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/socket.h>
  77 #include <sys/socketvar.h>
  78 #include <sys/sysctl.h>
  79
  80 #include <net/if.h>
  81 #include <net/if_var.h>
  82 #include <net/route.h>
  83 #include <net/vnet.h>
  84
  85 #include <netinet/in.h>
  86 #include <netinet/in_systm.h>
  87 #include <netinet/ip.h>
  88 #include <netinet/in_var.h>
  89 #include <netinet/in_pcb.h>
  90 #include <netinet/ip_var.h>
  91 #ifdef INET6
  92 #include <netinet/ip6.h>
  93 #include <netinet6/ip6_var.h>
  94 #endif
  95 #include <netinet/tcp.h>
  96 #include <netinet/tcp_var.h>
  97 #include <netinet/tcp_hostcache.h>
  98 #ifdef INET6
  99 #include <netinet6/tcp6_var.h>
 100 #endif
 101
 102 #include <vm/uma.h>
 103
 104 /* Arbitrary values */
 105 #define TCP_HOSTCACHE_HASHSIZE          512
 106 #define TCP_HOSTCACHE_BUCKETLIMIT       30
 107 #define TCP_HOSTCACHE_EXPIRE            60*60   /* one hour */
 108 #define TCP_HOSTCACHE_PRUNE             5*60    /* every 5 minutes */
 109
 110 static VNET_DEFINE(struct tcp_hostcache, tcp_hostcache);
 111 #define V_tcp_hostcache         VNET(tcp_hostcache)
 112
 113 static VNET_DEFINE(struct callout, tcp_hc_callout);
 114 #define V_tcp_hc_callout        VNET(tcp_hc_callout)
 115
 116 static struct hc_metrics *tcp_hc_lookup(struct in_conninfo *);
 117 static struct hc_metrics *tcp_hc_insert(struct in_conninfo *);
 118 static int sysctl_tcp_hc_list(SYSCTL_HANDLER_ARGS);
 119 static void tcp_hc_purge_internal(int);
 120 static void tcp_hc_purge(void *);
 121
 122 static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hostcache, CTLFLAG_RW, 0,
 123     "TCP Host cache");
 124
 125 SYSCTL_UINT(_net_inet_tcp_hostcache, OID_AUTO, cachelimit, CTLFLAG_VNET | CTLFLAG_RDTUN,
 126     &VNET_NAME(tcp_hostcache.cache_limit), 0,
 127     "Overall entry limit for hostcache");
 128
 129 SYSCTL_UINT(_net_inet_tcp_hostcache, OID_AUTO, hashsize, CTLFLAG_VNET | CTLFLAG_RDTUN,
 130     &VNET_NAME(tcp_hostcache.hashsize), 0,
 131     "Size of TCP hostcache hashtable");
 132
 133 SYSCTL_UINT(_net_inet_tcp_hostcache, OID_AUTO, bucketlimit,
 134     CTLFLAG_VNET | CTLFLAG_RDTUN, &VNET_NAME(tcp_hostcache.bucket_limit), 0,
 135     "Per-bucket hash limit for hostcache");
 136
 137 SYSCTL_UINT(_net_inet_tcp_hostcache, OID_AUTO, count, CTLFLAG_VNET | CTLFLAG_RD,
 138      &VNET_NAME(tcp_hostcache.cache_count), 0,
 139     "Current number of entries in hostcache");
 140
 141 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, expire, CTLFLAG_VNET | CTLFLAG_RW,
 142     &VNET_NAME(tcp_hostcache.expire), 0,
 143     "Expire time of TCP hostcache entries");
 144
 145 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, prune, CTLFLAG_VNET | CTLFLAG_RW,
 146     &VNET_NAME(tcp_hostcache.prune), 0,
 147     "Time between purge runs");
 148
 149 SYSCTL_INT(_net_inet_tcp_hostcache, OID_AUTO, purge, CTLFLAG_VNET | CTLFLAG_RW,
 150     &VNET_NAME(tcp_hostcache.purgeall), 0,
 151     "Expire all entires on next purge run");
 152
 153 SYSCTL_PROC(_net_inet_tcp_hostcache, OID_AUTO, list,
 154     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_SKIP, 0, 0,
 155     sysctl_tcp_hc_list, "A", "List of all hostcache entries");
 156
 157
 158 static MALLOC_DEFINE(M_HOSTCACHE, "hostcache", "TCP hostcache");
 159
 160 #define HOSTCACHE_HASH(ip) \
 161         (((ip)->s_addr ^ ((ip)->s_addr >> 7) ^ ((ip)->s_addr >> 17)) &  \
 162           V_tcp_hostcache.hashmask)
 163
 164 /* XXX: What is the recommended hash to get good entropy for IPv6 addresses? */
 165 #define HOSTCACHE_HASH6(ip6)                            \
 166         (((ip6)->s6_addr32[0] ^                         \
 167           (ip6)->s6_addr32[1] ^                         \
 168           (ip6)->s6_addr32[2] ^                         \
 169           (ip6)->s6_addr32[3]) &                        \
 170          V_tcp_hostcache.hashmask)
 171
 172 #define THC_LOCK(lp)            mtx_lock(lp)
 173 #define THC_UNLOCK(lp)          mtx_unlock(lp)
 174
 175 void
 176 tcp_hc_init(void)
 177 {
 178         u_int cache_limit;
 179         int i;
 180
 181         /*
 182          * Initialize hostcache structures.
 183          */
 184         V_tcp_hostcache.cache_count = 0;
 185         V_tcp_hostcache.hashsize = TCP_HOSTCACHE_HASHSIZE;
 186         V_tcp_hostcache.bucket_limit = TCP_HOSTCACHE_BUCKETLIMIT;
 187         V_tcp_hostcache.expire = TCP_HOSTCACHE_EXPIRE;
 188         V_tcp_hostcache.prune = TCP_HOSTCACHE_PRUNE;
 189
 190         TUNABLE_INT_FETCH("net.inet.tcp.hostcache.hashsize",
 191             &V_tcp_hostcache.hashsize);
 192         if (!powerof2(V_tcp_hostcache.hashsize)) {
 193                 printf("WARNING: hostcache hash size is not a power of 2.\n");
 194                 V_tcp_hostcache.hashsize = TCP_HOSTCACHE_HASHSIZE; /* default */
 195         }
 196         V_tcp_hostcache.hashmask = V_tcp_hostcache.hashsize - 1;
 197
 198         TUNABLE_INT_FETCH("net.inet.tcp.hostcache.bucketlimit",
 199             &V_tcp_hostcache.bucket_limit);
 200
 201         cache_limit = V_tcp_hostcache.hashsize * V_tcp_hostcache.bucket_limit;
 202         V_tcp_hostcache.cache_limit = cache_limit;
 203         TUNABLE_INT_FETCH("net.inet.tcp.hostcache.cachelimit",
 204             &V_tcp_hostcache.cache_limit);
 205         if (V_tcp_hostcache.cache_limit > cache_limit)
 206                 V_tcp_hostcache.cache_limit = cache_limit;
 207
 208         /*
 209          * Allocate the hash table.
 210          */
 211         V_tcp_hostcache.hashbase = (struct hc_head *)
 212             malloc(V_tcp_hostcache.hashsize * sizeof(struct hc_head),
 213                    M_HOSTCACHE, M_WAITOK | M_ZERO);
 214
 215         /*
 216          * Initialize the hash buckets.
 217          */
 218         for (i = 0; i < V_tcp_hostcache.hashsize; i++) {
 219                 TAILQ_INIT(&V_tcp_hostcache.hashbase[i].hch_bucket);
 220                 V_tcp_hostcache.hashbase[i].hch_length = 0;
 221                 mtx_init(&V_tcp_hostcache.hashbase[i].hch_mtx, "tcp_hc_entry",
 222                           NULL, MTX_DEF);
 223         }
 224
 225         /*
 226          * Allocate the hostcache entries.
 227          */
 228         V_tcp_hostcache.zone =
 229             uma_zcreate("hostcache", sizeof(struct hc_metrics),
 230             NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
 231         uma_zone_set_max(V_tcp_hostcache.zone, V_tcp_hostcache.cache_limit);
 232
 233         /*
 234          * Set up periodic cache cleanup.
 235          */
 236         callout_init(&V_tcp_hc_callout, CALLOUT_MPSAFE);
 237         callout_reset(&V_tcp_hc_callout, V_tcp_hostcache.prune * hz,
 238             tcp_hc_purge, curvnet);
 239 }
 240
 241 #ifdef VIMAGE
 242 void
 243 tcp_hc_destroy(void)
 244 {
 245         int i;
 246
 247         callout_drain(&V_tcp_hc_callout);
 248
 249         /* Purge all hc entries. */
 250         tcp_hc_purge_internal(1);
 251
 252         /* Free the uma zone and the allocated hash table. */
 253         uma_zdestroy(V_tcp_hostcache.zone);
 254
 255         for (i = 0; i < V_tcp_hostcache.hashsize; i++)
 256                 mtx_destroy(&V_tcp_hostcache.hashbase[i].hch_mtx);
 257         free(V_tcp_hostcache.hashbase, M_HOSTCACHE);
 258 }
 259 #endif
 260
 261 /*
 262  * Internal function: look up an entry in the hostcache or return NULL.
 263  *
 264  * If an entry has been returned, the caller becomes responsible for
 265  * unlocking the bucket row after he is done reading/modifying the entry.
 266  */
 267 static struct hc_metrics *
 268 tcp_hc_lookup(struct in_conninfo *inc)
 269 {
 270         int hash;
 271         struct hc_head *hc_head;
 272         struct hc_metrics *hc_entry;
 273
 274         KASSERT(inc != NULL, ("tcp_hc_lookup with NULL in_conninfo pointer"));
 275
 276         /*
 277          * Hash the foreign ip address.
 278          */
 279         if (inc->inc_flags & INC_ISIPV6)
 280                 hash = HOSTCACHE_HASH6(&inc->inc6_faddr);
 281         else
 282                 hash = HOSTCACHE_HASH(&inc->inc_faddr);
 283
 284         hc_head = &V_tcp_hostcache.hashbase[hash];
 285
 286         /*
 287          * Acquire lock for this bucket row; we release the lock if we don't
 288          * find an entry, otherwise the caller has to unlock after he is
 289          * done.
 290          */
 291         THC_LOCK(&hc_head->hch_mtx);
 292
 293         /*
 294          * Iterate through entries in bucket row looking for a match.
 295          */
 296         TAILQ_FOREACH(hc_entry, &hc_head->hch_bucket, rmx_q) {
 297                 if (inc->inc_flags & INC_ISIPV6) {
 298                         /* XXX: check ip6_zoneid */
 299                         if (memcmp(&inc->inc6_faddr, &hc_entry->ip6,
 300                             sizeof(inc->inc6_faddr)) == 0)
 301                                 return hc_entry;
 302                 } else {
 303                         if (memcmp(&inc->inc_faddr, &hc_entry->ip4,
 304                             sizeof(inc->inc_faddr)) == 0)
 305                                 return hc_entry;
 306                 }
 307         }
 308
 309         /*
 310          * We were unsuccessful and didn't find anything.
 311          */
 312         THC_UNLOCK(&hc_head->hch_mtx);
 313         return NULL;
 314 }
 315
 316 /*
 317  * Internal function: insert an entry into the hostcache or return NULL if
 318  * unable to allocate a new one.
 319  *
 320  * If an entry has been returned, the caller becomes responsible for
 321  * unlocking the bucket row after he is done reading/modifying the entry.
 322  */
 323 static struct hc_metrics *
 324 tcp_hc_insert(struct in_conninfo *inc)
 325 {
 326         int hash;
 327         struct hc_head *hc_head;
 328         struct hc_metrics *hc_entry;
 329
 330         KASSERT(inc != NULL, ("tcp_hc_insert with NULL in_conninfo pointer"));
 331
 332         /*
 333          * Hash the foreign ip address.
 334          */
 335         if (inc->inc_flags & INC_ISIPV6)
 336                 hash = HOSTCACHE_HASH6(&inc->inc6_faddr);
 337         else
 338                 hash = HOSTCACHE_HASH(&inc->inc_faddr);
 339
 340         hc_head = &V_tcp_hostcache.hashbase[hash];
 341
 342         /*
 343          * Acquire lock for this bucket row; we release the lock if we don't
 344          * find an entry, otherwise the caller has to unlock after he is
 345          * done.
 346          */
 347         THC_LOCK(&hc_head->hch_mtx);
 348
 349         /*
 350          * If the bucket limit is reached, reuse the least-used element.
 351          */
 352         if (hc_head->hch_length >= V_tcp_hostcache.bucket_limit ||
 353             V_tcp_hostcache.cache_count >= V_tcp_hostcache.cache_limit) {
 354                 hc_entry = TAILQ_LAST(&hc_head->hch_bucket, hc_qhead);
 355                 /*
 356                  * At first we were dropping the last element, just to
 357                  * reacquire it in the next two lines again, which isn't very
 358                  * efficient.  Instead just reuse the least used element.
 359                  * We may drop something that is still "in-use" but we can be
 360                  * "lossy".
 361                  * Just give up if this bucket row is empty and we don't have
 362                  * anything to replace.
 363                  */
 364                 if (hc_entry == NULL) {
 365                         THC_UNLOCK(&hc_head->hch_mtx);
 366                         return NULL;
 367                 }
 368                 TAILQ_REMOVE(&hc_head->hch_bucket, hc_entry, rmx_q);
 369                 V_tcp_hostcache.hashbase[hash].hch_length--;
 370                 V_tcp_hostcache.cache_count--;
 371                 TCPSTAT_INC(tcps_hc_bucketoverflow);
 372 #if 0
 373                 uma_zfree(V_tcp_hostcache.zone, hc_entry);
 374 #endif
 375         } else {
 376                 /*
 377                  * Allocate a new entry, or balk if not possible.
 378                  */
 379                 hc_entry = uma_zalloc(V_tcp_hostcache.zone, M_NOWAIT);
 380                 if (hc_entry == NULL) {
 381                         THC_UNLOCK(&hc_head->hch_mtx);
 382                         return NULL;
 383                 }
 384         }
 385
 386         /*
 387          * Initialize basic information of hostcache entry.
 388          */
 389         bzero(hc_entry, sizeof(*hc_entry));
 390         if (inc->inc_flags & INC_ISIPV6) {
 391                 hc_entry->ip6 = inc->inc6_faddr;
 392                 hc_entry->ip6_zoneid = inc->inc6_zoneid;
 393         } else
 394                 hc_entry->ip4 = inc->inc_faddr;
 395         hc_entry->rmx_head = hc_head;
 396         hc_entry->rmx_expire = V_tcp_hostcache.expire;
 397
 398         /*
 399          * Put it upfront.
 400          */
 401         TAILQ_INSERT_HEAD(&hc_head->hch_bucket, hc_entry, rmx_q);
 402         V_tcp_hostcache.hashbase[hash].hch_length++;
 403         V_tcp_hostcache.cache_count++;
 404         TCPSTAT_INC(tcps_hc_added);
 405
 406         return hc_entry;
 407 }
 408
 409 /*
 410  * External function: look up an entry in the hostcache and fill out the
 411  * supplied TCP metrics structure.  Fills in NULL when no entry was found or
 412  * a value is not set.
 413  */
 414 void
 415 tcp_hc_get(struct in_conninfo *inc, struct hc_metrics_lite *hc_metrics_lite)
 416 {
 417         struct hc_metrics *hc_entry;
 418
 419         /*
 420          * Find the right bucket.
 421          */
 422         hc_entry = tcp_hc_lookup(inc);
 423
 424         /*
 425          * If we don't have an existing object.
 426          */
 427         if (hc_entry == NULL) {
 428                 bzero(hc_metrics_lite, sizeof(*hc_metrics_lite));
 429                 return;
 430         }
 431         hc_entry->rmx_hits++;
 432         hc_entry->rmx_expire = V_tcp_hostcache.expire; /* start over again */
 433
 434         hc_metrics_lite->rmx_mtu = hc_entry->rmx_mtu;
 435         hc_metrics_lite->rmx_ssthresh = hc_entry->rmx_ssthresh;
 436         hc_metrics_lite->rmx_rtt = hc_entry->rmx_rtt;
 437         hc_metrics_lite->rmx_rttvar = hc_entry->rmx_rttvar;
 438         hc_metrics_lite->rmx_bandwidth = hc_entry->rmx_bandwidth;
 439         hc_metrics_lite->rmx_cwnd = hc_entry->rmx_cwnd;
 440         hc_metrics_lite->rmx_sendpipe = hc_entry->rmx_sendpipe;
 441         hc_metrics_lite->rmx_recvpipe = hc_entry->rmx_recvpipe;
 442
 443         /*
 444          * Unlock bucket row.
 445          */
 446         THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
 447 }
 448
 449 /*
 450  * External function: look up an entry in the hostcache and return the
 451  * discovered path MTU.  Returns NULL if no entry is found or value is not
 452  * set.
 453  */
 454 u_long
 455 tcp_hc_getmtu(struct in_conninfo *inc)
 456 {
 457         struct hc_metrics *hc_entry;
 458         u_long mtu;
 459
 460         hc_entry = tcp_hc_lookup(inc);
 461         if (hc_entry == NULL) {
 462                 return 0;
 463         }
 464         hc_entry->rmx_hits++;
 465         hc_entry->rmx_expire = V_tcp_hostcache.expire; /* start over again */
 466
 467         mtu = hc_entry->rmx_mtu;
 468         THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
 469         return mtu;
 470 }
 471
 472 /*
 473  * External function: update the MTU value of an entry in the hostcache.
 474  * Creates a new entry if none was found.
 475  */
 476 void
 477 tcp_hc_updatemtu(struct in_conninfo *inc, u_long mtu)
 478 {
 479         struct hc_metrics *hc_entry;
 480
 481         /*
 482          * Find the right bucket.
 483          */
 484         hc_entry = tcp_hc_lookup(inc);
 485
 486         /*
 487          * If we don't have an existing object, try to insert a new one.
 488          */
 489         if (hc_entry == NULL) {
 490                 hc_entry = tcp_hc_insert(inc);
 491                 if (hc_entry == NULL)
 492                         return;
 493         }
 494         hc_entry->rmx_updates++;
 495         hc_entry->rmx_expire = V_tcp_hostcache.expire; /* start over again */
 496
 497         hc_entry->rmx_mtu = mtu;
 498
 499         /*
 500          * Put it upfront so we find it faster next time.
 501          */
 502         TAILQ_REMOVE(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
 503         TAILQ_INSERT_HEAD(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
 504
 505         /*
 506          * Unlock bucket row.
 507          */
 508         THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
 509 }
 510
 511 /*
 512  * External function: update the TCP metrics of an entry in the hostcache.
 513  * Creates a new entry if none was found.
 514  */
 515 void
 516 tcp_hc_update(struct in_conninfo *inc, struct hc_metrics_lite *hcml)
 517 {
 518         struct hc_metrics *hc_entry;
 519
 520         hc_entry = tcp_hc_lookup(inc);
 521         if (hc_entry == NULL) {
 522                 hc_entry = tcp_hc_insert(inc);
 523                 if (hc_entry == NULL)
 524                         return;
 525         }
 526         hc_entry->rmx_updates++;
 527         hc_entry->rmx_expire = V_tcp_hostcache.expire; /* start over again */
 528
 529         if (hcml->rmx_rtt != 0) {
 530                 if (hc_entry->rmx_rtt == 0)
 531                         hc_entry->rmx_rtt = hcml->rmx_rtt;
 532                 else
 533                         hc_entry->rmx_rtt =
 534                             (hc_entry->rmx_rtt + hcml->rmx_rtt) / 2;
 535                 TCPSTAT_INC(tcps_cachedrtt);
 536         }
 537         if (hcml->rmx_rttvar != 0) {
 538                 if (hc_entry->rmx_rttvar == 0)
 539                         hc_entry->rmx_rttvar = hcml->rmx_rttvar;
 540                 else
 541                         hc_entry->rmx_rttvar =
 542                             (hc_entry->rmx_rttvar + hcml->rmx_rttvar) / 2;
 543                 TCPSTAT_INC(tcps_cachedrttvar);
 544         }
 545         if (hcml->rmx_ssthresh != 0) {
 546                 if (hc_entry->rmx_ssthresh == 0)
 547                         hc_entry->rmx_ssthresh = hcml->rmx_ssthresh;
 548                 else
 549                         hc_entry->rmx_ssthresh =
 550                             (hc_entry->rmx_ssthresh + hcml->rmx_ssthresh) / 2;
 551                 TCPSTAT_INC(tcps_cachedssthresh);
 552         }
 553         if (hcml->rmx_bandwidth != 0) {
 554                 if (hc_entry->rmx_bandwidth == 0)
 555                         hc_entry->rmx_bandwidth = hcml->rmx_bandwidth;
 556                 else
 557                         hc_entry->rmx_bandwidth =
 558                             (hc_entry->rmx_bandwidth + hcml->rmx_bandwidth) / 2;
 559                 /* TCPSTAT_INC(tcps_cachedbandwidth); */
 560         }
 561         if (hcml->rmx_cwnd != 0) {
 562                 if (hc_entry->rmx_cwnd == 0)
 563                         hc_entry->rmx_cwnd = hcml->rmx_cwnd;
 564                 else
 565                         hc_entry->rmx_cwnd =
 566                             (hc_entry->rmx_cwnd + hcml->rmx_cwnd) / 2;
 567                 /* TCPSTAT_INC(tcps_cachedcwnd); */
 568         }
 569         if (hcml->rmx_sendpipe != 0) {
 570                 if (hc_entry->rmx_sendpipe == 0)
 571                         hc_entry->rmx_sendpipe = hcml->rmx_sendpipe;
 572                 else
 573                         hc_entry->rmx_sendpipe =
 574                             (hc_entry->rmx_sendpipe + hcml->rmx_sendpipe) /2;
 575                 /* TCPSTAT_INC(tcps_cachedsendpipe); */
 576         }
 577         if (hcml->rmx_recvpipe != 0) {
 578                 if (hc_entry->rmx_recvpipe == 0)
 579                         hc_entry->rmx_recvpipe = hcml->rmx_recvpipe;
 580                 else
 581                         hc_entry->rmx_recvpipe =
 582                             (hc_entry->rmx_recvpipe + hcml->rmx_recvpipe) /2;
 583                 /* TCPSTAT_INC(tcps_cachedrecvpipe); */
 584         }
 585
 586         TAILQ_REMOVE(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
 587         TAILQ_INSERT_HEAD(&hc_entry->rmx_head->hch_bucket, hc_entry, rmx_q);
 588         THC_UNLOCK(&hc_entry->rmx_head->hch_mtx);
 589 }
 590
 591 /*
 592  * Sysctl function: prints the list and values of all hostcache entries in
 593  * unsorted order.
 594  */
 595 static int
 596 sysctl_tcp_hc_list(SYSCTL_HANDLER_ARGS)
 597 {
 598         int bufsize;
 599         int linesize = 128;
 600         char *p, *buf;
 601         int len, i, error;
 602         struct hc_metrics *hc_entry;
 603 #ifdef INET6
 604         char ip6buf[INET6_ADDRSTRLEN];
 605 #endif
 606
 607         bufsize = linesize * (V_tcp_hostcache.cache_count + 1);
 608
 609         p = buf = (char *)malloc(bufsize, M_TEMP, M_WAITOK|M_ZERO);
 610
 611         len = snprintf(p, linesize,
 612                 "\nIP address        MTU  SSTRESH      RTT   RTTVAR BANDWIDTH "
 613                 "    CWND SENDPIPE RECVPIPE HITS  UPD  EXP\n");
 614         p += len;
 615
 616 #define msec(u) (((u) + 500) / 1000)
 617         for (i = 0; i < V_tcp_hostcache.hashsize; i++) {
 618                 THC_LOCK(&V_tcp_hostcache.hashbase[i].hch_mtx);
 619                 TAILQ_FOREACH(hc_entry, &V_tcp_hostcache.hashbase[i].hch_bucket,
 620                               rmx_q) {
 621                         len = snprintf(p, linesize,
 622                             "%-15s %5lu %8lu %6lums %6lums %9lu %8lu %8lu %8lu "
 623                             "%4lu %4lu %4i\n",
 624                             hc_entry->ip4.s_addr ? inet_ntoa(hc_entry->ip4) :
 625 #ifdef INET6
 626                                 ip6_sprintf(ip6buf, &hc_entry->ip6),
 627 #else
 628                                 "IPv6?",
 629 #endif
 630                             hc_entry->rmx_mtu,
 631                             hc_entry->rmx_ssthresh,
 632                             msec(hc_entry->rmx_rtt *
 633                                 (RTM_RTTUNIT / (hz * TCP_RTT_SCALE))),
 634                             msec(hc_entry->rmx_rttvar *
 635                                 (RTM_RTTUNIT / (hz * TCP_RTTVAR_SCALE))),
 636                             hc_entry->rmx_bandwidth * 8,
 637                             hc_entry->rmx_cwnd,
 638                             hc_entry->rmx_sendpipe,
 639                             hc_entry->rmx_recvpipe,
 640                             hc_entry->rmx_hits,
 641                             hc_entry->rmx_updates,
 642                             hc_entry->rmx_expire);
 643                         p += len;
 644                 }
 645                 THC_UNLOCK(&V_tcp_hostcache.hashbase[i].hch_mtx);
 646         }
 647 #undef msec
 648         error = SYSCTL_OUT(req, buf, p - buf);
 649         free(buf, M_TEMP);
 650         return(error);
 651 }
 652
 653 /*
 654  * Caller has to make sure the curvnet is set properly.
 655  */
 656 static void
 657 tcp_hc_purge_internal(int all)
 658 {
 659         struct hc_metrics *hc_entry, *hc_next;
 660         int i;
 661
 662         for (i = 0; i < V_tcp_hostcache.hashsize; i++) {
 663                 THC_LOCK(&V_tcp_hostcache.hashbase[i].hch_mtx);
 664                 TAILQ_FOREACH_SAFE(hc_entry,
 665                     &V_tcp_hostcache.hashbase[i].hch_bucket, rmx_q, hc_next) {
 666                         if (all || hc_entry->rmx_expire <= 0) {
 667                                 TAILQ_REMOVE(&V_tcp_hostcache.hashbase[i].hch_bucket,
 668                                               hc_entry, rmx_q);
 669                                 uma_zfree(V_tcp_hostcache.zone, hc_entry);
 670                                 V_tcp_hostcache.hashbase[i].hch_length--;
 671                                 V_tcp_hostcache.cache_count--;
 672                         } else
 673                                 hc_entry->rmx_expire -= V_tcp_hostcache.prune;
 674                 }
 675                 THC_UNLOCK(&V_tcp_hostcache.hashbase[i].hch_mtx);
 676         }
 677 }
 678
 679 /*
 680  * Expire and purge (old|all) entries in the tcp_hostcache.  Runs
 681  * periodically from the callout.
 682  */
 683 static void
 684 tcp_hc_purge(void *arg)
 685 {
 686         CURVNET_SET((struct vnet *) arg);
 687         int all = 0;
 688
 689         if (V_tcp_hostcache.purgeall) {
 690                 all = 1;
 691                 V_tcp_hostcache.purgeall = 0;
 692         }
 693
 694         tcp_hc_purge_internal(all);
 695
 696         callout_reset(&V_tcp_hc_callout, V_tcp_hostcache.prune * hz,
 697             tcp_hc_purge, arg);
 698         CURVNET_RESTORE();
 699 }