1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * net/sched/ematch.c Extended Match API
5 * Authors: Thomas Graf <tgraf@suug.ch>
7 * ==========================================================================
9 * An extended match (ematch) is a small classification tool not worth
10 * writing a full classifier for. Ematches can be interconnected to form
11 * a logic expression and get attached to classifiers to extend their
14 * The userspace part transforms the logic expressions into an array
15 * consisting of multiple sequences of interconnected ematches separated
16 * by markers. Precedence is implemented by a special ematch kind
17 * referencing a sequence beyond the marker of the current sequence
18 * causing the current position in the sequence to be pushed onto a stack
19 * to allow the current position to be overwritten by the position referenced
20 * in the special ematch. Matching continues in the new sequence until a
21 * marker is reached causing the position to be restored from the stack.
24 * A AND (B1 OR B2) AND C AND D
26 * ------->-PUSH-------
27 * -->-- / -->-- \ -->--
29 * +-------+-------+-------+-------+-------+--------+
30 * | A AND | B AND | C AND | D END | B1 OR | B2 END |
31 * +-------+-------+-------+-------+-------+--------+
33 * --------<-POP---------
35 * where B is a virtual ematch referencing to sequence starting with B1.
37 * ==========================================================================
39 * How to write an ematch in 60 seconds
40 * ------------------------------------
42 * 1) Provide a matcher function:
43 * static int my_match(struct sk_buff *skb, struct tcf_ematch *m,
44 * struct tcf_pkt_info *info)
46 * struct mydata *d = (struct mydata *) m->data;
48 * if (...matching goes here...)
54 * 2) Fill out a struct tcf_ematch_ops:
55 * static struct tcf_ematch_ops my_ops = {
57 * .datalen = sizeof(struct mydata),
59 * .owner = THIS_MODULE,
62 * 3) Register/Unregister your ematch:
63 * static int __init init_my_ematch(void)
65 * return tcf_em_register(&my_ops);
68 * static void __exit exit_my_ematch(void)
70 * tcf_em_unregister(&my_ops);
73 * module_init(init_my_ematch);
74 * module_exit(exit_my_ematch);
76 * 4) By now you should have two more seconds left, barely enough to
77 * open up a beer to watch the compilation going.
80 #include <linux/module.h>
81 #include <linux/slab.h>
82 #include <linux/types.h>
83 #include <linux/kernel.h>
84 #include <linux/errno.h>
85 #include <linux/rtnetlink.h>
86 #include <linux/skbuff.h>
87 #include <net/pkt_cls.h>
89 static LIST_HEAD(ematch_ops);
90 static DEFINE_RWLOCK(ematch_mod_lock);
92 static struct tcf_ematch_ops *tcf_em_lookup(u16 kind)
94 struct tcf_ematch_ops *e = NULL;
96 read_lock(&ematch_mod_lock);
97 list_for_each_entry(e, &ematch_ops, link) {
98 if (kind == e->kind) {
99 if (!try_module_get(e->owner))
101 read_unlock(&ematch_mod_lock);
105 read_unlock(&ematch_mod_lock);
111 * tcf_em_register - register an extended match
113 * @ops: ematch operations lookup table
115 * This function must be called by ematches to announce their presence.
116 * The given @ops must have kind set to a unique identifier and the
117 * callback match() must be implemented. All other callbacks are optional
118 * and a fallback implementation is used instead.
120 * Returns -EEXISTS if an ematch of the same kind has already registered.
122 int tcf_em_register(struct tcf_ematch_ops *ops)
125 struct tcf_ematch_ops *e;
127 if (ops->match == NULL)
130 write_lock(&ematch_mod_lock);
131 list_for_each_entry(e, &ematch_ops, link)
132 if (ops->kind == e->kind)
135 list_add_tail(&ops->link, &ematch_ops);
138 write_unlock(&ematch_mod_lock);
141 EXPORT_SYMBOL(tcf_em_register);
144 * tcf_em_unregister - unregister and extended match
146 * @ops: ematch operations lookup table
148 * This function must be called by ematches to announce their disappearance
149 * for examples when the module gets unloaded. The @ops parameter must be
150 * the same as the one used for registration.
152 * Returns -ENOENT if no matching ematch was found.
154 void tcf_em_unregister(struct tcf_ematch_ops *ops)
156 write_lock(&ematch_mod_lock);
157 list_del(&ops->link);
158 write_unlock(&ematch_mod_lock);
160 EXPORT_SYMBOL(tcf_em_unregister);
162 static inline struct tcf_ematch *tcf_em_get_match(struct tcf_ematch_tree *tree,
165 return &tree->matches[index];
169 static int tcf_em_validate(struct tcf_proto *tp,
170 struct tcf_ematch_tree_hdr *tree_hdr,
171 struct tcf_ematch *em, struct nlattr *nla, int idx)
174 struct tcf_ematch_hdr *em_hdr = nla_data(nla);
175 int data_len = nla_len(nla) - sizeof(*em_hdr);
176 void *data = (void *) em_hdr + sizeof(*em_hdr);
177 struct net *net = tp->chain->block->net;
179 if (!TCF_EM_REL_VALID(em_hdr->flags))
182 if (em_hdr->kind == TCF_EM_CONTAINER) {
183 /* Special ematch called "container", carries an index
184 * referencing an external ematch sequence.
188 if (data_len < sizeof(ref))
192 if (ref >= tree_hdr->nmatches)
195 /* We do not allow backward jumps to avoid loops and jumps
196 * 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.
212 em->ops = tcf_em_lookup(em_hdr->kind);
214 if (em->ops == NULL) {
216 #ifdef CONFIG_MODULES
218 request_module("ematch-kind-%u", em_hdr->kind);
220 em->ops = tcf_em_lookup(em_hdr->kind);
222 /* We dropped the RTNL mutex in order to
223 * perform the module load. Tell the caller
224 * to replay the request.
226 module_put(em->ops->owner);
234 /* ematch module provides expected length of data, so we
235 * can do a basic sanity check.
237 if (em->ops->datalen && data_len < em->ops->datalen)
240 if (em->ops->change) {
242 if (em_hdr->flags & TCF_EM_SIMPLE)
244 err = em->ops->change(net, data, data_len, em);
247 } else if (data_len > 0) {
248 /* ematch module doesn't provide an own change
249 * procedure and expects us to allocate and copy
252 * TCF_EM_SIMPLE may be specified stating that the
253 * data only consists of a u32 integer and the module
254 * does not expected a memory reference but rather
257 if (em_hdr->flags & TCF_EM_SIMPLE) {
258 if (em->ops->datalen > 0)
260 if (data_len < sizeof(u32))
262 em->data = *(u32 *) data;
264 void *v = kmemdup(data, data_len, GFP_KERNEL);
269 em->data = (unsigned long) v;
271 em->datalen = data_len;
275 em->matchid = em_hdr->matchid;
276 em->flags = em_hdr->flags;
284 static const struct nla_policy em_policy[TCA_EMATCH_TREE_MAX + 1] = {
285 [TCA_EMATCH_TREE_HDR] = { .len = sizeof(struct tcf_ematch_tree_hdr) },
286 [TCA_EMATCH_TREE_LIST] = { .type = NLA_NESTED },
290 * tcf_em_tree_validate - validate ematch config TLV and build ematch tree
292 * @tp: classifier kind handle
293 * @nla: ematch tree configuration TLV
294 * @tree: destination ematch tree variable to store the resulting
297 * This function validates the given configuration TLV @nla and builds an
298 * ematch tree in @tree. The resulting tree must later be copied into
299 * the private classifier data using tcf_em_tree_change(). You MUST NOT
300 * provide the ematch tree variable of the private classifier data directly,
301 * the changes would not be locked properly.
303 * Returns a negative error code if the configuration TLV contains errors.
305 int tcf_em_tree_validate(struct tcf_proto *tp, struct nlattr *nla,
306 struct tcf_ematch_tree *tree)
308 int idx, list_len, matches_len, err;
309 struct nlattr *tb[TCA_EMATCH_TREE_MAX + 1];
310 struct nlattr *rt_match, *rt_hdr, *rt_list;
311 struct tcf_ematch_tree_hdr *tree_hdr;
312 struct tcf_ematch *em;
314 memset(tree, 0, sizeof(*tree));
318 err = nla_parse_nested_deprecated(tb, TCA_EMATCH_TREE_MAX, nla,
324 rt_hdr = tb[TCA_EMATCH_TREE_HDR];
325 rt_list = tb[TCA_EMATCH_TREE_LIST];
327 if (rt_hdr == NULL || rt_list == NULL)
330 tree_hdr = nla_data(rt_hdr);
331 memcpy(&tree->hdr, tree_hdr, sizeof(*tree_hdr));
333 rt_match = nla_data(rt_list);
334 list_len = nla_len(rt_list);
335 matches_len = tree_hdr->nmatches * sizeof(*em);
337 tree->matches = kzalloc(matches_len, GFP_KERNEL);
338 if (tree->matches == NULL)
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.
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.
350 for (idx = 0; nla_ok(rt_match, list_len); idx++) {
353 if (rt_match->nla_type != (idx + 1))
356 if (idx >= tree_hdr->nmatches)
359 if (nla_len(rt_match) < sizeof(struct tcf_ematch_hdr))
362 em = tcf_em_get_match(tree, idx);
364 err = tcf_em_validate(tp, tree_hdr, em, rt_match, idx);
368 rt_match = nla_next(rt_match, &list_len);
371 /* Check if the number of matches provided by userspace actually
372 * complies with the array of matches. The number was used for
373 * the validation of references and a mismatch could lead to
374 * undefined references during the matching process.
376 if (idx != tree_hdr->nmatches) {
386 tcf_em_tree_destroy(tree);
389 EXPORT_SYMBOL(tcf_em_tree_validate);
392 * tcf_em_tree_destroy - destroy an ematch tree
394 * @tree: ematch tree to be deleted
396 * This functions destroys an ematch tree previously created by
397 * tcf_em_tree_validate()/tcf_em_tree_change(). You must ensure that
398 * the ematch tree is not in use before calling this function.
400 void tcf_em_tree_destroy(struct tcf_ematch_tree *tree)
404 if (tree->matches == NULL)
407 for (i = 0; i < tree->hdr.nmatches; i++) {
408 struct tcf_ematch *em = tcf_em_get_match(tree, i);
411 if (em->ops->destroy)
412 em->ops->destroy(em);
413 else if (!tcf_em_is_simple(em))
414 kfree((void *) em->data);
415 module_put(em->ops->owner);
419 tree->hdr.nmatches = 0;
420 kfree(tree->matches);
421 tree->matches = NULL;
423 EXPORT_SYMBOL(tcf_em_tree_destroy);
426 * tcf_em_tree_dump - dump ematch tree into a rtnl message
428 * @skb: skb holding the rtnl message
429 * @tree: ematch tree to be dumped
430 * @tlv: TLV type to be used to encapsulate the tree
432 * This function dumps a ematch tree into a rtnl message. It is valid to
433 * call this function while the ematch tree is in use.
435 * Returns -1 if the skb tailroom is insufficient.
437 int tcf_em_tree_dump(struct sk_buff *skb, struct tcf_ematch_tree *tree, int tlv)
441 struct nlattr *top_start;
442 struct nlattr *list_start;
444 top_start = nla_nest_start_noflag(skb, tlv);
445 if (top_start == NULL)
446 goto nla_put_failure;
448 if (nla_put(skb, TCA_EMATCH_TREE_HDR, sizeof(tree->hdr), &tree->hdr))
449 goto nla_put_failure;
451 list_start = nla_nest_start_noflag(skb, TCA_EMATCH_TREE_LIST);
452 if (list_start == NULL)
453 goto nla_put_failure;
455 tail = skb_tail_pointer(skb);
456 for (i = 0; i < tree->hdr.nmatches; i++) {
457 struct nlattr *match_start = (struct nlattr *)tail;
458 struct tcf_ematch *em = tcf_em_get_match(tree, i);
459 struct tcf_ematch_hdr em_hdr = {
460 .kind = em->ops ? em->ops->kind : TCF_EM_CONTAINER,
461 .matchid = em->matchid,
465 if (nla_put(skb, i + 1, sizeof(em_hdr), &em_hdr))
466 goto nla_put_failure;
468 if (em->ops && em->ops->dump) {
469 if (em->ops->dump(skb, em) < 0)
470 goto nla_put_failure;
471 } else if (tcf_em_is_container(em) || tcf_em_is_simple(em)) {
473 nla_put_nohdr(skb, sizeof(u), &u);
474 } else if (em->datalen > 0)
475 nla_put_nohdr(skb, em->datalen, (void *) em->data);
477 tail = skb_tail_pointer(skb);
478 match_start->nla_len = tail - (u8 *)match_start;
481 nla_nest_end(skb, list_start);
482 nla_nest_end(skb, top_start);
489 EXPORT_SYMBOL(tcf_em_tree_dump);
491 static inline int tcf_em_match(struct sk_buff *skb, struct tcf_ematch *em,
492 struct tcf_pkt_info *info)
494 int r = em->ops->match(skb, em, info);
496 return tcf_em_is_inverted(em) ? !r : r;
499 /* Do not use this function directly, use tcf_em_tree_match instead */
500 int __tcf_em_tree_match(struct sk_buff *skb, struct tcf_ematch_tree *tree,
501 struct tcf_pkt_info *info)
503 int stackp = 0, match_idx = 0, res = 0;
504 struct tcf_ematch *cur_match;
505 int stack[CONFIG_NET_EMATCH_STACK];
508 while (match_idx < tree->hdr.nmatches) {
509 cur_match = tcf_em_get_match(tree, match_idx);
511 if (tcf_em_is_container(cur_match)) {
512 if (unlikely(stackp >= CONFIG_NET_EMATCH_STACK))
515 stack[stackp++] = match_idx;
516 match_idx = cur_match->data;
520 res = tcf_em_match(skb, cur_match, info);
522 if (tcf_em_early_end(cur_match, res))
530 match_idx = stack[--stackp];
531 cur_match = tcf_em_get_match(tree, match_idx);
533 if (tcf_em_is_inverted(cur_match))
536 if (tcf_em_early_end(cur_match, res)) {
547 net_warn_ratelimited("tc ematch: local stack overflow, increase NET_EMATCH_STACK\n");
550 EXPORT_SYMBOL(__tcf_em_tree_match);