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(rdr, BREAK),
139 PF_RULE_FIELD(nat, BREAK),
140 PF_RULE_FIELD(route, BREAK),
141 PF_RULE_FIELD(logif, BREAK),
142
143 /*
144 * Any fields not listed in this structure act as BREAK fields
145 */
146
147
148 /*
149 * These fields must not differ when we merge two rules together but
150 * their difference isn't enough to put the rules in different superblocks.
151 * There are no problems re-ordering any rules with these fields.
152 */
153 PF_RULE_FIELD(af, NOMERGE),
154 PF_RULE_FIELD(ifnot, NOMERGE),
155 PF_RULE_FIELD(ifname, NOMERGE), /* hack for IF groups */
156 PF_RULE_FIELD(match_tag_not, NOMERGE),
157 PF_RULE_FIELD(match_tagname, NOMERGE),
158 PF_RULE_FIELD(os_fingerprint, NOMERGE),
159 PF_RULE_FIELD(timeout, NOMERGE),
160 PF_RULE_FIELD(return_icmp, NOMERGE),
161 PF_RULE_FIELD(return_icmp6, NOMERGE),
162 PF_RULE_FIELD(uid, NOMERGE),
163 PF_RULE_FIELD(gid, NOMERGE),
164 PF_RULE_FIELD(direction, NOMERGE),
165 PF_RULE_FIELD(proto, NOMERGE),
166 PF_RULE_FIELD(type, NOMERGE),
167 PF_RULE_FIELD(code, NOMERGE),
168 PF_RULE_FIELD(flags, NOMERGE),
169 PF_RULE_FIELD(flagset, NOMERGE),
170 PF_RULE_FIELD(tos, NOMERGE),
171 PF_RULE_FIELD(src.port, NOMERGE),
172 PF_RULE_FIELD(dst.port, NOMERGE),
173 PF_RULE_FIELD(src.port_op, NOMERGE),
174 PF_RULE_FIELD(dst.port_op, NOMERGE),
175 PF_RULE_FIELD(src.neg, NOMERGE),
176 PF_RULE_FIELD(dst.neg, NOMERGE),
177 PF_RULE_FIELD(af, NOMERGE),
178
179 /* These fields can be merged */
180 PF_RULE_FIELD(src.addr, COMBINED),
181 PF_RULE_FIELD(dst.addr, COMBINED),
182
183 /* We just don't care about these fields. They're set by the kernel */
184 PF_RULE_FIELD(skip, DC),
185 PF_RULE_FIELD(evaluations, DC),
186 PF_RULE_FIELD(packets, DC),
187 PF_RULE_FIELD(bytes, DC),
188 PF_RULE_FIELD(kif, DC),
189 PF_RULE_FIELD(states_cur, DC),
190 PF_RULE_FIELD(states_tot, DC),
191 PF_RULE_FIELD(src_nodes, DC),
192 PF_RULE_FIELD(nr, DC),
193 PF_RULE_FIELD(entries, DC),
194 PF_RULE_FIELD(qid, DC),
195 PF_RULE_FIELD(pqid, DC),
196 PF_RULE_FIELD(anchor_relative, DC),
197 PF_RULE_FIELD(anchor_wildcard, DC),
198 PF_RULE_FIELD(tag, DC),
199 PF_RULE_FIELD(match_tag, DC),
200 PF_RULE_FIELD(overload_tbl, DC),
201
202 /* These fields should never be set in a PASS/BLOCK rule */
203 PF_RULE_FIELD(natpass, NEVER),
204 PF_RULE_FIELD(max_mss, NEVER),
205 PF_RULE_FIELD(min_ttl, NEVER),
206 PF_RULE_FIELD(set_tos, NEVER),
207 };
208
209
210
211 int add_opt_table(struct pfctl *, struct pf_opt_tbl **, sa_family_t,
212 struct pf_rule_addr *);
213 int addrs_combineable(struct pf_rule_addr *, struct pf_rule_addr *);
214 int addrs_equal(struct pf_rule_addr *, struct pf_rule_addr *);
215 int block_feedback(struct pfctl *, struct superblock *);
216 int combine_rules(struct pfctl *, struct superblock *);
217 void comparable_rule(struct pfctl_rule *, const struct pfctl_rule *, int);
218 int construct_superblocks(struct pfctl *, struct pf_opt_queue *,
219 struct superblocks *);
220 void exclude_supersets(struct pfctl_rule *, struct pfctl_rule *);
221 int interface_group(const char *);
222 int load_feedback_profile(struct pfctl *, struct superblocks *);
223 int optimize_superblock(struct pfctl *, struct superblock *);
224 int pf_opt_create_table(struct pfctl *, struct pf_opt_tbl *);
225 void remove_from_skipsteps(struct skiplist *, struct superblock *,
226 struct pf_opt_rule *, struct pf_skip_step *);
227 int remove_identical_rules(struct pfctl *, struct superblock *);
228 int reorder_rules(struct pfctl *, struct superblock *, int);
229 int rules_combineable(struct pfctl_rule *, struct pfctl_rule *);
230 void skip_append(struct superblock *, int, struct pf_skip_step *,
231 struct pf_opt_rule *);
232 int skip_compare(int, struct pf_skip_step *, struct pf_opt_rule *);
233 void skip_init(void);
234 int skip_cmp_af(struct pfctl_rule *, struct pfctl_rule *);
235 int skip_cmp_dir(struct pfctl_rule *, struct pfctl_rule *);
236 int skip_cmp_dst_addr(struct pfctl_rule *, struct pfctl_rule *);
237 int skip_cmp_dst_port(struct pfctl_rule *, struct pfctl_rule *);
238 int skip_cmp_ifp(struct pfctl_rule *, struct pfctl_rule *);
239 int skip_cmp_proto(struct pfctl_rule *, struct pfctl_rule *);
240 int skip_cmp_src_addr(struct pfctl_rule *, struct pfctl_rule *);
241 int skip_cmp_src_port(struct pfctl_rule *, struct pfctl_rule *);
242 int superblock_inclusive(struct superblock *, struct pf_opt_rule *);
243 void superblock_free(struct pfctl *, struct superblock *);
244
245
246 static int (*skip_comparitors[PF_SKIP_COUNT])(struct pfctl_rule *,
247 struct pfctl_rule *);
248 static const char *skip_comparitors_names[PF_SKIP_COUNT];
249 #define PF_SKIP_COMPARITORS { \
250 { "ifp", PF_SKIP_IFP, skip_cmp_ifp }, \
251 { "dir", PF_SKIP_DIR, skip_cmp_dir }, \
252 { "af", PF_SKIP_AF, skip_cmp_af }, \
253 { "proto", PF_SKIP_PROTO, skip_cmp_proto }, \
254 { "saddr", PF_SKIP_SRC_ADDR, skip_cmp_src_addr }, \
255 { "daddr", PF_SKIP_DST_ADDR, skip_cmp_dst_addr }, \
256 { "sport", PF_SKIP_SRC_PORT, skip_cmp_src_port }, \
257 { "dport", PF_SKIP_DST_PORT, skip_cmp_dst_port } \
258 }
259
260 static struct pfr_buffer table_buffer;
261 static int table_identifier;
262
263
264 int
pfctl_optimize_ruleset(struct pfctl * pf,struct pfctl_ruleset * rs)265 pfctl_optimize_ruleset(struct pfctl *pf, struct pfctl_ruleset *rs)
266 {
267 struct superblocks superblocks;
268 struct pf_opt_queue opt_queue;
269 struct superblock *block;
270 struct pf_opt_rule *por;
271 struct pfctl_rule *r;
272 struct pfctl_rulequeue *old_rules;
273
274 DEBUG("optimizing ruleset");
275 memset(&table_buffer, 0, sizeof(table_buffer));
276 skip_init();
277 TAILQ_INIT(&opt_queue);
278
279 old_rules = rs->rules[PF_RULESET_FILTER].active.ptr;
280 rs->rules[PF_RULESET_FILTER].active.ptr =
281 rs->rules[PF_RULESET_FILTER].inactive.ptr;
282 rs->rules[PF_RULESET_FILTER].inactive.ptr = old_rules;
283
284 /*
285 * XXX expanding the pf_opt_rule format throughout pfctl might allow
286 * us to avoid all this copying.
287 */
288 while ((r = TAILQ_FIRST(rs->rules[PF_RULESET_FILTER].inactive.ptr))
289 != NULL) {
290 TAILQ_REMOVE(rs->rules[PF_RULESET_FILTER].inactive.ptr, r,
291 entries);
292 if ((por = calloc(1, sizeof(*por))) == NULL)
293 err(1, "calloc");
294 memcpy(&por->por_rule, r, sizeof(*r));
295 if (TAILQ_FIRST(&r->rdr.list) != NULL) {
296 TAILQ_INIT(&por->por_rule.rdr.list);
297 pfctl_move_pool(&r->rdr, &por->por_rule.rdr);
298 } else
299 bzero(&por->por_rule.rdr,
300 sizeof(por->por_rule.rdr));
301 if (TAILQ_FIRST(&r->nat.list) != NULL) {
302 TAILQ_INIT(&por->por_rule.nat.list);
303 pfctl_move_pool(&r->nat, &por->por_rule.nat);
304 } else
305 bzero(&por->por_rule.nat,
306 sizeof(por->por_rule.nat));
307 if (TAILQ_FIRST(&r->route.list) != NULL) {
308 TAILQ_INIT(&por->por_rule.route.list);
309 pfctl_move_pool(&r->route, &por->por_rule.route);
310 } else
311 bzero(&por->por_rule.route,
312 sizeof(por->por_rule.route));
313
314 TAILQ_INSERT_TAIL(&opt_queue, por, por_entry);
315 }
316
317 TAILQ_INIT(&superblocks);
318 if (construct_superblocks(pf, &opt_queue, &superblocks))
319 goto error;
320
321 if (pf->optimize & PF_OPTIMIZE_PROFILE) {
322 if (load_feedback_profile(pf, &superblocks))
323 goto error;
324 }
325
326 TAILQ_FOREACH(block, &superblocks, sb_entry) {
327 if (optimize_superblock(pf, block))
328 goto error;
329 }
330
331 rs->anchor->refcnt = 0;
332 while ((block = TAILQ_FIRST(&superblocks))) {
333 TAILQ_REMOVE(&superblocks, block, sb_entry);
334
335 while ((por = TAILQ_FIRST(&block->sb_rules))) {
336 TAILQ_REMOVE(&block->sb_rules, por, por_entry);
337 por->por_rule.nr = rs->anchor->refcnt++;
338 if ((r = calloc(1, sizeof(*r))) == NULL)
339 err(1, "calloc");
340 memcpy(r, &por->por_rule, sizeof(*r));
341 TAILQ_INIT(&r->rdr.list);
342 pfctl_move_pool(&por->por_rule.rdr, &r->rdr);
343 TAILQ_INIT(&r->nat.list);
344 pfctl_move_pool(&por->por_rule.nat, &r->nat);
345 TAILQ_INSERT_TAIL(
346 rs->rules[PF_RULESET_FILTER].active.ptr,
347 r, entries);
348 free(por);
349 }
350 free(block);
351 }
352
353 return (0);
354
355 error:
356 while ((por = TAILQ_FIRST(&opt_queue))) {
357 TAILQ_REMOVE(&opt_queue, por, por_entry);
358 if (por->por_src_tbl) {
359 pfr_buf_clear(por->por_src_tbl->pt_buf);
360 free(por->por_src_tbl->pt_buf);
361 free(por->por_src_tbl);
362 }
363 if (por->por_dst_tbl) {
364 pfr_buf_clear(por->por_dst_tbl->pt_buf);
365 free(por->por_dst_tbl->pt_buf);
366 free(por->por_dst_tbl);
367 }
368 free(por);
369 }
370 while ((block = TAILQ_FIRST(&superblocks))) {
371 TAILQ_REMOVE(&superblocks, block, sb_entry);
372 superblock_free(pf, block);
373 }
374 return (1);
375 }
376
377
378 /*
379 * Go ahead and optimize a superblock
380 */
381 int
optimize_superblock(struct pfctl * pf,struct superblock * block)382 optimize_superblock(struct pfctl *pf, struct superblock *block)
383 {
384 #ifdef OPT_DEBUG
385 struct pf_opt_rule *por;
386 #endif /* OPT_DEBUG */
387
388 /* We have a few optimization passes:
389 * 1) remove duplicate rules or rules that are a subset of other
390 * rules
391 * 2) combine otherwise identical rules with different IP addresses
392 * into a single rule and put the addresses in a table.
393 * 3) re-order the rules to improve kernel skip steps
394 * 4) re-order the 'quick' rules based on feedback from the
395 * active ruleset statistics
396 *
397 * XXX combine_rules() doesn't combine v4 and v6 rules. would just
398 * have to keep af in the table container, make af 'COMBINE' and
399 * twiddle the af on the merged rule
400 * XXX maybe add a weighting to the metric on skipsteps when doing
401 * reordering. sometimes two sequential tables will be better
402 * that four consecutive interfaces.
403 * XXX need to adjust the skipstep count of everything after PROTO,
404 * since they aren't actually checked on a proto mismatch in
405 * pf_test_{tcp, udp, icmp}()
406 * XXX should i treat proto=0, af=0 or dir=0 special in skepstep
407 * calculation since they are a DC?
408 * XXX keep last skiplist of last superblock to influence this
409 * superblock. '5 inet6 log' should make '3 inet6' come before '4
410 * inet' in the next superblock.
411 * XXX would be useful to add tables for ports
412 * XXX we can also re-order some mutually exclusive superblocks to
413 * try merging superblocks before any of these optimization passes.
414 * for instance a single 'log in' rule in the middle of non-logging
415 * out rules.
416 */
417
418 /* shortcut. there will be a lot of 1-rule superblocks */
419 if (!TAILQ_NEXT(TAILQ_FIRST(&block->sb_rules), por_entry))
420 return (0);
421
422 #ifdef OPT_DEBUG
423 printf("--- Superblock ---\n");
424 TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
425 printf(" ");
426 print_rule(&por->por_rule, por->por_rule.anchor ?
427 por->por_rule.anchor->name : "", 1, 0);
428 }
429 #endif /* OPT_DEBUG */
430
431
432 if (remove_identical_rules(pf, block))
433 return (1);
434 if (combine_rules(pf, block))
435 return (1);
436 if ((pf->optimize & PF_OPTIMIZE_PROFILE) &&
437 TAILQ_FIRST(&block->sb_rules)->por_rule.quick &&
438 block->sb_profiled_block) {
439 if (block_feedback(pf, block))
440 return (1);
441 } else if (reorder_rules(pf, block, 0)) {
442 return (1);
443 }
444
445 /*
446 * Don't add any optimization passes below reorder_rules(). It will
447 * have divided superblocks into smaller blocks for further refinement
448 * and doesn't put them back together again. What once was a true
449 * superblock might have been split into multiple superblocks.
450 */
451
452 #ifdef OPT_DEBUG
453 printf("--- END Superblock ---\n");
454 #endif /* OPT_DEBUG */
455 return (0);
456 }
457
458
459 /*
460 * Optimization pass #1: remove identical rules
461 */
462 int
remove_identical_rules(struct pfctl * pf,struct superblock * block)463 remove_identical_rules(struct pfctl *pf, struct superblock *block)
464 {
465 struct pf_opt_rule *por1, *por2, *por_next, *por2_next;
466 struct pfctl_rule a, a2, b, b2;
467
468 for (por1 = TAILQ_FIRST(&block->sb_rules); por1; por1 = por_next) {
469 por_next = TAILQ_NEXT(por1, por_entry);
470 for (por2 = por_next; por2; por2 = por2_next) {
471 por2_next = TAILQ_NEXT(por2, por_entry);
472 comparable_rule(&a, &por1->por_rule, DC);
473 comparable_rule(&b, &por2->por_rule, DC);
474 memcpy(&a2, &a, sizeof(a2));
475 memcpy(&b2, &b, sizeof(b2));
476
477 exclude_supersets(&a, &b);
478 exclude_supersets(&b2, &a2);
479 if (memcmp(&a, &b, sizeof(a)) == 0) {
480 DEBUG("removing identical rule nr%d = *nr%d*",
481 por1->por_rule.nr, por2->por_rule.nr);
482 TAILQ_REMOVE(&block->sb_rules, por2, por_entry);
483 if (por_next == por2)
484 por_next = TAILQ_NEXT(por1, por_entry);
485 free(por2);
486 } else if (memcmp(&a2, &b2, sizeof(a2)) == 0) {
487 DEBUG("removing identical rule *nr%d* = nr%d",
488 por1->por_rule.nr, por2->por_rule.nr);
489 TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
490 free(por1);
491 break;
492 }
493 }
494 }
495
496 return (0);
497 }
498
499
500 /*
501 * Optimization pass #2: combine similar rules with different addresses
502 * into a single rule and a table
503 */
504 int
combine_rules(struct pfctl * pf,struct superblock * block)505 combine_rules(struct pfctl *pf, struct superblock *block)
506 {
507 struct pf_opt_rule *p1, *p2, *por_next;
508 int src_eq, dst_eq;
509
510 if ((pf->loadopt & PFCTL_FLAG_TABLE) == 0) {
511 warnx("Must enable table loading for optimizations");
512 return (1);
513 }
514
515 /* First we make a pass to combine the rules. O(n log n) */
516 TAILQ_FOREACH(p1, &block->sb_rules, por_entry) {
517 for (p2 = TAILQ_NEXT(p1, por_entry); p2; p2 = por_next) {
518 por_next = TAILQ_NEXT(p2, por_entry);
519
520 src_eq = addrs_equal(&p1->por_rule.src,
521 &p2->por_rule.src);
522 dst_eq = addrs_equal(&p1->por_rule.dst,
523 &p2->por_rule.dst);
524
525 if (src_eq && !dst_eq && p1->por_src_tbl == NULL &&
526 p2->por_dst_tbl == NULL &&
527 p2->por_src_tbl == NULL &&
528 rules_combineable(&p1->por_rule, &p2->por_rule) &&
529 addrs_combineable(&p1->por_rule.dst,
530 &p2->por_rule.dst)) {
531 DEBUG("can combine rules nr%d = nr%d",
532 p1->por_rule.nr, p2->por_rule.nr);
533 if (p1->por_dst_tbl == NULL &&
534 add_opt_table(pf, &p1->por_dst_tbl,
535 p1->por_rule.af, &p1->por_rule.dst))
536 return (1);
537 if (add_opt_table(pf, &p1->por_dst_tbl,
538 p1->por_rule.af, &p2->por_rule.dst))
539 return (1);
540 p2->por_dst_tbl = p1->por_dst_tbl;
541 if (p1->por_dst_tbl->pt_rulecount >=
542 TABLE_THRESHOLD) {
543 TAILQ_REMOVE(&block->sb_rules, p2,
544 por_entry);
545 free(p2);
546 }
547 } else if (!src_eq && dst_eq && p1->por_dst_tbl == NULL
548 && p2->por_src_tbl == NULL &&
549 p2->por_dst_tbl == NULL &&
550 rules_combineable(&p1->por_rule, &p2->por_rule) &&
551 addrs_combineable(&p1->por_rule.src,
552 &p2->por_rule.src)) {
553 DEBUG("can combine rules nr%d = nr%d",
554 p1->por_rule.nr, p2->por_rule.nr);
555 if (p1->por_src_tbl == NULL &&
556 add_opt_table(pf, &p1->por_src_tbl,
557 p1->por_rule.af, &p1->por_rule.src))
558 return (1);
559 if (add_opt_table(pf, &p1->por_src_tbl,
560 p1->por_rule.af, &p2->por_rule.src))
561 return (1);
562 p2->por_src_tbl = p1->por_src_tbl;
563 if (p1->por_src_tbl->pt_rulecount >=
564 TABLE_THRESHOLD) {
565 TAILQ_REMOVE(&block->sb_rules, p2,
566 por_entry);
567 free(p2);
568 }
569 }
570 }
571 }
572
573
574 /*
575 * Then we make a final pass to create a valid table name and
576 * insert the name into the rules.
577 */
578 for (p1 = TAILQ_FIRST(&block->sb_rules); p1; p1 = por_next) {
579 por_next = TAILQ_NEXT(p1, por_entry);
580 assert(p1->por_src_tbl == NULL || p1->por_dst_tbl == NULL);
581
582 if (p1->por_src_tbl && p1->por_src_tbl->pt_rulecount >=
583 TABLE_THRESHOLD) {
584 if (p1->por_src_tbl->pt_generated) {
585 /* This rule is included in a table */
586 TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
587 free(p1);
588 continue;
589 }
590 p1->por_src_tbl->pt_generated = 1;
591
592 if ((pf->opts & PF_OPT_NOACTION) == 0 &&
593 pf_opt_create_table(pf, p1->por_src_tbl))
594 return (1);
595
596 pf->tdirty = 1;
597
598 if (pf->opts & PF_OPT_VERBOSE)
599 print_tabledef(p1->por_src_tbl->pt_name,
600 PFR_TFLAG_CONST, 1,
601 &p1->por_src_tbl->pt_nodes);
602
603 memset(&p1->por_rule.src.addr, 0,
604 sizeof(p1->por_rule.src.addr));
605 p1->por_rule.src.addr.type = PF_ADDR_TABLE;
606 strlcpy(p1->por_rule.src.addr.v.tblname,
607 p1->por_src_tbl->pt_name,
608 sizeof(p1->por_rule.src.addr.v.tblname));
609
610 pfr_buf_clear(p1->por_src_tbl->pt_buf);
611 free(p1->por_src_tbl->pt_buf);
612 p1->por_src_tbl->pt_buf = NULL;
613 }
614 if (p1->por_dst_tbl && p1->por_dst_tbl->pt_rulecount >=
615 TABLE_THRESHOLD) {
616 if (p1->por_dst_tbl->pt_generated) {
617 /* This rule is included in a table */
618 TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
619 free(p1);
620 continue;
621 }
622 p1->por_dst_tbl->pt_generated = 1;
623
624 if ((pf->opts & PF_OPT_NOACTION) == 0 &&
625 pf_opt_create_table(pf, p1->por_dst_tbl))
626 return (1);
627 pf->tdirty = 1;
628
629 if (pf->opts & PF_OPT_VERBOSE)
630 print_tabledef(p1->por_dst_tbl->pt_name,
631 PFR_TFLAG_CONST, 1,
632 &p1->por_dst_tbl->pt_nodes);
633
634 memset(&p1->por_rule.dst.addr, 0,
635 sizeof(p1->por_rule.dst.addr));
636 p1->por_rule.dst.addr.type = PF_ADDR_TABLE;
637 strlcpy(p1->por_rule.dst.addr.v.tblname,
638 p1->por_dst_tbl->pt_name,
639 sizeof(p1->por_rule.dst.addr.v.tblname));
640
641 pfr_buf_clear(p1->por_dst_tbl->pt_buf);
642 free(p1->por_dst_tbl->pt_buf);
643 p1->por_dst_tbl->pt_buf = NULL;
644 }
645 }
646
647 return (0);
648 }
649
650
651 /*
652 * Optimization pass #3: re-order rules to improve skip steps
653 */
654 int
reorder_rules(struct pfctl * pf,struct superblock * block,int depth)655 reorder_rules(struct pfctl *pf, struct superblock *block, int depth)
656 {
657 struct superblock *newblock;
658 struct pf_skip_step *skiplist;
659 struct pf_opt_rule *por;
660 int i, largest, largest_list, rule_count = 0;
661 TAILQ_HEAD( , pf_opt_rule) head;
662
663 /*
664 * Calculate the best-case skip steps. We put each rule in a list
665 * of other rules with common fields
666 */
667 for (i = 0; i < PF_SKIP_COUNT; i++) {
668 TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
669 TAILQ_FOREACH(skiplist, &block->sb_skipsteps[i],
670 ps_entry) {
671 if (skip_compare(i, skiplist, por) == 0)
672 break;
673 }
674 if (skiplist == NULL) {
675 if ((skiplist = calloc(1, sizeof(*skiplist))) ==
676 NULL)
677 err(1, "calloc");
678 TAILQ_INIT(&skiplist->ps_rules);
679 TAILQ_INSERT_TAIL(&block->sb_skipsteps[i],
680 skiplist, ps_entry);
681 }
682 skip_append(block, i, skiplist, por);
683 }
684 }
685
686 TAILQ_FOREACH(por, &block->sb_rules, por_entry)
687 rule_count++;
688
689 /*
690 * Now we're going to ignore any fields that are identical between
691 * all of the rules in the superblock and those fields which differ
692 * between every rule in the superblock.
693 */
694 largest = 0;
695 for (i = 0; i < PF_SKIP_COUNT; i++) {
696 skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
697 if (skiplist->ps_count == rule_count) {
698 DEBUG("(%d) original skipstep '%s' is all rules",
699 depth, skip_comparitors_names[i]);
700 skiplist->ps_count = 0;
701 } else if (skiplist->ps_count == 1) {
702 skiplist->ps_count = 0;
703 } else {
704 DEBUG("(%d) original skipstep '%s' largest jump is %d",
705 depth, skip_comparitors_names[i],
706 skiplist->ps_count);
707 if (skiplist->ps_count > largest)
708 largest = skiplist->ps_count;
709 }
710 }
711 if (largest == 0) {
712 /* Ugh. There is NO commonality in the superblock on which
713 * optimize the skipsteps optimization.
714 */
715 goto done;
716 }
717
718 /*
719 * Now we're going to empty the superblock rule list and re-create
720 * it based on a more optimal skipstep order.
721 */
722 TAILQ_INIT(&head);
723 while ((por = TAILQ_FIRST(&block->sb_rules))) {
724 TAILQ_REMOVE(&block->sb_rules, por, por_entry);
725 TAILQ_INSERT_TAIL(&head, por, por_entry);
726 }
727
728
729 while (!TAILQ_EMPTY(&head)) {
730 largest = 1;
731
732 /*
733 * Find the most useful skip steps remaining
734 */
735 for (i = 0; i < PF_SKIP_COUNT; i++) {
736 skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
737 if (skiplist->ps_count > largest) {
738 largest = skiplist->ps_count;
739 largest_list = i;
740 }
741 }
742
743 if (largest <= 1) {
744 /*
745 * Nothing useful left. Leave remaining rules in order.
746 */
747 DEBUG("(%d) no more commonality for skip steps", depth);
748 while ((por = TAILQ_FIRST(&head))) {
749 TAILQ_REMOVE(&head, por, por_entry);
750 TAILQ_INSERT_TAIL(&block->sb_rules, por,
751 por_entry);
752 }
753 } else {
754 /*
755 * There is commonality. Extract those common rules
756 * and place them in the ruleset adjacent to each
757 * other.
758 */
759 skiplist = TAILQ_FIRST(&block->sb_skipsteps[
760 largest_list]);
761 DEBUG("(%d) skipstep '%s' largest jump is %d @ #%d",
762 depth, skip_comparitors_names[largest_list],
763 largest, TAILQ_FIRST(&TAILQ_FIRST(&block->
764 sb_skipsteps [largest_list])->ps_rules)->
765 por_rule.nr);
766 TAILQ_REMOVE(&block->sb_skipsteps[largest_list],
767 skiplist, ps_entry);
768
769
770 /*
771 * There may be further commonality inside these
772 * rules. So we'll split them off into they're own
773 * superblock and pass it back into the optimizer.
774 */
775 if (skiplist->ps_count > 2) {
776 if ((newblock = calloc(1, sizeof(*newblock)))
777 == NULL) {
778 warn("calloc");
779 return (1);
780 }
781 TAILQ_INIT(&newblock->sb_rules);
782 for (i = 0; i < PF_SKIP_COUNT; i++)
783 TAILQ_INIT(&newblock->sb_skipsteps[i]);
784 TAILQ_INSERT_BEFORE(block, newblock, sb_entry);
785 DEBUG("(%d) splitting off %d rules from superblock @ #%d",
786 depth, skiplist->ps_count,
787 TAILQ_FIRST(&skiplist->ps_rules)->
788 por_rule.nr);
789 } else {
790 newblock = block;
791 }
792
793 while ((por = TAILQ_FIRST(&skiplist->ps_rules))) {
794 TAILQ_REMOVE(&head, por, por_entry);
795 TAILQ_REMOVE(&skiplist->ps_rules, por,
796 por_skip_entry[largest_list]);
797 TAILQ_INSERT_TAIL(&newblock->sb_rules, por,
798 por_entry);
799
800 /* Remove this rule from all other skiplists */
801 remove_from_skipsteps(&block->sb_skipsteps[
802 largest_list], block, por, skiplist);
803 }
804 free(skiplist);
805 if (newblock != block)
806 if (reorder_rules(pf, newblock, depth + 1))
807 return (1);
808 }
809 }
810
811 done:
812 for (i = 0; i < PF_SKIP_COUNT; i++) {
813 while ((skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]))) {
814 TAILQ_REMOVE(&block->sb_skipsteps[i], skiplist,
815 ps_entry);
816 free(skiplist);
817 }
818 }
819
820 return (0);
821 }
822
823
824 /*
825 * Optimization pass #4: re-order 'quick' rules based on feedback from the
826 * currently running ruleset
827 */
828 int
block_feedback(struct pfctl * pf,struct superblock * block)829 block_feedback(struct pfctl *pf, struct superblock *block)
830 {
831 TAILQ_HEAD( , pf_opt_rule) queue;
832 struct pf_opt_rule *por1, *por2;
833 struct pfctl_rule a, b;
834
835
836 /*
837 * Walk through all of the profiled superblock's rules and copy
838 * the counters onto our rules.
839 */
840 TAILQ_FOREACH(por1, &block->sb_profiled_block->sb_rules, por_entry) {
841 comparable_rule(&a, &por1->por_rule, DC);
842 TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
843 if (por2->por_profile_count)
844 continue;
845 comparable_rule(&b, &por2->por_rule, DC);
846 if (memcmp(&a, &b, sizeof(a)) == 0) {
847 por2->por_profile_count =
848 por1->por_rule.packets[0] +
849 por1->por_rule.packets[1];
850 break;
851 }
852 }
853 }
854 superblock_free(pf, block->sb_profiled_block);
855 block->sb_profiled_block = NULL;
856
857 /*
858 * Now we pull all of the rules off the superblock and re-insert them
859 * in sorted order.
860 */
861
862 TAILQ_INIT(&queue);
863 while ((por1 = TAILQ_FIRST(&block->sb_rules)) != NULL) {
864 TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
865 TAILQ_INSERT_TAIL(&queue, por1, por_entry);
866 }
867
868 while ((por1 = TAILQ_FIRST(&queue)) != NULL) {
869 TAILQ_REMOVE(&queue, por1, por_entry);
870 /* XXX I should sort all of the unused rules based on skip steps */
871 TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
872 if (por1->por_profile_count > por2->por_profile_count) {
873 TAILQ_INSERT_BEFORE(por2, por1, por_entry);
874 break;
875 }
876 }
877 #ifdef __FreeBSD__
878 if (por2 == NULL)
879 #else
880 if (por2 == TAILQ_END(&block->sb_rules))
881 #endif
882 TAILQ_INSERT_TAIL(&block->sb_rules, por1, por_entry);
883 }
884
885 return (0);
886 }
887
888
889 /*
890 * Load the current ruleset from the kernel and try to associate them with
891 * the ruleset we're optimizing.
892 */
893 int
load_feedback_profile(struct pfctl * pf,struct superblocks * superblocks)894 load_feedback_profile(struct pfctl *pf, struct superblocks *superblocks)
895 {
896 char anchor_call[MAXPATHLEN] = "";
897 struct superblock *block, *blockcur;
898 struct superblocks prof_superblocks;
899 struct pf_opt_rule *por;
900 struct pf_opt_queue queue;
901 struct pfctl_rules_info rules;
902 struct pfctl_rule a, b, rule;
903 int nr, mnr;
904
905 TAILQ_INIT(&queue);
906 TAILQ_INIT(&prof_superblocks);
907
908 if (pfctl_get_rules_info_h(pf->h, &rules, PF_PASS, "")) {
909 warn("DIOCGETRULES");
910 return (1);
911 }
912 mnr = rules.nr;
913
914 DEBUG("Loading %d active rules for a feedback profile", mnr);
915 for (nr = 0; nr < mnr; ++nr) {
916 struct pfctl_ruleset *rs;
917 if ((por = calloc(1, sizeof(*por))) == NULL) {
918 warn("calloc");
919 return (1);
920 }
921
922 if (pfctl_get_rule_h(pf->h, nr, rules.ticket, "", PF_PASS,
923 &rule, anchor_call)) {
924 warn("DIOCGETRULENV");
925 return (1);
926 }
927 memcpy(&por->por_rule, &rule, sizeof(por->por_rule));
928 rs = pf_find_or_create_ruleset(anchor_call);
929 por->por_rule.anchor = rs->anchor;
930 if (TAILQ_EMPTY(&por->por_rule.rdr.list))
931 memset(&por->por_rule.rdr, 0,
932 sizeof(por->por_rule.rdr));
933 if (TAILQ_EMPTY(&por->por_rule.nat.list))
934 memset(&por->por_rule.nat, 0,
935 sizeof(por->por_rule.nat));
936 TAILQ_INSERT_TAIL(&queue, por, por_entry);
937
938 /* XXX pfctl_get_pool(pf->dev, &rule.rdr, nr, pr.ticket,
939 * PF_PASS, pf->anchor) ???
940 * ... pfctl_clear_pool(&rule.rdr)
941 */
942 }
943
944 if (construct_superblocks(pf, &queue, &prof_superblocks))
945 return (1);
946
947
948 /*
949 * Now we try to associate the active ruleset's superblocks with
950 * the superblocks we're compiling.
951 */
952 block = TAILQ_FIRST(superblocks);
953 blockcur = TAILQ_FIRST(&prof_superblocks);
954 while (block && blockcur) {
955 comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule,
956 BREAK);
957 comparable_rule(&b, &TAILQ_FIRST(&blockcur->sb_rules)->por_rule,
958 BREAK);
959 if (memcmp(&a, &b, sizeof(a)) == 0) {
960 /* The two superblocks lined up */
961 block->sb_profiled_block = blockcur;
962 } else {
963 DEBUG("superblocks don't line up between #%d and #%d",
964 TAILQ_FIRST(&block->sb_rules)->por_rule.nr,
965 TAILQ_FIRST(&blockcur->sb_rules)->por_rule.nr);
966 break;
967 }
968 block = TAILQ_NEXT(block, sb_entry);
969 blockcur = TAILQ_NEXT(blockcur, sb_entry);
970 }
971
972
973
974 /* Free any superblocks we couldn't link */
975 while (blockcur) {
976 block = TAILQ_NEXT(blockcur, sb_entry);
977 superblock_free(pf, blockcur);
978 blockcur = block;
979 }
980 return (0);
981 }
982
983
984 /*
985 * Compare a rule to a skiplist to see if the rule is a member
986 */
987 int
skip_compare(int skipnum,struct pf_skip_step * skiplist,struct pf_opt_rule * por)988 skip_compare(int skipnum, struct pf_skip_step *skiplist,
989 struct pf_opt_rule *por)
990 {
991 struct pfctl_rule *a, *b;
992 if (skipnum >= PF_SKIP_COUNT || skipnum < 0)
993 errx(1, "skip_compare() out of bounds");
994 a = &por->por_rule;
995 b = &TAILQ_FIRST(&skiplist->ps_rules)->por_rule;
996
997 return ((skip_comparitors[skipnum])(a, b));
998 }
999
1000
1001 /*
1002 * Add a rule to a skiplist
1003 */
1004 void
skip_append(struct superblock * superblock,int skipnum,struct pf_skip_step * skiplist,struct pf_opt_rule * por)1005 skip_append(struct superblock *superblock, int skipnum,
1006 struct pf_skip_step *skiplist, struct pf_opt_rule *por)
1007 {
1008 struct pf_skip_step *prev;
1009
1010 skiplist->ps_count++;
1011 TAILQ_INSERT_TAIL(&skiplist->ps_rules, por, por_skip_entry[skipnum]);
1012
1013 /* Keep the list of skiplists sorted by whichever is larger */
1014 while ((prev = TAILQ_PREV(skiplist, skiplist, ps_entry)) &&
1015 prev->ps_count < skiplist->ps_count) {
1016 TAILQ_REMOVE(&superblock->sb_skipsteps[skipnum],
1017 skiplist, ps_entry);
1018 TAILQ_INSERT_BEFORE(prev, skiplist, ps_entry);
1019 }
1020 }
1021
1022
1023 /*
1024 * Remove a rule from the other skiplist calculations.
1025 */
1026 void
remove_from_skipsteps(struct skiplist * head,struct superblock * block,struct pf_opt_rule * por,struct pf_skip_step * active_list)1027 remove_from_skipsteps(struct skiplist *head, struct superblock *block,
1028 struct pf_opt_rule *por, struct pf_skip_step *active_list)
1029 {
1030 struct pf_skip_step *sk, *next;
1031 struct pf_opt_rule *p2;
1032 int i, found;
1033
1034 for (i = 0; i < PF_SKIP_COUNT; i++) {
1035 sk = TAILQ_FIRST(&block->sb_skipsteps[i]);
1036 if (sk == NULL || sk == active_list || sk->ps_count <= 1)
1037 continue;
1038 found = 0;
1039 do {
1040 TAILQ_FOREACH(p2, &sk->ps_rules, por_skip_entry[i])
1041 if (p2 == por) {
1042 TAILQ_REMOVE(&sk->ps_rules, p2,
1043 por_skip_entry[i]);
1044 found = 1;
1045 sk->ps_count--;
1046 break;
1047 }
1048 } while (!found && (sk = TAILQ_NEXT(sk, ps_entry)));
1049 if (found && sk) {
1050 /* Does this change the sorting order? */
1051 while ((next = TAILQ_NEXT(sk, ps_entry)) &&
1052 next->ps_count > sk->ps_count) {
1053 TAILQ_REMOVE(head, sk, ps_entry);
1054 TAILQ_INSERT_AFTER(head, next, sk, ps_entry);
1055 }
1056 #ifdef OPT_DEBUG
1057 next = TAILQ_NEXT(sk, ps_entry);
1058 assert(next == NULL || next->ps_count <= sk->ps_count);
1059 #endif /* OPT_DEBUG */
1060 }
1061 }
1062 }
1063
1064
1065 /* Compare two rules AF field for skiplist construction */
1066 int
skip_cmp_af(struct pfctl_rule * a,struct pfctl_rule * b)1067 skip_cmp_af(struct pfctl_rule *a, struct pfctl_rule *b)
1068 {
1069 if (a->af != b->af || a->af == 0)
1070 return (1);
1071 return (0);
1072 }
1073
1074 /* Compare two rules DIRECTION field for skiplist construction */
1075 int
skip_cmp_dir(struct pfctl_rule * a,struct pfctl_rule * b)1076 skip_cmp_dir(struct pfctl_rule *a, struct pfctl_rule *b)
1077 {
1078 if (a->direction == 0 || a->direction != b->direction)
1079 return (1);
1080 return (0);
1081 }
1082
1083 /* Compare two rules DST Address field for skiplist construction */
1084 int
skip_cmp_dst_addr(struct pfctl_rule * a,struct pfctl_rule * b)1085 skip_cmp_dst_addr(struct pfctl_rule *a, struct pfctl_rule *b)
1086 {
1087 if (a->dst.neg != b->dst.neg ||
1088 a->dst.addr.type != b->dst.addr.type)
1089 return (1);
1090 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1091 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1092 * a->proto == IPPROTO_ICMP
1093 * return (1);
1094 */
1095 switch (a->dst.addr.type) {
1096 case PF_ADDR_ADDRMASK:
1097 if (memcmp(&a->dst.addr.v.a.addr, &b->dst.addr.v.a.addr,
1098 sizeof(a->dst.addr.v.a.addr)) ||
1099 memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask,
1100 sizeof(a->dst.addr.v.a.mask)) ||
1101 (a->dst.addr.v.a.addr.addr32[0] == 0 &&
1102 a->dst.addr.v.a.addr.addr32[1] == 0 &&
1103 a->dst.addr.v.a.addr.addr32[2] == 0 &&
1104 a->dst.addr.v.a.addr.addr32[3] == 0))
1105 return (1);
1106 return (0);
1107 case PF_ADDR_DYNIFTL:
1108 if (strcmp(a->dst.addr.v.ifname, b->dst.addr.v.ifname) != 0 ||
1109 a->dst.addr.iflags != b->dst.addr.iflags ||
1110 memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask,
1111 sizeof(a->dst.addr.v.a.mask)))
1112 return (1);
1113 return (0);
1114 case PF_ADDR_NOROUTE:
1115 case PF_ADDR_URPFFAILED:
1116 return (0);
1117 case PF_ADDR_TABLE:
1118 return (strcmp(a->dst.addr.v.tblname, b->dst.addr.v.tblname));
1119 }
1120 return (1);
1121 }
1122
1123 /* Compare two rules DST port field for skiplist construction */
1124 int
skip_cmp_dst_port(struct pfctl_rule * a,struct pfctl_rule * b)1125 skip_cmp_dst_port(struct pfctl_rule *a, struct pfctl_rule *b)
1126 {
1127 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1128 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1129 * a->proto == IPPROTO_ICMP
1130 * return (1);
1131 */
1132 if (a->dst.port_op == PF_OP_NONE || a->dst.port_op != b->dst.port_op ||
1133 a->dst.port[0] != b->dst.port[0] ||
1134 a->dst.port[1] != b->dst.port[1])
1135 return (1);
1136 return (0);
1137 }
1138
1139 /* Compare two rules IFP field for skiplist construction */
1140 int
skip_cmp_ifp(struct pfctl_rule * a,struct pfctl_rule * b)1141 skip_cmp_ifp(struct pfctl_rule *a, struct pfctl_rule *b)
1142 {
1143 if (strcmp(a->ifname, b->ifname) || a->ifname[0] == '\0')
1144 return (1);
1145 return (a->ifnot != b->ifnot);
1146 }
1147
1148 /* Compare two rules PROTO field for skiplist construction */
1149 int
skip_cmp_proto(struct pfctl_rule * a,struct pfctl_rule * b)1150 skip_cmp_proto(struct pfctl_rule *a, struct pfctl_rule *b)
1151 {
1152 return (a->proto != b->proto || a->proto == 0);
1153 }
1154
1155 /* Compare two rules SRC addr field for skiplist construction */
1156 int
skip_cmp_src_addr(struct pfctl_rule * a,struct pfctl_rule * b)1157 skip_cmp_src_addr(struct pfctl_rule *a, struct pfctl_rule *b)
1158 {
1159 if (a->src.neg != b->src.neg ||
1160 a->src.addr.type != b->src.addr.type)
1161 return (1);
1162 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1163 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1164 * a->proto == IPPROTO_ICMP
1165 * return (1);
1166 */
1167 switch (a->src.addr.type) {
1168 case PF_ADDR_ADDRMASK:
1169 if (memcmp(&a->src.addr.v.a.addr, &b->src.addr.v.a.addr,
1170 sizeof(a->src.addr.v.a.addr)) ||
1171 memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask,
1172 sizeof(a->src.addr.v.a.mask)) ||
1173 (a->src.addr.v.a.addr.addr32[0] == 0 &&
1174 a->src.addr.v.a.addr.addr32[1] == 0 &&
1175 a->src.addr.v.a.addr.addr32[2] == 0 &&
1176 a->src.addr.v.a.addr.addr32[3] == 0))
1177 return (1);
1178 return (0);
1179 case PF_ADDR_DYNIFTL:
1180 if (strcmp(a->src.addr.v.ifname, b->src.addr.v.ifname) != 0 ||
1181 a->src.addr.iflags != b->src.addr.iflags ||
1182 memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask,
1183 sizeof(a->src.addr.v.a.mask)))
1184 return (1);
1185 return (0);
1186 case PF_ADDR_NOROUTE:
1187 case PF_ADDR_URPFFAILED:
1188 return (0);
1189 case PF_ADDR_TABLE:
1190 return (strcmp(a->src.addr.v.tblname, b->src.addr.v.tblname));
1191 }
1192 return (1);
1193 }
1194
1195 /* Compare two rules SRC port field for skiplist construction */
1196 int
skip_cmp_src_port(struct pfctl_rule * a,struct pfctl_rule * b)1197 skip_cmp_src_port(struct pfctl_rule *a, struct pfctl_rule *b)
1198 {
1199 if (a->src.port_op == PF_OP_NONE || a->src.port_op != b->src.port_op ||
1200 a->src.port[0] != b->src.port[0] ||
1201 a->src.port[1] != b->src.port[1])
1202 return (1);
1203 /* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1204 * && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1205 * a->proto == IPPROTO_ICMP
1206 * return (1);
1207 */
1208 return (0);
1209 }
1210
1211
1212 void
skip_init(void)1213 skip_init(void)
1214 {
1215 struct {
1216 char *name;
1217 int skipnum;
1218 int (*func)(struct pfctl_rule *, struct pfctl_rule *);
1219 } comps[] = PF_SKIP_COMPARITORS;
1220 int skipnum, i;
1221
1222 for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++) {
1223 for (i = 0; i < sizeof(comps)/sizeof(*comps); i++)
1224 if (comps[i].skipnum == skipnum) {
1225 skip_comparitors[skipnum] = comps[i].func;
1226 skip_comparitors_names[skipnum] = comps[i].name;
1227 }
1228 }
1229 for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++)
1230 if (skip_comparitors[skipnum] == NULL)
1231 errx(1, "Need to add skip step comparitor to pfctl?!");
1232 }
1233
1234 /*
1235 * Add a host/netmask to a table
1236 */
1237 int
add_opt_table(struct pfctl * pf,struct pf_opt_tbl ** tbl,sa_family_t af,struct pf_rule_addr * addr)1238 add_opt_table(struct pfctl *pf, struct pf_opt_tbl **tbl, sa_family_t af,
1239 struct pf_rule_addr *addr)
1240 {
1241 #ifdef OPT_DEBUG
1242 char buf[128];
1243 #endif /* OPT_DEBUG */
1244 static int tablenum = 0;
1245 struct node_host node_host;
1246
1247 if (*tbl == NULL) {
1248 if ((*tbl = calloc(1, sizeof(**tbl))) == NULL ||
1249 ((*tbl)->pt_buf = calloc(1, sizeof(*(*tbl)->pt_buf))) ==
1250 NULL)
1251 err(1, "calloc");
1252 (*tbl)->pt_buf->pfrb_type = PFRB_ADDRS;
1253 SIMPLEQ_INIT(&(*tbl)->pt_nodes);
1254
1255 /* This is just a temporary table name */
1256 snprintf((*tbl)->pt_name, sizeof((*tbl)->pt_name), "%s%d",
1257 PF_OPT_TABLE_PREFIX, tablenum++);
1258 DEBUG("creating table <%s>", (*tbl)->pt_name);
1259 }
1260
1261 memset(&node_host, 0, sizeof(node_host));
1262 node_host.af = af;
1263 node_host.addr = addr->addr;
1264
1265 #ifdef OPT_DEBUG
1266 DEBUG("<%s> adding %s/%d", (*tbl)->pt_name, inet_ntop(af,
1267 &node_host.addr.v.a.addr, buf, sizeof(buf)),
1268 unmask(&node_host.addr.v.a.mask, af));
1269 #endif /* OPT_DEBUG */
1270
1271 if (append_addr_host((*tbl)->pt_buf, &node_host, 0, 0)) {
1272 warn("failed to add host");
1273 return (1);
1274 }
1275 if (pf->opts & PF_OPT_VERBOSE) {
1276 struct node_tinit *ti;
1277
1278 if ((ti = calloc(1, sizeof(*ti))) == NULL)
1279 err(1, "malloc");
1280 if ((ti->host = malloc(sizeof(*ti->host))) == NULL)
1281 err(1, "malloc");
1282 memcpy(ti->host, &node_host, sizeof(*ti->host));
1283 SIMPLEQ_INSERT_TAIL(&(*tbl)->pt_nodes, ti, entries);
1284 }
1285
1286 (*tbl)->pt_rulecount++;
1287 if ((*tbl)->pt_rulecount == TABLE_THRESHOLD)
1288 DEBUG("table <%s> now faster than skip steps", (*tbl)->pt_name);
1289
1290 return (0);
1291 }
1292
1293
1294 /*
1295 * Do the dirty work of choosing an unused table name and creating it.
1296 * (be careful with the table name, it might already be used in another anchor)
1297 */
1298 int
pf_opt_create_table(struct pfctl * pf,struct pf_opt_tbl * tbl)1299 pf_opt_create_table(struct pfctl *pf, struct pf_opt_tbl *tbl)
1300 {
1301 static int tablenum;
1302 struct pfr_table *t;
1303
1304 if (table_buffer.pfrb_type == 0) {
1305 /* Initialize the list of tables */
1306 table_buffer.pfrb_type = PFRB_TABLES;
1307 for (;;) {
1308 pfr_buf_grow(&table_buffer, table_buffer.pfrb_size);
1309 table_buffer.pfrb_size = table_buffer.pfrb_msize;
1310 if (pfr_get_tables(NULL, table_buffer.pfrb_caddr,
1311 &table_buffer.pfrb_size, PFR_FLAG_ALLRSETS))
1312 err(1, "pfr_get_tables");
1313 if (table_buffer.pfrb_size <= table_buffer.pfrb_msize)
1314 break;
1315 }
1316 table_identifier = arc4random();
1317 }
1318
1319 /* XXX would be *really* nice to avoid duplicating identical tables */
1320
1321 /* Now we have to pick a table name that isn't used */
1322 again:
1323 DEBUG("translating temporary table <%s> to <%s%x_%d>", tbl->pt_name,
1324 PF_OPT_TABLE_PREFIX, table_identifier, tablenum);
1325 snprintf(tbl->pt_name, sizeof(tbl->pt_name), "%s%x_%d",
1326 PF_OPT_TABLE_PREFIX, table_identifier, tablenum);
1327 PFRB_FOREACH(t, &table_buffer) {
1328 if (strcasecmp(t->pfrt_name, tbl->pt_name) == 0) {
1329 /* Collision. Try again */
1330 DEBUG("wow, table <%s> in use. trying again",
1331 tbl->pt_name);
1332 table_identifier = arc4random();
1333 goto again;
1334 }
1335 }
1336 tablenum++;
1337
1338
1339 if (pfctl_define_table(tbl->pt_name, PFR_TFLAG_CONST, 1,
1340 pf->astack[0]->name, tbl->pt_buf, pf->astack[0]->ruleset.tticket)) {
1341 warn("failed to create table %s in %s",
1342 tbl->pt_name, pf->astack[0]->name);
1343 return (1);
1344 }
1345 return (0);
1346 }
1347
1348 /*
1349 * Partition the flat ruleset into a list of distinct superblocks
1350 */
1351 int
construct_superblocks(struct pfctl * pf,struct pf_opt_queue * opt_queue,struct superblocks * superblocks)1352 construct_superblocks(struct pfctl *pf, struct pf_opt_queue *opt_queue,
1353 struct superblocks *superblocks)
1354 {
1355 struct superblock *block = NULL;
1356 struct pf_opt_rule *por;
1357 int i;
1358
1359 while (!TAILQ_EMPTY(opt_queue)) {
1360 por = TAILQ_FIRST(opt_queue);
1361 TAILQ_REMOVE(opt_queue, por, por_entry);
1362 if (block == NULL || !superblock_inclusive(block, por)) {
1363 if ((block = calloc(1, sizeof(*block))) == NULL) {
1364 warn("calloc");
1365 return (1);
1366 }
1367 TAILQ_INIT(&block->sb_rules);
1368 for (i = 0; i < PF_SKIP_COUNT; i++)
1369 TAILQ_INIT(&block->sb_skipsteps[i]);
1370 TAILQ_INSERT_TAIL(superblocks, block, sb_entry);
1371 }
1372 TAILQ_INSERT_TAIL(&block->sb_rules, por, por_entry);
1373 }
1374
1375 return (0);
1376 }
1377
1378
1379 /*
1380 * Compare two rule addresses
1381 */
1382 int
addrs_equal(struct pf_rule_addr * a,struct pf_rule_addr * b)1383 addrs_equal(struct pf_rule_addr *a, struct pf_rule_addr *b)
1384 {
1385 if (a->neg != b->neg)
1386 return (0);
1387 return (memcmp(&a->addr, &b->addr, sizeof(a->addr)) == 0);
1388 }
1389
1390
1391 /*
1392 * The addresses are not equal, but can we combine them into one table?
1393 */
1394 int
addrs_combineable(struct pf_rule_addr * a,struct pf_rule_addr * b)1395 addrs_combineable(struct pf_rule_addr *a, struct pf_rule_addr *b)
1396 {
1397 if (a->addr.type != PF_ADDR_ADDRMASK ||
1398 b->addr.type != PF_ADDR_ADDRMASK)
1399 return (0);
1400 if (a->neg != b->neg || a->port_op != b->port_op ||
1401 a->port[0] != b->port[0] || a->port[1] != b->port[1])
1402 return (0);
1403 return (1);
1404 }
1405
1406
1407 /*
1408 * Are we allowed to combine these two rules
1409 */
1410 int
rules_combineable(struct pfctl_rule * p1,struct pfctl_rule * p2)1411 rules_combineable(struct pfctl_rule *p1, struct pfctl_rule *p2)
1412 {
1413 struct pfctl_rule a, b;
1414
1415 comparable_rule(&a, p1, COMBINED);
1416 comparable_rule(&b, p2, COMBINED);
1417 return (memcmp(&a, &b, sizeof(a)) == 0);
1418 }
1419
1420
1421 /*
1422 * Can a rule be included inside a superblock
1423 */
1424 int
superblock_inclusive(struct superblock * block,struct pf_opt_rule * por)1425 superblock_inclusive(struct superblock *block, struct pf_opt_rule *por)
1426 {
1427 struct pfctl_rule a, b;
1428 int i, j;
1429
1430 /* First check for hard breaks */
1431 for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++) {
1432 if (pf_rule_desc[i].prf_type == BARRIER) {
1433 for (j = 0; j < pf_rule_desc[i].prf_size; j++)
1434 if (((char *)&por->por_rule)[j +
1435 pf_rule_desc[i].prf_offset] != 0)
1436 return (0);
1437 }
1438 }
1439
1440 /* per-rule src-track is also a hard break */
1441 if (por->por_rule.rule_flag & PFRULE_RULESRCTRACK)
1442 return (0);
1443
1444 /*
1445 * Have to handle interface groups separately. Consider the following
1446 * rules:
1447 * block on EXTIFS to any port 22
1448 * pass on em0 to any port 22
1449 * (where EXTIFS is an arbitrary interface group)
1450 * The optimizer may decide to re-order the pass rule in front of the
1451 * block rule. But what if EXTIFS includes em0??? Such a reordering
1452 * would change the meaning of the ruleset.
1453 * We can't just lookup the EXTIFS group and check if em0 is a member
1454 * because the user is allowed to add interfaces to a group during
1455 * runtime.
1456 * Ergo interface groups become a defacto superblock break :-(
1457 */
1458 if (interface_group(por->por_rule.ifname) ||
1459 interface_group(TAILQ_FIRST(&block->sb_rules)->por_rule.ifname)) {
1460 if (strcasecmp(por->por_rule.ifname,
1461 TAILQ_FIRST(&block->sb_rules)->por_rule.ifname) != 0)
1462 return (0);
1463 }
1464
1465 comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule, NOMERGE);
1466 comparable_rule(&b, &por->por_rule, NOMERGE);
1467 if (memcmp(&a, &b, sizeof(a)) == 0)
1468 return (1);
1469
1470 #ifdef OPT_DEBUG
1471 for (i = 0; i < sizeof(por->por_rule); i++) {
1472 int closest = -1;
1473 if (((u_int8_t *)&a)[i] != ((u_int8_t *)&b)[i]) {
1474 for (j = 0; j < sizeof(pf_rule_desc) /
1475 sizeof(*pf_rule_desc); j++) {
1476 if (i >= pf_rule_desc[j].prf_offset &&
1477 i < pf_rule_desc[j].prf_offset +
1478 pf_rule_desc[j].prf_size) {
1479 DEBUG("superblock break @ %d due to %s",
1480 por->por_rule.nr,
1481 pf_rule_desc[j].prf_name);
1482 return (0);
1483 }
1484 if (i > pf_rule_desc[j].prf_offset) {
1485 if (closest == -1 ||
1486 i-pf_rule_desc[j].prf_offset <
1487 i-pf_rule_desc[closest].prf_offset)
1488 closest = j;
1489 }
1490 }
1491
1492 if (closest >= 0)
1493 DEBUG("superblock break @ %d on %s+%zxh",
1494 por->por_rule.nr,
1495 pf_rule_desc[closest].prf_name,
1496 i - pf_rule_desc[closest].prf_offset -
1497 pf_rule_desc[closest].prf_size);
1498 else
1499 DEBUG("superblock break @ %d on field @ %d",
1500 por->por_rule.nr, i);
1501 return (0);
1502 }
1503 }
1504 #endif /* OPT_DEBUG */
1505
1506 return (0);
1507 }
1508
1509
1510 /*
1511 * Figure out if an interface name is an actual interface or actually a
1512 * group of interfaces.
1513 */
1514 int
interface_group(const char * ifname)1515 interface_group(const char *ifname)
1516 {
1517 int s;
1518 struct ifgroupreq ifgr;
1519
1520 if (ifname == NULL || !ifname[0])
1521 return (0);
1522
1523 s = get_query_socket();
1524
1525 memset(&ifgr, 0, sizeof(ifgr));
1526 strlcpy(ifgr.ifgr_name, ifname, IFNAMSIZ);
1527 if (ioctl(s, SIOCGIFGMEMB, (caddr_t)&ifgr) == -1) {
1528 if (errno == ENOENT)
1529 return (0);
1530 else
1531 err(1, "SIOCGIFGMEMB");
1532 }
1533
1534 return (1);
1535 }
1536
1537
1538 /*
1539 * Make a rule that can directly compared by memcmp()
1540 */
1541 void
comparable_rule(struct pfctl_rule * dst,const struct pfctl_rule * src,int type)1542 comparable_rule(struct pfctl_rule *dst, const struct pfctl_rule *src, int type)
1543 {
1544 int i;
1545 /*
1546 * To simplify the comparison, we just zero out the fields that are
1547 * allowed to be different and then do a simple memcmp()
1548 */
1549 memcpy(dst, src, sizeof(*dst));
1550 for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++)
1551 if (pf_rule_desc[i].prf_type >= type) {
1552 #ifdef OPT_DEBUG
1553 assert(pf_rule_desc[i].prf_type != NEVER ||
1554 *(((char *)dst) + pf_rule_desc[i].prf_offset) == 0);
1555 #endif /* OPT_DEBUG */
1556 memset(((char *)dst) + pf_rule_desc[i].prf_offset, 0,
1557 pf_rule_desc[i].prf_size);
1558 }
1559 }
1560
1561
1562 /*
1563 * Remove superset information from two rules so we can directly compare them
1564 * with memcmp()
1565 */
1566 void
exclude_supersets(struct pfctl_rule * super,struct pfctl_rule * sub)1567 exclude_supersets(struct pfctl_rule *super, struct pfctl_rule *sub)
1568 {
1569 if (super->ifname[0] == '\0')
1570 memset(sub->ifname, 0, sizeof(sub->ifname));
1571 if (super->direction == PF_INOUT)
1572 sub->direction = PF_INOUT;
1573 if ((super->proto == 0 || super->proto == sub->proto) &&
1574 super->flags == 0 && super->flagset == 0 && (sub->flags ||
1575 sub->flagset)) {
1576 sub->flags = super->flags;
1577 sub->flagset = super->flagset;
1578 }
1579 if (super->proto == 0)
1580 sub->proto = 0;
1581
1582 if (super->src.port_op == 0) {
1583 sub->src.port_op = 0;
1584 sub->src.port[0] = 0;
1585 sub->src.port[1] = 0;
1586 }
1587 if (super->dst.port_op == 0) {
1588 sub->dst.port_op = 0;
1589 sub->dst.port[0] = 0;
1590 sub->dst.port[1] = 0;
1591 }
1592
1593 if (super->src.addr.type == PF_ADDR_ADDRMASK && !super->src.neg &&
1594 !sub->src.neg && super->src.addr.v.a.mask.addr32[0] == 0 &&
1595 super->src.addr.v.a.mask.addr32[1] == 0 &&
1596 super->src.addr.v.a.mask.addr32[2] == 0 &&
1597 super->src.addr.v.a.mask.addr32[3] == 0)
1598 memset(&sub->src.addr, 0, sizeof(sub->src.addr));
1599 else if (super->src.addr.type == PF_ADDR_ADDRMASK &&
1600 sub->src.addr.type == PF_ADDR_ADDRMASK &&
1601 super->src.neg == sub->src.neg &&
1602 super->af == sub->af &&
1603 unmask(&super->src.addr.v.a.mask, super->af) <
1604 unmask(&sub->src.addr.v.a.mask, sub->af) &&
1605 super->src.addr.v.a.addr.addr32[0] ==
1606 (sub->src.addr.v.a.addr.addr32[0] &
1607 super->src.addr.v.a.mask.addr32[0]) &&
1608 super->src.addr.v.a.addr.addr32[1] ==
1609 (sub->src.addr.v.a.addr.addr32[1] &
1610 super->src.addr.v.a.mask.addr32[1]) &&
1611 super->src.addr.v.a.addr.addr32[2] ==
1612 (sub->src.addr.v.a.addr.addr32[2] &
1613 super->src.addr.v.a.mask.addr32[2]) &&
1614 super->src.addr.v.a.addr.addr32[3] ==
1615 (sub->src.addr.v.a.addr.addr32[3] &
1616 super->src.addr.v.a.mask.addr32[3])) {
1617 /* sub->src.addr is a subset of super->src.addr/mask */
1618 memcpy(&sub->src.addr, &super->src.addr, sizeof(sub->src.addr));
1619 }
1620
1621 if (super->dst.addr.type == PF_ADDR_ADDRMASK && !super->dst.neg &&
1622 !sub->dst.neg && super->dst.addr.v.a.mask.addr32[0] == 0 &&
1623 super->dst.addr.v.a.mask.addr32[1] == 0 &&
1624 super->dst.addr.v.a.mask.addr32[2] == 0 &&
1625 super->dst.addr.v.a.mask.addr32[3] == 0)
1626 memset(&sub->dst.addr, 0, sizeof(sub->dst.addr));
1627 else if (super->dst.addr.type == PF_ADDR_ADDRMASK &&
1628 sub->dst.addr.type == PF_ADDR_ADDRMASK &&
1629 super->dst.neg == sub->dst.neg &&
1630 super->af == sub->af &&
1631 unmask(&super->dst.addr.v.a.mask, super->af) <
1632 unmask(&sub->dst.addr.v.a.mask, sub->af) &&
1633 super->dst.addr.v.a.addr.addr32[0] ==
1634 (sub->dst.addr.v.a.addr.addr32[0] &
1635 super->dst.addr.v.a.mask.addr32[0]) &&
1636 super->dst.addr.v.a.addr.addr32[1] ==
1637 (sub->dst.addr.v.a.addr.addr32[1] &
1638 super->dst.addr.v.a.mask.addr32[1]) &&
1639 super->dst.addr.v.a.addr.addr32[2] ==
1640 (sub->dst.addr.v.a.addr.addr32[2] &
1641 super->dst.addr.v.a.mask.addr32[2]) &&
1642 super->dst.addr.v.a.addr.addr32[3] ==
1643 (sub->dst.addr.v.a.addr.addr32[3] &
1644 super->dst.addr.v.a.mask.addr32[3])) {
1645 /* sub->dst.addr is a subset of super->dst.addr/mask */
1646 memcpy(&sub->dst.addr, &super->dst.addr, sizeof(sub->dst.addr));
1647 }
1648
1649 if (super->af == 0)
1650 sub->af = 0;
1651 }
1652
1653
1654 void
superblock_free(struct pfctl * pf,struct superblock * block)1655 superblock_free(struct pfctl *pf, struct superblock *block)
1656 {
1657 struct pf_opt_rule *por;
1658 while ((por = TAILQ_FIRST(&block->sb_rules))) {
1659 TAILQ_REMOVE(&block->sb_rules, por, por_entry);
1660 if (por->por_src_tbl) {
1661 if (por->por_src_tbl->pt_buf) {
1662 pfr_buf_clear(por->por_src_tbl->pt_buf);
1663 free(por->por_src_tbl->pt_buf);
1664 }
1665 free(por->por_src_tbl);
1666 }
1667 if (por->por_dst_tbl) {
1668 if (por->por_dst_tbl->pt_buf) {
1669 pfr_buf_clear(por->por_dst_tbl->pt_buf);
1670 free(por->por_dst_tbl->pt_buf);
1671 }
1672 free(por->por_dst_tbl);
1673 }
1674 free(por);
1675 }
1676 if (block->sb_profiled_block)
1677 superblock_free(pf, block->sb_profiled_block);
1678 free(block);
1679 }
1680
1681