blob: 64b047c65568cafe0912b92f7bb18b7e56031c25 [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
84#include <linux/config.h>
85#include <linux/module.h>
86#include <linux/types.h>
87#include <linux/kernel.h>
88#include <linux/sched.h>
89#include <linux/mm.h>
90#include <linux/errno.h>
91#include <linux/interrupt.h>
92#include <linux/rtnetlink.h>
93#include <linux/skbuff.h>
94#include <net/pkt_cls.h>
95#include <config/net/ematch/stack.h>
96
97static LIST_HEAD(ematch_ops);
98static DEFINE_RWLOCK(ematch_mod_lock);
99
100static inline struct tcf_ematch_ops * tcf_em_lookup(u16 kind)
101{
102 struct tcf_ematch_ops *e = NULL;
103
104 read_lock(&ematch_mod_lock);
105 list_for_each_entry(e, &ematch_ops, link) {
106 if (kind == e->kind) {
107 if (!try_module_get(e->owner))
108 e = NULL;
109 read_unlock(&ematch_mod_lock);
110 return e;
111 }
112 }
113 read_unlock(&ematch_mod_lock);
114
115 return NULL;
116}
117
118/**
119 * tcf_em_register - register an extended match
120 *
121 * @ops: ematch operations lookup table
122 *
123 * This function must be called by ematches to announce their presence.
124 * The given @ops must have kind set to a unique identifier and the
125 * callback match() must be implemented. All other callbacks are optional
126 * and a fallback implementation is used instead.
127 *
128 * Returns -EEXISTS if an ematch of the same kind has already registered.
129 */
130int tcf_em_register(struct tcf_ematch_ops *ops)
131{
132 int err = -EEXIST;
133 struct tcf_ematch_ops *e;
134
135 if (ops->match == NULL)
136 return -EINVAL;
137
138 write_lock(&ematch_mod_lock);
139 list_for_each_entry(e, &ematch_ops, link)
140 if (ops->kind == e->kind)
141 goto errout;
142
143 list_add_tail(&ops->link, &ematch_ops);
144 err = 0;
145errout:
146 write_unlock(&ematch_mod_lock);
147 return err;
148}
149
150/**
151 * tcf_em_unregister - unregster and extended match
152 *
153 * @ops: ematch operations lookup table
154 *
155 * This function must be called by ematches to announce their disappearance
156 * for examples when the module gets unloaded. The @ops parameter must be
157 * the same as the one used for registration.
158 *
159 * Returns -ENOENT if no matching ematch was found.
160 */
161int tcf_em_unregister(struct tcf_ematch_ops *ops)
162{
163 int err = 0;
164 struct tcf_ematch_ops *e;
165
166 write_lock(&ematch_mod_lock);
167 list_for_each_entry(e, &ematch_ops, link) {
168 if (e == ops) {
169 list_del(&e->link);
170 goto out;
171 }
172 }
173
174 err = -ENOENT;
175out:
176 write_unlock(&ematch_mod_lock);
177 return err;
178}
179
180static inline struct tcf_ematch * tcf_em_get_match(struct tcf_ematch_tree *tree,
181 int index)
182{
183 return &tree->matches[index];
184}
185
186
187static int tcf_em_validate(struct tcf_proto *tp,
188 struct tcf_ematch_tree_hdr *tree_hdr,
189 struct tcf_ematch *em, struct rtattr *rta, int idx)
190{
191 int err = -EINVAL;
192 struct tcf_ematch_hdr *em_hdr = RTA_DATA(rta);
193 int data_len = RTA_PAYLOAD(rta) - sizeof(*em_hdr);
194 void *data = (void *) em_hdr + sizeof(*em_hdr);
195
196 if (!TCF_EM_REL_VALID(em_hdr->flags))
197 goto errout;
198
199 if (em_hdr->kind == TCF_EM_CONTAINER) {
200 /* Special ematch called "container", carries an index
201 * referencing an external ematch sequence. */
202 u32 ref;
203
204 if (data_len < sizeof(ref))
205 goto errout;
206 ref = *(u32 *) data;
207
208 if (ref >= tree_hdr->nmatches)
209 goto errout;
210
211 /* We do not allow backward jumps to avoid loops and jumps
212 * to our own position are of course illegal. */
213 if (ref <= idx)
214 goto errout;
215
216
217 em->data = ref;
218 } else {
219 /* Note: This lookup will increase the module refcnt
220 * of the ematch module referenced. In case of a failure,
221 * a destroy function is called by the underlying layer
222 * which automatically releases the reference again, therefore
223 * the module MUST not be given back under any circumstances
224 * here. Be aware, the destroy function assumes that the
225 * module is held if the ops field is non zero. */
226 em->ops = tcf_em_lookup(em_hdr->kind);
227
228 if (em->ops == NULL) {
229 err = -ENOENT;
230 goto errout;
231 }
232
233 /* ematch module provides expected length of data, so we
234 * can do a basic sanity check. */
235 if (em->ops->datalen && data_len < em->ops->datalen)
236 goto errout;
237
238 if (em->ops->change) {
239 err = em->ops->change(tp, data, data_len, em);
240 if (err < 0)
241 goto errout;
242 } else if (data_len > 0) {
243 /* ematch module doesn't provide an own change
244 * procedure and expects us to allocate and copy
245 * the ematch data.
246 *
247 * TCF_EM_SIMPLE may be specified stating that the
248 * data only consists of a u32 integer and the module
249 * does not expected a memory reference but rather
250 * the value carried. */
251 if (em_hdr->flags & TCF_EM_SIMPLE) {
252 if (data_len < sizeof(u32))
253 goto errout;
254 em->data = *(u32 *) data;
255 } else {
256 void *v = kmalloc(data_len, GFP_KERNEL);
257 if (v == NULL) {
258 err = -ENOBUFS;
259 goto errout;
260 }
261 memcpy(v, data, data_len);
262 em->data = (unsigned long) v;
263 }
264 }
265 }
266
267 em->matchid = em_hdr->matchid;
268 em->flags = em_hdr->flags;
269 em->datalen = data_len;
270
271 err = 0;
272errout:
273 return err;
274}
275
276/**
277 * tcf_em_tree_validate - validate ematch config TLV and build ematch tree
278 *
279 * @tp: classifier kind handle
280 * @rta: ematch tree configuration TLV
281 * @tree: destination ematch tree variable to store the resulting
282 * ematch tree.
283 *
284 * This function validates the given configuration TLV @rta and builds an
285 * ematch tree in @tree. The resulting tree must later be copied into
286 * the private classifier data using tcf_em_tree_change(). You MUST NOT
287 * provide the ematch tree variable of the private classifier data directly,
288 * the changes would not be locked properly.
289 *
290 * Returns a negative error code if the configuration TLV contains errors.
291 */
292int tcf_em_tree_validate(struct tcf_proto *tp, struct rtattr *rta,
293 struct tcf_ematch_tree *tree)
294{
295 int idx, list_len, matches_len, err = -EINVAL;
296 struct rtattr *tb[TCA_EMATCH_TREE_MAX];
297 struct rtattr *rt_match, *rt_hdr, *rt_list;
298 struct tcf_ematch_tree_hdr *tree_hdr;
299 struct tcf_ematch *em;
300
Thomas Grafb541ca22005-11-08 09:39:17 -0800301 if (!rta) {
302 memset(tree, 0, sizeof(*tree));
303 return 0;
304 }
305
Linus Torvalds1da177e2005-04-16 15:20:36 -0700306 if (rtattr_parse_nested(tb, TCA_EMATCH_TREE_MAX, rta) < 0)
307 goto errout;
308
309 rt_hdr = tb[TCA_EMATCH_TREE_HDR-1];
310 rt_list = tb[TCA_EMATCH_TREE_LIST-1];
311
312 if (rt_hdr == NULL || rt_list == NULL)
313 goto errout;
314
315 if (RTA_PAYLOAD(rt_hdr) < sizeof(*tree_hdr) ||
316 RTA_PAYLOAD(rt_list) < sizeof(*rt_match))
317 goto errout;
318
319 tree_hdr = RTA_DATA(rt_hdr);
320 memcpy(&tree->hdr, tree_hdr, sizeof(*tree_hdr));
321
322 rt_match = RTA_DATA(rt_list);
323 list_len = RTA_PAYLOAD(rt_list);
324 matches_len = tree_hdr->nmatches * sizeof(*em);
325
326 tree->matches = kmalloc(matches_len, GFP_KERNEL);
327 if (tree->matches == NULL)
328 goto errout;
329 memset(tree->matches, 0, matches_len);
330
331 /* We do not use rtattr_parse_nested here because the maximum
332 * number of attributes is unknown. This saves us the allocation
333 * for a tb buffer which would serve no purpose at all.
334 *
335 * The array of rt attributes is parsed in the order as they are
336 * provided, their type must be incremental from 1 to n. Even
337 * if it does not serve any real purpose, a failure of sticking
338 * to this policy will result in parsing failure. */
339 for (idx = 0; RTA_OK(rt_match, list_len); idx++) {
340 err = -EINVAL;
341
342 if (rt_match->rta_type != (idx + 1))
343 goto errout_abort;
344
345 if (idx >= tree_hdr->nmatches)
346 goto errout_abort;
347
348 if (RTA_PAYLOAD(rt_match) < sizeof(struct tcf_ematch_hdr))
349 goto errout_abort;
350
351 em = tcf_em_get_match(tree, idx);
352
353 err = tcf_em_validate(tp, tree_hdr, em, rt_match, idx);
354 if (err < 0)
355 goto errout_abort;
356
357 rt_match = RTA_NEXT(rt_match, list_len);
358 }
359
360 /* Check if the number of matches provided by userspace actually
361 * complies with the array of matches. The number was used for
362 * the validation of references and a mismatch could lead to
363 * undefined references during the matching process. */
364 if (idx != tree_hdr->nmatches) {
365 err = -EINVAL;
366 goto errout_abort;
367 }
368
369 err = 0;
370errout:
371 return err;
372
373errout_abort:
374 tcf_em_tree_destroy(tp, tree);
375 return err;
376}
377
378/**
379 * tcf_em_tree_destroy - destroy an ematch tree
380 *
381 * @tp: classifier kind handle
382 * @tree: ematch tree to be deleted
383 *
384 * This functions destroys an ematch tree previously created by
385 * tcf_em_tree_validate()/tcf_em_tree_change(). You must ensure that
386 * the ematch tree is not in use before calling this function.
387 */
388void tcf_em_tree_destroy(struct tcf_proto *tp, struct tcf_ematch_tree *tree)
389{
390 int i;
391
392 if (tree->matches == NULL)
393 return;
394
395 for (i = 0; i < tree->hdr.nmatches; i++) {
396 struct tcf_ematch *em = tcf_em_get_match(tree, i);
397
398 if (em->ops) {
399 if (em->ops->destroy)
400 em->ops->destroy(tp, em);
401 else if (!tcf_em_is_simple(em) && em->data)
402 kfree((void *) em->data);
403 module_put(em->ops->owner);
404 }
405 }
406
407 tree->hdr.nmatches = 0;
408 kfree(tree->matches);
409}
410
411/**
412 * tcf_em_tree_dump - dump ematch tree into a rtnl message
413 *
414 * @skb: skb holding the rtnl message
415 * @t: ematch tree to be dumped
416 * @tlv: TLV type to be used to encapsulate the tree
417 *
418 * This function dumps a ematch tree into a rtnl message. It is valid to
419 * call this function while the ematch tree is in use.
420 *
421 * Returns -1 if the skb tailroom is insufficient.
422 */
423int tcf_em_tree_dump(struct sk_buff *skb, struct tcf_ematch_tree *tree, int tlv)
424{
425 int i;
426 struct rtattr * top_start = (struct rtattr*) skb->tail;
427 struct rtattr * list_start;
428
429 RTA_PUT(skb, tlv, 0, NULL);
430 RTA_PUT(skb, TCA_EMATCH_TREE_HDR, sizeof(tree->hdr), &tree->hdr);
431
432 list_start = (struct rtattr *) skb->tail;
433 RTA_PUT(skb, TCA_EMATCH_TREE_LIST, 0, NULL);
434
435 for (i = 0; i < tree->hdr.nmatches; i++) {
436 struct rtattr *match_start = (struct rtattr*) skb->tail;
437 struct tcf_ematch *em = tcf_em_get_match(tree, i);
438 struct tcf_ematch_hdr em_hdr = {
439 .kind = em->ops ? em->ops->kind : TCF_EM_CONTAINER,
440 .matchid = em->matchid,
441 .flags = em->flags
442 };
443
444 RTA_PUT(skb, i+1, sizeof(em_hdr), &em_hdr);
445
446 if (em->ops && em->ops->dump) {
447 if (em->ops->dump(skb, em) < 0)
448 goto rtattr_failure;
449 } else if (tcf_em_is_container(em) || tcf_em_is_simple(em)) {
450 u32 u = em->data;
451 RTA_PUT_NOHDR(skb, sizeof(u), &u);
452 } else if (em->datalen > 0)
453 RTA_PUT_NOHDR(skb, em->datalen, (void *) em->data);
454
455 match_start->rta_len = skb->tail - (u8*) match_start;
456 }
457
458 list_start->rta_len = skb->tail - (u8 *) list_start;
459 top_start->rta_len = skb->tail - (u8 *) top_start;
460
461 return 0;
462
463rtattr_failure:
464 return -1;
465}
466
467static inline int tcf_em_match(struct sk_buff *skb, struct tcf_ematch *em,
468 struct tcf_pkt_info *info)
469{
470 int r = em->ops->match(skb, em, info);
471 return tcf_em_is_inverted(em) ? !r : r;
472}
473
474/* Do not use this function directly, use tcf_em_tree_match instead */
475int __tcf_em_tree_match(struct sk_buff *skb, struct tcf_ematch_tree *tree,
476 struct tcf_pkt_info *info)
477{
478 int stackp = 0, match_idx = 0, res = 0;
479 struct tcf_ematch *cur_match;
480 int stack[CONFIG_NET_EMATCH_STACK];
481
482proceed:
483 while (match_idx < tree->hdr.nmatches) {
484 cur_match = tcf_em_get_match(tree, match_idx);
485
486 if (tcf_em_is_container(cur_match)) {
487 if (unlikely(stackp >= CONFIG_NET_EMATCH_STACK))
488 goto stack_overflow;
489
490 stack[stackp++] = match_idx;
491 match_idx = cur_match->data;
492 goto proceed;
493 }
494
495 res = tcf_em_match(skb, cur_match, info);
496
497 if (tcf_em_early_end(cur_match, res))
498 break;
499
500 match_idx++;
501 }
502
503pop_stack:
504 if (stackp > 0) {
505 match_idx = stack[--stackp];
506 cur_match = tcf_em_get_match(tree, match_idx);
507
508 if (tcf_em_early_end(cur_match, res))
509 goto pop_stack;
510 else {
511 match_idx++;
512 goto proceed;
513 }
514 }
515
516 return res;
517
518stack_overflow:
519 if (net_ratelimit())
520 printk("Local stack overflow, increase NET_EMATCH_STACK\n");
521 return -1;
522}
523
524EXPORT_SYMBOL(tcf_em_register);
525EXPORT_SYMBOL(tcf_em_unregister);
526EXPORT_SYMBOL(tcf_em_tree_validate);
527EXPORT_SYMBOL(tcf_em_tree_destroy);
528EXPORT_SYMBOL(tcf_em_tree_dump);
529EXPORT_SYMBOL(__tcf_em_tree_match);