xref: /freebsd/sbin/pfctl/pfctl_optimize.c (revision d96700a6da2afa88607fbd7405ade439424d10d9)
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 <stddef.h>
37 #include <stdio.h>
38 #include <stdlib.h>
39 #include <string.h>
40 
41 #include "pfctl_parser.h"
42 #include "pfctl.h"
43 
44 /* The size at which a table becomes faster than individual rules */
45 #define TABLE_THRESHOLD		6
46 
47 
48 /* #define OPT_DEBUG	1 */
49 #ifdef OPT_DEBUG
50 # define DEBUG(str, v...) \
51 	printf("%s: " str "\n", __FUNCTION__ , ## v)
52 #else
53 # define DEBUG(str, v...) ((void)0)
54 #endif
55 
56 
57 /*
58  * A container that lets us sort a superblock to optimize the skip step jumps
59  */
60 struct pf_skip_step {
61 	int				ps_count;	/* number of items */
62 	TAILQ_HEAD( , pf_opt_rule)	ps_rules;
63 	TAILQ_ENTRY(pf_skip_step)	ps_entry;
64 };
65 
66 
67 /*
68  * A superblock is a block of adjacent rules of similar action.  If there
69  * are five PASS rules in a row, they all become members of a superblock.
70  * Once we have a superblock, we are free to re-order any rules within it
71  * in order to improve performance; if a packet is passed, it doesn't matter
72  * who passed it.
73  */
74 struct superblock {
75 	TAILQ_HEAD( , pf_opt_rule)		 sb_rules;
76 	TAILQ_ENTRY(superblock)			 sb_entry;
77 	struct superblock			*sb_profiled_block;
78 	TAILQ_HEAD(skiplist, pf_skip_step)	 sb_skipsteps[PF_SKIP_COUNT];
79 };
80 TAILQ_HEAD(superblocks, superblock);
81 
82 
83 /*
84  * Description of the PF rule structure.
85  */
86 enum {
87     BARRIER,	/* the presence of the field puts the rule in it's own block */
88     BREAK,	/* the field may not differ between rules in a superblock */
89     NOMERGE,	/* the field may not differ between rules when combined */
90     COMBINED,	/* the field may itself be combined with other rules */
91     DC,		/* we just don't care about the field */
92     NEVER};	/* we should never see this field set?!? */
93 static struct pf_rule_field {
94 	const char	*prf_name;
95 	int		 prf_type;
96 	size_t		 prf_offset;
97 	size_t		 prf_size;
98 } pf_rule_desc[] = {
99 #define PF_RULE_FIELD(field, ty)	\
100     {#field,				\
101     ty,					\
102     offsetof(struct pf_rule, field),	\
103     sizeof(((struct pf_rule *)0)->field)}
104 
105 
106     /*
107      * The presence of these fields in a rule put the rule in it's own
108      * superblock.  Thus it will not be optimized.  It also prevents the
109      * rule from being re-ordered at all.
110      */
111     PF_RULE_FIELD(label,		BARRIER),
112     PF_RULE_FIELD(prob,			BARRIER),
113     PF_RULE_FIELD(max_states,		BARRIER),
114     PF_RULE_FIELD(max_src_nodes,	BARRIER),
115     PF_RULE_FIELD(max_src_states,	BARRIER),
116     PF_RULE_FIELD(max_src_conn,		BARRIER),
117     PF_RULE_FIELD(max_src_conn_rate,	BARRIER),
118     PF_RULE_FIELD(anchor,		BARRIER),	/* for now */
119 
120     /*
121      * These fields must be the same between all rules in the same superblock.
122      * These rules are allowed to be re-ordered but only among like rules.
123      * For instance we can re-order all 'tag "foo"' rules because they have the
124      * same tag.  But we can not re-order between a 'tag "foo"' and a
125      * 'tag "bar"' since that would change the meaning of the ruleset.
126      */
127     PF_RULE_FIELD(tagname,		BREAK),
128     PF_RULE_FIELD(keep_state,		BREAK),
129     PF_RULE_FIELD(qname,		BREAK),
130     PF_RULE_FIELD(pqname,		BREAK),
131     PF_RULE_FIELD(rt,			BREAK),
132     PF_RULE_FIELD(allow_opts,		BREAK),
133     PF_RULE_FIELD(rule_flag,		BREAK),
134     PF_RULE_FIELD(action,		BREAK),
135     PF_RULE_FIELD(log,			BREAK),
136     PF_RULE_FIELD(quick,		BREAK),
137     PF_RULE_FIELD(return_ttl,		BREAK),
138     PF_RULE_FIELD(overload_tblname,	BREAK),
139     PF_RULE_FIELD(flush,		BREAK),
140     PF_RULE_FIELD(rpool,		BREAK),
141     PF_RULE_FIELD(logif,		BREAK),
142 
143     /*
144      * Any fields not listed in this structure act as BREAK fields
145      */
146 
147 
148     /*
149      * These fields must not differ when we merge two rules together but
150      * their difference isn't enough to put the rules in different superblocks.
151      * There are no problems re-ordering any rules with these fields.
152      */
153     PF_RULE_FIELD(af,			NOMERGE),
154     PF_RULE_FIELD(ifnot,		NOMERGE),
155     PF_RULE_FIELD(ifname,		NOMERGE),	/* hack for IF groups */
156     PF_RULE_FIELD(match_tag_not,	NOMERGE),
157     PF_RULE_FIELD(match_tagname,	NOMERGE),
158     PF_RULE_FIELD(os_fingerprint,	NOMERGE),
159     PF_RULE_FIELD(timeout,		NOMERGE),
160     PF_RULE_FIELD(return_icmp,		NOMERGE),
161     PF_RULE_FIELD(return_icmp6,		NOMERGE),
162     PF_RULE_FIELD(uid,			NOMERGE),
163     PF_RULE_FIELD(gid,			NOMERGE),
164     PF_RULE_FIELD(direction,		NOMERGE),
165     PF_RULE_FIELD(proto,		NOMERGE),
166     PF_RULE_FIELD(type,			NOMERGE),
167     PF_RULE_FIELD(code,			NOMERGE),
168     PF_RULE_FIELD(flags,		NOMERGE),
169     PF_RULE_FIELD(flagset,		NOMERGE),
170     PF_RULE_FIELD(tos,			NOMERGE),
171     PF_RULE_FIELD(src.port,		NOMERGE),
172     PF_RULE_FIELD(dst.port,		NOMERGE),
173     PF_RULE_FIELD(src.port_op,		NOMERGE),
174     PF_RULE_FIELD(dst.port_op,		NOMERGE),
175     PF_RULE_FIELD(src.neg,		NOMERGE),
176     PF_RULE_FIELD(dst.neg,		NOMERGE),
177 
178     /* These fields can be merged */
179     PF_RULE_FIELD(src.addr,		COMBINED),
180     PF_RULE_FIELD(dst.addr,		COMBINED),
181 
182     /* We just don't care about these fields.  They're set by the kernel */
183     PF_RULE_FIELD(skip,			DC),
184     PF_RULE_FIELD(evaluations,		DC),
185     PF_RULE_FIELD(packets,		DC),
186     PF_RULE_FIELD(bytes,		DC),
187     PF_RULE_FIELD(kif,			DC),
188     PF_RULE_FIELD(states_cur,		DC),
189     PF_RULE_FIELD(states_tot,		DC),
190     PF_RULE_FIELD(src_nodes,		DC),
191     PF_RULE_FIELD(nr,			DC),
192     PF_RULE_FIELD(entries,		DC),
193     PF_RULE_FIELD(qid,			DC),
194     PF_RULE_FIELD(pqid,			DC),
195     PF_RULE_FIELD(anchor_relative,	DC),
196     PF_RULE_FIELD(anchor_wildcard,	DC),
197     PF_RULE_FIELD(tag,			DC),
198     PF_RULE_FIELD(match_tag,		DC),
199     PF_RULE_FIELD(overload_tbl,		DC),
200 
201     /* These fields should never be set in a PASS/BLOCK rule */
202     PF_RULE_FIELD(natpass,		NEVER),
203     PF_RULE_FIELD(max_mss,		NEVER),
204     PF_RULE_FIELD(min_ttl,		NEVER),
205     PF_RULE_FIELD(set_tos,		NEVER),
206 };
207 
208 
209 
210 int	add_opt_table(struct pfctl *, struct pf_opt_tbl **, sa_family_t,
211 	    struct pf_rule_addr *);
212 int	addrs_combineable(struct pf_rule_addr *, struct pf_rule_addr *);
213 int	addrs_equal(struct pf_rule_addr *, struct pf_rule_addr *);
214 int	block_feedback(struct pfctl *, struct superblock *);
215 int	combine_rules(struct pfctl *, struct superblock *);
216 void	comparable_rule(struct pf_rule *, const struct pf_rule *, int);
217 int	construct_superblocks(struct pfctl *, struct pf_opt_queue *,
218 	    struct superblocks *);
219 void	exclude_supersets(struct pf_rule *, struct pf_rule *);
220 int	interface_group(const char *);
221 int	load_feedback_profile(struct pfctl *, struct superblocks *);
222 int	optimize_superblock(struct pfctl *, struct superblock *);
223 int	pf_opt_create_table(struct pfctl *, struct pf_opt_tbl *);
224 void	remove_from_skipsteps(struct skiplist *, struct superblock *,
225 	    struct pf_opt_rule *, struct pf_skip_step *);
226 int	remove_identical_rules(struct pfctl *, struct superblock *);
227 int	reorder_rules(struct pfctl *, struct superblock *, int);
228 int	rules_combineable(struct pf_rule *, struct pf_rule *);
229 void	skip_append(struct superblock *, int, struct pf_skip_step *,
230 	    struct pf_opt_rule *);
231 int	skip_compare(int, struct pf_skip_step *, struct pf_opt_rule *);
232 void	skip_init(void);
233 int	skip_cmp_af(struct pf_rule *, struct pf_rule *);
234 int	skip_cmp_dir(struct pf_rule *, struct pf_rule *);
235 int	skip_cmp_dst_addr(struct pf_rule *, struct pf_rule *);
236 int	skip_cmp_dst_port(struct pf_rule *, struct pf_rule *);
237 int	skip_cmp_ifp(struct pf_rule *, struct pf_rule *);
238 int	skip_cmp_proto(struct pf_rule *, struct pf_rule *);
239 int	skip_cmp_src_addr(struct pf_rule *, struct pf_rule *);
240 int	skip_cmp_src_port(struct pf_rule *, struct pf_rule *);
241 int	superblock_inclusive(struct superblock *, struct pf_opt_rule *);
242 void	superblock_free(struct pfctl *, struct superblock *);
243 
244 
245 static int (*skip_comparitors[PF_SKIP_COUNT])(struct pf_rule *,
246     struct pf_rule *);
247 static const char *skip_comparitors_names[PF_SKIP_COUNT];
248 #define PF_SKIP_COMPARITORS {				\
249     { "ifp", PF_SKIP_IFP, skip_cmp_ifp },		\
250     { "dir", PF_SKIP_DIR, skip_cmp_dir },		\
251     { "af", PF_SKIP_AF, skip_cmp_af },			\
252     { "proto", PF_SKIP_PROTO, skip_cmp_proto },		\
253     { "saddr", PF_SKIP_SRC_ADDR, skip_cmp_src_addr },	\
254     { "sport", PF_SKIP_SRC_PORT, skip_cmp_src_port },	\
255     { "daddr", PF_SKIP_DST_ADDR, skip_cmp_dst_addr },	\
256     { "dport", PF_SKIP_DST_PORT, skip_cmp_dst_port }	\
257 }
258 
259 static struct pfr_buffer table_buffer;
260 static int table_identifier;
261 
262 
263 int
264 pfctl_optimize_ruleset(struct pfctl *pf, struct pf_ruleset *rs)
265 {
266 	struct superblocks superblocks;
267 	struct pf_opt_queue opt_queue;
268 	struct superblock *block;
269 	struct pf_opt_rule *por;
270 	struct pf_rule *r;
271 	struct pf_rulequeue *old_rules;
272 
273 	DEBUG("optimizing ruleset");
274 	memset(&table_buffer, 0, sizeof(table_buffer));
275 	skip_init();
276 	TAILQ_INIT(&opt_queue);
277 
278 	old_rules = rs->rules[PF_RULESET_FILTER].active.ptr;
279 	rs->rules[PF_RULESET_FILTER].active.ptr =
280 	    rs->rules[PF_RULESET_FILTER].inactive.ptr;
281 	rs->rules[PF_RULESET_FILTER].inactive.ptr = old_rules;
282 
283 	/*
284 	 * XXX expanding the pf_opt_rule format throughout pfctl might allow
285 	 * us to avoid all this copying.
286 	 */
287 	while ((r = TAILQ_FIRST(rs->rules[PF_RULESET_FILTER].inactive.ptr))
288 	    != NULL) {
289 		TAILQ_REMOVE(rs->rules[PF_RULESET_FILTER].inactive.ptr, r,
290 		    entries);
291 		if ((por = calloc(1, sizeof(*por))) == NULL)
292 			err(1, "calloc");
293 		memcpy(&por->por_rule, r, sizeof(*r));
294 		if (TAILQ_FIRST(&r->rpool.list) != NULL) {
295 			TAILQ_INIT(&por->por_rule.rpool.list);
296 			pfctl_move_pool(&r->rpool, &por->por_rule.rpool);
297 		} else
298 			bzero(&por->por_rule.rpool,
299 			    sizeof(por->por_rule.rpool));
300 
301 
302 		TAILQ_INSERT_TAIL(&opt_queue, por, por_entry);
303 	}
304 
305 	TAILQ_INIT(&superblocks);
306 	if (construct_superblocks(pf, &opt_queue, &superblocks))
307 		goto error;
308 
309 	if (pf->optimize & PF_OPTIMIZE_PROFILE) {
310 		if (load_feedback_profile(pf, &superblocks))
311 			goto error;
312 	}
313 
314 	TAILQ_FOREACH(block, &superblocks, sb_entry) {
315 		if (optimize_superblock(pf, block))
316 			goto error;
317 	}
318 
319 	rs->anchor->refcnt = 0;
320 	while ((block = TAILQ_FIRST(&superblocks))) {
321 		TAILQ_REMOVE(&superblocks, block, sb_entry);
322 
323 		while ((por = TAILQ_FIRST(&block->sb_rules))) {
324 			TAILQ_REMOVE(&block->sb_rules, por, por_entry);
325 			por->por_rule.nr = rs->anchor->refcnt++;
326 			if ((r = calloc(1, sizeof(*r))) == NULL)
327 				err(1, "calloc");
328 			memcpy(r, &por->por_rule, sizeof(*r));
329 			TAILQ_INIT(&r->rpool.list);
330 			pfctl_move_pool(&por->por_rule.rpool, &r->rpool);
331 			TAILQ_INSERT_TAIL(
332 			    rs->rules[PF_RULESET_FILTER].active.ptr,
333 			    r, entries);
334 			free(por);
335 		}
336 		free(block);
337 	}
338 
339 	return (0);
340 
341 error:
342 	while ((por = TAILQ_FIRST(&opt_queue))) {
343 		TAILQ_REMOVE(&opt_queue, por, por_entry);
344 		if (por->por_src_tbl) {
345 			pfr_buf_clear(por->por_src_tbl->pt_buf);
346 			free(por->por_src_tbl->pt_buf);
347 			free(por->por_src_tbl);
348 		}
349 		if (por->por_dst_tbl) {
350 			pfr_buf_clear(por->por_dst_tbl->pt_buf);
351 			free(por->por_dst_tbl->pt_buf);
352 			free(por->por_dst_tbl);
353 		}
354 		free(por);
355 	}
356 	while ((block = TAILQ_FIRST(&superblocks))) {
357 		TAILQ_REMOVE(&superblocks, block, sb_entry);
358 		superblock_free(pf, block);
359 	}
360 	return (1);
361 }
362 
363 
364 /*
365  * Go ahead and optimize a superblock
366  */
367 int
368 optimize_superblock(struct pfctl *pf, struct superblock *block)
369 {
370 #ifdef OPT_DEBUG
371 	struct pf_opt_rule *por;
372 #endif /* OPT_DEBUG */
373 
374 	/* We have a few optimization passes:
375 	 *   1) remove duplicate rules or rules that are a subset of other
376 	 *      rules
377 	 *   2) combine otherwise identical rules with different IP addresses
378 	 *      into a single rule and put the addresses in a table.
379 	 *   3) re-order the rules to improve kernel skip steps
380 	 *   4) re-order the 'quick' rules based on feedback from the
381 	 *      active ruleset statistics
382 	 *
383 	 * XXX combine_rules() doesn't combine v4 and v6 rules.  would just
384 	 *     have to keep af in the table container, make af 'COMBINE' and
385 	 *     twiddle the af on the merged rule
386 	 * XXX maybe add a weighting to the metric on skipsteps when doing
387 	 *     reordering.  sometimes two sequential tables will be better
388 	 *     that four consecutive interfaces.
389 	 * XXX need to adjust the skipstep count of everything after PROTO,
390 	 *     since they aren't actually checked on a proto mismatch in
391 	 *     pf_test_{tcp, udp, icmp}()
392 	 * XXX should i treat proto=0, af=0 or dir=0 special in skepstep
393 	 *     calculation since they are a DC?
394 	 * XXX keep last skiplist of last superblock to influence this
395 	 *     superblock.  '5 inet6 log' should make '3 inet6' come before '4
396 	 *     inet' in the next superblock.
397 	 * XXX would be useful to add tables for ports
398 	 * XXX we can also re-order some mutually exclusive superblocks to
399 	 *     try merging superblocks before any of these optimization passes.
400 	 *     for instance a single 'log in' rule in the middle of non-logging
401 	 *     out rules.
402 	 */
403 
404 	/* shortcut.  there will be a lot of 1-rule superblocks */
405 	if (!TAILQ_NEXT(TAILQ_FIRST(&block->sb_rules), por_entry))
406 		return (0);
407 
408 #ifdef OPT_DEBUG
409 	printf("--- Superblock ---\n");
410 	TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
411 		printf("  ");
412 		print_rule(&por->por_rule, por->por_rule.anchor ?
413 		    por->por_rule.anchor->name : "", 1, 0);
414 	}
415 #endif /* OPT_DEBUG */
416 
417 
418 	if (remove_identical_rules(pf, block))
419 		return (1);
420 	if (combine_rules(pf, block))
421 		return (1);
422 	if ((pf->optimize & PF_OPTIMIZE_PROFILE) &&
423 	    TAILQ_FIRST(&block->sb_rules)->por_rule.quick &&
424 	    block->sb_profiled_block) {
425 		if (block_feedback(pf, block))
426 			return (1);
427 	} else if (reorder_rules(pf, block, 0)) {
428 		return (1);
429 	}
430 
431 	/*
432 	 * Don't add any optimization passes below reorder_rules().  It will
433 	 * have divided superblocks into smaller blocks for further refinement
434 	 * and doesn't put them back together again.  What once was a true
435 	 * superblock might have been split into multiple superblocks.
436 	 */
437 
438 #ifdef OPT_DEBUG
439 	printf("--- END Superblock ---\n");
440 #endif /* OPT_DEBUG */
441 	return (0);
442 }
443 
444 
445 /*
446  * Optimization pass #1: remove identical rules
447  */
448 int
449 remove_identical_rules(struct pfctl *pf, struct superblock *block)
450 {
451 	struct pf_opt_rule *por1, *por2, *por_next, *por2_next;
452 	struct pf_rule a, a2, b, b2;
453 
454 	for (por1 = TAILQ_FIRST(&block->sb_rules); por1; por1 = por_next) {
455 		por_next = TAILQ_NEXT(por1, por_entry);
456 		for (por2 = por_next; por2; por2 = por2_next) {
457 			por2_next = TAILQ_NEXT(por2, por_entry);
458 			comparable_rule(&a, &por1->por_rule, DC);
459 			comparable_rule(&b, &por2->por_rule, DC);
460 			memcpy(&a2, &a, sizeof(a2));
461 			memcpy(&b2, &b, sizeof(b2));
462 
463 			exclude_supersets(&a, &b);
464 			exclude_supersets(&b2, &a2);
465 			if (memcmp(&a, &b, sizeof(a)) == 0) {
466 				DEBUG("removing identical rule  nr%d = *nr%d*",
467 				    por1->por_rule.nr, por2->por_rule.nr);
468 				TAILQ_REMOVE(&block->sb_rules, por2, por_entry);
469 				if (por_next == por2)
470 					por_next = TAILQ_NEXT(por1, por_entry);
471 				free(por2);
472 			} else if (memcmp(&a2, &b2, sizeof(a2)) == 0) {
473 				DEBUG("removing identical rule  *nr%d* = nr%d",
474 				    por1->por_rule.nr, por2->por_rule.nr);
475 				TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
476 				free(por1);
477 				break;
478 			}
479 		}
480 	}
481 
482 	return (0);
483 }
484 
485 
486 /*
487  * Optimization pass #2: combine similar rules with different addresses
488  * into a single rule and a table
489  */
490 int
491 combine_rules(struct pfctl *pf, struct superblock *block)
492 {
493 	struct pf_opt_rule *p1, *p2, *por_next;
494 	int src_eq, dst_eq;
495 
496 	if ((pf->loadopt & PFCTL_FLAG_TABLE) == 0) {
497 		warnx("Must enable table loading for optimizations");
498 		return (1);
499 	}
500 
501 	/* First we make a pass to combine the rules.  O(n log n) */
502 	TAILQ_FOREACH(p1, &block->sb_rules, por_entry) {
503 		for (p2 = TAILQ_NEXT(p1, por_entry); p2; p2 = por_next) {
504 			por_next = TAILQ_NEXT(p2, por_entry);
505 
506 			src_eq = addrs_equal(&p1->por_rule.src,
507 			    &p2->por_rule.src);
508 			dst_eq = addrs_equal(&p1->por_rule.dst,
509 			    &p2->por_rule.dst);
510 
511 			if (src_eq && !dst_eq && p1->por_src_tbl == NULL &&
512 			    p2->por_dst_tbl == NULL &&
513 			    p2->por_src_tbl == NULL &&
514 			    rules_combineable(&p1->por_rule, &p2->por_rule) &&
515 			    addrs_combineable(&p1->por_rule.dst,
516 			    &p2->por_rule.dst)) {
517 				DEBUG("can combine rules  nr%d = nr%d",
518 				    p1->por_rule.nr, p2->por_rule.nr);
519 				if (p1->por_dst_tbl == NULL &&
520 				    add_opt_table(pf, &p1->por_dst_tbl,
521 				    p1->por_rule.af, &p1->por_rule.dst))
522 					return (1);
523 				if (add_opt_table(pf, &p1->por_dst_tbl,
524 				    p1->por_rule.af, &p2->por_rule.dst))
525 					return (1);
526 				p2->por_dst_tbl = p1->por_dst_tbl;
527 				if (p1->por_dst_tbl->pt_rulecount >=
528 				    TABLE_THRESHOLD) {
529 					TAILQ_REMOVE(&block->sb_rules, p2,
530 					    por_entry);
531 					free(p2);
532 				}
533 			} else if (!src_eq && dst_eq && p1->por_dst_tbl == NULL
534 			    && p2->por_src_tbl == NULL &&
535 			    p2->por_dst_tbl == NULL &&
536 			    rules_combineable(&p1->por_rule, &p2->por_rule) &&
537 			    addrs_combineable(&p1->por_rule.src,
538 			    &p2->por_rule.src)) {
539 				DEBUG("can combine rules  nr%d = nr%d",
540 				    p1->por_rule.nr, p2->por_rule.nr);
541 				if (p1->por_src_tbl == NULL &&
542 				    add_opt_table(pf, &p1->por_src_tbl,
543 				    p1->por_rule.af, &p1->por_rule.src))
544 					return (1);
545 				if (add_opt_table(pf, &p1->por_src_tbl,
546 				    p1->por_rule.af, &p2->por_rule.src))
547 					return (1);
548 				p2->por_src_tbl = p1->por_src_tbl;
549 				if (p1->por_src_tbl->pt_rulecount >=
550 				    TABLE_THRESHOLD) {
551 					TAILQ_REMOVE(&block->sb_rules, p2,
552 					    por_entry);
553 					free(p2);
554 				}
555 			}
556 		}
557 	}
558 
559 
560 	/*
561 	 * Then we make a final pass to create a valid table name and
562 	 * insert the name into the rules.
563 	 */
564 	for (p1 = TAILQ_FIRST(&block->sb_rules); p1; p1 = por_next) {
565 		por_next = TAILQ_NEXT(p1, por_entry);
566 		assert(p1->por_src_tbl == NULL || p1->por_dst_tbl == NULL);
567 
568 		if (p1->por_src_tbl && p1->por_src_tbl->pt_rulecount >=
569 		    TABLE_THRESHOLD) {
570 			if (p1->por_src_tbl->pt_generated) {
571 				/* This rule is included in a table */
572 				TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
573 				free(p1);
574 				continue;
575 			}
576 			p1->por_src_tbl->pt_generated = 1;
577 
578 			if ((pf->opts & PF_OPT_NOACTION) == 0 &&
579 			    pf_opt_create_table(pf, p1->por_src_tbl))
580 				return (1);
581 
582 			pf->tdirty = 1;
583 
584 			if (pf->opts & PF_OPT_VERBOSE)
585 				print_tabledef(p1->por_src_tbl->pt_name,
586 				    PFR_TFLAG_CONST, 1,
587 				    &p1->por_src_tbl->pt_nodes);
588 
589 			memset(&p1->por_rule.src.addr, 0,
590 			    sizeof(p1->por_rule.src.addr));
591 			p1->por_rule.src.addr.type = PF_ADDR_TABLE;
592 			strlcpy(p1->por_rule.src.addr.v.tblname,
593 			    p1->por_src_tbl->pt_name,
594 			    sizeof(p1->por_rule.src.addr.v.tblname));
595 
596 			pfr_buf_clear(p1->por_src_tbl->pt_buf);
597 			free(p1->por_src_tbl->pt_buf);
598 			p1->por_src_tbl->pt_buf = NULL;
599 		}
600 		if (p1->por_dst_tbl && p1->por_dst_tbl->pt_rulecount >=
601 		    TABLE_THRESHOLD) {
602 			if (p1->por_dst_tbl->pt_generated) {
603 				/* This rule is included in a table */
604 				TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
605 				free(p1);
606 				continue;
607 			}
608 			p1->por_dst_tbl->pt_generated = 1;
609 
610 			if ((pf->opts & PF_OPT_NOACTION) == 0 &&
611 			    pf_opt_create_table(pf, p1->por_dst_tbl))
612 				return (1);
613 			pf->tdirty = 1;
614 
615 			if (pf->opts & PF_OPT_VERBOSE)
616 				print_tabledef(p1->por_dst_tbl->pt_name,
617 				    PFR_TFLAG_CONST, 1,
618 				    &p1->por_dst_tbl->pt_nodes);
619 
620 			memset(&p1->por_rule.dst.addr, 0,
621 			    sizeof(p1->por_rule.dst.addr));
622 			p1->por_rule.dst.addr.type = PF_ADDR_TABLE;
623 			strlcpy(p1->por_rule.dst.addr.v.tblname,
624 			    p1->por_dst_tbl->pt_name,
625 			    sizeof(p1->por_rule.dst.addr.v.tblname));
626 
627 			pfr_buf_clear(p1->por_dst_tbl->pt_buf);
628 			free(p1->por_dst_tbl->pt_buf);
629 			p1->por_dst_tbl->pt_buf = NULL;
630 		}
631 	}
632 
633 	return (0);
634 }
635 
636 
637 /*
638  * Optimization pass #3: re-order rules to improve skip steps
639  */
640 int
641 reorder_rules(struct pfctl *pf, struct superblock *block, int depth)
642 {
643 	struct superblock *newblock;
644 	struct pf_skip_step *skiplist;
645 	struct pf_opt_rule *por;
646 	int i, largest, largest_list, rule_count = 0;
647 	TAILQ_HEAD( , pf_opt_rule) head;
648 
649 	/*
650 	 * Calculate the best-case skip steps.  We put each rule in a list
651 	 * of other rules with common fields
652 	 */
653 	for (i = 0; i < PF_SKIP_COUNT; i++) {
654 		TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
655 			TAILQ_FOREACH(skiplist, &block->sb_skipsteps[i],
656 			    ps_entry) {
657 				if (skip_compare(i, skiplist, por) == 0)
658 					break;
659 			}
660 			if (skiplist == NULL) {
661 				if ((skiplist = calloc(1, sizeof(*skiplist))) ==
662 				    NULL)
663 					err(1, "calloc");
664 				TAILQ_INIT(&skiplist->ps_rules);
665 				TAILQ_INSERT_TAIL(&block->sb_skipsteps[i],
666 				    skiplist, ps_entry);
667 			}
668 			skip_append(block, i, skiplist, por);
669 		}
670 	}
671 
672 	TAILQ_FOREACH(por, &block->sb_rules, por_entry)
673 		rule_count++;
674 
675 	/*
676 	 * Now we're going to ignore any fields that are identical between
677 	 * all of the rules in the superblock and those fields which differ
678 	 * between every rule in the superblock.
679 	 */
680 	largest = 0;
681 	for (i = 0; i < PF_SKIP_COUNT; i++) {
682 		skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
683 		if (skiplist->ps_count == rule_count) {
684 			DEBUG("(%d) original skipstep '%s' is all rules",
685 			    depth, skip_comparitors_names[i]);
686 			skiplist->ps_count = 0;
687 		} else if (skiplist->ps_count == 1) {
688 			skiplist->ps_count = 0;
689 		} else {
690 			DEBUG("(%d) original skipstep '%s' largest jump is %d",
691 			    depth, skip_comparitors_names[i],
692 			    skiplist->ps_count);
693 			if (skiplist->ps_count > largest)
694 				largest = skiplist->ps_count;
695 		}
696 	}
697 	if (largest == 0) {
698 		/* Ugh.  There is NO commonality in the superblock on which
699 		 * optimize the skipsteps optimization.
700 		 */
701 		goto done;
702 	}
703 
704 	/*
705 	 * Now we're going to empty the superblock rule list and re-create
706 	 * it based on a more optimal skipstep order.
707 	 */
708 	TAILQ_INIT(&head);
709 	while ((por = TAILQ_FIRST(&block->sb_rules))) {
710 		TAILQ_REMOVE(&block->sb_rules, por, por_entry);
711 		TAILQ_INSERT_TAIL(&head, por, por_entry);
712 	}
713 
714 
715 	while (!TAILQ_EMPTY(&head)) {
716 		largest = 1;
717 
718 		/*
719 		 * Find the most useful skip steps remaining
720 		 */
721 		for (i = 0; i < PF_SKIP_COUNT; i++) {
722 			skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
723 			if (skiplist->ps_count > largest) {
724 				largest = skiplist->ps_count;
725 				largest_list = i;
726 			}
727 		}
728 
729 		if (largest <= 1) {
730 			/*
731 			 * Nothing useful left.  Leave remaining rules in order.
732 			 */
733 			DEBUG("(%d) no more commonality for skip steps", depth);
734 			while ((por = TAILQ_FIRST(&head))) {
735 				TAILQ_REMOVE(&head, por, por_entry);
736 				TAILQ_INSERT_TAIL(&block->sb_rules, por,
737 				    por_entry);
738 			}
739 		} else {
740 			/*
741 			 * There is commonality.  Extract those common rules
742 			 * and place them in the ruleset adjacent to each
743 			 * other.
744 			 */
745 			skiplist = TAILQ_FIRST(&block->sb_skipsteps[
746 			    largest_list]);
747 			DEBUG("(%d) skipstep '%s' largest jump is %d @ #%d",
748 			    depth, skip_comparitors_names[largest_list],
749 			    largest, TAILQ_FIRST(&TAILQ_FIRST(&block->
750 			    sb_skipsteps [largest_list])->ps_rules)->
751 			    por_rule.nr);
752 			TAILQ_REMOVE(&block->sb_skipsteps[largest_list],
753 			    skiplist, ps_entry);
754 
755 
756 			/*
757 			 * There may be further commonality inside these
758 			 * rules.  So we'll split them off into they're own
759 			 * superblock and pass it back into the optimizer.
760 			 */
761 			if (skiplist->ps_count > 2) {
762 				if ((newblock = calloc(1, sizeof(*newblock)))
763 				    == NULL) {
764 					warn("calloc");
765 					return (1);
766 				}
767 				TAILQ_INIT(&newblock->sb_rules);
768 				for (i = 0; i < PF_SKIP_COUNT; i++)
769 					TAILQ_INIT(&newblock->sb_skipsteps[i]);
770 				TAILQ_INSERT_BEFORE(block, newblock, sb_entry);
771 				DEBUG("(%d) splitting off %d rules from superblock @ #%d",
772 				    depth, skiplist->ps_count,
773 				    TAILQ_FIRST(&skiplist->ps_rules)->
774 				    por_rule.nr);
775 			} else {
776 				newblock = block;
777 			}
778 
779 			while ((por = TAILQ_FIRST(&skiplist->ps_rules))) {
780 				TAILQ_REMOVE(&head, por, por_entry);
781 				TAILQ_REMOVE(&skiplist->ps_rules, por,
782 				    por_skip_entry[largest_list]);
783 				TAILQ_INSERT_TAIL(&newblock->sb_rules, por,
784 				    por_entry);
785 
786 				/* Remove this rule from all other skiplists */
787 				remove_from_skipsteps(&block->sb_skipsteps[
788 				    largest_list], block, por, skiplist);
789 			}
790 			free(skiplist);
791 			if (newblock != block)
792 				if (reorder_rules(pf, newblock, depth + 1))
793 					return (1);
794 		}
795 	}
796 
797 done:
798 	for (i = 0; i < PF_SKIP_COUNT; i++) {
799 		while ((skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]))) {
800 			TAILQ_REMOVE(&block->sb_skipsteps[i], skiplist,
801 			    ps_entry);
802 			free(skiplist);
803 		}
804 	}
805 
806 	return (0);
807 }
808 
809 
810 /*
811  * Optimization pass #4: re-order 'quick' rules based on feedback from the
812  * currently running ruleset
813  */
814 int
815 block_feedback(struct pfctl *pf, struct superblock *block)
816 {
817 	TAILQ_HEAD( , pf_opt_rule) queue;
818 	struct pf_opt_rule *por1, *por2;
819 	u_int64_t total_count = 0;
820 	struct pf_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 		total_count += por1->por_rule.packets[0] +
830 		    por1->por_rule.packets[1];
831 		TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
832 			if (por2->por_profile_count)
833 				continue;
834 			comparable_rule(&b, &por2->por_rule, DC);
835 			if (memcmp(&a, &b, sizeof(a)) == 0) {
836 				por2->por_profile_count =
837 				    por1->por_rule.packets[0] +
838 				    por1->por_rule.packets[1];
839 				break;
840 			}
841 		}
842 	}
843 	superblock_free(pf, block->sb_profiled_block);
844 	block->sb_profiled_block = NULL;
845 
846 	/*
847 	 * Now we pull all of the rules off the superblock and re-insert them
848 	 * in sorted order.
849 	 */
850 
851 	TAILQ_INIT(&queue);
852 	while ((por1 = TAILQ_FIRST(&block->sb_rules)) != NULL) {
853 		TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
854 		TAILQ_INSERT_TAIL(&queue, por1, por_entry);
855 	}
856 
857 	while ((por1 = TAILQ_FIRST(&queue)) != NULL) {
858 		TAILQ_REMOVE(&queue, por1, por_entry);
859 /* XXX I should sort all of the unused rules based on skip steps */
860 		TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
861 			if (por1->por_profile_count > por2->por_profile_count) {
862 				TAILQ_INSERT_BEFORE(por2, por1, por_entry);
863 				break;
864 			}
865 		}
866 #ifdef __FreeBSD__
867 		if (por2 == NULL)
868 #else
869 		if (por2 == TAILQ_END(&block->sb_rules))
870 #endif
871 			TAILQ_INSERT_TAIL(&block->sb_rules, por1, por_entry);
872 	}
873 
874 	return (0);
875 }
876 
877 
878 /*
879  * Load the current ruleset from the kernel and try to associate them with
880  * the ruleset we're optimizing.
881  */
882 int
883 load_feedback_profile(struct pfctl *pf, struct superblocks *superblocks)
884 {
885 	struct superblock *block, *blockcur;
886 	struct superblocks prof_superblocks;
887 	struct pf_opt_rule *por;
888 	struct pf_opt_queue queue;
889 	struct pfioc_rule pr;
890 	struct pf_rule a, b;
891 	int nr, mnr;
892 
893 	TAILQ_INIT(&queue);
894 	TAILQ_INIT(&prof_superblocks);
895 
896 	memset(&pr, 0, sizeof(pr));
897 	pr.rule.action = PF_PASS;
898 	if (ioctl(pf->dev, DIOCGETRULES, &pr)) {
899 		warn("DIOCGETRULES");
900 		return (1);
901 	}
902 	mnr = pr.nr;
903 
904 	DEBUG("Loading %d active rules for a feedback profile", mnr);
905 	for (nr = 0; nr < mnr; ++nr) {
906 		struct pf_ruleset *rs;
907 		if ((por = calloc(1, sizeof(*por))) == NULL) {
908 			warn("calloc");
909 			return (1);
910 		}
911 		pr.nr = nr;
912 		if (ioctl(pf->dev, DIOCGETRULE, &pr)) {
913 			warn("DIOCGETRULES");
914 			return (1);
915 		}
916 		memcpy(&por->por_rule, &pr.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, &pr.rule.rpool, nr, pr.ticket,
925 		 *         PF_PASS, pf->anchor) ???
926 		 * ... pfctl_clear_pool(&pr.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 pf_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 pf_rule *a, struct pf_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 pf_rule *a, struct pf_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 pf_rule *a, struct pf_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 != a->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 pf_rule *a, struct pf_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 pf_rule *a, struct pf_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 pf_rule *a, struct pf_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 pf_rule *a, struct pf_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 != a->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 pf_rule *a, struct pf_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 pf_rule *, struct pf_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 pf_rule *p1, struct pf_rule *p2)
1398 {
1399 	struct pf_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 pf_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+%xh",
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 	if (ifname == NULL || !ifname[0])
1504 		return (0);
1505 
1506 	/* Real interfaces must end in a number, interface groups do not */
1507 	if (isdigit(ifname[strlen(ifname) - 1]))
1508 		return (0);
1509 	else
1510 		return (1);
1511 }
1512 
1513 
1514 /*
1515  * Make a rule that can directly compared by memcmp()
1516  */
1517 void
1518 comparable_rule(struct pf_rule *dst, const struct pf_rule *src, int type)
1519 {
1520 	int i;
1521 	/*
1522 	 * To simplify the comparison, we just zero out the fields that are
1523 	 * allowed to be different and then do a simple memcmp()
1524 	 */
1525 	memcpy(dst, src, sizeof(*dst));
1526 	for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++)
1527 		if (pf_rule_desc[i].prf_type >= type) {
1528 #ifdef OPT_DEBUG
1529 			assert(pf_rule_desc[i].prf_type != NEVER ||
1530 			    *(((char *)dst) + pf_rule_desc[i].prf_offset) == 0);
1531 #endif /* OPT_DEBUG */
1532 			memset(((char *)dst) + pf_rule_desc[i].prf_offset, 0,
1533 			    pf_rule_desc[i].prf_size);
1534 		}
1535 }
1536 
1537 
1538 /*
1539  * Remove superset information from two rules so we can directly compare them
1540  * with memcmp()
1541  */
1542 void
1543 exclude_supersets(struct pf_rule *super, struct pf_rule *sub)
1544 {
1545 	if (super->ifname[0] == '\0')
1546 		memset(sub->ifname, 0, sizeof(sub->ifname));
1547 	if (super->direction == PF_INOUT)
1548 		sub->direction = PF_INOUT;
1549 	if ((super->proto == 0 || super->proto == sub->proto) &&
1550 	    super->flags == 0 && super->flagset == 0 && (sub->flags ||
1551 	    sub->flagset)) {
1552 		sub->flags = super->flags;
1553 		sub->flagset = super->flagset;
1554 	}
1555 	if (super->proto == 0)
1556 		sub->proto = 0;
1557 
1558 	if (super->src.port_op == 0) {
1559 		sub->src.port_op = 0;
1560 		sub->src.port[0] = 0;
1561 		sub->src.port[1] = 0;
1562 	}
1563 	if (super->dst.port_op == 0) {
1564 		sub->dst.port_op = 0;
1565 		sub->dst.port[0] = 0;
1566 		sub->dst.port[1] = 0;
1567 	}
1568 
1569 	if (super->src.addr.type == PF_ADDR_ADDRMASK && !super->src.neg &&
1570 	    !sub->src.neg && super->src.addr.v.a.mask.addr32[0] == 0 &&
1571 	    super->src.addr.v.a.mask.addr32[1] == 0 &&
1572 	    super->src.addr.v.a.mask.addr32[2] == 0 &&
1573 	    super->src.addr.v.a.mask.addr32[3] == 0)
1574 		memset(&sub->src.addr, 0, sizeof(sub->src.addr));
1575 	else if (super->src.addr.type == PF_ADDR_ADDRMASK &&
1576 	    sub->src.addr.type == PF_ADDR_ADDRMASK &&
1577 	    super->src.neg == sub->src.neg &&
1578 	    super->af == sub->af &&
1579 	    unmask(&super->src.addr.v.a.mask, super->af) <
1580 	    unmask(&sub->src.addr.v.a.mask, sub->af) &&
1581 	    super->src.addr.v.a.addr.addr32[0] ==
1582 	    (sub->src.addr.v.a.addr.addr32[0] &
1583 	    super->src.addr.v.a.mask.addr32[0]) &&
1584 	    super->src.addr.v.a.addr.addr32[1] ==
1585 	    (sub->src.addr.v.a.addr.addr32[1] &
1586 	    super->src.addr.v.a.mask.addr32[1]) &&
1587 	    super->src.addr.v.a.addr.addr32[2] ==
1588 	    (sub->src.addr.v.a.addr.addr32[2] &
1589 	    super->src.addr.v.a.mask.addr32[2]) &&
1590 	    super->src.addr.v.a.addr.addr32[3] ==
1591 	    (sub->src.addr.v.a.addr.addr32[3] &
1592 	    super->src.addr.v.a.mask.addr32[3])) {
1593 		/* sub->src.addr is a subset of super->src.addr/mask */
1594 		memcpy(&sub->src.addr, &super->src.addr, sizeof(sub->src.addr));
1595 	}
1596 
1597 	if (super->dst.addr.type == PF_ADDR_ADDRMASK && !super->dst.neg &&
1598 	    !sub->dst.neg && super->dst.addr.v.a.mask.addr32[0] == 0 &&
1599 	    super->dst.addr.v.a.mask.addr32[1] == 0 &&
1600 	    super->dst.addr.v.a.mask.addr32[2] == 0 &&
1601 	    super->dst.addr.v.a.mask.addr32[3] == 0)
1602 		memset(&sub->dst.addr, 0, sizeof(sub->dst.addr));
1603 	else if (super->dst.addr.type == PF_ADDR_ADDRMASK &&
1604 	    sub->dst.addr.type == PF_ADDR_ADDRMASK &&
1605 	    super->dst.neg == sub->dst.neg &&
1606 	    super->af == sub->af &&
1607 	    unmask(&super->dst.addr.v.a.mask, super->af) <
1608 	    unmask(&sub->dst.addr.v.a.mask, sub->af) &&
1609 	    super->dst.addr.v.a.addr.addr32[0] ==
1610 	    (sub->dst.addr.v.a.addr.addr32[0] &
1611 	    super->dst.addr.v.a.mask.addr32[0]) &&
1612 	    super->dst.addr.v.a.addr.addr32[1] ==
1613 	    (sub->dst.addr.v.a.addr.addr32[1] &
1614 	    super->dst.addr.v.a.mask.addr32[1]) &&
1615 	    super->dst.addr.v.a.addr.addr32[2] ==
1616 	    (sub->dst.addr.v.a.addr.addr32[2] &
1617 	    super->dst.addr.v.a.mask.addr32[2]) &&
1618 	    super->dst.addr.v.a.addr.addr32[3] ==
1619 	    (sub->dst.addr.v.a.addr.addr32[3] &
1620 	    super->dst.addr.v.a.mask.addr32[3])) {
1621 		/* sub->dst.addr is a subset of super->dst.addr/mask */
1622 		memcpy(&sub->dst.addr, &super->dst.addr, sizeof(sub->dst.addr));
1623 	}
1624 
1625 	if (super->af == 0)
1626 		sub->af = 0;
1627 }
1628 
1629 
1630 void
1631 superblock_free(struct pfctl *pf, struct superblock *block)
1632 {
1633 	struct pf_opt_rule *por;
1634 	while ((por = TAILQ_FIRST(&block->sb_rules))) {
1635 		TAILQ_REMOVE(&block->sb_rules, por, por_entry);
1636 		if (por->por_src_tbl) {
1637 			if (por->por_src_tbl->pt_buf) {
1638 				pfr_buf_clear(por->por_src_tbl->pt_buf);
1639 				free(por->por_src_tbl->pt_buf);
1640 			}
1641 			free(por->por_src_tbl);
1642 		}
1643 		if (por->por_dst_tbl) {
1644 			if (por->por_dst_tbl->pt_buf) {
1645 				pfr_buf_clear(por->por_dst_tbl->pt_buf);
1646 				free(por->por_dst_tbl->pt_buf);
1647 			}
1648 			free(por->por_dst_tbl);
1649 		}
1650 		free(por);
1651 	}
1652 	if (block->sb_profiled_block)
1653 		superblock_free(pf, block->sb_profiled_block);
1654 	free(block);
1655 }
1656 
1657