Merge master.kernel.org:/pub/scm/linux/kernel/git/dwmw2/audit-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/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
112 static ctl_table vs_vars_table[] = {
113         {
114                 .ctl_name       = NET_IPV4_VS_LBLC_EXPIRE,
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 ctl_table vs_table[] = {
125         {
126                 .ctl_name       = NET_IPV4_VS,
127                 .procname       = "vs",
128                 .mode           = 0555, 
129                 .child          = vs_vars_table
130         },
131         { .ctl_name = 0 }
132 };
133
134 static ctl_table ipvs_ipv4_table[] = {
135         {
136                 .ctl_name       = NET_IPV4,
137                 .procname       = "ipv4", 
138                 .mode           = 0555,
139                 .child          = vs_table
140         },
141         { .ctl_name = 0 }
142 };
143
144 static ctl_table lblc_root_table[] = {
145         {
146                 .ctl_name       = CTL_NET,
147                 .procname       = "net", 
148                 .mode           = 0555, 
149                 .child          = ipvs_ipv4_table
150         },
151         { .ctl_name = 0 }
152 };
153
154 static struct ctl_table_header * sysctl_header;
155
156 /*
157  *      new/free a ip_vs_lblc_entry, which is a mapping of a destionation
158  *      IP address to a server.
159  */
160 static inline struct ip_vs_lblc_entry *
161 ip_vs_lblc_new(__u32 daddr, struct ip_vs_dest *dest)
162 {
163         struct ip_vs_lblc_entry *en;
164
165         en = kmalloc(sizeof(struct ip_vs_lblc_entry), GFP_ATOMIC);
166         if (en == NULL) {
167                 IP_VS_ERR("ip_vs_lblc_new(): no memory\n");
168                 return NULL;
169         }
170
171         INIT_LIST_HEAD(&en->list);
172         en->addr = daddr;
173
174         atomic_inc(&dest->refcnt);
175         en->dest = dest;
176
177         return en;
178 }
179
180
181 static inline void ip_vs_lblc_free(struct ip_vs_lblc_entry *en)
182 {
183         list_del(&en->list);
184         /*
185          * We don't kfree dest because it is refered either by its service
186          * or the trash dest list.
187          */
188         atomic_dec(&en->dest->refcnt);
189         kfree(en);
190 }
191
192
193 /*
194  *      Returns hash value for IPVS LBLC entry
195  */
196 static inline unsigned ip_vs_lblc_hashkey(__u32 addr)
197 {
198         return (ntohl(addr)*2654435761UL) & IP_VS_LBLC_TAB_MASK;
199 }
200
201
202 /*
203  *      Hash an entry in the ip_vs_lblc_table.
204  *      returns bool success.
205  */
206 static int
207 ip_vs_lblc_hash(struct ip_vs_lblc_table *tbl, struct ip_vs_lblc_entry *en)
208 {
209         unsigned hash;
210
211         if (!list_empty(&en->list)) {
212                 IP_VS_ERR("ip_vs_lblc_hash(): request for already hashed, "
213                           "called from %p\n", __builtin_return_address(0));
214                 return 0;
215         }
216
217         /*
218          *      Hash by destination IP address
219          */
220         hash = ip_vs_lblc_hashkey(en->addr);
221
222         write_lock(&tbl->lock);
223         list_add(&en->list, &tbl->bucket[hash]);
224         atomic_inc(&tbl->entries);
225         write_unlock(&tbl->lock);
226
227         return 1;
228 }
229
230
231 #if 0000
232 /*
233  *      Unhash ip_vs_lblc_entry from ip_vs_lblc_table.
234  *      returns bool success.
235  */
236 static int ip_vs_lblc_unhash(struct ip_vs_lblc_table *tbl,
237                              struct ip_vs_lblc_entry *en)
238 {
239         if (list_empty(&en->list)) {
240                 IP_VS_ERR("ip_vs_lblc_unhash(): request for not hashed entry, "
241                           "called from %p\n", __builtin_return_address(0));
242                 return 0;
243         }
244
245         /*
246          * Remove it from the table
247          */
248         write_lock(&tbl->lock);
249         list_del(&en->list);
250         INIT_LIST_HEAD(&en->list);
251         write_unlock(&tbl->lock);
252
253         return 1;
254 }
255 #endif
256
257
258 /*
259  *  Get ip_vs_lblc_entry associated with supplied parameters.
260  */
261 static inline struct ip_vs_lblc_entry *
262 ip_vs_lblc_get(struct ip_vs_lblc_table *tbl, __u32 addr)
263 {
264         unsigned hash;
265         struct ip_vs_lblc_entry *en;
266
267         hash = ip_vs_lblc_hashkey(addr);
268
269         read_lock(&tbl->lock);
270
271         list_for_each_entry(en, &tbl->bucket[hash], list) {
272                 if (en->addr == addr) {
273                         /* HIT */
274                         read_unlock(&tbl->lock);
275                         return en;
276                 }
277         }
278
279         read_unlock(&tbl->lock);
280
281         return NULL;
282 }
283
284
285 /*
286  *      Flush all the entries of the specified table.
287  */
288 static void ip_vs_lblc_flush(struct ip_vs_lblc_table *tbl)
289 {
290         int i;
291         struct ip_vs_lblc_entry *en, *nxt;
292
293         for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
294                 write_lock(&tbl->lock);
295                 list_for_each_entry_safe(en, nxt, &tbl->bucket[i], list) {
296                         ip_vs_lblc_free(en);
297                         atomic_dec(&tbl->entries);
298                 }
299                 write_unlock(&tbl->lock);
300         }
301 }
302
303
304 static inline void ip_vs_lblc_full_check(struct ip_vs_lblc_table *tbl)
305 {
306         unsigned long now = jiffies;
307         int i, j;
308         struct ip_vs_lblc_entry *en, *nxt;
309
310         for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
311                 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
312
313                 write_lock(&tbl->lock);
314                 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
315                         if (time_before(now, 
316                                         en->lastuse + sysctl_ip_vs_lblc_expiration))
317                                 continue;
318
319                         ip_vs_lblc_free(en);
320                         atomic_dec(&tbl->entries);
321                 }
322                 write_unlock(&tbl->lock);
323         }
324         tbl->rover = j;
325 }
326
327
328 /*
329  *      Periodical timer handler for IPVS lblc table
330  *      It is used to collect stale entries when the number of entries
331  *      exceeds the maximum size of the table.
332  *
333  *      Fixme: we probably need more complicated algorithm to collect
334  *             entries that have not been used for a long time even
335  *             if the number of entries doesn't exceed the maximum size
336  *             of the table.
337  *      The full expiration check is for this purpose now.
338  */
339 static void ip_vs_lblc_check_expire(unsigned long data)
340 {
341         struct ip_vs_lblc_table *tbl;
342         unsigned long now = jiffies;
343         int goal;
344         int i, j;
345         struct ip_vs_lblc_entry *en, *nxt;
346
347         tbl = (struct ip_vs_lblc_table *)data;
348
349         if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) {
350                 /* do full expiration check */
351                 ip_vs_lblc_full_check(tbl);
352                 tbl->counter = 1;
353                 goto out;
354         }
355
356         if (atomic_read(&tbl->entries) <= tbl->max_size) {
357                 tbl->counter++;
358                 goto out;
359         }
360
361         goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3;
362         if (goal > tbl->max_size/2)
363                 goal = tbl->max_size/2;
364
365         for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
366                 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
367
368                 write_lock(&tbl->lock);
369                 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
370                         if (time_before(now, en->lastuse + ENTRY_TIMEOUT))
371                                 continue;
372
373                         ip_vs_lblc_free(en);
374                         atomic_dec(&tbl->entries);
375                         goal--;
376                 }
377                 write_unlock(&tbl->lock);
378                 if (goal <= 0)
379                         break;
380         }
381         tbl->rover = j;
382
383   out:
384         mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL);
385 }
386
387
388 static int ip_vs_lblc_init_svc(struct ip_vs_service *svc)
389 {
390         int i;
391         struct ip_vs_lblc_table *tbl;
392
393         /*
394          *    Allocate the ip_vs_lblc_table for this service
395          */
396         tbl = kmalloc(sizeof(struct ip_vs_lblc_table), GFP_ATOMIC);
397         if (tbl == NULL) {
398                 IP_VS_ERR("ip_vs_lblc_init_svc(): no memory\n");
399                 return -ENOMEM;
400         }
401         svc->sched_data = tbl;
402         IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) allocated for "
403                   "current service\n",
404                   sizeof(struct ip_vs_lblc_table));
405
406         /*
407          *    Initialize the hash buckets
408          */
409         for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
410                 INIT_LIST_HEAD(&tbl->bucket[i]);
411         }
412         rwlock_init(&tbl->lock);
413         tbl->max_size = IP_VS_LBLC_TAB_SIZE*16;
414         tbl->rover = 0;
415         tbl->counter = 1;
416
417         /*
418          *    Hook periodic timer for garbage collection
419          */
420         init_timer(&tbl->periodic_timer);
421         tbl->periodic_timer.data = (unsigned long)tbl;
422         tbl->periodic_timer.function = ip_vs_lblc_check_expire;
423         tbl->periodic_timer.expires = jiffies+CHECK_EXPIRE_INTERVAL;
424         add_timer(&tbl->periodic_timer);
425
426         return 0;
427 }
428
429
430 static int ip_vs_lblc_done_svc(struct ip_vs_service *svc)
431 {
432         struct ip_vs_lblc_table *tbl = svc->sched_data;
433
434         /* remove periodic timer */
435         del_timer_sync(&tbl->periodic_timer);
436
437         /* got to clean up table entries here */
438         ip_vs_lblc_flush(tbl);
439
440         /* release the table itself */
441         kfree(svc->sched_data);
442         IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) released\n",
443                   sizeof(struct ip_vs_lblc_table));
444
445         return 0;
446 }
447
448
449 static int ip_vs_lblc_update_svc(struct ip_vs_service *svc)
450 {
451         return 0;
452 }
453
454
455 static inline struct ip_vs_dest *
456 __ip_vs_wlc_schedule(struct ip_vs_service *svc, struct iphdr *iph)
457 {
458         struct ip_vs_dest *dest, *least;
459         int loh, doh;
460
461         /*
462          * We think the overhead of processing active connections is fifty
463          * times higher than that of inactive connections in average. (This
464          * fifty times might not be accurate, we will change it later.) We
465          * use the following formula to estimate the overhead:
466          *                dest->activeconns*50 + dest->inactconns
467          * and the load:
468          *                (dest overhead) / dest->weight
469          *
470          * Remember -- no floats in kernel mode!!!
471          * The comparison of h1*w2 > h2*w1 is equivalent to that of
472          *                h1/w1 > h2/w2
473          * if every weight is larger than zero.
474          *
475          * The server with weight=0 is quiesced and will not receive any
476          * new connection.
477          */
478         list_for_each_entry(dest, &svc->destinations, n_list) {
479                 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
480                         continue;
481                 if (atomic_read(&dest->weight) > 0) {
482                         least = dest;
483                         loh = atomic_read(&least->activeconns) * 50
484                                 + atomic_read(&least->inactconns);
485                         goto nextstage;
486                 }
487         }
488         return NULL;
489
490         /*
491          *    Find the destination with the least load.
492          */
493   nextstage:
494         list_for_each_entry_continue(dest, &svc->destinations, n_list) {
495                 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
496                         continue;
497
498                 doh = atomic_read(&dest->activeconns) * 50
499                         + atomic_read(&dest->inactconns);
500                 if (loh * atomic_read(&dest->weight) >
501                     doh * atomic_read(&least->weight)) {
502                         least = dest;
503                         loh = doh;
504                 }
505         }
506
507         IP_VS_DBG(6, "LBLC: server %d.%d.%d.%d:%d "
508                   "activeconns %d refcnt %d weight %d overhead %d\n",
509                   NIPQUAD(least->addr), ntohs(least->port),
510                   atomic_read(&least->activeconns),
511                   atomic_read(&least->refcnt),
512                   atomic_read(&least->weight), loh);
513
514         return least;
515 }
516
517
518 /*
519  *   If this destination server is overloaded and there is a less loaded
520  *   server, then return true.
521  */
522 static inline int
523 is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc)
524 {
525         if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) {
526                 struct ip_vs_dest *d;
527
528                 list_for_each_entry(d, &svc->destinations, n_list) {
529                         if (atomic_read(&d->activeconns)*2
530                             < atomic_read(&d->weight)) {
531                                 return 1;
532                         }
533                 }
534         }
535         return 0;
536 }
537
538
539 /*
540  *    Locality-Based (weighted) Least-Connection scheduling
541  */
542 static struct ip_vs_dest *
543 ip_vs_lblc_schedule(struct ip_vs_service *svc, const struct sk_buff *skb)
544 {
545         struct ip_vs_dest *dest;
546         struct ip_vs_lblc_table *tbl;
547         struct ip_vs_lblc_entry *en;
548         struct iphdr *iph = skb->nh.iph;
549
550         IP_VS_DBG(6, "ip_vs_lblc_schedule(): Scheduling...\n");
551
552         tbl = (struct ip_vs_lblc_table *)svc->sched_data;
553         en = ip_vs_lblc_get(tbl, iph->daddr);
554         if (en == NULL) {
555                 dest = __ip_vs_wlc_schedule(svc, iph);
556                 if (dest == NULL) {
557                         IP_VS_DBG(1, "no destination available\n");
558                         return NULL;
559                 }
560                 en = ip_vs_lblc_new(iph->daddr, dest);
561                 if (en == NULL) {
562                         return NULL;
563                 }
564                 ip_vs_lblc_hash(tbl, en);
565         } else {
566                 dest = en->dest;
567                 if (!(dest->flags & IP_VS_DEST_F_AVAILABLE)
568                     || atomic_read(&dest->weight) <= 0
569                     || is_overloaded(dest, svc)) {
570                         dest = __ip_vs_wlc_schedule(svc, iph);
571                         if (dest == NULL) {
572                                 IP_VS_DBG(1, "no destination available\n");
573                                 return NULL;
574                         }
575                         atomic_dec(&en->dest->refcnt);
576                         atomic_inc(&dest->refcnt);
577                         en->dest = dest;
578                 }
579         }
580         en->lastuse = jiffies;
581
582         IP_VS_DBG(6, "LBLC: destination IP address %u.%u.%u.%u "
583                   "--> server %u.%u.%u.%u:%d\n",
584                   NIPQUAD(en->addr),
585                   NIPQUAD(dest->addr),
586                   ntohs(dest->port));
587
588         return dest;
589 }
590
591
592 /*
593  *      IPVS LBLC Scheduler structure
594  */
595 static struct ip_vs_scheduler ip_vs_lblc_scheduler =
596 {
597         .name =                 "lblc",
598         .refcnt =               ATOMIC_INIT(0),
599         .module =               THIS_MODULE,
600         .init_service =         ip_vs_lblc_init_svc,
601         .done_service =         ip_vs_lblc_done_svc,
602         .update_service =       ip_vs_lblc_update_svc,
603         .schedule =             ip_vs_lblc_schedule,
604 };
605
606
607 static int __init ip_vs_lblc_init(void)
608 {
609         INIT_LIST_HEAD(&ip_vs_lblc_scheduler.n_list);
610         sysctl_header = register_sysctl_table(lblc_root_table, 0);
611         return register_ip_vs_scheduler(&ip_vs_lblc_scheduler);
612 }
613
614
615 static void __exit ip_vs_lblc_cleanup(void)
616 {
617         unregister_sysctl_table(sysctl_header);
618         unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler);
619 }
620
621
622 module_init(ip_vs_lblc_init);
623 module_exit(ip_vs_lblc_cleanup);
624 MODULE_LICENSE("GPL");