]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * IPVS: Locality-Based Least-Connection scheduling module | |
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 | * Martin Hamilton : fixed the terrible locking bugs | |
13 | * *lock(tbl->lock) ==> *lock(&tbl->lock) | |
421f91d2 | 14 | * Wensong Zhang : fixed the uninitialized tbl->lock bug |
1da177e4 LT |
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 | ||
9aada7ac HE |
42 | #define KMSG_COMPONENT "IPVS" |
43 | #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt | |
44 | ||
14c85021 | 45 | #include <linux/ip.h> |
5a0e3ad6 | 46 | #include <linux/slab.h> |
1da177e4 LT |
47 | #include <linux/module.h> |
48 | #include <linux/kernel.h> | |
14c85021 | 49 | #include <linux/skbuff.h> |
d7fe0f24 | 50 | #include <linux/jiffies.h> |
1da177e4 LT |
51 | |
52 | /* for sysctl */ | |
53 | #include <linux/fs.h> | |
54 | #include <linux/sysctl.h> | |
55 | ||
56 | #include <net/ip_vs.h> | |
57 | ||
58 | ||
59 | /* | |
60 | * It is for garbage collection of stale IPVS lblc 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 | /* | |
78 | * for IPVS lblc entry hash table | |
79 | */ | |
80 | #ifndef CONFIG_IP_VS_LBLC_TAB_BITS | |
81 | #define CONFIG_IP_VS_LBLC_TAB_BITS 10 | |
82 | #endif | |
83 | #define IP_VS_LBLC_TAB_BITS CONFIG_IP_VS_LBLC_TAB_BITS | |
84 | #define IP_VS_LBLC_TAB_SIZE (1 << IP_VS_LBLC_TAB_BITS) | |
85 | #define IP_VS_LBLC_TAB_MASK (IP_VS_LBLC_TAB_SIZE - 1) | |
86 | ||
87 | ||
88 | /* | |
89 | * IPVS lblc entry represents an association between destination | |
90 | * IP address and its destination server | |
91 | */ | |
92 | struct ip_vs_lblc_entry { | |
c2a4ffb7 | 93 | struct hlist_node list; |
44548375 JV |
94 | int af; /* address family */ |
95 | union nf_inet_addr addr; /* destination IP address */ | |
2f3d771a | 96 | struct ip_vs_dest *dest; /* real server (cache) */ |
1da177e4 | 97 | unsigned long lastuse; /* last used time */ |
c2a4ffb7 | 98 | struct rcu_head rcu_head; |
1da177e4 LT |
99 | }; |
100 | ||
101 | ||
102 | /* | |
103 | * IPVS lblc hash table | |
104 | */ | |
105 | struct ip_vs_lblc_table { | |
c2a4ffb7 | 106 | struct rcu_head rcu_head; |
f33c8b94 | 107 | struct hlist_head bucket[IP_VS_LBLC_TAB_SIZE]; /* hash bucket */ |
c2a4ffb7 | 108 | struct timer_list periodic_timer; /* collect stale entries */ |
1da177e4 LT |
109 | atomic_t entries; /* number of entries */ |
110 | int max_size; /* maximum size of entries */ | |
1da177e4 LT |
111 | int rover; /* rover for expire check */ |
112 | int counter; /* counter for no expire */ | |
c2a4ffb7 | 113 | bool dead; |
1da177e4 LT |
114 | }; |
115 | ||
116 | ||
117 | /* | |
118 | * IPVS LBLC sysctl table | |
119 | */ | |
fb1de432 | 120 | #ifdef CONFIG_SYSCTL |
fe2c6338 | 121 | static struct ctl_table vs_vars_table[] = { |
1da177e4 | 122 | { |
1da177e4 | 123 | .procname = "lblc_expiration", |
b6e885dd | 124 | .data = NULL, |
1da177e4 | 125 | .maxlen = sizeof(int), |
e905a9ed | 126 | .mode = 0644, |
6d9f239a | 127 | .proc_handler = proc_dointvec_jiffies, |
1da177e4 | 128 | }, |
f8572d8f | 129 | { } |
1da177e4 | 130 | }; |
fb1de432 | 131 | #endif |
1da177e4 | 132 | |
2f3d771a | 133 | static void ip_vs_lblc_rcu_free(struct rcu_head *head) |
1da177e4 | 134 | { |
2f3d771a JA |
135 | struct ip_vs_lblc_entry *en = container_of(head, |
136 | struct ip_vs_lblc_entry, | |
137 | rcu_head); | |
c2a4ffb7 | 138 | |
2f3d771a JA |
139 | ip_vs_dest_put(en->dest); |
140 | kfree(en); | |
1da177e4 LT |
141 | } |
142 | ||
2f3d771a JA |
143 | static inline void ip_vs_lblc_del(struct ip_vs_lblc_entry *en) |
144 | { | |
145 | hlist_del_rcu(&en->list); | |
146 | call_rcu(&en->rcu_head, ip_vs_lblc_rcu_free); | |
147 | } | |
1da177e4 LT |
148 | |
149 | /* | |
150 | * Returns hash value for IPVS LBLC entry | |
151 | */ | |
95c96174 | 152 | static inline unsigned int |
44548375 | 153 | ip_vs_lblc_hashkey(int af, const union nf_inet_addr *addr) |
1da177e4 | 154 | { |
44548375 JV |
155 | __be32 addr_fold = addr->ip; |
156 | ||
157 | #ifdef CONFIG_IP_VS_IPV6 | |
158 | if (af == AF_INET6) | |
159 | addr_fold = addr->ip6[0]^addr->ip6[1]^ | |
160 | addr->ip6[2]^addr->ip6[3]; | |
161 | #endif | |
162 | return (ntohl(addr_fold)*2654435761UL) & IP_VS_LBLC_TAB_MASK; | |
1da177e4 LT |
163 | } |
164 | ||
165 | ||
166 | /* | |
167 | * Hash an entry in the ip_vs_lblc_table. | |
168 | * returns bool success. | |
169 | */ | |
39ac50d0 | 170 | static void |
1da177e4 LT |
171 | ip_vs_lblc_hash(struct ip_vs_lblc_table *tbl, struct ip_vs_lblc_entry *en) |
172 | { | |
95c96174 | 173 | unsigned int hash = ip_vs_lblc_hashkey(en->af, &en->addr); |
1da177e4 | 174 | |
c2a4ffb7 | 175 | hlist_add_head_rcu(&en->list, &tbl->bucket[hash]); |
1da177e4 | 176 | atomic_inc(&tbl->entries); |
1da177e4 LT |
177 | } |
178 | ||
179 | ||
c2a4ffb7 | 180 | /* Get ip_vs_lblc_entry associated with supplied parameters. */ |
1da177e4 | 181 | static inline struct ip_vs_lblc_entry * |
44548375 JV |
182 | ip_vs_lblc_get(int af, struct ip_vs_lblc_table *tbl, |
183 | const union nf_inet_addr *addr) | |
1da177e4 | 184 | { |
95c96174 | 185 | unsigned int hash = ip_vs_lblc_hashkey(af, addr); |
1da177e4 LT |
186 | struct ip_vs_lblc_entry *en; |
187 | ||
c2a4ffb7 | 188 | hlist_for_each_entry_rcu(en, &tbl->bucket[hash], list) |
44548375 | 189 | if (ip_vs_addr_equal(af, &en->addr, addr)) |
39ac50d0 | 190 | return en; |
1da177e4 | 191 | |
39ac50d0 SW |
192 | return NULL; |
193 | } | |
1da177e4 | 194 | |
39ac50d0 SW |
195 | |
196 | /* | |
197 | * Create or update an ip_vs_lblc_entry, which is a mapping of a destination IP | |
ba3a3ce1 | 198 | * address to a server. Called under spin lock. |
39ac50d0 SW |
199 | */ |
200 | static inline struct ip_vs_lblc_entry * | |
44548375 | 201 | ip_vs_lblc_new(struct ip_vs_lblc_table *tbl, const union nf_inet_addr *daddr, |
39ac50d0 SW |
202 | struct ip_vs_dest *dest) |
203 | { | |
204 | struct ip_vs_lblc_entry *en; | |
205 | ||
44548375 | 206 | en = ip_vs_lblc_get(dest->af, tbl, daddr); |
2f3d771a JA |
207 | if (en) { |
208 | if (en->dest == dest) | |
209 | return en; | |
210 | ip_vs_lblc_del(en); | |
211 | } | |
212 | en = kmalloc(sizeof(*en), GFP_ATOMIC); | |
213 | if (!en) | |
214 | return NULL; | |
1da177e4 | 215 | |
2f3d771a JA |
216 | en->af = dest->af; |
217 | ip_vs_addr_copy(dest->af, &en->addr, daddr); | |
218 | en->lastuse = jiffies; | |
39ac50d0 | 219 | |
2f3d771a JA |
220 | ip_vs_dest_hold(dest); |
221 | en->dest = dest; | |
c2a4ffb7 | 222 | |
2f3d771a | 223 | ip_vs_lblc_hash(tbl, en); |
39ac50d0 SW |
224 | |
225 | return en; | |
1da177e4 LT |
226 | } |
227 | ||
228 | ||
229 | /* | |
230 | * Flush all the entries of the specified table. | |
231 | */ | |
c2a4ffb7 | 232 | static void ip_vs_lblc_flush(struct ip_vs_service *svc) |
1da177e4 | 233 | { |
c2a4ffb7 JA |
234 | struct ip_vs_lblc_table *tbl = svc->sched_data; |
235 | struct ip_vs_lblc_entry *en; | |
236 | struct hlist_node *next; | |
39ac50d0 | 237 | int i; |
1da177e4 | 238 | |
ba3a3ce1 | 239 | spin_lock_bh(&svc->sched_lock); |
c2a4ffb7 | 240 | tbl->dead = 1; |
1da177e4 | 241 | for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) { |
c2a4ffb7 | 242 | hlist_for_each_entry_safe(en, next, &tbl->bucket[i], list) { |
2f3d771a | 243 | ip_vs_lblc_del(en); |
1da177e4 LT |
244 | atomic_dec(&tbl->entries); |
245 | } | |
1da177e4 | 246 | } |
ba3a3ce1 | 247 | spin_unlock_bh(&svc->sched_lock); |
1da177e4 LT |
248 | } |
249 | ||
b27d777e SH |
250 | static int sysctl_lblc_expiration(struct ip_vs_service *svc) |
251 | { | |
252 | #ifdef CONFIG_SYSCTL | |
253 | struct netns_ipvs *ipvs = net_ipvs(svc->net); | |
254 | return ipvs->sysctl_lblc_expiration; | |
255 | #else | |
256 | return DEFAULT_EXPIRATION; | |
257 | #endif | |
258 | } | |
1da177e4 | 259 | |
39ac50d0 | 260 | static inline void ip_vs_lblc_full_check(struct ip_vs_service *svc) |
1da177e4 | 261 | { |
39ac50d0 | 262 | struct ip_vs_lblc_table *tbl = svc->sched_data; |
c2a4ffb7 JA |
263 | struct ip_vs_lblc_entry *en; |
264 | struct hlist_node *next; | |
1da177e4 LT |
265 | unsigned long now = jiffies; |
266 | int i, j; | |
1da177e4 LT |
267 | |
268 | for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) { | |
269 | j = (j + 1) & IP_VS_LBLC_TAB_MASK; | |
270 | ||
ba3a3ce1 | 271 | spin_lock(&svc->sched_lock); |
c2a4ffb7 | 272 | hlist_for_each_entry_safe(en, next, &tbl->bucket[j], list) { |
e905a9ed | 273 | if (time_before(now, |
b6e885dd | 274 | en->lastuse + |
b27d777e | 275 | sysctl_lblc_expiration(svc))) |
1da177e4 LT |
276 | continue; |
277 | ||
2f3d771a | 278 | ip_vs_lblc_del(en); |
1da177e4 LT |
279 | atomic_dec(&tbl->entries); |
280 | } | |
ba3a3ce1 | 281 | spin_unlock(&svc->sched_lock); |
1da177e4 LT |
282 | } |
283 | tbl->rover = j; | |
284 | } | |
285 | ||
286 | ||
287 | /* | |
288 | * Periodical timer handler for IPVS lblc table | |
289 | * It is used to collect stale entries when the number of entries | |
290 | * exceeds the maximum size of the table. | |
291 | * | |
292 | * Fixme: we probably need more complicated algorithm to collect | |
293 | * entries that have not been used for a long time even | |
294 | * if the number of entries doesn't exceed the maximum size | |
295 | * of the table. | |
296 | * The full expiration check is for this purpose now. | |
297 | */ | |
298 | static void ip_vs_lblc_check_expire(unsigned long data) | |
299 | { | |
39ac50d0 SW |
300 | struct ip_vs_service *svc = (struct ip_vs_service *) data; |
301 | struct ip_vs_lblc_table *tbl = svc->sched_data; | |
1da177e4 LT |
302 | unsigned long now = jiffies; |
303 | int goal; | |
304 | int i, j; | |
c2a4ffb7 JA |
305 | struct ip_vs_lblc_entry *en; |
306 | struct hlist_node *next; | |
1da177e4 | 307 | |
1da177e4 LT |
308 | if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) { |
309 | /* do full expiration check */ | |
39ac50d0 | 310 | ip_vs_lblc_full_check(svc); |
1da177e4 LT |
311 | tbl->counter = 1; |
312 | goto out; | |
313 | } | |
314 | ||
315 | if (atomic_read(&tbl->entries) <= tbl->max_size) { | |
316 | tbl->counter++; | |
317 | goto out; | |
318 | } | |
319 | ||
320 | goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3; | |
321 | if (goal > tbl->max_size/2) | |
322 | goal = tbl->max_size/2; | |
323 | ||
324 | for (i=0, j=tbl->rover; i<IP_VS_LBLC_TAB_SIZE; i++) { | |
325 | j = (j + 1) & IP_VS_LBLC_TAB_MASK; | |
326 | ||
ba3a3ce1 | 327 | spin_lock(&svc->sched_lock); |
c2a4ffb7 | 328 | hlist_for_each_entry_safe(en, next, &tbl->bucket[j], list) { |
1da177e4 LT |
329 | if (time_before(now, en->lastuse + ENTRY_TIMEOUT)) |
330 | continue; | |
331 | ||
2f3d771a | 332 | ip_vs_lblc_del(en); |
1da177e4 LT |
333 | atomic_dec(&tbl->entries); |
334 | goal--; | |
335 | } | |
ba3a3ce1 | 336 | spin_unlock(&svc->sched_lock); |
1da177e4 LT |
337 | if (goal <= 0) |
338 | break; | |
339 | } | |
340 | tbl->rover = j; | |
341 | ||
342 | out: | |
343 | mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL); | |
344 | } | |
345 | ||
346 | ||
347 | static int ip_vs_lblc_init_svc(struct ip_vs_service *svc) | |
348 | { | |
349 | int i; | |
350 | struct ip_vs_lblc_table *tbl; | |
351 | ||
352 | /* | |
353 | * Allocate the ip_vs_lblc_table for this service | |
354 | */ | |
748d845c | 355 | tbl = kmalloc(sizeof(*tbl), GFP_KERNEL); |
0a9ee813 | 356 | if (tbl == NULL) |
1da177e4 | 357 | return -ENOMEM; |
0a9ee813 | 358 | |
1da177e4 LT |
359 | svc->sched_data = tbl; |
360 | IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) allocated for " | |
39ac50d0 | 361 | "current service\n", sizeof(*tbl)); |
1da177e4 LT |
362 | |
363 | /* | |
364 | * Initialize the hash buckets | |
365 | */ | |
366 | for (i=0; i<IP_VS_LBLC_TAB_SIZE; i++) { | |
c2a4ffb7 | 367 | INIT_HLIST_HEAD(&tbl->bucket[i]); |
1da177e4 | 368 | } |
1da177e4 LT |
369 | tbl->max_size = IP_VS_LBLC_TAB_SIZE*16; |
370 | tbl->rover = 0; | |
371 | tbl->counter = 1; | |
c2a4ffb7 | 372 | tbl->dead = 0; |
1da177e4 LT |
373 | |
374 | /* | |
375 | * Hook periodic timer for garbage collection | |
376 | */ | |
b24b8a24 | 377 | setup_timer(&tbl->periodic_timer, ip_vs_lblc_check_expire, |
39ac50d0 SW |
378 | (unsigned long)svc); |
379 | mod_timer(&tbl->periodic_timer, jiffies + CHECK_EXPIRE_INTERVAL); | |
1da177e4 LT |
380 | |
381 | return 0; | |
382 | } | |
383 | ||
384 | ||
ed3ffc4e | 385 | static void ip_vs_lblc_done_svc(struct ip_vs_service *svc) |
1da177e4 LT |
386 | { |
387 | struct ip_vs_lblc_table *tbl = svc->sched_data; | |
388 | ||
389 | /* remove periodic timer */ | |
390 | del_timer_sync(&tbl->periodic_timer); | |
391 | ||
392 | /* got to clean up table entries here */ | |
c2a4ffb7 | 393 | ip_vs_lblc_flush(svc); |
1da177e4 LT |
394 | |
395 | /* release the table itself */ | |
c2a4ffb7 | 396 | kfree_rcu(tbl, rcu_head); |
1da177e4 | 397 | IP_VS_DBG(6, "LBLC hash table (memory=%Zdbytes) released\n", |
39ac50d0 | 398 | sizeof(*tbl)); |
1da177e4 LT |
399 | } |
400 | ||
401 | ||
1da177e4 | 402 | static inline struct ip_vs_dest * |
44548375 | 403 | __ip_vs_lblc_schedule(struct ip_vs_service *svc) |
1da177e4 LT |
404 | { |
405 | struct ip_vs_dest *dest, *least; | |
406 | int loh, doh; | |
407 | ||
408 | /* | |
b552f7e3 | 409 | * We use the following formula to estimate the load: |
1da177e4 LT |
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 | */ | |
c2a4ffb7 | 420 | list_for_each_entry_rcu(dest, &svc->destinations, n_list) { |
1da177e4 LT |
421 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) |
422 | continue; | |
423 | if (atomic_read(&dest->weight) > 0) { | |
424 | least = dest; | |
b552f7e3 | 425 | loh = ip_vs_dest_conn_overhead(least); |
1da177e4 LT |
426 | goto nextstage; |
427 | } | |
428 | } | |
429 | return NULL; | |
430 | ||
431 | /* | |
432 | * Find the destination with the least load. | |
433 | */ | |
434 | nextstage: | |
c2a4ffb7 | 435 | list_for_each_entry_continue_rcu(dest, &svc->destinations, n_list) { |
1da177e4 LT |
436 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) |
437 | continue; | |
438 | ||
b552f7e3 | 439 | doh = ip_vs_dest_conn_overhead(dest); |
c16526a7 SK |
440 | if ((__s64)loh * atomic_read(&dest->weight) > |
441 | (__s64)doh * atomic_read(&least->weight)) { | |
1da177e4 LT |
442 | least = dest; |
443 | loh = doh; | |
444 | } | |
445 | } | |
446 | ||
44548375 JV |
447 | IP_VS_DBG_BUF(6, "LBLC: server %s:%d " |
448 | "activeconns %d refcnt %d weight %d overhead %d\n", | |
449 | IP_VS_DBG_ADDR(least->af, &least->addr), | |
450 | ntohs(least->port), | |
451 | atomic_read(&least->activeconns), | |
452 | atomic_read(&least->refcnt), | |
453 | atomic_read(&least->weight), loh); | |
1da177e4 LT |
454 | |
455 | return least; | |
456 | } | |
457 | ||
458 | ||
459 | /* | |
460 | * If this destination server is overloaded and there is a less loaded | |
461 | * server, then return true. | |
462 | */ | |
463 | static inline int | |
464 | is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc) | |
465 | { | |
466 | if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) { | |
467 | struct ip_vs_dest *d; | |
468 | ||
c2a4ffb7 | 469 | list_for_each_entry_rcu(d, &svc->destinations, n_list) { |
1da177e4 LT |
470 | if (atomic_read(&d->activeconns)*2 |
471 | < atomic_read(&d->weight)) { | |
472 | return 1; | |
473 | } | |
474 | } | |
475 | } | |
476 | return 0; | |
477 | } | |
478 | ||
479 | ||
480 | /* | |
481 | * Locality-Based (weighted) Least-Connection scheduling | |
482 | */ | |
483 | static struct ip_vs_dest * | |
bba54de5 JA |
484 | ip_vs_lblc_schedule(struct ip_vs_service *svc, const struct sk_buff *skb, |
485 | struct ip_vs_iphdr *iph) | |
1da177e4 | 486 | { |
39ac50d0 | 487 | struct ip_vs_lblc_table *tbl = svc->sched_data; |
39ac50d0 SW |
488 | struct ip_vs_dest *dest = NULL; |
489 | struct ip_vs_lblc_entry *en; | |
1da177e4 | 490 | |
1e3e238e | 491 | IP_VS_DBG(6, "%s(): Scheduling...\n", __func__); |
1da177e4 | 492 | |
39ac50d0 | 493 | /* First look in our cache */ |
bba54de5 | 494 | en = ip_vs_lblc_get(svc->af, tbl, &iph->daddr); |
39ac50d0 SW |
495 | if (en) { |
496 | /* We only hold a read lock, but this is atomic */ | |
497 | en->lastuse = jiffies; | |
498 | ||
499 | /* | |
500 | * If the destination is not available, i.e. it's in the trash, | |
501 | * we must ignore it, as it may be removed from under our feet, | |
502 | * if someone drops our reference count. Our caller only makes | |
503 | * sure that destinations, that are not in the trash, are not | |
504 | * moved to the trash, while we are scheduling. But anyone can | |
505 | * free up entries from the trash at any time. | |
506 | */ | |
507 | ||
2f3d771a | 508 | dest = en->dest; |
c2a4ffb7 JA |
509 | if ((dest->flags & IP_VS_DEST_F_AVAILABLE) && |
510 | atomic_read(&dest->weight) > 0 && !is_overloaded(dest, svc)) | |
511 | goto out; | |
39ac50d0 | 512 | } |
39ac50d0 SW |
513 | |
514 | /* No cache entry or it is invalid, time to schedule */ | |
44548375 | 515 | dest = __ip_vs_lblc_schedule(svc); |
39ac50d0 | 516 | if (!dest) { |
41ac51ee | 517 | ip_vs_scheduler_err(svc, "no destination available"); |
39ac50d0 | 518 | return NULL; |
1da177e4 | 519 | } |
1da177e4 | 520 | |
39ac50d0 | 521 | /* If we fail to create a cache entry, we'll just use the valid dest */ |
ac69269a | 522 | spin_lock_bh(&svc->sched_lock); |
c2a4ffb7 | 523 | if (!tbl->dead) |
bba54de5 | 524 | ip_vs_lblc_new(tbl, &iph->daddr, dest); |
ac69269a | 525 | spin_unlock_bh(&svc->sched_lock); |
39ac50d0 SW |
526 | |
527 | out: | |
44548375 | 528 | IP_VS_DBG_BUF(6, "LBLC: destination IP address %s --> server %s:%d\n", |
bba54de5 | 529 | IP_VS_DBG_ADDR(svc->af, &iph->daddr), |
44548375 | 530 | IP_VS_DBG_ADDR(svc->af, &dest->addr), ntohs(dest->port)); |
1da177e4 LT |
531 | |
532 | return dest; | |
533 | } | |
534 | ||
535 | ||
536 | /* | |
537 | * IPVS LBLC Scheduler structure | |
538 | */ | |
539 | static struct ip_vs_scheduler ip_vs_lblc_scheduler = | |
540 | { | |
541 | .name = "lblc", | |
542 | .refcnt = ATOMIC_INIT(0), | |
543 | .module = THIS_MODULE, | |
d149ccc9 | 544 | .n_list = LIST_HEAD_INIT(ip_vs_lblc_scheduler.n_list), |
1da177e4 LT |
545 | .init_service = ip_vs_lblc_init_svc, |
546 | .done_service = ip_vs_lblc_done_svc, | |
1da177e4 LT |
547 | .schedule = ip_vs_lblc_schedule, |
548 | }; | |
549 | ||
61b1ab45 HS |
550 | /* |
551 | * per netns init. | |
552 | */ | |
fb1de432 | 553 | #ifdef CONFIG_SYSCTL |
61b1ab45 HS |
554 | static int __net_init __ip_vs_lblc_init(struct net *net) |
555 | { | |
b6e885dd HS |
556 | struct netns_ipvs *ipvs = net_ipvs(net); |
557 | ||
4b984cd5 HS |
558 | if (!ipvs) |
559 | return -ENOENT; | |
560 | ||
b6e885dd HS |
561 | if (!net_eq(net, &init_net)) { |
562 | ipvs->lblc_ctl_table = kmemdup(vs_vars_table, | |
563 | sizeof(vs_vars_table), | |
564 | GFP_KERNEL); | |
565 | if (ipvs->lblc_ctl_table == NULL) | |
0443929f | 566 | return -ENOMEM; |
464dc801 EB |
567 | |
568 | /* Don't export sysctls to unprivileged users */ | |
569 | if (net->user_ns != &init_user_ns) | |
570 | ipvs->lblc_ctl_table[0].procname = NULL; | |
571 | ||
b6e885dd HS |
572 | } else |
573 | ipvs->lblc_ctl_table = vs_vars_table; | |
b27d777e | 574 | ipvs->sysctl_lblc_expiration = DEFAULT_EXPIRATION; |
b6e885dd HS |
575 | ipvs->lblc_ctl_table[0].data = &ipvs->sysctl_lblc_expiration; |
576 | ||
577 | ipvs->lblc_ctl_header = | |
ec8f23ce | 578 | register_net_sysctl(net, "net/ipv4/vs", ipvs->lblc_ctl_table); |
0443929f SH |
579 | if (!ipvs->lblc_ctl_header) { |
580 | if (!net_eq(net, &init_net)) | |
e5ef39ed | 581 | kfree(ipvs->lblc_ctl_table); |
0443929f SH |
582 | return -ENOMEM; |
583 | } | |
61b1ab45 HS |
584 | |
585 | return 0; | |
586 | } | |
587 | ||
588 | static void __net_exit __ip_vs_lblc_exit(struct net *net) | |
589 | { | |
b6e885dd HS |
590 | struct netns_ipvs *ipvs = net_ipvs(net); |
591 | ||
592 | unregister_net_sysctl_table(ipvs->lblc_ctl_header); | |
61b1ab45 | 593 | |
b6e885dd HS |
594 | if (!net_eq(net, &init_net)) |
595 | kfree(ipvs->lblc_ctl_table); | |
61b1ab45 HS |
596 | } |
597 | ||
fb1de432 SH |
598 | #else |
599 | ||
600 | static int __net_init __ip_vs_lblc_init(struct net *net) { return 0; } | |
601 | static void __net_exit __ip_vs_lblc_exit(struct net *net) { } | |
602 | ||
603 | #endif | |
604 | ||
61b1ab45 HS |
605 | static struct pernet_operations ip_vs_lblc_ops = { |
606 | .init = __ip_vs_lblc_init, | |
607 | .exit = __ip_vs_lblc_exit, | |
608 | }; | |
1da177e4 LT |
609 | |
610 | static int __init ip_vs_lblc_init(void) | |
611 | { | |
a014bc8f PE |
612 | int ret; |
613 | ||
61b1ab45 HS |
614 | ret = register_pernet_subsys(&ip_vs_lblc_ops); |
615 | if (ret) | |
616 | return ret; | |
617 | ||
a014bc8f PE |
618 | ret = register_ip_vs_scheduler(&ip_vs_lblc_scheduler); |
619 | if (ret) | |
61b1ab45 | 620 | unregister_pernet_subsys(&ip_vs_lblc_ops); |
a014bc8f | 621 | return ret; |
1da177e4 LT |
622 | } |
623 | ||
1da177e4 LT |
624 | static void __exit ip_vs_lblc_cleanup(void) |
625 | { | |
1da177e4 | 626 | unregister_ip_vs_scheduler(&ip_vs_lblc_scheduler); |
61b1ab45 | 627 | unregister_pernet_subsys(&ip_vs_lblc_ops); |
2f3d771a | 628 | rcu_barrier(); |
1da177e4 LT |
629 | } |
630 | ||
631 | ||
632 | module_init(ip_vs_lblc_init); | |
633 | module_exit(ip_vs_lblc_cleanup); | |
634 | MODULE_LICENSE("GPL"); |