xref: /openbmc/linux/net/sched/ematch.c (revision f42b3800)
1 /*
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/module.h>
85 #include <linux/types.h>
86 #include <linux/kernel.h>
87 #include <linux/errno.h>
88 #include <linux/rtnetlink.h>
89 #include <linux/skbuff.h>
90 #include <net/pkt_cls.h>
91 
92 static LIST_HEAD(ematch_ops);
93 static DEFINE_RWLOCK(ematch_mod_lock);
94 
95 static inline struct tcf_ematch_ops * tcf_em_lookup(u16 kind)
96 {
97 	struct tcf_ematch_ops *e = NULL;
98 
99 	read_lock(&ematch_mod_lock);
100 	list_for_each_entry(e, &ematch_ops, link) {
101 		if (kind == e->kind) {
102 			if (!try_module_get(e->owner))
103 				e = NULL;
104 			read_unlock(&ematch_mod_lock);
105 			return e;
106 		}
107 	}
108 	read_unlock(&ematch_mod_lock);
109 
110 	return NULL;
111 }
112 
113 /**
114  * tcf_em_register - register an extended match
115  *
116  * @ops: ematch operations lookup table
117  *
118  * This function must be called by ematches to announce their presence.
119  * The given @ops must have kind set to a unique identifier and the
120  * callback match() must be implemented. All other callbacks are optional
121  * and a fallback implementation is used instead.
122  *
123  * Returns -EEXISTS if an ematch of the same kind has already registered.
124  */
125 int tcf_em_register(struct tcf_ematch_ops *ops)
126 {
127 	int err = -EEXIST;
128 	struct tcf_ematch_ops *e;
129 
130 	if (ops->match == NULL)
131 		return -EINVAL;
132 
133 	write_lock(&ematch_mod_lock);
134 	list_for_each_entry(e, &ematch_ops, link)
135 		if (ops->kind == e->kind)
136 			goto errout;
137 
138 	list_add_tail(&ops->link, &ematch_ops);
139 	err = 0;
140 errout:
141 	write_unlock(&ematch_mod_lock);
142 	return err;
143 }
144 EXPORT_SYMBOL(tcf_em_register);
145 
146 /**
147  * tcf_em_unregister - unregster and extended match
148  *
149  * @ops: ematch operations lookup table
150  *
151  * This function must be called by ematches to announce their disappearance
152  * for examples when the module gets unloaded. The @ops parameter must be
153  * the same as the one used for registration.
154  *
155  * Returns -ENOENT if no matching ematch was found.
156  */
157 int tcf_em_unregister(struct tcf_ematch_ops *ops)
158 {
159 	int err = 0;
160 	struct tcf_ematch_ops *e;
161 
162 	write_lock(&ematch_mod_lock);
163 	list_for_each_entry(e, &ematch_ops, link) {
164 		if (e == ops) {
165 			list_del(&e->link);
166 			goto out;
167 		}
168 	}
169 
170 	err = -ENOENT;
171 out:
172 	write_unlock(&ematch_mod_lock);
173 	return err;
174 }
175 EXPORT_SYMBOL(tcf_em_unregister);
176 
177 static inline struct tcf_ematch * tcf_em_get_match(struct tcf_ematch_tree *tree,
178 						   int index)
179 {
180 	return &tree->matches[index];
181 }
182 
183 
184 static int tcf_em_validate(struct tcf_proto *tp,
185 			   struct tcf_ematch_tree_hdr *tree_hdr,
186 			   struct tcf_ematch *em, struct nlattr *nla, int idx)
187 {
188 	int err = -EINVAL;
189 	struct tcf_ematch_hdr *em_hdr = nla_data(nla);
190 	int data_len = nla_len(nla) - sizeof(*em_hdr);
191 	void *data = (void *) em_hdr + sizeof(*em_hdr);
192 
193 	if (!TCF_EM_REL_VALID(em_hdr->flags))
194 		goto errout;
195 
196 	if (em_hdr->kind == TCF_EM_CONTAINER) {
197 		/* Special ematch called "container", carries an index
198 		 * referencing an external ematch sequence. */
199 		u32 ref;
200 
201 		if (data_len < sizeof(ref))
202 			goto errout;
203 		ref = *(u32 *) data;
204 
205 		if (ref >= tree_hdr->nmatches)
206 			goto errout;
207 
208 		/* We do not allow backward jumps to avoid loops and jumps
209 		 * to our own position are of course illegal. */
210 		if (ref <= idx)
211 			goto errout;
212 
213 
214 		em->data = ref;
215 	} else {
216 		/* Note: This lookup will increase the module refcnt
217 		 * of the ematch module referenced. In case of a failure,
218 		 * a destroy function is called by the underlying layer
219 		 * which automatically releases the reference again, therefore
220 		 * the module MUST not be given back under any circumstances
221 		 * here. Be aware, the destroy function assumes that the
222 		 * module is held if the ops field is non zero. */
223 		em->ops = tcf_em_lookup(em_hdr->kind);
224 
225 		if (em->ops == NULL) {
226 			err = -ENOENT;
227 #ifdef CONFIG_KMOD
228 			__rtnl_unlock();
229 			request_module("ematch-kind-%u", em_hdr->kind);
230 			rtnl_lock();
231 			em->ops = tcf_em_lookup(em_hdr->kind);
232 			if (em->ops) {
233 				/* We dropped the RTNL mutex in order to
234 				 * perform the module load. Tell the caller
235 				 * to replay the request. */
236 				module_put(em->ops->owner);
237 				err = -EAGAIN;
238 			}
239 #endif
240 			goto errout;
241 		}
242 
243 		/* ematch module provides expected length of data, so we
244 		 * can do a basic sanity check. */
245 		if (em->ops->datalen && data_len < em->ops->datalen)
246 			goto errout;
247 
248 		if (em->ops->change) {
249 			err = em->ops->change(tp, data, data_len, em);
250 			if (err < 0)
251 				goto errout;
252 		} else if (data_len > 0) {
253 			/* ematch module doesn't provide an own change
254 			 * procedure and expects us to allocate and copy
255 			 * the ematch data.
256 			 *
257 			 * TCF_EM_SIMPLE may be specified stating that the
258 			 * data only consists of a u32 integer and the module
259 			 * does not expected a memory reference but rather
260 			 * the value carried. */
261 			if (em_hdr->flags & TCF_EM_SIMPLE) {
262 				if (data_len < sizeof(u32))
263 					goto errout;
264 				em->data = *(u32 *) data;
265 			} else {
266 				void *v = kmemdup(data, data_len, GFP_KERNEL);
267 				if (v == NULL) {
268 					err = -ENOBUFS;
269 					goto errout;
270 				}
271 				em->data = (unsigned long) v;
272 			}
273 		}
274 	}
275 
276 	em->matchid = em_hdr->matchid;
277 	em->flags = em_hdr->flags;
278 	em->datalen = data_len;
279 
280 	err = 0;
281 errout:
282 	return err;
283 }
284 
285 static const struct nla_policy em_policy[TCA_EMATCH_TREE_MAX + 1] = {
286 	[TCA_EMATCH_TREE_HDR]	= { .len = sizeof(struct tcf_ematch_tree_hdr) },
287 	[TCA_EMATCH_TREE_LIST]	= { .type = NLA_NESTED },
288 };
289 
290 /**
291  * tcf_em_tree_validate - validate ematch config TLV and build ematch tree
292  *
293  * @tp: classifier kind handle
294  * @nla: ematch tree configuration TLV
295  * @tree: destination ematch tree variable to store the resulting
296  *        ematch tree.
297  *
298  * This function validates the given configuration TLV @nla and builds an
299  * ematch tree in @tree. The resulting tree must later be copied into
300  * the private classifier data using tcf_em_tree_change(). You MUST NOT
301  * provide the ematch tree variable of the private classifier data directly,
302  * the changes would not be locked properly.
303  *
304  * Returns a negative error code if the configuration TLV contains errors.
305  */
306 int tcf_em_tree_validate(struct tcf_proto *tp, struct nlattr *nla,
307 			 struct tcf_ematch_tree *tree)
308 {
309 	int idx, list_len, matches_len, err;
310 	struct nlattr *tb[TCA_EMATCH_TREE_MAX + 1];
311 	struct nlattr *rt_match, *rt_hdr, *rt_list;
312 	struct tcf_ematch_tree_hdr *tree_hdr;
313 	struct tcf_ematch *em;
314 
315 	memset(tree, 0, sizeof(*tree));
316 	if (!nla)
317 		return 0;
318 
319 	err = nla_parse_nested(tb, TCA_EMATCH_TREE_MAX, nla, em_policy);
320 	if (err < 0)
321 		goto errout;
322 
323 	err = -EINVAL;
324 	rt_hdr = tb[TCA_EMATCH_TREE_HDR];
325 	rt_list = tb[TCA_EMATCH_TREE_LIST];
326 
327 	if (rt_hdr == NULL || rt_list == NULL)
328 		goto errout;
329 
330 	tree_hdr = nla_data(rt_hdr);
331 	memcpy(&tree->hdr, tree_hdr, sizeof(*tree_hdr));
332 
333 	rt_match = nla_data(rt_list);
334 	list_len = nla_len(rt_list);
335 	matches_len = tree_hdr->nmatches * sizeof(*em);
336 
337 	tree->matches = kzalloc(matches_len, GFP_KERNEL);
338 	if (tree->matches == NULL)
339 		goto errout;
340 
341 	/* We do not use nla_parse_nested here because the maximum
342 	 * number of attributes is unknown. This saves us the allocation
343 	 * for a tb buffer which would serve no purpose at all.
344 	 *
345 	 * The array of rt attributes is parsed in the order as they are
346 	 * provided, their type must be incremental from 1 to n. Even
347 	 * if it does not serve any real purpose, a failure of sticking
348 	 * to this policy will result in parsing failure. */
349 	for (idx = 0; nla_ok(rt_match, list_len); idx++) {
350 		err = -EINVAL;
351 
352 		if (rt_match->nla_type != (idx + 1))
353 			goto errout_abort;
354 
355 		if (idx >= tree_hdr->nmatches)
356 			goto errout_abort;
357 
358 		if (nla_len(rt_match) < sizeof(struct tcf_ematch_hdr))
359 			goto errout_abort;
360 
361 		em = tcf_em_get_match(tree, idx);
362 
363 		err = tcf_em_validate(tp, tree_hdr, em, rt_match, idx);
364 		if (err < 0)
365 			goto errout_abort;
366 
367 		rt_match = nla_next(rt_match, &list_len);
368 	}
369 
370 	/* Check if the number of matches provided by userspace actually
371 	 * complies with the array of matches. The number was used for
372 	 * the validation of references and a mismatch could lead to
373 	 * undefined references during the matching process. */
374 	if (idx != tree_hdr->nmatches) {
375 		err = -EINVAL;
376 		goto errout_abort;
377 	}
378 
379 	err = 0;
380 errout:
381 	return err;
382 
383 errout_abort:
384 	tcf_em_tree_destroy(tp, tree);
385 	return err;
386 }
387 EXPORT_SYMBOL(tcf_em_tree_validate);
388 
389 /**
390  * tcf_em_tree_destroy - destroy an ematch tree
391  *
392  * @tp: classifier kind handle
393  * @tree: ematch tree to be deleted
394  *
395  * This functions destroys an ematch tree previously created by
396  * tcf_em_tree_validate()/tcf_em_tree_change(). You must ensure that
397  * the ematch tree is not in use before calling this function.
398  */
399 void tcf_em_tree_destroy(struct tcf_proto *tp, struct tcf_ematch_tree *tree)
400 {
401 	int i;
402 
403 	if (tree->matches == NULL)
404 		return;
405 
406 	for (i = 0; i < tree->hdr.nmatches; i++) {
407 		struct tcf_ematch *em = tcf_em_get_match(tree, i);
408 
409 		if (em->ops) {
410 			if (em->ops->destroy)
411 				em->ops->destroy(tp, em);
412 			else if (!tcf_em_is_simple(em))
413 				kfree((void *) em->data);
414 			module_put(em->ops->owner);
415 		}
416 	}
417 
418 	tree->hdr.nmatches = 0;
419 	kfree(tree->matches);
420 	tree->matches = NULL;
421 }
422 EXPORT_SYMBOL(tcf_em_tree_destroy);
423 
424 /**
425  * tcf_em_tree_dump - dump ematch tree into a rtnl message
426  *
427  * @skb: skb holding the rtnl message
428  * @t: ematch tree to be dumped
429  * @tlv: TLV type to be used to encapsulate the tree
430  *
431  * This function dumps a ematch tree into a rtnl message. It is valid to
432  * call this function while the ematch tree is in use.
433  *
434  * Returns -1 if the skb tailroom is insufficient.
435  */
436 int tcf_em_tree_dump(struct sk_buff *skb, struct tcf_ematch_tree *tree, int tlv)
437 {
438 	int i;
439 	u8 *tail;
440 	struct nlattr *top_start;
441 	struct nlattr *list_start;
442 
443 	top_start = nla_nest_start(skb, tlv);
444 	if (top_start == NULL)
445 		goto nla_put_failure;
446 
447 	NLA_PUT(skb, TCA_EMATCH_TREE_HDR, sizeof(tree->hdr), &tree->hdr);
448 
449 	list_start = nla_nest_start(skb, TCA_EMATCH_TREE_LIST);
450 	if (list_start == NULL)
451 		goto nla_put_failure;
452 
453 	tail = skb_tail_pointer(skb);
454 	for (i = 0; i < tree->hdr.nmatches; i++) {
455 		struct nlattr *match_start = (struct nlattr *)tail;
456 		struct tcf_ematch *em = tcf_em_get_match(tree, i);
457 		struct tcf_ematch_hdr em_hdr = {
458 			.kind = em->ops ? em->ops->kind : TCF_EM_CONTAINER,
459 			.matchid = em->matchid,
460 			.flags = em->flags
461 		};
462 
463 		NLA_PUT(skb, i+1, sizeof(em_hdr), &em_hdr);
464 
465 		if (em->ops && em->ops->dump) {
466 			if (em->ops->dump(skb, em) < 0)
467 				goto nla_put_failure;
468 		} else if (tcf_em_is_container(em) || tcf_em_is_simple(em)) {
469 			u32 u = em->data;
470 			nla_put_nohdr(skb, sizeof(u), &u);
471 		} else if (em->datalen > 0)
472 			nla_put_nohdr(skb, em->datalen, (void *) em->data);
473 
474 		tail = skb_tail_pointer(skb);
475 		match_start->nla_len = tail - (u8 *)match_start;
476 	}
477 
478 	nla_nest_end(skb, list_start);
479 	nla_nest_end(skb, top_start);
480 
481 	return 0;
482 
483 nla_put_failure:
484 	return -1;
485 }
486 EXPORT_SYMBOL(tcf_em_tree_dump);
487 
488 static inline int tcf_em_match(struct sk_buff *skb, struct tcf_ematch *em,
489 			       struct tcf_pkt_info *info)
490 {
491 	int r = em->ops->match(skb, em, info);
492 	return tcf_em_is_inverted(em) ? !r : r;
493 }
494 
495 /* Do not use this function directly, use tcf_em_tree_match instead */
496 int __tcf_em_tree_match(struct sk_buff *skb, struct tcf_ematch_tree *tree,
497 			struct tcf_pkt_info *info)
498 {
499 	int stackp = 0, match_idx = 0, res = 0;
500 	struct tcf_ematch *cur_match;
501 	int stack[CONFIG_NET_EMATCH_STACK];
502 
503 proceed:
504 	while (match_idx < tree->hdr.nmatches) {
505 		cur_match = tcf_em_get_match(tree, match_idx);
506 
507 		if (tcf_em_is_container(cur_match)) {
508 			if (unlikely(stackp >= CONFIG_NET_EMATCH_STACK))
509 				goto stack_overflow;
510 
511 			stack[stackp++] = match_idx;
512 			match_idx = cur_match->data;
513 			goto proceed;
514 		}
515 
516 		res = tcf_em_match(skb, cur_match, info);
517 
518 		if (tcf_em_early_end(cur_match, res))
519 			break;
520 
521 		match_idx++;
522 	}
523 
524 pop_stack:
525 	if (stackp > 0) {
526 		match_idx = stack[--stackp];
527 		cur_match = tcf_em_get_match(tree, match_idx);
528 
529 		if (tcf_em_early_end(cur_match, res))
530 			goto pop_stack;
531 		else {
532 			match_idx++;
533 			goto proceed;
534 		}
535 	}
536 
537 	return res;
538 
539 stack_overflow:
540 	if (net_ratelimit())
541 		printk("Local stack overflow, increase NET_EMATCH_STACK\n");
542 	return -1;
543 }
544 EXPORT_SYMBOL(__tcf_em_tree_match);
545