Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / 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 1.10 2002/09/15 08:14:08 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/ip.h>
45 #include <linux/module.h>
46 #include <linux/kernel.h>
47 #include <linux/skbuff.h>
48 #include <linux/jiffies.h>
49
50 /* for sysctl */
51 #include <linux/fs.h>
52 #include <linux/sysctl.h>
53
54 #include <net/ip_vs.h>
55
56
57 /*
58  *    It is for garbage collection of stale IPVS lblc entries,
59  *    when the table is full.
60  */
61 #define CHECK_EXPIRE_INTERVAL   (60*HZ)
62 #define ENTRY_TIMEOUT           (6*60*HZ)
63
64 /*
65  *    It is for full expiration check.
66  *    When there is no partial expiration check (garbage collection)
67  *    in a half hour, do a full expiration check to collect stale
68  *    entries that haven't been touched for a day.
69  */
70 #define COUNT_FOR_FULL_EXPIRATION   30
71 static int sysctl_ip_vs_lblc_expiration = 24*60*60*HZ;
72
73
74 /*
75  *     for IPVS lblc entry hash table
76  */
77 #ifndef CONFIG_IP_VS_LBLC_TAB_BITS
78 #define CONFIG_IP_VS_LBLC_TAB_BITS      10
79 #endif
80 #define IP_VS_LBLC_TAB_BITS     CONFIG_IP_VS_LBLC_TAB_BITS
81 #define IP_VS_LBLC_TAB_SIZE     (1 << IP_VS_LBLC_TAB_BITS)
82 #define IP_VS_LBLC_TAB_MASK     (IP_VS_LBLC_TAB_SIZE - 1)
83
84
85 /*
86  *      IPVS lblc entry represents an association between destination
87  *      IP address and its destination server
88  */
89 struct ip_vs_lblc_entry {
90         struct list_head        list;
91         __be32                  addr;           /* destination IP address */
92         struct ip_vs_dest       *dest;          /* real server (cache) */
93         unsigned long           lastuse;        /* last used time */
94 };
95
96
97 /*
98  *      IPVS lblc hash table
99  */
100 struct ip_vs_lblc_table {
101         rwlock_t                lock;           /* lock for this table */
102         struct list_head        bucket[IP_VS_LBLC_TAB_SIZE];  /* hash bucket */
103         atomic_t                entries;        /* number of entries */
104         int                     max_size;       /* maximum size of entries */
105         struct timer_list       periodic_timer; /* collect stale entries */
106         int                     rover;          /* rover for expire check */
107         int                     counter;        /* counter for no expire */
108 };
109
110
111 /*
112  *      IPVS LBLC sysctl table
113  */
114
115 static ctl_table vs_vars_table[] = {
116         {
117                 .ctl_name       = NET_IPV4_VS_LBLC_EXPIRE,
118                 .procname       = "lblc_expiration",
119                 .data           = &sysctl_ip_vs_lblc_expiration,
120                 .maxlen         = sizeof(int),
121                 .mode           = 0644, 
122                 .proc_handler   = &proc_dointvec_jiffies,
123         },
124         { .ctl_name = 0 }
125 };
126
127 static ctl_table vs_table[] = {
128         {
129                 .ctl_name       = NET_IPV4_VS,
130                 .procname       = "vs",
131                 .mode           = 0555, 
132                 .child          = vs_vars_table
133         },
134         { .ctl_name = 0 }
135 };
136
137 static ctl_table ipvs_ipv4_table[] = {
138         {
139                 .ctl_name       = NET_IPV4,
140                 .procname       = "ipv4", 
141                 .mode           = 0555,
142                 .child          = vs_table
143         },
144         { .ctl_name = 0 }
145 };
146
147 static ctl_table lblc_root_table[] = {
148         {
149                 .ctl_name       = CTL_NET,
150                 .procname       = "net", 
151                 .mode           = 0555, 
152                 .child          = ipvs_ipv4_table
153         },
154         { .ctl_name = 0 }
155 };
156
157 static struct ctl_table_header * sysctl_header;
158
159 /*
160  *      new/free a ip_vs_lblc_entry, which is a mapping of a destionation
161  *      IP address to a server.
162  */
163 static inline struct ip_vs_lblc_entry *
164 ip_vs_lblc_new(__be32 daddr, struct ip_vs_dest *dest)
165 {
166         struct ip_vs_lblc_entry *en;
167
168         en = kmalloc(sizeof(struct ip_vs_lblc_entry), GFP_ATOMIC);
169         if (en == NULL) {
170                 IP_VS_ERR("ip_vs_lblc_new(): no memory\n");
171                 return NULL;
172         }
173
174         INIT_LIST_HEAD(&en->list);
175         en->addr = daddr;
176
177         atomic_inc(&dest->refcnt);
178         en->dest = dest;
179
180         return en;
181 }
182
183
184 static inline void ip_vs_lblc_free(struct ip_vs_lblc_entry *en)
185 {
186         list_del(&en->list);
187         /*
188          * We don't kfree dest because it is refered either by its service
189          * or the trash dest list.
190          */
191         atomic_dec(&en->dest->refcnt);
192         kfree(en);
193 }
194
195
196 /*
197  *      Returns hash value for IPVS LBLC entry
198  */
199 static inline unsigned ip_vs_lblc_hashkey(__be32 addr)
200 {
201         return (ntohl(addr)*2654435761UL) & IP_VS_LBLC_TAB_MASK;
202 }
203
204
205 /*
206  *      Hash an entry in the ip_vs_lblc_table.
207  *      returns bool success.
208  */
209 static int
210 ip_vs_lblc_hash(struct ip_vs_lblc_table *tbl, struct ip_vs_lblc_entry *en)
211 {
212         unsigned hash;
213
214         if (!list_empty(&en->list)) {
215                 IP_VS_ERR("ip_vs_lblc_hash(): request for already hashed, "
216                           "called from %p\n", __builtin_return_address(0));
217                 return 0;
218         }
219
220         /*
221          *      Hash by destination IP address
222          */
223         hash = ip_vs_lblc_hashkey(en->addr);
224
225         write_lock(&tbl->lock);
226         list_add(&en->list, &tbl->bucket[hash]);
227         atomic_inc(&tbl->entries);
228         write_unlock(&tbl->lock);
229
230         return 1;
231 }
232
233
234 /*
235  *  Get ip_vs_lblc_entry associated with supplied parameters.
236  */
237 static inline struct ip_vs_lblc_entry *
238 ip_vs_lblc_get(struct ip_vs_lblc_table *tbl, __be32 addr)
239 {
240         unsigned hash;
241         struct ip_vs_lblc_entry *en;
242
243         hash = ip_vs_lblc_hashkey(addr);
244
245         read_lock(&tbl->lock);
246
247         list_for_each_entry(en, &tbl->bucket[hash], list) {
248                 if (en->addr == addr) {
249                         /* HIT */
250                         read_unlock(&tbl->lock);
251                         return en;
252                 }
253         }
254
255         read_unlock(&tbl->lock);
256
257         return NULL;
258 }
259
260
261 /*
262  *      Flush all the entries of the specified table.
263  */
264 static void ip_vs_lblc_flush(struct ip_vs_lblc_table *tbl)
265 {
266         int i;
267         struct ip_vs_lblc_entry *en, *nxt;
268
269         for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
270                 write_lock(&tbl->lock);
271                 list_for_each_entry_safe(en, nxt, &tbl->bucket[i], list) {
272                         ip_vs_lblc_free(en);
273                         atomic_dec(&tbl->entries);
274                 }
275                 write_unlock(&tbl->lock);
276         }
277 }
278
279
280 static inline void ip_vs_lblc_full_check(struct ip_vs_lblc_table *tbl)
281 {
282         unsigned long now = jiffies;
283         int i, j;
284         struct ip_vs_lblc_entry *en, *nxt;
285
286         for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
287                 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
288
289                 write_lock(&tbl->lock);
290                 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
291                         if (time_before(now, 
292                                         en->lastuse + sysctl_ip_vs_lblc_expiration))
293                                 continue;
294
295                         ip_vs_lblc_free(en);
296                         atomic_dec(&tbl->entries);
297                 }
298                 write_unlock(&tbl->lock);
299         }
300         tbl->rover = j;
301 }
302
303
304 /*
305  *      Periodical timer handler for IPVS lblc table
306  *      It is used to collect stale entries when the number of entries
307  *      exceeds the maximum size of the table.
308  *
309  *      Fixme: we probably need more complicated algorithm to collect
310  *             entries that have not been used for a long time even
311  *             if the number of entries doesn't exceed the maximum size
312  *             of the table.
313  *      The full expiration check is for this purpose now.
314  */
315 static void ip_vs_lblc_check_expire(unsigned long data)
316 {
317         struct ip_vs_lblc_table *tbl;
318         unsigned long now = jiffies;
319         int goal;
320         int i, j;
321         struct ip_vs_lblc_entry *en, *nxt;
322
323         tbl = (struct ip_vs_lblc_table *)data;
324
325         if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) {
326                 /* do full expiration check */
327                 ip_vs_lblc_full_check(tbl);
328                 tbl->counter = 1;
329                 goto out;
330         }
331
332         if (atomic_read(&tbl->entries) <= tbl->max_size) {
333                 tbl->counter++;
334                 goto out;
335         }
336
337         goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3;
338         if (goal > tbl->max_size/2)
339                 goal = tbl->max_size/2;
340
341         for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
342                 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
343
344                 write_lock(&tbl->lock);
345                 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
346                         if (time_before(now, en->lastuse + ENTRY_TIMEOUT))
347                                 continue;
348
349                         ip_vs_lblc_free(en);
350                         atomic_dec(&tbl->entries);
351                         goal--;
352                 }
353                 write_unlock(&tbl->lock);
354                 if (goal <= 0)
355                         break;
356         }
357         tbl->rover = j;
358
359   out:
360         mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL);
361 }
362
363
364 static int ip_vs_lblc_init_svc(struct ip_vs_service *svc)
365 {
366         int i;
367         struct ip_vs_lblc_table *tbl;
368
369         /*
370          *    Allocate the ip_vs_lblc_table for this service
371          */
372         tbl = kmalloc(sizeof(struct ip_vs_lblc_table), GFP_ATOMIC);
373         if (tbl == NULL) {
374                 IP_VS_ERR("ip_vs_lblc_init_svc(): no memory\n");
375                 return -ENOMEM;
376         }
377         svc->sched_data = tbl;
378         IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) allocated for "
379                   "current service\n",
380                   sizeof(struct ip_vs_lblc_table));
381
382         /*
383          *    Initialize the hash buckets
384          */
385         for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
386                 INIT_LIST_HEAD(&tbl->bucket[i]);
387         }
388         rwlock_init(&tbl->lock);
389         tbl->max_size = IP_VS_LBLC_TAB_SIZE*16;
390         tbl->rover = 0;
391         tbl->counter = 1;
392
393         /*
394          *    Hook periodic timer for garbage collection
395          */
396         init_timer(&tbl->periodic_timer);
397         tbl->periodic_timer.data = (unsigned long)tbl;
398         tbl->periodic_timer.function = ip_vs_lblc_check_expire;
399         tbl->periodic_timer.expires = jiffies+CHECK_EXPIRE_INTERVAL;
400         add_timer(&tbl->periodic_timer);
401
402         return 0;
403 }
404
405
406 static int ip_vs_lblc_done_svc(struct ip_vs_service *svc)
407 {
408         struct ip_vs_lblc_table *tbl = svc->sched_data;
409
410         /* remove periodic timer */
411         del_timer_sync(&tbl->periodic_timer);
412
413         /* got to clean up table entries here */
414         ip_vs_lblc_flush(tbl);
415
416         /* release the table itself */
417         kfree(svc->sched_data);
418         IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) released\n",
419                   sizeof(struct ip_vs_lblc_table));
420
421         return 0;
422 }
423
424
425 static int ip_vs_lblc_update_svc(struct ip_vs_service *svc)
426 {
427         return 0;
428 }
429
430
431 static inline struct ip_vs_dest *
432 __ip_vs_wlc_schedule(struct ip_vs_service *svc, struct iphdr *iph)
433 {
434         struct ip_vs_dest *dest, *least;
435         int loh, doh;
436
437         /*
438          * We think the overhead of processing active connections is fifty
439          * times higher than that of inactive connections in average. (This
440          * fifty times might not be accurate, we will change it later.) We
441          * use the following formula to estimate the overhead:
442          *                dest->activeconns*50 + dest->inactconns
443          * and the load:
444          *                (dest overhead) / dest->weight
445          *
446          * Remember -- no floats in kernel mode!!!
447          * The comparison of h1*w2 > h2*w1 is equivalent to that of
448          *                h1/w1 > h2/w2
449          * if every weight is larger than zero.
450          *
451          * The server with weight=0 is quiesced and will not receive any
452          * new connection.
453          */
454         list_for_each_entry(dest, &svc->destinations, n_list) {
455                 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
456                         continue;
457                 if (atomic_read(&dest->weight) > 0) {
458                         least = dest;
459                         loh = atomic_read(&least->activeconns) * 50
460                                 + atomic_read(&least->inactconns);
461                         goto nextstage;
462                 }
463         }
464         return NULL;
465
466         /*
467          *    Find the destination with the least load.
468          */
469   nextstage:
470         list_for_each_entry_continue(dest, &svc->destinations, n_list) {
471                 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
472                         continue;
473
474                 doh = atomic_read(&dest->activeconns) * 50
475                         + atomic_read(&dest->inactconns);
476                 if (loh * atomic_read(&dest->weight) >
477                     doh * atomic_read(&least->weight)) {
478                         least = dest;
479                         loh = doh;
480                 }
481         }
482
483         IP_VS_DBG(6, "LBLC: server %d.%d.%d.%d:%d "
484                   "activeconns %d refcnt %d weight %d overhead %d\n",
485                   NIPQUAD(least->addr), ntohs(least->port),
486                   atomic_read(&least->activeconns),
487                   atomic_read(&least->refcnt),
488                   atomic_read(&least->weight), loh);
489
490         return least;
491 }
492
493
494 /*
495  *   If this destination server is overloaded and there is a less loaded
496  *   server, then return true.
497  */
498 static inline int
499 is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc)
500 {
501         if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) {
502                 struct ip_vs_dest *d;
503
504                 list_for_each_entry(d, &svc->destinations, n_list) {
505                         if (atomic_read(&d->activeconns)*2
506                             < atomic_read(&d->weight)) {
507                                 return 1;
508                         }
509                 }
510         }
511         return 0;
512 }
513
514
515 /*
516  *    Locality-Based (weighted) Least-Connection scheduling
517  */
518 static struct ip_vs_dest *
519 ip_vs_lblc_schedule(struct ip_vs_service *svc, const struct sk_buff *skb)
520 {
521         struct ip_vs_dest *dest;
522         struct ip_vs_lblc_table *tbl;
523         struct ip_vs_lblc_entry *en;
524         struct iphdr *iph = skb->nh.iph;
525
526         IP_VS_DBG(6, "ip_vs_lblc_schedule(): Scheduling...\n");
527
528         tbl = (struct ip_vs_lblc_table *)svc->sched_data;
529         en = ip_vs_lblc_get(tbl, iph->daddr);
530         if (en == NULL) {
531                 dest = __ip_vs_wlc_schedule(svc, iph);
532                 if (dest == NULL) {
533                         IP_VS_DBG(1, "no destination available\n");
534                         return NULL;
535                 }
536                 en = ip_vs_lblc_new(iph->daddr, dest);
537                 if (en == NULL) {
538                         return NULL;
539                 }
540                 ip_vs_lblc_hash(tbl, en);
541         } else {
542                 dest = en->dest;
543                 if (!(dest->flags & IP_VS_DEST_F_AVAILABLE)
544                     || atomic_read(&dest->weight) <= 0
545                     || is_overloaded(dest, svc)) {
546                         dest = __ip_vs_wlc_schedule(svc, iph);
547                         if (dest == NULL) {
548                                 IP_VS_DBG(1, "no destination available\n");
549                                 return NULL;
550                         }
551                         atomic_dec(&en->dest->refcnt);
552                         atomic_inc(&dest->refcnt);
553                         en->dest = dest;
554                 }
555         }
556         en->lastuse = jiffies;
557
558         IP_VS_DBG(6, "LBLC: destination IP address %u.%u.%u.%u "
559                   "--> server %u.%u.%u.%u:%d\n",
560                   NIPQUAD(en->addr),
561                   NIPQUAD(dest->addr),
562                   ntohs(dest->port));
563
564         return dest;
565 }
566
567
568 /*
569  *      IPVS LBLC Scheduler structure
570  */
571 static struct ip_vs_scheduler ip_vs_lblc_scheduler =
572 {
573         .name =                 "lblc",
574         .refcnt =               ATOMIC_INIT(0),
575         .module =               THIS_MODULE,
576         .init_service =         ip_vs_lblc_init_svc,
577         .done_service =         ip_vs_lblc_done_svc,
578         .update_service =       ip_vs_lblc_update_svc,
579         .schedule =             ip_vs_lblc_schedule,
580 };
581
582
583 static int __init ip_vs_lblc_init(void)
584 {
585         INIT_LIST_HEAD(&ip_vs_lblc_scheduler.n_list);
586         sysctl_header = register_sysctl_table(lblc_root_table, 0);
587         return register_ip_vs_scheduler(&ip_vs_lblc_scheduler);
588 }
589
590
591 static void __exit ip_vs_lblc_cleanup(void)
592 {
593         unregister_sysctl_table(sysctl_header);
594         unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler);
595 }
596
597
598 module_init(ip_vs_lblc_init);
599 module_exit(ip_vs_lblc_cleanup);
600 MODULE_LICENSE("GPL");