xref: /netbsd-src/dist/pf/sbin/pfctl/pfctl_optimize.c (revision 23c8222edbfb0f0932d88a8351d3a0cf817dfb9e)
1 /*	$NetBSD: pfctl_optimize.c,v 1.2 2004/11/14 11:26:48 yamt Exp $	*/
2 /*	$OpenBSD: pfctl_optimize.c,v 1.2 2004/08/08 19:04:25 deraadt Exp $ */
3 
4 /*
5  * Copyright (c) 2004 Mike Frantzen <frantzen@openbsd.org>
6  *
7  * Permission to use, copy, modify, and distribute this software for any
8  * purpose with or without fee is hereby granted, provided that the above
9  * copyright notice and this permission notice appear in all copies.
10  *
11  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
12  * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
13  * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
14  * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
15  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
16  * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
17  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
18  */
19 
20 #include <sys/types.h>
21 #include <sys/ioctl.h>
22 #include <sys/socket.h>
23 
24 #include <net/if.h>
25 #include <netinet/in.h>
26 #include <net/pfvar.h>
27 
28 #include <arpa/inet.h>
29 
30 #include <assert.h>
31 #include <ctype.h>
32 #include <err.h>
33 #include <errno.h>
34 #include <stddef.h>
35 #include <stdio.h>
36 #include <stdlib.h>
37 #include <string.h>
38 
39 #include "pfctl_parser.h"
40 #include "pfctl.h"
41 
42 /* The size at which a table becomes faster than individual rules */
43 #define TABLE_THRESHOLD		6
44 
45 
46 /* #define OPT_DEBUG	1 */
47 #ifdef OPT_DEBUG
48 # define DEBUG(str, v...) \
49 	printf("%s: " str "\n", __FUNCTION__ , ## v)
50 #else
51 # define DEBUG(str, v...) ((void)0)
52 #endif
53 
54 
55 /*
56  * A container that lets us sort a superblock to optimize the skip step jumps
57  */
58 struct pf_skip_step {
59 	int				ps_count;	/* number of items */
60 	TAILQ_HEAD( , pf_opt_rule)	ps_rules;
61 	TAILQ_ENTRY(pf_skip_step)	ps_entry;
62 };
63 
64 
65 /*
66  * A superblock is a block of adjacent rules of similar action.  If there
67  * are five PASS rules in a row, they all become members of a superblock.
68  * Once we have a superblock, we are free to re-order any rules within it
69  * in order to improve performance; if a packet is passed, it doesn't matter
70  * who passed it.
71  */
72 struct superblock {
73 	TAILQ_HEAD( , pf_opt_rule)		 sb_rules;
74 	TAILQ_ENTRY(superblock)			 sb_entry;
75 	struct superblock			*sb_profiled_block;
76 	TAILQ_HEAD(skiplist, pf_skip_step)	 sb_skipsteps[PF_SKIP_COUNT];
77 };
78 TAILQ_HEAD(superblocks, superblock);
79 
80 
81 /*
82  * Description of the PF rule structure.
83  */
84 enum {
85     BARRIER,	/* the presence of the field puts the rule in it's own block */
86     BREAK,	/* the field may not differ between rules in a superblock */
87     NOMERGE,	/* the field may not differ between rules when combined */
88     COMBINED,	/* the field may itself be combined with other rules */
89     DC,		/* we just don't care about the field */
90     NEVER};	/* we should never see this field set?!? */
91 struct pf_rule_field {
92 	const char	*prf_name;
93 	int		 prf_type;
94 	size_t		 prf_offset;
95 	size_t		 prf_size;
96 } pf_rule_desc[] = {
97 #define PF_RULE_FIELD(field, ty)	\
98     {#field,				\
99     ty,					\
100     offsetof(struct pf_rule, field),	\
101     sizeof(((struct pf_rule *)0)->field)}
102 
103 
104     /*
105      * The presence of these fields in a rule put the rule in it's own
106      * superblock.  Thus it will not be optimized.  It also prevents the
107      * rule from being re-ordered at all.
108      */
109     PF_RULE_FIELD(label,		BARRIER),
110     PF_RULE_FIELD(prob,			BARRIER),
111     PF_RULE_FIELD(max_states,		BARRIER),
112     PF_RULE_FIELD(max_src_nodes,	BARRIER),
113 
114     /*
115      * These fields must be the same between all rules in the same superblock.
116      * These rules are allowed to be re-ordered but only among like rules.
117      * For instance we can re-order all 'tag "foo"' rules because they have the
118      * same tag.  But we can not re-order between a 'tag "foo"' and a
119      * 'tag "bar"' since that would change the meaning of the ruleset.
120      */
121     PF_RULE_FIELD(tagname,		BREAK),
122     PF_RULE_FIELD(keep_state,		BREAK),
123     PF_RULE_FIELD(qname,		BREAK),
124     PF_RULE_FIELD(rt,			BREAK),
125     PF_RULE_FIELD(allow_opts,		BREAK),
126     PF_RULE_FIELD(rule_flag,		BREAK),
127     PF_RULE_FIELD(action,		BREAK),
128 
129     /*
130      * Any fields not listed in this structure act as BREAK fields
131      */
132 
133 
134     /*
135      * These fields must not differ when we merge two rules together but
136      * their difference isn't enough to put the rules in different superblocks.
137      * There are no problems re-ordering any rules with these fields.
138      */
139     PF_RULE_FIELD(af,			NOMERGE),
140     PF_RULE_FIELD(ifnot,		NOMERGE),
141     PF_RULE_FIELD(ifname,		NOMERGE),
142     PF_RULE_FIELD(match_tag_not,	NOMERGE),
143     PF_RULE_FIELD(match_tagname,	NOMERGE),
144     PF_RULE_FIELD(os_fingerprint,	NOMERGE),
145     PF_RULE_FIELD(timeout,		NOMERGE),
146     PF_RULE_FIELD(return_icmp,		NOMERGE),
147     PF_RULE_FIELD(return_icmp6,		NOMERGE),
148     PF_RULE_FIELD(uid,			NOMERGE),
149     PF_RULE_FIELD(gid,			NOMERGE),
150     PF_RULE_FIELD(direction,		NOMERGE),
151     PF_RULE_FIELD(proto,		NOMERGE),
152     PF_RULE_FIELD(type,			NOMERGE),
153     PF_RULE_FIELD(code,			NOMERGE),
154     PF_RULE_FIELD(flags,		NOMERGE),
155     PF_RULE_FIELD(flagset,		NOMERGE),
156     PF_RULE_FIELD(tos,			NOMERGE),
157     PF_RULE_FIELD(src.port,		NOMERGE),
158     PF_RULE_FIELD(dst.port,		NOMERGE),
159     PF_RULE_FIELD(src.port_op,		NOMERGE),
160     PF_RULE_FIELD(dst.port_op,		NOMERGE),
161     PF_RULE_FIELD(src.neg,		NOMERGE),
162     PF_RULE_FIELD(dst.neg,		NOMERGE),
163 
164     /* These fields can be merged */
165     PF_RULE_FIELD(src.addr,		COMBINED),
166     PF_RULE_FIELD(dst.addr,		COMBINED),
167 
168     /* We just don't care about these fields.  They're set by the kernel */
169     PF_RULE_FIELD(skip,			DC),
170     PF_RULE_FIELD(evaluations,		DC),
171     PF_RULE_FIELD(packets,		DC),
172     PF_RULE_FIELD(bytes,		DC),
173     PF_RULE_FIELD(kif,			DC),
174     PF_RULE_FIELD(anchor,		DC),
175     PF_RULE_FIELD(states,		DC),
176     PF_RULE_FIELD(src_nodes,		DC),
177     PF_RULE_FIELD(nr,			DC),
178     PF_RULE_FIELD(entries,		DC),
179     PF_RULE_FIELD(qid,			DC),
180     PF_RULE_FIELD(pqid,			DC),
181     PF_RULE_FIELD(anchor_relative,	DC),
182     PF_RULE_FIELD(anchor_wildcard,	DC),
183 
184     /* These fields should never be set in a PASS/BLOCK rule */
185     PF_RULE_FIELD(natpass,		NEVER),
186     PF_RULE_FIELD(max_mss,		NEVER),
187     PF_RULE_FIELD(min_ttl,		NEVER),
188 };
189 
190 
191 
192 int	add_opt_table(struct pfctl *, struct pf_opt_tbl **, sa_family_t,
193 	    struct pf_rule_addr *);
194 int	addrs_combineable(struct pf_rule_addr *, struct pf_rule_addr *);
195 int	addrs_equal(struct pf_rule_addr *, struct pf_rule_addr *);
196 int	block_feedback(struct pfctl *, struct superblock *);
197 int	combine_rules(struct pfctl *, struct superblock *);
198 void	comparable_rule(struct pf_rule *, const struct pf_rule *, int);
199 int	construct_superblocks(struct pfctl *, struct pf_opt_queue *,
200 	    struct superblocks *);
201 void	exclude_supersets(struct pf_rule *, struct pf_rule *);
202 int	load_feedback_profile(struct pfctl *, struct superblocks *);
203 int	optimize_superblock(struct pfctl *, struct superblock *);
204 int	pf_opt_create_table(struct pfctl *, struct pf_opt_tbl *);
205 void	remove_from_skipsteps(struct skiplist *, struct superblock *,
206 	    struct pf_opt_rule *, struct pf_skip_step *);
207 int	remove_identical_rules(struct pfctl *, struct superblock *);
208 int	reorder_rules(struct pfctl *, struct superblock *, int);
209 int	rules_combineable(struct pf_rule *, struct pf_rule *);
210 void	skip_append(struct superblock *, int, struct pf_skip_step *,
211 	    struct pf_opt_rule *);
212 int	skip_compare(int, struct pf_skip_step *, struct pf_opt_rule *);
213 void	skip_init(void);
214 int	skip_cmp_af(struct pf_rule *, struct pf_rule *);
215 int	skip_cmp_dir(struct pf_rule *, struct pf_rule *);
216 int	skip_cmp_dst_addr(struct pf_rule *, struct pf_rule *);
217 int	skip_cmp_dst_port(struct pf_rule *, struct pf_rule *);
218 int	skip_cmp_ifp(struct pf_rule *, struct pf_rule *);
219 int	skip_cmp_proto(struct pf_rule *, struct pf_rule *);
220 int	skip_cmp_src_addr(struct pf_rule *, struct pf_rule *);
221 int	skip_cmp_src_port(struct pf_rule *, struct pf_rule *);
222 int	superblock_inclusive(struct superblock *, struct pf_opt_rule *);
223 void	superblock_free(struct pfctl *, struct superblock *);
224 
225 
226 int (*skip_comparitors[PF_SKIP_COUNT])(struct pf_rule *, struct pf_rule *);
227 const char *skip_comparitors_names[PF_SKIP_COUNT];
228 #define PF_SKIP_COMPARITORS {				\
229     { "ifp", PF_SKIP_IFP, skip_cmp_ifp },		\
230     { "dir", PF_SKIP_DIR, skip_cmp_dir },		\
231     { "af", PF_SKIP_AF, skip_cmp_af },			\
232     { "proto", PF_SKIP_PROTO, skip_cmp_proto },		\
233     { "saddr", PF_SKIP_SRC_ADDR, skip_cmp_src_addr },	\
234     { "sport", PF_SKIP_SRC_PORT, skip_cmp_src_port },	\
235     { "daddr", PF_SKIP_DST_ADDR, skip_cmp_dst_addr },	\
236     { "dport", PF_SKIP_DST_PORT, skip_cmp_dst_port }	\
237 }
238 
239 struct pfr_buffer table_buffer;
240 int table_identifier;
241 
242 
243 int
244 pfctl_optimize_rules(struct pfctl *pf)
245 {
246 	struct superblocks superblocks;
247 	struct superblock *block;
248 	struct pf_opt_rule *por;
249 	int nr;
250 
251 	DEBUG("optimizing ruleset");
252 	memset(&table_buffer, 0, sizeof(table_buffer));
253 	skip_init();
254 
255 	if (TAILQ_FIRST(&pf->opt_queue))
256 		nr = TAILQ_FIRST(&pf->opt_queue)->por_rule.nr;
257 
258 	TAILQ_INIT(&superblocks);
259 	if (construct_superblocks(pf, &pf->opt_queue, &superblocks))
260 		goto error;
261 
262 	if (pf->opts & PF_OPT_OPTIMIZE_PROFILE) {
263 		if (load_feedback_profile(pf, &superblocks))
264 			goto error;
265 	}
266 
267 	TAILQ_FOREACH(block, &superblocks, sb_entry) {
268 		if (optimize_superblock(pf, block))
269 			goto error;
270 	}
271 
272 
273 	/*
274 	 * Optimizations are done so we turn off the optimization flag and
275 	 * put the rules right back into the regular codepath.
276 	 */
277 	pf->opts &= ~PF_OPT_OPTIMIZE;
278 
279 	while ((block = TAILQ_FIRST(&superblocks))) {
280 		TAILQ_REMOVE(&superblocks, block, sb_entry);
281 
282 		while ((por = TAILQ_FIRST(&block->sb_rules))) {
283 			TAILQ_REMOVE(&block->sb_rules, por, por_entry);
284 			por->por_rule.nr = nr++;
285 			if (pfctl_add_rule(pf, &por->por_rule,
286 			    por->por_anchor)) {
287 				free(por);
288 				goto error;
289 			}
290 			free(por);
291 		}
292 		free(block);
293 	}
294 
295 	return (0);
296 
297 error:
298 	while ((por = TAILQ_FIRST(&pf->opt_queue))) {
299 		TAILQ_REMOVE(&pf->opt_queue, por, por_entry);
300 		if (por->por_src_tbl) {
301 			pfr_buf_clear(por->por_src_tbl->pt_buf);
302 			free(por->por_src_tbl->pt_buf);
303 			free(por->por_src_tbl);
304 		}
305 		if (por->por_dst_tbl) {
306 			pfr_buf_clear(por->por_dst_tbl->pt_buf);
307 			free(por->por_dst_tbl->pt_buf);
308 			free(por->por_dst_tbl);
309 		}
310 		free(por);
311 	}
312 	while ((block = TAILQ_FIRST(&superblocks))) {
313 		TAILQ_REMOVE(&superblocks, block, sb_entry);
314 		superblock_free(pf, block);
315 	}
316 	return (1);
317 }
318 
319 
320 /*
321  * Go ahead and optimize a superblock
322  */
323 int
324 optimize_superblock(struct pfctl *pf, struct superblock *block)
325 {
326 #ifdef OPT_DEBUG
327 	struct pf_opt_rule *por;
328 #endif /* OPT_DEBUG */
329 
330 	/* We have a few optimization passes:
331 	 *   1) remove duplicate rules or rules that are a subset of other
332 	 *      rules
333 	 *   2) combine otherwise identical rules with different IP addresses
334 	 *      into a single rule and put the addresses in a table.
335 	 *   3) re-order the rules to improve kernel skip steps
336 	 *   4) re-order the 'quick' rules based on feedback from the
337 	 *      active ruleset statistics
338 	 *
339 	 * XXX combine_rules() doesn't combine v4 and v6 rules.  would just
340 	 *     have to keep af in the table container, make af 'COMBINE' and
341 	 *     twiddle the af on the merged rule
342 	 * XXX maybe add a weighting to the metric on skipsteps when doing
343 	 *     reordering.  sometimes two sequential tables will be better
344 	 *     that four consecutive interfaces.
345 	 * XXX need to adjust the skipstep count of everything after PROTO,
346 	 *     since they aren't actually checked on a proto mismatch in
347 	 *     pf_test_{tcp, udp, icmp}()
348 	 * XXX should i treat proto=0, af=0 or dir=0 special in skepstep
349 	 *     calculation since they are a DC?
350 	 * XXX keep last skiplist of last superblock to influence this
351 	 *     superblock.  '5 inet6 log' should make '3 inet6' come before '4
352 	 *     inet' in the next superblock.
353 	 * XXX would be useful to add tables for ports
354 	 * XXX we can also re-order some mutually exclusive superblocks to
355 	 *     try merging superblocks before any of these optimization passes.
356 	 *     for instance a single 'log in' rule in the middle of non-logging
357 	 *     out rules.
358 	 */
359 
360 	/* shortcut.  there will be alot of 1-rule superblocks */
361 	if (!TAILQ_NEXT(TAILQ_FIRST(&block->sb_rules), por_entry))
362 		return (0);
363 
364 #ifdef OPT_DEBUG
365 	printf("--- Superblock ---\n");
366 	TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
367 		printf("  ");
368 		print_rule(&por->por_rule, por->por_anchor, 1);
369 	}
370 #endif /* OPT_DEBUG */
371 
372 
373 	if (remove_identical_rules(pf, block))
374 		return (1);
375 	if (combine_rules(pf, block))
376 		return (1);
377 	if ((pf->opts & PF_OPT_OPTIMIZE_PROFILE) &&
378 	    TAILQ_FIRST(&block->sb_rules)->por_rule.quick &&
379 	    block->sb_profiled_block) {
380 		if (block_feedback(pf, block))
381 			return (1);
382 	} else if (reorder_rules(pf, block, 0)) {
383 		return (1);
384 	}
385 
386 	/*
387 	 * Don't add any optimization passes below reorder_rules().  It will
388 	 * have divided superblocks into smaller blocks for further refinement
389 	 * and doesn't put them back together again.  What once was a true
390 	 * superblock might have been split into multiple superblocks.
391 	 */
392 
393 #ifdef OPT_DEBUG
394 	printf("--- END Superblock ---\n");
395 #endif /* OPT_DEBUG */
396 	return (0);
397 }
398 
399 
400 /*
401  * Optimization pass #1: remove identical rules
402  */
403 int
404 remove_identical_rules(struct pfctl *pf, struct superblock *block)
405 {
406 	struct pf_opt_rule *por1, *por2, *por_next, *por2_next;
407 	struct pf_rule a, a2, b, b2;
408 
409 	for (por1 = TAILQ_FIRST(&block->sb_rules); por1; por1 = por_next) {
410 		por_next = TAILQ_NEXT(por1, por_entry);
411 		for (por2 = por_next; por2; por2 = por2_next) {
412 			por2_next = TAILQ_NEXT(por2, por_entry);
413 			comparable_rule(&a, &por1->por_rule, DC);
414 			comparable_rule(&b, &por2->por_rule, DC);
415 			memcpy(&a2, &a, sizeof(a2));
416 			memcpy(&b2, &b, sizeof(b2));
417 
418 			exclude_supersets(&a, &b);
419 			exclude_supersets(&b2, &a2);
420 			if (memcmp(&a, &b, sizeof(a)) == 0) {
421 				DEBUG("removing identical rule  nr%d = *nr%d*",
422 				    por1->por_rule.nr, por2->por_rule.nr);
423 				TAILQ_REMOVE(&block->sb_rules, por2, por_entry);
424 				if (por_next == por2)
425 					por_next = TAILQ_NEXT(por1, por_entry);
426 				free(por2);
427 			} else if (memcmp(&a2, &b2, sizeof(a2)) == 0) {
428 				DEBUG("removing identical rule  *nr%d* = nr%d",
429 				    por1->por_rule.nr, por2->por_rule.nr);
430 				TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
431 				free(por1);
432 				break;
433 			}
434 		}
435 	}
436 
437 	return (0);
438 }
439 
440 
441 /*
442  * Optimization pass #2: combine similar rules with different addresses
443  * into a single rule and a table
444  */
445 int
446 combine_rules(struct pfctl *pf, struct superblock *block)
447 {
448 	struct pf_opt_rule *p1, *p2, *por_next;
449 	int src_eq, dst_eq;
450 
451 	if ((pf->loadopt & PFCTL_FLAG_TABLE) == 0) {
452 		warnx("Must enable table loading for optimizations");
453 		return (1);
454 	}
455 
456 	/* First we make a pass to combine the rules.  O(n log n) */
457 	TAILQ_FOREACH(p1, &block->sb_rules, por_entry) {
458 		for (p2 = TAILQ_NEXT(p1, por_entry); p2; p2 = por_next) {
459 			por_next = TAILQ_NEXT(p2, por_entry);
460 
461 			src_eq = addrs_equal(&p1->por_rule.src,
462 			    &p2->por_rule.src);
463 			dst_eq = addrs_equal(&p1->por_rule.dst,
464 			    &p2->por_rule.dst);
465 
466 			if (src_eq && !dst_eq && p1->por_src_tbl == NULL &&
467 			    p2->por_dst_tbl == NULL &&
468 			    rules_combineable(&p1->por_rule, &p2->por_rule) &&
469 			    addrs_combineable(&p1->por_rule.dst,
470 			    &p2->por_rule.dst)) {
471 				DEBUG("can combine rules  nr%d = nr%d",
472 				    p1->por_rule.nr, p2->por_rule.nr);
473 				if (p1->por_dst_tbl == NULL &&
474 				    add_opt_table(pf, &p1->por_dst_tbl,
475 				    p1->por_rule.af, &p1->por_rule.dst))
476 					return (1);
477 				if (add_opt_table(pf, &p1->por_dst_tbl,
478 				    p1->por_rule.af, &p2->por_rule.dst))
479 					return (1);
480 				p2->por_dst_tbl = p1->por_dst_tbl;
481 				if (p1->por_dst_tbl->pt_rulecount >=
482 				    TABLE_THRESHOLD) {
483 					TAILQ_REMOVE(&block->sb_rules, p2,
484 					    por_entry);
485 					free(p2);
486 				}
487 			} else if (!src_eq && dst_eq && p1->por_dst_tbl == NULL
488 			    && p2->por_src_tbl == NULL &&
489 			    rules_combineable(&p1->por_rule, &p2->por_rule) &&
490 			    addrs_combineable(&p1->por_rule.src,
491 			    &p2->por_rule.src)) {
492 				DEBUG("can combine rules  nr%d = nr%d",
493 				    p1->por_rule.nr, p2->por_rule.nr);
494 				if (p1->por_src_tbl == NULL &&
495 				    add_opt_table(pf, &p1->por_src_tbl,
496 				    p1->por_rule.af, &p1->por_rule.src))
497 					return (1);
498 				if (add_opt_table(pf, &p1->por_src_tbl,
499 				    p1->por_rule.af, &p2->por_rule.src))
500 					return (1);
501 				p2->por_src_tbl = p1->por_src_tbl;
502 				if (p1->por_src_tbl->pt_rulecount >=
503 				    TABLE_THRESHOLD) {
504 					TAILQ_REMOVE(&block->sb_rules, p2,
505 					    por_entry);
506 					free(p2);
507 				}
508 			}
509 		}
510 	}
511 
512 
513 	/*
514 	 * Then we make a final pass to create a valid table name and
515 	 * insert the name into the rules.
516 	 */
517 	for (p1 = TAILQ_FIRST(&block->sb_rules); p1; p1 = por_next) {
518 		por_next = TAILQ_NEXT(p1, por_entry);
519 		assert(p1->por_src_tbl == NULL || p1->por_dst_tbl == NULL);
520 
521 		if (p1->por_src_tbl && p1->por_src_tbl->pt_rulecount >=
522 		    TABLE_THRESHOLD) {
523 			if (p1->por_src_tbl->pt_generated) {
524 				/* This rule is included in a table */
525 				TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
526 				free(p1);
527 				continue;
528 			}
529 			p1->por_src_tbl->pt_generated = 1;
530 
531 			if ((pf->opts & PF_OPT_NOACTION) == 0 &&
532 			    pf_opt_create_table(pf, p1->por_src_tbl))
533 				return (1);
534 
535 			pf->tdirty = 1;
536 
537 			if (pf->opts & PF_OPT_VERBOSE)
538 				print_tabledef(p1->por_src_tbl->pt_name,
539 				    PFR_TFLAG_CONST, 1,
540 				    &p1->por_src_tbl->pt_nodes);
541 
542 			memset(&p1->por_rule.src.addr, 0,
543 			    sizeof(p1->por_rule.src.addr));
544 			p1->por_rule.src.addr.type = PF_ADDR_TABLE;
545 			strlcpy(p1->por_rule.src.addr.v.tblname,
546 			    p1->por_src_tbl->pt_name,
547 			    sizeof(p1->por_rule.src.addr.v.tblname));
548 
549 			pfr_buf_clear(p1->por_src_tbl->pt_buf);
550 			free(p1->por_src_tbl->pt_buf);
551 			p1->por_src_tbl->pt_buf = NULL;
552 		}
553 		if (p1->por_dst_tbl && p1->por_dst_tbl->pt_rulecount >=
554 		    TABLE_THRESHOLD) {
555 			if (p1->por_dst_tbl->pt_generated) {
556 				/* This rule is included in a table */
557 				TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
558 				free(p1);
559 				continue;
560 			}
561 			p1->por_dst_tbl->pt_generated = 1;
562 
563 			if ((pf->opts & PF_OPT_NOACTION) == 0 &&
564 			    pf_opt_create_table(pf, p1->por_dst_tbl))
565 				return (1);
566 			pf->tdirty = 1;
567 
568 			if (pf->opts & PF_OPT_VERBOSE)
569 				print_tabledef(p1->por_dst_tbl->pt_name,
570 				    PFR_TFLAG_CONST, 1,
571 				    &p1->por_dst_tbl->pt_nodes);
572 
573 			memset(&p1->por_rule.dst.addr, 0,
574 			    sizeof(p1->por_rule.dst.addr));
575 			p1->por_rule.dst.addr.type = PF_ADDR_TABLE;
576 			strlcpy(p1->por_rule.dst.addr.v.tblname,
577 			    p1->por_dst_tbl->pt_name,
578 			    sizeof(p1->por_rule.dst.addr.v.tblname));
579 
580 			pfr_buf_clear(p1->por_dst_tbl->pt_buf);
581 			free(p1->por_dst_tbl->pt_buf);
582 			p1->por_dst_tbl->pt_buf = NULL;
583 		}
584 	}
585 
586 	return (0);
587 }
588 
589 
590 /*
591  * Optimization pass #3: re-order rules to improve skip steps
592  */
593 int
594 reorder_rules(struct pfctl *pf, struct superblock *block, int depth)
595 {
596 	struct superblock *newblock;
597 	struct pf_skip_step *skiplist;
598 	struct pf_opt_rule *por;
599 	int i, largest, largest_list, rule_count = 0;
600 	TAILQ_HEAD( , pf_opt_rule) head;
601 
602 	/*
603 	 * Calculate the best-case skip steps.  We put each rule in a list
604 	 * of other rules with common fields
605 	 */
606 	for (i = 0; i < PF_SKIP_COUNT; i++) {
607 		TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
608 			TAILQ_FOREACH(skiplist, &block->sb_skipsteps[i],
609 			    ps_entry) {
610 				if (skip_compare(i, skiplist, por) == 0)
611 					break;
612 			}
613 			if (skiplist == NULL) {
614 				if ((skiplist = calloc(1, sizeof(*skiplist))) ==
615 				    NULL)
616 					err(1, "calloc");
617 				TAILQ_INIT(&skiplist->ps_rules);
618 				TAILQ_INSERT_TAIL(&block->sb_skipsteps[i],
619 				    skiplist, ps_entry);
620 			}
621 			skip_append(block, i, skiplist, por);
622 		}
623 	}
624 
625 	TAILQ_FOREACH(por, &block->sb_rules, por_entry)
626 		rule_count++;
627 
628 	/*
629 	 * Now we're going to ignore any fields that are identical between
630 	 * all of the rules in the superblock and those fields which differ
631 	 * between every rule in the superblock.
632 	 */
633 	largest = 0;
634 	for (i = 0; i < PF_SKIP_COUNT; i++) {
635 		skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
636 		if (skiplist->ps_count == rule_count) {
637 			DEBUG("(%d) original skipstep '%s' is all rules",
638 			    depth, skip_comparitors_names[i]);
639 			skiplist->ps_count = 0;
640 		} else if (skiplist->ps_count == 1) {
641 			skiplist->ps_count = 0;
642 		} else {
643 			DEBUG("(%d) original skipstep '%s' largest jump is %d",
644 			    depth, skip_comparitors_names[i],
645 			    skiplist->ps_count);
646 			if (skiplist->ps_count > largest)
647 				largest = skiplist->ps_count;
648 		}
649 	}
650 	if (largest == 0) {
651 		/* Ugh.  There is NO commonality in the superblock on which
652 		 * optimize the skipsteps optimization.
653 		 */
654 		goto done;
655 	}
656 
657 	/*
658 	 * Now we're going to empty the superblock rule list and re-create
659 	 * it based on a more optimal skipstep order.
660 	 */
661 	TAILQ_INIT(&head);
662 	while ((por = TAILQ_FIRST(&block->sb_rules))) {
663 		TAILQ_REMOVE(&block->sb_rules, por, por_entry);
664 		TAILQ_INSERT_TAIL(&head, por, por_entry);
665 	}
666 
667 
668 	while (!TAILQ_EMPTY(&head)) {
669 		largest = 1;
670 
671 		/*
672 		 * Find the most useful skip steps remaining
673 		 */
674 		for (i = 0; i < PF_SKIP_COUNT; i++) {
675 			skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
676 			if (skiplist->ps_count > largest) {
677 				largest = skiplist->ps_count;
678 				largest_list = i;
679 			}
680 		}
681 
682 		if (largest <= 1) {
683 			/*
684 			 * Nothing useful left.  Leave remaining rules in order.
685 			 */
686 			DEBUG("(%d) no more commonality for skip steps", depth);
687 			while ((por = TAILQ_FIRST(&head))) {
688 				TAILQ_REMOVE(&head, por, por_entry);
689 				TAILQ_INSERT_TAIL(&block->sb_rules, por,
690 				    por_entry);
691 			}
692 		} else {
693 			/*
694 			 * There is commonality.  Extract those common rules
695 			 * and place them in the ruleset adjacent to each
696 			 * other.
697 			 */
698 			skiplist = TAILQ_FIRST(&block->sb_skipsteps[
699 			    largest_list]);
700 			DEBUG("(%d) skipstep '%s' largest jump is %d @ #%d",
701 			    depth, skip_comparitors_names[largest_list],
702 			    largest, TAILQ_FIRST(&TAILQ_FIRST(&block->
703 			    sb_skipsteps [largest_list])->ps_rules)->
704 			    por_rule.nr);
705 			TAILQ_REMOVE(&block->sb_skipsteps[largest_list],
706 			    skiplist, ps_entry);
707 
708 
709 			/*
710 			 * There may be further commonality inside these
711 			 * rules.  So we'll split them off into they're own
712 			 * superblock and pass it back into the optimizer.
713 			 */
714 			if (skiplist->ps_count > 2) {
715 				if ((newblock = calloc(1, sizeof(*newblock)))
716 				    == NULL) {
717 					warn("calloc");
718 					return (1);
719 				}
720 				TAILQ_INIT(&newblock->sb_rules);
721 				for (i = 0; i < PF_SKIP_COUNT; i++)
722 					TAILQ_INIT(&newblock->sb_skipsteps[i]);
723 				TAILQ_INSERT_BEFORE(block, newblock, sb_entry);
724 				DEBUG("(%d) splitting off %d rules from superblock @ #%d",
725 				    depth, skiplist->ps_count,
726 				    TAILQ_FIRST(&skiplist->ps_rules)->
727 				    por_rule.nr);
728 			} else {
729 				newblock = block;
730 			}
731 
732 			while ((por = TAILQ_FIRST(&skiplist->ps_rules))) {
733 				TAILQ_REMOVE(&head, por, por_entry);
734 				TAILQ_REMOVE(&skiplist->ps_rules, por,
735 				    por_skip_entry[largest_list]);
736 				TAILQ_INSERT_TAIL(&newblock->sb_rules, por,
737 				    por_entry);
738 
739 				/* Remove this rule from all other skiplists */
740 				remove_from_skipsteps(&block->sb_skipsteps[
741 				    largest_list], block, por, skiplist);
742 			}
743 			free(skiplist);
744 			if (newblock != block)
745 				if (reorder_rules(pf, newblock, depth + 1))
746 					return (1);
747 		}
748 	}
749 
750 done:
751 	for (i = 0; i < PF_SKIP_COUNT; i++) {
752 		while ((skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]))) {
753 			TAILQ_REMOVE(&block->sb_skipsteps[i], skiplist,
754 			    ps_entry);
755 			free(skiplist);
756 		}
757 	}
758 
759 	return (0);
760 }
761 
762 
763 /*
764  * Optimization pass #4: re-order 'quick' rules based on feedback from the
765  * currently running ruleset
766  */
767 int
768 block_feedback(struct pfctl *pf, struct superblock *block)
769 {
770 	TAILQ_HEAD( , pf_opt_rule) queue;
771 	struct pf_opt_rule *por1, *por2;
772 	u_int64_t total_count = 0;
773 	struct pf_rule a, b;
774 
775 
776 	/*
777 	 * Walk through all of the profiled superblock's rules and copy
778 	 * the counters onto our rules.
779 	 */
780 	TAILQ_FOREACH(por1, &block->sb_profiled_block->sb_rules, por_entry) {
781 		comparable_rule(&a, &por1->por_rule, DC);
782 		total_count += por1->por_rule.packets;
783 		TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
784 			if (por2->por_profile_count)
785 				continue;
786 			comparable_rule(&b, &por2->por_rule, DC);
787 			if (memcmp(&a, &b, sizeof(a)) == 0) {
788 				por2->por_profile_count =
789 				    por1->por_rule.packets;
790 				break;
791 			}
792 		}
793 	}
794 	superblock_free(pf, block->sb_profiled_block);
795 	block->sb_profiled_block = NULL;
796 
797 	/*
798 	 * Now we pull all of the rules off the superblock and re-insert them
799 	 * in sorted order.
800 	 */
801 
802 	TAILQ_INIT(&queue);
803 	while ((por1 = TAILQ_FIRST(&block->sb_rules)) != NULL) {
804 		TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
805 		TAILQ_INSERT_TAIL(&queue, por1, por_entry);
806 	}
807 
808 	while ((por1 = TAILQ_FIRST(&queue)) != NULL) {
809 		TAILQ_REMOVE(&queue, por1, por_entry);
810 /* XXX I should sort all of the unused rules based on skip steps */
811 		TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
812 			if (por1->por_profile_count > por2->por_profile_count) {
813 				TAILQ_INSERT_BEFORE(por2, por1, por_entry);
814 				break;
815 			}
816 		}
817 		if (por2 == TAILQ_END(&block->sb_rules))
818 			TAILQ_INSERT_TAIL(&block->sb_rules, por1, por_entry);
819 	}
820 
821 	return (0);
822 }
823 
824 
825 /*
826  * Load the current ruleset from the kernel and try to associate them with
827  * the ruleset we're optimizing.
828  */
829 int
830 load_feedback_profile(struct pfctl *pf, struct superblocks *superblocks)
831 {
832 	struct superblock *block, *blockcur;
833 	struct superblocks prof_superblocks;
834 	struct pf_opt_rule *por;
835 	struct pf_opt_queue queue;
836 	struct pfioc_rule pr;
837 	struct pf_rule a, b;
838 	int nr, mnr;
839 
840 	TAILQ_INIT(&queue);
841 	TAILQ_INIT(&prof_superblocks);
842 
843 	memset(&pr, 0, sizeof(pr));
844 	pr.rule.action = PF_PASS;
845 	if (ioctl(pf->dev, DIOCGETRULES, &pr)) {
846 		warn("DIOCGETRULES");
847 		return (1);
848 	}
849 	mnr = pr.nr;
850 
851 	DEBUG("Loading %d active rules for a feedback profile", mnr);
852 	for (nr = 0; nr < mnr; ++nr) {
853 		if ((por = calloc(1, sizeof(*por))) == NULL) {
854 			warn("calloc");
855 			return (1);
856 		}
857 		pr.nr = nr;
858 		if (ioctl(pf->dev, DIOCGETRULE, &pr)) {
859 			warn("DIOCGETRULES");
860 			return (1);
861 		}
862 		memcpy(&por->por_rule, &pr.rule, sizeof(por->por_rule));
863 		strlcpy(por->por_anchor, pr.anchor_call,
864 		    sizeof(por->por_anchor));
865 		if (TAILQ_EMPTY(&por->por_rule.rpool.list))
866 			memset(&por->por_rule.rpool, 0,
867 			    sizeof(por->por_rule.rpool));
868 		TAILQ_INSERT_TAIL(&queue, por, por_entry);
869 
870 		/* XXX pfctl_get_pool(pf->dev, &pr.rule.rpool, nr, pr.ticket,
871 		 *         PF_PASS, pf->anchor) ???
872 		 * ... pfctl_clear_pool(&pr.rule.rpool)
873 		 */
874 	}
875 
876 	if (construct_superblocks(pf, &queue, &prof_superblocks))
877 		return (1);
878 
879 
880 	/*
881 	 * Now we try to associate the active ruleset's superblocks with
882 	 * the superblocks we're compiling.
883 	 */
884 	block = TAILQ_FIRST(superblocks);
885 	blockcur = TAILQ_FIRST(&prof_superblocks);
886 	while (block && blockcur) {
887 		comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule,
888 		    BREAK);
889 		comparable_rule(&b, &TAILQ_FIRST(&blockcur->sb_rules)->por_rule,
890 		    BREAK);
891 		if (memcmp(&a, &b, sizeof(a)) == 0) {
892 			/* The two superblocks lined up */
893 			block->sb_profiled_block = blockcur;
894 		} else {
895 			DEBUG("superblocks don't line up between #%d and #%d",
896 			    TAILQ_FIRST(&block->sb_rules)->por_rule.nr,
897 			    TAILQ_FIRST(&blockcur->sb_rules)->por_rule.nr);
898 			break;
899 		}
900 		block = TAILQ_NEXT(block, sb_entry);
901 		blockcur = TAILQ_NEXT(blockcur, sb_entry);
902 	}
903 
904 
905 
906 	/* Free any superblocks we couldn't link */
907 	while (blockcur) {
908 		block = TAILQ_NEXT(blockcur, sb_entry);
909 		superblock_free(pf, blockcur);
910 		blockcur = block;
911 	}
912 	return (0);
913 }
914 
915 
916 /*
917  * Compare a rule to a skiplist to see if the rule is a member
918  */
919 int
920 skip_compare(int skipnum, struct pf_skip_step *skiplist,
921     struct pf_opt_rule *por)
922 {
923 	struct pf_rule *a, *b;
924 	if (skipnum >= PF_SKIP_COUNT || skipnum < 0)
925 		errx(1, "skip_compare() out of bounds");
926 	a = &por->por_rule;
927 	b = &TAILQ_FIRST(&skiplist->ps_rules)->por_rule;
928 
929 	return ((skip_comparitors[skipnum])(a, b));
930 }
931 
932 
933 /*
934  * Add a rule to a skiplist
935  */
936 void
937 skip_append(struct superblock *superblock, int skipnum,
938     struct pf_skip_step *skiplist, struct pf_opt_rule *por)
939 {
940 	struct pf_skip_step *prev;
941 
942 	skiplist->ps_count++;
943 	TAILQ_INSERT_TAIL(&skiplist->ps_rules, por, por_skip_entry[skipnum]);
944 
945 	/* Keep the list of skiplists sorted by whichever is larger */
946 	while ((prev = TAILQ_PREV(skiplist, skiplist, ps_entry)) &&
947 	    prev->ps_count < skiplist->ps_count) {
948 		TAILQ_REMOVE(&superblock->sb_skipsteps[skipnum],
949 		    skiplist, ps_entry);
950 		TAILQ_INSERT_BEFORE(prev, skiplist, ps_entry);
951 	}
952 }
953 
954 
955 /*
956  * Remove a rule from the other skiplist calculations.
957  */
958 void
959 remove_from_skipsteps(struct skiplist *head, struct superblock *block,
960     struct pf_opt_rule *por, struct pf_skip_step *active_list)
961 {
962 	struct pf_skip_step *sk, *next;
963 	struct pf_opt_rule *p2;
964 	int i, found;
965 
966 	for (i = 0; i < PF_SKIP_COUNT; i++) {
967 		sk = TAILQ_FIRST(&block->sb_skipsteps[i]);
968 		if (sk == NULL || sk == active_list || sk->ps_count <= 1)
969 			continue;
970 		found = 0;
971 		do {
972 			TAILQ_FOREACH(p2, &sk->ps_rules, por_skip_entry[i])
973 				if (p2 == por) {
974 					TAILQ_REMOVE(&sk->ps_rules, p2,
975 					    por_skip_entry[i]);
976 					found = 1;
977 					sk->ps_count--;
978 					break;
979 				}
980 		} while (!found && (sk = TAILQ_NEXT(sk, ps_entry)));
981 		if (found && sk) {
982 			/* Does this change the sorting order? */
983 			while ((next = TAILQ_NEXT(sk, ps_entry)) &&
984 			    next->ps_count > sk->ps_count) {
985 				TAILQ_REMOVE(head, sk, ps_entry);
986 				TAILQ_INSERT_AFTER(head, next, sk, ps_entry);
987 			}
988 #ifdef OPT_DEBUG
989 			next = TAILQ_NEXT(sk, ps_entry);
990 			assert(next == NULL || next->ps_count <= sk->ps_count);
991 #endif /* OPT_DEBUG */
992 		}
993 	}
994 }
995 
996 
997 /* Compare two rules AF field for skiplist construction */
998 int
999 skip_cmp_af(struct pf_rule *a, struct pf_rule *b)
1000 {
1001 	if (a->af != b->af || a->af == 0)
1002 		return (1);
1003 	return (0);
1004 }
1005 
1006 /* Compare two rules DIRECTION field for skiplist construction */
1007 int
1008 skip_cmp_dir(struct pf_rule *a, struct pf_rule *b)
1009 {
1010 	if (a->direction == 0 || a->direction != b->direction)
1011 		return (1);
1012 	return (0);
1013 }
1014 
1015 /* Compare two rules DST Address field for skiplist construction */
1016 int
1017 skip_cmp_dst_addr(struct pf_rule *a, struct pf_rule *b)
1018 {
1019 	if (a->dst.neg != b->dst.neg ||
1020 	    a->dst.addr.type != b->dst.addr.type)
1021 		return (1);
1022 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1023 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1024 	 *    a->proto == IPPROTO_ICMP
1025 	 *	return (1);
1026 	 */
1027 	switch (a->dst.addr.type) {
1028 	case PF_ADDR_ADDRMASK:
1029 		if (memcmp(&a->dst.addr.v.a.addr, &b->dst.addr.v.a.addr,
1030 		    sizeof(a->dst.addr.v.a.addr)) ||
1031 		    memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask,
1032 		    sizeof(a->dst.addr.v.a.mask)) ||
1033 		    (a->dst.addr.v.a.addr.addr32[0] == 0 &&
1034 		    a->dst.addr.v.a.addr.addr32[1] == 0 &&
1035 		    a->dst.addr.v.a.addr.addr32[2] == 0 &&
1036 		    a->dst.addr.v.a.addr.addr32[3] == 0))
1037 			return (1);
1038 		return (0);
1039 	case PF_ADDR_DYNIFTL:
1040 		if (strcmp(a->dst.addr.v.ifname, b->dst.addr.v.ifname) != 0 ||
1041 		    a->dst.addr.iflags != a->dst.addr.iflags ||
1042 		    memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask,
1043 		    sizeof(a->dst.addr.v.a.mask)))
1044 			return (1);
1045 		return (0);
1046 	case PF_ADDR_NOROUTE:
1047 		return (0);
1048 	case PF_ADDR_TABLE:
1049 		return (strcmp(a->dst.addr.v.tblname, b->dst.addr.v.tblname));
1050 	}
1051 	return (1);
1052 }
1053 
1054 /* Compare two rules DST port field for skiplist construction */
1055 int
1056 skip_cmp_dst_port(struct pf_rule *a, struct pf_rule *b)
1057 {
1058 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1059 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1060 	 *    a->proto == IPPROTO_ICMP
1061 	 *	return (1);
1062 	 */
1063 	if (a->dst.port_op == PF_OP_NONE || a->dst.port_op != b->dst.port_op ||
1064 	    a->dst.port[0] != b->dst.port[0] ||
1065 	    a->dst.port[1] != b->dst.port[1])
1066 		return (1);
1067 	return (0);
1068 }
1069 
1070 /* Compare two rules IFP field for skiplist construction */
1071 int
1072 skip_cmp_ifp(struct pf_rule *a, struct pf_rule *b)
1073 {
1074 	if (strcmp(a->ifname, b->ifname) || a->ifname[0] == '\0')
1075 		return (1);
1076 	return (a->ifnot != b->ifnot);
1077 }
1078 
1079 /* Compare two rules PROTO field for skiplist construction */
1080 int
1081 skip_cmp_proto(struct pf_rule *a, struct pf_rule *b)
1082 {
1083 	return (a->proto != b->proto || a->proto == 0);
1084 }
1085 
1086 /* Compare two rules SRC addr field for skiplist construction */
1087 int
1088 skip_cmp_src_addr(struct pf_rule *a, struct pf_rule *b)
1089 {
1090 	if (a->src.neg != b->src.neg ||
1091 	    a->src.addr.type != b->src.addr.type)
1092 		return (1);
1093 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1094 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1095 	 *    a->proto == IPPROTO_ICMP
1096 	 *	return (1);
1097 	 */
1098 	switch (a->src.addr.type) {
1099 	case PF_ADDR_ADDRMASK:
1100 		if (memcmp(&a->src.addr.v.a.addr, &b->src.addr.v.a.addr,
1101 		    sizeof(a->src.addr.v.a.addr)) ||
1102 		    memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask,
1103 		    sizeof(a->src.addr.v.a.mask)) ||
1104 		    (a->src.addr.v.a.addr.addr32[0] == 0 &&
1105 		    a->src.addr.v.a.addr.addr32[1] == 0 &&
1106 		    a->src.addr.v.a.addr.addr32[2] == 0 &&
1107 		    a->src.addr.v.a.addr.addr32[3] == 0))
1108 			return (1);
1109 		return (0);
1110 	case PF_ADDR_DYNIFTL:
1111 		if (strcmp(a->src.addr.v.ifname, b->src.addr.v.ifname) != 0 ||
1112 		    a->src.addr.iflags != a->src.addr.iflags ||
1113 		    memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask,
1114 		    sizeof(a->src.addr.v.a.mask)))
1115 			return (1);
1116 		return (0);
1117 	case PF_ADDR_NOROUTE:
1118 		return (0);
1119 	case PF_ADDR_TABLE:
1120 		return (strcmp(a->src.addr.v.tblname, b->src.addr.v.tblname));
1121 	}
1122 	return (1);
1123 }
1124 
1125 /* Compare two rules SRC port field for skiplist construction */
1126 int
1127 skip_cmp_src_port(struct pf_rule *a, struct pf_rule *b)
1128 {
1129 	if (a->src.port_op == PF_OP_NONE || a->src.port_op != b->src.port_op ||
1130 	    a->src.port[0] != b->src.port[0] ||
1131 	    a->src.port[1] != b->src.port[1])
1132 		return (1);
1133 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1134 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1135 	 *    a->proto == IPPROTO_ICMP
1136 	 *	return (1);
1137 	 */
1138 	return (0);
1139 }
1140 
1141 
1142 void
1143 skip_init(void)
1144 {
1145 	struct {
1146 		char *name;
1147 		int skipnum;
1148 		int (*func)(struct pf_rule *, struct pf_rule *);
1149 	} comps[] = PF_SKIP_COMPARITORS;
1150 	int skipnum, i;
1151 
1152 	for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++) {
1153 		for (i = 0; i < sizeof(comps)/sizeof(*comps); i++)
1154 			if (comps[i].skipnum == skipnum) {
1155 				skip_comparitors[skipnum] = comps[i].func;
1156 				skip_comparitors_names[skipnum] = comps[i].name;
1157 			}
1158 	}
1159 	for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++)
1160 		if (skip_comparitors[skipnum] == NULL)
1161 			errx(1, "Need to add skip step comparitor to pfctl?!");
1162 }
1163 
1164 /*
1165  * Add a host/netmask to a table
1166  */
1167 int
1168 add_opt_table(struct pfctl *pf, struct pf_opt_tbl **tbl, sa_family_t af,
1169     struct pf_rule_addr *addr)
1170 {
1171 #ifdef OPT_DEBUG
1172 	char buf[128];
1173 #endif /* OPT_DEBUG */
1174 	static int tablenum = 0;
1175 	struct node_host node_host;
1176 
1177 	if (*tbl == NULL) {
1178 		if ((*tbl = calloc(1, sizeof(**tbl))) == NULL ||
1179 		    ((*tbl)->pt_buf = calloc(1, sizeof(*(*tbl)->pt_buf))) ==
1180 		    NULL)
1181 			err(1, "calloc");
1182 		(*tbl)->pt_buf->pfrb_type = PFRB_ADDRS;
1183 		SIMPLEQ_INIT(&(*tbl)->pt_nodes);
1184 
1185 		/* This is just a temporary table name */
1186 		snprintf((*tbl)->pt_name, sizeof((*tbl)->pt_name), "%s%d",
1187 		    PF_OPT_TABLE_PREFIX, tablenum++);
1188 		DEBUG("creating table <%s>", (*tbl)->pt_name);
1189 	}
1190 
1191 	memset(&node_host, 0, sizeof(node_host));
1192 	node_host.af = af;
1193 	node_host.addr = addr->addr;
1194 
1195 #ifdef OPT_DEBUG
1196 	DEBUG("<%s> adding %s/%d", (*tbl)->pt_name, inet_ntop(af,
1197 	    &node_host.addr.v.a.addr, buf, sizeof(buf)),
1198 	    unmask(&node_host.addr.v.a.mask, af));
1199 #endif /* OPT_DEBUG */
1200 
1201 	if (append_addr_host((*tbl)->pt_buf, &node_host, 0, 0))
1202 		return (1);
1203 	if (pf->opts & PF_OPT_VERBOSE) {
1204 		struct node_tinit *ti;
1205 
1206 		if ((ti = calloc(1, sizeof(*ti))) == NULL)
1207 			err(1, "malloc");
1208 		if ((ti->host = malloc(sizeof(*ti->host))) == NULL)
1209 			err(1, "malloc");
1210 		memcpy(ti->host, &node_host, sizeof(*ti->host));
1211 		SIMPLEQ_INSERT_TAIL(&(*tbl)->pt_nodes, ti, entries);
1212 	}
1213 
1214 	(*tbl)->pt_rulecount++;
1215 	if ((*tbl)->pt_rulecount == TABLE_THRESHOLD)
1216 		DEBUG("table <%s> now faster than skip steps", (*tbl)->pt_name);
1217 
1218 	return (0);
1219 }
1220 
1221 
1222 /*
1223  * Do the dirty work of choosing an unused table name and creating it.
1224  * (be careful with the table name, it might already be used in another anchor)
1225  */
1226 int
1227 pf_opt_create_table(struct pfctl *pf, struct pf_opt_tbl *tbl)
1228 {
1229 	static int tablenum;
1230 	struct pfr_table *t;
1231 
1232 	if (table_buffer.pfrb_type == 0) {
1233 		/* Initialize the list of tables */
1234 		table_buffer.pfrb_type = PFRB_TABLES;
1235 		for (;;) {
1236 			pfr_buf_grow(&table_buffer, table_buffer.pfrb_size);
1237 			table_buffer.pfrb_size = table_buffer.pfrb_msize;
1238 			if (pfr_get_tables(NULL, table_buffer.pfrb_caddr,
1239 			    &table_buffer.pfrb_size, PFR_FLAG_ALLRSETS))
1240 				err(1, "pfr_get_tables");
1241 			if (table_buffer.pfrb_size <= table_buffer.pfrb_msize)
1242 				break;
1243 		}
1244 		table_identifier = arc4random();
1245 	}
1246 
1247 	/* XXX would be *really* nice to avoid duplicating identical tables */
1248 
1249 	/* Now we have to pick a table name that isn't used */
1250 again:
1251 	DEBUG("translating temporary table <%s> to <%s%x_%d>", tbl->pt_name,
1252 	    PF_OPT_TABLE_PREFIX, table_identifier, tablenum);
1253 	snprintf(tbl->pt_name, sizeof(tbl->pt_name), "%s%x_%d",
1254 	    PF_OPT_TABLE_PREFIX, table_identifier, tablenum);
1255 	PFRB_FOREACH(t, &table_buffer) {
1256 		if (strcasecmp(t->pfrt_name, tbl->pt_name) == 0) {
1257 			/* Collision.  Try again */
1258 			DEBUG("wow, table <%s> in use.  trying again",
1259 			    tbl->pt_name);
1260 			table_identifier = arc4random();
1261 			goto again;
1262 		}
1263 	}
1264 	tablenum++;
1265 
1266 
1267 	if (pfctl_define_table(tbl->pt_name, PFR_TFLAG_CONST, 1, pf->anchor,
1268 	    tbl->pt_buf, pf->tticket))
1269 		return (1);
1270 	return (0);
1271 }
1272 
1273 /*
1274  * Partition the flat ruleset into a list of distinct superblocks
1275  */
1276 int
1277 construct_superblocks(struct pfctl *pf, struct pf_opt_queue *opt_queue,
1278     struct superblocks *superblocks)
1279 {
1280 	struct superblock *block = NULL;
1281 	struct pf_opt_rule *por;
1282 	int i;
1283 
1284 	while (!TAILQ_EMPTY(opt_queue)) {
1285 		por = TAILQ_FIRST(opt_queue);
1286 		TAILQ_REMOVE(opt_queue, por, por_entry);
1287 		if (block == NULL || !superblock_inclusive(block, por)) {
1288 			if ((block = calloc(1, sizeof(*block))) == NULL) {
1289 				warn("calloc");
1290 				return (1);
1291 			}
1292 			TAILQ_INIT(&block->sb_rules);
1293 			for (i = 0; i < PF_SKIP_COUNT; i++)
1294 				TAILQ_INIT(&block->sb_skipsteps[i]);
1295 			TAILQ_INSERT_TAIL(superblocks, block, sb_entry);
1296 		}
1297 		TAILQ_INSERT_TAIL(&block->sb_rules, por, por_entry);
1298 	}
1299 
1300 	return (0);
1301 }
1302 
1303 
1304 /*
1305  * Compare two rule addresses
1306  */
1307 int
1308 addrs_equal(struct pf_rule_addr *a, struct pf_rule_addr *b)
1309 {
1310 	if (a->neg != b->neg)
1311 		return (0);
1312 	return (memcmp(&a->addr, &b->addr, sizeof(a->addr)) == 0);
1313 }
1314 
1315 
1316 /*
1317  * The addresses are not equal, but can we combine them into one table?
1318  */
1319 int
1320 addrs_combineable(struct pf_rule_addr *a, struct pf_rule_addr *b)
1321 {
1322 	if (a->addr.type != PF_ADDR_ADDRMASK &&
1323 	    b->addr.type != PF_ADDR_ADDRMASK)
1324 		return (0);
1325 	if (a->neg != b->neg || a->port_op != b->port_op ||
1326 	    a->port[0] != b->port[0] || a->port[1] != b->port[1])
1327 		return (0);
1328 	return (1);
1329 }
1330 
1331 
1332 /*
1333  * Are we allowed to combine these two rules
1334  */
1335 int
1336 rules_combineable(struct pf_rule *p1, struct pf_rule *p2)
1337 {
1338 	struct pf_rule a, b;
1339 
1340 	comparable_rule(&a, p1, COMBINED);
1341 	comparable_rule(&b, p2, COMBINED);
1342 	return (memcmp(&a, &b, sizeof(a)) == 0);
1343 }
1344 
1345 
1346 /*
1347  * Can a rule be included inside a superblock
1348  */
1349 int
1350 superblock_inclusive(struct superblock *block, struct pf_opt_rule *por)
1351 {
1352 	struct pf_rule a, b;
1353 	int i, j;
1354 
1355 	/* First check for hard breaks */
1356 	for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++) {
1357 		if (pf_rule_desc[i].prf_type == BARRIER) {
1358 			for (j = 0; j < pf_rule_desc[i].prf_size; j++)
1359 				if (((char *)&por->por_rule)[j +
1360 				    pf_rule_desc[i].prf_offset] != 0)
1361 					return (0);
1362 		}
1363 	}
1364 
1365 	/* 'anchor' heads and per-rule src-track are also hard breaks */
1366 	if (por->por_anchor[0] != '\0' ||
1367 	    (por->por_rule.rule_flag & PFRULE_RULESRCTRACK))
1368 		return (0);
1369 
1370 	comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule, NOMERGE);
1371 	comparable_rule(&b, &por->por_rule, NOMERGE);
1372 	if (strcmp(TAILQ_FIRST(&block->sb_rules)->por_anchor,
1373 	    por->por_anchor) == 0 && memcmp(&a, &b, sizeof(a)) == 0)
1374 		return (1);
1375 
1376 #ifdef OPT_DEBUG
1377 	for (i = 0; i < sizeof(por->por_rule); i++) {
1378 		int closest = -1;
1379 		if (((u_int8_t *)&a)[i] != ((u_int8_t *)&b)[i]) {
1380 			for (j = 0; j < sizeof(pf_rule_desc) /
1381 			    sizeof(*pf_rule_desc); j++) {
1382 				if (i >= pf_rule_desc[j].prf_offset &&
1383 				    i < pf_rule_desc[j].prf_offset +
1384 				    pf_rule_desc[j].prf_size) {
1385 					DEBUG("superblock break @ %d due to %s",
1386 					    por->por_rule.nr,
1387 					    pf_rule_desc[j].prf_name);
1388 					return (0);
1389 				}
1390 				if (i > pf_rule_desc[j].prf_offset) {
1391 					if (closest == -1 ||
1392 					    i-pf_rule_desc[j].prf_offset <
1393 					    i-pf_rule_desc[closest].prf_offset)
1394 						closest = j;
1395 				}
1396 			}
1397 
1398 			if (closest >= 0)
1399 				DEBUG("superblock break @ %d on %s+%xh",
1400 				    por->por_rule.nr,
1401 				    pf_rule_desc[closest].prf_name,
1402 				    i - pf_rule_desc[closest].prf_offset -
1403 				    pf_rule_desc[closest].prf_size);
1404 			else
1405 				DEBUG("superblock break @ %d on field @ %d",
1406 				    por->por_rule.nr, i);
1407 			return (0);
1408 		}
1409 	}
1410 #endif /* OPT_DEBUG */
1411 
1412 	return (0);
1413 }
1414 
1415 
1416 /*
1417  * Make a rule that can directly compared by memcmp()
1418  */
1419 void
1420 comparable_rule(struct pf_rule *dst, const struct pf_rule *src, int type)
1421 {
1422 	int i;
1423 	/*
1424 	 * To simplify the comparison, we just zero out the fields that are
1425 	 * allowed to be different and then do a simple memcmp()
1426 	 */
1427 	memcpy(dst, src, sizeof(*dst));
1428 	for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++)
1429 		if (pf_rule_desc[i].prf_type >= type) {
1430 #ifdef OPT_DEBUG
1431 			assert(pf_rule_desc[i].prf_type != NEVER ||
1432 			    *(((char *)dst) + pf_rule_desc[i].prf_offset) == 0);
1433 #endif /* OPT_DEBUG */
1434 			memset(((char *)dst) + pf_rule_desc[i].prf_offset, 0,
1435 			    pf_rule_desc[i].prf_size);
1436 		}
1437 }
1438 
1439 
1440 /*
1441  * Remove superset information from two rules so we can directly compare them
1442  * with memcmp()
1443  */
1444 void
1445 exclude_supersets(struct pf_rule *super, struct pf_rule *sub)
1446 {
1447 	if (super->ifname[0] == '\0')
1448 		memset(sub->ifname, 0, sizeof(sub->ifname));
1449 	if (super->direction == PF_INOUT)
1450 		sub->direction = PF_INOUT;
1451 	if ((super->proto == 0 || super->proto == sub->proto) &&
1452 	    super->flags == 0 && super->flagset == 0 && (sub->flags ||
1453 	    sub->flagset)) {
1454 		sub->flags = super->flags;
1455 		sub->flagset = super->flagset;
1456 	}
1457 	if (super->proto == 0)
1458 		sub->proto = 0;
1459 
1460 	if (super->src.port_op == 0) {
1461 		sub->src.port_op = 0;
1462 		sub->src.port[0] = 0;
1463 		sub->src.port[1] = 0;
1464 	}
1465 	if (super->dst.port_op == 0) {
1466 		sub->dst.port_op = 0;
1467 		sub->dst.port[0] = 0;
1468 		sub->dst.port[1] = 0;
1469 	}
1470 
1471 	if (super->src.addr.type == PF_ADDR_ADDRMASK && !super->src.neg &&
1472 	    !sub->src.neg && super->src.addr.v.a.mask.addr32[0] == 0 &&
1473 	    super->src.addr.v.a.mask.addr32[1] == 0 &&
1474 	    super->src.addr.v.a.mask.addr32[2] == 0 &&
1475 	    super->src.addr.v.a.mask.addr32[3] == 0)
1476 		memset(&sub->src.addr, 0, sizeof(sub->src.addr));
1477 	else if (super->src.addr.type == PF_ADDR_ADDRMASK &&
1478 	    sub->src.addr.type == PF_ADDR_ADDRMASK &&
1479 	    super->src.neg == sub->src.neg &&
1480 	    super->af == sub->af &&
1481 	    unmask(&super->src.addr.v.a.mask, super->af) <
1482 	    unmask(&sub->src.addr.v.a.mask, sub->af) &&
1483 	    super->src.addr.v.a.addr.addr32[0] ==
1484 	    (sub->src.addr.v.a.addr.addr32[0] &
1485 	    super->src.addr.v.a.mask.addr32[0]) &&
1486 	    super->src.addr.v.a.addr.addr32[1] ==
1487 	    (sub->src.addr.v.a.addr.addr32[1] &
1488 	    super->src.addr.v.a.mask.addr32[1]) &&
1489 	    super->src.addr.v.a.addr.addr32[2] ==
1490 	    (sub->src.addr.v.a.addr.addr32[2] &
1491 	    super->src.addr.v.a.mask.addr32[2]) &&
1492 	    super->src.addr.v.a.addr.addr32[3] ==
1493 	    (sub->src.addr.v.a.addr.addr32[3] &
1494 	    super->src.addr.v.a.mask.addr32[3])) {
1495 		/* sub->src.addr is a subset of super->src.addr/mask */
1496 		memcpy(&sub->src.addr, &super->src.addr, sizeof(sub->src.addr));
1497 	}
1498 
1499 	if (super->dst.addr.type == PF_ADDR_ADDRMASK && !super->dst.neg &&
1500 	    !sub->dst.neg && super->dst.addr.v.a.mask.addr32[0] == 0 &&
1501 	    super->dst.addr.v.a.mask.addr32[1] == 0 &&
1502 	    super->dst.addr.v.a.mask.addr32[2] == 0 &&
1503 	    super->dst.addr.v.a.mask.addr32[3] == 0)
1504 		memset(&sub->dst.addr, 0, sizeof(sub->dst.addr));
1505 	else if (super->dst.addr.type == PF_ADDR_ADDRMASK &&
1506 	    sub->dst.addr.type == PF_ADDR_ADDRMASK &&
1507 	    super->dst.neg == sub->dst.neg &&
1508 	    super->af == sub->af &&
1509 	    unmask(&super->dst.addr.v.a.mask, super->af) <
1510 	    unmask(&sub->dst.addr.v.a.mask, sub->af) &&
1511 	    super->dst.addr.v.a.addr.addr32[0] ==
1512 	    (sub->dst.addr.v.a.addr.addr32[0] &
1513 	    super->dst.addr.v.a.mask.addr32[0]) &&
1514 	    super->dst.addr.v.a.addr.addr32[1] ==
1515 	    (sub->dst.addr.v.a.addr.addr32[1] &
1516 	    super->dst.addr.v.a.mask.addr32[1]) &&
1517 	    super->dst.addr.v.a.addr.addr32[2] ==
1518 	    (sub->dst.addr.v.a.addr.addr32[2] &
1519 	    super->dst.addr.v.a.mask.addr32[2]) &&
1520 	    super->dst.addr.v.a.addr.addr32[3] ==
1521 	    (sub->dst.addr.v.a.addr.addr32[3] &
1522 	    super->dst.addr.v.a.mask.addr32[3])) {
1523 		/* sub->dst.addr is a subset of super->dst.addr/mask */
1524 		memcpy(&sub->dst.addr, &super->dst.addr, sizeof(sub->dst.addr));
1525 	}
1526 
1527 	if (super->af == 0)
1528 		sub->af = 0;
1529 }
1530 
1531 
1532 void
1533 superblock_free(struct pfctl *pf, struct superblock *block)
1534 {
1535 	struct pf_opt_rule *por;
1536 	while ((por = TAILQ_FIRST(&block->sb_rules))) {
1537 		TAILQ_REMOVE(&block->sb_rules, por, por_entry);
1538 		if (por->por_src_tbl) {
1539 			if (por->por_src_tbl->pt_buf) {
1540 				pfr_buf_clear(por->por_src_tbl->pt_buf);
1541 				free(por->por_src_tbl->pt_buf);
1542 			}
1543 			free(por->por_src_tbl);
1544 		}
1545 		if (por->por_dst_tbl) {
1546 			if (por->por_dst_tbl->pt_buf) {
1547 				pfr_buf_clear(por->por_dst_tbl->pt_buf);
1548 				free(por->por_dst_tbl->pt_buf);
1549 			}
1550 			free(por->por_dst_tbl);
1551 		}
1552 		free(por);
1553 	}
1554 	if (block->sb_profiled_block)
1555 		superblock_free(pf, block->sb_profiled_block);
1556 	free(block);
1557 }
1558 
1559