net/ipv4/tcp_metrics.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/rcupdate.h>
   3 #include <linux/spinlock.h>
   4 #include <linux/jiffies.h>
   5 #include <linux/module.h>
   6 #include <linux/cache.h>
   7 #include <linux/slab.h>
   8 #include <linux/init.h>
   9 #include <linux/tcp.h>
  10 #include <linux/hash.h>
  11 #include <linux/tcp_metrics.h>
  12 #include <linux/vmalloc.h>
  13
  14 #include <net/inet_connection_sock.h>
  15 #include <net/net_namespace.h>
  16 #include <net/request_sock.h>
  17 #include <net/inetpeer.h>
  18 #include <net/sock.h>
  19 #include <net/ipv6.h>
  20 #include <net/dst.h>
  21 #include <net/tcp.h>
  22 #include <net/genetlink.h>
  23
  24 int sysctl_tcp_nometrics_save __read_mostly;
  25
  26 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
  27                                                    const struct inetpeer_addr *daddr,
  28                                                    struct net *net, unsigned int hash);
  29
  30 struct tcp_fastopen_metrics {
  31         u16     mss;
  32         u16     syn_loss:10,            /* Recurring Fast Open SYN losses */
  33                 try_exp:2;              /* Request w/ exp. option (once) */
  34         unsigned long   last_syn_loss;  /* Last Fast Open SYN loss */
  35         struct  tcp_fastopen_cookie     cookie;
  36 };
  37
  38 /* TCP_METRIC_MAX includes 2 extra fields for userspace compatibility
  39  * Kernel only stores RTT and RTTVAR in usec resolution
  40  */
  41 #define TCP_METRIC_MAX_KERNEL (TCP_METRIC_MAX - 2)
  42
  43 struct tcp_metrics_block {
  44         struct tcp_metrics_block __rcu  *tcpm_next;
  45         struct net                      *tcpm_net;
  46         struct inetpeer_addr            tcpm_saddr;
  47         struct inetpeer_addr            tcpm_daddr;
  48         unsigned long                   tcpm_stamp;
  49         u32                             tcpm_lock;
  50         u32                             tcpm_vals[TCP_METRIC_MAX_KERNEL + 1];
  51         struct tcp_fastopen_metrics     tcpm_fastopen;
  52
  53         struct rcu_head                 rcu_head;
  54 };
  55
  56 static inline struct net *tm_net(const struct tcp_metrics_block *tm)
  57 {
  58         /* Paired with the WRITE_ONCE() in tcpm_new() */
  59         return READ_ONCE(tm->tcpm_net);
  60 }
  61
  62 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
  63                               enum tcp_metric_index idx)
  64 {
  65         /* Paired with WRITE_ONCE() in tcpm_suck_dst() */
  66         return READ_ONCE(tm->tcpm_lock) & (1 << idx);
  67 }
  68
  69 static u32 tcp_metric_get(const struct tcp_metrics_block *tm,
  70                           enum tcp_metric_index idx)
  71 {
  72         /* Paired with WRITE_ONCE() in tcp_metric_set() */
  73         return READ_ONCE(tm->tcpm_vals[idx]);
  74 }
  75
  76 static void tcp_metric_set(struct tcp_metrics_block *tm,
  77                            enum tcp_metric_index idx,
  78                            u32 val)
  79 {
  80         /* Paired with READ_ONCE() in tcp_metric_get() */
  81         WRITE_ONCE(tm->tcpm_vals[idx], val);
  82 }
  83
  84 static bool addr_same(const struct inetpeer_addr *a,
  85                       const struct inetpeer_addr *b)
  86 {
  87         return (a->family == b->family) && !inetpeer_addr_cmp(a, b);
  88 }
  89
  90 struct tcpm_hash_bucket {
  91         struct tcp_metrics_block __rcu  *chain;
  92 };
  93
  94 static struct tcpm_hash_bucket  *tcp_metrics_hash __read_mostly;
  95 static unsigned int             tcp_metrics_hash_log __read_mostly;
  96
  97 static DEFINE_SPINLOCK(tcp_metrics_lock);
  98 static DEFINE_SEQLOCK(fastopen_seqlock);
  99
 100 static void tcpm_suck_dst(struct tcp_metrics_block *tm,
 101                           const struct dst_entry *dst,
 102                           bool fastopen_clear)
 103 {
 104         u32 msval;
 105         u32 val;
 106
 107         WRITE_ONCE(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         /* Paired with READ_ONCE() in tcp_metric_locked() */
 121         WRITE_ONCE(tm->tcpm_lock, val);
 122
 123         msval = dst_metric_raw(dst, RTAX_RTT);
 124         tcp_metric_set(tm, TCP_METRIC_RTT, msval * USEC_PER_MSEC);
 125
 126         msval = dst_metric_raw(dst, RTAX_RTTVAR);
 127         tcp_metric_set(tm, TCP_METRIC_RTTVAR, msval * USEC_PER_MSEC);
 128         tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 129                        dst_metric_raw(dst, RTAX_SSTHRESH));
 130         tcp_metric_set(tm, TCP_METRIC_CWND,
 131                        dst_metric_raw(dst, RTAX_CWND));
 132         tcp_metric_set(tm, TCP_METRIC_REORDERING,
 133                        dst_metric_raw(dst, RTAX_REORDERING));
 134         if (fastopen_clear) {
 135                 write_seqlock(&fastopen_seqlock);
 136                 tm->tcpm_fastopen.mss = 0;
 137                 tm->tcpm_fastopen.syn_loss = 0;
 138                 tm->tcpm_fastopen.try_exp = 0;
 139                 tm->tcpm_fastopen.cookie.exp = false;
 140                 tm->tcpm_fastopen.cookie.len = 0;
 141                 write_sequnlock(&fastopen_seqlock);
 142         }
 143 }
 144
 145 #define TCP_METRICS_TIMEOUT             (60 * 60 * HZ)
 146
 147 static void tcpm_check_stamp(struct tcp_metrics_block *tm,
 148                              const struct dst_entry *dst)
 149 {
 150         unsigned long limit;
 151
 152         if (!tm)
 153                 return;
 154         limit = READ_ONCE(tm->tcpm_stamp) + TCP_METRICS_TIMEOUT;
 155         if (unlikely(time_after(jiffies, limit)))
 156                 tcpm_suck_dst(tm, dst, false);
 157 }
 158
 159 #define TCP_METRICS_RECLAIM_DEPTH       5
 160 #define TCP_METRICS_RECLAIM_PTR         (struct tcp_metrics_block *) 0x1UL
 161
 162 #define deref_locked(p) \
 163         rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
 164
 165 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
 166                                           struct inetpeer_addr *saddr,
 167                                           struct inetpeer_addr *daddr,
 168                                           unsigned int hash)
 169 {
 170         struct tcp_metrics_block *tm;
 171         struct net *net;
 172         bool reclaim = false;
 173
 174         spin_lock_bh(&tcp_metrics_lock);
 175         net = dev_net(dst->dev);
 176
 177         /* While waiting for the spin-lock the cache might have been populated
 178          * with this entry and so we have to check again.
 179          */
 180         tm = __tcp_get_metrics(saddr, daddr, net, hash);
 181         if (tm == TCP_METRICS_RECLAIM_PTR) {
 182                 reclaim = true;
 183                 tm = NULL;
 184         }
 185         if (tm) {
 186                 tcpm_check_stamp(tm, dst);
 187                 goto out_unlock;
 188         }
 189
 190         if (unlikely(reclaim)) {
 191                 struct tcp_metrics_block *oldest;
 192
 193                 oldest = deref_locked(tcp_metrics_hash[hash].chain);
 194                 for (tm = deref_locked(oldest->tcpm_next); tm;
 195                      tm = deref_locked(tm->tcpm_next)) {
 196                         if (time_before(READ_ONCE(tm->tcpm_stamp),
 197                                         READ_ONCE(oldest->tcpm_stamp)))
 198                                 oldest = tm;
 199                 }
 200                 tm = oldest;
 201         } else {
 202                 tm = kzalloc(sizeof(*tm), GFP_ATOMIC);
 203                 if (!tm)
 204                         goto out_unlock;
 205         }
 206         /* Paired with the READ_ONCE() in tm_net() */
 207         WRITE_ONCE(tm->tcpm_net, net);
 208
 209         tm->tcpm_saddr = *saddr;
 210         tm->tcpm_daddr = *daddr;
 211
 212         tcpm_suck_dst(tm, dst, reclaim);
 213
 214         if (likely(!reclaim)) {
 215                 tm->tcpm_next = tcp_metrics_hash[hash].chain;
 216                 rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
 217         }
 218
 219 out_unlock:
 220         spin_unlock_bh(&tcp_metrics_lock);
 221         return tm;
 222 }
 223
 224 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
 225 {
 226         if (tm)
 227                 return tm;
 228         if (depth > TCP_METRICS_RECLAIM_DEPTH)
 229                 return TCP_METRICS_RECLAIM_PTR;
 230         return NULL;
 231 }
 232
 233 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
 234                                                    const struct inetpeer_addr *daddr,
 235                                                    struct net *net, unsigned int hash)
 236 {
 237         struct tcp_metrics_block *tm;
 238         int depth = 0;
 239
 240         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
 241              tm = rcu_dereference(tm->tcpm_next)) {
 242                 if (addr_same(&tm->tcpm_saddr, saddr) &&
 243                     addr_same(&tm->tcpm_daddr, daddr) &&
 244                     net_eq(tm_net(tm), net))
 245                         break;
 246                 depth++;
 247         }
 248         return tcp_get_encode(tm, depth);
 249 }
 250
 251 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
 252                                                        struct dst_entry *dst)
 253 {
 254         struct tcp_metrics_block *tm;
 255         struct inetpeer_addr saddr, daddr;
 256         unsigned int hash;
 257         struct net *net;
 258
 259         saddr.family = req->rsk_ops->family;
 260         daddr.family = req->rsk_ops->family;
 261         switch (daddr.family) {
 262         case AF_INET:
 263                 inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
 264                 inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
 265                 hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
 266                 break;
 267 #if IS_ENABLED(CONFIG_IPV6)
 268         case AF_INET6:
 269                 inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
 270                 inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr);
 271                 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
 272                 break;
 273 #endif
 274         default:
 275                 return NULL;
 276         }
 277
 278         net = dev_net(dst->dev);
 279         hash ^= net_hash_mix(net);
 280         hash = hash_32(hash, tcp_metrics_hash_log);
 281
 282         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
 283              tm = rcu_dereference(tm->tcpm_next)) {
 284                 if (addr_same(&tm->tcpm_saddr, &saddr) &&
 285                     addr_same(&tm->tcpm_daddr, &daddr) &&
 286                     net_eq(tm_net(tm), net))
 287                         break;
 288         }
 289         tcpm_check_stamp(tm, dst);
 290         return tm;
 291 }
 292
 293 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
 294                                                  struct dst_entry *dst,
 295                                                  bool create)
 296 {
 297         struct tcp_metrics_block *tm;
 298         struct inetpeer_addr saddr, daddr;
 299         unsigned int hash;
 300         struct net *net;
 301
 302         if (sk->sk_family == AF_INET) {
 303                 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
 304                 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
 305                 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
 306         }
 307 #if IS_ENABLED(CONFIG_IPV6)
 308         else if (sk->sk_family == AF_INET6) {
 309                 if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
 310                         inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
 311                         inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
 312                         hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
 313                 } else {
 314                         inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
 315                         inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
 316                         hash = ipv6_addr_hash(&sk->sk_v6_daddr);
 317                 }
 318         }
 319 #endif
 320         else
 321                 return NULL;
 322
 323         net = dev_net(dst->dev);
 324         hash ^= net_hash_mix(net);
 325         hash = hash_32(hash, tcp_metrics_hash_log);
 326
 327         tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
 328         if (tm == TCP_METRICS_RECLAIM_PTR)
 329                 tm = NULL;
 330         if (!tm && create)
 331                 tm = tcpm_new(dst, &saddr, &daddr, hash);
 332         else
 333                 tcpm_check_stamp(tm, dst);
 334
 335         return tm;
 336 }
 337
 338 /* Save metrics learned by this TCP session.  This function is called
 339  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
 340  * or goes from LAST-ACK to CLOSE.
 341  */
 342 void tcp_update_metrics(struct sock *sk)
 343 {
 344         const struct inet_connection_sock *icsk = inet_csk(sk);
 345         struct dst_entry *dst = __sk_dst_get(sk);
 346         struct tcp_sock *tp = tcp_sk(sk);
 347         struct net *net = sock_net(sk);
 348         struct tcp_metrics_block *tm;
 349         unsigned long rtt;
 350         u32 val;
 351         int m;
 352
 353         sk_dst_confirm(sk);
 354         if (sysctl_tcp_nometrics_save || !dst)
 355                 return;
 356
 357         rcu_read_lock();
 358         if (icsk->icsk_backoff || !tp->srtt_us) {
 359                 /* This session failed to estimate rtt. Why?
 360                  * Probably, no packets returned in time.  Reset our
 361                  * results.
 362                  */
 363                 tm = tcp_get_metrics(sk, dst, false);
 364                 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
 365                         tcp_metric_set(tm, TCP_METRIC_RTT, 0);
 366                 goto out_unlock;
 367         } else
 368                 tm = tcp_get_metrics(sk, dst, true);
 369
 370         if (!tm)
 371                 goto out_unlock;
 372
 373         rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
 374         m = rtt - tp->srtt_us;
 375
 376         /* If newly calculated rtt larger than stored one, store new
 377          * one. Otherwise, use EWMA. Remember, rtt overestimation is
 378          * always better than underestimation.
 379          */
 380         if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
 381                 if (m <= 0)
 382                         rtt = tp->srtt_us;
 383                 else
 384                         rtt -= (m >> 3);
 385                 tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
 386         }
 387
 388         if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
 389                 unsigned long var;
 390
 391                 if (m < 0)
 392                         m = -m;
 393
 394                 /* Scale deviation to rttvar fixed point */
 395                 m >>= 1;
 396                 if (m < tp->mdev_us)
 397                         m = tp->mdev_us;
 398
 399                 var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
 400                 if (m >= var)
 401                         var = m;
 402                 else
 403                         var -= (var - m) >> 2;
 404
 405                 tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
 406         }
 407
 408         if (tcp_in_initial_slowstart(tp)) {
 409                 /* Slow start still did not finish. */
 410                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 411                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 412                         if (val && (tp->snd_cwnd >> 1) > val)
 413                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 414                                                tp->snd_cwnd >> 1);
 415                 }
 416                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 417                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 418                         if (tp->snd_cwnd > val)
 419                                 tcp_metric_set(tm, TCP_METRIC_CWND,
 420                                                tp->snd_cwnd);
 421                 }
 422         } else if (!tcp_in_slow_start(tp) &&
 423                    icsk->icsk_ca_state == TCP_CA_Open) {
 424                 /* Cong. avoidance phase, cwnd is reliable. */
 425                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
 426                         tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 427                                        max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
 428                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 429                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 430                         tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
 431                 }
 432         } else {
 433                 /* Else slow start did not finish, cwnd is non-sense,
 434                  * ssthresh may be also invalid.
 435                  */
 436                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 437                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 438                         tcp_metric_set(tm, TCP_METRIC_CWND,
 439                                        (val + tp->snd_ssthresh) >> 1);
 440                 }
 441                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 442                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 443                         if (val && tp->snd_ssthresh > val)
 444                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 445                                                tp->snd_ssthresh);
 446                 }
 447                 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
 448                         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 449                         if (val < tp->reordering &&
 450                             tp->reordering != net->ipv4.sysctl_tcp_reordering)
 451                                 tcp_metric_set(tm, TCP_METRIC_REORDERING,
 452                                                tp->reordering);
 453                 }
 454         }
 455         WRITE_ONCE(tm->tcpm_stamp, jiffies);
 456 out_unlock:
 457         rcu_read_unlock();
 458 }
 459
 460 /* Initialize metrics on socket. */
 461
 462 void tcp_init_metrics(struct sock *sk)
 463 {
 464         struct dst_entry *dst = __sk_dst_get(sk);
 465         struct tcp_sock *tp = tcp_sk(sk);
 466         struct tcp_metrics_block *tm;
 467         u32 val, crtt = 0; /* cached RTT scaled by 8 */
 468
 469         sk_dst_confirm(sk);
 470         /* ssthresh may have been reduced unnecessarily during.
 471          * 3WHS. Restore it back to its initial default.
 472          */
 473         tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 474         if (!dst)
 475                 goto reset;
 476
 477         rcu_read_lock();
 478         tm = tcp_get_metrics(sk, dst, false);
 479         if (!tm) {
 480                 rcu_read_unlock();
 481                 goto reset;
 482         }
 483
 484         if (tcp_metric_locked(tm, TCP_METRIC_CWND))
 485                 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
 486
 487         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 488         if (val) {
 489                 tp->snd_ssthresh = val;
 490                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
 491                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
 492         }
 493         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 494         if (val && tp->reordering != val) {
 495                 tcp_disable_fack(tp);
 496                 tp->reordering = val;
 497         }
 498
 499         crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
 500         rcu_read_unlock();
 501 reset:
 502         /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
 503          * to seed the RTO for later data packets because SYN packets are
 504          * small. Use the per-dst cached values to seed the RTO but keep
 505          * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
 506          * Later the RTO will be updated immediately upon obtaining the first
 507          * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
 508          * influences the first RTO but not later RTT estimation.
 509          *
 510          * But if RTT is not available from the SYN (due to retransmits or
 511          * syn cookies) or the cache, force a conservative 3secs timeout.
 512          *
 513          * A bit of theory. RTT is time passed after "normal" sized packet
 514          * is sent until it is ACKed. In normal circumstances sending small
 515          * packets force peer to delay ACKs and calculation is correct too.
 516          * The algorithm is adaptive and, provided we follow specs, it
 517          * NEVER underestimate RTT. BUT! If peer tries to make some clever
 518          * tricks sort of "quick acks" for time long enough to decrease RTT
 519          * to low value, and then abruptly stops to do it and starts to delay
 520          * ACKs, wait for troubles.
 521          */
 522         if (crtt > tp->srtt_us) {
 523                 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
 524                 crtt /= 8 * USEC_PER_SEC / HZ;
 525                 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
 526         } else if (tp->srtt_us == 0) {
 527                 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
 528                  * 3WHS. This is most likely due to retransmission,
 529                  * including spurious one. Reset the RTO back to 3secs
 530                  * from the more aggressive 1sec to avoid more spurious
 531                  * retransmission.
 532                  */
 533                 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
 534                 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
 535
 536                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
 537         }
 538         /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
 539          * retransmitted. In light of RFC6298 more aggressive 1sec
 540          * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
 541          * retransmission has occurred.
 542          */
 543         if (tp->total_retrans > 1)
 544                 tp->snd_cwnd = 1;
 545         else
 546                 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
 547         tp->snd_cwnd_stamp = tcp_jiffies32;
 548 }
 549
 550 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
 551 {
 552         struct tcp_metrics_block *tm;
 553         bool ret;
 554
 555         if (!dst)
 556                 return false;
 557
 558         rcu_read_lock();
 559         tm = __tcp_get_metrics_req(req, dst);
 560         if (tm && tcp_metric_get(tm, TCP_METRIC_RTT))
 561                 ret = true;
 562         else
 563                 ret = false;
 564         rcu_read_unlock();
 565
 566         return ret;
 567 }
 568
 569 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
 570                             struct tcp_fastopen_cookie *cookie,
 571                             int *syn_loss, unsigned long *last_syn_loss)
 572 {
 573         struct tcp_metrics_block *tm;
 574
 575         rcu_read_lock();
 576         tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
 577         if (tm) {
 578                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 579                 unsigned int seq;
 580
 581                 do {
 582                         seq = read_seqbegin(&fastopen_seqlock);
 583                         if (tfom->mss)
 584                                 *mss = tfom->mss;
 585                         *cookie = tfom->cookie;
 586                         if (cookie->len <= 0 && tfom->try_exp == 1)
 587                                 cookie->exp = true;
 588                         *syn_loss = tfom->syn_loss;
 589                         *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
 590                 } while (read_seqretry(&fastopen_seqlock, seq));
 591         }
 592         rcu_read_unlock();
 593 }
 594
 595 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
 596                             struct tcp_fastopen_cookie *cookie, bool syn_lost,
 597                             u16 try_exp)
 598 {
 599         struct dst_entry *dst = __sk_dst_get(sk);
 600         struct tcp_metrics_block *tm;
 601
 602         if (!dst)
 603                 return;
 604         rcu_read_lock();
 605         tm = tcp_get_metrics(sk, dst, true);
 606         if (tm) {
 607                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 608
 609                 write_seqlock_bh(&fastopen_seqlock);
 610                 if (mss)
 611                         tfom->mss = mss;
 612                 if (cookie && cookie->len > 0)
 613                         tfom->cookie = *cookie;
 614                 else if (try_exp > tfom->try_exp &&
 615                          tfom->cookie.len <= 0 && !tfom->cookie.exp)
 616                         tfom->try_exp = try_exp;
 617                 if (syn_lost) {
 618                         ++tfom->syn_loss;
 619                         tfom->last_syn_loss = jiffies;
 620                 } else
 621                         tfom->syn_loss = 0;
 622                 write_sequnlock_bh(&fastopen_seqlock);
 623         }
 624         rcu_read_unlock();
 625 }
 626
 627 static struct genl_family tcp_metrics_nl_family;
 628
 629 static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
 630         [TCP_METRICS_ATTR_ADDR_IPV4]    = { .type = NLA_U32, },
 631         [TCP_METRICS_ATTR_ADDR_IPV6]    = { .type = NLA_BINARY,
 632                                             .len = sizeof(struct in6_addr), },
 633         /* Following attributes are not received for GET/DEL,
 634          * we keep them for reference
 635          */
 636 #if 0
 637         [TCP_METRICS_ATTR_AGE]          = { .type = NLA_MSECS, },
 638         [TCP_METRICS_ATTR_TW_TSVAL]     = { .type = NLA_U32, },
 639         [TCP_METRICS_ATTR_TW_TS_STAMP]  = { .type = NLA_S32, },
 640         [TCP_METRICS_ATTR_VALS]         = { .type = NLA_NESTED, },
 641         [TCP_METRICS_ATTR_FOPEN_MSS]    = { .type = NLA_U16, },
 642         [TCP_METRICS_ATTR_FOPEN_SYN_DROPS]      = { .type = NLA_U16, },
 643         [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]    = { .type = NLA_MSECS, },
 644         [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
 645                                             .len = TCP_FASTOPEN_COOKIE_MAX, },
 646 #endif
 647 };
 648
 649 /* Add attributes, caller cancels its header on failure */
 650 static int tcp_metrics_fill_info(struct sk_buff *msg,
 651                                  struct tcp_metrics_block *tm)
 652 {
 653         struct nlattr *nest;
 654         int i;
 655
 656         switch (tm->tcpm_daddr.family) {
 657         case AF_INET:
 658                 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
 659                                     inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
 660                         goto nla_put_failure;
 661                 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
 662                                     inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
 663                         goto nla_put_failure;
 664                 break;
 665         case AF_INET6:
 666                 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
 667                                      inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
 668                         goto nla_put_failure;
 669                 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
 670                                      inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
 671                         goto nla_put_failure;
 672                 break;
 673         default:
 674                 return -EAFNOSUPPORT;
 675         }
 676
 677         if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
 678                           jiffies - READ_ONCE(tm->tcpm_stamp),
 679                           TCP_METRICS_ATTR_PAD) < 0)
 680                 goto nla_put_failure;
 681
 682         {
 683                 int n = 0;
 684
 685                 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
 686                 if (!nest)
 687                         goto nla_put_failure;
 688                 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
 689                         u32 val = tcp_metric_get(tm, i);
 690
 691                         if (!val)
 692                                 continue;
 693                         if (i == TCP_METRIC_RTT) {
 694                                 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
 695                                                 val) < 0)
 696                                         goto nla_put_failure;
 697                                 n++;
 698                                 val = max(val / 1000, 1U);
 699                         }
 700                         if (i == TCP_METRIC_RTTVAR) {
 701                                 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
 702                                                 val) < 0)
 703                                         goto nla_put_failure;
 704                                 n++;
 705                                 val = max(val / 1000, 1U);
 706                         }
 707                         if (nla_put_u32(msg, i + 1, val) < 0)
 708                                 goto nla_put_failure;
 709                         n++;
 710                 }
 711                 if (n)
 712                         nla_nest_end(msg, nest);
 713                 else
 714                         nla_nest_cancel(msg, nest);
 715         }
 716
 717         {
 718                 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
 719                 unsigned int seq;
 720
 721                 do {
 722                         seq = read_seqbegin(&fastopen_seqlock);
 723                         tfom_copy[0] = tm->tcpm_fastopen;
 724                 } while (read_seqretry(&fastopen_seqlock, seq));
 725
 726                 tfom = tfom_copy;
 727                 if (tfom->mss &&
 728                     nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
 729                                 tfom->mss) < 0)
 730                         goto nla_put_failure;
 731                 if (tfom->syn_loss &&
 732                     (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
 733                                 tfom->syn_loss) < 0 ||
 734                      nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
 735                                 jiffies - tfom->last_syn_loss,
 736                                 TCP_METRICS_ATTR_PAD) < 0))
 737                         goto nla_put_failure;
 738                 if (tfom->cookie.len > 0 &&
 739                     nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
 740                             tfom->cookie.len, tfom->cookie.val) < 0)
 741                         goto nla_put_failure;
 742         }
 743
 744         return 0;
 745
 746 nla_put_failure:
 747         return -EMSGSIZE;
 748 }
 749
 750 static int tcp_metrics_dump_info(struct sk_buff *skb,
 751                                  struct netlink_callback *cb,
 752                                  struct tcp_metrics_block *tm)
 753 {
 754         void *hdr;
 755
 756         hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 757                           &tcp_metrics_nl_family, NLM_F_MULTI,
 758                           TCP_METRICS_CMD_GET);
 759         if (!hdr)
 760                 return -EMSGSIZE;
 761
 762         if (tcp_metrics_fill_info(skb, tm) < 0)
 763                 goto nla_put_failure;
 764
 765         genlmsg_end(skb, hdr);
 766         return 0;
 767
 768 nla_put_failure:
 769         genlmsg_cancel(skb, hdr);
 770         return -EMSGSIZE;
 771 }
 772
 773 static int tcp_metrics_nl_dump(struct sk_buff *skb,
 774                                struct netlink_callback *cb)
 775 {
 776         struct net *net = sock_net(skb->sk);
 777         unsigned int max_rows = 1U << tcp_metrics_hash_log;
 778         unsigned int row, s_row = cb->args[0];
 779         int s_col = cb->args[1], col = s_col;
 780
 781         for (row = s_row; row < max_rows; row++, s_col = 0) {
 782                 struct tcp_metrics_block *tm;
 783                 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
 784
 785                 rcu_read_lock();
 786                 for (col = 0, tm = rcu_dereference(hb->chain); tm;
 787                      tm = rcu_dereference(tm->tcpm_next), col++) {
 788                         if (!net_eq(tm_net(tm), net))
 789                                 continue;
 790                         if (col < s_col)
 791                                 continue;
 792                         if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
 793                                 rcu_read_unlock();
 794                                 goto done;
 795                         }
 796                 }
 797                 rcu_read_unlock();
 798         }
 799
 800 done:
 801         cb->args[0] = row;
 802         cb->args[1] = col;
 803         return skb->len;
 804 }
 805
 806 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 807                            unsigned int *hash, int optional, int v4, int v6)
 808 {
 809         struct nlattr *a;
 810
 811         a = info->attrs[v4];
 812         if (a) {
 813                 inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
 814                 if (hash)
 815                         *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
 816                 return 0;
 817         }
 818         a = info->attrs[v6];
 819         if (a) {
 820                 struct in6_addr in6;
 821
 822                 if (nla_len(a) != sizeof(struct in6_addr))
 823                         return -EINVAL;
 824                 in6 = nla_get_in6_addr(a);
 825                 inetpeer_set_addr_v6(addr, &in6);
 826                 if (hash)
 827                         *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
 828                 return 0;
 829         }
 830         return optional ? 1 : -EAFNOSUPPORT;
 831 }
 832
 833 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 834                          unsigned int *hash, int optional)
 835 {
 836         return __parse_nl_addr(info, addr, hash, optional,
 837                                TCP_METRICS_ATTR_ADDR_IPV4,
 838                                TCP_METRICS_ATTR_ADDR_IPV6);
 839 }
 840
 841 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
 842 {
 843         return __parse_nl_addr(info, addr, NULL, 0,
 844                                TCP_METRICS_ATTR_SADDR_IPV4,
 845                                TCP_METRICS_ATTR_SADDR_IPV6);
 846 }
 847
 848 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
 849 {
 850         struct tcp_metrics_block *tm;
 851         struct inetpeer_addr saddr, daddr;
 852         unsigned int hash;
 853         struct sk_buff *msg;
 854         struct net *net = genl_info_net(info);
 855         void *reply;
 856         int ret;
 857         bool src = true;
 858
 859         ret = parse_nl_addr(info, &daddr, &hash, 0);
 860         if (ret < 0)
 861                 return ret;
 862
 863         ret = parse_nl_saddr(info, &saddr);
 864         if (ret < 0)
 865                 src = false;
 866
 867         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 868         if (!msg)
 869                 return -ENOMEM;
 870
 871         reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
 872                                   info->genlhdr->cmd);
 873         if (!reply)
 874                 goto nla_put_failure;
 875
 876         hash ^= net_hash_mix(net);
 877         hash = hash_32(hash, tcp_metrics_hash_log);
 878         ret = -ESRCH;
 879         rcu_read_lock();
 880         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
 881              tm = rcu_dereference(tm->tcpm_next)) {
 882                 if (addr_same(&tm->tcpm_daddr, &daddr) &&
 883                     (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
 884                     net_eq(tm_net(tm), net)) {
 885                         ret = tcp_metrics_fill_info(msg, tm);
 886                         break;
 887                 }
 888         }
 889         rcu_read_unlock();
 890         if (ret < 0)
 891                 goto out_free;
 892
 893         genlmsg_end(msg, reply);
 894         return genlmsg_reply(msg, info);
 895
 896 nla_put_failure:
 897         ret = -EMSGSIZE;
 898
 899 out_free:
 900         nlmsg_free(msg);
 901         return ret;
 902 }
 903
 904 static void tcp_metrics_flush_all(struct net *net)
 905 {
 906         unsigned int max_rows = 1U << tcp_metrics_hash_log;
 907         struct tcpm_hash_bucket *hb = tcp_metrics_hash;
 908         struct tcp_metrics_block *tm;
 909         unsigned int row;
 910
 911         for (row = 0; row < max_rows; row++, hb++) {
 912                 struct tcp_metrics_block __rcu **pp;
 913                 spin_lock_bh(&tcp_metrics_lock);
 914                 pp = &hb->chain;
 915                 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
 916                         if (net_eq(tm_net(tm), net)) {
 917                                 *pp = tm->tcpm_next;
 918                                 kfree_rcu(tm, rcu_head);
 919                         } else {
 920                                 pp = &tm->tcpm_next;
 921                         }
 922                 }
 923                 spin_unlock_bh(&tcp_metrics_lock);
 924         }
 925 }
 926
 927 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
 928 {
 929         struct tcpm_hash_bucket *hb;
 930         struct tcp_metrics_block *tm;
 931         struct tcp_metrics_block __rcu **pp;
 932         struct inetpeer_addr saddr, daddr;
 933         unsigned int hash;
 934         struct net *net = genl_info_net(info);
 935         int ret;
 936         bool src = true, found = false;
 937
 938         ret = parse_nl_addr(info, &daddr, &hash, 1);
 939         if (ret < 0)
 940                 return ret;
 941         if (ret > 0) {
 942                 tcp_metrics_flush_all(net);
 943                 return 0;
 944         }
 945         ret = parse_nl_saddr(info, &saddr);
 946         if (ret < 0)
 947                 src = false;
 948
 949         hash ^= net_hash_mix(net);
 950         hash = hash_32(hash, tcp_metrics_hash_log);
 951         hb = tcp_metrics_hash + hash;
 952         pp = &hb->chain;
 953         spin_lock_bh(&tcp_metrics_lock);
 954         for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
 955                 if (addr_same(&tm->tcpm_daddr, &daddr) &&
 956                     (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
 957                     net_eq(tm_net(tm), net)) {
 958                         *pp = tm->tcpm_next;
 959                         kfree_rcu(tm, rcu_head);
 960                         found = true;
 961                 } else {
 962                         pp = &tm->tcpm_next;
 963                 }
 964         }
 965         spin_unlock_bh(&tcp_metrics_lock);
 966         if (!found)
 967                 return -ESRCH;
 968         return 0;
 969 }
 970
 971 static const struct genl_ops tcp_metrics_nl_ops[] = {
 972         {
 973                 .cmd = TCP_METRICS_CMD_GET,
 974                 .doit = tcp_metrics_nl_cmd_get,
 975                 .dumpit = tcp_metrics_nl_dump,
 976                 .policy = tcp_metrics_nl_policy,
 977         },
 978         {
 979                 .cmd = TCP_METRICS_CMD_DEL,
 980                 .doit = tcp_metrics_nl_cmd_del,
 981                 .policy = tcp_metrics_nl_policy,
 982                 .flags = GENL_ADMIN_PERM,
 983         },
 984 };
 985
 986 static struct genl_family tcp_metrics_nl_family __ro_after_init = {
 987         .hdrsize        = 0,
 988         .name           = TCP_METRICS_GENL_NAME,
 989         .version        = TCP_METRICS_GENL_VERSION,
 990         .maxattr        = TCP_METRICS_ATTR_MAX,
 991         .netnsok        = true,
 992         .module         = THIS_MODULE,
 993         .ops            = tcp_metrics_nl_ops,
 994         .n_ops          = ARRAY_SIZE(tcp_metrics_nl_ops),
 995 };
 996
 997 static unsigned int tcpmhash_entries;
 998 static int __init set_tcpmhash_entries(char *str)
 999 {
1000         ssize_t ret;
1001
1002         if (!str)
1003                 return 0;
1004
1005         ret = kstrtouint(str, 0, &tcpmhash_entries);
1006         if (ret)
1007                 return 0;
1008
1009         return 1;
1010 }
1011 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1012
1013 static int __net_init tcp_net_metrics_init(struct net *net)
1014 {
1015         size_t size;
1016         unsigned int slots;
1017
1018         if (!net_eq(net, &init_net))
1019                 return 0;
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         tcp_metrics_hash_log = order_base_2(slots);
1030         size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1031
1032         tcp_metrics_hash = kvzalloc(size, GFP_KERNEL);
1033         if (!tcp_metrics_hash)
1034                 return -ENOMEM;
1035
1036         return 0;
1037 }
1038
1039 static void __net_exit tcp_net_metrics_exit(struct net *net)
1040 {
1041         tcp_metrics_flush_all(net);
1042 }
1043
1044 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1045         .init   =       tcp_net_metrics_init,
1046         .exit   =       tcp_net_metrics_exit,
1047 };
1048
1049 void __init tcp_metrics_init(void)
1050 {
1051         int ret;
1052
1053         ret = register_pernet_subsys(&tcp_net_metrics_ops);
1054         if (ret < 0)
1055                 panic("Could not allocate the tcp_metrics hash table\n");
1056
1057         ret = genl_register_family(&tcp_metrics_nl_family);
1058         if (ret < 0)
1059                 panic("Could not register tcp_metrics generic netlink\n");
1060 }