xref: /illumos-gate/usr/src/uts/common/inet/ip/ip_ire.c (revision 63d2ef3c)
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 (c) 1991, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 1990 Mentat Inc.
24  */
25 
26 /*
27  * This file contains routines that manipulate Internet Routing Entries (IREs).
28  */
29 
30 #include <sys/types.h>
31 #include <sys/stream.h>
32 #include <sys/stropts.h>
33 #include <sys/strsun.h>
34 #include <sys/strsubr.h>
35 #include <sys/ddi.h>
36 #include <sys/cmn_err.h>
37 #include <sys/policy.h>
38 
39 #include <sys/systm.h>
40 #include <sys/kmem.h>
41 #include <sys/param.h>
42 #include <sys/socket.h>
43 #include <net/if.h>
44 #include <net/route.h>
45 #include <netinet/in.h>
46 #include <net/if_dl.h>
47 #include <netinet/ip6.h>
48 #include <netinet/icmp6.h>
49 
50 #include <inet/common.h>
51 #include <inet/mi.h>
52 #include <inet/ip.h>
53 #include <inet/ip6.h>
54 #include <inet/ip_ndp.h>
55 #include <inet/arp.h>
56 #include <inet/ip_if.h>
57 #include <inet/ip_ire.h>
58 #include <inet/ip_ftable.h>
59 #include <inet/ip_rts.h>
60 #include <inet/nd.h>
61 #include <inet/tunables.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 static void	ire_dep_incr_generation_locked(ire_t *);
234 
235 /*
236  * Following are the functions to increment/decrement the reference
237  * count of the IREs and IRBs (ire bucket).
238  *
239  * 1) We bump up the reference count of an IRE to make sure that
240  *    it does not get deleted and freed while we are using it.
241  *    Typically all the lookup functions hold the bucket lock,
242  *    and look for the IRE. If it finds an IRE, it bumps up the
243  *    reference count before dropping the lock. Sometimes we *may* want
244  *    to bump up the reference count after we *looked* up i.e without
245  *    holding the bucket lock. So, the ire_refhold function does not assert
246  *    on the bucket lock being held. Any thread trying to delete from
247  *    the hash bucket can still do so but cannot free the IRE if
248  *    ire_refcnt is not 0.
249  *
250  * 2) We bump up the reference count on the bucket where the IRE resides
251  *    (IRB), when we want to prevent the IREs getting deleted from a given
252  *    hash bucket. This makes life easier for ire_walk type functions which
253  *    wants to walk the IRE list, call a function, but needs to drop
254  *    the bucket lock to prevent recursive rw_enters. While the
255  *    lock is dropped, the list could be changed by other threads or
256  *    the same thread could end up deleting the ire or the ire pointed by
257  *    ire_next. ire_refholding the ire or ire_next is not sufficient as
258  *    a delete will still remove the ire from the bucket while we have
259  *    dropped the lock and hence the ire_next would be NULL. Thus, we
260  *    need a mechanism to prevent deletions from a given bucket.
261  *
262  *    To prevent deletions, we bump up the reference count on the
263  *    bucket. If the bucket is held, ire_delete just marks both
264  *    the ire and irb as CONDEMNED. When the
265  *    reference count on the bucket drops to zero, all the CONDEMNED ires
266  *    are deleted. We don't have to bump up the reference count on the
267  *    bucket if we are walking the bucket and never have to drop the bucket
268  *    lock. Note that irb_refhold does not prevent addition of new ires
269  *    in the list. It is okay because addition of new ires will not cause
270  *    ire_next to point to freed memory. We do irb_refhold only when
271  *    all of the 3 conditions are true :
272  *
273  *    1) The code needs to walk the IRE bucket from start to end.
274  *    2) It may have to drop the bucket lock sometimes while doing (1)
275  *    3) It does not want any ires to be deleted meanwhile.
276  */
277 
278 /*
279  * Bump up the reference count on the hash bucket - IRB to
280  * prevent ires from being deleted in this bucket.
281  */
282 void
irb_refhold(irb_t * irb)283 irb_refhold(irb_t *irb)
284 {
285 	rw_enter(&irb->irb_lock, RW_WRITER);
286 	irb->irb_refcnt++;
287 	ASSERT(irb->irb_refcnt != 0);
288 	rw_exit(&irb->irb_lock);
289 }
290 
291 void
irb_refhold_locked(irb_t * irb)292 irb_refhold_locked(irb_t *irb)
293 {
294 	ASSERT(RW_WRITE_HELD(&irb->irb_lock));
295 	irb->irb_refcnt++;
296 	ASSERT(irb->irb_refcnt != 0);
297 }
298 
299 /*
300  * Note: when IRB_MARK_DYNAMIC is not set the irb_t
301  * is statically allocated, so that when the irb_refcnt goes to 0,
302  * we simply clean up the ire list and continue.
303  */
304 void
irb_refrele(irb_t * irb)305 irb_refrele(irb_t *irb)
306 {
307 	if (irb->irb_marks & IRB_MARK_DYNAMIC) {
308 		irb_refrele_ftable(irb);
309 	} else {
310 		rw_enter(&irb->irb_lock, RW_WRITER);
311 		ASSERT(irb->irb_refcnt != 0);
312 		if (--irb->irb_refcnt	== 0 &&
313 		    (irb->irb_marks & IRB_MARK_CONDEMNED)) {
314 			ire_t *ire_list;
315 
316 			ire_list = ire_unlink(irb);
317 			rw_exit(&irb->irb_lock);
318 			ASSERT(ire_list != NULL);
319 			ire_cleanup(ire_list);
320 		} else {
321 			rw_exit(&irb->irb_lock);
322 		}
323 	}
324 }
325 
326 
327 /*
328  * Bump up the reference count on the IRE. We cannot assert that the
329  * bucket lock is being held as it is legal to bump up the reference
330  * count after the first lookup has returned the IRE without
331  * holding the lock.
332  */
333 void
ire_refhold(ire_t * ire)334 ire_refhold(ire_t *ire)
335 {
336 	atomic_inc_32(&(ire)->ire_refcnt);
337 	ASSERT((ire)->ire_refcnt != 0);
338 #ifdef DEBUG
339 	ire_trace_ref(ire);
340 #endif
341 }
342 
343 void
ire_refhold_notr(ire_t * ire)344 ire_refhold_notr(ire_t *ire)
345 {
346 	atomic_inc_32(&(ire)->ire_refcnt);
347 	ASSERT((ire)->ire_refcnt != 0);
348 }
349 
350 void
ire_refhold_locked(ire_t * ire)351 ire_refhold_locked(ire_t *ire)
352 {
353 #ifdef DEBUG
354 	ire_trace_ref(ire);
355 #endif
356 	ire->ire_refcnt++;
357 }
358 
359 /*
360  * Release a ref on an IRE.
361  *
362  * Must not be called while holding any locks. Otherwise if this is
363  * the last reference to be released there is a chance of recursive mutex
364  * panic due to ire_refrele -> ipif_ill_refrele_tail -> qwriter_ip trying
365  * to restart an ioctl. The one exception is when the caller is sure that
366  * this is not the last reference to be released. Eg. if the caller is
367  * sure that the ire has not been deleted and won't be deleted.
368  *
369  * In architectures e.g sun4u, where atomic_add_32_nv is just
370  * a cas, we need to maintain the right memory barrier semantics
371  * as that of mutex_exit i.e all the loads and stores should complete
372  * before the cas is executed. membar_exit() does that here.
373  */
374 void
ire_refrele(ire_t * ire)375 ire_refrele(ire_t *ire)
376 {
377 #ifdef DEBUG
378 	ire_untrace_ref(ire);
379 #endif
380 	ASSERT((ire)->ire_refcnt != 0);
381 	membar_exit();
382 	if (atomic_dec_32_nv(&(ire)->ire_refcnt) == 0)
383 		ire_inactive(ire);
384 }
385 
386 void
ire_refrele_notr(ire_t * ire)387 ire_refrele_notr(ire_t *ire)
388 {
389 	ASSERT((ire)->ire_refcnt != 0);
390 	membar_exit();
391 	if (atomic_dec_32_nv(&(ire)->ire_refcnt) == 0)
392 		ire_inactive(ire);
393 }
394 
395 /*
396  * This function is associated with the IP_IOC_IRE_DELETE[_NO_REPLY]
397  * IOCTL[s].  The NO_REPLY form is used by TCP to tell IP that it is
398  * having problems reaching a particular destination.
399  * This will make IP consider alternate routes (e.g., when there are
400  * muliple default routes), and it will also make IP discard any (potentially)
401  * stale redirect.
402  * Management processes may want to use the version that generates a reply.
403  *
404  * With the use of NUD like behavior for IPv4/ARP in addition to IPv6
405  * this function shouldn't be necessary for IP to recover from a bad redirect,
406  * a bad default router (when there are multiple default routers), or
407  * a stale ND/ARP entry. But we retain it in any case.
408  * For instance, this is helpful when TCP suspects a failure before NUD does.
409  */
410 int
ip_ire_delete(queue_t * q,mblk_t * mp,cred_t * ioc_cr)411 ip_ire_delete(queue_t *q, mblk_t *mp, cred_t *ioc_cr)
412 {
413 	uchar_t		*addr_ucp;
414 	uint_t		ipversion;
415 	sin_t		*sin;
416 	sin6_t		*sin6;
417 	ipaddr_t	v4addr;
418 	in6_addr_t	v6addr;
419 	ire_t		*ire;
420 	ipid_t		*ipid;
421 	zoneid_t	zoneid;
422 	ip_stack_t	*ipst;
423 
424 	ASSERT(q->q_next == NULL);
425 	zoneid = IPCL_ZONEID(Q_TO_CONN(q));
426 	ipst = CONNQ_TO_IPST(q);
427 
428 	/*
429 	 * Check privilege using the ioctl credential; if it is NULL
430 	 * then this is a kernel message and therefor privileged.
431 	 */
432 	if (ioc_cr != NULL && secpolicy_ip_config(ioc_cr, B_FALSE) != 0)
433 		return (EPERM);
434 
435 	ipid = (ipid_t *)mp->b_rptr;
436 
437 	addr_ucp = mi_offset_param(mp, ipid->ipid_addr_offset,
438 	    ipid->ipid_addr_length);
439 	if (addr_ucp == NULL || !OK_32PTR(addr_ucp))
440 		return (EINVAL);
441 	switch (ipid->ipid_addr_length) {
442 	case sizeof (sin_t):
443 		/*
444 		 * got complete (sockaddr) address - increment addr_ucp to point
445 		 * at the ip_addr field.
446 		 */
447 		sin = (sin_t *)addr_ucp;
448 		addr_ucp = (uchar_t *)&sin->sin_addr.s_addr;
449 		ipversion = IPV4_VERSION;
450 		break;
451 	case sizeof (sin6_t):
452 		/*
453 		 * got complete (sockaddr) address - increment addr_ucp to point
454 		 * at the ip_addr field.
455 		 */
456 		sin6 = (sin6_t *)addr_ucp;
457 		addr_ucp = (uchar_t *)&sin6->sin6_addr;
458 		ipversion = IPV6_VERSION;
459 		break;
460 	default:
461 		return (EINVAL);
462 	}
463 	if (ipversion == IPV4_VERSION) {
464 		/* Extract the destination address. */
465 		bcopy(addr_ucp, &v4addr, IP_ADDR_LEN);
466 
467 		ire = ire_ftable_lookup_v4(v4addr, 0, 0, 0, NULL,
468 		    zoneid, NULL, MATCH_IRE_DSTONLY, 0, ipst, NULL);
469 	} else {
470 		/* Extract the destination address. */
471 		bcopy(addr_ucp, &v6addr, IPV6_ADDR_LEN);
472 
473 		ire = ire_ftable_lookup_v6(&v6addr, NULL, NULL, 0, NULL,
474 		    zoneid, NULL, MATCH_IRE_DSTONLY, 0, ipst, NULL);
475 	}
476 	if (ire != NULL) {
477 		if (ipversion == IPV4_VERSION) {
478 			ip_rts_change(RTM_LOSING, ire->ire_addr,
479 			    ire->ire_gateway_addr, ire->ire_mask,
480 			    (Q_TO_CONN(q))->conn_laddr_v4,  0, 0, 0,
481 			    (RTA_DST | RTA_GATEWAY | RTA_NETMASK | RTA_IFA),
482 			    ire->ire_ipst);
483 		}
484 		(void) ire_no_good(ire);
485 		ire_refrele(ire);
486 	}
487 	return (0);
488 }
489 
490 /*
491  * Initialize the ire that is specific to IPv4 part and call
492  * ire_init_common to finish it.
493  * Returns zero or errno.
494  */
495 int
ire_init_v4(ire_t * ire,uchar_t * addr,uchar_t * mask,uchar_t * gateway,ushort_t type,ill_t * ill,zoneid_t zoneid,uint_t flags,tsol_gc_t * gc,ip_stack_t * ipst)496 ire_init_v4(ire_t *ire, uchar_t *addr, uchar_t *mask, uchar_t *gateway,
497     ushort_t type, ill_t *ill, zoneid_t zoneid, uint_t flags,
498     tsol_gc_t *gc, ip_stack_t *ipst)
499 {
500 	int error;
501 
502 	/*
503 	 * Reject IRE security attribute creation/initialization
504 	 * if system is not running in Trusted mode.
505 	 */
506 	if (gc != NULL && !is_system_labeled())
507 		return (EINVAL);
508 
509 	BUMP_IRE_STATS(ipst->ips_ire_stats_v4, ire_stats_alloced);
510 
511 	if (addr != NULL)
512 		bcopy(addr, &ire->ire_addr, IP_ADDR_LEN);
513 	if (gateway != NULL)
514 		bcopy(gateway, &ire->ire_gateway_addr, IP_ADDR_LEN);
515 
516 	/* Make sure we don't have stray values in some fields */
517 	switch (type) {
518 	case IRE_LOOPBACK:
519 	case IRE_HOST:
520 	case IRE_BROADCAST:
521 	case IRE_LOCAL:
522 	case IRE_IF_CLONE:
523 		ire->ire_mask = IP_HOST_MASK;
524 		ire->ire_masklen = IPV4_ABITS;
525 		break;
526 	case IRE_PREFIX:
527 	case IRE_DEFAULT:
528 	case IRE_IF_RESOLVER:
529 	case IRE_IF_NORESOLVER:
530 		if (mask != NULL) {
531 			bcopy(mask, &ire->ire_mask, IP_ADDR_LEN);
532 			ire->ire_masklen = ip_mask_to_plen(ire->ire_mask);
533 		}
534 		break;
535 	case IRE_MULTICAST:
536 	case IRE_NOROUTE:
537 		ASSERT(mask == NULL);
538 		break;
539 	default:
540 		ASSERT(0);
541 		return (EINVAL);
542 	}
543 
544 	error = ire_init_common(ire, type, ill, zoneid, flags, IPV4_VERSION,
545 	    gc, ipst);
546 	if (error != 0)
547 		return (error);
548 
549 	/* Determine which function pointers to use */
550 	ire->ire_postfragfn = ip_xmit;		/* Common case */
551 
552 	switch (ire->ire_type) {
553 	case IRE_LOCAL:
554 		ire->ire_sendfn = ire_send_local_v4;
555 		ire->ire_recvfn = ire_recv_local_v4;
556 		ASSERT(ire->ire_ill != NULL);
557 		if (ire->ire_ill->ill_flags & ILLF_NOACCEPT)
558 			ire->ire_recvfn = ire_recv_noaccept_v6;
559 		break;
560 	case IRE_LOOPBACK:
561 		ire->ire_sendfn = ire_send_local_v4;
562 		ire->ire_recvfn = ire_recv_loopback_v4;
563 		break;
564 	case IRE_BROADCAST:
565 		ire->ire_postfragfn = ip_postfrag_loopcheck;
566 		ire->ire_sendfn = ire_send_broadcast_v4;
567 		ire->ire_recvfn = ire_recv_broadcast_v4;
568 		break;
569 	case IRE_MULTICAST:
570 		ire->ire_postfragfn = ip_postfrag_loopcheck;
571 		ire->ire_sendfn = ire_send_multicast_v4;
572 		ire->ire_recvfn = ire_recv_multicast_v4;
573 		break;
574 	default:
575 		/*
576 		 * For IRE_IF_ALL and IRE_OFFLINK we forward received
577 		 * packets by default.
578 		 */
579 		ire->ire_sendfn = ire_send_wire_v4;
580 		ire->ire_recvfn = ire_recv_forward_v4;
581 		break;
582 	}
583 	if (ire->ire_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
584 		ire->ire_sendfn = ire_send_noroute_v4;
585 		ire->ire_recvfn = ire_recv_noroute_v4;
586 	} else if (ire->ire_flags & RTF_MULTIRT) {
587 		ire->ire_postfragfn = ip_postfrag_multirt_v4;
588 		ire->ire_sendfn = ire_send_multirt_v4;
589 		/* Multirt receive of broadcast uses ire_recv_broadcast_v4 */
590 		if (ire->ire_type != IRE_BROADCAST)
591 			ire->ire_recvfn = ire_recv_multirt_v4;
592 	}
593 	ire->ire_nce_capable = ire_determine_nce_capable(ire);
594 	return (0);
595 }
596 
597 /*
598  * Determine ire_nce_capable
599  */
600 boolean_t
ire_determine_nce_capable(ire_t * ire)601 ire_determine_nce_capable(ire_t *ire)
602 {
603 	int max_masklen;
604 
605 	if ((ire->ire_flags & (RTF_REJECT|RTF_BLACKHOLE)) ||
606 	    (ire->ire_type & IRE_MULTICAST))
607 		return (B_TRUE);
608 
609 	if (ire->ire_ipversion == IPV4_VERSION)
610 		max_masklen = IPV4_ABITS;
611 	else
612 		max_masklen = IPV6_ABITS;
613 
614 	if ((ire->ire_type & IRE_ONLINK) && ire->ire_masklen == max_masklen)
615 		return (B_TRUE);
616 	return (B_FALSE);
617 }
618 
619 /*
620  * ire_create is called to allocate and initialize a new IRE.
621  *
622  * NOTE : This is called as writer sometimes though not required
623  * by this function.
624  */
625 ire_t *
ire_create(uchar_t * addr,uchar_t * mask,uchar_t * gateway,ushort_t type,ill_t * ill,zoneid_t zoneid,uint_t flags,tsol_gc_t * gc,ip_stack_t * ipst)626 ire_create(uchar_t *addr, uchar_t *mask, uchar_t *gateway,
627     ushort_t type, ill_t *ill, zoneid_t zoneid, uint_t flags, tsol_gc_t *gc,
628     ip_stack_t *ipst)
629 {
630 	ire_t	*ire;
631 	int	error;
632 
633 	ire = kmem_cache_alloc(ire_cache, KM_NOSLEEP);
634 	if (ire == NULL) {
635 		DTRACE_PROBE(kmem__cache__alloc);
636 		return (NULL);
637 	}
638 	*ire = ire_null;
639 
640 	error = ire_init_v4(ire, addr, mask, gateway, type, ill, zoneid, flags,
641 	    gc, ipst);
642 	if (error != 0) {
643 		DTRACE_PROBE2(ire__init, ire_t *, ire, int, error);
644 		kmem_cache_free(ire_cache, ire);
645 		return (NULL);
646 	}
647 	return (ire);
648 }
649 
650 /*
651  * Common to IPv4 and IPv6
652  * Returns zero or errno.
653  */
654 int
ire_init_common(ire_t * ire,ushort_t type,ill_t * ill,zoneid_t zoneid,uint_t flags,uchar_t ipversion,tsol_gc_t * gc,ip_stack_t * ipst)655 ire_init_common(ire_t *ire, ushort_t type, ill_t *ill, zoneid_t zoneid,
656     uint_t flags, uchar_t ipversion, tsol_gc_t *gc, ip_stack_t *ipst)
657 {
658 	int error;
659 
660 #ifdef DEBUG
661 	if (ill != NULL) {
662 		if (ill->ill_isv6)
663 			ASSERT(ipversion == IPV6_VERSION);
664 		else
665 			ASSERT(ipversion == IPV4_VERSION);
666 	}
667 #endif /* DEBUG */
668 
669 	/*
670 	 * Create/initialize IRE security attribute only in Trusted mode;
671 	 * if the passed in gc is non-NULL, we expect that the caller
672 	 * has held a reference to it and will release it when this routine
673 	 * returns a failure, otherwise we own the reference.  We do this
674 	 * prior to initializing the rest IRE fields.
675 	 */
676 	if (is_system_labeled()) {
677 		if ((type & (IRE_LOCAL | IRE_LOOPBACK | IRE_BROADCAST |
678 		    IRE_IF_ALL | IRE_MULTICAST | IRE_NOROUTE)) != 0) {
679 			/* release references on behalf of caller */
680 			if (gc != NULL)
681 				GC_REFRELE(gc);
682 		} else {
683 			error = tsol_ire_init_gwattr(ire, ipversion, gc);
684 			if (error != 0)
685 				return (error);
686 		}
687 	}
688 
689 	ire->ire_type = type;
690 	ire->ire_flags = RTF_UP | flags;
691 	ire->ire_create_time = (uint32_t)gethrestime_sec();
692 	ire->ire_generation = IRE_GENERATION_INITIAL;
693 
694 	/*
695 	 * The ill_ire_cnt isn't increased until
696 	 * the IRE is added to ensure that a walker will find
697 	 * all IREs that hold a reference on an ill.
698 	 *
699 	 * Note that ill_ire_multicast doesn't hold a ref on the ill since
700 	 * ire_add() is not called for the IRE_MULTICAST.
701 	 */
702 	ire->ire_ill = ill;
703 	ire->ire_zoneid = zoneid;
704 	ire->ire_ipversion = ipversion;
705 
706 	mutex_init(&ire->ire_lock, NULL, MUTEX_DEFAULT, NULL);
707 	ire->ire_refcnt = 1;
708 	ire->ire_identical_ref = 1;	/* Number of ire_delete's needed */
709 	ire->ire_ipst = ipst;	/* No netstack_hold */
710 	ire->ire_trace_disable = B_FALSE;
711 
712 	return (0);
713 }
714 
715 /*
716  * This creates an IRE_BROADCAST based on the arguments.
717  * A mirror is ire_lookup_bcast().
718  *
719  * Any supression of unneeded ones is done in ire_add_v4.
720  * We add one IRE_BROADCAST per address. ire_send_broadcast_v4()
721  * takes care of generating a loopback copy of the packet.
722  */
723 ire_t **
ire_create_bcast(ill_t * ill,ipaddr_t addr,zoneid_t zoneid,ire_t ** irep)724 ire_create_bcast(ill_t *ill, ipaddr_t addr, zoneid_t zoneid, ire_t **irep)
725 {
726 	ip_stack_t	*ipst = ill->ill_ipst;
727 
728 	ASSERT(IAM_WRITER_ILL(ill));
729 
730 	*irep++ = ire_create(
731 	    (uchar_t *)&addr,			/* dest addr */
732 	    (uchar_t *)&ip_g_all_ones,		/* mask */
733 	    NULL,				/* no gateway */
734 	    IRE_BROADCAST,
735 	    ill,
736 	    zoneid,
737 	    RTF_KERNEL,
738 	    NULL,
739 	    ipst);
740 
741 	return (irep);
742 }
743 
744 /*
745  * This looks up an IRE_BROADCAST based on the arguments.
746  * Mirrors ire_create_bcast().
747  */
748 ire_t *
ire_lookup_bcast(ill_t * ill,ipaddr_t addr,zoneid_t zoneid)749 ire_lookup_bcast(ill_t *ill, ipaddr_t addr, zoneid_t zoneid)
750 {
751 	ire_t		*ire;
752 	int		match_args;
753 
754 	match_args = MATCH_IRE_TYPE | MATCH_IRE_ILL | MATCH_IRE_GW |
755 	    MATCH_IRE_MASK | MATCH_IRE_ZONEONLY;
756 
757 	if (IS_UNDER_IPMP(ill))
758 		match_args |= MATCH_IRE_TESTHIDDEN;
759 
760 	ire = ire_ftable_lookup_v4(
761 	    addr,				/* dest addr */
762 	    ip_g_all_ones,			/* mask */
763 	    0,					/* no gateway */
764 	    IRE_BROADCAST,
765 	    ill,
766 	    zoneid,
767 	    NULL,
768 	    match_args,
769 	    0,
770 	    ill->ill_ipst,
771 	    NULL);
772 	return (ire);
773 }
774 
775 /* Arrange to call the specified function for every IRE in the world. */
776 void
ire_walk(pfv_t func,void * arg,ip_stack_t * ipst)777 ire_walk(pfv_t func, void *arg, ip_stack_t *ipst)
778 {
779 	ire_walk_ipvers(func, arg, 0, ALL_ZONES, ipst);
780 }
781 
782 void
ire_walk_v4(pfv_t func,void * arg,zoneid_t zoneid,ip_stack_t * ipst)783 ire_walk_v4(pfv_t func, void *arg, zoneid_t zoneid, ip_stack_t *ipst)
784 {
785 	ire_walk_ipvers(func, arg, IPV4_VERSION, zoneid, ipst);
786 }
787 
788 void
ire_walk_v6(pfv_t func,void * arg,zoneid_t zoneid,ip_stack_t * ipst)789 ire_walk_v6(pfv_t func, void *arg, zoneid_t zoneid, ip_stack_t *ipst)
790 {
791 	ire_walk_ipvers(func, arg, IPV6_VERSION, zoneid, ipst);
792 }
793 
794 /*
795  * Walk a particular version. version == 0 means both v4 and v6.
796  */
797 static void
ire_walk_ipvers(pfv_t func,void * arg,uchar_t vers,zoneid_t zoneid,ip_stack_t * ipst)798 ire_walk_ipvers(pfv_t func, void *arg, uchar_t vers, zoneid_t zoneid,
799     ip_stack_t *ipst)
800 {
801 	if (vers != IPV6_VERSION) {
802 		/*
803 		 * ip_forwarding_table variable doesn't matter for IPv4 since
804 		 * ire_walk_ill_tables uses ips_ip_ftable for IPv4.
805 		 */
806 		ire_walk_ill_tables(0, 0, func, arg, IP_MASK_TABLE_SIZE,
807 		    0, NULL,
808 		    NULL, zoneid, ipst);
809 	}
810 	if (vers != IPV4_VERSION) {
811 		ire_walk_ill_tables(0, 0, func, arg, IP6_MASK_TABLE_SIZE,
812 		    ipst->ips_ip6_ftable_hash_size,
813 		    ipst->ips_ip_forwarding_table_v6,
814 		    NULL, zoneid, ipst);
815 	}
816 }
817 
818 /*
819  * Arrange to call the specified function for every IRE that matches the ill.
820  */
821 void
ire_walk_ill(uint_t match_flags,uint_t ire_type,pfv_t func,void * arg,ill_t * ill)822 ire_walk_ill(uint_t match_flags, uint_t ire_type, pfv_t func, void *arg,
823     ill_t *ill)
824 {
825 	uchar_t vers = (ill->ill_isv6 ? IPV6_VERSION : IPV4_VERSION);
826 
827 	ire_walk_ill_ipvers(match_flags, ire_type, func, arg, vers, ill);
828 }
829 
830 /*
831  * Walk a particular ill and version.
832  */
833 static void
ire_walk_ill_ipvers(uint_t match_flags,uint_t ire_type,pfv_t func,void * arg,uchar_t vers,ill_t * ill)834 ire_walk_ill_ipvers(uint_t match_flags, uint_t ire_type, pfv_t func,
835     void *arg, uchar_t vers, ill_t *ill)
836 {
837 	ip_stack_t	*ipst = ill->ill_ipst;
838 
839 	if (vers == IPV4_VERSION) {
840 		ire_walk_ill_tables(match_flags, ire_type, func, arg,
841 		    IP_MASK_TABLE_SIZE,
842 		    0, NULL,
843 		    ill, ALL_ZONES, ipst);
844 	}
845 	if (vers != IPV4_VERSION) {
846 		ire_walk_ill_tables(match_flags, ire_type, func, arg,
847 		    IP6_MASK_TABLE_SIZE, ipst->ips_ip6_ftable_hash_size,
848 		    ipst->ips_ip_forwarding_table_v6,
849 		    ill, ALL_ZONES, ipst);
850 	}
851 }
852 
853 /*
854  * Do the specific matching of IREs to shared-IP zones.
855  *
856  * We have the same logic as in ire_match_args but implemented slightly
857  * differently.
858  */
859 boolean_t
ire_walk_ill_match(uint_t match_flags,uint_t ire_type,ire_t * ire,ill_t * ill,zoneid_t zoneid,ip_stack_t * ipst)860 ire_walk_ill_match(uint_t match_flags, uint_t ire_type, ire_t *ire,
861     ill_t *ill, zoneid_t zoneid, ip_stack_t *ipst)
862 {
863 	ill_t *dst_ill = ire->ire_ill;
864 
865 	ASSERT(match_flags != 0 || zoneid != ALL_ZONES);
866 
867 	if (zoneid != ALL_ZONES && zoneid != ire->ire_zoneid &&
868 	    ire->ire_zoneid != ALL_ZONES) {
869 		/*
870 		 * We're walking the IREs for a specific zone. The only relevant
871 		 * IREs are:
872 		 * - all IREs with a matching ire_zoneid
873 		 * - IRE_IF_ALL IREs for interfaces with a usable source addr
874 		 *   with a matching zone
875 		 * - IRE_OFFLINK with a gateway reachable from the zone
876 		 * Note that ealier we only did the IRE_OFFLINK check for
877 		 * IRE_DEFAULT (and only when we had multiple IRE_DEFAULTs).
878 		 */
879 		if (ire->ire_type & IRE_ONLINK) {
880 			uint_t	ifindex;
881 
882 			/*
883 			 * Note there is no IRE_INTERFACE on vniN thus
884 			 * can't do an IRE lookup for a matching route.
885 			 */
886 			ifindex = dst_ill->ill_usesrc_ifindex;
887 			if (ifindex == 0)
888 				return (B_FALSE);
889 
890 			/*
891 			 * If there is a usable source address in the
892 			 * zone, then it's ok to return an
893 			 * IRE_INTERFACE
894 			 */
895 			if (!ipif_zone_avail(ifindex, dst_ill->ill_isv6,
896 			    zoneid, ipst)) {
897 				return (B_FALSE);
898 			}
899 		}
900 		if (dst_ill != NULL && (ire->ire_type & IRE_OFFLINK)) {
901 			ipif_t	*tipif;
902 
903 			mutex_enter(&dst_ill->ill_lock);
904 			for (tipif = dst_ill->ill_ipif;
905 			    tipif != NULL; tipif = tipif->ipif_next) {
906 				if (!IPIF_IS_CONDEMNED(tipif) &&
907 				    (tipif->ipif_flags & IPIF_UP) &&
908 				    (tipif->ipif_zoneid == zoneid ||
909 				    tipif->ipif_zoneid == ALL_ZONES))
910 					break;
911 			}
912 			mutex_exit(&dst_ill->ill_lock);
913 			if (tipif == NULL) {
914 				return (B_FALSE);
915 			}
916 		}
917 	}
918 	/*
919 	 * Except for ALL_ZONES, we only match the offlink routes
920 	 * where ire_gateway_addr has an IRE_INTERFACE for the zoneid.
921 	 * Since we can have leftover routes after the IP addresses have
922 	 * changed, the global zone will also match offlink routes where the
923 	 * gateway is unreachable from any zone.
924 	 */
925 	if ((ire->ire_type & IRE_OFFLINK) && zoneid != ALL_ZONES) {
926 		in6_addr_t gw_addr_v6;
927 		boolean_t reach;
928 
929 		if (ire->ire_ipversion == IPV4_VERSION) {
930 			reach = ire_gateway_ok_zone_v4(ire->ire_gateway_addr,
931 			    zoneid, dst_ill, NULL, ipst, B_FALSE);
932 		} else {
933 			ASSERT(ire->ire_ipversion == IPV6_VERSION);
934 			mutex_enter(&ire->ire_lock);
935 			gw_addr_v6 = ire->ire_gateway_addr_v6;
936 			mutex_exit(&ire->ire_lock);
937 
938 			reach = ire_gateway_ok_zone_v6(&gw_addr_v6, zoneid,
939 			    dst_ill, NULL, ipst, B_FALSE);
940 		}
941 		if (!reach) {
942 			if (zoneid != GLOBAL_ZONEID)
943 				return (B_FALSE);
944 
945 			/*
946 			 * Check if ALL_ZONES reachable - if not then let the
947 			 * global zone see it.
948 			 */
949 			if (ire->ire_ipversion == IPV4_VERSION) {
950 				reach = ire_gateway_ok_zone_v4(
951 				    ire->ire_gateway_addr, ALL_ZONES,
952 				    dst_ill, NULL, ipst, B_FALSE);
953 			} else {
954 				reach = ire_gateway_ok_zone_v6(&gw_addr_v6,
955 				    ALL_ZONES, dst_ill, NULL, ipst, B_FALSE);
956 			}
957 			if (reach) {
958 				/*
959 				 * Some other zone could see it, hence hide it
960 				 * in the global zone.
961 				 */
962 				return (B_FALSE);
963 			}
964 		}
965 	}
966 
967 	if (((!(match_flags & MATCH_IRE_TYPE)) ||
968 	    (ire->ire_type & ire_type)) &&
969 	    ((!(match_flags & MATCH_IRE_ILL)) ||
970 	    (dst_ill == ill ||
971 	    dst_ill != NULL && IS_IN_SAME_ILLGRP(dst_ill, ill)))) {
972 		return (B_TRUE);
973 	}
974 	return (B_FALSE);
975 }
976 
977 int
rtfunc(struct radix_node * rn,void * arg)978 rtfunc(struct radix_node *rn, void *arg)
979 {
980 	struct rtfuncarg *rtf = arg;
981 	struct rt_entry *rt;
982 	irb_t *irb;
983 	ire_t *ire;
984 	boolean_t ret;
985 
986 	rt = (struct rt_entry *)rn;
987 	ASSERT(rt != NULL);
988 	irb = &rt->rt_irb;
989 	for (ire = irb->irb_ire; ire != NULL; ire = ire->ire_next) {
990 		if ((rtf->rt_match_flags != 0) ||
991 		    (rtf->rt_zoneid != ALL_ZONES)) {
992 			ret = ire_walk_ill_match(rtf->rt_match_flags,
993 			    rtf->rt_ire_type, ire,
994 			    rtf->rt_ill, rtf->rt_zoneid, rtf->rt_ipst);
995 		} else {
996 			ret = B_TRUE;
997 		}
998 		if (ret)
999 			(*rtf->rt_func)(ire, rtf->rt_arg);
1000 	}
1001 	return (0);
1002 }
1003 
1004 /*
1005  * Walk the ftable entries that match the ill.
1006  */
1007 void
ire_walk_ill_tables(uint_t match_flags,uint_t ire_type,pfv_t func,void * arg,size_t ftbl_sz,size_t htbl_sz,irb_t ** ipftbl,ill_t * ill,zoneid_t zoneid,ip_stack_t * ipst)1008 ire_walk_ill_tables(uint_t match_flags, uint_t ire_type, pfv_t func,
1009     void *arg, size_t ftbl_sz, size_t htbl_sz, irb_t **ipftbl,
1010     ill_t *ill, zoneid_t zoneid,
1011     ip_stack_t *ipst)
1012 {
1013 	irb_t	*irb_ptr;
1014 	irb_t	*irb;
1015 	ire_t	*ire;
1016 	int i, j;
1017 	boolean_t ret;
1018 	struct rtfuncarg rtfarg;
1019 
1020 	ASSERT((!(match_flags & MATCH_IRE_ILL)) || (ill != NULL));
1021 	ASSERT(!(match_flags & MATCH_IRE_TYPE) || (ire_type != 0));
1022 
1023 	/* knobs such that routine is called only for v6 case */
1024 	if (ipftbl == ipst->ips_ip_forwarding_table_v6) {
1025 		for (i = (ftbl_sz - 1);  i >= 0; i--) {
1026 			if ((irb_ptr = ipftbl[i]) == NULL)
1027 				continue;
1028 			for (j = 0; j < htbl_sz; j++) {
1029 				irb = &irb_ptr[j];
1030 				if (irb->irb_ire == NULL)
1031 					continue;
1032 
1033 				irb_refhold(irb);
1034 				for (ire = irb->irb_ire; ire != NULL;
1035 				    ire = ire->ire_next) {
1036 					if (match_flags == 0 &&
1037 					    zoneid == ALL_ZONES) {
1038 						ret = B_TRUE;
1039 					} else {
1040 						ret =
1041 						    ire_walk_ill_match(
1042 						    match_flags,
1043 						    ire_type, ire, ill,
1044 						    zoneid, ipst);
1045 					}
1046 					if (ret)
1047 						(*func)(ire, arg);
1048 				}
1049 				irb_refrele(irb);
1050 			}
1051 		}
1052 	} else {
1053 		bzero(&rtfarg, sizeof (rtfarg));
1054 		rtfarg.rt_func = func;
1055 		rtfarg.rt_arg = arg;
1056 		if (match_flags != 0) {
1057 			rtfarg.rt_match_flags = match_flags;
1058 		}
1059 		rtfarg.rt_ire_type = ire_type;
1060 		rtfarg.rt_ill = ill;
1061 		rtfarg.rt_zoneid = zoneid;
1062 		rtfarg.rt_ipst = ipst;	/* No netstack_hold */
1063 		(void) ipst->ips_ip_ftable->rnh_walktree_mt(
1064 		    ipst->ips_ip_ftable,
1065 		    rtfunc, &rtfarg, irb_refhold_rn, irb_refrele_rn);
1066 	}
1067 }
1068 
1069 /*
1070  * This function takes a mask and returns
1071  * number of bits set in the mask. If no
1072  * bit is set it returns 0.
1073  * Assumes a contiguous mask.
1074  */
1075 int
ip_mask_to_plen(ipaddr_t mask)1076 ip_mask_to_plen(ipaddr_t mask)
1077 {
1078 	return (mask == 0 ? 0 : IP_ABITS - (ffs(ntohl(mask)) -1));
1079 }
1080 
1081 /*
1082  * Convert length for a mask to the mask.
1083  */
1084 ipaddr_t
ip_plen_to_mask(uint_t masklen)1085 ip_plen_to_mask(uint_t masklen)
1086 {
1087 	if (masklen == 0)
1088 		return (0);
1089 
1090 	return (htonl(IP_HOST_MASK << (IP_ABITS - masklen)));
1091 }
1092 
1093 void
ire_atomic_end(irb_t * irb_ptr,ire_t * ire)1094 ire_atomic_end(irb_t *irb_ptr, ire_t *ire)
1095 {
1096 	ill_t		*ill;
1097 
1098 	ill = ire->ire_ill;
1099 	if (ill != NULL)
1100 		mutex_exit(&ill->ill_lock);
1101 	rw_exit(&irb_ptr->irb_lock);
1102 }
1103 
1104 /*
1105  * ire_add_v[46] atomically make sure that the ill associated
1106  * with the new ire is not going away i.e., we check ILL_CONDEMNED.
1107  */
1108 int
ire_atomic_start(irb_t * irb_ptr,ire_t * ire)1109 ire_atomic_start(irb_t *irb_ptr, ire_t *ire)
1110 {
1111 	ill_t		*ill;
1112 
1113 	ill = ire->ire_ill;
1114 
1115 	rw_enter(&irb_ptr->irb_lock, RW_WRITER);
1116 	if (ill != NULL) {
1117 		mutex_enter(&ill->ill_lock);
1118 
1119 		/*
1120 		 * Don't allow IRE's to be created on dying ills, or on
1121 		 * ill's for which the last ipif is going down, or ones which
1122 		 * don't have even a single UP interface
1123 		 */
1124 		if ((ill->ill_state_flags &
1125 		    (ILL_CONDEMNED|ILL_DOWN_IN_PROGRESS)) != 0) {
1126 			ire_atomic_end(irb_ptr, ire);
1127 			DTRACE_PROBE1(ire__add__on__dying__ill, ire_t *, ire);
1128 			return (ENXIO);
1129 		}
1130 
1131 		if (IS_UNDER_IPMP(ill)) {
1132 			int	error = 0;
1133 			mutex_enter(&ill->ill_phyint->phyint_lock);
1134 			if (!ipmp_ill_is_active(ill) &&
1135 			    IRE_HIDDEN_TYPE(ire->ire_type) &&
1136 			    !ire->ire_testhidden) {
1137 				error = EINVAL;
1138 			}
1139 			mutex_exit(&ill->ill_phyint->phyint_lock);
1140 			if (error != 0) {
1141 				ire_atomic_end(irb_ptr, ire);
1142 				return (error);
1143 			}
1144 		}
1145 
1146 	}
1147 	return (0);
1148 }
1149 
1150 /*
1151  * Add a fully initialized IRE to the forwarding table.
1152  * This returns NULL on failure, or a held IRE on success.
1153  * Normally the returned IRE is the same as the argument. But a different
1154  * IRE will be returned if the added IRE is deemed identical to an existing
1155  * one. In that case ire_identical_ref will be increased.
1156  * The caller always needs to do an ire_refrele() on the returned IRE.
1157  */
1158 ire_t *
ire_add(ire_t * ire)1159 ire_add(ire_t *ire)
1160 {
1161 	if (IRE_HIDDEN_TYPE(ire->ire_type) &&
1162 	    ire->ire_ill != NULL && IS_UNDER_IPMP(ire->ire_ill)) {
1163 		/*
1164 		 * IREs hosted on interfaces that are under IPMP
1165 		 * should be hidden so that applications don't
1166 		 * accidentally end up sending packets with test
1167 		 * addresses as their source addresses, or
1168 		 * sending out interfaces that are e.g. IFF_INACTIVE.
1169 		 * Hide them here.
1170 		 */
1171 		ire->ire_testhidden = B_TRUE;
1172 	}
1173 
1174 	if (ire->ire_ipversion == IPV6_VERSION)
1175 		return (ire_add_v6(ire));
1176 	else
1177 		return (ire_add_v4(ire));
1178 }
1179 
1180 /*
1181  * Add a fully initialized IPv4 IRE to the forwarding table.
1182  * This returns NULL on failure, or a held IRE on success.
1183  * Normally the returned IRE is the same as the argument. But a different
1184  * IRE will be returned if the added IRE is deemed identical to an existing
1185  * one. In that case ire_identical_ref will be increased.
1186  * The caller always needs to do an ire_refrele() on the returned IRE.
1187  */
1188 static ire_t *
ire_add_v4(ire_t * ire)1189 ire_add_v4(ire_t *ire)
1190 {
1191 	ire_t	*ire1;
1192 	irb_t	*irb_ptr;
1193 	ire_t	**irep;
1194 	int	match_flags;
1195 	int	error;
1196 	ip_stack_t	*ipst = ire->ire_ipst;
1197 
1198 	if (ire->ire_ill != NULL)
1199 		ASSERT(!MUTEX_HELD(&ire->ire_ill->ill_lock));
1200 	ASSERT(ire->ire_ipversion == IPV4_VERSION);
1201 
1202 	/* Make sure the address is properly masked. */
1203 	ire->ire_addr &= ire->ire_mask;
1204 
1205 	match_flags = (MATCH_IRE_MASK | MATCH_IRE_TYPE | MATCH_IRE_GW);
1206 
1207 	if (ire->ire_ill != NULL) {
1208 		match_flags |= MATCH_IRE_ILL;
1209 	}
1210 	irb_ptr = ire_get_bucket(ire);
1211 	if (irb_ptr == NULL) {
1212 		printf("no bucket for %p\n", (void *)ire);
1213 		ire_delete(ire);
1214 		return (NULL);
1215 	}
1216 
1217 	/*
1218 	 * Start the atomic add of the ire. Grab the ill lock,
1219 	 * the bucket lock. Check for condemned.
1220 	 */
1221 	error = ire_atomic_start(irb_ptr, ire);
1222 	if (error != 0) {
1223 		printf("no ire_atomic_start for %p\n", (void *)ire);
1224 		ire_delete(ire);
1225 		irb_refrele(irb_ptr);
1226 		return (NULL);
1227 	}
1228 	/*
1229 	 * If we are creating a hidden IRE, make sure we search for
1230 	 * hidden IREs when searching for duplicates below.
1231 	 * Otherwise, we might find an IRE on some other interface
1232 	 * that's not marked hidden.
1233 	 */
1234 	if (ire->ire_testhidden)
1235 		match_flags |= MATCH_IRE_TESTHIDDEN;
1236 
1237 	/*
1238 	 * Atomically check for duplicate and insert in the table.
1239 	 */
1240 	for (ire1 = irb_ptr->irb_ire; ire1 != NULL; ire1 = ire1->ire_next) {
1241 		if (IRE_IS_CONDEMNED(ire1))
1242 			continue;
1243 		/*
1244 		 * Here we need an exact match on zoneid, i.e.,
1245 		 * ire_match_args doesn't fit.
1246 		 */
1247 		if (ire1->ire_zoneid != ire->ire_zoneid)
1248 			continue;
1249 
1250 		if (ire1->ire_type != ire->ire_type)
1251 			continue;
1252 
1253 		/*
1254 		 * Note: We do not allow multiple routes that differ only
1255 		 * in the gateway security attributes; such routes are
1256 		 * considered duplicates.
1257 		 * To change that we explicitly have to treat them as
1258 		 * different here.
1259 		 */
1260 		if (ire_match_args(ire1, ire->ire_addr, ire->ire_mask,
1261 		    ire->ire_gateway_addr, ire->ire_type, ire->ire_ill,
1262 		    ire->ire_zoneid, NULL, match_flags)) {
1263 			/*
1264 			 * Return the old ire after doing a REFHOLD.
1265 			 * As most of the callers continue to use the IRE
1266 			 * after adding, we return a held ire. This will
1267 			 * avoid a lookup in the caller again. If the callers
1268 			 * don't want to use it, they need to do a REFRELE.
1269 			 *
1270 			 * We only allow exactly one IRE_IF_CLONE for any dst,
1271 			 * so, if the is an IF_CLONE, return the ire without
1272 			 * an identical_ref, but with an ire_ref held.
1273 			 */
1274 			if (ire->ire_type != IRE_IF_CLONE) {
1275 				atomic_inc_32(&ire1->ire_identical_ref);
1276 				DTRACE_PROBE2(ire__add__exist, ire_t *, ire1,
1277 				    ire_t *, ire);
1278 			}
1279 			ire_refhold(ire1);
1280 			ire_atomic_end(irb_ptr, ire);
1281 			ire_delete(ire);
1282 			irb_refrele(irb_ptr);
1283 			return (ire1);
1284 		}
1285 	}
1286 
1287 	/*
1288 	 * Normally we do head insertion since most things do not care about
1289 	 * the order of the IREs in the bucket. Note that ip_cgtp_bcast_add
1290 	 * assumes we at least do head insertion so that its IRE_BROADCAST
1291 	 * arrive ahead of existing IRE_HOST for the same address.
1292 	 * However, due to shared-IP zones (and restrict_interzone_loopback)
1293 	 * we can have an IRE_LOCAL as well as IRE_IF_CLONE for the same
1294 	 * address. For that reason we do tail insertion for IRE_IF_CLONE.
1295 	 * Due to the IRE_BROADCAST on cgtp0, which must be last in the bucket,
1296 	 * we do tail insertion of IRE_BROADCASTs that do not have RTF_MULTIRT
1297 	 * set.
1298 	 */
1299 	irep = (ire_t **)irb_ptr;
1300 	if ((ire->ire_type & IRE_IF_CLONE) ||
1301 	    ((ire->ire_type & IRE_BROADCAST) &&
1302 	    !(ire->ire_flags & RTF_MULTIRT))) {
1303 		while ((ire1 = *irep) != NULL)
1304 			irep = &ire1->ire_next;
1305 	}
1306 	/* Insert at *irep */
1307 	ire1 = *irep;
1308 	if (ire1 != NULL)
1309 		ire1->ire_ptpn = &ire->ire_next;
1310 	ire->ire_next = ire1;
1311 	/* Link the new one in. */
1312 	ire->ire_ptpn = irep;
1313 
1314 	/*
1315 	 * ire_walk routines de-reference ire_next without holding
1316 	 * a lock. Before we point to the new ire, we want to make
1317 	 * sure the store that sets the ire_next of the new ire
1318 	 * reaches global visibility, so that ire_walk routines
1319 	 * don't see a truncated list of ires i.e if the ire_next
1320 	 * of the new ire gets set after we do "*irep = ire" due
1321 	 * to re-ordering, the ire_walk thread will see a NULL
1322 	 * once it accesses the ire_next of the new ire.
1323 	 * membar_producer() makes sure that the following store
1324 	 * happens *after* all of the above stores.
1325 	 */
1326 	membar_producer();
1327 	*irep = ire;
1328 	ire->ire_bucket = irb_ptr;
1329 	/*
1330 	 * We return a bumped up IRE above. Keep it symmetrical
1331 	 * so that the callers will always have to release. This
1332 	 * helps the callers of this function because they continue
1333 	 * to use the IRE after adding and hence they don't have to
1334 	 * lookup again after we return the IRE.
1335 	 *
1336 	 * NOTE : We don't have to use atomics as this is appearing
1337 	 * in the list for the first time and no one else can bump
1338 	 * up the reference count on this yet.
1339 	 */
1340 	ire_refhold_locked(ire);
1341 	BUMP_IRE_STATS(ipst->ips_ire_stats_v4, ire_stats_inserted);
1342 
1343 	irb_ptr->irb_ire_cnt++;
1344 	if (irb_ptr->irb_marks & IRB_MARK_DYNAMIC)
1345 		irb_ptr->irb_nire++;
1346 
1347 	if (ire->ire_ill != NULL) {
1348 		ire->ire_ill->ill_ire_cnt++;
1349 		ASSERT(ire->ire_ill->ill_ire_cnt != 0);	/* Wraparound */
1350 	}
1351 
1352 	ire_atomic_end(irb_ptr, ire);
1353 
1354 	/* Make any caching of the IREs be notified or updated */
1355 	ire_flush_cache_v4(ire, IRE_FLUSH_ADD);
1356 
1357 	if (ire->ire_ill != NULL)
1358 		ASSERT(!MUTEX_HELD(&ire->ire_ill->ill_lock));
1359 	irb_refrele(irb_ptr);
1360 	return (ire);
1361 }
1362 
1363 /*
1364  * irb_refrele is the only caller of the function. ire_unlink calls to
1365  * do the final cleanup for this ire.
1366  */
1367 void
ire_cleanup(ire_t * ire)1368 ire_cleanup(ire_t *ire)
1369 {
1370 	ire_t *ire_next;
1371 	ip_stack_t *ipst = ire->ire_ipst;
1372 
1373 	ASSERT(ire != NULL);
1374 
1375 	while (ire != NULL) {
1376 		ire_next = ire->ire_next;
1377 		if (ire->ire_ipversion == IPV4_VERSION) {
1378 			ire_delete_v4(ire);
1379 			BUMP_IRE_STATS(ipst->ips_ire_stats_v4,
1380 			    ire_stats_deleted);
1381 		} else {
1382 			ASSERT(ire->ire_ipversion == IPV6_VERSION);
1383 			ire_delete_v6(ire);
1384 			BUMP_IRE_STATS(ipst->ips_ire_stats_v6,
1385 			    ire_stats_deleted);
1386 		}
1387 		/*
1388 		 * Now it's really out of the list. Before doing the
1389 		 * REFRELE, set ire_next to NULL as ire_inactive asserts
1390 		 * so.
1391 		 */
1392 		ire->ire_next = NULL;
1393 		ire_refrele_notr(ire);
1394 		ire = ire_next;
1395 	}
1396 }
1397 
1398 /*
1399  * irb_refrele is the only caller of the function. It calls to unlink
1400  * all the CONDEMNED ires from this bucket.
1401  */
1402 ire_t *
ire_unlink(irb_t * irb)1403 ire_unlink(irb_t *irb)
1404 {
1405 	ire_t *ire;
1406 	ire_t *ire1;
1407 	ire_t **ptpn;
1408 	ire_t *ire_list = NULL;
1409 
1410 	ASSERT(RW_WRITE_HELD(&irb->irb_lock));
1411 	ASSERT(((irb->irb_marks & IRB_MARK_DYNAMIC) && irb->irb_refcnt == 1) ||
1412 	    (irb->irb_refcnt == 0));
1413 	ASSERT(irb->irb_marks & IRB_MARK_CONDEMNED);
1414 	ASSERT(irb->irb_ire != NULL);
1415 
1416 	for (ire = irb->irb_ire; ire != NULL; ire = ire1) {
1417 		ire1 = ire->ire_next;
1418 		if (IRE_IS_CONDEMNED(ire)) {
1419 			ptpn = ire->ire_ptpn;
1420 			ire1 = ire->ire_next;
1421 			if (ire1)
1422 				ire1->ire_ptpn = ptpn;
1423 			*ptpn = ire1;
1424 			ire->ire_ptpn = NULL;
1425 			ire->ire_next = NULL;
1426 
1427 			/*
1428 			 * We need to call ire_delete_v4 or ire_delete_v6 to
1429 			 * clean up dependents and the redirects pointing at
1430 			 * the default gateway. We need to drop the lock
1431 			 * as ire_flush_cache/ire_delete_host_redircts require
1432 			 * so. But we can't drop the lock, as ire_unlink needs
1433 			 * to atomically remove the ires from the list.
1434 			 * So, create a temporary list of CONDEMNED ires
1435 			 * for doing ire_delete_v4/ire_delete_v6 operations
1436 			 * later on.
1437 			 */
1438 			ire->ire_next = ire_list;
1439 			ire_list = ire;
1440 		}
1441 	}
1442 	irb->irb_marks &= ~IRB_MARK_CONDEMNED;
1443 	return (ire_list);
1444 }
1445 
1446 /*
1447  * Clean up the radix node for this ire. Must be called by irb_refrele
1448  * when there are no ire's left in the bucket. Returns TRUE if the bucket
1449  * is deleted and freed.
1450  */
1451 boolean_t
irb_inactive(irb_t * irb)1452 irb_inactive(irb_t *irb)
1453 {
1454 	struct rt_entry *rt;
1455 	struct radix_node *rn;
1456 	ip_stack_t *ipst = irb->irb_ipst;
1457 
1458 	ASSERT(irb->irb_ipst != NULL);
1459 
1460 	rt = IRB2RT(irb);
1461 	rn = (struct radix_node *)rt;
1462 
1463 	/* first remove it from the radix tree. */
1464 	RADIX_NODE_HEAD_WLOCK(ipst->ips_ip_ftable);
1465 	rw_enter(&irb->irb_lock, RW_WRITER);
1466 	if (irb->irb_refcnt == 1 && irb->irb_nire == 0) {
1467 		rn = ipst->ips_ip_ftable->rnh_deladdr(rn->rn_key, rn->rn_mask,
1468 		    ipst->ips_ip_ftable);
1469 		DTRACE_PROBE1(irb__free, rt_t *,  rt);
1470 		ASSERT((void *)rn == (void *)rt);
1471 		Free(rt, rt_entry_cache);
1472 		/* irb_lock is freed */
1473 		RADIX_NODE_HEAD_UNLOCK(ipst->ips_ip_ftable);
1474 		return (B_TRUE);
1475 	}
1476 	rw_exit(&irb->irb_lock);
1477 	RADIX_NODE_HEAD_UNLOCK(ipst->ips_ip_ftable);
1478 	return (B_FALSE);
1479 }
1480 
1481 /*
1482  * Delete the specified IRE.
1483  * We assume that if ire_bucket is not set then ire_ill->ill_ire_cnt was
1484  * not incremented i.e., that the insertion in the bucket and the increment
1485  * of that counter is done atomically.
1486  */
1487 void
ire_delete(ire_t * ire)1488 ire_delete(ire_t *ire)
1489 {
1490 	ire_t	*ire1;
1491 	ire_t	**ptpn;
1492 	irb_t	*irb;
1493 	ip_stack_t	*ipst = ire->ire_ipst;
1494 
1495 	if ((irb = ire->ire_bucket) == NULL) {
1496 		/*
1497 		 * It was never inserted in the list. Should call REFRELE
1498 		 * to free this IRE.
1499 		 */
1500 		ire_make_condemned(ire);
1501 		ire_refrele_notr(ire);
1502 		return;
1503 	}
1504 
1505 	/*
1506 	 * Move the use counts from an IRE_IF_CLONE to its parent
1507 	 * IRE_INTERFACE.
1508 	 * We need to do this before acquiring irb_lock.
1509 	 */
1510 	if (ire->ire_type & IRE_IF_CLONE) {
1511 		ire_t *parent;
1512 
1513 		rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
1514 		if ((parent = ire->ire_dep_parent) != NULL) {
1515 			parent->ire_ob_pkt_count += ire->ire_ob_pkt_count;
1516 			parent->ire_ib_pkt_count += ire->ire_ib_pkt_count;
1517 			ire->ire_ob_pkt_count = 0;
1518 			ire->ire_ib_pkt_count = 0;
1519 		}
1520 		rw_exit(&ipst->ips_ire_dep_lock);
1521 	}
1522 
1523 	rw_enter(&irb->irb_lock, RW_WRITER);
1524 	if (ire->ire_ptpn == NULL) {
1525 		/*
1526 		 * Some other thread has removed us from the list.
1527 		 * It should have done the REFRELE for us.
1528 		 */
1529 		rw_exit(&irb->irb_lock);
1530 		return;
1531 	}
1532 
1533 	if (!IRE_IS_CONDEMNED(ire)) {
1534 		/* Is this an IRE representing multiple duplicate entries? */
1535 		ASSERT(ire->ire_identical_ref >= 1);
1536 		if (atomic_dec_32_nv(&ire->ire_identical_ref) != 0) {
1537 			/* Removed one of the identical parties */
1538 			rw_exit(&irb->irb_lock);
1539 			return;
1540 		}
1541 
1542 		irb->irb_ire_cnt--;
1543 		ire_make_condemned(ire);
1544 	}
1545 
1546 	if (irb->irb_refcnt != 0) {
1547 		/*
1548 		 * The last thread to leave this bucket will
1549 		 * delete this ire.
1550 		 */
1551 		irb->irb_marks |= IRB_MARK_CONDEMNED;
1552 		rw_exit(&irb->irb_lock);
1553 		return;
1554 	}
1555 
1556 	/*
1557 	 * Normally to delete an ire, we walk the bucket. While we
1558 	 * walk the bucket, we normally bump up irb_refcnt and hence
1559 	 * we return from above where we mark CONDEMNED and the ire
1560 	 * gets deleted from ire_unlink. This case is where somebody
1561 	 * knows the ire e.g by doing a lookup, and wants to delete the
1562 	 * IRE. irb_refcnt would be 0 in this case if nobody is walking
1563 	 * the bucket.
1564 	 */
1565 	ptpn = ire->ire_ptpn;
1566 	ire1 = ire->ire_next;
1567 	if (ire1 != NULL)
1568 		ire1->ire_ptpn = ptpn;
1569 	ASSERT(ptpn != NULL);
1570 	*ptpn = ire1;
1571 	ire->ire_ptpn = NULL;
1572 	ire->ire_next = NULL;
1573 	if (ire->ire_ipversion == IPV6_VERSION) {
1574 		BUMP_IRE_STATS(ipst->ips_ire_stats_v6, ire_stats_deleted);
1575 	} else {
1576 		BUMP_IRE_STATS(ipst->ips_ire_stats_v4, ire_stats_deleted);
1577 	}
1578 	rw_exit(&irb->irb_lock);
1579 
1580 	/* Cleanup dependents and related stuff */
1581 	if (ire->ire_ipversion == IPV6_VERSION) {
1582 		ire_delete_v6(ire);
1583 	} else {
1584 		ire_delete_v4(ire);
1585 	}
1586 	/*
1587 	 * We removed it from the list. Decrement the
1588 	 * reference count.
1589 	 */
1590 	ire_refrele_notr(ire);
1591 }
1592 
1593 /*
1594  * Delete the specified IRE.
1595  * All calls should use ire_delete().
1596  * Sometimes called as writer though not required by this function.
1597  *
1598  * NOTE : This function is called only if the ire was added
1599  * in the list.
1600  */
1601 static void
ire_delete_v4(ire_t * ire)1602 ire_delete_v4(ire_t *ire)
1603 {
1604 	ip_stack_t	*ipst = ire->ire_ipst;
1605 
1606 	ASSERT(ire->ire_refcnt >= 1);
1607 	ASSERT(ire->ire_ipversion == IPV4_VERSION);
1608 
1609 	ire_flush_cache_v4(ire, IRE_FLUSH_DELETE);
1610 	if (ire->ire_type == IRE_DEFAULT) {
1611 		/*
1612 		 * when a default gateway is going away
1613 		 * delete all the host redirects pointing at that
1614 		 * gateway.
1615 		 */
1616 		ire_delete_host_redirects(ire->ire_gateway_addr, ipst);
1617 	}
1618 
1619 	/*
1620 	 * If we are deleting an IRE_INTERFACE then we make sure we also
1621 	 * delete any IRE_IF_CLONE that has been created from it.
1622 	 * Those are always in ire_dep_children.
1623 	 */
1624 	if ((ire->ire_type & IRE_INTERFACE) && ire->ire_dep_children != NULL)
1625 		ire_dep_delete_if_clone(ire);
1626 
1627 	/* Remove from parent dependencies and child */
1628 	rw_enter(&ipst->ips_ire_dep_lock, RW_WRITER);
1629 	if (ire->ire_dep_parent != NULL)
1630 		ire_dep_remove(ire);
1631 
1632 	while (ire->ire_dep_children != NULL)
1633 		ire_dep_remove(ire->ire_dep_children);
1634 	rw_exit(&ipst->ips_ire_dep_lock);
1635 }
1636 
1637 /*
1638  * ire_refrele is the only caller of the function. It calls
1639  * to free the ire when the reference count goes to zero.
1640  */
1641 void
ire_inactive(ire_t * ire)1642 ire_inactive(ire_t *ire)
1643 {
1644 	ill_t	*ill;
1645 	irb_t 	*irb;
1646 	ip_stack_t	*ipst = ire->ire_ipst;
1647 
1648 	ASSERT(ire->ire_refcnt == 0);
1649 	ASSERT(ire->ire_ptpn == NULL);
1650 	ASSERT(ire->ire_next == NULL);
1651 
1652 	/* Count how many condemned ires for kmem_cache callback */
1653 	ASSERT(IRE_IS_CONDEMNED(ire));
1654 	atomic_add_32(&ipst->ips_num_ire_condemned, -1);
1655 
1656 	if (ire->ire_gw_secattr != NULL) {
1657 		ire_gw_secattr_free(ire->ire_gw_secattr);
1658 		ire->ire_gw_secattr = NULL;
1659 	}
1660 
1661 	/*
1662 	 * ire_nce_cache is cleared in ire_delete, and we make sure we don't
1663 	 * set it once the ire is marked condemned.
1664 	 */
1665 	ASSERT(ire->ire_nce_cache == NULL);
1666 
1667 	/*
1668 	 * Since any parent would have a refhold on us they would already
1669 	 * have been removed.
1670 	 */
1671 	ASSERT(ire->ire_dep_parent == NULL);
1672 	ASSERT(ire->ire_dep_sib_next == NULL);
1673 	ASSERT(ire->ire_dep_sib_ptpn == NULL);
1674 
1675 	/*
1676 	 * Since any children would have a refhold on us they should have
1677 	 * already been removed.
1678 	 */
1679 	ASSERT(ire->ire_dep_children == NULL);
1680 
1681 	/*
1682 	 * ill_ire_ref is increased when the IRE is inserted in the
1683 	 * bucket - not when the IRE is created.
1684 	 */
1685 	irb = ire->ire_bucket;
1686 	ill = ire->ire_ill;
1687 	if (irb != NULL && ill != NULL) {
1688 		mutex_enter(&ill->ill_lock);
1689 		ASSERT(ill->ill_ire_cnt != 0);
1690 		DTRACE_PROBE3(ill__decr__cnt, (ill_t *), ill,
1691 		    (char *), "ire", (void *), ire);
1692 		ill->ill_ire_cnt--;
1693 		if (ILL_DOWN_OK(ill)) {
1694 			/* Drops the ill lock */
1695 			ipif_ill_refrele_tail(ill);
1696 		} else {
1697 			mutex_exit(&ill->ill_lock);
1698 		}
1699 	}
1700 	ire->ire_ill = NULL;
1701 
1702 	/* This should be true for both V4 and V6 */
1703 	if (irb != NULL && (irb->irb_marks & IRB_MARK_DYNAMIC)) {
1704 		rw_enter(&irb->irb_lock, RW_WRITER);
1705 		irb->irb_nire--;
1706 		/*
1707 		 * Instead of examining the conditions for freeing
1708 		 * the radix node here, we do it by calling
1709 		 * irb_refrele which is a single point in the code
1710 		 * that embeds that logic. Bump up the refcnt to
1711 		 * be able to call irb_refrele
1712 		 */
1713 		irb_refhold_locked(irb);
1714 		rw_exit(&irb->irb_lock);
1715 		irb_refrele(irb);
1716 	}
1717 
1718 #ifdef DEBUG
1719 	ire_trace_cleanup(ire);
1720 #endif
1721 	mutex_destroy(&ire->ire_lock);
1722 	if (ire->ire_ipversion == IPV6_VERSION) {
1723 		BUMP_IRE_STATS(ipst->ips_ire_stats_v6, ire_stats_freed);
1724 	} else {
1725 		BUMP_IRE_STATS(ipst->ips_ire_stats_v4, ire_stats_freed);
1726 	}
1727 	kmem_cache_free(ire_cache, ire);
1728 }
1729 
1730 /*
1731  * ire_update_generation is the callback function provided by
1732  * ire_get_bucket() to update the generation number of any
1733  * matching shorter route when a new route is added.
1734  *
1735  * This fucntion always returns a failure return (B_FALSE)
1736  * to force the caller (rn_matchaddr_args)
1737  * to back-track up the tree looking for shorter matches.
1738  */
1739 /* ARGSUSED */
1740 static boolean_t
ire_update_generation(struct radix_node * rn,void * arg)1741 ire_update_generation(struct radix_node *rn, void *arg)
1742 {
1743 	struct rt_entry *rt = (struct rt_entry *)rn;
1744 
1745 	/* We need to handle all in the same bucket */
1746 	irb_increment_generation(&rt->rt_irb);
1747 	return (B_FALSE);
1748 }
1749 
1750 /*
1751  * Take care of all the generation numbers in the bucket.
1752  */
1753 void
irb_increment_generation(irb_t * irb)1754 irb_increment_generation(irb_t *irb)
1755 {
1756 	ire_t *ire;
1757 	ip_stack_t *ipst;
1758 
1759 	if (irb == NULL || irb->irb_ire_cnt == 0)
1760 		return;
1761 
1762 	ipst = irb->irb_ipst;
1763 	/*
1764 	 * we cannot do an irb_refhold/irb_refrele here as the caller
1765 	 * already has the global RADIX_NODE_HEAD_WLOCK, and the irb_refrele
1766 	 * may result in an attempt to free the irb_t, which also needs
1767 	 * the RADIX_NODE_HEAD lock. However, since we want to traverse the
1768 	 * irb_ire list without fear of having a condemned ire removed from
1769 	 * the list, we acquire the irb_lock as WRITER. Moreover, since
1770 	 * the ire_generation increments are done under the ire_dep_lock,
1771 	 * acquire the locks in the prescribed lock order first.
1772 	 */
1773 	rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
1774 	rw_enter(&irb->irb_lock, RW_WRITER);
1775 	for (ire = irb->irb_ire; ire != NULL; ire = ire->ire_next) {
1776 		if (!IRE_IS_CONDEMNED(ire))
1777 			ire_increment_generation(ire);	/* Ourselves */
1778 		ire_dep_incr_generation_locked(ire);	/* Dependants */
1779 	}
1780 	rw_exit(&irb->irb_lock);
1781 	rw_exit(&ipst->ips_ire_dep_lock);
1782 }
1783 
1784 /*
1785  * When an IRE is added or deleted this routine is called to make sure
1786  * any caching of IRE information is notified or updated.
1787  *
1788  * The flag argument indicates if the flush request is due to addition
1789  * of new route (IRE_FLUSH_ADD), deletion of old route (IRE_FLUSH_DELETE),
1790  * or a change to ire_gateway_addr (IRE_FLUSH_GWCHANGE).
1791  */
1792 void
ire_flush_cache_v4(ire_t * ire,int flag)1793 ire_flush_cache_v4(ire_t *ire, int flag)
1794 {
1795 	irb_t *irb = ire->ire_bucket;
1796 	struct rt_entry *rt = IRB2RT(irb);
1797 	ip_stack_t *ipst = ire->ire_ipst;
1798 
1799 	/*
1800 	 * IRE_IF_CLONE ire's don't provide any new information
1801 	 * than the parent from which they are cloned, so don't
1802 	 * perturb the generation numbers.
1803 	 */
1804 	if (ire->ire_type & IRE_IF_CLONE)
1805 		return;
1806 
1807 	/*
1808 	 * Ensure that an ire_add during a lookup serializes the updates of the
1809 	 * generation numbers under the radix head lock so that the lookup gets
1810 	 * either the old ire and old generation number, or a new ire and new
1811 	 * generation number.
1812 	 */
1813 	RADIX_NODE_HEAD_WLOCK(ipst->ips_ip_ftable);
1814 
1815 	/*
1816 	 * If a route was just added, we need to notify everybody that
1817 	 * has cached an IRE_NOROUTE since there might now be a better
1818 	 * route for them.
1819 	 */
1820 	if (flag == IRE_FLUSH_ADD) {
1821 		ire_increment_generation(ipst->ips_ire_reject_v4);
1822 		ire_increment_generation(ipst->ips_ire_blackhole_v4);
1823 	}
1824 
1825 	/* Adding a default can't otherwise provide a better route */
1826 	if (ire->ire_type == IRE_DEFAULT && flag == IRE_FLUSH_ADD) {
1827 		RADIX_NODE_HEAD_UNLOCK(ipst->ips_ip_ftable);
1828 		return;
1829 	}
1830 
1831 	switch (flag) {
1832 	case IRE_FLUSH_DELETE:
1833 	case IRE_FLUSH_GWCHANGE:
1834 		/*
1835 		 * Update ire_generation for all ire_dep_children chains
1836 		 * starting with this IRE
1837 		 */
1838 		ire_dep_incr_generation(ire);
1839 		break;
1840 	case IRE_FLUSH_ADD:
1841 		/*
1842 		 * Update the generation numbers of all shorter matching routes.
1843 		 * ire_update_generation takes care of the dependants by
1844 		 * using ire_dep_incr_generation.
1845 		 */
1846 		(void) ipst->ips_ip_ftable->rnh_matchaddr_args(&rt->rt_dst,
1847 		    ipst->ips_ip_ftable, ire_update_generation, NULL);
1848 		break;
1849 	}
1850 	RADIX_NODE_HEAD_UNLOCK(ipst->ips_ip_ftable);
1851 }
1852 
1853 /*
1854  * Matches the arguments passed with the values in the ire.
1855  *
1856  * Note: for match types that match using "ill" passed in, ill
1857  * must be checked for non-NULL before calling this routine.
1858  */
1859 boolean_t
ire_match_args(ire_t * ire,ipaddr_t addr,ipaddr_t mask,ipaddr_t gateway,int type,const ill_t * ill,zoneid_t zoneid,const ts_label_t * tsl,int match_flags)1860 ire_match_args(ire_t *ire, ipaddr_t addr, ipaddr_t mask, ipaddr_t gateway,
1861     int type, const ill_t *ill, zoneid_t zoneid,
1862     const ts_label_t *tsl, int match_flags)
1863 {
1864 	ill_t *ire_ill = NULL, *dst_ill;
1865 	ip_stack_t *ipst = ire->ire_ipst;
1866 
1867 	ASSERT(ire->ire_ipversion == IPV4_VERSION);
1868 	ASSERT((ire->ire_addr & ~ire->ire_mask) == 0);
1869 	ASSERT((!(match_flags & (MATCH_IRE_ILL|MATCH_IRE_SRC_ILL))) ||
1870 	    (ill != NULL && !ill->ill_isv6));
1871 
1872 	/*
1873 	 * If MATCH_IRE_TESTHIDDEN is set, then only return the IRE if it is
1874 	 * in fact hidden, to ensure the caller gets the right one.
1875 	 */
1876 	if (ire->ire_testhidden) {
1877 		if (!(match_flags & MATCH_IRE_TESTHIDDEN))
1878 			return (B_FALSE);
1879 	}
1880 
1881 	if (zoneid != ALL_ZONES && zoneid != ire->ire_zoneid &&
1882 	    ire->ire_zoneid != ALL_ZONES) {
1883 		/*
1884 		 * If MATCH_IRE_ZONEONLY has been set and the supplied zoneid
1885 		 * does not match that of ire_zoneid, a failure to
1886 		 * match is reported at this point. Otherwise, since some IREs
1887 		 * that are available in the global zone can be used in local
1888 		 * zones, additional checks need to be performed:
1889 		 *
1890 		 * IRE_LOOPBACK
1891 		 *	entries should never be matched in this situation.
1892 		 *	Each zone has its own IRE_LOOPBACK.
1893 		 *
1894 		 * IRE_LOCAL
1895 		 *	We allow them for any zoneid. ire_route_recursive
1896 		 *	does additional checks when
1897 		 *	ip_restrict_interzone_loopback is set.
1898 		 *
1899 		 * If ill_usesrc_ifindex is set
1900 		 *	Then we check if the zone has a valid source address
1901 		 *	on the usesrc ill.
1902 		 *
1903 		 * If ire_ill is set, then check that the zone has an ipif
1904 		 *	on that ill.
1905 		 *
1906 		 * Outside of this function (in ire_round_robin) we check
1907 		 * that any IRE_OFFLINK has a gateway that reachable from the
1908 		 * zone when we have multiple choices (ECMP).
1909 		 */
1910 		if (match_flags & MATCH_IRE_ZONEONLY)
1911 			return (B_FALSE);
1912 		if (ire->ire_type & IRE_LOOPBACK)
1913 			return (B_FALSE);
1914 
1915 		if (ire->ire_type & IRE_LOCAL)
1916 			goto matchit;
1917 
1918 		/*
1919 		 * The normal case of IRE_ONLINK has a matching zoneid.
1920 		 * Here we handle the case when shared-IP zones have been
1921 		 * configured with IP addresses on vniN. In that case it
1922 		 * is ok for traffic from a zone to use IRE_ONLINK routes
1923 		 * if the ill has a usesrc pointing at vniN
1924 		 */
1925 		dst_ill = ire->ire_ill;
1926 		if (ire->ire_type & IRE_ONLINK) {
1927 			uint_t	ifindex;
1928 
1929 			/*
1930 			 * Note there is no IRE_INTERFACE on vniN thus
1931 			 * can't do an IRE lookup for a matching route.
1932 			 */
1933 			ifindex = dst_ill->ill_usesrc_ifindex;
1934 			if (ifindex == 0)
1935 				return (B_FALSE);
1936 
1937 			/*
1938 			 * If there is a usable source address in the
1939 			 * zone, then it's ok to return this IRE_INTERFACE
1940 			 */
1941 			if (!ipif_zone_avail(ifindex, dst_ill->ill_isv6,
1942 			    zoneid, ipst)) {
1943 				ip3dbg(("ire_match_args: no usrsrc for zone"
1944 				    " dst_ill %p\n", (void *)dst_ill));
1945 				return (B_FALSE);
1946 			}
1947 		}
1948 		/*
1949 		 * For example, with
1950 		 * route add 11.0.0.0 gw1 -ifp bge0
1951 		 * route add 11.0.0.0 gw2 -ifp bge1
1952 		 * this code would differentiate based on
1953 		 * where the sending zone has addresses.
1954 		 * Only if the zone has an address on bge0 can it use the first
1955 		 * route. It isn't clear if this behavior is documented
1956 		 * anywhere.
1957 		 */
1958 		if (dst_ill != NULL && (ire->ire_type & IRE_OFFLINK)) {
1959 			ipif_t	*tipif;
1960 
1961 			mutex_enter(&dst_ill->ill_lock);
1962 			for (tipif = dst_ill->ill_ipif;
1963 			    tipif != NULL; tipif = tipif->ipif_next) {
1964 				if (!IPIF_IS_CONDEMNED(tipif) &&
1965 				    (tipif->ipif_flags & IPIF_UP) &&
1966 				    (tipif->ipif_zoneid == zoneid ||
1967 				    tipif->ipif_zoneid == ALL_ZONES))
1968 					break;
1969 			}
1970 			mutex_exit(&dst_ill->ill_lock);
1971 			if (tipif == NULL) {
1972 				return (B_FALSE);
1973 			}
1974 		}
1975 	}
1976 
1977 matchit:
1978 	ire_ill = ire->ire_ill;
1979 	if (match_flags & MATCH_IRE_ILL) {
1980 
1981 		/*
1982 		 * If asked to match an ill, we *must* match
1983 		 * on the ire_ill for ipmp test addresses, or
1984 		 * any of the ill in the group for data addresses.
1985 		 * If we don't, we may as well fail.
1986 		 * However, we need an exception for IRE_LOCALs to ensure
1987 		 * we loopback packets even sent to test addresses on different
1988 		 * interfaces in the group.
1989 		 */
1990 		if ((match_flags & MATCH_IRE_TESTHIDDEN) &&
1991 		    !(ire->ire_type & IRE_LOCAL)) {
1992 			if (ire->ire_ill != ill)
1993 				return (B_FALSE);
1994 		} else  {
1995 			match_flags &= ~MATCH_IRE_TESTHIDDEN;
1996 			/*
1997 			 * We know that ill is not NULL, but ire_ill could be
1998 			 * NULL
1999 			 */
2000 			if (ire_ill == NULL || !IS_ON_SAME_LAN(ill, ire_ill))
2001 				return (B_FALSE);
2002 		}
2003 	}
2004 	if (match_flags & MATCH_IRE_SRC_ILL) {
2005 		if (ire_ill == NULL)
2006 			return (B_FALSE);
2007 		if (!IS_ON_SAME_LAN(ill, ire_ill)) {
2008 			if (ire_ill->ill_usesrc_ifindex == 0 ||
2009 			    (ire_ill->ill_usesrc_ifindex !=
2010 			    ill->ill_phyint->phyint_ifindex))
2011 				return (B_FALSE);
2012 		}
2013 	}
2014 
2015 	if ((ire->ire_addr == (addr & mask)) &&
2016 	    ((!(match_flags & MATCH_IRE_GW)) ||
2017 	    (ire->ire_gateway_addr == gateway)) &&
2018 	    ((!(match_flags & MATCH_IRE_DIRECT)) ||
2019 	    !(ire->ire_flags & RTF_INDIRECT)) &&
2020 	    ((!(match_flags & MATCH_IRE_TYPE)) || (ire->ire_type & type)) &&
2021 	    ((!(match_flags & MATCH_IRE_TESTHIDDEN)) || ire->ire_testhidden) &&
2022 	    ((!(match_flags & MATCH_IRE_MASK)) || (ire->ire_mask == mask)) &&
2023 	    ((!(match_flags & MATCH_IRE_SECATTR)) ||
2024 	    (!is_system_labeled()) ||
2025 	    (tsol_ire_match_gwattr(ire, tsl) == 0))) {
2026 		/* We found the matched IRE */
2027 		return (B_TRUE);
2028 	}
2029 	return (B_FALSE);
2030 }
2031 
2032 /*
2033  * Check if the IRE_LOCAL uses the same ill as another route would use.
2034  * If there is no alternate route, or the alternate is a REJECT or BLACKHOLE,
2035  * then we don't allow this IRE_LOCAL to be used.
2036  * We always return an IRE; will be RTF_REJECT if no route available.
2037  */
2038 ire_t *
ire_alt_local(ire_t * ire,zoneid_t zoneid,const ts_label_t * tsl,const ill_t * ill,uint_t * generationp)2039 ire_alt_local(ire_t *ire, zoneid_t zoneid, const ts_label_t *tsl,
2040     const ill_t *ill, uint_t *generationp)
2041 {
2042 	ip_stack_t	*ipst = ire->ire_ipst;
2043 	ire_t		*alt_ire;
2044 	uint_t		ire_type;
2045 	uint_t		generation;
2046 	uint_t		match_flags;
2047 
2048 	ASSERT(ire->ire_type & IRE_LOCAL);
2049 	ASSERT(ire->ire_ill != NULL);
2050 
2051 	/*
2052 	 * Need to match on everything but local.
2053 	 * This might result in the creation of a IRE_IF_CLONE for the
2054 	 * same address as the IRE_LOCAL when restrict_interzone_loopback is
2055 	 * set. ire_add_*() ensures that the IRE_IF_CLONE are tail inserted
2056 	 * to make sure the IRE_LOCAL is always found first.
2057 	 */
2058 	ire_type = (IRE_ONLINK | IRE_OFFLINK) & ~(IRE_LOCAL|IRE_LOOPBACK);
2059 	match_flags = MATCH_IRE_TYPE | MATCH_IRE_SECATTR;
2060 	if (ill != NULL)
2061 		match_flags |= MATCH_IRE_ILL;
2062 
2063 	if (ire->ire_ipversion == IPV4_VERSION) {
2064 		alt_ire = ire_route_recursive_v4(ire->ire_addr, ire_type,
2065 		    ill, zoneid, tsl, match_flags, IRR_ALLOCATE, 0, ipst, NULL,
2066 		    NULL, &generation);
2067 	} else {
2068 		alt_ire = ire_route_recursive_v6(&ire->ire_addr_v6, ire_type,
2069 		    ill, zoneid, tsl, match_flags, IRR_ALLOCATE, 0, ipst, NULL,
2070 		    NULL, &generation);
2071 	}
2072 	ASSERT(alt_ire != NULL);
2073 
2074 	if (alt_ire->ire_ill == ire->ire_ill) {
2075 		/* Going out the same ILL - ok to send to IRE_LOCAL */
2076 		ire_refrele(alt_ire);
2077 	} else {
2078 		/* Different ill - ignore IRE_LOCAL */
2079 		ire_refrele(ire);
2080 		ire = alt_ire;
2081 		if (generationp != NULL)
2082 			*generationp = generation;
2083 	}
2084 	return (ire);
2085 }
2086 
2087 boolean_t
ire_find_zoneid(struct radix_node * rn,void * arg)2088 ire_find_zoneid(struct radix_node *rn, void *arg)
2089 {
2090 	struct rt_entry *rt = (struct rt_entry *)rn;
2091 	irb_t *irb;
2092 	ire_t *ire;
2093 	ire_ftable_args_t *margs = arg;
2094 
2095 	ASSERT(rt != NULL);
2096 
2097 	irb = &rt->rt_irb;
2098 
2099 	if (irb->irb_ire_cnt == 0)
2100 		return (B_FALSE);
2101 
2102 	rw_enter(&irb->irb_lock, RW_READER);
2103 	for (ire = irb->irb_ire; ire != NULL; ire = ire->ire_next) {
2104 		if (IRE_IS_CONDEMNED(ire))
2105 			continue;
2106 
2107 		if (!(ire->ire_type & IRE_INTERFACE))
2108 			continue;
2109 
2110 		if (ire->ire_zoneid != ALL_ZONES &&
2111 		    ire->ire_zoneid != margs->ift_zoneid)
2112 			continue;
2113 
2114 		if (margs->ift_ill != NULL && margs->ift_ill != ire->ire_ill)
2115 			continue;
2116 
2117 		if (is_system_labeled() &&
2118 		    tsol_ire_match_gwattr(ire, margs->ift_tsl) != 0)
2119 			continue;
2120 
2121 		rw_exit(&irb->irb_lock);
2122 		return (B_TRUE);
2123 	}
2124 	rw_exit(&irb->irb_lock);
2125 	return (B_FALSE);
2126 }
2127 
2128 /*
2129  * Check if the zoneid (not ALL_ZONES) has an IRE_INTERFACE for the specified
2130  * gateway address. If ill is non-NULL we also match on it.
2131  * The caller must hold a read lock on RADIX_NODE_HEAD if lock_held is set.
2132  */
2133 boolean_t
ire_gateway_ok_zone_v4(ipaddr_t gateway,zoneid_t zoneid,ill_t * ill,const ts_label_t * tsl,ip_stack_t * ipst,boolean_t lock_held)2134 ire_gateway_ok_zone_v4(ipaddr_t gateway, zoneid_t zoneid, ill_t *ill,
2135     const ts_label_t *tsl, ip_stack_t *ipst, boolean_t lock_held)
2136 {
2137 	struct rt_sockaddr rdst;
2138 	struct rt_entry *rt;
2139 	ire_ftable_args_t margs;
2140 
2141 	ASSERT(ill == NULL || !ill->ill_isv6);
2142 	if (lock_held)
2143 		ASSERT(RW_READ_HELD(&ipst->ips_ip_ftable->rnh_lock));
2144 	else
2145 		RADIX_NODE_HEAD_RLOCK(ipst->ips_ip_ftable);
2146 
2147 	bzero(&rdst, sizeof (rdst));
2148 	rdst.rt_sin_len = sizeof (rdst);
2149 	rdst.rt_sin_family = AF_INET;
2150 	rdst.rt_sin_addr.s_addr = gateway;
2151 
2152 	/*
2153 	 * We only use margs for ill, zoneid, and tsl matching in
2154 	 * ire_find_zoneid
2155 	 */
2156 	bzero(&margs, sizeof (margs));
2157 	margs.ift_ill = ill;
2158 	margs.ift_zoneid = zoneid;
2159 	margs.ift_tsl = tsl;
2160 	rt = (struct rt_entry *)ipst->ips_ip_ftable->rnh_matchaddr_args(&rdst,
2161 	    ipst->ips_ip_ftable, ire_find_zoneid, (void *)&margs);
2162 
2163 	if (!lock_held)
2164 		RADIX_NODE_HEAD_UNLOCK(ipst->ips_ip_ftable);
2165 
2166 	return (rt != NULL);
2167 }
2168 
2169 /*
2170  * ire_walk routine to delete a fraction of redirect IREs and IRE_CLONE_IF IREs.
2171  * The fraction argument tells us what fraction of the IREs to delete.
2172  * Common for IPv4 and IPv6.
2173  * Used when memory backpressure.
2174  */
2175 static void
ire_delete_reclaim(ire_t * ire,char * arg)2176 ire_delete_reclaim(ire_t *ire, char *arg)
2177 {
2178 	ip_stack_t	*ipst = ire->ire_ipst;
2179 	uint_t		fraction = *(uint_t *)arg;
2180 	uint_t		rand;
2181 
2182 	if ((ire->ire_flags & RTF_DYNAMIC) ||
2183 	    (ire->ire_type & IRE_IF_CLONE)) {
2184 
2185 		/* Pick a random number */
2186 		rand = (uint_t)ddi_get_lbolt() +
2187 		    IRE_ADDR_HASH_V6(ire->ire_addr_v6, 256);
2188 
2189 		/* Use truncation */
2190 		if ((rand/fraction)*fraction == rand) {
2191 			IP_STAT(ipst, ip_ire_reclaim_deleted);
2192 			ire_delete(ire);
2193 		}
2194 	}
2195 
2196 }
2197 
2198 /*
2199  * kmem_cache callback to free up memory.
2200  *
2201  * Free a fraction (ips_ip_ire_reclaim_fraction) of things IP added dynamically
2202  * (RTF_DYNAMIC and IRE_IF_CLONE).
2203  */
2204 static void
ip_ire_reclaim_stack(ip_stack_t * ipst)2205 ip_ire_reclaim_stack(ip_stack_t *ipst)
2206 {
2207 	uint_t	fraction = ipst->ips_ip_ire_reclaim_fraction;
2208 
2209 	IP_STAT(ipst, ip_ire_reclaim_calls);
2210 
2211 	ire_walk(ire_delete_reclaim, &fraction, ipst);
2212 
2213 	/*
2214 	 * Walk all CONNs that can have a reference on an ire, nce or dce.
2215 	 * Get them to update any stale references to drop any refholds they
2216 	 * have.
2217 	 */
2218 	ipcl_walk(conn_ixa_cleanup, (void *)B_FALSE, ipst);
2219 }
2220 
2221 /*
2222  * Called by the memory allocator subsystem directly, when the system
2223  * is running low on memory.
2224  */
2225 /* ARGSUSED */
2226 void
ip_ire_reclaim(void * args)2227 ip_ire_reclaim(void *args)
2228 {
2229 	netstack_handle_t nh;
2230 	netstack_t *ns;
2231 	ip_stack_t *ipst;
2232 
2233 	netstack_next_init(&nh);
2234 	while ((ns = netstack_next(&nh)) != NULL) {
2235 		/*
2236 		 * netstack_next() can return a netstack_t with a NULL
2237 		 * netstack_ip at boot time.
2238 		 */
2239 		if ((ipst = ns->netstack_ip) == NULL) {
2240 			netstack_rele(ns);
2241 			continue;
2242 		}
2243 		ip_ire_reclaim_stack(ipst);
2244 		netstack_rele(ns);
2245 	}
2246 	netstack_next_fini(&nh);
2247 }
2248 
2249 static void
power2_roundup(uint32_t * value)2250 power2_roundup(uint32_t *value)
2251 {
2252 	int i;
2253 
2254 	for (i = 1; i < 31; i++) {
2255 		if (*value <= (1 << i))
2256 			break;
2257 	}
2258 	*value = (1 << i);
2259 }
2260 
2261 /* Global init for all zones */
2262 void
ip_ire_g_init()2263 ip_ire_g_init()
2264 {
2265 	/*
2266 	 * Create kmem_caches.  ip_ire_reclaim() and ip_nce_reclaim()
2267 	 * will give disposable IREs back to system when needed.
2268 	 * This needs to be done here before anything else, since
2269 	 * ire_add() expects the cache to be created.
2270 	 */
2271 	ire_cache = kmem_cache_create("ire_cache",
2272 	    sizeof (ire_t), 0, NULL, NULL,
2273 	    ip_ire_reclaim, NULL, NULL, 0);
2274 
2275 	ncec_cache = kmem_cache_create("ncec_cache",
2276 	    sizeof (ncec_t), 0, NULL, NULL,
2277 	    ip_nce_reclaim, NULL, NULL, 0);
2278 	nce_cache = kmem_cache_create("nce_cache",
2279 	    sizeof (nce_t), 0, NULL, NULL,
2280 	    NULL, NULL, NULL, 0);
2281 
2282 	rt_entry_cache = kmem_cache_create("rt_entry",
2283 	    sizeof (struct rt_entry), 0, NULL, NULL, NULL, NULL, NULL, 0);
2284 
2285 	/*
2286 	 * Have radix code setup kmem caches etc.
2287 	 */
2288 	rn_init();
2289 }
2290 
2291 void
ip_ire_init(ip_stack_t * ipst)2292 ip_ire_init(ip_stack_t *ipst)
2293 {
2294 	ire_t	*ire;
2295 	int	error;
2296 
2297 	mutex_init(&ipst->ips_ire_ft_init_lock, NULL, MUTEX_DEFAULT, 0);
2298 
2299 	(void) rn_inithead((void **)&ipst->ips_ip_ftable, 32);
2300 
2301 	/*
2302 	 * Make sure that the forwarding table size is a power of 2.
2303 	 * The IRE*_ADDR_HASH() macroes depend on that.
2304 	 */
2305 	ipst->ips_ip6_ftable_hash_size = ip6_ftable_hash_size;
2306 	power2_roundup(&ipst->ips_ip6_ftable_hash_size);
2307 
2308 	/*
2309 	 * Allocate/initialize a pair of IRE_NOROUTEs for each of IPv4 and IPv6.
2310 	 * The ire_reject_v* has RTF_REJECT set, and the ire_blackhole_v* has
2311 	 * RTF_BLACKHOLE set. We use the latter for transient errors such
2312 	 * as memory allocation failures and tripping on IRE_IS_CONDEMNED
2313 	 * entries.
2314 	 */
2315 	ire = kmem_cache_alloc(ire_cache, KM_SLEEP);
2316 	*ire = ire_null;
2317 	error = ire_init_v4(ire, 0, 0, 0, IRE_NOROUTE, NULL, ALL_ZONES,
2318 	    RTF_REJECT|RTF_UP, NULL, ipst);
2319 	ASSERT(error == 0);
2320 	ipst->ips_ire_reject_v4 = ire;
2321 
2322 	ire = kmem_cache_alloc(ire_cache, KM_SLEEP);
2323 	*ire = ire_null;
2324 	error = ire_init_v6(ire, 0, 0, 0, IRE_NOROUTE, NULL, ALL_ZONES,
2325 	    RTF_REJECT|RTF_UP, NULL, ipst);
2326 	ASSERT(error == 0);
2327 	ipst->ips_ire_reject_v6 = ire;
2328 
2329 	ire = kmem_cache_alloc(ire_cache, KM_SLEEP);
2330 	*ire = ire_null;
2331 	error = ire_init_v4(ire, 0, 0, 0, IRE_NOROUTE, NULL, ALL_ZONES,
2332 	    RTF_BLACKHOLE|RTF_UP, NULL, ipst);
2333 	ASSERT(error == 0);
2334 	ipst->ips_ire_blackhole_v4 = ire;
2335 
2336 	ire = kmem_cache_alloc(ire_cache, KM_SLEEP);
2337 	*ire = ire_null;
2338 	error = ire_init_v6(ire, 0, 0, 0, IRE_NOROUTE, NULL, ALL_ZONES,
2339 	    RTF_BLACKHOLE|RTF_UP, NULL, ipst);
2340 	ASSERT(error == 0);
2341 	ipst->ips_ire_blackhole_v6 = ire;
2342 
2343 	rw_init(&ipst->ips_ip6_ire_head_lock, NULL, RW_DEFAULT, NULL);
2344 	rw_init(&ipst->ips_ire_dep_lock, NULL, RW_DEFAULT, NULL);
2345 }
2346 
2347 void
ip_ire_g_fini(void)2348 ip_ire_g_fini(void)
2349 {
2350 	kmem_cache_destroy(ire_cache);
2351 	kmem_cache_destroy(ncec_cache);
2352 	kmem_cache_destroy(nce_cache);
2353 	kmem_cache_destroy(rt_entry_cache);
2354 
2355 	rn_fini();
2356 }
2357 
2358 void
ip_ire_fini(ip_stack_t * ipst)2359 ip_ire_fini(ip_stack_t *ipst)
2360 {
2361 	int i;
2362 
2363 	ire_make_condemned(ipst->ips_ire_reject_v6);
2364 	ire_refrele_notr(ipst->ips_ire_reject_v6);
2365 	ipst->ips_ire_reject_v6 = NULL;
2366 
2367 	ire_make_condemned(ipst->ips_ire_reject_v4);
2368 	ire_refrele_notr(ipst->ips_ire_reject_v4);
2369 	ipst->ips_ire_reject_v4 = NULL;
2370 
2371 	ire_make_condemned(ipst->ips_ire_blackhole_v6);
2372 	ire_refrele_notr(ipst->ips_ire_blackhole_v6);
2373 	ipst->ips_ire_blackhole_v6 = NULL;
2374 
2375 	ire_make_condemned(ipst->ips_ire_blackhole_v4);
2376 	ire_refrele_notr(ipst->ips_ire_blackhole_v4);
2377 	ipst->ips_ire_blackhole_v4 = NULL;
2378 
2379 	/*
2380 	 * Delete all IREs - assumes that the ill/ipifs have
2381 	 * been removed so what remains are just the ftable to handle.
2382 	 */
2383 	ire_walk(ire_delete, NULL, ipst);
2384 
2385 	rn_freehead(ipst->ips_ip_ftable);
2386 	ipst->ips_ip_ftable = NULL;
2387 
2388 	rw_destroy(&ipst->ips_ire_dep_lock);
2389 	rw_destroy(&ipst->ips_ip6_ire_head_lock);
2390 
2391 	mutex_destroy(&ipst->ips_ire_ft_init_lock);
2392 
2393 	for (i = 0; i < IP6_MASK_TABLE_SIZE; i++) {
2394 		irb_t *ptr;
2395 		int j;
2396 
2397 		if ((ptr = ipst->ips_ip_forwarding_table_v6[i]) == NULL)
2398 			continue;
2399 
2400 		for (j = 0; j < ipst->ips_ip6_ftable_hash_size; j++) {
2401 			ASSERT(ptr[j].irb_ire == NULL);
2402 			rw_destroy(&ptr[j].irb_lock);
2403 		}
2404 		mi_free(ptr);
2405 		ipst->ips_ip_forwarding_table_v6[i] = NULL;
2406 	}
2407 }
2408 
2409 #ifdef DEBUG
2410 void
ire_trace_ref(ire_t * ire)2411 ire_trace_ref(ire_t *ire)
2412 {
2413 	mutex_enter(&ire->ire_lock);
2414 	if (ire->ire_trace_disable) {
2415 		mutex_exit(&ire->ire_lock);
2416 		return;
2417 	}
2418 
2419 	if (th_trace_ref(ire, ire->ire_ipst)) {
2420 		mutex_exit(&ire->ire_lock);
2421 	} else {
2422 		ire->ire_trace_disable = B_TRUE;
2423 		mutex_exit(&ire->ire_lock);
2424 		ire_trace_cleanup(ire);
2425 	}
2426 }
2427 
2428 void
ire_untrace_ref(ire_t * ire)2429 ire_untrace_ref(ire_t *ire)
2430 {
2431 	mutex_enter(&ire->ire_lock);
2432 	if (!ire->ire_trace_disable)
2433 		th_trace_unref(ire);
2434 	mutex_exit(&ire->ire_lock);
2435 }
2436 
2437 static void
ire_trace_cleanup(const ire_t * ire)2438 ire_trace_cleanup(const ire_t *ire)
2439 {
2440 	th_trace_cleanup(ire, ire->ire_trace_disable);
2441 }
2442 #endif /* DEBUG */
2443 
2444 /*
2445  * Find, or create if needed, the nce_t pointer to the neighbor cache
2446  * entry ncec_t for an IPv4 address. The nce_t will be created on the ill_t
2447  * in the non-IPMP case, or on the cast-ill in the IPMP bcast/mcast case, or
2448  * on the next available under-ill (selected by the IPMP rotor) in the
2449  * unicast IPMP case.
2450  *
2451  * If a neighbor-cache entry has to be created (i.e., one does not already
2452  * exist in the nce list) the ncec_lladdr and ncec_state of the neighbor cache
2453  * entry are initialized in nce_add_v4(). The broadcast, multicast, and
2454  * link-layer type determine the contents of {ncec_state, ncec_lladdr} of
2455  * the ncec_t created. The ncec_lladdr is non-null for all link types with
2456  * non-zero ill_phys_addr_length, though the contents may be zero in cases
2457  * where the link-layer type is not known at the time of creation
2458  * (e.g., IRE_IFRESOLVER links)
2459  *
2460  * All IRE_BROADCAST entries have ncec_state = ND_REACHABLE, and the nce_lladr
2461  * has the physical broadcast address of the outgoing interface.
2462  * For unicast ire entries,
2463  *   - if the outgoing interface is of type IRE_IF_RESOLVER, a newly created
2464  *     ncec_t with 0 nce_lladr contents, and will be in the ND_INITIAL state.
2465  *   - if the outgoing interface is a IRE_IF_NORESOLVER interface, no link
2466  *     layer resolution is necessary, so that the ncec_t will be in the
2467  *     ND_REACHABLE state
2468  *
2469  * The link layer information needed for broadcast addresses, and for
2470  * packets sent on IRE_IF_NORESOLVER interfaces is a constant mapping that
2471  * never needs re-verification for the lifetime of the ncec_t. These are
2472  * therefore marked NCE_F_NONUD.
2473  *
2474  * The nce returned will be created such that the nce_ill == ill that
2475  * is passed in. Note that the nce itself may not have ncec_ill == ill
2476  * where IPMP links are involved.
2477  */
2478 static nce_t *
ire_nce_init(ill_t * ill,const void * addr,int ire_type)2479 ire_nce_init(ill_t *ill, const void *addr, int ire_type)
2480 {
2481 	int		err;
2482 	nce_t		*nce = NULL;
2483 	uint16_t	ncec_flags;
2484 	uchar_t		*hwaddr;
2485 	boolean_t	need_refrele = B_FALSE;
2486 	ill_t		*in_ill = ill;
2487 	boolean_t	is_unicast;
2488 	uint_t		hwaddr_len;
2489 
2490 	is_unicast = ((ire_type & (IRE_MULTICAST|IRE_BROADCAST)) == 0);
2491 	if (IS_IPMP(ill) ||
2492 	    ((ire_type & IRE_BROADCAST) && IS_UNDER_IPMP(ill))) {
2493 		if ((ill = ipmp_ill_hold_xmit_ill(ill, is_unicast)) == NULL)
2494 			return (NULL);
2495 		need_refrele = B_TRUE;
2496 	}
2497 	ncec_flags = (ill->ill_flags & ILLF_NONUD) ? NCE_F_NONUD : 0;
2498 
2499 	switch (ire_type) {
2500 	case IRE_BROADCAST:
2501 		ASSERT(!ill->ill_isv6);
2502 		ncec_flags |= (NCE_F_BCAST|NCE_F_NONUD);
2503 		break;
2504 	case IRE_MULTICAST:
2505 		ncec_flags |= (NCE_F_MCAST|NCE_F_NONUD);
2506 		break;
2507 	}
2508 
2509 	if (ill->ill_net_type == IRE_IF_NORESOLVER && is_unicast) {
2510 		hwaddr = ill->ill_dest_addr;
2511 	} else {
2512 		hwaddr = NULL;
2513 	}
2514 	hwaddr_len = ill->ill_phys_addr_length;
2515 
2516 retry:
2517 	/* nce_state will be computed by nce_add_common() */
2518 	if (!ill->ill_isv6) {
2519 		err = nce_lookup_then_add_v4(ill, hwaddr, hwaddr_len, addr,
2520 		    ncec_flags, ND_UNCHANGED, &nce);
2521 	} else {
2522 		err = nce_lookup_then_add_v6(ill, hwaddr, hwaddr_len, addr,
2523 		    ncec_flags, ND_UNCHANGED, &nce);
2524 	}
2525 
2526 	switch (err) {
2527 	case 0:
2528 		break;
2529 	case EEXIST:
2530 		/*
2531 		 * When subnets change or partially overlap what was once
2532 		 * a broadcast address could now be a unicast, or vice versa.
2533 		 */
2534 		if (((ncec_flags ^ nce->nce_common->ncec_flags) &
2535 		    NCE_F_BCAST) != 0) {
2536 			ASSERT(!ill->ill_isv6);
2537 			ncec_delete(nce->nce_common);
2538 			nce_refrele(nce);
2539 			goto retry;
2540 		}
2541 		break;
2542 	default:
2543 		DTRACE_PROBE2(nce__init__fail, ill_t *, ill, int, err);
2544 		if (need_refrele)
2545 			ill_refrele(ill);
2546 		return (NULL);
2547 	}
2548 	/*
2549 	 * If the ill was an under-ill of an IPMP group, we need to verify
2550 	 * that it is still active so that we select an active interface in
2551 	 * the group. However, since ipmp_ill_is_active ASSERTs for
2552 	 * IS_UNDER_IPMP(), we first need to verify that the ill is an
2553 	 * under-ill, and since this is being done in the data path, the
2554 	 * only way to ascertain this is by holding the ill_g_lock.
2555 	 */
2556 	rw_enter(&ill->ill_ipst->ips_ill_g_lock, RW_READER);
2557 	mutex_enter(&ill->ill_lock);
2558 	mutex_enter(&ill->ill_phyint->phyint_lock);
2559 	if (need_refrele && IS_UNDER_IPMP(ill) && !ipmp_ill_is_active(ill)) {
2560 		/*
2561 		 * need_refrele implies that the under ill was selected by
2562 		 * ipmp_ill_hold_xmit_ill() because either the in_ill was an
2563 		 * ipmp_ill, or we are sending a non-unicast packet on an
2564 		 * under_ill. However, when we get here, the ill selected by
2565 		 * ipmp_ill_hold_xmit_ill was pulled out of the active set
2566 		 * (for unicast) or cast_ill nomination (for !unicast) after
2567 		 * it was picked as the outgoing ill.  We have to pick an
2568 		 * active interface and/or cast_ill in the group.
2569 		 */
2570 		mutex_exit(&ill->ill_phyint->phyint_lock);
2571 		nce_delete(nce);
2572 		mutex_exit(&ill->ill_lock);
2573 		rw_exit(&ill->ill_ipst->ips_ill_g_lock);
2574 		nce_refrele(nce);
2575 		ill_refrele(ill);
2576 		if ((ill = ipmp_ill_hold_xmit_ill(in_ill, is_unicast)) == NULL)
2577 			return (NULL);
2578 		goto retry;
2579 	} else {
2580 		mutex_exit(&ill->ill_phyint->phyint_lock);
2581 		mutex_exit(&ill->ill_lock);
2582 		rw_exit(&ill->ill_ipst->ips_ill_g_lock);
2583 	}
2584 done:
2585 	ASSERT(nce->nce_ill == ill);
2586 	if (need_refrele)
2587 		ill_refrele(ill);
2588 	return (nce);
2589 }
2590 
2591 nce_t *
arp_nce_init(ill_t * ill,in_addr_t addr4,int ire_type)2592 arp_nce_init(ill_t *ill, in_addr_t addr4, int ire_type)
2593 {
2594 	return (ire_nce_init(ill, &addr4, ire_type));
2595 }
2596 
2597 nce_t *
ndp_nce_init(ill_t * ill,const in6_addr_t * addr6,int ire_type)2598 ndp_nce_init(ill_t *ill, const in6_addr_t *addr6, int ire_type)
2599 {
2600 	ASSERT((ire_type & IRE_BROADCAST) == 0);
2601 	return (ire_nce_init(ill, addr6, ire_type));
2602 }
2603 
2604 /*
2605  * The caller should hold irb_lock as a writer if the ire is in a bucket.
2606  * This routine will clear ire_nce_cache, and we make sure that we can never
2607  * set ire_nce_cache after the ire is marked condemned.
2608  */
2609 void
ire_make_condemned(ire_t * ire)2610 ire_make_condemned(ire_t *ire)
2611 {
2612 	ip_stack_t	*ipst = ire->ire_ipst;
2613 	nce_t		*nce;
2614 
2615 	mutex_enter(&ire->ire_lock);
2616 	ASSERT(ire->ire_bucket == NULL ||
2617 	    RW_WRITE_HELD(&ire->ire_bucket->irb_lock));
2618 	ASSERT(!IRE_IS_CONDEMNED(ire));
2619 	ire->ire_generation = IRE_GENERATION_CONDEMNED;
2620 	/* Count how many condemned ires for kmem_cache callback */
2621 	atomic_inc_32(&ipst->ips_num_ire_condemned);
2622 	nce = ire->ire_nce_cache;
2623 	ire->ire_nce_cache = NULL;
2624 	mutex_exit(&ire->ire_lock);
2625 	if (nce != NULL)
2626 		nce_refrele(nce);
2627 }
2628 
2629 /*
2630  * Increment the generation avoiding the special condemned value
2631  */
2632 void
ire_increment_generation(ire_t * ire)2633 ire_increment_generation(ire_t *ire)
2634 {
2635 	uint_t generation;
2636 
2637 	mutex_enter(&ire->ire_lock);
2638 	/*
2639 	 * Even though the caller has a hold it can't prevent a concurrent
2640 	 * ire_delete marking the IRE condemned
2641 	 */
2642 	if (!IRE_IS_CONDEMNED(ire)) {
2643 		generation = ire->ire_generation + 1;
2644 		if (generation == IRE_GENERATION_CONDEMNED)
2645 			generation = IRE_GENERATION_INITIAL;
2646 		ASSERT(generation != IRE_GENERATION_VERIFY);
2647 		ire->ire_generation = generation;
2648 	}
2649 	mutex_exit(&ire->ire_lock);
2650 }
2651 
2652 /*
2653  * Increment ire_generation on all the IRE_MULTICASTs
2654  * Used when the default multicast interface (as determined by
2655  * ill_lookup_multicast) might have changed.
2656  *
2657  * That includes the zoneid, IFF_ flags, the IPv6 scope of the address, and
2658  * ill unplumb.
2659  */
2660 void
ire_increment_multicast_generation(ip_stack_t * ipst,boolean_t isv6)2661 ire_increment_multicast_generation(ip_stack_t *ipst, boolean_t isv6)
2662 {
2663 	ill_t	*ill;
2664 	ill_walk_context_t ctx;
2665 
2666 	rw_enter(&ipst->ips_ill_g_lock, RW_READER);
2667 	if (isv6)
2668 		ill = ILL_START_WALK_V6(&ctx, ipst);
2669 	else
2670 		ill = ILL_START_WALK_V4(&ctx, ipst);
2671 	for (; ill != NULL; ill = ill_next(&ctx, ill)) {
2672 		if (ILL_IS_CONDEMNED(ill))
2673 			continue;
2674 		if (ill->ill_ire_multicast != NULL)
2675 			ire_increment_generation(ill->ill_ire_multicast);
2676 	}
2677 	rw_exit(&ipst->ips_ill_g_lock);
2678 }
2679 
2680 /*
2681  * Return a held IRE_NOROUTE with RTF_REJECT set
2682  */
2683 ire_t *
ire_reject(ip_stack_t * ipst,boolean_t isv6)2684 ire_reject(ip_stack_t *ipst, boolean_t isv6)
2685 {
2686 	ire_t *ire;
2687 
2688 	if (isv6)
2689 		ire = ipst->ips_ire_reject_v6;
2690 	else
2691 		ire = ipst->ips_ire_reject_v4;
2692 
2693 	ASSERT(ire->ire_generation != IRE_GENERATION_CONDEMNED);
2694 	ire_refhold(ire);
2695 	return (ire);
2696 }
2697 
2698 /*
2699  * Return a held IRE_NOROUTE with RTF_BLACKHOLE set
2700  */
2701 ire_t *
ire_blackhole(ip_stack_t * ipst,boolean_t isv6)2702 ire_blackhole(ip_stack_t *ipst, boolean_t isv6)
2703 {
2704 	ire_t *ire;
2705 
2706 	if (isv6)
2707 		ire = ipst->ips_ire_blackhole_v6;
2708 	else
2709 		ire = ipst->ips_ire_blackhole_v4;
2710 
2711 	ASSERT(ire->ire_generation != IRE_GENERATION_CONDEMNED);
2712 	ire_refhold(ire);
2713 	return (ire);
2714 }
2715 
2716 /*
2717  * Return a held IRE_MULTICAST.
2718  */
2719 ire_t *
ire_multicast(ill_t * ill)2720 ire_multicast(ill_t *ill)
2721 {
2722 	ire_t *ire = ill->ill_ire_multicast;
2723 
2724 	ASSERT(ire == NULL || ire->ire_generation != IRE_GENERATION_CONDEMNED);
2725 	if (ire == NULL)
2726 		ire = ire_blackhole(ill->ill_ipst, ill->ill_isv6);
2727 	else
2728 		ire_refhold(ire);
2729 	return (ire);
2730 }
2731 
2732 /*
2733  * Given an IRE return its nexthop IRE. The nexthop IRE is an IRE_ONLINK
2734  * that is an exact match (i.e., a /32 for IPv4 and /128 for IPv6).
2735  * This can return an RTF_REJECT|RTF_BLACKHOLE.
2736  * The returned IRE is held.
2737  * The assumption is that ip_select_route() has been called and returned the
2738  * IRE (thus ip_select_route would have set up the ire_dep* information.)
2739  * If some IRE is deleteted then ire_dep_remove() will have been called and
2740  * we might not find a nexthop IRE, in which case we return NULL.
2741  */
2742 ire_t *
ire_nexthop(ire_t * ire)2743 ire_nexthop(ire_t *ire)
2744 {
2745 	ip_stack_t	*ipst = ire->ire_ipst;
2746 
2747 	/* Acquire lock to walk ire_dep_parent */
2748 	rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
2749 	while (ire != NULL) {
2750 		if (ire->ire_flags & (RTF_REJECT|RTF_BLACKHOLE)) {
2751 			goto done;
2752 		}
2753 		/*
2754 		 * If we find an IRE_ONLINK we are done. This includes
2755 		 * the case of IRE_MULTICAST.
2756 		 * Note that in order to send packets we need a host-specific
2757 		 * IRE_IF_ALL first in the ire_dep_parent chain. Normally this
2758 		 * is done by inserting an IRE_IF_CLONE if the IRE_INTERFACE
2759 		 * was not host specific.
2760 		 * However, ip_rts_request doesn't want to send packets
2761 		 * hence doesn't want to allocate an IRE_IF_CLONE. Yet
2762 		 * it needs an IRE_IF_ALL to get to the ill. Thus
2763 		 * we return IRE_IF_ALL that are not host specific here.
2764 		 */
2765 		if (ire->ire_type & IRE_ONLINK)
2766 			goto done;
2767 		ire = ire->ire_dep_parent;
2768 	}
2769 	rw_exit(&ipst->ips_ire_dep_lock);
2770 	return (NULL);
2771 
2772 done:
2773 	ire_refhold(ire);
2774 	rw_exit(&ipst->ips_ire_dep_lock);
2775 	return (ire);
2776 }
2777 
2778 /*
2779  * Find the ill used to send packets. This will be NULL in case
2780  * of a reject or blackhole.
2781  * The returned ill is held; caller needs to do ill_refrele when done.
2782  */
2783 ill_t *
ire_nexthop_ill(ire_t * ire)2784 ire_nexthop_ill(ire_t *ire)
2785 {
2786 	ill_t		*ill;
2787 
2788 	ire = ire_nexthop(ire);
2789 	if (ire == NULL)
2790 		return (NULL);
2791 
2792 	/* ire_ill can not change for an existing ire */
2793 	ill = ire->ire_ill;
2794 	if (ill != NULL)
2795 		ill_refhold(ill);
2796 	ire_refrele(ire);
2797 	return (ill);
2798 }
2799 
2800 #ifdef DEBUG
2801 static boolean_t
parent_has_child(ire_t * parent,ire_t * child)2802 parent_has_child(ire_t *parent, ire_t *child)
2803 {
2804 	ire_t	*ire;
2805 	ire_t	*prev;
2806 
2807 	ire = parent->ire_dep_children;
2808 	prev = NULL;
2809 	while (ire != NULL) {
2810 		if (prev == NULL) {
2811 			ASSERT(ire->ire_dep_sib_ptpn ==
2812 			    &(parent->ire_dep_children));
2813 		} else {
2814 			ASSERT(ire->ire_dep_sib_ptpn ==
2815 			    &(prev->ire_dep_sib_next));
2816 		}
2817 		if (ire == child)
2818 			return (B_TRUE);
2819 		prev = ire;
2820 		ire = ire->ire_dep_sib_next;
2821 	}
2822 	return (B_FALSE);
2823 }
2824 
2825 static void
ire_dep_verify(ire_t * ire)2826 ire_dep_verify(ire_t *ire)
2827 {
2828 	ire_t		*parent = ire->ire_dep_parent;
2829 	ire_t		*child = ire->ire_dep_children;
2830 
2831 	ASSERT(ire->ire_ipversion == IPV4_VERSION ||
2832 	    ire->ire_ipversion == IPV6_VERSION);
2833 	if (parent != NULL) {
2834 		ASSERT(parent->ire_ipversion == IPV4_VERSION ||
2835 		    parent->ire_ipversion == IPV6_VERSION);
2836 		ASSERT(parent->ire_refcnt >= 1);
2837 		ASSERT(parent_has_child(parent, ire));
2838 	}
2839 	if (child != NULL) {
2840 		ASSERT(child->ire_ipversion == IPV4_VERSION ||
2841 		    child->ire_ipversion == IPV6_VERSION);
2842 		ASSERT(child->ire_dep_parent == ire);
2843 		ASSERT(child->ire_dep_sib_ptpn != NULL);
2844 		ASSERT(parent_has_child(ire, child));
2845 	}
2846 }
2847 #endif /* DEBUG */
2848 
2849 /*
2850  * Assumes ire_dep_parent is set. Remove this child from its parent's linkage.
2851  */
2852 void
ire_dep_remove(ire_t * ire)2853 ire_dep_remove(ire_t *ire)
2854 {
2855 	ip_stack_t	*ipst = ire->ire_ipst;
2856 	ire_t		*parent = ire->ire_dep_parent;
2857 	ire_t		*next;
2858 	nce_t		*nce;
2859 
2860 	ASSERT(RW_WRITE_HELD(&ipst->ips_ire_dep_lock));
2861 	ASSERT(ire->ire_dep_parent != NULL);
2862 	ASSERT(ire->ire_dep_sib_ptpn != NULL);
2863 
2864 #ifdef DEBUG
2865 	ire_dep_verify(ire);
2866 	ire_dep_verify(parent);
2867 #endif
2868 
2869 	next = ire->ire_dep_sib_next;
2870 	if (next != NULL)
2871 		next->ire_dep_sib_ptpn = ire->ire_dep_sib_ptpn;
2872 
2873 	ASSERT(*(ire->ire_dep_sib_ptpn) == ire);
2874 	*(ire->ire_dep_sib_ptpn) = ire->ire_dep_sib_next;
2875 
2876 	ire->ire_dep_sib_ptpn = NULL;
2877 	ire->ire_dep_sib_next = NULL;
2878 
2879 	mutex_enter(&ire->ire_lock);
2880 	parent = ire->ire_dep_parent;
2881 	ire->ire_dep_parent = NULL;
2882 	mutex_exit(&ire->ire_lock);
2883 
2884 	/*
2885 	 * Make sure all our children, grandchildren, etc set
2886 	 * ire_dep_parent_generation to IRE_GENERATION_VERIFY since
2887 	 * we can no longer guarantee than the children have a current
2888 	 * ire_nce_cache and ire_nexthop_ill().
2889 	 */
2890 	if (ire->ire_dep_children != NULL)
2891 		ire_dep_invalidate_children(ire->ire_dep_children);
2892 
2893 	/*
2894 	 * Since the parent is gone we make sure we clear ire_nce_cache.
2895 	 * We can clear it under ire_lock even if the IRE is used
2896 	 */
2897 	mutex_enter(&ire->ire_lock);
2898 	nce = ire->ire_nce_cache;
2899 	ire->ire_nce_cache = NULL;
2900 	mutex_exit(&ire->ire_lock);
2901 	if (nce != NULL)
2902 		nce_refrele(nce);
2903 
2904 #ifdef DEBUG
2905 	ire_dep_verify(ire);
2906 	ire_dep_verify(parent);
2907 #endif
2908 
2909 	ire_refrele_notr(parent);
2910 	ire_refrele_notr(ire);
2911 }
2912 
2913 /*
2914  * Insert the child in the linkage of the parent
2915  */
2916 static void
ire_dep_parent_insert(ire_t * child,ire_t * parent)2917 ire_dep_parent_insert(ire_t *child, ire_t *parent)
2918 {
2919 	ip_stack_t	*ipst = child->ire_ipst;
2920 	ire_t		*next;
2921 
2922 	ASSERT(RW_WRITE_HELD(&ipst->ips_ire_dep_lock));
2923 	ASSERT(child->ire_dep_parent == NULL);
2924 
2925 #ifdef DEBUG
2926 	ire_dep_verify(child);
2927 	ire_dep_verify(parent);
2928 #endif
2929 	/* No parents => no siblings */
2930 	ASSERT(child->ire_dep_sib_ptpn == NULL);
2931 	ASSERT(child->ire_dep_sib_next == NULL);
2932 
2933 	ire_refhold_notr(parent);
2934 	ire_refhold_notr(child);
2935 
2936 	/* Head insertion */
2937 	next = parent->ire_dep_children;
2938 	if (next != NULL) {
2939 		ASSERT(next->ire_dep_sib_ptpn == &(parent->ire_dep_children));
2940 		child->ire_dep_sib_next = next;
2941 		next->ire_dep_sib_ptpn = &(child->ire_dep_sib_next);
2942 	}
2943 	parent->ire_dep_children = child;
2944 	child->ire_dep_sib_ptpn = &(parent->ire_dep_children);
2945 
2946 	mutex_enter(&child->ire_lock);
2947 	child->ire_dep_parent = parent;
2948 	mutex_exit(&child->ire_lock);
2949 
2950 #ifdef DEBUG
2951 	ire_dep_verify(child);
2952 	ire_dep_verify(parent);
2953 #endif
2954 }
2955 
2956 
2957 /*
2958  * Given count worth of ires and generations, build ire_dep_* relationships
2959  * from ires[0] to ires[count-1]. Record generations[i+1] in
2960  * ire_dep_parent_generation for ires[i].
2961  * We graft onto an existing parent chain by making sure that we don't
2962  * touch ire_dep_parent for ires[count-1].
2963  *
2964  * We check for any condemned ire_generation count and return B_FALSE in
2965  * that case so that the caller can tear it apart.
2966  *
2967  * Note that generations[0] is not used. Caller handles that.
2968  */
2969 boolean_t
ire_dep_build(ire_t * ires[],uint_t generations[],uint_t count)2970 ire_dep_build(ire_t *ires[], uint_t generations[], uint_t count)
2971 {
2972 	ire_t		*ire = ires[0];
2973 	ip_stack_t	*ipst;
2974 	uint_t		i;
2975 
2976 	ASSERT(count > 0);
2977 	if (count == 1) {
2978 		/* No work to do */
2979 		return (B_TRUE);
2980 	}
2981 	ipst = ire->ire_ipst;
2982 	rw_enter(&ipst->ips_ire_dep_lock, RW_WRITER);
2983 	/*
2984 	 * Do not remove the linkage for any existing parent chain i.e.,
2985 	 * ires[count-1] is left alone.
2986 	 */
2987 	for (i = 0; i < count-1; i++) {
2988 		/* Remove existing parent if we need to change it */
2989 		if (ires[i]->ire_dep_parent != NULL &&
2990 		    ires[i]->ire_dep_parent != ires[i+1])
2991 			ire_dep_remove(ires[i]);
2992 	}
2993 
2994 	for (i = 0; i < count - 1; i++) {
2995 		ASSERT(ires[i]->ire_ipversion == IPV4_VERSION ||
2996 		    ires[i]->ire_ipversion == IPV6_VERSION);
2997 		/* Does it need to change? */
2998 		if (ires[i]->ire_dep_parent != ires[i+1])
2999 			ire_dep_parent_insert(ires[i], ires[i+1]);
3000 
3001 		mutex_enter(&ires[i+1]->ire_lock);
3002 		if (IRE_IS_CONDEMNED(ires[i+1])) {
3003 			mutex_exit(&ires[i+1]->ire_lock);
3004 			rw_exit(&ipst->ips_ire_dep_lock);
3005 			return (B_FALSE);
3006 		}
3007 		mutex_exit(&ires[i+1]->ire_lock);
3008 
3009 		mutex_enter(&ires[i]->ire_lock);
3010 		ires[i]->ire_dep_parent_generation = generations[i+1];
3011 		mutex_exit(&ires[i]->ire_lock);
3012 	}
3013 	rw_exit(&ipst->ips_ire_dep_lock);
3014 	return (B_TRUE);
3015 }
3016 
3017 /*
3018  * Given count worth of ires, unbuild ire_dep_* relationships
3019  * from ires[0] to ires[count-1].
3020  */
3021 void
ire_dep_unbuild(ire_t * ires[],uint_t count)3022 ire_dep_unbuild(ire_t *ires[], uint_t count)
3023 {
3024 	ip_stack_t	*ipst;
3025 	uint_t		i;
3026 
3027 	if (count == 0) {
3028 		/* No work to do */
3029 		return;
3030 	}
3031 	ipst = ires[0]->ire_ipst;
3032 	rw_enter(&ipst->ips_ire_dep_lock, RW_WRITER);
3033 	for (i = 0; i < count; i++) {
3034 		ASSERT(ires[i]->ire_ipversion == IPV4_VERSION ||
3035 		    ires[i]->ire_ipversion == IPV6_VERSION);
3036 		if (ires[i]->ire_dep_parent != NULL)
3037 			ire_dep_remove(ires[i]);
3038 		mutex_enter(&ires[i]->ire_lock);
3039 		ires[i]->ire_dep_parent_generation = IRE_GENERATION_VERIFY;
3040 		mutex_exit(&ires[i]->ire_lock);
3041 	}
3042 	rw_exit(&ipst->ips_ire_dep_lock);
3043 }
3044 
3045 /*
3046  * Both the forwarding and the outbound code paths can trip on
3047  * a condemned NCE, in which case we call this function.
3048  * We have two different behaviors: if the NCE was UNREACHABLE
3049  * it is an indication that something failed. In that case
3050  * we see if we should look for a different IRE (for example,
3051  * delete any matching redirect IRE, or try a different
3052  * IRE_DEFAULT (ECMP)). We mark the ire as bad so a hopefully
3053  * different IRE will be picked next time we send/forward.
3054  *
3055  * If we are called by the output path then fail_if_better is set
3056  * and we return NULL if there could be a better IRE. This is because the
3057  * output path retries the IRE lookup. (The input/forward path can not retry.)
3058  *
3059  * If the NCE was not unreachable then we pick/allocate a
3060  * new (most likely ND_INITIAL) NCE and proceed with it.
3061  *
3062  * ipha/ip6h are needed for multicast packets; ipha needs to be
3063  * set for IPv4 and ip6h needs to be set for IPv6 packets.
3064  */
3065 nce_t *
ire_handle_condemned_nce(nce_t * nce,ire_t * ire,ipha_t * ipha,ip6_t * ip6h,boolean_t fail_if_better)3066 ire_handle_condemned_nce(nce_t *nce, ire_t *ire, ipha_t *ipha, ip6_t *ip6h,
3067     boolean_t fail_if_better)
3068 {
3069 	if (nce->nce_common->ncec_state == ND_UNREACHABLE) {
3070 		if (ire_no_good(ire) && fail_if_better) {
3071 			/*
3072 			 * Did some changes, or ECMP likely to exist.
3073 			 * Make ip_output look for a different IRE
3074 			 */
3075 			return (NULL);
3076 		}
3077 	}
3078 	if (ire_revalidate_nce(ire) == ENETUNREACH) {
3079 		/* The ire_dep_parent chain went bad, or no memory? */
3080 		(void) ire_no_good(ire);
3081 		return (NULL);
3082 	}
3083 	if (ire->ire_ipversion == IPV4_VERSION) {
3084 		ASSERT(ipha != NULL);
3085 		nce = ire_to_nce(ire, ipha->ipha_dst, NULL);
3086 	} else {
3087 		ASSERT(ip6h != NULL);
3088 		nce = ire_to_nce(ire, INADDR_ANY, &ip6h->ip6_dst);
3089 	}
3090 
3091 	if (nce == NULL)
3092 		return (NULL);
3093 	if (nce->nce_is_condemned) {
3094 		nce_refrele(nce);
3095 		return (NULL);
3096 	}
3097 	return (nce);
3098 }
3099 
3100 /*
3101  * The caller has found that the ire is bad, either due to a reference to an NCE
3102  * in ND_UNREACHABLE state, or a MULTIRT route whose gateway can't be resolved.
3103  * We update things so a subsequent attempt to send to the destination
3104  * is likely to find different IRE, or that a new NCE would be created.
3105  *
3106  * Returns B_TRUE if it is likely that a subsequent ire_ftable_lookup would
3107  * find a different route (either due to having deleted a redirect, or there
3108  * being ECMP routes.)
3109  *
3110  * If we have a redirect (RTF_DYNAMIC) we delete it.
3111  * Otherwise we increment ire_badcnt and increment the generation number so
3112  * that a cached ixa_ire will redo the route selection. ire_badcnt is taken
3113  * into account in the route selection when we have multiple choices (multiple
3114  * default routes or ECMP in general).
3115  * Any time ip_select_route find an ire with a condemned ire_nce_cache
3116  * (e.g., if no equal cost route to the bad one) ip_select_route will make
3117  * sure the NCE is revalidated to avoid getting stuck on a
3118  * NCE_F_CONDMNED ncec that caused ire_no_good to be called.
3119  */
3120 boolean_t
ire_no_good(ire_t * ire)3121 ire_no_good(ire_t *ire)
3122 {
3123 	ip_stack_t	*ipst = ire->ire_ipst;
3124 	ire_t		*ire2;
3125 	nce_t		*nce;
3126 
3127 	if (ire->ire_flags & RTF_DYNAMIC) {
3128 		ire_delete(ire);
3129 		return (B_TRUE);
3130 	}
3131 	if (ire->ire_flags & RTF_INDIRECT) {
3132 		/* Check if next IRE is a redirect */
3133 		rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
3134 		if (ire->ire_dep_parent != NULL &&
3135 		    (ire->ire_dep_parent->ire_flags & RTF_DYNAMIC)) {
3136 			ire2 = ire->ire_dep_parent;
3137 			ire_refhold(ire2);
3138 		} else {
3139 			ire2 = NULL;
3140 		}
3141 		rw_exit(&ipst->ips_ire_dep_lock);
3142 		if (ire2 != NULL) {
3143 			ire_delete(ire2);
3144 			ire_refrele(ire2);
3145 			return (B_TRUE);
3146 		}
3147 	}
3148 	/*
3149 	 * No redirect involved. Increment badcnt so that if we have ECMP
3150 	 * routes we are likely to pick a different one for the next packet.
3151 	 *
3152 	 * If the NCE is unreachable and condemned we should drop the reference
3153 	 * to it so that a new NCE can be created.
3154 	 *
3155 	 * Finally we increment the generation number so that any ixa_ire
3156 	 * cache will be revalidated.
3157 	 */
3158 	mutex_enter(&ire->ire_lock);
3159 	ire->ire_badcnt++;
3160 	ire->ire_last_badcnt = TICK_TO_SEC(ddi_get_lbolt64());
3161 	nce = ire->ire_nce_cache;
3162 	if (nce != NULL && nce->nce_is_condemned &&
3163 	    nce->nce_common->ncec_state == ND_UNREACHABLE)
3164 		ire->ire_nce_cache = NULL;
3165 	else
3166 		nce = NULL;
3167 	mutex_exit(&ire->ire_lock);
3168 	if (nce != NULL)
3169 		nce_refrele(nce);
3170 
3171 	ire_increment_generation(ire);
3172 	ire_dep_incr_generation(ire);
3173 
3174 	return (ire->ire_bucket->irb_ire_cnt > 1);
3175 }
3176 
3177 /*
3178  * Walk ire_dep_parent chain and validate that ire_dep_parent->ire_generation ==
3179  * ire_dep_parent_generation.
3180  * If they all match we just return ire_generation from the topmost IRE.
3181  * Otherwise we propagate the mismatch by setting all ire_dep_parent_generation
3182  * above the mismatch to IRE_GENERATION_VERIFY and also returning
3183  * IRE_GENERATION_VERIFY.
3184  */
3185 uint_t
ire_dep_validate_generations(ire_t * ire)3186 ire_dep_validate_generations(ire_t *ire)
3187 {
3188 	ip_stack_t	*ipst = ire->ire_ipst;
3189 	uint_t		generation;
3190 	ire_t		*ire1;
3191 
3192 	rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
3193 	generation = ire->ire_generation;	/* Assuming things match */
3194 	for (ire1 = ire; ire1 != NULL; ire1 = ire1->ire_dep_parent) {
3195 		ASSERT(ire1->ire_ipversion == IPV4_VERSION ||
3196 		    ire1->ire_ipversion == IPV6_VERSION);
3197 		if (ire1->ire_dep_parent == NULL)
3198 			break;
3199 		if (ire1->ire_dep_parent_generation !=
3200 		    ire1->ire_dep_parent->ire_generation)
3201 			goto mismatch;
3202 	}
3203 	rw_exit(&ipst->ips_ire_dep_lock);
3204 	return (generation);
3205 
3206 mismatch:
3207 	generation = IRE_GENERATION_VERIFY;
3208 	/* Fill from top down to the mismatch with _VERIFY */
3209 	while (ire != ire1) {
3210 		ASSERT(ire->ire_ipversion == IPV4_VERSION ||
3211 		    ire->ire_ipversion == IPV6_VERSION);
3212 		mutex_enter(&ire->ire_lock);
3213 		ire->ire_dep_parent_generation = IRE_GENERATION_VERIFY;
3214 		mutex_exit(&ire->ire_lock);
3215 		ire = ire->ire_dep_parent;
3216 	}
3217 	rw_exit(&ipst->ips_ire_dep_lock);
3218 	return (generation);
3219 }
3220 
3221 /*
3222  * Used when we need to return an ire with ire_dep_parent, but we
3223  * know the chain is invalid for instance we didn't create an IRE_IF_CLONE
3224  * Using IRE_GENERATION_VERIFY means that next time we'll redo the
3225  * recursive lookup.
3226  */
3227 void
ire_dep_invalidate_generations(ire_t * ire)3228 ire_dep_invalidate_generations(ire_t *ire)
3229 {
3230 	ip_stack_t	*ipst = ire->ire_ipst;
3231 
3232 	rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
3233 	while (ire != NULL) {
3234 		ASSERT(ire->ire_ipversion == IPV4_VERSION ||
3235 		    ire->ire_ipversion == IPV6_VERSION);
3236 		mutex_enter(&ire->ire_lock);
3237 		ire->ire_dep_parent_generation = IRE_GENERATION_VERIFY;
3238 		mutex_exit(&ire->ire_lock);
3239 		ire = ire->ire_dep_parent;
3240 	}
3241 	rw_exit(&ipst->ips_ire_dep_lock);
3242 }
3243 
3244 /* Set _VERIFY ire_dep_parent_generation for all children recursively */
3245 static void
ire_dep_invalidate_children(ire_t * child)3246 ire_dep_invalidate_children(ire_t *child)
3247 {
3248 	ip_stack_t	*ipst = child->ire_ipst;
3249 
3250 	ASSERT(RW_WRITE_HELD(&ipst->ips_ire_dep_lock));
3251 	/* Depth first */
3252 	if (child->ire_dep_children != NULL)
3253 		ire_dep_invalidate_children(child->ire_dep_children);
3254 
3255 	while (child != NULL) {
3256 		mutex_enter(&child->ire_lock);
3257 		child->ire_dep_parent_generation = IRE_GENERATION_VERIFY;
3258 		mutex_exit(&child->ire_lock);
3259 		child = child->ire_dep_sib_next;
3260 	}
3261 }
3262 
3263 static void
ire_dep_increment_children(ire_t * child)3264 ire_dep_increment_children(ire_t *child)
3265 {
3266 	ip_stack_t	*ipst = child->ire_ipst;
3267 
3268 	ASSERT(RW_READ_HELD(&ipst->ips_ire_dep_lock));
3269 	/* Depth first */
3270 	if (child->ire_dep_children != NULL)
3271 		ire_dep_increment_children(child->ire_dep_children);
3272 
3273 	while (child != NULL) {
3274 		if (!IRE_IS_CONDEMNED(child))
3275 			ire_increment_generation(child);
3276 		child = child->ire_dep_sib_next;
3277 	}
3278 }
3279 
3280 /*
3281  * Walk all the children of this ire recursively and increment their
3282  * generation number.
3283  */
3284 static void
ire_dep_incr_generation_locked(ire_t * parent)3285 ire_dep_incr_generation_locked(ire_t *parent)
3286 {
3287 	ASSERT(RW_READ_HELD(&parent->ire_ipst->ips_ire_dep_lock));
3288 	if (parent->ire_dep_children != NULL)
3289 		ire_dep_increment_children(parent->ire_dep_children);
3290 }
3291 
3292 void
ire_dep_incr_generation(ire_t * parent)3293 ire_dep_incr_generation(ire_t *parent)
3294 {
3295 	ip_stack_t	*ipst = parent->ire_ipst;
3296 
3297 	rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
3298 	ire_dep_incr_generation_locked(parent);
3299 	rw_exit(&ipst->ips_ire_dep_lock);
3300 }
3301 
3302 /*
3303  * Get a new ire_nce_cache for this IRE as well as its nexthop.
3304  * Returns zero if it succeeds. Can fail due to lack of memory or when
3305  * the route has become unreachable. Returns ENOMEM and ENETUNREACH in those
3306  * cases.
3307  *
3308  * In the in.mpathd case, the ire will have ire_testhidden
3309  * set; so we should create the ncec for the underlying ill.
3310  *
3311  * Note that the error returned by ire_revalidate_nce() is ignored by most
3312  * callers except ire_handle_condemned_nce(), which handles the ENETUNREACH
3313  * error to mark potentially bad ire's. For all the other callers, an
3314  * error return could indicate a transient condition like ENOMEM, or could
3315  * be the result of an interface that is going down/unplumbing. In the former
3316  * case (transient error), we would leave the old stale ire/ire_nce_cache
3317  * in place, and possibly use incorrect link-layer information to send packets
3318  * but would eventually recover. In the latter case (ill down/replumb),
3319  * ire_revalidate_nce() might return a condemned nce back, but we would then
3320  * recover in the packet output path.
3321  */
3322 int
ire_revalidate_nce(ire_t * ire)3323 ire_revalidate_nce(ire_t *ire)
3324 {
3325 	nce_t		*nce, *old_nce;
3326 	ire_t		*nexthop;
3327 
3328 	/*
3329 	 * For multicast we conceptually have an NCE but we don't store it
3330 	 * in ire_nce_cache; when ire_to_nce is called we allocate the nce.
3331 	 */
3332 	if (ire->ire_type & IRE_MULTICAST)
3333 		return (0);
3334 
3335 	/* ire_testhidden should only be set on under-interfaces */
3336 	ASSERT(!ire->ire_testhidden || !IS_IPMP(ire->ire_ill));
3337 
3338 	nexthop = ire_nexthop(ire);
3339 	if (nexthop == NULL) {
3340 		/* The route is potentially bad */
3341 		(void) ire_no_good(ire);
3342 		return (ENETUNREACH);
3343 	}
3344 	if (ire->ire_type & (IRE_LOCAL|IRE_LOOPBACK)) {
3345 		ASSERT(ire->ire_ill != NULL);
3346 
3347 		if (ire->ire_ipversion == IPV4_VERSION)
3348 			nce = nce_lookup_v4(ire->ire_ill, &ire->ire_addr);
3349 		else
3350 			nce = nce_lookup_v6(ire->ire_ill, &ire->ire_addr_v6);
3351 	} else {
3352 		ASSERT(nexthop->ire_type & IRE_ONLINK);
3353 		if (ire->ire_ipversion == IPV4_VERSION) {
3354 			nce = arp_nce_init(nexthop->ire_ill, nexthop->ire_addr,
3355 			    nexthop->ire_type);
3356 		} else {
3357 			nce = ndp_nce_init(nexthop->ire_ill,
3358 			    &nexthop->ire_addr_v6, nexthop->ire_type);
3359 		}
3360 	}
3361 	if (nce == NULL) {
3362 		/*
3363 		 * Leave the old stale one in place to avoid a NULL
3364 		 * ire_nce_cache.
3365 		 */
3366 		ire_refrele(nexthop);
3367 		return (ENOMEM);
3368 	}
3369 
3370 	if (nexthop != ire) {
3371 		/* Update the nexthop ire */
3372 		mutex_enter(&nexthop->ire_lock);
3373 		old_nce = nexthop->ire_nce_cache;
3374 		if (!IRE_IS_CONDEMNED(nexthop)) {
3375 			nce_refhold(nce);
3376 			nexthop->ire_nce_cache = nce;
3377 		} else {
3378 			nexthop->ire_nce_cache = NULL;
3379 		}
3380 		mutex_exit(&nexthop->ire_lock);
3381 		if (old_nce != NULL)
3382 			nce_refrele(old_nce);
3383 	}
3384 	ire_refrele(nexthop);
3385 
3386 	mutex_enter(&ire->ire_lock);
3387 	old_nce = ire->ire_nce_cache;
3388 	if (!IRE_IS_CONDEMNED(ire)) {
3389 		nce_refhold(nce);
3390 		ire->ire_nce_cache = nce;
3391 	} else {
3392 		ire->ire_nce_cache = NULL;
3393 	}
3394 	mutex_exit(&ire->ire_lock);
3395 	if (old_nce != NULL)
3396 		nce_refrele(old_nce);
3397 
3398 	nce_refrele(nce);
3399 	return (0);
3400 }
3401 
3402 /*
3403  * Get a held nce for a given ire.
3404  * In the common case this is just from ire_nce_cache.
3405  * For IRE_MULTICAST this needs to do an explicit lookup since we do not
3406  * have an IRE_MULTICAST per address.
3407  * Note that this explicitly returns CONDEMNED NCEs. The caller needs those
3408  * so they can check whether the NCE went unreachable (as opposed to was
3409  * condemned for some other reason).
3410  */
3411 nce_t *
ire_to_nce(ire_t * ire,ipaddr_t v4nexthop,const in6_addr_t * v6nexthop)3412 ire_to_nce(ire_t *ire, ipaddr_t v4nexthop, const in6_addr_t *v6nexthop)
3413 {
3414 	nce_t	*nce;
3415 
3416 	if (ire->ire_flags & (RTF_REJECT|RTF_BLACKHOLE))
3417 		return (NULL);
3418 
3419 	/* ire_testhidden should only be set on under-interfaces */
3420 	ASSERT(!ire->ire_testhidden || !IS_IPMP(ire->ire_ill));
3421 
3422 	mutex_enter(&ire->ire_lock);
3423 	nce = ire->ire_nce_cache;
3424 	if (nce != NULL) {
3425 		nce_refhold(nce);
3426 		mutex_exit(&ire->ire_lock);
3427 		return (nce);
3428 	}
3429 	mutex_exit(&ire->ire_lock);
3430 
3431 	if (ire->ire_type & IRE_MULTICAST) {
3432 		ASSERT(ire->ire_ill != NULL);
3433 
3434 		if (ire->ire_ipversion == IPV4_VERSION) {
3435 			ASSERT(v6nexthop == NULL);
3436 
3437 			nce = arp_nce_init(ire->ire_ill, v4nexthop,
3438 			    ire->ire_type);
3439 		} else {
3440 			ASSERT(v6nexthop != NULL);
3441 			ASSERT(v4nexthop == 0);
3442 			nce = ndp_nce_init(ire->ire_ill, v6nexthop,
3443 			    ire->ire_type);
3444 		}
3445 		return (nce);
3446 	}
3447 	return (NULL);
3448 }
3449 
3450 nce_t *
ire_to_nce_pkt(ire_t * ire,mblk_t * mp)3451 ire_to_nce_pkt(ire_t *ire, mblk_t *mp)
3452 {
3453 	ipha_t		*ipha;
3454 	ip6_t		*ip6h;
3455 
3456 	if (IPH_HDR_VERSION(mp->b_rptr) == IPV4_VERSION) {
3457 		ipha = (ipha_t *)mp->b_rptr;
3458 		return (ire_to_nce(ire, ipha->ipha_dst, NULL));
3459 	} else {
3460 		ip6h = (ip6_t *)mp->b_rptr;
3461 		return (ire_to_nce(ire, INADDR_ANY, &ip6h->ip6_dst));
3462 	}
3463 }
3464 
3465 /*
3466  * Given an IRE_INTERFACE (that matches more than one address) create
3467  * and return an IRE_IF_CLONE for the specific address.
3468  * Return the generation number.
3469  * Returns NULL is no memory for the IRE.
3470  * Handles both IPv4 and IPv6.
3471  *
3472  * IRE_IF_CLONE entries may only be created adn added by calling
3473  * ire_create_if_clone(), and we depend on the fact that ire_add will
3474  * atomically ensure that attempts to add multiple identical IRE_IF_CLONE
3475  * entries will not result in duplicate (i.e., ire_identical_ref > 1)
3476  * CLONE entries, so that a single ire_delete is sufficient to remove the
3477  * CLONE.
3478  */
3479 ire_t *
ire_create_if_clone(ire_t * ire_if,const in6_addr_t * addr,uint_t * generationp)3480 ire_create_if_clone(ire_t *ire_if, const in6_addr_t *addr, uint_t *generationp)
3481 {
3482 	ire_t		*ire;
3483 	ire_t		*nire;
3484 
3485 	if (ire_if->ire_ipversion == IPV4_VERSION) {
3486 		ipaddr_t	v4addr;
3487 		ipaddr_t	mask = IP_HOST_MASK;
3488 
3489 		ASSERT(IN6_IS_ADDR_V4MAPPED(addr));
3490 		IN6_V4MAPPED_TO_IPADDR(addr, v4addr);
3491 
3492 		ire = ire_create(
3493 		    (uchar_t *)&v4addr,			/* dest address */
3494 		    (uchar_t *)&mask,			/* mask */
3495 		    (uchar_t *)&ire_if->ire_gateway_addr,
3496 		    IRE_IF_CLONE,			/* IRE type */
3497 		    ire_if->ire_ill,
3498 		    ire_if->ire_zoneid,
3499 		    ire_if->ire_flags | RTF_HOST,
3500 		    NULL,		/* No security attr for IRE_IF_ALL */
3501 		    ire_if->ire_ipst);
3502 	} else {
3503 		ASSERT(!IN6_IS_ADDR_V4MAPPED(addr));
3504 		ire = ire_create_v6(
3505 		    addr,				/* dest address */
3506 		    &ipv6_all_ones,			/* mask */
3507 		    &ire_if->ire_gateway_addr_v6,	/* gateway addr */
3508 		    IRE_IF_CLONE,			/* IRE type */
3509 		    ire_if->ire_ill,
3510 		    ire_if->ire_zoneid,
3511 		    ire_if->ire_flags | RTF_HOST,
3512 		    NULL,		/* No security attr for IRE_IF_ALL */
3513 		    ire_if->ire_ipst);
3514 	}
3515 	if (ire == NULL)
3516 		return (NULL);
3517 
3518 	/* Take the metrics, in particular the mtu, from the IRE_IF */
3519 	ire->ire_metrics = ire_if->ire_metrics;
3520 
3521 	nire = ire_add(ire);
3522 	if (nire == NULL) /* Some failure */
3523 		return (NULL);
3524 
3525 	if (generationp != NULL)
3526 		*generationp = nire->ire_generation;
3527 
3528 	return (nire);
3529 }
3530 
3531 /*
3532  * The argument is an IRE_INTERFACE. Delete all of IRE_IF_CLONE in the
3533  * ire_dep_children (just walk the ire_dep_sib_next since they are all
3534  * immediate children.)
3535  * Since we hold a lock while we remove them we need to defer the actual
3536  * calls to ire_delete() until we have dropped the lock. This makes things
3537  * less efficient since we restart at the top after dropping the lock. But
3538  * we only run when an IRE_INTERFACE is deleted which is infrquent.
3539  *
3540  * Note that ire_dep_children can be any mixture of offlink routes and
3541  * IRE_IF_CLONE entries.
3542  */
3543 void
ire_dep_delete_if_clone(ire_t * parent)3544 ire_dep_delete_if_clone(ire_t *parent)
3545 {
3546 	ip_stack_t	*ipst = parent->ire_ipst;
3547 	ire_t		*child, *next;
3548 
3549 restart:
3550 	rw_enter(&ipst->ips_ire_dep_lock, RW_READER);
3551 	if (parent->ire_dep_children == NULL) {
3552 		rw_exit(&ipst->ips_ire_dep_lock);
3553 		return;
3554 	}
3555 	child = parent->ire_dep_children;
3556 	while (child != NULL) {
3557 		next = child->ire_dep_sib_next;
3558 		if ((child->ire_type & IRE_IF_CLONE) &&
3559 		    !IRE_IS_CONDEMNED(child)) {
3560 			ire_refhold(child);
3561 			rw_exit(&ipst->ips_ire_dep_lock);
3562 			ire_delete(child);
3563 			ASSERT(IRE_IS_CONDEMNED(child));
3564 			ire_refrele(child);
3565 			goto restart;
3566 		}
3567 		child = next;
3568 	}
3569 	rw_exit(&ipst->ips_ire_dep_lock);
3570 }
3571 
3572 /*
3573  * In the preferred/strict src multihoming modes, unbound routes (i.e.,
3574  * ire_t entries with ire_unbound set to B_TRUE) are bound to an interface
3575  * by selecting the first available interface that has an interface route for
3576  * the ire_gateway. If that interface is subsequently brought down, ill_downi()
3577  * will call ire_rebind() so that the unbound route can be bound to some other
3578  * matching interface thereby preserving the intended reachability information
3579  * from the original unbound route.
3580  */
3581 void
ire_rebind(ire_t * ire)3582 ire_rebind(ire_t *ire)
3583 {
3584 	ire_t	*gw_ire, *new_ire;
3585 	int	match_flags = MATCH_IRE_TYPE;
3586 	ill_t	*gw_ill;
3587 	boolean_t isv6 = (ire->ire_ipversion == IPV6_VERSION);
3588 	ip_stack_t *ipst = ire->ire_ipst;
3589 
3590 	ASSERT(ire->ire_unbound);
3591 again:
3592 	if (isv6) {
3593 		gw_ire = ire_ftable_lookup_v6(&ire->ire_gateway_addr_v6, 0, 0,
3594 		    IRE_INTERFACE, NULL, ALL_ZONES, NULL, match_flags, 0,
3595 		    ipst, NULL);
3596 	} else {
3597 		gw_ire = ire_ftable_lookup_v4(ire->ire_gateway_addr, 0, 0,
3598 		    IRE_INTERFACE, NULL, ALL_ZONES, NULL, match_flags, 0,
3599 		    ipst, NULL);
3600 	}
3601 	if (gw_ire == NULL) {
3602 		/* see comments in ip_rt_add[_v6]() for IPMP */
3603 		if (match_flags & MATCH_IRE_TESTHIDDEN)
3604 			return;
3605 
3606 		match_flags |= MATCH_IRE_TESTHIDDEN;
3607 		goto again;
3608 	}
3609 	gw_ill = gw_ire->ire_ill;
3610 	if (isv6) {
3611 		new_ire = ire_create_v6(&ire->ire_addr_v6, &ire->ire_mask_v6,
3612 		    &ire->ire_gateway_addr_v6, ire->ire_type, gw_ill,
3613 		    ire->ire_zoneid, ire->ire_flags, NULL, ipst);
3614 	} else {
3615 		new_ire = ire_create((uchar_t *)&ire->ire_addr,
3616 		    (uchar_t *)&ire->ire_mask,
3617 		    (uchar_t *)&ire->ire_gateway_addr, ire->ire_type, gw_ill,
3618 		    ire->ire_zoneid, ire->ire_flags, NULL, ipst);
3619 	}
3620 	ire_refrele(gw_ire);
3621 	if (new_ire == NULL)
3622 		return;
3623 	new_ire->ire_unbound = B_TRUE;
3624 	new_ire = ire_add(new_ire);
3625 	if (new_ire != NULL)
3626 		ire_refrele(new_ire);
3627 }
3628