ad89644ef5d27e797f9a188451dc4f5c315cce70
[linux-2.6.git] / 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                 .procname       = "lblc_expiration",
118                 .data           = &sysctl_ip_vs_lblc_expiration,
119                 .maxlen         = sizeof(int),
120                 .mode           = 0644,
121                 .proc_handler   = &proc_dointvec_jiffies,
122         },
123         { .ctl_name = 0 }
124 };
125
126 static ctl_table vs_table[] = {
127         {
128                 .procname       = "vs",
129                 .mode           = 0555,
130                 .child          = vs_vars_table
131         },
132         { .ctl_name = 0 }
133 };
134
135 static ctl_table ipvs_ipv4_table[] = {
136         {
137                 .ctl_name       = NET_IPV4,
138                 .procname       = "ipv4",
139                 .mode           = 0555,
140                 .child          = vs_table
141         },
142         { .ctl_name = 0 }
143 };
144
145 static ctl_table lblc_root_table[] = {
146         {
147                 .ctl_name       = CTL_NET,
148                 .procname       = "net",
149                 .mode           = 0555,
150                 .child          = ipvs_ipv4_table
151         },
152         { .ctl_name = 0 }
153 };
154
155 static struct ctl_table_header * sysctl_header;
156
157 /*
158  *      new/free a ip_vs_lblc_entry, which is a mapping of a destionation
159  *      IP address to a server.
160  */
161 static inline struct ip_vs_lblc_entry *
162 ip_vs_lblc_new(__be32 daddr, struct ip_vs_dest *dest)
163 {
164         struct ip_vs_lblc_entry *en;
165
166         en = kmalloc(sizeof(struct ip_vs_lblc_entry), GFP_ATOMIC);
167         if (en == NULL) {
168                 IP_VS_ERR("ip_vs_lblc_new(): no memory\n");
169                 return NULL;
170         }
171
172         INIT_LIST_HEAD(&en->list);
173         en->addr = daddr;
174
175         atomic_inc(&dest->refcnt);
176         en->dest = dest;
177
178         return en;
179 }
180
181
182 static inline void ip_vs_lblc_free(struct ip_vs_lblc_entry *en)
183 {
184         list_del(&en->list);
185         /*
186          * We don't kfree dest because it is refered either by its service
187          * or the trash dest list.
188          */
189         atomic_dec(&en->dest->refcnt);
190         kfree(en);
191 }
192
193
194 /*
195  *      Returns hash value for IPVS LBLC entry
196  */
197 static inline unsigned ip_vs_lblc_hashkey(__be32 addr)
198 {
199         return (ntohl(addr)*2654435761UL) & IP_VS_LBLC_TAB_MASK;
200 }
201
202
203 /*
204  *      Hash an entry in the ip_vs_lblc_table.
205  *      returns bool success.
206  */
207 static int
208 ip_vs_lblc_hash(struct ip_vs_lblc_table *tbl, struct ip_vs_lblc_entry *en)
209 {
210         unsigned hash;
211
212         if (!list_empty(&en->list)) {
213                 IP_VS_ERR("ip_vs_lblc_hash(): request for already hashed, "
214                           "called from %p\n", __builtin_return_address(0));
215                 return 0;
216         }
217
218         /*
219          *      Hash by destination IP address
220          */
221         hash = ip_vs_lblc_hashkey(en->addr);
222
223         write_lock(&tbl->lock);
224         list_add(&en->list, &tbl->bucket[hash]);
225         atomic_inc(&tbl->entries);
226         write_unlock(&tbl->lock);
227
228         return 1;
229 }
230
231
232 /*
233  *  Get ip_vs_lblc_entry associated with supplied parameters.
234  */
235 static inline struct ip_vs_lblc_entry *
236 ip_vs_lblc_get(struct ip_vs_lblc_table *tbl, __be32 addr)
237 {
238         unsigned hash;
239         struct ip_vs_lblc_entry *en;
240
241         hash = ip_vs_lblc_hashkey(addr);
242
243         read_lock(&tbl->lock);
244
245         list_for_each_entry(en, &tbl->bucket[hash], list) {
246                 if (en->addr == addr) {
247                         /* HIT */
248                         read_unlock(&tbl->lock);
249                         return en;
250                 }
251         }
252
253         read_unlock(&tbl->lock);
254
255         return NULL;
256 }
257
258
259 /*
260  *      Flush all the entries of the specified table.
261  */
262 static void ip_vs_lblc_flush(struct ip_vs_lblc_table *tbl)
263 {
264         int i;
265         struct ip_vs_lblc_entry *en, *nxt;
266
267         for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
268                 write_lock(&tbl->lock);
269                 list_for_each_entry_safe(en, nxt, &tbl->bucket[i], list) {
270                         ip_vs_lblc_free(en);
271                         atomic_dec(&tbl->entries);
272                 }
273                 write_unlock(&tbl->lock);
274         }
275 }
276
277
278 static inline void ip_vs_lblc_full_check(struct ip_vs_lblc_table *tbl)
279 {
280         unsigned long now = jiffies;
281         int i, j;
282         struct ip_vs_lblc_entry *en, *nxt;
283
284         for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
285                 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
286
287                 write_lock(&tbl->lock);
288                 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
289                         if (time_before(now,
290                                         en->lastuse + sysctl_ip_vs_lblc_expiration))
291                                 continue;
292
293                         ip_vs_lblc_free(en);
294                         atomic_dec(&tbl->entries);
295                 }
296                 write_unlock(&tbl->lock);
297         }
298         tbl->rover = j;
299 }
300
301
302 /*
303  *      Periodical timer handler for IPVS lblc table
304  *      It is used to collect stale entries when the number of entries
305  *      exceeds the maximum size of the table.
306  *
307  *      Fixme: we probably need more complicated algorithm to collect
308  *             entries that have not been used for a long time even
309  *             if the number of entries doesn't exceed the maximum size
310  *             of the table.
311  *      The full expiration check is for this purpose now.
312  */
313 static void ip_vs_lblc_check_expire(unsigned long data)
314 {
315         struct ip_vs_lblc_table *tbl;
316         unsigned long now = jiffies;
317         int goal;
318         int i, j;
319         struct ip_vs_lblc_entry *en, *nxt;
320
321         tbl = (struct ip_vs_lblc_table *)data;
322
323         if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) {
324                 /* do full expiration check */
325                 ip_vs_lblc_full_check(tbl);
326                 tbl->counter = 1;
327                 goto out;
328         }
329
330         if (atomic_read(&tbl->entries) <= tbl->max_size) {
331                 tbl->counter++;
332                 goto out;
333         }
334
335         goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3;
336         if (goal > tbl->max_size/2)
337                 goal = tbl->max_size/2;
338
339         for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
340                 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
341
342                 write_lock(&tbl->lock);
343                 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
344                         if (time_before(now, en->lastuse + ENTRY_TIMEOUT))
345                                 continue;
346
347                         ip_vs_lblc_free(en);
348                         atomic_dec(&tbl->entries);
349                         goal--;
350                 }
351                 write_unlock(&tbl->lock);
352                 if (goal <= 0)
353                         break;
354         }
355         tbl->rover = j;
356
357   out:
358         mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL);
359 }
360
361
362 static int ip_vs_lblc_init_svc(struct ip_vs_service *svc)
363 {
364         int i;
365         struct ip_vs_lblc_table *tbl;
366
367         /*
368          *    Allocate the ip_vs_lblc_table for this service
369          */
370         tbl = kmalloc(sizeof(struct ip_vs_lblc_table), GFP_ATOMIC);
371         if (tbl == NULL) {
372                 IP_VS_ERR("ip_vs_lblc_init_svc(): no memory\n");
373                 return -ENOMEM;
374         }
375         svc->sched_data = tbl;
376         IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) allocated for "
377                   "current service\n",
378                   sizeof(struct ip_vs_lblc_table));
379
380         /*
381          *    Initialize the hash buckets
382          */
383         for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
384                 INIT_LIST_HEAD(&tbl->bucket[i]);
385         }
386         rwlock_init(&tbl->lock);
387         tbl->max_size = IP_VS_LBLC_TAB_SIZE*16;
388         tbl->rover = 0;
389         tbl->counter = 1;
390
391         /*
392          *    Hook periodic timer for garbage collection
393          */
394         init_timer(&tbl->periodic_timer);
395         tbl->periodic_timer.data = (unsigned long)tbl;
396         tbl->periodic_timer.function = ip_vs_lblc_check_expire;
397         tbl->periodic_timer.expires = jiffies+CHECK_EXPIRE_INTERVAL;
398         add_timer(&tbl->periodic_timer);
399
400         return 0;
401 }
402
403
404 static int ip_vs_lblc_done_svc(struct ip_vs_service *svc)
405 {
406         struct ip_vs_lblc_table *tbl = svc->sched_data;
407
408         /* remove periodic timer */
409         del_timer_sync(&tbl->periodic_timer);
410
411         /* got to clean up table entries here */
412         ip_vs_lblc_flush(tbl);
413
414         /* release the table itself */
415         kfree(svc->sched_data);
416         IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) released\n",
417                   sizeof(struct ip_vs_lblc_table));
418
419         return 0;
420 }
421
422
423 static int ip_vs_lblc_update_svc(struct ip_vs_service *svc)
424 {
425         return 0;
426 }
427
428
429 static inline struct ip_vs_dest *
430 __ip_vs_wlc_schedule(struct ip_vs_service *svc, struct iphdr *iph)
431 {
432         struct ip_vs_dest *dest, *least;
433         int loh, doh;
434
435         /*
436          * We think the overhead of processing active connections is fifty
437          * times higher than that of inactive connections in average. (This
438          * fifty times might not be accurate, we will change it later.) We
439          * use the following formula to estimate the overhead:
440          *                dest->activeconns*50 + dest->inactconns
441          * and the load:
442          *                (dest overhead) / dest->weight
443          *
444          * Remember -- no floats in kernel mode!!!
445          * The comparison of h1*w2 > h2*w1 is equivalent to that of
446          *                h1/w1 > h2/w2
447          * if every weight is larger than zero.
448          *
449          * The server with weight=0 is quiesced and will not receive any
450          * new connection.
451          */
452         list_for_each_entry(dest, &svc->destinations, n_list) {
453                 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
454                         continue;
455                 if (atomic_read(&dest->weight) > 0) {
456                         least = dest;
457                         loh = atomic_read(&least->activeconns) * 50
458                                 + atomic_read(&least->inactconns);
459                         goto nextstage;
460                 }
461         }
462         return NULL;
463
464         /*
465          *    Find the destination with the least load.
466          */
467   nextstage:
468         list_for_each_entry_continue(dest, &svc->destinations, n_list) {
469                 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
470                         continue;
471
472                 doh = atomic_read(&dest->activeconns) * 50
473                         + atomic_read(&dest->inactconns);
474                 if (loh * atomic_read(&dest->weight) >
475                     doh * atomic_read(&least->weight)) {
476                         least = dest;
477                         loh = doh;
478                 }
479         }
480
481         IP_VS_DBG(6, "LBLC: server %d.%d.%d.%d:%d "
482                   "activeconns %d refcnt %d weight %d overhead %d\n",
483                   NIPQUAD(least->addr), ntohs(least->port),
484                   atomic_read(&least->activeconns),
485                   atomic_read(&least->refcnt),
486                   atomic_read(&least->weight), loh);
487
488         return least;
489 }
490
491
492 /*
493  *   If this destination server is overloaded and there is a less loaded
494  *   server, then return true.
495  */
496 static inline int
497 is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc)
498 {
499         if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) {
500                 struct ip_vs_dest *d;
501
502                 list_for_each_entry(d, &svc->destinations, n_list) {
503                         if (atomic_read(&d->activeconns)*2
504                             < atomic_read(&d->weight)) {
505                                 return 1;
506                         }
507                 }
508         }
509         return 0;
510 }
511
512
513 /*
514  *    Locality-Based (weighted) Least-Connection scheduling
515  */
516 static struct ip_vs_dest *
517 ip_vs_lblc_schedule(struct ip_vs_service *svc, const struct sk_buff *skb)
518 {
519         struct ip_vs_dest *dest;
520         struct ip_vs_lblc_table *tbl;
521         struct ip_vs_lblc_entry *en;
522         struct iphdr *iph = ip_hdr(skb);
523
524         IP_VS_DBG(6, "ip_vs_lblc_schedule(): Scheduling...\n");
525
526         tbl = (struct ip_vs_lblc_table *)svc->sched_data;
527         en = ip_vs_lblc_get(tbl, iph->daddr);
528         if (en == NULL) {
529                 dest = __ip_vs_wlc_schedule(svc, iph);
530                 if (dest == NULL) {
531                         IP_VS_DBG(1, "no destination available\n");
532                         return NULL;
533                 }
534                 en = ip_vs_lblc_new(iph->daddr, dest);
535                 if (en == NULL) {
536                         return NULL;
537                 }
538                 ip_vs_lblc_hash(tbl, en);
539         } else {
540                 dest = en->dest;
541                 if (!(dest->flags & IP_VS_DEST_F_AVAILABLE)
542                     || atomic_read(&dest->weight) <= 0
543                     || is_overloaded(dest, svc)) {
544                         dest = __ip_vs_wlc_schedule(svc, iph);
545                         if (dest == NULL) {
546                                 IP_VS_DBG(1, "no destination available\n");
547                                 return NULL;
548                         }
549                         atomic_dec(&en->dest->refcnt);
550                         atomic_inc(&dest->refcnt);
551                         en->dest = dest;
552                 }
553         }
554         en->lastuse = jiffies;
555
556         IP_VS_DBG(6, "LBLC: destination IP address %u.%u.%u.%u "
557                   "--> server %u.%u.%u.%u:%d\n",
558                   NIPQUAD(en->addr),
559                   NIPQUAD(dest->addr),
560                   ntohs(dest->port));
561
562         return dest;
563 }
564
565
566 /*
567  *      IPVS LBLC Scheduler structure
568  */
569 static struct ip_vs_scheduler ip_vs_lblc_scheduler =
570 {
571         .name =                 "lblc",
572         .refcnt =               ATOMIC_INIT(0),
573         .module =               THIS_MODULE,
574         .init_service =         ip_vs_lblc_init_svc,
575         .done_service =         ip_vs_lblc_done_svc,
576         .update_service =       ip_vs_lblc_update_svc,
577         .schedule =             ip_vs_lblc_schedule,
578 };
579
580
581 static int __init ip_vs_lblc_init(void)
582 {
583         int ret;
584
585         INIT_LIST_HEAD(&ip_vs_lblc_scheduler.n_list);
586         sysctl_header = register_sysctl_table(lblc_root_table);
587         ret = register_ip_vs_scheduler(&ip_vs_lblc_scheduler);
588         if (ret)
589                 unregister_sysctl_table(sysctl_header);
590         return ret;
591 }
592
593
594 static void __exit ip_vs_lblc_cleanup(void)
595 {
596         unregister_sysctl_table(sysctl_header);
597         unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler);
598 }
599
600
601 module_init(ip_vs_lblc_init);
602 module_exit(ip_vs_lblc_cleanup);
603 MODULE_LICENSE("GPL");