xref: /freebsd-13.1/sys/netpfil/pf/pf_ioctl.c (revision a409ae58)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2001 Daniel Hartmeier
5  * Copyright (c) 2002,2003 Henning Brauer
6  * Copyright (c) 2012 Gleb Smirnoff <[email protected]>
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  *
13  *    - Redistributions of source code must retain the above copyright
14  *      notice, this list of conditions and the following disclaimer.
15  *    - Redistributions in binary form must reproduce the above
16  *      copyright notice, this list of conditions and the following
17  *      disclaimer in the documentation and/or other materials provided
18  *      with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
28  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
30  * ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
31  * POSSIBILITY OF SUCH DAMAGE.
32  *
33  * Effort sponsored in part by the Defense Advanced Research Projects
34  * Agency (DARPA) and Air Force Research Laboratory, Air Force
35  * Materiel Command, USAF, under agreement number F30602-01-2-0537.
36  *
37  *	$OpenBSD: pf_ioctl.c,v 1.213 2009/02/15 21:46:12 mbalmer Exp $
38  */
39 
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
42 
43 #include "opt_inet.h"
44 #include "opt_inet6.h"
45 #include "opt_bpf.h"
46 #include "opt_pf.h"
47 
48 #include <sys/param.h>
49 #include <sys/_bitset.h>
50 #include <sys/bitset.h>
51 #include <sys/bus.h>
52 #include <sys/conf.h>
53 #include <sys/endian.h>
54 #include <sys/fcntl.h>
55 #include <sys/filio.h>
56 #include <sys/hash.h>
57 #include <sys/interrupt.h>
58 #include <sys/jail.h>
59 #include <sys/kernel.h>
60 #include <sys/kthread.h>
61 #include <sys/lock.h>
62 #include <sys/mbuf.h>
63 #include <sys/module.h>
64 #include <sys/nv.h>
65 #include <sys/proc.h>
66 #include <sys/sdt.h>
67 #include <sys/smp.h>
68 #include <sys/socket.h>
69 #include <sys/sysctl.h>
70 #include <sys/md5.h>
71 #include <sys/ucred.h>
72 
73 #include <net/if.h>
74 #include <net/if_var.h>
75 #include <net/vnet.h>
76 #include <net/route.h>
77 #include <net/pfil.h>
78 #include <net/pfvar.h>
79 #include <net/if_pfsync.h>
80 #include <net/if_pflog.h>
81 
82 #include <netinet/in.h>
83 #include <netinet/ip.h>
84 #include <netinet/ip_var.h>
85 #include <netinet6/ip6_var.h>
86 #include <netinet/ip_icmp.h>
87 #include <netpfil/pf/pf_nv.h>
88 
89 #ifdef INET6
90 #include <netinet/ip6.h>
91 #endif /* INET6 */
92 
93 #ifdef ALTQ
94 #include <net/altq/altq.h>
95 #endif
96 
97 SDT_PROVIDER_DECLARE(pf);
98 SDT_PROBE_DEFINE3(pf, ioctl, ioctl, error, "int", "int", "int");
99 SDT_PROBE_DEFINE3(pf, ioctl, function, error, "char *", "int", "int");
100 SDT_PROBE_DEFINE2(pf, ioctl, addrule, error, "int", "int");
101 SDT_PROBE_DEFINE2(pf, ioctl, nvchk, error, "int", "int");
102 
103 static struct pf_kpool	*pf_get_kpool(const char *, u_int32_t, u_int8_t,
104 			    u_int32_t, u_int8_t, u_int8_t, u_int8_t);
105 
106 static void		 pf_mv_kpool(struct pf_kpalist *, struct pf_kpalist *);
107 static void		 pf_empty_kpool(struct pf_kpalist *);
108 static int		 pfioctl(struct cdev *, u_long, caddr_t, int,
109 			    struct thread *);
110 #ifdef ALTQ
111 static int		 pf_begin_altq(u_int32_t *);
112 static int		 pf_rollback_altq(u_int32_t);
113 static int		 pf_commit_altq(u_int32_t);
114 static int		 pf_enable_altq(struct pf_altq *);
115 static int		 pf_disable_altq(struct pf_altq *);
116 static uint16_t		 pf_qname2qid(const char *);
117 static void		 pf_qid_unref(uint16_t);
118 #endif /* ALTQ */
119 static int		 pf_begin_rules(u_int32_t *, int, const char *);
120 static int		 pf_rollback_rules(u_int32_t, int, char *);
121 static int		 pf_setup_pfsync_matching(struct pf_kruleset *);
122 static void		 pf_hash_rule(MD5_CTX *, struct pf_krule *);
123 static void		 pf_hash_rule_addr(MD5_CTX *, struct pf_rule_addr *);
124 static int		 pf_commit_rules(u_int32_t, int, char *);
125 static int		 pf_addr_setup(struct pf_kruleset *,
126 			    struct pf_addr_wrap *, sa_family_t);
127 static void		 pf_addr_copyout(struct pf_addr_wrap *);
128 static void		 pf_src_node_copy(const struct pf_ksrc_node *,
129 			    struct pf_src_node *);
130 #ifdef ALTQ
131 static int		 pf_export_kaltq(struct pf_altq *,
132 			    struct pfioc_altq_v1 *, size_t);
133 static int		 pf_import_kaltq(struct pfioc_altq_v1 *,
134 			    struct pf_altq *, size_t);
135 #endif /* ALTQ */
136 
137 VNET_DEFINE(struct pf_krule,	pf_default_rule);
138 
139 #ifdef ALTQ
140 VNET_DEFINE_STATIC(int,		pf_altq_running);
141 #define	V_pf_altq_running	VNET(pf_altq_running)
142 #endif
143 
144 #define	TAGID_MAX	 50000
145 struct pf_tagname {
146 	TAILQ_ENTRY(pf_tagname)	namehash_entries;
147 	TAILQ_ENTRY(pf_tagname)	taghash_entries;
148 	char			name[PF_TAG_NAME_SIZE];
149 	uint16_t		tag;
150 	int			ref;
151 };
152 
153 struct pf_tagset {
154 	TAILQ_HEAD(, pf_tagname)	*namehash;
155 	TAILQ_HEAD(, pf_tagname)	*taghash;
156 	unsigned int			 mask;
157 	uint32_t			 seed;
158 	BITSET_DEFINE(, TAGID_MAX)	 avail;
159 };
160 
161 VNET_DEFINE(struct pf_tagset, pf_tags);
162 #define	V_pf_tags	VNET(pf_tags)
163 static unsigned int	pf_rule_tag_hashsize;
164 #define	PF_RULE_TAG_HASH_SIZE_DEFAULT	128
165 SYSCTL_UINT(_net_pf, OID_AUTO, rule_tag_hashsize, CTLFLAG_RDTUN,
166     &pf_rule_tag_hashsize, PF_RULE_TAG_HASH_SIZE_DEFAULT,
167     "Size of pf(4) rule tag hashtable");
168 
169 #ifdef ALTQ
170 VNET_DEFINE(struct pf_tagset, pf_qids);
171 #define	V_pf_qids	VNET(pf_qids)
172 static unsigned int	pf_queue_tag_hashsize;
173 #define	PF_QUEUE_TAG_HASH_SIZE_DEFAULT	128
174 SYSCTL_UINT(_net_pf, OID_AUTO, queue_tag_hashsize, CTLFLAG_RDTUN,
175     &pf_queue_tag_hashsize, PF_QUEUE_TAG_HASH_SIZE_DEFAULT,
176     "Size of pf(4) queue tag hashtable");
177 #endif
178 VNET_DEFINE(uma_zone_t,	 pf_tag_z);
179 #define	V_pf_tag_z		 VNET(pf_tag_z)
180 static MALLOC_DEFINE(M_PFALTQ, "pf_altq", "pf(4) altq configuration db");
181 static MALLOC_DEFINE(M_PFRULE, "pf_rule", "pf(4) rules");
182 
183 #if (PF_QNAME_SIZE != PF_TAG_NAME_SIZE)
184 #error PF_QNAME_SIZE must be equal to PF_TAG_NAME_SIZE
185 #endif
186 
187 static void		 pf_init_tagset(struct pf_tagset *, unsigned int *,
188 			    unsigned int);
189 static void		 pf_cleanup_tagset(struct pf_tagset *);
190 static uint16_t		 tagname2hashindex(const struct pf_tagset *, const char *);
191 static uint16_t		 tag2hashindex(const struct pf_tagset *, uint16_t);
192 static u_int16_t	 tagname2tag(struct pf_tagset *, const char *);
193 static u_int16_t	 pf_tagname2tag(const char *);
194 static void		 tag_unref(struct pf_tagset *, u_int16_t);
195 
196 #define DPFPRINTF(n, x) if (V_pf_status.debug >= (n)) printf x
197 
198 struct cdev *pf_dev;
199 
200 /*
201  * XXX - These are new and need to be checked when moveing to a new version
202  */
203 static void		 pf_clear_all_states(void);
204 static unsigned int	 pf_clear_states(const struct pf_kstate_kill *);
205 static void		 pf_killstates(struct pf_kstate_kill *,
206 			    unsigned int *);
207 static int		 pf_killstates_row(struct pf_kstate_kill *,
208 			    struct pf_idhash *);
209 static int		 pf_killstates_nv(struct pfioc_nv *);
210 static int		 pf_clearstates_nv(struct pfioc_nv *);
211 static int		 pf_getstate(struct pfioc_nv *);
212 static int		 pf_getstatus(struct pfioc_nv *);
213 static int		 pf_clear_tables(void);
214 static void		 pf_clear_srcnodes(struct pf_ksrc_node *);
215 static void		 pf_kill_srcnodes(struct pfioc_src_node_kill *);
216 static int		 pf_keepcounters(struct pfioc_nv *);
217 static void		 pf_tbladdr_copyout(struct pf_addr_wrap *);
218 
219 /*
220  * Wrapper functions for pfil(9) hooks
221  */
222 #ifdef INET
223 static pfil_return_t pf_check_in(struct mbuf **m, struct ifnet *ifp,
224     int flags, void *ruleset __unused, struct inpcb *inp);
225 static pfil_return_t pf_check_out(struct mbuf **m, struct ifnet *ifp,
226     int flags, void *ruleset __unused, struct inpcb *inp);
227 #endif
228 #ifdef INET6
229 static pfil_return_t pf_check6_in(struct mbuf **m, struct ifnet *ifp,
230     int flags, void *ruleset __unused, struct inpcb *inp);
231 static pfil_return_t pf_check6_out(struct mbuf **m, struct ifnet *ifp,
232     int flags, void *ruleset __unused, struct inpcb *inp);
233 #endif
234 
235 static void		hook_pf(void);
236 static void		dehook_pf(void);
237 static int		shutdown_pf(void);
238 static int		pf_load(void);
239 static void		pf_unload(void);
240 
241 static struct cdevsw pf_cdevsw = {
242 	.d_ioctl =	pfioctl,
243 	.d_name =	PF_NAME,
244 	.d_version =	D_VERSION,
245 };
246 
247 volatile VNET_DEFINE_STATIC(int, pf_pfil_hooked);
248 #define V_pf_pfil_hooked	VNET(pf_pfil_hooked)
249 
250 /*
251  * We need a flag that is neither hooked nor running to know when
252  * the VNET is "valid".  We primarily need this to control (global)
253  * external event, e.g., eventhandlers.
254  */
255 VNET_DEFINE(int, pf_vnet_active);
256 #define V_pf_vnet_active	VNET(pf_vnet_active)
257 
258 int pf_end_threads;
259 struct proc *pf_purge_proc;
260 
261 struct rmlock			pf_rules_lock;
262 struct sx			pf_ioctl_lock;
263 struct sx			pf_end_lock;
264 
265 /* pfsync */
266 VNET_DEFINE(pfsync_state_import_t *, pfsync_state_import_ptr);
267 VNET_DEFINE(pfsync_insert_state_t *, pfsync_insert_state_ptr);
268 VNET_DEFINE(pfsync_update_state_t *, pfsync_update_state_ptr);
269 VNET_DEFINE(pfsync_delete_state_t *, pfsync_delete_state_ptr);
270 VNET_DEFINE(pfsync_clear_states_t *, pfsync_clear_states_ptr);
271 VNET_DEFINE(pfsync_defer_t *, pfsync_defer_ptr);
272 pfsync_detach_ifnet_t *pfsync_detach_ifnet_ptr;
273 
274 /* pflog */
275 pflog_packet_t			*pflog_packet_ptr = NULL;
276 
277 #define	ERROUT_FUNCTION(target, x)					\
278 	do {								\
279 		error = (x);						\
280 		SDT_PROBE3(pf, ioctl, function, error, __func__, error,	\
281 		    __LINE__);						\
282 		goto target;						\
283 	} while (0)
284 
285 /*
286  * Copy a user-provided string, returning an error if truncation would occur.
287  * Avoid scanning past "sz" bytes in the source string since there's no
288  * guarantee that it's nul-terminated.
289  */
290 static int
pf_user_strcpy(char * dst,const char * src,size_t sz)291 pf_user_strcpy(char *dst, const char *src, size_t sz)
292 {
293 	if (strnlen(src, sz) == sz)
294 		return (EINVAL);
295 	(void)strlcpy(dst, src, sz);
296 	return (0);
297 }
298 
299 static void
pfattach_vnet(void)300 pfattach_vnet(void)
301 {
302 	u_int32_t *my_timeout = V_pf_default_rule.timeout;
303 
304 	pf_initialize();
305 	pfr_initialize();
306 	pfi_initialize_vnet();
307 	pf_normalize_init();
308 	pf_syncookies_init();
309 
310 	V_pf_limits[PF_LIMIT_STATES].limit = PFSTATE_HIWAT;
311 	V_pf_limits[PF_LIMIT_SRC_NODES].limit = PFSNODE_HIWAT;
312 
313 	RB_INIT(&V_pf_anchors);
314 	pf_init_kruleset(&pf_main_ruleset);
315 
316 	/* default rule should never be garbage collected */
317 	V_pf_default_rule.entries.tqe_prev = &V_pf_default_rule.entries.tqe_next;
318 #ifdef PF_DEFAULT_TO_DROP
319 	V_pf_default_rule.action = PF_DROP;
320 #else
321 	V_pf_default_rule.action = PF_PASS;
322 #endif
323 	V_pf_default_rule.nr = -1;
324 	V_pf_default_rule.rtableid = -1;
325 
326 	pf_counter_u64_init(&V_pf_default_rule.evaluations, M_WAITOK);
327 	for (int i = 0; i < 2; i++) {
328 		pf_counter_u64_init(&V_pf_default_rule.packets[i], M_WAITOK);
329 		pf_counter_u64_init(&V_pf_default_rule.bytes[i], M_WAITOK);
330 	}
331 	V_pf_default_rule.states_cur = counter_u64_alloc(M_WAITOK);
332 	V_pf_default_rule.states_tot = counter_u64_alloc(M_WAITOK);
333 	V_pf_default_rule.src_nodes = counter_u64_alloc(M_WAITOK);
334 
335 #ifdef PF_WANT_32_TO_64_COUNTER
336 	V_pf_kifmarker = malloc(sizeof(*V_pf_kifmarker), PFI_MTYPE, M_WAITOK | M_ZERO);
337 	V_pf_rulemarker = malloc(sizeof(*V_pf_rulemarker), M_PFRULE, M_WAITOK | M_ZERO);
338 	PF_RULES_WLOCK();
339 	LIST_INSERT_HEAD(&V_pf_allkiflist, V_pf_kifmarker, pfik_allkiflist);
340 	LIST_INSERT_HEAD(&V_pf_allrulelist, &V_pf_default_rule, allrulelist);
341 	V_pf_allrulecount++;
342 	LIST_INSERT_HEAD(&V_pf_allrulelist, V_pf_rulemarker, allrulelist);
343 	PF_RULES_WUNLOCK();
344 #endif
345 
346 	/* initialize default timeouts */
347 	my_timeout[PFTM_TCP_FIRST_PACKET] = PFTM_TCP_FIRST_PACKET_VAL;
348 	my_timeout[PFTM_TCP_OPENING] = PFTM_TCP_OPENING_VAL;
349 	my_timeout[PFTM_TCP_ESTABLISHED] = PFTM_TCP_ESTABLISHED_VAL;
350 	my_timeout[PFTM_TCP_CLOSING] = PFTM_TCP_CLOSING_VAL;
351 	my_timeout[PFTM_TCP_FIN_WAIT] = PFTM_TCP_FIN_WAIT_VAL;
352 	my_timeout[PFTM_TCP_CLOSED] = PFTM_TCP_CLOSED_VAL;
353 	my_timeout[PFTM_UDP_FIRST_PACKET] = PFTM_UDP_FIRST_PACKET_VAL;
354 	my_timeout[PFTM_UDP_SINGLE] = PFTM_UDP_SINGLE_VAL;
355 	my_timeout[PFTM_UDP_MULTIPLE] = PFTM_UDP_MULTIPLE_VAL;
356 	my_timeout[PFTM_ICMP_FIRST_PACKET] = PFTM_ICMP_FIRST_PACKET_VAL;
357 	my_timeout[PFTM_ICMP_ERROR_REPLY] = PFTM_ICMP_ERROR_REPLY_VAL;
358 	my_timeout[PFTM_OTHER_FIRST_PACKET] = PFTM_OTHER_FIRST_PACKET_VAL;
359 	my_timeout[PFTM_OTHER_SINGLE] = PFTM_OTHER_SINGLE_VAL;
360 	my_timeout[PFTM_OTHER_MULTIPLE] = PFTM_OTHER_MULTIPLE_VAL;
361 	my_timeout[PFTM_FRAG] = PFTM_FRAG_VAL;
362 	my_timeout[PFTM_INTERVAL] = PFTM_INTERVAL_VAL;
363 	my_timeout[PFTM_SRC_NODE] = PFTM_SRC_NODE_VAL;
364 	my_timeout[PFTM_TS_DIFF] = PFTM_TS_DIFF_VAL;
365 	my_timeout[PFTM_ADAPTIVE_START] = PFSTATE_ADAPT_START;
366 	my_timeout[PFTM_ADAPTIVE_END] = PFSTATE_ADAPT_END;
367 
368 	bzero(&V_pf_status, sizeof(V_pf_status));
369 	V_pf_status.debug = PF_DEBUG_URGENT;
370 
371 	V_pf_pfil_hooked = 0;
372 
373 	/* XXX do our best to avoid a conflict */
374 	V_pf_status.hostid = arc4random();
375 
376 	for (int i = 0; i < PFRES_MAX; i++)
377 		V_pf_status.counters[i] = counter_u64_alloc(M_WAITOK);
378 	for (int i = 0; i < KLCNT_MAX; i++)
379 		V_pf_status.lcounters[i] = counter_u64_alloc(M_WAITOK);
380 	for (int i = 0; i < FCNT_MAX; i++)
381 		pf_counter_u64_init(&V_pf_status.fcounters[i], M_WAITOK);
382 	for (int i = 0; i < SCNT_MAX; i++)
383 		V_pf_status.scounters[i] = counter_u64_alloc(M_WAITOK);
384 
385 	if (swi_add(&V_pf_swi_ie, "pf send", pf_intr, curvnet, SWI_NET,
386 	    INTR_MPSAFE, &V_pf_swi_cookie) != 0)
387 		/* XXXGL: leaked all above. */
388 		return;
389 }
390 
391 static struct pf_kpool *
pf_get_kpool(const char * anchor,u_int32_t ticket,u_int8_t rule_action,u_int32_t rule_number,u_int8_t r_last,u_int8_t active,u_int8_t check_ticket)392 pf_get_kpool(const char *anchor, u_int32_t ticket, u_int8_t rule_action,
393     u_int32_t rule_number, u_int8_t r_last, u_int8_t active,
394     u_int8_t check_ticket)
395 {
396 	struct pf_kruleset	*ruleset;
397 	struct pf_krule		*rule;
398 	int			 rs_num;
399 
400 	ruleset = pf_find_kruleset(anchor);
401 	if (ruleset == NULL)
402 		return (NULL);
403 	rs_num = pf_get_ruleset_number(rule_action);
404 	if (rs_num >= PF_RULESET_MAX)
405 		return (NULL);
406 	if (active) {
407 		if (check_ticket && ticket !=
408 		    ruleset->rules[rs_num].active.ticket)
409 			return (NULL);
410 		if (r_last)
411 			rule = TAILQ_LAST(ruleset->rules[rs_num].active.ptr,
412 			    pf_krulequeue);
413 		else
414 			rule = TAILQ_FIRST(ruleset->rules[rs_num].active.ptr);
415 	} else {
416 		if (check_ticket && ticket !=
417 		    ruleset->rules[rs_num].inactive.ticket)
418 			return (NULL);
419 		if (r_last)
420 			rule = TAILQ_LAST(ruleset->rules[rs_num].inactive.ptr,
421 			    pf_krulequeue);
422 		else
423 			rule = TAILQ_FIRST(ruleset->rules[rs_num].inactive.ptr);
424 	}
425 	if (!r_last) {
426 		while ((rule != NULL) && (rule->nr != rule_number))
427 			rule = TAILQ_NEXT(rule, entries);
428 	}
429 	if (rule == NULL)
430 		return (NULL);
431 
432 	return (&rule->rpool);
433 }
434 
435 static void
pf_mv_kpool(struct pf_kpalist * poola,struct pf_kpalist * poolb)436 pf_mv_kpool(struct pf_kpalist *poola, struct pf_kpalist *poolb)
437 {
438 	struct pf_kpooladdr	*mv_pool_pa;
439 
440 	while ((mv_pool_pa = TAILQ_FIRST(poola)) != NULL) {
441 		TAILQ_REMOVE(poola, mv_pool_pa, entries);
442 		TAILQ_INSERT_TAIL(poolb, mv_pool_pa, entries);
443 	}
444 }
445 
446 static void
pf_empty_kpool(struct pf_kpalist * poola)447 pf_empty_kpool(struct pf_kpalist *poola)
448 {
449 	struct pf_kpooladdr *pa;
450 
451 	while ((pa = TAILQ_FIRST(poola)) != NULL) {
452 		switch (pa->addr.type) {
453 		case PF_ADDR_DYNIFTL:
454 			pfi_dynaddr_remove(pa->addr.p.dyn);
455 			break;
456 		case PF_ADDR_TABLE:
457 			/* XXX: this could be unfinished pooladdr on pabuf */
458 			if (pa->addr.p.tbl != NULL)
459 				pfr_detach_table(pa->addr.p.tbl);
460 			break;
461 		}
462 		if (pa->kif)
463 			pfi_kkif_unref(pa->kif);
464 		TAILQ_REMOVE(poola, pa, entries);
465 		free(pa, M_PFRULE);
466 	}
467 }
468 
469 static void
pf_unlink_rule(struct pf_krulequeue * rulequeue,struct pf_krule * rule)470 pf_unlink_rule(struct pf_krulequeue *rulequeue, struct pf_krule *rule)
471 {
472 
473 	PF_RULES_WASSERT();
474 
475 	TAILQ_REMOVE(rulequeue, rule, entries);
476 
477 	PF_UNLNKDRULES_LOCK();
478 	rule->rule_ref |= PFRULE_REFS;
479 	TAILQ_INSERT_TAIL(&V_pf_unlinked_rules, rule, entries);
480 	PF_UNLNKDRULES_UNLOCK();
481 }
482 
483 void
pf_free_rule(struct pf_krule * rule)484 pf_free_rule(struct pf_krule *rule)
485 {
486 
487 	PF_RULES_WASSERT();
488 
489 	if (rule->tag)
490 		tag_unref(&V_pf_tags, rule->tag);
491 	if (rule->match_tag)
492 		tag_unref(&V_pf_tags, rule->match_tag);
493 #ifdef ALTQ
494 	if (rule->pqid != rule->qid)
495 		pf_qid_unref(rule->pqid);
496 	pf_qid_unref(rule->qid);
497 #endif
498 	switch (rule->src.addr.type) {
499 	case PF_ADDR_DYNIFTL:
500 		pfi_dynaddr_remove(rule->src.addr.p.dyn);
501 		break;
502 	case PF_ADDR_TABLE:
503 		pfr_detach_table(rule->src.addr.p.tbl);
504 		break;
505 	}
506 	switch (rule->dst.addr.type) {
507 	case PF_ADDR_DYNIFTL:
508 		pfi_dynaddr_remove(rule->dst.addr.p.dyn);
509 		break;
510 	case PF_ADDR_TABLE:
511 		pfr_detach_table(rule->dst.addr.p.tbl);
512 		break;
513 	}
514 	if (rule->overload_tbl)
515 		pfr_detach_table(rule->overload_tbl);
516 	if (rule->kif)
517 		pfi_kkif_unref(rule->kif);
518 	pf_kanchor_remove(rule);
519 	pf_empty_kpool(&rule->rpool.list);
520 
521 	pf_krule_free(rule);
522 }
523 
524 static void
pf_init_tagset(struct pf_tagset * ts,unsigned int * tunable_size,unsigned int default_size)525 pf_init_tagset(struct pf_tagset *ts, unsigned int *tunable_size,
526     unsigned int default_size)
527 {
528 	unsigned int i;
529 	unsigned int hashsize;
530 
531 	if (*tunable_size == 0 || !powerof2(*tunable_size))
532 		*tunable_size = default_size;
533 
534 	hashsize = *tunable_size;
535 	ts->namehash = mallocarray(hashsize, sizeof(*ts->namehash), M_PFHASH,
536 	    M_WAITOK);
537 	ts->taghash = mallocarray(hashsize, sizeof(*ts->taghash), M_PFHASH,
538 	    M_WAITOK);
539 	ts->mask = hashsize - 1;
540 	ts->seed = arc4random();
541 	for (i = 0; i < hashsize; i++) {
542 		TAILQ_INIT(&ts->namehash[i]);
543 		TAILQ_INIT(&ts->taghash[i]);
544 	}
545 	BIT_FILL(TAGID_MAX, &ts->avail);
546 }
547 
548 static void
pf_cleanup_tagset(struct pf_tagset * ts)549 pf_cleanup_tagset(struct pf_tagset *ts)
550 {
551 	unsigned int i;
552 	unsigned int hashsize;
553 	struct pf_tagname *t, *tmp;
554 
555 	/*
556 	 * Only need to clean up one of the hashes as each tag is hashed
557 	 * into each table.
558 	 */
559 	hashsize = ts->mask + 1;
560 	for (i = 0; i < hashsize; i++)
561 		TAILQ_FOREACH_SAFE(t, &ts->namehash[i], namehash_entries, tmp)
562 			uma_zfree(V_pf_tag_z, t);
563 
564 	free(ts->namehash, M_PFHASH);
565 	free(ts->taghash, M_PFHASH);
566 }
567 
568 static uint16_t
tagname2hashindex(const struct pf_tagset * ts,const char * tagname)569 tagname2hashindex(const struct pf_tagset *ts, const char *tagname)
570 {
571 	size_t len;
572 
573 	len = strnlen(tagname, PF_TAG_NAME_SIZE - 1);
574 	return (murmur3_32_hash(tagname, len, ts->seed) & ts->mask);
575 }
576 
577 static uint16_t
tag2hashindex(const struct pf_tagset * ts,uint16_t tag)578 tag2hashindex(const struct pf_tagset *ts, uint16_t tag)
579 {
580 
581 	return (tag & ts->mask);
582 }
583 
584 static u_int16_t
tagname2tag(struct pf_tagset * ts,const char * tagname)585 tagname2tag(struct pf_tagset *ts, const char *tagname)
586 {
587 	struct pf_tagname	*tag;
588 	u_int32_t		 index;
589 	u_int16_t		 new_tagid;
590 
591 	PF_RULES_WASSERT();
592 
593 	index = tagname2hashindex(ts, tagname);
594 	TAILQ_FOREACH(tag, &ts->namehash[index], namehash_entries)
595 		if (strcmp(tagname, tag->name) == 0) {
596 			tag->ref++;
597 			return (tag->tag);
598 		}
599 
600 	/*
601 	 * new entry
602 	 *
603 	 * to avoid fragmentation, we do a linear search from the beginning
604 	 * and take the first free slot we find.
605 	 */
606 	new_tagid = BIT_FFS(TAGID_MAX, &ts->avail);
607 	/*
608 	 * Tags are 1-based, with valid tags in the range [1..TAGID_MAX].
609 	 * BIT_FFS() returns a 1-based bit number, with 0 indicating no bits
610 	 * set.  It may also return a bit number greater than TAGID_MAX due
611 	 * to rounding of the number of bits in the vector up to a multiple
612 	 * of the vector word size at declaration/allocation time.
613 	 */
614 	if ((new_tagid == 0) || (new_tagid > TAGID_MAX))
615 		return (0);
616 
617 	/* Mark the tag as in use.  Bits are 0-based for BIT_CLR() */
618 	BIT_CLR(TAGID_MAX, new_tagid - 1, &ts->avail);
619 
620 	/* allocate and fill new struct pf_tagname */
621 	tag = uma_zalloc(V_pf_tag_z, M_NOWAIT);
622 	if (tag == NULL)
623 		return (0);
624 	strlcpy(tag->name, tagname, sizeof(tag->name));
625 	tag->tag = new_tagid;
626 	tag->ref = 1;
627 
628 	/* Insert into namehash */
629 	TAILQ_INSERT_TAIL(&ts->namehash[index], tag, namehash_entries);
630 
631 	/* Insert into taghash */
632 	index = tag2hashindex(ts, new_tagid);
633 	TAILQ_INSERT_TAIL(&ts->taghash[index], tag, taghash_entries);
634 
635 	return (tag->tag);
636 }
637 
638 static void
tag_unref(struct pf_tagset * ts,u_int16_t tag)639 tag_unref(struct pf_tagset *ts, u_int16_t tag)
640 {
641 	struct pf_tagname	*t;
642 	uint16_t		 index;
643 
644 	PF_RULES_WASSERT();
645 
646 	index = tag2hashindex(ts, tag);
647 	TAILQ_FOREACH(t, &ts->taghash[index], taghash_entries)
648 		if (tag == t->tag) {
649 			if (--t->ref == 0) {
650 				TAILQ_REMOVE(&ts->taghash[index], t,
651 				    taghash_entries);
652 				index = tagname2hashindex(ts, t->name);
653 				TAILQ_REMOVE(&ts->namehash[index], t,
654 				    namehash_entries);
655 				/* Bits are 0-based for BIT_SET() */
656 				BIT_SET(TAGID_MAX, tag - 1, &ts->avail);
657 				uma_zfree(V_pf_tag_z, t);
658 			}
659 			break;
660 		}
661 }
662 
663 static uint16_t
pf_tagname2tag(const char * tagname)664 pf_tagname2tag(const char *tagname)
665 {
666 	return (tagname2tag(&V_pf_tags, tagname));
667 }
668 
669 #ifdef ALTQ
670 static uint16_t
pf_qname2qid(const char * qname)671 pf_qname2qid(const char *qname)
672 {
673 	return (tagname2tag(&V_pf_qids, qname));
674 }
675 
676 static void
pf_qid_unref(uint16_t qid)677 pf_qid_unref(uint16_t qid)
678 {
679 	tag_unref(&V_pf_qids, qid);
680 }
681 
682 static int
pf_begin_altq(u_int32_t * ticket)683 pf_begin_altq(u_int32_t *ticket)
684 {
685 	struct pf_altq	*altq, *tmp;
686 	int		 error = 0;
687 
688 	PF_RULES_WASSERT();
689 
690 	/* Purge the old altq lists */
691 	TAILQ_FOREACH_SAFE(altq, V_pf_altq_ifs_inactive, entries, tmp) {
692 		if ((altq->local_flags & PFALTQ_FLAG_IF_REMOVED) == 0) {
693 			/* detach and destroy the discipline */
694 			error = altq_remove(altq);
695 		}
696 		free(altq, M_PFALTQ);
697 	}
698 	TAILQ_INIT(V_pf_altq_ifs_inactive);
699 	TAILQ_FOREACH_SAFE(altq, V_pf_altqs_inactive, entries, tmp) {
700 		pf_qid_unref(altq->qid);
701 		free(altq, M_PFALTQ);
702 	}
703 	TAILQ_INIT(V_pf_altqs_inactive);
704 	if (error)
705 		return (error);
706 	*ticket = ++V_ticket_altqs_inactive;
707 	V_altqs_inactive_open = 1;
708 	return (0);
709 }
710 
711 static int
pf_rollback_altq(u_int32_t ticket)712 pf_rollback_altq(u_int32_t ticket)
713 {
714 	struct pf_altq	*altq, *tmp;
715 	int		 error = 0;
716 
717 	PF_RULES_WASSERT();
718 
719 	if (!V_altqs_inactive_open || ticket != V_ticket_altqs_inactive)
720 		return (0);
721 	/* Purge the old altq lists */
722 	TAILQ_FOREACH_SAFE(altq, V_pf_altq_ifs_inactive, entries, tmp) {
723 		if ((altq->local_flags & PFALTQ_FLAG_IF_REMOVED) == 0) {
724 			/* detach and destroy the discipline */
725 			error = altq_remove(altq);
726 		}
727 		free(altq, M_PFALTQ);
728 	}
729 	TAILQ_INIT(V_pf_altq_ifs_inactive);
730 	TAILQ_FOREACH_SAFE(altq, V_pf_altqs_inactive, entries, tmp) {
731 		pf_qid_unref(altq->qid);
732 		free(altq, M_PFALTQ);
733 	}
734 	TAILQ_INIT(V_pf_altqs_inactive);
735 	V_altqs_inactive_open = 0;
736 	return (error);
737 }
738 
739 static int
pf_commit_altq(u_int32_t ticket)740 pf_commit_altq(u_int32_t ticket)
741 {
742 	struct pf_altqqueue	*old_altqs, *old_altq_ifs;
743 	struct pf_altq		*altq, *tmp;
744 	int			 err, error = 0;
745 
746 	PF_RULES_WASSERT();
747 
748 	if (!V_altqs_inactive_open || ticket != V_ticket_altqs_inactive)
749 		return (EBUSY);
750 
751 	/* swap altqs, keep the old. */
752 	old_altqs = V_pf_altqs_active;
753 	old_altq_ifs = V_pf_altq_ifs_active;
754 	V_pf_altqs_active = V_pf_altqs_inactive;
755 	V_pf_altq_ifs_active = V_pf_altq_ifs_inactive;
756 	V_pf_altqs_inactive = old_altqs;
757 	V_pf_altq_ifs_inactive = old_altq_ifs;
758 	V_ticket_altqs_active = V_ticket_altqs_inactive;
759 
760 	/* Attach new disciplines */
761 	TAILQ_FOREACH(altq, V_pf_altq_ifs_active, entries) {
762 		if ((altq->local_flags & PFALTQ_FLAG_IF_REMOVED) == 0) {
763 			/* attach the discipline */
764 			error = altq_pfattach(altq);
765 			if (error == 0 && V_pf_altq_running)
766 				error = pf_enable_altq(altq);
767 			if (error != 0)
768 				return (error);
769 		}
770 	}
771 
772 	/* Purge the old altq lists */
773 	TAILQ_FOREACH_SAFE(altq, V_pf_altq_ifs_inactive, entries, tmp) {
774 		if ((altq->local_flags & PFALTQ_FLAG_IF_REMOVED) == 0) {
775 			/* detach and destroy the discipline */
776 			if (V_pf_altq_running)
777 				error = pf_disable_altq(altq);
778 			err = altq_pfdetach(altq);
779 			if (err != 0 && error == 0)
780 				error = err;
781 			err = altq_remove(altq);
782 			if (err != 0 && error == 0)
783 				error = err;
784 		}
785 		free(altq, M_PFALTQ);
786 	}
787 	TAILQ_INIT(V_pf_altq_ifs_inactive);
788 	TAILQ_FOREACH_SAFE(altq, V_pf_altqs_inactive, entries, tmp) {
789 		pf_qid_unref(altq->qid);
790 		free(altq, M_PFALTQ);
791 	}
792 	TAILQ_INIT(V_pf_altqs_inactive);
793 
794 	V_altqs_inactive_open = 0;
795 	return (error);
796 }
797 
798 static int
pf_enable_altq(struct pf_altq * altq)799 pf_enable_altq(struct pf_altq *altq)
800 {
801 	struct ifnet		*ifp;
802 	struct tb_profile	 tb;
803 	int			 error = 0;
804 
805 	if ((ifp = ifunit(altq->ifname)) == NULL)
806 		return (EINVAL);
807 
808 	if (ifp->if_snd.altq_type != ALTQT_NONE)
809 		error = altq_enable(&ifp->if_snd);
810 
811 	/* set tokenbucket regulator */
812 	if (error == 0 && ifp != NULL && ALTQ_IS_ENABLED(&ifp->if_snd)) {
813 		tb.rate = altq->ifbandwidth;
814 		tb.depth = altq->tbrsize;
815 		error = tbr_set(&ifp->if_snd, &tb);
816 	}
817 
818 	return (error);
819 }
820 
821 static int
pf_disable_altq(struct pf_altq * altq)822 pf_disable_altq(struct pf_altq *altq)
823 {
824 	struct ifnet		*ifp;
825 	struct tb_profile	 tb;
826 	int			 error;
827 
828 	if ((ifp = ifunit(altq->ifname)) == NULL)
829 		return (EINVAL);
830 
831 	/*
832 	 * when the discipline is no longer referenced, it was overridden
833 	 * by a new one.  if so, just return.
834 	 */
835 	if (altq->altq_disc != ifp->if_snd.altq_disc)
836 		return (0);
837 
838 	error = altq_disable(&ifp->if_snd);
839 
840 	if (error == 0) {
841 		/* clear tokenbucket regulator */
842 		tb.rate = 0;
843 		error = tbr_set(&ifp->if_snd, &tb);
844 	}
845 
846 	return (error);
847 }
848 
849 static int
pf_altq_ifnet_event_add(struct ifnet * ifp,int remove,u_int32_t ticket,struct pf_altq * altq)850 pf_altq_ifnet_event_add(struct ifnet *ifp, int remove, u_int32_t ticket,
851     struct pf_altq *altq)
852 {
853 	struct ifnet	*ifp1;
854 	int		 error = 0;
855 
856 	/* Deactivate the interface in question */
857 	altq->local_flags &= ~PFALTQ_FLAG_IF_REMOVED;
858 	if ((ifp1 = ifunit(altq->ifname)) == NULL ||
859 	    (remove && ifp1 == ifp)) {
860 		altq->local_flags |= PFALTQ_FLAG_IF_REMOVED;
861 	} else {
862 		error = altq_add(ifp1, altq);
863 
864 		if (ticket != V_ticket_altqs_inactive)
865 			error = EBUSY;
866 
867 		if (error)
868 			free(altq, M_PFALTQ);
869 	}
870 
871 	return (error);
872 }
873 
874 void
pf_altq_ifnet_event(struct ifnet * ifp,int remove)875 pf_altq_ifnet_event(struct ifnet *ifp, int remove)
876 {
877 	struct pf_altq	*a1, *a2, *a3;
878 	u_int32_t	 ticket;
879 	int		 error = 0;
880 
881 	/*
882 	 * No need to re-evaluate the configuration for events on interfaces
883 	 * that do not support ALTQ, as it's not possible for such
884 	 * interfaces to be part of the configuration.
885 	 */
886 	if (!ALTQ_IS_READY(&ifp->if_snd))
887 		return;
888 
889 	/* Interrupt userland queue modifications */
890 	if (V_altqs_inactive_open)
891 		pf_rollback_altq(V_ticket_altqs_inactive);
892 
893 	/* Start new altq ruleset */
894 	if (pf_begin_altq(&ticket))
895 		return;
896 
897 	/* Copy the current active set */
898 	TAILQ_FOREACH(a1, V_pf_altq_ifs_active, entries) {
899 		a2 = malloc(sizeof(*a2), M_PFALTQ, M_NOWAIT);
900 		if (a2 == NULL) {
901 			error = ENOMEM;
902 			break;
903 		}
904 		bcopy(a1, a2, sizeof(struct pf_altq));
905 
906 		error = pf_altq_ifnet_event_add(ifp, remove, ticket, a2);
907 		if (error)
908 			break;
909 
910 		TAILQ_INSERT_TAIL(V_pf_altq_ifs_inactive, a2, entries);
911 	}
912 	if (error)
913 		goto out;
914 	TAILQ_FOREACH(a1, V_pf_altqs_active, entries) {
915 		a2 = malloc(sizeof(*a2), M_PFALTQ, M_NOWAIT);
916 		if (a2 == NULL) {
917 			error = ENOMEM;
918 			break;
919 		}
920 		bcopy(a1, a2, sizeof(struct pf_altq));
921 
922 		if ((a2->qid = pf_qname2qid(a2->qname)) == 0) {
923 			error = EBUSY;
924 			free(a2, M_PFALTQ);
925 			break;
926 		}
927 		a2->altq_disc = NULL;
928 		TAILQ_FOREACH(a3, V_pf_altq_ifs_inactive, entries) {
929 			if (strncmp(a3->ifname, a2->ifname,
930 				IFNAMSIZ) == 0) {
931 				a2->altq_disc = a3->altq_disc;
932 				break;
933 			}
934 		}
935 		error = pf_altq_ifnet_event_add(ifp, remove, ticket, a2);
936 		if (error)
937 			break;
938 
939 		TAILQ_INSERT_TAIL(V_pf_altqs_inactive, a2, entries);
940 	}
941 
942 out:
943 	if (error != 0)
944 		pf_rollback_altq(ticket);
945 	else
946 		pf_commit_altq(ticket);
947 }
948 #endif /* ALTQ */
949 
950 static int
pf_begin_rules(u_int32_t * ticket,int rs_num,const char * anchor)951 pf_begin_rules(u_int32_t *ticket, int rs_num, const char *anchor)
952 {
953 	struct pf_kruleset	*rs;
954 	struct pf_krule		*rule;
955 
956 	PF_RULES_WASSERT();
957 
958 	if (rs_num < 0 || rs_num >= PF_RULESET_MAX)
959 		return (EINVAL);
960 	rs = pf_find_or_create_kruleset(anchor);
961 	if (rs == NULL)
962 		return (EINVAL);
963 	while ((rule = TAILQ_FIRST(rs->rules[rs_num].inactive.ptr)) != NULL) {
964 		pf_unlink_rule(rs->rules[rs_num].inactive.ptr, rule);
965 		rs->rules[rs_num].inactive.rcount--;
966 	}
967 	*ticket = ++rs->rules[rs_num].inactive.ticket;
968 	rs->rules[rs_num].inactive.open = 1;
969 	return (0);
970 }
971 
972 static int
pf_rollback_rules(u_int32_t ticket,int rs_num,char * anchor)973 pf_rollback_rules(u_int32_t ticket, int rs_num, char *anchor)
974 {
975 	struct pf_kruleset	*rs;
976 	struct pf_krule		*rule;
977 
978 	PF_RULES_WASSERT();
979 
980 	if (rs_num < 0 || rs_num >= PF_RULESET_MAX)
981 		return (EINVAL);
982 	rs = pf_find_kruleset(anchor);
983 	if (rs == NULL || !rs->rules[rs_num].inactive.open ||
984 	    rs->rules[rs_num].inactive.ticket != ticket)
985 		return (0);
986 	while ((rule = TAILQ_FIRST(rs->rules[rs_num].inactive.ptr)) != NULL) {
987 		pf_unlink_rule(rs->rules[rs_num].inactive.ptr, rule);
988 		rs->rules[rs_num].inactive.rcount--;
989 	}
990 	rs->rules[rs_num].inactive.open = 0;
991 	return (0);
992 }
993 
994 #define PF_MD5_UPD(st, elm)						\
995 		MD5Update(ctx, (u_int8_t *) &(st)->elm, sizeof((st)->elm))
996 
997 #define PF_MD5_UPD_STR(st, elm)						\
998 		MD5Update(ctx, (u_int8_t *) (st)->elm, strlen((st)->elm))
999 
1000 #define PF_MD5_UPD_HTONL(st, elm, stor) do {				\
1001 		(stor) = htonl((st)->elm);				\
1002 		MD5Update(ctx, (u_int8_t *) &(stor), sizeof(u_int32_t));\
1003 } while (0)
1004 
1005 #define PF_MD5_UPD_HTONS(st, elm, stor) do {				\
1006 		(stor) = htons((st)->elm);				\
1007 		MD5Update(ctx, (u_int8_t *) &(stor), sizeof(u_int16_t));\
1008 } while (0)
1009 
1010 static void
pf_hash_rule_addr(MD5_CTX * ctx,struct pf_rule_addr * pfr)1011 pf_hash_rule_addr(MD5_CTX *ctx, struct pf_rule_addr *pfr)
1012 {
1013 	PF_MD5_UPD(pfr, addr.type);
1014 	switch (pfr->addr.type) {
1015 		case PF_ADDR_DYNIFTL:
1016 			PF_MD5_UPD(pfr, addr.v.ifname);
1017 			PF_MD5_UPD(pfr, addr.iflags);
1018 			break;
1019 		case PF_ADDR_TABLE:
1020 			PF_MD5_UPD(pfr, addr.v.tblname);
1021 			break;
1022 		case PF_ADDR_ADDRMASK:
1023 			/* XXX ignore af? */
1024 			PF_MD5_UPD(pfr, addr.v.a.addr.addr32);
1025 			PF_MD5_UPD(pfr, addr.v.a.mask.addr32);
1026 			break;
1027 	}
1028 
1029 	PF_MD5_UPD(pfr, port[0]);
1030 	PF_MD5_UPD(pfr, port[1]);
1031 	PF_MD5_UPD(pfr, neg);
1032 	PF_MD5_UPD(pfr, port_op);
1033 }
1034 
1035 static void
pf_hash_rule(MD5_CTX * ctx,struct pf_krule * rule)1036 pf_hash_rule(MD5_CTX *ctx, struct pf_krule *rule)
1037 {
1038 	u_int16_t x;
1039 	u_int32_t y;
1040 
1041 	pf_hash_rule_addr(ctx, &rule->src);
1042 	pf_hash_rule_addr(ctx, &rule->dst);
1043 	for (int i = 0; i < PF_RULE_MAX_LABEL_COUNT; i++)
1044 		PF_MD5_UPD_STR(rule, label[i]);
1045 	PF_MD5_UPD_STR(rule, ifname);
1046 	PF_MD5_UPD_STR(rule, match_tagname);
1047 	PF_MD5_UPD_HTONS(rule, match_tag, x); /* dup? */
1048 	PF_MD5_UPD_HTONL(rule, os_fingerprint, y);
1049 	PF_MD5_UPD_HTONL(rule, prob, y);
1050 	PF_MD5_UPD_HTONL(rule, uid.uid[0], y);
1051 	PF_MD5_UPD_HTONL(rule, uid.uid[1], y);
1052 	PF_MD5_UPD(rule, uid.op);
1053 	PF_MD5_UPD_HTONL(rule, gid.gid[0], y);
1054 	PF_MD5_UPD_HTONL(rule, gid.gid[1], y);
1055 	PF_MD5_UPD(rule, gid.op);
1056 	PF_MD5_UPD_HTONL(rule, rule_flag, y);
1057 	PF_MD5_UPD(rule, action);
1058 	PF_MD5_UPD(rule, direction);
1059 	PF_MD5_UPD(rule, af);
1060 	PF_MD5_UPD(rule, quick);
1061 	PF_MD5_UPD(rule, ifnot);
1062 	PF_MD5_UPD(rule, match_tag_not);
1063 	PF_MD5_UPD(rule, natpass);
1064 	PF_MD5_UPD(rule, keep_state);
1065 	PF_MD5_UPD(rule, proto);
1066 	PF_MD5_UPD(rule, type);
1067 	PF_MD5_UPD(rule, code);
1068 	PF_MD5_UPD(rule, flags);
1069 	PF_MD5_UPD(rule, flagset);
1070 	PF_MD5_UPD(rule, allow_opts);
1071 	PF_MD5_UPD(rule, rt);
1072 	PF_MD5_UPD(rule, tos);
1073 }
1074 
1075 static bool
pf_krule_compare(struct pf_krule * a,struct pf_krule * b)1076 pf_krule_compare(struct pf_krule *a, struct pf_krule *b)
1077 {
1078 	MD5_CTX		ctx[2];
1079 	u_int8_t	digest[2][PF_MD5_DIGEST_LENGTH];
1080 
1081 	MD5Init(&ctx[0]);
1082 	MD5Init(&ctx[1]);
1083 	pf_hash_rule(&ctx[0], a);
1084 	pf_hash_rule(&ctx[1], b);
1085 	MD5Final(digest[0], &ctx[0]);
1086 	MD5Final(digest[1], &ctx[1]);
1087 
1088 	return (memcmp(digest[0], digest[1], PF_MD5_DIGEST_LENGTH) == 0);
1089 }
1090 
1091 static int
pf_commit_rules(u_int32_t ticket,int rs_num,char * anchor)1092 pf_commit_rules(u_int32_t ticket, int rs_num, char *anchor)
1093 {
1094 	struct pf_kruleset	*rs;
1095 	struct pf_krule		*rule, **old_array, *tail;
1096 	struct pf_krulequeue	*old_rules;
1097 	int			 error;
1098 	u_int32_t		 old_rcount;
1099 
1100 	PF_RULES_WASSERT();
1101 
1102 	if (rs_num < 0 || rs_num >= PF_RULESET_MAX)
1103 		return (EINVAL);
1104 	rs = pf_find_kruleset(anchor);
1105 	if (rs == NULL || !rs->rules[rs_num].inactive.open ||
1106 	    ticket != rs->rules[rs_num].inactive.ticket)
1107 		return (EBUSY);
1108 
1109 	/* Calculate checksum for the main ruleset */
1110 	if (rs == &pf_main_ruleset) {
1111 		error = pf_setup_pfsync_matching(rs);
1112 		if (error != 0)
1113 			return (error);
1114 	}
1115 
1116 	/* Swap rules, keep the old. */
1117 	old_rules = rs->rules[rs_num].active.ptr;
1118 	old_rcount = rs->rules[rs_num].active.rcount;
1119 	old_array = rs->rules[rs_num].active.ptr_array;
1120 
1121 	rs->rules[rs_num].active.ptr =
1122 	    rs->rules[rs_num].inactive.ptr;
1123 	rs->rules[rs_num].active.ptr_array =
1124 	    rs->rules[rs_num].inactive.ptr_array;
1125 	rs->rules[rs_num].active.rcount =
1126 	    rs->rules[rs_num].inactive.rcount;
1127 
1128 	/* Attempt to preserve counter information. */
1129 	if (V_pf_status.keep_counters) {
1130 		TAILQ_FOREACH(rule, rs->rules[rs_num].active.ptr,
1131 		    entries) {
1132 			tail = TAILQ_FIRST(old_rules);
1133 			while ((tail != NULL) && ! pf_krule_compare(tail, rule))
1134 				tail = TAILQ_NEXT(tail, entries);
1135 			if (tail != NULL) {
1136 				pf_counter_u64_critical_enter();
1137 				pf_counter_u64_add_protected(&rule->evaluations,
1138 				    pf_counter_u64_fetch(&tail->evaluations));
1139 				pf_counter_u64_add_protected(&rule->packets[0],
1140 				    pf_counter_u64_fetch(&tail->packets[0]));
1141 				pf_counter_u64_add_protected(&rule->packets[1],
1142 				    pf_counter_u64_fetch(&tail->packets[1]));
1143 				pf_counter_u64_add_protected(&rule->bytes[0],
1144 				    pf_counter_u64_fetch(&tail->bytes[0]));
1145 				pf_counter_u64_add_protected(&rule->bytes[1],
1146 				    pf_counter_u64_fetch(&tail->bytes[1]));
1147 				pf_counter_u64_critical_exit();
1148 			}
1149 		}
1150 	}
1151 
1152 	rs->rules[rs_num].inactive.ptr = old_rules;
1153 	rs->rules[rs_num].inactive.ptr_array = old_array;
1154 	rs->rules[rs_num].inactive.rcount = old_rcount;
1155 
1156 	rs->rules[rs_num].active.ticket =
1157 	    rs->rules[rs_num].inactive.ticket;
1158 	pf_calc_skip_steps(rs->rules[rs_num].active.ptr);
1159 
1160 	/* Purge the old rule list. */
1161 	while ((rule = TAILQ_FIRST(old_rules)) != NULL)
1162 		pf_unlink_rule(old_rules, rule);
1163 	if (rs->rules[rs_num].inactive.ptr_array)
1164 		free(rs->rules[rs_num].inactive.ptr_array, M_TEMP);
1165 	rs->rules[rs_num].inactive.ptr_array = NULL;
1166 	rs->rules[rs_num].inactive.rcount = 0;
1167 	rs->rules[rs_num].inactive.open = 0;
1168 	pf_remove_if_empty_kruleset(rs);
1169 
1170 	return (0);
1171 }
1172 
1173 static int
pf_setup_pfsync_matching(struct pf_kruleset * rs)1174 pf_setup_pfsync_matching(struct pf_kruleset *rs)
1175 {
1176 	MD5_CTX			 ctx;
1177 	struct pf_krule		*rule;
1178 	int			 rs_cnt;
1179 	u_int8_t		 digest[PF_MD5_DIGEST_LENGTH];
1180 
1181 	MD5Init(&ctx);
1182 	for (rs_cnt = 0; rs_cnt < PF_RULESET_MAX; rs_cnt++) {
1183 		/* XXX PF_RULESET_SCRUB as well? */
1184 		if (rs_cnt == PF_RULESET_SCRUB)
1185 			continue;
1186 
1187 		if (rs->rules[rs_cnt].inactive.ptr_array)
1188 			free(rs->rules[rs_cnt].inactive.ptr_array, M_TEMP);
1189 		rs->rules[rs_cnt].inactive.ptr_array = NULL;
1190 
1191 		if (rs->rules[rs_cnt].inactive.rcount) {
1192 			rs->rules[rs_cnt].inactive.ptr_array =
1193 			    malloc(sizeof(caddr_t) *
1194 			    rs->rules[rs_cnt].inactive.rcount,
1195 			    M_TEMP, M_NOWAIT);
1196 
1197 			if (!rs->rules[rs_cnt].inactive.ptr_array)
1198 				return (ENOMEM);
1199 		}
1200 
1201 		TAILQ_FOREACH(rule, rs->rules[rs_cnt].inactive.ptr,
1202 		    entries) {
1203 			pf_hash_rule(&ctx, rule);
1204 			(rs->rules[rs_cnt].inactive.ptr_array)[rule->nr] = rule;
1205 		}
1206 	}
1207 
1208 	MD5Final(digest, &ctx);
1209 	memcpy(V_pf_status.pf_chksum, digest, sizeof(V_pf_status.pf_chksum));
1210 	return (0);
1211 }
1212 
1213 static int
pf_addr_setup(struct pf_kruleset * ruleset,struct pf_addr_wrap * addr,sa_family_t af)1214 pf_addr_setup(struct pf_kruleset *ruleset, struct pf_addr_wrap *addr,
1215     sa_family_t af)
1216 {
1217 	int error = 0;
1218 
1219 	switch (addr->type) {
1220 	case PF_ADDR_TABLE:
1221 		addr->p.tbl = pfr_attach_table(ruleset, addr->v.tblname);
1222 		if (addr->p.tbl == NULL)
1223 			error = ENOMEM;
1224 		break;
1225 	case PF_ADDR_DYNIFTL:
1226 		error = pfi_dynaddr_setup(addr, af);
1227 		break;
1228 	}
1229 
1230 	return (error);
1231 }
1232 
1233 static void
pf_addr_copyout(struct pf_addr_wrap * addr)1234 pf_addr_copyout(struct pf_addr_wrap *addr)
1235 {
1236 
1237 	switch (addr->type) {
1238 	case PF_ADDR_DYNIFTL:
1239 		pfi_dynaddr_copyout(addr);
1240 		break;
1241 	case PF_ADDR_TABLE:
1242 		pf_tbladdr_copyout(addr);
1243 		break;
1244 	}
1245 }
1246 
1247 static void
pf_src_node_copy(const struct pf_ksrc_node * in,struct pf_src_node * out)1248 pf_src_node_copy(const struct pf_ksrc_node *in, struct pf_src_node *out)
1249 {
1250 	int	secs = time_uptime, diff;
1251 
1252 	bzero(out, sizeof(struct pf_src_node));
1253 
1254 	bcopy(&in->addr, &out->addr, sizeof(struct pf_addr));
1255 	bcopy(&in->raddr, &out->raddr, sizeof(struct pf_addr));
1256 
1257 	if (in->rule.ptr != NULL)
1258 		out->rule.nr = in->rule.ptr->nr;
1259 
1260 	for (int i = 0; i < 2; i++) {
1261 		out->bytes[i] = counter_u64_fetch(in->bytes[i]);
1262 		out->packets[i] = counter_u64_fetch(in->packets[i]);
1263 	}
1264 
1265 	out->states = in->states;
1266 	out->conn = in->conn;
1267 	out->af = in->af;
1268 	out->ruletype = in->ruletype;
1269 
1270 	out->creation = secs - in->creation;
1271 	if (out->expire > secs)
1272 		out->expire -= secs;
1273 	else
1274 		out->expire = 0;
1275 
1276 	/* Adjust the connection rate estimate. */
1277 	diff = secs - in->conn_rate.last;
1278 	if (diff >= in->conn_rate.seconds)
1279 		out->conn_rate.count = 0;
1280 	else
1281 		out->conn_rate.count -=
1282 		    in->conn_rate.count * diff /
1283 		    in->conn_rate.seconds;
1284 }
1285 
1286 #ifdef ALTQ
1287 /*
1288  * Handle export of struct pf_kaltq to user binaries that may be using any
1289  * version of struct pf_altq.
1290  */
1291 static int
pf_export_kaltq(struct pf_altq * q,struct pfioc_altq_v1 * pa,size_t ioc_size)1292 pf_export_kaltq(struct pf_altq *q, struct pfioc_altq_v1 *pa, size_t ioc_size)
1293 {
1294 	u_int32_t version;
1295 
1296 	if (ioc_size == sizeof(struct pfioc_altq_v0))
1297 		version = 0;
1298 	else
1299 		version = pa->version;
1300 
1301 	if (version > PFIOC_ALTQ_VERSION)
1302 		return (EINVAL);
1303 
1304 #define ASSIGN(x) exported_q->x = q->x
1305 #define COPY(x) \
1306 	bcopy(&q->x, &exported_q->x, min(sizeof(q->x), sizeof(exported_q->x)))
1307 #define SATU16(x) (u_int32_t)uqmin((x), USHRT_MAX)
1308 #define SATU32(x) (u_int32_t)uqmin((x), UINT_MAX)
1309 
1310 	switch (version) {
1311 	case 0: {
1312 		struct pf_altq_v0 *exported_q =
1313 		    &((struct pfioc_altq_v0 *)pa)->altq;
1314 
1315 		COPY(ifname);
1316 
1317 		ASSIGN(scheduler);
1318 		ASSIGN(tbrsize);
1319 		exported_q->tbrsize = SATU16(q->tbrsize);
1320 		exported_q->ifbandwidth = SATU32(q->ifbandwidth);
1321 
1322 		COPY(qname);
1323 		COPY(parent);
1324 		ASSIGN(parent_qid);
1325 		exported_q->bandwidth = SATU32(q->bandwidth);
1326 		ASSIGN(priority);
1327 		ASSIGN(local_flags);
1328 
1329 		ASSIGN(qlimit);
1330 		ASSIGN(flags);
1331 
1332 		if (q->scheduler == ALTQT_HFSC) {
1333 #define ASSIGN_OPT(x) exported_q->pq_u.hfsc_opts.x = q->pq_u.hfsc_opts.x
1334 #define ASSIGN_OPT_SATU32(x) exported_q->pq_u.hfsc_opts.x = \
1335 			    SATU32(q->pq_u.hfsc_opts.x)
1336 
1337 			ASSIGN_OPT_SATU32(rtsc_m1);
1338 			ASSIGN_OPT(rtsc_d);
1339 			ASSIGN_OPT_SATU32(rtsc_m2);
1340 
1341 			ASSIGN_OPT_SATU32(lssc_m1);
1342 			ASSIGN_OPT(lssc_d);
1343 			ASSIGN_OPT_SATU32(lssc_m2);
1344 
1345 			ASSIGN_OPT_SATU32(ulsc_m1);
1346 			ASSIGN_OPT(ulsc_d);
1347 			ASSIGN_OPT_SATU32(ulsc_m2);
1348 
1349 			ASSIGN_OPT(flags);
1350 
1351 #undef ASSIGN_OPT
1352 #undef ASSIGN_OPT_SATU32
1353 		} else
1354 			COPY(pq_u);
1355 
1356 		ASSIGN(qid);
1357 		break;
1358 	}
1359 	case 1:	{
1360 		struct pf_altq_v1 *exported_q =
1361 		    &((struct pfioc_altq_v1 *)pa)->altq;
1362 
1363 		COPY(ifname);
1364 
1365 		ASSIGN(scheduler);
1366 		ASSIGN(tbrsize);
1367 		ASSIGN(ifbandwidth);
1368 
1369 		COPY(qname);
1370 		COPY(parent);
1371 		ASSIGN(parent_qid);
1372 		ASSIGN(bandwidth);
1373 		ASSIGN(priority);
1374 		ASSIGN(local_flags);
1375 
1376 		ASSIGN(qlimit);
1377 		ASSIGN(flags);
1378 		COPY(pq_u);
1379 
1380 		ASSIGN(qid);
1381 		break;
1382 	}
1383 	default:
1384 		panic("%s: unhandled struct pfioc_altq version", __func__);
1385 		break;
1386 	}
1387 
1388 #undef ASSIGN
1389 #undef COPY
1390 #undef SATU16
1391 #undef SATU32
1392 
1393 	return (0);
1394 }
1395 
1396 /*
1397  * Handle import to struct pf_kaltq of struct pf_altq from user binaries
1398  * that may be using any version of it.
1399  */
1400 static int
pf_import_kaltq(struct pfioc_altq_v1 * pa,struct pf_altq * q,size_t ioc_size)1401 pf_import_kaltq(struct pfioc_altq_v1 *pa, struct pf_altq *q, size_t ioc_size)
1402 {
1403 	u_int32_t version;
1404 
1405 	if (ioc_size == sizeof(struct pfioc_altq_v0))
1406 		version = 0;
1407 	else
1408 		version = pa->version;
1409 
1410 	if (version > PFIOC_ALTQ_VERSION)
1411 		return (EINVAL);
1412 
1413 #define ASSIGN(x) q->x = imported_q->x
1414 #define COPY(x) \
1415 	bcopy(&imported_q->x, &q->x, min(sizeof(imported_q->x), sizeof(q->x)))
1416 
1417 	switch (version) {
1418 	case 0: {
1419 		struct pf_altq_v0 *imported_q =
1420 		    &((struct pfioc_altq_v0 *)pa)->altq;
1421 
1422 		COPY(ifname);
1423 
1424 		ASSIGN(scheduler);
1425 		ASSIGN(tbrsize); /* 16-bit -> 32-bit */
1426 		ASSIGN(ifbandwidth); /* 32-bit -> 64-bit */
1427 
1428 		COPY(qname);
1429 		COPY(parent);
1430 		ASSIGN(parent_qid);
1431 		ASSIGN(bandwidth); /* 32-bit -> 64-bit */
1432 		ASSIGN(priority);
1433 		ASSIGN(local_flags);
1434 
1435 		ASSIGN(qlimit);
1436 		ASSIGN(flags);
1437 
1438 		if (imported_q->scheduler == ALTQT_HFSC) {
1439 #define ASSIGN_OPT(x) q->pq_u.hfsc_opts.x = imported_q->pq_u.hfsc_opts.x
1440 
1441 			/*
1442 			 * The m1 and m2 parameters are being copied from
1443 			 * 32-bit to 64-bit.
1444 			 */
1445 			ASSIGN_OPT(rtsc_m1);
1446 			ASSIGN_OPT(rtsc_d);
1447 			ASSIGN_OPT(rtsc_m2);
1448 
1449 			ASSIGN_OPT(lssc_m1);
1450 			ASSIGN_OPT(lssc_d);
1451 			ASSIGN_OPT(lssc_m2);
1452 
1453 			ASSIGN_OPT(ulsc_m1);
1454 			ASSIGN_OPT(ulsc_d);
1455 			ASSIGN_OPT(ulsc_m2);
1456 
1457 			ASSIGN_OPT(flags);
1458 
1459 #undef ASSIGN_OPT
1460 		} else
1461 			COPY(pq_u);
1462 
1463 		ASSIGN(qid);
1464 		break;
1465 	}
1466 	case 1: {
1467 		struct pf_altq_v1 *imported_q =
1468 		    &((struct pfioc_altq_v1 *)pa)->altq;
1469 
1470 		COPY(ifname);
1471 
1472 		ASSIGN(scheduler);
1473 		ASSIGN(tbrsize);
1474 		ASSIGN(ifbandwidth);
1475 
1476 		COPY(qname);
1477 		COPY(parent);
1478 		ASSIGN(parent_qid);
1479 		ASSIGN(bandwidth);
1480 		ASSIGN(priority);
1481 		ASSIGN(local_flags);
1482 
1483 		ASSIGN(qlimit);
1484 		ASSIGN(flags);
1485 		COPY(pq_u);
1486 
1487 		ASSIGN(qid);
1488 		break;
1489 	}
1490 	default:
1491 		panic("%s: unhandled struct pfioc_altq version", __func__);
1492 		break;
1493 	}
1494 
1495 #undef ASSIGN
1496 #undef COPY
1497 
1498 	return (0);
1499 }
1500 
1501 static struct pf_altq *
pf_altq_get_nth_active(u_int32_t n)1502 pf_altq_get_nth_active(u_int32_t n)
1503 {
1504 	struct pf_altq		*altq;
1505 	u_int32_t		 nr;
1506 
1507 	nr = 0;
1508 	TAILQ_FOREACH(altq, V_pf_altq_ifs_active, entries) {
1509 		if (nr == n)
1510 			return (altq);
1511 		nr++;
1512 	}
1513 
1514 	TAILQ_FOREACH(altq, V_pf_altqs_active, entries) {
1515 		if (nr == n)
1516 			return (altq);
1517 		nr++;
1518 	}
1519 
1520 	return (NULL);
1521 }
1522 #endif /* ALTQ */
1523 
1524 struct pf_krule *
pf_krule_alloc(void)1525 pf_krule_alloc(void)
1526 {
1527 	struct pf_krule *rule;
1528 
1529 	rule = malloc(sizeof(struct pf_krule), M_PFRULE, M_WAITOK | M_ZERO);
1530 	mtx_init(&rule->rpool.mtx, "pf_krule_pool", NULL, MTX_DEF);
1531 	return (rule);
1532 }
1533 
1534 void
pf_krule_free(struct pf_krule * rule)1535 pf_krule_free(struct pf_krule *rule)
1536 {
1537 #ifdef PF_WANT_32_TO_64_COUNTER
1538 	bool wowned;
1539 #endif
1540 
1541 	if (rule == NULL)
1542 		return;
1543 
1544 #ifdef PF_WANT_32_TO_64_COUNTER
1545 	if (rule->allrulelinked) {
1546 		wowned = PF_RULES_WOWNED();
1547 		if (!wowned)
1548 			PF_RULES_WLOCK();
1549 		LIST_REMOVE(rule, allrulelist);
1550 		V_pf_allrulecount--;
1551 		if (!wowned)
1552 			PF_RULES_WUNLOCK();
1553 	}
1554 #endif
1555 
1556 	pf_counter_u64_deinit(&rule->evaluations);
1557 	for (int i = 0; i < 2; i++) {
1558 		pf_counter_u64_deinit(&rule->packets[i]);
1559 		pf_counter_u64_deinit(&rule->bytes[i]);
1560 	}
1561 	counter_u64_free(rule->states_cur);
1562 	counter_u64_free(rule->states_tot);
1563 	counter_u64_free(rule->src_nodes);
1564 
1565 	mtx_destroy(&rule->rpool.mtx);
1566 	free(rule, M_PFRULE);
1567 }
1568 
1569 static void
pf_kpooladdr_to_pooladdr(const struct pf_kpooladdr * kpool,struct pf_pooladdr * pool)1570 pf_kpooladdr_to_pooladdr(const struct pf_kpooladdr *kpool,
1571     struct pf_pooladdr *pool)
1572 {
1573 
1574 	bzero(pool, sizeof(*pool));
1575 	bcopy(&kpool->addr, &pool->addr, sizeof(pool->addr));
1576 	strlcpy(pool->ifname, kpool->ifname, sizeof(pool->ifname));
1577 }
1578 
1579 static int
pf_pooladdr_to_kpooladdr(const struct pf_pooladdr * pool,struct pf_kpooladdr * kpool)1580 pf_pooladdr_to_kpooladdr(const struct pf_pooladdr *pool,
1581     struct pf_kpooladdr *kpool)
1582 {
1583 	int ret;
1584 
1585 	bzero(kpool, sizeof(*kpool));
1586 	bcopy(&pool->addr, &kpool->addr, sizeof(kpool->addr));
1587 	ret = pf_user_strcpy(kpool->ifname, pool->ifname,
1588 	    sizeof(kpool->ifname));
1589 	return (ret);
1590 }
1591 
1592 static void
pf_kpool_to_pool(const struct pf_kpool * kpool,struct pf_pool * pool)1593 pf_kpool_to_pool(const struct pf_kpool *kpool, struct pf_pool *pool)
1594 {
1595 	bzero(pool, sizeof(*pool));
1596 
1597 	bcopy(&kpool->key, &pool->key, sizeof(pool->key));
1598 	bcopy(&kpool->counter, &pool->counter, sizeof(pool->counter));
1599 
1600 	pool->tblidx = kpool->tblidx;
1601 	pool->proxy_port[0] = kpool->proxy_port[0];
1602 	pool->proxy_port[1] = kpool->proxy_port[1];
1603 	pool->opts = kpool->opts;
1604 }
1605 
1606 static void
pf_pool_to_kpool(const struct pf_pool * pool,struct pf_kpool * kpool)1607 pf_pool_to_kpool(const struct pf_pool *pool, struct pf_kpool *kpool)
1608 {
1609 	_Static_assert(sizeof(pool->key) == sizeof(kpool->key), "");
1610 	_Static_assert(sizeof(pool->counter) == sizeof(kpool->counter), "");
1611 
1612 	bcopy(&pool->key, &kpool->key, sizeof(kpool->key));
1613 	bcopy(&pool->counter, &kpool->counter, sizeof(kpool->counter));
1614 
1615 	kpool->tblidx = pool->tblidx;
1616 	kpool->proxy_port[0] = pool->proxy_port[0];
1617 	kpool->proxy_port[1] = pool->proxy_port[1];
1618 	kpool->opts = pool->opts;
1619 }
1620 
1621 static void
pf_krule_to_rule(const struct pf_krule * krule,struct pf_rule * rule)1622 pf_krule_to_rule(const struct pf_krule *krule, struct pf_rule *rule)
1623 {
1624 
1625 	bzero(rule, sizeof(*rule));
1626 
1627 	bcopy(&krule->src, &rule->src, sizeof(rule->src));
1628 	bcopy(&krule->dst, &rule->dst, sizeof(rule->dst));
1629 
1630 	for (int i = 0; i < PF_SKIP_COUNT; ++i) {
1631 		if (rule->skip[i].ptr == NULL)
1632 			rule->skip[i].nr = -1;
1633 		else
1634 			rule->skip[i].nr = krule->skip[i].ptr->nr;
1635 	}
1636 
1637 	strlcpy(rule->label, krule->label[0], sizeof(rule->label));
1638 	strlcpy(rule->ifname, krule->ifname, sizeof(rule->ifname));
1639 	strlcpy(rule->qname, krule->qname, sizeof(rule->qname));
1640 	strlcpy(rule->pqname, krule->pqname, sizeof(rule->pqname));
1641 	strlcpy(rule->tagname, krule->tagname, sizeof(rule->tagname));
1642 	strlcpy(rule->match_tagname, krule->match_tagname,
1643 	    sizeof(rule->match_tagname));
1644 	strlcpy(rule->overload_tblname, krule->overload_tblname,
1645 	    sizeof(rule->overload_tblname));
1646 
1647 	pf_kpool_to_pool(&krule->rpool, &rule->rpool);
1648 
1649 	rule->evaluations = pf_counter_u64_fetch(&krule->evaluations);
1650 	for (int i = 0; i < 2; i++) {
1651 		rule->packets[i] = pf_counter_u64_fetch(&krule->packets[i]);
1652 		rule->bytes[i] = pf_counter_u64_fetch(&krule->bytes[i]);
1653 	}
1654 
1655 	/* kif, anchor, overload_tbl are not copied over. */
1656 
1657 	rule->os_fingerprint = krule->os_fingerprint;
1658 
1659 	rule->rtableid = krule->rtableid;
1660 	bcopy(krule->timeout, rule->timeout, sizeof(krule->timeout));
1661 	rule->max_states = krule->max_states;
1662 	rule->max_src_nodes = krule->max_src_nodes;
1663 	rule->max_src_states = krule->max_src_states;
1664 	rule->max_src_conn = krule->max_src_conn;
1665 	rule->max_src_conn_rate.limit = krule->max_src_conn_rate.limit;
1666 	rule->max_src_conn_rate.seconds = krule->max_src_conn_rate.seconds;
1667 	rule->qid = krule->qid;
1668 	rule->pqid = krule->pqid;
1669 	rule->nr = krule->nr;
1670 	rule->prob = krule->prob;
1671 	rule->cuid = krule->cuid;
1672 	rule->cpid = krule->cpid;
1673 
1674 	rule->return_icmp = krule->return_icmp;
1675 	rule->return_icmp6 = krule->return_icmp6;
1676 	rule->max_mss = krule->max_mss;
1677 	rule->tag = krule->tag;
1678 	rule->match_tag = krule->match_tag;
1679 	rule->scrub_flags = krule->scrub_flags;
1680 
1681 	bcopy(&krule->uid, &rule->uid, sizeof(krule->uid));
1682 	bcopy(&krule->gid, &rule->gid, sizeof(krule->gid));
1683 
1684 	rule->rule_flag = krule->rule_flag;
1685 	rule->action = krule->action;
1686 	rule->direction = krule->direction;
1687 	rule->log = krule->log;
1688 	rule->logif = krule->logif;
1689 	rule->quick = krule->quick;
1690 	rule->ifnot = krule->ifnot;
1691 	rule->match_tag_not = krule->match_tag_not;
1692 	rule->natpass = krule->natpass;
1693 
1694 	rule->keep_state = krule->keep_state;
1695 	rule->af = krule->af;
1696 	rule->proto = krule->proto;
1697 	rule->type = krule->type;
1698 	rule->code = krule->code;
1699 	rule->flags = krule->flags;
1700 	rule->flagset = krule->flagset;
1701 	rule->min_ttl = krule->min_ttl;
1702 	rule->allow_opts = krule->allow_opts;
1703 	rule->rt = krule->rt;
1704 	rule->return_ttl = krule->return_ttl;
1705 	rule->tos = krule->tos;
1706 	rule->set_tos = krule->set_tos;
1707 	rule->anchor_relative = krule->anchor_relative;
1708 	rule->anchor_wildcard = krule->anchor_wildcard;
1709 
1710 	rule->flush = krule->flush;
1711 	rule->prio = krule->prio;
1712 	rule->set_prio[0] = krule->set_prio[0];
1713 	rule->set_prio[1] = krule->set_prio[1];
1714 
1715 	bcopy(&krule->divert, &rule->divert, sizeof(krule->divert));
1716 
1717 	rule->u_states_cur = counter_u64_fetch(krule->states_cur);
1718 	rule->u_states_tot = counter_u64_fetch(krule->states_tot);
1719 	rule->u_src_nodes = counter_u64_fetch(krule->src_nodes);
1720 }
1721 
1722 static int
pf_rule_to_krule(const struct pf_rule * rule,struct pf_krule * krule)1723 pf_rule_to_krule(const struct pf_rule *rule, struct pf_krule *krule)
1724 {
1725 	int ret;
1726 
1727 #ifndef INET
1728 	if (rule->af == AF_INET) {
1729 		return (EAFNOSUPPORT);
1730 	}
1731 #endif /* INET */
1732 #ifndef INET6
1733 	if (rule->af == AF_INET6) {
1734 		return (EAFNOSUPPORT);
1735 	}
1736 #endif /* INET6 */
1737 
1738 	ret = pf_check_rule_addr(&rule->src);
1739 	if (ret != 0)
1740 		return (ret);
1741 	ret = pf_check_rule_addr(&rule->dst);
1742 	if (ret != 0)
1743 		return (ret);
1744 
1745 	bcopy(&rule->src, &krule->src, sizeof(rule->src));
1746 	bcopy(&rule->dst, &krule->dst, sizeof(rule->dst));
1747 
1748 	ret = pf_user_strcpy(krule->label[0], rule->label, sizeof(rule->label));
1749 	if (ret != 0)
1750 		return (ret);
1751 	ret = pf_user_strcpy(krule->ifname, rule->ifname, sizeof(rule->ifname));
1752 	if (ret != 0)
1753 		return (ret);
1754 	ret = pf_user_strcpy(krule->qname, rule->qname, sizeof(rule->qname));
1755 	if (ret != 0)
1756 		return (ret);
1757 	ret = pf_user_strcpy(krule->pqname, rule->pqname, sizeof(rule->pqname));
1758 	if (ret != 0)
1759 		return (ret);
1760 	ret = pf_user_strcpy(krule->tagname, rule->tagname,
1761 	    sizeof(rule->tagname));
1762 	if (ret != 0)
1763 		return (ret);
1764 	ret = pf_user_strcpy(krule->match_tagname, rule->match_tagname,
1765 	    sizeof(rule->match_tagname));
1766 	if (ret != 0)
1767 		return (ret);
1768 	ret = pf_user_strcpy(krule->overload_tblname, rule->overload_tblname,
1769 	    sizeof(rule->overload_tblname));
1770 	if (ret != 0)
1771 		return (ret);
1772 
1773 	pf_pool_to_kpool(&rule->rpool, &krule->rpool);
1774 
1775 	/* Don't allow userspace to set evaulations, packets or bytes. */
1776 	/* kif, anchor, overload_tbl are not copied over. */
1777 
1778 	krule->os_fingerprint = rule->os_fingerprint;
1779 
1780 	krule->rtableid = rule->rtableid;
1781 	bcopy(rule->timeout, krule->timeout, sizeof(krule->timeout));
1782 	krule->max_states = rule->max_states;
1783 	krule->max_src_nodes = rule->max_src_nodes;
1784 	krule->max_src_states = rule->max_src_states;
1785 	krule->max_src_conn = rule->max_src_conn;
1786 	krule->max_src_conn_rate.limit = rule->max_src_conn_rate.limit;
1787 	krule->max_src_conn_rate.seconds = rule->max_src_conn_rate.seconds;
1788 	krule->qid = rule->qid;
1789 	krule->pqid = rule->pqid;
1790 	krule->nr = rule->nr;
1791 	krule->prob = rule->prob;
1792 	krule->cuid = rule->cuid;
1793 	krule->cpid = rule->cpid;
1794 
1795 	krule->return_icmp = rule->return_icmp;
1796 	krule->return_icmp6 = rule->return_icmp6;
1797 	krule->max_mss = rule->max_mss;
1798 	krule->tag = rule->tag;
1799 	krule->match_tag = rule->match_tag;
1800 	krule->scrub_flags = rule->scrub_flags;
1801 
1802 	bcopy(&rule->uid, &krule->uid, sizeof(krule->uid));
1803 	bcopy(&rule->gid, &krule->gid, sizeof(krule->gid));
1804 
1805 	krule->rule_flag = rule->rule_flag;
1806 	krule->action = rule->action;
1807 	krule->direction = rule->direction;
1808 	krule->log = rule->log;
1809 	krule->logif = rule->logif;
1810 	krule->quick = rule->quick;
1811 	krule->ifnot = rule->ifnot;
1812 	krule->match_tag_not = rule->match_tag_not;
1813 	krule->natpass = rule->natpass;
1814 
1815 	krule->keep_state = rule->keep_state;
1816 	krule->af = rule->af;
1817 	krule->proto = rule->proto;
1818 	krule->type = rule->type;
1819 	krule->code = rule->code;
1820 	krule->flags = rule->flags;
1821 	krule->flagset = rule->flagset;
1822 	krule->min_ttl = rule->min_ttl;
1823 	krule->allow_opts = rule->allow_opts;
1824 	krule->rt = rule->rt;
1825 	krule->return_ttl = rule->return_ttl;
1826 	krule->tos = rule->tos;
1827 	krule->set_tos = rule->set_tos;
1828 
1829 	krule->flush = rule->flush;
1830 	krule->prio = rule->prio;
1831 	krule->set_prio[0] = rule->set_prio[0];
1832 	krule->set_prio[1] = rule->set_prio[1];
1833 
1834 	bcopy(&rule->divert, &krule->divert, sizeof(krule->divert));
1835 
1836 	return (0);
1837 }
1838 
1839 static bool
pf_label_match(const struct pf_krule * rule,const char * label)1840 pf_label_match(const struct pf_krule *rule, const char *label)
1841 {
1842 	int i = 0;
1843 
1844 	while (*rule->label[i]) {
1845 		if (strcmp(rule->label[i], label) == 0)
1846 			return (true);
1847 		i++;
1848 	}
1849 
1850 	return (false);
1851 }
1852 
1853 static unsigned int
pf_kill_matching_state(struct pf_state_key_cmp * key,int dir)1854 pf_kill_matching_state(struct pf_state_key_cmp *key, int dir)
1855 {
1856 	struct pf_kstate *match;
1857 	int more = 0;
1858 	unsigned int killed = 0;
1859 
1860 	/* Call with unlocked hashrow */
1861 
1862 	match = pf_find_state_all(key, dir, &more);
1863 	if (match && !more) {
1864 		pf_unlink_state(match, 0);
1865 		killed++;
1866 	}
1867 
1868 	return (killed);
1869 }
1870 
1871 static int
pf_killstates_row(struct pf_kstate_kill * psk,struct pf_idhash * ih)1872 pf_killstates_row(struct pf_kstate_kill *psk, struct pf_idhash *ih)
1873 {
1874 	struct pf_kstate		*s;
1875 	struct pf_state_key	*sk;
1876 	struct pf_addr		*srcaddr, *dstaddr;
1877 	struct pf_state_key_cmp	 match_key;
1878 	int			 idx, killed = 0;
1879 	unsigned int		 dir;
1880 	u_int16_t		 srcport, dstport;
1881 	struct pfi_kkif		*kif;
1882 
1883 relock_DIOCKILLSTATES:
1884 	PF_HASHROW_LOCK(ih);
1885 	LIST_FOREACH(s, &ih->states, entry) {
1886 		/* For floating states look at the original kif. */
1887 		kif = s->kif == V_pfi_all ? s->orig_kif : s->kif;
1888 
1889 		sk = s->key[PF_SK_WIRE];
1890 		if (s->direction == PF_OUT) {
1891 			srcaddr = &sk->addr[1];
1892 			dstaddr = &sk->addr[0];
1893 			srcport = sk->port[1];
1894 			dstport = sk->port[0];
1895 		} else {
1896 			srcaddr = &sk->addr[0];
1897 			dstaddr = &sk->addr[1];
1898 			srcport = sk->port[0];
1899 			dstport = sk->port[1];
1900 		}
1901 
1902 		if (psk->psk_af && sk->af != psk->psk_af)
1903 			continue;
1904 
1905 		if (psk->psk_proto && psk->psk_proto != sk->proto)
1906 			continue;
1907 
1908 		if (! PF_MATCHA(psk->psk_src.neg, &psk->psk_src.addr.v.a.addr,
1909 		    &psk->psk_src.addr.v.a.mask, srcaddr, sk->af))
1910 			continue;
1911 
1912 		if (! PF_MATCHA(psk->psk_dst.neg, &psk->psk_dst.addr.v.a.addr,
1913 		    &psk->psk_dst.addr.v.a.mask, dstaddr, sk->af))
1914 			continue;
1915 
1916 		if (!  PF_MATCHA(psk->psk_rt_addr.neg,
1917 		    &psk->psk_rt_addr.addr.v.a.addr,
1918 		    &psk->psk_rt_addr.addr.v.a.mask,
1919 		    &s->rt_addr, sk->af))
1920 			continue;
1921 
1922 		if (psk->psk_src.port_op != 0 &&
1923 		    ! pf_match_port(psk->psk_src.port_op,
1924 		    psk->psk_src.port[0], psk->psk_src.port[1], srcport))
1925 			continue;
1926 
1927 		if (psk->psk_dst.port_op != 0 &&
1928 		    ! pf_match_port(psk->psk_dst.port_op,
1929 		    psk->psk_dst.port[0], psk->psk_dst.port[1], dstport))
1930 			continue;
1931 
1932 		if (psk->psk_label[0] &&
1933 		    ! pf_label_match(s->rule.ptr, psk->psk_label))
1934 			continue;
1935 
1936 		if (psk->psk_ifname[0] && strcmp(psk->psk_ifname,
1937 		    kif->pfik_name))
1938 			continue;
1939 
1940 		if (psk->psk_kill_match) {
1941 			/* Create the key to find matching states, with lock
1942 			 * held. */
1943 
1944 			bzero(&match_key, sizeof(match_key));
1945 
1946 			if (s->direction == PF_OUT) {
1947 				dir = PF_IN;
1948 				idx = PF_SK_STACK;
1949 			} else {
1950 				dir = PF_OUT;
1951 				idx = PF_SK_WIRE;
1952 			}
1953 
1954 			match_key.af = s->key[idx]->af;
1955 			match_key.proto = s->key[idx]->proto;
1956 			PF_ACPY(&match_key.addr[0],
1957 			    &s->key[idx]->addr[1], match_key.af);
1958 			match_key.port[0] = s->key[idx]->port[1];
1959 			PF_ACPY(&match_key.addr[1],
1960 			    &s->key[idx]->addr[0], match_key.af);
1961 			match_key.port[1] = s->key[idx]->port[0];
1962 		}
1963 
1964 		pf_unlink_state(s, PF_ENTER_LOCKED);
1965 		killed++;
1966 
1967 		if (psk->psk_kill_match)
1968 			killed += pf_kill_matching_state(&match_key, dir);
1969 
1970 		goto relock_DIOCKILLSTATES;
1971 	}
1972 	PF_HASHROW_UNLOCK(ih);
1973 
1974 	return (killed);
1975 }
1976 
1977 static int
pf_state_kill_to_kstate_kill(const struct pfioc_state_kill * psk,struct pf_kstate_kill * kill)1978 pf_state_kill_to_kstate_kill(const struct pfioc_state_kill *psk,
1979     struct pf_kstate_kill *kill)
1980 {
1981 	int ret;
1982 
1983 	bzero(kill, sizeof(*kill));
1984 
1985 	bcopy(&psk->psk_pfcmp, &kill->psk_pfcmp, sizeof(kill->psk_pfcmp));
1986 	kill->psk_af = psk->psk_af;
1987 	kill->psk_proto = psk->psk_proto;
1988 	bcopy(&psk->psk_src, &kill->psk_src, sizeof(kill->psk_src));
1989 	bcopy(&psk->psk_dst, &kill->psk_dst, sizeof(kill->psk_dst));
1990 	ret = pf_user_strcpy(kill->psk_ifname, psk->psk_ifname,
1991 	    sizeof(kill->psk_ifname));
1992 	if (ret != 0)
1993 		return (ret);
1994 	ret = pf_user_strcpy(kill->psk_label, psk->psk_label,
1995 	    sizeof(kill->psk_label));
1996 	if (ret != 0)
1997 		return (ret);
1998 
1999 	return (0);
2000 }
2001 
2002 static int
pf_ioctl_addrule(struct pf_krule * rule,uint32_t ticket,uint32_t pool_ticket,const char * anchor,const char * anchor_call,struct thread * td)2003 pf_ioctl_addrule(struct pf_krule *rule, uint32_t ticket,
2004     uint32_t pool_ticket, const char *anchor, const char *anchor_call,
2005     struct thread *td)
2006 {
2007 	struct pf_kruleset	*ruleset;
2008 	struct pf_krule		*tail;
2009 	struct pf_kpooladdr	*pa;
2010 	struct pfi_kkif		*kif = NULL;
2011 	int			 rs_num;
2012 	int			 error = 0;
2013 
2014 	if ((rule->return_icmp >> 8) > ICMP_MAXTYPE) {
2015 		error = EINVAL;
2016 		goto errout_unlocked;
2017 	}
2018 
2019 #define	ERROUT(x)	ERROUT_FUNCTION(errout, x)
2020 
2021 	if (rule->ifname[0])
2022 		kif = pf_kkif_create(M_WAITOK);
2023 	pf_counter_u64_init(&rule->evaluations, M_WAITOK);
2024 	for (int i = 0; i < 2; i++) {
2025 		pf_counter_u64_init(&rule->packets[i], M_WAITOK);
2026 		pf_counter_u64_init(&rule->bytes[i], M_WAITOK);
2027 	}
2028 	rule->states_cur = counter_u64_alloc(M_WAITOK);
2029 	rule->states_tot = counter_u64_alloc(M_WAITOK);
2030 	rule->src_nodes = counter_u64_alloc(M_WAITOK);
2031 	rule->cuid = td->td_ucred->cr_ruid;
2032 	rule->cpid = td->td_proc ? td->td_proc->p_pid : 0;
2033 	TAILQ_INIT(&rule->rpool.list);
2034 
2035 	PF_RULES_WLOCK();
2036 #ifdef PF_WANT_32_TO_64_COUNTER
2037 	LIST_INSERT_HEAD(&V_pf_allrulelist, rule, allrulelist);
2038 	MPASS(!rule->allrulelinked);
2039 	rule->allrulelinked = true;
2040 	V_pf_allrulecount++;
2041 #endif
2042 	ruleset = pf_find_kruleset(anchor);
2043 	if (ruleset == NULL)
2044 		ERROUT(EINVAL);
2045 	rs_num = pf_get_ruleset_number(rule->action);
2046 	if (rs_num >= PF_RULESET_MAX)
2047 		ERROUT(EINVAL);
2048 	if (ticket != ruleset->rules[rs_num].inactive.ticket) {
2049 		DPFPRINTF(PF_DEBUG_MISC,
2050 		    ("ticket: %d != [%d]%d\n", ticket, rs_num,
2051 		    ruleset->rules[rs_num].inactive.ticket));
2052 		ERROUT(EBUSY);
2053 	}
2054 	if (pool_ticket != V_ticket_pabuf) {
2055 		DPFPRINTF(PF_DEBUG_MISC,
2056 		    ("pool_ticket: %d != %d\n", pool_ticket,
2057 		    V_ticket_pabuf));
2058 		ERROUT(EBUSY);
2059 	}
2060 
2061 	tail = TAILQ_LAST(ruleset->rules[rs_num].inactive.ptr,
2062 	    pf_krulequeue);
2063 	if (tail)
2064 		rule->nr = tail->nr + 1;
2065 	else
2066 		rule->nr = 0;
2067 	if (rule->ifname[0]) {
2068 		rule->kif = pfi_kkif_attach(kif, rule->ifname);
2069 		kif = NULL;
2070 		pfi_kkif_ref(rule->kif);
2071 	} else
2072 		rule->kif = NULL;
2073 
2074 	if (rule->rtableid > 0 && rule->rtableid >= rt_numfibs)
2075 		error = EBUSY;
2076 
2077 #ifdef ALTQ
2078 	/* set queue IDs */
2079 	if (rule->qname[0] != 0) {
2080 		if ((rule->qid = pf_qname2qid(rule->qname)) == 0)
2081 			error = EBUSY;
2082 		else if (rule->pqname[0] != 0) {
2083 			if ((rule->pqid =
2084 			    pf_qname2qid(rule->pqname)) == 0)
2085 				error = EBUSY;
2086 		} else
2087 			rule->pqid = rule->qid;
2088 	}
2089 #endif
2090 	if (rule->tagname[0])
2091 		if ((rule->tag = pf_tagname2tag(rule->tagname)) == 0)
2092 			error = EBUSY;
2093 	if (rule->match_tagname[0])
2094 		if ((rule->match_tag =
2095 		    pf_tagname2tag(rule->match_tagname)) == 0)
2096 			error = EBUSY;
2097 	if (rule->rt && !rule->direction)
2098 		error = EINVAL;
2099 	if (!rule->log)
2100 		rule->logif = 0;
2101 	if (rule->logif >= PFLOGIFS_MAX)
2102 		error = EINVAL;
2103 	if (pf_addr_setup(ruleset, &rule->src.addr, rule->af))
2104 		error = ENOMEM;
2105 	if (pf_addr_setup(ruleset, &rule->dst.addr, rule->af))
2106 		error = ENOMEM;
2107 	if (pf_kanchor_setup(rule, ruleset, anchor_call))
2108 		error = EINVAL;
2109 	if (rule->scrub_flags & PFSTATE_SETPRIO &&
2110 	    (rule->set_prio[0] > PF_PRIO_MAX ||
2111 	    rule->set_prio[1] > PF_PRIO_MAX))
2112 		error = EINVAL;
2113 	TAILQ_FOREACH(pa, &V_pf_pabuf, entries)
2114 		if (pa->addr.type == PF_ADDR_TABLE) {
2115 			pa->addr.p.tbl = pfr_attach_table(ruleset,
2116 			    pa->addr.v.tblname);
2117 			if (pa->addr.p.tbl == NULL)
2118 				error = ENOMEM;
2119 		}
2120 
2121 	rule->overload_tbl = NULL;
2122 	if (rule->overload_tblname[0]) {
2123 		if ((rule->overload_tbl = pfr_attach_table(ruleset,
2124 		    rule->overload_tblname)) == NULL)
2125 			error = EINVAL;
2126 		else
2127 			rule->overload_tbl->pfrkt_flags |=
2128 			    PFR_TFLAG_ACTIVE;
2129 	}
2130 
2131 	pf_mv_kpool(&V_pf_pabuf, &rule->rpool.list);
2132 	if (((((rule->action == PF_NAT) || (rule->action == PF_RDR) ||
2133 	    (rule->action == PF_BINAT)) && rule->anchor == NULL) ||
2134 	    (rule->rt > PF_NOPFROUTE)) &&
2135 	    (TAILQ_FIRST(&rule->rpool.list) == NULL))
2136 		error = EINVAL;
2137 
2138 	if (error) {
2139 		pf_free_rule(rule);
2140 		rule = NULL;
2141 		ERROUT(error);
2142 	}
2143 
2144 	rule->rpool.cur = TAILQ_FIRST(&rule->rpool.list);
2145 	pf_counter_u64_zero(&rule->evaluations);
2146 	for (int i = 0; i < 2; i++) {
2147 		pf_counter_u64_zero(&rule->packets[i]);
2148 		pf_counter_u64_zero(&rule->bytes[i]);
2149 	}
2150 	TAILQ_INSERT_TAIL(ruleset->rules[rs_num].inactive.ptr,
2151 	    rule, entries);
2152 	ruleset->rules[rs_num].inactive.rcount++;
2153 
2154 	PF_RULES_WUNLOCK();
2155 
2156 	return (0);
2157 
2158 #undef ERROUT
2159 errout:
2160 	PF_RULES_WUNLOCK();
2161 errout_unlocked:
2162 	pf_kkif_free(kif);
2163 	pf_krule_free(rule);
2164 	return (error);
2165 }
2166 
2167 static int
pfioctl(struct cdev * dev,u_long cmd,caddr_t addr,int flags,struct thread * td)2168 pfioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flags, struct thread *td)
2169 {
2170 	int			 error = 0;
2171 	PF_RULES_RLOCK_TRACKER;
2172 
2173 #define	ERROUT_IOCTL(target, x)					\
2174     do {								\
2175 	    error = (x);						\
2176 	    SDT_PROBE3(pf, ioctl, ioctl, error, cmd, error, __LINE__);	\
2177 	    goto target;						\
2178     } while (0)
2179 
2180 
2181 	/* XXX keep in sync with switch() below */
2182 	if (securelevel_gt(td->td_ucred, 2))
2183 		switch (cmd) {
2184 		case DIOCGETRULES:
2185 		case DIOCGETRULE:
2186 		case DIOCGETRULENV:
2187 		case DIOCGETADDRS:
2188 		case DIOCGETADDR:
2189 		case DIOCGETSTATE:
2190 		case DIOCGETSTATENV:
2191 		case DIOCSETSTATUSIF:
2192 		case DIOCGETSTATUS:
2193 		case DIOCGETSTATUSNV:
2194 		case DIOCCLRSTATUS:
2195 		case DIOCNATLOOK:
2196 		case DIOCSETDEBUG:
2197 		case DIOCGETSTATES:
2198 		case DIOCGETSTATESV2:
2199 		case DIOCGETTIMEOUT:
2200 		case DIOCCLRRULECTRS:
2201 		case DIOCGETLIMIT:
2202 		case DIOCGETALTQSV0:
2203 		case DIOCGETALTQSV1:
2204 		case DIOCGETALTQV0:
2205 		case DIOCGETALTQV1:
2206 		case DIOCGETQSTATSV0:
2207 		case DIOCGETQSTATSV1:
2208 		case DIOCGETRULESETS:
2209 		case DIOCGETRULESET:
2210 		case DIOCRGETTABLES:
2211 		case DIOCRGETTSTATS:
2212 		case DIOCRCLRTSTATS:
2213 		case DIOCRCLRADDRS:
2214 		case DIOCRADDADDRS:
2215 		case DIOCRDELADDRS:
2216 		case DIOCRSETADDRS:
2217 		case DIOCRGETADDRS:
2218 		case DIOCRGETASTATS:
2219 		case DIOCRCLRASTATS:
2220 		case DIOCRTSTADDRS:
2221 		case DIOCOSFPGET:
2222 		case DIOCGETSRCNODES:
2223 		case DIOCCLRSRCNODES:
2224 		case DIOCGETSYNCOOKIES:
2225 		case DIOCIGETIFACES:
2226 		case DIOCGIFSPEEDV0:
2227 		case DIOCGIFSPEEDV1:
2228 		case DIOCSETIFFLAG:
2229 		case DIOCCLRIFFLAG:
2230 			break;
2231 		case DIOCRCLRTABLES:
2232 		case DIOCRADDTABLES:
2233 		case DIOCRDELTABLES:
2234 		case DIOCRSETTFLAGS:
2235 			if (((struct pfioc_table *)addr)->pfrio_flags &
2236 			    PFR_FLAG_DUMMY)
2237 				break; /* dummy operation ok */
2238 			return (EPERM);
2239 		default:
2240 			return (EPERM);
2241 		}
2242 
2243 	if (!(flags & FWRITE))
2244 		switch (cmd) {
2245 		case DIOCGETRULES:
2246 		case DIOCGETADDRS:
2247 		case DIOCGETADDR:
2248 		case DIOCGETSTATE:
2249 		case DIOCGETSTATENV:
2250 		case DIOCGETSTATUS:
2251 		case DIOCGETSTATUSNV:
2252 		case DIOCGETSTATES:
2253 		case DIOCGETSTATESV2:
2254 		case DIOCGETTIMEOUT:
2255 		case DIOCGETLIMIT:
2256 		case DIOCGETALTQSV0:
2257 		case DIOCGETALTQSV1:
2258 		case DIOCGETALTQV0:
2259 		case DIOCGETALTQV1:
2260 		case DIOCGETQSTATSV0:
2261 		case DIOCGETQSTATSV1:
2262 		case DIOCGETRULESETS:
2263 		case DIOCGETRULESET:
2264 		case DIOCNATLOOK:
2265 		case DIOCRGETTABLES:
2266 		case DIOCRGETTSTATS:
2267 		case DIOCRGETADDRS:
2268 		case DIOCRGETASTATS:
2269 		case DIOCRTSTADDRS:
2270 		case DIOCOSFPGET:
2271 		case DIOCGETSRCNODES:
2272 		case DIOCGETSYNCOOKIES:
2273 		case DIOCIGETIFACES:
2274 		case DIOCGIFSPEEDV1:
2275 		case DIOCGIFSPEEDV0:
2276 		case DIOCGETRULENV:
2277 			break;
2278 		case DIOCRCLRTABLES:
2279 		case DIOCRADDTABLES:
2280 		case DIOCRDELTABLES:
2281 		case DIOCRCLRTSTATS:
2282 		case DIOCRCLRADDRS:
2283 		case DIOCRADDADDRS:
2284 		case DIOCRDELADDRS:
2285 		case DIOCRSETADDRS:
2286 		case DIOCRSETTFLAGS:
2287 			if (((struct pfioc_table *)addr)->pfrio_flags &
2288 			    PFR_FLAG_DUMMY) {
2289 				flags |= FWRITE; /* need write lock for dummy */
2290 				break; /* dummy operation ok */
2291 			}
2292 			return (EACCES);
2293 		case DIOCGETRULE:
2294 			if (((struct pfioc_rule *)addr)->action ==
2295 			    PF_GET_CLR_CNTR)
2296 				return (EACCES);
2297 			break;
2298 		default:
2299 			return (EACCES);
2300 		}
2301 
2302 	CURVNET_SET(TD_TO_VNET(td));
2303 
2304 	switch (cmd) {
2305 	case DIOCSTART:
2306 		sx_xlock(&pf_ioctl_lock);
2307 		if (V_pf_status.running)
2308 			error = EEXIST;
2309 		else {
2310 			int cpu;
2311 
2312 			hook_pf();
2313 			V_pf_status.running = 1;
2314 			V_pf_status.since = time_second;
2315 
2316 			CPU_FOREACH(cpu)
2317 				V_pf_stateid[cpu] = time_second;
2318 
2319 			DPFPRINTF(PF_DEBUG_MISC, ("pf: started\n"));
2320 		}
2321 		break;
2322 
2323 	case DIOCSTOP:
2324 		sx_xlock(&pf_ioctl_lock);
2325 		if (!V_pf_status.running)
2326 			error = ENOENT;
2327 		else {
2328 			V_pf_status.running = 0;
2329 			dehook_pf();
2330 			V_pf_status.since = time_second;
2331 			DPFPRINTF(PF_DEBUG_MISC, ("pf: stopped\n"));
2332 		}
2333 		break;
2334 
2335 	case DIOCADDRULENV: {
2336 		struct pfioc_nv	*nv = (struct pfioc_nv *)addr;
2337 		nvlist_t	*nvl = NULL;
2338 		void		*nvlpacked = NULL;
2339 		struct pf_krule	*rule = NULL;
2340 		const char	*anchor = "", *anchor_call = "";
2341 		uint32_t	 ticket = 0, pool_ticket = 0;
2342 
2343 #define	ERROUT(x)	ERROUT_IOCTL(DIOCADDRULENV_error, x)
2344 
2345 		if (nv->len > pf_ioctl_maxcount)
2346 			ERROUT(ENOMEM);
2347 
2348 		nvlpacked = malloc(nv->len, M_TEMP, M_WAITOK);
2349 		error = copyin(nv->data, nvlpacked, nv->len);
2350 		if (error)
2351 			ERROUT(error);
2352 
2353 		nvl = nvlist_unpack(nvlpacked, nv->len, 0);
2354 		if (nvl == NULL)
2355 			ERROUT(EBADMSG);
2356 
2357 		if (! nvlist_exists_number(nvl, "ticket"))
2358 			ERROUT(EINVAL);
2359 		ticket = nvlist_get_number(nvl, "ticket");
2360 
2361 		if (! nvlist_exists_number(nvl, "pool_ticket"))
2362 			ERROUT(EINVAL);
2363 		pool_ticket = nvlist_get_number(nvl, "pool_ticket");
2364 
2365 		if (! nvlist_exists_nvlist(nvl, "rule"))
2366 			ERROUT(EINVAL);
2367 
2368 		rule = pf_krule_alloc();
2369 		error = pf_nvrule_to_krule(nvlist_get_nvlist(nvl, "rule"),
2370 		    rule);
2371 		if (error)
2372 			ERROUT(error);
2373 
2374 		if (nvlist_exists_string(nvl, "anchor"))
2375 			anchor = nvlist_get_string(nvl, "anchor");
2376 		if (nvlist_exists_string(nvl, "anchor_call"))
2377 			anchor_call = nvlist_get_string(nvl, "anchor_call");
2378 
2379 		if ((error = nvlist_error(nvl)))
2380 			ERROUT(error);
2381 
2382 		/* Frees rule on error */
2383 		error = pf_ioctl_addrule(rule, ticket, pool_ticket, anchor,
2384 		    anchor_call, td);
2385 
2386 		nvlist_destroy(nvl);
2387 		free(nvlpacked, M_TEMP);
2388 		break;
2389 #undef ERROUT
2390 DIOCADDRULENV_error:
2391 		pf_krule_free(rule);
2392 		nvlist_destroy(nvl);
2393 		free(nvlpacked, M_TEMP);
2394 
2395 		break;
2396 	}
2397 	case DIOCADDRULE: {
2398 		struct pfioc_rule	*pr = (struct pfioc_rule *)addr;
2399 		struct pf_krule		*rule;
2400 
2401 		rule = pf_krule_alloc();
2402 		error = pf_rule_to_krule(&pr->rule, rule);
2403 		if (error != 0) {
2404 			pf_krule_free(rule);
2405 			break;
2406 		}
2407 
2408 		pr->anchor[sizeof(pr->anchor) - 1] = 0;
2409 
2410 		/* Frees rule on error */
2411 		error = pf_ioctl_addrule(rule, pr->ticket, pr->pool_ticket,
2412 		    pr->anchor, pr->anchor_call, td);
2413 		break;
2414 	}
2415 
2416 	case DIOCGETRULES: {
2417 		struct pfioc_rule	*pr = (struct pfioc_rule *)addr;
2418 		struct pf_kruleset	*ruleset;
2419 		struct pf_krule		*tail;
2420 		int			 rs_num;
2421 
2422 		pr->anchor[sizeof(pr->anchor) - 1] = 0;
2423 
2424 		PF_RULES_WLOCK();
2425 		ruleset = pf_find_kruleset(pr->anchor);
2426 		if (ruleset == NULL) {
2427 			PF_RULES_WUNLOCK();
2428 			error = EINVAL;
2429 			break;
2430 		}
2431 		rs_num = pf_get_ruleset_number(pr->rule.action);
2432 		if (rs_num >= PF_RULESET_MAX) {
2433 			PF_RULES_WUNLOCK();
2434 			error = EINVAL;
2435 			break;
2436 		}
2437 		tail = TAILQ_LAST(ruleset->rules[rs_num].active.ptr,
2438 		    pf_krulequeue);
2439 		if (tail)
2440 			pr->nr = tail->nr + 1;
2441 		else
2442 			pr->nr = 0;
2443 		pr->ticket = ruleset->rules[rs_num].active.ticket;
2444 		PF_RULES_WUNLOCK();
2445 		break;
2446 	}
2447 
2448 	case DIOCGETRULE: {
2449 		struct pfioc_rule	*pr = (struct pfioc_rule *)addr;
2450 		struct pf_kruleset	*ruleset;
2451 		struct pf_krule		*rule;
2452 		int			 rs_num;
2453 
2454 		pr->anchor[sizeof(pr->anchor) - 1] = 0;
2455 
2456 		PF_RULES_WLOCK();
2457 		ruleset = pf_find_kruleset(pr->anchor);
2458 		if (ruleset == NULL) {
2459 			PF_RULES_WUNLOCK();
2460 			error = EINVAL;
2461 			break;
2462 		}
2463 		rs_num = pf_get_ruleset_number(pr->rule.action);
2464 		if (rs_num >= PF_RULESET_MAX) {
2465 			PF_RULES_WUNLOCK();
2466 			error = EINVAL;
2467 			break;
2468 		}
2469 		if (pr->ticket != ruleset->rules[rs_num].active.ticket) {
2470 			PF_RULES_WUNLOCK();
2471 			error = EBUSY;
2472 			break;
2473 		}
2474 		rule = TAILQ_FIRST(ruleset->rules[rs_num].active.ptr);
2475 		while ((rule != NULL) && (rule->nr != pr->nr))
2476 			rule = TAILQ_NEXT(rule, entries);
2477 		if (rule == NULL) {
2478 			PF_RULES_WUNLOCK();
2479 			error = EBUSY;
2480 			break;
2481 		}
2482 
2483 		pf_krule_to_rule(rule, &pr->rule);
2484 
2485 		if (pf_kanchor_copyout(ruleset, rule, pr)) {
2486 			PF_RULES_WUNLOCK();
2487 			error = EBUSY;
2488 			break;
2489 		}
2490 		pf_addr_copyout(&pr->rule.src.addr);
2491 		pf_addr_copyout(&pr->rule.dst.addr);
2492 
2493 		if (pr->action == PF_GET_CLR_CNTR) {
2494 			pf_counter_u64_zero(&rule->evaluations);
2495 			for (int i = 0; i < 2; i++) {
2496 				pf_counter_u64_zero(&rule->packets[i]);
2497 				pf_counter_u64_zero(&rule->bytes[i]);
2498 			}
2499 			counter_u64_zero(rule->states_tot);
2500 		}
2501 		PF_RULES_WUNLOCK();
2502 		break;
2503 	}
2504 
2505 	case DIOCGETRULENV: {
2506 		struct pfioc_nv		*nv = (struct pfioc_nv *)addr;
2507 		nvlist_t		*nvrule = NULL;
2508 		nvlist_t		*nvl = NULL;
2509 		struct pf_kruleset	*ruleset;
2510 		struct pf_krule		*rule;
2511 		void			*nvlpacked = NULL;
2512 		int			 rs_num, nr;
2513 		bool			 clear_counter = false;
2514 
2515 #define	ERROUT(x)	ERROUT_IOCTL(DIOCGETRULENV_error, x)
2516 
2517 		if (nv->len > pf_ioctl_maxcount)
2518 			ERROUT(ENOMEM);
2519 
2520 		/* Copy the request in */
2521 		nvlpacked = malloc(nv->len, M_NVLIST, M_WAITOK);
2522 		if (nvlpacked == NULL)
2523 			ERROUT(ENOMEM);
2524 
2525 		error = copyin(nv->data, nvlpacked, nv->len);
2526 		if (error)
2527 			ERROUT(error);
2528 
2529 		nvl = nvlist_unpack(nvlpacked, nv->len, 0);
2530 		if (nvl == NULL)
2531 			ERROUT(EBADMSG);
2532 
2533 		if (! nvlist_exists_string(nvl, "anchor"))
2534 			ERROUT(EBADMSG);
2535 		if (! nvlist_exists_number(nvl, "ruleset"))
2536 			ERROUT(EBADMSG);
2537 		if (! nvlist_exists_number(nvl, "ticket"))
2538 			ERROUT(EBADMSG);
2539 		if (! nvlist_exists_number(nvl, "nr"))
2540 			ERROUT(EBADMSG);
2541 
2542 		if (nvlist_exists_bool(nvl, "clear_counter"))
2543 			clear_counter = nvlist_get_bool(nvl, "clear_counter");
2544 
2545 		if (clear_counter && !(flags & FWRITE))
2546 			ERROUT(EACCES);
2547 
2548 		nr = nvlist_get_number(nvl, "nr");
2549 
2550 		PF_RULES_WLOCK();
2551 		ruleset = pf_find_kruleset(nvlist_get_string(nvl, "anchor"));
2552 		if (ruleset == NULL) {
2553 			PF_RULES_WUNLOCK();
2554 			ERROUT(ENOENT);
2555 		}
2556 
2557 		rs_num = pf_get_ruleset_number(nvlist_get_number(nvl, "ruleset"));
2558 		if (rs_num >= PF_RULESET_MAX) {
2559 			PF_RULES_WUNLOCK();
2560 			ERROUT(EINVAL);
2561 		}
2562 
2563 		if (nvlist_get_number(nvl, "ticket") !=
2564 		    ruleset->rules[rs_num].active.ticket) {
2565 			PF_RULES_WUNLOCK();
2566 			ERROUT(EBUSY);
2567 		}
2568 
2569 		if ((error = nvlist_error(nvl))) {
2570 			PF_RULES_WUNLOCK();
2571 			ERROUT(error);
2572 		}
2573 
2574 		rule = TAILQ_FIRST(ruleset->rules[rs_num].active.ptr);
2575 		while ((rule != NULL) && (rule->nr != nr))
2576 			rule = TAILQ_NEXT(rule, entries);
2577 		if (rule == NULL) {
2578 			PF_RULES_WUNLOCK();
2579 			ERROUT(EBUSY);
2580 		}
2581 
2582 		nvrule = pf_krule_to_nvrule(rule);
2583 
2584 		nvlist_destroy(nvl);
2585 		nvl = nvlist_create(0);
2586 		if (nvl == NULL) {
2587 			PF_RULES_WUNLOCK();
2588 			ERROUT(ENOMEM);
2589 		}
2590 		nvlist_add_number(nvl, "nr", nr);
2591 		nvlist_add_nvlist(nvl, "rule", nvrule);
2592 		nvlist_destroy(nvrule);
2593 		nvrule = NULL;
2594 		if (pf_kanchor_nvcopyout(ruleset, rule, nvl)) {
2595 			PF_RULES_WUNLOCK();
2596 			ERROUT(EBUSY);
2597 		}
2598 
2599 		free(nvlpacked, M_NVLIST);
2600 		nvlpacked = nvlist_pack(nvl, &nv->len);
2601 		if (nvlpacked == NULL) {
2602 			PF_RULES_WUNLOCK();
2603 			ERROUT(ENOMEM);
2604 		}
2605 
2606 		if (nv->size == 0) {
2607 			PF_RULES_WUNLOCK();
2608 			ERROUT(0);
2609 		}
2610 		else if (nv->size < nv->len) {
2611 			PF_RULES_WUNLOCK();
2612 			ERROUT(ENOSPC);
2613 		}
2614 
2615 		if (clear_counter) {
2616 			pf_counter_u64_zero(&rule->evaluations);
2617 			for (int i = 0; i < 2; i++) {
2618 				pf_counter_u64_zero(&rule->packets[i]);
2619 				pf_counter_u64_zero(&rule->bytes[i]);
2620 			}
2621 			counter_u64_zero(rule->states_tot);
2622 		}
2623 		PF_RULES_WUNLOCK();
2624 
2625 		error = copyout(nvlpacked, nv->data, nv->len);
2626 
2627 #undef ERROUT
2628 DIOCGETRULENV_error:
2629 		free(nvlpacked, M_NVLIST);
2630 		nvlist_destroy(nvrule);
2631 		nvlist_destroy(nvl);
2632 
2633 		break;
2634 	}
2635 
2636 	case DIOCCHANGERULE: {
2637 		struct pfioc_rule	*pcr = (struct pfioc_rule *)addr;
2638 		struct pf_kruleset	*ruleset;
2639 		struct pf_krule		*oldrule = NULL, *newrule = NULL;
2640 		struct pfi_kkif		*kif = NULL;
2641 		struct pf_kpooladdr	*pa;
2642 		u_int32_t		 nr = 0;
2643 		int			 rs_num;
2644 
2645 		pcr->anchor[sizeof(pcr->anchor) - 1] = 0;
2646 
2647 		if (pcr->action < PF_CHANGE_ADD_HEAD ||
2648 		    pcr->action > PF_CHANGE_GET_TICKET) {
2649 			error = EINVAL;
2650 			break;
2651 		}
2652 		if (pcr->rule.return_icmp >> 8 > ICMP_MAXTYPE) {
2653 			error = EINVAL;
2654 			break;
2655 		}
2656 
2657 		if (pcr->action != PF_CHANGE_REMOVE) {
2658 			newrule = pf_krule_alloc();
2659 			error = pf_rule_to_krule(&pcr->rule, newrule);
2660 			if (error != 0) {
2661 				free(newrule, M_PFRULE);
2662 				break;
2663 			}
2664 
2665 			if (newrule->ifname[0])
2666 				kif = pf_kkif_create(M_WAITOK);
2667 			pf_counter_u64_init(&newrule->evaluations, M_WAITOK);
2668 			for (int i = 0; i < 2; i++) {
2669 				pf_counter_u64_init(&newrule->packets[i], M_WAITOK);
2670 				pf_counter_u64_init(&newrule->bytes[i], M_WAITOK);
2671 			}
2672 			newrule->states_cur = counter_u64_alloc(M_WAITOK);
2673 			newrule->states_tot = counter_u64_alloc(M_WAITOK);
2674 			newrule->src_nodes = counter_u64_alloc(M_WAITOK);
2675 			newrule->cuid = td->td_ucred->cr_ruid;
2676 			newrule->cpid = td->td_proc ? td->td_proc->p_pid : 0;
2677 			TAILQ_INIT(&newrule->rpool.list);
2678 		}
2679 #define	ERROUT(x)	{ error = (x); goto DIOCCHANGERULE_error; }
2680 
2681 		PF_RULES_WLOCK();
2682 #ifdef PF_WANT_32_TO_64_COUNTER
2683 		if (newrule != NULL) {
2684 			LIST_INSERT_HEAD(&V_pf_allrulelist, newrule, allrulelist);
2685 			newrule->allrulelinked = true;
2686 			V_pf_allrulecount++;
2687 		}
2688 #endif
2689 
2690 		if (!(pcr->action == PF_CHANGE_REMOVE ||
2691 		    pcr->action == PF_CHANGE_GET_TICKET) &&
2692 		    pcr->pool_ticket != V_ticket_pabuf)
2693 			ERROUT(EBUSY);
2694 
2695 		ruleset = pf_find_kruleset(pcr->anchor);
2696 		if (ruleset == NULL)
2697 			ERROUT(EINVAL);
2698 
2699 		rs_num = pf_get_ruleset_number(pcr->rule.action);
2700 		if (rs_num >= PF_RULESET_MAX)
2701 			ERROUT(EINVAL);
2702 
2703 		if (pcr->action == PF_CHANGE_GET_TICKET) {
2704 			pcr->ticket = ++ruleset->rules[rs_num].active.ticket;
2705 			ERROUT(0);
2706 		} else if (pcr->ticket !=
2707 			    ruleset->rules[rs_num].active.ticket)
2708 				ERROUT(EINVAL);
2709 
2710 		if (pcr->action != PF_CHANGE_REMOVE) {
2711 			if (newrule->ifname[0]) {
2712 				newrule->kif = pfi_kkif_attach(kif,
2713 				    newrule->ifname);
2714 				kif = NULL;
2715 				pfi_kkif_ref(newrule->kif);
2716 			} else
2717 				newrule->kif = NULL;
2718 
2719 			if (newrule->rtableid > 0 &&
2720 			    newrule->rtableid >= rt_numfibs)
2721 				error = EBUSY;
2722 
2723 #ifdef ALTQ
2724 			/* set queue IDs */
2725 			if (newrule->qname[0] != 0) {
2726 				if ((newrule->qid =
2727 				    pf_qname2qid(newrule->qname)) == 0)
2728 					error = EBUSY;
2729 				else if (newrule->pqname[0] != 0) {
2730 					if ((newrule->pqid =
2731 					    pf_qname2qid(newrule->pqname)) == 0)
2732 						error = EBUSY;
2733 				} else
2734 					newrule->pqid = newrule->qid;
2735 			}
2736 #endif /* ALTQ */
2737 			if (newrule->tagname[0])
2738 				if ((newrule->tag =
2739 				    pf_tagname2tag(newrule->tagname)) == 0)
2740 					error = EBUSY;
2741 			if (newrule->match_tagname[0])
2742 				if ((newrule->match_tag = pf_tagname2tag(
2743 				    newrule->match_tagname)) == 0)
2744 					error = EBUSY;
2745 			if (newrule->rt && !newrule->direction)
2746 				error = EINVAL;
2747 			if (!newrule->log)
2748 				newrule->logif = 0;
2749 			if (newrule->logif >= PFLOGIFS_MAX)
2750 				error = EINVAL;
2751 			if (pf_addr_setup(ruleset, &newrule->src.addr, newrule->af))
2752 				error = ENOMEM;
2753 			if (pf_addr_setup(ruleset, &newrule->dst.addr, newrule->af))
2754 				error = ENOMEM;
2755 			if (pf_kanchor_setup(newrule, ruleset, pcr->anchor_call))
2756 				error = EINVAL;
2757 			TAILQ_FOREACH(pa, &V_pf_pabuf, entries)
2758 				if (pa->addr.type == PF_ADDR_TABLE) {
2759 					pa->addr.p.tbl =
2760 					    pfr_attach_table(ruleset,
2761 					    pa->addr.v.tblname);
2762 					if (pa->addr.p.tbl == NULL)
2763 						error = ENOMEM;
2764 				}
2765 
2766 			newrule->overload_tbl = NULL;
2767 			if (newrule->overload_tblname[0]) {
2768 				if ((newrule->overload_tbl = pfr_attach_table(
2769 				    ruleset, newrule->overload_tblname)) ==
2770 				    NULL)
2771 					error = EINVAL;
2772 				else
2773 					newrule->overload_tbl->pfrkt_flags |=
2774 					    PFR_TFLAG_ACTIVE;
2775 			}
2776 
2777 			pf_mv_kpool(&V_pf_pabuf, &newrule->rpool.list);
2778 			if (((((newrule->action == PF_NAT) ||
2779 			    (newrule->action == PF_RDR) ||
2780 			    (newrule->action == PF_BINAT) ||
2781 			    (newrule->rt > PF_NOPFROUTE)) &&
2782 			    !newrule->anchor)) &&
2783 			    (TAILQ_FIRST(&newrule->rpool.list) == NULL))
2784 				error = EINVAL;
2785 
2786 			if (error) {
2787 				pf_free_rule(newrule);
2788 				PF_RULES_WUNLOCK();
2789 				break;
2790 			}
2791 
2792 			newrule->rpool.cur = TAILQ_FIRST(&newrule->rpool.list);
2793 		}
2794 		pf_empty_kpool(&V_pf_pabuf);
2795 
2796 		if (pcr->action == PF_CHANGE_ADD_HEAD)
2797 			oldrule = TAILQ_FIRST(
2798 			    ruleset->rules[rs_num].active.ptr);
2799 		else if (pcr->action == PF_CHANGE_ADD_TAIL)
2800 			oldrule = TAILQ_LAST(
2801 			    ruleset->rules[rs_num].active.ptr, pf_krulequeue);
2802 		else {
2803 			oldrule = TAILQ_FIRST(
2804 			    ruleset->rules[rs_num].active.ptr);
2805 			while ((oldrule != NULL) && (oldrule->nr != pcr->nr))
2806 				oldrule = TAILQ_NEXT(oldrule, entries);
2807 			if (oldrule == NULL) {
2808 				if (newrule != NULL)
2809 					pf_free_rule(newrule);
2810 				PF_RULES_WUNLOCK();
2811 				error = EINVAL;
2812 				break;
2813 			}
2814 		}
2815 
2816 		if (pcr->action == PF_CHANGE_REMOVE) {
2817 			pf_unlink_rule(ruleset->rules[rs_num].active.ptr,
2818 			    oldrule);
2819 			ruleset->rules[rs_num].active.rcount--;
2820 		} else {
2821 			if (oldrule == NULL)
2822 				TAILQ_INSERT_TAIL(
2823 				    ruleset->rules[rs_num].active.ptr,
2824 				    newrule, entries);
2825 			else if (pcr->action == PF_CHANGE_ADD_HEAD ||
2826 			    pcr->action == PF_CHANGE_ADD_BEFORE)
2827 				TAILQ_INSERT_BEFORE(oldrule, newrule, entries);
2828 			else
2829 				TAILQ_INSERT_AFTER(
2830 				    ruleset->rules[rs_num].active.ptr,
2831 				    oldrule, newrule, entries);
2832 			ruleset->rules[rs_num].active.rcount++;
2833 		}
2834 
2835 		nr = 0;
2836 		TAILQ_FOREACH(oldrule,
2837 		    ruleset->rules[rs_num].active.ptr, entries)
2838 			oldrule->nr = nr++;
2839 
2840 		ruleset->rules[rs_num].active.ticket++;
2841 
2842 		pf_calc_skip_steps(ruleset->rules[rs_num].active.ptr);
2843 		pf_remove_if_empty_kruleset(ruleset);
2844 
2845 		PF_RULES_WUNLOCK();
2846 		break;
2847 
2848 #undef ERROUT
2849 DIOCCHANGERULE_error:
2850 		PF_RULES_WUNLOCK();
2851 		pf_krule_free(newrule);
2852 		pf_kkif_free(kif);
2853 		break;
2854 	}
2855 
2856 	case DIOCCLRSTATES: {
2857 		struct pfioc_state_kill *psk = (struct pfioc_state_kill *)addr;
2858 		struct pf_kstate_kill	 kill;
2859 
2860 		error = pf_state_kill_to_kstate_kill(psk, &kill);
2861 		if (error)
2862 			break;
2863 
2864 		psk->psk_killed = pf_clear_states(&kill);
2865 		break;
2866 	}
2867 
2868 	case DIOCCLRSTATESNV: {
2869 		error = pf_clearstates_nv((struct pfioc_nv *)addr);
2870 		break;
2871 	}
2872 
2873 	case DIOCKILLSTATES: {
2874 		struct pfioc_state_kill	*psk = (struct pfioc_state_kill *)addr;
2875 		struct pf_kstate_kill	 kill;
2876 
2877 		error = pf_state_kill_to_kstate_kill(psk, &kill);
2878 		if (error)
2879 			break;
2880 
2881 		psk->psk_killed = 0;
2882 		pf_killstates(&kill, &psk->psk_killed);
2883 		break;
2884 	}
2885 
2886 	case DIOCKILLSTATESNV: {
2887 		error = pf_killstates_nv((struct pfioc_nv *)addr);
2888 		break;
2889 	}
2890 
2891 	case DIOCADDSTATE: {
2892 		struct pfioc_state	*ps = (struct pfioc_state *)addr;
2893 		struct pfsync_state	*sp = &ps->state;
2894 
2895 		if (sp->timeout >= PFTM_MAX) {
2896 			error = EINVAL;
2897 			break;
2898 		}
2899 		if (V_pfsync_state_import_ptr != NULL) {
2900 			PF_RULES_RLOCK();
2901 			error = V_pfsync_state_import_ptr(sp, PFSYNC_SI_IOCTL);
2902 			PF_RULES_RUNLOCK();
2903 		} else
2904 			error = EOPNOTSUPP;
2905 		break;
2906 	}
2907 
2908 	case DIOCGETSTATE: {
2909 		struct pfioc_state	*ps = (struct pfioc_state *)addr;
2910 		struct pf_kstate	*s;
2911 
2912 		s = pf_find_state_byid(ps->state.id, ps->state.creatorid);
2913 		if (s == NULL) {
2914 			error = ENOENT;
2915 			break;
2916 		}
2917 
2918 		pfsync_state_export(&ps->state, s);
2919 		PF_STATE_UNLOCK(s);
2920 		break;
2921 	}
2922 
2923 	case DIOCGETSTATENV: {
2924 		error = pf_getstate((struct pfioc_nv *)addr);
2925 		break;
2926 	}
2927 
2928 	case DIOCGETSTATES: {
2929 		struct pfioc_states	*ps = (struct pfioc_states *)addr;
2930 		struct pf_kstate	*s;
2931 		struct pfsync_state	*pstore, *p;
2932 		int			 i, nr;
2933 		size_t			 slice_count = 16, count;
2934 		void			*out;
2935 
2936 		if (ps->ps_len <= 0) {
2937 			nr = uma_zone_get_cur(V_pf_state_z);
2938 			ps->ps_len = sizeof(struct pfsync_state) * nr;
2939 			break;
2940 		}
2941 
2942 		out = ps->ps_states;
2943 		pstore = mallocarray(slice_count,
2944 		    sizeof(struct pfsync_state), M_TEMP, M_WAITOK | M_ZERO);
2945 		nr = 0;
2946 
2947 		for (i = 0; i <= pf_hashmask; i++) {
2948 			struct pf_idhash *ih = &V_pf_idhash[i];
2949 
2950 DIOCGETSTATES_retry:
2951 			p = pstore;
2952 
2953 			if (LIST_EMPTY(&ih->states))
2954 				continue;
2955 
2956 			PF_HASHROW_LOCK(ih);
2957 			count = 0;
2958 			LIST_FOREACH(s, &ih->states, entry) {
2959 				if (s->timeout == PFTM_UNLINKED)
2960 					continue;
2961 				count++;
2962 			}
2963 
2964 			if (count > slice_count) {
2965 				PF_HASHROW_UNLOCK(ih);
2966 				free(pstore, M_TEMP);
2967 				slice_count = count * 2;
2968 				pstore = mallocarray(slice_count,
2969 				    sizeof(struct pfsync_state), M_TEMP,
2970 				    M_WAITOK | M_ZERO);
2971 				goto DIOCGETSTATES_retry;
2972 			}
2973 
2974 			if ((nr+count) * sizeof(*p) > ps->ps_len) {
2975 				PF_HASHROW_UNLOCK(ih);
2976 				goto DIOCGETSTATES_full;
2977 			}
2978 
2979 			LIST_FOREACH(s, &ih->states, entry) {
2980 				if (s->timeout == PFTM_UNLINKED)
2981 					continue;
2982 
2983 				pfsync_state_export(p, s);
2984 				p++;
2985 				nr++;
2986 			}
2987 			PF_HASHROW_UNLOCK(ih);
2988 			error = copyout(pstore, out,
2989 			    sizeof(struct pfsync_state) * count);
2990 			if (error)
2991 				break;
2992 			out = ps->ps_states + nr;
2993 		}
2994 DIOCGETSTATES_full:
2995 		ps->ps_len = sizeof(struct pfsync_state) * nr;
2996 		free(pstore, M_TEMP);
2997 
2998 		break;
2999 	}
3000 
3001 	case DIOCGETSTATESV2: {
3002 		struct pfioc_states_v2	*ps = (struct pfioc_states_v2 *)addr;
3003 		struct pf_kstate	*s;
3004 		struct pf_state_export	*pstore, *p;
3005 		int i, nr;
3006 		size_t slice_count = 16, count;
3007 		void *out;
3008 
3009 		if (ps->ps_req_version > PF_STATE_VERSION) {
3010 			error = ENOTSUP;
3011 			break;
3012 		}
3013 
3014 		if (ps->ps_len <= 0) {
3015 			nr = uma_zone_get_cur(V_pf_state_z);
3016 			ps->ps_len = sizeof(struct pf_state_export) * nr;
3017 			break;
3018 		}
3019 
3020 		out = ps->ps_states;
3021 		pstore = mallocarray(slice_count,
3022 		    sizeof(struct pf_state_export), M_TEMP, M_WAITOK | M_ZERO);
3023 		nr = 0;
3024 
3025 		for (i = 0; i <= pf_hashmask; i++) {
3026 			struct pf_idhash *ih = &V_pf_idhash[i];
3027 
3028 DIOCGETSTATESV2_retry:
3029 			p = pstore;
3030 
3031 			if (LIST_EMPTY(&ih->states))
3032 				continue;
3033 
3034 			PF_HASHROW_LOCK(ih);
3035 			count = 0;
3036 			LIST_FOREACH(s, &ih->states, entry) {
3037 				if (s->timeout == PFTM_UNLINKED)
3038 					continue;
3039 				count++;
3040 			}
3041 
3042 			if (count > slice_count) {
3043 				PF_HASHROW_UNLOCK(ih);
3044 				free(pstore, M_TEMP);
3045 				slice_count = count * 2;
3046 				pstore = mallocarray(slice_count,
3047 				    sizeof(struct pf_state_export), M_TEMP,
3048 				    M_WAITOK | M_ZERO);
3049 				goto DIOCGETSTATESV2_retry;
3050 			}
3051 
3052 			if ((nr+count) * sizeof(*p) > ps->ps_len) {
3053 				PF_HASHROW_UNLOCK(ih);
3054 				goto DIOCGETSTATESV2_full;
3055 			}
3056 
3057 			LIST_FOREACH(s, &ih->states, entry) {
3058 				if (s->timeout == PFTM_UNLINKED)
3059 					continue;
3060 
3061 				pf_state_export(p, s);
3062 				p++;
3063 				nr++;
3064 			}
3065 			PF_HASHROW_UNLOCK(ih);
3066 			error = copyout(pstore, out,
3067 			    sizeof(struct pf_state_export) * count);
3068 			if (error)
3069 				break;
3070 			out = ps->ps_states + nr;
3071 		}
3072 DIOCGETSTATESV2_full:
3073 		ps->ps_len = nr * sizeof(struct pf_state_export);
3074 		free(pstore, M_TEMP);
3075 
3076 		break;
3077 	}
3078 
3079 	case DIOCGETSTATUS: {
3080 		struct pf_status *s = (struct pf_status *)addr;
3081 
3082 		PF_RULES_RLOCK();
3083 		s->running = V_pf_status.running;
3084 		s->since   = V_pf_status.since;
3085 		s->debug   = V_pf_status.debug;
3086 		s->hostid  = V_pf_status.hostid;
3087 		s->states  = V_pf_status.states;
3088 		s->src_nodes = V_pf_status.src_nodes;
3089 
3090 		for (int i = 0; i < PFRES_MAX; i++)
3091 			s->counters[i] =
3092 			    counter_u64_fetch(V_pf_status.counters[i]);
3093 		for (int i = 0; i < LCNT_MAX; i++)
3094 			s->lcounters[i] =
3095 			    counter_u64_fetch(V_pf_status.lcounters[i]);
3096 		for (int i = 0; i < FCNT_MAX; i++)
3097 			s->fcounters[i] =
3098 			    pf_counter_u64_fetch(&V_pf_status.fcounters[i]);
3099 		for (int i = 0; i < SCNT_MAX; i++)
3100 			s->scounters[i] =
3101 			    counter_u64_fetch(V_pf_status.scounters[i]);
3102 
3103 		bcopy(V_pf_status.ifname, s->ifname, IFNAMSIZ);
3104 		bcopy(V_pf_status.pf_chksum, s->pf_chksum,
3105 		    PF_MD5_DIGEST_LENGTH);
3106 
3107 		pfi_update_status(s->ifname, s);
3108 		PF_RULES_RUNLOCK();
3109 		break;
3110 	}
3111 
3112 	case DIOCGETSTATUSNV: {
3113 		error = pf_getstatus((struct pfioc_nv *)addr);
3114 		break;
3115 	}
3116 
3117 	case DIOCSETSTATUSIF: {
3118 		struct pfioc_if	*pi = (struct pfioc_if *)addr;
3119 
3120 		if (pi->ifname[0] == 0) {
3121 			bzero(V_pf_status.ifname, IFNAMSIZ);
3122 			break;
3123 		}
3124 		PF_RULES_WLOCK();
3125 		error = pf_user_strcpy(V_pf_status.ifname, pi->ifname, IFNAMSIZ);
3126 		PF_RULES_WUNLOCK();
3127 		break;
3128 	}
3129 
3130 	case DIOCCLRSTATUS: {
3131 		PF_RULES_WLOCK();
3132 		for (int i = 0; i < PFRES_MAX; i++)
3133 			counter_u64_zero(V_pf_status.counters[i]);
3134 		for (int i = 0; i < FCNT_MAX; i++)
3135 			pf_counter_u64_zero(&V_pf_status.fcounters[i]);
3136 		for (int i = 0; i < SCNT_MAX; i++)
3137 			counter_u64_zero(V_pf_status.scounters[i]);
3138 		for (int i = 0; i < KLCNT_MAX; i++)
3139 			counter_u64_zero(V_pf_status.lcounters[i]);
3140 		V_pf_status.since = time_second;
3141 		if (*V_pf_status.ifname)
3142 			pfi_update_status(V_pf_status.ifname, NULL);
3143 		PF_RULES_WUNLOCK();
3144 		break;
3145 	}
3146 
3147 	case DIOCNATLOOK: {
3148 		struct pfioc_natlook	*pnl = (struct pfioc_natlook *)addr;
3149 		struct pf_state_key	*sk;
3150 		struct pf_kstate	*state;
3151 		struct pf_state_key_cmp	 key;
3152 		int			 m = 0, direction = pnl->direction;
3153 		int			 sidx, didx;
3154 
3155 		/* NATLOOK src and dst are reversed, so reverse sidx/didx */
3156 		sidx = (direction == PF_IN) ? 1 : 0;
3157 		didx = (direction == PF_IN) ? 0 : 1;
3158 
3159 		if (!pnl->proto ||
3160 		    PF_AZERO(&pnl->saddr, pnl->af) ||
3161 		    PF_AZERO(&pnl->daddr, pnl->af) ||
3162 		    ((pnl->proto == IPPROTO_TCP ||
3163 		    pnl->proto == IPPROTO_UDP) &&
3164 		    (!pnl->dport || !pnl->sport)))
3165 			error = EINVAL;
3166 		else {
3167 			bzero(&key, sizeof(key));
3168 			key.af = pnl->af;
3169 			key.proto = pnl->proto;
3170 			PF_ACPY(&key.addr[sidx], &pnl->saddr, pnl->af);
3171 			key.port[sidx] = pnl->sport;
3172 			PF_ACPY(&key.addr[didx], &pnl->daddr, pnl->af);
3173 			key.port[didx] = pnl->dport;
3174 
3175 			state = pf_find_state_all(&key, direction, &m);
3176 
3177 			if (m > 1)
3178 				error = E2BIG;	/* more than one state */
3179 			else if (state != NULL) {
3180 				/* XXXGL: not locked read */
3181 				sk = state->key[sidx];
3182 				PF_ACPY(&pnl->rsaddr, &sk->addr[sidx], sk->af);
3183 				pnl->rsport = sk->port[sidx];
3184 				PF_ACPY(&pnl->rdaddr, &sk->addr[didx], sk->af);
3185 				pnl->rdport = sk->port[didx];
3186 			} else
3187 				error = ENOENT;
3188 		}
3189 		break;
3190 	}
3191 
3192 	case DIOCSETTIMEOUT: {
3193 		struct pfioc_tm	*pt = (struct pfioc_tm *)addr;
3194 		int		 old;
3195 
3196 		if (pt->timeout < 0 || pt->timeout >= PFTM_MAX ||
3197 		    pt->seconds < 0) {
3198 			error = EINVAL;
3199 			break;
3200 		}
3201 		PF_RULES_WLOCK();
3202 		old = V_pf_default_rule.timeout[pt->timeout];
3203 		if (pt->timeout == PFTM_INTERVAL && pt->seconds == 0)
3204 			pt->seconds = 1;
3205 		V_pf_default_rule.timeout[pt->timeout] = pt->seconds;
3206 		if (pt->timeout == PFTM_INTERVAL && pt->seconds < old)
3207 			wakeup(pf_purge_thread);
3208 		pt->seconds = old;
3209 		PF_RULES_WUNLOCK();
3210 		break;
3211 	}
3212 
3213 	case DIOCGETTIMEOUT: {
3214 		struct pfioc_tm	*pt = (struct pfioc_tm *)addr;
3215 
3216 		if (pt->timeout < 0 || pt->timeout >= PFTM_MAX) {
3217 			error = EINVAL;
3218 			break;
3219 		}
3220 		PF_RULES_RLOCK();
3221 		pt->seconds = V_pf_default_rule.timeout[pt->timeout];
3222 		PF_RULES_RUNLOCK();
3223 		break;
3224 	}
3225 
3226 	case DIOCGETLIMIT: {
3227 		struct pfioc_limit	*pl = (struct pfioc_limit *)addr;
3228 
3229 		if (pl->index < 0 || pl->index >= PF_LIMIT_MAX) {
3230 			error = EINVAL;
3231 			break;
3232 		}
3233 		PF_RULES_RLOCK();
3234 		pl->limit = V_pf_limits[pl->index].limit;
3235 		PF_RULES_RUNLOCK();
3236 		break;
3237 	}
3238 
3239 	case DIOCSETLIMIT: {
3240 		struct pfioc_limit	*pl = (struct pfioc_limit *)addr;
3241 		int			 old_limit;
3242 
3243 		PF_RULES_WLOCK();
3244 		if (pl->index < 0 || pl->index >= PF_LIMIT_MAX ||
3245 		    V_pf_limits[pl->index].zone == NULL) {
3246 			PF_RULES_WUNLOCK();
3247 			error = EINVAL;
3248 			break;
3249 		}
3250 		uma_zone_set_max(V_pf_limits[pl->index].zone, pl->limit);
3251 		old_limit = V_pf_limits[pl->index].limit;
3252 		V_pf_limits[pl->index].limit = pl->limit;
3253 		pl->limit = old_limit;
3254 		PF_RULES_WUNLOCK();
3255 		break;
3256 	}
3257 
3258 	case DIOCSETDEBUG: {
3259 		u_int32_t	*level = (u_int32_t *)addr;
3260 
3261 		PF_RULES_WLOCK();
3262 		V_pf_status.debug = *level;
3263 		PF_RULES_WUNLOCK();
3264 		break;
3265 	}
3266 
3267 	case DIOCCLRRULECTRS: {
3268 		/* obsoleted by DIOCGETRULE with action=PF_GET_CLR_CNTR */
3269 		struct pf_kruleset	*ruleset = &pf_main_ruleset;
3270 		struct pf_krule		*rule;
3271 
3272 		PF_RULES_WLOCK();
3273 		TAILQ_FOREACH(rule,
3274 		    ruleset->rules[PF_RULESET_FILTER].active.ptr, entries) {
3275 			pf_counter_u64_zero(&rule->evaluations);
3276 			for (int i = 0; i < 2; i++) {
3277 				pf_counter_u64_zero(&rule->packets[i]);
3278 				pf_counter_u64_zero(&rule->bytes[i]);
3279 			}
3280 		}
3281 		PF_RULES_WUNLOCK();
3282 		break;
3283 	}
3284 
3285 	case DIOCGIFSPEEDV0:
3286 	case DIOCGIFSPEEDV1: {
3287 		struct pf_ifspeed_v1	*psp = (struct pf_ifspeed_v1 *)addr;
3288 		struct pf_ifspeed_v1	ps;
3289 		struct ifnet		*ifp;
3290 
3291 		if (psp->ifname[0] == '\0') {
3292 			error = EINVAL;
3293 			break;
3294 		}
3295 
3296 		error = pf_user_strcpy(ps.ifname, psp->ifname, IFNAMSIZ);
3297 		if (error != 0)
3298 			break;
3299 		ifp = ifunit(ps.ifname);
3300 		if (ifp != NULL) {
3301 			psp->baudrate32 =
3302 			    (u_int32_t)uqmin(ifp->if_baudrate, UINT_MAX);
3303 			if (cmd == DIOCGIFSPEEDV1)
3304 				psp->baudrate = ifp->if_baudrate;
3305 		} else {
3306 			error = EINVAL;
3307 		}
3308 		break;
3309 	}
3310 
3311 #ifdef ALTQ
3312 	case DIOCSTARTALTQ: {
3313 		struct pf_altq		*altq;
3314 
3315 		PF_RULES_WLOCK();
3316 		/* enable all altq interfaces on active list */
3317 		TAILQ_FOREACH(altq, V_pf_altq_ifs_active, entries) {
3318 			if ((altq->local_flags & PFALTQ_FLAG_IF_REMOVED) == 0) {
3319 				error = pf_enable_altq(altq);
3320 				if (error != 0)
3321 					break;
3322 			}
3323 		}
3324 		if (error == 0)
3325 			V_pf_altq_running = 1;
3326 		PF_RULES_WUNLOCK();
3327 		DPFPRINTF(PF_DEBUG_MISC, ("altq: started\n"));
3328 		break;
3329 	}
3330 
3331 	case DIOCSTOPALTQ: {
3332 		struct pf_altq		*altq;
3333 
3334 		PF_RULES_WLOCK();
3335 		/* disable all altq interfaces on active list */
3336 		TAILQ_FOREACH(altq, V_pf_altq_ifs_active, entries) {
3337 			if ((altq->local_flags & PFALTQ_FLAG_IF_REMOVED) == 0) {
3338 				error = pf_disable_altq(altq);
3339 				if (error != 0)
3340 					break;
3341 			}
3342 		}
3343 		if (error == 0)
3344 			V_pf_altq_running = 0;
3345 		PF_RULES_WUNLOCK();
3346 		DPFPRINTF(PF_DEBUG_MISC, ("altq: stopped\n"));
3347 		break;
3348 	}
3349 
3350 	case DIOCADDALTQV0:
3351 	case DIOCADDALTQV1: {
3352 		struct pfioc_altq_v1	*pa = (struct pfioc_altq_v1 *)addr;
3353 		struct pf_altq		*altq, *a;
3354 		struct ifnet		*ifp;
3355 
3356 		altq = malloc(sizeof(*altq), M_PFALTQ, M_WAITOK | M_ZERO);
3357 		error = pf_import_kaltq(pa, altq, IOCPARM_LEN(cmd));
3358 		if (error)
3359 			break;
3360 		altq->local_flags = 0;
3361 
3362 		PF_RULES_WLOCK();
3363 		if (pa->ticket != V_ticket_altqs_inactive) {
3364 			PF_RULES_WUNLOCK();
3365 			free(altq, M_PFALTQ);
3366 			error = EBUSY;
3367 			break;
3368 		}
3369 
3370 		/*
3371 		 * if this is for a queue, find the discipline and
3372 		 * copy the necessary fields
3373 		 */
3374 		if (altq->qname[0] != 0) {
3375 			if ((altq->qid = pf_qname2qid(altq->qname)) == 0) {
3376 				PF_RULES_WUNLOCK();
3377 				error = EBUSY;
3378 				free(altq, M_PFALTQ);
3379 				break;
3380 			}
3381 			altq->altq_disc = NULL;
3382 			TAILQ_FOREACH(a, V_pf_altq_ifs_inactive, entries) {
3383 				if (strncmp(a->ifname, altq->ifname,
3384 				    IFNAMSIZ) == 0) {
3385 					altq->altq_disc = a->altq_disc;
3386 					break;
3387 				}
3388 			}
3389 		}
3390 
3391 		if ((ifp = ifunit(altq->ifname)) == NULL)
3392 			altq->local_flags |= PFALTQ_FLAG_IF_REMOVED;
3393 		else
3394 			error = altq_add(ifp, altq);
3395 
3396 		if (error) {
3397 			PF_RULES_WUNLOCK();
3398 			free(altq, M_PFALTQ);
3399 			break;
3400 		}
3401 
3402 		if (altq->qname[0] != 0)
3403 			TAILQ_INSERT_TAIL(V_pf_altqs_inactive, altq, entries);
3404 		else
3405 			TAILQ_INSERT_TAIL(V_pf_altq_ifs_inactive, altq, entries);
3406 		/* version error check done on import above */
3407 		pf_export_kaltq(altq, pa, IOCPARM_LEN(cmd));
3408 		PF_RULES_WUNLOCK();
3409 		break;
3410 	}
3411 
3412 	case DIOCGETALTQSV0:
3413 	case DIOCGETALTQSV1: {
3414 		struct pfioc_altq_v1	*pa = (struct pfioc_altq_v1 *)addr;
3415 		struct pf_altq		*altq;
3416 
3417 		PF_RULES_RLOCK();
3418 		pa->nr = 0;
3419 		TAILQ_FOREACH(altq, V_pf_altq_ifs_active, entries)
3420 			pa->nr++;
3421 		TAILQ_FOREACH(altq, V_pf_altqs_active, entries)
3422 			pa->nr++;
3423 		pa->ticket = V_ticket_altqs_active;
3424 		PF_RULES_RUNLOCK();
3425 		break;
3426 	}
3427 
3428 	case DIOCGETALTQV0:
3429 	case DIOCGETALTQV1: {
3430 		struct pfioc_altq_v1	*pa = (struct pfioc_altq_v1 *)addr;
3431 		struct pf_altq		*altq;
3432 
3433 		PF_RULES_RLOCK();
3434 		if (pa->ticket != V_ticket_altqs_active) {
3435 			PF_RULES_RUNLOCK();
3436 			error = EBUSY;
3437 			break;
3438 		}
3439 		altq = pf_altq_get_nth_active(pa->nr);
3440 		if (altq == NULL) {
3441 			PF_RULES_RUNLOCK();
3442 			error = EBUSY;
3443 			break;
3444 		}
3445 		pf_export_kaltq(altq, pa, IOCPARM_LEN(cmd));
3446 		PF_RULES_RUNLOCK();
3447 		break;
3448 	}
3449 
3450 	case DIOCCHANGEALTQV0:
3451 	case DIOCCHANGEALTQV1:
3452 		/* CHANGEALTQ not supported yet! */
3453 		error = ENODEV;
3454 		break;
3455 
3456 	case DIOCGETQSTATSV0:
3457 	case DIOCGETQSTATSV1: {
3458 		struct pfioc_qstats_v1	*pq = (struct pfioc_qstats_v1 *)addr;
3459 		struct pf_altq		*altq;
3460 		int			 nbytes;
3461 		u_int32_t		 version;
3462 
3463 		PF_RULES_RLOCK();
3464 		if (pq->ticket != V_ticket_altqs_active) {
3465 			PF_RULES_RUNLOCK();
3466 			error = EBUSY;
3467 			break;
3468 		}
3469 		nbytes = pq->nbytes;
3470 		altq = pf_altq_get_nth_active(pq->nr);
3471 		if (altq == NULL) {
3472 			PF_RULES_RUNLOCK();
3473 			error = EBUSY;
3474 			break;
3475 		}
3476 
3477 		if ((altq->local_flags & PFALTQ_FLAG_IF_REMOVED) != 0) {
3478 			PF_RULES_RUNLOCK();
3479 			error = ENXIO;
3480 			break;
3481 		}
3482 		PF_RULES_RUNLOCK();
3483 		if (cmd == DIOCGETQSTATSV0)
3484 			version = 0;  /* DIOCGETQSTATSV0 means stats struct v0 */
3485 		else
3486 			version = pq->version;
3487 		error = altq_getqstats(altq, pq->buf, &nbytes, version);
3488 		if (error == 0) {
3489 			pq->scheduler = altq->scheduler;
3490 			pq->nbytes = nbytes;
3491 		}
3492 		break;
3493 	}
3494 #endif /* ALTQ */
3495 
3496 	case DIOCBEGINADDRS: {
3497 		struct pfioc_pooladdr	*pp = (struct pfioc_pooladdr *)addr;
3498 
3499 		PF_RULES_WLOCK();
3500 		pf_empty_kpool(&V_pf_pabuf);
3501 		pp->ticket = ++V_ticket_pabuf;
3502 		PF_RULES_WUNLOCK();
3503 		break;
3504 	}
3505 
3506 	case DIOCADDADDR: {
3507 		struct pfioc_pooladdr	*pp = (struct pfioc_pooladdr *)addr;
3508 		struct pf_kpooladdr	*pa;
3509 		struct pfi_kkif		*kif = NULL;
3510 
3511 #ifndef INET
3512 		if (pp->af == AF_INET) {
3513 			error = EAFNOSUPPORT;
3514 			break;
3515 		}
3516 #endif /* INET */
3517 #ifndef INET6
3518 		if (pp->af == AF_INET6) {
3519 			error = EAFNOSUPPORT;
3520 			break;
3521 		}
3522 #endif /* INET6 */
3523 		if (pp->addr.addr.type != PF_ADDR_ADDRMASK &&
3524 		    pp->addr.addr.type != PF_ADDR_DYNIFTL &&
3525 		    pp->addr.addr.type != PF_ADDR_TABLE) {
3526 			error = EINVAL;
3527 			break;
3528 		}
3529 		if (pp->addr.addr.p.dyn != NULL) {
3530 			error = EINVAL;
3531 			break;
3532 		}
3533 		pa = malloc(sizeof(*pa), M_PFRULE, M_WAITOK);
3534 		error = pf_pooladdr_to_kpooladdr(&pp->addr, pa);
3535 		if (error != 0)
3536 			break;
3537 		if (pa->ifname[0])
3538 			kif = pf_kkif_create(M_WAITOK);
3539 		PF_RULES_WLOCK();
3540 		if (pp->ticket != V_ticket_pabuf) {
3541 			PF_RULES_WUNLOCK();
3542 			if (pa->ifname[0])
3543 				pf_kkif_free(kif);
3544 			free(pa, M_PFRULE);
3545 			error = EBUSY;
3546 			break;
3547 		}
3548 		if (pa->ifname[0]) {
3549 			pa->kif = pfi_kkif_attach(kif, pa->ifname);
3550 			kif = NULL;
3551 			pfi_kkif_ref(pa->kif);
3552 		} else
3553 			pa->kif = NULL;
3554 		if (pa->addr.type == PF_ADDR_DYNIFTL && ((error =
3555 		    pfi_dynaddr_setup(&pa->addr, pp->af)) != 0)) {
3556 			if (pa->ifname[0])
3557 				pfi_kkif_unref(pa->kif);
3558 			PF_RULES_WUNLOCK();
3559 			free(pa, M_PFRULE);
3560 			break;
3561 		}
3562 		TAILQ_INSERT_TAIL(&V_pf_pabuf, pa, entries);
3563 		PF_RULES_WUNLOCK();
3564 		break;
3565 	}
3566 
3567 	case DIOCGETADDRS: {
3568 		struct pfioc_pooladdr	*pp = (struct pfioc_pooladdr *)addr;
3569 		struct pf_kpool		*pool;
3570 		struct pf_kpooladdr	*pa;
3571 
3572 		pp->anchor[sizeof(pp->anchor) - 1] = 0;
3573 		pp->nr = 0;
3574 
3575 		PF_RULES_RLOCK();
3576 		pool = pf_get_kpool(pp->anchor, pp->ticket, pp->r_action,
3577 		    pp->r_num, 0, 1, 0);
3578 		if (pool == NULL) {
3579 			PF_RULES_RUNLOCK();
3580 			error = EBUSY;
3581 			break;
3582 		}
3583 		TAILQ_FOREACH(pa, &pool->list, entries)
3584 			pp->nr++;
3585 		PF_RULES_RUNLOCK();
3586 		break;
3587 	}
3588 
3589 	case DIOCGETADDR: {
3590 		struct pfioc_pooladdr	*pp = (struct pfioc_pooladdr *)addr;
3591 		struct pf_kpool		*pool;
3592 		struct pf_kpooladdr	*pa;
3593 		u_int32_t		 nr = 0;
3594 
3595 		pp->anchor[sizeof(pp->anchor) - 1] = 0;
3596 
3597 		PF_RULES_RLOCK();
3598 		pool = pf_get_kpool(pp->anchor, pp->ticket, pp->r_action,
3599 		    pp->r_num, 0, 1, 1);
3600 		if (pool == NULL) {
3601 			PF_RULES_RUNLOCK();
3602 			error = EBUSY;
3603 			break;
3604 		}
3605 		pa = TAILQ_FIRST(&pool->list);
3606 		while ((pa != NULL) && (nr < pp->nr)) {
3607 			pa = TAILQ_NEXT(pa, entries);
3608 			nr++;
3609 		}
3610 		if (pa == NULL) {
3611 			PF_RULES_RUNLOCK();
3612 			error = EBUSY;
3613 			break;
3614 		}
3615 		pf_kpooladdr_to_pooladdr(pa, &pp->addr);
3616 		pf_addr_copyout(&pp->addr.addr);
3617 		PF_RULES_RUNLOCK();
3618 		break;
3619 	}
3620 
3621 	case DIOCCHANGEADDR: {
3622 		struct pfioc_pooladdr	*pca = (struct pfioc_pooladdr *)addr;
3623 		struct pf_kpool		*pool;
3624 		struct pf_kpooladdr	*oldpa = NULL, *newpa = NULL;
3625 		struct pf_kruleset	*ruleset;
3626 		struct pfi_kkif		*kif = NULL;
3627 
3628 		pca->anchor[sizeof(pca->anchor) - 1] = 0;
3629 
3630 		if (pca->action < PF_CHANGE_ADD_HEAD ||
3631 		    pca->action > PF_CHANGE_REMOVE) {
3632 			error = EINVAL;
3633 			break;
3634 		}
3635 		if (pca->addr.addr.type != PF_ADDR_ADDRMASK &&
3636 		    pca->addr.addr.type != PF_ADDR_DYNIFTL &&
3637 		    pca->addr.addr.type != PF_ADDR_TABLE) {
3638 			error = EINVAL;
3639 			break;
3640 		}
3641 		if (pca->addr.addr.p.dyn != NULL) {
3642 			error = EINVAL;
3643 			break;
3644 		}
3645 
3646 		if (pca->action != PF_CHANGE_REMOVE) {
3647 #ifndef INET
3648 			if (pca->af == AF_INET) {
3649 				error = EAFNOSUPPORT;
3650 				break;
3651 			}
3652 #endif /* INET */
3653 #ifndef INET6
3654 			if (pca->af == AF_INET6) {
3655 				error = EAFNOSUPPORT;
3656 				break;
3657 			}
3658 #endif /* INET6 */
3659 			newpa = malloc(sizeof(*newpa), M_PFRULE, M_WAITOK);
3660 			bcopy(&pca->addr, newpa, sizeof(struct pf_pooladdr));
3661 			if (newpa->ifname[0])
3662 				kif = pf_kkif_create(M_WAITOK);
3663 			newpa->kif = NULL;
3664 		}
3665 #define	ERROUT(x)	ERROUT_IOCTL(DIOCCHANGEADDR_error, x)
3666 		PF_RULES_WLOCK();
3667 		ruleset = pf_find_kruleset(pca->anchor);
3668 		if (ruleset == NULL)
3669 			ERROUT(EBUSY);
3670 
3671 		pool = pf_get_kpool(pca->anchor, pca->ticket, pca->r_action,
3672 		    pca->r_num, pca->r_last, 1, 1);
3673 		if (pool == NULL)
3674 			ERROUT(EBUSY);
3675 
3676 		if (pca->action != PF_CHANGE_REMOVE) {
3677 			if (newpa->ifname[0]) {
3678 				newpa->kif = pfi_kkif_attach(kif, newpa->ifname);
3679 				pfi_kkif_ref(newpa->kif);
3680 				kif = NULL;
3681 			}
3682 
3683 			switch (newpa->addr.type) {
3684 			case PF_ADDR_DYNIFTL:
3685 				error = pfi_dynaddr_setup(&newpa->addr,
3686 				    pca->af);
3687 				break;
3688 			case PF_ADDR_TABLE:
3689 				newpa->addr.p.tbl = pfr_attach_table(ruleset,
3690 				    newpa->addr.v.tblname);
3691 				if (newpa->addr.p.tbl == NULL)
3692 					error = ENOMEM;
3693 				break;
3694 			}
3695 			if (error)
3696 				goto DIOCCHANGEADDR_error;
3697 		}
3698 
3699 		switch (pca->action) {
3700 		case PF_CHANGE_ADD_HEAD:
3701 			oldpa = TAILQ_FIRST(&pool->list);
3702 			break;
3703 		case PF_CHANGE_ADD_TAIL:
3704 			oldpa = TAILQ_LAST(&pool->list, pf_kpalist);
3705 			break;
3706 		default:
3707 			oldpa = TAILQ_FIRST(&pool->list);
3708 			for (int i = 0; oldpa && i < pca->nr; i++)
3709 				oldpa = TAILQ_NEXT(oldpa, entries);
3710 
3711 			if (oldpa == NULL)
3712 				ERROUT(EINVAL);
3713 		}
3714 
3715 		if (pca->action == PF_CHANGE_REMOVE) {
3716 			TAILQ_REMOVE(&pool->list, oldpa, entries);
3717 			switch (oldpa->addr.type) {
3718 			case PF_ADDR_DYNIFTL:
3719 				pfi_dynaddr_remove(oldpa->addr.p.dyn);
3720 				break;
3721 			case PF_ADDR_TABLE:
3722 				pfr_detach_table(oldpa->addr.p.tbl);
3723 				break;
3724 			}
3725 			if (oldpa->kif)
3726 				pfi_kkif_unref(oldpa->kif);
3727 			free(oldpa, M_PFRULE);
3728 		} else {
3729 			if (oldpa == NULL)
3730 				TAILQ_INSERT_TAIL(&pool->list, newpa, entries);
3731 			else if (pca->action == PF_CHANGE_ADD_HEAD ||
3732 			    pca->action == PF_CHANGE_ADD_BEFORE)
3733 				TAILQ_INSERT_BEFORE(oldpa, newpa, entries);
3734 			else
3735 				TAILQ_INSERT_AFTER(&pool->list, oldpa,
3736 				    newpa, entries);
3737 		}
3738 
3739 		pool->cur = TAILQ_FIRST(&pool->list);
3740 		PF_ACPY(&pool->counter, &pool->cur->addr.v.a.addr, pca->af);
3741 		PF_RULES_WUNLOCK();
3742 		break;
3743 
3744 #undef ERROUT
3745 DIOCCHANGEADDR_error:
3746 		if (newpa != NULL) {
3747 			if (newpa->kif)
3748 				pfi_kkif_unref(newpa->kif);
3749 			free(newpa, M_PFRULE);
3750 		}
3751 		PF_RULES_WUNLOCK();
3752 		pf_kkif_free(kif);
3753 		break;
3754 	}
3755 
3756 	case DIOCGETRULESETS: {
3757 		struct pfioc_ruleset	*pr = (struct pfioc_ruleset *)addr;
3758 		struct pf_kruleset	*ruleset;
3759 		struct pf_kanchor	*anchor;
3760 
3761 		pr->path[sizeof(pr->path) - 1] = 0;
3762 
3763 		PF_RULES_RLOCK();
3764 		if ((ruleset = pf_find_kruleset(pr->path)) == NULL) {
3765 			PF_RULES_RUNLOCK();
3766 			error = ENOENT;
3767 			break;
3768 		}
3769 		pr->nr = 0;
3770 		if (ruleset->anchor == NULL) {
3771 			/* XXX kludge for pf_main_ruleset */
3772 			RB_FOREACH(anchor, pf_kanchor_global, &V_pf_anchors)
3773 				if (anchor->parent == NULL)
3774 					pr->nr++;
3775 		} else {
3776 			RB_FOREACH(anchor, pf_kanchor_node,
3777 			    &ruleset->anchor->children)
3778 				pr->nr++;
3779 		}
3780 		PF_RULES_RUNLOCK();
3781 		break;
3782 	}
3783 
3784 	case DIOCGETRULESET: {
3785 		struct pfioc_ruleset	*pr = (struct pfioc_ruleset *)addr;
3786 		struct pf_kruleset	*ruleset;
3787 		struct pf_kanchor	*anchor;
3788 		u_int32_t		 nr = 0;
3789 
3790 		pr->path[sizeof(pr->path) - 1] = 0;
3791 
3792 		PF_RULES_RLOCK();
3793 		if ((ruleset = pf_find_kruleset(pr->path)) == NULL) {
3794 			PF_RULES_RUNLOCK();
3795 			error = ENOENT;
3796 			break;
3797 		}
3798 		pr->name[0] = 0;
3799 		if (ruleset->anchor == NULL) {
3800 			/* XXX kludge for pf_main_ruleset */
3801 			RB_FOREACH(anchor, pf_kanchor_global, &V_pf_anchors)
3802 				if (anchor->parent == NULL && nr++ == pr->nr) {
3803 					strlcpy(pr->name, anchor->name,
3804 					    sizeof(pr->name));
3805 					break;
3806 				}
3807 		} else {
3808 			RB_FOREACH(anchor, pf_kanchor_node,
3809 			    &ruleset->anchor->children)
3810 				if (nr++ == pr->nr) {
3811 					strlcpy(pr->name, anchor->name,
3812 					    sizeof(pr->name));
3813 					break;
3814 				}
3815 		}
3816 		if (!pr->name[0])
3817 			error = EBUSY;
3818 		PF_RULES_RUNLOCK();
3819 		break;
3820 	}
3821 
3822 	case DIOCRCLRTABLES: {
3823 		struct pfioc_table *io = (struct pfioc_table *)addr;
3824 
3825 		if (io->pfrio_esize != 0) {
3826 			error = ENODEV;
3827 			break;
3828 		}
3829 		PF_RULES_WLOCK();
3830 		error = pfr_clr_tables(&io->pfrio_table, &io->pfrio_ndel,
3831 		    io->pfrio_flags | PFR_FLAG_USERIOCTL);
3832 		PF_RULES_WUNLOCK();
3833 		break;
3834 	}
3835 
3836 	case DIOCRADDTABLES: {
3837 		struct pfioc_table *io = (struct pfioc_table *)addr;
3838 		struct pfr_table *pfrts;
3839 		size_t totlen;
3840 
3841 		if (io->pfrio_esize != sizeof(struct pfr_table)) {
3842 			error = ENODEV;
3843 			break;
3844 		}
3845 
3846 		if (io->pfrio_size < 0 || io->pfrio_size > pf_ioctl_maxcount ||
3847 		    WOULD_OVERFLOW(io->pfrio_size, sizeof(struct pfr_table))) {
3848 			error = ENOMEM;
3849 			break;
3850 		}
3851 
3852 		totlen = io->pfrio_size * sizeof(struct pfr_table);
3853 		pfrts = mallocarray(io->pfrio_size, sizeof(struct pfr_table),
3854 		    M_TEMP, M_WAITOK);
3855 		error = copyin(io->pfrio_buffer, pfrts, totlen);
3856 		if (error) {
3857 			free(pfrts, M_TEMP);
3858 			break;
3859 		}
3860 		PF_RULES_WLOCK();
3861 		error = pfr_add_tables(pfrts, io->pfrio_size,
3862 		    &io->pfrio_nadd, io->pfrio_flags | PFR_FLAG_USERIOCTL);
3863 		PF_RULES_WUNLOCK();
3864 		free(pfrts, M_TEMP);
3865 		break;
3866 	}
3867 
3868 	case DIOCRDELTABLES: {
3869 		struct pfioc_table *io = (struct pfioc_table *)addr;
3870 		struct pfr_table *pfrts;
3871 		size_t totlen;
3872 
3873 		if (io->pfrio_esize != sizeof(struct pfr_table)) {
3874 			error = ENODEV;
3875 			break;
3876 		}
3877 
3878 		if (io->pfrio_size < 0 || io->pfrio_size > pf_ioctl_maxcount ||
3879 		    WOULD_OVERFLOW(io->pfrio_size, sizeof(struct pfr_table))) {
3880 			error = ENOMEM;
3881 			break;
3882 		}
3883 
3884 		totlen = io->pfrio_size * sizeof(struct pfr_table);
3885 		pfrts = mallocarray(io->pfrio_size, sizeof(struct pfr_table),
3886 		    M_TEMP, M_WAITOK);
3887 		error = copyin(io->pfrio_buffer, pfrts, totlen);
3888 		if (error) {
3889 			free(pfrts, M_TEMP);
3890 			break;
3891 		}
3892 		PF_RULES_WLOCK();
3893 		error = pfr_del_tables(pfrts, io->pfrio_size,
3894 		    &io->pfrio_ndel, io->pfrio_flags | PFR_FLAG_USERIOCTL);
3895 		PF_RULES_WUNLOCK();
3896 		free(pfrts, M_TEMP);
3897 		break;
3898 	}
3899 
3900 	case DIOCRGETTABLES: {
3901 		struct pfioc_table *io = (struct pfioc_table *)addr;
3902 		struct pfr_table *pfrts;
3903 		size_t totlen;
3904 		int n;
3905 
3906 		if (io->pfrio_esize != sizeof(struct pfr_table)) {
3907 			error = ENODEV;
3908 			break;
3909 		}
3910 		PF_RULES_RLOCK();
3911 		n = pfr_table_count(&io->pfrio_table, io->pfrio_flags);
3912 		if (n < 0) {
3913 			PF_RULES_RUNLOCK();
3914 			error = EINVAL;
3915 			break;
3916 		}
3917 		io->pfrio_size = min(io->pfrio_size, n);
3918 
3919 		totlen = io->pfrio_size * sizeof(struct pfr_table);
3920 
3921 		pfrts = mallocarray(io->pfrio_size, sizeof(struct pfr_table),
3922 		    M_TEMP, M_NOWAIT | M_ZERO);
3923 		if (pfrts == NULL) {
3924 			error = ENOMEM;
3925 			PF_RULES_RUNLOCK();
3926 			break;
3927 		}
3928 		error = pfr_get_tables(&io->pfrio_table, pfrts,
3929 		    &io->pfrio_size, io->pfrio_flags | PFR_FLAG_USERIOCTL);
3930 		PF_RULES_RUNLOCK();
3931 		if (error == 0)
3932 			error = copyout(pfrts, io->pfrio_buffer, totlen);
3933 		free(pfrts, M_TEMP);
3934 		break;
3935 	}
3936 
3937 	case DIOCRGETTSTATS: {
3938 		struct pfioc_table *io = (struct pfioc_table *)addr;
3939 		struct pfr_tstats *pfrtstats;
3940 		size_t totlen;
3941 		int n;
3942 
3943 		if (io->pfrio_esize != sizeof(struct pfr_tstats)) {
3944 			error = ENODEV;
3945 			break;
3946 		}
3947 		PF_TABLE_STATS_LOCK();
3948 		PF_RULES_RLOCK();
3949 		n = pfr_table_count(&io->pfrio_table, io->pfrio_flags);
3950 		if (n < 0) {
3951 			PF_RULES_RUNLOCK();
3952 			PF_TABLE_STATS_UNLOCK();
3953 			error = EINVAL;
3954 			break;
3955 		}
3956 		io->pfrio_size = min(io->pfrio_size, n);
3957 
3958 		totlen = io->pfrio_size * sizeof(struct pfr_tstats);
3959 		pfrtstats = mallocarray(io->pfrio_size,
3960 		    sizeof(struct pfr_tstats), M_TEMP, M_NOWAIT | M_ZERO);
3961 		if (pfrtstats == NULL) {
3962 			error = ENOMEM;
3963 			PF_RULES_RUNLOCK();
3964 			PF_TABLE_STATS_UNLOCK();
3965 			break;
3966 		}
3967 		error = pfr_get_tstats(&io->pfrio_table, pfrtstats,
3968 		    &io->pfrio_size, io->pfrio_flags | PFR_FLAG_USERIOCTL);
3969 		PF_RULES_RUNLOCK();
3970 		PF_TABLE_STATS_UNLOCK();
3971 		if (error == 0)
3972 			error = copyout(pfrtstats, io->pfrio_buffer, totlen);
3973 		free(pfrtstats, M_TEMP);
3974 		break;
3975 	}
3976 
3977 	case DIOCRCLRTSTATS: {
3978 		struct pfioc_table *io = (struct pfioc_table *)addr;
3979 		struct pfr_table *pfrts;
3980 		size_t totlen;
3981 
3982 		if (io->pfrio_esize != sizeof(struct pfr_table)) {
3983 			error = ENODEV;
3984 			break;
3985 		}
3986 
3987 		if (io->pfrio_size < 0 || io->pfrio_size > pf_ioctl_maxcount ||
3988 		    WOULD_OVERFLOW(io->pfrio_size, sizeof(struct pfr_table))) {
3989 			/* We used to count tables and use the minimum required
3990 			 * size, so we didn't fail on overly large requests.
3991 			 * Keep doing so. */
3992 			io->pfrio_size = pf_ioctl_maxcount;
3993 			break;
3994 		}
3995 
3996 		totlen = io->pfrio_size * sizeof(struct pfr_table);
3997 		pfrts = mallocarray(io->pfrio_size, sizeof(struct pfr_table),
3998 		    M_TEMP, M_WAITOK);
3999 		error = copyin(io->pfrio_buffer, pfrts, totlen);
4000 		if (error) {
4001 			free(pfrts, M_TEMP);
4002 			break;
4003 		}
4004 
4005 		PF_TABLE_STATS_LOCK();
4006 		PF_RULES_RLOCK();
4007 		error = pfr_clr_tstats(pfrts, io->pfrio_size,
4008 		    &io->pfrio_nzero, io->pfrio_flags | PFR_FLAG_USERIOCTL);
4009 		PF_RULES_RUNLOCK();
4010 		PF_TABLE_STATS_UNLOCK();
4011 		free(pfrts, M_TEMP);
4012 		break;
4013 	}
4014 
4015 	case DIOCRSETTFLAGS: {
4016 		struct pfioc_table *io = (struct pfioc_table *)addr;
4017 		struct pfr_table *pfrts;
4018 		size_t totlen;
4019 		int n;
4020 
4021 		if (io->pfrio_esize != sizeof(struct pfr_table)) {
4022 			error = ENODEV;
4023 			break;
4024 		}
4025 
4026 		PF_RULES_RLOCK();
4027 		n = pfr_table_count(&io->pfrio_table, io->pfrio_flags);
4028 		if (n < 0) {
4029 			PF_RULES_RUNLOCK();
4030 			error = EINVAL;
4031 			break;
4032 		}
4033 
4034 		io->pfrio_size = min(io->pfrio_size, n);
4035 		PF_RULES_RUNLOCK();
4036 
4037 		totlen = io->pfrio_size * sizeof(struct pfr_table);
4038 		pfrts = mallocarray(io->pfrio_size, sizeof(struct pfr_table),
4039 		    M_TEMP, M_WAITOK);
4040 		error = copyin(io->pfrio_buffer, pfrts, totlen);
4041 		if (error) {
4042 			free(pfrts, M_TEMP);
4043 			break;
4044 		}
4045 		PF_RULES_WLOCK();
4046 		error = pfr_set_tflags(pfrts, io->pfrio_size,
4047 		    io->pfrio_setflag, io->pfrio_clrflag, &io->pfrio_nchange,
4048 		    &io->pfrio_ndel, io->pfrio_flags | PFR_FLAG_USERIOCTL);
4049 		PF_RULES_WUNLOCK();
4050 		free(pfrts, M_TEMP);
4051 		break;
4052 	}
4053 
4054 	case DIOCRCLRADDRS: {
4055 		struct pfioc_table *io = (struct pfioc_table *)addr;
4056 
4057 		if (io->pfrio_esize != 0) {
4058 			error = ENODEV;
4059 			break;
4060 		}
4061 		PF_RULES_WLOCK();
4062 		error = pfr_clr_addrs(&io->pfrio_table, &io->pfrio_ndel,
4063 		    io->pfrio_flags | PFR_FLAG_USERIOCTL);
4064 		PF_RULES_WUNLOCK();
4065 		break;
4066 	}
4067 
4068 	case DIOCRADDADDRS: {
4069 		struct pfioc_table *io = (struct pfioc_table *)addr;
4070 		struct pfr_addr *pfras;
4071 		size_t totlen;
4072 
4073 		if (io->pfrio_esize != sizeof(struct pfr_addr)) {
4074 			error = ENODEV;
4075 			break;
4076 		}
4077 		if (io->pfrio_size < 0 ||
4078 		    io->pfrio_size > pf_ioctl_maxcount ||
4079 		    WOULD_OVERFLOW(io->pfrio_size, sizeof(struct pfr_addr))) {
4080 			error = EINVAL;
4081 			break;
4082 		}
4083 		totlen = io->pfrio_size * sizeof(struct pfr_addr);
4084 		pfras = mallocarray(io->pfrio_size, sizeof(struct pfr_addr),
4085 		    M_TEMP, M_WAITOK);
4086 		error = copyin(io->pfrio_buffer, pfras, totlen);
4087 		if (error) {
4088 			free(pfras, M_TEMP);
4089 			break;
4090 		}
4091 		PF_RULES_WLOCK();
4092 		error = pfr_add_addrs(&io->pfrio_table, pfras,
4093 		    io->pfrio_size, &io->pfrio_nadd, io->pfrio_flags |
4094 		    PFR_FLAG_USERIOCTL);
4095 		PF_RULES_WUNLOCK();
4096 		if (error == 0 && io->pfrio_flags & PFR_FLAG_FEEDBACK)
4097 			error = copyout(pfras, io->pfrio_buffer, totlen);
4098 		free(pfras, M_TEMP);
4099 		break;
4100 	}
4101 
4102 	case DIOCRDELADDRS: {
4103 		struct pfioc_table *io = (struct pfioc_table *)addr;
4104 		struct pfr_addr *pfras;
4105 		size_t totlen;
4106 
4107 		if (io->pfrio_esize != sizeof(struct pfr_addr)) {
4108 			error = ENODEV;
4109 			break;
4110 		}
4111 		if (io->pfrio_size < 0 ||
4112 		    io->pfrio_size > pf_ioctl_maxcount ||
4113 		    WOULD_OVERFLOW(io->pfrio_size, sizeof(struct pfr_addr))) {
4114 			error = EINVAL;
4115 			break;
4116 		}
4117 		totlen = io->pfrio_size * sizeof(struct pfr_addr);
4118 		pfras = mallocarray(io->pfrio_size, sizeof(struct pfr_addr),
4119 		    M_TEMP, M_WAITOK);
4120 		error = copyin(io->pfrio_buffer, pfras, totlen);
4121 		if (error) {
4122 			free(pfras, M_TEMP);
4123 			break;
4124 		}
4125 		PF_RULES_WLOCK();
4126 		error = pfr_del_addrs(&io->pfrio_table, pfras,
4127 		    io->pfrio_size, &io->pfrio_ndel, io->pfrio_flags |
4128 		    PFR_FLAG_USERIOCTL);
4129 		PF_RULES_WUNLOCK();
4130 		if (error == 0 && io->pfrio_flags & PFR_FLAG_FEEDBACK)
4131 			error = copyout(pfras, io->pfrio_buffer, totlen);
4132 		free(pfras, M_TEMP);
4133 		break;
4134 	}
4135 
4136 	case DIOCRSETADDRS: {
4137 		struct pfioc_table *io = (struct pfioc_table *)addr;
4138 		struct pfr_addr *pfras;
4139 		size_t totlen, count;
4140 
4141 		if (io->pfrio_esize != sizeof(struct pfr_addr)) {
4142 			error = ENODEV;
4143 			break;
4144 		}
4145 		if (io->pfrio_size < 0 || io->pfrio_size2 < 0) {
4146 			error = EINVAL;
4147 			break;
4148 		}
4149 		count = max(io->pfrio_size, io->pfrio_size2);
4150 		if (count > pf_ioctl_maxcount ||
4151 		    WOULD_OVERFLOW(count, sizeof(struct pfr_addr))) {
4152 			error = EINVAL;
4153 			break;
4154 		}
4155 		totlen = count * sizeof(struct pfr_addr);
4156 		pfras = mallocarray(count, sizeof(struct pfr_addr), M_TEMP,
4157 		    M_WAITOK);
4158 		error = copyin(io->pfrio_buffer, pfras, totlen);
4159 		if (error) {
4160 			free(pfras, M_TEMP);
4161 			break;
4162 		}
4163 		PF_RULES_WLOCK();
4164 		error = pfr_set_addrs(&io->pfrio_table, pfras,
4165 		    io->pfrio_size, &io->pfrio_size2, &io->pfrio_nadd,
4166 		    &io->pfrio_ndel, &io->pfrio_nchange, io->pfrio_flags |
4167 		    PFR_FLAG_USERIOCTL, 0);
4168 		PF_RULES_WUNLOCK();
4169 		if (error == 0 && io->pfrio_flags & PFR_FLAG_FEEDBACK)
4170 			error = copyout(pfras, io->pfrio_buffer, totlen);
4171 		free(pfras, M_TEMP);
4172 		break;
4173 	}
4174 
4175 	case DIOCRGETADDRS: {
4176 		struct pfioc_table *io = (struct pfioc_table *)addr;
4177 		struct pfr_addr *pfras;
4178 		size_t totlen;
4179 
4180 		if (io->pfrio_esize != sizeof(struct pfr_addr)) {
4181 			error = ENODEV;
4182 			break;
4183 		}
4184 		if (io->pfrio_size < 0 ||
4185 		    io->pfrio_size > pf_ioctl_maxcount ||
4186 		    WOULD_OVERFLOW(io->pfrio_size, sizeof(struct pfr_addr))) {
4187 			error = EINVAL;
4188 			break;
4189 		}
4190 		totlen = io->pfrio_size * sizeof(struct pfr_addr);
4191 		pfras = mallocarray(io->pfrio_size, sizeof(struct pfr_addr),
4192 		    M_TEMP, M_WAITOK | M_ZERO);
4193 		PF_RULES_RLOCK();
4194 		error = pfr_get_addrs(&io->pfrio_table, pfras,
4195 		    &io->pfrio_size, io->pfrio_flags | PFR_FLAG_USERIOCTL);
4196 		PF_RULES_RUNLOCK();
4197 		if (error == 0)
4198 			error = copyout(pfras, io->pfrio_buffer, totlen);
4199 		free(pfras, M_TEMP);
4200 		break;
4201 	}
4202 
4203 	case DIOCRGETASTATS: {
4204 		struct pfioc_table *io = (struct pfioc_table *)addr;
4205 		struct pfr_astats *pfrastats;
4206 		size_t totlen;
4207 
4208 		if (io->pfrio_esize != sizeof(struct pfr_astats)) {
4209 			error = ENODEV;
4210 			break;
4211 		}
4212 		if (io->pfrio_size < 0 ||
4213 		    io->pfrio_size > pf_ioctl_maxcount ||
4214 		    WOULD_OVERFLOW(io->pfrio_size, sizeof(struct pfr_astats))) {
4215 			error = EINVAL;
4216 			break;
4217 		}
4218 		totlen = io->pfrio_size * sizeof(struct pfr_astats);
4219 		pfrastats = mallocarray(io->pfrio_size,
4220 		    sizeof(struct pfr_astats), M_TEMP, M_WAITOK | M_ZERO);
4221 		PF_RULES_RLOCK();
4222 		error = pfr_get_astats(&io->pfrio_table, pfrastats,
4223 		    &io->pfrio_size, io->pfrio_flags | PFR_FLAG_USERIOCTL);
4224 		PF_RULES_RUNLOCK();
4225 		if (error == 0)
4226 			error = copyout(pfrastats, io->pfrio_buffer, totlen);
4227 		free(pfrastats, M_TEMP);
4228 		break;
4229 	}
4230 
4231 	case DIOCRCLRASTATS: {
4232 		struct pfioc_table *io = (struct pfioc_table *)addr;
4233 		struct pfr_addr *pfras;
4234 		size_t totlen;
4235 
4236 		if (io->pfrio_esize != sizeof(struct pfr_addr)) {
4237 			error = ENODEV;
4238 			break;
4239 		}
4240 		if (io->pfrio_size < 0 ||
4241 		    io->pfrio_size > pf_ioctl_maxcount ||
4242 		    WOULD_OVERFLOW(io->pfrio_size, sizeof(struct pfr_addr))) {
4243 			error = EINVAL;
4244 			break;
4245 		}
4246 		totlen = io->pfrio_size * sizeof(struct pfr_addr);
4247 		pfras = mallocarray(io->pfrio_size, sizeof(struct pfr_addr),
4248 		    M_TEMP, M_WAITOK);
4249 		error = copyin(io->pfrio_buffer, pfras, totlen);
4250 		if (error) {
4251 			free(pfras, M_TEMP);
4252 			break;
4253 		}
4254 		PF_RULES_WLOCK();
4255 		error = pfr_clr_astats(&io->pfrio_table, pfras,
4256 		    io->pfrio_size, &io->pfrio_nzero, io->pfrio_flags |
4257 		    PFR_FLAG_USERIOCTL);
4258 		PF_RULES_WUNLOCK();
4259 		if (error == 0 && io->pfrio_flags & PFR_FLAG_FEEDBACK)
4260 			error = copyout(pfras, io->pfrio_buffer, totlen);
4261 		free(pfras, M_TEMP);
4262 		break;
4263 	}
4264 
4265 	case DIOCRTSTADDRS: {
4266 		struct pfioc_table *io = (struct pfioc_table *)addr;
4267 		struct pfr_addr *pfras;
4268 		size_t totlen;
4269 
4270 		if (io->pfrio_esize != sizeof(struct pfr_addr)) {
4271 			error = ENODEV;
4272 			break;
4273 		}
4274 		if (io->pfrio_size < 0 ||
4275 		    io->pfrio_size > pf_ioctl_maxcount ||
4276 		    WOULD_OVERFLOW(io->pfrio_size, sizeof(struct pfr_addr))) {
4277 			error = EINVAL;
4278 			break;
4279 		}
4280 		totlen = io->pfrio_size * sizeof(struct pfr_addr);
4281 		pfras = mallocarray(io->pfrio_size, sizeof(struct pfr_addr),
4282 		    M_TEMP, M_WAITOK);
4283 		error = copyin(io->pfrio_buffer, pfras, totlen);
4284 		if (error) {
4285 			free(pfras, M_TEMP);
4286 			break;
4287 		}
4288 		PF_RULES_RLOCK();
4289 		error = pfr_tst_addrs(&io->pfrio_table, pfras,
4290 		    io->pfrio_size, &io->pfrio_nmatch, io->pfrio_flags |
4291 		    PFR_FLAG_USERIOCTL);
4292 		PF_RULES_RUNLOCK();
4293 		if (error == 0)
4294 			error = copyout(pfras, io->pfrio_buffer, totlen);
4295 		free(pfras, M_TEMP);
4296 		break;
4297 	}
4298 
4299 	case DIOCRINADEFINE: {
4300 		struct pfioc_table *io = (struct pfioc_table *)addr;
4301 		struct pfr_addr *pfras;
4302 		size_t totlen;
4303 
4304 		if (io->pfrio_esize != sizeof(struct pfr_addr)) {
4305 			error = ENODEV;
4306 			break;
4307 		}
4308 		if (io->pfrio_size < 0 ||
4309 		    io->pfrio_size > pf_ioctl_maxcount ||
4310 		    WOULD_OVERFLOW(io->pfrio_size, sizeof(struct pfr_addr))) {
4311 			error = EINVAL;
4312 			break;
4313 		}
4314 		totlen = io->pfrio_size * sizeof(struct pfr_addr);
4315 		pfras = mallocarray(io->pfrio_size, sizeof(struct pfr_addr),
4316 		    M_TEMP, M_WAITOK);
4317 		error = copyin(io->pfrio_buffer, pfras, totlen);
4318 		if (error) {
4319 			free(pfras, M_TEMP);
4320 			break;
4321 		}
4322 		PF_RULES_WLOCK();
4323 		error = pfr_ina_define(&io->pfrio_table, pfras,
4324 		    io->pfrio_size, &io->pfrio_nadd, &io->pfrio_naddr,
4325 		    io->pfrio_ticket, io->pfrio_flags | PFR_FLAG_USERIOCTL);
4326 		PF_RULES_WUNLOCK();
4327 		free(pfras, M_TEMP);
4328 		break;
4329 	}
4330 
4331 	case DIOCOSFPADD: {
4332 		struct pf_osfp_ioctl *io = (struct pf_osfp_ioctl *)addr;
4333 		PF_RULES_WLOCK();
4334 		error = pf_osfp_add(io);
4335 		PF_RULES_WUNLOCK();
4336 		break;
4337 	}
4338 
4339 	case DIOCOSFPGET: {
4340 		struct pf_osfp_ioctl *io = (struct pf_osfp_ioctl *)addr;
4341 		PF_RULES_RLOCK();
4342 		error = pf_osfp_get(io);
4343 		PF_RULES_RUNLOCK();
4344 		break;
4345 	}
4346 
4347 	case DIOCXBEGIN: {
4348 		struct pfioc_trans	*io = (struct pfioc_trans *)addr;
4349 		struct pfioc_trans_e	*ioes, *ioe;
4350 		size_t			 totlen;
4351 		int			 i;
4352 
4353 		if (io->esize != sizeof(*ioe)) {
4354 			error = ENODEV;
4355 			break;
4356 		}
4357 		if (io->size < 0 ||
4358 		    io->size > pf_ioctl_maxcount ||
4359 		    WOULD_OVERFLOW(io->size, sizeof(struct pfioc_trans_e))) {
4360 			error = EINVAL;
4361 			break;
4362 		}
4363 		totlen = sizeof(struct pfioc_trans_e) * io->size;
4364 		ioes = mallocarray(io->size, sizeof(struct pfioc_trans_e),
4365 		    M_TEMP, M_WAITOK);
4366 		error = copyin(io->array, ioes, totlen);
4367 		if (error) {
4368 			free(ioes, M_TEMP);
4369 			break;
4370 		}
4371 		PF_RULES_WLOCK();
4372 		for (i = 0, ioe = ioes; i < io->size; i++, ioe++) {
4373 			ioe->anchor[sizeof(ioe->anchor) - 1] = '\0';
4374 			switch (ioe->rs_num) {
4375 #ifdef ALTQ
4376 			case PF_RULESET_ALTQ:
4377 				if (ioe->anchor[0]) {
4378 					PF_RULES_WUNLOCK();
4379 					free(ioes, M_TEMP);
4380 					error = EINVAL;
4381 					goto fail;
4382 				}
4383 				if ((error = pf_begin_altq(&ioe->ticket))) {
4384 					PF_RULES_WUNLOCK();
4385 					free(ioes, M_TEMP);
4386 					goto fail;
4387 				}
4388 				break;
4389 #endif /* ALTQ */
4390 			case PF_RULESET_TABLE:
4391 			    {
4392 				struct pfr_table table;
4393 
4394 				bzero(&table, sizeof(table));
4395 				strlcpy(table.pfrt_anchor, ioe->anchor,
4396 				    sizeof(table.pfrt_anchor));
4397 				if ((error = pfr_ina_begin(&table,
4398 				    &ioe->ticket, NULL, 0))) {
4399 					PF_RULES_WUNLOCK();
4400 					free(ioes, M_TEMP);
4401 					goto fail;
4402 				}
4403 				break;
4404 			    }
4405 			default:
4406 				if ((error = pf_begin_rules(&ioe->ticket,
4407 				    ioe->rs_num, ioe->anchor))) {
4408 					PF_RULES_WUNLOCK();
4409 					free(ioes, M_TEMP);
4410 					goto fail;
4411 				}
4412 				break;
4413 			}
4414 		}
4415 		PF_RULES_WUNLOCK();
4416 		error = copyout(ioes, io->array, totlen);
4417 		free(ioes, M_TEMP);
4418 		break;
4419 	}
4420 
4421 	case DIOCXROLLBACK: {
4422 		struct pfioc_trans	*io = (struct pfioc_trans *)addr;
4423 		struct pfioc_trans_e	*ioe, *ioes;
4424 		size_t			 totlen;
4425 		int			 i;
4426 
4427 		if (io->esize != sizeof(*ioe)) {
4428 			error = ENODEV;
4429 			break;
4430 		}
4431 		if (io->size < 0 ||
4432 		    io->size > pf_ioctl_maxcount ||
4433 		    WOULD_OVERFLOW(io->size, sizeof(struct pfioc_trans_e))) {
4434 			error = EINVAL;
4435 			break;
4436 		}
4437 		totlen = sizeof(struct pfioc_trans_e) * io->size;
4438 		ioes = mallocarray(io->size, sizeof(struct pfioc_trans_e),
4439 		    M_TEMP, M_WAITOK);
4440 		error = copyin(io->array, ioes, totlen);
4441 		if (error) {
4442 			free(ioes, M_TEMP);
4443 			break;
4444 		}
4445 		PF_RULES_WLOCK();
4446 		for (i = 0, ioe = ioes; i < io->size; i++, ioe++) {
4447 			ioe->anchor[sizeof(ioe->anchor) - 1] = '\0';
4448 			switch (ioe->rs_num) {
4449 #ifdef ALTQ
4450 			case PF_RULESET_ALTQ:
4451 				if (ioe->anchor[0]) {
4452 					PF_RULES_WUNLOCK();
4453 					free(ioes, M_TEMP);
4454 					error = EINVAL;
4455 					goto fail;
4456 				}
4457 				if ((error = pf_rollback_altq(ioe->ticket))) {
4458 					PF_RULES_WUNLOCK();
4459 					free(ioes, M_TEMP);
4460 					goto fail; /* really bad */
4461 				}
4462 				break;
4463 #endif /* ALTQ */
4464 			case PF_RULESET_TABLE:
4465 			    {
4466 				struct pfr_table table;
4467 
4468 				bzero(&table, sizeof(table));
4469 				strlcpy(table.pfrt_anchor, ioe->anchor,
4470 				    sizeof(table.pfrt_anchor));
4471 				if ((error = pfr_ina_rollback(&table,
4472 				    ioe->ticket, NULL, 0))) {
4473 					PF_RULES_WUNLOCK();
4474 					free(ioes, M_TEMP);
4475 					goto fail; /* really bad */
4476 				}
4477 				break;
4478 			    }
4479 			default:
4480 				if ((error = pf_rollback_rules(ioe->ticket,
4481 				    ioe->rs_num, ioe->anchor))) {
4482 					PF_RULES_WUNLOCK();
4483 					free(ioes, M_TEMP);
4484 					goto fail; /* really bad */
4485 				}
4486 				break;
4487 			}
4488 		}
4489 		PF_RULES_WUNLOCK();
4490 		free(ioes, M_TEMP);
4491 		break;
4492 	}
4493 
4494 	case DIOCXCOMMIT: {
4495 		struct pfioc_trans	*io = (struct pfioc_trans *)addr;
4496 		struct pfioc_trans_e	*ioe, *ioes;
4497 		struct pf_kruleset	*rs;
4498 		size_t			 totlen;
4499 		int			 i;
4500 
4501 		if (io->esize != sizeof(*ioe)) {
4502 			error = ENODEV;
4503 			break;
4504 		}
4505 
4506 		if (io->size < 0 ||
4507 		    io->size > pf_ioctl_maxcount ||
4508 		    WOULD_OVERFLOW(io->size, sizeof(struct pfioc_trans_e))) {
4509 			error = EINVAL;
4510 			break;
4511 		}
4512 
4513 		totlen = sizeof(struct pfioc_trans_e) * io->size;
4514 		ioes = mallocarray(io->size, sizeof(struct pfioc_trans_e),
4515 		    M_TEMP, M_WAITOK);
4516 		error = copyin(io->array, ioes, totlen);
4517 		if (error) {
4518 			free(ioes, M_TEMP);
4519 			break;
4520 		}
4521 		PF_RULES_WLOCK();
4522 		/* First makes sure everything will succeed. */
4523 		for (i = 0, ioe = ioes; i < io->size; i++, ioe++) {
4524 			ioe->anchor[sizeof(ioe->anchor) - 1] = 0;
4525 			switch (ioe->rs_num) {
4526 #ifdef ALTQ
4527 			case PF_RULESET_ALTQ:
4528 				if (ioe->anchor[0]) {
4529 					PF_RULES_WUNLOCK();
4530 					free(ioes, M_TEMP);
4531 					error = EINVAL;
4532 					goto fail;
4533 				}
4534 				if (!V_altqs_inactive_open || ioe->ticket !=
4535 				    V_ticket_altqs_inactive) {
4536 					PF_RULES_WUNLOCK();
4537 					free(ioes, M_TEMP);
4538 					error = EBUSY;
4539 					goto fail;
4540 				}
4541 				break;
4542 #endif /* ALTQ */
4543 			case PF_RULESET_TABLE:
4544 				rs = pf_find_kruleset(ioe->anchor);
4545 				if (rs == NULL || !rs->topen || ioe->ticket !=
4546 				    rs->tticket) {
4547 					PF_RULES_WUNLOCK();
4548 					free(ioes, M_TEMP);
4549 					error = EBUSY;
4550 					goto fail;
4551 				}
4552 				break;
4553 			default:
4554 				if (ioe->rs_num < 0 || ioe->rs_num >=
4555 				    PF_RULESET_MAX) {
4556 					PF_RULES_WUNLOCK();
4557 					free(ioes, M_TEMP);
4558 					error = EINVAL;
4559 					goto fail;
4560 				}
4561 				rs = pf_find_kruleset(ioe->anchor);
4562 				if (rs == NULL ||
4563 				    !rs->rules[ioe->rs_num].inactive.open ||
4564 				    rs->rules[ioe->rs_num].inactive.ticket !=
4565 				    ioe->ticket) {
4566 					PF_RULES_WUNLOCK();
4567 					free(ioes, M_TEMP);
4568 					error = EBUSY;
4569 					goto fail;
4570 				}
4571 				break;
4572 			}
4573 		}
4574 		/* Now do the commit - no errors should happen here. */
4575 		for (i = 0, ioe = ioes; i < io->size; i++, ioe++) {
4576 			switch (ioe->rs_num) {
4577 #ifdef ALTQ
4578 			case PF_RULESET_ALTQ:
4579 				if ((error = pf_commit_altq(ioe->ticket))) {
4580 					PF_RULES_WUNLOCK();
4581 					free(ioes, M_TEMP);
4582 					goto fail; /* really bad */
4583 				}
4584 				break;
4585 #endif /* ALTQ */
4586 			case PF_RULESET_TABLE:
4587 			    {
4588 				struct pfr_table table;
4589 
4590 				bzero(&table, sizeof(table));
4591 				(void)strlcpy(table.pfrt_anchor, ioe->anchor,
4592 				    sizeof(table.pfrt_anchor));
4593 				if ((error = pfr_ina_commit(&table,
4594 				    ioe->ticket, NULL, NULL, 0))) {
4595 					PF_RULES_WUNLOCK();
4596 					free(ioes, M_TEMP);
4597 					goto fail; /* really bad */
4598 				}
4599 				break;
4600 			    }
4601 			default:
4602 				if ((error = pf_commit_rules(ioe->ticket,
4603 				    ioe->rs_num, ioe->anchor))) {
4604 					PF_RULES_WUNLOCK();
4605 					free(ioes, M_TEMP);
4606 					goto fail; /* really bad */
4607 				}
4608 				break;
4609 			}
4610 		}
4611 		PF_RULES_WUNLOCK();
4612 		free(ioes, M_TEMP);
4613 		break;
4614 	}
4615 
4616 	case DIOCGETSRCNODES: {
4617 		struct pfioc_src_nodes	*psn = (struct pfioc_src_nodes *)addr;
4618 		struct pf_srchash	*sh;
4619 		struct pf_ksrc_node	*n;
4620 		struct pf_src_node	*p, *pstore;
4621 		uint32_t		 i, nr = 0;
4622 
4623 		for (i = 0, sh = V_pf_srchash; i <= pf_srchashmask;
4624 				i++, sh++) {
4625 			PF_HASHROW_LOCK(sh);
4626 			LIST_FOREACH(n, &sh->nodes, entry)
4627 				nr++;
4628 			PF_HASHROW_UNLOCK(sh);
4629 		}
4630 
4631 		psn->psn_len = min(psn->psn_len,
4632 		    sizeof(struct pf_src_node) * nr);
4633 
4634 		if (psn->psn_len == 0) {
4635 			psn->psn_len = sizeof(struct pf_src_node) * nr;
4636 			break;
4637 		}
4638 
4639 		nr = 0;
4640 
4641 		p = pstore = malloc(psn->psn_len, M_TEMP, M_WAITOK | M_ZERO);
4642 		for (i = 0, sh = V_pf_srchash; i <= pf_srchashmask;
4643 		    i++, sh++) {
4644 		    PF_HASHROW_LOCK(sh);
4645 		    LIST_FOREACH(n, &sh->nodes, entry) {
4646 
4647 			if ((nr + 1) * sizeof(*p) > (unsigned)psn->psn_len)
4648 				break;
4649 
4650 			pf_src_node_copy(n, p);
4651 
4652 			p++;
4653 			nr++;
4654 		    }
4655 		    PF_HASHROW_UNLOCK(sh);
4656 		}
4657 		error = copyout(pstore, psn->psn_src_nodes,
4658 		    sizeof(struct pf_src_node) * nr);
4659 		if (error) {
4660 			free(pstore, M_TEMP);
4661 			break;
4662 		}
4663 		psn->psn_len = sizeof(struct pf_src_node) * nr;
4664 		free(pstore, M_TEMP);
4665 		break;
4666 	}
4667 
4668 	case DIOCCLRSRCNODES: {
4669 		pf_clear_srcnodes(NULL);
4670 		pf_purge_expired_src_nodes();
4671 		break;
4672 	}
4673 
4674 	case DIOCKILLSRCNODES:
4675 		pf_kill_srcnodes((struct pfioc_src_node_kill *)addr);
4676 		break;
4677 
4678 	case DIOCKEEPCOUNTERS:
4679 		error = pf_keepcounters((struct pfioc_nv *)addr);
4680 		break;
4681 
4682 	case DIOCGETSYNCOOKIES:
4683 		error = pf_get_syncookies((struct pfioc_nv *)addr);
4684 		break;
4685 
4686 	case DIOCSETSYNCOOKIES:
4687 		error = pf_set_syncookies((struct pfioc_nv *)addr);
4688 		break;
4689 
4690 	case DIOCSETHOSTID: {
4691 		u_int32_t	*hostid = (u_int32_t *)addr;
4692 
4693 		PF_RULES_WLOCK();
4694 		if (*hostid == 0)
4695 			V_pf_status.hostid = arc4random();
4696 		else
4697 			V_pf_status.hostid = *hostid;
4698 		PF_RULES_WUNLOCK();
4699 		break;
4700 	}
4701 
4702 	case DIOCOSFPFLUSH:
4703 		PF_RULES_WLOCK();
4704 		pf_osfp_flush();
4705 		PF_RULES_WUNLOCK();
4706 		break;
4707 
4708 	case DIOCIGETIFACES: {
4709 		struct pfioc_iface *io = (struct pfioc_iface *)addr;
4710 		struct pfi_kif *ifstore;
4711 		size_t bufsiz;
4712 
4713 		if (io->pfiio_esize != sizeof(struct pfi_kif)) {
4714 			error = ENODEV;
4715 			break;
4716 		}
4717 
4718 		if (io->pfiio_size < 0 ||
4719 		    io->pfiio_size > pf_ioctl_maxcount ||
4720 		    WOULD_OVERFLOW(io->pfiio_size, sizeof(struct pfi_kif))) {
4721 			error = EINVAL;
4722 			break;
4723 		}
4724 
4725 		bufsiz = io->pfiio_size * sizeof(struct pfi_kif);
4726 		ifstore = mallocarray(io->pfiio_size, sizeof(struct pfi_kif),
4727 		    M_TEMP, M_WAITOK | M_ZERO);
4728 
4729 		PF_RULES_RLOCK();
4730 		pfi_get_ifaces(io->pfiio_name, ifstore, &io->pfiio_size);
4731 		PF_RULES_RUNLOCK();
4732 		error = copyout(ifstore, io->pfiio_buffer, bufsiz);
4733 		free(ifstore, M_TEMP);
4734 		break;
4735 	}
4736 
4737 	case DIOCSETIFFLAG: {
4738 		struct pfioc_iface *io = (struct pfioc_iface *)addr;
4739 
4740 		PF_RULES_WLOCK();
4741 		error = pfi_set_flags(io->pfiio_name, io->pfiio_flags);
4742 		PF_RULES_WUNLOCK();
4743 		break;
4744 	}
4745 
4746 	case DIOCCLRIFFLAG: {
4747 		struct pfioc_iface *io = (struct pfioc_iface *)addr;
4748 
4749 		PF_RULES_WLOCK();
4750 		error = pfi_clear_flags(io->pfiio_name, io->pfiio_flags);
4751 		PF_RULES_WUNLOCK();
4752 		break;
4753 	}
4754 
4755 	default:
4756 		error = ENODEV;
4757 		break;
4758 	}
4759 fail:
4760 	if (sx_xlocked(&pf_ioctl_lock))
4761 		sx_xunlock(&pf_ioctl_lock);
4762 	CURVNET_RESTORE();
4763 
4764 #undef ERROUT_IOCTL
4765 
4766 	return (error);
4767 }
4768 
4769 void
pfsync_state_export(struct pfsync_state * sp,struct pf_kstate * st)4770 pfsync_state_export(struct pfsync_state *sp, struct pf_kstate *st)
4771 {
4772 	bzero(sp, sizeof(struct pfsync_state));
4773 
4774 	/* copy from state key */
4775 	sp->key[PF_SK_WIRE].addr[0] = st->key[PF_SK_WIRE]->addr[0];
4776 	sp->key[PF_SK_WIRE].addr[1] = st->key[PF_SK_WIRE]->addr[1];
4777 	sp->key[PF_SK_WIRE].port[0] = st->key[PF_SK_WIRE]->port[0];
4778 	sp->key[PF_SK_WIRE].port[1] = st->key[PF_SK_WIRE]->port[1];
4779 	sp->key[PF_SK_STACK].addr[0] = st->key[PF_SK_STACK]->addr[0];
4780 	sp->key[PF_SK_STACK].addr[1] = st->key[PF_SK_STACK]->addr[1];
4781 	sp->key[PF_SK_STACK].port[0] = st->key[PF_SK_STACK]->port[0];
4782 	sp->key[PF_SK_STACK].port[1] = st->key[PF_SK_STACK]->port[1];
4783 	sp->proto = st->key[PF_SK_WIRE]->proto;
4784 	sp->af = st->key[PF_SK_WIRE]->af;
4785 
4786 	/* copy from state */
4787 	strlcpy(sp->ifname, st->kif->pfik_name, sizeof(sp->ifname));
4788 	bcopy(&st->rt_addr, &sp->rt_addr, sizeof(sp->rt_addr));
4789 	sp->creation = htonl(time_uptime - st->creation);
4790 	sp->expire = pf_state_expires(st);
4791 	if (sp->expire <= time_uptime)
4792 		sp->expire = htonl(0);
4793 	else
4794 		sp->expire = htonl(sp->expire - time_uptime);
4795 
4796 	sp->direction = st->direction;
4797 	sp->log = st->log;
4798 	sp->timeout = st->timeout;
4799 	sp->state_flags = st->state_flags;
4800 	if (st->src_node)
4801 		sp->sync_flags |= PFSYNC_FLAG_SRCNODE;
4802 	if (st->nat_src_node)
4803 		sp->sync_flags |= PFSYNC_FLAG_NATSRCNODE;
4804 
4805 	sp->id = st->id;
4806 	sp->creatorid = st->creatorid;
4807 	pf_state_peer_hton(&st->src, &sp->src);
4808 	pf_state_peer_hton(&st->dst, &sp->dst);
4809 
4810 	if (st->rule.ptr == NULL)
4811 		sp->rule = htonl(-1);
4812 	else
4813 		sp->rule = htonl(st->rule.ptr->nr);
4814 	if (st->anchor.ptr == NULL)
4815 		sp->anchor = htonl(-1);
4816 	else
4817 		sp->anchor = htonl(st->anchor.ptr->nr);
4818 	if (st->nat_rule.ptr == NULL)
4819 		sp->nat_rule = htonl(-1);
4820 	else
4821 		sp->nat_rule = htonl(st->nat_rule.ptr->nr);
4822 
4823 	pf_state_counter_hton(st->packets[0], sp->packets[0]);
4824 	pf_state_counter_hton(st->packets[1], sp->packets[1]);
4825 	pf_state_counter_hton(st->bytes[0], sp->bytes[0]);
4826 	pf_state_counter_hton(st->bytes[1], sp->bytes[1]);
4827 }
4828 
4829 void
pf_state_export(struct pf_state_export * sp,struct pf_kstate * st)4830 pf_state_export(struct pf_state_export *sp, struct pf_kstate *st)
4831 {
4832 	bzero(sp, sizeof(*sp));
4833 
4834 	sp->version = PF_STATE_VERSION;
4835 
4836 	/* copy from state key */
4837 	sp->key[PF_SK_WIRE].addr[0] = st->key[PF_SK_WIRE]->addr[0];
4838 	sp->key[PF_SK_WIRE].addr[1] = st->key[PF_SK_WIRE]->addr[1];
4839 	sp->key[PF_SK_WIRE].port[0] = st->key[PF_SK_WIRE]->port[0];
4840 	sp->key[PF_SK_WIRE].port[1] = st->key[PF_SK_WIRE]->port[1];
4841 	sp->key[PF_SK_STACK].addr[0] = st->key[PF_SK_STACK]->addr[0];
4842 	sp->key[PF_SK_STACK].addr[1] = st->key[PF_SK_STACK]->addr[1];
4843 	sp->key[PF_SK_STACK].port[0] = st->key[PF_SK_STACK]->port[0];
4844 	sp->key[PF_SK_STACK].port[1] = st->key[PF_SK_STACK]->port[1];
4845 	sp->proto = st->key[PF_SK_WIRE]->proto;
4846 	sp->af = st->key[PF_SK_WIRE]->af;
4847 
4848 	/* copy from state */
4849 	strlcpy(sp->ifname, st->kif->pfik_name, sizeof(sp->ifname));
4850 	strlcpy(sp->orig_ifname, st->orig_kif->pfik_name,
4851 	    sizeof(sp->orig_ifname));
4852 	bcopy(&st->rt_addr, &sp->rt_addr, sizeof(sp->rt_addr));
4853 	sp->creation = htonl(time_uptime - st->creation);
4854 	sp->expire = pf_state_expires(st);
4855 	if (sp->expire <= time_uptime)
4856 		sp->expire = htonl(0);
4857 	else
4858 		sp->expire = htonl(sp->expire - time_uptime);
4859 
4860 	sp->direction = st->direction;
4861 	sp->log = st->log;
4862 	sp->timeout = st->timeout;
4863 	sp->state_flags = st->state_flags;
4864 	if (st->src_node)
4865 		sp->sync_flags |= PFSYNC_FLAG_SRCNODE;
4866 	if (st->nat_src_node)
4867 		sp->sync_flags |= PFSYNC_FLAG_NATSRCNODE;
4868 
4869 	sp->id = st->id;
4870 	sp->creatorid = st->creatorid;
4871 	pf_state_peer_hton(&st->src, &sp->src);
4872 	pf_state_peer_hton(&st->dst, &sp->dst);
4873 
4874 	if (st->rule.ptr == NULL)
4875 		sp->rule = htonl(-1);
4876 	else
4877 		sp->rule = htonl(st->rule.ptr->nr);
4878 	if (st->anchor.ptr == NULL)
4879 		sp->anchor = htonl(-1);
4880 	else
4881 		sp->anchor = htonl(st->anchor.ptr->nr);
4882 	if (st->nat_rule.ptr == NULL)
4883 		sp->nat_rule = htonl(-1);
4884 	else
4885 		sp->nat_rule = htonl(st->nat_rule.ptr->nr);
4886 
4887 	sp->packets[0] = st->packets[0];
4888 	sp->packets[1] = st->packets[1];
4889 	sp->bytes[0] = st->bytes[0];
4890 	sp->bytes[1] = st->bytes[1];
4891 }
4892 
4893 static void
pf_tbladdr_copyout(struct pf_addr_wrap * aw)4894 pf_tbladdr_copyout(struct pf_addr_wrap *aw)
4895 {
4896 	struct pfr_ktable *kt;
4897 
4898 	KASSERT(aw->type == PF_ADDR_TABLE, ("%s: type %u", __func__, aw->type));
4899 
4900 	kt = aw->p.tbl;
4901 	if (!(kt->pfrkt_flags & PFR_TFLAG_ACTIVE) && kt->pfrkt_root != NULL)
4902 		kt = kt->pfrkt_root;
4903 	aw->p.tbl = NULL;
4904 	aw->p.tblcnt = (kt->pfrkt_flags & PFR_TFLAG_ACTIVE) ?
4905 		kt->pfrkt_cnt : -1;
4906 }
4907 
4908 static int
pf_add_status_counters(nvlist_t * nvl,const char * name,counter_u64_t * counters,size_t number,char ** names)4909 pf_add_status_counters(nvlist_t *nvl, const char *name, counter_u64_t *counters,
4910     size_t number, char **names)
4911 {
4912 	nvlist_t        *nvc;
4913 
4914 	nvc = nvlist_create(0);
4915 	if (nvc == NULL)
4916 		return (ENOMEM);
4917 
4918 	for (int i = 0; i < number; i++) {
4919 		nvlist_append_number_array(nvc, "counters",
4920 		    counter_u64_fetch(counters[i]));
4921 		nvlist_append_string_array(nvc, "names",
4922 		    names[i]);
4923 		nvlist_append_number_array(nvc, "ids",
4924 		    i);
4925 	}
4926 	nvlist_add_nvlist(nvl, name, nvc);
4927 	nvlist_destroy(nvc);
4928 
4929 	return (0);
4930 }
4931 
4932 static int
pf_getstatus(struct pfioc_nv * nv)4933 pf_getstatus(struct pfioc_nv *nv)
4934 {
4935 	nvlist_t        *nvl = NULL, *nvc = NULL;
4936 	void            *nvlpacked = NULL;
4937 	int              error;
4938 	struct pf_status s;
4939 	char *pf_reasons[PFRES_MAX+1] = PFRES_NAMES;
4940 	char *pf_lcounter[KLCNT_MAX+1] = KLCNT_NAMES;
4941 	char *pf_fcounter[FCNT_MAX+1] = FCNT_NAMES;
4942 	PF_RULES_RLOCK_TRACKER;
4943 
4944 #define ERROUT(x)      ERROUT_FUNCTION(errout, x)
4945 
4946 	PF_RULES_RLOCK();
4947 
4948 	nvl = nvlist_create(0);
4949 	if (nvl == NULL)
4950 		ERROUT(ENOMEM);
4951 
4952 	nvlist_add_bool(nvl, "running", V_pf_status.running);
4953 	nvlist_add_number(nvl, "since", V_pf_status.since);
4954 	nvlist_add_number(nvl, "debug", V_pf_status.debug);
4955 	nvlist_add_number(nvl, "hostid", V_pf_status.hostid);
4956 	nvlist_add_number(nvl, "states", V_pf_status.states);
4957 	nvlist_add_number(nvl, "src_nodes", V_pf_status.src_nodes);
4958 
4959 	/* counters */
4960 	error = pf_add_status_counters(nvl, "counters", V_pf_status.counters,
4961 	    PFRES_MAX, pf_reasons);
4962 	if (error != 0)
4963 		ERROUT(error);
4964 
4965 	/* lcounters */
4966 	error = pf_add_status_counters(nvl, "lcounters", V_pf_status.lcounters,
4967 	    KLCNT_MAX, pf_lcounter);
4968 	if (error != 0)
4969 		ERROUT(error);
4970 
4971 	/* fcounters */
4972 	nvc = nvlist_create(0);
4973 	if (nvc == NULL)
4974 		ERROUT(ENOMEM);
4975 
4976 	for (int i = 0; i < FCNT_MAX; i++) {
4977 		nvlist_append_number_array(nvc, "counters",
4978 		    pf_counter_u64_fetch(&V_pf_status.fcounters[i]));
4979 		nvlist_append_string_array(nvc, "names",
4980 		    pf_fcounter[i]);
4981 		nvlist_append_number_array(nvc, "ids",
4982 		    i);
4983 	}
4984 	nvlist_add_nvlist(nvl, "fcounters", nvc);
4985 	nvlist_destroy(nvc);
4986 	nvc = NULL;
4987 
4988 	/* scounters */
4989 	error = pf_add_status_counters(nvl, "scounters", V_pf_status.scounters,
4990 	    SCNT_MAX, pf_fcounter);
4991 	if (error != 0)
4992 		ERROUT(error);
4993 
4994 	nvlist_add_string(nvl, "ifname", V_pf_status.ifname);
4995 	nvlist_add_binary(nvl, "chksum", V_pf_status.pf_chksum,
4996 	    PF_MD5_DIGEST_LENGTH);
4997 
4998 	pfi_update_status(V_pf_status.ifname, &s);
4999 
5000 	/* pcounters / bcounters */
5001 	for (int i = 0; i < 2; i++) {
5002 		for (int j = 0; j < 2; j++) {
5003 			for (int k = 0; k < 2; k++) {
5004 				nvlist_append_number_array(nvl, "pcounters",
5005 				    s.pcounters[i][j][k]);
5006 			}
5007 			nvlist_append_number_array(nvl, "bcounters",
5008 			    s.bcounters[i][j]);
5009 		}
5010 	}
5011 
5012 	nvlpacked = nvlist_pack(nvl, &nv->len);
5013 	if (nvlpacked == NULL)
5014 		ERROUT(ENOMEM);
5015 
5016 	if (nv->size == 0)
5017 		ERROUT(0);
5018 	else if (nv->size < nv->len)
5019 		ERROUT(ENOSPC);
5020 
5021 	PF_RULES_RUNLOCK();
5022 	error = copyout(nvlpacked, nv->data, nv->len);
5023 	goto done;
5024 
5025 #undef ERROUT
5026 errout:
5027 	PF_RULES_RUNLOCK();
5028 done:
5029 	free(nvlpacked, M_NVLIST);
5030 	nvlist_destroy(nvc);
5031 	nvlist_destroy(nvl);
5032 
5033 	return (error);
5034 }
5035 
5036 /*
5037  * XXX - Check for version missmatch!!!
5038  */
5039 static void
pf_clear_all_states(void)5040 pf_clear_all_states(void)
5041 {
5042 	struct pf_kstate	*s;
5043 	u_int i;
5044 
5045 	for (i = 0; i <= pf_hashmask; i++) {
5046 		struct pf_idhash *ih = &V_pf_idhash[i];
5047 relock:
5048 		PF_HASHROW_LOCK(ih);
5049 		LIST_FOREACH(s, &ih->states, entry) {
5050 			s->timeout = PFTM_PURGE;
5051 			/* Don't send out individual delete messages. */
5052 			s->state_flags |= PFSTATE_NOSYNC;
5053 			pf_unlink_state(s, PF_ENTER_LOCKED);
5054 			goto relock;
5055 		}
5056 		PF_HASHROW_UNLOCK(ih);
5057 	}
5058 }
5059 
5060 static int
pf_clear_tables(void)5061 pf_clear_tables(void)
5062 {
5063 	struct pfioc_table io;
5064 	int error;
5065 
5066 	bzero(&io, sizeof(io));
5067 
5068 	error = pfr_clr_tables(&io.pfrio_table, &io.pfrio_ndel,
5069 	    io.pfrio_flags);
5070 
5071 	return (error);
5072 }
5073 
5074 static void
pf_clear_srcnodes(struct pf_ksrc_node * n)5075 pf_clear_srcnodes(struct pf_ksrc_node *n)
5076 {
5077 	struct pf_kstate *s;
5078 	int i;
5079 
5080 	for (i = 0; i <= pf_hashmask; i++) {
5081 		struct pf_idhash *ih = &V_pf_idhash[i];
5082 
5083 		PF_HASHROW_LOCK(ih);
5084 		LIST_FOREACH(s, &ih->states, entry) {
5085 			if (n == NULL || n == s->src_node)
5086 				s->src_node = NULL;
5087 			if (n == NULL || n == s->nat_src_node)
5088 				s->nat_src_node = NULL;
5089 		}
5090 		PF_HASHROW_UNLOCK(ih);
5091 	}
5092 
5093 	if (n == NULL) {
5094 		struct pf_srchash *sh;
5095 
5096 		for (i = 0, sh = V_pf_srchash; i <= pf_srchashmask;
5097 		    i++, sh++) {
5098 			PF_HASHROW_LOCK(sh);
5099 			LIST_FOREACH(n, &sh->nodes, entry) {
5100 				n->expire = 1;
5101 				n->states = 0;
5102 			}
5103 			PF_HASHROW_UNLOCK(sh);
5104 		}
5105 	} else {
5106 		/* XXX: hash slot should already be locked here. */
5107 		n->expire = 1;
5108 		n->states = 0;
5109 	}
5110 }
5111 
5112 static void
pf_kill_srcnodes(struct pfioc_src_node_kill * psnk)5113 pf_kill_srcnodes(struct pfioc_src_node_kill *psnk)
5114 {
5115 	struct pf_ksrc_node_list	 kill;
5116 
5117 	LIST_INIT(&kill);
5118 	for (int i = 0; i <= pf_srchashmask; i++) {
5119 		struct pf_srchash *sh = &V_pf_srchash[i];
5120 		struct pf_ksrc_node *sn, *tmp;
5121 
5122 		PF_HASHROW_LOCK(sh);
5123 		LIST_FOREACH_SAFE(sn, &sh->nodes, entry, tmp)
5124 			if (PF_MATCHA(psnk->psnk_src.neg,
5125 			      &psnk->psnk_src.addr.v.a.addr,
5126 			      &psnk->psnk_src.addr.v.a.mask,
5127 			      &sn->addr, sn->af) &&
5128 			    PF_MATCHA(psnk->psnk_dst.neg,
5129 			      &psnk->psnk_dst.addr.v.a.addr,
5130 			      &psnk->psnk_dst.addr.v.a.mask,
5131 			      &sn->raddr, sn->af)) {
5132 				pf_unlink_src_node(sn);
5133 				LIST_INSERT_HEAD(&kill, sn, entry);
5134 				sn->expire = 1;
5135 			}
5136 		PF_HASHROW_UNLOCK(sh);
5137 	}
5138 
5139 	for (int i = 0; i <= pf_hashmask; i++) {
5140 		struct pf_idhash *ih = &V_pf_idhash[i];
5141 		struct pf_kstate *s;
5142 
5143 		PF_HASHROW_LOCK(ih);
5144 		LIST_FOREACH(s, &ih->states, entry) {
5145 			if (s->src_node && s->src_node->expire == 1)
5146 				s->src_node = NULL;
5147 			if (s->nat_src_node && s->nat_src_node->expire == 1)
5148 				s->nat_src_node = NULL;
5149 		}
5150 		PF_HASHROW_UNLOCK(ih);
5151 	}
5152 
5153 	psnk->psnk_killed = pf_free_src_nodes(&kill);
5154 }
5155 
5156 static int
pf_keepcounters(struct pfioc_nv * nv)5157 pf_keepcounters(struct pfioc_nv *nv)
5158 {
5159 	nvlist_t	*nvl = NULL;
5160 	void		*nvlpacked = NULL;
5161 	int		 error = 0;
5162 
5163 #define	ERROUT(x)	ERROUT_FUNCTION(on_error, x)
5164 
5165 	if (nv->len > pf_ioctl_maxcount)
5166 		ERROUT(ENOMEM);
5167 
5168 	nvlpacked = malloc(nv->len, M_TEMP, M_WAITOK);
5169 	if (nvlpacked == NULL)
5170 		ERROUT(ENOMEM);
5171 
5172 	error = copyin(nv->data, nvlpacked, nv->len);
5173 	if (error)
5174 		ERROUT(error);
5175 
5176 	nvl = nvlist_unpack(nvlpacked, nv->len, 0);
5177 	if (nvl == NULL)
5178 		ERROUT(EBADMSG);
5179 
5180 	if (! nvlist_exists_bool(nvl, "keep_counters"))
5181 		ERROUT(EBADMSG);
5182 
5183 	V_pf_status.keep_counters = nvlist_get_bool(nvl, "keep_counters");
5184 
5185 on_error:
5186 	nvlist_destroy(nvl);
5187 	free(nvlpacked, M_TEMP);
5188 	return (error);
5189 }
5190 
5191 static unsigned int
pf_clear_states(const struct pf_kstate_kill * kill)5192 pf_clear_states(const struct pf_kstate_kill *kill)
5193 {
5194 	struct pf_state_key_cmp	 match_key;
5195 	struct pf_kstate	*s;
5196 	struct pfi_kkif	*kif;
5197 	int		 idx;
5198 	unsigned int	 killed = 0, dir;
5199 
5200 	for (unsigned int i = 0; i <= pf_hashmask; i++) {
5201 		struct pf_idhash *ih = &V_pf_idhash[i];
5202 
5203 relock_DIOCCLRSTATES:
5204 		PF_HASHROW_LOCK(ih);
5205 		LIST_FOREACH(s, &ih->states, entry) {
5206 			/* For floating states look at the original kif. */
5207 			kif = s->kif == V_pfi_all ? s->orig_kif : s->kif;
5208 
5209 			if (kill->psk_ifname[0] &&
5210 			    strcmp(kill->psk_ifname,
5211 			    kif->pfik_name))
5212 				continue;
5213 
5214 			if (kill->psk_kill_match) {
5215 				bzero(&match_key, sizeof(match_key));
5216 
5217 				if (s->direction == PF_OUT) {
5218 					dir = PF_IN;
5219 					idx = PF_SK_STACK;
5220 				} else {
5221 					dir = PF_OUT;
5222 					idx = PF_SK_WIRE;
5223 				}
5224 
5225 				match_key.af = s->key[idx]->af;
5226 				match_key.proto = s->key[idx]->proto;
5227 				PF_ACPY(&match_key.addr[0],
5228 				    &s->key[idx]->addr[1], match_key.af);
5229 				match_key.port[0] = s->key[idx]->port[1];
5230 				PF_ACPY(&match_key.addr[1],
5231 				    &s->key[idx]->addr[0], match_key.af);
5232 				match_key.port[1] = s->key[idx]->port[0];
5233 			}
5234 
5235 			/*
5236 			 * Don't send out individual
5237 			 * delete messages.
5238 			 */
5239 			s->state_flags |= PFSTATE_NOSYNC;
5240 			pf_unlink_state(s, PF_ENTER_LOCKED);
5241 			killed++;
5242 
5243 			if (kill->psk_kill_match)
5244 				killed += pf_kill_matching_state(&match_key,
5245 				    dir);
5246 
5247 			goto relock_DIOCCLRSTATES;
5248 		}
5249 		PF_HASHROW_UNLOCK(ih);
5250 	}
5251 
5252 	if (V_pfsync_clear_states_ptr != NULL)
5253 		V_pfsync_clear_states_ptr(V_pf_status.hostid, kill->psk_ifname);
5254 
5255 	return (killed);
5256 }
5257 
5258 static void
pf_killstates(struct pf_kstate_kill * kill,unsigned int * killed)5259 pf_killstates(struct pf_kstate_kill *kill, unsigned int *killed)
5260 {
5261 	struct pf_kstate	*s;
5262 
5263 	if (kill->psk_pfcmp.id) {
5264 		if (kill->psk_pfcmp.creatorid == 0)
5265 			kill->psk_pfcmp.creatorid = V_pf_status.hostid;
5266 		if ((s = pf_find_state_byid(kill->psk_pfcmp.id,
5267 		    kill->psk_pfcmp.creatorid))) {
5268 			pf_unlink_state(s, PF_ENTER_LOCKED);
5269 			*killed = 1;
5270 		}
5271 		return;
5272 	}
5273 
5274 	for (unsigned int i = 0; i <= pf_hashmask; i++)
5275 		*killed += pf_killstates_row(kill, &V_pf_idhash[i]);
5276 
5277 	return;
5278 }
5279 
5280 static int
pf_killstates_nv(struct pfioc_nv * nv)5281 pf_killstates_nv(struct pfioc_nv *nv)
5282 {
5283 	struct pf_kstate_kill	 kill;
5284 	nvlist_t		*nvl = NULL;
5285 	void			*nvlpacked = NULL;
5286 	int			 error = 0;
5287 	unsigned int		 killed = 0;
5288 
5289 #define ERROUT(x)	ERROUT_FUNCTION(on_error, x)
5290 
5291 	if (nv->len > pf_ioctl_maxcount)
5292 		ERROUT(ENOMEM);
5293 
5294 	nvlpacked = malloc(nv->len, M_NVLIST, M_WAITOK);
5295 	if (nvlpacked == NULL)
5296 		ERROUT(ENOMEM);
5297 
5298 	error = copyin(nv->data, nvlpacked, nv->len);
5299 	if (error)
5300 		ERROUT(error);
5301 
5302 	nvl = nvlist_unpack(nvlpacked, nv->len, 0);
5303 	if (nvl == NULL)
5304 		ERROUT(EBADMSG);
5305 
5306 	error = pf_nvstate_kill_to_kstate_kill(nvl, &kill);
5307 	if (error)
5308 		ERROUT(error);
5309 
5310 	pf_killstates(&kill, &killed);
5311 
5312 	free(nvlpacked, M_NVLIST);
5313 	nvlpacked = NULL;
5314 	nvlist_destroy(nvl);
5315 	nvl = nvlist_create(0);
5316 	if (nvl == NULL)
5317 		ERROUT(ENOMEM);
5318 
5319 	nvlist_add_number(nvl, "killed", killed);
5320 
5321 	nvlpacked = nvlist_pack(nvl, &nv->len);
5322 	if (nvlpacked == NULL)
5323 		ERROUT(ENOMEM);
5324 
5325 	if (nv->size == 0)
5326 		ERROUT(0);
5327 	else if (nv->size < nv->len)
5328 		ERROUT(ENOSPC);
5329 
5330 	error = copyout(nvlpacked, nv->data, nv->len);
5331 
5332 on_error:
5333 	nvlist_destroy(nvl);
5334 	free(nvlpacked, M_NVLIST);
5335 	return (error);
5336 }
5337 
5338 static int
pf_clearstates_nv(struct pfioc_nv * nv)5339 pf_clearstates_nv(struct pfioc_nv *nv)
5340 {
5341 	struct pf_kstate_kill	 kill;
5342 	nvlist_t		*nvl = NULL;
5343 	void			*nvlpacked = NULL;
5344 	int			 error = 0;
5345 	unsigned int		 killed;
5346 
5347 #define ERROUT(x)	ERROUT_FUNCTION(on_error, x)
5348 
5349 	if (nv->len > pf_ioctl_maxcount)
5350 		ERROUT(ENOMEM);
5351 
5352 	nvlpacked = malloc(nv->len, M_NVLIST, M_WAITOK);
5353 	if (nvlpacked == NULL)
5354 		ERROUT(ENOMEM);
5355 
5356 	error = copyin(nv->data, nvlpacked, nv->len);
5357 	if (error)
5358 		ERROUT(error);
5359 
5360 	nvl = nvlist_unpack(nvlpacked, nv->len, 0);
5361 	if (nvl == NULL)
5362 		ERROUT(EBADMSG);
5363 
5364 	error = pf_nvstate_kill_to_kstate_kill(nvl, &kill);
5365 	if (error)
5366 		ERROUT(error);
5367 
5368 	killed = pf_clear_states(&kill);
5369 
5370 	free(nvlpacked, M_NVLIST);
5371 	nvlpacked = NULL;
5372 	nvlist_destroy(nvl);
5373 	nvl = nvlist_create(0);
5374 	if (nvl == NULL)
5375 		ERROUT(ENOMEM);
5376 
5377 	nvlist_add_number(nvl, "killed", killed);
5378 
5379 	nvlpacked = nvlist_pack(nvl, &nv->len);
5380 	if (nvlpacked == NULL)
5381 		ERROUT(ENOMEM);
5382 
5383 	if (nv->size == 0)
5384 		ERROUT(0);
5385 	else if (nv->size < nv->len)
5386 		ERROUT(ENOSPC);
5387 
5388 	error = copyout(nvlpacked, nv->data, nv->len);
5389 
5390 #undef ERROUT
5391 on_error:
5392 	nvlist_destroy(nvl);
5393 	free(nvlpacked, M_NVLIST);
5394 	return (error);
5395 }
5396 
5397 static int
pf_getstate(struct pfioc_nv * nv)5398 pf_getstate(struct pfioc_nv *nv)
5399 {
5400 	nvlist_t		*nvl = NULL, *nvls;
5401 	void			*nvlpacked = NULL;
5402 	struct pf_kstate	*s = NULL;
5403 	int			 error = 0;
5404 	uint64_t		 id, creatorid;
5405 
5406 #define ERROUT(x)	ERROUT_FUNCTION(errout, x)
5407 
5408 	if (nv->len > pf_ioctl_maxcount)
5409 		ERROUT(ENOMEM);
5410 
5411 	nvlpacked = malloc(nv->len, M_NVLIST, M_WAITOK);
5412 	if (nvlpacked == NULL)
5413 		ERROUT(ENOMEM);
5414 
5415 	error = copyin(nv->data, nvlpacked, nv->len);
5416 	if (error)
5417 		ERROUT(error);
5418 
5419 	nvl = nvlist_unpack(nvlpacked, nv->len, 0);
5420 	if (nvl == NULL)
5421 		ERROUT(EBADMSG);
5422 
5423 	PFNV_CHK(pf_nvuint64(nvl, "id", &id));
5424 	PFNV_CHK(pf_nvuint64(nvl, "creatorid", &creatorid));
5425 
5426 	s = pf_find_state_byid(id, creatorid);
5427 	if (s == NULL)
5428 		ERROUT(ENOENT);
5429 
5430 	free(nvlpacked, M_NVLIST);
5431 	nvlpacked = NULL;
5432 	nvlist_destroy(nvl);
5433 	nvl = nvlist_create(0);
5434 	if (nvl == NULL)
5435 		ERROUT(ENOMEM);
5436 
5437 	nvls = pf_state_to_nvstate(s);
5438 	if (nvls == NULL)
5439 		ERROUT(ENOMEM);
5440 
5441 	nvlist_add_nvlist(nvl, "state", nvls);
5442 	nvlist_destroy(nvls);
5443 
5444 	nvlpacked = nvlist_pack(nvl, &nv->len);
5445 	if (nvlpacked == NULL)
5446 		ERROUT(ENOMEM);
5447 
5448 	if (nv->size == 0)
5449 		ERROUT(0);
5450 	else if (nv->size < nv->len)
5451 		ERROUT(ENOSPC);
5452 
5453 	error = copyout(nvlpacked, nv->data, nv->len);
5454 
5455 #undef ERROUT
5456 errout:
5457 	if (s != NULL)
5458 		PF_STATE_UNLOCK(s);
5459 	free(nvlpacked, M_NVLIST);
5460 	nvlist_destroy(nvl);
5461 	return (error);
5462 }
5463 
5464 /*
5465  * XXX - Check for version missmatch!!!
5466  */
5467 
5468 /*
5469  * Duplicate pfctl -Fa operation to get rid of as much as we can.
5470  */
5471 static int
shutdown_pf(void)5472 shutdown_pf(void)
5473 {
5474 	int error = 0;
5475 	u_int32_t t[5];
5476 	char nn = '\0';
5477 
5478 	do {
5479 		if ((error = pf_begin_rules(&t[0], PF_RULESET_SCRUB, &nn))
5480 		    != 0) {
5481 			DPFPRINTF(PF_DEBUG_MISC, ("shutdown_pf: SCRUB\n"));
5482 			break;
5483 		}
5484 		if ((error = pf_begin_rules(&t[1], PF_RULESET_FILTER, &nn))
5485 		    != 0) {
5486 			DPFPRINTF(PF_DEBUG_MISC, ("shutdown_pf: FILTER\n"));
5487 			break;		/* XXX: rollback? */
5488 		}
5489 		if ((error = pf_begin_rules(&t[2], PF_RULESET_NAT, &nn))
5490 		    != 0) {
5491 			DPFPRINTF(PF_DEBUG_MISC, ("shutdown_pf: NAT\n"));
5492 			break;		/* XXX: rollback? */
5493 		}
5494 		if ((error = pf_begin_rules(&t[3], PF_RULESET_BINAT, &nn))
5495 		    != 0) {
5496 			DPFPRINTF(PF_DEBUG_MISC, ("shutdown_pf: BINAT\n"));
5497 			break;		/* XXX: rollback? */
5498 		}
5499 		if ((error = pf_begin_rules(&t[4], PF_RULESET_RDR, &nn))
5500 		    != 0) {
5501 			DPFPRINTF(PF_DEBUG_MISC, ("shutdown_pf: RDR\n"));
5502 			break;		/* XXX: rollback? */
5503 		}
5504 
5505 		/* XXX: these should always succeed here */
5506 		pf_commit_rules(t[0], PF_RULESET_SCRUB, &nn);
5507 		pf_commit_rules(t[1], PF_RULESET_FILTER, &nn);
5508 		pf_commit_rules(t[2], PF_RULESET_NAT, &nn);
5509 		pf_commit_rules(t[3], PF_RULESET_BINAT, &nn);
5510 		pf_commit_rules(t[4], PF_RULESET_RDR, &nn);
5511 
5512 		if ((error = pf_clear_tables()) != 0)
5513 			break;
5514 
5515 #ifdef ALTQ
5516 		if ((error = pf_begin_altq(&t[0])) != 0) {
5517 			DPFPRINTF(PF_DEBUG_MISC, ("shutdown_pf: ALTQ\n"));
5518 			break;
5519 		}
5520 		pf_commit_altq(t[0]);
5521 #endif
5522 
5523 		pf_clear_all_states();
5524 
5525 		pf_clear_srcnodes(NULL);
5526 
5527 		/* status does not use malloced mem so no need to cleanup */
5528 		/* fingerprints and interfaces have their own cleanup code */
5529 	} while(0);
5530 
5531 	return (error);
5532 }
5533 
5534 static pfil_return_t
pf_check_return(int chk,struct mbuf ** m)5535 pf_check_return(int chk, struct mbuf **m)
5536 {
5537 
5538 	switch (chk) {
5539 	case PF_PASS:
5540 		if (*m == NULL)
5541 			return (PFIL_CONSUMED);
5542 		else
5543 			return (PFIL_PASS);
5544 		break;
5545 	default:
5546 		if (*m != NULL) {
5547 			m_freem(*m);
5548 			*m = NULL;
5549 		}
5550 		return (PFIL_DROPPED);
5551 	}
5552 }
5553 
5554 #ifdef INET
5555 static pfil_return_t
pf_check_in(struct mbuf ** m,struct ifnet * ifp,int flags,void * ruleset __unused,struct inpcb * inp)5556 pf_check_in(struct mbuf **m, struct ifnet *ifp, int flags,
5557     void *ruleset __unused, struct inpcb *inp)
5558 {
5559 	int chk;
5560 
5561 	chk = pf_test(PF_IN, flags, ifp, m, inp);
5562 
5563 	return (pf_check_return(chk, m));
5564 }
5565 
5566 static pfil_return_t
pf_check_out(struct mbuf ** m,struct ifnet * ifp,int flags,void * ruleset __unused,struct inpcb * inp)5567 pf_check_out(struct mbuf **m, struct ifnet *ifp, int flags,
5568     void *ruleset __unused,  struct inpcb *inp)
5569 {
5570 	int chk;
5571 
5572 	chk = pf_test(PF_OUT, flags, ifp, m, inp);
5573 
5574 	return (pf_check_return(chk, m));
5575 }
5576 #endif
5577 
5578 #ifdef INET6
5579 static pfil_return_t
pf_check6_in(struct mbuf ** m,struct ifnet * ifp,int flags,void * ruleset __unused,struct inpcb * inp)5580 pf_check6_in(struct mbuf **m, struct ifnet *ifp, int flags,
5581     void *ruleset __unused,  struct inpcb *inp)
5582 {
5583 	int chk;
5584 
5585 	/*
5586 	 * In case of loopback traffic IPv6 uses the real interface in
5587 	 * order to support scoped addresses. In order to support stateful
5588 	 * filtering we have change this to lo0 as it is the case in IPv4.
5589 	 */
5590 	CURVNET_SET(ifp->if_vnet);
5591 	chk = pf_test6(PF_IN, flags, (*m)->m_flags & M_LOOP ? V_loif : ifp, m, inp);
5592 	CURVNET_RESTORE();
5593 
5594 	return (pf_check_return(chk, m));
5595 }
5596 
5597 static pfil_return_t
pf_check6_out(struct mbuf ** m,struct ifnet * ifp,int flags,void * ruleset __unused,struct inpcb * inp)5598 pf_check6_out(struct mbuf **m, struct ifnet *ifp, int flags,
5599     void *ruleset __unused,  struct inpcb *inp)
5600 {
5601 	int chk;
5602 
5603 	CURVNET_SET(ifp->if_vnet);
5604 	chk = pf_test6(PF_OUT, flags, ifp, m, inp);
5605 	CURVNET_RESTORE();
5606 
5607 	return (pf_check_return(chk, m));
5608 }
5609 #endif /* INET6 */
5610 
5611 #ifdef INET
5612 VNET_DEFINE_STATIC(pfil_hook_t, pf_ip4_in_hook);
5613 VNET_DEFINE_STATIC(pfil_hook_t, pf_ip4_out_hook);
5614 #define	V_pf_ip4_in_hook	VNET(pf_ip4_in_hook)
5615 #define	V_pf_ip4_out_hook	VNET(pf_ip4_out_hook)
5616 #endif
5617 #ifdef INET6
5618 VNET_DEFINE_STATIC(pfil_hook_t, pf_ip6_in_hook);
5619 VNET_DEFINE_STATIC(pfil_hook_t, pf_ip6_out_hook);
5620 #define	V_pf_ip6_in_hook	VNET(pf_ip6_in_hook)
5621 #define	V_pf_ip6_out_hook	VNET(pf_ip6_out_hook)
5622 #endif
5623 
5624 static void
hook_pf(void)5625 hook_pf(void)
5626 {
5627 	struct pfil_hook_args pha;
5628 	struct pfil_link_args pla;
5629 	int ret;
5630 
5631 	if (V_pf_pfil_hooked)
5632 		return;
5633 
5634 	pha.pa_version = PFIL_VERSION;
5635 	pha.pa_modname = "pf";
5636 	pha.pa_ruleset = NULL;
5637 
5638 	pla.pa_version = PFIL_VERSION;
5639 
5640 #ifdef INET
5641 	pha.pa_type = PFIL_TYPE_IP4;
5642 	pha.pa_func = pf_check_in;
5643 	pha.pa_flags = PFIL_IN;
5644 	pha.pa_rulname = "default-in";
5645 	V_pf_ip4_in_hook = pfil_add_hook(&pha);
5646 	pla.pa_flags = PFIL_IN | PFIL_HEADPTR | PFIL_HOOKPTR;
5647 	pla.pa_head = V_inet_pfil_head;
5648 	pla.pa_hook = V_pf_ip4_in_hook;
5649 	ret = pfil_link(&pla);
5650 	MPASS(ret == 0);
5651 	pha.pa_func = pf_check_out;
5652 	pha.pa_flags = PFIL_OUT;
5653 	pha.pa_rulname = "default-out";
5654 	V_pf_ip4_out_hook = pfil_add_hook(&pha);
5655 	pla.pa_flags = PFIL_OUT | PFIL_HEADPTR | PFIL_HOOKPTR;
5656 	pla.pa_head = V_inet_pfil_head;
5657 	pla.pa_hook = V_pf_ip4_out_hook;
5658 	ret = pfil_link(&pla);
5659 	MPASS(ret == 0);
5660 #endif
5661 #ifdef INET6
5662 	pha.pa_type = PFIL_TYPE_IP6;
5663 	pha.pa_func = pf_check6_in;
5664 	pha.pa_flags = PFIL_IN;
5665 	pha.pa_rulname = "default-in6";
5666 	V_pf_ip6_in_hook = pfil_add_hook(&pha);
5667 	pla.pa_flags = PFIL_IN | PFIL_HEADPTR | PFIL_HOOKPTR;
5668 	pla.pa_head = V_inet6_pfil_head;
5669 	pla.pa_hook = V_pf_ip6_in_hook;
5670 	ret = pfil_link(&pla);
5671 	MPASS(ret == 0);
5672 	pha.pa_func = pf_check6_out;
5673 	pha.pa_rulname = "default-out6";
5674 	pha.pa_flags = PFIL_OUT;
5675 	V_pf_ip6_out_hook = pfil_add_hook(&pha);
5676 	pla.pa_flags = PFIL_OUT | PFIL_HEADPTR | PFIL_HOOKPTR;
5677 	pla.pa_head = V_inet6_pfil_head;
5678 	pla.pa_hook = V_pf_ip6_out_hook;
5679 	ret = pfil_link(&pla);
5680 	MPASS(ret == 0);
5681 #endif
5682 
5683 	V_pf_pfil_hooked = 1;
5684 }
5685 
5686 static void
dehook_pf(void)5687 dehook_pf(void)
5688 {
5689 
5690 	if (V_pf_pfil_hooked == 0)
5691 		return;
5692 
5693 #ifdef INET
5694 	pfil_remove_hook(V_pf_ip4_in_hook);
5695 	pfil_remove_hook(V_pf_ip4_out_hook);
5696 #endif
5697 #ifdef INET6
5698 	pfil_remove_hook(V_pf_ip6_in_hook);
5699 	pfil_remove_hook(V_pf_ip6_out_hook);
5700 #endif
5701 
5702 	V_pf_pfil_hooked = 0;
5703 }
5704 
5705 static void
pf_load_vnet(void)5706 pf_load_vnet(void)
5707 {
5708 	V_pf_tag_z = uma_zcreate("pf tags", sizeof(struct pf_tagname),
5709 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
5710 
5711 	pf_init_tagset(&V_pf_tags, &pf_rule_tag_hashsize,
5712 	    PF_RULE_TAG_HASH_SIZE_DEFAULT);
5713 #ifdef ALTQ
5714 	pf_init_tagset(&V_pf_qids, &pf_queue_tag_hashsize,
5715 	    PF_QUEUE_TAG_HASH_SIZE_DEFAULT);
5716 #endif
5717 
5718 	pfattach_vnet();
5719 	V_pf_vnet_active = 1;
5720 }
5721 
5722 static int
pf_load(void)5723 pf_load(void)
5724 {
5725 	int error;
5726 
5727 	rm_init_flags(&pf_rules_lock, "pf rulesets", RM_RECURSE);
5728 	sx_init(&pf_ioctl_lock, "pf ioctl");
5729 	sx_init(&pf_end_lock, "pf end thread");
5730 
5731 	pf_mtag_initialize();
5732 
5733 	pf_dev = make_dev(&pf_cdevsw, 0, UID_ROOT, GID_WHEEL, 0600, PF_NAME);
5734 	if (pf_dev == NULL)
5735 		return (ENOMEM);
5736 
5737 	pf_end_threads = 0;
5738 	error = kproc_create(pf_purge_thread, NULL, &pf_purge_proc, 0, 0, "pf purge");
5739 	if (error != 0)
5740 		return (error);
5741 
5742 	pfi_initialize();
5743 
5744 	return (0);
5745 }
5746 
5747 static void
pf_unload_vnet(void)5748 pf_unload_vnet(void)
5749 {
5750 	int ret;
5751 
5752 	V_pf_vnet_active = 0;
5753 	V_pf_status.running = 0;
5754 	dehook_pf();
5755 
5756 	PF_RULES_WLOCK();
5757 	pf_syncookies_cleanup();
5758 	shutdown_pf();
5759 	PF_RULES_WUNLOCK();
5760 
5761 	ret = swi_remove(V_pf_swi_cookie);
5762 	MPASS(ret == 0);
5763 	ret = intr_event_destroy(V_pf_swi_ie);
5764 	MPASS(ret == 0);
5765 
5766 	pf_unload_vnet_purge();
5767 
5768 	pf_normalize_cleanup();
5769 	PF_RULES_WLOCK();
5770 	pfi_cleanup_vnet();
5771 	PF_RULES_WUNLOCK();
5772 	pfr_cleanup();
5773 	pf_osfp_flush();
5774 	pf_cleanup();
5775 	if (IS_DEFAULT_VNET(curvnet))
5776 		pf_mtag_cleanup();
5777 
5778 	pf_cleanup_tagset(&V_pf_tags);
5779 #ifdef ALTQ
5780 	pf_cleanup_tagset(&V_pf_qids);
5781 #endif
5782 	uma_zdestroy(V_pf_tag_z);
5783 
5784 #ifdef PF_WANT_32_TO_64_COUNTER
5785 	PF_RULES_WLOCK();
5786 	LIST_REMOVE(V_pf_kifmarker, pfik_allkiflist);
5787 
5788 	MPASS(LIST_EMPTY(&V_pf_allkiflist));
5789 	MPASS(V_pf_allkifcount == 0);
5790 
5791 	LIST_REMOVE(&V_pf_default_rule, allrulelist);
5792 	V_pf_allrulecount--;
5793 	LIST_REMOVE(V_pf_rulemarker, allrulelist);
5794 
5795 	/*
5796 	 * There are known pf rule leaks when running the test suite.
5797 	 */
5798 #ifdef notyet
5799 	MPASS(LIST_EMPTY(&V_pf_allrulelist));
5800 	MPASS(V_pf_allrulecount == 0);
5801 #endif
5802 
5803 	PF_RULES_WUNLOCK();
5804 
5805 	free(V_pf_kifmarker, PFI_MTYPE);
5806 	free(V_pf_rulemarker, M_PFRULE);
5807 #endif
5808 
5809 	/* Free counters last as we updated them during shutdown. */
5810 	pf_counter_u64_deinit(&V_pf_default_rule.evaluations);
5811 	for (int i = 0; i < 2; i++) {
5812 		pf_counter_u64_deinit(&V_pf_default_rule.packets[i]);
5813 		pf_counter_u64_deinit(&V_pf_default_rule.bytes[i]);
5814 	}
5815 	counter_u64_free(V_pf_default_rule.states_cur);
5816 	counter_u64_free(V_pf_default_rule.states_tot);
5817 	counter_u64_free(V_pf_default_rule.src_nodes);
5818 
5819 	for (int i = 0; i < PFRES_MAX; i++)
5820 		counter_u64_free(V_pf_status.counters[i]);
5821 	for (int i = 0; i < KLCNT_MAX; i++)
5822 		counter_u64_free(V_pf_status.lcounters[i]);
5823 	for (int i = 0; i < FCNT_MAX; i++)
5824 		pf_counter_u64_deinit(&V_pf_status.fcounters[i]);
5825 	for (int i = 0; i < SCNT_MAX; i++)
5826 		counter_u64_free(V_pf_status.scounters[i]);
5827 }
5828 
5829 static void
pf_unload(void)5830 pf_unload(void)
5831 {
5832 
5833 	sx_xlock(&pf_end_lock);
5834 	pf_end_threads = 1;
5835 	while (pf_end_threads < 2) {
5836 		wakeup_one(pf_purge_thread);
5837 		sx_sleep(pf_purge_proc, &pf_end_lock, 0, "pftmo", 0);
5838 	}
5839 	sx_xunlock(&pf_end_lock);
5840 
5841 	if (pf_dev != NULL)
5842 		destroy_dev(pf_dev);
5843 
5844 	pfi_cleanup();
5845 
5846 	rm_destroy(&pf_rules_lock);
5847 	sx_destroy(&pf_ioctl_lock);
5848 	sx_destroy(&pf_end_lock);
5849 }
5850 
5851 static void
vnet_pf_init(void * unused __unused)5852 vnet_pf_init(void *unused __unused)
5853 {
5854 
5855 	pf_load_vnet();
5856 }
5857 VNET_SYSINIT(vnet_pf_init, SI_SUB_PROTO_FIREWALL, SI_ORDER_THIRD,
5858     vnet_pf_init, NULL);
5859 
5860 static void
vnet_pf_uninit(const void * unused __unused)5861 vnet_pf_uninit(const void *unused __unused)
5862 {
5863 
5864 	pf_unload_vnet();
5865 }
5866 SYSUNINIT(pf_unload, SI_SUB_PROTO_FIREWALL, SI_ORDER_SECOND, pf_unload, NULL);
5867 VNET_SYSUNINIT(vnet_pf_uninit, SI_SUB_PROTO_FIREWALL, SI_ORDER_THIRD,
5868     vnet_pf_uninit, NULL);
5869 
5870 static int
pf_modevent(module_t mod,int type,void * data)5871 pf_modevent(module_t mod, int type, void *data)
5872 {
5873 	int error = 0;
5874 
5875 	switch(type) {
5876 	case MOD_LOAD:
5877 		error = pf_load();
5878 		break;
5879 	case MOD_UNLOAD:
5880 		/* Handled in SYSUNINIT(pf_unload) to ensure it's done after
5881 		 * the vnet_pf_uninit()s */
5882 		break;
5883 	default:
5884 		error = EINVAL;
5885 		break;
5886 	}
5887 
5888 	return (error);
5889 }
5890 
5891 static moduledata_t pf_mod = {
5892 	"pf",
5893 	pf_modevent,
5894 	0
5895 };
5896 
5897 DECLARE_MODULE(pf, pf_mod, SI_SUB_PROTO_FIREWALL, SI_ORDER_SECOND);
5898 MODULE_VERSION(pf, PF_MODVER);
5899