lofs - zygaena/hammerhead

root / base / usr / src / uts / common / fs / lofs
lofs Plain Text 2673 lines 65.7 KB
/*
 * 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 2008 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */
/*
 * Copyright (c) 2017 by Delphix. All rights reserved.
 */

/*
 * The idea behind composition-based stacked filesystems is to add a
 * vnode to the stack of vnodes for each mount. These vnodes have their
 * own set of mount options and filesystem-specific functions, so they
 * can modify data or operations before they are passed along. Such a
 * filesystem must maintain a mapping from the underlying vnodes to its
 * interposing vnodes.
 *
 * In lofs, this mapping is implemented by a hashtable. Each bucket
 * contains a count of the number of nodes currently contained, the
 * chain of vnodes, and a lock to protect the list of vnodes. The
 * hashtable dynamically grows if the number of vnodes in the table as a
 * whole exceeds the size of the table left-shifted by
 * lo_resize_threshold. In order to minimize lock contention, there is
 * no global lock protecting the hashtable, hence obtaining the
 * per-bucket locks consists of a dance to make sure we've actually
 * locked the correct bucket. Acquiring a bucket lock doesn't involve
 * locking the hashtable itself, so we refrain from freeing old
 * hashtables, and store them in a linked list of retired hashtables;
 * the list is freed when the filesystem is unmounted.
 */

#include <sys/param.h>
#include <sys/kmem.h>
#include <sys/vfs.h>
#include <sys/vnode.h>
#include <sys/cmn_err.h>
#include <sys/systm.h>
#include <sys/t_lock.h>
#include <sys/debug.h>
#include <sys/atomic.h>

#include <sys/fs/lofs_node.h>
#include <sys/fs/lofs_info.h>
/*
 * Due to the hashing algorithm, the size of the hash table needs to be a
 * power of 2.
 */
#define	LOFS_DEFAULT_HTSIZE	(1 << 6)

#define	ltablehash(vp, tblsz)	((((intptr_t)(vp))>>10) & ((tblsz)-1))

/*
 * The following macros can only be safely used when the desired bucket
 * is already locked.
 */
/*
 * The lock in the hashtable associated with the given vnode.
 */
#define	TABLE_LOCK(vp, li)      \
	(&(li)->li_hashtable[ltablehash((vp), (li)->li_htsize)].lh_lock)

/*
 * The bucket in the hashtable that the given vnode hashes to.
 */
#define	TABLE_BUCKET(vp, li)    \
	((li)->li_hashtable[ltablehash((vp), (li)->li_htsize)].lh_chain)

/*
 * Number of elements currently in the bucket that the vnode hashes to.
 */
#define	TABLE_COUNT(vp, li)	\
	((li)->li_hashtable[ltablehash((vp), (li)->li_htsize)].lh_count)

/*
 * Grab/Drop the lock for the bucket this vnode hashes to.
 */
#define	TABLE_LOCK_ENTER(vp, li)	table_lock_enter(vp, li)
#define	TABLE_LOCK_EXIT(vp, li)		\
	mutex_exit(&(li)->li_hashtable[ltablehash((vp),	\
	    (li)->li_htsize)].lh_lock)

static lnode_t *lfind(struct vnode *, struct loinfo *);
static void lsave(lnode_t *, struct loinfo *);
static struct vfs *makelfsnode(struct vfs *, struct loinfo *);
static struct lfsnode *lfsfind(struct vfs *, struct loinfo *);

uint_t lo_resize_threshold = 1;
uint_t lo_resize_factor = 2;

static kmem_cache_t *lnode_cache;

/*
 * Since the hashtable itself isn't protected by a lock, obtaining a
 * per-bucket lock proceeds as follows:
 *
 * (a) li->li_htlock protects li->li_hashtable, li->li_htsize, and
 * li->li_retired.
 *
 * (b) Per-bucket locks (lh_lock) protect the contents of the bucket.
 *
 * (c) Locking order for resizing the hashtable is li_htlock then
 * lh_lock.
 *
 * To grab the bucket lock we:
 *
 * (1) Stash away the htsize and the pointer to the hashtable to make
 * sure neither change while we're using them.
 *
 * (2) lgrow() updates the pointer to the hashtable before it updates
 * the size: the worst case scenario is that we have the wrong size (but
 * the correct table), so we hash to the wrong bucket, grab the wrong
 * lock, and then realize that things have changed, rewind and start
 * again. If both the size and the table changed since we loaded them,
 * we'll realize that too and restart.
 *
 * (3) The protocol for growing the hashtable involves holding *all* the
 * locks in the table, hence the unlocking code (TABLE_LOCK_EXIT())
 * doesn't need to do any dances, since neither the table nor the size
 * can change while any bucket lock is held.
 *
 * (4) If the hashtable is growing (by thread t1) while another thread
 * (t2) is trying to grab a bucket lock, t2 might have a stale reference
 * to li->li_htsize:
 *
 * - t1 grabs all locks in lgrow()
 * 	- t2 loads li->li_htsize and li->li_hashtable
 * - t1 changes li->hashtable
 * 	- t2 loads from an offset in the "stale" hashtable and tries to grab
 * 	the relevant mutex.
 *
 * If t1 had free'd the stale hashtable, t2 would be in trouble. Hence,
 * stale hashtables are not freed but stored in a list of "retired"
 * hashtables, which is emptied when the filesystem is unmounted.
 */
static void
table_lock_enter(vnode_t *vp, struct loinfo *li)
{
	struct lobucket *chain;
	uint_t htsize;
	uint_t hash;

	for (;;) {
		htsize = li->li_htsize;
		membar_consumer();
		chain = (struct lobucket *)li->li_hashtable;
		hash = ltablehash(vp, htsize);
		mutex_enter(&chain[hash].lh_lock);
		if (li->li_hashtable == chain && li->li_htsize == htsize)
			break;
		mutex_exit(&chain[hash].lh_lock);
	}
}

void
lofs_subrinit(void)
{
	/*
	 * Initialize the cache.
	 */
	lnode_cache = kmem_cache_create("lnode_cache", sizeof (lnode_t),
	    0, NULL, NULL, NULL, NULL, NULL, 0);
}

void
lofs_subrfini(void)
{
	kmem_cache_destroy(lnode_cache);
}

/*
 * Initialize a (struct loinfo), and initialize the hashtable to have
 * htsize buckets.
 */
void
lsetup(struct loinfo *li, uint_t htsize)
{
	li->li_refct = 0;
	li->li_lfs = NULL;
	if (htsize == 0)
		htsize = LOFS_DEFAULT_HTSIZE;
	li->li_htsize = htsize;
	li->li_hashtable = kmem_zalloc(htsize * sizeof (*li->li_hashtable),
	    KM_SLEEP);
	mutex_init(&li->li_lfslock, NULL, MUTEX_DEFAULT, NULL);
	mutex_init(&li->li_htlock, NULL, MUTEX_DEFAULT, NULL);
	li->li_retired = NULL;
}

/*
 * Destroy a (struct loinfo)
 */
void
ldestroy(struct loinfo *li)
{
	uint_t i, htsize;
	struct lobucket *table;
	struct lo_retired_ht *lrhp, *trhp;

	mutex_destroy(&li->li_htlock);
	mutex_destroy(&li->li_lfslock);
	htsize = li->li_htsize;
	table = li->li_hashtable;
	for (i = 0; i < htsize; i++)
		mutex_destroy(&table[i].lh_lock);
	kmem_free(table, htsize * sizeof (*li->li_hashtable));

	/*
	 * Free the retired hashtables.
	 */
	lrhp = li->li_retired;
	while (lrhp != NULL) {
		trhp = lrhp;
		lrhp = lrhp->lrh_next;
		kmem_free(trhp->lrh_table,
		    trhp->lrh_size * sizeof (*li->li_hashtable));
		kmem_free(trhp, sizeof (*trhp));
	}
	li->li_retired = NULL;
}

/*
 * Return a looped back vnode for the given vnode.
 * If no lnode exists for this vnode create one and put it
 * in a table hashed by vnode.  If the lnode for
 * this vnode is already in the table return it (ref count is
 * incremented by lfind).  The lnode will be flushed from the
 * table when lo_inactive calls freelonode.  The creation of
 * a new lnode can be forced via the LOF_FORCE flag even if
 * the vnode exists in the table.  This is used in the creation
 * of a terminating lnode when looping is detected.  A unique
 * lnode is required for the correct evaluation of the current
 * working directory.
 * NOTE: vp is assumed to be a held vnode.
 */
struct vnode *
makelonode(struct vnode *vp, struct loinfo *li, int flag)
{
	lnode_t *lp, *tlp;
	struct vfs *vfsp;
	vnode_t *nvp;

	lp = NULL;
	TABLE_LOCK_ENTER(vp, li);
	if (flag != LOF_FORCE)
		lp = lfind(vp, li);
	if ((flag == LOF_FORCE) || (lp == NULL)) {
		/*
		 * Optimistically assume that we won't need to sleep.
		 */
		lp = kmem_cache_alloc(lnode_cache, KM_NOSLEEP);
		nvp = vn_alloc(KM_NOSLEEP);
		if (lp == NULL || nvp == NULL) {
			TABLE_LOCK_EXIT(vp, li);
			/* The lnode allocation may have succeeded, save it */
			tlp = lp;
			if (tlp == NULL) {
				tlp = kmem_cache_alloc(lnode_cache, KM_SLEEP);
			}
			if (nvp == NULL) {
				nvp = vn_alloc(KM_SLEEP);
			}
			lp = NULL;
			TABLE_LOCK_ENTER(vp, li);
			if (flag != LOF_FORCE)
				lp = lfind(vp, li);
			if (lp != NULL) {
				kmem_cache_free(lnode_cache, tlp);
				vn_free(nvp);
				VN_RELE(vp);
				goto found_lnode;
			}
			lp = tlp;
		}
		atomic_inc_32(&li->li_refct);
		vfsp = makelfsnode(vp->v_vfsp, li);
		lp->lo_vnode = nvp;
		VN_SET_VFS_TYPE_DEV(nvp, vfsp, vp->v_type, vp->v_rdev);
		nvp->v_flag |= (vp->v_flag & (VNOMOUNT|VNOMAP|VDIROPEN));
		vn_setops(nvp, lo_vnodeops);
		nvp->v_data = (caddr_t)lp;
		lp->lo_vp = vp;
		lp->lo_looping = 0;
		lsave(lp, li);
		vn_exists(vp);
	} else {
		VN_RELE(vp);
	}

found_lnode:
	TABLE_LOCK_EXIT(vp, li);
	return (ltov(lp));
}

/*
 * Get/Make vfs structure for given real vfs
 */
static struct vfs *
makelfsnode(struct vfs *vfsp, struct loinfo *li)
{
	struct lfsnode *lfs;
	struct lfsnode *tlfs;

	/*
	 * Don't grab any locks for the fast (common) case.
	 */
	if (vfsp == li->li_realvfs)
		return (li->li_mountvfs);
	ASSERT(li->li_refct > 0);
	mutex_enter(&li->li_lfslock);
	if ((lfs = lfsfind(vfsp, li)) == NULL) {
		mutex_exit(&li->li_lfslock);
		lfs = kmem_zalloc(sizeof (*lfs), KM_SLEEP);
		mutex_enter(&li->li_lfslock);
		if ((tlfs = lfsfind(vfsp, li)) != NULL) {
			kmem_free(lfs, sizeof (*lfs));
			lfs = tlfs;
			goto found_lfs;
		}
		lfs->lfs_realvfs = vfsp;

		/*
		 * Even though the lfsnode is strictly speaking a private
		 * implementation detail of lofs, it should behave as a regular
		 * vfs_t for the benefit of the rest of the kernel.
		 */
		VFS_INIT(&lfs->lfs_vfs, lo_vfsops, (caddr_t)li);
		lfs->lfs_vfs.vfs_fstype = li->li_mountvfs->vfs_fstype;
		lfs->lfs_vfs.vfs_flag =
		    ((vfsp->vfs_flag | li->li_mflag) & ~li->li_dflag) &
		    INHERIT_VFS_FLAG;
		lfs->lfs_vfs.vfs_bsize = vfsp->vfs_bsize;
		lfs->lfs_vfs.vfs_dev = vfsp->vfs_dev;
		lfs->lfs_vfs.vfs_fsid = vfsp->vfs_fsid;

		if (vfsp->vfs_mntpt != NULL) {
			lfs->lfs_vfs.vfs_mntpt = vfs_getmntpoint(vfsp);
			/* Leave a reference to the mountpoint */
		}

		(void) VFS_ROOT(vfsp, &lfs->lfs_realrootvp);

		/*
		 * We use 1 instead of 0 as the value to associate with
		 * an idle lfs_vfs.  This is to prevent VFS_RELE()
		 * trying to kmem_free() our lfs_t (which is the wrong
		 * size).
		 */
		VFS_HOLD(&lfs->lfs_vfs);
		lfs->lfs_next = li->li_lfs;
		li->li_lfs = lfs;
		vfs_propagate_features(vfsp, &lfs->lfs_vfs);
	}

found_lfs:
	VFS_HOLD(&lfs->lfs_vfs);
	mutex_exit(&li->li_lfslock);
	return (&lfs->lfs_vfs);
}

/*
 * Free lfs node since no longer in use
 */
static void
freelfsnode(struct lfsnode *lfs, struct loinfo *li)
{
	struct lfsnode *prev = NULL;
	struct lfsnode *this;

	ASSERT(MUTEX_HELD(&li->li_lfslock));
	ASSERT(li->li_refct > 0);
	for (this = li->li_lfs; this != NULL; this = this->lfs_next) {
		if (this == lfs) {
			ASSERT(lfs->lfs_vfs.vfs_count == 1);
			if (prev == NULL)
				li->li_lfs = lfs->lfs_next;
			else
				prev->lfs_next = lfs->lfs_next;
			if (lfs->lfs_realrootvp != NULL) {
				VN_RELE(lfs->lfs_realrootvp);
			}
			if (lfs->lfs_vfs.vfs_mntpt != NULL)
				refstr_rele(lfs->lfs_vfs.vfs_mntpt);
			if (lfs->lfs_vfs.vfs_implp != NULL) {
				ASSERT(lfs->lfs_vfs.vfs_femhead == NULL);
				ASSERT(lfs->lfs_vfs.vfs_vskap == NULL);
				ASSERT(lfs->lfs_vfs.vfs_fstypevsp == NULL);
				kmem_free(lfs->lfs_vfs.vfs_implp,
				    sizeof (vfs_impl_t));
			}
			sema_destroy(&lfs->lfs_vfs.vfs_reflock);
			kmem_free(lfs, sizeof (struct lfsnode));
			return;
		}
		prev = this;
	}
	panic("freelfsnode");
	/*NOTREACHED*/
}

/*
 * Find lfs given real vfs and mount instance(li)
 */
static struct lfsnode *
lfsfind(struct vfs *vfsp, struct loinfo *li)
{
	struct lfsnode *lfs;

	ASSERT(MUTEX_HELD(&li->li_lfslock));

	/*
	 * We need to handle the case where a UFS filesystem was forced
	 * unmounted and then a subsequent mount got the same vfs
	 * structure.  If the new mount lies in the lofs hierarchy, then
	 * this will confuse lofs, because the original vfsp (of the
	 * forced unmounted filesystem) is still around. We check for
	 * this condition here.
	 *
	 * If we find a cache vfsp hit, then we check to see if the
	 * cached filesystem was forced unmounted. Skip all such
	 * entries. This should be safe to do since no
	 * makelonode()->makelfsnode()->lfsfind() calls should be
	 * generated for such force-unmounted filesystems (because (ufs)
	 * lookup would've returned an error).
	 */
	for (lfs = li->li_lfs; lfs != NULL; lfs = lfs->lfs_next) {
		if (lfs->lfs_realvfs == vfsp) {
			struct vnode *realvp;

			realvp = lfs->lfs_realrootvp;
			if (realvp == NULL)
				continue;
			if (realvp->v_vfsp == NULL || realvp->v_type == VBAD)
				continue;
			return (lfs);
		}
	}
	return (NULL);
}

/*
 * Find real vfs given loopback vfs
 */
struct vfs *
lo_realvfs(struct vfs *vfsp, struct vnode **realrootvpp)
{
	struct loinfo *li = vtoli(vfsp);
	struct lfsnode *lfs;

	ASSERT(li->li_refct > 0);
	if (vfsp == li->li_mountvfs) {
		if (realrootvpp != NULL)
			*realrootvpp = vtol(li->li_rootvp)->lo_vp;
		return (li->li_realvfs);
	}
	mutex_enter(&li->li_lfslock);
	for (lfs = li->li_lfs; lfs != NULL; lfs = lfs->lfs_next) {
		if (vfsp == &lfs->lfs_vfs) {
			if (realrootvpp != NULL)
				*realrootvpp = lfs->lfs_realrootvp;
			mutex_exit(&li->li_lfslock);
			return (lfs->lfs_realvfs);
		}
	}
	panic("lo_realvfs");
	/*NOTREACHED*/
}

/*
 * Lnode lookup stuff.
 * These routines maintain a table of lnodes hashed by vp so
 * that the lnode for a vp can be found if it already exists.
 *
 * NB: A lofs shadow vnode causes exactly one VN_HOLD() on the
 * underlying vnode.
 */

/*
 * Retire old hashtables.
 */
static void
lretire(struct loinfo *li, struct lobucket *table, uint_t size)
{
	struct lo_retired_ht *lrhp;

	lrhp = kmem_alloc(sizeof (*lrhp), KM_SLEEP);
	lrhp->lrh_table = table;
	lrhp->lrh_size = size;

	mutex_enter(&li->li_htlock);
	lrhp->lrh_next = li->li_retired;
	li->li_retired = lrhp;
	mutex_exit(&li->li_htlock);
}

/*
 * Grow the hashtable.
 */
static void
lgrow(struct loinfo *li, uint_t newsize)
{
	uint_t oldsize;
	uint_t i;
	struct lobucket *oldtable, *newtable;

	/*
	 * It's OK to not have enough memory to resize the hashtable.
	 * We'll go down this path the next time we add something to the
	 * table, and retry the allocation then.
	 */
	if ((newtable = kmem_zalloc(newsize * sizeof (*li->li_hashtable),
	    KM_NOSLEEP)) == NULL)
		return;

	mutex_enter(&li->li_htlock);
	if (newsize <= li->li_htsize) {
		mutex_exit(&li->li_htlock);
		kmem_free(newtable, newsize * sizeof (*li->li_hashtable));
		return;
	}
	oldsize = li->li_htsize;
	oldtable = li->li_hashtable;

	/*
	 * Grab all locks so TABLE_LOCK_ENTER() calls block until the
	 * resize is complete.
	 */
	for (i = 0; i < oldsize; i++)
		mutex_enter(&oldtable[i].lh_lock);
	/*
	 * li->li_hashtable gets set before li->li_htsize, so in the
	 * time between the two assignments, callers of
	 * TABLE_LOCK_ENTER() cannot hash to a bucket beyond oldsize,
	 * hence we only need to grab the locks up to oldsize.
	 */
	for (i = 0; i < oldsize; i++)
		mutex_enter(&newtable[i].lh_lock);
	/*
	 * Rehash.
	 */
	for (i = 0; i < oldsize; i++) {
		lnode_t *tlp, *nlp;

		for (tlp = oldtable[i].lh_chain; tlp != NULL; tlp = nlp) {
			uint_t hash = ltablehash(tlp->lo_vp, newsize);

			nlp = tlp->lo_next;
			tlp->lo_next = newtable[hash].lh_chain;
			newtable[hash].lh_chain = tlp;
			newtable[hash].lh_count++;
		}
	}

	/*
	 * As soon as we store the new hashtable, future locking operations
	 * will use it.  Therefore, we must ensure that all the state we've
	 * just established reaches global visibility before the new hashtable
	 * does.
	 */
	membar_producer();
	li->li_hashtable = newtable;

	/*
	 * table_lock_enter() relies on the fact that li->li_hashtable
	 * is set to its new value before li->li_htsize.
	 */
	membar_producer();
	li->li_htsize = newsize;

	/*
	 * The new state is consistent now, so we can drop all the locks.
	 */
	for (i = 0; i < oldsize; i++) {
		mutex_exit(&newtable[i].lh_lock);
		mutex_exit(&oldtable[i].lh_lock);
	}
	mutex_exit(&li->li_htlock);

	lretire(li, oldtable, oldsize);
}

/*
 * Put a lnode in the table
 */
static void
lsave(lnode_t *lp, struct loinfo *li)
{
	ASSERT(lp->lo_vp);
	ASSERT(MUTEX_HELD(TABLE_LOCK(lp->lo_vp, li)));

#ifdef LODEBUG
	lo_dprint(4, "lsave lp %p hash %d\n",
	    lp, ltablehash(lp->lo_vp, li));
#endif

	TABLE_COUNT(lp->lo_vp, li)++;
	lp->lo_next = TABLE_BUCKET(lp->lo_vp, li);
	TABLE_BUCKET(lp->lo_vp, li) = lp;

	if (li->li_refct > (li->li_htsize << lo_resize_threshold)) {
		TABLE_LOCK_EXIT(lp->lo_vp, li);
		lgrow(li, li->li_htsize << lo_resize_factor);
		TABLE_LOCK_ENTER(lp->lo_vp, li);
	}
}

/*
 * Our version of vfs_rele() that stops at 1 instead of 0, and calls
 * freelfsnode() instead of kmem_free().
 */
static void
lfs_rele(struct lfsnode *lfs, struct loinfo *li)
{
	vfs_t *vfsp = &lfs->lfs_vfs;

	ASSERT(MUTEX_HELD(&li->li_lfslock));
	ASSERT(vfsp->vfs_count > 1);
	if (atomic_dec_32_nv(&vfsp->vfs_count) == 1)
		freelfsnode(lfs, li);
}

/*
 * Remove a lnode from the table
 */
void
freelonode(lnode_t *lp)
{
	lnode_t *lt;
	lnode_t *ltprev = NULL;
	struct lfsnode *lfs, *nextlfs;
	struct vfs *vfsp;
	struct vnode *vp = ltov(lp);
	struct vnode *realvp = realvp(vp);
	struct loinfo *li = vtoli(vp->v_vfsp);

#ifdef LODEBUG
	lo_dprint(4, "freelonode lp %p hash %d\n",
	    lp, ltablehash(lp->lo_vp, li));
#endif
	TABLE_LOCK_ENTER(lp->lo_vp, li);

	mutex_enter(&vp->v_lock);
	if (vp->v_count > 1) {
		VN_RELE_LOCKED(vp);
		mutex_exit(&vp->v_lock);
		TABLE_LOCK_EXIT(lp->lo_vp, li);
		return;
	}
	mutex_exit(&vp->v_lock);

	for (lt = TABLE_BUCKET(lp->lo_vp, li); lt != NULL;
	    ltprev = lt, lt = lt->lo_next) {
		if (lt == lp) {
#ifdef LODEBUG
			lo_dprint(4, "freeing %p, vfsp %p\n",
			    vp, vp->v_vfsp);
#endif
			atomic_dec_32(&li->li_refct);
			vfsp = vp->v_vfsp;
			vn_invalid(vp);
			if (vfsp != li->li_mountvfs) {
				mutex_enter(&li->li_lfslock);
				/*
				 * Check for unused lfs
				 */
				lfs = li->li_lfs;
				while (lfs != NULL) {
					nextlfs = lfs->lfs_next;
					if (vfsp == &lfs->lfs_vfs) {
						lfs_rele(lfs, li);
						break;
					}
					if (lfs->lfs_vfs.vfs_count == 1) {
						/*
						 * Lfs is idle
						 */
						freelfsnode(lfs, li);
					}
					lfs = nextlfs;
				}
				mutex_exit(&li->li_lfslock);
			}
			if (ltprev == NULL) {
				TABLE_BUCKET(lt->lo_vp, li) = lt->lo_next;
			} else {
				ltprev->lo_next = lt->lo_next;
			}
			TABLE_COUNT(lt->lo_vp, li)--;
			TABLE_LOCK_EXIT(lt->lo_vp, li);
			kmem_cache_free(lnode_cache, lt);
			vn_free(vp);
			VN_RELE(realvp);
			return;
		}
	}
	panic("freelonode");
	/*NOTREACHED*/
}

/*
 * Lookup a lnode by vp
 */
static lnode_t *
lfind(struct vnode *vp, struct loinfo *li)
{
	lnode_t *lt;

	ASSERT(MUTEX_HELD(TABLE_LOCK(vp, li)));

	lt = TABLE_BUCKET(vp, li);
	while (lt != NULL) {
		if (lt->lo_vp == vp) {
			VN_HOLD(ltov(lt));
			return (lt);
		}
		lt = lt->lo_next;
	}
	return (NULL);
}

#ifdef	LODEBUG
static int lofsdebug;
#endif	/* LODEBUG */

/*
 * Utilities used by both client and server
 * Standard levels:
 * 0) no debugging
 * 1) hard failures
 * 2) soft failures
 * 3) current test software
 * 4) main procedure entry points
 * 5) main procedure exit points
 * 6) utility procedure entry points
 * 7) utility procedure exit points
 * 8) obscure procedure entry points
 * 9) obscure procedure exit points
 * 10) random stuff
 * 11) all <= 1
 * 12) all <= 2
 * 13) all <= 3
 * ...
 */

#ifdef LODEBUG
/*VARARGS2*/
lo_dprint(int level, char *str, int a1, int a2, int a3, int a4, int a5, int a6,
    int a7, int a8, int a9)
{

	if (lofsdebug == level || (lofsdebug > 10 && (lofsdebug - 10) >= level))
		printf(str, a1, a2, a3, a4, a5, a6, a7, a8, a9);
}
#endif
/*
 * 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 (c) 1991, 2010, Oracle and/or its affiliates. All rights reserved.
 * Copyright 2024 Oxide Computer Company
 */

#include <sys/param.h>
#include <sys/errno.h>
#include <sys/vfs.h>
#include <sys/vfs_opreg.h>
#include <sys/vnode.h>
#include <sys/uio.h>
#include <sys/pathname.h>
#include <sys/kmem.h>
#include <sys/cred.h>
#include <sys/statvfs.h>
#include <sys/fs/lofs_info.h>
#include <sys/fs/lofs_node.h>
#include <sys/mount.h>
#include <sys/mntent.h>
#include <sys/mkdev.h>
#include <sys/priv.h>
#include <sys/sysmacros.h>
#include <sys/systm.h>
#include <sys/cmn_err.h>
#include <sys/policy.h>
#include <sys/tsol/label.h>
#include "fs/fs_subr.h"

/*
 * This is the loadable module wrapper.
 */
#include <sys/modctl.h>

static mntopts_t lofs_mntopts;

static int lofsinit(int, char *);

static vfsdef_t vfw = {
	VFSDEF_VERSION,
	"lofs",
	lofsinit,
	VSW_HASPROTO|VSW_STATS|VSW_ZMOUNT,
	&lofs_mntopts
};

/*
 * LOFS mount options table
 */
static char *xattr_cancel[] = { MNTOPT_NOXATTR, NULL };
static char *noxattr_cancel[] = { MNTOPT_XATTR, NULL };
static char *sub_cancel[] = { MNTOPT_LOFS_NOSUB, NULL };
static char *nosub_cancel[] = { MNTOPT_LOFS_SUB, NULL };

static mntopt_t mntopts[] = {
/*
 *	option name		cancel option	default arg	flags
 *		private data
 */
	{ MNTOPT_XATTR,		xattr_cancel,	NULL,		0,
		(void *)0 },
	{ MNTOPT_NOXATTR,	noxattr_cancel,	NULL,		0,
		(void *)0 },
	{ MNTOPT_LOFS_SUB,	sub_cancel,	NULL,		0,
		(void *)0 },
	{ MNTOPT_LOFS_NOSUB,	nosub_cancel,	NULL,		0,
		(void *)0 },
};

static mntopts_t lofs_mntopts = {
	sizeof (mntopts) / sizeof (mntopt_t),
	mntopts
};

/*
 * Module linkage information for the kernel.
 */

static struct modlfs modlfs = {
	&mod_fsops, "filesystem for lofs", &vfw
};

static struct modlinkage modlinkage = {
	MODREV_1, (void *)&modlfs, NULL
};

/*
 * This is the module initialization routine.
 */

int
_init(void)
{
	int status;

	lofs_subrinit();
	status = mod_install(&modlinkage);
	if (status != 0) {
		/*
		 * Cleanup previously initialized work.
		 */
		lofs_subrfini();
	}

	return (status);
}

/*
 * Don't allow the lofs module to be unloaded for now.
 * There is a memory leak if it gets unloaded.
 */

int
_fini(void)
{
	return (EBUSY);
}

int
_info(struct modinfo *modinfop)
{
	return (mod_info(&modlinkage, modinfop));
}


static int lofsfstype;
vfsops_t *lo_vfsops;

/*
 * lo mount vfsop
 * Set up mount info record and attach it to vfs struct.
 */
/*ARGSUSED*/
static int
lo_mount(struct vfs *vfsp, struct vnode *vp, struct mounta *uap,
    struct cred *cr)
{
	int error;
	struct vnode *srootvp = NULL;	/* the server's root */
	struct vnode *realrootvp;
	struct loinfo *li;
	int nodev;

	nodev = vfs_optionisset(vfsp, MNTOPT_NODEVICES, NULL);

	if ((error = secpolicy_fs_mount(cr, vp, vfsp)) != 0)
		return (EPERM);

	/*
	 * Loopback devices which get "nodevices" added can be done without
	 * "nodevices" set because we cannot import devices into a zone
	 * with loopback.  Note that we have all zone privileges when
	 * this happens; if not, we'd have gotten "nosuid".
	 */
	if (!nodev && vfs_optionisset(vfsp, MNTOPT_NODEVICES, NULL))
		vfs_setmntopt(vfsp, MNTOPT_DEVICES, NULL, VFS_NODISPLAY);

	mutex_enter(&vp->v_lock);
	if (!(uap->flags & MS_OVERLAY) &&
	    (vp->v_count != 1 || (vp->v_flag & VROOT))) {
		mutex_exit(&vp->v_lock);
		return (EBUSY);
	}
	mutex_exit(&vp->v_lock);

	/*
	 * Find real root, and make vfs point to real vfs
	 */

	if (error = lookupname(uap->spec, (uap->flags & MS_SYSSPACE) ?
	    UIO_SYSSPACE : UIO_USERSPACE, FOLLOW, NULLVPP, &realrootvp))
		return (error);

	/*
	 * Enforce MAC policy if needed.
	 *
	 * Loopback mounts must not allow writing up. The dominance test
	 * is intended to prevent a global zone caller from accidentally
	 * creating write-up conditions between two labeled zones.
	 * Local zones can't violate MAC on their own without help from
	 * the global zone because they can't name a pathname that
	 * they don't already have.
	 *
	 * The special case check for the NET_MAC_AWARE process flag is
	 * to support the case of the automounter in the global zone. We
	 * permit automounting of local zone directories such as home
	 * directories, into the global zone as required by setlabel,
	 * zonecopy, and saving of desktop sessions. Such mounts are
	 * trusted not to expose the contents of one zone's directories
	 * to another by leaking them through the global zone.
	 */
	if (is_system_labeled() && crgetzoneid(cr) == GLOBAL_ZONEID) {
		char	specname[MAXPATHLEN];
		zone_t	*from_zptr;
		zone_t	*to_zptr;

		if (vnodetopath(NULL, realrootvp, specname,
		    sizeof (specname), CRED()) != 0) {
			VN_RELE(realrootvp);
			return (EACCES);
		}

		from_zptr = zone_find_by_path(specname);
		to_zptr = zone_find_by_path(refstr_value(vfsp->vfs_mntpt));

		/*
		 * Special case for scratch zones used for Live Upgrade:
		 * this is used to mount the zone's root from /root to /a in
		 * the scratch zone.  As with the other special case, this
		 * appears to be outside of the zone because it's not under
		 * the zone rootpath, which is $ZONEPATH/lu in the scratch
		 * zone case.
		 */

		if (from_zptr != to_zptr &&
		    !(to_zptr->zone_flags & ZF_IS_SCRATCH)) {
			/*
			 * We know at this point that the labels aren't equal
			 * because the zone pointers aren't equal, and zones
			 * can't share a label.
			 *
			 * If the source is the global zone then making
			 * it available to a local zone must be done in
			 * read-only mode as the label will become admin_low.
			 *
			 * If it is a mount between local zones then if
			 * the current process is in the global zone and has
			 * the NET_MAC_AWARE flag, then regular read-write
			 * access is allowed.  If it's in some other zone, but
			 * the label on the mount point dominates the original
			 * source, then allow the mount as read-only
			 * ("read-down").
			 */
			if (from_zptr->zone_id == GLOBAL_ZONEID) {
				/* make the mount read-only */
				vfs_setmntopt(vfsp, MNTOPT_RO, NULL, 0);
			} else { /* cross-zone mount */
				if (to_zptr->zone_id == GLOBAL_ZONEID &&
				    /* LINTED: no consequent */
				    getpflags(NET_MAC_AWARE, cr) != 0) {
					/* Allow the mount as read-write */
				} else if (bldominates(
				    label2bslabel(to_zptr->zone_slabel),
				    label2bslabel(from_zptr->zone_slabel))) {
					/* make the mount read-only */
					vfs_setmntopt(vfsp, MNTOPT_RO, NULL, 0);
				} else {
					VN_RELE(realrootvp);
					zone_rele(to_zptr);
					zone_rele(from_zptr);
					return (EACCES);
				}
			}
		}
		zone_rele(to_zptr);
		zone_rele(from_zptr);
	}

	/*
	 * realrootvp may be an AUTOFS node, in which case we perform a
	 * VOP_ACCESS() to trigger the mount of the intended filesystem.
	 * This causes a loopback mount of the intended filesystem instead
	 * of the AUTOFS filesystem.
	 *
	 * If a lofs mount creates a mount loop (such that a lofs vfs is
	 * mounted on an autofs node and that lofs vfs points back to the
	 * autofs node which it is mounted on) then a VOP_ACCESS call will
	 * create a deadlock. Once this deadlock is released, VOP_ACCESS will
	 * return EINTR. In such a case we don't want the lofs vfs to be
	 * created as the loop could panic the system.
	 */
	if ((error = VOP_ACCESS(realrootvp, 0, 0, cr, NULL)) != 0) {
		VN_RELE(realrootvp);
		return (error);
	}

	/*
	 * We're interested in the top most filesystem.
	 * This is specially important when uap->spec is a trigger
	 * AUTOFS node, since we're really interested in mounting the
	 * filesystem AUTOFS mounted as result of the VOP_ACCESS()
	 * call not the AUTOFS node itself.
	 */
	if (vn_mountedvfs(realrootvp) != NULL) {
		if (error = traverse(&realrootvp)) {
			VN_RELE(realrootvp);
			return (error);
		}
	}

	/*
	 * Allocate a vfs info struct and attach it
	 */
	li = kmem_zalloc(sizeof (struct loinfo), KM_SLEEP);
	li->li_realvfs = realrootvp->v_vfsp;
	li->li_mountvfs = vfsp;

	/*
	 * Set mount flags to be inherited by loopback vfs's
	 */
	if (vfs_optionisset(vfsp, MNTOPT_RO, NULL)) {
		li->li_mflag |= VFS_RDONLY;
	}
	if (vfs_optionisset(vfsp, MNTOPT_NOSUID, NULL)) {
		li->li_mflag |= (VFS_NOSETUID|VFS_NODEVICES);
	}
	if (vfs_optionisset(vfsp, MNTOPT_NODEVICES, NULL)) {
		li->li_mflag |= VFS_NODEVICES;
	}
	if (vfs_optionisset(vfsp, MNTOPT_NOSETUID, NULL)) {
		li->li_mflag |= VFS_NOSETUID;
	}
	/*
	 * Permissive flags are added to the "deny" bitmap.
	 */
	if (vfs_optionisset(vfsp, MNTOPT_NOXATTR, NULL)) {
		li->li_dflag |= VFS_XATTR;
	}
	if (vfs_optionisset(vfsp, MNTOPT_NONBMAND, NULL)) {
		li->li_dflag |= VFS_NBMAND;
	}

	/*
	 * Propagate inheritable mount flags from the real vfs.
	 */
	if ((li->li_realvfs->vfs_flag & VFS_RDONLY) &&
	    !vfs_optionisset(vfsp, MNTOPT_RO, NULL))
		vfs_setmntopt(vfsp, MNTOPT_RO, NULL,
		    VFS_NODISPLAY);
	if ((li->li_realvfs->vfs_flag & VFS_NOSETUID) &&
	    !vfs_optionisset(vfsp, MNTOPT_NOSETUID, NULL))
		vfs_setmntopt(vfsp, MNTOPT_NOSETUID, NULL,
		    VFS_NODISPLAY);
	if ((li->li_realvfs->vfs_flag & VFS_NODEVICES) &&
	    !vfs_optionisset(vfsp, MNTOPT_NODEVICES, NULL))
		vfs_setmntopt(vfsp, MNTOPT_NODEVICES, NULL,
		    VFS_NODISPLAY);
	/*
	 * Permissive flags such as VFS_XATTR, as opposed to restrictive flags
	 * such as VFS_RDONLY, are handled differently.  An explicit
	 * MNTOPT_NOXATTR should override the underlying filesystem's VFS_XATTR.
	 */
	if ((li->li_realvfs->vfs_flag & VFS_XATTR) &&
	    !vfs_optionisset(vfsp, MNTOPT_NOXATTR, NULL) &&
	    !vfs_optionisset(vfsp, MNTOPT_XATTR, NULL))
		vfs_setmntopt(vfsp, MNTOPT_XATTR, NULL,
		    VFS_NODISPLAY);
	if ((li->li_realvfs->vfs_flag & VFS_NBMAND) &&
	    !vfs_optionisset(vfsp, MNTOPT_NBMAND, NULL) &&
	    !vfs_optionisset(vfsp, MNTOPT_NONBMAND, NULL))
		vfs_setmntopt(vfsp, MNTOPT_NBMAND, NULL,
		    VFS_NODISPLAY);

	li->li_refct = 0;
	vfsp->vfs_data = (caddr_t)li;
	vfsp->vfs_bcount = 0;
	vfsp->vfs_fstype = lofsfstype;
	vfsp->vfs_bsize = li->li_realvfs->vfs_bsize;

	vfsp->vfs_dev = li->li_realvfs->vfs_dev;
	vfsp->vfs_fsid.val[0] = li->li_realvfs->vfs_fsid.val[0];
	vfsp->vfs_fsid.val[1] = li->li_realvfs->vfs_fsid.val[1];

	if (vfs_optionisset(vfsp, MNTOPT_LOFS_NOSUB, NULL)) {
		li->li_flag |= LO_NOSUB;
	}

	/*
	 * Propagate any VFS features
	 */

	vfs_propagate_features(li->li_realvfs, vfsp);

	/*
	 * Setup the hashtable. If the root of this mount isn't a directory,
	 * there's no point in allocating a large hashtable. A table with one
	 * bucket is sufficient.
	 */
	if (realrootvp->v_type != VDIR)
		lsetup(li, 1);
	else
		lsetup(li, 0);

	/*
	 * Make the root vnode
	 */
	srootvp = makelonode(realrootvp, li, 0);
	srootvp->v_flag |= VROOT;
	li->li_rootvp = srootvp;

#ifdef LODEBUG
	lo_dprint(4, "lo_mount: vfs %p realvfs %p root %p realroot %p li %p\n",
	    vfsp, li->li_realvfs, srootvp, realrootvp, li);
#endif
	return (0);
}

/*
 * Undo loopback mount
 */
static int
lo_unmount(struct vfs *vfsp, int flag, struct cred *cr)
{
	struct loinfo *li;

	if (secpolicy_fs_unmount(cr, vfsp) != 0)
		return (EPERM);

	/*
	 * Forced unmount is not supported by this file system
	 * and thus, ENOTSUP, is being returned.
	 */
	if (flag & MS_FORCE)
		return (ENOTSUP);

	li = vtoli(vfsp);
#ifdef LODEBUG
	lo_dprint(4, "lo_unmount(%p) li %p\n", vfsp, li);
#endif
	if (li->li_refct != 1 || li->li_rootvp->v_count != 1) {
#ifdef LODEBUG
		lo_dprint(4, "refct %d v_ct %d\n", li->li_refct,
		    li->li_rootvp->v_count);
#endif
		return (EBUSY);
	}
	VN_RELE(li->li_rootvp);
	return (0);
}

/*
 * Find root of lofs mount.
 */
static int
lo_root(struct vfs *vfsp, struct vnode **vpp)
{
	*vpp = vtoli(vfsp)->li_rootvp;
#ifdef LODEBUG
	lo_dprint(4, "lo_root(0x%p) = %p\n", vfsp, *vpp);
#endif
	/*
	 * If the root of the filesystem is a special file, return the specvp
	 * version of the vnode. We don't save the specvp vnode in our
	 * hashtable since that's exclusively for lnodes.
	 */
	if (IS_DEVVP(*vpp)) {
		struct vnode *svp;

		svp = specvp(*vpp, (*vpp)->v_rdev, (*vpp)->v_type, kcred);
		if (svp == NULL)
			return (ENOSYS);
		*vpp = svp;
	} else {
		VN_HOLD(*vpp);
	}

	return (0);
}

/*
 * Get file system statistics.
 */
static int
lo_statvfs(register struct vfs *vfsp, struct statvfs64 *sbp)
{
	vnode_t *realrootvp;

#ifdef LODEBUG
	lo_dprint(4, "lostatvfs %p\n", vfsp);
#endif
	/*
	 * Using realrootvp->v_vfsp (instead of the realvfsp that was
	 * cached) is necessary to make lofs work woth forced UFS unmounts.
	 * In the case of a forced unmount, UFS stores a set of dummy vfsops
	 * in all the (i)vnodes in the filesystem. The dummy ops simply
	 * returns back EIO.
	 */
	(void) lo_realvfs(vfsp, &realrootvp);
	if (realrootvp != NULL)
		return (VFS_STATVFS(realrootvp->v_vfsp, sbp));
	else
		return (EIO);
}

/*
 * LOFS doesn't have any data or metadata to flush, pending I/O on the
 * underlying filesystem will be flushed when such filesystem is synched.
 */
/* ARGSUSED */
static int
lo_sync(struct vfs *vfsp, short flag, struct cred *cr)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_sync: %p\n", vfsp);
#endif
	return (0);
}

/*
 * While the general sync(2) entry point above assumes that the underlying fs
 * will be synced, we treat this as a directed blocking sync on the file system
 * which means we should attempt the underlying file system.
 */
static int
lo_syncfs(vfs_t *vfsp, uint64_t flags, cred_t *cr)
{
	vfs_t *realvfs;

#ifdef LODEBUG
	lo_dprint(4, "lo_syncfs: %p\n", vfsp);
#endif
	realvfs = lo_realvfs(vfsp, NULL);
	if (realvfs != NULL) {
		return (VFS_SYNCFS(realvfs, flags, cr));
	} else {
		return (EIO);
	}

}

/*
 * Obtain the vnode from the underlying filesystem.
 */
static int
lo_vget(struct vfs *vfsp, struct vnode **vpp, struct fid *fidp)
{
	vnode_t *realrootvp;

#ifdef LODEBUG
	lo_dprint(4, "lo_vget: %p\n", vfsp);
#endif
	(void) lo_realvfs(vfsp, &realrootvp);
	if (realrootvp != NULL)
		return (VFS_VGET(realrootvp->v_vfsp, vpp, fidp));
	else
		return (EIO);
}

/*
 * Free mount-specific data.
 */
static void
lo_freevfs(struct vfs *vfsp)
{
	struct loinfo *li = vtoli(vfsp);

	ldestroy(li);
	kmem_free(li, sizeof (struct loinfo));
}

static int
lofsinit(int fstyp, char *name)
{
	static const fs_operation_def_t lo_vfsops_template[] = {
		VFSNAME_MOUNT,		{ .vfs_mount = lo_mount },
		VFSNAME_UNMOUNT,	{ .vfs_unmount = lo_unmount },
		VFSNAME_ROOT,		{ .vfs_root = lo_root },
		VFSNAME_STATVFS,	{ .vfs_statvfs = lo_statvfs },
		VFSNAME_SYNC,		{ .vfs_sync = lo_sync },
		VFSNAME_VGET,		{ .vfs_vget = lo_vget },
		VFSNAME_FREEVFS,	{ .vfs_freevfs = lo_freevfs },
		VFSNAME_SYNCFS,		{ .vfs_syncfs = lo_syncfs },
		NULL,			NULL
	};
	int error;

	error = vfs_setfsops(fstyp, lo_vfsops_template, &lo_vfsops);
	if (error != 0) {
		cmn_err(CE_WARN, "lofsinit: bad vfs ops template");
		return (error);
	}

	error = vn_make_ops(name, lo_vnodeops_template, &lo_vnodeops);
	if (error != 0) {
		(void) vfs_freevfsops_by_type(fstyp);
		cmn_err(CE_WARN, "lofsinit: bad vnode ops template");
		return (error);
	}

	lofsfstype = fstyp;

	return (0);
}
/*
 * 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 2008 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 * Copyright 2015 Joyent, Inc.
 */

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/errno.h>
#include <sys/vnode.h>
#include <sys/vfs.h>
#include <sys/vfs_opreg.h>
#include <sys/uio.h>
#include <sys/cred.h>
#include <sys/pathname.h>
#include <sys/debug.h>
#include <sys/fs/lofs_node.h>
#include <sys/fs/lofs_info.h>
#include <fs/fs_subr.h>
#include <vm/as.h>
#include <vm/seg.h>

/*
 * These are the vnode ops routines which implement the vnode interface to
 * the looped-back file system.  These routines just take their parameters,
 * and then calling the appropriate real vnode routine(s) to do the work.
 */

static int
lo_open(vnode_t **vpp, int flag, struct cred *cr, caller_context_t *ct)
{
	vnode_t *vp = *vpp;
	vnode_t *rvp;
	vnode_t *oldvp;
	int error;

#ifdef LODEBUG
	lo_dprint(4, "lo_open vp %p cnt=%d realvp %p cnt=%d\n",
	    vp, vp->v_count, realvp(vp), realvp(vp)->v_count);
#endif

	oldvp = vp;
	vp = rvp = realvp(vp);
	/*
	 * Need to hold new reference to vp since VOP_OPEN() may
	 * decide to release it.
	 */
	VN_HOLD(vp);
	error = VOP_OPEN(&rvp, flag, cr, ct);

	if (!error && rvp != vp) {
		/*
		 * the FS which we called should have released the
		 * new reference on vp
		 */
		*vpp = makelonode(rvp, vtoli(oldvp->v_vfsp), 0);
		if ((*vpp)->v_type == VDIR) {
			/*
			 * Copy over any looping flags to the new lnode.
			 */
			(vtol(*vpp))->lo_looping |= (vtol(oldvp))->lo_looping;
		}
		if (IS_DEVVP(*vpp)) {
			vnode_t *svp;

			svp = specvp(*vpp, (*vpp)->v_rdev, (*vpp)->v_type, cr);
			VN_RELE(*vpp);
			if (svp == NULL)
				error = ENOSYS;
			else
				*vpp = svp;
		}
		VN_RELE(oldvp);
	} else {
		ASSERT(rvp->v_count > 1);
		VN_RELE(rvp);
	}

	return (error);
}

static int
lo_close(
	vnode_t *vp,
	int flag,
	int count,
	offset_t offset,
	struct cred *cr,
	caller_context_t *ct)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_close vp %p realvp %p\n", vp, realvp(vp));
#endif
	vp = realvp(vp);
	return (VOP_CLOSE(vp, flag, count, offset, cr, ct));
}

static int
lo_read(vnode_t *vp, struct uio *uiop, int ioflag, struct cred *cr,
	caller_context_t *ct)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_read vp %p realvp %p\n", vp, realvp(vp));
#endif
	vp = realvp(vp);
	return (VOP_READ(vp, uiop, ioflag, cr, ct));
}

static int
lo_write(vnode_t *vp, struct uio *uiop, int ioflag, struct cred *cr,
	caller_context_t *ct)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_write vp %p realvp %p\n", vp, realvp(vp));
#endif
	vp = realvp(vp);
	return (VOP_WRITE(vp, uiop, ioflag, cr, ct));
}

static int
lo_ioctl(
	vnode_t *vp,
	int cmd,
	intptr_t arg,
	int flag,
	struct cred *cr,
	int *rvalp,
	caller_context_t *ct)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_ioctl vp %p realvp %p\n", vp, realvp(vp));
#endif
	vp = realvp(vp);
	return (VOP_IOCTL(vp, cmd, arg, flag, cr, rvalp, ct));
}

static int
lo_setfl(vnode_t *vp, int oflags, int nflags, cred_t *cr, caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_SETFL(vp, oflags, nflags, cr, ct));
}

static int
lo_getattr(
	vnode_t *vp,
	struct vattr *vap,
	int flags,
	struct cred *cr,
	caller_context_t *ct)
{
	int error;

#ifdef LODEBUG
	lo_dprint(4, "lo_getattr vp %p realvp %p\n", vp, realvp(vp));
#endif
	if (error = VOP_GETATTR(realvp(vp), vap, flags, cr, ct))
		return (error);

	return (0);
}

static int
lo_setattr(
	vnode_t *vp,
	struct vattr *vap,
	int flags,
	struct cred *cr,
	caller_context_t *ct)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_setattr vp %p realvp %p\n", vp, realvp(vp));
#endif
	vp = realvp(vp);
	return (VOP_SETATTR(vp, vap, flags, cr, ct));
}

static int
lo_access(
	vnode_t *vp,
	int mode,
	int flags,
	struct cred *cr,
	caller_context_t *ct)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_access vp %p realvp %p\n", vp, realvp(vp));
#endif
	if (mode & VWRITE) {
		if (vp->v_type == VREG && vn_is_readonly(vp))
			return (EROFS);
	}
	vp = realvp(vp);
	return (VOP_ACCESS(vp, mode, flags, cr, ct));
}

static int
lo_fsync(vnode_t *vp, int syncflag, struct cred *cr, caller_context_t *ct)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_fsync vp %p realvp %p\n", vp, realvp(vp));
#endif
	vp = realvp(vp);
	return (VOP_FSYNC(vp, syncflag, cr, ct));
}

/*ARGSUSED*/
static void
lo_inactive(vnode_t *vp, struct cred *cr, caller_context_t *ct)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_inactive %p, realvp %p\n", vp, realvp(vp));
#endif
	freelonode(vtol(vp));
}

/* ARGSUSED */
static int
lo_fid(vnode_t *vp, struct fid *fidp, caller_context_t *ct)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_fid %p, realvp %p\n", vp, realvp(vp));
#endif
	vp = realvp(vp);
	return (VOP_FID(vp, fidp, ct));
}

/*
 * Given a vnode of lofs type, lookup nm name and
 * return a shadow vnode (of lofs type) of the
 * real vnode found.
 *
 * Due to the nature of lofs, there is a potential
 * looping in path traversal.
 *
 * starting from the mount point of an lofs;
 * a loop is defined to be a traversal path
 * where the mount point or the real vnode of
 * the root of this lofs is encountered twice.
 * Once at the start of traversal and second
 * when the looping is found.
 *
 * When a loop is encountered, a shadow of the
 * covered vnode is returned to stop the looping.
 *
 * This normally works, but with the advent of
 * the new automounter, returning the shadow of the
 * covered vnode (autonode, in this case) does not
 * stop the loop.  Because further lookup on this
 * lonode will cause the autonode to call lo_lookup()
 * on the lonode covering it.
 *
 * example "/net/jurassic/net/jurassic" is a loop.
 * returning the shadow of the autonode corresponding to
 * "/net/jurassic/net/jurassic" will not terminate the
 * loop.   To solve this problem we allow the loop to go
 * through one more level component lookup.  Whichever
 * directory is then looked up in "/net/jurassic/net/jurassic"
 * the vnode returned is the vnode covered by the autonode
 * "net" and this will terminate the loop.
 *
 * Lookup for dot dot has to be dealt with separately.
 * It will be nice to have a "one size fits all" kind
 * of solution, so that we don't have so many ifs statement
 * in the lo_lookup() to handle dotdot.  But, since
 * there are so many special cases to handle different
 * kinds looping above, we need special codes to handle
 * dotdot lookup as well.
 */
static int
lo_lookup(
	vnode_t *dvp,
	char *nm,
	vnode_t **vpp,
	struct pathname *pnp,
	int flags,
	vnode_t *rdir,
	struct cred *cr,
	caller_context_t *ct,
	int *direntflags,
	pathname_t *realpnp)
{
	vnode_t *vp = NULL, *tvp = NULL, *nonlovp;
	int error, is_indirectloop;
	vnode_t *realdvp = realvp(dvp);
	struct loinfo *li = vtoli(dvp->v_vfsp);
	int looping = 0;
	int autoloop = 0;
	int doingdotdot = 0;
	int nosub = 0;
	int mkflag = 0;

	/*
	 * If name is empty and no XATTR flags are set, then return
	 * dvp (empty name == lookup ".").  If an XATTR flag is set
	 * then we need to call VOP_LOOKUP to get the xattr dir.
	 */
	if (nm[0] == '\0' && ! (flags & (CREATE_XATTR_DIR|LOOKUP_XATTR))) {
		VN_HOLD(dvp);
		*vpp = dvp;
		return (0);
	}

	if (nm[0] == '.' && nm[1] == '.' && nm[2] == '\0') {
		doingdotdot++;
		/*
		 * Handle ".." out of mounted filesystem
		 */
		while ((realdvp->v_flag & VROOT) && realdvp != rootdir) {
			realdvp = realdvp->v_vfsp->vfs_vnodecovered;
			ASSERT(realdvp != NULL);
		}
	}

	*vpp = NULL;	/* default(error) case */

	/*
	 * Do the normal lookup
	 */
	if (error = VOP_LOOKUP(realdvp, nm, &vp, pnp, flags, rdir, cr,
	    ct, direntflags, realpnp)) {
		vp = NULL;
		goto out;
	}

	/*
	 * We do this check here to avoid returning a stale file handle to the
	 * caller.
	 */
	if (nm[0] == '.' && nm[1] == '\0') {
		ASSERT(vp == realdvp);
		VN_HOLD(dvp);
		VN_RELE(vp);
		*vpp = dvp;
		return (0);
	}

	if (doingdotdot) {
		if ((vtol(dvp))->lo_looping & LO_LOOPING) {
			vfs_t *vfsp;

			error = vn_vfsrlock_wait(realdvp);
			if (error)
				goto out;
			vfsp = vn_mountedvfs(realdvp);
			/*
			 * In the standard case if the looping flag is set and
			 * performing dotdot we would be returning from a
			 * covered vnode, implying vfsp could not be null. The
			 * exceptions being if we have looping and overlay
			 * mounts or looping and covered file systems.
			 */
			if (vfsp == NULL) {
				/*
				 * Overlay mount or covered file system,
				 * so just make the shadow node.
				 */
				vn_vfsunlock(realdvp);
				*vpp = makelonode(vp, li, 0);
				(vtol(*vpp))->lo_looping |= LO_LOOPING;
				return (0);
			}
			/*
			 * When looping get the actual found vnode
			 * instead of the vnode covered.
			 * Here we have to hold the lock for realdvp
			 * since an unmount during the traversal to the
			 * root vnode would turn *vfsp into garbage
			 * which would be fatal.
			 */
			error = VFS_ROOT(vfsp, &tvp);
			vn_vfsunlock(realdvp);

			if (error)
				goto out;

			if ((tvp == li->li_rootvp) && (vp == realvp(tvp))) {
				/*
				 * we're back at the real vnode
				 * of the rootvp
				 *
				 * return the rootvp
				 * Ex: /mnt/mnt/..
				 * where / has been lofs-mounted
				 * onto /mnt.  Return the lofs
				 * node mounted at /mnt.
				 */
				*vpp = tvp;
				VN_RELE(vp);
				return (0);
			} else {
				/*
				 * We are returning from a covered
				 * node whose vfs_mountedhere is
				 * not pointing to vfs of the current
				 * root vnode.
				 * This is a condn where in we
				 * returned a covered node say Zc
				 * but Zc is not the cover of current
				 * root.
				 * i.e.., if X is the root vnode
				 * lookup(Zc,"..") is taking us to
				 * X.
				 * Ex: /net/X/net/X/Y
				 *
				 * If LO_AUTOLOOP (autofs/lofs looping detected)
				 * has been set then we are encountering the
				 * cover of Y (Y being any directory vnode
				 * under /net/X/net/X/).
				 * When performing a dotdot set the
				 * returned vp to the vnode covered
				 * by the mounted lofs, ie /net/X/net/X
				 */
				VN_RELE(tvp);
				if ((vtol(dvp))->lo_looping & LO_AUTOLOOP) {
					VN_RELE(vp);
					vp = li->li_rootvp;
					vp = vp->v_vfsp->vfs_vnodecovered;
					VN_HOLD(vp);
					*vpp = makelonode(vp, li, 0);
					(vtol(*vpp))->lo_looping |= LO_LOOPING;
					return (0);
				}
			}
		} else {
			/*
			 * No frills just make the shadow node.
			 */
			*vpp = makelonode(vp, li, 0);
			return (0);
		}
	}

	nosub = (vtoli(dvp->v_vfsp)->li_flag & LO_NOSUB);

	/*
	 * If this vnode is mounted on, then we
	 * traverse to the vnode which is the root of
	 * the mounted file system.
	 */
	if (!nosub && (error = traverse(&vp)))
		goto out;

	/*
	 * Make a lnode for the real vnode.
	 */
	if (vp->v_type != VDIR || nosub) {
		*vpp = makelonode(vp, li, 0);
		if (IS_DEVVP(*vpp)) {
			vnode_t *svp;

			svp = specvp(*vpp, (*vpp)->v_rdev, (*vpp)->v_type, cr);
			VN_RELE(*vpp);
			if (svp == NULL)
				error = ENOSYS;
			else
				*vpp = svp;
		}
		return (error);
	}

	/*
	 * if the found vnode (vp) is not of type lofs
	 * then we're just going to make a shadow of that
	 * vp and get out.
	 *
	 * If the found vnode (vp) is of lofs type, and
	 * we're not doing dotdot, check if we are
	 * looping.
	 */
	if (!doingdotdot && vfs_matchops(vp->v_vfsp, lo_vfsops)) {
		/*
		 * Check if we're looping, i.e.
		 * vp equals the root vp of the lofs, directly
		 * or indirectly, return the covered node.
		 */

		if (!((vtol(dvp))->lo_looping & LO_LOOPING)) {
			if (vp == li->li_rootvp) {
				/*
				 * Direct looping condn.
				 * Ex:- X is / mounted directory so lookup of
				 * /X/X is a direct looping condn.
				 */
				tvp = vp;
				vp = vp->v_vfsp->vfs_vnodecovered;
				VN_HOLD(vp);
				VN_RELE(tvp);
				looping++;
			} else {
				/*
				 * Indirect looping can be defined as
				 * real lookup returning rootvp of the current
				 * tree in any level of recursion.
				 *
				 * This check is useful if there are multiple
				 * levels of lofs indirections. Suppose vnode X
				 * in the current lookup has as its real vnode
				 * another lofs node. Y = realvp(X) Y should be
				 * a lofs node for the check to continue or Y
				 * is not the rootvp of X.
				 * Ex:- say X and Y are two vnodes
				 * say real(Y) is X and real(X) is Z
				 * parent vnode for X and Y is Z
				 * lookup(Y,"path") say we are looking for Y
				 * again under Y and we have to return Yc.
				 * but the lookup of Y under Y doesnot return
				 * Y the root vnode again here is why.
				 * 1. lookup(Y,"path of Y") will go to
				 * 2. lookup(real(Y),"path of Y") and then to
				 * 3. lookup(real(X),"path of Y").
				 * and now what lookup level 1 sees is the
				 * outcome of 2 but the vnode Y is due to
				 * lookup(Z,"path of Y") so we have to skip
				 * intermediate levels to find if in any level
				 * there is a looping.
				 */
				is_indirectloop = 0;
				nonlovp = vp;
				while (
				    vfs_matchops(nonlovp->v_vfsp, lo_vfsops) &&
				    !(is_indirectloop)) {
					if (li->li_rootvp  == nonlovp) {
						is_indirectloop++;
						break;
					}
					nonlovp = realvp(nonlovp);
				}

				if (is_indirectloop) {
					VN_RELE(vp);
					vp = nonlovp;
					vp = vp->v_vfsp->vfs_vnodecovered;
					VN_HOLD(vp);
					looping++;
				}
			}
		} else {
			/*
			 * come here only because of the interaction between
			 * the autofs and lofs.
			 *
			 * Lookup of "/net/X/net/X" will return a shadow of
			 * an autonode X_a which we call X_l.
			 *
			 * Lookup of anything under X_l, will trigger a call to
			 * auto_lookup(X_a,nm) which will eventually call
			 * lo_lookup(X_lr,nm) where X_lr is the root vnode of
			 * the current lofs.
			 *
			 * We come here only when we are called with X_l as dvp
			 * and look for something underneath.
			 *
			 * Now that an autofs/lofs looping condition has been
			 * identified any directory vnode contained within
			 * dvp will be set to the vnode covered by the
			 * mounted autofs. Thus all directories within dvp
			 * will appear empty hence teminating the looping.
			 * The LO_AUTOLOOP flag is set on the returned lonode
			 * to indicate the termination of the autofs/lofs
			 * looping. This is required for the correct behaviour
			 * when performing a dotdot.
			 */
			realdvp = realvp(dvp);
			while (vfs_matchops(realdvp->v_vfsp, lo_vfsops)) {
				realdvp = realvp(realdvp);
			}

			error = VFS_ROOT(realdvp->v_vfsp, &tvp);
			if (error)
				goto out;
			/*
			 * tvp now contains the rootvp of the vfs of the
			 * real vnode of dvp. The directory vnode vp is set
			 * to the covered vnode to terminate looping. No
			 * distinction is made between any vp as all directory
			 * vnodes contained in dvp are returned as the covered
			 * vnode.
			 */
			VN_RELE(vp);
			vp = tvp;	/* possibly is an autonode */

			/*
			 * Need to find the covered vnode
			 */
			if (vp->v_vfsp->vfs_vnodecovered == NULL) {
				/*
				 * We don't have a covered vnode so this isn't
				 * an autonode. To find the autonode simply
				 * find the vnode covered by the lofs rootvp.
				 */
				vp = li->li_rootvp;
				vp = vp->v_vfsp->vfs_vnodecovered;
				VN_RELE(tvp);
				error = VFS_ROOT(vp->v_vfsp, &tvp);
				if (error)
					goto out;
				vp = tvp;	/* now this is an autonode */
				if (vp->v_vfsp->vfs_vnodecovered == NULL) {
					/*
					 * Still can't find a covered vnode.
					 * Fail the lookup, or we'd loop.
					 */
					error = ENOENT;
					goto out;
				}
			}
			vp = vp->v_vfsp->vfs_vnodecovered;
			VN_HOLD(vp);
			VN_RELE(tvp);
			/*
			 * Force the creation of a new lnode even if the hash
			 * table contains a lnode that references this vnode.
			 */
			mkflag = LOF_FORCE;
			autoloop++;
		}
	}
	*vpp = makelonode(vp, li, mkflag);

	if ((looping) ||
	    (((vtol(dvp))->lo_looping & LO_LOOPING) && !doingdotdot)) {
		(vtol(*vpp))->lo_looping |= LO_LOOPING;
	}

	if (autoloop) {
		(vtol(*vpp))->lo_looping |= LO_AUTOLOOP;
	}

out:
	if (error != 0 && vp != NULL)
		VN_RELE(vp);
#ifdef LODEBUG
	lo_dprint(4,
	"lo_lookup dvp %x realdvp %x nm '%s' newvp %x real vp %x error %d\n",
	    dvp, realvp(dvp), nm, *vpp, vp, error);
#endif
	return (error);
}

/*ARGSUSED*/
static int
lo_create(
	vnode_t *dvp,
	char *nm,
	struct vattr *va,
	enum vcexcl exclusive,
	int mode,
	vnode_t **vpp,
	struct cred *cr,
	int flag,
	caller_context_t *ct,
	vsecattr_t *vsecp)
{
	int error;
	vnode_t *vp = NULL;

#ifdef LODEBUG
	lo_dprint(4, "lo_create vp %p realvp %p\n", dvp, realvp(dvp));
#endif
	if (*nm == '\0') {
		ASSERT(vpp && dvp == *vpp);
		vp = realvp(*vpp);
	}

	error = VOP_CREATE(realvp(dvp), nm, va, exclusive, mode, &vp, cr, flag,
	    ct, vsecp);
	if (!error) {
		*vpp = makelonode(vp, vtoli(dvp->v_vfsp), 0);
		if (IS_DEVVP(*vpp)) {
			vnode_t *svp;

			svp = specvp(*vpp, (*vpp)->v_rdev, (*vpp)->v_type, cr);
			VN_RELE(*vpp);
			if (svp == NULL)
				error = ENOSYS;
			else
				*vpp = svp;
		}
	} else if (error == ENOSYS && exclusive == NONEXCL &&
	    dvp == vtoli(dvp->v_vfsp)->li_rootvp &&
	    realvp(dvp)->v_type == VREG) {
		/*
		 * We have a single regular file lofs mounted, thus the file is
		 * the root vnode (the directory vp is the file vp). Some
		 * underlying file systems (e.g. tmpfs or ufs) properly handle
		 * this style of create but at least zfs won't support create
		 * this way (see zfs_fvnodeops_template which has fs_nosys for
		 * the vop_create entry because zfs_create doesn't work
		 * properly for this case).
		 */
		if ((error = VOP_ACCESS(dvp, mode, 0, cr, NULL)) == 0) {
			/*
			 * Since we already know the vnode for the existing
			 * file we can handle create as a no-op, as expected,
			 * truncating the file if necessary.
			 */
			struct vattr vattr;

			vattr.va_size = 0;
			vattr.va_mask = AT_SIZE;

			if ((va->va_mask & AT_SIZE) != 0 && va->va_size == 0 &&
			    VOP_SETATTR(dvp, &vattr, 0, CRED(), NULL) != 0)
				return (error);

			/*
			 * vn_createat will do a vn_rele on the file if it is
			 * pre-existing, which it is in the case of a single
			 * file mounted as the root. Thus, when we eventually
			 * close the file the count will already be 1 so the
			 * vnode would be freed. To prevent that, we add an
			 * extra hold here.
			 */
			VN_HOLD(dvp);
			*vpp = dvp;
			error = 0;
		}
	}

	return (error);
}

static int
lo_remove(
	vnode_t *dvp,
	char *nm,
	struct cred *cr,
	caller_context_t *ct,
	int flags)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_remove vp %p realvp %p\n", dvp, realvp(dvp));
#endif
	dvp = realvp(dvp);
	return (VOP_REMOVE(dvp, nm, cr, ct, flags));
}

static int
lo_link(
	vnode_t *tdvp,
	vnode_t *vp,
	char *tnm,
	struct cred *cr,
	caller_context_t *ct,
	int flags)
{
	vnode_t *realvp;

#ifdef LODEBUG
	lo_dprint(4, "lo_link vp %p realvp %p\n", vp, realvp(vp));
#endif

	/*
	 * The source and destination vnodes may be in different lofs
	 * filesystems sharing the same underlying filesystem, so we need to
	 * make sure that the filesystem containing the source vnode is not
	 * mounted read-only (vn_link() has already checked the target vnode).
	 *
	 * In a situation such as:
	 *
	 * /data	- regular filesystem
	 * /foo		- lofs mount of /data/foo
	 * /bar		- read-only lofs mount of /data/bar
	 *
	 * This disallows a link from /bar/somefile to /foo/somefile,
	 * which would otherwise allow changes to somefile on the read-only
	 * mounted /bar.
	 */

	if (vn_is_readonly(vp)) {
		return (EROFS);
	}
	while (vn_matchops(vp, lo_vnodeops)) {
		vp = realvp(vp);
	}

	/*
	 * In the case where the source vnode is on another stacking
	 * filesystem (such as specfs), the loop above will
	 * terminate before finding the true underlying vnode.
	 *
	 * We use VOP_REALVP here to continue the search.
	 */
	if (VOP_REALVP(vp, &realvp, ct) == 0)
		vp = realvp;

	while (vn_matchops(tdvp, lo_vnodeops)) {
		tdvp = realvp(tdvp);
	}
	if (vp->v_vfsp != tdvp->v_vfsp)
		return (EXDEV);
	return (VOP_LINK(tdvp, vp, tnm, cr, ct, flags));
}

static int
lo_rename(
	vnode_t *odvp,
	char *onm,
	vnode_t *ndvp,
	char *nnm,
	struct cred *cr,
	caller_context_t *ct,
	int flags)
{
	vnode_t *tnvp;

#ifdef LODEBUG
	lo_dprint(4, "lo_rename vp %p realvp %p\n", odvp, realvp(odvp));
#endif
	/*
	 * If we are coming from a loop back mounted fs, that has been
	 * mounted in the same filesystem as where we want to move to,
	 * and that filesystem is read/write, but the lofs filesystem is
	 * read only, we don't want to allow a rename of the file. The
	 * vn_rename code checks to be sure the target is read/write already
	 * so that is not necessary here. However, consider the following
	 * example:
	 *		/ - regular root fs
	 *		/foo - directory in root
	 *		/foo/bar - file in foo directory(in root fs)
	 *		/baz - directory in root
	 *		mount -F lofs -o ro /foo /baz - all still in root
	 *			directory
	 * The fact that we mounted /foo on /baz read only should stop us
	 * from renaming the file /foo/bar /bar, but it doesn't since
	 * / is read/write. We are still renaming here since we are still
	 * in the same filesystem, it is just that we do not check to see
	 * if the filesystem we are coming from in this case is read only.
	 */
	if (odvp->v_vfsp->vfs_flag & VFS_RDONLY)
		return (EROFS);
	/*
	 * We need to make sure we're not trying to remove a mount point for a
	 * filesystem mounted on top of lofs, which only we know about.
	 */
	if (vn_matchops(ndvp, lo_vnodeops))	/* Not our problem. */
		goto rename;

	/*
	 * XXXci - Once case-insensitive behavior is implemented, it should
	 * be added here.
	 */
	if (VOP_LOOKUP(ndvp, nnm, &tnvp, NULL, 0, NULL, cr,
	    ct, NULL, NULL) != 0)
		goto rename;
	if (tnvp->v_type != VDIR) {
		VN_RELE(tnvp);
		goto rename;
	}
	if (vn_mountedvfs(tnvp)) {
		VN_RELE(tnvp);
		return (EBUSY);
	}
	VN_RELE(tnvp);
rename:
	/*
	 * Since the case we're dealing with above can happen at any layer in
	 * the stack of lofs filesystems, we need to recurse down the stack,
	 * checking to see if there are any instances of a filesystem mounted on
	 * top of lofs. In order to keep on using the lofs version of
	 * VOP_RENAME(), we make sure that while the target directory is of type
	 * lofs, the source directory (the one used for getting the fs-specific
	 * version of VOP_RENAME()) is also of type lofs.
	 */
	if (vn_matchops(ndvp, lo_vnodeops)) {
		ndvp = realvp(ndvp);	/* Check the next layer */
	} else {
		/*
		 * We can go fast here
		 */
		while (vn_matchops(odvp, lo_vnodeops)) {
			odvp = realvp(odvp);
		}
		if (odvp->v_vfsp != ndvp->v_vfsp)
			return (EXDEV);
	}
	return (VOP_RENAME(odvp, onm, ndvp, nnm, cr, ct, flags));
}

static int
lo_mkdir(
	vnode_t *dvp,
	char *nm,
	struct vattr *va,
	vnode_t **vpp,
	struct cred *cr,
	caller_context_t *ct,
	int flags,
	vsecattr_t *vsecp)
{
	int error;

#ifdef LODEBUG
	lo_dprint(4, "lo_mkdir vp %p realvp %p\n", dvp, realvp(dvp));
#endif
	error = VOP_MKDIR(realvp(dvp), nm, va, vpp, cr, ct, flags, vsecp);
	if (!error)
		*vpp = makelonode(*vpp, vtoli(dvp->v_vfsp), 0);
	return (error);
}

static int
lo_realvp(vnode_t *vp, vnode_t **vpp, caller_context_t *ct)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_realvp %p\n", vp);
#endif
	while (vn_matchops(vp, lo_vnodeops))
		vp = realvp(vp);

	if (VOP_REALVP(vp, vpp, ct) != 0)
		*vpp = vp;
	return (0);
}

static int
lo_rmdir(
	vnode_t *dvp,
	char *nm,
	vnode_t *cdir,
	struct cred *cr,
	caller_context_t *ct,
	int flags)
{
	vnode_t *rvp = cdir;

#ifdef LODEBUG
	lo_dprint(4, "lo_rmdir vp %p realvp %p\n", dvp, realvp(dvp));
#endif
	/* if cdir is lofs vnode ptr get its real vnode ptr */
	if (vn_matchops(dvp, vn_getops(rvp)))
		(void) lo_realvp(cdir, &rvp, ct);
	dvp = realvp(dvp);
	return (VOP_RMDIR(dvp, nm, rvp, cr, ct, flags));
}

static int
lo_symlink(
	vnode_t *dvp,
	char *lnm,
	struct vattr *tva,
	char *tnm,
	struct cred *cr,
	caller_context_t *ct,
	int flags)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_symlink vp %p realvp %p\n", dvp, realvp(dvp));
#endif
	dvp = realvp(dvp);
	return (VOP_SYMLINK(dvp, lnm, tva, tnm, cr, ct, flags));
}

static int
lo_readlink(
	vnode_t *vp,
	struct uio *uiop,
	struct cred *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_READLINK(vp, uiop, cr, ct));
}

static int
lo_readdir(
	vnode_t *vp,
	struct uio *uiop,
	struct cred *cr,
	int *eofp,
	caller_context_t *ct,
	int flags)
{
#ifdef LODEBUG
	lo_dprint(4, "lo_readdir vp %p realvp %p\n", vp, realvp(vp));
#endif
	vp = realvp(vp);
	return (VOP_READDIR(vp, uiop, cr, eofp, ct, flags));
}

static int
lo_rwlock(vnode_t *vp, int write_lock, caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_RWLOCK(vp, write_lock, ct));
}

static void
lo_rwunlock(vnode_t *vp, int write_lock, caller_context_t *ct)
{
	vp = realvp(vp);
	VOP_RWUNLOCK(vp, write_lock, ct);
}

static int
lo_seek(vnode_t *vp, offset_t ooff, offset_t *noffp, caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_SEEK(vp, ooff, noffp, ct));
}

static int
lo_cmp(vnode_t *vp1, vnode_t *vp2, caller_context_t *ct)
{
	while (vn_matchops(vp1, lo_vnodeops))
		vp1 = realvp(vp1);
	while (vn_matchops(vp2, lo_vnodeops))
		vp2 = realvp(vp2);
	return (VOP_CMP(vp1, vp2, ct));
}

static int
lo_frlock(
	vnode_t *vp,
	int cmd,
	struct flock64 *bfp,
	int flag,
	offset_t offset,
	struct flk_callback *flk_cbp,
	cred_t *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_FRLOCK(vp, cmd, bfp, flag, offset, flk_cbp, cr, ct));
}

static int
lo_space(
	vnode_t *vp,
	int cmd,
	struct flock64 *bfp,
	int flag,
	offset_t offset,
	struct cred *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_SPACE(vp, cmd, bfp, flag, offset, cr, ct));
}

static int
lo_getpage(
	vnode_t *vp,
	offset_t off,
	size_t len,
	uint_t *prot,
	struct page *parr[],
	size_t psz,
	struct seg *seg,
	caddr_t addr,
	enum seg_rw rw,
	struct cred *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_GETPAGE(vp, off, len, prot, parr, psz, seg, addr, rw, cr,
	    ct));
}

static int
lo_putpage(
	vnode_t *vp,
	offset_t off,
	size_t len,
	int flags,
	struct cred *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_PUTPAGE(vp, off, len, flags, cr, ct));
}

static int
lo_map(
	vnode_t *vp,
	offset_t off,
	struct as *as,
	caddr_t *addrp,
	size_t len,
	uchar_t prot,
	uchar_t maxprot,
	uint_t flags,
	struct cred *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_MAP(vp, off, as, addrp, len, prot, maxprot, flags, cr, ct));
}

static int
lo_addmap(
	vnode_t *vp,
	offset_t off,
	struct as *as,
	caddr_t addr,
	size_t len,
	uchar_t prot,
	uchar_t maxprot,
	uint_t flags,
	struct cred *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_ADDMAP(vp, off, as, addr, len, prot, maxprot, flags, cr,
	    ct));
}

static int
lo_delmap(
	vnode_t *vp,
	offset_t off,
	struct as *as,
	caddr_t addr,
	size_t len,
	uint_t prot,
	uint_t maxprot,
	uint_t flags,
	struct cred *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_DELMAP(vp, off, as, addr, len, prot, maxprot, flags, cr,
	    ct));
}

static int
lo_poll(
	vnode_t *vp,
	short events,
	int anyyet,
	short *reventsp,
	struct pollhead **phpp,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_POLL(vp, events, anyyet, reventsp, phpp, ct));
}

static int
lo_dump(vnode_t *vp, caddr_t addr, offset_t bn, offset_t count,
    caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_DUMP(vp, addr, bn, count, ct));
}

static int
lo_pathconf(
	vnode_t *vp,
	int cmd,
	ulong_t *valp,
	struct cred *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_PATHCONF(vp, cmd, valp, cr, ct));
}

static int
lo_pageio(
	vnode_t *vp,
	struct page *pp,
	u_offset_t io_off,
	size_t io_len,
	int flags,
	cred_t *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_PAGEIO(vp, pp, io_off, io_len, flags, cr, ct));
}

static void
lo_dispose(
	vnode_t *vp,
	page_t *pp,
	int fl,
	int dn,
	cred_t *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	if (vp != NULL && !VN_ISKAS(vp))
		VOP_DISPOSE(vp, pp, fl, dn, cr, ct);
}

static int
lo_setsecattr(
	vnode_t *vp,
	vsecattr_t *secattr,
	int flags,
	struct cred *cr,
	caller_context_t *ct)
{
	if (vn_is_readonly(vp))
		return (EROFS);
	vp = realvp(vp);
	return (VOP_SETSECATTR(vp, secattr, flags, cr, ct));
}

static int
lo_getsecattr(
	vnode_t *vp,
	vsecattr_t *secattr,
	int flags,
	struct cred *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_GETSECATTR(vp, secattr, flags, cr, ct));
}

static int
lo_shrlock(
	vnode_t *vp,
	int cmd,
	struct shrlock *shr,
	int flag,
	cred_t *cr,
	caller_context_t *ct)
{
	vp = realvp(vp);
	return (VOP_SHRLOCK(vp, cmd, shr, flag, cr, ct));
}

/*
 * Loopback vnode operations vector.
 */

struct vnodeops *lo_vnodeops;

const fs_operation_def_t lo_vnodeops_template[] = {
	VOPNAME_OPEN,		{ .vop_open = lo_open },
	VOPNAME_CLOSE,		{ .vop_close = lo_close },
	VOPNAME_READ,		{ .vop_read = lo_read },
	VOPNAME_WRITE,		{ .vop_write = lo_write },
	VOPNAME_IOCTL,		{ .vop_ioctl = lo_ioctl },
	VOPNAME_SETFL,		{ .vop_setfl = lo_setfl },
	VOPNAME_GETATTR,	{ .vop_getattr = lo_getattr },
	VOPNAME_SETATTR,	{ .vop_setattr = lo_setattr },
	VOPNAME_ACCESS,		{ .vop_access = lo_access },
	VOPNAME_LOOKUP,		{ .vop_lookup = lo_lookup },
	VOPNAME_CREATE,		{ .vop_create = lo_create },
	VOPNAME_REMOVE,		{ .vop_remove = lo_remove },
	VOPNAME_LINK,		{ .vop_link = lo_link },
	VOPNAME_RENAME,		{ .vop_rename = lo_rename },
	VOPNAME_MKDIR,		{ .vop_mkdir = lo_mkdir },
	VOPNAME_RMDIR,		{ .vop_rmdir = lo_rmdir },
	VOPNAME_READDIR,	{ .vop_readdir = lo_readdir },
	VOPNAME_SYMLINK,	{ .vop_symlink = lo_symlink },
	VOPNAME_READLINK,	{ .vop_readlink = lo_readlink },
	VOPNAME_FSYNC,		{ .vop_fsync = lo_fsync },
	VOPNAME_INACTIVE,	{ .vop_inactive = lo_inactive },
	VOPNAME_FID,		{ .vop_fid = lo_fid },
	VOPNAME_RWLOCK,		{ .vop_rwlock = lo_rwlock },
	VOPNAME_RWUNLOCK,	{ .vop_rwunlock = lo_rwunlock },
	VOPNAME_SEEK,		{ .vop_seek = lo_seek },
	VOPNAME_CMP,		{ .vop_cmp = lo_cmp },
	VOPNAME_FRLOCK,		{ .vop_frlock = lo_frlock },
	VOPNAME_SPACE,		{ .vop_space = lo_space },
	VOPNAME_REALVP,		{ .vop_realvp = lo_realvp },
	VOPNAME_GETPAGE,	{ .vop_getpage = lo_getpage },
	VOPNAME_PUTPAGE,	{ .vop_putpage = lo_putpage },
	VOPNAME_MAP,		{ .vop_map = lo_map },
	VOPNAME_ADDMAP,		{ .vop_addmap = lo_addmap },
	VOPNAME_DELMAP,		{ .vop_delmap = lo_delmap },
	VOPNAME_POLL,		{ .vop_poll = lo_poll },
	VOPNAME_DUMP,		{ .vop_dump = lo_dump },
	VOPNAME_DUMPCTL,	{ .error = fs_error },	/* XXX - why? */
	VOPNAME_PATHCONF,	{ .vop_pathconf = lo_pathconf },
	VOPNAME_PAGEIO,		{ .vop_pageio = lo_pageio },
	VOPNAME_DISPOSE,	{ .vop_dispose = lo_dispose },
	VOPNAME_SETSECATTR,	{ .vop_setsecattr = lo_setsecattr },
	VOPNAME_GETSECATTR,	{ .vop_getsecattr = lo_getsecattr },
	VOPNAME_SHRLOCK,	{ .vop_shrlock = lo_shrlock },
	NULL,			NULL
};