HAMMER VFS - Fix deadlock which can occur under severe filesystem pressure

[dragonfly.git] / sys / vfs / hammer / hammer_inode.c

/*
 * Copyright (c) 2007-2008 The DragonFly Project.  All rights reserved.
 * 
 * This code is derived from software contributed to The DragonFly Project
 * by Matthew Dillon <dillon@backplane.com>
 * 
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in
 *    the documentation and/or other materials provided with the
 *    distribution.
 * 3. Neither the name of The DragonFly Project nor the names of its
 *    contributors may be used to endorse or promote products derived
 *    from this software without specific, prior written permission.
 * 
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 * 
 * $DragonFly: src/sys/vfs/hammer/hammer_inode.c,v 1.114 2008/09/24 00:53:51 dillon Exp $
 */

#include "hammer.h"
#include <vm/vm_extern.h>

static int	hammer_unload_inode(struct hammer_inode *ip);
static void	hammer_free_inode(hammer_inode_t ip);
static void	hammer_flush_inode_core(hammer_inode_t ip,
					hammer_flush_group_t flg, int flags);
static int	hammer_setup_child_callback(hammer_record_t rec, void *data);
#if 0
static int	hammer_syncgrp_child_callback(hammer_record_t rec, void *data);
#endif
static int	hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
					hammer_flush_group_t flg);
static int	hammer_setup_parent_inodes_helper(hammer_record_t record,
					int depth, hammer_flush_group_t flg);
static void	hammer_inode_wakereclaims(hammer_inode_t ip);
static struct hammer_inostats *hammer_inode_inostats(hammer_mount_t hmp,
					pid_t pid);

#ifdef DEBUG_TRUNCATE
extern struct hammer_inode *HammerTruncIp;
#endif

/*
 * RB-Tree support for inode structures
 */
int
hammer_ino_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
{
	if (ip1->obj_localization < ip2->obj_localization)
		return(-1);
	if (ip1->obj_localization > ip2->obj_localization)
		return(1);
	if (ip1->obj_id < ip2->obj_id)
		return(-1);
	if (ip1->obj_id > ip2->obj_id)
		return(1);
	if (ip1->obj_asof < ip2->obj_asof)
		return(-1);
	if (ip1->obj_asof > ip2->obj_asof)
		return(1);
	return(0);
}

int
hammer_redo_rb_compare(hammer_inode_t ip1, hammer_inode_t ip2)
{
	if (ip1->redo_fifo_start < ip2->redo_fifo_start)
		return(-1);
	if (ip1->redo_fifo_start > ip2->redo_fifo_start)
		return(1);
	return(0);
}

/*
 * RB-Tree support for inode structures / special LOOKUP_INFO
 */
static int
hammer_inode_info_cmp(hammer_inode_info_t info, hammer_inode_t ip)
{
	if (info->obj_localization < ip->obj_localization)
		return(-1);
	if (info->obj_localization > ip->obj_localization)
		return(1);
	if (info->obj_id < ip->obj_id)
		return(-1);
	if (info->obj_id > ip->obj_id)
		return(1);
	if (info->obj_asof < ip->obj_asof)
		return(-1);
	if (info->obj_asof > ip->obj_asof)
		return(1);
	return(0);
}

/*
 * Used by hammer_scan_inode_snapshots() to locate all of an object's
 * snapshots.  Note that the asof field is not tested, which we can get
 * away with because it is the lowest-priority field.
 */
static int
hammer_inode_info_cmp_all_history(hammer_inode_t ip, void *data)
{
	hammer_inode_info_t info = data;

	if (ip->obj_localization > info->obj_localization)
		return(1);
	if (ip->obj_localization < info->obj_localization)
		return(-1);
	if (ip->obj_id > info->obj_id)
		return(1);
	if (ip->obj_id < info->obj_id)
		return(-1);
	return(0);
}

/*
 * Used by hammer_unload_pseudofs() to locate all inodes associated with
 * a particular PFS.
 */
static int
hammer_inode_pfs_cmp(hammer_inode_t ip, void *data)
{
	u_int32_t localization = *(u_int32_t *)data;
	if (ip->obj_localization > localization)
		return(1);
	if (ip->obj_localization < localization)
		return(-1);
	return(0);
}

/*
 * RB-Tree support for pseudofs structures
 */
static int
hammer_pfs_rb_compare(hammer_pseudofs_inmem_t p1, hammer_pseudofs_inmem_t p2)
{
	if (p1->localization < p2->localization)
		return(-1);
	if (p1->localization > p2->localization)
		return(1);
	return(0);
}


RB_GENERATE(hammer_ino_rb_tree, hammer_inode, rb_node, hammer_ino_rb_compare);
RB_GENERATE_XLOOKUP(hammer_ino_rb_tree, INFO, hammer_inode, rb_node,
		hammer_inode_info_cmp, hammer_inode_info_t);
RB_GENERATE2(hammer_pfs_rb_tree, hammer_pseudofs_inmem, rb_node,
             hammer_pfs_rb_compare, u_int32_t, localization);

/*
 * The kernel is not actively referencing this vnode but is still holding
 * it cached.
 *
 * This is called from the frontend.
 *
 * MPALMOSTSAFE
 */
int
hammer_vop_inactive(struct vop_inactive_args *ap)
{
	struct hammer_inode *ip = VTOI(ap->a_vp);
	hammer_mount_t hmp;

	/*
	 * Degenerate case
	 */
	if (ip == NULL) {
		vrecycle(ap->a_vp);
		return(0);
	}

	/*
	 * If the inode no longer has visibility in the filesystem try to
	 * recycle it immediately, even if the inode is dirty.  Recycling
	 * it quickly allows the system to reclaim buffer cache and VM
	 * resources which can matter a lot in a heavily loaded system.
	 *
	 * This can deadlock in vfsync() if we aren't careful.
	 * 
	 * Do not queue the inode to the flusher if we still have visibility,
	 * otherwise namespace calls such as chmod will unnecessarily generate
	 * multiple inode updates.
	 */
	if (ip->ino_data.nlinks == 0) {
		hmp = ip->hmp;
		lwkt_gettoken(&hmp->fs_token);
		hammer_inode_unloadable_check(ip, 0);
		if (ip->flags & HAMMER_INODE_MODMASK)
			hammer_flush_inode(ip, 0);
		lwkt_reltoken(&hmp->fs_token);
		vrecycle(ap->a_vp);
	}
	return(0);
}

/*
 * Release the vnode association.  This is typically (but not always)
 * the last reference on the inode.
 *
 * Once the association is lost we are on our own with regards to
 * flushing the inode.
 *
 * We must interlock ip->vp so hammer_get_vnode() can avoid races.
 */
int
hammer_vop_reclaim(struct vop_reclaim_args *ap)
{
	struct hammer_inode *ip;
	hammer_mount_t hmp;
	struct vnode *vp;

	vp = ap->a_vp;

	if ((ip = vp->v_data) != NULL) {
		hmp = ip->hmp;
		lwkt_gettoken(&hmp->fs_token);
		hammer_lock_ex(&ip->lock);
		vp->v_data = NULL;
		ip->vp = NULL;

		if ((ip->flags & HAMMER_INODE_RECLAIM) == 0) {
			++hammer_count_reclaiming;
			++hmp->inode_reclaims;
			ip->flags |= HAMMER_INODE_RECLAIM;
		}
		hammer_unlock(&ip->lock);
		hammer_rel_inode(ip, 1);
		lwkt_reltoken(&hmp->fs_token);
	}
	return(0);
}

/*
 * Return a locked vnode for the specified inode.  The inode must be
 * referenced but NOT LOCKED on entry and will remain referenced on
 * return.
 *
 * Called from the frontend.
 */
int
hammer_get_vnode(struct hammer_inode *ip, struct vnode **vpp)
{
	hammer_mount_t hmp;
	struct vnode *vp;
	int error = 0;
	u_int8_t obj_type;

	hmp = ip->hmp;

	for (;;) {
		if ((vp = ip->vp) == NULL) {
			error = getnewvnode(VT_HAMMER, hmp->mp, vpp, 0, 0);
			if (error)
				break;
			hammer_lock_ex(&ip->lock);
			if (ip->vp != NULL) {
				hammer_unlock(&ip->lock);
				vp = *vpp;
				vp->v_type = VBAD;
				vx_put(vp);
				continue;
			}
			hammer_ref(&ip->lock);
			vp = *vpp;
			ip->vp = vp;

			obj_type = ip->ino_data.obj_type;
			vp->v_type = hammer_get_vnode_type(obj_type);

			hammer_inode_wakereclaims(ip);

			switch(ip->ino_data.obj_type) {
			case HAMMER_OBJTYPE_CDEV:
			case HAMMER_OBJTYPE_BDEV:
				vp->v_ops = &hmp->mp->mnt_vn_spec_ops;
				addaliasu(vp, ip->ino_data.rmajor,
					  ip->ino_data.rminor);
				break;
			case HAMMER_OBJTYPE_FIFO:
				vp->v_ops = &hmp->mp->mnt_vn_fifo_ops;
				break;
			case HAMMER_OBJTYPE_REGFILE:
				break;
			default:
				break;
			}

			/*
			 * Only mark as the root vnode if the ip is not
			 * historical, otherwise the VFS cache will get
			 * confused.  The other half of the special handling
			 * is in hammer_vop_nlookupdotdot().
			 *
			 * Pseudo-filesystem roots can be accessed via
			 * non-root filesystem paths and setting VROOT may
			 * confuse the namecache.  Set VPFSROOT instead.
			 */
			if (ip->obj_id == HAMMER_OBJID_ROOT &&
			    ip->obj_asof == hmp->asof) {
				if (ip->obj_localization == 0)
					vsetflags(vp, VROOT);
				else
					vsetflags(vp, VPFSROOT);
			}

			vp->v_data = (void *)ip;
			/* vnode locked by getnewvnode() */
			/* make related vnode dirty if inode dirty? */
			hammer_unlock(&ip->lock);
			if (vp->v_type == VREG) {
				vinitvmio(vp, ip->ino_data.size,
					  hammer_blocksize(ip->ino_data.size),
					  hammer_blockoff(ip->ino_data.size));
			}
			break;
		}

		/*
		 * Interlock vnode clearing.  This does not prevent the
		 * vnode from going into a reclaimed state but it does
		 * prevent it from being destroyed or reused so the vget()
		 * will properly fail.
		 */
		hammer_lock_ex(&ip->lock);
		if ((vp = ip->vp) == NULL) {
			hammer_unlock(&ip->lock);
			continue;
		}
		vhold_interlocked(vp);
		hammer_unlock(&ip->lock);

		/*
		 * loop if the vget fails (aka races), or if the vp
		 * no longer matches ip->vp.
		 */
		if (vget(vp, LK_EXCLUSIVE) == 0) {
			if (vp == ip->vp) {
				vdrop(vp);
				break;
			}
			vput(vp);
		}
		vdrop(vp);
	}
	*vpp = vp;
	return(error);
}

/*
 * Locate all copies of the inode for obj_id compatible with the specified
 * asof, reference, and issue the related call-back.  This routine is used
 * for direct-io invalidation and does not create any new inodes.
 */
void
hammer_scan_inode_snapshots(hammer_mount_t hmp, hammer_inode_info_t iinfo,
		            int (*callback)(hammer_inode_t ip, void *data),
			    void *data)
{
	hammer_ino_rb_tree_RB_SCAN(&hmp->rb_inos_root,
				   hammer_inode_info_cmp_all_history,
				   callback, iinfo);
}

/*
 * Acquire a HAMMER inode.  The returned inode is not locked.  These functions
 * do not attach or detach the related vnode (use hammer_get_vnode() for
 * that).
 *
 * The flags argument is only applied for newly created inodes, and only
 * certain flags are inherited.
 *
 * Called from the frontend.
 */
struct hammer_inode *
hammer_get_inode(hammer_transaction_t trans, hammer_inode_t dip,
		 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
		 int flags, int *errorp)
{
	hammer_mount_t hmp = trans->hmp;
	struct hammer_node_cache *cachep;
	struct hammer_inode_info iinfo;
	struct hammer_cursor cursor;
	struct hammer_inode *ip;


	/*
	 * Determine if we already have an inode cached.  If we do then
	 * we are golden.
	 *
	 * If we find an inode with no vnode we have to mark the
	 * transaction such that hammer_inode_waitreclaims() is
	 * called later on to avoid building up an infinite number
	 * of inodes.  Otherwise we can continue to * add new inodes
	 * faster then they can be disposed of, even with the tsleep
	 * delay.
	 *
	 * If we find a dummy inode we return a failure so dounlink
	 * (which does another lookup) doesn't try to mess with the
	 * link count.  hammer_vop_nresolve() uses hammer_get_dummy_inode()
	 * to ref dummy inodes.
	 */
	iinfo.obj_id = obj_id;
	iinfo.obj_asof = asof;
	iinfo.obj_localization = localization;
loop:
	ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
	if (ip) {
		if (ip->flags & HAMMER_INODE_DUMMY) {
			*errorp = ENOENT;
			return(NULL);
		}
		hammer_ref(&ip->lock);
		*errorp = 0;
		return(ip);
	}

	/*
	 * Allocate a new inode structure and deal with races later.
	 */
	ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
	++hammer_count_inodes;
	++hmp->count_inodes;
	ip->obj_id = obj_id;
	ip->obj_asof = iinfo.obj_asof;
	ip->obj_localization = localization;
	ip->hmp = hmp;
	ip->flags = flags & HAMMER_INODE_RO;
	ip->cache[0].ip = ip;
	ip->cache[1].ip = ip;
	ip->cache[2].ip = ip;
	ip->cache[3].ip = ip;
	if (hmp->ronly)
		ip->flags |= HAMMER_INODE_RO;
	ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
		0x7FFFFFFFFFFFFFFFLL;
	RB_INIT(&ip->rec_tree);
	TAILQ_INIT(&ip->target_list);
	hammer_ref(&ip->lock);

	/*
	 * Locate the on-disk inode.  If this is a PFS root we always
	 * access the current version of the root inode and (if it is not
	 * a master) always access information under it with a snapshot
	 * TID.
	 *
	 * We cache recent inode lookups in this directory in dip->cache[2].
	 * If we can't find it we assume the inode we are looking for is
	 * close to the directory inode.
	 */
retry:
	cachep = NULL;
	if (dip) {
		if (dip->cache[2].node)
			cachep = &dip->cache[2];
		else
			cachep = &dip->cache[0];
	}
	hammer_init_cursor(trans, &cursor, cachep, NULL);
	cursor.key_beg.localization = localization + HAMMER_LOCALIZE_INODE;
	cursor.key_beg.obj_id = ip->obj_id;
	cursor.key_beg.key = 0;
	cursor.key_beg.create_tid = 0;
	cursor.key_beg.delete_tid = 0;
	cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
	cursor.key_beg.obj_type = 0;

	cursor.asof = iinfo.obj_asof;
	cursor.flags = HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_GET_DATA |
		       HAMMER_CURSOR_ASOF;

	*errorp = hammer_btree_lookup(&cursor);
	if (*errorp == EDEADLK) {
		hammer_done_cursor(&cursor);
		goto retry;
	}

	/*
	 * On success the B-Tree lookup will hold the appropriate
	 * buffer cache buffers and provide a pointer to the requested
	 * information.  Copy the information to the in-memory inode
	 * and cache the B-Tree node to improve future operations.
	 */
	if (*errorp == 0) {
		ip->ino_leaf = cursor.node->ondisk->elms[cursor.index].leaf;
		ip->ino_data = cursor.data->inode;

		/*
		 * cache[0] tries to cache the location of the object inode.
		 * The assumption is that it is near the directory inode.
		 *
		 * cache[1] tries to cache the location of the object data.
		 * We might have something in the governing directory from
		 * scan optimizations (see the strategy code in
		 * hammer_vnops.c).
		 *
		 * We update dip->cache[2], if possible, with the location
		 * of the object inode for future directory shortcuts.
		 */
		hammer_cache_node(&ip->cache[0], cursor.node);
		if (dip) {
			if (dip->cache[3].node) {
				hammer_cache_node(&ip->cache[1],
						  dip->cache[3].node);
			}
			hammer_cache_node(&dip->cache[2], cursor.node);
		}

		/*
		 * The file should not contain any data past the file size
		 * stored in the inode.  Setting save_trunc_off to the
		 * file size instead of max reduces B-Tree lookup overheads
		 * on append by allowing the flusher to avoid checking for
		 * record overwrites.
		 */
		ip->save_trunc_off = ip->ino_data.size;

		/*
		 * Locate and assign the pseudofs management structure to
		 * the inode.
		 */
		if (dip && dip->obj_localization == ip->obj_localization) {
			ip->pfsm = dip->pfsm;
			hammer_ref(&ip->pfsm->lock);
		} else {
			ip->pfsm = hammer_load_pseudofs(trans,
							ip->obj_localization,
							errorp);
			*errorp = 0;	/* ignore ENOENT */
		}
	}

	/*
	 * The inode is placed on the red-black tree and will be synced to
	 * the media when flushed or by the filesystem sync.  If this races
	 * another instantiation/lookup the insertion will fail.
	 */
	if (*errorp == 0) {
		if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
			hammer_free_inode(ip);
			hammer_done_cursor(&cursor);
			goto loop;
		}
		ip->flags |= HAMMER_INODE_ONDISK;
	} else {
		if (ip->flags & HAMMER_INODE_RSV_INODES) {
			ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
			--hmp->rsv_inodes;
		}

		hammer_free_inode(ip);
		ip = NULL;
	}
	hammer_done_cursor(&cursor);

	/*
	 * NEWINODE is only set if the inode becomes dirty later,
	 * setting it here just leads to unnecessary stalls.
	 *
	 * trans->flags |= HAMMER_TRANSF_NEWINODE;
	 */
	return (ip);
}

/*
 * Get a dummy inode to placemark a broken directory entry.
 */
struct hammer_inode *
hammer_get_dummy_inode(hammer_transaction_t trans, hammer_inode_t dip,
		 int64_t obj_id, hammer_tid_t asof, u_int32_t localization,
		 int flags, int *errorp)
{
	hammer_mount_t hmp = trans->hmp;
	struct hammer_inode_info iinfo;
	struct hammer_inode *ip;

	/*
	 * Determine if we already have an inode cached.  If we do then
	 * we are golden.
	 *
	 * If we find an inode with no vnode we have to mark the
	 * transaction such that hammer_inode_waitreclaims() is
	 * called later on to avoid building up an infinite number
	 * of inodes.  Otherwise we can continue to * add new inodes
	 * faster then they can be disposed of, even with the tsleep
	 * delay.
	 *
	 * If we find a non-fake inode we return an error.  Only fake
	 * inodes can be returned by this routine.
	 */
	iinfo.obj_id = obj_id;
	iinfo.obj_asof = asof;
	iinfo.obj_localization = localization;
loop:
	*errorp = 0;
	ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
	if (ip) {
		if ((ip->flags & HAMMER_INODE_DUMMY) == 0) {
			*errorp = ENOENT;
			return(NULL);
		}
		hammer_ref(&ip->lock);
		return(ip);
	}

	/*
	 * Allocate a new inode structure and deal with races later.
	 */
	ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
	++hammer_count_inodes;
	++hmp->count_inodes;
	ip->obj_id = obj_id;
	ip->obj_asof = iinfo.obj_asof;
	ip->obj_localization = localization;
	ip->hmp = hmp;
	ip->flags = flags | HAMMER_INODE_RO | HAMMER_INODE_DUMMY;
	ip->cache[0].ip = ip;
	ip->cache[1].ip = ip;
	ip->cache[2].ip = ip;
	ip->cache[3].ip = ip;
	ip->sync_trunc_off = ip->trunc_off = ip->save_trunc_off =
		0x7FFFFFFFFFFFFFFFLL;
	RB_INIT(&ip->rec_tree);
	TAILQ_INIT(&ip->target_list);
	hammer_ref(&ip->lock);

	/*
	 * Populate the dummy inode.  Leave everything zero'd out.
	 *
	 * (ip->ino_leaf and ip->ino_data)
	 *
	 * Make the dummy inode a FIFO object which most copy programs
	 * will properly ignore.
	 */
	ip->save_trunc_off = ip->ino_data.size;
	ip->ino_data.obj_type = HAMMER_OBJTYPE_FIFO;

	/*
	 * Locate and assign the pseudofs management structure to
	 * the inode.
	 */
	if (dip && dip->obj_localization == ip->obj_localization) {
		ip->pfsm = dip->pfsm;
		hammer_ref(&ip->pfsm->lock);
	} else {
		ip->pfsm = hammer_load_pseudofs(trans, ip->obj_localization,
						errorp);
		*errorp = 0;	/* ignore ENOENT */
	}

	/*
	 * The inode is placed on the red-black tree and will be synced to
	 * the media when flushed or by the filesystem sync.  If this races
	 * another instantiation/lookup the insertion will fail.
	 *
	 * NOTE: Do not set HAMMER_INODE_ONDISK.  The inode is a fake.
	 */
	if (*errorp == 0) {
		if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
			hammer_free_inode(ip);
			goto loop;
		}
	} else {
		if (ip->flags & HAMMER_INODE_RSV_INODES) {
			ip->flags &= ~HAMMER_INODE_RSV_INODES; /* sanity */
			--hmp->rsv_inodes;
		}
		hammer_free_inode(ip);
		ip = NULL;
	}
	trans->flags |= HAMMER_TRANSF_NEWINODE;
	return (ip);
}

/*
 * Return a referenced inode only if it is in our inode cache.
 *
 * Dummy inodes do not count.
 */
struct hammer_inode *
hammer_find_inode(hammer_transaction_t trans, int64_t obj_id,
		  hammer_tid_t asof, u_int32_t localization)
{
	hammer_mount_t hmp = trans->hmp;
	struct hammer_inode_info iinfo;
	struct hammer_inode *ip;

	iinfo.obj_id = obj_id;
	iinfo.obj_asof = asof;
	iinfo.obj_localization = localization;

	ip = hammer_ino_rb_tree_RB_LOOKUP_INFO(&hmp->rb_inos_root, &iinfo);
	if (ip) {
		if (ip->flags & HAMMER_INODE_DUMMY)
			ip = NULL;
		else
			hammer_ref(&ip->lock);
	}
	return(ip);
}

/*
 * Create a new filesystem object, returning the inode in *ipp.  The
 * returned inode will be referenced.  The inode is created in-memory.
 *
 * If pfsm is non-NULL the caller wishes to create the root inode for
 * a master PFS.
 */
int
hammer_create_inode(hammer_transaction_t trans, struct vattr *vap,
		    struct ucred *cred,
		    hammer_inode_t dip, const char *name, int namelen,
		    hammer_pseudofs_inmem_t pfsm, struct hammer_inode **ipp)
{
	hammer_mount_t hmp;
	hammer_inode_t ip;
	uid_t xuid;
	int error;
	int64_t namekey;
	u_int32_t dummy;

	hmp = trans->hmp;

	ip = kmalloc(sizeof(*ip), hmp->m_inodes, M_WAITOK|M_ZERO);
	++hammer_count_inodes;
	++hmp->count_inodes;
	trans->flags |= HAMMER_TRANSF_NEWINODE;

	if (pfsm) {
		KKASSERT(pfsm->localization != 0);
		ip->obj_id = HAMMER_OBJID_ROOT;
		ip->obj_localization = pfsm->localization;
	} else {
		KKASSERT(dip != NULL);
		namekey = hammer_directory_namekey(dip, name, namelen, &dummy);
		ip->obj_id = hammer_alloc_objid(hmp, dip, namekey);
		ip->obj_localization = dip->obj_localization;
	}

	KKASSERT(ip->obj_id != 0);
	ip->obj_asof = hmp->asof;
	ip->hmp = hmp;
	ip->flush_state = HAMMER_FST_IDLE;
	ip->flags = HAMMER_INODE_DDIRTY |
		    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME;
	ip->cache[0].ip = ip;
	ip->cache[1].ip = ip;
	ip->cache[2].ip = ip;
	ip->cache[3].ip = ip;

	ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
	/* ip->save_trunc_off = 0; (already zero) */
	RB_INIT(&ip->rec_tree);
	TAILQ_INIT(&ip->target_list);

	ip->ino_data.atime = trans->time;
	ip->ino_data.mtime = trans->time;
	ip->ino_data.size = 0;
	ip->ino_data.nlinks = 0;

	/*
	 * A nohistory designator on the parent directory is inherited by
	 * the child.  We will do this even for pseudo-fs creation... the
	 * sysad can turn it off.
	 */
	if (dip) {
		ip->ino_data.uflags = dip->ino_data.uflags &
				      (SF_NOHISTORY|UF_NOHISTORY|UF_NODUMP);
	}

	ip->ino_leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
	ip->ino_leaf.base.localization = ip->obj_localization +
					 HAMMER_LOCALIZE_INODE;
	ip->ino_leaf.base.obj_id = ip->obj_id;
	ip->ino_leaf.base.key = 0;
	ip->ino_leaf.base.create_tid = 0;
	ip->ino_leaf.base.delete_tid = 0;
	ip->ino_leaf.base.rec_type = HAMMER_RECTYPE_INODE;
	ip->ino_leaf.base.obj_type = hammer_get_obj_type(vap->va_type);

	ip->ino_data.obj_type = ip->ino_leaf.base.obj_type;
	ip->ino_data.version = HAMMER_INODE_DATA_VERSION;
	ip->ino_data.mode = vap->va_mode;
	ip->ino_data.ctime = trans->time;

	/*
	 * If we are running version 2 or greater directory entries are
	 * inode-localized instead of data-localized.
	 */
	if (trans->hmp->version >= HAMMER_VOL_VERSION_TWO) {
		if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
			ip->ino_data.cap_flags |=
				HAMMER_INODE_CAP_DIR_LOCAL_INO;
		}
	}
	if (trans->hmp->version >= HAMMER_VOL_VERSION_SIX) {
		if (ip->ino_leaf.base.obj_type == HAMMER_OBJTYPE_DIRECTORY) {
			ip->ino_data.cap_flags |=
				HAMMER_INODE_CAP_DIRHASH_ALG1;
		}
	}

	/*
	 * Setup the ".." pointer.  This only needs to be done for directories
	 * but we do it for all objects as a recovery aid.
	 */
	if (dip)
		ip->ino_data.parent_obj_id = dip->ino_leaf.base.obj_id;
#if 0
	/*
	 * The parent_obj_localization field only applies to pseudo-fs roots.
	 * XXX this is no longer applicable, PFSs are no longer directly
	 * tied into the parent's directory structure.
	 */
	if (ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY &&
	    ip->obj_id == HAMMER_OBJID_ROOT) {
		ip->ino_data.ext.obj.parent_obj_localization = 
						dip->obj_localization;
	}
#endif

	switch(ip->ino_leaf.base.obj_type) {
	case HAMMER_OBJTYPE_CDEV:
	case HAMMER_OBJTYPE_BDEV:
		ip->ino_data.rmajor = vap->va_rmajor;
		ip->ino_data.rminor = vap->va_rminor;
		break;
	default:
		break;
	}

	/*
	 * Calculate default uid/gid and overwrite with information from
	 * the vap.
	 */
	if (dip) {
		xuid = hammer_to_unix_xid(&dip->ino_data.uid);
		xuid = vop_helper_create_uid(hmp->mp, dip->ino_data.mode,
					     xuid, cred, &vap->va_mode);
	} else {
		xuid = 0;
	}
	ip->ino_data.mode = vap->va_mode;

	if (vap->va_vaflags & VA_UID_UUID_VALID)
		ip->ino_data.uid = vap->va_uid_uuid;
	else if (vap->va_uid != (uid_t)VNOVAL)
		hammer_guid_to_uuid(&ip->ino_data.uid, vap->va_uid);
	else
		hammer_guid_to_uuid(&ip->ino_data.uid, xuid);

	if (vap->va_vaflags & VA_GID_UUID_VALID)
		ip->ino_data.gid = vap->va_gid_uuid;
	else if (vap->va_gid != (gid_t)VNOVAL)
		hammer_guid_to_uuid(&ip->ino_data.gid, vap->va_gid);
	else if (dip)
		ip->ino_data.gid = dip->ino_data.gid;

	hammer_ref(&ip->lock);

	if (pfsm) {
		ip->pfsm = pfsm;
		hammer_ref(&pfsm->lock);
		error = 0;
	} else if (dip->obj_localization == ip->obj_localization) {
		ip->pfsm = dip->pfsm;
		hammer_ref(&ip->pfsm->lock);
		error = 0;
	} else {
		ip->pfsm = hammer_load_pseudofs(trans,
						ip->obj_localization,
						&error);
		error = 0;	/* ignore ENOENT */
	}

	if (error) {
		hammer_free_inode(ip);
		ip = NULL;
	} else if (RB_INSERT(hammer_ino_rb_tree, &hmp->rb_inos_root, ip)) {
		panic("hammer_create_inode: duplicate obj_id %llx",
		      (long long)ip->obj_id);
		/* not reached */
		hammer_free_inode(ip);
	}
	*ipp = ip;
	return(error);
}

/*
 * Final cleanup / freeing of an inode structure
 */
static void
hammer_free_inode(hammer_inode_t ip)
{
	struct hammer_mount *hmp;

	hmp = ip->hmp;
	KKASSERT(hammer_oneref(&ip->lock));
	hammer_uncache_node(&ip->cache[0]);
	hammer_uncache_node(&ip->cache[1]);
	hammer_uncache_node(&ip->cache[2]);
	hammer_uncache_node(&ip->cache[3]);
	hammer_inode_wakereclaims(ip);
	if (ip->objid_cache)
		hammer_clear_objid(ip);
	--hammer_count_inodes;
	--hmp->count_inodes;
	if (ip->pfsm) {
		hammer_rel_pseudofs(hmp, ip->pfsm);
		ip->pfsm = NULL;
	}
	kfree(ip, hmp->m_inodes);
	ip = NULL;
}

/*
 * Retrieve pseudo-fs data.  NULL will never be returned.
 *
 * If an error occurs *errorp will be set and a default template is returned,
 * otherwise *errorp is set to 0.  Typically when an error occurs it will
 * be ENOENT.
 */
hammer_pseudofs_inmem_t
hammer_load_pseudofs(hammer_transaction_t trans,
		     u_int32_t localization, int *errorp)
{
	hammer_mount_t hmp = trans->hmp;
	hammer_inode_t ip;
	hammer_pseudofs_inmem_t pfsm;
	struct hammer_cursor cursor;
	int bytes;

retry:
	pfsm = RB_LOOKUP(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, localization);
	if (pfsm) {
		hammer_ref(&pfsm->lock);
		*errorp = 0;
		return(pfsm);
	}

	/*
	 * PFS records are stored in the root inode (not the PFS root inode,
	 * but the real root).  Avoid an infinite recursion if loading
	 * the PFS for the real root.
	 */
	if (localization) {
		ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT,
				      HAMMER_MAX_TID,
				      HAMMER_DEF_LOCALIZATION, 0, errorp);
	} else {
		ip = NULL;
	}

	pfsm = kmalloc(sizeof(*pfsm), hmp->m_misc, M_WAITOK | M_ZERO);
	pfsm->localization = localization;
	pfsm->pfsd.unique_uuid = trans->rootvol->ondisk->vol_fsid;
	pfsm->pfsd.shared_uuid = pfsm->pfsd.unique_uuid;

	hammer_init_cursor(trans, &cursor, (ip ? &ip->cache[1] : NULL), ip);
	cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION +
				      HAMMER_LOCALIZE_MISC;
	cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
	cursor.key_beg.create_tid = 0;
	cursor.key_beg.delete_tid = 0;
	cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
	cursor.key_beg.obj_type = 0;
	cursor.key_beg.key = localization;
	cursor.asof = HAMMER_MAX_TID;
	cursor.flags |= HAMMER_CURSOR_ASOF;

	if (ip)
		*errorp = hammer_ip_lookup(&cursor);
	else
		*errorp = hammer_btree_lookup(&cursor);
	if (*errorp == 0) {
		*errorp = hammer_ip_resolve_data(&cursor);
		if (*errorp == 0) {
			if (cursor.data->pfsd.mirror_flags &
			    HAMMER_PFSD_DELETED) {
				*errorp = ENOENT;
			} else {
				bytes = cursor.leaf->data_len;
				if (bytes > sizeof(pfsm->pfsd))
					bytes = sizeof(pfsm->pfsd);
				bcopy(cursor.data, &pfsm->pfsd, bytes);
			}
		}
	}
	hammer_done_cursor(&cursor);

	pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
	hammer_ref(&pfsm->lock);
	if (ip)
		hammer_rel_inode(ip, 0);
	if (RB_INSERT(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm)) {
		kfree(pfsm, hmp->m_misc);
		goto retry;
	}
	return(pfsm);
}

/*
 * Store pseudo-fs data.  The backend will automatically delete any prior
 * on-disk pseudo-fs data but we have to delete in-memory versions.
 */
int
hammer_save_pseudofs(hammer_transaction_t trans, hammer_pseudofs_inmem_t pfsm)
{
	struct hammer_cursor cursor;
	hammer_record_t record;
	hammer_inode_t ip;
	int error;

	ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
			      HAMMER_DEF_LOCALIZATION, 0, &error);
retry:
	pfsm->fsid_udev = hammer_fsid_to_udev(&pfsm->pfsd.shared_uuid);
	hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
	cursor.key_beg.localization = ip->obj_localization +
				      HAMMER_LOCALIZE_MISC;
	cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
	cursor.key_beg.create_tid = 0;
	cursor.key_beg.delete_tid = 0;
	cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
	cursor.key_beg.obj_type = 0;
	cursor.key_beg.key = pfsm->localization;
	cursor.asof = HAMMER_MAX_TID;
	cursor.flags |= HAMMER_CURSOR_ASOF;

	/*
	 * Replace any in-memory version of the record.
	 */
	error = hammer_ip_lookup(&cursor);
	if (error == 0 && hammer_cursor_inmem(&cursor)) {
		record = cursor.iprec;
		if (record->flags & HAMMER_RECF_INTERLOCK_BE) {
			KKASSERT(cursor.deadlk_rec == NULL);
			hammer_ref(&record->lock);
			cursor.deadlk_rec = record;
			error = EDEADLK;
		} else {
			record->flags |= HAMMER_RECF_DELETED_FE;
			error = 0;
		}
	}

	/*
	 * Allocate replacement general record.  The backend flush will
	 * delete any on-disk version of the record.
	 */
	if (error == 0 || error == ENOENT) {
		record = hammer_alloc_mem_record(ip, sizeof(pfsm->pfsd));
		record->type = HAMMER_MEM_RECORD_GENERAL;

		record->leaf.base.localization = ip->obj_localization +
						 HAMMER_LOCALIZE_MISC;
		record->leaf.base.rec_type = HAMMER_RECTYPE_PFS;
		record->leaf.base.key = pfsm->localization;
		record->leaf.data_len = sizeof(pfsm->pfsd);
		bcopy(&pfsm->pfsd, record->data, sizeof(pfsm->pfsd));
		error = hammer_ip_add_record(trans, record);
	}
	hammer_done_cursor(&cursor);
	if (error == EDEADLK)
		goto retry;
	hammer_rel_inode(ip, 0);
	return(error);
}

/*
 * Create a root directory for a PFS if one does not alredy exist.
 *
 * The PFS root stands alone so we must also bump the nlinks count
 * to prevent it from being destroyed on release.
 */
int
hammer_mkroot_pseudofs(hammer_transaction_t trans, struct ucred *cred,
		       hammer_pseudofs_inmem_t pfsm)
{
	hammer_inode_t ip;
	struct vattr vap;
	int error;

	ip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
			      pfsm->localization, 0, &error);
	if (ip == NULL) {
		vattr_null(&vap);
		vap.va_mode = 0755;
		vap.va_type = VDIR;
		error = hammer_create_inode(trans, &vap, cred,
					    NULL, NULL, 0,
					    pfsm, &ip);
		if (error == 0) {
			++ip->ino_data.nlinks;
			hammer_modify_inode(trans, ip, HAMMER_INODE_DDIRTY);
		}
	}
	if (ip)
		hammer_rel_inode(ip, 0);
	return(error);
}

/*
 * Unload any vnodes & inodes associated with a PFS, return ENOTEMPTY
 * if we are unable to disassociate all the inodes.
 */
static
int
hammer_unload_pseudofs_callback(hammer_inode_t ip, void *data)
{
	int res;

	hammer_ref(&ip->lock);
	if (hammer_isactive(&ip->lock) == 2 && ip->vp)
		vclean_unlocked(ip->vp);
	if (hammer_isactive(&ip->lock) == 1 && ip->vp == NULL)
		res = 0;
	else
		res = -1;	/* stop, someone is using the inode */
	hammer_rel_inode(ip, 0);
	return(res);
}

int
hammer_unload_pseudofs(hammer_transaction_t trans, u_int32_t localization)
{
	int res;
	int try;

	for (try = res = 0; try < 4; ++try) {
		res = hammer_ino_rb_tree_RB_SCAN(&trans->hmp->rb_inos_root,
					   hammer_inode_pfs_cmp,
					   hammer_unload_pseudofs_callback,
					   &localization);
		if (res == 0 && try > 1)
			break;
		hammer_flusher_sync(trans->hmp);
	}
	if (res != 0)
		res = ENOTEMPTY;
	return(res);
}


/*
 * Release a reference on a PFS
 */
void
hammer_rel_pseudofs(hammer_mount_t hmp, hammer_pseudofs_inmem_t pfsm)
{
	hammer_rel(&pfsm->lock);
	if (hammer_norefs(&pfsm->lock)) {
		RB_REMOVE(hammer_pfs_rb_tree, &hmp->rb_pfsm_root, pfsm);
		kfree(pfsm, hmp->m_misc);
	}
}

/*
 * Called by hammer_sync_inode().
 */
static int
hammer_update_inode(hammer_cursor_t cursor, hammer_inode_t ip)
{
	hammer_transaction_t trans = cursor->trans;
	hammer_record_t record;
	int error;
	int redirty;

retry:
	error = 0;

	/*
	 * If the inode has a presence on-disk then locate it and mark
	 * it deleted, setting DELONDISK.
	 *
	 * The record may or may not be physically deleted, depending on
	 * the retention policy.
	 */
	if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) ==
	    HAMMER_INODE_ONDISK) {
		hammer_normalize_cursor(cursor);
		cursor->key_beg.localization = ip->obj_localization + 
					       HAMMER_LOCALIZE_INODE;
		cursor->key_beg.obj_id = ip->obj_id;
		cursor->key_beg.key = 0;
		cursor->key_beg.create_tid = 0;
		cursor->key_beg.delete_tid = 0;
		cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
		cursor->key_beg.obj_type = 0;
		cursor->asof = ip->obj_asof;
		cursor->flags &= ~HAMMER_CURSOR_INITMASK;
		cursor->flags |= HAMMER_CURSOR_GET_LEAF | HAMMER_CURSOR_ASOF;
		cursor->flags |= HAMMER_CURSOR_BACKEND;

		error = hammer_btree_lookup(cursor);
		if (hammer_debug_inode)
			kprintf("IPDEL %p %08x %d", ip, ip->flags, error);

		if (error == 0) {
			error = hammer_ip_delete_record(cursor, ip, trans->tid);
			if (hammer_debug_inode)
				kprintf(" error %d\n", error);
			if (error == 0) {
				ip->flags |= HAMMER_INODE_DELONDISK;
			}
			if (cursor->node)
				hammer_cache_node(&ip->cache[0], cursor->node);
		}
		if (error == EDEADLK) {
			hammer_done_cursor(cursor);
			error = hammer_init_cursor(trans, cursor,
						   &ip->cache[0], ip);
			if (hammer_debug_inode)
				kprintf("IPDED %p %d\n", ip, error);
			if (error == 0)
				goto retry;
		}
	}

	/*
	 * Ok, write out the initial record or a new record (after deleting
	 * the old one), unless the DELETED flag is set.  This routine will
	 * clear DELONDISK if it writes out a record.
	 *
	 * Update our inode statistics if this is the first application of
	 * the inode on-disk.
	 */
	if (error == 0 && (ip->flags & HAMMER_INODE_DELETED) == 0) {
		/*
		 * Generate a record and write it to the media.  We clean-up
		 * the state before releasing so we do not have to set-up
		 * a flush_group.
		 */
		record = hammer_alloc_mem_record(ip, 0);
		record->type = HAMMER_MEM_RECORD_INODE;
		record->flush_state = HAMMER_FST_FLUSH;
		record->leaf = ip->sync_ino_leaf;
		record->leaf.base.create_tid = trans->tid;
		record->leaf.data_len = sizeof(ip->sync_ino_data);
		record->leaf.create_ts = trans->time32;
		record->data = (void *)&ip->sync_ino_data;
		record->flags |= HAMMER_RECF_INTERLOCK_BE;

		/*
		 * If this flag is set we cannot sync the new file size
		 * because we haven't finished related truncations.  The
		 * inode will be flushed in another flush group to finish
		 * the job.
		 */
		if ((ip->flags & HAMMER_INODE_WOULDBLOCK) &&
		    ip->sync_ino_data.size != ip->ino_data.size) {
			redirty = 1;
			ip->sync_ino_data.size = ip->ino_data.size;
		} else {
			redirty = 0;
		}

		for (;;) {
			error = hammer_ip_sync_record_cursor(cursor, record);
			if (hammer_debug_inode)
				kprintf("GENREC %p rec %08x %d\n",	
					ip, record->flags, error);
			if (error != EDEADLK)
				break;
			hammer_done_cursor(cursor);
			error = hammer_init_cursor(trans, cursor,
						   &ip->cache[0], ip);
			if (hammer_debug_inode)
				kprintf("GENREC reinit %d\n", error);
			if (error)
				break;
		}

		/*
		 * Note:  The record was never on the inode's record tree
		 * so just wave our hands importantly and destroy it.
		 */
		record->flags |= HAMMER_RECF_COMMITTED;
		record->flags &= ~HAMMER_RECF_INTERLOCK_BE;
		record->flush_state = HAMMER_FST_IDLE;
		++ip->rec_generation;
		hammer_rel_mem_record(record);

		/*
		 * Finish up.
		 */
		if (error == 0) {
			if (hammer_debug_inode)
				kprintf("CLEANDELOND %p %08x\n", ip, ip->flags);
			ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
					    HAMMER_INODE_SDIRTY |
					    HAMMER_INODE_ATIME |
					    HAMMER_INODE_MTIME);
			ip->flags &= ~HAMMER_INODE_DELONDISK;
			if (redirty)
				ip->sync_flags |= HAMMER_INODE_DDIRTY;

			/*
			 * Root volume count of inodes
			 */
			hammer_sync_lock_sh(trans);
			if ((ip->flags & HAMMER_INODE_ONDISK) == 0) {
				hammer_modify_volume_field(trans,
							   trans->rootvol,
							   vol0_stat_inodes);
				++ip->hmp->rootvol->ondisk->vol0_stat_inodes;
				hammer_modify_volume_done(trans->rootvol);
				ip->flags |= HAMMER_INODE_ONDISK;
				if (hammer_debug_inode)
					kprintf("NOWONDISK %p\n", ip);
			}
			hammer_sync_unlock(trans);
		}
	}

	/*
	 * If the inode has been destroyed, clean out any left-over flags
	 * that may have been set by the frontend.
	 */
	if (error == 0 && (ip->flags & HAMMER_INODE_DELETED)) { 
		ip->sync_flags &= ~(HAMMER_INODE_DDIRTY |
				    HAMMER_INODE_SDIRTY |
				    HAMMER_INODE_ATIME |
				    HAMMER_INODE_MTIME);
	}
	return(error);
}

/*
 * Update only the itimes fields.
 *
 * ATIME can be updated without generating any UNDO.  MTIME is updated
 * with UNDO so it is guaranteed to be synchronized properly in case of
 * a crash.
 *
 * Neither field is included in the B-Tree leaf element's CRC, which is how
 * we can get away with updating ATIME the way we do.
 */
static int
hammer_update_itimes(hammer_cursor_t cursor, hammer_inode_t ip)
{
	hammer_transaction_t trans = cursor->trans;
	int error;

retry:
	if ((ip->flags & (HAMMER_INODE_ONDISK|HAMMER_INODE_DELONDISK)) !=
	    HAMMER_INODE_ONDISK) {
		return(0);
	}

	hammer_normalize_cursor(cursor);
	cursor->key_beg.localization = ip->obj_localization + 
				       HAMMER_LOCALIZE_INODE;
	cursor->key_beg.obj_id = ip->obj_id;
	cursor->key_beg.key = 0;
	cursor->key_beg.create_tid = 0;
	cursor->key_beg.delete_tid = 0;
	cursor->key_beg.rec_type = HAMMER_RECTYPE_INODE;
	cursor->key_beg.obj_type = 0;
	cursor->asof = ip->obj_asof;
	cursor->flags &= ~HAMMER_CURSOR_INITMASK;
	cursor->flags |= HAMMER_CURSOR_ASOF;
	cursor->flags |= HAMMER_CURSOR_GET_LEAF;
	cursor->flags |= HAMMER_CURSOR_GET_DATA;
	cursor->flags |= HAMMER_CURSOR_BACKEND;

	error = hammer_btree_lookup(cursor);
	if (error == 0) {
		hammer_cache_node(&ip->cache[0], cursor->node);
		if (ip->sync_flags & HAMMER_INODE_MTIME) {
			/*
			 * Updating MTIME requires an UNDO.  Just cover
			 * both atime and mtime.
			 */
			hammer_sync_lock_sh(trans);
			hammer_modify_buffer(trans, cursor->data_buffer,
				     HAMMER_ITIMES_BASE(&cursor->data->inode),
				     HAMMER_ITIMES_BYTES);
			cursor->data->inode.atime = ip->sync_ino_data.atime;
			cursor->data->inode.mtime = ip->sync_ino_data.mtime;
			hammer_modify_buffer_done(cursor->data_buffer);
			hammer_sync_unlock(trans);
		} else if (ip->sync_flags & HAMMER_INODE_ATIME) {
			/*
			 * Updating atime only can be done in-place with
			 * no UNDO.
			 */
			hammer_sync_lock_sh(trans);
			hammer_modify_buffer(trans, cursor->data_buffer,
					     NULL, 0);
			cursor->data->inode.atime = ip->sync_ino_data.atime;
			hammer_modify_buffer_done(cursor->data_buffer);
			hammer_sync_unlock(trans);
		}
		ip->sync_flags &= ~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME);
	}
	if (error == EDEADLK) {
		hammer_done_cursor(cursor);
		error = hammer_init_cursor(trans, cursor,
					   &ip->cache[0], ip);
		if (error == 0)
			goto retry;
	}
	return(error);
}

/*
 * Release a reference on an inode, flush as requested.
 *
 * On the last reference we queue the inode to the flusher for its final
 * disposition.
 */
void
hammer_rel_inode(struct hammer_inode *ip, int flush)
{
	/*hammer_mount_t hmp = ip->hmp;*/

	/*
	 * Handle disposition when dropping the last ref.
	 */
	for (;;) {
		if (hammer_oneref(&ip->lock)) {
			/*
			 * Determine whether on-disk action is needed for
			 * the inode's final disposition.
			 */
			KKASSERT(ip->vp == NULL);
			hammer_inode_unloadable_check(ip, 0);
			if (ip->flags & HAMMER_INODE_MODMASK) {
				hammer_flush_inode(ip, 0);
			} else if (hammer_oneref(&ip->lock)) {
				hammer_unload_inode(ip);
				break;
			}
		} else {
			if (flush)
				hammer_flush_inode(ip, 0);

			/*
			 * The inode still has multiple refs, try to drop
			 * one ref.
			 */
			KKASSERT(hammer_isactive(&ip->lock) >= 1);
			if (hammer_isactive(&ip->lock) > 1) {
				hammer_rel(&ip->lock);
				break;
			}
		}
	}
}

/*
 * Unload and destroy the specified inode.  Must be called with one remaining
 * reference.  The reference is disposed of.
 *
 * The inode must be completely clean.
 */
static int
hammer_unload_inode(struct hammer_inode *ip)
{
	hammer_mount_t hmp = ip->hmp;

	KASSERT(hammer_oneref(&ip->lock),
		("hammer_unload_inode: %d refs\n", hammer_isactive(&ip->lock)));
	KKASSERT(ip->vp == NULL);
	KKASSERT(ip->flush_state == HAMMER_FST_IDLE);
	KKASSERT(ip->cursor_ip_refs == 0);
	KKASSERT(hammer_notlocked(&ip->lock));
	KKASSERT((ip->flags & HAMMER_INODE_MODMASK) == 0);

	KKASSERT(RB_EMPTY(&ip->rec_tree));
	KKASSERT(TAILQ_EMPTY(&ip->target_list));

	if (ip->flags & HAMMER_INODE_RDIRTY) {
		RB_REMOVE(hammer_redo_rb_tree, &hmp->rb_redo_root, ip);
		ip->flags &= ~HAMMER_INODE_RDIRTY;
	}
	RB_REMOVE(hammer_ino_rb_tree, &hmp->rb_inos_root, ip);

	hammer_free_inode(ip);
	return(0);
}

/*
 * Called during unmounting if a critical error occured.  The in-memory
 * inode and all related structures are destroyed.
 *
 * If a critical error did not occur the unmount code calls the standard
 * release and asserts that the inode is gone.
 */
int
hammer_destroy_inode_callback(struct hammer_inode *ip, void *data __unused)
{
	hammer_record_t rec;

	/*
	 * Get rid of the inodes in-memory records, regardless of their
	 * state, and clear the mod-mask.
	 */
	while ((rec = TAILQ_FIRST(&ip->target_list)) != NULL) {
		TAILQ_REMOVE(&ip->target_list, rec, target_entry);
		rec->target_ip = NULL;
		if (rec->flush_state == HAMMER_FST_SETUP)
			rec->flush_state = HAMMER_FST_IDLE;
	}
	while ((rec = RB_ROOT(&ip->rec_tree)) != NULL) {
		if (rec->flush_state == HAMMER_FST_FLUSH)
			--rec->flush_group->refs;
		else
			hammer_ref(&rec->lock);
		KKASSERT(hammer_oneref(&rec->lock));
		rec->flush_state = HAMMER_FST_IDLE;
		rec->flush_group = NULL;
		rec->flags |= HAMMER_RECF_DELETED_FE; /* wave hands */
		rec->flags |= HAMMER_RECF_DELETED_BE; /* wave hands */
		++ip->rec_generation;
		hammer_rel_mem_record(rec);
	}
	ip->flags &= ~HAMMER_INODE_MODMASK;
	ip->sync_flags &= ~HAMMER_INODE_MODMASK;
	KKASSERT(ip->vp == NULL);

	/*
	 * Remove the inode from any flush group, force it idle.  FLUSH
	 * and SETUP states have an inode ref.
	 */
	switch(ip->flush_state) {
	case HAMMER_FST_FLUSH:
		RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
		--ip->flush_group->refs;
		ip->flush_group = NULL;
		/* fall through */
	case HAMMER_FST_SETUP:
		hammer_rel(&ip->lock);
		ip->flush_state = HAMMER_FST_IDLE;
		/* fall through */
	case HAMMER_FST_IDLE:
		break;
	}

	/*
	 * There shouldn't be any associated vnode.  The unload needs at
	 * least one ref, if we do have a vp steal its ip ref.
	 */
	if (ip->vp) {
		kprintf("hammer_destroy_inode_callback: Unexpected "
			"vnode association ip %p vp %p\n", ip, ip->vp);
		ip->vp->v_data = NULL;
		ip->vp = NULL;
	} else {
		hammer_ref(&ip->lock);
	}
	hammer_unload_inode(ip);
	return(0);
}

/*
 * Called on mount -u when switching from RW to RO or vise-versa.  Adjust
 * the read-only flag for cached inodes.
 *
 * This routine is called from a RB_SCAN().
 */
int
hammer_reload_inode(hammer_inode_t ip, void *arg __unused)
{
	hammer_mount_t hmp = ip->hmp;

	if (hmp->ronly || hmp->asof != HAMMER_MAX_TID)
		ip->flags |= HAMMER_INODE_RO;
	else
		ip->flags &= ~HAMMER_INODE_RO;
	return(0);
}

/*
 * A transaction has modified an inode, requiring updates as specified by
 * the passed flags.
 *
 * HAMMER_INODE_DDIRTY: Inode data has been updated, not incl mtime/atime,
 *			and not including size changes due to write-append
 *			(but other size changes are included).
 * HAMMER_INODE_SDIRTY: Inode data has been updated, size changes due to
 *			write-append.
 * HAMMER_INODE_XDIRTY: Dirty in-memory records
 * HAMMER_INODE_BUFS:   Dirty buffer cache buffers
 * HAMMER_INODE_DELETED: Inode record/data must be deleted
 * HAMMER_INODE_ATIME/MTIME: mtime/atime has been updated
 */
void
hammer_modify_inode(hammer_transaction_t trans, hammer_inode_t ip, int flags)
{
	/* 
	 * ronly of 0 or 2 does not trigger assertion.
	 * 2 is a special error state 
	 */
	KKASSERT(ip->hmp->ronly != 1 ||
		  (flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY | 
			    HAMMER_INODE_SDIRTY |
			    HAMMER_INODE_BUFS | HAMMER_INODE_DELETED |
			    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) == 0);
	if ((ip->flags & HAMMER_INODE_RSV_INODES) == 0) {
		ip->flags |= HAMMER_INODE_RSV_INODES;
		++ip->hmp->rsv_inodes;
	}

	/*
	 * Set the NEWINODE flag in the transaction if the inode
	 * transitions to a dirty state.  This is used to track
	 * the load on the inode cache.
	 */
	if (trans &&
	    (ip->flags & HAMMER_INODE_MODMASK) == 0 &&
	    (flags & HAMMER_INODE_MODMASK)) {
		trans->flags |= HAMMER_TRANSF_NEWINODE;
	}

	ip->flags |= flags;
}

/*
 * Request that an inode be flushed.  This whole mess cannot block and may
 * recurse (if not synchronous).  Once requested HAMMER will attempt to
 * actively flush the inode until the flush can be done.
 *
 * The inode may already be flushing, or may be in a setup state.  We can
 * place the inode in a flushing state if it is currently idle and flag it
 * to reflush if it is currently flushing.
 *
 * Upon return if the inode could not be flushed due to a setup
 * dependancy, then it will be automatically flushed when the dependancy
 * is satisfied.
 */
void
hammer_flush_inode(hammer_inode_t ip, int flags)
{
	hammer_mount_t hmp;
	hammer_flush_group_t flg;
	int good;

	/*
	 * fill_flush_group is the first flush group we may be able to
	 * continue filling, it may be open or closed but it will always
	 * be past the currently flushing (running) flg.
	 *
	 * next_flush_group is the next open flush group.
	 */
	hmp = ip->hmp;
	while ((flg = hmp->fill_flush_group) != NULL) {
		KKASSERT(flg->running == 0);
		if (flg->total_count + flg->refs <= ip->hmp->undo_rec_limit &&
		    flg->total_count <= hammer_autoflush) {
			break;
		}
		hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
		hammer_flusher_async(ip->hmp, flg);
	}
	if (flg == NULL) {
		flg = kmalloc(sizeof(*flg), hmp->m_misc, M_WAITOK|M_ZERO);
		flg->seq = hmp->flusher.next++;
		if (hmp->next_flush_group == NULL)
			hmp->next_flush_group = flg;
		if (hmp->fill_flush_group == NULL)
			hmp->fill_flush_group = flg;
		RB_INIT(&flg->flush_tree);
		TAILQ_INSERT_TAIL(&hmp->flush_group_list, flg, flush_entry);
	}

	/*
	 * Trivial 'nothing to flush' case.  If the inode is in a SETUP
	 * state we have to put it back into an IDLE state so we can
	 * drop the extra ref.
	 *
	 * If we have a parent dependancy we must still fall through
	 * so we can run it.
	 */
	if ((ip->flags & HAMMER_INODE_MODMASK) == 0) {
		if (ip->flush_state == HAMMER_FST_SETUP &&
		    TAILQ_EMPTY(&ip->target_list)) {
			ip->flush_state = HAMMER_FST_IDLE;
			hammer_rel_inode(ip, 0);
		}
		if (ip->flush_state == HAMMER_FST_IDLE)
			return;
	}

	/*
	 * Our flush action will depend on the current state.
	 */
	switch(ip->flush_state) {
	case HAMMER_FST_IDLE:
		/*
		 * We have no dependancies and can flush immediately.  Some
		 * our children may not be flushable so we have to re-test
		 * with that additional knowledge.
		 */
		hammer_flush_inode_core(ip, flg, flags);
		break;
	case HAMMER_FST_SETUP:
		/*
		 * Recurse upwards through dependancies via target_list
		 * and start their flusher actions going if possible.
		 *
		 * 'good' is our connectivity.  -1 means we have none and
		 * can't flush, 0 means there weren't any dependancies, and
		 * 1 means we have good connectivity.
		 */
		good = hammer_setup_parent_inodes(ip, 0, flg);

		if (good >= 0) {
			/*
			 * We can continue if good >= 0.  Determine how 
			 * many records under our inode can be flushed (and
			 * mark them).
			 */
			hammer_flush_inode_core(ip, flg, flags);
		} else {
			/*
			 * Parent has no connectivity, tell it to flush
			 * us as soon as it does.
			 *
			 * The REFLUSH flag is also needed to trigger
			 * dependancy wakeups.
			 */
			ip->flags |= HAMMER_INODE_CONN_DOWN |
				     HAMMER_INODE_REFLUSH;
			if (flags & HAMMER_FLUSH_SIGNAL) {
				ip->flags |= HAMMER_INODE_RESIGNAL;
				hammer_flusher_async(ip->hmp, flg);
			}
		}
		break;
	case HAMMER_FST_FLUSH:
		/*
		 * We are already flushing, flag the inode to reflush
		 * if needed after it completes its current flush.
		 *
		 * The REFLUSH flag is also needed to trigger
		 * dependancy wakeups.
		 */
		if ((ip->flags & HAMMER_INODE_REFLUSH) == 0)
			ip->flags |= HAMMER_INODE_REFLUSH;
		if (flags & HAMMER_FLUSH_SIGNAL) {
			ip->flags |= HAMMER_INODE_RESIGNAL;
			hammer_flusher_async(ip->hmp, flg);
		}
		break;
	}
}

/*
 * Scan ip->target_list, which is a list of records owned by PARENTS to our
 * ip which reference our ip.
 *
 * XXX This is a huge mess of recursive code, but not one bit of it blocks
 *     so for now do not ref/deref the structures.  Note that if we use the
 *     ref/rel code later, the rel CAN block.
 */
static int
hammer_setup_parent_inodes(hammer_inode_t ip, int depth,
			   hammer_flush_group_t flg)
{
	hammer_record_t depend;
	int good;
	int r;

	/*
	 * If we hit our recursion limit and we have parent dependencies
	 * We cannot continue.  Returning < 0 will cause us to be flagged
	 * for reflush.  Returning -2 cuts off additional dependency checks
	 * because they are likely to also hit the depth limit.
	 *
	 * We cannot return < 0 if there are no dependencies or there might
	 * not be anything to wakeup (ip).
	 */
	if (depth == 20 && TAILQ_FIRST(&ip->target_list)) {
		kprintf("HAMMER Warning: depth limit reached on "
			"setup recursion, inode %p %016llx\n",
			ip, (long long)ip->obj_id);
		return(-2);
	}

	/*
	 * Scan dependencies
	 */
	good = 0;
	TAILQ_FOREACH(depend, &ip->target_list, target_entry) {
		r = hammer_setup_parent_inodes_helper(depend, depth, flg);
		KKASSERT(depend->target_ip == ip);
		if (r < 0 && good == 0)
			good = -1;
		if (r > 0)
			good = 1;

		/*
		 * If we failed due to the recursion depth limit then stop
		 * now.
		 */
		if (r == -2)
			break;
	}
	return(good);
}

/*
 * This helper function takes a record representing the dependancy between
 * the parent inode and child inode.
 *
 * record->ip		= parent inode
 * record->target_ip	= child inode
 * 
 * We are asked to recurse upwards and convert the record from SETUP
 * to FLUSH if possible.
 *
 * Return 1 if the record gives us connectivity
 *
 * Return 0 if the record is not relevant 
 *
 * Return -1 if we can't resolve the dependancy and there is no connectivity.
 */
static int
hammer_setup_parent_inodes_helper(hammer_record_t record, int depth,
				  hammer_flush_group_t flg)
{
	hammer_mount_t hmp;
	hammer_inode_t pip;
	int good;

	KKASSERT(record->flush_state != HAMMER_FST_IDLE);
	pip = record->ip;
	hmp = pip->hmp;

	/*
	 * If the record is already flushing, is it in our flush group?
	 *
	 * If it is in our flush group but it is a general record or a 
	 * delete-on-disk, it does not improve our connectivity (return 0),
	 * and if the target inode is not trying to destroy itself we can't
	 * allow the operation yet anyway (the second return -1).
	 */
	if (record->flush_state == HAMMER_FST_FLUSH) {
		/*
		 * If not in our flush group ask the parent to reflush
		 * us as soon as possible.
		 */
		if (record->flush_group != flg) {
			pip->flags |= HAMMER_INODE_REFLUSH;
			record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
			return(-1);
		}

		/*
		 * If in our flush group everything is already set up,
		 * just return whether the record will improve our
		 * visibility or not.
		 */
		if (record->type == HAMMER_MEM_RECORD_ADD)
			return(1);
		return(0);
	}

	/*
	 * It must be a setup record.  Try to resolve the setup dependancies
	 * by recursing upwards so we can place ip on the flush list.
	 *
	 * Limit ourselves to 20 levels of recursion to avoid blowing out
	 * the kernel stack.  If we hit the recursion limit we can't flush
	 * until the parent flushes.  The parent will flush independantly
	 * on its own and ultimately a deep recursion will be resolved.
	 */
	KKASSERT(record->flush_state == HAMMER_FST_SETUP);

	good = hammer_setup_parent_inodes(pip, depth + 1, flg);

	/*
	 * If good < 0 the parent has no connectivity and we cannot safely
	 * flush the directory entry, which also means we can't flush our
	 * ip.  Flag us for downward recursion once the parent's
	 * connectivity is resolved.  Flag the parent for [re]flush or it
	 * may not check for downward recursions.
	 */
	if (good < 0) {
		pip->flags |= HAMMER_INODE_REFLUSH;
		record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
		return(good);
	}

	/*
	 * We are go, place the parent inode in a flushing state so we can
	 * place its record in a flushing state.  Note that the parent
	 * may already be flushing.  The record must be in the same flush
	 * group as the parent.
	 */
	if (pip->flush_state != HAMMER_FST_FLUSH)
		hammer_flush_inode_core(pip, flg, HAMMER_FLUSH_RECURSION);
	KKASSERT(pip->flush_state == HAMMER_FST_FLUSH);

	/*
	 * It is possible for a rename to create a loop in the recursion
	 * and revisit a record.  This will result in the record being
	 * placed in a flush state unexpectedly.  This check deals with
	 * the case.
	 */
	if (record->flush_state == HAMMER_FST_FLUSH) {
		if (record->type == HAMMER_MEM_RECORD_ADD)
			return(1);
		return(0);
	}

	KKASSERT(record->flush_state == HAMMER_FST_SETUP);

#if 0
	if (record->type == HAMMER_MEM_RECORD_DEL &&
	    (record->target_ip->flags & (HAMMER_INODE_DELETED|HAMMER_INODE_DELONDISK)) == 0) {
		/*
		 * Regardless of flushing state we cannot sync this path if the
		 * record represents a delete-on-disk but the target inode
		 * is not ready to sync its own deletion.
		 *
		 * XXX need to count effective nlinks to determine whether
		 * the flush is ok, otherwise removing a hardlink will
		 * just leave the DEL record to rot.
		 */
		record->target_ip->flags |= HAMMER_INODE_REFLUSH;
		return(-1);
	} else
#endif
	if (pip->flush_group == flg) {
		/*
		 * Because we have not calculated nlinks yet we can just
		 * set records to the flush state if the parent is in
		 * the same flush group as we are.
		 */
		record->flush_state = HAMMER_FST_FLUSH;
		record->flush_group = flg;
		++record->flush_group->refs;
		hammer_ref(&record->lock);

		/*
		 * A general directory-add contributes to our visibility.
		 *
		 * Otherwise it is probably a directory-delete or 
		 * delete-on-disk record and does not contribute to our
		 * visbility (but we can still flush it).
		 */
		if (record->type == HAMMER_MEM_RECORD_ADD)
			return(1);
		return(0);
	} else {
		/*
		 * If the parent is not in our flush group we cannot
		 * flush this record yet, there is no visibility.
		 * We tell the parent to reflush and mark ourselves
		 * so the parent knows it should flush us too.
		 */
		pip->flags |= HAMMER_INODE_REFLUSH;
		record->target_ip->flags |= HAMMER_INODE_CONN_DOWN;
		return(-1);
	}
}

/*
 * This is the core routine placing an inode into the FST_FLUSH state.
 */
static void
hammer_flush_inode_core(hammer_inode_t ip, hammer_flush_group_t flg, int flags)
{
	hammer_mount_t hmp = ip->hmp;
	int go_count;

	/*
	 * Set flush state and prevent the flusher from cycling into
	 * the next flush group.  Do not place the ip on the list yet.
	 * Inodes not in the idle state get an extra reference.
	 */
	KKASSERT(ip->flush_state != HAMMER_FST_FLUSH);
	if (ip->flush_state == HAMMER_FST_IDLE)
		hammer_ref(&ip->lock);
	ip->flush_state = HAMMER_FST_FLUSH;
	ip->flush_group = flg;
	++hmp->flusher.group_lock;
	++hmp->count_iqueued;
	++hammer_count_iqueued;
	++flg->total_count;
	hammer_redo_fifo_start_flush(ip);

#if 0
	/*
	 * We need to be able to vfsync/truncate from the backend.
	 *
	 * XXX Any truncation from the backend will acquire the vnode
	 *     independently.
	 */
	KKASSERT((ip->flags & HAMMER_INODE_VHELD) == 0);
	if (ip->vp && (ip->vp->v_flag & VINACTIVE) == 0) {
		ip->flags |= HAMMER_INODE_VHELD;
		vref(ip->vp);
	}
#endif

	/*
	 * Figure out how many in-memory records we can actually flush
	 * (not including inode meta-data, buffers, etc).
	 */
	KKASSERT((ip->flags & HAMMER_INODE_WOULDBLOCK) == 0);
	if (flags & HAMMER_FLUSH_RECURSION) {
		/*
		 * If this is a upwards recursion we do not want to
		 * recurse down again!
		 */
		go_count = 1;
#if 0
	} else if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
		/*
		 * No new records are added if we must complete a flush
		 * from a previous cycle, but we do have to move the records
		 * from the previous cycle to the current one.
		 */
#if 0
		go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
				   hammer_syncgrp_child_callback, NULL);
#endif
		go_count = 1;
#endif
	} else {
		/*
		 * Normal flush, scan records and bring them into the flush.
		 * Directory adds and deletes are usually skipped (they are
		 * grouped with the related inode rather then with the
		 * directory).
		 *
		 * go_count can be negative, which means the scan aborted
		 * due to the flush group being over-full and we should
		 * flush what we have.
		 */
		go_count = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
				   hammer_setup_child_callback, NULL);
	}

	/*
	 * This is a more involved test that includes go_count.  If we
	 * can't flush, flag the inode and return.  If go_count is 0 we
	 * were are unable to flush any records in our rec_tree and
	 * must ignore the XDIRTY flag.
	 */
	if (go_count == 0) {
		if ((ip->flags & HAMMER_INODE_MODMASK_NOXDIRTY) == 0) {
			--hmp->count_iqueued;
			--hammer_count_iqueued;

			--flg->total_count;
			ip->flush_state = HAMMER_FST_SETUP;
			ip->flush_group = NULL;
			if (flags & HAMMER_FLUSH_SIGNAL) {
				ip->flags |= HAMMER_INODE_REFLUSH |
					     HAMMER_INODE_RESIGNAL;
			} else {
				ip->flags |= HAMMER_INODE_REFLUSH;
			}
#if 0
			if (ip->flags & HAMMER_INODE_VHELD) {
				ip->flags &= ~HAMMER_INODE_VHELD;
				vrele(ip->vp);
			}
#endif

			/*
			 * REFLUSH is needed to trigger dependancy wakeups
			 * when an inode is in SETUP.
			 */
			ip->flags |= HAMMER_INODE_REFLUSH;
			if (--hmp->flusher.group_lock == 0)
				wakeup(&hmp->flusher.group_lock);
			return;
		}
	}

	/*
	 * Snapshot the state of the inode for the backend flusher.
	 *
	 * We continue to retain save_trunc_off even when all truncations
	 * have been resolved as an optimization to determine if we can
	 * skip the B-Tree lookup for overwrite deletions.
	 *
	 * NOTE: The DELETING flag is a mod flag, but it is also sticky,
	 * and stays in ip->flags.  Once set, it stays set until the
	 * inode is destroyed.
	 */
	if (ip->flags & HAMMER_INODE_TRUNCATED) {
		KKASSERT((ip->sync_flags & HAMMER_INODE_TRUNCATED) == 0);
		ip->sync_trunc_off = ip->trunc_off;
		ip->trunc_off = 0x7FFFFFFFFFFFFFFFLL;
		ip->flags &= ~HAMMER_INODE_TRUNCATED;
		ip->sync_flags |= HAMMER_INODE_TRUNCATED;

		/*
		 * The save_trunc_off used to cache whether the B-Tree
		 * holds any records past that point is not used until
		 * after the truncation has succeeded, so we can safely
		 * set it now.
		 */
		if (ip->save_trunc_off > ip->sync_trunc_off)
			ip->save_trunc_off = ip->sync_trunc_off;
	}
	ip->sync_flags |= (ip->flags & HAMMER_INODE_MODMASK &
			   ~HAMMER_INODE_TRUNCATED);
	ip->sync_ino_leaf = ip->ino_leaf;
	ip->sync_ino_data = ip->ino_data;
	ip->flags &= ~HAMMER_INODE_MODMASK | HAMMER_INODE_TRUNCATED;
#ifdef DEBUG_TRUNCATE
	if ((ip->sync_flags & HAMMER_INODE_TRUNCATED) && ip == HammerTruncIp)
		kprintf("truncateS %016llx\n", ip->sync_trunc_off);
#endif

	/*
	 * The flusher list inherits our inode and reference.
	 */
	KKASSERT(flg->running == 0);
	RB_INSERT(hammer_fls_rb_tree, &flg->flush_tree, ip);
	if (--hmp->flusher.group_lock == 0)
		wakeup(&hmp->flusher.group_lock);

	/*
	 * Auto-flush the group if it grows too large.  Make sure the
	 * inode reclaim wait pipeline continues to work.
	 */
	if (flg->total_count >= hammer_autoflush ||
	    flg->total_count >= hammer_limit_reclaim / 4) {
		if (hmp->fill_flush_group == flg)
			hmp->fill_flush_group = TAILQ_NEXT(flg, flush_entry);
		hammer_flusher_async(hmp, flg);
	}
}

/*
 * Callback for scan of ip->rec_tree.  Try to include each record in our
 * flush.  ip->flush_group has been set but the inode has not yet been
 * moved into a flushing state.
 *
 * If we get stuck on a record we have to set HAMMER_INODE_REFLUSH on
 * both inodes.
 *
 * We return 1 for any record placed or found in FST_FLUSH, which prevents
 * the caller from shortcutting the flush.
 */
static int
hammer_setup_child_callback(hammer_record_t rec, void *data)
{
	hammer_flush_group_t flg;
	hammer_inode_t target_ip;
	hammer_inode_t ip;
	int r;

	/*
	 * Records deleted or committed by the backend are ignored.
	 * Note that the flush detects deleted frontend records at
	 * multiple points to deal with races.  This is just the first
	 * line of defense.  The only time HAMMER_RECF_DELETED_FE cannot
	 * be set is when HAMMER_RECF_INTERLOCK_BE is set, because it
	 * messes up link-count calculations.
	 *
	 * NOTE: Don't get confused between record deletion and, say,
	 * directory entry deletion.  The deletion of a directory entry
	 * which is on-media has nothing to do with the record deletion
	 * flags.
	 */
	if (rec->flags & (HAMMER_RECF_DELETED_FE | HAMMER_RECF_DELETED_BE |
			  HAMMER_RECF_COMMITTED)) {
		if (rec->flush_state == HAMMER_FST_FLUSH) {
			KKASSERT(rec->flush_group == rec->ip->flush_group);
			r = 1;
		} else {
			r = 0;
		}
		return(r);
	}

	/*
	 * If the record is in an idle state it has no dependancies and
	 * can be flushed.
	 */
	ip = rec->ip;
	flg = ip->flush_group;
	r = 0;

	switch(rec->flush_state) {
	case HAMMER_FST_IDLE:
		/*
		 * The record has no setup dependancy, we can flush it.
		 */
		KKASSERT(rec->target_ip == NULL);
		rec->flush_state = HAMMER_FST_FLUSH;
		rec->flush_group = flg;
		++flg->refs;
		hammer_ref(&rec->lock);
		r = 1;
		break;
	case HAMMER_FST_SETUP:
		/*
		 * The record has a setup dependancy.  These are typically
		 * directory entry adds and deletes.  Such entries will be
		 * flushed when their inodes are flushed so we do not
		 * usually have to add them to the flush here.  However,
		 * if the target_ip has set HAMMER_INODE_CONN_DOWN then
		 * it is asking us to flush this record (and it).
		 */
		target_ip = rec->target_ip;
		KKASSERT(target_ip != NULL);
		KKASSERT(target_ip->flush_state != HAMMER_FST_IDLE);

		/*
		 * If the target IP is already flushing in our group
		 * we could associate the record, but target_ip has
		 * already synced ino_data to sync_ino_data and we
		 * would also have to adjust nlinks.   Plus there are
		 * ordering issues for adds and deletes.
		 *
		 * Reflush downward if this is an ADD, and upward if
		 * this is a DEL.
		 */
		if (target_ip->flush_state == HAMMER_FST_FLUSH) {
			if (rec->flush_state == HAMMER_MEM_RECORD_ADD)
				ip->flags |= HAMMER_INODE_REFLUSH;
			else
				target_ip->flags |= HAMMER_INODE_REFLUSH;
			break;
		} 

		/*
		 * Target IP is not yet flushing.  This can get complex
		 * because we have to be careful about the recursion.
		 *
		 * Directories create an issue for us in that if a flush
		 * of a directory is requested the expectation is to flush
		 * any pending directory entries, but this will cause the
		 * related inodes to recursively flush as well.  We can't
		 * really defer the operation so just get as many as we
		 * can and
		 */
#if 0
		if ((target_ip->flags & HAMMER_INODE_RECLAIM) == 0 &&
		    (target_ip->flags & HAMMER_INODE_CONN_DOWN) == 0) {
			/*
			 * We aren't reclaiming and the target ip was not
			 * previously prevented from flushing due to this
			 * record dependancy.  Do not flush this record.
			 */
			/*r = 0;*/
		} else
#endif
		if (flg->total_count + flg->refs >
			   ip->hmp->undo_rec_limit) {
			/*
			 * Our flush group is over-full and we risk blowing
			 * out the UNDO FIFO.  Stop the scan, flush what we
			 * have, then reflush the directory.
			 *
			 * The directory may be forced through multiple
			 * flush groups before it can be completely
			 * flushed.
			 */
			ip->flags |= HAMMER_INODE_RESIGNAL |
				     HAMMER_INODE_REFLUSH;
			r = -1;
		} else if (rec->type == HAMMER_MEM_RECORD_ADD) {
			/*
			 * If the target IP is not flushing we can force
			 * it to flush, even if it is unable to write out
			 * any of its own records we have at least one in
			 * hand that we CAN deal with.
			 */
			rec->flush_state = HAMMER_FST_FLUSH;
			rec->flush_group = flg;
			++flg->refs;
			hammer_ref(&rec->lock);
			hammer_flush_inode_core(target_ip, flg,
						HAMMER_FLUSH_RECURSION);
			r = 1;
		} else {
			/*
			 * General or delete-on-disk record.
			 *
			 * XXX this needs help.  If a delete-on-disk we could
			 * disconnect the target.  If the target has its own
			 * dependancies they really need to be flushed.
			 *
			 * XXX
			 */
			rec->flush_state = HAMMER_FST_FLUSH;
			rec->flush_group = flg;
			++flg->refs;
			hammer_ref(&rec->lock);
			hammer_flush_inode_core(target_ip, flg,
						HAMMER_FLUSH_RECURSION);
			r = 1;
		}
		break;
	case HAMMER_FST_FLUSH:
		/* 
		 * The record could be part of a previous flush group if the
		 * inode is a directory (the record being a directory entry).
		 * Once the flush group was closed a hammer_test_inode()
		 * function can cause a new flush group to be setup, placing
		 * the directory inode itself in a new flush group.
		 *
		 * When associated with a previous flush group we count it
		 * as if it were in our current flush group, since it will
		 * effectively be flushed by the time we flush our current
		 * flush group.
		 */
		KKASSERT(
		    rec->ip->ino_data.obj_type == HAMMER_OBJTYPE_DIRECTORY ||
		    rec->flush_group == flg);
		r = 1;
		break;
	}
	return(r);
}

#if 0
/*
 * This version just moves records already in a flush state to the new
 * flush group and that is it.
 */
static int
hammer_syncgrp_child_callback(hammer_record_t rec, void *data)
{
	hammer_inode_t ip = rec->ip;

	switch(rec->flush_state) {
	case HAMMER_FST_FLUSH:
		KKASSERT(rec->flush_group == ip->flush_group);
		break;
	default:
		break;
	}
	return(0);
}
#endif

/*
 * Wait for a previously queued flush to complete.
 *
 * If a critical error occured we don't try to wait.
 */
void
hammer_wait_inode(hammer_inode_t ip)
{
	hammer_flush_group_t flg;

	flg = NULL;
	if ((ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
		while (ip->flush_state != HAMMER_FST_IDLE &&
		       (ip->hmp->flags & HAMMER_MOUNT_CRITICAL_ERROR) == 0) {
			if (ip->flush_state == HAMMER_FST_SETUP)
				hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);

			/*
			 * If the inode was already being flushed its flg
			 * may not have been queued to the backend.  We have
			 * to make sure it gets queued or we can wind up
			 * blocked or deadlocked (particularly if we are
			 * the vnlru thread).
			 */
			KKASSERT(ip->flush_group);
			if (ip->flush_group->closed == 0) {
				kprintf("hammer: debug: forcing async "
					"flush ip %016jx\n",
					(intmax_t)ip->obj_id);
				hammer_flusher_async(ip->hmp, ip->flush_group);
			}
			if (ip->flush_state != HAMMER_FST_IDLE) {
				ip->flags |= HAMMER_INODE_FLUSHW;
				tsleep(&ip->flags, 0, "hmrwin", 0);
			}
		}
	}
}

/*
 * Called by the backend code when a flush has been completed.
 * The inode has already been removed from the flush list.
 *
 * A pipelined flush can occur, in which case we must re-enter the
 * inode on the list and re-copy its fields.
 */
void
hammer_flush_inode_done(hammer_inode_t ip, int error)
{
	hammer_mount_t hmp;
	int dorel;

	KKASSERT(ip->flush_state == HAMMER_FST_FLUSH);

	hmp = ip->hmp;

	/*
	 * Auto-reflush if the backend could not completely flush
	 * the inode.  This fixes a case where a deferred buffer flush
	 * could cause fsync to return early.
	 */
	if (ip->sync_flags & HAMMER_INODE_MODMASK)
		ip->flags |= HAMMER_INODE_REFLUSH;

	/*
	 * Merge left-over flags back into the frontend and fix the state.
	 * Incomplete truncations are retained by the backend.
	 */
	ip->error = error;
	ip->flags |= ip->sync_flags & ~HAMMER_INODE_TRUNCATED;
	ip->sync_flags &= HAMMER_INODE_TRUNCATED;

	/*
	 * The backend may have adjusted nlinks, so if the adjusted nlinks
	 * does not match the fronttend set the frontend's DDIRTY flag again.
	 */
	if (ip->ino_data.nlinks != ip->sync_ino_data.nlinks)
		ip->flags |= HAMMER_INODE_DDIRTY;

	/*
	 * Fix up the dirty buffer status.
	 */
	if (ip->vp && RB_ROOT(&ip->vp->v_rbdirty_tree)) {
		ip->flags |= HAMMER_INODE_BUFS;
	}
	hammer_redo_fifo_end_flush(ip);

	/*
	 * Re-set the XDIRTY flag if some of the inode's in-memory records
	 * could not be flushed.
	 */
	KKASSERT((RB_EMPTY(&ip->rec_tree) &&
		  (ip->flags & HAMMER_INODE_XDIRTY) == 0) ||
		 (!RB_EMPTY(&ip->rec_tree) &&
		  (ip->flags & HAMMER_INODE_XDIRTY) != 0));

	/*
	 * Do not lose track of inodes which no longer have vnode
	 * assocations, otherwise they may never get flushed again.
	 *
	 * The reflush flag can be set superfluously, causing extra pain
	 * for no reason.  If the inode is no longer modified it no longer
	 * needs to be flushed.
	 */
	if (ip->flags & HAMMER_INODE_MODMASK) {
		if (ip->vp == NULL)
			ip->flags |= HAMMER_INODE_REFLUSH;
	} else {
		ip->flags &= ~HAMMER_INODE_REFLUSH;
	}

	/*
	 * Adjust the flush state.
	 */
	if (ip->flags & HAMMER_INODE_WOULDBLOCK) {
		/*
		 * We were unable to flush out all our records, leave the
		 * inode in a flush state and in the current flush group.
		 * The flush group will be re-run.
		 *
		 * This occurs if the UNDO block gets too full or there is
		 * too much dirty meta-data and allows the flusher to
		 * finalize the UNDO block and then re-flush.
		 */
		ip->flags &= ~HAMMER_INODE_WOULDBLOCK;
		dorel = 0;
	} else {
		/*
		 * Remove from the flush_group
		 */
		RB_REMOVE(hammer_fls_rb_tree, &ip->flush_group->flush_tree, ip);
		ip->flush_group = NULL;

#if 0
		/*
		 * Clean up the vnode ref and tracking counts.
		 */
		if (ip->flags & HAMMER_INODE_VHELD) {
			ip->flags &= ~HAMMER_INODE_VHELD;
			vrele(ip->vp);
		}
#endif
		--hmp->count_iqueued;
		--hammer_count_iqueued;

		/*
		 * And adjust the state.
		 */
		if (TAILQ_EMPTY(&ip->target_list) && RB_EMPTY(&ip->rec_tree)) {
			ip->flush_state = HAMMER_FST_IDLE;
			dorel = 1;
		} else {
			ip->flush_state = HAMMER_FST_SETUP;
			dorel = 0;
		}

		/*
		 * If the frontend is waiting for a flush to complete,
		 * wake it up.
		 */
		if (ip->flags & HAMMER_INODE_FLUSHW) {
			ip->flags &= ~HAMMER_INODE_FLUSHW;
			wakeup(&ip->flags);
		}

		/*
		 * If the frontend made more changes and requested another
		 * flush, then try to get it running.
		 *
		 * Reflushes are aborted when the inode is errored out.
		 */
		if (ip->flags & HAMMER_INODE_REFLUSH) {
			ip->flags &= ~HAMMER_INODE_REFLUSH;
			if (ip->flags & HAMMER_INODE_RESIGNAL) {
				ip->flags &= ~HAMMER_INODE_RESIGNAL;
				hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
			} else {
				hammer_flush_inode(ip, 0);
			}
		}
	}

	/*
	 * If we have no parent dependancies we can clear CONN_DOWN
	 */
	if (TAILQ_EMPTY(&ip->target_list))
		ip->flags &= ~HAMMER_INODE_CONN_DOWN;

	/*
	 * If the inode is now clean drop the space reservation.
	 */
	if ((ip->flags & HAMMER_INODE_MODMASK) == 0 &&
	    (ip->flags & HAMMER_INODE_RSV_INODES)) {
		ip->flags &= ~HAMMER_INODE_RSV_INODES;
		--hmp->rsv_inodes;
	}

	if (dorel)
		hammer_rel_inode(ip, 0);
}

/*
 * Called from hammer_sync_inode() to synchronize in-memory records
 * to the media.
 */
static int
hammer_sync_record_callback(hammer_record_t record, void *data)
{
	hammer_cursor_t cursor = data;
	hammer_transaction_t trans = cursor->trans;
	hammer_mount_t hmp = trans->hmp;
	int error;

	/*
	 * Skip records that do not belong to the current flush.
	 */
	++hammer_stats_record_iterations;
	if (record->flush_state != HAMMER_FST_FLUSH)
		return(0);

#if 1
	if (record->flush_group != record->ip->flush_group) {
		kprintf("sync_record %p ip %p bad flush group %p %p\n", record, record->ip, record->flush_group ,record->ip->flush_group);
		if (hammer_debug_critical)
			Debugger("blah2");
		return(0);
	}
#endif
	KKASSERT(record->flush_group == record->ip->flush_group);

	/*
	 * Interlock the record using the BE flag.  Once BE is set the
	 * frontend cannot change the state of FE.
	 *
	 * NOTE: If FE is set prior to us setting BE we still sync the
	 * record out, but the flush completion code converts it to 
	 * a delete-on-disk record instead of destroying it.
	 */
	KKASSERT((record->flags & HAMMER_RECF_INTERLOCK_BE) == 0);
	record->flags |= HAMMER_RECF_INTERLOCK_BE;

	/*
	 * The backend has already disposed of the record.
	 */
	if (record->flags & (HAMMER_RECF_DELETED_BE | HAMMER_RECF_COMMITTED)) {
		error = 0;
		goto done;
	}

	/*
	 * If the whole inode is being deleting all on-disk records will
	 * be deleted very soon, we can't sync any new records to disk
	 * because they will be deleted in the same transaction they were
	 * created in (delete_tid == create_tid), which will assert.
	 *
	 * XXX There may be a case with RECORD_ADD with DELETED_FE set
	 * that we currently panic on.
	 */
	if (record->ip->sync_flags & HAMMER_INODE_DELETING) {
		switch(record->type) {
		case HAMMER_MEM_RECORD_DATA:
			/*
			 * We don't have to do anything, if the record was
			 * committed the space will have been accounted for
			 * in the blockmap.
			 */
			/* fall through */
		case HAMMER_MEM_RECORD_GENERAL:
			/*
			 * Set deleted-by-backend flag.  Do not set the
			 * backend committed flag, because we are throwing
			 * the record away.
			 */
			record->flags |= HAMMER_RECF_DELETED_BE;
			++record->ip->rec_generation;
			error = 0;
			goto done;
		case HAMMER_MEM_RECORD_ADD:
			panic("hammer_sync_record_callback: illegal add "
			      "during inode deletion record %p", record);
			break; /* NOT REACHED */
		case HAMMER_MEM_RECORD_INODE:
			panic("hammer_sync_record_callback: attempt to "
			      "sync inode record %p?", record);
			break; /* NOT REACHED */
		case HAMMER_MEM_RECORD_DEL:
			/* 
			 * Follow through and issue the on-disk deletion
			 */
			break;
		}
	}

	/*
	 * If DELETED_FE is set special handling is needed for directory
	 * entries.  Dependant pieces related to the directory entry may
	 * have already been synced to disk.  If this occurs we have to
	 * sync the directory entry and then change the in-memory record
	 * from an ADD to a DELETE to cover the fact that it's been
	 * deleted by the frontend.
	 *
	 * A directory delete covering record (MEM_RECORD_DEL) can never
	 * be deleted by the frontend.
	 *
	 * Any other record type (aka DATA) can be deleted by the frontend.
	 * XXX At the moment the flusher must skip it because there may
	 * be another data record in the flush group for the same block,
	 * meaning that some frontend data changes can leak into the backend's
	 * synchronization point.
	 */
	if (record->flags & HAMMER_RECF_DELETED_FE) {
		if (record->type == HAMMER_MEM_RECORD_ADD) {
			/*
			 * Convert a front-end deleted directory-add to
			 * a directory-delete entry later.
			 */
			record->flags |= HAMMER_RECF_CONVERT_DELETE;
		} else {
			/*
			 * Dispose of the record (race case).  Mark as
			 * deleted by backend (and not committed).
			 */
			KKASSERT(record->type != HAMMER_MEM_RECORD_DEL);
			record->flags |= HAMMER_RECF_DELETED_BE;
			++record->ip->rec_generation;
			error = 0;
			goto done;
		}
	}

	/*
	 * Assign the create_tid for new records.  Deletions already
	 * have the record's entire key properly set up.
	 */
	if (record->type != HAMMER_MEM_RECORD_DEL) {
		record->leaf.base.create_tid = trans->tid;
		record->leaf.create_ts = trans->time32;
	}

	/*
	 * This actually moves the record to the on-media B-Tree.  We
	 * must also generate REDO_TERM entries in the UNDO/REDO FIFO
	 * indicating that the related REDO_WRITE(s) have been committed.
	 *
	 * During recovery any REDO_TERM's within the nominal recovery span
	 * are ignored since the related meta-data is being undone, causing
	 * any matching REDO_WRITEs to execute.  The REDO_TERMs outside
	 * the nominal recovery span will match against REDO_WRITEs and
	 * prevent them from being executed (because the meta-data has
	 * already been synchronized).
	 */
	if (record->flags & HAMMER_RECF_REDO) {
		KKASSERT(record->type == HAMMER_MEM_RECORD_DATA);
		hammer_generate_redo(trans, record->ip,
				     record->leaf.base.key -
					 record->leaf.data_len,
				     HAMMER_REDO_TERM_WRITE,
				     NULL,
				     record->leaf.data_len);
	}
	for (;;) {
		error = hammer_ip_sync_record_cursor(cursor, record);
		if (error != EDEADLK)
			break;
		hammer_done_cursor(cursor);
		error = hammer_init_cursor(trans, cursor, &record->ip->cache[0],
					   record->ip);
		if (error)
			break;
	}
	record->flags &= ~HAMMER_RECF_CONVERT_DELETE;

	if (error)
		error = -error;
done:
	hammer_flush_record_done(record, error);

	/*
	 * Do partial finalization if we have built up too many dirty
	 * buffers.  Otherwise a buffer cache deadlock can occur when
	 * doing things like creating tens of thousands of tiny files.
	 *
	 * We must release our cursor lock to avoid a 3-way deadlock
	 * due to the exclusive sync lock the finalizer must get.
	 *
	 * WARNING: See warnings in hammer_unlock_cursor() function.
	 */
        if (hammer_flusher_meta_limit(hmp)) {
		hammer_unlock_cursor(cursor);
                hammer_flusher_finalize(trans, 0);
		hammer_lock_cursor(cursor);
	}

	return(error);
}

/*
 * Backend function called by the flusher to sync an inode to media.
 */
int
hammer_sync_inode(hammer_transaction_t trans, hammer_inode_t ip)
{
	struct hammer_cursor cursor;
	hammer_node_t tmp_node;
	hammer_record_t depend;
	hammer_record_t next;
	int error, tmp_error;
	u_int64_t nlinks;

	if ((ip->sync_flags & HAMMER_INODE_MODMASK) == 0)
		return(0);

	error = hammer_init_cursor(trans, &cursor, &ip->cache[1], ip);
	if (error)
		goto done;

	/*
	 * Any directory records referencing this inode which are not in
	 * our current flush group must adjust our nlink count for the
	 * purposes of synchronizating to disk.
	 *
	 * Records which are in our flush group can be unlinked from our
	 * inode now, potentially allowing the inode to be physically
	 * deleted.
	 *
	 * This cannot block.
	 */
	nlinks = ip->ino_data.nlinks;
	next = TAILQ_FIRST(&ip->target_list);
	while ((depend = next) != NULL) {
		next = TAILQ_NEXT(depend, target_entry);
		if (depend->flush_state == HAMMER_FST_FLUSH &&
		    depend->flush_group == ip->flush_group) {
			/*
			 * If this is an ADD that was deleted by the frontend
			 * the frontend nlinks count will have already been
			 * decremented, but the backend is going to sync its
			 * directory entry and must account for it.  The
			 * record will be converted to a delete-on-disk when
			 * it gets synced.
			 *
			 * If the ADD was not deleted by the frontend we
			 * can remove the dependancy from our target_list.
			 */
			if (depend->flags & HAMMER_RECF_DELETED_FE) {
				++nlinks;
			} else {
				TAILQ_REMOVE(&ip->target_list, depend,
					     target_entry);
				depend->target_ip = NULL;
			}
		} else if ((depend->flags & HAMMER_RECF_DELETED_FE) == 0) {
			/*
			 * Not part of our flush group and not deleted by
			 * the front-end, adjust the link count synced to
			 * the media (undo what the frontend did when it
			 * queued the record).
			 */
			KKASSERT((depend->flags & HAMMER_RECF_DELETED_BE) == 0);
			switch(depend->type) {
			case HAMMER_MEM_RECORD_ADD:
				--nlinks;
				break;
			case HAMMER_MEM_RECORD_DEL:
				++nlinks;
				break;
			default:
				break;
			}
		}
	}

	/*
	 * Set dirty if we had to modify the link count.
	 */
	if (ip->sync_ino_data.nlinks != nlinks) {
		KKASSERT((int64_t)nlinks >= 0);
		ip->sync_ino_data.nlinks = nlinks;
		ip->sync_flags |= HAMMER_INODE_DDIRTY;
	}

	/*
	 * If there is a trunction queued destroy any data past the (aligned)
	 * truncation point.  Userland will have dealt with the buffer
	 * containing the truncation point for us.
	 *
	 * We don't flush pending frontend data buffers until after we've
	 * dealt with the truncation.
	 */
	if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
		/*
		 * Interlock trunc_off.  The VOP front-end may continue to
		 * make adjustments to it while we are blocked.
		 */
		off_t trunc_off;
		off_t aligned_trunc_off;
		int blkmask;

		trunc_off = ip->sync_trunc_off;
		blkmask = hammer_blocksize(trunc_off) - 1;
		aligned_trunc_off = (trunc_off + blkmask) & ~(int64_t)blkmask;

		/*
		 * Delete any whole blocks on-media.  The front-end has
		 * already cleaned out any partial block and made it
		 * pending.  The front-end may have updated trunc_off
		 * while we were blocked so we only use sync_trunc_off.
		 *
		 * This operation can blow out the buffer cache, EWOULDBLOCK
		 * means we were unable to complete the deletion.  The
		 * deletion will update sync_trunc_off in that case.
		 */
		error = hammer_ip_delete_range(&cursor, ip,
						aligned_trunc_off,
						0x7FFFFFFFFFFFFFFFLL, 2);
		if (error == EWOULDBLOCK) {
			ip->flags |= HAMMER_INODE_WOULDBLOCK;
			error = 0;
			goto defer_buffer_flush;
		}

		if (error)
			goto done;

		/*
		 * Generate a REDO_TERM_TRUNC entry in the UNDO/REDO FIFO.
		 *
		 * XXX we do this even if we did not previously generate
		 * a REDO_TRUNC record.  This operation may enclosed the
		 * range for multiple prior truncation entries in the REDO
		 * log.
		 */
		if (trans->hmp->version >= HAMMER_VOL_VERSION_FOUR &&
		    (ip->flags & HAMMER_INODE_RDIRTY)) {
			hammer_generate_redo(trans, ip, aligned_trunc_off,
					     HAMMER_REDO_TERM_TRUNC,
					     NULL, 0);
		}

		/*
		 * Clear the truncation flag on the backend after we have
		 * completed the deletions.  Backend data is now good again
		 * (including new records we are about to sync, below).
		 *
		 * Leave sync_trunc_off intact.  As we write additional
		 * records the backend will update sync_trunc_off.  This
		 * tells the backend whether it can skip the overwrite
		 * test.  This should work properly even when the backend
		 * writes full blocks where the truncation point straddles
		 * the block because the comparison is against the base
		 * offset of the record.
		 */
		ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
		/* ip->sync_trunc_off = 0x7FFFFFFFFFFFFFFFLL; */
	} else {
		error = 0;
	}

	/*
	 * Now sync related records.  These will typically be directory
	 * entries, records tracking direct-writes, or delete-on-disk records.
	 */
	if (error == 0) {
		tmp_error = RB_SCAN(hammer_rec_rb_tree, &ip->rec_tree, NULL,
				    hammer_sync_record_callback, &cursor);
		if (tmp_error < 0)
			tmp_error = -error;
		if (tmp_error)
			error = tmp_error;
	}
	hammer_cache_node(&ip->cache[1], cursor.node);

	/*
	 * Re-seek for inode update, assuming our cache hasn't been ripped
	 * out from under us.
	 */
	if (error == 0) {
		tmp_node = hammer_ref_node_safe(trans, &ip->cache[0], &error);
		if (tmp_node) {
			hammer_cursor_downgrade(&cursor);
			hammer_lock_sh(&tmp_node->lock);
			if ((tmp_node->flags & HAMMER_NODE_DELETED) == 0)
				hammer_cursor_seek(&cursor, tmp_node, 0);
			hammer_unlock(&tmp_node->lock);
			hammer_rel_node(tmp_node);
		}
		error = 0;
	}

	/*
	 * If we are deleting the inode the frontend had better not have
	 * any active references on elements making up the inode.
	 *
	 * The call to hammer_ip_delete_clean() cleans up auxillary records
	 * but not DB or DATA records.  Those must have already been deleted
	 * by the normal truncation mechanic.
	 */
	if (error == 0 && ip->sync_ino_data.nlinks == 0 &&
		RB_EMPTY(&ip->rec_tree)  &&
	    (ip->sync_flags & HAMMER_INODE_DELETING) &&
	    (ip->flags & HAMMER_INODE_DELETED) == 0) {
		int count1 = 0;

		error = hammer_ip_delete_clean(&cursor, ip, &count1);
		if (error == 0) {
			ip->flags |= HAMMER_INODE_DELETED;
			ip->sync_flags &= ~HAMMER_INODE_DELETING;
			ip->sync_flags &= ~HAMMER_INODE_TRUNCATED;
			KKASSERT(RB_EMPTY(&ip->rec_tree));

			/*
			 * Set delete_tid in both the frontend and backend
			 * copy of the inode record.  The DELETED flag handles
			 * this, do not set DDIRTY.
			 */
			ip->ino_leaf.base.delete_tid = trans->tid;
			ip->sync_ino_leaf.base.delete_tid = trans->tid;
			ip->ino_leaf.delete_ts = trans->time32;
			ip->sync_ino_leaf.delete_ts = trans->time32;


			/*
			 * Adjust the inode count in the volume header
			 */
			hammer_sync_lock_sh(trans);
			if (ip->flags & HAMMER_INODE_ONDISK) {
				hammer_modify_volume_field(trans,
							   trans->rootvol,
							   vol0_stat_inodes);
				--ip->hmp->rootvol->ondisk->vol0_stat_inodes;
				hammer_modify_volume_done(trans->rootvol);
			}
			hammer_sync_unlock(trans);
		}
	}

	if (error)
		goto done;
	ip->sync_flags &= ~HAMMER_INODE_BUFS;

defer_buffer_flush:
	/*
	 * Now update the inode's on-disk inode-data and/or on-disk record.
	 * DELETED and ONDISK are managed only in ip->flags.
	 *
	 * In the case of a defered buffer flush we still update the on-disk
	 * inode to satisfy visibility requirements if there happen to be
	 * directory dependancies.
	 */
	switch(ip->flags & (HAMMER_INODE_DELETED | HAMMER_INODE_ONDISK)) {
	case HAMMER_INODE_DELETED|HAMMER_INODE_ONDISK:
		/*
		 * If deleted and on-disk, don't set any additional flags.
		 * the delete flag takes care of things.
		 *
		 * Clear flags which may have been set by the frontend.
		 */
		ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
				    HAMMER_INODE_SDIRTY |
				    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
				    HAMMER_INODE_DELETING);
		break;
	case HAMMER_INODE_DELETED:
		/*
		 * Take care of the case where a deleted inode was never
		 * flushed to the disk in the first place.
		 *
		 * Clear flags which may have been set by the frontend.
		 */
		ip->sync_flags &= ~(HAMMER_INODE_DDIRTY | HAMMER_INODE_XDIRTY |
				    HAMMER_INODE_SDIRTY |
				    HAMMER_INODE_ATIME | HAMMER_INODE_MTIME |
				    HAMMER_INODE_DELETING);
		while (RB_ROOT(&ip->rec_tree)) {
			hammer_record_t record = RB_ROOT(&ip->rec_tree);
			hammer_ref(&record->lock);
			KKASSERT(hammer_oneref(&record->lock));
			record->flags |= HAMMER_RECF_DELETED_BE;
			++record->ip->rec_generation;
			hammer_rel_mem_record(record);
		}
		break;
	case HAMMER_INODE_ONDISK:
		/*
		 * If already on-disk, do not set any additional flags.
		 */
		break;
	default:
		/*
		 * If not on-disk and not deleted, set DDIRTY to force
		 * an initial record to be written.
		 *
		 * Also set the create_tid in both the frontend and backend
		 * copy of the inode record.
		 */
		ip->ino_leaf.base.create_tid = trans->tid;
		ip->ino_leaf.create_ts = trans->time32;
		ip->sync_ino_leaf.base.create_tid = trans->tid;
		ip->sync_ino_leaf.create_ts = trans->time32;
		ip->sync_flags |= HAMMER_INODE_DDIRTY;
		break;
	}

	/*
	 * If DDIRTY or SDIRTY is set, write out a new record.
	 * If the inode is already on-disk the old record is marked as
	 * deleted.
	 *
	 * If DELETED is set hammer_update_inode() will delete the existing
	 * record without writing out a new one.
	 *
	 * If *ONLY* the ITIMES flag is set we can update the record in-place.
	 */
	if (ip->flags & HAMMER_INODE_DELETED) {
		error = hammer_update_inode(&cursor, ip);
	} else 
	if (!(ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY)) &&
	    (ip->sync_flags & (HAMMER_INODE_ATIME | HAMMER_INODE_MTIME))) {
		error = hammer_update_itimes(&cursor, ip);
	} else
	if (ip->sync_flags & (HAMMER_INODE_DDIRTY | HAMMER_INODE_SDIRTY |
			      HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
		error = hammer_update_inode(&cursor, ip);
	}
done:
	if (error) {
		hammer_critical_error(ip->hmp, ip, error,
				      "while syncing inode");
	}
	hammer_done_cursor(&cursor);
	return(error);
}

/*
 * This routine is called when the OS is no longer actively referencing
 * the inode (but might still be keeping it cached), or when releasing
 * the last reference to an inode.
 *
 * At this point if the inode's nlinks count is zero we want to destroy
 * it, which may mean destroying it on-media too.
 */
void
hammer_inode_unloadable_check(hammer_inode_t ip, int getvp)
{
	struct vnode *vp;

	/*
	 * Set the DELETING flag when the link count drops to 0 and the
	 * OS no longer has any opens on the inode.
	 *
	 * The backend will clear DELETING (a mod flag) and set DELETED
	 * (a state flag) when it is actually able to perform the
	 * operation.
	 *
	 * Don't reflag the deletion if the flusher is currently syncing
	 * one that was already flagged.  A previously set DELETING flag
	 * may bounce around flags and sync_flags until the operation is
	 * completely done.
	 *
	 * Do not attempt to modify a snapshot inode (one set to read-only).
	 */
	if (ip->ino_data.nlinks == 0 &&
	    ((ip->flags | ip->sync_flags) & (HAMMER_INODE_RO|HAMMER_INODE_DELETING|HAMMER_INODE_DELETED)) == 0) {
		ip->flags |= HAMMER_INODE_DELETING;
		ip->flags |= HAMMER_INODE_TRUNCATED;
		ip->trunc_off = 0;
		vp = NULL;
		if (getvp) {
			if (hammer_get_vnode(ip, &vp) != 0)
				return;
		}

		/*
		 * Final cleanup
		 */
		if (ip->vp)
			nvtruncbuf(ip->vp, 0, HAMMER_BUFSIZE, 0);
		if (getvp)
			vput(vp);
	}
}

/*
 * After potentially resolving a dependancy the inode is tested
 * to determine whether it needs to be reflushed.
 */
void
hammer_test_inode(hammer_inode_t ip)
{
	if (ip->flags & HAMMER_INODE_REFLUSH) {
		ip->flags &= ~HAMMER_INODE_REFLUSH;
		hammer_ref(&ip->lock);
		if (ip->flags & HAMMER_INODE_RESIGNAL) {
			ip->flags &= ~HAMMER_INODE_RESIGNAL;
			hammer_flush_inode(ip, HAMMER_FLUSH_SIGNAL);
		} else {
			hammer_flush_inode(ip, 0);
		}
		hammer_rel_inode(ip, 0);
	}
}

/*
 * Clear the RECLAIM flag on an inode.  This occurs when the inode is
 * reassociated with a vp or just before it gets freed.
 *
 * Pipeline wakeups to threads blocked due to an excessive number of
 * detached inodes.  This typically occurs when atime updates accumulate
 * while scanning a directory tree.
 */
static void
hammer_inode_wakereclaims(hammer_inode_t ip)
{
	struct hammer_reclaim *reclaim;
	hammer_mount_t hmp = ip->hmp;

	if ((ip->flags & HAMMER_INODE_RECLAIM) == 0)
		return;

	--hammer_count_reclaiming;
	--hmp->inode_reclaims;
	ip->flags &= ~HAMMER_INODE_RECLAIM;

	if ((reclaim = TAILQ_FIRST(&hmp->reclaim_list)) != NULL) {
		KKASSERT(reclaim->count > 0);
		if (--reclaim->count == 0) {
			TAILQ_REMOVE(&hmp->reclaim_list, reclaim, entry);
			wakeup(reclaim);
		}
	}
}

/*
 * Setup our reclaim pipeline.  We only let so many detached (and dirty)
 * inodes build up before we start blocking.  This routine is called
 * if a new inode is created or an inode is loaded from media.
 *
 * When we block we don't care *which* inode has finished reclaiming,
 * as lone as one does.
 *
 * The reclaim pipeline is primary governed by the auto-flush which is
 * 1/4 hammer_limit_reclaim.  We don't want to block if the count is
 * less than 1/2 hammer_limit_reclaim.  From 1/2 to full count is
 * dynamically governed.
 */
void
hammer_inode_waitreclaims(hammer_transaction_t trans)
{
	hammer_mount_t hmp = trans->hmp;
	struct hammer_reclaim reclaim;
	int lower_limit;

	/*
	 * Track inode load, delay if the number of reclaiming inodes is
	 * between 2/4 and 4/4 hammer_limit_reclaim, depending.
	 */
	if (curthread->td_proc) {
		struct hammer_inostats *stats;

		stats = hammer_inode_inostats(hmp, curthread->td_proc->p_pid);
		++stats->count;

		if (stats->count > hammer_limit_reclaim / 2)
			stats->count = hammer_limit_reclaim / 2;
		lower_limit = hammer_limit_reclaim - stats->count;
		if (hammer_debug_general & 0x10000) {
			kprintf("pid %5d limit %d\n",
				(int)curthread->td_proc->p_pid, lower_limit);
		}
	} else {
		lower_limit = hammer_limit_reclaim * 3 / 4;
	}
	if (hmp->inode_reclaims >= lower_limit) {
		reclaim.count = 1;
		TAILQ_INSERT_TAIL(&hmp->reclaim_list, &reclaim, entry);
		tsleep(&reclaim, 0, "hmrrcm", hz);
		if (reclaim.count > 0)
			TAILQ_REMOVE(&hmp->reclaim_list, &reclaim, entry);
	}
}

/*
 * Keep track of reclaim statistics on a per-pid basis using a loose
 * 4-way set associative hash table.  Collisions inherit the count of
 * the previous entry.
 *
 * NOTE: We want to be careful here to limit the chain size.  If the chain
 *	 size is too large a pid will spread its stats out over too many
 *	 entries under certain types of heavy filesystem activity and
 *	 wind up not delaying long enough.
 */
static
struct hammer_inostats *
hammer_inode_inostats(hammer_mount_t hmp, pid_t pid)
{
	struct hammer_inostats *stats;
	int delta;
	int chain;
	static volatile int iterator;	/* we don't care about MP races */

	/*
	 * Chain up to 4 times to find our entry.
	 */
	for (chain = 0; chain < 4; ++chain) {
		stats = &hmp->inostats[(pid + chain) & HAMMER_INOSTATS_HMASK];
		if (stats->pid == pid)
			break;
	}

	/*
	 * Replace one of the four chaining entries with our new entry.
	 */
	if (chain == 4) {
		stats = &hmp->inostats[(pid + (iterator++ & 3)) &
				       HAMMER_INOSTATS_HMASK];
		stats->pid = pid;
	}

	/*
	 * Decay the entry
	 */
	if (stats->count && stats->ltick != ticks) {
		delta = ticks - stats->ltick;
		stats->ltick = ticks;
		if (delta <= 0 || delta > hz * 60)
			stats->count = 0;
		else
			stats->count = stats->count * hz / (hz + delta);
	}
	if (hammer_debug_general & 0x10000)
		kprintf("pid %5d stats %d\n", (int)pid, stats->count);
	return (stats);
}

#if 0

/*
 * XXX not used, doesn't work very well due to the large batching nature
 * of flushes.
 *
 * A larger then normal backlog of inodes is sitting in the flusher,
 * enforce a general slowdown to let it catch up.  This routine is only
 * called on completion of a non-flusher-related transaction which
 * performed B-Tree node I/O.
 *
 * It is possible for the flusher to stall in a continuous load.
 * blogbench -i1000 -o seems to do a good job generating this sort of load.
 * If the flusher is unable to catch up the inode count can bloat until
 * we run out of kvm.
 *
 * This is a bit of a hack.
 */
void
hammer_inode_waithard(hammer_mount_t hmp)
{
	/*
	 * Hysteresis.
	 */
	if (hmp->flags & HAMMER_MOUNT_FLUSH_RECOVERY) {
		if (hmp->inode_reclaims < hammer_limit_reclaim / 2 &&
		    hmp->count_iqueued < hmp->count_inodes / 20) {
			hmp->flags &= ~HAMMER_MOUNT_FLUSH_RECOVERY;
			return;
		}
	} else {
		if (hmp->inode_reclaims < hammer_limit_reclaim ||
		    hmp->count_iqueued < hmp->count_inodes / 10) {
			return;
		}
		hmp->flags |= HAMMER_MOUNT_FLUSH_RECOVERY;
	}

	/*
	 * Block for one flush cycle.
	 */
	hammer_flusher_wait_next(hmp);
}

#endif