sys/netinet/tcp_hostcache.c

   1 /*-
   2  * Copyright (c) 2002 Andre Oppermann, Internet Business Solutions AG
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  * 3. The name of the author may not be used to endorse or promote
  14  *    products derived from this software without specific prior written
  15  *    permission.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  */
  29
  30 /*
  31  * The tcp_hostcache moves the tcp-specific cached metrics from the routing
  32  * table to a dedicated structure indexed by the remote IP address.  It keeps
  33  * information on the measured TCP parameters of past TCP sessions to allow
  34  * better initial start values to be used with later connections to/from the
  35  * same source.  Depending on the network parameters (delay, bandwidth, max
  36  * MTU, congestion window) between local and remote sites, this can lead to
  37  * significant speed-ups for new TCP connections after the first one.
  38  *
  39  * Due to the tcp_hostcache, all TCP-specific metrics information in the
  40  * routing table have been removed.  The inpcb no longer keeps a pointer to
  41  * the routing entry, and protocol-initiated route cloning has been removed
  42  * as well.  With these changes, the routing table has gone back to being
  43  * more lightwight and only carries information related to packet forwarding.
  44  *
  45  * tcp_hostcache is designed for multiple concurrent access in SMP
  46  * environments and high contention.  All bucket rows have their own lock and
  47  * thus multiple lookups and modifies can be done at the same time as long as
  48  * they are in different bucket rows.  If a request for insertion of a new
  49  * record can't be satisfied, it simply returns an empty structure.  Nobody
  50  * and nothing outside of tcp_hostcache.c will ever point directly to any
  51  * entry in the tcp_hostcache.  All communication is done in an
  52  * object-oriented way and only functions of tcp_hostcache will manipulate
  53  * hostcache entries.  Otherwise, we are unable to achieve good behaviour in
  54  * concurrent access situations.  Since tcp_hostcache is only caching
  55  * information, there are no fatal consequences if we either can't satisfy
  56  * any particular request or have to drop/overwrite an existing entry because
  57  * of bucket limit memory constrains.
  58  */
  59
  60 /*
  61  * Many thanks to jlemon for basic structure of tcp_syncache which is being
  62  * followed here.
  63  */
  64
  65 #include <sys/cdefs.h>
  66 __FBSDID("$FreeBSD$");
  67
  68 #include "opt_inet6.h"
  69
  70 #include <sys/param.h>
  71 #include <sys/systm.h>
  72 #include <sys/kernel.h>
  73 #include <sys/lock.h>
  74 #include <sys/mutex.h>
  75 #include <sys/malloc.h>
  76 #include <sys/sbuf.h>
  77 #include <sys/socket.h>
  78 #include <sys/socketvar.h>
  79 #include <sys/sysctl.h>
  80
  81 #include <net/if.h>
  82 #include <net/if_var.h>
  83 #include <net/route.h>
  84 #include <net/vnet.h>
  85
  86 #include <netinet/in.h>
  87 #include <netinet/in_systm.h>
  88 #include <netinet/ip.h>
  89 #include <netinet/in_var.h>
  90 #include <netinet/in_pcb.h>
  91 #include <netinet/ip_var.h>
  92 #ifdef INET6
  93 #include <netinet/ip6.h>
  94 #include <netinet6/ip6_var.h>
  95 #endif
  96 #include <netinet/tcp.h>
  97 #include <netinet/tcp_var.h>
  98 #include <netinet/tcp_hostcache.h>
  99 #ifdef INET6
 100 #include <netinet6/tcp6_var.h>
 101 #endif
 102
 103 #include <vm/uma.h>
 104
 105 /* Arbitrary values */
 106 #define TCP_HOSTCACHE_HASHSIZE          512
 107 #define TCP_HOSTCACHE_BUCKETLIMIT       30
 108 #define TCP_HOSTCACHE_EXPIRE            60*60   /* one hour */
 109 #define TCP_HOSTCACHE_PRUNE             5*60    /* every 5 minutes */
 110
 111 static VNET_DEFINE(struct tcp_hostcache, tcp_hostcache);
 112 #define V_tcp_hostcache         VNET(tcp_hostcache)
 113
 114 static VNET_DEFINE(struct callout, tcp_hc_callout);
 115 #define V_tcp_hc_callout        VNET(tcp_hc_callout)
 116
 117 static struct hc_metrics *tcp_hc_lookup(struct in_conninfo *);
 118 static struct hc_metrics *tcp_hc_insert(struct in_conninfo *);
 119 static int sysctl_tcp_hc_list(SYSCTL_HANDLER_ARGS);
 120 static void tcp_hc_purge_internal(int);
 121 static void tcp_hc_purge(void *);
 122
 123 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hostcache, CTLFLAG_RW, 0,
 124     "TCP Host cache");
 125
 126 SYSCTL_UINT(_net_inet_tcp_hostcache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
 127     &VNET_NAME(tcp_hostcache.cache_limit), 0,
 128     "Overall entry limit for hostcache");
 129
 130 SYSCTL_UINT(_net_inet_tcp_hostcache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
 131     &VNET_NAME(tcp_hostcache.hashsize), 0,
 132     "Size of TCP hostcache hashtable");
 133
 134 SYSCTL_UINT(_net_inet_tcp_hostcache, OID_AUTO, bucketlimit,
 135     CTLFLAG_VNET | CTLFLAG_RDTUN, &VNET_NAME(tcp_hostcache.bucket_limit), 0,
 136     "Per-bucket hash limit for hostcache");
 137
 138 SYSCTL_UINT(_net_inet_tcp_hostcache, OID_AUTO, count, CTLFLAG_VNET | CTLFLAG_RD,
 139      &VNET_NAME(tcp_hostcache.cache_count), 0,
 140     "Current number of entries in hostcache");
 141
 142 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, expire, CTLFLAG_VNET | CTLFLAG_RW,
 143     &VNET_NAME(tcp_hostcache.expire), 0,
 144     "Expire time of TCP hostcache entries");
 145
 146 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, prune, CTLFLAG_VNET | CTLFLAG_RW,
 147     &VNET_NAME(tcp_hostcache.prune), 0,
 148     "Time between purge runs");
 149
 150 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, purge, CTLFLAG_VNET | CTLFLAG_RW,
 151     &VNET_NAME(tcp_hostcache.purgeall), 0,
 152     "Expire all entires on next purge run");
 153
 154 SYSCTL_PROC(_net_inet_tcp_hostcache, OID_AUTO, list,
 155     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_SKIP, 0, 0,
 156     sysctl_tcp_hc_list, "A", "List of all hostcache entries");
 157
 158
 159 static MALLOC_DEFINE(M_HOSTCACHE, "hostcache", "TCP hostcache");
 160
 161 #define HOSTCACHE_HASH(ip) \
 162         (((ip)->s_addr ^ ((ip)->s_addr >> 7) ^ ((ip)->s_addr >> 17)) &  \
 163           V_tcp_hostcache.hashmask)
 164
 165 /* XXX: What is the recommended hash to get good entropy for IPv6 addresses? */
 166 #define HOSTCACHE_HASH6(ip6)                            \
 167         (((ip6)->s6_addr32[0] ^                         \
 168           (ip6)->s6_addr32[1] ^                         \
 169           (ip6)->s6_addr32[2] ^                         \
 170           (ip6)->s6_addr32[3]) &                        \
 171          V_tcp_hostcache.hashmask)
 172
 173 #define THC_LOCK(lp)            mtx_lock(lp)
 174 #define THC_UNLOCK(lp)          mtx_unlock(lp)
 175
 176 void
 177 tcp_hc_init(void)
 178 {
 179         u_int cache_limit;
 180         int i;
 181
 182         /*
 183          * Initialize hostcache structures.
 184          */
 185         V_tcp_hostcache.cache_count = 0;
 186         V_tcp_hostcache.hashsize = TCP_HOSTCACHE_HASHSIZE;
 187         V_tcp_hostcache.bucket_limit = TCP_HOSTCACHE_BUCKETLIMIT;
 188         V_tcp_hostcache.expire = TCP_HOSTCACHE_EXPIRE;
 189         V_tcp_hostcache.prune = TCP_HOSTCACHE_PRUNE;
 190
 191         TUNABLE_INT_FETCH("net.inet.tcp.hostcache.hashsize",
 192             &V_tcp_hostcache.hashsize);
 193         if (!powerof2(V_tcp_hostcache.hashsize)) {
 194                 printf("WARNING: hostcache hash size is not a power of 2.\n");
 195                 V_tcp_hostcache.hashsize = TCP_HOSTCACHE_HASHSIZE; /* default */
 196         }
 197         V_tcp_hostcache.hashmask = V_tcp_hostcache.hashsize - 1;
 198
 199         TUNABLE_INT_FETCH("net.inet.tcp.hostcache.bucketlimit",
 200             &V_tcp_hostcache.bucket_limit);
 201
 202         cache_limit = V_tcp_hostcache.hashsize * V_tcp_hostcache.bucket_limit;
 203         V_tcp_hostcache.cache_limit = cache_limit;
 204         TUNABLE_INT_FETCH("net.inet.tcp.hostcache.cachelimit",
 205             &V_tcp_hostcache.cache_limit);
 206         if (V_tcp_hostcache.cache_limit > cache_limit)
 207                 V_tcp_hostcache.cache_limit = cache_limit;
 208
 209         /*
 210          * Allocate the hash table.
 211          */
 212         V_tcp_hostcache.hashbase = (struct hc_head *)
 213             malloc(V_tcp_hostcache.hashsize * sizeof(struct hc_head),
 214                    M_HOSTCACHE, M_WAITOK | M_ZERO);
 215
 216         /*
 217          * Initialize the hash buckets.
 218          */
 219         for (i = 0; i < V_tcp_hostcache.hashsize; i++) {
 220                 TAILQ_INIT(&V_tcp_hostcache.hashbase[i].hch_bucket);
 221                 V_tcp_hostcache.hashbase[i].hch_length = 0;
 222                 mtx_init(&V_tcp_hostcache.hashbase[i].hch_mtx, "tcp_hc_entry",
 223                           NULL, MTX_DEF);
 224         }
 225
 226         /*
 227          * Allocate the hostcache entries.
 228          */
 229         V_tcp_hostcache.zone =
 230             uma_zcreate("hostcache", sizeof(struct hc_metrics),
 231             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
 232         uma_zone_set_max(V_tcp_hostcache.zone, V_tcp_hostcache.cache_limit);
 233
 234         /*
 235          * Set up periodic cache cleanup.
 236          */
 237         callout_init(&V_tcp_hc_callout, CALLOUT_MPSAFE);
 238         callout_reset(&V_tcp_hc_callout, V_tcp_hostcache.prune * hz,
 239             tcp_hc_purge, curvnet);
 240 }
 241
 242 #ifdef VIMAGE
 243 void
 244 tcp_hc_destroy(void)
 245 {
 246         int i;
 247
 248         callout_drain(&V_tcp_hc_callout);
 249
 250         /* Purge all hc entries. */
 251         tcp_hc_purge_internal(1);
 252
 253         /* Free the uma zone and the allocated hash table. */
 254         uma_zdestroy(V_tcp_hostcache.zone);
 255
 256         for (i = 0; i < V_tcp_hostcache.hashsize; i++)
 257                 mtx_destroy(&V_tcp_hostcache.hashbase[i].hch_mtx);
 258         free(V_tcp_hostcache.hashbase, M_HOSTCACHE);
 259 }
 260 #endif
 261
 262 /*
 263  * Internal function: look up an entry in the hostcache or return NULL.
 264  *
 265  * If an entry has been returned, the caller becomes responsible for
 266  * unlocking the bucket row after he is done reading/modifying the entry.
 267  */
 268 static struct hc_metrics *
 269 tcp_hc_lookup(struct in_conninfo *inc)
 270 {
 271         int hash;
 272         struct hc_head *hc_head;
 273         struct hc_metrics *hc_entry;
 274
 275         KASSERT(inc != NULL, ("tcp_hc_lookup with NULL in_conninfo pointer"));
 276
 277         /*
 278          * Hash the foreign ip address.
 279          */
 280         if (inc->inc_flags & INC_ISIPV6)
 281                 hash = HOSTCACHE_HASH6(&inc->inc6_faddr);
 282         else
 283                 hash = HOSTCACHE_HASH(&inc->inc_faddr);
 284
 285         hc_head = &V_tcp_hostcache.hashbase[hash];
 286
 287         /*
 288          * Acquire lock for this bucket row; we release the lock if we don't
 289          * find an entry, otherwise the caller has to unlock after he is
 290          * done.
 291          */
 292         THC_LOCK(&hc_head->hch_mtx);
 293
 294         /*
 295          * Iterate through entries in bucket row looking for a match.
 296          */
 297         TAILQ_FOREACH(hc_entry, &hc_head->hch_bucket, rmx_q) {
 298                 if (inc->inc_flags & INC_ISIPV6) {
 299                         /* XXX: check ip6_zoneid */
 300                         if (memcmp(&inc->inc6_faddr, &hc_entry->ip6,
 301                             sizeof(inc->inc6_faddr)) == 0)
 302                                 return hc_entry;
 303                 } else {
 304                         if (memcmp(&inc->inc_faddr, &hc_entry->ip4,
 305                             sizeof(inc->inc_faddr)) == 0)
 306                                 return hc_entry;
 307                 }
 308         }
 309
 310         /*
 311          * We were unsuccessful and didn't find anything.
 312          */
 313         THC_UNLOCK(&hc_head->hch_mtx);
 314         return NULL;
 315 }
 316
 317 /*
 318  * Internal function: insert an entry into the hostcache or return NULL if
 319  * unable to allocate a new one.
 320  *
 321  * If an entry has been returned, the caller becomes responsible for
 322  * unlocking the bucket row after he is done reading/modifying the entry.
 323  */
 324 static struct hc_metrics *
 325 tcp_hc_insert(struct in_conninfo *inc)
 326 {
 327         int hash;
 328         struct hc_head *hc_head;
 329         struct hc_metrics *hc_entry;
 330
 331         KASSERT(inc != NULL, ("tcp_hc_insert with NULL in_conninfo pointer"));
 332
 333         /*
 334          * Hash the foreign ip address.
 335          */
 336         if (inc->inc_flags & INC_ISIPV6)
 337                 hash = HOSTCACHE_HASH6(&inc->inc6_faddr);
 338         else
 339                 hash = HOSTCACHE_HASH(&inc->inc_faddr);
 340
 341         hc_head = &V_tcp_hostcache.hashbase[hash];
 342
 343         /*
 344          * Acquire lock for this bucket row; we release the lock if we don't
 345          * find an entry, otherwise the caller has to unlock after he is
 346          * done.
 347          */
 348         THC_LOCK(&hc_head->hch_mtx);
 349
 350         /*
 351          * If the bucket limit is reached, reuse the least-used element.
 352          */
 353         if (hc_head->hch_length >= V_tcp_hostcache.bucket_limit ||
 354             V_tcp_hostcache.cache_count >= V_tcp_hostcache.cache_limit) {
 355                 hc_entry = TAILQ_LAST(&hc_head->hch_bucket, hc_qhead);
 356                 /*
 357                  * At first we were dropping the last element, just to
 358                  * reacquire it in the next two lines again, which isn't very
 359                  * efficient.  Instead just reuse the least used element.
 360                  * We may drop something that is still "in-use" but we can be
 361                  * "lossy".
 362                  * Just give up if this bucket row is empty and we don't have
 363                  * anything to replace.
 364                  */
 365                 if (hc_entry == NULL) {
 366                         THC_UNLOCK(&hc_head->hch_mtx);
 367                         return NULL;
 368                 }
 369                 TAILQ_REMOVE(&hc_head->hch_bucket, hc_entry, rmx_q);
 370                 V_tcp_hostcache.hashbase[hash].hch_length--;
 371                 V_tcp_hostcache.cache_count--;
 372                 TCPSTAT_INC(tcps_hc_bucketoverflow);
 373 #if 0
 374                 uma_zfree(V_tcp_hostcache.zone, hc_entry);
 375 #endif
 376         } else {
 377                 /*
 378                  * Allocate a new entry, or balk if not possible.
 379                  */
 380                 hc_entry = uma_zalloc(V_tcp_hostcache.zone, M_NOWAIT);
 381                 if (hc_entry == NULL) {
 382                         THC_UNLOCK(&hc_head->hch_mtx);
 383                         return NULL;
 384                 }
 385         }
 386
 387         /*
 388          * Initialize basic information of hostcache entry.
 389          */
 390         bzero(hc_entry, sizeof(*hc_entry));
 391         if (inc->inc_flags & INC_ISIPV6) {
 392                 hc_entry->ip6 = inc->inc6_faddr;
 393                 hc_entry->ip6_zoneid = inc->inc6_zoneid;
 394         } else
 395                 hc_entry->ip4 = inc->inc_faddr;
 396         hc_entry->rmx_head = hc_head;
 397         hc_entry->rmx_expire = V_tcp_hostcache.expire;
 398
 399         /*
 400          * Put it upfront.
 401          */
 402         TAILQ_INSERT_HEAD(&hc_head->hch_bucket, hc_entry, rmx_q);
 403         V_tcp_hostcache.hashbase[hash].hch_length++;
 404         V_tcp_hostcache.cache_count++;
 405         TCPSTAT_INC(tcps_hc_added);
 406
 407         return hc_entry;
 408 }
 409
 410 /*
 411  * External function: look up an entry in the hostcache and fill out the
 412  * supplied TCP metrics structure.  Fills in NULL when no entry was found or
 413  * a value is not set.
 414  */
 415 void
 416 tcp_hc_get(struct in_conninfo *inc, struct hc_metrics_lite *hc_metrics_lite)
 417 {
 418         struct hc_metrics *hc_entry;
 419
 420         /*
 421          * Find the right bucket.
 422          */
 423         hc_entry = tcp_hc_lookup(inc);
 424
 425         /*
 426          * If we don't have an existing object.
 427          */
 428         if (hc_entry == NULL) {
 429                 bzero(hc_metrics_lite, sizeof(*hc_metrics_lite));
 430                 return;
 431         }
 432         hc_entry->rmx_hits++;
 433         hc_entry->rmx_expire = V_tcp_hostcache.expire; /* start over again */
 434
 435         hc_metrics_lite->rmx_mtu = hc_entry->rmx_mtu;
 436         hc_metrics_lite->rmx_ssthresh = hc_entry->rmx_ssthresh;
 437         hc_metrics_lite->rmx_rtt = hc_entry->rmx_rtt;
 438         hc_metrics_lite->rmx_rttvar = hc_entry->rmx_rttvar;
 439         hc_metrics_lite->rmx_bandwidth = hc_entry->rmx_bandwidth;
 440         hc_metrics_lite->rmx_cwnd = hc_entry->rmx_cwnd;
 441         hc_metrics_lite->rmx_sendpipe = hc_entry->rmx_sendpipe;
 442         hc_metrics_lite->rmx_recvpipe = hc_entry->rmx_recvpipe;
 443
 444         /*
 445          * Unlock bucket row.
 446          */
 447         THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
 448 }
 449
 450 /*
 451  * External function: look up an entry in the hostcache and return the
 452  * discovered path MTU.  Returns NULL if no entry is found or value is not
 453  * set.
 454  */
 455 u_long
 456 tcp_hc_getmtu(struct in_conninfo *inc)
 457 {
 458         struct hc_metrics *hc_entry;
 459         u_long mtu;
 460
 461         hc_entry = tcp_hc_lookup(inc);
 462         if (hc_entry == NULL) {
 463                 return 0;
 464         }
 465         hc_entry->rmx_hits++;
 466         hc_entry->rmx_expire = V_tcp_hostcache.expire; /* start over again */
 467
 468         mtu = hc_entry->rmx_mtu;
 469         THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
 470         return mtu;
 471 }
 472
 473 /*
 474  * External function: update the MTU value of an entry in the hostcache.
 475  * Creates a new entry if none was found.
 476  */
 477 void
 478 tcp_hc_updatemtu(struct in_conninfo *inc, u_long mtu)
 479 {
 480         struct hc_metrics *hc_entry;
 481
 482         /*
 483          * Find the right bucket.
 484          */
 485         hc_entry = tcp_hc_lookup(inc);
 486
 487         /*
 488          * If we don't have an existing object, try to insert a new one.
 489          */
 490         if (hc_entry == NULL) {
 491                 hc_entry = tcp_hc_insert(inc);
 492                 if (hc_entry == NULL)
 493                         return;
 494         }
 495         hc_entry->rmx_updates++;
 496         hc_entry->rmx_expire = V_tcp_hostcache.expire; /* start over again */
 497
 498         hc_entry->rmx_mtu = mtu;
 499
 500         /*
 501          * Put it upfront so we find it faster next time.
 502          */
 503         TAILQ_REMOVE(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
 504         TAILQ_INSERT_HEAD(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
 505
 506         /*
 507          * Unlock bucket row.
 508          */
 509         THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
 510 }
 511
 512 /*
 513  * External function: update the TCP metrics of an entry in the hostcache.
 514  * Creates a new entry if none was found.
 515  */
 516 void
 517 tcp_hc_update(struct in_conninfo *inc, struct hc_metrics_lite *hcml)
 518 {
 519         struct hc_metrics *hc_entry;
 520
 521         hc_entry = tcp_hc_lookup(inc);
 522         if (hc_entry == NULL) {
 523                 hc_entry = tcp_hc_insert(inc);
 524                 if (hc_entry == NULL)
 525                         return;
 526         }
 527         hc_entry->rmx_updates++;
 528         hc_entry->rmx_expire = V_tcp_hostcache.expire; /* start over again */
 529
 530         if (hcml->rmx_rtt != 0) {
 531                 if (hc_entry->rmx_rtt == 0)
 532                         hc_entry->rmx_rtt = hcml->rmx_rtt;
 533                 else
 534                         hc_entry->rmx_rtt =
 535                             (hc_entry->rmx_rtt + hcml->rmx_rtt) / 2;
 536                 TCPSTAT_INC(tcps_cachedrtt);
 537         }
 538         if (hcml->rmx_rttvar != 0) {
 539                 if (hc_entry->rmx_rttvar == 0)
 540                         hc_entry->rmx_rttvar = hcml->rmx_rttvar;
 541                 else
 542                         hc_entry->rmx_rttvar =
 543                             (hc_entry->rmx_rttvar + hcml->rmx_rttvar) / 2;
 544                 TCPSTAT_INC(tcps_cachedrttvar);
 545         }
 546         if (hcml->rmx_ssthresh != 0) {
 547                 if (hc_entry->rmx_ssthresh == 0)
 548                         hc_entry->rmx_ssthresh = hcml->rmx_ssthresh;
 549                 else
 550                         hc_entry->rmx_ssthresh =
 551                             (hc_entry->rmx_ssthresh + hcml->rmx_ssthresh) / 2;
 552                 TCPSTAT_INC(tcps_cachedssthresh);
 553         }
 554         if (hcml->rmx_bandwidth != 0) {
 555                 if (hc_entry->rmx_bandwidth == 0)
 556                         hc_entry->rmx_bandwidth = hcml->rmx_bandwidth;
 557                 else
 558                         hc_entry->rmx_bandwidth =
 559                             (hc_entry->rmx_bandwidth + hcml->rmx_bandwidth) / 2;
 560                 /* TCPSTAT_INC(tcps_cachedbandwidth); */
 561         }
 562         if (hcml->rmx_cwnd != 0) {
 563                 if (hc_entry->rmx_cwnd == 0)
 564                         hc_entry->rmx_cwnd = hcml->rmx_cwnd;
 565                 else
 566                         hc_entry->rmx_cwnd =
 567                             (hc_entry->rmx_cwnd + hcml->rmx_cwnd) / 2;
 568                 /* TCPSTAT_INC(tcps_cachedcwnd); */
 569         }
 570         if (hcml->rmx_sendpipe != 0) {
 571                 if (hc_entry->rmx_sendpipe == 0)
 572                         hc_entry->rmx_sendpipe = hcml->rmx_sendpipe;
 573                 else
 574                         hc_entry->rmx_sendpipe =
 575                             (hc_entry->rmx_sendpipe + hcml->rmx_sendpipe) /2;
 576                 /* TCPSTAT_INC(tcps_cachedsendpipe); */
 577         }
 578         if (hcml->rmx_recvpipe != 0) {
 579                 if (hc_entry->rmx_recvpipe == 0)
 580                         hc_entry->rmx_recvpipe = hcml->rmx_recvpipe;
 581                 else
 582                         hc_entry->rmx_recvpipe =
 583                             (hc_entry->rmx_recvpipe + hcml->rmx_recvpipe) /2;
 584                 /* TCPSTAT_INC(tcps_cachedrecvpipe); */
 585         }
 586
 587         TAILQ_REMOVE(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
 588         TAILQ_INSERT_HEAD(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
 589         THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
 590 }
 591
 592 /*
 593  * Sysctl function: prints the list and values of all hostcache entries in
 594  * unsorted order.
 595  */
 596 static int
 597 sysctl_tcp_hc_list(SYSCTL_HANDLER_ARGS)
 598 {
 599         int linesize = 128;
 600         struct sbuf sb;
 601         int i, error;
 602         struct hc_metrics *hc_entry;
 603 #ifdef INET6
 604         char ip6buf[INET6_ADDRSTRLEN];
 605 #endif
 606
 607         sbuf_new(&sb, NULL, linesize * (V_tcp_hostcache.cache_count + 1),
 608             SBUF_FIXEDLEN);
 609
 610         sbuf_printf(&sb,
 611                 "\nIP address        MTU  SSTRESH      RTT   RTTVAR BANDWIDTH "
 612                 "    CWND SENDPIPE RECVPIPE HITS  UPD  EXP\n");
 613
 614 #define msec(u) (((u) + 500) / 1000)
 615         for (i = 0; i < V_tcp_hostcache.hashsize; i++) {
 616                 THC_LOCK(&V_tcp_hostcache.hashbase[i].hch_mtx);
 617                 TAILQ_FOREACH(hc_entry, &V_tcp_hostcache.hashbase[i].hch_bucket,
 618                               rmx_q) {
 619                         sbuf_printf(&sb,
 620                             "%-15s %5lu %8lu %6lums %6lums %9lu %8lu %8lu %8lu "
 621                             "%4lu %4lu %4i\n",
 622                             hc_entry->ip4.s_addr ? inet_ntoa(hc_entry->ip4) :
 623 #ifdef INET6
 624                                 ip6_sprintf(ip6buf, &hc_entry->ip6),
 625 #else
 626                                 "IPv6?",
 627 #endif
 628                             hc_entry->rmx_mtu,
 629                             hc_entry->rmx_ssthresh,
 630                             msec(hc_entry->rmx_rtt *
 631                                 (RTM_RTTUNIT / (hz * TCP_RTT_SCALE))),
 632                             msec(hc_entry->rmx_rttvar *
 633                                 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE))),
 634                             hc_entry->rmx_bandwidth * 8,
 635                             hc_entry->rmx_cwnd,
 636                             hc_entry->rmx_sendpipe,
 637                             hc_entry->rmx_recvpipe,
 638                             hc_entry->rmx_hits,
 639                             hc_entry->rmx_updates,
 640                             hc_entry->rmx_expire);
 641                 }
 642                 THC_UNLOCK(&V_tcp_hostcache.hashbase[i].hch_mtx);
 643         }
 644 #undef msec
 645         sbuf_finish(&sb);
 646         error = SYSCTL_OUT(req, sbuf_data(&sb), sbuf_len(&sb));
 647         sbuf_delete(&sb);
 648         return(error);
 649 }
 650
 651 /*
 652  * Caller has to make sure the curvnet is set properly.
 653  */
 654 static void
 655 tcp_hc_purge_internal(int all)
 656 {
 657         struct hc_metrics *hc_entry, *hc_next;
 658         int i;
 659
 660         for (i = 0; i < V_tcp_hostcache.hashsize; i++) {
 661                 THC_LOCK(&V_tcp_hostcache.hashbase[i].hch_mtx);
 662                 TAILQ_FOREACH_SAFE(hc_entry,
 663                     &V_tcp_hostcache.hashbase[i].hch_bucket, rmx_q, hc_next) {
 664                         if (all || hc_entry->rmx_expire <= 0) {
 665                                 TAILQ_REMOVE(&V_tcp_hostcache.hashbase[i].hch_bucket,
 666                                               hc_entry, rmx_q);
 667                                 uma_zfree(V_tcp_hostcache.zone, hc_entry);
 668                                 V_tcp_hostcache.hashbase[i].hch_length--;
 669                                 V_tcp_hostcache.cache_count--;
 670                         } else
 671                                 hc_entry->rmx_expire -= V_tcp_hostcache.prune;
 672                 }
 673                 THC_UNLOCK(&V_tcp_hostcache.hashbase[i].hch_mtx);
 674         }
 675 }
 676
 677 /*
 678  * Expire and purge (old|all) entries in the tcp_hostcache.  Runs
 679  * periodically from the callout.
 680  */
 681 static void
 682 tcp_hc_purge(void *arg)
 683 {
 684         CURVNET_SET((struct vnet *) arg);
 685         int all = 0;
 686
 687         if (V_tcp_hostcache.purgeall) {
 688                 all = 1;
 689                 V_tcp_hostcache.purgeall = 0;
 690         }
 691
 692         tcp_hc_purge_internal(all);
 693
 694         callout_reset(&V_tcp_hc_callout, V_tcp_hostcache.prune * hz,
 695             tcp_hc_purge, arg);
 696         CURVNET_RESTORE();
 697 }