net/ipv4/ipvs/ip_vs_lblc.c

   1 /*
   2  * IPVS:        Locality-Based Least-Connection scheduling module
   3  *
   4  * Version:     $Id: ip_vs_lblc.c,v 1.9 2002/03/25 12:44:35 wensong Exp $
   5  *
   6  * Authors:     Wensong Zhang <wensong@gnuchina.org>
   7  *
   8  *              This program is free software; you can redistribute it and/or
   9  *              modify it under the terms of the GNU General Public License
  10  *              as published by the Free Software Foundation; either version
  11  *              2 of the License, or (at your option) any later version.
  12  *
  13  * Changes:
  14  *     Martin Hamilton         :    fixed the terrible locking bugs
  15  *                                   *lock(tbl->lock) ==> *lock(&tbl->lock)
  16  *     Wensong Zhang           :    fixed the uninitilized tbl->lock bug
  17  *     Wensong Zhang           :    added doing full expiration check to
  18  *                                   collect stale entries of 24+ hours when
  19  *                                   no partial expire check in a half hour
  20  *     Julian Anastasov        :    replaced del_timer call with del_timer_sync
  21  *                                   to avoid the possible race between timer
  22  *                                   handler and del_timer thread in SMP
  23  *
  24  */
  25
  26 /*
  27  * The lblc algorithm is as follows (pseudo code):
  28  *
  29  *       if cachenode[dest_ip] is null then
  30  *               n, cachenode[dest_ip] <- {weighted least-conn node};
  31  *       else
  32  *               n <- cachenode[dest_ip];
  33  *               if (n is dead) OR
  34  *                  (n.conns>n.weight AND
  35  *                   there is a node m with m.conns<m.weight/2) then
  36  *                 n, cachenode[dest_ip] <- {weighted least-conn node};
  37  *
  38  *       return n;
  39  *
  40  * Thanks must go to Wenzhuo Zhang for talking WCCP to me and pushing
  41  * me to write this module.
  42  */
  43
  44 #include <linux/module.h>
  45 #include <linux/kernel.h>
  46
  47 /* for sysctl */
  48 #include <linux/fs.h>
  49 #include <linux/sysctl.h>
  50
  51 #include <net/ip_vs.h>
  52
  53
  54 /*
  55  *    It is for garbage collection of stale IPVS lblc entries,
  56  *    when the table is full.
  57  */
  58 #define CHECK_EXPIRE_INTERVAL   (60*HZ)
  59 #define ENTRY_TIMEOUT           (6*60*HZ)
  60
  61 /*
  62  *    It is for full expiration check.
  63  *    When there is no partial expiration check (garbage collection)
  64  *    in a half hour, do a full expiration check to collect stale
  65  *    entries that haven't been touched for a day.
  66  */
  67 #define COUNT_FOR_FULL_EXPIRATION   30
  68 static int sysctl_ip_vs_lblc_expiration = 24*60*60*HZ;
  69
  70
  71 /*
  72  *     for IPVS lblc entry hash table
  73  */
  74 #ifndef CONFIG_IP_VS_LBLC_TAB_BITS
  75 #define CONFIG_IP_VS_LBLC_TAB_BITS      10
  76 #endif
  77 #define IP_VS_LBLC_TAB_BITS     CONFIG_IP_VS_LBLC_TAB_BITS
  78 #define IP_VS_LBLC_TAB_SIZE     (1 << IP_VS_LBLC_TAB_BITS)
  79 #define IP_VS_LBLC_TAB_MASK     (IP_VS_LBLC_TAB_SIZE - 1)
  80
  81
  82 /*
  83  *      IPVS lblc entry represents an association between destination
  84  *      IP address and its destination server
  85  */
  86 struct ip_vs_lblc_entry {
  87         struct list_head        list;
  88         __u32                   addr;           /* destination IP address */
  89         struct ip_vs_dest       *dest;          /* real server (cache) */
  90         unsigned long           lastuse;        /* last used time */
  91 };
  92
  93
  94 /*
  95  *      IPVS lblc hash table
  96  */
  97 struct ip_vs_lblc_table {
  98         rwlock_t                lock;           /* lock for this table */
  99         struct list_head        bucket[IP_VS_LBLC_TAB_SIZE];  /* hash bucket */
 100         atomic_t                entries;        /* number of entries */
 101         int                     max_size;       /* maximum size of entries */
 102         struct timer_list       periodic_timer; /* collect stale entries */
 103         int                     rover;          /* rover for expire check */
 104         int                     counter;        /* counter for no expire */
 105 };
 106
 107
 108 /*
 109  *      IPVS LBLC sysctl table
 110  */
 111 struct ip_vs_lblc_sysctl_table {
 112         struct ctl_table_header *sysctl_header;
 113         ctl_table vs_vars[2];
 114         ctl_table vs_dir[2];
 115         ctl_table ipv4_dir[2];
 116         ctl_table root_dir[2];
 117 };
 118
 119
 120 static struct ip_vs_lblc_sysctl_table lblc_sysctl_table = {
 121         NULL,
 122         {{NET_IPV4_VS_LBLC_EXPIRE, "lblc_expiration",
 123           &sysctl_ip_vs_lblc_expiration,
 124           sizeof(int), 0644, NULL, &proc_dointvec_jiffies},
 125          {0}},
 126         {{NET_IPV4_VS, "vs", NULL, 0, 0555, lblc_sysctl_table.vs_vars},
 127          {0}},
 128         {{NET_IPV4, "ipv4", NULL, 0, 0555, lblc_sysctl_table.vs_dir},
 129          {0}},
 130         {{CTL_NET, "net", NULL, 0, 0555, lblc_sysctl_table.ipv4_dir},
 131          {0}}
 132 };
 133
 134
 135 /*
 136  *      new/free a ip_vs_lblc_entry, which is a mapping of a destionation
 137  *      IP address to a server.
 138  */
 139 static inline struct ip_vs_lblc_entry *
 140 ip_vs_lblc_new(__u32 daddr, struct ip_vs_dest *dest)
 141 {
 142         struct ip_vs_lblc_entry *en;
 143
 144         en = kmalloc(sizeof(struct ip_vs_lblc_entry), GFP_ATOMIC);
 145         if (en == NULL) {
 146                 IP_VS_ERR("ip_vs_lblc_new(): no memory\n");
 147                 return NULL;
 148         }
 149
 150         INIT_LIST_HEAD(&en->list);
 151         en->addr = daddr;
 152
 153         atomic_inc(&dest->refcnt);
 154         en->dest = dest;
 155
 156         return en;
 157 }
 158
 159
 160 static inline void ip_vs_lblc_free(struct ip_vs_lblc_entry *en)
 161 {
 162         list_del(&en->list);
 163         /*
 164          * We don't kfree dest because it is refered either by its service
 165          * or the trash dest list.
 166          */
 167         atomic_dec(&en->dest->refcnt);
 168         kfree(en);
 169 }
 170
 171
 172 /*
 173  *      Returns hash value for IPVS LBLC entry
 174  */
 175 static inline unsigned ip_vs_lblc_hashkey(__u32 addr)
 176 {
 177         return (ntohl(addr)*2654435761UL) & IP_VS_LBLC_TAB_MASK;
 178 }
 179
 180
 181 /*
 182  *      Hash an entry in the ip_vs_lblc_table.
 183  *      returns bool success.
 184  */
 185 static int
 186 ip_vs_lblc_hash(struct ip_vs_lblc_table *tbl, struct ip_vs_lblc_entry *en)
 187 {
 188         unsigned hash;
 189
 190         if (!list_empty(&en->list)) {
 191                 IP_VS_ERR("ip_vs_lblc_hash(): request for already hashed, "
 192                           "called from %p\n", __builtin_return_address(0));
 193                 return 0;
 194         }
 195
 196         /*
 197          *      Hash by destination IP address
 198          */
 199         hash = ip_vs_lblc_hashkey(en->addr);
 200
 201         write_lock(&tbl->lock);
 202         list_add(&en->list, &tbl->bucket[hash]);
 203         atomic_inc(&tbl->entries);
 204         write_unlock(&tbl->lock);
 205
 206         return 1;
 207 }
 208
 209
 210 #if 0000
 211 /*
 212  *      Unhash ip_vs_lblc_entry from ip_vs_lblc_table.
 213  *      returns bool success.
 214  */
 215 static int ip_vs_lblc_unhash(struct ip_vs_lblc_table *tbl,
 216                              struct ip_vs_lblc_entry *en)
 217 {
 218         if (list_empty(&en->list)) {
 219                 IP_VS_ERR("ip_vs_lblc_unhash(): request for not hashed entry, "
 220                           "called from %p\n", __builtin_return_address(0));
 221                 return 0;
 222         }
 223
 224         /*
 225          * Remove it from the table
 226          */
 227         write_lock(&tbl->lock);
 228         list_del(&en->list);
 229         INIT_LIST_HEAD(&en->list);
 230         write_unlock(&tbl->lock);
 231
 232         return 1;
 233 }
 234 #endif
 235
 236
 237 /*
 238  *  Get ip_vs_lblc_entry associated with supplied parameters.
 239  */
 240 static inline struct ip_vs_lblc_entry *
 241 ip_vs_lblc_get(struct ip_vs_lblc_table *tbl, __u32 addr)
 242 {
 243         unsigned hash;
 244         struct ip_vs_lblc_entry *en;
 245         struct list_head *l,*e;
 246
 247         hash = ip_vs_lblc_hashkey(addr);
 248         l = &tbl->bucket[hash];
 249
 250         read_lock(&tbl->lock);
 251
 252         for (e=l->next; e!=l; e=e->next) {
 253                 en = list_entry(e, struct ip_vs_lblc_entry, list);
 254                 if (en->addr == addr) {
 255                         /* HIT */
 256                         read_unlock(&tbl->lock);
 257                         return en;
 258                 }
 259         }
 260
 261         read_unlock(&tbl->lock);
 262
 263         return NULL;
 264 }
 265
 266
 267 /*
 268  *      Flush all the entries of the specified table.
 269  */
 270 static void ip_vs_lblc_flush(struct ip_vs_lblc_table *tbl)
 271 {
 272         int i;
 273         struct list_head *l;
 274         struct ip_vs_lblc_entry *en;
 275
 276         for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
 277                 write_lock(&tbl->lock);
 278                 for (l=&tbl->bucket[i]; l->next!=l; ) {
 279                         en = list_entry(l->next,
 280                                         struct ip_vs_lblc_entry, list);
 281                         ip_vs_lblc_free(en);
 282                         atomic_dec(&tbl->entries);
 283                 }
 284                 write_unlock(&tbl->lock);
 285         }
 286 }
 287
 288
 289 static inline void ip_vs_lblc_full_check(struct ip_vs_lblc_table *tbl)
 290 {
 291         unsigned long now = jiffies;
 292         int i, j;
 293         struct list_head *l, *e;
 294         struct ip_vs_lblc_entry *en;
 295
 296         for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
 297                 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
 298                 e = l = &tbl->bucket[j];
 299                 write_lock(&tbl->lock);
 300                 while (e->next != l) {
 301                         en = list_entry(e->next,
 302                                         struct ip_vs_lblc_entry, list);
 303                         if ((now - en->lastuse) <
 304                             sysctl_ip_vs_lblc_expiration) {
 305                                 e = e->next;
 306                                 continue;
 307                         }
 308                         ip_vs_lblc_free(en);
 309                         atomic_dec(&tbl->entries);
 310                 }
 311                 write_unlock(&tbl->lock);
 312         }
 313         tbl->rover = j;
 314 }
 315
 316
 317 /*
 318  *      Periodical timer handler for IPVS lblc table
 319  *      It is used to collect stale entries when the number of entries
 320  *      exceeds the maximum size of the table.
 321  *
 322  *      Fixme: we probably need more complicated algorithm to collect
 323  *             entries that have not been used for a long time even
 324  *             if the number of entries doesn't exceed the maximum size
 325  *             of the table.
 326  *      The full expiration check is for this purpose now.
 327  */
 328 static void ip_vs_lblc_check_expire(unsigned long data)
 329 {
 330         struct ip_vs_lblc_table *tbl;
 331         unsigned long now = jiffies;
 332         int goal;
 333         int i, j;
 334         struct list_head *l, *e;
 335         struct ip_vs_lblc_entry *en;
 336
 337         tbl = (struct ip_vs_lblc_table *)data;
 338
 339         if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) {
 340                 /* do full expiration check */
 341                 ip_vs_lblc_full_check(tbl);
 342                 tbl->counter = 1;
 343                 goto out;
 344         }
 345
 346         if (atomic_read(&tbl->entries) <= tbl->max_size) {
 347                 tbl->counter++;
 348                 goto out;
 349         }
 350
 351         goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3;
 352         if (goal > tbl->max_size/2)
 353                 goal = tbl->max_size/2;
 354
 355         for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
 356                 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
 357                 e = l = &tbl->bucket[j];
 358                 write_lock(&tbl->lock);
 359                 while (e->next != l) {
 360                         en = list_entry(e->next,
 361                                         struct ip_vs_lblc_entry, list);
 362                         if ((now - en->lastuse) < ENTRY_TIMEOUT) {
 363                                 e = e->next;
 364                                 continue;
 365                         }
 366                         ip_vs_lblc_free(en);
 367                         atomic_dec(&tbl->entries);
 368                         goal--;
 369                 }
 370                 write_unlock(&tbl->lock);
 371                 if (goal <= 0)
 372                         break;
 373         }
 374         tbl->rover = j;
 375
 376   out:
 377         mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL);
 378 }
 379
 380
 381 static int ip_vs_lblc_init_svc(struct ip_vs_service *svc)
 382 {
 383         int i;
 384         struct ip_vs_lblc_table *tbl;
 385
 386         /*
 387          *    Allocate the ip_vs_lblc_table for this service
 388          */
 389         tbl = kmalloc(sizeof(struct ip_vs_lblc_table), GFP_ATOMIC);
 390         if (tbl == NULL) {
 391                 IP_VS_ERR("ip_vs_lblc_init_svc(): no memory\n");
 392                 return -ENOMEM;
 393         }
 394         svc->sched_data = tbl;
 395         IP_VS_DBG(6, "LBLC hash table (memory=%dbytes) allocated for "
 396                   "current service\n",
 397                   sizeof(struct ip_vs_lblc_table));
 398
 399         /*
 400          *    Initialize the hash buckets
 401          */
 402         for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
 403                 INIT_LIST_HEAD(&tbl->bucket[i]);
 404         }
 405         tbl->lock = RW_LOCK_UNLOCKED;
 406         tbl->max_size = IP_VS_LBLC_TAB_SIZE*16;
 407         tbl->rover = 0;
 408         tbl->counter = 1;
 409
 410         /*
 411          *    Hook periodic timer for garbage collection
 412          */
 413         init_timer(&tbl->periodic_timer);
 414         tbl->periodic_timer.data = (unsigned long)tbl;
 415         tbl->periodic_timer.function = ip_vs_lblc_check_expire;
 416         tbl->periodic_timer.expires = jiffies+CHECK_EXPIRE_INTERVAL;
 417         add_timer(&tbl->periodic_timer);
 418
 419         return 0;
 420 }
 421
 422
 423 static int ip_vs_lblc_done_svc(struct ip_vs_service *svc)
 424 {
 425         struct ip_vs_lblc_table *tbl = svc->sched_data;
 426
 427         /* remove periodic timer */
 428         del_timer_sync(&tbl->periodic_timer);
 429
 430         /* got to clean up table entries here */
 431         ip_vs_lblc_flush(tbl);
 432
 433         /* release the table itself */
 434         kfree(svc->sched_data);
 435         IP_VS_DBG(6, "LBLC hash table (memory=%dbytes) released\n",
 436                   sizeof(struct ip_vs_lblc_table));
 437
 438         return 0;
 439 }
 440
 441
 442 static int ip_vs_lblc_update_svc(struct ip_vs_service *svc)
 443 {
 444         return 0;
 445 }
 446
 447
 448 static inline struct ip_vs_dest *
 449 __ip_vs_wlc_schedule(struct ip_vs_service *svc, struct iphdr *iph)
 450 {
 451         register struct list_head *l, *e;
 452         struct ip_vs_dest *dest, *least;
 453         int loh, doh;
 454
 455         /*
 456          * We think the overhead of processing active connections is fifty
 457          * times higher than that of inactive connections in average. (This
 458          * fifty times might not be accurate, we will change it later.) We
 459          * use the following formula to estimate the overhead:
 460          *                dest->activeconns*50 + dest->inactconns
 461          * and the load:
 462          *                (dest overhead) / dest->weight
 463          *
 464          * Remember -- no floats in kernel mode!!!
 465          * The comparison of h1*w2 > h2*w1 is equivalent to that of
 466          *                h1/w1 > h2/w2
 467          * if every weight is larger than zero.
 468          *
 469          * The server with weight=0 is quiesced and will not receive any
 470          * new connection.
 471          */
 472
 473         l = &svc->destinations;
 474         for (e=l->next; e!=l; e=e->next) {
 475                 least = list_entry(e, struct ip_vs_dest, n_list);
 476                 if (atomic_read(&least->weight) > 0) {
 477                         loh = atomic_read(&least->activeconns) * 50
 478                                 + atomic_read(&least->inactconns);
 479                         goto nextstage;
 480                 }
 481         }
 482         return NULL;
 483
 484         /*
 485          *    Find the destination with the least load.
 486          */
 487   nextstage:
 488         for (e=e->next; e!=l; e=e->next) {
 489                 dest = list_entry(e, struct ip_vs_dest, n_list);
 490                 doh = atomic_read(&dest->activeconns) * 50
 491                         + atomic_read(&dest->inactconns);
 492                 if (loh * atomic_read(&dest->weight) >
 493                     doh * atomic_read(&least->weight)) {
 494                         least = dest;
 495                         loh = doh;
 496                 }
 497         }
 498
 499         IP_VS_DBG(6, "LBLC: server %d.%d.%d.%d:%d "
 500                   "activeconns %d refcnt %d weight %d overhead %d\n",
 501                   NIPQUAD(least->addr), ntohs(least->port),
 502                   atomic_read(&least->activeconns),
 503                   atomic_read(&least->refcnt),
 504                   atomic_read(&least->weight), loh);
 505
 506         return least;
 507 }
 508
 509
 510 /*
 511  *   If this destination server is overloaded and there is a less loaded
 512  *   server, then return true.
 513  */
 514 static inline int
 515 is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc)
 516 {
 517         if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) {
 518                 register struct list_head *l, *e;
 519                 struct ip_vs_dest *d;
 520
 521                 l = &svc->destinations;
 522                 for (e=l->next; e!=l; e=e->next) {
 523                         d = list_entry(e, struct ip_vs_dest, n_list);
 524                         if (atomic_read(&d->activeconns)*2
 525                             < atomic_read(&d->weight)) {
 526                                 return 1;
 527                         }
 528                 }
 529         }
 530         return 0;
 531 }
 532
 533
 534 /*
 535  *    Locality-Based (weighted) Least-Connection scheduling
 536  */
 537 static struct ip_vs_dest *
 538 ip_vs_lblc_schedule(struct ip_vs_service *svc, struct iphdr *iph)
 539 {
 540         struct ip_vs_dest *dest;
 541         struct ip_vs_lblc_table *tbl;
 542         struct ip_vs_lblc_entry *en;
 543
 544         IP_VS_DBG(6, "ip_vs_lblc_schedule(): Scheduling...\n");
 545
 546         tbl = (struct ip_vs_lblc_table *)svc->sched_data;
 547         en = ip_vs_lblc_get(tbl, iph->daddr);
 548         if (en == NULL) {
 549                 dest = __ip_vs_wlc_schedule(svc, iph);
 550                 if (dest == NULL) {
 551                         IP_VS_DBG(1, "no destination available\n");
 552                         return NULL;
 553                 }
 554                 en = ip_vs_lblc_new(iph->daddr, dest);
 555                 if (en == NULL) {
 556                         return NULL;
 557                 }
 558                 ip_vs_lblc_hash(tbl, en);
 559         } else {
 560                 dest = en->dest;
 561                 if (!(dest->flags & IP_VS_DEST_F_AVAILABLE)
 562                     || atomic_read(&dest->weight) <= 0
 563                     || is_overloaded(dest, svc)) {
 564                         dest = __ip_vs_wlc_schedule(svc, iph);
 565                         if (dest == NULL) {
 566                                 IP_VS_DBG(1, "no destination available\n");
 567                                 return NULL;
 568                         }
 569                         atomic_dec(&en->dest->refcnt);
 570                         atomic_inc(&dest->refcnt);
 571                         en->dest = dest;
 572                 }
 573         }
 574         en->lastuse = jiffies;
 575
 576         IP_VS_DBG(6, "LBLC: destination IP address %u.%u.%u.%u "
 577                   "--> server %u.%u.%u.%u:%d\n",
 578                   NIPQUAD(en->addr),
 579                   NIPQUAD(dest->addr),
 580                   ntohs(dest->port));
 581
 582         return dest;
 583 }
 584
 585
 586 /*
 587  *      IPVS LBLC Scheduler structure
 588  */
 589 static struct ip_vs_scheduler ip_vs_lblc_scheduler =
 590 {
 591         {0},                    /* n_list */
 592         "lblc",                 /* name */
 593         ATOMIC_INIT(0),         /* refcnt */
 594         THIS_MODULE,            /* this module */
 595         ip_vs_lblc_init_svc,    /* service initializer */
 596         ip_vs_lblc_done_svc,    /* service done */
 597         ip_vs_lblc_update_svc,  /* service updater */
 598         ip_vs_lblc_schedule,    /* select a server from the destination list */
 599 };
 600
 601
 602 static int __init ip_vs_lblc_init(void)
 603 {
 604         INIT_LIST_HEAD(&ip_vs_lblc_scheduler.n_list);
 605         lblc_sysctl_table.sysctl_header =
 606                 register_sysctl_table(lblc_sysctl_table.root_dir, 0);
 607         return register_ip_vs_scheduler(&ip_vs_lblc_scheduler);
 608 }
 609
 610
 611 static void __exit ip_vs_lblc_cleanup(void)
 612 {
 613         unregister_sysctl_table(lblc_sysctl_table.sysctl_header);
 614         unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler);
 615 }
 616
 617
 618 module_init(ip_vs_lblc_init);
 619 module_exit(ip_vs_lblc_cleanup);
 620 MODULE_LICENSE("GPL");