xref: /titanic_44/usr/src/uts/common/inet/ip/ip_ire.c (revision 888e055994b8b0dc77b98c53dd97026237caec5d)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 /* Copyright (c) 1990 Mentat Inc. */
26 
27 /*
28  * This file contains routines that manipulate Internet Routing Entries (IREs).
29  */
30 
31 #include <sys/types.h>
32 #include <sys/stream.h>
33 #include <sys/stropts.h>
34 #include <sys/strsun.h>
35 #include <sys/strsubr.h>
36 #include <sys/ddi.h>
37 #include <sys/cmn_err.h>
38 #include <sys/policy.h>
39 
40 #include <sys/systm.h>
41 #include <sys/kmem.h>
42 #include <sys/param.h>
43 #include <sys/socket.h>
44 #include <net/if.h>
45 #include <net/route.h>
46 #include <netinet/in.h>
47 #include <net/if_dl.h>
48 #include <netinet/ip6.h>
49 #include <netinet/icmp6.h>
50 
51 #include <inet/common.h>
52 #include <inet/mi.h>
53 #include <inet/ip.h>
54 #include <inet/ip6.h>
55 #include <inet/ip_ndp.h>
56 #include <inet/arp.h>
57 #include <inet/ip_if.h>
58 #include <inet/ip_ire.h>
59 #include <inet/ip_ftable.h>
60 #include <inet/ip_rts.h>
61 #include <inet/nd.h>
62 
63 #include <inet/tcp.h>
64 #include <inet/ipclassifier.h>
65 #include <sys/zone.h>
66 #include <sys/cpuvar.h>
67 
68 #include <sys/tsol/label.h>
69 #include <sys/tsol/tnet.h>
70 
71 struct kmem_cache *rt_entry_cache;
72 
73 typedef struct nce_clookup_s {
74 	ipaddr_t ncecl_addr;
75 	boolean_t ncecl_found;
76 } nce_clookup_t;
77 
78 /*
79  * Synchronization notes:
80  *
81  * The fields of the ire_t struct are protected in the following way :
82  *
83  * ire_next/ire_ptpn
84  *
85  *	- bucket lock of the forwarding table in which is ire stored.
86  *
87  * ire_ill, ire_u *except* ire_gateway_addr[v6], ire_mask,
88  * ire_type, ire_create_time, ire_masklen, ire_ipversion, ire_flags,
89  * ire_bucket
90  *
91  *	- Set in ire_create_v4/v6 and never changes after that. Thus,
92  *	  we don't need a lock whenever these fields are accessed.
93  *
94  *	- ire_bucket and ire_masklen (also set in ire_create) is set in
95  *        ire_add before inserting in the bucket and never
96  *        changes after that. Thus we don't need a lock whenever these
97  *	  fields are accessed.
98  *
99  * ire_gateway_addr_v4[v6]
100  *
101  *	- ire_gateway_addr_v4[v6] is set during ire_create and later modified
102  *	  by rts_setgwr[v6]. As ire_gateway_addr is a uint32_t, updates to
103  *	  it assumed to be atomic and hence the other parts of the code
104  *	  does not use any locks. ire_gateway_addr_v6 updates are not atomic
105  *	  and hence any access to it uses ire_lock to get/set the right value.
106  *
107  * ire_refcnt, ire_identical_ref
108  *
109  *	- Updated atomically using atomic_add_32
110  *
111  * ire_ssthresh, ire_rtt_sd, ire_rtt, ire_ib_pkt_count, ire_ob_pkt_count
112  *
113  *	- Assumes that 32 bit writes are atomic. No locks. ire_lock is
114  *	  used to serialize updates to ire_ssthresh, ire_rtt_sd, ire_rtt.
115  *
116  * ire_generation
117  *	- Under ire_lock
118  *
119  * ire_nce_cache
120  *	- Under ire_lock
121  *
122  * ire_dep_parent (To next IRE in recursive lookup chain)
123  *	- Under ips_ire_dep_lock. Write held when modifying. Read held when
124  *	  walking. We also hold ire_lock when modifying to allow the data path
125  *	  to only acquire ire_lock.
126  *
127  * ire_dep_parent_generation (Generation number from ire_dep_parent)
128  *	- Under ips_ire_dep_lock and/or ire_lock. (A read claim on the dep_lock
129  *	  and ire_lock held when modifying)
130  *
131  * ire_dep_children (From parent to first child)
132  * ire_dep_sib_next (linked list of siblings)
133  * ire_dep_sib_ptpn (linked list of siblings)
134  *	- Under ips_ire_dep_lock. Write held when modifying. Read held when
135  *	  walking.
136  *
137  * As we always hold the bucket locks in all the places while accessing
138  * the above values, it is natural to use them for protecting them.
139  *
140  * We have a forwarding table for IPv4 and IPv6. The IPv6 forwarding table
141  * (ip_forwarding_table_v6) is an array of pointers to arrays of irb_t
142  * structures. ip_forwarding_table_v6 is allocated dynamically in
143  * ire_add_v6. ire_ft_init_lock is used to serialize multiple threads
144  * initializing the same bucket. Once a bucket is initialized, it is never
145  * de-alloacted. This assumption enables us to access
146  * ip_forwarding_table_v6[i] without any locks.
147  *
148  * The forwarding table for IPv4 is a radix tree whose leaves
149  * are rt_entry structures containing the irb_t for the rt_dst. The irb_t
150  * for IPv4 is dynamically allocated and freed.
151  *
152  * Each irb_t - ire bucket structure has a lock to protect
153  * a bucket and the ires residing in the bucket have a back pointer to
154  * the bucket structure. It also has a reference count for the number
155  * of threads walking the bucket - irb_refcnt which is bumped up
156  * using the irb_refhold function. The flags irb_marks can be
157  * set to IRB_MARK_CONDEMNED indicating that there are some ires
158  * in this bucket that are IRE_IS_CONDEMNED and the
159  * last thread to leave the bucket should delete the ires. Usually
160  * this is done by the irb_refrele function which is used to decrement
161  * the reference count on a bucket. See comments above irb_t structure
162  * definition in ip.h for further details.
163  *
164  * The ire_refhold/ire_refrele functions operate on the ire which increments/
165  * decrements the reference count, ire_refcnt, atomically on the ire.
166  * ire_refcnt is modified only using those functions. Operations on the IRE
167  * could be described as follows :
168  *
169  * CREATE an ire with reference count initialized to 1.
170  *
171  * ADDITION of an ire holds the bucket lock, checks for duplicates
172  * and then adds the ire. ire_add returns the ire after
173  * bumping up once more i.e the reference count is 2. This is to avoid
174  * an extra lookup in the functions calling ire_add which wants to
175  * work with the ire after adding.
176  *
177  * LOOKUP of an ire bumps up the reference count using ire_refhold
178  * function. It is valid to bump up the referece count of the IRE,
179  * after the lookup has returned an ire. Following are the lookup
180  * functions that return an HELD ire :
181  *
182  * ire_ftable_lookup[_v6], ire_lookup_multi_ill[_v6]
183  *
184  * DELETION of an ire holds the bucket lock, removes it from the list
185  * and then decrements the reference count for having removed from the list
186  * by using the ire_refrele function. If some other thread has looked up
187  * the ire, the reference count would have been bumped up and hence
188  * this ire will not be freed once deleted. It will be freed once the
189  * reference count drops to zero.
190  *
191  * Add and Delete acquires the bucket lock as RW_WRITER, while all the
192  * lookups acquire the bucket lock as RW_READER.
193  *
194  * The general rule is to do the ire_refrele in the function
195  * that is passing the ire as an argument.
196  *
197  * In trying to locate ires the following points are to be noted.
198  *
199  * IRE_IS_CONDEMNED signifies that the ire has been logically deleted and is
200  * to be ignored when walking the ires using ire_next.
201  *
202  * Zones note:
203  *	Walking IREs within a given zone also walks certain ires in other
204  *	zones.  This is done intentionally.  IRE walks with a specified
205  *	zoneid are used only when doing informational reports, and
206  *	zone users want to see things that they can access. See block
207  *	comment in ire_walk_ill_match().
208  */
209 
210 /*
211  * The size of the forwarding table.  We will make sure that it is a
212  * power of 2 in ip_ire_init().
213  * Setable in /etc/system
214  */
215 uint32_t ip6_ftable_hash_size = IP6_FTABLE_HASH_SIZE;
216 
217 struct	kmem_cache	*ire_cache;
218 struct	kmem_cache	*ncec_cache;
219 struct	kmem_cache	*nce_cache;
220 
221 static ire_t	ire_null;
222 
223 static ire_t	*ire_add_v4(ire_t *ire);
224 static void	ire_delete_v4(ire_t *ire);
225 static void	ire_dep_invalidate_children(ire_t *child);
226 static void	ire_walk_ipvers(pfv_t func, void *arg, uchar_t vers,
227     zoneid_t zoneid, ip_stack_t *);
228 static void	ire_walk_ill_ipvers(uint_t match_flags, uint_t ire_type,
229     pfv_t func, void *arg, uchar_t vers, ill_t *ill);
230 #ifdef DEBUG
231 static void	ire_trace_cleanup(const ire_t *);
232 #endif
233 
234 /*
235  * Following are the functions to increment/decrement the reference
236  * count of the IREs and IRBs (ire bucket).
237  *
238  * 1) We bump up the reference count of an IRE to make sure that
239  *    it does not get deleted and freed while we are using it.
240  *    Typically all the lookup functions hold the bucket lock,
241  *    and look for the IRE. If it finds an IRE, it bumps up the
242  *    reference count before dropping the lock. Sometimes we *may* want
243  *    to bump up the reference count after we *looked* up i.e without
244  *    holding the bucket lock. So, the ire_refhold function does not assert
245  *    on the bucket lock being held. Any thread trying to delete from
246  *    the hash bucket can still do so but cannot free the IRE if
247  *    ire_refcnt is not 0.
248  *
249  * 2) We bump up the reference count on the bucket where the IRE resides
250  *    (IRB), when we want to prevent the IREs getting deleted from a given
251  *    hash bucket. This makes life easier for ire_walk type functions which
252  *    wants to walk the IRE list, call a function, but needs to drop
253  *    the bucket lock to prevent recursive rw_enters. While the
254  *    lock is dropped, the list could be changed by other threads or
255  *    the same thread could end up deleting the ire or the ire pointed by
256  *    ire_next. ire_refholding the ire or ire_next is not sufficient as
257  *    a delete will still remove the ire from the bucket while we have
258  *    dropped the lock and hence the ire_next would be NULL. Thus, we
259  *    need a mechanism to prevent deletions from a given bucket.
260  *
261  *    To prevent deletions, we bump up the reference count on the
262  *    bucket. If the bucket is held, ire_delete just marks both
263  *    the ire and irb as CONDEMNED. When the
264  *    reference count on the bucket drops to zero, all the CONDEMNED ires
265  *    are deleted. We don't have to bump up the reference count on the
266  *    bucket if we are walking the bucket and never have to drop the bucket
267  *    lock. Note that irb_refhold does not prevent addition of new ires
268  *    in the list. It is okay because addition of new ires will not cause
269  *    ire_next to point to freed memory. We do irb_refhold only when
270  *    all of the 3 conditions are true :
271  *
272  *    1) The code needs to walk the IRE bucket from start to end.
273  *    2) It may have to drop the bucket lock sometimes while doing (1)
274  *    3) It does not want any ires to be deleted meanwhile.
275  */
276 
277 /*
278  * Bump up the reference count on the hash bucket - IRB to
279  * prevent ires from being deleted in this bucket.
280  */
281 void
282 irb_refhold(irb_t *irb)
283 {
284 	rw_enter(&irb->irb_lock, RW_WRITER);
285 	irb->irb_refcnt++;
286 	ASSERT(irb->irb_refcnt != 0);
287 	rw_exit(&irb->irb_lock);
288 }
289 
290 void
291 irb_refhold_locked(irb_t *irb)
292 {
293 	ASSERT(RW_WRITE_HELD(&irb->irb_lock));
294 	irb->irb_refcnt++;
295 	ASSERT(irb->irb_refcnt != 0);
296 }
297 
298 /*
299  * Note: when IRB_MARK_DYNAMIC is not set the irb_t
300  * is statically allocated, so that when the irb_refcnt goes to 0,
301  * we simply clean up the ire list and continue.
302  */
303 void
304 irb_refrele(irb_t *irb)
305 {
306 	if (irb->irb_marks & IRB_MARK_DYNAMIC) {
307 		irb_refrele_ftable(irb);
308 	} else {
309 		rw_enter(&irb->irb_lock, RW_WRITER);
310 		ASSERT(irb->irb_refcnt != 0);
311 		if (--irb->irb_refcnt	== 0 &&
312 		    (irb->irb_marks & IRB_MARK_CONDEMNED)) {
313 			ire_t *ire_list;
314 
315 			ire_list = ire_unlink(irb);
316 			rw_exit(&irb->irb_lock);
317 			ASSERT(ire_list != NULL);
318 			ire_cleanup(ire_list);
319 		} else {
320 			rw_exit(&irb->irb_lock);
321 		}
322 	}
323 }
324 
325 
326 /*
327  * Bump up the reference count on the IRE. We cannot assert that the
328  * bucket lock is being held as it is legal to bump up the reference
329  * count after the first lookup has returned the IRE without
330  * holding the lock.
331  */
332 void
333 ire_refhold(ire_t *ire)
334 {
335 	atomic_add_32(&(ire)->ire_refcnt, 1);
336 	ASSERT((ire)->ire_refcnt != 0);
337 #ifdef DEBUG
338 	ire_trace_ref(ire);
339 #endif
340 }
341 
342 void
343 ire_refhold_notr(ire_t *ire)
344 {
345 	atomic_add_32(&(ire)->ire_refcnt, 1);
346 	ASSERT((ire)->ire_refcnt != 0);
347 }
348 
349 void
350 ire_refhold_locked(ire_t *ire)
351 {
352 #ifdef DEBUG
353 	ire_trace_ref(ire);
354 #endif
355 	ire->ire_refcnt++;
356 }
357 
358 /*
359  * Release a ref on an IRE.
360  *
361  * Must not be called while holding any locks. Otherwise if this is
362  * the last reference to be released there is a chance of recursive mutex
363  * panic due to ire_refrele -> ipif_ill_refrele_tail -> qwriter_ip trying
364  * to restart an ioctl. The one exception is when the caller is sure that
365  * this is not the last reference to be released. Eg. if the caller is
366  * sure that the ire has not been deleted and won't be deleted.
367  *
368  * In architectures e.g sun4u, where atomic_add_32_nv is just
369  * a cas, we need to maintain the right memory barrier semantics
370  * as that of mutex_exit i.e all the loads and stores should complete
371  * before the cas is executed. membar_exit() does that here.
372  */
373 void
374 ire_refrele(ire_t *ire)
375 {
376 #ifdef DEBUG
377 	ire_untrace_ref(ire);
378 #endif
379 	ASSERT((ire)->ire_refcnt != 0);
380 	membar_exit();
381 	if (atomic_add_32_nv(&(ire)->ire_refcnt, -1) == 0)
382 		ire_inactive(ire);
383 }
384 
385 void
386 ire_refrele_notr(ire_t *ire)
387 {
388 	ASSERT((ire)->ire_refcnt != 0);
389 	membar_exit();
390 	if (atomic_add_32_nv(&(ire)->ire_refcnt, -1) == 0)
391 		ire_inactive(ire);
392 }
393 
394 /*
395  * This function is associated with the IP_IOC_IRE_DELETE[_NO_REPLY]
396  * IOCTL[s].  The NO_REPLY form is used by TCP to tell IP that it is
397  * having problems reaching a particular destination.
398  * This will make IP consider alternate routes (e.g., when there are
399  * muliple default routes), and it will also make IP discard any (potentially)
400  * stale redirect.
401  * Management processes may want to use the version that generates a reply.
402  *
403  * With the use of NUD like behavior for IPv4/ARP in addition to IPv6
404  * this function shouldn't be necessary for IP to recover from a bad redirect,
405  * a bad default router (when there are multiple default routers), or
406  * a stale ND/ARP entry. But we retain it in any case.
407  * For instance, this is helpful when TCP suspects a failure before NUD does.
408  */
409 int
410 ip_ire_delete(queue_t *q, mblk_t *mp, cred_t *ioc_cr)
411 {
412 	uchar_t		*addr_ucp;
413 	uint_t		ipversion;
414 	sin_t		*sin;
415 	sin6_t		*sin6;
416 	ipaddr_t	v4addr;
417 	in6_addr_t	v6addr;
418 	ire_t		*ire;
419 	ipid_t		*ipid;
420 	zoneid_t	zoneid;
421 	ip_stack_t	*ipst;
422 
423 	ASSERT(q->q_next == NULL);
424 	zoneid = IPCL_ZONEID(Q_TO_CONN(q));
425 	ipst = CONNQ_TO_IPST(q);
426 
427 	/*
428 	 * Check privilege using the ioctl credential; if it is NULL
429 	 * then this is a kernel message and therefor privileged.
430 	 */
431 	if (ioc_cr != NULL && secpolicy_ip_config(ioc_cr, B_FALSE) != 0)
432 		return (EPERM);
433 
434 	ipid = (ipid_t *)mp->b_rptr;
435 
436 	addr_ucp = mi_offset_param(mp, ipid->ipid_addr_offset,
437 	    ipid->ipid_addr_length);
438 	if (addr_ucp == NULL || !OK_32PTR(addr_ucp))
439 		return (EINVAL);
440 	switch (ipid->ipid_addr_length) {
441 	case sizeof (sin_t):
442 		/*
443 		 * got complete (sockaddr) address - increment addr_ucp to point
444 		 * at the ip_addr field.
445 		 */
446 		sin = (sin_t *)addr_ucp;
447 		addr_ucp = (uchar_t *)&sin->sin_addr.s_addr;
448 		ipversion = IPV4_VERSION;
449 		break;
450 	case sizeof (sin6_t):
451 		/*
452 		 * got complete (sockaddr) address - increment addr_ucp to point
453 		 * at the ip_addr field.
454 		 */
455 		sin6 = (sin6_t *)addr_ucp;
456 		addr_ucp = (uchar_t *)&sin6->sin6_addr;
457 		ipversion = IPV6_VERSION;
458 		break;
459 	default:
460 		return (EINVAL);
461 	}
462 	if (ipversion == IPV4_VERSION) {
463 		/* Extract the destination address. */
464 		bcopy(addr_ucp, &v4addr, IP_ADDR_LEN);
465 
466 		ire = ire_ftable_lookup_v4(v4addr, 0, 0, 0, NULL,
467 		    zoneid, NULL, MATCH_IRE_DSTONLY, 0, ipst, NULL);
468 	} else {
469 		/* Extract the destination address. */
470 		bcopy(addr_ucp, &v6addr, IPV6_ADDR_LEN);
471 
472 		ire = ire_ftable_lookup_v6(&v6addr, NULL, NULL, 0, NULL,
473 		    zoneid, NULL, MATCH_IRE_DSTONLY, 0, ipst, NULL);
474 	}
475 	if (ire != NULL) {
476 		if (ipversion == IPV4_VERSION) {
477 			ip_rts_change(RTM_LOSING, ire->ire_addr,
478 			    ire->ire_gateway_addr, ire->ire_mask,
479 			    (Q_TO_CONN(q))->conn_laddr_v4,  0, 0, 0,
480 			    (RTA_DST | RTA_GATEWAY | RTA_NETMASK | RTA_IFA),
481 			    ire->ire_ipst);
482 		}
483 		(void) ire_no_good(ire);
484 		ire_refrele(ire);
485 	}
486 	return (0);
487 }
488 
489 /*
490  * Initialize the ire that is specific to IPv4 part and call
491  * ire_init_common to finish it.
492  * Returns zero or errno.
493  */
494 int
495 ire_init_v4(ire_t *ire, uchar_t *addr, uchar_t *mask, uchar_t *gateway,
496     ushort_t type, ill_t *ill, zoneid_t zoneid, uint_t flags,
497     tsol_gc_t *gc, ip_stack_t *ipst)
498 {
499 	int error;
500 
501 	/*
502 	 * Reject IRE security attribute creation/initialization
503 	 * if system is not running in Trusted mode.
504 	 */
505 	if (gc != NULL && !is_system_labeled())
506 		return (EINVAL);
507 
508 	BUMP_IRE_STATS(ipst->ips_ire_stats_v4, ire_stats_alloced);
509 
510 	if (addr != NULL)
511 		bcopy(addr, &ire->ire_addr, IP_ADDR_LEN);
512 	if (gateway != NULL)
513 		bcopy(gateway, &ire->ire_gateway_addr, IP_ADDR_LEN);
514 
515 	/* Make sure we don't have stray values in some fields */
516 	switch (type) {
517 	case IRE_LOOPBACK:
518 		bcopy(&ire->ire_addr, &ire->ire_gateway_addr, IP_ADDR_LEN);
519 		/* FALLTHRU */
520 	case IRE_HOST:
521 	case IRE_BROADCAST:
522 	case IRE_LOCAL:
523 	case IRE_IF_CLONE:
524 		ire->ire_mask = IP_HOST_MASK;
525 		ire->ire_masklen = IPV4_ABITS;
526 		break;
527 	case IRE_PREFIX:
528 	case IRE_DEFAULT:
529 	case IRE_IF_RESOLVER:
530 	case IRE_IF_NORESOLVER:
531 		if (mask != NULL) {
532 			bcopy(mask, &ire->ire_mask, IP_ADDR_LEN);
533 			ire->ire_masklen = ip_mask_to_plen(ire->ire_mask);
534 		}
535 		break;
536 	case IRE_MULTICAST:
537 	case IRE_NOROUTE:
538 		ASSERT(mask == NULL);
539 		break;
540 	default:
541 		ASSERT(0);
542 		return (EINVAL);
543 	}
544 
545 	error = ire_init_common(ire, type, ill, zoneid, flags, IPV4_VERSION,
546 	    gc, ipst);
547 	if (error != NULL)
548 		return (error);
549 
550 	/* Determine which function pointers to use */
551 	ire->ire_postfragfn = ip_xmit;		/* Common case */
552 
553 	switch (ire->ire_type) {
554 	case IRE_LOCAL:
555 		ire->ire_sendfn = ire_send_local_v4;
556 		ire->ire_recvfn = ire_recv_local_v4;
557 #ifdef SO_VRRP
558 		ASSERT(ire->ire_ill != NULL);
559 		if (ire->ire_ill->ill_flags & ILLF_NOACCEPT) {
560 			ire->ire_noaccept = B_TRUE;
561 			ire->ire_recvfn = ire_recv_noaccept_v6;
562 		}
563 #endif
564 		break;
565 	case IRE_LOOPBACK:
566 		ire->ire_sendfn = ire_send_local_v4;
567 		ire->ire_recvfn = ire_recv_loopback_v4;
568 		break;
569 	case IRE_BROADCAST:
570 		ire->ire_postfragfn = ip_postfrag_loopcheck;
571 		ire->ire_sendfn = ire_send_broadcast_v4;
572 		ire->ire_recvfn = ire_recv_broadcast_v4;
573 		break;
574 	case IRE_MULTICAST:
575 		ire->ire_postfragfn = ip_postfrag_loopcheck;
576 		ire->ire_sendfn = ire_send_multicast_v4;
577 		ire->ire_recvfn = ire_recv_multicast_v4;
578 		break;
579 	default:
580 		/*
581 		 * For IRE_IF_ALL and IRE_OFFLINK we forward received
582 		 * packets by default.
583 		 */
584 		ire->ire_sendfn = ire_send_wire_v4;
585 		ire->ire_recvfn = ire_recv_forward_v4;
586 		break;
587 	}
588 	if (ire->ire_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
589 		ire->ire_sendfn = ire_send_noroute_v4;
590 		ire->ire_recvfn = ire_recv_noroute_v4;
591 	} else if (ire->ire_flags & RTF_MULTIRT) {
592 		ire->ire_postfragfn = ip_postfrag_multirt_v4;
593 		ire->ire_sendfn = ire_send_multirt_v4;
594 		/* Multirt receive of broadcast uses ire_recv_broadcast_v4 */
595 		if (ire->ire_type != IRE_BROADCAST)
596 			ire->ire_recvfn = ire_recv_multirt_v4;
597 	}
598 	ire->ire_nce_capable = ire_determine_nce_capable(ire);
599 	return (0);
600 }
601 
602 /*
603  * Determine ire_nce_capable
604  */
605 boolean_t
606 ire_determine_nce_capable(ire_t *ire)
607 {
608 	int max_masklen;
609 
610 	if ((ire->ire_flags & (RTF_REJECT|RTF_BLACKHOLE)) ||
611 	    (ire->ire_type & IRE_MULTICAST))
612 		return (B_TRUE);
613 
614 	if (ire->ire_ipversion == IPV4_VERSION)
615 		max_masklen = IPV4_ABITS;
616 	else
617 		max_masklen = IPV6_ABITS;
618 
619 	if ((ire->ire_type & IRE_ONLINK) && ire->ire_masklen == max_masklen)
620 		return (B_TRUE);
621 	return (B_FALSE);
622 }
623 
624 /*
625  * ire_create is called to allocate and initialize a new IRE.
626  *
627  * NOTE : This is called as writer sometimes though not required
628  * by this function.
629  */
630 ire_t *
631 ire_create(uchar_t *addr, uchar_t *mask, uchar_t *gateway,
632     ushort_t type, ill_t *ill, zoneid_t zoneid, uint_t flags, tsol_gc_t *gc,
633     ip_stack_t *ipst)
634 {
635 	ire_t	*ire;
636 	int	error;
637 
638 	ire = kmem_cache_alloc(ire_cache, KM_NOSLEEP);
639 	if (ire == NULL) {
640 		DTRACE_PROBE(kmem__cache__alloc);
641 		return (NULL);
642 	}
643 	*ire = ire_null;
644 
645 	error = ire_init_v4(ire, addr, mask, gateway, type, ill, zoneid, flags,
646 	    gc, ipst);
647 	if (error != 0) {
648 		DTRACE_PROBE2(ire__init, ire_t *, ire, int, error);
649 		kmem_cache_free(ire_cache, ire);
650 		return (NULL);
651 	}
652 	return (ire);
653 }
654 
655 /*
656  * Common to IPv4 and IPv6
657  * Returns zero or errno.
658  */
659 int
660 ire_init_common(ire_t *ire, ushort_t type, ill_t *ill, zoneid_t zoneid,
661     uint_t flags, uchar_t ipversion, tsol_gc_t *gc, ip_stack_t *ipst)
662 {
663 	int error;
664 
665 #ifdef DEBUG
666 	if (ill != NULL) {
667 		if (ill->ill_isv6)
668 			ASSERT(ipversion == IPV6_VERSION);
669 		else
670 			ASSERT(ipversion == IPV4_VERSION);
671 	}
672 #endif /* DEBUG */
673 
674 	/*
675 	 * Create/initialize IRE security attribute only in Trusted mode;
676 	 * if the passed in gc is non-NULL, we expect that the caller
677 	 * has held a reference to it and will release it when this routine
678 	 * returns a failure, otherwise we own the reference.  We do this
679 	 * prior to initializing the rest IRE fields.
680 	 */
681 	if (is_system_labeled()) {
682 		if ((type & (IRE_LOCAL | IRE_LOOPBACK | IRE_BROADCAST |
683 		    IRE_IF_ALL | IRE_MULTICAST | IRE_NOROUTE)) != 0) {
684 			/* release references on behalf of caller */
685 			if (gc != NULL)
686 				GC_REFRELE(gc);
687 		} else {
688 			error = tsol_ire_init_gwattr(ire, ipversion, gc);
689 			if (error != 0)
690 				return (error);
691 		}
692 	}
693 
694 	ire->ire_type = type;
695 	ire->ire_flags = RTF_UP | flags;
696 	ire->ire_create_time = (uint32_t)gethrestime_sec();
697 	ire->ire_generation = IRE_GENERATION_INITIAL;
698 
699 	/*
700 	 * The ill_ire_cnt isn't increased until
701 	 * the IRE is added to ensure that a walker will find
702 	 * all IREs that hold a reference on an ill.
703 	 *
704 	 * Note that ill_ire_multicast doesn't hold a ref on the ill since
705 	 * ire_add() is not called for the IRE_MULTICAST.
706 	 */
707 	ire->ire_ill = ill;
708 	ire->ire_zoneid = zoneid;
709 	ire->ire_ipversion = ipversion;
710 
711 	mutex_init(&ire->ire_lock, NULL, MUTEX_DEFAULT, NULL);
712 	ire->ire_refcnt = 1;
713 	ire->ire_identical_ref = 1;	/* Number of ire_delete's needed */
714 	ire->ire_ipst = ipst;	/* No netstack_hold */
715 	ire->ire_trace_disable = B_FALSE;
716 
717 	return (0);
718 }
719 
720 /*
721  * This creates an IRE_BROADCAST based on the arguments.
722  * A mirror is ire_lookup_bcast().
723  *
724  * Any supression of unneeded ones is done in ire_add_v4.
725  * We add one IRE_BROADCAST per address. ire_send_broadcast_v4()
726  * takes care of generating a loopback copy of the packet.
727  */
728 ire_t **
729 ire_create_bcast(ill_t *ill, ipaddr_t addr, zoneid_t zoneid, ire_t **irep)
730 {
731 	ip_stack_t	*ipst = ill->ill_ipst;
732 
733 	ASSERT(IAM_WRITER_ILL(ill));
734 
735 	*irep++ = ire_create(
736 	    (uchar_t *)&addr,			/* dest addr */
737 	    (uchar_t *)&ip_g_all_ones,		/* mask */
738 	    NULL,				/* no gateway */
739 	    IRE_BROADCAST,
740 	    ill,
741 	    zoneid,
742 	    RTF_KERNEL,
743 	    NULL,
744 	    ipst);
745 
746 	return (irep);
747 }
748 
749 /*
750  * This looks up an IRE_BROADCAST based on the arguments.
751  * Mirrors ire_create_bcast().
752  */
753 ire_t *
754 ire_lookup_bcast(ill_t *ill, ipaddr_t addr, zoneid_t zoneid)
755 {
756 	ire_t		*ire;
757 	int		match_args;
758 
759 	match_args = MATCH_IRE_TYPE | MATCH_IRE_ILL | MATCH_IRE_GW |
760 	    MATCH_IRE_MASK | MATCH_IRE_ZONEONLY;
761 
762 	if (IS_UNDER_IPMP(ill))
763 		match_args |= MATCH_IRE_TESTHIDDEN;
764 
765 	ire = ire_ftable_lookup_v4(
766 	    addr,				/* dest addr */
767 	    ip_g_all_ones,			/* mask */
768 	    0,					/* no gateway */
769 	    IRE_BROADCAST,
770 	    ill,
771 	    zoneid,
772 	    NULL,
773 	    match_args,
774 	    0,
775 	    ill->ill_ipst,
776 	    NULL);
777 	return (ire);
778 }
779 
780 /* Arrange to call the specified function for every IRE in the world. */
781 void
782 ire_walk(pfv_t func, void *arg, ip_stack_t *ipst)
783 {
784 	ire_walk_ipvers(func, arg, 0, ALL_ZONES, ipst);
785 }
786 
787 void
788 ire_walk_v4(pfv_t func, void *arg, zoneid_t zoneid, ip_stack_t *ipst)
789 {
790 	ire_walk_ipvers(func, arg, IPV4_VERSION, zoneid, ipst);
791 }
792 
793 void
794 ire_walk_v6(pfv_t func, void *arg, zoneid_t zoneid, ip_stack_t *ipst)
795 {
796 	ire_walk_ipvers(func, arg, IPV6_VERSION, zoneid, ipst);
797 }
798 
799 /*
800  * Walk a particular version. version == 0 means both v4 and v6.
801  */
802 static void
803 ire_walk_ipvers(pfv_t func, void *arg, uchar_t vers, zoneid_t zoneid,
804     ip_stack_t *ipst)
805 {
806 	if (vers != IPV6_VERSION) {
807 		/*
808 		 * ip_forwarding_table variable doesn't matter for IPv4 since
809 		 * ire_walk_ill_tables uses ips_ip_ftable for IPv4.
810 		 */
811 		ire_walk_ill_tables(0, 0, func, arg, IP_MASK_TABLE_SIZE,
812 		    0, NULL,
813 		    NULL, zoneid, ipst);
814 	}
815 	if (vers != IPV4_VERSION) {
816 		ire_walk_ill_tables(0, 0, func, arg, IP6_MASK_TABLE_SIZE,
817 		    ipst->ips_ip6_ftable_hash_size,
818 		    ipst->ips_ip_forwarding_table_v6,
819 		    NULL, zoneid, ipst);
820 	}
821 }
822 
823 /*
824  * Arrange to call the specified function for every IRE that matches the ill.
825  */
826 void
827 ire_walk_ill(uint_t match_flags, uint_t ire_type, pfv_t func, void *arg,
828     ill_t *ill)
829 {
830 	uchar_t vers = (ill->ill_isv6 ? IPV6_VERSION : IPV4_VERSION);
831 
832 	ire_walk_ill_ipvers(match_flags, ire_type, func, arg, vers, ill);
833 }
834 
835 /*
836  * Walk a particular ill and version.
837  */
838 static void
839 ire_walk_ill_ipvers(uint_t match_flags, uint_t ire_type, pfv_t func,
840     void *arg, uchar_t vers, ill_t *ill)
841 {
842 	ip_stack_t	*ipst = ill->ill_ipst;
843 
844 	if (vers == IPV4_VERSION) {
845 		ire_walk_ill_tables(match_flags, ire_type, func, arg,
846 		    IP_MASK_TABLE_SIZE,
847 		    0, NULL,
848 		    ill, ALL_ZONES, ipst);
849 	}
850 	if (vers != IPV4_VERSION) {
851 		ire_walk_ill_tables(match_flags, ire_type, func, arg,
852 		    IP6_MASK_TABLE_SIZE, ipst->ips_ip6_ftable_hash_size,
853 		    ipst->ips_ip_forwarding_table_v6,
854 		    ill, ALL_ZONES, ipst);
855 	}
856 }
857 
858 /*
859  * Do the specific matching of IREs to shared-IP zones.
860  *
861  * We have the same logic as in ire_match_args but implemented slightly
862  * differently.
863  */
864 boolean_t
865 ire_walk_ill_match(uint_t match_flags, uint_t ire_type, ire_t *ire,
866     ill_t *ill, zoneid_t zoneid, ip_stack_t *ipst)
867 {
868 	ill_t *dst_ill = NULL;
869 
870 	ASSERT(match_flags != 0 || zoneid != ALL_ZONES);
871 	if (match_flags & MATCH_IRE_ILL) {
872 		dst_ill = ire->ire_ill;
873 	}
874 
875 	if (zoneid != ALL_ZONES && zoneid != ire->ire_zoneid &&
876 	    ire->ire_zoneid != ALL_ZONES) {
877 		/*
878 		 * We're walking the IREs for a specific zone. The only relevant
879 		 * IREs are:
880 		 * - all IREs with a matching ire_zoneid
881 		 * - IRE_IF_ALL IREs for interfaces with a usable source addr
882 		 *   with a matching zone
883 		 * - IRE_OFFLINK with a gateway reachable from the zone
884 		 * Note that ealier we only did the IRE_OFFLINK check for
885 		 * IRE_DEFAULT (and only when we had multiple IRE_DEFAULTs).
886 		 */
887 		dst_ill = ire->ire_ill;
888 
889 		if (ire->ire_type & IRE_ONLINK) {
890 			uint_t	ifindex;
891 
892 			/*
893 			 * Note there is no IRE_INTERFACE on vniN thus
894 			 * can't do an IRE lookup for a matching route.
895 			 */
896 			ifindex = dst_ill->ill_usesrc_ifindex;
897 			if (ifindex == 0)
898 				return (B_FALSE);
899 
900 			/*
901 			 * If there is a usable source address in the
902 			 * zone, then it's ok to return an
903 			 * IRE_INTERFACE
904 			 */
905 			if (!ipif_zone_avail(ifindex, dst_ill->ill_isv6,
906 			    zoneid, ipst)) {
907 				return (B_FALSE);
908 			}
909 		}
910 
911 		if (dst_ill != NULL && (ire->ire_type & IRE_OFFLINK)) {
912 			ipif_t	*tipif;
913 
914 			mutex_enter(&dst_ill->ill_lock);
915 			for (tipif = dst_ill->ill_ipif;
916 			    tipif != NULL; tipif = tipif->ipif_next) {
917 				if (!IPIF_IS_CONDEMNED(tipif) &&
918 				    (tipif->ipif_flags & IPIF_UP) &&
919 				    (tipif->ipif_zoneid == zoneid ||
920 				    tipif->ipif_zoneid == ALL_ZONES))
921 					break;
922 			}
923 			mutex_exit(&dst_ill->ill_lock);
924 			if (tipif == NULL) {
925 				return (B_FALSE);
926 			}
927 		}
928 
929 		/*
930 		 * Match all offlink routes from the global zone, irrespective
931 		 * of reachability. For a non-global zone only match those
932 		 * where ire_gateway_addr has an IRE_INTERFACE for the zoneid.
933 		 */
934 		if ((ire->ire_type & IRE_OFFLINK) && zoneid != GLOBAL_ZONEID &&
935 		    zoneid != ALL_ZONES) {
936 			in6_addr_t gw_addr_v6;
937 
938 			if (ire->ire_ipversion == IPV4_VERSION) {
939 				if (!ire_gateway_ok_zone_v4(
940 				    ire->ire_gateway_addr, zoneid,
941 				    dst_ill, NULL, ipst, B_FALSE))
942 					return (B_FALSE);
943 			} else {
944 				ASSERT(ire->ire_ipversion == IPV6_VERSION);
945 				mutex_enter(&ire->ire_lock);
946 				gw_addr_v6 = ire->ire_gateway_addr_v6;
947 				mutex_exit(&ire->ire_lock);
948 
949 				if (!ire_gateway_ok_zone_v6(&gw_addr_v6, zoneid,
950 				    dst_ill, NULL, ipst, B_FALSE))
951 					return (B_FALSE);
952 			}
953 		}
954 	}
955 
956 	if (((!(match_flags & MATCH_IRE_TYPE)) ||
957 	    (ire->ire_type & ire_type)) &&
958 	    ((!(match_flags & MATCH_IRE_ILL)) ||
959 	    (dst_ill == ill ||
960 	    dst_ill != NULL && IS_IN_SAME_ILLGRP(dst_ill, ill)))) {
961 		return (B_TRUE);
962 	}
963 	return (B_FALSE);
964 }
965 
966 int
967 rtfunc(struct radix_node *rn, void *arg)
968 {
969 	struct rtfuncarg *rtf = arg;
970 	struct rt_entry *rt;
971 	irb_t *irb;
972 	ire_t *ire;
973 	boolean_t ret;
974 
975 	rt = (struct rt_entry *)rn;
976 	ASSERT(rt != NULL);
977 	irb = &rt->rt_irb;
978 	for (ire = irb->irb_ire; ire != NULL; ire = ire->ire_next) {
979 		if ((rtf->rt_match_flags != 0) ||
980 		    (rtf->rt_zoneid != ALL_ZONES)) {
981 			ret = ire_walk_ill_match(rtf->rt_match_flags,
982 			    rtf->rt_ire_type, ire,
983 			    rtf->rt_ill, rtf->rt_zoneid, rtf->rt_ipst);
984 		} else {
985 			ret = B_TRUE;
986 		}
987 		if (ret)
988 			(*rtf->rt_func)(ire, rtf->rt_arg);
989 	}
990 	return (0);
991 }
992 
993 /*
994  * Walk the ftable entries that match the ill.
995  */
996 void
997 ire_walk_ill_tables(uint_t match_flags, uint_t ire_type, pfv_t func,
998     void *arg, size_t ftbl_sz, size_t htbl_sz, irb_t **ipftbl,
999     ill_t *ill, zoneid_t zoneid,
1000     ip_stack_t *ipst)
1001 {
1002 	irb_t	*irb_ptr;
1003 	irb_t	*irb;
1004 	ire_t	*ire;
1005 	int i, j;
1006 	boolean_t ret;
1007 	struct rtfuncarg rtfarg;
1008 
1009 	ASSERT((!(match_flags & MATCH_IRE_ILL)) || (ill != NULL));
1010 	ASSERT(!(match_flags & MATCH_IRE_TYPE) || (ire_type != 0));
1011 
1012 	/* knobs such that routine is called only for v6 case */
1013 	if (ipftbl == ipst->ips_ip_forwarding_table_v6) {
1014 		for (i = (ftbl_sz - 1);  i >= 0; i--) {
1015 			if ((irb_ptr = ipftbl[i]) == NULL)
1016 				continue;
1017 			for (j = 0; j < htbl_sz; j++) {
1018 				irb = &irb_ptr[j];
1019 				if (irb->irb_ire == NULL)
1020 					continue;
1021 
1022 				irb_refhold(irb);
1023 				for (ire = irb->irb_ire; ire != NULL;
1024 				    ire = ire->ire_next) {
1025 					if (match_flags == 0 &&
1026 					    zoneid == ALL_ZONES) {
1027 						ret = B_TRUE;
1028 					} else {
1029 						ret =
1030 						    ire_walk_ill_match(
1031 						    match_flags,
1032 						    ire_type, ire, ill,
1033 						    zoneid, ipst);
1034 					}
1035 					if (ret)
1036 						(*func)(ire, arg);
1037 				}
1038 				irb_refrele(irb);
1039 			}
1040 		}
1041 	} else {
1042 		(void) memset(&rtfarg, 0, sizeof (rtfarg));
1043 		rtfarg.rt_func = func;
1044 		rtfarg.rt_arg = arg;
1045 		if (match_flags != 0) {
1046 			rtfarg.rt_match_flags = match_flags;
1047 		}
1048 		rtfarg.rt_ire_type = ire_type;
1049 		rtfarg.rt_ill = ill;
1050 		rtfarg.rt_zoneid = zoneid;
1051 		rtfarg.rt_ipst = ipst;	/* No netstack_hold */
1052 		(void) ipst->ips_ip_ftable->rnh_walktree_mt(
1053 		    ipst->ips_ip_ftable,
1054 		    rtfunc, &rtfarg, irb_refhold_rn, irb_refrele_rn);
1055 	}
1056 }
1057 
1058 /*
1059  * This function takes a mask and returns
1060  * number of bits set in the mask. If no
1061  * bit is set it returns 0.
1062  * Assumes a contiguous mask.
1063  */
1064 int
1065 ip_mask_to_plen(ipaddr_t mask)
1066 {
1067 	return (mask == 0 ? 0 : IP_ABITS - (ffs(ntohl(mask)) -1));
1068 }
1069 
1070 /*
1071  * Convert length for a mask to the mask.
1072  */
1073 ipaddr_t
1074 ip_plen_to_mask(uint_t masklen)
1075 {
1076 	if (masklen == 0)
1077 		return (0);
1078 
1079 	return (htonl(IP_HOST_MASK << (IP_ABITS - masklen)));
1080 }
1081 
1082 void
1083 ire_atomic_end(irb_t *irb_ptr, ire_t *ire)
1084 {
1085 	ill_t		*ill;
1086 
1087 	ill = ire->ire_ill;
1088 	if (ill != NULL)
1089 		mutex_exit(&ill->ill_lock);
1090 	rw_exit(&irb_ptr->irb_lock);
1091 }
1092 
1093 /*
1094  * ire_add_v[46] atomically make sure that the ill associated
1095  * with the new ire is not going away i.e., we check ILL_CONDEMNED.
1096  */
1097 int
1098 ire_atomic_start(irb_t *irb_ptr, ire_t *ire)
1099 {
1100 	ill_t		*ill;
1101 
1102 	ill = ire->ire_ill;
1103 
1104 	rw_enter(&irb_ptr->irb_lock, RW_WRITER);
1105 	if (ill != NULL) {
1106 		mutex_enter(&ill->ill_lock);
1107 
1108 		/*
1109 		 * Don't allow IRE's to be created on dying ills.
1110 		 */
1111 		if (ill->ill_state_flags & ILL_CONDEMNED) {
1112 			ire_atomic_end(irb_ptr, ire);
1113 			return (ENXIO);
1114 		}
1115 
1116 		if (IS_UNDER_IPMP(ill)) {
1117 			int	error = 0;
1118 			mutex_enter(&ill->ill_phyint->phyint_lock);
1119 			if (!ipmp_ill_is_active(ill) &&
1120 			    IRE_HIDDEN_TYPE(ire->ire_type) &&
1121 			    !ire->ire_testhidden) {
1122 				error = EINVAL;
1123 			}
1124 			mutex_exit(&ill->ill_phyint->phyint_lock);
1125 			if (error != 0) {
1126 				ire_atomic_end(irb_ptr, ire);
1127 				return (error);
1128 			}
1129 		}
1130 
1131 	}
1132 	return (0);
1133 }
1134 
1135 /*
1136  * Add a fully initialized IRE to the forwarding table.
1137  * This returns NULL on failure, or a held IRE on success.
1138  * Normally the returned IRE is the same as the argument. But a different
1139  * IRE will be returned if the added IRE is deemed identical to an existing
1140  * one. In that case ire_identical_ref will be increased.
1141  * The caller always needs to do an ire_refrele() on the returned IRE.
1142  */
1143 ire_t *
1144 ire_add(ire_t *ire)
1145 {
1146 	if (IRE_HIDDEN_TYPE(ire->ire_type) &&
1147 	    ire->ire_ill != NULL && IS_UNDER_IPMP(ire->ire_ill)) {
1148 		/*
1149 		 * IREs hosted on interfaces that are under IPMP
1150 		 * should be hidden so that applications don't
1151 		 * accidentally end up sending packets with test
1152 		 * addresses as their source addresses, or
1153 		 * sending out interfaces that are e.g. IFF_INACTIVE.
1154 		 * Hide them here.
1155 		 */
1156 		ire->ire_testhidden = B_TRUE;
1157 	}
1158 
1159 	if (ire->ire_ipversion == IPV6_VERSION)
1160 		return (ire_add_v6(ire));
1161 	else
1162 		return (ire_add_v4(ire));
1163 }
1164 
1165 /*
1166  * Add a fully initialized IPv4 IRE to the forwarding table.
1167  * This returns NULL on failure, or a held IRE on success.
1168  * Normally the returned IRE is the same as the argument. But a different
1169  * IRE will be returned if the added IRE is deemed identical to an existing
1170  * one. In that case ire_identical_ref will be increased.
1171  * The caller always needs to do an ire_refrele() on the returned IRE.
1172  */
1173 static ire_t *
1174 ire_add_v4(ire_t *ire)
1175 {
1176 	ire_t	*ire1;
1177 	irb_t	*irb_ptr;
1178 	ire_t	**irep;
1179 	int	match_flags;
1180 	int	error;
1181 	ip_stack_t	*ipst = ire->ire_ipst;
1182 
1183 	if (ire->ire_ill != NULL)
1184 		ASSERT(!MUTEX_HELD(&ire->ire_ill->ill_lock));
1185 	ASSERT(ire->ire_ipversion == IPV4_VERSION);
1186 
1187 	/* Make sure the address is properly masked. */
1188 	ire->ire_addr &= ire->ire_mask;
1189 
1190 	match_flags = (MATCH_IRE_MASK | MATCH_IRE_TYPE | MATCH_IRE_GW);
1191 
1192 	if (ire->ire_ill != NULL) {
1193 		match_flags |= MATCH_IRE_ILL;
1194 	}
1195 	irb_ptr = ire_get_bucket(ire);
1196 	if (irb_ptr == NULL) {
1197 		printf("no bucket for %p\n", (void *)ire);
1198 		ire_delete(ire);
1199 		return (NULL);
1200 	}
1201 
1202 	/*
1203 	 * Start the atomic add of the ire. Grab the ill lock,
1204 	 * the bucket lock. Check for condemned.
1205 	 */
1206 	error = ire_atomic_start(irb_ptr, ire);
1207 	if (error != 0) {
1208 		printf("no ire_atomic_start for %p\n", (void *)ire);
1209 		ire_delete(ire);
1210 		irb_refrele(irb_ptr);
1211 		return (NULL);
1212 	}
1213 	/*
1214 	 * If we are creating a hidden IRE, make sure we search for
1215 	 * hidden IREs when searching for duplicates below.
1216 	 * Otherwise, we might find an IRE on some other interface
1217 	 * that's not marked hidden.
1218 	 */
1219 	if (ire->ire_testhidden)
1220 		match_flags |= MATCH_IRE_TESTHIDDEN;
1221 
1222 	/*
1223 	 * Atomically check for duplicate and insert in the table.
1224 	 */
1225 	for (ire1 = irb_ptr->irb_ire; ire1 != NULL; ire1 = ire1->ire_next) {
1226 		if (IRE_IS_CONDEMNED(ire1))
1227 			continue;
1228 		/*
1229 		 * Here we need an exact match on zoneid, i.e.,
1230 		 * ire_match_args doesn't fit.
1231 		 */
1232 		if (ire1->ire_zoneid != ire->ire_zoneid)
1233 			continue;
1234 
1235 		if (ire1->ire_type != ire->ire_type)
1236 			continue;
1237 
1238 		/*
1239 		 * Note: We do not allow multiple routes that differ only
1240 		 * in the gateway security attributes; such routes are
1241 		 * considered duplicates.
1242 		 * To change that we explicitly have to treat them as
1243 		 * different here.
1244 		 */
1245 		if (ire_match_args(ire1, ire->ire_addr, ire->ire_mask,
1246 		    ire->ire_gateway_addr, ire->ire_type, ire->ire_ill,
1247 		    ire->ire_zoneid, NULL, match_flags)) {
1248 			/*
1249 			 * Return the old ire after doing a REFHOLD.
1250 			 * As most of the callers continue to use the IRE
1251 			 * after adding, we return a held ire. This will
1252 			 * avoid a lookup in the caller again. If the callers
1253 			 * don't want to use it, they need to do a REFRELE.
1254 			 */
1255 			atomic_add_32(&ire1->ire_identical_ref, 1);
1256 			DTRACE_PROBE2(ire__add__exist, ire_t *, ire1,
1257 			    ire_t *, ire);
1258 			ire_refhold(ire1);
1259 			ire_atomic_end(irb_ptr, ire);
1260 			ire_delete(ire);
1261 			irb_refrele(irb_ptr);
1262 			return (ire1);
1263 		}
1264 	}
1265 
1266 	/*
1267 	 * Normally we do head insertion since most things do not care about
1268 	 * the order of the IREs in the bucket. Note that ip_cgtp_bcast_add
1269 	 * assumes we at least do head insertion so that its IRE_BROADCAST
1270 	 * arrive ahead of existing IRE_HOST for the same address.
1271 	 * However, due to shared-IP zones (and restrict_interzone_loopback)
1272 	 * we can have an IRE_LOCAL as well as IRE_IF_CLONE for the same
1273 	 * address. For that reason we do tail insertion for IRE_IF_CLONE.
1274 	 * Due to the IRE_BROADCAST on cgtp0, which must be last in the bucket,
1275 	 * we do tail insertion of IRE_BROADCASTs that do not have RTF_MULTIRT
1276 	 * set.
1277 	 */
1278 	irep = (ire_t **)irb_ptr;
1279 	if ((ire->ire_type & IRE_IF_CLONE) ||
1280 	    ((ire->ire_type & IRE_BROADCAST) &&
1281 	    !(ire->ire_flags & RTF_MULTIRT))) {
1282 		while ((ire1 = *irep) != NULL)
1283 			irep = &ire1->ire_next;
1284 	}
1285 	/* Insert at *irep */
1286 	ire1 = *irep;
1287 	if (ire1 != NULL)
1288 		ire1->ire_ptpn = &ire->ire_next;
1289 	ire->ire_next = ire1;
1290 	/* Link the new one in. */
1291 	ire->ire_ptpn = irep;
1292 
1293 	/*
1294 	 * ire_walk routines de-reference ire_next without holding
1295 	 * a lock. Before we point to the new ire, we want to make
1296 	 * sure the store that sets the ire_next of the new ire
1297 	 * reaches global visibility, so that ire_walk routines
1298 	 * don't see a truncated list of ires i.e if the ire_next
1299 	 * of the new ire gets set after we do "*irep = ire" due
1300 	 * to re-ordering, the ire_walk thread will see a NULL
1301 	 * once it accesses the ire_next of the new ire.
1302 	 * membar_producer() makes sure that the following store
1303 	 * happens *after* all of the above stores.
1304 	 */
1305 	membar_producer();
1306 	*irep = ire;
1307 	ire->ire_bucket = irb_ptr;
1308 	/*
1309 	 * We return a bumped up IRE above. Keep it symmetrical
1310 	 * so that the callers will always have to release. This
1311 	 * helps the callers of this function because they continue
1312 	 * to use the IRE after adding and hence they don't have to
1313 	 * lookup again after we return the IRE.
1314 	 *
1315 	 * NOTE : We don't have to use atomics as this is appearing
1316 	 * in the list for the first time and no one else can bump
1317 	 * up the reference count on this yet.
1318 	 */
1319 	ire_refhold_locked(ire);
1320 	BUMP_IRE_STATS(ipst->ips_ire_stats_v4, ire_stats_inserted);
1321 
1322 	irb_ptr->irb_ire_cnt++;
1323 	if (irb_ptr->irb_marks & IRB_MARK_DYNAMIC)
1324 		irb_ptr->irb_nire++;
1325 
1326 	if (ire->ire_ill != NULL) {
1327 		ire->ire_ill->ill_ire_cnt++;
1328 		ASSERT(ire->ire_ill->ill_ire_cnt != 0);	/* Wraparound */
1329 	}
1330 
1331 	ire_atomic_end(irb_ptr, ire);
1332 
1333 	/* Make any caching of the IREs be notified or updated */
1334 	ire_flush_cache_v4(ire, IRE_FLUSH_ADD);
1335 
1336 	if (ire->ire_ill != NULL)
1337 		ASSERT(!MUTEX_HELD(&ire->ire_ill->ill_lock));
1338 	irb_refrele(irb_ptr);
1339 	return (ire);
1340 }
1341 
1342 /*
1343  * irb_refrele is the only caller of the function. ire_unlink calls to
1344  * do the final cleanup for this ire.
1345  */
1346 void
1347 ire_cleanup(ire_t *ire)
1348 {
1349 	ire_t *ire_next;
1350 	ip_stack_t *ipst = ire->ire_ipst;
1351 
1352 	ASSERT(ire != NULL);
1353 
1354 	while (ire != NULL) {
1355 		ire_next = ire->ire_next;
1356 		if (ire->ire_ipversion == IPV4_VERSION) {
1357 			ire_delete_v4(ire);
1358 			BUMP_IRE_STATS(ipst->ips_ire_stats_v4,
1359 			    ire_stats_deleted);
1360 		} else {
1361 			ASSERT(ire->ire_ipversion == IPV6_VERSION);
1362 			ire_delete_v6(ire);
1363 			BUMP_IRE_STATS(ipst->ips_ire_stats_v6,
1364 			    ire_stats_deleted);
1365 		}
1366 		/*
1367 		 * Now it's really out of the list. Before doing the
1368 		 * REFRELE, set ire_next to NULL as ire_inactive asserts
1369 		 * so.
1370 		 */
1371 		ire->ire_next = NULL;
1372 		ire_refrele_notr(ire);
1373 		ire = ire_next;
1374 	}
1375 }
1376 
1377 /*
1378  * irb_refrele is the only caller of the function. It calls to unlink
1379  * all the CONDEMNED ires from this bucket.
1380  */
1381 ire_t *
1382 ire_unlink(irb_t *irb)
1383 {
1384 	ire_t *ire;
1385 	ire_t *ire1;
1386 	ire_t **ptpn;
1387 	ire_t *ire_list = NULL;
1388 
1389 	ASSERT(RW_WRITE_HELD(&irb->irb_lock));
1390 	ASSERT(((irb->irb_marks & IRB_MARK_DYNAMIC) && irb->irb_refcnt == 1) ||
1391 	    (irb->irb_refcnt == 0));
1392 	ASSERT(irb->irb_marks & IRB_MARK_CONDEMNED);
1393 	ASSERT(irb->irb_ire != NULL);
1394 
1395 	for (ire = irb->irb_ire; ire != NULL; ire = ire1) {
1396 		ire1 = ire->ire_next;
1397 		if (IRE_IS_CONDEMNED(ire)) {
1398 			ptpn = ire->ire_ptpn;
1399 			ire1 = ire->ire_next;
1400 			if (ire1)
1401 				ire1->ire_ptpn = ptpn;
1402 			*ptpn = ire1;
1403 			ire->ire_ptpn = NULL;
1404 			ire->ire_next = NULL;
1405 
1406 			/*
1407 			 * We need to call ire_delete_v4 or ire_delete_v6 to
1408 			 * clean up dependents and the redirects pointing at
1409 			 * the default gateway. We need to drop the lock
1410 			 * as ire_flush_cache/ire_delete_host_redircts require
1411 			 * so. But we can't drop the lock, as ire_unlink needs
1412 			 * to atomically remove the ires from the list.
1413 			 * So, create a temporary list of CONDEMNED ires
1414 			 * for doing ire_delete_v4/ire_delete_v6 operations
1415 			 * later on.
1416 			 */
1417 			ire->ire_next = ire_list;
1418 			ire_list = ire;
1419 		}
1420 	}
1421 	irb->irb_marks &= ~IRB_MARK_CONDEMNED;
1422 	return (ire_list);
1423 }
1424 
1425 /*
1426  * Clean up the radix node for this ire. Must be called by irb_refrele
1427  * when there are no ire's left in the bucket. Returns TRUE if the bucket
1428  * is deleted and freed.
1429  */
1430 boolean_t
1431 irb_inactive(irb_t *irb)
1432 {
1433 	struct rt_entry *rt;
1434 	struct radix_node *rn;
1435 	ip_stack_t *ipst = irb->irb_ipst;
1436 
1437 	ASSERT(irb->irb_ipst != NULL);
1438 
1439 	rt = IRB2RT(irb);
1440 	rn = (struct radix_node *)rt;
1441 
1442 	/* first remove it from the radix tree. */
1443 	RADIX_NODE_HEAD_WLOCK(ipst->ips_ip_ftable);
1444 	rw_enter(&irb->irb_lock, RW_WRITER);
1445 	if (irb->irb_refcnt == 1 && irb->irb_nire == 0) {
1446 		rn = ipst->ips_ip_ftable->rnh_deladdr(rn->rn_key, rn->rn_mask,
1447 		    ipst->ips_ip_ftable);
1448 		DTRACE_PROBE1(irb__free, rt_t *,  rt);
1449 		ASSERT((void *)rn == (void *)rt);
1450 		Free(rt, rt_entry_cache);
1451 		/* irb_lock is freed */
1452 		RADIX_NODE_HEAD_UNLOCK(ipst->ips_ip_ftable);
1453 		return (B_TRUE);
1454 	}
1455 	rw_exit(&irb->irb_lock);
1456 	RADIX_NODE_HEAD_UNLOCK(ipst->ips_ip_ftable);
1457 	return (B_FALSE);
1458 }
1459 
1460 /*
1461  * Delete the specified IRE.
1462  * We assume that if ire_bucket is not set then ire_ill->ill_ire_cnt was
1463  * not incremented i.e., that the insertion in the bucket and the increment
1464  * of that counter is done atomically.
1465  */
1466 void
1467 ire_delete(ire_t *ire)
1468 {
1469 	ire_t	*ire1;
1470 	ire_t	**ptpn;
1471 	irb_t	*irb;
1472 	nce_t	*nce;
1473 	ip_stack_t	*ipst = ire->ire_ipst;
1474 
1475 	/* We can clear ire_nce_cache under ire_lock even if the IRE is used */
1476 	mutex_enter(&ire->ire_lock);
1477 	nce = ire->ire_nce_cache;
1478 	ire->ire_nce_cache = NULL;
1479 	mutex_exit(&ire->ire_lock);
1480 	if (nce != NULL)
1481 		nce_refrele(nce);
1482 
1483 	if ((irb = ire->ire_bucket) == NULL) {
1484 		/*
1485 		 * It was never inserted in the list. Should call REFRELE
1486 		 * to free this IRE.
1487 		 */
1488 		ire_refrele_notr(ire);
1489 		return;
1490 	}
1491 
1492 	/*
1493 	 * Move the use counts from an IRE_IF_CLONE to its parent
1494 	 * IRE_INTERFACE.
1495 	 * We need to do this before acquiring irb_lock.
1496 	 */
1497 	if (ire->ire_type & IRE_IF_CLONE) {
1498 		ire_t *parent;
1499 
1500 		rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
1501 		if ((parent = ire->ire_dep_parent) != NULL) {
1502 			parent->ire_ob_pkt_count += ire->ire_ob_pkt_count;
1503 			parent->ire_ib_pkt_count += ire->ire_ib_pkt_count;
1504 			ire->ire_ob_pkt_count = 0;
1505 			ire->ire_ib_pkt_count = 0;
1506 		}
1507 		rw_exit(&ipst->ips_ire_dep_lock);
1508 	}
1509 
1510 	rw_enter(&irb->irb_lock, RW_WRITER);
1511 	if (ire->ire_ptpn == NULL) {
1512 		/*
1513 		 * Some other thread has removed us from the list.
1514 		 * It should have done the REFRELE for us.
1515 		 */
1516 		rw_exit(&irb->irb_lock);
1517 		return;
1518 	}
1519 
1520 	if (!IRE_IS_CONDEMNED(ire)) {
1521 		/* Is this an IRE representing multiple duplicate entries? */
1522 		ASSERT(ire->ire_identical_ref >= 1);
1523 		if (atomic_add_32_nv(&ire->ire_identical_ref, -1) != 0) {
1524 			/* Removed one of the identical parties */
1525 			rw_exit(&irb->irb_lock);
1526 			return;
1527 		}
1528 
1529 		irb->irb_ire_cnt--;
1530 		ire_make_condemned(ire);
1531 	}
1532 
1533 	if (irb->irb_refcnt != 0) {
1534 		/*
1535 		 * The last thread to leave this bucket will
1536 		 * delete this ire.
1537 		 */
1538 		irb->irb_marks |= IRB_MARK_CONDEMNED;
1539 		rw_exit(&irb->irb_lock);
1540 		return;
1541 	}
1542 
1543 	/*
1544 	 * Normally to delete an ire, we walk the bucket. While we
1545 	 * walk the bucket, we normally bump up irb_refcnt and hence
1546 	 * we return from above where we mark CONDEMNED and the ire
1547 	 * gets deleted from ire_unlink. This case is where somebody
1548 	 * knows the ire e.g by doing a lookup, and wants to delete the
1549 	 * IRE. irb_refcnt would be 0 in this case if nobody is walking
1550 	 * the bucket.
1551 	 */
1552 	ptpn = ire->ire_ptpn;
1553 	ire1 = ire->ire_next;
1554 	if (ire1 != NULL)
1555 		ire1->ire_ptpn = ptpn;
1556 	ASSERT(ptpn != NULL);
1557 	*ptpn = ire1;
1558 	ire->ire_ptpn = NULL;
1559 	ire->ire_next = NULL;
1560 	if (ire->ire_ipversion == IPV6_VERSION) {
1561 		BUMP_IRE_STATS(ipst->ips_ire_stats_v6, ire_stats_deleted);
1562 	} else {
1563 		BUMP_IRE_STATS(ipst->ips_ire_stats_v4, ire_stats_deleted);
1564 	}
1565 	rw_exit(&irb->irb_lock);
1566 
1567 	/* Cleanup dependents and related stuff */
1568 	if (ire->ire_ipversion == IPV6_VERSION) {
1569 		ire_delete_v6(ire);
1570 	} else {
1571 		ire_delete_v4(ire);
1572 	}
1573 	/*
1574 	 * We removed it from the list. Decrement the
1575 	 * reference count.
1576 	 */
1577 	ire_refrele_notr(ire);
1578 }
1579 
1580 /*
1581  * Delete the specified IRE.
1582  * All calls should use ire_delete().
1583  * Sometimes called as writer though not required by this function.
1584  *
1585  * NOTE : This function is called only if the ire was added
1586  * in the list.
1587  */
1588 static void
1589 ire_delete_v4(ire_t *ire)
1590 {
1591 	ip_stack_t	*ipst = ire->ire_ipst;
1592 
1593 	ASSERT(ire->ire_refcnt >= 1);
1594 	ASSERT(ire->ire_ipversion == IPV4_VERSION);
1595 
1596 	ire_flush_cache_v4(ire, IRE_FLUSH_DELETE);
1597 	if (ire->ire_type == IRE_DEFAULT) {
1598 		/*
1599 		 * when a default gateway is going away
1600 		 * delete all the host redirects pointing at that
1601 		 * gateway.
1602 		 */
1603 		ire_delete_host_redirects(ire->ire_gateway_addr, ipst);
1604 	}
1605 
1606 	/*
1607 	 * If we are deleting an IRE_INTERFACE then we make sure we also
1608 	 * delete any IRE_IF_CLONE that has been created from it.
1609 	 * Those are always in ire_dep_children.
1610 	 */
1611 	if ((ire->ire_type & IRE_INTERFACE) && ire->ire_dep_children != NULL)
1612 		ire_dep_delete_if_clone(ire);
1613 
1614 	/* Remove from parent dependencies and child */
1615 	rw_enter(&ipst->ips_ire_dep_lock, RW_WRITER);
1616 	if (ire->ire_dep_parent != NULL)
1617 		ire_dep_remove(ire);
1618 
1619 	while (ire->ire_dep_children != NULL)
1620 		ire_dep_remove(ire->ire_dep_children);
1621 	rw_exit(&ipst->ips_ire_dep_lock);
1622 }
1623 
1624 /*
1625  * ire_refrele is the only caller of the function. It calls
1626  * to free the ire when the reference count goes to zero.
1627  */
1628 void
1629 ire_inactive(ire_t *ire)
1630 {
1631 	ill_t	*ill;
1632 	irb_t 	*irb;
1633 	ip_stack_t	*ipst = ire->ire_ipst;
1634 
1635 	ASSERT(ire->ire_refcnt == 0);
1636 	ASSERT(ire->ire_ptpn == NULL);
1637 	ASSERT(ire->ire_next == NULL);
1638 
1639 	/* Count how many condemned ires for kmem_cache callback */
1640 	if (IRE_IS_CONDEMNED(ire))
1641 		atomic_add_32(&ipst->ips_num_ire_condemned, -1);
1642 
1643 	if (ire->ire_gw_secattr != NULL) {
1644 		ire_gw_secattr_free(ire->ire_gw_secattr);
1645 		ire->ire_gw_secattr = NULL;
1646 	}
1647 
1648 	/*
1649 	 * ire_nce_cache is cleared in ire_delete, and we make sure we don't
1650 	 * set it once the ire is marked condemned.
1651 	 */
1652 	ASSERT(ire->ire_nce_cache == NULL);
1653 
1654 	/*
1655 	 * Since any parent would have a refhold on us they would already
1656 	 * have been removed.
1657 	 */
1658 	ASSERT(ire->ire_dep_parent == NULL);
1659 	ASSERT(ire->ire_dep_sib_next == NULL);
1660 	ASSERT(ire->ire_dep_sib_ptpn == NULL);
1661 
1662 	/*
1663 	 * Since any children would have a refhold on us they should have
1664 	 * already been removed.
1665 	 */
1666 	ASSERT(ire->ire_dep_children == NULL);
1667 
1668 	/*
1669 	 * ill_ire_ref is increased when the IRE is inserted in the
1670 	 * bucket - not when the IRE is created.
1671 	 */
1672 	irb = ire->ire_bucket;
1673 	ill = ire->ire_ill;
1674 	if (irb != NULL && ill != NULL) {
1675 		mutex_enter(&ill->ill_lock);
1676 		ASSERT(ill->ill_ire_cnt != 0);
1677 		DTRACE_PROBE3(ill__decr__cnt, (ill_t *), ill,
1678 		    (char *), "ire", (void *), ire);
1679 		ill->ill_ire_cnt--;
1680 		if (ILL_DOWN_OK(ill)) {
1681 			/* Drops the ill lock */
1682 			ipif_ill_refrele_tail(ill);
1683 		} else {
1684 			mutex_exit(&ill->ill_lock);
1685 		}
1686 	}
1687 	ire->ire_ill = NULL;
1688 
1689 	/* This should be true for both V4 and V6 */
1690 	if (irb != NULL && (irb->irb_marks & IRB_MARK_DYNAMIC)) {
1691 		rw_enter(&irb->irb_lock, RW_WRITER);
1692 		irb->irb_nire--;
1693 		/*
1694 		 * Instead of examining the conditions for freeing
1695 		 * the radix node here, we do it by calling
1696 		 * irb_refrele which is a single point in the code
1697 		 * that embeds that logic. Bump up the refcnt to
1698 		 * be able to call irb_refrele
1699 		 */
1700 		irb_refhold_locked(irb);
1701 		rw_exit(&irb->irb_lock);
1702 		irb_refrele(irb);
1703 	}
1704 
1705 #ifdef DEBUG
1706 	ire_trace_cleanup(ire);
1707 #endif
1708 	mutex_destroy(&ire->ire_lock);
1709 	if (ire->ire_ipversion == IPV6_VERSION) {
1710 		BUMP_IRE_STATS(ipst->ips_ire_stats_v6, ire_stats_freed);
1711 	} else {
1712 		BUMP_IRE_STATS(ipst->ips_ire_stats_v4, ire_stats_freed);
1713 	}
1714 	kmem_cache_free(ire_cache, ire);
1715 }
1716 
1717 /*
1718  * ire_update_generation is the callback function provided by
1719  * ire_get_bucket() to update the generation number of any
1720  * matching shorter route when a new route is added.
1721  *
1722  * This fucntion always returns a failure return (B_FALSE)
1723  * to force the caller (rn_matchaddr_args)
1724  * to back-track up the tree looking for shorter matches.
1725  */
1726 /* ARGSUSED */
1727 static boolean_t
1728 ire_update_generation(struct radix_node *rn, void *arg)
1729 {
1730 	struct rt_entry *rt = (struct rt_entry *)rn;
1731 
1732 	/* We need to handle all in the same bucket */
1733 	irb_increment_generation(&rt->rt_irb);
1734 	return (B_FALSE);
1735 }
1736 
1737 /*
1738  * Take care of all the generation numbers in the bucket.
1739  */
1740 void
1741 irb_increment_generation(irb_t *irb)
1742 {
1743 	ire_t *ire;
1744 
1745 	if (irb == NULL || irb->irb_ire_cnt == 0)
1746 		return;
1747 
1748 	irb_refhold(irb);
1749 	for (ire = irb->irb_ire; ire != NULL; ire = ire->ire_next) {
1750 		if (!IRE_IS_CONDEMNED(ire))
1751 			ire_increment_generation(ire);	/* Ourselves */
1752 		ire_dep_incr_generation(ire);	/* Dependants */
1753 	}
1754 	irb_refrele(irb);
1755 }
1756 
1757 /*
1758  * When an IRE is added or deleted this routine is called to make sure
1759  * any caching of IRE information is notified or updated.
1760  *
1761  * The flag argument indicates if the flush request is due to addition
1762  * of new route (IRE_FLUSH_ADD), deletion of old route (IRE_FLUSH_DELETE),
1763  * or a change to ire_gateway_addr (IRE_FLUSH_GWCHANGE).
1764  */
1765 void
1766 ire_flush_cache_v4(ire_t *ire, int flag)
1767 {
1768 	irb_t *irb = ire->ire_bucket;
1769 	struct rt_entry *rt = IRB2RT(irb);
1770 	ip_stack_t *ipst = ire->ire_ipst;
1771 
1772 	/*
1773 	 * IRE_IF_CLONE ire's don't provide any new information
1774 	 * than the parent from which they are cloned, so don't
1775 	 * perturb the generation numbers.
1776 	 */
1777 	if (ire->ire_type & IRE_IF_CLONE)
1778 		return;
1779 
1780 	/*
1781 	 * Ensure that an ire_add during a lookup serializes the updates of the
1782 	 * generation numbers under the radix head lock so that the lookup gets
1783 	 * either the old ire and old generation number, or a new ire and new
1784 	 * generation number.
1785 	 */
1786 	RADIX_NODE_HEAD_WLOCK(ipst->ips_ip_ftable);
1787 
1788 	/*
1789 	 * If a route was just added, we need to notify everybody that
1790 	 * has cached an IRE_NOROUTE since there might now be a better
1791 	 * route for them.
1792 	 */
1793 	if (flag == IRE_FLUSH_ADD) {
1794 		ire_increment_generation(ipst->ips_ire_reject_v4);
1795 		ire_increment_generation(ipst->ips_ire_blackhole_v4);
1796 	}
1797 
1798 	/* Adding a default can't otherwise provide a better route */
1799 	if (ire->ire_type == IRE_DEFAULT && flag == IRE_FLUSH_ADD) {
1800 		RADIX_NODE_HEAD_UNLOCK(ipst->ips_ip_ftable);
1801 		return;
1802 	}
1803 
1804 	switch (flag) {
1805 	case IRE_FLUSH_DELETE:
1806 	case IRE_FLUSH_GWCHANGE:
1807 		/*
1808 		 * Update ire_generation for all ire_dep_children chains
1809 		 * starting with this IRE
1810 		 */
1811 		ire_dep_incr_generation(ire);
1812 		break;
1813 	case IRE_FLUSH_ADD:
1814 		/*
1815 		 * Update the generation numbers of all shorter matching routes.
1816 		 * ire_update_generation takes care of the dependants by
1817 		 * using ire_dep_incr_generation.
1818 		 */
1819 		(void) ipst->ips_ip_ftable->rnh_matchaddr_args(&rt->rt_dst,
1820 		    ipst->ips_ip_ftable, ire_update_generation, NULL);
1821 		break;
1822 	}
1823 	RADIX_NODE_HEAD_UNLOCK(ipst->ips_ip_ftable);
1824 }
1825 
1826 /*
1827  * Matches the arguments passed with the values in the ire.
1828  *
1829  * Note: for match types that match using "ill" passed in, ill
1830  * must be checked for non-NULL before calling this routine.
1831  */
1832 boolean_t
1833 ire_match_args(ire_t *ire, ipaddr_t addr, ipaddr_t mask, ipaddr_t gateway,
1834     int type, const ill_t *ill, zoneid_t zoneid,
1835     const ts_label_t *tsl, int match_flags)
1836 {
1837 	ill_t *ire_ill = NULL, *dst_ill;
1838 	ip_stack_t *ipst = ire->ire_ipst;
1839 
1840 	ASSERT(ire->ire_ipversion == IPV4_VERSION);
1841 	ASSERT((ire->ire_addr & ~ire->ire_mask) == 0);
1842 	ASSERT((!(match_flags & MATCH_IRE_ILL)) ||
1843 	    (ill != NULL && !ill->ill_isv6));
1844 
1845 	/*
1846 	 * If MATCH_IRE_TESTHIDDEN is set, then only return the IRE if it is
1847 	 * in fact hidden, to ensure the caller gets the right one.
1848 	 */
1849 	if (ire->ire_testhidden) {
1850 		if (!(match_flags & MATCH_IRE_TESTHIDDEN))
1851 			return (B_FALSE);
1852 	}
1853 
1854 	if (zoneid != ALL_ZONES && zoneid != ire->ire_zoneid &&
1855 	    ire->ire_zoneid != ALL_ZONES) {
1856 		/*
1857 		 * If MATCH_IRE_ZONEONLY has been set and the supplied zoneid
1858 		 * does not match that of ire_zoneid, a failure to
1859 		 * match is reported at this point. Otherwise, since some IREs
1860 		 * that are available in the global zone can be used in local
1861 		 * zones, additional checks need to be performed:
1862 		 *
1863 		 * IRE_LOOPBACK
1864 		 *	entries should never be matched in this situation.
1865 		 *	Each zone has its own IRE_LOOPBACK.
1866 		 *
1867 		 * IRE_LOCAL
1868 		 *	We allow them for any zoneid. ire_route_recursive
1869 		 *	does additional checks when
1870 		 *	ip_restrict_interzone_loopback is set.
1871 		 *
1872 		 * If ill_usesrc_ifindex is set
1873 		 *	Then we check if the zone has a valid source address
1874 		 *	on the usesrc ill.
1875 		 *
1876 		 * If ire_ill is set, then check that the zone has an ipif
1877 		 *	on that ill.
1878 		 *
1879 		 * Outside of this function (in ire_round_robin) we check
1880 		 * that any IRE_OFFLINK has a gateway that reachable from the
1881 		 * zone when we have multiple choices (ECMP).
1882 		 */
1883 		if (match_flags & MATCH_IRE_ZONEONLY)
1884 			return (B_FALSE);
1885 		if (ire->ire_type & IRE_LOOPBACK)
1886 			return (B_FALSE);
1887 
1888 		if (ire->ire_type & IRE_LOCAL)
1889 			goto matchit;
1890 
1891 		/*
1892 		 * The normal case of IRE_ONLINK has a matching zoneid.
1893 		 * Here we handle the case when shared-IP zones have been
1894 		 * configured with IP addresses on vniN. In that case it
1895 		 * is ok for traffic from a zone to use IRE_ONLINK routes
1896 		 * if the ill has a usesrc pointing at vniN
1897 		 */
1898 		dst_ill = ire->ire_ill;
1899 		if (ire->ire_type & IRE_ONLINK) {
1900 			uint_t	ifindex;
1901 
1902 			/*
1903 			 * Note there is no IRE_INTERFACE on vniN thus
1904 			 * can't do an IRE lookup for a matching route.
1905 			 */
1906 			ifindex = dst_ill->ill_usesrc_ifindex;
1907 			if (ifindex == 0)
1908 				return (B_FALSE);
1909 
1910 			/*
1911 			 * If there is a usable source address in the
1912 			 * zone, then it's ok to return this IRE_INTERFACE
1913 			 */
1914 			if (!ipif_zone_avail(ifindex, dst_ill->ill_isv6,
1915 			    zoneid, ipst)) {
1916 				ip3dbg(("ire_match_args: no usrsrc for zone"
1917 				    " dst_ill %p\n", (void *)dst_ill));
1918 				return (B_FALSE);
1919 			}
1920 		}
1921 		/*
1922 		 * For exampe, with
1923 		 * route add 11.0.0.0 gw1 -ifp bge0
1924 		 * route add 11.0.0.0 gw2 -ifp bge1
1925 		 * this code would differentiate based on
1926 		 * where the sending zone has addresses.
1927 		 * Only if the zone has an address on bge0 can it use the first
1928 		 * route. It isn't clear if this behavior is documented
1929 		 * anywhere.
1930 		 */
1931 		if (dst_ill != NULL && (ire->ire_type & IRE_OFFLINK)) {
1932 			ipif_t	*tipif;
1933 
1934 			mutex_enter(&dst_ill->ill_lock);
1935 			for (tipif = dst_ill->ill_ipif;
1936 			    tipif != NULL; tipif = tipif->ipif_next) {
1937 				if (!IPIF_IS_CONDEMNED(tipif) &&
1938 				    (tipif->ipif_flags & IPIF_UP) &&
1939 				    (tipif->ipif_zoneid == zoneid ||
1940 				    tipif->ipif_zoneid == ALL_ZONES))
1941 					break;
1942 			}
1943 			mutex_exit(&dst_ill->ill_lock);
1944 			if (tipif == NULL) {
1945 				return (B_FALSE);
1946 			}
1947 		}
1948 	}
1949 
1950 matchit:
1951 	if (match_flags & MATCH_IRE_ILL) {
1952 		ire_ill = ire->ire_ill;
1953 
1954 		/*
1955 		 * If asked to match an ill, we *must* match
1956 		 * on the ire_ill for ipmp test addresses, or
1957 		 * any of the ill in the group for data addresses.
1958 		 * If we don't, we may as well fail.
1959 		 * However, we need an exception for IRE_LOCALs to ensure
1960 		 * we loopback packets even sent to test addresses on different
1961 		 * interfaces in the group.
1962 		 */
1963 		if ((match_flags & MATCH_IRE_TESTHIDDEN) &&
1964 		    !(ire->ire_type & IRE_LOCAL)) {
1965 			if (ire->ire_ill != ill)
1966 				return (B_FALSE);
1967 		} else  {
1968 			match_flags &= ~MATCH_IRE_TESTHIDDEN;
1969 			/*
1970 			 * We know that ill is not NULL, but ire_ill could be
1971 			 * NULL
1972 			 */
1973 			if (ire_ill == NULL || !IS_ON_SAME_LAN(ill, ire_ill))
1974 				return (B_FALSE);
1975 		}
1976 	}
1977 
1978 	if ((ire->ire_addr == (addr & mask)) &&
1979 	    ((!(match_flags & MATCH_IRE_GW)) ||
1980 	    (ire->ire_gateway_addr == gateway)) &&
1981 	    ((!(match_flags & MATCH_IRE_TYPE)) || (ire->ire_type & type)) &&
1982 	    ((!(match_flags & MATCH_IRE_TESTHIDDEN)) || ire->ire_testhidden) &&
1983 	    ((!(match_flags & MATCH_IRE_MASK)) || (ire->ire_mask == mask)) &&
1984 	    ((!(match_flags & MATCH_IRE_SECATTR)) ||
1985 	    (!is_system_labeled()) ||
1986 	    (tsol_ire_match_gwattr(ire, tsl) == 0))) {
1987 		/* We found the matched IRE */
1988 		return (B_TRUE);
1989 	}
1990 	return (B_FALSE);
1991 }
1992 
1993 /*
1994  * Check if the IRE_LOCAL uses the same ill as another route would use.
1995  * If there is no alternate route, or the alternate is a REJECT or BLACKHOLE,
1996  * then we don't allow this IRE_LOCAL to be used.
1997  * We always return an IRE; will be RTF_REJECT if no route available.
1998  */
1999 ire_t *
2000 ire_alt_local(ire_t *ire, zoneid_t zoneid, const ts_label_t *tsl,
2001     const ill_t *ill, uint_t *generationp)
2002 {
2003 	ip_stack_t	*ipst = ire->ire_ipst;
2004 	ire_t		*alt_ire;
2005 	uint_t		ire_type;
2006 	uint_t		generation;
2007 	uint_t		match_flags;
2008 
2009 	ASSERT(ire->ire_type & IRE_LOCAL);
2010 	ASSERT(ire->ire_ill != NULL);
2011 
2012 	/*
2013 	 * Need to match on everything but local.
2014 	 * This might result in the creation of a IRE_IF_CLONE for the
2015 	 * same address as the IRE_LOCAL when restrict_interzone_loopback is
2016 	 * set. ire_add_*() ensures that the IRE_IF_CLONE are tail inserted
2017 	 * to make sure the IRE_LOCAL is always found first.
2018 	 */
2019 	ire_type = (IRE_ONLINK | IRE_OFFLINK) & ~(IRE_LOCAL|IRE_LOOPBACK);
2020 	match_flags = MATCH_IRE_TYPE | MATCH_IRE_SECATTR;
2021 	if (ill != NULL)
2022 		match_flags |= MATCH_IRE_ILL;
2023 
2024 	if (ire->ire_ipversion == IPV4_VERSION) {
2025 		alt_ire = ire_route_recursive_v4(ire->ire_addr, ire_type,
2026 		    ill, zoneid, tsl, match_flags, B_TRUE, 0, ipst, NULL, NULL,
2027 		    &generation);
2028 	} else {
2029 		alt_ire = ire_route_recursive_v6(&ire->ire_addr_v6, ire_type,
2030 		    ill, zoneid, tsl, match_flags, B_TRUE, 0, ipst, NULL, NULL,
2031 		    &generation);
2032 	}
2033 	ASSERT(alt_ire != NULL);
2034 
2035 	if (alt_ire->ire_ill == ire->ire_ill) {
2036 		/* Going out the same ILL - ok to send to IRE_LOCAL */
2037 		ire_refrele(alt_ire);
2038 	} else {
2039 		/* Different ill - ignore IRE_LOCAL */
2040 		ire_refrele(ire);
2041 		ire = alt_ire;
2042 		if (generationp != NULL)
2043 			*generationp = generation;
2044 	}
2045 	return (ire);
2046 }
2047 
2048 boolean_t
2049 ire_find_zoneid(struct radix_node *rn, void *arg)
2050 {
2051 	struct rt_entry *rt = (struct rt_entry *)rn;
2052 	irb_t *irb;
2053 	ire_t *ire;
2054 	ire_ftable_args_t *margs = arg;
2055 
2056 	ASSERT(rt != NULL);
2057 
2058 	irb = &rt->rt_irb;
2059 
2060 	if (irb->irb_ire_cnt == 0)
2061 		return (B_FALSE);
2062 
2063 	rw_enter(&irb->irb_lock, RW_READER);
2064 	for (ire = irb->irb_ire; ire != NULL; ire = ire->ire_next) {
2065 		if (IRE_IS_CONDEMNED(ire))
2066 			continue;
2067 
2068 		if (ire->ire_zoneid != ALL_ZONES &&
2069 		    ire->ire_zoneid != margs->ift_zoneid)
2070 			continue;
2071 
2072 		if (margs->ift_ill != NULL && margs->ift_ill != ire->ire_ill)
2073 			continue;
2074 
2075 		if (is_system_labeled() &&
2076 		    tsol_ire_match_gwattr(ire, margs->ift_tsl) != 0)
2077 			continue;
2078 
2079 		rw_exit(&irb->irb_lock);
2080 		return (B_TRUE);
2081 	}
2082 	rw_exit(&irb->irb_lock);
2083 	return (B_FALSE);
2084 }
2085 
2086 /*
2087  * Check if the zoneid (not ALL_ZONES) has an IRE_INTERFACE for the specified
2088  * gateway address. If ill is non-NULL we also match on it.
2089  * The caller must hold a read lock on RADIX_NODE_HEAD if lock_held is set.
2090  */
2091 boolean_t
2092 ire_gateway_ok_zone_v4(ipaddr_t gateway, zoneid_t zoneid, ill_t *ill,
2093     const ts_label_t *tsl, ip_stack_t *ipst, boolean_t lock_held)
2094 {
2095 	struct rt_sockaddr rdst;
2096 	struct rt_entry *rt;
2097 	ire_ftable_args_t margs;
2098 
2099 	ASSERT(ill == NULL || !ill->ill_isv6);
2100 	if (lock_held)
2101 		ASSERT(RW_READ_HELD(&ipst->ips_ip_ftable->rnh_lock));
2102 	else
2103 		RADIX_NODE_HEAD_RLOCK(ipst->ips_ip_ftable);
2104 
2105 	rdst.rt_sin_len = sizeof (rdst);
2106 	rdst.rt_sin_family = AF_INET;
2107 	rdst.rt_sin_addr.s_addr = gateway;
2108 
2109 	/*
2110 	 * We only use margs for ill, zoneid, and tsl matching in
2111 	 * ire_find_zoneid
2112 	 */
2113 	(void) memset(&margs, 0, sizeof (margs));
2114 	margs.ift_ill = ill;
2115 	margs.ift_zoneid = zoneid;
2116 	margs.ift_tsl = tsl;
2117 	rt = (struct rt_entry *)ipst->ips_ip_ftable->rnh_matchaddr_args(&rdst,
2118 	    ipst->ips_ip_ftable, ire_find_zoneid, (void *)&margs);
2119 
2120 	if (!lock_held)
2121 		RADIX_NODE_HEAD_UNLOCK(ipst->ips_ip_ftable);
2122 
2123 	return (rt != NULL);
2124 }
2125 
2126 /*
2127  * ire_walk routine to delete a fraction of redirect IREs and IRE_CLONE_IF IREs.
2128  * The fraction argument tells us what fraction of the IREs to delete.
2129  * Common for IPv4 and IPv6.
2130  * Used when memory backpressure.
2131  */
2132 static void
2133 ire_delete_reclaim(ire_t *ire, char *arg)
2134 {
2135 	ip_stack_t	*ipst = ire->ire_ipst;
2136 	uint_t		fraction = *(uint_t *)arg;
2137 	uint_t		rand;
2138 
2139 	if ((ire->ire_flags & RTF_DYNAMIC) ||
2140 	    (ire->ire_type & IRE_IF_CLONE)) {
2141 
2142 		/* Pick a random number */
2143 		rand = (uint_t)lbolt +
2144 		    IRE_ADDR_HASH_V6(ire->ire_addr_v6, 256);
2145 
2146 		/* Use truncation */
2147 		if ((rand/fraction)*fraction == rand) {
2148 			IP_STAT(ipst, ip_ire_reclaim_deleted);
2149 			ire_delete(ire);
2150 		}
2151 	}
2152 
2153 }
2154 
2155 /*
2156  * kmem_cache callback to free up memory.
2157  *
2158  * Free a fraction (ips_ip_ire_reclaim_fraction) of things IP added dynamically
2159  * (RTF_DYNAMIC and IRE_IF_CLONE).
2160  */
2161 static void
2162 ip_ire_reclaim_stack(ip_stack_t *ipst)
2163 {
2164 	uint_t	fraction = ipst->ips_ip_ire_reclaim_fraction;
2165 
2166 	IP_STAT(ipst, ip_ire_reclaim_calls);
2167 
2168 	ire_walk(ire_delete_reclaim, &fraction, ipst);
2169 
2170 	/*
2171 	 * Walk all CONNs that can have a reference on an ire, nce or dce.
2172 	 * Get them to update any stale references to drop any refholds they
2173 	 * have.
2174 	 */
2175 	ipcl_walk(conn_ixa_cleanup, (void *)B_FALSE, ipst);
2176 }
2177 
2178 /*
2179  * Called by the memory allocator subsystem directly, when the system
2180  * is running low on memory.
2181  */
2182 /* ARGSUSED */
2183 void
2184 ip_ire_reclaim(void *args)
2185 {
2186 	netstack_handle_t nh;
2187 	netstack_t *ns;
2188 
2189 	netstack_next_init(&nh);
2190 	while ((ns = netstack_next(&nh)) != NULL) {
2191 		ip_ire_reclaim_stack(ns->netstack_ip);
2192 		netstack_rele(ns);
2193 	}
2194 	netstack_next_fini(&nh);
2195 }
2196 
2197 static void
2198 power2_roundup(uint32_t *value)
2199 {
2200 	int i;
2201 
2202 	for (i = 1; i < 31; i++) {
2203 		if (*value <= (1 << i))
2204 			break;
2205 	}
2206 	*value = (1 << i);
2207 }
2208 
2209 /* Global init for all zones */
2210 void
2211 ip_ire_g_init()
2212 {
2213 	/*
2214 	 * Create kmem_caches.  ip_ire_reclaim() and ip_nce_reclaim()
2215 	 * will give disposable IREs back to system when needed.
2216 	 * This needs to be done here before anything else, since
2217 	 * ire_add() expects the cache to be created.
2218 	 */
2219 	ire_cache = kmem_cache_create("ire_cache",
2220 	    sizeof (ire_t), 0, NULL, NULL,
2221 	    ip_ire_reclaim, NULL, NULL, 0);
2222 
2223 	ncec_cache = kmem_cache_create("ncec_cache",
2224 	    sizeof (ncec_t), 0, NULL, NULL,
2225 	    ip_nce_reclaim, NULL, NULL, 0);
2226 	nce_cache = kmem_cache_create("nce_cache",
2227 	    sizeof (nce_t), 0, NULL, NULL,
2228 	    NULL, NULL, NULL, 0);
2229 
2230 	rt_entry_cache = kmem_cache_create("rt_entry",
2231 	    sizeof (struct rt_entry), 0, NULL, NULL, NULL, NULL, NULL, 0);
2232 
2233 	/*
2234 	 * Have radix code setup kmem caches etc.
2235 	 */
2236 	rn_init();
2237 }
2238 
2239 void
2240 ip_ire_init(ip_stack_t *ipst)
2241 {
2242 	ire_t	*ire;
2243 	int	error;
2244 
2245 	mutex_init(&ipst->ips_ire_ft_init_lock, NULL, MUTEX_DEFAULT, 0);
2246 
2247 	(void) rn_inithead((void **)&ipst->ips_ip_ftable, 32);
2248 
2249 	/*
2250 	 * Make sure that the forwarding table size is a power of 2.
2251 	 * The IRE*_ADDR_HASH() macroes depend on that.
2252 	 */
2253 	ipst->ips_ip6_ftable_hash_size = ip6_ftable_hash_size;
2254 	power2_roundup(&ipst->ips_ip6_ftable_hash_size);
2255 
2256 	/*
2257 	 * Allocate/initialize a pair of IRE_NOROUTEs for each of IPv4 and IPv6.
2258 	 * The ire_reject_v* has RTF_REJECT set, and the ire_blackhole_v* has
2259 	 * RTF_BLACKHOLE set. We use the latter for transient errors such
2260 	 * as memory allocation failures and tripping on IRE_IS_CONDEMNED
2261 	 * entries.
2262 	 */
2263 	ire = kmem_cache_alloc(ire_cache, KM_SLEEP);
2264 	*ire = ire_null;
2265 	error = ire_init_v4(ire, 0, 0, 0, IRE_NOROUTE, NULL, ALL_ZONES,
2266 	    RTF_REJECT|RTF_UP, NULL, ipst);
2267 	ASSERT(error == 0);
2268 	ipst->ips_ire_reject_v4 = ire;
2269 
2270 	ire = kmem_cache_alloc(ire_cache, KM_SLEEP);
2271 	*ire = ire_null;
2272 	error = ire_init_v6(ire, 0, 0, 0, IRE_NOROUTE, NULL, ALL_ZONES,
2273 	    RTF_REJECT|RTF_UP, NULL, ipst);
2274 	ASSERT(error == 0);
2275 	ipst->ips_ire_reject_v6 = ire;
2276 
2277 	ire = kmem_cache_alloc(ire_cache, KM_SLEEP);
2278 	*ire = ire_null;
2279 	error = ire_init_v4(ire, 0, 0, 0, IRE_NOROUTE, NULL, ALL_ZONES,
2280 	    RTF_BLACKHOLE|RTF_UP, NULL, ipst);
2281 	ASSERT(error == 0);
2282 	ipst->ips_ire_blackhole_v4 = ire;
2283 
2284 	ire = kmem_cache_alloc(ire_cache, KM_SLEEP);
2285 	*ire = ire_null;
2286 	error = ire_init_v6(ire, 0, 0, 0, IRE_NOROUTE, NULL, ALL_ZONES,
2287 	    RTF_BLACKHOLE|RTF_UP, NULL, ipst);
2288 	ASSERT(error == 0);
2289 	ipst->ips_ire_blackhole_v6 = ire;
2290 
2291 	rw_init(&ipst->ips_ip6_ire_head_lock, NULL, RW_DEFAULT, NULL);
2292 	rw_init(&ipst->ips_ire_dep_lock, NULL, RW_DEFAULT, NULL);
2293 }
2294 
2295 void
2296 ip_ire_g_fini(void)
2297 {
2298 	kmem_cache_destroy(ire_cache);
2299 	kmem_cache_destroy(ncec_cache);
2300 	kmem_cache_destroy(nce_cache);
2301 	kmem_cache_destroy(rt_entry_cache);
2302 
2303 	rn_fini();
2304 }
2305 
2306 void
2307 ip_ire_fini(ip_stack_t *ipst)
2308 {
2309 	int i;
2310 
2311 	rw_destroy(&ipst->ips_ire_dep_lock);
2312 	rw_destroy(&ipst->ips_ip6_ire_head_lock);
2313 
2314 	ire_refrele_notr(ipst->ips_ire_reject_v6);
2315 	ipst->ips_ire_reject_v6 = NULL;
2316 	ire_refrele_notr(ipst->ips_ire_reject_v4);
2317 	ipst->ips_ire_reject_v4 = NULL;
2318 	ire_refrele_notr(ipst->ips_ire_blackhole_v6);
2319 	ipst->ips_ire_blackhole_v6 = NULL;
2320 	ire_refrele_notr(ipst->ips_ire_blackhole_v4);
2321 	ipst->ips_ire_blackhole_v4 = NULL;
2322 
2323 	/*
2324 	 * Delete all IREs - assumes that the ill/ipifs have
2325 	 * been removed so what remains are just the ftable to handle.
2326 	 */
2327 	ire_walk(ire_delete, NULL, ipst);
2328 
2329 	rn_freehead(ipst->ips_ip_ftable);
2330 	ipst->ips_ip_ftable = NULL;
2331 
2332 	mutex_destroy(&ipst->ips_ire_ft_init_lock);
2333 
2334 	for (i = 0; i < IP6_MASK_TABLE_SIZE; i++) {
2335 		irb_t *ptr;
2336 		int j;
2337 
2338 		if ((ptr = ipst->ips_ip_forwarding_table_v6[i]) == NULL)
2339 			continue;
2340 
2341 		for (j = 0; j < ipst->ips_ip6_ftable_hash_size; j++) {
2342 			ASSERT(ptr[j].irb_ire == NULL);
2343 			rw_destroy(&ptr[j].irb_lock);
2344 		}
2345 		mi_free(ptr);
2346 		ipst->ips_ip_forwarding_table_v6[i] = NULL;
2347 	}
2348 }
2349 
2350 #ifdef DEBUG
2351 void
2352 ire_trace_ref(ire_t *ire)
2353 {
2354 	mutex_enter(&ire->ire_lock);
2355 	if (ire->ire_trace_disable) {
2356 		mutex_exit(&ire->ire_lock);
2357 		return;
2358 	}
2359 
2360 	if (th_trace_ref(ire, ire->ire_ipst)) {
2361 		mutex_exit(&ire->ire_lock);
2362 	} else {
2363 		ire->ire_trace_disable = B_TRUE;
2364 		mutex_exit(&ire->ire_lock);
2365 		ire_trace_cleanup(ire);
2366 	}
2367 }
2368 
2369 void
2370 ire_untrace_ref(ire_t *ire)
2371 {
2372 	mutex_enter(&ire->ire_lock);
2373 	if (!ire->ire_trace_disable)
2374 		th_trace_unref(ire);
2375 	mutex_exit(&ire->ire_lock);
2376 }
2377 
2378 static void
2379 ire_trace_cleanup(const ire_t *ire)
2380 {
2381 	th_trace_cleanup(ire, ire->ire_trace_disable);
2382 }
2383 #endif /* DEBUG */
2384 
2385 /*
2386  * Find, or create if needed, the nce_t pointer to the neighbor cache
2387  * entry ncec_t for an IPv4 address. The nce_t will be created on the ill_t
2388  * in the non-IPMP case, or on the cast-ill in the IPMP bcast/mcast case, or
2389  * on the next available under-ill (selected by the IPMP rotor) in the
2390  * unicast IPMP case.
2391  *
2392  * If a neighbor-cache entry has to be created (i.e., one does not already
2393  * exist in the nce list) the ncec_lladdr and ncec_state of the neighbor cache
2394  * entry are initialized in nce_add_v4(). The broadcast, multicast, and
2395  * link-layer type determine the contents of {ncec_state, ncec_lladdr} of
2396  * the ncec_t created. The ncec_lladdr is non-null for all link types with
2397  * non-zero ill_phys_addr_length, though the contents may be zero in cases
2398  * where the link-layer type is not known at the time of creation
2399  * (e.g., IRE_IFRESOLVER links)
2400  *
2401  * All IRE_BROADCAST entries have ncec_state = ND_REACHABLE, and the nce_lladr
2402  * has the physical broadcast address of the outgoing interface.
2403  * For unicast ire entries,
2404  *   - if the outgoing interface is of type IRE_IF_RESOLVER, a newly created
2405  *     ncec_t with 0 nce_lladr contents, and will be in the ND_INITIAL state.
2406  *   - if the outgoing interface is a IRE_IF_NORESOLVER interface, no link
2407  *     layer resolution is necessary, so that the ncec_t will be in the
2408  *     ND_REACHABLE state
2409  *
2410  * The link layer information needed for broadcast addresses, and for
2411  * packets sent on IRE_IF_NORESOLVER interfaces is a constant mapping that
2412  * never needs re-verification for the lifetime of the ncec_t. These are
2413  * therefore marked NCE_F_NONUD.
2414  *
2415  * The nce returned will be created such that the nce_ill == ill that
2416  * is passed in. Note that the nce itself may not have ncec_ill == ill
2417  * where IPMP links are involved.
2418  */
2419 static nce_t *
2420 ire_nce_init(ill_t *ill, const void *addr, int ire_type)
2421 {
2422 	int		err;
2423 	nce_t		*nce = NULL;
2424 	uint16_t	ncec_flags;
2425 	uchar_t		*hwaddr;
2426 	boolean_t	need_refrele = B_FALSE;
2427 	ill_t		*in_ill = ill;
2428 	boolean_t	is_unicast;
2429 	uint_t		hwaddr_len;
2430 
2431 	is_unicast = ((ire_type & (IRE_MULTICAST|IRE_BROADCAST)) == 0);
2432 	if (IS_IPMP(ill) ||
2433 	    ((ire_type & IRE_BROADCAST) && IS_UNDER_IPMP(ill))) {
2434 		if ((ill = ipmp_ill_get_xmit_ill(ill, is_unicast)) == NULL)
2435 			return (NULL);
2436 		need_refrele = B_TRUE;
2437 	}
2438 	ncec_flags = (ill->ill_flags & ILLF_NONUD) ? NCE_F_NONUD : 0;
2439 
2440 	switch (ire_type) {
2441 	case IRE_BROADCAST:
2442 		ASSERT(!ill->ill_isv6);
2443 		ncec_flags |= (NCE_F_BCAST|NCE_F_NONUD);
2444 		break;
2445 	case IRE_MULTICAST:
2446 		ncec_flags |= (NCE_F_MCAST|NCE_F_NONUD);
2447 		break;
2448 	}
2449 
2450 	if (ill->ill_net_type == IRE_IF_NORESOLVER && is_unicast) {
2451 		hwaddr = ill->ill_dest_addr;
2452 	} else {
2453 		hwaddr = NULL;
2454 	}
2455 	hwaddr_len = ill->ill_phys_addr_length;
2456 
2457 retry:
2458 	/* nce_state will be computed by nce_add_common() */
2459 	if (!ill->ill_isv6) {
2460 		err = nce_lookup_then_add_v4(ill, hwaddr, hwaddr_len, addr,
2461 		    ncec_flags, ND_UNCHANGED, &nce);
2462 	} else {
2463 		err = nce_lookup_then_add_v6(ill, hwaddr, hwaddr_len, addr,
2464 		    ncec_flags, ND_UNCHANGED, &nce);
2465 	}
2466 
2467 	switch (err) {
2468 	case 0:
2469 		break;
2470 	case EEXIST:
2471 		/*
2472 		 * When subnets change or partially overlap what was once
2473 		 * a broadcast address could now be a unicast, or vice versa.
2474 		 */
2475 		if (((ncec_flags ^ nce->nce_common->ncec_flags) &
2476 		    NCE_F_BCAST) != 0) {
2477 			ASSERT(!ill->ill_isv6);
2478 			ncec_delete(nce->nce_common);
2479 			nce_refrele(nce);
2480 			goto retry;
2481 		}
2482 		break;
2483 	default:
2484 		DTRACE_PROBE2(nce__init__fail, ill_t *, ill, int, err);
2485 		if (need_refrele)
2486 			ill_refrele(ill);
2487 		return (NULL);
2488 	}
2489 	/*
2490 	 * If the ill was an under-ill of an IPMP group, we need to verify
2491 	 * that it is still active so that we select an active interface in
2492 	 * the group. However, since ipmp_ill_is_active ASSERTs for
2493 	 * IS_UNDER_IPMP(), we first need to verify that the ill is an
2494 	 * under-ill, and since this is being done in the data path, the
2495 	 * only way to ascertain this is by holding the ill_g_lock.
2496 	 */
2497 	rw_enter(&ill->ill_ipst->ips_ill_g_lock, RW_READER);
2498 	mutex_enter(&ill->ill_lock);
2499 	mutex_enter(&ill->ill_phyint->phyint_lock);
2500 	if (need_refrele && IS_UNDER_IPMP(ill) && !ipmp_ill_is_active(ill)) {
2501 		/*
2502 		 * need_refrele implies that the under ill was selected by
2503 		 * ipmp_ill_get_xmit_ill() because either the in_ill was an
2504 		 * ipmp_ill, or we are sending a non-unicast packet on
2505 		 * an under_ill. However, when we get here, the ill selected by
2506 		 * ipmp_ill_get_xmit_ill  was pulled out of the active set
2507 		 * (for unicast)  or cast_ill nomination (for
2508 		 * !unicast) after it was  picked as the outgoing ill.
2509 		 * We have to pick an active interface and/or cast_ill in the
2510 		 * group.
2511 		 */
2512 		mutex_exit(&ill->ill_phyint->phyint_lock);
2513 		nce_delete(nce);
2514 		mutex_exit(&ill->ill_lock);
2515 		rw_exit(&ill->ill_ipst->ips_ill_g_lock);
2516 		nce_refrele(nce);
2517 		ill_refrele(ill);
2518 		if ((ill = ipmp_ill_get_xmit_ill(in_ill, is_unicast)) == NULL)
2519 			return (NULL);
2520 		goto retry;
2521 	} else {
2522 		mutex_exit(&ill->ill_phyint->phyint_lock);
2523 		mutex_exit(&ill->ill_lock);
2524 		rw_exit(&ill->ill_ipst->ips_ill_g_lock);
2525 	}
2526 done:
2527 	ASSERT(nce->nce_ill == ill);
2528 	if (need_refrele)
2529 		ill_refrele(ill);
2530 	return (nce);
2531 }
2532 
2533 nce_t *
2534 arp_nce_init(ill_t *ill, in_addr_t addr4, int ire_type)
2535 {
2536 	return (ire_nce_init(ill, &addr4, ire_type));
2537 }
2538 
2539 nce_t *
2540 ndp_nce_init(ill_t *ill, const in6_addr_t *addr6, int ire_type)
2541 {
2542 	ASSERT((ire_type & IRE_BROADCAST) == 0);
2543 	return (ire_nce_init(ill, addr6, ire_type));
2544 }
2545 
2546 /*
2547  * The caller should hold irb_lock as a writer if the ire is in a bucket.
2548  */
2549 void
2550 ire_make_condemned(ire_t *ire)
2551 {
2552 	ip_stack_t	*ipst = ire->ire_ipst;
2553 
2554 	mutex_enter(&ire->ire_lock);
2555 	ASSERT(ire->ire_bucket == NULL ||
2556 	    RW_WRITE_HELD(&ire->ire_bucket->irb_lock));
2557 	ASSERT(!IRE_IS_CONDEMNED(ire));
2558 	ire->ire_generation = IRE_GENERATION_CONDEMNED;
2559 	/* Count how many condemned ires for kmem_cache callback */
2560 	atomic_add_32(&ipst->ips_num_ire_condemned, 1);
2561 	mutex_exit(&ire->ire_lock);
2562 }
2563 
2564 /*
2565  * Increment the generation avoiding the special condemned value
2566  */
2567 void
2568 ire_increment_generation(ire_t *ire)
2569 {
2570 	uint_t generation;
2571 
2572 	mutex_enter(&ire->ire_lock);
2573 	/*
2574 	 * Even though the caller has a hold it can't prevent a concurrent
2575 	 * ire_delete marking the IRE condemned
2576 	 */
2577 	if (!IRE_IS_CONDEMNED(ire)) {
2578 		generation = ire->ire_generation + 1;
2579 		if (generation == IRE_GENERATION_CONDEMNED)
2580 			generation = IRE_GENERATION_INITIAL;
2581 		ASSERT(generation != IRE_GENERATION_VERIFY);
2582 		ire->ire_generation = generation;
2583 	}
2584 	mutex_exit(&ire->ire_lock);
2585 }
2586 
2587 /*
2588  * Increment ire_generation on all the IRE_MULTICASTs
2589  * Used when the default multicast interface (as determined by
2590  * ill_lookup_multicast) might have changed.
2591  *
2592  * That includes the zoneid, IFF_ flags, the IPv6 scope of the address, and
2593  * ill unplumb.
2594  */
2595 void
2596 ire_increment_multicast_generation(ip_stack_t *ipst, boolean_t isv6)
2597 {
2598 	ill_t	*ill;
2599 	ill_walk_context_t ctx;
2600 
2601 	rw_enter(&ipst->ips_ill_g_lock, RW_READER);
2602 	if (isv6)
2603 		ill = ILL_START_WALK_V6(&ctx, ipst);
2604 	else
2605 		ill = ILL_START_WALK_V4(&ctx, ipst);
2606 	for (; ill != NULL; ill = ill_next(&ctx, ill)) {
2607 		if (ILL_IS_CONDEMNED(ill))
2608 			continue;
2609 		if (ill->ill_ire_multicast != NULL)
2610 			ire_increment_generation(ill->ill_ire_multicast);
2611 	}
2612 	rw_exit(&ipst->ips_ill_g_lock);
2613 }
2614 
2615 /*
2616  * Return a held IRE_NOROUTE with RTF_REJECT set
2617  */
2618 ire_t *
2619 ire_reject(ip_stack_t *ipst, boolean_t isv6)
2620 {
2621 	ire_t *ire;
2622 
2623 	if (isv6)
2624 		ire = ipst->ips_ire_reject_v6;
2625 	else
2626 		ire = ipst->ips_ire_reject_v4;
2627 
2628 	ASSERT(ire->ire_generation != IRE_GENERATION_CONDEMNED);
2629 	ire_refhold(ire);
2630 	return (ire);
2631 }
2632 
2633 /*
2634  * Return a held IRE_NOROUTE with RTF_BLACKHOLE set
2635  */
2636 ire_t *
2637 ire_blackhole(ip_stack_t *ipst, boolean_t isv6)
2638 {
2639 	ire_t *ire;
2640 
2641 	if (isv6)
2642 		ire = ipst->ips_ire_blackhole_v6;
2643 	else
2644 		ire = ipst->ips_ire_blackhole_v4;
2645 
2646 	ASSERT(ire->ire_generation != IRE_GENERATION_CONDEMNED);
2647 	ire_refhold(ire);
2648 	return (ire);
2649 }
2650 
2651 /*
2652  * Return a held IRE_MULTICAST.
2653  */
2654 ire_t *
2655 ire_multicast(ill_t *ill)
2656 {
2657 	ire_t *ire = ill->ill_ire_multicast;
2658 
2659 	ASSERT(ire == NULL || ire->ire_generation != IRE_GENERATION_CONDEMNED);
2660 	if (ire == NULL)
2661 		ire = ire_blackhole(ill->ill_ipst, ill->ill_isv6);
2662 	else
2663 		ire_refhold(ire);
2664 	return (ire);
2665 }
2666 
2667 /*
2668  * Given an IRE return its nexthop IRE. The nexthop IRE is an IRE_ONLINK
2669  * that is an exact match (i.e., a /32 for IPv4 and /128 for IPv6).
2670  * This can return an RTF_REJECT|RTF_BLACKHOLE.
2671  * The returned IRE is held.
2672  * The assumption is that ip_select_route() has been called and returned the
2673  * IRE (thus ip_select_route would have set up the ire_dep* information.)
2674  * If some IRE is deleteted then ire_dep_remove() will have been called and
2675  * we might not find a nexthop IRE, in which case we return NULL.
2676  */
2677 ire_t *
2678 ire_nexthop(ire_t *ire)
2679 {
2680 	ip_stack_t	*ipst = ire->ire_ipst;
2681 
2682 	/* Acquire lock to walk ire_dep_parent */
2683 	rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
2684 	while (ire != NULL) {
2685 		if (ire->ire_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
2686 			goto done;
2687 		}
2688 		/*
2689 		 * If we find an IRE_ONLINK we are done. This includes
2690 		 * the case of IRE_MULTICAST.
2691 		 * Note that in order to send packets we need a host-specific
2692 		 * IRE_IF_ALL first in the ire_dep_parent chain. Normally this
2693 		 * is done by inserting an IRE_IF_CLONE if the IRE_INTERFACE
2694 		 * was not host specific.
2695 		 * However, ip_rts_request doesn't want to send packets
2696 		 * hence doesn't want to allocate an IRE_IF_CLONE. Yet
2697 		 * it needs an IRE_IF_ALL to get to the ill. Thus
2698 		 * we return IRE_IF_ALL that are not host specific here.
2699 		 */
2700 		if (ire->ire_type & IRE_ONLINK)
2701 			goto done;
2702 		ire = ire->ire_dep_parent;
2703 	}
2704 	rw_exit(&ipst->ips_ire_dep_lock);
2705 	return (NULL);
2706 
2707 done:
2708 	ire_refhold(ire);
2709 	rw_exit(&ipst->ips_ire_dep_lock);
2710 	return (ire);
2711 }
2712 
2713 /*
2714  * Find the ill used to send packets. This will be NULL in case
2715  * of a reject or blackhole.
2716  * The returned ill is held; caller needs to do ill_refrele when done.
2717  */
2718 ill_t *
2719 ire_nexthop_ill(ire_t *ire)
2720 {
2721 	ill_t		*ill;
2722 
2723 	ire = ire_nexthop(ire);
2724 	if (ire == NULL)
2725 		return (NULL);
2726 
2727 	/* ire_ill can not change for an existing ire */
2728 	ill = ire->ire_ill;
2729 	if (ill != NULL)
2730 		ill_refhold(ill);
2731 	ire_refrele(ire);
2732 	return (ill);
2733 }
2734 
2735 #ifdef DEBUG
2736 static boolean_t
2737 parent_has_child(ire_t *parent, ire_t *child)
2738 {
2739 	ire_t	*ire;
2740 	ire_t	*prev;
2741 
2742 	ire = parent->ire_dep_children;
2743 	prev = NULL;
2744 	while (ire != NULL) {
2745 		if (prev == NULL) {
2746 			ASSERT(ire->ire_dep_sib_ptpn ==
2747 			    &(parent->ire_dep_children));
2748 		} else {
2749 			ASSERT(ire->ire_dep_sib_ptpn ==
2750 			    &(prev->ire_dep_sib_next));
2751 		}
2752 		if (ire == child)
2753 			return (B_TRUE);
2754 		prev = ire;
2755 		ire = ire->ire_dep_sib_next;
2756 	}
2757 	return (B_FALSE);
2758 }
2759 
2760 static void
2761 ire_dep_verify(ire_t *ire)
2762 {
2763 	ire_t		*parent = ire->ire_dep_parent;
2764 	ire_t		*child = ire->ire_dep_children;
2765 
2766 	ASSERT(ire->ire_ipversion == IPV4_VERSION ||
2767 	    ire->ire_ipversion == IPV6_VERSION);
2768 	if (parent != NULL) {
2769 		ASSERT(parent->ire_ipversion == IPV4_VERSION ||
2770 		    parent->ire_ipversion == IPV6_VERSION);
2771 		ASSERT(parent->ire_refcnt >= 1);
2772 		ASSERT(parent_has_child(parent, ire));
2773 	}
2774 	if (child != NULL) {
2775 		ASSERT(child->ire_ipversion == IPV4_VERSION ||
2776 		    child->ire_ipversion == IPV6_VERSION);
2777 		ASSERT(child->ire_dep_parent == ire);
2778 		ASSERT(child->ire_dep_sib_ptpn != NULL);
2779 		ASSERT(parent_has_child(ire, child));
2780 	}
2781 }
2782 #endif /* DEBUG */
2783 
2784 /*
2785  * Assumes ire_dep_parent is set. Remove this child from its parent's linkage.
2786  */
2787 void
2788 ire_dep_remove(ire_t *ire)
2789 {
2790 	ip_stack_t	*ipst = ire->ire_ipst;
2791 	ire_t		*parent = ire->ire_dep_parent;
2792 	ire_t		*next;
2793 	nce_t		*nce;
2794 
2795 	ASSERT(RW_WRITE_HELD(&ipst->ips_ire_dep_lock));
2796 	ASSERT(ire->ire_dep_parent != NULL);
2797 	ASSERT(ire->ire_dep_sib_ptpn != NULL);
2798 
2799 #ifdef DEBUG
2800 	ire_dep_verify(ire);
2801 	ire_dep_verify(parent);
2802 #endif
2803 
2804 	next = ire->ire_dep_sib_next;
2805 	if (next != NULL)
2806 		next->ire_dep_sib_ptpn = ire->ire_dep_sib_ptpn;
2807 
2808 	ASSERT(*(ire->ire_dep_sib_ptpn) == ire);
2809 	*(ire->ire_dep_sib_ptpn) = ire->ire_dep_sib_next;
2810 
2811 	ire->ire_dep_sib_ptpn = NULL;
2812 	ire->ire_dep_sib_next = NULL;
2813 
2814 	mutex_enter(&ire->ire_lock);
2815 	parent = ire->ire_dep_parent;
2816 	ire->ire_dep_parent = NULL;
2817 	mutex_exit(&ire->ire_lock);
2818 
2819 	/*
2820 	 * Make sure all our children, grandchildren, etc set
2821 	 * ire_dep_parent_generation to IRE_GENERATION_VERIFY since
2822 	 * we can no longer guarantee than the children have a current
2823 	 * ire_nce_cache and ire_nexthop_ill().
2824 	 */
2825 	if (ire->ire_dep_children != NULL)
2826 		ire_dep_invalidate_children(ire->ire_dep_children);
2827 
2828 	/*
2829 	 * Since the parent is gone we make sure we clear ire_nce_cache.
2830 	 * We can clear it under ire_lock even if the IRE is used
2831 	 */
2832 	mutex_enter(&ire->ire_lock);
2833 	nce = ire->ire_nce_cache;
2834 	ire->ire_nce_cache = NULL;
2835 	mutex_exit(&ire->ire_lock);
2836 	if (nce != NULL)
2837 		nce_refrele(nce);
2838 
2839 #ifdef DEBUG
2840 	ire_dep_verify(ire);
2841 	ire_dep_verify(parent);
2842 #endif
2843 
2844 	ire_refrele_notr(parent);
2845 	ire_refrele_notr(ire);
2846 }
2847 
2848 /*
2849  * Insert the child in the linkage of the parent
2850  */
2851 static void
2852 ire_dep_parent_insert(ire_t *child, ire_t *parent)
2853 {
2854 	ip_stack_t	*ipst = child->ire_ipst;
2855 	ire_t		*next;
2856 
2857 	ASSERT(RW_WRITE_HELD(&ipst->ips_ire_dep_lock));
2858 	ASSERT(child->ire_dep_parent == NULL);
2859 
2860 #ifdef DEBUG
2861 	ire_dep_verify(child);
2862 	ire_dep_verify(parent);
2863 #endif
2864 	/* No parents => no siblings */
2865 	ASSERT(child->ire_dep_sib_ptpn == NULL);
2866 	ASSERT(child->ire_dep_sib_next == NULL);
2867 
2868 	ire_refhold_notr(parent);
2869 	ire_refhold_notr(child);
2870 
2871 	/* Head insertion */
2872 	next = parent->ire_dep_children;
2873 	if (next != NULL) {
2874 		ASSERT(next->ire_dep_sib_ptpn == &(parent->ire_dep_children));
2875 		child->ire_dep_sib_next = next;
2876 		next->ire_dep_sib_ptpn = &(child->ire_dep_sib_next);
2877 	}
2878 	parent->ire_dep_children = child;
2879 	child->ire_dep_sib_ptpn = &(parent->ire_dep_children);
2880 
2881 	mutex_enter(&child->ire_lock);
2882 	child->ire_dep_parent = parent;
2883 	mutex_exit(&child->ire_lock);
2884 
2885 #ifdef DEBUG
2886 	ire_dep_verify(child);
2887 	ire_dep_verify(parent);
2888 #endif
2889 }
2890 
2891 
2892 /*
2893  * Given count worth of ires and generations, build ire_dep_* relationships
2894  * from ires[0] to ires[count-1]. Record generations[i+1] in
2895  * ire_dep_parent_generation for ires[i].
2896  * We graft onto an existing parent chain by making sure that we don't
2897  * touch ire_dep_parent for ires[count-1].
2898  *
2899  * We check for any condemned ire_generation count and return B_FALSE in
2900  * that case so that the caller can tear it apart.
2901  *
2902  * Note that generations[0] is not used. Caller handles that.
2903  */
2904 boolean_t
2905 ire_dep_build(ire_t *ires[], uint_t generations[], uint_t count)
2906 {
2907 	ire_t		*ire = ires[0];
2908 	ip_stack_t	*ipst;
2909 	uint_t		i;
2910 
2911 	ASSERT(count > 0);
2912 	if (count == 1) {
2913 		/* No work to do */
2914 		return (B_TRUE);
2915 	}
2916 	ipst = ire->ire_ipst;
2917 	rw_enter(&ipst->ips_ire_dep_lock, RW_WRITER);
2918 	/*
2919 	 * Do not remove the linkage for any existing parent chain i.e.,
2920 	 * ires[count-1] is left alone.
2921 	 */
2922 	for (i = 0; i < count-1; i++) {
2923 		/* Remove existing parent if we need to change it */
2924 		if (ires[i]->ire_dep_parent != NULL &&
2925 		    ires[i]->ire_dep_parent != ires[i+1])
2926 			ire_dep_remove(ires[i]);
2927 	}
2928 
2929 	for (i = 0; i < count - 1; i++) {
2930 		ASSERT(ires[i]->ire_ipversion == IPV4_VERSION ||
2931 		    ires[i]->ire_ipversion == IPV6_VERSION);
2932 		/* Does it need to change? */
2933 		if (ires[i]->ire_dep_parent != ires[i+1])
2934 			ire_dep_parent_insert(ires[i], ires[i+1]);
2935 
2936 		mutex_enter(&ires[i+1]->ire_lock);
2937 		if (IRE_IS_CONDEMNED(ires[i+1])) {
2938 			mutex_exit(&ires[i+1]->ire_lock);
2939 			rw_exit(&ipst->ips_ire_dep_lock);
2940 			return (B_FALSE);
2941 		}
2942 		mutex_exit(&ires[i+1]->ire_lock);
2943 
2944 		mutex_enter(&ires[i]->ire_lock);
2945 		ires[i]->ire_dep_parent_generation = generations[i+1];
2946 		mutex_exit(&ires[i]->ire_lock);
2947 	}
2948 	rw_exit(&ipst->ips_ire_dep_lock);
2949 	return (B_TRUE);
2950 }
2951 
2952 /*
2953  * Given count worth of ires, unbuild ire_dep_* relationships
2954  * from ires[0] to ires[count-1].
2955  */
2956 void
2957 ire_dep_unbuild(ire_t *ires[], uint_t count)
2958 {
2959 	ip_stack_t	*ipst;
2960 	uint_t		i;
2961 
2962 	if (count == 0) {
2963 		/* No work to do */
2964 		return;
2965 	}
2966 	ipst = ires[0]->ire_ipst;
2967 	rw_enter(&ipst->ips_ire_dep_lock, RW_WRITER);
2968 	for (i = 0; i < count; i++) {
2969 		ASSERT(ires[i]->ire_ipversion == IPV4_VERSION ||
2970 		    ires[i]->ire_ipversion == IPV6_VERSION);
2971 		if (ires[i]->ire_dep_parent != NULL)
2972 			ire_dep_remove(ires[i]);
2973 		mutex_enter(&ires[i]->ire_lock);
2974 		ires[i]->ire_dep_parent_generation = IRE_GENERATION_VERIFY;
2975 		mutex_exit(&ires[i]->ire_lock);
2976 	}
2977 	rw_exit(&ipst->ips_ire_dep_lock);
2978 }
2979 
2980 /*
2981  * Both the forwarding and the outbound code paths can trip on
2982  * a condemned NCE, in which case we call this function.
2983  * We have two different behaviors: if the NCE was UNREACHABLE
2984  * it is an indication that something failed. In that case
2985  * we see if we should look for a different IRE (for example,
2986  * delete any matching redirect IRE, or try a different
2987  * IRE_DEFAULT (ECMP)). We mark the ire as bad so a hopefully
2988  * different IRE will be picked next time we send/forward.
2989  *
2990  * If we are called by the output path then fail_if_better is set
2991  * and we return NULL if there could be a better IRE. This is because the
2992  * output path retries the IRE lookup. (The input/forward path can not retry.)
2993  *
2994  * If the NCE was not unreachable then we pick/allocate a
2995  * new (most likely ND_INITIAL) NCE and proceed with it.
2996  *
2997  * ipha/ip6h are needed for multicast packets; ipha needs to be
2998  * set for IPv4 and ip6h needs to be set for IPv6 packets.
2999  */
3000 nce_t *
3001 ire_handle_condemned_nce(nce_t *nce, ire_t *ire, ipha_t *ipha, ip6_t *ip6h,
3002     boolean_t fail_if_better)
3003 {
3004 	if (nce->nce_common->ncec_state == ND_UNREACHABLE) {
3005 		if (ire_no_good(ire) && fail_if_better) {
3006 			/*
3007 			 * Did some changes, or ECMP likely to exist.
3008 			 * Make ip_output look for a different IRE
3009 			 */
3010 			return (NULL);
3011 		}
3012 	}
3013 	if (ire_revalidate_nce(ire) == ENETUNREACH) {
3014 		/* The ire_dep_parent chain went bad, or no memory? */
3015 		(void) ire_no_good(ire);
3016 		return (NULL);
3017 	}
3018 	if (ire->ire_ipversion == IPV4_VERSION) {
3019 		ASSERT(ipha != NULL);
3020 		nce = ire_to_nce(ire, ipha->ipha_dst, NULL);
3021 	} else {
3022 		ASSERT(ip6h != NULL);
3023 		nce = ire_to_nce(ire, INADDR_ANY, &ip6h->ip6_dst);
3024 	}
3025 
3026 	if (nce == NULL)
3027 		return (NULL);
3028 	if (nce->nce_is_condemned) {
3029 		nce_refrele(nce);
3030 		return (NULL);
3031 	}
3032 	return (nce);
3033 }
3034 
3035 /*
3036  * The caller has found that the ire is bad, either due to a reference to an NCE
3037  * in ND_UNREACHABLE state, or a MULTIRT route whose gateway can't be resolved.
3038  * We update things so a subsequent attempt to send to the destination
3039  * is likely to find different IRE, or that a new NCE would be created.
3040  *
3041  * Returns B_TRUE if it is likely that a subsequent ire_ftable_lookup would
3042  * find a different route (either due to having deleted a redirect, or there
3043  * being ECMP routes.)
3044  *
3045  * If we have a redirect (RTF_DYNAMIC) we delete it.
3046  * Otherwise we increment ire_badcnt and increment the generation number so
3047  * that a cached ixa_ire will redo the route selection. ire_badcnt is taken
3048  * into account in the route selection when we have multiple choices (multiple
3049  * default routes or ECMP in general).
3050  * Any time ip_select_route find an ire with a condemned ire_nce_cache
3051  * (e.g., if no equal cost route to the bad one) ip_select_route will make
3052  * sure the NCE is revalidated to avoid getting stuck on a
3053  * NCE_F_CONDMNED ncec that caused ire_no_good to be called.
3054  */
3055 boolean_t
3056 ire_no_good(ire_t *ire)
3057 {
3058 	ip_stack_t	*ipst = ire->ire_ipst;
3059 	ire_t		*ire2;
3060 	nce_t		*nce;
3061 
3062 	if (ire->ire_flags & RTF_DYNAMIC) {
3063 		ire_delete(ire);
3064 		return (B_TRUE);
3065 	}
3066 	if (ire->ire_flags & RTF_INDIRECT) {
3067 		/* Check if next IRE is a redirect */
3068 		rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
3069 		if (ire->ire_dep_parent != NULL &&
3070 		    (ire->ire_dep_parent->ire_flags & RTF_DYNAMIC)) {
3071 			ire2 = ire->ire_dep_parent;
3072 			ire_refhold(ire2);
3073 		} else {
3074 			ire2 = NULL;
3075 		}
3076 		rw_exit(&ipst->ips_ire_dep_lock);
3077 		if (ire2 != NULL) {
3078 			ire_delete(ire2);
3079 			ire_refrele(ire2);
3080 			return (B_TRUE);
3081 		}
3082 	}
3083 	/*
3084 	 * No redirect involved. Increment badcnt so that if we have ECMP
3085 	 * routes we are likely to pick a different one for the next packet.
3086 	 *
3087 	 * If the NCE is unreachable and condemned we should drop the reference
3088 	 * to it so that a new NCE can be created.
3089 	 *
3090 	 * Finally we increment the generation number so that any ixa_ire
3091 	 * cache will be revalidated.
3092 	 */
3093 	mutex_enter(&ire->ire_lock);
3094 	ire->ire_badcnt++;
3095 	ire->ire_last_badcnt = TICK_TO_SEC(lbolt64);
3096 	nce = ire->ire_nce_cache;
3097 	if (nce != NULL && nce->nce_is_condemned &&
3098 	    nce->nce_common->ncec_state == ND_UNREACHABLE)
3099 		ire->ire_nce_cache = NULL;
3100 	else
3101 		nce = NULL;
3102 	mutex_exit(&ire->ire_lock);
3103 	if (nce != NULL)
3104 		nce_refrele(nce);
3105 
3106 	ire_increment_generation(ire);
3107 	ire_dep_incr_generation(ire);
3108 
3109 	return (ire->ire_bucket->irb_ire_cnt > 1);
3110 }
3111 
3112 /*
3113  * Walk ire_dep_parent chain and validate that ire_dep_parent->ire_generation ==
3114  * ire_dep_parent_generation.
3115  * If they all match we just return ire_generation from the topmost IRE.
3116  * Otherwise we propagate the mismatch by setting all ire_dep_parent_generation
3117  * above the mismatch to IRE_GENERATION_VERIFY and also returning
3118  * IRE_GENERATION_VERIFY.
3119  */
3120 uint_t
3121 ire_dep_validate_generations(ire_t *ire)
3122 {
3123 	ip_stack_t	*ipst = ire->ire_ipst;
3124 	uint_t		generation;
3125 	ire_t		*ire1;
3126 
3127 	rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
3128 	generation = ire->ire_generation;	/* Assuming things match */
3129 	for (ire1 = ire; ire1 != NULL; ire1 = ire1->ire_dep_parent) {
3130 		ASSERT(ire1->ire_ipversion == IPV4_VERSION ||
3131 		    ire1->ire_ipversion == IPV6_VERSION);
3132 		if (ire1->ire_dep_parent == NULL)
3133 			break;
3134 		if (ire1->ire_dep_parent_generation !=
3135 		    ire1->ire_dep_parent->ire_generation)
3136 			goto mismatch;
3137 	}
3138 	rw_exit(&ipst->ips_ire_dep_lock);
3139 	return (generation);
3140 
3141 mismatch:
3142 	generation = IRE_GENERATION_VERIFY;
3143 	/* Fill from top down to the mismatch with _VERIFY */
3144 	while (ire != ire1) {
3145 		ASSERT(ire->ire_ipversion == IPV4_VERSION ||
3146 		    ire->ire_ipversion == IPV6_VERSION);
3147 		mutex_enter(&ire->ire_lock);
3148 		ire->ire_dep_parent_generation = IRE_GENERATION_VERIFY;
3149 		mutex_exit(&ire->ire_lock);
3150 		ire = ire->ire_dep_parent;
3151 	}
3152 	rw_exit(&ipst->ips_ire_dep_lock);
3153 	return (generation);
3154 }
3155 
3156 /*
3157  * Used when we need to return an ire with ire_dep_parent, but we
3158  * know the chain is invalid for instance we didn't create an IRE_IF_CLONE
3159  * Using IRE_GENERATION_VERIFY means that next time we'll redo the
3160  * recursive lookup.
3161  */
3162 void
3163 ire_dep_invalidate_generations(ire_t *ire)
3164 {
3165 	ip_stack_t	*ipst = ire->ire_ipst;
3166 
3167 	rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
3168 	while (ire != NULL) {
3169 		ASSERT(ire->ire_ipversion == IPV4_VERSION ||
3170 		    ire->ire_ipversion == IPV6_VERSION);
3171 		mutex_enter(&ire->ire_lock);
3172 		ire->ire_dep_parent_generation = IRE_GENERATION_VERIFY;
3173 		mutex_exit(&ire->ire_lock);
3174 		ire = ire->ire_dep_parent;
3175 	}
3176 	rw_exit(&ipst->ips_ire_dep_lock);
3177 }
3178 
3179 /* Set _VERIFY ire_dep_parent_generation for all children recursively */
3180 static void
3181 ire_dep_invalidate_children(ire_t *child)
3182 {
3183 	ip_stack_t	*ipst = child->ire_ipst;
3184 
3185 	ASSERT(RW_WRITE_HELD(&ipst->ips_ire_dep_lock));
3186 	/* Depth first */
3187 	if (child->ire_dep_children != NULL)
3188 		ire_dep_invalidate_children(child->ire_dep_children);
3189 
3190 	while (child != NULL) {
3191 		mutex_enter(&child->ire_lock);
3192 		child->ire_dep_parent_generation = IRE_GENERATION_VERIFY;
3193 		mutex_exit(&child->ire_lock);
3194 		child = child->ire_dep_sib_next;
3195 	}
3196 }
3197 
3198 static void
3199 ire_dep_increment_children(ire_t *child)
3200 {
3201 	ip_stack_t	*ipst = child->ire_ipst;
3202 
3203 	ASSERT(RW_READ_HELD(&ipst->ips_ire_dep_lock));
3204 	/* Depth first */
3205 	if (child->ire_dep_children != NULL)
3206 		ire_dep_increment_children(child->ire_dep_children);
3207 
3208 	while (child != NULL) {
3209 		if (!IRE_IS_CONDEMNED(child))
3210 			ire_increment_generation(child);
3211 		child = child->ire_dep_sib_next;
3212 	}
3213 }
3214 
3215 /*
3216  * Walk all the children of this ire recursively and increment their
3217  * generation number.
3218  */
3219 void
3220 ire_dep_incr_generation(ire_t *parent)
3221 {
3222 	ip_stack_t	*ipst = parent->ire_ipst;
3223 
3224 	rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
3225 	if (parent->ire_dep_children != NULL)
3226 		ire_dep_increment_children(parent->ire_dep_children);
3227 	rw_exit(&ipst->ips_ire_dep_lock);
3228 }
3229 
3230 /*
3231  * Get a new ire_nce_cache for this IRE as well as its nexthop.
3232  * Returns zero if it succeeds. Can fail due to lack of memory or when
3233  * the route has become unreachable. Returns ENOMEM and ENETUNREACH in those
3234  * cases.
3235  *
3236  * In the in.mpathd case, the ire will have ire_testhidden
3237  * set; so we should create the ncec for the underlying ill.
3238  *
3239  * Note that the error returned by ire_revalidate_nce() is ignored by most
3240  * callers except ire_handle_condemned_nce(), which handles the ENETUNREACH
3241  * error to mark potentially bad ire's. For all the other callers, an
3242  * error return could indicate a transient condition like ENOMEM, or could
3243  * be the result of an interface that is going down/unplumbing. In the former
3244  * case (transient error), we would leave the old stale ire/ire_nce_cache
3245  * in place, and possibly use incorrect link-layer information to send packets
3246  * but would eventually recover. In the latter case (ill down/replumb),
3247  * ire_revalidate_nce() might return a condemned nce back, but we would then
3248  * recover in the packet output path.
3249  */
3250 int
3251 ire_revalidate_nce(ire_t *ire)
3252 {
3253 	nce_t		*nce, *old_nce;
3254 	ire_t		*nexthop;
3255 
3256 	/*
3257 	 * For multicast we conceptually have an NCE but we don't store it
3258 	 * in ire_nce_cache; when ire_to_nce is called we allocate the nce.
3259 	 */
3260 	if (ire->ire_type & IRE_MULTICAST)
3261 		return (0);
3262 
3263 	/* ire_testhidden should only be set on under-interfaces */
3264 	ASSERT(!ire->ire_testhidden || !IS_IPMP(ire->ire_ill));
3265 
3266 	nexthop = ire_nexthop(ire);
3267 	if (nexthop == NULL) {
3268 		/* The route is potentially bad */
3269 		(void) ire_no_good(ire);
3270 		return (ENETUNREACH);
3271 	}
3272 	if (ire->ire_type & (IRE_LOCAL|IRE_LOOPBACK)) {
3273 		ASSERT(ire->ire_ill != NULL);
3274 
3275 		if (ire->ire_ipversion == IPV4_VERSION)
3276 			nce = nce_lookup_v4(ire->ire_ill, &ire->ire_addr);
3277 		else
3278 			nce = nce_lookup_v6(ire->ire_ill, &ire->ire_addr_v6);
3279 	} else {
3280 		ASSERT(nexthop->ire_type & IRE_ONLINK);
3281 		if (ire->ire_ipversion == IPV4_VERSION) {
3282 			nce = arp_nce_init(nexthop->ire_ill, nexthop->ire_addr,
3283 			    nexthop->ire_type);
3284 		} else {
3285 			nce = ndp_nce_init(nexthop->ire_ill,
3286 			    &nexthop->ire_addr_v6, nexthop->ire_type);
3287 		}
3288 	}
3289 	if (nce == NULL) {
3290 		/*
3291 		 * Leave the old stale one in place to avoid a NULL
3292 		 * ire_nce_cache.
3293 		 */
3294 		ire_refrele(nexthop);
3295 		return (ENOMEM);
3296 	}
3297 
3298 	if (nexthop != ire) {
3299 		/* Update the nexthop ire */
3300 		mutex_enter(&nexthop->ire_lock);
3301 		old_nce = nexthop->ire_nce_cache;
3302 		if (!IRE_IS_CONDEMNED(nexthop)) {
3303 			nce_refhold(nce);
3304 			nexthop->ire_nce_cache = nce;
3305 		} else {
3306 			nexthop->ire_nce_cache = NULL;
3307 		}
3308 		mutex_exit(&nexthop->ire_lock);
3309 		if (old_nce != NULL)
3310 			nce_refrele(old_nce);
3311 	}
3312 	ire_refrele(nexthop);
3313 
3314 	mutex_enter(&ire->ire_lock);
3315 	old_nce = ire->ire_nce_cache;
3316 	if (!IRE_IS_CONDEMNED(ire)) {
3317 		nce_refhold(nce);
3318 		ire->ire_nce_cache = nce;
3319 	} else {
3320 		ire->ire_nce_cache = NULL;
3321 	}
3322 	mutex_exit(&ire->ire_lock);
3323 	if (old_nce != NULL)
3324 		nce_refrele(old_nce);
3325 
3326 	nce_refrele(nce);
3327 	return (0);
3328 }
3329 
3330 /*
3331  * Get a held nce for a given ire.
3332  * In the common case this is just from ire_nce_cache.
3333  * For IRE_MULTICAST this needs to do an explicit lookup since we do not
3334  * have an IRE_MULTICAST per address.
3335  * Note that this explicitly returns CONDEMNED NCEs. The caller needs those
3336  * so they can check whether the NCE went unreachable (as opposed to was
3337  * condemned for some other reason).
3338  */
3339 nce_t *
3340 ire_to_nce(ire_t *ire, ipaddr_t v4nexthop, const in6_addr_t *v6nexthop)
3341 {
3342 	nce_t	*nce;
3343 
3344 	if (ire->ire_flags & (RTF_REJECT|RTF_BLACKHOLE))
3345 		return (NULL);
3346 
3347 	/* ire_testhidden should only be set on under-interfaces */
3348 	ASSERT(!ire->ire_testhidden || !IS_IPMP(ire->ire_ill));
3349 
3350 	mutex_enter(&ire->ire_lock);
3351 	nce = ire->ire_nce_cache;
3352 	if (nce != NULL) {
3353 		nce_refhold(nce);
3354 		mutex_exit(&ire->ire_lock);
3355 		return (nce);
3356 	}
3357 	mutex_exit(&ire->ire_lock);
3358 
3359 	if (ire->ire_type & IRE_MULTICAST) {
3360 		ASSERT(ire->ire_ill != NULL);
3361 
3362 		if (ire->ire_ipversion == IPV4_VERSION) {
3363 			ASSERT(v6nexthop == NULL);
3364 
3365 			nce = arp_nce_init(ire->ire_ill, v4nexthop,
3366 			    ire->ire_type);
3367 		} else {
3368 			ASSERT(v6nexthop != NULL);
3369 			ASSERT(v4nexthop == 0);
3370 			nce = ndp_nce_init(ire->ire_ill, v6nexthop,
3371 			    ire->ire_type);
3372 		}
3373 		return (nce);
3374 	}
3375 	return (NULL);
3376 }
3377 
3378 nce_t *
3379 ire_to_nce_pkt(ire_t *ire, mblk_t *mp)
3380 {
3381 	ipha_t		*ipha;
3382 	ip6_t		*ip6h;
3383 
3384 	if (IPH_HDR_VERSION(mp->b_rptr) == IPV4_VERSION) {
3385 		ipha = (ipha_t *)mp->b_rptr;
3386 		return (ire_to_nce(ire, ipha->ipha_dst, NULL));
3387 	} else {
3388 		ip6h = (ip6_t *)mp->b_rptr;
3389 		return (ire_to_nce(ire, INADDR_ANY, &ip6h->ip6_dst));
3390 	}
3391 }
3392 
3393 /*
3394  * Given an IRE_INTERFACE (that matches more than one address) create
3395  * and return an IRE_IF_CLONE for the specific address.
3396  * Return the generation number.
3397  * Returns NULL is no memory for the IRE.
3398  * Handles both IPv4 and IPv6.
3399  */
3400 ire_t *
3401 ire_create_if_clone(ire_t *ire_if, const in6_addr_t *addr, uint_t *generationp)
3402 {
3403 	ire_t		*ire;
3404 	ire_t		*nire;
3405 
3406 	if (ire_if->ire_ipversion == IPV4_VERSION) {
3407 		ipaddr_t	v4addr;
3408 		ipaddr_t	mask = IP_HOST_MASK;
3409 
3410 		ASSERT(IN6_IS_ADDR_V4MAPPED(addr));
3411 		IN6_V4MAPPED_TO_IPADDR(addr, v4addr);
3412 
3413 		ire = ire_create(
3414 		    (uchar_t *)&v4addr,			/* dest address */
3415 		    (uchar_t *)&mask,			/* mask */
3416 		    (uchar_t *)&ire_if->ire_gateway_addr,
3417 		    IRE_IF_CLONE,			/* IRE type */
3418 		    ire_if->ire_ill,
3419 		    ire_if->ire_zoneid,
3420 		    ire_if->ire_flags | RTF_HOST,
3421 		    NULL,		/* No security attr for IRE_IF_ALL */
3422 		    ire_if->ire_ipst);
3423 	} else {
3424 		ASSERT(!IN6_IS_ADDR_V4MAPPED(addr));
3425 		ire = ire_create_v6(
3426 		    addr,				/* dest address */
3427 		    &ipv6_all_ones,			/* mask */
3428 		    &ire_if->ire_gateway_addr_v6,	/* gateway addr */
3429 		    IRE_IF_CLONE,			/* IRE type */
3430 		    ire_if->ire_ill,
3431 		    ire_if->ire_zoneid,
3432 		    ire_if->ire_flags | RTF_HOST,
3433 		    NULL,		/* No security attr for IRE_IF_ALL */
3434 		    ire_if->ire_ipst);
3435 	}
3436 	if (ire == NULL)
3437 		return (NULL);
3438 
3439 	/* Take the metrics, in particular the mtu, from the IRE_IF */
3440 	ire->ire_metrics = ire_if->ire_metrics;
3441 
3442 	nire = ire_add(ire);
3443 	if (nire == NULL) /* Some failure */
3444 		return (NULL);
3445 
3446 	if (generationp != NULL)
3447 		*generationp = nire->ire_generation;
3448 
3449 	/*
3450 	 * Make sure races don't add a duplicate by
3451 	 * catching the case when an identical was returned.
3452 	 */
3453 	if (nire != ire) {
3454 		ASSERT(nire->ire_identical_ref > 1);
3455 		ire_delete(nire);
3456 	}
3457 	return (nire);
3458 }
3459 
3460 /*
3461  * The argument is an IRE_INTERFACE. Delete all of IRE_IF_CLONE in the
3462  * ire_dep_children (just walk the ire_dep_sib_next since they are all
3463  * immediate children.)
3464  * Since we hold a lock while we remove them we need to defer the actual
3465  * calls to ire_delete() until we have dropped the lock. This makes things
3466  * less efficient since we restart at the top after dropping the lock. But
3467  * we only run when an IRE_INTERFACE is deleted which is infrquent.
3468  *
3469  * Note that ire_dep_children can be any mixture of offlink routes and
3470  * IRE_IF_CLONE entries.
3471  */
3472 void
3473 ire_dep_delete_if_clone(ire_t *parent)
3474 {
3475 	ip_stack_t	*ipst = parent->ire_ipst;
3476 	ire_t		*child, *next;
3477 
3478 restart:
3479 	rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
3480 	if (parent->ire_dep_children == NULL) {
3481 		rw_exit(&ipst->ips_ire_dep_lock);
3482 		return;
3483 	}
3484 	child = parent->ire_dep_children;
3485 	while (child != NULL) {
3486 		next = child->ire_dep_sib_next;
3487 		if ((child->ire_type & IRE_IF_CLONE) &&
3488 		    !IRE_IS_CONDEMNED(child)) {
3489 			ire_refhold(child);
3490 			rw_exit(&ipst->ips_ire_dep_lock);
3491 			ire_delete(child);
3492 			ASSERT(IRE_IS_CONDEMNED(child));
3493 			ire_refrele(child);
3494 			goto restart;
3495 		}
3496 		child = next;
3497 	}
3498 	rw_exit(&ipst->ips_ire_dep_lock);
3499 }
3500 
3501 /*
3502  * ire_pref() is used in recursive route-resolution for a destination to
3503  * determine the preference of an ire, where "preference" is determined
3504  * based on the level of indirection to the destination of the ire.
3505  * A higher preference indicates that fewer lookups are needed to complete
3506  * recursive route lookup. Thus
3507  * ire_pref(RTF_INDIRECT) < ire_pref(IRE_IF_RESOLVER) < ire_pref(IRE_PREF_CLONE)
3508  */
3509 int
3510 ire_pref(ire_t *ire)
3511 {
3512 	if (ire->ire_flags & RTF_INDIRECT)
3513 		return (1);
3514 	if (ire->ire_type & IRE_OFFLINK)
3515 		return (2);
3516 	if (ire->ire_type & (IRE_IF_RESOLVER|IRE_IF_NORESOLVER))
3517 		return (3);
3518 	if (ire->ire_type & IRE_IF_CLONE)
3519 		return (4);
3520 	if (ire->ire_type & (IRE_LOCAL|IRE_LOOPBACK|IRE_BROADCAST))
3521 		return (5);
3522 	return (-1); /* unknown ire_type */
3523 }
3524