blob: 524751e031de5203c12735d0aaa3c588dce675cd [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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
Arnaldo Carvalho de Melo14c85022005-12-27 02:43:12 -020044#include <linux/ip.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070045#include <linux/module.h>
46#include <linux/kernel.h>
Arnaldo Carvalho de Melo14c85022005-12-27 02:43:12 -020047#include <linux/skbuff.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070048
49/* for sysctl */
50#include <linux/fs.h>
51#include <linux/sysctl.h>
52
53#include <net/ip_vs.h>
54
55
56/*
57 * It is for garbage collection of stale IPVS lblc entries,
58 * when the table is full.
59 */
60#define CHECK_EXPIRE_INTERVAL (60*HZ)
61#define ENTRY_TIMEOUT (6*60*HZ)
62
63/*
64 * It is for full expiration check.
65 * When there is no partial expiration check (garbage collection)
66 * in a half hour, do a full expiration check to collect stale
67 * entries that haven't been touched for a day.
68 */
69#define COUNT_FOR_FULL_EXPIRATION 30
70static int sysctl_ip_vs_lblc_expiration = 24*60*60*HZ;
71
72
73/*
74 * for IPVS lblc entry hash table
75 */
76#ifndef CONFIG_IP_VS_LBLC_TAB_BITS
77#define CONFIG_IP_VS_LBLC_TAB_BITS 10
78#endif
79#define IP_VS_LBLC_TAB_BITS CONFIG_IP_VS_LBLC_TAB_BITS
80#define IP_VS_LBLC_TAB_SIZE (1 << IP_VS_LBLC_TAB_BITS)
81#define IP_VS_LBLC_TAB_MASK (IP_VS_LBLC_TAB_SIZE - 1)
82
83
84/*
85 * IPVS lblc entry represents an association between destination
86 * IP address and its destination server
87 */
88struct ip_vs_lblc_entry {
89 struct list_head list;
Al Viro014d7302006-09-28 14:29:52 -070090 __be32 addr; /* destination IP address */
Linus Torvalds1da177e2005-04-16 15:20:36 -070091 struct ip_vs_dest *dest; /* real server (cache) */
92 unsigned long lastuse; /* last used time */
93};
94
95
96/*
97 * IPVS lblc hash table
98 */
99struct ip_vs_lblc_table {
100 rwlock_t lock; /* lock for this table */
101 struct list_head bucket[IP_VS_LBLC_TAB_SIZE]; /* hash bucket */
102 atomic_t entries; /* number of entries */
103 int max_size; /* maximum size of entries */
104 struct timer_list periodic_timer; /* collect stale entries */
105 int rover; /* rover for expire check */
106 int counter; /* counter for no expire */
107};
108
109
110/*
111 * IPVS LBLC sysctl table
112 */
113
114static ctl_table vs_vars_table[] = {
115 {
116 .ctl_name = NET_IPV4_VS_LBLC_EXPIRE,
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
126static ctl_table vs_table[] = {
127 {
128 .ctl_name = NET_IPV4_VS,
129 .procname = "vs",
130 .mode = 0555,
131 .child = vs_vars_table
132 },
133 { .ctl_name = 0 }
134};
135
David S. Millerbf0ff9e2005-08-19 16:37:30 -0700136static ctl_table ipvs_ipv4_table[] = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137 {
138 .ctl_name = NET_IPV4,
139 .procname = "ipv4",
140 .mode = 0555,
141 .child = vs_table
142 },
143 { .ctl_name = 0 }
144};
145
146static ctl_table lblc_root_table[] = {
147 {
148 .ctl_name = CTL_NET,
149 .procname = "net",
150 .mode = 0555,
David S. Millerbf0ff9e2005-08-19 16:37:30 -0700151 .child = ipvs_ipv4_table
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 },
153 { .ctl_name = 0 }
154};
155
156static struct ctl_table_header * sysctl_header;
157
158/*
159 * new/free a ip_vs_lblc_entry, which is a mapping of a destionation
160 * IP address to a server.
161 */
162static inline struct ip_vs_lblc_entry *
Al Viro014d7302006-09-28 14:29:52 -0700163ip_vs_lblc_new(__be32 daddr, struct ip_vs_dest *dest)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164{
165 struct ip_vs_lblc_entry *en;
166
167 en = kmalloc(sizeof(struct ip_vs_lblc_entry), GFP_ATOMIC);
168 if (en == NULL) {
169 IP_VS_ERR("ip_vs_lblc_new(): no memory\n");
170 return NULL;
171 }
172
173 INIT_LIST_HEAD(&en->list);
174 en->addr = daddr;
175
176 atomic_inc(&dest->refcnt);
177 en->dest = dest;
178
179 return en;
180}
181
182
183static inline void ip_vs_lblc_free(struct ip_vs_lblc_entry *en)
184{
185 list_del(&en->list);
186 /*
187 * We don't kfree dest because it is refered either by its service
188 * or the trash dest list.
189 */
190 atomic_dec(&en->dest->refcnt);
191 kfree(en);
192}
193
194
195/*
196 * Returns hash value for IPVS LBLC entry
197 */
Al Viro014d7302006-09-28 14:29:52 -0700198static inline unsigned ip_vs_lblc_hashkey(__be32 addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199{
200 return (ntohl(addr)*2654435761UL) & IP_VS_LBLC_TAB_MASK;
201}
202
203
204/*
205 * Hash an entry in the ip_vs_lblc_table.
206 * returns bool success.
207 */
208static int
209ip_vs_lblc_hash(struct ip_vs_lblc_table *tbl, struct ip_vs_lblc_entry *en)
210{
211 unsigned hash;
212
213 if (!list_empty(&en->list)) {
214 IP_VS_ERR("ip_vs_lblc_hash(): request for already hashed, "
215 "called from %p\n", __builtin_return_address(0));
216 return 0;
217 }
218
219 /*
220 * Hash by destination IP address
221 */
222 hash = ip_vs_lblc_hashkey(en->addr);
223
224 write_lock(&tbl->lock);
225 list_add(&en->list, &tbl->bucket[hash]);
226 atomic_inc(&tbl->entries);
227 write_unlock(&tbl->lock);
228
229 return 1;
230}
231
232
Linus Torvalds1da177e2005-04-16 15:20:36 -0700233/*
234 * Get ip_vs_lblc_entry associated with supplied parameters.
235 */
236static inline struct ip_vs_lblc_entry *
Al Viro014d7302006-09-28 14:29:52 -0700237ip_vs_lblc_get(struct ip_vs_lblc_table *tbl, __be32 addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238{
239 unsigned hash;
240 struct ip_vs_lblc_entry *en;
241
242 hash = ip_vs_lblc_hashkey(addr);
243
244 read_lock(&tbl->lock);
245
246 list_for_each_entry(en, &tbl->bucket[hash], list) {
247 if (en->addr == addr) {
248 /* HIT */
249 read_unlock(&tbl->lock);
250 return en;
251 }
252 }
253
254 read_unlock(&tbl->lock);
255
256 return NULL;
257}
258
259
260/*
261 * Flush all the entries of the specified table.
262 */
263static void ip_vs_lblc_flush(struct ip_vs_lblc_table *tbl)
264{
265 int i;
266 struct ip_vs_lblc_entry *en, *nxt;
267
268 for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
269 write_lock(&tbl->lock);
270 list_for_each_entry_safe(en, nxt, &tbl->bucket[i], list) {
271 ip_vs_lblc_free(en);
272 atomic_dec(&tbl->entries);
273 }
274 write_unlock(&tbl->lock);
275 }
276}
277
278
279static inline void ip_vs_lblc_full_check(struct ip_vs_lblc_table *tbl)
280{
281 unsigned long now = jiffies;
282 int i, j;
283 struct ip_vs_lblc_entry *en, *nxt;
284
285 for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
286 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
287
288 write_lock(&tbl->lock);
289 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
290 if (time_before(now,
291 en->lastuse + sysctl_ip_vs_lblc_expiration))
292 continue;
293
294 ip_vs_lblc_free(en);
295 atomic_dec(&tbl->entries);
296 }
297 write_unlock(&tbl->lock);
298 }
299 tbl->rover = j;
300}
301
302
303/*
304 * Periodical timer handler for IPVS lblc table
305 * It is used to collect stale entries when the number of entries
306 * exceeds the maximum size of the table.
307 *
308 * Fixme: we probably need more complicated algorithm to collect
309 * entries that have not been used for a long time even
310 * if the number of entries doesn't exceed the maximum size
311 * of the table.
312 * The full expiration check is for this purpose now.
313 */
314static void ip_vs_lblc_check_expire(unsigned long data)
315{
316 struct ip_vs_lblc_table *tbl;
317 unsigned long now = jiffies;
318 int goal;
319 int i, j;
320 struct ip_vs_lblc_entry *en, *nxt;
321
322 tbl = (struct ip_vs_lblc_table *)data;
323
324 if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) {
325 /* do full expiration check */
326 ip_vs_lblc_full_check(tbl);
327 tbl->counter = 1;
328 goto out;
329 }
330
331 if (atomic_read(&tbl->entries) <= tbl->max_size) {
332 tbl->counter++;
333 goto out;
334 }
335
336 goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3;
337 if (goal > tbl->max_size/2)
338 goal = tbl->max_size/2;
339
340 for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) {
341 j = (j + 1) & IP_VS_LBLC_TAB_MASK;
342
343 write_lock(&tbl->lock);
344 list_for_each_entry_safe(en, nxt, &tbl->bucket[j], list) {
345 if (time_before(now, en->lastuse + ENTRY_TIMEOUT))
346 continue;
347
348 ip_vs_lblc_free(en);
349 atomic_dec(&tbl->entries);
350 goal--;
351 }
352 write_unlock(&tbl->lock);
353 if (goal <= 0)
354 break;
355 }
356 tbl->rover = j;
357
358 out:
359 mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL);
360}
361
362
363static int ip_vs_lblc_init_svc(struct ip_vs_service *svc)
364{
365 int i;
366 struct ip_vs_lblc_table *tbl;
367
368 /*
369 * Allocate the ip_vs_lblc_table for this service
370 */
371 tbl = kmalloc(sizeof(struct ip_vs_lblc_table), GFP_ATOMIC);
372 if (tbl == NULL) {
373 IP_VS_ERR("ip_vs_lblc_init_svc(): no memory\n");
374 return -ENOMEM;
375 }
376 svc->sched_data = tbl;
377 IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) allocated for "
378 "current service\n",
379 sizeof(struct ip_vs_lblc_table));
380
381 /*
382 * Initialize the hash buckets
383 */
384 for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) {
385 INIT_LIST_HEAD(&tbl->bucket[i]);
386 }
387 rwlock_init(&tbl->lock);
388 tbl->max_size = IP_VS_LBLC_TAB_SIZE*16;
389 tbl->rover = 0;
390 tbl->counter = 1;
391
392 /*
393 * Hook periodic timer for garbage collection
394 */
395 init_timer(&tbl->periodic_timer);
396 tbl->periodic_timer.data = (unsigned long)tbl;
397 tbl->periodic_timer.function = ip_vs_lblc_check_expire;
398 tbl->periodic_timer.expires = jiffies+CHECK_EXPIRE_INTERVAL;
399 add_timer(&tbl->periodic_timer);
400
401 return 0;
402}
403
404
405static int ip_vs_lblc_done_svc(struct ip_vs_service *svc)
406{
407 struct ip_vs_lblc_table *tbl = svc->sched_data;
408
409 /* remove periodic timer */
410 del_timer_sync(&tbl->periodic_timer);
411
412 /* got to clean up table entries here */
413 ip_vs_lblc_flush(tbl);
414
415 /* release the table itself */
416 kfree(svc->sched_data);
417 IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) released\n",
418 sizeof(struct ip_vs_lblc_table));
419
420 return 0;
421}
422
423
424static int ip_vs_lblc_update_svc(struct ip_vs_service *svc)
425{
426 return 0;
427}
428
429
430static inline struct ip_vs_dest *
431__ip_vs_wlc_schedule(struct ip_vs_service *svc, struct iphdr *iph)
432{
433 struct ip_vs_dest *dest, *least;
434 int loh, doh;
435
436 /*
437 * We think the overhead of processing active connections is fifty
438 * times higher than that of inactive connections in average. (This
439 * fifty times might not be accurate, we will change it later.) We
440 * use the following formula to estimate the overhead:
441 * dest->activeconns*50 + dest->inactconns
442 * and the load:
443 * (dest overhead) / dest->weight
444 *
445 * Remember -- no floats in kernel mode!!!
446 * The comparison of h1*w2 > h2*w1 is equivalent to that of
447 * h1/w1 > h2/w2
448 * if every weight is larger than zero.
449 *
450 * The server with weight=0 is quiesced and will not receive any
451 * new connection.
452 */
453 list_for_each_entry(dest, &svc->destinations, n_list) {
454 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
455 continue;
456 if (atomic_read(&dest->weight) > 0) {
457 least = dest;
458 loh = atomic_read(&least->activeconns) * 50
459 + atomic_read(&least->inactconns);
460 goto nextstage;
461 }
462 }
463 return NULL;
464
465 /*
466 * Find the destination with the least load.
467 */
468 nextstage:
469 list_for_each_entry_continue(dest, &svc->destinations, n_list) {
470 if (dest->flags & IP_VS_DEST_F_OVERLOAD)
471 continue;
472
473 doh = atomic_read(&dest->activeconns) * 50
474 + atomic_read(&dest->inactconns);
475 if (loh * atomic_read(&dest->weight) >
476 doh * atomic_read(&least->weight)) {
477 least = dest;
478 loh = doh;
479 }
480 }
481
482 IP_VS_DBG(6, "LBLC: server %d.%d.%d.%d:%d "
483 "activeconns %d refcnt %d weight %d overhead %d\n",
484 NIPQUAD(least->addr), ntohs(least->port),
485 atomic_read(&least->activeconns),
486 atomic_read(&least->refcnt),
487 atomic_read(&least->weight), loh);
488
489 return least;
490}
491
492
493/*
494 * If this destination server is overloaded and there is a less loaded
495 * server, then return true.
496 */
497static inline int
498is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc)
499{
500 if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) {
501 struct ip_vs_dest *d;
502
503 list_for_each_entry(d, &svc->destinations, n_list) {
504 if (atomic_read(&d->activeconns)*2
505 < atomic_read(&d->weight)) {
506 return 1;
507 }
508 }
509 }
510 return 0;
511}
512
513
514/*
515 * Locality-Based (weighted) Least-Connection scheduling
516 */
517static struct ip_vs_dest *
518ip_vs_lblc_schedule(struct ip_vs_service *svc, const struct sk_buff *skb)
519{
520 struct ip_vs_dest *dest;
521 struct ip_vs_lblc_table *tbl;
522 struct ip_vs_lblc_entry *en;
523 struct iphdr *iph = skb->nh.iph;
524
525 IP_VS_DBG(6, "ip_vs_lblc_schedule(): Scheduling...\n");
526
527 tbl = (struct ip_vs_lblc_table *)svc->sched_data;
528 en = ip_vs_lblc_get(tbl, iph->daddr);
529 if (en == NULL) {
530 dest = __ip_vs_wlc_schedule(svc, iph);
531 if (dest == NULL) {
532 IP_VS_DBG(1, "no destination available\n");
533 return NULL;
534 }
535 en = ip_vs_lblc_new(iph->daddr, dest);
536 if (en == NULL) {
537 return NULL;
538 }
539 ip_vs_lblc_hash(tbl, en);
540 } else {
541 dest = en->dest;
542 if (!(dest->flags & IP_VS_DEST_F_AVAILABLE)
543 || atomic_read(&dest->weight) <= 0
544 || is_overloaded(dest, svc)) {
545 dest = __ip_vs_wlc_schedule(svc, iph);
546 if (dest == NULL) {
547 IP_VS_DBG(1, "no destination available\n");
548 return NULL;
549 }
550 atomic_dec(&en->dest->refcnt);
551 atomic_inc(&dest->refcnt);
552 en->dest = dest;
553 }
554 }
555 en->lastuse = jiffies;
556
557 IP_VS_DBG(6, "LBLC: destination IP address %u.%u.%u.%u "
558 "--> server %u.%u.%u.%u:%d\n",
559 NIPQUAD(en->addr),
560 NIPQUAD(dest->addr),
561 ntohs(dest->port));
562
563 return dest;
564}
565
566
567/*
568 * IPVS LBLC Scheduler structure
569 */
570static struct ip_vs_scheduler ip_vs_lblc_scheduler =
571{
572 .name = "lblc",
573 .refcnt = ATOMIC_INIT(0),
574 .module = THIS_MODULE,
575 .init_service = ip_vs_lblc_init_svc,
576 .done_service = ip_vs_lblc_done_svc,
577 .update_service = ip_vs_lblc_update_svc,
578 .schedule = ip_vs_lblc_schedule,
579};
580
581
582static int __init ip_vs_lblc_init(void)
583{
584 INIT_LIST_HEAD(&ip_vs_lblc_scheduler.n_list);
585 sysctl_header = register_sysctl_table(lblc_root_table, 0);
586 return register_ip_vs_scheduler(&ip_vs_lblc_scheduler);
587}
588
589
590static void __exit ip_vs_lblc_cleanup(void)
591{
592 unregister_sysctl_table(sysctl_header);
593 unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler);
594}
595
596
597module_init(ip_vs_lblc_init);
598module_exit(ip_vs_lblc_cleanup);
599MODULE_LICENSE("GPL");