1 /* $NetBSD: pfctl_optimize.c,v 1.5 2005/07/01 12:43:50 peter Exp $ */ 2 /* $OpenBSD: pfctl_optimize.c,v 1.5 2005/01/03 15:18:10 frantzen Exp $ */ 3 4 /* 5 * Copyright (c) 2004 Mike Frantzen <frantzen@openbsd.org> 6 * 7 * Permission to use, copy, modify, and distribute this software for any 8 * purpose with or without fee is hereby granted, provided that the above 9 * copyright notice and this permission notice appear in all copies. 10 * 11 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES 12 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF 13 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR 14 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES 15 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN 16 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF 17 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. 18 */ 19 20 #include <sys/types.h> 21 #include <sys/ioctl.h> 22 #include <sys/socket.h> 23 24 #include <net/if.h> 25 #include <netinet/in.h> 26 #include <net/pfvar.h> 27 28 #include <arpa/inet.h> 29 30 #include <assert.h> 31 #include <ctype.h> 32 #include <err.h> 33 #include <errno.h> 34 #include <stddef.h> 35 #include <stdio.h> 36 #include <stdlib.h> 37 #include <string.h> 38 39 #include "pfctl_parser.h" 40 #include "pfctl.h" 41 42 /* The size at which a table becomes faster than individual rules */ 43 #define TABLE_THRESHOLD 6 44 45 46 /* #define OPT_DEBUG 1 */ 47 #ifdef OPT_DEBUG 48 # define DEBUG(str, v...) \ 49 printf("%s: " str "\n", __FUNCTION__ , ## v) 50 #else 51 # define DEBUG(str, v...) ((void)0) 52 #endif 53 54 55 /* 56 * A container that lets us sort a superblock to optimize the skip step jumps 57 */ 58 struct pf_skip_step { 59 int ps_count; /* number of items */ 60 TAILQ_HEAD( , pf_opt_rule) ps_rules; 61 TAILQ_ENTRY(pf_skip_step) ps_entry; 62 }; 63 64 65 /* 66 * A superblock is a block of adjacent rules of similar action. If there 67 * are five PASS rules in a row, they all become members of a superblock. 68 * Once we have a superblock, we are free to re-order any rules within it 69 * in order to improve performance; if a packet is passed, it doesn't matter 70 * who passed it. 71 */ 72 struct superblock { 73 TAILQ_HEAD( , pf_opt_rule) sb_rules; 74 TAILQ_ENTRY(superblock) sb_entry; 75 struct superblock *sb_profiled_block; 76 TAILQ_HEAD(skiplist, pf_skip_step) sb_skipsteps[PF_SKIP_COUNT]; 77 }; 78 TAILQ_HEAD(superblocks, superblock); 79 80 81 /* 82 * Description of the PF rule structure. 83 */ 84 enum { 85 BARRIER, /* the presence of the field puts the rule in it's own block */ 86 BREAK, /* the field may not differ between rules in a superblock */ 87 NOMERGE, /* the field may not differ between rules when combined */ 88 COMBINED, /* the field may itself be combined with other rules */ 89 DC, /* we just don't care about the field */ 90 NEVER}; /* we should never see this field set?!? */ 91 struct pf_rule_field { 92 const char *prf_name; 93 int prf_type; 94 size_t prf_offset; 95 size_t prf_size; 96 } pf_rule_desc[] = { 97 #define PF_RULE_FIELD(field, ty) \ 98 {#field, \ 99 ty, \ 100 offsetof(struct pf_rule, field), \ 101 sizeof(((struct pf_rule *)0)->field)} 102 103 104 /* 105 * The presence of these fields in a rule put the rule in it's own 106 * superblock. Thus it will not be optimized. It also prevents the 107 * rule from being re-ordered at all. 108 */ 109 PF_RULE_FIELD(label, BARRIER), 110 PF_RULE_FIELD(prob, BARRIER), 111 PF_RULE_FIELD(max_states, BARRIER), 112 PF_RULE_FIELD(max_src_nodes, BARRIER), 113 114 /* 115 * These fields must be the same between all rules in the same superblock. 116 * These rules are allowed to be re-ordered but only among like rules. 117 * For instance we can re-order all 'tag "foo"' rules because they have the 118 * same tag. But we can not re-order between a 'tag "foo"' and a 119 * 'tag "bar"' since that would change the meaning of the ruleset. 120 */ 121 PF_RULE_FIELD(tagname, BREAK), 122 PF_RULE_FIELD(keep_state, BREAK), 123 PF_RULE_FIELD(qname, BREAK), 124 PF_RULE_FIELD(rt, BREAK), 125 PF_RULE_FIELD(allow_opts, BREAK), 126 PF_RULE_FIELD(rule_flag, BREAK), 127 PF_RULE_FIELD(action, BREAK), 128 129 /* 130 * Any fields not listed in this structure act as BREAK fields 131 */ 132 133 134 /* 135 * These fields must not differ when we merge two rules together but 136 * their difference isn't enough to put the rules in different superblocks. 137 * There are no problems re-ordering any rules with these fields. 138 */ 139 PF_RULE_FIELD(af, NOMERGE), 140 PF_RULE_FIELD(ifnot, NOMERGE), 141 PF_RULE_FIELD(ifname, NOMERGE), 142 PF_RULE_FIELD(match_tag_not, NOMERGE), 143 PF_RULE_FIELD(match_tagname, NOMERGE), 144 PF_RULE_FIELD(os_fingerprint, NOMERGE), 145 PF_RULE_FIELD(timeout, NOMERGE), 146 PF_RULE_FIELD(return_icmp, NOMERGE), 147 PF_RULE_FIELD(return_icmp6, NOMERGE), 148 PF_RULE_FIELD(uid, NOMERGE), 149 PF_RULE_FIELD(gid, NOMERGE), 150 PF_RULE_FIELD(direction, NOMERGE), 151 PF_RULE_FIELD(proto, NOMERGE), 152 PF_RULE_FIELD(type, NOMERGE), 153 PF_RULE_FIELD(code, NOMERGE), 154 PF_RULE_FIELD(flags, NOMERGE), 155 PF_RULE_FIELD(flagset, NOMERGE), 156 PF_RULE_FIELD(tos, NOMERGE), 157 PF_RULE_FIELD(src.port, NOMERGE), 158 PF_RULE_FIELD(dst.port, NOMERGE), 159 PF_RULE_FIELD(src.port_op, NOMERGE), 160 PF_RULE_FIELD(dst.port_op, NOMERGE), 161 PF_RULE_FIELD(src.neg, NOMERGE), 162 PF_RULE_FIELD(dst.neg, NOMERGE), 163 164 /* These fields can be merged */ 165 PF_RULE_FIELD(src.addr, COMBINED), 166 PF_RULE_FIELD(dst.addr, COMBINED), 167 168 /* We just don't care about these fields. They're set by the kernel */ 169 PF_RULE_FIELD(skip, DC), 170 PF_RULE_FIELD(evaluations, DC), 171 PF_RULE_FIELD(packets, DC), 172 PF_RULE_FIELD(bytes, DC), 173 PF_RULE_FIELD(kif, DC), 174 PF_RULE_FIELD(anchor, DC), 175 PF_RULE_FIELD(states, DC), 176 PF_RULE_FIELD(src_nodes, DC), 177 PF_RULE_FIELD(nr, DC), 178 PF_RULE_FIELD(entries, DC), 179 PF_RULE_FIELD(qid, DC), 180 PF_RULE_FIELD(pqid, DC), 181 PF_RULE_FIELD(anchor_relative, DC), 182 PF_RULE_FIELD(anchor_wildcard, DC), 183 184 /* These fields should never be set in a PASS/BLOCK rule */ 185 PF_RULE_FIELD(natpass, NEVER), 186 PF_RULE_FIELD(max_mss, NEVER), 187 PF_RULE_FIELD(min_ttl, NEVER), 188 }; 189 190 191 192 int add_opt_table(struct pfctl *, struct pf_opt_tbl **, sa_family_t, 193 struct pf_rule_addr *); 194 int addrs_combineable(struct pf_rule_addr *, struct pf_rule_addr *); 195 int addrs_equal(struct pf_rule_addr *, struct pf_rule_addr *); 196 int block_feedback(struct pfctl *, struct superblock *); 197 int combine_rules(struct pfctl *, struct superblock *); 198 void comparable_rule(struct pf_rule *, const struct pf_rule *, int); 199 int construct_superblocks(struct pfctl *, struct pf_opt_queue *, 200 struct superblocks *); 201 void exclude_supersets(struct pf_rule *, struct pf_rule *); 202 int load_feedback_profile(struct pfctl *, struct superblocks *); 203 int optimize_superblock(struct pfctl *, struct superblock *); 204 int pf_opt_create_table(struct pfctl *, struct pf_opt_tbl *); 205 void remove_from_skipsteps(struct skiplist *, struct superblock *, 206 struct pf_opt_rule *, struct pf_skip_step *); 207 int remove_identical_rules(struct pfctl *, struct superblock *); 208 int reorder_rules(struct pfctl *, struct superblock *, int); 209 int rules_combineable(struct pf_rule *, struct pf_rule *); 210 void skip_append(struct superblock *, int, struct pf_skip_step *, 211 struct pf_opt_rule *); 212 int skip_compare(int, struct pf_skip_step *, struct pf_opt_rule *); 213 void skip_init(void); 214 int skip_cmp_af(struct pf_rule *, struct pf_rule *); 215 int skip_cmp_dir(struct pf_rule *, struct pf_rule *); 216 int skip_cmp_dst_addr(struct pf_rule *, struct pf_rule *); 217 int skip_cmp_dst_port(struct pf_rule *, struct pf_rule *); 218 int skip_cmp_ifp(struct pf_rule *, struct pf_rule *); 219 int skip_cmp_proto(struct pf_rule *, struct pf_rule *); 220 int skip_cmp_src_addr(struct pf_rule *, struct pf_rule *); 221 int skip_cmp_src_port(struct pf_rule *, struct pf_rule *); 222 int superblock_inclusive(struct superblock *, struct pf_opt_rule *); 223 void superblock_free(struct pfctl *, struct superblock *); 224 225 226 int (*skip_comparitors[PF_SKIP_COUNT])(struct pf_rule *, struct pf_rule *); 227 const char *skip_comparitors_names[PF_SKIP_COUNT]; 228 #define PF_SKIP_COMPARITORS { \ 229 { "ifp", PF_SKIP_IFP, skip_cmp_ifp }, \ 230 { "dir", PF_SKIP_DIR, skip_cmp_dir }, \ 231 { "af", PF_SKIP_AF, skip_cmp_af }, \ 232 { "proto", PF_SKIP_PROTO, skip_cmp_proto }, \ 233 { "saddr", PF_SKIP_SRC_ADDR, skip_cmp_src_addr }, \ 234 { "sport", PF_SKIP_SRC_PORT, skip_cmp_src_port }, \ 235 { "daddr", PF_SKIP_DST_ADDR, skip_cmp_dst_addr }, \ 236 { "dport", PF_SKIP_DST_PORT, skip_cmp_dst_port } \ 237 } 238 239 struct pfr_buffer table_buffer; 240 int table_identifier; 241 242 243 int 244 pfctl_optimize_rules(struct pfctl *pf) 245 { 246 struct superblocks superblocks; 247 struct superblock *block; 248 struct pf_opt_rule *por; 249 int nr = 0; 250 251 DEBUG("optimizing ruleset"); 252 memset(&table_buffer, 0, sizeof(table_buffer)); 253 skip_init(); 254 255 if (TAILQ_FIRST(&pf->opt_queue)) 256 nr = TAILQ_FIRST(&pf->opt_queue)->por_rule.nr; 257 258 TAILQ_INIT(&superblocks); 259 if (construct_superblocks(pf, &pf->opt_queue, &superblocks)) 260 goto error; 261 262 if (pf->opts & PF_OPT_OPTIMIZE_PROFILE) { 263 if (load_feedback_profile(pf, &superblocks)) 264 goto error; 265 } 266 267 TAILQ_FOREACH(block, &superblocks, sb_entry) { 268 if (optimize_superblock(pf, block)) 269 goto error; 270 } 271 272 273 /* 274 * Optimizations are done so we turn off the optimization flag and 275 * put the rules right back into the regular codepath. 276 */ 277 pf->opts &= ~PF_OPT_OPTIMIZE; 278 279 while ((block = TAILQ_FIRST(&superblocks))) { 280 TAILQ_REMOVE(&superblocks, block, sb_entry); 281 282 while ((por = TAILQ_FIRST(&block->sb_rules))) { 283 TAILQ_REMOVE(&block->sb_rules, por, por_entry); 284 por->por_rule.nr = nr++; 285 if (pfctl_add_rule(pf, &por->por_rule, 286 por->por_anchor)) { 287 free(por); 288 goto error; 289 } 290 free(por); 291 } 292 free(block); 293 } 294 295 return (0); 296 297 error: 298 while ((por = TAILQ_FIRST(&pf->opt_queue))) { 299 TAILQ_REMOVE(&pf->opt_queue, por, por_entry); 300 if (por->por_src_tbl) { 301 pfr_buf_clear(por->por_src_tbl->pt_buf); 302 free(por->por_src_tbl->pt_buf); 303 free(por->por_src_tbl); 304 } 305 if (por->por_dst_tbl) { 306 pfr_buf_clear(por->por_dst_tbl->pt_buf); 307 free(por->por_dst_tbl->pt_buf); 308 free(por->por_dst_tbl); 309 } 310 free(por); 311 } 312 while ((block = TAILQ_FIRST(&superblocks))) { 313 TAILQ_REMOVE(&superblocks, block, sb_entry); 314 superblock_free(pf, block); 315 } 316 return (1); 317 } 318 319 320 /* 321 * Go ahead and optimize a superblock 322 */ 323 int 324 optimize_superblock(struct pfctl *pf, struct superblock *block) 325 { 326 #ifdef OPT_DEBUG 327 struct pf_opt_rule *por; 328 #endif /* OPT_DEBUG */ 329 330 /* We have a few optimization passes: 331 * 1) remove duplicate rules or rules that are a subset of other 332 * rules 333 * 2) combine otherwise identical rules with different IP addresses 334 * into a single rule and put the addresses in a table. 335 * 3) re-order the rules to improve kernel skip steps 336 * 4) re-order the 'quick' rules based on feedback from the 337 * active ruleset statistics 338 * 339 * XXX combine_rules() doesn't combine v4 and v6 rules. would just 340 * have to keep af in the table container, make af 'COMBINE' and 341 * twiddle the af on the merged rule 342 * XXX maybe add a weighting to the metric on skipsteps when doing 343 * reordering. sometimes two sequential tables will be better 344 * that four consecutive interfaces. 345 * XXX need to adjust the skipstep count of everything after PROTO, 346 * since they aren't actually checked on a proto mismatch in 347 * pf_test_{tcp, udp, icmp}() 348 * XXX should i treat proto=0, af=0 or dir=0 special in skepstep 349 * calculation since they are a DC? 350 * XXX keep last skiplist of last superblock to influence this 351 * superblock. '5 inet6 log' should make '3 inet6' come before '4 352 * inet' in the next superblock. 353 * XXX would be useful to add tables for ports 354 * XXX we can also re-order some mutually exclusive superblocks to 355 * try merging superblocks before any of these optimization passes. 356 * for instance a single 'log in' rule in the middle of non-logging 357 * out rules. 358 */ 359 360 /* shortcut. there will be alot of 1-rule superblocks */ 361 if (!TAILQ_NEXT(TAILQ_FIRST(&block->sb_rules), por_entry)) 362 return (0); 363 364 #ifdef OPT_DEBUG 365 printf("--- Superblock ---\n"); 366 TAILQ_FOREACH(por, &block->sb_rules, por_entry) { 367 printf(" "); 368 print_rule(&por->por_rule, por->por_anchor, 1); 369 } 370 #endif /* OPT_DEBUG */ 371 372 373 if (remove_identical_rules(pf, block)) 374 return (1); 375 if (combine_rules(pf, block)) 376 return (1); 377 if ((pf->opts & PF_OPT_OPTIMIZE_PROFILE) && 378 TAILQ_FIRST(&block->sb_rules)->por_rule.quick && 379 block->sb_profiled_block) { 380 if (block_feedback(pf, block)) 381 return (1); 382 } else if (reorder_rules(pf, block, 0)) { 383 return (1); 384 } 385 386 /* 387 * Don't add any optimization passes below reorder_rules(). It will 388 * have divided superblocks into smaller blocks for further refinement 389 * and doesn't put them back together again. What once was a true 390 * superblock might have been split into multiple superblocks. 391 */ 392 393 #ifdef OPT_DEBUG 394 printf("--- END Superblock ---\n"); 395 #endif /* OPT_DEBUG */ 396 return (0); 397 } 398 399 400 /* 401 * Optimization pass #1: remove identical rules 402 */ 403 int 404 remove_identical_rules(struct pfctl *pf, struct superblock *block) 405 { 406 struct pf_opt_rule *por1, *por2, *por_next, *por2_next; 407 struct pf_rule a, a2, b, b2; 408 409 for (por1 = TAILQ_FIRST(&block->sb_rules); por1; por1 = por_next) { 410 por_next = TAILQ_NEXT(por1, por_entry); 411 for (por2 = por_next; por2; por2 = por2_next) { 412 por2_next = TAILQ_NEXT(por2, por_entry); 413 comparable_rule(&a, &por1->por_rule, DC); 414 comparable_rule(&b, &por2->por_rule, DC); 415 memcpy(&a2, &a, sizeof(a2)); 416 memcpy(&b2, &b, sizeof(b2)); 417 418 exclude_supersets(&a, &b); 419 exclude_supersets(&b2, &a2); 420 if (memcmp(&a, &b, sizeof(a)) == 0) { 421 DEBUG("removing identical rule nr%d = *nr%d*", 422 por1->por_rule.nr, por2->por_rule.nr); 423 TAILQ_REMOVE(&block->sb_rules, por2, por_entry); 424 if (por_next == por2) 425 por_next = TAILQ_NEXT(por1, por_entry); 426 free(por2); 427 } else if (memcmp(&a2, &b2, sizeof(a2)) == 0) { 428 DEBUG("removing identical rule *nr%d* = nr%d", 429 por1->por_rule.nr, por2->por_rule.nr); 430 TAILQ_REMOVE(&block->sb_rules, por1, por_entry); 431 free(por1); 432 break; 433 } 434 } 435 } 436 437 return (0); 438 } 439 440 441 /* 442 * Optimization pass #2: combine similar rules with different addresses 443 * into a single rule and a table 444 */ 445 int 446 combine_rules(struct pfctl *pf, struct superblock *block) 447 { 448 struct pf_opt_rule *p1, *p2, *por_next; 449 int src_eq, dst_eq; 450 451 if ((pf->loadopt & PFCTL_FLAG_TABLE) == 0) { 452 warnx("Must enable table loading for optimizations"); 453 return (1); 454 } 455 456 /* First we make a pass to combine the rules. O(n log n) */ 457 TAILQ_FOREACH(p1, &block->sb_rules, por_entry) { 458 for (p2 = TAILQ_NEXT(p1, por_entry); p2; p2 = por_next) { 459 por_next = TAILQ_NEXT(p2, por_entry); 460 461 src_eq = addrs_equal(&p1->por_rule.src, 462 &p2->por_rule.src); 463 dst_eq = addrs_equal(&p1->por_rule.dst, 464 &p2->por_rule.dst); 465 466 if (src_eq && !dst_eq && p1->por_src_tbl == NULL && 467 p2->por_dst_tbl == NULL && 468 p2->por_src_tbl == NULL && 469 rules_combineable(&p1->por_rule, &p2->por_rule) && 470 addrs_combineable(&p1->por_rule.dst, 471 &p2->por_rule.dst)) { 472 DEBUG("can combine rules nr%d = nr%d", 473 p1->por_rule.nr, p2->por_rule.nr); 474 if (p1->por_dst_tbl == NULL && 475 add_opt_table(pf, &p1->por_dst_tbl, 476 p1->por_rule.af, &p1->por_rule.dst)) 477 return (1); 478 if (add_opt_table(pf, &p1->por_dst_tbl, 479 p1->por_rule.af, &p2->por_rule.dst)) 480 return (1); 481 p2->por_dst_tbl = p1->por_dst_tbl; 482 if (p1->por_dst_tbl->pt_rulecount >= 483 TABLE_THRESHOLD) { 484 TAILQ_REMOVE(&block->sb_rules, p2, 485 por_entry); 486 free(p2); 487 } 488 } else if (!src_eq && dst_eq && p1->por_dst_tbl == NULL 489 && p2->por_src_tbl == NULL && 490 p2->por_dst_tbl == NULL && 491 rules_combineable(&p1->por_rule, &p2->por_rule) && 492 addrs_combineable(&p1->por_rule.src, 493 &p2->por_rule.src)) { 494 DEBUG("can combine rules nr%d = nr%d", 495 p1->por_rule.nr, p2->por_rule.nr); 496 if (p1->por_src_tbl == NULL && 497 add_opt_table(pf, &p1->por_src_tbl, 498 p1->por_rule.af, &p1->por_rule.src)) 499 return (1); 500 if (add_opt_table(pf, &p1->por_src_tbl, 501 p1->por_rule.af, &p2->por_rule.src)) 502 return (1); 503 p2->por_src_tbl = p1->por_src_tbl; 504 if (p1->por_src_tbl->pt_rulecount >= 505 TABLE_THRESHOLD) { 506 TAILQ_REMOVE(&block->sb_rules, p2, 507 por_entry); 508 free(p2); 509 } 510 } 511 } 512 } 513 514 515 /* 516 * Then we make a final pass to create a valid table name and 517 * insert the name into the rules. 518 */ 519 for (p1 = TAILQ_FIRST(&block->sb_rules); p1; p1 = por_next) { 520 por_next = TAILQ_NEXT(p1, por_entry); 521 assert(p1->por_src_tbl == NULL || p1->por_dst_tbl == NULL); 522 523 if (p1->por_src_tbl && p1->por_src_tbl->pt_rulecount >= 524 TABLE_THRESHOLD) { 525 if (p1->por_src_tbl->pt_generated) { 526 /* This rule is included in a table */ 527 TAILQ_REMOVE(&block->sb_rules, p1, por_entry); 528 free(p1); 529 continue; 530 } 531 p1->por_src_tbl->pt_generated = 1; 532 533 if ((pf->opts & PF_OPT_NOACTION) == 0 && 534 pf_opt_create_table(pf, p1->por_src_tbl)) 535 return (1); 536 537 pf->tdirty = 1; 538 539 if (pf->opts & PF_OPT_VERBOSE) 540 print_tabledef(p1->por_src_tbl->pt_name, 541 PFR_TFLAG_CONST, 1, 542 &p1->por_src_tbl->pt_nodes); 543 544 memset(&p1->por_rule.src.addr, 0, 545 sizeof(p1->por_rule.src.addr)); 546 p1->por_rule.src.addr.type = PF_ADDR_TABLE; 547 strlcpy(p1->por_rule.src.addr.v.tblname, 548 p1->por_src_tbl->pt_name, 549 sizeof(p1->por_rule.src.addr.v.tblname)); 550 551 pfr_buf_clear(p1->por_src_tbl->pt_buf); 552 free(p1->por_src_tbl->pt_buf); 553 p1->por_src_tbl->pt_buf = NULL; 554 } 555 if (p1->por_dst_tbl && p1->por_dst_tbl->pt_rulecount >= 556 TABLE_THRESHOLD) { 557 if (p1->por_dst_tbl->pt_generated) { 558 /* This rule is included in a table */ 559 TAILQ_REMOVE(&block->sb_rules, p1, por_entry); 560 free(p1); 561 continue; 562 } 563 p1->por_dst_tbl->pt_generated = 1; 564 565 if ((pf->opts & PF_OPT_NOACTION) == 0 && 566 pf_opt_create_table(pf, p1->por_dst_tbl)) 567 return (1); 568 pf->tdirty = 1; 569 570 if (pf->opts & PF_OPT_VERBOSE) 571 print_tabledef(p1->por_dst_tbl->pt_name, 572 PFR_TFLAG_CONST, 1, 573 &p1->por_dst_tbl->pt_nodes); 574 575 memset(&p1->por_rule.dst.addr, 0, 576 sizeof(p1->por_rule.dst.addr)); 577 p1->por_rule.dst.addr.type = PF_ADDR_TABLE; 578 strlcpy(p1->por_rule.dst.addr.v.tblname, 579 p1->por_dst_tbl->pt_name, 580 sizeof(p1->por_rule.dst.addr.v.tblname)); 581 582 pfr_buf_clear(p1->por_dst_tbl->pt_buf); 583 free(p1->por_dst_tbl->pt_buf); 584 p1->por_dst_tbl->pt_buf = NULL; 585 } 586 } 587 588 return (0); 589 } 590 591 592 /* 593 * Optimization pass #3: re-order rules to improve skip steps 594 */ 595 int 596 reorder_rules(struct pfctl *pf, struct superblock *block, int depth) 597 { 598 struct superblock *newblock; 599 struct pf_skip_step *skiplist; 600 struct pf_opt_rule *por; 601 int i, largest, largest_list = -1, rule_count = 0; 602 TAILQ_HEAD( , pf_opt_rule) head; 603 604 /* 605 * Calculate the best-case skip steps. We put each rule in a list 606 * of other rules with common fields 607 */ 608 for (i = 0; i < PF_SKIP_COUNT; i++) { 609 TAILQ_FOREACH(por, &block->sb_rules, por_entry) { 610 TAILQ_FOREACH(skiplist, &block->sb_skipsteps[i], 611 ps_entry) { 612 if (skip_compare(i, skiplist, por) == 0) 613 break; 614 } 615 if (skiplist == NULL) { 616 if ((skiplist = calloc(1, sizeof(*skiplist))) == 617 NULL) 618 err(1, "calloc"); 619 TAILQ_INIT(&skiplist->ps_rules); 620 TAILQ_INSERT_TAIL(&block->sb_skipsteps[i], 621 skiplist, ps_entry); 622 } 623 skip_append(block, i, skiplist, por); 624 } 625 } 626 627 TAILQ_FOREACH(por, &block->sb_rules, por_entry) 628 rule_count++; 629 630 /* 631 * Now we're going to ignore any fields that are identical between 632 * all of the rules in the superblock and those fields which differ 633 * between every rule in the superblock. 634 */ 635 largest = 0; 636 for (i = 0; i < PF_SKIP_COUNT; i++) { 637 skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]); 638 if (skiplist->ps_count == rule_count) { 639 DEBUG("(%d) original skipstep '%s' is all rules", 640 depth, skip_comparitors_names[i]); 641 skiplist->ps_count = 0; 642 } else if (skiplist->ps_count == 1) { 643 skiplist->ps_count = 0; 644 } else { 645 DEBUG("(%d) original skipstep '%s' largest jump is %d", 646 depth, skip_comparitors_names[i], 647 skiplist->ps_count); 648 if (skiplist->ps_count > largest) 649 largest = skiplist->ps_count; 650 } 651 } 652 if (largest == 0) { 653 /* Ugh. There is NO commonality in the superblock on which 654 * optimize the skipsteps optimization. 655 */ 656 goto done; 657 } 658 659 /* 660 * Now we're going to empty the superblock rule list and re-create 661 * it based on a more optimal skipstep order. 662 */ 663 TAILQ_INIT(&head); 664 while ((por = TAILQ_FIRST(&block->sb_rules))) { 665 TAILQ_REMOVE(&block->sb_rules, por, por_entry); 666 TAILQ_INSERT_TAIL(&head, por, por_entry); 667 } 668 669 670 while (!TAILQ_EMPTY(&head)) { 671 largest = 1; 672 673 /* 674 * Find the most useful skip steps remaining 675 */ 676 for (i = 0; i < PF_SKIP_COUNT; i++) { 677 skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]); 678 if (skiplist->ps_count > largest) { 679 largest = skiplist->ps_count; 680 largest_list = i; 681 } 682 } 683 684 if (largest <= 1) { 685 /* 686 * Nothing useful left. Leave remaining rules in order. 687 */ 688 DEBUG("(%d) no more commonality for skip steps", depth); 689 while ((por = TAILQ_FIRST(&head))) { 690 TAILQ_REMOVE(&head, por, por_entry); 691 TAILQ_INSERT_TAIL(&block->sb_rules, por, 692 por_entry); 693 } 694 } else { 695 /* 696 * There is commonality. Extract those common rules 697 * and place them in the ruleset adjacent to each 698 * other. 699 */ 700 skiplist = TAILQ_FIRST(&block->sb_skipsteps[ 701 largest_list]); 702 DEBUG("(%d) skipstep '%s' largest jump is %d @ #%d", 703 depth, skip_comparitors_names[largest_list], 704 largest, TAILQ_FIRST(&TAILQ_FIRST(&block-> 705 sb_skipsteps [largest_list])->ps_rules)-> 706 por_rule.nr); 707 TAILQ_REMOVE(&block->sb_skipsteps[largest_list], 708 skiplist, ps_entry); 709 710 711 /* 712 * There may be further commonality inside these 713 * rules. So we'll split them off into they're own 714 * superblock and pass it back into the optimizer. 715 */ 716 if (skiplist->ps_count > 2) { 717 if ((newblock = calloc(1, sizeof(*newblock))) 718 == NULL) { 719 warn("calloc"); 720 return (1); 721 } 722 TAILQ_INIT(&newblock->sb_rules); 723 for (i = 0; i < PF_SKIP_COUNT; i++) 724 TAILQ_INIT(&newblock->sb_skipsteps[i]); 725 TAILQ_INSERT_BEFORE(block, newblock, sb_entry); 726 DEBUG("(%d) splitting off %d rules from superblock @ #%d", 727 depth, skiplist->ps_count, 728 TAILQ_FIRST(&skiplist->ps_rules)-> 729 por_rule.nr); 730 } else { 731 newblock = block; 732 } 733 734 while ((por = TAILQ_FIRST(&skiplist->ps_rules))) { 735 TAILQ_REMOVE(&head, por, por_entry); 736 TAILQ_REMOVE(&skiplist->ps_rules, por, 737 por_skip_entry[largest_list]); 738 TAILQ_INSERT_TAIL(&newblock->sb_rules, por, 739 por_entry); 740 741 /* Remove this rule from all other skiplists */ 742 remove_from_skipsteps(&block->sb_skipsteps[ 743 largest_list], block, por, skiplist); 744 } 745 free(skiplist); 746 if (newblock != block) 747 if (reorder_rules(pf, newblock, depth + 1)) 748 return (1); 749 } 750 } 751 752 done: 753 for (i = 0; i < PF_SKIP_COUNT; i++) { 754 while ((skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]))) { 755 TAILQ_REMOVE(&block->sb_skipsteps[i], skiplist, 756 ps_entry); 757 free(skiplist); 758 } 759 } 760 761 return (0); 762 } 763 764 765 /* 766 * Optimization pass #4: re-order 'quick' rules based on feedback from the 767 * currently running ruleset 768 */ 769 int 770 block_feedback(struct pfctl *pf, struct superblock *block) 771 { 772 TAILQ_HEAD( , pf_opt_rule) queue; 773 struct pf_opt_rule *por1, *por2; 774 u_int64_t total_count = 0; 775 struct pf_rule a, b; 776 777 778 /* 779 * Walk through all of the profiled superblock's rules and copy 780 * the counters onto our rules. 781 */ 782 TAILQ_FOREACH(por1, &block->sb_profiled_block->sb_rules, por_entry) { 783 comparable_rule(&a, &por1->por_rule, DC); 784 total_count += por1->por_rule.packets; 785 TAILQ_FOREACH(por2, &block->sb_rules, por_entry) { 786 if (por2->por_profile_count) 787 continue; 788 comparable_rule(&b, &por2->por_rule, DC); 789 if (memcmp(&a, &b, sizeof(a)) == 0) { 790 por2->por_profile_count = 791 por1->por_rule.packets; 792 break; 793 } 794 } 795 } 796 superblock_free(pf, block->sb_profiled_block); 797 block->sb_profiled_block = NULL; 798 799 /* 800 * Now we pull all of the rules off the superblock and re-insert them 801 * in sorted order. 802 */ 803 804 TAILQ_INIT(&queue); 805 while ((por1 = TAILQ_FIRST(&block->sb_rules)) != NULL) { 806 TAILQ_REMOVE(&block->sb_rules, por1, por_entry); 807 TAILQ_INSERT_TAIL(&queue, por1, por_entry); 808 } 809 810 while ((por1 = TAILQ_FIRST(&queue)) != NULL) { 811 TAILQ_REMOVE(&queue, por1, por_entry); 812 /* XXX I should sort all of the unused rules based on skip steps */ 813 TAILQ_FOREACH(por2, &block->sb_rules, por_entry) { 814 if (por1->por_profile_count > por2->por_profile_count) { 815 TAILQ_INSERT_BEFORE(por2, por1, por_entry); 816 break; 817 } 818 } 819 if (por2 == TAILQ_END(&block->sb_rules)) 820 TAILQ_INSERT_TAIL(&block->sb_rules, por1, por_entry); 821 } 822 823 return (0); 824 } 825 826 827 /* 828 * Load the current ruleset from the kernel and try to associate them with 829 * the ruleset we're optimizing. 830 */ 831 int 832 load_feedback_profile(struct pfctl *pf, struct superblocks *superblocks) 833 { 834 struct superblock *block, *blockcur; 835 struct superblocks prof_superblocks; 836 struct pf_opt_rule *por; 837 struct pf_opt_queue queue; 838 struct pfioc_rule pr; 839 struct pf_rule a, b; 840 int nr, mnr; 841 842 TAILQ_INIT(&queue); 843 TAILQ_INIT(&prof_superblocks); 844 845 memset(&pr, 0, sizeof(pr)); 846 pr.rule.action = PF_PASS; 847 if (ioctl(pf->dev, DIOCGETRULES, &pr)) { 848 warn("DIOCGETRULES"); 849 return (1); 850 } 851 mnr = pr.nr; 852 853 DEBUG("Loading %d active rules for a feedback profile", mnr); 854 for (nr = 0; nr < mnr; ++nr) { 855 if ((por = calloc(1, sizeof(*por))) == NULL) { 856 warn("calloc"); 857 return (1); 858 } 859 pr.nr = nr; 860 if (ioctl(pf->dev, DIOCGETRULE, &pr)) { 861 warn("DIOCGETRULES"); 862 return (1); 863 } 864 memcpy(&por->por_rule, &pr.rule, sizeof(por->por_rule)); 865 strlcpy(por->por_anchor, pr.anchor_call, 866 sizeof(por->por_anchor)); 867 if (TAILQ_EMPTY(&por->por_rule.rpool.list)) 868 memset(&por->por_rule.rpool, 0, 869 sizeof(por->por_rule.rpool)); 870 TAILQ_INSERT_TAIL(&queue, por, por_entry); 871 872 /* XXX pfctl_get_pool(pf->dev, &pr.rule.rpool, nr, pr.ticket, 873 * PF_PASS, pf->anchor) ??? 874 * ... pfctl_clear_pool(&pr.rule.rpool) 875 */ 876 } 877 878 if (construct_superblocks(pf, &queue, &prof_superblocks)) 879 return (1); 880 881 882 /* 883 * Now we try to associate the active ruleset's superblocks with 884 * the superblocks we're compiling. 885 */ 886 block = TAILQ_FIRST(superblocks); 887 blockcur = TAILQ_FIRST(&prof_superblocks); 888 while (block && blockcur) { 889 comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule, 890 BREAK); 891 comparable_rule(&b, &TAILQ_FIRST(&blockcur->sb_rules)->por_rule, 892 BREAK); 893 if (memcmp(&a, &b, sizeof(a)) == 0) { 894 /* The two superblocks lined up */ 895 block->sb_profiled_block = blockcur; 896 } else { 897 DEBUG("superblocks don't line up between #%d and #%d", 898 TAILQ_FIRST(&block->sb_rules)->por_rule.nr, 899 TAILQ_FIRST(&blockcur->sb_rules)->por_rule.nr); 900 break; 901 } 902 block = TAILQ_NEXT(block, sb_entry); 903 blockcur = TAILQ_NEXT(blockcur, sb_entry); 904 } 905 906 907 908 /* Free any superblocks we couldn't link */ 909 while (blockcur) { 910 block = TAILQ_NEXT(blockcur, sb_entry); 911 superblock_free(pf, blockcur); 912 blockcur = block; 913 } 914 return (0); 915 } 916 917 918 /* 919 * Compare a rule to a skiplist to see if the rule is a member 920 */ 921 int 922 skip_compare(int skipnum, struct pf_skip_step *skiplist, 923 struct pf_opt_rule *por) 924 { 925 struct pf_rule *a, *b; 926 if (skipnum >= PF_SKIP_COUNT || skipnum < 0) 927 errx(1, "skip_compare() out of bounds"); 928 a = &por->por_rule; 929 b = &TAILQ_FIRST(&skiplist->ps_rules)->por_rule; 930 931 return ((skip_comparitors[skipnum])(a, b)); 932 } 933 934 935 /* 936 * Add a rule to a skiplist 937 */ 938 void 939 skip_append(struct superblock *superblock, int skipnum, 940 struct pf_skip_step *skiplist, struct pf_opt_rule *por) 941 { 942 struct pf_skip_step *prev; 943 944 skiplist->ps_count++; 945 TAILQ_INSERT_TAIL(&skiplist->ps_rules, por, por_skip_entry[skipnum]); 946 947 /* Keep the list of skiplists sorted by whichever is larger */ 948 while ((prev = TAILQ_PREV(skiplist, skiplist, ps_entry)) && 949 prev->ps_count < skiplist->ps_count) { 950 TAILQ_REMOVE(&superblock->sb_skipsteps[skipnum], 951 skiplist, ps_entry); 952 TAILQ_INSERT_BEFORE(prev, skiplist, ps_entry); 953 } 954 } 955 956 957 /* 958 * Remove a rule from the other skiplist calculations. 959 */ 960 void 961 remove_from_skipsteps(struct skiplist *head, struct superblock *block, 962 struct pf_opt_rule *por, struct pf_skip_step *active_list) 963 { 964 struct pf_skip_step *sk, *next; 965 struct pf_opt_rule *p2; 966 int i, found; 967 968 for (i = 0; i < PF_SKIP_COUNT; i++) { 969 sk = TAILQ_FIRST(&block->sb_skipsteps[i]); 970 if (sk == NULL || sk == active_list || sk->ps_count <= 1) 971 continue; 972 found = 0; 973 do { 974 TAILQ_FOREACH(p2, &sk->ps_rules, por_skip_entry[i]) 975 if (p2 == por) { 976 TAILQ_REMOVE(&sk->ps_rules, p2, 977 por_skip_entry[i]); 978 found = 1; 979 sk->ps_count--; 980 break; 981 } 982 } while (!found && (sk = TAILQ_NEXT(sk, ps_entry))); 983 if (found && sk) { 984 /* Does this change the sorting order? */ 985 while ((next = TAILQ_NEXT(sk, ps_entry)) && 986 next->ps_count > sk->ps_count) { 987 TAILQ_REMOVE(head, sk, ps_entry); 988 TAILQ_INSERT_AFTER(head, next, sk, ps_entry); 989 } 990 #ifdef OPT_DEBUG 991 next = TAILQ_NEXT(sk, ps_entry); 992 assert(next == NULL || next->ps_count <= sk->ps_count); 993 #endif /* OPT_DEBUG */ 994 } 995 } 996 } 997 998 999 /* Compare two rules AF field for skiplist construction */ 1000 int 1001 skip_cmp_af(struct pf_rule *a, struct pf_rule *b) 1002 { 1003 if (a->af != b->af || a->af == 0) 1004 return (1); 1005 return (0); 1006 } 1007 1008 /* Compare two rules DIRECTION field for skiplist construction */ 1009 int 1010 skip_cmp_dir(struct pf_rule *a, struct pf_rule *b) 1011 { 1012 if (a->direction == 0 || a->direction != b->direction) 1013 return (1); 1014 return (0); 1015 } 1016 1017 /* Compare two rules DST Address field for skiplist construction */ 1018 int 1019 skip_cmp_dst_addr(struct pf_rule *a, struct pf_rule *b) 1020 { 1021 if (a->dst.neg != b->dst.neg || 1022 a->dst.addr.type != b->dst.addr.type) 1023 return (1); 1024 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0 1025 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP || 1026 * a->proto == IPPROTO_ICMP 1027 * return (1); 1028 */ 1029 switch (a->dst.addr.type) { 1030 case PF_ADDR_ADDRMASK: 1031 if (memcmp(&a->dst.addr.v.a.addr, &b->dst.addr.v.a.addr, 1032 sizeof(a->dst.addr.v.a.addr)) || 1033 memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask, 1034 sizeof(a->dst.addr.v.a.mask)) || 1035 (a->dst.addr.v.a.addr.addr32[0] == 0 && 1036 a->dst.addr.v.a.addr.addr32[1] == 0 && 1037 a->dst.addr.v.a.addr.addr32[2] == 0 && 1038 a->dst.addr.v.a.addr.addr32[3] == 0)) 1039 return (1); 1040 return (0); 1041 case PF_ADDR_DYNIFTL: 1042 if (strcmp(a->dst.addr.v.ifname, b->dst.addr.v.ifname) != 0 || 1043 a->dst.addr.iflags != a->dst.addr.iflags || 1044 memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask, 1045 sizeof(a->dst.addr.v.a.mask))) 1046 return (1); 1047 return (0); 1048 case PF_ADDR_NOROUTE: 1049 return (0); 1050 case PF_ADDR_TABLE: 1051 return (strcmp(a->dst.addr.v.tblname, b->dst.addr.v.tblname)); 1052 } 1053 return (1); 1054 } 1055 1056 /* Compare two rules DST port field for skiplist construction */ 1057 int 1058 skip_cmp_dst_port(struct pf_rule *a, struct pf_rule *b) 1059 { 1060 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0 1061 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP || 1062 * a->proto == IPPROTO_ICMP 1063 * return (1); 1064 */ 1065 if (a->dst.port_op == PF_OP_NONE || a->dst.port_op != b->dst.port_op || 1066 a->dst.port[0] != b->dst.port[0] || 1067 a->dst.port[1] != b->dst.port[1]) 1068 return (1); 1069 return (0); 1070 } 1071 1072 /* Compare two rules IFP field for skiplist construction */ 1073 int 1074 skip_cmp_ifp(struct pf_rule *a, struct pf_rule *b) 1075 { 1076 if (strcmp(a->ifname, b->ifname) || a->ifname[0] == '\0') 1077 return (1); 1078 return (a->ifnot != b->ifnot); 1079 } 1080 1081 /* Compare two rules PROTO field for skiplist construction */ 1082 int 1083 skip_cmp_proto(struct pf_rule *a, struct pf_rule *b) 1084 { 1085 return (a->proto != b->proto || a->proto == 0); 1086 } 1087 1088 /* Compare two rules SRC addr field for skiplist construction */ 1089 int 1090 skip_cmp_src_addr(struct pf_rule *a, struct pf_rule *b) 1091 { 1092 if (a->src.neg != b->src.neg || 1093 a->src.addr.type != b->src.addr.type) 1094 return (1); 1095 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0 1096 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP || 1097 * a->proto == IPPROTO_ICMP 1098 * return (1); 1099 */ 1100 switch (a->src.addr.type) { 1101 case PF_ADDR_ADDRMASK: 1102 if (memcmp(&a->src.addr.v.a.addr, &b->src.addr.v.a.addr, 1103 sizeof(a->src.addr.v.a.addr)) || 1104 memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask, 1105 sizeof(a->src.addr.v.a.mask)) || 1106 (a->src.addr.v.a.addr.addr32[0] == 0 && 1107 a->src.addr.v.a.addr.addr32[1] == 0 && 1108 a->src.addr.v.a.addr.addr32[2] == 0 && 1109 a->src.addr.v.a.addr.addr32[3] == 0)) 1110 return (1); 1111 return (0); 1112 case PF_ADDR_DYNIFTL: 1113 if (strcmp(a->src.addr.v.ifname, b->src.addr.v.ifname) != 0 || 1114 a->src.addr.iflags != a->src.addr.iflags || 1115 memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask, 1116 sizeof(a->src.addr.v.a.mask))) 1117 return (1); 1118 return (0); 1119 case PF_ADDR_NOROUTE: 1120 return (0); 1121 case PF_ADDR_TABLE: 1122 return (strcmp(a->src.addr.v.tblname, b->src.addr.v.tblname)); 1123 } 1124 return (1); 1125 } 1126 1127 /* Compare two rules SRC port field for skiplist construction */ 1128 int 1129 skip_cmp_src_port(struct pf_rule *a, struct pf_rule *b) 1130 { 1131 if (a->src.port_op == PF_OP_NONE || a->src.port_op != b->src.port_op || 1132 a->src.port[0] != b->src.port[0] || 1133 a->src.port[1] != b->src.port[1]) 1134 return (1); 1135 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0 1136 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP || 1137 * a->proto == IPPROTO_ICMP 1138 * return (1); 1139 */ 1140 return (0); 1141 } 1142 1143 1144 void 1145 skip_init(void) 1146 { 1147 struct { 1148 char *name; 1149 int skipnum; 1150 int (*func)(struct pf_rule *, struct pf_rule *); 1151 } comps[] = PF_SKIP_COMPARITORS; 1152 int skipnum, i; 1153 1154 for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++) { 1155 for (i = 0; i < sizeof(comps)/sizeof(*comps); i++) 1156 if (comps[i].skipnum == skipnum) { 1157 skip_comparitors[skipnum] = comps[i].func; 1158 skip_comparitors_names[skipnum] = comps[i].name; 1159 } 1160 } 1161 for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++) 1162 if (skip_comparitors[skipnum] == NULL) 1163 errx(1, "Need to add skip step comparitor to pfctl?!"); 1164 } 1165 1166 /* 1167 * Add a host/netmask to a table 1168 */ 1169 int 1170 add_opt_table(struct pfctl *pf, struct pf_opt_tbl **tbl, sa_family_t af, 1171 struct pf_rule_addr *addr) 1172 { 1173 #ifdef OPT_DEBUG 1174 char buf[128]; 1175 #endif /* OPT_DEBUG */ 1176 static int tablenum = 0; 1177 struct node_host node_host; 1178 1179 if (*tbl == NULL) { 1180 if ((*tbl = calloc(1, sizeof(**tbl))) == NULL || 1181 ((*tbl)->pt_buf = calloc(1, sizeof(*(*tbl)->pt_buf))) == 1182 NULL) 1183 err(1, "calloc"); 1184 (*tbl)->pt_buf->pfrb_type = PFRB_ADDRS; 1185 SIMPLEQ_INIT(&(*tbl)->pt_nodes); 1186 1187 /* This is just a temporary table name */ 1188 snprintf((*tbl)->pt_name, sizeof((*tbl)->pt_name), "%s%d", 1189 PF_OPT_TABLE_PREFIX, tablenum++); 1190 DEBUG("creating table <%s>", (*tbl)->pt_name); 1191 } 1192 1193 memset(&node_host, 0, sizeof(node_host)); 1194 node_host.af = af; 1195 node_host.addr = addr->addr; 1196 1197 #ifdef OPT_DEBUG 1198 DEBUG("<%s> adding %s/%d", (*tbl)->pt_name, inet_ntop(af, 1199 &node_host.addr.v.a.addr, buf, sizeof(buf)), 1200 unmask(&node_host.addr.v.a.mask, af)); 1201 #endif /* OPT_DEBUG */ 1202 1203 if (append_addr_host((*tbl)->pt_buf, &node_host, 0, 0)) { 1204 warn("failed to add host"); 1205 return (1); 1206 } 1207 if (pf->opts & PF_OPT_VERBOSE) { 1208 struct node_tinit *ti; 1209 1210 if ((ti = calloc(1, sizeof(*ti))) == NULL) 1211 err(1, "malloc"); 1212 if ((ti->host = malloc(sizeof(*ti->host))) == NULL) 1213 err(1, "malloc"); 1214 memcpy(ti->host, &node_host, sizeof(*ti->host)); 1215 SIMPLEQ_INSERT_TAIL(&(*tbl)->pt_nodes, ti, entries); 1216 } 1217 1218 (*tbl)->pt_rulecount++; 1219 if ((*tbl)->pt_rulecount == TABLE_THRESHOLD) 1220 DEBUG("table <%s> now faster than skip steps", (*tbl)->pt_name); 1221 1222 return (0); 1223 } 1224 1225 1226 /* 1227 * Do the dirty work of choosing an unused table name and creating it. 1228 * (be careful with the table name, it might already be used in another anchor) 1229 */ 1230 int 1231 pf_opt_create_table(struct pfctl *pf, struct pf_opt_tbl *tbl) 1232 { 1233 static int tablenum; 1234 struct pfr_table *t; 1235 1236 if (table_buffer.pfrb_type == 0) { 1237 /* Initialize the list of tables */ 1238 table_buffer.pfrb_type = PFRB_TABLES; 1239 for (;;) { 1240 pfr_buf_grow(&table_buffer, table_buffer.pfrb_size); 1241 table_buffer.pfrb_size = table_buffer.pfrb_msize; 1242 if (pfr_get_tables(NULL, table_buffer.pfrb_caddr, 1243 &table_buffer.pfrb_size, PFR_FLAG_ALLRSETS)) 1244 err(1, "pfr_get_tables"); 1245 if (table_buffer.pfrb_size <= table_buffer.pfrb_msize) 1246 break; 1247 } 1248 table_identifier = arc4random(); 1249 } 1250 1251 /* XXX would be *really* nice to avoid duplicating identical tables */ 1252 1253 /* Now we have to pick a table name that isn't used */ 1254 again: 1255 DEBUG("translating temporary table <%s> to <%s%x_%d>", tbl->pt_name, 1256 PF_OPT_TABLE_PREFIX, table_identifier, tablenum); 1257 snprintf(tbl->pt_name, sizeof(tbl->pt_name), "%s%x_%d", 1258 PF_OPT_TABLE_PREFIX, table_identifier, tablenum); 1259 PFRB_FOREACH(t, &table_buffer) { 1260 if (strcasecmp(t->pfrt_name, tbl->pt_name) == 0) { 1261 /* Collision. Try again */ 1262 DEBUG("wow, table <%s> in use. trying again", 1263 tbl->pt_name); 1264 table_identifier = arc4random(); 1265 goto again; 1266 } 1267 } 1268 tablenum++; 1269 1270 1271 if (pfctl_define_table(tbl->pt_name, PFR_TFLAG_CONST, 1, pf->anchor, 1272 tbl->pt_buf, pf->tticket)) { 1273 warn("failed to create table %s", tbl->pt_name); 1274 return (1); 1275 } 1276 return (0); 1277 } 1278 1279 /* 1280 * Partition the flat ruleset into a list of distinct superblocks 1281 */ 1282 int 1283 construct_superblocks(struct pfctl *pf, struct pf_opt_queue *opt_queue, 1284 struct superblocks *superblocks) 1285 { 1286 struct superblock *block = NULL; 1287 struct pf_opt_rule *por; 1288 int i; 1289 1290 while (!TAILQ_EMPTY(opt_queue)) { 1291 por = TAILQ_FIRST(opt_queue); 1292 TAILQ_REMOVE(opt_queue, por, por_entry); 1293 if (block == NULL || !superblock_inclusive(block, por)) { 1294 if ((block = calloc(1, sizeof(*block))) == NULL) { 1295 warn("calloc"); 1296 return (1); 1297 } 1298 TAILQ_INIT(&block->sb_rules); 1299 for (i = 0; i < PF_SKIP_COUNT; i++) 1300 TAILQ_INIT(&block->sb_skipsteps[i]); 1301 TAILQ_INSERT_TAIL(superblocks, block, sb_entry); 1302 } 1303 TAILQ_INSERT_TAIL(&block->sb_rules, por, por_entry); 1304 } 1305 1306 return (0); 1307 } 1308 1309 1310 /* 1311 * Compare two rule addresses 1312 */ 1313 int 1314 addrs_equal(struct pf_rule_addr *a, struct pf_rule_addr *b) 1315 { 1316 if (a->neg != b->neg) 1317 return (0); 1318 return (memcmp(&a->addr, &b->addr, sizeof(a->addr)) == 0); 1319 } 1320 1321 1322 /* 1323 * The addresses are not equal, but can we combine them into one table? 1324 */ 1325 int 1326 addrs_combineable(struct pf_rule_addr *a, struct pf_rule_addr *b) 1327 { 1328 if (a->addr.type != PF_ADDR_ADDRMASK || 1329 b->addr.type != PF_ADDR_ADDRMASK) 1330 return (0); 1331 if (a->neg != b->neg || a->port_op != b->port_op || 1332 a->port[0] != b->port[0] || a->port[1] != b->port[1]) 1333 return (0); 1334 return (1); 1335 } 1336 1337 1338 /* 1339 * Are we allowed to combine these two rules 1340 */ 1341 int 1342 rules_combineable(struct pf_rule *p1, struct pf_rule *p2) 1343 { 1344 struct pf_rule a, b; 1345 1346 comparable_rule(&a, p1, COMBINED); 1347 comparable_rule(&b, p2, COMBINED); 1348 return (memcmp(&a, &b, sizeof(a)) == 0); 1349 } 1350 1351 1352 /* 1353 * Can a rule be included inside a superblock 1354 */ 1355 int 1356 superblock_inclusive(struct superblock *block, struct pf_opt_rule *por) 1357 { 1358 struct pf_rule a, b; 1359 int i, j; 1360 1361 /* First check for hard breaks */ 1362 for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++) { 1363 if (pf_rule_desc[i].prf_type == BARRIER) { 1364 for (j = 0; j < pf_rule_desc[i].prf_size; j++) 1365 if (((char *)&por->por_rule)[j + 1366 pf_rule_desc[i].prf_offset] != 0) 1367 return (0); 1368 } 1369 } 1370 1371 /* 'anchor' heads and per-rule src-track are also hard breaks */ 1372 if (por->por_anchor[0] != '\0' || 1373 (por->por_rule.rule_flag & PFRULE_RULESRCTRACK)) 1374 return (0); 1375 1376 comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule, NOMERGE); 1377 comparable_rule(&b, &por->por_rule, NOMERGE); 1378 if (strcmp(TAILQ_FIRST(&block->sb_rules)->por_anchor, 1379 por->por_anchor) == 0 && memcmp(&a, &b, sizeof(a)) == 0) 1380 return (1); 1381 1382 #ifdef OPT_DEBUG 1383 for (i = 0; i < sizeof(por->por_rule); i++) { 1384 int closest = -1; 1385 if (((u_int8_t *)&a)[i] != ((u_int8_t *)&b)[i]) { 1386 for (j = 0; j < sizeof(pf_rule_desc) / 1387 sizeof(*pf_rule_desc); j++) { 1388 if (i >= pf_rule_desc[j].prf_offset && 1389 i < pf_rule_desc[j].prf_offset + 1390 pf_rule_desc[j].prf_size) { 1391 DEBUG("superblock break @ %d due to %s", 1392 por->por_rule.nr, 1393 pf_rule_desc[j].prf_name); 1394 return (0); 1395 } 1396 if (i > pf_rule_desc[j].prf_offset) { 1397 if (closest == -1 || 1398 i-pf_rule_desc[j].prf_offset < 1399 i-pf_rule_desc[closest].prf_offset) 1400 closest = j; 1401 } 1402 } 1403 1404 if (closest >= 0) 1405 DEBUG("superblock break @ %d on %s+%xh", 1406 por->por_rule.nr, 1407 pf_rule_desc[closest].prf_name, 1408 i - pf_rule_desc[closest].prf_offset - 1409 pf_rule_desc[closest].prf_size); 1410 else 1411 DEBUG("superblock break @ %d on field @ %d", 1412 por->por_rule.nr, i); 1413 return (0); 1414 } 1415 } 1416 #endif /* OPT_DEBUG */ 1417 1418 return (0); 1419 } 1420 1421 1422 /* 1423 * Make a rule that can directly compared by memcmp() 1424 */ 1425 void 1426 comparable_rule(struct pf_rule *dst, const struct pf_rule *src, int type) 1427 { 1428 int i; 1429 /* 1430 * To simplify the comparison, we just zero out the fields that are 1431 * allowed to be different and then do a simple memcmp() 1432 */ 1433 memcpy(dst, src, sizeof(*dst)); 1434 for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++) 1435 if (pf_rule_desc[i].prf_type >= type) { 1436 #ifdef OPT_DEBUG 1437 assert(pf_rule_desc[i].prf_type != NEVER || 1438 *(((char *)dst) + pf_rule_desc[i].prf_offset) == 0); 1439 #endif /* OPT_DEBUG */ 1440 memset(((char *)dst) + pf_rule_desc[i].prf_offset, 0, 1441 pf_rule_desc[i].prf_size); 1442 } 1443 } 1444 1445 1446 /* 1447 * Remove superset information from two rules so we can directly compare them 1448 * with memcmp() 1449 */ 1450 void 1451 exclude_supersets(struct pf_rule *super, struct pf_rule *sub) 1452 { 1453 if (super->ifname[0] == '\0') 1454 memset(sub->ifname, 0, sizeof(sub->ifname)); 1455 if (super->direction == PF_INOUT) 1456 sub->direction = PF_INOUT; 1457 if ((super->proto == 0 || super->proto == sub->proto) && 1458 super->flags == 0 && super->flagset == 0 && (sub->flags || 1459 sub->flagset)) { 1460 sub->flags = super->flags; 1461 sub->flagset = super->flagset; 1462 } 1463 if (super->proto == 0) 1464 sub->proto = 0; 1465 1466 if (super->src.port_op == 0) { 1467 sub->src.port_op = 0; 1468 sub->src.port[0] = 0; 1469 sub->src.port[1] = 0; 1470 } 1471 if (super->dst.port_op == 0) { 1472 sub->dst.port_op = 0; 1473 sub->dst.port[0] = 0; 1474 sub->dst.port[1] = 0; 1475 } 1476 1477 if (super->src.addr.type == PF_ADDR_ADDRMASK && !super->src.neg && 1478 !sub->src.neg && super->src.addr.v.a.mask.addr32[0] == 0 && 1479 super->src.addr.v.a.mask.addr32[1] == 0 && 1480 super->src.addr.v.a.mask.addr32[2] == 0 && 1481 super->src.addr.v.a.mask.addr32[3] == 0) 1482 memset(&sub->src.addr, 0, sizeof(sub->src.addr)); 1483 else if (super->src.addr.type == PF_ADDR_ADDRMASK && 1484 sub->src.addr.type == PF_ADDR_ADDRMASK && 1485 super->src.neg == sub->src.neg && 1486 super->af == sub->af && 1487 unmask(&super->src.addr.v.a.mask, super->af) < 1488 unmask(&sub->src.addr.v.a.mask, sub->af) && 1489 super->src.addr.v.a.addr.addr32[0] == 1490 (sub->src.addr.v.a.addr.addr32[0] & 1491 super->src.addr.v.a.mask.addr32[0]) && 1492 super->src.addr.v.a.addr.addr32[1] == 1493 (sub->src.addr.v.a.addr.addr32[1] & 1494 super->src.addr.v.a.mask.addr32[1]) && 1495 super->src.addr.v.a.addr.addr32[2] == 1496 (sub->src.addr.v.a.addr.addr32[2] & 1497 super->src.addr.v.a.mask.addr32[2]) && 1498 super->src.addr.v.a.addr.addr32[3] == 1499 (sub->src.addr.v.a.addr.addr32[3] & 1500 super->src.addr.v.a.mask.addr32[3])) { 1501 /* sub->src.addr is a subset of super->src.addr/mask */ 1502 memcpy(&sub->src.addr, &super->src.addr, sizeof(sub->src.addr)); 1503 } 1504 1505 if (super->dst.addr.type == PF_ADDR_ADDRMASK && !super->dst.neg && 1506 !sub->dst.neg && super->dst.addr.v.a.mask.addr32[0] == 0 && 1507 super->dst.addr.v.a.mask.addr32[1] == 0 && 1508 super->dst.addr.v.a.mask.addr32[2] == 0 && 1509 super->dst.addr.v.a.mask.addr32[3] == 0) 1510 memset(&sub->dst.addr, 0, sizeof(sub->dst.addr)); 1511 else if (super->dst.addr.type == PF_ADDR_ADDRMASK && 1512 sub->dst.addr.type == PF_ADDR_ADDRMASK && 1513 super->dst.neg == sub->dst.neg && 1514 super->af == sub->af && 1515 unmask(&super->dst.addr.v.a.mask, super->af) < 1516 unmask(&sub->dst.addr.v.a.mask, sub->af) && 1517 super->dst.addr.v.a.addr.addr32[0] == 1518 (sub->dst.addr.v.a.addr.addr32[0] & 1519 super->dst.addr.v.a.mask.addr32[0]) && 1520 super->dst.addr.v.a.addr.addr32[1] == 1521 (sub->dst.addr.v.a.addr.addr32[1] & 1522 super->dst.addr.v.a.mask.addr32[1]) && 1523 super->dst.addr.v.a.addr.addr32[2] == 1524 (sub->dst.addr.v.a.addr.addr32[2] & 1525 super->dst.addr.v.a.mask.addr32[2]) && 1526 super->dst.addr.v.a.addr.addr32[3] == 1527 (sub->dst.addr.v.a.addr.addr32[3] & 1528 super->dst.addr.v.a.mask.addr32[3])) { 1529 /* sub->dst.addr is a subset of super->dst.addr/mask */ 1530 memcpy(&sub->dst.addr, &super->dst.addr, sizeof(sub->dst.addr)); 1531 } 1532 1533 if (super->af == 0) 1534 sub->af = 0; 1535 } 1536 1537 1538 void 1539 superblock_free(struct pfctl *pf, struct superblock *block) 1540 { 1541 struct pf_opt_rule *por; 1542 while ((por = TAILQ_FIRST(&block->sb_rules))) { 1543 TAILQ_REMOVE(&block->sb_rules, por, por_entry); 1544 if (por->por_src_tbl) { 1545 if (por->por_src_tbl->pt_buf) { 1546 pfr_buf_clear(por->por_src_tbl->pt_buf); 1547 free(por->por_src_tbl->pt_buf); 1548 } 1549 free(por->por_src_tbl); 1550 } 1551 if (por->por_dst_tbl) { 1552 if (por->por_dst_tbl->pt_buf) { 1553 pfr_buf_clear(por->por_dst_tbl->pt_buf); 1554 free(por->por_dst_tbl->pt_buf); 1555 } 1556 free(por->por_dst_tbl); 1557 } 1558 free(por); 1559 } 1560 if (block->sb_profiled_block) 1561 superblock_free(pf, block->sb_profiled_block); 1562 free(block); 1563 } 1564 1565