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         possible_net_t                  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(struct tcp_metrics_block *tm)
  57 {
  58         return read_pnet(&tm->tcpm_net);
  59 }
  60
  61 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
  62                               enum tcp_metric_index idx)
  63 {
  64         return tm->tcpm_lock & (1 << idx);
  65 }
  66
  67 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
  68                           enum tcp_metric_index idx)
  69 {
  70         return 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         tm->tcpm_vals[idx] = val;
  78 }
  79
  80 static bool addr_same(const struct inetpeer_addr *a,
  81                       const struct inetpeer_addr *b)
  82 {
  83         return inetpeer_addr_cmp(a, b) == 0;
  84 }
  85
  86 struct tcpm_hash_bucket {
  87         struct tcp_metrics_block __rcu  *chain;
  88 };
  89
  90 static struct tcpm_hash_bucket  *tcp_metrics_hash __read_mostly;
  91 static unsigned int             tcp_metrics_hash_log __read_mostly;
  92
  93 static DEFINE_SPINLOCK(tcp_metrics_lock);
  94
  95 static void tcpm_suck_dst(struct tcp_metrics_block *tm,
  96                           const struct dst_entry *dst,
  97                           bool fastopen_clear)
  98 {
  99         u32 msval;
 100         u32 val;
 101
 102         tm->tcpm_stamp = jiffies;
 103
 104         val = 0;
 105         if (dst_metric_locked(dst, RTAX_RTT))
 106                 val |= 1 << TCP_METRIC_RTT;
 107         if (dst_metric_locked(dst, RTAX_RTTVAR))
 108                 val |= 1 << TCP_METRIC_RTTVAR;
 109         if (dst_metric_locked(dst, RTAX_SSTHRESH))
 110                 val |= 1 << TCP_METRIC_SSTHRESH;
 111         if (dst_metric_locked(dst, RTAX_CWND))
 112                 val |= 1 << TCP_METRIC_CWND;
 113         if (dst_metric_locked(dst, RTAX_REORDERING))
 114                 val |= 1 << TCP_METRIC_REORDERING;
 115         tm->tcpm_lock = val;
 116
 117         msval = dst_metric_raw(dst, RTAX_RTT);
 118         tm->tcpm_vals[TCP_METRIC_RTT] = msval * USEC_PER_MSEC;
 119
 120         msval = dst_metric_raw(dst, RTAX_RTTVAR);
 121         tm->tcpm_vals[TCP_METRIC_RTTVAR] = msval * USEC_PER_MSEC;
 122         tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
 123         tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
 124         tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
 125         if (fastopen_clear) {
 126                 tm->tcpm_fastopen.mss = 0;
 127                 tm->tcpm_fastopen.syn_loss = 0;
 128                 tm->tcpm_fastopen.try_exp = 0;
 129                 tm->tcpm_fastopen.cookie.exp = false;
 130                 tm->tcpm_fastopen.cookie.len = 0;
 131         }
 132 }
 133
 134 #define TCP_METRICS_TIMEOUT             (60 * 60 * HZ)
 135
 136 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
 137 {
 138         if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
 139                 tcpm_suck_dst(tm, dst, false);
 140 }
 141
 142 #define TCP_METRICS_RECLAIM_DEPTH       5
 143 #define TCP_METRICS_RECLAIM_PTR         (struct tcp_metrics_block *) 0x1UL
 144
 145 #define deref_locked(p) \
 146         rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
 147
 148 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
 149                                           struct inetpeer_addr *saddr,
 150                                           struct inetpeer_addr *daddr,
 151                                           unsigned int hash)
 152 {
 153         struct tcp_metrics_block *tm;
 154         struct net *net;
 155         bool reclaim = false;
 156
 157         spin_lock_bh(&tcp_metrics_lock);
 158         net = dev_net(dst->dev);
 159
 160         /* While waiting for the spin-lock the cache might have been populated
 161          * with this entry and so we have to check again.
 162          */
 163         tm = __tcp_get_metrics(saddr, daddr, net, hash);
 164         if (tm == TCP_METRICS_RECLAIM_PTR) {
 165                 reclaim = true;
 166                 tm = NULL;
 167         }
 168         if (tm) {
 169                 tcpm_check_stamp(tm, dst);
 170                 goto out_unlock;
 171         }
 172
 173         if (unlikely(reclaim)) {
 174                 struct tcp_metrics_block *oldest;
 175
 176                 oldest = deref_locked(tcp_metrics_hash[hash].chain);
 177                 for (tm = deref_locked(oldest->tcpm_next); tm;
 178                      tm = deref_locked(tm->tcpm_next)) {
 179                         if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
 180                                 oldest = tm;
 181                 }
 182                 tm = oldest;
 183         } else {
 184                 tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
 185                 if (!tm)
 186                         goto out_unlock;
 187         }
 188         write_pnet(&tm->tcpm_net, net);
 189         tm->tcpm_saddr = *saddr;
 190         tm->tcpm_daddr = *daddr;
 191
 192         tcpm_suck_dst(tm, dst, true);
 193
 194         if (likely(!reclaim)) {
 195                 tm->tcpm_next = tcp_metrics_hash[hash].chain;
 196                 rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
 197         }
 198
 199 out_unlock:
 200         spin_unlock_bh(&tcp_metrics_lock);
 201         return tm;
 202 }
 203
 204 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
 205 {
 206         if (tm)
 207                 return tm;
 208         if (depth > TCP_METRICS_RECLAIM_DEPTH)
 209                 return TCP_METRICS_RECLAIM_PTR;
 210         return NULL;
 211 }
 212
 213 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
 214                                                    const struct inetpeer_addr *daddr,
 215                                                    struct net *net, unsigned int hash)
 216 {
 217         struct tcp_metrics_block *tm;
 218         int depth = 0;
 219
 220         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
 221              tm = rcu_dereference(tm->tcpm_next)) {
 222                 if (addr_same(&tm->tcpm_saddr, saddr) &&
 223                     addr_same(&tm->tcpm_daddr, daddr) &&
 224                     net_eq(tm_net(tm), net))
 225                         break;
 226                 depth++;
 227         }
 228         return tcp_get_encode(tm, depth);
 229 }
 230
 231 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
 232                                                        struct dst_entry *dst)
 233 {
 234         struct tcp_metrics_block *tm;
 235         struct inetpeer_addr saddr, daddr;
 236         unsigned int hash;
 237         struct net *net;
 238
 239         saddr.family = req->rsk_ops->family;
 240         daddr.family = req->rsk_ops->family;
 241         switch (daddr.family) {
 242         case AF_INET:
 243                 inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
 244                 inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
 245                 hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
 246                 break;
 247 #if IS_ENABLED(CONFIG_IPV6)
 248         case AF_INET6:
 249                 inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
 250                 inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr);
 251                 hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
 252                 break;
 253 #endif
 254         default:
 255                 return NULL;
 256         }
 257
 258         net = dev_net(dst->dev);
 259         hash ^= net_hash_mix(net);
 260         hash = hash_32(hash, tcp_metrics_hash_log);
 261
 262         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
 263              tm = rcu_dereference(tm->tcpm_next)) {
 264                 if (addr_same(&tm->tcpm_saddr, &saddr) &&
 265                     addr_same(&tm->tcpm_daddr, &daddr) &&
 266                     net_eq(tm_net(tm), net))
 267                         break;
 268         }
 269         tcpm_check_stamp(tm, dst);
 270         return tm;
 271 }
 272
 273 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
 274                                                  struct dst_entry *dst,
 275                                                  bool create)
 276 {
 277         struct tcp_metrics_block *tm;
 278         struct inetpeer_addr saddr, daddr;
 279         unsigned int hash;
 280         struct net *net;
 281
 282         if (sk->sk_family == AF_INET) {
 283                 inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
 284                 inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
 285                 hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
 286         }
 287 #if IS_ENABLED(CONFIG_IPV6)
 288         else if (sk->sk_family == AF_INET6) {
 289                 if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
 290                         inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
 291                         inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
 292                         hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
 293                 } else {
 294                         inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
 295                         inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
 296                         hash = ipv6_addr_hash(&sk->sk_v6_daddr);
 297                 }
 298         }
 299 #endif
 300         else
 301                 return NULL;
 302
 303         net = dev_net(dst->dev);
 304         hash ^= net_hash_mix(net);
 305         hash = hash_32(hash, tcp_metrics_hash_log);
 306
 307         tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
 308         if (tm == TCP_METRICS_RECLAIM_PTR)
 309                 tm = NULL;
 310         if (!tm && create)
 311                 tm = tcpm_new(dst, &saddr, &daddr, hash);
 312         else
 313                 tcpm_check_stamp(tm, dst);
 314
 315         return tm;
 316 }
 317
 318 /* Save metrics learned by this TCP session.  This function is called
 319  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
 320  * or goes from LAST-ACK to CLOSE.
 321  */
 322 void tcp_update_metrics(struct sock *sk)
 323 {
 324         const struct inet_connection_sock *icsk = inet_csk(sk);
 325         struct dst_entry *dst = __sk_dst_get(sk);
 326         struct tcp_sock *tp = tcp_sk(sk);
 327         struct net *net = sock_net(sk);
 328         struct tcp_metrics_block *tm;
 329         unsigned long rtt;
 330         u32 val;
 331         int m;
 332
 333         sk_dst_confirm(sk);
 334         if (sysctl_tcp_nometrics_save || !dst)
 335                 return;
 336
 337         rcu_read_lock();
 338         if (icsk->icsk_backoff || !tp->srtt_us) {
 339                 /* This session failed to estimate rtt. Why?
 340                  * Probably, no packets returned in time.  Reset our
 341                  * results.
 342                  */
 343                 tm = tcp_get_metrics(sk, dst, false);
 344                 if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
 345                         tcp_metric_set(tm, TCP_METRIC_RTT, 0);
 346                 goto out_unlock;
 347         } else
 348                 tm = tcp_get_metrics(sk, dst, true);
 349
 350         if (!tm)
 351                 goto out_unlock;
 352
 353         rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
 354         m = rtt - tp->srtt_us;
 355
 356         /* If newly calculated rtt larger than stored one, store new
 357          * one. Otherwise, use EWMA. Remember, rtt overestimation is
 358          * always better than underestimation.
 359          */
 360         if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
 361                 if (m <= 0)
 362                         rtt = tp->srtt_us;
 363                 else
 364                         rtt -= (m >> 3);
 365                 tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
 366         }
 367
 368         if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
 369                 unsigned long var;
 370
 371                 if (m < 0)
 372                         m = -m;
 373
 374                 /* Scale deviation to rttvar fixed point */
 375                 m >>= 1;
 376                 if (m < tp->mdev_us)
 377                         m = tp->mdev_us;
 378
 379                 var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
 380                 if (m >= var)
 381                         var = m;
 382                 else
 383                         var -= (var - m) >> 2;
 384
 385                 tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
 386         }
 387
 388         if (tcp_in_initial_slowstart(tp)) {
 389                 /* Slow start still did not finish. */
 390                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 391                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 392                         if (val && (tp->snd_cwnd >> 1) > val)
 393                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 394                                                tp->snd_cwnd >> 1);
 395                 }
 396                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 397                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 398                         if (tp->snd_cwnd > val)
 399                                 tcp_metric_set(tm, TCP_METRIC_CWND,
 400                                                tp->snd_cwnd);
 401                 }
 402         } else if (!tcp_in_slow_start(tp) &&
 403                    icsk->icsk_ca_state == TCP_CA_Open) {
 404                 /* Cong. avoidance phase, cwnd is reliable. */
 405                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
 406                         tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 407                                        max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
 408                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 409                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 410                         tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
 411                 }
 412         } else {
 413                 /* Else slow start did not finish, cwnd is non-sense,
 414                  * ssthresh may be also invalid.
 415                  */
 416                 if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
 417                         val = tcp_metric_get(tm, TCP_METRIC_CWND);
 418                         tcp_metric_set(tm, TCP_METRIC_CWND,
 419                                        (val + tp->snd_ssthresh) >> 1);
 420                 }
 421                 if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
 422                         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 423                         if (val && tp->snd_ssthresh > val)
 424                                 tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
 425                                                tp->snd_ssthresh);
 426                 }
 427                 if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
 428                         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 429                         if (val < tp->reordering &&
 430                             tp->reordering != net->ipv4.sysctl_tcp_reordering)
 431                                 tcp_metric_set(tm, TCP_METRIC_REORDERING,
 432                                                tp->reordering);
 433                 }
 434         }
 435         tm->tcpm_stamp = jiffies;
 436 out_unlock:
 437         rcu_read_unlock();
 438 }
 439
 440 /* Initialize metrics on socket. */
 441
 442 void tcp_init_metrics(struct sock *sk)
 443 {
 444         struct dst_entry *dst = __sk_dst_get(sk);
 445         struct tcp_sock *tp = tcp_sk(sk);
 446         struct tcp_metrics_block *tm;
 447         u32 val, crtt = 0; /* cached RTT scaled by 8 */
 448
 449         sk_dst_confirm(sk);
 450         if (!dst)
 451                 goto reset;
 452
 453         rcu_read_lock();
 454         tm = tcp_get_metrics(sk, dst, true);
 455         if (!tm) {
 456                 rcu_read_unlock();
 457                 goto reset;
 458         }
 459
 460         if (tcp_metric_locked(tm, TCP_METRIC_CWND))
 461                 tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
 462
 463         val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
 464         if (val) {
 465                 tp->snd_ssthresh = val;
 466                 if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
 467                         tp->snd_ssthresh = tp->snd_cwnd_clamp;
 468         } else {
 469                 /* ssthresh may have been reduced unnecessarily during.
 470                  * 3WHS. Restore it back to its initial default.
 471                  */
 472                 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
 473         }
 474         val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
 475         if (val && tp->reordering != val) {
 476                 tcp_disable_fack(tp);
 477                 tp->reordering = val;
 478         }
 479
 480         crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
 481         rcu_read_unlock();
 482 reset:
 483         /* The initial RTT measurement from the SYN/SYN-ACK is not ideal
 484          * to seed the RTO for later data packets because SYN packets are
 485          * small. Use the per-dst cached values to seed the RTO but keep
 486          * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
 487          * Later the RTO will be updated immediately upon obtaining the first
 488          * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
 489          * influences the first RTO but not later RTT estimation.
 490          *
 491          * But if RTT is not available from the SYN (due to retransmits or
 492          * syn cookies) or the cache, force a conservative 3secs timeout.
 493          *
 494          * A bit of theory. RTT is time passed after "normal" sized packet
 495          * is sent until it is ACKed. In normal circumstances sending small
 496          * packets force peer to delay ACKs and calculation is correct too.
 497          * The algorithm is adaptive and, provided we follow specs, it
 498          * NEVER underestimate RTT. BUT! If peer tries to make some clever
 499          * tricks sort of "quick acks" for time long enough to decrease RTT
 500          * to low value, and then abruptly stops to do it and starts to delay
 501          * ACKs, wait for troubles.
 502          */
 503         if (crtt > tp->srtt_us) {
 504                 /* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
 505                 crtt /= 8 * USEC_PER_SEC / HZ;
 506                 inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
 507         } else if (tp->srtt_us == 0) {
 508                 /* RFC6298: 5.7 We've failed to get a valid RTT sample from
 509                  * 3WHS. This is most likely due to retransmission,
 510                  * including spurious one. Reset the RTO back to 3secs
 511                  * from the more aggressive 1sec to avoid more spurious
 512                  * retransmission.
 513                  */
 514                 tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
 515                 tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
 516
 517                 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
 518         }
 519         /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
 520          * retransmitted. In light of RFC6298 more aggressive 1sec
 521          * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
 522          * retransmission has occurred.
 523          */
 524         if (tp->total_retrans > 1)
 525                 tp->snd_cwnd = 1;
 526         else
 527                 tp->snd_cwnd = tcp_init_cwnd(tp, dst);
 528         tp->snd_cwnd_stamp = tcp_jiffies32;
 529 }
 530
 531 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
 532 {
 533         struct tcp_metrics_block *tm;
 534         bool ret;
 535
 536         if (!dst)
 537                 return false;
 538
 539         rcu_read_lock();
 540         tm = __tcp_get_metrics_req(req, dst);
 541         if (tm && tcp_metric_get(tm, TCP_METRIC_RTT))
 542                 ret = true;
 543         else
 544                 ret = false;
 545         rcu_read_unlock();
 546
 547         return ret;
 548 }
 549
 550 static DEFINE_SEQLOCK(fastopen_seqlock);
 551
 552 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
 553                             struct tcp_fastopen_cookie *cookie,
 554                             int *syn_loss, unsigned long *last_syn_loss)
 555 {
 556         struct tcp_metrics_block *tm;
 557
 558         rcu_read_lock();
 559         tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
 560         if (tm) {
 561                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 562                 unsigned int seq;
 563
 564                 do {
 565                         seq = read_seqbegin(&fastopen_seqlock);
 566                         if (tfom->mss)
 567                                 *mss = tfom->mss;
 568                         *cookie = tfom->cookie;
 569                         if (cookie->len <= 0 && tfom->try_exp == 1)
 570                                 cookie->exp = true;
 571                         *syn_loss = tfom->syn_loss;
 572                         *last_syn_loss = *syn_loss ? tfom->last_syn_loss : 0;
 573                 } while (read_seqretry(&fastopen_seqlock, seq));
 574         }
 575         rcu_read_unlock();
 576 }
 577
 578 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
 579                             struct tcp_fastopen_cookie *cookie, bool syn_lost,
 580                             u16 try_exp)
 581 {
 582         struct dst_entry *dst = __sk_dst_get(sk);
 583         struct tcp_metrics_block *tm;
 584
 585         if (!dst)
 586                 return;
 587         rcu_read_lock();
 588         tm = tcp_get_metrics(sk, dst, true);
 589         if (tm) {
 590                 struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
 591
 592                 write_seqlock_bh(&fastopen_seqlock);
 593                 if (mss)
 594                         tfom->mss = mss;
 595                 if (cookie && cookie->len > 0)
 596                         tfom->cookie = *cookie;
 597                 else if (try_exp > tfom->try_exp &&
 598                          tfom->cookie.len <= 0 && !tfom->cookie.exp)
 599                         tfom->try_exp = try_exp;
 600                 if (syn_lost) {
 601                         ++tfom->syn_loss;
 602                         tfom->last_syn_loss = jiffies;
 603                 } else
 604                         tfom->syn_loss = 0;
 605                 write_sequnlock_bh(&fastopen_seqlock);
 606         }
 607         rcu_read_unlock();
 608 }
 609
 610 static struct genl_family tcp_metrics_nl_family;
 611
 612 static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
 613         [TCP_METRICS_ATTR_ADDR_IPV4]    = { .type = NLA_U32, },
 614         [TCP_METRICS_ATTR_ADDR_IPV6]    = { .type = NLA_BINARY,
 615                                             .len = sizeof(struct in6_addr), },
 616         /* Following attributes are not received for GET/DEL,
 617          * we keep them for reference
 618          */
 619 #if 0
 620         [TCP_METRICS_ATTR_AGE]          = { .type = NLA_MSECS, },
 621         [TCP_METRICS_ATTR_TW_TSVAL]     = { .type = NLA_U32, },
 622         [TCP_METRICS_ATTR_TW_TS_STAMP]  = { .type = NLA_S32, },
 623         [TCP_METRICS_ATTR_VALS]         = { .type = NLA_NESTED, },
 624         [TCP_METRICS_ATTR_FOPEN_MSS]    = { .type = NLA_U16, },
 625         [TCP_METRICS_ATTR_FOPEN_SYN_DROPS]      = { .type = NLA_U16, },
 626         [TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]    = { .type = NLA_MSECS, },
 627         [TCP_METRICS_ATTR_FOPEN_COOKIE] = { .type = NLA_BINARY,
 628                                             .len = TCP_FASTOPEN_COOKIE_MAX, },
 629 #endif
 630 };
 631
 632 /* Add attributes, caller cancels its header on failure */
 633 static int tcp_metrics_fill_info(struct sk_buff *msg,
 634                                  struct tcp_metrics_block *tm)
 635 {
 636         struct nlattr *nest;
 637         int i;
 638
 639         switch (tm->tcpm_daddr.family) {
 640         case AF_INET:
 641                 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
 642                                     inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
 643                         goto nla_put_failure;
 644                 if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
 645                                     inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
 646                         goto nla_put_failure;
 647                 break;
 648         case AF_INET6:
 649                 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
 650                                      inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
 651                         goto nla_put_failure;
 652                 if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
 653                                      inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
 654                         goto nla_put_failure;
 655                 break;
 656         default:
 657                 return -EAFNOSUPPORT;
 658         }
 659
 660         if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
 661                           jiffies - tm->tcpm_stamp,
 662                           TCP_METRICS_ATTR_PAD) < 0)
 663                 goto nla_put_failure;
 664
 665         {
 666                 int n = 0;
 667
 668                 nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
 669                 if (!nest)
 670                         goto nla_put_failure;
 671                 for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
 672                         u32 val = tm->tcpm_vals[i];
 673
 674                         if (!val)
 675                                 continue;
 676                         if (i == TCP_METRIC_RTT) {
 677                                 if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
 678                                                 val) < 0)
 679                                         goto nla_put_failure;
 680                                 n++;
 681                                 val = max(val / 1000, 1U);
 682                         }
 683                         if (i == TCP_METRIC_RTTVAR) {
 684                                 if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
 685                                                 val) < 0)
 686                                         goto nla_put_failure;
 687                                 n++;
 688                                 val = max(val / 1000, 1U);
 689                         }
 690                         if (nla_put_u32(msg, i + 1, val) < 0)
 691                                 goto nla_put_failure;
 692                         n++;
 693                 }
 694                 if (n)
 695                         nla_nest_end(msg, nest);
 696                 else
 697                         nla_nest_cancel(msg, nest);
 698         }
 699
 700         {
 701                 struct tcp_fastopen_metrics tfom_copy[1], *tfom;
 702                 unsigned int seq;
 703
 704                 do {
 705                         seq = read_seqbegin(&fastopen_seqlock);
 706                         tfom_copy[0] = tm->tcpm_fastopen;
 707                 } while (read_seqretry(&fastopen_seqlock, seq));
 708
 709                 tfom = tfom_copy;
 710                 if (tfom->mss &&
 711                     nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
 712                                 tfom->mss) < 0)
 713                         goto nla_put_failure;
 714                 if (tfom->syn_loss &&
 715                     (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
 716                                 tfom->syn_loss) < 0 ||
 717                      nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
 718                                 jiffies - tfom->last_syn_loss,
 719                                 TCP_METRICS_ATTR_PAD) < 0))
 720                         goto nla_put_failure;
 721                 if (tfom->cookie.len > 0 &&
 722                     nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
 723                             tfom->cookie.len, tfom->cookie.val) < 0)
 724                         goto nla_put_failure;
 725         }
 726
 727         return 0;
 728
 729 nla_put_failure:
 730         return -EMSGSIZE;
 731 }
 732
 733 static int tcp_metrics_dump_info(struct sk_buff *skb,
 734                                  struct netlink_callback *cb,
 735                                  struct tcp_metrics_block *tm)
 736 {
 737         void *hdr;
 738
 739         hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
 740                           &tcp_metrics_nl_family, NLM_F_MULTI,
 741                           TCP_METRICS_CMD_GET);
 742         if (!hdr)
 743                 return -EMSGSIZE;
 744
 745         if (tcp_metrics_fill_info(skb, tm) < 0)
 746                 goto nla_put_failure;
 747
 748         genlmsg_end(skb, hdr);
 749         return 0;
 750
 751 nla_put_failure:
 752         genlmsg_cancel(skb, hdr);
 753         return -EMSGSIZE;
 754 }
 755
 756 static int tcp_metrics_nl_dump(struct sk_buff *skb,
 757                                struct netlink_callback *cb)
 758 {
 759         struct net *net = sock_net(skb->sk);
 760         unsigned int max_rows = 1U << tcp_metrics_hash_log;
 761         unsigned int row, s_row = cb->args[0];
 762         int s_col = cb->args[1], col = s_col;
 763
 764         for (row = s_row; row < max_rows; row++, s_col = 0) {
 765                 struct tcp_metrics_block *tm;
 766                 struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
 767
 768                 rcu_read_lock();
 769                 for (col = 0, tm = rcu_dereference(hb->chain); tm;
 770                      tm = rcu_dereference(tm->tcpm_next), col++) {
 771                         if (!net_eq(tm_net(tm), net))
 772                                 continue;
 773                         if (col < s_col)
 774                                 continue;
 775                         if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
 776                                 rcu_read_unlock();
 777                                 goto done;
 778                         }
 779                 }
 780                 rcu_read_unlock();
 781         }
 782
 783 done:
 784         cb->args[0] = row;
 785         cb->args[1] = col;
 786         return skb->len;
 787 }
 788
 789 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 790                            unsigned int *hash, int optional, int v4, int v6)
 791 {
 792         struct nlattr *a;
 793
 794         a = info->attrs[v4];
 795         if (a) {
 796                 inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
 797                 if (hash)
 798                         *hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
 799                 return 0;
 800         }
 801         a = info->attrs[v6];
 802         if (a) {
 803                 struct in6_addr in6;
 804
 805                 if (nla_len(a) != sizeof(struct in6_addr))
 806                         return -EINVAL;
 807                 in6 = nla_get_in6_addr(a);
 808                 inetpeer_set_addr_v6(addr, &in6);
 809                 if (hash)
 810                         *hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
 811                 return 0;
 812         }
 813         return optional ? 1 : -EAFNOSUPPORT;
 814 }
 815
 816 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
 817                          unsigned int *hash, int optional)
 818 {
 819         return __parse_nl_addr(info, addr, hash, optional,
 820                                TCP_METRICS_ATTR_ADDR_IPV4,
 821                                TCP_METRICS_ATTR_ADDR_IPV6);
 822 }
 823
 824 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
 825 {
 826         return __parse_nl_addr(info, addr, NULL, 0,
 827                                TCP_METRICS_ATTR_SADDR_IPV4,
 828                                TCP_METRICS_ATTR_SADDR_IPV6);
 829 }
 830
 831 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
 832 {
 833         struct tcp_metrics_block *tm;
 834         struct inetpeer_addr saddr, daddr;
 835         unsigned int hash;
 836         struct sk_buff *msg;
 837         struct net *net = genl_info_net(info);
 838         void *reply;
 839         int ret;
 840         bool src = true;
 841
 842         ret = parse_nl_addr(info, &daddr, &hash, 0);
 843         if (ret < 0)
 844                 return ret;
 845
 846         ret = parse_nl_saddr(info, &saddr);
 847         if (ret < 0)
 848                 src = false;
 849
 850         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 851         if (!msg)
 852                 return -ENOMEM;
 853
 854         reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
 855                                   info->genlhdr->cmd);
 856         if (!reply)
 857                 goto nla_put_failure;
 858
 859         hash ^= net_hash_mix(net);
 860         hash = hash_32(hash, tcp_metrics_hash_log);
 861         ret = -ESRCH;
 862         rcu_read_lock();
 863         for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
 864              tm = rcu_dereference(tm->tcpm_next)) {
 865                 if (addr_same(&tm->tcpm_daddr, &daddr) &&
 866                     (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
 867                     net_eq(tm_net(tm), net)) {
 868                         ret = tcp_metrics_fill_info(msg, tm);
 869                         break;
 870                 }
 871         }
 872         rcu_read_unlock();
 873         if (ret < 0)
 874                 goto out_free;
 875
 876         genlmsg_end(msg, reply);
 877         return genlmsg_reply(msg, info);
 878
 879 nla_put_failure:
 880         ret = -EMSGSIZE;
 881
 882 out_free:
 883         nlmsg_free(msg);
 884         return ret;
 885 }
 886
 887 static void tcp_metrics_flush_all(struct net *net)
 888 {
 889         unsigned int max_rows = 1U << tcp_metrics_hash_log;
 890         struct tcpm_hash_bucket *hb = tcp_metrics_hash;
 891         struct tcp_metrics_block *tm;
 892         unsigned int row;
 893
 894         for (row = 0; row < max_rows; row++, hb++) {
 895                 struct tcp_metrics_block __rcu **pp;
 896                 spin_lock_bh(&tcp_metrics_lock);
 897                 pp = &hb->chain;
 898                 for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
 899                         if (net_eq(tm_net(tm), net)) {
 900                                 *pp = tm->tcpm_next;
 901                                 kfree_rcu(tm, rcu_head);
 902                         } else {
 903                                 pp = &tm->tcpm_next;
 904                         }
 905                 }
 906                 spin_unlock_bh(&tcp_metrics_lock);
 907         }
 908 }
 909
 910 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
 911 {
 912         struct tcpm_hash_bucket *hb;
 913         struct tcp_metrics_block *tm;
 914         struct tcp_metrics_block __rcu **pp;
 915         struct inetpeer_addr saddr, daddr;
 916         unsigned int hash;
 917         struct net *net = genl_info_net(info);
 918         int ret;
 919         bool src = true, found = false;
 920
 921         ret = parse_nl_addr(info, &daddr, &hash, 1);
 922         if (ret < 0)
 923                 return ret;
 924         if (ret > 0) {
 925                 tcp_metrics_flush_all(net);
 926                 return 0;
 927         }
 928         ret = parse_nl_saddr(info, &saddr);
 929         if (ret < 0)
 930                 src = false;
 931
 932         hash ^= net_hash_mix(net);
 933         hash = hash_32(hash, tcp_metrics_hash_log);
 934         hb = tcp_metrics_hash + hash;
 935         pp = &hb->chain;
 936         spin_lock_bh(&tcp_metrics_lock);
 937         for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
 938                 if (addr_same(&tm->tcpm_daddr, &daddr) &&
 939                     (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
 940                     net_eq(tm_net(tm), net)) {
 941                         *pp = tm->tcpm_next;
 942                         kfree_rcu(tm, rcu_head);
 943                         found = true;
 944                 } else {
 945                         pp = &tm->tcpm_next;
 946                 }
 947         }
 948         spin_unlock_bh(&tcp_metrics_lock);
 949         if (!found)
 950                 return -ESRCH;
 951         return 0;
 952 }
 953
 954 static const struct genl_ops tcp_metrics_nl_ops[] = {
 955         {
 956                 .cmd = TCP_METRICS_CMD_GET,
 957                 .doit = tcp_metrics_nl_cmd_get,
 958                 .dumpit = tcp_metrics_nl_dump,
 959                 .policy = tcp_metrics_nl_policy,
 960         },
 961         {
 962                 .cmd = TCP_METRICS_CMD_DEL,
 963                 .doit = tcp_metrics_nl_cmd_del,
 964                 .policy = tcp_metrics_nl_policy,
 965                 .flags = GENL_ADMIN_PERM,
 966         },
 967 };
 968
 969 static struct genl_family tcp_metrics_nl_family __ro_after_init = {
 970         .hdrsize        = 0,
 971         .name           = TCP_METRICS_GENL_NAME,
 972         .version        = TCP_METRICS_GENL_VERSION,
 973         .maxattr        = TCP_METRICS_ATTR_MAX,
 974         .netnsok        = true,
 975         .module         = THIS_MODULE,
 976         .ops            = tcp_metrics_nl_ops,
 977         .n_ops          = ARRAY_SIZE(tcp_metrics_nl_ops),
 978 };
 979
 980 static unsigned int tcpmhash_entries;
 981 static int __init set_tcpmhash_entries(char *str)
 982 {
 983         ssize_t ret;
 984
 985         if (!str)
 986                 return 0;
 987
 988         ret = kstrtouint(str, 0, &tcpmhash_entries);
 989         if (ret)
 990                 return 0;
 991
 992         return 1;
 993 }
 994 __setup("tcpmhash_entries=", set_tcpmhash_entries);
 995
 996 static int __net_init tcp_net_metrics_init(struct net *net)
 997 {
 998         size_t size;
 999         unsigned int slots;
1000
1001         if (!net_eq(net, &init_net))
1002                 return 0;
1003
1004         slots = tcpmhash_entries;
1005         if (!slots) {
1006                 if (totalram_pages >= 128 * 1024)
1007                         slots = 16 * 1024;
1008                 else
1009                         slots = 8 * 1024;
1010         }
1011
1012         tcp_metrics_hash_log = order_base_2(slots);
1013         size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1014
1015         tcp_metrics_hash = kvzalloc(size, GFP_KERNEL);
1016         if (!tcp_metrics_hash)
1017                 return -ENOMEM;
1018
1019         return 0;
1020 }
1021
1022 static void __net_exit tcp_net_metrics_exit(struct net *net)
1023 {
1024         tcp_metrics_flush_all(net);
1025 }
1026
1027 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1028         .init   =       tcp_net_metrics_init,
1029         .exit   =       tcp_net_metrics_exit,
1030 };
1031
1032 void __init tcp_metrics_init(void)
1033 {
1034         int ret;
1035
1036         ret = register_pernet_subsys(&tcp_net_metrics_ops);
1037         if (ret < 0)
1038                 panic("Could not allocate the tcp_metrics hash table\n");
1039
1040         ret = genl_register_family(&tcp_metrics_nl_family);
1041         if (ret < 0)
1042                 panic("Could not register tcp_metrics generic netlink\n");
1043 }