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 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
  25                                                    const struct inetpeer_addr *daddr,
  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                 try_exp:2;              /* Request w/ exp. option (once) */
  32         unsigned long   last_syn_loss;  /* Last Fast Open SYN loss */
  33         struct  tcp_fastopen_cookie     cookie;
  34 };
  35
  36 /* TCP_METRIC_MAX includes 2 extra fields for userspace compatibility
  37  * Kernel only stores RTT and RTTVAR in usec resolution
  38  */
  39 #define TCP_METRIC_MAX_KERNEL (TCP_METRIC_MAX - 2)
  40
  41 struct tcp_metrics_block {
  42         struct tcp_metrics_block __rcu  *tcpm_next;
  43         possible_net_t                  tcpm_net;
  44         struct inetpeer_addr            tcpm_saddr;
  45         struct inetpeer_addr            tcpm_daddr;
  46         unsigned long                   tcpm_stamp;
  47         u32                             tcpm_lock;
  48         u32                             tcpm_vals[TCP_METRIC_MAX_KERNEL + 1];
  49         struct tcp_fastopen_metrics     tcpm_fastopen;
  50
  51         struct rcu_head                 rcu_head;
  52 };
  53
  54 static inline struct net *tm_net(struct tcp_metrics_block *tm)
  55 {
  56         return read_pnet(&tm->tcpm_net);
  57 }
  58
  59 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
  60                               enum tcp_metric_index idx)
  61 {
  62         /* Paired with WRITE_ONCE() in tcpm_suck_dst() */
  63         return READ_ONCE(tm->tcpm_lock) & (1 << idx);
  64 }
  65
  66 static u32 tcp_metric_get(const struct tcp_metrics_block *tm,
  67                           enum tcp_metric_index idx)
  68 {
  69         /* Paired with WRITE_ONCE() in tcp_metric_set() */
  70         return READ_ONCE(tm->tcpm_vals[idx]);
  71 }
  72
  73 static void tcp_metric_set(struct tcp_metrics_block *tm,
  74                            enum tcp_metric_index idx,
  75                            u32 val)
  76 {
  77         /* Paired with READ_ONCE() in tcp_metric_get() */
  78         WRITE_ONCE(tm->tcpm_vals[idx], val);
  79 }
  80
  81 static bool addr_same(const struct inetpeer_addr *a,
  82                       const struct inetpeer_addr *b)
  83 {
  84         return (a->family == b->family) && !inetpeer_addr_cmp(a, b);
  85 }
  86
  87 struct tcpm_hash_bucket {
  88         struct tcp_metrics_block __rcu  *chain;
  89 };
  90
  91 static struct tcpm_hash_bucket  *tcp_metrics_hash __read_mostly;
  92 static unsigned int             tcp_metrics_hash_log __read_mostly;
  93
  94 static DEFINE_SPINLOCK(tcp_metrics_lock);
  95
  96 static void tcpm_suck_dst(struct tcp_metrics_block *tm,
  97                           const struct dst_entry *dst,
  98                           bool fastopen_clear)
  99 {
 100         u32 msval;
 101         u32 val;
 102
 103         WRITE_ONCE(tm->tcpm_stamp, jiffies);
 104
 105         val = 0;
 106         if (dst_metric_locked(dst, RTAX_RTT))
 107                 val |= 1 << TCP_METRIC_RTT;
 108         if (dst_metric_locked(dst, RTAX_RTTVAR))
 109                 val |= 1 << TCP_METRIC_RTTVAR;
 110         if (dst_metric_locked(dst, RTAX_SSTHRESH))
 111                 val |= 1 << TCP_METRIC_SSTHRESH;
 112         if (dst_metric_locked(dst, RTAX_CWND))
 113                 val |= 1 << TCP_METRIC_CWND;
 114         if (dst_metric_locked(dst, RTAX_REORDERING))
 115                 val |= 1 << TCP_METRIC_REORDERING;
 116         /* Paired with READ_ONCE() in tcp_metric_locked() */
 117         WRITE_ONCE(tm->tcpm_lock, val);
 118
 119         msval = dst_metric_raw(dst, RTAX_RTT);
 120         tcp_metric_set(tm, TCP_METRIC_RTT, msval * USEC_PER_MSEC);
 121
 122         msval = dst_metric_raw(dst, RTAX_RTTVAR);
 123         tcp_metric_set(tm, TCP_METRIC_RTTVAR, msval * USEC_PER_MSEC);
 124         tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 125                        dst_metric_raw(dst, RTAX_SSTHRESH));
 126         tcp_metric_set(tm, TCP_METRIC_CWND,
 127                        dst_metric_raw(dst, RTAX_CWND));
 128         tcp_metric_set(tm, TCP_METRIC_REORDERING,
 129                        dst_metric_raw(dst, RTAX_REORDERING));
 130         if (fastopen_clear) {
 131                 tm->tcpm_fastopen.mss = 0;
 132                 tm->tcpm_fastopen.syn_loss = 0;
 133                 tm->tcpm_fastopen.try_exp = 0;
 134                 tm->tcpm_fastopen.cookie.exp = false;
 135                 tm->tcpm_fastopen.cookie.len = 0;
 136         }
 137 }
 138
 139 #define TCP_METRICS_TIMEOUT             (60 * 60 * HZ)
 140
 141 static void tcpm_check_stamp(struct tcp_metrics_block *tm,
 142                              const struct dst_entry *dst)
 143 {
 144         unsigned long limit;
 145
 146         if (!tm)
 147                 return;
 148         limit = READ_ONCE(tm->tcpm_stamp) + TCP_METRICS_TIMEOUT;
 149         if (unlikely(time_after(jiffies, limit)))
 150                 tcpm_suck_dst(tm, dst, false);
 151 }
 152
 153 #define TCP_METRICS_RECLAIM_DEPTH       5
 154 #define TCP_METRICS_RECLAIM_PTR         (struct tcp_metrics_block *) 0x1UL
 155
 156 #define deref_locked(p) \
 157         rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
 158
 159 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
 160                                           struct inetpeer_addr *saddr,
 161                                           struct inetpeer_addr *daddr,
 162                                           unsigned int hash)
 163 {
 164         struct tcp_metrics_block *tm;
 165         struct net *net;
 166         bool reclaim = false;
 167
 168         spin_lock_bh(&tcp_metrics_lock);
 169         net = dev_net(dst->dev);
 170
 171         /* While waiting for the spin-lock the cache might have been populated
 172          * with this entry and so we have to check again.
 173          */
 174         tm = __tcp_get_metrics(saddr, daddr, net, hash);
 175         if (tm == TCP_METRICS_RECLAIM_PTR) {
 176                 reclaim = true;
 177                 tm = NULL;
 178         }
 179         if (tm) {
 180                 tcpm_check_stamp(tm, dst);
 181                 goto out_unlock;
 182         }
 183
 184         if (unlikely(reclaim)) {
 185                 struct tcp_metrics_block *oldest;
 186
 187                 oldest = deref_locked(tcp_metrics_hash[hash].chain);
 188                 for (tm = deref_locked(oldest->tcpm_next); tm;
 189                      tm = deref_locked(tm->tcpm_next)) {
 190                         if (time_before(READ_ONCE(tm->tcpm_stamp),
 191                                         READ_ONCE(oldest->tcpm_stamp)))
 192                                 oldest = tm;
 193                 }
 194                 tm = oldest;
 195         } else {
 196                 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
 197                 if (!tm)
 198                         goto out_unlock;
 199         }
 200         write_pnet(&tm->tcpm_net, net);
 201         tm->tcpm_saddr = *saddr;
 202         tm->tcpm_daddr = *daddr;
 203
 204         tcpm_suck_dst(tm, dst, true);
 205
 206         if (likely(!reclaim)) {
 207                 tm->tcpm_next = tcp_metrics_hash[hash].chain;
 208                 rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
 209         }
 210
 211 out_unlock:
 212         spin_unlock_bh(&tcp_metrics_lock);
 213         return tm;
 214 }
 215
 216 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
 217 {
 218         if (tm)
 219                 return tm;
 220         if (depth > TCP_METRICS_RECLAIM_DEPTH)
 221                 return TCP_METRICS_RECLAIM_PTR;
 222         return NULL;
 223 }
 224
 225 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
 226                                                    const struct inetpeer_addr *daddr,
 227                                                    struct net *net, unsigned int hash)
 228 {
 229         struct tcp_metrics_block *tm;
 230         int depth = 0;
 231
 232         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
 233              tm = rcu_dereference(tm->tcpm_next)) {
 234                 if (addr_same(&tm->tcpm_saddr, saddr) &&
 235                     addr_same(&tm->tcpm_daddr, daddr) &&
 236                     net_eq(tm_net(tm), net))
 237                         break;
 238                 depth++;
 239         }
 240         return tcp_get_encode(tm, depth);
 241 }
 242
 243 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
 244                                                        struct dst_entry *dst)
 245 {
 246         struct tcp_metrics_block *tm;
 247         struct inetpeer_addr saddr, daddr;
 248         unsigned int hash;
 249         struct net *net;
 250
 251         saddr.family = req->rsk_ops->family;
 252         daddr.family = req->rsk_ops->family;
 253         switch (daddr.family) {
 254         case AF_INET:
 255                 inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
 256                 inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
 257                 hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
 258                 break;
 259 #if IS_ENABLED(CONFIG_IPV6)
 260         case AF_INET6:
 261                 inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
 262                 inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr);
 263                 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
 264                 break;
 265 #endif
 266         default:
 267                 return NULL;
 268         }
 269
 270         net = dev_net(dst->dev);
 271         hash ^= net_hash_mix(net);
 272         hash = hash_32(hash, tcp_metrics_hash_log);
 273
 274         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
 275              tm = rcu_dereference(tm->tcpm_next)) {
 276                 if (addr_same(&tm->tcpm_saddr, &saddr) &&
 277                     addr_same(&tm->tcpm_daddr, &daddr) &&
 278                     net_eq(tm_net(tm), net))
 279                         break;
 280         }
 281         tcpm_check_stamp(tm, dst);
 282         return tm;
 283 }
 284
 285 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
 286                                                  struct dst_entry *dst,
 287                                                  bool create)
 288 {
 289         struct tcp_metrics_block *tm;
 290         struct inetpeer_addr saddr, daddr;
 291         unsigned int hash;
 292         struct net *net;
 293
 294         if (sk->sk_family == AF_INET) {
 295                 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
 296                 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
 297                 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
 298         }
 299 #if IS_ENABLED(CONFIG_IPV6)
 300         else if (sk->sk_family == AF_INET6) {
 301                 if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
 302                         inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
 303                         inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
 304                         hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
 305                 } else {
 306                         inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
 307                         inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
 308                         hash = ipv6_addr_hash(&sk->sk_v6_daddr);
 309                 }
 310         }
 311 #endif
 312         else
 313                 return NULL;
 314
 315         net = dev_net(dst->dev);
 316         hash ^= net_hash_mix(net);
 317         hash = hash_32(hash, tcp_metrics_hash_log);
 318
 319         tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
 320         if (tm == TCP_METRICS_RECLAIM_PTR)
 321                 tm = NULL;
 322         if (!tm && create)
 323                 tm = tcpm_new(dst, &saddr, &daddr, hash);
 324         else
 325                 tcpm_check_stamp(tm, dst);
 326
 327         return tm;
 328 }
 329
 330 /* Save metrics learned by this TCP session.  This function is called
 331  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
 332  * or goes from LAST-ACK to CLOSE.
 333  */
 334 void tcp_update_metrics(struct sock *sk)
 335 {
 336         const struct inet_connection_sock *icsk = inet_csk(sk);
 337         struct dst_entry *dst = __sk_dst_get(sk);
 338         struct tcp_sock *tp = tcp_sk(sk);
 339         struct net *net = sock_net(sk);
 340         struct tcp_metrics_block *tm;
 341         unsigned long rtt;
 342         u32 val;
 343         int m;
 344
 345         sk_dst_confirm(sk);
 346         if (READ_ONCE(net->ipv4.sysctl_tcp_nometrics_save) || !dst)
 347                 return;
 348
 349         rcu_read_lock();
 350         if (icsk->icsk_backoff || !tp->srtt_us) {
 351                 /* This session failed to estimate rtt. Why?
 352                  * Probably, no packets returned in time.  Reset our
 353                  * results.
 354                  */
 355                 tm = tcp_get_metrics(sk, dst, false);
 356                 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
 357                         tcp_metric_set(tm, TCP_METRIC_RTT, 0);
 358                 goto out_unlock;
 359         } else
 360                 tm = tcp_get_metrics(sk, dst, true);
 361
 362         if (!tm)
 363                 goto out_unlock;
 364
 365         rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
 366         m = rtt - tp->srtt_us;
 367
 368         /* If newly calculated rtt larger than stored one, store new
 369          * one. Otherwise, use EWMA. Remember, rtt overestimation is
 370          * always better than underestimation.
 371          */
 372         if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
 373                 if (m <= 0)
 374                         rtt = tp->srtt_us;
 375                 else
 376                         rtt -= (m >> 3);
 377                 tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
 378         }
 379
 380         if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
 381                 unsigned long var;
 382
 383                 if (m < 0)
 384                         m = -m;
 385
 386                 /* Scale deviation to rttvar fixed point */
 387                 m >>= 1;
 388                 if (m < tp->mdev_us)
 389                         m = tp->mdev_us;
 390
 391                 var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
 392                 if (m >= var)
 393                         var = m;
 394                 else
 395                         var -= (var - m) >> 2;
 396
 397                 tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
 398         }
 399
 400         if (tcp_in_initial_slowstart(tp)) {
 401                 /* Slow start still did not finish. */
 402                 if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
 403                     !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 404                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 405                         if (val && (tcp_snd_cwnd(tp) >> 1) > val)
 406                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 407                                                tcp_snd_cwnd(tp) >> 1);
 408                 }
 409                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 410                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 411                         if (tcp_snd_cwnd(tp) > val)
 412                                 tcp_metric_set(tm, TCP_METRIC_CWND,
 413                                                tcp_snd_cwnd(tp));
 414                 }
 415         } else if (!tcp_in_slow_start(tp) &&
 416                    icsk->icsk_ca_state == TCP_CA_Open) {
 417                 /* Cong. avoidance phase, cwnd is reliable. */
 418                 if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
 419                     !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
 420                         tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 421                                        max(tcp_snd_cwnd(tp) >> 1, tp->snd_ssthresh));
 422                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 423                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 424                         tcp_metric_set(tm, TCP_METRIC_CWND, (val + tcp_snd_cwnd(tp)) >> 1);
 425                 }
 426         } else {
 427                 /* Else slow start did not finish, cwnd is non-sense,
 428                  * ssthresh may be also invalid.
 429                  */
 430                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 431                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 432                         tcp_metric_set(tm, TCP_METRIC_CWND,
 433                                        (val + tp->snd_ssthresh) >> 1);
 434                 }
 435                 if (!READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) &&
 436                     !tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 437                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 438                         if (val && tp->snd_ssthresh > val)
 439                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 440                                                tp->snd_ssthresh);
 441                 }
 442                 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
 443                         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 444                         if (val < tp->reordering &&
 445                             tp->reordering !=
 446                             READ_ONCE(net->ipv4.sysctl_tcp_reordering))
 447                                 tcp_metric_set(tm, TCP_METRIC_REORDERING,
 448                                                tp->reordering);
 449                 }
 450         }
 451         WRITE_ONCE(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 net *net = sock_net(sk);
 463         struct tcp_metrics_block *tm;
 464         u32 val, crtt = 0; /* cached RTT scaled by 8 */
 465
 466         sk_dst_confirm(sk);
 467         if (!dst)
 468                 goto reset;
 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 = READ_ONCE(net->ipv4.sysctl_tcp_no_ssthresh_metrics_save) ?
 481               0 : tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 482         if (val) {
 483                 tp->snd_ssthresh = val;
 484                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
 485                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
 486         } else {
 487                 /* ssthresh may have been reduced unnecessarily during.
 488                  * 3WHS. Restore it back to its initial default.
 489                  */
 490                 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 491         }
 492         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 493         if (val && tp->reordering != val)
 494                 tp->reordering = val;
 495
 496         crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
 497         rcu_read_unlock();
 498 reset:
 499         /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
 500          * to seed the RTO for later data packets because SYN packets are
 501          * small. Use the per-dst cached values to seed the RTO but keep
 502          * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
 503          * Later the RTO will be updated immediately upon obtaining the first
 504          * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
 505          * influences the first RTO but not later RTT estimation.
 506          *
 507          * But if RTT is not available from the SYN (due to retransmits or
 508          * syn cookies) or the cache, force a conservative 3secs timeout.
 509          *
 510          * A bit of theory. RTT is time passed after "normal" sized packet
 511          * is sent until it is ACKed. In normal circumstances sending small
 512          * packets force peer to delay ACKs and calculation is correct too.
 513          * The algorithm is adaptive and, provided we follow specs, it
 514          * NEVER underestimate RTT. BUT! If peer tries to make some clever
 515          * tricks sort of "quick acks" for time long enough to decrease RTT
 516          * to low value, and then abruptly stops to do it and starts to delay
 517          * ACKs, wait for troubles.
 518          */
 519         if (crtt > tp->srtt_us) {
 520                 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
 521                 crtt /= 8 * USEC_PER_SEC / HZ;
 522                 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
 523         } else if (tp->srtt_us == 0) {
 524                 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
 525                  * 3WHS. This is most likely due to retransmission,
 526                  * including spurious one. Reset the RTO back to 3secs
 527                  * from the more aggressive 1sec to avoid more spurious
 528                  * retransmission.
 529                  */
 530                 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
 531                 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
 532
 533                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
 534         }
 535 }
 536
 537 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
 538 {
 539         struct tcp_metrics_block *tm;
 540         bool ret;
 541
 542         if (!dst)
 543                 return false;
 544
 545         rcu_read_lock();
 546         tm = __tcp_get_metrics_req(req, dst);
 547         if (tm && tcp_metric_get(tm, TCP_METRIC_RTT))
 548                 ret = true;
 549         else
 550                 ret = false;
 551         rcu_read_unlock();
 552
 553         return ret;
 554 }
 555
 556 static DEFINE_SEQLOCK(fastopen_seqlock);
 557
 558 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
 559                             struct tcp_fastopen_cookie *cookie)
 560 {
 561         struct tcp_metrics_block *tm;
 562
 563         rcu_read_lock();
 564         tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
 565         if (tm) {
 566                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 567                 unsigned int seq;
 568
 569                 do {
 570                         seq = read_seqbegin(&fastopen_seqlock);
 571                         if (tfom->mss)
 572                                 *mss = tfom->mss;
 573                         *cookie = tfom->cookie;
 574                         if (cookie->len <= 0 && tfom->try_exp == 1)
 575                                 cookie->exp = true;
 576                 } while (read_seqretry(&fastopen_seqlock, seq));
 577         }
 578         rcu_read_unlock();
 579 }
 580
 581 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
 582                             struct tcp_fastopen_cookie *cookie, bool syn_lost,
 583                             u16 try_exp)
 584 {
 585         struct dst_entry *dst = __sk_dst_get(sk);
 586         struct tcp_metrics_block *tm;
 587
 588         if (!dst)
 589                 return;
 590         rcu_read_lock();
 591         tm = tcp_get_metrics(sk, dst, true);
 592         if (tm) {
 593                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 594
 595                 write_seqlock_bh(&fastopen_seqlock);
 596                 if (mss)
 597                         tfom->mss = mss;
 598                 if (cookie && cookie->len > 0)
 599                         tfom->cookie = *cookie;
 600                 else if (try_exp > tfom->try_exp &&
 601                          tfom->cookie.len <= 0 && !tfom->cookie.exp)
 602                         tfom->try_exp = try_exp;
 603                 if (syn_lost) {
 604                         ++tfom->syn_loss;
 605                         tfom->last_syn_loss = jiffies;
 606                 } else
 607                         tfom->syn_loss = 0;
 608                 write_sequnlock_bh(&fastopen_seqlock);
 609         }
 610         rcu_read_unlock();
 611 }
 612
 613 static struct genl_family tcp_metrics_nl_family;
 614
 615 static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
 616         [TCP_METRICS_ATTR_ADDR_IPV4]    = { .type = NLA_U32, },
 617         [TCP_METRICS_ATTR_ADDR_IPV6]    = { .type = NLA_BINARY,
 618                                             .len = sizeof(struct in6_addr), },
 619         /* Following attributes are not received for GET/DEL,
 620          * we keep them for reference
 621          */
 622 #if 0
 623         [TCP_METRICS_ATTR_AGE]          = { .type = NLA_MSECS, },
 624         [TCP_METRICS_ATTR_TW_TSVAL]     = { .type = NLA_U32, },
 625         [TCP_METRICS_ATTR_TW_TS_STAMP]  = { .type = NLA_S32, },
 626         [TCP_METRICS_ATTR_VALS]         = { .type = NLA_NESTED, },
 627         [TCP_METRICS_ATTR_FOPEN_MSS]    = { .type = NLA_U16, },
 628         [TCP_METRICS_ATTR_FOPEN_SYN_DROPS]      = { .type = NLA_U16, },
 629         [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]    = { .type = NLA_MSECS, },
 630         [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
 631                                             .len = TCP_FASTOPEN_COOKIE_MAX, },
 632 #endif
 633 };
 634
 635 /* Add attributes, caller cancels its header on failure */
 636 static int tcp_metrics_fill_info(struct sk_buff *msg,
 637                                  struct tcp_metrics_block *tm)
 638 {
 639         struct nlattr *nest;
 640         int i;
 641
 642         switch (tm->tcpm_daddr.family) {
 643         case AF_INET:
 644                 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
 645                                     inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
 646                         goto nla_put_failure;
 647                 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
 648                                     inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
 649                         goto nla_put_failure;
 650                 break;
 651         case AF_INET6:
 652                 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
 653                                      inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
 654                         goto nla_put_failure;
 655                 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
 656                                      inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
 657                         goto nla_put_failure;
 658                 break;
 659         default:
 660                 return -EAFNOSUPPORT;
 661         }
 662
 663         if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
 664                           jiffies - READ_ONCE(tm->tcpm_stamp),
 665                           TCP_METRICS_ATTR_PAD) < 0)
 666                 goto nla_put_failure;
 667
 668         {
 669                 int n = 0;
 670
 671                 nest = nla_nest_start_noflag(msg, TCP_METRICS_ATTR_VALS);
 672                 if (!nest)
 673                         goto nla_put_failure;
 674                 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
 675                         u32 val = tcp_metric_get(tm, i);
 676
 677                         if (!val)
 678                                 continue;
 679                         if (i == TCP_METRIC_RTT) {
 680                                 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
 681                                                 val) < 0)
 682                                         goto nla_put_failure;
 683                                 n++;
 684                                 val = max(val / 1000, 1U);
 685                         }
 686                         if (i == TCP_METRIC_RTTVAR) {
 687                                 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
 688                                                 val) < 0)
 689                                         goto nla_put_failure;
 690                                 n++;
 691                                 val = max(val / 1000, 1U);
 692                         }
 693                         if (nla_put_u32(msg, i + 1, val) < 0)
 694                                 goto nla_put_failure;
 695                         n++;
 696                 }
 697                 if (n)
 698                         nla_nest_end(msg, nest);
 699                 else
 700                         nla_nest_cancel(msg, nest);
 701         }
 702
 703         {
 704                 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
 705                 unsigned int seq;
 706
 707                 do {
 708                         seq = read_seqbegin(&fastopen_seqlock);
 709                         tfom_copy[0] = tm->tcpm_fastopen;
 710                 } while (read_seqretry(&fastopen_seqlock, seq));
 711
 712                 tfom = tfom_copy;
 713                 if (tfom->mss &&
 714                     nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
 715                                 tfom->mss) < 0)
 716                         goto nla_put_failure;
 717                 if (tfom->syn_loss &&
 718                     (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
 719                                 tfom->syn_loss) < 0 ||
 720                      nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
 721                                 jiffies - tfom->last_syn_loss,
 722                                 TCP_METRICS_ATTR_PAD) < 0))
 723                         goto nla_put_failure;
 724                 if (tfom->cookie.len > 0 &&
 725                     nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
 726                             tfom->cookie.len, tfom->cookie.val) < 0)
 727                         goto nla_put_failure;
 728         }
 729
 730         return 0;
 731
 732 nla_put_failure:
 733         return -EMSGSIZE;
 734 }
 735
 736 static int tcp_metrics_dump_info(struct sk_buff *skb,
 737                                  struct netlink_callback *cb,
 738                                  struct tcp_metrics_block *tm)
 739 {
 740         void *hdr;
 741
 742         hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 743                           &tcp_metrics_nl_family, NLM_F_MULTI,
 744                           TCP_METRICS_CMD_GET);
 745         if (!hdr)
 746                 return -EMSGSIZE;
 747
 748         if (tcp_metrics_fill_info(skb, tm) < 0)
 749                 goto nla_put_failure;
 750
 751         genlmsg_end(skb, hdr);
 752         return 0;
 753
 754 nla_put_failure:
 755         genlmsg_cancel(skb, hdr);
 756         return -EMSGSIZE;
 757 }
 758
 759 static int tcp_metrics_nl_dump(struct sk_buff *skb,
 760                                struct netlink_callback *cb)
 761 {
 762         struct net *net = sock_net(skb->sk);
 763         unsigned int max_rows = 1U << tcp_metrics_hash_log;
 764         unsigned int row, s_row = cb->args[0];
 765         int s_col = cb->args[1], col = s_col;
 766
 767         for (row = s_row; row < max_rows; row++, s_col = 0) {
 768                 struct tcp_metrics_block *tm;
 769                 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
 770
 771                 rcu_read_lock();
 772                 for (col = 0, tm = rcu_dereference(hb->chain); tm;
 773                      tm = rcu_dereference(tm->tcpm_next), col++) {
 774                         if (!net_eq(tm_net(tm), net))
 775                                 continue;
 776                         if (col < s_col)
 777                                 continue;
 778                         if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
 779                                 rcu_read_unlock();
 780                                 goto done;
 781                         }
 782                 }
 783                 rcu_read_unlock();
 784         }
 785
 786 done:
 787         cb->args[0] = row;
 788         cb->args[1] = col;
 789         return skb->len;
 790 }
 791
 792 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 793                            unsigned int *hash, int optional, int v4, int v6)
 794 {
 795         struct nlattr *a;
 796
 797         a = info->attrs[v4];
 798         if (a) {
 799                 inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
 800                 if (hash)
 801                         *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
 802                 return 0;
 803         }
 804         a = info->attrs[v6];
 805         if (a) {
 806                 struct in6_addr in6;
 807
 808                 if (nla_len(a) != sizeof(struct in6_addr))
 809                         return -EINVAL;
 810                 in6 = nla_get_in6_addr(a);
 811                 inetpeer_set_addr_v6(addr, &in6);
 812                 if (hash)
 813                         *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
 814                 return 0;
 815         }
 816         return optional ? 1 : -EAFNOSUPPORT;
 817 }
 818
 819 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 820                          unsigned int *hash, int optional)
 821 {
 822         return __parse_nl_addr(info, addr, hash, optional,
 823                                TCP_METRICS_ATTR_ADDR_IPV4,
 824                                TCP_METRICS_ATTR_ADDR_IPV6);
 825 }
 826
 827 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
 828 {
 829         return __parse_nl_addr(info, addr, NULL, 0,
 830                                TCP_METRICS_ATTR_SADDR_IPV4,
 831                                TCP_METRICS_ATTR_SADDR_IPV6);
 832 }
 833
 834 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
 835 {
 836         struct tcp_metrics_block *tm;
 837         struct inetpeer_addr saddr, daddr;
 838         unsigned int hash;
 839         struct sk_buff *msg;
 840         struct net *net = genl_info_net(info);
 841         void *reply;
 842         int ret;
 843         bool src = true;
 844
 845         ret = parse_nl_addr(info, &daddr, &hash, 0);
 846         if (ret < 0)
 847                 return ret;
 848
 849         ret = parse_nl_saddr(info, &saddr);
 850         if (ret < 0)
 851                 src = false;
 852
 853         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 854         if (!msg)
 855                 return -ENOMEM;
 856
 857         reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
 858                                   info->genlhdr->cmd);
 859         if (!reply)
 860                 goto nla_put_failure;
 861
 862         hash ^= net_hash_mix(net);
 863         hash = hash_32(hash, tcp_metrics_hash_log);
 864         ret = -ESRCH;
 865         rcu_read_lock();
 866         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
 867              tm = rcu_dereference(tm->tcpm_next)) {
 868                 if (addr_same(&tm->tcpm_daddr, &daddr) &&
 869                     (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
 870                     net_eq(tm_net(tm), net)) {
 871                         ret = tcp_metrics_fill_info(msg, tm);
 872                         break;
 873                 }
 874         }
 875         rcu_read_unlock();
 876         if (ret < 0)
 877                 goto out_free;
 878
 879         genlmsg_end(msg, reply);
 880         return genlmsg_reply(msg, info);
 881
 882 nla_put_failure:
 883         ret = -EMSGSIZE;
 884
 885 out_free:
 886         nlmsg_free(msg);
 887         return ret;
 888 }
 889
 890 static void tcp_metrics_flush_all(struct net *net)
 891 {
 892         unsigned int max_rows = 1U << tcp_metrics_hash_log;
 893         struct tcpm_hash_bucket *hb = tcp_metrics_hash;
 894         struct tcp_metrics_block *tm;
 895         unsigned int row;
 896
 897         for (row = 0; row < max_rows; row++, hb++) {
 898                 struct tcp_metrics_block __rcu **pp;
 899                 bool match;
 900
 901                 spin_lock_bh(&tcp_metrics_lock);
 902                 pp = &hb->chain;
 903                 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
 904                         match = net ? net_eq(tm_net(tm), net) :
 905                                 !refcount_read(&tm_net(tm)->ns.count);
 906                         if (match) {
 907                                 *pp = tm->tcpm_next;
 908                                 kfree_rcu(tm, rcu_head);
 909                         } else {
 910                                 pp = &tm->tcpm_next;
 911                         }
 912                 }
 913                 spin_unlock_bh(&tcp_metrics_lock);
 914         }
 915 }
 916
 917 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
 918 {
 919         struct tcpm_hash_bucket *hb;
 920         struct tcp_metrics_block *tm;
 921         struct tcp_metrics_block __rcu **pp;
 922         struct inetpeer_addr saddr, daddr;
 923         unsigned int hash;
 924         struct net *net = genl_info_net(info);
 925         int ret;
 926         bool src = true, found = false;
 927
 928         ret = parse_nl_addr(info, &daddr, &hash, 1);
 929         if (ret < 0)
 930                 return ret;
 931         if (ret > 0) {
 932                 tcp_metrics_flush_all(net);
 933                 return 0;
 934         }
 935         ret = parse_nl_saddr(info, &saddr);
 936         if (ret < 0)
 937                 src = false;
 938
 939         hash ^= net_hash_mix(net);
 940         hash = hash_32(hash, tcp_metrics_hash_log);
 941         hb = tcp_metrics_hash + hash;
 942         pp = &hb->chain;
 943         spin_lock_bh(&tcp_metrics_lock);
 944         for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
 945                 if (addr_same(&tm->tcpm_daddr, &daddr) &&
 946                     (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
 947                     net_eq(tm_net(tm), net)) {
 948                         *pp = tm->tcpm_next;
 949                         kfree_rcu(tm, rcu_head);
 950                         found = true;
 951                 } else {
 952                         pp = &tm->tcpm_next;
 953                 }
 954         }
 955         spin_unlock_bh(&tcp_metrics_lock);
 956         if (!found)
 957                 return -ESRCH;
 958         return 0;
 959 }
 960
 961 static const struct genl_small_ops tcp_metrics_nl_ops[] = {
 962         {
 963                 .cmd = TCP_METRICS_CMD_GET,
 964                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 965                 .doit = tcp_metrics_nl_cmd_get,
 966                 .dumpit = tcp_metrics_nl_dump,
 967         },
 968         {
 969                 .cmd = TCP_METRICS_CMD_DEL,
 970                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 971                 .doit = tcp_metrics_nl_cmd_del,
 972                 .flags = GENL_ADMIN_PERM,
 973         },
 974 };
 975
 976 static struct genl_family tcp_metrics_nl_family __ro_after_init = {
 977         .hdrsize        = 0,
 978         .name           = TCP_METRICS_GENL_NAME,
 979         .version        = TCP_METRICS_GENL_VERSION,
 980         .maxattr        = TCP_METRICS_ATTR_MAX,
 981         .policy = tcp_metrics_nl_policy,
 982         .netnsok        = true,
 983         .module         = THIS_MODULE,
 984         .small_ops      = tcp_metrics_nl_ops,
 985         .n_small_ops    = ARRAY_SIZE(tcp_metrics_nl_ops),
 986         .resv_start_op  = TCP_METRICS_CMD_DEL + 1,
 987 };
 988
 989 static unsigned int tcpmhash_entries;
 990 static int __init set_tcpmhash_entries(char *str)
 991 {
 992         ssize_t ret;
 993
 994         if (!str)
 995                 return 0;
 996
 997         ret = kstrtouint(str, 0, &tcpmhash_entries);
 998         if (ret)
 999                 return 0;
1000
1001         return 1;
1002 }
1003 __setup("tcpmhash_entries=", set_tcpmhash_entries);
1004
1005 static int __net_init tcp_net_metrics_init(struct net *net)
1006 {
1007         size_t size;
1008         unsigned int slots;
1009
1010         if (!net_eq(net, &init_net))
1011                 return 0;
1012
1013         slots = tcpmhash_entries;
1014         if (!slots) {
1015                 if (totalram_pages() >= 128 * 1024)
1016                         slots = 16 * 1024;
1017                 else
1018                         slots = 8 * 1024;
1019         }
1020
1021         tcp_metrics_hash_log = order_base_2(slots);
1022         size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1023
1024         tcp_metrics_hash = kvzalloc(size, GFP_KERNEL);
1025         if (!tcp_metrics_hash)
1026                 return -ENOMEM;
1027
1028         return 0;
1029 }
1030
1031 static void __net_exit tcp_net_metrics_exit_batch(struct list_head *net_exit_list)
1032 {
1033         tcp_metrics_flush_all(NULL);
1034 }
1035
1036 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1037         .init           =       tcp_net_metrics_init,
1038         .exit_batch     =       tcp_net_metrics_exit_batch,
1039 };
1040
1041 void __init tcp_metrics_init(void)
1042 {
1043         int ret;
1044
1045         ret = register_pernet_subsys(&tcp_net_metrics_ops);
1046         if (ret < 0)
1047                 panic("Could not allocate the tcp_metrics hash table\n");
1048
1049         ret = genl_register_family(&tcp_metrics_nl_family);
1050         if (ret < 0)
1051                 panic("Could not register tcp_metrics generic netlink\n");
1052 }