usr/src/uts/common/fs/nfs/nfs_subr.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 2006 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 *
 *	Copyright (c) 1983,1984,1985,1986,1987,1988,1989  AT&T.
 *	All rights reserved.
 */

#pragma ident	"%Z%%M%	%I%	%E% SMI"

#include <sys/param.h>
#include <sys/types.h>
#include <sys/systm.h>
#include <sys/cred_impl.h>
#include <sys/proc.h>
#include <sys/user.h>
#include <sys/time.h>
#include <sys/buf.h>
#include <sys/vfs.h>
#include <sys/vnode.h>
#include <sys/socket.h>
#include <sys/uio.h>
#include <sys/tiuser.h>
#include <sys/swap.h>
#include <sys/errno.h>
#include <sys/debug.h>
#include <sys/kmem.h>
#include <sys/kstat.h>
#include <sys/cmn_err.h>
#include <sys/vtrace.h>
#include <sys/session.h>
#include <sys/dnlc.h>
#include <sys/bitmap.h>
#include <sys/acl.h>
#include <sys/ddi.h>
#include <sys/pathname.h>
#include <sys/flock.h>
#include <sys/dirent.h>
#include <sys/flock.h>
#include <sys/callb.h>
#include <sys/atomic.h>
#include <sys/list.h>
#include <sys/tsol/tnet.h>
#include <sys/priv.h>

#include <inet/ip6.h>

#include <rpc/types.h>
#include <rpc/xdr.h>
#include <rpc/auth.h>
#include <rpc/clnt.h>

#include <nfs/nfs.h>
#include <nfs/nfs4.h>
#include <nfs/nfs_clnt.h>
#include <nfs/rnode.h>
#include <nfs/nfs_acl.h>

/*
 * The hash queues for the access to active and cached rnodes
 * are organized as doubly linked lists.  A reader/writer lock
 * for each hash bucket is used to control access and to synchronize
 * lookups, additions, and deletions from the hash queue.
 *
 * The rnode freelist is organized as a doubly linked list with
 * a head pointer.  Additions and deletions are synchronized via
 * a single mutex.
 *
 * In order to add an rnode to the free list, it must be hashed into
 * a hash queue and the exclusive lock to the hash queue be held.
 * If an rnode is not hashed into a hash queue, then it is destroyed
 * because it represents no valuable information that can be reused
 * about the file.  The exclusive lock to the hash queue must be
 * held in order to prevent a lookup in the hash queue from finding
 * the rnode and using it and assuming that the rnode is not on the
 * freelist.  The lookup in the hash queue will have the hash queue
 * locked, either exclusive or shared.
 *
 * The vnode reference count for each rnode is not allowed to drop
 * below 1.  This prevents external entities, such as the VM
 * subsystem, from acquiring references to vnodes already on the
 * freelist and then trying to place them back on the freelist
 * when their reference is released.  This means that the when an
 * rnode is looked up in the hash queues, then either the rnode
 * is removed from the freelist and that reference is tranfered to
 * the new reference or the vnode reference count must be incremented
 * accordingly.  The mutex for the freelist must be held in order to
 * accurately test to see if the rnode is on the freelist or not.
 * The hash queue lock might be held shared and it is possible that
 * two different threads may race to remove the rnode from the
 * freelist.  This race can be resolved by holding the mutex for the
 * freelist.  Please note that the mutex for the freelist does not
 * need to held if the rnode is not on the freelist.  It can not be
 * placed on the freelist due to the requirement that the thread
 * putting the rnode on the freelist must hold the exclusive lock
 * to the hash queue and the thread doing the lookup in the hash
 * queue is holding either a shared or exclusive lock to the hash
 * queue.
 *
 * The lock ordering is:
 *
 *	hash bucket lock -> vnode lock
 *	hash bucket lock -> freelist lock
 */
static rhashq_t *rtable;

static kmutex_t rpfreelist_lock;
static rnode_t *rpfreelist = NULL;
static long rnew = 0;
long nrnode = 0;

static int rtablesize;
static int rtablemask;

static int hashlen = 4;

static struct kmem_cache *rnode_cache;

/*
 * Mutex to protect the following variables:
 *	nfs_major
 *	nfs_minor
 */
kmutex_t nfs_minor_lock;
int nfs_major;
int nfs_minor;

/* Do we allow preepoch (negative) time values otw? */
bool_t nfs_allow_preepoch_time = FALSE;	/* default: do not allow preepoch */

/*
 * Access cache
 */
static acache_hash_t *acache;
static long nacache;	/* used strictly to size the number of hash queues */

static int acachesize;
static int acachemask;
static struct kmem_cache *acache_cache;

/*
 * Client side utilities
 */

/*
 * client side statistics
 */
static const struct clstat clstat_tmpl = {
	{ "calls",	KSTAT_DATA_UINT64 },
	{ "badcalls",	KSTAT_DATA_UINT64 },
	{ "clgets",	KSTAT_DATA_UINT64 },
	{ "cltoomany",	KSTAT_DATA_UINT64 },
#ifdef DEBUG
	{ "clalloc",	KSTAT_DATA_UINT64 },
	{ "noresponse",	KSTAT_DATA_UINT64 },
	{ "failover",	KSTAT_DATA_UINT64 },
	{ "remap",	KSTAT_DATA_UINT64 },
#endif
};

/*
 * The following are statistics that describe behavior of the system as a whole
 * and doesn't correspond to any one particular zone.
 */
#ifdef DEBUG
static struct clstat_debug {
	kstat_named_t	nrnode;			/* number of allocated rnodes */
	kstat_named_t	access;			/* size of access cache */
	kstat_named_t	dirent;			/* size of readdir cache */
	kstat_named_t	dirents;		/* size of readdir buf cache */
	kstat_named_t	reclaim;		/* number of reclaims */
	kstat_named_t	clreclaim;		/* number of cl reclaims */
	kstat_named_t	f_reclaim;		/* number of free reclaims */
	kstat_named_t	a_reclaim;		/* number of active reclaims */
	kstat_named_t	r_reclaim;		/* number of rnode reclaims */
	kstat_named_t	rpath;			/* bytes used to store rpaths */
} clstat_debug = {
	{ "nrnode",	KSTAT_DATA_UINT64 },
	{ "access",	KSTAT_DATA_UINT64 },
	{ "dirent",	KSTAT_DATA_UINT64 },
	{ "dirents",	KSTAT_DATA_UINT64 },
	{ "reclaim",	KSTAT_DATA_UINT64 },
	{ "clreclaim",	KSTAT_DATA_UINT64 },
	{ "f_reclaim",	KSTAT_DATA_UINT64 },
	{ "a_reclaim",	KSTAT_DATA_UINT64 },
	{ "r_reclaim",	KSTAT_DATA_UINT64 },
	{ "r_path",	KSTAT_DATA_UINT64 },
};
#endif	/* DEBUG */

/*
 * We keep a global list of per-zone client data, so we can clean up all zones
 * if we get low on memory.
 */
static list_t nfs_clnt_list;
static kmutex_t nfs_clnt_list_lock;
static zone_key_t nfsclnt_zone_key;

static struct kmem_cache *chtab_cache;

/*
 * Some servers do not properly update the attributes of the
 * directory when changes are made.  To allow interoperability
 * with these broken servers, the nfs_disable_rddir_cache
 * parameter must be set in /etc/system
 */
int nfs_disable_rddir_cache = 0;

int		clget(clinfo_t *, servinfo_t *, cred_t *, CLIENT **,
		    struct chtab **);
void		clfree(CLIENT *, struct chtab *);
static int	acl_clget(mntinfo_t *, servinfo_t *, cred_t *, CLIENT **,
		    struct chtab **, struct nfs_clnt *);
static int	nfs_clget(mntinfo_t *, servinfo_t *, cred_t *, CLIENT **,
		    struct chtab **, struct nfs_clnt *);
static void	clreclaim(void *);
static int	nfs_feedback(int, int, mntinfo_t *);
static int	rfscall(mntinfo_t *, rpcproc_t, xdrproc_t, caddr_t, xdrproc_t,
		    caddr_t, cred_t *, int *, enum clnt_stat *, int,
		    failinfo_t *);
static int	aclcall(mntinfo_t *, rpcproc_t, xdrproc_t, caddr_t, xdrproc_t,
		    caddr_t, cred_t *, int *, int, failinfo_t *);
static void	rinactive(rnode_t *, cred_t *);
static int	rtablehash(nfs_fhandle *);
static vnode_t	*make_rnode(nfs_fhandle *, rhashq_t *, struct vfs *,
		    struct vnodeops *,
		    int (*)(vnode_t *, page_t *, u_offset_t *, size_t *, int,
			cred_t *),
		    int (*)(const void *, const void *), int *, cred_t *,
		    char *, char *);
static void	rp_rmfree(rnode_t *);
static void	rp_addhash(rnode_t *);
static void	rp_rmhash_locked(rnode_t *);
static rnode_t	*rfind(rhashq_t *, nfs_fhandle *, struct vfs *);
static void	destroy_rnode(rnode_t *);
static void	rddir_cache_free(rddir_cache *);
static int	nfs_free_data_reclaim(rnode_t *);
static int	nfs_active_data_reclaim(rnode_t *);
static int	nfs_free_reclaim(void);
static int	nfs_active_reclaim(void);
static int	nfs_rnode_reclaim(void);
static void	nfs_reclaim(void *);
static int	failover_safe(failinfo_t *);
static void	failover_newserver(mntinfo_t *mi);
static void	failover_thread(mntinfo_t *mi);
static int	failover_wait(mntinfo_t *);
static int	failover_remap(failinfo_t *);
static int	failover_lookup(char *, vnode_t *,
		    int (*)(vnode_t *, char *, vnode_t **,
			struct pathname *, int, vnode_t *, cred_t *, int),
		    int (*)(vnode_t *, vnode_t **, bool_t, cred_t *, int),
		    vnode_t **);
static void	nfs_free_r_path(rnode_t *);
static void	nfs_set_vroot(vnode_t *);
static char	*nfs_getsrvnames(mntinfo_t *, size_t *);

/*
 * from rpcsec module (common/rpcsec)
 */
extern int sec_clnt_geth(CLIENT *, struct sec_data *, cred_t *, AUTH **);
extern void sec_clnt_freeh(AUTH *);
extern void sec_clnt_freeinfo(struct sec_data *);

/*
 * used in mount policy
 */
extern ts_label_t *getflabel_cipso(vfs_t *);

/*
 * EIO or EINTR are not recoverable errors.
 */
#define	IS_RECOVERABLE_ERROR(error)	!((error == EINTR) || (error == EIO))

/*
 * Common handle get program for NFS, NFS ACL, and NFS AUTH client.
 */
static int
clget_impl(clinfo_t *ci, servinfo_t *svp, cred_t *cr, CLIENT **newcl,
    struct chtab **chp, struct nfs_clnt *nfscl)
{
	struct chhead *ch, *newch;
	struct chhead **plistp;
	struct chtab *cp;
	int error;
	k_sigset_t smask;

	if (newcl == NULL || chp == NULL || ci == NULL)
		return (EINVAL);

	*newcl = NULL;
	*chp = NULL;

	/*
	 * Find an unused handle or create one
	 */
	newch = NULL;
	nfscl->nfscl_stat.clgets.value.ui64++;
top:
	/*
	 * Find the correct entry in the cache to check for free
	 * client handles.  The search is based on the RPC program
	 * number, program version number, dev_t for the transport
	 * device, and the protocol family.
	 */
	mutex_enter(&nfscl->nfscl_chtable_lock);
	plistp = &nfscl->nfscl_chtable;
	for (ch = nfscl->nfscl_chtable; ch != NULL; ch = ch->ch_next) {
		if (ch->ch_prog == ci->cl_prog &&
		    ch->ch_vers == ci->cl_vers &&
		    ch->ch_dev == svp->sv_knconf->knc_rdev &&
		    (strcmp(ch->ch_protofmly,
			svp->sv_knconf->knc_protofmly) == 0))
			break;
		plistp = &ch->ch_next;
	}

	/*
	 * If we didn't find a cache entry for this quadruple, then
	 * create one.  If we don't have one already preallocated,
	 * then drop the cache lock, create one, and then start over.
	 * If we did have a preallocated entry, then just add it to
	 * the front of the list.
	 */
	if (ch == NULL) {
		if (newch == NULL) {
			mutex_exit(&nfscl->nfscl_chtable_lock);
			newch = kmem_alloc(sizeof (*newch), KM_SLEEP);
			newch->ch_timesused = 0;
			newch->ch_prog = ci->cl_prog;
			newch->ch_vers = ci->cl_vers;
			newch->ch_dev = svp->sv_knconf->knc_rdev;
			newch->ch_protofmly = kmem_alloc(
			    strlen(svp->sv_knconf->knc_protofmly) + 1,
			    KM_SLEEP);
			(void) strcpy(newch->ch_protofmly,
			    svp->sv_knconf->knc_protofmly);
			newch->ch_list = NULL;
			goto top;
		}
		ch = newch;
		newch = NULL;
		ch->ch_next = nfscl->nfscl_chtable;
		nfscl->nfscl_chtable = ch;
	/*
	 * We found a cache entry, but if it isn't on the front of the
	 * list, then move it to the front of the list to try to take
	 * advantage of locality of operations.
	 */
	} else if (ch != nfscl->nfscl_chtable) {
		*plistp = ch->ch_next;
		ch->ch_next = nfscl->nfscl_chtable;
		nfscl->nfscl_chtable = ch;
	}

	/*
	 * If there was a free client handle cached, then remove it
	 * from the list, init it, and use it.
	 */
	if (ch->ch_list != NULL) {
		cp = ch->ch_list;
		ch->ch_list = cp->ch_list;
		mutex_exit(&nfscl->nfscl_chtable_lock);
		if (newch != NULL) {
			kmem_free(newch->ch_protofmly,
			    strlen(newch->ch_protofmly) + 1);
			kmem_free(newch, sizeof (*newch));
		}
		(void) clnt_tli_kinit(cp->ch_client, svp->sv_knconf,
		    &svp->sv_addr, ci->cl_readsize, ci->cl_retrans, cr);
		error = sec_clnt_geth(cp->ch_client, svp->sv_secdata, cr,
		    &cp->ch_client->cl_auth);
		if (error || cp->ch_client->cl_auth == NULL) {
			CLNT_DESTROY(cp->ch_client);
			kmem_cache_free(chtab_cache, cp);
			return ((error != 0) ? error : EINTR);
		}
		ch->ch_timesused++;
		*newcl = cp->ch_client;
		*chp = cp;
		return (0);
	}

	/*
	 * There weren't any free client handles which fit, so allocate
	 * a new one and use that.
	 */
#ifdef DEBUG
	atomic_add_64(&nfscl->nfscl_stat.clalloc.value.ui64, 1);
#endif
	mutex_exit(&nfscl->nfscl_chtable_lock);

	nfscl->nfscl_stat.cltoomany.value.ui64++;
	if (newch != NULL) {
		kmem_free(newch->ch_protofmly, strlen(newch->ch_protofmly) + 1);
		kmem_free(newch, sizeof (*newch));
	}

	cp = kmem_cache_alloc(chtab_cache, KM_SLEEP);
	cp->ch_head = ch;

	sigintr(&smask, (int)ci->cl_flags & MI_INT);
	error = clnt_tli_kcreate(svp->sv_knconf, &svp->sv_addr, ci->cl_prog,
	    ci->cl_vers, ci->cl_readsize, ci->cl_retrans, cr, &cp->ch_client);
	sigunintr(&smask);

	if (error != 0) {
		kmem_cache_free(chtab_cache, cp);
#ifdef DEBUG
		atomic_add_64(&nfscl->nfscl_stat.clalloc.value.ui64, -1);
#endif
		/*
		 * Warning is unnecessary if error is EINTR.
		 */
		if (error != EINTR) {
			nfs_cmn_err(error, CE_WARN,
			    "clget: couldn't create handle: %m\n");
		}
		return (error);
	}
	(void) CLNT_CONTROL(cp->ch_client, CLSET_PROGRESS, NULL);
	auth_destroy(cp->ch_client->cl_auth);
	error = sec_clnt_geth(cp->ch_client, svp->sv_secdata, cr,
	    &cp->ch_client->cl_auth);
	if (error || cp->ch_client->cl_auth == NULL) {
		CLNT_DESTROY(cp->ch_client);
		kmem_cache_free(chtab_cache, cp);
#ifdef DEBUG
		atomic_add_64(&nfscl->nfscl_stat.clalloc.value.ui64, -1);
#endif
		return ((error != 0) ? error : EINTR);
	}
	ch->ch_timesused++;
	*newcl = cp->ch_client;
	ASSERT(cp->ch_client->cl_nosignal == FALSE);
	*chp = cp;
	return (0);
}

int
clget(clinfo_t *ci, servinfo_t *svp, cred_t *cr, CLIENT **newcl,
    struct chtab **chp)
{
	struct nfs_clnt *nfscl;

	nfscl = zone_getspecific(nfsclnt_zone_key, nfs_zone());
	ASSERT(nfscl != NULL);

	return (clget_impl(ci, svp, cr, newcl, chp, nfscl));
}

static int
acl_clget(mntinfo_t *mi, servinfo_t *svp, cred_t *cr, CLIENT **newcl,
    struct chtab **chp, struct nfs_clnt *nfscl)
{
	clinfo_t ci;
	int error;

	/*
	 * Set read buffer size to rsize
	 * and add room for RPC headers.
	 */
	ci.cl_readsize = mi->mi_tsize;
	if (ci.cl_readsize != 0)
		ci.cl_readsize += (RPC_MAXDATASIZE - NFS_MAXDATA);

	/*
	 * If soft mount and server is down just try once.
	 * meaning: do not retransmit.
	 */
	if (!(mi->mi_flags & MI_HARD) && (mi->mi_flags & MI_DOWN))
		ci.cl_retrans = 0;
	else
		ci.cl_retrans = mi->mi_retrans;

	ci.cl_prog = NFS_ACL_PROGRAM;
	ci.cl_vers = mi->mi_vers;
	ci.cl_flags = mi->mi_flags;

	/*
	 * clget calls sec_clnt_geth() to get an auth handle. For RPCSEC_GSS
	 * security flavor, the client tries to establish a security context
	 * by contacting the server. If the connection is timed out or reset,
	 * e.g. server reboot, we will try again.
	 */
	do {
		error = clget_impl(&ci, svp, cr, newcl, chp, nfscl);

		if (error == 0)
			break;

		/*
		 * For forced unmount or zone shutdown, bail out, no retry.
		 */
		if (FS_OR_ZONE_GONE(mi->mi_vfsp)) {
			error = EIO;
			break;
		}

		/* do not retry for softmount */
		if (!(mi->mi_flags & MI_HARD))
			break;

		/* let the caller deal with the failover case */
		if (FAILOVER_MOUNT(mi))
			break;

	} while (error == ETIMEDOUT || error == ECONNRESET);

	return (error);
}

static int
nfs_clget(mntinfo_t *mi, servinfo_t *svp, cred_t *cr, CLIENT **newcl,
    struct chtab **chp, struct nfs_clnt *nfscl)
{
	clinfo_t ci;
	int error;

	/*
	 * Set read buffer size to rsize
	 * and add room for RPC headers.
	 */
	ci.cl_readsize = mi->mi_tsize;
	if (ci.cl_readsize != 0)
		ci.cl_readsize += (RPC_MAXDATASIZE - NFS_MAXDATA);

	/*
	 * If soft mount and server is down just try once.
	 * meaning: do not retransmit.
	 */
	if (!(mi->mi_flags & MI_HARD) && (mi->mi_flags & MI_DOWN))
		ci.cl_retrans = 0;
	else
		ci.cl_retrans = mi->mi_retrans;

	ci.cl_prog = mi->mi_prog;
	ci.cl_vers = mi->mi_vers;
	ci.cl_flags = mi->mi_flags;

	/*
	 * clget calls sec_clnt_geth() to get an auth handle. For RPCSEC_GSS
	 * security flavor, the client tries to establish a security context
	 * by contacting the server. If the connection is timed out or reset,
	 * e.g. server reboot, we will try again.
	 */
	do {
		error = clget_impl(&ci, svp, cr, newcl, chp, nfscl);

		if (error == 0)
			break;

		/*
		 * For forced unmount or zone shutdown, bail out, no retry.
		 */
		if (FS_OR_ZONE_GONE(mi->mi_vfsp)) {
			error = EIO;
			break;
		}

		/* do not retry for softmount */
		if (!(mi->mi_flags & MI_HARD))
			break;

		/* let the caller deal with the failover case */
		if (FAILOVER_MOUNT(mi))
			break;

	} while (error == ETIMEDOUT || error == ECONNRESET);

	return (error);
}

static void
clfree_impl(CLIENT *cl, struct chtab *cp, struct nfs_clnt *nfscl)
{
	if (cl->cl_auth != NULL) {
		sec_clnt_freeh(cl->cl_auth);
		cl->cl_auth = NULL;
	}

	/*
	 * Timestamp this cache entry so that we know when it was last
	 * used.
	 */
	cp->ch_freed = gethrestime_sec();

	/*
	 * Add the free client handle to the front of the list.
	 * This way, the list will be sorted in youngest to oldest
	 * order.
	 */
	mutex_enter(&nfscl->nfscl_chtable_lock);
	cp->ch_list = cp->ch_head->ch_list;
	cp->ch_head->ch_list = cp;
	mutex_exit(&nfscl->nfscl_chtable_lock);
}

void
clfree(CLIENT *cl, struct chtab *cp)
{
	struct nfs_clnt *nfscl;

	nfscl = zone_getspecific(nfsclnt_zone_key, nfs_zone());
	ASSERT(nfscl != NULL);

	clfree_impl(cl, cp, nfscl);
}

#define	CL_HOLDTIME	60	/* time to hold client handles */

static void
clreclaim_zone(struct nfs_clnt *nfscl, uint_t cl_holdtime)
{
	struct chhead *ch;
	struct chtab *cp;	/* list of objects that can be reclaimed */
	struct chtab *cpe;
	struct chtab *cpl;
	struct chtab **cpp;
#ifdef DEBUG
	int n = 0;
#endif

	/*
	 * Need to reclaim some memory, so step through the cache
	 * looking through the lists for entries which can be freed.
	 */
	cp = NULL;

	mutex_enter(&nfscl->nfscl_chtable_lock);

	/*
	 * Here we step through each non-NULL quadruple and start to
	 * construct the reclaim list pointed to by cp.  Note that
	 * cp will contain all eligible chtab entries.  When this traversal
	 * completes, chtab entries from the last quadruple will be at the
	 * front of cp and entries from previously inspected quadruples have
	 * been appended to the rear of cp.
	 */
	for (ch = nfscl->nfscl_chtable; ch != NULL; ch = ch->ch_next) {
		if (ch->ch_list == NULL)
			continue;
		/*
		 * Search each list for entries older then
		 * cl_holdtime seconds.  The lists are maintained
		 * in youngest to oldest order so that when the
		 * first entry is found which is old enough, then
		 * all of the rest of the entries on the list will
		 * be old enough as well.
		 */
		cpl = ch->ch_list;
		cpp = &ch->ch_list;
		while (cpl != NULL &&
			cpl->ch_freed + cl_holdtime > gethrestime_sec()) {
			cpp = &cpl->ch_list;
			cpl = cpl->ch_list;
		}
		if (cpl != NULL) {
			*cpp = NULL;
			if (cp != NULL) {
				cpe = cpl;
				while (cpe->ch_list != NULL)
					cpe = cpe->ch_list;
				cpe->ch_list = cp;
			}
			cp = cpl;
		}
	}

	mutex_exit(&nfscl->nfscl_chtable_lock);

	/*
	 * If cp is empty, then there is nothing to reclaim here.
	 */
	if (cp == NULL)
		return;

	/*
	 * Step through the list of entries to free, destroying each client
	 * handle and kmem_free'ing the memory for each entry.
	 */
	while (cp != NULL) {
#ifdef DEBUG
		n++;
#endif
		CLNT_DESTROY(cp->ch_client);
		cpl = cp->ch_list;
		kmem_cache_free(chtab_cache, cp);
		cp = cpl;
	}

#ifdef DEBUG
	/*
	 * Update clalloc so that nfsstat shows the current number
	 * of allocated client handles.
	 */
	atomic_add_64(&nfscl->nfscl_stat.clalloc.value.ui64, -n);
#endif
}

/* ARGSUSED */
static void
clreclaim(void *all)
{
	struct nfs_clnt *nfscl;

#ifdef DEBUG
	clstat_debug.clreclaim.value.ui64++;
#endif
	/*
	 * The system is low on memory; go through and try to reclaim some from
	 * every zone on the system.
	 */
	mutex_enter(&nfs_clnt_list_lock);
	nfscl = list_head(&nfs_clnt_list);
	for (; nfscl != NULL; nfscl = list_next(&nfs_clnt_list, nfscl))
		clreclaim_zone(nfscl, CL_HOLDTIME);
	mutex_exit(&nfs_clnt_list_lock);
}

/*
 * Minimum time-out values indexed by call type
 * These units are in "eights" of a second to avoid multiplies
 */
static unsigned int minimum_timeo[] = {
	6, 7, 10
};

/*
 * Back off for retransmission timeout, MAXTIMO is in hz of a sec
 */
#define	MAXTIMO	(20*hz)
#define	backoff(tim)	(((tim) < MAXTIMO) ? dobackoff(tim) : (tim))
#define	dobackoff(tim)	((((tim) << 1) > MAXTIMO) ? MAXTIMO : ((tim) << 1))

#define	MIN_NFS_TSIZE 512	/* minimum "chunk" of NFS IO */
#define	REDUCE_NFS_TIME (hz/2)	/* rtxcur we try to keep under */
#define	INCREASE_NFS_TIME (hz/3*8) /* srtt we try to keep under (scaled*8) */

/*
 * Function called when rfscall notices that we have been
 * re-transmitting, or when we get a response without retransmissions.
 * Return 1 if the transfer size was adjusted down - 0 if no change.
 */
static int
nfs_feedback(int flag, int which, mntinfo_t *mi)
{
	int kind;
	int r = 0;

	mutex_enter(&mi->mi_lock);
	if (flag == FEEDBACK_REXMIT1) {
		if (mi->mi_timers[NFS_CALLTYPES].rt_rtxcur != 0 &&
		    mi->mi_timers[NFS_CALLTYPES].rt_rtxcur < REDUCE_NFS_TIME)
			goto done;
		if (mi->mi_curread > MIN_NFS_TSIZE) {
			mi->mi_curread /= 2;
			if (mi->mi_curread < MIN_NFS_TSIZE)
				mi->mi_curread = MIN_NFS_TSIZE;
			r = 1;
		}

		if (mi->mi_curwrite > MIN_NFS_TSIZE) {
			mi->mi_curwrite /= 2;
			if (mi->mi_curwrite < MIN_NFS_TSIZE)
				mi->mi_curwrite = MIN_NFS_TSIZE;
			r = 1;
		}
	} else if (flag == FEEDBACK_OK) {
		kind = mi->mi_timer_type[which];
		if (kind == 0 ||
		    mi->mi_timers[kind].rt_srtt >= INCREASE_NFS_TIME)
			goto done;
		if (kind == 1) {
			if (mi->mi_curread >= mi->mi_tsize)
				goto done;
			mi->mi_curread +=  MIN_NFS_TSIZE;
			if (mi->mi_curread > mi->mi_tsize/2)
				mi->mi_curread = mi->mi_tsize;
		} else if (kind == 2) {
			if (mi->mi_curwrite >= mi->mi_stsize)
				goto done;
			mi->mi_curwrite += MIN_NFS_TSIZE;
			if (mi->mi_curwrite > mi->mi_stsize/2)
				mi->mi_curwrite = mi->mi_stsize;
		}
	}
done:
	mutex_exit(&mi->mi_lock);
	return (r);
}

#ifdef DEBUG
static int rfs2call_hits = 0;
static int rfs2call_misses = 0;
#endif

int
rfs2call(mntinfo_t *mi, rpcproc_t which, xdrproc_t xdrargs, caddr_t argsp,
    xdrproc_t xdrres, caddr_t resp, cred_t *cr, int *douprintf,
    enum nfsstat *statusp, int flags, failinfo_t *fi)
{
	int rpcerror;
	enum clnt_stat rpc_status;

	ASSERT(statusp != NULL);

	rpcerror = rfscall(mi, which, xdrargs, argsp, xdrres, resp,
	    cr, douprintf, &rpc_status, flags, fi);
	if (!rpcerror) {
		/*
		 * See crnetadjust() for comments.
		 */
		if (*statusp == NFSERR_ACCES &&
		    (cr = crnetadjust(cr)) != NULL) {
#ifdef DEBUG
			rfs2call_hits++;
#endif
			rpcerror = rfscall(mi, which, xdrargs, argsp, xdrres,
			    resp, cr, douprintf, NULL, flags, fi);
			crfree(cr);
#ifdef DEBUG
			if (*statusp == NFSERR_ACCES)
				rfs2call_misses++;
#endif
		}
	} else if (rpc_status == RPC_PROCUNAVAIL) {
		*statusp = NFSERR_OPNOTSUPP;
		rpcerror = 0;
	}

	return (rpcerror);
}

#define	NFS3_JUKEBOX_DELAY	10 * hz

static clock_t nfs3_jukebox_delay = 0;

#ifdef DEBUG
static int rfs3call_hits = 0;
static int rfs3call_misses = 0;
#endif

int
rfs3call(mntinfo_t *mi, rpcproc_t which, xdrproc_t xdrargs, caddr_t argsp,
    xdrproc_t xdrres, caddr_t resp, cred_t *cr, int *douprintf,
    nfsstat3 *statusp, int flags, failinfo_t *fi)
{
	int rpcerror;
	int user_informed;

	user_informed = 0;
	do {
		rpcerror = rfscall(mi, which, xdrargs, argsp, xdrres, resp,
		    cr, douprintf, NULL, flags, fi);
		if (!rpcerror) {
			cred_t *crr;
			if (*statusp == NFS3ERR_JUKEBOX) {
				if (ttoproc(curthread) == &p0) {
					rpcerror = EAGAIN;
					break;
				}
				if (!user_informed) {
					user_informed = 1;
					uprintf(
		"file temporarily unavailable on the server, retrying...\n");
				}
				delay(nfs3_jukebox_delay);
			}
			/*
			 * See crnetadjust() for comments.
			 */
			else if (*statusp == NFS3ERR_ACCES &&
			    (crr = crnetadjust(cr)) != NULL) {
#ifdef DEBUG
				rfs3call_hits++;
#endif
				rpcerror = rfscall(mi, which, xdrargs, argsp,
				    xdrres, resp, crr, douprintf,
				    NULL, flags, fi);

				crfree(crr);
#ifdef DEBUG
				if (*statusp == NFS3ERR_ACCES)
					rfs3call_misses++;
#endif
			}
		}
	} while (!rpcerror && *statusp == NFS3ERR_JUKEBOX);

	return (rpcerror);
}

#define	VALID_FH(fi)	(VTOR(fi->vp)->r_server == VTOMI(fi->vp)->mi_curr_serv)
#define	INC_READERS(mi)		{ \
	mi->mi_readers++; \
}
#define	DEC_READERS(mi)		{ \
	mi->mi_readers--; \
	if (mi->mi_readers == 0) \
		cv_broadcast(&mi->mi_failover_cv); \
}

static int
rfscall(mntinfo_t *mi, rpcproc_t which, xdrproc_t xdrargs, caddr_t argsp,
    xdrproc_t xdrres, caddr_t resp, cred_t *icr, int *douprintf,
    enum clnt_stat *rpc_status, int flags, failinfo_t *fi)
{
	CLIENT *client;
	struct chtab *ch;
	cred_t *cr = icr;
	enum clnt_stat status;
	struct rpc_err rpcerr;
	struct timeval wait;
	int timeo;		/* in units of hz */
	int my_rsize, my_wsize;
	bool_t tryagain;
	bool_t cred_cloned = FALSE;
	k_sigset_t smask;
	servinfo_t *svp;
	struct nfs_clnt *nfscl;
	zoneid_t zoneid = getzoneid();
#ifdef DEBUG
	char *bufp;
#endif


	TRACE_2(TR_FAC_NFS, TR_RFSCALL_START,
		"rfscall_start:which %d mi %p", which, mi);

	nfscl = zone_getspecific(nfsclnt_zone_key, nfs_zone());
	ASSERT(nfscl != NULL);

	nfscl->nfscl_stat.calls.value.ui64++;
	mi->mi_reqs[which].value.ui64++;

	rpcerr.re_status = RPC_SUCCESS;

	/*
	 * In case of forced unmount or zone shutdown, return EIO.
	 */

	if (FS_OR_ZONE_GONE(mi->mi_vfsp)) {
		rpcerr.re_status = RPC_FAILED;
		rpcerr.re_errno = EIO;
		return (rpcerr.re_errno);
	}

	/*
	 * Remember the transfer sizes in case
	 * nfs_feedback changes them underneath us.
	 */
	my_rsize = mi->mi_curread;
	my_wsize = mi->mi_curwrite;

	/*
	 * NFS client failover support
	 *
	 * If this rnode is not in sync with the current server (VALID_FH),
	 * we'd like to do a remap to get in sync.  We can be interrupted
	 * in failover_remap(), and if so we'll bail.  Otherwise, we'll
	 * use the best info we have to try the RPC.  Part of that is
	 * unconditionally updating the filehandle copy kept for V3.
	 *
	 * Locking: INC_READERS/DEC_READERS is a poor man's interrruptible
	 * rw_enter(); we're trying to keep the current server from being
	 * changed on us until we're done with the remapping and have a
	 * matching client handle.  We don't want to sending a filehandle
	 * to the wrong host.
	 */
failoverretry:
	if (FAILOVER_MOUNT(mi)) {
		mutex_enter(&mi->mi_lock);
		if (!(flags & RFSCALL_SOFT) && failover_safe(fi)) {
			if (failover_wait(mi)) {
				mutex_exit(&mi->mi_lock);
				return (EINTR);
			}
		}
		INC_READERS(mi);
		mutex_exit(&mi->mi_lock);
		if (fi) {
			if (!VALID_FH(fi) &&
			    !(flags & RFSCALL_SOFT) && failover_safe(fi)) {
				int remaperr;

				svp = mi->mi_curr_serv;
				remaperr = failover_remap(fi);
				if (remaperr != 0) {
#ifdef DEBUG
					if (remaperr != EINTR)
						nfs_cmn_err(remaperr, CE_WARN,
					    "rfscall couldn't failover: %m");
#endif
					mutex_enter(&mi->mi_lock);
					DEC_READERS(mi);
					mutex_exit(&mi->mi_lock);
					/*
					 * If failover_remap returns ETIMEDOUT
					 * and the filesystem is hard mounted
					 * we have to retry the call with a new
					 * server.
					 */
					if ((mi->mi_flags & MI_HARD) &&
					    IS_RECOVERABLE_ERROR(remaperr)) {
						if (svp == mi->mi_curr_serv)
							failover_newserver(mi);
						rpcerr.re_status = RPC_SUCCESS;
						goto failoverretry;
					}
					rpcerr.re_errno = remaperr;
					return (remaperr);
				}
			}
			if (fi->fhp && fi->copyproc)
				(*fi->copyproc)(fi->fhp, fi->vp);
		}
	}

	/* For TSOL, use a new cred which has net_mac_aware flag */
	if (!cred_cloned && is_system_labeled()) {
		cred_cloned = TRUE;
		cr = crdup(icr);
		(void) setpflags(NET_MAC_AWARE, 1, cr);
	}

	/*
	 * clget() calls clnt_tli_kinit() which clears the xid, so we
	 * are guaranteed to reprocess the retry as a new request.
	 */
	svp = mi->mi_curr_serv;
	rpcerr.re_errno = nfs_clget(mi, svp, cr, &client, &ch, nfscl);

	if (FAILOVER_MOUNT(mi)) {
		mutex_enter(&mi->mi_lock);
		DEC_READERS(mi);
		mutex_exit(&mi->mi_lock);

		if ((rpcerr.re_errno == ETIMEDOUT ||
				rpcerr.re_errno == ECONNRESET) &&
				failover_safe(fi)) {
			if (svp == mi->mi_curr_serv)
				failover_newserver(mi);
			goto failoverretry;
		}
	}
	if (rpcerr.re_errno != 0)
		return (rpcerr.re_errno);

	if (svp->sv_knconf->knc_semantics == NC_TPI_COTS_ORD ||
	    svp->sv_knconf->knc_semantics == NC_TPI_COTS) {
		timeo = (mi->mi_timeo * hz) / 10;
	} else {
		mutex_enter(&mi->mi_lock);
		timeo = CLNT_SETTIMERS(client,
		    &(mi->mi_timers[mi->mi_timer_type[which]]),
		    &(mi->mi_timers[NFS_CALLTYPES]),
		    (minimum_timeo[mi->mi_call_type[which]]*hz)>>3,
		    (void (*)())NULL, (caddr_t)mi, 0);
		mutex_exit(&mi->mi_lock);
	}

	/*
	 * If hard mounted fs, retry call forever unless hard error occurs.
	 */
	do {
		tryagain = FALSE;

		if (FS_OR_ZONE_GONE(mi->mi_vfsp)) {
			status = RPC_FAILED;
			rpcerr.re_status = RPC_FAILED;
			rpcerr.re_errno = EIO;
			break;
		}

		TICK_TO_TIMEVAL(timeo, &wait);

		/*
		 * Mask out all signals except SIGHUP, SIGINT, SIGQUIT
		 * and SIGTERM. (Preserving the existing masks).
		 * Mask out SIGINT if mount option nointr is specified.
		 */
		sigintr(&smask, (int)mi->mi_flags & MI_INT);
		if (!(mi->mi_flags & MI_INT))
			client->cl_nosignal = TRUE;

		/*
		 * If there is a current signal, then don't bother
		 * even trying to send out the request because we
		 * won't be able to block waiting for the response.
		 * Simply assume RPC_INTR and get on with it.
		 */
		if (ttolwp(curthread) != NULL && ISSIG(curthread, JUSTLOOKING))
			status = RPC_INTR;
		else {
			status = CLNT_CALL(client, which, xdrargs, argsp,
			    xdrres, resp, wait);
		}

		if (!(mi->mi_flags & MI_INT))
			client->cl_nosignal = FALSE;
		/*
		 * restore original signal mask
		 */
		sigunintr(&smask);

		switch (status) {
		case RPC_SUCCESS:
			if ((mi->mi_flags & MI_DYNAMIC) &&
			    mi->mi_timer_type[which] != 0 &&
			    (mi->mi_curread != my_rsize ||
			    mi->mi_curwrite != my_wsize))
				(void) nfs_feedback(FEEDBACK_OK, which, mi);
			break;

		case RPC_INTR:
			/*
			 * There is no way to recover from this error,
			 * even if mount option nointr is specified.
			 * SIGKILL, for example, cannot be blocked.
			 */
			rpcerr.re_status = RPC_INTR;
			rpcerr.re_errno = EINTR;
			break;

		case RPC_UDERROR:
			/*
			 * If the NFS server is local (vold) and
			 * it goes away then we get RPC_UDERROR.
			 * This is a retryable error, so we would
			 * loop, so check to see if the specific
			 * error was ECONNRESET, indicating that
			 * target did not exist at all.  If so,
			 * return with RPC_PROGUNAVAIL and
			 * ECONNRESET to indicate why.
			 */
			CLNT_GETERR(client, &rpcerr);
			if (rpcerr.re_errno == ECONNRESET) {
				rpcerr.re_status = RPC_PROGUNAVAIL;
				rpcerr.re_errno = ECONNRESET;
				break;
			}
			/*FALLTHROUGH*/

		default:		/* probably RPC_TIMEDOUT */
			if (IS_UNRECOVERABLE_RPC(status))
				break;

			/*
			 * increment server not responding count
			 */
			mutex_enter(&mi->mi_lock);
			mi->mi_noresponse++;
			mutex_exit(&mi->mi_lock);
#ifdef DEBUG
			nfscl->nfscl_stat.noresponse.value.ui64++;
#endif

			if (!(mi->mi_flags & MI_HARD)) {
				if (!(mi->mi_flags & MI_SEMISOFT) ||
				    (mi->mi_ss_call_type[which] == 0))
					break;
			}

			/*
			 * The call is in progress (over COTS).
			 * Try the CLNT_CALL again, but don't
			 * print a noisy error message.
			 */
			if (status == RPC_INPROGRESS) {
				tryagain = TRUE;
				break;
			}

			if (flags & RFSCALL_SOFT)
				break;

			/*
			 * On zone shutdown, just move on.
			 */
			if (zone_status_get(curproc->p_zone) >=
			    ZONE_IS_SHUTTING_DOWN) {
				rpcerr.re_status = RPC_FAILED;
				rpcerr.re_errno = EIO;
				break;
			}

			/*
			 * NFS client failover support
			 *
			 * If the current server just failed us, we'll
			 * start the process of finding a new server.
			 * After that, we can just retry.
			 */
			if (FAILOVER_MOUNT(mi) && failover_safe(fi)) {
				if (svp == mi->mi_curr_serv)
					failover_newserver(mi);
				clfree_impl(client, ch, nfscl);
				goto failoverretry;
			}

			tryagain = TRUE;
			timeo = backoff(timeo);
			mutex_enter(&mi->mi_lock);
			if (!(mi->mi_flags & MI_PRINTED)) {
				mi->mi_flags |= MI_PRINTED;
				mutex_exit(&mi->mi_lock);
#ifdef DEBUG
				zprintf(zoneid,
			"NFS%d server %s not responding still trying\n",
				    mi->mi_vers, svp->sv_hostname);
#else
				zprintf(zoneid,
			"NFS server %s not responding still trying\n",
				    svp->sv_hostname);
#endif
			} else
				mutex_exit(&mi->mi_lock);
			if (*douprintf && curproc->p_sessp->s_vp != NULL) {
				*douprintf = 0;
				if (!(mi->mi_flags & MI_NOPRINT))
#ifdef DEBUG
					uprintf(
			    "NFS%d server %s not responding still trying\n",
					    mi->mi_vers, svp->sv_hostname);
#else
					uprintf(
			    "NFS server %s not responding still trying\n",
					    svp->sv_hostname);
#endif
			}

			/*
			 * If doing dynamic adjustment of transfer
			 * size and if it's a read or write call
			 * and if the transfer size changed while
			 * retransmitting or if the feedback routine
			 * changed the transfer size,
			 * then exit rfscall so that the transfer
			 * size can be adjusted at the vnops level.
			 */
			if ((mi->mi_flags & MI_DYNAMIC) &&
			    mi->mi_timer_type[which] != 0 &&
			    (mi->mi_curread != my_rsize ||
			    mi->mi_curwrite != my_wsize ||
			    nfs_feedback(FEEDBACK_REXMIT1, which, mi))) {
				/*
				 * On read or write calls, return
				 * back to the vnode ops level if
				 * the transfer size changed.
				 */
				clfree_impl(client, ch, nfscl);
				if (cred_cloned)
					crfree(cr);
				return (ENFS_TRYAGAIN);
			}
		}
	} while (tryagain);

	if (status != RPC_SUCCESS) {
		/*
		 * Let soft mounts use the timed out message.
		 */
		if (status == RPC_INPROGRESS)
			status = RPC_TIMEDOUT;
		nfscl->nfscl_stat.badcalls.value.ui64++;
		if (status != RPC_INTR) {
			mutex_enter(&mi->mi_lock);
			mi->mi_flags |= MI_DOWN;
			mutex_exit(&mi->mi_lock);
			CLNT_GETERR(client, &rpcerr);
#ifdef DEBUG
			bufp = clnt_sperror(client, svp->sv_hostname);
			zprintf(zoneid, "NFS%d %s failed for %s\n",
			    mi->mi_vers, mi->mi_rfsnames[which], bufp);
			if (curproc->p_sessp->s_vp != NULL) {
				if (!(mi->mi_flags & MI_NOPRINT)) {
					uprintf("NFS%d %s failed for %s\n",
					    mi->mi_vers, mi->mi_rfsnames[which],
					    bufp);
				}
			}
			kmem_free(bufp, MAXPATHLEN);
#else
			zprintf(zoneid,
			    "NFS %s failed for server %s: error %d (%s)\n",
			    mi->mi_rfsnames[which], svp->sv_hostname,
			    status, clnt_sperrno(status));
			if (curproc->p_sessp->s_vp != NULL) {
				if (!(mi->mi_flags & MI_NOPRINT)) {
					uprintf(
				"NFS %s failed for server %s: error %d (%s)\n",
					    mi->mi_rfsnames[which],
					    svp->sv_hostname, status,
					    clnt_sperrno(status));
				}
			}
#endif
			/*
			 * when CLNT_CALL() fails with RPC_AUTHERROR,
			 * re_errno is set appropriately depending on
			 * the authentication error
			 */
			if (status == RPC_VERSMISMATCH ||
			    status == RPC_PROGVERSMISMATCH)
				rpcerr.re_errno = EIO;
		}
	} else {
		/*
		 * Test the value of mi_down and mi_printed without
		 * holding the mi_lock mutex.  If they are both zero,
		 * then it is okay to skip the down and printed
		 * processing.  This saves on a mutex_enter and
		 * mutex_exit pair for a normal, successful RPC.
		 * This was just complete overhead.
		 */
		if (mi->mi_flags & (MI_DOWN | MI_PRINTED)) {
			mutex_enter(&mi->mi_lock);
			mi->mi_flags &= ~MI_DOWN;
			if (mi->mi_flags & MI_PRINTED) {
				mi->mi_flags &= ~MI_PRINTED;
				mutex_exit(&mi->mi_lock);
#ifdef DEBUG
			if (!(mi->mi_vfsp->vfs_flag & VFS_UNMOUNTED))
				zprintf(zoneid, "NFS%d server %s ok\n",
				    mi->mi_vers, svp->sv_hostname);
#else
			if (!(mi->mi_vfsp->vfs_flag & VFS_UNMOUNTED))
				zprintf(zoneid, "NFS server %s ok\n",
				    svp->sv_hostname);
#endif
			} else
				mutex_exit(&mi->mi_lock);
		}

		if (*douprintf == 0) {
			if (!(mi->mi_flags & MI_NOPRINT))
#ifdef DEBUG
				if (!(mi->mi_vfsp->vfs_flag & VFS_UNMOUNTED))
					uprintf("NFS%d server %s ok\n",
					    mi->mi_vers, svp->sv_hostname);
#else
			if (!(mi->mi_vfsp->vfs_flag & VFS_UNMOUNTED))
				uprintf("NFS server %s ok\n", svp->sv_hostname);
#endif
			*douprintf = 1;
		}
	}

	clfree_impl(client, ch, nfscl);
	if (cred_cloned)
		crfree(cr);

	ASSERT(rpcerr.re_status == RPC_SUCCESS || rpcerr.re_errno != 0);

	if (rpc_status != NULL)
		*rpc_status = rpcerr.re_status;

	TRACE_1(TR_FAC_NFS, TR_RFSCALL_END, "rfscall_end:errno %d",
	    rpcerr.re_errno);

	return (rpcerr.re_errno);
}

#ifdef DEBUG
static int acl2call_hits = 0;
static int acl2call_misses = 0;
#endif

int
acl2call(mntinfo_t *mi, rpcproc_t which, xdrproc_t xdrargs, caddr_t argsp,
    xdrproc_t xdrres, caddr_t resp, cred_t *cr, int *douprintf,
    enum nfsstat *statusp, int flags, failinfo_t *fi)
{
	int rpcerror;

	rpcerror = aclcall(mi, which, xdrargs, argsp, xdrres, resp,
	    cr, douprintf, flags, fi);
	if (!rpcerror) {
		/*
		 * See comments with crnetadjust().
		 */
		if (*statusp == NFSERR_ACCES &&
		    (cr = crnetadjust(cr)) != NULL) {
#ifdef DEBUG
			acl2call_hits++;
#endif
			rpcerror = aclcall(mi, which, xdrargs, argsp, xdrres,
			    resp, cr, douprintf, flags, fi);
			crfree(cr);
#ifdef DEBUG
			if (*statusp == NFSERR_ACCES)
				acl2call_misses++;
#endif
		}
	}

	return (rpcerror);
}

#ifdef DEBUG
static int acl3call_hits = 0;
static int acl3call_misses = 0;
#endif

int
acl3call(mntinfo_t *mi, rpcproc_t which, xdrproc_t xdrargs, caddr_t argsp,
    xdrproc_t xdrres, caddr_t resp, cred_t *cr, int *douprintf,
    nfsstat3 *statusp, int flags, failinfo_t *fi)
{
	int rpcerror;
	int user_informed;

	user_informed = 0;

	do {
		rpcerror = aclcall(mi, which, xdrargs, argsp, xdrres, resp,
		    cr, douprintf, flags, fi);
		if (!rpcerror) {
			cred_t *crr;
			if (*statusp == NFS3ERR_JUKEBOX) {
				if (!user_informed) {
					user_informed = 1;
					uprintf(
		"file temporarily unavailable on the server, retrying...\n");
				}
				delay(nfs3_jukebox_delay);
			}
			/*
			 * See crnetadjust() for comments.
			 */
			else if (*statusp == NFS3ERR_ACCES &&
			    (crr = crnetadjust(cr)) != NULL) {
#ifdef DEBUG
				acl3call_hits++;
#endif
				rpcerror = aclcall(mi, which, xdrargs, argsp,
				    xdrres, resp, crr, douprintf, flags, fi);

				crfree(crr);
#ifdef DEBUG
				if (*statusp == NFS3ERR_ACCES)
					acl3call_misses++;
#endif
			}
		}
	} while (!rpcerror && *statusp == NFS3ERR_JUKEBOX);

	return (rpcerror);
}

static int
aclcall(mntinfo_t *mi, rpcproc_t which, xdrproc_t xdrargs, caddr_t argsp,
    xdrproc_t xdrres, caddr_t resp, cred_t *icr, int *douprintf,
    int flags, failinfo_t *fi)
{
	CLIENT *client;
	struct chtab *ch;
	cred_t *cr = icr;
	bool_t cred_cloned = FALSE;
	enum clnt_stat status;
	struct rpc_err rpcerr;
	struct timeval wait;
	int timeo;		/* in units of hz */
#if 0 /* notyet */
	int my_rsize, my_wsize;
#endif
	bool_t tryagain;
	k_sigset_t smask;
	servinfo_t *svp;
	struct nfs_clnt *nfscl;
	zoneid_t zoneid = getzoneid();
#ifdef DEBUG
	char *bufp;
#endif

#if 0 /* notyet */
	TRACE_2(TR_FAC_NFS, TR_RFSCALL_START,
		"rfscall_start:which %d mi %p", which, mi);
#endif

	nfscl = zone_getspecific(nfsclnt_zone_key, nfs_zone());
	ASSERT(nfscl != NULL);

	nfscl->nfscl_stat.calls.value.ui64++;
	mi->mi_aclreqs[which].value.ui64++;

	rpcerr.re_status = RPC_SUCCESS;

	if (FS_OR_ZONE_GONE(mi->mi_vfsp)) {
		rpcerr.re_status = RPC_FAILED;
		rpcerr.re_errno = EIO;
		return (rpcerr.re_errno);
	}

#if 0 /* notyet */
	/*
	 * Remember the transfer sizes in case
	 * nfs_feedback changes them underneath us.
	 */
	my_rsize = mi->mi_curread;
	my_wsize = mi->mi_curwrite;
#endif

	/*
	 * NFS client failover support
	 *
	 * If this rnode is not in sync with the current server (VALID_FH),
	 * we'd like to do a remap to get in sync.  We can be interrupted
	 * in failover_remap(), and if so we'll bail.  Otherwise, we'll
	 * use the best info we have to try the RPC.  Part of that is
	 * unconditionally updating the filehandle copy kept for V3.
	 *
	 * Locking: INC_READERS/DEC_READERS is a poor man's interrruptible
	 * rw_enter(); we're trying to keep the current server from being
	 * changed on us until we're done with the remapping and have a
	 * matching client handle.  We don't want to sending a filehandle
	 * to the wrong host.
	 */
failoverretry:
	if (FAILOVER_MOUNT(mi)) {
		mutex_enter(&mi->mi_lock);
		if (!(flags & RFSCALL_SOFT) && failover_safe(fi)) {
			if (failover_wait(mi)) {
				mutex_exit(&mi->mi_lock);
				return (EINTR);
			}
		}
		INC_READERS(mi);
		mutex_exit(&mi->mi_lock);
		if (fi) {
			if (!VALID_FH(fi) &&
			    !(flags & RFSCALL_SOFT) && failover_safe(fi)) {
				int remaperr;

				svp = mi->mi_curr_serv;
				remaperr = failover_remap(fi);
				if (remaperr != 0) {
#ifdef DEBUG
					if (remaperr != EINTR)
						nfs_cmn_err(remaperr, CE_WARN,
					    "aclcall couldn't failover: %m");
#endif
					mutex_enter(&mi->mi_lock);
					DEC_READERS(mi);
					mutex_exit(&mi->mi_lock);

					/*
					 * If failover_remap returns ETIMEDOUT
					 * and the filesystem is hard mounted
					 * we have to retry the call with a new
					 * server.
					 */
					if ((mi->mi_flags & MI_HARD) &&
					    IS_RECOVERABLE_ERROR(remaperr)) {
						if (svp == mi->mi_curr_serv)
							failover_newserver(mi);
						rpcerr.re_status = RPC_SUCCESS;
						goto failoverretry;
					}
					return (remaperr);
				}
			}
			if (fi->fhp && fi->copyproc)
				(*fi->copyproc)(fi->fhp, fi->vp);
		}
	}

	/* For TSOL, use a new cred which has net_mac_aware flag */
	if (!cred_cloned && is_system_labeled()) {
		cred_cloned = TRUE;
		cr = crdup(icr);
		(void) setpflags(NET_MAC_AWARE, 1, cr);
	}

	/*
	 * acl_clget() calls clnt_tli_kinit() which clears the xid, so we
	 * are guaranteed to reprocess the retry as a new request.
	 */
	svp = mi->mi_curr_serv;
	rpcerr.re_errno = acl_clget(mi, svp, cr, &client, &ch, nfscl);
	if (FAILOVER_MOUNT(mi)) {
		mutex_enter(&mi->mi_lock);
		DEC_READERS(mi);
		mutex_exit(&mi->mi_lock);

		if ((rpcerr.re_errno == ETIMEDOUT ||
				rpcerr.re_errno == ECONNRESET) &&
				failover_safe(fi)) {
			if (svp == mi->mi_curr_serv)
				failover_newserver(mi);
			goto failoverretry;
		}
	}
	if (rpcerr.re_errno != 0) {
		if (cred_cloned)
			crfree(cr);
		return (rpcerr.re_errno);
	}

	if (svp->sv_knconf->knc_semantics == NC_TPI_COTS_ORD ||
	    svp->sv_knconf->knc_semantics == NC_TPI_COTS) {
		timeo = (mi->mi_timeo * hz) / 10;
	} else {
		mutex_enter(&mi->mi_lock);
		timeo = CLNT_SETTIMERS(client,
		    &(mi->mi_timers[mi->mi_acl_timer_type[which]]),
		    &(mi->mi_timers[NFS_CALLTYPES]),
		    (minimum_timeo[mi->mi_acl_call_type[which]]*hz)>>3,
		    (void (*)()) 0, (caddr_t)mi, 0);
		mutex_exit(&mi->mi_lock);
	}

	/*
	 * If hard mounted fs, retry call forever unless hard error occurs.
	 */
	do {
		tryagain = FALSE;

		if (FS_OR_ZONE_GONE(mi->mi_vfsp)) {
			status = RPC_FAILED;
			rpcerr.re_status = RPC_FAILED;
			rpcerr.re_errno = EIO;
			break;
		}

		TICK_TO_TIMEVAL(timeo, &wait);

		/*
		 * Mask out all signals except SIGHUP, SIGINT, SIGQUIT
		 * and SIGTERM. (Preserving the existing masks).
		 * Mask out SIGINT if mount option nointr is specified.
		 */
		sigintr(&smask, (int)mi->mi_flags & MI_INT);
		if (!(mi->mi_flags & MI_INT))
			client->cl_nosignal = TRUE;

		/*
		 * If there is a current signal, then don't bother
		 * even trying to send out the request because we
		 * won't be able to block waiting for the response.
		 * Simply assume RPC_INTR and get on with it.
		 */
		if (ttolwp(curthread) != NULL && ISSIG(curthread, JUSTLOOKING))
			status = RPC_INTR;
		else {
			status = CLNT_CALL(client, which, xdrargs, argsp,
			    xdrres, resp, wait);
		}

		if (!(mi->mi_flags & MI_INT))
			client->cl_nosignal = FALSE;
		/*
		 * restore original signal mask
		 */
		sigunintr(&smask);

		switch (status) {
		case RPC_SUCCESS:
#if 0 /* notyet */
			if ((mi->mi_flags & MI_DYNAMIC) &&
			    mi->mi_timer_type[which] != 0 &&
			    (mi->mi_curread != my_rsize ||
			    mi->mi_curwrite != my_wsize))
				(void) nfs_feedback(FEEDBACK_OK, which, mi);
#endif
			break;

		/*
		 * Unfortunately, there are servers in the world which
		 * are not coded correctly.  They are not prepared to
		 * handle RPC requests to the NFS port which are not
		 * NFS requests.  Thus, they may try to process the
		 * NFS_ACL request as if it were an NFS request.  This
		 * does not work.  Generally, an error will be generated
		 * on the client because it will not be able to decode
		 * the response from the server.  However, it seems
		 * possible that the server may not be able to decode
		 * the arguments.  Thus, the criteria for deciding
		 * whether the server supports NFS_ACL or not is whether
		 * the following RPC errors are returned from CLNT_CALL.
		 */
		case RPC_CANTDECODERES:
		case RPC_PROGUNAVAIL:
		case RPC_CANTDECODEARGS:
		case RPC_PROGVERSMISMATCH:
			mutex_enter(&mi->mi_lock);
			mi->mi_flags &= ~(MI_ACL | MI_EXTATTR);
			mutex_exit(&mi->mi_lock);
			break;

		/*
		 * If the server supports NFS_ACL but not the new ops
		 * for extended attributes, make sure we don't retry.
		 */
		case RPC_PROCUNAVAIL:
			mutex_enter(&mi->mi_lock);
			mi->mi_flags &= ~MI_EXTATTR;
			mutex_exit(&mi->mi_lock);
			break;

		case RPC_INTR:
			/*
			 * There is no way to recover from this error,
			 * even if mount option nointr is specified.
			 * SIGKILL, for example, cannot be blocked.
			 */
			rpcerr.re_status = RPC_INTR;
			rpcerr.re_errno = EINTR;
			break;

		case RPC_UDERROR:
			/*
			 * If the NFS server is local (vold) and
			 * it goes away then we get RPC_UDERROR.
			 * This is a retryable error, so we would
			 * loop, so check to see if the specific
			 * error was ECONNRESET, indicating that
			 * target did not exist at all.  If so,
			 * return with RPC_PROGUNAVAIL and
			 * ECONNRESET to indicate why.
			 */
			CLNT_GETERR(client, &rpcerr);
			if (rpcerr.re_errno == ECONNRESET) {
				rpcerr.re_status = RPC_PROGUNAVAIL;
				rpcerr.re_errno = ECONNRESET;
				break;
			}
			/*FALLTHROUGH*/

		default:		/* probably RPC_TIMEDOUT */
			if (IS_UNRECOVERABLE_RPC(status))
				break;

			/*
			 * increment server not responding count
			 */
			mutex_enter(&mi->mi_lock);
			mi->mi_noresponse++;
			mutex_exit(&mi->mi_lock);
#ifdef DEBUG
			nfscl->nfscl_stat.noresponse.value.ui64++;
#endif

			if (!(mi->mi_flags & MI_HARD)) {
				if (!(mi->mi_flags & MI_SEMISOFT) ||
				    (mi->mi_acl_ss_call_type[which] == 0))
					break;
			}

			/*
			 * The call is in progress (over COTS).
			 * Try the CLNT_CALL again, but don't
			 * print a noisy error message.
			 */
			if (status == RPC_INPROGRESS) {
				tryagain = TRUE;
				break;
			}

			if (flags & RFSCALL_SOFT)
				break;

			/*
			 * On zone shutdown, just move on.
			 */
			if (zone_status_get(curproc->p_zone) >=
			    ZONE_IS_SHUTTING_DOWN) {
				rpcerr.re_status = RPC_FAILED;
				rpcerr.re_errno = EIO;
				break;
			}

			/*
			 * NFS client failover support
			 *
			 * If the current server just failed us, we'll
			 * start the process of finding a new server.
			 * After that, we can just retry.
			 */
			if (FAILOVER_MOUNT(mi) && failover_safe(fi)) {
				if (svp == mi->mi_curr_serv)
					failover_newserver(mi);
				clfree_impl(client, ch, nfscl);
				goto failoverretry;
			}

			tryagain = TRUE;
			timeo = backoff(timeo);
			mutex_enter(&mi->mi_lock);
			if (!(mi->mi_flags & MI_PRINTED)) {
				mi->mi_flags |= MI_PRINTED;
				mutex_exit(&mi->mi_lock);
#ifdef DEBUG
				zprintf(zoneid,
			"NFS_ACL%d server %s not responding still trying\n",
				    mi->mi_vers, svp->sv_hostname);
#else
				zprintf(zoneid,
			    "NFS server %s not responding still trying\n",
				    svp->sv_hostname);
#endif
			} else
				mutex_exit(&mi->mi_lock);
			if (*douprintf && curproc->p_sessp->s_vp != NULL) {
				*douprintf = 0;
				if (!(mi->mi_flags & MI_NOPRINT))
#ifdef DEBUG
					uprintf(
			"NFS_ACL%d server %s not responding still trying\n",
					    mi->mi_vers, svp->sv_hostname);
#else
					uprintf(
			    "NFS server %s not responding still trying\n",
					    svp->sv_hostname);
#endif
			}

#if 0 /* notyet */
			/*
			 * If doing dynamic adjustment of transfer
			 * size and if it's a read or write call
			 * and if the transfer size changed while
			 * retransmitting or if the feedback routine
			 * changed the transfer size,
			 * then exit rfscall so that the transfer
			 * size can be adjusted at the vnops level.
			 */
			if ((mi->mi_flags & MI_DYNAMIC) &&
			    mi->mi_acl_timer_type[which] != 0 &&
			    (mi->mi_curread != my_rsize ||
			    mi->mi_curwrite != my_wsize ||
			    nfs_feedback(FEEDBACK_REXMIT1, which, mi))) {
				/*
				 * On read or write calls, return
				 * back to the vnode ops level if
				 * the transfer size changed.
				 */
				clfree_impl(client, ch, nfscl);
				if (cred_cloned)
					crfree(cr);
				return (ENFS_TRYAGAIN);
			}
#endif
		}
	} while (tryagain);

	if (status != RPC_SUCCESS) {
		/*
		 * Let soft mounts use the timed out message.
		 */
		if (status == RPC_INPROGRESS)
			status = RPC_TIMEDOUT;
		nfscl->nfscl_stat.badcalls.value.ui64++;
		if (status == RPC_CANTDECODERES ||
		    status == RPC_PROGUNAVAIL ||
		    status == RPC_PROCUNAVAIL ||
		    status == RPC_CANTDECODEARGS ||
		    status == RPC_PROGVERSMISMATCH)
			CLNT_GETERR(client, &rpcerr);
		else if (status != RPC_INTR) {
			mutex_enter(&mi->mi_lock);
			mi->mi_flags |= MI_DOWN;
			mutex_exit(&mi->mi_lock);
			CLNT_GETERR(client, &rpcerr);
#ifdef DEBUG
			bufp = clnt_sperror(client, svp->sv_hostname);
			zprintf(zoneid, "NFS_ACL%d %s failed for %s\n",
			    mi->mi_vers, mi->mi_aclnames[which], bufp);
			if (curproc->p_sessp->s_vp != NULL) {
				if (!(mi->mi_flags & MI_NOPRINT)) {
					uprintf("NFS_ACL%d %s failed for %s\n",
					    mi->mi_vers, mi->mi_aclnames[which],
					    bufp);
				}
			}
			kmem_free(bufp, MAXPATHLEN);
#else
			zprintf(zoneid,
			    "NFS %s failed for server %s: error %d (%s)\n",
			    mi->mi_aclnames[which], svp->sv_hostname,
			    status, clnt_sperrno(status));
			if (curproc->p_sessp->s_vp != NULL) {
				if (!(mi->mi_flags & MI_NOPRINT))
					uprintf(
				"NFS %s failed for server %s: error %d (%s)\n",
					    mi->mi_aclnames[which],
					    svp->sv_hostname, status,
					    clnt_sperrno(status));
			}
#endif
			/*
			 * when CLNT_CALL() fails with RPC_AUTHERROR,
			 * re_errno is set appropriately depending on
			 * the authentication error
			 */
			if (status == RPC_VERSMISMATCH ||
			    status == RPC_PROGVERSMISMATCH)
				rpcerr.re_errno = EIO;
		}
	} else {
		/*
		 * Test the value of mi_down and mi_printed without
		 * holding the mi_lock mutex.  If they are both zero,
		 * then it is okay to skip the down and printed
		 * processing.  This saves on a mutex_enter and
		 * mutex_exit pair for a normal, successful RPC.
		 * This was just complete overhead.
		 */
		if (mi->mi_flags & (MI_DOWN | MI_PRINTED)) {
			mutex_enter(&mi->mi_lock);
			mi->mi_flags &= ~MI_DOWN;
			if (mi->mi_flags & MI_PRINTED) {
				mi->mi_flags &= ~MI_PRINTED;
				mutex_exit(&mi->mi_lock);
#ifdef DEBUG
				zprintf(zoneid, "NFS_ACL%d server %s ok\n",
				    mi->mi_vers, svp->sv_hostname);
#else
				zprintf(zoneid, "NFS server %s ok\n",
				    svp->sv_hostname);
#endif
			} else
				mutex_exit(&mi->mi_lock);
		}

		if (*douprintf == 0) {
			if (!(mi->mi_flags & MI_NOPRINT))
#ifdef DEBUG
				uprintf("NFS_ACL%d server %s ok\n",
				    mi->mi_vers, svp->sv_hostname);
#else
				uprintf("NFS server %s ok\n", svp->sv_hostname);
#endif
			*douprintf = 1;
		}
	}

	clfree_impl(client, ch, nfscl);
	if (cred_cloned)
		crfree(cr);

	ASSERT(rpcerr.re_status == RPC_SUCCESS || rpcerr.re_errno != 0);

#if 0 /* notyet */
	TRACE_1(TR_FAC_NFS, TR_RFSCALL_END, "rfscall_end:errno %d",
	    rpcerr.re_errno);
#endif

	return (rpcerr.re_errno);
}

int
vattr_to_sattr(struct vattr *vap, struct nfssattr *sa)
{
	uint_t mask = vap->va_mask;

	if (!(mask & AT_MODE))
		sa->sa_mode = (uint32_t)-1;
	else
		sa->sa_mode = vap->va_mode;
	if (!(mask & AT_UID))
		sa->sa_uid = (uint32_t)-1;
	else
		sa->sa_uid = (uint32_t)vap->va_uid;
	if (!(mask & AT_GID))
		sa->sa_gid = (uint32_t)-1;
	else
		sa->sa_gid = (uint32_t)vap->va_gid;
	if (!(mask & AT_SIZE))
		sa->sa_size = (uint32_t)-1;
	else
		sa->sa_size = (uint32_t)vap->va_size;
	if (!(mask & AT_ATIME))
		sa->sa_atime.tv_sec = sa->sa_atime.tv_usec = (int32_t)-1;
	else {
		/* check time validity */
		if (! NFS_TIME_T_OK(vap->va_atime.tv_sec)) {
			return (EOVERFLOW);
		}
		sa->sa_atime.tv_sec = vap->va_atime.tv_sec;
		sa->sa_atime.tv_usec = vap->va_atime.tv_nsec / 1000;
	}
	if (!(mask & AT_MTIME))
		sa->sa_mtime.tv_sec = sa->sa_mtime.tv_usec = (int32_t)-1;
	else {
		/* check time validity */
		if (! NFS_TIME_T_OK(vap->va_mtime.tv_sec)) {
			return (EOVERFLOW);
		}
		sa->sa_mtime.tv_sec = vap->va_mtime.tv_sec;
		sa->sa_mtime.tv_usec = vap->va_mtime.tv_nsec / 1000;
	}
	return (0);
}

int
vattr_to_sattr3(struct vattr *vap, sattr3 *sa)
{
	uint_t mask = vap->va_mask;

	if (!(mask & AT_MODE))
		sa->mode.set_it = FALSE;
	else {
		sa->mode.set_it = TRUE;
		sa->mode.mode = (mode3)vap->va_mode;
	}
	if (!(mask & AT_UID))
		sa->uid.set_it = FALSE;
	else {
		sa->uid.set_it = TRUE;
		sa->uid.uid = (uid3)vap->va_uid;
	}
	if (!(mask & AT_GID))
		sa->gid.set_it = FALSE;
	else {
		sa->gid.set_it = TRUE;
		sa->gid.gid = (gid3)vap->va_gid;
	}
	if (!(mask & AT_SIZE))
		sa->size.set_it = FALSE;
	else {
		sa->size.set_it = TRUE;
		sa->size.size = (size3)vap->va_size;
	}
	if (!(mask & AT_ATIME))
		sa->atime.set_it = DONT_CHANGE;
	else {
		/* check time validity */
		if (! NFS_TIME_T_OK(vap->va_atime.tv_sec)) {
			return (EOVERFLOW);
		}
		sa->atime.set_it = SET_TO_CLIENT_TIME;
		sa->atime.atime.seconds = (uint32)vap->va_atime.tv_sec;
		sa->atime.atime.nseconds = (uint32)vap->va_atime.tv_nsec;
	}
	if (!(mask & AT_MTIME))
		sa->mtime.set_it = DONT_CHANGE;
	else {
		/* check time validity */
		if (! NFS_TIME_T_OK(vap->va_mtime.tv_sec)) {
			return (EOVERFLOW);
		}
		sa->mtime.set_it = SET_TO_CLIENT_TIME;
		sa->mtime.mtime.seconds = (uint32)vap->va_mtime.tv_sec;
		sa->mtime.mtime.nseconds = (uint32)vap->va_mtime.tv_nsec;
	}
	return (0);
}

void
setdiropargs(struct nfsdiropargs *da, char *nm, vnode_t *dvp)
{

	da->da_fhandle = VTOFH(dvp);
	da->da_name = nm;
	da->da_flags = 0;
}

void
setdiropargs3(diropargs3 *da, char *nm, vnode_t *dvp)
{

	da->dirp = VTOFH3(dvp);
	da->name = nm;
}

int
setdirgid(vnode_t *dvp, gid_t *gidp, cred_t *cr)
{
	int error;
	rnode_t *rp;
	struct vattr va;

	va.va_mask = AT_MODE | AT_GID;
	error = VOP_GETATTR(dvp, &va, 0, cr);
	if (error)
		return (error);

	/*
	 * To determine the expected group-id of the created file:
	 *  1)	If the filesystem was not mounted with the Old-BSD-compatible
	 *	GRPID option, and the directory's set-gid bit is clear,
	 *	then use the process's gid.
	 *  2)	Otherwise, set the group-id to the gid of the parent directory.
	 */
	rp = VTOR(dvp);
	mutex_enter(&rp->r_statelock);
	if (!(VTOMI(dvp)->mi_flags & MI_GRPID) && !(va.va_mode & VSGID))
		*gidp = crgetgid(cr);
	else
		*gidp = va.va_gid;
	mutex_exit(&rp->r_statelock);
	return (0);
}

int
setdirmode(vnode_t *dvp, mode_t *omp, cred_t *cr)
{
	int error;
	struct vattr va;

	va.va_mask = AT_MODE;
	error = VOP_GETATTR(dvp, &va, 0, cr);
	if (error)
		return (error);

	/*
	 * Modify the expected mode (om) so that the set-gid bit matches
	 * that of the parent directory (dvp).
	 */
	if (va.va_mode & VSGID)
		*omp |= VSGID;
	else
		*omp &= ~VSGID;
	return (0);
}

void
nfs_setswaplike(vnode_t *vp, vattr_t *vap)
{

	if (vp->v_type == VREG && (vap->va_mode & (VEXEC | VSVTX)) == VSVTX) {
		if (!(vp->v_flag & VSWAPLIKE)) {
			mutex_enter(&vp->v_lock);
			vp->v_flag |= VSWAPLIKE;
			mutex_exit(&vp->v_lock);
		}
	} else {
		if (vp->v_flag & VSWAPLIKE) {
			mutex_enter(&vp->v_lock);
			vp->v_flag &= ~VSWAPLIKE;
			mutex_exit(&vp->v_lock);
		}
	}
}

/*
 * Free the resources associated with an rnode.
 */
static void
rinactive(rnode_t *rp, cred_t *cr)
{
	vnode_t *vp;
	cred_t *cred;
	char *contents;
	int size;
	vsecattr_t *vsp;
	int error;
	nfs3_pathconf_info *info;

	/*
	 * Before freeing anything, wait until all asynchronous
	 * activity is done on this rnode.  This will allow all
	 * asynchronous read ahead and write behind i/o's to
	 * finish.
	 */
	mutex_enter(&rp->r_statelock);
	while (rp->r_count > 0)
		cv_wait(&rp->r_cv, &rp->r_statelock);
	mutex_exit(&rp->r_statelock);

	/*
	 * Flush and invalidate all pages associated with the vnode.
	 */
	vp = RTOV(rp);
	if (vn_has_cached_data(vp)) {
		ASSERT(vp->v_type != VCHR);
		if ((rp->r_flags & RDIRTY) && !rp->r_error) {
			error = VOP_PUTPAGE(vp, (u_offset_t)0, 0, 0, cr);
			if (error && (error == ENOSPC || error == EDQUOT)) {
				mutex_enter(&rp->r_statelock);
				if (!rp->r_error)
					rp->r_error = error;
				mutex_exit(&rp->r_statelock);
			}
		}
		nfs_invalidate_pages(vp, (u_offset_t)0, cr);
	}

	/*
	 * Free any held credentials and caches which may be associated
	 * with this rnode.
	 */
	mutex_enter(&rp->r_statelock);
	cred = rp->r_cred;
	rp->r_cred = NULL;
	contents = rp->r_symlink.contents;
	size = rp->r_symlink.size;
	rp->r_symlink.contents = NULL;
	vsp = rp->r_secattr;
	rp->r_secattr = NULL;
	info = rp->r_pathconf;
	rp->r_pathconf = NULL;
	mutex_exit(&rp->r_statelock);

	/*
	 * Free the held credential.
	 */
	if (cred != NULL)
		crfree(cred);

	/*
	 * Free the access cache entries.
	 */
	(void) nfs_access_purge_rp(rp);

	/*
	 * Free the readdir cache entries.
	 */
	if (HAVE_RDDIR_CACHE(rp))
		nfs_purge_rddir_cache(vp);

	/*
	 * Free the symbolic link cache.
	 */
	if (contents != NULL) {

		kmem_free((void *)contents, size);
	}

	/*
	 * Free any cached ACL.
	 */
	if (vsp != NULL)
		nfs_acl_free(vsp);

	/*
	 * Free any cached pathconf information.
	 */
	if (info != NULL)
		kmem_free(info, sizeof (*info));
}

/*
 * Return a vnode for the given NFS Version 2 file handle.
 * If no rnode exists for this fhandle, create one and put it
 * into the hash queues.  If the rnode for this fhandle
 * already exists, return it.
 *
 * Note: make_rnode() may upgrade the hash bucket lock to exclusive.
 */
vnode_t *
makenfsnode(fhandle_t *fh, struct nfsfattr *attr, struct vfs *vfsp,
    hrtime_t t, cred_t *cr, char *dnm, char *nm)
{
	int newnode;
	int index;
	vnode_t *vp;
	nfs_fhandle nfh;
	vattr_t va;

	nfh.fh_len = NFS_FHSIZE;
	bcopy(fh, nfh.fh_buf, NFS_FHSIZE);

	index = rtablehash(&nfh);
	rw_enter(&rtable[index].r_lock, RW_READER);

	vp = make_rnode(&nfh, &rtable[index], vfsp, nfs_vnodeops,
	    nfs_putapage, nfs_rddir_compar, &newnode, cr, dnm, nm);

	if (attr != NULL) {
		if (!newnode) {
			rw_exit(&rtable[index].r_lock);
			(void) nfs_cache_fattr(vp, attr, &va, t, cr);
		} else {
			if (attr->na_type < NFNON || attr->na_type > NFSOC)
				vp->v_type = VBAD;
			else
				vp->v_type = n2v_type(attr);
			/*
			 * A translation here seems to be necessary
			 * because this function can be called
			 * with `attr' that has come from the wire,
			 * and been operated on by vattr_to_nattr().
			 * See nfsrootvp()->VOP_GETTATTR()->nfsgetattr()
			 * ->nfs_getattr_otw()->rfscall()->vattr_to_nattr()
			 * ->makenfsnode().
			 */
			if ((attr->na_rdev & 0xffff0000) == 0)
				vp->v_rdev = nfsv2_expdev(attr->na_rdev);
			else
				vp->v_rdev = expldev(n2v_rdev(attr));
			nfs_attrcache(vp, attr, t);
			rw_exit(&rtable[index].r_lock);
		}
	} else {
		if (newnode) {
			PURGE_ATTRCACHE(vp);
		}
		rw_exit(&rtable[index].r_lock);
	}

	return (vp);
}

/*
 * Return a vnode for the given NFS Version 3 file handle.
 * If no rnode exists for this fhandle, create one and put it
 * into the hash queues.  If the rnode for this fhandle
 * already exists, return it.
 *
 * Note: make_rnode() may upgrade the hash bucket lock to exclusive.
 */
vnode_t *
makenfs3node_va(nfs_fh3 *fh, vattr_t *vap, struct vfs *vfsp, hrtime_t t,
    cred_t *cr, char *dnm, char *nm)
{
	int newnode;
	int index;
	vnode_t *vp;

	index = rtablehash((nfs_fhandle *)fh);
	rw_enter(&rtable[index].r_lock, RW_READER);

	vp = make_rnode((nfs_fhandle *)fh, &rtable[index], vfsp,
	    nfs3_vnodeops, nfs3_putapage, nfs3_rddir_compar, &newnode, cr,
	    dnm, nm);

	if (vap == NULL) {
		if (newnode) {
			PURGE_ATTRCACHE(vp);
		}
		rw_exit(&rtable[index].r_lock);
		return (vp);
	}

	if (!newnode) {
		rw_exit(&rtable[index].r_lock);
		nfs_attr_cache(vp, vap, t, cr);
	} else {
		rnode_t *rp = VTOR(vp);

		vp->v_type = vap->va_type;
		vp->v_rdev = vap->va_rdev;

		mutex_enter(&rp->r_statelock);
		if (rp->r_mtime <= t)
			nfs_attrcache_va(vp, vap);
		mutex_exit(&rp->r_statelock);
		rw_exit(&rtable[index].r_lock);
	}

	return (vp);
}

vnode_t *
makenfs3node(nfs_fh3 *fh, fattr3 *attr, struct vfs *vfsp, hrtime_t t,
    cred_t *cr, char *dnm, char *nm)
{
	int newnode;
	int index;
	vnode_t *vp;
	vattr_t va;

	index = rtablehash((nfs_fhandle *)fh);
	rw_enter(&rtable[index].r_lock, RW_READER);

	vp = make_rnode((nfs_fhandle *)fh, &rtable[index], vfsp,
	    nfs3_vnodeops, nfs3_putapage, nfs3_rddir_compar, &newnode, cr,
	    dnm, nm);

	if (attr == NULL) {
		if (newnode) {
			PURGE_ATTRCACHE(vp);
		}
		rw_exit(&rtable[index].r_lock);
		return (vp);
	}

	if (!newnode) {
		rw_exit(&rtable[index].r_lock);
		(void) nfs3_cache_fattr3(vp, attr, &va, t, cr);
	} else {
		if (attr->type < NF3REG || attr->type > NF3FIFO)
			vp->v_type = VBAD;
		else
			vp->v_type = nf3_to_vt[attr->type];
		vp->v_rdev = makedevice(attr->rdev.specdata1,
			    attr->rdev.specdata2);
		nfs3_attrcache(vp, attr, t);
		rw_exit(&rtable[index].r_lock);
	}

	return (vp);
}

/*
 * Read this comment before making changes to rtablehash()!
 * This is a hash function in which seemingly obvious and harmless
 * changes can cause escalations costing million dollars!
 * Know what you are doing.
 *
 * rtablehash() implements Jenkins' one-at-a-time hash algorithm.  The
 * algorithm is currently detailed here:
 *
 *   http://burtleburtle.net/bob/hash/doobs.html
 *
 * Of course, the above link may not be valid by the time you are reading
 * this, but suffice it to say that the one-at-a-time algorithm works well in
 * almost all cases.  If you are changing the algorithm be sure to verify that
 * the hash algorithm still provides even distribution in all cases and with
 * any server returning filehandles in whatever order (sequential or random).
 */
static int
rtablehash(nfs_fhandle *fh)
{
	ulong_t hash, len, i;
	char *key;

	key = fh->fh_buf;
	len = (ulong_t)fh->fh_len;
	for (hash = 0, i = 0; i < len; i++) {
		hash += key[i];
		hash += (hash << 10);
		hash ^= (hash >> 6);
	}
	hash += (hash << 3);
	hash ^= (hash >> 11);
	hash += (hash << 15);
	return (hash & rtablemask);
}

static vnode_t *
make_rnode(nfs_fhandle *fh, rhashq_t *rhtp, struct vfs *vfsp,
    struct vnodeops *vops,
    int (*putapage)(vnode_t *, page_t *, u_offset_t *, size_t *, int, cred_t *),
    int (*compar)(const void *, const void *),
    int *newnode, cred_t *cr, char *dnm, char *nm)
{
	rnode_t *rp;
	rnode_t *trp;
	vnode_t *vp;
	mntinfo_t *mi;

	ASSERT(RW_READ_HELD(&rhtp->r_lock));

	mi = VFTOMI(vfsp);
start:
	if ((rp = rfind(rhtp, fh, vfsp)) != NULL) {
		vp = RTOV(rp);
		nfs_set_vroot(vp);
		*newnode = 0;
		return (vp);
	}
	rw_exit(&rhtp->r_lock);

	mutex_enter(&rpfreelist_lock);
	if (rpfreelist != NULL && rnew >= nrnode) {
		rp = rpfreelist;
		rp_rmfree(rp);
		mutex_exit(&rpfreelist_lock);

		vp = RTOV(rp);

		if (rp->r_flags & RHASHED) {
			rw_enter(&rp->r_hashq->r_lock, RW_WRITER);
			mutex_enter(&vp->v_lock);
			if (vp->v_count > 1) {
				vp->v_count--;
				mutex_exit(&vp->v_lock);
				rw_exit(&rp->r_hashq->r_lock);
				rw_enter(&rhtp->r_lock, RW_READER);
				goto start;
			}
			mutex_exit(&vp->v_lock);
			rp_rmhash_locked(rp);
			rw_exit(&rp->r_hashq->r_lock);
		}

		rinactive(rp, cr);

		mutex_enter(&vp->v_lock);
		if (vp->v_count > 1) {
			vp->v_count--;
			mutex_exit(&vp->v_lock);
			rw_enter(&rhtp->r_lock, RW_READER);
			goto start;
		}
		mutex_exit(&vp->v_lock);
		vn_invalid(vp);
		/*
		 * destroy old locks before bzero'ing and
		 * recreating the locks below.
		 */
		nfs_rw_destroy(&rp->r_rwlock);
		nfs_rw_destroy(&rp->r_lkserlock);
		mutex_destroy(&rp->r_statelock);
		cv_destroy(&rp->r_cv);
		cv_destroy(&rp->r_commit.c_cv);
		nfs_free_r_path(rp);
		avl_destroy(&rp->r_dir);
		/*
		 * Make sure that if rnode is recycled then
		 * VFS count is decremented properly before
		 * reuse.
		 */
		VFS_RELE(vp->v_vfsp);
		vn_reinit(vp);
	} else {
		vnode_t *new_vp;

		mutex_exit(&rpfreelist_lock);

		rp = kmem_cache_alloc(rnode_cache, KM_SLEEP);
		new_vp = vn_alloc(KM_SLEEP);

		atomic_add_long((ulong_t *)&rnew, 1);
#ifdef DEBUG
		clstat_debug.nrnode.value.ui64++;
#endif
		vp = new_vp;
	}

	bzero(rp, sizeof (*rp));
	rp->r_vnode = vp;
	nfs_rw_init(&rp->r_rwlock, NULL, RW_DEFAULT, NULL);
	nfs_rw_init(&rp->r_lkserlock, NULL, RW_DEFAULT, NULL);
	mutex_init(&rp->r_statelock, NULL, MUTEX_DEFAULT, NULL);
	cv_init(&rp->r_cv, NULL, CV_DEFAULT, NULL);
	cv_init(&rp->r_commit.c_cv, NULL, CV_DEFAULT, NULL);
	rp->r_fh.fh_len = fh->fh_len;
	bcopy(fh->fh_buf, rp->r_fh.fh_buf, fh->fh_len);
	rp->r_server = mi->mi_curr_serv;
	if (FAILOVER_MOUNT(mi)) {
		/*
		 * If replicated servers, stash pathnames
		 */
		if (dnm != NULL && nm != NULL) {
			char *s, *p;
			uint_t len;

			len = (uint_t)(strlen(dnm) + strlen(nm) + 2);
			rp->r_path = kmem_alloc(len, KM_SLEEP);
#ifdef DEBUG
			clstat_debug.rpath.value.ui64 += len;
#endif
			s = rp->r_path;
			for (p = dnm; *p; p++)
				*s++ = *p;
			*s++ = '/';
			for (p = nm; *p; p++)
				*s++ = *p;
			*s = '\0';
		} else {
			/* special case for root */
			rp->r_path = kmem_alloc(2, KM_SLEEP);
#ifdef DEBUG
			clstat_debug.rpath.value.ui64 += 2;
#endif
			*rp->r_path = '.';
			*(rp->r_path + 1) = '\0';
		}
	}
	VFS_HOLD(vfsp);
	rp->r_putapage = putapage;
	rp->r_hashq = rhtp;
	rp->r_flags = RREADDIRPLUS;
	avl_create(&rp->r_dir, compar, sizeof (rddir_cache),
	    offsetof(rddir_cache, tree));
	vn_setops(vp, vops);
	vp->v_data = (caddr_t)rp;
	vp->v_vfsp = vfsp;
	vp->v_type = VNON;
	nfs_set_vroot(vp);

	/*
	 * There is a race condition if someone else
	 * alloc's the rnode while no locks are held, so we
	 * check again and recover if found.
	 */
	rw_enter(&rhtp->r_lock, RW_WRITER);
	if ((trp = rfind(rhtp, fh, vfsp)) != NULL) {
		vp = RTOV(trp);
		nfs_set_vroot(vp);
		*newnode = 0;
		rw_exit(&rhtp->r_lock);
		rp_addfree(rp, cr);
		rw_enter(&rhtp->r_lock, RW_READER);
		return (vp);
	}
	rp_addhash(rp);
	*newnode = 1;
	return (vp);
}

static void
nfs_set_vroot(vnode_t *vp)
{
	rnode_t *rp;
	nfs_fhandle *rootfh;

	rp = VTOR(vp);
	rootfh = &rp->r_server->sv_fhandle;
	if (rootfh->fh_len == rp->r_fh.fh_len &&
	    bcmp(rootfh->fh_buf, rp->r_fh.fh_buf, rp->r_fh.fh_len) == 0) {
		if (!(vp->v_flag & VROOT)) {
			mutex_enter(&vp->v_lock);
			vp->v_flag |= VROOT;
			mutex_exit(&vp->v_lock);
		}
	}
}

static void
nfs_free_r_path(rnode_t *rp)
{
	char *path;
	size_t len;

	path = rp->r_path;
	if (path) {
		rp->r_path = NULL;
		len = strlen(path) + 1;
		kmem_free(path, len);
#ifdef DEBUG
		clstat_debug.rpath.value.ui64 -= len;
#endif
	}
}

/*
 * Put an rnode on the free list.
 *
 * Rnodes which were allocated above and beyond the normal limit
 * are immediately freed.
 */
void
rp_addfree(rnode_t *rp, cred_t *cr)
{
	vnode_t *vp;
	struct vfs *vfsp;

	vp = RTOV(rp);
	ASSERT(vp->v_count >= 1);
	ASSERT(rp->r_freef == NULL && rp->r_freeb == NULL);

	/*
	 * If we have too many rnodes allocated and there are no
	 * references to this rnode, or if the rnode is no longer
	 * accessible by it does not reside in the hash queues,
	 * or if an i/o error occurred while writing to the file,
	 * then just free it instead of putting it on the rnode
	 * freelist.
	 */
	vfsp = vp->v_vfsp;
	if (((rnew > nrnode || !(rp->r_flags & RHASHED) || rp->r_error ||
	    (vfsp->vfs_flag & VFS_UNMOUNTED)) && rp->r_count == 0)) {
		if (rp->r_flags & RHASHED) {
			rw_enter(&rp->r_hashq->r_lock, RW_WRITER);
			mutex_enter(&vp->v_lock);
			if (vp->v_count > 1) {
				vp->v_count--;
				mutex_exit(&vp->v_lock);
				rw_exit(&rp->r_hashq->r_lock);
				return;
			}
			mutex_exit(&vp->v_lock);
			rp_rmhash_locked(rp);
			rw_exit(&rp->r_hashq->r_lock);
		}

		rinactive(rp, cr);

		/*
		 * Recheck the vnode reference count.  We need to
		 * make sure that another reference has not been
		 * acquired while we were not holding v_lock.  The
		 * rnode is not in the rnode hash queues, so the
		 * only way for a reference to have been acquired
		 * is for a VOP_PUTPAGE because the rnode was marked
		 * with RDIRTY or for a modified page.  This
		 * reference may have been acquired before our call
		 * to rinactive.  The i/o may have been completed,
		 * thus allowing rinactive to complete, but the
		 * reference to the vnode may not have been released
		 * yet.  In any case, the rnode can not be destroyed
		 * until the other references to this vnode have been
		 * released.  The other references will take care of
		 * either destroying the rnode or placing it on the
		 * rnode freelist.  If there are no other references,
		 * then the rnode may be safely destroyed.
		 */
		mutex_enter(&vp->v_lock);
		if (vp->v_count > 1) {
			vp->v_count--;
			mutex_exit(&vp->v_lock);
			return;
		}
		mutex_exit(&vp->v_lock);

		destroy_rnode(rp);
		return;
	}

	/*
	 * Lock the hash queue and then recheck the reference count
	 * to ensure that no other threads have acquired a reference
	 * to indicate that the rnode should not be placed on the
	 * freelist.  If another reference has been acquired, then
	 * just release this one and let the other thread complete
	 * the processing of adding this rnode to the freelist.
	 */
	rw_enter(&rp->r_hashq->r_lock, RW_WRITER);

	mutex_enter(&vp->v_lock);
	if (vp->v_count > 1) {
		vp->v_count--;
		mutex_exit(&vp->v_lock);
		rw_exit(&rp->r_hashq->r_lock);
		return;
	}
	mutex_exit(&vp->v_lock);

	/*
	 * If there is no cached data or metadata for this file, then
	 * put the rnode on the front of the freelist so that it will
	 * be reused before other rnodes which may have cached data or
	 * metadata associated with them.
	 */
	mutex_enter(&rpfreelist_lock);
	if (rpfreelist == NULL) {
		rp->r_freef = rp;
		rp->r_freeb = rp;
		rpfreelist = rp;
	} else {
		rp->r_freef = rpfreelist;
		rp->r_freeb = rpfreelist->r_freeb;
		rpfreelist->r_freeb->r_freef = rp;
		rpfreelist->r_freeb = rp;
		if (!vn_has_cached_data(vp) &&
		    !HAVE_RDDIR_CACHE(rp) &&
		    rp->r_symlink.contents == NULL &&
		    rp->r_secattr == NULL &&
		    rp->r_pathconf == NULL)
			rpfreelist = rp;
	}
	mutex_exit(&rpfreelist_lock);

	rw_exit(&rp->r_hashq->r_lock);
}

/*
 * Remove an rnode from the free list.
 *
 * The caller must be holding rpfreelist_lock and the rnode
 * must be on the freelist.
 */
static void
rp_rmfree(rnode_t *rp)
{

	ASSERT(MUTEX_HELD(&rpfreelist_lock));
	ASSERT(rp->r_freef != NULL && rp->r_freeb != NULL);

	if (rp == rpfreelist) {
		rpfreelist = rp->r_freef;
		if (rp == rpfreelist)
			rpfreelist = NULL;
	}

	rp->r_freeb->r_freef = rp->r_freef;
	rp->r_freef->r_freeb = rp->r_freeb;

	rp->r_freef = rp->r_freeb = NULL;
}

/*
 * Put a rnode in the hash table.
 *
 * The caller must be holding the exclusive hash queue lock.
 */
static void
rp_addhash(rnode_t *rp)
{

	ASSERT(RW_WRITE_HELD(&rp->r_hashq->r_lock));
	ASSERT(!(rp->r_flags & RHASHED));

	rp->r_hashf = rp->r_hashq->r_hashf;
	rp->