Merge branch 'master' of /home/cbou/linux-2.6
[linux-2.6] / net / ipv4 / ipvs / ip_vs_lblc.c
1 /*
2  * IPVS:        Locality-Based Least-Connection scheduling module
3  *
4  * Authors:     Wensong Zhang <wensong@gnuchina.org>
5  *
6  *              This program is free software; you can redistribute it and/or
7  *              modify it under the terms of the GNU General Public License
8  *              as published by the Free Software Foundation; either version
9  *              2 of the License, or (at your option) any later version.
10  *
11  * Changes:
12  *     Martin Hamilton         :    fixed the terrible locking bugs
13  *                                   *lock(tbl->lock) ==> *lock(&tbl->lock)
14  *     Wensong Zhang           :    fixed the uninitilized tbl->lock bug
15  *     Wensong Zhang           :    added doing full expiration check to
16  *                                   collect stale entries of 24+ hours when
17  *                                   no partial expire check in a half hour
18  *     Julian Anastasov        :    replaced del_timer call with del_timer_sync
19  *                                   to avoid the possible race between timer
20  *                                   handler and del_timer thread in SMP
21  *
22  */
23
24 /*
25  * The lblc algorithm is as follows (pseudo code):
26  *
27  *       if cachenode[dest_ip] is null then
28  *               n, cachenode[dest_ip] <- {weighted least-conn node};
29  *       else
30  *               n <- cachenode[dest_ip];
31  *               if (n is dead) OR
32  *                  (n.conns>n.weight AND
33  *                   there is a node m with m.conns<m.weight/2) then
34  *                 n, cachenode[dest_ip] <- {weighted least-conn node};
35  *
36  *       return n;
37  *
38  * Thanks must go to Wenzhuo Zhang for talking WCCP to me and pushing
39  * me to write this module.
40  */
41
42 #include <linux/ip.h>
43 #include <linux/module.h>
44 #include <linux/kernel.h>
45 #include <linux/skbuff.h>
46 #include <linux/jiffies.h>
47
48 /* for sysctl */
49 #include <linux/fs.h>
50 #include <linux/sysctl.h>
51
52 #include <net/ip_vs.h>
53
54
55 /*
56  *    It is for garbage collection of stale IPVS lblc entries,
57  *    when the table is full.
58  */
59 #define CHECK_EXPIRE_INTERVAL   (60*HZ)
60 #define ENTRY_TIMEOUT           (6*60*HZ)
61
62 /*
63  *    It is for full expiration check.
64  *    When there is no partial expiration check (garbage collection)
65  *    in a half hour, do a full expiration check to collect stale
66  *    entries that haven't been touched for a day.
67  */
68 #define COUNT_FOR_FULL_EXPIRATION   30
69 static int sysctl_ip_vs_lblc_expiration = 24*60*60*HZ;
70
71
72 /*
73  *     for IPVS lblc entry hash table
74  */
75 #ifndef CONFIG_IP_VS_LBLC_TAB_BITS
76 #define CONFIG_IP_VS_LBLC_TAB_BITS      10
77 #endif
78 #define IP_VS_LBLC_TAB_BITS     CONFIG_IP_VS_LBLC_TAB_BITS
79 #define IP_VS_LBLC_TAB_SIZE     (1 << IP_VS_LBLC_TAB_BITS)
80 #define IP_VS_LBLC_TAB_MASK     (IP_VS_LBLC_TAB_SIZE - 1)
81
82
83 /*
84  *      IPVS lblc entry represents an association between destination
85  *      IP address and its destination server
86  */
87 struct ip_vs_lblc_entry {
88         struct list_head        list;
89         __be32                  addr;           /* destination IP address */
90         struct ip_vs_dest       *dest;          /* real server (cache) */
91         unsigned long           lastuse;        /* last used time */
92 };
93
94
95 /*
96  *      IPVS lblc hash table
97  */
98 struct ip_vs_lblc_table {
99         rwlock_t                lock;           /* lock for this table */
100         struct list_head        bucket[IP_VS_LBLC_TAB_SIZE];  /* hash bucket */
101         atomic_t                entries;        /* number of entries */
102         int                     max_size;       /* maximum size of entries */
103         struct timer_list       periodic_timer; /* collect stale entries */
104         int                     rover;          /* rover for expire check */
105         int                     counter;        /* counter for no expire */
106 };
107
108
109 /*
110  *      IPVS LBLC sysctl table
111  */
112
113 static ctl_table vs_vars_table[] = {
114         {
115                 .procname       = "lblc_expiration",
116                 .data           = &sysctl_ip_vs_lblc_expiration,
117                 .maxlen         = sizeof(int),
118                 .mode           = 0644,
119                 .proc_handler   = &proc_dointvec_jiffies,
120         },
121         { .ctl_name = 0 }
122 };
123
124 static struct ctl_table_header * sysctl_header;
125
126 /*
127  *      new/free a ip_vs_lblc_entry, which is a mapping of a destionation
128  *      IP address to a server.
129  */
130 static inline struct ip_vs_lblc_entry *
131 ip_vs_lblc_new(__be32 daddr, struct ip_vs_dest *dest)
132 {
133         struct ip_vs_lblc_entry *en;
134
135         en = kmalloc(sizeof(struct ip_vs_lblc_entry), GFP_ATOMIC);
136         if (en == NULL) {
137                 IP_VS_ERR("ip_vs_lblc_new(): no memory\n");
138                 return NULL;
139         }
140
141         INIT_LIST_HEAD(&en->list);
142         en->addr = daddr;
143
144         atomic_inc(&dest->refcnt);
145         en->dest = dest;
146
147         return en;
148 }
149
150
151 static inline void ip_vs_lblc_free(struct ip_vs_lblc_entry *en)
152 {
153         list_del(&en->list);
154         /*
155          * We don't kfree dest because it is refered either by its service
156          * or the trash dest list.
157          */
158         atomic_dec(&en->dest->refcnt);
159         kfree(en);
160 }
161
162
163 /*
164  *      Returns hash value for IPVS LBLC entry
165  */
166 static inline unsigned ip_vs_lblc_hashkey(__be32 addr)
167 {
168         return (ntohl(addr)*2654435761UL) & IP_VS_LBLC_TAB_MASK;
169 }
170
171
172 /*
173  *      Hash an entry in the ip_vs_lblc_table.
174  *      returns bool success.
175  */
176 static int
177 ip_vs_lblc_hash(struct ip_vs_lblc_table *tbl, struct ip_vs_lblc_entry *en)
178 {
179         unsigned hash;
180
181         if (!list_empty(&en->list)) {
182                 IP_VS_ERR("ip_vs_lblc_hash(): request for already hashed, "
183                           "called from %p\n", __builtin_return_address(0));
184                 return 0;
185         }
186
187         /*
188          *      Hash by destination IP address
189          */
190         hash = ip_vs_lblc_hashkey(en->addr);
191
192         write_lock(&tbl->lock);
193         list_add(&en->list, &tbl->bucket[hash]);
194         atomic_inc(&tbl->entries);
195         write_unlock(&tbl->lock);
196
197         return 1;
198 }
199
200
201 /*
202  *  Get ip_vs_lblc_entry associated with supplied parameters.
203  */
204 static inline struct ip_vs_lblc_entry *
205 ip_vs_lblc_get(struct ip_vs_lblc_table *tbl, __be32 addr)
206 {
207         unsigned hash;
208         struct ip_vs_lblc_entry *en;
209
210         hash = ip_vs_lblc_hashkey(addr);
211
212         read_lock(&tbl->lock);
213
214         list_for_each_entry(en, &tbl->bucket[hash], list) {
215                 if (en->addr == addr) {
216                         /* HIT */
217                         read_unlock(&tbl->lock);
218                         return en;
219                 }
220         }
221
222         read_unlock(&tbl->lock);
223
224         return NULL;
225 }
226
227
228 /*
229  *      Flush all the entries of the specified table.
230  */
231 static void ip_vs_lblc_flush(struct ip_vs_lblc_table *tbl)
232 {
233         int i;
234         struct ip_vs_lblc_entry *en, *nxt;
235
236         for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
237                 write_lock(&tbl->lock);
238                 list_for_each_entry_safe(en, nxt, &tbl->bucket[i], list) {
239                         ip_vs_lblc_free(en);
240                         atomic_dec(&tbl->entries);
241                 }
242                 write_unlock(&tbl->lock);
243         }
244 }
245
246
247 static inline void ip_vs_lblc_full_check(struct ip_vs_lblc_table *tbl)
248 {
249         unsigned long now = jiffies;
250         int i, j;
251         struct ip_vs_lblc_entry *en, *nxt;
252
253         for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
254                 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
255
256                 write_lock(&tbl->lock);
257                 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
258                         if (time_before(now,
259                                         en->lastuse + sysctl_ip_vs_lblc_expiration))
260                                 continue;
261
262                         ip_vs_lblc_free(en);
263                         atomic_dec(&tbl->entries);
264                 }
265                 write_unlock(&tbl->lock);
266         }
267         tbl->rover = j;
268 }
269
270
271 /*
272  *      Periodical timer handler for IPVS lblc table
273  *      It is used to collect stale entries when the number of entries
274  *      exceeds the maximum size of the table.
275  *
276  *      Fixme: we probably need more complicated algorithm to collect
277  *             entries that have not been used for a long time even
278  *             if the number of entries doesn't exceed the maximum size
279  *             of the table.
280  *      The full expiration check is for this purpose now.
281  */
282 static void ip_vs_lblc_check_expire(unsigned long data)
283 {
284         struct ip_vs_lblc_table *tbl;
285         unsigned long now = jiffies;
286         int goal;
287         int i, j;
288         struct ip_vs_lblc_entry *en, *nxt;
289
290         tbl = (struct ip_vs_lblc_table *)data;
291
292         if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) {
293                 /* do full expiration check */
294                 ip_vs_lblc_full_check(tbl);
295                 tbl->counter = 1;
296                 goto out;
297         }
298
299         if (atomic_read(&tbl->entries) <= tbl->max_size) {
300                 tbl->counter++;
301                 goto out;
302         }
303
304         goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3;
305         if (goal > tbl->max_size/2)
306                 goal = tbl->max_size/2;
307
308         for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
309                 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
310
311                 write_lock(&tbl->lock);
312                 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
313                         if (time_before(now, en->lastuse + ENTRY_TIMEOUT))
314                                 continue;
315
316                         ip_vs_lblc_free(en);
317                         atomic_dec(&tbl->entries);
318                         goal--;
319                 }
320                 write_unlock(&tbl->lock);
321                 if (goal <= 0)
322                         break;
323         }
324         tbl->rover = j;
325
326   out:
327         mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL);
328 }
329
330
331 static int ip_vs_lblc_init_svc(struct ip_vs_service *svc)
332 {
333         int i;
334         struct ip_vs_lblc_table *tbl;
335
336         /*
337          *    Allocate the ip_vs_lblc_table for this service
338          */
339         tbl = kmalloc(sizeof(struct ip_vs_lblc_table), GFP_ATOMIC);
340         if (tbl == NULL) {
341                 IP_VS_ERR("ip_vs_lblc_init_svc(): no memory\n");
342                 return -ENOMEM;
343         }
344         svc->sched_data = tbl;
345         IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) allocated for "
346                   "current service\n",
347                   sizeof(struct ip_vs_lblc_table));
348
349         /*
350          *    Initialize the hash buckets
351          */
352         for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
353                 INIT_LIST_HEAD(&tbl->bucket[i]);
354         }
355         rwlock_init(&tbl->lock);
356         tbl->max_size = IP_VS_LBLC_TAB_SIZE*16;
357         tbl->rover = 0;
358         tbl->counter = 1;
359
360         /*
361          *    Hook periodic timer for garbage collection
362          */
363         setup_timer(&tbl->periodic_timer, ip_vs_lblc_check_expire,
364                         (unsigned long)tbl);
365         tbl->periodic_timer.expires = jiffies+CHECK_EXPIRE_INTERVAL;
366         add_timer(&tbl->periodic_timer);
367
368         return 0;
369 }
370
371
372 static int ip_vs_lblc_done_svc(struct ip_vs_service *svc)
373 {
374         struct ip_vs_lblc_table *tbl = svc->sched_data;
375
376         /* remove periodic timer */
377         del_timer_sync(&tbl->periodic_timer);
378
379         /* got to clean up table entries here */
380         ip_vs_lblc_flush(tbl);
381
382         /* release the table itself */
383         kfree(svc->sched_data);
384         IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) released\n",
385                   sizeof(struct ip_vs_lblc_table));
386
387         return 0;
388 }
389
390
391 static int ip_vs_lblc_update_svc(struct ip_vs_service *svc)
392 {
393         return 0;
394 }
395
396
397 static inline struct ip_vs_dest *
398 __ip_vs_wlc_schedule(struct ip_vs_service *svc, struct iphdr *iph)
399 {
400         struct ip_vs_dest *dest, *least;
401         int loh, doh;
402
403         /*
404          * We think the overhead of processing active connections is fifty
405          * times higher than that of inactive connections in average. (This
406          * fifty times might not be accurate, we will change it later.) We
407          * use the following formula to estimate the overhead:
408          *                dest->activeconns*50 + dest->inactconns
409          * and the load:
410          *                (dest overhead) / dest->weight
411          *
412          * Remember -- no floats in kernel mode!!!
413          * The comparison of h1*w2 > h2*w1 is equivalent to that of
414          *                h1/w1 > h2/w2
415          * if every weight is larger than zero.
416          *
417          * The server with weight=0 is quiesced and will not receive any
418          * new connection.
419          */
420         list_for_each_entry(dest, &svc->destinations, n_list) {
421                 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
422                         continue;
423                 if (atomic_read(&dest->weight) > 0) {
424                         least = dest;
425                         loh = atomic_read(&least->activeconns) * 50
426                                 + atomic_read(&least->inactconns);
427                         goto nextstage;
428                 }
429         }
430         return NULL;
431
432         /*
433          *    Find the destination with the least load.
434          */
435   nextstage:
436         list_for_each_entry_continue(dest, &svc->destinations, n_list) {
437                 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
438                         continue;
439
440                 doh = atomic_read(&dest->activeconns) * 50
441                         + atomic_read(&dest->inactconns);
442                 if (loh * atomic_read(&dest->weight) >
443                     doh * atomic_read(&least->weight)) {
444                         least = dest;
445                         loh = doh;
446                 }
447         }
448
449         IP_VS_DBG(6, "LBLC: server %d.%d.%d.%d:%d "
450                   "activeconns %d refcnt %d weight %d overhead %d\n",
451                   NIPQUAD(least->addr), ntohs(least->port),
452                   atomic_read(&least->activeconns),
453                   atomic_read(&least->refcnt),
454                   atomic_read(&least->weight), loh);
455
456         return least;
457 }
458
459
460 /*
461  *   If this destination server is overloaded and there is a less loaded
462  *   server, then return true.
463  */
464 static inline int
465 is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc)
466 {
467         if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) {
468                 struct ip_vs_dest *d;
469
470                 list_for_each_entry(d, &svc->destinations, n_list) {
471                         if (atomic_read(&d->activeconns)*2
472                             < atomic_read(&d->weight)) {
473                                 return 1;
474                         }
475                 }
476         }
477         return 0;
478 }
479
480
481 /*
482  *    Locality-Based (weighted) Least-Connection scheduling
483  */
484 static struct ip_vs_dest *
485 ip_vs_lblc_schedule(struct ip_vs_service *svc, const struct sk_buff *skb)
486 {
487         struct ip_vs_dest *dest;
488         struct ip_vs_lblc_table *tbl;
489         struct ip_vs_lblc_entry *en;
490         struct iphdr *iph = ip_hdr(skb);
491
492         IP_VS_DBG(6, "ip_vs_lblc_schedule(): Scheduling...\n");
493
494         tbl = (struct ip_vs_lblc_table *)svc->sched_data;
495         en = ip_vs_lblc_get(tbl, iph->daddr);
496         if (en == NULL) {
497                 dest = __ip_vs_wlc_schedule(svc, iph);
498                 if (dest == NULL) {
499                         IP_VS_DBG(1, "no destination available\n");
500                         return NULL;
501                 }
502                 en = ip_vs_lblc_new(iph->daddr, dest);
503                 if (en == NULL) {
504                         return NULL;
505                 }
506                 ip_vs_lblc_hash(tbl, en);
507         } else {
508                 dest = en->dest;
509                 if (!(dest->flags & IP_VS_DEST_F_AVAILABLE)
510                     || atomic_read(&dest->weight) <= 0
511                     || is_overloaded(dest, svc)) {
512                         dest = __ip_vs_wlc_schedule(svc, iph);
513                         if (dest == NULL) {
514                                 IP_VS_DBG(1, "no destination available\n");
515                                 return NULL;
516                         }
517                         atomic_dec(&en->dest->refcnt);
518                         atomic_inc(&dest->refcnt);
519                         en->dest = dest;
520                 }
521         }
522         en->lastuse = jiffies;
523
524         IP_VS_DBG(6, "LBLC: destination IP address %u.%u.%u.%u "
525                   "--> server %u.%u.%u.%u:%d\n",
526                   NIPQUAD(en->addr),
527                   NIPQUAD(dest->addr),
528                   ntohs(dest->port));
529
530         return dest;
531 }
532
533
534 /*
535  *      IPVS LBLC Scheduler structure
536  */
537 static struct ip_vs_scheduler ip_vs_lblc_scheduler =
538 {
539         .name =                 "lblc",
540         .refcnt =               ATOMIC_INIT(0),
541         .module =               THIS_MODULE,
542         .init_service =         ip_vs_lblc_init_svc,
543         .done_service =         ip_vs_lblc_done_svc,
544         .update_service =       ip_vs_lblc_update_svc,
545         .schedule =             ip_vs_lblc_schedule,
546 };
547
548
549 static int __init ip_vs_lblc_init(void)
550 {
551         int ret;
552
553         INIT_LIST_HEAD(&ip_vs_lblc_scheduler.n_list);
554         sysctl_header = register_sysctl_paths(net_vs_ctl_path, vs_vars_table);
555         ret = register_ip_vs_scheduler(&ip_vs_lblc_scheduler);
556         if (ret)
557                 unregister_sysctl_table(sysctl_header);
558         return ret;
559 }
560
561
562 static void __exit ip_vs_lblc_cleanup(void)
563 {
564         unregister_sysctl_table(sysctl_header);
565         unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler);
566 }
567
568
569 module_init(ip_vs_lblc_init);
570 module_exit(ip_vs_lblc_cleanup);
571 MODULE_LICENSE("GPL");