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/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>
92 static LIST_HEAD(ematch_ops);
93 static DEFINE_RWLOCK(ematch_mod_lock);
95 static inline struct tcf_ematch_ops * tcf_em_lookup(u16 kind)
97 struct tcf_ematch_ops *e = NULL;
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))
104 read_unlock(&ematch_mod_lock);
108 read_unlock(&ematch_mod_lock);
114 * tcf_em_register - register an extended match
116 * @ops: ematch operations lookup table
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.
123 * Returns -EEXISTS if an ematch of the same kind has already registered.
125 int tcf_em_register(struct tcf_ematch_ops *ops)
128 struct tcf_ematch_ops *e;
130 if (ops->match == NULL)
133 write_lock(&ematch_mod_lock);
134 list_for_each_entry(e, &ematch_ops, link)
135 if (ops->kind == e->kind)
138 list_add_tail(&ops->link, &ematch_ops);
141 write_unlock(&ematch_mod_lock);
144 EXPORT_SYMBOL(tcf_em_register);
147 * tcf_em_unregister - unregster and extended match
149 * @ops: ematch operations lookup table
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.
155 * Returns -ENOENT if no matching ematch was found.
157 void tcf_em_unregister(struct tcf_ematch_ops *ops)
159 write_lock(&ematch_mod_lock);
160 list_del(&ops->link);
161 write_unlock(&ematch_mod_lock);
163 EXPORT_SYMBOL(tcf_em_unregister);
165 static inline struct tcf_ematch * tcf_em_get_match(struct tcf_ematch_tree *tree,
168 return &tree->matches[index];
172 static int tcf_em_validate(struct tcf_proto *tp,
173 struct tcf_ematch_tree_hdr *tree_hdr,
174 struct tcf_ematch *em, struct nlattr *nla, int idx)
177 struct tcf_ematch_hdr *em_hdr = nla_data(nla);
178 int data_len = nla_len(nla) - sizeof(*em_hdr);
179 void *data = (void *) em_hdr + sizeof(*em_hdr);
181 if (!TCF_EM_REL_VALID(em_hdr->flags))
184 if (em_hdr->kind == TCF_EM_CONTAINER) {
185 /* Special ematch called "container", carries an index
186 * referencing an external ematch sequence. */
189 if (data_len < sizeof(ref))
193 if (ref >= tree_hdr->nmatches)
196 /* We do not allow backward jumps to avoid loops and jumps
197 * to our own position are of course illegal. */
204 /* Note: This lookup will increase the module refcnt
205 * of the ematch module referenced. In case of a failure,
206 * a destroy function is called by the underlying layer
207 * which automatically releases the reference again, therefore
208 * the module MUST not be given back under any circumstances
209 * here. Be aware, the destroy function assumes that the
210 * module is held if the ops field is non zero. */
211 em->ops = tcf_em_lookup(em_hdr->kind);
213 if (em->ops == NULL) {
215 #ifdef CONFIG_MODULES
217 request_module("ematch-kind-%u", em_hdr->kind);
219 em->ops = tcf_em_lookup(em_hdr->kind);
221 /* We dropped the RTNL mutex in order to
222 * perform the module load. Tell the caller
223 * to replay the request. */
224 module_put(em->ops->owner);
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)
236 if (em->ops->change) {
237 err = em->ops->change(tp, data, data_len, em);
240 } else if (data_len > 0) {
241 /* ematch module doesn't provide an own change
242 * procedure and expects us to allocate and copy
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))
252 em->data = *(u32 *) data;
254 void *v = kmemdup(data, data_len, GFP_KERNEL);
259 em->data = (unsigned long) v;
264 em->matchid = em_hdr->matchid;
265 em->flags = em_hdr->flags;
266 em->datalen = data_len;
273 static const struct nla_policy em_policy[TCA_EMATCH_TREE_MAX + 1] = {
274 [TCA_EMATCH_TREE_HDR] = { .len = sizeof(struct tcf_ematch_tree_hdr) },
275 [TCA_EMATCH_TREE_LIST] = { .type = NLA_NESTED },
279 * tcf_em_tree_validate - validate ematch config TLV and build ematch tree
281 * @tp: classifier kind handle
282 * @nla: ematch tree configuration TLV
283 * @tree: destination ematch tree variable to store the resulting
286 * This function validates the given configuration TLV @nla and builds an
287 * ematch tree in @tree. The resulting tree must later be copied into
288 * the private classifier data using tcf_em_tree_change(). You MUST NOT
289 * provide the ematch tree variable of the private classifier data directly,
290 * the changes would not be locked properly.
292 * Returns a negative error code if the configuration TLV contains errors.
294 int tcf_em_tree_validate(struct tcf_proto *tp, struct nlattr *nla,
295 struct tcf_ematch_tree *tree)
297 int idx, list_len, matches_len, err;
298 struct nlattr *tb[TCA_EMATCH_TREE_MAX + 1];
299 struct nlattr *rt_match, *rt_hdr, *rt_list;
300 struct tcf_ematch_tree_hdr *tree_hdr;
301 struct tcf_ematch *em;
303 memset(tree, 0, sizeof(*tree));
307 err = nla_parse_nested(tb, TCA_EMATCH_TREE_MAX, nla, em_policy);
312 rt_hdr = tb[TCA_EMATCH_TREE_HDR];
313 rt_list = tb[TCA_EMATCH_TREE_LIST];
315 if (rt_hdr == NULL || rt_list == NULL)
318 tree_hdr = nla_data(rt_hdr);
319 memcpy(&tree->hdr, tree_hdr, sizeof(*tree_hdr));
321 rt_match = nla_data(rt_list);
322 list_len = nla_len(rt_list);
323 matches_len = tree_hdr->nmatches * sizeof(*em);
325 tree->matches = kzalloc(matches_len, GFP_KERNEL);
326 if (tree->matches == NULL)
329 /* We do not use nla_parse_nested here because the maximum
330 * number of attributes is unknown. This saves us the allocation
331 * for a tb buffer which would serve no purpose at all.
333 * The array of rt attributes is parsed in the order as they are
334 * provided, their type must be incremental from 1 to n. Even
335 * if it does not serve any real purpose, a failure of sticking
336 * to this policy will result in parsing failure. */
337 for (idx = 0; nla_ok(rt_match, list_len); idx++) {
340 if (rt_match->nla_type != (idx + 1))
343 if (idx >= tree_hdr->nmatches)
346 if (nla_len(rt_match) < sizeof(struct tcf_ematch_hdr))
349 em = tcf_em_get_match(tree, idx);
351 err = tcf_em_validate(tp, tree_hdr, em, rt_match, idx);
355 rt_match = nla_next(rt_match, &list_len);
358 /* Check if the number of matches provided by userspace actually
359 * complies with the array of matches. The number was used for
360 * the validation of references and a mismatch could lead to
361 * undefined references during the matching process. */
362 if (idx != tree_hdr->nmatches) {
372 tcf_em_tree_destroy(tp, tree);
375 EXPORT_SYMBOL(tcf_em_tree_validate);
378 * tcf_em_tree_destroy - destroy an ematch tree
380 * @tp: classifier kind handle
381 * @tree: ematch tree to be deleted
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.
387 void tcf_em_tree_destroy(struct tcf_proto *tp, struct tcf_ematch_tree *tree)
391 if (tree->matches == NULL)
394 for (i = 0; i < tree->hdr.nmatches; i++) {
395 struct tcf_ematch *em = tcf_em_get_match(tree, i);
398 if (em->ops->destroy)
399 em->ops->destroy(tp, em);
400 else if (!tcf_em_is_simple(em))
401 kfree((void *) em->data);
402 module_put(em->ops->owner);
406 tree->hdr.nmatches = 0;
407 kfree(tree->matches);
408 tree->matches = NULL;
410 EXPORT_SYMBOL(tcf_em_tree_destroy);
413 * tcf_em_tree_dump - dump ematch tree into a rtnl message
415 * @skb: skb holding the rtnl message
416 * @t: ematch tree to be dumped
417 * @tlv: TLV type to be used to encapsulate the tree
419 * This function dumps a ematch tree into a rtnl message. It is valid to
420 * call this function while the ematch tree is in use.
422 * Returns -1 if the skb tailroom is insufficient.
424 int tcf_em_tree_dump(struct sk_buff *skb, struct tcf_ematch_tree *tree, int tlv)
428 struct nlattr *top_start;
429 struct nlattr *list_start;
431 top_start = nla_nest_start(skb, tlv);
432 if (top_start == NULL)
433 goto nla_put_failure;
435 NLA_PUT(skb, TCA_EMATCH_TREE_HDR, sizeof(tree->hdr), &tree->hdr);
437 list_start = nla_nest_start(skb, TCA_EMATCH_TREE_LIST);
438 if (list_start == NULL)
439 goto nla_put_failure;
441 tail = skb_tail_pointer(skb);
442 for (i = 0; i < tree->hdr.nmatches; i++) {
443 struct nlattr *match_start = (struct nlattr *)tail;
444 struct tcf_ematch *em = tcf_em_get_match(tree, i);
445 struct tcf_ematch_hdr em_hdr = {
446 .kind = em->ops ? em->ops->kind : TCF_EM_CONTAINER,
447 .matchid = em->matchid,
451 NLA_PUT(skb, i+1, sizeof(em_hdr), &em_hdr);
453 if (em->ops && em->ops->dump) {
454 if (em->ops->dump(skb, em) < 0)
455 goto nla_put_failure;
456 } else if (tcf_em_is_container(em) || tcf_em_is_simple(em)) {
458 nla_put_nohdr(skb, sizeof(u), &u);
459 } else if (em->datalen > 0)
460 nla_put_nohdr(skb, em->datalen, (void *) em->data);
462 tail = skb_tail_pointer(skb);
463 match_start->nla_len = tail - (u8 *)match_start;
466 nla_nest_end(skb, list_start);
467 nla_nest_end(skb, top_start);
474 EXPORT_SYMBOL(tcf_em_tree_dump);
476 static inline int tcf_em_match(struct sk_buff *skb, struct tcf_ematch *em,
477 struct tcf_pkt_info *info)
479 int r = em->ops->match(skb, em, info);
480 return tcf_em_is_inverted(em) ? !r : r;
483 /* Do not use this function directly, use tcf_em_tree_match instead */
484 int __tcf_em_tree_match(struct sk_buff *skb, struct tcf_ematch_tree *tree,
485 struct tcf_pkt_info *info)
487 int stackp = 0, match_idx = 0, res = 0;
488 struct tcf_ematch *cur_match;
489 int stack[CONFIG_NET_EMATCH_STACK];
492 while (match_idx < tree->hdr.nmatches) {
493 cur_match = tcf_em_get_match(tree, match_idx);
495 if (tcf_em_is_container(cur_match)) {
496 if (unlikely(stackp >= CONFIG_NET_EMATCH_STACK))
499 stack[stackp++] = match_idx;
500 match_idx = cur_match->data;
504 res = tcf_em_match(skb, cur_match, info);
506 if (tcf_em_early_end(cur_match, res))
514 match_idx = stack[--stackp];
515 cur_match = tcf_em_get_match(tree, match_idx);
517 if (tcf_em_early_end(cur_match, res))
529 printk("Local stack overflow, increase NET_EMATCH_STACK\n");
532 EXPORT_SYMBOL(__tcf_em_tree_match);