]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * ip6_flowlabel.c IPv6 flowlabel manager. | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or | |
5 | * modify it under the terms of the GNU General Public License | |
6 | * as published by the Free Software Foundation; either version | |
7 | * 2 of the License, or (at your option) any later version. | |
8 | * | |
9 | * Authors: Alexey Kuznetsov, <kuznet@ms2.inr.ac.ru> | |
10 | */ | |
11 | ||
4fc268d2 | 12 | #include <linux/capability.h> |
1da177e4 LT |
13 | #include <linux/errno.h> |
14 | #include <linux/types.h> | |
15 | #include <linux/socket.h> | |
16 | #include <linux/net.h> | |
17 | #include <linux/netdevice.h> | |
18 | #include <linux/if_arp.h> | |
19 | #include <linux/in6.h> | |
20 | #include <linux/route.h> | |
21 | #include <linux/proc_fs.h> | |
22 | #include <linux/seq_file.h> | |
5a0e3ad6 | 23 | #include <linux/slab.h> |
bc3b2d7f | 24 | #include <linux/export.h> |
4f82f457 | 25 | #include <linux/pid_namespace.h> |
1da177e4 | 26 | |
457c4cbc | 27 | #include <net/net_namespace.h> |
1da177e4 LT |
28 | #include <net/sock.h> |
29 | ||
30 | #include <net/ipv6.h> | |
31 | #include <net/ndisc.h> | |
32 | #include <net/protocol.h> | |
33 | #include <net/ip6_route.h> | |
34 | #include <net/addrconf.h> | |
35 | #include <net/rawv6.h> | |
36 | #include <net/icmp.h> | |
37 | #include <net/transp_v6.h> | |
38 | ||
39 | #include <asm/uaccess.h> | |
40 | ||
41 | #define FL_MIN_LINGER 6 /* Minimal linger. It is set to 6sec specified | |
42 | in old IPv6 RFC. Well, it was reasonable value. | |
43 | */ | |
44 | #define FL_MAX_LINGER 60 /* Maximal linger timeout */ | |
45 | ||
46 | /* FL hash table */ | |
47 | ||
48 | #define FL_MAX_PER_SOCK 32 | |
49 | #define FL_MAX_SIZE 4096 | |
50 | #define FL_HASH_MASK 255 | |
51 | #define FL_HASH(l) (ntohl(l)&FL_HASH_MASK) | |
52 | ||
53 | static atomic_t fl_size = ATOMIC_INIT(0); | |
d3aedd5e | 54 | static struct ip6_flowlabel __rcu *fl_ht[FL_HASH_MASK+1]; |
1da177e4 LT |
55 | |
56 | static void ip6_fl_gc(unsigned long dummy); | |
8d06afab | 57 | static DEFINE_TIMER(ip6_fl_gc_timer, ip6_fl_gc, 0, 0); |
1da177e4 LT |
58 | |
59 | /* FL hash table lock: it protects only of GC */ | |
60 | ||
d3aedd5e | 61 | static DEFINE_SPINLOCK(ip6_fl_lock); |
1da177e4 LT |
62 | |
63 | /* Big socket sock */ | |
64 | ||
18367681 | 65 | static DEFINE_SPINLOCK(ip6_sk_fl_lock); |
1da177e4 | 66 | |
d3aedd5e | 67 | #define for_each_fl_rcu(hash, fl) \ |
6a98dcf0 | 68 | for (fl = rcu_dereference_bh(fl_ht[(hash)]); \ |
d3aedd5e | 69 | fl != NULL; \ |
6a98dcf0 | 70 | fl = rcu_dereference_bh(fl->next)) |
d3aedd5e | 71 | #define for_each_fl_continue_rcu(fl) \ |
6a98dcf0 | 72 | for (fl = rcu_dereference_bh(fl->next); \ |
d3aedd5e | 73 | fl != NULL; \ |
6a98dcf0 | 74 | fl = rcu_dereference_bh(fl->next)) |
1da177e4 | 75 | |
18367681 YH |
76 | #define for_each_sk_fl_rcu(np, sfl) \ |
77 | for (sfl = rcu_dereference_bh(np->ipv6_fl_list); \ | |
78 | sfl != NULL; \ | |
79 | sfl = rcu_dereference_bh(sfl->next)) | |
80 | ||
60e8fbc4 | 81 | static inline struct ip6_flowlabel *__fl_lookup(struct net *net, __be32 label) |
1da177e4 LT |
82 | { |
83 | struct ip6_flowlabel *fl; | |
84 | ||
d3aedd5e | 85 | for_each_fl_rcu(FL_HASH(label), fl) { |
09ad9bc7 | 86 | if (fl->label == label && net_eq(fl->fl_net, net)) |
1da177e4 LT |
87 | return fl; |
88 | } | |
89 | return NULL; | |
90 | } | |
91 | ||
60e8fbc4 | 92 | static struct ip6_flowlabel *fl_lookup(struct net *net, __be32 label) |
1da177e4 LT |
93 | { |
94 | struct ip6_flowlabel *fl; | |
95 | ||
d3aedd5e | 96 | rcu_read_lock_bh(); |
60e8fbc4 | 97 | fl = __fl_lookup(net, label); |
d3aedd5e YH |
98 | if (fl && !atomic_inc_not_zero(&fl->users)) |
99 | fl = NULL; | |
100 | rcu_read_unlock_bh(); | |
1da177e4 LT |
101 | return fl; |
102 | } | |
103 | ||
104 | ||
105 | static void fl_free(struct ip6_flowlabel *fl) | |
106 | { | |
60e8fbc4 | 107 | if (fl) { |
898132ae DC |
108 | if (fl->share == IPV6_FL_S_PROCESS) |
109 | put_pid(fl->owner.pid); | |
60e8fbc4 | 110 | release_net(fl->fl_net); |
1da177e4 | 111 | kfree(fl->opt); |
d3aedd5e | 112 | kfree_rcu(fl, rcu); |
60e8fbc4 | 113 | } |
1da177e4 LT |
114 | } |
115 | ||
116 | static void fl_release(struct ip6_flowlabel *fl) | |
117 | { | |
d3aedd5e | 118 | spin_lock_bh(&ip6_fl_lock); |
1da177e4 LT |
119 | |
120 | fl->lastuse = jiffies; | |
121 | if (atomic_dec_and_test(&fl->users)) { | |
122 | unsigned long ttd = fl->lastuse + fl->linger; | |
123 | if (time_after(ttd, fl->expires)) | |
124 | fl->expires = ttd; | |
125 | ttd = fl->expires; | |
126 | if (fl->opt && fl->share == IPV6_FL_S_EXCL) { | |
127 | struct ipv6_txoptions *opt = fl->opt; | |
128 | fl->opt = NULL; | |
129 | kfree(opt); | |
130 | } | |
131 | if (!timer_pending(&ip6_fl_gc_timer) || | |
132 | time_after(ip6_fl_gc_timer.expires, ttd)) | |
133 | mod_timer(&ip6_fl_gc_timer, ttd); | |
134 | } | |
d3aedd5e | 135 | spin_unlock_bh(&ip6_fl_lock); |
1da177e4 LT |
136 | } |
137 | ||
138 | static void ip6_fl_gc(unsigned long dummy) | |
139 | { | |
140 | int i; | |
141 | unsigned long now = jiffies; | |
142 | unsigned long sched = 0; | |
143 | ||
d3aedd5e | 144 | spin_lock(&ip6_fl_lock); |
1da177e4 LT |
145 | |
146 | for (i=0; i<=FL_HASH_MASK; i++) { | |
7f0e44ac ED |
147 | struct ip6_flowlabel *fl; |
148 | struct ip6_flowlabel __rcu **flp; | |
149 | ||
1da177e4 | 150 | flp = &fl_ht[i]; |
d3aedd5e YH |
151 | while ((fl = rcu_dereference_protected(*flp, |
152 | lockdep_is_held(&ip6_fl_lock))) != NULL) { | |
1da177e4 LT |
153 | if (atomic_read(&fl->users) == 0) { |
154 | unsigned long ttd = fl->lastuse + fl->linger; | |
155 | if (time_after(ttd, fl->expires)) | |
156 | fl->expires = ttd; | |
157 | ttd = fl->expires; | |
158 | if (time_after_eq(now, ttd)) { | |
159 | *flp = fl->next; | |
160 | fl_free(fl); | |
161 | atomic_dec(&fl_size); | |
162 | continue; | |
163 | } | |
164 | if (!sched || time_before(ttd, sched)) | |
165 | sched = ttd; | |
166 | } | |
167 | flp = &fl->next; | |
168 | } | |
169 | } | |
170 | if (!sched && atomic_read(&fl_size)) | |
171 | sched = now + FL_MAX_LINGER; | |
172 | if (sched) { | |
60e8fbc4 | 173 | mod_timer(&ip6_fl_gc_timer, sched); |
1da177e4 | 174 | } |
d3aedd5e | 175 | spin_unlock(&ip6_fl_lock); |
1da177e4 LT |
176 | } |
177 | ||
2c8c1e72 | 178 | static void __net_exit ip6_fl_purge(struct net *net) |
60e8fbc4 BT |
179 | { |
180 | int i; | |
181 | ||
d3aedd5e | 182 | spin_lock(&ip6_fl_lock); |
60e8fbc4 | 183 | for (i = 0; i <= FL_HASH_MASK; i++) { |
7f0e44ac ED |
184 | struct ip6_flowlabel *fl; |
185 | struct ip6_flowlabel __rcu **flp; | |
186 | ||
60e8fbc4 | 187 | flp = &fl_ht[i]; |
d3aedd5e YH |
188 | while ((fl = rcu_dereference_protected(*flp, |
189 | lockdep_is_held(&ip6_fl_lock))) != NULL) { | |
09ad9bc7 OP |
190 | if (net_eq(fl->fl_net, net) && |
191 | atomic_read(&fl->users) == 0) { | |
60e8fbc4 BT |
192 | *flp = fl->next; |
193 | fl_free(fl); | |
194 | atomic_dec(&fl_size); | |
195 | continue; | |
196 | } | |
197 | flp = &fl->next; | |
198 | } | |
199 | } | |
d3aedd5e | 200 | spin_unlock(&ip6_fl_lock); |
60e8fbc4 BT |
201 | } |
202 | ||
203 | static struct ip6_flowlabel *fl_intern(struct net *net, | |
204 | struct ip6_flowlabel *fl, __be32 label) | |
1da177e4 | 205 | { |
78c2e502 PE |
206 | struct ip6_flowlabel *lfl; |
207 | ||
1da177e4 LT |
208 | fl->label = label & IPV6_FLOWLABEL_MASK; |
209 | ||
d3aedd5e | 210 | spin_lock_bh(&ip6_fl_lock); |
1da177e4 LT |
211 | if (label == 0) { |
212 | for (;;) { | |
213 | fl->label = htonl(net_random())&IPV6_FLOWLABEL_MASK; | |
214 | if (fl->label) { | |
60e8fbc4 | 215 | lfl = __fl_lookup(net, fl->label); |
1da177e4 LT |
216 | if (lfl == NULL) |
217 | break; | |
218 | } | |
219 | } | |
78c2e502 PE |
220 | } else { |
221 | /* | |
222 | * we dropper the ip6_fl_lock, so this entry could reappear | |
223 | * and we need to recheck with it. | |
224 | * | |
225 | * OTOH no need to search the active socket first, like it is | |
226 | * done in ipv6_flowlabel_opt - sock is locked, so new entry | |
227 | * with the same label can only appear on another sock | |
228 | */ | |
60e8fbc4 | 229 | lfl = __fl_lookup(net, fl->label); |
78c2e502 PE |
230 | if (lfl != NULL) { |
231 | atomic_inc(&lfl->users); | |
d3aedd5e | 232 | spin_unlock_bh(&ip6_fl_lock); |
78c2e502 PE |
233 | return lfl; |
234 | } | |
1da177e4 LT |
235 | } |
236 | ||
237 | fl->lastuse = jiffies; | |
238 | fl->next = fl_ht[FL_HASH(fl->label)]; | |
d3aedd5e | 239 | rcu_assign_pointer(fl_ht[FL_HASH(fl->label)], fl); |
1da177e4 | 240 | atomic_inc(&fl_size); |
d3aedd5e | 241 | spin_unlock_bh(&ip6_fl_lock); |
78c2e502 | 242 | return NULL; |
1da177e4 LT |
243 | } |
244 | ||
245 | ||
246 | ||
247 | /* Socket flowlabel lists */ | |
248 | ||
90bcaf7b | 249 | struct ip6_flowlabel * fl6_sock_lookup(struct sock *sk, __be32 label) |
1da177e4 LT |
250 | { |
251 | struct ipv6_fl_socklist *sfl; | |
252 | struct ipv6_pinfo *np = inet6_sk(sk); | |
253 | ||
254 | label &= IPV6_FLOWLABEL_MASK; | |
255 | ||
18367681 YH |
256 | rcu_read_lock_bh(); |
257 | for_each_sk_fl_rcu(np, sfl) { | |
1da177e4 LT |
258 | struct ip6_flowlabel *fl = sfl->fl; |
259 | if (fl->label == label) { | |
260 | fl->lastuse = jiffies; | |
261 | atomic_inc(&fl->users); | |
18367681 | 262 | rcu_read_unlock_bh(); |
1da177e4 LT |
263 | return fl; |
264 | } | |
265 | } | |
18367681 | 266 | rcu_read_unlock_bh(); |
1da177e4 LT |
267 | return NULL; |
268 | } | |
269 | ||
3cf3dc6c ACM |
270 | EXPORT_SYMBOL_GPL(fl6_sock_lookup); |
271 | ||
1da177e4 LT |
272 | void fl6_free_socklist(struct sock *sk) |
273 | { | |
274 | struct ipv6_pinfo *np = inet6_sk(sk); | |
275 | struct ipv6_fl_socklist *sfl; | |
276 | ||
18367681 | 277 | if (!rcu_access_pointer(np->ipv6_fl_list)) |
f256dc59 YH |
278 | return; |
279 | ||
18367681 YH |
280 | spin_lock_bh(&ip6_sk_fl_lock); |
281 | while ((sfl = rcu_dereference_protected(np->ipv6_fl_list, | |
282 | lockdep_is_held(&ip6_sk_fl_lock))) != NULL) { | |
283 | np->ipv6_fl_list = sfl->next; | |
284 | spin_unlock_bh(&ip6_sk_fl_lock); | |
f256dc59 | 285 | |
1da177e4 | 286 | fl_release(sfl->fl); |
18367681 YH |
287 | kfree_rcu(sfl, rcu); |
288 | ||
289 | spin_lock_bh(&ip6_sk_fl_lock); | |
1da177e4 | 290 | } |
18367681 | 291 | spin_unlock_bh(&ip6_sk_fl_lock); |
1da177e4 LT |
292 | } |
293 | ||
294 | /* Service routines */ | |
295 | ||
296 | ||
297 | /* | |
298 | It is the only difficult place. flowlabel enforces equal headers | |
299 | before and including routing header, however user may supply options | |
300 | following rthdr. | |
301 | */ | |
302 | ||
303 | struct ipv6_txoptions *fl6_merge_options(struct ipv6_txoptions * opt_space, | |
304 | struct ip6_flowlabel * fl, | |
305 | struct ipv6_txoptions * fopt) | |
306 | { | |
df9890c3 | 307 | struct ipv6_txoptions * fl_opt = fl->opt; |
1ab1457c | 308 | |
df9890c3 YH |
309 | if (fopt == NULL || fopt->opt_flen == 0) |
310 | return fl_opt; | |
1ab1457c | 311 | |
1da177e4 LT |
312 | if (fl_opt != NULL) { |
313 | opt_space->hopopt = fl_opt->hopopt; | |
df9890c3 | 314 | opt_space->dst0opt = fl_opt->dst0opt; |
1da177e4 LT |
315 | opt_space->srcrt = fl_opt->srcrt; |
316 | opt_space->opt_nflen = fl_opt->opt_nflen; | |
317 | } else { | |
318 | if (fopt->opt_nflen == 0) | |
319 | return fopt; | |
320 | opt_space->hopopt = NULL; | |
321 | opt_space->dst0opt = NULL; | |
322 | opt_space->srcrt = NULL; | |
323 | opt_space->opt_nflen = 0; | |
324 | } | |
325 | opt_space->dst1opt = fopt->dst1opt; | |
1da177e4 LT |
326 | opt_space->opt_flen = fopt->opt_flen; |
327 | return opt_space; | |
328 | } | |
a495f836 | 329 | EXPORT_SYMBOL_GPL(fl6_merge_options); |
1da177e4 LT |
330 | |
331 | static unsigned long check_linger(unsigned long ttl) | |
332 | { | |
333 | if (ttl < FL_MIN_LINGER) | |
334 | return FL_MIN_LINGER*HZ; | |
335 | if (ttl > FL_MAX_LINGER && !capable(CAP_NET_ADMIN)) | |
336 | return 0; | |
337 | return ttl*HZ; | |
338 | } | |
339 | ||
340 | static int fl6_renew(struct ip6_flowlabel *fl, unsigned long linger, unsigned long expires) | |
341 | { | |
342 | linger = check_linger(linger); | |
343 | if (!linger) | |
344 | return -EPERM; | |
345 | expires = check_linger(expires); | |
346 | if (!expires) | |
347 | return -EPERM; | |
348 | fl->lastuse = jiffies; | |
349 | if (time_before(fl->linger, linger)) | |
350 | fl->linger = linger; | |
351 | if (time_before(expires, fl->linger)) | |
352 | expires = fl->linger; | |
353 | if (time_before(fl->expires, fl->lastuse + expires)) | |
354 | fl->expires = fl->lastuse + expires; | |
355 | return 0; | |
356 | } | |
357 | ||
358 | static struct ip6_flowlabel * | |
ec0506db MŻ |
359 | fl_create(struct net *net, struct sock *sk, struct in6_flowlabel_req *freq, |
360 | char __user *optval, int optlen, int *err_p) | |
1da177e4 | 361 | { |
684de409 | 362 | struct ip6_flowlabel *fl = NULL; |
1da177e4 LT |
363 | int olen; |
364 | int addr_type; | |
365 | int err; | |
366 | ||
684de409 DM |
367 | olen = optlen - CMSG_ALIGN(sizeof(*freq)); |
368 | err = -EINVAL; | |
369 | if (olen > 64 * 1024) | |
370 | goto done; | |
371 | ||
1da177e4 | 372 | err = -ENOMEM; |
0c600eda | 373 | fl = kzalloc(sizeof(*fl), GFP_KERNEL); |
1da177e4 LT |
374 | if (fl == NULL) |
375 | goto done; | |
1da177e4 | 376 | |
1da177e4 LT |
377 | if (olen > 0) { |
378 | struct msghdr msg; | |
4c9483b2 | 379 | struct flowi6 flowi6; |
1da177e4 LT |
380 | int junk; |
381 | ||
382 | err = -ENOMEM; | |
383 | fl->opt = kmalloc(sizeof(*fl->opt) + olen, GFP_KERNEL); | |
384 | if (fl->opt == NULL) | |
385 | goto done; | |
386 | ||
387 | memset(fl->opt, 0, sizeof(*fl->opt)); | |
388 | fl->opt->tot_len = sizeof(*fl->opt) + olen; | |
389 | err = -EFAULT; | |
390 | if (copy_from_user(fl->opt+1, optval+CMSG_ALIGN(sizeof(*freq)), olen)) | |
391 | goto done; | |
392 | ||
393 | msg.msg_controllen = olen; | |
394 | msg.msg_control = (void*)(fl->opt+1); | |
4c9483b2 | 395 | memset(&flowi6, 0, sizeof(flowi6)); |
1da177e4 | 396 | |
73df66f8 TP |
397 | err = ip6_datagram_send_ctl(net, sk, &msg, &flowi6, fl->opt, |
398 | &junk, &junk, &junk); | |
1da177e4 LT |
399 | if (err) |
400 | goto done; | |
401 | err = -EINVAL; | |
402 | if (fl->opt->opt_flen) | |
403 | goto done; | |
404 | if (fl->opt->opt_nflen == 0) { | |
405 | kfree(fl->opt); | |
406 | fl->opt = NULL; | |
407 | } | |
408 | } | |
409 | ||
60e8fbc4 | 410 | fl->fl_net = hold_net(net); |
1da177e4 LT |
411 | fl->expires = jiffies; |
412 | err = fl6_renew(fl, freq->flr_linger, freq->flr_expires); | |
413 | if (err) | |
414 | goto done; | |
415 | fl->share = freq->flr_share; | |
416 | addr_type = ipv6_addr_type(&freq->flr_dst); | |
35700212 JP |
417 | if ((addr_type & IPV6_ADDR_MAPPED) || |
418 | addr_type == IPV6_ADDR_ANY) { | |
c6817e4c | 419 | err = -EINVAL; |
1da177e4 | 420 | goto done; |
c6817e4c | 421 | } |
4e3fd7a0 | 422 | fl->dst = freq->flr_dst; |
1da177e4 LT |
423 | atomic_set(&fl->users, 1); |
424 | switch (fl->share) { | |
425 | case IPV6_FL_S_EXCL: | |
426 | case IPV6_FL_S_ANY: | |
427 | break; | |
428 | case IPV6_FL_S_PROCESS: | |
4f82f457 | 429 | fl->owner.pid = get_task_pid(current, PIDTYPE_PID); |
1da177e4 LT |
430 | break; |
431 | case IPV6_FL_S_USER: | |
4f82f457 | 432 | fl->owner.uid = current_euid(); |
1da177e4 LT |
433 | break; |
434 | default: | |
435 | err = -EINVAL; | |
436 | goto done; | |
437 | } | |
438 | return fl; | |
439 | ||
440 | done: | |
441 | fl_free(fl); | |
442 | *err_p = err; | |
443 | return NULL; | |
444 | } | |
445 | ||
446 | static int mem_check(struct sock *sk) | |
447 | { | |
448 | struct ipv6_pinfo *np = inet6_sk(sk); | |
449 | struct ipv6_fl_socklist *sfl; | |
450 | int room = FL_MAX_SIZE - atomic_read(&fl_size); | |
451 | int count = 0; | |
452 | ||
453 | if (room > FL_MAX_SIZE - FL_MAX_PER_SOCK) | |
454 | return 0; | |
455 | ||
18367681 | 456 | for_each_sk_fl_rcu(np, sfl) |
1da177e4 LT |
457 | count++; |
458 | ||
459 | if (room <= 0 || | |
460 | ((count >= FL_MAX_PER_SOCK || | |
35700212 JP |
461 | (count > 0 && room < FL_MAX_SIZE/2) || room < FL_MAX_SIZE/4) && |
462 | !capable(CAP_NET_ADMIN))) | |
1da177e4 LT |
463 | return -ENOBUFS; |
464 | ||
465 | return 0; | |
466 | } | |
467 | ||
04028045 PE |
468 | static inline void fl_link(struct ipv6_pinfo *np, struct ipv6_fl_socklist *sfl, |
469 | struct ip6_flowlabel *fl) | |
470 | { | |
18367681 | 471 | spin_lock_bh(&ip6_sk_fl_lock); |
04028045 PE |
472 | sfl->fl = fl; |
473 | sfl->next = np->ipv6_fl_list; | |
18367681 YH |
474 | rcu_assign_pointer(np->ipv6_fl_list, sfl); |
475 | spin_unlock_bh(&ip6_sk_fl_lock); | |
04028045 PE |
476 | } |
477 | ||
3fdfa5ff FF |
478 | int ipv6_flowlabel_opt_get(struct sock *sk, struct in6_flowlabel_req *freq) |
479 | { | |
480 | struct ipv6_pinfo *np = inet6_sk(sk); | |
481 | struct ipv6_fl_socklist *sfl; | |
482 | ||
483 | rcu_read_lock_bh(); | |
484 | ||
485 | for_each_sk_fl_rcu(np, sfl) { | |
486 | if (sfl->fl->label == (np->flow_label & IPV6_FLOWLABEL_MASK)) { | |
487 | spin_lock_bh(&ip6_fl_lock); | |
488 | freq->flr_label = sfl->fl->label; | |
489 | freq->flr_dst = sfl->fl->dst; | |
490 | freq->flr_share = sfl->fl->share; | |
491 | freq->flr_expires = (sfl->fl->expires - jiffies) / HZ; | |
492 | freq->flr_linger = sfl->fl->linger / HZ; | |
493 | ||
494 | spin_unlock_bh(&ip6_fl_lock); | |
495 | rcu_read_unlock_bh(); | |
496 | return 0; | |
497 | } | |
498 | } | |
499 | rcu_read_unlock_bh(); | |
500 | ||
501 | return -ENOENT; | |
502 | } | |
503 | ||
1da177e4 LT |
504 | int ipv6_flowlabel_opt(struct sock *sk, char __user *optval, int optlen) |
505 | { | |
55205d40 | 506 | int uninitialized_var(err); |
60e8fbc4 | 507 | struct net *net = sock_net(sk); |
1da177e4 LT |
508 | struct ipv6_pinfo *np = inet6_sk(sk); |
509 | struct in6_flowlabel_req freq; | |
510 | struct ipv6_fl_socklist *sfl1=NULL; | |
7f0e44ac ED |
511 | struct ipv6_fl_socklist *sfl; |
512 | struct ipv6_fl_socklist __rcu **sflp; | |
78c2e502 PE |
513 | struct ip6_flowlabel *fl, *fl1 = NULL; |
514 | ||
1da177e4 LT |
515 | |
516 | if (optlen < sizeof(freq)) | |
517 | return -EINVAL; | |
518 | ||
519 | if (copy_from_user(&freq, optval, sizeof(freq))) | |
520 | return -EFAULT; | |
521 | ||
522 | switch (freq.flr_action) { | |
523 | case IPV6_FL_A_PUT: | |
18367681 YH |
524 | spin_lock_bh(&ip6_sk_fl_lock); |
525 | for (sflp = &np->ipv6_fl_list; | |
526 | (sfl = rcu_dereference(*sflp))!=NULL; | |
527 | sflp = &sfl->next) { | |
1da177e4 LT |
528 | if (sfl->fl->label == freq.flr_label) { |
529 | if (freq.flr_label == (np->flow_label&IPV6_FLOWLABEL_MASK)) | |
530 | np->flow_label &= ~IPV6_FLOWLABEL_MASK; | |
18367681 YH |
531 | *sflp = rcu_dereference(sfl->next); |
532 | spin_unlock_bh(&ip6_sk_fl_lock); | |
1da177e4 | 533 | fl_release(sfl->fl); |
18367681 | 534 | kfree_rcu(sfl, rcu); |
1da177e4 LT |
535 | return 0; |
536 | } | |
537 | } | |
18367681 | 538 | spin_unlock_bh(&ip6_sk_fl_lock); |
1da177e4 LT |
539 | return -ESRCH; |
540 | ||
541 | case IPV6_FL_A_RENEW: | |
18367681 YH |
542 | rcu_read_lock_bh(); |
543 | for_each_sk_fl_rcu(np, sfl) { | |
1da177e4 LT |
544 | if (sfl->fl->label == freq.flr_label) { |
545 | err = fl6_renew(sfl->fl, freq.flr_linger, freq.flr_expires); | |
18367681 | 546 | rcu_read_unlock_bh(); |
1da177e4 LT |
547 | return err; |
548 | } | |
549 | } | |
18367681 | 550 | rcu_read_unlock_bh(); |
1da177e4 | 551 | |
af31f412 EB |
552 | if (freq.flr_share == IPV6_FL_S_NONE && |
553 | ns_capable(net->user_ns, CAP_NET_ADMIN)) { | |
60e8fbc4 | 554 | fl = fl_lookup(net, freq.flr_label); |
1da177e4 LT |
555 | if (fl) { |
556 | err = fl6_renew(fl, freq.flr_linger, freq.flr_expires); | |
557 | fl_release(fl); | |
558 | return err; | |
559 | } | |
560 | } | |
561 | return -ESRCH; | |
562 | ||
563 | case IPV6_FL_A_GET: | |
564 | if (freq.flr_label & ~IPV6_FLOWLABEL_MASK) | |
565 | return -EINVAL; | |
566 | ||
ec0506db | 567 | fl = fl_create(net, sk, &freq, optval, optlen, &err); |
1da177e4 LT |
568 | if (fl == NULL) |
569 | return err; | |
570 | sfl1 = kmalloc(sizeof(*sfl1), GFP_KERNEL); | |
571 | ||
572 | if (freq.flr_label) { | |
1da177e4 | 573 | err = -EEXIST; |
18367681 YH |
574 | rcu_read_lock_bh(); |
575 | for_each_sk_fl_rcu(np, sfl) { | |
1da177e4 LT |
576 | if (sfl->fl->label == freq.flr_label) { |
577 | if (freq.flr_flags&IPV6_FL_F_EXCL) { | |
18367681 | 578 | rcu_read_unlock_bh(); |
1da177e4 LT |
579 | goto done; |
580 | } | |
581 | fl1 = sfl->fl; | |
4ea6a804 | 582 | atomic_inc(&fl1->users); |
1da177e4 LT |
583 | break; |
584 | } | |
585 | } | |
18367681 | 586 | rcu_read_unlock_bh(); |
1da177e4 LT |
587 | |
588 | if (fl1 == NULL) | |
60e8fbc4 | 589 | fl1 = fl_lookup(net, freq.flr_label); |
1da177e4 | 590 | if (fl1) { |
78c2e502 | 591 | recheck: |
1da177e4 LT |
592 | err = -EEXIST; |
593 | if (freq.flr_flags&IPV6_FL_F_EXCL) | |
594 | goto release; | |
595 | err = -EPERM; | |
596 | if (fl1->share == IPV6_FL_S_EXCL || | |
597 | fl1->share != fl->share || | |
4f82f457 EB |
598 | ((fl1->share == IPV6_FL_S_PROCESS) && |
599 | (fl1->owner.pid == fl->owner.pid)) || | |
600 | ((fl1->share == IPV6_FL_S_USER) && | |
601 | uid_eq(fl1->owner.uid, fl->owner.uid))) | |
1da177e4 LT |
602 | goto release; |
603 | ||
1da177e4 LT |
604 | err = -ENOMEM; |
605 | if (sfl1 == NULL) | |
606 | goto release; | |
607 | if (fl->linger > fl1->linger) | |
608 | fl1->linger = fl->linger; | |
609 | if ((long)(fl->expires - fl1->expires) > 0) | |
610 | fl1->expires = fl->expires; | |
04028045 | 611 | fl_link(np, sfl1, fl1); |
1da177e4 LT |
612 | fl_free(fl); |
613 | return 0; | |
614 | ||
615 | release: | |
616 | fl_release(fl1); | |
617 | goto done; | |
618 | } | |
619 | } | |
620 | err = -ENOENT; | |
621 | if (!(freq.flr_flags&IPV6_FL_F_CREATE)) | |
622 | goto done; | |
623 | ||
624 | err = -ENOMEM; | |
625 | if (sfl1 == NULL || (err = mem_check(sk)) != 0) | |
626 | goto done; | |
627 | ||
60e8fbc4 | 628 | fl1 = fl_intern(net, fl, freq.flr_label); |
78c2e502 PE |
629 | if (fl1 != NULL) |
630 | goto recheck; | |
1da177e4 | 631 | |
6c94d361 DM |
632 | if (!freq.flr_label) { |
633 | if (copy_to_user(&((struct in6_flowlabel_req __user *) optval)->flr_label, | |
634 | &fl->label, sizeof(fl->label))) { | |
635 | /* Intentionally ignore fault. */ | |
636 | } | |
637 | } | |
1da177e4 | 638 | |
04028045 | 639 | fl_link(np, sfl1, fl); |
1da177e4 LT |
640 | return 0; |
641 | ||
642 | default: | |
643 | return -EINVAL; | |
644 | } | |
645 | ||
646 | done: | |
647 | fl_free(fl); | |
648 | kfree(sfl1); | |
649 | return err; | |
650 | } | |
651 | ||
652 | #ifdef CONFIG_PROC_FS | |
653 | ||
654 | struct ip6fl_iter_state { | |
5983a3df | 655 | struct seq_net_private p; |
4f82f457 | 656 | struct pid_namespace *pid_ns; |
1da177e4 LT |
657 | int bucket; |
658 | }; | |
659 | ||
660 | #define ip6fl_seq_private(seq) ((struct ip6fl_iter_state *)(seq)->private) | |
661 | ||
662 | static struct ip6_flowlabel *ip6fl_get_first(struct seq_file *seq) | |
663 | { | |
664 | struct ip6_flowlabel *fl = NULL; | |
665 | struct ip6fl_iter_state *state = ip6fl_seq_private(seq); | |
5983a3df | 666 | struct net *net = seq_file_net(seq); |
1da177e4 LT |
667 | |
668 | for (state->bucket = 0; state->bucket <= FL_HASH_MASK; ++state->bucket) { | |
d3aedd5e YH |
669 | for_each_fl_rcu(state->bucket, fl) { |
670 | if (net_eq(fl->fl_net, net)) | |
671 | goto out; | |
672 | } | |
1da177e4 | 673 | } |
d3aedd5e YH |
674 | fl = NULL; |
675 | out: | |
1da177e4 LT |
676 | return fl; |
677 | } | |
678 | ||
679 | static struct ip6_flowlabel *ip6fl_get_next(struct seq_file *seq, struct ip6_flowlabel *fl) | |
680 | { | |
681 | struct ip6fl_iter_state *state = ip6fl_seq_private(seq); | |
5983a3df | 682 | struct net *net = seq_file_net(seq); |
1da177e4 | 683 | |
d3aedd5e YH |
684 | for_each_fl_continue_rcu(fl) { |
685 | if (net_eq(fl->fl_net, net)) | |
686 | goto out; | |
687 | } | |
688 | ||
5983a3df | 689 | try_again: |
d3aedd5e YH |
690 | if (++state->bucket <= FL_HASH_MASK) { |
691 | for_each_fl_rcu(state->bucket, fl) { | |
692 | if (net_eq(fl->fl_net, net)) | |
693 | goto out; | |
694 | } | |
695 | goto try_again; | |
1da177e4 | 696 | } |
d3aedd5e YH |
697 | fl = NULL; |
698 | ||
699 | out: | |
1da177e4 LT |
700 | return fl; |
701 | } | |
702 | ||
703 | static struct ip6_flowlabel *ip6fl_get_idx(struct seq_file *seq, loff_t pos) | |
704 | { | |
705 | struct ip6_flowlabel *fl = ip6fl_get_first(seq); | |
706 | if (fl) | |
707 | while (pos && (fl = ip6fl_get_next(seq, fl)) != NULL) | |
708 | --pos; | |
709 | return pos ? NULL : fl; | |
710 | } | |
711 | ||
712 | static void *ip6fl_seq_start(struct seq_file *seq, loff_t *pos) | |
d3aedd5e | 713 | __acquires(RCU) |
1da177e4 | 714 | { |
d3aedd5e | 715 | rcu_read_lock_bh(); |
1da177e4 LT |
716 | return *pos ? ip6fl_get_idx(seq, *pos - 1) : SEQ_START_TOKEN; |
717 | } | |
718 | ||
719 | static void *ip6fl_seq_next(struct seq_file *seq, void *v, loff_t *pos) | |
720 | { | |
721 | struct ip6_flowlabel *fl; | |
722 | ||
723 | if (v == SEQ_START_TOKEN) | |
724 | fl = ip6fl_get_first(seq); | |
725 | else | |
726 | fl = ip6fl_get_next(seq, v); | |
727 | ++*pos; | |
728 | return fl; | |
729 | } | |
730 | ||
731 | static void ip6fl_seq_stop(struct seq_file *seq, void *v) | |
d3aedd5e | 732 | __releases(RCU) |
1da177e4 | 733 | { |
d3aedd5e | 734 | rcu_read_unlock_bh(); |
1da177e4 LT |
735 | } |
736 | ||
1b7c2dbc | 737 | static int ip6fl_seq_show(struct seq_file *seq, void *v) |
1da177e4 | 738 | { |
4f82f457 | 739 | struct ip6fl_iter_state *state = ip6fl_seq_private(seq); |
1b7c2dbc JM |
740 | if (v == SEQ_START_TOKEN) |
741 | seq_printf(seq, "%-5s %-1s %-6s %-6s %-6s %-8s %-32s %s\n", | |
742 | "Label", "S", "Owner", "Users", "Linger", "Expires", "Dst", "Opt"); | |
743 | else { | |
744 | struct ip6_flowlabel *fl = v; | |
1da177e4 | 745 | seq_printf(seq, |
4b7a4274 | 746 | "%05X %-1d %-6d %-6d %-6ld %-8ld %pi6 %-4d\n", |
95c96174 | 747 | (unsigned int)ntohl(fl->label), |
1da177e4 | 748 | fl->share, |
4f82f457 EB |
749 | ((fl->share == IPV6_FL_S_PROCESS) ? |
750 | pid_nr_ns(fl->owner.pid, state->pid_ns) : | |
751 | ((fl->share == IPV6_FL_S_USER) ? | |
752 | from_kuid_munged(seq_user_ns(seq), fl->owner.uid) : | |
753 | 0)), | |
1da177e4 LT |
754 | atomic_read(&fl->users), |
755 | fl->linger/HZ, | |
756 | (long)(fl->expires - jiffies)/HZ, | |
b071195d | 757 | &fl->dst, |
1da177e4 | 758 | fl->opt ? fl->opt->opt_nflen : 0); |
1da177e4 | 759 | } |
1da177e4 LT |
760 | return 0; |
761 | } | |
762 | ||
56b3d975 | 763 | static const struct seq_operations ip6fl_seq_ops = { |
1da177e4 LT |
764 | .start = ip6fl_seq_start, |
765 | .next = ip6fl_seq_next, | |
766 | .stop = ip6fl_seq_stop, | |
767 | .show = ip6fl_seq_show, | |
768 | }; | |
769 | ||
770 | static int ip6fl_seq_open(struct inode *inode, struct file *file) | |
771 | { | |
4f82f457 EB |
772 | struct seq_file *seq; |
773 | struct ip6fl_iter_state *state; | |
774 | int err; | |
775 | ||
776 | err = seq_open_net(inode, file, &ip6fl_seq_ops, | |
777 | sizeof(struct ip6fl_iter_state)); | |
778 | ||
779 | if (!err) { | |
780 | seq = file->private_data; | |
781 | state = ip6fl_seq_private(seq); | |
782 | rcu_read_lock(); | |
783 | state->pid_ns = get_pid_ns(task_active_pid_ns(current)); | |
784 | rcu_read_unlock(); | |
785 | } | |
786 | return err; | |
787 | } | |
788 | ||
789 | static int ip6fl_seq_release(struct inode *inode, struct file *file) | |
790 | { | |
791 | struct seq_file *seq = file->private_data; | |
792 | struct ip6fl_iter_state *state = ip6fl_seq_private(seq); | |
793 | put_pid_ns(state->pid_ns); | |
794 | return seq_release_net(inode, file); | |
1da177e4 LT |
795 | } |
796 | ||
9a32144e | 797 | static const struct file_operations ip6fl_seq_fops = { |
1da177e4 LT |
798 | .owner = THIS_MODULE, |
799 | .open = ip6fl_seq_open, | |
800 | .read = seq_read, | |
801 | .llseek = seq_lseek, | |
4f82f457 | 802 | .release = ip6fl_seq_release, |
1da177e4 | 803 | }; |
1da177e4 | 804 | |
2c8c1e72 | 805 | static int __net_init ip6_flowlabel_proc_init(struct net *net) |
0a3e78ac | 806 | { |
d4beaa66 G |
807 | if (!proc_create("ip6_flowlabel", S_IRUGO, net->proc_net, |
808 | &ip6fl_seq_fops)) | |
0a3e78ac DL |
809 | return -ENOMEM; |
810 | return 0; | |
811 | } | |
1da177e4 | 812 | |
2c8c1e72 | 813 | static void __net_exit ip6_flowlabel_proc_fini(struct net *net) |
1da177e4 | 814 | { |
ece31ffd | 815 | remove_proc_entry("ip6_flowlabel", net->proc_net); |
0a3e78ac DL |
816 | } |
817 | #else | |
818 | static inline int ip6_flowlabel_proc_init(struct net *net) | |
819 | { | |
820 | return 0; | |
821 | } | |
822 | static inline void ip6_flowlabel_proc_fini(struct net *net) | |
823 | { | |
0a3e78ac | 824 | } |
1da177e4 | 825 | #endif |
0a3e78ac | 826 | |
2c8c1e72 | 827 | static void __net_exit ip6_flowlabel_net_exit(struct net *net) |
60e8fbc4 BT |
828 | { |
829 | ip6_fl_purge(net); | |
5983a3df | 830 | ip6_flowlabel_proc_fini(net); |
60e8fbc4 BT |
831 | } |
832 | ||
833 | static struct pernet_operations ip6_flowlabel_net_ops = { | |
5983a3df | 834 | .init = ip6_flowlabel_proc_init, |
60e8fbc4 BT |
835 | .exit = ip6_flowlabel_net_exit, |
836 | }; | |
837 | ||
0a3e78ac DL |
838 | int ip6_flowlabel_init(void) |
839 | { | |
5983a3df | 840 | return register_pernet_subsys(&ip6_flowlabel_net_ops); |
1da177e4 LT |
841 | } |
842 | ||
843 | void ip6_flowlabel_cleanup(void) | |
844 | { | |
845 | del_timer(&ip6_fl_gc_timer); | |
60e8fbc4 | 846 | unregister_pernet_subsys(&ip6_flowlabel_net_ops); |
1da177e4 | 847 | } |