xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision daf1cffce2e07931f27c6c6998652e90df6ba87e)
1 /*
2  * Copyright 1998 Marshall Kirk McKusick. All Rights Reserved.
3  *
4  * The soft updates code is derived from the appendix of a University
5  * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
6  * "Soft Updates: A Solution to the Metadata Update Problem in File
7  * Systems", CSE-TR-254-95, August 1995).
8  *
9  * The following are the copyrights and redistribution conditions that
10  * apply to this copy of the soft update software. For a license
11  * to use, redistribute or sell the soft update software under
12  * conditions other than those described here, please contact the
13  * author at one of the following addresses:
14  *
15  *	Marshall Kirk McKusick		mckusick@mckusick.com
16  *	1614 Oxford Street		+1-510-843-9542
17  *	Berkeley, CA 94709-1608
18  *	USA
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  * 1. Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  * 2. Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in the
28  *    documentation and/or other materials provided with the distribution.
29  * 3. None of the names of McKusick, Ganger, Patt, or the University of
30  *    Michigan may be used to endorse or promote products derived from
31  *    this software without specific prior written permission.
32  * 4. Redistributions in any form must be accompanied by information on
33  *    how to obtain complete source code for any accompanying software
34  *    that uses this software. This source code must either be included
35  *    in the distribution or be available for no more than the cost of
36  *    distribution plus a nominal fee, and must be freely redistributable
37  *    under reasonable conditions. For an executable file, complete
38  *    source code means the source code for all modules it contains.
39  *    It does not mean source code for modules or files that typically
40  *    accompany the operating system on which the executable file runs,
41  *    e.g., standard library modules or system header files.
42  *
43  * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
44  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
45  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
46  * DISCLAIMED.  IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
47  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
48  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
49  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
50  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
51  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
52  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
53  * SUCH DAMAGE.
54  *
55  *	from: @(#)ffs_softdep.c	9.56 (McKusick) 1/17/00
56  * $FreeBSD$
57  */
58 
59 /*
60  * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
61  */
62 #ifndef DIAGNOSTIC
63 #define DIAGNOSTIC
64 #endif
65 #ifndef DEBUG
66 #define DEBUG
67 #endif
68 
69 #include <sys/param.h>
70 #include <sys/kernel.h>
71 #include <sys/systm.h>
72 #include <sys/buf.h>
73 #include <sys/malloc.h>
74 #include <sys/mount.h>
75 #include <sys/proc.h>
76 #include <sys/syslog.h>
77 #include <sys/vnode.h>
78 #include <sys/conf.h>
79 #include <ufs/ufs/dir.h>
80 #include <ufs/ufs/quota.h>
81 #include <ufs/ufs/inode.h>
82 #include <ufs/ufs/ufsmount.h>
83 #include <ufs/ffs/fs.h>
84 #include <ufs/ffs/softdep.h>
85 #include <ufs/ffs/ffs_extern.h>
86 #include <ufs/ufs/ufs_extern.h>
87 
88 /*
89  * These definitions need to be adapted to the system to which
90  * this file is being ported.
91  */
92 /*
93  * malloc types defined for the softdep system.
94  */
95 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
96 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
97 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
98 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
99 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
100 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
101 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
102 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
103 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
104 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
105 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
106 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
107 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
108 
109 #define	D_PAGEDEP	0
110 #define	D_INODEDEP	1
111 #define	D_NEWBLK	2
112 #define	D_BMSAFEMAP	3
113 #define	D_ALLOCDIRECT	4
114 #define	D_INDIRDEP	5
115 #define	D_ALLOCINDIR	6
116 #define	D_FREEFRAG	7
117 #define	D_FREEBLKS	8
118 #define	D_FREEFILE	9
119 #define	D_DIRADD	10
120 #define	D_MKDIR		11
121 #define	D_DIRREM	12
122 #define D_LAST		D_DIRREM
123 
124 /*
125  * translate from workitem type to memory type
126  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
127  */
128 static struct malloc_type *memtype[] = {
129 	M_PAGEDEP,
130 	M_INODEDEP,
131 	M_NEWBLK,
132 	M_BMSAFEMAP,
133 	M_ALLOCDIRECT,
134 	M_INDIRDEP,
135 	M_ALLOCINDIR,
136 	M_FREEFRAG,
137 	M_FREEBLKS,
138 	M_FREEFILE,
139 	M_DIRADD,
140 	M_MKDIR,
141 	M_DIRREM
142 };
143 
144 #define DtoM(type) (memtype[type])
145 
146 /*
147  * Names of malloc types.
148  */
149 #define TYPENAME(type)  \
150 	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
151 #define CURPROC curproc
152 /*
153  * End system adaptaion definitions.
154  */
155 
156 /*
157  * Internal function prototypes.
158  */
159 static	void softdep_error __P((char *, int));
160 static	void drain_output __P((struct vnode *, int));
161 static	int getdirtybuf __P((struct buf **, int));
162 static	void clear_remove __P((struct proc *));
163 static	void clear_inodedeps __P((struct proc *));
164 static	int flush_pagedep_deps __P((struct vnode *, struct mount *,
165 	    struct diraddhd *));
166 static	int flush_inodedep_deps __P((struct fs *, ino_t));
167 static	int handle_written_filepage __P((struct pagedep *, struct buf *));
168 static  void diradd_inode_written __P((struct diradd *, struct inodedep *));
169 static	int handle_written_inodeblock __P((struct inodedep *, struct buf *));
170 static	void handle_allocdirect_partdone __P((struct allocdirect *));
171 static	void handle_allocindir_partdone __P((struct allocindir *));
172 static	void initiate_write_filepage __P((struct pagedep *, struct buf *));
173 static	void handle_written_mkdir __P((struct mkdir *, int));
174 static	void initiate_write_inodeblock __P((struct inodedep *, struct buf *));
175 static	void handle_workitem_freefile __P((struct freefile *));
176 static	void handle_workitem_remove __P((struct dirrem *));
177 static	struct dirrem *newdirrem __P((struct buf *, struct inode *,
178 	    struct inode *, int, struct dirrem **));
179 static	void free_diradd __P((struct diradd *));
180 static	void free_allocindir __P((struct allocindir *, struct inodedep *));
181 static	int indir_trunc __P((struct inode *, ufs_daddr_t, int, ufs_lbn_t,
182 	    long *));
183 static	void deallocate_dependencies __P((struct buf *, struct inodedep *));
184 static	void free_allocdirect __P((struct allocdirectlst *,
185 	    struct allocdirect *, int));
186 static	int check_inode_unwritten __P((struct inodedep *));
187 static	int free_inodedep __P((struct inodedep *));
188 static	void handle_workitem_freeblocks __P((struct freeblks *));
189 static	void merge_inode_lists __P((struct inodedep *));
190 static	void setup_allocindir_phase2 __P((struct buf *, struct inode *,
191 	    struct allocindir *));
192 static	struct allocindir *newallocindir __P((struct inode *, int, ufs_daddr_t,
193 	    ufs_daddr_t));
194 static	void handle_workitem_freefrag __P((struct freefrag *));
195 static	struct freefrag *newfreefrag __P((struct inode *, ufs_daddr_t, long));
196 static	void allocdirect_merge __P((struct allocdirectlst *,
197 	    struct allocdirect *, struct allocdirect *));
198 static	struct bmsafemap *bmsafemap_lookup __P((struct buf *));
199 static	int newblk_lookup __P((struct fs *, ufs_daddr_t, int,
200 	    struct newblk **));
201 static	int inodedep_lookup __P((struct fs *, ino_t, int, struct inodedep **));
202 static	int pagedep_lookup __P((struct inode *, ufs_lbn_t, int,
203 	    struct pagedep **));
204 static	void pause_timer __P((void *));
205 static	int request_cleanup __P((int, int));
206 static	void add_to_worklist __P((struct worklist *));
207 
208 /*
209  * Exported softdep operations.
210  */
211 static	void softdep_disk_io_initiation __P((struct buf *));
212 static	void softdep_disk_write_complete __P((struct buf *));
213 static	void softdep_deallocate_dependencies __P((struct buf *));
214 static	int softdep_fsync __P((struct vnode *));
215 static	int softdep_process_worklist __P((struct mount *));
216 static	void softdep_move_dependencies __P((struct buf *, struct buf *));
217 static	int softdep_count_dependencies __P((struct buf *bp, int));
218 
219 struct bio_ops bioops = {
220 	softdep_disk_io_initiation,		/* io_start */
221 	softdep_disk_write_complete,		/* io_complete */
222 	softdep_deallocate_dependencies,	/* io_deallocate */
223 	softdep_fsync,				/* io_fsync */
224 	softdep_process_worklist,		/* io_sync */
225 	softdep_move_dependencies,		/* io_movedeps */
226 	softdep_count_dependencies,		/* io_countdeps */
227 };
228 
229 /*
230  * Locking primitives.
231  *
232  * For a uniprocessor, all we need to do is protect against disk
233  * interrupts. For a multiprocessor, this lock would have to be
234  * a mutex. A single mutex is used throughout this file, though
235  * finer grain locking could be used if contention warranted it.
236  *
237  * For a multiprocessor, the sleep call would accept a lock and
238  * release it after the sleep processing was complete. In a uniprocessor
239  * implementation there is no such interlock, so we simple mark
240  * the places where it needs to be done with the `interlocked' form
241  * of the lock calls. Since the uniprocessor sleep already interlocks
242  * the spl, there is nothing that really needs to be done.
243  */
244 #ifndef /* NOT */ DEBUG
245 static struct lockit {
246 	int	lkt_spl;
247 } lk = { 0 };
248 #define ACQUIRE_LOCK(lk)		(lk)->lkt_spl = splbio()
249 #define FREE_LOCK(lk)			splx((lk)->lkt_spl)
250 #define ACQUIRE_LOCK_INTERLOCKED(lk)
251 #define FREE_LOCK_INTERLOCKED(lk)
252 
253 #else /* DEBUG */
254 static struct lockit {
255 	int	lkt_spl;
256 	pid_t	lkt_held;
257 } lk = { 0, -1 };
258 static int lockcnt;
259 
260 static	void acquire_lock __P((struct lockit *));
261 static	void free_lock __P((struct lockit *));
262 static	void acquire_lock_interlocked __P((struct lockit *));
263 static	void free_lock_interlocked __P((struct lockit *));
264 
265 #define ACQUIRE_LOCK(lk)		acquire_lock(lk)
266 #define FREE_LOCK(lk)			free_lock(lk)
267 #define ACQUIRE_LOCK_INTERLOCKED(lk)	acquire_lock_interlocked(lk)
268 #define FREE_LOCK_INTERLOCKED(lk)	free_lock_interlocked(lk)
269 
270 static void
271 acquire_lock(lk)
272 	struct lockit *lk;
273 {
274 
275 	if (lk->lkt_held != -1) {
276 		if (lk->lkt_held == CURPROC->p_pid)
277 			panic("softdep_lock: locking against myself");
278 		else
279 			panic("softdep_lock: lock held by %d", lk->lkt_held);
280 	}
281 	lk->lkt_spl = splbio();
282 	lk->lkt_held = CURPROC->p_pid;
283 	lockcnt++;
284 }
285 
286 static void
287 free_lock(lk)
288 	struct lockit *lk;
289 {
290 
291 	if (lk->lkt_held == -1)
292 		panic("softdep_unlock: lock not held");
293 	lk->lkt_held = -1;
294 	splx(lk->lkt_spl);
295 }
296 
297 static void
298 acquire_lock_interlocked(lk)
299 	struct lockit *lk;
300 {
301 
302 	if (lk->lkt_held != -1) {
303 		if (lk->lkt_held == CURPROC->p_pid)
304 			panic("softdep_lock_interlocked: locking against self");
305 		else
306 			panic("softdep_lock_interlocked: lock held by %d",
307 			    lk->lkt_held);
308 	}
309 	lk->lkt_held = CURPROC->p_pid;
310 	lockcnt++;
311 }
312 
313 static void
314 free_lock_interlocked(lk)
315 	struct lockit *lk;
316 {
317 
318 	if (lk->lkt_held == -1)
319 		panic("softdep_unlock_interlocked: lock not held");
320 	lk->lkt_held = -1;
321 }
322 #endif /* DEBUG */
323 
324 /*
325  * Place holder for real semaphores.
326  */
327 struct sema {
328 	int	value;
329 	pid_t	holder;
330 	char	*name;
331 	int	prio;
332 	int	timo;
333 };
334 static	void sema_init __P((struct sema *, char *, int, int));
335 static	int sema_get __P((struct sema *, struct lockit *));
336 static	void sema_release __P((struct sema *));
337 
338 static void
339 sema_init(semap, name, prio, timo)
340 	struct sema *semap;
341 	char *name;
342 	int prio, timo;
343 {
344 
345 	semap->holder = -1;
346 	semap->value = 0;
347 	semap->name = name;
348 	semap->prio = prio;
349 	semap->timo = timo;
350 }
351 
352 static int
353 sema_get(semap, interlock)
354 	struct sema *semap;
355 	struct lockit *interlock;
356 {
357 
358 	if (semap->value++ > 0) {
359 		if (interlock != NULL)
360 			FREE_LOCK_INTERLOCKED(interlock);
361 		tsleep((caddr_t)semap, semap->prio, semap->name, semap->timo);
362 		if (interlock != NULL) {
363 			ACQUIRE_LOCK_INTERLOCKED(interlock);
364 			FREE_LOCK(interlock);
365 		}
366 		return (0);
367 	}
368 	semap->holder = CURPROC->p_pid;
369 	if (interlock != NULL)
370 		FREE_LOCK(interlock);
371 	return (1);
372 }
373 
374 static void
375 sema_release(semap)
376 	struct sema *semap;
377 {
378 
379 	if (semap->value <= 0 || semap->holder != CURPROC->p_pid)
380 		panic("sema_release: not held");
381 	if (--semap->value > 0) {
382 		semap->value = 0;
383 		wakeup(semap);
384 	}
385 	semap->holder = -1;
386 }
387 
388 /*
389  * Worklist queue management.
390  * These routines require that the lock be held.
391  */
392 #ifndef /* NOT */ DEBUG
393 #define WORKLIST_INSERT(head, item) do {	\
394 	(item)->wk_state |= ONWORKLIST;		\
395 	LIST_INSERT_HEAD(head, item, wk_list);	\
396 } while (0)
397 #define WORKLIST_REMOVE(item) do {		\
398 	(item)->wk_state &= ~ONWORKLIST;	\
399 	LIST_REMOVE(item, wk_list);		\
400 } while (0)
401 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
402 
403 #else /* DEBUG */
404 static	void worklist_insert __P((struct workhead *, struct worklist *));
405 static	void worklist_remove __P((struct worklist *));
406 static	void workitem_free __P((struct worklist *, int));
407 
408 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
409 #define WORKLIST_REMOVE(item) worklist_remove(item)
410 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
411 
412 static void
413 worklist_insert(head, item)
414 	struct workhead *head;
415 	struct worklist *item;
416 {
417 
418 	if (lk.lkt_held == -1)
419 		panic("worklist_insert: lock not held");
420 	if (item->wk_state & ONWORKLIST)
421 		panic("worklist_insert: already on list");
422 	item->wk_state |= ONWORKLIST;
423 	LIST_INSERT_HEAD(head, item, wk_list);
424 }
425 
426 static void
427 worklist_remove(item)
428 	struct worklist *item;
429 {
430 
431 	if (lk.lkt_held == -1)
432 		panic("worklist_remove: lock not held");
433 	if ((item->wk_state & ONWORKLIST) == 0)
434 		panic("worklist_remove: not on list");
435 	item->wk_state &= ~ONWORKLIST;
436 	LIST_REMOVE(item, wk_list);
437 }
438 
439 static void
440 workitem_free(item, type)
441 	struct worklist *item;
442 	int type;
443 {
444 
445 	if (item->wk_state & ONWORKLIST)
446 		panic("workitem_free: still on list");
447 	if (item->wk_type != type)
448 		panic("workitem_free: type mismatch");
449 	FREE(item, DtoM(type));
450 }
451 #endif /* DEBUG */
452 
453 /*
454  * Workitem queue management
455  */
456 static struct workhead softdep_workitem_pending;
457 static int softdep_worklist_busy;
458 static int max_softdeps;	/* maximum number of structs before slowdown */
459 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
460 static int proc_waiting;	/* tracks whether we have a timeout posted */
461 static struct proc *filesys_syncer; /* proc of filesystem syncer process */
462 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
463 #define FLUSH_INODES	1
464 static int req_clear_remove;	/* syncer process flush some freeblks */
465 #define FLUSH_REMOVE	2
466 /*
467  * runtime statistics
468  */
469 static int stat_blk_limit_push;	/* number of times block limit neared */
470 static int stat_ino_limit_push;	/* number of times inode limit neared */
471 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
472 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
473 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
474 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
475 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
476 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
477 #ifdef DEBUG
478 #include <vm/vm.h>
479 #include <sys/sysctl.h>
480 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
481 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
482 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
483 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
484 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
485 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
486 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
487 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
488 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
489 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
490 #endif /* DEBUG */
491 
492 /*
493  * Add an item to the end of the work queue.
494  * This routine requires that the lock be held.
495  * This is the only routine that adds items to the list.
496  * The following routine is the only one that removes items
497  * and does so in order from first to last.
498  */
499 static void
500 add_to_worklist(wk)
501 	struct worklist *wk;
502 {
503 	static struct worklist *worklist_tail;
504 
505 	if (wk->wk_state & ONWORKLIST)
506 		panic("add_to_worklist: already on list");
507 	wk->wk_state |= ONWORKLIST;
508 	if (LIST_FIRST(&softdep_workitem_pending) == NULL)
509 		LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
510 	else
511 		LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
512 	worklist_tail = wk;
513 }
514 
515 /*
516  * Process that runs once per second to handle items in the background queue.
517  *
518  * Note that we ensure that everything is done in the order in which they
519  * appear in the queue. The code below depends on this property to ensure
520  * that blocks of a file are freed before the inode itself is freed. This
521  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
522  * until all the old ones have been purged from the dependency lists.
523  */
524 static int
525 softdep_process_worklist(matchmnt)
526 	struct mount *matchmnt;
527 {
528 	struct proc *p = CURPROC;
529 	struct worklist *wk;
530 	struct fs *matchfs;
531 	int matchcnt, loopcount;
532 
533 	/*
534 	 * Record the process identifier of our caller so that we can give
535 	 * this process preferential treatment in request_cleanup below.
536 	 */
537 	filesys_syncer = p;
538 	matchcnt = 0;
539 	matchfs = NULL;
540 	if (matchmnt != NULL)
541 		matchfs = VFSTOUFS(matchmnt)->um_fs;
542 	/*
543 	 * There is no danger of having multiple processes run this
544 	 * code. It is single threaded solely so that softdep_flushfiles
545 	 * (below) can get an accurate count of the number of items
546 	 * related to its mount point that are in the list.
547 	 */
548 	if (softdep_worklist_busy && matchmnt == NULL)
549 		return (-1);
550 	/*
551 	 * If requested, try removing inode or removal dependencies.
552 	 */
553 	if (req_clear_inodedeps) {
554 		clear_inodedeps(p);
555 		req_clear_inodedeps = 0;
556 		wakeup(&proc_waiting);
557 	}
558 	if (req_clear_remove) {
559 		clear_remove(p);
560 		req_clear_remove = 0;
561 		wakeup(&proc_waiting);
562 	}
563 	ACQUIRE_LOCK(&lk);
564 	loopcount = 1;
565 	while ((wk = LIST_FIRST(&softdep_workitem_pending)) != 0) {
566 		WORKLIST_REMOVE(wk);
567 		FREE_LOCK(&lk);
568 		switch (wk->wk_type) {
569 
570 		case D_DIRREM:
571 			/* removal of a directory entry */
572 			if (WK_DIRREM(wk)->dm_mnt == matchmnt)
573 				matchcnt += 1;
574 			handle_workitem_remove(WK_DIRREM(wk));
575 			break;
576 
577 		case D_FREEBLKS:
578 			/* releasing blocks and/or fragments from a file */
579 			if (WK_FREEBLKS(wk)->fb_fs == matchfs)
580 				matchcnt += 1;
581 			handle_workitem_freeblocks(WK_FREEBLKS(wk));
582 			break;
583 
584 		case D_FREEFRAG:
585 			/* releasing a fragment when replaced as a file grows */
586 			if (WK_FREEFRAG(wk)->ff_fs == matchfs)
587 				matchcnt += 1;
588 			handle_workitem_freefrag(WK_FREEFRAG(wk));
589 			break;
590 
591 		case D_FREEFILE:
592 			/* releasing an inode when its link count drops to 0 */
593 			if (WK_FREEFILE(wk)->fx_fs == matchfs)
594 				matchcnt += 1;
595 			handle_workitem_freefile(WK_FREEFILE(wk));
596 			break;
597 
598 		default:
599 			panic("%s_process_worklist: Unknown type %s",
600 			    "softdep", TYPENAME(wk->wk_type));
601 			/* NOTREACHED */
602 		}
603 		if (softdep_worklist_busy && matchmnt == NULL)
604 			return (-1);
605 		/*
606 		 * If requested, try removing inode or removal dependencies.
607 		 */
608 		if (req_clear_inodedeps) {
609 			clear_inodedeps(p);
610 			req_clear_inodedeps = 0;
611 			wakeup(&proc_waiting);
612 		}
613 		if (req_clear_remove) {
614 			clear_remove(p);
615 			req_clear_remove = 0;
616 			wakeup(&proc_waiting);
617 		}
618 		/*
619 		 * We do not generally want to stop for buffer space, but if
620 		 * we are really being a buffer hog, we will stop and wait.
621 		 */
622 		if (loopcount++ % 128 == 0)
623 			bwillwrite();
624 		ACQUIRE_LOCK(&lk);
625 	}
626 	FREE_LOCK(&lk);
627 	return (matchcnt);
628 }
629 
630 /*
631  * Move dependencies from one buffer to another.
632  */
633 static void
634 softdep_move_dependencies(oldbp, newbp)
635 	struct buf *oldbp;
636 	struct buf *newbp;
637 {
638 	struct worklist *wk, *wktail;
639 
640 	if (LIST_FIRST(&newbp->b_dep) != NULL)
641 		panic("softdep_move_dependencies: need merge code");
642 	wktail = 0;
643 	ACQUIRE_LOCK(&lk);
644 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
645 		LIST_REMOVE(wk, wk_list);
646 		if (wktail == 0)
647 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
648 		else
649 			LIST_INSERT_AFTER(wktail, wk, wk_list);
650 		wktail = wk;
651 	}
652 	FREE_LOCK(&lk);
653 }
654 
655 /*
656  * Purge the work list of all items associated with a particular mount point.
657  */
658 int
659 softdep_flushfiles(oldmnt, flags, p)
660 	struct mount *oldmnt;
661 	int flags;
662 	struct proc *p;
663 {
664 	struct vnode *devvp;
665 	int error, loopcnt;
666 
667 	/*
668 	 * Await our turn to clear out the queue.
669 	 */
670 	while (softdep_worklist_busy)
671 		tsleep(&lbolt, PRIBIO, "softflush", 0);
672 	softdep_worklist_busy = 1;
673 	if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0) {
674 		softdep_worklist_busy = 0;
675 		return (error);
676 	}
677 	/*
678 	 * Alternately flush the block device associated with the mount
679 	 * point and process any dependencies that the flushing
680 	 * creates. In theory, this loop can happen at most twice,
681 	 * but we give it a few extra just to be sure.
682 	 */
683 	devvp = VFSTOUFS(oldmnt)->um_devvp;
684 	for (loopcnt = 10; loopcnt > 0; ) {
685 		if (softdep_process_worklist(oldmnt) == 0) {
686 			loopcnt--;
687 			/*
688 			 * Do another flush in case any vnodes were brought in
689 			 * as part of the cleanup operations.
690 			 */
691 			if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0)
692 				break;
693 			/*
694 			 * If we still found nothing to do, we are really done.
695 			 */
696 			if (softdep_process_worklist(oldmnt) == 0)
697 				break;
698 		}
699 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, p);
700 		error = VOP_FSYNC(devvp, p->p_ucred, MNT_WAIT, p);
701 		VOP_UNLOCK(devvp, 0, p);
702 		if (error)
703 			break;
704 	}
705 	softdep_worklist_busy = 0;
706 	/*
707 	 * If we are unmounting then it is an error to fail. If we
708 	 * are simply trying to downgrade to read-only, then filesystem
709 	 * activity can keep us busy forever, so we just fail with EBUSY.
710 	 */
711 	if (loopcnt == 0) {
712 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
713 			panic("softdep_flushfiles: looping");
714 		error = EBUSY;
715 	}
716 	return (error);
717 }
718 
719 /*
720  * Structure hashing.
721  *
722  * There are three types of structures that can be looked up:
723  *	1) pagedep structures identified by mount point, inode number,
724  *	   and logical block.
725  *	2) inodedep structures identified by mount point and inode number.
726  *	3) newblk structures identified by mount point and
727  *	   physical block number.
728  *
729  * The "pagedep" and "inodedep" dependency structures are hashed
730  * separately from the file blocks and inodes to which they correspond.
731  * This separation helps when the in-memory copy of an inode or
732  * file block must be replaced. It also obviates the need to access
733  * an inode or file page when simply updating (or de-allocating)
734  * dependency structures. Lookup of newblk structures is needed to
735  * find newly allocated blocks when trying to associate them with
736  * their allocdirect or allocindir structure.
737  *
738  * The lookup routines optionally create and hash a new instance when
739  * an existing entry is not found.
740  */
741 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
742 
743 /*
744  * Structures and routines associated with pagedep caching.
745  */
746 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
747 u_long	pagedep_hash;		/* size of hash table - 1 */
748 #define	PAGEDEP_HASH(mp, inum, lbn) \
749 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
750 	    pagedep_hash])
751 static struct sema pagedep_in_progress;
752 
753 /*
754  * Look up a pagedep. Return 1 if found, 0 if not found.
755  * If not found, allocate if DEPALLOC flag is passed.
756  * Found or allocated entry is returned in pagedeppp.
757  * This routine must be called with splbio interrupts blocked.
758  */
759 static int
760 pagedep_lookup(ip, lbn, flags, pagedeppp)
761 	struct inode *ip;
762 	ufs_lbn_t lbn;
763 	int flags;
764 	struct pagedep **pagedeppp;
765 {
766 	struct pagedep *pagedep;
767 	struct pagedep_hashhead *pagedephd;
768 	struct mount *mp;
769 	int i;
770 
771 #ifdef DEBUG
772 	if (lk.lkt_held == -1)
773 		panic("pagedep_lookup: lock not held");
774 #endif
775 	mp = ITOV(ip)->v_mount;
776 	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
777 top:
778 	for (pagedep = LIST_FIRST(pagedephd); pagedep;
779 	     pagedep = LIST_NEXT(pagedep, pd_hash))
780 		if (ip->i_number == pagedep->pd_ino &&
781 		    lbn == pagedep->pd_lbn &&
782 		    mp == pagedep->pd_mnt)
783 			break;
784 	if (pagedep) {
785 		*pagedeppp = pagedep;
786 		return (1);
787 	}
788 	if ((flags & DEPALLOC) == 0) {
789 		*pagedeppp = NULL;
790 		return (0);
791 	}
792 	if (sema_get(&pagedep_in_progress, &lk) == 0) {
793 		ACQUIRE_LOCK(&lk);
794 		goto top;
795 	}
796 	MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
797 		M_WAITOK);
798 	bzero(pagedep, sizeof(struct pagedep));
799 	pagedep->pd_list.wk_type = D_PAGEDEP;
800 	pagedep->pd_mnt = mp;
801 	pagedep->pd_ino = ip->i_number;
802 	pagedep->pd_lbn = lbn;
803 	LIST_INIT(&pagedep->pd_dirremhd);
804 	LIST_INIT(&pagedep->pd_pendinghd);
805 	for (i = 0; i < DAHASHSZ; i++)
806 		LIST_INIT(&pagedep->pd_diraddhd[i]);
807 	ACQUIRE_LOCK(&lk);
808 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
809 	sema_release(&pagedep_in_progress);
810 	*pagedeppp = pagedep;
811 	return (0);
812 }
813 
814 /*
815  * Structures and routines associated with inodedep caching.
816  */
817 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
818 static u_long	inodedep_hash;	/* size of hash table - 1 */
819 static long	num_inodedep;	/* number of inodedep allocated */
820 #define	INODEDEP_HASH(fs, inum) \
821       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
822 static struct sema inodedep_in_progress;
823 
824 /*
825  * Look up a inodedep. Return 1 if found, 0 if not found.
826  * If not found, allocate if DEPALLOC flag is passed.
827  * Found or allocated entry is returned in inodedeppp.
828  * This routine must be called with splbio interrupts blocked.
829  */
830 static int
831 inodedep_lookup(fs, inum, flags, inodedeppp)
832 	struct fs *fs;
833 	ino_t inum;
834 	int flags;
835 	struct inodedep **inodedeppp;
836 {
837 	struct inodedep *inodedep;
838 	struct inodedep_hashhead *inodedephd;
839 	int firsttry;
840 
841 #ifdef DEBUG
842 	if (lk.lkt_held == -1)
843 		panic("inodedep_lookup: lock not held");
844 #endif
845 	firsttry = 1;
846 	inodedephd = INODEDEP_HASH(fs, inum);
847 top:
848 	for (inodedep = LIST_FIRST(inodedephd); inodedep;
849 	     inodedep = LIST_NEXT(inodedep, id_hash))
850 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
851 			break;
852 	if (inodedep) {
853 		*inodedeppp = inodedep;
854 		return (1);
855 	}
856 	if ((flags & DEPALLOC) == 0) {
857 		*inodedeppp = NULL;
858 		return (0);
859 	}
860 	/*
861 	 * If we are over our limit, try to improve the situation.
862 	 */
863 	if (num_inodedep > max_softdeps && firsttry && speedup_syncer() == 0 &&
864 	    request_cleanup(FLUSH_INODES, 1)) {
865 		firsttry = 0;
866 		goto top;
867 	}
868 	if (sema_get(&inodedep_in_progress, &lk) == 0) {
869 		ACQUIRE_LOCK(&lk);
870 		goto top;
871 	}
872 	num_inodedep += 1;
873 	MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
874 		M_INODEDEP, M_WAITOK);
875 	inodedep->id_list.wk_type = D_INODEDEP;
876 	inodedep->id_fs = fs;
877 	inodedep->id_ino = inum;
878 	inodedep->id_state = ALLCOMPLETE;
879 	inodedep->id_nlinkdelta = 0;
880 	inodedep->id_savedino = NULL;
881 	inodedep->id_savedsize = -1;
882 	inodedep->id_buf = NULL;
883 	LIST_INIT(&inodedep->id_pendinghd);
884 	LIST_INIT(&inodedep->id_inowait);
885 	LIST_INIT(&inodedep->id_bufwait);
886 	TAILQ_INIT(&inodedep->id_inoupdt);
887 	TAILQ_INIT(&inodedep->id_newinoupdt);
888 	ACQUIRE_LOCK(&lk);
889 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
890 	sema_release(&inodedep_in_progress);
891 	*inodedeppp = inodedep;
892 	return (0);
893 }
894 
895 /*
896  * Structures and routines associated with newblk caching.
897  */
898 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
899 u_long	newblk_hash;		/* size of hash table - 1 */
900 #define	NEWBLK_HASH(fs, inum) \
901 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
902 static struct sema newblk_in_progress;
903 
904 /*
905  * Look up a newblk. Return 1 if found, 0 if not found.
906  * If not found, allocate if DEPALLOC flag is passed.
907  * Found or allocated entry is returned in newblkpp.
908  */
909 static int
910 newblk_lookup(fs, newblkno, flags, newblkpp)
911 	struct fs *fs;
912 	ufs_daddr_t newblkno;
913 	int flags;
914 	struct newblk **newblkpp;
915 {
916 	struct newblk *newblk;
917 	struct newblk_hashhead *newblkhd;
918 
919 	newblkhd = NEWBLK_HASH(fs, newblkno);
920 top:
921 	for (newblk = LIST_FIRST(newblkhd); newblk;
922 	     newblk = LIST_NEXT(newblk, nb_hash))
923 		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
924 			break;
925 	if (newblk) {
926 		*newblkpp = newblk;
927 		return (1);
928 	}
929 	if ((flags & DEPALLOC) == 0) {
930 		*newblkpp = NULL;
931 		return (0);
932 	}
933 	if (sema_get(&newblk_in_progress, 0) == 0)
934 		goto top;
935 	MALLOC(newblk, struct newblk *, sizeof(struct newblk),
936 		M_NEWBLK, M_WAITOK);
937 	newblk->nb_state = 0;
938 	newblk->nb_fs = fs;
939 	newblk->nb_newblkno = newblkno;
940 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
941 	sema_release(&newblk_in_progress);
942 	*newblkpp = newblk;
943 	return (0);
944 }
945 
946 /*
947  * Executed during filesystem system initialization before
948  * mounting any file systems.
949  */
950 void
951 softdep_initialize()
952 {
953 
954 	LIST_INIT(&mkdirlisthd);
955 	LIST_INIT(&softdep_workitem_pending);
956 	max_softdeps = desiredvnodes * 8;
957 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
958 	    &pagedep_hash);
959 	sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0);
960 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
961 	sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0);
962 	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
963 	sema_init(&newblk_in_progress, "newblk", PRIBIO, 0);
964 }
965 
966 /*
967  * Called at mount time to notify the dependency code that a
968  * filesystem wishes to use it.
969  */
970 int
971 softdep_mount(devvp, mp, fs, cred)
972 	struct vnode *devvp;
973 	struct mount *mp;
974 	struct fs *fs;
975 	struct ucred *cred;
976 {
977 	struct csum cstotal;
978 	struct cg *cgp;
979 	struct buf *bp;
980 	int error, cyl;
981 
982 	mp->mnt_flag &= ~MNT_ASYNC;
983 	mp->mnt_flag |= MNT_SOFTDEP;
984 	/*
985 	 * When doing soft updates, the counters in the
986 	 * superblock may have gotten out of sync, so we have
987 	 * to scan the cylinder groups and recalculate them.
988 	 */
989 	if (fs->fs_clean != 0)
990 		return (0);
991 	bzero(&cstotal, sizeof cstotal);
992 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
993 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
994 		    fs->fs_cgsize, cred, &bp)) != 0) {
995 			brelse(bp);
996 			return (error);
997 		}
998 		cgp = (struct cg *)bp->b_data;
999 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1000 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1001 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1002 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1003 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
1004 		brelse(bp);
1005 	}
1006 #ifdef DEBUG
1007 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1008 		printf("ffs_mountfs: superblock updated for soft updates\n");
1009 #endif
1010 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1011 	return (0);
1012 }
1013 
1014 /*
1015  * Protecting the freemaps (or bitmaps).
1016  *
1017  * To eliminate the need to execute fsck before mounting a file system
1018  * after a power failure, one must (conservatively) guarantee that the
1019  * on-disk copy of the bitmaps never indicate that a live inode or block is
1020  * free.  So, when a block or inode is allocated, the bitmap should be
1021  * updated (on disk) before any new pointers.  When a block or inode is
1022  * freed, the bitmap should not be updated until all pointers have been
1023  * reset.  The latter dependency is handled by the delayed de-allocation
1024  * approach described below for block and inode de-allocation.  The former
1025  * dependency is handled by calling the following procedure when a block or
1026  * inode is allocated. When an inode is allocated an "inodedep" is created
1027  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1028  * Each "inodedep" is also inserted into the hash indexing structure so
1029  * that any additional link additions can be made dependent on the inode
1030  * allocation.
1031  *
1032  * The ufs file system maintains a number of free block counts (e.g., per
1033  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1034  * in addition to the bitmaps.  These counts are used to improve efficiency
1035  * during allocation and therefore must be consistent with the bitmaps.
1036  * There is no convenient way to guarantee post-crash consistency of these
1037  * counts with simple update ordering, for two main reasons: (1) The counts
1038  * and bitmaps for a single cylinder group block are not in the same disk
1039  * sector.  If a disk write is interrupted (e.g., by power failure), one may
1040  * be written and the other not.  (2) Some of the counts are located in the
1041  * superblock rather than the cylinder group block. So, we focus our soft
1042  * updates implementation on protecting the bitmaps. When mounting a
1043  * filesystem, we recompute the auxiliary counts from the bitmaps.
1044  */
1045 
1046 /*
1047  * Called just after updating the cylinder group block to allocate an inode.
1048  */
1049 void
1050 softdep_setup_inomapdep(bp, ip, newinum)
1051 	struct buf *bp;		/* buffer for cylgroup block with inode map */
1052 	struct inode *ip;	/* inode related to allocation */
1053 	ino_t newinum;		/* new inode number being allocated */
1054 {
1055 	struct inodedep *inodedep;
1056 	struct bmsafemap *bmsafemap;
1057 
1058 	/*
1059 	 * Create a dependency for the newly allocated inode.
1060 	 * Panic if it already exists as something is seriously wrong.
1061 	 * Otherwise add it to the dependency list for the buffer holding
1062 	 * the cylinder group map from which it was allocated.
1063 	 */
1064 	ACQUIRE_LOCK(&lk);
1065 	if (inodedep_lookup(ip->i_fs, newinum, DEPALLOC, &inodedep) != 0)
1066 		panic("softdep_setup_inomapdep: found inode");
1067 	inodedep->id_buf = bp;
1068 	inodedep->id_state &= ~DEPCOMPLETE;
1069 	bmsafemap = bmsafemap_lookup(bp);
1070 	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1071 	FREE_LOCK(&lk);
1072 }
1073 
1074 /*
1075  * Called just after updating the cylinder group block to
1076  * allocate block or fragment.
1077  */
1078 void
1079 softdep_setup_blkmapdep(bp, fs, newblkno)
1080 	struct buf *bp;		/* buffer for cylgroup block with block map */
1081 	struct fs *fs;		/* filesystem doing allocation */
1082 	ufs_daddr_t newblkno;	/* number of newly allocated block */
1083 {
1084 	struct newblk *newblk;
1085 	struct bmsafemap *bmsafemap;
1086 
1087 	/*
1088 	 * Create a dependency for the newly allocated block.
1089 	 * Add it to the dependency list for the buffer holding
1090 	 * the cylinder group map from which it was allocated.
1091 	 */
1092 	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1093 		panic("softdep_setup_blkmapdep: found block");
1094 	ACQUIRE_LOCK(&lk);
1095 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1096 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1097 	FREE_LOCK(&lk);
1098 }
1099 
1100 /*
1101  * Find the bmsafemap associated with a cylinder group buffer.
1102  * If none exists, create one. The buffer must be locked when
1103  * this routine is called and this routine must be called with
1104  * splbio interrupts blocked.
1105  */
1106 static struct bmsafemap *
1107 bmsafemap_lookup(bp)
1108 	struct buf *bp;
1109 {
1110 	struct bmsafemap *bmsafemap;
1111 	struct worklist *wk;
1112 
1113 #ifdef DEBUG
1114 	if (lk.lkt_held == -1)
1115 		panic("bmsafemap_lookup: lock not held");
1116 #endif
1117 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = LIST_NEXT(wk, wk_list))
1118 		if (wk->wk_type == D_BMSAFEMAP)
1119 			return (WK_BMSAFEMAP(wk));
1120 	FREE_LOCK(&lk);
1121 	MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1122 		M_BMSAFEMAP, M_WAITOK);
1123 	bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1124 	bmsafemap->sm_list.wk_state = 0;
1125 	bmsafemap->sm_buf = bp;
1126 	LIST_INIT(&bmsafemap->sm_allocdirecthd);
1127 	LIST_INIT(&bmsafemap->sm_allocindirhd);
1128 	LIST_INIT(&bmsafemap->sm_inodedephd);
1129 	LIST_INIT(&bmsafemap->sm_newblkhd);
1130 	ACQUIRE_LOCK(&lk);
1131 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1132 	return (bmsafemap);
1133 }
1134 
1135 /*
1136  * Direct block allocation dependencies.
1137  *
1138  * When a new block is allocated, the corresponding disk locations must be
1139  * initialized (with zeros or new data) before the on-disk inode points to
1140  * them.  Also, the freemap from which the block was allocated must be
1141  * updated (on disk) before the inode's pointer. These two dependencies are
1142  * independent of each other and are needed for all file blocks and indirect
1143  * blocks that are pointed to directly by the inode.  Just before the
1144  * "in-core" version of the inode is updated with a newly allocated block
1145  * number, a procedure (below) is called to setup allocation dependency
1146  * structures.  These structures are removed when the corresponding
1147  * dependencies are satisfied or when the block allocation becomes obsolete
1148  * (i.e., the file is deleted, the block is de-allocated, or the block is a
1149  * fragment that gets upgraded).  All of these cases are handled in
1150  * procedures described later.
1151  *
1152  * When a file extension causes a fragment to be upgraded, either to a larger
1153  * fragment or to a full block, the on-disk location may change (if the
1154  * previous fragment could not simply be extended). In this case, the old
1155  * fragment must be de-allocated, but not until after the inode's pointer has
1156  * been updated. In most cases, this is handled by later procedures, which
1157  * will construct a "freefrag" structure to be added to the workitem queue
1158  * when the inode update is complete (or obsolete).  The main exception to
1159  * this is when an allocation occurs while a pending allocation dependency
1160  * (for the same block pointer) remains.  This case is handled in the main
1161  * allocation dependency setup procedure by immediately freeing the
1162  * unreferenced fragments.
1163  */
1164 void
1165 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1166 	struct inode *ip;	/* inode to which block is being added */
1167 	ufs_lbn_t lbn;		/* block pointer within inode */
1168 	ufs_daddr_t newblkno;	/* disk block number being added */
1169 	ufs_daddr_t oldblkno;	/* previous block number, 0 unless frag */
1170 	long newsize;		/* size of new block */
1171 	long oldsize;		/* size of new block */
1172 	struct buf *bp;		/* bp for allocated block */
1173 {
1174 	struct allocdirect *adp, *oldadp;
1175 	struct allocdirectlst *adphead;
1176 	struct bmsafemap *bmsafemap;
1177 	struct inodedep *inodedep;
1178 	struct pagedep *pagedep;
1179 	struct newblk *newblk;
1180 
1181 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1182 		M_ALLOCDIRECT, M_WAITOK);
1183 	bzero(adp, sizeof(struct allocdirect));
1184 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1185 	adp->ad_lbn = lbn;
1186 	adp->ad_newblkno = newblkno;
1187 	adp->ad_oldblkno = oldblkno;
1188 	adp->ad_newsize = newsize;
1189 	adp->ad_oldsize = oldsize;
1190 	adp->ad_state = ATTACHED;
1191 	if (newblkno == oldblkno)
1192 		adp->ad_freefrag = NULL;
1193 	else
1194 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1195 
1196 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1197 		panic("softdep_setup_allocdirect: lost block");
1198 
1199 	ACQUIRE_LOCK(&lk);
1200 	(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
1201 	adp->ad_inodedep = inodedep;
1202 
1203 	if (newblk->nb_state == DEPCOMPLETE) {
1204 		adp->ad_state |= DEPCOMPLETE;
1205 		adp->ad_buf = NULL;
1206 	} else {
1207 		bmsafemap = newblk->nb_bmsafemap;
1208 		adp->ad_buf = bmsafemap->sm_buf;
1209 		LIST_REMOVE(newblk, nb_deps);
1210 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1211 	}
1212 	LIST_REMOVE(newblk, nb_hash);
1213 	FREE(newblk, M_NEWBLK);
1214 
1215 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1216 	if (lbn >= NDADDR) {
1217 		/* allocating an indirect block */
1218 		if (oldblkno != 0)
1219 			panic("softdep_setup_allocdirect: non-zero indir");
1220 	} else {
1221 		/*
1222 		 * Allocating a direct block.
1223 		 *
1224 		 * If we are allocating a directory block, then we must
1225 		 * allocate an associated pagedep to track additions and
1226 		 * deletions.
1227 		 */
1228 		if ((ip->i_mode & IFMT) == IFDIR &&
1229 		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1230 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1231 	}
1232 	/*
1233 	 * The list of allocdirects must be kept in sorted and ascending
1234 	 * order so that the rollback routines can quickly determine the
1235 	 * first uncommitted block (the size of the file stored on disk
1236 	 * ends at the end of the lowest committed fragment, or if there
1237 	 * are no fragments, at the end of the highest committed block).
1238 	 * Since files generally grow, the typical case is that the new
1239 	 * block is to be added at the end of the list. We speed this
1240 	 * special case by checking against the last allocdirect in the
1241 	 * list before laboriously traversing the list looking for the
1242 	 * insertion point.
1243 	 */
1244 	adphead = &inodedep->id_newinoupdt;
1245 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1246 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1247 		/* insert at end of list */
1248 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1249 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1250 			allocdirect_merge(adphead, adp, oldadp);
1251 		FREE_LOCK(&lk);
1252 		return;
1253 	}
1254 	for (oldadp = TAILQ_FIRST(adphead); oldadp;
1255 	     oldadp = TAILQ_NEXT(oldadp, ad_next)) {
1256 		if (oldadp->ad_lbn >= lbn)
1257 			break;
1258 	}
1259 	if (oldadp == NULL)
1260 		panic("softdep_setup_allocdirect: lost entry");
1261 	/* insert in middle of list */
1262 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1263 	if (oldadp->ad_lbn == lbn)
1264 		allocdirect_merge(adphead, adp, oldadp);
1265 	FREE_LOCK(&lk);
1266 }
1267 
1268 /*
1269  * Replace an old allocdirect dependency with a newer one.
1270  * This routine must be called with splbio interrupts blocked.
1271  */
1272 static void
1273 allocdirect_merge(adphead, newadp, oldadp)
1274 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
1275 	struct allocdirect *newadp;	/* allocdirect being added */
1276 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
1277 {
1278 	struct freefrag *freefrag;
1279 
1280 #ifdef DEBUG
1281 	if (lk.lkt_held == -1)
1282 		panic("allocdirect_merge: lock not held");
1283 #endif
1284 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1285 	    newadp->ad_oldsize != oldadp->ad_newsize ||
1286 	    newadp->ad_lbn >= NDADDR)
1287 		panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1288 		    newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1289 		    NDADDR);
1290 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
1291 	newadp->ad_oldsize = oldadp->ad_oldsize;
1292 	/*
1293 	 * If the old dependency had a fragment to free or had never
1294 	 * previously had a block allocated, then the new dependency
1295 	 * can immediately post its freefrag and adopt the old freefrag.
1296 	 * This action is done by swapping the freefrag dependencies.
1297 	 * The new dependency gains the old one's freefrag, and the
1298 	 * old one gets the new one and then immediately puts it on
1299 	 * the worklist when it is freed by free_allocdirect. It is
1300 	 * not possible to do this swap when the old dependency had a
1301 	 * non-zero size but no previous fragment to free. This condition
1302 	 * arises when the new block is an extension of the old block.
1303 	 * Here, the first part of the fragment allocated to the new
1304 	 * dependency is part of the block currently claimed on disk by
1305 	 * the old dependency, so cannot legitimately be freed until the
1306 	 * conditions for the new dependency are fulfilled.
1307 	 */
1308 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1309 		freefrag = newadp->ad_freefrag;
1310 		newadp->ad_freefrag = oldadp->ad_freefrag;
1311 		oldadp->ad_freefrag = freefrag;
1312 	}
1313 	free_allocdirect(adphead, oldadp, 0);
1314 }
1315 
1316 /*
1317  * Allocate a new freefrag structure if needed.
1318  */
1319 static struct freefrag *
1320 newfreefrag(ip, blkno, size)
1321 	struct inode *ip;
1322 	ufs_daddr_t blkno;
1323 	long size;
1324 {
1325 	struct freefrag *freefrag;
1326 	struct fs *fs;
1327 
1328 	if (blkno == 0)
1329 		return (NULL);
1330 	fs = ip->i_fs;
1331 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1332 		panic("newfreefrag: frag size");
1333 	MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1334 		M_FREEFRAG, M_WAITOK);
1335 	freefrag->ff_list.wk_type = D_FREEFRAG;
1336 	freefrag->ff_state = ip->i_uid & ~ONWORKLIST;	/* XXX - used below */
1337 	freefrag->ff_inum = ip->i_number;
1338 	freefrag->ff_fs = fs;
1339 	freefrag->ff_devvp = ip->i_devvp;
1340 	freefrag->ff_blkno = blkno;
1341 	freefrag->ff_fragsize = size;
1342 	return (freefrag);
1343 }
1344 
1345 /*
1346  * This workitem de-allocates fragments that were replaced during
1347  * file block allocation.
1348  */
1349 static void
1350 handle_workitem_freefrag(freefrag)
1351 	struct freefrag *freefrag;
1352 {
1353 	struct inode tip;
1354 
1355 	tip.i_fs = freefrag->ff_fs;
1356 	tip.i_devvp = freefrag->ff_devvp;
1357 	tip.i_dev = freefrag->ff_devvp->v_rdev;
1358 	tip.i_number = freefrag->ff_inum;
1359 	tip.i_uid = freefrag->ff_state & ~ONWORKLIST;	/* XXX - set above */
1360 	ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1361 	FREE(freefrag, M_FREEFRAG);
1362 }
1363 
1364 /*
1365  * Indirect block allocation dependencies.
1366  *
1367  * The same dependencies that exist for a direct block also exist when
1368  * a new block is allocated and pointed to by an entry in a block of
1369  * indirect pointers. The undo/redo states described above are also
1370  * used here. Because an indirect block contains many pointers that
1371  * may have dependencies, a second copy of the entire in-memory indirect
1372  * block is kept. The buffer cache copy is always completely up-to-date.
1373  * The second copy, which is used only as a source for disk writes,
1374  * contains only the safe pointers (i.e., those that have no remaining
1375  * update dependencies). The second copy is freed when all pointers
1376  * are safe. The cache is not allowed to replace indirect blocks with
1377  * pending update dependencies. If a buffer containing an indirect
1378  * block with dependencies is written, these routines will mark it
1379  * dirty again. It can only be successfully written once all the
1380  * dependencies are removed. The ffs_fsync routine in conjunction with
1381  * softdep_sync_metadata work together to get all the dependencies
1382  * removed so that a file can be successfully written to disk. Three
1383  * procedures are used when setting up indirect block pointer
1384  * dependencies. The division is necessary because of the organization
1385  * of the "balloc" routine and because of the distinction between file
1386  * pages and file metadata blocks.
1387  */
1388 
1389 /*
1390  * Allocate a new allocindir structure.
1391  */
1392 static struct allocindir *
1393 newallocindir(ip, ptrno, newblkno, oldblkno)
1394 	struct inode *ip;	/* inode for file being extended */
1395 	int ptrno;		/* offset of pointer in indirect block */
1396 	ufs_daddr_t newblkno;	/* disk block number being added */
1397 	ufs_daddr_t oldblkno;	/* previous block number, 0 if none */
1398 {
1399 	struct allocindir *aip;
1400 
1401 	MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1402 		M_ALLOCINDIR, M_WAITOK);
1403 	bzero(aip, sizeof(struct allocindir));
1404 	aip->ai_list.wk_type = D_ALLOCINDIR;
1405 	aip->ai_state = ATTACHED;
1406 	aip->ai_offset = ptrno;
1407 	aip->ai_newblkno = newblkno;
1408 	aip->ai_oldblkno = oldblkno;
1409 	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1410 	return (aip);
1411 }
1412 
1413 /*
1414  * Called just before setting an indirect block pointer
1415  * to a newly allocated file page.
1416  */
1417 void
1418 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1419 	struct inode *ip;	/* inode for file being extended */
1420 	ufs_lbn_t lbn;		/* allocated block number within file */
1421 	struct buf *bp;		/* buffer with indirect blk referencing page */
1422 	int ptrno;		/* offset of pointer in indirect block */
1423 	ufs_daddr_t newblkno;	/* disk block number being added */
1424 	ufs_daddr_t oldblkno;	/* previous block number, 0 if none */
1425 	struct buf *nbp;	/* buffer holding allocated page */
1426 {
1427 	struct allocindir *aip;
1428 	struct pagedep *pagedep;
1429 
1430 	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1431 	ACQUIRE_LOCK(&lk);
1432 	/*
1433 	 * If we are allocating a directory page, then we must
1434 	 * allocate an associated pagedep to track additions and
1435 	 * deletions.
1436 	 */
1437 	if ((ip->i_mode & IFMT) == IFDIR &&
1438 	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1439 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1440 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1441 	FREE_LOCK(&lk);
1442 	setup_allocindir_phase2(bp, ip, aip);
1443 }
1444 
1445 /*
1446  * Called just before setting an indirect block pointer to a
1447  * newly allocated indirect block.
1448  */
1449 void
1450 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1451 	struct buf *nbp;	/* newly allocated indirect block */
1452 	struct inode *ip;	/* inode for file being extended */
1453 	struct buf *bp;		/* indirect block referencing allocated block */
1454 	int ptrno;		/* offset of pointer in indirect block */
1455 	ufs_daddr_t newblkno;	/* disk block number being added */
1456 {
1457 	struct allocindir *aip;
1458 
1459 	aip = newallocindir(ip, ptrno, newblkno, 0);
1460 	ACQUIRE_LOCK(&lk);
1461 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1462 	FREE_LOCK(&lk);
1463 	setup_allocindir_phase2(bp, ip, aip);
1464 }
1465 
1466 /*
1467  * Called to finish the allocation of the "aip" allocated
1468  * by one of the two routines above.
1469  */
1470 static void
1471 setup_allocindir_phase2(bp, ip, aip)
1472 	struct buf *bp;		/* in-memory copy of the indirect block */
1473 	struct inode *ip;	/* inode for file being extended */
1474 	struct allocindir *aip;	/* allocindir allocated by the above routines */
1475 {
1476 	struct worklist *wk;
1477 	struct indirdep *indirdep, *newindirdep;
1478 	struct bmsafemap *bmsafemap;
1479 	struct allocindir *oldaip;
1480 	struct freefrag *freefrag;
1481 	struct newblk *newblk;
1482 
1483 	if (bp->b_lblkno >= 0)
1484 		panic("setup_allocindir_phase2: not indir blk");
1485 	for (indirdep = NULL, newindirdep = NULL; ; ) {
1486 		ACQUIRE_LOCK(&lk);
1487 		for (wk = LIST_FIRST(&bp->b_dep); wk;
1488 		     wk = LIST_NEXT(wk, wk_list)) {
1489 			if (wk->wk_type != D_INDIRDEP)
1490 				continue;
1491 			indirdep = WK_INDIRDEP(wk);
1492 			break;
1493 		}
1494 		if (indirdep == NULL && newindirdep) {
1495 			indirdep = newindirdep;
1496 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1497 			newindirdep = NULL;
1498 		}
1499 		FREE_LOCK(&lk);
1500 		if (indirdep) {
1501 			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1502 			    &newblk) == 0)
1503 				panic("setup_allocindir: lost block");
1504 			ACQUIRE_LOCK(&lk);
1505 			if (newblk->nb_state == DEPCOMPLETE) {
1506 				aip->ai_state |= DEPCOMPLETE;
1507 				aip->ai_buf = NULL;
1508 			} else {
1509 				bmsafemap = newblk->nb_bmsafemap;
1510 				aip->ai_buf = bmsafemap->sm_buf;
1511 				LIST_REMOVE(newblk, nb_deps);
1512 				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1513 				    aip, ai_deps);
1514 			}
1515 			LIST_REMOVE(newblk, nb_hash);
1516 			FREE(newblk, M_NEWBLK);
1517 			aip->ai_indirdep = indirdep;
1518 			/*
1519 			 * Check to see if there is an existing dependency
1520 			 * for this block. If there is, merge the old
1521 			 * dependency into the new one.
1522 			 */
1523 			if (aip->ai_oldblkno == 0)
1524 				oldaip = NULL;
1525 			else
1526 				for (oldaip=LIST_FIRST(&indirdep->ir_deplisthd);
1527 				    oldaip; oldaip = LIST_NEXT(oldaip, ai_next))
1528 					if (oldaip->ai_offset == aip->ai_offset)
1529 						break;
1530 			if (oldaip != NULL) {
1531 				if (oldaip->ai_newblkno != aip->ai_oldblkno)
1532 					panic("setup_allocindir_phase2: blkno");
1533 				aip->ai_oldblkno = oldaip->ai_oldblkno;
1534 				freefrag = oldaip->ai_freefrag;
1535 				oldaip->ai_freefrag = aip->ai_freefrag;
1536 				aip->ai_freefrag = freefrag;
1537 				free_allocindir(oldaip, NULL);
1538 			}
1539 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1540 			((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1541 			    [aip->ai_offset] = aip->ai_oldblkno;
1542 			FREE_LOCK(&lk);
1543 		}
1544 		if (newindirdep) {
1545 			if (indirdep->ir_savebp != NULL)
1546 				brelse(newindirdep->ir_savebp);
1547 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1548 		}
1549 		if (indirdep)
1550 			break;
1551 		MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1552 			M_INDIRDEP, M_WAITOK);
1553 		newindirdep->ir_list.wk_type = D_INDIRDEP;
1554 		newindirdep->ir_state = ATTACHED;
1555 		LIST_INIT(&newindirdep->ir_deplisthd);
1556 		LIST_INIT(&newindirdep->ir_donehd);
1557 		if (bp->b_blkno == bp->b_lblkno) {
1558 			VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
1559 				NULL, NULL);
1560 		}
1561 		newindirdep->ir_savebp =
1562 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0);
1563 		BUF_KERNPROC(newindirdep->ir_savebp);
1564 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1565 	}
1566 }
1567 
1568 /*
1569  * Block de-allocation dependencies.
1570  *
1571  * When blocks are de-allocated, the on-disk pointers must be nullified before
1572  * the blocks are made available for use by other files.  (The true
1573  * requirement is that old pointers must be nullified before new on-disk
1574  * pointers are set.  We chose this slightly more stringent requirement to
1575  * reduce complexity.) Our implementation handles this dependency by updating
1576  * the inode (or indirect block) appropriately but delaying the actual block
1577  * de-allocation (i.e., freemap and free space count manipulation) until
1578  * after the updated versions reach stable storage.  After the disk is
1579  * updated, the blocks can be safely de-allocated whenever it is convenient.
1580  * This implementation handles only the common case of reducing a file's
1581  * length to zero. Other cases are handled by the conventional synchronous
1582  * write approach.
1583  *
1584  * The ffs implementation with which we worked double-checks
1585  * the state of the block pointers and file size as it reduces
1586  * a file's length.  Some of this code is replicated here in our
1587  * soft updates implementation.  The freeblks->fb_chkcnt field is
1588  * used to transfer a part of this information to the procedure
1589  * that eventually de-allocates the blocks.
1590  *
1591  * This routine should be called from the routine that shortens
1592  * a file's length, before the inode's size or block pointers
1593  * are modified. It will save the block pointer information for
1594  * later release and zero the inode so that the calling routine
1595  * can release it.
1596  */
1597 void
1598 softdep_setup_freeblocks(ip, length)
1599 	struct inode *ip;	/* The inode whose length is to be reduced */
1600 	off_t length;		/* The new length for the file */
1601 {
1602 	struct freeblks *freeblks;
1603 	struct inodedep *inodedep;
1604 	struct allocdirect *adp;
1605 	struct vnode *vp;
1606 	struct buf *bp;
1607 	struct fs *fs;
1608 	int i, error;
1609 
1610 	fs = ip->i_fs;
1611 	if (length != 0)
1612 		panic("softde_setup_freeblocks: non-zero length");
1613 	MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1614 		M_FREEBLKS, M_WAITOK);
1615 	bzero(freeblks, sizeof(struct freeblks));
1616 	freeblks->fb_list.wk_type = D_FREEBLKS;
1617 	freeblks->fb_uid = ip->i_uid;
1618 	freeblks->fb_previousinum = ip->i_number;
1619 	freeblks->fb_devvp = ip->i_devvp;
1620 	freeblks->fb_fs = fs;
1621 	freeblks->fb_oldsize = ip->i_size;
1622 	freeblks->fb_newsize = length;
1623 	freeblks->fb_chkcnt = ip->i_blocks;
1624 	for (i = 0; i < NDADDR; i++) {
1625 		freeblks->fb_dblks[i] = ip->i_db[i];
1626 		ip->i_db[i] = 0;
1627 	}
1628 	for (i = 0; i < NIADDR; i++) {
1629 		freeblks->fb_iblks[i] = ip->i_ib[i];
1630 		ip->i_ib[i] = 0;
1631 	}
1632 	ip->i_blocks = 0;
1633 	ip->i_size = 0;
1634 	/*
1635 	 * Push the zero'ed inode to to its disk buffer so that we are free
1636 	 * to delete its dependencies below. Once the dependencies are gone
1637 	 * the buffer can be safely released.
1638 	 */
1639 	if ((error = bread(ip->i_devvp,
1640 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
1641 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0)
1642 		softdep_error("softdep_setup_freeblocks", error);
1643 	*((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
1644 	    ip->i_din;
1645 	/*
1646 	 * Find and eliminate any inode dependencies.
1647 	 */
1648 	ACQUIRE_LOCK(&lk);
1649 	(void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1650 	if ((inodedep->id_state & IOSTARTED) != 0)
1651 		panic("softdep_setup_freeblocks: inode busy");
1652 	/*
1653 	 * Because the file length has been truncated to zero, any
1654 	 * pending block allocation dependency structures associated
1655 	 * with this inode are obsolete and can simply be de-allocated.
1656 	 * We must first merge the two dependency lists to get rid of
1657 	 * any duplicate freefrag structures, then purge the merged list.
1658 	 */
1659 	merge_inode_lists(inodedep);
1660 	while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1661 		free_allocdirect(&inodedep->id_inoupdt, adp, 1);
1662 	FREE_LOCK(&lk);
1663 	bdwrite(bp);
1664 	/*
1665 	 * We must wait for any I/O in progress to finish so that
1666 	 * all potential buffers on the dirty list will be visible.
1667 	 * Once they are all there, walk the list and get rid of
1668 	 * any dependencies.
1669 	 */
1670 	vp = ITOV(ip);
1671 	ACQUIRE_LOCK(&lk);
1672 	drain_output(vp, 1);
1673 	while (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) {
1674 		bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
1675 		(void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
1676 		deallocate_dependencies(bp, inodedep);
1677 		bp->b_flags |= B_INVAL | B_NOCACHE;
1678 		FREE_LOCK(&lk);
1679 		brelse(bp);
1680 		ACQUIRE_LOCK(&lk);
1681 	}
1682 	/*
1683 	 * Add the freeblks structure to the list of operations that
1684 	 * must await the zero'ed inode being written to disk. If we
1685 	 * still have a bitmap dependency, then the inode has never been
1686 	 * written to disk, so we can process the freeblks immediately.
1687 	 * If the inodedep does not exist, then the zero'ed inode has
1688 	 * been written and we can also proceed.
1689 	 */
1690 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0 ||
1691 	    free_inodedep(inodedep) ||
1692 	    (inodedep->id_state & DEPCOMPLETE) == 0) {
1693 		FREE_LOCK(&lk);
1694 		handle_workitem_freeblocks(freeblks);
1695 	} else {
1696 		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1697 		FREE_LOCK(&lk);
1698 	}
1699 }
1700 
1701 /*
1702  * Reclaim any dependency structures from a buffer that is about to
1703  * be reallocated to a new vnode. The buffer must be locked, thus,
1704  * no I/O completion operations can occur while we are manipulating
1705  * its associated dependencies. The mutex is held so that other I/O's
1706  * associated with related dependencies do not occur.
1707  */
1708 static void
1709 deallocate_dependencies(bp, inodedep)
1710 	struct buf *bp;
1711 	struct inodedep *inodedep;
1712 {
1713 	struct worklist *wk;
1714 	struct indirdep *indirdep;
1715 	struct allocindir *aip;
1716 	struct pagedep *pagedep;
1717 	struct dirrem *dirrem;
1718 	struct diradd *dap;
1719 	int i;
1720 
1721 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1722 		switch (wk->wk_type) {
1723 
1724 		case D_INDIRDEP:
1725 			indirdep = WK_INDIRDEP(wk);
1726 			/*
1727 			 * None of the indirect pointers will ever be visible,
1728 			 * so they can simply be tossed. GOINGAWAY ensures
1729 			 * that allocated pointers will be saved in the buffer
1730 			 * cache until they are freed. Note that they will
1731 			 * only be able to be found by their physical address
1732 			 * since the inode mapping the logical address will
1733 			 * be gone. The save buffer used for the safe copy
1734 			 * was allocated in setup_allocindir_phase2 using
1735 			 * the physical address so it could be used for this
1736 			 * purpose. Hence we swap the safe copy with the real
1737 			 * copy, allowing the safe copy to be freed and holding
1738 			 * on to the real copy for later use in indir_trunc.
1739 			 */
1740 			if (indirdep->ir_state & GOINGAWAY)
1741 				panic("deallocate_dependencies: already gone");
1742 			indirdep->ir_state |= GOINGAWAY;
1743 			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
1744 				free_allocindir(aip, inodedep);
1745 			if (bp->b_lblkno >= 0 ||
1746 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno)
1747 				panic("deallocate_dependencies: not indir");
1748 			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
1749 			    bp->b_bcount);
1750 			WORKLIST_REMOVE(wk);
1751 			WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
1752 			continue;
1753 
1754 		case D_PAGEDEP:
1755 			pagedep = WK_PAGEDEP(wk);
1756 			/*
1757 			 * None of the directory additions will ever be
1758 			 * visible, so they can simply be tossed.
1759 			 */
1760 			for (i = 0; i < DAHASHSZ; i++)
1761 				while ((dap =
1762 				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
1763 					free_diradd(dap);
1764 			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
1765 				free_diradd(dap);
1766 			/*
1767 			 * Copy any directory remove dependencies to the list
1768 			 * to be processed after the zero'ed inode is written.
1769 			 * If the inode has already been written, then they
1770 			 * can be dumped directly onto the work list.
1771 			 */
1772 			for (dirrem = LIST_FIRST(&pagedep->pd_dirremhd); dirrem;
1773 			     dirrem = LIST_NEXT(dirrem, dm_next)) {
1774 				LIST_REMOVE(dirrem, dm_next);
1775 				dirrem->dm_dirinum = pagedep->pd_ino;
1776 				if (inodedep == NULL ||
1777 				    (inodedep->id_state & ALLCOMPLETE) ==
1778 				     ALLCOMPLETE)
1779 					add_to_worklist(&dirrem->dm_list);
1780 				else
1781 					WORKLIST_INSERT(&inodedep->id_bufwait,
1782 					    &dirrem->dm_list);
1783 			}
1784 			WORKLIST_REMOVE(&pagedep->pd_list);
1785 			LIST_REMOVE(pagedep, pd_hash);
1786 			WORKITEM_FREE(pagedep, D_PAGEDEP);
1787 			continue;
1788 
1789 		case D_ALLOCINDIR:
1790 			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
1791 			continue;
1792 
1793 		case D_ALLOCDIRECT:
1794 		case D_INODEDEP:
1795 			panic("deallocate_dependencies: Unexpected type %s",
1796 			    TYPENAME(wk->wk_type));
1797 			/* NOTREACHED */
1798 
1799 		default:
1800 			panic("deallocate_dependencies: Unknown type %s",
1801 			    TYPENAME(wk->wk_type));
1802 			/* NOTREACHED */
1803 		}
1804 	}
1805 }
1806 
1807 /*
1808  * Free an allocdirect. Generate a new freefrag work request if appropriate.
1809  * This routine must be called with splbio interrupts blocked.
1810  */
1811 static void
1812 free_allocdirect(adphead, adp, delay)
1813 	struct allocdirectlst *adphead;
1814 	struct allocdirect *adp;
1815 	int delay;
1816 {
1817 
1818 #ifdef DEBUG
1819 	if (lk.lkt_held == -1)
1820 		panic("free_allocdirect: lock not held");
1821 #endif
1822 	if ((adp->ad_state & DEPCOMPLETE) == 0)
1823 		LIST_REMOVE(adp, ad_deps);
1824 	TAILQ_REMOVE(adphead, adp, ad_next);
1825 	if ((adp->ad_state & COMPLETE) == 0)
1826 		WORKLIST_REMOVE(&adp->ad_list);
1827 	if (adp->ad_freefrag != NULL) {
1828 		if (delay)
1829 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
1830 			    &adp->ad_freefrag->ff_list);
1831 		else
1832 			add_to_worklist(&adp->ad_freefrag->ff_list);
1833 	}
1834 	WORKITEM_FREE(adp, D_ALLOCDIRECT);
1835 }
1836 
1837 /*
1838  * Prepare an inode to be freed. The actual free operation is not
1839  * done until the zero'ed inode has been written to disk.
1840  */
1841 void
1842 softdep_freefile(pvp, ino, mode)
1843 		struct vnode *pvp;
1844 		ino_t ino;
1845 		int mode;
1846 {
1847 	struct inode *ip = VTOI(pvp);
1848 	struct inodedep *inodedep;
1849 	struct freefile *freefile;
1850 
1851 	/*
1852 	 * This sets up the inode de-allocation dependency.
1853 	 */
1854 	MALLOC(freefile, struct freefile *, sizeof(struct freefile),
1855 		M_FREEFILE, M_WAITOK);
1856 	freefile->fx_list.wk_type = D_FREEFILE;
1857 	freefile->fx_list.wk_state = 0;
1858 	freefile->fx_mode = mode;
1859 	freefile->fx_oldinum = ino;
1860 	freefile->fx_devvp = ip->i_devvp;
1861 	freefile->fx_fs = ip->i_fs;
1862 
1863 	/*
1864 	 * If the inodedep does not exist, then the zero'ed inode has
1865 	 * been written to disk. If the allocated inode has never been
1866 	 * written to disk, then the on-disk inode is zero'ed. In either
1867 	 * case we can free the file immediately.
1868 	 */
1869 	ACQUIRE_LOCK(&lk);
1870 	if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
1871 	    check_inode_unwritten(inodedep)) {
1872 		FREE_LOCK(&lk);
1873 		handle_workitem_freefile(freefile);
1874 		return;
1875 	}
1876 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
1877 	FREE_LOCK(&lk);
1878 }
1879 
1880 /*
1881  * Check to see if an inode has never been written to disk. If
1882  * so free the inodedep and return success, otherwise return failure.
1883  * This routine must be called with splbio interrupts blocked.
1884  *
1885  * If we still have a bitmap dependency, then the inode has never
1886  * been written to disk. Drop the dependency as it is no longer
1887  * necessary since the inode is being deallocated. We set the
1888  * ALLCOMPLETE flags since the bitmap now properly shows that the
1889  * inode is not allocated. Even if the inode is actively being
1890  * written, it has been rolled back to its zero'ed state, so we
1891  * are ensured that a zero inode is what is on the disk. For short
1892  * lived files, this change will usually result in removing all the
1893  * dependencies from the inode so that it can be freed immediately.
1894  */
1895 static int
1896 check_inode_unwritten(inodedep)
1897 	struct inodedep *inodedep;
1898 {
1899 
1900 	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
1901 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
1902 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
1903 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
1904 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
1905 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
1906 	    inodedep->id_nlinkdelta != 0)
1907 		return (0);
1908 	inodedep->id_state |= ALLCOMPLETE;
1909 	LIST_REMOVE(inodedep, id_deps);
1910 	inodedep->id_buf = NULL;
1911 	WORKLIST_REMOVE(&inodedep->id_list);
1912 	if (inodedep->id_savedino != NULL) {
1913 		FREE(inodedep->id_savedino, M_INODEDEP);
1914 		inodedep->id_savedino = NULL;
1915 	}
1916 	if (free_inodedep(inodedep) == 0)
1917 		panic("check_inode_unwritten: busy inode");
1918 	return (1);
1919 }
1920 
1921 /*
1922  * Try to free an inodedep structure. Return 1 if it could be freed.
1923  */
1924 static int
1925 free_inodedep(inodedep)
1926 	struct inodedep *inodedep;
1927 {
1928 
1929 	if ((inodedep->id_state & ONWORKLIST) != 0 ||
1930 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
1931 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
1932 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
1933 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
1934 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
1935 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
1936 	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
1937 		return (0);
1938 	LIST_REMOVE(inodedep, id_hash);
1939 	WORKITEM_FREE(inodedep, D_INODEDEP);
1940 	num_inodedep -= 1;
1941 	return (1);
1942 }
1943 
1944 /*
1945  * This workitem routine performs the block de-allocation.
1946  * The workitem is added to the pending list after the updated
1947  * inode block has been written to disk.  As mentioned above,
1948  * checks regarding the number of blocks de-allocated (compared
1949  * to the number of blocks allocated for the file) are also
1950  * performed in this function.
1951  */
1952 static void
1953 handle_workitem_freeblocks(freeblks)
1954 	struct freeblks *freeblks;
1955 {
1956 	struct inode tip;
1957 	ufs_daddr_t bn;
1958 	struct fs *fs;
1959 	int i, level, bsize;
1960 	long nblocks, blocksreleased = 0;
1961 	int error, allerror = 0;
1962 	ufs_lbn_t baselbns[NIADDR], tmpval;
1963 
1964 	tip.i_number = freeblks->fb_previousinum;
1965 	tip.i_devvp = freeblks->fb_devvp;
1966 	tip.i_dev = freeblks->fb_devvp->v_rdev;
1967 	tip.i_fs = freeblks->fb_fs;
1968 	tip.i_size = freeblks->fb_oldsize;
1969 	tip.i_uid = freeblks->fb_uid;
1970 	fs = freeblks->fb_fs;
1971 	tmpval = 1;
1972 	baselbns[0] = NDADDR;
1973 	for (i = 1; i < NIADDR; i++) {
1974 		tmpval *= NINDIR(fs);
1975 		baselbns[i] = baselbns[i - 1] + tmpval;
1976 	}
1977 	nblocks = btodb(fs->fs_bsize);
1978 	blocksreleased = 0;
1979 	/*
1980 	 * Indirect blocks first.
1981 	 */
1982 	for (level = (NIADDR - 1); level >= 0; level--) {
1983 		if ((bn = freeblks->fb_iblks[level]) == 0)
1984 			continue;
1985 		if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level,
1986 		    baselbns[level], &blocksreleased)) == 0)
1987 			allerror = error;
1988 		ffs_blkfree(&tip, bn, fs->fs_bsize);
1989 		blocksreleased += nblocks;
1990 	}
1991 	/*
1992 	 * All direct blocks or frags.
1993 	 */
1994 	for (i = (NDADDR - 1); i >= 0; i--) {
1995 		if ((bn = freeblks->fb_dblks[i]) == 0)
1996 			continue;
1997 		bsize = blksize(fs, &tip, i);
1998 		ffs_blkfree(&tip, bn, bsize);
1999 		blocksreleased += btodb(bsize);
2000 	}
2001 
2002 #ifdef DIAGNOSTIC
2003 	if (freeblks->fb_chkcnt != blocksreleased)
2004 		panic("handle_workitem_freeblocks: block count");
2005 	if (allerror)
2006 		softdep_error("handle_workitem_freeblks", allerror);
2007 #endif /* DIAGNOSTIC */
2008 	WORKITEM_FREE(freeblks, D_FREEBLKS);
2009 }
2010 
2011 /*
2012  * Release blocks associated with the inode ip and stored in the indirect
2013  * block dbn. If level is greater than SINGLE, the block is an indirect block
2014  * and recursive calls to indirtrunc must be used to cleanse other indirect
2015  * blocks.
2016  */
2017 static int
2018 indir_trunc(ip, dbn, level, lbn, countp)
2019 	struct inode *ip;
2020 	ufs_daddr_t dbn;
2021 	int level;
2022 	ufs_lbn_t lbn;
2023 	long *countp;
2024 {
2025 	struct buf *bp;
2026 	ufs_daddr_t *bap;
2027 	ufs_daddr_t nb;
2028 	struct fs *fs;
2029 	struct worklist *wk;
2030 	struct indirdep *indirdep;
2031 	int i, lbnadd, nblocks;
2032 	int error, allerror = 0;
2033 
2034 	fs = ip->i_fs;
2035 	lbnadd = 1;
2036 	for (i = level; i > 0; i--)
2037 		lbnadd *= NINDIR(fs);
2038 	/*
2039 	 * Get buffer of block pointers to be freed. This routine is not
2040 	 * called until the zero'ed inode has been written, so it is safe
2041 	 * to free blocks as they are encountered. Because the inode has
2042 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
2043 	 * have to use the on-disk address and the block device for the
2044 	 * filesystem to look them up. If the file was deleted before its
2045 	 * indirect blocks were all written to disk, the routine that set
2046 	 * us up (deallocate_dependencies) will have arranged to leave
2047 	 * a complete copy of the indirect block in memory for our use.
2048 	 * Otherwise we have to read the blocks in from the disk.
2049 	 */
2050 	ACQUIRE_LOCK(&lk);
2051 	if ((bp = incore(ip->i_devvp, dbn)) != NULL &&
2052 	    (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2053 		if (wk->wk_type != D_INDIRDEP ||
2054 		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2055 		    (indirdep->ir_state & GOINGAWAY) == 0)
2056 			panic("indir_trunc: lost indirdep");
2057 		WORKLIST_REMOVE(wk);
2058 		WORKITEM_FREE(indirdep, D_INDIRDEP);
2059 		if (LIST_FIRST(&bp->b_dep) != NULL)
2060 			panic("indir_trunc: dangling dep");
2061 		FREE_LOCK(&lk);
2062 	} else {
2063 		FREE_LOCK(&lk);
2064 		error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, NOCRED, &bp);
2065 		if (error)
2066 			return (error);
2067 	}
2068 	/*
2069 	 * Recursively free indirect blocks.
2070 	 */
2071 	bap = (ufs_daddr_t *)bp->b_data;
2072 	nblocks = btodb(fs->fs_bsize);
2073 	for (i = NINDIR(fs) - 1; i >= 0; i--) {
2074 		if ((nb = bap[i]) == 0)
2075 			continue;
2076 		if (level != 0) {
2077 			if ((error = indir_trunc(ip, fsbtodb(fs, nb),
2078 			     level - 1, lbn + (i * lbnadd), countp)) != 0)
2079 				allerror = error;
2080 		}
2081 		ffs_blkfree(ip, nb, fs->fs_bsize);
2082 		*countp += nblocks;
2083 	}
2084 	bp->b_flags |= B_INVAL | B_NOCACHE;
2085 	brelse(bp);
2086 	return (allerror);
2087 }
2088 
2089 /*
2090  * Free an allocindir.
2091  * This routine must be called with splbio interrupts blocked.
2092  */
2093 static void
2094 free_allocindir(aip, inodedep)
2095 	struct allocindir *aip;
2096 	struct inodedep *inodedep;
2097 {
2098 	struct freefrag *freefrag;
2099 
2100 #ifdef DEBUG
2101 	if (lk.lkt_held == -1)
2102 		panic("free_allocindir: lock not held");
2103 #endif
2104 	if ((aip->ai_state & DEPCOMPLETE) == 0)
2105 		LIST_REMOVE(aip, ai_deps);
2106 	if (aip->ai_state & ONWORKLIST)
2107 		WORKLIST_REMOVE(&aip->ai_list);
2108 	LIST_REMOVE(aip, ai_next);
2109 	if ((freefrag = aip->ai_freefrag) != NULL) {
2110 		if (inodedep == NULL)
2111 			add_to_worklist(&freefrag->ff_list);
2112 		else
2113 			WORKLIST_INSERT(&inodedep->id_bufwait,
2114 			    &freefrag->ff_list);
2115 	}
2116 	WORKITEM_FREE(aip, D_ALLOCINDIR);
2117 }
2118 
2119 /*
2120  * Directory entry addition dependencies.
2121  *
2122  * When adding a new directory entry, the inode (with its incremented link
2123  * count) must be written to disk before the directory entry's pointer to it.
2124  * Also, if the inode is newly allocated, the corresponding freemap must be
2125  * updated (on disk) before the directory entry's pointer. These requirements
2126  * are met via undo/redo on the directory entry's pointer, which consists
2127  * simply of the inode number.
2128  *
2129  * As directory entries are added and deleted, the free space within a
2130  * directory block can become fragmented.  The ufs file system will compact
2131  * a fragmented directory block to make space for a new entry. When this
2132  * occurs, the offsets of previously added entries change. Any "diradd"
2133  * dependency structures corresponding to these entries must be updated with
2134  * the new offsets.
2135  */
2136 
2137 /*
2138  * This routine is called after the in-memory inode's link
2139  * count has been incremented, but before the directory entry's
2140  * pointer to the inode has been set.
2141  */
2142 void
2143 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp)
2144 	struct buf *bp;		/* buffer containing directory block */
2145 	struct inode *dp;	/* inode for directory */
2146 	off_t diroffset;	/* offset of new entry in directory */
2147 	long newinum;		/* inode referenced by new directory entry */
2148 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
2149 {
2150 	int offset;		/* offset of new entry within directory block */
2151 	ufs_lbn_t lbn;		/* block in directory containing new entry */
2152 	struct fs *fs;
2153 	struct diradd *dap;
2154 	struct pagedep *pagedep;
2155 	struct inodedep *inodedep;
2156 	struct mkdir *mkdir1, *mkdir2;
2157 
2158 	/*
2159 	 * Whiteouts have no dependencies.
2160 	 */
2161 	if (newinum == WINO) {
2162 		if (newdirbp != NULL)
2163 			bdwrite(newdirbp);
2164 		return;
2165 	}
2166 
2167 	fs = dp->i_fs;
2168 	lbn = lblkno(fs, diroffset);
2169 	offset = blkoff(fs, diroffset);
2170 	MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, M_WAITOK);
2171 	bzero(dap, sizeof(struct diradd));
2172 	dap->da_list.wk_type = D_DIRADD;
2173 	dap->da_offset = offset;
2174 	dap->da_newinum = newinum;
2175 	dap->da_state = ATTACHED;
2176 	if (newdirbp == NULL) {
2177 		dap->da_state |= DEPCOMPLETE;
2178 		ACQUIRE_LOCK(&lk);
2179 	} else {
2180 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2181 		MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2182 		    M_WAITOK);
2183 		mkdir1->md_list.wk_type = D_MKDIR;
2184 		mkdir1->md_state = MKDIR_BODY;
2185 		mkdir1->md_diradd = dap;
2186 		MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2187 		    M_WAITOK);
2188 		mkdir2->md_list.wk_type = D_MKDIR;
2189 		mkdir2->md_state = MKDIR_PARENT;
2190 		mkdir2->md_diradd = dap;
2191 		/*
2192 		 * Dependency on "." and ".." being written to disk.
2193 		 */
2194 		mkdir1->md_buf = newdirbp;
2195 		ACQUIRE_LOCK(&lk);
2196 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2197 		WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2198 		FREE_LOCK(&lk);
2199 		bdwrite(newdirbp);
2200 		/*
2201 		 * Dependency on link count increase for parent directory
2202 		 */
2203 		ACQUIRE_LOCK(&lk);
2204 		if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2205 		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2206 			dap->da_state &= ~MKDIR_PARENT;
2207 			WORKITEM_FREE(mkdir2, D_MKDIR);
2208 		} else {
2209 			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2210 			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2211 		}
2212 	}
2213 	/*
2214 	 * Link into parent directory pagedep to await its being written.
2215 	 */
2216 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2217 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2218 	dap->da_pagedep = pagedep;
2219 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2220 	    da_pdlist);
2221 	/*
2222 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
2223 	 * is not yet written. If it is written, do the post-inode write
2224 	 * processing to put it on the id_pendinghd list.
2225 	 */
2226 	(void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2227 	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2228 		diradd_inode_written(dap, inodedep);
2229 	else
2230 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2231 	FREE_LOCK(&lk);
2232 }
2233 
2234 /*
2235  * This procedure is called to change the offset of a directory
2236  * entry when compacting a directory block which must be owned
2237  * exclusively by the caller. Note that the actual entry movement
2238  * must be done in this procedure to ensure that no I/O completions
2239  * occur while the move is in progress.
2240  */
2241 void
2242 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2243 	struct inode *dp;	/* inode for directory */
2244 	caddr_t base;		/* address of dp->i_offset */
2245 	caddr_t oldloc;		/* address of old directory location */
2246 	caddr_t newloc;		/* address of new directory location */
2247 	int entrysize;		/* size of directory entry */
2248 {
2249 	int offset, oldoffset, newoffset;
2250 	struct pagedep *pagedep;
2251 	struct diradd *dap;
2252 	ufs_lbn_t lbn;
2253 
2254 	ACQUIRE_LOCK(&lk);
2255 	lbn = lblkno(dp->i_fs, dp->i_offset);
2256 	offset = blkoff(dp->i_fs, dp->i_offset);
2257 	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2258 		goto done;
2259 	oldoffset = offset + (oldloc - base);
2260 	newoffset = offset + (newloc - base);
2261 	for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(oldoffset)]);
2262 	     dap; dap = LIST_NEXT(dap, da_pdlist)) {
2263 		if (dap->da_offset != oldoffset)
2264 			continue;
2265 		dap->da_offset = newoffset;
2266 		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2267 			break;
2268 		LIST_REMOVE(dap, da_pdlist);
2269 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2270 		    dap, da_pdlist);
2271 		break;
2272 	}
2273 	if (dap == NULL) {
2274 		for (dap = LIST_FIRST(&pagedep->pd_pendinghd);
2275 		     dap; dap = LIST_NEXT(dap, da_pdlist)) {
2276 			if (dap->da_offset == oldoffset) {
2277 				dap->da_offset = newoffset;
2278 				break;
2279 			}
2280 		}
2281 	}
2282 done:
2283 	bcopy(oldloc, newloc, entrysize);
2284 	FREE_LOCK(&lk);
2285 }
2286 
2287 /*
2288  * Free a diradd dependency structure. This routine must be called
2289  * with splbio interrupts blocked.
2290  */
2291 static void
2292 free_diradd(dap)
2293 	struct diradd *dap;
2294 {
2295 	struct dirrem *dirrem;
2296 	struct pagedep *pagedep;
2297 	struct inodedep *inodedep;
2298 	struct mkdir *mkdir, *nextmd;
2299 
2300 #ifdef DEBUG
2301 	if (lk.lkt_held == -1)
2302 		panic("free_diradd: lock not held");
2303 #endif
2304 	WORKLIST_REMOVE(&dap->da_list);
2305 	LIST_REMOVE(dap, da_pdlist);
2306 	if ((dap->da_state & DIRCHG) == 0) {
2307 		pagedep = dap->da_pagedep;
2308 	} else {
2309 		dirrem = dap->da_previous;
2310 		pagedep = dirrem->dm_pagedep;
2311 		dirrem->dm_dirinum = pagedep->pd_ino;
2312 		add_to_worklist(&dirrem->dm_list);
2313 	}
2314 	if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2315 	    0, &inodedep) != 0)
2316 		(void) free_inodedep(inodedep);
2317 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2318 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2319 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
2320 			if (mkdir->md_diradd != dap)
2321 				continue;
2322 			dap->da_state &= ~mkdir->md_state;
2323 			WORKLIST_REMOVE(&mkdir->md_list);
2324 			LIST_REMOVE(mkdir, md_mkdirs);
2325 			WORKITEM_FREE(mkdir, D_MKDIR);
2326 		}
2327 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
2328 			panic("free_diradd: unfound ref");
2329 	}
2330 	WORKITEM_FREE(dap, D_DIRADD);
2331 }
2332 
2333 /*
2334  * Directory entry removal dependencies.
2335  *
2336  * When removing a directory entry, the entry's inode pointer must be
2337  * zero'ed on disk before the corresponding inode's link count is decremented
2338  * (possibly freeing the inode for re-use). This dependency is handled by
2339  * updating the directory entry but delaying the inode count reduction until
2340  * after the directory block has been written to disk. After this point, the
2341  * inode count can be decremented whenever it is convenient.
2342  */
2343 
2344 /*
2345  * This routine should be called immediately after removing
2346  * a directory entry.  The inode's link count should not be
2347  * decremented by the calling procedure -- the soft updates
2348  * code will do this task when it is safe.
2349  */
2350 void
2351 softdep_setup_remove(bp, dp, ip, isrmdir)
2352 	struct buf *bp;		/* buffer containing directory block */
2353 	struct inode *dp;	/* inode for the directory being modified */
2354 	struct inode *ip;	/* inode for directory entry being removed */
2355 	int isrmdir;		/* indicates if doing RMDIR */
2356 {
2357 	struct dirrem *dirrem, *prevdirrem;
2358 
2359 	/*
2360 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2361 	 */
2362 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2363 
2364 	/*
2365 	 * If the COMPLETE flag is clear, then there were no active
2366 	 * entries and we want to roll back to a zeroed entry until
2367 	 * the new inode is committed to disk. If the COMPLETE flag is
2368 	 * set then we have deleted an entry that never made it to
2369 	 * disk. If the entry we deleted resulted from a name change,
2370 	 * then the old name still resides on disk. We cannot delete
2371 	 * its inode (returned to us in prevdirrem) until the zeroed
2372 	 * directory entry gets to disk. The new inode has never been
2373 	 * referenced on the disk, so can be deleted immediately.
2374 	 */
2375 	if ((dirrem->dm_state & COMPLETE) == 0) {
2376 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2377 		    dm_next);
2378 		FREE_LOCK(&lk);
2379 	} else {
2380 		if (prevdirrem != NULL)
2381 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2382 			    prevdirrem, dm_next);
2383 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2384 		FREE_LOCK(&lk);
2385 		handle_workitem_remove(dirrem);
2386 	}
2387 }
2388 
2389 /*
2390  * Allocate a new dirrem if appropriate and return it along with
2391  * its associated pagedep. Called without a lock, returns with lock.
2392  */
2393 static long num_dirrem;		/* number of dirrem allocated */
2394 static struct dirrem *
2395 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2396 	struct buf *bp;		/* buffer containing directory block */
2397 	struct inode *dp;	/* inode for the directory being modified */
2398 	struct inode *ip;	/* inode for directory entry being removed */
2399 	int isrmdir;		/* indicates if doing RMDIR */
2400 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
2401 {
2402 	int offset;
2403 	ufs_lbn_t lbn;
2404 	struct diradd *dap;
2405 	struct dirrem *dirrem;
2406 	struct pagedep *pagedep;
2407 
2408 	/*
2409 	 * Whiteouts have no deletion dependencies.
2410 	 */
2411 	if (ip == NULL)
2412 		panic("newdirrem: whiteout");
2413 	/*
2414 	 * If we are over our limit, try to improve the situation.
2415 	 * Limiting the number of dirrem structures will also limit
2416 	 * the number of freefile and freeblks structures.
2417 	 */
2418 	if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2419 		(void) request_cleanup(FLUSH_REMOVE, 0);
2420 	num_dirrem += 1;
2421 	MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
2422 		M_DIRREM, M_WAITOK);
2423 	bzero(dirrem, sizeof(struct dirrem));
2424 	dirrem->dm_list.wk_type = D_DIRREM;
2425 	dirrem->dm_state = isrmdir ? RMDIR : 0;
2426 	dirrem->dm_mnt = ITOV(ip)->v_mount;
2427 	dirrem->dm_oldinum = ip->i_number;
2428 	*prevdirremp = NULL;
2429 
2430 	ACQUIRE_LOCK(&lk);
2431 	lbn = lblkno(dp->i_fs, dp->i_offset);
2432 	offset = blkoff(dp->i_fs, dp->i_offset);
2433 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2434 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2435 	dirrem->dm_pagedep = pagedep;
2436 	/*
2437 	 * Check for a diradd dependency for the same directory entry.
2438 	 * If present, then both dependencies become obsolete and can
2439 	 * be de-allocated. Check for an entry on both the pd_dirraddhd
2440 	 * list and the pd_pendinghd list.
2441 	 */
2442 	for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(offset)]);
2443 	     dap; dap = LIST_NEXT(dap, da_pdlist))
2444 		if (dap->da_offset == offset)
2445 			break;
2446 	if (dap == NULL) {
2447 		for (dap = LIST_FIRST(&pagedep->pd_pendinghd);
2448 		     dap; dap = LIST_NEXT(dap, da_pdlist))
2449 			if (dap->da_offset == offset)
2450 				break;
2451 		if (dap == NULL)
2452 			return (dirrem);
2453 	}
2454 	/*
2455 	 * Must be ATTACHED at this point.
2456 	 */
2457 	if ((dap->da_state & ATTACHED) == 0)
2458 		panic("newdirrem: not ATTACHED");
2459 	if (dap->da_newinum != ip->i_number)
2460 		panic("newdirrem: inum %d should be %d",
2461 		    ip->i_number, dap->da_newinum);
2462 	/*
2463 	 * If we are deleting a changed name that never made it to disk,
2464 	 * then return the dirrem describing the previous inode (which
2465 	 * represents the inode currently referenced from this entry on disk).
2466 	 */
2467 	if ((dap->da_state & DIRCHG) != 0) {
2468 		*prevdirremp = dap->da_previous;
2469 		dap->da_state &= ~DIRCHG;
2470 		dap->da_pagedep = pagedep;
2471 	}
2472 	/*
2473 	 * We are deleting an entry that never made it to disk.
2474 	 * Mark it COMPLETE so we can delete its inode immediately.
2475 	 */
2476 	dirrem->dm_state |= COMPLETE;
2477 	free_diradd(dap);
2478 	return (dirrem);
2479 }
2480 
2481 /*
2482  * Directory entry change dependencies.
2483  *
2484  * Changing an existing directory entry requires that an add operation
2485  * be completed first followed by a deletion. The semantics for the addition
2486  * are identical to the description of adding a new entry above except
2487  * that the rollback is to the old inode number rather than zero. Once
2488  * the addition dependency is completed, the removal is done as described
2489  * in the removal routine above.
2490  */
2491 
2492 /*
2493  * This routine should be called immediately after changing
2494  * a directory entry.  The inode's link count should not be
2495  * decremented by the calling procedure -- the soft updates
2496  * code will perform this task when it is safe.
2497  */
2498 void
2499 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
2500 	struct buf *bp;		/* buffer containing directory block */
2501 	struct inode *dp;	/* inode for the directory being modified */
2502 	struct inode *ip;	/* inode for directory entry being removed */
2503 	long newinum;		/* new inode number for changed entry */
2504 	int isrmdir;		/* indicates if doing RMDIR */
2505 {
2506 	int offset;
2507 	struct diradd *dap = NULL;
2508 	struct dirrem *dirrem, *prevdirrem;
2509 	struct pagedep *pagedep;
2510 	struct inodedep *inodedep;
2511 
2512 	offset = blkoff(dp->i_fs, dp->i_offset);
2513 
2514 	/*
2515 	 * Whiteouts do not need diradd dependencies.
2516 	 */
2517 	if (newinum != WINO) {
2518 		MALLOC(dap, struct diradd *, sizeof(struct diradd),
2519 		    M_DIRADD, M_WAITOK);
2520 		bzero(dap, sizeof(struct diradd));
2521 		dap->da_list.wk_type = D_DIRADD;
2522 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2523 		dap->da_offset = offset;
2524 		dap->da_newinum = newinum;
2525 	}
2526 
2527 	/*
2528 	 * Allocate a new dirrem and ACQUIRE_LOCK.
2529 	 */
2530 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2531 	pagedep = dirrem->dm_pagedep;
2532 	/*
2533 	 * The possible values for isrmdir:
2534 	 *	0 - non-directory file rename
2535 	 *	1 - directory rename within same directory
2536 	 *   inum - directory rename to new directory of given inode number
2537 	 * When renaming to a new directory, we are both deleting and
2538 	 * creating a new directory entry, so the link count on the new
2539 	 * directory should not change. Thus we do not need the followup
2540 	 * dirrem which is usually done in handle_workitem_remove. We set
2541 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
2542 	 * followup dirrem.
2543 	 */
2544 	if (isrmdir > 1)
2545 		dirrem->dm_state |= DIRCHG;
2546 
2547 	/*
2548 	 * Whiteouts have no additional dependencies,
2549 	 * so just put the dirrem on the correct list.
2550 	 */
2551 	if (newinum == WINO) {
2552 		if ((dirrem->dm_state & COMPLETE) == 0) {
2553 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2554 			    dm_next);
2555 		} else {
2556 			dirrem->dm_dirinum = pagedep->pd_ino;
2557 			add_to_worklist(&dirrem->dm_list);
2558 		}
2559 		FREE_LOCK(&lk);
2560 		return;
2561 	}
2562 
2563 	/*
2564 	 * If the COMPLETE flag is clear, then there were no active
2565 	 * entries and we want to roll back to the previous inode until
2566 	 * the new inode is committed to disk. If the COMPLETE flag is
2567 	 * set, then we have deleted an entry that never made it to disk.
2568 	 * If the entry we deleted resulted from a name change, then the old
2569 	 * inode reference still resides on disk. Any rollback that we do
2570 	 * needs to be to that old inode (returned to us in prevdirrem). If
2571 	 * the entry we deleted resulted from a create, then there is
2572 	 * no entry on the disk, so we want to roll back to zero rather
2573 	 * than the uncommitted inode. In either of the COMPLETE cases we
2574 	 * want to immediately free the unwritten and unreferenced inode.
2575 	 */
2576 	if ((dirrem->dm_state & COMPLETE) == 0) {
2577 		dap->da_previous = dirrem;
2578 	} else {
2579 		if (prevdirrem != NULL) {
2580 			dap->da_previous = prevdirrem;
2581 		} else {
2582 			dap->da_state &= ~DIRCHG;
2583 			dap->da_pagedep = pagedep;
2584 		}
2585 		dirrem->dm_dirinum = pagedep->pd_ino;
2586 		add_to_worklist(&dirrem->dm_list);
2587 	}
2588 	/*
2589 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
2590 	 * is not yet written. If it is written, do the post-inode write
2591 	 * processing to put it on the id_pendinghd list.
2592 	 */
2593 	if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2594 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2595 		dap->da_state |= COMPLETE;
2596 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2597 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2598 	} else {
2599 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2600 		    dap, da_pdlist);
2601 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2602 	}
2603 	FREE_LOCK(&lk);
2604 }
2605 
2606 /*
2607  * Called whenever the link count on an inode is changed.
2608  * It creates an inode dependency so that the new reference(s)
2609  * to the inode cannot be committed to disk until the updated
2610  * inode has been written.
2611  */
2612 void
2613 softdep_change_linkcnt(ip)
2614 	struct inode *ip;	/* the inode with the increased link count */
2615 {
2616 	struct inodedep *inodedep;
2617 
2618 	ACQUIRE_LOCK(&lk);
2619 	(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2620 	if (ip->i_nlink < ip->i_effnlink)
2621 		panic("softdep_change_linkcnt: bad delta");
2622 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2623 	FREE_LOCK(&lk);
2624 }
2625 
2626 /*
2627  * This workitem decrements the inode's link count.
2628  * If the link count reaches zero, the file is removed.
2629  */
2630 static void
2631 handle_workitem_remove(dirrem)
2632 	struct dirrem *dirrem;
2633 {
2634 	struct proc *p = CURPROC;	/* XXX */
2635 	struct inodedep *inodedep;
2636 	struct vnode *vp;
2637 	struct inode *ip;
2638 	ino_t oldinum;
2639 	int error;
2640 
2641 	if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) {
2642 		softdep_error("handle_workitem_remove: vget", error);
2643 		return;
2644 	}
2645 	ip = VTOI(vp);
2646 	ACQUIRE_LOCK(&lk);
2647 	if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0)
2648 		panic("handle_workitem_remove: lost inodedep");
2649 	/*
2650 	 * Normal file deletion.
2651 	 */
2652 	if ((dirrem->dm_state & RMDIR) == 0) {
2653 		ip->i_nlink--;
2654 		ip->i_flag |= IN_CHANGE;
2655 		if (ip->i_nlink < ip->i_effnlink)
2656 			panic("handle_workitem_remove: bad file delta");
2657 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2658 		FREE_LOCK(&lk);
2659 		vput(vp);
2660 		num_dirrem -= 1;
2661 		WORKITEM_FREE(dirrem, D_DIRREM);
2662 		return;
2663 	}
2664 	/*
2665 	 * Directory deletion. Decrement reference count for both the
2666 	 * just deleted parent directory entry and the reference for ".".
2667 	 * Next truncate the directory to length zero. When the
2668 	 * truncation completes, arrange to have the reference count on
2669 	 * the parent decremented to account for the loss of "..".
2670 	 */
2671 	ip->i_nlink -= 2;
2672 	ip->i_flag |= IN_CHANGE;
2673 	if (ip->i_nlink < ip->i_effnlink)
2674 		panic("handle_workitem_remove: bad dir delta");
2675 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2676 	FREE_LOCK(&lk);
2677 	if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, p->p_ucred, p)) != 0)
2678 		softdep_error("handle_workitem_remove: truncate", error);
2679 	/*
2680 	 * Rename a directory to a new parent. Since, we are both deleting
2681 	 * and creating a new directory entry, the link count on the new
2682 	 * directory should not change. Thus we skip the followup dirrem.
2683 	 */
2684 	if (dirrem->dm_state & DIRCHG) {
2685 		vput(vp);
2686 		num_dirrem -= 1;
2687 		WORKITEM_FREE(dirrem, D_DIRREM);
2688 		return;
2689 	}
2690 	/*
2691 	 * If the inodedep does not exist, then the zero'ed inode has
2692 	 * been written to disk. If the allocated inode has never been
2693 	 * written to disk, then the on-disk inode is zero'ed. In either
2694 	 * case we can remove the file immediately.
2695 	 */
2696 	ACQUIRE_LOCK(&lk);
2697 	dirrem->dm_state = 0;
2698 	oldinum = dirrem->dm_oldinum;
2699 	dirrem->dm_oldinum = dirrem->dm_dirinum;
2700 	if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
2701 	    check_inode_unwritten(inodedep)) {
2702 		FREE_LOCK(&lk);
2703 		vput(vp);
2704 		handle_workitem_remove(dirrem);
2705 		return;
2706 	}
2707 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
2708 	FREE_LOCK(&lk);
2709 	vput(vp);
2710 }
2711 
2712 /*
2713  * Inode de-allocation dependencies.
2714  *
2715  * When an inode's link count is reduced to zero, it can be de-allocated. We
2716  * found it convenient to postpone de-allocation until after the inode is
2717  * written to disk with its new link count (zero).  At this point, all of the
2718  * on-disk inode's block pointers are nullified and, with careful dependency
2719  * list ordering, all dependencies related to the inode will be satisfied and
2720  * the corresponding dependency structures de-allocated.  So, if/when the
2721  * inode is reused, there will be no mixing of old dependencies with new
2722  * ones.  This artificial dependency is set up by the block de-allocation
2723  * procedure above (softdep_setup_freeblocks) and completed by the
2724  * following procedure.
2725  */
2726 static void
2727 handle_workitem_freefile(freefile)
2728 	struct freefile *freefile;
2729 {
2730 	struct vnode vp;
2731 	struct inode tip;
2732 	struct inodedep *idp;
2733 	int error;
2734 
2735 #ifdef DEBUG
2736 	ACQUIRE_LOCK(&lk);
2737 	if (inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp))
2738 		panic("handle_workitem_freefile: inodedep survived");
2739 	FREE_LOCK(&lk);
2740 #endif
2741 	tip.i_devvp = freefile->fx_devvp;
2742 	tip.i_dev = freefile->fx_devvp->v_rdev;
2743 	tip.i_fs = freefile->fx_fs;
2744 	vp.v_data = &tip;
2745 	if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
2746 		softdep_error("handle_workitem_freefile", error);
2747 	WORKITEM_FREE(freefile, D_FREEFILE);
2748 }
2749 
2750 /*
2751  * Disk writes.
2752  *
2753  * The dependency structures constructed above are most actively used when file
2754  * system blocks are written to disk.  No constraints are placed on when a
2755  * block can be written, but unsatisfied update dependencies are made safe by
2756  * modifying (or replacing) the source memory for the duration of the disk
2757  * write.  When the disk write completes, the memory block is again brought
2758  * up-to-date.
2759  *
2760  * In-core inode structure reclamation.
2761  *
2762  * Because there are a finite number of "in-core" inode structures, they are
2763  * reused regularly.  By transferring all inode-related dependencies to the
2764  * in-memory inode block and indexing them separately (via "inodedep"s), we
2765  * can allow "in-core" inode structures to be reused at any time and avoid
2766  * any increase in contention.
2767  *
2768  * Called just before entering the device driver to initiate a new disk I/O.
2769  * The buffer must be locked, thus, no I/O completion operations can occur
2770  * while we are manipulating its associated dependencies.
2771  */
2772 static void
2773 softdep_disk_io_initiation(bp)
2774 	struct buf *bp;		/* structure describing disk write to occur */
2775 {
2776 	struct worklist *wk, *nextwk;
2777 	struct indirdep *indirdep;
2778 
2779 	/*
2780 	 * We only care about write operations. There should never
2781 	 * be dependencies for reads.
2782 	 */
2783 	if (bp->b_flags & B_READ)
2784 		panic("softdep_disk_io_initiation: read");
2785 	/*
2786 	 * Do any necessary pre-I/O processing.
2787 	 */
2788 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
2789 		nextwk = LIST_NEXT(wk, wk_list);
2790 		switch (wk->wk_type) {
2791 
2792 		case D_PAGEDEP:
2793 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
2794 			continue;
2795 
2796 		case D_INODEDEP:
2797 			initiate_write_inodeblock(WK_INODEDEP(wk), bp);
2798 			continue;
2799 
2800 		case D_INDIRDEP:
2801 			indirdep = WK_INDIRDEP(wk);
2802 			if (indirdep->ir_state & GOINGAWAY)
2803 				panic("disk_io_initiation: indirdep gone");
2804 			/*
2805 			 * If there are no remaining dependencies, this
2806 			 * will be writing the real pointers, so the
2807 			 * dependency can be freed.
2808 			 */
2809 			if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
2810 				indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
2811 				brelse(indirdep->ir_savebp);
2812 				/* inline expand WORKLIST_REMOVE(wk); */
2813 				wk->wk_state &= ~ONWORKLIST;
2814 				LIST_REMOVE(wk, wk_list);
2815 				WORKITEM_FREE(indirdep, D_INDIRDEP);
2816 				continue;
2817 			}
2818 			/*
2819 			 * Replace up-to-date version with safe version.
2820 			 */
2821 			ACQUIRE_LOCK(&lk);
2822 			indirdep->ir_state &= ~ATTACHED;
2823 			indirdep->ir_state |= UNDONE;
2824 			MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
2825 			    M_INDIRDEP, M_WAITOK);
2826 			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
2827 			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
2828 			    bp->b_bcount);
2829 			FREE_LOCK(&lk);
2830 			continue;
2831 
2832 		case D_MKDIR:
2833 		case D_BMSAFEMAP:
2834 		case D_ALLOCDIRECT:
2835 		case D_ALLOCINDIR:
2836 			continue;
2837 
2838 		default:
2839 			panic("handle_disk_io_initiation: Unexpected type %s",
2840 			    TYPENAME(wk->wk_type));
2841 			/* NOTREACHED */
2842 		}
2843 	}
2844 }
2845 
2846 /*
2847  * Called from within the procedure above to deal with unsatisfied
2848  * allocation dependencies in a directory. The buffer must be locked,
2849  * thus, no I/O completion operations can occur while we are
2850  * manipulating its associated dependencies.
2851  */
2852 static void
2853 initiate_write_filepage(pagedep, bp)
2854 	struct pagedep *pagedep;
2855 	struct buf *bp;
2856 {
2857 	struct diradd *dap;
2858 	struct direct *ep;
2859 	int i;
2860 
2861 	if (pagedep->pd_state & IOSTARTED) {
2862 		/*
2863 		 * This can only happen if there is a driver that does not
2864 		 * understand chaining. Here biodone will reissue the call
2865 		 * to strategy for the incomplete buffers.
2866 		 */
2867 		printf("initiate_write_filepage: already started\n");
2868 		return;
2869 	}
2870 	pagedep->pd_state |= IOSTARTED;
2871 	ACQUIRE_LOCK(&lk);
2872 	for (i = 0; i < DAHASHSZ; i++) {
2873 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
2874 		     dap = LIST_NEXT(dap, da_pdlist)) {
2875 			ep = (struct direct *)
2876 			    ((char *)bp->b_data + dap->da_offset);
2877 			if (ep->d_ino != dap->da_newinum)
2878 				panic("%s: dir inum %d != new %d",
2879 				    "initiate_write_filepage",
2880 				    ep->d_ino, dap->da_newinum);
2881 			if (dap->da_state & DIRCHG)
2882 				ep->d_ino = dap->da_previous->dm_oldinum;
2883 			else
2884 				ep->d_ino = 0;
2885 			dap->da_state &= ~ATTACHED;
2886 			dap->da_state |= UNDONE;
2887 		}
2888 	}
2889 	FREE_LOCK(&lk);
2890 }
2891 
2892 /*
2893  * Called from within the procedure above to deal with unsatisfied
2894  * allocation dependencies in an inodeblock. The buffer must be
2895  * locked, thus, no I/O completion operations can occur while we
2896  * are manipulating its associated dependencies.
2897  */
2898 static void
2899 initiate_write_inodeblock(inodedep, bp)
2900 	struct inodedep *inodedep;
2901 	struct buf *bp;			/* The inode block */
2902 {
2903 	struct allocdirect *adp, *lastadp;
2904 	struct dinode *dp;
2905 	struct fs *fs;
2906 	ufs_lbn_t prevlbn = 0;
2907 	int i, deplist;
2908 
2909 	if (inodedep->id_state & IOSTARTED)
2910 		panic("initiate_write_inodeblock: already started");
2911 	inodedep->id_state |= IOSTARTED;
2912 	fs = inodedep->id_fs;
2913 	dp = (struct dinode *)bp->b_data +
2914 	    ino_to_fsbo(fs, inodedep->id_ino);
2915 	/*
2916 	 * If the bitmap is not yet written, then the allocated
2917 	 * inode cannot be written to disk.
2918 	 */
2919 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
2920 		if (inodedep->id_savedino != NULL)
2921 			panic("initiate_write_inodeblock: already doing I/O");
2922 		MALLOC(inodedep->id_savedino, struct dinode *,
2923 		    sizeof(struct dinode), M_INODEDEP, M_WAITOK);
2924 		*inodedep->id_savedino = *dp;
2925 		bzero((caddr_t)dp, sizeof(struct dinode));
2926 		return;
2927 	}
2928 	/*
2929 	 * If no dependencies, then there is nothing to roll back.
2930 	 */
2931 	inodedep->id_savedsize = dp->di_size;
2932 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
2933 		return;
2934 	/*
2935 	 * Set the dependencies to busy.
2936 	 */
2937 	ACQUIRE_LOCK(&lk);
2938 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
2939 	     adp = TAILQ_NEXT(adp, ad_next)) {
2940 #ifdef DIAGNOSTIC
2941 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
2942 			panic("softdep_write_inodeblock: lbn order");
2943 		prevlbn = adp->ad_lbn;
2944 		if (adp->ad_lbn < NDADDR &&
2945 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno)
2946 			panic("%s: direct pointer #%ld mismatch %d != %d",
2947 			    "softdep_write_inodeblock", adp->ad_lbn,
2948 			    dp->di_db[adp->ad_lbn], adp->ad_newblkno);
2949 		if (adp->ad_lbn >= NDADDR &&
2950 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno)
2951 			panic("%s: indirect pointer #%ld mismatch %d != %d",
2952 			    "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
2953 			    dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
2954 		deplist |= 1 << adp->ad_lbn;
2955 		if ((adp->ad_state & ATTACHED) == 0)
2956 			panic("softdep_write_inodeblock: Unknown state 0x%x",
2957 			    adp->ad_state);
2958 #endif /* DIAGNOSTIC */
2959 		adp->ad_state &= ~ATTACHED;
2960 		adp->ad_state |= UNDONE;
2961 	}
2962 	/*
2963 	 * The on-disk inode cannot claim to be any larger than the last
2964 	 * fragment that has been written. Otherwise, the on-disk inode
2965 	 * might have fragments that were not the last block in the file
2966 	 * which would corrupt the filesystem.
2967 	 */
2968 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
2969 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
2970 		if (adp->ad_lbn >= NDADDR)
2971 			break;
2972 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
2973 		/* keep going until hitting a rollback to a frag */
2974 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
2975 			continue;
2976 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
2977 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
2978 #ifdef DIAGNOSTIC
2979 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
2980 				panic("softdep_write_inodeblock: lost dep1");
2981 #endif /* DIAGNOSTIC */
2982 			dp->di_db[i] = 0;
2983 		}
2984 		for (i = 0; i < NIADDR; i++) {
2985 #ifdef DIAGNOSTIC
2986 			if (dp->di_ib[i] != 0 &&
2987 			    (deplist & ((1 << NDADDR) << i)) == 0)
2988 				panic("softdep_write_inodeblock: lost dep2");
2989 #endif /* DIAGNOSTIC */
2990 			dp->di_ib[i] = 0;
2991 		}
2992 		FREE_LOCK(&lk);
2993 		return;
2994 	}
2995 	/*
2996 	 * If we have zero'ed out the last allocated block of the file,
2997 	 * roll back the size to the last currently allocated block.
2998 	 * We know that this last allocated block is a full-sized as
2999 	 * we already checked for fragments in the loop above.
3000 	 */
3001 	if (lastadp != NULL &&
3002 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3003 		for (i = lastadp->ad_lbn; i >= 0; i--)
3004 			if (dp->di_db[i] != 0)
3005 				break;
3006 		dp->di_size = (i + 1) * fs->fs_bsize;
3007 	}
3008 	/*
3009 	 * The only dependencies are for indirect blocks.
3010 	 *
3011 	 * The file size for indirect block additions is not guaranteed.
3012 	 * Such a guarantee would be non-trivial to achieve. The conventional
3013 	 * synchronous write implementation also does not make this guarantee.
3014 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3015 	 * can be over-estimated without destroying integrity when the file
3016 	 * moves into the indirect blocks (i.e., is large). If we want to
3017 	 * postpone fsck, we are stuck with this argument.
3018 	 */
3019 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3020 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3021 	FREE_LOCK(&lk);
3022 }
3023 
3024 /*
3025  * This routine is called during the completion interrupt
3026  * service routine for a disk write (from the procedure called
3027  * by the device driver to inform the file system caches of
3028  * a request completion).  It should be called early in this
3029  * procedure, before the block is made available to other
3030  * processes or other routines are called.
3031  */
3032 static void
3033 softdep_disk_write_complete(bp)
3034 	struct buf *bp;		/* describes the completed disk write */
3035 {
3036 	struct worklist *wk;
3037 	struct workhead reattach;
3038 	struct newblk *newblk;
3039 	struct allocindir *aip;
3040 	struct allocdirect *adp;
3041 	struct indirdep *indirdep;
3042 	struct inodedep *inodedep;
3043 	struct bmsafemap *bmsafemap;
3044 
3045 #ifdef DEBUG
3046 	if (lk.lkt_held != -1)
3047 		panic("softdep_disk_write_complete: lock is held");
3048 	lk.lkt_held = -2;
3049 #endif
3050 	LIST_INIT(&reattach);
3051 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3052 		WORKLIST_REMOVE(wk);
3053 		switch (wk->wk_type) {
3054 
3055 		case D_PAGEDEP:
3056 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3057 				WORKLIST_INSERT(&reattach, wk);
3058 			continue;
3059 
3060 		case D_INODEDEP:
3061 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3062 				WORKLIST_INSERT(&reattach, wk);
3063 			continue;
3064 
3065 		case D_BMSAFEMAP:
3066 			bmsafemap = WK_BMSAFEMAP(wk);
3067 			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3068 				newblk->nb_state |= DEPCOMPLETE;
3069 				newblk->nb_bmsafemap = NULL;
3070 				LIST_REMOVE(newblk, nb_deps);
3071 			}
3072 			while ((adp =
3073 			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3074 				adp->ad_state |= DEPCOMPLETE;
3075 				adp->ad_buf = NULL;
3076 				LIST_REMOVE(adp, ad_deps);
3077 				handle_allocdirect_partdone(adp);
3078 			}
3079 			while ((aip =
3080 			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3081 				aip->ai_state |= DEPCOMPLETE;
3082 				aip->ai_buf = NULL;
3083 				LIST_REMOVE(aip, ai_deps);
3084 				handle_allocindir_partdone(aip);
3085 			}
3086 			while ((inodedep =
3087 			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3088 				inodedep->id_state |= DEPCOMPLETE;
3089 				LIST_REMOVE(inodedep, id_deps);
3090 				inodedep->id_buf = NULL;
3091 			}
3092 			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3093 			continue;
3094 
3095 		case D_MKDIR:
3096 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3097 			continue;
3098 
3099 		case D_ALLOCDIRECT:
3100 			adp = WK_ALLOCDIRECT(wk);
3101 			adp->ad_state |= COMPLETE;
3102 			handle_allocdirect_partdone(adp);
3103 			continue;
3104 
3105 		case D_ALLOCINDIR:
3106 			aip = WK_ALLOCINDIR(wk);
3107 			aip->ai_state |= COMPLETE;
3108 			handle_allocindir_partdone(aip);
3109 			continue;
3110 
3111 		case D_INDIRDEP:
3112 			indirdep = WK_INDIRDEP(wk);
3113 			if (indirdep->ir_state & GOINGAWAY)
3114 				panic("disk_write_complete: indirdep gone");
3115 			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3116 			FREE(indirdep->ir_saveddata, M_INDIRDEP);
3117 			indirdep->ir_saveddata = 0;
3118 			indirdep->ir_state &= ~UNDONE;
3119 			indirdep->ir_state |= ATTACHED;
3120 			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3121 				handle_allocindir_partdone(aip);
3122 				if (aip == LIST_FIRST(&indirdep->ir_donehd))
3123 					panic("disk_write_complete: not gone");
3124 			}
3125 			WORKLIST_INSERT(&reattach, wk);
3126 			if ((bp->b_flags & B_DELWRI) == 0)
3127 				stat_indir_blk_ptrs++;
3128 			bdirty(bp);
3129 			continue;
3130 
3131 		default:
3132 			panic("handle_disk_write_complete: Unknown type %s",
3133 			    TYPENAME(wk->wk_type));
3134 			/* NOTREACHED */
3135 		}
3136 	}
3137 	/*
3138 	 * Reattach any requests that must be redone.
3139 	 */
3140 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
3141 		WORKLIST_REMOVE(wk);
3142 		WORKLIST_INSERT(&bp->b_dep, wk);
3143 	}
3144 #ifdef DEBUG
3145 	if (lk.lkt_held != -2)
3146 		panic("softdep_disk_write_complete: lock lost");
3147 	lk.lkt_held = -1;
3148 #endif
3149 }
3150 
3151 /*
3152  * Called from within softdep_disk_write_complete above. Note that
3153  * this routine is always called from interrupt level with further
3154  * splbio interrupts blocked.
3155  */
3156 static void
3157 handle_allocdirect_partdone(adp)
3158 	struct allocdirect *adp;	/* the completed allocdirect */
3159 {
3160 	struct allocdirect *listadp;
3161 	struct inodedep *inodedep;
3162 	long bsize;
3163 
3164 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3165 		return;
3166 	if (adp->ad_buf != NULL)
3167 		panic("handle_allocdirect_partdone: dangling dep");
3168 	/*
3169 	 * The on-disk inode cannot claim to be any larger than the last
3170 	 * fragment that has been written. Otherwise, the on-disk inode
3171 	 * might have fragments that were not the last block in the file
3172 	 * which would corrupt the filesystem. Thus, we cannot free any
3173 	 * allocdirects after one whose ad_oldblkno claims a fragment as
3174 	 * these blocks must be rolled back to zero before writing the inode.
3175 	 * We check the currently active set of allocdirects in id_inoupdt.
3176 	 */
3177 	inodedep = adp->ad_inodedep;
3178 	bsize = inodedep->id_fs->fs_bsize;
3179 	for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp;
3180 	     listadp = TAILQ_NEXT(listadp, ad_next)) {
3181 		/* found our block */
3182 		if (listadp == adp)
3183 			break;
3184 		/* continue if ad_oldlbn is not a fragment */
3185 		if (listadp->ad_oldsize == 0 ||
3186 		    listadp->ad_oldsize == bsize)
3187 			continue;
3188 		/* hit a fragment */
3189 		return;
3190 	}
3191 	/*
3192 	 * If we have reached the end of the current list without
3193 	 * finding the just finished dependency, then it must be
3194 	 * on the future dependency list. Future dependencies cannot
3195 	 * be freed until they are moved to the current list.
3196 	 */
3197 	if (listadp == NULL) {
3198 #ifdef DEBUG
3199 		for (listadp = TAILQ_FIRST(&inodedep->id_newinoupdt); listadp;
3200 		     listadp = TAILQ_NEXT(listadp, ad_next))
3201 			/* found our block */
3202 			if (listadp == adp)
3203 				break;
3204 		if (listadp == NULL)
3205 			panic("handle_allocdirect_partdone: lost dep");
3206 #endif /* DEBUG */
3207 		return;
3208 	}
3209 	/*
3210 	 * If we have found the just finished dependency, then free
3211 	 * it along with anything that follows it that is complete.
3212 	 */
3213 	for (; adp; adp = listadp) {
3214 		listadp = TAILQ_NEXT(adp, ad_next);
3215 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3216 			return;
3217 		free_allocdirect(&inodedep->id_inoupdt, adp, 1);
3218 	}
3219 }
3220 
3221 /*
3222  * Called from within softdep_disk_write_complete above. Note that
3223  * this routine is always called from interrupt level with further
3224  * splbio interrupts blocked.
3225  */
3226 static void
3227 handle_allocindir_partdone(aip)
3228 	struct allocindir *aip;		/* the completed allocindir */
3229 {
3230 	struct indirdep *indirdep;
3231 
3232 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
3233 		return;
3234 	if (aip->ai_buf != NULL)
3235 		panic("handle_allocindir_partdone: dangling dependency");
3236 	indirdep = aip->ai_indirdep;
3237 	if (indirdep->ir_state & UNDONE) {
3238 		LIST_REMOVE(aip, ai_next);
3239 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
3240 		return;
3241 	}
3242 	((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
3243 	    aip->ai_newblkno;
3244 	LIST_REMOVE(aip, ai_next);
3245 	if (aip->ai_freefrag != NULL)
3246 		add_to_worklist(&aip->ai_freefrag->ff_list);
3247 	WORKITEM_FREE(aip, D_ALLOCINDIR);
3248 }
3249 
3250 /*
3251  * Called from within softdep_disk_write_complete above to restore
3252  * in-memory inode block contents to their most up-to-date state. Note
3253  * that this routine is always called from interrupt level with further
3254  * splbio interrupts blocked.
3255  */
3256 static int
3257 handle_written_inodeblock(inodedep, bp)
3258 	struct inodedep *inodedep;
3259 	struct buf *bp;		/* buffer containing the inode block */
3260 {
3261 	struct worklist *wk, *filefree;
3262 	struct allocdirect *adp, *nextadp;
3263 	struct dinode *dp;
3264 	int hadchanges;
3265 
3266 	if ((inodedep->id_state & IOSTARTED) == 0)
3267 		panic("handle_written_inodeblock: not started");
3268 	inodedep->id_state &= ~IOSTARTED;
3269 	inodedep->id_state |= COMPLETE;
3270 	dp = (struct dinode *)bp->b_data +
3271 	    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
3272 	/*
3273 	 * If we had to rollback the inode allocation because of
3274 	 * bitmaps being incomplete, then simply restore it.
3275 	 * Keep the block dirty so that it will not be reclaimed until
3276 	 * all associated dependencies have been cleared and the
3277 	 * corresponding updates written to disk.
3278 	 */
3279 	if (inodedep->id_savedino != NULL) {
3280 		*dp = *inodedep->id_savedino;
3281 		FREE(inodedep->id_savedino, M_INODEDEP);
3282 		inodedep->id_savedino = NULL;
3283 		if ((bp->b_flags & B_DELWRI) == 0)
3284 			stat_inode_bitmap++;
3285 		bdirty(bp);
3286 		return (1);
3287 	}
3288 	/*
3289 	 * Roll forward anything that had to be rolled back before
3290 	 * the inode could be updated.
3291 	 */
3292 	hadchanges = 0;
3293 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
3294 		nextadp = TAILQ_NEXT(adp, ad_next);
3295 		if (adp->ad_state & ATTACHED)
3296 			panic("handle_written_inodeblock: new entry");
3297 		if (adp->ad_lbn < NDADDR) {
3298 			if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno)
3299 				panic("%s: %s #%ld mismatch %d != %d",
3300 				    "handle_written_inodeblock",
3301 				    "direct pointer", adp->ad_lbn,
3302 				    dp->di_db[adp->ad_lbn], adp->ad_oldblkno);
3303 			dp->di_db[adp->ad_lbn] = adp->ad_newblkno;
3304 		} else {
3305 			if (dp->di_ib[adp->ad_lbn - NDADDR] != 0)
3306 				panic("%s: %s #%ld allocated as %d",
3307 				    "handle_written_inodeblock",
3308 				    "indirect pointer", adp->ad_lbn - NDADDR,
3309 				    dp->di_ib[adp->ad_lbn - NDADDR]);
3310 			dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno;
3311 		}
3312 		adp->ad_state &= ~UNDONE;
3313 		adp->ad_state |= ATTACHED;
3314 		hadchanges = 1;
3315 	}
3316 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
3317 		stat_direct_blk_ptrs++;
3318 	/*
3319 	 * Reset the file size to its most up-to-date value.
3320 	 */
3321 	if (inodedep->id_savedsize == -1)
3322 		panic("handle_written_inodeblock: bad size");
3323 	if (dp->di_size != inodedep->id_savedsize) {
3324 		dp->di_size = inodedep->id_savedsize;
3325 		hadchanges = 1;
3326 	}
3327 	inodedep->id_savedsize = -1;
3328 	/*
3329 	 * If there were any rollbacks in the inode block, then it must be
3330 	 * marked dirty so that its will eventually get written back in
3331 	 * its correct form.
3332 	 */
3333 	if (hadchanges)
3334 		bdirty(bp);
3335 	/*
3336 	 * Process any allocdirects that completed during the update.
3337 	 */
3338 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
3339 		handle_allocdirect_partdone(adp);
3340 	/*
3341 	 * Process deallocations that were held pending until the
3342 	 * inode had been written to disk. Freeing of the inode
3343 	 * is delayed until after all blocks have been freed to
3344 	 * avoid creation of new <vfsid, inum, lbn> triples
3345 	 * before the old ones have been deleted.
3346 	 */
3347 	filefree = NULL;
3348 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
3349 		WORKLIST_REMOVE(wk);
3350 		switch (wk->wk_type) {
3351 
3352 		case D_FREEFILE:
3353 			/*
3354 			 * We defer adding filefree to the worklist until
3355 			 * all other additions have been made to ensure
3356 			 * that it will be done after all the old blocks
3357 			 * have been freed.
3358 			 */
3359 			if (filefree != NULL)
3360 				panic("handle_written_inodeblock: filefree");
3361 			filefree = wk;
3362 			continue;
3363 
3364 		case D_MKDIR:
3365 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
3366 			continue;
3367 
3368 		case D_DIRADD:
3369 			diradd_inode_written(WK_DIRADD(wk), inodedep);
3370 			continue;
3371 
3372 		case D_FREEBLKS:
3373 		case D_FREEFRAG:
3374 		case D_DIRREM:
3375 			add_to_worklist(wk);
3376 			continue;
3377 
3378 		default:
3379 			panic("handle_written_inodeblock: Unknown type %s",
3380 			    TYPENAME(wk->wk_type));
3381 			/* NOTREACHED */
3382 		}
3383 	}
3384 	if (filefree != NULL) {
3385 		if (free_inodedep(inodedep) == 0)
3386 			panic("handle_written_inodeblock: live inodedep");
3387 		add_to_worklist(filefree);
3388 		return (0);
3389 	}
3390 
3391 	/*
3392 	 * If no outstanding dependencies, free it.
3393 	 */
3394 	if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0)
3395 		return (0);
3396 	return (hadchanges);
3397 }
3398 
3399 /*
3400  * Process a diradd entry after its dependent inode has been written.
3401  * This routine must be called with splbio interrupts blocked.
3402  */
3403 static void
3404 diradd_inode_written(dap, inodedep)
3405 	struct diradd *dap;
3406 	struct inodedep *inodedep;
3407 {
3408 	struct pagedep *pagedep;
3409 
3410 	dap->da_state |= COMPLETE;
3411 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3412 		if (dap->da_state & DIRCHG)
3413 			pagedep = dap->da_previous->dm_pagedep;
3414 		else
3415 			pagedep = dap->da_pagedep;
3416 		LIST_REMOVE(dap, da_pdlist);
3417 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3418 	}
3419 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3420 }
3421 
3422 /*
3423  * Handle the completion of a mkdir dependency.
3424  */
3425 static void
3426 handle_written_mkdir(mkdir, type)
3427 	struct mkdir *mkdir;
3428 	int type;
3429 {
3430 	struct diradd *dap;
3431 	struct pagedep *pagedep;
3432 
3433 	if (mkdir->md_state != type)
3434 		panic("handle_written_mkdir: bad type");
3435 	dap = mkdir->md_diradd;
3436 	dap->da_state &= ~type;
3437 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
3438 		dap->da_state |= DEPCOMPLETE;
3439 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3440 		if (dap->da_state & DIRCHG)
3441 			pagedep = dap->da_previous->dm_pagedep;
3442 		else
3443 			pagedep = dap->da_pagedep;
3444 		LIST_REMOVE(dap, da_pdlist);
3445 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3446 	}
3447 	LIST_REMOVE(mkdir, md_mkdirs);
3448 	WORKITEM_FREE(mkdir, D_MKDIR);
3449 }
3450 
3451 /*
3452  * Called from within softdep_disk_write_complete above.
3453  * A write operation was just completed. Removed inodes can
3454  * now be freed and associated block pointers may be committed.
3455  * Note that this routine is always called from interrupt level
3456  * with further splbio interrupts blocked.
3457  */
3458 static int
3459 handle_written_filepage(pagedep, bp)
3460 	struct pagedep *pagedep;
3461 	struct buf *bp;		/* buffer containing the written page */
3462 {
3463 	struct dirrem *dirrem;
3464 	struct diradd *dap, *nextdap;
3465 	struct direct *ep;
3466 	int i, chgs;
3467 
3468 	if ((pagedep->pd_state & IOSTARTED) == 0)
3469 		panic("handle_written_filepage: not started");
3470 	pagedep->pd_state &= ~IOSTARTED;
3471 	/*
3472 	 * Process any directory removals that have been committed.
3473 	 */
3474 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
3475 		LIST_REMOVE(dirrem, dm_next);
3476 		dirrem->dm_dirinum = pagedep->pd_ino;
3477 		add_to_worklist(&dirrem->dm_list);
3478 	}
3479 	/*
3480 	 * Free any directory additions that have been committed.
3481 	 */
3482 	while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
3483 		free_diradd(dap);
3484 	/*
3485 	 * Uncommitted directory entries must be restored.
3486 	 */
3487 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
3488 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
3489 		     dap = nextdap) {
3490 			nextdap = LIST_NEXT(dap, da_pdlist);
3491 			if (dap->da_state & ATTACHED)
3492 				panic("handle_written_filepage: attached");
3493 			ep = (struct direct *)
3494 			    ((char *)bp->b_data + dap->da_offset);
3495 			ep->d_ino = dap->da_newinum;
3496 			dap->da_state &= ~UNDONE;
3497 			dap->da_state |= ATTACHED;
3498 			chgs = 1;
3499 			/*
3500 			 * If the inode referenced by the directory has
3501 			 * been written out, then the dependency can be
3502 			 * moved to the pending list.
3503 			 */
3504 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3505 				LIST_REMOVE(dap, da_pdlist);
3506 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
3507 				    da_pdlist);
3508 			}
3509 		}
3510 	}
3511 	/*
3512 	 * If there were any rollbacks in the directory, then it must be
3513 	 * marked dirty so that its will eventually get written back in
3514 	 * its correct form.
3515 	 */
3516 	if (chgs) {
3517 		if ((bp->b_flags & B_DELWRI) == 0)
3518 			stat_dir_entry++;
3519 		bdirty(bp);
3520 	}
3521 	/*
3522 	 * If no dependencies remain, the pagedep will be freed.
3523 	 * Otherwise it will remain to update the page before it
3524 	 * is written back to disk.
3525 	 */
3526 	if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) {
3527 		for (i = 0; i < DAHASHSZ; i++)
3528 			if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
3529 				break;
3530 		if (i == DAHASHSZ) {
3531 			LIST_REMOVE(pagedep, pd_hash);
3532 			WORKITEM_FREE(pagedep, D_PAGEDEP);
3533 			return (0);
3534 		}
3535 	}
3536 	return (1);
3537 }
3538 
3539 /*
3540  * Writing back in-core inode structures.
3541  *
3542  * The file system only accesses an inode's contents when it occupies an
3543  * "in-core" inode structure.  These "in-core" structures are separate from
3544  * the page frames used to cache inode blocks.  Only the latter are
3545  * transferred to/from the disk.  So, when the updated contents of the
3546  * "in-core" inode structure are copied to the corresponding in-memory inode
3547  * block, the dependencies are also transferred.  The following procedure is
3548  * called when copying a dirty "in-core" inode to a cached inode block.
3549  */
3550 
3551 /*
3552  * Called when an inode is loaded from disk. If the effective link count
3553  * differed from the actual link count when it was last flushed, then we
3554  * need to ensure that the correct effective link count is put back.
3555  */
3556 void
3557 softdep_load_inodeblock(ip)
3558 	struct inode *ip;	/* the "in_core" copy of the inode */
3559 {
3560 	struct inodedep *inodedep;
3561 
3562 	/*
3563 	 * Check for alternate nlink count.
3564 	 */
3565 	ip->i_effnlink = ip->i_nlink;
3566 	ACQUIRE_LOCK(&lk);
3567 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3568 		FREE_LOCK(&lk);
3569 		return;
3570 	}
3571 	ip->i_effnlink -= inodedep->id_nlinkdelta;
3572 	FREE_LOCK(&lk);
3573 }
3574 
3575 /*
3576  * This routine is called just before the "in-core" inode
3577  * information is to be copied to the in-memory inode block.
3578  * Recall that an inode block contains several inodes. If
3579  * the force flag is set, then the dependencies will be
3580  * cleared so that the update can always be made. Note that
3581  * the buffer is locked when this routine is called, so we
3582  * will never be in the middle of writing the inode block
3583  * to disk.
3584  */
3585 void
3586 softdep_update_inodeblock(ip, bp, waitfor)
3587 	struct inode *ip;	/* the "in_core" copy of the inode */
3588 	struct buf *bp;		/* the buffer containing the inode block */
3589 	int waitfor;		/* nonzero => update must be allowed */
3590 {
3591 	struct inodedep *inodedep;
3592 	struct worklist *wk;
3593 	int error, gotit;
3594 
3595 	/*
3596 	 * If the effective link count is not equal to the actual link
3597 	 * count, then we must track the difference in an inodedep while
3598 	 * the inode is (potentially) tossed out of the cache. Otherwise,
3599 	 * if there is no existing inodedep, then there are no dependencies
3600 	 * to track.
3601 	 */
3602 	ACQUIRE_LOCK(&lk);
3603 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3604 		if (ip->i_effnlink != ip->i_nlink)
3605 			panic("softdep_update_inodeblock: bad link count");
3606 		FREE_LOCK(&lk);
3607 		return;
3608 	}
3609 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
3610 		panic("softdep_update_inodeblock: bad delta");
3611 	/*
3612 	 * Changes have been initiated. Anything depending on these
3613 	 * changes cannot occur until this inode has been written.
3614 	 */
3615 	inodedep->id_state &= ~COMPLETE;
3616 	if ((inodedep->id_state & ONWORKLIST) == 0)
3617 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
3618 	/*
3619 	 * Any new dependencies associated with the incore inode must
3620 	 * now be moved to the list associated with the buffer holding
3621 	 * the in-memory copy of the inode. Once merged process any
3622 	 * allocdirects that are completed by the merger.
3623 	 */
3624 	merge_inode_lists(inodedep);
3625 	if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
3626 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
3627 	/*
3628 	 * Now that the inode has been pushed into the buffer, the
3629 	 * operations dependent on the inode being written to disk
3630 	 * can be moved to the id_bufwait so that they will be
3631 	 * processed when the buffer I/O completes.
3632 	 */
3633 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
3634 		WORKLIST_REMOVE(wk);
3635 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
3636 	}
3637 	/*
3638 	 * Newly allocated inodes cannot be written until the bitmap
3639 	 * that allocates them have been written (indicated by
3640 	 * DEPCOMPLETE being set in id_state). If we are doing a
3641 	 * forced sync (e.g., an fsync on a file), we force the bitmap
3642 	 * to be written so that the update can be done.
3643 	 */
3644 	if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
3645 		FREE_LOCK(&lk);
3646 		return;
3647 	}
3648 	gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
3649 	FREE_LOCK(&lk);
3650 	if (gotit &&
3651 	    (error = VOP_BWRITE(inodedep->id_buf->b_vp, inodedep->id_buf)) != 0)
3652 		softdep_error("softdep_update_inodeblock: bwrite", error);
3653 	if ((inodedep->id_state & DEPCOMPLETE) == 0)
3654 		panic("softdep_update_inodeblock: update failed");
3655 }
3656 
3657 /*
3658  * Merge the new inode dependency list (id_newinoupdt) into the old
3659  * inode dependency list (id_inoupdt). This routine must be called
3660  * with splbio interrupts blocked.
3661  */
3662 static void
3663 merge_inode_lists(inodedep)
3664 	struct inodedep *inodedep;
3665 {
3666 	struct allocdirect *listadp, *newadp;
3667 
3668 	newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3669 	for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) {
3670 		if (listadp->ad_lbn < newadp->ad_lbn) {
3671 			listadp = TAILQ_NEXT(listadp, ad_next);
3672 			continue;
3673 		}
3674 		TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3675 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
3676 		if (listadp->ad_lbn == newadp->ad_lbn) {
3677 			allocdirect_merge(&inodedep->id_inoupdt, newadp,
3678 			    listadp);
3679 			listadp = newadp;
3680 		}
3681 		newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3682 	}
3683 	while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) {
3684 		TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3685 		TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next);
3686 	}
3687 }
3688 
3689 /*
3690  * If we are doing an fsync, then we must ensure that any directory
3691  * entries for the inode have been written after the inode gets to disk.
3692  */
3693 static int
3694 softdep_fsync(vp)
3695 	struct vnode *vp;	/* the "in_core" copy of the inode */
3696 {
3697 	struct inodedep *inodedep;
3698 	struct pagedep *pagedep;
3699 	struct worklist *wk;
3700 	struct diradd *dap;
3701 	struct mount *mnt;
3702 	struct vnode *pvp;
3703 	struct inode *ip;
3704 	struct buf *bp;
3705 	struct fs *fs;
3706 	struct proc *p = CURPROC;		/* XXX */
3707 	int error, flushparent;
3708 	ino_t parentino;
3709 	ufs_lbn_t lbn;
3710 
3711 	ip = VTOI(vp);
3712 	fs = ip->i_fs;
3713 	ACQUIRE_LOCK(&lk);
3714 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
3715 		FREE_LOCK(&lk);
3716 		return (0);
3717 	}
3718 	if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
3719 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
3720 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
3721 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL)
3722 		panic("softdep_fsync: pending ops");
3723 	for (error = 0, flushparent = 0; ; ) {
3724 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
3725 			break;
3726 		if (wk->wk_type != D_DIRADD)
3727 			panic("softdep_fsync: Unexpected type %s",
3728 			    TYPENAME(wk->wk_type));
3729 		dap = WK_DIRADD(wk);
3730 		/*
3731 		 * Flush our parent if this directory entry
3732 		 * has a MKDIR_PARENT dependency.
3733 		 */
3734 		if (dap->da_state & DIRCHG)
3735 			pagedep = dap->da_previous->dm_pagedep;
3736 		else
3737 			pagedep = dap->da_pagedep;
3738 		mnt = pagedep->pd_mnt;
3739 		parentino = pagedep->pd_ino;
3740 		lbn = pagedep->pd_lbn;
3741 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
3742 			panic("softdep_fsync: dirty");
3743 		flushparent = dap->da_state & MKDIR_PARENT;
3744 		/*
3745 		 * If we are being fsync'ed as part of vgone'ing this vnode,
3746 		 * then we will not be able to release and recover the
3747 		 * vnode below, so we just have to give up on writing its
3748 		 * directory entry out. It will eventually be written, just
3749 		 * not now, but then the user was not asking to have it
3750 		 * written, so we are not breaking any promises.
3751 		 */
3752 		if (vp->v_flag & VXLOCK)
3753 			break;
3754 		/*
3755 		 * We prevent deadlock by always fetching inodes from the
3756 		 * root, moving down the directory tree. Thus, when fetching
3757 		 * our parent directory, we must unlock ourselves before
3758 		 * requesting the lock on our parent. See the comment in
3759 		 * ufs_lookup for details on possible races.
3760 		 */
3761 		FREE_LOCK(&lk);
3762 		VOP_UNLOCK(vp, 0, p);
3763 		error = VFS_VGET(mnt, parentino, &pvp);
3764 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p);
3765 		if (error != 0)
3766 			return (error);
3767 		if (flushparent) {
3768 			if ((error = UFS_UPDATE(pvp, 1)) != 0) {
3769 				vput(pvp);
3770 				return (error);
3771 			}
3772 		}
3773 		/*
3774 		 * Flush directory page containing the inode's name.
3775 		 */
3776 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), p->p_ucred,
3777 		    &bp);
3778 		if (error == 0)
3779 			error = VOP_BWRITE(bp->b_vp, bp);
3780 		vput(pvp);
3781 		if (error != 0)
3782 			return (error);
3783 		ACQUIRE_LOCK(&lk);
3784 		if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
3785 			break;
3786 	}
3787 	FREE_LOCK(&lk);
3788 	return (0);
3789 }
3790 
3791 /*
3792  * Flush all the dirty bitmaps associated with the block device
3793  * before flushing the rest of the dirty blocks so as to reduce
3794  * the number of dependencies that will have to be rolled back.
3795  */
3796 void
3797 softdep_fsync_mountdev(vp)
3798 	struct vnode *vp;
3799 {
3800 	struct buf *bp, *nbp;
3801 	struct worklist *wk;
3802 
3803 	if (!vn_isdisk(vp, NULL))
3804 		panic("softdep_fsync_mountdev: vnode not a disk");
3805 	ACQUIRE_LOCK(&lk);
3806 	for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
3807 		nbp = TAILQ_NEXT(bp, b_vnbufs);
3808 		/*
3809 		 * If it is already scheduled, skip to the next buffer.
3810 		 */
3811 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
3812 			continue;
3813 		if ((bp->b_flags & B_DELWRI) == 0)
3814 			panic("softdep_fsync_mountdev: not dirty");
3815 		/*
3816 		 * We are only interested in bitmaps with outstanding
3817 		 * dependencies.
3818 		 */
3819 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
3820 		    wk->wk_type != D_BMSAFEMAP ||
3821 		    (bp->b_xflags & BX_BKGRDINPROG)) {
3822 			BUF_UNLOCK(bp);
3823 			continue;
3824 		}
3825 		bremfree(bp);
3826 		FREE_LOCK(&lk);
3827 		(void) bawrite(bp);
3828 		ACQUIRE_LOCK(&lk);
3829 		/*
3830 		 * Since we may have slept during the I/O, we need
3831 		 * to start from a known point.
3832 		 */
3833 		nbp = TAILQ_FIRST(&vp->v_dirtyblkhd);
3834 	}
3835 	drain_output(vp, 1);
3836 	FREE_LOCK(&lk);
3837 }
3838 
3839 /*
3840  * This routine is called when we are trying to synchronously flush a
3841  * file. This routine must eliminate any filesystem metadata dependencies
3842  * so that the syncing routine can succeed by pushing the dirty blocks
3843  * associated with the file. If any I/O errors occur, they are returned.
3844  */
3845 int
3846 softdep_sync_metadata(ap)
3847 	struct vop_fsync_args /* {
3848 		struct vnode *a_vp;
3849 		struct ucred *a_cred;
3850 		int a_waitfor;
3851 		struct proc *a_p;
3852 	} */ *ap;
3853 {
3854 	struct vnode *vp = ap->a_vp;
3855 	struct pagedep *pagedep;
3856 	struct allocdirect *adp;
3857 	struct allocindir *aip;
3858 	struct buf *bp, *nbp;
3859 	struct worklist *wk;
3860 	int i, error, waitfor;
3861 
3862 	/*
3863 	 * Check whether this vnode is involved in a filesystem
3864 	 * that is doing soft dependency processing.
3865 	 */
3866 	if (!vn_isdisk(vp, NULL)) {
3867 		if (!DOINGSOFTDEP(vp))
3868 			return (0);
3869 	} else
3870 		if (vp->v_specmountpoint == NULL ||
3871 		    (vp->v_specmountpoint->mnt_flag & MNT_SOFTDEP) == 0)
3872 			return (0);
3873 	/*
3874 	 * Ensure that any direct block dependencies have been cleared.
3875 	 */
3876 	ACQUIRE_LOCK(&lk);
3877 	if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
3878 		FREE_LOCK(&lk);
3879 		return (error);
3880 	}
3881 	/*
3882 	 * For most files, the only metadata dependencies are the
3883 	 * cylinder group maps that allocate their inode or blocks.
3884 	 * The block allocation dependencies can be found by traversing
3885 	 * the dependency lists for any buffers that remain on their
3886 	 * dirty buffer list. The inode allocation dependency will
3887 	 * be resolved when the inode is updated with MNT_WAIT.
3888 	 * This work is done in two passes. The first pass grabs most
3889 	 * of the buffers and begins asynchronously writing them. The
3890 	 * only way to wait for these asynchronous writes is to sleep
3891 	 * on the filesystem vnode which may stay busy for a long time
3892 	 * if the filesystem is active. So, instead, we make a second
3893 	 * pass over the dependencies blocking on each write. In the
3894 	 * usual case we will be blocking against a write that we
3895 	 * initiated, so when it is done the dependency will have been
3896 	 * resolved. Thus the second pass is expected to end quickly.
3897 	 */
3898 	waitfor = MNT_NOWAIT;
3899 top:
3900 	if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) {
3901 		FREE_LOCK(&lk);
3902 		return (0);
3903 	}
3904 	bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
3905 loop:
3906 	/*
3907 	 * As we hold the buffer locked, none of its dependencies
3908 	 * will disappear.
3909 	 */
3910 	for (wk = LIST_FIRST(&bp->b_dep); wk;
3911 	     wk = LIST_NEXT(wk, wk_list)) {
3912 		switch (wk->wk_type) {
3913 
3914 		case D_ALLOCDIRECT:
3915 			adp = WK_ALLOCDIRECT(wk);
3916 			if (adp->ad_state & DEPCOMPLETE)
3917 				break;
3918 			nbp = adp->ad_buf;
3919 			if (getdirtybuf(&nbp, waitfor) == 0)
3920 				break;
3921 			FREE_LOCK(&lk);
3922 			if (waitfor == MNT_NOWAIT) {
3923 				bawrite(nbp);
3924 			} else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
3925 				bawrite(bp);
3926 				return (error);
3927 			}
3928 			ACQUIRE_LOCK(&lk);
3929 			break;
3930 
3931 		case D_ALLOCINDIR:
3932 			aip = WK_ALLOCINDIR(wk);
3933 			if (aip->ai_state & DEPCOMPLETE)
3934 				break;
3935 			nbp = aip->ai_buf;
3936 			if (getdirtybuf(&nbp, waitfor) == 0)
3937 				break;
3938 			FREE_LOCK(&lk);
3939 			if (waitfor == MNT_NOWAIT) {
3940 				bawrite(nbp);
3941 			} else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
3942 				bawrite(bp);
3943 				return (error);
3944 			}
3945 			ACQUIRE_LOCK(&lk);
3946 			break;
3947 
3948 		case D_INDIRDEP:
3949 		restart:
3950 			for (aip = LIST_FIRST(&WK_INDIRDEP(wk)->ir_deplisthd);
3951 			     aip; aip = LIST_NEXT(aip, ai_next)) {
3952 				if (aip->ai_state & DEPCOMPLETE)
3953 					continue;
3954 				nbp = aip->ai_buf;
3955 				if (getdirtybuf(&nbp, MNT_WAIT) == 0)
3956 					goto restart;
3957 				FREE_LOCK(&lk);
3958 				if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
3959 					bawrite(bp);
3960 					return (error);
3961 				}
3962 				ACQUIRE_LOCK(&lk);
3963 				goto restart;
3964 			}
3965 			break;
3966 
3967 		case D_INODEDEP:
3968 			if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
3969 			    WK_INODEDEP(wk)->id_ino)) != 0) {
3970 				FREE_LOCK(&lk);
3971 				bawrite(bp);
3972 				return (error);
3973 			}
3974 			break;
3975 
3976 		case D_PAGEDEP:
3977 			/*
3978 			 * We are trying to sync a directory that may
3979 			 * have dependencies on both its own metadata
3980 			 * and/or dependencies on the inodes of any
3981 			 * recently allocated files. We walk its diradd
3982 			 * lists pushing out the associated inode.
3983 			 */
3984 			pagedep = WK_PAGEDEP(wk);
3985 			for (i = 0; i < DAHASHSZ; i++) {
3986 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
3987 					continue;
3988 				if ((error =
3989 				    flush_pagedep_deps(vp, pagedep->pd_mnt,
3990 						&pagedep->pd_diraddhd[i]))) {
3991 					FREE_LOCK(&lk);
3992 					bawrite(bp);
3993 					return (error);
3994 				}
3995 			}
3996 			break;
3997 
3998 		case D_MKDIR:
3999 			/*
4000 			 * This case should never happen if the vnode has
4001 			 * been properly sync'ed. However, if this function
4002 			 * is used at a place where the vnode has not yet
4003 			 * been sync'ed, this dependency can show up. So,
4004 			 * rather than panic, just flush it.
4005 			 */
4006 			nbp = WK_MKDIR(wk)->md_buf;
4007 			if (getdirtybuf(&nbp, waitfor) == 0)
4008 				break;
4009 			FREE_LOCK(&lk);
4010 			if (waitfor == MNT_NOWAIT) {
4011 				bawrite(nbp);
4012 			} else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4013 				bawrite(bp);
4014 				return (error);
4015 			}
4016 			ACQUIRE_LOCK(&lk);
4017 			break;
4018 
4019 		case D_BMSAFEMAP:
4020 			/*
4021 			 * This case should never happen if the vnode has
4022 			 * been properly sync'ed. However, if this function
4023 			 * is used at a place where the vnode has not yet
4024 			 * been sync'ed, this dependency can show up. So,
4025 			 * rather than panic, just flush it.
4026 			 */
4027 			nbp = WK_BMSAFEMAP(wk)->sm_buf;
4028 			if (getdirtybuf(&nbp, waitfor) == 0)
4029 				break;
4030 			FREE_LOCK(&lk);
4031 			if (waitfor == MNT_NOWAIT) {
4032 				bawrite(nbp);
4033 			} else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4034 				bawrite(bp);
4035 				return (error);
4036 			}
4037 			ACQUIRE_LOCK(&lk);
4038 			break;
4039 
4040 		default:
4041 			panic("softdep_sync_metadata: Unknown type %s",
4042 			    TYPENAME(wk->wk_type));
4043 			/* NOTREACHED */
4044 		}
4045 	}
4046 	(void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT);
4047 	nbp = TAILQ_NEXT(bp, b_vnbufs);
4048 	FREE_LOCK(&lk);
4049 	bawrite(bp);
4050 	ACQUIRE_LOCK(&lk);
4051 	if (nbp != NULL) {
4052 		bp = nbp;
4053 		goto loop;
4054 	}
4055 	/*
4056 	 * We must wait for any I/O in progress to finish so that
4057 	 * all potential buffers on the dirty list will be visible.
4058 	 * Once they are all there, proceed with the second pass
4059 	 * which will wait for the I/O as per above.
4060 	 */
4061 	drain_output(vp, 1);
4062 	/*
4063 	 * The brief unlock is to allow any pent up dependency
4064 	 * processing to be done.
4065 	 */
4066 	if (waitfor == MNT_NOWAIT) {
4067 		waitfor = MNT_WAIT;
4068 		FREE_LOCK(&lk);
4069 		ACQUIRE_LOCK(&lk);
4070 		goto top;
4071 	}
4072 
4073 	/*
4074 	 * If we have managed to get rid of all the dirty buffers,
4075 	 * then we are done. For certain directories and block
4076 	 * devices, we may need to do further work.
4077 	 */
4078 	if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) {
4079 		FREE_LOCK(&lk);
4080 		return (0);
4081 	}
4082 
4083 	FREE_LOCK(&lk);
4084 	/*
4085 	 * If we are trying to sync a block device, some of its buffers may
4086 	 * contain metadata that cannot be written until the contents of some
4087 	 * partially written files have been written to disk. The only easy
4088 	 * way to accomplish this is to sync the entire filesystem (luckily
4089 	 * this happens rarely).
4090 	 */
4091 	if (vn_isdisk(vp, NULL) &&
4092 	    vp->v_specmountpoint && !VOP_ISLOCKED(vp, NULL) &&
4093 	    (error = VFS_SYNC(vp->v_specmountpoint, MNT_WAIT, ap->a_cred,
4094 	     ap->a_p)) != 0)
4095 		return (error);
4096 	return (0);
4097 }
4098 
4099 /*
4100  * Flush the dependencies associated with an inodedep.
4101  * Called with splbio blocked.
4102  */
4103 static int
4104 flush_inodedep_deps(fs, ino)
4105 	struct fs *fs;
4106 	ino_t ino;
4107 {
4108 	struct inodedep *inodedep;
4109 	struct allocdirect *adp;
4110 	int error, waitfor;
4111 	struct buf *bp;
4112 
4113 	/*
4114 	 * This work is done in two passes. The first pass grabs most
4115 	 * of the buffers and begins asynchronously writing them. The
4116 	 * only way to wait for these asynchronous writes is to sleep
4117 	 * on the filesystem vnode which may stay busy for a long time
4118 	 * if the filesystem is active. So, instead, we make a second
4119 	 * pass over the dependencies blocking on each write. In the
4120 	 * usual case we will be blocking against a write that we
4121 	 * initiated, so when it is done the dependency will have been
4122 	 * resolved. Thus the second pass is expected to end quickly.
4123 	 * We give a brief window at the top of the loop to allow
4124 	 * any pending I/O to complete.
4125 	 */
4126 	for (waitfor = MNT_NOWAIT; ; ) {
4127 		FREE_LOCK(&lk);
4128 		ACQUIRE_LOCK(&lk);
4129 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4130 			return (0);
4131 		for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
4132 		     adp = TAILQ_NEXT(adp, ad_next)) {
4133 			if (adp->ad_state & DEPCOMPLETE)
4134 				continue;
4135 			bp = adp->ad_buf;
4136 			if (getdirtybuf(&bp, waitfor) == 0) {
4137 				if (waitfor == MNT_NOWAIT)
4138 					continue;
4139 				break;
4140 			}
4141 			FREE_LOCK(&lk);
4142 			if (waitfor == MNT_NOWAIT) {
4143 				bawrite(bp);
4144 			} else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) {
4145 				ACQUIRE_LOCK(&lk);
4146 				return (error);
4147 			}
4148 			ACQUIRE_LOCK(&lk);
4149 			break;
4150 		}
4151 		if (adp != NULL)
4152 			continue;
4153 		for (adp = TAILQ_FIRST(&inodedep->id_newinoupdt); adp;
4154 		     adp = TAILQ_NEXT(adp, ad_next)) {
4155 			if (adp->ad_state & DEPCOMPLETE)
4156 				continue;
4157 			bp = adp->ad_buf;
4158 			if (getdirtybuf(&bp, waitfor) == 0) {
4159 				if (waitfor == MNT_NOWAIT)
4160 					continue;
4161 				break;
4162 			}
4163 			FREE_LOCK(&lk);
4164 			if (waitfor == MNT_NOWAIT) {
4165 				bawrite(bp);
4166 			} else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) {
4167 				ACQUIRE_LOCK(&lk);
4168 				return (error);
4169 			}
4170 			ACQUIRE_LOCK(&lk);
4171 			break;
4172 		}
4173 		if (adp != NULL)
4174 			continue;
4175 		/*
4176 		 * If pass2, we are done, otherwise do pass 2.
4177 		 */
4178 		if (waitfor == MNT_WAIT)
4179 			break;
4180 		waitfor = MNT_WAIT;
4181 	}
4182 	/*
4183 	 * Try freeing inodedep in case all dependencies have been removed.
4184 	 */
4185 	if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
4186 		(void) free_inodedep(inodedep);
4187 	return (0);
4188 }
4189 
4190 /*
4191  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
4192  * Called with splbio blocked.
4193  */
4194 static int
4195 flush_pagedep_deps(pvp, mp, diraddhdp)
4196 	struct vnode *pvp;
4197 	struct mount *mp;
4198 	struct diraddhd *diraddhdp;
4199 {
4200 	struct proc *p = CURPROC;	/* XXX */
4201 	struct inodedep *inodedep;
4202 	struct ufsmount *ump;
4203 	struct diradd *dap;
4204 	struct vnode *vp;
4205 	int gotit, error = 0;
4206 	struct buf *bp;
4207 	ino_t inum;
4208 
4209 	ump = VFSTOUFS(mp);
4210 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
4211 		/*
4212 		 * Flush ourselves if this directory entry
4213 		 * has a MKDIR_PARENT dependency.
4214 		 */
4215 		if (dap->da_state & MKDIR_PARENT) {
4216 			FREE_LOCK(&lk);
4217 			if ((error = UFS_UPDATE(pvp, 1)) != 0)
4218 				break;
4219 			ACQUIRE_LOCK(&lk);
4220 			/*
4221 			 * If that cleared dependencies, go on to next.
4222 			 */
4223 			if (dap != LIST_FIRST(diraddhdp))
4224 				continue;
4225 			if (dap->da_state & MKDIR_PARENT)
4226 				panic("flush_pagedep_deps: MKDIR_PARENT");
4227 		}
4228 		/*
4229 		 * A newly allocated directory must have its "." and
4230 		 * ".." entries written out before its name can be
4231 		 * committed in its parent. We do not want or need
4232 		 * the full semantics of a synchronous VOP_FSYNC as
4233 		 * that may end up here again, once for each directory
4234 		 * level in the filesystem. Instead, we push the blocks
4235 		 * and wait for them to clear. We have to fsync twice
4236 		 * because the first call may choose to defer blocks
4237 		 * that still have dependencies, but deferral will
4238 		 * happen at most once.
4239 		 */
4240 		inum = dap->da_newinum;
4241 		if (dap->da_state & MKDIR_BODY) {
4242 			FREE_LOCK(&lk);
4243 			if ((error = VFS_VGET(mp, inum, &vp)) != 0)
4244 				break;
4245 			if ((error=VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)) ||
4246 			    (error=VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) {
4247 				vput(vp);
4248 				break;
4249 			}
4250 			drain_output(vp, 0);
4251 			vput(vp);
4252 			ACQUIRE_LOCK(&lk);
4253 			/*
4254 			 * If that cleared dependencies, go on to next.
4255 			 */
4256 			if (dap != LIST_FIRST(diraddhdp))
4257 				continue;
4258 			if (dap->da_state & MKDIR_BODY)
4259 				panic("flush_pagedep_deps: MKDIR_BODY");
4260 		}
4261 		/*
4262 		 * Flush the inode on which the directory entry depends.
4263 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
4264 		 * the only remaining dependency is that the updated inode
4265 		 * count must get pushed to disk. The inode has already
4266 		 * been pushed into its inode buffer (via VOP_UPDATE) at
4267 		 * the time of the reference count change. So we need only
4268 		 * locate that buffer, ensure that there will be no rollback
4269 		 * caused by a bitmap dependency, then write the inode buffer.
4270 		 */
4271 		if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0)
4272 			panic("flush_pagedep_deps: lost inode");
4273 		/*
4274 		 * If the inode still has bitmap dependencies,
4275 		 * push them to disk.
4276 		 */
4277 		if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4278 			gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4279 			FREE_LOCK(&lk);
4280 			if (gotit &&
4281 			    (error = VOP_BWRITE(inodedep->id_buf->b_vp,
4282 			     inodedep->id_buf)) != 0)
4283 				break;
4284 			ACQUIRE_LOCK(&lk);
4285 			if (dap != LIST_FIRST(diraddhdp))
4286 				continue;
4287 		}
4288 		/*
4289 		 * If the inode is still sitting in a buffer waiting
4290 		 * to be written, push it to disk.
4291 		 */
4292 		FREE_LOCK(&lk);
4293 		if ((error = bread(ump->um_devvp,
4294 		    fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
4295 		    (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0)
4296 			break;
4297 		if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0)
4298 			break;
4299 		ACQUIRE_LOCK(&lk);
4300 		/*
4301 		 * If we have failed to get rid of all the dependencies
4302 		 * then something is seriously wrong.
4303 		 */
4304 		if (dap == LIST_FIRST(diraddhdp))
4305 			panic("flush_pagedep_deps: flush failed");
4306 	}
4307 	if (error)
4308 		ACQUIRE_LOCK(&lk);
4309 	return (error);
4310 }
4311 
4312 /*
4313  * A large burst of file addition or deletion activity can drive the
4314  * memory load excessively high. Therefore we deliberately slow things
4315  * down and speed up the I/O processing if we find ourselves with too
4316  * many dependencies in progress.
4317  */
4318 static int
4319 request_cleanup(resource, islocked)
4320 	int resource;
4321 	int islocked;
4322 {
4323 	struct callout_handle handle;
4324 	struct proc *p = CURPROC;
4325 
4326 	/*
4327 	 * We never hold up the filesystem syncer process.
4328 	 */
4329 	if (p == filesys_syncer)
4330 		return (0);
4331 	/*
4332 	 * If we are resource constrained on inode dependencies, try
4333 	 * flushing some dirty inodes. Otherwise, we are constrained
4334 	 * by file deletions, so try accelerating flushes of directories
4335 	 * with removal dependencies. We would like to do the cleanup
4336 	 * here, but we probably hold an inode locked at this point and
4337 	 * that might deadlock against one that we try to clean. So,
4338 	 * the best that we can do is request the syncer daemon to do
4339 	 * the cleanup for us.
4340 	 */
4341 	switch (resource) {
4342 
4343 	case FLUSH_INODES:
4344 		stat_ino_limit_push += 1;
4345 		req_clear_inodedeps = 1;
4346 		break;
4347 
4348 	case FLUSH_REMOVE:
4349 		stat_blk_limit_push += 1;
4350 		req_clear_remove = 1;
4351 		break;
4352 
4353 	default:
4354 		panic("request_cleanup: unknown type");
4355 	}
4356 	/*
4357 	 * Hopefully the syncer daemon will catch up and awaken us.
4358 	 * We wait at most tickdelay before proceeding in any case.
4359 	 */
4360 	if (islocked == 0)
4361 		ACQUIRE_LOCK(&lk);
4362 	if (proc_waiting == 0) {
4363 		proc_waiting = 1;
4364 		handle = timeout(pause_timer, NULL,
4365 		    tickdelay > 2 ? tickdelay : 2);
4366 	}
4367 	FREE_LOCK_INTERLOCKED(&lk);
4368 	(void) tsleep((caddr_t)&proc_waiting, PPAUSE, "softupdate", 0);
4369 	ACQUIRE_LOCK_INTERLOCKED(&lk);
4370 	if (proc_waiting) {
4371 		untimeout(pause_timer, NULL, handle);
4372 		proc_waiting = 0;
4373 	} else {
4374 		switch (resource) {
4375 
4376 		case FLUSH_INODES:
4377 			stat_ino_limit_hit += 1;
4378 			break;
4379 
4380 		case FLUSH_REMOVE:
4381 			stat_blk_limit_hit += 1;
4382 			break;
4383 		}
4384 	}
4385 	if (islocked == 0)
4386 		FREE_LOCK(&lk);
4387 	return (1);
4388 }
4389 
4390 /*
4391  * Awaken processes pausing in request_cleanup and clear proc_waiting
4392  * to indicate that there is no longer a timer running.
4393  */
4394 void
4395 pause_timer(arg)
4396 	void *arg;
4397 {
4398 
4399 	proc_waiting = 0;
4400 	wakeup(&proc_waiting);
4401 }
4402 
4403 /*
4404  * Flush out a directory with at least one removal dependency in an effort to
4405  * reduce the number of dirrem, freefile, and freeblks dependency structures.
4406  */
4407 static void
4408 clear_remove(p)
4409 	struct proc *p;
4410 {
4411 	struct pagedep_hashhead *pagedephd;
4412 	struct pagedep *pagedep;
4413 	static int next = 0;
4414 	struct mount *mp;
4415 	struct vnode *vp;
4416 	int error, cnt;
4417 	ino_t ino;
4418 
4419 	ACQUIRE_LOCK(&lk);
4420 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
4421 		pagedephd = &pagedep_hashtbl[next++];
4422 		if (next >= pagedep_hash)
4423 			next = 0;
4424 		for (pagedep = LIST_FIRST(pagedephd); pagedep;
4425 		     pagedep = LIST_NEXT(pagedep, pd_hash)) {
4426 			if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
4427 				continue;
4428 			mp = pagedep->pd_mnt;
4429 			ino = pagedep->pd_ino;
4430 			FREE_LOCK(&lk);
4431 			if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4432 				softdep_error("clear_remove: vget", error);
4433 				return;
4434 			}
4435 			if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)))
4436 				softdep_error("clear_remove: fsync", error);
4437 			drain_output(vp, 0);
4438 			vput(vp);
4439 			return;
4440 		}
4441 	}
4442 	FREE_LOCK(&lk);
4443 }
4444 
4445 /*
4446  * Clear out a block of dirty inodes in an effort to reduce
4447  * the number of inodedep dependency structures.
4448  */
4449 static void
4450 clear_inodedeps(p)
4451 	struct proc *p;
4452 {
4453 	struct inodedep_hashhead *inodedephd;
4454 	struct inodedep *inodedep;
4455 	static int next = 0;
4456 	struct mount *mp;
4457 	struct vnode *vp;
4458 	struct fs *fs;
4459 	int error, cnt;
4460 	ino_t firstino, lastino, ino;
4461 
4462 	ACQUIRE_LOCK(&lk);
4463 	/*
4464 	 * Pick a random inode dependency to be cleared.
4465 	 * We will then gather up all the inodes in its block
4466 	 * that have dependencies and flush them out.
4467 	 */
4468 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
4469 		inodedephd = &inodedep_hashtbl[next++];
4470 		if (next >= inodedep_hash)
4471 			next = 0;
4472 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
4473 			break;
4474 	}
4475 	/*
4476 	 * Ugly code to find mount point given pointer to superblock.
4477 	 */
4478 	fs = inodedep->id_fs;
4479 	TAILQ_FOREACH(mp, &mountlist, mnt_list)
4480 		if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs)
4481 			break;
4482 	/*
4483 	 * Find the last inode in the block with dependencies.
4484 	 */
4485 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
4486 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
4487 		if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
4488 			break;
4489 	/*
4490 	 * Asynchronously push all but the last inode with dependencies.
4491 	 * Synchronously push the last inode with dependencies to ensure
4492 	 * that the inode block gets written to free up the inodedeps.
4493 	 */
4494 	for (ino = firstino; ino <= lastino; ino++) {
4495 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4496 			continue;
4497 		FREE_LOCK(&lk);
4498 		if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4499 			softdep_error("clear_inodedeps: vget", error);
4500 			return;
4501 		}
4502 		if (ino == lastino) {
4503 			if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_WAIT, p)))
4504 				softdep_error("clear_inodedeps: fsync1", error);
4505 		} else {
4506 			if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)))
4507 				softdep_error("clear_inodedeps: fsync2", error);
4508 			drain_output(vp, 0);
4509 		}
4510 		vput(vp);
4511 		ACQUIRE_LOCK(&lk);
4512 	}
4513 	FREE_LOCK(&lk);
4514 }
4515 
4516 /*
4517  * Function to determine if the buffer has outstanding dependencies
4518  * that will cause a roll-back if the buffer is written. If wantcount
4519  * is set, return number of dependencies, otherwise just yes or no.
4520  */
4521 static int
4522 softdep_count_dependencies(bp, wantcount)
4523 	struct buf *bp;
4524 	int wantcount;
4525 {
4526 	struct worklist *wk;
4527 	struct inodedep *inodedep;
4528 	struct indirdep *indirdep;
4529 	struct allocindir *aip;
4530 	struct pagedep *pagedep;
4531 	struct diradd *dap;
4532 	int i, retval;
4533 
4534 	retval = 0;
4535 	ACQUIRE_LOCK(&lk);
4536 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = LIST_NEXT(wk, wk_list)) {
4537 		switch (wk->wk_type) {
4538 
4539 		case D_INODEDEP:
4540 			inodedep = WK_INODEDEP(wk);
4541 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4542 				/* bitmap allocation dependency */
4543 				retval += 1;
4544 				if (!wantcount)
4545 					goto out;
4546 			}
4547 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
4548 				/* direct block pointer dependency */
4549 				retval += 1;
4550 				if (!wantcount)
4551 					goto out;
4552 			}
4553 			continue;
4554 
4555 		case D_INDIRDEP:
4556 			indirdep = WK_INDIRDEP(wk);
4557 			for (aip = LIST_FIRST(&indirdep->ir_deplisthd);
4558 			     aip; aip = LIST_NEXT(aip, ai_next)) {
4559 				/* indirect block pointer dependency */
4560 				retval += 1;
4561 				if (!wantcount)
4562 					goto out;
4563 			}
4564 			continue;
4565 
4566 		case D_PAGEDEP:
4567 			pagedep = WK_PAGEDEP(wk);
4568 			for (i = 0; i < DAHASHSZ; i++) {
4569 				for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]);
4570 				     dap; dap = LIST_NEXT(dap, da_pdlist)) {
4571 					/* directory entry dependency */
4572 					retval += 1;
4573 					if (!wantcount)
4574 						goto out;
4575 				}
4576 			}
4577 			continue;
4578 
4579 		case D_BMSAFEMAP:
4580 		case D_ALLOCDIRECT:
4581 		case D_ALLOCINDIR:
4582 		case D_MKDIR:
4583 			/* never a dependency on these blocks */
4584 			continue;
4585 
4586 		default:
4587 			panic("softdep_check_for_rollback: Unexpected type %s",
4588 			    TYPENAME(wk->wk_type));
4589 			/* NOTREACHED */
4590 		}
4591 	}
4592 out:
4593 	FREE_LOCK(&lk);
4594 	return retval;
4595 }
4596 
4597 /*
4598  * Acquire exclusive access to a buffer.
4599  * Must be called with splbio blocked.
4600  * Return 1 if buffer was acquired.
4601  */
4602 static int
4603 getdirtybuf(bpp, waitfor)
4604 	struct buf **bpp;
4605 	int waitfor;
4606 {
4607 	struct buf *bp;
4608 
4609 	for (;;) {
4610 		if ((bp = *bpp) == NULL)
4611 			return (0);
4612 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
4613 			if ((bp->b_xflags & BX_BKGRDINPROG) == 0)
4614 				break;
4615 			BUF_UNLOCK(bp);
4616 			if (waitfor != MNT_WAIT)
4617 				return (0);
4618 			bp->b_xflags |= BX_BKGRDWAIT;
4619 			FREE_LOCK_INTERLOCKED(&lk);
4620 			tsleep(&bp->b_xflags, PRIBIO, "getbuf", 0);
4621 			ACQUIRE_LOCK_INTERLOCKED(&lk);
4622 			continue;
4623 		}
4624 		if (waitfor != MNT_WAIT)
4625 			return (0);
4626 		FREE_LOCK_INTERLOCKED(&lk);
4627 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL) != ENOLCK)
4628 			panic("getdirtybuf: inconsistent lock");
4629 		ACQUIRE_LOCK_INTERLOCKED(&lk);
4630 	}
4631 	if ((bp->b_flags & B_DELWRI) == 0) {
4632 		BUF_UNLOCK(bp);
4633 		return (0);
4634 	}
4635 	bremfree(bp);
4636 	return (1);
4637 }
4638 
4639 /*
4640  * Wait for pending output on a vnode to complete.
4641  * Must be called with vnode locked.
4642  */
4643 static void
4644 drain_output(vp, islocked)
4645 	struct vnode *vp;
4646 	int islocked;
4647 {
4648 
4649 	if (!islocked)
4650 		ACQUIRE_LOCK(&lk);
4651 	while (vp->v_numoutput) {
4652 		vp->v_flag |= VBWAIT;
4653 		FREE_LOCK_INTERLOCKED(&lk);
4654 		tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, "drainvp", 0);
4655 		ACQUIRE_LOCK_INTERLOCKED(&lk);
4656 	}
4657 	if (!islocked)
4658 		FREE_LOCK(&lk);
4659 }
4660 
4661 /*
4662  * Called whenever a buffer that is being invalidated or reallocated
4663  * contains dependencies. This should only happen if an I/O error has
4664  * occurred. The routine is called with the buffer locked.
4665  */
4666 static void
4667 softdep_deallocate_dependencies(bp)
4668 	struct buf *bp;
4669 {
4670 
4671 	if ((bp->b_flags & B_ERROR) == 0)
4672 		panic("softdep_deallocate_dependencies: dangling deps");
4673 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
4674 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
4675 }
4676 
4677 /*
4678  * Function to handle asynchronous write errors in the filesystem.
4679  */
4680 void
4681 softdep_error(func, error)
4682 	char *func;
4683 	int error;
4684 {
4685 
4686 	/* XXX should do something better! */
4687 	printf("%s: got error %d while accessing filesystem\n", func, error);
4688 }
4689