1 /* $OpenBSD: pfctl_optimize.c,v 1.17 2008/05/06 03:45:21 mpf 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/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 <libpfctl.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 its 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 static 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 pfctl_rule, field), \
101 sizeof(((struct pfctl_rule *)0)->field)}
102
103
104 /*
105 * The presence of these fields in a rule put the rule in its 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 PF_RULE_FIELD(max_src_states, BARRIER),
114 PF_RULE_FIELD(max_src_conn, BARRIER),
115 PF_RULE_FIELD(max_src_conn_rate, BARRIER),
116 PF_RULE_FIELD(anchor, BARRIER), /* for now */
117
118 /*
119 * These fields must be the same between all rules in the same superblock.
120 * These rules are allowed to be re-ordered but only among like rules.
121 * For instance we can re-order all 'tag "foo"' rules because they have the
122 * same tag. But we can not re-order between a 'tag "foo"' and a
123 * 'tag "bar"' since that would change the meaning of the ruleset.
124 */
125 PF_RULE_FIELD(tagname, BREAK),
126 PF_RULE_FIELD(keep_state, BREAK),
127 PF_RULE_FIELD(qname, BREAK),
128 PF_RULE_FIELD(pqname, BREAK),
129 PF_RULE_FIELD(rt, BREAK),
130 PF_RULE_FIELD(allow_opts, BREAK),
131 PF_RULE_FIELD(rule_flag, BREAK),
132 PF_RULE_FIELD(action, BREAK),
133 PF_RULE_FIELD(log, BREAK),
134 PF_RULE_FIELD(quick, BREAK),
135 PF_RULE_FIELD(return_ttl, BREAK),
136 PF_RULE_FIELD(overload_tblname, BREAK),
137 PF_RULE_FIELD(flush, BREAK),
138 PF_RULE_FIELD(rpool, BREAK),
139 PF_RULE_FIELD(logif, BREAK),
140
141 /*
142 * Any fields not listed in this structure act as BREAK fields
143 */
144
145
146 /*
147 * These fields must not differ when we merge two rules together but
148 * their difference isn't enough to put the rules in different superblocks.
149 * There are no problems re-ordering any rules with these fields.
150 */
151 PF_RULE_FIELD(af, NOMERGE),
152 PF_RULE_FIELD(ifnot, NOMERGE),
153 PF_RULE_FIELD(ifname, NOMERGE), /* hack for IF groups */
154 PF_RULE_FIELD(match_tag_not, NOMERGE),
155 PF_RULE_FIELD(match_tagname, NOMERGE),
156 PF_RULE_FIELD(os_fingerprint, NOMERGE),
157 PF_RULE_FIELD(timeout, NOMERGE),
158 PF_RULE_FIELD(return_icmp, NOMERGE),
159 PF_RULE_FIELD(return_icmp6, NOMERGE),
160 PF_RULE_FIELD(uid, NOMERGE),
161 PF_RULE_FIELD(gid, NOMERGE),
162 PF_RULE_FIELD(direction, NOMERGE),
163 PF_RULE_FIELD(proto, NOMERGE),
164 PF_RULE_FIELD(type, NOMERGE),
165 PF_RULE_FIELD(code, NOMERGE),
166 PF_RULE_FIELD(flags, NOMERGE),
167 PF_RULE_FIELD(flagset, NOMERGE),
168 PF_RULE_FIELD(tos, NOMERGE),
169 PF_RULE_FIELD(src.port, NOMERGE),
170 PF_RULE_FIELD(dst.port, NOMERGE),
171 PF_RULE_FIELD(src.port_op, NOMERGE),
172 PF_RULE_FIELD(dst.port_op, NOMERGE),
173 PF_RULE_FIELD(src.neg, NOMERGE),
174 PF_RULE_FIELD(dst.neg, NOMERGE),
175
176 /* These fields can be merged */
177 PF_RULE_FIELD(src.addr, COMBINED),
178 PF_RULE_FIELD(dst.addr, COMBINED),
179
180 /* We just don't care about these fields. They're set by the kernel */
181 PF_RULE_FIELD(skip, DC),
182 PF_RULE_FIELD(evaluations, DC),
183 PF_RULE_FIELD(packets, DC),
184 PF_RULE_FIELD(bytes, DC),
185 PF_RULE_FIELD(kif, DC),
186 PF_RULE_FIELD(states_cur, DC),
187 PF_RULE_FIELD(states_tot, DC),
188 PF_RULE_FIELD(src_nodes, DC),
189 PF_RULE_FIELD(nr, DC),
190 PF_RULE_FIELD(entries, DC),
191 PF_RULE_FIELD(qid, DC),
192 PF_RULE_FIELD(pqid, DC),
193 PF_RULE_FIELD(anchor_relative, DC),
194 PF_RULE_FIELD(anchor_wildcard, DC),
195 PF_RULE_FIELD(tag, DC),
196 PF_RULE_FIELD(match_tag, DC),
197 PF_RULE_FIELD(overload_tbl, DC),
198
199 /* These fields should never be set in a PASS/BLOCK rule */
200 PF_RULE_FIELD(natpass, NEVER),
201 PF_RULE_FIELD(max_mss, NEVER),
202 PF_RULE_FIELD(min_ttl, NEVER),
203 PF_RULE_FIELD(set_tos, NEVER),
204 };
205
206
207
208 int add_opt_table(struct pfctl *, struct pf_opt_tbl **, sa_family_t,
209 struct pf_rule_addr *);
210 int addrs_combineable(struct pf_rule_addr *, struct pf_rule_addr *);
211 int addrs_equal(struct pf_rule_addr *, struct pf_rule_addr *);
212 int block_feedback(struct pfctl *, struct superblock *);
213 int combine_rules(struct pfctl *, struct superblock *);
214 void comparable_rule(struct pfctl_rule *, const struct pfctl_rule *, int);
215 int construct_superblocks(struct pfctl *, struct pf_opt_queue *,
216 struct superblocks *);
217 void exclude_supersets(struct pfctl_rule *, struct pfctl_rule *);
218 int interface_group(const char *);
219 int load_feedback_profile(struct pfctl *, struct superblocks *);
220 int optimize_superblock(struct pfctl *, struct superblock *);
221 int pf_opt_create_table(struct pfctl *, struct pf_opt_tbl *);
222 void remove_from_skipsteps(struct skiplist *, struct superblock *,
223 struct pf_opt_rule *, struct pf_skip_step *);
224 int remove_identical_rules(struct pfctl *, struct superblock *);
225 int reorder_rules(struct pfctl *, struct superblock *, int);
226 int rules_combineable(struct pfctl_rule *, struct pfctl_rule *);
227 void skip_append(struct superblock *, int, struct pf_skip_step *,
228 struct pf_opt_rule *);
229 int skip_compare(int, struct pf_skip_step *, struct pf_opt_rule *);
230 void skip_init(void);
231 int skip_cmp_af(struct pfctl_rule *, struct pfctl_rule *);
232 int skip_cmp_dir(struct pfctl_rule *, struct pfctl_rule *);
233 int skip_cmp_dst_addr(struct pfctl_rule *, struct pfctl_rule *);
234 int skip_cmp_dst_port(struct pfctl_rule *, struct pfctl_rule *);
235 int skip_cmp_ifp(struct pfctl_rule *, struct pfctl_rule *);
236 int skip_cmp_proto(struct pfctl_rule *, struct pfctl_rule *);
237 int skip_cmp_src_addr(struct pfctl_rule *, struct pfctl_rule *);
238 int skip_cmp_src_port(struct pfctl_rule *, struct pfctl_rule *);
239 int superblock_inclusive(struct superblock *, struct pf_opt_rule *);
240 void superblock_free(struct pfctl *, struct superblock *);
241
242
243 static int (*skip_comparitors[PF_SKIP_COUNT])(struct pfctl_rule *,
244 struct pfctl_rule *);
245 static const char *skip_comparitors_names[PF_SKIP_COUNT];
246 #define PF_SKIP_COMPARITORS { \
247 { "ifp", PF_SKIP_IFP, skip_cmp_ifp }, \
248 { "dir", PF_SKIP_DIR, skip_cmp_dir }, \
249 { "af", PF_SKIP_AF, skip_cmp_af }, \
250 { "proto", PF_SKIP_PROTO, skip_cmp_proto }, \
251 { "saddr", PF_SKIP_SRC_ADDR, skip_cmp_src_addr }, \
252 { "daddr", PF_SKIP_DST_ADDR, skip_cmp_dst_addr }, \
253 { "sport", PF_SKIP_SRC_PORT, skip_cmp_src_port }, \
254 { "dport", PF_SKIP_DST_PORT, skip_cmp_dst_port } \
255 }
256
257 static struct pfr_buffer table_buffer;
258 static int table_identifier;
259
260
261 int
pfctl_optimize_ruleset(struct pfctl * pf,struct pfctl_ruleset * rs)262 pfctl_optimize_ruleset(struct pfctl *pf, struct pfctl_ruleset *rs)
263 {
264 struct superblocks superblocks;
265 struct pf_opt_queue opt_queue;
266 struct superblock *block;
267 struct pf_opt_rule *por;
268 struct pfctl_rule *r;
269 struct pfctl_rulequeue *old_rules;
270
271 DEBUG("optimizing ruleset");
272 memset(&table_buffer, 0, sizeof(table_buffer));
273 skip_init();
274 TAILQ_INIT(&opt_queue);
275
276 old_rules = rs->rules[PF_RULESET_FILTER].active.ptr;
277 rs->rules[PF_RULESET_FILTER].active.ptr =
278 rs->rules[PF_RULESET_FILTER].inactive.ptr;
279 rs->rules[PF_RULESET_FILTER].inactive.ptr = old_rules;
280
281 /*
282 * XXX expanding the pf_opt_rule format throughout pfctl might allow
283 * us to avoid all this copying.
284 */
285 while ((r = TAILQ_FIRST(rs->rules[PF_RULESET_FILTER].inactive.ptr))
286 != NULL) {
287 TAILQ_REMOVE(rs->rules[PF_RULESET_FILTER].inactive.ptr, r,
288 entries);
289 if ((por = calloc(1, sizeof(*por))) == NULL)
290 err(1, "calloc");
291 memcpy(&por->por_rule, r, sizeof(*r));
292 if (TAILQ_FIRST(&r->rpool.list) != NULL) {
293 TAILQ_INIT(&por->por_rule.rpool.list);
294 pfctl_move_pool(&r->rpool, &por->por_rule.rpool);
295 } else
296 bzero(&por->por_rule.rpool,
297 sizeof(por->por_rule.rpool));
298
299
300 TAILQ_INSERT_TAIL(&opt_queue, por, por_entry);
301 }
302
303 TAILQ_INIT(&superblocks);
304 if (construct_superblocks(pf, &opt_queue, &superblocks))
305 goto error;
306
307 if (pf->optimize & PF_OPTIMIZE_PROFILE) {
308 if (load_feedback_profile(pf, &superblocks))
309 goto error;
310 }
311
312 TAILQ_FOREACH(block, &superblocks, sb_entry) {
313 if (optimize_superblock(pf, block))
314 goto error;
315 }
316
317 rs->anchor->refcnt = 0;
318 while ((block = TAILQ_FIRST(&superblocks))) {
319 TAILQ_REMOVE(&superblocks, block, sb_entry);
320
321 while ((por = TAILQ_FIRST(&block->sb_rules))) {
322 TAILQ_REMOVE(&block->sb_rules, por, por_entry);
323 por->por_rule.nr = rs->anchor->refcnt++;
324 if ((r = calloc(1, sizeof(*r))) == NULL)
325 err(1, "calloc");
326 memcpy(r, &por->por_rule, sizeof(*r));
327 TAILQ_INIT(&r->rpool.list);
328 pfctl_move_pool(&por->por_rule.rpool, &r->rpool);
329 TAILQ_INSERT_TAIL(
330 rs->rules[PF_RULESET_FILTER].active.ptr,
331 r, entries);
332 free(por);
333 }
334 free(block);
335 }
336
337 return (0);
338
339 error:
340 while ((por = TAILQ_FIRST(&opt_queue))) {
341 TAILQ_REMOVE(&opt_queue, por, por_entry);
342 if (por->por_src_tbl) {
343 pfr_buf_clear(por->por_src_tbl->pt_buf);
344 free(por->por_src_tbl->pt_buf);
345 free(por->por_src_tbl);
346 }
347 if (por->por_dst_tbl) {
348 pfr_buf_clear(por->por_dst_tbl->pt_buf);
349 free(por->por_dst_tbl->pt_buf);
350 free(por->por_dst_tbl);
351 }
352 free(por);
353 }
354 while ((block = TAILQ_FIRST(&superblocks))) {
355 TAILQ_REMOVE(&superblocks, block, sb_entry);
356 superblock_free(pf, block);
357 }
358 return (1);
359 }
360
361
362 /*
363 * Go ahead and optimize a superblock
364 */
365 int
optimize_superblock(struct pfctl * pf,struct superblock * block)366 optimize_superblock(struct pfctl *pf, struct superblock *block)
367 {
368 #ifdef OPT_DEBUG
369 struct pf_opt_rule *por;
370 #endif /* OPT_DEBUG */
371
372 /* We have a few optimization passes:
373 * 1) remove duplicate rules or rules that are a subset of other
374 * rules
375 * 2) combine otherwise identical rules with different IP addresses
376 * into a single rule and put the addresses in a table.
377 * 3) re-order the rules to improve kernel skip steps
378 * 4) re-order the 'quick' rules based on feedback from the
379 * active ruleset statistics
380 *
381 * XXX combine_rules() doesn't combine v4 and v6 rules. would just
382 * have to keep af in the table container, make af 'COMBINE' and
383 * twiddle the af on the merged rule
384 * XXX maybe add a weighting to the metric on skipsteps when doing
385 * reordering. sometimes two sequential tables will be better
386 * that four consecutive interfaces.
387 * XXX need to adjust the skipstep count of everything after PROTO,
388 * since they aren't actually checked on a proto mismatch in
389 * pf_test_{tcp, udp, icmp}()
390 * XXX should i treat proto=0, af=0 or dir=0 special in skepstep
391 * calculation since they are a DC?
392 * XXX keep last skiplist of last superblock to influence this
393 * superblock. '5 inet6 log' should make '3 inet6' come before '4
394 * inet' in the next superblock.
395 * XXX would be useful to add tables for ports
396 * XXX we can also re-order some mutually exclusive superblocks to
397 * try merging superblocks before any of these optimization passes.
398 * for instance a single 'log in' rule in the middle of non-logging
399 * out rules.
400 */
401
402 /* shortcut. there will be a lot of 1-rule superblocks */
403 if (!TAILQ_NEXT(TAILQ_FIRST(&block->sb_rules), por_entry))
404 return (0);
405
406 #ifdef OPT_DEBUG
407 printf("--- Superblock ---\n");
408 TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
409 printf(" ");
410 print_rule(&por->por_rule, por->por_rule.anchor ?
411 por->por_rule.anchor->name : "", 1, 0);
412 }
413 #endif /* OPT_DEBUG */
414
415
416 if (remove_identical_rules(pf, block))
417 return (1);
418 if (combine_rules(pf, block))
419 return (1);
420 if ((pf->optimize & PF_OPTIMIZE_PROFILE) &&
421 TAILQ_FIRST(&block->sb_rules)->por_rule.quick &&
422 block->sb_profiled_block) {
423 if (block_feedback(pf, block))
424 return (1);
425 } else if (reorder_rules(pf, block, 0)) {
426 return (1);
427 }
428
429 /*
430 * Don't add any optimization passes below reorder_rules(). It will
431 * have divided superblocks into smaller blocks for further refinement
432 * and doesn't put them back together again. What once was a true
433 * superblock might have been split into multiple superblocks.
434 */
435
436 #ifdef OPT_DEBUG
437 printf("--- END Superblock ---\n");
438 #endif /* OPT_DEBUG */
439 return (0);
440 }
441
442
443 /*
444 * Optimization pass #1: remove identical rules
445 */
446 int
remove_identical_rules(struct pfctl * pf,struct superblock * block)447 remove_identical_rules(struct pfctl *pf, struct superblock *block)
448 {
449 struct pf_opt_rule *por1, *por2, *por_next, *por2_next;
450 struct pfctl_rule a, a2, b, b2;
451
452 for (por1 = TAILQ_FIRST(&block->sb_rules); por1; por1 = por_next) {
453 por_next = TAILQ_NEXT(por1, por_entry);
454 for (por2 = por_next; por2; por2 = por2_next) {
455 por2_next = TAILQ_NEXT(por2, por_entry);
456 comparable_rule(&a, &por1->por_rule, DC);
457 comparable_rule(&b, &por2->por_rule, DC);
458 memcpy(&a2, &a, sizeof(a2));
459 memcpy(&b2, &b, sizeof(b2));
460
461 exclude_supersets(&a, &b);
462 exclude_supersets(&b2, &a2);
463 if (memcmp(&a, &b, sizeof(a)) == 0) {
464 DEBUG("removing identical rule nr%d = *nr%d*",
465 por1->por_rule.nr, por2->por_rule.nr);
466 TAILQ_REMOVE(&block->sb_rules, por2, por_entry);
467 if (por_next == por2)
468 por_next = TAILQ_NEXT(por1, por_entry);
469 free(por2);
470 } else if (memcmp(&a2, &b2, sizeof(a2)) == 0) {
471 DEBUG("removing identical rule *nr%d* = nr%d",
472 por1->por_rule.nr, por2->por_rule.nr);
473 TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
474 free(por1);
475 break;
476 }
477 }
478 }
479
480 return (0);
481 }
482
483
484 /*
485 * Optimization pass #2: combine similar rules with different addresses
486 * into a single rule and a table
487 */
488 int
combine_rules(struct pfctl * pf,struct superblock * block)489 combine_rules(struct pfctl *pf, struct superblock *block)
490 {
491 struct pf_opt_rule *p1, *p2, *por_next;
492 int src_eq, dst_eq;
493
494 if ((pf->loadopt & PFCTL_FLAG_TABLE) == 0) {
495 warnx("Must enable table loading for optimizations");
496 return (1);
497 }
498
499 /* First we make a pass to combine the rules. O(n log n) */
500 TAILQ_FOREACH(p1, &block->sb_rules, por_entry) {
501 for (p2 = TAILQ_NEXT(p1, por_entry); p2; p2 = por_next) {
502 por_next = TAILQ_NEXT(p2, por_entry);
503
504 src_eq = addrs_equal(&p1->por_rule.src,
505 &p2->por_rule.src);
506 dst_eq = addrs_equal(&p1->por_rule.dst,
507 &p2->por_rule.dst);
508
509 if (src_eq && !dst_eq && p1->por_src_tbl == NULL &&
510 p2->por_dst_tbl == NULL &&
511 p2->por_src_tbl == NULL &&
512 rules_combineable(&p1->por_rule, &p2->por_rule) &&
513 addrs_combineable(&p1->por_rule.dst,
514 &p2->por_rule.dst)) {
515 DEBUG("can combine rules nr%d = nr%d",
516 p1->por_rule.nr, p2->por_rule.nr);
517 if (p1->por_dst_tbl == NULL &&
518 add_opt_table(pf, &p1->por_dst_tbl,
519 p1->por_rule.af, &p1->por_rule.dst))
520 return (1);
521 if (add_opt_table(pf, &p1->por_dst_tbl,
522 p1->por_rule.af, &p2->por_rule.dst))
523 return (1);
524 p2->por_dst_tbl = p1->por_dst_tbl;
525 if (p1->por_dst_tbl->pt_rulecount >=
526 TABLE_THRESHOLD) {
527 TAILQ_REMOVE(&block->sb_rules, p2,
528 por_entry);
529 free(p2);
530 }
531 } else if (!src_eq && dst_eq && p1->por_dst_tbl == NULL
532 && p2->por_src_tbl == NULL &&
533 p2->por_dst_tbl == NULL &&
534 rules_combineable(&p1->por_rule, &p2->por_rule) &&
535 addrs_combineable(&p1->por_rule.src,
536 &p2->por_rule.src)) {
537 DEBUG("can combine rules nr%d = nr%d",
538 p1->por_rule.nr, p2->por_rule.nr);
539 if (p1->por_src_tbl == NULL &&
540 add_opt_table(pf, &p1->por_src_tbl,
541 p1->por_rule.af, &p1->por_rule.src))
542 return (1);
543 if (add_opt_table(pf, &p1->por_src_tbl,
544 p1->por_rule.af, &p2->por_rule.src))
545 return (1);
546 p2->por_src_tbl = p1->por_src_tbl;
547 if (p1->por_src_tbl->pt_rulecount >=
548 TABLE_THRESHOLD) {
549 TAILQ_REMOVE(&block->sb_rules, p2,
550 por_entry);
551 free(p2);
552 }
553 }
554 }
555 }
556
557
558 /*
559 * Then we make a final pass to create a valid table name and
560 * insert the name into the rules.
561 */
562 for (p1 = TAILQ_FIRST(&block->sb_rules); p1; p1 = por_next) {
563 por_next = TAILQ_NEXT(p1, por_entry);
564 assert(p1->por_src_tbl == NULL || p1->por_dst_tbl == NULL);
565
566 if (p1->por_src_tbl && p1->por_src_tbl->pt_rulecount >=
567 TABLE_THRESHOLD) {
568 if (p1->por_src_tbl->pt_generated) {
569 /* This rule is included in a table */
570 TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
571 free(p1);
572 continue;
573 }
574 p1->por_src_tbl->pt_generated = 1;
575
576 if ((pf->opts & PF_OPT_NOACTION) == 0 &&
577 pf_opt_create_table(pf, p1->por_src_tbl))
578 return (1);
579
580 pf->tdirty = 1;
581
582 if (pf->opts & PF_OPT_VERBOSE)
583 print_tabledef(p1->por_src_tbl->pt_name,
584 PFR_TFLAG_CONST, 1,
585 &p1->por_src_tbl->pt_nodes);
586
587 memset(&p1->por_rule.src.addr, 0,
588 sizeof(p1->por_rule.src.addr));
589 p1->por_rule.src.addr.type = PF_ADDR_TABLE;
590 strlcpy(p1->por_rule.src.addr.v.tblname,
591 p1->por_src_tbl->pt_name,
592 sizeof(p1->por_rule.src.addr.v.tblname));
593
594 pfr_buf_clear(p1->por_src_tbl->pt_buf);
595 free(p1->por_src_tbl->pt_buf);
596 p1->por_src_tbl->pt_buf = NULL;
597 }
598 if (p1->por_dst_tbl && p1->por_dst_tbl->pt_rulecount >=
599 TABLE_THRESHOLD) {
600 if (p1->por_dst_tbl->pt_generated) {
601 /* This rule is included in a table */
602 TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
603 free(p1);
604 continue;
605 }
606 p1->por_dst_tbl->pt_generated = 1;
607
608 if ((pf->opts & PF_OPT_NOACTION) == 0 &&
609 pf_opt_create_table(pf, p1->por_dst_tbl))
610 return (1);
611 pf->tdirty = 1;
612
613 if (pf->opts & PF_OPT_VERBOSE)
614 print_tabledef(p1->por_dst_tbl->pt_name,
615 PFR_TFLAG_CONST, 1,
616 &p1->por_dst_tbl->pt_nodes);
617
618 memset(&p1->por_rule.dst.addr, 0,
619 sizeof(p1->por_rule.dst.addr));
620 p1->por_rule.dst.addr.type = PF_ADDR_TABLE;
621 strlcpy(p1->por_rule.dst.addr.v.tblname,
622 p1->por_dst_tbl->pt_name,
623 sizeof(p1->por_rule.dst.addr.v.tblname));
624
625 pfr_buf_clear(p1->por_dst_tbl->pt_buf);
626 free(p1->por_dst_tbl->pt_buf);
627 p1->por_dst_tbl->pt_buf = NULL;
628 }
629 }
630
631 return (0);
632 }
633
634
635 /*
636 * Optimization pass #3: re-order rules to improve skip steps
637 */
638 int
reorder_rules(struct pfctl * pf,struct superblock * block,int depth)639 reorder_rules(struct pfctl *pf, struct superblock *block, int depth)
640 {
641 struct superblock *newblock;
642 struct pf_skip_step *skiplist;
643 struct pf_opt_rule *por;
644 int i, largest, largest_list, rule_count = 0;
645 TAILQ_HEAD( , pf_opt_rule) head;
646
647 /*
648 * Calculate the best-case skip steps. We put each rule in a list
649 * of other rules with common fields
650 */
651 for (i = 0; i < PF_SKIP_COUNT; i++) {
652 TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
653 TAILQ_FOREACH(skiplist, &block->sb_skipsteps[i],
654 ps_entry) {
655 if (skip_compare(i, skiplist, por) == 0)
656 break;
657 }
658 if (skiplist == NULL) {
659 if ((skiplist = calloc(1, sizeof(*skiplist))) ==
660 NULL)
661 err(1, "calloc");
662 TAILQ_INIT(&skiplist->ps_rules);
663 TAILQ_INSERT_TAIL(&block->sb_skipsteps[i],
664 skiplist, ps_entry);
665 }
666 skip_append(block, i, skiplist, por);
667 }
668 }
669
670 TAILQ_FOREACH(por, &block->sb_rules, por_entry)
671 rule_count++;
672
673 /*
674 * Now we're going to ignore any fields that are identical between
675 * all of the rules in the superblock and those fields which differ
676 * between every rule in the superblock.
677 */
678 largest = 0;
679 for (i = 0; i < PF_SKIP_COUNT; i++) {
680 skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
681 if (skiplist->ps_count == rule_count) {
682 DEBUG("(%d) original skipstep '%s' is all rules",
683 depth, skip_comparitors_names[i]);
684 skiplist->ps_count = 0;
685 } else if (skiplist->ps_count == 1) {
686 skiplist->ps_count = 0;
687 } else {
688 DEBUG("(%d) original skipstep '%s' largest jump is %d",
689 depth, skip_comparitors_names[i],
690 skiplist->ps_count);
691 if (skiplist->ps_count > largest)
692 largest = skiplist->ps_count;
693 }
694 }
695 if (largest == 0) {
696 /* Ugh. There is NO commonality in the superblock on which
697 * optimize the skipsteps optimization.
698 */
699 goto done;
700 }
701
702 /*
703 * Now we're going to empty the superblock rule list and re-create
704 * it based on a more optimal skipstep order.
705 */
706 TAILQ_INIT(&head);
707 while ((por = TAILQ_FIRST(&block->sb_rules))) {
708 TAILQ_REMOVE(&block->sb_rules, por, por_entry);
709 TAILQ_INSERT_TAIL(&head, por, por_entry);
710 }
711
712
713 while (!TAILQ_EMPTY(&head)) {
714 largest = 1;
715
716 /*
717 * Find the most useful skip steps remaining
718 */
719 for (i = 0; i < PF_SKIP_COUNT; i++) {
720 skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
721 if (skiplist->ps_count > largest) {
722 largest = skiplist->ps_count;
723 largest_list = i;
724 }
725 }
726
727 if (largest <= 1) {
728 /*
729 * Nothing useful left. Leave remaining rules in order.
730 */
731 DEBUG("(%d) no more commonality for skip steps", depth);
732 while ((por = TAILQ_FIRST(&head))) {
733 TAILQ_REMOVE(&head, por, por_entry);
734 TAILQ_INSERT_TAIL(&block->sb_rules, por,
735 por_entry);
736 }
737 } else {
738 /*
739 * There is commonality. Extract those common rules
740 * and place them in the ruleset adjacent to each
741 * other.
742 */
743 skiplist = TAILQ_FIRST(&block->sb_skipsteps[
744 largest_list]);
745 DEBUG("(%d) skipstep '%s' largest jump is %d @ #%d",
746 depth, skip_comparitors_names[largest_list],
747 largest, TAILQ_FIRST(&TAILQ_FIRST(&block->
748 sb_skipsteps [largest_list])->ps_rules)->
749 por_rule.nr);
750 TAILQ_REMOVE(&block->sb_skipsteps[largest_list],
751 skiplist, ps_entry);
752
753
754 /*
755 * There may be further commonality inside these
756 * rules. So we'll split them off into they're own
757 * superblock and pass it back into the optimizer.
758 */
759 if (skiplist->ps_count > 2) {
760 if ((newblock = calloc(1, sizeof(*newblock)))
761 == NULL) {
762 warn("calloc");
763 return (1);
764 }
765 TAILQ_INIT(&newblock->sb_rules);
766 for (i = 0; i < PF_SKIP_COUNT; i++)
767 TAILQ_INIT(&newblock->sb_skipsteps[i]);
768 TAILQ_INSERT_BEFORE(block, newblock, sb_entry);
769 DEBUG("(%d) splitting off %d rules from superblock @ #%d",
770 depth, skiplist->ps_count,
771 TAILQ_FIRST(&skiplist->ps_rules)->
772 por_rule.nr);
773 } else {
774 newblock = block;
775 }
776
777 while ((por = TAILQ_FIRST(&skiplist->ps_rules))) {
778 TAILQ_REMOVE(&head, por, por_entry);
779 TAILQ_REMOVE(&skiplist->ps_rules, por,
780 por_skip_entry[largest_list]);
781 TAILQ_INSERT_TAIL(&newblock->sb_rules, por,
782 por_entry);
783
784 /* Remove this rule from all other skiplists */
785 remove_from_skipsteps(&block->sb_skipsteps[
786 largest_list], block, por, skiplist);
787 }
788 free(skiplist);
789 if (newblock != block)
790 if (reorder_rules(pf, newblock, depth + 1))
791 return (1);
792 }
793 }
794
795 done:
796 for (i = 0; i < PF_SKIP_COUNT; i++) {
797 while ((skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]))) {
798 TAILQ_REMOVE(&block->sb_skipsteps[i], skiplist,
799 ps_entry);
800 free(skiplist);
801 }
802 }
803
804 return (0);
805 }
806
807
808 /*
809 * Optimization pass #4: re-order 'quick' rules based on feedback from the
810 * currently running ruleset
811 */
812 int
block_feedback(struct pfctl * pf,struct superblock * block)813 block_feedback(struct pfctl *pf, struct superblock *block)
814 {
815 TAILQ_HEAD( , pf_opt_rule) queue;
816 struct pf_opt_rule *por1, *por2;
817 struct pfctl_rule a, b;
818
819
820 /*
821 * Walk through all of the profiled superblock's rules and copy
822 * the counters onto our rules.
823 */
824 TAILQ_FOREACH(por1, &block->sb_profiled_block->sb_rules, por_entry) {
825 comparable_rule(&a, &por1->por_rule, DC);
826 TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
827 if (por2->por_profile_count)
828 continue;
829 comparable_rule(&b, &por2->por_rule, DC);
830 if (memcmp(&a, &b, sizeof(a)) == 0) {
831 por2->por_profile_count =
832 por1->por_rule.packets[0] +
833 por1->por_rule.packets[1];
834 break;
835 }
836 }
837 }
838 superblock_free(pf, block->sb_profiled_block);
839 block->sb_profiled_block = NULL;
840
841 /*
842 * Now we pull all of the rules off the superblock and re-insert them
843 * in sorted order.
844 */
845
846 TAILQ_INIT(&queue);
847 while ((por1 = TAILQ_FIRST(&block->sb_rules)) != NULL) {
848 TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
849 TAILQ_INSERT_TAIL(&queue, por1, por_entry);
850 }
851
852 while ((por1 = TAILQ_FIRST(&queue)) != NULL) {
853 TAILQ_REMOVE(&queue, por1, por_entry);
854 /* XXX I should sort all of the unused rules based on skip steps */
855 TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
856 if (por1->por_profile_count > por2->por_profile_count) {
857 TAILQ_INSERT_BEFORE(por2, por1, por_entry);
858 break;
859 }
860 }
861 #ifdef __FreeBSD__
862 if (por2 == NULL)
863 #else
864 if (por2 == TAILQ_END(&block->sb_rules))
865 #endif
866 TAILQ_INSERT_TAIL(&block->sb_rules, por1, por_entry);
867 }
868
869 return (0);
870 }
871
872
873 /*
874 * Load the current ruleset from the kernel and try to associate them with
875 * the ruleset we're optimizing.
876 */
877 int
load_feedback_profile(struct pfctl * pf,struct superblocks * superblocks)878 load_feedback_profile(struct pfctl *pf, struct superblocks *superblocks)
879 {
880 char anchor_call[MAXPATHLEN] = "";
881 struct superblock *block, *blockcur;
882 struct superblocks prof_superblocks;
883 struct pf_opt_rule *por;
884 struct pf_opt_queue queue;
885 struct pfctl_rules_info rules;
886 struct pfctl_rule a, b, rule;
887 int nr, mnr;
888
889 TAILQ_INIT(&queue);
890 TAILQ_INIT(&prof_superblocks);
891
892 if (pfctl_get_rules_info_h(pf->h, &rules, PF_PASS, "")) {
893 warn("DIOCGETRULES");
894 return (1);
895 }
896 mnr = rules.nr;
897
898 DEBUG("Loading %d active rules for a feedback profile", mnr);
899 for (nr = 0; nr < mnr; ++nr) {
900 struct pfctl_ruleset *rs;
901 if ((por = calloc(1, sizeof(*por))) == NULL) {
902 warn("calloc");
903 return (1);
904 }
905
906 if (pfctl_get_rule_h(pf->h, nr, rules.ticket, "", PF_PASS,
907 &rule, anchor_call)) {
908 warn("DIOCGETRULENV");
909 return (1);
910 }
911 memcpy(&por->por_rule, &rule, sizeof(por->por_rule));
912 rs = pf_find_or_create_ruleset(anchor_call);
913 por->por_rule.anchor = rs->anchor;
914 if (TAILQ_EMPTY(&por->por_rule.rpool.list))
915 memset(&por->por_rule.rpool, 0,
916 sizeof(por->por_rule.rpool));
917 TAILQ_INSERT_TAIL(&queue, por, por_entry);
918
919 /* XXX pfctl_get_pool(pf->dev, &rule.rpool, nr, pr.ticket,
920 * PF_PASS, pf->anchor) ???
921 * ... pfctl_clear_pool(&rule.rpool)
922 */
923 }
924
925 if (construct_superblocks(pf, &queue, &prof_superblocks))
926 return (1);
927
928
929 /*
930 * Now we try to associate the active ruleset's superblocks with
931 * the superblocks we're compiling.
932 */
933 block = TAILQ_FIRST(superblocks);
934 blockcur = TAILQ_FIRST(&prof_superblocks);
935 while (block && blockcur) {
936 comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule,
937 BREAK);
938 comparable_rule(&b, &TAILQ_FIRST(&blockcur->sb_rules)->por_rule,
939 BREAK);
940 if (memcmp(&a, &b, sizeof(a)) == 0) {
941 /* The two superblocks lined up */
942 block->sb_profiled_block = blockcur;
943 } else {
944 DEBUG("superblocks don't line up between #%d and #%d",
945 TAILQ_FIRST(&block->sb_rules)->por_rule.nr,
946 TAILQ_FIRST(&blockcur->sb_rules)->por_rule.nr);
947 break;
948 }
949 block = TAILQ_NEXT(block, sb_entry);
950 blockcur = TAILQ_NEXT(blockcur, sb_entry);
951 }
952
953
954
955 /* Free any superblocks we couldn't link */
956 while (blockcur) {
957 block = TAILQ_NEXT(blockcur, sb_entry);
958 superblock_free(pf, blockcur);
959 blockcur = block;
960 }
961 return (0);
962 }
963
964
965 /*
966 * Compare a rule to a skiplist to see if the rule is a member
967 */
968 int
skip_compare(int skipnum,struct pf_skip_step * skiplist,struct pf_opt_rule * por)969 skip_compare(int skipnum, struct pf_skip_step *skiplist,
970 struct pf_opt_rule *por)
971 {
972 struct pfctl_rule *a, *b;
973 if (skipnum >= PF_SKIP_COUNT || skipnum < 0)
974 errx(1, "skip_compare() out of bounds");
975 a = &por->por_rule;
976 b = &TAILQ_FIRST(&skiplist->ps_rules)->por_rule;
977
978 return ((skip_comparitors[skipnum])(a, b));
979 }
980
981
982 /*
983 * Add a rule to a skiplist
984 */
985 void
skip_append(struct superblock * superblock,int skipnum,struct pf_skip_step * skiplist,struct pf_opt_rule * por)986 skip_append(struct superblock *superblock, int skipnum,
987 struct pf_skip_step *skiplist, struct pf_opt_rule *por)
988 {
989 struct pf_skip_step *prev;
990
991 skiplist->ps_count++;
992 TAILQ_INSERT_TAIL(&skiplist->ps_rules, por, por_skip_entry[skipnum]);
993
994 /* Keep the list of skiplists sorted by whichever is larger */
995 while ((prev = TAILQ_PREV(skiplist, skiplist, ps_entry)) &&
996 prev->ps_count < skiplist->ps_count) {
997 TAILQ_REMOVE(&superblock->sb_skipsteps[skipnum],
998 skiplist, ps_entry);
999 TAILQ_INSERT_BEFORE(prev, skiplist, ps_entry);
1000 }
1001 }
1002
1003
1004 /*
1005 * Remove a rule from the other skiplist calculations.
1006 */
1007 void
remove_from_skipsteps(struct skiplist * head,struct superblock * block,struct pf_opt_rule * por,struct pf_skip_step * active_list)1008 remove_from_skipsteps(struct skiplist *head, struct superblock *block,
1009 struct pf_opt_rule *por, struct pf_skip_step *active_list)
1010 {
1011 struct pf_skip_step *sk, *next;
1012 struct pf_opt_rule *p2;
1013 int i, found;
1014
1015 for (i = 0; i < PF_SKIP_COUNT; i++) {
1016 sk = TAILQ_FIRST(&block->sb_skipsteps[i]);
1017 if (sk == NULL || sk == active_list || sk->ps_count <= 1)
1018 continue;
1019 found = 0;
1020 do {
1021 TAILQ_FOREACH(p2, &sk->ps_rules, por_skip_entry[i])
1022 if (p2 == por) {
1023 TAILQ_REMOVE(&sk->ps_rules, p2,
1024 por_skip_entry[i]);
1025 found = 1;
1026 sk->ps_count--;
1027 break;
1028 }
1029 } while (!found && (sk = TAILQ_NEXT(sk, ps_entry)));
1030 if (found && sk) {
1031 /* Does this change the sorting order? */
1032 while ((next = TAILQ_NEXT(sk, ps_entry)) &&
1033 next->ps_count > sk->ps_count) {
1034 TAILQ_REMOVE(head, sk, ps_entry);
1035 TAILQ_INSERT_AFTER(head, next, sk, ps_entry);
1036 }
1037 #ifdef OPT_DEBUG
1038 next = TAILQ_NEXT(sk, ps_entry);
1039 assert(next == NULL || next->ps_count <= sk->ps_count);
1040 #endif /* OPT_DEBUG */
1041 }
1042 }
1043 }
1044
1045
1046 /* Compare two rules AF field for skiplist construction */
1047 int
skip_cmp_af(struct pfctl_rule * a,struct pfctl_rule * b)1048 skip_cmp_af(struct pfctl_rule *a, struct pfctl_rule *b)
1049 {
1050 if (a->af != b->af || a->af == 0)
1051 return (1);
1052 return (0);
1053 }
1054
1055 /* Compare two rules DIRECTION field for skiplist construction */
1056 int
skip_cmp_dir(struct pfctl_rule * a,struct pfctl_rule * b)1057 skip_cmp_dir(struct pfctl_rule *a, struct pfctl_rule *b)
1058 {
1059 if (a->direction == 0 || a->direction != b->direction)
1060 return (1);
1061 return (0);
1062 }
1063
1064 /* Compare two rules DST Address field for skiplist construction */
1065 int
skip_cmp_dst_addr(struct pfctl_rule * a,struct pfctl_rule * b)1066 skip_cmp_dst_addr(struct pfctl_rule *a, struct pfctl_rule *b)
1067 {
1068 if (a->dst.neg != b->dst.neg ||
1069 a->dst.addr.type != b->dst.addr.type)
1070 return (1);
1071 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1072 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1073 * a->proto == IPPROTO_ICMP
1074 * return (1);
1075 */
1076 switch (a->dst.addr.type) {
1077 case PF_ADDR_ADDRMASK:
1078 if (memcmp(&a->dst.addr.v.a.addr, &b->dst.addr.v.a.addr,
1079 sizeof(a->dst.addr.v.a.addr)) ||
1080 memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask,
1081 sizeof(a->dst.addr.v.a.mask)) ||
1082 (a->dst.addr.v.a.addr.addr32[0] == 0 &&
1083 a->dst.addr.v.a.addr.addr32[1] == 0 &&
1084 a->dst.addr.v.a.addr.addr32[2] == 0 &&
1085 a->dst.addr.v.a.addr.addr32[3] == 0))
1086 return (1);
1087 return (0);
1088 case PF_ADDR_DYNIFTL:
1089 if (strcmp(a->dst.addr.v.ifname, b->dst.addr.v.ifname) != 0 ||
1090 a->dst.addr.iflags != b->dst.addr.iflags ||
1091 memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask,
1092 sizeof(a->dst.addr.v.a.mask)))
1093 return (1);
1094 return (0);
1095 case PF_ADDR_NOROUTE:
1096 case PF_ADDR_URPFFAILED:
1097 return (0);
1098 case PF_ADDR_TABLE:
1099 return (strcmp(a->dst.addr.v.tblname, b->dst.addr.v.tblname));
1100 }
1101 return (1);
1102 }
1103
1104 /* Compare two rules DST port field for skiplist construction */
1105 int
skip_cmp_dst_port(struct pfctl_rule * a,struct pfctl_rule * b)1106 skip_cmp_dst_port(struct pfctl_rule *a, struct pfctl_rule *b)
1107 {
1108 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1109 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1110 * a->proto == IPPROTO_ICMP
1111 * return (1);
1112 */
1113 if (a->dst.port_op == PF_OP_NONE || a->dst.port_op != b->dst.port_op ||
1114 a->dst.port[0] != b->dst.port[0] ||
1115 a->dst.port[1] != b->dst.port[1])
1116 return (1);
1117 return (0);
1118 }
1119
1120 /* Compare two rules IFP field for skiplist construction */
1121 int
skip_cmp_ifp(struct pfctl_rule * a,struct pfctl_rule * b)1122 skip_cmp_ifp(struct pfctl_rule *a, struct pfctl_rule *b)
1123 {
1124 if (strcmp(a->ifname, b->ifname) || a->ifname[0] == '\0')
1125 return (1);
1126 return (a->ifnot != b->ifnot);
1127 }
1128
1129 /* Compare two rules PROTO field for skiplist construction */
1130 int
skip_cmp_proto(struct pfctl_rule * a,struct pfctl_rule * b)1131 skip_cmp_proto(struct pfctl_rule *a, struct pfctl_rule *b)
1132 {
1133 return (a->proto != b->proto || a->proto == 0);
1134 }
1135
1136 /* Compare two rules SRC addr field for skiplist construction */
1137 int
skip_cmp_src_addr(struct pfctl_rule * a,struct pfctl_rule * b)1138 skip_cmp_src_addr(struct pfctl_rule *a, struct pfctl_rule *b)
1139 {
1140 if (a->src.neg != b->src.neg ||
1141 a->src.addr.type != b->src.addr.type)
1142 return (1);
1143 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1144 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1145 * a->proto == IPPROTO_ICMP
1146 * return (1);
1147 */
1148 switch (a->src.addr.type) {
1149 case PF_ADDR_ADDRMASK:
1150 if (memcmp(&a->src.addr.v.a.addr, &b->src.addr.v.a.addr,
1151 sizeof(a->src.addr.v.a.addr)) ||
1152 memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask,
1153 sizeof(a->src.addr.v.a.mask)) ||
1154 (a->src.addr.v.a.addr.addr32[0] == 0 &&
1155 a->src.addr.v.a.addr.addr32[1] == 0 &&
1156 a->src.addr.v.a.addr.addr32[2] == 0 &&
1157 a->src.addr.v.a.addr.addr32[3] == 0))
1158 return (1);
1159 return (0);
1160 case PF_ADDR_DYNIFTL:
1161 if (strcmp(a->src.addr.v.ifname, b->src.addr.v.ifname) != 0 ||
1162 a->src.addr.iflags != b->src.addr.iflags ||
1163 memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask,
1164 sizeof(a->src.addr.v.a.mask)))
1165 return (1);
1166 return (0);
1167 case PF_ADDR_NOROUTE:
1168 case PF_ADDR_URPFFAILED:
1169 return (0);
1170 case PF_ADDR_TABLE:
1171 return (strcmp(a->src.addr.v.tblname, b->src.addr.v.tblname));
1172 }
1173 return (1);
1174 }
1175
1176 /* Compare two rules SRC port field for skiplist construction */
1177 int
skip_cmp_src_port(struct pfctl_rule * a,struct pfctl_rule * b)1178 skip_cmp_src_port(struct pfctl_rule *a, struct pfctl_rule *b)
1179 {
1180 if (a->src.port_op == PF_OP_NONE || a->src.port_op != b->src.port_op ||
1181 a->src.port[0] != b->src.port[0] ||
1182 a->src.port[1] != b->src.port[1])
1183 return (1);
1184 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1185 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1186 * a->proto == IPPROTO_ICMP
1187 * return (1);
1188 */
1189 return (0);
1190 }
1191
1192
1193 void
skip_init(void)1194 skip_init(void)
1195 {
1196 struct {
1197 char *name;
1198 int skipnum;
1199 int (*func)(struct pfctl_rule *, struct pfctl_rule *);
1200 } comps[] = PF_SKIP_COMPARITORS;
1201 int skipnum, i;
1202
1203 for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++) {
1204 for (i = 0; i < sizeof(comps)/sizeof(*comps); i++)
1205 if (comps[i].skipnum == skipnum) {
1206 skip_comparitors[skipnum] = comps[i].func;
1207 skip_comparitors_names[skipnum] = comps[i].name;
1208 }
1209 }
1210 for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++)
1211 if (skip_comparitors[skipnum] == NULL)
1212 errx(1, "Need to add skip step comparitor to pfctl?!");
1213 }
1214
1215 /*
1216 * Add a host/netmask to a table
1217 */
1218 int
add_opt_table(struct pfctl * pf,struct pf_opt_tbl ** tbl,sa_family_t af,struct pf_rule_addr * addr)1219 add_opt_table(struct pfctl *pf, struct pf_opt_tbl **tbl, sa_family_t af,
1220 struct pf_rule_addr *addr)
1221 {
1222 #ifdef OPT_DEBUG
1223 char buf[128];
1224 #endif /* OPT_DEBUG */
1225 static int tablenum = 0;
1226 struct node_host node_host;
1227
1228 if (*tbl == NULL) {
1229 if ((*tbl = calloc(1, sizeof(**tbl))) == NULL ||
1230 ((*tbl)->pt_buf = calloc(1, sizeof(*(*tbl)->pt_buf))) ==
1231 NULL)
1232 err(1, "calloc");
1233 (*tbl)->pt_buf->pfrb_type = PFRB_ADDRS;
1234 SIMPLEQ_INIT(&(*tbl)->pt_nodes);
1235
1236 /* This is just a temporary table name */
1237 snprintf((*tbl)->pt_name, sizeof((*tbl)->pt_name), "%s%d",
1238 PF_OPT_TABLE_PREFIX, tablenum++);
1239 DEBUG("creating table <%s>", (*tbl)->pt_name);
1240 }
1241
1242 memset(&node_host, 0, sizeof(node_host));
1243 node_host.af = af;
1244 node_host.addr = addr->addr;
1245
1246 #ifdef OPT_DEBUG
1247 DEBUG("<%s> adding %s/%d", (*tbl)->pt_name, inet_ntop(af,
1248 &node_host.addr.v.a.addr, buf, sizeof(buf)),
1249 unmask(&node_host.addr.v.a.mask, af));
1250 #endif /* OPT_DEBUG */
1251
1252 if (append_addr_host((*tbl)->pt_buf, &node_host, 0, 0)) {
1253 warn("failed to add host");
1254 return (1);
1255 }
1256 if (pf->opts & PF_OPT_VERBOSE) {
1257 struct node_tinit *ti;
1258
1259 if ((ti = calloc(1, sizeof(*ti))) == NULL)
1260 err(1, "malloc");
1261 if ((ti->host = malloc(sizeof(*ti->host))) == NULL)
1262 err(1, "malloc");
1263 memcpy(ti->host, &node_host, sizeof(*ti->host));
1264 SIMPLEQ_INSERT_TAIL(&(*tbl)->pt_nodes, ti, entries);
1265 }
1266
1267 (*tbl)->pt_rulecount++;
1268 if ((*tbl)->pt_rulecount == TABLE_THRESHOLD)
1269 DEBUG("table <%s> now faster than skip steps", (*tbl)->pt_name);
1270
1271 return (0);
1272 }
1273
1274
1275 /*
1276 * Do the dirty work of choosing an unused table name and creating it.
1277 * (be careful with the table name, it might already be used in another anchor)
1278 */
1279 int
pf_opt_create_table(struct pfctl * pf,struct pf_opt_tbl * tbl)1280 pf_opt_create_table(struct pfctl *pf, struct pf_opt_tbl *tbl)
1281 {
1282 static int tablenum;
1283 struct pfr_table *t;
1284
1285 if (table_buffer.pfrb_type == 0) {
1286 /* Initialize the list of tables */
1287 table_buffer.pfrb_type = PFRB_TABLES;
1288 for (;;) {
1289 pfr_buf_grow(&table_buffer, table_buffer.pfrb_size);
1290 table_buffer.pfrb_size = table_buffer.pfrb_msize;
1291 if (pfr_get_tables(NULL, table_buffer.pfrb_caddr,
1292 &table_buffer.pfrb_size, PFR_FLAG_ALLRSETS))
1293 err(1, "pfr_get_tables");
1294 if (table_buffer.pfrb_size <= table_buffer.pfrb_msize)
1295 break;
1296 }
1297 table_identifier = arc4random();
1298 }
1299
1300 /* XXX would be *really* nice to avoid duplicating identical tables */
1301
1302 /* Now we have to pick a table name that isn't used */
1303 again:
1304 DEBUG("translating temporary table <%s> to <%s%x_%d>", tbl->pt_name,
1305 PF_OPT_TABLE_PREFIX, table_identifier, tablenum);
1306 snprintf(tbl->pt_name, sizeof(tbl->pt_name), "%s%x_%d",
1307 PF_OPT_TABLE_PREFIX, table_identifier, tablenum);
1308 PFRB_FOREACH(t, &table_buffer) {
1309 if (strcasecmp(t->pfrt_name, tbl->pt_name) == 0) {
1310 /* Collision. Try again */
1311 DEBUG("wow, table <%s> in use. trying again",
1312 tbl->pt_name);
1313 table_identifier = arc4random();
1314 goto again;
1315 }
1316 }
1317 tablenum++;
1318
1319
1320 if (pfctl_define_table(tbl->pt_name, PFR_TFLAG_CONST, 1,
1321 pf->astack[0]->name, tbl->pt_buf, pf->astack[0]->ruleset.tticket)) {
1322 warn("failed to create table %s in %s",
1323 tbl->pt_name, pf->astack[0]->name);
1324 return (1);
1325 }
1326 return (0);
1327 }
1328
1329 /*
1330 * Partition the flat ruleset into a list of distinct superblocks
1331 */
1332 int
construct_superblocks(struct pfctl * pf,struct pf_opt_queue * opt_queue,struct superblocks * superblocks)1333 construct_superblocks(struct pfctl *pf, struct pf_opt_queue *opt_queue,
1334 struct superblocks *superblocks)
1335 {
1336 struct superblock *block = NULL;
1337 struct pf_opt_rule *por;
1338 int i;
1339
1340 while (!TAILQ_EMPTY(opt_queue)) {
1341 por = TAILQ_FIRST(opt_queue);
1342 TAILQ_REMOVE(opt_queue, por, por_entry);
1343 if (block == NULL || !superblock_inclusive(block, por)) {
1344 if ((block = calloc(1, sizeof(*block))) == NULL) {
1345 warn("calloc");
1346 return (1);
1347 }
1348 TAILQ_INIT(&block->sb_rules);
1349 for (i = 0; i < PF_SKIP_COUNT; i++)
1350 TAILQ_INIT(&block->sb_skipsteps[i]);
1351 TAILQ_INSERT_TAIL(superblocks, block, sb_entry);
1352 }
1353 TAILQ_INSERT_TAIL(&block->sb_rules, por, por_entry);
1354 }
1355
1356 return (0);
1357 }
1358
1359
1360 /*
1361 * Compare two rule addresses
1362 */
1363 int
addrs_equal(struct pf_rule_addr * a,struct pf_rule_addr * b)1364 addrs_equal(struct pf_rule_addr *a, struct pf_rule_addr *b)
1365 {
1366 if (a->neg != b->neg)
1367 return (0);
1368 return (memcmp(&a->addr, &b->addr, sizeof(a->addr)) == 0);
1369 }
1370
1371
1372 /*
1373 * The addresses are not equal, but can we combine them into one table?
1374 */
1375 int
addrs_combineable(struct pf_rule_addr * a,struct pf_rule_addr * b)1376 addrs_combineable(struct pf_rule_addr *a, struct pf_rule_addr *b)
1377 {
1378 if (a->addr.type != PF_ADDR_ADDRMASK ||
1379 b->addr.type != PF_ADDR_ADDRMASK)
1380 return (0);
1381 if (a->neg != b->neg || a->port_op != b->port_op ||
1382 a->port[0] != b->port[0] || a->port[1] != b->port[1])
1383 return (0);
1384 return (1);
1385 }
1386
1387
1388 /*
1389 * Are we allowed to combine these two rules
1390 */
1391 int
rules_combineable(struct pfctl_rule * p1,struct pfctl_rule * p2)1392 rules_combineable(struct pfctl_rule *p1, struct pfctl_rule *p2)
1393 {
1394 struct pfctl_rule a, b;
1395
1396 comparable_rule(&a, p1, COMBINED);
1397 comparable_rule(&b, p2, COMBINED);
1398 return (memcmp(&a, &b, sizeof(a)) == 0);
1399 }
1400
1401
1402 /*
1403 * Can a rule be included inside a superblock
1404 */
1405 int
superblock_inclusive(struct superblock * block,struct pf_opt_rule * por)1406 superblock_inclusive(struct superblock *block, struct pf_opt_rule *por)
1407 {
1408 struct pfctl_rule a, b;
1409 int i, j;
1410
1411 /* First check for hard breaks */
1412 for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++) {
1413 if (pf_rule_desc[i].prf_type == BARRIER) {
1414 for (j = 0; j < pf_rule_desc[i].prf_size; j++)
1415 if (((char *)&por->por_rule)[j +
1416 pf_rule_desc[i].prf_offset] != 0)
1417 return (0);
1418 }
1419 }
1420
1421 /* per-rule src-track is also a hard break */
1422 if (por->por_rule.rule_flag & PFRULE_RULESRCTRACK)
1423 return (0);
1424
1425 /*
1426 * Have to handle interface groups separately. Consider the following
1427 * rules:
1428 * block on EXTIFS to any port 22
1429 * pass on em0 to any port 22
1430 * (where EXTIFS is an arbitrary interface group)
1431 * The optimizer may decide to re-order the pass rule in front of the
1432 * block rule. But what if EXTIFS includes em0??? Such a reordering
1433 * would change the meaning of the ruleset.
1434 * We can't just lookup the EXTIFS group and check if em0 is a member
1435 * because the user is allowed to add interfaces to a group during
1436 * runtime.
1437 * Ergo interface groups become a defacto superblock break :-(
1438 */
1439 if (interface_group(por->por_rule.ifname) ||
1440 interface_group(TAILQ_FIRST(&block->sb_rules)->por_rule.ifname)) {
1441 if (strcasecmp(por->por_rule.ifname,
1442 TAILQ_FIRST(&block->sb_rules)->por_rule.ifname) != 0)
1443 return (0);
1444 }
1445
1446 comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule, NOMERGE);
1447 comparable_rule(&b, &por->por_rule, NOMERGE);
1448 if (memcmp(&a, &b, sizeof(a)) == 0)
1449 return (1);
1450
1451 #ifdef OPT_DEBUG
1452 for (i = 0; i < sizeof(por->por_rule); i++) {
1453 int closest = -1;
1454 if (((u_int8_t *)&a)[i] != ((u_int8_t *)&b)[i]) {
1455 for (j = 0; j < sizeof(pf_rule_desc) /
1456 sizeof(*pf_rule_desc); j++) {
1457 if (i >= pf_rule_desc[j].prf_offset &&
1458 i < pf_rule_desc[j].prf_offset +
1459 pf_rule_desc[j].prf_size) {
1460 DEBUG("superblock break @ %d due to %s",
1461 por->por_rule.nr,
1462 pf_rule_desc[j].prf_name);
1463 return (0);
1464 }
1465 if (i > pf_rule_desc[j].prf_offset) {
1466 if (closest == -1 ||
1467 i-pf_rule_desc[j].prf_offset <
1468 i-pf_rule_desc[closest].prf_offset)
1469 closest = j;
1470 }
1471 }
1472
1473 if (closest >= 0)
1474 DEBUG("superblock break @ %d on %s+%zxh",
1475 por->por_rule.nr,
1476 pf_rule_desc[closest].prf_name,
1477 i - pf_rule_desc[closest].prf_offset -
1478 pf_rule_desc[closest].prf_size);
1479 else
1480 DEBUG("superblock break @ %d on field @ %d",
1481 por->por_rule.nr, i);
1482 return (0);
1483 }
1484 }
1485 #endif /* OPT_DEBUG */
1486
1487 return (0);
1488 }
1489
1490
1491 /*
1492 * Figure out if an interface name is an actual interface or actually a
1493 * group of interfaces.
1494 */
1495 int
interface_group(const char * ifname)1496 interface_group(const char *ifname)
1497 {
1498 int s;
1499 struct ifgroupreq ifgr;
1500
1501 if (ifname == NULL || !ifname[0])
1502 return (0);
1503
1504 s = get_query_socket();
1505
1506 memset(&ifgr, 0, sizeof(ifgr));
1507 strlcpy(ifgr.ifgr_name, ifname, IFNAMSIZ);
1508 if (ioctl(s, SIOCGIFGMEMB, (caddr_t)&ifgr) == -1) {
1509 if (errno == ENOENT)
1510 return (0);
1511 else
1512 err(1, "SIOCGIFGMEMB");
1513 }
1514
1515 return (1);
1516 }
1517
1518
1519 /*
1520 * Make a rule that can directly compared by memcmp()
1521 */
1522 void
comparable_rule(struct pfctl_rule * dst,const struct pfctl_rule * src,int type)1523 comparable_rule(struct pfctl_rule *dst, const struct pfctl_rule *src, int type)
1524 {
1525 int i;
1526 /*
1527 * To simplify the comparison, we just zero out the fields that are
1528 * allowed to be different and then do a simple memcmp()
1529 */
1530 memcpy(dst, src, sizeof(*dst));
1531 for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++)
1532 if (pf_rule_desc[i].prf_type >= type) {
1533 #ifdef OPT_DEBUG
1534 assert(pf_rule_desc[i].prf_type != NEVER ||
1535 *(((char *)dst) + pf_rule_desc[i].prf_offset) == 0);
1536 #endif /* OPT_DEBUG */
1537 memset(((char *)dst) + pf_rule_desc[i].prf_offset, 0,
1538 pf_rule_desc[i].prf_size);
1539 }
1540 }
1541
1542
1543 /*
1544 * Remove superset information from two rules so we can directly compare them
1545 * with memcmp()
1546 */
1547 void
exclude_supersets(struct pfctl_rule * super,struct pfctl_rule * sub)1548 exclude_supersets(struct pfctl_rule *super, struct pfctl_rule *sub)
1549 {
1550 if (super->ifname[0] == '\0')
1551 memset(sub->ifname, 0, sizeof(sub->ifname));
1552 if (super->direction == PF_INOUT)
1553 sub->direction = PF_INOUT;
1554 if ((super->proto == 0 || super->proto == sub->proto) &&
1555 super->flags == 0 && super->flagset == 0 && (sub->flags ||
1556 sub->flagset)) {
1557 sub->flags = super->flags;
1558 sub->flagset = super->flagset;
1559 }
1560 if (super->proto == 0)
1561 sub->proto = 0;
1562
1563 if (super->src.port_op == 0) {
1564 sub->src.port_op = 0;
1565 sub->src.port[0] = 0;
1566 sub->src.port[1] = 0;
1567 }
1568 if (super->dst.port_op == 0) {
1569 sub->dst.port_op = 0;
1570 sub->dst.port[0] = 0;
1571 sub->dst.port[1] = 0;
1572 }
1573
1574 if (super->src.addr.type == PF_ADDR_ADDRMASK && !super->src.neg &&
1575 !sub->src.neg && super->src.addr.v.a.mask.addr32[0] == 0 &&
1576 super->src.addr.v.a.mask.addr32[1] == 0 &&
1577 super->src.addr.v.a.mask.addr32[2] == 0 &&
1578 super->src.addr.v.a.mask.addr32[3] == 0)
1579 memset(&sub->src.addr, 0, sizeof(sub->src.addr));
1580 else if (super->src.addr.type == PF_ADDR_ADDRMASK &&
1581 sub->src.addr.type == PF_ADDR_ADDRMASK &&
1582 super->src.neg == sub->src.neg &&
1583 super->af == sub->af &&
1584 unmask(&super->src.addr.v.a.mask, super->af) <
1585 unmask(&sub->src.addr.v.a.mask, sub->af) &&
1586 super->src.addr.v.a.addr.addr32[0] ==
1587 (sub->src.addr.v.a.addr.addr32[0] &
1588 super->src.addr.v.a.mask.addr32[0]) &&
1589 super->src.addr.v.a.addr.addr32[1] ==
1590 (sub->src.addr.v.a.addr.addr32[1] &
1591 super->src.addr.v.a.mask.addr32[1]) &&
1592 super->src.addr.v.a.addr.addr32[2] ==
1593 (sub->src.addr.v.a.addr.addr32[2] &
1594 super->src.addr.v.a.mask.addr32[2]) &&
1595 super->src.addr.v.a.addr.addr32[3] ==
1596 (sub->src.addr.v.a.addr.addr32[3] &
1597 super->src.addr.v.a.mask.addr32[3])) {
1598 /* sub->src.addr is a subset of super->src.addr/mask */
1599 memcpy(&sub->src.addr, &super->src.addr, sizeof(sub->src.addr));
1600 }
1601
1602 if (super->dst.addr.type == PF_ADDR_ADDRMASK && !super->dst.neg &&
1603 !sub->dst.neg && super->dst.addr.v.a.mask.addr32[0] == 0 &&
1604 super->dst.addr.v.a.mask.addr32[1] == 0 &&
1605 super->dst.addr.v.a.mask.addr32[2] == 0 &&
1606 super->dst.addr.v.a.mask.addr32[3] == 0)
1607 memset(&sub->dst.addr, 0, sizeof(sub->dst.addr));
1608 else if (super->dst.addr.type == PF_ADDR_ADDRMASK &&
1609 sub->dst.addr.type == PF_ADDR_ADDRMASK &&
1610 super->dst.neg == sub->dst.neg &&
1611 super->af == sub->af &&
1612 unmask(&super->dst.addr.v.a.mask, super->af) <
1613 unmask(&sub->dst.addr.v.a.mask, sub->af) &&
1614 super->dst.addr.v.a.addr.addr32[0] ==
1615 (sub->dst.addr.v.a.addr.addr32[0] &
1616 super->dst.addr.v.a.mask.addr32[0]) &&
1617 super->dst.addr.v.a.addr.addr32[1] ==
1618 (sub->dst.addr.v.a.addr.addr32[1] &
1619 super->dst.addr.v.a.mask.addr32[1]) &&
1620 super->dst.addr.v.a.addr.addr32[2] ==
1621 (sub->dst.addr.v.a.addr.addr32[2] &
1622 super->dst.addr.v.a.mask.addr32[2]) &&
1623 super->dst.addr.v.a.addr.addr32[3] ==
1624 (sub->dst.addr.v.a.addr.addr32[3] &
1625 super->dst.addr.v.a.mask.addr32[3])) {
1626 /* sub->dst.addr is a subset of super->dst.addr/mask */
1627 memcpy(&sub->dst.addr, &super->dst.addr, sizeof(sub->dst.addr));
1628 }
1629
1630 if (super->af == 0)
1631 sub->af = 0;
1632 }
1633
1634
1635 void
superblock_free(struct pfctl * pf,struct superblock * block)1636 superblock_free(struct pfctl *pf, struct superblock *block)
1637 {
1638 struct pf_opt_rule *por;
1639 while ((por = TAILQ_FIRST(&block->sb_rules))) {
1640 TAILQ_REMOVE(&block->sb_rules, por, por_entry);
1641 if (por->por_src_tbl) {
1642 if (por->por_src_tbl->pt_buf) {
1643 pfr_buf_clear(por->por_src_tbl->pt_buf);
1644 free(por->por_src_tbl->pt_buf);
1645 }
1646 free(por->por_src_tbl);
1647 }
1648 if (por->por_dst_tbl) {
1649 if (por->por_dst_tbl->pt_buf) {
1650 pfr_buf_clear(por->por_dst_tbl->pt_buf);
1651 free(por->por_dst_tbl->pt_buf);
1652 }
1653 free(por->por_dst_tbl);
1654 }
1655 free(por);
1656 }
1657 if (block->sb_profiled_block)
1658 superblock_free(pf, block->sb_profiled_block);
1659 free(block);
1660 }
1661
1662