2 * net/sched/ematch.c Extended Match API
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.
9 * Authors: Thomas Graf <tgraf@suug.ch>
11 * ==========================================================================
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
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.
28 * A AND (B1 OR B2) AND C AND D
30 * ------->-PUSH-------
31 * -->-- / -->-- \ -->--
33 * +-------+-------+-------+-------+-------+--------+
34 * | A AND | B AND | C AND | D END | B1 OR | B2 END |
35 * +-------+-------+-------+-------+-------+--------+
37 * --------<-POP---------
39 * where B is a virtual ematch referencing to sequence starting with B1.
41 * ==========================================================================
43 * How to write an ematch in 60 seconds
44 * ------------------------------------
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)
50 * struct mydata *d = (struct mydata *) m->data;
52 * if (...matching goes here...)
58 * 2) Fill out a struct tcf_ematch_ops:
59 * static struct tcf_ematch_ops my_ops = {
61 * .datalen = sizeof(struct mydata),
63 * .owner = THIS_MODULE,
66 * 3) Register/Unregister your ematch:
67 * static int __init init_my_ematch(void)
69 * return tcf_em_register(&my_ops);
72 * static void __exit exit_my_ematch(void)
74 * tcf_em_unregister(&my_ops);
77 * module_init(init_my_ematch);
78 * module_exit(exit_my_ematch);
80 * 4) By now you should have two more seconds left, barely enough to
81 * open up a beer to watch the compilation going.
84 #include <linux/module.h>
85 #include <linux/slab.h>
86 #include <linux/types.h>
87 #include <linux/kernel.h>
88 #include <linux/errno.h>
89 #include <linux/rtnetlink.h>
90 #include <linux/skbuff.h>
91 #include <net/pkt_cls.h>
93 static LIST_HEAD(ematch_ops);
94 static DEFINE_RWLOCK(ematch_mod_lock);
96 static struct tcf_ematch_ops *tcf_em_lookup(u16 kind)
98 struct tcf_ematch_ops *e = NULL;
100 read_lock(&ematch_mod_lock);
101 list_for_each_entry(e, &ematch_ops, link) {
102 if (kind == e->kind) {
103 if (!try_module_get(e->owner))
105 read_unlock(&ematch_mod_lock);
109 read_unlock(&ematch_mod_lock);
115 * tcf_em_register - register an extended match
117 * @ops: ematch operations lookup table
119 * This function must be called by ematches to announce their presence.
120 * The given @ops must have kind set to a unique identifier and the
121 * callback match() must be implemented. All other callbacks are optional
122 * and a fallback implementation is used instead.
124 * Returns -EEXISTS if an ematch of the same kind has already registered.
126 int tcf_em_register(struct tcf_ematch_ops *ops)
129 struct tcf_ematch_ops *e;
131 if (ops->match == NULL)
134 write_lock(&ematch_mod_lock);
135 list_for_each_entry(e, &ematch_ops, link)
136 if (ops->kind == e->kind)
139 list_add_tail(&ops->link, &ematch_ops);
142 write_unlock(&ematch_mod_lock);
145 EXPORT_SYMBOL(tcf_em_register);
148 * tcf_em_unregister - unregster and extended match
150 * @ops: ematch operations lookup table
152 * This function must be called by ematches to announce their disappearance
153 * for examples when the module gets unloaded. The @ops parameter must be
154 * the same as the one used for registration.
156 * Returns -ENOENT if no matching ematch was found.
158 void tcf_em_unregister(struct tcf_ematch_ops *ops)
160 write_lock(&ematch_mod_lock);
161 list_del(&ops->link);
162 write_unlock(&ematch_mod_lock);
164 EXPORT_SYMBOL(tcf_em_unregister);
166 static inline struct tcf_ematch *tcf_em_get_match(struct tcf_ematch_tree *tree,
169 return &tree->matches[index];
173 static int tcf_em_validate(struct tcf_proto *tp,
174 struct tcf_ematch_tree_hdr *tree_hdr,
175 struct tcf_ematch *em, struct nlattr *nla, int idx)
178 struct tcf_ematch_hdr *em_hdr = nla_data(nla);
179 int data_len = nla_len(nla) - sizeof(*em_hdr);
180 void *data = (void *) em_hdr + sizeof(*em_hdr);
181 struct net *net = dev_net(qdisc_dev(tp->q));
183 if (!TCF_EM_REL_VALID(em_hdr->flags))
186 if (em_hdr->kind == TCF_EM_CONTAINER) {
187 /* Special ematch called "container", carries an index
188 * referencing an external ematch sequence.
192 if (data_len < sizeof(ref))
196 if (ref >= tree_hdr->nmatches)
199 /* We do not allow backward jumps to avoid loops and jumps
200 * to our own position are of course illegal.
208 /* Note: This lookup will increase the module refcnt
209 * of the ematch module referenced. In case of a failure,
210 * a destroy function is called by the underlying layer
211 * which automatically releases the reference again, therefore
212 * the module MUST not be given back under any circumstances
213 * here. Be aware, the destroy function assumes that the
214 * module is held if the ops field is non zero.
216 em->ops = tcf_em_lookup(em_hdr->kind);
218 if (em->ops == NULL) {
220 #ifdef CONFIG_MODULES
222 request_module("ematch-kind-%u", em_hdr->kind);
224 em->ops = tcf_em_lookup(em_hdr->kind);
226 /* We dropped the RTNL mutex in order to
227 * perform the module load. Tell the caller
228 * to replay the request.
230 module_put(em->ops->owner);
238 /* ematch module provides expected length of data, so we
239 * can do a basic sanity check.
241 if (em->ops->datalen && data_len < em->ops->datalen)
244 if (em->ops->change) {
245 err = em->ops->change(net, data, data_len, em);
248 } else if (data_len > 0) {
249 /* ematch module doesn't provide an own change
250 * procedure and expects us to allocate and copy
253 * TCF_EM_SIMPLE may be specified stating that the
254 * data only consists of a u32 integer and the module
255 * does not expected a memory reference but rather
258 if (em_hdr->flags & TCF_EM_SIMPLE) {
259 if (data_len < sizeof(u32))
261 em->data = *(u32 *) data;
263 void *v = kmemdup(data, data_len, GFP_KERNEL);
268 em->data = (unsigned long) v;
273 em->matchid = em_hdr->matchid;
274 em->flags = em_hdr->flags;
275 em->datalen = data_len;
283 static const struct nla_policy em_policy[TCA_EMATCH_TREE_MAX + 1] = {
284 [TCA_EMATCH_TREE_HDR] = { .len = sizeof(struct tcf_ematch_tree_hdr) },
285 [TCA_EMATCH_TREE_LIST] = { .type = NLA_NESTED },
289 * tcf_em_tree_validate - validate ematch config TLV and build ematch tree
291 * @tp: classifier kind handle
292 * @nla: ematch tree configuration TLV
293 * @tree: destination ematch tree variable to store the resulting
296 * This function validates the given configuration TLV @nla and builds an
297 * ematch tree in @tree. The resulting tree must later be copied into
298 * the private classifier data using tcf_em_tree_change(). You MUST NOT
299 * provide the ematch tree variable of the private classifier data directly,
300 * the changes would not be locked properly.
302 * Returns a negative error code if the configuration TLV contains errors.
304 int tcf_em_tree_validate(struct tcf_proto *tp, struct nlattr *nla,
305 struct tcf_ematch_tree *tree)
307 int idx, list_len, matches_len, err;
308 struct nlattr *tb[TCA_EMATCH_TREE_MAX + 1];
309 struct nlattr *rt_match, *rt_hdr, *rt_list;
310 struct tcf_ematch_tree_hdr *tree_hdr;
311 struct tcf_ematch *em;
313 memset(tree, 0, sizeof(*tree));
317 err = nla_parse_nested(tb, TCA_EMATCH_TREE_MAX, nla, em_policy);
322 rt_hdr = tb[TCA_EMATCH_TREE_HDR];
323 rt_list = tb[TCA_EMATCH_TREE_LIST];
325 if (rt_hdr == NULL || rt_list == NULL)
328 tree_hdr = nla_data(rt_hdr);
329 memcpy(&tree->hdr, tree_hdr, sizeof(*tree_hdr));
331 rt_match = nla_data(rt_list);
332 list_len = nla_len(rt_list);
333 matches_len = tree_hdr->nmatches * sizeof(*em);
335 tree->matches = kzalloc(matches_len, GFP_KERNEL);
336 if (tree->matches == NULL)
339 /* We do not use nla_parse_nested here because the maximum
340 * number of attributes is unknown. This saves us the allocation
341 * for a tb buffer which would serve no purpose at all.
343 * The array of rt attributes is parsed in the order as they are
344 * provided, their type must be incremental from 1 to n. Even
345 * if it does not serve any real purpose, a failure of sticking
346 * to this policy will result in parsing failure.
348 for (idx = 0; nla_ok(rt_match, list_len); idx++) {
351 if (rt_match->nla_type != (idx + 1))
354 if (idx >= tree_hdr->nmatches)
357 if (nla_len(rt_match) < sizeof(struct tcf_ematch_hdr))
360 em = tcf_em_get_match(tree, idx);
362 err = tcf_em_validate(tp, tree_hdr, em, rt_match, idx);
366 rt_match = nla_next(rt_match, &list_len);
369 /* Check if the number of matches provided by userspace actually
370 * complies with the array of matches. The number was used for
371 * the validation of references and a mismatch could lead to
372 * undefined references during the matching process.
374 if (idx != tree_hdr->nmatches) {
384 tcf_em_tree_destroy(tree);
387 EXPORT_SYMBOL(tcf_em_tree_validate);
390 * tcf_em_tree_destroy - destroy an ematch tree
392 * @tp: classifier kind handle
393 * @tree: ematch tree to be deleted
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.
399 void tcf_em_tree_destroy(struct tcf_ematch_tree *tree)
403 if (tree->matches == NULL)
406 for (i = 0; i < tree->hdr.nmatches; i++) {
407 struct tcf_ematch *em = tcf_em_get_match(tree, i);
410 if (em->ops->destroy)
411 em->ops->destroy(em);
412 else if (!tcf_em_is_simple(em))
413 kfree((void *) em->data);
414 module_put(em->ops->owner);
418 tree->hdr.nmatches = 0;
419 kfree(tree->matches);
420 tree->matches = NULL;
422 EXPORT_SYMBOL(tcf_em_tree_destroy);
425 * tcf_em_tree_dump - dump ematch tree into a rtnl message
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
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.
434 * Returns -1 if the skb tailroom is insufficient.
436 int tcf_em_tree_dump(struct sk_buff *skb, struct tcf_ematch_tree *tree, int tlv)
440 struct nlattr *top_start;
441 struct nlattr *list_start;
443 top_start = nla_nest_start(skb, tlv);
444 if (top_start == NULL)
445 goto nla_put_failure;
447 if (nla_put(skb, TCA_EMATCH_TREE_HDR, sizeof(tree->hdr), &tree->hdr))
448 goto nla_put_failure;
450 list_start = nla_nest_start(skb, TCA_EMATCH_TREE_LIST);
451 if (list_start == NULL)
452 goto nla_put_failure;
454 tail = skb_tail_pointer(skb);
455 for (i = 0; i < tree->hdr.nmatches; i++) {
456 struct nlattr *match_start = (struct nlattr *)tail;
457 struct tcf_ematch *em = tcf_em_get_match(tree, i);
458 struct tcf_ematch_hdr em_hdr = {
459 .kind = em->ops ? em->ops->kind : TCF_EM_CONTAINER,
460 .matchid = em->matchid,
464 if (nla_put(skb, i + 1, sizeof(em_hdr), &em_hdr))
465 goto nla_put_failure;
467 if (em->ops && em->ops->dump) {
468 if (em->ops->dump(skb, em) < 0)
469 goto nla_put_failure;
470 } else if (tcf_em_is_container(em) || tcf_em_is_simple(em)) {
472 nla_put_nohdr(skb, sizeof(u), &u);
473 } else if (em->datalen > 0)
474 nla_put_nohdr(skb, em->datalen, (void *) em->data);
476 tail = skb_tail_pointer(skb);
477 match_start->nla_len = tail - (u8 *)match_start;
480 nla_nest_end(skb, list_start);
481 nla_nest_end(skb, top_start);
488 EXPORT_SYMBOL(tcf_em_tree_dump);
490 static inline int tcf_em_match(struct sk_buff *skb, struct tcf_ematch *em,
491 struct tcf_pkt_info *info)
493 int r = em->ops->match(skb, em, info);
495 return tcf_em_is_inverted(em) ? !r : r;
498 /* Do not use this function directly, use tcf_em_tree_match instead */
499 int __tcf_em_tree_match(struct sk_buff *skb, struct tcf_ematch_tree *tree,
500 struct tcf_pkt_info *info)
502 int stackp = 0, match_idx = 0, res = 0;
503 struct tcf_ematch *cur_match;
504 int stack[CONFIG_NET_EMATCH_STACK];
507 while (match_idx < tree->hdr.nmatches) {
508 cur_match = tcf_em_get_match(tree, match_idx);
510 if (tcf_em_is_container(cur_match)) {
511 if (unlikely(stackp >= CONFIG_NET_EMATCH_STACK))
514 stack[stackp++] = match_idx;
515 match_idx = cur_match->data;
519 res = tcf_em_match(skb, cur_match, info);
521 if (tcf_em_early_end(cur_match, res))
529 match_idx = stack[--stackp];
530 cur_match = tcf_em_get_match(tree, match_idx);
532 if (tcf_em_is_inverted(cur_match))
535 if (tcf_em_early_end(cur_match, res)) {
546 net_warn_ratelimited("tc ematch: local stack overflow, increase NET_EMATCH_STACK\n");
549 EXPORT_SYMBOL(__tcf_em_tree_match);