blob: 0fd0768a17c6cc9ab5b627bbfd94bee9c03b9657 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * net/sched/ematch.c Extended Match API
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: Thomas Graf <tgraf@suug.ch>
10 *
11 * ==========================================================================
12 *
13 * An extended match (ematch) is a small classification tool not worth
14 * writing a full classifier for. Ematches can be interconnected to form
15 * a logic expression and get attached to classifiers to extend their
16 * functionatlity.
17 *
18 * The userspace part transforms the logic expressions into an array
19 * consisting of multiple sequences of interconnected ematches separated
20 * by markers. Precedence is implemented by a special ematch kind
21 * referencing a sequence beyond the marker of the current sequence
22 * causing the current position in the sequence to be pushed onto a stack
23 * to allow the current position to be overwritten by the position referenced
24 * in the special ematch. Matching continues in the new sequence until a
25 * marker is reached causing the position to be restored from the stack.
26 *
27 * Example:
28 * A AND (B1 OR B2) AND C AND D
29 *
30 * ------->-PUSH-------
31 * -->-- / -->-- \ -->--
32 * / \ / / \ \ / \
33 * +-------+-------+-------+-------+-------+--------+
34 * | A AND | B AND | C AND | D END | B1 OR | B2 END |
35 * +-------+-------+-------+-------+-------+--------+
36 * \ /
37 * --------<-POP---------
38 *
39 * where B is a virtual ematch referencing to sequence starting with B1.
40 *
41 * ==========================================================================
42 *
43 * How to write an ematch in 60 seconds
44 * ------------------------------------
45 *
46 * 1) Provide a matcher function:
47 * static int my_match(struct sk_buff *skb, struct tcf_ematch *m,
48 * struct tcf_pkt_info *info)
49 * {
50 * struct mydata *d = (struct mydata *) m->data;
51 *
52 * if (...matching goes here...)
53 * return 1;
54 * else
55 * return 0;
56 * }
57 *
58 * 2) Fill out a struct tcf_ematch_ops:
59 * static struct tcf_ematch_ops my_ops = {
60 * .kind = unique id,
61 * .datalen = sizeof(struct mydata),
62 * .match = my_match,
63 * .owner = THIS_MODULE,
64 * };
65 *
66 * 3) Register/Unregister your ematch:
67 * static int __init init_my_ematch(void)
68 * {
69 * return tcf_em_register(&my_ops);
70 * }
71 *
72 * static void __exit exit_my_ematch(void)
73 * {
74 * return tcf_em_unregister(&my_ops);
75 * }
76 *
77 * module_init(init_my_ematch);
78 * module_exit(exit_my_ematch);
79 *
80 * 4) By now you should have two more seconds left, barely enough to
81 * open up a beer to watch the compilation going.
82 */
83
Linus Torvalds1da177e2005-04-16 15:20:36 -070084#include <linux/module.h>
85#include <linux/types.h>
86#include <linux/kernel.h>
87#include <linux/sched.h>
88#include <linux/mm.h>
89#include <linux/errno.h>
90#include <linux/interrupt.h>
91#include <linux/rtnetlink.h>
92#include <linux/skbuff.h>
93#include <net/pkt_cls.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070094
95static LIST_HEAD(ematch_ops);
96static DEFINE_RWLOCK(ematch_mod_lock);
97
98static inline struct tcf_ematch_ops * tcf_em_lookup(u16 kind)
99{
100 struct tcf_ematch_ops *e = NULL;
101
102 read_lock(&ematch_mod_lock);
103 list_for_each_entry(e, &ematch_ops, link) {
104 if (kind == e->kind) {
105 if (!try_module_get(e->owner))
106 e = NULL;
107 read_unlock(&ematch_mod_lock);
108 return e;
109 }
110 }
111 read_unlock(&ematch_mod_lock);
112
113 return NULL;
114}
115
116/**
117 * tcf_em_register - register an extended match
118 *
119 * @ops: ematch operations lookup table
120 *
121 * This function must be called by ematches to announce their presence.
122 * The given @ops must have kind set to a unique identifier and the
123 * callback match() must be implemented. All other callbacks are optional
124 * and a fallback implementation is used instead.
125 *
126 * Returns -EEXISTS if an ematch of the same kind has already registered.
127 */
128int tcf_em_register(struct tcf_ematch_ops *ops)
129{
130 int err = -EEXIST;
131 struct tcf_ematch_ops *e;
132
133 if (ops->match == NULL)
134 return -EINVAL;
135
136 write_lock(&ematch_mod_lock);
137 list_for_each_entry(e, &ematch_ops, link)
138 if (ops->kind == e->kind)
139 goto errout;
140
141 list_add_tail(&ops->link, &ematch_ops);
142 err = 0;
143errout:
144 write_unlock(&ematch_mod_lock);
145 return err;
146}
147
148/**
149 * tcf_em_unregister - unregster and extended match
150 *
151 * @ops: ematch operations lookup table
152 *
153 * This function must be called by ematches to announce their disappearance
154 * for examples when the module gets unloaded. The @ops parameter must be
155 * the same as the one used for registration.
156 *
157 * Returns -ENOENT if no matching ematch was found.
158 */
159int tcf_em_unregister(struct tcf_ematch_ops *ops)
160{
161 int err = 0;
162 struct tcf_ematch_ops *e;
163
164 write_lock(&ematch_mod_lock);
165 list_for_each_entry(e, &ematch_ops, link) {
166 if (e == ops) {
167 list_del(&e->link);
168 goto out;
169 }
170 }
171
172 err = -ENOENT;
173out:
174 write_unlock(&ematch_mod_lock);
175 return err;
176}
177
178static inline struct tcf_ematch * tcf_em_get_match(struct tcf_ematch_tree *tree,
179 int index)
180{
181 return &tree->matches[index];
182}
183
184
185static int tcf_em_validate(struct tcf_proto *tp,
186 struct tcf_ematch_tree_hdr *tree_hdr,
187 struct tcf_ematch *em, struct rtattr *rta, int idx)
188{
189 int err = -EINVAL;
190 struct tcf_ematch_hdr *em_hdr = RTA_DATA(rta);
191 int data_len = RTA_PAYLOAD(rta) - sizeof(*em_hdr);
192 void *data = (void *) em_hdr + sizeof(*em_hdr);
193
194 if (!TCF_EM_REL_VALID(em_hdr->flags))
195 goto errout;
196
197 if (em_hdr->kind == TCF_EM_CONTAINER) {
198 /* Special ematch called "container", carries an index
199 * referencing an external ematch sequence. */
200 u32 ref;
201
202 if (data_len < sizeof(ref))
203 goto errout;
204 ref = *(u32 *) data;
205
206 if (ref >= tree_hdr->nmatches)
207 goto errout;
208
209 /* We do not allow backward jumps to avoid loops and jumps
210 * to our own position are of course illegal. */
211 if (ref <= idx)
212 goto errout;
213
214
215 em->data = ref;
216 } else {
217 /* Note: This lookup will increase the module refcnt
218 * of the ematch module referenced. In case of a failure,
219 * a destroy function is called by the underlying layer
220 * which automatically releases the reference again, therefore
221 * the module MUST not be given back under any circumstances
222 * here. Be aware, the destroy function assumes that the
223 * module is held if the ops field is non zero. */
224 em->ops = tcf_em_lookup(em_hdr->kind);
225
226 if (em->ops == NULL) {
227 err = -ENOENT;
228 goto errout;
229 }
230
231 /* ematch module provides expected length of data, so we
232 * can do a basic sanity check. */
233 if (em->ops->datalen && data_len < em->ops->datalen)
234 goto errout;
235
236 if (em->ops->change) {
237 err = em->ops->change(tp, data, data_len, em);
238 if (err < 0)
239 goto errout;
240 } else if (data_len > 0) {
241 /* ematch module doesn't provide an own change
242 * procedure and expects us to allocate and copy
243 * the ematch data.
244 *
245 * TCF_EM_SIMPLE may be specified stating that the
246 * data only consists of a u32 integer and the module
247 * does not expected a memory reference but rather
248 * the value carried. */
249 if (em_hdr->flags & TCF_EM_SIMPLE) {
250 if (data_len < sizeof(u32))
251 goto errout;
252 em->data = *(u32 *) data;
253 } else {
254 void *v = kmalloc(data_len, GFP_KERNEL);
255 if (v == NULL) {
256 err = -ENOBUFS;
257 goto errout;
258 }
259 memcpy(v, data, data_len);
260 em->data = (unsigned long) v;
261 }
262 }
263 }
264
265 em->matchid = em_hdr->matchid;
266 em->flags = em_hdr->flags;
267 em->datalen = data_len;
268
269 err = 0;
270errout:
271 return err;
272}
273
274/**
275 * tcf_em_tree_validate - validate ematch config TLV and build ematch tree
276 *
277 * @tp: classifier kind handle
278 * @rta: ematch tree configuration TLV
279 * @tree: destination ematch tree variable to store the resulting
280 * ematch tree.
281 *
282 * This function validates the given configuration TLV @rta and builds an
283 * ematch tree in @tree. The resulting tree must later be copied into
284 * the private classifier data using tcf_em_tree_change(). You MUST NOT
285 * provide the ematch tree variable of the private classifier data directly,
286 * the changes would not be locked properly.
287 *
288 * Returns a negative error code if the configuration TLV contains errors.
289 */
290int tcf_em_tree_validate(struct tcf_proto *tp, struct rtattr *rta,
291 struct tcf_ematch_tree *tree)
292{
293 int idx, list_len, matches_len, err = -EINVAL;
294 struct rtattr *tb[TCA_EMATCH_TREE_MAX];
295 struct rtattr *rt_match, *rt_hdr, *rt_list;
296 struct tcf_ematch_tree_hdr *tree_hdr;
297 struct tcf_ematch *em;
298
Thomas Grafb541ca22005-11-08 09:39:17 -0800299 if (!rta) {
300 memset(tree, 0, sizeof(*tree));
301 return 0;
302 }
303
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 if (rtattr_parse_nested(tb, TCA_EMATCH_TREE_MAX, rta) < 0)
305 goto errout;
306
307 rt_hdr = tb[TCA_EMATCH_TREE_HDR-1];
308 rt_list = tb[TCA_EMATCH_TREE_LIST-1];
309
310 if (rt_hdr == NULL || rt_list == NULL)
311 goto errout;
312
313 if (RTA_PAYLOAD(rt_hdr) < sizeof(*tree_hdr) ||
314 RTA_PAYLOAD(rt_list) < sizeof(*rt_match))
315 goto errout;
316
317 tree_hdr = RTA_DATA(rt_hdr);
318 memcpy(&tree->hdr, tree_hdr, sizeof(*tree_hdr));
319
320 rt_match = RTA_DATA(rt_list);
321 list_len = RTA_PAYLOAD(rt_list);
322 matches_len = tree_hdr->nmatches * sizeof(*em);
323
Panagiotis Issaris0da974f2006-07-21 14:51:30 -0700324 tree->matches = kzalloc(matches_len, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 if (tree->matches == NULL)
326 goto errout;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327
328 /* We do not use rtattr_parse_nested here because the maximum
329 * number of attributes is unknown. This saves us the allocation
330 * for a tb buffer which would serve no purpose at all.
331 *
332 * The array of rt attributes is parsed in the order as they are
333 * provided, their type must be incremental from 1 to n. Even
334 * if it does not serve any real purpose, a failure of sticking
335 * to this policy will result in parsing failure. */
336 for (idx = 0; RTA_OK(rt_match, list_len); idx++) {
337 err = -EINVAL;
338
339 if (rt_match->rta_type != (idx + 1))
340 goto errout_abort;
341
342 if (idx >= tree_hdr->nmatches)
343 goto errout_abort;
344
345 if (RTA_PAYLOAD(rt_match) < sizeof(struct tcf_ematch_hdr))
346 goto errout_abort;
347
348 em = tcf_em_get_match(tree, idx);
349
350 err = tcf_em_validate(tp, tree_hdr, em, rt_match, idx);
351 if (err < 0)
352 goto errout_abort;
353
354 rt_match = RTA_NEXT(rt_match, list_len);
355 }
356
357 /* Check if the number of matches provided by userspace actually
358 * complies with the array of matches. The number was used for
359 * the validation of references and a mismatch could lead to
360 * undefined references during the matching process. */
361 if (idx != tree_hdr->nmatches) {
362 err = -EINVAL;
363 goto errout_abort;
364 }
365
366 err = 0;
367errout:
368 return err;
369
370errout_abort:
371 tcf_em_tree_destroy(tp, tree);
372 return err;
373}
374
375/**
376 * tcf_em_tree_destroy - destroy an ematch tree
377 *
378 * @tp: classifier kind handle
379 * @tree: ematch tree to be deleted
380 *
381 * This functions destroys an ematch tree previously created by
382 * tcf_em_tree_validate()/tcf_em_tree_change(). You must ensure that
383 * the ematch tree is not in use before calling this function.
384 */
385void tcf_em_tree_destroy(struct tcf_proto *tp, struct tcf_ematch_tree *tree)
386{
387 int i;
388
389 if (tree->matches == NULL)
390 return;
391
392 for (i = 0; i < tree->hdr.nmatches; i++) {
393 struct tcf_ematch *em = tcf_em_get_match(tree, i);
394
395 if (em->ops) {
396 if (em->ops->destroy)
397 em->ops->destroy(tp, em);
398 else if (!tcf_em_is_simple(em) && em->data)
399 kfree((void *) em->data);
400 module_put(em->ops->owner);
401 }
402 }
403
404 tree->hdr.nmatches = 0;
405 kfree(tree->matches);
406}
407
408/**
409 * tcf_em_tree_dump - dump ematch tree into a rtnl message
410 *
411 * @skb: skb holding the rtnl message
412 * @t: ematch tree to be dumped
413 * @tlv: TLV type to be used to encapsulate the tree
414 *
415 * This function dumps a ematch tree into a rtnl message. It is valid to
416 * call this function while the ematch tree is in use.
417 *
418 * Returns -1 if the skb tailroom is insufficient.
419 */
420int tcf_em_tree_dump(struct sk_buff *skb, struct tcf_ematch_tree *tree, int tlv)
421{
422 int i;
423 struct rtattr * top_start = (struct rtattr*) skb->tail;
424 struct rtattr * list_start;
425
426 RTA_PUT(skb, tlv, 0, NULL);
427 RTA_PUT(skb, TCA_EMATCH_TREE_HDR, sizeof(tree->hdr), &tree->hdr);
428
429 list_start = (struct rtattr *) skb->tail;
430 RTA_PUT(skb, TCA_EMATCH_TREE_LIST, 0, NULL);
431
432 for (i = 0; i < tree->hdr.nmatches; i++) {
433 struct rtattr *match_start = (struct rtattr*) skb->tail;
434 struct tcf_ematch *em = tcf_em_get_match(tree, i);
435 struct tcf_ematch_hdr em_hdr = {
436 .kind = em->ops ? em->ops->kind : TCF_EM_CONTAINER,
437 .matchid = em->matchid,
438 .flags = em->flags
439 };
440
441 RTA_PUT(skb, i+1, sizeof(em_hdr), &em_hdr);
442
443 if (em->ops && em->ops->dump) {
444 if (em->ops->dump(skb, em) < 0)
445 goto rtattr_failure;
446 } else if (tcf_em_is_container(em) || tcf_em_is_simple(em)) {
447 u32 u = em->data;
448 RTA_PUT_NOHDR(skb, sizeof(u), &u);
449 } else if (em->datalen > 0)
450 RTA_PUT_NOHDR(skb, em->datalen, (void *) em->data);
451
452 match_start->rta_len = skb->tail - (u8*) match_start;
453 }
454
455 list_start->rta_len = skb->tail - (u8 *) list_start;
456 top_start->rta_len = skb->tail - (u8 *) top_start;
457
458 return 0;
459
460rtattr_failure:
461 return -1;
462}
463
464static inline int tcf_em_match(struct sk_buff *skb, struct tcf_ematch *em,
465 struct tcf_pkt_info *info)
466{
467 int r = em->ops->match(skb, em, info);
468 return tcf_em_is_inverted(em) ? !r : r;
469}
470
471/* Do not use this function directly, use tcf_em_tree_match instead */
472int __tcf_em_tree_match(struct sk_buff *skb, struct tcf_ematch_tree *tree,
473 struct tcf_pkt_info *info)
474{
475 int stackp = 0, match_idx = 0, res = 0;
476 struct tcf_ematch *cur_match;
477 int stack[CONFIG_NET_EMATCH_STACK];
478
479proceed:
480 while (match_idx < tree->hdr.nmatches) {
481 cur_match = tcf_em_get_match(tree, match_idx);
482
483 if (tcf_em_is_container(cur_match)) {
484 if (unlikely(stackp >= CONFIG_NET_EMATCH_STACK))
485 goto stack_overflow;
486
487 stack[stackp++] = match_idx;
488 match_idx = cur_match->data;
489 goto proceed;
490 }
491
492 res = tcf_em_match(skb, cur_match, info);
493
494 if (tcf_em_early_end(cur_match, res))
495 break;
496
497 match_idx++;
498 }
499
500pop_stack:
501 if (stackp > 0) {
502 match_idx = stack[--stackp];
503 cur_match = tcf_em_get_match(tree, match_idx);
504
505 if (tcf_em_early_end(cur_match, res))
506 goto pop_stack;
507 else {
508 match_idx++;
509 goto proceed;
510 }
511 }
512
513 return res;
514
515stack_overflow:
516 if (net_ratelimit())
517 printk("Local stack overflow, increase NET_EMATCH_STACK\n");
518 return -1;
519}
520
521EXPORT_SYMBOL(tcf_em_register);
522EXPORT_SYMBOL(tcf_em_unregister);
523EXPORT_SYMBOL(tcf_em_tree_validate);
524EXPORT_SYMBOL(tcf_em_tree_destroy);
525EXPORT_SYMBOL(tcf_em_tree_dump);
526EXPORT_SYMBOL(__tcf_em_tree_match);