xref: /openbsd-src/sbin/pfctl/pfctl_optimize.c (revision 9f11ffb7133c203312a01e4b986886bc88c7d74b)
1 /*	$OpenBSD: pfctl_optimize.c,v 1.40 2019/01/03 22:49:00 kn 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 <netinet/in.h>
24 #include <arpa/inet.h>
25 #include <net/if.h>
26 #include <net/pfvar.h>
27 
28 #include <assert.h>
29 #include <ctype.h>
30 #include <err.h>
31 #include <errno.h>
32 #include <stddef.h>
33 #include <stdio.h>
34 #include <stdlib.h>
35 #include <string.h>
36 
37 #include "pfctl_parser.h"
38 #include "pfctl.h"
39 
40 /* The size at which a table becomes faster than individual rules */
41 #define TABLE_THRESHOLD		6
42 
43 
44 /* #define OPT_DEBUG	1 */
45 #ifdef OPT_DEBUG
46 # define DEBUG(str, v...) \
47 	printf("%s: " str "\n", __FUNCTION__ , ## v)
48 #else
49 # define DEBUG(str, v...) ((void)0)
50 #endif
51 
52 
53 /*
54  * A container that lets us sort a superblock to optimize the skip step jumps
55  */
56 struct pf_skip_step {
57 	int				ps_count;	/* number of items */
58 	TAILQ_HEAD( , pf_opt_rule)	ps_rules;
59 	TAILQ_ENTRY(pf_skip_step)	ps_entry;
60 };
61 
62 
63 /*
64  * A superblock is a block of adjacent rules of similar action.  If there
65  * are five PASS rules in a row, they all become members of a superblock.
66  * Once we have a superblock, we are free to re-order any rules within it
67  * in order to improve performance; if a packet is passed, it doesn't matter
68  * who passed it.
69  */
70 struct superblock {
71 	TAILQ_HEAD( , pf_opt_rule)		 sb_rules;
72 	TAILQ_ENTRY(superblock)			 sb_entry;
73 	struct superblock			*sb_profiled_block;
74 	TAILQ_HEAD(skiplist, pf_skip_step)	 sb_skipsteps[PF_SKIP_COUNT];
75 };
76 TAILQ_HEAD(superblocks, superblock);
77 
78 
79 /*
80  * Description of the PF rule structure.
81  */
82 enum {
83     BARRIER,	/* the presence of the field puts the rule in it's own block */
84     BREAK,	/* the field may not differ between rules in a superblock */
85     NOMERGE,	/* the field may not differ between rules when combined */
86     COMBINED,	/* the field may itself be combined with other rules */
87     DC,		/* we just don't care about the field */
88     NEVER};	/* we should never see this field set?!? */
89 struct pf_rule_field {
90 	const char	*prf_name;
91 	int		 prf_type;
92 	size_t		 prf_offset;
93 	size_t		 prf_size;
94 } pf_rule_desc[] = {
95 #define PF_RULE_FIELD(field, ty)	\
96     {#field,				\
97     ty,					\
98     offsetof(struct pf_rule, field),	\
99     sizeof(((struct pf_rule *)0)->field)}
100 
101 
102     /*
103      * The presence of these fields in a rule put the rule in it's own
104      * superblock.  Thus it will not be optimized.  It also prevents the
105      * rule from being re-ordered at all.
106      */
107     PF_RULE_FIELD(label,		BARRIER),
108     PF_RULE_FIELD(prob,			BARRIER),
109     PF_RULE_FIELD(max_states,		BARRIER),
110     PF_RULE_FIELD(max_src_nodes,	BARRIER),
111     PF_RULE_FIELD(max_src_states,	BARRIER),
112     PF_RULE_FIELD(max_src_conn,		BARRIER),
113     PF_RULE_FIELD(max_src_conn_rate,	BARRIER),
114     PF_RULE_FIELD(anchor,		BARRIER),	/* for now */
115 
116     /*
117      * These fields must be the same between all rules in the same superblock.
118      * These rules are allowed to be re-ordered but only among like rules.
119      * For instance we can re-order all 'tag "foo"' rules because they have the
120      * same tag.  But we can not re-order between a 'tag "foo"' and a
121      * 'tag "bar"' since that would change the meaning of the ruleset.
122      */
123     PF_RULE_FIELD(tagname,		BREAK),
124     PF_RULE_FIELD(keep_state,		BREAK),
125     PF_RULE_FIELD(qname,		BREAK),
126     PF_RULE_FIELD(pqname,		BREAK),
127     PF_RULE_FIELD(rt,			BREAK),
128     PF_RULE_FIELD(allow_opts,		BREAK),
129     PF_RULE_FIELD(rule_flag,		BREAK),
130     PF_RULE_FIELD(action,		BREAK),
131     PF_RULE_FIELD(log,			BREAK),
132     PF_RULE_FIELD(quick,		BREAK),
133     PF_RULE_FIELD(return_ttl,		BREAK),
134     PF_RULE_FIELD(overload_tblname,	BREAK),
135     PF_RULE_FIELD(flush,		BREAK),
136     PF_RULE_FIELD(rdr,			BREAK),
137     PF_RULE_FIELD(nat,			BREAK),
138     PF_RULE_FIELD(logif,		BREAK),
139     PF_RULE_FIELD(route,		BREAK),
140     PF_RULE_FIELD(rtableid,		BREAK),
141 
142     /*
143      * Any fields not listed in this structure act as BREAK fields
144      */
145 
146 
147     /*
148      * These fields must not differ when we merge two rules together but
149      * their difference isn't enough to put the rules in different superblocks.
150      * There are no problems re-ordering any rules with these fields.
151      */
152     PF_RULE_FIELD(af,			NOMERGE),
153     PF_RULE_FIELD(ifnot,		NOMERGE),
154     PF_RULE_FIELD(ifname,		NOMERGE),	/* hack for IF groups */
155     PF_RULE_FIELD(match_tag_not,	NOMERGE),
156     PF_RULE_FIELD(match_tagname,	NOMERGE),
157     PF_RULE_FIELD(os_fingerprint,	NOMERGE),
158     PF_RULE_FIELD(timeout,		NOMERGE),
159     PF_RULE_FIELD(return_icmp,		NOMERGE),
160     PF_RULE_FIELD(return_icmp6,		NOMERGE),
161     PF_RULE_FIELD(uid,			NOMERGE),
162     PF_RULE_FIELD(gid,			NOMERGE),
163     PF_RULE_FIELD(direction,		NOMERGE),
164     PF_RULE_FIELD(proto,		NOMERGE),
165     PF_RULE_FIELD(type,			NOMERGE),
166     PF_RULE_FIELD(code,			NOMERGE),
167     PF_RULE_FIELD(flags,		NOMERGE),
168     PF_RULE_FIELD(flagset,		NOMERGE),
169     PF_RULE_FIELD(tos,			NOMERGE),
170     PF_RULE_FIELD(src.port,		NOMERGE),
171     PF_RULE_FIELD(dst.port,		NOMERGE),
172     PF_RULE_FIELD(src.port_op,		NOMERGE),
173     PF_RULE_FIELD(dst.port_op,		NOMERGE),
174     PF_RULE_FIELD(src.neg,		NOMERGE),
175     PF_RULE_FIELD(dst.neg,		NOMERGE),
176     PF_RULE_FIELD(onrdomain,		NOMERGE),
177     PF_RULE_FIELD(naf,			NOMERGE),
178 
179     /* These fields can be merged */
180     PF_RULE_FIELD(src.addr,		COMBINED),
181     PF_RULE_FIELD(dst.addr,		COMBINED),
182 
183     /* We just don't care about these fields.  They're set by the kernel */
184     PF_RULE_FIELD(skip,			DC),
185     PF_RULE_FIELD(evaluations,		DC),
186     PF_RULE_FIELD(packets,		DC),
187     PF_RULE_FIELD(bytes,		DC),
188     PF_RULE_FIELD(kif,			DC),
189     PF_RULE_FIELD(states_cur,		DC),
190     PF_RULE_FIELD(states_tot,		DC),
191     PF_RULE_FIELD(src_nodes,		DC),
192     PF_RULE_FIELD(nr,			DC),
193     PF_RULE_FIELD(entries,		DC),
194     PF_RULE_FIELD(qid,			DC),
195     PF_RULE_FIELD(pqid,			DC),
196     PF_RULE_FIELD(anchor_relative,	DC),
197     PF_RULE_FIELD(anchor_wildcard,	DC),
198     PF_RULE_FIELD(tag,			DC),
199     PF_RULE_FIELD(match_tag,		DC),
200     PF_RULE_FIELD(overload_tbl,		DC),
201 
202     /* These fields should never be set in a PASS/BLOCK rule XXX fix*/
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	addrs_combineable(struct pf_rule_addr *, struct pf_rule_addr *);
211 int	addrs_equal(struct pf_rule_addr *, struct pf_rule_addr *);
212 int	block_feedback(struct pfctl *, struct superblock *);
213 int	combine_rules(struct pfctl *, struct superblock *);
214 void	comparable_rule(struct pf_rule *, const struct pf_rule *, int);
215 int	construct_superblocks(struct pfctl *, struct pf_opt_queue *,
216 	    struct superblocks *);
217 void	exclude_supersets(struct pf_rule *, struct pf_rule *);
218 int	interface_group(const char *);
219 int	load_feedback_profile(struct pfctl *, struct superblocks *);
220 int	optimize_superblock(struct pfctl *, struct superblock *);
221 void	remove_from_skipsteps(struct skiplist *, struct superblock *,
222 	    struct pf_opt_rule *, struct pf_skip_step *);
223 int	remove_identical_rules(struct pfctl *, struct superblock *);
224 int	reorder_rules(struct pfctl *, struct superblock *, int);
225 int	rules_combineable(struct pf_rule *, struct pf_rule *);
226 void	skip_append(struct superblock *, int, struct pf_skip_step *,
227 	    struct pf_opt_rule *);
228 int	skip_compare(int, struct pf_skip_step *, struct pf_opt_rule *);
229 void	skip_init(void);
230 int	skip_cmp_af(struct pf_rule *, struct pf_rule *);
231 int	skip_cmp_dir(struct pf_rule *, struct pf_rule *);
232 int	skip_cmp_rdom(struct pf_rule *, struct pf_rule *);
233 int	skip_cmp_dst_addr(struct pf_rule *, struct pf_rule *);
234 int	skip_cmp_dst_port(struct pf_rule *, struct pf_rule *);
235 int	skip_cmp_ifp(struct pf_rule *, struct pf_rule *);
236 int	skip_cmp_proto(struct pf_rule *, struct pf_rule *);
237 int	skip_cmp_src_addr(struct pf_rule *, struct pf_rule *);
238 int	skip_cmp_src_port(struct pf_rule *, struct pf_rule *);
239 int	superblock_inclusive(struct superblock *, struct pf_opt_rule *);
240 void	superblock_free(struct pfctl *, struct superblock *);
241 struct	pf_opt_tbl *pf_opt_table_ref(struct pf_opt_tbl *);
242 void	pf_opt_table_unref(struct pf_opt_tbl *);
243 
244 
245 int (*skip_comparitors[PF_SKIP_COUNT])(struct pf_rule *, struct pf_rule *);
246 const char *skip_comparitors_names[PF_SKIP_COUNT];
247 #define PF_SKIP_COMPARITORS {				\
248     { "ifp", PF_SKIP_IFP, skip_cmp_ifp },		\
249     { "dir", PF_SKIP_DIR, skip_cmp_dir },		\
250     { "rdomain", PF_SKIP_RDOM, skip_cmp_rdom },		\
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     { "daddr", PF_SKIP_DST_ADDR, skip_cmp_dst_addr },	\
255     { "sport", PF_SKIP_SRC_PORT, skip_cmp_src_port },	\
256     { "dport", PF_SKIP_DST_PORT, skip_cmp_dst_port }	\
257 }
258 
259 struct pfr_buffer table_buffer;
260 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.active.ptr;
279 	rs->rules.active.ptr = rs->rules.inactive.ptr;
280 	rs->rules.inactive.ptr = old_rules;
281 
282 	/*
283 	 * XXX expanding the pf_opt_rule format throughout pfctl might allow
284 	 * us to avoid all this copying.
285 	 */
286 	while ((r = TAILQ_FIRST(rs->rules.inactive.ptr)) != NULL) {
287 		TAILQ_REMOVE(rs->rules.inactive.ptr, r, entries);
288 		if ((por = calloc(1, sizeof(*por))) == NULL)
289 			err(1, "calloc");
290 		memcpy(&por->por_rule, r, sizeof(*r));
291 
292 		TAILQ_INSERT_TAIL(&opt_queue, por, por_entry);
293 	}
294 
295 	TAILQ_INIT(&superblocks);
296 	if (construct_superblocks(pf, &opt_queue, &superblocks))
297 		goto error;
298 
299 	if (pf->optimize & PF_OPTIMIZE_PROFILE) {
300 		if (load_feedback_profile(pf, &superblocks))
301 			goto error;
302 	}
303 
304 	TAILQ_FOREACH(block, &superblocks, sb_entry) {
305 		if (optimize_superblock(pf, block))
306 			goto error;
307 	}
308 
309 	rs->anchor->refcnt = 0;
310 	while ((block = TAILQ_FIRST(&superblocks))) {
311 		TAILQ_REMOVE(&superblocks, block, sb_entry);
312 
313 		while ((por = TAILQ_FIRST(&block->sb_rules))) {
314 			TAILQ_REMOVE(&block->sb_rules, por, por_entry);
315 			por->por_rule.nr = rs->anchor->refcnt++;
316 			if ((r = calloc(1, sizeof(*r))) == NULL)
317 				err(1, "calloc");
318 			memcpy(r, &por->por_rule, sizeof(*r));
319 			TAILQ_INSERT_TAIL(rs->rules.active.ptr, r, entries);
320 			pf_opt_table_unref(por->por_src_tbl);
321 			pf_opt_table_unref(por->por_dst_tbl);
322 			free(por);
323 		}
324 		superblock_free(pf, block);
325 	}
326 
327 	return (0);
328 
329 error:
330 	while ((por = TAILQ_FIRST(&opt_queue))) {
331 		TAILQ_REMOVE(&opt_queue, por, por_entry);
332 		pf_opt_table_unref(por->por_src_tbl);
333 		pf_opt_table_unref(por->por_dst_tbl);
334 		free(por);
335 	}
336 	while ((block = TAILQ_FIRST(&superblocks))) {
337 		TAILQ_REMOVE(&superblocks, block, sb_entry);
338 		superblock_free(pf, block);
339 	}
340 	return (1);
341 }
342 
343 
344 /*
345  * Go ahead and optimize a superblock
346  */
347 int
348 optimize_superblock(struct pfctl *pf, struct superblock *block)
349 {
350 #ifdef OPT_DEBUG
351 	struct pf_opt_rule *por;
352 #endif /* OPT_DEBUG */
353 
354 	/* We have a few optimization passes:
355 	 *   1) remove duplicate rules or rules that are a subset of other
356 	 *      rules
357 	 *   2) combine otherwise identical rules with different IP addresses
358 	 *      into a single rule and put the addresses in a table.
359 	 *   3) re-order the rules to improve kernel skip steps
360 	 *   4) re-order the 'quick' rules based on feedback from the
361 	 *      active ruleset statistics
362 	 *
363 	 * XXX combine_rules() doesn't combine v4 and v6 rules.  would just
364 	 *     have to keep af in the table container, make af 'COMBINE' and
365 	 *     twiddle the af on the merged rule
366 	 * XXX maybe add a weighting to the metric on skipsteps when doing
367 	 *     reordering.  sometimes two sequential tables will be better
368 	 *     that four consecutive interfaces.
369 	 * XXX need to adjust the skipstep count of everything after PROTO,
370 	 *     since they aren't actually checked on a proto mismatch in
371 	 *     pf_test_{tcp, udp, icmp}()
372 	 * XXX should i treat proto=0, af=0 or dir=0 special in skepstep
373 	 *     calculation since they are a DC?
374 	 * XXX keep last skiplist of last superblock to influence this
375 	 *     superblock.  '5 inet6 log' should make '3 inet6' come before '4
376 	 *     inet' in the next superblock.
377 	 * XXX would be useful to add tables for ports
378 	 * XXX we can also re-order some mutually exclusive superblocks to
379 	 *     try merging superblocks before any of these optimization passes.
380 	 *     for instance a single 'log in' rule in the middle of non-logging
381 	 *     out rules.
382 	 */
383 
384 	/* shortcut.  there will be a lot of 1-rule superblocks */
385 	if (!TAILQ_NEXT(TAILQ_FIRST(&block->sb_rules), por_entry))
386 		return (0);
387 
388 #ifdef OPT_DEBUG
389 	printf("--- Superblock ---\n");
390 	TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
391 		printf("  ");
392 		print_rule(&por->por_rule, por->por_rule.anchor ?
393 		    por->por_rule.anchor->name : "", PF_OPT_DEBUG);
394 	}
395 #endif /* OPT_DEBUG */
396 
397 
398 	if (remove_identical_rules(pf, block))
399 		return (1);
400 	if (combine_rules(pf, block))
401 		return (1);
402 	if ((pf->optimize & PF_OPTIMIZE_PROFILE) &&
403 	    TAILQ_FIRST(&block->sb_rules)->por_rule.quick &&
404 	    block->sb_profiled_block) {
405 		if (block_feedback(pf, block))
406 			return (1);
407 	} else if (reorder_rules(pf, block, 0)) {
408 		return (1);
409 	}
410 
411 	/*
412 	 * Don't add any optimization passes below reorder_rules().  It will
413 	 * have divided superblocks into smaller blocks for further refinement
414 	 * and doesn't put them back together again.  What once was a true
415 	 * superblock might have been split into multiple superblocks.
416 	 */
417 
418 #ifdef OPT_DEBUG
419 	printf("--- END Superblock ---\n");
420 #endif /* OPT_DEBUG */
421 	return (0);
422 }
423 
424 
425 /*
426  * Optimization pass #1: remove identical rules
427  */
428 int
429 remove_identical_rules(struct pfctl *pf, struct superblock *block)
430 {
431 	struct pf_opt_rule *por1, *por2, *por_next, *por2_next;
432 	struct pf_rule a, a2, b, b2;
433 
434 	for (por1 = TAILQ_FIRST(&block->sb_rules); por1; por1 = por_next) {
435 		por_next = TAILQ_NEXT(por1, por_entry);
436 		for (por2 = por_next; por2; por2 = por2_next) {
437 			por2_next = TAILQ_NEXT(por2, por_entry);
438 			comparable_rule(&a, &por1->por_rule, DC);
439 			comparable_rule(&b, &por2->por_rule, DC);
440 			memcpy(&a2, &a, sizeof(a2));
441 			memcpy(&b2, &b, sizeof(b2));
442 
443 			exclude_supersets(&a, &b);
444 			exclude_supersets(&b2, &a2);
445 			if (memcmp(&a, &b, sizeof(a)) == 0) {
446 				DEBUG("removing identical rule  nr%d = *nr%d*",
447 				    por1->por_rule.nr, por2->por_rule.nr);
448 				TAILQ_REMOVE(&block->sb_rules, por2, por_entry);
449 				if (por_next == por2)
450 					por_next = TAILQ_NEXT(por1, por_entry);
451 				free(por2);
452 			} else if (memcmp(&a2, &b2, sizeof(a2)) == 0) {
453 				DEBUG("removing identical rule  *nr%d* = nr%d",
454 				    por1->por_rule.nr, por2->por_rule.nr);
455 				TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
456 				free(por1);
457 				break;
458 			}
459 		}
460 	}
461 
462 	return (0);
463 }
464 
465 
466 /*
467  * Optimization pass #2: combine similar rules with different addresses
468  * into a single rule and a table
469  */
470 int
471 combine_rules(struct pfctl *pf, struct superblock *block)
472 {
473 	struct pf_opt_rule *p1, *p2, *por_next;
474 	int src_eq, dst_eq;
475 
476 	/* First we make a pass to combine the rules.  O(n log n) */
477 	TAILQ_FOREACH(p1, &block->sb_rules, por_entry) {
478 		for (p2 = TAILQ_NEXT(p1, por_entry); p2; p2 = por_next) {
479 			por_next = TAILQ_NEXT(p2, por_entry);
480 
481 			src_eq = addrs_equal(&p1->por_rule.src,
482 			    &p2->por_rule.src);
483 			dst_eq = addrs_equal(&p1->por_rule.dst,
484 			    &p2->por_rule.dst);
485 
486 			if (src_eq && !dst_eq && p1->por_src_tbl == NULL &&
487 			    p2->por_dst_tbl == NULL &&
488 			    p2->por_src_tbl == NULL &&
489 			    rules_combineable(&p1->por_rule, &p2->por_rule) &&
490 			    addrs_combineable(&p1->por_rule.dst,
491 			    &p2->por_rule.dst)) {
492 				DEBUG("can combine rules  nr%d = nr%d",
493 				    p1->por_rule.nr, p2->por_rule.nr);
494 				if (p1->por_dst_tbl == NULL &&
495 				    add_opt_table(pf, &p1->por_dst_tbl,
496 				    p1->por_rule.af, &p1->por_rule.dst, NULL))
497 					return (1);
498 				if (add_opt_table(pf, &p1->por_dst_tbl,
499 				    p1->por_rule.af, &p2->por_rule.dst, NULL))
500 					return (1);
501 				if (p1->por_dst_tbl->pt_rulecount >=
502 				    TABLE_THRESHOLD) {
503 					TAILQ_REMOVE(&block->sb_rules, p2,
504 					    por_entry);
505 					free(p2);
506 				} else
507 					p2->por_dst_tbl =
508 					    pf_opt_table_ref(p1->por_dst_tbl);
509 			} else if (!src_eq && dst_eq && p1->por_dst_tbl == NULL
510 			    && p2->por_src_tbl == NULL &&
511 			    p2->por_dst_tbl == NULL &&
512 			    rules_combineable(&p1->por_rule, &p2->por_rule) &&
513 			    addrs_combineable(&p1->por_rule.src,
514 			    &p2->por_rule.src)) {
515 				DEBUG("can combine rules  nr%d = nr%d",
516 				    p1->por_rule.nr, p2->por_rule.nr);
517 				if (p1->por_src_tbl == NULL &&
518 				    add_opt_table(pf, &p1->por_src_tbl,
519 				    p1->por_rule.af, &p1->por_rule.src, NULL))
520 					return (1);
521 				if (add_opt_table(pf, &p1->por_src_tbl,
522 				    p1->por_rule.af, &p2->por_rule.src, NULL))
523 					return (1);
524 				if (p1->por_src_tbl->pt_rulecount >=
525 				    TABLE_THRESHOLD) {
526 					TAILQ_REMOVE(&block->sb_rules, p2,
527 					    por_entry);
528 					free(p2);
529 				} else
530 					p2->por_src_tbl =
531 					    pf_opt_table_ref(p1->por_src_tbl);
532 			}
533 		}
534 	}
535 
536 
537 	/*
538 	 * Then we make a final pass to create a valid table name and
539 	 * insert the name into the rules.
540 	 * Convert translation/routing mapping pools to tables as well.
541 	 */
542 	for (p1 = TAILQ_FIRST(&block->sb_rules); p1; p1 = por_next) {
543 		por_next = TAILQ_NEXT(p1, por_entry);
544 		assert(p1->por_src_tbl == NULL || p1->por_dst_tbl == NULL);
545 
546 		if (p1->por_src_tbl && p1->por_src_tbl->pt_rulecount >=
547 		    TABLE_THRESHOLD) {
548 			if (p1->por_src_tbl->pt_generated) {
549 				/* This rule is included in a table */
550 				TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
551 				free(p1);
552 				continue;
553 			}
554 			p1->por_src_tbl->pt_generated = 1;
555 
556 			if ((pf->opts & PF_OPT_NOACTION) == 0 &&
557 			    pf_opt_create_table(pf, p1->por_src_tbl))
558 				return (1);
559 
560 			pf->tdirty = 1;
561 
562 			if (pf->opts & PF_OPT_VERBOSE)
563 				print_tabledef(p1->por_src_tbl->pt_name,
564 				    PFR_TFLAG_CONST, 1,
565 				    &p1->por_src_tbl->pt_nodes);
566 
567 			memset(&p1->por_rule.src.addr, 0,
568 			    sizeof(p1->por_rule.src.addr));
569 			p1->por_rule.src.addr.type = PF_ADDR_TABLE;
570 			strlcpy(p1->por_rule.src.addr.v.tblname,
571 			    p1->por_src_tbl->pt_name,
572 			    sizeof(p1->por_rule.src.addr.v.tblname));
573 
574 			pfr_buf_clear(p1->por_src_tbl->pt_buf);
575 			free(p1->por_src_tbl->pt_buf);
576 			p1->por_src_tbl->pt_buf = NULL;
577 		}
578 		if (p1->por_dst_tbl && p1->por_dst_tbl->pt_rulecount >=
579 		    TABLE_THRESHOLD) {
580 			if (p1->por_dst_tbl->pt_generated) {
581 				/* This rule is included in a table */
582 				TAILQ_REMOVE(&block->sb_rules, p1, por_entry);
583 				free(p1);
584 				continue;
585 			}
586 			p1->por_dst_tbl->pt_generated = 1;
587 
588 			if ((pf->opts & PF_OPT_NOACTION) == 0 &&
589 			    pf_opt_create_table(pf, p1->por_dst_tbl))
590 				return (1);
591 			pf->tdirty = 1;
592 
593 			if (pf->opts & PF_OPT_VERBOSE)
594 				print_tabledef(p1->por_dst_tbl->pt_name,
595 				    PFR_TFLAG_CONST, 1,
596 				    &p1->por_dst_tbl->pt_nodes);
597 
598 			memset(&p1->por_rule.dst.addr, 0,
599 			    sizeof(p1->por_rule.dst.addr));
600 			p1->por_rule.dst.addr.type = PF_ADDR_TABLE;
601 			strlcpy(p1->por_rule.dst.addr.v.tblname,
602 			    p1->por_dst_tbl->pt_name,
603 			    sizeof(p1->por_rule.dst.addr.v.tblname));
604 
605 			pfr_buf_clear(p1->por_dst_tbl->pt_buf);
606 			free(p1->por_dst_tbl->pt_buf);
607 			p1->por_dst_tbl->pt_buf = NULL;
608 		}
609 	}
610 
611 	return (0);
612 }
613 
614 
615 /*
616  * Optimization pass #3: re-order rules to improve skip steps
617  */
618 int
619 reorder_rules(struct pfctl *pf, struct superblock *block, int depth)
620 {
621 	struct superblock *newblock;
622 	struct pf_skip_step *skiplist;
623 	struct pf_opt_rule *por;
624 	int i, largest, largest_list, rule_count = 0;
625 	TAILQ_HEAD( , pf_opt_rule) head;
626 
627 	/*
628 	 * Calculate the best-case skip steps.  We put each rule in a list
629 	 * of other rules with common fields
630 	 */
631 	for (i = 0; i < PF_SKIP_COUNT; i++) {
632 		TAILQ_FOREACH(por, &block->sb_rules, por_entry) {
633 			TAILQ_FOREACH(skiplist, &block->sb_skipsteps[i],
634 			    ps_entry) {
635 				if (skip_compare(i, skiplist, por) == 0)
636 					break;
637 			}
638 			if (skiplist == NULL) {
639 				if ((skiplist = calloc(1, sizeof(*skiplist))) ==
640 				    NULL)
641 					err(1, "calloc");
642 				TAILQ_INIT(&skiplist->ps_rules);
643 				TAILQ_INSERT_TAIL(&block->sb_skipsteps[i],
644 				    skiplist, ps_entry);
645 			}
646 			skip_append(block, i, skiplist, por);
647 		}
648 	}
649 
650 	TAILQ_FOREACH(por, &block->sb_rules, por_entry)
651 		rule_count++;
652 
653 	/*
654 	 * Now we're going to ignore any fields that are identical between
655 	 * all of the rules in the superblock and those fields which differ
656 	 * between every rule in the superblock.
657 	 */
658 	largest = 0;
659 	for (i = 0; i < PF_SKIP_COUNT; i++) {
660 		skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
661 		if (skiplist->ps_count == rule_count) {
662 			DEBUG("(%d) original skipstep '%s' is all rules",
663 			    depth, skip_comparitors_names[i]);
664 			skiplist->ps_count = 0;
665 		} else if (skiplist->ps_count == 1) {
666 			skiplist->ps_count = 0;
667 		} else {
668 			DEBUG("(%d) original skipstep '%s' largest jump is %d",
669 			    depth, skip_comparitors_names[i],
670 			    skiplist->ps_count);
671 			if (skiplist->ps_count > largest)
672 				largest = skiplist->ps_count;
673 		}
674 	}
675 	if (largest == 0) {
676 		/* Ugh.  There is NO commonality in the superblock on which
677 		 * optimize the skipsteps optimization.
678 		 */
679 		goto done;
680 	}
681 
682 	/*
683 	 * Now we're going to empty the superblock rule list and re-create
684 	 * it based on a more optimal skipstep order.
685 	 */
686 	TAILQ_INIT(&head);
687 	while ((por = TAILQ_FIRST(&block->sb_rules))) {
688 		TAILQ_REMOVE(&block->sb_rules, por, por_entry);
689 		TAILQ_INSERT_TAIL(&head, por, por_entry);
690 	}
691 
692 
693 	while (!TAILQ_EMPTY(&head)) {
694 		largest = 1;
695 
696 		/*
697 		 * Find the most useful skip steps remaining
698 		 */
699 		for (i = 0; i < PF_SKIP_COUNT; i++) {
700 			skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]);
701 			if (skiplist->ps_count > largest) {
702 				largest = skiplist->ps_count;
703 				largest_list = i;
704 			}
705 		}
706 
707 		if (largest <= 1) {
708 			/*
709 			 * Nothing useful left.  Leave remaining rules in order.
710 			 */
711 			DEBUG("(%d) no more commonality for skip steps", depth);
712 			while ((por = TAILQ_FIRST(&head))) {
713 				TAILQ_REMOVE(&head, por, por_entry);
714 				TAILQ_INSERT_TAIL(&block->sb_rules, por,
715 				    por_entry);
716 			}
717 		} else {
718 			/*
719 			 * There is commonality.  Extract those common rules
720 			 * and place them in the ruleset adjacent to each
721 			 * other.
722 			 */
723 			skiplist = TAILQ_FIRST(&block->sb_skipsteps[
724 			    largest_list]);
725 			DEBUG("(%d) skipstep '%s' largest jump is %d @ #%d",
726 			    depth, skip_comparitors_names[largest_list],
727 			    largest, TAILQ_FIRST(&TAILQ_FIRST(&block->
728 			    sb_skipsteps [largest_list])->ps_rules)->
729 			    por_rule.nr);
730 			TAILQ_REMOVE(&block->sb_skipsteps[largest_list],
731 			    skiplist, ps_entry);
732 
733 
734 			/*
735 			 * There may be further commonality inside these
736 			 * rules.  So we'll split them off into they're own
737 			 * superblock and pass it back into the optimizer.
738 			 */
739 			if (skiplist->ps_count > 2) {
740 				if ((newblock = calloc(1, sizeof(*newblock)))
741 				    == NULL) {
742 					warn("calloc");
743 					return (1);
744 				}
745 				TAILQ_INIT(&newblock->sb_rules);
746 				for (i = 0; i < PF_SKIP_COUNT; i++)
747 					TAILQ_INIT(&newblock->sb_skipsteps[i]);
748 				TAILQ_INSERT_BEFORE(block, newblock, sb_entry);
749 				DEBUG("(%d) splitting off %d rules from superblock @ #%d",
750 				    depth, skiplist->ps_count,
751 				    TAILQ_FIRST(&skiplist->ps_rules)->
752 				    por_rule.nr);
753 			} else {
754 				newblock = block;
755 			}
756 
757 			while ((por = TAILQ_FIRST(&skiplist->ps_rules))) {
758 				TAILQ_REMOVE(&head, por, por_entry);
759 				TAILQ_REMOVE(&skiplist->ps_rules, por,
760 				    por_skip_entry[largest_list]);
761 				TAILQ_INSERT_TAIL(&newblock->sb_rules, por,
762 				    por_entry);
763 
764 				/* Remove this rule from all other skiplists */
765 				remove_from_skipsteps(&block->sb_skipsteps[
766 				    largest_list], block, por, skiplist);
767 			}
768 			free(skiplist);
769 			if (newblock != block)
770 				if (reorder_rules(pf, newblock, depth + 1))
771 					return (1);
772 		}
773 	}
774 
775 done:
776 	for (i = 0; i < PF_SKIP_COUNT; i++) {
777 		while ((skiplist = TAILQ_FIRST(&block->sb_skipsteps[i]))) {
778 			TAILQ_REMOVE(&block->sb_skipsteps[i], skiplist,
779 			    ps_entry);
780 			free(skiplist);
781 		}
782 	}
783 
784 	return (0);
785 }
786 
787 
788 /*
789  * Optimization pass #4: re-order 'quick' rules based on feedback from the
790  * currently running ruleset
791  */
792 int
793 block_feedback(struct pfctl *pf, struct superblock *block)
794 {
795 	TAILQ_HEAD( , pf_opt_rule) queue;
796 	struct pf_opt_rule *por1, *por2;
797 	u_int64_t total_count = 0;
798 	struct pf_rule a, b;
799 
800 
801 	/*
802 	 * Walk through all of the profiled superblock's rules and copy
803 	 * the counters onto our rules.
804 	 */
805 	TAILQ_FOREACH(por1, &block->sb_profiled_block->sb_rules, por_entry) {
806 		comparable_rule(&a, &por1->por_rule, DC);
807 		total_count += por1->por_rule.packets[0] +
808 		    por1->por_rule.packets[1];
809 		TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
810 			if (por2->por_profile_count)
811 				continue;
812 			comparable_rule(&b, &por2->por_rule, DC);
813 			if (memcmp(&a, &b, sizeof(a)) == 0) {
814 				por2->por_profile_count =
815 				    por1->por_rule.packets[0] +
816 				    por1->por_rule.packets[1];
817 				break;
818 			}
819 		}
820 	}
821 	superblock_free(pf, block->sb_profiled_block);
822 	block->sb_profiled_block = NULL;
823 
824 	/*
825 	 * Now we pull all of the rules off the superblock and re-insert them
826 	 * in sorted order.
827 	 */
828 
829 	TAILQ_INIT(&queue);
830 	while ((por1 = TAILQ_FIRST(&block->sb_rules)) != NULL) {
831 		TAILQ_REMOVE(&block->sb_rules, por1, por_entry);
832 		TAILQ_INSERT_TAIL(&queue, por1, por_entry);
833 	}
834 
835 	while ((por1 = TAILQ_FIRST(&queue)) != NULL) {
836 		TAILQ_REMOVE(&queue, por1, por_entry);
837 /* XXX I should sort all of the unused rules based on skip steps */
838 		TAILQ_FOREACH(por2, &block->sb_rules, por_entry) {
839 			if (por1->por_profile_count > por2->por_profile_count) {
840 				TAILQ_INSERT_BEFORE(por2, por1, por_entry);
841 				break;
842 			}
843 		}
844 		if (por2 == NULL)
845 			TAILQ_INSERT_TAIL(&block->sb_rules, por1, por_entry);
846 	}
847 
848 	return (0);
849 }
850 
851 
852 /*
853  * Load the current ruleset from the kernel and try to associate them with
854  * the ruleset we're optimizing.
855  */
856 int
857 load_feedback_profile(struct pfctl *pf, struct superblocks *superblocks)
858 {
859 	struct superblock *block, *blockcur;
860 	struct superblocks prof_superblocks;
861 	struct pf_opt_rule *por;
862 	struct pf_opt_queue queue;
863 	struct pfioc_rule pr;
864 	struct pf_rule a, b;
865 	int nr, mnr;
866 
867 	TAILQ_INIT(&queue);
868 	TAILQ_INIT(&prof_superblocks);
869 
870 	memset(&pr, 0, sizeof(pr));
871 	pr.rule.action = PF_PASS;
872 	if (ioctl(pf->dev, DIOCGETRULES, &pr)) {
873 		warn("DIOCGETRULES");
874 		return (1);
875 	}
876 	mnr = pr.nr;
877 
878 	DEBUG("Loading %d active rules for a feedback profile", mnr);
879 	for (nr = 0; nr < mnr; ++nr) {
880 		struct pf_ruleset *rs;
881 		if ((por = calloc(1, sizeof(*por))) == NULL) {
882 			warn("calloc");
883 			return (1);
884 		}
885 		pr.nr = nr;
886 		if (ioctl(pf->dev, DIOCGETRULE, &pr)) {
887 			warn("DIOCGETRULES");
888 			free(por);
889 			return (1);
890 		}
891 		memcpy(&por->por_rule, &pr.rule, sizeof(por->por_rule));
892 		rs = pf_find_or_create_ruleset(pr.anchor_call);
893 		por->por_rule.anchor = rs->anchor;
894 		TAILQ_INSERT_TAIL(&queue, por, por_entry);
895 
896 		/* XXX pfctl_get_pool(pf->dev, &pr.rule.rpool, nr, pr.ticket,
897 		 *         PF_PASS, pf->anchor) ???
898 		 * ... pfctl_clear_pool(&pr.rule.rpool)
899 		 */
900 	}
901 
902 	if (construct_superblocks(pf, &queue, &prof_superblocks))
903 		return (1);
904 
905 
906 	/*
907 	 * Now we try to associate the active ruleset's superblocks with
908 	 * the superblocks we're compiling.
909 	 */
910 	block = TAILQ_FIRST(superblocks);
911 	blockcur = TAILQ_FIRST(&prof_superblocks);
912 	while (block && blockcur) {
913 		comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule,
914 		    BREAK);
915 		comparable_rule(&b, &TAILQ_FIRST(&blockcur->sb_rules)->por_rule,
916 		    BREAK);
917 		if (memcmp(&a, &b, sizeof(a)) == 0) {
918 			/* The two superblocks lined up */
919 			block->sb_profiled_block = blockcur;
920 		} else {
921 			DEBUG("superblocks don't line up between #%d and #%d",
922 			    TAILQ_FIRST(&block->sb_rules)->por_rule.nr,
923 			    TAILQ_FIRST(&blockcur->sb_rules)->por_rule.nr);
924 			break;
925 		}
926 		block = TAILQ_NEXT(block, sb_entry);
927 		blockcur = TAILQ_NEXT(blockcur, sb_entry);
928 	}
929 
930 
931 
932 	/* Free any superblocks we couldn't link */
933 	while (blockcur) {
934 		block = TAILQ_NEXT(blockcur, sb_entry);
935 		superblock_free(pf, blockcur);
936 		blockcur = block;
937 	}
938 	return (0);
939 }
940 
941 
942 /*
943  * Compare a rule to a skiplist to see if the rule is a member
944  */
945 int
946 skip_compare(int skipnum, struct pf_skip_step *skiplist,
947     struct pf_opt_rule *por)
948 {
949 	struct pf_rule *a, *b;
950 	if (skipnum >= PF_SKIP_COUNT || skipnum < 0)
951 		errx(1, "skip_compare() out of bounds");
952 	a = &por->por_rule;
953 	b = &TAILQ_FIRST(&skiplist->ps_rules)->por_rule;
954 
955 	return ((skip_comparitors[skipnum])(a, b));
956 }
957 
958 
959 /*
960  * Add a rule to a skiplist
961  */
962 void
963 skip_append(struct superblock *superblock, int skipnum,
964     struct pf_skip_step *skiplist, struct pf_opt_rule *por)
965 {
966 	struct pf_skip_step *prev;
967 
968 	skiplist->ps_count++;
969 	TAILQ_INSERT_TAIL(&skiplist->ps_rules, por, por_skip_entry[skipnum]);
970 
971 	/* Keep the list of skiplists sorted by whichever is larger */
972 	while ((prev = TAILQ_PREV(skiplist, skiplist, ps_entry)) &&
973 	    prev->ps_count < skiplist->ps_count) {
974 		TAILQ_REMOVE(&superblock->sb_skipsteps[skipnum],
975 		    skiplist, ps_entry);
976 		TAILQ_INSERT_BEFORE(prev, skiplist, ps_entry);
977 	}
978 }
979 
980 
981 /*
982  * Remove a rule from the other skiplist calculations.
983  */
984 void
985 remove_from_skipsteps(struct skiplist *head, struct superblock *block,
986     struct pf_opt_rule *por, struct pf_skip_step *active_list)
987 {
988 	struct pf_skip_step *sk, *next;
989 	struct pf_opt_rule *p2;
990 	int i, found;
991 
992 	for (i = 0; i < PF_SKIP_COUNT; i++) {
993 		sk = TAILQ_FIRST(&block->sb_skipsteps[i]);
994 		if (sk == NULL || sk == active_list || sk->ps_count <= 1)
995 			continue;
996 		found = 0;
997 		do {
998 			TAILQ_FOREACH(p2, &sk->ps_rules, por_skip_entry[i])
999 				if (p2 == por) {
1000 					TAILQ_REMOVE(&sk->ps_rules, p2,
1001 					    por_skip_entry[i]);
1002 					found = 1;
1003 					sk->ps_count--;
1004 					break;
1005 				}
1006 		} while (!found && (sk = TAILQ_NEXT(sk, ps_entry)));
1007 		if (found && sk) {
1008 			/* Does this change the sorting order? */
1009 			while ((next = TAILQ_NEXT(sk, ps_entry)) &&
1010 			    next->ps_count > sk->ps_count) {
1011 				TAILQ_REMOVE(head, sk, ps_entry);
1012 				TAILQ_INSERT_AFTER(head, next, sk, ps_entry);
1013 			}
1014 #ifdef OPT_DEBUG
1015 			next = TAILQ_NEXT(sk, ps_entry);
1016 			assert(next == NULL || next->ps_count <= sk->ps_count);
1017 #endif /* OPT_DEBUG */
1018 		}
1019 	}
1020 }
1021 
1022 
1023 /* Compare two rules AF field for skiplist construction */
1024 int
1025 skip_cmp_af(struct pf_rule *a, struct pf_rule *b)
1026 {
1027 	if (a->af != b->af || a->af == 0)
1028 		return (1);
1029 	return (0);
1030 }
1031 
1032 /* Compare two rules DIRECTION field for skiplist construction */
1033 int
1034 skip_cmp_dir(struct pf_rule *a, struct pf_rule *b)
1035 {
1036 	if (a->direction == 0 || a->direction != b->direction)
1037 		return (1);
1038 	return (0);
1039 }
1040 
1041 /* Compare two rules ON RDOMAIN field for skiplist construction */
1042 int
1043 skip_cmp_rdom(struct pf_rule *a, struct pf_rule *b)
1044 {
1045 	if (a->onrdomain == -1 || a->onrdomain != b->onrdomain)
1046 		return (1);
1047 	return (a->ifnot != b->ifnot);
1048 }
1049 
1050 /* Compare two rules DST Address field for skiplist construction */
1051 int
1052 skip_cmp_dst_addr(struct pf_rule *a, struct pf_rule *b)
1053 {
1054 	if (a->dst.neg != b->dst.neg ||
1055 	    a->dst.addr.type != b->dst.addr.type)
1056 		return (1);
1057 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1058 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1059 	 *    a->proto == IPPROTO_ICMP
1060 	 *	return (1);
1061 	 */
1062 	switch (a->dst.addr.type) {
1063 	case PF_ADDR_ADDRMASK:
1064 		if (memcmp(&a->dst.addr.v.a.addr, &b->dst.addr.v.a.addr,
1065 		    sizeof(a->dst.addr.v.a.addr)) ||
1066 		    memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask,
1067 		    sizeof(a->dst.addr.v.a.mask)) ||
1068 		    (a->dst.addr.v.a.addr.addr32[0] == 0 &&
1069 		    a->dst.addr.v.a.addr.addr32[1] == 0 &&
1070 		    a->dst.addr.v.a.addr.addr32[2] == 0 &&
1071 		    a->dst.addr.v.a.addr.addr32[3] == 0))
1072 			return (1);
1073 		return (0);
1074 	case PF_ADDR_DYNIFTL:
1075 		if (strcmp(a->dst.addr.v.ifname, b->dst.addr.v.ifname) != 0 ||
1076 		    a->dst.addr.iflags != b->dst.addr.iflags ||
1077 		    memcmp(&a->dst.addr.v.a.mask, &b->dst.addr.v.a.mask,
1078 		    sizeof(a->dst.addr.v.a.mask)))
1079 			return (1);
1080 		return (0);
1081 	case PF_ADDR_NOROUTE:
1082 	case PF_ADDR_URPFFAILED:
1083 		return (0);
1084 	case PF_ADDR_TABLE:
1085 		return (strcmp(a->dst.addr.v.tblname, b->dst.addr.v.tblname));
1086 	}
1087 	return (1);
1088 }
1089 
1090 /* Compare two rules DST port field for skiplist construction */
1091 int
1092 skip_cmp_dst_port(struct pf_rule *a, struct pf_rule *b)
1093 {
1094 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1095 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1096 	 *    a->proto == IPPROTO_ICMP
1097 	 *	return (1);
1098 	 */
1099 	if (a->dst.port_op == PF_OP_NONE || a->dst.port_op != b->dst.port_op ||
1100 	    a->dst.port[0] != b->dst.port[0] ||
1101 	    a->dst.port[1] != b->dst.port[1])
1102 		return (1);
1103 	return (0);
1104 }
1105 
1106 /* Compare two rules IFP field for skiplist construction */
1107 int
1108 skip_cmp_ifp(struct pf_rule *a, struct pf_rule *b)
1109 {
1110 	if (strcmp(a->ifname, b->ifname) || a->ifname[0] == '\0')
1111 		return (1);
1112 	return (a->ifnot != b->ifnot);
1113 }
1114 
1115 /* Compare two rules PROTO field for skiplist construction */
1116 int
1117 skip_cmp_proto(struct pf_rule *a, struct pf_rule *b)
1118 {
1119 	return (a->proto != b->proto || a->proto == 0);
1120 }
1121 
1122 /* Compare two rules SRC addr field for skiplist construction */
1123 int
1124 skip_cmp_src_addr(struct pf_rule *a, struct pf_rule *b)
1125 {
1126 	if (a->src.neg != b->src.neg ||
1127 	    a->src.addr.type != b->src.addr.type)
1128 		return (1);
1129 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1130 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1131 	 *    a->proto == IPPROTO_ICMP
1132 	 *	return (1);
1133 	 */
1134 	switch (a->src.addr.type) {
1135 	case PF_ADDR_ADDRMASK:
1136 		if (memcmp(&a->src.addr.v.a.addr, &b->src.addr.v.a.addr,
1137 		    sizeof(a->src.addr.v.a.addr)) ||
1138 		    memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask,
1139 		    sizeof(a->src.addr.v.a.mask)) ||
1140 		    (a->src.addr.v.a.addr.addr32[0] == 0 &&
1141 		    a->src.addr.v.a.addr.addr32[1] == 0 &&
1142 		    a->src.addr.v.a.addr.addr32[2] == 0 &&
1143 		    a->src.addr.v.a.addr.addr32[3] == 0))
1144 			return (1);
1145 		return (0);
1146 	case PF_ADDR_DYNIFTL:
1147 		if (strcmp(a->src.addr.v.ifname, b->src.addr.v.ifname) != 0 ||
1148 		    a->src.addr.iflags != b->src.addr.iflags ||
1149 		    memcmp(&a->src.addr.v.a.mask, &b->src.addr.v.a.mask,
1150 		    sizeof(a->src.addr.v.a.mask)))
1151 			return (1);
1152 		return (0);
1153 	case PF_ADDR_NOROUTE:
1154 	case PF_ADDR_URPFFAILED:
1155 		return (0);
1156 	case PF_ADDR_TABLE:
1157 		return (strcmp(a->src.addr.v.tblname, b->src.addr.v.tblname));
1158 	}
1159 	return (1);
1160 }
1161 
1162 /* Compare two rules SRC port field for skiplist construction */
1163 int
1164 skip_cmp_src_port(struct pf_rule *a, struct pf_rule *b)
1165 {
1166 	if (a->src.port_op == PF_OP_NONE || a->src.port_op != b->src.port_op ||
1167 	    a->src.port[0] != b->src.port[0] ||
1168 	    a->src.port[1] != b->src.port[1])
1169 		return (1);
1170 	/* XXX if (a->proto != b->proto && a->proto != 0 && b->proto != 0
1171 	 *    && (a->proto == IPPROTO_TCP || a->proto == IPPROTO_UDP ||
1172 	 *    a->proto == IPPROTO_ICMP
1173 	 *	return (1);
1174 	 */
1175 	return (0);
1176 }
1177 
1178 
1179 void
1180 skip_init(void)
1181 {
1182 	struct {
1183 		char *name;
1184 		int skipnum;
1185 		int (*func)(struct pf_rule *, struct pf_rule *);
1186 	} comps[] = PF_SKIP_COMPARITORS;
1187 	int skipnum, i;
1188 
1189 	for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++) {
1190 		for (i = 0; i < sizeof(comps)/sizeof(*comps); i++)
1191 			if (comps[i].skipnum == skipnum) {
1192 				skip_comparitors[skipnum] = comps[i].func;
1193 				skip_comparitors_names[skipnum] = comps[i].name;
1194 			}
1195 	}
1196 	for (skipnum = 0; skipnum < PF_SKIP_COUNT; skipnum++)
1197 		if (skip_comparitors[skipnum] == NULL)
1198 			errx(1, "Need to add skip step comparitor to pfctl?!");
1199 }
1200 
1201 /*
1202  * Add a host/netmask to a table
1203  */
1204 int
1205 add_opt_table(struct pfctl *pf, struct pf_opt_tbl **tbl, sa_family_t af,
1206     struct pf_rule_addr *addr, char *ifname)
1207 {
1208 #ifdef OPT_DEBUG
1209 	char buf[128];
1210 #endif /* OPT_DEBUG */
1211 	static int tablenum = 0;
1212 	struct node_host node_host;
1213 
1214 	if (*tbl == NULL) {
1215 		if ((*tbl = calloc(1, sizeof(**tbl))) == NULL ||
1216 		    ((*tbl)->pt_buf = calloc(1, sizeof(*(*tbl)->pt_buf))) ==
1217 		    NULL)
1218 			err(1, "calloc");
1219 		(*tbl)->pt_refcnt = 1;
1220 		(*tbl)->pt_buf->pfrb_type = PFRB_ADDRS;
1221 		SIMPLEQ_INIT(&(*tbl)->pt_nodes);
1222 
1223 		/* This is just a temporary table name */
1224 		snprintf((*tbl)->pt_name, sizeof((*tbl)->pt_name), "%s%d",
1225 		    PF_OPT_TABLE_PREFIX, tablenum++);
1226 		DEBUG("creating table <%s>", (*tbl)->pt_name);
1227 	}
1228 
1229 	memset(&node_host, 0, sizeof(node_host));
1230 	node_host.af = af;
1231 	node_host.addr = addr->addr;
1232 	node_host.ifname = ifname;
1233 	node_host.weight = addr->weight;
1234 
1235 	DEBUG("<%s> adding %s/%d", (*tbl)->pt_name, inet_ntop(af,
1236 	    &node_host.addr.v.a.addr, buf, sizeof(buf)),
1237 	    unmask(&node_host.addr.v.a.mask));
1238 
1239 	if (append_addr_host((*tbl)->pt_buf, &node_host, 0, 0)) {
1240 		warn("failed to add host");
1241 		return (1);
1242 	}
1243 	if (pf->opts & PF_OPT_VERBOSE) {
1244 		struct node_tinit *ti;
1245 
1246 		if ((ti = calloc(1, sizeof(*ti))) == NULL)
1247 			err(1, "malloc");
1248 		if ((ti->host = malloc(sizeof(*ti->host))) == NULL)
1249 			err(1, "malloc");
1250 		memcpy(ti->host, &node_host, sizeof(*ti->host));
1251 		SIMPLEQ_INSERT_TAIL(&(*tbl)->pt_nodes, ti, entries);
1252 	}
1253 
1254 	(*tbl)->pt_rulecount++;
1255 	if ((*tbl)->pt_rulecount == TABLE_THRESHOLD)
1256 		DEBUG("table <%s> now faster than skip steps", (*tbl)->pt_name);
1257 
1258 	return (0);
1259 }
1260 
1261 
1262 /*
1263  * Do the dirty work of choosing an unused table name and creating it.
1264  * (be careful with the table name, it might already be used in another anchor)
1265  */
1266 int
1267 pf_opt_create_table(struct pfctl *pf, struct pf_opt_tbl *tbl)
1268 {
1269 	static int tablenum;
1270 	struct pfr_table *t;
1271 
1272 	if (table_buffer.pfrb_type == 0) {
1273 		/* Initialize the list of tables */
1274 		table_buffer.pfrb_type = PFRB_TABLES;
1275 		for (;;) {
1276 			pfr_buf_grow(&table_buffer, table_buffer.pfrb_size);
1277 			table_buffer.pfrb_size = table_buffer.pfrb_msize;
1278 			if (pfr_get_tables(NULL, table_buffer.pfrb_caddr,
1279 			    &table_buffer.pfrb_size, PFR_FLAG_ALLRSETS))
1280 				err(1, "pfr_get_tables");
1281 			if (table_buffer.pfrb_size <= table_buffer.pfrb_msize)
1282 				break;
1283 		}
1284 		table_identifier = arc4random();
1285 	}
1286 
1287 	/* XXX would be *really* nice to avoid duplicating identical tables */
1288 
1289 	/* Now we have to pick a table name that isn't used */
1290 again:
1291 	DEBUG("translating temporary table <%s> to <%s%x_%d>", tbl->pt_name,
1292 	    PF_OPT_TABLE_PREFIX, table_identifier, tablenum);
1293 	snprintf(tbl->pt_name, sizeof(tbl->pt_name), "%s%x_%d",
1294 	    PF_OPT_TABLE_PREFIX, table_identifier, tablenum);
1295 	PFRB_FOREACH(t, &table_buffer) {
1296 		if (strcasecmp(t->pfrt_name, tbl->pt_name) == 0) {
1297 			/* Collision.  Try again */
1298 			DEBUG("wow, table <%s> in use.  trying again",
1299 			    tbl->pt_name);
1300 			table_identifier = arc4random();
1301 			goto again;
1302 		}
1303 	}
1304 	tablenum++;
1305 
1306 	if (pfctl_define_table(tbl->pt_name, PFR_TFLAG_CONST | tbl->pt_flags, 1,
1307 	    pf->astack[0]->path, tbl->pt_buf, pf->astack[0]->ruleset.tticket)) {
1308 		warn("failed to create table %s in %s",
1309 		    tbl->pt_name, pf->astack[0]->name);
1310 		return (1);
1311 	}
1312 	return (0);
1313 }
1314 
1315 /*
1316  * Partition the flat ruleset into a list of distinct superblocks
1317  */
1318 int
1319 construct_superblocks(struct pfctl *pf, struct pf_opt_queue *opt_queue,
1320     struct superblocks *superblocks)
1321 {
1322 	struct superblock *block = NULL;
1323 	struct pf_opt_rule *por;
1324 	int i;
1325 
1326 	while (!TAILQ_EMPTY(opt_queue)) {
1327 		por = TAILQ_FIRST(opt_queue);
1328 		TAILQ_REMOVE(opt_queue, por, por_entry);
1329 		if (block == NULL || !superblock_inclusive(block, por)) {
1330 			if ((block = calloc(1, sizeof(*block))) == NULL) {
1331 				warn("calloc");
1332 				return (1);
1333 			}
1334 			TAILQ_INIT(&block->sb_rules);
1335 			for (i = 0; i < PF_SKIP_COUNT; i++)
1336 				TAILQ_INIT(&block->sb_skipsteps[i]);
1337 			TAILQ_INSERT_TAIL(superblocks, block, sb_entry);
1338 		}
1339 		TAILQ_INSERT_TAIL(&block->sb_rules, por, por_entry);
1340 	}
1341 
1342 	return (0);
1343 }
1344 
1345 
1346 /*
1347  * Compare two rule addresses
1348  */
1349 int
1350 addrs_equal(struct pf_rule_addr *a, struct pf_rule_addr *b)
1351 {
1352 	if (a->neg != b->neg)
1353 		return (0);
1354 	return (memcmp(&a->addr, &b->addr, sizeof(a->addr)) == 0);
1355 }
1356 
1357 
1358 /*
1359  * The addresses are not equal, but can we combine them into one table?
1360  */
1361 int
1362 addrs_combineable(struct pf_rule_addr *a, struct pf_rule_addr *b)
1363 {
1364 	if (a->addr.type != PF_ADDR_ADDRMASK ||
1365 	    b->addr.type != PF_ADDR_ADDRMASK)
1366 		return (0);
1367 	if (a->neg != b->neg || a->port_op != b->port_op ||
1368 	    a->port[0] != b->port[0] || a->port[1] != b->port[1])
1369 		return (0);
1370 	return (1);
1371 }
1372 
1373 
1374 /*
1375  * Are we allowed to combine these two rules
1376  */
1377 int
1378 rules_combineable(struct pf_rule *p1, struct pf_rule *p2)
1379 {
1380 	struct pf_rule a, b;
1381 
1382 	comparable_rule(&a, p1, COMBINED);
1383 	comparable_rule(&b, p2, COMBINED);
1384 	return (memcmp(&a, &b, sizeof(a)) == 0);
1385 }
1386 
1387 
1388 /*
1389  * Can a rule be included inside a superblock
1390  */
1391 int
1392 superblock_inclusive(struct superblock *block, struct pf_opt_rule *por)
1393 {
1394 	struct pf_rule a, b;
1395 	int i, j;
1396 
1397 	/* First check for hard breaks */
1398 	for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++) {
1399 		if (pf_rule_desc[i].prf_type == BARRIER) {
1400 			for (j = 0; j < pf_rule_desc[i].prf_size; j++)
1401 				if (((char *)&por->por_rule)[j +
1402 				    pf_rule_desc[i].prf_offset] != 0)
1403 					return (0);
1404 		}
1405 	}
1406 
1407 	/* per-rule src-track is also a hard break */
1408 	if (por->por_rule.rule_flag & PFRULE_RULESRCTRACK)
1409 		return (0);
1410 
1411 	/*
1412 	 * Have to handle interface groups separately.  Consider the following
1413 	 * rules:
1414 	 *	block on EXTIFS to any port 22
1415 	 *	pass  on em0 to any port 22
1416 	 * (where EXTIFS is an arbitrary interface group)
1417 	 * The optimizer may decide to re-order the pass rule in front of the
1418 	 * block rule.  But what if EXTIFS includes em0???  Such a reordering
1419 	 * would change the meaning of the ruleset.
1420 	 * We can't just lookup the EXTIFS group and check if em0 is a member
1421 	 * because the user is allowed to add interfaces to a group during
1422 	 * runtime.
1423 	 * Ergo interface groups become a defacto superblock break :-(
1424 	 */
1425 	if (interface_group(por->por_rule.ifname) ||
1426 	    interface_group(TAILQ_FIRST(&block->sb_rules)->por_rule.ifname)) {
1427 		if (strcasecmp(por->por_rule.ifname,
1428 		    TAILQ_FIRST(&block->sb_rules)->por_rule.ifname) != 0)
1429 			return (0);
1430 	}
1431 
1432 	comparable_rule(&a, &TAILQ_FIRST(&block->sb_rules)->por_rule, NOMERGE);
1433 	comparable_rule(&b, &por->por_rule, NOMERGE);
1434 	if (memcmp(&a, &b, sizeof(a)) == 0)
1435 		return (1);
1436 
1437 #ifdef OPT_DEBUG
1438 	for (i = 0; i < sizeof(por->por_rule); i++) {
1439 		int closest = -1;
1440 		if (((u_int8_t *)&a)[i] != ((u_int8_t *)&b)[i]) {
1441 			for (j = 0; j < sizeof(pf_rule_desc) /
1442 			    sizeof(*pf_rule_desc); j++) {
1443 				if (i >= pf_rule_desc[j].prf_offset &&
1444 				    i < pf_rule_desc[j].prf_offset +
1445 				    pf_rule_desc[j].prf_size) {
1446 					DEBUG("superblock break @ %d due to %s",
1447 					    por->por_rule.nr,
1448 					    pf_rule_desc[j].prf_name);
1449 					return (0);
1450 				}
1451 				if (i > pf_rule_desc[j].prf_offset) {
1452 					if (closest == -1 ||
1453 					    i-pf_rule_desc[j].prf_offset <
1454 					    i-pf_rule_desc[closest].prf_offset)
1455 						closest = j;
1456 				}
1457 			}
1458 
1459 			if (closest >= 0)
1460 				DEBUG("superblock break @ %d on %s+%lxh",
1461 				    por->por_rule.nr,
1462 				    pf_rule_desc[closest].prf_name,
1463 				    i - pf_rule_desc[closest].prf_offset -
1464 				    pf_rule_desc[closest].prf_size);
1465 			else
1466 				DEBUG("superblock break @ %d on field @ %d",
1467 				    por->por_rule.nr, i);
1468 			return (0);
1469 		}
1470 	}
1471 #endif /* OPT_DEBUG */
1472 
1473 	return (0);
1474 }
1475 
1476 
1477 /*
1478  * Figure out if an interface name is an actual interface or actually a
1479  * group of interfaces.
1480  */
1481 int
1482 interface_group(const char *ifname)
1483 {
1484 	if (ifname == NULL || !ifname[0])
1485 		return (0);
1486 
1487 	/* Real interfaces must end in a number, interface groups do not */
1488 	if (isdigit((unsigned char)ifname[strlen(ifname) - 1]))
1489 		return (0);
1490 	else
1491 		return (1);
1492 }
1493 
1494 
1495 /*
1496  * Make a rule that can directly compared by memcmp()
1497  */
1498 void
1499 comparable_rule(struct pf_rule *dst, const struct pf_rule *src, int type)
1500 {
1501 	int i;
1502 	/*
1503 	 * To simplify the comparison, we just zero out the fields that are
1504 	 * allowed to be different and then do a simple memcmp()
1505 	 */
1506 	memcpy(dst, src, sizeof(*dst));
1507 	for (i = 0; i < sizeof(pf_rule_desc)/sizeof(*pf_rule_desc); i++)
1508 		if (pf_rule_desc[i].prf_type >= type) {
1509 #ifdef OPT_DEBUG
1510 			assert(pf_rule_desc[i].prf_type != NEVER ||
1511 			    *(((char *)dst) + pf_rule_desc[i].prf_offset) == 0);
1512 #endif /* OPT_DEBUG */
1513 			memset(((char *)dst) + pf_rule_desc[i].prf_offset, 0,
1514 			    pf_rule_desc[i].prf_size);
1515 		}
1516 }
1517 
1518 
1519 /*
1520  * Remove superset information from two rules so we can directly compare them
1521  * with memcmp()
1522  */
1523 void
1524 exclude_supersets(struct pf_rule *super, struct pf_rule *sub)
1525 {
1526 	if (super->ifname[0] == '\0')
1527 		memset(sub->ifname, 0, sizeof(sub->ifname));
1528 	if (super->direction == PF_INOUT)
1529 		sub->direction = PF_INOUT;
1530 	if ((super->proto == 0 || super->proto == sub->proto) &&
1531 	    super->flags == 0 && super->flagset == 0 && (sub->flags ||
1532 	    sub->flagset)) {
1533 		sub->flags = super->flags;
1534 		sub->flagset = super->flagset;
1535 	}
1536 	if (super->proto == 0)
1537 		sub->proto = 0;
1538 
1539 	if (super->src.port_op == 0) {
1540 		sub->src.port_op = 0;
1541 		sub->src.port[0] = 0;
1542 		sub->src.port[1] = 0;
1543 	}
1544 	if (super->dst.port_op == 0) {
1545 		sub->dst.port_op = 0;
1546 		sub->dst.port[0] = 0;
1547 		sub->dst.port[1] = 0;
1548 	}
1549 
1550 	if (super->src.addr.type == PF_ADDR_ADDRMASK && !super->src.neg &&
1551 	    !sub->src.neg && super->src.addr.v.a.mask.addr32[0] == 0 &&
1552 	    super->src.addr.v.a.mask.addr32[1] == 0 &&
1553 	    super->src.addr.v.a.mask.addr32[2] == 0 &&
1554 	    super->src.addr.v.a.mask.addr32[3] == 0)
1555 		memset(&sub->src.addr, 0, sizeof(sub->src.addr));
1556 	else if (super->src.addr.type == PF_ADDR_ADDRMASK &&
1557 	    sub->src.addr.type == PF_ADDR_ADDRMASK &&
1558 	    super->src.neg == sub->src.neg &&
1559 	    super->af == sub->af &&
1560 	    unmask(&super->src.addr.v.a.mask) <
1561 	    unmask(&sub->src.addr.v.a.mask) &&
1562 	    super->src.addr.v.a.addr.addr32[0] ==
1563 	    (sub->src.addr.v.a.addr.addr32[0] &
1564 	    super->src.addr.v.a.mask.addr32[0]) &&
1565 	    super->src.addr.v.a.addr.addr32[1] ==
1566 	    (sub->src.addr.v.a.addr.addr32[1] &
1567 	    super->src.addr.v.a.mask.addr32[1]) &&
1568 	    super->src.addr.v.a.addr.addr32[2] ==
1569 	    (sub->src.addr.v.a.addr.addr32[2] &
1570 	    super->src.addr.v.a.mask.addr32[2]) &&
1571 	    super->src.addr.v.a.addr.addr32[3] ==
1572 	    (sub->src.addr.v.a.addr.addr32[3] &
1573 	    super->src.addr.v.a.mask.addr32[3])) {
1574 		/* sub->src.addr is a subset of super->src.addr/mask */
1575 		memcpy(&sub->src.addr, &super->src.addr, sizeof(sub->src.addr));
1576 	}
1577 
1578 	if (super->dst.addr.type == PF_ADDR_ADDRMASK && !super->dst.neg &&
1579 	    !sub->dst.neg && super->dst.addr.v.a.mask.addr32[0] == 0 &&
1580 	    super->dst.addr.v.a.mask.addr32[1] == 0 &&
1581 	    super->dst.addr.v.a.mask.addr32[2] == 0 &&
1582 	    super->dst.addr.v.a.mask.addr32[3] == 0)
1583 		memset(&sub->dst.addr, 0, sizeof(sub->dst.addr));
1584 	else if (super->dst.addr.type == PF_ADDR_ADDRMASK &&
1585 	    sub->dst.addr.type == PF_ADDR_ADDRMASK &&
1586 	    super->dst.neg == sub->dst.neg &&
1587 	    super->af == sub->af &&
1588 	    unmask(&super->dst.addr.v.a.mask) <
1589 	    unmask(&sub->dst.addr.v.a.mask) &&
1590 	    super->dst.addr.v.a.addr.addr32[0] ==
1591 	    (sub->dst.addr.v.a.addr.addr32[0] &
1592 	    super->dst.addr.v.a.mask.addr32[0]) &&
1593 	    super->dst.addr.v.a.addr.addr32[1] ==
1594 	    (sub->dst.addr.v.a.addr.addr32[1] &
1595 	    super->dst.addr.v.a.mask.addr32[1]) &&
1596 	    super->dst.addr.v.a.addr.addr32[2] ==
1597 	    (sub->dst.addr.v.a.addr.addr32[2] &
1598 	    super->dst.addr.v.a.mask.addr32[2]) &&
1599 	    super->dst.addr.v.a.addr.addr32[3] ==
1600 	    (sub->dst.addr.v.a.addr.addr32[3] &
1601 	    super->dst.addr.v.a.mask.addr32[3])) {
1602 		/* sub->dst.addr is a subset of super->dst.addr/mask */
1603 		memcpy(&sub->dst.addr, &super->dst.addr, sizeof(sub->dst.addr));
1604 	}
1605 
1606 	if (super->af == 0)
1607 		sub->af = 0;
1608 }
1609 
1610 
1611 void
1612 superblock_free(struct pfctl *pf, struct superblock *block)
1613 {
1614 	struct pf_opt_rule *por;
1615 	while ((por = TAILQ_FIRST(&block->sb_rules))) {
1616 		TAILQ_REMOVE(&block->sb_rules, por, por_entry);
1617 		pf_opt_table_unref(por->por_src_tbl);
1618 		pf_opt_table_unref(por->por_dst_tbl);
1619 		free(por);
1620 	}
1621 	if (block->sb_profiled_block)
1622 		superblock_free(pf, block->sb_profiled_block);
1623 	free(block);
1624 }
1625 
1626 struct	pf_opt_tbl *
1627 pf_opt_table_ref(struct pf_opt_tbl *pt)
1628 {
1629 	/* parser does not run concurrently, we don't need atomic ops. */
1630 	if (pt != NULL)
1631 		pt->pt_refcnt++;
1632 
1633 	return (pt);
1634 }
1635 
1636 void
1637 pf_opt_table_unref(struct pf_opt_tbl *pt)
1638 {
1639 	if ((pt != NULL) && ((--pt->pt_refcnt) == 0)) {
1640 		if (pt->pt_buf != NULL) {
1641 			pfr_buf_clear(pt->pt_buf);
1642 			free(pt->pt_buf);
1643 		}
1644 		free(pt);
1645 	}
1646 }
1647