1 | // SPDX-License-Identifier: GPL-2.0-or-later |
2 | /* |
3 | * IPVS: Destination Hashing scheduling module |
4 | * |
5 | * Authors: Wensong Zhang <wensong@gnuchina.org> |
6 | * |
7 | * Inspired by the consistent hashing scheduler patch from |
8 | * Thomas Proell <proellt@gmx.de> |
9 | * |
10 | * Changes: |
11 | */ |
12 | |
13 | /* |
14 | * The dh algorithm is to select server by the hash key of destination IP |
15 | * address. The pseudo code is as follows: |
16 | * |
17 | * n <- servernode[dest_ip]; |
18 | * if (n is dead) OR |
19 | * (n is overloaded) OR (n.weight <= 0) then |
20 | * return NULL; |
21 | * |
22 | * return n; |
23 | * |
24 | * Notes that servernode is a 256-bucket hash table that maps the hash |
25 | * index derived from packet destination IP address to the current server |
26 | * array. If the dh scheduler is used in cache cluster, it is good to |
27 | * combine it with cache_bypass feature. When the statically assigned |
28 | * server is dead or overloaded, the load balancer can bypass the cache |
29 | * server and send requests to the original server directly. |
30 | * |
31 | */ |
32 | |
33 | #define KMSG_COMPONENT "IPVS" |
34 | #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt |
35 | |
36 | #include <linux/ip.h> |
37 | #include <linux/slab.h> |
38 | #include <linux/module.h> |
39 | #include <linux/kernel.h> |
40 | #include <linux/skbuff.h> |
41 | #include <linux/hash.h> |
42 | |
43 | #include <net/ip_vs.h> |
44 | |
45 | |
46 | /* |
47 | * IPVS DH bucket |
48 | */ |
49 | struct ip_vs_dh_bucket { |
50 | struct ip_vs_dest __rcu *dest; /* real server (cache) */ |
51 | }; |
52 | |
53 | /* |
54 | * for IPVS DH entry hash table |
55 | */ |
56 | #ifndef CONFIG_IP_VS_DH_TAB_BITS |
57 | #define CONFIG_IP_VS_DH_TAB_BITS 8 |
58 | #endif |
59 | #define IP_VS_DH_TAB_BITS CONFIG_IP_VS_DH_TAB_BITS |
60 | #define IP_VS_DH_TAB_SIZE (1 << IP_VS_DH_TAB_BITS) |
61 | #define IP_VS_DH_TAB_MASK (IP_VS_DH_TAB_SIZE - 1) |
62 | |
63 | struct ip_vs_dh_state { |
64 | struct ip_vs_dh_bucket buckets[IP_VS_DH_TAB_SIZE]; |
65 | struct rcu_head rcu_head; |
66 | }; |
67 | |
68 | /* |
69 | * Returns hash value for IPVS DH entry |
70 | */ |
71 | static inline unsigned int ip_vs_dh_hashkey(int af, const union nf_inet_addr *addr) |
72 | { |
73 | __be32 addr_fold = addr->ip; |
74 | |
75 | #ifdef CONFIG_IP_VS_IPV6 |
76 | if (af == AF_INET6) |
77 | addr_fold = addr->ip6[0]^addr->ip6[1]^ |
78 | addr->ip6[2]^addr->ip6[3]; |
79 | #endif |
80 | return hash_32(ntohl(addr_fold), IP_VS_DH_TAB_BITS); |
81 | } |
82 | |
83 | |
84 | /* |
85 | * Get ip_vs_dest associated with supplied parameters. |
86 | */ |
87 | static inline struct ip_vs_dest * |
88 | ip_vs_dh_get(int af, struct ip_vs_dh_state *s, const union nf_inet_addr *addr) |
89 | { |
90 | return rcu_dereference(s->buckets[ip_vs_dh_hashkey(af, addr)].dest); |
91 | } |
92 | |
93 | |
94 | /* |
95 | * Assign all the hash buckets of the specified table with the service. |
96 | */ |
97 | static int |
98 | ip_vs_dh_reassign(struct ip_vs_dh_state *s, struct ip_vs_service *svc) |
99 | { |
100 | int i; |
101 | struct ip_vs_dh_bucket *b; |
102 | struct list_head *p; |
103 | struct ip_vs_dest *dest; |
104 | bool empty; |
105 | |
106 | b = &s->buckets[0]; |
107 | p = &svc->destinations; |
108 | empty = list_empty(head: p); |
109 | for (i=0; i<IP_VS_DH_TAB_SIZE; i++) { |
110 | dest = rcu_dereference_protected(b->dest, 1); |
111 | if (dest) |
112 | ip_vs_dest_put(dest); |
113 | if (empty) |
114 | RCU_INIT_POINTER(b->dest, NULL); |
115 | else { |
116 | if (p == &svc->destinations) |
117 | p = p->next; |
118 | |
119 | dest = list_entry(p, struct ip_vs_dest, n_list); |
120 | ip_vs_dest_hold(dest); |
121 | RCU_INIT_POINTER(b->dest, dest); |
122 | |
123 | p = p->next; |
124 | } |
125 | b++; |
126 | } |
127 | return 0; |
128 | } |
129 | |
130 | |
131 | /* |
132 | * Flush all the hash buckets of the specified table. |
133 | */ |
134 | static void ip_vs_dh_flush(struct ip_vs_dh_state *s) |
135 | { |
136 | int i; |
137 | struct ip_vs_dh_bucket *b; |
138 | struct ip_vs_dest *dest; |
139 | |
140 | b = &s->buckets[0]; |
141 | for (i=0; i<IP_VS_DH_TAB_SIZE; i++) { |
142 | dest = rcu_dereference_protected(b->dest, 1); |
143 | if (dest) { |
144 | ip_vs_dest_put(dest); |
145 | RCU_INIT_POINTER(b->dest, NULL); |
146 | } |
147 | b++; |
148 | } |
149 | } |
150 | |
151 | |
152 | static int ip_vs_dh_init_svc(struct ip_vs_service *svc) |
153 | { |
154 | struct ip_vs_dh_state *s; |
155 | |
156 | /* allocate the DH table for this service */ |
157 | s = kzalloc(size: sizeof(struct ip_vs_dh_state), GFP_KERNEL); |
158 | if (s == NULL) |
159 | return -ENOMEM; |
160 | |
161 | svc->sched_data = s; |
162 | IP_VS_DBG(6, "DH hash table (memory=%zdbytes) allocated for " |
163 | "current service\n" , |
164 | sizeof(struct ip_vs_dh_bucket)*IP_VS_DH_TAB_SIZE); |
165 | |
166 | /* assign the hash buckets with current dests */ |
167 | ip_vs_dh_reassign(s, svc); |
168 | |
169 | return 0; |
170 | } |
171 | |
172 | |
173 | static void ip_vs_dh_done_svc(struct ip_vs_service *svc) |
174 | { |
175 | struct ip_vs_dh_state *s = svc->sched_data; |
176 | |
177 | /* got to clean up hash buckets here */ |
178 | ip_vs_dh_flush(s); |
179 | |
180 | /* release the table itself */ |
181 | kfree_rcu(s, rcu_head); |
182 | IP_VS_DBG(6, "DH hash table (memory=%zdbytes) released\n" , |
183 | sizeof(struct ip_vs_dh_bucket)*IP_VS_DH_TAB_SIZE); |
184 | } |
185 | |
186 | |
187 | static int ip_vs_dh_dest_changed(struct ip_vs_service *svc, |
188 | struct ip_vs_dest *dest) |
189 | { |
190 | struct ip_vs_dh_state *s = svc->sched_data; |
191 | |
192 | /* assign the hash buckets with the updated service */ |
193 | ip_vs_dh_reassign(s, svc); |
194 | |
195 | return 0; |
196 | } |
197 | |
198 | |
199 | /* |
200 | * If the dest flags is set with IP_VS_DEST_F_OVERLOAD, |
201 | * consider that the server is overloaded here. |
202 | */ |
203 | static inline int is_overloaded(struct ip_vs_dest *dest) |
204 | { |
205 | return dest->flags & IP_VS_DEST_F_OVERLOAD; |
206 | } |
207 | |
208 | |
209 | /* |
210 | * Destination hashing scheduling |
211 | */ |
212 | static struct ip_vs_dest * |
213 | ip_vs_dh_schedule(struct ip_vs_service *svc, const struct sk_buff *skb, |
214 | struct ip_vs_iphdr *iph) |
215 | { |
216 | struct ip_vs_dest *dest; |
217 | struct ip_vs_dh_state *s; |
218 | |
219 | IP_VS_DBG(6, "%s(): Scheduling...\n" , __func__); |
220 | |
221 | s = (struct ip_vs_dh_state *) svc->sched_data; |
222 | dest = ip_vs_dh_get(af: svc->af, s, addr: &iph->daddr); |
223 | if (!dest |
224 | || !(dest->flags & IP_VS_DEST_F_AVAILABLE) |
225 | || atomic_read(v: &dest->weight) <= 0 |
226 | || is_overloaded(dest)) { |
227 | ip_vs_scheduler_err(svc, msg: "no destination available" ); |
228 | return NULL; |
229 | } |
230 | |
231 | IP_VS_DBG_BUF(6, "DH: destination IP address %s --> server %s:%d\n" , |
232 | IP_VS_DBG_ADDR(svc->af, &iph->daddr), |
233 | IP_VS_DBG_ADDR(dest->af, &dest->addr), |
234 | ntohs(dest->port)); |
235 | |
236 | return dest; |
237 | } |
238 | |
239 | |
240 | /* |
241 | * IPVS DH Scheduler structure |
242 | */ |
243 | static struct ip_vs_scheduler ip_vs_dh_scheduler = |
244 | { |
245 | .name = "dh" , |
246 | .refcnt = ATOMIC_INIT(0), |
247 | .module = THIS_MODULE, |
248 | .n_list = LIST_HEAD_INIT(ip_vs_dh_scheduler.n_list), |
249 | .init_service = ip_vs_dh_init_svc, |
250 | .done_service = ip_vs_dh_done_svc, |
251 | .add_dest = ip_vs_dh_dest_changed, |
252 | .del_dest = ip_vs_dh_dest_changed, |
253 | .schedule = ip_vs_dh_schedule, |
254 | }; |
255 | |
256 | |
257 | static int __init ip_vs_dh_init(void) |
258 | { |
259 | return register_ip_vs_scheduler(scheduler: &ip_vs_dh_scheduler); |
260 | } |
261 | |
262 | |
263 | static void __exit ip_vs_dh_cleanup(void) |
264 | { |
265 | unregister_ip_vs_scheduler(scheduler: &ip_vs_dh_scheduler); |
266 | synchronize_rcu(); |
267 | } |
268 | |
269 | |
270 | module_init(ip_vs_dh_init); |
271 | module_exit(ip_vs_dh_cleanup); |
272 | MODULE_LICENSE("GPL" ); |
273 | MODULE_DESCRIPTION("ipvs destination hashing scheduler" ); |
274 | |