blob: 5cb956b721e8e05a452262decdc114061cb0a0a0 [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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070095
96static LIST_HEAD(ematch_ops);
97static DEFINE_RWLOCK(ematch_mod_lock);
98
99static inline struct tcf_ematch_ops * tcf_em_lookup(u16 kind)
100{
101 struct tcf_ematch_ops *e = NULL;
102
103 read_lock(&ematch_mod_lock);
104 list_for_each_entry(e, &ematch_ops, link) {
105 if (kind == e->kind) {
106 if (!try_module_get(e->owner))
107 e = NULL;
108 read_unlock(&ematch_mod_lock);
109 return e;
110 }
111 }
112 read_unlock(&ematch_mod_lock);
113
114 return NULL;
115}
116
117/**
118 * tcf_em_register - register an extended match
119 *
120 * @ops: ematch operations lookup table
121 *
122 * This function must be called by ematches to announce their presence.
123 * The given @ops must have kind set to a unique identifier and the
124 * callback match() must be implemented. All other callbacks are optional
125 * and a fallback implementation is used instead.
126 *
127 * Returns -EEXISTS if an ematch of the same kind has already registered.
128 */
129int tcf_em_register(struct tcf_ematch_ops *ops)
130{
131 int err = -EEXIST;
132 struct tcf_ematch_ops *e;
133
134 if (ops->match == NULL)
135 return -EINVAL;
136
137 write_lock(&ematch_mod_lock);
138 list_for_each_entry(e, &ematch_ops, link)
139 if (ops->kind == e->kind)
140 goto errout;
141
142 list_add_tail(&ops->link, &ematch_ops);
143 err = 0;
144errout:
145 write_unlock(&ematch_mod_lock);
146 return err;
147}
148
149/**
150 * tcf_em_unregister - unregster and extended match
151 *
152 * @ops: ematch operations lookup table
153 *
154 * This function must be called by ematches to announce their disappearance
155 * for examples when the module gets unloaded. The @ops parameter must be
156 * the same as the one used for registration.
157 *
158 * Returns -ENOENT if no matching ematch was found.
159 */
160int tcf_em_unregister(struct tcf_ematch_ops *ops)
161{
162 int err = 0;
163 struct tcf_ematch_ops *e;
164
165 write_lock(&ematch_mod_lock);
166 list_for_each_entry(e, &ematch_ops, link) {
167 if (e == ops) {
168 list_del(&e->link);
169 goto out;
170 }
171 }
172
173 err = -ENOENT;
174out:
175 write_unlock(&ematch_mod_lock);
176 return err;
177}
178
179static inline struct tcf_ematch * tcf_em_get_match(struct tcf_ematch_tree *tree,
180 int index)
181{
182 return &tree->matches[index];
183}
184
185
186static int tcf_em_validate(struct tcf_proto *tp,
187 struct tcf_ematch_tree_hdr *tree_hdr,
188 struct tcf_ematch *em, struct rtattr *rta, int idx)
189{
190 int err = -EINVAL;
191 struct tcf_ematch_hdr *em_hdr = RTA_DATA(rta);
192 int data_len = RTA_PAYLOAD(rta) - sizeof(*em_hdr);
193 void *data = (void *) em_hdr + sizeof(*em_hdr);
194
195 if (!TCF_EM_REL_VALID(em_hdr->flags))
196 goto errout;
197
198 if (em_hdr->kind == TCF_EM_CONTAINER) {
199 /* Special ematch called "container", carries an index
200 * referencing an external ematch sequence. */
201 u32 ref;
202
203 if (data_len < sizeof(ref))
204 goto errout;
205 ref = *(u32 *) data;
206
207 if (ref >= tree_hdr->nmatches)
208 goto errout;
209
210 /* We do not allow backward jumps to avoid loops and jumps
211 * to our own position are of course illegal. */
212 if (ref <= idx)
213 goto errout;
214
215
216 em->data = ref;
217 } else {
218 /* Note: This lookup will increase the module refcnt
219 * of the ematch module referenced. In case of a failure,
220 * a destroy function is called by the underlying layer
221 * which automatically releases the reference again, therefore
222 * the module MUST not be given back under any circumstances
223 * here. Be aware, the destroy function assumes that the
224 * module is held if the ops field is non zero. */
225 em->ops = tcf_em_lookup(em_hdr->kind);
226
227 if (em->ops == NULL) {
228 err = -ENOENT;
229 goto errout;
230 }
231
232 /* ematch module provides expected length of data, so we
233 * can do a basic sanity check. */
234 if (em->ops->datalen && data_len < em->ops->datalen)
235 goto errout;
236
237 if (em->ops->change) {
238 err = em->ops->change(tp, data, data_len, em);
239 if (err < 0)
240 goto errout;
241 } else if (data_len > 0) {
242 /* ematch module doesn't provide an own change
243 * procedure and expects us to allocate and copy
244 * the ematch data.
245 *
246 * TCF_EM_SIMPLE may be specified stating that the
247 * data only consists of a u32 integer and the module
248 * does not expected a memory reference but rather
249 * the value carried. */
250 if (em_hdr->flags & TCF_EM_SIMPLE) {
251 if (data_len < sizeof(u32))
252 goto errout;
253 em->data = *(u32 *) data;
254 } else {
255 void *v = kmalloc(data_len, GFP_KERNEL);
256 if (v == NULL) {
257 err = -ENOBUFS;
258 goto errout;
259 }
260 memcpy(v, data, data_len);
261 em->data = (unsigned long) v;
262 }
263 }
264 }
265
266 em->matchid = em_hdr->matchid;
267 em->flags = em_hdr->flags;
268 em->datalen = data_len;
269
270 err = 0;
271errout:
272 return err;
273}
274
275/**
276 * tcf_em_tree_validate - validate ematch config TLV and build ematch tree
277 *
278 * @tp: classifier kind handle
279 * @rta: ematch tree configuration TLV
280 * @tree: destination ematch tree variable to store the resulting
281 * ematch tree.
282 *
283 * This function validates the given configuration TLV @rta and builds an
284 * ematch tree in @tree. The resulting tree must later be copied into
285 * the private classifier data using tcf_em_tree_change(). You MUST NOT
286 * provide the ematch tree variable of the private classifier data directly,
287 * the changes would not be locked properly.
288 *
289 * Returns a negative error code if the configuration TLV contains errors.
290 */
291int tcf_em_tree_validate(struct tcf_proto *tp, struct rtattr *rta,
292 struct tcf_ematch_tree *tree)
293{
294 int idx, list_len, matches_len, err = -EINVAL;
295 struct rtattr *tb[TCA_EMATCH_TREE_MAX];
296 struct rtattr *rt_match, *rt_hdr, *rt_list;
297 struct tcf_ematch_tree_hdr *tree_hdr;
298 struct tcf_ematch *em;
299
Thomas Grafb541ca22005-11-08 09:39:17 -0800300 if (!rta) {
301 memset(tree, 0, sizeof(*tree));
302 return 0;
303 }
304
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 if (rtattr_parse_nested(tb, TCA_EMATCH_TREE_MAX, rta) < 0)
306 goto errout;
307
308 rt_hdr = tb[TCA_EMATCH_TREE_HDR-1];
309 rt_list = tb[TCA_EMATCH_TREE_LIST-1];
310
311 if (rt_hdr == NULL || rt_list == NULL)
312 goto errout;
313
314 if (RTA_PAYLOAD(rt_hdr) < sizeof(*tree_hdr) ||
315 RTA_PAYLOAD(rt_list) < sizeof(*rt_match))
316 goto errout;
317
318 tree_hdr = RTA_DATA(rt_hdr);
319 memcpy(&tree->hdr, tree_hdr, sizeof(*tree_hdr));
320
321 rt_match = RTA_DATA(rt_list);
322 list_len = RTA_PAYLOAD(rt_list);
323 matches_len = tree_hdr->nmatches * sizeof(*em);
324
325 tree->matches = kmalloc(matches_len, GFP_KERNEL);
326 if (tree->matches == NULL)
327 goto errout;
328 memset(tree->matches, 0, matches_len);
329
330 /* We do not use rtattr_parse_nested here because the maximum
331 * number of attributes is unknown. This saves us the allocation
332 * for a tb buffer which would serve no purpose at all.
333 *
334 * The array of rt attributes is parsed in the order as they are
335 * provided, their type must be incremental from 1 to n. Even
336 * if it does not serve any real purpose, a failure of sticking
337 * to this policy will result in parsing failure. */
338 for (idx = 0; RTA_OK(rt_match, list_len); idx++) {
339 err = -EINVAL;
340
341 if (rt_match->rta_type != (idx + 1))
342 goto errout_abort;
343
344 if (idx >= tree_hdr->nmatches)
345 goto errout_abort;
346
347 if (RTA_PAYLOAD(rt_match) < sizeof(struct tcf_ematch_hdr))
348 goto errout_abort;
349
350 em = tcf_em_get_match(tree, idx);
351
352 err = tcf_em_validate(tp, tree_hdr, em, rt_match, idx);
353 if (err < 0)
354 goto errout_abort;
355
356 rt_match = RTA_NEXT(rt_match, list_len);
357 }
358
359 /* Check if the number of matches provided by userspace actually
360 * complies with the array of matches. The number was used for
361 * the validation of references and a mismatch could lead to
362 * undefined references during the matching process. */
363 if (idx != tree_hdr->nmatches) {
364 err = -EINVAL;
365 goto errout_abort;
366 }
367
368 err = 0;
369errout:
370 return err;
371
372errout_abort:
373 tcf_em_tree_destroy(tp, tree);
374 return err;
375}
376
377/**
378 * tcf_em_tree_destroy - destroy an ematch tree
379 *
380 * @tp: classifier kind handle
381 * @tree: ematch tree to be deleted
382 *
383 * This functions destroys an ematch tree previously created by
384 * tcf_em_tree_validate()/tcf_em_tree_change(). You must ensure that
385 * the ematch tree is not in use before calling this function.
386 */
387void tcf_em_tree_destroy(struct tcf_proto *tp, struct tcf_ematch_tree *tree)
388{
389 int i;
390
391 if (tree->matches == NULL)
392 return;
393
394 for (i = 0; i < tree->hdr.nmatches; i++) {
395 struct tcf_ematch *em = tcf_em_get_match(tree, i);
396
397 if (em->ops) {
398 if (em->ops->destroy)
399 em->ops->destroy(tp, em);
400 else if (!tcf_em_is_simple(em) && em->data)
401 kfree((void *) em->data);
402 module_put(em->ops->owner);
403 }
404 }
405
406 tree->hdr.nmatches = 0;
407 kfree(tree->matches);
408}
409
410/**
411 * tcf_em_tree_dump - dump ematch tree into a rtnl message
412 *
413 * @skb: skb holding the rtnl message
414 * @t: ematch tree to be dumped
415 * @tlv: TLV type to be used to encapsulate the tree
416 *
417 * This function dumps a ematch tree into a rtnl message. It is valid to
418 * call this function while the ematch tree is in use.
419 *
420 * Returns -1 if the skb tailroom is insufficient.
421 */
422int tcf_em_tree_dump(struct sk_buff *skb, struct tcf_ematch_tree *tree, int tlv)
423{
424 int i;
425 struct rtattr * top_start = (struct rtattr*) skb->tail;
426 struct rtattr * list_start;
427
428 RTA_PUT(skb, tlv, 0, NULL);
429 RTA_PUT(skb, TCA_EMATCH_TREE_HDR, sizeof(tree->hdr), &tree->hdr);
430
431 list_start = (struct rtattr *) skb->tail;
432 RTA_PUT(skb, TCA_EMATCH_TREE_LIST, 0, NULL);
433
434 for (i = 0; i < tree->hdr.nmatches; i++) {
435 struct rtattr *match_start = (struct rtattr*) skb->tail;
436 struct tcf_ematch *em = tcf_em_get_match(tree, i);
437 struct tcf_ematch_hdr em_hdr = {
438 .kind = em->ops ? em->ops->kind : TCF_EM_CONTAINER,
439 .matchid = em->matchid,
440 .flags = em->flags
441 };
442
443 RTA_PUT(skb, i+1, sizeof(em_hdr), &em_hdr);
444
445 if (em->ops && em->ops->dump) {
446 if (em->ops->dump(skb, em) < 0)
447 goto rtattr_failure;
448 } else if (tcf_em_is_container(em) || tcf_em_is_simple(em)) {
449 u32 u = em->data;
450 RTA_PUT_NOHDR(skb, sizeof(u), &u);
451 } else if (em->datalen > 0)
452 RTA_PUT_NOHDR(skb, em->datalen, (void *) em->data);
453
454 match_start->rta_len = skb->tail - (u8*) match_start;
455 }
456
457 list_start->rta_len = skb->tail - (u8 *) list_start;
458 top_start->rta_len = skb->tail - (u8 *) top_start;
459
460 return 0;
461
462rtattr_failure:
463 return -1;
464}
465
466static inline int tcf_em_match(struct sk_buff *skb, struct tcf_ematch *em,
467 struct tcf_pkt_info *info)
468{
469 int r = em->ops->match(skb, em, info);
470 return tcf_em_is_inverted(em) ? !r : r;
471}
472
473/* Do not use this function directly, use tcf_em_tree_match instead */
474int __tcf_em_tree_match(struct sk_buff *skb, struct tcf_ematch_tree *tree,
475 struct tcf_pkt_info *info)
476{
477 int stackp = 0, match_idx = 0, res = 0;
478 struct tcf_ematch *cur_match;
479 int stack[CONFIG_NET_EMATCH_STACK];
480
481proceed:
482 while (match_idx < tree->hdr.nmatches) {
483 cur_match = tcf_em_get_match(tree, match_idx);
484
485 if (tcf_em_is_container(cur_match)) {
486 if (unlikely(stackp >= CONFIG_NET_EMATCH_STACK))
487 goto stack_overflow;
488
489 stack[stackp++] = match_idx;
490 match_idx = cur_match->data;
491 goto proceed;
492 }
493
494 res = tcf_em_match(skb, cur_match, info);
495
496 if (tcf_em_early_end(cur_match, res))
497 break;
498
499 match_idx++;
500 }
501
502pop_stack:
503 if (stackp > 0) {
504 match_idx = stack[--stackp];
505 cur_match = tcf_em_get_match(tree, match_idx);
506
507 if (tcf_em_early_end(cur_match, res))
508 goto pop_stack;
509 else {
510 match_idx++;
511 goto proceed;
512 }
513 }
514
515 return res;
516
517stack_overflow:
518 if (net_ratelimit())
519 printk("Local stack overflow, increase NET_EMATCH_STACK\n");
520 return -1;
521}
522
523EXPORT_SYMBOL(tcf_em_register);
524EXPORT_SYMBOL(tcf_em_unregister);
525EXPORT_SYMBOL(tcf_em_tree_validate);
526EXPORT_SYMBOL(tcf_em_tree_destroy);
527EXPORT_SYMBOL(tcf_em_tree_dump);
528EXPORT_SYMBOL(__tcf_em_tree_match);