xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision b601c69bdbe8755d26570261d7fd4c02ee4eff74)
1 /*
2  * Copyright 1998, 2000 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  * Further information about soft updates can be obtained from:
10  *
11  *	Marshall Kirk McKusick		http://www.mckusick.com/softdep/
12  *	1614 Oxford Street		mckusick@mckusick.com
13  *	Berkeley, CA 94709-1608		+1-510-843-9542
14  *	USA
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  *
20  * 1. Redistributions of source code must retain the above copyright
21  *    notice, this list of conditions and the following disclaimer.
22  * 2. Redistributions in binary form must reproduce the above copyright
23  *    notice, this list of conditions and the following disclaimer in the
24  *    documentation and/or other materials provided with the distribution.
25  *
26  * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
27  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
28  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
29  * DISCLAIMED.  IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
30  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)ffs_softdep.c	9.59 (McKusick) 6/21/00
39  * $FreeBSD$
40  */
41 
42 /*
43  * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
44  */
45 #ifndef DIAGNOSTIC
46 #define DIAGNOSTIC
47 #endif
48 #ifndef DEBUG
49 #define DEBUG
50 #endif
51 
52 #include <sys/param.h>
53 #include <sys/kernel.h>
54 #include <sys/systm.h>
55 #include <sys/bio.h>
56 #include <sys/buf.h>
57 #include <sys/malloc.h>
58 #include <sys/mount.h>
59 #include <sys/proc.h>
60 #include <sys/syslog.h>
61 #include <sys/vnode.h>
62 #include <sys/conf.h>
63 #include <ufs/ufs/dir.h>
64 #include <ufs/ufs/extattr.h>
65 #include <ufs/ufs/quota.h>
66 #include <ufs/ufs/inode.h>
67 #include <ufs/ufs/ufsmount.h>
68 #include <ufs/ffs/fs.h>
69 #include <ufs/ffs/softdep.h>
70 #include <ufs/ffs/ffs_extern.h>
71 #include <ufs/ufs/ufs_extern.h>
72 
73 /*
74  * These definitions need to be adapted to the system to which
75  * this file is being ported.
76  */
77 /*
78  * malloc types defined for the softdep system.
79  */
80 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
81 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
82 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
83 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
84 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
85 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
86 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
87 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
88 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
89 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
90 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
91 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
92 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
93 
94 #define	D_PAGEDEP	0
95 #define	D_INODEDEP	1
96 #define	D_NEWBLK	2
97 #define	D_BMSAFEMAP	3
98 #define	D_ALLOCDIRECT	4
99 #define	D_INDIRDEP	5
100 #define	D_ALLOCINDIR	6
101 #define	D_FREEFRAG	7
102 #define	D_FREEBLKS	8
103 #define	D_FREEFILE	9
104 #define	D_DIRADD	10
105 #define	D_MKDIR		11
106 #define	D_DIRREM	12
107 #define D_LAST		D_DIRREM
108 
109 /*
110  * translate from workitem type to memory type
111  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
112  */
113 static struct malloc_type *memtype[] = {
114 	M_PAGEDEP,
115 	M_INODEDEP,
116 	M_NEWBLK,
117 	M_BMSAFEMAP,
118 	M_ALLOCDIRECT,
119 	M_INDIRDEP,
120 	M_ALLOCINDIR,
121 	M_FREEFRAG,
122 	M_FREEBLKS,
123 	M_FREEFILE,
124 	M_DIRADD,
125 	M_MKDIR,
126 	M_DIRREM
127 };
128 
129 #define DtoM(type) (memtype[type])
130 
131 /*
132  * Names of malloc types.
133  */
134 #define TYPENAME(type)  \
135 	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
136 #define CURPROC curproc
137 /*
138  * End system adaptaion definitions.
139  */
140 
141 /*
142  * Internal function prototypes.
143  */
144 static	void softdep_error __P((char *, int));
145 static	void drain_output __P((struct vnode *, int));
146 static	int getdirtybuf __P((struct buf **, int));
147 static	void clear_remove __P((struct proc *));
148 static	void clear_inodedeps __P((struct proc *));
149 static	int flush_pagedep_deps __P((struct vnode *, struct mount *,
150 	    struct diraddhd *));
151 static	int flush_inodedep_deps __P((struct fs *, ino_t));
152 static	int handle_written_filepage __P((struct pagedep *, struct buf *));
153 static  void diradd_inode_written __P((struct diradd *, struct inodedep *));
154 static	int handle_written_inodeblock __P((struct inodedep *, struct buf *));
155 static	void handle_allocdirect_partdone __P((struct allocdirect *));
156 static	void handle_allocindir_partdone __P((struct allocindir *));
157 static	void initiate_write_filepage __P((struct pagedep *, struct buf *));
158 static	void handle_written_mkdir __P((struct mkdir *, int));
159 static	void initiate_write_inodeblock __P((struct inodedep *, struct buf *));
160 static	void handle_workitem_freefile __P((struct freefile *));
161 static	void handle_workitem_remove __P((struct dirrem *));
162 static	struct dirrem *newdirrem __P((struct buf *, struct inode *,
163 	    struct inode *, int, struct dirrem **));
164 static	void free_diradd __P((struct diradd *));
165 static	void free_allocindir __P((struct allocindir *, struct inodedep *));
166 static	int indir_trunc __P((struct inode *, ufs_daddr_t, int, ufs_lbn_t,
167 	    long *));
168 static	void deallocate_dependencies __P((struct buf *, struct inodedep *));
169 static	void free_allocdirect __P((struct allocdirectlst *,
170 	    struct allocdirect *, int));
171 static	int check_inode_unwritten __P((struct inodedep *));
172 static	int free_inodedep __P((struct inodedep *));
173 static	void handle_workitem_freeblocks __P((struct freeblks *));
174 static	void merge_inode_lists __P((struct inodedep *));
175 static	void setup_allocindir_phase2 __P((struct buf *, struct inode *,
176 	    struct allocindir *));
177 static	struct allocindir *newallocindir __P((struct inode *, int, ufs_daddr_t,
178 	    ufs_daddr_t));
179 static	void handle_workitem_freefrag __P((struct freefrag *));
180 static	struct freefrag *newfreefrag __P((struct inode *, ufs_daddr_t, long));
181 static	void allocdirect_merge __P((struct allocdirectlst *,
182 	    struct allocdirect *, struct allocdirect *));
183 static	struct bmsafemap *bmsafemap_lookup __P((struct buf *));
184 static	int newblk_lookup __P((struct fs *, ufs_daddr_t, int,
185 	    struct newblk **));
186 static	int inodedep_lookup __P((struct fs *, ino_t, int, struct inodedep **));
187 static	int pagedep_lookup __P((struct inode *, ufs_lbn_t, int,
188 	    struct pagedep **));
189 static	void pause_timer __P((void *));
190 static	int request_cleanup __P((int, int));
191 static	void add_to_worklist __P((struct worklist *));
192 
193 /*
194  * Exported softdep operations.
195  */
196 static	void softdep_disk_io_initiation __P((struct buf *));
197 static	void softdep_disk_write_complete __P((struct buf *));
198 static	void softdep_deallocate_dependencies __P((struct buf *));
199 static	void softdep_move_dependencies __P((struct buf *, struct buf *));
200 static	int softdep_count_dependencies __P((struct buf *bp, int));
201 
202 struct bio_ops bioops = {
203 	softdep_disk_io_initiation,		/* io_start */
204 	softdep_disk_write_complete,		/* io_complete */
205 	softdep_deallocate_dependencies,	/* io_deallocate */
206 	softdep_move_dependencies,		/* io_movedeps */
207 	softdep_count_dependencies,		/* io_countdeps */
208 };
209 
210 /*
211  * Locking primitives.
212  *
213  * For a uniprocessor, all we need to do is protect against disk
214  * interrupts. For a multiprocessor, this lock would have to be
215  * a mutex. A single mutex is used throughout this file, though
216  * finer grain locking could be used if contention warranted it.
217  *
218  * For a multiprocessor, the sleep call would accept a lock and
219  * release it after the sleep processing was complete. In a uniprocessor
220  * implementation there is no such interlock, so we simple mark
221  * the places where it needs to be done with the `interlocked' form
222  * of the lock calls. Since the uniprocessor sleep already interlocks
223  * the spl, there is nothing that really needs to be done.
224  */
225 #ifndef /* NOT */ DEBUG
226 static struct lockit {
227 	int	lkt_spl;
228 } lk = { 0 };
229 #define ACQUIRE_LOCK(lk)		(lk)->lkt_spl = splbio()
230 #define FREE_LOCK(lk)			splx((lk)->lkt_spl)
231 #define ACQUIRE_LOCK_INTERLOCKED(lk)
232 #define FREE_LOCK_INTERLOCKED(lk)
233 
234 #else /* DEBUG */
235 static struct lockit {
236 	int	lkt_spl;
237 	pid_t	lkt_held;
238 } lk = { 0, -1 };
239 static int lockcnt;
240 
241 static	void acquire_lock __P((struct lockit *));
242 static	void free_lock __P((struct lockit *));
243 static	void acquire_lock_interlocked __P((struct lockit *));
244 static	void free_lock_interlocked __P((struct lockit *));
245 
246 #define ACQUIRE_LOCK(lk)		acquire_lock(lk)
247 #define FREE_LOCK(lk)			free_lock(lk)
248 #define ACQUIRE_LOCK_INTERLOCKED(lk)	acquire_lock_interlocked(lk)
249 #define FREE_LOCK_INTERLOCKED(lk)	free_lock_interlocked(lk)
250 
251 static void
252 acquire_lock(lk)
253 	struct lockit *lk;
254 {
255 
256 	if (lk->lkt_held != -1) {
257 		if (lk->lkt_held == CURPROC->p_pid)
258 			panic("softdep_lock: locking against myself");
259 		else
260 			panic("softdep_lock: lock held by %d", lk->lkt_held);
261 	}
262 	lk->lkt_spl = splbio();
263 	lk->lkt_held = CURPROC->p_pid;
264 	lockcnt++;
265 }
266 
267 static void
268 free_lock(lk)
269 	struct lockit *lk;
270 {
271 
272 	if (lk->lkt_held == -1)
273 		panic("softdep_unlock: lock not held");
274 	lk->lkt_held = -1;
275 	splx(lk->lkt_spl);
276 }
277 
278 static void
279 acquire_lock_interlocked(lk)
280 	struct lockit *lk;
281 {
282 
283 	if (lk->lkt_held != -1) {
284 		if (lk->lkt_held == CURPROC->p_pid)
285 			panic("softdep_lock_interlocked: locking against self");
286 		else
287 			panic("softdep_lock_interlocked: lock held by %d",
288 			    lk->lkt_held);
289 	}
290 	lk->lkt_held = CURPROC->p_pid;
291 	lockcnt++;
292 }
293 
294 static void
295 free_lock_interlocked(lk)
296 	struct lockit *lk;
297 {
298 
299 	if (lk->lkt_held == -1)
300 		panic("softdep_unlock_interlocked: lock not held");
301 	lk->lkt_held = -1;
302 }
303 #endif /* DEBUG */
304 
305 /*
306  * Place holder for real semaphores.
307  */
308 struct sema {
309 	int	value;
310 	pid_t	holder;
311 	char	*name;
312 	int	prio;
313 	int	timo;
314 };
315 static	void sema_init __P((struct sema *, char *, int, int));
316 static	int sema_get __P((struct sema *, struct lockit *));
317 static	void sema_release __P((struct sema *));
318 
319 static void
320 sema_init(semap, name, prio, timo)
321 	struct sema *semap;
322 	char *name;
323 	int prio, timo;
324 {
325 
326 	semap->holder = -1;
327 	semap->value = 0;
328 	semap->name = name;
329 	semap->prio = prio;
330 	semap->timo = timo;
331 }
332 
333 static int
334 sema_get(semap, interlock)
335 	struct sema *semap;
336 	struct lockit *interlock;
337 {
338 
339 	if (semap->value++ > 0) {
340 		if (interlock != NULL)
341 			FREE_LOCK_INTERLOCKED(interlock);
342 		tsleep((caddr_t)semap, semap->prio, semap->name, semap->timo);
343 		if (interlock != NULL) {
344 			ACQUIRE_LOCK_INTERLOCKED(interlock);
345 			FREE_LOCK(interlock);
346 		}
347 		return (0);
348 	}
349 	semap->holder = CURPROC->p_pid;
350 	if (interlock != NULL)
351 		FREE_LOCK(interlock);
352 	return (1);
353 }
354 
355 static void
356 sema_release(semap)
357 	struct sema *semap;
358 {
359 
360 	if (semap->value <= 0 || semap->holder != CURPROC->p_pid)
361 		panic("sema_release: not held");
362 	if (--semap->value > 0) {
363 		semap->value = 0;
364 		wakeup(semap);
365 	}
366 	semap->holder = -1;
367 }
368 
369 /*
370  * Worklist queue management.
371  * These routines require that the lock be held.
372  */
373 #ifndef /* NOT */ DEBUG
374 #define WORKLIST_INSERT(head, item) do {	\
375 	(item)->wk_state |= ONWORKLIST;		\
376 	LIST_INSERT_HEAD(head, item, wk_list);	\
377 } while (0)
378 #define WORKLIST_REMOVE(item) do {		\
379 	(item)->wk_state &= ~ONWORKLIST;	\
380 	LIST_REMOVE(item, wk_list);		\
381 } while (0)
382 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
383 
384 #else /* DEBUG */
385 static	void worklist_insert __P((struct workhead *, struct worklist *));
386 static	void worklist_remove __P((struct worklist *));
387 static	void workitem_free __P((struct worklist *, int));
388 
389 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
390 #define WORKLIST_REMOVE(item) worklist_remove(item)
391 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
392 
393 static void
394 worklist_insert(head, item)
395 	struct workhead *head;
396 	struct worklist *item;
397 {
398 
399 	if (lk.lkt_held == -1)
400 		panic("worklist_insert: lock not held");
401 	if (item->wk_state & ONWORKLIST)
402 		panic("worklist_insert: already on list");
403 	item->wk_state |= ONWORKLIST;
404 	LIST_INSERT_HEAD(head, item, wk_list);
405 }
406 
407 static void
408 worklist_remove(item)
409 	struct worklist *item;
410 {
411 
412 	if (lk.lkt_held == -1)
413 		panic("worklist_remove: lock not held");
414 	if ((item->wk_state & ONWORKLIST) == 0)
415 		panic("worklist_remove: not on list");
416 	item->wk_state &= ~ONWORKLIST;
417 	LIST_REMOVE(item, wk_list);
418 }
419 
420 static void
421 workitem_free(item, type)
422 	struct worklist *item;
423 	int type;
424 {
425 
426 	if (item->wk_state & ONWORKLIST)
427 		panic("workitem_free: still on list");
428 	if (item->wk_type != type)
429 		panic("workitem_free: type mismatch");
430 	FREE(item, DtoM(type));
431 }
432 #endif /* DEBUG */
433 
434 /*
435  * Workitem queue management
436  */
437 static struct workhead softdep_workitem_pending;
438 static int softdep_worklist_busy;
439 static int max_softdeps;	/* maximum number of structs before slowdown */
440 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
441 static int proc_waiting;	/* tracks whether we have a timeout posted */
442 static struct proc *filesys_syncer; /* proc of filesystem syncer process */
443 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
444 #define FLUSH_INODES	1
445 static int req_clear_remove;	/* syncer process flush some freeblks */
446 #define FLUSH_REMOVE	2
447 /*
448  * runtime statistics
449  */
450 static int stat_blk_limit_push;	/* number of times block limit neared */
451 static int stat_ino_limit_push;	/* number of times inode limit neared */
452 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
453 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
454 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
455 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
456 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
457 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
458 #ifdef DEBUG
459 #include <vm/vm.h>
460 #include <sys/sysctl.h>
461 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
462 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
463 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
464 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
465 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
466 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
467 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
468 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
469 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
470 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
471 #endif /* DEBUG */
472 
473 /*
474  * Add an item to the end of the work queue.
475  * This routine requires that the lock be held.
476  * This is the only routine that adds items to the list.
477  * The following routine is the only one that removes items
478  * and does so in order from first to last.
479  */
480 static void
481 add_to_worklist(wk)
482 	struct worklist *wk;
483 {
484 	static struct worklist *worklist_tail;
485 
486 	if (wk->wk_state & ONWORKLIST)
487 		panic("add_to_worklist: already on list");
488 	wk->wk_state |= ONWORKLIST;
489 	if (LIST_FIRST(&softdep_workitem_pending) == NULL)
490 		LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
491 	else
492 		LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
493 	worklist_tail = wk;
494 }
495 
496 /*
497  * Process that runs once per second to handle items in the background queue.
498  *
499  * Note that we ensure that everything is done in the order in which they
500  * appear in the queue. The code below depends on this property to ensure
501  * that blocks of a file are freed before the inode itself is freed. This
502  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
503  * until all the old ones have been purged from the dependency lists.
504  */
505 int
506 softdep_process_worklist(matchmnt)
507 	struct mount *matchmnt;
508 {
509 	struct proc *p = CURPROC;
510 	struct worklist *wk;
511 	struct mount *mp;
512 	int matchcnt, loopcount;
513 
514 	/*
515 	 * Record the process identifier of our caller so that we can give
516 	 * this process preferential treatment in request_cleanup below.
517 	 */
518 	filesys_syncer = p;
519 	matchcnt = 0;
520 	/*
521 	 * There is no danger of having multiple processes run this
522 	 * code. It is single threaded solely so that softdep_flushfiles
523 	 * (below) can get an accurate count of the number of items
524 	 * related to its mount point that are in the list.
525 	 */
526 	if (softdep_worklist_busy && matchmnt == NULL)
527 		return (-1);
528 	/*
529 	 * If requested, try removing inode or removal dependencies.
530 	 */
531 	if (req_clear_inodedeps) {
532 		clear_inodedeps(p);
533 		req_clear_inodedeps = 0;
534 		wakeup(&proc_waiting);
535 	}
536 	if (req_clear_remove) {
537 		clear_remove(p);
538 		req_clear_remove = 0;
539 		wakeup(&proc_waiting);
540 	}
541 	ACQUIRE_LOCK(&lk);
542 	loopcount = 1;
543 	while ((wk = LIST_FIRST(&softdep_workitem_pending)) != 0) {
544 		WORKLIST_REMOVE(wk);
545 		FREE_LOCK(&lk);
546 		switch (wk->wk_type) {
547 
548 		case D_DIRREM:
549 			/* removal of a directory entry */
550 			mp = WK_DIRREM(wk)->dm_mnt;
551 			if (mp == matchmnt)
552 				matchcnt += 1;
553 			vn_start_write(NULL, &mp, V_WAIT);
554 			handle_workitem_remove(WK_DIRREM(wk));
555 			vn_finished_write(mp);
556 			break;
557 
558 		case D_FREEBLKS:
559 			/* releasing blocks and/or fragments from a file */
560 			mp = WK_FREEBLKS(wk)->fb_mnt;
561 			if (mp == matchmnt)
562 				matchcnt += 1;
563 			vn_start_write(NULL, &mp, V_WAIT);
564 			handle_workitem_freeblocks(WK_FREEBLKS(wk));
565 			vn_finished_write(mp);
566 			break;
567 
568 		case D_FREEFRAG:
569 			/* releasing a fragment when replaced as a file grows */
570 			mp = WK_FREEFRAG(wk)->ff_mnt;
571 			if (mp == matchmnt)
572 				matchcnt += 1;
573 			vn_start_write(NULL, &mp, V_WAIT);
574 			handle_workitem_freefrag(WK_FREEFRAG(wk));
575 			vn_finished_write(mp);
576 			break;
577 
578 		case D_FREEFILE:
579 			/* releasing an inode when its link count drops to 0 */
580 			mp = WK_FREEFILE(wk)->fx_mnt;
581 			if (mp == matchmnt)
582 				matchcnt += 1;
583 			vn_start_write(NULL, &mp, V_WAIT);
584 			handle_workitem_freefile(WK_FREEFILE(wk));
585 			vn_finished_write(mp);
586 			break;
587 
588 		default:
589 			panic("%s_process_worklist: Unknown type %s",
590 			    "softdep", TYPENAME(wk->wk_type));
591 			/* NOTREACHED */
592 		}
593 		if (softdep_worklist_busy && matchmnt == NULL)
594 			return (-1);
595 		/*
596 		 * If requested, try removing inode or removal dependencies.
597 		 */
598 		if (req_clear_inodedeps) {
599 			clear_inodedeps(p);
600 			req_clear_inodedeps = 0;
601 			wakeup(&proc_waiting);
602 		}
603 		if (req_clear_remove) {
604 			clear_remove(p);
605 			req_clear_remove = 0;
606 			wakeup(&proc_waiting);
607 		}
608 		/*
609 		 * We do not generally want to stop for buffer space, but if
610 		 * we are really being a buffer hog, we will stop and wait.
611 		 */
612 		if (loopcount++ % 128 == 0)
613 			bwillwrite();
614 		ACQUIRE_LOCK(&lk);
615 	}
616 	FREE_LOCK(&lk);
617 	return (matchcnt);
618 }
619 
620 /*
621  * Move dependencies from one buffer to another.
622  */
623 static void
624 softdep_move_dependencies(oldbp, newbp)
625 	struct buf *oldbp;
626 	struct buf *newbp;
627 {
628 	struct worklist *wk, *wktail;
629 
630 	if (LIST_FIRST(&newbp->b_dep) != NULL)
631 		panic("softdep_move_dependencies: need merge code");
632 	wktail = 0;
633 	ACQUIRE_LOCK(&lk);
634 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
635 		LIST_REMOVE(wk, wk_list);
636 		if (wktail == 0)
637 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
638 		else
639 			LIST_INSERT_AFTER(wktail, wk, wk_list);
640 		wktail = wk;
641 	}
642 	FREE_LOCK(&lk);
643 }
644 
645 /*
646  * Purge the work list of all items associated with a particular mount point.
647  */
648 int
649 softdep_flushfiles(oldmnt, flags, p)
650 	struct mount *oldmnt;
651 	int flags;
652 	struct proc *p;
653 {
654 	struct vnode *devvp;
655 	int error, loopcnt;
656 
657 	/*
658 	 * Await our turn to clear out the queue.
659 	 */
660 	while (softdep_worklist_busy)
661 		tsleep(&lbolt, PRIBIO, "softflush", 0);
662 	softdep_worklist_busy = 1;
663 	if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0) {
664 		softdep_worklist_busy = 0;
665 		return (error);
666 	}
667 	/*
668 	 * Alternately flush the block device associated with the mount
669 	 * point and process any dependencies that the flushing
670 	 * creates. In theory, this loop can happen at most twice,
671 	 * but we give it a few extra just to be sure.
672 	 */
673 	devvp = VFSTOUFS(oldmnt)->um_devvp;
674 	for (loopcnt = 10; loopcnt > 0; ) {
675 		if (softdep_process_worklist(oldmnt) == 0) {
676 			loopcnt--;
677 			/*
678 			 * Do another flush in case any vnodes were brought in
679 			 * as part of the cleanup operations.
680 			 */
681 			if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0)
682 				break;
683 			/*
684 			 * If we still found nothing to do, we are really done.
685 			 */
686 			if (softdep_process_worklist(oldmnt) == 0)
687 				break;
688 		}
689 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, p);
690 		error = VOP_FSYNC(devvp, p->p_ucred, MNT_WAIT, p);
691 		VOP_UNLOCK(devvp, 0, p);
692 		if (error)
693 			break;
694 	}
695 	softdep_worklist_busy = 0;
696 	/*
697 	 * If we are unmounting then it is an error to fail. If we
698 	 * are simply trying to downgrade to read-only, then filesystem
699 	 * activity can keep us busy forever, so we just fail with EBUSY.
700 	 */
701 	if (loopcnt == 0) {
702 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
703 			panic("softdep_flushfiles: looping");
704 		error = EBUSY;
705 	}
706 	return (error);
707 }
708 
709 /*
710  * Structure hashing.
711  *
712  * There are three types of structures that can be looked up:
713  *	1) pagedep structures identified by mount point, inode number,
714  *	   and logical block.
715  *	2) inodedep structures identified by mount point and inode number.
716  *	3) newblk structures identified by mount point and
717  *	   physical block number.
718  *
719  * The "pagedep" and "inodedep" dependency structures are hashed
720  * separately from the file blocks and inodes to which they correspond.
721  * This separation helps when the in-memory copy of an inode or
722  * file block must be replaced. It also obviates the need to access
723  * an inode or file page when simply updating (or de-allocating)
724  * dependency structures. Lookup of newblk structures is needed to
725  * find newly allocated blocks when trying to associate them with
726  * their allocdirect or allocindir structure.
727  *
728  * The lookup routines optionally create and hash a new instance when
729  * an existing entry is not found.
730  */
731 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
732 
733 /*
734  * Structures and routines associated with pagedep caching.
735  */
736 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
737 u_long	pagedep_hash;		/* size of hash table - 1 */
738 #define	PAGEDEP_HASH(mp, inum, lbn) \
739 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
740 	    pagedep_hash])
741 static struct sema pagedep_in_progress;
742 
743 /*
744  * Look up a pagedep. Return 1 if found, 0 if not found.
745  * If not found, allocate if DEPALLOC flag is passed.
746  * Found or allocated entry is returned in pagedeppp.
747  * This routine must be called with splbio interrupts blocked.
748  */
749 static int
750 pagedep_lookup(ip, lbn, flags, pagedeppp)
751 	struct inode *ip;
752 	ufs_lbn_t lbn;
753 	int flags;
754 	struct pagedep **pagedeppp;
755 {
756 	struct pagedep *pagedep;
757 	struct pagedep_hashhead *pagedephd;
758 	struct mount *mp;
759 	int i;
760 
761 #ifdef DEBUG
762 	if (lk.lkt_held == -1)
763 		panic("pagedep_lookup: lock not held");
764 #endif
765 	mp = ITOV(ip)->v_mount;
766 	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
767 top:
768 	for (pagedep = LIST_FIRST(pagedephd); pagedep;
769 	     pagedep = LIST_NEXT(pagedep, pd_hash))
770 		if (ip->i_number == pagedep->pd_ino &&
771 		    lbn == pagedep->pd_lbn &&
772 		    mp == pagedep->pd_mnt)
773 			break;
774 	if (pagedep) {
775 		*pagedeppp = pagedep;
776 		return (1);
777 	}
778 	if ((flags & DEPALLOC) == 0) {
779 		*pagedeppp = NULL;
780 		return (0);
781 	}
782 	if (sema_get(&pagedep_in_progress, &lk) == 0) {
783 		ACQUIRE_LOCK(&lk);
784 		goto top;
785 	}
786 	MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
787 		M_WAITOK);
788 	bzero(pagedep, sizeof(struct pagedep));
789 	pagedep->pd_list.wk_type = D_PAGEDEP;
790 	pagedep->pd_mnt = mp;
791 	pagedep->pd_ino = ip->i_number;
792 	pagedep->pd_lbn = lbn;
793 	LIST_INIT(&pagedep->pd_dirremhd);
794 	LIST_INIT(&pagedep->pd_pendinghd);
795 	for (i = 0; i < DAHASHSZ; i++)
796 		LIST_INIT(&pagedep->pd_diraddhd[i]);
797 	ACQUIRE_LOCK(&lk);
798 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
799 	sema_release(&pagedep_in_progress);
800 	*pagedeppp = pagedep;
801 	return (0);
802 }
803 
804 /*
805  * Structures and routines associated with inodedep caching.
806  */
807 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
808 static u_long	inodedep_hash;	/* size of hash table - 1 */
809 static long	num_inodedep;	/* number of inodedep allocated */
810 #define	INODEDEP_HASH(fs, inum) \
811       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
812 static struct sema inodedep_in_progress;
813 
814 /*
815  * Look up a inodedep. Return 1 if found, 0 if not found.
816  * If not found, allocate if DEPALLOC flag is passed.
817  * Found or allocated entry is returned in inodedeppp.
818  * This routine must be called with splbio interrupts blocked.
819  */
820 static int
821 inodedep_lookup(fs, inum, flags, inodedeppp)
822 	struct fs *fs;
823 	ino_t inum;
824 	int flags;
825 	struct inodedep **inodedeppp;
826 {
827 	struct inodedep *inodedep;
828 	struct inodedep_hashhead *inodedephd;
829 	int firsttry;
830 
831 #ifdef DEBUG
832 	if (lk.lkt_held == -1)
833 		panic("inodedep_lookup: lock not held");
834 #endif
835 	firsttry = 1;
836 	inodedephd = INODEDEP_HASH(fs, inum);
837 top:
838 	for (inodedep = LIST_FIRST(inodedephd); inodedep;
839 	     inodedep = LIST_NEXT(inodedep, id_hash))
840 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
841 			break;
842 	if (inodedep) {
843 		*inodedeppp = inodedep;
844 		return (1);
845 	}
846 	if ((flags & DEPALLOC) == 0) {
847 		*inodedeppp = NULL;
848 		return (0);
849 	}
850 	/*
851 	 * If we are over our limit, try to improve the situation.
852 	 */
853 	if (num_inodedep > max_softdeps && firsttry && speedup_syncer() == 0 &&
854 	    request_cleanup(FLUSH_INODES, 1)) {
855 		firsttry = 0;
856 		goto top;
857 	}
858 	if (sema_get(&inodedep_in_progress, &lk) == 0) {
859 		ACQUIRE_LOCK(&lk);
860 		goto top;
861 	}
862 	num_inodedep += 1;
863 	MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
864 		M_INODEDEP, M_WAITOK);
865 	inodedep->id_list.wk_type = D_INODEDEP;
866 	inodedep->id_fs = fs;
867 	inodedep->id_ino = inum;
868 	inodedep->id_state = ALLCOMPLETE;
869 	inodedep->id_nlinkdelta = 0;
870 	inodedep->id_savedino = NULL;
871 	inodedep->id_savedsize = -1;
872 	inodedep->id_buf = NULL;
873 	LIST_INIT(&inodedep->id_pendinghd);
874 	LIST_INIT(&inodedep->id_inowait);
875 	LIST_INIT(&inodedep->id_bufwait);
876 	TAILQ_INIT(&inodedep->id_inoupdt);
877 	TAILQ_INIT(&inodedep->id_newinoupdt);
878 	ACQUIRE_LOCK(&lk);
879 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
880 	sema_release(&inodedep_in_progress);
881 	*inodedeppp = inodedep;
882 	return (0);
883 }
884 
885 /*
886  * Structures and routines associated with newblk caching.
887  */
888 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
889 u_long	newblk_hash;		/* size of hash table - 1 */
890 #define	NEWBLK_HASH(fs, inum) \
891 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
892 static struct sema newblk_in_progress;
893 
894 /*
895  * Look up a newblk. Return 1 if found, 0 if not found.
896  * If not found, allocate if DEPALLOC flag is passed.
897  * Found or allocated entry is returned in newblkpp.
898  */
899 static int
900 newblk_lookup(fs, newblkno, flags, newblkpp)
901 	struct fs *fs;
902 	ufs_daddr_t newblkno;
903 	int flags;
904 	struct newblk **newblkpp;
905 {
906 	struct newblk *newblk;
907 	struct newblk_hashhead *newblkhd;
908 
909 	newblkhd = NEWBLK_HASH(fs, newblkno);
910 top:
911 	for (newblk = LIST_FIRST(newblkhd); newblk;
912 	     newblk = LIST_NEXT(newblk, nb_hash))
913 		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
914 			break;
915 	if (newblk) {
916 		*newblkpp = newblk;
917 		return (1);
918 	}
919 	if ((flags & DEPALLOC) == 0) {
920 		*newblkpp = NULL;
921 		return (0);
922 	}
923 	if (sema_get(&newblk_in_progress, 0) == 0)
924 		goto top;
925 	MALLOC(newblk, struct newblk *, sizeof(struct newblk),
926 		M_NEWBLK, M_WAITOK);
927 	newblk->nb_state = 0;
928 	newblk->nb_fs = fs;
929 	newblk->nb_newblkno = newblkno;
930 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
931 	sema_release(&newblk_in_progress);
932 	*newblkpp = newblk;
933 	return (0);
934 }
935 
936 /*
937  * Executed during filesystem system initialization before
938  * mounting any file systems.
939  */
940 void
941 softdep_initialize()
942 {
943 
944 	LIST_INIT(&mkdirlisthd);
945 	LIST_INIT(&softdep_workitem_pending);
946 	max_softdeps = desiredvnodes * 8;
947 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
948 	    &pagedep_hash);
949 	sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0);
950 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
951 	sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0);
952 	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
953 	sema_init(&newblk_in_progress, "newblk", PRIBIO, 0);
954 }
955 
956 /*
957  * Called at mount time to notify the dependency code that a
958  * filesystem wishes to use it.
959  */
960 int
961 softdep_mount(devvp, mp, fs, cred)
962 	struct vnode *devvp;
963 	struct mount *mp;
964 	struct fs *fs;
965 	struct ucred *cred;
966 {
967 	struct csum cstotal;
968 	struct cg *cgp;
969 	struct buf *bp;
970 	int error, cyl;
971 
972 	mp->mnt_flag &= ~MNT_ASYNC;
973 	mp->mnt_flag |= MNT_SOFTDEP;
974 	/*
975 	 * When doing soft updates, the counters in the
976 	 * superblock may have gotten out of sync, so we have
977 	 * to scan the cylinder groups and recalculate them.
978 	 */
979 	if (fs->fs_clean != 0)
980 		return (0);
981 	bzero(&cstotal, sizeof cstotal);
982 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
983 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
984 		    fs->fs_cgsize, cred, &bp)) != 0) {
985 			brelse(bp);
986 			return (error);
987 		}
988 		cgp = (struct cg *)bp->b_data;
989 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
990 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
991 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
992 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
993 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
994 		brelse(bp);
995 	}
996 #ifdef DEBUG
997 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
998 		printf("ffs_mountfs: superblock updated for soft updates\n");
999 #endif
1000 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1001 	return (0);
1002 }
1003 
1004 /*
1005  * Protecting the freemaps (or bitmaps).
1006  *
1007  * To eliminate the need to execute fsck before mounting a file system
1008  * after a power failure, one must (conservatively) guarantee that the
1009  * on-disk copy of the bitmaps never indicate that a live inode or block is
1010  * free.  So, when a block or inode is allocated, the bitmap should be
1011  * updated (on disk) before any new pointers.  When a block or inode is
1012  * freed, the bitmap should not be updated until all pointers have been
1013  * reset.  The latter dependency is handled by the delayed de-allocation
1014  * approach described below for block and inode de-allocation.  The former
1015  * dependency is handled by calling the following procedure when a block or
1016  * inode is allocated. When an inode is allocated an "inodedep" is created
1017  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1018  * Each "inodedep" is also inserted into the hash indexing structure so
1019  * that any additional link additions can be made dependent on the inode
1020  * allocation.
1021  *
1022  * The ufs file system maintains a number of free block counts (e.g., per
1023  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1024  * in addition to the bitmaps.  These counts are used to improve efficiency
1025  * during allocation and therefore must be consistent with the bitmaps.
1026  * There is no convenient way to guarantee post-crash consistency of these
1027  * counts with simple update ordering, for two main reasons: (1) The counts
1028  * and bitmaps for a single cylinder group block are not in the same disk
1029  * sector.  If a disk write is interrupted (e.g., by power failure), one may
1030  * be written and the other not.  (2) Some of the counts are located in the
1031  * superblock rather than the cylinder group block. So, we focus our soft
1032  * updates implementation on protecting the bitmaps. When mounting a
1033  * filesystem, we recompute the auxiliary counts from the bitmaps.
1034  */
1035 
1036 /*
1037  * Called just after updating the cylinder group block to allocate an inode.
1038  */
1039 void
1040 softdep_setup_inomapdep(bp, ip, newinum)
1041 	struct buf *bp;		/* buffer for cylgroup block with inode map */
1042 	struct inode *ip;	/* inode related to allocation */
1043 	ino_t newinum;		/* new inode number being allocated */
1044 {
1045 	struct inodedep *inodedep;
1046 	struct bmsafemap *bmsafemap;
1047 
1048 	/*
1049 	 * Create a dependency for the newly allocated inode.
1050 	 * Panic if it already exists as something is seriously wrong.
1051 	 * Otherwise add it to the dependency list for the buffer holding
1052 	 * the cylinder group map from which it was allocated.
1053 	 */
1054 	ACQUIRE_LOCK(&lk);
1055 	if (inodedep_lookup(ip->i_fs, newinum, DEPALLOC, &inodedep) != 0)
1056 		panic("softdep_setup_inomapdep: found inode");
1057 	inodedep->id_buf = bp;
1058 	inodedep->id_state &= ~DEPCOMPLETE;
1059 	bmsafemap = bmsafemap_lookup(bp);
1060 	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1061 	FREE_LOCK(&lk);
1062 }
1063 
1064 /*
1065  * Called just after updating the cylinder group block to
1066  * allocate block or fragment.
1067  */
1068 void
1069 softdep_setup_blkmapdep(bp, fs, newblkno)
1070 	struct buf *bp;		/* buffer for cylgroup block with block map */
1071 	struct fs *fs;		/* filesystem doing allocation */
1072 	ufs_daddr_t newblkno;	/* number of newly allocated block */
1073 {
1074 	struct newblk *newblk;
1075 	struct bmsafemap *bmsafemap;
1076 
1077 	/*
1078 	 * Create a dependency for the newly allocated block.
1079 	 * Add it to the dependency list for the buffer holding
1080 	 * the cylinder group map from which it was allocated.
1081 	 */
1082 	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1083 		panic("softdep_setup_blkmapdep: found block");
1084 	ACQUIRE_LOCK(&lk);
1085 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1086 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1087 	FREE_LOCK(&lk);
1088 }
1089 
1090 /*
1091  * Find the bmsafemap associated with a cylinder group buffer.
1092  * If none exists, create one. The buffer must be locked when
1093  * this routine is called and this routine must be called with
1094  * splbio interrupts blocked.
1095  */
1096 static struct bmsafemap *
1097 bmsafemap_lookup(bp)
1098 	struct buf *bp;
1099 {
1100 	struct bmsafemap *bmsafemap;
1101 	struct worklist *wk;
1102 
1103 #ifdef DEBUG
1104 	if (lk.lkt_held == -1)
1105 		panic("bmsafemap_lookup: lock not held");
1106 #endif
1107 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = LIST_NEXT(wk, wk_list))
1108 		if (wk->wk_type == D_BMSAFEMAP)
1109 			return (WK_BMSAFEMAP(wk));
1110 	FREE_LOCK(&lk);
1111 	MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1112 		M_BMSAFEMAP, M_WAITOK);
1113 	bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1114 	bmsafemap->sm_list.wk_state = 0;
1115 	bmsafemap->sm_buf = bp;
1116 	LIST_INIT(&bmsafemap->sm_allocdirecthd);
1117 	LIST_INIT(&bmsafemap->sm_allocindirhd);
1118 	LIST_INIT(&bmsafemap->sm_inodedephd);
1119 	LIST_INIT(&bmsafemap->sm_newblkhd);
1120 	ACQUIRE_LOCK(&lk);
1121 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1122 	return (bmsafemap);
1123 }
1124 
1125 /*
1126  * Direct block allocation dependencies.
1127  *
1128  * When a new block is allocated, the corresponding disk locations must be
1129  * initialized (with zeros or new data) before the on-disk inode points to
1130  * them.  Also, the freemap from which the block was allocated must be
1131  * updated (on disk) before the inode's pointer. These two dependencies are
1132  * independent of each other and are needed for all file blocks and indirect
1133  * blocks that are pointed to directly by the inode.  Just before the
1134  * "in-core" version of the inode is updated with a newly allocated block
1135  * number, a procedure (below) is called to setup allocation dependency
1136  * structures.  These structures are removed when the corresponding
1137  * dependencies are satisfied or when the block allocation becomes obsolete
1138  * (i.e., the file is deleted, the block is de-allocated, or the block is a
1139  * fragment that gets upgraded).  All of these cases are handled in
1140  * procedures described later.
1141  *
1142  * When a file extension causes a fragment to be upgraded, either to a larger
1143  * fragment or to a full block, the on-disk location may change (if the
1144  * previous fragment could not simply be extended). In this case, the old
1145  * fragment must be de-allocated, but not until after the inode's pointer has
1146  * been updated. In most cases, this is handled by later procedures, which
1147  * will construct a "freefrag" structure to be added to the workitem queue
1148  * when the inode update is complete (or obsolete).  The main exception to
1149  * this is when an allocation occurs while a pending allocation dependency
1150  * (for the same block pointer) remains.  This case is handled in the main
1151  * allocation dependency setup procedure by immediately freeing the
1152  * unreferenced fragments.
1153  */
1154 void
1155 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1156 	struct inode *ip;	/* inode to which block is being added */
1157 	ufs_lbn_t lbn;		/* block pointer within inode */
1158 	ufs_daddr_t newblkno;	/* disk block number being added */
1159 	ufs_daddr_t oldblkno;	/* previous block number, 0 unless frag */
1160 	long newsize;		/* size of new block */
1161 	long oldsize;		/* size of new block */
1162 	struct buf *bp;		/* bp for allocated block */
1163 {
1164 	struct allocdirect *adp, *oldadp;
1165 	struct allocdirectlst *adphead;
1166 	struct bmsafemap *bmsafemap;
1167 	struct inodedep *inodedep;
1168 	struct pagedep *pagedep;
1169 	struct newblk *newblk;
1170 
1171 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1172 		M_ALLOCDIRECT, M_WAITOK);
1173 	bzero(adp, sizeof(struct allocdirect));
1174 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1175 	adp->ad_lbn = lbn;
1176 	adp->ad_newblkno = newblkno;
1177 	adp->ad_oldblkno = oldblkno;
1178 	adp->ad_newsize = newsize;
1179 	adp->ad_oldsize = oldsize;
1180 	adp->ad_state = ATTACHED;
1181 	if (newblkno == oldblkno)
1182 		adp->ad_freefrag = NULL;
1183 	else
1184 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1185 
1186 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1187 		panic("softdep_setup_allocdirect: lost block");
1188 
1189 	ACQUIRE_LOCK(&lk);
1190 	(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
1191 	adp->ad_inodedep = inodedep;
1192 
1193 	if (newblk->nb_state == DEPCOMPLETE) {
1194 		adp->ad_state |= DEPCOMPLETE;
1195 		adp->ad_buf = NULL;
1196 	} else {
1197 		bmsafemap = newblk->nb_bmsafemap;
1198 		adp->ad_buf = bmsafemap->sm_buf;
1199 		LIST_REMOVE(newblk, nb_deps);
1200 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1201 	}
1202 	LIST_REMOVE(newblk, nb_hash);
1203 	FREE(newblk, M_NEWBLK);
1204 
1205 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1206 	if (lbn >= NDADDR) {
1207 		/* allocating an indirect block */
1208 		if (oldblkno != 0)
1209 			panic("softdep_setup_allocdirect: non-zero indir");
1210 	} else {
1211 		/*
1212 		 * Allocating a direct block.
1213 		 *
1214 		 * If we are allocating a directory block, then we must
1215 		 * allocate an associated pagedep to track additions and
1216 		 * deletions.
1217 		 */
1218 		if ((ip->i_mode & IFMT) == IFDIR &&
1219 		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1220 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1221 	}
1222 	/*
1223 	 * The list of allocdirects must be kept in sorted and ascending
1224 	 * order so that the rollback routines can quickly determine the
1225 	 * first uncommitted block (the size of the file stored on disk
1226 	 * ends at the end of the lowest committed fragment, or if there
1227 	 * are no fragments, at the end of the highest committed block).
1228 	 * Since files generally grow, the typical case is that the new
1229 	 * block is to be added at the end of the list. We speed this
1230 	 * special case by checking against the last allocdirect in the
1231 	 * list before laboriously traversing the list looking for the
1232 	 * insertion point.
1233 	 */
1234 	adphead = &inodedep->id_newinoupdt;
1235 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1236 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1237 		/* insert at end of list */
1238 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1239 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1240 			allocdirect_merge(adphead, adp, oldadp);
1241 		FREE_LOCK(&lk);
1242 		return;
1243 	}
1244 	for (oldadp = TAILQ_FIRST(adphead); oldadp;
1245 	     oldadp = TAILQ_NEXT(oldadp, ad_next)) {
1246 		if (oldadp->ad_lbn >= lbn)
1247 			break;
1248 	}
1249 	if (oldadp == NULL)
1250 		panic("softdep_setup_allocdirect: lost entry");
1251 	/* insert in middle of list */
1252 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1253 	if (oldadp->ad_lbn == lbn)
1254 		allocdirect_merge(adphead, adp, oldadp);
1255 	FREE_LOCK(&lk);
1256 }
1257 
1258 /*
1259  * Replace an old allocdirect dependency with a newer one.
1260  * This routine must be called with splbio interrupts blocked.
1261  */
1262 static void
1263 allocdirect_merge(adphead, newadp, oldadp)
1264 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
1265 	struct allocdirect *newadp;	/* allocdirect being added */
1266 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
1267 {
1268 	struct freefrag *freefrag;
1269 
1270 #ifdef DEBUG
1271 	if (lk.lkt_held == -1)
1272 		panic("allocdirect_merge: lock not held");
1273 #endif
1274 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1275 	    newadp->ad_oldsize != oldadp->ad_newsize ||
1276 	    newadp->ad_lbn >= NDADDR)
1277 		panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1278 		    newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1279 		    NDADDR);
1280 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
1281 	newadp->ad_oldsize = oldadp->ad_oldsize;
1282 	/*
1283 	 * If the old dependency had a fragment to free or had never
1284 	 * previously had a block allocated, then the new dependency
1285 	 * can immediately post its freefrag and adopt the old freefrag.
1286 	 * This action is done by swapping the freefrag dependencies.
1287 	 * The new dependency gains the old one's freefrag, and the
1288 	 * old one gets the new one and then immediately puts it on
1289 	 * the worklist when it is freed by free_allocdirect. It is
1290 	 * not possible to do this swap when the old dependency had a
1291 	 * non-zero size but no previous fragment to free. This condition
1292 	 * arises when the new block is an extension of the old block.
1293 	 * Here, the first part of the fragment allocated to the new
1294 	 * dependency is part of the block currently claimed on disk by
1295 	 * the old dependency, so cannot legitimately be freed until the
1296 	 * conditions for the new dependency are fulfilled.
1297 	 */
1298 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1299 		freefrag = newadp->ad_freefrag;
1300 		newadp->ad_freefrag = oldadp->ad_freefrag;
1301 		oldadp->ad_freefrag = freefrag;
1302 	}
1303 	free_allocdirect(adphead, oldadp, 0);
1304 }
1305 
1306 /*
1307  * Allocate a new freefrag structure if needed.
1308  */
1309 static struct freefrag *
1310 newfreefrag(ip, blkno, size)
1311 	struct inode *ip;
1312 	ufs_daddr_t blkno;
1313 	long size;
1314 {
1315 	struct freefrag *freefrag;
1316 	struct fs *fs;
1317 
1318 	if (blkno == 0)
1319 		return (NULL);
1320 	fs = ip->i_fs;
1321 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1322 		panic("newfreefrag: frag size");
1323 	MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1324 		M_FREEFRAG, M_WAITOK);
1325 	freefrag->ff_list.wk_type = D_FREEFRAG;
1326 	freefrag->ff_state = ip->i_uid & ~ONWORKLIST;	/* XXX - used below */
1327 	freefrag->ff_inum = ip->i_number;
1328 	freefrag->ff_mnt = ITOV(ip)->v_mount;
1329 	freefrag->ff_devvp = ip->i_devvp;
1330 	freefrag->ff_blkno = blkno;
1331 	freefrag->ff_fragsize = size;
1332 	return (freefrag);
1333 }
1334 
1335 /*
1336  * This workitem de-allocates fragments that were replaced during
1337  * file block allocation.
1338  */
1339 static void
1340 handle_workitem_freefrag(freefrag)
1341 	struct freefrag *freefrag;
1342 {
1343 	struct inode tip;
1344 
1345 	tip.i_vnode = NULL;
1346 	tip.i_fs = VFSTOUFS(freefrag->ff_mnt)->um_fs;
1347 	tip.i_devvp = freefrag->ff_devvp;
1348 	tip.i_dev = freefrag->ff_devvp->v_rdev;
1349 	tip.i_number = freefrag->ff_inum;
1350 	tip.i_uid = freefrag->ff_state & ~ONWORKLIST;	/* XXX - set above */
1351 	ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1352 	FREE(freefrag, M_FREEFRAG);
1353 }
1354 
1355 /*
1356  * Indirect block allocation dependencies.
1357  *
1358  * The same dependencies that exist for a direct block also exist when
1359  * a new block is allocated and pointed to by an entry in a block of
1360  * indirect pointers. The undo/redo states described above are also
1361  * used here. Because an indirect block contains many pointers that
1362  * may have dependencies, a second copy of the entire in-memory indirect
1363  * block is kept. The buffer cache copy is always completely up-to-date.
1364  * The second copy, which is used only as a source for disk writes,
1365  * contains only the safe pointers (i.e., those that have no remaining
1366  * update dependencies). The second copy is freed when all pointers
1367  * are safe. The cache is not allowed to replace indirect blocks with
1368  * pending update dependencies. If a buffer containing an indirect
1369  * block with dependencies is written, these routines will mark it
1370  * dirty again. It can only be successfully written once all the
1371  * dependencies are removed. The ffs_fsync routine in conjunction with
1372  * softdep_sync_metadata work together to get all the dependencies
1373  * removed so that a file can be successfully written to disk. Three
1374  * procedures are used when setting up indirect block pointer
1375  * dependencies. The division is necessary because of the organization
1376  * of the "balloc" routine and because of the distinction between file
1377  * pages and file metadata blocks.
1378  */
1379 
1380 /*
1381  * Allocate a new allocindir structure.
1382  */
1383 static struct allocindir *
1384 newallocindir(ip, ptrno, newblkno, oldblkno)
1385 	struct inode *ip;	/* inode for file being extended */
1386 	int ptrno;		/* offset of pointer in indirect block */
1387 	ufs_daddr_t newblkno;	/* disk block number being added */
1388 	ufs_daddr_t oldblkno;	/* previous block number, 0 if none */
1389 {
1390 	struct allocindir *aip;
1391 
1392 	MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1393 		M_ALLOCINDIR, M_WAITOK);
1394 	bzero(aip, sizeof(struct allocindir));
1395 	aip->ai_list.wk_type = D_ALLOCINDIR;
1396 	aip->ai_state = ATTACHED;
1397 	aip->ai_offset = ptrno;
1398 	aip->ai_newblkno = newblkno;
1399 	aip->ai_oldblkno = oldblkno;
1400 	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1401 	return (aip);
1402 }
1403 
1404 /*
1405  * Called just before setting an indirect block pointer
1406  * to a newly allocated file page.
1407  */
1408 void
1409 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1410 	struct inode *ip;	/* inode for file being extended */
1411 	ufs_lbn_t lbn;		/* allocated block number within file */
1412 	struct buf *bp;		/* buffer with indirect blk referencing page */
1413 	int ptrno;		/* offset of pointer in indirect block */
1414 	ufs_daddr_t newblkno;	/* disk block number being added */
1415 	ufs_daddr_t oldblkno;	/* previous block number, 0 if none */
1416 	struct buf *nbp;	/* buffer holding allocated page */
1417 {
1418 	struct allocindir *aip;
1419 	struct pagedep *pagedep;
1420 
1421 	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1422 	ACQUIRE_LOCK(&lk);
1423 	/*
1424 	 * If we are allocating a directory page, then we must
1425 	 * allocate an associated pagedep to track additions and
1426 	 * deletions.
1427 	 */
1428 	if ((ip->i_mode & IFMT) == IFDIR &&
1429 	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1430 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1431 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1432 	FREE_LOCK(&lk);
1433 	setup_allocindir_phase2(bp, ip, aip);
1434 }
1435 
1436 /*
1437  * Called just before setting an indirect block pointer to a
1438  * newly allocated indirect block.
1439  */
1440 void
1441 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1442 	struct buf *nbp;	/* newly allocated indirect block */
1443 	struct inode *ip;	/* inode for file being extended */
1444 	struct buf *bp;		/* indirect block referencing allocated block */
1445 	int ptrno;		/* offset of pointer in indirect block */
1446 	ufs_daddr_t newblkno;	/* disk block number being added */
1447 {
1448 	struct allocindir *aip;
1449 
1450 	aip = newallocindir(ip, ptrno, newblkno, 0);
1451 	ACQUIRE_LOCK(&lk);
1452 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1453 	FREE_LOCK(&lk);
1454 	setup_allocindir_phase2(bp, ip, aip);
1455 }
1456 
1457 /*
1458  * Called to finish the allocation of the "aip" allocated
1459  * by one of the two routines above.
1460  */
1461 static void
1462 setup_allocindir_phase2(bp, ip, aip)
1463 	struct buf *bp;		/* in-memory copy of the indirect block */
1464 	struct inode *ip;	/* inode for file being extended */
1465 	struct allocindir *aip;	/* allocindir allocated by the above routines */
1466 {
1467 	struct worklist *wk;
1468 	struct indirdep *indirdep, *newindirdep;
1469 	struct bmsafemap *bmsafemap;
1470 	struct allocindir *oldaip;
1471 	struct freefrag *freefrag;
1472 	struct newblk *newblk;
1473 
1474 	if (bp->b_lblkno >= 0)
1475 		panic("setup_allocindir_phase2: not indir blk");
1476 	for (indirdep = NULL, newindirdep = NULL; ; ) {
1477 		ACQUIRE_LOCK(&lk);
1478 		for (wk = LIST_FIRST(&bp->b_dep); wk;
1479 		     wk = LIST_NEXT(wk, wk_list)) {
1480 			if (wk->wk_type != D_INDIRDEP)
1481 				continue;
1482 			indirdep = WK_INDIRDEP(wk);
1483 			break;
1484 		}
1485 		if (indirdep == NULL && newindirdep) {
1486 			indirdep = newindirdep;
1487 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1488 			newindirdep = NULL;
1489 		}
1490 		FREE_LOCK(&lk);
1491 		if (indirdep) {
1492 			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1493 			    &newblk) == 0)
1494 				panic("setup_allocindir: lost block");
1495 			ACQUIRE_LOCK(&lk);
1496 			if (newblk->nb_state == DEPCOMPLETE) {
1497 				aip->ai_state |= DEPCOMPLETE;
1498 				aip->ai_buf = NULL;
1499 			} else {
1500 				bmsafemap = newblk->nb_bmsafemap;
1501 				aip->ai_buf = bmsafemap->sm_buf;
1502 				LIST_REMOVE(newblk, nb_deps);
1503 				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1504 				    aip, ai_deps);
1505 			}
1506 			LIST_REMOVE(newblk, nb_hash);
1507 			FREE(newblk, M_NEWBLK);
1508 			aip->ai_indirdep = indirdep;
1509 			/*
1510 			 * Check to see if there is an existing dependency
1511 			 * for this block. If there is, merge the old
1512 			 * dependency into the new one.
1513 			 */
1514 			if (aip->ai_oldblkno == 0)
1515 				oldaip = NULL;
1516 			else
1517 				for (oldaip=LIST_FIRST(&indirdep->ir_deplisthd);
1518 				    oldaip; oldaip = LIST_NEXT(oldaip, ai_next))
1519 					if (oldaip->ai_offset == aip->ai_offset)
1520 						break;
1521 			freefrag = NULL;
1522 			if (oldaip != NULL) {
1523 				if (oldaip->ai_newblkno != aip->ai_oldblkno)
1524 					panic("setup_allocindir_phase2: blkno");
1525 				aip->ai_oldblkno = oldaip->ai_oldblkno;
1526 				freefrag = aip->ai_freefrag;
1527 				aip->ai_freefrag = oldaip->ai_freefrag;
1528 				oldaip->ai_freefrag = NULL;
1529 				free_allocindir(oldaip, NULL);
1530 			}
1531 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1532 			((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1533 			    [aip->ai_offset] = aip->ai_oldblkno;
1534 			FREE_LOCK(&lk);
1535 			if (freefrag != NULL)
1536 				handle_workitem_freefrag(freefrag);
1537 		}
1538 		if (newindirdep) {
1539 			if (indirdep->ir_savebp != NULL)
1540 				brelse(newindirdep->ir_savebp);
1541 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1542 		}
1543 		if (indirdep)
1544 			break;
1545 		MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1546 			M_INDIRDEP, M_WAITOK);
1547 		newindirdep->ir_list.wk_type = D_INDIRDEP;
1548 		newindirdep->ir_state = ATTACHED;
1549 		LIST_INIT(&newindirdep->ir_deplisthd);
1550 		LIST_INIT(&newindirdep->ir_donehd);
1551 		if (bp->b_blkno == bp->b_lblkno) {
1552 			VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
1553 				NULL, NULL);
1554 		}
1555 		newindirdep->ir_savebp =
1556 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0);
1557 		BUF_KERNPROC(newindirdep->ir_savebp);
1558 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1559 	}
1560 }
1561 
1562 /*
1563  * Block de-allocation dependencies.
1564  *
1565  * When blocks are de-allocated, the on-disk pointers must be nullified before
1566  * the blocks are made available for use by other files.  (The true
1567  * requirement is that old pointers must be nullified before new on-disk
1568  * pointers are set.  We chose this slightly more stringent requirement to
1569  * reduce complexity.) Our implementation handles this dependency by updating
1570  * the inode (or indirect block) appropriately but delaying the actual block
1571  * de-allocation (i.e., freemap and free space count manipulation) until
1572  * after the updated versions reach stable storage.  After the disk is
1573  * updated, the blocks can be safely de-allocated whenever it is convenient.
1574  * This implementation handles only the common case of reducing a file's
1575  * length to zero. Other cases are handled by the conventional synchronous
1576  * write approach.
1577  *
1578  * The ffs implementation with which we worked double-checks
1579  * the state of the block pointers and file size as it reduces
1580  * a file's length.  Some of this code is replicated here in our
1581  * soft updates implementation.  The freeblks->fb_chkcnt field is
1582  * used to transfer a part of this information to the procedure
1583  * that eventually de-allocates the blocks.
1584  *
1585  * This routine should be called from the routine that shortens
1586  * a file's length, before the inode's size or block pointers
1587  * are modified. It will save the block pointer information for
1588  * later release and zero the inode so that the calling routine
1589  * can release it.
1590  */
1591 void
1592 softdep_setup_freeblocks(ip, length)
1593 	struct inode *ip;	/* The inode whose length is to be reduced */
1594 	off_t length;		/* The new length for the file */
1595 {
1596 	struct freeblks *freeblks;
1597 	struct inodedep *inodedep;
1598 	struct allocdirect *adp;
1599 	struct vnode *vp;
1600 	struct buf *bp;
1601 	struct fs *fs;
1602 	int i, delay, error;
1603 
1604 	fs = ip->i_fs;
1605 	if (length != 0)
1606 		panic("softde_setup_freeblocks: non-zero length");
1607 	MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1608 		M_FREEBLKS, M_WAITOK);
1609 	bzero(freeblks, sizeof(struct freeblks));
1610 	freeblks->fb_list.wk_type = D_FREEBLKS;
1611 	freeblks->fb_uid = ip->i_uid;
1612 	freeblks->fb_previousinum = ip->i_number;
1613 	freeblks->fb_devvp = ip->i_devvp;
1614 	freeblks->fb_mnt = ITOV(ip)->v_mount;
1615 	freeblks->fb_oldsize = ip->i_size;
1616 	freeblks->fb_newsize = length;
1617 	freeblks->fb_chkcnt = ip->i_blocks;
1618 	for (i = 0; i < NDADDR; i++) {
1619 		freeblks->fb_dblks[i] = ip->i_db[i];
1620 		ip->i_db[i] = 0;
1621 	}
1622 	for (i = 0; i < NIADDR; i++) {
1623 		freeblks->fb_iblks[i] = ip->i_ib[i];
1624 		ip->i_ib[i] = 0;
1625 	}
1626 	ip->i_blocks = 0;
1627 	ip->i_size = 0;
1628 	/*
1629 	 * Push the zero'ed inode to to its disk buffer so that we are free
1630 	 * to delete its dependencies below. Once the dependencies are gone
1631 	 * the buffer can be safely released.
1632 	 */
1633 	if ((error = bread(ip->i_devvp,
1634 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
1635 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0)
1636 		softdep_error("softdep_setup_freeblocks", error);
1637 	*((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
1638 	    ip->i_din;
1639 	/*
1640 	 * Find and eliminate any inode dependencies.
1641 	 */
1642 	ACQUIRE_LOCK(&lk);
1643 	(void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1644 	if ((inodedep->id_state & IOSTARTED) != 0)
1645 		panic("softdep_setup_freeblocks: inode busy");
1646 	/*
1647 	 * Because the file length has been truncated to zero, any
1648 	 * pending block allocation dependency structures associated
1649 	 * with this inode are obsolete and can simply be de-allocated.
1650 	 * We must first merge the two dependency lists to get rid of
1651 	 * any duplicate freefrag structures, then purge the merged list.
1652 	 * If we still have a bitmap dependency, then the inode has never
1653 	 * been written to disk, so we can free any fragments without delay.
1654 	 */
1655 	merge_inode_lists(inodedep);
1656 	delay = (inodedep->id_state & DEPCOMPLETE);
1657 	while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1658 		free_allocdirect(&inodedep->id_inoupdt, adp, delay);
1659 	FREE_LOCK(&lk);
1660 	bdwrite(bp);
1661 	/*
1662 	 * We must wait for any I/O in progress to finish so that
1663 	 * all potential buffers on the dirty list will be visible.
1664 	 * Once they are all there, walk the list and get rid of
1665 	 * any dependencies.
1666 	 */
1667 	vp = ITOV(ip);
1668 	ACQUIRE_LOCK(&lk);
1669 	drain_output(vp, 1);
1670 	while (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) {
1671 		bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
1672 		(void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
1673 		deallocate_dependencies(bp, inodedep);
1674 		bp->b_flags |= B_INVAL | B_NOCACHE;
1675 		FREE_LOCK(&lk);
1676 		brelse(bp);
1677 		ACQUIRE_LOCK(&lk);
1678 	}
1679 	/*
1680 	 * Add the freeblks structure to the list of operations that
1681 	 * must await the zero'ed inode being written to disk. If we
1682 	 * still have a bitmap dependency, then the inode has never been
1683 	 * written to disk, so we can process the freeblks immediately.
1684 	 * If the inodedep does not exist, then the zero'ed inode has
1685 	 * been written and we can also proceed.
1686 	 */
1687 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0 ||
1688 	    free_inodedep(inodedep) ||
1689 	    (inodedep->id_state & DEPCOMPLETE) == 0) {
1690 		FREE_LOCK(&lk);
1691 		handle_workitem_freeblocks(freeblks);
1692 	} else {
1693 		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1694 		FREE_LOCK(&lk);
1695 	}
1696 }
1697 
1698 /*
1699  * Reclaim any dependency structures from a buffer that is about to
1700  * be reallocated to a new vnode. The buffer must be locked, thus,
1701  * no I/O completion operations can occur while we are manipulating
1702  * its associated dependencies. The mutex is held so that other I/O's
1703  * associated with related dependencies do not occur.
1704  */
1705 static void
1706 deallocate_dependencies(bp, inodedep)
1707 	struct buf *bp;
1708 	struct inodedep *inodedep;
1709 {
1710 	struct worklist *wk;
1711 	struct indirdep *indirdep;
1712 	struct allocindir *aip;
1713 	struct pagedep *pagedep;
1714 	struct dirrem *dirrem;
1715 	struct diradd *dap;
1716 	int i;
1717 
1718 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1719 		switch (wk->wk_type) {
1720 
1721 		case D_INDIRDEP:
1722 			indirdep = WK_INDIRDEP(wk);
1723 			/*
1724 			 * None of the indirect pointers will ever be visible,
1725 			 * so they can simply be tossed. GOINGAWAY ensures
1726 			 * that allocated pointers will be saved in the buffer
1727 			 * cache until they are freed. Note that they will
1728 			 * only be able to be found by their physical address
1729 			 * since the inode mapping the logical address will
1730 			 * be gone. The save buffer used for the safe copy
1731 			 * was allocated in setup_allocindir_phase2 using
1732 			 * the physical address so it could be used for this
1733 			 * purpose. Hence we swap the safe copy with the real
1734 			 * copy, allowing the safe copy to be freed and holding
1735 			 * on to the real copy for later use in indir_trunc.
1736 			 */
1737 			if (indirdep->ir_state & GOINGAWAY)
1738 				panic("deallocate_dependencies: already gone");
1739 			indirdep->ir_state |= GOINGAWAY;
1740 			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
1741 				free_allocindir(aip, inodedep);
1742 			if (bp->b_lblkno >= 0 ||
1743 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno)
1744 				panic("deallocate_dependencies: not indir");
1745 			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
1746 			    bp->b_bcount);
1747 			WORKLIST_REMOVE(wk);
1748 			WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
1749 			continue;
1750 
1751 		case D_PAGEDEP:
1752 			pagedep = WK_PAGEDEP(wk);
1753 			/*
1754 			 * None of the directory additions will ever be
1755 			 * visible, so they can simply be tossed.
1756 			 */
1757 			for (i = 0; i < DAHASHSZ; i++)
1758 				while ((dap =
1759 				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
1760 					free_diradd(dap);
1761 			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
1762 				free_diradd(dap);
1763 			/*
1764 			 * Copy any directory remove dependencies to the list
1765 			 * to be processed after the zero'ed inode is written.
1766 			 * If the inode has already been written, then they
1767 			 * can be dumped directly onto the work list.
1768 			 */
1769 			for (dirrem = LIST_FIRST(&pagedep->pd_dirremhd); dirrem;
1770 			     dirrem = LIST_NEXT(dirrem, dm_next)) {
1771 				LIST_REMOVE(dirrem, dm_next);
1772 				dirrem->dm_dirinum = pagedep->pd_ino;
1773 				if (inodedep == NULL ||
1774 				    (inodedep->id_state & ALLCOMPLETE) ==
1775 				     ALLCOMPLETE)
1776 					add_to_worklist(&dirrem->dm_list);
1777 				else
1778 					WORKLIST_INSERT(&inodedep->id_bufwait,
1779 					    &dirrem->dm_list);
1780 			}
1781 			WORKLIST_REMOVE(&pagedep->pd_list);
1782 			LIST_REMOVE(pagedep, pd_hash);
1783 			WORKITEM_FREE(pagedep, D_PAGEDEP);
1784 			continue;
1785 
1786 		case D_ALLOCINDIR:
1787 			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
1788 			continue;
1789 
1790 		case D_ALLOCDIRECT:
1791 		case D_INODEDEP:
1792 			panic("deallocate_dependencies: Unexpected type %s",
1793 			    TYPENAME(wk->wk_type));
1794 			/* NOTREACHED */
1795 
1796 		default:
1797 			panic("deallocate_dependencies: Unknown type %s",
1798 			    TYPENAME(wk->wk_type));
1799 			/* NOTREACHED */
1800 		}
1801 	}
1802 }
1803 
1804 /*
1805  * Free an allocdirect. Generate a new freefrag work request if appropriate.
1806  * This routine must be called with splbio interrupts blocked.
1807  */
1808 static void
1809 free_allocdirect(adphead, adp, delay)
1810 	struct allocdirectlst *adphead;
1811 	struct allocdirect *adp;
1812 	int delay;
1813 {
1814 
1815 #ifdef DEBUG
1816 	if (lk.lkt_held == -1)
1817 		panic("free_allocdirect: lock not held");
1818 #endif
1819 	if ((adp->ad_state & DEPCOMPLETE) == 0)
1820 		LIST_REMOVE(adp, ad_deps);
1821 	TAILQ_REMOVE(adphead, adp, ad_next);
1822 	if ((adp->ad_state & COMPLETE) == 0)
1823 		WORKLIST_REMOVE(&adp->ad_list);
1824 	if (adp->ad_freefrag != NULL) {
1825 		if (delay)
1826 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
1827 			    &adp->ad_freefrag->ff_list);
1828 		else
1829 			add_to_worklist(&adp->ad_freefrag->ff_list);
1830 	}
1831 	WORKITEM_FREE(adp, D_ALLOCDIRECT);
1832 }
1833 
1834 /*
1835  * Prepare an inode to be freed. The actual free operation is not
1836  * done until the zero'ed inode has been written to disk.
1837  */
1838 void
1839 softdep_freefile(pvp, ino, mode)
1840 		struct vnode *pvp;
1841 		ino_t ino;
1842 		int mode;
1843 {
1844 	struct inode *ip = VTOI(pvp);
1845 	struct inodedep *inodedep;
1846 	struct freefile *freefile;
1847 
1848 	/*
1849 	 * This sets up the inode de-allocation dependency.
1850 	 */
1851 	MALLOC(freefile, struct freefile *, sizeof(struct freefile),
1852 		M_FREEFILE, M_WAITOK);
1853 	freefile->fx_list.wk_type = D_FREEFILE;
1854 	freefile->fx_list.wk_state = 0;
1855 	freefile->fx_mode = mode;
1856 	freefile->fx_oldinum = ino;
1857 	freefile->fx_devvp = ip->i_devvp;
1858 	freefile->fx_mnt = ITOV(ip)->v_mount;
1859 
1860 	/*
1861 	 * If the inodedep does not exist, then the zero'ed inode has
1862 	 * been written to disk. If the allocated inode has never been
1863 	 * written to disk, then the on-disk inode is zero'ed. In either
1864 	 * case we can free the file immediately.
1865 	 */
1866 	ACQUIRE_LOCK(&lk);
1867 	if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
1868 	    check_inode_unwritten(inodedep)) {
1869 		FREE_LOCK(&lk);
1870 		handle_workitem_freefile(freefile);
1871 		return;
1872 	}
1873 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
1874 	FREE_LOCK(&lk);
1875 }
1876 
1877 /*
1878  * Check to see if an inode has never been written to disk. If
1879  * so free the inodedep and return success, otherwise return failure.
1880  * This routine must be called with splbio interrupts blocked.
1881  *
1882  * If we still have a bitmap dependency, then the inode has never
1883  * been written to disk. Drop the dependency as it is no longer
1884  * necessary since the inode is being deallocated. We set the
1885  * ALLCOMPLETE flags since the bitmap now properly shows that the
1886  * inode is not allocated. Even if the inode is actively being
1887  * written, it has been rolled back to its zero'ed state, so we
1888  * are ensured that a zero inode is what is on the disk. For short
1889  * lived files, this change will usually result in removing all the
1890  * dependencies from the inode so that it can be freed immediately.
1891  */
1892 static int
1893 check_inode_unwritten(inodedep)
1894 	struct inodedep *inodedep;
1895 {
1896 
1897 	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
1898 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
1899 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
1900 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
1901 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
1902 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
1903 	    inodedep->id_nlinkdelta != 0)
1904 		return (0);
1905 	inodedep->id_state |= ALLCOMPLETE;
1906 	LIST_REMOVE(inodedep, id_deps);
1907 	inodedep->id_buf = NULL;
1908 	if (inodedep->id_state & ONWORKLIST)
1909 		WORKLIST_REMOVE(&inodedep->id_list);
1910 	if (inodedep->id_savedino != NULL) {
1911 		FREE(inodedep->id_savedino, M_INODEDEP);
1912 		inodedep->id_savedino = NULL;
1913 	}
1914 	if (free_inodedep(inodedep) == 0)
1915 		panic("check_inode_unwritten: busy inode");
1916 	return (1);
1917 }
1918 
1919 /*
1920  * Try to free an inodedep structure. Return 1 if it could be freed.
1921  */
1922 static int
1923 free_inodedep(inodedep)
1924 	struct inodedep *inodedep;
1925 {
1926 
1927 	if ((inodedep->id_state & ONWORKLIST) != 0 ||
1928 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
1929 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
1930 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
1931 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
1932 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
1933 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
1934 	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
1935 		return (0);
1936 	LIST_REMOVE(inodedep, id_hash);
1937 	WORKITEM_FREE(inodedep, D_INODEDEP);
1938 	num_inodedep -= 1;
1939 	return (1);
1940 }
1941 
1942 /*
1943  * This workitem routine performs the block de-allocation.
1944  * The workitem is added to the pending list after the updated
1945  * inode block has been written to disk.  As mentioned above,
1946  * checks regarding the number of blocks de-allocated (compared
1947  * to the number of blocks allocated for the file) are also
1948  * performed in this function.
1949  */
1950 static void
1951 handle_workitem_freeblocks(freeblks)
1952 	struct freeblks *freeblks;
1953 {
1954 	struct inode tip;
1955 	ufs_daddr_t bn;
1956 	struct fs *fs;
1957 	int i, level, bsize;
1958 	long nblocks, blocksreleased = 0;
1959 	int error, allerror = 0;
1960 	ufs_lbn_t baselbns[NIADDR], tmpval;
1961 
1962 	tip.i_fs = fs = VFSTOUFS(freeblks->fb_mnt)->um_fs;
1963 	tip.i_number = freeblks->fb_previousinum;
1964 	tip.i_devvp = freeblks->fb_devvp;
1965 	tip.i_dev = freeblks->fb_devvp->v_rdev;
1966 	tip.i_size = freeblks->fb_oldsize;
1967 	tip.i_uid = freeblks->fb_uid;
1968 	tip.i_vnode = NULL;
1969 	tmpval = 1;
1970 	baselbns[0] = NDADDR;
1971 	for (i = 1; i < NIADDR; i++) {
1972 		tmpval *= NINDIR(fs);
1973 		baselbns[i] = baselbns[i - 1] + tmpval;
1974 	}
1975 	nblocks = btodb(fs->fs_bsize);
1976 	blocksreleased = 0;
1977 	/*
1978 	 * Indirect blocks first.
1979 	 */
1980 	for (level = (NIADDR - 1); level >= 0; level--) {
1981 		if ((bn = freeblks->fb_iblks[level]) == 0)
1982 			continue;
1983 		if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level,
1984 		    baselbns[level], &blocksreleased)) == 0)
1985 			allerror = error;
1986 		ffs_blkfree(&tip, bn, fs->fs_bsize);
1987 		blocksreleased += nblocks;
1988 	}
1989 	/*
1990 	 * All direct blocks or frags.
1991 	 */
1992 	for (i = (NDADDR - 1); i >= 0; i--) {
1993 		if ((bn = freeblks->fb_dblks[i]) == 0)
1994 			continue;
1995 		bsize = blksize(fs, &tip, i);
1996 		ffs_blkfree(&tip, bn, bsize);
1997 		blocksreleased += btodb(bsize);
1998 	}
1999 
2000 #ifdef DIAGNOSTIC
2001 	if (freeblks->fb_chkcnt != blocksreleased)
2002 		panic("handle_workitem_freeblocks: block count");
2003 	if (allerror)
2004 		softdep_error("handle_workitem_freeblks", allerror);
2005 #endif /* DIAGNOSTIC */
2006 	WORKITEM_FREE(freeblks, D_FREEBLKS);
2007 }
2008 
2009 /*
2010  * Release blocks associated with the inode ip and stored in the indirect
2011  * block dbn. If level is greater than SINGLE, the block is an indirect block
2012  * and recursive calls to indirtrunc must be used to cleanse other indirect
2013  * blocks.
2014  */
2015 static int
2016 indir_trunc(ip, dbn, level, lbn, countp)
2017 	struct inode *ip;
2018 	ufs_daddr_t dbn;
2019 	int level;
2020 	ufs_lbn_t lbn;
2021 	long *countp;
2022 {
2023 	struct buf *bp;
2024 	ufs_daddr_t *bap;
2025 	ufs_daddr_t nb;
2026 	struct fs *fs;
2027 	struct worklist *wk;
2028 	struct indirdep *indirdep;
2029 	int i, lbnadd, nblocks;
2030 	int error, allerror = 0;
2031 
2032 	fs = ip->i_fs;
2033 	lbnadd = 1;
2034 	for (i = level; i > 0; i--)
2035 		lbnadd *= NINDIR(fs);
2036 	/*
2037 	 * Get buffer of block pointers to be freed. This routine is not
2038 	 * called until the zero'ed inode has been written, so it is safe
2039 	 * to free blocks as they are encountered. Because the inode has
2040 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
2041 	 * have to use the on-disk address and the block device for the
2042 	 * filesystem to look them up. If the file was deleted before its
2043 	 * indirect blocks were all written to disk, the routine that set
2044 	 * us up (deallocate_dependencies) will have arranged to leave
2045 	 * a complete copy of the indirect block in memory for our use.
2046 	 * Otherwise we have to read the blocks in from the disk.
2047 	 */
2048 	ACQUIRE_LOCK(&lk);
2049 	if ((bp = incore(ip->i_devvp, dbn)) != NULL &&
2050 	    (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2051 		if (wk->wk_type != D_INDIRDEP ||
2052 		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2053 		    (indirdep->ir_state & GOINGAWAY) == 0)
2054 			panic("indir_trunc: lost indirdep");
2055 		WORKLIST_REMOVE(wk);
2056 		WORKITEM_FREE(indirdep, D_INDIRDEP);
2057 		if (LIST_FIRST(&bp->b_dep) != NULL)
2058 			panic("indir_trunc: dangling dep");
2059 		FREE_LOCK(&lk);
2060 	} else {
2061 		FREE_LOCK(&lk);
2062 		error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, NOCRED, &bp);
2063 		if (error)
2064 			return (error);
2065 	}
2066 	/*
2067 	 * Recursively free indirect blocks.
2068 	 */
2069 	bap = (ufs_daddr_t *)bp->b_data;
2070 	nblocks = btodb(fs->fs_bsize);
2071 	for (i = NINDIR(fs) - 1; i >= 0; i--) {
2072 		if ((nb = bap[i]) == 0)
2073 			continue;
2074 		if (level != 0) {
2075 			if ((error = indir_trunc(ip, fsbtodb(fs, nb),
2076 			     level - 1, lbn + (i * lbnadd), countp)) != 0)
2077 				allerror = error;
2078 		}
2079 		ffs_blkfree(ip, nb, fs->fs_bsize);
2080 		*countp += nblocks;
2081 	}
2082 	bp->b_flags |= B_INVAL | B_NOCACHE;
2083 	brelse(bp);
2084 	return (allerror);
2085 }
2086 
2087 /*
2088  * Free an allocindir.
2089  * This routine must be called with splbio interrupts blocked.
2090  */
2091 static void
2092 free_allocindir(aip, inodedep)
2093 	struct allocindir *aip;
2094 	struct inodedep *inodedep;
2095 {
2096 	struct freefrag *freefrag;
2097 
2098 #ifdef DEBUG
2099 	if (lk.lkt_held == -1)
2100 		panic("free_allocindir: lock not held");
2101 #endif
2102 	if ((aip->ai_state & DEPCOMPLETE) == 0)
2103 		LIST_REMOVE(aip, ai_deps);
2104 	if (aip->ai_state & ONWORKLIST)
2105 		WORKLIST_REMOVE(&aip->ai_list);
2106 	LIST_REMOVE(aip, ai_next);
2107 	if ((freefrag = aip->ai_freefrag) != NULL) {
2108 		if (inodedep == NULL)
2109 			add_to_worklist(&freefrag->ff_list);
2110 		else
2111 			WORKLIST_INSERT(&inodedep->id_bufwait,
2112 			    &freefrag->ff_list);
2113 	}
2114 	WORKITEM_FREE(aip, D_ALLOCINDIR);
2115 }
2116 
2117 /*
2118  * Directory entry addition dependencies.
2119  *
2120  * When adding a new directory entry, the inode (with its incremented link
2121  * count) must be written to disk before the directory entry's pointer to it.
2122  * Also, if the inode is newly allocated, the corresponding freemap must be
2123  * updated (on disk) before the directory entry's pointer. These requirements
2124  * are met via undo/redo on the directory entry's pointer, which consists
2125  * simply of the inode number.
2126  *
2127  * As directory entries are added and deleted, the free space within a
2128  * directory block can become fragmented.  The ufs file system will compact
2129  * a fragmented directory block to make space for a new entry. When this
2130  * occurs, the offsets of previously added entries change. Any "diradd"
2131  * dependency structures corresponding to these entries must be updated with
2132  * the new offsets.
2133  */
2134 
2135 /*
2136  * This routine is called after the in-memory inode's link
2137  * count has been incremented, but before the directory entry's
2138  * pointer to the inode has been set.
2139  */
2140 void
2141 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp)
2142 	struct buf *bp;		/* buffer containing directory block */
2143 	struct inode *dp;	/* inode for directory */
2144 	off_t diroffset;	/* offset of new entry in directory */
2145 	long newinum;		/* inode referenced by new directory entry */
2146 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
2147 {
2148 	int offset;		/* offset of new entry within directory block */
2149 	ufs_lbn_t lbn;		/* block in directory containing new entry */
2150 	struct fs *fs;
2151 	struct diradd *dap;
2152 	struct pagedep *pagedep;
2153 	struct inodedep *inodedep;
2154 	struct mkdir *mkdir1, *mkdir2;
2155 
2156 	/*
2157 	 * Whiteouts have no dependencies.
2158 	 */
2159 	if (newinum == WINO) {
2160 		if (newdirbp != NULL)
2161 			bdwrite(newdirbp);
2162 		return;
2163 	}
2164 
2165 	fs = dp->i_fs;
2166 	lbn = lblkno(fs, diroffset);
2167 	offset = blkoff(fs, diroffset);
2168 	MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, M_WAITOK);
2169 	bzero(dap, sizeof(struct diradd));
2170 	dap->da_list.wk_type = D_DIRADD;
2171 	dap->da_offset = offset;
2172 	dap->da_newinum = newinum;
2173 	dap->da_state = ATTACHED;
2174 	if (newdirbp == NULL) {
2175 		dap->da_state |= DEPCOMPLETE;
2176 		ACQUIRE_LOCK(&lk);
2177 	} else {
2178 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2179 		MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2180 		    M_WAITOK);
2181 		mkdir1->md_list.wk_type = D_MKDIR;
2182 		mkdir1->md_state = MKDIR_BODY;
2183 		mkdir1->md_diradd = dap;
2184 		MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2185 		    M_WAITOK);
2186 		mkdir2->md_list.wk_type = D_MKDIR;
2187 		mkdir2->md_state = MKDIR_PARENT;
2188 		mkdir2->md_diradd = dap;
2189 		/*
2190 		 * Dependency on "." and ".." being written to disk.
2191 		 */
2192 		mkdir1->md_buf = newdirbp;
2193 		ACQUIRE_LOCK(&lk);
2194 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2195 		WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2196 		FREE_LOCK(&lk);
2197 		bdwrite(newdirbp);
2198 		/*
2199 		 * Dependency on link count increase for parent directory
2200 		 */
2201 		ACQUIRE_LOCK(&lk);
2202 		if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2203 		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2204 			dap->da_state &= ~MKDIR_PARENT;
2205 			WORKITEM_FREE(mkdir2, D_MKDIR);
2206 		} else {
2207 			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2208 			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2209 		}
2210 	}
2211 	/*
2212 	 * Link into parent directory pagedep to await its being written.
2213 	 */
2214 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2215 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2216 	dap->da_pagedep = pagedep;
2217 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2218 	    da_pdlist);
2219 	/*
2220 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
2221 	 * is not yet written. If it is written, do the post-inode write
2222 	 * processing to put it on the id_pendinghd list.
2223 	 */
2224 	(void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2225 	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2226 		diradd_inode_written(dap, inodedep);
2227 	else
2228 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2229 	FREE_LOCK(&lk);
2230 }
2231 
2232 /*
2233  * This procedure is called to change the offset of a directory
2234  * entry when compacting a directory block which must be owned
2235  * exclusively by the caller. Note that the actual entry movement
2236  * must be done in this procedure to ensure that no I/O completions
2237  * occur while the move is in progress.
2238  */
2239 void
2240 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2241 	struct inode *dp;	/* inode for directory */
2242 	caddr_t base;		/* address of dp->i_offset */
2243 	caddr_t oldloc;		/* address of old directory location */
2244 	caddr_t newloc;		/* address of new directory location */
2245 	int entrysize;		/* size of directory entry */
2246 {
2247 	int offset, oldoffset, newoffset;
2248 	struct pagedep *pagedep;
2249 	struct diradd *dap;
2250 	ufs_lbn_t lbn;
2251 
2252 	ACQUIRE_LOCK(&lk);
2253 	lbn = lblkno(dp->i_fs, dp->i_offset);
2254 	offset = blkoff(dp->i_fs, dp->i_offset);
2255 	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2256 		goto done;
2257 	oldoffset = offset + (oldloc - base);
2258 	newoffset = offset + (newloc - base);
2259 	for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(oldoffset)]);
2260 	     dap; dap = LIST_NEXT(dap, da_pdlist)) {
2261 		if (dap->da_offset != oldoffset)
2262 			continue;
2263 		dap->da_offset = newoffset;
2264 		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2265 			break;
2266 		LIST_REMOVE(dap, da_pdlist);
2267 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2268 		    dap, da_pdlist);
2269 		break;
2270 	}
2271 	if (dap == NULL) {
2272 		for (dap = LIST_FIRST(&pagedep->pd_pendinghd);
2273 		     dap; dap = LIST_NEXT(dap, da_pdlist)) {
2274 			if (dap->da_offset == oldoffset) {
2275 				dap->da_offset = newoffset;
2276 				break;
2277 			}
2278 		}
2279 	}
2280 done:
2281 	bcopy(oldloc, newloc, entrysize);
2282 	FREE_LOCK(&lk);
2283 }
2284 
2285 /*
2286  * Free a diradd dependency structure. This routine must be called
2287  * with splbio interrupts blocked.
2288  */
2289 static void
2290 free_diradd(dap)
2291 	struct diradd *dap;
2292 {
2293 	struct dirrem *dirrem;
2294 	struct pagedep *pagedep;
2295 	struct inodedep *inodedep;
2296 	struct mkdir *mkdir, *nextmd;
2297 
2298 #ifdef DEBUG
2299 	if (lk.lkt_held == -1)
2300 		panic("free_diradd: lock not held");
2301 #endif
2302 	WORKLIST_REMOVE(&dap->da_list);
2303 	LIST_REMOVE(dap, da_pdlist);
2304 	if ((dap->da_state & DIRCHG) == 0) {
2305 		pagedep = dap->da_pagedep;
2306 	} else {
2307 		dirrem = dap->da_previous;
2308 		pagedep = dirrem->dm_pagedep;
2309 		dirrem->dm_dirinum = pagedep->pd_ino;
2310 		add_to_worklist(&dirrem->dm_list);
2311 	}
2312 	if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2313 	    0, &inodedep) != 0)
2314 		(void) free_inodedep(inodedep);
2315 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2316 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2317 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
2318 			if (mkdir->md_diradd != dap)
2319 				continue;
2320 			dap->da_state &= ~mkdir->md_state;
2321 			WORKLIST_REMOVE(&mkdir->md_list);
2322 			LIST_REMOVE(mkdir, md_mkdirs);
2323 			WORKITEM_FREE(mkdir, D_MKDIR);
2324 		}
2325 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
2326 			panic("free_diradd: unfound ref");
2327 	}
2328 	WORKITEM_FREE(dap, D_DIRADD);
2329 }
2330 
2331 /*
2332  * Directory entry removal dependencies.
2333  *
2334  * When removing a directory entry, the entry's inode pointer must be
2335  * zero'ed on disk before the corresponding inode's link count is decremented
2336  * (possibly freeing the inode for re-use). This dependency is handled by
2337  * updating the directory entry but delaying the inode count reduction until
2338  * after the directory block has been written to disk. After this point, the
2339  * inode count can be decremented whenever it is convenient.
2340  */
2341 
2342 /*
2343  * This routine should be called immediately after removing
2344  * a directory entry.  The inode's link count should not be
2345  * decremented by the calling procedure -- the soft updates
2346  * code will do this task when it is safe.
2347  */
2348 void
2349 softdep_setup_remove(bp, dp, ip, isrmdir)
2350 	struct buf *bp;		/* buffer containing directory block */
2351 	struct inode *dp;	/* inode for the directory being modified */
2352 	struct inode *ip;	/* inode for directory entry being removed */
2353 	int isrmdir;		/* indicates if doing RMDIR */
2354 {
2355 	struct dirrem *dirrem, *prevdirrem;
2356 
2357 	/*
2358 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2359 	 */
2360 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2361 
2362 	/*
2363 	 * If the COMPLETE flag is clear, then there were no active
2364 	 * entries and we want to roll back to a zeroed entry until
2365 	 * the new inode is committed to disk. If the COMPLETE flag is
2366 	 * set then we have deleted an entry that never made it to
2367 	 * disk. If the entry we deleted resulted from a name change,
2368 	 * then the old name still resides on disk. We cannot delete
2369 	 * its inode (returned to us in prevdirrem) until the zeroed
2370 	 * directory entry gets to disk. The new inode has never been
2371 	 * referenced on the disk, so can be deleted immediately.
2372 	 */
2373 	if ((dirrem->dm_state & COMPLETE) == 0) {
2374 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2375 		    dm_next);
2376 		FREE_LOCK(&lk);
2377 	} else {
2378 		if (prevdirrem != NULL)
2379 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2380 			    prevdirrem, dm_next);
2381 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2382 		FREE_LOCK(&lk);
2383 		handle_workitem_remove(dirrem);
2384 	}
2385 }
2386 
2387 /*
2388  * Allocate a new dirrem if appropriate and return it along with
2389  * its associated pagedep. Called without a lock, returns with lock.
2390  */
2391 static long num_dirrem;		/* number of dirrem allocated */
2392 static struct dirrem *
2393 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2394 	struct buf *bp;		/* buffer containing directory block */
2395 	struct inode *dp;	/* inode for the directory being modified */
2396 	struct inode *ip;	/* inode for directory entry being removed */
2397 	int isrmdir;		/* indicates if doing RMDIR */
2398 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
2399 {
2400 	int offset;
2401 	ufs_lbn_t lbn;
2402 	struct diradd *dap;
2403 	struct dirrem *dirrem;
2404 	struct pagedep *pagedep;
2405 
2406 	/*
2407 	 * Whiteouts have no deletion dependencies.
2408 	 */
2409 	if (ip == NULL)
2410 		panic("newdirrem: whiteout");
2411 	/*
2412 	 * If we are over our limit, try to improve the situation.
2413 	 * Limiting the number of dirrem structures will also limit
2414 	 * the number of freefile and freeblks structures.
2415 	 */
2416 	if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2417 		(void) request_cleanup(FLUSH_REMOVE, 0);
2418 	num_dirrem += 1;
2419 	MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
2420 		M_DIRREM, M_WAITOK);
2421 	bzero(dirrem, sizeof(struct dirrem));
2422 	dirrem->dm_list.wk_type = D_DIRREM;
2423 	dirrem->dm_state = isrmdir ? RMDIR : 0;
2424 	dirrem->dm_mnt = ITOV(ip)->v_mount;
2425 	dirrem->dm_oldinum = ip->i_number;
2426 	*prevdirremp = NULL;
2427 
2428 	ACQUIRE_LOCK(&lk);
2429 	lbn = lblkno(dp->i_fs, dp->i_offset);
2430 	offset = blkoff(dp->i_fs, dp->i_offset);
2431 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2432 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2433 	dirrem->dm_pagedep = pagedep;
2434 	/*
2435 	 * Check for a diradd dependency for the same directory entry.
2436 	 * If present, then both dependencies become obsolete and can
2437 	 * be de-allocated. Check for an entry on both the pd_dirraddhd
2438 	 * list and the pd_pendinghd list.
2439 	 */
2440 	for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(offset)]);
2441 	     dap; dap = LIST_NEXT(dap, da_pdlist))
2442 		if (dap->da_offset == offset)
2443 			break;
2444 	if (dap == NULL) {
2445 		for (dap = LIST_FIRST(&pagedep->pd_pendinghd);
2446 		     dap; dap = LIST_NEXT(dap, da_pdlist))
2447 			if (dap->da_offset == offset)
2448 				break;
2449 		if (dap == NULL)
2450 			return (dirrem);
2451 	}
2452 	/*
2453 	 * Must be ATTACHED at this point.
2454 	 */
2455 	if ((dap->da_state & ATTACHED) == 0)
2456 		panic("newdirrem: not ATTACHED");
2457 	if (dap->da_newinum != ip->i_number)
2458 		panic("newdirrem: inum %d should be %d",
2459 		    ip->i_number, dap->da_newinum);
2460 	/*
2461 	 * If we are deleting a changed name that never made it to disk,
2462 	 * then return the dirrem describing the previous inode (which
2463 	 * represents the inode currently referenced from this entry on disk).
2464 	 */
2465 	if ((dap->da_state & DIRCHG) != 0) {
2466 		*prevdirremp = dap->da_previous;
2467 		dap->da_state &= ~DIRCHG;
2468 		dap->da_pagedep = pagedep;
2469 	}
2470 	/*
2471 	 * We are deleting an entry that never made it to disk.
2472 	 * Mark it COMPLETE so we can delete its inode immediately.
2473 	 */
2474 	dirrem->dm_state |= COMPLETE;
2475 	free_diradd(dap);
2476 	return (dirrem);
2477 }
2478 
2479 /*
2480  * Directory entry change dependencies.
2481  *
2482  * Changing an existing directory entry requires that an add operation
2483  * be completed first followed by a deletion. The semantics for the addition
2484  * are identical to the description of adding a new entry above except
2485  * that the rollback is to the old inode number rather than zero. Once
2486  * the addition dependency is completed, the removal is done as described
2487  * in the removal routine above.
2488  */
2489 
2490 /*
2491  * This routine should be called immediately after changing
2492  * a directory entry.  The inode's link count should not be
2493  * decremented by the calling procedure -- the soft updates
2494  * code will perform this task when it is safe.
2495  */
2496 void
2497 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
2498 	struct buf *bp;		/* buffer containing directory block */
2499 	struct inode *dp;	/* inode for the directory being modified */
2500 	struct inode *ip;	/* inode for directory entry being removed */
2501 	long newinum;		/* new inode number for changed entry */
2502 	int isrmdir;		/* indicates if doing RMDIR */
2503 {
2504 	int offset;
2505 	struct diradd *dap = NULL;
2506 	struct dirrem *dirrem, *prevdirrem;
2507 	struct pagedep *pagedep;
2508 	struct inodedep *inodedep;
2509 
2510 	offset = blkoff(dp->i_fs, dp->i_offset);
2511 
2512 	/*
2513 	 * Whiteouts do not need diradd dependencies.
2514 	 */
2515 	if (newinum != WINO) {
2516 		MALLOC(dap, struct diradd *, sizeof(struct diradd),
2517 		    M_DIRADD, M_WAITOK);
2518 		bzero(dap, sizeof(struct diradd));
2519 		dap->da_list.wk_type = D_DIRADD;
2520 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2521 		dap->da_offset = offset;
2522 		dap->da_newinum = newinum;
2523 	}
2524 
2525 	/*
2526 	 * Allocate a new dirrem and ACQUIRE_LOCK.
2527 	 */
2528 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2529 	pagedep = dirrem->dm_pagedep;
2530 	/*
2531 	 * The possible values for isrmdir:
2532 	 *	0 - non-directory file rename
2533 	 *	1 - directory rename within same directory
2534 	 *   inum - directory rename to new directory of given inode number
2535 	 * When renaming to a new directory, we are both deleting and
2536 	 * creating a new directory entry, so the link count on the new
2537 	 * directory should not change. Thus we do not need the followup
2538 	 * dirrem which is usually done in handle_workitem_remove. We set
2539 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
2540 	 * followup dirrem.
2541 	 */
2542 	if (isrmdir > 1)
2543 		dirrem->dm_state |= DIRCHG;
2544 
2545 	/*
2546 	 * Whiteouts have no additional dependencies,
2547 	 * so just put the dirrem on the correct list.
2548 	 */
2549 	if (newinum == WINO) {
2550 		if ((dirrem->dm_state & COMPLETE) == 0) {
2551 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2552 			    dm_next);
2553 		} else {
2554 			dirrem->dm_dirinum = pagedep->pd_ino;
2555 			add_to_worklist(&dirrem->dm_list);
2556 		}
2557 		FREE_LOCK(&lk);
2558 		return;
2559 	}
2560 
2561 	/*
2562 	 * If the COMPLETE flag is clear, then there were no active
2563 	 * entries and we want to roll back to the previous inode until
2564 	 * the new inode is committed to disk. If the COMPLETE flag is
2565 	 * set, then we have deleted an entry that never made it to disk.
2566 	 * If the entry we deleted resulted from a name change, then the old
2567 	 * inode reference still resides on disk. Any rollback that we do
2568 	 * needs to be to that old inode (returned to us in prevdirrem). If
2569 	 * the entry we deleted resulted from a create, then there is
2570 	 * no entry on the disk, so we want to roll back to zero rather
2571 	 * than the uncommitted inode. In either of the COMPLETE cases we
2572 	 * want to immediately free the unwritten and unreferenced inode.
2573 	 */
2574 	if ((dirrem->dm_state & COMPLETE) == 0) {
2575 		dap->da_previous = dirrem;
2576 	} else {
2577 		if (prevdirrem != NULL) {
2578 			dap->da_previous = prevdirrem;
2579 		} else {
2580 			dap->da_state &= ~DIRCHG;
2581 			dap->da_pagedep = pagedep;
2582 		}
2583 		dirrem->dm_dirinum = pagedep->pd_ino;
2584 		add_to_worklist(&dirrem->dm_list);
2585 	}
2586 	/*
2587 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
2588 	 * is not yet written. If it is written, do the post-inode write
2589 	 * processing to put it on the id_pendinghd list.
2590 	 */
2591 	if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2592 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2593 		dap->da_state |= COMPLETE;
2594 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2595 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2596 	} else {
2597 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2598 		    dap, da_pdlist);
2599 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2600 	}
2601 	FREE_LOCK(&lk);
2602 }
2603 
2604 /*
2605  * Called whenever the link count on an inode is changed.
2606  * It creates an inode dependency so that the new reference(s)
2607  * to the inode cannot be committed to disk until the updated
2608  * inode has been written.
2609  */
2610 void
2611 softdep_change_linkcnt(ip)
2612 	struct inode *ip;	/* the inode with the increased link count */
2613 {
2614 	struct inodedep *inodedep;
2615 
2616 	ACQUIRE_LOCK(&lk);
2617 	(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2618 	if (ip->i_nlink < ip->i_effnlink)
2619 		panic("softdep_change_linkcnt: bad delta");
2620 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2621 	FREE_LOCK(&lk);
2622 }
2623 
2624 /*
2625  * This workitem decrements the inode's link count.
2626  * If the link count reaches zero, the file is removed.
2627  */
2628 static void
2629 handle_workitem_remove(dirrem)
2630 	struct dirrem *dirrem;
2631 {
2632 	struct proc *p = CURPROC;	/* XXX */
2633 	struct inodedep *inodedep;
2634 	struct vnode *vp;
2635 	struct inode *ip;
2636 	ino_t oldinum;
2637 	int error;
2638 
2639 	if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) {
2640 		softdep_error("handle_workitem_remove: vget", error);
2641 		return;
2642 	}
2643 	ip = VTOI(vp);
2644 	ACQUIRE_LOCK(&lk);
2645 	if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0)
2646 		panic("handle_workitem_remove: lost inodedep");
2647 	/*
2648 	 * Normal file deletion.
2649 	 */
2650 	if ((dirrem->dm_state & RMDIR) == 0) {
2651 		ip->i_nlink--;
2652 		ip->i_flag |= IN_CHANGE;
2653 		if (ip->i_nlink < ip->i_effnlink)
2654 			panic("handle_workitem_remove: bad file delta");
2655 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2656 		FREE_LOCK(&lk);
2657 		vput(vp);
2658 		num_dirrem -= 1;
2659 		WORKITEM_FREE(dirrem, D_DIRREM);
2660 		return;
2661 	}
2662 	/*
2663 	 * Directory deletion. Decrement reference count for both the
2664 	 * just deleted parent directory entry and the reference for ".".
2665 	 * Next truncate the directory to length zero. When the
2666 	 * truncation completes, arrange to have the reference count on
2667 	 * the parent decremented to account for the loss of "..".
2668 	 */
2669 	ip->i_nlink -= 2;
2670 	ip->i_flag |= IN_CHANGE;
2671 	if (ip->i_nlink < ip->i_effnlink)
2672 		panic("handle_workitem_remove: bad dir delta");
2673 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2674 	FREE_LOCK(&lk);
2675 	if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, p->p_ucred, p)) != 0)
2676 		softdep_error("handle_workitem_remove: truncate", error);
2677 	/*
2678 	 * Rename a directory to a new parent. Since, we are both deleting
2679 	 * and creating a new directory entry, the link count on the new
2680 	 * directory should not change. Thus we skip the followup dirrem.
2681 	 */
2682 	if (dirrem->dm_state & DIRCHG) {
2683 		vput(vp);
2684 		num_dirrem -= 1;
2685 		WORKITEM_FREE(dirrem, D_DIRREM);
2686 		return;
2687 	}
2688 	/*
2689 	 * If the inodedep does not exist, then the zero'ed inode has
2690 	 * been written to disk. If the allocated inode has never been
2691 	 * written to disk, then the on-disk inode is zero'ed. In either
2692 	 * case we can remove the file immediately.
2693 	 */
2694 	ACQUIRE_LOCK(&lk);
2695 	dirrem->dm_state = 0;
2696 	oldinum = dirrem->dm_oldinum;
2697 	dirrem->dm_oldinum = dirrem->dm_dirinum;
2698 	if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
2699 	    check_inode_unwritten(inodedep)) {
2700 		FREE_LOCK(&lk);
2701 		vput(vp);
2702 		handle_workitem_remove(dirrem);
2703 		return;
2704 	}
2705 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
2706 	FREE_LOCK(&lk);
2707 	vput(vp);
2708 }
2709 
2710 /*
2711  * Inode de-allocation dependencies.
2712  *
2713  * When an inode's link count is reduced to zero, it can be de-allocated. We
2714  * found it convenient to postpone de-allocation until after the inode is
2715  * written to disk with its new link count (zero).  At this point, all of the
2716  * on-disk inode's block pointers are nullified and, with careful dependency
2717  * list ordering, all dependencies related to the inode will be satisfied and
2718  * the corresponding dependency structures de-allocated.  So, if/when the
2719  * inode is reused, there will be no mixing of old dependencies with new
2720  * ones.  This artificial dependency is set up by the block de-allocation
2721  * procedure above (softdep_setup_freeblocks) and completed by the
2722  * following procedure.
2723  */
2724 static void
2725 handle_workitem_freefile(freefile)
2726 	struct freefile *freefile;
2727 {
2728 	struct fs *fs;
2729 	struct vnode vp;
2730 	struct inode tip;
2731 	struct inodedep *idp;
2732 	int error;
2733 
2734 	fs = VFSTOUFS(freefile->fx_mnt)->um_fs;
2735 #ifdef DEBUG
2736 	ACQUIRE_LOCK(&lk);
2737 	if (inodedep_lookup(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 = fs;
2744 	tip.i_vnode = &vp;
2745 	vp.v_data = &tip;
2746 	if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
2747 		softdep_error("handle_workitem_freefile", error);
2748 	WORKITEM_FREE(freefile, D_FREEFILE);
2749 }
2750 
2751 /*
2752  * Disk writes.
2753  *
2754  * The dependency structures constructed above are most actively used when file
2755  * system blocks are written to disk.  No constraints are placed on when a
2756  * block can be written, but unsatisfied update dependencies are made safe by
2757  * modifying (or replacing) the source memory for the duration of the disk
2758  * write.  When the disk write completes, the memory block is again brought
2759  * up-to-date.
2760  *
2761  * In-core inode structure reclamation.
2762  *
2763  * Because there are a finite number of "in-core" inode structures, they are
2764  * reused regularly.  By transferring all inode-related dependencies to the
2765  * in-memory inode block and indexing them separately (via "inodedep"s), we
2766  * can allow "in-core" inode structures to be reused at any time and avoid
2767  * any increase in contention.
2768  *
2769  * Called just before entering the device driver to initiate a new disk I/O.
2770  * The buffer must be locked, thus, no I/O completion operations can occur
2771  * while we are manipulating its associated dependencies.
2772  */
2773 static void
2774 softdep_disk_io_initiation(bp)
2775 	struct buf *bp;		/* structure describing disk write to occur */
2776 {
2777 	struct worklist *wk, *nextwk;
2778 	struct indirdep *indirdep;
2779 
2780 	/*
2781 	 * We only care about write operations. There should never
2782 	 * be dependencies for reads.
2783 	 */
2784 	if (bp->b_iocmd == BIO_READ)
2785 		panic("softdep_disk_io_initiation: read");
2786 	/*
2787 	 * Do any necessary pre-I/O processing.
2788 	 */
2789 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
2790 		nextwk = LIST_NEXT(wk, wk_list);
2791 		switch (wk->wk_type) {
2792 
2793 		case D_PAGEDEP:
2794 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
2795 			continue;
2796 
2797 		case D_INODEDEP:
2798 			initiate_write_inodeblock(WK_INODEDEP(wk), bp);
2799 			continue;
2800 
2801 		case D_INDIRDEP:
2802 			indirdep = WK_INDIRDEP(wk);
2803 			if (indirdep->ir_state & GOINGAWAY)
2804 				panic("disk_io_initiation: indirdep gone");
2805 			/*
2806 			 * If there are no remaining dependencies, this
2807 			 * will be writing the real pointers, so the
2808 			 * dependency can be freed.
2809 			 */
2810 			if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
2811 				indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
2812 				brelse(indirdep->ir_savebp);
2813 				/* inline expand WORKLIST_REMOVE(wk); */
2814 				wk->wk_state &= ~ONWORKLIST;
2815 				LIST_REMOVE(wk, wk_list);
2816 				WORKITEM_FREE(indirdep, D_INDIRDEP);
2817 				continue;
2818 			}
2819 			/*
2820 			 * Replace up-to-date version with safe version.
2821 			 */
2822 			ACQUIRE_LOCK(&lk);
2823 			indirdep->ir_state &= ~ATTACHED;
2824 			indirdep->ir_state |= UNDONE;
2825 			MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
2826 			    M_INDIRDEP, M_WAITOK);
2827 			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
2828 			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
2829 			    bp->b_bcount);
2830 			FREE_LOCK(&lk);
2831 			continue;
2832 
2833 		case D_MKDIR:
2834 		case D_BMSAFEMAP:
2835 		case D_ALLOCDIRECT:
2836 		case D_ALLOCINDIR:
2837 			continue;
2838 
2839 		default:
2840 			panic("handle_disk_io_initiation: Unexpected type %s",
2841 			    TYPENAME(wk->wk_type));
2842 			/* NOTREACHED */
2843 		}
2844 	}
2845 }
2846 
2847 /*
2848  * Called from within the procedure above to deal with unsatisfied
2849  * allocation dependencies in a directory. The buffer must be locked,
2850  * thus, no I/O completion operations can occur while we are
2851  * manipulating its associated dependencies.
2852  */
2853 static void
2854 initiate_write_filepage(pagedep, bp)
2855 	struct pagedep *pagedep;
2856 	struct buf *bp;
2857 {
2858 	struct diradd *dap;
2859 	struct direct *ep;
2860 	int i;
2861 
2862 	if (pagedep->pd_state & IOSTARTED) {
2863 		/*
2864 		 * This can only happen if there is a driver that does not
2865 		 * understand chaining. Here biodone will reissue the call
2866 		 * to strategy for the incomplete buffers.
2867 		 */
2868 		printf("initiate_write_filepage: already started\n");
2869 		return;
2870 	}
2871 	pagedep->pd_state |= IOSTARTED;
2872 	ACQUIRE_LOCK(&lk);
2873 	for (i = 0; i < DAHASHSZ; i++) {
2874 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
2875 		     dap = LIST_NEXT(dap, da_pdlist)) {
2876 			ep = (struct direct *)
2877 			    ((char *)bp->b_data + dap->da_offset);
2878 			if (ep->d_ino != dap->da_newinum)
2879 				panic("%s: dir inum %d != new %d",
2880 				    "initiate_write_filepage",
2881 				    ep->d_ino, dap->da_newinum);
2882 			if (dap->da_state & DIRCHG)
2883 				ep->d_ino = dap->da_previous->dm_oldinum;
2884 			else
2885 				ep->d_ino = 0;
2886 			dap->da_state &= ~ATTACHED;
2887 			dap->da_state |= UNDONE;
2888 		}
2889 	}
2890 	FREE_LOCK(&lk);
2891 }
2892 
2893 /*
2894  * Called from within the procedure above to deal with unsatisfied
2895  * allocation dependencies in an inodeblock. The buffer must be
2896  * locked, thus, no I/O completion operations can occur while we
2897  * are manipulating its associated dependencies.
2898  */
2899 static void
2900 initiate_write_inodeblock(inodedep, bp)
2901 	struct inodedep *inodedep;
2902 	struct buf *bp;			/* The inode block */
2903 {
2904 	struct allocdirect *adp, *lastadp;
2905 	struct dinode *dp;
2906 	struct fs *fs;
2907 	ufs_lbn_t prevlbn = 0;
2908 	int i, deplist;
2909 
2910 	if (inodedep->id_state & IOSTARTED)
2911 		panic("initiate_write_inodeblock: already started");
2912 	inodedep->id_state |= IOSTARTED;
2913 	fs = inodedep->id_fs;
2914 	dp = (struct dinode *)bp->b_data +
2915 	    ino_to_fsbo(fs, inodedep->id_ino);
2916 	/*
2917 	 * If the bitmap is not yet written, then the allocated
2918 	 * inode cannot be written to disk.
2919 	 */
2920 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
2921 		if (inodedep->id_savedino != NULL)
2922 			panic("initiate_write_inodeblock: already doing I/O");
2923 		MALLOC(inodedep->id_savedino, struct dinode *,
2924 		    sizeof(struct dinode), M_INODEDEP, M_WAITOK);
2925 		*inodedep->id_savedino = *dp;
2926 		bzero((caddr_t)dp, sizeof(struct dinode));
2927 		return;
2928 	}
2929 	/*
2930 	 * If no dependencies, then there is nothing to roll back.
2931 	 */
2932 	inodedep->id_savedsize = dp->di_size;
2933 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
2934 		return;
2935 	/*
2936 	 * Set the dependencies to busy.
2937 	 */
2938 	ACQUIRE_LOCK(&lk);
2939 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
2940 	     adp = TAILQ_NEXT(adp, ad_next)) {
2941 #ifdef DIAGNOSTIC
2942 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
2943 			panic("softdep_write_inodeblock: lbn order");
2944 		prevlbn = adp->ad_lbn;
2945 		if (adp->ad_lbn < NDADDR &&
2946 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno)
2947 			panic("%s: direct pointer #%ld mismatch %d != %d",
2948 			    "softdep_write_inodeblock", adp->ad_lbn,
2949 			    dp->di_db[adp->ad_lbn], adp->ad_newblkno);
2950 		if (adp->ad_lbn >= NDADDR &&
2951 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno)
2952 			panic("%s: indirect pointer #%ld mismatch %d != %d",
2953 			    "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
2954 			    dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
2955 		deplist |= 1 << adp->ad_lbn;
2956 		if ((adp->ad_state & ATTACHED) == 0)
2957 			panic("softdep_write_inodeblock: Unknown state 0x%x",
2958 			    adp->ad_state);
2959 #endif /* DIAGNOSTIC */
2960 		adp->ad_state &= ~ATTACHED;
2961 		adp->ad_state |= UNDONE;
2962 	}
2963 	/*
2964 	 * The on-disk inode cannot claim to be any larger than the last
2965 	 * fragment that has been written. Otherwise, the on-disk inode
2966 	 * might have fragments that were not the last block in the file
2967 	 * which would corrupt the filesystem.
2968 	 */
2969 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
2970 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
2971 		if (adp->ad_lbn >= NDADDR)
2972 			break;
2973 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
2974 		/* keep going until hitting a rollback to a frag */
2975 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
2976 			continue;
2977 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
2978 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
2979 #ifdef DIAGNOSTIC
2980 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
2981 				panic("softdep_write_inodeblock: lost dep1");
2982 #endif /* DIAGNOSTIC */
2983 			dp->di_db[i] = 0;
2984 		}
2985 		for (i = 0; i < NIADDR; i++) {
2986 #ifdef DIAGNOSTIC
2987 			if (dp->di_ib[i] != 0 &&
2988 			    (deplist & ((1 << NDADDR) << i)) == 0)
2989 				panic("softdep_write_inodeblock: lost dep2");
2990 #endif /* DIAGNOSTIC */
2991 			dp->di_ib[i] = 0;
2992 		}
2993 		FREE_LOCK(&lk);
2994 		return;
2995 	}
2996 	/*
2997 	 * If we have zero'ed out the last allocated block of the file,
2998 	 * roll back the size to the last currently allocated block.
2999 	 * We know that this last allocated block is a full-sized as
3000 	 * we already checked for fragments in the loop above.
3001 	 */
3002 	if (lastadp != NULL &&
3003 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3004 		for (i = lastadp->ad_lbn; i >= 0; i--)
3005 			if (dp->di_db[i] != 0)
3006 				break;
3007 		dp->di_size = (i + 1) * fs->fs_bsize;
3008 	}
3009 	/*
3010 	 * The only dependencies are for indirect blocks.
3011 	 *
3012 	 * The file size for indirect block additions is not guaranteed.
3013 	 * Such a guarantee would be non-trivial to achieve. The conventional
3014 	 * synchronous write implementation also does not make this guarantee.
3015 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3016 	 * can be over-estimated without destroying integrity when the file
3017 	 * moves into the indirect blocks (i.e., is large). If we want to
3018 	 * postpone fsck, we are stuck with this argument.
3019 	 */
3020 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3021 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3022 	FREE_LOCK(&lk);
3023 }
3024 
3025 /*
3026  * This routine is called during the completion interrupt
3027  * service routine for a disk write (from the procedure called
3028  * by the device driver to inform the file system caches of
3029  * a request completion).  It should be called early in this
3030  * procedure, before the block is made available to other
3031  * processes or other routines are called.
3032  */
3033 static void
3034 softdep_disk_write_complete(bp)
3035 	struct buf *bp;		/* describes the completed disk write */
3036 {
3037 	struct worklist *wk;
3038 	struct workhead reattach;
3039 	struct newblk *newblk;
3040 	struct allocindir *aip;
3041 	struct allocdirect *adp;
3042 	struct indirdep *indirdep;
3043 	struct inodedep *inodedep;
3044 	struct bmsafemap *bmsafemap;
3045 
3046 #ifdef DEBUG
3047 	if (lk.lkt_held != -1)
3048 		panic("softdep_disk_write_complete: lock is held");
3049 	lk.lkt_held = -2;
3050 #endif
3051 	LIST_INIT(&reattach);
3052 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3053 		WORKLIST_REMOVE(wk);
3054 		switch (wk->wk_type) {
3055 
3056 		case D_PAGEDEP:
3057 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3058 				WORKLIST_INSERT(&reattach, wk);
3059 			continue;
3060 
3061 		case D_INODEDEP:
3062 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3063 				WORKLIST_INSERT(&reattach, wk);
3064 			continue;
3065 
3066 		case D_BMSAFEMAP:
3067 			bmsafemap = WK_BMSAFEMAP(wk);
3068 			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3069 				newblk->nb_state |= DEPCOMPLETE;
3070 				newblk->nb_bmsafemap = NULL;
3071 				LIST_REMOVE(newblk, nb_deps);
3072 			}
3073 			while ((adp =
3074 			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3075 				adp->ad_state |= DEPCOMPLETE;
3076 				adp->ad_buf = NULL;
3077 				LIST_REMOVE(adp, ad_deps);
3078 				handle_allocdirect_partdone(adp);
3079 			}
3080 			while ((aip =
3081 			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3082 				aip->ai_state |= DEPCOMPLETE;
3083 				aip->ai_buf = NULL;
3084 				LIST_REMOVE(aip, ai_deps);
3085 				handle_allocindir_partdone(aip);
3086 			}
3087 			while ((inodedep =
3088 			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3089 				inodedep->id_state |= DEPCOMPLETE;
3090 				LIST_REMOVE(inodedep, id_deps);
3091 				inodedep->id_buf = NULL;
3092 			}
3093 			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3094 			continue;
3095 
3096 		case D_MKDIR:
3097 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3098 			continue;
3099 
3100 		case D_ALLOCDIRECT:
3101 			adp = WK_ALLOCDIRECT(wk);
3102 			adp->ad_state |= COMPLETE;
3103 			handle_allocdirect_partdone(adp);
3104 			continue;
3105 
3106 		case D_ALLOCINDIR:
3107 			aip = WK_ALLOCINDIR(wk);
3108 			aip->ai_state |= COMPLETE;
3109 			handle_allocindir_partdone(aip);
3110 			continue;
3111 
3112 		case D_INDIRDEP:
3113 			indirdep = WK_INDIRDEP(wk);
3114 			if (indirdep->ir_state & GOINGAWAY)
3115 				panic("disk_write_complete: indirdep gone");
3116 			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3117 			FREE(indirdep->ir_saveddata, M_INDIRDEP);
3118 			indirdep->ir_saveddata = 0;
3119 			indirdep->ir_state &= ~UNDONE;
3120 			indirdep->ir_state |= ATTACHED;
3121 			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3122 				handle_allocindir_partdone(aip);
3123 				if (aip == LIST_FIRST(&indirdep->ir_donehd))
3124 					panic("disk_write_complete: not gone");
3125 			}
3126 			WORKLIST_INSERT(&reattach, wk);
3127 			if ((bp->b_flags & B_DELWRI) == 0)
3128 				stat_indir_blk_ptrs++;
3129 			bdirty(bp);
3130 			continue;
3131 
3132 		default:
3133 			panic("handle_disk_write_complete: Unknown type %s",
3134 			    TYPENAME(wk->wk_type));
3135 			/* NOTREACHED */
3136 		}
3137 	}
3138 	/*
3139 	 * Reattach any requests that must be redone.
3140 	 */
3141 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
3142 		WORKLIST_REMOVE(wk);
3143 		WORKLIST_INSERT(&bp->b_dep, wk);
3144 	}
3145 #ifdef DEBUG
3146 	if (lk.lkt_held != -2)
3147 		panic("softdep_disk_write_complete: lock lost");
3148 	lk.lkt_held = -1;
3149 #endif
3150 }
3151 
3152 /*
3153  * Called from within softdep_disk_write_complete above. Note that
3154  * this routine is always called from interrupt level with further
3155  * splbio interrupts blocked.
3156  */
3157 static void
3158 handle_allocdirect_partdone(adp)
3159 	struct allocdirect *adp;	/* the completed allocdirect */
3160 {
3161 	struct allocdirect *listadp;
3162 	struct inodedep *inodedep;
3163 	long bsize, delay;
3164 
3165 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3166 		return;
3167 	if (adp->ad_buf != NULL)
3168 		panic("handle_allocdirect_partdone: dangling dep");
3169 	/*
3170 	 * The on-disk inode cannot claim to be any larger than the last
3171 	 * fragment that has been written. Otherwise, the on-disk inode
3172 	 * might have fragments that were not the last block in the file
3173 	 * which would corrupt the filesystem. Thus, we cannot free any
3174 	 * allocdirects after one whose ad_oldblkno claims a fragment as
3175 	 * these blocks must be rolled back to zero before writing the inode.
3176 	 * We check the currently active set of allocdirects in id_inoupdt.
3177 	 */
3178 	inodedep = adp->ad_inodedep;
3179 	bsize = inodedep->id_fs->fs_bsize;
3180 	for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp;
3181 	     listadp = TAILQ_NEXT(listadp, ad_next)) {
3182 		/* found our block */
3183 		if (listadp == adp)
3184 			break;
3185 		/* continue if ad_oldlbn is not a fragment */
3186 		if (listadp->ad_oldsize == 0 ||
3187 		    listadp->ad_oldsize == bsize)
3188 			continue;
3189 		/* hit a fragment */
3190 		return;
3191 	}
3192 	/*
3193 	 * If we have reached the end of the current list without
3194 	 * finding the just finished dependency, then it must be
3195 	 * on the future dependency list. Future dependencies cannot
3196 	 * be freed until they are moved to the current list.
3197 	 */
3198 	if (listadp == NULL) {
3199 #ifdef DEBUG
3200 		for (listadp = TAILQ_FIRST(&inodedep->id_newinoupdt); listadp;
3201 		     listadp = TAILQ_NEXT(listadp, ad_next))
3202 			/* found our block */
3203 			if (listadp == adp)
3204 				break;
3205 		if (listadp == NULL)
3206 			panic("handle_allocdirect_partdone: lost dep");
3207 #endif /* DEBUG */
3208 		return;
3209 	}
3210 	/*
3211 	 * If we have found the just finished dependency, then free
3212 	 * it along with anything that follows it that is complete.
3213 	 * If the inode still has a bitmap dependency, then it has
3214 	 * never been written to disk, hence the on-disk inode cannot
3215 	 * reference the old fragment so we can free it without delay.
3216 	 */
3217 	delay = (inodedep->id_state & DEPCOMPLETE);
3218 	for (; adp; adp = listadp) {
3219 		listadp = TAILQ_NEXT(adp, ad_next);
3220 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3221 			return;
3222 		free_allocdirect(&inodedep->id_inoupdt, adp, delay);
3223 	}
3224 }
3225 
3226 /*
3227  * Called from within softdep_disk_write_complete above. Note that
3228  * this routine is always called from interrupt level with further
3229  * splbio interrupts blocked.
3230  */
3231 static void
3232 handle_allocindir_partdone(aip)
3233 	struct allocindir *aip;		/* the completed allocindir */
3234 {
3235 	struct indirdep *indirdep;
3236 
3237 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
3238 		return;
3239 	if (aip->ai_buf != NULL)
3240 		panic("handle_allocindir_partdone: dangling dependency");
3241 	indirdep = aip->ai_indirdep;
3242 	if (indirdep->ir_state & UNDONE) {
3243 		LIST_REMOVE(aip, ai_next);
3244 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
3245 		return;
3246 	}
3247 	((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
3248 	    aip->ai_newblkno;
3249 	LIST_REMOVE(aip, ai_next);
3250 	if (aip->ai_freefrag != NULL)
3251 		add_to_worklist(&aip->ai_freefrag->ff_list);
3252 	WORKITEM_FREE(aip, D_ALLOCINDIR);
3253 }
3254 
3255 /*
3256  * Called from within softdep_disk_write_complete above to restore
3257  * in-memory inode block contents to their most up-to-date state. Note
3258  * that this routine is always called from interrupt level with further
3259  * splbio interrupts blocked.
3260  */
3261 static int
3262 handle_written_inodeblock(inodedep, bp)
3263 	struct inodedep *inodedep;
3264 	struct buf *bp;		/* buffer containing the inode block */
3265 {
3266 	struct worklist *wk, *filefree;
3267 	struct allocdirect *adp, *nextadp;
3268 	struct dinode *dp;
3269 	int hadchanges;
3270 
3271 	if ((inodedep->id_state & IOSTARTED) == 0)
3272 		panic("handle_written_inodeblock: not started");
3273 	inodedep->id_state &= ~IOSTARTED;
3274 	inodedep->id_state |= COMPLETE;
3275 	dp = (struct dinode *)bp->b_data +
3276 	    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
3277 	/*
3278 	 * If we had to rollback the inode allocation because of
3279 	 * bitmaps being incomplete, then simply restore it.
3280 	 * Keep the block dirty so that it will not be reclaimed until
3281 	 * all associated dependencies have been cleared and the
3282 	 * corresponding updates written to disk.
3283 	 */
3284 	if (inodedep->id_savedino != NULL) {
3285 		*dp = *inodedep->id_savedino;
3286 		FREE(inodedep->id_savedino, M_INODEDEP);
3287 		inodedep->id_savedino = NULL;
3288 		if ((bp->b_flags & B_DELWRI) == 0)
3289 			stat_inode_bitmap++;
3290 		bdirty(bp);
3291 		return (1);
3292 	}
3293 	/*
3294 	 * Roll forward anything that had to be rolled back before
3295 	 * the inode could be updated.
3296 	 */
3297 	hadchanges = 0;
3298 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
3299 		nextadp = TAILQ_NEXT(adp, ad_next);
3300 		if (adp->ad_state & ATTACHED)
3301 			panic("handle_written_inodeblock: new entry");
3302 		if (adp->ad_lbn < NDADDR) {
3303 			if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno)
3304 				panic("%s: %s #%ld mismatch %d != %d",
3305 				    "handle_written_inodeblock",
3306 				    "direct pointer", adp->ad_lbn,
3307 				    dp->di_db[adp->ad_lbn], adp->ad_oldblkno);
3308 			dp->di_db[adp->ad_lbn] = adp->ad_newblkno;
3309 		} else {
3310 			if (dp->di_ib[adp->ad_lbn - NDADDR] != 0)
3311 				panic("%s: %s #%ld allocated as %d",
3312 				    "handle_written_inodeblock",
3313 				    "indirect pointer", adp->ad_lbn - NDADDR,
3314 				    dp->di_ib[adp->ad_lbn - NDADDR]);
3315 			dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno;
3316 		}
3317 		adp->ad_state &= ~UNDONE;
3318 		adp->ad_state |= ATTACHED;
3319 		hadchanges = 1;
3320 	}
3321 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
3322 		stat_direct_blk_ptrs++;
3323 	/*
3324 	 * Reset the file size to its most up-to-date value.
3325 	 */
3326 	if (inodedep->id_savedsize == -1)
3327 		panic("handle_written_inodeblock: bad size");
3328 	if (dp->di_size != inodedep->id_savedsize) {
3329 		dp->di_size = inodedep->id_savedsize;
3330 		hadchanges = 1;
3331 	}
3332 	inodedep->id_savedsize = -1;
3333 	/*
3334 	 * If there were any rollbacks in the inode block, then it must be
3335 	 * marked dirty so that its will eventually get written back in
3336 	 * its correct form.
3337 	 */
3338 	if (hadchanges)
3339 		bdirty(bp);
3340 	/*
3341 	 * Process any allocdirects that completed during the update.
3342 	 */
3343 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
3344 		handle_allocdirect_partdone(adp);
3345 	/*
3346 	 * Process deallocations that were held pending until the
3347 	 * inode had been written to disk. Freeing of the inode
3348 	 * is delayed until after all blocks have been freed to
3349 	 * avoid creation of new <vfsid, inum, lbn> triples
3350 	 * before the old ones have been deleted.
3351 	 */
3352 	filefree = NULL;
3353 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
3354 		WORKLIST_REMOVE(wk);
3355 		switch (wk->wk_type) {
3356 
3357 		case D_FREEFILE:
3358 			/*
3359 			 * We defer adding filefree to the worklist until
3360 			 * all other additions have been made to ensure
3361 			 * that it will be done after all the old blocks
3362 			 * have been freed.
3363 			 */
3364 			if (filefree != NULL)
3365 				panic("handle_written_inodeblock: filefree");
3366 			filefree = wk;
3367 			continue;
3368 
3369 		case D_MKDIR:
3370 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
3371 			continue;
3372 
3373 		case D_DIRADD:
3374 			diradd_inode_written(WK_DIRADD(wk), inodedep);
3375 			continue;
3376 
3377 		case D_FREEBLKS:
3378 		case D_FREEFRAG:
3379 		case D_DIRREM:
3380 			add_to_worklist(wk);
3381 			continue;
3382 
3383 		default:
3384 			panic("handle_written_inodeblock: Unknown type %s",
3385 			    TYPENAME(wk->wk_type));
3386 			/* NOTREACHED */
3387 		}
3388 	}
3389 	if (filefree != NULL) {
3390 		if (free_inodedep(inodedep) == 0)
3391 			panic("handle_written_inodeblock: live inodedep");
3392 		add_to_worklist(filefree);
3393 		return (0);
3394 	}
3395 
3396 	/*
3397 	 * If no outstanding dependencies, free it.
3398 	 */
3399 	if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0)
3400 		return (0);
3401 	return (hadchanges);
3402 }
3403 
3404 /*
3405  * Process a diradd entry after its dependent inode has been written.
3406  * This routine must be called with splbio interrupts blocked.
3407  */
3408 static void
3409 diradd_inode_written(dap, inodedep)
3410 	struct diradd *dap;
3411 	struct inodedep *inodedep;
3412 {
3413 	struct pagedep *pagedep;
3414 
3415 	dap->da_state |= COMPLETE;
3416 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3417 		if (dap->da_state & DIRCHG)
3418 			pagedep = dap->da_previous->dm_pagedep;
3419 		else
3420 			pagedep = dap->da_pagedep;
3421 		LIST_REMOVE(dap, da_pdlist);
3422 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3423 	}
3424 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3425 }
3426 
3427 /*
3428  * Handle the completion of a mkdir dependency.
3429  */
3430 static void
3431 handle_written_mkdir(mkdir, type)
3432 	struct mkdir *mkdir;
3433 	int type;
3434 {
3435 	struct diradd *dap;
3436 	struct pagedep *pagedep;
3437 
3438 	if (mkdir->md_state != type)
3439 		panic("handle_written_mkdir: bad type");
3440 	dap = mkdir->md_diradd;
3441 	dap->da_state &= ~type;
3442 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
3443 		dap->da_state |= DEPCOMPLETE;
3444 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3445 		if (dap->da_state & DIRCHG)
3446 			pagedep = dap->da_previous->dm_pagedep;
3447 		else
3448 			pagedep = dap->da_pagedep;
3449 		LIST_REMOVE(dap, da_pdlist);
3450 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3451 	}
3452 	LIST_REMOVE(mkdir, md_mkdirs);
3453 	WORKITEM_FREE(mkdir, D_MKDIR);
3454 }
3455 
3456 /*
3457  * Called from within softdep_disk_write_complete above.
3458  * A write operation was just completed. Removed inodes can
3459  * now be freed and associated block pointers may be committed.
3460  * Note that this routine is always called from interrupt level
3461  * with further splbio interrupts blocked.
3462  */
3463 static int
3464 handle_written_filepage(pagedep, bp)
3465 	struct pagedep *pagedep;
3466 	struct buf *bp;		/* buffer containing the written page */
3467 {
3468 	struct dirrem *dirrem;
3469 	struct diradd *dap, *nextdap;
3470 	struct direct *ep;
3471 	int i, chgs;
3472 
3473 	if ((pagedep->pd_state & IOSTARTED) == 0)
3474 		panic("handle_written_filepage: not started");
3475 	pagedep->pd_state &= ~IOSTARTED;
3476 	/*
3477 	 * Process any directory removals that have been committed.
3478 	 */
3479 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
3480 		LIST_REMOVE(dirrem, dm_next);
3481 		dirrem->dm_dirinum = pagedep->pd_ino;
3482 		add_to_worklist(&dirrem->dm_list);
3483 	}
3484 	/*
3485 	 * Free any directory additions that have been committed.
3486 	 */
3487 	while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
3488 		free_diradd(dap);
3489 	/*
3490 	 * Uncommitted directory entries must be restored.
3491 	 */
3492 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
3493 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
3494 		     dap = nextdap) {
3495 			nextdap = LIST_NEXT(dap, da_pdlist);
3496 			if (dap->da_state & ATTACHED)
3497 				panic("handle_written_filepage: attached");
3498 			ep = (struct direct *)
3499 			    ((char *)bp->b_data + dap->da_offset);
3500 			ep->d_ino = dap->da_newinum;
3501 			dap->da_state &= ~UNDONE;
3502 			dap->da_state |= ATTACHED;
3503 			chgs = 1;
3504 			/*
3505 			 * If the inode referenced by the directory has
3506 			 * been written out, then the dependency can be
3507 			 * moved to the pending list.
3508 			 */
3509 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3510 				LIST_REMOVE(dap, da_pdlist);
3511 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
3512 				    da_pdlist);
3513 			}
3514 		}
3515 	}
3516 	/*
3517 	 * If there were any rollbacks in the directory, then it must be
3518 	 * marked dirty so that its will eventually get written back in
3519 	 * its correct form.
3520 	 */
3521 	if (chgs) {
3522 		if ((bp->b_flags & B_DELWRI) == 0)
3523 			stat_dir_entry++;
3524 		bdirty(bp);
3525 	}
3526 	/*
3527 	 * If no dependencies remain, the pagedep will be freed.
3528 	 * Otherwise it will remain to update the page before it
3529 	 * is written back to disk.
3530 	 */
3531 	if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) {
3532 		for (i = 0; i < DAHASHSZ; i++)
3533 			if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
3534 				break;
3535 		if (i == DAHASHSZ) {
3536 			LIST_REMOVE(pagedep, pd_hash);
3537 			WORKITEM_FREE(pagedep, D_PAGEDEP);
3538 			return (0);
3539 		}
3540 	}
3541 	return (1);
3542 }
3543 
3544 /*
3545  * Writing back in-core inode structures.
3546  *
3547  * The file system only accesses an inode's contents when it occupies an
3548  * "in-core" inode structure.  These "in-core" structures are separate from
3549  * the page frames used to cache inode blocks.  Only the latter are
3550  * transferred to/from the disk.  So, when the updated contents of the
3551  * "in-core" inode structure are copied to the corresponding in-memory inode
3552  * block, the dependencies are also transferred.  The following procedure is
3553  * called when copying a dirty "in-core" inode to a cached inode block.
3554  */
3555 
3556 /*
3557  * Called when an inode is loaded from disk. If the effective link count
3558  * differed from the actual link count when it was last flushed, then we
3559  * need to ensure that the correct effective link count is put back.
3560  */
3561 void
3562 softdep_load_inodeblock(ip)
3563 	struct inode *ip;	/* the "in_core" copy of the inode */
3564 {
3565 	struct inodedep *inodedep;
3566 
3567 	/*
3568 	 * Check for alternate nlink count.
3569 	 */
3570 	ip->i_effnlink = ip->i_nlink;
3571 	ACQUIRE_LOCK(&lk);
3572 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3573 		FREE_LOCK(&lk);
3574 		return;
3575 	}
3576 	ip->i_effnlink -= inodedep->id_nlinkdelta;
3577 	FREE_LOCK(&lk);
3578 }
3579 
3580 /*
3581  * This routine is called just before the "in-core" inode
3582  * information is to be copied to the in-memory inode block.
3583  * Recall that an inode block contains several inodes. If
3584  * the force flag is set, then the dependencies will be
3585  * cleared so that the update can always be made. Note that
3586  * the buffer is locked when this routine is called, so we
3587  * will never be in the middle of writing the inode block
3588  * to disk.
3589  */
3590 void
3591 softdep_update_inodeblock(ip, bp, waitfor)
3592 	struct inode *ip;	/* the "in_core" copy of the inode */
3593 	struct buf *bp;		/* the buffer containing the inode block */
3594 	int waitfor;		/* nonzero => update must be allowed */
3595 {
3596 	struct inodedep *inodedep;
3597 	struct worklist *wk;
3598 	int error, gotit;
3599 
3600 	/*
3601 	 * If the effective link count is not equal to the actual link
3602 	 * count, then we must track the difference in an inodedep while
3603 	 * the inode is (potentially) tossed out of the cache. Otherwise,
3604 	 * if there is no existing inodedep, then there are no dependencies
3605 	 * to track.
3606 	 */
3607 	ACQUIRE_LOCK(&lk);
3608 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3609 		if (ip->i_effnlink != ip->i_nlink)
3610 			panic("softdep_update_inodeblock: bad link count");
3611 		FREE_LOCK(&lk);
3612 		return;
3613 	}
3614 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
3615 		panic("softdep_update_inodeblock: bad delta");
3616 	/*
3617 	 * Changes have been initiated. Anything depending on these
3618 	 * changes cannot occur until this inode has been written.
3619 	 */
3620 	inodedep->id_state &= ~COMPLETE;
3621 	if ((inodedep->id_state & ONWORKLIST) == 0)
3622 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
3623 	/*
3624 	 * Any new dependencies associated with the incore inode must
3625 	 * now be moved to the list associated with the buffer holding
3626 	 * the in-memory copy of the inode. Once merged process any
3627 	 * allocdirects that are completed by the merger.
3628 	 */
3629 	merge_inode_lists(inodedep);
3630 	if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
3631 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
3632 	/*
3633 	 * Now that the inode has been pushed into the buffer, the
3634 	 * operations dependent on the inode being written to disk
3635 	 * can be moved to the id_bufwait so that they will be
3636 	 * processed when the buffer I/O completes.
3637 	 */
3638 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
3639 		WORKLIST_REMOVE(wk);
3640 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
3641 	}
3642 	/*
3643 	 * Newly allocated inodes cannot be written until the bitmap
3644 	 * that allocates them have been written (indicated by
3645 	 * DEPCOMPLETE being set in id_state). If we are doing a
3646 	 * forced sync (e.g., an fsync on a file), we force the bitmap
3647 	 * to be written so that the update can be done.
3648 	 */
3649 	if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
3650 		FREE_LOCK(&lk);
3651 		return;
3652 	}
3653 	gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
3654 	FREE_LOCK(&lk);
3655 	if (gotit &&
3656 	    (error = BUF_WRITE(inodedep->id_buf)) != 0)
3657 		softdep_error("softdep_update_inodeblock: bwrite", error);
3658 	if ((inodedep->id_state & DEPCOMPLETE) == 0)
3659 		panic("softdep_update_inodeblock: update failed");
3660 }
3661 
3662 /*
3663  * Merge the new inode dependency list (id_newinoupdt) into the old
3664  * inode dependency list (id_inoupdt). This routine must be called
3665  * with splbio interrupts blocked.
3666  */
3667 static void
3668 merge_inode_lists(inodedep)
3669 	struct inodedep *inodedep;
3670 {
3671 	struct allocdirect *listadp, *newadp;
3672 
3673 	newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3674 	for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) {
3675 		if (listadp->ad_lbn < newadp->ad_lbn) {
3676 			listadp = TAILQ_NEXT(listadp, ad_next);
3677 			continue;
3678 		}
3679 		TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3680 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
3681 		if (listadp->ad_lbn == newadp->ad_lbn) {
3682 			allocdirect_merge(&inodedep->id_inoupdt, newadp,
3683 			    listadp);
3684 			listadp = newadp;
3685 		}
3686 		newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3687 	}
3688 	while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) {
3689 		TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3690 		TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next);
3691 	}
3692 }
3693 
3694 /*
3695  * If we are doing an fsync, then we must ensure that any directory
3696  * entries for the inode have been written after the inode gets to disk.
3697  */
3698 int
3699 softdep_fsync(vp)
3700 	struct vnode *vp;	/* the "in_core" copy of the inode */
3701 {
3702 	struct inodedep *inodedep;
3703 	struct pagedep *pagedep;
3704 	struct worklist *wk;
3705 	struct diradd *dap;
3706 	struct mount *mnt;
3707 	struct vnode *pvp;
3708 	struct inode *ip;
3709 	struct buf *bp;
3710 	struct fs *fs;
3711 	struct proc *p = CURPROC;		/* XXX */
3712 	int error, flushparent;
3713 	ino_t parentino;
3714 	ufs_lbn_t lbn;
3715 
3716 	ip = VTOI(vp);
3717 	fs = ip->i_fs;
3718 	ACQUIRE_LOCK(&lk);
3719 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
3720 		FREE_LOCK(&lk);
3721 		return (0);
3722 	}
3723 	if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
3724 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
3725 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
3726 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL)
3727 		panic("softdep_fsync: pending ops");
3728 	for (error = 0, flushparent = 0; ; ) {
3729 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
3730 			break;
3731 		if (wk->wk_type != D_DIRADD)
3732 			panic("softdep_fsync: Unexpected type %s",
3733 			    TYPENAME(wk->wk_type));
3734 		dap = WK_DIRADD(wk);
3735 		/*
3736 		 * Flush our parent if this directory entry
3737 		 * has a MKDIR_PARENT dependency.
3738 		 */
3739 		if (dap->da_state & DIRCHG)
3740 			pagedep = dap->da_previous->dm_pagedep;
3741 		else
3742 			pagedep = dap->da_pagedep;
3743 		mnt = pagedep->pd_mnt;
3744 		parentino = pagedep->pd_ino;
3745 		lbn = pagedep->pd_lbn;
3746 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
3747 			panic("softdep_fsync: dirty");
3748 		flushparent = dap->da_state & MKDIR_PARENT;
3749 		/*
3750 		 * If we are being fsync'ed as part of vgone'ing this vnode,
3751 		 * then we will not be able to release and recover the
3752 		 * vnode below, so we just have to give up on writing its
3753 		 * directory entry out. It will eventually be written, just
3754 		 * not now, but then the user was not asking to have it
3755 		 * written, so we are not breaking any promises.
3756 		 */
3757 		if (vp->v_flag & VXLOCK)
3758 			break;
3759 		/*
3760 		 * We prevent deadlock by always fetching inodes from the
3761 		 * root, moving down the directory tree. Thus, when fetching
3762 		 * our parent directory, we must unlock ourselves before
3763 		 * requesting the lock on our parent. See the comment in
3764 		 * ufs_lookup for details on possible races.
3765 		 */
3766 		FREE_LOCK(&lk);
3767 		VOP_UNLOCK(vp, 0, p);
3768 		error = VFS_VGET(mnt, parentino, &pvp);
3769 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p);
3770 		if (error != 0)
3771 			return (error);
3772 		if (flushparent) {
3773 			if ((error = UFS_UPDATE(pvp, 1)) != 0) {
3774 				vput(pvp);
3775 				return (error);
3776 			}
3777 		}
3778 		/*
3779 		 * Flush directory page containing the inode's name.
3780 		 */
3781 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), p->p_ucred,
3782 		    &bp);
3783 		if (error == 0)
3784 			error = BUF_WRITE(bp);
3785 		vput(pvp);
3786 		if (error != 0)
3787 			return (error);
3788 		ACQUIRE_LOCK(&lk);
3789 		if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
3790 			break;
3791 	}
3792 	FREE_LOCK(&lk);
3793 	return (0);
3794 }
3795 
3796 /*
3797  * Flush all the dirty bitmaps associated with the block device
3798  * before flushing the rest of the dirty blocks so as to reduce
3799  * the number of dependencies that will have to be rolled back.
3800  */
3801 void
3802 softdep_fsync_mountdev(vp)
3803 	struct vnode *vp;
3804 {
3805 	struct buf *bp, *nbp;
3806 	struct worklist *wk;
3807 
3808 	if (!vn_isdisk(vp, NULL))
3809 		panic("softdep_fsync_mountdev: vnode not a disk");
3810 	ACQUIRE_LOCK(&lk);
3811 	for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
3812 		nbp = TAILQ_NEXT(bp, b_vnbufs);
3813 		/*
3814 		 * If it is already scheduled, skip to the next buffer.
3815 		 */
3816 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
3817 			continue;
3818 		if ((bp->b_flags & B_DELWRI) == 0)
3819 			panic("softdep_fsync_mountdev: not dirty");
3820 		/*
3821 		 * We are only interested in bitmaps with outstanding
3822 		 * dependencies.
3823 		 */
3824 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
3825 		    wk->wk_type != D_BMSAFEMAP ||
3826 		    (bp->b_xflags & BX_BKGRDINPROG)) {
3827 			BUF_UNLOCK(bp);
3828 			continue;
3829 		}
3830 		bremfree(bp);
3831 		FREE_LOCK(&lk);
3832 		(void) bawrite(bp);
3833 		ACQUIRE_LOCK(&lk);
3834 		/*
3835 		 * Since we may have slept during the I/O, we need
3836 		 * to start from a known point.
3837 		 */
3838 		nbp = TAILQ_FIRST(&vp->v_dirtyblkhd);
3839 	}
3840 	drain_output(vp, 1);
3841 	FREE_LOCK(&lk);
3842 }
3843 
3844 /*
3845  * This routine is called when we are trying to synchronously flush a
3846  * file. This routine must eliminate any filesystem metadata dependencies
3847  * so that the syncing routine can succeed by pushing the dirty blocks
3848  * associated with the file. If any I/O errors occur, they are returned.
3849  */
3850 int
3851 softdep_sync_metadata(ap)
3852 	struct vop_fsync_args /* {
3853 		struct vnode *a_vp;
3854 		struct ucred *a_cred;
3855 		int a_waitfor;
3856 		struct proc *a_p;
3857 	} */ *ap;
3858 {
3859 	struct vnode *vp = ap->a_vp;
3860 	struct pagedep *pagedep;
3861 	struct allocdirect *adp;
3862 	struct allocindir *aip;
3863 	struct buf *bp, *nbp;
3864 	struct worklist *wk;
3865 	int i, error, waitfor;
3866 
3867 	/*
3868 	 * Check whether this vnode is involved in a filesystem
3869 	 * that is doing soft dependency processing.
3870 	 */
3871 	if (!vn_isdisk(vp, NULL)) {
3872 		if (!DOINGSOFTDEP(vp))
3873 			return (0);
3874 	} else
3875 		if (vp->v_specmountpoint == NULL ||
3876 		    (vp->v_specmountpoint->mnt_flag & MNT_SOFTDEP) == 0)
3877 			return (0);
3878 	/*
3879 	 * Ensure that any direct block dependencies have been cleared.
3880 	 */
3881 	ACQUIRE_LOCK(&lk);
3882 	if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
3883 		FREE_LOCK(&lk);
3884 		return (error);
3885 	}
3886 	/*
3887 	 * For most files, the only metadata dependencies are the
3888 	 * cylinder group maps that allocate their inode or blocks.
3889 	 * The block allocation dependencies can be found by traversing
3890 	 * the dependency lists for any buffers that remain on their
3891 	 * dirty buffer list. The inode allocation dependency will
3892 	 * be resolved when the inode is updated with MNT_WAIT.
3893 	 * This work is done in two passes. The first pass grabs most
3894 	 * of the buffers and begins asynchronously writing them. The
3895 	 * only way to wait for these asynchronous writes is to sleep
3896 	 * on the filesystem vnode which may stay busy for a long time
3897 	 * if the filesystem is active. So, instead, we make a second
3898 	 * pass over the dependencies blocking on each write. In the
3899 	 * usual case we will be blocking against a write that we
3900 	 * initiated, so when it is done the dependency will have been
3901 	 * resolved. Thus the second pass is expected to end quickly.
3902 	 */
3903 	waitfor = MNT_NOWAIT;
3904 top:
3905 	if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) {
3906 		FREE_LOCK(&lk);
3907 		return (0);
3908 	}
3909 	bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
3910 loop:
3911 	/*
3912 	 * As we hold the buffer locked, none of its dependencies
3913 	 * will disappear.
3914 	 */
3915 	for (wk = LIST_FIRST(&bp->b_dep); wk;
3916 	     wk = LIST_NEXT(wk, wk_list)) {
3917 		switch (wk->wk_type) {
3918 
3919 		case D_ALLOCDIRECT:
3920 			adp = WK_ALLOCDIRECT(wk);
3921 			if (adp->ad_state & DEPCOMPLETE)
3922 				break;
3923 			nbp = adp->ad_buf;
3924 			if (getdirtybuf(&nbp, waitfor) == 0)
3925 				break;
3926 			FREE_LOCK(&lk);
3927 			if (waitfor == MNT_NOWAIT) {
3928 				bawrite(nbp);
3929 			} else if ((error = BUF_WRITE(nbp)) != 0) {
3930 				bawrite(bp);
3931 				return (error);
3932 			}
3933 			ACQUIRE_LOCK(&lk);
3934 			break;
3935 
3936 		case D_ALLOCINDIR:
3937 			aip = WK_ALLOCINDIR(wk);
3938 			if (aip->ai_state & DEPCOMPLETE)
3939 				break;
3940 			nbp = aip->ai_buf;
3941 			if (getdirtybuf(&nbp, waitfor) == 0)
3942 				break;
3943 			FREE_LOCK(&lk);
3944 			if (waitfor == MNT_NOWAIT) {
3945 				bawrite(nbp);
3946 			} else if ((error = BUF_WRITE(nbp)) != 0) {
3947 				bawrite(bp);
3948 				return (error);
3949 			}
3950 			ACQUIRE_LOCK(&lk);
3951 			break;
3952 
3953 		case D_INDIRDEP:
3954 		restart:
3955 			for (aip = LIST_FIRST(&WK_INDIRDEP(wk)->ir_deplisthd);
3956 			     aip; aip = LIST_NEXT(aip, ai_next)) {
3957 				if (aip->ai_state & DEPCOMPLETE)
3958 					continue;
3959 				nbp = aip->ai_buf;
3960 				if (getdirtybuf(&nbp, MNT_WAIT) == 0)
3961 					goto restart;
3962 				FREE_LOCK(&lk);
3963 				if ((error = BUF_WRITE(nbp)) != 0) {
3964 					bawrite(bp);
3965 					return (error);
3966 				}
3967 				ACQUIRE_LOCK(&lk);
3968 				goto restart;
3969 			}
3970 			break;
3971 
3972 		case D_INODEDEP:
3973 			if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
3974 			    WK_INODEDEP(wk)->id_ino)) != 0) {
3975 				FREE_LOCK(&lk);
3976 				bawrite(bp);
3977 				return (error);
3978 			}
3979 			break;
3980 
3981 		case D_PAGEDEP:
3982 			/*
3983 			 * We are trying to sync a directory that may
3984 			 * have dependencies on both its own metadata
3985 			 * and/or dependencies on the inodes of any
3986 			 * recently allocated files. We walk its diradd
3987 			 * lists pushing out the associated inode.
3988 			 */
3989 			pagedep = WK_PAGEDEP(wk);
3990 			for (i = 0; i < DAHASHSZ; i++) {
3991 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
3992 					continue;
3993 				if ((error =
3994 				    flush_pagedep_deps(vp, pagedep->pd_mnt,
3995 						&pagedep->pd_diraddhd[i]))) {
3996 					FREE_LOCK(&lk);
3997 					bawrite(bp);
3998 					return (error);
3999 				}
4000 			}
4001 			break;
4002 
4003 		case D_MKDIR:
4004 			/*
4005 			 * This case should never happen if the vnode has
4006 			 * been properly sync'ed. However, if this function
4007 			 * is used at a place where the vnode has not yet
4008 			 * been sync'ed, this dependency can show up. So,
4009 			 * rather than panic, just flush it.
4010 			 */
4011 			nbp = WK_MKDIR(wk)->md_buf;
4012 			if (getdirtybuf(&nbp, waitfor) == 0)
4013 				break;
4014 			FREE_LOCK(&lk);
4015 			if (waitfor == MNT_NOWAIT) {
4016 				bawrite(nbp);
4017 			} else if ((error = BUF_WRITE(nbp)) != 0) {
4018 				bawrite(bp);
4019 				return (error);
4020 			}
4021 			ACQUIRE_LOCK(&lk);
4022 			break;
4023 
4024 		case D_BMSAFEMAP:
4025 			/*
4026 			 * This case should never happen if the vnode has
4027 			 * been properly sync'ed. However, if this function
4028 			 * is used at a place where the vnode has not yet
4029 			 * been sync'ed, this dependency can show up. So,
4030 			 * rather than panic, just flush it.
4031 			 */
4032 			nbp = WK_BMSAFEMAP(wk)->sm_buf;
4033 			if (getdirtybuf(&nbp, waitfor) == 0)
4034 				break;
4035 			FREE_LOCK(&lk);
4036 			if (waitfor == MNT_NOWAIT) {
4037 				bawrite(nbp);
4038 			} else if ((error = BUF_WRITE(nbp)) != 0) {
4039 				bawrite(bp);
4040 				return (error);
4041 			}
4042 			ACQUIRE_LOCK(&lk);
4043 			break;
4044 
4045 		default:
4046 			panic("softdep_sync_metadata: Unknown type %s",
4047 			    TYPENAME(wk->wk_type));
4048 			/* NOTREACHED */
4049 		}
4050 	}
4051 	(void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT);
4052 	nbp = TAILQ_NEXT(bp, b_vnbufs);
4053 	FREE_LOCK(&lk);
4054 	bawrite(bp);
4055 	ACQUIRE_LOCK(&lk);
4056 	if (nbp != NULL) {
4057 		bp = nbp;
4058 		goto loop;
4059 	}
4060 	/*
4061 	 * We must wait for any I/O in progress to finish so that
4062 	 * all potential buffers on the dirty list will be visible.
4063 	 * Once they are all there, proceed with the second pass
4064 	 * which will wait for the I/O as per above.
4065 	 */
4066 	drain_output(vp, 1);
4067 	/*
4068 	 * The brief unlock is to allow any pent up dependency
4069 	 * processing to be done.
4070 	 */
4071 	if (waitfor == MNT_NOWAIT) {
4072 		waitfor = MNT_WAIT;
4073 		FREE_LOCK(&lk);
4074 		ACQUIRE_LOCK(&lk);
4075 		goto top;
4076 	}
4077 
4078 	/*
4079 	 * If we have managed to get rid of all the dirty buffers,
4080 	 * then we are done. For certain directories and block
4081 	 * devices, we may need to do further work.
4082 	 */
4083 	if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) {
4084 		FREE_LOCK(&lk);
4085 		return (0);
4086 	}
4087 
4088 	FREE_LOCK(&lk);
4089 	/*
4090 	 * If we are trying to sync a block device, some of its buffers may
4091 	 * contain metadata that cannot be written until the contents of some
4092 	 * partially written files have been written to disk. The only easy
4093 	 * way to accomplish this is to sync the entire filesystem (luckily
4094 	 * this happens rarely).
4095 	 */
4096 	if (vn_isdisk(vp, NULL) &&
4097 	    vp->v_specmountpoint && !VOP_ISLOCKED(vp, NULL) &&
4098 	    (error = VFS_SYNC(vp->v_specmountpoint, MNT_WAIT, ap->a_cred,
4099 	     ap->a_p)) != 0)
4100 		return (error);
4101 	return (0);
4102 }
4103 
4104 /*
4105  * Flush the dependencies associated with an inodedep.
4106  * Called with splbio blocked.
4107  */
4108 static int
4109 flush_inodedep_deps(fs, ino)
4110 	struct fs *fs;
4111 	ino_t ino;
4112 {
4113 	struct inodedep *inodedep;
4114 	struct allocdirect *adp;
4115 	int error, waitfor;
4116 	struct buf *bp;
4117 
4118 	/*
4119 	 * This work is done in two passes. The first pass grabs most
4120 	 * of the buffers and begins asynchronously writing them. The
4121 	 * only way to wait for these asynchronous writes is to sleep
4122 	 * on the filesystem vnode which may stay busy for a long time
4123 	 * if the filesystem is active. So, instead, we make a second
4124 	 * pass over the dependencies blocking on each write. In the
4125 	 * usual case we will be blocking against a write that we
4126 	 * initiated, so when it is done the dependency will have been
4127 	 * resolved. Thus the second pass is expected to end quickly.
4128 	 * We give a brief window at the top of the loop to allow
4129 	 * any pending I/O to complete.
4130 	 */
4131 	for (waitfor = MNT_NOWAIT; ; ) {
4132 		FREE_LOCK(&lk);
4133 		ACQUIRE_LOCK(&lk);
4134 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4135 			return (0);
4136 		for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
4137 		     adp = TAILQ_NEXT(adp, ad_next)) {
4138 			if (adp->ad_state & DEPCOMPLETE)
4139 				continue;
4140 			bp = adp->ad_buf;
4141 			if (getdirtybuf(&bp, waitfor) == 0) {
4142 				if (waitfor == MNT_NOWAIT)
4143 					continue;
4144 				break;
4145 			}
4146 			FREE_LOCK(&lk);
4147 			if (waitfor == MNT_NOWAIT) {
4148 				bawrite(bp);
4149 			} else if ((error = BUF_WRITE(bp)) != 0) {
4150 				ACQUIRE_LOCK(&lk);
4151 				return (error);
4152 			}
4153 			ACQUIRE_LOCK(&lk);
4154 			break;
4155 		}
4156 		if (adp != NULL)
4157 			continue;
4158 		for (adp = TAILQ_FIRST(&inodedep->id_newinoupdt); adp;
4159 		     adp = TAILQ_NEXT(adp, ad_next)) {
4160 			if (adp->ad_state & DEPCOMPLETE)
4161 				continue;
4162 			bp = adp->ad_buf;
4163 			if (getdirtybuf(&bp, waitfor) == 0) {
4164 				if (waitfor == MNT_NOWAIT)
4165 					continue;
4166 				break;
4167 			}
4168 			FREE_LOCK(&lk);
4169 			if (waitfor == MNT_NOWAIT) {
4170 				bawrite(bp);
4171 			} else if ((error = BUF_WRITE(bp)) != 0) {
4172 				ACQUIRE_LOCK(&lk);
4173 				return (error);
4174 			}
4175 			ACQUIRE_LOCK(&lk);
4176 			break;
4177 		}
4178 		if (adp != NULL)
4179 			continue;
4180 		/*
4181 		 * If pass2, we are done, otherwise do pass 2.
4182 		 */
4183 		if (waitfor == MNT_WAIT)
4184 			break;
4185 		waitfor = MNT_WAIT;
4186 	}
4187 	/*
4188 	 * Try freeing inodedep in case all dependencies have been removed.
4189 	 */
4190 	if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
4191 		(void) free_inodedep(inodedep);
4192 	return (0);
4193 }
4194 
4195 /*
4196  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
4197  * Called with splbio blocked.
4198  */
4199 static int
4200 flush_pagedep_deps(pvp, mp, diraddhdp)
4201 	struct vnode *pvp;
4202 	struct mount *mp;
4203 	struct diraddhd *diraddhdp;
4204 {
4205 	struct proc *p = CURPROC;	/* XXX */
4206 	struct inodedep *inodedep;
4207 	struct ufsmount *ump;
4208 	struct diradd *dap;
4209 	struct vnode *vp;
4210 	int gotit, error = 0;
4211 	struct buf *bp;
4212 	ino_t inum;
4213 
4214 	ump = VFSTOUFS(mp);
4215 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
4216 		/*
4217 		 * Flush ourselves if this directory entry
4218 		 * has a MKDIR_PARENT dependency.
4219 		 */
4220 		if (dap->da_state & MKDIR_PARENT) {
4221 			FREE_LOCK(&lk);
4222 			if ((error = UFS_UPDATE(pvp, 1)) != 0)
4223 				break;
4224 			ACQUIRE_LOCK(&lk);
4225 			/*
4226 			 * If that cleared dependencies, go on to next.
4227 			 */
4228 			if (dap != LIST_FIRST(diraddhdp))
4229 				continue;
4230 			if (dap->da_state & MKDIR_PARENT)
4231 				panic("flush_pagedep_deps: MKDIR_PARENT");
4232 		}
4233 		/*
4234 		 * A newly allocated directory must have its "." and
4235 		 * ".." entries written out before its name can be
4236 		 * committed in its parent. We do not want or need
4237 		 * the full semantics of a synchronous VOP_FSYNC as
4238 		 * that may end up here again, once for each directory
4239 		 * level in the filesystem. Instead, we push the blocks
4240 		 * and wait for them to clear. We have to fsync twice
4241 		 * because the first call may choose to defer blocks
4242 		 * that still have dependencies, but deferral will
4243 		 * happen at most once.
4244 		 */
4245 		inum = dap->da_newinum;
4246 		if (dap->da_state & MKDIR_BODY) {
4247 			FREE_LOCK(&lk);
4248 			if ((error = VFS_VGET(mp, inum, &vp)) != 0)
4249 				break;
4250 			if ((error=VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)) ||
4251 			    (error=VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) {
4252 				vput(vp);
4253 				break;
4254 			}
4255 			drain_output(vp, 0);
4256 			vput(vp);
4257 			ACQUIRE_LOCK(&lk);
4258 			/*
4259 			 * If that cleared dependencies, go on to next.
4260 			 */
4261 			if (dap != LIST_FIRST(diraddhdp))
4262 				continue;
4263 			if (dap->da_state & MKDIR_BODY)
4264 				panic("flush_pagedep_deps: MKDIR_BODY");
4265 		}
4266 		/*
4267 		 * Flush the inode on which the directory entry depends.
4268 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
4269 		 * the only remaining dependency is that the updated inode
4270 		 * count must get pushed to disk. The inode has already
4271 		 * been pushed into its inode buffer (via VOP_UPDATE) at
4272 		 * the time of the reference count change. So we need only
4273 		 * locate that buffer, ensure that there will be no rollback
4274 		 * caused by a bitmap dependency, then write the inode buffer.
4275 		 */
4276 		if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0)
4277 			panic("flush_pagedep_deps: lost inode");
4278 		/*
4279 		 * If the inode still has bitmap dependencies,
4280 		 * push them to disk.
4281 		 */
4282 		if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4283 			gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4284 			FREE_LOCK(&lk);
4285 			if (gotit &&
4286 			    (error = BUF_WRITE(inodedep->id_buf)) != 0)
4287 				break;
4288 			ACQUIRE_LOCK(&lk);
4289 			if (dap != LIST_FIRST(diraddhdp))
4290 				continue;
4291 		}
4292 		/*
4293 		 * If the inode is still sitting in a buffer waiting
4294 		 * to be written, push it to disk.
4295 		 */
4296 		FREE_LOCK(&lk);
4297 		if ((error = bread(ump->um_devvp,
4298 		    fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
4299 		    (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0)
4300 			break;
4301 		if ((error = BUF_WRITE(bp)) != 0)
4302 			break;
4303 		ACQUIRE_LOCK(&lk);
4304 		/*
4305 		 * If we have failed to get rid of all the dependencies
4306 		 * then something is seriously wrong.
4307 		 */
4308 		if (dap == LIST_FIRST(diraddhdp))
4309 			panic("flush_pagedep_deps: flush failed");
4310 	}
4311 	if (error)
4312 		ACQUIRE_LOCK(&lk);
4313 	return (error);
4314 }
4315 
4316 /*
4317  * A large burst of file addition or deletion activity can drive the
4318  * memory load excessively high. Therefore we deliberately slow things
4319  * down and speed up the I/O processing if we find ourselves with too
4320  * many dependencies in progress.
4321  */
4322 static int
4323 request_cleanup(resource, islocked)
4324 	int resource;
4325 	int islocked;
4326 {
4327 	struct callout_handle handle;
4328 	struct proc *p = CURPROC;
4329 
4330 	/*
4331 	 * We never hold up the filesystem syncer process.
4332 	 */
4333 	if (p == filesys_syncer)
4334 		return (0);
4335 	/*
4336 	 * If we are resource constrained on inode dependencies, try
4337 	 * flushing some dirty inodes. Otherwise, we are constrained
4338 	 * by file deletions, so try accelerating flushes of directories
4339 	 * with removal dependencies. We would like to do the cleanup
4340 	 * here, but we probably hold an inode locked at this point and
4341 	 * that might deadlock against one that we try to clean. So,
4342 	 * the best that we can do is request the syncer daemon to do
4343 	 * the cleanup for us.
4344 	 */
4345 	switch (resource) {
4346 
4347 	case FLUSH_INODES:
4348 		stat_ino_limit_push += 1;
4349 		req_clear_inodedeps = 1;
4350 		break;
4351 
4352 	case FLUSH_REMOVE:
4353 		stat_blk_limit_push += 1;
4354 		req_clear_remove = 1;
4355 		break;
4356 
4357 	default:
4358 		panic("request_cleanup: unknown type");
4359 	}
4360 	/*
4361 	 * Hopefully the syncer daemon will catch up and awaken us.
4362 	 * We wait at most tickdelay before proceeding in any case.
4363 	 */
4364 	if (islocked == 0)
4365 		ACQUIRE_LOCK(&lk);
4366 	if (proc_waiting == 0) {
4367 		proc_waiting = 1;
4368 		handle = timeout(pause_timer, NULL,
4369 		    tickdelay > 2 ? tickdelay : 2);
4370 	}
4371 	FREE_LOCK_INTERLOCKED(&lk);
4372 	(void) tsleep((caddr_t)&proc_waiting, PPAUSE, "softupdate", 0);
4373 	ACQUIRE_LOCK_INTERLOCKED(&lk);
4374 	if (proc_waiting) {
4375 		untimeout(pause_timer, NULL, handle);
4376 		proc_waiting = 0;
4377 	} else {
4378 		switch (resource) {
4379 
4380 		case FLUSH_INODES:
4381 			stat_ino_limit_hit += 1;
4382 			break;
4383 
4384 		case FLUSH_REMOVE:
4385 			stat_blk_limit_hit += 1;
4386 			break;
4387 		}
4388 	}
4389 	if (islocked == 0)
4390 		FREE_LOCK(&lk);
4391 	return (1);
4392 }
4393 
4394 /*
4395  * Awaken processes pausing in request_cleanup and clear proc_waiting
4396  * to indicate that there is no longer a timer running.
4397  */
4398 void
4399 pause_timer(arg)
4400 	void *arg;
4401 {
4402 
4403 	proc_waiting = 0;
4404 	wakeup(&proc_waiting);
4405 }
4406 
4407 /*
4408  * Flush out a directory with at least one removal dependency in an effort to
4409  * reduce the number of dirrem, freefile, and freeblks dependency structures.
4410  */
4411 static void
4412 clear_remove(p)
4413 	struct proc *p;
4414 {
4415 	struct pagedep_hashhead *pagedephd;
4416 	struct pagedep *pagedep;
4417 	static int next = 0;
4418 	struct mount *mp;
4419 	struct vnode *vp;
4420 	int error, cnt;
4421 	ino_t ino;
4422 
4423 	ACQUIRE_LOCK(&lk);
4424 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
4425 		pagedephd = &pagedep_hashtbl[next++];
4426 		if (next >= pagedep_hash)
4427 			next = 0;
4428 		for (pagedep = LIST_FIRST(pagedephd); pagedep;
4429 		     pagedep = LIST_NEXT(pagedep, pd_hash)) {
4430 			if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
4431 				continue;
4432 			mp = pagedep->pd_mnt;
4433 			ino = pagedep->pd_ino;
4434 			FREE_LOCK(&lk);
4435 			if (vn_start_write(NULL, &mp, V_WAIT | PCATCH) != 0)
4436 				return;
4437 			if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4438 				softdep_error("clear_remove: vget", error);
4439 				vn_finished_write(mp);
4440 				return;
4441 			}
4442 			if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)))
4443 				softdep_error("clear_remove: fsync", error);
4444 			drain_output(vp, 0);
4445 			vput(vp);
4446 			vn_finished_write(mp);
4447 			return;
4448 		}
4449 	}
4450 	FREE_LOCK(&lk);
4451 }
4452 
4453 /*
4454  * Clear out a block of dirty inodes in an effort to reduce
4455  * the number of inodedep dependency structures.
4456  */
4457 static void
4458 clear_inodedeps(p)
4459 	struct proc *p;
4460 {
4461 	struct inodedep_hashhead *inodedephd;
4462 	struct inodedep *inodedep;
4463 	static int next = 0;
4464 	struct mount *mp;
4465 	struct vnode *vp;
4466 	struct fs *fs;
4467 	int error, cnt;
4468 	ino_t firstino, lastino, ino;
4469 
4470 	ACQUIRE_LOCK(&lk);
4471 	/*
4472 	 * Pick a random inode dependency to be cleared.
4473 	 * We will then gather up all the inodes in its block
4474 	 * that have dependencies and flush them out.
4475 	 */
4476 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
4477 		inodedephd = &inodedep_hashtbl[next++];
4478 		if (next >= inodedep_hash)
4479 			next = 0;
4480 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
4481 			break;
4482 	}
4483 	/*
4484 	 * Ugly code to find mount point given pointer to superblock.
4485 	 */
4486 	fs = inodedep->id_fs;
4487 	TAILQ_FOREACH(mp, &mountlist, mnt_list)
4488 		if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs)
4489 			break;
4490 	/*
4491 	 * Find the last inode in the block with dependencies.
4492 	 */
4493 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
4494 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
4495 		if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
4496 			break;
4497 	/*
4498 	 * Asynchronously push all but the last inode with dependencies.
4499 	 * Synchronously push the last inode with dependencies to ensure
4500 	 * that the inode block gets written to free up the inodedeps.
4501 	 */
4502 	for (ino = firstino; ino <= lastino; ino++) {
4503 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4504 			continue;
4505 		FREE_LOCK(&lk);
4506 		if (vn_start_write(NULL, &mp, V_WAIT | PCATCH) != 0)
4507 			return;
4508 		if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4509 			softdep_error("clear_inodedeps: vget", error);
4510 			vn_finished_write(mp);
4511 			return;
4512 		}
4513 		if (ino == lastino) {
4514 			if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_WAIT, p)))
4515 				softdep_error("clear_inodedeps: fsync1", error);
4516 		} else {
4517 			if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)))
4518 				softdep_error("clear_inodedeps: fsync2", error);
4519 			drain_output(vp, 0);
4520 		}
4521 		vput(vp);
4522 		vn_finished_write(mp);
4523 		ACQUIRE_LOCK(&lk);
4524 	}
4525 	FREE_LOCK(&lk);
4526 }
4527 
4528 /*
4529  * Function to determine if the buffer has outstanding dependencies
4530  * that will cause a roll-back if the buffer is written. If wantcount
4531  * is set, return number of dependencies, otherwise just yes or no.
4532  */
4533 static int
4534 softdep_count_dependencies(bp, wantcount)
4535 	struct buf *bp;
4536 	int wantcount;
4537 {
4538 	struct worklist *wk;
4539 	struct inodedep *inodedep;
4540 	struct indirdep *indirdep;
4541 	struct allocindir *aip;
4542 	struct pagedep *pagedep;
4543 	struct diradd *dap;
4544 	int i, retval;
4545 
4546 	retval = 0;
4547 	ACQUIRE_LOCK(&lk);
4548 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = LIST_NEXT(wk, wk_list)) {
4549 		switch (wk->wk_type) {
4550 
4551 		case D_INODEDEP:
4552 			inodedep = WK_INODEDEP(wk);
4553 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4554 				/* bitmap allocation dependency */
4555 				retval += 1;
4556 				if (!wantcount)
4557 					goto out;
4558 			}
4559 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
4560 				/* direct block pointer dependency */
4561 				retval += 1;
4562 				if (!wantcount)
4563 					goto out;
4564 			}
4565 			continue;
4566 
4567 		case D_INDIRDEP:
4568 			indirdep = WK_INDIRDEP(wk);
4569 			for (aip = LIST_FIRST(&indirdep->ir_deplisthd);
4570 			     aip; aip = LIST_NEXT(aip, ai_next)) {
4571 				/* indirect block pointer dependency */
4572 				retval += 1;
4573 				if (!wantcount)
4574 					goto out;
4575 			}
4576 			continue;
4577 
4578 		case D_PAGEDEP:
4579 			pagedep = WK_PAGEDEP(wk);
4580 			for (i = 0; i < DAHASHSZ; i++) {
4581 				for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]);
4582 				     dap; dap = LIST_NEXT(dap, da_pdlist)) {
4583 					/* directory entry dependency */
4584 					retval += 1;
4585 					if (!wantcount)
4586 						goto out;
4587 				}
4588 			}
4589 			continue;
4590 
4591 		case D_BMSAFEMAP:
4592 		case D_ALLOCDIRECT:
4593 		case D_ALLOCINDIR:
4594 		case D_MKDIR:
4595 			/* never a dependency on these blocks */
4596 			continue;
4597 
4598 		default:
4599 			panic("softdep_check_for_rollback: Unexpected type %s",
4600 			    TYPENAME(wk->wk_type));
4601 			/* NOTREACHED */
4602 		}
4603 	}
4604 out:
4605 	FREE_LOCK(&lk);
4606 	return retval;
4607 }
4608 
4609 /*
4610  * Acquire exclusive access to a buffer.
4611  * Must be called with splbio blocked.
4612  * Return 1 if buffer was acquired.
4613  */
4614 static int
4615 getdirtybuf(bpp, waitfor)
4616 	struct buf **bpp;
4617 	int waitfor;
4618 {
4619 	struct buf *bp;
4620 
4621 	for (;;) {
4622 		if ((bp = *bpp) == NULL)
4623 			return (0);
4624 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
4625 			if ((bp->b_xflags & BX_BKGRDINPROG) == 0)
4626 				break;
4627 			BUF_UNLOCK(bp);
4628 			if (waitfor != MNT_WAIT)
4629 				return (0);
4630 			bp->b_xflags |= BX_BKGRDWAIT;
4631 			FREE_LOCK_INTERLOCKED(&lk);
4632 			tsleep(&bp->b_xflags, PRIBIO, "getbuf", 0);
4633 			ACQUIRE_LOCK_INTERLOCKED(&lk);
4634 			continue;
4635 		}
4636 		if (waitfor != MNT_WAIT)
4637 			return (0);
4638 		FREE_LOCK_INTERLOCKED(&lk);
4639 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL) != ENOLCK)
4640 			panic("getdirtybuf: inconsistent lock");
4641 		ACQUIRE_LOCK_INTERLOCKED(&lk);
4642 	}
4643 	if ((bp->b_flags & B_DELWRI) == 0) {
4644 		BUF_UNLOCK(bp);
4645 		return (0);
4646 	}
4647 	bremfree(bp);
4648 	return (1);
4649 }
4650 
4651 /*
4652  * Wait for pending output on a vnode to complete.
4653  * Must be called with vnode locked.
4654  */
4655 static void
4656 drain_output(vp, islocked)
4657 	struct vnode *vp;
4658 	int islocked;
4659 {
4660 
4661 	if (!islocked)
4662 		ACQUIRE_LOCK(&lk);
4663 	while (vp->v_numoutput) {
4664 		vp->v_flag |= VBWAIT;
4665 		FREE_LOCK_INTERLOCKED(&lk);
4666 		tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, "drainvp", 0);
4667 		ACQUIRE_LOCK_INTERLOCKED(&lk);
4668 	}
4669 	if (!islocked)
4670 		FREE_LOCK(&lk);
4671 }
4672 
4673 /*
4674  * Called whenever a buffer that is being invalidated or reallocated
4675  * contains dependencies. This should only happen if an I/O error has
4676  * occurred. The routine is called with the buffer locked.
4677  */
4678 static void
4679 softdep_deallocate_dependencies(bp)
4680 	struct buf *bp;
4681 {
4682 
4683 	if ((bp->b_ioflags & BIO_ERROR) == 0)
4684 		panic("softdep_deallocate_dependencies: dangling deps");
4685 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
4686 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
4687 }
4688 
4689 /*
4690  * Function to handle asynchronous write errors in the filesystem.
4691  */
4692 void
4693 softdep_error(func, error)
4694 	char *func;
4695 	int error;
4696 {
4697 
4698 	/* XXX should do something better! */
4699 	printf("%s: got error %d while accessing filesystem\n", func, error);
4700 }
4701