1 /* $OpenBSD: pfctl_optimize.c,v 1.13 2006/10/31 14:17:45 mcbride Exp $ */ 2 3 /* 4 * Copyright (c) 2004 Mike Frantzen <frantzen@openbsd.org> 5 * 6 * Permission to use, copy, modify, and distribute this software for any 7 * purpose with or without fee is hereby granted, provided that the above 8 * copyright notice and this permission notice appear in all copies. 9 * 10 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES 11 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF 12 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR 13 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES 14 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN 15 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF 16 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. 17 */ 18 19 #include <sys/types.h> 20 #include <sys/ioctl.h> 21 #include <sys/socket.h> 22 23 #include <net/if.h> 24 #include <net/pf/pfvar.h> 25 26 #include <netinet/in.h> 27 #include <arpa/inet.h> 28 29 #include <assert.h> 30 #include <ctype.h> 31 #include <err.h> 32 #include <errno.h> 33 #include <stddef.h> 34 #include <stdio.h> 35 #include <stdlib.h> 36 #include <string.h> 37 38 #include "pfctl_parser.h" 39 #include "pfctl.h" 40 41 /* The size at which a table becomes faster than individual rules */ 42 #define TABLE_THRESHOLD 6 43 44 45 /* #define OPT_DEBUG 1 */ 46 #ifdef OPT_DEBUG 47 # define DEBUG(str, v...) \ 48 printf("%s: " str "\n", __FUNCTION__ , ## v) 49 #else 50 # define DEBUG(str, v...) ((void)0) 51 #endif 52 53 54 /* 55 * A container that lets us sort a superblock to optimize the skip step jumps 56 */ 57 struct pf_skip_step { 58 int ps_count; /* number of items */ 59 TAILQ_HEAD( , pf_opt_rule) ps_rules; 60 TAILQ_ENTRY(pf_skip_step) ps_entry; 61 }; 62 63 64 /* 65 * A superblock is a block of adjacent rules of similar action. If there 66 * are five PASS rules in a row, they all become members of a superblock. 67 * Once we have a superblock, we are free to re-order any rules within it 68 * in order to improve performance; if a packet is passed, it doesn't matter 69 * who passed it. 70 */ 71 struct superblock { 72 TAILQ_HEAD( , pf_opt_rule) sb_rules; 73 TAILQ_ENTRY(superblock) sb_entry; 74 struct superblock *sb_profiled_block; 75 TAILQ_HEAD(skiplist, pf_skip_step) sb_skipsteps[PF_SKIP_COUNT]; 76 }; 77 TAILQ_HEAD(superblocks, superblock); 78 79 80 /* 81 * Description of the PF rule structure. 82 */ 83 enum { 84 BARRIER, /* the presence of the field puts the rule in it's own block */ 85 BREAK, /* the field may not differ between rules in a superblock */ 86 NOMERGE, /* the field may not differ between rules when combined */ 87 COMBINED, /* the field may itself be combined with other rules */ 88 DC, /* we just don't care about the field */ 89 NEVER}; /* we should never see this field set?!? */ 90 struct pf_rule_field { 91 const char *prf_name; 92 int prf_type; 93 size_t prf_offset; 94 size_t prf_size; 95 } pf_rule_desc[] = { 96 #define PF_RULE_FIELD(field, ty) \ 97 {#field, \ 98 ty, \ 99 offsetof(struct pf_rule, field), \ 100 sizeof(((struct pf_rule *)0)->field)} 101 102 103 /* 104 * The presence of these fields in a rule put the rule in it's own 105 * superblock. Thus it will not be optimized. It also prevents the 106 * rule from being re-ordered at all. 107 */ 108 PF_RULE_FIELD(label, BARRIER), 109 PF_RULE_FIELD(prob, BARRIER), 110 PF_RULE_FIELD(max_states, BARRIER), 111 PF_RULE_FIELD(max_src_nodes, BARRIER), 112 PF_RULE_FIELD(max_src_states, BARRIER), 113 PF_RULE_FIELD(max_src_conn, BARRIER), 114 PF_RULE_FIELD(max_src_conn_rate, BARRIER), 115 PF_RULE_FIELD(anchor, BARRIER), /* for now */ 116 117 /* 118 * These fields must be the same between all rules in the same superblock. 119 * These rules are allowed to be re-ordered but only among like rules. 120 * For instance we can re-order all 'tag "foo"' rules because they have the 121 * same tag. But we can not re-order between a 'tag "foo"' and a 122 * 'tag "bar"' since that would change the meaning of the ruleset. 123 */ 124 PF_RULE_FIELD(tagname, BREAK), 125 PF_RULE_FIELD(keep_state, BREAK), 126 PF_RULE_FIELD(qname, BREAK), 127 PF_RULE_FIELD(pqname, BREAK), 128 PF_RULE_FIELD(rt, BREAK), 129 PF_RULE_FIELD(allow_opts, BREAK), 130 PF_RULE_FIELD(rule_flag, BREAK), 131 PF_RULE_FIELD(action, BREAK), 132 PF_RULE_FIELD(log, BREAK), 133 PF_RULE_FIELD(quick, BREAK), 134 PF_RULE_FIELD(return_ttl, BREAK), 135 PF_RULE_FIELD(overload_tblname, BREAK), 136 PF_RULE_FIELD(flush, BREAK), 137 PF_RULE_FIELD(rpool, BREAK), 138 PF_RULE_FIELD(logif, BREAK), 139 140 /* 141 * Any fields not listed in this structure act as BREAK fields 142 */ 143 144 145 /* 146 * These fields must not differ when we merge two rules together but 147 * their difference isn't enough to put the rules in different superblocks. 148 * There are no problems re-ordering any rules with these fields. 149 */ 150 PF_RULE_FIELD(af, NOMERGE), 151 PF_RULE_FIELD(ifnot, NOMERGE), 152 PF_RULE_FIELD(ifname, NOMERGE), /* hack for IF groups */ 153 PF_RULE_FIELD(match_tag_not, NOMERGE), 154 PF_RULE_FIELD(match_tagname, NOMERGE), 155 PF_RULE_FIELD(os_fingerprint, NOMERGE), 156 PF_RULE_FIELD(timeout, NOMERGE), 157 PF_RULE_FIELD(return_icmp, NOMERGE), 158 PF_RULE_FIELD(return_icmp6, NOMERGE), 159 PF_RULE_FIELD(uid, NOMERGE), 160 PF_RULE_FIELD(gid, NOMERGE), 161 PF_RULE_FIELD(direction, NOMERGE), 162 PF_RULE_FIELD(proto, NOMERGE), 163 PF_RULE_FIELD(type, NOMERGE), 164 PF_RULE_FIELD(code, NOMERGE), 165 PF_RULE_FIELD(flags, NOMERGE), 166 PF_RULE_FIELD(flagset, NOMERGE), 167 PF_RULE_FIELD(tos, NOMERGE), 168 PF_RULE_FIELD(src.port, NOMERGE), 169 PF_RULE_FIELD(dst.port, NOMERGE), 170 PF_RULE_FIELD(src.port_op, NOMERGE), 171 PF_RULE_FIELD(dst.port_op, NOMERGE), 172 PF_RULE_FIELD(src.neg, NOMERGE), 173 PF_RULE_FIELD(dst.neg, NOMERGE), 174 175 /* These fields can be merged */ 176 PF_RULE_FIELD(src.addr, COMBINED), 177 PF_RULE_FIELD(dst.addr, COMBINED), 178 179 /* We just don't care about these fields. They're set by the kernel */ 180 PF_RULE_FIELD(skip, DC), 181 PF_RULE_FIELD(evaluations, DC), 182 PF_RULE_FIELD(packets, DC), 183 PF_RULE_FIELD(bytes, DC), 184 PF_RULE_FIELD(kif, DC), 185 PF_RULE_FIELD(states, DC), 186 PF_RULE_FIELD(src_nodes, DC), 187 PF_RULE_FIELD(nr, DC), 188 PF_RULE_FIELD(entries, DC), 189 PF_RULE_FIELD(qid, DC), 190 PF_RULE_FIELD(pqid, DC), 191 PF_RULE_FIELD(anchor_relative, DC), 192 PF_RULE_FIELD(anchor_wildcard, DC), 193 PF_RULE_FIELD(tag, DC), 194 PF_RULE_FIELD(match_tag, DC), 195 PF_RULE_FIELD(overload_tbl, DC), 196 197 /* These fields should never be set in a PASS/BLOCK rule */ 198 PF_RULE_FIELD(natpass, NEVER), 199 PF_RULE_FIELD(max_mss, NEVER), 200 PF_RULE_FIELD(min_ttl, NEVER), 201 }; 202 203 204 205 int add_opt_table(struct pfctl *, struct pf_opt_tbl **, sa_family_t, 206 struct pf_rule_addr *); 207 int addrs_combineable(struct pf_rule_addr *, struct pf_rule_addr *); 208 int addrs_equal(struct pf_rule_addr *, struct pf_rule_addr *); 209 int block_feedback(struct pfctl *, struct superblock *); 210 int combine_rules(struct pfctl *, struct superblock *); 211 void comparable_rule(struct pf_rule *, const struct pf_rule *, int); 212 int construct_superblocks(struct pfctl *, struct pf_opt_queue *, 213 struct superblocks *); 214 void exclude_supersets(struct pf_rule *, struct pf_rule *); 215 int interface_group(const char *); 216 int load_feedback_profile(struct pfctl *, struct superblocks *); 217 int optimize_superblock(struct pfctl *, struct superblock *); 218 int pf_opt_create_table(struct pfctl *, struct pf_opt_tbl *); 219 void remove_from_skipsteps(struct skiplist *, struct superblock *, 220 struct pf_opt_rule *, struct pf_skip_step *); 221 int remove_identical_rules(struct pfctl *, struct superblock *); 222 int reorder_rules(struct pfctl *, struct superblock *, int); 223 int rules_combineable(struct pf_rule *, struct pf_rule *); 224 void skip_append(struct superblock *, int, struct pf_skip_step *, 225 struct pf_opt_rule *); 226 int skip_compare(int, struct pf_skip_step *, struct pf_opt_rule *); 227 void skip_init(void); 228 int skip_cmp_af(struct pf_rule *, struct pf_rule *); 229 int skip_cmp_dir(struct pf_rule *, struct pf_rule *); 230 int skip_cmp_dst_addr(struct pf_rule *, struct pf_rule *); 231 int skip_cmp_dst_port(struct pf_rule *, struct pf_rule *); 232 int skip_cmp_ifp(struct pf_rule *, struct pf_rule *); 233 int skip_cmp_proto(struct pf_rule *, struct pf_rule *); 234 int skip_cmp_src_addr(struct pf_rule *, struct pf_rule *); 235 int skip_cmp_src_port(struct pf_rule *, struct pf_rule *); 236 int superblock_inclusive(struct superblock *, struct pf_opt_rule *); 237 void superblock_free(struct pfctl *, struct superblock *); 238 239 240 int (*skip_comparitors[PF_SKIP_COUNT])(struct pf_rule *, struct pf_rule *); 241 const char *skip_comparitors_names[PF_SKIP_COUNT]; 242 #define PF_SKIP_COMPARITORS { \ 243 { "ifp", PF_SKIP_IFP, skip_cmp_ifp }, \ 244 { "dir", PF_SKIP_DIR, skip_cmp_dir }, \ 245 { "af", PF_SKIP_AF, skip_cmp_af }, \ 246 { "proto", PF_SKIP_PROTO, skip_cmp_proto }, \ 247 { "saddr", PF_SKIP_SRC_ADDR, skip_cmp_src_addr }, \ 248 { "sport", PF_SKIP_SRC_PORT, skip_cmp_src_port }, \ 249 { "daddr", PF_SKIP_DST_ADDR, skip_cmp_dst_addr }, \ 250 { "dport", PF_SKIP_DST_PORT, skip_cmp_dst_port } \ 251 } 252 253 struct pfr_buffer table_buffer; 254 int table_identifier; 255 256 257 int 258 pfctl_optimize_ruleset(struct pfctl *pf, struct pf_ruleset *rs) 259 { 260 struct superblocks superblocks; 261 struct pf_opt_queue opt_queue; 262 struct superblock *block; 263 struct pf_opt_rule *por; 264 struct pf_rule *r; 265 struct pf_rulequeue *old_rules; 266 267 DEBUG("optimizing ruleset"); 268 memset(&table_buffer, 0, sizeof(table_buffer)); 269 skip_init(); 270 TAILQ_INIT(&opt_queue); 271 272 old_rules = rs->rules[PF_RULESET_FILTER].active.ptr; 273 rs->rules[PF_RULESET_FILTER].active.ptr = 274 rs->rules[PF_RULESET_FILTER].inactive.ptr; 275 rs->rules[PF_RULESET_FILTER].inactive.ptr = old_rules; 276 277 /* 278 * XXX expanding the pf_opt_rule format throughout pfctl might allow 279 * us to avoid all this copying. 280 */ 281 while ((r = TAILQ_FIRST(rs->rules[PF_RULESET_FILTER].inactive.ptr)) 282 != NULL) { 283 TAILQ_REMOVE(rs->rules[PF_RULESET_FILTER].inactive.ptr, r, 284 entries); 285 if ((por = calloc(1, sizeof(*por))) == NULL) 286 err(1, "calloc"); 287 memcpy(&por->por_rule, r, sizeof(*r)); 288 if (TAILQ_FIRST(&r->rpool.list) != NULL) { 289 TAILQ_INIT(&por->por_rule.rpool.list); 290 pfctl_move_pool(&r->rpool, &por->por_rule.rpool); 291 } else 292 bzero(&por->por_rule.rpool, 293 sizeof(por->por_rule.rpool)); 294 295 296 TAILQ_INSERT_TAIL(&opt_queue, por, por_entry); 297 } 298 299 TAILQ_INIT(&superblocks); 300 if (construct_superblocks(pf, &opt_queue, &superblocks)) 301 goto error; 302 303 if (pf->optimize & PF_OPTIMIZE_PROFILE) { 304 if (load_feedback_profile(pf, &superblocks)) 305 goto error; 306 } 307 308 TAILQ_FOREACH(block, &superblocks, sb_entry) { 309 if (optimize_superblock(pf, block)) 310 goto error; 311 } 312 313 rs->anchor->refcnt = 0; 314 while ((block = TAILQ_FIRST(&superblocks))) { 315 TAILQ_REMOVE(&superblocks, block, sb_entry); 316 317 while ((por = TAILQ_FIRST(&block->sb_rules))) { 318 TAILQ_REMOVE(&block->sb_rules, por, por_entry); 319 por->por_rule.nr = rs->anchor->refcnt++; 320 if ((r = calloc(1, sizeof(*r))) == NULL) 321 err(1, "calloc"); 322 memcpy(r, &por->por_rule, sizeof(*r)); 323 TAILQ_INIT(&r->rpool.list); 324 pfctl_move_pool(&por->por_rule.rpool, &r->rpool); 325 TAILQ_INSERT_TAIL( 326 rs->rules[PF_RULESET_FILTER].active.ptr, 327 r, entries); 328 free(por); 329 } 330 free(block); 331 } 332 333 return (0); 334 335 error: 336 while ((por = TAILQ_FIRST(&opt_queue))) { 337 TAILQ_REMOVE(&opt_queue, por, por_entry); 338 if (por->por_src_tbl) { 339 pfr_buf_clear(por->por_src_tbl->pt_buf); 340 free(por->por_src_tbl->pt_buf); 341 free(por->por_src_tbl); 342 } 343 if (por->por_dst_tbl) { 344 pfr_buf_clear(por->por_dst_tbl->pt_buf); 345 free(por->por_dst_tbl->pt_buf); 346 free(por->por_dst_tbl); 347 } 348 free(por); 349 } 350 while ((block = TAILQ_FIRST(&superblocks))) { 351 TAILQ_REMOVE(&superblocks, block, sb_entry); 352 superblock_free(pf, block); 353 } 354 return (1); 355 } 356 357 358 /* 359 * Go ahead and optimize a superblock 360 */ 361 int 362 optimize_superblock(struct pfctl *pf, struct superblock *block) 363 { 364 #ifdef OPT_DEBUG 365 struct pf_opt_rule *por; 366 #endif /* OPT_DEBUG */ 367 368 /* We have a few optimization passes: 369 * 1) remove duplicate rules or rules that are a subset of other 370 * rules 371 * 2) combine otherwise identical rules with different IP addresses 372 * into a single rule and put the addresses in a table. 373 * 3) re-order the rules to improve kernel skip steps 374 * 4) re-order the 'quick' rules based on feedback from the 375 * active ruleset statistics 376 * 377 * XXX combine_rules() doesn't combine v4 and v6 rules. would just 378 * have to keep af in the table container, make af 'COMBINE' and 379 * twiddle the af on the merged rule 380 * XXX maybe add a weighting to the metric on skipsteps when doing 381 * reordering. sometimes two sequential tables will be better 382 * that four consecutive interfaces. 383 * XXX need to adjust the skipstep count of everything after PROTO, 384 * since they aren't actually checked on a proto mismatch in 385 * pf_test_{tcp, udp, icmp}() 386 * XXX should i treat proto=0, af=0 or dir=0 special in skepstep 387 * calculation since they are a DC? 388 * XXX keep last skiplist of last superblock to influence this 389 * superblock. '5 inet6 log' should make '3 inet6' come before '4 390 * inet' in the next superblock. 391 * XXX would be useful to add tables for ports 392 * XXX we can also re-order some mutually exclusive superblocks to 393 * try merging superblocks before any of these optimization passes. 394 * for instance a single 'log in' rule in the middle of non-logging 395 * out rules. 396 */ 397 398 /* shortcut. there will be alot of 1-rule superblocks */ 399 if (!TAILQ_NEXT(TAILQ_FIRST(&block->sb_rules), por_entry)) 400 return (0); 401 402 #ifdef OPT_DEBUG 403 printf("--- Superblock ---\n"); 404 TAILQ_FOREACH(por, &block->sb_rules, por_entry) { 405 printf(" "); 406 print_rule(&por->por_rule, por->por_rule.anchor ? 407 por->por_rule.anchor->name : "", 1); 408 } 409 #endif /* OPT_DEBUG */ 410 411 412 if (remove_identical_rules(pf, block)) 413 return (1); 414 if (combine_rules(pf, block)) 415 return (1); 416 if ((pf->optimize & PF_OPTIMIZE_PROFILE) && 417 TAILQ_FIRST(&block->sb_rules)->por_rule.quick && 418 block->sb_profiled_block) { 419 if (block_feedback(pf, block)) 420 return (1); 421 } else if (reorder_rules(pf, block, 0)) { 422 return (1); 423 } 424 425 /* 426 * Don't add any optimization passes below reorder_rules(). It will 427 * have divided superblocks into smaller blocks for further refinement 428 * and doesn't put them back together again. What once was a true 429 * superblock might have been split into multiple superblocks. 430 */ 431 432 #ifdef OPT_DEBUG 433 printf("--- END Superblock ---\n"); 434 #endif /* OPT_DEBUG */ 435 return (0); 436 } 437 438 439 /* 440 * Optimization pass #1: remove identical rules 441 */ 442 int 443 remove_identical_rules(struct pfctl *pf, struct superblock *block) 444 { 445 struct pf_opt_rule *por1, *por2, *por_next, *por2_next; 446 struct pf_rule a, a2, b, b2; 447 448 for (por1 = TAILQ_FIRST(&block->sb_rules); por1; por1 = por_next) { 449 por_next = TAILQ_NEXT(por1, por_entry); 450 for (por2 = por_next; por2; por2 = por2_next) { 451 por2_next = TAILQ_NEXT(por2, por_entry); 452 comparable_rule(&a, &por1->por_rule, DC); 453 comparable_rule(&b, &por2->por_rule, DC); 454 memcpy(&a2, &a, sizeof(a2)); 455 memcpy(&b2, &b, sizeof(b2)); 456 457 exclude_supersets(&a, &b); 458 exclude_supersets(&b2, &a2); 459 if (memcmp(&a, &b, sizeof(a)) == 0) { 460 DEBUG("removing identical rule nr%d = *nr%d*", 461 por1->por_rule.nr, por2->por_rule.nr); 462 TAILQ_REMOVE(&block->sb_rules, por2, por_entry); 463 if (por_next == por2) 464 por_next = TAILQ_NEXT(por1, por_entry); 465 free(por2); 466 } else if (memcmp(&a2, &b2, sizeof(a2)) == 0) { 467 DEBUG("removing identical rule *nr%d* = nr%d", 468 por1->por_rule.nr, por2->por_rule.nr); 469 TAILQ_REMOVE(&block->sb_rules, por1, por_entry); 470 free(por1); 471 break; 472 } 473 } 474 } 475 476 return (0); 477 } 478 479 480 /* 481 * Optimization pass #2: combine similar rules with different addresses 482 * into a single rule and a table 483 */ 484 int 485 combine_rules(struct pfctl *pf, struct superblock *block) 486 { 487 struct pf_opt_rule *p1, *p2, *por_next; 488 int src_eq, dst_eq; 489 490 if ((pf->loadopt & PFCTL_FLAG_TABLE) == 0) { 491 warnx("Must enable table loading for optimizations"); 492 return (1); 493 } 494 495 /* First we make a pass to combine the rules. O(n log n) */ 496 TAILQ_FOREACH(p1, &block->sb_rules, por_entry) { 497 for (p2 = TAILQ_NEXT(p1, por_entry); p2; p2 = por_next) { 498 por_next = TAILQ_NEXT(p2, por_entry); 499 500 src_eq = addrs_equal(&p1->por_rule.src, 501 &p2->por_rule.src); 502 dst_eq = addrs_equal(&p1->por_rule.dst, 503 &p2->por_rule.dst); 504 505 if (src_eq && !dst_eq && p1->por_src_tbl == NULL && 506 p2->por_dst_tbl == NULL && 507 p2->por_src_tbl == NULL && 508 rules_combineable(&p1->por_rule, &p2->por_rule) && 509 addrs_combineable(&p1->por_rule.dst, 510 &p2->por_rule.dst)) { 511 DEBUG("can combine rules nr%d = nr%d", 512 p1->por_rule.nr, p2->por_rule.nr); 513 if (p1->por_dst_tbl == NULL && 514 add_opt_table(pf, &p1->por_dst_tbl, 515 p1->por_rule.af, &p1->por_rule.dst)) 516 return (1); 517 if (add_opt_table(pf, &p1->por_dst_tbl, 518 p1->por_rule.af, &p2->por_rule.dst)) 519 return (1); 520 p2->por_dst_tbl = p1->por_dst_tbl; 521 if (p1->por_dst_tbl->pt_rulecount >= 522 TABLE_THRESHOLD) { 523 TAILQ_REMOVE(&block->sb_rules, p2, 524 por_entry); 525 free(p2); 526 } 527 } else if (!src_eq && dst_eq && p1->por_dst_tbl == NULL 528 && p2->por_src_tbl == NULL && 529 p2->por_dst_tbl == NULL && 530 rules_combineable(&p1->por_rule, &p2->por_rule) && 531 addrs_combineable(&p1->por_rule.src, 532 &p2->por_rule.src)) { 533 DEBUG("can combine rules nr%d = nr%d", 534 p1->por_rule.nr, p2->por_rule.nr); 535 if (p1->por_src_tbl == NULL && 536 add_opt_table(pf, &p1->por_src_tbl, 537 p1->por_rule.af, &p1->por_rule.src)) 538 return (1); 539 if (add_opt_table(pf, &p1->por_src_tbl, 540 p1->por_rule.af, &p2->por_rule.src)) 541 return (1); 542 p2->por_src_tbl = p1->por_src_tbl; 543 if (p1->por_src_tbl->pt_rulecount >= 544 TABLE_THRESHOLD) { 545 TAILQ_REMOVE(&block->sb_rules, p2, 546 por_entry); 547 free(p2); 548 } 549 } 550 } 551 } 552 553 554 /* 555 * Then we make a final pass to create a valid table name and 556 * insert the name into the rules. 557 */ 558 for (p1 = TAILQ_FIRST(&block->sb_rules); p1; p1 = por_next) { 559 por_next = TAILQ_NEXT(p1, por_entry); 560 assert(p1->por_src_tbl == NULL || p1->por_dst_tbl == NULL); 561 562 if (p1->por_src_tbl && p1->por_src_tbl->pt_rulecount >= 563 TABLE_THRESHOLD) { 564 if (p1->por_src_tbl->pt_generated) { 565 /* This rule is included in a table */ 566 TAILQ_REMOVE(&block->sb_rules, p1, por_entry); 567 free(p1); 568 continue; 569 } 570 p1->por_src_tbl->pt_generated = 1; 571 572 if ((pf->opts & PF_OPT_NOACTION) == 0 && 573 pf_opt_create_table(pf, p1->por_src_tbl)) 574 return (1); 575 576 pf->tdirty = 1; 577 578 if (pf->opts & PF_OPT_VERBOSE) 579 print_tabledef(p1->por_src_tbl->pt_name, 580 PFR_TFLAG_CONST, 1, 581 &p1->por_src_tbl->pt_nodes); 582 583 memset(&p1->por_rule.src.addr, 0, 584 sizeof(p1->por_rule.src.addr)); 585 p1->por_rule.src.addr.type = PF_ADDR_TABLE; 586 strlcpy(p1->por_rule.src.addr.v.tblname, 587 p1->por_src_tbl->pt_name, 588 sizeof(p1->por_rule.src.addr.v.tblname)); 589 590 pfr_buf_clear(p1->por_src_tbl->pt_buf); 591 free(p1->por_src_tbl->pt_buf); 592 p1->por_src_tbl->pt_buf = NULL; 593 } 594 if (p1->por_dst_tbl && p1->por_dst_tbl->pt_rulecount >= 595 TABLE_THRESHOLD) { 596 if (p1->por_dst_tbl->pt_generated) { 597 /* This rule is included in a table */ 598 TAILQ_REMOVE(&block->sb_rules, p1, por_entry); 599 free(p1); 600 continue; 601 } 602 p1->por_dst_tbl->pt_generated = 1; 603 604 if ((pf->opts & PF_OPT_NOACTION) == 0 && 605 pf_opt_create_table(pf, p1->por_dst_tbl)) 606 return (1); 607 pf->tdirty = 1; 608 609 if (pf->opts & PF_OPT_VERBOSE) 610 print_tabledef(p1->por_dst_tbl->pt_name, 611 PFR_TFLAG_CONST, 1, 612 &p1->por_dst_tbl->pt_nodes); 613 614 memset(&p1->por_rule.dst.addr, 0, 615 sizeof(p1->por_rule.dst.addr)); 616 p1->por_rule.dst.addr.type = PF_ADDR_TABLE; 617 strlcpy(p1->por_rule.dst.addr.v.tblname, 618 p1->por_dst_tbl->pt_name, 619 sizeof(p1->por_rule.dst.addr.v.tblname)); 620 621 pfr_buf_clear(p1->por_dst_tbl->pt_buf); 622 free(p1->por_dst_tbl->pt_buf); 623 p1->por_dst_tbl->pt_buf = NULL; 624 } 625 } 626 627 return (0); 628 } 629 630 631 /* 632 * Optimization pass #3: re-order rules to improve skip steps 633 */ 634 int 635 reorder_rules(struct pfctl *pf, struct superblock *block, int depth) 636 { 637 struct superblock *newblock; 638 struct pf_skip_step *skiplist; 639 struct pf_opt_rule *por; 640 int i, largest, largest_list, rule_count = 0; 641 TAILQ_HEAD( , pf_opt_rule) head; 642 643 /* 644 * Calculate the best-case skip steps. We put each rule in a list 645 * of other rules with common fields 646 */ 647 for (i = 0; i < PF_SKIP_COUNT; i++) { 648 TAILQ_FOREACH(por, &block->sb_rules, por_entry) { 649 TAILQ_FOREACH(skiplist, &block->sb_skipsteps[i], 650 ps_entry) { 651 if (skip_compare(i, skiplist, por) == 0) 652 break; 653 } 654 if (skiplist == NULL) { 655 if ((skiplist = calloc(1, sizeof(*skiplist))) == 656 NULL) 657 err(1, "calloc"); 658 TAILQ_INIT(&skiplist->ps_rules); 659 TAILQ_INSERT_TAIL(&block->sb_skipsteps[i], 660 skiplist, ps_entry); 661 } 662 skip_append(block, i, skiplist, por); 663 } 664 } 665 666 TAILQ_FOREACH(por, &block->sb_rules, por_entry) 667 rule_count++; 668 669 /* 670 * Now we're going to ignore any fields that are identical between 671 * all of the rules in the superblock and those fields which differ 672 * between every rule in the superblock. 673 */ 674 largest = 0; 675 for (i = 0; i < PF_SKIP_COUNT; i++) { 676 skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]); 677 if (skiplist->ps_count == rule_count) { 678 DEBUG("(%d) original skipstep '%s' is all rules", 679 depth, skip_comparitors_names[i]); 680 skiplist->ps_count = 0; 681 } else if (skiplist->ps_count == 1) { 682 skiplist->ps_count = 0; 683 } else { 684 DEBUG("(%d) original skipstep '%s' largest jump is %d", 685 depth, skip_comparitors_names[i], 686 skiplist->ps_count); 687 if (skiplist->ps_count > largest) 688 largest = skiplist->ps_count; 689 } 690 } 691 if (largest == 0) { 692 /* Ugh. There is NO commonality in the superblock on which 693 * optimize the skipsteps optimization. 694 */ 695 goto done; 696 } 697 698 /* 699 * Now we're going to empty the superblock rule list and re-create 700 * it based on a more optimal skipstep order. 701 */ 702 TAILQ_INIT(&head); 703 while ((por = TAILQ_FIRST(&block->sb_rules))) { 704 TAILQ_REMOVE(&block->sb_rules, por, por_entry); 705 TAILQ_INSERT_TAIL(&head, por, por_entry); 706 } 707 708 709 while (!TAILQ_EMPTY(&head)) { 710 largest = 1; 711 712 /* 713 * Find the most useful skip steps remaining 714 */ 715 for (i = 0; i < PF_SKIP_COUNT; i++) { 716 skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]); 717 if (skiplist->ps_count > largest) { 718 largest = skiplist->ps_count; 719 largest_list = i; 720 } 721 } 722 723 if (largest <= 1) { 724 /* 725 * Nothing useful left. Leave remaining rules in order. 726 */ 727 DEBUG("(%d) no more commonality for skip steps", depth); 728 while ((por = TAILQ_FIRST(&head))) { 729 TAILQ_REMOVE(&head, por, por_entry); 730 TAILQ_INSERT_TAIL(&block->sb_rules, por, 731 por_entry); 732 } 733 } else { 734 /* 735 * There is commonality. Extract those common rules 736 * and place them in the ruleset adjacent to each 737 * other. 738 */ 739 skiplist = TAILQ_FIRST(&block->sb_skipsteps[ 740 largest_list]); 741 DEBUG("(%d) skipstep '%s' largest jump is %d @ #%d", 742 depth, skip_comparitors_names[largest_list], 743 largest, TAILQ_FIRST(&TAILQ_FIRST(&block-> 744 sb_skipsteps [largest_list])->ps_rules)-> 745 por_rule.nr); 746 TAILQ_REMOVE(&block->sb_skipsteps[largest_list], 747 skiplist, ps_entry); 748 749 750 /* 751 * There may be further commonality inside these 752 * rules. So we'll split them off into they're own 753 * superblock and pass it back into the optimizer. 754 */ 755 if (skiplist->ps_count > 2) { 756 if ((newblock = calloc(1, sizeof(*newblock))) 757 == NULL) { 758 warn("calloc"); 759 return (1); 760 } 761 TAILQ_INIT(&newblock->sb_rules); 762 for (i = 0; i < PF_SKIP_COUNT; i++) 763 TAILQ_INIT(&newblock->sb_skipsteps[i]); 764 TAILQ_INSERT_BEFORE(block, newblock, sb_entry); 765 DEBUG("(%d) splitting off %d rules from superblock @ #%d", 766 depth, skiplist->ps_count, 767 TAILQ_FIRST(&skiplist->ps_rules)-> 768 por_rule.nr); 769 } else { 770 newblock = block; 771 } 772 773 while ((por = TAILQ_FIRST(&skiplist->ps_rules))) { 774 TAILQ_REMOVE(&head, por, por_entry); 775 TAILQ_REMOVE(&skiplist->ps_rules, por, 776 por_skip_entry[largest_list]); 777 TAILQ_INSERT_TAIL(&newblock->sb_rules, por, 778 por_entry); 779 780 /* Remove this rule from all other skiplists */ 781 remove_from_skipsteps(&block->sb_skipsteps[ 782 largest_list], block, por, skiplist); 783 } 784 free(skiplist); 785 if (newblock != block) 786 if (reorder_rules(pf, newblock, depth + 1)) 787 return (1); 788 } 789 } 790 791 done: 792 for (i = 0; i < PF_SKIP_COUNT; i++) { 793 while ((skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]))) { 794 TAILQ_REMOVE(&block->sb_skipsteps[i], skiplist, 795 ps_entry); 796 free(skiplist); 797 } 798 } 799 800 return (0); 801 } 802 803 804 /* 805 * Optimization pass #4: re-order 'quick' rules based on feedback from the 806 * currently running ruleset 807 */ 808 int 809 block_feedback(struct pfctl *pf, struct superblock *block) 810 { 811 TAILQ_HEAD( , pf_opt_rule) queue; 812 struct pf_opt_rule *por1, *por2; 813 u_int64_t total_count = 0; 814 struct pf_rule a, b; 815 816 817 /* 818 * Walk through all of the profiled superblock's rules and copy 819 * the counters onto our rules. 820 */ 821 TAILQ_FOREACH(por1, &block->sb_profiled_block->sb_rules, por_entry) { 822 comparable_rule(&a, &por1->por_rule, DC); 823 total_count += por1->por_rule.packets[0] + 824 por1->por_rule.packets[1]; 825 TAILQ_FOREACH(por2, &block->sb_rules, por_entry) { 826 if (por2->por_profile_count) 827 continue; 828 comparable_rule(&b, &por2->por_rule, DC); 829 if (memcmp(&a, &b, sizeof(a)) == 0) { 830 por2->por_profile_count = 831 por1->por_rule.packets[0] + 832 por1->por_rule.packets[1]; 833 break; 834 } 835 } 836 } 837 superblock_free(pf, block->sb_profiled_block); 838 block->sb_profiled_block = NULL; 839 840 /* 841 * Now we pull all of the rules off the superblock and re-insert them 842 * in sorted order. 843 */ 844 845 TAILQ_INIT(&queue); 846 while ((por1 = TAILQ_FIRST(&block->sb_rules)) != NULL) { 847 TAILQ_REMOVE(&block->sb_rules, por1, por_entry); 848 TAILQ_INSERT_TAIL(&queue, por1, por_entry); 849 } 850 851 while ((por1 = TAILQ_FIRST(&queue)) != NULL) { 852 TAILQ_REMOVE(&queue, por1, por_entry); 853 /* XXX I should sort all of the unused rules based on skip steps */ 854 TAILQ_FOREACH(por2, &block->sb_rules, por_entry) { 855 if (por1->por_profile_count > por2->por_profile_count) { 856 TAILQ_INSERT_BEFORE(por2, por1, por_entry); 857 break; 858 } 859 } 860 if (por2 == NULL) 861 TAILQ_INSERT_TAIL(&block->sb_rules, por1, por_entry); 862 } 863 864 return (0); 865 } 866 867 868 /* 869 * Load the current ruleset from the kernel and try to associate them with 870 * the ruleset we're optimizing. 871 */ 872 int 873 load_feedback_profile(struct pfctl *pf, struct superblocks *superblocks) 874 { 875 struct superblock *block, *blockcur; 876 struct superblocks prof_superblocks; 877 struct pf_opt_rule *por; 878 struct pf_opt_queue queue; 879 struct pfioc_rule pr; 880 struct pf_rule a, b; 881 int nr, mnr; 882 883 TAILQ_INIT(&queue); 884 TAILQ_INIT(&prof_superblocks); 885 886 memset(&pr, 0, sizeof(pr)); 887 pr.rule.action = PF_PASS; 888 if (ioctl(pf->dev, DIOCGETRULES, &pr)) { 889 warn("DIOCGETRULES"); 890 return (1); 891 } 892 mnr = pr.nr; 893 894 DEBUG("Loading %d active rules for a feedback profile", mnr); 895 for (nr = 0; nr < mnr; ++nr) { 896 struct pf_ruleset *rs; 897 if ((por = calloc(1, sizeof(*por))) == NULL) { 898 warn("calloc"); 899 return (1); 900 } 901 pr.nr = nr; 902 if (ioctl(pf->dev, DIOCGETRULE, &pr)) { 903 warn("DIOCGETRULES"); 904 return (1); 905 } 906 memcpy(&por->por_rule, &pr.rule, sizeof(por->por_rule)); 907 rs = pf_find_or_create_ruleset(pr.anchor_call); 908 por->por_rule.anchor = rs->anchor; 909 if (TAILQ_EMPTY(&por->por_rule.rpool.list)) 910 memset(&por->por_rule.rpool, 0, 911 sizeof(por->por_rule.rpool)); 912 TAILQ_INSERT_TAIL(&queue, por, por_entry); 913 914 /* XXX pfctl_get_pool(pf->dev, &pr.rule.rpool, nr, pr.ticket, 915 * PF_PASS, pf->anchor) ??? 916 * ... pfctl_clear_pool(&pr.rule.rpool) 917 */ 918 } 919 920 if (construct_superblocks(pf, &queue, &prof_superblocks)) 921 return (1); 922 923 924 /* 925 * Now we try to associate the active ruleset's superblocks with 926 * the superblocks we're compiling. 927 */ 928 block = TAILQ_FIRST(superblocks); 929 blockcur = TAILQ_FIRST(&prof_superblocks); 930 while (block && blockcur) { 931 comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule, 932 BREAK); 933 comparable_rule(&b, &TAILQ_FIRST(&blockcur->sb_rules)->por_rule, 934 BREAK); 935 if (memcmp(&a, &b, sizeof(a)) == 0) { 936 /* The two superblocks lined up */ 937 block->sb_profiled_block = blockcur; 938 } else { 939 DEBUG("superblocks don't line up between #%d and #%d", 940 TAILQ_FIRST(&block->sb_rules)->por_rule.nr, 941 TAILQ_FIRST(&blockcur->sb_rules)->por_rule.nr); 942 break; 943 } 944 block = TAILQ_NEXT(block, sb_entry); 945 blockcur = TAILQ_NEXT(blockcur, sb_entry); 946 } 947 948 949 950 /* Free any superblocks we couldn't link */ 951 while (blockcur) { 952 block = TAILQ_NEXT(blockcur, sb_entry); 953 superblock_free(pf, blockcur); 954 blockcur = block; 955 } 956 return (0); 957 } 958 959 960 /* 961 * Compare a rule to a skiplist to see if the rule is a member 962 */ 963 int 964 skip_compare(int skipnum, struct pf_skip_step *skiplist, 965 struct pf_opt_rule *por) 966 { 967 struct pf_rule *a, *b; 968 if (skipnum >= PF_SKIP_COUNT || skipnum < 0) 969 errx(1, "skip_compare() out of bounds"); 970 a = &por->por_rule; 971 b = &TAILQ_FIRST(&skiplist->ps_rules)->por_rule; 972 973 return ((skip_comparitors[skipnum])(a, b)); 974 } 975 976 977 /* 978 * Add a rule to a skiplist 979 */ 980 void 981 skip_append(struct superblock *superblock, int skipnum, 982 struct pf_skip_step *skiplist, struct pf_opt_rule *por) 983 { 984 struct pf_skip_step *prev; 985 986 skiplist->ps_count++; 987 TAILQ_INSERT_TAIL(&skiplist->ps_rules, por, por_skip_entry[skipnum]); 988 989 /* Keep the list of skiplists sorted by whichever is larger */ 990 while ((prev = TAILQ_PREV(skiplist, skiplist, ps_entry)) && 991 prev->ps_count < skiplist->ps_count) { 992 TAILQ_REMOVE(&superblock->sb_skipsteps[skipnum], 993 skiplist, ps_entry); 994 TAILQ_INSERT_BEFORE(prev, skiplist, ps_entry); 995 } 996 } 997 998 999 /* 1000 * Remove a rule from the other skiplist calculations. 1001 */ 1002 void 1003 remove_from_skipsteps(struct skiplist *head, struct superblock *block, 1004 struct pf_opt_rule *por, struct pf_skip_step *active_list) 1005 { 1006 struct pf_skip_step *sk, *next; 1007 struct pf_opt_rule *p2; 1008 int i, found; 1009 1010 for (i = 0; i < PF_SKIP_COUNT; i++) { 1011 sk = TAILQ_FIRST(&block->sb_skipsteps[i]); 1012 if (sk == NULL || sk == active_list || sk->ps_count <= 1) 1013 continue; 1014 found = 0; 1015 do { 1016 TAILQ_FOREACH(p2, &sk->ps_rules, por_skip_entry[i]) 1017 if (p2 == por) { 1018 TAILQ_REMOVE(&sk->ps_rules, p2, 1019 por_skip_entry[i]); 1020 found = 1; 1021 sk->ps_count--; 1022 break; 1023 } 1024 } while (!found && (sk = TAILQ_NEXT(sk, ps_entry))); 1025 if (found && sk) { 1026 /* Does this change the sorting order? */ 1027 while ((next = TAILQ_NEXT(sk, ps_entry)) && 1028 next->ps_count > sk->ps_count) { 1029 TAILQ_REMOVE(head, sk, ps_entry); 1030 TAILQ_INSERT_AFTER(head, next, sk, ps_entry); 1031 } 1032 #ifdef OPT_DEBUG 1033 next = TAILQ_NEXT(sk, ps_entry); 1034 assert(next == NULL || next->ps_count <= sk->ps_count); 1035 #endif /* OPT_DEBUG */ 1036 } 1037 } 1038 } 1039 1040 1041 /* Compare two rules AF field for skiplist construction */ 1042 int 1043 skip_cmp_af(struct pf_rule *a, struct pf_rule *b) 1044 { 1045 if (a->af != b->af || a->af == 0) 1046 return (1); 1047 return (0); 1048 } 1049 1050 /* Compare two rules DIRECTION field for skiplist construction */ 1051 int 1052 skip_cmp_dir(struct pf_rule *a, struct pf_rule *b) 1053 { 1054 if (a->direction == 0 || a->direction != b->direction) 1055 return (1); 1056 return (0); 1057 } 1058 1059 /* Compare two rules DST Address field for skiplist construction */ 1060 int 1061 skip_cmp_dst_addr(struct pf_rule *a, struct pf_rule *b) 1062 { 1063 if (a->dst.neg != b->dst.neg || 1064 a->dst.addr.type != b->dst.addr.type) 1065 return (1); 1066 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0 1067 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP || 1068 * a->proto == IPPROTO_ICMP 1069 * return (1); 1070 */ 1071 switch (a->dst.addr.type) { 1072 case PF_ADDR_ADDRMASK: 1073 if (memcmp(&a->dst.addr.v.a.addr, &b->dst.addr.v.a.addr, 1074 sizeof(a->dst.addr.v.a.addr)) || 1075 memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask, 1076 sizeof(a->dst.addr.v.a.mask)) || 1077 (a->dst.addr.v.a.addr.addr32[0] == 0 && 1078 a->dst.addr.v.a.addr.addr32[1] == 0 && 1079 a->dst.addr.v.a.addr.addr32[2] == 0 && 1080 a->dst.addr.v.a.addr.addr32[3] == 0)) 1081 return (1); 1082 return (0); 1083 case PF_ADDR_DYNIFTL: 1084 if (strcmp(a->dst.addr.v.ifname, b->dst.addr.v.ifname) != 0 || 1085 a->dst.addr.iflags != a->dst.addr.iflags || 1086 memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask, 1087 sizeof(a->dst.addr.v.a.mask))) 1088 return (1); 1089 return (0); 1090 case PF_ADDR_NOROUTE: 1091 case PF_ADDR_URPFFAILED: 1092 return (0); 1093 case PF_ADDR_TABLE: 1094 return (strcmp(a->dst.addr.v.tblname, b->dst.addr.v.tblname)); 1095 } 1096 return (1); 1097 } 1098 1099 /* Compare two rules DST port field for skiplist construction */ 1100 int 1101 skip_cmp_dst_port(struct pf_rule *a, struct pf_rule *b) 1102 { 1103 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0 1104 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP || 1105 * a->proto == IPPROTO_ICMP 1106 * return (1); 1107 */ 1108 if (a->dst.port_op == PF_OP_NONE || a->dst.port_op != b->dst.port_op || 1109 a->dst.port[0] != b->dst.port[0] || 1110 a->dst.port[1] != b->dst.port[1]) 1111 return (1); 1112 return (0); 1113 } 1114 1115 /* Compare two rules IFP field for skiplist construction */ 1116 int 1117 skip_cmp_ifp(struct pf_rule *a, struct pf_rule *b) 1118 { 1119 if (strcmp(a->ifname, b->ifname) || a->ifname[0] == '\0') 1120 return (1); 1121 return (a->ifnot != b->ifnot); 1122 } 1123 1124 /* Compare two rules PROTO field for skiplist construction */ 1125 int 1126 skip_cmp_proto(struct pf_rule *a, struct pf_rule *b) 1127 { 1128 return (a->proto != b->proto || a->proto == 0); 1129 } 1130 1131 /* Compare two rules SRC addr field for skiplist construction */ 1132 int 1133 skip_cmp_src_addr(struct pf_rule *a, struct pf_rule *b) 1134 { 1135 if (a->src.neg != b->src.neg || 1136 a->src.addr.type != b->src.addr.type) 1137 return (1); 1138 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0 1139 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP || 1140 * a->proto == IPPROTO_ICMP 1141 * return (1); 1142 */ 1143 switch (a->src.addr.type) { 1144 case PF_ADDR_ADDRMASK: 1145 if (memcmp(&a->src.addr.v.a.addr, &b->src.addr.v.a.addr, 1146 sizeof(a->src.addr.v.a.addr)) || 1147 memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask, 1148 sizeof(a->src.addr.v.a.mask)) || 1149 (a->src.addr.v.a.addr.addr32[0] == 0 && 1150 a->src.addr.v.a.addr.addr32[1] == 0 && 1151 a->src.addr.v.a.addr.addr32[2] == 0 && 1152 a->src.addr.v.a.addr.addr32[3] == 0)) 1153 return (1); 1154 return (0); 1155 case PF_ADDR_DYNIFTL: 1156 if (strcmp(a->src.addr.v.ifname, b->src.addr.v.ifname) != 0 || 1157 a->src.addr.iflags != a->src.addr.iflags || 1158 memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask, 1159 sizeof(a->src.addr.v.a.mask))) 1160 return (1); 1161 return (0); 1162 case PF_ADDR_NOROUTE: 1163 case PF_ADDR_URPFFAILED: 1164 return (0); 1165 case PF_ADDR_TABLE: 1166 return (strcmp(a->src.addr.v.tblname, b->src.addr.v.tblname)); 1167 } 1168 return (1); 1169 } 1170 1171 /* Compare two rules SRC port field for skiplist construction */ 1172 int 1173 skip_cmp_src_port(struct pf_rule *a, struct pf_rule *b) 1174 { 1175 if (a->src.port_op == PF_OP_NONE || a->src.port_op != b->src.port_op || 1176 a->src.port[0] != b->src.port[0] || 1177 a->src.port[1] != b->src.port[1]) 1178 return (1); 1179 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0 1180 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP || 1181 * a->proto == IPPROTO_ICMP 1182 * return (1); 1183 */ 1184 return (0); 1185 } 1186 1187 1188 void 1189 skip_init(void) 1190 { 1191 struct { 1192 char *name; 1193 int skipnum; 1194 int (*func)(struct pf_rule *, struct pf_rule *); 1195 } comps[] = PF_SKIP_COMPARITORS; 1196 int skipnum, i; 1197 1198 for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++) { 1199 for (i = 0; i < sizeof(comps)/sizeof(*comps); i++) 1200 if (comps[i].skipnum == skipnum) { 1201 skip_comparitors[skipnum] = comps[i].func; 1202 skip_comparitors_names[skipnum] = comps[i].name; 1203 } 1204 } 1205 for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++) 1206 if (skip_comparitors[skipnum] == NULL) 1207 errx(1, "Need to add skip step comparitor to pfctl?!"); 1208 } 1209 1210 /* 1211 * Add a host/netmask to a table 1212 */ 1213 int 1214 add_opt_table(struct pfctl *pf, struct pf_opt_tbl **tbl, sa_family_t af, 1215 struct pf_rule_addr *addr) 1216 { 1217 #ifdef OPT_DEBUG 1218 char buf[128]; 1219 #endif /* OPT_DEBUG */ 1220 static int tablenum = 0; 1221 struct node_host node_host; 1222 1223 if (*tbl == NULL) { 1224 if ((*tbl = calloc(1, sizeof(**tbl))) == NULL || 1225 ((*tbl)->pt_buf = calloc(1, sizeof(*(*tbl)->pt_buf))) == 1226 NULL) 1227 err(1, "calloc"); 1228 (*tbl)->pt_buf->pfrb_type = PFRB_ADDRS; 1229 SIMPLEQ_INIT(&(*tbl)->pt_nodes); 1230 1231 /* This is just a temporary table name */ 1232 snprintf((*tbl)->pt_name, sizeof((*tbl)->pt_name), "%s%d", 1233 PF_OPT_TABLE_PREFIX, tablenum++); 1234 DEBUG("creating table <%s>", (*tbl)->pt_name); 1235 } 1236 1237 memset(&node_host, 0, sizeof(node_host)); 1238 node_host.af = af; 1239 node_host.addr = addr->addr; 1240 1241 #ifdef OPT_DEBUG 1242 DEBUG("<%s> adding %s/%d", (*tbl)->pt_name, inet_ntop(af, 1243 &node_host.addr.v.a.addr, buf, sizeof(buf)), 1244 unmask(&node_host.addr.v.a.mask, af)); 1245 #endif /* OPT_DEBUG */ 1246 1247 if (append_addr_host((*tbl)->pt_buf, &node_host, 0, 0)) { 1248 warn("failed to add host"); 1249 return (1); 1250 } 1251 if (pf->opts & PF_OPT_VERBOSE) { 1252 struct node_tinit *ti; 1253 1254 if ((ti = calloc(1, sizeof(*ti))) == NULL) 1255 err(1, "malloc"); 1256 if ((ti->host = malloc(sizeof(*ti->host))) == NULL) 1257 err(1, "malloc"); 1258 memcpy(ti->host, &node_host, sizeof(*ti->host)); 1259 SIMPLEQ_INSERT_TAIL(&(*tbl)->pt_nodes, ti, entries); 1260 } 1261 1262 (*tbl)->pt_rulecount++; 1263 if ((*tbl)->pt_rulecount == TABLE_THRESHOLD) 1264 DEBUG("table <%s> now faster than skip steps", (*tbl)->pt_name); 1265 1266 return (0); 1267 } 1268 1269 1270 /* 1271 * Do the dirty work of choosing an unused table name and creating it. 1272 * (be careful with the table name, it might already be used in another anchor) 1273 */ 1274 int 1275 pf_opt_create_table(struct pfctl *pf, struct pf_opt_tbl *tbl) 1276 { 1277 static int tablenum; 1278 const struct pfr_table *t; 1279 1280 if (table_buffer.pfrb_type == 0) { 1281 /* Initialize the list of tables */ 1282 table_buffer.pfrb_type = PFRB_TABLES; 1283 for (;;) { 1284 pfr_buf_grow(&table_buffer, table_buffer.pfrb_size); 1285 table_buffer.pfrb_size = table_buffer.pfrb_msize; 1286 if (pfr_get_tables(NULL, table_buffer.pfrb_caddr, 1287 &table_buffer.pfrb_size, PFR_FLAG_ALLRSETS)) 1288 err(1, "pfr_get_tables"); 1289 if (table_buffer.pfrb_size <= table_buffer.pfrb_msize) 1290 break; 1291 } 1292 table_identifier = arc4random(); 1293 } 1294 1295 /* XXX would be *really* nice to avoid duplicating identical tables */ 1296 1297 /* Now we have to pick a table name that isn't used */ 1298 again: 1299 DEBUG("translating temporary table <%s> to <%s%x_%d>", tbl->pt_name, 1300 PF_OPT_TABLE_PREFIX, table_identifier, tablenum); 1301 snprintf(tbl->pt_name, sizeof(tbl->pt_name), "%s%x_%d", 1302 PF_OPT_TABLE_PREFIX, table_identifier, tablenum); 1303 PFRB_FOREACH(t, &table_buffer) { 1304 if (strcasecmp(t->pfrt_name, tbl->pt_name) == 0) { 1305 /* Collision. Try again */ 1306 DEBUG("wow, table <%s> in use. trying again", 1307 tbl->pt_name); 1308 table_identifier = arc4random(); 1309 goto again; 1310 } 1311 } 1312 tablenum++; 1313 1314 1315 if (pfctl_define_table(tbl->pt_name, PFR_TFLAG_CONST, 1, 1316 pf->anchor->name, tbl->pt_buf, pf->anchor->ruleset.tticket)) { 1317 warn("failed to create table %s", tbl->pt_name); 1318 return (1); 1319 } 1320 return (0); 1321 } 1322 1323 /* 1324 * Partition the flat ruleset into a list of distinct superblocks 1325 */ 1326 int 1327 construct_superblocks(struct pfctl *pf, struct pf_opt_queue *opt_queue, 1328 struct superblocks *superblocks) 1329 { 1330 struct superblock *block = NULL; 1331 struct pf_opt_rule *por; 1332 int i; 1333 1334 while (!TAILQ_EMPTY(opt_queue)) { 1335 por = TAILQ_FIRST(opt_queue); 1336 TAILQ_REMOVE(opt_queue, por, por_entry); 1337 if (block == NULL || !superblock_inclusive(block, por)) { 1338 if ((block = calloc(1, sizeof(*block))) == NULL) { 1339 warn("calloc"); 1340 return (1); 1341 } 1342 TAILQ_INIT(&block->sb_rules); 1343 for (i = 0; i < PF_SKIP_COUNT; i++) 1344 TAILQ_INIT(&block->sb_skipsteps[i]); 1345 TAILQ_INSERT_TAIL(superblocks, block, sb_entry); 1346 } 1347 TAILQ_INSERT_TAIL(&block->sb_rules, por, por_entry); 1348 } 1349 1350 return (0); 1351 } 1352 1353 1354 /* 1355 * Compare two rule addresses 1356 */ 1357 int 1358 addrs_equal(struct pf_rule_addr *a, struct pf_rule_addr *b) 1359 { 1360 if (a->neg != b->neg) 1361 return (0); 1362 return (memcmp(&a->addr, &b->addr, sizeof(a->addr)) == 0); 1363 } 1364 1365 1366 /* 1367 * The addresses are not equal, but can we combine them into one table? 1368 */ 1369 int 1370 addrs_combineable(struct pf_rule_addr *a, struct pf_rule_addr *b) 1371 { 1372 if (a->addr.type != PF_ADDR_ADDRMASK || 1373 b->addr.type != PF_ADDR_ADDRMASK) 1374 return (0); 1375 if (a->neg != b->neg || a->port_op != b->port_op || 1376 a->port[0] != b->port[0] || a->port[1] != b->port[1]) 1377 return (0); 1378 return (1); 1379 } 1380 1381 1382 /* 1383 * Are we allowed to combine these two rules 1384 */ 1385 int 1386 rules_combineable(struct pf_rule *p1, struct pf_rule *p2) 1387 { 1388 struct pf_rule a, b; 1389 1390 comparable_rule(&a, p1, COMBINED); 1391 comparable_rule(&b, p2, COMBINED); 1392 return (memcmp(&a, &b, sizeof(a)) == 0); 1393 } 1394 1395 1396 /* 1397 * Can a rule be included inside a superblock 1398 */ 1399 int 1400 superblock_inclusive(struct superblock *block, struct pf_opt_rule *por) 1401 { 1402 struct pf_rule a, b; 1403 int i, j; 1404 1405 /* First check for hard breaks */ 1406 for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++) { 1407 if (pf_rule_desc[i].prf_type == BARRIER) { 1408 for (j = 0; j < pf_rule_desc[i].prf_size; j++) 1409 if (((char *)&por->por_rule)[j + 1410 pf_rule_desc[i].prf_offset] != 0) 1411 return (0); 1412 } 1413 } 1414 1415 /* per-rule src-track is also a hard break */ 1416 if (por->por_rule.rule_flag & PFRULE_RULESRCTRACK) 1417 return (0); 1418 1419 /* 1420 * Have to handle interface groups seperately. Consider the following 1421 * rules: 1422 * block on EXTIFS to any port 22 1423 * pass on em0 to any port 22 1424 * (where EXTIFS is an arbitrary interface group) 1425 * The optimizer may decide to re-order the pass rule in front of the 1426 * block rule. But what if EXTIFS includes em0??? Such a reordering 1427 * would change the meaning of the ruleset. 1428 * We can't just lookup the EXTIFS group and check if em0 is a member 1429 * because the user is allowed to add interfaces to a group during 1430 * runtime. 1431 * Ergo interface groups become a defacto superblock break :-( 1432 */ 1433 if (interface_group(por->por_rule.ifname) || 1434 interface_group(TAILQ_FIRST(&block->sb_rules)->por_rule.ifname)) { 1435 if (strcasecmp(por->por_rule.ifname, 1436 TAILQ_FIRST(&block->sb_rules)->por_rule.ifname) != 0) 1437 return (0); 1438 } 1439 1440 comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule, NOMERGE); 1441 comparable_rule(&b, &por->por_rule, NOMERGE); 1442 if (memcmp(&a, &b, sizeof(a)) == 0) 1443 return (1); 1444 1445 #ifdef OPT_DEBUG 1446 for (i = 0; i < sizeof(por->por_rule); i++) { 1447 int closest = -1; 1448 if (((u_int8_t *)&a)[i] != ((u_int8_t *)&b)[i]) { 1449 for (j = 0; j < sizeof(pf_rule_desc) / 1450 sizeof(*pf_rule_desc); j++) { 1451 if (i >= pf_rule_desc[j].prf_offset && 1452 i < pf_rule_desc[j].prf_offset + 1453 pf_rule_desc[j].prf_size) { 1454 DEBUG("superblock break @ %d due to %s", 1455 por->por_rule.nr, 1456 pf_rule_desc[j].prf_name); 1457 return (0); 1458 } 1459 if (i > pf_rule_desc[j].prf_offset) { 1460 if (closest == -1 || 1461 i-pf_rule_desc[j].prf_offset < 1462 i-pf_rule_desc[closest].prf_offset) 1463 closest = j; 1464 } 1465 } 1466 1467 if (closest >= 0) 1468 DEBUG("superblock break @ %d on %s+%xh", 1469 por->por_rule.nr, 1470 pf_rule_desc[closest].prf_name, 1471 i - pf_rule_desc[closest].prf_offset - 1472 pf_rule_desc[closest].prf_size); 1473 else 1474 DEBUG("superblock break @ %d on field @ %d", 1475 por->por_rule.nr, i); 1476 return (0); 1477 } 1478 } 1479 #endif /* OPT_DEBUG */ 1480 1481 return (0); 1482 } 1483 1484 1485 /* 1486 * Figure out if an interface name is an actual interface or actually a 1487 * group of interfaces. 1488 */ 1489 int 1490 interface_group(const char *ifname) 1491 { 1492 if (ifname == NULL || !ifname[0]) 1493 return (0); 1494 1495 /* Real interfaces must end in a number, interface groups do not */ 1496 if (isdigit(ifname[strlen(ifname) - 1])) 1497 return (0); 1498 else 1499 return (1); 1500 } 1501 1502 1503 /* 1504 * Make a rule that can directly compared by memcmp() 1505 */ 1506 void 1507 comparable_rule(struct pf_rule *dst, const struct pf_rule *src, int type) 1508 { 1509 int i; 1510 /* 1511 * To simplify the comparison, we just zero out the fields that are 1512 * allowed to be different and then do a simple memcmp() 1513 */ 1514 memcpy(dst, src, sizeof(*dst)); 1515 for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++) 1516 if (pf_rule_desc[i].prf_type >= type) { 1517 #ifdef OPT_DEBUG 1518 assert(pf_rule_desc[i].prf_type != NEVER || 1519 *(((char *)dst) + pf_rule_desc[i].prf_offset) == 0); 1520 #endif /* OPT_DEBUG */ 1521 memset(((char *)dst) + pf_rule_desc[i].prf_offset, 0, 1522 pf_rule_desc[i].prf_size); 1523 } 1524 } 1525 1526 1527 /* 1528 * Remove superset information from two rules so we can directly compare them 1529 * with memcmp() 1530 */ 1531 void 1532 exclude_supersets(struct pf_rule *super, struct pf_rule *sub) 1533 { 1534 if (super->ifname[0] == '\0') 1535 memset(sub->ifname, 0, sizeof(sub->ifname)); 1536 if (super->direction == PF_INOUT) 1537 sub->direction = PF_INOUT; 1538 if ((super->proto == 0 || super->proto == sub->proto) && 1539 super->flags == 0 && super->flagset == 0 && (sub->flags || 1540 sub->flagset)) { 1541 sub->flags = super->flags; 1542 sub->flagset = super->flagset; 1543 } 1544 if (super->proto == 0) 1545 sub->proto = 0; 1546 1547 if (super->src.port_op == 0) { 1548 sub->src.port_op = 0; 1549 sub->src.port[0] = 0; 1550 sub->src.port[1] = 0; 1551 } 1552 if (super->dst.port_op == 0) { 1553 sub->dst.port_op = 0; 1554 sub->dst.port[0] = 0; 1555 sub->dst.port[1] = 0; 1556 } 1557 1558 if (super->src.addr.type == PF_ADDR_ADDRMASK && !super->src.neg && 1559 !sub->src.neg && super->src.addr.v.a.mask.addr32[0] == 0 && 1560 super->src.addr.v.a.mask.addr32[1] == 0 && 1561 super->src.addr.v.a.mask.addr32[2] == 0 && 1562 super->src.addr.v.a.mask.addr32[3] == 0) 1563 memset(&sub->src.addr, 0, sizeof(sub->src.addr)); 1564 else if (super->src.addr.type == PF_ADDR_ADDRMASK && 1565 sub->src.addr.type == PF_ADDR_ADDRMASK && 1566 super->src.neg == sub->src.neg && 1567 super->af == sub->af && 1568 unmask(&super->src.addr.v.a.mask, super->af) < 1569 unmask(&sub->src.addr.v.a.mask, sub->af) && 1570 super->src.addr.v.a.addr.addr32[0] == 1571 (sub->src.addr.v.a.addr.addr32[0] & 1572 super->src.addr.v.a.mask.addr32[0]) && 1573 super->src.addr.v.a.addr.addr32[1] == 1574 (sub->src.addr.v.a.addr.addr32[1] & 1575 super->src.addr.v.a.mask.addr32[1]) && 1576 super->src.addr.v.a.addr.addr32[2] == 1577 (sub->src.addr.v.a.addr.addr32[2] & 1578 super->src.addr.v.a.mask.addr32[2]) && 1579 super->src.addr.v.a.addr.addr32[3] == 1580 (sub->src.addr.v.a.addr.addr32[3] & 1581 super->src.addr.v.a.mask.addr32[3])) { 1582 /* sub->src.addr is a subset of super->src.addr/mask */ 1583 memcpy(&sub->src.addr, &super->src.addr, sizeof(sub->src.addr)); 1584 } 1585 1586 if (super->dst.addr.type == PF_ADDR_ADDRMASK && !super->dst.neg && 1587 !sub->dst.neg && super->dst.addr.v.a.mask.addr32[0] == 0 && 1588 super->dst.addr.v.a.mask.addr32[1] == 0 && 1589 super->dst.addr.v.a.mask.addr32[2] == 0 && 1590 super->dst.addr.v.a.mask.addr32[3] == 0) 1591 memset(&sub->dst.addr, 0, sizeof(sub->dst.addr)); 1592 else if (super->dst.addr.type == PF_ADDR_ADDRMASK && 1593 sub->dst.addr.type == PF_ADDR_ADDRMASK && 1594 super->dst.neg == sub->dst.neg && 1595 super->af == sub->af && 1596 unmask(&super->dst.addr.v.a.mask, super->af) < 1597 unmask(&sub->dst.addr.v.a.mask, sub->af) && 1598 super->dst.addr.v.a.addr.addr32[0] == 1599 (sub->dst.addr.v.a.addr.addr32[0] & 1600 super->dst.addr.v.a.mask.addr32[0]) && 1601 super->dst.addr.v.a.addr.addr32[1] == 1602 (sub->dst.addr.v.a.addr.addr32[1] & 1603 super->dst.addr.v.a.mask.addr32[1]) && 1604 super->dst.addr.v.a.addr.addr32[2] == 1605 (sub->dst.addr.v.a.addr.addr32[2] & 1606 super->dst.addr.v.a.mask.addr32[2]) && 1607 super->dst.addr.v.a.addr.addr32[3] == 1608 (sub->dst.addr.v.a.addr.addr32[3] & 1609 super->dst.addr.v.a.mask.addr32[3])) { 1610 /* sub->dst.addr is a subset of super->dst.addr/mask */ 1611 memcpy(&sub->dst.addr, &super->dst.addr, sizeof(sub->dst.addr)); 1612 } 1613 1614 if (super->af == 0) 1615 sub->af = 0; 1616 } 1617 1618 1619 void 1620 superblock_free(struct pfctl *pf, struct superblock *block) 1621 { 1622 struct pf_opt_rule *por; 1623 while ((por = TAILQ_FIRST(&block->sb_rules))) { 1624 TAILQ_REMOVE(&block->sb_rules, por, por_entry); 1625 if (por->por_src_tbl) { 1626 if (por->por_src_tbl->pt_buf) { 1627 pfr_buf_clear(por->por_src_tbl->pt_buf); 1628 free(por->por_src_tbl->pt_buf); 1629 } 1630 free(por->por_src_tbl); 1631 } 1632 if (por->por_dst_tbl) { 1633 if (por->por_dst_tbl->pt_buf) { 1634 pfr_buf_clear(por->por_dst_tbl->pt_buf); 1635 free(por->por_dst_tbl->pt_buf); 1636 } 1637 free(por->por_dst_tbl); 1638 } 1639 free(por); 1640 } 1641 if (block->sb_profiled_block) 1642 superblock_free(pf, block->sb_profiled_block); 1643 free(block); 1644 } 1645 1646