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 | * | |
8 | * Version: $Id: ip_vs_conn.c,v 1.31 2003/04/18 09:03:16 wensong Exp $ | |
9 | * | |
10 | * Authors: Wensong Zhang <wensong@linuxvirtualserver.org> | |
11 | * Peter Kese <peter.kese@ijs.si> | |
12 | * Julian Anastasov <ja@ssi.bg> | |
13 | * | |
14 | * This program is free software; you can redistribute it and/or | |
15 | * modify it under the terms of the GNU General Public License | |
16 | * as published by the Free Software Foundation; either version | |
17 | * 2 of the License, or (at your option) any later version. | |
18 | * | |
19 | * The IPVS code for kernel 2.2 was done by Wensong Zhang and Peter Kese, | |
20 | * with changes/fixes from Julian Anastasov, Lars Marowsky-Bree, Horms | |
21 | * and others. Many code here is taken from IP MASQ code of kernel 2.2. | |
22 | * | |
23 | * Changes: | |
24 | * | |
25 | */ | |
26 | ||
27 | #include <linux/kernel.h> | |
28 | #include <linux/vmalloc.h> | |
29 | #include <linux/proc_fs.h> /* for proc_net_* */ | |
30 | #include <linux/seq_file.h> | |
31 | #include <linux/jhash.h> | |
32 | #include <linux/random.h> | |
33 | ||
34 | #include <net/ip_vs.h> | |
35 | ||
36 | ||
37 | /* | |
38 | * Connection hash table: for input and output packets lookups of IPVS | |
39 | */ | |
40 | static struct list_head *ip_vs_conn_tab; | |
41 | ||
42 | /* SLAB cache for IPVS connections */ | |
ba89966c | 43 | static kmem_cache_t *ip_vs_conn_cachep __read_mostly; |
1da177e4 LT |
44 | |
45 | /* counter for current IPVS connections */ | |
46 | static atomic_t ip_vs_conn_count = ATOMIC_INIT(0); | |
47 | ||
48 | /* counter for no client port connections */ | |
49 | static atomic_t ip_vs_conn_no_cport_cnt = ATOMIC_INIT(0); | |
50 | ||
51 | /* random value for IPVS connection hash */ | |
52 | static unsigned int ip_vs_conn_rnd; | |
53 | ||
54 | /* | |
55 | * Fine locking granularity for big connection hash table | |
56 | */ | |
57 | #define CT_LOCKARRAY_BITS 4 | |
58 | #define CT_LOCKARRAY_SIZE (1<<CT_LOCKARRAY_BITS) | |
59 | #define CT_LOCKARRAY_MASK (CT_LOCKARRAY_SIZE-1) | |
60 | ||
61 | struct ip_vs_aligned_lock | |
62 | { | |
63 | rwlock_t l; | |
64 | } __attribute__((__aligned__(SMP_CACHE_BYTES))); | |
65 | ||
66 | /* lock array for conn table */ | |
67 | static struct ip_vs_aligned_lock | |
68 | __ip_vs_conntbl_lock_array[CT_LOCKARRAY_SIZE] __cacheline_aligned; | |
69 | ||
70 | static inline void ct_read_lock(unsigned key) | |
71 | { | |
72 | read_lock(&__ip_vs_conntbl_lock_array[key&CT_LOCKARRAY_MASK].l); | |
73 | } | |
74 | ||
75 | static inline void ct_read_unlock(unsigned key) | |
76 | { | |
77 | read_unlock(&__ip_vs_conntbl_lock_array[key&CT_LOCKARRAY_MASK].l); | |
78 | } | |
79 | ||
80 | static inline void ct_write_lock(unsigned key) | |
81 | { | |
82 | write_lock(&__ip_vs_conntbl_lock_array[key&CT_LOCKARRAY_MASK].l); | |
83 | } | |
84 | ||
85 | static inline void ct_write_unlock(unsigned key) | |
86 | { | |
87 | write_unlock(&__ip_vs_conntbl_lock_array[key&CT_LOCKARRAY_MASK].l); | |
88 | } | |
89 | ||
90 | static inline void ct_read_lock_bh(unsigned key) | |
91 | { | |
92 | read_lock_bh(&__ip_vs_conntbl_lock_array[key&CT_LOCKARRAY_MASK].l); | |
93 | } | |
94 | ||
95 | static inline void ct_read_unlock_bh(unsigned key) | |
96 | { | |
97 | read_unlock_bh(&__ip_vs_conntbl_lock_array[key&CT_LOCKARRAY_MASK].l); | |
98 | } | |
99 | ||
100 | static inline void ct_write_lock_bh(unsigned key) | |
101 | { | |
102 | write_lock_bh(&__ip_vs_conntbl_lock_array[key&CT_LOCKARRAY_MASK].l); | |
103 | } | |
104 | ||
105 | static inline void ct_write_unlock_bh(unsigned key) | |
106 | { | |
107 | write_unlock_bh(&__ip_vs_conntbl_lock_array[key&CT_LOCKARRAY_MASK].l); | |
108 | } | |
109 | ||
110 | ||
111 | /* | |
112 | * Returns hash value for IPVS connection entry | |
113 | */ | |
114 | static unsigned int ip_vs_conn_hashkey(unsigned proto, __u32 addr, __u16 port) | |
115 | { | |
116 | return jhash_3words(addr, port, proto, ip_vs_conn_rnd) | |
117 | & IP_VS_CONN_TAB_MASK; | |
118 | } | |
119 | ||
120 | ||
121 | /* | |
122 | * Hashes ip_vs_conn in ip_vs_conn_tab by proto,addr,port. | |
123 | * returns bool success. | |
124 | */ | |
125 | static inline int ip_vs_conn_hash(struct ip_vs_conn *cp) | |
126 | { | |
127 | unsigned hash; | |
128 | int ret; | |
129 | ||
130 | /* Hash by protocol, client address and port */ | |
131 | hash = ip_vs_conn_hashkey(cp->protocol, cp->caddr, cp->cport); | |
132 | ||
133 | ct_write_lock(hash); | |
134 | ||
135 | if (!(cp->flags & IP_VS_CONN_F_HASHED)) { | |
136 | list_add(&cp->c_list, &ip_vs_conn_tab[hash]); | |
137 | cp->flags |= IP_VS_CONN_F_HASHED; | |
138 | atomic_inc(&cp->refcnt); | |
139 | ret = 1; | |
140 | } else { | |
141 | IP_VS_ERR("ip_vs_conn_hash(): request for already hashed, " | |
142 | "called from %p\n", __builtin_return_address(0)); | |
143 | ret = 0; | |
144 | } | |
145 | ||
146 | ct_write_unlock(hash); | |
147 | ||
148 | return ret; | |
149 | } | |
150 | ||
151 | ||
152 | /* | |
153 | * UNhashes ip_vs_conn from ip_vs_conn_tab. | |
154 | * returns bool success. | |
155 | */ | |
156 | static inline int ip_vs_conn_unhash(struct ip_vs_conn *cp) | |
157 | { | |
158 | unsigned hash; | |
159 | int ret; | |
160 | ||
161 | /* unhash it and decrease its reference counter */ | |
162 | hash = ip_vs_conn_hashkey(cp->protocol, cp->caddr, cp->cport); | |
163 | ||
164 | ct_write_lock(hash); | |
165 | ||
166 | if (cp->flags & IP_VS_CONN_F_HASHED) { | |
167 | list_del(&cp->c_list); | |
168 | cp->flags &= ~IP_VS_CONN_F_HASHED; | |
169 | atomic_dec(&cp->refcnt); | |
170 | ret = 1; | |
171 | } else | |
172 | ret = 0; | |
173 | ||
174 | ct_write_unlock(hash); | |
175 | ||
176 | return ret; | |
177 | } | |
178 | ||
179 | ||
180 | /* | |
181 | * Gets ip_vs_conn associated with supplied parameters in the ip_vs_conn_tab. | |
182 | * Called for pkts coming from OUTside-to-INside. | |
183 | * s_addr, s_port: pkt source address (foreign host) | |
184 | * d_addr, d_port: pkt dest address (load balancer) | |
185 | */ | |
186 | static inline struct ip_vs_conn *__ip_vs_conn_in_get | |
187 | (int protocol, __u32 s_addr, __u16 s_port, __u32 d_addr, __u16 d_port) | |
188 | { | |
189 | unsigned hash; | |
190 | struct ip_vs_conn *cp; | |
191 | ||
192 | hash = ip_vs_conn_hashkey(protocol, s_addr, s_port); | |
193 | ||
194 | ct_read_lock(hash); | |
195 | ||
196 | list_for_each_entry(cp, &ip_vs_conn_tab[hash], c_list) { | |
197 | if (s_addr==cp->caddr && s_port==cp->cport && | |
198 | d_port==cp->vport && d_addr==cp->vaddr && | |
87375ab4 | 199 | ((!s_port) ^ (!(cp->flags & IP_VS_CONN_F_NO_CPORT))) && |
1da177e4 LT |
200 | protocol==cp->protocol) { |
201 | /* HIT */ | |
202 | atomic_inc(&cp->refcnt); | |
203 | ct_read_unlock(hash); | |
204 | return cp; | |
205 | } | |
206 | } | |
207 | ||
208 | ct_read_unlock(hash); | |
209 | ||
210 | return NULL; | |
211 | } | |
212 | ||
213 | struct ip_vs_conn *ip_vs_conn_in_get | |
214 | (int protocol, __u32 s_addr, __u16 s_port, __u32 d_addr, __u16 d_port) | |
215 | { | |
216 | struct ip_vs_conn *cp; | |
217 | ||
218 | cp = __ip_vs_conn_in_get(protocol, s_addr, s_port, d_addr, d_port); | |
219 | if (!cp && atomic_read(&ip_vs_conn_no_cport_cnt)) | |
220 | cp = __ip_vs_conn_in_get(protocol, s_addr, 0, d_addr, d_port); | |
221 | ||
222 | IP_VS_DBG(7, "lookup/in %s %u.%u.%u.%u:%d->%u.%u.%u.%u:%d %s\n", | |
223 | ip_vs_proto_name(protocol), | |
224 | NIPQUAD(s_addr), ntohs(s_port), | |
225 | NIPQUAD(d_addr), ntohs(d_port), | |
226 | cp?"hit":"not hit"); | |
227 | ||
228 | return cp; | |
229 | } | |
230 | ||
87375ab4 JA |
231 | /* Get reference to connection template */ |
232 | struct ip_vs_conn *ip_vs_ct_in_get | |
233 | (int protocol, __u32 s_addr, __u16 s_port, __u32 d_addr, __u16 d_port) | |
234 | { | |
235 | unsigned hash; | |
236 | struct ip_vs_conn *cp; | |
237 | ||
238 | hash = ip_vs_conn_hashkey(protocol, s_addr, s_port); | |
239 | ||
240 | ct_read_lock(hash); | |
241 | ||
242 | list_for_each_entry(cp, &ip_vs_conn_tab[hash], c_list) { | |
243 | if (s_addr==cp->caddr && s_port==cp->cport && | |
244 | d_port==cp->vport && d_addr==cp->vaddr && | |
245 | cp->flags & IP_VS_CONN_F_TEMPLATE && | |
246 | protocol==cp->protocol) { | |
247 | /* HIT */ | |
248 | atomic_inc(&cp->refcnt); | |
249 | goto out; | |
250 | } | |
251 | } | |
252 | cp = NULL; | |
253 | ||
254 | out: | |
255 | ct_read_unlock(hash); | |
256 | ||
257 | IP_VS_DBG(7, "template lookup/in %s %u.%u.%u.%u:%d->%u.%u.%u.%u:%d %s\n", | |
258 | ip_vs_proto_name(protocol), | |
259 | NIPQUAD(s_addr), ntohs(s_port), | |
260 | NIPQUAD(d_addr), ntohs(d_port), | |
261 | cp?"hit":"not hit"); | |
262 | ||
263 | return cp; | |
264 | } | |
1da177e4 LT |
265 | |
266 | /* | |
267 | * Gets ip_vs_conn associated with supplied parameters in the ip_vs_conn_tab. | |
268 | * Called for pkts coming from inside-to-OUTside. | |
269 | * s_addr, s_port: pkt source address (inside host) | |
270 | * d_addr, d_port: pkt dest address (foreign host) | |
271 | */ | |
272 | struct ip_vs_conn *ip_vs_conn_out_get | |
273 | (int protocol, __u32 s_addr, __u16 s_port, __u32 d_addr, __u16 d_port) | |
274 | { | |
275 | unsigned hash; | |
276 | struct ip_vs_conn *cp, *ret=NULL; | |
277 | ||
278 | /* | |
279 | * Check for "full" addressed entries | |
280 | */ | |
281 | hash = ip_vs_conn_hashkey(protocol, d_addr, d_port); | |
282 | ||
283 | ct_read_lock(hash); | |
284 | ||
285 | list_for_each_entry(cp, &ip_vs_conn_tab[hash], c_list) { | |
286 | if (d_addr == cp->caddr && d_port == cp->cport && | |
287 | s_port == cp->dport && s_addr == cp->daddr && | |
288 | protocol == cp->protocol) { | |
289 | /* HIT */ | |
290 | atomic_inc(&cp->refcnt); | |
291 | ret = cp; | |
292 | break; | |
293 | } | |
294 | } | |
295 | ||
296 | ct_read_unlock(hash); | |
297 | ||
298 | IP_VS_DBG(7, "lookup/out %s %u.%u.%u.%u:%d->%u.%u.%u.%u:%d %s\n", | |
299 | ip_vs_proto_name(protocol), | |
300 | NIPQUAD(s_addr), ntohs(s_port), | |
301 | NIPQUAD(d_addr), ntohs(d_port), | |
302 | ret?"hit":"not hit"); | |
303 | ||
304 | return ret; | |
305 | } | |
306 | ||
307 | ||
308 | /* | |
309 | * Put back the conn and restart its timer with its timeout | |
310 | */ | |
311 | void ip_vs_conn_put(struct ip_vs_conn *cp) | |
312 | { | |
313 | /* reset it expire in its timeout */ | |
314 | mod_timer(&cp->timer, jiffies+cp->timeout); | |
315 | ||
316 | __ip_vs_conn_put(cp); | |
317 | } | |
318 | ||
319 | ||
320 | /* | |
321 | * Fill a no_client_port connection with a client port number | |
322 | */ | |
323 | void ip_vs_conn_fill_cport(struct ip_vs_conn *cp, __u16 cport) | |
324 | { | |
325 | if (ip_vs_conn_unhash(cp)) { | |
326 | spin_lock(&cp->lock); | |
327 | if (cp->flags & IP_VS_CONN_F_NO_CPORT) { | |
328 | atomic_dec(&ip_vs_conn_no_cport_cnt); | |
329 | cp->flags &= ~IP_VS_CONN_F_NO_CPORT; | |
330 | cp->cport = cport; | |
331 | } | |
332 | spin_unlock(&cp->lock); | |
333 | ||
334 | /* hash on new dport */ | |
335 | ip_vs_conn_hash(cp); | |
336 | } | |
337 | } | |
338 | ||
339 | ||
340 | /* | |
341 | * Bind a connection entry with the corresponding packet_xmit. | |
342 | * Called by ip_vs_conn_new. | |
343 | */ | |
344 | static inline void ip_vs_bind_xmit(struct ip_vs_conn *cp) | |
345 | { | |
346 | switch (IP_VS_FWD_METHOD(cp)) { | |
347 | case IP_VS_CONN_F_MASQ: | |
348 | cp->packet_xmit = ip_vs_nat_xmit; | |
349 | break; | |
350 | ||
351 | case IP_VS_CONN_F_TUNNEL: | |
352 | cp->packet_xmit = ip_vs_tunnel_xmit; | |
353 | break; | |
354 | ||
355 | case IP_VS_CONN_F_DROUTE: | |
356 | cp->packet_xmit = ip_vs_dr_xmit; | |
357 | break; | |
358 | ||
359 | case IP_VS_CONN_F_LOCALNODE: | |
360 | cp->packet_xmit = ip_vs_null_xmit; | |
361 | break; | |
362 | ||
363 | case IP_VS_CONN_F_BYPASS: | |
364 | cp->packet_xmit = ip_vs_bypass_xmit; | |
365 | break; | |
366 | } | |
367 | } | |
368 | ||
369 | ||
370 | static inline int ip_vs_dest_totalconns(struct ip_vs_dest *dest) | |
371 | { | |
372 | return atomic_read(&dest->activeconns) | |
373 | + atomic_read(&dest->inactconns); | |
374 | } | |
375 | ||
376 | /* | |
377 | * Bind a connection entry with a virtual service destination | |
378 | * Called just after a new connection entry is created. | |
379 | */ | |
380 | static inline void | |
381 | ip_vs_bind_dest(struct ip_vs_conn *cp, struct ip_vs_dest *dest) | |
382 | { | |
383 | /* if dest is NULL, then return directly */ | |
384 | if (!dest) | |
385 | return; | |
386 | ||
387 | /* Increase the refcnt counter of the dest */ | |
388 | atomic_inc(&dest->refcnt); | |
389 | ||
390 | /* Bind with the destination and its corresponding transmitter */ | |
391 | cp->flags |= atomic_read(&dest->conn_flags); | |
392 | cp->dest = dest; | |
393 | ||
394 | IP_VS_DBG(9, "Bind-dest %s c:%u.%u.%u.%u:%d v:%u.%u.%u.%u:%d " | |
395 | "d:%u.%u.%u.%u:%d fwd:%c s:%u flg:%X cnt:%d destcnt:%d\n", | |
396 | ip_vs_proto_name(cp->protocol), | |
397 | NIPQUAD(cp->caddr), ntohs(cp->cport), | |
398 | NIPQUAD(cp->vaddr), ntohs(cp->vport), | |
399 | NIPQUAD(cp->daddr), ntohs(cp->dport), | |
400 | ip_vs_fwd_tag(cp), cp->state, | |
401 | cp->flags, atomic_read(&cp->refcnt), | |
402 | atomic_read(&dest->refcnt)); | |
403 | ||
404 | /* Update the connection counters */ | |
87375ab4 | 405 | if (!(cp->flags & IP_VS_CONN_F_TEMPLATE)) { |
1da177e4 LT |
406 | /* It is a normal connection, so increase the inactive |
407 | connection counter because it is in TCP SYNRECV | |
408 | state (inactive) or other protocol inacive state */ | |
409 | atomic_inc(&dest->inactconns); | |
410 | } else { | |
411 | /* It is a persistent connection/template, so increase | |
412 | the peristent connection counter */ | |
413 | atomic_inc(&dest->persistconns); | |
414 | } | |
415 | ||
416 | if (dest->u_threshold != 0 && | |
417 | ip_vs_dest_totalconns(dest) >= dest->u_threshold) | |
418 | dest->flags |= IP_VS_DEST_F_OVERLOAD; | |
419 | } | |
420 | ||
421 | ||
422 | /* | |
423 | * Unbind a connection entry with its VS destination | |
424 | * Called by the ip_vs_conn_expire function. | |
425 | */ | |
426 | static inline void ip_vs_unbind_dest(struct ip_vs_conn *cp) | |
427 | { | |
428 | struct ip_vs_dest *dest = cp->dest; | |
429 | ||
430 | if (!dest) | |
431 | return; | |
432 | ||
433 | IP_VS_DBG(9, "Unbind-dest %s c:%u.%u.%u.%u:%d v:%u.%u.%u.%u:%d " | |
434 | "d:%u.%u.%u.%u:%d fwd:%c s:%u flg:%X cnt:%d destcnt:%d\n", | |
435 | ip_vs_proto_name(cp->protocol), | |
436 | NIPQUAD(cp->caddr), ntohs(cp->cport), | |
437 | NIPQUAD(cp->vaddr), ntohs(cp->vport), | |
438 | NIPQUAD(cp->daddr), ntohs(cp->dport), | |
439 | ip_vs_fwd_tag(cp), cp->state, | |
440 | cp->flags, atomic_read(&cp->refcnt), | |
441 | atomic_read(&dest->refcnt)); | |
442 | ||
443 | /* Update the connection counters */ | |
87375ab4 | 444 | if (!(cp->flags & IP_VS_CONN_F_TEMPLATE)) { |
1da177e4 LT |
445 | /* It is a normal connection, so decrease the inactconns |
446 | or activeconns counter */ | |
447 | if (cp->flags & IP_VS_CONN_F_INACTIVE) { | |
448 | atomic_dec(&dest->inactconns); | |
449 | } else { | |
450 | atomic_dec(&dest->activeconns); | |
451 | } | |
452 | } else { | |
453 | /* It is a persistent connection/template, so decrease | |
454 | the peristent connection counter */ | |
455 | atomic_dec(&dest->persistconns); | |
456 | } | |
457 | ||
458 | if (dest->l_threshold != 0) { | |
459 | if (ip_vs_dest_totalconns(dest) < dest->l_threshold) | |
460 | dest->flags &= ~IP_VS_DEST_F_OVERLOAD; | |
461 | } else if (dest->u_threshold != 0) { | |
462 | if (ip_vs_dest_totalconns(dest) * 4 < dest->u_threshold * 3) | |
463 | dest->flags &= ~IP_VS_DEST_F_OVERLOAD; | |
464 | } else { | |
465 | if (dest->flags & IP_VS_DEST_F_OVERLOAD) | |
466 | dest->flags &= ~IP_VS_DEST_F_OVERLOAD; | |
467 | } | |
468 | ||
469 | /* | |
470 | * Simply decrease the refcnt of the dest, because the | |
471 | * dest will be either in service's destination list | |
472 | * or in the trash. | |
473 | */ | |
474 | atomic_dec(&dest->refcnt); | |
475 | } | |
476 | ||
477 | ||
478 | /* | |
479 | * Checking if the destination of a connection template is available. | |
480 | * If available, return 1, otherwise invalidate this connection | |
481 | * template and return 0. | |
482 | */ | |
483 | int ip_vs_check_template(struct ip_vs_conn *ct) | |
484 | { | |
485 | struct ip_vs_dest *dest = ct->dest; | |
486 | ||
487 | /* | |
488 | * Checking the dest server status. | |
489 | */ | |
490 | if ((dest == NULL) || | |
491 | !(dest->flags & IP_VS_DEST_F_AVAILABLE) || | |
492 | (sysctl_ip_vs_expire_quiescent_template && | |
493 | (atomic_read(&dest->weight) == 0))) { | |
494 | IP_VS_DBG(9, "check_template: dest not available for " | |
495 | "protocol %s s:%u.%u.%u.%u:%d v:%u.%u.%u.%u:%d " | |
496 | "-> d:%u.%u.%u.%u:%d\n", | |
497 | ip_vs_proto_name(ct->protocol), | |
498 | NIPQUAD(ct->caddr), ntohs(ct->cport), | |
499 | NIPQUAD(ct->vaddr), ntohs(ct->vport), | |
500 | NIPQUAD(ct->daddr), ntohs(ct->dport)); | |
501 | ||
502 | /* | |
503 | * Invalidate the connection template | |
504 | */ | |
f5e229db | 505 | if (ct->vport != 65535) { |
1da177e4 LT |
506 | if (ip_vs_conn_unhash(ct)) { |
507 | ct->dport = 65535; | |
508 | ct->vport = 65535; | |
509 | ct->cport = 0; | |
510 | ip_vs_conn_hash(ct); | |
511 | } | |
512 | } | |
513 | ||
514 | /* | |
515 | * Simply decrease the refcnt of the template, | |
516 | * don't restart its timer. | |
517 | */ | |
518 | atomic_dec(&ct->refcnt); | |
519 | return 0; | |
520 | } | |
521 | return 1; | |
522 | } | |
523 | ||
524 | static void ip_vs_conn_expire(unsigned long data) | |
525 | { | |
526 | struct ip_vs_conn *cp = (struct ip_vs_conn *)data; | |
527 | ||
528 | cp->timeout = 60*HZ; | |
529 | ||
530 | /* | |
531 | * hey, I'm using it | |
532 | */ | |
533 | atomic_inc(&cp->refcnt); | |
534 | ||
535 | /* | |
536 | * do I control anybody? | |
537 | */ | |
538 | if (atomic_read(&cp->n_control)) | |
539 | goto expire_later; | |
540 | ||
541 | /* | |
542 | * unhash it if it is hashed in the conn table | |
543 | */ | |
544 | if (!ip_vs_conn_unhash(cp)) | |
545 | goto expire_later; | |
546 | ||
547 | /* | |
548 | * refcnt==1 implies I'm the only one referrer | |
549 | */ | |
550 | if (likely(atomic_read(&cp->refcnt) == 1)) { | |
551 | /* delete the timer if it is activated by other users */ | |
552 | if (timer_pending(&cp->timer)) | |
553 | del_timer(&cp->timer); | |
554 | ||
555 | /* does anybody control me? */ | |
556 | if (cp->control) | |
557 | ip_vs_control_del(cp); | |
558 | ||
559 | if (unlikely(cp->app != NULL)) | |
560 | ip_vs_unbind_app(cp); | |
561 | ip_vs_unbind_dest(cp); | |
562 | if (cp->flags & IP_VS_CONN_F_NO_CPORT) | |
563 | atomic_dec(&ip_vs_conn_no_cport_cnt); | |
564 | atomic_dec(&ip_vs_conn_count); | |
565 | ||
566 | kmem_cache_free(ip_vs_conn_cachep, cp); | |
567 | return; | |
568 | } | |
569 | ||
570 | /* hash it back to the table */ | |
571 | ip_vs_conn_hash(cp); | |
572 | ||
573 | expire_later: | |
574 | IP_VS_DBG(7, "delayed: refcnt-1=%d conn.n_control=%d\n", | |
575 | atomic_read(&cp->refcnt)-1, | |
576 | atomic_read(&cp->n_control)); | |
577 | ||
578 | ip_vs_conn_put(cp); | |
579 | } | |
580 | ||
581 | ||
582 | void ip_vs_conn_expire_now(struct ip_vs_conn *cp) | |
583 | { | |
584 | if (del_timer(&cp->timer)) | |
585 | mod_timer(&cp->timer, jiffies); | |
1da177e4 LT |
586 | } |
587 | ||
588 | ||
589 | /* | |
590 | * Create a new connection entry and hash it into the ip_vs_conn_tab | |
591 | */ | |
592 | struct ip_vs_conn * | |
593 | ip_vs_conn_new(int proto, __u32 caddr, __u16 cport, __u32 vaddr, __u16 vport, | |
594 | __u32 daddr, __u16 dport, unsigned flags, | |
595 | struct ip_vs_dest *dest) | |
596 | { | |
597 | struct ip_vs_conn *cp; | |
598 | struct ip_vs_protocol *pp = ip_vs_proto_get(proto); | |
599 | ||
600 | cp = kmem_cache_alloc(ip_vs_conn_cachep, GFP_ATOMIC); | |
601 | if (cp == NULL) { | |
602 | IP_VS_ERR_RL("ip_vs_conn_new: no memory available.\n"); | |
603 | return NULL; | |
604 | } | |
605 | ||
606 | memset(cp, 0, sizeof(*cp)); | |
607 | INIT_LIST_HEAD(&cp->c_list); | |
608 | init_timer(&cp->timer); | |
609 | cp->timer.data = (unsigned long)cp; | |
610 | cp->timer.function = ip_vs_conn_expire; | |
611 | cp->protocol = proto; | |
612 | cp->caddr = caddr; | |
613 | cp->cport = cport; | |
614 | cp->vaddr = vaddr; | |
615 | cp->vport = vport; | |
616 | cp->daddr = daddr; | |
617 | cp->dport = dport; | |
618 | cp->flags = flags; | |
619 | spin_lock_init(&cp->lock); | |
620 | ||
621 | /* | |
622 | * Set the entry is referenced by the current thread before hashing | |
623 | * it in the table, so that other thread run ip_vs_random_dropentry | |
624 | * but cannot drop this entry. | |
625 | */ | |
626 | atomic_set(&cp->refcnt, 1); | |
627 | ||
628 | atomic_set(&cp->n_control, 0); | |
629 | atomic_set(&cp->in_pkts, 0); | |
630 | ||
631 | atomic_inc(&ip_vs_conn_count); | |
632 | if (flags & IP_VS_CONN_F_NO_CPORT) | |
633 | atomic_inc(&ip_vs_conn_no_cport_cnt); | |
634 | ||
635 | /* Bind the connection with a destination server */ | |
636 | ip_vs_bind_dest(cp, dest); | |
637 | ||
638 | /* Set its state and timeout */ | |
639 | cp->state = 0; | |
640 | cp->timeout = 3*HZ; | |
641 | ||
642 | /* Bind its packet transmitter */ | |
643 | ip_vs_bind_xmit(cp); | |
644 | ||
645 | if (unlikely(pp && atomic_read(&pp->appcnt))) | |
646 | ip_vs_bind_app(cp, pp); | |
647 | ||
648 | /* Hash it in the ip_vs_conn_tab finally */ | |
649 | ip_vs_conn_hash(cp); | |
650 | ||
651 | return cp; | |
652 | } | |
653 | ||
654 | ||
655 | /* | |
656 | * /proc/net/ip_vs_conn entries | |
657 | */ | |
658 | #ifdef CONFIG_PROC_FS | |
659 | ||
660 | static void *ip_vs_conn_array(struct seq_file *seq, loff_t pos) | |
661 | { | |
662 | int idx; | |
663 | struct ip_vs_conn *cp; | |
664 | ||
665 | for(idx = 0; idx < IP_VS_CONN_TAB_SIZE; idx++) { | |
666 | ct_read_lock_bh(idx); | |
667 | list_for_each_entry(cp, &ip_vs_conn_tab[idx], c_list) { | |
668 | if (pos-- == 0) { | |
669 | seq->private = &ip_vs_conn_tab[idx]; | |
670 | return cp; | |
671 | } | |
672 | } | |
673 | ct_read_unlock_bh(idx); | |
674 | } | |
675 | ||
676 | return NULL; | |
677 | } | |
678 | ||
679 | static void *ip_vs_conn_seq_start(struct seq_file *seq, loff_t *pos) | |
680 | { | |
681 | seq->private = NULL; | |
682 | return *pos ? ip_vs_conn_array(seq, *pos - 1) :SEQ_START_TOKEN; | |
683 | } | |
684 | ||
685 | static void *ip_vs_conn_seq_next(struct seq_file *seq, void *v, loff_t *pos) | |
686 | { | |
687 | struct ip_vs_conn *cp = v; | |
688 | struct list_head *e, *l = seq->private; | |
689 | int idx; | |
690 | ||
691 | ++*pos; | |
692 | if (v == SEQ_START_TOKEN) | |
693 | return ip_vs_conn_array(seq, 0); | |
694 | ||
695 | /* more on same hash chain? */ | |
696 | if ((e = cp->c_list.next) != l) | |
697 | return list_entry(e, struct ip_vs_conn, c_list); | |
698 | ||
699 | idx = l - ip_vs_conn_tab; | |
700 | ct_read_unlock_bh(idx); | |
701 | ||
702 | while (++idx < IP_VS_CONN_TAB_SIZE) { | |
703 | ct_read_lock_bh(idx); | |
704 | list_for_each_entry(cp, &ip_vs_conn_tab[idx], c_list) { | |
705 | seq->private = &ip_vs_conn_tab[idx]; | |
706 | return cp; | |
707 | } | |
708 | ct_read_unlock_bh(idx); | |
709 | } | |
710 | seq->private = NULL; | |
711 | return NULL; | |
712 | } | |
713 | ||
714 | static void ip_vs_conn_seq_stop(struct seq_file *seq, void *v) | |
715 | { | |
716 | struct list_head *l = seq->private; | |
717 | ||
718 | if (l) | |
719 | ct_read_unlock_bh(l - ip_vs_conn_tab); | |
720 | } | |
721 | ||
722 | static int ip_vs_conn_seq_show(struct seq_file *seq, void *v) | |
723 | { | |
724 | ||
725 | if (v == SEQ_START_TOKEN) | |
726 | seq_puts(seq, | |
727 | "Pro FromIP FPrt ToIP TPrt DestIP DPrt State Expires\n"); | |
728 | else { | |
729 | const struct ip_vs_conn *cp = v; | |
730 | ||
731 | seq_printf(seq, | |
732 | "%-3s %08X %04X %08X %04X %08X %04X %-11s %7lu\n", | |
733 | ip_vs_proto_name(cp->protocol), | |
734 | ntohl(cp->caddr), ntohs(cp->cport), | |
735 | ntohl(cp->vaddr), ntohs(cp->vport), | |
736 | ntohl(cp->daddr), ntohs(cp->dport), | |
737 | ip_vs_state_name(cp->protocol, cp->state), | |
738 | (cp->timer.expires-jiffies)/HZ); | |
739 | } | |
740 | return 0; | |
741 | } | |
742 | ||
743 | static struct seq_operations ip_vs_conn_seq_ops = { | |
744 | .start = ip_vs_conn_seq_start, | |
745 | .next = ip_vs_conn_seq_next, | |
746 | .stop = ip_vs_conn_seq_stop, | |
747 | .show = ip_vs_conn_seq_show, | |
748 | }; | |
749 | ||
750 | static int ip_vs_conn_open(struct inode *inode, struct file *file) | |
751 | { | |
752 | return seq_open(file, &ip_vs_conn_seq_ops); | |
753 | } | |
754 | ||
755 | static struct file_operations ip_vs_conn_fops = { | |
756 | .owner = THIS_MODULE, | |
757 | .open = ip_vs_conn_open, | |
758 | .read = seq_read, | |
759 | .llseek = seq_lseek, | |
760 | .release = seq_release, | |
761 | }; | |
762 | #endif | |
763 | ||
764 | ||
765 | /* | |
766 | * Randomly drop connection entries before running out of memory | |
767 | */ | |
768 | static inline int todrop_entry(struct ip_vs_conn *cp) | |
769 | { | |
770 | /* | |
771 | * The drop rate array needs tuning for real environments. | |
772 | * Called from timer bh only => no locking | |
773 | */ | |
774 | static char todrop_rate[9] = {0, 1, 2, 3, 4, 5, 6, 7, 8}; | |
775 | static char todrop_counter[9] = {0}; | |
776 | int i; | |
777 | ||
778 | /* if the conn entry hasn't lasted for 60 seconds, don't drop it. | |
779 | This will leave enough time for normal connection to get | |
780 | through. */ | |
781 | if (time_before(cp->timeout + jiffies, cp->timer.expires + 60*HZ)) | |
782 | return 0; | |
783 | ||
784 | /* Don't drop the entry if its number of incoming packets is not | |
785 | located in [0, 8] */ | |
786 | i = atomic_read(&cp->in_pkts); | |
787 | if (i > 8 || i < 0) return 0; | |
788 | ||
789 | if (!todrop_rate[i]) return 0; | |
790 | if (--todrop_counter[i] > 0) return 0; | |
791 | ||
792 | todrop_counter[i] = todrop_rate[i]; | |
793 | return 1; | |
794 | } | |
795 | ||
af9debd4 | 796 | /* Called from keventd and must protect itself from softirqs */ |
1da177e4 LT |
797 | void ip_vs_random_dropentry(void) |
798 | { | |
799 | int idx; | |
800 | struct ip_vs_conn *cp; | |
1da177e4 LT |
801 | |
802 | /* | |
803 | * Randomly scan 1/32 of the whole table every second | |
804 | */ | |
805 | for (idx = 0; idx < (IP_VS_CONN_TAB_SIZE>>5); idx++) { | |
806 | unsigned hash = net_random() & IP_VS_CONN_TAB_MASK; | |
807 | ||
808 | /* | |
809 | * Lock is actually needed in this loop. | |
810 | */ | |
af9debd4 | 811 | ct_write_lock_bh(hash); |
1da177e4 LT |
812 | |
813 | list_for_each_entry(cp, &ip_vs_conn_tab[hash], c_list) { | |
87375ab4 | 814 | if (cp->flags & IP_VS_CONN_F_TEMPLATE) |
1da177e4 LT |
815 | /* connection template */ |
816 | continue; | |
817 | ||
818 | if (cp->protocol == IPPROTO_TCP) { | |
819 | switch(cp->state) { | |
820 | case IP_VS_TCP_S_SYN_RECV: | |
821 | case IP_VS_TCP_S_SYNACK: | |
822 | break; | |
823 | ||
824 | case IP_VS_TCP_S_ESTABLISHED: | |
825 | if (todrop_entry(cp)) | |
826 | break; | |
827 | continue; | |
828 | ||
829 | default: | |
830 | continue; | |
831 | } | |
832 | } else { | |
833 | if (!todrop_entry(cp)) | |
834 | continue; | |
835 | } | |
836 | ||
1da177e4 LT |
837 | IP_VS_DBG(4, "del connection\n"); |
838 | ip_vs_conn_expire_now(cp); | |
fb3d8949 | 839 | if (cp->control) { |
1da177e4 | 840 | IP_VS_DBG(4, "del conn template\n"); |
fb3d8949 | 841 | ip_vs_conn_expire_now(cp->control); |
1da177e4 | 842 | } |
1da177e4 | 843 | } |
af9debd4 | 844 | ct_write_unlock_bh(hash); |
1da177e4 LT |
845 | } |
846 | } | |
847 | ||
848 | ||
849 | /* | |
850 | * Flush all the connection entries in the ip_vs_conn_tab | |
851 | */ | |
852 | static void ip_vs_conn_flush(void) | |
853 | { | |
854 | int idx; | |
855 | struct ip_vs_conn *cp; | |
1da177e4 LT |
856 | |
857 | flush_again: | |
858 | for (idx=0; idx<IP_VS_CONN_TAB_SIZE; idx++) { | |
859 | /* | |
860 | * Lock is actually needed in this loop. | |
861 | */ | |
862 | ct_write_lock_bh(idx); | |
863 | ||
864 | list_for_each_entry(cp, &ip_vs_conn_tab[idx], c_list) { | |
1da177e4 | 865 | |
1da177e4 LT |
866 | IP_VS_DBG(4, "del connection\n"); |
867 | ip_vs_conn_expire_now(cp); | |
fb3d8949 | 868 | if (cp->control) { |
1da177e4 | 869 | IP_VS_DBG(4, "del conn template\n"); |
fb3d8949 | 870 | ip_vs_conn_expire_now(cp->control); |
1da177e4 | 871 | } |
1da177e4 LT |
872 | } |
873 | ct_write_unlock_bh(idx); | |
874 | } | |
875 | ||
876 | /* the counter may be not NULL, because maybe some conn entries | |
877 | are run by slow timer handler or unhashed but still referred */ | |
878 | if (atomic_read(&ip_vs_conn_count) != 0) { | |
879 | schedule(); | |
880 | goto flush_again; | |
881 | } | |
882 | } | |
883 | ||
884 | ||
885 | int ip_vs_conn_init(void) | |
886 | { | |
887 | int idx; | |
888 | ||
889 | /* | |
890 | * Allocate the connection hash table and initialize its list heads | |
891 | */ | |
892 | ip_vs_conn_tab = vmalloc(IP_VS_CONN_TAB_SIZE*sizeof(struct list_head)); | |
893 | if (!ip_vs_conn_tab) | |
894 | return -ENOMEM; | |
895 | ||
896 | /* Allocate ip_vs_conn slab cache */ | |
897 | ip_vs_conn_cachep = kmem_cache_create("ip_vs_conn", | |
898 | sizeof(struct ip_vs_conn), 0, | |
899 | SLAB_HWCACHE_ALIGN, NULL, NULL); | |
900 | if (!ip_vs_conn_cachep) { | |
901 | vfree(ip_vs_conn_tab); | |
902 | return -ENOMEM; | |
903 | } | |
904 | ||
905 | IP_VS_INFO("Connection hash table configured " | |
906 | "(size=%d, memory=%ldKbytes)\n", | |
907 | IP_VS_CONN_TAB_SIZE, | |
908 | (long)(IP_VS_CONN_TAB_SIZE*sizeof(struct list_head))/1024); | |
909 | IP_VS_DBG(0, "Each connection entry needs %Zd bytes at least\n", | |
910 | sizeof(struct ip_vs_conn)); | |
911 | ||
912 | for (idx = 0; idx < IP_VS_CONN_TAB_SIZE; idx++) { | |
913 | INIT_LIST_HEAD(&ip_vs_conn_tab[idx]); | |
914 | } | |
915 | ||
916 | for (idx = 0; idx < CT_LOCKARRAY_SIZE; idx++) { | |
917 | rwlock_init(&__ip_vs_conntbl_lock_array[idx].l); | |
918 | } | |
919 | ||
920 | proc_net_fops_create("ip_vs_conn", 0, &ip_vs_conn_fops); | |
921 | ||
922 | /* calculate the random value for connection hash */ | |
923 | get_random_bytes(&ip_vs_conn_rnd, sizeof(ip_vs_conn_rnd)); | |
924 | ||
925 | return 0; | |
926 | } | |
927 | ||
928 | ||
929 | void ip_vs_conn_cleanup(void) | |
930 | { | |
931 | /* flush all the connection entries first */ | |
932 | ip_vs_conn_flush(); | |
933 | ||
934 | /* Release the empty cache */ | |
935 | kmem_cache_destroy(ip_vs_conn_cachep); | |
936 | proc_net_remove("ip_vs_conn"); | |
937 | vfree(ip_vs_conn_tab); | |
938 | } |