 {
 102         u32 val;
 103
 104         tm->tcpm_stamp = jiffies;
 105
 106         val = 0;
 107         if (dst_metric_locked(dst, RTAX_RTT))
 108                 val |= 1 << TCP_METRIC_RTT;
 109         if (dst_metric_locked(dst, RTAX_RTTVAR))
 110                 val |= 1 << TCP_METRIC_RTTVAR;
 111         if (dst_metric_locked(dst, RTAX_SSTHRESH))
 112                 val |= 1 << TCP_METRIC_SSTHRESH;
 113         if (dst_metric_locked(dst, RTAX_CWND))
 114                 val |= 1 << TCP_METRIC_CWND;
 115         if (dst_metric_locked(dst, RTAX_REORDERING))
 116                 val |= 1 << TCP_METRIC_REORDERING;
 117         tm->tcpm_lock = val;
 118
 119         tm->tcpm_vals[TCP_METRIC_RTT] = dst_metric_raw(dst, RTAX_RTT);
 120         tm->tcpm_vals[TCP_METRIC_RTTVAR] = dst_metric_raw(dst, RTAX_RTTVAR);
 121         tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
 122         tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
 123         tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
 124         tm->tcpm_ts = 0;
 125         tm->tcpm_ts_stamp = 0;
 126         if (fastopen_clear) {
 127                 tm->tcpm_fastopen.mss = 0;
 128                 tm->tcpm_fastopen.syn_loss = 0;
 129                 tm->tcpm_fastopen.cookie.len = 0;
 130         }
 131 }
 132
 133 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
 134                                           struct inetpeer_addr *addr,
 135                                           unsigned int hash,
 136                                           bool reclaim)
 137 {
 138         struct tcp_metrics_block *tm;
 139         struct net *net;
 140
 141         spin_lock_bh(&tcp_metrics_lock);
 142         net = dev_net(dst->dev);
 143         if (unlikely(reclaim)) {
 144                 struct tcp_metrics_block *oldest;
 145
 146                 oldest = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain);
 147                 for (tm = rcu_dereference(oldest->tcpm_next); tm;
 148                      tm = rcu_dereference(tm->tcpm_next)) {
 149                         if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
 150                                 oldest = tm;
 151                 }
 152                 tm = oldest;
 153         } else {
 154                 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
 155                 if (!tm)
 156                         goto out_unlock;
 157         }
 158         tm->tcpm_addr = *addr;
 159
 160         tcpm_suck_dst(tm, dst, true);
 161
 162         if (likely(!reclaim)) {
 163                 tm->tcpm_next = net->ipv4.tcp_metrics_hash[hash].chain;
 164                 rcu_assign_pointer(net->ipv4.tcp_metrics_hash[hash].chain, tm);
 165         }
 166
 167 out_unlock:
 168         spin_unlock_bh(&tcp_metrics_lock);
 169         return tm;
 170 }
 171
 172 #define TCP_METRICS_TIMEOUT             (60 * 60 * HZ)
 173
 174 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
 175 {
 176         if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
 177                 tcpm_suck_dst(tm, dst, false);
 178 }
 179
 180 #define TCP_METRICS_RECLAIM_DEPTH       5
 181 #define TCP_METRICS_RECLAIM_PTR         (struct tcp_metrics_block *) 0x1UL
 182
 183 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
 184 {
 185         if (tm)
 186                 return tm;
 187         if (depth > TCP_METRICS_RECLAIM_DEPTH)
 188                 return TCP_METRICS_RECLAIM_PTR;
 189         return NULL;
 190 }
 191
 192 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *addr,
 193                                                    struct net *net, unsigned int hash)
 194 {
 195         struct tcp_metrics_block *tm;
 196         int depth = 0;
 197
 198         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 199              tm = rcu_dereference(tm->tcpm_next)) {
 200                 if (addr_same(&tm->tcpm_addr, addr))
 201                         break;
 202                 depth++;
 203         }
 204         return tcp_get_encode(tm, depth);
 205 }
 206
 207 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
 208                                                        struct dst_entry *dst)
 209 {
 210         struct tcp_metrics_block *tm;
 211         struct inetpeer_addr addr;
 212         unsigned int hash;
 213         struct net *net;
 214
 215         addr.family = req->rsk_ops->family;
 216         switch (addr.family) {
 217         case AF_INET:
 218                 addr.addr.a4 = inet_rsk(req)->rmt_addr;
 219                 hash = (__force unsigned int) addr.addr.a4;
 220                 break;
 221         case AF_INET6:
 222                 *(struct in6_addr *)addr.addr.a6 = inet6_rsk(req)->rmt_addr;
 223                 hash = ipv6_addr_hash(&inet6_rsk(req)->rmt_addr);
 224                 break;
 225         default:
 226                 return NULL;
 227         }
 228
 229         net = dev_net(dst->dev);
 230         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 231
 232         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 233              tm = rcu_dereference(tm->tcpm_next)) {
 234                 if (addr_same(&tm->tcpm_addr, &addr))
 235                         break;
 236         }
 237         tcpm_check_stamp(tm, dst);
 238         return tm;
 239 }
 240
 241 static struct tcp_metrics_block *__tcp_get_metrics_tw(struct inet_timewait_sock *tw)
 242 {
 243         struct inet6_timewait_sock *tw6;
 244         struct tcp_metrics_block *tm;
 245         struct inetpeer_addr addr;
 246         unsigned int hash;
 247         struct net *net;
 248
 249         addr.family = tw->tw_family;
 250         switch (addr.family) {
 251         case AF_INET:
 252                 addr.addr.a4 = tw->tw_daddr;
 253                 hash = (__force unsigned int) addr.addr.a4;
 254                 break;
 255         case AF_INET6:
 256                 tw6 = inet6_twsk((struct sock *)tw);
 257                 *(struct in6_addr *)addr.addr.a6 = tw6->tw_v6_daddr;
 258                 hash = ipv6_addr_hash(&tw6->tw_v6_daddr);
 259                 break;
 260         default:
 261                 return NULL;
 262         }
 263
 264         net = twsk_net(tw);
 265         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 266
 267         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 268              tm = rcu_dereference(tm->tcpm_next)) {
 269                 if (addr_same(&tm->tcpm_addr, &addr))
 270                         break;
 271         }
 272         return tm;
 273 }
 274
 275 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
 276                                                  struct dst_entry *dst,
 277                                                  bool create)
 278 {
 279         struct tcp_metrics_block *tm;
 280         struct inetpeer_addr addr;
 281         unsigned int hash;
 282         struct net *net;
 283         bool reclaim;
 284
 285         addr.family = sk->sk_family;
 286         switch (addr.family) {
 287         case AF_INET:
 288                 addr.addr.a4 = inet_sk(sk)->inet_daddr;
 289                 hash = (__force unsigned int) addr.addr.a4;
 290                 break;
 291         case AF_INET6:
 292                 *(struct in6_addr *)addr.addr.a6 = inet6_sk(sk)->daddr;
 293                 hash = ipv6_addr_hash(&inet6_sk(sk)->daddr);
 294                 break;
 295         default:
 296                 return NULL;
 297         }
 298
 299         net = dev_net(dst->dev);
 300         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 301
 302         tm = __tcp_get_metrics(&addr, net, hash);
 303         reclaim = false;
 304         if (tm == TCP_METRICS_RECLAIM_PTR) {
 305                 reclaim = true;
 306                 tm = NULL;
 307         }
 308         if (!tm && create)
 309                 tm = tcpm_new(dst, &addr, hash, reclaim);
 310         else
 311                 tcpm_check_stamp(tm, dst);
 312
 313         return tm;
 314 }
 315
 316 /* Save metrics learned by this TCP session.  This function is called
 317  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
 318  * or goes from LAST-ACK to CLOSE.
 319  */
 320 void tcp_update_metrics(struct sock *sk)
 321 {
 322         const struct inet_connection_sock *icsk = inet_csk(sk);
 323         struct dst_entry *dst = __sk_dst_get(sk);
 324         struct tcp_sock *tp = tcp_sk(sk);
 325         struct tcp_metrics_block *tm;
 326         unsigned long rtt;
 327         u32 val;
 328         int m;
 329
 330         if (sysctl_tcp_nometrics_save || !dst)
 331                 return;
 332
 333         if (dst->flags & DST_HOST)
 334                 dst_confirm(dst);
 335
 336         rcu_read_lock();
 337         if (icsk->icsk_backoff || !tp->srtt) {
 338                 /* This session failed to estimate rtt. Why?
 339                  * Probably, no packets returned in time.  Reset our
 340                  * results.
 341                  */
 342                 tm = tcp_get_metrics(sk, dst, false);
 343                 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
 344                         tcp_metric_set(tm, TCP_METRIC_RTT, 0);
 345                 goto out_unlock;
 346         } else
 347                 tm = tcp_get_metrics(sk, dst, true);
 348
 349         if (!tm)
 350                 goto out_unlock;
 351
 352         rtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
 353         m = rtt - tp->srtt;
 354
 355         /* If newly calculated rtt larger than stored one, store new
 356          * one. Otherwise, use EWMA. Remember, rtt overestimation is
 357          * always better than underestimation.
 358          */
 359         if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
 360                 if (m <= 0)
 361                         rtt = tp->srtt;
 362                 else
 363                         rtt -= (m >> 3);
 364                 tcp_metric_set_msecs(tm, TCP_METRIC_RTT, rtt);
 365         }
 366
 367         if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
 368                 unsigned long var;
 369
 370                 if (m < 0)
 371                         m = -m;
 372
 373                 /* Scale deviation to rttvar fixed point */
 374                 m >>= 1;
 375                 if (m < tp->mdev)
 376                         m = tp->mdev;
 377
 378                 var = tcp_metric_get_jiffies(tm, TCP_METRIC_RTTVAR);
 379                 if (m >= var)
 380                         var = m;
 381                 else
 382                         var -= (var - m) >> 2;
 383
 384                 tcp_metric_set_msecs(tm, TCP_METRIC_RTTVAR, var);
 385         }
 386
 387         if (tcp_in_initial_slowstart(tp)) {
 388                 /* Slow start still did not finish. */
 389                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 390                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 391                         if (val && (tp->snd_cwnd >> 1) > val)
 392                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 393                                                tp->snd_cwnd >> 1);
 394                 }
 395                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 396                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 397                         if (tp->snd_cwnd > val)
 398                                 tcp_metric_set(tm, TCP_METRIC_CWND,
 399                                                tp->snd_cwnd);
 400                 }
 401         } else if (tp->snd_cwnd > tp->snd_ssthresh &&
 402                    icsk->icsk_ca_state == TCP_CA_Open) {
 403                 /* Cong. avoidance phase, cwnd is reliable. */
 404                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
 405                         tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 406                                        max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
 407                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 408                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 409                         tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
 410                 }
 411         } else {
 412                 /* Else slow start did not finish, cwnd is non-sense,
 413                  * ssthresh may be also invalid.
 414                  */
 415                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 416                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 417                         tcp_metric_set(tm, TCP_METRIC_CWND,
 418                                        (val + tp->snd_ssthresh) >> 1);
 419                 }
 420                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 421                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 422                         if (val && tp->snd_ssthresh > val)
 423                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 424                                                tp->snd_ssthresh);
 425                 }
 426                 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
 427                         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 428                         if (val < tp->reordering &&
 429                             tp->reordering != sysctl_tcp_reordering)
 430                                 tcp_metric_set(tm, TCP_METRIC_REORDERING,
 431                                                tp->reordering);
 432                 }
 433         }
 434         tm->tcpm_stamp = jiffies;
 435 out_unlock:
 436         rcu_read_unlock();
 437 }
 438
 439 /* Initialize metrics on socket. */
 440
 441 void tcp_init_metrics(struct sock *sk)
 442 {
 443         struct dst_entry *dst = __sk_dst_get(sk);
 444         struct tcp_sock *tp = tcp_sk(sk);
 445         struct tcp_metrics_block *tm;
 446         u32 val, crtt = 0; /* cached RTT scaled by 8 */
 447
 448         if (dst == NULL)
 449                 goto reset;
 450
 451         dst_confirm(dst);
 452
 453         rcu_read_lock();
 454         tm = tcp_get_metrics(sk, dst, true);
 455         if (!tm) {
 456                 rcu_read_unlock();
 457                 goto reset;
 458         }
 459
 460         if (tcp_metric_locked(tm, TCP_METRIC_CWND))
 461                 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
 462
 463         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 464         if (val) {
 465                 tp->snd_ssthresh = val;
 466                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
 467                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
 468         } else {
 469                 /* ssthresh may have been reduced unnecessarily during.
 470                  * 3WHS. Restore it back to its initial default.
 471                  */
 472                 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 473         }
 474         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 475         if (val && tp->reordering != val) {
 476                 tcp_disable_fack(tp);
 477                 tcp_disable_early_retrans(tp);
 478                 tp->reordering = val;
 479         }
 480
 481         crtt = tcp_metric_get_jiffies(tm, TCP_METRIC_RTT);
 482         rcu_read_unlock();
 483 reset:
 484         /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
 485          * to seed the RTO for later data packets because SYN packets are
 486          * small. Use the per-dst cached values to seed the RTO but keep
 487          * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
 488          * Later the RTO will be updated immediately upon obtaining the first
 489          * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
 490          * influences the first RTO but not later RTT estimation.
 491          *
 492          * But if RTT is not available from the SYN (due to retransmits or
 493          * syn cookies) or the cache, force a conservative 3secs timeout.
 494          *
 495          * A bit of theory. RTT is time passed after "normal" sized packet
 496          * is sent until it is ACKed. In normal circumstances sending small
 497          * packets force peer to delay ACKs and calculation is correct too.
 498          * The algorithm is adaptive and, provided we follow specs, it
 499          * NEVER underestimate RTT. BUT! If peer tries to make some clever
 500          * tricks sort of "quick acks" for time long enough to decrease RTT
 501          * to low value, and then abruptly stops to do it and starts to delay
 502          * ACKs, wait for troubles.
 503          */
 504         if (crtt > tp->srtt) {
 505                 inet_csk(sk)->icsk_rto = crtt + max(crtt >> 2, tcp_rto_min(sk));
 506         } else if (tp->srtt == 0) {
 507                 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
 508                  * 3WHS. This is most likely due to retransmission,
 509                  * including spurious one. Reset the RTO back to 3secs
 510                  * from the more aggressive 1sec to avoid more spurious
 511                  * retransmission.
 512                  */
 513                 tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK;
 514                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
 515         }
 516         /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
 517          * retransmitted. In light of RFC6298 more aggressive 1sec
 518          * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
 519          * retransmission has occurred.
 520          */
 521         if (tp->total_retrans > 1)
 522                 tp->snd_cwnd = 1;
 523         else
 524                 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
 525         tp->snd_cwnd_stamp = tcp_time_stamp;
 526 }
 527
 528 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst, bool paws_check)
 529 {
 530         struct tcp_metrics_block *tm;
 531         bool ret;
 532
 533         if (!dst)
 534                 return false;
 535
 536         rcu_read_lock();
 537         tm = __tcp_get_metrics_req(req, dst);
 538         if (paws_check) {
 539                 if (tm &&
 540                     (u32)get_seconds() - tm->tcpm_ts_stamp < TCP_PAWS_MSL &&
 541                     (s32)(tm->tcpm_ts - req->ts_recent) > TCP_PAWS_WINDOW)
 542                         ret = false;
 543                 else
 544                         ret = true;
 545         } else {
 546                 if (tm && tcp_metric_get(tm, TCP_METRIC_RTT) && tm->tcpm_ts_stamp)
 547                         ret = true;
 548                 else
 549                         ret = false;
 550         }
 551         rcu_read_unlock();
 552
 553         return ret;
 554 }
 555 EXPORT_SYMBOL_GPL(tcp_peer_is_proven);
 556
 557 void tcp_fetch_timewait_stamp(struct sock *sk, struct dst_entry *dst)
 558 {
 559         struct tcp_metrics_block *tm;
 560
 561         rcu_read_lock();
 562         tm = tcp_get_metrics(sk, dst, true);
 563         if (tm) {
 564                 struct tcp_sock *tp = tcp_sk(sk);
 565
 566                 if ((u32)get_seconds() - tm->tcpm_ts_stamp <= TCP_PAWS_MSL) {
 567                         tp->rx_opt.ts_recent_stamp = tm->tcpm_ts_stamp;
 568                         tp->rx_opt.ts_recent = tm->tcpm_ts;
 569                 }
 570         }
 571         rcu_read_unlock();
 572 }
 573 EXPORT_SYMBOL_GPL(tcp_fetch_timewait_stamp);
 574
 575 /* VJ's idea. Save last timestamp seen from this destination and hold
 576  * it at least for normal timewait interval to use for duplicate
 577  * segment detection in subsequent connections, before they enter
 578  * synchronized state.
 579  */
 580 bool tcp_remember_stamp(struct sock *sk)
 581 {
 582         struct dst_entry *dst = __sk_dst_get(sk);
 583         bool ret = false;
 584
 585         if (dst) {
 586                 struct tcp_metrics_block *tm;
 587
 588                 rcu_read_lock();
 589                 tm = tcp_get_metrics(sk, dst, true);
 590                 if (tm) {
 591                         struct tcp_sock *tp = tcp_sk(sk);
 592
 593                         if ((s32)(tm->tcpm_ts - tp->rx_opt.ts_recent) <= 0 ||
 594                             ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 595                              tm->tcpm_ts_stamp <= (u32)tp->rx_opt.ts_recent_stamp)) {
 596                                 tm->tcpm_ts_stamp = (u32)tp->rx_opt.ts_recent_stamp;
 597                                 tm->tcpm_ts = tp->rx_opt.ts_recent;
 598                         }
 599                         ret = true;
 600                 }
 601                 rcu_read_unlock();
 602         }
 603         return ret;
 604 }
 605
 606 bool tcp_tw_remember_stamp(struct inet_timewait_sock *tw)
 607 {
 608         struct tcp_metrics_block *tm;
 609         bool ret = false;
 610
 611         rcu_read_lock();
 612         tm = __tcp_get_metrics_tw(tw);
 613         if (tm) {
 614                 const struct tcp_timewait_sock *tcptw;
 615                 struct sock *sk = (struct sock *) tw;
 616
 617                 tcptw = tcp_twsk(sk);
 618                 if ((s32)(tm->tcpm_ts - tcptw->tw_ts_recent) <= 0 ||
 619                     ((u32)get_seconds() - tm->tcpm_ts_stamp > TCP_PAWS_MSL &&
 620                      tm->tcpm_ts_stamp <= (u32)tcptw->tw_ts_recent_stamp)) {
 621                         tm->tcpm_ts_stamp = (u32)tcptw->tw_ts_recent_stamp;
 622                         tm->tcpm_ts        = tcptw->tw_ts_recent;
 623                 }
 624                 ret = true;
 625         }
 626         rcu_read_unlock();
 627
 628         return ret;
 629 }
 630
 631 static DEFINE_SEQLOCK(fastopen_seqlock);
 632
 633 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
 634                             struct tcp_fastopen_cookie *cookie,
 635                             int *syn_loss, unsigned long *last_syn_loss)
 636 {
 637         struct tcp_metrics_block *tm;
 638
 639         rcu_read_lock();
 640         tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
 641         if (tm) {
 642                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 643                 unsigned int seq;
 644
 645                 do {
 646                         seq = read_seqbegin(&fastopen_seqlock);
 647                         if (tfom->mss)
 648                                 *mss = tfom->mss;
 649                         *cookie = tfom->cookie;
 650                         *syn_loss = tfom->syn_loss;
 651                         *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
 652                 } while (read_seqretry(&fastopen_seqlock, seq));
 653         }
 654         rcu_read_unlock();
 655 }
 656
 657 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
 658                             struct tcp_fastopen_cookie *cookie, bool syn_lost)
 659 {
 660         struct tcp_metrics_block *tm;
 661
 662         rcu_read_lock();
 663         tm = tcp_get_metrics(sk, __sk_dst_get(sk), true);
 664         if (tm) {
 665                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 666
 667                 write_seqlock_bh(&fastopen_seqlock);
 668                 tfom->mss = mss;
 669                 if (cookie->len > 0)
 670                         tfom->cookie = *cookie;
 671                 if (syn_lost) {
 672                         ++tfom->syn_loss;
 673                         tfom->last_syn_loss = jiffies;
 674                 } else
 675                         tfom->syn_loss = 0;
 676                 write_sequnlock_bh(&fastopen_seqlock);
 677         }
 678         rcu_read_unlock();
 679 }
 680
 681 static struct genl_family tcp_metrics_nl_family = {
 682         .id             = GENL_ID_GENERATE,
 683         .hdrsize        = 0,
 684         .name           = TCP_METRICS_GENL_NAME,
 685         .version        = TCP_METRICS_GENL_VERSION,
 686         .maxattr        = TCP_METRICS_ATTR_MAX,
 687         .netnsok        = true,
 688 };
 689
 690 static struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
 691         [TCP_METRICS_ATTR_ADDR_IPV4]    = { .type = NLA_U32, },
 692         [TCP_METRICS_ATTR_ADDR_IPV6]    = { .type = NLA_BINARY,
 693                                             .len = sizeof(struct in6_addr), },
 694         /* Following attributes are not received for GET/DEL,
 695          * we keep them for reference
 696          */
 697 #if 0
 698         [TCP_METRICS_ATTR_AGE]          = { .type = NLA_MSECS, },
 699         [TCP_METRICS_ATTR_TW_TSVAL]     = { .type = NLA_U32, },
 700         [TCP_METRICS_ATTR_TW_TS_STAMP]  = { .type = NLA_S32, },
 701         [TCP_METRICS_ATTR_VALS]         = { .type = NLA_NESTED, },
 702         [TCP_METRICS_ATTR_FOPEN_MSS]    = { .type = NLA_U16, },
 703         [TCP_METRICS_ATTR_FOPEN_SYN_DROPS]      = { .type = NLA_U16, },
 704         [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]    = { .type = NLA_MSECS, },
 705         [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
 706                                             .len = TCP_FASTOPEN_COOKIE_MAX, },
 707 #endif
 708 };
 709
 710 /* Add attributes, caller cancels its header on failure */
 711 static int tcp_metrics_fill_info(struct sk_buff *msg,
 712                                  struct tcp_metrics_block *tm)
 713 {
 714         struct nlattr *nest;
 715         int i;
 716
 717         switch (tm->tcpm_addr.family) {
 718         case AF_INET:
 719                 if (nla_put_be32(msg, TCP_METRICS_ATTR_ADDR_IPV4,
 720                                 tm->tcpm_addr.addr.a4) < 0)
 721                         goto nla_put_failure;
 722                 break;
 723         case AF_INET6:
 724                 if (nla_put(msg, TCP_METRICS_ATTR_ADDR_IPV6, 16,
 725                             tm->tcpm_addr.addr.a6) < 0)
 726                         goto nla_put_failure;
 727                 break;
 728         default:
 729                 return -EAFNOSUPPORT;
 730         }
 731
 732         if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
 733                           jiffies - tm->tcpm_stamp) < 0)
 734                 goto nla_put_failure;
 735         if (tm->tcpm_ts_stamp) {
 736                 if (nla_put_s32(msg, TCP_METRICS_ATTR_TW_TS_STAMP,
 737                                 (s32) (get_seconds() - tm->tcpm_ts_stamp)) < 0)
 738                         goto nla_put_failure;
 739                 if (nla_put_u32(msg, TCP_METRICS_ATTR_TW_TSVAL,
 740                                 tm->tcpm_ts) < 0)
 741                         goto nla_put_failure;
 742         }
 743
 744         {
 745                 int n = 0;
 746
 747                 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
 748                 if (!nest)
 749                         goto nla_put_failure;
 750                 for (i = 0; i < TCP_METRIC_MAX + 1; i++) {
 751                         if (!tm->tcpm_vals[i])
 752                                 continue;
 753                         if (nla_put_u32(msg, i + 1, tm->tcpm_vals[i]) < 0)
 754                                 goto nla_put_failure;
 755                         n++;
 756                 }
 757                 if (n)
 758                         nla_nest_end(msg, nest);
 759                 else
 760                         nla_nest_cancel(msg, nest);
 761         }
 762
 763         {
 764                 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
 765                 unsigned int seq;
 766
 767                 do {
 768                         seq = read_seqbegin(&fastopen_seqlock);
 769                         tfom_copy[0] = tm->tcpm_fastopen;
 770                 } while (read_seqretry(&fastopen_seqlock, seq));
 771
 772                 tfom = tfom_copy;
 773                 if (tfom->mss &&
 774                     nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
 775                                 tfom->mss) < 0)
 776                         goto nla_put_failure;
 777                 if (tfom->syn_loss &&
 778                     (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
 779                                 tfom->syn_loss) < 0 ||
 780                      nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
 781                                 jiffies - tfom->last_syn_loss) < 0))
 782                         goto nla_put_failure;
 783                 if (tfom->cookie.len > 0 &&
 784                     nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
 785                             tfom->cookie.len, tfom->cookie.val) < 0)
 786                         goto nla_put_failure;
 787         }
 788
 789         return 0;
 790
 791 nla_put_failure:
 792         return -EMSGSIZE;
 793 }
 794
 795 static int tcp_metrics_dump_info(struct sk_buff *skb,
 796                                  struct netlink_callback *cb,
 797                                  struct tcp_metrics_block *tm)
 798 {
 799         void *hdr;
 800
 801         hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 802                           &tcp_metrics_nl_family, NLM_F_MULTI,
 803                           TCP_METRICS_CMD_GET);
 804         if (!hdr)
 805                 return -EMSGSIZE;
 806
 807         if (tcp_metrics_fill_info(skb, tm) < 0)
 808                 goto nla_put_failure;
 809
 810         return genlmsg_end(skb, hdr);
 811
 812 nla_put_failure:
 813         genlmsg_cancel(skb, hdr);
 814         return -EMSGSIZE;
 815 }
 816
 817 static int tcp_metrics_nl_dump(struct sk_buff *skb,
 818                                struct netlink_callback *cb)
 819 {
 820         struct net *net = sock_net(skb->sk);
 821         unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
 822         unsigned int row, s_row = cb->args[0];
 823         int s_col = cb->args[1], col = s_col;
 824
 825         for (row = s_row; row < max_rows; row++, s_col = 0) {
 826                 struct tcp_metrics_block *tm;
 827                 struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash + row;
 828
 829                 rcu_read_lock();
 830                 for (col = 0, tm = rcu_dereference(hb->chain); tm;
 831                      tm = rcu_dereference(tm->tcpm_next), col++) {
 832                         if (col < s_col)
 833                                 continue;
 834                         if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
 835                                 rcu_read_unlock();
 836                                 goto done;
 837                         }
 838                 }
 839                 rcu_read_unlock();
 840         }
 841
 842 done:
 843         cb->args[0] = row;
 844         cb->args[1] = col;
 845         return skb->len;
 846 }
 847
 848 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 849                          unsigned int *hash, int optional)
 850 {
 851         struct nlattr *a;
 852
 853         a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV4];
 854         if (a) {
 855                 addr->family = AF_INET;
 856                 addr->addr.a4 = nla_get_be32(a);
 857                 *hash = (__force unsigned int) addr->addr.a4;
 858                 return 0;
 859         }
 860         a = info->attrs[TCP_METRICS_ATTR_ADDR_IPV6];
 861         if (a) {
 862                 if (nla_len(a) != sizeof(struct in6_addr))
 863                         return -EINVAL;
 864                 addr->family = AF_INET6;
 865                 memcpy(addr->addr.a6, nla_data(a), sizeof(addr->addr.a6));
 866                 *hash = ipv6_addr_hash((struct in6_addr *) addr->addr.a6);
 867                 return 0;
 868         }
 869         return optional ? 1 : -EAFNOSUPPORT;
 870 }
 871
 872 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
 873 {
 874         struct tcp_metrics_block *tm;
 875         struct inetpeer_addr addr;
 876         unsigned int hash;
 877         struct sk_buff *msg;
 878         struct net *net = genl_info_net(info);
 879         void *reply;
 880         int ret;
 881
 882         ret = parse_nl_addr(info, &addr, &hash, 0);
 883         if (ret < 0)
 884                 return ret;
 885
 886         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 887         if (!msg)
 888                 return -ENOMEM;
 889
 890         reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
 891                                   info->genlhdr->cmd);
 892         if (!reply)
 893                 goto nla_put_failure;
 894
 895         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 896         ret = -ESRCH;
 897         rcu_read_lock();
 898         for (tm = rcu_dereference(net->ipv4.tcp_metrics_hash[hash].chain); tm;
 899              tm = rcu_dereference(tm->tcpm_next)) {
 900                 if (addr_same(&tm->tcpm_addr, &addr)) {
 901                         ret = tcp_metrics_fill_info(msg, tm);
 902                         break;
 903                 }
 904         }
 905         rcu_read_unlock();
 906         if (ret < 0)
 907                 goto out_free;
 908
 909         genlmsg_end(msg, reply);
 910         return genlmsg_reply(msg, info);
 911
 912 nla_put_failure:
 913         ret = -EMSGSIZE;
 914
 915 out_free:
 916         nlmsg_free(msg);
 917         return ret;
 918 }
 919
 920 #define deref_locked_genl(p)    \
 921         rcu_dereference_protected(p, lockdep_genl_is_held() && \
 922                                      lockdep_is_held(&tcp_metrics_lock))
 923
 924 #define deref_genl(p)   rcu_dereference_protected(p, lockdep_genl_is_held())
 925
 926 static int tcp_metrics_flush_all(struct net *net)
 927 {
 928         unsigned int max_rows = 1U << net->ipv4.tcp_metrics_hash_log;
 929         struct tcpm_hash_bucket *hb = net->ipv4.tcp_metrics_hash;
 930         struct tcp_metrics_block *tm;
 931         unsigned int row;
 932
 933         for (row = 0; row < max_rows; row++, hb++) {
 934                 spin_lock_bh(&tcp_metrics_lock);
 935                 tm = deref_locked_genl(hb->chain);
 936                 if (tm)
 937                         hb->chain = NULL;
 938                 spin_unlock_bh(&tcp_metrics_lock);
 939                 while (tm) {
 940                         struct tcp_metrics_block *next;
 941
 942                         next = deref_genl(tm->tcpm_next);
 943                         kfree_rcu(tm, rcu_head);
 944                         tm = next;
 945                 }
 946         }
 947         return 0;
 948 }
 949
 950 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
 951 {
 952         struct tcpm_hash_bucket *hb;
 953         struct tcp_metrics_block *tm;
 954         struct tcp_metrics_block __rcu **pp;
 955         struct inetpeer_addr addr;
 956         unsigned int hash;
 957         struct net *net = genl_info_net(info);
 958         int ret;
 959
 960         ret = parse_nl_addr(info, &addr, &hash, 1);
 961         if (ret < 0)
 962                 return ret;
 963         if (ret > 0)
 964                 return tcp_metrics_flush_all(net);
 965
 966         hash = hash_32(hash, net->ipv4.tcp_metrics_hash_log);
 967         hb = net->ipv4.tcp_metrics_hash + hash;
 968         pp = &hb->chain;
 969         spin_lock_bh(&tcp_metrics_lock);
 970         for (tm = deref_locked_genl(*pp); tm;
 971              pp = &tm->tcpm_next, tm = deref_locked_genl(*pp)) {
 972                 if (addr_same(&tm->tcpm_addr, &addr)) {
 973                         *pp = tm->tcpm_next;
 974                         break;
 975                 }
 976         }
 977         spin_unlock_bh(&tcp_metrics_lock);
 978         if (!tm)
 979                 return -ESRCH;
 980         kfree_rcu(tm, rcu_head);
 981         return 0;
 982 }
 983
 984 static struct genl_ops tcp_metrics_nl_ops[] = {
 985         {
 986                 .cmd = TCP_METRICS_CMD_GET,
 987                 .doit = tcp_metrics_nl_cmd_get,
 988                 .dumpit = tcp_metrics_nl_dump,
 989                 .policy = tcp_metrics_nl_policy,
 990                 .flags = GENL_ADMIN_PERM,
 991         },
 992         {
 993                 .cmd = TCP_METRICS_CMD_DEL,
 994                 .doit = tcp_metrics_nl_cmd_del,
 995                 .policy = tcp_metrics_nl_policy,
 996                 .flags = GENL_ADMIN_PERM,
 997         },
 998 };
 999
1000 static unsigned int tcpmhash_entries;
1001 static int __init set_tcpmhash_entries(char *str)
1002 {
1003         ssize_t ret;
1004
1005         if (!str)
1006                 return 0;
1007
1008         ret = kstrtouint(str, 0, &tcpmhash_entries);
1009         if (ret)
1010                 return 0;
1011
1012         return 1;
1013 }
1014 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1015
1016 static int __net_init tcp_net_metrics_init(struct net *net)
1017 {
1018         size_t size;
1019         unsigned int slots;
1020
1021         slots = tcpmhash_entries;
1022         if (!slots) {
1023                 if (totalram_pages >= 128 * 1024)
1024                         slots = 16 * 1024;
1025                 else
1026                         slots = 8 * 1024;
1027         }
1028
1029         net->ipv4.tcp_metrics_hash_log = order_base_2(slots);
1030         size = sizeof(struct tcpm_hash_bucket) << net->ipv4.tcp_metrics_hash_log;
1031
1032         net->ipv4.tcp_metrics_hash = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
1033         if (!net->ipv4.tcp_metrics_hash)
1034                 net->ipv4.tcp_metrics_hash = vzalloc(size);
1035
1036         if (!net->ipv4.tcp_metrics_hash)
1037                 return -ENOMEM;
1038
1039         return 0;
1040 }
1041
1042 static void __net_exit tcp_net_metrics_exit(struct net *net)
1043 {
1044         unsigned int i;
1045
1046         for (i = 0; i < (1U << net->ipv4.tcp_metrics_hash_log) ; i++) {
1047                 struct tcp_metrics_block *tm, *next;
1048
1049                 tm = rcu_dereference_protected(net->ipv4.tcp_metrics_hash[i].chain, 1);
1050                 while (tm) {
1051                         next = rcu_dereference_protected(tm->tcpm_next, 1);
1052                         kfree(tm);
1053                         tm = next;
1054                 }
1055         }
1056         if (is_vmalloc_addr(net->ipv4.tcp_metrics_hash))
1057                 vfree(net->ipv4.tcp_metrics_hash);
1058         else
1059                 kfree(net->ipv4.tcp_metrics_hash);
1060 }
1061
1062 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1063         .init   =       tcp_net_metrics_init,
1064         .exit   =       tcp_net_metrics_exit,
1065 };
1066
1067 void __init tcp_metrics_init(void)
1068 {
1069         int ret;
1070
1071         ret = register_pernet_subsys(&tcp_net_metrics_ops);
1072         if (ret < 0)
1073                 goto cleanup;
1074         ret = genl_register_family_with_ops(&tcp_metrics_nl_family,
1075                                             tcp_metrics_nl_ops,
1076                                             ARRAY_SIZE(tcp_metrics_nl_ops));
1077         if (ret < 0)
1078                 goto cleanup_subsys;
1079         return;
1080
1081 cleanup_subsys:
1082         unregister_pernet_subsys(&tcp_net_metrics_ops);
1083
1084 cleanup:
1085         return;
1086 }