xref: /freebsd/sbin/pfctl/pfctl_optimize.c (revision b3e7694832e81d7a904a10f525f8797b753bf0d3)
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/cdefs.h>
20 __FBSDID("$FreeBSD$");
21 
22 #include <sys/types.h>
23 #include <sys/ioctl.h>
24 #include <sys/socket.h>
25 
26 #include <net/if.h>
27 #include <net/pfvar.h>
28 
29 #include <netinet/in.h>
30 #include <arpa/inet.h>
31 
32 #include <assert.h>
33 #include <ctype.h>
34 #include <err.h>
35 #include <errno.h>
36 #include <libpfctl.h>
37 #include <stddef.h>
38 #include <stdio.h>
39 #include <stdlib.h>
40 #include <string.h>
41 
42 #include "pfctl_parser.h"
43 #include "pfctl.h"
44 
45 /* The size at which a table becomes faster than individual rules */
46 #define TABLE_THRESHOLD		6
47 
48 
49 /* #define OPT_DEBUG	1 */
50 #ifdef OPT_DEBUG
51 # define DEBUG(str, v...) \
52 	printf("%s: " str "\n", __FUNCTION__ , ## v)
53 #else
54 # define DEBUG(str, v...) ((void)0)
55 #endif
56 
57 
58 /*
59  * A container that lets us sort a superblock to optimize the skip step jumps
60  */
61 struct pf_skip_step {
62 	int				ps_count;	/* number of items */
63 	TAILQ_HEAD( , pf_opt_rule)	ps_rules;
64 	TAILQ_ENTRY(pf_skip_step)	ps_entry;
65 };
66 
67 
68 /*
69  * A superblock is a block of adjacent rules of similar action.  If there
70  * are five PASS rules in a row, they all become members of a superblock.
71  * Once we have a superblock, we are free to re-order any rules within it
72  * in order to improve performance; if a packet is passed, it doesn't matter
73  * who passed it.
74  */
75 struct superblock {
76 	TAILQ_HEAD( , pf_opt_rule)		 sb_rules;
77 	TAILQ_ENTRY(superblock)			 sb_entry;
78 	struct superblock			*sb_profiled_block;
79 	TAILQ_HEAD(skiplist, pf_skip_step)	 sb_skipsteps[PF_SKIP_COUNT];
80 };
81 TAILQ_HEAD(superblocks, superblock);
82 
83 
84 /*
85  * Description of the PF rule structure.
86  */
87 enum {
88     BARRIER,	/* the presence of the field puts the rule in its own block */
89     BREAK,	/* the field may not differ between rules in a superblock */
90     NOMERGE,	/* the field may not differ between rules when combined */
91     COMBINED,	/* the field may itself be combined with other rules */
92     DC,		/* we just don't care about the field */
93     NEVER};	/* we should never see this field set?!? */
94 static struct pf_rule_field {
95 	const char	*prf_name;
96 	int		 prf_type;
97 	size_t		 prf_offset;
98 	size_t		 prf_size;
99 } pf_rule_desc[] = {
100 #define PF_RULE_FIELD(field, ty)	\
101     {#field,				\
102     ty,					\
103     offsetof(struct pfctl_rule, field),	\
104     sizeof(((struct pfctl_rule *)0)->field)}
105 
106 
107     /*
108      * The presence of these fields in a rule put the rule in its own
109      * superblock.  Thus it will not be optimized.  It also prevents the
110      * rule from being re-ordered at all.
111      */
112     PF_RULE_FIELD(label,		BARRIER),
113     PF_RULE_FIELD(prob,			BARRIER),
114     PF_RULE_FIELD(max_states,		BARRIER),
115     PF_RULE_FIELD(max_src_nodes,	BARRIER),
116     PF_RULE_FIELD(max_src_states,	BARRIER),
117     PF_RULE_FIELD(max_src_conn,		BARRIER),
118     PF_RULE_FIELD(max_src_conn_rate,	BARRIER),
119     PF_RULE_FIELD(anchor,		BARRIER),	/* for now */
120 
121     /*
122      * These fields must be the same between all rules in the same superblock.
123      * These rules are allowed to be re-ordered but only among like rules.
124      * For instance we can re-order all 'tag "foo"' rules because they have the
125      * same tag.  But we can not re-order between a 'tag "foo"' and a
126      * 'tag "bar"' since that would change the meaning of the ruleset.
127      */
128     PF_RULE_FIELD(tagname,		BREAK),
129     PF_RULE_FIELD(keep_state,		BREAK),
130     PF_RULE_FIELD(qname,		BREAK),
131     PF_RULE_FIELD(pqname,		BREAK),
132     PF_RULE_FIELD(rt,			BREAK),
133     PF_RULE_FIELD(allow_opts,		BREAK),
134     PF_RULE_FIELD(rule_flag,		BREAK),
135     PF_RULE_FIELD(action,		BREAK),
136     PF_RULE_FIELD(log,			BREAK),
137     PF_RULE_FIELD(quick,		BREAK),
138     PF_RULE_FIELD(return_ttl,		BREAK),
139     PF_RULE_FIELD(overload_tblname,	BREAK),
140     PF_RULE_FIELD(flush,		BREAK),
141     PF_RULE_FIELD(rpool,		BREAK),
142     PF_RULE_FIELD(logif,		BREAK),
143 
144     /*
145      * Any fields not listed in this structure act as BREAK fields
146      */
147 
148 
149     /*
150      * These fields must not differ when we merge two rules together but
151      * their difference isn't enough to put the rules in different superblocks.
152      * There are no problems re-ordering any rules with these fields.
153      */
154     PF_RULE_FIELD(af,			NOMERGE),
155     PF_RULE_FIELD(ifnot,		NOMERGE),
156     PF_RULE_FIELD(ifname,		NOMERGE),	/* hack for IF groups */
157     PF_RULE_FIELD(match_tag_not,	NOMERGE),
158     PF_RULE_FIELD(match_tagname,	NOMERGE),
159     PF_RULE_FIELD(os_fingerprint,	NOMERGE),
160     PF_RULE_FIELD(timeout,		NOMERGE),
161     PF_RULE_FIELD(return_icmp,		NOMERGE),
162     PF_RULE_FIELD(return_icmp6,		NOMERGE),
163     PF_RULE_FIELD(uid,			NOMERGE),
164     PF_RULE_FIELD(gid,			NOMERGE),
165     PF_RULE_FIELD(direction,		NOMERGE),
166     PF_RULE_FIELD(proto,		NOMERGE),
167     PF_RULE_FIELD(type,			NOMERGE),
168     PF_RULE_FIELD(code,			NOMERGE),
169     PF_RULE_FIELD(flags,		NOMERGE),
170     PF_RULE_FIELD(flagset,		NOMERGE),
171     PF_RULE_FIELD(tos,			NOMERGE),
172     PF_RULE_FIELD(src.port,		NOMERGE),
173     PF_RULE_FIELD(dst.port,		NOMERGE),
174     PF_RULE_FIELD(src.port_op,		NOMERGE),
175     PF_RULE_FIELD(dst.port_op,		NOMERGE),
176     PF_RULE_FIELD(src.neg,		NOMERGE),
177     PF_RULE_FIELD(dst.neg,		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     { "sport", PF_SKIP_SRC_PORT, skip_cmp_src_port },	\
256     { "daddr", PF_SKIP_DST_ADDR, skip_cmp_dst_addr },	\
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
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->rpool.list) != NULL) {
296 			TAILQ_INIT(&por->por_rule.rpool.list);
297 			pfctl_move_pool(&r->rpool, &por->por_rule.rpool);
298 		} else
299 			bzero(&por->por_rule.rpool,
300 			    sizeof(por->por_rule.rpool));
301 
302 
303 		TAILQ_INSERT_TAIL(&opt_queue, por, por_entry);
304 	}
305 
306 	TAILQ_INIT(&superblocks);
307 	if (construct_superblocks(pf, &opt_queue, &superblocks))
308 		goto error;
309 
310 	if (pf->optimize & PF_OPTIMIZE_PROFILE) {
311 		if (load_feedback_profile(pf, &superblocks))
312 			goto error;
313 	}
314 
315 	TAILQ_FOREACH(block, &superblocks, sb_entry) {
316 		if (optimize_superblock(pf, block))
317 			goto error;
318 	}
319 
320 	rs->anchor->refcnt = 0;
321 	while ((block = TAILQ_FIRST(&superblocks))) {
322 		TAILQ_REMOVE(&superblocks, block, sb_entry);
323 
324 		while ((por = TAILQ_FIRST(&block->sb_rules))) {
325 			TAILQ_REMOVE(&block->sb_rules, por, por_entry);
326 			por->por_rule.nr = rs->anchor->refcnt++;
327 			if ((r = calloc(1, sizeof(*r))) == NULL)
328 				err(1, "calloc");
329 			memcpy(r, &por->por_rule, sizeof(*r));
330 			TAILQ_INIT(&r->rpool.list);
331 			pfctl_move_pool(&por->por_rule.rpool, &r->rpool);
332 			TAILQ_INSERT_TAIL(
333 			    rs->rules[PF_RULESET_FILTER].active.ptr,
334 			    r, entries);
335 			free(por);
336 		}
337 		free(block);
338 	}
339 
340 	return (0);
341 
342 error:
343 	while ((por = TAILQ_FIRST(&opt_queue))) {
344 		TAILQ_REMOVE(&opt_queue, por, por_entry);
345 		if (por->por_src_tbl) {
346 			pfr_buf_clear(por->por_src_tbl->pt_buf);
347 			free(por->por_src_tbl->pt_buf);
348 			free(por->por_src_tbl);
349 		}
350 		if (por->por_dst_tbl) {
351 			pfr_buf_clear(por->por_dst_tbl->pt_buf);
352 			free(por->por_dst_tbl->pt_buf);
353 			free(por->por_dst_tbl);
354 		}
355 		free(por);
356 	}
357 	while ((block = TAILQ_FIRST(&superblocks))) {
358 		TAILQ_REMOVE(&superblocks, block, sb_entry);
359 		superblock_free(pf, block);
360 	}
361 	return (1);
362 }
363 
364 
365 /*
366  * Go ahead and optimize a superblock
367  */
368 int
369 optimize_superblock(struct pfctl *pf, struct superblock *block)
370 {
371 #ifdef OPT_DEBUG
372 	struct pf_opt_rule *por;
373 #endif /* OPT_DEBUG */
374 
375 	/* We have a few optimization passes:
376 	 *   1) remove duplicate rules or rules that are a subset of other
377 	 *      rules
378 	 *   2) combine otherwise identical rules with different IP addresses
379 	 *      into a single rule and put the addresses in a table.
380 	 *   3) re-order the rules to improve kernel skip steps
381 	 *   4) re-order the 'quick' rules based on feedback from the
382 	 *      active ruleset statistics
383 	 *
384 	 * XXX combine_rules() doesn't combine v4 and v6 rules.  would just
385 	 *     have to keep af in the table container, make af 'COMBINE' and
386 	 *     twiddle the af on the merged rule
387 	 * XXX maybe add a weighting to the metric on skipsteps when doing
388 	 *     reordering.  sometimes two sequential tables will be better
389 	 *     that four consecutive interfaces.
390 	 * XXX need to adjust the skipstep count of everything after PROTO,
391 	 *     since they aren't actually checked on a proto mismatch in
392 	 *     pf_test_{tcp, udp, icmp}()
393 	 * XXX should i treat proto=0, af=0 or dir=0 special in skepstep
394 	 *     calculation since they are a DC?
395 	 * XXX keep last skiplist of last superblock to influence this
396 	 *     superblock.  '5 inet6 log' should make '3 inet6' come before '4
397 	 *     inet' in the next superblock.
398 	 * XXX would be useful to add tables for ports
399 	 * XXX we can also re-order some mutually exclusive superblocks to
400 	 *     try merging superblocks before any of these optimization passes.
401 	 *     for instance a single 'log in' rule in the middle of non-logging
402 	 *     out rules.
403 	 */
404 
405 	/* shortcut.  there will be a lot of 1-rule superblocks */
406 	if (!TAILQ_NEXT(TAILQ_FIRST(&block->sb_rules), por_entry))
407 		return (0);
408 
409 #ifdef OPT_DEBUG
410 	printf("--- Superblock ---\n");
411 	TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
412 		printf("  ");
413 		print_rule(&por->por_rule, por->por_rule.anchor ?
414 		    por->por_rule.anchor->name : "", 1, 0);
415 	}
416 #endif /* OPT_DEBUG */
417 
418 
419 	if (remove_identical_rules(pf, block))
420 		return (1);
421 	if (combine_rules(pf, block))
422 		return (1);
423 	if ((pf->optimize & PF_OPTIMIZE_PROFILE) &&
424 	    TAILQ_FIRST(&block->sb_rules)->por_rule.quick &&
425 	    block->sb_profiled_block) {
426 		if (block_feedback(pf, block))
427 			return (1);
428 	} else if (reorder_rules(pf, block, 0)) {
429 		return (1);
430 	}
431 
432 	/*
433 	 * Don't add any optimization passes below reorder_rules().  It will
434 	 * have divided superblocks into smaller blocks for further refinement
435 	 * and doesn't put them back together again.  What once was a true
436 	 * superblock might have been split into multiple superblocks.
437 	 */
438 
439 #ifdef OPT_DEBUG
440 	printf("--- END Superblock ---\n");
441 #endif /* OPT_DEBUG */
442 	return (0);
443 }
444 
445 
446 /*
447  * Optimization pass #1: remove identical rules
448  */
449 int
450 remove_identical_rules(struct pfctl *pf, struct superblock *block)
451 {
452 	struct pf_opt_rule *por1, *por2, *por_next, *por2_next;
453 	struct pfctl_rule a, a2, b, b2;
454 
455 	for (por1 = TAILQ_FIRST(&block->sb_rules); por1; por1 = por_next) {
456 		por_next = TAILQ_NEXT(por1, por_entry);
457 		for (por2 = por_next; por2; por2 = por2_next) {
458 			por2_next = TAILQ_NEXT(por2, por_entry);
459 			comparable_rule(&a, &por1->por_rule, DC);
460 			comparable_rule(&b, &por2->por_rule, DC);
461 			memcpy(&a2, &a, sizeof(a2));
462 			memcpy(&b2, &b, sizeof(b2));
463 
464 			exclude_supersets(&a, &b);
465 			exclude_supersets(&b2, &a2);
466 			if (memcmp(&a, &b, sizeof(a)) == 0) {
467 				DEBUG("removing identical rule  nr%d = *nr%d*",
468 				    por1->por_rule.nr, por2->por_rule.nr);
469 				TAILQ_REMOVE(&block->sb_rules, por2, por_entry);
470 				if (por_next == por2)
471 					por_next = TAILQ_NEXT(por1, por_entry);
472 				free(por2);
473 			} else if (memcmp(&a2, &b2, sizeof(a2)) == 0) {
474 				DEBUG("removing identical rule  *nr%d* = nr%d",
475 				    por1->por_rule.nr, por2->por_rule.nr);
476 				TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
477 				free(por1);
478 				break;
479 			}
480 		}
481 	}
482 
483 	return (0);
484 }
485 
486 
487 /*
488  * Optimization pass #2: combine similar rules with different addresses
489  * into a single rule and a table
490  */
491 int
492 combine_rules(struct pfctl *pf, struct superblock *block)
493 {
494 	struct pf_opt_rule *p1, *p2, *por_next;
495 	int src_eq, dst_eq;
496 
497 	if ((pf->loadopt & PFCTL_FLAG_TABLE) == 0) {
498 		warnx("Must enable table loading for optimizations");
499 		return (1);
500 	}
501 
502 	/* First we make a pass to combine the rules.  O(n log n) */
503 	TAILQ_FOREACH(p1, &block->sb_rules, por_entry) {
504 		for (p2 = TAILQ_NEXT(p1, por_entry); p2; p2 = por_next) {
505 			por_next = TAILQ_NEXT(p2, por_entry);
506 
507 			src_eq = addrs_equal(&p1->por_rule.src,
508 			    &p2->por_rule.src);
509 			dst_eq = addrs_equal(&p1->por_rule.dst,
510 			    &p2->por_rule.dst);
511 
512 			if (src_eq && !dst_eq && p1->por_src_tbl == NULL &&
513 			    p2->por_dst_tbl == NULL &&
514 			    p2->por_src_tbl == NULL &&
515 			    rules_combineable(&p1->por_rule, &p2->por_rule) &&
516 			    addrs_combineable(&p1->por_rule.dst,
517 			    &p2->por_rule.dst)) {
518 				DEBUG("can combine rules  nr%d = nr%d",
519 				    p1->por_rule.nr, p2->por_rule.nr);
520 				if (p1->por_dst_tbl == NULL &&
521 				    add_opt_table(pf, &p1->por_dst_tbl,
522 				    p1->por_rule.af, &p1->por_rule.dst))
523 					return (1);
524 				if (add_opt_table(pf, &p1->por_dst_tbl,
525 				    p1->por_rule.af, &p2->por_rule.dst))
526 					return (1);
527 				p2->por_dst_tbl = p1->por_dst_tbl;
528 				if (p1->por_dst_tbl->pt_rulecount >=
529 				    TABLE_THRESHOLD) {
530 					TAILQ_REMOVE(&block->sb_rules, p2,
531 					    por_entry);
532 					free(p2);
533 				}
534 			} else if (!src_eq && dst_eq && p1->por_dst_tbl == NULL
535 			    && p2->por_src_tbl == NULL &&
536 			    p2->por_dst_tbl == NULL &&
537 			    rules_combineable(&p1->por_rule, &p2->por_rule) &&
538 			    addrs_combineable(&p1->por_rule.src,
539 			    &p2->por_rule.src)) {
540 				DEBUG("can combine rules  nr%d = nr%d",
541 				    p1->por_rule.nr, p2->por_rule.nr);
542 				if (p1->por_src_tbl == NULL &&
543 				    add_opt_table(pf, &p1->por_src_tbl,
544 				    p1->por_rule.af, &p1->por_rule.src))
545 					return (1);
546 				if (add_opt_table(pf, &p1->por_src_tbl,
547 				    p1->por_rule.af, &p2->por_rule.src))
548 					return (1);
549 				p2->por_src_tbl = p1->por_src_tbl;
550 				if (p1->por_src_tbl->pt_rulecount >=
551 				    TABLE_THRESHOLD) {
552 					TAILQ_REMOVE(&block->sb_rules, p2,
553 					    por_entry);
554 					free(p2);
555 				}
556 			}
557 		}
558 	}
559 
560 
561 	/*
562 	 * Then we make a final pass to create a valid table name and
563 	 * insert the name into the rules.
564 	 */
565 	for (p1 = TAILQ_FIRST(&block->sb_rules); p1; p1 = por_next) {
566 		por_next = TAILQ_NEXT(p1, por_entry);
567 		assert(p1->por_src_tbl == NULL || p1->por_dst_tbl == NULL);
568 
569 		if (p1->por_src_tbl && p1->por_src_tbl->pt_rulecount >=
570 		    TABLE_THRESHOLD) {
571 			if (p1->por_src_tbl->pt_generated) {
572 				/* This rule is included in a table */
573 				TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
574 				free(p1);
575 				continue;
576 			}
577 			p1->por_src_tbl->pt_generated = 1;
578 
579 			if ((pf->opts & PF_OPT_NOACTION) == 0 &&
580 			    pf_opt_create_table(pf, p1->por_src_tbl))
581 				return (1);
582 
583 			pf->tdirty = 1;
584 
585 			if (pf->opts & PF_OPT_VERBOSE)
586 				print_tabledef(p1->por_src_tbl->pt_name,
587 				    PFR_TFLAG_CONST, 1,
588 				    &p1->por_src_tbl->pt_nodes);
589 
590 			memset(&p1->por_rule.src.addr, 0,
591 			    sizeof(p1->por_rule.src.addr));
592 			p1->por_rule.src.addr.type = PF_ADDR_TABLE;
593 			strlcpy(p1->por_rule.src.addr.v.tblname,
594 			    p1->por_src_tbl->pt_name,
595 			    sizeof(p1->por_rule.src.addr.v.tblname));
596 
597 			pfr_buf_clear(p1->por_src_tbl->pt_buf);
598 			free(p1->por_src_tbl->pt_buf);
599 			p1->por_src_tbl->pt_buf = NULL;
600 		}
601 		if (p1->por_dst_tbl && p1->por_dst_tbl->pt_rulecount >=
602 		    TABLE_THRESHOLD) {
603 			if (p1->por_dst_tbl->pt_generated) {
604 				/* This rule is included in a table */
605 				TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
606 				free(p1);
607 				continue;
608 			}
609 			p1->por_dst_tbl->pt_generated = 1;
610 
611 			if ((pf->opts & PF_OPT_NOACTION) == 0 &&
612 			    pf_opt_create_table(pf, p1->por_dst_tbl))
613 				return (1);
614 			pf->tdirty = 1;
615 
616 			if (pf->opts & PF_OPT_VERBOSE)
617 				print_tabledef(p1->por_dst_tbl->pt_name,
618 				    PFR_TFLAG_CONST, 1,
619 				    &p1->por_dst_tbl->pt_nodes);
620 
621 			memset(&p1->por_rule.dst.addr, 0,
622 			    sizeof(p1->por_rule.dst.addr));
623 			p1->por_rule.dst.addr.type = PF_ADDR_TABLE;
624 			strlcpy(p1->por_rule.dst.addr.v.tblname,
625 			    p1->por_dst_tbl->pt_name,
626 			    sizeof(p1->por_rule.dst.addr.v.tblname));
627 
628 			pfr_buf_clear(p1->por_dst_tbl->pt_buf);
629 			free(p1->por_dst_tbl->pt_buf);
630 			p1->por_dst_tbl->pt_buf = NULL;
631 		}
632 	}
633 
634 	return (0);
635 }
636 
637 
638 /*
639  * Optimization pass #3: re-order rules to improve skip steps
640  */
641 int
642 reorder_rules(struct pfctl *pf, struct superblock *block, int depth)
643 {
644 	struct superblock *newblock;
645 	struct pf_skip_step *skiplist;
646 	struct pf_opt_rule *por;
647 	int i, largest, largest_list, rule_count = 0;
648 	TAILQ_HEAD( , pf_opt_rule) head;
649 
650 	/*
651 	 * Calculate the best-case skip steps.  We put each rule in a list
652 	 * of other rules with common fields
653 	 */
654 	for (i = 0; i < PF_SKIP_COUNT; i++) {
655 		TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
656 			TAILQ_FOREACH(skiplist, &block->sb_skipsteps[i],
657 			    ps_entry) {
658 				if (skip_compare(i, skiplist, por) == 0)
659 					break;
660 			}
661 			if (skiplist == NULL) {
662 				if ((skiplist = calloc(1, sizeof(*skiplist))) ==
663 				    NULL)
664 					err(1, "calloc");
665 				TAILQ_INIT(&skiplist->ps_rules);
666 				TAILQ_INSERT_TAIL(&block->sb_skipsteps[i],
667 				    skiplist, ps_entry);
668 			}
669 			skip_append(block, i, skiplist, por);
670 		}
671 	}
672 
673 	TAILQ_FOREACH(por, &block->sb_rules, por_entry)
674 		rule_count++;
675 
676 	/*
677 	 * Now we're going to ignore any fields that are identical between
678 	 * all of the rules in the superblock and those fields which differ
679 	 * between every rule in the superblock.
680 	 */
681 	largest = 0;
682 	for (i = 0; i < PF_SKIP_COUNT; i++) {
683 		skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
684 		if (skiplist->ps_count == rule_count) {
685 			DEBUG("(%d) original skipstep '%s' is all rules",
686 			    depth, skip_comparitors_names[i]);
687 			skiplist->ps_count = 0;
688 		} else if (skiplist->ps_count == 1) {
689 			skiplist->ps_count = 0;
690 		} else {
691 			DEBUG("(%d) original skipstep '%s' largest jump is %d",
692 			    depth, skip_comparitors_names[i],
693 			    skiplist->ps_count);
694 			if (skiplist->ps_count > largest)
695 				largest = skiplist->ps_count;
696 		}
697 	}
698 	if (largest == 0) {
699 		/* Ugh.  There is NO commonality in the superblock on which
700 		 * optimize the skipsteps optimization.
701 		 */
702 		goto done;
703 	}
704 
705 	/*
706 	 * Now we're going to empty the superblock rule list and re-create
707 	 * it based on a more optimal skipstep order.
708 	 */
709 	TAILQ_INIT(&head);
710 	while ((por = TAILQ_FIRST(&block->sb_rules))) {
711 		TAILQ_REMOVE(&block->sb_rules, por, por_entry);
712 		TAILQ_INSERT_TAIL(&head, por, por_entry);
713 	}
714 
715 
716 	while (!TAILQ_EMPTY(&head)) {
717 		largest = 1;
718 
719 		/*
720 		 * Find the most useful skip steps remaining
721 		 */
722 		for (i = 0; i < PF_SKIP_COUNT; i++) {
723 			skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
724 			if (skiplist->ps_count > largest) {
725 				largest = skiplist->ps_count;
726 				largest_list = i;
727 			}
728 		}
729 
730 		if (largest <= 1) {
731 			/*
732 			 * Nothing useful left.  Leave remaining rules in order.
733 			 */
734 			DEBUG("(%d) no more commonality for skip steps", depth);
735 			while ((por = TAILQ_FIRST(&head))) {
736 				TAILQ_REMOVE(&head, por, por_entry);
737 				TAILQ_INSERT_TAIL(&block->sb_rules, por,
738 				    por_entry);
739 			}
740 		} else {
741 			/*
742 			 * There is commonality.  Extract those common rules
743 			 * and place them in the ruleset adjacent to each
744 			 * other.
745 			 */
746 			skiplist = TAILQ_FIRST(&block->sb_skipsteps[
747 			    largest_list]);
748 			DEBUG("(%d) skipstep '%s' largest jump is %d @ #%d",
749 			    depth, skip_comparitors_names[largest_list],
750 			    largest, TAILQ_FIRST(&TAILQ_FIRST(&block->
751 			    sb_skipsteps [largest_list])->ps_rules)->
752 			    por_rule.nr);
753 			TAILQ_REMOVE(&block->sb_skipsteps[largest_list],
754 			    skiplist, ps_entry);
755 
756 
757 			/*
758 			 * There may be further commonality inside these
759 			 * rules.  So we'll split them off into they're own
760 			 * superblock and pass it back into the optimizer.
761 			 */
762 			if (skiplist->ps_count > 2) {
763 				if ((newblock = calloc(1, sizeof(*newblock)))
764 				    == NULL) {
765 					warn("calloc");
766 					return (1);
767 				}
768 				TAILQ_INIT(&newblock->sb_rules);
769 				for (i = 0; i < PF_SKIP_COUNT; i++)
770 					TAILQ_INIT(&newblock->sb_skipsteps[i]);
771 				TAILQ_INSERT_BEFORE(block, newblock, sb_entry);
772 				DEBUG("(%d) splitting off %d rules from superblock @ #%d",
773 				    depth, skiplist->ps_count,
774 				    TAILQ_FIRST(&skiplist->ps_rules)->
775 				    por_rule.nr);
776 			} else {
777 				newblock = block;
778 			}
779 
780 			while ((por = TAILQ_FIRST(&skiplist->ps_rules))) {
781 				TAILQ_REMOVE(&head, por, por_entry);
782 				TAILQ_REMOVE(&skiplist->ps_rules, por,
783 				    por_skip_entry[largest_list]);
784 				TAILQ_INSERT_TAIL(&newblock->sb_rules, por,
785 				    por_entry);
786 
787 				/* Remove this rule from all other skiplists */
788 				remove_from_skipsteps(&block->sb_skipsteps[
789 				    largest_list], block, por, skiplist);
790 			}
791 			free(skiplist);
792 			if (newblock != block)
793 				if (reorder_rules(pf, newblock, depth + 1))
794 					return (1);
795 		}
796 	}
797 
798 done:
799 	for (i = 0; i < PF_SKIP_COUNT; i++) {
800 		while ((skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]))) {
801 			TAILQ_REMOVE(&block->sb_skipsteps[i], skiplist,
802 			    ps_entry);
803 			free(skiplist);
804 		}
805 	}
806 
807 	return (0);
808 }
809 
810 
811 /*
812  * Optimization pass #4: re-order 'quick' rules based on feedback from the
813  * currently running ruleset
814  */
815 int
816 block_feedback(struct pfctl *pf, struct superblock *block)
817 {
818 	TAILQ_HEAD( , pf_opt_rule) queue;
819 	struct pf_opt_rule *por1, *por2;
820 	struct pfctl_rule a, b;
821 
822 
823 	/*
824 	 * Walk through all of the profiled superblock's rules and copy
825 	 * the counters onto our rules.
826 	 */
827 	TAILQ_FOREACH(por1, &block->sb_profiled_block->sb_rules, por_entry) {
828 		comparable_rule(&a, &por1->por_rule, DC);
829 		TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
830 			if (por2->por_profile_count)
831 				continue;
832 			comparable_rule(&b, &por2->por_rule, DC);
833 			if (memcmp(&a, &b, sizeof(a)) == 0) {
834 				por2->por_profile_count =
835 				    por1->por_rule.packets[0] +
836 				    por1->por_rule.packets[1];
837 				break;
838 			}
839 		}
840 	}
841 	superblock_free(pf, block->sb_profiled_block);
842 	block->sb_profiled_block = NULL;
843 
844 	/*
845 	 * Now we pull all of the rules off the superblock and re-insert them
846 	 * in sorted order.
847 	 */
848 
849 	TAILQ_INIT(&queue);
850 	while ((por1 = TAILQ_FIRST(&block->sb_rules)) != NULL) {
851 		TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
852 		TAILQ_INSERT_TAIL(&queue, por1, por_entry);
853 	}
854 
855 	while ((por1 = TAILQ_FIRST(&queue)) != NULL) {
856 		TAILQ_REMOVE(&queue, por1, por_entry);
857 /* XXX I should sort all of the unused rules based on skip steps */
858 		TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
859 			if (por1->por_profile_count > por2->por_profile_count) {
860 				TAILQ_INSERT_BEFORE(por2, por1, por_entry);
861 				break;
862 			}
863 		}
864 #ifdef __FreeBSD__
865 		if (por2 == NULL)
866 #else
867 		if (por2 == TAILQ_END(&block->sb_rules))
868 #endif
869 			TAILQ_INSERT_TAIL(&block->sb_rules, por1, por_entry);
870 	}
871 
872 	return (0);
873 }
874 
875 
876 /*
877  * Load the current ruleset from the kernel and try to associate them with
878  * the ruleset we're optimizing.
879  */
880 int
881 load_feedback_profile(struct pfctl *pf, struct superblocks *superblocks)
882 {
883 	struct superblock *block, *blockcur;
884 	struct superblocks prof_superblocks;
885 	struct pf_opt_rule *por;
886 	struct pf_opt_queue queue;
887 	struct pfioc_rule pr;
888 	struct pfctl_rule a, b, rule;
889 	int nr, mnr;
890 
891 	TAILQ_INIT(&queue);
892 	TAILQ_INIT(&prof_superblocks);
893 
894 	memset(&pr, 0, sizeof(pr));
895 	pr.rule.action = PF_PASS;
896 	if (ioctl(pf->dev, DIOCGETRULES, &pr)) {
897 		warn("DIOCGETRULES");
898 		return (1);
899 	}
900 	mnr = pr.nr;
901 
902 	DEBUG("Loading %d active rules for a feedback profile", mnr);
903 	for (nr = 0; nr < mnr; ++nr) {
904 		struct pfctl_ruleset *rs;
905 		if ((por = calloc(1, sizeof(*por))) == NULL) {
906 			warn("calloc");
907 			return (1);
908 		}
909 		pr.nr = nr;
910 
911 		if (pfctl_get_rule(pf->dev, nr, pr.ticket, "", PF_PASS,
912 		    &rule, pr.anchor_call)) {
913 			warn("DIOCGETRULENV");
914 			return (1);
915 		}
916 		memcpy(&por->por_rule, &rule, sizeof(por->por_rule));
917 		rs = pf_find_or_create_ruleset(pr.anchor_call);
918 		por->por_rule.anchor = rs->anchor;
919 		if (TAILQ_EMPTY(&por->por_rule.rpool.list))
920 			memset(&por->por_rule.rpool, 0,
921 			    sizeof(por->por_rule.rpool));
922 		TAILQ_INSERT_TAIL(&queue, por, por_entry);
923 
924 		/* XXX pfctl_get_pool(pf->dev, &rule.rpool, nr, pr.ticket,
925 		 *         PF_PASS, pf->anchor) ???
926 		 * ... pfctl_clear_pool(&rule.rpool)
927 		 */
928 	}
929 
930 	if (construct_superblocks(pf, &queue, &prof_superblocks))
931 		return (1);
932 
933 
934 	/*
935 	 * Now we try to associate the active ruleset's superblocks with
936 	 * the superblocks we're compiling.
937 	 */
938 	block = TAILQ_FIRST(superblocks);
939 	blockcur = TAILQ_FIRST(&prof_superblocks);
940 	while (block && blockcur) {
941 		comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule,
942 		    BREAK);
943 		comparable_rule(&b, &TAILQ_FIRST(&blockcur->sb_rules)->por_rule,
944 		    BREAK);
945 		if (memcmp(&a, &b, sizeof(a)) == 0) {
946 			/* The two superblocks lined up */
947 			block->sb_profiled_block = blockcur;
948 		} else {
949 			DEBUG("superblocks don't line up between #%d and #%d",
950 			    TAILQ_FIRST(&block->sb_rules)->por_rule.nr,
951 			    TAILQ_FIRST(&blockcur->sb_rules)->por_rule.nr);
952 			break;
953 		}
954 		block = TAILQ_NEXT(block, sb_entry);
955 		blockcur = TAILQ_NEXT(blockcur, sb_entry);
956 	}
957 
958 
959 
960 	/* Free any superblocks we couldn't link */
961 	while (blockcur) {
962 		block = TAILQ_NEXT(blockcur, sb_entry);
963 		superblock_free(pf, blockcur);
964 		blockcur = block;
965 	}
966 	return (0);
967 }
968 
969 
970 /*
971  * Compare a rule to a skiplist to see if the rule is a member
972  */
973 int
974 skip_compare(int skipnum, struct pf_skip_step *skiplist,
975     struct pf_opt_rule *por)
976 {
977 	struct pfctl_rule *a, *b;
978 	if (skipnum >= PF_SKIP_COUNT || skipnum < 0)
979 		errx(1, "skip_compare() out of bounds");
980 	a = &por->por_rule;
981 	b = &TAILQ_FIRST(&skiplist->ps_rules)->por_rule;
982 
983 	return ((skip_comparitors[skipnum])(a, b));
984 }
985 
986 
987 /*
988  * Add a rule to a skiplist
989  */
990 void
991 skip_append(struct superblock *superblock, int skipnum,
992     struct pf_skip_step *skiplist, struct pf_opt_rule *por)
993 {
994 	struct pf_skip_step *prev;
995 
996 	skiplist->ps_count++;
997 	TAILQ_INSERT_TAIL(&skiplist->ps_rules, por, por_skip_entry[skipnum]);
998 
999 	/* Keep the list of skiplists sorted by whichever is larger */
1000 	while ((prev = TAILQ_PREV(skiplist, skiplist, ps_entry)) &&
1001 	    prev->ps_count < skiplist->ps_count) {
1002 		TAILQ_REMOVE(&superblock->sb_skipsteps[skipnum],
1003 		    skiplist, ps_entry);
1004 		TAILQ_INSERT_BEFORE(prev, skiplist, ps_entry);
1005 	}
1006 }
1007 
1008 
1009 /*
1010  * Remove a rule from the other skiplist calculations.
1011  */
1012 void
1013 remove_from_skipsteps(struct skiplist *head, struct superblock *block,
1014     struct pf_opt_rule *por, struct pf_skip_step *active_list)
1015 {
1016 	struct pf_skip_step *sk, *next;
1017 	struct pf_opt_rule *p2;
1018 	int i, found;
1019 
1020 	for (i = 0; i < PF_SKIP_COUNT; i++) {
1021 		sk = TAILQ_FIRST(&block->sb_skipsteps[i]);
1022 		if (sk == NULL || sk == active_list || sk->ps_count <= 1)
1023 			continue;
1024 		found = 0;
1025 		do {
1026 			TAILQ_FOREACH(p2, &sk->ps_rules, por_skip_entry[i])
1027 				if (p2 == por) {
1028 					TAILQ_REMOVE(&sk->ps_rules, p2,
1029 					    por_skip_entry[i]);
1030 					found = 1;
1031 					sk->ps_count--;
1032 					break;
1033 				}
1034 		} while (!found && (sk = TAILQ_NEXT(sk, ps_entry)));
1035 		if (found && sk) {
1036 			/* Does this change the sorting order? */
1037 			while ((next = TAILQ_NEXT(sk, ps_entry)) &&
1038 			    next->ps_count > sk->ps_count) {
1039 				TAILQ_REMOVE(head, sk, ps_entry);
1040 				TAILQ_INSERT_AFTER(head, next, sk, ps_entry);
1041 			}
1042 #ifdef OPT_DEBUG
1043 			next = TAILQ_NEXT(sk, ps_entry);
1044 			assert(next == NULL || next->ps_count <= sk->ps_count);
1045 #endif /* OPT_DEBUG */
1046 		}
1047 	}
1048 }
1049 
1050 
1051 /* Compare two rules AF field for skiplist construction */
1052 int
1053 skip_cmp_af(struct pfctl_rule *a, struct pfctl_rule *b)
1054 {
1055 	if (a->af != b->af || a->af == 0)
1056 		return (1);
1057 	return (0);
1058 }
1059 
1060 /* Compare two rules DIRECTION field for skiplist construction */
1061 int
1062 skip_cmp_dir(struct pfctl_rule *a, struct pfctl_rule *b)
1063 {
1064 	if (a->direction == 0 || a->direction != b->direction)
1065 		return (1);
1066 	return (0);
1067 }
1068 
1069 /* Compare two rules DST Address field for skiplist construction */
1070 int
1071 skip_cmp_dst_addr(struct pfctl_rule *a, struct pfctl_rule *b)
1072 {
1073 	if (a->dst.neg != b->dst.neg ||
1074 	    a->dst.addr.type != b->dst.addr.type)
1075 		return (1);
1076 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1077 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1078 	 *    a->proto == IPPROTO_ICMP
1079 	 *	return (1);
1080 	 */
1081 	switch (a->dst.addr.type) {
1082 	case PF_ADDR_ADDRMASK:
1083 		if (memcmp(&a->dst.addr.v.a.addr, &b->dst.addr.v.a.addr,
1084 		    sizeof(a->dst.addr.v.a.addr)) ||
1085 		    memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask,
1086 		    sizeof(a->dst.addr.v.a.mask)) ||
1087 		    (a->dst.addr.v.a.addr.addr32[0] == 0 &&
1088 		    a->dst.addr.v.a.addr.addr32[1] == 0 &&
1089 		    a->dst.addr.v.a.addr.addr32[2] == 0 &&
1090 		    a->dst.addr.v.a.addr.addr32[3] == 0))
1091 			return (1);
1092 		return (0);
1093 	case PF_ADDR_DYNIFTL:
1094 		if (strcmp(a->dst.addr.v.ifname, b->dst.addr.v.ifname) != 0 ||
1095 		    a->dst.addr.iflags != b->dst.addr.iflags ||
1096 		    memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask,
1097 		    sizeof(a->dst.addr.v.a.mask)))
1098 			return (1);
1099 		return (0);
1100 	case PF_ADDR_NOROUTE:
1101 	case PF_ADDR_URPFFAILED:
1102 		return (0);
1103 	case PF_ADDR_TABLE:
1104 		return (strcmp(a->dst.addr.v.tblname, b->dst.addr.v.tblname));
1105 	}
1106 	return (1);
1107 }
1108 
1109 /* Compare two rules DST port field for skiplist construction */
1110 int
1111 skip_cmp_dst_port(struct pfctl_rule *a, struct pfctl_rule *b)
1112 {
1113 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1114 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1115 	 *    a->proto == IPPROTO_ICMP
1116 	 *	return (1);
1117 	 */
1118 	if (a->dst.port_op == PF_OP_NONE || a->dst.port_op != b->dst.port_op ||
1119 	    a->dst.port[0] != b->dst.port[0] ||
1120 	    a->dst.port[1] != b->dst.port[1])
1121 		return (1);
1122 	return (0);
1123 }
1124 
1125 /* Compare two rules IFP field for skiplist construction */
1126 int
1127 skip_cmp_ifp(struct pfctl_rule *a, struct pfctl_rule *b)
1128 {
1129 	if (strcmp(a->ifname, b->ifname) || a->ifname[0] == '\0')
1130 		return (1);
1131 	return (a->ifnot != b->ifnot);
1132 }
1133 
1134 /* Compare two rules PROTO field for skiplist construction */
1135 int
1136 skip_cmp_proto(struct pfctl_rule *a, struct pfctl_rule *b)
1137 {
1138 	return (a->proto != b->proto || a->proto == 0);
1139 }
1140 
1141 /* Compare two rules SRC addr field for skiplist construction */
1142 int
1143 skip_cmp_src_addr(struct pfctl_rule *a, struct pfctl_rule *b)
1144 {
1145 	if (a->src.neg != b->src.neg ||
1146 	    a->src.addr.type != b->src.addr.type)
1147 		return (1);
1148 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1149 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1150 	 *    a->proto == IPPROTO_ICMP
1151 	 *	return (1);
1152 	 */
1153 	switch (a->src.addr.type) {
1154 	case PF_ADDR_ADDRMASK:
1155 		if (memcmp(&a->src.addr.v.a.addr, &b->src.addr.v.a.addr,
1156 		    sizeof(a->src.addr.v.a.addr)) ||
1157 		    memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask,
1158 		    sizeof(a->src.addr.v.a.mask)) ||
1159 		    (a->src.addr.v.a.addr.addr32[0] == 0 &&
1160 		    a->src.addr.v.a.addr.addr32[1] == 0 &&
1161 		    a->src.addr.v.a.addr.addr32[2] == 0 &&
1162 		    a->src.addr.v.a.addr.addr32[3] == 0))
1163 			return (1);
1164 		return (0);
1165 	case PF_ADDR_DYNIFTL:
1166 		if (strcmp(a->src.addr.v.ifname, b->src.addr.v.ifname) != 0 ||
1167 		    a->src.addr.iflags != b->src.addr.iflags ||
1168 		    memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask,
1169 		    sizeof(a->src.addr.v.a.mask)))
1170 			return (1);
1171 		return (0);
1172 	case PF_ADDR_NOROUTE:
1173 	case PF_ADDR_URPFFAILED:
1174 		return (0);
1175 	case PF_ADDR_TABLE:
1176 		return (strcmp(a->src.addr.v.tblname, b->src.addr.v.tblname));
1177 	}
1178 	return (1);
1179 }
1180 
1181 /* Compare two rules SRC port field for skiplist construction */
1182 int
1183 skip_cmp_src_port(struct pfctl_rule *a, struct pfctl_rule *b)
1184 {
1185 	if (a->src.port_op == PF_OP_NONE || a->src.port_op != b->src.port_op ||
1186 	    a->src.port[0] != b->src.port[0] ||
1187 	    a->src.port[1] != b->src.port[1])
1188 		return (1);
1189 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1190 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1191 	 *    a->proto == IPPROTO_ICMP
1192 	 *	return (1);
1193 	 */
1194 	return (0);
1195 }
1196 
1197 
1198 void
1199 skip_init(void)
1200 {
1201 	struct {
1202 		char *name;
1203 		int skipnum;
1204 		int (*func)(struct pfctl_rule *, struct pfctl_rule *);
1205 	} comps[] = PF_SKIP_COMPARITORS;
1206 	int skipnum, i;
1207 
1208 	for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++) {
1209 		for (i = 0; i < sizeof(comps)/sizeof(*comps); i++)
1210 			if (comps[i].skipnum == skipnum) {
1211 				skip_comparitors[skipnum] = comps[i].func;
1212 				skip_comparitors_names[skipnum] = comps[i].name;
1213 			}
1214 	}
1215 	for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++)
1216 		if (skip_comparitors[skipnum] == NULL)
1217 			errx(1, "Need to add skip step comparitor to pfctl?!");
1218 }
1219 
1220 /*
1221  * Add a host/netmask to a table
1222  */
1223 int
1224 add_opt_table(struct pfctl *pf, struct pf_opt_tbl **tbl, sa_family_t af,
1225     struct pf_rule_addr *addr)
1226 {
1227 #ifdef OPT_DEBUG
1228 	char buf[128];
1229 #endif /* OPT_DEBUG */
1230 	static int tablenum = 0;
1231 	struct node_host node_host;
1232 
1233 	if (*tbl == NULL) {
1234 		if ((*tbl = calloc(1, sizeof(**tbl))) == NULL ||
1235 		    ((*tbl)->pt_buf = calloc(1, sizeof(*(*tbl)->pt_buf))) ==
1236 		    NULL)
1237 			err(1, "calloc");
1238 		(*tbl)->pt_buf->pfrb_type = PFRB_ADDRS;
1239 		SIMPLEQ_INIT(&(*tbl)->pt_nodes);
1240 
1241 		/* This is just a temporary table name */
1242 		snprintf((*tbl)->pt_name, sizeof((*tbl)->pt_name), "%s%d",
1243 		    PF_OPT_TABLE_PREFIX, tablenum++);
1244 		DEBUG("creating table <%s>", (*tbl)->pt_name);
1245 	}
1246 
1247 	memset(&node_host, 0, sizeof(node_host));
1248 	node_host.af = af;
1249 	node_host.addr = addr->addr;
1250 
1251 #ifdef OPT_DEBUG
1252 	DEBUG("<%s> adding %s/%d", (*tbl)->pt_name, inet_ntop(af,
1253 	    &node_host.addr.v.a.addr, buf, sizeof(buf)),
1254 	    unmask(&node_host.addr.v.a.mask, af));
1255 #endif /* OPT_DEBUG */
1256 
1257 	if (append_addr_host((*tbl)->pt_buf, &node_host, 0, 0)) {
1258 		warn("failed to add host");
1259 		return (1);
1260 	}
1261 	if (pf->opts & PF_OPT_VERBOSE) {
1262 		struct node_tinit *ti;
1263 
1264 		if ((ti = calloc(1, sizeof(*ti))) == NULL)
1265 			err(1, "malloc");
1266 		if ((ti->host = malloc(sizeof(*ti->host))) == NULL)
1267 			err(1, "malloc");
1268 		memcpy(ti->host, &node_host, sizeof(*ti->host));
1269 		SIMPLEQ_INSERT_TAIL(&(*tbl)->pt_nodes, ti, entries);
1270 	}
1271 
1272 	(*tbl)->pt_rulecount++;
1273 	if ((*tbl)->pt_rulecount == TABLE_THRESHOLD)
1274 		DEBUG("table <%s> now faster than skip steps", (*tbl)->pt_name);
1275 
1276 	return (0);
1277 }
1278 
1279 
1280 /*
1281  * Do the dirty work of choosing an unused table name and creating it.
1282  * (be careful with the table name, it might already be used in another anchor)
1283  */
1284 int
1285 pf_opt_create_table(struct pfctl *pf, struct pf_opt_tbl *tbl)
1286 {
1287 	static int tablenum;
1288 	struct pfr_table *t;
1289 
1290 	if (table_buffer.pfrb_type == 0) {
1291 		/* Initialize the list of tables */
1292 		table_buffer.pfrb_type = PFRB_TABLES;
1293 		for (;;) {
1294 			pfr_buf_grow(&table_buffer, table_buffer.pfrb_size);
1295 			table_buffer.pfrb_size = table_buffer.pfrb_msize;
1296 			if (pfr_get_tables(NULL, table_buffer.pfrb_caddr,
1297 			    &table_buffer.pfrb_size, PFR_FLAG_ALLRSETS))
1298 				err(1, "pfr_get_tables");
1299 			if (table_buffer.pfrb_size <= table_buffer.pfrb_msize)
1300 				break;
1301 		}
1302 		table_identifier = arc4random();
1303 	}
1304 
1305 	/* XXX would be *really* nice to avoid duplicating identical tables */
1306 
1307 	/* Now we have to pick a table name that isn't used */
1308 again:
1309 	DEBUG("translating temporary table <%s> to <%s%x_%d>", tbl->pt_name,
1310 	    PF_OPT_TABLE_PREFIX, table_identifier, tablenum);
1311 	snprintf(tbl->pt_name, sizeof(tbl->pt_name), "%s%x_%d",
1312 	    PF_OPT_TABLE_PREFIX, table_identifier, tablenum);
1313 	PFRB_FOREACH(t, &table_buffer) {
1314 		if (strcasecmp(t->pfrt_name, tbl->pt_name) == 0) {
1315 			/* Collision.  Try again */
1316 			DEBUG("wow, table <%s> in use.  trying again",
1317 			    tbl->pt_name);
1318 			table_identifier = arc4random();
1319 			goto again;
1320 		}
1321 	}
1322 	tablenum++;
1323 
1324 
1325 	if (pfctl_define_table(tbl->pt_name, PFR_TFLAG_CONST, 1,
1326 	    pf->astack[0]->name, tbl->pt_buf, pf->astack[0]->ruleset.tticket)) {
1327 		warn("failed to create table %s in %s",
1328 		    tbl->pt_name, pf->astack[0]->name);
1329 		return (1);
1330 	}
1331 	return (0);
1332 }
1333 
1334 /*
1335  * Partition the flat ruleset into a list of distinct superblocks
1336  */
1337 int
1338 construct_superblocks(struct pfctl *pf, struct pf_opt_queue *opt_queue,
1339     struct superblocks *superblocks)
1340 {
1341 	struct superblock *block = NULL;
1342 	struct pf_opt_rule *por;
1343 	int i;
1344 
1345 	while (!TAILQ_EMPTY(opt_queue)) {
1346 		por = TAILQ_FIRST(opt_queue);
1347 		TAILQ_REMOVE(opt_queue, por, por_entry);
1348 		if (block == NULL || !superblock_inclusive(block, por)) {
1349 			if ((block = calloc(1, sizeof(*block))) == NULL) {
1350 				warn("calloc");
1351 				return (1);
1352 			}
1353 			TAILQ_INIT(&block->sb_rules);
1354 			for (i = 0; i < PF_SKIP_COUNT; i++)
1355 				TAILQ_INIT(&block->sb_skipsteps[i]);
1356 			TAILQ_INSERT_TAIL(superblocks, block, sb_entry);
1357 		}
1358 		TAILQ_INSERT_TAIL(&block->sb_rules, por, por_entry);
1359 	}
1360 
1361 	return (0);
1362 }
1363 
1364 
1365 /*
1366  * Compare two rule addresses
1367  */
1368 int
1369 addrs_equal(struct pf_rule_addr *a, struct pf_rule_addr *b)
1370 {
1371 	if (a->neg != b->neg)
1372 		return (0);
1373 	return (memcmp(&a->addr, &b->addr, sizeof(a->addr)) == 0);
1374 }
1375 
1376 
1377 /*
1378  * The addresses are not equal, but can we combine them into one table?
1379  */
1380 int
1381 addrs_combineable(struct pf_rule_addr *a, struct pf_rule_addr *b)
1382 {
1383 	if (a->addr.type != PF_ADDR_ADDRMASK ||
1384 	    b->addr.type != PF_ADDR_ADDRMASK)
1385 		return (0);
1386 	if (a->neg != b->neg || a->port_op != b->port_op ||
1387 	    a->port[0] != b->port[0] || a->port[1] != b->port[1])
1388 		return (0);
1389 	return (1);
1390 }
1391 
1392 
1393 /*
1394  * Are we allowed to combine these two rules
1395  */
1396 int
1397 rules_combineable(struct pfctl_rule *p1, struct pfctl_rule *p2)
1398 {
1399 	struct pfctl_rule a, b;
1400 
1401 	comparable_rule(&a, p1, COMBINED);
1402 	comparable_rule(&b, p2, COMBINED);
1403 	return (memcmp(&a, &b, sizeof(a)) == 0);
1404 }
1405 
1406 
1407 /*
1408  * Can a rule be included inside a superblock
1409  */
1410 int
1411 superblock_inclusive(struct superblock *block, struct pf_opt_rule *por)
1412 {
1413 	struct pfctl_rule a, b;
1414 	int i, j;
1415 
1416 	/* First check for hard breaks */
1417 	for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++) {
1418 		if (pf_rule_desc[i].prf_type == BARRIER) {
1419 			for (j = 0; j < pf_rule_desc[i].prf_size; j++)
1420 				if (((char *)&por->por_rule)[j +
1421 				    pf_rule_desc[i].prf_offset] != 0)
1422 					return (0);
1423 		}
1424 	}
1425 
1426 	/* per-rule src-track is also a hard break */
1427 	if (por->por_rule.rule_flag & PFRULE_RULESRCTRACK)
1428 		return (0);
1429 
1430 	/*
1431 	 * Have to handle interface groups separately.  Consider the following
1432 	 * rules:
1433 	 *	block on EXTIFS to any port 22
1434 	 *	pass  on em0 to any port 22
1435 	 * (where EXTIFS is an arbitrary interface group)
1436 	 * The optimizer may decide to re-order the pass rule in front of the
1437 	 * block rule.  But what if EXTIFS includes em0???  Such a reordering
1438 	 * would change the meaning of the ruleset.
1439 	 * We can't just lookup the EXTIFS group and check if em0 is a member
1440 	 * because the user is allowed to add interfaces to a group during
1441 	 * runtime.
1442 	 * Ergo interface groups become a defacto superblock break :-(
1443 	 */
1444 	if (interface_group(por->por_rule.ifname) ||
1445 	    interface_group(TAILQ_FIRST(&block->sb_rules)->por_rule.ifname)) {
1446 		if (strcasecmp(por->por_rule.ifname,
1447 		    TAILQ_FIRST(&block->sb_rules)->por_rule.ifname) != 0)
1448 			return (0);
1449 	}
1450 
1451 	comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule, NOMERGE);
1452 	comparable_rule(&b, &por->por_rule, NOMERGE);
1453 	if (memcmp(&a, &b, sizeof(a)) == 0)
1454 		return (1);
1455 
1456 #ifdef OPT_DEBUG
1457 	for (i = 0; i < sizeof(por->por_rule); i++) {
1458 		int closest = -1;
1459 		if (((u_int8_t *)&a)[i] != ((u_int8_t *)&b)[i]) {
1460 			for (j = 0; j < sizeof(pf_rule_desc) /
1461 			    sizeof(*pf_rule_desc); j++) {
1462 				if (i >= pf_rule_desc[j].prf_offset &&
1463 				    i < pf_rule_desc[j].prf_offset +
1464 				    pf_rule_desc[j].prf_size) {
1465 					DEBUG("superblock break @ %d due to %s",
1466 					    por->por_rule.nr,
1467 					    pf_rule_desc[j].prf_name);
1468 					return (0);
1469 				}
1470 				if (i > pf_rule_desc[j].prf_offset) {
1471 					if (closest == -1 ||
1472 					    i-pf_rule_desc[j].prf_offset <
1473 					    i-pf_rule_desc[closest].prf_offset)
1474 						closest = j;
1475 				}
1476 			}
1477 
1478 			if (closest >= 0)
1479 				DEBUG("superblock break @ %d on %s+%zxh",
1480 				    por->por_rule.nr,
1481 				    pf_rule_desc[closest].prf_name,
1482 				    i - pf_rule_desc[closest].prf_offset -
1483 				    pf_rule_desc[closest].prf_size);
1484 			else
1485 				DEBUG("superblock break @ %d on field @ %d",
1486 				    por->por_rule.nr, i);
1487 			return (0);
1488 		}
1489 	}
1490 #endif /* OPT_DEBUG */
1491 
1492 	return (0);
1493 }
1494 
1495 
1496 /*
1497  * Figure out if an interface name is an actual interface or actually a
1498  * group of interfaces.
1499  */
1500 int
1501 interface_group(const char *ifname)
1502 {
1503 	int			s;
1504 	struct ifgroupreq	ifgr;
1505 
1506 	if (ifname == NULL || !ifname[0])
1507 		return (0);
1508 
1509 	s = get_query_socket();
1510 
1511 	memset(&ifgr, 0, sizeof(ifgr));
1512 	strlcpy(ifgr.ifgr_name, ifname, IFNAMSIZ);
1513 	if (ioctl(s, SIOCGIFGMEMB, (caddr_t)&ifgr) == -1) {
1514 		if (errno == ENOENT)
1515 			return (0);
1516 		else
1517 			err(1, "SIOCGIFGMEMB");
1518 	}
1519 
1520 	return (1);
1521 }
1522 
1523 
1524 /*
1525  * Make a rule that can directly compared by memcmp()
1526  */
1527 void
1528 comparable_rule(struct pfctl_rule *dst, const struct pfctl_rule *src, int type)
1529 {
1530 	int i;
1531 	/*
1532 	 * To simplify the comparison, we just zero out the fields that are
1533 	 * allowed to be different and then do a simple memcmp()
1534 	 */
1535 	memcpy(dst, src, sizeof(*dst));
1536 	for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++)
1537 		if (pf_rule_desc[i].prf_type >= type) {
1538 #ifdef OPT_DEBUG
1539 			assert(pf_rule_desc[i].prf_type != NEVER ||
1540 			    *(((char *)dst) + pf_rule_desc[i].prf_offset) == 0);
1541 #endif /* OPT_DEBUG */
1542 			memset(((char *)dst) + pf_rule_desc[i].prf_offset, 0,
1543 			    pf_rule_desc[i].prf_size);
1544 		}
1545 }
1546 
1547 
1548 /*
1549  * Remove superset information from two rules so we can directly compare them
1550  * with memcmp()
1551  */
1552 void
1553 exclude_supersets(struct pfctl_rule *super, struct pfctl_rule *sub)
1554 {
1555 	if (super->ifname[0] == '\0')
1556 		memset(sub->ifname, 0, sizeof(sub->ifname));
1557 	if (super->direction == PF_INOUT)
1558 		sub->direction = PF_INOUT;
1559 	if ((super->proto == 0 || super->proto == sub->proto) &&
1560 	    super->flags == 0 && super->flagset == 0 && (sub->flags ||
1561 	    sub->flagset)) {
1562 		sub->flags = super->flags;
1563 		sub->flagset = super->flagset;
1564 	}
1565 	if (super->proto == 0)
1566 		sub->proto = 0;
1567 
1568 	if (super->src.port_op == 0) {
1569 		sub->src.port_op = 0;
1570 		sub->src.port[0] = 0;
1571 		sub->src.port[1] = 0;
1572 	}
1573 	if (super->dst.port_op == 0) {
1574 		sub->dst.port_op = 0;
1575 		sub->dst.port[0] = 0;
1576 		sub->dst.port[1] = 0;
1577 	}
1578 
1579 	if (super->src.addr.type == PF_ADDR_ADDRMASK && !super->src.neg &&
1580 	    !sub->src.neg && super->src.addr.v.a.mask.addr32[0] == 0 &&
1581 	    super->src.addr.v.a.mask.addr32[1] == 0 &&
1582 	    super->src.addr.v.a.mask.addr32[2] == 0 &&
1583 	    super->src.addr.v.a.mask.addr32[3] == 0)
1584 		memset(&sub->src.addr, 0, sizeof(sub->src.addr));
1585 	else if (super->src.addr.type == PF_ADDR_ADDRMASK &&
1586 	    sub->src.addr.type == PF_ADDR_ADDRMASK &&
1587 	    super->src.neg == sub->src.neg &&
1588 	    super->af == sub->af &&
1589 	    unmask(&super->src.addr.v.a.mask, super->af) <
1590 	    unmask(&sub->src.addr.v.a.mask, sub->af) &&
1591 	    super->src.addr.v.a.addr.addr32[0] ==
1592 	    (sub->src.addr.v.a.addr.addr32[0] &
1593 	    super->src.addr.v.a.mask.addr32[0]) &&
1594 	    super->src.addr.v.a.addr.addr32[1] ==
1595 	    (sub->src.addr.v.a.addr.addr32[1] &
1596 	    super->src.addr.v.a.mask.addr32[1]) &&
1597 	    super->src.addr.v.a.addr.addr32[2] ==
1598 	    (sub->src.addr.v.a.addr.addr32[2] &
1599 	    super->src.addr.v.a.mask.addr32[2]) &&
1600 	    super->src.addr.v.a.addr.addr32[3] ==
1601 	    (sub->src.addr.v.a.addr.addr32[3] &
1602 	    super->src.addr.v.a.mask.addr32[3])) {
1603 		/* sub->src.addr is a subset of super->src.addr/mask */
1604 		memcpy(&sub->src.addr, &super->src.addr, sizeof(sub->src.addr));
1605 	}
1606 
1607 	if (super->dst.addr.type == PF_ADDR_ADDRMASK && !super->dst.neg &&
1608 	    !sub->dst.neg && super->dst.addr.v.a.mask.addr32[0] == 0 &&
1609 	    super->dst.addr.v.a.mask.addr32[1] == 0 &&
1610 	    super->dst.addr.v.a.mask.addr32[2] == 0 &&
1611 	    super->dst.addr.v.a.mask.addr32[3] == 0)
1612 		memset(&sub->dst.addr, 0, sizeof(sub->dst.addr));
1613 	else if (super->dst.addr.type == PF_ADDR_ADDRMASK &&
1614 	    sub->dst.addr.type == PF_ADDR_ADDRMASK &&
1615 	    super->dst.neg == sub->dst.neg &&
1616 	    super->af == sub->af &&
1617 	    unmask(&super->dst.addr.v.a.mask, super->af) <
1618 	    unmask(&sub->dst.addr.v.a.mask, sub->af) &&
1619 	    super->dst.addr.v.a.addr.addr32[0] ==
1620 	    (sub->dst.addr.v.a.addr.addr32[0] &
1621 	    super->dst.addr.v.a.mask.addr32[0]) &&
1622 	    super->dst.addr.v.a.addr.addr32[1] ==
1623 	    (sub->dst.addr.v.a.addr.addr32[1] &
1624 	    super->dst.addr.v.a.mask.addr32[1]) &&
1625 	    super->dst.addr.v.a.addr.addr32[2] ==
1626 	    (sub->dst.addr.v.a.addr.addr32[2] &
1627 	    super->dst.addr.v.a.mask.addr32[2]) &&
1628 	    super->dst.addr.v.a.addr.addr32[3] ==
1629 	    (sub->dst.addr.v.a.addr.addr32[3] &
1630 	    super->dst.addr.v.a.mask.addr32[3])) {
1631 		/* sub->dst.addr is a subset of super->dst.addr/mask */
1632 		memcpy(&sub->dst.addr, &super->dst.addr, sizeof(sub->dst.addr));
1633 	}
1634 
1635 	if (super->af == 0)
1636 		sub->af = 0;
1637 }
1638 
1639 
1640 void
1641 superblock_free(struct pfctl *pf, struct superblock *block)
1642 {
1643 	struct pf_opt_rule *por;
1644 	while ((por = TAILQ_FIRST(&block->sb_rules))) {
1645 		TAILQ_REMOVE(&block->sb_rules, por, por_entry);
1646 		if (por->por_src_tbl) {
1647 			if (por->por_src_tbl->pt_buf) {
1648 				pfr_buf_clear(por->por_src_tbl->pt_buf);
1649 				free(por->por_src_tbl->pt_buf);
1650 			}
1651 			free(por->por_src_tbl);
1652 		}
1653 		if (por->por_dst_tbl) {
1654 			if (por->por_dst_tbl->pt_buf) {
1655 				pfr_buf_clear(por->por_dst_tbl->pt_buf);
1656 				free(por->por_dst_tbl->pt_buf);
1657 			}
1658 			free(por->por_dst_tbl);
1659 		}
1660 		free(por);
1661 	}
1662 	if (block->sb_profiled_block)
1663 		superblock_free(pf, block->sb_profiled_block);
1664 	free(block);
1665 }
1666 
1667