usr/src/uts/common/inet/ip/spd.c

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */
/*
 * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

/*
 * IPsec Security Policy Database.
 *
 * This module maintains the SPD and provides routines used by ip and ip6
 * to apply IPsec policy to inbound and outbound datagrams.
 */

#include <sys/types.h>
#include <sys/stream.h>
#include <sys/stropts.h>
#include <sys/sysmacros.h>
#include <sys/strsubr.h>
#include <sys/strsun.h>
#include <sys/strlog.h>
#include <sys/strsun.h>
#include <sys/cmn_err.h>
#include <sys/zone.h>

#include <sys/systm.h>
#include <sys/param.h>
#include <sys/kmem.h>
#include <sys/ddi.h>

#include <sys/crypto/api.h>

#include <inet/common.h>
#include <inet/mi.h>

#include <netinet/ip6.h>
#include <netinet/icmp6.h>
#include <netinet/udp.h>

#include <inet/ip.h>
#include <inet/ip6.h>

#include <net/pfkeyv2.h>
#include <net/pfpolicy.h>
#include <inet/sadb.h>
#include <inet/ipsec_impl.h>

#include <inet/ip_impl.h>	/* For IP_MOD_ID */

#include <inet/ipsecah.h>
#include <inet/ipsecesp.h>
#include <inet/ipdrop.h>
#include <inet/ipclassifier.h>
#include <inet/iptun.h>
#include <inet/iptun/iptun_impl.h>

static void ipsec_update_present_flags(ipsec_stack_t *);
static ipsec_act_t *ipsec_act_wildcard_expand(ipsec_act_t *, uint_t *,
    netstack_t *);
static mblk_t *ipsec_check_ipsecin_policy(mblk_t *, ipsec_policy_t *,
    ipha_t *, ip6_t *, uint64_t, ip_recv_attr_t *, netstack_t *);
static void ipsec_action_free_table(ipsec_action_t *);
static void ipsec_action_reclaim(void *);
static void ipsec_action_reclaim_stack(netstack_t *);
static void ipsid_init(netstack_t *);
static void ipsid_fini(netstack_t *);

/* sel_flags values for ipsec_init_inbound_sel(). */
#define	SEL_NONE	0x0000
#define	SEL_PORT_POLICY	0x0001
#define	SEL_IS_ICMP	0x0002
#define	SEL_TUNNEL_MODE	0x0004
#define	SEL_POST_FRAG	0x0008

/* Return values for ipsec_init_inbound_sel(). */
typedef enum { SELRET_NOMEM, SELRET_BADPKT, SELRET_SUCCESS, SELRET_TUNFRAG}
    selret_t;

static selret_t ipsec_init_inbound_sel(ipsec_selector_t *, mblk_t *,
    ipha_t *, ip6_t *, uint8_t);

static boolean_t ipsec_check_ipsecin_action(ip_recv_attr_t *, mblk_t *,
    struct ipsec_action_s *, ipha_t *ipha, ip6_t *ip6h, const char **,
    kstat_named_t **, netstack_t *);
static void ipsec_unregister_prov_update(void);
static void ipsec_prov_update_callback_stack(uint32_t, void *, netstack_t *);
static boolean_t ipsec_compare_action(ipsec_policy_t *, ipsec_policy_t *);
static uint32_t selector_hash(ipsec_selector_t *, ipsec_policy_root_t *);
static boolean_t ipsec_kstat_init(ipsec_stack_t *);
static void ipsec_kstat_destroy(ipsec_stack_t *);
static int ipsec_free_tables(ipsec_stack_t *);
static int tunnel_compare(const void *, const void *);
static void ipsec_freemsg_chain(mblk_t *);
static void ip_drop_packet_chain(mblk_t *, boolean_t, ill_t *,
    struct kstat_named *, ipdropper_t *);
static boolean_t ipsec_kstat_init(ipsec_stack_t *);
static void ipsec_kstat_destroy(ipsec_stack_t *);
static int ipsec_free_tables(ipsec_stack_t *);
static int tunnel_compare(const void *, const void *);
static void ipsec_freemsg_chain(mblk_t *);

/*
 * Selector hash table is statically sized at module load time.
 * we default to 251 buckets, which is the largest prime number under 255
 */

#define	IPSEC_SPDHASH_DEFAULT 251

/* SPD hash-size tunable per tunnel. */
#define	TUN_SPDHASH_DEFAULT 5

uint32_t ipsec_spd_hashsize;
uint32_t tun_spd_hashsize;

#define	IPSEC_SEL_NOHASH ((uint32_t)(~0))

/*
 * Handle global across all stack instances
 */
static crypto_notify_handle_t prov_update_handle = NULL;

static kmem_cache_t *ipsec_action_cache;
static kmem_cache_t *ipsec_sel_cache;
static kmem_cache_t *ipsec_pol_cache;

/* Frag cache prototypes */
static void ipsec_fragcache_clean(ipsec_fragcache_t *, ipsec_stack_t *);
static ipsec_fragcache_entry_t *fragcache_delentry(int,
    ipsec_fragcache_entry_t *, ipsec_fragcache_t *, ipsec_stack_t *);
boolean_t ipsec_fragcache_init(ipsec_fragcache_t *);
void ipsec_fragcache_uninit(ipsec_fragcache_t *, ipsec_stack_t *ipss);
mblk_t *ipsec_fragcache_add(ipsec_fragcache_t *, mblk_t *, mblk_t *,
    int, ipsec_stack_t *);

int ipsec_hdr_pullup_needed = 0;
int ipsec_weird_null_inbound_policy = 0;

#define	ALGBITS_ROUND_DOWN(x, align)	(((x)/(align))*(align))
#define	ALGBITS_ROUND_UP(x, align)	ALGBITS_ROUND_DOWN((x)+(align)-1, align)

/*
 * Inbound traffic should have matching identities for both SA's.
 */

#define	SA_IDS_MATCH(sa1, sa2) 						\
	(((sa1) == NULL) || ((sa2) == NULL) ||				\
	(((sa1)->ipsa_src_cid == (sa2)->ipsa_src_cid) &&		\
	    (((sa1)->ipsa_dst_cid == (sa2)->ipsa_dst_cid))))

/*
 * IPv6 Fragments
 */
#define	IS_V6_FRAGMENT(ipp)	(ipp.ipp_fields & IPPF_FRAGHDR)

/*
 * Policy failure messages.
 */
static char *ipsec_policy_failure_msgs[] = {

	/* IPSEC_POLICY_NOT_NEEDED */
	"%s: Dropping the datagram because the incoming packet "
	"is %s, but the recipient expects clear; Source %s, "
	"Destination %s.\n",

	/* IPSEC_POLICY_MISMATCH */
	"%s: Policy Failure for the incoming packet (%s); Source %s, "
	"Destination %s.\n",

	/* IPSEC_POLICY_AUTH_NOT_NEEDED	*/
	"%s: Authentication present while not expected in the "
	"incoming %s packet; Source %s, Destination %s.\n",

	/* IPSEC_POLICY_ENCR_NOT_NEEDED */
	"%s: Encryption present while not expected in the "
	"incoming %s packet; Source %s, Destination %s.\n",

	/* IPSEC_POLICY_SE_NOT_NEEDED */
	"%s: Self-Encapsulation present while not expected in the "
	"incoming %s packet; Source %s, Destination %s.\n",
};

/*
 * General overviews:
 *
 * Locking:
 *
 *	All of the system policy structures are protected by a single
 *	rwlock.  These structures are threaded in a
 *	fairly complex fashion and are not expected to change on a
 *	regular basis, so this should not cause scaling/contention
 *	problems.  As a result, policy checks should (hopefully) be MT-hot.
 *
 * Allocation policy:
 *
 *	We use custom kmem cache types for the various
 *	bits & pieces of the policy data structures.  All allocations
 *	use KM_NOSLEEP instead of KM_SLEEP for policy allocation.  The
 *	policy table is of potentially unbounded size, so we don't
 *	want to provide a way to hog all system memory with policy
 *	entries..
 */

/* Convenient functions for freeing or dropping a b_next linked mblk chain */

/* Free all messages in an mblk chain */
static void
ipsec_freemsg_chain(mblk_t *mp)
{
	mblk_t *mpnext;
	while (mp != NULL) {
		ASSERT(mp->b_prev == NULL);
		mpnext = mp->b_next;
		mp->b_next = NULL;
		freemsg(mp);
		mp = mpnext;
	}
}

/*
 * ip_drop all messages in an mblk chain
 * Can handle a b_next chain of ip_recv_attr_t mblks, or just a b_next chain
 * of data.
 */
static void
ip_drop_packet_chain(mblk_t *mp, boolean_t inbound, ill_t *ill,
    struct kstat_named *counter, ipdropper_t *who_called)
{
	mblk_t *mpnext;
	while (mp != NULL) {
		ASSERT(mp->b_prev == NULL);
		mpnext = mp->b_next;
		mp->b_next = NULL;
		if (ip_recv_attr_is_mblk(mp))
			mp = ip_recv_attr_free_mblk(mp);
		ip_drop_packet(mp, inbound, ill, counter, who_called);
		mp = mpnext;
	}
}

/*
 * AVL tree comparison function.
 * the in-kernel avl assumes unique keys for all objects.
 * Since sometimes policy will duplicate rules, we may insert
 * multiple rules with the same rule id, so we need a tie-breaker.
 */
static int
ipsec_policy_cmpbyid(const void *a, const void *b)
{
	const ipsec_policy_t *ipa, *ipb;
	uint64_t idxa, idxb;

	ipa = (const ipsec_policy_t *)a;
	ipb = (const ipsec_policy_t *)b;
	idxa = ipa->ipsp_index;
	idxb = ipb->ipsp_index;

	if (idxa < idxb)
		return (-1);
	if (idxa > idxb)
		return (1);
	/*
	 * Tie-breaker #1: All installed policy rules have a non-NULL
	 * ipsl_sel (selector set), so an entry with a NULL ipsp_sel is not
	 * actually in-tree but rather a template node being used in
	 * an avl_find query; see ipsec_policy_delete().  This gives us
	 * a placeholder in the ordering just before the first entry with
	 * a key >= the one we're looking for, so we can walk forward from
	 * that point to get the remaining entries with the same id.
	 */
	if ((ipa->ipsp_sel == NULL) && (ipb->ipsp_sel != NULL))
		return (-1);
	if ((ipb->ipsp_sel == NULL) && (ipa->ipsp_sel != NULL))
		return (1);
	/*
	 * At most one of the arguments to the comparison should have a
	 * NULL selector pointer; if not, the tree is broken.
	 */
	ASSERT(ipa->ipsp_sel != NULL);
	ASSERT(ipb->ipsp_sel != NULL);
	/*
	 * Tie-breaker #2: use the virtual address of the policy node
	 * to arbitrarily break ties.  Since we use the new tree node in
	 * the avl_find() in ipsec_insert_always, the new node will be
	 * inserted into the tree in the right place in the sequence.
	 */
	if (ipa < ipb)
		return (-1);
	if (ipa > ipb)
		return (1);
	return (0);
}

/*
 * Free what ipsec_alloc_table allocated.
 */
void
ipsec_polhead_free_table(ipsec_policy_head_t *iph)
{
	int dir;
	int i;

	for (dir = 0; dir < IPSEC_NTYPES; dir++) {
		ipsec_policy_root_t *ipr = &iph->iph_root[dir];

		if (ipr->ipr_hash == NULL)
			continue;

		for (i = 0; i < ipr->ipr_nchains; i++) {
			ASSERT(ipr->ipr_hash[i].hash_head == NULL);
		}
		kmem_free(ipr->ipr_hash, ipr->ipr_nchains *
		    sizeof (ipsec_policy_hash_t));
		ipr->ipr_hash = NULL;
	}
}

void
ipsec_polhead_destroy(ipsec_policy_head_t *iph)
{
	int dir;

	avl_destroy(&iph->iph_rulebyid);
	rw_destroy(&iph->iph_lock);

	for (dir = 0; dir < IPSEC_NTYPES; dir++) {
		ipsec_policy_root_t *ipr = &iph->iph_root[dir];
		int chain;

		for (chain = 0; chain < ipr->ipr_nchains; chain++)
			mutex_destroy(&(ipr->ipr_hash[chain].hash_lock));

	}
	ipsec_polhead_free_table(iph);
}

/*
 * Free the IPsec stack instance.
 */
/* ARGSUSED */
static void
ipsec_stack_fini(netstackid_t stackid, void *arg)
{
	ipsec_stack_t	*ipss = (ipsec_stack_t *)arg;
	void *cookie;
	ipsec_tun_pol_t *node;
	netstack_t	*ns = ipss->ipsec_netstack;
	int		i;
	ipsec_algtype_t	algtype;

	ipsec_loader_destroy(ipss);

	rw_enter(&ipss->ipsec_tunnel_policy_lock, RW_WRITER);
	/*
	 * It's possible we can just ASSERT() the tree is empty.  After all,
	 * we aren't called until IP is ready to unload (and presumably all
	 * tunnels have been unplumbed).  But we'll play it safe for now, the
	 * loop will just exit immediately if it's empty.
	 */
	cookie = NULL;
	while ((node = (ipsec_tun_pol_t *)
	    avl_destroy_nodes(&ipss->ipsec_tunnel_policies,
	    &cookie)) != NULL) {
		ITP_REFRELE(node, ns);
	}
	avl_destroy(&ipss->ipsec_tunnel_policies);
	rw_exit(&ipss->ipsec_tunnel_policy_lock);
	rw_destroy(&ipss->ipsec_tunnel_policy_lock);

	ipsec_config_flush(ns);

	ipsec_kstat_destroy(ipss);

	ip_drop_unregister(&ipss->ipsec_dropper);

	ip_drop_unregister(&ipss->ipsec_spd_dropper);
	ip_drop_destroy(ipss);
	/*
	 * Globals start with ref == 1 to prevent IPPH_REFRELE() from
	 * attempting to free them, hence they should have 1 now.
	 */
	ipsec_polhead_destroy(&ipss->ipsec_system_policy);
	ASSERT(ipss->ipsec_system_policy.iph_refs == 1);
	ipsec_polhead_destroy(&ipss->ipsec_inactive_policy);
	ASSERT(ipss->ipsec_inactive_policy.iph_refs == 1);

	for (i = 0; i < IPSEC_ACTION_HASH_SIZE; i++) {
		ipsec_action_free_table(ipss->ipsec_action_hash[i].hash_head);
		ipss->ipsec_action_hash[i].hash_head = NULL;
		mutex_destroy(&(ipss->ipsec_action_hash[i].hash_lock));
	}

	for (i = 0; i < ipss->ipsec_spd_hashsize; i++) {
		ASSERT(ipss->ipsec_sel_hash[i].hash_head == NULL);
		mutex_destroy(&(ipss->ipsec_sel_hash[i].hash_lock));
	}

	mutex_enter(&ipss->ipsec_alg_lock);
	for (algtype = 0; algtype < IPSEC_NALGTYPES; algtype ++) {
		int nalgs = ipss->ipsec_nalgs[algtype];

		for (i = 0; i < nalgs; i++) {
			if (ipss->ipsec_alglists[algtype][i] != NULL)
				ipsec_alg_unreg(algtype, i, ns);
		}
	}
	mutex_exit(&ipss->ipsec_alg_lock);
	mutex_destroy(&ipss->ipsec_alg_lock);

	ipsid_gc(ns);
	ipsid_fini(ns);

	(void) ipsec_free_tables(ipss);
	kmem_free(ipss, sizeof (*ipss));
}

void
ipsec_policy_g_destroy(void)
{
	kmem_cache_destroy(ipsec_action_cache);
	kmem_cache_destroy(ipsec_sel_cache);
	kmem_cache_destroy(ipsec_pol_cache);

	ipsec_unregister_prov_update();

	netstack_unregister(NS_IPSEC);
}


/*
 * Free what ipsec_alloc_tables allocated.
 * Called when table allocation fails to free the table.
 */
static int
ipsec_free_tables(ipsec_stack_t *ipss)
{
	int i;

	if (ipss->ipsec_sel_hash != NULL) {
		for (i = 0; i < ipss->ipsec_spd_hashsize; i++) {
			ASSERT(ipss->ipsec_sel_hash[i].hash_head == NULL);
		}
		kmem_free(ipss->ipsec_sel_hash, ipss->ipsec_spd_hashsize *
		    sizeof (*ipss->ipsec_sel_hash));
		ipss->ipsec_sel_hash = NULL;
		ipss->ipsec_spd_hashsize = 0;
	}
	ipsec_polhead_free_table(&ipss->ipsec_system_policy);
	ipsec_polhead_free_table(&ipss->ipsec_inactive_policy);

	return (ENOMEM);
}

/*
 * Attempt to allocate the tables in a single policy head.
 * Return nonzero on failure after cleaning up any work in progress.
 */
int
ipsec_alloc_table(ipsec_policy_head_t *iph, int nchains, int kmflag,
    boolean_t global_cleanup, netstack_t *ns)
{
	int dir;

	for (dir = 0; dir < IPSEC_NTYPES; dir++) {
		ipsec_policy_root_t *ipr = &iph->iph_root[dir];

		ipr->ipr_nchains = nchains;
		ipr->ipr_hash = kmem_zalloc(nchains *
		    sizeof (ipsec_policy_hash_t), kmflag);
		if (ipr->ipr_hash == NULL)
			return (global_cleanup ?
			    ipsec_free_tables(ns->netstack_ipsec) :
			    ENOMEM);
	}
	return (0);
}

/*
 * Attempt to allocate the various tables.  Return nonzero on failure
 * after cleaning up any work in progress.
 */
static int
ipsec_alloc_tables(int kmflag, netstack_t *ns)
{
	int error;
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	error = ipsec_alloc_table(&ipss->ipsec_system_policy,
	    ipss->ipsec_spd_hashsize, kmflag, B_TRUE, ns);
	if (error != 0)
		return (error);

	error = ipsec_alloc_table(&ipss->ipsec_inactive_policy,
	    ipss->ipsec_spd_hashsize, kmflag, B_TRUE, ns);
	if (error != 0)
		return (error);

	ipss->ipsec_sel_hash = kmem_zalloc(ipss->ipsec_spd_hashsize *
	    sizeof (*ipss->ipsec_sel_hash), kmflag);

	if (ipss->ipsec_sel_hash == NULL)
		return (ipsec_free_tables(ipss));

	return (0);
}

/*
 * After table allocation, initialize a policy head.
 */
void
ipsec_polhead_init(ipsec_policy_head_t *iph, int nchains)
{
	int dir, chain;

	rw_init(&iph->iph_lock, NULL, RW_DEFAULT, NULL);
	avl_create(&iph->iph_rulebyid, ipsec_policy_cmpbyid,
	    sizeof (ipsec_policy_t), offsetof(ipsec_policy_t, ipsp_byid));

	for (dir = 0; dir < IPSEC_NTYPES; dir++) {
		ipsec_policy_root_t *ipr = &iph->iph_root[dir];
		ipr->ipr_nchains = nchains;

		for (chain = 0; chain < nchains; chain++) {
			mutex_init(&(ipr->ipr_hash[chain].hash_lock),
			    NULL, MUTEX_DEFAULT, NULL);
		}
	}
}

static boolean_t
ipsec_kstat_init(ipsec_stack_t *ipss)
{
	ipss->ipsec_ksp = kstat_create_netstack("ip", 0, "ipsec_stat", "net",
	    KSTAT_TYPE_NAMED, sizeof (ipsec_kstats_t) / sizeof (kstat_named_t),
	    KSTAT_FLAG_PERSISTENT, ipss->ipsec_netstack->netstack_stackid);

	if (ipss->ipsec_ksp == NULL || ipss->ipsec_ksp->ks_data == NULL)
		return (B_FALSE);

	ipss->ipsec_kstats = ipss->ipsec_ksp->ks_data;

#define	KI(x) kstat_named_init(&ipss->ipsec_kstats->x, #x, KSTAT_DATA_UINT64)
	KI(esp_stat_in_requests);
	KI(esp_stat_in_discards);
	KI(esp_stat_lookup_failure);
	KI(ah_stat_in_requests);
	KI(ah_stat_in_discards);
	KI(ah_stat_lookup_failure);
	KI(sadb_acquire_maxpackets);
	KI(sadb_acquire_qhiwater);
#undef KI

	kstat_install(ipss->ipsec_ksp);
	return (B_TRUE);
}

static void
ipsec_kstat_destroy(ipsec_stack_t *ipss)
{
	kstat_delete_netstack(ipss->ipsec_ksp,
	    ipss->ipsec_netstack->netstack_stackid);
	ipss->ipsec_kstats = NULL;

}

/*
 * Initialize the IPsec stack instance.
 */
/* ARGSUSED */
static void *
ipsec_stack_init(netstackid_t stackid, netstack_t *ns)
{
	ipsec_stack_t	*ipss;
	int i;

	ipss = (ipsec_stack_t *)kmem_zalloc(sizeof (*ipss), KM_SLEEP);
	ipss->ipsec_netstack = ns;

	/*
	 * FIXME: netstack_ipsec is used by some of the routines we call
	 * below, but it isn't set until this routine returns.
	 * Either we introduce optional xxx_stack_alloc() functions
	 * that will be called by the netstack framework before xxx_stack_init,
	 * or we switch spd.c and sadb.c to operate on ipsec_stack_t
	 * (latter has some include file order issues for sadb.h, but makes
	 * sense if we merge some of the ipsec related stack_t's together.
	 */
	ns->netstack_ipsec = ipss;

	/*
	 * Make two attempts to allocate policy hash tables; try it at
	 * the "preferred" size (may be set in /etc/system) first,
	 * then fall back to the default size.
	 */
	ipss->ipsec_spd_hashsize = (ipsec_spd_hashsize == 0) ?
	    IPSEC_SPDHASH_DEFAULT : ipsec_spd_hashsize;

	if (ipsec_alloc_tables(KM_NOSLEEP, ns) != 0) {
		cmn_err(CE_WARN,
		    "Unable to allocate %d entry IPsec policy hash table",
		    ipss->ipsec_spd_hashsize);
		ipss->ipsec_spd_hashsize = IPSEC_SPDHASH_DEFAULT;
		cmn_err(CE_WARN, "Falling back to %d entries",
		    ipss->ipsec_spd_hashsize);
		(void) ipsec_alloc_tables(KM_SLEEP, ns);
	}

	/* Just set a default for tunnels. */
	ipss->ipsec_tun_spd_hashsize = (tun_spd_hashsize == 0) ?
	    TUN_SPDHASH_DEFAULT : tun_spd_hashsize;

	ipsid_init(ns);
	/*
	 * Globals need ref == 1 to prevent IPPH_REFRELE() from attempting
	 * to free them.
	 */
	ipss->ipsec_system_policy.iph_refs = 1;
	ipss->ipsec_inactive_policy.iph_refs = 1;
	ipsec_polhead_init(&ipss->ipsec_system_policy,
	    ipss->ipsec_spd_hashsize);
	ipsec_polhead_init(&ipss->ipsec_inactive_policy,
	    ipss->ipsec_spd_hashsize);
	rw_init(&ipss->ipsec_tunnel_policy_lock, NULL, RW_DEFAULT, NULL);
	avl_create(&ipss->ipsec_tunnel_policies, tunnel_compare,
	    sizeof (ipsec_tun_pol_t), 0);

	ipss->ipsec_next_policy_index = 1;

	rw_init(&ipss->ipsec_system_policy.iph_lock, NULL, RW_DEFAULT, NULL);
	rw_init(&ipss->ipsec_inactive_policy.iph_lock, NULL, RW_DEFAULT, NULL);

	for (i = 0; i < IPSEC_ACTION_HASH_SIZE; i++)
		mutex_init(&(ipss->ipsec_action_hash[i].hash_lock),
		    NULL, MUTEX_DEFAULT, NULL);

	for (i = 0; i < ipss->ipsec_spd_hashsize; i++)
		mutex_init(&(ipss->ipsec_sel_hash[i].hash_lock),
		    NULL, MUTEX_DEFAULT, NULL);

	mutex_init(&ipss->ipsec_alg_lock, NULL, MUTEX_DEFAULT, NULL);
	for (i = 0; i < IPSEC_NALGTYPES; i++) {
		ipss->ipsec_nalgs[i] = 0;
	}

	ip_drop_init(ipss);
	ip_drop_register(&ipss->ipsec_spd_dropper, "IPsec SPD");

	/* IP's IPsec code calls the packet dropper */
	ip_drop_register(&ipss->ipsec_dropper, "IP IPsec processing");

	(void) ipsec_kstat_init(ipss);

	ipsec_loader_init(ipss);
	ipsec_loader_start(ipss);

	return (ipss);
}

/* Global across all stack instances */
void
ipsec_policy_g_init(void)
{
	ipsec_action_cache = kmem_cache_create("ipsec_actions",
	    sizeof (ipsec_action_t), _POINTER_ALIGNMENT, NULL, NULL,
	    ipsec_action_reclaim, NULL, NULL, 0);
	ipsec_sel_cache = kmem_cache_create("ipsec_selectors",
	    sizeof (ipsec_sel_t), _POINTER_ALIGNMENT, NULL, NULL,
	    NULL, NULL, NULL, 0);
	ipsec_pol_cache = kmem_cache_create("ipsec_policy",
	    sizeof (ipsec_policy_t), _POINTER_ALIGNMENT, NULL, NULL,
	    NULL, NULL, NULL, 0);

	/*
	 * We want to be informed each time a stack is created or
	 * destroyed in the kernel, so we can maintain the
	 * set of ipsec_stack_t's.
	 */
	netstack_register(NS_IPSEC, ipsec_stack_init, NULL, ipsec_stack_fini);
}

/*
 * Sort algorithm lists.
 *
 * I may need to split this based on
 * authentication/encryption, and I may wish to have an administrator
 * configure this list.  Hold on to some NDD variables...
 *
 * XXX For now, sort on minimum key size (GAG!).  While minimum key size is
 * not the ideal metric, it's the only quantifiable measure available.
 * We need a better metric for sorting algorithms by preference.
 */
static void
alg_insert_sortlist(enum ipsec_algtype at, uint8_t algid, netstack_t *ns)
{
	ipsec_stack_t	*ipss = ns->netstack_ipsec;
	ipsec_alginfo_t *ai = ipss->ipsec_alglists[at][algid];
	uint8_t holder, swap;
	uint_t i;
	uint_t count = ipss->ipsec_nalgs[at];
	ASSERT(ai != NULL);
	ASSERT(algid == ai->alg_id);

	ASSERT(MUTEX_HELD(&ipss->ipsec_alg_lock));

	holder = algid;

	for (i = 0; i < count - 1; i++) {
		ipsec_alginfo_t *alt;

		alt = ipss->ipsec_alglists[at][ipss->ipsec_sortlist[at][i]];
		/*
		 * If you want to give precedence to newly added algs,
		 * add the = in the > comparison.
		 */
		if ((holder != algid) || (ai->alg_minbits > alt->alg_minbits)) {
			/* Swap sortlist[i] and holder. */
			swap = ipss->ipsec_sortlist[at][i];
			ipss->ipsec_sortlist[at][i] = holder;
			holder = swap;
			ai = alt;
		} /* Else just continue. */
	}

	/* Store holder in last slot. */
	ipss->ipsec_sortlist[at][i] = holder;
}

/*
 * Remove an algorithm from a sorted algorithm list.
 * This should be considerably easier, even with complex sorting.
 */
static void
alg_remove_sortlist(enum ipsec_algtype at, uint8_t algid, netstack_t *ns)
{
	boolean_t copyback = B_FALSE;
	int i;
	ipsec_stack_t	*ipss = ns->netstack_ipsec;
	int newcount = ipss->ipsec_nalgs[at];

	ASSERT(MUTEX_HELD(&ipss->ipsec_alg_lock));

	for (i = 0; i <= newcount; i++) {
		if (copyback) {
			ipss->ipsec_sortlist[at][i-1] =
			    ipss->ipsec_sortlist[at][i];
		} else if (ipss->ipsec_sortlist[at][i] == algid) {
			copyback = B_TRUE;
		}
	}
}

/*
 * Add the specified algorithm to the algorithm tables.
 * Must be called while holding the algorithm table writer lock.
 */
void
ipsec_alg_reg(ipsec_algtype_t algtype, ipsec_alginfo_t *alg, netstack_t *ns)
{
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	ASSERT(MUTEX_HELD(&ipss->ipsec_alg_lock));

	ASSERT(ipss->ipsec_alglists[algtype][alg->alg_id] == NULL);
	ipsec_alg_fix_min_max(alg, algtype, ns);
	ipss->ipsec_alglists[algtype][alg->alg_id] = alg;

	ipss->ipsec_nalgs[algtype]++;
	alg_insert_sortlist(algtype, alg->alg_id, ns);
}

/*
 * Remove the specified algorithm from the algorithm tables.
 * Must be called while holding the algorithm table writer lock.
 */
void
ipsec_alg_unreg(ipsec_algtype_t algtype, uint8_t algid, netstack_t *ns)
{
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	ASSERT(MUTEX_HELD(&ipss->ipsec_alg_lock));

	ASSERT(ipss->ipsec_alglists[algtype][algid] != NULL);
	ipsec_alg_free(ipss->ipsec_alglists[algtype][algid]);
	ipss->ipsec_alglists[algtype][algid] = NULL;

	ipss->ipsec_nalgs[algtype]--;
	alg_remove_sortlist(algtype, algid, ns);
}

/*
 * Hooks for spdsock to get a grip on system policy.
 */

ipsec_policy_head_t *
ipsec_system_policy(netstack_t *ns)
{
	ipsec_stack_t	*ipss = ns->netstack_ipsec;
	ipsec_policy_head_t *h = &ipss->ipsec_system_policy;

	IPPH_REFHOLD(h);
	return (h);
}

ipsec_policy_head_t *
ipsec_inactive_policy(netstack_t *ns)
{
	ipsec_stack_t	*ipss = ns->netstack_ipsec;
	ipsec_policy_head_t *h = &ipss->ipsec_inactive_policy;

	IPPH_REFHOLD(h);
	return (h);
}

/*
 * Lock inactive policy, then active policy, then exchange policy root
 * pointers.
 */
void
ipsec_swap_policy(ipsec_policy_head_t *active, ipsec_policy_head_t *inactive,
    netstack_t *ns)
{
	int af, dir;
	avl_tree_t r1, r2;

	rw_enter(&inactive->iph_lock, RW_WRITER);
	rw_enter(&active->iph_lock, RW_WRITER);

	r1 = active->iph_rulebyid;
	r2 = inactive->iph_rulebyid;
	active->iph_rulebyid = r2;
	inactive->iph_rulebyid = r1;

	for (dir = 0; dir < IPSEC_NTYPES; dir++) {
		ipsec_policy_hash_t *h1, *h2;

		h1 = active->iph_root[dir].ipr_hash;
		h2 = inactive->iph_root[dir].ipr_hash;
		active->iph_root[dir].ipr_hash = h2;
		inactive->iph_root[dir].ipr_hash = h1;

		for (af = 0; af < IPSEC_NAF; af++) {
			ipsec_policy_t *t1, *t2;

			t1 = active->iph_root[dir].ipr_nonhash[af];
			t2 = inactive->iph_root[dir].ipr_nonhash[af];
			active->iph_root[dir].ipr_nonhash[af] = t2;
			inactive->iph_root[dir].ipr_nonhash[af] = t1;
			if (t1 != NULL) {
				t1->ipsp_hash.hash_pp =
				    &(inactive->iph_root[dir].ipr_nonhash[af]);
			}
			if (t2 != NULL) {
				t2->ipsp_hash.hash_pp =
				    &(active->iph_root[dir].ipr_nonhash[af]);
			}

		}
	}
	active->iph_gen++;
	inactive->iph_gen++;
	ipsec_update_present_flags(ns->netstack_ipsec);
	rw_exit(&active->iph_lock);
	rw_exit(&inactive->iph_lock);
}

/*
 * Swap global policy primary/secondary.
 */
void
ipsec_swap_global_policy(netstack_t *ns)
{
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	ipsec_swap_policy(&ipss->ipsec_system_policy,
	    &ipss->ipsec_inactive_policy, ns);
}

/*
 * Clone one policy rule..
 */
static ipsec_policy_t *
ipsec_copy_policy(const ipsec_policy_t *src)
{
	ipsec_policy_t *dst = kmem_cache_alloc(ipsec_pol_cache, KM_NOSLEEP);

	if (dst == NULL)
		return (NULL);

	/*
	 * Adjust refcounts of cloned state.
	 */
	IPACT_REFHOLD(src->ipsp_act);
	src->ipsp_sel->ipsl_refs++;

	HASH_NULL(dst, ipsp_hash);
	dst->ipsp_netstack = src->ipsp_netstack;
	dst->ipsp_refs = 1;
	dst->ipsp_sel = src->ipsp_sel;
	dst->ipsp_act = src->ipsp_act;
	dst->ipsp_prio = src->ipsp_prio;
	dst->ipsp_index = src->ipsp_index;

	return (dst);
}

void
ipsec_insert_always(avl_tree_t *tree, void *new_node)
{
	void *node;
	avl_index_t where;

	node = avl_find(tree, new_node, &where);
	ASSERT(node == NULL);
	avl_insert(tree, new_node, where);
}


static int
ipsec_copy_chain(ipsec_policy_head_t *dph, ipsec_policy_t *src,
    ipsec_policy_t **dstp)
{
	for (; src != NULL; src = src->ipsp_hash.hash_next) {
		ipsec_policy_t *dst = ipsec_copy_policy(src);
		if (dst == NULL)
			return (ENOMEM);

		HASHLIST_INSERT(dst, ipsp_hash, *dstp);
		ipsec_insert_always(&dph->iph_rulebyid, dst);
	}
	return (0);
}



/*
 * Make one policy head look exactly like another.
 *
 * As with ipsec_swap_policy, we lock the destination policy head first, then
 * the source policy head. Note that we only need to read-lock the source
 * policy head as we are not changing it.
 */
int
ipsec_copy_polhead(ipsec_policy_head_t *sph, ipsec_policy_head_t *dph,
    netstack_t *ns)
{
	int af, dir, chain, nchains;

	rw_enter(&dph->iph_lock, RW_WRITER);

	ipsec_polhead_flush(dph, ns);

	rw_enter(&sph->iph_lock, RW_READER);

	for (dir = 0; dir < IPSEC_NTYPES; dir++) {
		ipsec_policy_root_t *dpr = &dph->iph_root[dir];
		ipsec_policy_root_t *spr = &sph->iph_root[dir];
		nchains = dpr->ipr_nchains;

		ASSERT(dpr->ipr_nchains == spr->ipr_nchains);

		for (af = 0; af < IPSEC_NAF; af++) {
			if (ipsec_copy_chain(dph, spr->ipr_nonhash[af],
			    &dpr->ipr_nonhash[af]))
				goto abort_copy;
		}

		for (chain = 0; chain < nchains; chain++) {
			if (ipsec_copy_chain(dph,
			    spr->ipr_hash[chain].hash_head,
			    &dpr->ipr_hash[chain].hash_head))
				goto abort_copy;
		}
	}

	dph->iph_gen++;

	rw_exit(&sph->iph_lock);
	rw_exit(&dph->iph_lock);
	return (0);

abort_copy:
	ipsec_polhead_flush(dph, ns);
	rw_exit(&sph->iph_lock);
	rw_exit(&dph->iph_lock);
	return (ENOMEM);
}

/*
 * Clone currently active policy to the inactive policy list.
 */
int
ipsec_clone_system_policy(netstack_t *ns)
{
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	return (ipsec_copy_polhead(&ipss->ipsec_system_policy,
	    &ipss->ipsec_inactive_policy, ns));
}

/*
 * Extract the string from ipsec_policy_failure_msgs[type] and
 * log it.
 *
 */
void
ipsec_log_policy_failure(int type, char *func_name, ipha_t *ipha, ip6_t *ip6h,
    boolean_t secure, netstack_t *ns)
{
	char	sbuf[INET6_ADDRSTRLEN];
	char	dbuf[INET6_ADDRSTRLEN];
	char	*s;
	char	*d;
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	ASSERT((ipha == NULL && ip6h != NULL) ||
	    (ip6h == NULL && ipha != NULL));

	if (ipha != NULL) {
		s = inet_ntop(AF_INET, &ipha->ipha_src, sbuf, sizeof (sbuf));
		d = inet_ntop(AF_INET, &ipha->ipha_dst, dbuf, sizeof (dbuf));
	} else {
		s = inet_ntop(AF_INET6, &ip6h->ip6_src, sbuf, sizeof (sbuf));
		d = inet_ntop(AF_INET6, &ip6h->ip6_dst, dbuf, sizeof (dbuf));

	}

	/* Always bump the policy failure counter. */
	ipss->ipsec_policy_failure_count[type]++;

	ipsec_rl_strlog(ns, IP_MOD_ID, 0, 0, SL_ERROR|SL_WARN|SL_CONSOLE,
	    ipsec_policy_failure_msgs[type], func_name,
	    (secure ? "secure" : "not secure"), s, d);
}

/*
 * Rate-limiting front-end to strlog() for AH and ESP.	Uses the ndd variables
 * in /dev/ip and the same rate-limiting clock so that there's a single
 * knob to turn to throttle the rate of messages.
 */
void
ipsec_rl_strlog(netstack_t *ns, short mid, short sid, char level, ushort_t sl,
    char *fmt, ...)
{
	va_list adx;
	hrtime_t current = gethrtime();
	ip_stack_t	*ipst = ns->netstack_ip;
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	sl |= SL_CONSOLE;
	/*
	 * Throttle logging to stop syslog from being swamped. If variable
	 * 'ipsec_policy_log_interval' is zero, don't log any messages at
	 * all, otherwise log only one message every 'ipsec_policy_log_interval'
	 * msec. Convert interval (in msec) to hrtime (in nsec).
	 */

	if (ipst->ips_ipsec_policy_log_interval) {
		if (ipss->ipsec_policy_failure_last +
		    ((hrtime_t)ipst->ips_ipsec_policy_log_interval *
		    (hrtime_t)1000000) <= current) {
			va_start(adx, fmt);
			(void) vstrlog(mid, sid, level, sl, fmt, adx);
			va_end(adx);
			ipss->ipsec_policy_failure_last = current;
		}
	}
}

void
ipsec_config_flush(netstack_t *ns)
{
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	rw_enter(&ipss->ipsec_system_policy.iph_lock, RW_WRITER);
	ipsec_polhead_flush(&ipss->ipsec_system_policy, ns);
	ipss->ipsec_next_policy_index = 1;
	rw_exit(&ipss->ipsec_system_policy.iph_lock);
	ipsec_action_reclaim_stack(ns);
}

/*
 * Clip a policy's min/max keybits vs. the capabilities of the
 * algorithm.
 */
static void
act_alg_adjust(uint_t algtype, uint_t algid,
    uint16_t *minbits, uint16_t *maxbits, netstack_t *ns)
{
	ipsec_stack_t	*ipss = ns->netstack_ipsec;
	ipsec_alginfo_t *algp = ipss->ipsec_alglists[algtype][algid];

	if (algp != NULL) {
		/*
		 * If passed-in minbits is zero, we assume the caller trusts
		 * us with setting the minimum key size.  We pick the
		 * algorithms DEFAULT key size for the minimum in this case.
		 */
		if (*minbits == 0) {
			*minbits = algp->alg_default_bits;
			ASSERT(*minbits >= algp->alg_minbits);
		} else {
			*minbits = MAX(MIN(*minbits, algp->alg_maxbits),
			    algp->alg_minbits);
		}
		if (*maxbits == 0)
			*maxbits = algp->alg_maxbits;
		else
			*maxbits = MIN(MAX(*maxbits, algp->alg_minbits),
			    algp->alg_maxbits);
		ASSERT(*minbits <= *maxbits);
	} else {
		*minbits = 0;
		*maxbits = 0;
	}
}

/*
 * Check an action's requested algorithms against the algorithms currently
 * loaded in the system.
 */
boolean_t
ipsec_check_action(ipsec_act_t *act, int *diag, netstack_t *ns)
{
	ipsec_prot_t *ipp;
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	ipp = &act->ipa_apply;

	if (ipp->ipp_use_ah &&
	    ipss->ipsec_alglists[IPSEC_ALG_AUTH][ipp->ipp_auth_alg] == NULL) {
		*diag = SPD_DIAGNOSTIC_UNSUPP_AH_ALG;
		return (B_FALSE);
	}
	if (ipp->ipp_use_espa &&
	    ipss->ipsec_alglists[IPSEC_ALG_AUTH][ipp->ipp_esp_auth_alg] ==
	    NULL) {
		*diag = SPD_DIAGNOSTIC_UNSUPP_ESP_AUTH_ALG;
		return (B_FALSE);
	}
	if (ipp->ipp_use_esp &&
	    ipss->ipsec_alglists[IPSEC_ALG_ENCR][ipp->ipp_encr_alg] == NULL) {
		*diag = SPD_DIAGNOSTIC_UNSUPP_ESP_ENCR_ALG;
		return (B_FALSE);
	}

	act_alg_adjust(IPSEC_ALG_AUTH, ipp->ipp_auth_alg,
	    &ipp->ipp_ah_minbits, &ipp->ipp_ah_maxbits, ns);
	act_alg_adjust(IPSEC_ALG_AUTH, ipp->ipp_esp_auth_alg,
	    &ipp->ipp_espa_minbits, &ipp->ipp_espa_maxbits, ns);
	act_alg_adjust(IPSEC_ALG_ENCR, ipp->ipp_encr_alg,
	    &ipp->ipp_espe_minbits, &ipp->ipp_espe_maxbits, ns);

	if (ipp->ipp_ah_minbits > ipp->ipp_ah_maxbits) {
		*diag = SPD_DIAGNOSTIC_UNSUPP_AH_KEYSIZE;
		return (B_FALSE);
	}
	if (ipp->ipp_espa_minbits > ipp->ipp_espa_maxbits) {
		*diag = SPD_DIAGNOSTIC_UNSUPP_ESP_AUTH_KEYSIZE;
		return (B_FALSE);
	}
	if (ipp->ipp_espe_minbits > ipp->ipp_espe_maxbits) {
		*diag = SPD_DIAGNOSTIC_UNSUPP_ESP_ENCR_KEYSIZE;
		return (B_FALSE);
	}
	/* TODO: sanity check lifetimes */
	return (B_TRUE);
}

/*
 * Set up a single action during wildcard expansion..
 */
static void
ipsec_setup_act(ipsec_act_t *outact, ipsec_act_t *act,
    uint_t auth_alg, uint_t encr_alg, uint_t eauth_alg, netstack_t *ns)
{
	ipsec_prot_t *ipp;

	*outact = *act;
	ipp = &outact->ipa_apply;
	ipp->ipp_auth_alg = (uint8_t)auth_alg;
	ipp->ipp_encr_alg = (uint8_t)encr_alg;
	ipp->ipp_esp_auth_alg = (uint8_t)eauth_alg;

	act_alg_adjust(IPSEC_ALG_AUTH, auth_alg,
	    &ipp->ipp_ah_minbits, &ipp->ipp_ah_maxbits, ns);
	act_alg_adjust(IPSEC_ALG_AUTH, eauth_alg,
	    &ipp->ipp_espa_minbits, &ipp->ipp_espa_maxbits, ns);
	act_alg_adjust(IPSEC_ALG_ENCR, encr_alg,
	    &ipp->ipp_espe_minbits, &ipp->ipp_espe_maxbits, ns);
}

/*
 * combinatoric expansion time: expand a wildcarded action into an
 * array of wildcarded actions; we return the exploded action list,
 * and return a count in *nact (output only).
 */
static ipsec_act_t *
ipsec_act_wildcard_expand(ipsec_act_t *act, uint_t *nact, netstack_t *ns)
{
	boolean_t use_ah, use_esp, use_espa;
	boolean_t wild_auth, wild_encr, wild_eauth;
	uint_t	auth_alg, auth_idx, auth_min, auth_max;
	uint_t	eauth_alg, eauth_idx, eauth_min, eauth_max;
	uint_t  encr_alg, encr_idx, encr_min, encr_max;
	uint_t	action_count, ai;
	ipsec_act_t *outact;
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	if (act->ipa_type != IPSEC_ACT_APPLY) {
		outact = kmem_alloc(sizeof (*act), KM_NOSLEEP);
		*nact = 1;
		if (outact != NULL)
			bcopy(act, outact, sizeof (*act));
		return (outact);
	}
	/*
	 * compute the combinatoric explosion..
	 *
	 * we assume a request for encr if esp_req is PREF_REQUIRED
	 * we assume a request for ah auth if ah_req is PREF_REQUIRED.
	 * we assume a request for esp auth if !ah and esp_req is PREF_REQUIRED
	 */

	use_ah = act->ipa_apply.ipp_use_ah;
	use_esp = act->ipa_apply.ipp_use_esp;
	use_espa = act->ipa_apply.ipp_use_espa;
	auth_alg = act->ipa_apply.ipp_auth_alg;
	eauth_alg = act->ipa_apply.ipp_esp_auth_alg;
	encr_alg = act->ipa_apply.ipp_encr_alg;

	wild_auth = use_ah && (auth_alg == 0);
	wild_eauth = use_espa && (eauth_alg == 0);
	wild_encr = use_esp && (encr_alg == 0);

	action_count = 1;
	auth_min = auth_max = auth_alg;
	eauth_min = eauth_max = eauth_alg;
	encr_min = encr_max = encr_alg;

	/*
	 * set up for explosion.. for each dimension, expand output
	 * size by the explosion factor.
	 *
	 * Don't include the "any" algorithms, if defined, as no
	 * kernel policies should be set for these algorithms.
	 */

#define	SET_EXP_MINMAX(type, wild, alg, min, max, ipss)		\
	if (wild) {						\
		int nalgs = ipss->ipsec_nalgs[type];		\
		if (ipss->ipsec_alglists[type][alg] != NULL)	\
			nalgs--;				\
		action_count *= nalgs;				\
		min = 0;					\
		max = ipss->ipsec_nalgs[type] - 1;		\
	}

	SET_EXP_MINMAX(IPSEC_ALG_AUTH, wild_auth, SADB_AALG_NONE,
	    auth_min, auth_max, ipss);
	SET_EXP_MINMAX(IPSEC_ALG_AUTH, wild_eauth, SADB_AALG_NONE,
	    eauth_min, eauth_max, ipss);
	SET_EXP_MINMAX(IPSEC_ALG_ENCR, wild_encr, SADB_EALG_NONE,
	    encr_min, encr_max, ipss);

#undef	SET_EXP_MINMAX

	/*
	 * ok, allocate the whole mess..
	 */

	outact = kmem_alloc(sizeof (*outact) * action_count, KM_NOSLEEP);
	if (outact == NULL)
		return (NULL);

	/*
	 * Now compute all combinations.  Note that non-wildcarded
	 * dimensions just get a single value from auth_min, while
	 * wildcarded dimensions indirect through the sortlist.
	 *
	 * We do encryption outermost since, at this time, there's
	 * greater difference in security and performance between
	 * encryption algorithms vs. authentication algorithms.
	 */

	ai = 0;

#define	WHICH_ALG(type, wild, idx, ipss) \
	((wild)?(ipss->ipsec_sortlist[type][idx]):(idx))

	for (encr_idx = encr_min; encr_idx <= encr_max; encr_idx++) {
		encr_alg = WHICH_ALG(IPSEC_ALG_ENCR, wild_encr, encr_idx, ipss);
		if (wild_encr && encr_alg == SADB_EALG_NONE)
			continue;
		for (auth_idx = auth_min; auth_idx <= auth_max; auth_idx++) {
			auth_alg = WHICH_ALG(IPSEC_ALG_AUTH, wild_auth,
			    auth_idx, ipss);
			if (wild_auth && auth_alg == SADB_AALG_NONE)
				continue;
			for (eauth_idx = eauth_min; eauth_idx <= eauth_max;
			    eauth_idx++) {
				eauth_alg = WHICH_ALG(IPSEC_ALG_AUTH,
				    wild_eauth, eauth_idx, ipss);
				if (wild_eauth && eauth_alg == SADB_AALG_NONE)
					continue;

				ipsec_setup_act(&outact[ai], act,
				    auth_alg, encr_alg, eauth_alg, ns);
				ai++;
			}
		}
	}

#undef WHICH_ALG

	ASSERT(ai == action_count);
	*nact = action_count;
	return (outact);
}

/*
 * Extract the parts of an ipsec_prot_t from an old-style ipsec_req_t.
 */
static void
ipsec_prot_from_req(const ipsec_req_t *req, ipsec_prot_t *ipp)
{
	bzero(ipp, sizeof (*ipp));
	/*
	 * ipp_use_* are bitfields.  Look at "!!" in the following as a
	 * "boolean canonicalization" operator.
	 */
	ipp->ipp_use_ah = !!(req->ipsr_ah_req & IPSEC_PREF_REQUIRED);
	ipp->ipp_use_esp = !!(req->ipsr_esp_req & IPSEC_PREF_REQUIRED);
	ipp->ipp_use_espa = !!(req->ipsr_esp_auth_alg);
	ipp->ipp_use_se = !!(req->ipsr_self_encap_req & IPSEC_PREF_REQUIRED);
	ipp->ipp_use_unique = !!((req->ipsr_ah_req|req->ipsr_esp_req) &
	    IPSEC_PREF_UNIQUE);
	ipp->ipp_encr_alg = req->ipsr_esp_alg;
	/*
	 * SADB_AALG_ANY is a placeholder to distinguish "any" from
	 * "none" above.  If auth is required, as determined above,
	 * SADB_AALG_ANY becomes 0, which is the representation
	 * of "any" and "none" in PF_KEY v2.
	 */
	ipp->ipp_auth_alg = (req->ipsr_auth_alg != SADB_AALG_ANY) ?
	    req->ipsr_auth_alg : 0;
	ipp->ipp_esp_auth_alg = (req->ipsr_esp_auth_alg != SADB_AALG_ANY) ?
	    req->ipsr_esp_auth_alg : 0;
}

/*
 * Extract a new-style action from a request.
 */
void
ipsec_actvec_from_req(const ipsec_req_t *req, ipsec_act_t **actp, uint_t *nactp,
    netstack_t *ns)
{
	struct ipsec_act act;

	bzero(&act, sizeof (act));
	if ((req->ipsr_ah_req & IPSEC_PREF_NEVER) &&
	    (req->ipsr_esp_req & IPSEC_PREF_NEVER)) {
		act.ipa_type = IPSEC_ACT_BYPASS;
	} else {
		act.ipa_type = IPSEC_ACT_APPLY;
		ipsec_prot_from_req(req, &act.ipa_apply);
	}
	*actp = ipsec_act_wildcard_expand(&act, nactp, ns);
}

/*
 * Convert a new-style "prot" back to an ipsec_req_t (more backwards compat).
 * We assume caller has already zero'ed *req for us.
 */
static int
ipsec_req_from_prot(ipsec_prot_t *ipp, ipsec_req_t *req)
{
	req->ipsr_esp_alg = ipp->ipp_encr_alg;
	req->ipsr_auth_alg = ipp->ipp_auth_alg;
	req->ipsr_esp_auth_alg = ipp->ipp_esp_auth_alg;

	if (ipp->ipp_use_unique) {
		req->ipsr_ah_req |= IPSEC_PREF_UNIQUE;
		req->ipsr_esp_req |= IPSEC_PREF_UNIQUE;
	}
	if (ipp->ipp_use_se)
		req->ipsr_self_encap_req |= IPSEC_PREF_REQUIRED;
	if (ipp->ipp_use_ah)
		req->ipsr_ah_req |= IPSEC_PREF_REQUIRED;
	if (ipp->ipp_use_esp)
		req->ipsr_esp_req |= IPSEC_PREF_REQUIRED;
	return (sizeof (*req));
}

/*
 * Convert a new-style action back to an ipsec_req_t (more backwards compat).
 * We assume caller has already zero'ed *req for us.
 */
static int
ipsec_req_from_act(ipsec_action_t *ap, ipsec_req_t *req)
{
	switch (ap->ipa_act.ipa_type) {
	case IPSEC_ACT_BYPASS:
		req->ipsr_ah_req = IPSEC_PREF_NEVER;
		req->ipsr_esp_req = IPSEC_PREF_NEVER;
		return (sizeof (*req));
	case IPSEC_ACT_APPLY:
		return (ipsec_req_from_prot(&ap->ipa_act.ipa_apply, req));
	}
	return (sizeof (*req));
}

/*
 * Convert a new-style action back to an ipsec_req_t (more backwards compat).
 * We assume caller has already zero'ed *req for us.
 */
int
ipsec_req_from_head(ipsec_policy_head_t *ph, ipsec_req_t *req, int af)
{
	ipsec_policy_t *p;

	/*
	 * FULL-PERSOCK: consult hash table, too?
	 */
	for (p = ph->iph_root[IPSEC_INBOUND].ipr_nonhash[af];
	    p != NULL;
	    p = p->ipsp_hash.hash_next) {
		if ((p->ipsp_sel->ipsl_key.ipsl_valid & IPSL_WILDCARD) == 0)
			return (ipsec_req_from_act(p->ipsp_act, req));
	}
	return (sizeof (*req));
}

/*
 * Based on per-socket or latched policy, convert to an appropriate
 * IP_SEC_OPT ipsec_req_t for the socket option; return size so we can
 * be tail-called from ip.
 */
int
ipsec_req_from_conn(conn_t *connp, ipsec_req_t *req, int af)
{
	ipsec_latch_t *ipl;
	int rv = sizeof (ipsec_req_t);

	bzero(req, sizeof (*req));

	ASSERT(MUTEX_HELD(&connp->conn_lock));
	ipl = connp->conn_latch;

	/*
	 * Find appropriate policy.  First choice is latched action;
	 * failing that, see latched policy; failing that,
	 * look at configured policy.
	 */
	if (ipl != NULL) {
		if (connp->conn_latch_in_action != NULL) {
			rv = ipsec_req_from_act(connp->conn_latch_in_action,
			    req);
			goto done;
		}
		if (connp->conn_latch_in_policy != NULL) {
			rv = ipsec_req_from_act(
			    connp->conn_latch_in_policy->ipsp_act, req);
			goto done;
		}
	}
	if (connp->conn_policy != NULL)
		rv = ipsec_req_from_head(connp->conn_policy, req, af);
done:
	return (rv);
}

void
ipsec_actvec_free(ipsec_act_t *act, uint_t nact)
{
	kmem_free(act, nact * sizeof (*act));
}

/*
 * Consumes a reference to ipsp.
 */
static mblk_t *
ipsec_check_loopback_policy(mblk_t *data_mp, ip_recv_attr_t *ira,
    ipsec_policy_t *ipsp)
{
	if (!(ira->ira_flags & IRAF_IPSEC_SECURE))
		return (data_mp);

	ASSERT(ira->ira_flags & IRAF_LOOPBACK);

	IPPOL_REFRELE(ipsp);

	/*
	 * We should do an actual policy check here.  Revisit this
	 * when we revisit the IPsec API.  (And pass a conn_t in when we
	 * get there.)
	 */

	return (data_mp);
}

/*
 * Check that packet's inbound ports & proto match the selectors
 * expected by the SAs it traversed on the way in.
 */
static boolean_t
ipsec_check_ipsecin_unique(ip_recv_attr_t *ira, const char **reason,
    kstat_named_t **counter, uint64_t pkt_unique, netstack_t *ns)
{
	uint64_t ah_mask, esp_mask;
	ipsa_t *ah_assoc;
	ipsa_t *esp_assoc;
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	ASSERT(ira->ira_flags & IRAF_IPSEC_SECURE);
	ASSERT(!(ira->ira_flags & IRAF_LOOPBACK));

	ah_assoc = ira->ira_ipsec_ah_sa;
	esp_assoc = ira->ira_ipsec_esp_sa;
	ASSERT((ah_assoc != NULL) || (esp_assoc != NULL));

	ah_mask = (ah_assoc != NULL) ? ah_assoc->ipsa_unique_mask : 0;
	esp_mask = (esp_assoc != NULL) ? esp_assoc->ipsa_unique_mask : 0;

	if ((ah_mask == 0) && (esp_mask == 0))
		return (B_TRUE);

	/*
	 * The pkt_unique check will also check for tunnel mode on the SA
	 * vs. the tunneled_packet boolean.  "Be liberal in what you receive"
	 * should not apply in this case.  ;)
	 */

	if (ah_mask != 0 &&
	    ah_assoc->ipsa_unique_id != (pkt_unique & ah_mask)) {
		*reason = "AH inner header mismatch";
		*counter = DROPPER(ipss, ipds_spd_ah_innermismatch);
		return (B_FALSE);
	}
	if (esp_mask != 0 &&
	    esp_assoc->ipsa_unique_id != (pkt_unique & esp_mask)) {
		*reason = "ESP inner header mismatch";
		*counter = DROPPER(ipss, ipds_spd_esp_innermismatch);
		return (B_FALSE);
	}
	return (B_TRUE);
}

static boolean_t
ipsec_check_ipsecin_action(ip_recv_attr_t *ira, mblk_t *mp, ipsec_action_t *ap,
    ipha_t *ipha, ip6_t *ip6h, const char **reason, kstat_named_t **counter,
    netstack_t *ns)
{
	boolean_t ret = B_TRUE;
	ipsec_prot_t *ipp;
	ipsa_t *ah_assoc;
	ipsa_t *esp_assoc;
	boolean_t decaps;
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	ASSERT((ipha == NULL && ip6h != NULL) ||
	    (ip6h == NULL && ipha != NULL));

	if (ira->ira_flags & IRAF_LOOPBACK) {
		/*
		 * Besides accepting pointer-equivalent actions, we also
		 * accept any ICMP errors we generated for ourselves,
		 * regardless of policy.  If we do not wish to make this
		 * assumption in the future, check here, and where
		 * IXAF_TRUSTED_ICMP is initialized in ip.c and ip6.c.
		 */
		if (ap == ira->ira_ipsec_action ||
		    (ira->ira_flags & IRAF_TRUSTED_ICMP))
			return (B_TRUE);

		/* Deep compare necessary here?? */
		*counter = DROPPER(ipss, ipds_spd_loopback_mismatch);
		*reason = "loopback policy mismatch";
		return (B_FALSE);
	}
	ASSERT(!(ira->ira_flags & IRAF_TRUSTED_ICMP));
	ASSERT(ira->ira_flags & IRAF_IPSEC_SECURE);

	ah_assoc = ira->ira_ipsec_ah_sa;
	esp_assoc = ira->ira_ipsec_esp_sa;

	decaps = (ira->ira_flags & IRAF_IPSEC_DECAPS);

	switch (ap->ipa_act.ipa_type) {
	case IPSEC_ACT_DISCARD:
	case IPSEC_ACT_REJECT:
		/* Should "fail hard" */
		*counter = DROPPER(ipss, ipds_spd_explicit);
		*reason = "blocked by policy";
		return (B_FALSE);

	case IPSEC_ACT_BYPASS:
	case IPSEC_ACT_CLEAR:
		*counter = DROPPER(ipss, ipds_spd_got_secure);
		*reason = "expected clear, got protected";
		return (B_FALSE);

	case IPSEC_ACT_APPLY:
		ipp = &ap->ipa_act.ipa_apply;
		/*
		 * As of now we do the simple checks of whether
		 * the datagram has gone through the required IPSEC
		 * protocol constraints or not. We might have more
		 * in the future like sensitive levels, key bits, etc.
		 * If it fails the constraints, check whether we would
		 * have accepted this if it had come in clear.
		 */
		if (ipp->ipp_use_ah) {
			if (ah_assoc == NULL) {
				ret = ipsec_inbound_accept_clear(mp, ipha,
				    ip6h);
				*counter = DROPPER(ipss, ipds_spd_got_clear);
				*reason = "unprotected not accepted";
				break;
			}
			ASSERT(ah_assoc != NULL);
			ASSERT(ipp->ipp_auth_alg != 0);

			if (ah_assoc->ipsa_auth_alg !=
			    ipp->ipp_auth_alg) {
				*counter = DROPPER(ipss, ipds_spd_bad_ahalg);
				*reason = "unacceptable ah alg";
				ret = B_FALSE;
				break;
			}
		} else if (ah_assoc != NULL) {
			/*
			 * Don't allow this. Check IPSEC NOTE above
			 * ip_fanout_proto().
			 */
			*counter = DROPPER(ipss, ipds_spd_got_ah);
			*reason = "unexpected AH";
			ret = B_FALSE;
			break;
		}
		if (ipp->ipp_use_esp) {
			if (esp_assoc == NULL) {
				ret = ipsec_inbound_accept_clear(mp, ipha,
				    ip6h);
				*counter = DROPPER(ipss, ipds_spd_got_clear);
				*reason = "unprotected not accepted";
				break;
			}
			ASSERT(esp_assoc != NULL);
			ASSERT(ipp->ipp_encr_alg != 0);

			if (esp_assoc->ipsa_encr_alg !=
			    ipp->ipp_encr_alg) {
				*counter = DROPPER(ipss, ipds_spd_bad_espealg);
				*reason = "unacceptable esp alg";
				ret = B_FALSE;
				break;
			}
			/*
			 * If the client does not need authentication,
			 * we don't verify the alogrithm.
			 */
			if (ipp->ipp_use_espa) {
				if (esp_assoc->ipsa_auth_alg !=
				    ipp->ipp_esp_auth_alg) {
					*counter = DROPPER(ipss,
					    ipds_spd_bad_espaalg);
					*reason = "unacceptable esp auth alg";
					ret = B_FALSE;
					break;
				}
			}
		} else if (esp_assoc != NULL) {
			/*
			 * Don't allow this. Check IPSEC NOTE above
			 * ip_fanout_proto().
			 */
			*counter = DROPPER(ipss, ipds_spd_got_esp);
			*reason = "unexpected ESP";
			ret = B_FALSE;
			break;
		}
		if (ipp->ipp_use_se) {
			if (!decaps) {
				ret = ipsec_inbound_accept_clear(mp, ipha,
				    ip6h);
				if (!ret) {
					/* XXX mutant? */
					*counter = DROPPER(ipss,
					    ipds_spd_bad_selfencap);
					*reason = "self encap not found";
					break;
				}
			}
		} else if (decaps) {
			/*
			 * XXX If the packet comes in tunneled and the
			 * recipient does not expect it to be tunneled, it
			 * is okay. But we drop to be consistent with the
			 * other cases.
			 */
			*counter = DROPPER(ipss, ipds_spd_got_selfencap);
			*reason = "unexpected self encap";
			ret = B_FALSE;
			break;
		}
		if (ira->ira_ipsec_action != NULL) {
			/*
			 * This can happen if we do a double policy-check on
			 * a packet
			 * XXX XXX should fix this case!
			 */
			IPACT_REFRELE(ira->ira_ipsec_action);
		}
		ASSERT(ira->ira_flags & IRAF_IPSEC_SECURE);
		ASSERT(ira->ira_ipsec_action == NULL);
		IPACT_REFHOLD(ap);
		ira->ira_ipsec_action = ap;
		break;	/* from switch */
	}
	return (ret);
}

static boolean_t
spd_match_inbound_ids(ipsec_latch_t *ipl, ipsa_t *sa)
{
	ASSERT(ipl->ipl_ids_latched == B_TRUE);
	return ipsid_equal(ipl->ipl_remote_cid, sa->ipsa_src_cid) &&
	    ipsid_equal(ipl->ipl_local_cid, sa->ipsa_dst_cid);
}

/*
 * Takes a latched conn and an inbound packet and returns a unique_id suitable
 * for SA comparisons.  Most of the time we will copy from the conn_t, but
 * there are cases when the conn_t is latched but it has wildcard selectors,
 * and then we need to fallback to scooping them out of the packet.
 *
 * Assume we'll never have 0 with a conn_t present, so use 0 as a failure.  We
 * can get away with this because we only have non-zero ports/proto for
 * latched conn_ts.
 *
 * Ideal candidate for an "inline" keyword, as we're JUST convoluted enough
 * to not be a nice macro.
 */
static uint64_t
conn_to_unique(conn_t *connp, mblk_t *data_mp, ipha_t *ipha, ip6_t *ip6h)
{
	ipsec_selector_t sel;
	uint8_t ulp = connp->conn_proto;

	ASSERT(connp->conn_latch_in_policy != NULL);

	if ((ulp == IPPROTO_TCP || ulp == IPPROTO_UDP || ulp == IPPROTO_SCTP) &&
	    (connp->conn_fport == 0 || connp->conn_lport == 0)) {
		/* Slow path - we gotta grab from the packet. */
		if (ipsec_init_inbound_sel(&sel, data_mp, ipha, ip6h,
		    SEL_NONE) != SELRET_SUCCESS) {
			/* Failure -> have caller free packet with ENOMEM. */
			return (0);
		}
		return (SA_UNIQUE_ID(sel.ips_remote_port, sel.ips_local_port,
		    sel.ips_protocol, 0));
	}

#ifdef DEBUG_NOT_UNTIL_6478464
	if (ipsec_init_inbound_sel(&sel, data_mp, ipha, ip6h, SEL_NONE) ==
	    SELRET_SUCCESS) {
		ASSERT(sel.ips_local_port == connp->conn_lport);
		ASSERT(sel.ips_remote_port == connp->conn_fport);
		ASSERT(sel.ips_protocol == connp->conn_proto);
	}
	ASSERT(connp->conn_proto != 0);
#endif

	return (SA_UNIQUE_ID(connp->conn_fport, connp->conn_lport, ulp, 0));
}

/*
 * Called to check policy on a latched connection.
 * Note that we don't dereference conn_latch or conn_ihere since the conn might
 * be closing. The caller passes a held ipsec_latch_t instead.
 */
static boolean_t
ipsec_check_ipsecin_latch(ip_recv_attr_t *ira, mblk_t *mp, ipsec_latch_t *ipl,
    ipsec_action_t *ap, ipha_t *ipha, ip6_t *ip6h, const char **reason,
    kstat_named_t **counter, conn_t *connp, netstack_t *ns)
{
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	ASSERT(ipl->ipl_ids_latched == B_TRUE);
	ASSERT(ira->ira_flags & IRAF_IPSEC_SECURE);

	if (!(ira->ira_flags & IRAF_LOOPBACK)) {
		/*
		 * Over loopback, there aren't real security associations,
		 * so there are neither identities nor "unique" values
		 * for us to check the packet against.
		 */
		if (ira->ira_ipsec_ah_sa != NULL) {
			if (!spd_match_inbound_ids(ipl,
			    ira->ira_ipsec_ah_sa)) {
				*counter = DROPPER(ipss, ipds_spd_ah_badid);
				*reason = "AH identity mismatch";
				return (B_FALSE);
			}
		}

		if (ira->ira_ipsec_esp_sa != NULL) {
			if (!spd_match_inbound_ids(ipl,
			    ira->ira_ipsec_esp_sa)) {
				*counter = DROPPER(ipss, ipds_spd_esp_badid);
				*reason = "ESP identity mismatch";
				return (B_FALSE);
			}
		}

		/*
		 * Can fudge pkt_unique from connp because we're latched.
		 * In DEBUG kernels (see conn_to_unique()'s implementation),
		 * verify this even if it REALLY slows things down.
		 */
		if (!ipsec_check_ipsecin_unique(ira, reason, counter,
		    conn_to_unique(connp, mp, ipha, ip6h), ns)) {
			return (B_FALSE);
		}
	}
	return (ipsec_check_ipsecin_action(ira, mp, ap, ipha, ip6h, reason,
	    counter, ns));
}

/*
 * Check to see whether this secured datagram meets the policy
 * constraints specified in ipsp.
 *
 * Called from ipsec_check_global_policy, and ipsec_check_inbound_policy.
 *
 * Consumes a reference to ipsp.
 * Returns the mblk if ok.
 */
static mblk_t *
ipsec_check_ipsecin_policy(mblk_t *data_mp, ipsec_policy_t *ipsp,
    ipha_t *ipha, ip6_t *ip6h, uint64_t pkt_unique, ip_recv_attr_t *ira,
    netstack_t *ns)
{
	ipsec_action_t *ap;
	const char *reason = "no policy actions found";
	ip_stack_t	*ipst = ns->netstack_ip;
	ipsec_stack_t	*ipss = ns->netstack_ipsec;
	kstat_named_t *counter;

	counter = DROPPER(ipss, ipds_spd_got_secure);

	ASSERT(ipsp != NULL);

	ASSERT((ipha == NULL && ip6h != NULL) ||
	    (ip6h == NULL && ipha != NULL));

	if (ira->ira_flags & IRAF_LOOPBACK)
		return (ipsec_check_loopback_policy(data_mp, ira, ipsp));

	ASSERT(ira->ira_flags & IRAF_IPSEC_SECURE);

	if (ira->ira_ipsec_action != NULL) {
		/*
		 * this can happen if we do a double policy-check on a packet
		 * Would be nice to be able to delete this test..
		 */
		IPACT_REFRELE(ira->ira_ipsec_action);
	}
	ASSERT(ira->ira_ipsec_action == NULL);

	if (!SA_IDS_MATCH(ira->ira_ipsec_ah_sa, ira->ira_ipsec_esp_sa)) {
		reason = "inbound AH and ESP identities differ";
		counter = DROPPER(ipss, ipds_spd_ahesp_diffid);
		goto drop;
	}

	if (!ipsec_check_ipsecin_unique(ira, &reason, &counter, pkt_unique,
	    ns))
		goto drop;

	/*
	 * Ok, now loop through the possible actions and see if any
	 * of them work for us.
	 */

	for (ap = ipsp->ipsp_act; ap != NULL; ap = ap->ipa_next) {
		if (ipsec_check_ipsecin_action(ira, data_mp, ap,
		    ipha, ip6h, &reason, &counter, ns)) {
			BUMP_MIB(&ipst->ips_ip_mib, ipsecInSucceeded);
			IPPOL_REFRELE(ipsp);
			return (data_mp);
		}
	}
drop:
	ipsec_rl_strlog(ns, IP_MOD_ID, 0, 0, SL_ERROR|SL_WARN|SL_CONSOLE,
	    "ipsec inbound policy mismatch: %s, packet dropped\n",
	    reason);
	IPPOL_REFRELE(ipsp);
	ASSERT(ira->ira_ipsec_action == NULL);
	BUMP_MIB(&ipst->ips_ip_mib, ipsecInFailed);
	ip_drop_packet(data_mp, B_TRUE, NULL, counter,
	    &ipss->ipsec_spd_dropper);
	return (NULL);
}

/*
 * sleazy prefix-length-based compare.
 * another inlining candidate..
 */
boolean_t
ip_addr_match(uint8_t *addr1, int pfxlen, in6_addr_t *addr2p)
{
	int offset = pfxlen>>3;
	int bitsleft = pfxlen & 7;
	uint8_t *addr2 = (uint8_t *)addr2p;

	/*
	 * and there was much evil..
	 * XXX should inline-expand the bcmp here and do this 32 bits
	 * or 64 bits at a time..
	 */
	return ((bcmp(addr1, addr2, offset) == 0) &&
	    ((bitsleft == 0) ||
	    (((addr1[offset] ^ addr2[offset]) & (0xff<<(8-bitsleft))) == 0)));
}

static ipsec_policy_t *
ipsec_find_policy_chain(ipsec_policy_t *best, ipsec_policy_t *chain,
    ipsec_selector_t *sel, boolean_t is_icmp_inv_acq)
{
	ipsec_selkey_t *isel;
	ipsec_policy_t *p;
	int bpri = best ? best->ipsp_prio : 0;

	for (p = chain; p != NULL; p = p->ipsp_hash.hash_next) {
		uint32_t valid;

		if (p->ipsp_prio <= bpri)
			continue;
		isel = &p->ipsp_sel->ipsl_key;
		valid = isel->ipsl_valid;

		if ((valid & IPSL_PROTOCOL) &&
		    (isel->ipsl_proto != sel->ips_protocol))
			continue;

		if ((valid & IPSL_REMOTE_ADDR) &&
		    !ip_addr_match((uint8_t *)&isel->ipsl_remote,
		    isel->ipsl_remote_pfxlen, &sel->ips_remote_addr_v6))
			continue;

		if ((valid & IPSL_LOCAL_ADDR) &&
		    !ip_addr_match((uint8_t *)&isel->ipsl_local,
		    isel->ipsl_local_pfxlen, &sel->ips_local_addr_v6))
			continue;

		if ((valid & IPSL_REMOTE_PORT) &&
		    isel->ipsl_rport != sel->ips_remote_port)
			continue;

		if ((valid & IPSL_LOCAL_PORT) &&
		    isel->ipsl_lport != sel->ips_local_port)
			continue;

		if (!is_icmp_inv_acq) {
			if ((valid & IPSL_ICMP_TYPE) &&
			    (isel->ipsl_icmp_type > sel->ips_icmp_type ||
			    isel->ipsl_icmp_type_end < sel->ips_icmp_type)) {
				continue;
			}

			if ((valid & IPSL_ICMP_CODE) &&
			    (isel->ipsl_icmp_code > sel->ips_icmp_code ||
			    isel->ipsl_icmp_code_end <
			    sel->ips_icmp_code)) {
				continue;
			}
		} else {
			/*
			 * special case for icmp inverse acquire
			 * we only want policies that aren't drop/pass
			 */
			if (p->ipsp_act->ipa_act.ipa_type != IPSEC_ACT_APPLY)
				continue;
		}

		/* we matched all the packet-port-field selectors! */
		best = p;
		bpri = p->ipsp_prio;
	}

	return (best);
}

/*
 * Try to find and return the best policy entry under a given policy
 * root for a given set of selectors; the first parameter "best" is
 * the current best policy so far.  If "best" is non-null, we have a
 * reference to it.  We return a reference to a policy; if that policy
 * is not the original "best", we need to release that reference
 * before returning.
 */
ipsec_policy_t *
ipsec_find_policy_head(ipsec_policy_t *best, ipsec_policy_head_t *head,
    int direction, ipsec_selector_t *sel)
{
	ipsec_policy_t *curbest;
	ipsec_policy_root_t *root;
	uint8_t is_icmp_inv_acq = sel->ips_is_icmp_inv_acq;
	int af = sel->ips_isv4 ? IPSEC_AF_V4 : IPSEC_AF_V6;

	curbest = best;
	root = &head->iph_root[direction];

#ifdef DEBUG
	if (is_icmp_inv_acq) {
		if (sel->ips_isv4) {
			if (sel->ips_protocol != IPPROTO_ICMP) {
				cmn_err(CE_WARN, "ipsec_find_policy_head:"
				    " expecting icmp, got %d",
				    sel->ips_protocol);
			}
		} else {
			if (sel->ips_protocol != IPPROTO_ICMPV6) {
				cmn_err(CE_WARN, "ipsec_find_policy_head:"
				    " expecting icmpv6, got %d",
				    sel->ips_protocol);
			}
		}
	}
#endif

	rw_enter(&head->iph_lock, RW_READER);

	if (root->ipr_nchains > 0) {
		curbest = ipsec_find_policy_chain(curbest,
		    root->ipr_hash[selector_hash(sel, root)].hash_head, sel,
		    is_icmp_inv_acq);
	}
	curbest = ipsec_find_policy_chain(curbest, root->ipr_nonhash[af], sel,
	    is_icmp_inv_acq);

	/*
	 * Adjust reference counts if we found anything new.
	 */
	if (curbest != best) {
		ASSERT(curbest != NULL);
		IPPOL_REFHOLD(curbest);

		if (best != NULL) {
			IPPOL_REFRELE(best);
		}
	}

	rw_exit(&head->iph_lock);

	return (curbest);
}

/*
 * Find the best system policy (either global or per-interface) which
 * applies to the given selector; look in all the relevant policy roots
 * to figure out which policy wins.
 *
 * Returns a reference to a policy; caller must release this
 * reference when done.
 */
ipsec_policy_t *
ipsec_find_policy(int direction, const conn_t *connp, ipsec_selector_t *sel,
    netstack_t *ns)
{
	ipsec_policy_t *p;
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	p = ipsec_find_policy_head(NULL, &ipss->ipsec_system_policy,
	    direction, sel);
	if ((connp != NULL) && (connp->conn_policy != NULL)) {
		p = ipsec_find_policy_head(p, connp->conn_policy,
		    direction, sel);
	}

	return (p);
}

/*
 * Check with global policy and see whether this inbound
 * packet meets the policy constraints.
 *
 * Locate appropriate policy from global policy, supplemented by the
 * conn's configured and/or cached policy if the conn is supplied.
 *
 * Dispatch to ipsec_check_ipsecin_policy if we have policy and an
 * encrypted packet to see if they match.
 *
 * Otherwise, see if the policy allows cleartext; if not, drop it on the
 * floor.
 */
mblk_t *
ipsec_check_global_policy(mblk_t *data_mp, conn_t *connp,
    ipha_t *ipha, ip6_t *ip6h, ip_recv_attr_t *ira, netstack_t *ns)
{
	ipsec_policy_t *p;
	ipsec_selector_t sel;
	boolean_t policy_present;
	kstat_named_t *counter;
	uint64_t pkt_unique;
	ip_stack_t	*ipst = ns->netstack_ip;
	ipsec_stack_t	*ipss = ns->netstack_ipsec;

	sel.ips_is_icmp_inv_acq = 0;

	ASSERT((ipha == NULL && ip6h != NULL) ||
	    (ip6h == NULL && ipha != NULL));

	if (ipha != NULL)
		policy_present = ipss->ipsec_inbound_v4_policy_present;
	else
		policy_present = ipss->ipsec_inbound_v6_policy_present;

	if (!policy_present && connp == NULL) {
		/*
		 * No global policy and no per-socket policy;
		 * just pass it back (but we shouldn't get here in that case)
		 */
		return (data_mp);
	}

	/*
	 * If we have cached policy, use it.
	 * Otherwise consult system policy.
	 */
	if ((connp != NULL) && (connp->conn_latch != NULL)) {
		p = connp->conn_latch_in_policy;
		if (p != NULL) {
			IPPOL_REFHOLD(p);
		}
		/*
		 * Fudge sel for UNIQUE_ID setting below.
		 */
		pkt_unique = conn_to_unique(connp, data_mp, ipha, ip6h);
	} else {
		/* Initialize the ports in the selector */
		if (ipsec_init_inbound_sel(&sel, data_mp, ipha, ip6h,
		    SEL_NONE) == SELRET_NOMEM) {
			/*
			 * Technically not a policy mismatch, but it is
			 * an internal failure.
			 */
			ipsec_log_policy_failure(IPSEC_POLICY_MISMATCH,
			    "ipsec_init_inbound_sel", ipha, ip6h, B_TRUE, ns);
			counter = DROPPER(ipss, ipds_spd_nomem);
			goto fail;
		}

		/*
		 * Find the policy which best applies.
		 *
		 * If we find global policy, we should look at both
		 * local policy and global policy and see which is
		 * stronger and match accordingly.
		 *
		 * If we don't find a global policy, check with
		 * local policy alone.
		 */

		p = ipsec_find_policy(IPSEC_TYPE_INBOUND, connp, &sel, ns);
		pkt_unique = SA_UNIQUE_ID(sel.ips_remote_port,
		    sel.ips_local_port, sel.ips_protocol, 0);
	}

	if (p == NULL) {
		if (!(ira->ira_flags & IRAF_IPSEC_SECURE)) {
			/*
			 * We have no policy; default to succeeding.
			 * XXX paranoid system design doesn't do this.
			 */
			BUMP_MIB(&ipst->ips_ip_mib, ipsecInSucceeded);
			return (data_mp);
		} else {
			counter = DROPPER(ipss, ipds_spd_got_secure);
			ipsec_log_policy_failure(IPSEC_POLICY_NOT_NEEDED,
			    "ipsec_check_global_policy", ipha, ip6h, B_TRUE,
			    ns);
			goto fail;
		}
	}
	if (ira->ira_flags & IRAF_IPSEC_SECURE) {
		return (ipsec_check_ipsecin_policy(data_mp, p, ipha, ip6h,
		    pkt_unique, ira, ns));
	}
	if (p->ipsp_act->ipa_allow_clear) {
		BUMP_MIB(&ipst->ips_ip_mib, ipsecInSucceeded);
		IPPOL_REFRELE(p);
		return (data_mp);
	}
	IPPOL_REFRELE(p);
	/*
	 * If we reach here, we will drop the packet because it failed the
	 * global policy check because the packet was cleartext, and it
	 * should not have been.
	 */
	ipsec_log_policy_failure(IPSEC_POLICY_MISMATCH,
	    "ipsec_check_global_policy", ipha, ip6h, B_FALSE, ns);
	counter = DROPPER(ipss, ipds_spd_got_clear);

fail:
	ip_drop_packet(data_mp, B_TRUE, NULL, counter,
	    &ipss->ipsec_spd_dropper);
	BUMP_MIB(&ipst->ips_ip_mib, ipsecInFailed);
	return (NULL);
}

/*
 * We check whether an inbound datagram is a valid one
 * to accept in clear. If it is secure, it is the job
 * of IPSEC to log information appropriately if it
 * suspects that it may not be the real one.
 *
 * It is called only while fanning out to the ULP
 * where ULP accepts only secure data and the incoming
 * is clear. Usually we never accept clear datagrams in
 * such cases. ICMP is the only exception.
 *
 * NOTE : We don't call this function if the client (ULP)
 * is willing to accept things in clear.
 */
boolean_t
ipsec_inbound_accept_clear(mblk_t *mp, ipha_t *ipha, ip6_t *ip6h)
{
	ushort_t iph_hdr_length;
	icmph_t *icmph;
	icmp6_t *icmp6;
	uint8_t *nexthdrp;

	ASSERT((ipha != NULL && ip6h == NULL) ||
	    (ipha == NULL && ip6h != NULL));

	if (ip6h != NULL) {
		iph_hdr_length = ip_hdr_length_v6(mp, ip6h);
		if (!ip_hdr_length_nexthdr_v6(mp, ip6h, &iph_hdr_length,
		    &nexthdrp)) {
			return (B_FALSE);
		}
		if (*nexthdrp != IPPROTO_ICMPV6)
			return (B_FALSE);
		icmp6 = (icmp6_t *)(&mp->b_rptr[iph_hdr_length]);
		/* Match IPv6 ICMP policy as closely as IPv4 as possible. */
		switch (icmp6->icmp6_type) {
		case ICMP6_PARAM_PROB:
			/* Corresponds to port/proto unreach in IPv4. */
		case ICMP6_ECHO_REQUEST:
			/* Just like IPv4. */
			return (B_FALSE);

		case MLD_LISTENER_QUERY:
		case MLD_LISTENER_REPORT:
		case MLD_LISTENER_REDUCTION:
			/*
			 * XXX Seperate NDD in IPv4 what about here?
			 * Plus, mcast is important to ND.
			 */
		case ICMP6_DST_UNREACH:
			/* Corresponds to HOST/NET unreachable in IPv4. */
		case ICMP6_PACKET_TOO_BIG:
		case ICMP6_ECHO_REPLY:
			/* These are trusted in IPv4. */
		case ND_ROUTER_SOLICIT:
		case ND_ROUTER_ADVERT:
		case ND_NEIGHBOR_SOLICIT:
		case ND_NEIGHBOR_ADVERT:
		case ND_REDIRECT:
			/* Trust ND messages for now. */
		case ICMP6_TIME_EXCEEDED:
		default:
			return (B_TRUE);
		}
	} else {
		/*
		 * If it is not ICMP, fail this request.
		 */
		if (ipha->ipha_protocol != IPPROTO_ICMP) {
#ifdef FRAGCACHE_DEBUG
			cmn_err(CE_WARN, "Dropping - ipha_proto = %d\n",
			    ipha->ipha_protocol);
#endif
			return (B_FALSE);
		}
		iph_hdr_length = IPH_HDR_LENGTH(ipha);
		icmph = (icmph_t *)&mp->b_rptr[iph_hdr_length];
		/*
		 * It is an insecure icmp message. Check to see whether we are
		 * willing to accept this one.
		 */

		switch (icmph->icmph_type) {
		case ICMP_ECHO_REPLY:
		case ICMP_TIME_STAMP_REPLY:
		case ICMP_INFO_REPLY:
		case ICMP_ROUTER_ADVERTISEMENT:
			/*
			 * We should not encourage clear replies if this
			 * client expects secure. If somebody is replying
			 * in clear some mailicious user watching both the
			 * request and reply, can do chosen-plain-text attacks.
			 * With global policy we might be just expecting secure
			 * but sending out clear. We don't know what the right
			 * thing is. We can't do much here as we can't control
			 * the sender here. Till we are sure of what to do,
			 * accept them.
			 */
			return (B_TRUE);
		case ICMP_ECHO_REQUEST:
		case ICMP_TIME_STAMP_REQUEST:
		case ICMP_INFO_REQUEST:
		case ICMP_ADDRESS_MASK_REQUEST:
		case ICMP_ROUTER_SOLICITATION:
		case ICMP_ADDRESS_MASK_REPLY:
			/*
			 * Don't accept this as somebody could be sending
			 * us plain text to get encrypted data. If we reply,
			 * it will lead to chosen plain text attack.
			 */
			return (B_FALSE);
		case ICMP_DEST_UNREACHABLE:
			switch (icmph->icmph_code) {
			case ICMP_FRAGMENTATION_NEEDED:
				/*
				 * Be in sync with icmp_inbound, where we have
				 * already set dce_pmtu
				 */
#ifdef FRAGCACHE_DEBUG
			cmn_err(CE_WARN, "ICMP frag needed\n");
#endif
				return (B_TRUE);
			case ICMP_HOST_UNREACHABLE:
			case ICMP_NET_UNREACHABLE:
				/*
				 * By accepting, we could reset a connection.
				 * How do we solve the problem of some
				 * intermediate router sending in-secure ICMP
				 * messages ?
				 */
				return (B_TRUE);
			case ICMP_PORT_UNREACHABLE:
			case ICMP_PROTOCOL_UNREACHABLE:
			default :
				return (B_FALSE);
			}
		case ICMP_SOURCE_QUENCH:
			/*
			 * If this is an attack, TCP will slow start
			 * because of this. Is it very harmful ?
			 */
			return (B_TRUE);
		case ICMP_PARAM_PROBLEM:
			return (B_FALSE);
		case ICMP_TIME_EXCEEDED:
			return (B_TRUE);
		case ICMP_REDIRECT:
			return (B_FALSE);
		default :
			return (B_FALSE);
		}
	}
}

void
ipsec_latch_ids(ipsec_latch_t *ipl, ipsid_t *local, ipsid_t *remote)
{
	mutex_enter(&ipl->ipl_lock);

	if (ipl->ipl_ids_latched) {
		/* I lost, someone else got here before me */
		mutex_exit(&ipl->ipl_lock);
		return;
	}

	if (local != NULL)
		IPSID_REFHOLD(local);
	if (remote != NULL)
		IPSID_REFHOLD(remote);

	ipl->ipl_local_cid = local;
	ipl->ipl_remote_cid = remote;
	ipl->ipl_ids_latched = B_TRUE;
	mutex_exit(&ipl->ipl_lock);
}

void
ipsec_latch_inbound(conn_t *connp, ip_recv_attr_t *ira)
{
	ipsa_t *sa;
	ipsec_latch_t *ipl = connp->conn_latch;

	if (!ipl->ipl_ids_latched) {
		ipsid_t *local = NULL;
		ipsid_t *remote = NULL;

		if (!(ira->ira_flags & IRAF_LOOPBACK)) {
			ASSERT(ira->ira_flags & IRAF_IPSEC_SECURE);
			if (ira->ira_ipsec_esp_sa != NULL)
				sa = ira->ira_ipsec_esp_sa;
			else
				sa = ira->ira_ipsec_ah_sa;
			ASSERT(sa != NULL);
			local = sa->ipsa_dst_cid;
			remote = sa->ipsa_src_cid;
		}
		ipsec_latch_ids(ipl, local, remote);
	}
	if (ira->ira_flags & IRAF_IPSEC_SECURE) {
		if (connp->conn_latch_in_action != NULL) {
			/*
			 * Previously cached action.  This is probably
			 * harmless, but in DEBUG kernels, check for
			 * action equality.
			 *
			 * Preserve the existing action to preserve latch
			 * invariance.
			 */
			ASSERT(connp->conn_latch_in_action ==
			    ira->ira_ipsec_action);
			return;
		}
		connp->conn_latch_in_action = ira->ira_ipsec_action;
		IPACT_REFHOLD(connp->conn_latch_in_action);
	}
}

/*
 * Check whether the policy constraints are met either for an
 * inbound datagram; called from IP in numerous places.
 *
 * Note that this is not a chokepoint for inbound policy checks;
 * see also ipsec_check_ipsecin_latch() and ipsec_check_global_policy()
 */
mblk_t *
ipsec_check_inbound_policy(mblk_t *mp, conn_t *connp,
    ipha_t *ipha, ip6_t *ip6h, ip_recv_attr_t *ira)
{
	boolean_t	ret;
	ipsec_latch_t	*ipl;
	ipsec_action_t	*ap;
	uint64_t	unique_id;
	ipsec_stack_t	*ipss;
	ip_stack_t	*ipst;
	netstack_t	*ns;
	ipsec_policy_head_t *policy_head;
	ipsec_policy_t	*p = NULL;

	ASSERT(connp != NULL);
	ns = connp->conn_netstack;
	ipss = ns->netstack_ipsec;
	ipst = ns->netstack_ip;

	if (!(ira->ira_flags & IRAF_IPSEC_SECURE)) {
		/*
		 * This is the case where the incoming datagram is
		 * cleartext and we need to see whether this client
		 * would like to receive such untrustworthy things from
		 * the wire.
		 */
		ASSERT(mp != NULL);

		mutex_enter(&connp->conn_lock);
		if (connp->conn_state_flags & CONN_CONDEMNED) {
			mutex_exit(&connp->conn_lock);
			ip_drop_packet(mp, B_TRUE, NULL,
			    DROPPER(ipss, ipds_spd_got_clear),
			    &ipss->ipsec_spd_dropper);
			BUMP_MIB(&ipst->ips_ip_mib, ipsecInFailed);
			return (NULL);
		}
		if (connp->conn_latch != NULL) {
			/* Hold a reference in case the conn is closing */
			p = connp->conn_latch_in_policy;
			if (p != NULL)
				IPPOL_REFHOLD(p);
			mutex_exit(&connp->conn_lock);
			/*
			 * Policy is cached in the conn.
			 */
			if (p != NULL && !p->ipsp_act->ipa_allow_clear) {
				ret = ipsec_inbound_accept_clear(mp,
				    ipha, ip6h);
				if (ret) {
					BUMP_MIB(&ipst->ips_ip_mib,
					    ipsecInSucceeded);
					IPPOL_REFRELE(p);
					return (mp);
				} else {
					ipsec_log_policy_failure(
					    IPSEC_POLICY_MISMATCH,
					    "ipsec_check_inbound_policy", ipha,
					    ip6h, B_FALSE, ns);
					ip_drop_packet(mp, B_TRUE, NULL,
					    DROPPER(ipss, ipds_spd_got_clear),
					    &ipss->ipsec_spd_dropper);
					BUMP_MIB(&ipst->ips_ip_mib,
					    ipsecInFailed);
					IPPOL_REFRELE(p);
					return (NULL);
				}
			} else {
				BUMP_MIB(&ipst->ips_ip_mib, ipsecInSucceeded);
				if (p != NULL)
					IPPOL_REFRELE(p);
				return (mp);
			}
		} else {
			policy_head = connp->conn_policy;

			/* Hold a reference in case the conn is closing */
			if (policy_head != NULL)
				IPPH_REFHOLD(policy_head);
			mutex_exit(&connp->conn_lock);
			/*
			 * As this is a non-hardbound connection we need
			 * to look at both per-socket policy and global
			 * policy.
			 */
			mp = ipsec_check_global_policy(mp, connp,
			    ipha, ip6h, ira, ns);
			if (policy_head != NULL)
				IPPH_REFRELE(policy_head, ns);
			return (mp);
		}
	}

	mutex_enter(&connp->conn_lock);
	/* Connection is closing */
	if (connp->conn_state_flags & CONN_CONDEMNED) {
		mutex_exit(&connp->conn_lock);
		ip_drop_packet(mp, B_TRUE, NULL,
		    DROPPER(ipss, ipds_spd_got_clear),
		    &ipss->ipsec_spd_dropper);
		BUMP_MIB(&ipst->ips_ip_mib, ipsecInFailed);
		return (NULL);
	}

	/*
	 * Once a connection is latched it remains so for life, the conn_latch
	 * pointer on the conn has not changed, simply initializing ipl here
	 * as the earlier initialization was done only in the cleartext case.
	 */
	if ((ipl = connp->conn_latch) == NULL) {
		mblk_t *retmp;
		policy_head = connp->conn_policy;

		/* Hold a reference in case the conn is closing */
		if (policy_head != NULL)
			IPPH_REFHOLD(policy_head);
		mutex_exit(&connp->conn_lock);
		/*
		 * We don't have policies cached in the conn
		 * for this stream. So, look at the global
		 * policy. It will check against conn or global
		 * depending on whichever is stronger.
		 */
		retmp = ipsec_check_global_policy(mp, connp,
		    ipha, ip6h, ira, ns);
		if (policy_head != NULL)
			IPPH_REFRELE(policy_head, ns);
		return (retmp);
	}

	IPLATCH_REFHOLD(ipl);
	/* Hold reference on conn_latch_in_action in case conn is closing */
	ap = connp->conn_latch_in_action;
	if (ap != NULL)
		IPACT_REFHOLD(ap);
	mutex_exit(&connp->conn_lock);

	if (ap != NULL) {
		/* Policy is cached & latched; fast(er) path */
		const char *reason;
		kstat_named_t *counter;

		if (ipsec_check_ipsecin_latch(ira, mp, ipl, ap,
		    ipha, ip6h, &reason, &counter, connp, ns)) {
			BUMP_MIB(&ipst->ips_ip_mib, ipsecInSucceeded);
			IPLATCH_REFRELE(ipl);
			IPACT_REFRELE(ap);
			return (mp);
		}
		ipsec_rl_strlog(ns, IP_MOD_ID, 0, 0,
		    SL_ERROR|SL_WARN|SL_CONSOLE,
		    "ipsec inbound policy mismatch: %s, packet dropped\n",
		    reason);
		ip_drop_packet(mp, B_TRUE, NULL, counter,
		    &ipss->ipsec_spd_dropper);
		BUMP_MIB(&ipst->ips_ip_mib, ipsecInFailed);
		IPLATCH_REFRELE(ipl);
		IPACT_REFRELE(ap);
		return (NULL);
	}
	if ((p = connp->conn_latch_in_policy) == NULL) {
		ipsec_weird_null_inbound_policy++;
		IPLATCH_REFRELE(ipl);
		return (mp);
	}

	unique_id = conn_to_unique(connp, mp, ipha, ip6h);
	IPPOL_REFHOLD(p);
	mp = ipsec_check_ipsecin_policy(mp, p, ipha, ip6h, unique_id, ira, ns);
	/*
	 * NOTE: ipsecIn{Failed,Succeeeded} bumped by
	 * ipsec_check_ipsecin_policy().
	 */
	if (mp != NULL)
		ipsec_latch_inbound(connp, ira);
	IPLATCH_REFRELE(ipl);
	return (mp);
}

/*
 * Handle all sorts of cases like tunnel-mode and ICMP.
 */
static int
prepended_length(mblk_t *mp, uintptr_t hptr)
{
	int rc = 0;

	while (mp != NULL) {
		if (hptr >= (uintptr_t)mp->b_rptr && hptr <
		    (uintptr_t)mp->b_wptr) {
			rc += (int)(hptr - (uintptr_t)mp->b_rptr);
			break;	/* out of while loop */
		}
		rc += (int)MBLKL(mp);
		mp = mp->b_cont;
	}

	if (mp == NULL) {
		/*
		 * IF (big IF) we make it here by naturally exiting the loop,
		 * then ip6h isn't in the mblk chain "mp" at all.
		 *
		 * The only case where this happens is with a reversed IP
		 * header that gets passed up by inbound ICMP processing.
		 * This unfortunately triggers longstanding bug 6478464.  For
		 * now, just pass up 0 for the answer.
		 */
#ifdef DEBUG_NOT_UNTIL_6478464
		ASSERT(mp != NULL);
#endif
		rc = 0;
	}

	return (rc);
}

/*
 * Returns:
 *
 * SELRET_NOMEM --> msgpullup() needed to gather things failed.
 * SELRET_BADPKT --> If we're being called after tunnel-mode fragment
 *		     gathering, the initial fragment is too short for
 *		     useful data.  Only returned if SEL_TUNNEL_FIRSTFRAG is
 *		     set.
 * SELRET_SUCCESS --> "sel" now has initialized IPsec selector data.
 * SELRET_TUNFRAG --> This is a fragment in a tunnel-mode packet.  Caller
 *		      should put this packet in a fragment-gathering queue.
 *		      Only returned if SEL_TUNNEL_MODE and SEL_PORT_POLICY
 *		      is set.
 *
 * Note that ipha/ip6h can be in a different mblk (mp->b_cont) in the case
 * of tunneled packets.
 * Also, mp->b_rptr can be an ICMP error where ipha/ip6h is the packet in
 * error past the ICMP error.
 */
static selret_t
ipsec_init_inbound_sel(ipsec_selector_t *sel, mblk_t *mp, ipha_t *ipha,
    ip6_t *ip6h, uint8_t sel_flags)
{
	uint16_t *ports;
	int outer_hdr_len = 0;	/* For ICMP or tunnel-mode cases... */
	ushort_t hdr_len;
	mblk_t *spare_mp = NULL;
	uint8_t *nexthdrp, *transportp;
	uint8_t nexthdr;
	uint8_t icmp_proto;
	ip_pkt_t ipp;
	boolean_t port_policy_present = (sel_flags & SEL_PORT_POLICY);
	boolean_t is_icmp = (sel_flags & SEL_IS_ICMP);
	boolean_t tunnel_mode = (sel_flags & SEL_TUNNEL_MODE);
	boolean_t post_frag = (sel_flags & SEL_POST_FRAG);

	ASSERT((ipha == NULL && ip6h != NULL) ||
	    (ipha != NULL && ip6h == NULL));

	if (ip6h != NULL) {
		outer_hdr_len = prepended_length(mp, (uintptr_t)ip6h);
		nexthdr = ip6h->ip6_nxt;
		icmp_proto = IPPROTO_ICMPV6;
		sel->ips_isv4 = B_FALSE;
		sel->ips_local_addr_v6 = ip6h->ip6_dst;
		sel->ips_remote_addr_v6 = ip6h->ip6_src;

		bzero(&ipp, sizeof (ipp));
		(void) ip_find_hdr_v6(mp, ip6h, B_FALSE, &ipp, NULL);

		switch (nexthdr) {
		case IPPROTO_HOPOPTS:
		case IPPROTO_ROUTING:
		case IPPROTO_DSTOPTS:
		case IPPROTO_FRAGMENT:
			/*
			 * Use ip_hdr_length_nexthdr_v6().  And have a spare
			 * mblk that's contiguous to feed it
			 */
			if ((spare_mp = msgpullup(mp, -1)) == NULL)
				return (SELRET_NOMEM);
			if (!ip_hdr_length_nexthdr_v6(spare_mp,
			    (ip6_t *)(spare_mp->b_rptr + outer_hdr_len),
			    &hdr_len, &nexthdrp)) {
				/* Malformed packet - caller frees. */
				ipsec_freemsg_chain(spare_mp);
				return (SELRET_BADPKT);
			}
			nexthdr = *nexthdrp;
			/* We can just extract based on hdr_len now. */
			break;
		default:
			hdr_len = IPV6_HDR_LEN;
			break;
		}

		if (port_policy_present && IS_V6_FRAGMENT(ipp) && !is_icmp) {
			/* IPv6 Fragment */
			ipsec_freemsg_chain(spare_mp);
			return (SELRET_TUNFRAG);
		}
		transportp = (uint8_t *)ip6h + hdr_len;
	} else {
		outer_hdr_len = prepended_length(mp, (uintptr_t)ipha);
		icmp_proto = IPPROTO_ICMP;
		sel->ips_isv4 = B_TRUE;
		sel->ips_local_addr_v4 = ipha->ipha_dst;
		sel->ips_remote_addr_v4 = ipha->ipha_src;
		nexthdr = ipha->ipha_protocol;
		hdr_len = IPH_HDR_LENGTH(ipha);

		if (port_policy_present &&
		    IS_V4_FRAGMENT(ipha->ipha_fragment_offset_and_flags) &&
		    !is_icmp) {
			/* IPv4 Fragment */
			ipsec_freemsg_chain(spare_mp);
			return (SELRET_TUNFRAG);
		}
		transportp = (uint8_t *)ipha + hdr_len;
	}
	sel->ips_protocol = nexthdr;

	if ((nexthdr != IPPROTO_TCP && nexthdr != IPPROTO_UDP &&
	    nexthdr != IPPROTO_SCTP && nexthdr != icmp_proto) ||
	    (!port_policy_present && !post_frag && tunnel_mode)) {
		sel->ips_remote_port = sel->ips_local_port = 0;
		ipsec_freemsg_chain(spare_mp);
		return (SELRET_SUCCESS);
	}

	if (transportp + 4 > mp->b_wptr) {
		/* If we didn't pullup a copy already, do so now. */
		/*
		 * XXX performance, will upper-layers frequently split TCP/UDP
		 * apart from IP or options?  If so, perhaps we should revisit
		 * the spare_mp strategy.
		 */
		ipsec_hdr_pullup_needed++;
		if (spare_mp == NULL &&
		    (spare_mp = msgpullup(mp, -1)) == NULL) {
			return (SELRET_NOMEM);
		}
		transportp = &spare_mp->b_rptr[hdr_len + outer_hdr_len];