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