]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * IPVS: Locality-Based Least-Connection with Replication scheduler | |
3 | * | |
1da177e4 LT |
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 | * Julian Anastasov : Added the missing (dest->weight>0) | |
13 | * condition in the ip_vs_dest_set_max. | |
14 | * | |
15 | */ | |
16 | ||
17 | /* | |
18 | * The lblc/r algorithm is as follows (pseudo code): | |
19 | * | |
20 | * if serverSet[dest_ip] is null then | |
21 | * n, serverSet[dest_ip] <- {weighted least-conn node}; | |
22 | * else | |
23 | * n <- {least-conn (alive) node in serverSet[dest_ip]}; | |
24 | * if (n is null) OR | |
25 | * (n.conns>n.weight AND | |
26 | * there is a node m with m.conns<m.weight/2) then | |
27 | * n <- {weighted least-conn node}; | |
28 | * add n to serverSet[dest_ip]; | |
29 | * if |serverSet[dest_ip]| > 1 AND | |
30 | * now - serverSet[dest_ip].lastMod > T then | |
31 | * m <- {most conn node in serverSet[dest_ip]}; | |
32 | * remove m from serverSet[dest_ip]; | |
33 | * if serverSet[dest_ip] changed then | |
34 | * serverSet[dest_ip].lastMod <- now; | |
35 | * | |
36 | * return n; | |
37 | * | |
38 | */ | |
39 | ||
9aada7ac HE |
40 | #define KMSG_COMPONENT "IPVS" |
41 | #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt | |
42 | ||
14c85021 | 43 | #include <linux/ip.h> |
1da177e4 LT |
44 | #include <linux/module.h> |
45 | #include <linux/kernel.h> | |
14c85021 | 46 | #include <linux/skbuff.h> |
d7fe0f24 | 47 | #include <linux/jiffies.h> |
51f0bc78 | 48 | #include <linux/list.h> |
5a0e3ad6 | 49 | #include <linux/slab.h> |
1da177e4 LT |
50 | |
51 | /* for sysctl */ | |
52 | #include <linux/fs.h> | |
53 | #include <linux/sysctl.h> | |
457c4cbc | 54 | #include <net/net_namespace.h> |
1da177e4 LT |
55 | |
56 | #include <net/ip_vs.h> | |
57 | ||
58 | ||
59 | /* | |
60 | * It is for garbage collection of stale IPVS lblcr entries, | |
61 | * when the table is full. | |
62 | */ | |
63 | #define CHECK_EXPIRE_INTERVAL (60*HZ) | |
64 | #define ENTRY_TIMEOUT (6*60*HZ) | |
65 | ||
b27d777e SH |
66 | #define DEFAULT_EXPIRATION (24*60*60*HZ) |
67 | ||
1da177e4 LT |
68 | /* |
69 | * It is for full expiration check. | |
70 | * When there is no partial expiration check (garbage collection) | |
71 | * in a half hour, do a full expiration check to collect stale | |
72 | * entries that haven't been touched for a day. | |
73 | */ | |
74 | #define COUNT_FOR_FULL_EXPIRATION 30 | |
1da177e4 LT |
75 | |
76 | /* | |
77 | * for IPVS lblcr entry hash table | |
78 | */ | |
79 | #ifndef CONFIG_IP_VS_LBLCR_TAB_BITS | |
80 | #define CONFIG_IP_VS_LBLCR_TAB_BITS 10 | |
81 | #endif | |
82 | #define IP_VS_LBLCR_TAB_BITS CONFIG_IP_VS_LBLCR_TAB_BITS | |
83 | #define IP_VS_LBLCR_TAB_SIZE (1 << IP_VS_LBLCR_TAB_BITS) | |
84 | #define IP_VS_LBLCR_TAB_MASK (IP_VS_LBLCR_TAB_SIZE - 1) | |
85 | ||
86 | ||
87 | /* | |
88 | * IPVS destination set structure and operations | |
89 | */ | |
51f0bc78 SH |
90 | struct ip_vs_dest_set_elem { |
91 | struct list_head list; /* list link */ | |
c5549571 JA |
92 | struct ip_vs_dest __rcu *dest; /* destination server */ |
93 | struct rcu_head rcu_head; | |
1da177e4 LT |
94 | }; |
95 | ||
96 | struct ip_vs_dest_set { | |
97 | atomic_t size; /* set size */ | |
98 | unsigned long lastmod; /* last modified time */ | |
51f0bc78 | 99 | struct list_head list; /* destination list */ |
1da177e4 LT |
100 | }; |
101 | ||
102 | ||
c5549571 JA |
103 | static void ip_vs_dest_set_insert(struct ip_vs_dest_set *set, |
104 | struct ip_vs_dest *dest, bool check) | |
1da177e4 | 105 | { |
51f0bc78 | 106 | struct ip_vs_dest_set_elem *e; |
1da177e4 | 107 | |
c5549571 JA |
108 | if (check) { |
109 | list_for_each_entry(e, &set->list, list) { | |
110 | struct ip_vs_dest *d; | |
111 | ||
112 | d = rcu_dereference_protected(e->dest, 1); | |
113 | if (d == dest) | |
114 | /* already existed */ | |
115 | return; | |
116 | } | |
1da177e4 LT |
117 | } |
118 | ||
f728bafb | 119 | e = kmalloc(sizeof(*e), GFP_ATOMIC); |
0a9ee813 | 120 | if (e == NULL) |
c5549571 | 121 | return; |
1da177e4 | 122 | |
c5549571 JA |
123 | ip_vs_dest_hold(dest); |
124 | RCU_INIT_POINTER(e->dest, dest); | |
1da177e4 | 125 | |
c5549571 | 126 | list_add_rcu(&e->list, &set->list); |
1da177e4 | 127 | atomic_inc(&set->size); |
1da177e4 LT |
128 | |
129 | set->lastmod = jiffies; | |
1da177e4 LT |
130 | } |
131 | ||
132 | static void | |
133 | ip_vs_dest_set_erase(struct ip_vs_dest_set *set, struct ip_vs_dest *dest) | |
134 | { | |
51f0bc78 | 135 | struct ip_vs_dest_set_elem *e; |
1da177e4 | 136 | |
51f0bc78 | 137 | list_for_each_entry(e, &set->list, list) { |
c5549571 JA |
138 | struct ip_vs_dest *d; |
139 | ||
140 | d = rcu_dereference_protected(e->dest, 1); | |
141 | if (d == dest) { | |
1da177e4 | 142 | /* HIT */ |
1da177e4 LT |
143 | atomic_dec(&set->size); |
144 | set->lastmod = jiffies; | |
c5549571 JA |
145 | ip_vs_dest_put(dest); |
146 | list_del_rcu(&e->list); | |
147 | kfree_rcu(e, rcu_head); | |
1da177e4 LT |
148 | break; |
149 | } | |
1da177e4 | 150 | } |
1da177e4 LT |
151 | } |
152 | ||
153 | static void ip_vs_dest_set_eraseall(struct ip_vs_dest_set *set) | |
154 | { | |
51f0bc78 | 155 | struct ip_vs_dest_set_elem *e, *ep; |
1da177e4 | 156 | |
51f0bc78 | 157 | list_for_each_entry_safe(e, ep, &set->list, list) { |
c5549571 JA |
158 | struct ip_vs_dest *d; |
159 | ||
160 | d = rcu_dereference_protected(e->dest, 1); | |
1da177e4 | 161 | /* |
25985edc | 162 | * We don't kfree dest because it is referred either |
1da177e4 LT |
163 | * by its service or by the trash dest list. |
164 | */ | |
c5549571 JA |
165 | ip_vs_dest_put(d); |
166 | list_del_rcu(&e->list); | |
167 | kfree_rcu(e, rcu_head); | |
1da177e4 | 168 | } |
1da177e4 LT |
169 | } |
170 | ||
171 | /* get weighted least-connection node in the destination set */ | |
172 | static inline struct ip_vs_dest *ip_vs_dest_set_min(struct ip_vs_dest_set *set) | |
173 | { | |
51f0bc78 | 174 | register struct ip_vs_dest_set_elem *e; |
1da177e4 LT |
175 | struct ip_vs_dest *dest, *least; |
176 | int loh, doh; | |
177 | ||
178 | if (set == NULL) | |
179 | return NULL; | |
180 | ||
1da177e4 | 181 | /* select the first destination server, whose weight > 0 */ |
c5549571 JA |
182 | list_for_each_entry_rcu(e, &set->list, list) { |
183 | least = rcu_dereference(e->dest); | |
1da177e4 LT |
184 | if (least->flags & IP_VS_DEST_F_OVERLOAD) |
185 | continue; | |
186 | ||
187 | if ((atomic_read(&least->weight) > 0) | |
188 | && (least->flags & IP_VS_DEST_F_AVAILABLE)) { | |
b552f7e3 | 189 | loh = ip_vs_dest_conn_overhead(least); |
1da177e4 LT |
190 | goto nextstage; |
191 | } | |
192 | } | |
1da177e4 LT |
193 | return NULL; |
194 | ||
195 | /* find the destination with the weighted least load */ | |
196 | nextstage: | |
c5549571 JA |
197 | list_for_each_entry_continue_rcu(e, &set->list, list) { |
198 | dest = rcu_dereference(e->dest); | |
1da177e4 LT |
199 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) |
200 | continue; | |
201 | ||
b552f7e3 | 202 | doh = ip_vs_dest_conn_overhead(dest); |
1da177e4 LT |
203 | if ((loh * atomic_read(&dest->weight) > |
204 | doh * atomic_read(&least->weight)) | |
205 | && (dest->flags & IP_VS_DEST_F_AVAILABLE)) { | |
206 | least = dest; | |
207 | loh = doh; | |
208 | } | |
209 | } | |
1da177e4 | 210 | |
1e3e238e | 211 | IP_VS_DBG_BUF(6, "%s(): server %s:%d " |
44548375 | 212 | "activeconns %d refcnt %d weight %d overhead %d\n", |
1e3e238e | 213 | __func__, |
44548375 JV |
214 | IP_VS_DBG_ADDR(least->af, &least->addr), |
215 | ntohs(least->port), | |
216 | atomic_read(&least->activeconns), | |
217 | atomic_read(&least->refcnt), | |
218 | atomic_read(&least->weight), loh); | |
1da177e4 LT |
219 | return least; |
220 | } | |
221 | ||
222 | ||
223 | /* get weighted most-connection node in the destination set */ | |
224 | static inline struct ip_vs_dest *ip_vs_dest_set_max(struct ip_vs_dest_set *set) | |
225 | { | |
51f0bc78 | 226 | register struct ip_vs_dest_set_elem *e; |
1da177e4 LT |
227 | struct ip_vs_dest *dest, *most; |
228 | int moh, doh; | |
229 | ||
230 | if (set == NULL) | |
231 | return NULL; | |
232 | ||
1da177e4 | 233 | /* select the first destination server, whose weight > 0 */ |
51f0bc78 | 234 | list_for_each_entry(e, &set->list, list) { |
c5549571 | 235 | most = rcu_dereference_protected(e->dest, 1); |
1da177e4 | 236 | if (atomic_read(&most->weight) > 0) { |
b552f7e3 | 237 | moh = ip_vs_dest_conn_overhead(most); |
1da177e4 LT |
238 | goto nextstage; |
239 | } | |
240 | } | |
1da177e4 LT |
241 | return NULL; |
242 | ||
243 | /* find the destination with the weighted most load */ | |
244 | nextstage: | |
c5549571 JA |
245 | list_for_each_entry_continue(e, &set->list, list) { |
246 | dest = rcu_dereference_protected(e->dest, 1); | |
b552f7e3 | 247 | doh = ip_vs_dest_conn_overhead(dest); |
1da177e4 LT |
248 | /* moh/mw < doh/dw ==> moh*dw < doh*mw, where mw,dw>0 */ |
249 | if ((moh * atomic_read(&dest->weight) < | |
250 | doh * atomic_read(&most->weight)) | |
251 | && (atomic_read(&dest->weight) > 0)) { | |
252 | most = dest; | |
253 | moh = doh; | |
254 | } | |
255 | } | |
1da177e4 | 256 | |
1e3e238e | 257 | IP_VS_DBG_BUF(6, "%s(): server %s:%d " |
44548375 | 258 | "activeconns %d refcnt %d weight %d overhead %d\n", |
1e3e238e | 259 | __func__, |
44548375 JV |
260 | IP_VS_DBG_ADDR(most->af, &most->addr), ntohs(most->port), |
261 | atomic_read(&most->activeconns), | |
262 | atomic_read(&most->refcnt), | |
263 | atomic_read(&most->weight), moh); | |
1da177e4 LT |
264 | return most; |
265 | } | |
266 | ||
267 | ||
268 | /* | |
269 | * IPVS lblcr entry represents an association between destination | |
270 | * IP address and its destination server set | |
271 | */ | |
272 | struct ip_vs_lblcr_entry { | |
c5549571 | 273 | struct hlist_node list; |
44548375 JV |
274 | int af; /* address family */ |
275 | union nf_inet_addr addr; /* destination IP address */ | |
1da177e4 LT |
276 | struct ip_vs_dest_set set; /* destination server set */ |
277 | unsigned long lastuse; /* last used time */ | |
c5549571 | 278 | struct rcu_head rcu_head; |
1da177e4 LT |
279 | }; |
280 | ||
281 | ||
282 | /* | |
283 | * IPVS lblcr hash table | |
284 | */ | |
285 | struct ip_vs_lblcr_table { | |
c5549571 | 286 | struct rcu_head rcu_head; |
f33c8b94 | 287 | struct hlist_head bucket[IP_VS_LBLCR_TAB_SIZE]; /* hash bucket */ |
1da177e4 LT |
288 | atomic_t entries; /* number of entries */ |
289 | int max_size; /* maximum size of entries */ | |
290 | struct timer_list periodic_timer; /* collect stale entries */ | |
291 | int rover; /* rover for expire check */ | |
292 | int counter; /* counter for no expire */ | |
c5549571 | 293 | bool dead; |
1da177e4 LT |
294 | }; |
295 | ||
296 | ||
fb1de432 | 297 | #ifdef CONFIG_SYSCTL |
1da177e4 LT |
298 | /* |
299 | * IPVS LBLCR sysctl table | |
300 | */ | |
301 | ||
302 | static ctl_table vs_vars_table[] = { | |
303 | { | |
1da177e4 | 304 | .procname = "lblcr_expiration", |
d0a1eef9 | 305 | .data = NULL, |
1da177e4 | 306 | .maxlen = sizeof(int), |
e905a9ed | 307 | .mode = 0644, |
6d9f239a | 308 | .proc_handler = proc_dointvec_jiffies, |
1da177e4 | 309 | }, |
f8572d8f | 310 | { } |
1da177e4 | 311 | }; |
fb1de432 | 312 | #endif |
1da177e4 | 313 | |
1da177e4 LT |
314 | static inline void ip_vs_lblcr_free(struct ip_vs_lblcr_entry *en) |
315 | { | |
c5549571 | 316 | hlist_del_rcu(&en->list); |
1da177e4 | 317 | ip_vs_dest_set_eraseall(&en->set); |
c5549571 | 318 | kfree_rcu(en, rcu_head); |
1da177e4 LT |
319 | } |
320 | ||
321 | ||
322 | /* | |
323 | * Returns hash value for IPVS LBLCR entry | |
324 | */ | |
95c96174 | 325 | static inline unsigned int |
44548375 | 326 | ip_vs_lblcr_hashkey(int af, const union nf_inet_addr *addr) |
1da177e4 | 327 | { |
44548375 JV |
328 | __be32 addr_fold = addr->ip; |
329 | ||
330 | #ifdef CONFIG_IP_VS_IPV6 | |
331 | if (af == AF_INET6) | |
332 | addr_fold = addr->ip6[0]^addr->ip6[1]^ | |
333 | addr->ip6[2]^addr->ip6[3]; | |
334 | #endif | |
335 | return (ntohl(addr_fold)*2654435761UL) & IP_VS_LBLCR_TAB_MASK; | |
1da177e4 LT |
336 | } |
337 | ||
338 | ||
339 | /* | |
340 | * Hash an entry in the ip_vs_lblcr_table. | |
341 | * returns bool success. | |
342 | */ | |
f728bafb | 343 | static void |
1da177e4 LT |
344 | ip_vs_lblcr_hash(struct ip_vs_lblcr_table *tbl, struct ip_vs_lblcr_entry *en) |
345 | { | |
95c96174 | 346 | unsigned int hash = ip_vs_lblcr_hashkey(en->af, &en->addr); |
1da177e4 | 347 | |
c5549571 | 348 | hlist_add_head_rcu(&en->list, &tbl->bucket[hash]); |
1da177e4 | 349 | atomic_inc(&tbl->entries); |
1da177e4 LT |
350 | } |
351 | ||
352 | ||
c5549571 | 353 | /* Get ip_vs_lblcr_entry associated with supplied parameters. */ |
1da177e4 | 354 | static inline struct ip_vs_lblcr_entry * |
44548375 JV |
355 | ip_vs_lblcr_get(int af, struct ip_vs_lblcr_table *tbl, |
356 | const union nf_inet_addr *addr) | |
1da177e4 | 357 | { |
95c96174 | 358 | unsigned int hash = ip_vs_lblcr_hashkey(af, addr); |
1da177e4 LT |
359 | struct ip_vs_lblcr_entry *en; |
360 | ||
c5549571 | 361 | hlist_for_each_entry_rcu(en, &tbl->bucket[hash], list) |
44548375 | 362 | if (ip_vs_addr_equal(af, &en->addr, addr)) |
f728bafb SW |
363 | return en; |
364 | ||
365 | return NULL; | |
366 | } | |
1da177e4 | 367 | |
1da177e4 | 368 | |
f728bafb SW |
369 | /* |
370 | * Create or update an ip_vs_lblcr_entry, which is a mapping of a destination | |
ba3a3ce1 | 371 | * IP address to a server. Called under spin lock. |
f728bafb SW |
372 | */ |
373 | static inline struct ip_vs_lblcr_entry * | |
44548375 | 374 | ip_vs_lblcr_new(struct ip_vs_lblcr_table *tbl, const union nf_inet_addr *daddr, |
f728bafb SW |
375 | struct ip_vs_dest *dest) |
376 | { | |
377 | struct ip_vs_lblcr_entry *en; | |
378 | ||
44548375 | 379 | en = ip_vs_lblcr_get(dest->af, tbl, daddr); |
f728bafb SW |
380 | if (!en) { |
381 | en = kmalloc(sizeof(*en), GFP_ATOMIC); | |
0a9ee813 | 382 | if (!en) |
f728bafb | 383 | return NULL; |
f728bafb | 384 | |
44548375 JV |
385 | en->af = dest->af; |
386 | ip_vs_addr_copy(dest->af, &en->addr, daddr); | |
f728bafb SW |
387 | en->lastuse = jiffies; |
388 | ||
421f91d2 | 389 | /* initialize its dest set */ |
f728bafb | 390 | atomic_set(&(en->set.size), 0); |
51f0bc78 | 391 | INIT_LIST_HEAD(&en->set.list); |
c5549571 JA |
392 | |
393 | ip_vs_dest_set_insert(&en->set, dest, false); | |
f728bafb SW |
394 | |
395 | ip_vs_lblcr_hash(tbl, en); | |
c5549571 | 396 | return en; |
1da177e4 LT |
397 | } |
398 | ||
c5549571 | 399 | ip_vs_dest_set_insert(&en->set, dest, true); |
1da177e4 | 400 | |
f728bafb | 401 | return en; |
1da177e4 LT |
402 | } |
403 | ||
404 | ||
405 | /* | |
406 | * Flush all the entries of the specified table. | |
407 | */ | |
c5549571 | 408 | static void ip_vs_lblcr_flush(struct ip_vs_service *svc) |
1da177e4 | 409 | { |
c5549571 | 410 | struct ip_vs_lblcr_table *tbl = svc->sched_data; |
1da177e4 | 411 | int i; |
c5549571 JA |
412 | struct ip_vs_lblcr_entry *en; |
413 | struct hlist_node *next; | |
1da177e4 | 414 | |
ba3a3ce1 | 415 | spin_lock_bh(&svc->sched_lock); |
c5549571 | 416 | tbl->dead = 1; |
1da177e4 | 417 | for (i=0; i<IP_VS_LBLCR_TAB_SIZE; i++) { |
c5549571 | 418 | hlist_for_each_entry_safe(en, next, &tbl->bucket[i], list) { |
1da177e4 | 419 | ip_vs_lblcr_free(en); |
1da177e4 | 420 | } |
1da177e4 | 421 | } |
ba3a3ce1 | 422 | spin_unlock_bh(&svc->sched_lock); |
1da177e4 LT |
423 | } |
424 | ||
b27d777e SH |
425 | static int sysctl_lblcr_expiration(struct ip_vs_service *svc) |
426 | { | |
427 | #ifdef CONFIG_SYSCTL | |
428 | struct netns_ipvs *ipvs = net_ipvs(svc->net); | |
429 | return ipvs->sysctl_lblcr_expiration; | |
430 | #else | |
431 | return DEFAULT_EXPIRATION; | |
432 | #endif | |
433 | } | |
1da177e4 | 434 | |
f728bafb | 435 | static inline void ip_vs_lblcr_full_check(struct ip_vs_service *svc) |
1da177e4 | 436 | { |
f728bafb | 437 | struct ip_vs_lblcr_table *tbl = svc->sched_data; |
1da177e4 LT |
438 | unsigned long now = jiffies; |
439 | int i, j; | |
c5549571 JA |
440 | struct ip_vs_lblcr_entry *en; |
441 | struct hlist_node *next; | |
1da177e4 LT |
442 | |
443 | for (i=0, j=tbl->rover; i<IP_VS_LBLCR_TAB_SIZE; i++) { | |
444 | j = (j + 1) & IP_VS_LBLCR_TAB_MASK; | |
445 | ||
ba3a3ce1 | 446 | spin_lock(&svc->sched_lock); |
c5549571 | 447 | hlist_for_each_entry_safe(en, next, &tbl->bucket[j], list) { |
b27d777e SH |
448 | if (time_after(en->lastuse + |
449 | sysctl_lblcr_expiration(svc), now)) | |
1da177e4 LT |
450 | continue; |
451 | ||
452 | ip_vs_lblcr_free(en); | |
453 | atomic_dec(&tbl->entries); | |
454 | } | |
ba3a3ce1 | 455 | spin_unlock(&svc->sched_lock); |
1da177e4 LT |
456 | } |
457 | tbl->rover = j; | |
458 | } | |
459 | ||
460 | ||
461 | /* | |
462 | * Periodical timer handler for IPVS lblcr table | |
463 | * It is used to collect stale entries when the number of entries | |
464 | * exceeds the maximum size of the table. | |
465 | * | |
466 | * Fixme: we probably need more complicated algorithm to collect | |
467 | * entries that have not been used for a long time even | |
468 | * if the number of entries doesn't exceed the maximum size | |
469 | * of the table. | |
470 | * The full expiration check is for this purpose now. | |
471 | */ | |
472 | static void ip_vs_lblcr_check_expire(unsigned long data) | |
473 | { | |
f728bafb SW |
474 | struct ip_vs_service *svc = (struct ip_vs_service *) data; |
475 | struct ip_vs_lblcr_table *tbl = svc->sched_data; | |
1da177e4 LT |
476 | unsigned long now = jiffies; |
477 | int goal; | |
478 | int i, j; | |
c5549571 JA |
479 | struct ip_vs_lblcr_entry *en; |
480 | struct hlist_node *next; | |
1da177e4 | 481 | |
1da177e4 LT |
482 | if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) { |
483 | /* do full expiration check */ | |
f728bafb | 484 | ip_vs_lblcr_full_check(svc); |
1da177e4 LT |
485 | tbl->counter = 1; |
486 | goto out; | |
487 | } | |
488 | ||
489 | if (atomic_read(&tbl->entries) <= tbl->max_size) { | |
490 | tbl->counter++; | |
491 | goto out; | |
492 | } | |
493 | ||
494 | goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3; | |
495 | if (goal > tbl->max_size/2) | |
496 | goal = tbl->max_size/2; | |
497 | ||
498 | for (i=0, j=tbl->rover; i<IP_VS_LBLCR_TAB_SIZE; i++) { | |
499 | j = (j + 1) & IP_VS_LBLCR_TAB_MASK; | |
500 | ||
ba3a3ce1 | 501 | spin_lock(&svc->sched_lock); |
c5549571 | 502 | hlist_for_each_entry_safe(en, next, &tbl->bucket[j], list) { |
1da177e4 LT |
503 | if (time_before(now, en->lastuse+ENTRY_TIMEOUT)) |
504 | continue; | |
505 | ||
506 | ip_vs_lblcr_free(en); | |
507 | atomic_dec(&tbl->entries); | |
508 | goal--; | |
509 | } | |
ba3a3ce1 | 510 | spin_unlock(&svc->sched_lock); |
1da177e4 LT |
511 | if (goal <= 0) |
512 | break; | |
513 | } | |
514 | tbl->rover = j; | |
515 | ||
516 | out: | |
517 | mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL); | |
518 | } | |
519 | ||
1da177e4 LT |
520 | static int ip_vs_lblcr_init_svc(struct ip_vs_service *svc) |
521 | { | |
522 | int i; | |
523 | struct ip_vs_lblcr_table *tbl; | |
524 | ||
525 | /* | |
526 | * Allocate the ip_vs_lblcr_table for this service | |
527 | */ | |
45d4e71a | 528 | tbl = kmalloc(sizeof(*tbl), GFP_KERNEL); |
0a9ee813 | 529 | if (tbl == NULL) |
1da177e4 | 530 | return -ENOMEM; |
0a9ee813 | 531 | |
1da177e4 LT |
532 | svc->sched_data = tbl; |
533 | IP_VS_DBG(6, "LBLCR hash table (memory=%Zdbytes) allocated for " | |
f728bafb | 534 | "current service\n", sizeof(*tbl)); |
1da177e4 LT |
535 | |
536 | /* | |
537 | * Initialize the hash buckets | |
538 | */ | |
539 | for (i=0; i<IP_VS_LBLCR_TAB_SIZE; i++) { | |
c5549571 | 540 | INIT_HLIST_HEAD(&tbl->bucket[i]); |
1da177e4 | 541 | } |
1da177e4 LT |
542 | tbl->max_size = IP_VS_LBLCR_TAB_SIZE*16; |
543 | tbl->rover = 0; | |
544 | tbl->counter = 1; | |
c5549571 | 545 | tbl->dead = 0; |
1da177e4 LT |
546 | |
547 | /* | |
548 | * Hook periodic timer for garbage collection | |
549 | */ | |
b24b8a24 | 550 | setup_timer(&tbl->periodic_timer, ip_vs_lblcr_check_expire, |
f728bafb SW |
551 | (unsigned long)svc); |
552 | mod_timer(&tbl->periodic_timer, jiffies + CHECK_EXPIRE_INTERVAL); | |
1da177e4 | 553 | |
1da177e4 LT |
554 | return 0; |
555 | } | |
556 | ||
557 | ||
ed3ffc4e | 558 | static void ip_vs_lblcr_done_svc(struct ip_vs_service *svc) |
1da177e4 LT |
559 | { |
560 | struct ip_vs_lblcr_table *tbl = svc->sched_data; | |
561 | ||
562 | /* remove periodic timer */ | |
563 | del_timer_sync(&tbl->periodic_timer); | |
564 | ||
565 | /* got to clean up table entries here */ | |
c5549571 | 566 | ip_vs_lblcr_flush(svc); |
1da177e4 LT |
567 | |
568 | /* release the table itself */ | |
c5549571 | 569 | kfree_rcu(tbl, rcu_head); |
1da177e4 | 570 | IP_VS_DBG(6, "LBLCR hash table (memory=%Zdbytes) released\n", |
f728bafb | 571 | sizeof(*tbl)); |
1da177e4 LT |
572 | } |
573 | ||
574 | ||
1da177e4 | 575 | static inline struct ip_vs_dest * |
44548375 | 576 | __ip_vs_lblcr_schedule(struct ip_vs_service *svc) |
1da177e4 LT |
577 | { |
578 | struct ip_vs_dest *dest, *least; | |
579 | int loh, doh; | |
580 | ||
581 | /* | |
b552f7e3 | 582 | * We use the following formula to estimate the load: |
1da177e4 LT |
583 | * (dest overhead) / dest->weight |
584 | * | |
585 | * Remember -- no floats in kernel mode!!! | |
586 | * The comparison of h1*w2 > h2*w1 is equivalent to that of | |
587 | * h1/w1 > h2/w2 | |
588 | * if every weight is larger than zero. | |
589 | * | |
590 | * The server with weight=0 is quiesced and will not receive any | |
591 | * new connection. | |
592 | */ | |
c5549571 | 593 | list_for_each_entry_rcu(dest, &svc->destinations, n_list) { |
1da177e4 LT |
594 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) |
595 | continue; | |
596 | ||
597 | if (atomic_read(&dest->weight) > 0) { | |
598 | least = dest; | |
b552f7e3 | 599 | loh = ip_vs_dest_conn_overhead(least); |
1da177e4 LT |
600 | goto nextstage; |
601 | } | |
602 | } | |
603 | return NULL; | |
604 | ||
605 | /* | |
606 | * Find the destination with the least load. | |
607 | */ | |
608 | nextstage: | |
c5549571 | 609 | list_for_each_entry_continue_rcu(dest, &svc->destinations, n_list) { |
1da177e4 LT |
610 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) |
611 | continue; | |
612 | ||
b552f7e3 | 613 | doh = ip_vs_dest_conn_overhead(dest); |
1da177e4 LT |
614 | if (loh * atomic_read(&dest->weight) > |
615 | doh * atomic_read(&least->weight)) { | |
616 | least = dest; | |
617 | loh = doh; | |
618 | } | |
619 | } | |
620 | ||
44548375 JV |
621 | IP_VS_DBG_BUF(6, "LBLCR: server %s:%d " |
622 | "activeconns %d refcnt %d weight %d overhead %d\n", | |
623 | IP_VS_DBG_ADDR(least->af, &least->addr), | |
624 | ntohs(least->port), | |
625 | atomic_read(&least->activeconns), | |
626 | atomic_read(&least->refcnt), | |
627 | atomic_read(&least->weight), loh); | |
1da177e4 LT |
628 | |
629 | return least; | |
630 | } | |
631 | ||
632 | ||
633 | /* | |
634 | * If this destination server is overloaded and there is a less loaded | |
635 | * server, then return true. | |
636 | */ | |
637 | static inline int | |
638 | is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc) | |
639 | { | |
640 | if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) { | |
641 | struct ip_vs_dest *d; | |
642 | ||
c5549571 | 643 | list_for_each_entry_rcu(d, &svc->destinations, n_list) { |
1da177e4 LT |
644 | if (atomic_read(&d->activeconns)*2 |
645 | < atomic_read(&d->weight)) { | |
646 | return 1; | |
647 | } | |
648 | } | |
649 | } | |
650 | return 0; | |
651 | } | |
652 | ||
653 | ||
654 | /* | |
655 | * Locality-Based (weighted) Least-Connection scheduling | |
656 | */ | |
657 | static struct ip_vs_dest * | |
658 | ip_vs_lblcr_schedule(struct ip_vs_service *svc, const struct sk_buff *skb) | |
659 | { | |
f728bafb | 660 | struct ip_vs_lblcr_table *tbl = svc->sched_data; |
44548375 | 661 | struct ip_vs_iphdr iph; |
c5549571 | 662 | struct ip_vs_dest *dest; |
f728bafb | 663 | struct ip_vs_lblcr_entry *en; |
1da177e4 | 664 | |
63dca2c0 | 665 | ip_vs_fill_iph_addr_only(svc->af, skb, &iph); |
44548375 | 666 | |
1e3e238e | 667 | IP_VS_DBG(6, "%s(): Scheduling...\n", __func__); |
1da177e4 | 668 | |
f728bafb | 669 | /* First look in our cache */ |
44548375 | 670 | en = ip_vs_lblcr_get(svc->af, tbl, &iph.daddr); |
f728bafb | 671 | if (en) { |
f728bafb SW |
672 | en->lastuse = jiffies; |
673 | ||
674 | /* Get the least loaded destination */ | |
1da177e4 | 675 | dest = ip_vs_dest_set_min(&en->set); |
f728bafb SW |
676 | |
677 | /* More than one destination + enough time passed by, cleanup */ | |
1da177e4 | 678 | if (atomic_read(&en->set.size) > 1 && |
c5549571 | 679 | time_after(jiffies, en->set.lastmod + |
b27d777e | 680 | sysctl_lblcr_expiration(svc))) { |
ac69269a | 681 | spin_lock_bh(&svc->sched_lock); |
c5549571 JA |
682 | if (atomic_read(&en->set.size) > 1) { |
683 | struct ip_vs_dest *m; | |
f728bafb | 684 | |
c5549571 JA |
685 | m = ip_vs_dest_set_max(&en->set); |
686 | if (m) | |
687 | ip_vs_dest_set_erase(&en->set, m); | |
688 | } | |
ac69269a | 689 | spin_unlock_bh(&svc->sched_lock); |
f728bafb SW |
690 | } |
691 | ||
692 | /* If the destination is not overloaded, use it */ | |
c5549571 | 693 | if (dest && !is_overloaded(dest, svc)) |
f728bafb | 694 | goto out; |
f728bafb SW |
695 | |
696 | /* The cache entry is invalid, time to schedule */ | |
44548375 | 697 | dest = __ip_vs_lblcr_schedule(svc); |
f728bafb | 698 | if (!dest) { |
41ac51ee | 699 | ip_vs_scheduler_err(svc, "no destination available"); |
f728bafb | 700 | return NULL; |
1da177e4 | 701 | } |
f728bafb SW |
702 | |
703 | /* Update our cache entry */ | |
ac69269a | 704 | spin_lock_bh(&svc->sched_lock); |
c5549571 JA |
705 | if (!tbl->dead) |
706 | ip_vs_dest_set_insert(&en->set, dest, true); | |
ac69269a | 707 | spin_unlock_bh(&svc->sched_lock); |
f728bafb | 708 | goto out; |
c5549571 | 709 | } |
f728bafb SW |
710 | |
711 | /* No cache entry, time to schedule */ | |
44548375 | 712 | dest = __ip_vs_lblcr_schedule(svc); |
f728bafb SW |
713 | if (!dest) { |
714 | IP_VS_DBG(1, "no destination available\n"); | |
715 | return NULL; | |
1da177e4 | 716 | } |
1da177e4 | 717 | |
f728bafb | 718 | /* If we fail to create a cache entry, we'll just use the valid dest */ |
ac69269a | 719 | spin_lock_bh(&svc->sched_lock); |
c5549571 JA |
720 | if (!tbl->dead) |
721 | ip_vs_lblcr_new(tbl, &iph.daddr, dest); | |
ac69269a | 722 | spin_unlock_bh(&svc->sched_lock); |
f728bafb SW |
723 | |
724 | out: | |
44548375 JV |
725 | IP_VS_DBG_BUF(6, "LBLCR: destination IP address %s --> server %s:%d\n", |
726 | IP_VS_DBG_ADDR(svc->af, &iph.daddr), | |
727 | IP_VS_DBG_ADDR(svc->af, &dest->addr), ntohs(dest->port)); | |
1da177e4 LT |
728 | |
729 | return dest; | |
730 | } | |
731 | ||
732 | ||
733 | /* | |
734 | * IPVS LBLCR Scheduler structure | |
735 | */ | |
736 | static struct ip_vs_scheduler ip_vs_lblcr_scheduler = | |
737 | { | |
738 | .name = "lblcr", | |
739 | .refcnt = ATOMIC_INIT(0), | |
740 | .module = THIS_MODULE, | |
d149ccc9 | 741 | .n_list = LIST_HEAD_INIT(ip_vs_lblcr_scheduler.n_list), |
1da177e4 LT |
742 | .init_service = ip_vs_lblcr_init_svc, |
743 | .done_service = ip_vs_lblcr_done_svc, | |
1da177e4 LT |
744 | .schedule = ip_vs_lblcr_schedule, |
745 | }; | |
746 | ||
61b1ab45 HS |
747 | /* |
748 | * per netns init. | |
749 | */ | |
fb1de432 | 750 | #ifdef CONFIG_SYSCTL |
61b1ab45 HS |
751 | static int __net_init __ip_vs_lblcr_init(struct net *net) |
752 | { | |
d0a1eef9 HS |
753 | struct netns_ipvs *ipvs = net_ipvs(net); |
754 | ||
4b984cd5 HS |
755 | if (!ipvs) |
756 | return -ENOENT; | |
757 | ||
d0a1eef9 HS |
758 | if (!net_eq(net, &init_net)) { |
759 | ipvs->lblcr_ctl_table = kmemdup(vs_vars_table, | |
760 | sizeof(vs_vars_table), | |
761 | GFP_KERNEL); | |
762 | if (ipvs->lblcr_ctl_table == NULL) | |
0443929f | 763 | return -ENOMEM; |
464dc801 EB |
764 | |
765 | /* Don't export sysctls to unprivileged users */ | |
766 | if (net->user_ns != &init_user_ns) | |
767 | ipvs->lblcr_ctl_table[0].procname = NULL; | |
d0a1eef9 HS |
768 | } else |
769 | ipvs->lblcr_ctl_table = vs_vars_table; | |
b27d777e | 770 | ipvs->sysctl_lblcr_expiration = DEFAULT_EXPIRATION; |
d0a1eef9 HS |
771 | ipvs->lblcr_ctl_table[0].data = &ipvs->sysctl_lblcr_expiration; |
772 | ||
773 | ipvs->lblcr_ctl_header = | |
ec8f23ce | 774 | register_net_sysctl(net, "net/ipv4/vs", ipvs->lblcr_ctl_table); |
0443929f SH |
775 | if (!ipvs->lblcr_ctl_header) { |
776 | if (!net_eq(net, &init_net)) | |
777 | kfree(ipvs->lblcr_ctl_table); | |
778 | return -ENOMEM; | |
779 | } | |
61b1ab45 HS |
780 | |
781 | return 0; | |
782 | } | |
783 | ||
784 | static void __net_exit __ip_vs_lblcr_exit(struct net *net) | |
785 | { | |
d0a1eef9 HS |
786 | struct netns_ipvs *ipvs = net_ipvs(net); |
787 | ||
788 | unregister_net_sysctl_table(ipvs->lblcr_ctl_header); | |
61b1ab45 | 789 | |
d0a1eef9 HS |
790 | if (!net_eq(net, &init_net)) |
791 | kfree(ipvs->lblcr_ctl_table); | |
61b1ab45 HS |
792 | } |
793 | ||
fb1de432 SH |
794 | #else |
795 | ||
796 | static int __net_init __ip_vs_lblcr_init(struct net *net) { return 0; } | |
797 | static void __net_exit __ip_vs_lblcr_exit(struct net *net) { } | |
798 | ||
799 | #endif | |
800 | ||
61b1ab45 HS |
801 | static struct pernet_operations ip_vs_lblcr_ops = { |
802 | .init = __ip_vs_lblcr_init, | |
803 | .exit = __ip_vs_lblcr_exit, | |
804 | }; | |
1da177e4 LT |
805 | |
806 | static int __init ip_vs_lblcr_init(void) | |
807 | { | |
a014bc8f PE |
808 | int ret; |
809 | ||
61b1ab45 HS |
810 | ret = register_pernet_subsys(&ip_vs_lblcr_ops); |
811 | if (ret) | |
812 | return ret; | |
813 | ||
a014bc8f PE |
814 | ret = register_ip_vs_scheduler(&ip_vs_lblcr_scheduler); |
815 | if (ret) | |
61b1ab45 | 816 | unregister_pernet_subsys(&ip_vs_lblcr_ops); |
a014bc8f | 817 | return ret; |
1da177e4 LT |
818 | } |
819 | ||
1da177e4 LT |
820 | static void __exit ip_vs_lblcr_cleanup(void) |
821 | { | |
1da177e4 | 822 | unregister_ip_vs_scheduler(&ip_vs_lblcr_scheduler); |
61b1ab45 | 823 | unregister_pernet_subsys(&ip_vs_lblcr_ops); |
ceec4c38 | 824 | synchronize_rcu(); |
1da177e4 LT |
825 | } |
826 | ||
827 | ||
828 | module_init(ip_vs_lblcr_init); | |
829 | module_exit(ip_vs_lblcr_cleanup); | |
830 | MODULE_LICENSE("GPL"); |