Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 1 | /* |
| 2 | * IPVS: Locality-Based Least-Connection with Replication scheduler |
| 3 | * |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 | |
Hannes Eder | 9aada7a | 2009-07-30 14:29:44 -0700 | [diff] [blame] | 40 | #define KMSG_COMPONENT "IPVS" |
| 41 | #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt |
| 42 | |
Arnaldo Carvalho de Melo | 14c8502 | 2005-12-27 02:43:12 -0200 | [diff] [blame] | 43 | #include <linux/ip.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 44 | #include <linux/module.h> |
| 45 | #include <linux/kernel.h> |
Arnaldo Carvalho de Melo | 14c8502 | 2005-12-27 02:43:12 -0200 | [diff] [blame] | 46 | #include <linux/skbuff.h> |
Al Viro | d7fe0f2 | 2006-12-03 23:15:30 -0500 | [diff] [blame] | 47 | #include <linux/jiffies.h> |
Simon Horman | 51f0bc7 | 2010-02-26 17:45:14 +0100 | [diff] [blame] | 48 | #include <linux/list.h> |
Tejun Heo | 5a0e3ad | 2010-03-24 17:04:11 +0900 | [diff] [blame] | 49 | #include <linux/slab.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 50 | |
| 51 | /* for sysctl */ |
| 52 | #include <linux/fs.h> |
| 53 | #include <linux/sysctl.h> |
Eric W. Biederman | 457c4cb | 2007-09-12 12:01:34 +0200 | [diff] [blame] | 54 | #include <net/net_namespace.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 | |
Simon Horman | b27d777 | 2011-02-04 18:33:01 +0900 | [diff] [blame] | 66 | #define DEFAULT_EXPIRATION (24*60*60*HZ) |
| 67 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 | */ |
Simon Horman | 51f0bc7 | 2010-02-26 17:45:14 +0100 | [diff] [blame] | 90 | struct ip_vs_dest_set_elem { |
| 91 | struct list_head list; /* list link */ |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 92 | struct ip_vs_dest *dest; /* destination server */ |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 93 | struct rcu_head rcu_head; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 94 | }; |
| 95 | |
| 96 | struct ip_vs_dest_set { |
| 97 | atomic_t size; /* set size */ |
| 98 | unsigned long lastmod; /* last modified time */ |
Simon Horman | 51f0bc7 | 2010-02-26 17:45:14 +0100 | [diff] [blame] | 99 | struct list_head list; /* destination list */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 100 | }; |
| 101 | |
| 102 | |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 103 | static void ip_vs_dest_set_insert(struct ip_vs_dest_set *set, |
| 104 | struct ip_vs_dest *dest, bool check) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 105 | { |
Simon Horman | 51f0bc7 | 2010-02-26 17:45:14 +0100 | [diff] [blame] | 106 | struct ip_vs_dest_set_elem *e; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 107 | |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 108 | if (check) { |
| 109 | list_for_each_entry(e, &set->list, list) { |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 110 | if (e->dest == dest) |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 111 | return; |
| 112 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 113 | } |
| 114 | |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 115 | e = kmalloc(sizeof(*e), GFP_ATOMIC); |
Joe Perches | 0a9ee81 | 2011-08-29 14:17:25 -0700 | [diff] [blame] | 116 | if (e == NULL) |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 117 | return; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 118 | |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 119 | ip_vs_dest_hold(dest); |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 120 | e->dest = dest; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 121 | |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 122 | list_add_rcu(&e->list, &set->list); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 123 | atomic_inc(&set->size); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 124 | |
| 125 | set->lastmod = jiffies; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 126 | } |
| 127 | |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 128 | static void ip_vs_lblcr_elem_rcu_free(struct rcu_head *head) |
| 129 | { |
| 130 | struct ip_vs_dest_set_elem *e; |
| 131 | |
| 132 | e = container_of(head, struct ip_vs_dest_set_elem, rcu_head); |
Julian Anastasov | 9e4e948 | 2013-10-09 09:24:27 +0300 | [diff] [blame] | 133 | ip_vs_dest_put_and_free(e->dest); |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 134 | kfree(e); |
| 135 | } |
| 136 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 137 | static void |
| 138 | ip_vs_dest_set_erase(struct ip_vs_dest_set *set, struct ip_vs_dest *dest) |
| 139 | { |
Simon Horman | 51f0bc7 | 2010-02-26 17:45:14 +0100 | [diff] [blame] | 140 | struct ip_vs_dest_set_elem *e; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 141 | |
Simon Horman | 51f0bc7 | 2010-02-26 17:45:14 +0100 | [diff] [blame] | 142 | list_for_each_entry(e, &set->list, list) { |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 143 | if (e->dest == dest) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 144 | /* HIT */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 145 | atomic_dec(&set->size); |
| 146 | set->lastmod = jiffies; |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 147 | list_del_rcu(&e->list); |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 148 | call_rcu(&e->rcu_head, ip_vs_lblcr_elem_rcu_free); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 149 | break; |
| 150 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 151 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 152 | } |
| 153 | |
| 154 | static void ip_vs_dest_set_eraseall(struct ip_vs_dest_set *set) |
| 155 | { |
Simon Horman | 51f0bc7 | 2010-02-26 17:45:14 +0100 | [diff] [blame] | 156 | struct ip_vs_dest_set_elem *e, *ep; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 157 | |
Simon Horman | 51f0bc7 | 2010-02-26 17:45:14 +0100 | [diff] [blame] | 158 | list_for_each_entry_safe(e, ep, &set->list, list) { |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 159 | list_del_rcu(&e->list); |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 160 | call_rcu(&e->rcu_head, ip_vs_lblcr_elem_rcu_free); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 161 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 162 | } |
| 163 | |
| 164 | /* get weighted least-connection node in the destination set */ |
| 165 | static inline struct ip_vs_dest *ip_vs_dest_set_min(struct ip_vs_dest_set *set) |
| 166 | { |
Simon Horman | 51f0bc7 | 2010-02-26 17:45:14 +0100 | [diff] [blame] | 167 | register struct ip_vs_dest_set_elem *e; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 168 | struct ip_vs_dest *dest, *least; |
| 169 | int loh, doh; |
| 170 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 171 | /* select the first destination server, whose weight > 0 */ |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 172 | list_for_each_entry_rcu(e, &set->list, list) { |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 173 | least = e->dest; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 174 | if (least->flags & IP_VS_DEST_F_OVERLOAD) |
| 175 | continue; |
| 176 | |
| 177 | if ((atomic_read(&least->weight) > 0) |
| 178 | && (least->flags & IP_VS_DEST_F_AVAILABLE)) { |
Changli Gao | b552f7e | 2011-02-19 17:32:28 +0800 | [diff] [blame] | 179 | loh = ip_vs_dest_conn_overhead(least); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 180 | goto nextstage; |
| 181 | } |
| 182 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 183 | return NULL; |
| 184 | |
| 185 | /* find the destination with the weighted least load */ |
| 186 | nextstage: |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 187 | list_for_each_entry_continue_rcu(e, &set->list, list) { |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 188 | dest = e->dest; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 189 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) |
| 190 | continue; |
| 191 | |
Changli Gao | b552f7e | 2011-02-19 17:32:28 +0800 | [diff] [blame] | 192 | doh = ip_vs_dest_conn_overhead(dest); |
Simon Kirby | c16526a | 2013-08-10 01:26:18 -0700 | [diff] [blame] | 193 | if (((__s64)loh * atomic_read(&dest->weight) > |
| 194 | (__s64)doh * atomic_read(&least->weight)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 195 | && (dest->flags & IP_VS_DEST_F_AVAILABLE)) { |
| 196 | least = dest; |
| 197 | loh = doh; |
| 198 | } |
| 199 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 200 | |
Hannes Eder | 1e3e238 | 2009-08-02 11:05:41 +0000 | [diff] [blame] | 201 | IP_VS_DBG_BUF(6, "%s(): server %s:%d " |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 202 | "activeconns %d refcnt %d weight %d overhead %d\n", |
Hannes Eder | 1e3e238 | 2009-08-02 11:05:41 +0000 | [diff] [blame] | 203 | __func__, |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 204 | IP_VS_DBG_ADDR(least->af, &least->addr), |
| 205 | ntohs(least->port), |
| 206 | atomic_read(&least->activeconns), |
| 207 | atomic_read(&least->refcnt), |
| 208 | atomic_read(&least->weight), loh); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 209 | return least; |
| 210 | } |
| 211 | |
| 212 | |
| 213 | /* get weighted most-connection node in the destination set */ |
| 214 | static inline struct ip_vs_dest *ip_vs_dest_set_max(struct ip_vs_dest_set *set) |
| 215 | { |
Simon Horman | 51f0bc7 | 2010-02-26 17:45:14 +0100 | [diff] [blame] | 216 | register struct ip_vs_dest_set_elem *e; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 217 | struct ip_vs_dest *dest, *most; |
| 218 | int moh, doh; |
| 219 | |
| 220 | if (set == NULL) |
| 221 | return NULL; |
| 222 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 223 | /* select the first destination server, whose weight > 0 */ |
Simon Horman | 51f0bc7 | 2010-02-26 17:45:14 +0100 | [diff] [blame] | 224 | list_for_each_entry(e, &set->list, list) { |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 225 | most = e->dest; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 226 | if (atomic_read(&most->weight) > 0) { |
Changli Gao | b552f7e | 2011-02-19 17:32:28 +0800 | [diff] [blame] | 227 | moh = ip_vs_dest_conn_overhead(most); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 228 | goto nextstage; |
| 229 | } |
| 230 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 231 | return NULL; |
| 232 | |
| 233 | /* find the destination with the weighted most load */ |
| 234 | nextstage: |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 235 | list_for_each_entry_continue(e, &set->list, list) { |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 236 | dest = e->dest; |
Changli Gao | b552f7e | 2011-02-19 17:32:28 +0800 | [diff] [blame] | 237 | doh = ip_vs_dest_conn_overhead(dest); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 238 | /* moh/mw < doh/dw ==> moh*dw < doh*mw, where mw,dw>0 */ |
Simon Kirby | c16526a | 2013-08-10 01:26:18 -0700 | [diff] [blame] | 239 | if (((__s64)moh * atomic_read(&dest->weight) < |
| 240 | (__s64)doh * atomic_read(&most->weight)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 241 | && (atomic_read(&dest->weight) > 0)) { |
| 242 | most = dest; |
| 243 | moh = doh; |
| 244 | } |
| 245 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 246 | |
Hannes Eder | 1e3e238 | 2009-08-02 11:05:41 +0000 | [diff] [blame] | 247 | IP_VS_DBG_BUF(6, "%s(): server %s:%d " |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 248 | "activeconns %d refcnt %d weight %d overhead %d\n", |
Hannes Eder | 1e3e238 | 2009-08-02 11:05:41 +0000 | [diff] [blame] | 249 | __func__, |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 250 | IP_VS_DBG_ADDR(most->af, &most->addr), ntohs(most->port), |
| 251 | atomic_read(&most->activeconns), |
| 252 | atomic_read(&most->refcnt), |
| 253 | atomic_read(&most->weight), moh); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 254 | return most; |
| 255 | } |
| 256 | |
| 257 | |
| 258 | /* |
| 259 | * IPVS lblcr entry represents an association between destination |
| 260 | * IP address and its destination server set |
| 261 | */ |
| 262 | struct ip_vs_lblcr_entry { |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 263 | struct hlist_node list; |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 264 | int af; /* address family */ |
| 265 | union nf_inet_addr addr; /* destination IP address */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 266 | struct ip_vs_dest_set set; /* destination server set */ |
| 267 | unsigned long lastuse; /* last used time */ |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 268 | struct rcu_head rcu_head; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 269 | }; |
| 270 | |
| 271 | |
| 272 | /* |
| 273 | * IPVS lblcr hash table |
| 274 | */ |
| 275 | struct ip_vs_lblcr_table { |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 276 | struct rcu_head rcu_head; |
Julian Anastasov | f33c8b9 | 2013-04-17 23:50:47 +0300 | [diff] [blame] | 277 | struct hlist_head bucket[IP_VS_LBLCR_TAB_SIZE]; /* hash bucket */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 278 | atomic_t entries; /* number of entries */ |
| 279 | int max_size; /* maximum size of entries */ |
| 280 | struct timer_list periodic_timer; /* collect stale entries */ |
| 281 | int rover; /* rover for expire check */ |
| 282 | int counter; /* counter for no expire */ |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 283 | bool dead; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 284 | }; |
| 285 | |
| 286 | |
Simon Horman | fb1de43 | 2011-02-04 18:33:02 +0900 | [diff] [blame] | 287 | #ifdef CONFIG_SYSCTL |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 288 | /* |
| 289 | * IPVS LBLCR sysctl table |
| 290 | */ |
| 291 | |
Joe Perches | fe2c633 | 2013-06-11 23:04:25 -0700 | [diff] [blame] | 292 | static struct ctl_table vs_vars_table[] = { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 293 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 294 | .procname = "lblcr_expiration", |
Hans Schillstrom | d0a1eef | 2011-01-03 14:44:44 +0100 | [diff] [blame] | 295 | .data = NULL, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 296 | .maxlen = sizeof(int), |
YOSHIFUJI Hideaki | e905a9e | 2007-02-09 23:24:47 +0900 | [diff] [blame] | 297 | .mode = 0644, |
Alexey Dobriyan | 6d9f239 | 2008-11-03 18:21:05 -0800 | [diff] [blame] | 298 | .proc_handler = proc_dointvec_jiffies, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 299 | }, |
Eric W. Biederman | f8572d8 | 2009-11-05 13:32:03 -0800 | [diff] [blame] | 300 | { } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 301 | }; |
Simon Horman | fb1de43 | 2011-02-04 18:33:02 +0900 | [diff] [blame] | 302 | #endif |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 303 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 304 | static inline void ip_vs_lblcr_free(struct ip_vs_lblcr_entry *en) |
| 305 | { |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 306 | hlist_del_rcu(&en->list); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 307 | ip_vs_dest_set_eraseall(&en->set); |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 308 | kfree_rcu(en, rcu_head); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 309 | } |
| 310 | |
| 311 | |
| 312 | /* |
| 313 | * Returns hash value for IPVS LBLCR entry |
| 314 | */ |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 315 | static inline unsigned int |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 316 | ip_vs_lblcr_hashkey(int af, const union nf_inet_addr *addr) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 317 | { |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 318 | __be32 addr_fold = addr->ip; |
| 319 | |
| 320 | #ifdef CONFIG_IP_VS_IPV6 |
| 321 | if (af == AF_INET6) |
| 322 | addr_fold = addr->ip6[0]^addr->ip6[1]^ |
| 323 | addr->ip6[2]^addr->ip6[3]; |
| 324 | #endif |
| 325 | return (ntohl(addr_fold)*2654435761UL) & IP_VS_LBLCR_TAB_MASK; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 326 | } |
| 327 | |
| 328 | |
| 329 | /* |
| 330 | * Hash an entry in the ip_vs_lblcr_table. |
| 331 | * returns bool success. |
| 332 | */ |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 333 | static void |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 334 | ip_vs_lblcr_hash(struct ip_vs_lblcr_table *tbl, struct ip_vs_lblcr_entry *en) |
| 335 | { |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 336 | unsigned int hash = ip_vs_lblcr_hashkey(en->af, &en->addr); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 337 | |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 338 | hlist_add_head_rcu(&en->list, &tbl->bucket[hash]); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 339 | atomic_inc(&tbl->entries); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 340 | } |
| 341 | |
| 342 | |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 343 | /* Get ip_vs_lblcr_entry associated with supplied parameters. */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 344 | static inline struct ip_vs_lblcr_entry * |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 345 | ip_vs_lblcr_get(int af, struct ip_vs_lblcr_table *tbl, |
| 346 | const union nf_inet_addr *addr) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 347 | { |
Eric Dumazet | 95c9617 | 2012-04-15 05:58:06 +0000 | [diff] [blame] | 348 | unsigned int hash = ip_vs_lblcr_hashkey(af, addr); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 349 | struct ip_vs_lblcr_entry *en; |
| 350 | |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 351 | hlist_for_each_entry_rcu(en, &tbl->bucket[hash], list) |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 352 | if (ip_vs_addr_equal(af, &en->addr, addr)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 353 | return en; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 354 | |
| 355 | return NULL; |
| 356 | } |
| 357 | |
| 358 | |
| 359 | /* |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 360 | * Create or update an ip_vs_lblcr_entry, which is a mapping of a destination |
Julian Anastasov | ba3a3ce | 2013-03-22 11:46:51 +0200 | [diff] [blame] | 361 | * IP address to a server. Called under spin lock. |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 362 | */ |
| 363 | static inline struct ip_vs_lblcr_entry * |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 364 | ip_vs_lblcr_new(struct ip_vs_lblcr_table *tbl, const union nf_inet_addr *daddr, |
Julian Anastasov | cf34e64 | 2014-09-09 16:40:30 -0700 | [diff] [blame] | 365 | u16 af, struct ip_vs_dest *dest) |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 366 | { |
| 367 | struct ip_vs_lblcr_entry *en; |
| 368 | |
Julian Anastasov | cf34e64 | 2014-09-09 16:40:30 -0700 | [diff] [blame] | 369 | en = ip_vs_lblcr_get(af, tbl, daddr); |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 370 | if (!en) { |
| 371 | en = kmalloc(sizeof(*en), GFP_ATOMIC); |
Joe Perches | 0a9ee81 | 2011-08-29 14:17:25 -0700 | [diff] [blame] | 372 | if (!en) |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 373 | return NULL; |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 374 | |
Julian Anastasov | cf34e64 | 2014-09-09 16:40:30 -0700 | [diff] [blame] | 375 | en->af = af; |
| 376 | ip_vs_addr_copy(af, &en->addr, daddr); |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 377 | en->lastuse = jiffies; |
| 378 | |
Uwe Kleine-König | 421f91d | 2010-06-11 12:17:00 +0200 | [diff] [blame] | 379 | /* initialize its dest set */ |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 380 | atomic_set(&(en->set.size), 0); |
Simon Horman | 51f0bc7 | 2010-02-26 17:45:14 +0100 | [diff] [blame] | 381 | INIT_LIST_HEAD(&en->set.list); |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 382 | |
| 383 | ip_vs_dest_set_insert(&en->set, dest, false); |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 384 | |
| 385 | ip_vs_lblcr_hash(tbl, en); |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 386 | return en; |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 387 | } |
| 388 | |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 389 | ip_vs_dest_set_insert(&en->set, dest, true); |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 390 | |
| 391 | return en; |
| 392 | } |
| 393 | |
| 394 | |
| 395 | /* |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 396 | * Flush all the entries of the specified table. |
| 397 | */ |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 398 | static void ip_vs_lblcr_flush(struct ip_vs_service *svc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 399 | { |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 400 | struct ip_vs_lblcr_table *tbl = svc->sched_data; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 401 | int i; |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 402 | struct ip_vs_lblcr_entry *en; |
| 403 | struct hlist_node *next; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 404 | |
Julian Anastasov | ba3a3ce | 2013-03-22 11:46:51 +0200 | [diff] [blame] | 405 | spin_lock_bh(&svc->sched_lock); |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 406 | tbl->dead = 1; |
Dragos Foianu | 70e3ca7 | 2013-07-11 09:45:42 +0300 | [diff] [blame] | 407 | for (i = 0; i < IP_VS_LBLCR_TAB_SIZE; i++) { |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 408 | hlist_for_each_entry_safe(en, next, &tbl->bucket[i], list) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 409 | ip_vs_lblcr_free(en); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 410 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 411 | } |
Julian Anastasov | ba3a3ce | 2013-03-22 11:46:51 +0200 | [diff] [blame] | 412 | spin_unlock_bh(&svc->sched_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 413 | } |
| 414 | |
Simon Horman | b27d777 | 2011-02-04 18:33:01 +0900 | [diff] [blame] | 415 | static int sysctl_lblcr_expiration(struct ip_vs_service *svc) |
| 416 | { |
| 417 | #ifdef CONFIG_SYSCTL |
Eric W. Biederman | 3109d2f | 2015-09-21 13:01:44 -0500 | [diff] [blame] | 418 | return svc->ipvs->sysctl_lblcr_expiration; |
Simon Horman | b27d777 | 2011-02-04 18:33:01 +0900 | [diff] [blame] | 419 | #else |
| 420 | return DEFAULT_EXPIRATION; |
| 421 | #endif |
| 422 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 423 | |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 424 | static inline void ip_vs_lblcr_full_check(struct ip_vs_service *svc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 425 | { |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 426 | struct ip_vs_lblcr_table *tbl = svc->sched_data; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 427 | unsigned long now = jiffies; |
| 428 | int i, j; |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 429 | struct ip_vs_lblcr_entry *en; |
| 430 | struct hlist_node *next; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 431 | |
Dragos Foianu | 70e3ca7 | 2013-07-11 09:45:42 +0300 | [diff] [blame] | 432 | for (i = 0, j = tbl->rover; i < IP_VS_LBLCR_TAB_SIZE; i++) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 433 | j = (j + 1) & IP_VS_LBLCR_TAB_MASK; |
| 434 | |
Julian Anastasov | ba3a3ce | 2013-03-22 11:46:51 +0200 | [diff] [blame] | 435 | spin_lock(&svc->sched_lock); |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 436 | hlist_for_each_entry_safe(en, next, &tbl->bucket[j], list) { |
Simon Horman | b27d777 | 2011-02-04 18:33:01 +0900 | [diff] [blame] | 437 | if (time_after(en->lastuse + |
| 438 | sysctl_lblcr_expiration(svc), now)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 439 | continue; |
| 440 | |
| 441 | ip_vs_lblcr_free(en); |
| 442 | atomic_dec(&tbl->entries); |
| 443 | } |
Julian Anastasov | ba3a3ce | 2013-03-22 11:46:51 +0200 | [diff] [blame] | 444 | spin_unlock(&svc->sched_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 445 | } |
| 446 | tbl->rover = j; |
| 447 | } |
| 448 | |
| 449 | |
| 450 | /* |
| 451 | * Periodical timer handler for IPVS lblcr table |
| 452 | * It is used to collect stale entries when the number of entries |
| 453 | * exceeds the maximum size of the table. |
| 454 | * |
| 455 | * Fixme: we probably need more complicated algorithm to collect |
| 456 | * entries that have not been used for a long time even |
| 457 | * if the number of entries doesn't exceed the maximum size |
| 458 | * of the table. |
| 459 | * The full expiration check is for this purpose now. |
| 460 | */ |
| 461 | static void ip_vs_lblcr_check_expire(unsigned long data) |
| 462 | { |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 463 | struct ip_vs_service *svc = (struct ip_vs_service *) data; |
| 464 | struct ip_vs_lblcr_table *tbl = svc->sched_data; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 465 | unsigned long now = jiffies; |
| 466 | int goal; |
| 467 | int i, j; |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 468 | struct ip_vs_lblcr_entry *en; |
| 469 | struct hlist_node *next; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 470 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 471 | if ((tbl->counter % COUNT_FOR_FULL_EXPIRATION) == 0) { |
| 472 | /* do full expiration check */ |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 473 | ip_vs_lblcr_full_check(svc); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 474 | tbl->counter = 1; |
| 475 | goto out; |
| 476 | } |
| 477 | |
| 478 | if (atomic_read(&tbl->entries) <= tbl->max_size) { |
| 479 | tbl->counter++; |
| 480 | goto out; |
| 481 | } |
| 482 | |
| 483 | goal = (atomic_read(&tbl->entries) - tbl->max_size)*4/3; |
| 484 | if (goal > tbl->max_size/2) |
| 485 | goal = tbl->max_size/2; |
| 486 | |
Dragos Foianu | 70e3ca7 | 2013-07-11 09:45:42 +0300 | [diff] [blame] | 487 | for (i = 0, j = tbl->rover; i < IP_VS_LBLCR_TAB_SIZE; i++) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 488 | j = (j + 1) & IP_VS_LBLCR_TAB_MASK; |
| 489 | |
Julian Anastasov | ba3a3ce | 2013-03-22 11:46:51 +0200 | [diff] [blame] | 490 | spin_lock(&svc->sched_lock); |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 491 | hlist_for_each_entry_safe(en, next, &tbl->bucket[j], list) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 492 | if (time_before(now, en->lastuse+ENTRY_TIMEOUT)) |
| 493 | continue; |
| 494 | |
| 495 | ip_vs_lblcr_free(en); |
| 496 | atomic_dec(&tbl->entries); |
| 497 | goal--; |
| 498 | } |
Julian Anastasov | ba3a3ce | 2013-03-22 11:46:51 +0200 | [diff] [blame] | 499 | spin_unlock(&svc->sched_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 500 | if (goal <= 0) |
| 501 | break; |
| 502 | } |
| 503 | tbl->rover = j; |
| 504 | |
| 505 | out: |
| 506 | mod_timer(&tbl->periodic_timer, jiffies+CHECK_EXPIRE_INTERVAL); |
| 507 | } |
| 508 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 509 | static int ip_vs_lblcr_init_svc(struct ip_vs_service *svc) |
| 510 | { |
| 511 | int i; |
| 512 | struct ip_vs_lblcr_table *tbl; |
| 513 | |
| 514 | /* |
| 515 | * Allocate the ip_vs_lblcr_table for this service |
| 516 | */ |
Julian Anastasov | 45d4e71 | 2012-04-13 16:49:41 +0300 | [diff] [blame] | 517 | tbl = kmalloc(sizeof(*tbl), GFP_KERNEL); |
Joe Perches | 0a9ee81 | 2011-08-29 14:17:25 -0700 | [diff] [blame] | 518 | if (tbl == NULL) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 519 | return -ENOMEM; |
Joe Perches | 0a9ee81 | 2011-08-29 14:17:25 -0700 | [diff] [blame] | 520 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 521 | svc->sched_data = tbl; |
| 522 | IP_VS_DBG(6, "LBLCR hash table (memory=%Zdbytes) allocated for " |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 523 | "current service\n", sizeof(*tbl)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 524 | |
| 525 | /* |
| 526 | * Initialize the hash buckets |
| 527 | */ |
Dragos Foianu | 70e3ca7 | 2013-07-11 09:45:42 +0300 | [diff] [blame] | 528 | for (i = 0; i < IP_VS_LBLCR_TAB_SIZE; i++) { |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 529 | INIT_HLIST_HEAD(&tbl->bucket[i]); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 530 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 531 | tbl->max_size = IP_VS_LBLCR_TAB_SIZE*16; |
| 532 | tbl->rover = 0; |
| 533 | tbl->counter = 1; |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 534 | tbl->dead = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 535 | |
| 536 | /* |
| 537 | * Hook periodic timer for garbage collection |
| 538 | */ |
Pavel Emelyanov | b24b8a2 | 2008-01-23 21:20:07 -0800 | [diff] [blame] | 539 | setup_timer(&tbl->periodic_timer, ip_vs_lblcr_check_expire, |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 540 | (unsigned long)svc); |
| 541 | mod_timer(&tbl->periodic_timer, jiffies + CHECK_EXPIRE_INTERVAL); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 542 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 543 | return 0; |
| 544 | } |
| 545 | |
| 546 | |
Julian Anastasov | ed3ffc4 | 2013-03-22 11:46:50 +0200 | [diff] [blame] | 547 | static void ip_vs_lblcr_done_svc(struct ip_vs_service *svc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 548 | { |
| 549 | struct ip_vs_lblcr_table *tbl = svc->sched_data; |
| 550 | |
| 551 | /* remove periodic timer */ |
| 552 | del_timer_sync(&tbl->periodic_timer); |
| 553 | |
| 554 | /* got to clean up table entries here */ |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 555 | ip_vs_lblcr_flush(svc); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 556 | |
| 557 | /* release the table itself */ |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 558 | kfree_rcu(tbl, rcu_head); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 559 | IP_VS_DBG(6, "LBLCR hash table (memory=%Zdbytes) released\n", |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 560 | sizeof(*tbl)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 561 | } |
| 562 | |
| 563 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 564 | static inline struct ip_vs_dest * |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 565 | __ip_vs_lblcr_schedule(struct ip_vs_service *svc) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 566 | { |
| 567 | struct ip_vs_dest *dest, *least; |
| 568 | int loh, doh; |
| 569 | |
| 570 | /* |
Changli Gao | b552f7e | 2011-02-19 17:32:28 +0800 | [diff] [blame] | 571 | * We use the following formula to estimate the load: |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 572 | * (dest overhead) / dest->weight |
| 573 | * |
| 574 | * Remember -- no floats in kernel mode!!! |
| 575 | * The comparison of h1*w2 > h2*w1 is equivalent to that of |
| 576 | * h1/w1 > h2/w2 |
| 577 | * if every weight is larger than zero. |
| 578 | * |
| 579 | * The server with weight=0 is quiesced and will not receive any |
| 580 | * new connection. |
| 581 | */ |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 582 | list_for_each_entry_rcu(dest, &svc->destinations, n_list) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 583 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) |
| 584 | continue; |
| 585 | |
| 586 | if (atomic_read(&dest->weight) > 0) { |
| 587 | least = dest; |
Changli Gao | b552f7e | 2011-02-19 17:32:28 +0800 | [diff] [blame] | 588 | loh = ip_vs_dest_conn_overhead(least); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 589 | goto nextstage; |
| 590 | } |
| 591 | } |
| 592 | return NULL; |
| 593 | |
| 594 | /* |
| 595 | * Find the destination with the least load. |
| 596 | */ |
| 597 | nextstage: |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 598 | list_for_each_entry_continue_rcu(dest, &svc->destinations, n_list) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 599 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) |
| 600 | continue; |
| 601 | |
Changli Gao | b552f7e | 2011-02-19 17:32:28 +0800 | [diff] [blame] | 602 | doh = ip_vs_dest_conn_overhead(dest); |
Simon Kirby | c16526a | 2013-08-10 01:26:18 -0700 | [diff] [blame] | 603 | if ((__s64)loh * atomic_read(&dest->weight) > |
| 604 | (__s64)doh * atomic_read(&least->weight)) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 605 | least = dest; |
| 606 | loh = doh; |
| 607 | } |
| 608 | } |
| 609 | |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 610 | IP_VS_DBG_BUF(6, "LBLCR: server %s:%d " |
| 611 | "activeconns %d refcnt %d weight %d overhead %d\n", |
| 612 | IP_VS_DBG_ADDR(least->af, &least->addr), |
| 613 | ntohs(least->port), |
| 614 | atomic_read(&least->activeconns), |
| 615 | atomic_read(&least->refcnt), |
| 616 | atomic_read(&least->weight), loh); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 617 | |
| 618 | return least; |
| 619 | } |
| 620 | |
| 621 | |
| 622 | /* |
| 623 | * If this destination server is overloaded and there is a less loaded |
| 624 | * server, then return true. |
| 625 | */ |
| 626 | static inline int |
| 627 | is_overloaded(struct ip_vs_dest *dest, struct ip_vs_service *svc) |
| 628 | { |
| 629 | if (atomic_read(&dest->activeconns) > atomic_read(&dest->weight)) { |
| 630 | struct ip_vs_dest *d; |
| 631 | |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 632 | list_for_each_entry_rcu(d, &svc->destinations, n_list) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 633 | if (atomic_read(&d->activeconns)*2 |
| 634 | < atomic_read(&d->weight)) { |
| 635 | return 1; |
| 636 | } |
| 637 | } |
| 638 | } |
| 639 | return 0; |
| 640 | } |
| 641 | |
| 642 | |
| 643 | /* |
| 644 | * Locality-Based (weighted) Least-Connection scheduling |
| 645 | */ |
| 646 | static struct ip_vs_dest * |
Julian Anastasov | bba54de | 2013-06-16 09:09:36 +0300 | [diff] [blame] | 647 | ip_vs_lblcr_schedule(struct ip_vs_service *svc, const struct sk_buff *skb, |
| 648 | struct ip_vs_iphdr *iph) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 649 | { |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 650 | struct ip_vs_lblcr_table *tbl = svc->sched_data; |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 651 | struct ip_vs_dest *dest; |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 652 | struct ip_vs_lblcr_entry *en; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 653 | |
Hannes Eder | 1e3e238 | 2009-08-02 11:05:41 +0000 | [diff] [blame] | 654 | IP_VS_DBG(6, "%s(): Scheduling...\n", __func__); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 655 | |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 656 | /* First look in our cache */ |
Julian Anastasov | bba54de | 2013-06-16 09:09:36 +0300 | [diff] [blame] | 657 | en = ip_vs_lblcr_get(svc->af, tbl, &iph->daddr); |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 658 | if (en) { |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 659 | en->lastuse = jiffies; |
| 660 | |
| 661 | /* Get the least loaded destination */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 662 | dest = ip_vs_dest_set_min(&en->set); |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 663 | |
| 664 | /* More than one destination + enough time passed by, cleanup */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 665 | if (atomic_read(&en->set.size) > 1 && |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 666 | time_after(jiffies, en->set.lastmod + |
Simon Horman | b27d777 | 2011-02-04 18:33:01 +0900 | [diff] [blame] | 667 | sysctl_lblcr_expiration(svc))) { |
Julian Anastasov | ac69269 | 2013-03-22 11:46:54 +0200 | [diff] [blame] | 668 | spin_lock_bh(&svc->sched_lock); |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 669 | if (atomic_read(&en->set.size) > 1) { |
| 670 | struct ip_vs_dest *m; |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 671 | |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 672 | m = ip_vs_dest_set_max(&en->set); |
| 673 | if (m) |
| 674 | ip_vs_dest_set_erase(&en->set, m); |
| 675 | } |
Julian Anastasov | ac69269 | 2013-03-22 11:46:54 +0200 | [diff] [blame] | 676 | spin_unlock_bh(&svc->sched_lock); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 677 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 678 | |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 679 | /* If the destination is not overloaded, use it */ |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 680 | if (dest && !is_overloaded(dest, svc)) |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 681 | goto out; |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 682 | |
| 683 | /* The cache entry is invalid, time to schedule */ |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 684 | dest = __ip_vs_lblcr_schedule(svc); |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 685 | if (!dest) { |
Patrick Schaaf | 41ac51e | 2011-02-11 14:01:12 +0100 | [diff] [blame] | 686 | ip_vs_scheduler_err(svc, "no destination available"); |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 687 | return NULL; |
| 688 | } |
| 689 | |
| 690 | /* Update our cache entry */ |
Julian Anastasov | ac69269 | 2013-03-22 11:46:54 +0200 | [diff] [blame] | 691 | spin_lock_bh(&svc->sched_lock); |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 692 | if (!tbl->dead) |
| 693 | ip_vs_dest_set_insert(&en->set, dest, true); |
Julian Anastasov | ac69269 | 2013-03-22 11:46:54 +0200 | [diff] [blame] | 694 | spin_unlock_bh(&svc->sched_lock); |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 695 | goto out; |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 696 | } |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 697 | |
| 698 | /* No cache entry, time to schedule */ |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 699 | dest = __ip_vs_lblcr_schedule(svc); |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 700 | if (!dest) { |
| 701 | IP_VS_DBG(1, "no destination available\n"); |
| 702 | return NULL; |
| 703 | } |
| 704 | |
| 705 | /* If we fail to create a cache entry, we'll just use the valid dest */ |
Julian Anastasov | ac69269 | 2013-03-22 11:46:54 +0200 | [diff] [blame] | 706 | spin_lock_bh(&svc->sched_lock); |
Julian Anastasov | c554957 | 2013-03-22 11:46:41 +0200 | [diff] [blame] | 707 | if (!tbl->dead) |
Julian Anastasov | cf34e64 | 2014-09-09 16:40:30 -0700 | [diff] [blame] | 708 | ip_vs_lblcr_new(tbl, &iph->daddr, svc->af, dest); |
Julian Anastasov | ac69269 | 2013-03-22 11:46:54 +0200 | [diff] [blame] | 709 | spin_unlock_bh(&svc->sched_lock); |
Sven Wegener | f728baf | 2008-08-19 08:16:19 +0200 | [diff] [blame] | 710 | |
| 711 | out: |
Julius Volz | 4454837 | 2008-11-03 17:08:28 -0800 | [diff] [blame] | 712 | IP_VS_DBG_BUF(6, "LBLCR: destination IP address %s --> server %s:%d\n", |
Julian Anastasov | bba54de | 2013-06-16 09:09:36 +0300 | [diff] [blame] | 713 | IP_VS_DBG_ADDR(svc->af, &iph->daddr), |
Julian Anastasov | cf34e64 | 2014-09-09 16:40:30 -0700 | [diff] [blame] | 714 | IP_VS_DBG_ADDR(dest->af, &dest->addr), ntohs(dest->port)); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 715 | |
| 716 | return dest; |
| 717 | } |
| 718 | |
| 719 | |
| 720 | /* |
| 721 | * IPVS LBLCR Scheduler structure |
| 722 | */ |
| 723 | static struct ip_vs_scheduler ip_vs_lblcr_scheduler = |
| 724 | { |
| 725 | .name = "lblcr", |
| 726 | .refcnt = ATOMIC_INIT(0), |
| 727 | .module = THIS_MODULE, |
Sven Wegener | d149ccc | 2008-08-10 09:18:02 +0000 | [diff] [blame] | 728 | .n_list = LIST_HEAD_INIT(ip_vs_lblcr_scheduler.n_list), |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 729 | .init_service = ip_vs_lblcr_init_svc, |
| 730 | .done_service = ip_vs_lblcr_done_svc, |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 731 | .schedule = ip_vs_lblcr_schedule, |
| 732 | }; |
| 733 | |
Hans Schillstrom | 61b1ab4 | 2011-01-03 14:44:42 +0100 | [diff] [blame] | 734 | /* |
| 735 | * per netns init. |
| 736 | */ |
Simon Horman | fb1de43 | 2011-02-04 18:33:02 +0900 | [diff] [blame] | 737 | #ifdef CONFIG_SYSCTL |
Hans Schillstrom | 61b1ab4 | 2011-01-03 14:44:42 +0100 | [diff] [blame] | 738 | static int __net_init __ip_vs_lblcr_init(struct net *net) |
| 739 | { |
Hans Schillstrom | d0a1eef | 2011-01-03 14:44:44 +0100 | [diff] [blame] | 740 | struct netns_ipvs *ipvs = net_ipvs(net); |
Hans Schillstrom | 61b1ab4 | 2011-01-03 14:44:42 +0100 | [diff] [blame] | 741 | |
Hans Schillstrom | 4b984cd | 2012-04-26 09:45:34 +0200 | [diff] [blame] | 742 | if (!ipvs) |
| 743 | return -ENOENT; |
| 744 | |
Hans Schillstrom | d0a1eef | 2011-01-03 14:44:44 +0100 | [diff] [blame] | 745 | if (!net_eq(net, &init_net)) { |
| 746 | ipvs->lblcr_ctl_table = kmemdup(vs_vars_table, |
| 747 | sizeof(vs_vars_table), |
| 748 | GFP_KERNEL); |
| 749 | if (ipvs->lblcr_ctl_table == NULL) |
Simon Horman | 0443929 | 2011-02-01 18:29:04 +0100 | [diff] [blame] | 750 | return -ENOMEM; |
Eric W. Biederman | 464dc80 | 2012-11-16 03:02:59 +0000 | [diff] [blame] | 751 | |
| 752 | /* Don't export sysctls to unprivileged users */ |
| 753 | if (net->user_ns != &init_user_ns) |
| 754 | ipvs->lblcr_ctl_table[0].procname = NULL; |
Hans Schillstrom | d0a1eef | 2011-01-03 14:44:44 +0100 | [diff] [blame] | 755 | } else |
| 756 | ipvs->lblcr_ctl_table = vs_vars_table; |
Simon Horman | b27d777 | 2011-02-04 18:33:01 +0900 | [diff] [blame] | 757 | ipvs->sysctl_lblcr_expiration = DEFAULT_EXPIRATION; |
Hans Schillstrom | d0a1eef | 2011-01-03 14:44:44 +0100 | [diff] [blame] | 758 | ipvs->lblcr_ctl_table[0].data = &ipvs->sysctl_lblcr_expiration; |
| 759 | |
| 760 | ipvs->lblcr_ctl_header = |
Eric W. Biederman | ec8f23c | 2012-04-19 13:44:49 +0000 | [diff] [blame] | 761 | register_net_sysctl(net, "net/ipv4/vs", ipvs->lblcr_ctl_table); |
Simon Horman | 0443929 | 2011-02-01 18:29:04 +0100 | [diff] [blame] | 762 | if (!ipvs->lblcr_ctl_header) { |
| 763 | if (!net_eq(net, &init_net)) |
| 764 | kfree(ipvs->lblcr_ctl_table); |
| 765 | return -ENOMEM; |
| 766 | } |
Hans Schillstrom | 61b1ab4 | 2011-01-03 14:44:42 +0100 | [diff] [blame] | 767 | |
| 768 | return 0; |
| 769 | } |
| 770 | |
| 771 | static void __net_exit __ip_vs_lblcr_exit(struct net *net) |
| 772 | { |
Hans Schillstrom | d0a1eef | 2011-01-03 14:44:44 +0100 | [diff] [blame] | 773 | struct netns_ipvs *ipvs = net_ipvs(net); |
Hans Schillstrom | 61b1ab4 | 2011-01-03 14:44:42 +0100 | [diff] [blame] | 774 | |
Hans Schillstrom | d0a1eef | 2011-01-03 14:44:44 +0100 | [diff] [blame] | 775 | unregister_net_sysctl_table(ipvs->lblcr_ctl_header); |
| 776 | |
| 777 | if (!net_eq(net, &init_net)) |
| 778 | kfree(ipvs->lblcr_ctl_table); |
Hans Schillstrom | 61b1ab4 | 2011-01-03 14:44:42 +0100 | [diff] [blame] | 779 | } |
| 780 | |
Simon Horman | fb1de43 | 2011-02-04 18:33:02 +0900 | [diff] [blame] | 781 | #else |
| 782 | |
| 783 | static int __net_init __ip_vs_lblcr_init(struct net *net) { return 0; } |
| 784 | static void __net_exit __ip_vs_lblcr_exit(struct net *net) { } |
| 785 | |
| 786 | #endif |
| 787 | |
Hans Schillstrom | 61b1ab4 | 2011-01-03 14:44:42 +0100 | [diff] [blame] | 788 | static struct pernet_operations ip_vs_lblcr_ops = { |
| 789 | .init = __ip_vs_lblcr_init, |
| 790 | .exit = __ip_vs_lblcr_exit, |
| 791 | }; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 792 | |
| 793 | static int __init ip_vs_lblcr_init(void) |
| 794 | { |
Pavel Emelyanov | a014bc8 | 2007-12-04 00:43:24 -0800 | [diff] [blame] | 795 | int ret; |
| 796 | |
Hans Schillstrom | 61b1ab4 | 2011-01-03 14:44:42 +0100 | [diff] [blame] | 797 | ret = register_pernet_subsys(&ip_vs_lblcr_ops); |
| 798 | if (ret) |
| 799 | return ret; |
| 800 | |
Pavel Emelyanov | a014bc8 | 2007-12-04 00:43:24 -0800 | [diff] [blame] | 801 | ret = register_ip_vs_scheduler(&ip_vs_lblcr_scheduler); |
| 802 | if (ret) |
Hans Schillstrom | 61b1ab4 | 2011-01-03 14:44:42 +0100 | [diff] [blame] | 803 | unregister_pernet_subsys(&ip_vs_lblcr_ops); |
Pavel Emelyanov | a014bc8 | 2007-12-04 00:43:24 -0800 | [diff] [blame] | 804 | return ret; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 805 | } |
| 806 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 807 | static void __exit ip_vs_lblcr_cleanup(void) |
| 808 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 809 | unregister_ip_vs_scheduler(&ip_vs_lblcr_scheduler); |
Hans Schillstrom | 61b1ab4 | 2011-01-03 14:44:42 +0100 | [diff] [blame] | 810 | unregister_pernet_subsys(&ip_vs_lblcr_ops); |
Julian Anastasov | 742617b | 2013-09-12 11:21:09 +0300 | [diff] [blame] | 811 | rcu_barrier(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 812 | } |
| 813 | |
| 814 | |
| 815 | module_init(ip_vs_lblcr_init); |
| 816 | module_exit(ip_vs_lblcr_cleanup); |
| 817 | MODULE_LICENSE("GPL"); |