xref: /openbsd-src/sbin/pfctl/pfctl_optimize.c (revision 50b7afb2c2c0993b0894d4e34bf857cb13ed9c80)
1 /*	$OpenBSD: pfctl_optimize.c,v 1.33 2013/11/22 04:12:48 deraadt Exp $ */
2 
3 /*
4  * Copyright (c) 2004 Mike Frantzen <frantzen@openbsd.org>
5  *
6  * Permission to use, copy, modify, and distribute this software for any
7  * purpose with or without fee is hereby granted, provided that the above
8  * copyright notice and this permission notice appear in all copies.
9  *
10  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
11  * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
12  * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
13  * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
14  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
15  * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
16  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
17  */
18 
19 #include <sys/types.h>
20 #include <sys/ioctl.h>
21 #include <sys/socket.h>
22 
23 #include <net/if.h>
24 #include <net/pfvar.h>
25 
26 #include <netinet/in.h>
27 #include <arpa/inet.h>
28 
29 #include <assert.h>
30 #include <ctype.h>
31 #include <err.h>
32 #include <errno.h>
33 #include <stddef.h>
34 #include <stdio.h>
35 #include <stdlib.h>
36 #include <string.h>
37 
38 #include "pfctl_parser.h"
39 #include "pfctl.h"
40 
41 /* The size at which a table becomes faster than individual rules */
42 #define TABLE_THRESHOLD		6
43 
44 
45 /* #define OPT_DEBUG	1 */
46 #ifdef OPT_DEBUG
47 # define DEBUG(str, v...) \
48 	printf("%s: " str "\n", __FUNCTION__ , ## v)
49 #else
50 # define DEBUG(str, v...) ((void)0)
51 #endif
52 
53 
54 /*
55  * A container that lets us sort a superblock to optimize the skip step jumps
56  */
57 struct pf_skip_step {
58 	int				ps_count;	/* number of items */
59 	TAILQ_HEAD( , pf_opt_rule)	ps_rules;
60 	TAILQ_ENTRY(pf_skip_step)	ps_entry;
61 };
62 
63 
64 /*
65  * A superblock is a block of adjacent rules of similar action.  If there
66  * are five PASS rules in a row, they all become members of a superblock.
67  * Once we have a superblock, we are free to re-order any rules within it
68  * in order to improve performance; if a packet is passed, it doesn't matter
69  * who passed it.
70  */
71 struct superblock {
72 	TAILQ_HEAD( , pf_opt_rule)		 sb_rules;
73 	TAILQ_ENTRY(superblock)			 sb_entry;
74 	struct superblock			*sb_profiled_block;
75 	TAILQ_HEAD(skiplist, pf_skip_step)	 sb_skipsteps[PF_SKIP_COUNT];
76 };
77 TAILQ_HEAD(superblocks, superblock);
78 
79 
80 /*
81  * Description of the PF rule structure.
82  */
83 enum {
84     BARRIER,	/* the presence of the field puts the rule in it's own block */
85     BREAK,	/* the field may not differ between rules in a superblock */
86     NOMERGE,	/* the field may not differ between rules when combined */
87     COMBINED,	/* the field may itself be combined with other rules */
88     DC,		/* we just don't care about the field */
89     NEVER};	/* we should never see this field set?!? */
90 struct pf_rule_field {
91 	const char	*prf_name;
92 	int		 prf_type;
93 	size_t		 prf_offset;
94 	size_t		 prf_size;
95 } pf_rule_desc[] = {
96 #define PF_RULE_FIELD(field, ty)	\
97     {#field,				\
98     ty,					\
99     offsetof(struct pf_rule, field),	\
100     sizeof(((struct pf_rule *)0)->field)}
101 
102 
103     /*
104      * The presence of these fields in a rule put the rule in it's own
105      * superblock.  Thus it will not be optimized.  It also prevents the
106      * rule from being re-ordered at all.
107      */
108     PF_RULE_FIELD(label,		BARRIER),
109     PF_RULE_FIELD(prob,			BARRIER),
110     PF_RULE_FIELD(max_states,		BARRIER),
111     PF_RULE_FIELD(max_src_nodes,	BARRIER),
112     PF_RULE_FIELD(max_src_states,	BARRIER),
113     PF_RULE_FIELD(max_src_conn,		BARRIER),
114     PF_RULE_FIELD(max_src_conn_rate,	BARRIER),
115     PF_RULE_FIELD(anchor,		BARRIER),	/* for now */
116 
117     /*
118      * These fields must be the same between all rules in the same superblock.
119      * These rules are allowed to be re-ordered but only among like rules.
120      * For instance we can re-order all 'tag "foo"' rules because they have the
121      * same tag.  But we can not re-order between a 'tag "foo"' and a
122      * 'tag "bar"' since that would change the meaning of the ruleset.
123      */
124     PF_RULE_FIELD(tagname,		BREAK),
125     PF_RULE_FIELD(keep_state,		BREAK),
126     PF_RULE_FIELD(qname,		BREAK),
127     PF_RULE_FIELD(pqname,		BREAK),
128     PF_RULE_FIELD(rt,			BREAK),
129     PF_RULE_FIELD(allow_opts,		BREAK),
130     PF_RULE_FIELD(rule_flag,		BREAK),
131     PF_RULE_FIELD(action,		BREAK),
132     PF_RULE_FIELD(log,			BREAK),
133     PF_RULE_FIELD(quick,		BREAK),
134     PF_RULE_FIELD(return_ttl,		BREAK),
135     PF_RULE_FIELD(overload_tblname,	BREAK),
136     PF_RULE_FIELD(flush,		BREAK),
137     PF_RULE_FIELD(rdr,			BREAK),
138     PF_RULE_FIELD(nat,			BREAK),
139     PF_RULE_FIELD(logif,		BREAK),
140     PF_RULE_FIELD(route,		BREAK),
141     PF_RULE_FIELD(rtableid,		BREAK),
142 
143     /*
144      * Any fields not listed in this structure act as BREAK fields
145      */
146 
147 
148     /*
149      * These fields must not differ when we merge two rules together but
150      * their difference isn't enough to put the rules in different superblocks.
151      * There are no problems re-ordering any rules with these fields.
152      */
153     PF_RULE_FIELD(af,			NOMERGE),
154     PF_RULE_FIELD(ifnot,		NOMERGE),
155     PF_RULE_FIELD(ifname,		NOMERGE),	/* hack for IF groups */
156     PF_RULE_FIELD(match_tag_not,	NOMERGE),
157     PF_RULE_FIELD(match_tagname,	NOMERGE),
158     PF_RULE_FIELD(os_fingerprint,	NOMERGE),
159     PF_RULE_FIELD(timeout,		NOMERGE),
160     PF_RULE_FIELD(return_icmp,		NOMERGE),
161     PF_RULE_FIELD(return_icmp6,		NOMERGE),
162     PF_RULE_FIELD(uid,			NOMERGE),
163     PF_RULE_FIELD(gid,			NOMERGE),
164     PF_RULE_FIELD(direction,		NOMERGE),
165     PF_RULE_FIELD(proto,		NOMERGE),
166     PF_RULE_FIELD(type,			NOMERGE),
167     PF_RULE_FIELD(code,			NOMERGE),
168     PF_RULE_FIELD(flags,		NOMERGE),
169     PF_RULE_FIELD(flagset,		NOMERGE),
170     PF_RULE_FIELD(tos,			NOMERGE),
171     PF_RULE_FIELD(src.port,		NOMERGE),
172     PF_RULE_FIELD(dst.port,		NOMERGE),
173     PF_RULE_FIELD(src.port_op,		NOMERGE),
174     PF_RULE_FIELD(dst.port_op,		NOMERGE),
175     PF_RULE_FIELD(src.neg,		NOMERGE),
176     PF_RULE_FIELD(dst.neg,		NOMERGE),
177     PF_RULE_FIELD(onrdomain,		NOMERGE),
178     PF_RULE_FIELD(naf,			NOMERGE),
179 
180     /* These fields can be merged */
181     PF_RULE_FIELD(src.addr,		COMBINED),
182     PF_RULE_FIELD(dst.addr,		COMBINED),
183 
184     /* We just don't care about these fields.  They're set by the kernel */
185     PF_RULE_FIELD(skip,			DC),
186     PF_RULE_FIELD(evaluations,		DC),
187     PF_RULE_FIELD(packets,		DC),
188     PF_RULE_FIELD(bytes,		DC),
189     PF_RULE_FIELD(kif,			DC),
190     PF_RULE_FIELD(states_cur,		DC),
191     PF_RULE_FIELD(states_tot,		DC),
192     PF_RULE_FIELD(src_nodes,		DC),
193     PF_RULE_FIELD(nr,			DC),
194     PF_RULE_FIELD(entries,		DC),
195     PF_RULE_FIELD(qid,			DC),
196     PF_RULE_FIELD(pqid,			DC),
197     PF_RULE_FIELD(anchor_relative,	DC),
198     PF_RULE_FIELD(anchor_wildcard,	DC),
199     PF_RULE_FIELD(tag,			DC),
200     PF_RULE_FIELD(match_tag,		DC),
201     PF_RULE_FIELD(overload_tbl,		DC),
202 
203     /* These fields should never be set in a PASS/BLOCK rule XXX fix*/
204     PF_RULE_FIELD(max_mss,		NEVER),
205     PF_RULE_FIELD(min_ttl,		NEVER),
206     PF_RULE_FIELD(set_tos,		NEVER),
207 };
208 
209 
210 
211 int	addrs_combineable(struct pf_rule_addr *, struct pf_rule_addr *);
212 int	addrs_equal(struct pf_rule_addr *, struct pf_rule_addr *);
213 int	block_feedback(struct pfctl *, struct superblock *);
214 int	combine_rules(struct pfctl *, struct superblock *);
215 void	comparable_rule(struct pf_rule *, const struct pf_rule *, int);
216 int	construct_superblocks(struct pfctl *, struct pf_opt_queue *,
217 	    struct superblocks *);
218 void	exclude_supersets(struct pf_rule *, struct pf_rule *);
219 int	interface_group(const char *);
220 int	load_feedback_profile(struct pfctl *, struct superblocks *);
221 int	optimize_superblock(struct pfctl *, struct superblock *);
222 void	remove_from_skipsteps(struct skiplist *, struct superblock *,
223 	    struct pf_opt_rule *, struct pf_skip_step *);
224 int	remove_identical_rules(struct pfctl *, struct superblock *);
225 int	reorder_rules(struct pfctl *, struct superblock *, int);
226 int	rules_combineable(struct pf_rule *, struct pf_rule *);
227 void	skip_append(struct superblock *, int, struct pf_skip_step *,
228 	    struct pf_opt_rule *);
229 int	skip_compare(int, struct pf_skip_step *, struct pf_opt_rule *);
230 void	skip_init(void);
231 int	skip_cmp_af(struct pf_rule *, struct pf_rule *);
232 int	skip_cmp_dir(struct pf_rule *, struct pf_rule *);
233 int	skip_cmp_rdom(struct pf_rule *, struct pf_rule *);
234 int	skip_cmp_dst_addr(struct pf_rule *, struct pf_rule *);
235 int	skip_cmp_dst_port(struct pf_rule *, struct pf_rule *);
236 int	skip_cmp_ifp(struct pf_rule *, struct pf_rule *);
237 int	skip_cmp_proto(struct pf_rule *, struct pf_rule *);
238 int	skip_cmp_src_addr(struct pf_rule *, struct pf_rule *);
239 int	skip_cmp_src_port(struct pf_rule *, struct pf_rule *);
240 int	superblock_inclusive(struct superblock *, struct pf_opt_rule *);
241 void	superblock_free(struct pfctl *, struct superblock *);
242 
243 
244 int (*skip_comparitors[PF_SKIP_COUNT])(struct pf_rule *, struct pf_rule *);
245 const char *skip_comparitors_names[PF_SKIP_COUNT];
246 #define PF_SKIP_COMPARITORS {				\
247     { "ifp", PF_SKIP_IFP, skip_cmp_ifp },		\
248     { "dir", PF_SKIP_DIR, skip_cmp_dir },		\
249     { "rdomain", PF_SKIP_RDOM, skip_cmp_rdom },		\
250     { "af", PF_SKIP_AF, skip_cmp_af },			\
251     { "proto", PF_SKIP_PROTO, skip_cmp_proto },		\
252     { "saddr", PF_SKIP_SRC_ADDR, skip_cmp_src_addr },	\
253     { "daddr", PF_SKIP_DST_ADDR, skip_cmp_dst_addr },	\
254     { "sport", PF_SKIP_SRC_PORT, skip_cmp_src_port },	\
255     { "dport", PF_SKIP_DST_PORT, skip_cmp_dst_port }	\
256 }
257 
258 struct pfr_buffer table_buffer;
259 int table_identifier;
260 
261 
262 int
263 pfctl_optimize_ruleset(struct pfctl *pf, struct pf_ruleset *rs)
264 {
265 	struct superblocks superblocks;
266 	struct pf_opt_queue opt_queue;
267 	struct superblock *block;
268 	struct pf_opt_rule *por;
269 	struct pf_rule *r;
270 	struct pf_rulequeue *old_rules;
271 
272 	DEBUG("optimizing ruleset");
273 	memset(&table_buffer, 0, sizeof(table_buffer));
274 	skip_init();
275 	TAILQ_INIT(&opt_queue);
276 
277 	old_rules = rs->rules.active.ptr;
278 	rs->rules.active.ptr = rs->rules.inactive.ptr;
279 	rs->rules.inactive.ptr = old_rules;
280 
281 	/*
282 	 * XXX expanding the pf_opt_rule format throughout pfctl might allow
283 	 * us to avoid all this copying.
284 	 */
285 	while ((r = TAILQ_FIRST(rs->rules.inactive.ptr)) != NULL) {
286 		TAILQ_REMOVE(rs->rules.inactive.ptr, r, entries);
287 		if ((por = calloc(1, sizeof(*por))) == NULL)
288 			err(1, "calloc");
289 		memcpy(&por->por_rule, r, sizeof(*r));
290 
291 		TAILQ_INSERT_TAIL(&opt_queue, por, por_entry);
292 	}
293 
294 	TAILQ_INIT(&superblocks);
295 	if (construct_superblocks(pf, &opt_queue, &superblocks))
296 		goto error;
297 
298 	if (pf->optimize & PF_OPTIMIZE_PROFILE) {
299 		if (load_feedback_profile(pf, &superblocks))
300 			goto error;
301 	}
302 
303 	TAILQ_FOREACH(block, &superblocks, sb_entry) {
304 		if (optimize_superblock(pf, block))
305 			goto error;
306 	}
307 
308 	rs->anchor->refcnt = 0;
309 	while ((block = TAILQ_FIRST(&superblocks))) {
310 		TAILQ_REMOVE(&superblocks, block, sb_entry);
311 
312 		while ((por = TAILQ_FIRST(&block->sb_rules))) {
313 			TAILQ_REMOVE(&block->sb_rules, por, por_entry);
314 			por->por_rule.nr = rs->anchor->refcnt++;
315 			if ((r = calloc(1, sizeof(*r))) == NULL)
316 				err(1, "calloc");
317 			memcpy(r, &por->por_rule, sizeof(*r));
318 			TAILQ_INSERT_TAIL(rs->rules.active.ptr, r, entries);
319 			free(por);
320 		}
321 		free(block);
322 	}
323 
324 	return (0);
325 
326 error:
327 	while ((por = TAILQ_FIRST(&opt_queue))) {
328 		TAILQ_REMOVE(&opt_queue, por, por_entry);
329 		if (por->por_src_tbl) {
330 			pfr_buf_clear(por->por_src_tbl->pt_buf);
331 			free(por->por_src_tbl->pt_buf);
332 			free(por->por_src_tbl);
333 		}
334 		if (por->por_dst_tbl) {
335 			pfr_buf_clear(por->por_dst_tbl->pt_buf);
336 			free(por->por_dst_tbl->pt_buf);
337 			free(por->por_dst_tbl);
338 		}
339 		free(por);
340 	}
341 	while ((block = TAILQ_FIRST(&superblocks))) {
342 		TAILQ_REMOVE(&superblocks, block, sb_entry);
343 		superblock_free(pf, block);
344 	}
345 	return (1);
346 }
347 
348 
349 /*
350  * Go ahead and optimize a superblock
351  */
352 int
353 optimize_superblock(struct pfctl *pf, struct superblock *block)
354 {
355 #ifdef OPT_DEBUG
356 	struct pf_opt_rule *por;
357 #endif /* OPT_DEBUG */
358 
359 	/* We have a few optimization passes:
360 	 *   1) remove duplicate rules or rules that are a subset of other
361 	 *      rules
362 	 *   2) combine otherwise identical rules with different IP addresses
363 	 *      into a single rule and put the addresses in a table.
364 	 *   3) re-order the rules to improve kernel skip steps
365 	 *   4) re-order the 'quick' rules based on feedback from the
366 	 *      active ruleset statistics
367 	 *
368 	 * XXX combine_rules() doesn't combine v4 and v6 rules.  would just
369 	 *     have to keep af in the table container, make af 'COMBINE' and
370 	 *     twiddle the af on the merged rule
371 	 * XXX maybe add a weighting to the metric on skipsteps when doing
372 	 *     reordering.  sometimes two sequential tables will be better
373 	 *     that four consecutive interfaces.
374 	 * XXX need to adjust the skipstep count of everything after PROTO,
375 	 *     since they aren't actually checked on a proto mismatch in
376 	 *     pf_test_{tcp, udp, icmp}()
377 	 * XXX should i treat proto=0, af=0 or dir=0 special in skepstep
378 	 *     calculation since they are a DC?
379 	 * XXX keep last skiplist of last superblock to influence this
380 	 *     superblock.  '5 inet6 log' should make '3 inet6' come before '4
381 	 *     inet' in the next superblock.
382 	 * XXX would be useful to add tables for ports
383 	 * XXX we can also re-order some mutually exclusive superblocks to
384 	 *     try merging superblocks before any of these optimization passes.
385 	 *     for instance a single 'log in' rule in the middle of non-logging
386 	 *     out rules.
387 	 */
388 
389 	/* shortcut.  there will be a lot of 1-rule superblocks */
390 	if (!TAILQ_NEXT(TAILQ_FIRST(&block->sb_rules), por_entry))
391 		return (0);
392 
393 #ifdef OPT_DEBUG
394 	printf("--- Superblock ---\n");
395 	TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
396 		printf("  ");
397 		print_rule(&por->por_rule, por->por_rule.anchor ?
398 		    por->por_rule.anchor->name : "", PF_OPT_DEBUG);
399 	}
400 #endif /* OPT_DEBUG */
401 
402 
403 	if (remove_identical_rules(pf, block))
404 		return (1);
405 	if (combine_rules(pf, block))
406 		return (1);
407 	if ((pf->optimize & PF_OPTIMIZE_PROFILE) &&
408 	    TAILQ_FIRST(&block->sb_rules)->por_rule.quick &&
409 	    block->sb_profiled_block) {
410 		if (block_feedback(pf, block))
411 			return (1);
412 	} else if (reorder_rules(pf, block, 0)) {
413 		return (1);
414 	}
415 
416 	/*
417 	 * Don't add any optimization passes below reorder_rules().  It will
418 	 * have divided superblocks into smaller blocks for further refinement
419 	 * and doesn't put them back together again.  What once was a true
420 	 * superblock might have been split into multiple superblocks.
421 	 */
422 
423 #ifdef OPT_DEBUG
424 	printf("--- END Superblock ---\n");
425 #endif /* OPT_DEBUG */
426 	return (0);
427 }
428 
429 
430 /*
431  * Optimization pass #1: remove identical rules
432  */
433 int
434 remove_identical_rules(struct pfctl *pf, struct superblock *block)
435 {
436 	struct pf_opt_rule *por1, *por2, *por_next, *por2_next;
437 	struct pf_rule a, a2, b, b2;
438 
439 	for (por1 = TAILQ_FIRST(&block->sb_rules); por1; por1 = por_next) {
440 		por_next = TAILQ_NEXT(por1, por_entry);
441 		for (por2 = por_next; por2; por2 = por2_next) {
442 			por2_next = TAILQ_NEXT(por2, por_entry);
443 			comparable_rule(&a, &por1->por_rule, DC);
444 			comparable_rule(&b, &por2->por_rule, DC);
445 			memcpy(&a2, &a, sizeof(a2));
446 			memcpy(&b2, &b, sizeof(b2));
447 
448 			exclude_supersets(&a, &b);
449 			exclude_supersets(&b2, &a2);
450 			if (memcmp(&a, &b, sizeof(a)) == 0) {
451 				DEBUG("removing identical rule  nr%d = *nr%d*",
452 				    por1->por_rule.nr, por2->por_rule.nr);
453 				TAILQ_REMOVE(&block->sb_rules, por2, por_entry);
454 				if (por_next == por2)
455 					por_next = TAILQ_NEXT(por1, por_entry);
456 				free(por2);
457 			} else if (memcmp(&a2, &b2, sizeof(a2)) == 0) {
458 				DEBUG("removing identical rule  *nr%d* = nr%d",
459 				    por1->por_rule.nr, por2->por_rule.nr);
460 				TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
461 				free(por1);
462 				break;
463 			}
464 		}
465 	}
466 
467 	return (0);
468 }
469 
470 
471 /*
472  * Optimization pass #2: combine similar rules with different addresses
473  * into a single rule and a table
474  */
475 int
476 combine_rules(struct pfctl *pf, struct superblock *block)
477 {
478 	struct pf_opt_rule *p1, *p2, *por_next;
479 	int src_eq, dst_eq;
480 
481 	/* First we make a pass to combine the rules.  O(n log n) */
482 	TAILQ_FOREACH(p1, &block->sb_rules, por_entry) {
483 		for (p2 = TAILQ_NEXT(p1, por_entry); p2; p2 = por_next) {
484 			por_next = TAILQ_NEXT(p2, por_entry);
485 
486 			src_eq = addrs_equal(&p1->por_rule.src,
487 			    &p2->por_rule.src);
488 			dst_eq = addrs_equal(&p1->por_rule.dst,
489 			    &p2->por_rule.dst);
490 
491 			if (src_eq && !dst_eq && p1->por_src_tbl == NULL &&
492 			    p2->por_dst_tbl == NULL &&
493 			    p2->por_src_tbl == NULL &&
494 			    rules_combineable(&p1->por_rule, &p2->por_rule) &&
495 			    addrs_combineable(&p1->por_rule.dst,
496 			    &p2->por_rule.dst)) {
497 				DEBUG("can combine rules  nr%d = nr%d",
498 				    p1->por_rule.nr, p2->por_rule.nr);
499 				if (p1->por_dst_tbl == NULL &&
500 				    add_opt_table(pf, &p1->por_dst_tbl,
501 				    p1->por_rule.af, &p1->por_rule.dst, NULL))
502 					return (1);
503 				if (add_opt_table(pf, &p1->por_dst_tbl,
504 				    p1->por_rule.af, &p2->por_rule.dst, NULL))
505 					return (1);
506 				p2->por_dst_tbl = p1->por_dst_tbl;
507 				if (p1->por_dst_tbl->pt_rulecount >=
508 				    TABLE_THRESHOLD) {
509 					TAILQ_REMOVE(&block->sb_rules, p2,
510 					    por_entry);
511 					free(p2);
512 				}
513 			} else if (!src_eq && dst_eq && p1->por_dst_tbl == NULL
514 			    && p2->por_src_tbl == NULL &&
515 			    p2->por_dst_tbl == NULL &&
516 			    rules_combineable(&p1->por_rule, &p2->por_rule) &&
517 			    addrs_combineable(&p1->por_rule.src,
518 			    &p2->por_rule.src)) {
519 				DEBUG("can combine rules  nr%d = nr%d",
520 				    p1->por_rule.nr, p2->por_rule.nr);
521 				if (p1->por_src_tbl == NULL &&
522 				    add_opt_table(pf, &p1->por_src_tbl,
523 				    p1->por_rule.af, &p1->por_rule.src, NULL))
524 					return (1);
525 				if (add_opt_table(pf, &p1->por_src_tbl,
526 				    p1->por_rule.af, &p2->por_rule.src, NULL))
527 					return (1);
528 				p2->por_src_tbl = p1->por_src_tbl;
529 				if (p1->por_src_tbl->pt_rulecount >=
530 				    TABLE_THRESHOLD) {
531 					TAILQ_REMOVE(&block->sb_rules, p2,
532 					    por_entry);
533 					free(p2);
534 				}
535 			}
536 		}
537 	}
538 
539 
540 	/*
541 	 * Then we make a final pass to create a valid table name and
542 	 * insert the name into the rules.
543 	 * Convert translation/routing mapping pools to tables as well.
544 	 */
545 	for (p1 = TAILQ_FIRST(&block->sb_rules); p1; p1 = por_next) {
546 		por_next = TAILQ_NEXT(p1, por_entry);
547 		assert(p1->por_src_tbl == NULL || p1->por_dst_tbl == NULL);
548 
549 		if (p1->por_src_tbl && p1->por_src_tbl->pt_rulecount >=
550 		    TABLE_THRESHOLD) {
551 			if (p1->por_src_tbl->pt_generated) {
552 				/* This rule is included in a table */
553 				TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
554 				free(p1);
555 				continue;
556 			}
557 			p1->por_src_tbl->pt_generated = 1;
558 
559 			if ((pf->opts & PF_OPT_NOACTION) == 0 &&
560 			    pf_opt_create_table(pf, p1->por_src_tbl))
561 				return (1);
562 
563 			pf->tdirty = 1;
564 
565 			if (pf->opts & PF_OPT_VERBOSE)
566 				print_tabledef(p1->por_src_tbl->pt_name,
567 				    PFR_TFLAG_CONST, 1,
568 				    &p1->por_src_tbl->pt_nodes);
569 
570 			memset(&p1->por_rule.src.addr, 0,
571 			    sizeof(p1->por_rule.src.addr));
572 			p1->por_rule.src.addr.type = PF_ADDR_TABLE;
573 			strlcpy(p1->por_rule.src.addr.v.tblname,
574 			    p1->por_src_tbl->pt_name,
575 			    sizeof(p1->por_rule.src.addr.v.tblname));
576 
577 			pfr_buf_clear(p1->por_src_tbl->pt_buf);
578 			free(p1->por_src_tbl->pt_buf);
579 			p1->por_src_tbl->pt_buf = NULL;
580 		}
581 		if (p1->por_dst_tbl && p1->por_dst_tbl->pt_rulecount >=
582 		    TABLE_THRESHOLD) {
583 			if (p1->por_dst_tbl->pt_generated) {
584 				/* This rule is included in a table */
585 				TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
586 				free(p1);
587 				continue;
588 			}
589 			p1->por_dst_tbl->pt_generated = 1;
590 
591 			if ((pf->opts & PF_OPT_NOACTION) == 0 &&
592 			    pf_opt_create_table(pf, p1->por_dst_tbl))
593 				return (1);
594 			pf->tdirty = 1;
595 
596 			if (pf->opts & PF_OPT_VERBOSE)
597 				print_tabledef(p1->por_dst_tbl->pt_name,
598 				    PFR_TFLAG_CONST, 1,
599 				    &p1->por_dst_tbl->pt_nodes);
600 
601 			memset(&p1->por_rule.dst.addr, 0,
602 			    sizeof(p1->por_rule.dst.addr));
603 			p1->por_rule.dst.addr.type = PF_ADDR_TABLE;
604 			strlcpy(p1->por_rule.dst.addr.v.tblname,
605 			    p1->por_dst_tbl->pt_name,
606 			    sizeof(p1->por_rule.dst.addr.v.tblname));
607 
608 			pfr_buf_clear(p1->por_dst_tbl->pt_buf);
609 			free(p1->por_dst_tbl->pt_buf);
610 			p1->por_dst_tbl->pt_buf = NULL;
611 		}
612 	}
613 
614 	return (0);
615 }
616 
617 
618 /*
619  * Optimization pass #3: re-order rules to improve skip steps
620  */
621 int
622 reorder_rules(struct pfctl *pf, struct superblock *block, int depth)
623 {
624 	struct superblock *newblock;
625 	struct pf_skip_step *skiplist;
626 	struct pf_opt_rule *por;
627 	int i, largest, largest_list, rule_count = 0;
628 	TAILQ_HEAD( , pf_opt_rule) head;
629 
630 	/*
631 	 * Calculate the best-case skip steps.  We put each rule in a list
632 	 * of other rules with common fields
633 	 */
634 	for (i = 0; i < PF_SKIP_COUNT; i++) {
635 		TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
636 			TAILQ_FOREACH(skiplist, &block->sb_skipsteps[i],
637 			    ps_entry) {
638 				if (skip_compare(i, skiplist, por) == 0)
639 					break;
640 			}
641 			if (skiplist == NULL) {
642 				if ((skiplist = calloc(1, sizeof(*skiplist))) ==
643 				    NULL)
644 					err(1, "calloc");
645 				TAILQ_INIT(&skiplist->ps_rules);
646 				TAILQ_INSERT_TAIL(&block->sb_skipsteps[i],
647 				    skiplist, ps_entry);
648 			}
649 			skip_append(block, i, skiplist, por);
650 		}
651 	}
652 
653 	TAILQ_FOREACH(por, &block->sb_rules, por_entry)
654 		rule_count++;
655 
656 	/*
657 	 * Now we're going to ignore any fields that are identical between
658 	 * all of the rules in the superblock and those fields which differ
659 	 * between every rule in the superblock.
660 	 */
661 	largest = 0;
662 	for (i = 0; i < PF_SKIP_COUNT; i++) {
663 		skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
664 		if (skiplist->ps_count == rule_count) {
665 			DEBUG("(%d) original skipstep '%s' is all rules",
666 			    depth, skip_comparitors_names[i]);
667 			skiplist->ps_count = 0;
668 		} else if (skiplist->ps_count == 1) {
669 			skiplist->ps_count = 0;
670 		} else {
671 			DEBUG("(%d) original skipstep '%s' largest jump is %d",
672 			    depth, skip_comparitors_names[i],
673 			    skiplist->ps_count);
674 			if (skiplist->ps_count > largest)
675 				largest = skiplist->ps_count;
676 		}
677 	}
678 	if (largest == 0) {
679 		/* Ugh.  There is NO commonality in the superblock on which
680 		 * optimize the skipsteps optimization.
681 		 */
682 		goto done;
683 	}
684 
685 	/*
686 	 * Now we're going to empty the superblock rule list and re-create
687 	 * it based on a more optimal skipstep order.
688 	 */
689 	TAILQ_INIT(&head);
690 	while ((por = TAILQ_FIRST(&block->sb_rules))) {
691 		TAILQ_REMOVE(&block->sb_rules, por, por_entry);
692 		TAILQ_INSERT_TAIL(&head, por, por_entry);
693 	}
694 
695 
696 	while (!TAILQ_EMPTY(&head)) {
697 		largest = 1;
698 
699 		/*
700 		 * Find the most useful skip steps remaining
701 		 */
702 		for (i = 0; i < PF_SKIP_COUNT; i++) {
703 			skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
704 			if (skiplist->ps_count > largest) {
705 				largest = skiplist->ps_count;
706 				largest_list = i;
707 			}
708 		}
709 
710 		if (largest <= 1) {
711 			/*
712 			 * Nothing useful left.  Leave remaining rules in order.
713 			 */
714 			DEBUG("(%d) no more commonality for skip steps", depth);
715 			while ((por = TAILQ_FIRST(&head))) {
716 				TAILQ_REMOVE(&head, por, por_entry);
717 				TAILQ_INSERT_TAIL(&block->sb_rules, por,
718 				    por_entry);
719 			}
720 		} else {
721 			/*
722 			 * There is commonality.  Extract those common rules
723 			 * and place them in the ruleset adjacent to each
724 			 * other.
725 			 */
726 			skiplist = TAILQ_FIRST(&block->sb_skipsteps[
727 			    largest_list]);
728 			DEBUG("(%d) skipstep '%s' largest jump is %d @ #%d",
729 			    depth, skip_comparitors_names[largest_list],
730 			    largest, TAILQ_FIRST(&TAILQ_FIRST(&block->
731 			    sb_skipsteps [largest_list])->ps_rules)->
732 			    por_rule.nr);
733 			TAILQ_REMOVE(&block->sb_skipsteps[largest_list],
734 			    skiplist, ps_entry);
735 
736 
737 			/*
738 			 * There may be further commonality inside these
739 			 * rules.  So we'll split them off into they're own
740 			 * superblock and pass it back into the optimizer.
741 			 */
742 			if (skiplist->ps_count > 2) {
743 				if ((newblock = calloc(1, sizeof(*newblock)))
744 				    == NULL) {
745 					warn("calloc");
746 					return (1);
747 				}
748 				TAILQ_INIT(&newblock->sb_rules);
749 				for (i = 0; i < PF_SKIP_COUNT; i++)
750 					TAILQ_INIT(&newblock->sb_skipsteps[i]);
751 				TAILQ_INSERT_BEFORE(block, newblock, sb_entry);
752 				DEBUG("(%d) splitting off %d rules from superblock @ #%d",
753 				    depth, skiplist->ps_count,
754 				    TAILQ_FIRST(&skiplist->ps_rules)->
755 				    por_rule.nr);
756 			} else {
757 				newblock = block;
758 			}
759 
760 			while ((por = TAILQ_FIRST(&skiplist->ps_rules))) {
761 				TAILQ_REMOVE(&head, por, por_entry);
762 				TAILQ_REMOVE(&skiplist->ps_rules, por,
763 				    por_skip_entry[largest_list]);
764 				TAILQ_INSERT_TAIL(&newblock->sb_rules, por,
765 				    por_entry);
766 
767 				/* Remove this rule from all other skiplists */
768 				remove_from_skipsteps(&block->sb_skipsteps[
769 				    largest_list], block, por, skiplist);
770 			}
771 			free(skiplist);
772 			if (newblock != block)
773 				if (reorder_rules(pf, newblock, depth + 1))
774 					return (1);
775 		}
776 	}
777 
778 done:
779 	for (i = 0; i < PF_SKIP_COUNT; i++) {
780 		while ((skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]))) {
781 			TAILQ_REMOVE(&block->sb_skipsteps[i], skiplist,
782 			    ps_entry);
783 			free(skiplist);
784 		}
785 	}
786 
787 	return (0);
788 }
789 
790 
791 /*
792  * Optimization pass #4: re-order 'quick' rules based on feedback from the
793  * currently running ruleset
794  */
795 int
796 block_feedback(struct pfctl *pf, struct superblock *block)
797 {
798 	TAILQ_HEAD( , pf_opt_rule) queue;
799 	struct pf_opt_rule *por1, *por2;
800 	u_int64_t total_count = 0;
801 	struct pf_rule a, b;
802 
803 
804 	/*
805 	 * Walk through all of the profiled superblock's rules and copy
806 	 * the counters onto our rules.
807 	 */
808 	TAILQ_FOREACH(por1, &block->sb_profiled_block->sb_rules, por_entry) {
809 		comparable_rule(&a, &por1->por_rule, DC);
810 		total_count += por1->por_rule.packets[0] +
811 		    por1->por_rule.packets[1];
812 		TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
813 			if (por2->por_profile_count)
814 				continue;
815 			comparable_rule(&b, &por2->por_rule, DC);
816 			if (memcmp(&a, &b, sizeof(a)) == 0) {
817 				por2->por_profile_count =
818 				    por1->por_rule.packets[0] +
819 				    por1->por_rule.packets[1];
820 				break;
821 			}
822 		}
823 	}
824 	superblock_free(pf, block->sb_profiled_block);
825 	block->sb_profiled_block = NULL;
826 
827 	/*
828 	 * Now we pull all of the rules off the superblock and re-insert them
829 	 * in sorted order.
830 	 */
831 
832 	TAILQ_INIT(&queue);
833 	while ((por1 = TAILQ_FIRST(&block->sb_rules)) != NULL) {
834 		TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
835 		TAILQ_INSERT_TAIL(&queue, por1, por_entry);
836 	}
837 
838 	while ((por1 = TAILQ_FIRST(&queue)) != NULL) {
839 		TAILQ_REMOVE(&queue, por1, por_entry);
840 /* XXX I should sort all of the unused rules based on skip steps */
841 		TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
842 			if (por1->por_profile_count > por2->por_profile_count) {
843 				TAILQ_INSERT_BEFORE(por2, por1, por_entry);
844 				break;
845 			}
846 		}
847 		if (por2 == TAILQ_END(&block->sb_rules))
848 			TAILQ_INSERT_TAIL(&block->sb_rules, por1, por_entry);
849 	}
850 
851 	return (0);
852 }
853 
854 
855 /*
856  * Load the current ruleset from the kernel and try to associate them with
857  * the ruleset we're optimizing.
858  */
859 int
860 load_feedback_profile(struct pfctl *pf, struct superblocks *superblocks)
861 {
862 	struct superblock *block, *blockcur;
863 	struct superblocks prof_superblocks;
864 	struct pf_opt_rule *por;
865 	struct pf_opt_queue queue;
866 	struct pfioc_rule pr;
867 	struct pf_rule a, b;
868 	int nr, mnr;
869 
870 	TAILQ_INIT(&queue);
871 	TAILQ_INIT(&prof_superblocks);
872 
873 	memset(&pr, 0, sizeof(pr));
874 	pr.rule.action = PF_PASS;
875 	if (ioctl(pf->dev, DIOCGETRULES, &pr)) {
876 		warn("DIOCGETRULES");
877 		return (1);
878 	}
879 	mnr = pr.nr;
880 
881 	DEBUG("Loading %d active rules for a feedback profile", mnr);
882 	for (nr = 0; nr < mnr; ++nr) {
883 		struct pf_ruleset *rs;
884 		if ((por = calloc(1, sizeof(*por))) == NULL) {
885 			warn("calloc");
886 			return (1);
887 		}
888 		pr.nr = nr;
889 		if (ioctl(pf->dev, DIOCGETRULE, &pr)) {
890 			warn("DIOCGETRULES");
891 			free(por);
892 			return (1);
893 		}
894 		memcpy(&por->por_rule, &pr.rule, sizeof(por->por_rule));
895 		rs = pf_find_or_create_ruleset(pr.anchor_call);
896 		por->por_rule.anchor = rs->anchor;
897 		TAILQ_INSERT_TAIL(&queue, por, por_entry);
898 
899 		/* XXX pfctl_get_pool(pf->dev, &pr.rule.rpool, nr, pr.ticket,
900 		 *         PF_PASS, pf->anchor) ???
901 		 * ... pfctl_clear_pool(&pr.rule.rpool)
902 		 */
903 	}
904 
905 	if (construct_superblocks(pf, &queue, &prof_superblocks))
906 		return (1);
907 
908 
909 	/*
910 	 * Now we try to associate the active ruleset's superblocks with
911 	 * the superblocks we're compiling.
912 	 */
913 	block = TAILQ_FIRST(superblocks);
914 	blockcur = TAILQ_FIRST(&prof_superblocks);
915 	while (block && blockcur) {
916 		comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule,
917 		    BREAK);
918 		comparable_rule(&b, &TAILQ_FIRST(&blockcur->sb_rules)->por_rule,
919 		    BREAK);
920 		if (memcmp(&a, &b, sizeof(a)) == 0) {
921 			/* The two superblocks lined up */
922 			block->sb_profiled_block = blockcur;
923 		} else {
924 			DEBUG("superblocks don't line up between #%d and #%d",
925 			    TAILQ_FIRST(&block->sb_rules)->por_rule.nr,
926 			    TAILQ_FIRST(&blockcur->sb_rules)->por_rule.nr);
927 			break;
928 		}
929 		block = TAILQ_NEXT(block, sb_entry);
930 		blockcur = TAILQ_NEXT(blockcur, sb_entry);
931 	}
932 
933 
934 
935 	/* Free any superblocks we couldn't link */
936 	while (blockcur) {
937 		block = TAILQ_NEXT(blockcur, sb_entry);
938 		superblock_free(pf, blockcur);
939 		blockcur = block;
940 	}
941 	return (0);
942 }
943 
944 
945 /*
946  * Compare a rule to a skiplist to see if the rule is a member
947  */
948 int
949 skip_compare(int skipnum, struct pf_skip_step *skiplist,
950     struct pf_opt_rule *por)
951 {
952 	struct pf_rule *a, *b;
953 	if (skipnum >= PF_SKIP_COUNT || skipnum < 0)
954 		errx(1, "skip_compare() out of bounds");
955 	a = &por->por_rule;
956 	b = &TAILQ_FIRST(&skiplist->ps_rules)->por_rule;
957 
958 	return ((skip_comparitors[skipnum])(a, b));
959 }
960 
961 
962 /*
963  * Add a rule to a skiplist
964  */
965 void
966 skip_append(struct superblock *superblock, int skipnum,
967     struct pf_skip_step *skiplist, struct pf_opt_rule *por)
968 {
969 	struct pf_skip_step *prev;
970 
971 	skiplist->ps_count++;
972 	TAILQ_INSERT_TAIL(&skiplist->ps_rules, por, por_skip_entry[skipnum]);
973 
974 	/* Keep the list of skiplists sorted by whichever is larger */
975 	while ((prev = TAILQ_PREV(skiplist, skiplist, ps_entry)) &&
976 	    prev->ps_count < skiplist->ps_count) {
977 		TAILQ_REMOVE(&superblock->sb_skipsteps[skipnum],
978 		    skiplist, ps_entry);
979 		TAILQ_INSERT_BEFORE(prev, skiplist, ps_entry);
980 	}
981 }
982 
983 
984 /*
985  * Remove a rule from the other skiplist calculations.
986  */
987 void
988 remove_from_skipsteps(struct skiplist *head, struct superblock *block,
989     struct pf_opt_rule *por, struct pf_skip_step *active_list)
990 {
991 	struct pf_skip_step *sk, *next;
992 	struct pf_opt_rule *p2;
993 	int i, found;
994 
995 	for (i = 0; i < PF_SKIP_COUNT; i++) {
996 		sk = TAILQ_FIRST(&block->sb_skipsteps[i]);
997 		if (sk == NULL || sk == active_list || sk->ps_count <= 1)
998 			continue;
999 		found = 0;
1000 		do {
1001 			TAILQ_FOREACH(p2, &sk->ps_rules, por_skip_entry[i])
1002 				if (p2 == por) {
1003 					TAILQ_REMOVE(&sk->ps_rules, p2,
1004 					    por_skip_entry[i]);
1005 					found = 1;
1006 					sk->ps_count--;
1007 					break;
1008 				}
1009 		} while (!found && (sk = TAILQ_NEXT(sk, ps_entry)));
1010 		if (found && sk) {
1011 			/* Does this change the sorting order? */
1012 			while ((next = TAILQ_NEXT(sk, ps_entry)) &&
1013 			    next->ps_count > sk->ps_count) {
1014 				TAILQ_REMOVE(head, sk, ps_entry);
1015 				TAILQ_INSERT_AFTER(head, next, sk, ps_entry);
1016 			}
1017 #ifdef OPT_DEBUG
1018 			next = TAILQ_NEXT(sk, ps_entry);
1019 			assert(next == NULL || next->ps_count <= sk->ps_count);
1020 #endif /* OPT_DEBUG */
1021 		}
1022 	}
1023 }
1024 
1025 
1026 /* Compare two rules AF field for skiplist construction */
1027 int
1028 skip_cmp_af(struct pf_rule *a, struct pf_rule *b)
1029 {
1030 	if (a->af != b->af || a->af == 0)
1031 		return (1);
1032 	return (0);
1033 }
1034 
1035 /* Compare two rules DIRECTION field for skiplist construction */
1036 int
1037 skip_cmp_dir(struct pf_rule *a, struct pf_rule *b)
1038 {
1039 	if (a->direction == 0 || a->direction != b->direction)
1040 		return (1);
1041 	return (0);
1042 }
1043 
1044 /* Compare two rules ON RDOMAIN field for skiplist construction */
1045 int
1046 skip_cmp_rdom(struct pf_rule *a, struct pf_rule *b)
1047 {
1048 	if (a->onrdomain == -1 || a->onrdomain != b->onrdomain)
1049 		return (1);
1050 	return (a->ifnot != b->ifnot);
1051 }
1052 
1053 /* Compare two rules DST Address field for skiplist construction */
1054 int
1055 skip_cmp_dst_addr(struct pf_rule *a, struct pf_rule *b)
1056 {
1057 	if (a->dst.neg != b->dst.neg ||
1058 	    a->dst.addr.type != b->dst.addr.type)
1059 		return (1);
1060 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1061 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1062 	 *    a->proto == IPPROTO_ICMP
1063 	 *	return (1);
1064 	 */
1065 	switch (a->dst.addr.type) {
1066 	case PF_ADDR_ADDRMASK:
1067 		if (memcmp(&a->dst.addr.v.a.addr, &b->dst.addr.v.a.addr,
1068 		    sizeof(a->dst.addr.v.a.addr)) ||
1069 		    memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask,
1070 		    sizeof(a->dst.addr.v.a.mask)) ||
1071 		    (a->dst.addr.v.a.addr.addr32[0] == 0 &&
1072 		    a->dst.addr.v.a.addr.addr32[1] == 0 &&
1073 		    a->dst.addr.v.a.addr.addr32[2] == 0 &&
1074 		    a->dst.addr.v.a.addr.addr32[3] == 0))
1075 			return (1);
1076 		return (0);
1077 	case PF_ADDR_DYNIFTL:
1078 		if (strcmp(a->dst.addr.v.ifname, b->dst.addr.v.ifname) != 0 ||
1079 		    a->dst.addr.iflags != a->dst.addr.iflags ||
1080 		    memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask,
1081 		    sizeof(a->dst.addr.v.a.mask)))
1082 			return (1);
1083 		return (0);
1084 	case PF_ADDR_NOROUTE:
1085 	case PF_ADDR_URPFFAILED:
1086 		return (0);
1087 	case PF_ADDR_TABLE:
1088 		return (strcmp(a->dst.addr.v.tblname, b->dst.addr.v.tblname));
1089 	}
1090 	return (1);
1091 }
1092 
1093 /* Compare two rules DST port field for skiplist construction */
1094 int
1095 skip_cmp_dst_port(struct pf_rule *a, struct pf_rule *b)
1096 {
1097 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1098 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1099 	 *    a->proto == IPPROTO_ICMP
1100 	 *	return (1);
1101 	 */
1102 	if (a->dst.port_op == PF_OP_NONE || a->dst.port_op != b->dst.port_op ||
1103 	    a->dst.port[0] != b->dst.port[0] ||
1104 	    a->dst.port[1] != b->dst.port[1])
1105 		return (1);
1106 	return (0);
1107 }
1108 
1109 /* Compare two rules IFP field for skiplist construction */
1110 int
1111 skip_cmp_ifp(struct pf_rule *a, struct pf_rule *b)
1112 {
1113 	if (strcmp(a->ifname, b->ifname) || a->ifname[0] == '\0')
1114 		return (1);
1115 	return (a->ifnot != b->ifnot);
1116 }
1117 
1118 /* Compare two rules PROTO field for skiplist construction */
1119 int
1120 skip_cmp_proto(struct pf_rule *a, struct pf_rule *b)
1121 {
1122 	return (a->proto != b->proto || a->proto == 0);
1123 }
1124 
1125 /* Compare two rules SRC addr field for skiplist construction */
1126 int
1127 skip_cmp_src_addr(struct pf_rule *a, struct pf_rule *b)
1128 {
1129 	if (a->src.neg != b->src.neg ||
1130 	    a->src.addr.type != b->src.addr.type)
1131 		return (1);
1132 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1133 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1134 	 *    a->proto == IPPROTO_ICMP
1135 	 *	return (1);
1136 	 */
1137 	switch (a->src.addr.type) {
1138 	case PF_ADDR_ADDRMASK:
1139 		if (memcmp(&a->src.addr.v.a.addr, &b->src.addr.v.a.addr,
1140 		    sizeof(a->src.addr.v.a.addr)) ||
1141 		    memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask,
1142 		    sizeof(a->src.addr.v.a.mask)) ||
1143 		    (a->src.addr.v.a.addr.addr32[0] == 0 &&
1144 		    a->src.addr.v.a.addr.addr32[1] == 0 &&
1145 		    a->src.addr.v.a.addr.addr32[2] == 0 &&
1146 		    a->src.addr.v.a.addr.addr32[3] == 0))
1147 			return (1);
1148 		return (0);
1149 	case PF_ADDR_DYNIFTL:
1150 		if (strcmp(a->src.addr.v.ifname, b->src.addr.v.ifname) != 0 ||
1151 		    a->src.addr.iflags != a->src.addr.iflags ||
1152 		    memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask,
1153 		    sizeof(a->src.addr.v.a.mask)))
1154 			return (1);
1155 		return (0);
1156 	case PF_ADDR_NOROUTE:
1157 	case PF_ADDR_URPFFAILED:
1158 		return (0);
1159 	case PF_ADDR_TABLE:
1160 		return (strcmp(a->src.addr.v.tblname, b->src.addr.v.tblname));
1161 	}
1162 	return (1);
1163 }
1164 
1165 /* Compare two rules SRC port field for skiplist construction */
1166 int
1167 skip_cmp_src_port(struct pf_rule *a, struct pf_rule *b)
1168 {
1169 	if (a->src.port_op == PF_OP_NONE || a->src.port_op != b->src.port_op ||
1170 	    a->src.port[0] != b->src.port[0] ||
1171 	    a->src.port[1] != b->src.port[1])
1172 		return (1);
1173 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1174 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1175 	 *    a->proto == IPPROTO_ICMP
1176 	 *	return (1);
1177 	 */
1178 	return (0);
1179 }
1180 
1181 
1182 void
1183 skip_init(void)
1184 {
1185 	struct {
1186 		char *name;
1187 		int skipnum;
1188 		int (*func)(struct pf_rule *, struct pf_rule *);
1189 	} comps[] = PF_SKIP_COMPARITORS;
1190 	int skipnum, i;
1191 
1192 	for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++) {
1193 		for (i = 0; i < sizeof(comps)/sizeof(*comps); i++)
1194 			if (comps[i].skipnum == skipnum) {
1195 				skip_comparitors[skipnum] = comps[i].func;
1196 				skip_comparitors_names[skipnum] = comps[i].name;
1197 			}
1198 	}
1199 	for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++)
1200 		if (skip_comparitors[skipnum] == NULL)
1201 			errx(1, "Need to add skip step comparitor to pfctl?!");
1202 }
1203 
1204 /*
1205  * Add a host/netmask to a table
1206  */
1207 int
1208 add_opt_table(struct pfctl *pf, struct pf_opt_tbl **tbl, sa_family_t af,
1209     struct pf_rule_addr *addr, char *ifname)
1210 {
1211 #ifdef OPT_DEBUG
1212 	char buf[128];
1213 #endif /* OPT_DEBUG */
1214 	static int tablenum = 0;
1215 	struct node_host node_host;
1216 
1217 	if (*tbl == NULL) {
1218 		if ((*tbl = calloc(1, sizeof(**tbl))) == NULL ||
1219 		    ((*tbl)->pt_buf = calloc(1, sizeof(*(*tbl)->pt_buf))) ==
1220 		    NULL)
1221 			err(1, "calloc");
1222 		(*tbl)->pt_buf->pfrb_type = PFRB_ADDRS;
1223 		SIMPLEQ_INIT(&(*tbl)->pt_nodes);
1224 
1225 		/* This is just a temporary table name */
1226 		snprintf((*tbl)->pt_name, sizeof((*tbl)->pt_name), "%s%d",
1227 		    PF_OPT_TABLE_PREFIX, tablenum++);
1228 		DEBUG("creating table <%s>", (*tbl)->pt_name);
1229 	}
1230 
1231 	memset(&node_host, 0, sizeof(node_host));
1232 	node_host.af = af;
1233 	node_host.addr = addr->addr;
1234 	node_host.ifname = ifname;
1235 	node_host.weight = addr->weight;
1236 
1237 #ifdef OPT_DEBUG
1238 	DEBUG("<%s> adding %s/%d", (*tbl)->pt_name, inet_ntop(af,
1239 	    &node_host.addr.v.a.addr, buf, sizeof(buf)),
1240 	    unmask(&node_host.addr.v.a.mask, af));
1241 #endif /* OPT_DEBUG */
1242 
1243 	if (append_addr_host((*tbl)->pt_buf, &node_host, 0, 0)) {
1244 		warn("failed to add host");
1245 		return (1);
1246 	}
1247 	if (pf->opts & PF_OPT_VERBOSE) {
1248 		struct node_tinit *ti;
1249 
1250 		if ((ti = calloc(1, sizeof(*ti))) == NULL)
1251 			err(1, "malloc");
1252 		if ((ti->host = malloc(sizeof(*ti->host))) == NULL)
1253 			err(1, "malloc");
1254 		memcpy(ti->host, &node_host, sizeof(*ti->host));
1255 		SIMPLEQ_INSERT_TAIL(&(*tbl)->pt_nodes, ti, entries);
1256 	}
1257 
1258 	(*tbl)->pt_rulecount++;
1259 	if ((*tbl)->pt_rulecount == TABLE_THRESHOLD)
1260 		DEBUG("table <%s> now faster than skip steps", (*tbl)->pt_name);
1261 
1262 	return (0);
1263 }
1264 
1265 
1266 /*
1267  * Do the dirty work of choosing an unused table name and creating it.
1268  * (be careful with the table name, it might already be used in another anchor)
1269  */
1270 int
1271 pf_opt_create_table(struct pfctl *pf, struct pf_opt_tbl *tbl)
1272 {
1273 	static int tablenum;
1274 	struct pfr_table *t;
1275 
1276 	if (table_buffer.pfrb_type == 0) {
1277 		/* Initialize the list of tables */
1278 		table_buffer.pfrb_type = PFRB_TABLES;
1279 		for (;;) {
1280 			pfr_buf_grow(&table_buffer, table_buffer.pfrb_size);
1281 			table_buffer.pfrb_size = table_buffer.pfrb_msize;
1282 			if (pfr_get_tables(NULL, table_buffer.pfrb_caddr,
1283 			    &table_buffer.pfrb_size, PFR_FLAG_ALLRSETS))
1284 				err(1, "pfr_get_tables");
1285 			if (table_buffer.pfrb_size <= table_buffer.pfrb_msize)
1286 				break;
1287 		}
1288 		table_identifier = arc4random();
1289 	}
1290 
1291 	/* XXX would be *really* nice to avoid duplicating identical tables */
1292 
1293 	/* Now we have to pick a table name that isn't used */
1294 again:
1295 	DEBUG("translating temporary table <%s> to <%s%x_%d>", tbl->pt_name,
1296 	    PF_OPT_TABLE_PREFIX, table_identifier, tablenum);
1297 	snprintf(tbl->pt_name, sizeof(tbl->pt_name), "%s%x_%d",
1298 	    PF_OPT_TABLE_PREFIX, table_identifier, tablenum);
1299 	PFRB_FOREACH(t, &table_buffer) {
1300 		if (strcasecmp(t->pfrt_name, tbl->pt_name) == 0) {
1301 			/* Collision.  Try again */
1302 			DEBUG("wow, table <%s> in use.  trying again",
1303 			    tbl->pt_name);
1304 			table_identifier = arc4random();
1305 			goto again;
1306 		}
1307 	}
1308 	tablenum++;
1309 
1310 	if (pfctl_define_table(tbl->pt_name, PFR_TFLAG_CONST | tbl->pt_flags, 1,
1311 	    pf->astack[0]->name, tbl->pt_buf, pf->astack[0]->ruleset.tticket)) {
1312 		warn("failed to create table %s in %s",
1313 		    tbl->pt_name, pf->astack[0]->name);
1314 		return (1);
1315 	}
1316 	return (0);
1317 }
1318 
1319 /*
1320  * Partition the flat ruleset into a list of distinct superblocks
1321  */
1322 int
1323 construct_superblocks(struct pfctl *pf, struct pf_opt_queue *opt_queue,
1324     struct superblocks *superblocks)
1325 {
1326 	struct superblock *block = NULL;
1327 	struct pf_opt_rule *por;
1328 	int i;
1329 
1330 	while (!TAILQ_EMPTY(opt_queue)) {
1331 		por = TAILQ_FIRST(opt_queue);
1332 		TAILQ_REMOVE(opt_queue, por, por_entry);
1333 		if (block == NULL || !superblock_inclusive(block, por)) {
1334 			if ((block = calloc(1, sizeof(*block))) == NULL) {
1335 				warn("calloc");
1336 				return (1);
1337 			}
1338 			TAILQ_INIT(&block->sb_rules);
1339 			for (i = 0; i < PF_SKIP_COUNT; i++)
1340 				TAILQ_INIT(&block->sb_skipsteps[i]);
1341 			TAILQ_INSERT_TAIL(superblocks, block, sb_entry);
1342 		}
1343 		TAILQ_INSERT_TAIL(&block->sb_rules, por, por_entry);
1344 	}
1345 
1346 	return (0);
1347 }
1348 
1349 
1350 /*
1351  * Compare two rule addresses
1352  */
1353 int
1354 addrs_equal(struct pf_rule_addr *a, struct pf_rule_addr *b)
1355 {
1356 	if (a->neg != b->neg)
1357 		return (0);
1358 	return (memcmp(&a->addr, &b->addr, sizeof(a->addr)) == 0);
1359 }
1360 
1361 
1362 /*
1363  * The addresses are not equal, but can we combine them into one table?
1364  */
1365 int
1366 addrs_combineable(struct pf_rule_addr *a, struct pf_rule_addr *b)
1367 {
1368 	if (a->addr.type != PF_ADDR_ADDRMASK ||
1369 	    b->addr.type != PF_ADDR_ADDRMASK)
1370 		return (0);
1371 	if (a->neg != b->neg || a->port_op != b->port_op ||
1372 	    a->port[0] != b->port[0] || a->port[1] != b->port[1])
1373 		return (0);
1374 	return (1);
1375 }
1376 
1377 
1378 /*
1379  * Are we allowed to combine these two rules
1380  */
1381 int
1382 rules_combineable(struct pf_rule *p1, struct pf_rule *p2)
1383 {
1384 	struct pf_rule a, b;
1385 
1386 	comparable_rule(&a, p1, COMBINED);
1387 	comparable_rule(&b, p2, COMBINED);
1388 	return (memcmp(&a, &b, sizeof(a)) == 0);
1389 }
1390 
1391 
1392 /*
1393  * Can a rule be included inside a superblock
1394  */
1395 int
1396 superblock_inclusive(struct superblock *block, struct pf_opt_rule *por)
1397 {
1398 	struct pf_rule a, b;
1399 	int i, j;
1400 
1401 	/* First check for hard breaks */
1402 	for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++) {
1403 		if (pf_rule_desc[i].prf_type == BARRIER) {
1404 			for (j = 0; j < pf_rule_desc[i].prf_size; j++)
1405 				if (((char *)&por->por_rule)[j +
1406 				    pf_rule_desc[i].prf_offset] != 0)
1407 					return (0);
1408 		}
1409 	}
1410 
1411 	/* per-rule src-track is also a hard break */
1412 	if (por->por_rule.rule_flag & PFRULE_RULESRCTRACK)
1413 		return (0);
1414 
1415 	/*
1416 	 * Have to handle interface groups separately.  Consider the following
1417 	 * rules:
1418 	 *	block on EXTIFS to any port 22
1419 	 *	pass  on em0 to any port 22
1420 	 * (where EXTIFS is an arbitrary interface group)
1421 	 * The optimizer may decide to re-order the pass rule in front of the
1422 	 * block rule.  But what if EXTIFS includes em0???  Such a reordering
1423 	 * would change the meaning of the ruleset.
1424 	 * We can't just lookup the EXTIFS group and check if em0 is a member
1425 	 * because the user is allowed to add interfaces to a group during
1426 	 * runtime.
1427 	 * Ergo interface groups become a defacto superblock break :-(
1428 	 */
1429 	if (interface_group(por->por_rule.ifname) ||
1430 	    interface_group(TAILQ_FIRST(&block->sb_rules)->por_rule.ifname)) {
1431 		if (strcasecmp(por->por_rule.ifname,
1432 		    TAILQ_FIRST(&block->sb_rules)->por_rule.ifname) != 0)
1433 			return (0);
1434 	}
1435 
1436 	comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule, NOMERGE);
1437 	comparable_rule(&b, &por->por_rule, NOMERGE);
1438 	if (memcmp(&a, &b, sizeof(a)) == 0)
1439 		return (1);
1440 
1441 #ifdef OPT_DEBUG
1442 	for (i = 0; i < sizeof(por->por_rule); i++) {
1443 		int closest = -1;
1444 		if (((u_int8_t *)&a)[i] != ((u_int8_t *)&b)[i]) {
1445 			for (j = 0; j < sizeof(pf_rule_desc) /
1446 			    sizeof(*pf_rule_desc); j++) {
1447 				if (i >= pf_rule_desc[j].prf_offset &&
1448 				    i < pf_rule_desc[j].prf_offset +
1449 				    pf_rule_desc[j].prf_size) {
1450 					DEBUG("superblock break @ %d due to %s",
1451 					    por->por_rule.nr,
1452 					    pf_rule_desc[j].prf_name);
1453 					return (0);
1454 				}
1455 				if (i > pf_rule_desc[j].prf_offset) {
1456 					if (closest == -1 ||
1457 					    i-pf_rule_desc[j].prf_offset <
1458 					    i-pf_rule_desc[closest].prf_offset)
1459 						closest = j;
1460 				}
1461 			}
1462 
1463 			if (closest >= 0)
1464 				DEBUG("superblock break @ %d on %s+%lxh",
1465 				    por->por_rule.nr,
1466 				    pf_rule_desc[closest].prf_name,
1467 				    i - pf_rule_desc[closest].prf_offset -
1468 				    pf_rule_desc[closest].prf_size);
1469 			else
1470 				DEBUG("superblock break @ %d on field @ %d",
1471 				    por->por_rule.nr, i);
1472 			return (0);
1473 		}
1474 	}
1475 #endif /* OPT_DEBUG */
1476 
1477 	return (0);
1478 }
1479 
1480 
1481 /*
1482  * Figure out if an interface name is an actual interface or actually a
1483  * group of interfaces.
1484  */
1485 int
1486 interface_group(const char *ifname)
1487 {
1488 	if (ifname == NULL || !ifname[0])
1489 		return (0);
1490 
1491 	/* Real interfaces must end in a number, interface groups do not */
1492 	if (isdigit((unsigned char)ifname[strlen(ifname) - 1]))
1493 		return (0);
1494 	else
1495 		return (1);
1496 }
1497 
1498 
1499 /*
1500  * Make a rule that can directly compared by memcmp()
1501  */
1502 void
1503 comparable_rule(struct pf_rule *dst, const struct pf_rule *src, int type)
1504 {
1505 	int i;
1506 	/*
1507 	 * To simplify the comparison, we just zero out the fields that are
1508 	 * allowed to be different and then do a simple memcmp()
1509 	 */
1510 	memcpy(dst, src, sizeof(*dst));
1511 	for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++)
1512 		if (pf_rule_desc[i].prf_type >= type) {
1513 #ifdef OPT_DEBUG
1514 			assert(pf_rule_desc[i].prf_type != NEVER ||
1515 			    *(((char *)dst) + pf_rule_desc[i].prf_offset) == 0);
1516 #endif /* OPT_DEBUG */
1517 			memset(((char *)dst) + pf_rule_desc[i].prf_offset, 0,
1518 			    pf_rule_desc[i].prf_size);
1519 		}
1520 }
1521 
1522 
1523 /*
1524  * Remove superset information from two rules so we can directly compare them
1525  * with memcmp()
1526  */
1527 void
1528 exclude_supersets(struct pf_rule *super, struct pf_rule *sub)
1529 {
1530 	if (super->ifname[0] == '\0')
1531 		memset(sub->ifname, 0, sizeof(sub->ifname));
1532 	if (super->direction == PF_INOUT)
1533 		sub->direction = PF_INOUT;
1534 	if ((super->proto == 0 || super->proto == sub->proto) &&
1535 	    super->flags == 0 && super->flagset == 0 && (sub->flags ||
1536 	    sub->flagset)) {
1537 		sub->flags = super->flags;
1538 		sub->flagset = super->flagset;
1539 	}
1540 	if (super->proto == 0)
1541 		sub->proto = 0;
1542 
1543 	if (super->src.port_op == 0) {
1544 		sub->src.port_op = 0;
1545 		sub->src.port[0] = 0;
1546 		sub->src.port[1] = 0;
1547 	}
1548 	if (super->dst.port_op == 0) {
1549 		sub->dst.port_op = 0;
1550 		sub->dst.port[0] = 0;
1551 		sub->dst.port[1] = 0;
1552 	}
1553 
1554 	if (super->src.addr.type == PF_ADDR_ADDRMASK && !super->src.neg &&
1555 	    !sub->src.neg && super->src.addr.v.a.mask.addr32[0] == 0 &&
1556 	    super->src.addr.v.a.mask.addr32[1] == 0 &&
1557 	    super->src.addr.v.a.mask.addr32[2] == 0 &&
1558 	    super->src.addr.v.a.mask.addr32[3] == 0)
1559 		memset(&sub->src.addr, 0, sizeof(sub->src.addr));
1560 	else if (super->src.addr.type == PF_ADDR_ADDRMASK &&
1561 	    sub->src.addr.type == PF_ADDR_ADDRMASK &&
1562 	    super->src.neg == sub->src.neg &&
1563 	    super->af == sub->af &&
1564 	    unmask(&super->src.addr.v.a.mask, super->af) <
1565 	    unmask(&sub->src.addr.v.a.mask, sub->af) &&
1566 	    super->src.addr.v.a.addr.addr32[0] ==
1567 	    (sub->src.addr.v.a.addr.addr32[0] &
1568 	    super->src.addr.v.a.mask.addr32[0]) &&
1569 	    super->src.addr.v.a.addr.addr32[1] ==
1570 	    (sub->src.addr.v.a.addr.addr32[1] &
1571 	    super->src.addr.v.a.mask.addr32[1]) &&
1572 	    super->src.addr.v.a.addr.addr32[2] ==
1573 	    (sub->src.addr.v.a.addr.addr32[2] &
1574 	    super->src.addr.v.a.mask.addr32[2]) &&
1575 	    super->src.addr.v.a.addr.addr32[3] ==
1576 	    (sub->src.addr.v.a.addr.addr32[3] &
1577 	    super->src.addr.v.a.mask.addr32[3])) {
1578 		/* sub->src.addr is a subset of super->src.addr/mask */
1579 		memcpy(&sub->src.addr, &super->src.addr, sizeof(sub->src.addr));
1580 	}
1581 
1582 	if (super->dst.addr.type == PF_ADDR_ADDRMASK && !super->dst.neg &&
1583 	    !sub->dst.neg && super->dst.addr.v.a.mask.addr32[0] == 0 &&
1584 	    super->dst.addr.v.a.mask.addr32[1] == 0 &&
1585 	    super->dst.addr.v.a.mask.addr32[2] == 0 &&
1586 	    super->dst.addr.v.a.mask.addr32[3] == 0)
1587 		memset(&sub->dst.addr, 0, sizeof(sub->dst.addr));
1588 	else if (super->dst.addr.type == PF_ADDR_ADDRMASK &&
1589 	    sub->dst.addr.type == PF_ADDR_ADDRMASK &&
1590 	    super->dst.neg == sub->dst.neg &&
1591 	    super->af == sub->af &&
1592 	    unmask(&super->dst.addr.v.a.mask, super->af) <
1593 	    unmask(&sub->dst.addr.v.a.mask, sub->af) &&
1594 	    super->dst.addr.v.a.addr.addr32[0] ==
1595 	    (sub->dst.addr.v.a.addr.addr32[0] &
1596 	    super->dst.addr.v.a.mask.addr32[0]) &&
1597 	    super->dst.addr.v.a.addr.addr32[1] ==
1598 	    (sub->dst.addr.v.a.addr.addr32[1] &
1599 	    super->dst.addr.v.a.mask.addr32[1]) &&
1600 	    super->dst.addr.v.a.addr.addr32[2] ==
1601 	    (sub->dst.addr.v.a.addr.addr32[2] &
1602 	    super->dst.addr.v.a.mask.addr32[2]) &&
1603 	    super->dst.addr.v.a.addr.addr32[3] ==
1604 	    (sub->dst.addr.v.a.addr.addr32[3] &
1605 	    super->dst.addr.v.a.mask.addr32[3])) {
1606 		/* sub->dst.addr is a subset of super->dst.addr/mask */
1607 		memcpy(&sub->dst.addr, &super->dst.addr, sizeof(sub->dst.addr));
1608 	}
1609 
1610 	if (super->af == 0)
1611 		sub->af = 0;
1612 }
1613 
1614 
1615 void
1616 superblock_free(struct pfctl *pf, struct superblock *block)
1617 {
1618 	struct pf_opt_rule *por;
1619 	while ((por = TAILQ_FIRST(&block->sb_rules))) {
1620 		TAILQ_REMOVE(&block->sb_rules, por, por_entry);
1621 		if (por->por_src_tbl) {
1622 			if (por->por_src_tbl->pt_buf) {
1623 				pfr_buf_clear(por->por_src_tbl->pt_buf);
1624 				free(por->por_src_tbl->pt_buf);
1625 			}
1626 			free(por->por_src_tbl);
1627 		}
1628 		if (por->por_dst_tbl) {
1629 			if (por->por_dst_tbl->pt_buf) {
1630 				pfr_buf_clear(por->por_dst_tbl->pt_buf);
1631 				free(por->por_dst_tbl->pt_buf);
1632 			}
1633 			free(por->por_dst_tbl);
1634 		}
1635 		free(por);
1636 	}
1637 	if (block->sb_profiled_block)
1638 		superblock_free(pf, block->sb_profiled_block);
1639 	free(block);
1640 }
1641 
1642