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 * 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 void tcf_em_unregister(struct tcf_ematch_ops *ops) 158 { 159 write_lock(&ematch_mod_lock); 160 list_del(&ops->link); 161 write_unlock(&ematch_mod_lock); 162 } 163 EXPORT_SYMBOL(tcf_em_unregister); 164 165 static inline struct tcf_ematch * tcf_em_get_match(struct tcf_ematch_tree *tree, 166 int index) 167 { 168 return &tree->matches[index]; 169 } 170 171 172 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) 175 { 176 int err = -EINVAL; 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); 180 181 if (!TCF_EM_REL_VALID(em_hdr->flags)) 182 goto errout; 183 184 if (em_hdr->kind == TCF_EM_CONTAINER) { 185 /* Special ematch called "container", carries an index 186 * referencing an external ematch sequence. */ 187 u32 ref; 188 189 if (data_len < sizeof(ref)) 190 goto errout; 191 ref = *(u32 *) data; 192 193 if (ref >= tree_hdr->nmatches) 194 goto errout; 195 196 /* We do not allow backward jumps to avoid loops and jumps 197 * to our own position are of course illegal. */ 198 if (ref <= idx) 199 goto errout; 200 201 202 em->data = ref; 203 } else { 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); 212 213 if (em->ops == NULL) { 214 err = -ENOENT; 215 #ifdef CONFIG_MODULES 216 __rtnl_unlock(); 217 request_module("ematch-kind-%u", em_hdr->kind); 218 rtnl_lock(); 219 em->ops = tcf_em_lookup(em_hdr->kind); 220 if (em->ops) { 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); 225 err = -EAGAIN; 226 } 227 #endif 228 goto errout; 229 } 230 231 /* 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) 234 goto errout; 235 236 if (em->ops->change) { 237 err = em->ops->change(tp, data, data_len, em); 238 if (err < 0) 239 goto errout; 240 } else if (data_len > 0) { 241 /* ematch module doesn't provide an own change 242 * procedure and expects us to allocate and copy 243 * the ematch data. 244 * 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)) 251 goto errout; 252 em->data = *(u32 *) data; 253 } else { 254 void *v = kmemdup(data, data_len, GFP_KERNEL); 255 if (v == NULL) { 256 err = -ENOBUFS; 257 goto errout; 258 } 259 em->data = (unsigned long) v; 260 } 261 } 262 } 263 264 em->matchid = em_hdr->matchid; 265 em->flags = em_hdr->flags; 266 em->datalen = data_len; 267 268 err = 0; 269 errout: 270 return err; 271 } 272 273 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 }, 276 }; 277 278 /** 279 * tcf_em_tree_validate - validate ematch config TLV and build ematch tree 280 * 281 * @tp: classifier kind handle 282 * @nla: ematch tree configuration TLV 283 * @tree: destination ematch tree variable to store the resulting 284 * ematch tree. 285 * 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. 291 * 292 * Returns a negative error code if the configuration TLV contains errors. 293 */ 294 int tcf_em_tree_validate(struct tcf_proto *tp, struct nlattr *nla, 295 struct tcf_ematch_tree *tree) 296 { 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; 302 303 memset(tree, 0, sizeof(*tree)); 304 if (!nla) 305 return 0; 306 307 err = nla_parse_nested(tb, TCA_EMATCH_TREE_MAX, nla, em_policy); 308 if (err < 0) 309 goto errout; 310 311 err = -EINVAL; 312 rt_hdr = tb[TCA_EMATCH_TREE_HDR]; 313 rt_list = tb[TCA_EMATCH_TREE_LIST]; 314 315 if (rt_hdr == NULL || rt_list == NULL) 316 goto errout; 317 318 tree_hdr = nla_data(rt_hdr); 319 memcpy(&tree->hdr, tree_hdr, sizeof(*tree_hdr)); 320 321 rt_match = nla_data(rt_list); 322 list_len = nla_len(rt_list); 323 matches_len = tree_hdr->nmatches * sizeof(*em); 324 325 tree->matches = kzalloc(matches_len, GFP_KERNEL); 326 if (tree->matches == NULL) 327 goto errout; 328 329 /* 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. 332 * 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++) { 338 err = -EINVAL; 339 340 if (rt_match->nla_type != (idx + 1)) 341 goto errout_abort; 342 343 if (idx >= tree_hdr->nmatches) 344 goto errout_abort; 345 346 if (nla_len(rt_match) < sizeof(struct tcf_ematch_hdr)) 347 goto errout_abort; 348 349 em = tcf_em_get_match(tree, idx); 350 351 err = tcf_em_validate(tp, tree_hdr, em, rt_match, idx); 352 if (err < 0) 353 goto errout_abort; 354 355 rt_match = nla_next(rt_match, &list_len); 356 } 357 358 /* 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) { 363 err = -EINVAL; 364 goto errout_abort; 365 } 366 367 err = 0; 368 errout: 369 return err; 370 371 errout_abort: 372 tcf_em_tree_destroy(tp, tree); 373 return err; 374 } 375 EXPORT_SYMBOL(tcf_em_tree_validate); 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 */ 387 void 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)) 401 kfree((void *) em->data); 402 module_put(em->ops->owner); 403 } 404 } 405 406 tree->hdr.nmatches = 0; 407 kfree(tree->matches); 408 tree->matches = NULL; 409 } 410 EXPORT_SYMBOL(tcf_em_tree_destroy); 411 412 /** 413 * tcf_em_tree_dump - dump ematch tree into a rtnl message 414 * 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 418 * 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. 421 * 422 * Returns -1 if the skb tailroom is insufficient. 423 */ 424 int tcf_em_tree_dump(struct sk_buff *skb, struct tcf_ematch_tree *tree, int tlv) 425 { 426 int i; 427 u8 *tail; 428 struct nlattr *top_start; 429 struct nlattr *list_start; 430 431 top_start = nla_nest_start(skb, tlv); 432 if (top_start == NULL) 433 goto nla_put_failure; 434 435 NLA_PUT(skb, TCA_EMATCH_TREE_HDR, sizeof(tree->hdr), &tree->hdr); 436 437 list_start = nla_nest_start(skb, TCA_EMATCH_TREE_LIST); 438 if (list_start == NULL) 439 goto nla_put_failure; 440 441 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, 448 .flags = em->flags 449 }; 450 451 NLA_PUT(skb, i+1, sizeof(em_hdr), &em_hdr); 452 453 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)) { 457 u32 u = em->data; 458 nla_put_nohdr(skb, sizeof(u), &u); 459 } else if (em->datalen > 0) 460 nla_put_nohdr(skb, em->datalen, (void *) em->data); 461 462 tail = skb_tail_pointer(skb); 463 match_start->nla_len = tail - (u8 *)match_start; 464 } 465 466 nla_nest_end(skb, list_start); 467 nla_nest_end(skb, top_start); 468 469 return 0; 470 471 nla_put_failure: 472 return -1; 473 } 474 EXPORT_SYMBOL(tcf_em_tree_dump); 475 476 static inline int tcf_em_match(struct sk_buff *skb, struct tcf_ematch *em, 477 struct tcf_pkt_info *info) 478 { 479 int r = em->ops->match(skb, em, info); 480 return tcf_em_is_inverted(em) ? !r : r; 481 } 482 483 /* 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) 486 { 487 int stackp = 0, match_idx = 0, res = 0; 488 struct tcf_ematch *cur_match; 489 int stack[CONFIG_NET_EMATCH_STACK]; 490 491 proceed: 492 while (match_idx < tree->hdr.nmatches) { 493 cur_match = tcf_em_get_match(tree, match_idx); 494 495 if (tcf_em_is_container(cur_match)) { 496 if (unlikely(stackp >= CONFIG_NET_EMATCH_STACK)) 497 goto stack_overflow; 498 499 stack[stackp++] = match_idx; 500 match_idx = cur_match->data; 501 goto proceed; 502 } 503 504 res = tcf_em_match(skb, cur_match, info); 505 506 if (tcf_em_early_end(cur_match, res)) 507 break; 508 509 match_idx++; 510 } 511 512 pop_stack: 513 if (stackp > 0) { 514 match_idx = stack[--stackp]; 515 cur_match = tcf_em_get_match(tree, match_idx); 516 517 if (tcf_em_early_end(cur_match, res)) 518 goto pop_stack; 519 else { 520 match_idx++; 521 goto proceed; 522 } 523 } 524 525 return res; 526 527 stack_overflow: 528 if (net_ratelimit()) 529 printk("Local stack overflow, increase NET_EMATCH_STACK\n"); 530 return -1; 531 } 532 EXPORT_SYMBOL(__tcf_em_tree_match); 533