Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * IPVS An implementation of the IP virtual server support for the | |
3 | * LINUX operating system. IPVS is now implemented as a module | |
4 | * over the NetFilter framework. IPVS can be used to build a | |
5 | * high-performance and highly available server based on a | |
6 | * cluster of servers. | |
7 | * | |
1da177e4 LT |
8 | * Authors: Wensong Zhang <wensong@linuxvirtualserver.org> |
9 | * Peter Kese <peter.kese@ijs.si> | |
10 | * Julian Anastasov <ja@ssi.bg> | |
11 | * | |
12 | * This program is free software; you can redistribute it and/or | |
13 | * modify it under the terms of the GNU General Public License | |
14 | * as published by the Free Software Foundation; either version | |
15 | * 2 of the License, or (at your option) any later version. | |
16 | * | |
17 | * Changes: | |
18 | * | |
19 | */ | |
20 | ||
9aada7ac HE |
21 | #define KMSG_COMPONENT "IPVS" |
22 | #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt | |
23 | ||
1da177e4 LT |
24 | #include <linux/module.h> |
25 | #include <linux/init.h> | |
26 | #include <linux/types.h> | |
4fc268d2 | 27 | #include <linux/capability.h> |
1da177e4 LT |
28 | #include <linux/fs.h> |
29 | #include <linux/sysctl.h> | |
30 | #include <linux/proc_fs.h> | |
31 | #include <linux/workqueue.h> | |
32 | #include <linux/swap.h> | |
1da177e4 | 33 | #include <linux/seq_file.h> |
5a0e3ad6 | 34 | #include <linux/slab.h> |
1da177e4 LT |
35 | |
36 | #include <linux/netfilter.h> | |
37 | #include <linux/netfilter_ipv4.h> | |
14cc3e2b | 38 | #include <linux/mutex.h> |
1da177e4 | 39 | |
457c4cbc | 40 | #include <net/net_namespace.h> |
9330419d | 41 | #include <linux/nsproxy.h> |
1da177e4 | 42 | #include <net/ip.h> |
09571c7a VB |
43 | #ifdef CONFIG_IP_VS_IPV6 |
44 | #include <net/ipv6.h> | |
45 | #include <net/ip6_route.h> | |
46 | #endif | |
14c85021 | 47 | #include <net/route.h> |
1da177e4 | 48 | #include <net/sock.h> |
9a812198 | 49 | #include <net/genetlink.h> |
1da177e4 LT |
50 | |
51 | #include <asm/uaccess.h> | |
52 | ||
53 | #include <net/ip_vs.h> | |
54 | ||
55 | /* semaphore for IPVS sockopts. And, [gs]etsockopt may sleep. */ | |
14cc3e2b | 56 | static DEFINE_MUTEX(__ip_vs_mutex); |
1da177e4 | 57 | |
1da177e4 | 58 | /* sysctl variables */ |
1da177e4 LT |
59 | |
60 | #ifdef CONFIG_IP_VS_DEBUG | |
61 | static int sysctl_ip_vs_debug_level = 0; | |
62 | ||
63 | int ip_vs_get_debug_level(void) | |
64 | { | |
65 | return sysctl_ip_vs_debug_level; | |
66 | } | |
67 | #endif | |
68 | ||
7a4f0761 HS |
69 | |
70 | /* Protos */ | |
578bc3ef | 71 | static void __ip_vs_del_service(struct ip_vs_service *svc, bool cleanup); |
7a4f0761 HS |
72 | |
73 | ||
09571c7a VB |
74 | #ifdef CONFIG_IP_VS_IPV6 |
75 | /* Taken from rt6_fill_node() in net/ipv6/route.c, is there a better way? */ | |
c24584c0 ED |
76 | static bool __ip_vs_addr_is_local_v6(struct net *net, |
77 | const struct in6_addr *addr) | |
09571c7a | 78 | { |
4c9483b2 DM |
79 | struct flowi6 fl6 = { |
80 | .daddr = *addr, | |
09571c7a | 81 | }; |
c24584c0 ED |
82 | struct dst_entry *dst = ip6_route_output(net, NULL, &fl6); |
83 | bool is_local; | |
09571c7a | 84 | |
c24584c0 | 85 | is_local = !dst->error && dst->dev && (dst->dev->flags & IFF_LOOPBACK); |
09571c7a | 86 | |
c24584c0 ED |
87 | dst_release(dst); |
88 | return is_local; | |
09571c7a VB |
89 | } |
90 | #endif | |
14e40546 SH |
91 | |
92 | #ifdef CONFIG_SYSCTL | |
1da177e4 | 93 | /* |
af9debd4 JA |
94 | * update_defense_level is called from keventd and from sysctl, |
95 | * so it needs to protect itself from softirqs | |
1da177e4 | 96 | */ |
9330419d | 97 | static void update_defense_level(struct netns_ipvs *ipvs) |
1da177e4 LT |
98 | { |
99 | struct sysinfo i; | |
100 | static int old_secure_tcp = 0; | |
101 | int availmem; | |
102 | int nomem; | |
103 | int to_change = -1; | |
104 | ||
105 | /* we only count free and buffered memory (in pages) */ | |
106 | si_meminfo(&i); | |
107 | availmem = i.freeram + i.bufferram; | |
108 | /* however in linux 2.5 the i.bufferram is total page cache size, | |
109 | we need adjust it */ | |
110 | /* si_swapinfo(&i); */ | |
111 | /* availmem = availmem - (i.totalswap - i.freeswap); */ | |
112 | ||
a0840e2e | 113 | nomem = (availmem < ipvs->sysctl_amemthresh); |
1da177e4 | 114 | |
af9debd4 JA |
115 | local_bh_disable(); |
116 | ||
1da177e4 | 117 | /* drop_entry */ |
a0840e2e HS |
118 | spin_lock(&ipvs->dropentry_lock); |
119 | switch (ipvs->sysctl_drop_entry) { | |
1da177e4 | 120 | case 0: |
a0840e2e | 121 | atomic_set(&ipvs->dropentry, 0); |
1da177e4 LT |
122 | break; |
123 | case 1: | |
124 | if (nomem) { | |
a0840e2e HS |
125 | atomic_set(&ipvs->dropentry, 1); |
126 | ipvs->sysctl_drop_entry = 2; | |
1da177e4 | 127 | } else { |
a0840e2e | 128 | atomic_set(&ipvs->dropentry, 0); |
1da177e4 LT |
129 | } |
130 | break; | |
131 | case 2: | |
132 | if (nomem) { | |
a0840e2e | 133 | atomic_set(&ipvs->dropentry, 1); |
1da177e4 | 134 | } else { |
a0840e2e HS |
135 | atomic_set(&ipvs->dropentry, 0); |
136 | ipvs->sysctl_drop_entry = 1; | |
1da177e4 LT |
137 | }; |
138 | break; | |
139 | case 3: | |
a0840e2e | 140 | atomic_set(&ipvs->dropentry, 1); |
1da177e4 LT |
141 | break; |
142 | } | |
a0840e2e | 143 | spin_unlock(&ipvs->dropentry_lock); |
1da177e4 LT |
144 | |
145 | /* drop_packet */ | |
a0840e2e HS |
146 | spin_lock(&ipvs->droppacket_lock); |
147 | switch (ipvs->sysctl_drop_packet) { | |
1da177e4 | 148 | case 0: |
a0840e2e | 149 | ipvs->drop_rate = 0; |
1da177e4 LT |
150 | break; |
151 | case 1: | |
152 | if (nomem) { | |
a0840e2e HS |
153 | ipvs->drop_rate = ipvs->drop_counter |
154 | = ipvs->sysctl_amemthresh / | |
155 | (ipvs->sysctl_amemthresh-availmem); | |
156 | ipvs->sysctl_drop_packet = 2; | |
1da177e4 | 157 | } else { |
a0840e2e | 158 | ipvs->drop_rate = 0; |
1da177e4 LT |
159 | } |
160 | break; | |
161 | case 2: | |
162 | if (nomem) { | |
a0840e2e HS |
163 | ipvs->drop_rate = ipvs->drop_counter |
164 | = ipvs->sysctl_amemthresh / | |
165 | (ipvs->sysctl_amemthresh-availmem); | |
1da177e4 | 166 | } else { |
a0840e2e HS |
167 | ipvs->drop_rate = 0; |
168 | ipvs->sysctl_drop_packet = 1; | |
1da177e4 LT |
169 | } |
170 | break; | |
171 | case 3: | |
a0840e2e | 172 | ipvs->drop_rate = ipvs->sysctl_am_droprate; |
1da177e4 LT |
173 | break; |
174 | } | |
a0840e2e | 175 | spin_unlock(&ipvs->droppacket_lock); |
1da177e4 LT |
176 | |
177 | /* secure_tcp */ | |
a0840e2e HS |
178 | spin_lock(&ipvs->securetcp_lock); |
179 | switch (ipvs->sysctl_secure_tcp) { | |
1da177e4 LT |
180 | case 0: |
181 | if (old_secure_tcp >= 2) | |
182 | to_change = 0; | |
183 | break; | |
184 | case 1: | |
185 | if (nomem) { | |
186 | if (old_secure_tcp < 2) | |
187 | to_change = 1; | |
a0840e2e | 188 | ipvs->sysctl_secure_tcp = 2; |
1da177e4 LT |
189 | } else { |
190 | if (old_secure_tcp >= 2) | |
191 | to_change = 0; | |
192 | } | |
193 | break; | |
194 | case 2: | |
195 | if (nomem) { | |
196 | if (old_secure_tcp < 2) | |
197 | to_change = 1; | |
198 | } else { | |
199 | if (old_secure_tcp >= 2) | |
200 | to_change = 0; | |
a0840e2e | 201 | ipvs->sysctl_secure_tcp = 1; |
1da177e4 LT |
202 | } |
203 | break; | |
204 | case 3: | |
205 | if (old_secure_tcp < 2) | |
206 | to_change = 1; | |
207 | break; | |
208 | } | |
a0840e2e | 209 | old_secure_tcp = ipvs->sysctl_secure_tcp; |
1da177e4 | 210 | if (to_change >= 0) |
9330419d | 211 | ip_vs_protocol_timeout_change(ipvs, |
a0840e2e HS |
212 | ipvs->sysctl_secure_tcp > 1); |
213 | spin_unlock(&ipvs->securetcp_lock); | |
af9debd4 JA |
214 | |
215 | local_bh_enable(); | |
1da177e4 LT |
216 | } |
217 | ||
218 | ||
219 | /* | |
220 | * Timer for checking the defense | |
221 | */ | |
222 | #define DEFENSE_TIMER_PERIOD 1*HZ | |
1da177e4 | 223 | |
c4028958 | 224 | static void defense_work_handler(struct work_struct *work) |
1da177e4 | 225 | { |
f6340ee0 HS |
226 | struct netns_ipvs *ipvs = |
227 | container_of(work, struct netns_ipvs, defense_work.work); | |
9330419d HS |
228 | |
229 | update_defense_level(ipvs); | |
a0840e2e | 230 | if (atomic_read(&ipvs->dropentry)) |
423b5595 | 231 | ip_vs_random_dropentry(ipvs); |
f6340ee0 | 232 | schedule_delayed_work(&ipvs->defense_work, DEFENSE_TIMER_PERIOD); |
1da177e4 | 233 | } |
14e40546 | 234 | #endif |
1da177e4 LT |
235 | |
236 | int | |
237 | ip_vs_use_count_inc(void) | |
238 | { | |
239 | return try_module_get(THIS_MODULE); | |
240 | } | |
241 | ||
242 | void | |
243 | ip_vs_use_count_dec(void) | |
244 | { | |
245 | module_put(THIS_MODULE); | |
246 | } | |
247 | ||
248 | ||
249 | /* | |
250 | * Hash table: for virtual service lookups | |
251 | */ | |
252 | #define IP_VS_SVC_TAB_BITS 8 | |
253 | #define IP_VS_SVC_TAB_SIZE (1 << IP_VS_SVC_TAB_BITS) | |
254 | #define IP_VS_SVC_TAB_MASK (IP_VS_SVC_TAB_SIZE - 1) | |
255 | ||
256 | /* the service table hashed by <protocol, addr, port> */ | |
ceec4c38 | 257 | static struct hlist_head ip_vs_svc_table[IP_VS_SVC_TAB_SIZE]; |
1da177e4 | 258 | /* the service table hashed by fwmark */ |
ceec4c38 | 259 | static struct hlist_head ip_vs_svc_fwm_table[IP_VS_SVC_TAB_SIZE]; |
1da177e4 | 260 | |
1da177e4 LT |
261 | |
262 | /* | |
263 | * Returns hash value for virtual service | |
264 | */ | |
95c96174 | 265 | static inline unsigned int |
ba61f390 | 266 | ip_vs_svc_hashkey(struct netns_ipvs *ipvs, int af, unsigned int proto, |
fc723250 | 267 | const union nf_inet_addr *addr, __be16 port) |
1da177e4 | 268 | { |
95c96174 | 269 | register unsigned int porth = ntohs(port); |
b18610de | 270 | __be32 addr_fold = addr->ip; |
e9836f24 | 271 | __u32 ahash; |
1da177e4 | 272 | |
b18610de JV |
273 | #ifdef CONFIG_IP_VS_IPV6 |
274 | if (af == AF_INET6) | |
275 | addr_fold = addr->ip6[0]^addr->ip6[1]^ | |
276 | addr->ip6[2]^addr->ip6[3]; | |
277 | #endif | |
e9836f24 | 278 | ahash = ntohl(addr_fold); |
ba61f390 | 279 | ahash ^= ((size_t) ipvs >> 8); |
b18610de | 280 | |
e9836f24 JA |
281 | return (proto ^ ahash ^ (porth >> IP_VS_SVC_TAB_BITS) ^ porth) & |
282 | IP_VS_SVC_TAB_MASK; | |
1da177e4 LT |
283 | } |
284 | ||
285 | /* | |
286 | * Returns hash value of fwmark for virtual service lookup | |
287 | */ | |
f6510b24 | 288 | static inline unsigned int ip_vs_svc_fwm_hashkey(struct netns_ipvs *ipvs, __u32 fwmark) |
1da177e4 | 289 | { |
f6510b24 | 290 | return (((size_t)ipvs>>8) ^ fwmark) & IP_VS_SVC_TAB_MASK; |
1da177e4 LT |
291 | } |
292 | ||
293 | /* | |
fc723250 | 294 | * Hashes a service in the ip_vs_svc_table by <netns,proto,addr,port> |
1da177e4 LT |
295 | * or in the ip_vs_svc_fwm_table by fwmark. |
296 | * Should be called with locked tables. | |
297 | */ | |
298 | static int ip_vs_svc_hash(struct ip_vs_service *svc) | |
299 | { | |
95c96174 | 300 | unsigned int hash; |
1da177e4 LT |
301 | |
302 | if (svc->flags & IP_VS_SVC_F_HASHED) { | |
1e3e238e HE |
303 | pr_err("%s(): request for already hashed, called from %pF\n", |
304 | __func__, __builtin_return_address(0)); | |
1da177e4 LT |
305 | return 0; |
306 | } | |
307 | ||
308 | if (svc->fwmark == 0) { | |
309 | /* | |
fc723250 | 310 | * Hash it by <netns,protocol,addr,port> in ip_vs_svc_table |
1da177e4 | 311 | */ |
ba61f390 | 312 | hash = ip_vs_svc_hashkey(svc->ipvs, svc->af, svc->protocol, |
fc723250 | 313 | &svc->addr, svc->port); |
ceec4c38 | 314 | hlist_add_head_rcu(&svc->s_list, &ip_vs_svc_table[hash]); |
1da177e4 LT |
315 | } else { |
316 | /* | |
fc723250 | 317 | * Hash it by fwmark in svc_fwm_table |
1da177e4 | 318 | */ |
f6510b24 | 319 | hash = ip_vs_svc_fwm_hashkey(svc->ipvs, svc->fwmark); |
ceec4c38 | 320 | hlist_add_head_rcu(&svc->f_list, &ip_vs_svc_fwm_table[hash]); |
1da177e4 LT |
321 | } |
322 | ||
323 | svc->flags |= IP_VS_SVC_F_HASHED; | |
324 | /* increase its refcnt because it is referenced by the svc table */ | |
325 | atomic_inc(&svc->refcnt); | |
326 | return 1; | |
327 | } | |
328 | ||
329 | ||
330 | /* | |
fc723250 | 331 | * Unhashes a service from svc_table / svc_fwm_table. |
1da177e4 LT |
332 | * Should be called with locked tables. |
333 | */ | |
334 | static int ip_vs_svc_unhash(struct ip_vs_service *svc) | |
335 | { | |
336 | if (!(svc->flags & IP_VS_SVC_F_HASHED)) { | |
1e3e238e HE |
337 | pr_err("%s(): request for unhash flagged, called from %pF\n", |
338 | __func__, __builtin_return_address(0)); | |
1da177e4 LT |
339 | return 0; |
340 | } | |
341 | ||
342 | if (svc->fwmark == 0) { | |
fc723250 | 343 | /* Remove it from the svc_table table */ |
ceec4c38 | 344 | hlist_del_rcu(&svc->s_list); |
1da177e4 | 345 | } else { |
fc723250 | 346 | /* Remove it from the svc_fwm_table table */ |
ceec4c38 | 347 | hlist_del_rcu(&svc->f_list); |
1da177e4 LT |
348 | } |
349 | ||
350 | svc->flags &= ~IP_VS_SVC_F_HASHED; | |
351 | atomic_dec(&svc->refcnt); | |
352 | return 1; | |
353 | } | |
354 | ||
355 | ||
356 | /* | |
fc723250 | 357 | * Get service by {netns, proto,addr,port} in the service table. |
1da177e4 | 358 | */ |
b18610de | 359 | static inline struct ip_vs_service * |
bb2e2a8c | 360 | __ip_vs_service_find(struct netns_ipvs *ipvs, int af, __u16 protocol, |
fc723250 | 361 | const union nf_inet_addr *vaddr, __be16 vport) |
1da177e4 | 362 | { |
95c96174 | 363 | unsigned int hash; |
1da177e4 LT |
364 | struct ip_vs_service *svc; |
365 | ||
366 | /* Check for "full" addressed entries */ | |
ba61f390 | 367 | hash = ip_vs_svc_hashkey(ipvs, af, protocol, vaddr, vport); |
1da177e4 | 368 | |
ceec4c38 | 369 | hlist_for_each_entry_rcu(svc, &ip_vs_svc_table[hash], s_list) { |
b18610de JV |
370 | if ((svc->af == af) |
371 | && ip_vs_addr_equal(af, &svc->addr, vaddr) | |
1da177e4 | 372 | && (svc->port == vport) |
fc723250 | 373 | && (svc->protocol == protocol) |
3109d2f2 | 374 | && (svc->ipvs == ipvs)) { |
1da177e4 | 375 | /* HIT */ |
1da177e4 LT |
376 | return svc; |
377 | } | |
378 | } | |
379 | ||
380 | return NULL; | |
381 | } | |
382 | ||
383 | ||
384 | /* | |
385 | * Get service by {fwmark} in the service table. | |
386 | */ | |
b18610de | 387 | static inline struct ip_vs_service * |
1ed8b947 | 388 | __ip_vs_svc_fwm_find(struct netns_ipvs *ipvs, int af, __u32 fwmark) |
1da177e4 | 389 | { |
95c96174 | 390 | unsigned int hash; |
1da177e4 LT |
391 | struct ip_vs_service *svc; |
392 | ||
393 | /* Check for fwmark addressed entries */ | |
f6510b24 | 394 | hash = ip_vs_svc_fwm_hashkey(ipvs, fwmark); |
1da177e4 | 395 | |
ceec4c38 | 396 | hlist_for_each_entry_rcu(svc, &ip_vs_svc_fwm_table[hash], f_list) { |
fc723250 | 397 | if (svc->fwmark == fwmark && svc->af == af |
3109d2f2 | 398 | && (svc->ipvs == ipvs)) { |
1da177e4 | 399 | /* HIT */ |
1da177e4 LT |
400 | return svc; |
401 | } | |
402 | } | |
403 | ||
404 | return NULL; | |
405 | } | |
406 | ||
ceec4c38 | 407 | /* Find service, called under RCU lock */ |
1da177e4 | 408 | struct ip_vs_service * |
0a4fd6ce | 409 | ip_vs_service_find(struct netns_ipvs *ipvs, int af, __u32 fwmark, __u16 protocol, |
ceec4c38 | 410 | const union nf_inet_addr *vaddr, __be16 vport) |
1da177e4 LT |
411 | { |
412 | struct ip_vs_service *svc; | |
3c2e0505 | 413 | |
1da177e4 LT |
414 | /* |
415 | * Check the table hashed by fwmark first | |
416 | */ | |
097fc76a | 417 | if (fwmark) { |
1ed8b947 | 418 | svc = __ip_vs_svc_fwm_find(ipvs, af, fwmark); |
097fc76a JA |
419 | if (svc) |
420 | goto out; | |
421 | } | |
1da177e4 LT |
422 | |
423 | /* | |
424 | * Check the table hashed by <protocol,addr,port> | |
425 | * for "full" addressed entries | |
426 | */ | |
bb2e2a8c | 427 | svc = __ip_vs_service_find(ipvs, af, protocol, vaddr, vport); |
1da177e4 LT |
428 | |
429 | if (svc == NULL | |
430 | && protocol == IPPROTO_TCP | |
763f8d0e | 431 | && atomic_read(&ipvs->ftpsvc_counter) |
1da177e4 LT |
432 | && (vport == FTPDATA || ntohs(vport) >= PROT_SOCK)) { |
433 | /* | |
434 | * Check if ftp service entry exists, the packet | |
435 | * might belong to FTP data connections. | |
436 | */ | |
bb2e2a8c | 437 | svc = __ip_vs_service_find(ipvs, af, protocol, vaddr, FTPPORT); |
1da177e4 LT |
438 | } |
439 | ||
440 | if (svc == NULL | |
763f8d0e | 441 | && atomic_read(&ipvs->nullsvc_counter)) { |
1da177e4 LT |
442 | /* |
443 | * Check if the catch-all port (port zero) exists | |
444 | */ | |
bb2e2a8c | 445 | svc = __ip_vs_service_find(ipvs, af, protocol, vaddr, 0); |
1da177e4 LT |
446 | } |
447 | ||
448 | out: | |
3c2e0505 JV |
449 | IP_VS_DBG_BUF(9, "lookup service: fwm %u %s %s:%u %s\n", |
450 | fwmark, ip_vs_proto_name(protocol), | |
451 | IP_VS_DBG_ADDR(af, vaddr), ntohs(vport), | |
452 | svc ? "hit" : "not hit"); | |
1da177e4 LT |
453 | |
454 | return svc; | |
455 | } | |
456 | ||
457 | ||
458 | static inline void | |
459 | __ip_vs_bind_svc(struct ip_vs_dest *dest, struct ip_vs_service *svc) | |
460 | { | |
461 | atomic_inc(&svc->refcnt); | |
bcbde4c0 | 462 | rcu_assign_pointer(dest->svc, svc); |
1da177e4 LT |
463 | } |
464 | ||
ceec4c38 JA |
465 | static void ip_vs_service_free(struct ip_vs_service *svc) |
466 | { | |
982f4051 | 467 | free_percpu(svc->stats.cpustats); |
ceec4c38 JA |
468 | kfree(svc); |
469 | } | |
470 | ||
bcbde4c0 | 471 | static void ip_vs_service_rcu_free(struct rcu_head *head) |
1da177e4 | 472 | { |
bcbde4c0 JA |
473 | struct ip_vs_service *svc; |
474 | ||
475 | svc = container_of(head, struct ip_vs_service, rcu_head); | |
476 | ip_vs_service_free(svc); | |
477 | } | |
1da177e4 | 478 | |
bcbde4c0 JA |
479 | static void __ip_vs_svc_put(struct ip_vs_service *svc, bool do_delay) |
480 | { | |
26c15cfd | 481 | if (atomic_dec_and_test(&svc->refcnt)) { |
ceec4c38 | 482 | IP_VS_DBG_BUF(3, "Removing service %u/%s:%u\n", |
26c15cfd JA |
483 | svc->fwmark, |
484 | IP_VS_DBG_ADDR(svc->af, &svc->addr), | |
ceec4c38 | 485 | ntohs(svc->port)); |
bcbde4c0 JA |
486 | if (do_delay) |
487 | call_rcu(&svc->rcu_head, ip_vs_service_rcu_free); | |
488 | else | |
489 | ip_vs_service_free(svc); | |
26c15cfd | 490 | } |
1da177e4 LT |
491 | } |
492 | ||
493 | ||
494 | /* | |
495 | * Returns hash value for real service | |
496 | */ | |
95c96174 | 497 | static inline unsigned int ip_vs_rs_hashkey(int af, |
7937df15 JV |
498 | const union nf_inet_addr *addr, |
499 | __be16 port) | |
1da177e4 | 500 | { |
95c96174 | 501 | register unsigned int porth = ntohs(port); |
7937df15 JV |
502 | __be32 addr_fold = addr->ip; |
503 | ||
504 | #ifdef CONFIG_IP_VS_IPV6 | |
505 | if (af == AF_INET6) | |
506 | addr_fold = addr->ip6[0]^addr->ip6[1]^ | |
507 | addr->ip6[2]^addr->ip6[3]; | |
508 | #endif | |
1da177e4 | 509 | |
7937df15 | 510 | return (ntohl(addr_fold)^(porth>>IP_VS_RTAB_BITS)^porth) |
1da177e4 LT |
511 | & IP_VS_RTAB_MASK; |
512 | } | |
513 | ||
276472ea JA |
514 | /* Hash ip_vs_dest in rs_table by <proto,addr,port>. */ |
515 | static void ip_vs_rs_hash(struct netns_ipvs *ipvs, struct ip_vs_dest *dest) | |
1da177e4 | 516 | { |
95c96174 | 517 | unsigned int hash; |
1da177e4 | 518 | |
276472ea JA |
519 | if (dest->in_rs_table) |
520 | return; | |
1da177e4 LT |
521 | |
522 | /* | |
523 | * Hash by proto,addr,port, | |
524 | * which are the parameters of the real service. | |
525 | */ | |
7937df15 JV |
526 | hash = ip_vs_rs_hashkey(dest->af, &dest->addr, dest->port); |
527 | ||
276472ea JA |
528 | hlist_add_head_rcu(&dest->d_list, &ipvs->rs_table[hash]); |
529 | dest->in_rs_table = 1; | |
1da177e4 LT |
530 | } |
531 | ||
276472ea JA |
532 | /* Unhash ip_vs_dest from rs_table. */ |
533 | static void ip_vs_rs_unhash(struct ip_vs_dest *dest) | |
1da177e4 LT |
534 | { |
535 | /* | |
fc723250 | 536 | * Remove it from the rs_table table. |
1da177e4 | 537 | */ |
276472ea JA |
538 | if (dest->in_rs_table) { |
539 | hlist_del_rcu(&dest->d_list); | |
540 | dest->in_rs_table = 0; | |
1da177e4 | 541 | } |
1da177e4 LT |
542 | } |
543 | ||
276472ea | 544 | /* Check if real service by <proto,addr,port> is present */ |
48aed1b0 | 545 | bool ip_vs_has_real_service(struct netns_ipvs *ipvs, int af, __u16 protocol, |
276472ea | 546 | const union nf_inet_addr *daddr, __be16 dport) |
1da177e4 | 547 | { |
95c96174 | 548 | unsigned int hash; |
1da177e4 LT |
549 | struct ip_vs_dest *dest; |
550 | ||
276472ea | 551 | /* Check for "full" addressed entries */ |
7937df15 | 552 | hash = ip_vs_rs_hashkey(af, daddr, dport); |
1da177e4 | 553 | |
276472ea JA |
554 | rcu_read_lock(); |
555 | hlist_for_each_entry_rcu(dest, &ipvs->rs_table[hash], d_list) { | |
556 | if (dest->port == dport && | |
557 | dest->af == af && | |
558 | ip_vs_addr_equal(af, &dest->addr, daddr) && | |
559 | (dest->protocol == protocol || dest->vfwmark)) { | |
1da177e4 | 560 | /* HIT */ |
276472ea JA |
561 | rcu_read_unlock(); |
562 | return true; | |
1da177e4 LT |
563 | } |
564 | } | |
276472ea | 565 | rcu_read_unlock(); |
1da177e4 | 566 | |
276472ea | 567 | return false; |
1da177e4 LT |
568 | } |
569 | ||
413c2d04 JA |
570 | /* Lookup destination by {addr,port} in the given service |
571 | * Called under RCU lock. | |
1da177e4 LT |
572 | */ |
573 | static struct ip_vs_dest * | |
655eef10 AG |
574 | ip_vs_lookup_dest(struct ip_vs_service *svc, int dest_af, |
575 | const union nf_inet_addr *daddr, __be16 dport) | |
1da177e4 LT |
576 | { |
577 | struct ip_vs_dest *dest; | |
578 | ||
579 | /* | |
580 | * Find the destination for the given service | |
581 | */ | |
413c2d04 | 582 | list_for_each_entry_rcu(dest, &svc->destinations, n_list) { |
655eef10 AG |
583 | if ((dest->af == dest_af) && |
584 | ip_vs_addr_equal(dest_af, &dest->addr, daddr) && | |
585 | (dest->port == dport)) { | |
1da177e4 LT |
586 | /* HIT */ |
587 | return dest; | |
588 | } | |
589 | } | |
590 | ||
591 | return NULL; | |
592 | } | |
593 | ||
1e356f9c RB |
594 | /* |
595 | * Find destination by {daddr,dport,vaddr,protocol} | |
413c2d04 | 596 | * Created to be used in ip_vs_process_message() in |
1e356f9c RB |
597 | * the backup synchronization daemon. It finds the |
598 | * destination to be bound to the received connection | |
599 | * on the backup. | |
413c2d04 | 600 | * Called under RCU lock, no refcnt is returned. |
1e356f9c | 601 | */ |
dc2add6f | 602 | struct ip_vs_dest *ip_vs_find_dest(struct netns_ipvs *ipvs, int svc_af, int dest_af, |
fc723250 | 603 | const union nf_inet_addr *daddr, |
7937df15 JV |
604 | __be16 dport, |
605 | const union nf_inet_addr *vaddr, | |
52793dbe JA |
606 | __be16 vport, __u16 protocol, __u32 fwmark, |
607 | __u32 flags) | |
1e356f9c RB |
608 | { |
609 | struct ip_vs_dest *dest; | |
610 | struct ip_vs_service *svc; | |
52793dbe | 611 | __be16 port = dport; |
1e356f9c | 612 | |
dc2add6f | 613 | svc = ip_vs_service_find(ipvs, svc_af, fwmark, protocol, vaddr, vport); |
1e356f9c RB |
614 | if (!svc) |
615 | return NULL; | |
52793dbe JA |
616 | if (fwmark && (flags & IP_VS_CONN_F_FWD_MASK) != IP_VS_CONN_F_MASQ) |
617 | port = 0; | |
655eef10 | 618 | dest = ip_vs_lookup_dest(svc, dest_af, daddr, port); |
52793dbe | 619 | if (!dest) |
655eef10 | 620 | dest = ip_vs_lookup_dest(svc, dest_af, daddr, port ^ dport); |
1e356f9c RB |
621 | return dest; |
622 | } | |
1da177e4 | 623 | |
026ace06 JA |
624 | void ip_vs_dest_dst_rcu_free(struct rcu_head *head) |
625 | { | |
626 | struct ip_vs_dest_dst *dest_dst = container_of(head, | |
627 | struct ip_vs_dest_dst, | |
628 | rcu_head); | |
629 | ||
630 | dst_release(dest_dst->dst_cache); | |
631 | kfree(dest_dst); | |
632 | } | |
633 | ||
634 | /* Release dest_dst and dst_cache for dest in user context */ | |
d1deae4d JA |
635 | static void __ip_vs_dst_cache_reset(struct ip_vs_dest *dest) |
636 | { | |
026ace06 | 637 | struct ip_vs_dest_dst *old; |
d1deae4d | 638 | |
026ace06 JA |
639 | old = rcu_dereference_protected(dest->dest_dst, 1); |
640 | if (old) { | |
641 | RCU_INIT_POINTER(dest->dest_dst, NULL); | |
642 | call_rcu(&old->rcu_head, ip_vs_dest_dst_rcu_free); | |
643 | } | |
d1deae4d JA |
644 | } |
645 | ||
1da177e4 LT |
646 | /* |
647 | * Lookup dest by {svc,addr,port} in the destination trash. | |
648 | * The destination trash is used to hold the destinations that are removed | |
649 | * from the service table but are still referenced by some conn entries. | |
650 | * The reason to add the destination trash is when the dest is temporary | |
651 | * down (either by administrator or by monitor program), the dest can be | |
652 | * picked back from the trash, the remaining connections to the dest can | |
653 | * continue, and the counting information of the dest is also useful for | |
654 | * scheduling. | |
655 | */ | |
656 | static struct ip_vs_dest * | |
ad147aa4 AG |
657 | ip_vs_trash_get_dest(struct ip_vs_service *svc, int dest_af, |
658 | const union nf_inet_addr *daddr, __be16 dport) | |
1da177e4 | 659 | { |
578bc3ef | 660 | struct ip_vs_dest *dest; |
3109d2f2 | 661 | struct netns_ipvs *ipvs = svc->ipvs; |
1da177e4 LT |
662 | |
663 | /* | |
664 | * Find the destination in trash | |
665 | */ | |
578bc3ef JA |
666 | spin_lock_bh(&ipvs->dest_trash_lock); |
667 | list_for_each_entry(dest, &ipvs->dest_trash, t_list) { | |
7937df15 JV |
668 | IP_VS_DBG_BUF(3, "Destination %u/%s:%u still in trash, " |
669 | "dest->refcnt=%d\n", | |
670 | dest->vfwmark, | |
ad147aa4 | 671 | IP_VS_DBG_ADDR(dest->af, &dest->addr), |
7937df15 JV |
672 | ntohs(dest->port), |
673 | atomic_read(&dest->refcnt)); | |
ad147aa4 AG |
674 | if (dest->af == dest_af && |
675 | ip_vs_addr_equal(dest_af, &dest->addr, daddr) && | |
1da177e4 LT |
676 | dest->port == dport && |
677 | dest->vfwmark == svc->fwmark && | |
678 | dest->protocol == svc->protocol && | |
679 | (svc->fwmark || | |
7937df15 | 680 | (ip_vs_addr_equal(svc->af, &dest->vaddr, &svc->addr) && |
1da177e4 LT |
681 | dest->vport == svc->port))) { |
682 | /* HIT */ | |
578bc3ef JA |
683 | list_del(&dest->t_list); |
684 | ip_vs_dest_hold(dest); | |
685 | goto out; | |
1da177e4 LT |
686 | } |
687 | } | |
688 | ||
578bc3ef JA |
689 | dest = NULL; |
690 | ||
691 | out: | |
692 | spin_unlock_bh(&ipvs->dest_trash_lock); | |
693 | ||
694 | return dest; | |
1da177e4 LT |
695 | } |
696 | ||
578bc3ef JA |
697 | static void ip_vs_dest_free(struct ip_vs_dest *dest) |
698 | { | |
bcbde4c0 JA |
699 | struct ip_vs_service *svc = rcu_dereference_protected(dest->svc, 1); |
700 | ||
578bc3ef | 701 | __ip_vs_dst_cache_reset(dest); |
bcbde4c0 | 702 | __ip_vs_svc_put(svc, false); |
578bc3ef | 703 | free_percpu(dest->stats.cpustats); |
9e4e948a | 704 | ip_vs_dest_put_and_free(dest); |
578bc3ef | 705 | } |
1da177e4 LT |
706 | |
707 | /* | |
708 | * Clean up all the destinations in the trash | |
709 | * Called by the ip_vs_control_cleanup() | |
710 | * | |
711 | * When the ip_vs_control_clearup is activated by ipvs module exit, | |
712 | * the service tables must have been flushed and all the connections | |
713 | * are expired, and the refcnt of each destination in the trash must | |
578bc3ef | 714 | * be 0, so we simply release them here. |
1da177e4 | 715 | */ |
6c0e14f5 | 716 | static void ip_vs_trash_cleanup(struct netns_ipvs *ipvs) |
1da177e4 LT |
717 | { |
718 | struct ip_vs_dest *dest, *nxt; | |
719 | ||
578bc3ef JA |
720 | del_timer_sync(&ipvs->dest_trash_timer); |
721 | /* No need to use dest_trash_lock */ | |
722 | list_for_each_entry_safe(dest, nxt, &ipvs->dest_trash, t_list) { | |
723 | list_del(&dest->t_list); | |
724 | ip_vs_dest_free(dest); | |
1da177e4 LT |
725 | } |
726 | } | |
727 | ||
55a3d4e1 | 728 | static void |
cd67cd5e | 729 | ip_vs_copy_stats(struct ip_vs_kstats *dst, struct ip_vs_stats *src) |
55a3d4e1 | 730 | { |
cd67cd5e | 731 | #define IP_VS_SHOW_STATS_COUNTER(c) dst->c = src->kstats.c - src->kstats0.c |
55a3d4e1 JA |
732 | |
733 | spin_lock_bh(&src->lock); | |
734 | ||
735 | IP_VS_SHOW_STATS_COUNTER(conns); | |
736 | IP_VS_SHOW_STATS_COUNTER(inpkts); | |
737 | IP_VS_SHOW_STATS_COUNTER(outpkts); | |
738 | IP_VS_SHOW_STATS_COUNTER(inbytes); | |
739 | IP_VS_SHOW_STATS_COUNTER(outbytes); | |
740 | ||
ea9f22cc | 741 | ip_vs_read_estimator(dst, src); |
55a3d4e1 JA |
742 | |
743 | spin_unlock_bh(&src->lock); | |
744 | } | |
1da177e4 | 745 | |
cd67cd5e JA |
746 | static void |
747 | ip_vs_export_stats_user(struct ip_vs_stats_user *dst, struct ip_vs_kstats *src) | |
748 | { | |
749 | dst->conns = (u32)src->conns; | |
750 | dst->inpkts = (u32)src->inpkts; | |
751 | dst->outpkts = (u32)src->outpkts; | |
752 | dst->inbytes = src->inbytes; | |
753 | dst->outbytes = src->outbytes; | |
754 | dst->cps = (u32)src->cps; | |
755 | dst->inpps = (u32)src->inpps; | |
756 | dst->outpps = (u32)src->outpps; | |
757 | dst->inbps = (u32)src->inbps; | |
758 | dst->outbps = (u32)src->outbps; | |
759 | } | |
760 | ||
1da177e4 LT |
761 | static void |
762 | ip_vs_zero_stats(struct ip_vs_stats *stats) | |
763 | { | |
764 | spin_lock_bh(&stats->lock); | |
e93615d0 | 765 | |
55a3d4e1 JA |
766 | /* get current counters as zero point, rates are zeroed */ |
767 | ||
cd67cd5e | 768 | #define IP_VS_ZERO_STATS_COUNTER(c) stats->kstats0.c = stats->kstats.c |
55a3d4e1 JA |
769 | |
770 | IP_VS_ZERO_STATS_COUNTER(conns); | |
771 | IP_VS_ZERO_STATS_COUNTER(inpkts); | |
772 | IP_VS_ZERO_STATS_COUNTER(outpkts); | |
773 | IP_VS_ZERO_STATS_COUNTER(inbytes); | |
774 | IP_VS_ZERO_STATS_COUNTER(outbytes); | |
775 | ||
1da177e4 | 776 | ip_vs_zero_estimator(stats); |
e93615d0 | 777 | |
3a14a313 | 778 | spin_unlock_bh(&stats->lock); |
1da177e4 LT |
779 | } |
780 | ||
781 | /* | |
782 | * Update a destination in the given service | |
783 | */ | |
784 | static void | |
26c15cfd JA |
785 | __ip_vs_update_dest(struct ip_vs_service *svc, struct ip_vs_dest *dest, |
786 | struct ip_vs_dest_user_kern *udest, int add) | |
1da177e4 | 787 | { |
3109d2f2 | 788 | struct netns_ipvs *ipvs = svc->ipvs; |
bcbde4c0 | 789 | struct ip_vs_service *old_svc; |
ceec4c38 | 790 | struct ip_vs_scheduler *sched; |
1da177e4 LT |
791 | int conn_flags; |
792 | ||
391f503d AG |
793 | /* We cannot modify an address and change the address family */ |
794 | BUG_ON(!add && udest->af != dest->af); | |
795 | ||
796 | if (add && udest->af != svc->af) | |
797 | ipvs->mixed_address_family_dests++; | |
798 | ||
1da177e4 LT |
799 | /* set the weight and the flags */ |
800 | atomic_set(&dest->weight, udest->weight); | |
3575792e JA |
801 | conn_flags = udest->conn_flags & IP_VS_CONN_F_DEST_MASK; |
802 | conn_flags |= IP_VS_CONN_F_INACTIVE; | |
1da177e4 | 803 | |
1da177e4 | 804 | /* set the IP_VS_CONN_F_NOOUTPUT flag if not masquerading/NAT */ |
3575792e | 805 | if ((conn_flags & IP_VS_CONN_F_FWD_MASK) != IP_VS_CONN_F_MASQ) { |
1da177e4 LT |
806 | conn_flags |= IP_VS_CONN_F_NOOUTPUT; |
807 | } else { | |
808 | /* | |
fc723250 | 809 | * Put the real service in rs_table if not present. |
1da177e4 LT |
810 | * For now only for NAT! |
811 | */ | |
fc723250 | 812 | ip_vs_rs_hash(ipvs, dest); |
1da177e4 LT |
813 | } |
814 | atomic_set(&dest->conn_flags, conn_flags); | |
815 | ||
816 | /* bind the service */ | |
bcbde4c0 JA |
817 | old_svc = rcu_dereference_protected(dest->svc, 1); |
818 | if (!old_svc) { | |
1da177e4 LT |
819 | __ip_vs_bind_svc(dest, svc); |
820 | } else { | |
bcbde4c0 | 821 | if (old_svc != svc) { |
1da177e4 LT |
822 | ip_vs_zero_stats(&dest->stats); |
823 | __ip_vs_bind_svc(dest, svc); | |
bcbde4c0 | 824 | __ip_vs_svc_put(old_svc, true); |
1da177e4 LT |
825 | } |
826 | } | |
827 | ||
828 | /* set the dest status flags */ | |
829 | dest->flags |= IP_VS_DEST_F_AVAILABLE; | |
830 | ||
831 | if (udest->u_threshold == 0 || udest->u_threshold > dest->u_threshold) | |
832 | dest->flags &= ~IP_VS_DEST_F_OVERLOAD; | |
833 | dest->u_threshold = udest->u_threshold; | |
834 | dest->l_threshold = udest->l_threshold; | |
26c15cfd | 835 | |
6cff339b AG |
836 | dest->af = udest->af; |
837 | ||
ff75f40f | 838 | spin_lock_bh(&dest->dst_lock); |
d1deae4d | 839 | __ip_vs_dst_cache_reset(dest); |
ff75f40f | 840 | spin_unlock_bh(&dest->dst_lock); |
fc604767 | 841 | |
26c15cfd | 842 | if (add) { |
0f34d54b | 843 | ip_vs_start_estimator(svc->ipvs, &dest->stats); |
413c2d04 | 844 | list_add_rcu(&dest->n_list, &svc->destinations); |
26c15cfd | 845 | svc->num_dests++; |
05f00505 JA |
846 | sched = rcu_dereference_protected(svc->scheduler, 1); |
847 | if (sched && sched->add_dest) | |
ceec4c38 | 848 | sched->add_dest(svc, dest); |
6b6df466 | 849 | } else { |
05f00505 JA |
850 | sched = rcu_dereference_protected(svc->scheduler, 1); |
851 | if (sched && sched->upd_dest) | |
ceec4c38 | 852 | sched->upd_dest(svc, dest); |
26c15cfd | 853 | } |
1da177e4 LT |
854 | } |
855 | ||
856 | ||
857 | /* | |
858 | * Create a destination for the given service | |
859 | */ | |
860 | static int | |
c860c6b1 | 861 | ip_vs_new_dest(struct ip_vs_service *svc, struct ip_vs_dest_user_kern *udest, |
1da177e4 LT |
862 | struct ip_vs_dest **dest_p) |
863 | { | |
864 | struct ip_vs_dest *dest; | |
827da44c | 865 | unsigned int atype, i; |
1da177e4 LT |
866 | |
867 | EnterFunction(2); | |
868 | ||
09571c7a | 869 | #ifdef CONFIG_IP_VS_IPV6 |
6cff339b | 870 | if (udest->af == AF_INET6) { |
09571c7a | 871 | atype = ipv6_addr_type(&udest->addr.in6); |
3bfb92f4 SW |
872 | if ((!(atype & IPV6_ADDR_UNICAST) || |
873 | atype & IPV6_ADDR_LINKLOCAL) && | |
3109d2f2 | 874 | !__ip_vs_addr_is_local_v6(svc->ipvs->net, &udest->addr.in6)) |
09571c7a VB |
875 | return -EINVAL; |
876 | } else | |
877 | #endif | |
878 | { | |
3109d2f2 | 879 | atype = inet_addr_type(svc->ipvs->net, udest->addr.ip); |
09571c7a VB |
880 | if (atype != RTN_LOCAL && atype != RTN_UNICAST) |
881 | return -EINVAL; | |
882 | } | |
1da177e4 | 883 | |
dee06e47 | 884 | dest = kzalloc(sizeof(struct ip_vs_dest), GFP_KERNEL); |
0a9ee813 | 885 | if (dest == NULL) |
1da177e4 | 886 | return -ENOMEM; |
0a9ee813 | 887 | |
b17fc996 | 888 | dest->stats.cpustats = alloc_percpu(struct ip_vs_cpu_stats); |
0a9ee813 | 889 | if (!dest->stats.cpustats) |
b17fc996 | 890 | goto err_alloc; |
1da177e4 | 891 | |
827da44c JS |
892 | for_each_possible_cpu(i) { |
893 | struct ip_vs_cpu_stats *ip_vs_dest_stats; | |
894 | ip_vs_dest_stats = per_cpu_ptr(dest->stats.cpustats, i); | |
895 | u64_stats_init(&ip_vs_dest_stats->syncp); | |
896 | } | |
897 | ||
6cff339b | 898 | dest->af = udest->af; |
1da177e4 | 899 | dest->protocol = svc->protocol; |
c860c6b1 | 900 | dest->vaddr = svc->addr; |
1da177e4 LT |
901 | dest->vport = svc->port; |
902 | dest->vfwmark = svc->fwmark; | |
6cff339b | 903 | ip_vs_addr_copy(udest->af, &dest->addr, &udest->addr); |
1da177e4 LT |
904 | dest->port = udest->port; |
905 | ||
906 | atomic_set(&dest->activeconns, 0); | |
907 | atomic_set(&dest->inactconns, 0); | |
908 | atomic_set(&dest->persistconns, 0); | |
26c15cfd | 909 | atomic_set(&dest->refcnt, 1); |
1da177e4 | 910 | |
276472ea | 911 | INIT_HLIST_NODE(&dest->d_list); |
1da177e4 LT |
912 | spin_lock_init(&dest->dst_lock); |
913 | spin_lock_init(&dest->stats.lock); | |
26c15cfd | 914 | __ip_vs_update_dest(svc, dest, udest, 1); |
1da177e4 LT |
915 | |
916 | *dest_p = dest; | |
917 | ||
918 | LeaveFunction(2); | |
919 | return 0; | |
b17fc996 HS |
920 | |
921 | err_alloc: | |
922 | kfree(dest); | |
923 | return -ENOMEM; | |
1da177e4 LT |
924 | } |
925 | ||
926 | ||
927 | /* | |
928 | * Add a destination into an existing service | |
929 | */ | |
930 | static int | |
c860c6b1 | 931 | ip_vs_add_dest(struct ip_vs_service *svc, struct ip_vs_dest_user_kern *udest) |
1da177e4 LT |
932 | { |
933 | struct ip_vs_dest *dest; | |
c860c6b1 | 934 | union nf_inet_addr daddr; |
014d730d | 935 | __be16 dport = udest->port; |
1da177e4 LT |
936 | int ret; |
937 | ||
938 | EnterFunction(2); | |
939 | ||
940 | if (udest->weight < 0) { | |
1e3e238e | 941 | pr_err("%s(): server weight less than zero\n", __func__); |
1da177e4 LT |
942 | return -ERANGE; |
943 | } | |
944 | ||
945 | if (udest->l_threshold > udest->u_threshold) { | |
1e3e238e HE |
946 | pr_err("%s(): lower threshold is higher than upper threshold\n", |
947 | __func__); | |
1da177e4 LT |
948 | return -ERANGE; |
949 | } | |
950 | ||
6cff339b | 951 | ip_vs_addr_copy(udest->af, &daddr, &udest->addr); |
c860c6b1 | 952 | |
413c2d04 JA |
953 | /* We use function that requires RCU lock */ |
954 | rcu_read_lock(); | |
655eef10 | 955 | dest = ip_vs_lookup_dest(svc, udest->af, &daddr, dport); |
413c2d04 | 956 | rcu_read_unlock(); |
7937df15 | 957 | |
1da177e4 | 958 | if (dest != NULL) { |
1e3e238e | 959 | IP_VS_DBG(1, "%s(): dest already exists\n", __func__); |
1da177e4 LT |
960 | return -EEXIST; |
961 | } | |
962 | ||
963 | /* | |
964 | * Check if the dest already exists in the trash and | |
965 | * is from the same service | |
966 | */ | |
ad147aa4 | 967 | dest = ip_vs_trash_get_dest(svc, udest->af, &daddr, dport); |
7937df15 | 968 | |
1da177e4 | 969 | if (dest != NULL) { |
cfc78c5a JV |
970 | IP_VS_DBG_BUF(3, "Get destination %s:%u from trash, " |
971 | "dest->refcnt=%d, service %u/%s:%u\n", | |
6cff339b | 972 | IP_VS_DBG_ADDR(udest->af, &daddr), ntohs(dport), |
cfc78c5a JV |
973 | atomic_read(&dest->refcnt), |
974 | dest->vfwmark, | |
975 | IP_VS_DBG_ADDR(svc->af, &dest->vaddr), | |
976 | ntohs(dest->vport)); | |
977 | ||
26c15cfd JA |
978 | __ip_vs_update_dest(svc, dest, udest, 1); |
979 | ret = 0; | |
980 | } else { | |
1da177e4 | 981 | /* |
26c15cfd | 982 | * Allocate and initialize the dest structure |
1da177e4 | 983 | */ |
26c15cfd | 984 | ret = ip_vs_new_dest(svc, udest, &dest); |
1da177e4 | 985 | } |
1da177e4 LT |
986 | LeaveFunction(2); |
987 | ||
26c15cfd | 988 | return ret; |
1da177e4 LT |
989 | } |
990 | ||
991 | ||
992 | /* | |
993 | * Edit a destination in the given service | |
994 | */ | |
995 | static int | |
c860c6b1 | 996 | ip_vs_edit_dest(struct ip_vs_service *svc, struct ip_vs_dest_user_kern *udest) |
1da177e4 LT |
997 | { |
998 | struct ip_vs_dest *dest; | |
c860c6b1 | 999 | union nf_inet_addr daddr; |
014d730d | 1000 | __be16 dport = udest->port; |
1da177e4 LT |
1001 | |
1002 | EnterFunction(2); | |
1003 | ||
1004 | if (udest->weight < 0) { | |
1e3e238e | 1005 | pr_err("%s(): server weight less than zero\n", __func__); |
1da177e4 LT |
1006 | return -ERANGE; |
1007 | } | |
1008 | ||
1009 | if (udest->l_threshold > udest->u_threshold) { | |
1e3e238e HE |
1010 | pr_err("%s(): lower threshold is higher than upper threshold\n", |
1011 | __func__); | |
1da177e4 LT |
1012 | return -ERANGE; |
1013 | } | |
1014 | ||
6cff339b | 1015 | ip_vs_addr_copy(udest->af, &daddr, &udest->addr); |
c860c6b1 | 1016 | |
413c2d04 JA |
1017 | /* We use function that requires RCU lock */ |
1018 | rcu_read_lock(); | |
655eef10 | 1019 | dest = ip_vs_lookup_dest(svc, udest->af, &daddr, dport); |
413c2d04 | 1020 | rcu_read_unlock(); |
7937df15 | 1021 | |
1da177e4 | 1022 | if (dest == NULL) { |
1e3e238e | 1023 | IP_VS_DBG(1, "%s(): dest doesn't exist\n", __func__); |
1da177e4 LT |
1024 | return -ENOENT; |
1025 | } | |
1026 | ||
26c15cfd | 1027 | __ip_vs_update_dest(svc, dest, udest, 0); |
1da177e4 LT |
1028 | LeaveFunction(2); |
1029 | ||
1030 | return 0; | |
1031 | } | |
1032 | ||
1da177e4 LT |
1033 | /* |
1034 | * Delete a destination (must be already unlinked from the service) | |
1035 | */ | |
79ac82e0 | 1036 | static void __ip_vs_del_dest(struct netns_ipvs *ipvs, struct ip_vs_dest *dest, |
578bc3ef | 1037 | bool cleanup) |
1da177e4 | 1038 | { |
0f34d54b | 1039 | ip_vs_stop_estimator(ipvs, &dest->stats); |
1da177e4 LT |
1040 | |
1041 | /* | |
1042 | * Remove it from the d-linked list with the real services. | |
1043 | */ | |
1da177e4 | 1044 | ip_vs_rs_unhash(dest); |
1da177e4 | 1045 | |
578bc3ef JA |
1046 | spin_lock_bh(&ipvs->dest_trash_lock); |
1047 | IP_VS_DBG_BUF(3, "Moving dest %s:%u into trash, dest->refcnt=%d\n", | |
1048 | IP_VS_DBG_ADDR(dest->af, &dest->addr), ntohs(dest->port), | |
1049 | atomic_read(&dest->refcnt)); | |
1050 | if (list_empty(&ipvs->dest_trash) && !cleanup) | |
1051 | mod_timer(&ipvs->dest_trash_timer, | |
bcbde4c0 | 1052 | jiffies + (IP_VS_DEST_TRASH_PERIOD >> 1)); |
578bc3ef JA |
1053 | /* dest lives in trash without reference */ |
1054 | list_add(&dest->t_list, &ipvs->dest_trash); | |
bcbde4c0 | 1055 | dest->idle_start = 0; |
578bc3ef JA |
1056 | spin_unlock_bh(&ipvs->dest_trash_lock); |
1057 | ip_vs_dest_put(dest); | |
1da177e4 LT |
1058 | } |
1059 | ||
1060 | ||
1061 | /* | |
1062 | * Unlink a destination from the given service | |
1063 | */ | |
1064 | static void __ip_vs_unlink_dest(struct ip_vs_service *svc, | |
1065 | struct ip_vs_dest *dest, | |
1066 | int svcupd) | |
1067 | { | |
1068 | dest->flags &= ~IP_VS_DEST_F_AVAILABLE; | |
1069 | ||
1070 | /* | |
1071 | * Remove it from the d-linked destination list. | |
1072 | */ | |
413c2d04 | 1073 | list_del_rcu(&dest->n_list); |
1da177e4 | 1074 | svc->num_dests--; |
82dfb6f3 | 1075 | |
391f503d | 1076 | if (dest->af != svc->af) |
3109d2f2 | 1077 | svc->ipvs->mixed_address_family_dests--; |
391f503d | 1078 | |
ceec4c38 JA |
1079 | if (svcupd) { |
1080 | struct ip_vs_scheduler *sched; | |
6b6df466 | 1081 | |
ceec4c38 | 1082 | sched = rcu_dereference_protected(svc->scheduler, 1); |
05f00505 | 1083 | if (sched && sched->del_dest) |
ceec4c38 JA |
1084 | sched->del_dest(svc, dest); |
1085 | } | |
1da177e4 LT |
1086 | } |
1087 | ||
1088 | ||
1089 | /* | |
1090 | * Delete a destination server in the given service | |
1091 | */ | |
1092 | static int | |
c860c6b1 | 1093 | ip_vs_del_dest(struct ip_vs_service *svc, struct ip_vs_dest_user_kern *udest) |
1da177e4 LT |
1094 | { |
1095 | struct ip_vs_dest *dest; | |
014d730d | 1096 | __be16 dport = udest->port; |
1da177e4 LT |
1097 | |
1098 | EnterFunction(2); | |
1099 | ||
413c2d04 JA |
1100 | /* We use function that requires RCU lock */ |
1101 | rcu_read_lock(); | |
655eef10 | 1102 | dest = ip_vs_lookup_dest(svc, udest->af, &udest->addr, dport); |
413c2d04 | 1103 | rcu_read_unlock(); |
c860c6b1 | 1104 | |
1da177e4 | 1105 | if (dest == NULL) { |
1e3e238e | 1106 | IP_VS_DBG(1, "%s(): destination not found!\n", __func__); |
1da177e4 LT |
1107 | return -ENOENT; |
1108 | } | |
1109 | ||
1da177e4 LT |
1110 | /* |
1111 | * Unlink dest from the service | |
1112 | */ | |
1113 | __ip_vs_unlink_dest(svc, dest, 1); | |
1114 | ||
1da177e4 LT |
1115 | /* |
1116 | * Delete the destination | |
1117 | */ | |
79ac82e0 | 1118 | __ip_vs_del_dest(svc->ipvs, dest, false); |
1da177e4 LT |
1119 | |
1120 | LeaveFunction(2); | |
1121 | ||
1122 | return 0; | |
1123 | } | |
1124 | ||
578bc3ef JA |
1125 | static void ip_vs_dest_trash_expire(unsigned long data) |
1126 | { | |
8e743f1b | 1127 | struct netns_ipvs *ipvs = (struct netns_ipvs *)data; |
578bc3ef | 1128 | struct ip_vs_dest *dest, *next; |
bcbde4c0 | 1129 | unsigned long now = jiffies; |
578bc3ef JA |
1130 | |
1131 | spin_lock(&ipvs->dest_trash_lock); | |
1132 | list_for_each_entry_safe(dest, next, &ipvs->dest_trash, t_list) { | |
578bc3ef JA |
1133 | if (atomic_read(&dest->refcnt) > 0) |
1134 | continue; | |
bcbde4c0 JA |
1135 | if (dest->idle_start) { |
1136 | if (time_before(now, dest->idle_start + | |
1137 | IP_VS_DEST_TRASH_PERIOD)) | |
1138 | continue; | |
1139 | } else { | |
1140 | dest->idle_start = max(1UL, now); | |
1141 | continue; | |
1142 | } | |
578bc3ef JA |
1143 | IP_VS_DBG_BUF(3, "Removing destination %u/%s:%u from trash\n", |
1144 | dest->vfwmark, | |
bcbde4c0 | 1145 | IP_VS_DBG_ADDR(dest->af, &dest->addr), |
578bc3ef JA |
1146 | ntohs(dest->port)); |
1147 | list_del(&dest->t_list); | |
1148 | ip_vs_dest_free(dest); | |
1149 | } | |
1150 | if (!list_empty(&ipvs->dest_trash)) | |
1151 | mod_timer(&ipvs->dest_trash_timer, | |
bcbde4c0 | 1152 | jiffies + (IP_VS_DEST_TRASH_PERIOD >> 1)); |
578bc3ef JA |
1153 | spin_unlock(&ipvs->dest_trash_lock); |
1154 | } | |
1da177e4 LT |
1155 | |
1156 | /* | |
1157 | * Add a service into the service hash table | |
1158 | */ | |
1159 | static int | |
5060bd83 | 1160 | ip_vs_add_service(struct netns_ipvs *ipvs, struct ip_vs_service_user_kern *u, |
c860c6b1 | 1161 | struct ip_vs_service **svc_p) |
1da177e4 | 1162 | { |
827da44c | 1163 | int ret = 0, i; |
1da177e4 | 1164 | struct ip_vs_scheduler *sched = NULL; |
0d1e71b0 | 1165 | struct ip_vs_pe *pe = NULL; |
1da177e4 LT |
1166 | struct ip_vs_service *svc = NULL; |
1167 | ||
1168 | /* increase the module use count */ | |
1169 | ip_vs_use_count_inc(); | |
1170 | ||
1171 | /* Lookup the scheduler by 'u->sched_name' */ | |
05f00505 JA |
1172 | if (strcmp(u->sched_name, "none")) { |
1173 | sched = ip_vs_scheduler_get(u->sched_name); | |
1174 | if (!sched) { | |
1175 | pr_info("Scheduler module ip_vs_%s not found\n", | |
1176 | u->sched_name); | |
1177 | ret = -ENOENT; | |
1178 | goto out_err; | |
1179 | } | |
1da177e4 LT |
1180 | } |
1181 | ||
0d1e71b0 | 1182 | if (u->pe_name && *u->pe_name) { |
e9e5eee8 | 1183 | pe = ip_vs_pe_getbyname(u->pe_name); |
0d1e71b0 SH |
1184 | if (pe == NULL) { |
1185 | pr_info("persistence engine module ip_vs_pe_%s " | |
1186 | "not found\n", u->pe_name); | |
1187 | ret = -ENOENT; | |
1188 | goto out_err; | |
1189 | } | |
1190 | } | |
1191 | ||
f94fd041 | 1192 | #ifdef CONFIG_IP_VS_IPV6 |
0a925864 JA |
1193 | if (u->af == AF_INET6) { |
1194 | __u32 plen = (__force __u32) u->netmask; | |
1195 | ||
1196 | if (plen < 1 || plen > 128) { | |
1197 | ret = -EINVAL; | |
1198 | goto out_err; | |
1199 | } | |
f94fd041 JV |
1200 | } |
1201 | #endif | |
1202 | ||
dee06e47 | 1203 | svc = kzalloc(sizeof(struct ip_vs_service), GFP_KERNEL); |
1da177e4 | 1204 | if (svc == NULL) { |
1e3e238e | 1205 | IP_VS_DBG(1, "%s(): no memory\n", __func__); |
1da177e4 LT |
1206 | ret = -ENOMEM; |
1207 | goto out_err; | |
1208 | } | |
b17fc996 | 1209 | svc->stats.cpustats = alloc_percpu(struct ip_vs_cpu_stats); |
0a54e939 JL |
1210 | if (!svc->stats.cpustats) { |
1211 | ret = -ENOMEM; | |
b17fc996 | 1212 | goto out_err; |
0a54e939 | 1213 | } |
1da177e4 | 1214 | |
827da44c JS |
1215 | for_each_possible_cpu(i) { |
1216 | struct ip_vs_cpu_stats *ip_vs_stats; | |
1217 | ip_vs_stats = per_cpu_ptr(svc->stats.cpustats, i); | |
1218 | u64_stats_init(&ip_vs_stats->syncp); | |
1219 | } | |
1220 | ||
1221 | ||
1da177e4 | 1222 | /* I'm the first user of the service */ |
1da177e4 LT |
1223 | atomic_set(&svc->refcnt, 0); |
1224 | ||
c860c6b1 | 1225 | svc->af = u->af; |
1da177e4 | 1226 | svc->protocol = u->protocol; |
c860c6b1 | 1227 | ip_vs_addr_copy(svc->af, &svc->addr, &u->addr); |
1da177e4 LT |
1228 | svc->port = u->port; |
1229 | svc->fwmark = u->fwmark; | |
1230 | svc->flags = u->flags; | |
1231 | svc->timeout = u->timeout * HZ; | |
1232 | svc->netmask = u->netmask; | |
3109d2f2 | 1233 | svc->ipvs = ipvs; |
1da177e4 LT |
1234 | |
1235 | INIT_LIST_HEAD(&svc->destinations); | |
ba3a3ce1 | 1236 | spin_lock_init(&svc->sched_lock); |
1da177e4 LT |
1237 | spin_lock_init(&svc->stats.lock); |
1238 | ||
1239 | /* Bind the scheduler */ | |
05f00505 JA |
1240 | if (sched) { |
1241 | ret = ip_vs_bind_scheduler(svc, sched); | |
1242 | if (ret) | |
1243 | goto out_err; | |
1244 | sched = NULL; | |
1245 | } | |
1da177e4 | 1246 | |
0d1e71b0 | 1247 | /* Bind the ct retriever */ |
ceec4c38 | 1248 | RCU_INIT_POINTER(svc->pe, pe); |
0d1e71b0 SH |
1249 | pe = NULL; |
1250 | ||
1da177e4 LT |
1251 | /* Update the virtual service counters */ |
1252 | if (svc->port == FTPPORT) | |
763f8d0e | 1253 | atomic_inc(&ipvs->ftpsvc_counter); |
1da177e4 | 1254 | else if (svc->port == 0) |
763f8d0e | 1255 | atomic_inc(&ipvs->nullsvc_counter); |
1da177e4 | 1256 | |
0f34d54b | 1257 | ip_vs_start_estimator(ipvs, &svc->stats); |
f94fd041 JV |
1258 | |
1259 | /* Count only IPv4 services for old get/setsockopt interface */ | |
1260 | if (svc->af == AF_INET) | |
a0840e2e | 1261 | ipvs->num_services++; |
1da177e4 LT |
1262 | |
1263 | /* Hash the service into the service table */ | |
1da177e4 | 1264 | ip_vs_svc_hash(svc); |
1da177e4 LT |
1265 | |
1266 | *svc_p = svc; | |
7a4f0761 HS |
1267 | /* Now there is a service - full throttle */ |
1268 | ipvs->enable = 1; | |
1da177e4 LT |
1269 | return 0; |
1270 | ||
b17fc996 | 1271 | |
6e08bfb8 | 1272 | out_err: |
1da177e4 | 1273 | if (svc != NULL) { |
ceec4c38 JA |
1274 | ip_vs_unbind_scheduler(svc, sched); |
1275 | ip_vs_service_free(svc); | |
1da177e4 LT |
1276 | } |
1277 | ip_vs_scheduler_put(sched); | |
0d1e71b0 | 1278 | ip_vs_pe_put(pe); |
1da177e4 | 1279 | |
1da177e4 LT |
1280 | /* decrease the module use count */ |
1281 | ip_vs_use_count_dec(); | |
1282 | ||
1283 | return ret; | |
1284 | } | |
1285 | ||
1286 | ||
1287 | /* | |
1288 | * Edit a service and bind it with a new scheduler | |
1289 | */ | |
1290 | static int | |
c860c6b1 | 1291 | ip_vs_edit_service(struct ip_vs_service *svc, struct ip_vs_service_user_kern *u) |
1da177e4 | 1292 | { |
05f00505 | 1293 | struct ip_vs_scheduler *sched = NULL, *old_sched; |
0d1e71b0 | 1294 | struct ip_vs_pe *pe = NULL, *old_pe = NULL; |
1da177e4 LT |
1295 | int ret = 0; |
1296 | ||
1297 | /* | |
1298 | * Lookup the scheduler, by 'u->sched_name' | |
1299 | */ | |
05f00505 JA |
1300 | if (strcmp(u->sched_name, "none")) { |
1301 | sched = ip_vs_scheduler_get(u->sched_name); | |
1302 | if (!sched) { | |
1303 | pr_info("Scheduler module ip_vs_%s not found\n", | |
1304 | u->sched_name); | |
1305 | return -ENOENT; | |
1306 | } | |
1da177e4 LT |
1307 | } |
1308 | old_sched = sched; | |
1309 | ||
0d1e71b0 | 1310 | if (u->pe_name && *u->pe_name) { |
e9e5eee8 | 1311 | pe = ip_vs_pe_getbyname(u->pe_name); |
0d1e71b0 SH |
1312 | if (pe == NULL) { |
1313 | pr_info("persistence engine module ip_vs_pe_%s " | |
1314 | "not found\n", u->pe_name); | |
1315 | ret = -ENOENT; | |
1316 | goto out; | |
1317 | } | |
1318 | old_pe = pe; | |
1319 | } | |
1320 | ||
f94fd041 | 1321 | #ifdef CONFIG_IP_VS_IPV6 |
0a925864 JA |
1322 | if (u->af == AF_INET6) { |
1323 | __u32 plen = (__force __u32) u->netmask; | |
1324 | ||
1325 | if (plen < 1 || plen > 128) { | |
1326 | ret = -EINVAL; | |
1327 | goto out; | |
1328 | } | |
f94fd041 JV |
1329 | } |
1330 | #endif | |
1331 | ||
ceec4c38 JA |
1332 | old_sched = rcu_dereference_protected(svc->scheduler, 1); |
1333 | if (sched != old_sched) { | |
05f00505 JA |
1334 | if (old_sched) { |
1335 | ip_vs_unbind_scheduler(svc, old_sched); | |
1336 | RCU_INIT_POINTER(svc->scheduler, NULL); | |
1337 | /* Wait all svc->sched_data users */ | |
1338 | synchronize_rcu(); | |
1339 | } | |
ceec4c38 | 1340 | /* Bind the new scheduler */ |
05f00505 JA |
1341 | if (sched) { |
1342 | ret = ip_vs_bind_scheduler(svc, sched); | |
1343 | if (ret) { | |
1344 | ip_vs_scheduler_put(sched); | |
1345 | goto out; | |
1346 | } | |
ceec4c38 | 1347 | } |
ceec4c38 | 1348 | } |
1da177e4 LT |
1349 | |
1350 | /* | |
1351 | * Set the flags and timeout value | |
1352 | */ | |
1353 | svc->flags = u->flags | IP_VS_SVC_F_HASHED; | |
1354 | svc->timeout = u->timeout * HZ; | |
1355 | svc->netmask = u->netmask; | |
1356 | ||
ceec4c38 JA |
1357 | old_pe = rcu_dereference_protected(svc->pe, 1); |
1358 | if (pe != old_pe) | |
1359 | rcu_assign_pointer(svc->pe, pe); | |
1da177e4 | 1360 | |
552ad65a | 1361 | out: |
6e08bfb8 | 1362 | ip_vs_scheduler_put(old_sched); |
0d1e71b0 | 1363 | ip_vs_pe_put(old_pe); |
1da177e4 LT |
1364 | return ret; |
1365 | } | |
1366 | ||
1da177e4 LT |
1367 | /* |
1368 | * Delete a service from the service list | |
1369 | * - The service must be unlinked, unlocked and not referenced! | |
1370 | * - We are called under _bh lock | |
1371 | */ | |
578bc3ef | 1372 | static void __ip_vs_del_service(struct ip_vs_service *svc, bool cleanup) |
1da177e4 LT |
1373 | { |
1374 | struct ip_vs_dest *dest, *nxt; | |
1375 | struct ip_vs_scheduler *old_sched; | |
0d1e71b0 | 1376 | struct ip_vs_pe *old_pe; |
3109d2f2 | 1377 | struct netns_ipvs *ipvs = svc->ipvs; |
0d1e71b0 SH |
1378 | |
1379 | pr_info("%s: enter\n", __func__); | |
1da177e4 | 1380 | |
f94fd041 JV |
1381 | /* Count only IPv4 services for old get/setsockopt interface */ |
1382 | if (svc->af == AF_INET) | |
a0840e2e | 1383 | ipvs->num_services--; |
f94fd041 | 1384 | |
0f34d54b | 1385 | ip_vs_stop_estimator(svc->ipvs, &svc->stats); |
1da177e4 LT |
1386 | |
1387 | /* Unbind scheduler */ | |
ceec4c38 JA |
1388 | old_sched = rcu_dereference_protected(svc->scheduler, 1); |
1389 | ip_vs_unbind_scheduler(svc, old_sched); | |
6e08bfb8 | 1390 | ip_vs_scheduler_put(old_sched); |
1da177e4 | 1391 | |
ceec4c38 JA |
1392 | /* Unbind persistence engine, keep svc->pe */ |
1393 | old_pe = rcu_dereference_protected(svc->pe, 1); | |
0d1e71b0 SH |
1394 | ip_vs_pe_put(old_pe); |
1395 | ||
1da177e4 LT |
1396 | /* |
1397 | * Unlink the whole destination list | |
1398 | */ | |
1399 | list_for_each_entry_safe(dest, nxt, &svc->destinations, n_list) { | |
1400 | __ip_vs_unlink_dest(svc, dest, 0); | |
79ac82e0 | 1401 | __ip_vs_del_dest(svc->ipvs, dest, cleanup); |
1da177e4 LT |
1402 | } |
1403 | ||
1404 | /* | |
1405 | * Update the virtual service counters | |
1406 | */ | |
1407 | if (svc->port == FTPPORT) | |
763f8d0e | 1408 | atomic_dec(&ipvs->ftpsvc_counter); |
1da177e4 | 1409 | else if (svc->port == 0) |
763f8d0e | 1410 | atomic_dec(&ipvs->nullsvc_counter); |
1da177e4 LT |
1411 | |
1412 | /* | |
1413 | * Free the service if nobody refers to it | |
1414 | */ | |
bcbde4c0 | 1415 | __ip_vs_svc_put(svc, true); |
1da177e4 LT |
1416 | |
1417 | /* decrease the module use count */ | |
1418 | ip_vs_use_count_dec(); | |
1419 | } | |
1420 | ||
1421 | /* | |
26c15cfd | 1422 | * Unlink a service from list and try to delete it if its refcnt reached 0 |
1da177e4 | 1423 | */ |
578bc3ef | 1424 | static void ip_vs_unlink_service(struct ip_vs_service *svc, bool cleanup) |
1da177e4 | 1425 | { |
ceec4c38 JA |
1426 | /* Hold svc to avoid double release from dest_trash */ |
1427 | atomic_inc(&svc->refcnt); | |
1da177e4 LT |
1428 | /* |
1429 | * Unhash it from the service table | |
1430 | */ | |
1da177e4 LT |
1431 | ip_vs_svc_unhash(svc); |
1432 | ||
578bc3ef | 1433 | __ip_vs_del_service(svc, cleanup); |
26c15cfd JA |
1434 | } |
1435 | ||
1436 | /* | |
1437 | * Delete a service from the service list | |
1438 | */ | |
1439 | static int ip_vs_del_service(struct ip_vs_service *svc) | |
1440 | { | |
1441 | if (svc == NULL) | |
1442 | return -EEXIST; | |
578bc3ef | 1443 | ip_vs_unlink_service(svc, false); |
1da177e4 LT |
1444 | |
1445 | return 0; | |
1446 | } | |
1447 | ||
1448 | ||
1449 | /* | |
1450 | * Flush all the virtual services | |
1451 | */ | |
ef7c599d | 1452 | static int ip_vs_flush(struct netns_ipvs *ipvs, bool cleanup) |
1da177e4 LT |
1453 | { |
1454 | int idx; | |
ceec4c38 JA |
1455 | struct ip_vs_service *svc; |
1456 | struct hlist_node *n; | |
1da177e4 LT |
1457 | |
1458 | /* | |
fc723250 | 1459 | * Flush the service table hashed by <netns,protocol,addr,port> |
1da177e4 LT |
1460 | */ |
1461 | for(idx = 0; idx < IP_VS_SVC_TAB_SIZE; idx++) { | |
ceec4c38 JA |
1462 | hlist_for_each_entry_safe(svc, n, &ip_vs_svc_table[idx], |
1463 | s_list) { | |
3109d2f2 | 1464 | if (svc->ipvs == ipvs) |
578bc3ef | 1465 | ip_vs_unlink_service(svc, cleanup); |
1da177e4 LT |
1466 | } |
1467 | } | |
1468 | ||
1469 | /* | |
1470 | * Flush the service table hashed by fwmark | |
1471 | */ | |
1472 | for(idx = 0; idx < IP_VS_SVC_TAB_SIZE; idx++) { | |
ceec4c38 JA |
1473 | hlist_for_each_entry_safe(svc, n, &ip_vs_svc_fwm_table[idx], |
1474 | f_list) { | |
3109d2f2 | 1475 | if (svc->ipvs == ipvs) |
578bc3ef | 1476 | ip_vs_unlink_service(svc, cleanup); |
1da177e4 LT |
1477 | } |
1478 | } | |
1479 | ||
1480 | return 0; | |
1481 | } | |
1482 | ||
7a4f0761 HS |
1483 | /* |
1484 | * Delete service by {netns} in the service table. | |
1485 | * Called by __ip_vs_cleanup() | |
1486 | */ | |
56d2169b | 1487 | void ip_vs_service_net_cleanup(struct netns_ipvs *ipvs) |
7a4f0761 HS |
1488 | { |
1489 | EnterFunction(2); | |
1490 | /* Check for "full" addressed entries */ | |
1491 | mutex_lock(&__ip_vs_mutex); | |
ef7c599d | 1492 | ip_vs_flush(ipvs, true); |
7a4f0761 HS |
1493 | mutex_unlock(&__ip_vs_mutex); |
1494 | LeaveFunction(2); | |
1495 | } | |
d1deae4d JA |
1496 | |
1497 | /* Put all references for device (dst_cache) */ | |
7a4f0761 | 1498 | static inline void |
d1deae4d | 1499 | ip_vs_forget_dev(struct ip_vs_dest *dest, struct net_device *dev) |
7a4f0761 | 1500 | { |
d717bb2a JA |
1501 | struct ip_vs_dest_dst *dest_dst; |
1502 | ||
7a4f0761 | 1503 | spin_lock_bh(&dest->dst_lock); |
d717bb2a JA |
1504 | dest_dst = rcu_dereference_protected(dest->dest_dst, 1); |
1505 | if (dest_dst && dest_dst->dst_cache->dev == dev) { | |
7a4f0761 HS |
1506 | IP_VS_DBG_BUF(3, "Reset dev:%s dest %s:%u ,dest->refcnt=%d\n", |
1507 | dev->name, | |
1508 | IP_VS_DBG_ADDR(dest->af, &dest->addr), | |
1509 | ntohs(dest->port), | |
1510 | atomic_read(&dest->refcnt)); | |
d1deae4d | 1511 | __ip_vs_dst_cache_reset(dest); |
7a4f0761 HS |
1512 | } |
1513 | spin_unlock_bh(&dest->dst_lock); | |
1514 | ||
1515 | } | |
313eae63 JA |
1516 | /* Netdev event receiver |
1517 | * Currently only NETDEV_DOWN is handled to release refs to cached dsts | |
7a4f0761 HS |
1518 | */ |
1519 | static int ip_vs_dst_event(struct notifier_block *this, unsigned long event, | |
351638e7 | 1520 | void *ptr) |
7a4f0761 | 1521 | { |
351638e7 | 1522 | struct net_device *dev = netdev_notifier_info_to_dev(ptr); |
7a4f0761 | 1523 | struct net *net = dev_net(dev); |
283283c4 | 1524 | struct netns_ipvs *ipvs = net_ipvs(net); |
7a4f0761 HS |
1525 | struct ip_vs_service *svc; |
1526 | struct ip_vs_dest *dest; | |
1527 | unsigned int idx; | |
1528 | ||
313eae63 | 1529 | if (event != NETDEV_DOWN || !ipvs) |
7a4f0761 HS |
1530 | return NOTIFY_DONE; |
1531 | IP_VS_DBG(3, "%s() dev=%s\n", __func__, dev->name); | |
1532 | EnterFunction(2); | |
1533 | mutex_lock(&__ip_vs_mutex); | |
1534 | for (idx = 0; idx < IP_VS_SVC_TAB_SIZE; idx++) { | |
ceec4c38 | 1535 | hlist_for_each_entry(svc, &ip_vs_svc_table[idx], s_list) { |
3109d2f2 | 1536 | if (svc->ipvs == ipvs) { |
7a4f0761 HS |
1537 | list_for_each_entry(dest, &svc->destinations, |
1538 | n_list) { | |
d1deae4d | 1539 | ip_vs_forget_dev(dest, dev); |
7a4f0761 HS |
1540 | } |
1541 | } | |
1542 | } | |
1543 | ||
ceec4c38 | 1544 | hlist_for_each_entry(svc, &ip_vs_svc_fwm_table[idx], f_list) { |
3109d2f2 | 1545 | if (svc->ipvs == ipvs) { |
7a4f0761 HS |
1546 | list_for_each_entry(dest, &svc->destinations, |
1547 | n_list) { | |
d1deae4d | 1548 | ip_vs_forget_dev(dest, dev); |
7a4f0761 HS |
1549 | } |
1550 | } | |
1551 | ||
1552 | } | |
1553 | } | |
1554 | ||
578bc3ef JA |
1555 | spin_lock_bh(&ipvs->dest_trash_lock); |
1556 | list_for_each_entry(dest, &ipvs->dest_trash, t_list) { | |
d1deae4d | 1557 | ip_vs_forget_dev(dest, dev); |
7a4f0761 | 1558 | } |
578bc3ef | 1559 | spin_unlock_bh(&ipvs->dest_trash_lock); |
7a4f0761 HS |
1560 | mutex_unlock(&__ip_vs_mutex); |
1561 | LeaveFunction(2); | |
1562 | return NOTIFY_DONE; | |
1563 | } | |
1da177e4 LT |
1564 | |
1565 | /* | |
1566 | * Zero counters in a service or all services | |
1567 | */ | |
1568 | static int ip_vs_zero_service(struct ip_vs_service *svc) | |
1569 | { | |
1570 | struct ip_vs_dest *dest; | |
1571 | ||
1da177e4 LT |
1572 | list_for_each_entry(dest, &svc->destinations, n_list) { |
1573 | ip_vs_zero_stats(&dest->stats); | |
1574 | } | |
1575 | ip_vs_zero_stats(&svc->stats); | |
1da177e4 LT |
1576 | return 0; |
1577 | } | |
1578 | ||
c60856c6 | 1579 | static int ip_vs_zero_all(struct netns_ipvs *ipvs) |
1da177e4 LT |
1580 | { |
1581 | int idx; | |
1582 | struct ip_vs_service *svc; | |
1583 | ||
1584 | for(idx = 0; idx < IP_VS_SVC_TAB_SIZE; idx++) { | |
ceec4c38 | 1585 | hlist_for_each_entry(svc, &ip_vs_svc_table[idx], s_list) { |
3109d2f2 | 1586 | if (svc->ipvs == ipvs) |
fc723250 | 1587 | ip_vs_zero_service(svc); |
1da177e4 LT |
1588 | } |
1589 | } | |
1590 | ||
1591 | for(idx = 0; idx < IP_VS_SVC_TAB_SIZE; idx++) { | |
ceec4c38 | 1592 | hlist_for_each_entry(svc, &ip_vs_svc_fwm_table[idx], f_list) { |
3109d2f2 | 1593 | if (svc->ipvs == ipvs) |
fc723250 | 1594 | ip_vs_zero_service(svc); |
1da177e4 LT |
1595 | } |
1596 | } | |
1597 | ||
3109d2f2 | 1598 | ip_vs_zero_stats(&ipvs->tot_stats); |
1da177e4 LT |
1599 | return 0; |
1600 | } | |
1601 | ||
14e40546 | 1602 | #ifdef CONFIG_SYSCTL |
749c42b6 JA |
1603 | |
1604 | static int zero; | |
1605 | static int three = 3; | |
1606 | ||
1da177e4 | 1607 | static int |
fe2c6338 | 1608 | proc_do_defense_mode(struct ctl_table *table, int write, |
1da177e4 LT |
1609 | void __user *buffer, size_t *lenp, loff_t *ppos) |
1610 | { | |
717e917d | 1611 | struct netns_ipvs *ipvs = table->extra2; |
1da177e4 LT |
1612 | int *valp = table->data; |
1613 | int val = *valp; | |
1614 | int rc; | |
1615 | ||
8d65af78 | 1616 | rc = proc_dointvec(table, write, buffer, lenp, ppos); |
1da177e4 LT |
1617 | if (write && (*valp != val)) { |
1618 | if ((*valp < 0) || (*valp > 3)) { | |
1619 | /* Restore the correct value */ | |
1620 | *valp = val; | |
1621 | } else { | |
717e917d | 1622 | update_defense_level(ipvs); |
1da177e4 LT |
1623 | } |
1624 | } | |
1625 | return rc; | |
1626 | } | |
1627 | ||
1da177e4 | 1628 | static int |
fe2c6338 | 1629 | proc_do_sync_threshold(struct ctl_table *table, int write, |
1da177e4 LT |
1630 | void __user *buffer, size_t *lenp, loff_t *ppos) |
1631 | { | |
1632 | int *valp = table->data; | |
1633 | int val[2]; | |
1634 | int rc; | |
1635 | ||
1636 | /* backup the value first */ | |
1637 | memcpy(val, valp, sizeof(val)); | |
1638 | ||
8d65af78 | 1639 | rc = proc_dointvec(table, write, buffer, lenp, ppos); |
749c42b6 JA |
1640 | if (write && (valp[0] < 0 || valp[1] < 0 || |
1641 | (valp[0] >= valp[1] && valp[1]))) { | |
1da177e4 LT |
1642 | /* Restore the correct value */ |
1643 | memcpy(valp, val, sizeof(val)); | |
1644 | } | |
1645 | return rc; | |
1646 | } | |
1647 | ||
b880c1f0 | 1648 | static int |
fe2c6338 | 1649 | proc_do_sync_mode(struct ctl_table *table, int write, |
b880c1f0 HS |
1650 | void __user *buffer, size_t *lenp, loff_t *ppos) |
1651 | { | |
1652 | int *valp = table->data; | |
1653 | int val = *valp; | |
1654 | int rc; | |
1655 | ||
1656 | rc = proc_dointvec(table, write, buffer, lenp, ppos); | |
1657 | if (write && (*valp != val)) { | |
1658 | if ((*valp < 0) || (*valp > 1)) { | |
1659 | /* Restore the correct value */ | |
1660 | *valp = val; | |
f73181c8 PNA |
1661 | } |
1662 | } | |
1663 | return rc; | |
1664 | } | |
1665 | ||
1666 | static int | |
fe2c6338 | 1667 | proc_do_sync_ports(struct ctl_table *table, int write, |
f73181c8 PNA |
1668 | void __user *buffer, size_t *lenp, loff_t *ppos) |
1669 | { | |
1670 | int *valp = table->data; | |
1671 | int val = *valp; | |
1672 | int rc; | |
1673 | ||
1674 | rc = proc_dointvec(table, write, buffer, lenp, ppos); | |
1675 | if (write && (*valp != val)) { | |
1676 | if (*valp < 1 || !is_power_of_2(*valp)) { | |
1677 | /* Restore the correct value */ | |
1678 | *valp = val; | |
b880c1f0 HS |
1679 | } |
1680 | } | |
1681 | return rc; | |
1682 | } | |
1da177e4 LT |
1683 | |
1684 | /* | |
1685 | * IPVS sysctl table (under the /proc/sys/net/ipv4/vs/) | |
a0840e2e | 1686 | * Do not change order or insert new entries without |
503cf15a | 1687 | * align with netns init in ip_vs_control_net_init() |
1da177e4 LT |
1688 | */ |
1689 | ||
1690 | static struct ctl_table vs_vars[] = { | |
1691 | { | |
1da177e4 | 1692 | .procname = "amemthresh", |
1da177e4 LT |
1693 | .maxlen = sizeof(int), |
1694 | .mode = 0644, | |
6d9f239a | 1695 | .proc_handler = proc_dointvec, |
1da177e4 | 1696 | }, |
1da177e4 | 1697 | { |
1da177e4 | 1698 | .procname = "am_droprate", |
1da177e4 LT |
1699 | .maxlen = sizeof(int), |
1700 | .mode = 0644, | |
6d9f239a | 1701 | .proc_handler = proc_dointvec, |
1da177e4 LT |
1702 | }, |
1703 | { | |
1da177e4 | 1704 | .procname = "drop_entry", |
1da177e4 LT |
1705 | .maxlen = sizeof(int), |
1706 | .mode = 0644, | |
6d9f239a | 1707 | .proc_handler = proc_do_defense_mode, |
1da177e4 LT |
1708 | }, |
1709 | { | |
1da177e4 | 1710 | .procname = "drop_packet", |
1da177e4 LT |
1711 | .maxlen = sizeof(int), |
1712 | .mode = 0644, | |
6d9f239a | 1713 | .proc_handler = proc_do_defense_mode, |
1da177e4 | 1714 | }, |
f4bc17cd JA |
1715 | #ifdef CONFIG_IP_VS_NFCT |
1716 | { | |
1717 | .procname = "conntrack", | |
f4bc17cd JA |
1718 | .maxlen = sizeof(int), |
1719 | .mode = 0644, | |
1720 | .proc_handler = &proc_dointvec, | |
1721 | }, | |
1722 | #endif | |
1da177e4 | 1723 | { |
1da177e4 | 1724 | .procname = "secure_tcp", |
1da177e4 LT |
1725 | .maxlen = sizeof(int), |
1726 | .mode = 0644, | |
6d9f239a | 1727 | .proc_handler = proc_do_defense_mode, |
1da177e4 | 1728 | }, |
8a803040 JA |
1729 | { |
1730 | .procname = "snat_reroute", | |
8a803040 JA |
1731 | .maxlen = sizeof(int), |
1732 | .mode = 0644, | |
1733 | .proc_handler = &proc_dointvec, | |
1734 | }, | |
b880c1f0 HS |
1735 | { |
1736 | .procname = "sync_version", | |
b880c1f0 HS |
1737 | .maxlen = sizeof(int), |
1738 | .mode = 0644, | |
1739 | .proc_handler = &proc_do_sync_mode, | |
1740 | }, | |
f73181c8 PNA |
1741 | { |
1742 | .procname = "sync_ports", | |
1743 | .maxlen = sizeof(int), | |
1744 | .mode = 0644, | |
1745 | .proc_handler = &proc_do_sync_ports, | |
1746 | }, | |
4d0c875d JA |
1747 | { |
1748 | .procname = "sync_persist_mode", | |
1749 | .maxlen = sizeof(int), | |
1750 | .mode = 0644, | |
1751 | .proc_handler = proc_dointvec, | |
1752 | }, | |
1c003b15 PNA |
1753 | { |
1754 | .procname = "sync_qlen_max", | |
07995674 | 1755 | .maxlen = sizeof(unsigned long), |
1c003b15 | 1756 | .mode = 0644, |
07995674 | 1757 | .proc_handler = proc_doulongvec_minmax, |
1c003b15 PNA |
1758 | }, |
1759 | { | |
1760 | .procname = "sync_sock_size", | |
1761 | .maxlen = sizeof(int), | |
1762 | .mode = 0644, | |
1763 | .proc_handler = proc_dointvec, | |
1764 | }, | |
a0840e2e HS |
1765 | { |
1766 | .procname = "cache_bypass", | |
1767 | .maxlen = sizeof(int), | |
1768 | .mode = 0644, | |
1769 | .proc_handler = proc_dointvec, | |
1770 | }, | |
1771 | { | |
1772 | .procname = "expire_nodest_conn", | |
1773 | .maxlen = sizeof(int), | |
1774 | .mode = 0644, | |
1775 | .proc_handler = proc_dointvec, | |
1776 | }, | |
c6c96c18 AF |
1777 | { |
1778 | .procname = "sloppy_tcp", | |
1779 | .maxlen = sizeof(int), | |
1780 | .mode = 0644, | |
1781 | .proc_handler = proc_dointvec, | |
1782 | }, | |
1783 | { | |
1784 | .procname = "sloppy_sctp", | |
1785 | .maxlen = sizeof(int), | |
1786 | .mode = 0644, | |
1787 | .proc_handler = proc_dointvec, | |
1788 | }, | |
a0840e2e HS |
1789 | { |
1790 | .procname = "expire_quiescent_template", | |
1791 | .maxlen = sizeof(int), | |
1792 | .mode = 0644, | |
1793 | .proc_handler = proc_dointvec, | |
1794 | }, | |
1795 | { | |
1796 | .procname = "sync_threshold", | |
1797 | .maxlen = | |
1798 | sizeof(((struct netns_ipvs *)0)->sysctl_sync_threshold), | |
1799 | .mode = 0644, | |
1800 | .proc_handler = proc_do_sync_threshold, | |
1801 | }, | |
749c42b6 JA |
1802 | { |
1803 | .procname = "sync_refresh_period", | |
1804 | .maxlen = sizeof(int), | |
1805 | .mode = 0644, | |
1806 | .proc_handler = proc_dointvec_jiffies, | |
1807 | }, | |
1808 | { | |
1809 | .procname = "sync_retries", | |
1810 | .maxlen = sizeof(int), | |
1811 | .mode = 0644, | |
1812 | .proc_handler = proc_dointvec_minmax, | |
1813 | .extra1 = &zero, | |
1814 | .extra2 = &three, | |
1815 | }, | |
a0840e2e HS |
1816 | { |
1817 | .procname = "nat_icmp_send", | |
1818 | .maxlen = sizeof(int), | |
1819 | .mode = 0644, | |
1820 | .proc_handler = proc_dointvec, | |
1821 | }, | |
3654e611 JA |
1822 | { |
1823 | .procname = "pmtu_disc", | |
1824 | .maxlen = sizeof(int), | |
1825 | .mode = 0644, | |
1826 | .proc_handler = proc_dointvec, | |
1827 | }, | |
0c12582f JA |
1828 | { |
1829 | .procname = "backup_only", | |
1830 | .maxlen = sizeof(int), | |
1831 | .mode = 0644, | |
1832 | .proc_handler = proc_dointvec, | |
1833 | }, | |
d752c364 MRL |
1834 | { |
1835 | .procname = "conn_reuse_mode", | |
1836 | .maxlen = sizeof(int), | |
1837 | .mode = 0644, | |
1838 | .proc_handler = proc_dointvec, | |
1839 | }, | |
94485fed AG |
1840 | { |
1841 | .procname = "schedule_icmp", | |
1842 | .maxlen = sizeof(int), | |
1843 | .mode = 0644, | |
1844 | .proc_handler = proc_dointvec, | |
1845 | }, | |
4e478098 AG |
1846 | { |
1847 | .procname = "ignore_tunneled", | |
1848 | .maxlen = sizeof(int), | |
1849 | .mode = 0644, | |
1850 | .proc_handler = proc_dointvec, | |
1851 | }, | |
a0840e2e HS |
1852 | #ifdef CONFIG_IP_VS_DEBUG |
1853 | { | |
1854 | .procname = "debug_level", | |
1855 | .data = &sysctl_ip_vs_debug_level, | |
1856 | .maxlen = sizeof(int), | |
1857 | .mode = 0644, | |
1858 | .proc_handler = proc_dointvec, | |
1859 | }, | |
1da177e4 | 1860 | #endif |
f8572d8f | 1861 | { } |
1da177e4 LT |
1862 | }; |
1863 | ||
14e40546 | 1864 | #endif |
1da177e4 | 1865 | |
1da177e4 LT |
1866 | #ifdef CONFIG_PROC_FS |
1867 | ||
1868 | struct ip_vs_iter { | |
fc723250 | 1869 | struct seq_net_private p; /* Do not move this, netns depends upon it*/ |
ceec4c38 | 1870 | struct hlist_head *table; |
1da177e4 LT |
1871 | int bucket; |
1872 | }; | |
1873 | ||
1874 | /* | |
1875 | * Write the contents of the VS rule table to a PROCfs file. | |
1876 | * (It is kept just for backward compatibility) | |
1877 | */ | |
95c96174 | 1878 | static inline const char *ip_vs_fwd_name(unsigned int flags) |
1da177e4 LT |
1879 | { |
1880 | switch (flags & IP_VS_CONN_F_FWD_MASK) { | |
1881 | case IP_VS_CONN_F_LOCALNODE: | |
1882 | return "Local"; | |
1883 | case IP_VS_CONN_F_TUNNEL: | |
1884 | return "Tunnel"; | |
1885 | case IP_VS_CONN_F_DROUTE: | |
1886 | return "Route"; | |
1887 | default: | |
1888 | return "Masq"; | |
1889 | } | |
1890 | } | |
1891 | ||
1892 | ||
1893 | /* Get the Nth entry in the two lists */ | |
1894 | static struct ip_vs_service *ip_vs_info_array(struct seq_file *seq, loff_t pos) | |
1895 | { | |
fc723250 | 1896 | struct net *net = seq_file_net(seq); |
3109d2f2 | 1897 | struct netns_ipvs *ipvs = net_ipvs(net); |
1da177e4 LT |
1898 | struct ip_vs_iter *iter = seq->private; |
1899 | int idx; | |
1900 | struct ip_vs_service *svc; | |
1901 | ||
1902 | /* look in hash by protocol */ | |
1903 | for (idx = 0; idx < IP_VS_SVC_TAB_SIZE; idx++) { | |
ceec4c38 | 1904 | hlist_for_each_entry_rcu(svc, &ip_vs_svc_table[idx], s_list) { |
3109d2f2 | 1905 | if ((svc->ipvs == ipvs) && pos-- == 0) { |
1da177e4 LT |
1906 | iter->table = ip_vs_svc_table; |
1907 | iter->bucket = idx; | |
1908 | return svc; | |
1909 | } | |
1910 | } | |
1911 | } | |
1912 | ||
1913 | /* keep looking in fwmark */ | |
1914 | for (idx = 0; idx < IP_VS_SVC_TAB_SIZE; idx++) { | |
ceec4c38 JA |
1915 | hlist_for_each_entry_rcu(svc, &ip_vs_svc_fwm_table[idx], |
1916 | f_list) { | |
3109d2f2 | 1917 | if ((svc->ipvs == ipvs) && pos-- == 0) { |
1da177e4 LT |
1918 | iter->table = ip_vs_svc_fwm_table; |
1919 | iter->bucket = idx; | |
1920 | return svc; | |
1921 | } | |
1922 | } | |
1923 | } | |
1924 | ||
1925 | return NULL; | |
1926 | } | |
1927 | ||
1928 | static void *ip_vs_info_seq_start(struct seq_file *seq, loff_t *pos) | |
371990ee | 1929 | __acquires(RCU) |
1da177e4 | 1930 | { |
ceec4c38 | 1931 | rcu_read_lock(); |
1da177e4 LT |
1932 | return *pos ? ip_vs_info_array(seq, *pos - 1) : SEQ_START_TOKEN; |
1933 | } | |
1934 | ||
1935 | ||
1936 | static void *ip_vs_info_seq_next(struct seq_file *seq, void *v, loff_t *pos) | |
1937 | { | |
ceec4c38 | 1938 | struct hlist_node *e; |
1da177e4 LT |
1939 | struct ip_vs_iter *iter; |
1940 | struct ip_vs_service *svc; | |
1941 | ||
1942 | ++*pos; | |
1943 | if (v == SEQ_START_TOKEN) | |
1944 | return ip_vs_info_array(seq,0); | |
1945 | ||
1946 | svc = v; | |
1947 | iter = seq->private; | |
1948 | ||
1949 | if (iter->table == ip_vs_svc_table) { | |
1950 | /* next service in table hashed by protocol */ | |
ceec4c38 JA |
1951 | e = rcu_dereference(hlist_next_rcu(&svc->s_list)); |
1952 | if (e) | |
1953 | return hlist_entry(e, struct ip_vs_service, s_list); | |
1da177e4 LT |
1954 | |
1955 | while (++iter->bucket < IP_VS_SVC_TAB_SIZE) { | |
ceec4c38 JA |
1956 | hlist_for_each_entry_rcu(svc, |
1957 | &ip_vs_svc_table[iter->bucket], | |
1958 | s_list) { | |
1da177e4 LT |
1959 | return svc; |
1960 | } | |
1961 | } | |
1962 | ||
1963 | iter->table = ip_vs_svc_fwm_table; | |
1964 | iter->bucket = -1; | |
1965 | goto scan_fwmark; | |
1966 | } | |
1967 | ||
1968 | /* next service in hashed by fwmark */ | |
ceec4c38 JA |
1969 | e = rcu_dereference(hlist_next_rcu(&svc->f_list)); |
1970 | if (e) | |
1971 | return hlist_entry(e, struct ip_vs_service, f_list); | |
1da177e4 LT |
1972 | |
1973 | scan_fwmark: | |
1974 | while (++iter->bucket < IP_VS_SVC_TAB_SIZE) { | |
ceec4c38 JA |
1975 | hlist_for_each_entry_rcu(svc, |
1976 | &ip_vs_svc_fwm_table[iter->bucket], | |
1977 | f_list) | |
1da177e4 LT |
1978 | return svc; |
1979 | } | |
1980 | ||
1981 | return NULL; | |
1982 | } | |
1983 | ||
1984 | static void ip_vs_info_seq_stop(struct seq_file *seq, void *v) | |
371990ee | 1985 | __releases(RCU) |
1da177e4 | 1986 | { |
ceec4c38 | 1987 | rcu_read_unlock(); |
1da177e4 LT |
1988 | } |
1989 | ||
1990 | ||
1991 | static int ip_vs_info_seq_show(struct seq_file *seq, void *v) | |
1992 | { | |
1993 | if (v == SEQ_START_TOKEN) { | |
1994 | seq_printf(seq, | |
1995 | "IP Virtual Server version %d.%d.%d (size=%d)\n", | |
6f7edb48 | 1996 | NVERSION(IP_VS_VERSION_CODE), ip_vs_conn_tab_size); |
1da177e4 LT |
1997 | seq_puts(seq, |
1998 | "Prot LocalAddress:Port Scheduler Flags\n"); | |
1999 | seq_puts(seq, | |
2000 | " -> RemoteAddress:Port Forward Weight ActiveConn InActConn\n"); | |
2001 | } else { | |
2002 | const struct ip_vs_service *svc = v; | |
2003 | const struct ip_vs_iter *iter = seq->private; | |
2004 | const struct ip_vs_dest *dest; | |
ceec4c38 | 2005 | struct ip_vs_scheduler *sched = rcu_dereference(svc->scheduler); |
05f00505 | 2006 | char *sched_name = sched ? sched->name : "none"; |
1da177e4 | 2007 | |
667a5f18 VB |
2008 | if (iter->table == ip_vs_svc_table) { |
2009 | #ifdef CONFIG_IP_VS_IPV6 | |
2010 | if (svc->af == AF_INET6) | |
5b095d98 | 2011 | seq_printf(seq, "%s [%pI6]:%04X %s ", |
667a5f18 | 2012 | ip_vs_proto_name(svc->protocol), |
38ff4fa4 | 2013 | &svc->addr.in6, |
667a5f18 | 2014 | ntohs(svc->port), |
05f00505 | 2015 | sched_name); |
667a5f18 VB |
2016 | else |
2017 | #endif | |
26ec037f | 2018 | seq_printf(seq, "%s %08X:%04X %s %s ", |
667a5f18 VB |
2019 | ip_vs_proto_name(svc->protocol), |
2020 | ntohl(svc->addr.ip), | |
2021 | ntohs(svc->port), | |
05f00505 | 2022 | sched_name, |
26ec037f | 2023 | (svc->flags & IP_VS_SVC_F_ONEPACKET)?"ops ":""); |
667a5f18 | 2024 | } else { |
26ec037f | 2025 | seq_printf(seq, "FWM %08X %s %s", |
05f00505 | 2026 | svc->fwmark, sched_name, |
26ec037f | 2027 | (svc->flags & IP_VS_SVC_F_ONEPACKET)?"ops ":""); |
667a5f18 | 2028 | } |
1da177e4 LT |
2029 | |
2030 | if (svc->flags & IP_VS_SVC_F_PERSISTENT) | |
2031 | seq_printf(seq, "persistent %d %08X\n", | |
2032 | svc->timeout, | |
2033 | ntohl(svc->netmask)); | |
2034 | else | |
2035 | seq_putc(seq, '\n'); | |
2036 | ||
413c2d04 | 2037 | list_for_each_entry_rcu(dest, &svc->destinations, n_list) { |
667a5f18 VB |
2038 | #ifdef CONFIG_IP_VS_IPV6 |
2039 | if (dest->af == AF_INET6) | |
2040 | seq_printf(seq, | |
5b095d98 | 2041 | " -> [%pI6]:%04X" |
667a5f18 | 2042 | " %-7s %-6d %-10d %-10d\n", |
38ff4fa4 | 2043 | &dest->addr.in6, |
667a5f18 VB |
2044 | ntohs(dest->port), |
2045 | ip_vs_fwd_name(atomic_read(&dest->conn_flags)), | |
2046 | atomic_read(&dest->weight), | |
2047 | atomic_read(&dest->activeconns), | |
2048 | atomic_read(&dest->inactconns)); | |
2049 | else | |
2050 | #endif | |
2051 | seq_printf(seq, | |
2052 | " -> %08X:%04X " | |
2053 | "%-7s %-6d %-10d %-10d\n", | |
2054 | ntohl(dest->addr.ip), | |
2055 | ntohs(dest->port), | |
2056 | ip_vs_fwd_name(atomic_read(&dest->conn_flags)), | |
2057 | atomic_read(&dest->weight), | |
2058 | atomic_read(&dest->activeconns), | |
2059 | atomic_read(&dest->inactconns)); | |
2060 | ||
1da177e4 LT |
2061 | } |
2062 | } | |
2063 | return 0; | |
2064 | } | |
2065 | ||
56b3d975 | 2066 | static const struct seq_operations ip_vs_info_seq_ops = { |
1da177e4 LT |
2067 | .start = ip_vs_info_seq_start, |
2068 | .next = ip_vs_info_seq_next, | |
2069 | .stop = ip_vs_info_seq_stop, | |
2070 | .show = ip_vs_info_seq_show, | |
2071 | }; | |
2072 | ||
2073 | static int ip_vs_info_open(struct inode *inode, struct file *file) | |
2074 | { | |
fc723250 | 2075 | return seq_open_net(inode, file, &ip_vs_info_seq_ops, |
cf7732e4 | 2076 | sizeof(struct ip_vs_iter)); |
1da177e4 LT |
2077 | } |
2078 | ||
9a32144e | 2079 | static const struct file_operations ip_vs_info_fops = { |
1da177e4 LT |
2080 | .owner = THIS_MODULE, |
2081 | .open = ip_vs_info_open, | |
2082 | .read = seq_read, | |
2083 | .llseek = seq_lseek, | |
0f08190f | 2084 | .release = seq_release_net, |
1da177e4 LT |
2085 | }; |
2086 | ||
1da177e4 LT |
2087 | static int ip_vs_stats_show(struct seq_file *seq, void *v) |
2088 | { | |
b17fc996 | 2089 | struct net *net = seq_file_single_net(seq); |
cd67cd5e | 2090 | struct ip_vs_kstats show; |
1da177e4 LT |
2091 | |
2092 | /* 01234567 01234567 01234567 0123456701234567 0123456701234567 */ | |
2093 | seq_puts(seq, | |
2094 | " Total Incoming Outgoing Incoming Outgoing\n"); | |
2095 | seq_printf(seq, | |
2096 | " Conns Packets Packets Bytes Bytes\n"); | |
2097 | ||
55a3d4e1 | 2098 | ip_vs_copy_stats(&show, &net_ipvs(net)->tot_stats); |
cd67cd5e JA |
2099 | seq_printf(seq, "%8LX %8LX %8LX %16LX %16LX\n\n", |
2100 | (unsigned long long)show.conns, | |
2101 | (unsigned long long)show.inpkts, | |
2102 | (unsigned long long)show.outpkts, | |
2103 | (unsigned long long)show.inbytes, | |
2104 | (unsigned long long)show.outbytes); | |
2105 | ||
2106 | /* 01234567 01234567 01234567 0123456701234567 0123456701234567*/ | |
1da177e4 | 2107 | seq_puts(seq, |
cd67cd5e JA |
2108 | " Conns/s Pkts/s Pkts/s Bytes/s Bytes/s\n"); |
2109 | seq_printf(seq, "%8LX %8LX %8LX %16LX %16LX\n", | |
2110 | (unsigned long long)show.cps, | |
2111 | (unsigned long long)show.inpps, | |
2112 | (unsigned long long)show.outpps, | |
2113 | (unsigned long long)show.inbps, | |
2114 | (unsigned long long)show.outbps); | |
1da177e4 LT |
2115 | |
2116 | return 0; | |
2117 | } | |
2118 | ||
2119 | static int ip_vs_stats_seq_open(struct inode *inode, struct file *file) | |
2120 | { | |
fc723250 | 2121 | return single_open_net(inode, file, ip_vs_stats_show); |
1da177e4 LT |
2122 | } |
2123 | ||
9a32144e | 2124 | static const struct file_operations ip_vs_stats_fops = { |
1da177e4 LT |
2125 | .owner = THIS_MODULE, |
2126 | .open = ip_vs_stats_seq_open, | |
2127 | .read = seq_read, | |
2128 | .llseek = seq_lseek, | |
0f08190f | 2129 | .release = single_release_net, |
1da177e4 LT |
2130 | }; |
2131 | ||
b17fc996 HS |
2132 | static int ip_vs_stats_percpu_show(struct seq_file *seq, void *v) |
2133 | { | |
2134 | struct net *net = seq_file_single_net(seq); | |
2a0751af | 2135 | struct ip_vs_stats *tot_stats = &net_ipvs(net)->tot_stats; |
371990ee | 2136 | struct ip_vs_cpu_stats __percpu *cpustats = tot_stats->cpustats; |
cd67cd5e | 2137 | struct ip_vs_kstats kstats; |
b17fc996 HS |
2138 | int i; |
2139 | ||
2140 | /* 01234567 01234567 01234567 0123456701234567 0123456701234567 */ | |
2141 | seq_puts(seq, | |
2142 | " Total Incoming Outgoing Incoming Outgoing\n"); | |
2143 | seq_printf(seq, | |
2144 | "CPU Conns Packets Packets Bytes Bytes\n"); | |
2145 | ||
2146 | for_each_possible_cpu(i) { | |
2a0751af JA |
2147 | struct ip_vs_cpu_stats *u = per_cpu_ptr(cpustats, i); |
2148 | unsigned int start; | |
cd67cd5e | 2149 | u64 conns, inpkts, outpkts, inbytes, outbytes; |
2a0751af JA |
2150 | |
2151 | do { | |
57a7744e | 2152 | start = u64_stats_fetch_begin_irq(&u->syncp); |
cd67cd5e JA |
2153 | conns = u->cnt.conns; |
2154 | inpkts = u->cnt.inpkts; | |
2155 | outpkts = u->cnt.outpkts; | |
2156 | inbytes = u->cnt.inbytes; | |
2157 | outbytes = u->cnt.outbytes; | |
57a7744e | 2158 | } while (u64_stats_fetch_retry_irq(&u->syncp, start)); |
2a0751af | 2159 | |
cd67cd5e JA |
2160 | seq_printf(seq, "%3X %8LX %8LX %8LX %16LX %16LX\n", |
2161 | i, (u64)conns, (u64)inpkts, | |
2162 | (u64)outpkts, (u64)inbytes, | |
2163 | (u64)outbytes); | |
b17fc996 HS |
2164 | } |
2165 | ||
cd67cd5e | 2166 | ip_vs_copy_stats(&kstats, tot_stats); |
ea9f22cc | 2167 | |
cd67cd5e JA |
2168 | seq_printf(seq, " ~ %8LX %8LX %8LX %16LX %16LX\n\n", |
2169 | (unsigned long long)kstats.conns, | |
2170 | (unsigned long long)kstats.inpkts, | |
2171 | (unsigned long long)kstats.outpkts, | |
2172 | (unsigned long long)kstats.inbytes, | |
2173 | (unsigned long long)kstats.outbytes); | |
ea9f22cc | 2174 | |
cd67cd5e | 2175 | /* ... 01234567 01234567 01234567 0123456701234567 0123456701234567 */ |
b17fc996 | 2176 | seq_puts(seq, |
cd67cd5e JA |
2177 | " Conns/s Pkts/s Pkts/s Bytes/s Bytes/s\n"); |
2178 | seq_printf(seq, " %8LX %8LX %8LX %16LX %16LX\n", | |
2179 | kstats.cps, | |
2180 | kstats.inpps, | |
2181 | kstats.outpps, | |
2182 | kstats.inbps, | |
2183 | kstats.outbps); | |
b17fc996 HS |
2184 | |
2185 | return 0; | |
2186 | } | |
2187 | ||
2188 | static int ip_vs_stats_percpu_seq_open(struct inode *inode, struct file *file) | |
2189 | { | |
2190 | return single_open_net(inode, file, ip_vs_stats_percpu_show); | |
2191 | } | |
2192 | ||
2193 | static const struct file_operations ip_vs_stats_percpu_fops = { | |
2194 | .owner = THIS_MODULE, | |
2195 | .open = ip_vs_stats_percpu_seq_open, | |
2196 | .read = seq_read, | |
2197 | .llseek = seq_lseek, | |
0f08190f | 2198 | .release = single_release_net, |
b17fc996 | 2199 | }; |
1da177e4 LT |
2200 | #endif |
2201 | ||
2202 | /* | |
2203 | * Set timeout values for tcp tcpfin udp in the timeout_table. | |
2204 | */ | |
f1faa1e7 | 2205 | static int ip_vs_set_timeout(struct netns_ipvs *ipvs, struct ip_vs_timeout_user *u) |
1da177e4 | 2206 | { |
091bb34c | 2207 | #if defined(CONFIG_IP_VS_PROTO_TCP) || defined(CONFIG_IP_VS_PROTO_UDP) |
9330419d | 2208 | struct ip_vs_proto_data *pd; |
091bb34c | 2209 | #endif |
9330419d | 2210 | |
1da177e4 LT |
2211 | IP_VS_DBG(2, "Setting timeout tcp:%d tcpfin:%d udp:%d\n", |
2212 | u->tcp_timeout, | |
2213 | u->tcp_fin_timeout, | |
2214 | u->udp_timeout); | |
2215 | ||
2216 | #ifdef CONFIG_IP_VS_PROTO_TCP | |
2217 | if (u->tcp_timeout) { | |
18d6ade6 | 2218 | pd = ip_vs_proto_data_get(ipvs, IPPROTO_TCP); |
9330419d | 2219 | pd->timeout_table[IP_VS_TCP_S_ESTABLISHED] |
1da177e4 LT |
2220 | = u->tcp_timeout * HZ; |
2221 | } | |
2222 | ||
2223 | if (u->tcp_fin_timeout) { | |
18d6ade6 | 2224 | pd = ip_vs_proto_data_get(ipvs, IPPROTO_TCP); |
9330419d | 2225 | pd->timeout_table[IP_VS_TCP_S_FIN_WAIT] |
1da177e4 LT |
2226 | = u->tcp_fin_timeout * HZ; |
2227 | } | |
2228 | #endif | |
2229 | ||
2230 | #ifdef CONFIG_IP_VS_PROTO_UDP | |
2231 | if (u->udp_timeout) { | |
18d6ade6 | 2232 | pd = ip_vs_proto_data_get(ipvs, IPPROTO_UDP); |
9330419d | 2233 | pd->timeout_table[IP_VS_UDP_S_NORMAL] |
1da177e4 LT |
2234 | = u->udp_timeout * HZ; |
2235 | } | |
2236 | #endif | |
2237 | return 0; | |
2238 | } | |
2239 | ||
5fcf0cf6 | 2240 | #define CMDID(cmd) (cmd - IP_VS_BASE_CTL) |
1da177e4 | 2241 | |
5fcf0cf6 JA |
2242 | struct ip_vs_svcdest_user { |
2243 | struct ip_vs_service_user s; | |
2244 | struct ip_vs_dest_user d; | |
1da177e4 LT |
2245 | }; |
2246 | ||
5fcf0cf6 JA |
2247 | static const unsigned char set_arglen[CMDID(IP_VS_SO_SET_MAX) + 1] = { |
2248 | [CMDID(IP_VS_SO_SET_ADD)] = sizeof(struct ip_vs_service_user), | |
2249 | [CMDID(IP_VS_SO_SET_EDIT)] = sizeof(struct ip_vs_service_user), | |
2250 | [CMDID(IP_VS_SO_SET_DEL)] = sizeof(struct ip_vs_service_user), | |
2251 | [CMDID(IP_VS_SO_SET_ADDDEST)] = sizeof(struct ip_vs_svcdest_user), | |
2252 | [CMDID(IP_VS_SO_SET_DELDEST)] = sizeof(struct ip_vs_svcdest_user), | |
2253 | [CMDID(IP_VS_SO_SET_EDITDEST)] = sizeof(struct ip_vs_svcdest_user), | |
2254 | [CMDID(IP_VS_SO_SET_TIMEOUT)] = sizeof(struct ip_vs_timeout_user), | |
2255 | [CMDID(IP_VS_SO_SET_STARTDAEMON)] = sizeof(struct ip_vs_daemon_user), | |
2256 | [CMDID(IP_VS_SO_SET_STOPDAEMON)] = sizeof(struct ip_vs_daemon_user), | |
2257 | [CMDID(IP_VS_SO_SET_ZERO)] = sizeof(struct ip_vs_service_user), | |
2258 | }; | |
2259 | ||
2260 | union ip_vs_set_arglen { | |
2261 | struct ip_vs_service_user field_IP_VS_SO_SET_ADD; | |
2262 | struct ip_vs_service_user field_IP_VS_SO_SET_EDIT; | |
2263 | struct ip_vs_service_user field_IP_VS_SO_SET_DEL; | |
2264 | struct ip_vs_svcdest_user field_IP_VS_SO_SET_ADDDEST; | |
2265 | struct ip_vs_svcdest_user field_IP_VS_SO_SET_DELDEST; | |
2266 | struct ip_vs_svcdest_user field_IP_VS_SO_SET_EDITDEST; | |
2267 | struct ip_vs_timeout_user field_IP_VS_SO_SET_TIMEOUT; | |
2268 | struct ip_vs_daemon_user field_IP_VS_SO_SET_STARTDAEMON; | |
2269 | struct ip_vs_daemon_user field_IP_VS_SO_SET_STOPDAEMON; | |
2270 | struct ip_vs_service_user field_IP_VS_SO_SET_ZERO; | |
2271 | }; | |
2272 | ||
2273 | #define MAX_SET_ARGLEN sizeof(union ip_vs_set_arglen) | |
2274 | ||
c860c6b1 JV |
2275 | static void ip_vs_copy_usvc_compat(struct ip_vs_service_user_kern *usvc, |
2276 | struct ip_vs_service_user *usvc_compat) | |
2277 | { | |
0d1e71b0 SH |
2278 | memset(usvc, 0, sizeof(*usvc)); |
2279 | ||
c860c6b1 JV |
2280 | usvc->af = AF_INET; |
2281 | usvc->protocol = usvc_compat->protocol; | |
2282 | usvc->addr.ip = usvc_compat->addr; | |
2283 | usvc->port = usvc_compat->port; | |
2284 | usvc->fwmark = usvc_compat->fwmark; | |
2285 | ||
2286 | /* Deep copy of sched_name is not needed here */ | |
2287 | usvc->sched_name = usvc_compat->sched_name; | |
2288 | ||
2289 | usvc->flags = usvc_compat->flags; | |
2290 | usvc->timeout = usvc_compat->timeout; | |
2291 | usvc->netmask = usvc_compat->netmask; | |
2292 | } | |
2293 | ||
2294 | static void ip_vs_copy_udest_compat(struct ip_vs_dest_user_kern *udest, | |
2295 | struct ip_vs_dest_user *udest_compat) | |
2296 | { | |
0d1e71b0 SH |
2297 | memset(udest, 0, sizeof(*udest)); |
2298 | ||
c860c6b1 JV |
2299 | udest->addr.ip = udest_compat->addr; |
2300 | udest->port = udest_compat->port; | |
2301 | udest->conn_flags = udest_compat->conn_flags; | |
2302 | udest->weight = udest_compat->weight; | |
2303 | udest->u_threshold = udest_compat->u_threshold; | |
2304 | udest->l_threshold = udest_compat->l_threshold; | |
6cff339b | 2305 | udest->af = AF_INET; |
c860c6b1 JV |
2306 | } |
2307 | ||
1da177e4 LT |
2308 | static int |
2309 | do_ip_vs_set_ctl(struct sock *sk, int cmd, void __user *user, unsigned int len) | |
2310 | { | |
fc723250 | 2311 | struct net *net = sock_net(sk); |
1da177e4 | 2312 | int ret; |
5fcf0cf6 | 2313 | unsigned char arg[MAX_SET_ARGLEN]; |
c860c6b1 JV |
2314 | struct ip_vs_service_user *usvc_compat; |
2315 | struct ip_vs_service_user_kern usvc; | |
1da177e4 | 2316 | struct ip_vs_service *svc; |
c860c6b1 JV |
2317 | struct ip_vs_dest_user *udest_compat; |
2318 | struct ip_vs_dest_user_kern udest; | |
ae1d48b2 | 2319 | struct netns_ipvs *ipvs = net_ipvs(net); |
1da177e4 | 2320 | |
5fcf0cf6 | 2321 | BUILD_BUG_ON(sizeof(arg) > 255); |
df008c91 | 2322 | if (!ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN)) |
1da177e4 LT |
2323 | return -EPERM; |
2324 | ||
04bcef2a AV |
2325 | if (cmd < IP_VS_BASE_CTL || cmd > IP_VS_SO_SET_MAX) |
2326 | return -EINVAL; | |
5fcf0cf6 JA |
2327 | if (len != set_arglen[CMDID(cmd)]) { |
2328 | IP_VS_DBG(1, "set_ctl: len %u != %u\n", | |
2329 | len, set_arglen[CMDID(cmd)]); | |
1da177e4 LT |
2330 | return -EINVAL; |
2331 | } | |
2332 | ||
2333 | if (copy_from_user(arg, user, len) != 0) | |
2334 | return -EFAULT; | |
2335 | ||
2336 | /* increase the module use count */ | |
2337 | ip_vs_use_count_inc(); | |
2338 | ||
ae1d48b2 HS |
2339 | /* Handle daemons since they have another lock */ |
2340 | if (cmd == IP_VS_SO_SET_STARTDAEMON || | |
2341 | cmd == IP_VS_SO_SET_STOPDAEMON) { | |
2342 | struct ip_vs_daemon_user *dm = (struct ip_vs_daemon_user *)arg; | |
2343 | ||
e0b26cc9 | 2344 | if (cmd == IP_VS_SO_SET_STARTDAEMON) { |
e4ff6751 JA |
2345 | struct ipvs_sync_daemon_cfg cfg; |
2346 | ||
2347 | memset(&cfg, 0, sizeof(cfg)); | |
2348 | strlcpy(cfg.mcast_ifn, dm->mcast_ifn, | |
2349 | sizeof(cfg.mcast_ifn)); | |
2350 | cfg.syncid = dm->syncid; | |
e0b26cc9 JA |
2351 | rtnl_lock(); |
2352 | mutex_lock(&ipvs->sync_mutex); | |
6ac121d7 | 2353 | ret = start_sync_thread(ipvs, &cfg, dm->state); |
e0b26cc9 JA |
2354 | mutex_unlock(&ipvs->sync_mutex); |
2355 | rtnl_unlock(); | |
2356 | } else { | |
2357 | mutex_lock(&ipvs->sync_mutex); | |
b3cf3cbf | 2358 | ret = stop_sync_thread(ipvs, dm->state); |
e0b26cc9 JA |
2359 | mutex_unlock(&ipvs->sync_mutex); |
2360 | } | |
ae1d48b2 HS |
2361 | goto out_dec; |
2362 | } | |
2363 | ||
7926dbfa | 2364 | mutex_lock(&__ip_vs_mutex); |
1da177e4 LT |
2365 | if (cmd == IP_VS_SO_SET_FLUSH) { |
2366 | /* Flush the virtual service */ | |
ef7c599d | 2367 | ret = ip_vs_flush(ipvs, false); |
1da177e4 LT |
2368 | goto out_unlock; |
2369 | } else if (cmd == IP_VS_SO_SET_TIMEOUT) { | |
2370 | /* Set timeout values for (tcp tcpfin udp) */ | |
f1faa1e7 | 2371 | ret = ip_vs_set_timeout(ipvs, (struct ip_vs_timeout_user *)arg); |
1da177e4 | 2372 | goto out_unlock; |
1da177e4 LT |
2373 | } |
2374 | ||
c860c6b1 JV |
2375 | usvc_compat = (struct ip_vs_service_user *)arg; |
2376 | udest_compat = (struct ip_vs_dest_user *)(usvc_compat + 1); | |
2377 | ||
2378 | /* We only use the new structs internally, so copy userspace compat | |
2379 | * structs to extended internal versions */ | |
2380 | ip_vs_copy_usvc_compat(&usvc, usvc_compat); | |
2381 | ip_vs_copy_udest_compat(&udest, udest_compat); | |
1da177e4 LT |
2382 | |
2383 | if (cmd == IP_VS_SO_SET_ZERO) { | |
2384 | /* if no service address is set, zero counters in all */ | |
c860c6b1 | 2385 | if (!usvc.fwmark && !usvc.addr.ip && !usvc.port) { |
c60856c6 | 2386 | ret = ip_vs_zero_all(ipvs); |
1da177e4 LT |
2387 | goto out_unlock; |
2388 | } | |
2389 | } | |
2390 | ||
2906f66a VMR |
2391 | /* Check for valid protocol: TCP or UDP or SCTP, even for fwmark!=0 */ |
2392 | if (usvc.protocol != IPPROTO_TCP && usvc.protocol != IPPROTO_UDP && | |
2393 | usvc.protocol != IPPROTO_SCTP) { | |
1e3e238e HE |
2394 | pr_err("set_ctl: invalid protocol: %d %pI4:%d %s\n", |
2395 | usvc.protocol, &usvc.addr.ip, | |
2396 | ntohs(usvc.port), usvc.sched_name); | |
1da177e4 LT |
2397 | ret = -EFAULT; |
2398 | goto out_unlock; | |
2399 | } | |
2400 | ||
2401 | /* Lookup the exact service by <protocol, addr, port> or fwmark */ | |
ceec4c38 | 2402 | rcu_read_lock(); |
c860c6b1 | 2403 | if (usvc.fwmark == 0) |
bb2e2a8c | 2404 | svc = __ip_vs_service_find(ipvs, usvc.af, usvc.protocol, |
26c15cfd | 2405 | &usvc.addr, usvc.port); |
1da177e4 | 2406 | else |
1ed8b947 | 2407 | svc = __ip_vs_svc_fwm_find(ipvs, usvc.af, usvc.fwmark); |
ceec4c38 | 2408 | rcu_read_unlock(); |
1da177e4 LT |
2409 | |
2410 | if (cmd != IP_VS_SO_SET_ADD | |
c860c6b1 | 2411 | && (svc == NULL || svc->protocol != usvc.protocol)) { |
1da177e4 | 2412 | ret = -ESRCH; |
26c15cfd | 2413 | goto out_unlock; |
1da177e4 LT |
2414 | } |
2415 | ||
2416 | switch (cmd) { | |
2417 | case IP_VS_SO_SET_ADD: | |
2418 | if (svc != NULL) | |
2419 | ret = -EEXIST; | |
2420 | else | |
5060bd83 | 2421 | ret = ip_vs_add_service(ipvs, &usvc, &svc); |
1da177e4 LT |
2422 | break; |
2423 | case IP_VS_SO_SET_EDIT: | |
c860c6b1 | 2424 | ret = ip_vs_edit_service(svc, &usvc); |
1da177e4 LT |
2425 | break; |
2426 | case IP_VS_SO_SET_DEL: | |
2427 | ret = ip_vs_del_service(svc); | |
2428 | if (!ret) | |
2429 | goto out_unlock; | |
2430 | break; | |
2431 | case IP_VS_SO_SET_ZERO: | |
2432 | ret = ip_vs_zero_service(svc); | |
2433 | break; | |
2434 | case IP_VS_SO_SET_ADDDEST: | |
c860c6b1 | 2435 | ret = ip_vs_add_dest(svc, &udest); |
1da177e4 LT |
2436 | break; |
2437 | case IP_VS_SO_SET_EDITDEST: | |
c860c6b1 | 2438 | ret = ip_vs_edit_dest(svc, &udest); |
1da177e4 LT |
2439 | break; |
2440 | case IP_VS_SO_SET_DELDEST: | |
c860c6b1 | 2441 | ret = ip_vs_del_dest(svc, &udest); |
1da177e4 LT |
2442 | break; |
2443 | default: | |
2444 | ret = -EINVAL; | |
2445 | } | |
2446 | ||
1da177e4 | 2447 | out_unlock: |
14cc3e2b | 2448 | mutex_unlock(&__ip_vs_mutex); |
1da177e4 LT |
2449 | out_dec: |
2450 | /* decrease the module use count */ | |
2451 | ip_vs_use_count_dec(); | |
2452 | ||
2453 | return ret; | |
2454 | } | |
2455 | ||
2456 | ||
1da177e4 LT |
2457 | static void |
2458 | ip_vs_copy_service(struct ip_vs_service_entry *dst, struct ip_vs_service *src) | |
2459 | { | |
ceec4c38 | 2460 | struct ip_vs_scheduler *sched; |
cd67cd5e | 2461 | struct ip_vs_kstats kstats; |
05f00505 | 2462 | char *sched_name; |
ceec4c38 JA |
2463 | |
2464 | sched = rcu_dereference_protected(src->scheduler, 1); | |
05f00505 | 2465 | sched_name = sched ? sched->name : "none"; |
1da177e4 | 2466 | dst->protocol = src->protocol; |
e7ade46a | 2467 | dst->addr = src->addr.ip; |
1da177e4 LT |
2468 | dst->port = src->port; |
2469 | dst->fwmark = src->fwmark; | |
05f00505 | 2470 | strlcpy(dst->sched_name, sched_name, sizeof(dst->sched_name)); |
1da177e4 LT |
2471 | dst->flags = src->flags; |
2472 | dst->timeout = src->timeout / HZ; | |
2473 | dst->netmask = src->netmask; | |
2474 | dst->num_dests = src->num_dests; | |
cd67cd5e JA |
2475 | ip_vs_copy_stats(&kstats, &src->stats); |
2476 | ip_vs_export_stats_user(&dst->stats, &kstats); | |
1da177e4 LT |
2477 | } |
2478 | ||
2479 | static inline int | |
b2876b77 | 2480 | __ip_vs_get_service_entries(struct netns_ipvs *ipvs, |
fc723250 | 2481 | const struct ip_vs_get_services *get, |
1da177e4 LT |
2482 | struct ip_vs_get_services __user *uptr) |
2483 | { | |
2484 | int idx, count=0; | |
2485 | struct ip_vs_service *svc; | |
2486 | struct ip_vs_service_entry entry; | |
2487 | int ret = 0; | |
2488 | ||
2489 | for (idx = 0; idx < IP_VS_SVC_TAB_SIZE; idx++) { | |
ceec4c38 | 2490 | hlist_for_each_entry(svc, &ip_vs_svc_table[idx], s_list) { |
f94fd041 | 2491 | /* Only expose IPv4 entries to old interface */ |
3109d2f2 | 2492 | if (svc->af != AF_INET || (svc->ipvs != ipvs)) |
f94fd041 JV |
2493 | continue; |
2494 | ||
1da177e4 LT |
2495 | if (count >= get->num_services) |
2496 | goto out; | |
4da62fc7 | 2497 | memset(&entry, 0, sizeof(entry)); |
1da177e4 LT |
2498 | ip_vs_copy_service(&entry, svc); |
2499 | if (copy_to_user(&uptr->entrytable[count], | |
2500 | &entry, sizeof(entry))) { | |
2501 | ret = -EFAULT; | |
2502 | goto out; | |
2503 | } | |
2504 | count++; | |
2505 | } | |
2506 | } | |
2507 | ||
2508 | for (idx = 0; idx < IP_VS_SVC_TAB_SIZE; idx++) { | |
ceec4c38 | 2509 | hlist_for_each_entry(svc, &ip_vs_svc_fwm_table[idx], f_list) { |
f94fd041 | 2510 | /* Only expose IPv4 entries to old interface */ |
3109d2f2 | 2511 | if (svc->af != AF_INET || (svc->ipvs != ipvs)) |
f94fd041 JV |
2512 | continue; |
2513 | ||
1da177e4 LT |
2514 | if (count >= get->num_services) |
2515 | goto out; | |
4da62fc7 | 2516 | memset(&entry, 0, sizeof(entry)); |
1da177e4 LT |
2517 | ip_vs_copy_service(&entry, svc); |
2518 | if (copy_to_user(&uptr->entrytable[count], | |
2519 | &entry, sizeof(entry))) { | |
2520 | ret = -EFAULT; | |
2521 | goto out; | |
2522 | } | |
2523 | count++; | |
2524 | } | |
2525 | } | |
552ad65a | 2526 | out: |
1da177e4 LT |
2527 | return ret; |
2528 | } | |
2529 | ||
2530 | static inline int | |
08fff4c3 | 2531 | __ip_vs_get_dest_entries(struct netns_ipvs *ipvs, const struct ip_vs_get_dests *get, |
1da177e4 LT |
2532 | struct ip_vs_get_dests __user *uptr) |
2533 | { | |
2534 | struct ip_vs_service *svc; | |
b18610de | 2535 | union nf_inet_addr addr = { .ip = get->addr }; |
1da177e4 LT |
2536 | int ret = 0; |
2537 | ||
ceec4c38 | 2538 | rcu_read_lock(); |
1da177e4 | 2539 | if (get->fwmark) |
1ed8b947 | 2540 | svc = __ip_vs_svc_fwm_find(ipvs, AF_INET, get->fwmark); |
1da177e4 | 2541 | else |
bb2e2a8c | 2542 | svc = __ip_vs_service_find(ipvs, AF_INET, get->protocol, &addr, |
26c15cfd | 2543 | get->port); |
ceec4c38 | 2544 | rcu_read_unlock(); |
b18610de | 2545 | |
1da177e4 LT |
2546 | if (svc) { |
2547 | int count = 0; | |
2548 | struct ip_vs_dest *dest; | |
2549 | struct ip_vs_dest_entry entry; | |
cd67cd5e | 2550 | struct ip_vs_kstats kstats; |
1da177e4 | 2551 | |
a8241c63 | 2552 | memset(&entry, 0, sizeof(entry)); |
1da177e4 LT |
2553 | list_for_each_entry(dest, &svc->destinations, n_list) { |
2554 | if (count >= get->num_dests) | |
2555 | break; | |
2556 | ||
6cff339b AG |
2557 | /* Cannot expose heterogeneous members via sockopt |
2558 | * interface | |
2559 | */ | |
2560 | if (dest->af != svc->af) | |
2561 | continue; | |
2562 | ||
e7ade46a | 2563 | entry.addr = dest->addr.ip; |
1da177e4 LT |
2564 | entry.port = dest->port; |
2565 | entry.conn_flags = atomic_read(&dest->conn_flags); | |
2566 | entry.weight = atomic_read(&dest->weight); | |
2567 | entry.u_threshold = dest->u_threshold; | |
2568 | entry.l_threshold = dest->l_threshold; | |
2569 | entry.activeconns = atomic_read(&dest->activeconns); | |
2570 | entry.inactconns = atomic_read(&dest->inactconns); | |
2571 | entry.persistconns = atomic_read(&dest->persistconns); | |
cd67cd5e JA |
2572 | ip_vs_copy_stats(&kstats, &dest->stats); |
2573 | ip_vs_export_stats_user(&entry.stats, &kstats); | |
1da177e4 LT |
2574 | if (copy_to_user(&uptr->entrytable[count], |
2575 | &entry, sizeof(entry))) { | |
2576 | ret = -EFAULT; | |
2577 | break; | |
2578 | } | |
2579 | count++; | |
2580 | } | |
1da177e4 LT |
2581 | } else |
2582 | ret = -ESRCH; | |
2583 | return ret; | |
2584 | } | |
2585 | ||
2586 | static inline void | |
af540341 | 2587 | __ip_vs_get_timeouts(struct netns_ipvs *ipvs, struct ip_vs_timeout_user *u) |
1da177e4 | 2588 | { |
091bb34c | 2589 | #if defined(CONFIG_IP_VS_PROTO_TCP) || defined(CONFIG_IP_VS_PROTO_UDP) |
9330419d | 2590 | struct ip_vs_proto_data *pd; |
091bb34c | 2591 | #endif |
9330419d | 2592 | |
b61a602e AB |
2593 | memset(u, 0, sizeof (*u)); |
2594 | ||
1da177e4 | 2595 | #ifdef CONFIG_IP_VS_PROTO_TCP |
18d6ade6 | 2596 | pd = ip_vs_proto_data_get(ipvs, IPPROTO_TCP); |
9330419d HS |
2597 | u->tcp_timeout = pd->timeout_table[IP_VS_TCP_S_ESTABLISHED] / HZ; |
2598 | u->tcp_fin_timeout = pd->timeout_table[IP_VS_TCP_S_FIN_WAIT] / HZ; | |
1da177e4 LT |
2599 | #endif |
2600 | #ifdef CONFIG_IP_VS_PROTO_UDP | |
18d6ade6 | 2601 | pd = ip_vs_proto_data_get(ipvs, IPPROTO_UDP); |
1da177e4 | 2602 | u->udp_timeout = |
9330419d | 2603 | pd->timeout_table[IP_VS_UDP_S_NORMAL] / HZ; |
1da177e4 LT |
2604 | #endif |
2605 | } | |
2606 | ||
5fcf0cf6 JA |
2607 | static const unsigned char get_arglen[CMDID(IP_VS_SO_GET_MAX) + 1] = { |
2608 | [CMDID(IP_VS_SO_GET_VERSION)] = 64, | |
2609 | [CMDID(IP_VS_SO_GET_INFO)] = sizeof(struct ip_vs_getinfo), | |
2610 | [CMDID(IP_VS_SO_GET_SERVICES)] = sizeof(struct ip_vs_get_services), | |
2611 | [CMDID(IP_VS_SO_GET_SERVICE)] = sizeof(struct ip_vs_service_entry), | |
2612 | [CMDID(IP_VS_SO_GET_DESTS)] = sizeof(struct ip_vs_get_dests), | |
2613 | [CMDID(IP_VS_SO_GET_TIMEOUT)] = sizeof(struct ip_vs_timeout_user), | |
2614 | [CMDID(IP_VS_SO_GET_DAEMON)] = 2 * sizeof(struct ip_vs_daemon_user), | |
2615 | }; | |
1da177e4 | 2616 | |
5fcf0cf6 JA |
2617 | union ip_vs_get_arglen { |
2618 | char field_IP_VS_SO_GET_VERSION[64]; | |
2619 | struct ip_vs_getinfo field_IP_VS_SO_GET_INFO; | |
2620 | struct ip_vs_get_services field_IP_VS_SO_GET_SERVICES; | |
2621 | struct ip_vs_service_entry field_IP_VS_SO_GET_SERVICE; | |
2622 | struct ip_vs_get_dests field_IP_VS_SO_GET_DESTS; | |
2623 | struct ip_vs_timeout_user field_IP_VS_SO_GET_TIMEOUT; | |
2624 | struct ip_vs_daemon_user field_IP_VS_SO_GET_DAEMON[2]; | |
1da177e4 LT |
2625 | }; |
2626 | ||
5fcf0cf6 JA |
2627 | #define MAX_GET_ARGLEN sizeof(union ip_vs_get_arglen) |
2628 | ||
1da177e4 LT |
2629 | static int |
2630 | do_ip_vs_get_ctl(struct sock *sk, int cmd, void __user *user, int *len) | |
2631 | { | |
5fcf0cf6 | 2632 | unsigned char arg[MAX_GET_ARGLEN]; |
1da177e4 | 2633 | int ret = 0; |
04bcef2a | 2634 | unsigned int copylen; |
fc723250 | 2635 | struct net *net = sock_net(sk); |
f131315f | 2636 | struct netns_ipvs *ipvs = net_ipvs(net); |
1da177e4 | 2637 | |
fc723250 | 2638 | BUG_ON(!net); |
5fcf0cf6 | 2639 | BUILD_BUG_ON(sizeof(arg) > 255); |
df008c91 | 2640 | if (!ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN)) |
1da177e4 LT |
2641 | return -EPERM; |
2642 | ||
04bcef2a AV |
2643 | if (cmd < IP_VS_BASE_CTL || cmd > IP_VS_SO_GET_MAX) |
2644 | return -EINVAL; | |
2645 | ||
5fcf0cf6 JA |
2646 | copylen = get_arglen[CMDID(cmd)]; |
2647 | if (*len < (int) copylen) { | |
2648 | IP_VS_DBG(1, "get_ctl: len %d < %u\n", *len, copylen); | |
1da177e4 LT |
2649 | return -EINVAL; |
2650 | } | |
2651 | ||
04bcef2a | 2652 | if (copy_from_user(arg, user, copylen) != 0) |
1da177e4 | 2653 | return -EFAULT; |
ae1d48b2 HS |
2654 | /* |
2655 | * Handle daemons first since it has its own locking | |
2656 | */ | |
2657 | if (cmd == IP_VS_SO_GET_DAEMON) { | |
2658 | struct ip_vs_daemon_user d[2]; | |
2659 | ||
2660 | memset(&d, 0, sizeof(d)); | |
7926dbfa | 2661 | mutex_lock(&ipvs->sync_mutex); |
ae1d48b2 HS |
2662 | if (ipvs->sync_state & IP_VS_STATE_MASTER) { |
2663 | d[0].state = IP_VS_STATE_MASTER; | |
e4ff6751 | 2664 | strlcpy(d[0].mcast_ifn, ipvs->mcfg.mcast_ifn, |
ae1d48b2 | 2665 | sizeof(d[0].mcast_ifn)); |
e4ff6751 | 2666 | d[0].syncid = ipvs->mcfg.syncid; |
ae1d48b2 HS |
2667 | } |
2668 | if (ipvs->sync_state & IP_VS_STATE_BACKUP) { | |
2669 | d[1].state = IP_VS_STATE_BACKUP; | |
e4ff6751 | 2670 | strlcpy(d[1].mcast_ifn, ipvs->bcfg.mcast_ifn, |
ae1d48b2 | 2671 | sizeof(d[1].mcast_ifn)); |
e4ff6751 | 2672 | d[1].syncid = ipvs->bcfg.syncid; |
ae1d48b2 HS |
2673 | } |
2674 | if (copy_to_user(user, &d, sizeof(d)) != 0) | |
2675 | ret = -EFAULT; | |
2676 | mutex_unlock(&ipvs->sync_mutex); | |
2677 | return ret; | |
2678 | } | |
1da177e4 | 2679 | |
7926dbfa | 2680 | mutex_lock(&__ip_vs_mutex); |
1da177e4 LT |
2681 | switch (cmd) { |
2682 | case IP_VS_SO_GET_VERSION: | |
2683 | { | |
2684 | char buf[64]; | |
2685 | ||
2686 | sprintf(buf, "IP Virtual Server version %d.%d.%d (size=%d)", | |
6f7edb48 | 2687 | NVERSION(IP_VS_VERSION_CODE), ip_vs_conn_tab_size); |
1da177e4 LT |
2688 | if (copy_to_user(user, buf, strlen(buf)+1) != 0) { |
2689 | ret = -EFAULT; | |
2690 | goto out; | |
2691 | } | |
2692 | *len = strlen(buf)+1; | |
2693 | } | |
2694 | break; | |
2695 | ||
2696 | case IP_VS_SO_GET_INFO: | |
2697 | { | |
2698 | struct ip_vs_getinfo info; | |
2699 | info.version = IP_VS_VERSION_CODE; | |
6f7edb48 | 2700 | info.size = ip_vs_conn_tab_size; |
a0840e2e | 2701 | info.num_services = ipvs->num_services; |
1da177e4 LT |
2702 | if (copy_to_user(user, &info, sizeof(info)) != 0) |
2703 | ret = -EFAULT; | |
2704 | } | |
2705 | break; | |
2706 | ||
2707 | case IP_VS_SO_GET_SERVICES: | |
2708 | { | |
2709 | struct ip_vs_get_services *get; | |
2710 | int size; | |
2711 | ||
2712 | get = (struct ip_vs_get_services *)arg; | |
2713 | size = sizeof(*get) + | |
2714 | sizeof(struct ip_vs_service_entry) * get->num_services; | |
2715 | if (*len != size) { | |
1e3e238e | 2716 | pr_err("length: %u != %u\n", *len, size); |
1da177e4 LT |
2717 | ret = -EINVAL; |
2718 | goto out; | |
2719 | } | |
b2876b77 | 2720 | ret = __ip_vs_get_service_entries(ipvs, get, user); |
1da177e4 LT |
2721 | } |
2722 | break; | |
2723 | ||
2724 | case IP_VS_SO_GET_SERVICE: | |
2725 | { | |
2726 | struct ip_vs_service_entry *entry; | |
2727 | struct ip_vs_service *svc; | |
b18610de | 2728 | union nf_inet_addr addr; |
1da177e4 LT |
2729 | |
2730 | entry = (struct ip_vs_service_entry *)arg; | |
b18610de | 2731 | addr.ip = entry->addr; |
ceec4c38 | 2732 | rcu_read_lock(); |
1da177e4 | 2733 | if (entry->fwmark) |
1ed8b947 | 2734 | svc = __ip_vs_svc_fwm_find(ipvs, AF_INET, entry->fwmark); |
1da177e4 | 2735 | else |
bb2e2a8c | 2736 | svc = __ip_vs_service_find(ipvs, AF_INET, |
fc723250 HS |
2737 | entry->protocol, &addr, |
2738 | entry->port); | |
ceec4c38 | 2739 | rcu_read_unlock(); |
1da177e4 LT |
2740 | if (svc) { |
2741 | ip_vs_copy_service(entry, svc); | |
2742 | if (copy_to_user(user, entry, sizeof(*entry)) != 0) | |
2743 | ret = -EFAULT; | |
1da177e4 LT |
2744 | } else |
2745 | ret = -ESRCH; | |
2746 | } | |
2747 | break; | |
2748 | ||
2749 | case IP_VS_SO_GET_DESTS: | |
2750 | { | |
2751 | struct ip_vs_get_dests *get; | |
2752 | int size; | |
2753 | ||
2754 | get = (struct ip_vs_get_dests *)arg; | |
2755 | size = sizeof(*get) + | |
2756 | sizeof(struct ip_vs_dest_entry) * get->num_dests; | |
2757 | if (*len != size) { | |
1e3e238e | 2758 | pr_err("length: %u != %u\n", *len, size); |
1da177e4 LT |
2759 | ret = -EINVAL; |
2760 | goto out; | |
2761 | } | |
08fff4c3 | 2762 | ret = __ip_vs_get_dest_entries(ipvs, get, user); |
1da177e4 LT |
2763 | } |
2764 | break; | |
2765 | ||
2766 | case IP_VS_SO_GET_TIMEOUT: | |
2767 | { | |
2768 | struct ip_vs_timeout_user t; | |
2769 | ||
af540341 | 2770 | __ip_vs_get_timeouts(ipvs, &t); |
1da177e4 LT |
2771 | if (copy_to_user(user, &t, sizeof(t)) != 0) |
2772 | ret = -EFAULT; | |
2773 | } | |
2774 | break; | |
2775 | ||
1da177e4 LT |
2776 | default: |
2777 | ret = -EINVAL; | |
2778 | } | |
2779 | ||
552ad65a | 2780 | out: |
14cc3e2b | 2781 | mutex_unlock(&__ip_vs_mutex); |
1da177e4 LT |
2782 | return ret; |
2783 | } | |
2784 | ||
2785 | ||
2786 | static struct nf_sockopt_ops ip_vs_sockopts = { | |
2787 | .pf = PF_INET, | |
2788 | .set_optmin = IP_VS_BASE_CTL, | |
2789 | .set_optmax = IP_VS_SO_SET_MAX+1, | |
2790 | .set = do_ip_vs_set_ctl, | |
2791 | .get_optmin = IP_VS_BASE_CTL, | |
2792 | .get_optmax = IP_VS_SO_GET_MAX+1, | |
2793 | .get = do_ip_vs_get_ctl, | |
16fcec35 | 2794 | .owner = THIS_MODULE, |
1da177e4 LT |
2795 | }; |
2796 | ||
9a812198 JV |
2797 | /* |
2798 | * Generic Netlink interface | |
2799 | */ | |
2800 | ||
2801 | /* IPVS genetlink family */ | |
2802 | static struct genl_family ip_vs_genl_family = { | |
2803 | .id = GENL_ID_GENERATE, | |
2804 | .hdrsize = 0, | |
2805 | .name = IPVS_GENL_NAME, | |
2806 | .version = IPVS_GENL_VERSION, | |
2807 | .maxattr = IPVS_CMD_MAX, | |
c6d2d445 | 2808 | .netnsok = true, /* Make ipvsadm to work on netns */ |
9a812198 JV |
2809 | }; |
2810 | ||
2811 | /* Policy used for first-level command attributes */ | |
2812 | static const struct nla_policy ip_vs_cmd_policy[IPVS_CMD_ATTR_MAX + 1] = { | |
2813 | [IPVS_CMD_ATTR_SERVICE] = { .type = NLA_NESTED }, | |
2814 | [IPVS_CMD_ATTR_DEST] = { .type = NLA_NESTED }, | |
2815 | [IPVS_CMD_ATTR_DAEMON] = { .type = NLA_NESTED }, | |
2816 | [IPVS_CMD_ATTR_TIMEOUT_TCP] = { .type = NLA_U32 }, | |
2817 | [IPVS_CMD_ATTR_TIMEOUT_TCP_FIN] = { .type = NLA_U32 }, | |
2818 | [IPVS_CMD_ATTR_TIMEOUT_UDP] = { .type = NLA_U32 }, | |
2819 | }; | |
2820 | ||
2821 | /* Policy used for attributes in nested attribute IPVS_CMD_ATTR_DAEMON */ | |
2822 | static const struct nla_policy ip_vs_daemon_policy[IPVS_DAEMON_ATTR_MAX + 1] = { | |
2823 | [IPVS_DAEMON_ATTR_STATE] = { .type = NLA_U32 }, | |
2824 | [IPVS_DAEMON_ATTR_MCAST_IFN] = { .type = NLA_NUL_STRING, | |
2825 | .len = IP_VS_IFNAME_MAXLEN }, | |
2826 | [IPVS_DAEMON_ATTR_SYNC_ID] = { .type = NLA_U32 }, | |
e4ff6751 | 2827 | [IPVS_DAEMON_ATTR_SYNC_MAXLEN] = { .type = NLA_U16 }, |
d3328817 JA |
2828 | [IPVS_DAEMON_ATTR_MCAST_GROUP] = { .type = NLA_U32 }, |
2829 | [IPVS_DAEMON_ATTR_MCAST_GROUP6] = { .len = sizeof(struct in6_addr) }, | |
2830 | [IPVS_DAEMON_ATTR_MCAST_PORT] = { .type = NLA_U16 }, | |
2831 | [IPVS_DAEMON_ATTR_MCAST_TTL] = { .type = NLA_U8 }, | |
9a812198 JV |
2832 | }; |
2833 | ||
2834 | /* Policy used for attributes in nested attribute IPVS_CMD_ATTR_SERVICE */ | |
2835 | static const struct nla_policy ip_vs_svc_policy[IPVS_SVC_ATTR_MAX + 1] = { | |
2836 | [IPVS_SVC_ATTR_AF] = { .type = NLA_U16 }, | |
2837 | [IPVS_SVC_ATTR_PROTOCOL] = { .type = NLA_U16 }, | |
2838 | [IPVS_SVC_ATTR_ADDR] = { .type = NLA_BINARY, | |
2839 | .len = sizeof(union nf_inet_addr) }, | |
2840 | [IPVS_SVC_ATTR_PORT] = { .type = NLA_U16 }, | |
2841 | [IPVS_SVC_ATTR_FWMARK] = { .type = NLA_U32 }, | |
2842 | [IPVS_SVC_ATTR_SCHED_NAME] = { .type = NLA_NUL_STRING, | |
2843 | .len = IP_VS_SCHEDNAME_MAXLEN }, | |
0d1e71b0 SH |
2844 | [IPVS_SVC_ATTR_PE_NAME] = { .type = NLA_NUL_STRING, |
2845 | .len = IP_VS_PENAME_MAXLEN }, | |
9a812198 JV |
2846 | [IPVS_SVC_ATTR_FLAGS] = { .type = NLA_BINARY, |
2847 | .len = sizeof(struct ip_vs_flags) }, | |
2848 | [IPVS_SVC_ATTR_TIMEOUT] = { .type = NLA_U32 }, | |
2849 | [IPVS_SVC_ATTR_NETMASK] = { .type = NLA_U32 }, | |
2850 | [IPVS_SVC_ATTR_STATS] = { .type = NLA_NESTED }, | |
2851 | }; | |
2852 | ||
2853 | /* Policy used for attributes in nested attribute IPVS_CMD_ATTR_DEST */ | |
2854 | static const struct nla_policy ip_vs_dest_policy[IPVS_DEST_ATTR_MAX + 1] = { | |
2855 | [IPVS_DEST_ATTR_ADDR] = { .type = NLA_BINARY, | |
2856 | .len = sizeof(union nf_inet_addr) }, | |
2857 | [IPVS_DEST_ATTR_PORT] = { .type = NLA_U16 }, | |
2858 | [IPVS_DEST_ATTR_FWD_METHOD] = { .type = NLA_U32 }, | |
2859 | [IPVS_DEST_ATTR_WEIGHT] = { .type = NLA_U32 }, | |
2860 | [IPVS_DEST_ATTR_U_THRESH] = { .type = NLA_U32 }, | |
2861 | [IPVS_DEST_ATTR_L_THRESH] = { .type = NLA_U32 }, | |
2862 | [IPVS_DEST_ATTR_ACTIVE_CONNS] = { .type = NLA_U32 }, | |
2863 | [IPVS_DEST_ATTR_INACT_CONNS] = { .type = NLA_U32 }, | |
2864 | [IPVS_DEST_ATTR_PERSIST_CONNS] = { .type = NLA_U32 }, | |
2865 | [IPVS_DEST_ATTR_STATS] = { .type = NLA_NESTED }, | |
6cff339b | 2866 | [IPVS_DEST_ATTR_ADDR_FAMILY] = { .type = NLA_U16 }, |
9a812198 JV |
2867 | }; |
2868 | ||
2869 | static int ip_vs_genl_fill_stats(struct sk_buff *skb, int container_type, | |
cd67cd5e JA |
2870 | struct ip_vs_kstats *kstats) |
2871 | { | |
2872 | struct nlattr *nl_stats = nla_nest_start(skb, container_type); | |
2873 | ||
2874 | if (!nl_stats) | |
2875 | return -EMSGSIZE; | |
2876 | ||
2877 | if (nla_put_u32(skb, IPVS_STATS_ATTR_CONNS, (u32)kstats->conns) || | |
2878 | nla_put_u32(skb, IPVS_STATS_ATTR_INPKTS, (u32)kstats->inpkts) || | |
2879 | nla_put_u32(skb, IPVS_STATS_ATTR_OUTPKTS, (u32)kstats->outpkts) || | |
2880 | nla_put_u64(skb, IPVS_STATS_ATTR_INBYTES, kstats->inbytes) || | |
2881 | nla_put_u64(skb, IPVS_STATS_ATTR_OUTBYTES, kstats->outbytes) || | |
2882 | nla_put_u32(skb, IPVS_STATS_ATTR_CPS, (u32)kstats->cps) || | |
2883 | nla_put_u32(skb, IPVS_STATS_ATTR_INPPS, (u32)kstats->inpps) || | |
2884 | nla_put_u32(skb, IPVS_STATS_ATTR_OUTPPS, (u32)kstats->outpps) || | |
2885 | nla_put_u32(skb, IPVS_STATS_ATTR_INBPS, (u32)kstats->inbps) || | |
2886 | nla_put_u32(skb, IPVS_STATS_ATTR_OUTBPS, (u32)kstats->outbps)) | |
2887 | goto nla_put_failure; | |
2888 | nla_nest_end(skb, nl_stats); | |
2889 | ||
2890 | return 0; | |
2891 | ||
2892 | nla_put_failure: | |
2893 | nla_nest_cancel(skb, nl_stats); | |
2894 | return -EMSGSIZE; | |
2895 | } | |
2896 | ||
2897 | static int ip_vs_genl_fill_stats64(struct sk_buff *skb, int container_type, | |
2898 | struct ip_vs_kstats *kstats) | |
9a812198 JV |
2899 | { |
2900 | struct nlattr *nl_stats = nla_nest_start(skb, container_type); | |
cd67cd5e | 2901 | |
9a812198 JV |
2902 | if (!nl_stats) |
2903 | return -EMSGSIZE; | |
2904 | ||
cd67cd5e JA |
2905 | if (nla_put_u64(skb, IPVS_STATS_ATTR_CONNS, kstats->conns) || |
2906 | nla_put_u64(skb, IPVS_STATS_ATTR_INPKTS, kstats->inpkts) || | |
2907 | nla_put_u64(skb, IPVS_STATS_ATTR_OUTPKTS, kstats->outpkts) || | |
2908 | nla_put_u64(skb, IPVS_STATS_ATTR_INBYTES, kstats->inbytes) || | |
2909 | nla_put_u64(skb, IPVS_STATS_ATTR_OUTBYTES, kstats->outbytes) || | |
2910 | nla_put_u64(skb, IPVS_STATS_ATTR_CPS, kstats->cps) || | |
2911 | nla_put_u64(skb, IPVS_STATS_ATTR_INPPS, kstats->inpps) || | |
2912 | nla_put_u64(skb, IPVS_STATS_ATTR_OUTPPS, kstats->outpps) || | |
2913 | nla_put_u64(skb, IPVS_STATS_ATTR_INBPS, kstats->inbps) || | |
2914 | nla_put_u64(skb, IPVS_STATS_ATTR_OUTBPS, kstats->outbps)) | |
969e8e25 | 2915 | goto nla_put_failure; |
9a812198 JV |
2916 | nla_nest_end(skb, nl_stats); |
2917 | ||
2918 | return 0; | |
2919 | ||
2920 | nla_put_failure: | |
9a812198 JV |
2921 | nla_nest_cancel(skb, nl_stats); |
2922 | return -EMSGSIZE; | |
2923 | } | |
2924 | ||
2925 | static int ip_vs_genl_fill_service(struct sk_buff *skb, | |
2926 | struct ip_vs_service *svc) | |
2927 | { | |
ceec4c38 | 2928 | struct ip_vs_scheduler *sched; |
371990ee | 2929 | struct ip_vs_pe *pe; |
9a812198 JV |
2930 | struct nlattr *nl_service; |
2931 | struct ip_vs_flags flags = { .flags = svc->flags, | |
2932 | .mask = ~0 }; | |
cd67cd5e | 2933 | struct ip_vs_kstats kstats; |
05f00505 | 2934 | char *sched_name; |
9a812198 JV |
2935 | |
2936 | nl_service = nla_nest_start(skb, IPVS_CMD_ATTR_SERVICE); | |
2937 | if (!nl_service) | |
2938 | return -EMSGSIZE; | |
2939 | ||
969e8e25 DM |
2940 | if (nla_put_u16(skb, IPVS_SVC_ATTR_AF, svc->af)) |
2941 | goto nla_put_failure; | |
9a812198 | 2942 | if (svc->fwmark) { |
969e8e25 DM |
2943 | if (nla_put_u32(skb, IPVS_SVC_ATTR_FWMARK, svc->fwmark)) |
2944 | goto nla_put_failure; | |
9a812198 | 2945 | } else { |
969e8e25 DM |
2946 | if (nla_put_u16(skb, IPVS_SVC_ATTR_PROTOCOL, svc->protocol) || |
2947 | nla_put(skb, IPVS_SVC_ATTR_ADDR, sizeof(svc->addr), &svc->addr) || | |
0a925864 | 2948 | nla_put_be16(skb, IPVS_SVC_ATTR_PORT, svc->port)) |
969e8e25 | 2949 | goto nla_put_failure; |
9a812198 JV |
2950 | } |
2951 | ||
ceec4c38 | 2952 | sched = rcu_dereference_protected(svc->scheduler, 1); |
05f00505 | 2953 | sched_name = sched ? sched->name : "none"; |
371990ee | 2954 | pe = rcu_dereference_protected(svc->pe, 1); |
05f00505 | 2955 | if (nla_put_string(skb, IPVS_SVC_ATTR_SCHED_NAME, sched_name) || |
371990ee | 2956 | (pe && nla_put_string(skb, IPVS_SVC_ATTR_PE_NAME, pe->name)) || |
969e8e25 DM |
2957 | nla_put(skb, IPVS_SVC_ATTR_FLAGS, sizeof(flags), &flags) || |
2958 | nla_put_u32(skb, IPVS_SVC_ATTR_TIMEOUT, svc->timeout / HZ) || | |
0a925864 | 2959 | nla_put_be32(skb, IPVS_SVC_ATTR_NETMASK, svc->netmask)) |
969e8e25 | 2960 | goto nla_put_failure; |
cd67cd5e JA |
2961 | ip_vs_copy_stats(&kstats, &svc->stats); |
2962 | if (ip_vs_genl_fill_stats(skb, IPVS_SVC_ATTR_STATS, &kstats)) | |
2963 | goto nla_put_failure; | |
2964 | if (ip_vs_genl_fill_stats64(skb, IPVS_SVC_ATTR_STATS64, &kstats)) | |
9a812198 JV |
2965 | goto nla_put_failure; |
2966 | ||
2967 | nla_nest_end(skb, nl_service); | |
2968 | ||
2969 | return 0; | |
2970 | ||
2971 | nla_put_failure: | |
2972 | nla_nest_cancel(skb, nl_service); | |
2973 | return -EMSGSIZE; | |
2974 | } | |
2975 | ||
2976 | static int ip_vs_genl_dump_service(struct sk_buff *skb, | |
2977 | struct ip_vs_service *svc, | |
2978 | struct netlink_callback *cb) | |
2979 | { | |
2980 | void *hdr; | |
2981 | ||
15e47304 | 2982 | hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq, |
9a812198 JV |
2983 | &ip_vs_genl_family, NLM_F_MULTI, |
2984 | IPVS_CMD_NEW_SERVICE); | |
2985 | if (!hdr) | |
2986 | return -EMSGSIZE; | |
2987 | ||
2988 | if (ip_vs_genl_fill_service(skb, svc) < 0) | |
2989 | goto nla_put_failure; | |
2990 | ||
053c095a JB |
2991 | genlmsg_end(skb, hdr); |
2992 | return 0; | |
9a812198 JV |
2993 | |
2994 | nla_put_failure: | |
2995 | genlmsg_cancel(skb, hdr); | |
2996 | return -EMSGSIZE; | |
2997 | } | |
2998 | ||
2999 | static int ip_vs_genl_dump_services(struct sk_buff *skb, | |
3000 | struct netlink_callback *cb) | |
3001 | { | |
3002 | int idx = 0, i; | |
3003 | int start = cb->args[0]; | |
3004 | struct ip_vs_service *svc; | |
3109d2f2 | 3005 | struct netns_ipvs *ipvs = net_ipvs(skb_sknet(skb)); |
9a812198 JV |
3006 | |
3007 | mutex_lock(&__ip_vs_mutex); | |
3008 | for (i = 0; i < IP_VS_SVC_TAB_SIZE; i++) { | |
ceec4c38 | 3009 | hlist_for_each_entry(svc, &ip_vs_svc_table[i], s_list) { |
3109d2f2 | 3010 | if (++idx <= start || (svc->ipvs != ipvs)) |
9a812198 JV |
3011 | continue; |
3012 | if (ip_vs_genl_dump_service(skb, svc, cb) < 0) { | |
3013 | idx--; | |
3014 | goto nla_put_failure; | |
3015 | } | |
3016 | } | |
3017 | } | |
3018 | ||
3019 | for (i = 0; i < IP_VS_SVC_TAB_SIZE; i++) { | |
ceec4c38 | 3020 | hlist_for_each_entry(svc, &ip_vs_svc_fwm_table[i], f_list) { |
3109d2f2 | 3021 | if (++idx <= start || (svc->ipvs != ipvs)) |
9a812198 JV |
3022 | continue; |
3023 | if (ip_vs_genl_dump_service(skb, svc, cb) < 0) { | |
3024 | idx--; | |
3025 | goto nla_put_failure; | |
3026 | } | |
3027 | } | |
3028 | } | |
3029 | ||
3030 | nla_put_failure: | |
3031 | mutex_unlock(&__ip_vs_mutex); | |
3032 | cb->args[0] = idx; | |
3033 | ||
3034 | return skb->len; | |
3035 | } | |
3036 | ||
613fb830 | 3037 | static int ip_vs_genl_parse_service(struct netns_ipvs *ipvs, |
fc723250 | 3038 | struct ip_vs_service_user_kern *usvc, |
26c15cfd JA |
3039 | struct nlattr *nla, int full_entry, |
3040 | struct ip_vs_service **ret_svc) | |
9a812198 JV |
3041 | { |
3042 | struct nlattr *attrs[IPVS_SVC_ATTR_MAX + 1]; | |
3043 | struct nlattr *nla_af, *nla_port, *nla_fwmark, *nla_protocol, *nla_addr; | |
26c15cfd | 3044 | struct ip_vs_service *svc; |
9a812198 JV |
3045 | |
3046 | /* Parse mandatory identifying service fields first */ | |
3047 | if (nla == NULL || | |
3048 | nla_parse_nested(attrs, IPVS_SVC_ATTR_MAX, nla, ip_vs_svc_policy)) | |
3049 | return -EINVAL; | |
3050 | ||
3051 | nla_af = attrs[IPVS_SVC_ATTR_AF]; | |
3052 | nla_protocol = attrs[IPVS_SVC_ATTR_PROTOCOL]; | |
3053 | nla_addr = attrs[IPVS_SVC_ATTR_ADDR]; | |
3054 | nla_port = attrs[IPVS_SVC_ATTR_PORT]; | |
3055 | nla_fwmark = attrs[IPVS_SVC_ATTR_FWMARK]; | |
3056 | ||
3057 | if (!(nla_af && (nla_fwmark || (nla_port && nla_protocol && nla_addr)))) | |
3058 | return -EINVAL; | |
3059 | ||
258c8893 SH |
3060 | memset(usvc, 0, sizeof(*usvc)); |
3061 | ||
c860c6b1 | 3062 | usvc->af = nla_get_u16(nla_af); |
f94fd041 JV |
3063 | #ifdef CONFIG_IP_VS_IPV6 |
3064 | if (usvc->af != AF_INET && usvc->af != AF_INET6) | |
3065 | #else | |
3066 | if (usvc->af != AF_INET) | |
3067 | #endif | |
9a812198 JV |
3068 | return -EAFNOSUPPORT; |
3069 | ||
3070 | if (nla_fwmark) { | |
3071 | usvc->protocol = IPPROTO_TCP; | |
3072 | usvc->fwmark = nla_get_u32(nla_fwmark); | |
3073 | } else { | |
3074 | usvc->protocol = nla_get_u16(nla_protocol); | |
3075 | nla_memcpy(&usvc->addr, nla_addr, sizeof(usvc->addr)); | |
0a925864 | 3076 | usvc->port = nla_get_be16(nla_port); |
9a812198 JV |
3077 | usvc->fwmark = 0; |
3078 | } | |
3079 | ||
ceec4c38 | 3080 | rcu_read_lock(); |
26c15cfd | 3081 | if (usvc->fwmark) |
1ed8b947 | 3082 | svc = __ip_vs_svc_fwm_find(ipvs, usvc->af, usvc->fwmark); |
26c15cfd | 3083 | else |
bb2e2a8c | 3084 | svc = __ip_vs_service_find(ipvs, usvc->af, usvc->protocol, |
26c15cfd | 3085 | &usvc->addr, usvc->port); |
ceec4c38 | 3086 | rcu_read_unlock(); |
26c15cfd JA |
3087 | *ret_svc = svc; |
3088 | ||
9a812198 JV |
3089 | /* If a full entry was requested, check for the additional fields */ |
3090 | if (full_entry) { | |
0d1e71b0 | 3091 | struct nlattr *nla_sched, *nla_flags, *nla_pe, *nla_timeout, |
9a812198 JV |
3092 | *nla_netmask; |
3093 | struct ip_vs_flags flags; | |
9a812198 JV |
3094 | |
3095 | nla_sched = attrs[IPVS_SVC_ATTR_SCHED_NAME]; | |
0d1e71b0 | 3096 | nla_pe = attrs[IPVS_SVC_ATTR_PE_NAME]; |
9a812198 JV |
3097 | nla_flags = attrs[IPVS_SVC_ATTR_FLAGS]; |
3098 | nla_timeout = attrs[IPVS_SVC_ATTR_TIMEOUT]; | |
3099 | nla_netmask = attrs[IPVS_SVC_ATTR_NETMASK]; | |
3100 | ||
3101 | if (!(nla_sched && nla_flags && nla_timeout && nla_netmask)) | |
3102 | return -EINVAL; | |
3103 | ||
3104 | nla_memcpy(&flags, nla_flags, sizeof(flags)); | |
3105 | ||
3106 | /* prefill flags from service if it already exists */ | |
26c15cfd | 3107 | if (svc) |
9a812198 | 3108 | usvc->flags = svc->flags; |
9a812198 JV |
3109 | |
3110 | /* set new flags from userland */ | |
3111 | usvc->flags = (usvc->flags & ~flags.mask) | | |
3112 | (flags.flags & flags.mask); | |
c860c6b1 | 3113 | usvc->sched_name = nla_data(nla_sched); |
0d1e71b0 | 3114 | usvc->pe_name = nla_pe ? nla_data(nla_pe) : NULL; |
9a812198 | 3115 | usvc->timeout = nla_get_u32(nla_timeout); |
0a925864 | 3116 | usvc->netmask = nla_get_be32(nla_netmask); |
9a812198 JV |
3117 | } |
3118 | ||
3119 | return 0; | |
3120 | } | |
3121 | ||
34c2f514 | 3122 | static struct ip_vs_service *ip_vs_genl_find_service(struct netns_ipvs *ipvs, |
fc723250 | 3123 | struct nlattr *nla) |
9a812198 | 3124 | { |
c860c6b1 | 3125 | struct ip_vs_service_user_kern usvc; |
26c15cfd | 3126 | struct ip_vs_service *svc; |
9a812198 JV |
3127 | int ret; |
3128 | ||
613fb830 | 3129 | ret = ip_vs_genl_parse_service(ipvs, &usvc, nla, 0, &svc); |
26c15cfd | 3130 | return ret ? ERR_PTR(ret) : svc; |
9a812198 JV |
3131 | } |
3132 | ||
3133 | static int ip_vs_genl_fill_dest(struct sk_buff *skb, struct ip_vs_dest *dest) | |
3134 | { | |
3135 | struct nlattr *nl_dest; | |
cd67cd5e | 3136 | struct ip_vs_kstats kstats; |
9a812198 JV |
3137 | |
3138 | nl_dest = nla_nest_start(skb, IPVS_CMD_ATTR_DEST); | |
3139 | if (!nl_dest) | |
3140 | return -EMSGSIZE; | |
3141 | ||
969e8e25 | 3142 | if (nla_put(skb, IPVS_DEST_ATTR_ADDR, sizeof(dest->addr), &dest->addr) || |
0a925864 | 3143 | nla_put_be16(skb, IPVS_DEST_ATTR_PORT, dest->port) || |
969e8e25 DM |
3144 | nla_put_u32(skb, IPVS_DEST_ATTR_FWD_METHOD, |
3145 | (atomic_read(&dest->conn_flags) & | |
3146 | IP_VS_CONN_F_FWD_MASK)) || | |
3147 | nla_put_u32(skb, IPVS_DEST_ATTR_WEIGHT, | |
3148 | atomic_read(&dest->weight)) || | |
3149 | nla_put_u32(skb, IPVS_DEST_ATTR_U_THRESH, dest->u_threshold) || | |
3150 | nla_put_u32(skb, IPVS_DEST_ATTR_L_THRESH, dest->l_threshold) || | |
3151 | nla_put_u32(skb, IPVS_DEST_ATTR_ACTIVE_CONNS, | |
3152 | atomic_read(&dest->activeconns)) || | |
3153 | nla_put_u32(skb, IPVS_DEST_ATTR_INACT_CONNS, | |
3154 | atomic_read(&dest->inactconns)) || | |
3155 | nla_put_u32(skb, IPVS_DEST_ATTR_PERSIST_CONNS, | |
6cff339b AG |
3156 | atomic_read(&dest->persistconns)) || |
3157 | nla_put_u16(skb, IPVS_DEST_ATTR_ADDR_FAMILY, dest->af)) | |
969e8e25 | 3158 | goto nla_put_failure; |
cd67cd5e JA |
3159 | ip_vs_copy_stats(&kstats, &dest->stats); |
3160 | if (ip_vs_genl_fill_stats(skb, IPVS_DEST_ATTR_STATS, &kstats)) | |
3161 | goto nla_put_failure; | |
3162 | if (ip_vs_genl_fill_stats64(skb, IPVS_DEST_ATTR_STATS64, &kstats)) | |
9a812198 JV |
3163 | goto nla_put_failure; |
3164 | ||
3165 | nla_nest_end(skb, nl_dest); | |
3166 | ||
3167 | return 0; | |
3168 | ||
3169 | nla_put_failure: | |
3170 | nla_nest_cancel(skb, nl_dest); | |
3171 | return -EMSGSIZE; | |
3172 | } | |
3173 | ||
3174 | static int ip_vs_genl_dump_dest(struct sk_buff *skb, struct ip_vs_dest *dest, | |
3175 | struct netlink_callback *cb) | |
3176 | { | |
3177 | void *hdr; | |
3178 | ||
15e47304 | 3179 | hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq, |
9a812198 JV |
3180 | &ip_vs_genl_family, NLM_F_MULTI, |
3181 | IPVS_CMD_NEW_DEST); | |
3182 | if (!hdr) | |
3183 | return -EMSGSIZE; | |
3184 | ||
3185 | if (ip_vs_genl_fill_dest(skb, dest) < 0) | |
3186 | goto nla_put_failure; | |
3187 | ||
053c095a JB |
3188 | genlmsg_end(skb, hdr); |
3189 | return 0; | |
9a812198 JV |
3190 | |
3191 | nla_put_failure: | |
3192 | genlmsg_cancel(skb, hdr); | |
3193 | return -EMSGSIZE; | |
3194 | } | |
3195 | ||
3196 | static int ip_vs_genl_dump_dests(struct sk_buff *skb, | |
3197 | struct netlink_callback *cb) | |
3198 | { | |
3199 | int idx = 0; | |
3200 | int start = cb->args[0]; | |
3201 | struct ip_vs_service *svc; | |
3202 | struct ip_vs_dest *dest; | |
3203 | struct nlattr *attrs[IPVS_CMD_ATTR_MAX + 1]; | |
a0840e2e | 3204 | struct net *net = skb_sknet(skb); |
34c2f514 | 3205 | struct netns_ipvs *ipvs = net_ipvs(net); |
9a812198 JV |
3206 | |
3207 | mutex_lock(&__ip_vs_mutex); | |
3208 | ||
3209 | /* Try to find the service for which to dump destinations */ | |
3210 | if (nlmsg_parse(cb->nlh, GENL_HDRLEN, attrs, | |
3211 | IPVS_CMD_ATTR_MAX, ip_vs_cmd_policy)) | |
3212 | goto out_err; | |
3213 | ||
a0840e2e | 3214 | |
34c2f514 | 3215 | svc = ip_vs_genl_find_service(ipvs, attrs[IPVS_CMD_ATTR_SERVICE]); |
9a812198 JV |
3216 | if (IS_ERR(svc) || svc == NULL) |
3217 | goto out_err; | |
3218 | ||
3219 | /* Dump the destinations */ | |
3220 | list_for_each_entry(dest, &svc->destinations, n_list) { | |
3221 | if (++idx <= start) | |
3222 | continue; | |
3223 | if (ip_vs_genl_dump_dest(skb, dest, cb) < 0) { | |
3224 | idx--; | |
3225 | goto nla_put_failure; | |
3226 | } | |
3227 | } | |
3228 | ||
3229 | nla_put_failure: | |
3230 | cb->args[0] = idx; | |
9a812198 JV |
3231 | |
3232 | out_err: | |
3233 | mutex_unlock(&__ip_vs_mutex); | |
3234 | ||
3235 | return skb->len; | |
3236 | } | |
3237 | ||
c860c6b1 | 3238 | static int ip_vs_genl_parse_dest(struct ip_vs_dest_user_kern *udest, |
9a812198 JV |
3239 | struct nlattr *nla, int full_entry) |
3240 | { | |
3241 | struct nlattr *attrs[IPVS_DEST_ATTR_MAX + 1]; | |
3242 | struct nlattr *nla_addr, *nla_port; | |
6cff339b | 3243 | struct nlattr *nla_addr_family; |
9a812198 JV |
3244 | |
3245 | /* Parse mandatory identifying destination fields first */ | |
3246 | if (nla == NULL || | |
3247 | nla_parse_nested(attrs, IPVS_DEST_ATTR_MAX, nla, ip_vs_dest_policy)) | |
3248 | return -EINVAL; | |
3249 | ||
3250 | nla_addr = attrs[IPVS_DEST_ATTR_ADDR]; | |
3251 | nla_port = attrs[IPVS_DEST_ATTR_PORT]; | |
6cff339b | 3252 | nla_addr_family = attrs[IPVS_DEST_ATTR_ADDR_FAMILY]; |
9a812198 JV |
3253 | |
3254 | if (!(nla_addr && nla_port)) | |
3255 | return -EINVAL; | |
3256 | ||
258c8893 SH |
3257 | memset(udest, 0, sizeof(*udest)); |
3258 | ||
9a812198 | 3259 | nla_memcpy(&udest->addr, nla_addr, sizeof(udest->addr)); |
0a925864 | 3260 | udest->port = nla_get_be16(nla_port); |
9a812198 | 3261 | |
6cff339b AG |
3262 | if (nla_addr_family) |
3263 | udest->af = nla_get_u16(nla_addr_family); | |
3264 | else | |
3265 | udest->af = 0; | |
3266 | ||
9a812198 JV |
3267 | /* If a full entry was requested, check for the additional fields */ |
3268 | if (full_entry) { | |
3269 | struct nlattr *nla_fwd, *nla_weight, *nla_u_thresh, | |
3270 | *nla_l_thresh; | |
3271 | ||
3272 | nla_fwd = attrs[IPVS_DEST_ATTR_FWD_METHOD]; | |
3273 | nla_weight = attrs[IPVS_DEST_ATTR_WEIGHT]; | |
3274 | nla_u_thresh = attrs[IPVS_DEST_ATTR_U_THRESH]; | |
3275 | nla_l_thresh = attrs[IPVS_DEST_ATTR_L_THRESH]; | |
3276 | ||
3277 | if (!(nla_fwd && nla_weight && nla_u_thresh && nla_l_thresh)) | |
3278 | return -EINVAL; | |
3279 | ||
3280 | udest->conn_flags = nla_get_u32(nla_fwd) | |
3281 | & IP_VS_CONN_F_FWD_MASK; | |
3282 | udest->weight = nla_get_u32(nla_weight); | |
3283 | udest->u_threshold = nla_get_u32(nla_u_thresh); | |
3284 | udest->l_threshold = nla_get_u32(nla_l_thresh); | |
3285 | } | |
3286 | ||
3287 | return 0; | |
3288 | } | |
3289 | ||
0a925864 | 3290 | static int ip_vs_genl_fill_daemon(struct sk_buff *skb, __u32 state, |
e4ff6751 | 3291 | struct ipvs_sync_daemon_cfg *c) |
9a812198 JV |
3292 | { |
3293 | struct nlattr *nl_daemon; | |
3294 | ||
3295 | nl_daemon = nla_nest_start(skb, IPVS_CMD_ATTR_DAEMON); | |
3296 | if (!nl_daemon) | |
3297 | return -EMSGSIZE; | |
3298 | ||
969e8e25 | 3299 | if (nla_put_u32(skb, IPVS_DAEMON_ATTR_STATE, state) || |
e4ff6751 JA |
3300 | nla_put_string(skb, IPVS_DAEMON_ATTR_MCAST_IFN, c->mcast_ifn) || |
3301 | nla_put_u32(skb, IPVS_DAEMON_ATTR_SYNC_ID, c->syncid) || | |
d3328817 JA |
3302 | nla_put_u16(skb, IPVS_DAEMON_ATTR_SYNC_MAXLEN, c->sync_maxlen) || |
3303 | nla_put_u16(skb, IPVS_DAEMON_ATTR_MCAST_PORT, c->mcast_port) || | |
3304 | nla_put_u8(skb, IPVS_DAEMON_ATTR_MCAST_TTL, c->mcast_ttl)) | |
969e8e25 | 3305 | goto nla_put_failure; |
d3328817 JA |
3306 | #ifdef CONFIG_IP_VS_IPV6 |
3307 | if (c->mcast_af == AF_INET6) { | |
3308 | if (nla_put_in6_addr(skb, IPVS_DAEMON_ATTR_MCAST_GROUP6, | |
3309 | &c->mcast_group.in6)) | |
3310 | goto nla_put_failure; | |
3311 | } else | |
3312 | #endif | |
3313 | if (c->mcast_af == AF_INET && | |
3314 | nla_put_in_addr(skb, IPVS_DAEMON_ATTR_MCAST_GROUP, | |
3315 | c->mcast_group.ip)) | |
3316 | goto nla_put_failure; | |
9a812198 JV |
3317 | nla_nest_end(skb, nl_daemon); |
3318 | ||
3319 | return 0; | |
3320 | ||
3321 | nla_put_failure: | |
3322 | nla_nest_cancel(skb, nl_daemon); | |
3323 | return -EMSGSIZE; | |
3324 | } | |
3325 | ||
0a925864 | 3326 | static int ip_vs_genl_dump_daemon(struct sk_buff *skb, __u32 state, |
e4ff6751 | 3327 | struct ipvs_sync_daemon_cfg *c, |
9a812198 JV |
3328 | struct netlink_callback *cb) |
3329 | { | |
3330 | void *hdr; | |
15e47304 | 3331 | hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq, |
9a812198 JV |
3332 | &ip_vs_genl_family, NLM_F_MULTI, |
3333 | IPVS_CMD_NEW_DAEMON); | |
3334 | if (!hdr) | |
3335 | return -EMSGSIZE; | |
3336 | ||
e4ff6751 | 3337 | if (ip_vs_genl_fill_daemon(skb, state, c)) |
9a812198 JV |
3338 | goto nla_put_failure; |
3339 | ||
053c095a JB |
3340 | genlmsg_end(skb, hdr); |
3341 | return 0; | |
9a812198 JV |
3342 | |
3343 | nla_put_failure: | |
3344 | genlmsg_cancel(skb, hdr); | |
3345 | return -EMSGSIZE; | |
3346 | } | |
3347 | ||
3348 | static int ip_vs_genl_dump_daemons(struct sk_buff *skb, | |
3349 | struct netlink_callback *cb) | |
3350 | { | |
a09d1977 | 3351 | struct net *net = skb_sknet(skb); |
f131315f HS |
3352 | struct netns_ipvs *ipvs = net_ipvs(net); |
3353 | ||
ae1d48b2 | 3354 | mutex_lock(&ipvs->sync_mutex); |
f131315f | 3355 | if ((ipvs->sync_state & IP_VS_STATE_MASTER) && !cb->args[0]) { |
9a812198 | 3356 | if (ip_vs_genl_dump_daemon(skb, IP_VS_STATE_MASTER, |
e4ff6751 | 3357 | &ipvs->mcfg, cb) < 0) |
9a812198 JV |
3358 | goto nla_put_failure; |
3359 | ||
3360 | cb->args[0] = 1; | |
3361 | } | |
3362 | ||
f131315f | 3363 | if ((ipvs->sync_state & IP_VS_STATE_BACKUP) && !cb->args[1]) { |
9a812198 | 3364 | if (ip_vs_genl_dump_daemon(skb, IP_VS_STATE_BACKUP, |
e4ff6751 | 3365 | &ipvs->bcfg, cb) < 0) |
9a812198 JV |
3366 | goto nla_put_failure; |
3367 | ||
3368 | cb->args[1] = 1; | |
3369 | } | |
3370 | ||
3371 | nla_put_failure: | |
ae1d48b2 | 3372 | mutex_unlock(&ipvs->sync_mutex); |
9a812198 JV |
3373 | |
3374 | return skb->len; | |
3375 | } | |
3376 | ||
d8443c5f | 3377 | static int ip_vs_genl_new_daemon(struct netns_ipvs *ipvs, struct nlattr **attrs) |
9a812198 | 3378 | { |
e4ff6751 JA |
3379 | struct ipvs_sync_daemon_cfg c; |
3380 | struct nlattr *a; | |
e0b26cc9 JA |
3381 | int ret; |
3382 | ||
e4ff6751 | 3383 | memset(&c, 0, sizeof(c)); |
9a812198 JV |
3384 | if (!(attrs[IPVS_DAEMON_ATTR_STATE] && |
3385 | attrs[IPVS_DAEMON_ATTR_MCAST_IFN] && | |
3386 | attrs[IPVS_DAEMON_ATTR_SYNC_ID])) | |
3387 | return -EINVAL; | |
e4ff6751 JA |
3388 | strlcpy(c.mcast_ifn, nla_data(attrs[IPVS_DAEMON_ATTR_MCAST_IFN]), |
3389 | sizeof(c.mcast_ifn)); | |
3390 | c.syncid = nla_get_u32(attrs[IPVS_DAEMON_ATTR_SYNC_ID]); | |
3391 | ||
3392 | a = attrs[IPVS_DAEMON_ATTR_SYNC_MAXLEN]; | |
3393 | if (a) | |
3394 | c.sync_maxlen = nla_get_u16(a); | |
9a812198 | 3395 | |
d3328817 JA |
3396 | a = attrs[IPVS_DAEMON_ATTR_MCAST_GROUP]; |
3397 | if (a) { | |
3398 | c.mcast_af = AF_INET; | |
3399 | c.mcast_group.ip = nla_get_in_addr(a); | |
3400 | if (!ipv4_is_multicast(c.mcast_group.ip)) | |
3401 | return -EINVAL; | |
3402 | } else { | |
3403 | a = attrs[IPVS_DAEMON_ATTR_MCAST_GROUP6]; | |
3404 | if (a) { | |
3405 | #ifdef CONFIG_IP_VS_IPV6 | |
3406 | int addr_type; | |
3407 | ||
3408 | c.mcast_af = AF_INET6; | |
3409 | c.mcast_group.in6 = nla_get_in6_addr(a); | |
3410 | addr_type = ipv6_addr_type(&c.mcast_group.in6); | |
3411 | if (!(addr_type & IPV6_ADDR_MULTICAST)) | |
3412 | return -EINVAL; | |
3413 | #else | |
3414 | return -EAFNOSUPPORT; | |
3415 | #endif | |
3416 | } | |
3417 | } | |
3418 | ||
3419 | a = attrs[IPVS_DAEMON_ATTR_MCAST_PORT]; | |
3420 | if (a) | |
3421 | c.mcast_port = nla_get_u16(a); | |
3422 | ||
3423 | a = attrs[IPVS_DAEMON_ATTR_MCAST_TTL]; | |
3424 | if (a) | |
3425 | c.mcast_ttl = nla_get_u8(a); | |
3426 | ||
391f503d AG |
3427 | /* The synchronization protocol is incompatible with mixed family |
3428 | * services | |
3429 | */ | |
e4ff6751 | 3430 | if (ipvs->mixed_address_family_dests > 0) |
391f503d AG |
3431 | return -EINVAL; |
3432 | ||
e0b26cc9 JA |
3433 | rtnl_lock(); |
3434 | mutex_lock(&ipvs->sync_mutex); | |
6ac121d7 | 3435 | ret = start_sync_thread(ipvs, &c, |
e4ff6751 | 3436 | nla_get_u32(attrs[IPVS_DAEMON_ATTR_STATE])); |
e0b26cc9 JA |
3437 | mutex_unlock(&ipvs->sync_mutex); |
3438 | rtnl_unlock(); | |
3439 | return ret; | |
9a812198 JV |
3440 | } |
3441 | ||
df04ffb7 | 3442 | static int ip_vs_genl_del_daemon(struct netns_ipvs *ipvs, struct nlattr **attrs) |
9a812198 | 3443 | { |
e0b26cc9 JA |
3444 | int ret; |
3445 | ||
9a812198 JV |
3446 | if (!attrs[IPVS_DAEMON_ATTR_STATE]) |
3447 | return -EINVAL; | |
3448 | ||
e0b26cc9 | 3449 | mutex_lock(&ipvs->sync_mutex); |
b3cf3cbf | 3450 | ret = stop_sync_thread(ipvs, |
e0b26cc9 JA |
3451 | nla_get_u32(attrs[IPVS_DAEMON_ATTR_STATE])); |
3452 | mutex_unlock(&ipvs->sync_mutex); | |
3453 | return ret; | |
9a812198 JV |
3454 | } |
3455 | ||
cacd1e60 | 3456 | static int ip_vs_genl_set_config(struct netns_ipvs *ipvs, struct nlattr **attrs) |
9a812198 JV |
3457 | { |
3458 | struct ip_vs_timeout_user t; | |
3459 | ||
af540341 | 3460 | __ip_vs_get_timeouts(ipvs, &t); |
9a812198 JV |
3461 | |
3462 | if (attrs[IPVS_CMD_ATTR_TIMEOUT_TCP]) | |
3463 | t.tcp_timeout = nla_get_u32(attrs[IPVS_CMD_ATTR_TIMEOUT_TCP]); | |
3464 | ||
3465 | if (attrs[IPVS_CMD_ATTR_TIMEOUT_TCP_FIN]) | |
3466 | t.tcp_fin_timeout = | |
3467 | nla_get_u32(attrs[IPVS_CMD_ATTR_TIMEOUT_TCP_FIN]); | |
3468 | ||
3469 | if (attrs[IPVS_CMD_ATTR_TIMEOUT_UDP]) | |
3470 | t.udp_timeout = nla_get_u32(attrs[IPVS_CMD_ATTR_TIMEOUT_UDP]); | |
3471 | ||
f1faa1e7 | 3472 | return ip_vs_set_timeout(ipvs, &t); |
9a812198 JV |
3473 | } |
3474 | ||
ae1d48b2 | 3475 | static int ip_vs_genl_set_daemon(struct sk_buff *skb, struct genl_info *info) |
9a812198 | 3476 | { |
e0b26cc9 | 3477 | int ret = -EINVAL, cmd; |
fc723250 | 3478 | struct net *net; |
a0840e2e | 3479 | struct netns_ipvs *ipvs; |
9a812198 | 3480 | |
fc723250 | 3481 | net = skb_sknet(skb); |
a0840e2e | 3482 | ipvs = net_ipvs(net); |
9a812198 JV |
3483 | cmd = info->genlhdr->cmd; |
3484 | ||
ae1d48b2 | 3485 | if (cmd == IPVS_CMD_NEW_DAEMON || cmd == IPVS_CMD_DEL_DAEMON) { |
9a812198 JV |
3486 | struct nlattr *daemon_attrs[IPVS_DAEMON_ATTR_MAX + 1]; |
3487 | ||
3488 | if (!info->attrs[IPVS_CMD_ATTR_DAEMON] || | |
3489 | nla_parse_nested(daemon_attrs, IPVS_DAEMON_ATTR_MAX, | |
3490 | info->attrs[IPVS_CMD_ATTR_DAEMON], | |
e0b26cc9 | 3491 | ip_vs_daemon_policy)) |
9a812198 | 3492 | goto out; |
9a812198 JV |
3493 | |
3494 | if (cmd == IPVS_CMD_NEW_DAEMON) | |
d8443c5f | 3495 | ret = ip_vs_genl_new_daemon(ipvs, daemon_attrs); |
9a812198 | 3496 | else |
df04ffb7 | 3497 | ret = ip_vs_genl_del_daemon(ipvs, daemon_attrs); |
ae1d48b2 | 3498 | } |
e0b26cc9 JA |
3499 | |
3500 | out: | |
ae1d48b2 HS |
3501 | return ret; |
3502 | } | |
3503 | ||
3504 | static int ip_vs_genl_set_cmd(struct sk_buff *skb, struct genl_info *info) | |
3505 | { | |
3506 | struct ip_vs_service *svc = NULL; | |
3507 | struct ip_vs_service_user_kern usvc; | |
3508 | struct ip_vs_dest_user_kern udest; | |
3509 | int ret = 0, cmd; | |
3510 | int need_full_svc = 0, need_full_dest = 0; | |
3511 | struct net *net; | |
cd58278b | 3512 | struct netns_ipvs *ipvs; |
ae1d48b2 HS |
3513 | |
3514 | net = skb_sknet(skb); | |
cd58278b | 3515 | ipvs = net_ipvs(net); |
ae1d48b2 HS |
3516 | cmd = info->genlhdr->cmd; |
3517 | ||
3518 | mutex_lock(&__ip_vs_mutex); | |
3519 | ||
3520 | if (cmd == IPVS_CMD_FLUSH) { | |
ef7c599d | 3521 | ret = ip_vs_flush(ipvs, false); |
ae1d48b2 HS |
3522 | goto out; |
3523 | } else if (cmd == IPVS_CMD_SET_CONFIG) { | |
cacd1e60 | 3524 | ret = ip_vs_genl_set_config(ipvs, info->attrs); |
9a812198 JV |
3525 | goto out; |
3526 | } else if (cmd == IPVS_CMD_ZERO && | |
3527 | !info->attrs[IPVS_CMD_ATTR_SERVICE]) { | |
c60856c6 | 3528 | ret = ip_vs_zero_all(ipvs); |
9a812198 JV |
3529 | goto out; |
3530 | } | |
3531 | ||
3532 | /* All following commands require a service argument, so check if we | |
3533 | * received a valid one. We need a full service specification when | |
3534 | * adding / editing a service. Only identifying members otherwise. */ | |
3535 | if (cmd == IPVS_CMD_NEW_SERVICE || cmd == IPVS_CMD_SET_SERVICE) | |
3536 | need_full_svc = 1; | |
3537 | ||
613fb830 | 3538 | ret = ip_vs_genl_parse_service(ipvs, &usvc, |
9a812198 | 3539 | info->attrs[IPVS_CMD_ATTR_SERVICE], |
26c15cfd | 3540 | need_full_svc, &svc); |
9a812198 JV |
3541 | if (ret) |
3542 | goto out; | |
3543 | ||
9a812198 JV |
3544 | /* Unless we're adding a new service, the service must already exist */ |
3545 | if ((cmd != IPVS_CMD_NEW_SERVICE) && (svc == NULL)) { | |
3546 | ret = -ESRCH; | |
3547 | goto out; | |
3548 | } | |
3549 | ||
3550 | /* Destination commands require a valid destination argument. For | |
3551 | * adding / editing a destination, we need a full destination | |
3552 | * specification. */ | |
3553 | if (cmd == IPVS_CMD_NEW_DEST || cmd == IPVS_CMD_SET_DEST || | |
3554 | cmd == IPVS_CMD_DEL_DEST) { | |
3555 | if (cmd != IPVS_CMD_DEL_DEST) | |
3556 | need_full_dest = 1; | |
3557 | ||
3558 | ret = ip_vs_genl_parse_dest(&udest, | |
3559 | info->attrs[IPVS_CMD_ATTR_DEST], | |
3560 | need_full_dest); | |
3561 | if (ret) | |
3562 | goto out; | |
6cff339b AG |
3563 | |
3564 | /* Old protocols did not allow the user to specify address | |
3565 | * family, so we set it to zero instead. We also didn't | |
3566 | * allow heterogeneous pools in the old code, so it's safe | |
3567 | * to assume that this will have the same address family as | |
3568 | * the service. | |
3569 | */ | |
3570 | if (udest.af == 0) | |
3571 | udest.af = svc->af; | |
bc18d37f | 3572 | |
dd3733b3 | 3573 | if (udest.af != svc->af && cmd != IPVS_CMD_DEL_DEST) { |
bc18d37f AG |
3574 | /* The synchronization protocol is incompatible |
3575 | * with mixed family services | |
3576 | */ | |
cd58278b | 3577 | if (ipvs->sync_state) { |
bc18d37f AG |
3578 | ret = -EINVAL; |
3579 | goto out; | |
3580 | } | |
3581 | ||
3582 | /* Which connection types do we support? */ | |
3583 | switch (udest.conn_flags) { | |
3584 | case IP_VS_CONN_F_TUNNEL: | |
3585 | /* We are able to forward this */ | |
3586 | break; | |
3587 | default: | |
3588 | ret = -EINVAL; | |
3589 | goto out; | |
3590 | } | |
3591 | } | |
9a812198 JV |
3592 | } |
3593 | ||
3594 | switch (cmd) { | |
3595 | case IPVS_CMD_NEW_SERVICE: | |
3596 | if (svc == NULL) | |
5060bd83 | 3597 | ret = ip_vs_add_service(ipvs, &usvc, &svc); |
9a812198 JV |
3598 | else |
3599 | ret = -EEXIST; | |
3600 | break; | |
3601 | case IPVS_CMD_SET_SERVICE: | |
3602 | ret = ip_vs_edit_service(svc, &usvc); | |
3603 | break; | |
3604 | case IPVS_CMD_DEL_SERVICE: | |
3605 | ret = ip_vs_del_service(svc); | |
26c15cfd | 3606 | /* do not use svc, it can be freed */ |
9a812198 JV |
3607 | break; |
3608 | case IPVS_CMD_NEW_DEST: | |
3609 | ret = ip_vs_add_dest(svc, &udest); | |
3610 | break; | |
3611 | case IPVS_CMD_SET_DEST: | |
3612 | ret = ip_vs_edit_dest(svc, &udest); | |
3613 | break; | |
3614 | case IPVS_CMD_DEL_DEST: | |
3615 | ret = ip_vs_del_dest(svc, &udest); | |
3616 | break; | |
3617 | case IPVS_CMD_ZERO: | |
3618 | ret = ip_vs_zero_service(svc); | |
3619 | break; | |
3620 | default: | |
3621 | ret = -EINVAL; | |
3622 | } | |
3623 | ||
3624 | out: | |
9a812198 JV |
3625 | mutex_unlock(&__ip_vs_mutex); |
3626 | ||
3627 | return ret; | |
3628 | } | |
3629 | ||
3630 | static int ip_vs_genl_get_cmd(struct sk_buff *skb, struct genl_info *info) | |
3631 | { | |
3632 | struct sk_buff *msg; | |
3633 | void *reply; | |
3634 | int ret, cmd, reply_cmd; | |
fc723250 | 3635 | struct net *net; |
af540341 | 3636 | struct netns_ipvs *ipvs; |
9a812198 | 3637 | |
fc723250 | 3638 | net = skb_sknet(skb); |
af540341 | 3639 | ipvs = net_ipvs(net); |
9a812198 JV |
3640 | cmd = info->genlhdr->cmd; |
3641 | ||
3642 | if (cmd == IPVS_CMD_GET_SERVICE) | |
3643 | reply_cmd = IPVS_CMD_NEW_SERVICE; | |
3644 | else if (cmd == IPVS_CMD_GET_INFO) | |
3645 | reply_cmd = IPVS_CMD_SET_INFO; | |
3646 | else if (cmd == IPVS_CMD_GET_CONFIG) | |
3647 | reply_cmd = IPVS_CMD_SET_CONFIG; | |
3648 | else { | |
1e3e238e | 3649 | pr_err("unknown Generic Netlink command\n"); |
9a812198 JV |
3650 | return -EINVAL; |
3651 | } | |
3652 | ||
3653 | msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL); | |
3654 | if (!msg) | |
3655 | return -ENOMEM; | |
3656 | ||
3657 | mutex_lock(&__ip_vs_mutex); | |
3658 | ||
3659 | reply = genlmsg_put_reply(msg, info, &ip_vs_genl_family, 0, reply_cmd); | |
3660 | if (reply == NULL) | |
3661 | goto nla_put_failure; | |
3662 | ||
3663 | switch (cmd) { | |
3664 | case IPVS_CMD_GET_SERVICE: | |
3665 | { | |
3666 | struct ip_vs_service *svc; | |
3667 | ||
34c2f514 | 3668 | svc = ip_vs_genl_find_service(ipvs, |
fc723250 | 3669 | info->attrs[IPVS_CMD_ATTR_SERVICE]); |
9a812198 JV |
3670 | if (IS_ERR(svc)) { |
3671 | ret = PTR_ERR(svc); | |
3672 | goto out_err; | |
3673 | } else if (svc) { | |
3674 | ret = ip_vs_genl_fill_service(msg, svc); | |
9a812198 JV |
3675 | if (ret) |
3676 | goto nla_put_failure; | |
3677 | } else { | |
3678 | ret = -ESRCH; | |
3679 | goto out_err; | |
3680 | } | |
3681 | ||
3682 | break; | |
3683 | } | |
3684 | ||
3685 | case IPVS_CMD_GET_CONFIG: | |
3686 | { | |
3687 | struct ip_vs_timeout_user t; | |
3688 | ||
af540341 | 3689 | __ip_vs_get_timeouts(ipvs, &t); |
9a812198 | 3690 | #ifdef CONFIG_IP_VS_PROTO_TCP |
969e8e25 DM |
3691 | if (nla_put_u32(msg, IPVS_CMD_ATTR_TIMEOUT_TCP, |
3692 | t.tcp_timeout) || | |
3693 | nla_put_u32(msg, IPVS_CMD_ATTR_TIMEOUT_TCP_FIN, | |
3694 | t.tcp_fin_timeout)) | |
3695 | goto nla_put_failure; | |
9a812198 JV |
3696 | #endif |
3697 | #ifdef CONFIG_IP_VS_PROTO_UDP | |
969e8e25 DM |
3698 | if (nla_put_u32(msg, IPVS_CMD_ATTR_TIMEOUT_UDP, t.udp_timeout)) |
3699 | goto nla_put_failure; | |
9a812198 JV |
3700 | #endif |
3701 | ||
3702 | break; | |
3703 | } | |
3704 | ||
3705 | case IPVS_CMD_GET_INFO: | |
969e8e25 DM |
3706 | if (nla_put_u32(msg, IPVS_INFO_ATTR_VERSION, |
3707 | IP_VS_VERSION_CODE) || | |
3708 | nla_put_u32(msg, IPVS_INFO_ATTR_CONN_TAB_SIZE, | |
3709 | ip_vs_conn_tab_size)) | |
3710 | goto nla_put_failure; | |
9a812198 JV |
3711 | break; |
3712 | } | |
3713 | ||
3714 | genlmsg_end(msg, reply); | |
134e6375 | 3715 | ret = genlmsg_reply(msg, info); |
9a812198 JV |
3716 | goto out; |
3717 | ||
3718 | nla_put_failure: | |
1e3e238e | 3719 | pr_err("not enough space in Netlink message\n"); |
9a812198 JV |
3720 | ret = -EMSGSIZE; |
3721 | ||
3722 | out_err: | |
3723 | nlmsg_free(msg); | |
3724 | out: | |
3725 | mutex_unlock(&__ip_vs_mutex); | |
3726 | ||
3727 | return ret; | |
3728 | } | |
3729 | ||
3730 | ||
c61b0c13 | 3731 | static const struct genl_ops ip_vs_genl_ops[] = { |
9a812198 JV |
3732 | { |
3733 | .cmd = IPVS_CMD_NEW_SERVICE, | |
3734 | .flags = GENL_ADMIN_PERM, | |
3735 | .policy = ip_vs_cmd_policy, | |
3736 | .doit = ip_vs_genl_set_cmd, | |
3737 | }, | |
3738 | { | |
3739 | .cmd = IPVS_CMD_SET_SERVICE, | |
3740 | .flags = GENL_ADMIN_PERM, | |
3741 | .policy = ip_vs_cmd_policy, | |
3742 | .doit = ip_vs_genl_set_cmd, | |
3743 | }, | |
3744 | { | |
3745 | .cmd = IPVS_CMD_DEL_SERVICE, | |
3746 | .flags = GENL_ADMIN_PERM, | |
3747 | .policy = ip_vs_cmd_policy, | |
3748 | .doit = ip_vs_genl_set_cmd, | |
3749 | }, | |
3750 | { | |
3751 | .cmd = IPVS_CMD_GET_SERVICE, | |
3752 | .flags = GENL_ADMIN_PERM, | |
3753 | .doit = ip_vs_genl_get_cmd, | |
3754 | .dumpit = ip_vs_genl_dump_services, | |
3755 | .policy = ip_vs_cmd_policy, | |
3756 | }, | |
3757 | { | |
3758 | .cmd = IPVS_CMD_NEW_DEST, | |
3759 | .flags = GENL_ADMIN_PERM, | |
3760 | .policy = ip_vs_cmd_policy, | |
3761 | .doit = ip_vs_genl_set_cmd, | |
3762 | }, | |
3763 | { | |
3764 | .cmd = IPVS_CMD_SET_DEST, | |
3765 | .flags = GENL_ADMIN_PERM, | |
3766 | .policy = ip_vs_cmd_policy, | |
3767 | .doit = ip_vs_genl_set_cmd, | |
3768 | }, | |
3769 | { | |
3770 | .cmd = IPVS_CMD_DEL_DEST, | |
3771 | .flags = GENL_ADMIN_PERM, | |
3772 | .policy = ip_vs_cmd_policy, | |
3773 | .doit = ip_vs_genl_set_cmd, | |
3774 | }, | |
3775 | { | |
3776 | .cmd = IPVS_CMD_GET_DEST, | |
3777 | .flags = GENL_ADMIN_PERM, | |
3778 | .policy = ip_vs_cmd_policy, | |
3779 | .dumpit = ip_vs_genl_dump_dests, | |
3780 | }, | |
3781 | { | |
3782 | .cmd = IPVS_CMD_NEW_DAEMON, | |
3783 | .flags = GENL_ADMIN_PERM, | |
3784 | .policy = ip_vs_cmd_policy, | |
ae1d48b2 | 3785 | .doit = ip_vs_genl_set_daemon, |
9a812198 JV |
3786 | }, |
3787 | { | |
3788 | .cmd = IPVS_CMD_DEL_DAEMON, | |
3789 | .flags = GENL_ADMIN_PERM, | |
3790 | .policy = ip_vs_cmd_policy, | |
ae1d48b2 | 3791 | .doit = ip_vs_genl_set_daemon, |
9a812198 JV |
3792 | }, |
3793 | { | |
3794 | .cmd = IPVS_CMD_GET_DAEMON, | |
3795 | .flags = GENL_ADMIN_PERM, | |
3796 | .dumpit = ip_vs_genl_dump_daemons, | |
3797 | }, | |
3798 | { | |
3799 | .cmd = IPVS_CMD_SET_CONFIG, | |
3800 | .flags = GENL_ADMIN_PERM, | |
3801 | .policy = ip_vs_cmd_policy, | |
3802 | .doit = ip_vs_genl_set_cmd, | |
3803 | }, | |
3804 | { | |
3805 | .cmd = IPVS_CMD_GET_CONFIG, | |
3806 | .flags = GENL_ADMIN_PERM, | |
3807 | .doit = ip_vs_genl_get_cmd, | |
3808 | }, | |
3809 | { | |
3810 | .cmd = IPVS_CMD_GET_INFO, | |
3811 | .flags = GENL_ADMIN_PERM, | |
3812 | .doit = ip_vs_genl_get_cmd, | |
3813 | }, | |
3814 | { | |
3815 | .cmd = IPVS_CMD_ZERO, | |
3816 | .flags = GENL_ADMIN_PERM, | |
3817 | .policy = ip_vs_cmd_policy, | |
3818 | .doit = ip_vs_genl_set_cmd, | |
3819 | }, | |
3820 | { | |
3821 | .cmd = IPVS_CMD_FLUSH, | |
3822 | .flags = GENL_ADMIN_PERM, | |
3823 | .doit = ip_vs_genl_set_cmd, | |
3824 | }, | |
3825 | }; | |
3826 | ||
3827 | static int __init ip_vs_genl_register(void) | |
3828 | { | |
8f698d54 | 3829 | return genl_register_family_with_ops(&ip_vs_genl_family, |
c53ed742 | 3830 | ip_vs_genl_ops); |
9a812198 JV |
3831 | } |
3832 | ||
3833 | static void ip_vs_genl_unregister(void) | |
3834 | { | |
3835 | genl_unregister_family(&ip_vs_genl_family); | |
3836 | } | |
3837 | ||
3838 | /* End of Generic Netlink interface definitions */ | |
3839 | ||
61b1ab45 HS |
3840 | /* |
3841 | * per netns intit/exit func. | |
3842 | */ | |
14e40546 | 3843 | #ifdef CONFIG_SYSCTL |
8b8237a5 | 3844 | static int __net_init ip_vs_control_net_init_sysctl(struct netns_ipvs *ipvs) |
61b1ab45 | 3845 | { |
8b8237a5 | 3846 | struct net *net = ipvs->net; |
fc723250 | 3847 | int idx; |
a0840e2e | 3848 | struct ctl_table *tbl; |
fc723250 | 3849 | |
a0840e2e HS |
3850 | atomic_set(&ipvs->dropentry, 0); |
3851 | spin_lock_init(&ipvs->dropentry_lock); | |
3852 | spin_lock_init(&ipvs->droppacket_lock); | |
3853 | spin_lock_init(&ipvs->securetcp_lock); | |
a0840e2e HS |
3854 | |
3855 | if (!net_eq(net, &init_net)) { | |
3856 | tbl = kmemdup(vs_vars, sizeof(vs_vars), GFP_KERNEL); | |
3857 | if (tbl == NULL) | |
14e40546 | 3858 | return -ENOMEM; |
464dc801 EB |
3859 | |
3860 | /* Don't export sysctls to unprivileged users */ | |
3861 | if (net->user_ns != &init_user_ns) | |
3862 | tbl[0].procname = NULL; | |
a0840e2e HS |
3863 | } else |
3864 | tbl = vs_vars; | |
3865 | /* Initialize sysctl defaults */ | |
717e917d EB |
3866 | for (idx = 0; idx < ARRAY_SIZE(vs_vars); idx++) { |
3867 | if (tbl[idx].proc_handler == proc_do_defense_mode) | |
3868 | tbl[idx].extra2 = ipvs; | |
3869 | } | |
a0840e2e HS |
3870 | idx = 0; |
3871 | ipvs->sysctl_amemthresh = 1024; | |
3872 | tbl[idx++].data = &ipvs->sysctl_amemthresh; | |
3873 | ipvs->sysctl_am_droprate = 10; | |
3874 | tbl[idx++].data = &ipvs->sysctl_am_droprate; | |
3875 | tbl[idx++].data = &ipvs->sysctl_drop_entry; | |
3876 | tbl[idx++].data = &ipvs->sysctl_drop_packet; | |
3877 | #ifdef CONFIG_IP_VS_NFCT | |
3878 | tbl[idx++].data = &ipvs->sysctl_conntrack; | |
3879 | #endif | |
3880 | tbl[idx++].data = &ipvs->sysctl_secure_tcp; | |
3881 | ipvs->sysctl_snat_reroute = 1; | |
3882 | tbl[idx++].data = &ipvs->sysctl_snat_reroute; | |
3883 | ipvs->sysctl_sync_ver = 1; | |
3884 | tbl[idx++].data = &ipvs->sysctl_sync_ver; | |
f73181c8 PNA |
3885 | ipvs->sysctl_sync_ports = 1; |
3886 | tbl[idx++].data = &ipvs->sysctl_sync_ports; | |
4d0c875d | 3887 | tbl[idx++].data = &ipvs->sysctl_sync_persist_mode; |
1c003b15 PNA |
3888 | ipvs->sysctl_sync_qlen_max = nr_free_buffer_pages() / 32; |
3889 | tbl[idx++].data = &ipvs->sysctl_sync_qlen_max; | |
3890 | ipvs->sysctl_sync_sock_size = 0; | |
3891 | tbl[idx++].data = &ipvs->sysctl_sync_sock_size; | |
a0840e2e HS |
3892 | tbl[idx++].data = &ipvs->sysctl_cache_bypass; |
3893 | tbl[idx++].data = &ipvs->sysctl_expire_nodest_conn; | |
c6c96c18 AF |
3894 | tbl[idx++].data = &ipvs->sysctl_sloppy_tcp; |
3895 | tbl[idx++].data = &ipvs->sysctl_sloppy_sctp; | |
a0840e2e | 3896 | tbl[idx++].data = &ipvs->sysctl_expire_quiescent_template; |
59e0350e SH |
3897 | ipvs->sysctl_sync_threshold[0] = DEFAULT_SYNC_THRESHOLD; |
3898 | ipvs->sysctl_sync_threshold[1] = DEFAULT_SYNC_PERIOD; | |
a0840e2e HS |
3899 | tbl[idx].data = &ipvs->sysctl_sync_threshold; |
3900 | tbl[idx++].maxlen = sizeof(ipvs->sysctl_sync_threshold); | |
749c42b6 JA |
3901 | ipvs->sysctl_sync_refresh_period = DEFAULT_SYNC_REFRESH_PERIOD; |
3902 | tbl[idx++].data = &ipvs->sysctl_sync_refresh_period; | |
3903 | ipvs->sysctl_sync_retries = clamp_t(int, DEFAULT_SYNC_RETRIES, 0, 3); | |
3904 | tbl[idx++].data = &ipvs->sysctl_sync_retries; | |
a0840e2e | 3905 | tbl[idx++].data = &ipvs->sysctl_nat_icmp_send; |
3654e611 JA |
3906 | ipvs->sysctl_pmtu_disc = 1; |
3907 | tbl[idx++].data = &ipvs->sysctl_pmtu_disc; | |
0c12582f | 3908 | tbl[idx++].data = &ipvs->sysctl_backup_only; |
d752c364 MRL |
3909 | ipvs->sysctl_conn_reuse_mode = 1; |
3910 | tbl[idx++].data = &ipvs->sysctl_conn_reuse_mode; | |
94485fed | 3911 | tbl[idx++].data = &ipvs->sysctl_schedule_icmp; |
4e478098 | 3912 | tbl[idx++].data = &ipvs->sysctl_ignore_tunneled; |
a0840e2e | 3913 | |
ec8f23ce | 3914 | ipvs->sysctl_hdr = register_net_sysctl(net, "net/ipv4/vs", tbl); |
0443929f SH |
3915 | if (ipvs->sysctl_hdr == NULL) { |
3916 | if (!net_eq(net, &init_net)) | |
3917 | kfree(tbl); | |
14e40546 | 3918 | return -ENOMEM; |
0443929f | 3919 | } |
0f34d54b | 3920 | ip_vs_start_estimator(ipvs, &ipvs->tot_stats); |
a0840e2e | 3921 | ipvs->sysctl_tbl = tbl; |
f6340ee0 HS |
3922 | /* Schedule defense work */ |
3923 | INIT_DELAYED_WORK(&ipvs->defense_work, defense_work_handler); | |
3924 | schedule_delayed_work(&ipvs->defense_work, DEFENSE_TIMER_PERIOD); | |
61b1ab45 | 3925 | |
61b1ab45 | 3926 | return 0; |
61b1ab45 HS |
3927 | } |
3928 | ||
8b8237a5 | 3929 | static void __net_exit ip_vs_control_net_cleanup_sysctl(struct netns_ipvs *ipvs) |
61b1ab45 | 3930 | { |
8b8237a5 | 3931 | struct net *net = ipvs->net; |
b17fc996 | 3932 | |
f2431e6e HS |
3933 | cancel_delayed_work_sync(&ipvs->defense_work); |
3934 | cancel_work_sync(&ipvs->defense_work.work); | |
a0840e2e | 3935 | unregister_net_sysctl_table(ipvs->sysctl_hdr); |
0f34d54b | 3936 | ip_vs_stop_estimator(ipvs, &ipvs->tot_stats); |
f30bf2a5 TR |
3937 | |
3938 | if (!net_eq(net, &init_net)) | |
3939 | kfree(ipvs->sysctl_tbl); | |
14e40546 SH |
3940 | } |
3941 | ||
3942 | #else | |
3943 | ||
8b8237a5 EB |
3944 | static int __net_init ip_vs_control_net_init_sysctl(struct netns_ipvs *ipvs) { return 0; } |
3945 | static void __net_exit ip_vs_control_net_cleanup_sysctl(struct netns_ipvs *ipvs) { } | |
14e40546 | 3946 | |
0443929f | 3947 | #endif |
14e40546 | 3948 | |
7a4f0761 HS |
3949 | static struct notifier_block ip_vs_dst_notifier = { |
3950 | .notifier_call = ip_vs_dst_event, | |
3951 | }; | |
3952 | ||
503cf15a | 3953 | int __net_init ip_vs_control_net_init(struct net *net) |
14e40546 | 3954 | { |
827da44c | 3955 | int i, idx; |
14e40546 SH |
3956 | struct netns_ipvs *ipvs = net_ipvs(net); |
3957 | ||
14e40546 SH |
3958 | /* Initialize rs_table */ |
3959 | for (idx = 0; idx < IP_VS_RTAB_SIZE; idx++) | |
276472ea | 3960 | INIT_HLIST_HEAD(&ipvs->rs_table[idx]); |
14e40546 SH |
3961 | |
3962 | INIT_LIST_HEAD(&ipvs->dest_trash); | |
578bc3ef JA |
3963 | spin_lock_init(&ipvs->dest_trash_lock); |
3964 | setup_timer(&ipvs->dest_trash_timer, ip_vs_dest_trash_expire, | |
8e743f1b | 3965 | (unsigned long) ipvs); |
14e40546 SH |
3966 | atomic_set(&ipvs->ftpsvc_counter, 0); |
3967 | atomic_set(&ipvs->nullsvc_counter, 0); | |
3968 | ||
3969 | /* procfs stats */ | |
3970 | ipvs->tot_stats.cpustats = alloc_percpu(struct ip_vs_cpu_stats); | |
0a9ee813 | 3971 | if (!ipvs->tot_stats.cpustats) |
14e40546 | 3972 | return -ENOMEM; |
0a9ee813 | 3973 | |
827da44c JS |
3974 | for_each_possible_cpu(i) { |
3975 | struct ip_vs_cpu_stats *ipvs_tot_stats; | |
3976 | ipvs_tot_stats = per_cpu_ptr(ipvs->tot_stats.cpustats, i); | |
3977 | u64_stats_init(&ipvs_tot_stats->syncp); | |
3978 | } | |
3979 | ||
14e40546 SH |
3980 | spin_lock_init(&ipvs->tot_stats.lock); |
3981 | ||
d4beaa66 G |
3982 | proc_create("ip_vs", 0, net->proc_net, &ip_vs_info_fops); |
3983 | proc_create("ip_vs_stats", 0, net->proc_net, &ip_vs_stats_fops); | |
3984 | proc_create("ip_vs_stats_percpu", 0, net->proc_net, | |
3985 | &ip_vs_stats_percpu_fops); | |
14e40546 | 3986 | |
8b8237a5 | 3987 | if (ip_vs_control_net_init_sysctl(ipvs)) |
14e40546 SH |
3988 | goto err; |
3989 | ||
3990 | return 0; | |
3991 | ||
3992 | err: | |
2a0751af | 3993 | free_percpu(ipvs->tot_stats.cpustats); |
61b1ab45 HS |
3994 | return -ENOMEM; |
3995 | } | |
3996 | ||
503cf15a | 3997 | void __net_exit ip_vs_control_net_cleanup(struct net *net) |
61b1ab45 | 3998 | { |
b17fc996 HS |
3999 | struct netns_ipvs *ipvs = net_ipvs(net); |
4000 | ||
6c0e14f5 | 4001 | ip_vs_trash_cleanup(ipvs); |
8b8237a5 | 4002 | ip_vs_control_net_cleanup_sysctl(ipvs); |
ece31ffd G |
4003 | remove_proc_entry("ip_vs_stats_percpu", net->proc_net); |
4004 | remove_proc_entry("ip_vs_stats", net->proc_net); | |
4005 | remove_proc_entry("ip_vs", net->proc_net); | |
2a0751af | 4006 | free_percpu(ipvs->tot_stats.cpustats); |
61b1ab45 HS |
4007 | } |
4008 | ||
8537de8a | 4009 | int __init ip_vs_register_nl_ioctl(void) |
1da177e4 | 4010 | { |
fc723250 | 4011 | int ret; |
1da177e4 | 4012 | |
1da177e4 LT |
4013 | ret = nf_register_sockopt(&ip_vs_sockopts); |
4014 | if (ret) { | |
1e3e238e | 4015 | pr_err("cannot register sockopt.\n"); |
7a4f0761 | 4016 | goto err_sock; |
1da177e4 LT |
4017 | } |
4018 | ||
9a812198 JV |
4019 | ret = ip_vs_genl_register(); |
4020 | if (ret) { | |
1e3e238e | 4021 | pr_err("cannot register Generic Netlink interface.\n"); |
7a4f0761 | 4022 | goto err_genl; |
9a812198 | 4023 | } |
1da177e4 | 4024 | return 0; |
fc723250 | 4025 | |
7a4f0761 HS |
4026 | err_genl: |
4027 | nf_unregister_sockopt(&ip_vs_sockopts); | |
4028 | err_sock: | |
fc723250 | 4029 | return ret; |
1da177e4 LT |
4030 | } |
4031 | ||
8537de8a HS |
4032 | void ip_vs_unregister_nl_ioctl(void) |
4033 | { | |
4034 | ip_vs_genl_unregister(); | |
4035 | nf_unregister_sockopt(&ip_vs_sockopts); | |
4036 | } | |
4037 | ||
4038 | int __init ip_vs_control_init(void) | |
4039 | { | |
4040 | int idx; | |
4041 | int ret; | |
4042 | ||
4043 | EnterFunction(2); | |
4044 | ||
276472ea | 4045 | /* Initialize svc_table, ip_vs_svc_fwm_table */ |
8537de8a | 4046 | for (idx = 0; idx < IP_VS_SVC_TAB_SIZE; idx++) { |
ceec4c38 JA |
4047 | INIT_HLIST_HEAD(&ip_vs_svc_table[idx]); |
4048 | INIT_HLIST_HEAD(&ip_vs_svc_fwm_table[idx]); | |
8537de8a HS |
4049 | } |
4050 | ||
4051 | smp_wmb(); /* Do we really need it now ? */ | |
4052 | ||
4053 | ret = register_netdevice_notifier(&ip_vs_dst_notifier); | |
4054 | if (ret < 0) | |
4055 | return ret; | |
4056 | ||
4057 | LeaveFunction(2); | |
4058 | return 0; | |
4059 | } | |
4060 | ||
1da177e4 LT |
4061 | |
4062 | void ip_vs_control_cleanup(void) | |
4063 | { | |
4064 | EnterFunction(2); | |
7676e345 | 4065 | unregister_netdevice_notifier(&ip_vs_dst_notifier); |
1da177e4 LT |
4066 | LeaveFunction(2); |
4067 | } |