kernel - Do a better job with the filesystem background sync

[dragonfly.git] / sys / vfs / ufs / ffs_softdep.c

/*
 * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
 *
 * The soft updates code is derived from the appendix of a University
 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
 * "Soft Updates: A Solution to the Metadata Update Problem in File
 * Systems", CSE-TR-254-95, August 1995).
 *
 * Further information about soft updates can be obtained from:
 *
 *	Marshall Kirk McKusick		http://www.mckusick.com/softdep/
 *	1614 Oxford Street		mckusick@mckusick.com
 *	Berkeley, CA 94709-1608		+1-510-843-9542
 *	USA
 *
 * 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.
 *
 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``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 MARSHALL KIRK MCKUSICK 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.
 *
 *	from: @(#)ffs_softdep.c	9.59 (McKusick) 6/21/00
 * $FreeBSD: src/sys/ufs/ffs/ffs_softdep.c,v 1.57.2.11 2002/02/05 18:46:53 dillon Exp $
 * $DragonFly: src/sys/vfs/ufs/ffs_softdep.c,v 1.57 2008/06/28 17:59:51 dillon Exp $
 */

/*
 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
 */
#ifndef DIAGNOSTIC
#define DIAGNOSTIC
#endif
#ifndef DEBUG
#define DEBUG
#endif

#include <sys/param.h>
#include <sys/kernel.h>
#include <sys/systm.h>
#include <sys/buf.h>
#include <sys/malloc.h>
#include <sys/mount.h>
#include <sys/proc.h>
#include <sys/syslog.h>
#include <sys/vnode.h>
#include <sys/conf.h>
#include <machine/inttypes.h>
#include "dir.h"
#include "quota.h"
#include "inode.h"
#include "ufsmount.h"
#include "fs.h"
#include "softdep.h"
#include "ffs_extern.h"
#include "ufs_extern.h"

#include <sys/buf2.h>
#include <sys/mplock2.h>
#include <sys/thread2.h>

/*
 * These definitions need to be adapted to the system to which
 * this file is being ported.
 */
/*
 * malloc types defined for the softdep system.
 */
MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");

#define M_SOFTDEP_FLAGS		(M_WAITOK | M_USE_RESERVE)

#define	D_PAGEDEP	0
#define	D_INODEDEP	1
#define	D_NEWBLK	2
#define	D_BMSAFEMAP	3
#define	D_ALLOCDIRECT	4
#define	D_INDIRDEP	5
#define	D_ALLOCINDIR	6
#define	D_FREEFRAG	7
#define	D_FREEBLKS	8
#define	D_FREEFILE	9
#define	D_DIRADD	10
#define	D_MKDIR		11
#define	D_DIRREM	12
#define D_LAST		D_DIRREM

/* 
 * translate from workitem type to memory type
 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
 */
static struct malloc_type *memtype[] = {
	M_PAGEDEP,
	M_INODEDEP,
	M_NEWBLK,
	M_BMSAFEMAP,
	M_ALLOCDIRECT,
	M_INDIRDEP,
	M_ALLOCINDIR,
	M_FREEFRAG,
	M_FREEBLKS,
	M_FREEFILE,
	M_DIRADD,
	M_MKDIR,
	M_DIRREM
};

#define DtoM(type) (memtype[type])

/*
 * Names of malloc types.
 */
#define TYPENAME(type)  \
	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
/*
 * End system adaptaion definitions.
 */

/*
 * Internal function prototypes.
 */
static	void softdep_error(char *, int);
static	void drain_output(struct vnode *, int);
static	int getdirtybuf(struct buf **, int);
static	void clear_remove(struct thread *);
static	void clear_inodedeps(struct thread *);
static	int flush_pagedep_deps(struct vnode *, struct mount *,
	    struct diraddhd *);
static	int flush_inodedep_deps(struct fs *, ino_t);
static	int handle_written_filepage(struct pagedep *, struct buf *);
static  void diradd_inode_written(struct diradd *, struct inodedep *);
static	int handle_written_inodeblock(struct inodedep *, struct buf *);
static	void handle_allocdirect_partdone(struct allocdirect *);
static	void handle_allocindir_partdone(struct allocindir *);
static	void initiate_write_filepage(struct pagedep *, struct buf *);
static	void handle_written_mkdir(struct mkdir *, int);
static	void initiate_write_inodeblock(struct inodedep *, struct buf *);
static	void handle_workitem_freefile(struct freefile *);
static	void handle_workitem_remove(struct dirrem *);
static	struct dirrem *newdirrem(struct buf *, struct inode *,
	    struct inode *, int, struct dirrem **);
static	void free_diradd(struct diradd *);
static	void free_allocindir(struct allocindir *, struct inodedep *);
static	int indir_trunc (struct inode *, off_t, int, ufs_lbn_t, long *);
static	void deallocate_dependencies(struct buf *, struct inodedep *);
static	void free_allocdirect(struct allocdirectlst *,
	    struct allocdirect *, int);
static	int check_inode_unwritten(struct inodedep *);
static	int free_inodedep(struct inodedep *);
static	void handle_workitem_freeblocks(struct freeblks *);
static	void merge_inode_lists(struct inodedep *);
static	void setup_allocindir_phase2(struct buf *, struct inode *,
	    struct allocindir *);
static	struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t,
	    ufs_daddr_t);
static	void handle_workitem_freefrag(struct freefrag *);
static	struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long);
static	void allocdirect_merge(struct allocdirectlst *,
	    struct allocdirect *, struct allocdirect *);
static	struct bmsafemap *bmsafemap_lookup(struct buf *);
static	int newblk_lookup(struct fs *, ufs_daddr_t, int,
	    struct newblk **);
static	int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
static	int pagedep_lookup(struct inode *, ufs_lbn_t, int,
	    struct pagedep **);
static	void pause_timer(void *);
static	int request_cleanup(int, int);
static	int process_worklist_item(struct mount *, int);
static	void add_to_worklist(struct worklist *);

/*
 * Exported softdep operations.
 */
static	void softdep_disk_io_initiation(struct buf *);
static	void softdep_disk_write_complete(struct buf *);
static	void softdep_deallocate_dependencies(struct buf *);
static	int softdep_fsync(struct vnode *);
static	int softdep_process_worklist(struct mount *);
static	void softdep_move_dependencies(struct buf *, struct buf *);
static	int softdep_count_dependencies(struct buf *bp, int);
static  int softdep_checkread(struct buf *bp);
static  int softdep_checkwrite(struct buf *bp);

static struct bio_ops softdep_bioops = {
	.io_start = softdep_disk_io_initiation,
	.io_complete = softdep_disk_write_complete,
	.io_deallocate = softdep_deallocate_dependencies,
	.io_fsync = softdep_fsync,
	.io_sync = softdep_process_worklist,
	.io_movedeps = softdep_move_dependencies,
	.io_countdeps = softdep_count_dependencies,
	.io_checkread = softdep_checkread,
	.io_checkwrite = softdep_checkwrite
};

/*
 * Locking primitives.
 *
 * For a uniprocessor, all we need to do is protect against disk
 * interrupts. For a multiprocessor, this lock would have to be
 * a mutex. A single mutex is used throughout this file, though
 * finer grain locking could be used if contention warranted it.
 *
 * For a multiprocessor, the sleep call would accept a lock and
 * release it after the sleep processing was complete. In a uniprocessor
 * implementation there is no such interlock, so we simple mark
 * the places where it needs to be done with the `interlocked' form
 * of the lock calls. Since the uniprocessor sleep already interlocks
 * the spl, there is nothing that really needs to be done.
 */
#ifndef /* NOT */ DEBUG
static struct lockit {
} lk = { 0 };
#define ACQUIRE_LOCK(lk)		crit_enter_id("softupdates");
#define FREE_LOCK(lk)			crit_exit_id("softupdates");

#else /* DEBUG */
#define NOHOLDER	((struct thread *)-1)
#define SPECIAL_FLAG	((struct thread *)-2)
static struct lockit {
	int	lkt_spl;
	struct thread *lkt_held;
} lk = { 0, NOHOLDER };
static int lockcnt;

static	void acquire_lock(struct lockit *);
static	void free_lock(struct lockit *);
void	softdep_panic(char *);

#define ACQUIRE_LOCK(lk)		acquire_lock(lk)
#define FREE_LOCK(lk)			free_lock(lk)

static void
acquire_lock(struct lockit *lk)
{
	thread_t holder;

	if (lk->lkt_held != NOHOLDER) {
		holder = lk->lkt_held;
		FREE_LOCK(lk);
		if (holder == curthread)
			panic("softdep_lock: locking against myself");
		else
			panic("softdep_lock: lock held by %p", holder);
	}
	crit_enter_id("softupdates");
	lk->lkt_held = curthread;
	lockcnt++;
}

static void
free_lock(struct lockit *lk)
{

	if (lk->lkt_held == NOHOLDER)
		panic("softdep_unlock: lock not held");
	lk->lkt_held = NOHOLDER;
	crit_exit_id("softupdates");
}

/*
 * Function to release soft updates lock and panic.
 */
void
softdep_panic(char *msg)
{

	if (lk.lkt_held != NOHOLDER)
		FREE_LOCK(&lk);
	panic(msg);
}
#endif /* DEBUG */

static	int interlocked_sleep(struct lockit *, int, void *, int,
	    const char *, int);

/*
 * When going to sleep, we must save our SPL so that it does
 * not get lost if some other process uses the lock while we
 * are sleeping. We restore it after we have slept. This routine
 * wraps the interlocking with functions that sleep. The list
 * below enumerates the available set of operations.
 */
#define	UNKNOWN		0
#define	SLEEP		1
#define	LOCKBUF		2

static int
interlocked_sleep(struct lockit *lk, int op, void *ident, int flags,
		  const char *wmesg, int timo)
{
	thread_t holder;
	int s, retval;

	s = lk->lkt_spl;
#	ifdef DEBUG
	if (lk->lkt_held == NOHOLDER)
		panic("interlocked_sleep: lock not held");
	lk->lkt_held = NOHOLDER;
#	endif /* DEBUG */
	switch (op) {
	case SLEEP:
		retval = tsleep(ident, flags, wmesg, timo);
		break;
	case LOCKBUF:
		retval = BUF_LOCK((struct buf *)ident, flags);
		break;
	default:
		panic("interlocked_sleep: unknown operation");
	}
#	ifdef DEBUG
	if (lk->lkt_held != NOHOLDER) {
		holder = lk->lkt_held;
		FREE_LOCK(lk);
		if (holder == curthread)
			panic("interlocked_sleep: locking against self");
		else
			panic("interlocked_sleep: lock held by %p", holder);
	}
	lk->lkt_held = curthread;
	lockcnt++;
#	endif /* DEBUG */
	lk->lkt_spl = s;
	return (retval);
}

/*
 * Place holder for real semaphores.
 */
struct sema {
	int	value;
	thread_t holder;
	char	*name;
	int	prio;
	int	timo;
};
static	void sema_init(struct sema *, char *, int, int);
static	int sema_get(struct sema *, struct lockit *);
static	void sema_release(struct sema *);

static void
sema_init(struct sema *semap, char *name, int prio, int timo)
{

	semap->holder = NOHOLDER;
	semap->value = 0;
	semap->name = name;
	semap->prio = prio;
	semap->timo = timo;
}

static int
sema_get(struct sema *semap, struct lockit *interlock)
{

	if (semap->value++ > 0) {
		if (interlock != NULL) {
			interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
			    semap->prio, semap->name, semap->timo);
			FREE_LOCK(interlock);
		} else {
			tsleep((caddr_t)semap, semap->prio, semap->name,
			    semap->timo);
		}
		return (0);
	}
	semap->holder = curthread;
	if (interlock != NULL)
		FREE_LOCK(interlock);
	return (1);
}

static void
sema_release(struct sema *semap)
{

	if (semap->value <= 0 || semap->holder != curthread) {
		if (lk.lkt_held != NOHOLDER)
			FREE_LOCK(&lk);
		panic("sema_release: not held");
	}
	if (--semap->value > 0) {
		semap->value = 0;
		wakeup(semap);
	}
	semap->holder = NOHOLDER;
}

/*
 * Worklist queue management.
 * These routines require that the lock be held.
 */
#ifndef /* NOT */ DEBUG
#define WORKLIST_INSERT(head, item) do {	\
	(item)->wk_state |= ONWORKLIST;		\
	LIST_INSERT_HEAD(head, item, wk_list);	\
} while (0)

#define WORKLIST_INSERT_BP(bp, item) do {	\
	(item)->wk_state |= ONWORKLIST;		\
	(bp)->b_ops = &softdep_bioops;		\
	LIST_INSERT_HEAD(&(bp)->b_dep, item, wk_list);	\
} while (0)

#define WORKLIST_REMOVE(item) do {		\
	(item)->wk_state &= ~ONWORKLIST;	\
	LIST_REMOVE(item, wk_list);		\
} while (0)

#define WORKITEM_FREE(item, type) FREE(item, DtoM(type))

#else /* DEBUG */
static	void worklist_insert(struct workhead *, struct worklist *);
static	void worklist_remove(struct worklist *);
static	void workitem_free(struct worklist *, int);

#define WORKLIST_INSERT_BP(bp, item) do {	\
	(bp)->b_ops = &softdep_bioops;		\
	worklist_insert(&(bp)->b_dep, item);	\
} while (0)

#define WORKLIST_INSERT(head, item) worklist_insert(head, item)
#define WORKLIST_REMOVE(item) worklist_remove(item)
#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)

static void
worklist_insert(struct workhead *head, struct worklist *item)
{

	if (lk.lkt_held == NOHOLDER)
		panic("worklist_insert: lock not held");
	if (item->wk_state & ONWORKLIST) {
		FREE_LOCK(&lk);
		panic("worklist_insert: already on list");
	}
	item->wk_state |= ONWORKLIST;
	LIST_INSERT_HEAD(head, item, wk_list);
}

static void
worklist_remove(struct worklist *item)
{

	if (lk.lkt_held == NOHOLDER)
		panic("worklist_remove: lock not held");
	if ((item->wk_state & ONWORKLIST) == 0) {
		FREE_LOCK(&lk);
		panic("worklist_remove: not on list");
	}
	item->wk_state &= ~ONWORKLIST;
	LIST_REMOVE(item, wk_list);
}

static void
workitem_free(struct worklist *item, int type)
{

	if (item->wk_state & ONWORKLIST) {
		if (lk.lkt_held != NOHOLDER)
			FREE_LOCK(&lk);
		panic("workitem_free: still on list");
	}
	if (item->wk_type != type) {
		if (lk.lkt_held != NOHOLDER)
			FREE_LOCK(&lk);
		panic("workitem_free: type mismatch");
	}
	FREE(item, DtoM(type));
}
#endif /* DEBUG */

/*
 * Workitem queue management
 */
static struct workhead softdep_workitem_pending;
static int num_on_worklist;	/* number of worklist items to be processed */
static int softdep_worklist_busy; /* 1 => trying to do unmount */
static int softdep_worklist_req; /* serialized waiters */
static int max_softdeps;	/* maximum number of structs before slowdown */
static int tickdelay = 2;	/* number of ticks to pause during slowdown */
static int *stat_countp;	/* statistic to count in proc_waiting timeout */
static int proc_waiting;	/* tracks whether we have a timeout posted */
static struct callout handle; /* handle on posted proc_waiting timeout */
static struct thread *filesys_syncer; /* proc of filesystem syncer process */
static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
#define FLUSH_INODES	1
static int req_clear_remove;	/* syncer process flush some freeblks */
#define FLUSH_REMOVE	2
/*
 * runtime statistics
 */
static int stat_worklist_push;	/* number of worklist cleanups */
static int stat_blk_limit_push;	/* number of times block limit neared */
static int stat_ino_limit_push;	/* number of times inode limit neared */
static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
static int stat_sync_limit_hit;	/* number of synchronous slowdowns imposed */
static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
#ifdef DEBUG
#include <vm/vm.h>
#include <sys/sysctl.h>
SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0,
    "Maximum soft dependencies before slowdown occurs");
SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0,
    "Ticks to delay before allocating during slowdown");
SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,
    "Number of worklist cleanups");
SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,
    "Number of times block limit neared");
SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,
    "Number of times inode limit neared");
SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0,
    "Number of times block slowdown imposed");
SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0,
    "Number of times inode slowdown imposed ");
SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0,
    "Number of synchronous slowdowns imposed");
SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0,
    "Bufs redirtied as indir ptrs not written");
SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0,
    "Bufs redirtied as inode bitmap not written");
SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0,
    "Bufs redirtied as direct ptrs not written");
SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0,
    "Bufs redirtied as dir entry cannot write");
#endif /* DEBUG */

/*
 * Add an item to the end of the work queue.
 * This routine requires that the lock be held.
 * This is the only routine that adds items to the list.
 * The following routine is the only one that removes items
 * and does so in order from first to last.
 */
static void
add_to_worklist(struct worklist *wk)
{
	static struct worklist *worklist_tail;

	if (wk->wk_state & ONWORKLIST) {
		if (lk.lkt_held != NOHOLDER)
			FREE_LOCK(&lk);
		panic("add_to_worklist: already on list");
	}
	wk->wk_state |= ONWORKLIST;
	if (LIST_FIRST(&softdep_workitem_pending) == NULL)
		LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
	else
		LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
	worklist_tail = wk;
	num_on_worklist += 1;
}

/*
 * Process that runs once per second to handle items in the background queue.
 *
 * Note that we ensure that everything is done in the order in which they
 * appear in the queue. The code below depends on this property to ensure
 * that blocks of a file are freed before the inode itself is freed. This
 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
 * until all the old ones have been purged from the dependency lists.
 *
 * bioops callback - hold io_token
 */
static int 
softdep_process_worklist(struct mount *matchmnt)
{
	thread_t td = curthread;
	int matchcnt, loopcount;
	long starttime;

	get_mplock();

	/*
	 * Record the process identifier of our caller so that we can give
	 * this process preferential treatment in request_cleanup below.
	 */
	filesys_syncer = td;
	matchcnt = 0;

	/*
	 * There is no danger of having multiple processes run this
	 * code, but we have to single-thread it when softdep_flushfiles()
	 * is in operation to get an accurate count of the number of items
	 * related to its mount point that are in the list.
	 */
	if (matchmnt == NULL) {
		if (softdep_worklist_busy < 0) {
			matchcnt = -1;
			goto done;
		}
		softdep_worklist_busy += 1;
	}

	/*
	 * If requested, try removing inode or removal dependencies.
	 */
	if (req_clear_inodedeps) {
		clear_inodedeps(td);
		req_clear_inodedeps -= 1;
		wakeup_one(&proc_waiting);
	}
	if (req_clear_remove) {
		clear_remove(td);
		req_clear_remove -= 1;
		wakeup_one(&proc_waiting);
	}
	loopcount = 1;
	starttime = time_second;
	while (num_on_worklist > 0) {
		matchcnt += process_worklist_item(matchmnt, 0);

		/*
		 * If a umount operation wants to run the worklist
		 * accurately, abort.
		 */
		if (softdep_worklist_req && matchmnt == NULL) {
			matchcnt = -1;
			break;
		}

		/*
		 * If requested, try removing inode or removal dependencies.
		 */
		if (req_clear_inodedeps) {
			clear_inodedeps(td);
			req_clear_inodedeps -= 1;
			wakeup_one(&proc_waiting);
		}
		if (req_clear_remove) {
			clear_remove(td);
			req_clear_remove -= 1;
			wakeup_one(&proc_waiting);
		}
		/*
		 * We do not generally want to stop for buffer space, but if
		 * we are really being a buffer hog, we will stop and wait.
		 */
		if (loopcount++ % 128 == 0)
			bwillinode(1);
		/*
		 * Never allow processing to run for more than one
		 * second. Otherwise the other syncer tasks may get
		 * excessively backlogged.
		 */
		if (starttime != time_second && matchmnt == NULL) {
			matchcnt = -1;
			break;
		}
	}
	if (matchmnt == NULL) {
		--softdep_worklist_busy;
		if (softdep_worklist_req && softdep_worklist_busy == 0)
			wakeup(&softdep_worklist_req);
	}
done:
	rel_mplock();
	return (matchcnt);
}

/*
 * Process one item on the worklist.
 */
static int
process_worklist_item(struct mount *matchmnt, int flags)
{
	struct worklist *wk;
	struct dirrem *dirrem;
	struct fs *matchfs;
	struct vnode *vp;
	int matchcnt = 0;

	matchfs = NULL;
	if (matchmnt != NULL)
		matchfs = VFSTOUFS(matchmnt)->um_fs;
	ACQUIRE_LOCK(&lk);
	/*
	 * Normally we just process each item on the worklist in order.
	 * However, if we are in a situation where we cannot lock any
	 * inodes, we have to skip over any dirrem requests whose
	 * vnodes are resident and locked.
	 */
	LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
		if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
			break;
		dirrem = WK_DIRREM(wk);
		vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev,
		    dirrem->dm_oldinum);
		if (vp == NULL || !vn_islocked(vp))
			break;
	}
	if (wk == 0) {
		FREE_LOCK(&lk);
		return (0);
	}
	WORKLIST_REMOVE(wk);
	num_on_worklist -= 1;
	FREE_LOCK(&lk);
	switch (wk->wk_type) {

	case D_DIRREM:
		/* removal of a directory entry */
		if (WK_DIRREM(wk)->dm_mnt == matchmnt)
			matchcnt += 1;
		handle_workitem_remove(WK_DIRREM(wk));
		break;

	case D_FREEBLKS:
		/* releasing blocks and/or fragments from a file */
		if (WK_FREEBLKS(wk)->fb_fs == matchfs)
			matchcnt += 1;
		handle_workitem_freeblocks(WK_FREEBLKS(wk));
		break;

	case D_FREEFRAG:
		/* releasing a fragment when replaced as a file grows */
		if (WK_FREEFRAG(wk)->ff_fs == matchfs)
			matchcnt += 1;
		handle_workitem_freefrag(WK_FREEFRAG(wk));
		break;

	case D_FREEFILE:
		/* releasing an inode when its link count drops to 0 */
		if (WK_FREEFILE(wk)->fx_fs == matchfs)
			matchcnt += 1;
		handle_workitem_freefile(WK_FREEFILE(wk));
		break;

	default:
		panic("%s_process_worklist: Unknown type %s",
		    "softdep", TYPENAME(wk->wk_type));
		/* NOTREACHED */
	}
	return (matchcnt);
}

/*
 * Move dependencies from one buffer to another.
 *
 * bioops callback - hold io_token
 */
static void
softdep_move_dependencies(struct buf *oldbp, struct buf *newbp)
{
	struct worklist *wk, *wktail;

	get_mplock();
	if (LIST_FIRST(&newbp->b_dep) != NULL)
		panic("softdep_move_dependencies: need merge code");
	wktail = NULL;
	ACQUIRE_LOCK(&lk);
	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
		LIST_REMOVE(wk, wk_list);
		if (wktail == NULL)
			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
		else
			LIST_INSERT_AFTER(wktail, wk, wk_list);
		wktail = wk;
		newbp->b_ops = &softdep_bioops;
	}
	FREE_LOCK(&lk);
	rel_mplock();
}

/*
 * Purge the work list of all items associated with a particular mount point.
 */
int
softdep_flushfiles(struct mount *oldmnt, int flags)
{
	struct vnode *devvp;
	int error, loopcnt;

	/*
	 * Await our turn to clear out the queue, then serialize access.
	 */
	while (softdep_worklist_busy != 0) {
		softdep_worklist_req += 1;
		tsleep(&softdep_worklist_req, 0, "softflush", 0);
		softdep_worklist_req -= 1;
	}
	softdep_worklist_busy = -1;

	if ((error = ffs_flushfiles(oldmnt, flags)) != 0) {
		softdep_worklist_busy = 0;
		if (softdep_worklist_req)
			wakeup(&softdep_worklist_req);
		return (error);
	}
	/*
	 * Alternately flush the block device associated with the mount
	 * point and process any dependencies that the flushing
	 * creates. In theory, this loop can happen at most twice,
	 * but we give it a few extra just to be sure.
	 */
	devvp = VFSTOUFS(oldmnt)->um_devvp;
	for (loopcnt = 10; loopcnt > 0; ) {
		if (softdep_process_worklist(oldmnt) == 0) {
			loopcnt--;
			/*
			 * Do another flush in case any vnodes were brought in
			 * as part of the cleanup operations.
			 */
			if ((error = ffs_flushfiles(oldmnt, flags)) != 0)
				break;
			/*
			 * If we still found nothing to do, we are really done.
			 */
			if (softdep_process_worklist(oldmnt) == 0)
				break;
		}
		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
		error = VOP_FSYNC(devvp, MNT_WAIT, 0);
		vn_unlock(devvp);
		if (error)
			break;
	}
	softdep_worklist_busy = 0;
	if (softdep_worklist_req)
		wakeup(&softdep_worklist_req);

	/*
	 * If we are unmounting then it is an error to fail. If we
	 * are simply trying to downgrade to read-only, then filesystem
	 * activity can keep us busy forever, so we just fail with EBUSY.
	 */
	if (loopcnt == 0) {
		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
			panic("softdep_flushfiles: looping");
		error = EBUSY;
	}
	return (error);
}

/*
 * Structure hashing.
 * 
 * There are three types of structures that can be looked up:
 *	1) pagedep structures identified by mount point, inode number,
 *	   and logical block.
 *	2) inodedep structures identified by mount point and inode number.
 *	3) newblk structures identified by mount point and
 *	   physical block number.
 *
 * The "pagedep" and "inodedep" dependency structures are hashed
 * separately from the file blocks and inodes to which they correspond.
 * This separation helps when the in-memory copy of an inode or
 * file block must be replaced. It also obviates the need to access
 * an inode or file page when simply updating (or de-allocating)
 * dependency structures. Lookup of newblk structures is needed to
 * find newly allocated blocks when trying to associate them with
 * their allocdirect or allocindir structure.
 *
 * The lookup routines optionally create and hash a new instance when
 * an existing entry is not found.
 */
#define DEPALLOC	0x0001	/* allocate structure if lookup fails */
#define NODELAY		0x0002	/* cannot do background work */

/*
 * Structures and routines associated with pagedep caching.
 */
LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
u_long	pagedep_hash;		/* size of hash table - 1 */
#define	PAGEDEP_HASH(mp, inum, lbn) \
	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
	    pagedep_hash])
static struct sema pagedep_in_progress;

/*
 * Helper routine for pagedep_lookup()
 */
static __inline
struct pagedep *
pagedep_find(struct pagedep_hashhead *pagedephd, ino_t ino, ufs_lbn_t lbn,
	     struct mount *mp)
{
	struct pagedep *pagedep;

	LIST_FOREACH(pagedep, pagedephd, pd_hash) {
		if (ino == pagedep->pd_ino &&
		    lbn == pagedep->pd_lbn &&
		    mp == pagedep->pd_mnt) {
			return (pagedep);
		}
	}
	return(NULL);
}

/*
 * Look up a pagedep. Return 1 if found, 0 if not found.
 * If not found, allocate if DEPALLOC flag is passed.
 * Found or allocated entry is returned in pagedeppp.
 * This routine must be called with splbio interrupts blocked.
 */
static int
pagedep_lookup(struct inode *ip, ufs_lbn_t lbn, int flags,
	       struct pagedep **pagedeppp)
{
	struct pagedep *pagedep;
	struct pagedep_hashhead *pagedephd;
	struct mount *mp;
	int i;

#ifdef DEBUG
	if (lk.lkt_held == NOHOLDER)
		panic("pagedep_lookup: lock not held");
#endif
	mp = ITOV(ip)->v_mount;
	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
top:
	*pagedeppp = pagedep_find(pagedephd, ip->i_number, lbn, mp);
	if (*pagedeppp)
		return(1);
	if ((flags & DEPALLOC) == 0)
		return (0);
	if (sema_get(&pagedep_in_progress, &lk) == 0) {
		ACQUIRE_LOCK(&lk);
		goto top;
	}
	MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
		M_SOFTDEP_FLAGS | M_ZERO);

	if (pagedep_find(pagedephd, ip->i_number, lbn, mp)) {
		kprintf("pagedep_lookup: blocking race avoided\n");
		ACQUIRE_LOCK(&lk);
		sema_release(&pagedep_in_progress);
		kfree(pagedep, M_PAGEDEP);
		goto top;
	}

	pagedep->pd_list.wk_type = D_PAGEDEP;
	pagedep->pd_mnt = mp;
	pagedep->pd_ino = ip->i_number;
	pagedep->pd_lbn = lbn;
	LIST_INIT(&pagedep->pd_dirremhd);
	LIST_INIT(&pagedep->pd_pendinghd);
	for (i = 0; i < DAHASHSZ; i++)
		LIST_INIT(&pagedep->pd_diraddhd[i]);
	ACQUIRE_LOCK(&lk);
	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
	sema_release(&pagedep_in_progress);
	*pagedeppp = pagedep;
	return (0);
}

/*
 * Structures and routines associated with inodedep caching.
 */
LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
static u_long	inodedep_hash;	/* size of hash table - 1 */
static long	num_inodedep;	/* number of inodedep allocated */
#define	INODEDEP_HASH(fs, inum) \
      (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
static struct sema inodedep_in_progress;

/*
 * Helper routine for inodedep_lookup()
 */
static __inline
struct inodedep *
inodedep_find(struct inodedep_hashhead *inodedephd, struct fs *fs, ino_t inum)
{
	struct inodedep *inodedep;

	LIST_FOREACH(inodedep, inodedephd, id_hash) {
		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
			return(inodedep);
	}
	return (NULL);
}

/*
 * Look up a inodedep. Return 1 if found, 0 if not found.
 * If not found, allocate if DEPALLOC flag is passed.
 * Found or allocated entry is returned in inodedeppp.
 * This routine must be called with splbio interrupts blocked.
 */
static int
inodedep_lookup(struct fs *fs, ino_t inum, int flags,
	        struct inodedep **inodedeppp)
{
	struct inodedep *inodedep;
	struct inodedep_hashhead *inodedephd;
	int firsttry;

#ifdef DEBUG
	if (lk.lkt_held == NOHOLDER)
		panic("inodedep_lookup: lock not held");
#endif
	firsttry = 1;
	inodedephd = INODEDEP_HASH(fs, inum);
top:
	*inodedeppp = inodedep_find(inodedephd, fs, inum);
	if (*inodedeppp)
		return (1);
	if ((flags & DEPALLOC) == 0)
		return (0);
	/*
	 * If we are over our limit, try to improve the situation.
	 */
	if (num_inodedep > max_softdeps && firsttry && 
	    speedup_syncer() == 0 && (flags & NODELAY) == 0 &&
	    request_cleanup(FLUSH_INODES, 1)) {
		firsttry = 0;
		goto top;
	}
	if (sema_get(&inodedep_in_progress, &lk) == 0) {
		ACQUIRE_LOCK(&lk);
		goto top;
	}
	MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
		M_INODEDEP, M_SOFTDEP_FLAGS | M_ZERO);
	if (inodedep_find(inodedephd, fs, inum)) {
		kprintf("inodedep_lookup: blocking race avoided\n");
		ACQUIRE_LOCK(&lk);
		sema_release(&inodedep_in_progress);
		kfree(inodedep, M_INODEDEP);
		goto top;
	}
	inodedep->id_list.wk_type = D_INODEDEP;
	inodedep->id_fs = fs;
	inodedep->id_ino = inum;
	inodedep->id_state = ALLCOMPLETE;
	inodedep->id_nlinkdelta = 0;
	inodedep->id_savedino = NULL;
	inodedep->id_savedsize = -1;
	inodedep->id_buf = NULL;
	LIST_INIT(&inodedep->id_pendinghd);
	LIST_INIT(&inodedep->id_inowait);
	LIST_INIT(&inodedep->id_bufwait);
	TAILQ_INIT(&inodedep->id_inoupdt);
	TAILQ_INIT(&inodedep->id_newinoupdt);
	ACQUIRE_LOCK(&lk);
	num_inodedep += 1;
	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
	sema_release(&inodedep_in_progress);
	*inodedeppp = inodedep;
	return (0);
}

/*
 * Structures and routines associated with newblk caching.
 */
LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
u_long	newblk_hash;		/* size of hash table - 1 */
#define	NEWBLK_HASH(fs, inum) \
	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
static struct sema newblk_in_progress;

/*
 * Helper routine for newblk_lookup()
 */
static __inline
struct newblk *
newblk_find(struct newblk_hashhead *newblkhd, struct fs *fs, 
	    ufs_daddr_t newblkno)
{
	struct newblk *newblk;

	LIST_FOREACH(newblk, newblkhd, nb_hash) {
		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
			return (newblk);
	}
	return(NULL);
}

/*
 * Look up a newblk. Return 1 if found, 0 if not found.
 * If not found, allocate if DEPALLOC flag is passed.
 * Found or allocated entry is returned in newblkpp.
 */
static int
newblk_lookup(struct fs *fs, ufs_daddr_t newblkno, int flags,
	      struct newblk **newblkpp)
{
	struct newblk *newblk;
	struct newblk_hashhead *newblkhd;

	newblkhd = NEWBLK_HASH(fs, newblkno);
top:
	*newblkpp = newblk_find(newblkhd, fs, newblkno);
	if (*newblkpp)
		return(1);
	if ((flags & DEPALLOC) == 0)
		return (0);
	if (sema_get(&newblk_in_progress, 0) == 0)
		goto top;
	MALLOC(newblk, struct newblk *, sizeof(struct newblk),
		M_NEWBLK, M_SOFTDEP_FLAGS | M_ZERO);

	if (newblk_find(newblkhd, fs, newblkno)) {
		kprintf("newblk_lookup: blocking race avoided\n");
		sema_release(&pagedep_in_progress);
		kfree(newblk, M_NEWBLK);
		goto top;
	}
	newblk->nb_state = 0;
	newblk->nb_fs = fs;
	newblk->nb_newblkno = newblkno;
	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
	sema_release(&newblk_in_progress);
	*newblkpp = newblk;
	return (0);
}

/*
 * Executed during filesystem system initialization before
 * mounting any filesystems.
 */
void 
softdep_initialize(void)
{
	callout_init(&handle);

	LIST_INIT(&mkdirlisthd);
	LIST_INIT(&softdep_workitem_pending);
	max_softdeps = min(desiredvnodes * 8,
		M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep)));
	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
	    &pagedep_hash);
	sema_init(&pagedep_in_progress, "pagedep", 0, 0);
	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
	sema_init(&inodedep_in_progress, "inodedep", 0, 0);
	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
	sema_init(&newblk_in_progress, "newblk", 0, 0);
	add_bio_ops(&softdep_bioops);
}

/*
 * Called at mount time to notify the dependency code that a
 * filesystem wishes to use it.
 */
int
softdep_mount(struct vnode *devvp, struct mount *mp, struct fs *fs)
{
	struct csum cstotal;
	struct cg *cgp;
	struct buf *bp;
	int error, cyl;

	mp->mnt_flag &= ~MNT_ASYNC;
	mp->mnt_flag |= MNT_SOFTDEP;
	mp->mnt_bioops = &softdep_bioops;
	/*
	 * When doing soft updates, the counters in the
	 * superblock may have gotten out of sync, so we have
	 * to scan the cylinder groups and recalculate them.
	 */
	if (fs->fs_clean != 0)
		return (0);
	bzero(&cstotal, sizeof cstotal);
	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
		if ((error = bread(devvp, fsbtodoff(fs, cgtod(fs, cyl)),
				   fs->fs_cgsize, &bp)) != 0) {
			brelse(bp);
			return (error);
		}
		cgp = (struct cg *)bp->b_data;
		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
		fs->fs_cs(fs, cyl) = cgp->cg_cs;
		brelse(bp);
	}
#ifdef DEBUG
	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
		kprintf("ffs_mountfs: superblock updated for soft updates\n");
#endif
	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
	return (0);
}

/*
 * Protecting the freemaps (or bitmaps).
 * 
 * To eliminate the need to execute fsck before mounting a filesystem
 * after a power failure, one must (conservatively) guarantee that the
 * on-disk copy of the bitmaps never indicate that a live inode or block is
 * free.  So, when a block or inode is allocated, the bitmap should be
 * updated (on disk) before any new pointers.  When a block or inode is
 * freed, the bitmap should not be updated until all pointers have been
 * reset.  The latter dependency is handled by the delayed de-allocation
 * approach described below for block and inode de-allocation.  The former
 * dependency is handled by calling the following procedure when a block or
 * inode is allocated. When an inode is allocated an "inodedep" is created
 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
 * Each "inodedep" is also inserted into the hash indexing structure so
 * that any additional link additions can be made dependent on the inode
 * allocation.
 * 
 * The ufs filesystem maintains a number of free block counts (e.g., per
 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
 * in addition to the bitmaps.  These counts are used to improve efficiency
 * during allocation and therefore must be consistent with the bitmaps.
 * There is no convenient way to guarantee post-crash consistency of these
 * counts with simple update ordering, for two main reasons: (1) The counts
 * and bitmaps for a single cylinder group block are not in the same disk
 * sector.  If a disk write is interrupted (e.g., by power failure), one may
 * be written and the other not.  (2) Some of the counts are located in the
 * superblock rather than the cylinder group block. So, we focus our soft
 * updates implementation on protecting the bitmaps. When mounting a
 * filesystem, we recompute the auxiliary counts from the bitmaps.
 */

/*
 * Called just after updating the cylinder group block to allocate an inode.
 *
 * Parameters:
 *	bp:		buffer for cylgroup block with inode map
 *	ip:		inode related to allocation
 *	newinum:	new inode number being allocated
 */
void
softdep_setup_inomapdep(struct buf *bp, struct inode *ip, ino_t newinum)
{
	struct inodedep *inodedep;
	struct bmsafemap *bmsafemap;

	/*
	 * Create a dependency for the newly allocated inode.
	 * Panic if it already exists as something is seriously wrong.
	 * Otherwise add it to the dependency list for the buffer holding
	 * the cylinder group map from which it was allocated.
	 */
	ACQUIRE_LOCK(&lk);
	if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
		FREE_LOCK(&lk);
		panic("softdep_setup_inomapdep: found inode");
	}
	inodedep->id_buf = bp;
	inodedep->id_state &= ~DEPCOMPLETE;
	bmsafemap = bmsafemap_lookup(bp);
	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
	FREE_LOCK(&lk);
}

/*
 * Called just after updating the cylinder group block to
 * allocate block or fragment.
 *
 * Parameters:
 *	bp:		buffer for cylgroup block with block map
 *	fs:		filesystem doing allocation
 *	newblkno:	number of newly allocated block
 */
void
softdep_setup_blkmapdep(struct buf *bp, struct fs *fs,
			ufs_daddr_t newblkno)
{
	struct newblk *newblk;
	struct bmsafemap *bmsafemap;

	/*
	 * Create a dependency for the newly allocated block.
	 * Add it to the dependency list for the buffer holding
	 * the cylinder group map from which it was allocated.
	 */
	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
		panic("softdep_setup_blkmapdep: found block");
	ACQUIRE_LOCK(&lk);
	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
	FREE_LOCK(&lk);
}

/*
 * Find the bmsafemap associated with a cylinder group buffer.
 * If none exists, create one. The buffer must be locked when
 * this routine is called and this routine must be called with
 * splbio interrupts blocked.
 */
static struct bmsafemap *
bmsafemap_lookup(struct buf *bp)
{
	struct bmsafemap *bmsafemap;
	struct worklist *wk;

#ifdef DEBUG
	if (lk.lkt_held == NOHOLDER)
		panic("bmsafemap_lookup: lock not held");
#endif
	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
		if (wk->wk_type == D_BMSAFEMAP)
			return (WK_BMSAFEMAP(wk));
	}
	FREE_LOCK(&lk);
	MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
		M_BMSAFEMAP, M_SOFTDEP_FLAGS);
	bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
	bmsafemap->sm_list.wk_state = 0;
	bmsafemap->sm_buf = bp;
	LIST_INIT(&bmsafemap->sm_allocdirecthd);
	LIST_INIT(&bmsafemap->sm_allocindirhd);
	LIST_INIT(&bmsafemap->sm_inodedephd);
	LIST_INIT(&bmsafemap->sm_newblkhd);
	ACQUIRE_LOCK(&lk);
	WORKLIST_INSERT_BP(bp, &bmsafemap->sm_list);
	return (bmsafemap);
}

/*
 * Direct block allocation dependencies.
 * 
 * When a new block is allocated, the corresponding disk locations must be
 * initialized (with zeros or new data) before the on-disk inode points to
 * them.  Also, the freemap from which the block was allocated must be
 * updated (on disk) before the inode's pointer. These two dependencies are
 * independent of each other and are needed for all file blocks and indirect
 * blocks that are pointed to directly by the inode.  Just before the
 * "in-core" version of the inode is updated with a newly allocated block
 * number, a procedure (below) is called to setup allocation dependency
 * structures.  These structures are removed when the corresponding
 * dependencies are satisfied or when the block allocation becomes obsolete
 * (i.e., the file is deleted, the block is de-allocated, or the block is a
 * fragment that gets upgraded).  All of these cases are handled in
 * procedures described later.
 * 
 * When a file extension causes a fragment to be upgraded, either to a larger
 * fragment or to a full block, the on-disk location may change (if the
 * previous fragment could not simply be extended). In this case, the old
 * fragment must be de-allocated, but not until after the inode's pointer has
 * been updated. In most cases, this is handled by later procedures, which
 * will construct a "freefrag" structure to be added to the workitem queue
 * when the inode update is complete (or obsolete).  The main exception to
 * this is when an allocation occurs while a pending allocation dependency
 * (for the same block pointer) remains.  This case is handled in the main
 * allocation dependency setup procedure by immediately freeing the
 * unreferenced fragments.
 *
 * Parameters:
 *	ip:		inode to which block is being added
 *	lbn:		block pointer within inode
 *	newblkno:	disk block number being added
 *	oldblkno:	previous block number, 0 unless frag
 *	newsize:	size of new block
 *	oldsize:	size of new block
 *	bp:		bp for allocated block
 */ 
void 
softdep_setup_allocdirect(struct inode *ip, ufs_lbn_t lbn, ufs_daddr_t newblkno,
			  ufs_daddr_t oldblkno, long newsize, long oldsize,
			  struct buf *bp)
{
	struct allocdirect *adp, *oldadp;
	struct allocdirectlst *adphead;
	struct bmsafemap *bmsafemap;
	struct inodedep *inodedep;
	struct pagedep *pagedep;
	struct newblk *newblk;

	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
		M_ALLOCDIRECT, M_SOFTDEP_FLAGS | M_ZERO);
	adp->ad_list.wk_type = D_ALLOCDIRECT;
	adp->ad_lbn = lbn;
	adp->ad_newblkno = newblkno;
	adp->ad_oldblkno = oldblkno;
	adp->ad_newsize = newsize;
	adp->ad_oldsize = oldsize;
	adp->ad_state = ATTACHED;
	if (newblkno == oldblkno)
		adp->ad_freefrag = NULL;
	else
		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);

	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
		panic("softdep_setup_allocdirect: lost block");

	ACQUIRE_LOCK(&lk);
	inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
	adp->ad_inodedep = inodedep;

	if (newblk->nb_state == DEPCOMPLETE) {
		adp->ad_state |= DEPCOMPLETE;
		adp->ad_buf = NULL;
	} else {
		bmsafemap = newblk->nb_bmsafemap;
		adp->ad_buf = bmsafemap->sm_buf;
		LIST_REMOVE(newblk, nb_deps);
		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
	}
	LIST_REMOVE(newblk, nb_hash);
	FREE(newblk, M_NEWBLK);

	WORKLIST_INSERT_BP(bp, &adp->ad_list);
	if (lbn >= NDADDR) {
		/* allocating an indirect block */
		if (oldblkno != 0) {
			FREE_LOCK(&lk);
			panic("softdep_setup_allocdirect: non-zero indir");
		}
	} else {
		/*
		 * Allocating a direct block.
		 *
		 * If we are allocating a directory block, then we must
		 * allocate an associated pagedep to track additions and
		 * deletions.
		 */
		if ((ip->i_mode & IFMT) == IFDIR &&
		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) {
			WORKLIST_INSERT_BP(bp, &pagedep->pd_list);
		}
	}
	/*
	 * The list of allocdirects must be kept in sorted and ascending
	 * order so that the rollback routines can quickly determine the
	 * first uncommitted block (the size of the file stored on disk
	 * ends at the end of the lowest committed fragment, or if there
	 * are no fragments, at the end of the highest committed block).
	 * Since files generally grow, the typical case is that the new
	 * block is to be added at the end of the list. We speed this
	 * special case by checking against the last allocdirect in the
	 * list before laboriously traversing the list looking for the
	 * insertion point.
	 */
	adphead = &inodedep->id_newinoupdt;
	oldadp = TAILQ_LAST(adphead, allocdirectlst);
	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
		/* insert at end of list */
		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
		if (oldadp != NULL && oldadp->ad_lbn == lbn)
			allocdirect_merge(adphead, adp, oldadp);
		FREE_LOCK(&lk);
		return;
	}
	TAILQ_FOREACH(oldadp, adphead, ad_next) {
		if (oldadp->ad_lbn >= lbn)
			break;
	}
	if (oldadp == NULL) {
		FREE_LOCK(&lk);
		panic("softdep_setup_allocdirect: lost entry");
	}
	/* insert in middle of list */
	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
	if (oldadp->ad_lbn == lbn)
		allocdirect_merge(adphead, adp, oldadp);
	FREE_LOCK(&lk);
}

/*
 * Replace an old allocdirect dependency with a newer one.
 * This routine must be called with splbio interrupts blocked.
 *
 * Parameters:
 *	adphead:	head of list holding allocdirects
 *	newadp:		allocdirect being added
 *	oldadp:		existing allocdirect being checked
 */
static void
allocdirect_merge(struct allocdirectlst *adphead,
		  struct allocdirect *newadp,
		  struct allocdirect *oldadp)
{
	struct freefrag *freefrag;

#ifdef DEBUG
	if (lk.lkt_held == NOHOLDER)
		panic("allocdirect_merge: lock not held");
#endif
	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
	    newadp->ad_oldsize != oldadp->ad_newsize ||
	    newadp->ad_lbn >= NDADDR) {
		FREE_LOCK(&lk);
		panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
		    newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
		    NDADDR);
	}
	newadp->ad_oldblkno = oldadp->ad_oldblkno;
	newadp->ad_oldsize = oldadp->ad_oldsize;
	/*
	 * If the old dependency had a fragment to free or had never
	 * previously had a block allocated, then the new dependency
	 * can immediately post its freefrag and adopt the old freefrag.
	 * This action is done by swapping the freefrag dependencies.
	 * The new dependency gains the old one's freefrag, and the
	 * old one gets the new one and then immediately puts it on
	 * the worklist when it is freed by free_allocdirect. It is
	 * not possible to do this swap when the old dependency had a
	 * non-zero size but no previous fragment to free. This condition
	 * arises when the new block is an extension of the old block.
	 * Here, the first part of the fragment allocated to the new
	 * dependency is part of the block currently claimed on disk by
	 * the old dependency, so cannot legitimately be freed until the
	 * conditions for the new dependency are fulfilled.
	 */
	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
		freefrag = newadp->ad_freefrag;
		newadp->ad_freefrag = oldadp->ad_freefrag;
		oldadp->ad_freefrag = freefrag;
	}
	free_allocdirect(adphead, oldadp, 0);
}
		
/*
 * Allocate a new freefrag structure if needed.
 */
static struct freefrag *
newfreefrag(struct inode *ip, ufs_daddr_t blkno, long size)
{
	struct freefrag *freefrag;
	struct fs *fs;

	if (blkno == 0)
		return (NULL);
	fs = ip->i_fs;
	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
		panic("newfreefrag: frag size");
	MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
		M_FREEFRAG, M_SOFTDEP_FLAGS);
	freefrag->ff_list.wk_type = D_FREEFRAG;
	freefrag->ff_state = ip->i_uid & ~ONWORKLIST;	/* XXX - used below */
	freefrag->ff_inum = ip->i_number;
	freefrag->ff_fs = fs;
	freefrag->ff_devvp = ip->i_devvp;
	freefrag->ff_blkno = blkno;
	freefrag->ff_fragsize = size;
	return (freefrag);
}

/*
 * This workitem de-allocates fragments that were replaced during
 * file block allocation.
 */
static void 
handle_workitem_freefrag(struct freefrag *freefrag)
{
	struct inode tip;

	tip.i_fs = freefrag->ff_fs;
	tip.i_devvp = freefrag->ff_devvp;
	tip.i_dev = freefrag->ff_devvp->v_rdev;
	tip.i_number = freefrag->ff_inum;
	tip.i_uid = freefrag->ff_state & ~ONWORKLIST;	/* XXX - set above */
	ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
	FREE(freefrag, M_FREEFRAG);
}

/*
 * Indirect block allocation dependencies.
 * 
 * The same dependencies that exist for a direct block also exist when
 * a new block is allocated and pointed to by an entry in a block of
 * indirect pointers. The undo/redo states described above are also
 * used here. Because an indirect block contains many pointers that
 * may have dependencies, a second copy of the entire in-memory indirect
 * block is kept. The buffer cache copy is always completely up-to-date.
 * The second copy, which is used only as a source for disk writes,
 * contains only the safe pointers (i.e., those that have no remaining
 * update dependencies). The second copy is freed when all pointers
 * are safe. The cache is not allowed to replace indirect blocks with
 * pending update dependencies. If a buffer containing an indirect
 * block with dependencies is written, these routines will mark it
 * dirty again. It can only be successfully written once all the
 * dependencies are removed. The ffs_fsync routine in conjunction with
 * softdep_sync_metadata work together to get all the dependencies
 * removed so that a file can be successfully written to disk. Three
 * procedures are used when setting up indirect block pointer
 * dependencies. The division is necessary because of the organization
 * of the "balloc" routine and because of the distinction between file
 * pages and file metadata blocks.
 */

/*
 * Allocate a new allocindir structure.
 *
 * Parameters:
 *	ip:		inode for file being extended
 *	ptrno:		offset of pointer in indirect block
 *	newblkno:	disk block number being added
 *	oldblkno:	previous block number, 0 if none
 */
static struct allocindir *
newallocindir(struct inode *ip, int ptrno, ufs_daddr_t newblkno,
	      ufs_daddr_t oldblkno)
{
	struct allocindir *aip;

	MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
		M_ALLOCINDIR, M_SOFTDEP_FLAGS | M_ZERO);
	aip->ai_list.wk_type = D_ALLOCINDIR;
	aip->ai_state = ATTACHED;
	aip->ai_offset = ptrno;
	aip->ai_newblkno = newblkno;
	aip->ai_oldblkno = oldblkno;
	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
	return (aip);
}

/*
 * Called just before setting an indirect block pointer
 * to a newly allocated file page.
 *
 * Parameters:
 *	ip:		inode for file being extended
 *	lbn:		allocated block number within file
 *	bp:		buffer with indirect blk referencing page
 *	ptrno:		offset of pointer in indirect block
 *	newblkno:	disk block number being added
 *	oldblkno:	previous block number, 0 if none
 *	nbp:		buffer holding allocated page
 */
void
softdep_setup_allocindir_page(struct inode *ip, ufs_lbn_t lbn,
			      struct buf *bp, int ptrno,
			      ufs_daddr_t newblkno, ufs_daddr_t oldblkno,
			      struct buf *nbp)
{
	struct allocindir *aip;
	struct pagedep *pagedep;

	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
	ACQUIRE_LOCK(&lk);
	/*
	 * If we are allocating a directory page, then we must
	 * allocate an associated pagedep to track additions and
	 * deletions.
	 */
	if ((ip->i_mode & IFMT) == IFDIR &&
	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
		WORKLIST_INSERT_BP(nbp, &pagedep->pd_list);
	WORKLIST_INSERT_BP(nbp, &aip->ai_list);
	FREE_LOCK(&lk);
	setup_allocindir_phase2(bp, ip, aip);
}

/*
 * Called just before setting an indirect block pointer to a
 * newly allocated indirect block.
 * Parameters:
 *	nbp:		newly allocated indirect block
 *	ip:		inode for file being extended
 *	bp:		indirect block referencing allocated block
 *	ptrno:		offset of pointer in indirect block
 *	newblkno:	disk block number being added
 */
void
softdep_setup_allocindir_meta(struct buf *nbp, struct inode *ip,
			      struct buf *bp, int ptrno,
			      ufs_daddr_t newblkno)
{
	struct allocindir *aip;

	aip = newallocindir(ip, ptrno, newblkno, 0);
	ACQUIRE_LOCK(&lk);
	WORKLIST_INSERT_BP(nbp, &aip->ai_list);
	FREE_LOCK(&lk);
	setup_allocindir_phase2(bp, ip, aip);
}

/*
 * Called to finish the allocation of the "aip" allocated
 * by one of the two routines above.
 *
 * Parameters:
 *	bp:	in-memory copy of the indirect block
 *	ip:	inode for file being extended
 *	aip:	allocindir allocated by the above routines
 */
static void 
setup_allocindir_phase2(struct buf *bp, struct inode *ip,
			struct allocindir *aip)
{
	struct worklist *wk;
	struct indirdep *indirdep, *newindirdep;
	struct bmsafemap *bmsafemap;
	struct allocindir *oldaip;
	struct freefrag *freefrag;
	struct newblk *newblk;

	if (bp->b_loffset >= 0)
		panic("setup_allocindir_phase2: not indir blk");
	for (indirdep = NULL, newindirdep = NULL; ; ) {
		ACQUIRE_LOCK(&lk);
		LIST_FOREACH(wk, &bp->b_dep, wk_list) {
			if (wk->wk_type != D_INDIRDEP)
				continue;
			indirdep = WK_INDIRDEP(wk);
			break;
		}
		if (indirdep == NULL && newindirdep) {
			indirdep = newindirdep;
			WORKLIST_INSERT_BP(bp, &indirdep->ir_list);
			newindirdep = NULL;
		}
		FREE_LOCK(&lk);
		if (indirdep) {
			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
			    &newblk) == 0)
				panic("setup_allocindir: lost block");
			ACQUIRE_LOCK(&lk);
			if (newblk->nb_state == DEPCOMPLETE) {
				aip->ai_state |= DEPCOMPLETE;
				aip->ai_buf = NULL;
			} else {
				bmsafemap = newblk->nb_bmsafemap;
				aip->ai_buf = bmsafemap->sm_buf;
				LIST_REMOVE(newblk, nb_deps);
				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
				    aip, ai_deps);
			}
			LIST_REMOVE(newblk, nb_hash);
			FREE(newblk, M_NEWBLK);
			aip->ai_indirdep = indirdep;
			/*
			 * Check to see if there is an existing dependency
			 * for this block. If there is, merge the old
			 * dependency into the new one.
			 */
			if (aip->ai_oldblkno == 0)
				oldaip = NULL;
			else

				LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
					if (oldaip->ai_offset == aip->ai_offset)
						break;
			if (oldaip != NULL) {
				if (oldaip->ai_newblkno != aip->ai_oldblkno) {
					FREE_LOCK(&lk);
					panic("setup_allocindir_phase2: blkno");
				}
				aip->ai_oldblkno = oldaip->ai_oldblkno;
				freefrag = oldaip->ai_freefrag;
				oldaip->ai_freefrag = aip->ai_freefrag;
				aip->ai_freefrag = freefrag;
				free_allocindir(oldaip, NULL);
			}
			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
			((ufs_daddr_t *)indirdep->ir_savebp->b_data)
			    [aip->ai_offset] = aip->ai_oldblkno;
			FREE_LOCK(&lk);
		}
		if (newindirdep) {
			/*
			 * Avoid any possibility of data corruption by 
			 * ensuring that our old version is thrown away.
			 */
			newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
			brelse(newindirdep->ir_savebp);
			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
		}
		if (indirdep)
			break;
		MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
			M_INDIRDEP, M_SOFTDEP_FLAGS);
		newindirdep->ir_list.wk_type = D_INDIRDEP;
		newindirdep->ir_state = ATTACHED;
		LIST_INIT(&newindirdep->ir_deplisthd);
		LIST_INIT(&newindirdep->ir_donehd);
		if (bp->b_bio2.bio_offset == NOOFFSET) {
			VOP_BMAP(bp->b_vp, bp->b_bio1.bio_offset, 
				 &bp->b_bio2.bio_offset, NULL, NULL,
				 BUF_CMD_WRITE);
		}
		KKASSERT(bp->b_bio2.bio_offset != NOOFFSET);
		newindirdep->ir_savebp = getblk(ip->i_devvp,
						bp->b_bio2.bio_offset,
					        bp->b_bcount, 0, 0);
		BUF_KERNPROC(newindirdep->ir_savebp);
		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
	}
}

/*
 * Block de-allocation dependencies.
 * 
 * When blocks are de-allocated, the on-disk pointers must be nullified before
 * the blocks are made available for use by other files.  (The true
 * requirement is that old pointers must be nullified before new on-disk
 * pointers are set.  We chose this slightly more stringent requirement to
 * reduce complexity.) Our implementation handles this dependency by updating
 * the inode (or indirect block) appropriately but delaying the actual block
 * de-allocation (i.e., freemap and free space count manipulation) until
 * after the updated versions reach stable storage.  After the disk is
 * updated, the blocks can be safely de-allocated whenever it is convenient.
 * This implementation handles only the common case of reducing a file's
 * length to zero. Other cases are handled by the conventional synchronous
 * write approach.
 *
 * The ffs implementation with which we worked double-checks
 * the state of the block pointers and file size as it reduces
 * a file's length.  Some of this code is replicated here in our
 * soft updates implementation.  The freeblks->fb_chkcnt field is
 * used to transfer a part of this information to the procedure
 * that eventually de-allocates the blocks.
 *
 * This routine should be called from the routine that shortens
 * a file's length, before the inode's size or block pointers
 * are modified. It will save the block pointer information for
 * later release and zero the inode so that the calling routine
 * can release it.
 */
struct softdep_setup_freeblocks_info {
	struct fs *fs;
	struct inode *ip;
};

static int softdep_setup_freeblocks_bp(struct buf *bp, void *data);

/*
 * Parameters:
 *	ip:	The inode whose length is to be reduced
 *	length:	The new length for the file
 */
void
softdep_setup_freeblocks(struct inode *ip, off_t length)
{
	struct softdep_setup_freeblocks_info info;
	struct freeblks *freeblks;
	struct inodedep *inodedep;
	struct allocdirect *adp;
	struct vnode *vp;
	struct buf *bp;
	struct fs *fs;
	int i, error, delay;
	int count;

	fs = ip->i_fs;
	if (length != 0)
		panic("softde_setup_freeblocks: non-zero length");
	MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
		M_FREEBLKS, M_SOFTDEP_FLAGS | M_ZERO);
	freeblks->fb_list.wk_type = D_FREEBLKS;
	freeblks->fb_state = ATTACHED;
	freeblks->fb_uid = ip->i_uid;
	freeblks->fb_previousinum = ip->i_number;
	freeblks->fb_devvp = ip->i_devvp;
	freeblks->fb_fs = fs;
	freeblks->fb_oldsize = ip->i_size;
	freeblks->fb_newsize = length;
	freeblks->fb_chkcnt = ip->i_blocks;
	for (i = 0; i < NDADDR; i++) {
		freeblks->fb_dblks[i] = ip->i_db[i];
		ip->i_db[i] = 0;
	}
	for (i = 0; i < NIADDR; i++) {
		freeblks->fb_iblks[i] = ip->i_ib[i];
		ip->i_ib[i] = 0;
	}
	ip->i_blocks = 0;
	ip->i_size = 0;
	/*
	 * Push the zero'ed inode to to its disk buffer so that we are free
	 * to delete its dependencies below. Once the dependencies are gone
	 * the buffer can be safely released.
	 */
	if ((error = bread(ip->i_devvp,
			    fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)),
	    (int)fs->fs_bsize, &bp)) != 0)
		softdep_error("softdep_setup_freeblocks", error);
	*((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
	    ip->i_din;
	/*
	 * Find and eliminate any inode dependencies.
	 */
	ACQUIRE_LOCK(&lk);
	(void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
	if ((inodedep->id_state & IOSTARTED) != 0) {
		FREE_LOCK(&lk);
		panic("softdep_setup_freeblocks: inode busy");
	}
	/*
	 * Add the freeblks structure to the list of operations that
	 * must await the zero'ed inode being written to disk. If we
	 * still have a bitmap dependency (delay == 0), then the inode
	 * has never been written to disk, so we can process the
	 * freeblks below once we have deleted the dependencies.
	 */
	delay = (inodedep->id_state & DEPCOMPLETE);
	if (delay)
		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
	/*
	 * Because the file length has been truncated to zero, any
	 * pending block allocation dependency structures associated
	 * with this inode are obsolete and can simply be de-allocated.
	 * We must first merge the two dependency lists to get rid of
	 * any duplicate freefrag structures, then purge the merged list.
	 */
	merge_inode_lists(inodedep);
	while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
		free_allocdirect(&inodedep->id_inoupdt, adp, 1);
	FREE_LOCK(&lk);
	bdwrite(bp);
	/*
	 * We must wait for any I/O in progress to finish so that
	 * all potential buffers on the dirty list will be visible.
	 * Once they are all there, walk the list and get rid of
	 * any dependencies.
	 */
	vp = ITOV(ip);
	ACQUIRE_LOCK(&lk);
	drain_output(vp, 1);

	info.fs = fs;
	info.ip = ip;
	lwkt_gettoken(&vp->v_token);
	do {
		count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL, 
				softdep_setup_freeblocks_bp, &info);
	} while (count != 0);
	lwkt_reltoken(&vp->v_token);

	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
		(void)free_inodedep(inodedep);

	if (delay) {
		freeblks->fb_state |= DEPCOMPLETE;
		/*
		 * If the inode with zeroed block pointers is now on disk
		 * we can start freeing blocks. Add freeblks to the worklist
		 * instead of calling  handle_workitem_freeblocks directly as
		 * it is more likely that additional IO is needed to complete
		 * the request here than in the !delay case.
		 */
		if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
			add_to_worklist(&freeblks->fb_list);
	}

	FREE_LOCK(&lk);
	/*
	 * If the inode has never been written to disk (delay == 0),
	 * then we can process the freeblks now that we have deleted
	 * the dependencies.
	 */
	if (!delay)
		handle_workitem_freeblocks(freeblks);
}

static int
softdep_setup_freeblocks_bp(struct buf *bp, void *data)
{
	struct softdep_setup_freeblocks_info *info = data;
	struct inodedep *inodedep;

	if (getdirtybuf(&bp, MNT_WAIT) == 0) {
		kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp);
		return(-1);
	}
	if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) {
		kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp);
		BUF_UNLOCK(bp);
		return(-1);
	}
	(void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep);
	deallocate_dependencies(bp, inodedep);
	bp->b_flags |= B_INVAL | B_NOCACHE;
	FREE_LOCK(&lk);
	brelse(bp);
	ACQUIRE_LOCK(&lk);
	return(1);
}

/*
 * Reclaim any dependency structures from a buffer that is about to
 * be reallocated to a new vnode. The buffer must be locked, thus,
 * no I/O completion operations can occur while we are manipulating
 * its associated dependencies. The mutex is held so that other I/O's
 * associated with related dependencies do not occur.
 */
static void
deallocate_dependencies(struct buf *bp, struct inodedep *inodedep)
{
	struct worklist *wk;
	struct indirdep *indirdep;
	struct allocindir *aip;
	struct pagedep *pagedep;
	struct dirrem *dirrem;
	struct diradd *dap;
	int i;

	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
		switch (wk->wk_type) {

		case D_INDIRDEP:
			indirdep = WK_INDIRDEP(wk);
			/*
			 * None of the indirect pointers will ever be visible,
			 * so they can simply be tossed. GOINGAWAY ensures
			 * that allocated pointers will be saved in the buffer
			 * cache until they are freed. Note that they will
			 * only be able to be found by their physical address
			 * since the inode mapping the logical address will
			 * be gone. The save buffer used for the safe copy
			 * was allocated in setup_allocindir_phase2 using
			 * the physical address so it could be used for this
			 * purpose. Hence we swap the safe copy with the real
			 * copy, allowing the safe copy to be freed and holding
			 * on to the real copy for later use in indir_trunc.
			 *
			 * NOTE: ir_savebp is relative to the block device
			 * so b_bio1 contains the device block number.
			 */
			if (indirdep->ir_state & GOINGAWAY) {
				FREE_LOCK(&lk);
				panic("deallocate_dependencies: already gone");
			}
			indirdep->ir_state |= GOINGAWAY;
			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
				free_allocindir(aip, inodedep);
			if (bp->b_bio1.bio_offset >= 0 ||
			    bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) {
				FREE_LOCK(&lk);
				panic("deallocate_dependencies: not indir");
			}
			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
			    bp->b_bcount);
			WORKLIST_REMOVE(wk);
			WORKLIST_INSERT_BP(indirdep->ir_savebp, wk);
			continue;

		case D_PAGEDEP:
			pagedep = WK_PAGEDEP(wk);
			/*
			 * None of the directory additions will ever be
			 * visible, so they can simply be tossed.
			 */
			for (i = 0; i < DAHASHSZ; i++)
				while ((dap =
				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
					free_diradd(dap);
			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
				free_diradd(dap);
			/*
			 * Copy any directory remove dependencies to the list
			 * to be processed after the zero'ed inode is written.
			 * If the inode has already been written, then they 
			 * can be dumped directly onto the work list.
			 */
			LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
				LIST_REMOVE(dirrem, dm_next);
				dirrem->dm_dirinum = pagedep->pd_ino;
				if (inodedep == NULL ||
				    (inodedep->id_state & ALLCOMPLETE) ==
				     ALLCOMPLETE)
					add_to_worklist(&dirrem->dm_list);
				else
					WORKLIST_INSERT(&inodedep->id_bufwait,
					    &dirrem->dm_list);
			}
			WORKLIST_REMOVE(&pagedep->pd_list);
			LIST_REMOVE(pagedep, pd_hash);
			WORKITEM_FREE(pagedep, D_PAGEDEP);
			continue;

		case D_ALLOCINDIR:
			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
			continue;

		case D_ALLOCDIRECT:
		case D_INODEDEP:
			FREE_LOCK(&lk);
			panic("deallocate_dependencies: Unexpected type %s",
			    TYPENAME(wk->wk_type));
			/* NOTREACHED */

		default:
			FREE_LOCK(&lk);
			panic("deallocate_dependencies: Unknown type %s",
			    TYPENAME(wk->wk_type));
			/* NOTREACHED */
		}
	}
}

/*
 * Free an allocdirect. Generate a new freefrag work request if appropriate.
 * This routine must be called with splbio interrupts blocked.
 */
static void
free_allocdirect(struct allocdirectlst *adphead,
		 struct allocdirect *adp, int delay)
{

#ifdef DEBUG
	if (lk.lkt_held == NOHOLDER)
		panic("free_allocdirect: lock not held");
#endif
	if ((adp->ad_state & DEPCOMPLETE) == 0)
		LIST_REMOVE(adp, ad_deps);
	TAILQ_REMOVE(adphead, adp, ad_next);
	if ((adp->ad_state & COMPLETE) == 0)
		WORKLIST_REMOVE(&adp->ad_list);
	if (adp->ad_freefrag != NULL) {
		if (delay)
			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
			    &adp->ad_freefrag->ff_list);
		else
			add_to_worklist(&adp->ad_freefrag->ff_list);
	}
	WORKITEM_FREE(adp, D_ALLOCDIRECT);
}

/*
 * Prepare an inode to be freed. The actual free operation is not
 * done until the zero'ed inode has been written to disk.
 */
void
softdep_freefile(struct vnode *pvp, ino_t ino, int mode)
{
	struct inode *ip = VTOI(pvp);
	struct inodedep *inodedep;
	struct freefile *freefile;

	/*
	 * This sets up the inode de-allocation dependency.
	 */
	MALLOC(freefile, struct freefile *, sizeof(struct freefile),
		M_FREEFILE, M_SOFTDEP_FLAGS);
	freefile->fx_list.wk_type = D_FREEFILE;
	freefile->fx_list.wk_state = 0;
	freefile->fx_mode = mode;
	freefile->fx_oldinum = ino;
	freefile->fx_devvp = ip->i_devvp;
	freefile->fx_fs = ip->i_fs;

	/*
	 * If the inodedep does not exist, then the zero'ed inode has
	 * been written to disk. If the allocated inode has never been
	 * written to disk, then the on-disk inode is zero'ed. In either
	 * case we can free the file immediately.
	 */
	ACQUIRE_LOCK(&lk);
	if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
	    check_inode_unwritten(inodedep)) {
		FREE_LOCK(&lk);
		handle_workitem_freefile(freefile);
		return;
	}
	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
	FREE_LOCK(&lk);
}

/*
 * Check to see if an inode has never been written to disk. If
 * so free the inodedep and return success, otherwise return failure.
 * This routine must be called with splbio interrupts blocked.
 *
 * If we still have a bitmap dependency, then the inode has never
 * been written to disk. Drop the dependency as it is no longer
 * necessary since the inode is being deallocated. We set the
 * ALLCOMPLETE flags since the bitmap now properly shows that the
 * inode is not allocated. Even if the inode is actively being
 * written, it has been rolled back to its zero'ed state, so we
 * are ensured that a zero inode is what is on the disk. For short
 * lived files, this change will usually result in removing all the
 * dependencies from the inode so that it can be freed immediately.
 */
static int
check_inode_unwritten(struct inodedep *inodedep)
{

	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
	    inodedep->id_nlinkdelta != 0)
		return (0);

	/*
	 * Another process might be in initiate_write_inodeblock
	 * trying to allocate memory without holding "Softdep Lock".
	 */
	if ((inodedep->id_state & IOSTARTED) != 0 &&
	    inodedep->id_savedino == NULL)
		return(0);

	inodedep->id_state |= ALLCOMPLETE;
	LIST_REMOVE(inodedep, id_deps);
	inodedep->id_buf = NULL;
	if (inodedep->id_state & ONWORKLIST)
		WORKLIST_REMOVE(&inodedep->id_list);
	if (inodedep->id_savedino != NULL) {
		FREE(inodedep->id_savedino, M_INODEDEP);
		inodedep->id_savedino = NULL;
	}
	if (free_inodedep(inodedep) == 0) {
		FREE_LOCK(&lk);
		panic("check_inode_unwritten: busy inode");
	}
	return (1);
}

/*
 * Try to free an inodedep structure. Return 1 if it could be freed.
 */
static int
free_inodedep(struct inodedep *inodedep)
{

	if ((inodedep->id_state & ONWORKLIST) != 0 ||
	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
		return (0);
	LIST_REMOVE(inodedep, id_hash);
	WORKITEM_FREE(inodedep, D_INODEDEP);
	num_inodedep -= 1;
	return (1);
}

/*
 * This workitem routine performs the block de-allocation.
 * The workitem is added to the pending list after the updated
 * inode block has been written to disk.  As mentioned above,
 * checks regarding the number of blocks de-allocated (compared
 * to the number of blocks allocated for the file) are also
 * performed in this function.
 */
static void
handle_workitem_freeblocks(struct freeblks *freeblks)
{
	struct inode tip;
	ufs_daddr_t bn;
	struct fs *fs;
	int i, level, bsize;
	long nblocks, blocksreleased = 0;
	int error, allerror = 0;
	ufs_lbn_t baselbns[NIADDR], tmpval;

	tip.i_number = freeblks->fb_previousinum;
	tip.i_devvp = freeblks->fb_devvp;
	tip.i_dev = freeblks->fb_devvp->v_rdev;
	tip.i_fs = freeblks->fb_fs;
	tip.i_size = freeblks->fb_oldsize;
	tip.i_uid = freeblks->fb_uid;
	fs = freeblks->fb_fs;
	tmpval = 1;
	baselbns[0] = NDADDR;
	for (i = 1; i < NIADDR; i++) {
		tmpval *= NINDIR(fs);
		baselbns[i] = baselbns[i - 1] + tmpval;
	}
	nblocks = btodb(fs->fs_bsize);
	blocksreleased = 0;
	/*
	 * Indirect blocks first.
	 */
	for (level = (NIADDR - 1); level >= 0; level--) {
		if ((bn = freeblks->fb_iblks[level]) == 0)
			continue;
		if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level,
		    baselbns[level], &blocksreleased)) == 0)
			allerror = error;
		ffs_blkfree(&tip, bn, fs->fs_bsize);
		blocksreleased += nblocks;
	}
	/*
	 * All direct blocks or frags.
	 */
	for (i = (NDADDR - 1); i >= 0; i--) {
		if ((bn = freeblks->fb_dblks[i]) == 0)
			continue;
		bsize = blksize(fs, &tip, i);
		ffs_blkfree(&tip, bn, bsize);
		blocksreleased += btodb(bsize);
	}

#ifdef DIAGNOSTIC
	if (freeblks->fb_chkcnt != blocksreleased)
		kprintf("handle_workitem_freeblocks: block count\n");
	if (allerror)
		softdep_error("handle_workitem_freeblks", allerror);
#endif /* DIAGNOSTIC */
	WORKITEM_FREE(freeblks, D_FREEBLKS);
}

/*
 * Release blocks associated with the inode ip and stored in the indirect
 * block at doffset. If level is greater than SINGLE, the block is an
 * indirect block and recursive calls to indirtrunc must be used to
 * cleanse other indirect blocks.
 */
static int
indir_trunc(struct inode *ip, off_t doffset, int level, ufs_lbn_t lbn,
	    long *countp)
{
	struct buf *bp;
	ufs_daddr_t *bap;
	ufs_daddr_t nb;
	struct fs *fs;
	struct worklist *wk;
	struct indirdep *indirdep;
	int i, lbnadd, nblocks;
	int error, allerror = 0;

	fs = ip->i_fs;
	lbnadd = 1;
	for (i = level; i > 0; i--)
		lbnadd *= NINDIR(fs);
	/*
	 * Get buffer of block pointers to be freed. This routine is not
	 * called until the zero'ed inode has been written, so it is safe
	 * to free blocks as they are encountered. Because the inode has
	 * been zero'ed, calls to bmap on these blocks will fail. So, we
	 * have to use the on-disk address and the block device for the
	 * filesystem to look them up. If the file was deleted before its
	 * indirect blocks were all written to disk, the routine that set
	 * us up (deallocate_dependencies) will have arranged to leave
	 * a complete copy of the indirect block in memory for our use.
	 * Otherwise we have to read the blocks in from the disk.
	 */
	ACQUIRE_LOCK(&lk);
	if ((bp = findblk(ip->i_devvp, doffset, FINDBLK_TEST)) != NULL &&
	    (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
		/*
		 * bp must be ir_savebp, which is held locked for our use.
		 */
		if (wk->wk_type != D_INDIRDEP ||
		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
		    (indirdep->ir_state & GOINGAWAY) == 0) {
			FREE_LOCK(&lk);
			panic("indir_trunc: lost indirdep");
		}
		WORKLIST_REMOVE(wk);
		WORKITEM_FREE(indirdep, D_INDIRDEP);
		if (LIST_FIRST(&bp->b_dep) != NULL) {
			FREE_LOCK(&lk);
			panic("indir_trunc: dangling dep");
		}
		FREE_LOCK(&lk);
	} else {
		FREE_LOCK(&lk);
		error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp);
		if (error)
			return (error);
	}
	/*
	 * Recursively free indirect blocks.
	 */
	bap = (ufs_daddr_t *)bp->b_data;
	nblocks = btodb(fs->fs_bsize);
	for (i = NINDIR(fs) - 1; i >= 0; i--) {
		if ((nb = bap[i]) == 0)
			continue;
		if (level != 0) {
			if ((error = indir_trunc(ip, fsbtodoff(fs, nb),
			     level - 1, lbn + (i * lbnadd), countp)) != 0)
				allerror = error;
		}
		ffs_blkfree(ip, nb, fs->fs_bsize);
		*countp += nblocks;
	}
	bp->b_flags |= B_INVAL | B_NOCACHE;
	brelse(bp);
	return (allerror);
}

/*
 * Free an allocindir.
 * This routine must be called with splbio interrupts blocked.
 */
static void
free_allocindir(struct allocindir *aip, struct inodedep *inodedep)
{
	struct freefrag *freefrag;

#ifdef DEBUG
	if (lk.lkt_held == NOHOLDER)
		panic("free_allocindir: lock not held");
#endif
	if ((aip->ai_state & DEPCOMPLETE) == 0)
		LIST_REMOVE(aip, ai_deps);
	if (aip->ai_state & ONWORKLIST)
		WORKLIST_REMOVE(&aip->ai_list);
	LIST_REMOVE(aip, ai_next);
	if ((freefrag = aip->ai_freefrag) != NULL) {
		if (inodedep == NULL)
			add_to_worklist(&freefrag->ff_list);
		else
			WORKLIST_INSERT(&inodedep->id_bufwait,
			    &freefrag->ff_list);
	}
	WORKITEM_FREE(aip, D_ALLOCINDIR);
}

/*
 * Directory entry addition dependencies.
 * 
 * When adding a new directory entry, the inode (with its incremented link
 * count) must be written to disk before the directory entry's pointer to it.
 * Also, if the inode is newly allocated, the corresponding freemap must be
 * updated (on disk) before the directory entry's pointer. These requirements
 * are met via undo/redo on the directory entry's pointer, which consists
 * simply of the inode number.
 * 
 * As directory entries are added and deleted, the free space within a
 * directory block can become fragmented.  The ufs filesystem will compact
 * a fragmented directory block to make space for a new entry. When this
 * occurs, the offsets of previously added entries change. Any "diradd"
 * dependency structures corresponding to these entries must be updated with
 * the new offsets.
 */

/*
 * This routine is called after the in-memory inode's link
 * count has been incremented, but before the directory entry's
 * pointer to the inode has been set.
 *
 * Parameters:
 *	bp:		buffer containing directory block
 *	dp:		inode for directory
 *	diroffset:	offset of new entry in directory
 *	newinum:	inode referenced by new directory entry
 *	newdirbp:	non-NULL => contents of new mkdir
 */
void 
softdep_setup_directory_add(struct buf *bp, struct inode *dp, off_t diroffset,
			    ino_t newinum, struct buf *newdirbp)
{
	int offset;		/* offset of new entry within directory block */
	ufs_lbn_t lbn;		/* block in directory containing new entry */
	struct fs *fs;
	struct diradd *dap;
	struct pagedep *pagedep;
	struct inodedep *inodedep;
	struct mkdir *mkdir1, *mkdir2;

	/*
	 * Whiteouts have no dependencies.
	 */
	if (newinum == WINO) {
		if (newdirbp != NULL)
			bdwrite(newdirbp);
		return;
	}

	fs = dp->i_fs;
	lbn = lblkno(fs, diroffset);
	offset = blkoff(fs, diroffset);
	MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
	    M_SOFTDEP_FLAGS | M_ZERO);
	dap->da_list.wk_type = D_DIRADD;
	dap->da_offset = offset;
	dap->da_newinum = newinum;
	dap->da_state = ATTACHED;
	if (newdirbp == NULL) {
		dap->da_state |= DEPCOMPLETE;
		ACQUIRE_LOCK(&lk);
	} else {
		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
		MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
		    M_SOFTDEP_FLAGS);
		mkdir1->md_list.wk_type = D_MKDIR;
		mkdir1->md_state = MKDIR_BODY;
		mkdir1->md_diradd = dap;
		MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
		    M_SOFTDEP_FLAGS);
		mkdir2->md_list.wk_type = D_MKDIR;
		mkdir2->md_state = MKDIR_PARENT;
		mkdir2->md_diradd = dap;
		/*
		 * Dependency on "." and ".." being written to disk.
		 */
		mkdir1->md_buf = newdirbp;
		ACQUIRE_LOCK(&lk);
		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
		WORKLIST_INSERT_BP(newdirbp, &mkdir1->md_list);
		FREE_LOCK(&lk);
		bdwrite(newdirbp);
		/*
		 * Dependency on link count increase for parent directory
		 */
		ACQUIRE_LOCK(&lk);
		if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
			dap->da_state &= ~MKDIR_PARENT;
			WORKITEM_FREE(mkdir2, D_MKDIR);
		} else {
			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
		}
	}
	/*
	 * Link into parent directory pagedep to await its being written.
	 */
	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
		WORKLIST_INSERT_BP(bp, &pagedep->pd_list);
	dap->da_pagedep = pagedep;
	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
	    da_pdlist);
	/*
	 * Link into its inodedep. Put it on the id_bufwait list if the inode
	 * is not yet written. If it is written, do the post-inode write
	 * processing to put it on the id_pendinghd list.
	 */
	(void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
		diradd_inode_written(dap, inodedep);
	else
		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
	FREE_LOCK(&lk);
}

/*
 * This procedure is called to change the offset of a directory
 * entry when compacting a directory block which must be owned
 * exclusively by the caller. Note that the actual entry movement
 * must be done in this procedure to ensure that no I/O completions
 * occur while the move is in progress.
 *
 * Parameters:
 *	dp:	inode for directory
 *	base:		address of dp->i_offset
 *	oldloc:		address of old directory location
 *	newloc:		address of new directory location
 *	entrysize:	size of directory entry
 */
void 
softdep_change_directoryentry_offset(struct inode *dp, caddr_t base,
				     caddr_t oldloc, caddr_t newloc,
				     int entrysize)
{
	int offset, oldoffset, newoffset;
	struct pagedep *pagedep;
	struct diradd *dap;
	ufs_lbn_t lbn;

	ACQUIRE_LOCK(&lk);
	lbn = lblkno(dp->i_fs, dp->i_offset);
	offset = blkoff(dp->i_fs, dp->i_offset);
	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
		goto done;
	oldoffset = offset + (oldloc - base);
	newoffset = offset + (newloc - base);

	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
		if (dap->da_offset != oldoffset)
			continue;
		dap->da_offset = newoffset;
		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
			break;
		LIST_REMOVE(dap, da_pdlist);
		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
		    dap, da_pdlist);
		break;
	}
	if (dap == NULL) {

		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
			if (dap->da_offset == oldoffset) {
				dap->da_offset = newoffset;
				break;
			}
		}
	}
done:
	bcopy(oldloc, newloc, entrysize);
	FREE_LOCK(&lk);
}

/*
 * Free a diradd dependency structure. This routine must be called
 * with splbio interrupts blocked.
 */
static void
free_diradd(struct diradd *dap)
{
	struct dirrem *dirrem;
	struct pagedep *pagedep;
	struct inodedep *inodedep;
	struct mkdir *mkdir, *nextmd;

#ifdef DEBUG
	if (lk.lkt_held == NOHOLDER)
		panic("free_diradd: lock not held");
#endif
	WORKLIST_REMOVE(&dap->da_list);
	LIST_REMOVE(dap, da_pdlist);
	if ((dap->da_state & DIRCHG) == 0) {
		pagedep = dap->da_pagedep;
	} else {
		dirrem = dap->da_previous;
		pagedep = dirrem->dm_pagedep;
		dirrem->dm_dirinum = pagedep->pd_ino;
		add_to_worklist(&dirrem->dm_list);
	}
	if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
	    0, &inodedep) != 0)
		(void) free_inodedep(inodedep);
	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
			nextmd = LIST_NEXT(mkdir, md_mkdirs);
			if (mkdir->md_diradd != dap)
				continue;
			dap->da_state &= ~mkdir->md_state;
			WORKLIST_REMOVE(&mkdir->md_list);
			LIST_REMOVE(mkdir, md_mkdirs);
			WORKITEM_FREE(mkdir, D_MKDIR);
		}
		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
			FREE_LOCK(&lk);
			panic("free_diradd: unfound ref");
		}
	}
	WORKITEM_FREE(dap, D_DIRADD);
}

/*
 * Directory entry removal dependencies.
 * 
 * When removing a directory entry, the entry's inode pointer must be
 * zero'ed on disk before the corresponding inode's link count is decremented
 * (possibly freeing the inode for re-use). This dependency is handled by
 * updating the directory entry but delaying the inode count reduction until
 * after the directory block has been written to disk. After this point, the
 * inode count can be decremented whenever it is convenient.
 */

/*
 * This routine should be called immediately after removing
 * a directory entry.  The inode's link count should not be
 * decremented by the calling procedure -- the soft updates
 * code will do this task when it is safe.
 *
 * Parameters:
 *	bp:		buffer containing directory block
 *	dp:		inode for the directory being modified
 *	ip:		inode for directory entry being removed
 *	isrmdir:	indicates if doing RMDIR
 */
void 
softdep_setup_remove(struct buf *bp, struct inode *dp, struct inode *ip,
		     int isrmdir)
{
	struct dirrem *dirrem, *prevdirrem;

	/*
	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
	 */
	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);

	/*
	 * If the COMPLETE flag is clear, then there were no active
	 * entries and we want to roll back to a zeroed entry until
	 * the new inode is committed to disk. If the COMPLETE flag is
	 * set then we have deleted an entry that never made it to
	 * disk. If the entry we deleted resulted from a name change,
	 * then the old name still resides on disk. We cannot delete
	 * its inode (returned to us in prevdirrem) until the zeroed
	 * directory entry gets to disk. The new inode has never been
	 * referenced on the disk, so can be deleted immediately.
	 */
	if ((dirrem->dm_state & COMPLETE) == 0) {
		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
		    dm_next);
		FREE_LOCK(&lk);
	} else {
		if (prevdirrem != NULL)
			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
			    prevdirrem, dm_next);
		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
		FREE_LOCK(&lk);
		handle_workitem_remove(dirrem);
	}
}

/*
 * Allocate a new dirrem if appropriate and return it along with
 * its associated pagedep. Called without a lock, returns with lock.
 */
static long num_dirrem;		/* number of dirrem allocated */

/*
 * Parameters:
 *	bp:		buffer containing directory block
 *	dp:		inode for the directory being modified
 *	ip:		inode for directory entry being removed
 *	isrmdir:	indicates if doing RMDIR
 *	prevdirremp:	previously referenced inode, if any
 */
static struct dirrem *
newdirrem(struct buf *bp, struct inode *dp, struct inode *ip,
	  int isrmdir, struct dirrem **prevdirremp)
{
	int offset;
	ufs_lbn_t lbn;
	struct diradd *dap;
	struct dirrem *dirrem;
	struct pagedep *pagedep;

	/*
	 * Whiteouts have no deletion dependencies.
	 */
	if (ip == NULL)
		panic("newdirrem: whiteout");
	/*
	 * If we are over our limit, try to improve the situation.
	 * Limiting the number of dirrem structures will also limit
	 * the number of freefile and freeblks structures.
	 */
	if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
		(void) request_cleanup(FLUSH_REMOVE, 0);
	num_dirrem += 1;
	MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
		M_DIRREM, M_SOFTDEP_FLAGS | M_ZERO);
	dirrem->dm_list.wk_type = D_DIRREM;
	dirrem->dm_state = isrmdir ? RMDIR : 0;
	dirrem->dm_mnt = ITOV(ip)->v_mount;
	dirrem->dm_oldinum = ip->i_number;
	*prevdirremp = NULL;

	ACQUIRE_LOCK(&lk);
	lbn = lblkno(dp->i_fs, dp->i_offset);
	offset = blkoff(dp->i_fs, dp->i_offset);
	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
		WORKLIST_INSERT_BP(bp, &pagedep->pd_list);
	dirrem->dm_pagedep = pagedep;
	/*
	 * Check for a diradd dependency for the same directory entry.
	 * If present, then both dependencies become obsolete and can
	 * be de-allocated. Check for an entry on both the pd_dirraddhd
	 * list and the pd_pendinghd list.
	 */

	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
		if (dap->da_offset == offset)
			break;
	if (dap == NULL) {

		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
			if (dap->da_offset == offset)
				break;
		if (dap == NULL)
			return (dirrem);
	}
	/*
	 * Must be ATTACHED at this point.
	 */
	if ((dap->da_state & ATTACHED) == 0) {
		FREE_LOCK(&lk);
		panic("newdirrem: not ATTACHED");
	}
	if (dap->da_newinum != ip->i_number) {
		FREE_LOCK(&lk);
		panic("newdirrem: inum %"PRId64" should be %"PRId64,
		    ip->i_number, dap->da_newinum);
	}
	/*
	 * If we are deleting a changed name that never made it to disk,
	 * then return the dirrem describing the previous inode (which
	 * represents the inode currently referenced from this entry on disk).
	 */
	if ((dap->da_state & DIRCHG) != 0) {
		*prevdirremp = dap->da_previous;
		dap->da_state &= ~DIRCHG;
		dap->da_pagedep = pagedep;
	}
	/*
	 * We are deleting an entry that never made it to disk.
	 * Mark it COMPLETE so we can delete its inode immediately.
	 */
	dirrem->dm_state |= COMPLETE;
	free_diradd(dap);
	return (dirrem);
}

/*
 * Directory entry change dependencies.
 * 
 * Changing an existing directory entry requires that an add operation
 * be completed first followed by a deletion. The semantics for the addition
 * are identical to the description of adding a new entry above except
 * that the rollback is to the old inode number rather than zero. Once
 * the addition dependency is completed, the removal is done as described
 * in the removal routine above.
 */

/*
 * This routine should be called immediately after changing
 * a directory entry.  The inode's link count should not be
 * decremented by the calling procedure -- the soft updates
 * code will perform this task when it is safe.
 *
 * Parameters:
 *	bp:		buffer containing directory block
 *	dp:		inode for the directory being modified
 *	ip:		inode for directory entry being removed
 *	newinum:	new inode number for changed entry
 *	isrmdir:	indicates if doing RMDIR
 */
void 
softdep_setup_directory_change(struct buf *bp, struct inode *dp,
			       struct inode *ip, ino_t newinum,
			       int isrmdir)
{
	int offset;
	struct diradd *dap = NULL;
	struct dirrem *dirrem, *prevdirrem;
	struct pagedep *pagedep;
	struct inodedep *inodedep;

	offset = blkoff(dp->i_fs, dp->i_offset);

	/*
	 * Whiteouts do not need diradd dependencies.
	 */
	if (newinum != WINO) {
		MALLOC(dap, struct diradd *, sizeof(struct diradd),
		    M_DIRADD, M_SOFTDEP_FLAGS | M_ZERO);
		dap->da_list.wk_type = D_DIRADD;
		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
		dap->da_offset = offset;
		dap->da_newinum = newinum;
	}

	/*
	 * Allocate a new dirrem and ACQUIRE_LOCK.
	 */
	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
	pagedep = dirrem->dm_pagedep;
	/*
	 * The possible values for isrmdir:
	 *	0 - non-directory file rename
	 *	1 - directory rename within same directory
	 *   inum - directory rename to new directory of given inode number
	 * When renaming to a new directory, we are both deleting and
	 * creating a new directory entry, so the link count on the new
	 * directory should not change. Thus we do not need the followup
	 * dirrem which is usually done in handle_workitem_remove. We set
	 * the DIRCHG flag to tell handle_workitem_remove to skip the 
	 * followup dirrem.
	 */
	if (isrmdir > 1)
		dirrem->dm_state |= DIRCHG;

	/*
	 * Whiteouts have no additional dependencies,
	 * so just put the dirrem on the correct list.
	 */
	if (newinum == WINO) {
		if ((dirrem->dm_state & COMPLETE) == 0) {
			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
			    dm_next);
		} else {
			dirrem->dm_dirinum = pagedep->pd_ino;
			add_to_worklist(&dirrem->dm_list);
		}
		FREE_LOCK(&lk);
		return;
	}

	/*
	 * If the COMPLETE flag is clear, then there were no active
	 * entries and we want to roll back to the previous inode until
	 * the new inode is committed to disk. If the COMPLETE flag is
	 * set, then we have deleted an entry that never made it to disk.
	 * If the entry we deleted resulted from a name change, then the old
	 * inode reference still resides on disk. Any rollback that we do
	 * needs to be to that old inode (returned to us in prevdirrem). If
	 * the entry we deleted resulted from a create, then there is
	 * no entry on the disk, so we want to roll back to zero rather
	 * than the uncommitted inode. In either of the COMPLETE cases we
	 * want to immediately free the unwritten and unreferenced inode.
	 */
	if ((dirrem->dm_state & COMPLETE) == 0) {
		dap->da_previous = dirrem;
	} else {
		if (prevdirrem != NULL) {
			dap->da_previous = prevdirrem;
		} else {
			dap->da_state &= ~DIRCHG;
			dap->da_pagedep = pagedep;
		}
		dirrem->dm_dirinum = pagedep->pd_ino;
		add_to_worklist(&dirrem->dm_list);
	}
	/*
	 * Link into its inodedep. Put it on the id_bufwait list if the inode
	 * is not yet written. If it is written, do the post-inode write
	 * processing to put it on the id_pendinghd list.
	 */
	if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
		dap->da_state |= COMPLETE;
		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
	} else {
		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
		    dap, da_pdlist);
		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
	}
	FREE_LOCK(&lk);
}

/*
 * Called whenever the link count on an inode is changed.
 * It creates an inode dependency so that the new reference(s)
 * to the inode cannot be committed to disk until the updated
 * inode has been written.
 *
 * Parameters:
 *	ip:	the inode with the increased link count
 */
void
softdep_change_linkcnt(struct inode *ip)
{
	struct inodedep *inodedep;

	ACQUIRE_LOCK(&lk);
	(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
	if (ip->i_nlink < ip->i_effnlink) {
		FREE_LOCK(&lk);
		panic("softdep_change_linkcnt: bad delta");
	}
	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
	FREE_LOCK(&lk);
}

/*
 * This workitem decrements the inode's link count.
 * If the link count reaches zero, the file is removed.
 */
static void 
handle_workitem_remove(struct dirrem *dirrem)
{
	struct inodedep *inodedep;
	struct vnode *vp;
	struct inode *ip;
	ino_t oldinum;
	int error;

	error = VFS_VGET(dirrem->dm_mnt, NULL, dirrem->dm_oldinum, &vp);
	if (error) {
		softdep_error("handle_workitem_remove: vget", error);
		return;
	}
	ip = VTOI(vp);
	ACQUIRE_LOCK(&lk);
	if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
		FREE_LOCK(&lk);
		panic("handle_workitem_remove: lost inodedep");
	}
	/*
	 * Normal file deletion.
	 */
	if ((dirrem->dm_state & RMDIR) == 0) {
		ip->i_nlink--;
		ip->i_flag |= IN_CHANGE;
		if (ip->i_nlink < ip->i_effnlink) {
			FREE_LOCK(&lk);
			panic("handle_workitem_remove: bad file delta");
		}
		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
		FREE_LOCK(&lk);
		vput(vp);
		num_dirrem -= 1;
		WORKITEM_FREE(dirrem, D_DIRREM);
		return;
	}
	/*
	 * Directory deletion. Decrement reference count for both the
	 * just deleted parent directory entry and the reference for ".".
	 * Next truncate the directory to length zero. When the
	 * truncation completes, arrange to have the reference count on
	 * the parent decremented to account for the loss of "..".
	 */
	ip->i_nlink -= 2;
	ip->i_flag |= IN_CHANGE;
	if (ip->i_nlink < ip->i_effnlink) {
		FREE_LOCK(&lk);
		panic("handle_workitem_remove: bad dir delta");
	}
	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
	FREE_LOCK(&lk);
	if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0)
		softdep_error("handle_workitem_remove: truncate", error);
	/*
	 * Rename a directory to a new parent. Since, we are both deleting
	 * and creating a new directory entry, the link count on the new
	 * directory should not change. Thus we skip the followup dirrem.
	 */
	if (dirrem->dm_state & DIRCHG) {
		vput(vp);
		num_dirrem -= 1;
		WORKITEM_FREE(dirrem, D_DIRREM);
		return;
	}
	/*
	 * If the inodedep does not exist, then the zero'ed inode has
	 * been written to disk. If the allocated inode has never been
	 * written to disk, then the on-disk inode is zero'ed. In either
	 * case we can remove the file immediately.
	 */
	ACQUIRE_LOCK(&lk);
	dirrem->dm_state = 0;
	oldinum = dirrem->dm_oldinum;
	dirrem->dm_oldinum = dirrem->dm_dirinum;
	if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
	    check_inode_unwritten(inodedep)) {
		FREE_LOCK(&lk);
		vput(vp);
		handle_workitem_remove(dirrem);
		return;
	}
	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
	FREE_LOCK(&lk);
	ip->i_flag |= IN_CHANGE;
	ffs_update(vp, 0);
	vput(vp);
}

/*
 * Inode de-allocation dependencies.
 * 
 * When an inode's link count is reduced to zero, it can be de-allocated. We
 * found it convenient to postpone de-allocation until after the inode is
 * written to disk with its new link count (zero).  At this point, all of the
 * on-disk inode's block pointers are nullified and, with careful dependency
 * list ordering, all dependencies related to the inode will be satisfied and
 * the corresponding dependency structures de-allocated.  So, if/when the
 * inode is reused, there will be no mixing of old dependencies with new
 * ones.  This artificial dependency is set up by the block de-allocation
 * procedure above (softdep_setup_freeblocks) and completed by the
 * following procedure.
 */
static void 
handle_workitem_freefile(struct freefile *freefile)
{
	struct vnode vp;
	struct inode tip;
	struct inodedep *idp;
	int error;

#ifdef DEBUG
	ACQUIRE_LOCK(&lk);
	error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp);
	FREE_LOCK(&lk);
	if (error)
		panic("handle_workitem_freefile: inodedep survived");
#endif
	tip.i_devvp = freefile->fx_devvp;
	tip.i_dev = freefile->fx_devvp->v_rdev;
	tip.i_fs = freefile->fx_fs;
	vp.v_data = &tip;
	if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
		softdep_error("handle_workitem_freefile", error);
	WORKITEM_FREE(freefile, D_FREEFILE);
}

/*
 * Helper function which unlinks marker element from work list and returns
 * the next element on the list.
 */
static __inline struct worklist *
markernext(struct worklist *marker)
{
	struct worklist *next;

	next = LIST_NEXT(marker, wk_list);
	LIST_REMOVE(marker, wk_list);
	return next;
}

/*
 * checkread, checkwrite
 *
 * bioops callback - hold io_token
 */
static  int
softdep_checkread(struct buf *bp)
{
	/* nothing to do, mp lock not needed */
	return(0);
}

/*
 * bioops callback - hold io_token
 */
static  int
softdep_checkwrite(struct buf *bp)
{
	/* nothing to do, mp lock not needed */
	return(0);
}

/*
 * Disk writes.
 * 
 * The dependency structures constructed above are most actively used when file
 * system blocks are written to disk.  No constraints are placed on when a
 * block can be written, but unsatisfied update dependencies are made safe by
 * modifying (or replacing) the source memory for the duration of the disk
 * write.  When the disk write completes, the memory block is again brought
 * up-to-date.
 *
 * In-core inode structure reclamation.
 * 
 * Because there are a finite number of "in-core" inode structures, they are
 * reused regularly.  By transferring all inode-related dependencies to the
 * in-memory inode block and indexing them separately (via "inodedep"s), we
 * can allow "in-core" inode structures to be reused at any time and avoid
 * any increase in contention.
 *
 * Called just before entering the device driver to initiate a new disk I/O.
 * The buffer must be locked, thus, no I/O completion operations can occur
 * while we are manipulating its associated dependencies.
 *
 * bioops callback - hold io_token
 *
 * Parameters:
 *	bp:	structure describing disk write to occur
 */
static void 
softdep_disk_io_initiation(struct buf *bp)
{
	struct worklist *wk;
	struct worklist marker;
	struct indirdep *indirdep;

	/*
	 * We only care about write operations. There should never
	 * be dependencies for reads.
	 */
	if (bp->b_cmd == BUF_CMD_READ)
		panic("softdep_disk_io_initiation: read");

	get_mplock();
	marker.wk_type = D_LAST + 1;	/* Not a normal workitem */
	
	/*
	 * Do any necessary pre-I/O processing.
	 */
	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) {
		LIST_INSERT_AFTER(wk, &marker, wk_list);

		switch (wk->wk_type) {
		case D_PAGEDEP:
			initiate_write_filepage(WK_PAGEDEP(wk), bp);
			continue;

		case D_INODEDEP:
			initiate_write_inodeblock(WK_INODEDEP(wk), bp);
			continue;

		case D_INDIRDEP:
			indirdep = WK_INDIRDEP(wk);
			if (indirdep->ir_state & GOINGAWAY)
				panic("disk_io_initiation: indirdep gone");
			/*
			 * If there are no remaining dependencies, this
			 * will be writing the real pointers, so the
			 * dependency can be freed.
			 */
			if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
				indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
				brelse(indirdep->ir_savebp);
				/* inline expand WORKLIST_REMOVE(wk); */
				wk->wk_state &= ~ONWORKLIST;
				LIST_REMOVE(wk, wk_list);
				WORKITEM_FREE(indirdep, D_INDIRDEP);
				continue;
			}
			/*
			 * Replace up-to-date version with safe version.
			 */
			MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
			    M_INDIRDEP, M_SOFTDEP_FLAGS);
			ACQUIRE_LOCK(&lk);
			indirdep->ir_state &= ~ATTACHED;
			indirdep->ir_state |= UNDONE;
			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
			    bp->b_bcount);
			FREE_LOCK(&lk);
			continue;

		case D_MKDIR:
		case D_BMSAFEMAP:
		case D_ALLOCDIRECT:
		case D_ALLOCINDIR:
			continue;

		default:
			panic("handle_disk_io_initiation: Unexpected type %s",
			    TYPENAME(wk->wk_type));
			/* NOTREACHED */
		}
	}
	rel_mplock();
}

/*
 * Called from within the procedure above to deal with unsatisfied
 * allocation dependencies in a directory. The buffer must be locked,
 * thus, no I/O completion operations can occur while we are
 * manipulating its associated dependencies.
 */
static void
initiate_write_filepage(struct pagedep *pagedep, struct buf *bp)
{
	struct diradd *dap;
	struct direct *ep;
	int i;

	if (pagedep->pd_state & IOSTARTED) {
		/*
		 * This can only happen if there is a driver that does not
		 * understand chaining. Here biodone will reissue the call
		 * to strategy for the incomplete buffers.
		 */
		kprintf("initiate_write_filepage: already started\n");
		return;
	}
	pagedep->pd_state |= IOSTARTED;
	ACQUIRE_LOCK(&lk);
	for (i = 0; i < DAHASHSZ; i++) {
		LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
			ep = (struct direct *)
			    ((char *)bp->b_data + dap->da_offset);
			if (ep->d_ino != dap->da_newinum) {
				FREE_LOCK(&lk);
				panic("%s: dir inum %d != new %"PRId64,
				    "initiate_write_filepage",
				    ep->d_ino, dap->da_newinum);
			}
			if (dap->da_state & DIRCHG)
				ep->d_ino = dap->da_previous->dm_oldinum;
			else
				ep->d_ino = 0;
			dap->da_state &= ~ATTACHED;
			dap->da_state |= UNDONE;
		}
	}
	FREE_LOCK(&lk);
}

/*
 * Called from within the procedure above to deal with unsatisfied
 * allocation dependencies in an inodeblock. The buffer must be
 * locked, thus, no I/O completion operations can occur while we
 * are manipulating its associated dependencies.
 *
 * Parameters:
 *	bp:	The inode block
 */
static void 
initiate_write_inodeblock(struct inodedep *inodedep, struct buf *bp)
{
	struct allocdirect *adp, *lastadp;
	struct ufs1_dinode *dp;
	struct ufs1_dinode *sip;
	struct fs *fs;
	ufs_lbn_t prevlbn = 0;
	int i, deplist;

	if (inodedep->id_state & IOSTARTED)
		panic("initiate_write_inodeblock: already started");
	inodedep->id_state |= IOSTARTED;
	fs = inodedep->id_fs;
	dp = (struct ufs1_dinode *)bp->b_data +
	    ino_to_fsbo(fs, inodedep->id_ino);
	/*
	 * If the bitmap is not yet written, then the allocated
	 * inode cannot be written to disk.
	 */
	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
		if (inodedep->id_savedino != NULL)
			panic("initiate_write_inodeblock: already doing I/O");
		MALLOC(sip, struct ufs1_dinode *,
		    sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
		inodedep->id_savedino = sip;
		*inodedep->id_savedino = *dp;
		bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
		dp->di_gen = inodedep->id_savedino->di_gen;
		return;
	}
	/*
	 * If no dependencies, then there is nothing to roll back.
	 */
	inodedep->id_savedsize = dp->di_size;
	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
		return;
	/*
	 * Set the dependencies to busy.
	 */
	ACQUIRE_LOCK(&lk);
	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
	     adp = TAILQ_NEXT(adp, ad_next)) {
#ifdef DIAGNOSTIC
		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
			FREE_LOCK(&lk);
			panic("softdep_write_inodeblock: lbn order");
		}
		prevlbn = adp->ad_lbn;
		if (adp->ad_lbn < NDADDR &&
		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
			FREE_LOCK(&lk);
			panic("%s: direct pointer #%ld mismatch %d != %d",
			    "softdep_write_inodeblock", adp->ad_lbn,
			    dp->di_db[adp->ad_lbn], adp->ad_newblkno);
		}
		if (adp->ad_lbn >= NDADDR &&
		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
			FREE_LOCK(&lk);
			panic("%s: indirect pointer #%ld mismatch %d != %d",
			    "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
			    dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
		}
		deplist |= 1 << adp->ad_lbn;
		if ((adp->ad_state & ATTACHED) == 0) {
			FREE_LOCK(&lk);
			panic("softdep_write_inodeblock: Unknown state 0x%x",
			    adp->ad_state);
		}
#endif /* DIAGNOSTIC */
		adp->ad_state &= ~ATTACHED;
		adp->ad_state |= UNDONE;
	}
	/*
	 * The on-disk inode cannot claim to be any larger than the last
	 * fragment that has been written. Otherwise, the on-disk inode
	 * might have fragments that were not the last block in the file
	 * which would corrupt the filesystem.
	 */
	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
		if (adp->ad_lbn >= NDADDR)
			break;
		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
		/* keep going until hitting a rollback to a frag */
		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
			continue;
		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
#ifdef DIAGNOSTIC
			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
				FREE_LOCK(&lk);
				panic("softdep_write_inodeblock: lost dep1");
			}
#endif /* DIAGNOSTIC */
			dp->di_db[i] = 0;
		}
		for (i = 0; i < NIADDR; i++) {
#ifdef DIAGNOSTIC
			if (dp->di_ib[i] != 0 &&
			    (deplist & ((1 << NDADDR) << i)) == 0) {
				FREE_LOCK(&lk);
				panic("softdep_write_inodeblock: lost dep2");
			}
#endif /* DIAGNOSTIC */
			dp->di_ib[i] = 0;
		}
		FREE_LOCK(&lk);
		return;
	}
	/*
	 * If we have zero'ed out the last allocated block of the file,
	 * roll back the size to the last currently allocated block.
	 * We know that this last allocated block is a full-sized as
	 * we already checked for fragments in the loop above.
	 */
	if (lastadp != NULL &&
	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
		for (i = lastadp->ad_lbn; i >= 0; i--)
			if (dp->di_db[i] != 0)
				break;
		dp->di_size = (i + 1) * fs->fs_bsize;
	}
	/*
	 * The only dependencies are for indirect blocks.
	 *
	 * The file size for indirect block additions is not guaranteed.
	 * Such a guarantee would be non-trivial to achieve. The conventional
	 * synchronous write implementation also does not make this guarantee.
	 * Fsck should catch and fix discrepancies. Arguably, the file size
	 * can be over-estimated without destroying integrity when the file
	 * moves into the indirect blocks (i.e., is large). If we want to
	 * postpone fsck, we are stuck with this argument.
	 */
	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
	FREE_LOCK(&lk);
}

/*
 * This routine is called during the completion interrupt
 * service routine for a disk write (from the procedure called
 * by the device driver to inform the filesystem caches of
 * a request completion).  It should be called early in this
 * procedure, before the block is made available to other
 * processes or other routines are called.
 *
 * bioops callback - hold io_token
 *
 * Parameters:
 *	bp:	describes the completed disk write
 */
static void 
softdep_disk_write_complete(struct buf *bp)
{
	struct worklist *wk;
	struct workhead reattach;
	struct newblk *newblk;
	struct allocindir *aip;
	struct allocdirect *adp;
	struct indirdep *indirdep;
	struct inodedep *inodedep;
	struct bmsafemap *bmsafemap;

	get_mplock();
#ifdef DEBUG
	if (lk.lkt_held != NOHOLDER)
		panic("softdep_disk_write_complete: lock is held");
	lk.lkt_held = SPECIAL_FLAG;
#endif
	LIST_INIT(&reattach);
	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
		WORKLIST_REMOVE(wk);
		switch (wk->wk_type) {

		case D_PAGEDEP:
			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
				WORKLIST_INSERT(&reattach, wk);
			continue;

		case D_INODEDEP:
			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
				WORKLIST_INSERT(&reattach, wk);
			continue;

		case D_BMSAFEMAP:
			bmsafemap = WK_BMSAFEMAP(wk);
			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
				newblk->nb_state |= DEPCOMPLETE;
				newblk->nb_bmsafemap = NULL;
				LIST_REMOVE(newblk, nb_deps);
			}
			while ((adp =
			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
				adp->ad_state |= DEPCOMPLETE;
				adp->ad_buf = NULL;
				LIST_REMOVE(adp, ad_deps);
				handle_allocdirect_partdone(adp);
			}
			while ((aip =
			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
				aip->ai_state |= DEPCOMPLETE;
				aip->ai_buf = NULL;
				LIST_REMOVE(aip, ai_deps);
				handle_allocindir_partdone(aip);
			}
			while ((inodedep =
			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
				inodedep->id_state |= DEPCOMPLETE;
				LIST_REMOVE(inodedep, id_deps);
				inodedep->id_buf = NULL;
			}
			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
			continue;

		case D_MKDIR:
			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
			continue;

		case D_ALLOCDIRECT:
			adp = WK_ALLOCDIRECT(wk);
			adp->ad_state |= COMPLETE;
			handle_allocdirect_partdone(adp);
			continue;

		case D_ALLOCINDIR:
			aip = WK_ALLOCINDIR(wk);
			aip->ai_state |= COMPLETE;
			handle_allocindir_partdone(aip);
			continue;

		case D_INDIRDEP:
			indirdep = WK_INDIRDEP(wk);
			if (indirdep->ir_state & GOINGAWAY) {
				lk.lkt_held = NOHOLDER;
				panic("disk_write_complete: indirdep gone");
			}
			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
			FREE(indirdep->ir_saveddata, M_INDIRDEP);
			indirdep->ir_saveddata = 0;
			indirdep->ir_state &= ~UNDONE;
			indirdep->ir_state |= ATTACHED;
			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
				handle_allocindir_partdone(aip);
				if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
					lk.lkt_held = NOHOLDER;
					panic("disk_write_complete: not gone");
				}
			}
			WORKLIST_INSERT(&reattach, wk);
			if ((bp->b_flags & B_DELWRI) == 0)
				stat_indir_blk_ptrs++;
			bdirty(bp);
			continue;

		default:
			lk.lkt_held = NOHOLDER;
			panic("handle_disk_write_complete: Unknown type %s",
			    TYPENAME(wk->wk_type));
			/* NOTREACHED */
		}
	}
	/*
	 * Reattach any requests that must be redone.
	 */
	while ((wk = LIST_FIRST(&reattach)) != NULL) {
		WORKLIST_REMOVE(wk);
		WORKLIST_INSERT_BP(bp, wk);
	}
#ifdef DEBUG
	if (lk.lkt_held != SPECIAL_FLAG)
		panic("softdep_disk_write_complete: lock lost");
	lk.lkt_held = NOHOLDER;
#endif
	rel_mplock();
}

/*
 * Called from within softdep_disk_write_complete above. Note that
 * this routine is always called from interrupt level with further
 * splbio interrupts blocked.
 *
 * Parameters:
 *	adp:	the completed allocdirect
 */
static void 
handle_allocdirect_partdone(struct allocdirect *adp)
{
	struct allocdirect *listadp;
	struct inodedep *inodedep;
	long bsize;

	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
		return;
	if (adp->ad_buf != NULL) {
		lk.lkt_held = NOHOLDER;
		panic("handle_allocdirect_partdone: dangling dep");
	}
	/*
	 * The on-disk inode cannot claim to be any larger than the last
	 * fragment that has been written. Otherwise, the on-disk inode
	 * might have fragments that were not the last block in the file
	 * which would corrupt the filesystem. Thus, we cannot free any
	 * allocdirects after one whose ad_oldblkno claims a fragment as
	 * these blocks must be rolled back to zero before writing the inode.
	 * We check the currently active set of allocdirects in id_inoupdt.
	 */
	inodedep = adp->ad_inodedep;
	bsize = inodedep->id_fs->fs_bsize;
	TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) {
		/* found our block */
		if (listadp == adp)
			break;
		/* continue if ad_oldlbn is not a fragment */
		if (listadp->ad_oldsize == 0 ||
		    listadp->ad_oldsize == bsize)
			continue;
		/* hit a fragment */
		return;
	}
	/*
	 * If we have reached the end of the current list without
	 * finding the just finished dependency, then it must be
	 * on the future dependency list. Future dependencies cannot
	 * be freed until they are moved to the current list.
	 */
	if (listadp == NULL) {
#ifdef DEBUG
		TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next)
			/* found our block */
			if (listadp == adp)
				break;
		if (listadp == NULL) {
			lk.lkt_held = NOHOLDER;
			panic("handle_allocdirect_partdone: lost dep");
		}
#endif /* DEBUG */
		return;
	}
	/*
	 * If we have found the just finished dependency, then free
	 * it along with anything that follows it that is complete.
	 */
	for (; adp; adp = listadp) {
		listadp = TAILQ_NEXT(adp, ad_next);
		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
			return;
		free_allocdirect(&inodedep->id_inoupdt, adp, 1);
	}
}

/*
 * Called from within softdep_disk_write_complete above. Note that
 * this routine is always called from interrupt level with further