xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision c4f6a2a9e1b1879b618c436ab4f56ff75c73a0f5)
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  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 /*
45  * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
46  */
47 #ifndef DIAGNOSTIC
48 #define DIAGNOSTIC
49 #endif
50 #ifndef DEBUG
51 #define DEBUG
52 #endif
53 
54 #include <sys/param.h>
55 #include <sys/kernel.h>
56 #include <sys/systm.h>
57 #include <sys/stdint.h>
58 #include <sys/bio.h>
59 #include <sys/buf.h>
60 #include <sys/malloc.h>
61 #include <sys/mount.h>
62 #include <sys/proc.h>
63 #include <sys/stat.h>
64 #include <sys/syslog.h>
65 #include <sys/vnode.h>
66 #include <sys/conf.h>
67 #include <ufs/ufs/dir.h>
68 #include <ufs/ufs/extattr.h>
69 #include <ufs/ufs/quota.h>
70 #include <ufs/ufs/inode.h>
71 #include <ufs/ufs/ufsmount.h>
72 #include <ufs/ffs/fs.h>
73 #include <ufs/ffs/softdep.h>
74 #include <ufs/ffs/ffs_extern.h>
75 #include <ufs/ufs/ufs_extern.h>
76 
77 /*
78  * These definitions need to be adapted to the system to which
79  * this file is being ported.
80  */
81 /*
82  * malloc types defined for the softdep system.
83  */
84 static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
85 static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
86 static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
87 static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
88 static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
89 static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
90 static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
91 static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
92 static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
93 static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
94 static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
95 static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
96 static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
97 static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block");
98 
99 #define M_SOFTDEP_FLAGS	(M_WAITOK | M_USE_RESERVE)
100 
101 #define	D_PAGEDEP	0
102 #define	D_INODEDEP	1
103 #define	D_NEWBLK	2
104 #define	D_BMSAFEMAP	3
105 #define	D_ALLOCDIRECT	4
106 #define	D_INDIRDEP	5
107 #define	D_ALLOCINDIR	6
108 #define	D_FREEFRAG	7
109 #define	D_FREEBLKS	8
110 #define	D_FREEFILE	9
111 #define	D_DIRADD	10
112 #define	D_MKDIR		11
113 #define	D_DIRREM	12
114 #define	D_NEWDIRBLK	13
115 #define	D_LAST		D_NEWDIRBLK
116 
117 /*
118  * translate from workitem type to memory type
119  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
120  */
121 static struct malloc_type *memtype[] = {
122 	M_PAGEDEP,
123 	M_INODEDEP,
124 	M_NEWBLK,
125 	M_BMSAFEMAP,
126 	M_ALLOCDIRECT,
127 	M_INDIRDEP,
128 	M_ALLOCINDIR,
129 	M_FREEFRAG,
130 	M_FREEBLKS,
131 	M_FREEFILE,
132 	M_DIRADD,
133 	M_MKDIR,
134 	M_DIRREM,
135 	M_NEWDIRBLK
136 };
137 
138 #define DtoM(type) (memtype[type])
139 
140 /*
141  * Names of malloc types.
142  */
143 #define TYPENAME(type)  \
144 	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
145 /*
146  * End system adaptaion definitions.
147  */
148 
149 /*
150  * Internal function prototypes.
151  */
152 static	void softdep_error(char *, int);
153 static	void drain_output(struct vnode *, int);
154 static	int getdirtybuf(struct buf **, int);
155 static	void clear_remove(struct thread *);
156 static	void clear_inodedeps(struct thread *);
157 static	int flush_pagedep_deps(struct vnode *, struct mount *,
158 	    struct diraddhd *);
159 static	int flush_inodedep_deps(struct fs *, ino_t);
160 static	int flush_deplist(struct allocdirectlst *, int, int *);
161 static	int handle_written_filepage(struct pagedep *, struct buf *);
162 static  void diradd_inode_written(struct diradd *, struct inodedep *);
163 static	int handle_written_inodeblock(struct inodedep *, struct buf *);
164 static	void handle_allocdirect_partdone(struct allocdirect *);
165 static	void handle_allocindir_partdone(struct allocindir *);
166 static	void initiate_write_filepage(struct pagedep *, struct buf *);
167 static	void handle_written_mkdir(struct mkdir *, int);
168 static	void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
169 static	void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
170 static	void handle_workitem_freefile(struct freefile *);
171 static	void handle_workitem_remove(struct dirrem *, struct vnode *);
172 static	struct dirrem *newdirrem(struct buf *, struct inode *,
173 	    struct inode *, int, struct dirrem **);
174 static	void free_diradd(struct diradd *);
175 static	void free_allocindir(struct allocindir *, struct inodedep *);
176 static	void free_newdirblk(struct newdirblk *);
177 static	int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t,
178 	    ufs2_daddr_t *);
179 static	void deallocate_dependencies(struct buf *, struct inodedep *);
180 static	void free_allocdirect(struct allocdirectlst *,
181 	    struct allocdirect *, int);
182 static	int check_inode_unwritten(struct inodedep *);
183 static	int free_inodedep(struct inodedep *);
184 static	void handle_workitem_freeblocks(struct freeblks *, int);
185 static	void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
186 static	void setup_allocindir_phase2(struct buf *, struct inode *,
187 	    struct allocindir *);
188 static	struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
189 	    ufs2_daddr_t);
190 static	void handle_workitem_freefrag(struct freefrag *);
191 static	struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long);
192 static	void allocdirect_merge(struct allocdirectlst *,
193 	    struct allocdirect *, struct allocdirect *);
194 static	struct bmsafemap *bmsafemap_lookup(struct buf *);
195 static	int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **);
196 static	int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
197 static	int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **);
198 static	void pause_timer(void *);
199 static	int request_cleanup(int, int);
200 static	int process_worklist_item(struct mount *, int);
201 static	void add_to_worklist(struct worklist *);
202 
203 /*
204  * Exported softdep operations.
205  */
206 static	void softdep_disk_io_initiation(struct buf *);
207 static	void softdep_disk_write_complete(struct buf *);
208 static	void softdep_deallocate_dependencies(struct buf *);
209 static	void softdep_move_dependencies(struct buf *, struct buf *);
210 static	int softdep_count_dependencies(struct buf *bp, int);
211 
212 /*
213  * Locking primitives.
214  *
215  * For a uniprocessor, all we need to do is protect against disk
216  * interrupts. For a multiprocessor, this lock would have to be
217  * a mutex. A single mutex is used throughout this file, though
218  * finer grain locking could be used if contention warranted it.
219  *
220  * For a multiprocessor, the sleep call would accept a lock and
221  * release it after the sleep processing was complete. In a uniprocessor
222  * implementation there is no such interlock, so we simple mark
223  * the places where it needs to be done with the `interlocked' form
224  * of the lock calls. Since the uniprocessor sleep already interlocks
225  * the spl, there is nothing that really needs to be done.
226  */
227 #ifndef /* NOT */ DEBUG
228 static struct lockit {
229 	int	lkt_spl;
230 } lk = { 0 };
231 #define ACQUIRE_LOCK(lk)		(lk)->lkt_spl = splbio()
232 #define FREE_LOCK(lk)			splx((lk)->lkt_spl)
233 
234 #else /* DEBUG */
235 #define NOHOLDER	((struct thread *)-1)
236 #define SPECIAL_FLAG	((struct thread *)-2)
237 static struct lockit {
238 	int	lkt_spl;
239 	struct	thread *lkt_held;
240 } lk = { 0, NOHOLDER };
241 static int lockcnt;
242 
243 static	void acquire_lock(struct lockit *);
244 static	void free_lock(struct lockit *);
245 void	softdep_panic(char *);
246 
247 #define ACQUIRE_LOCK(lk)		acquire_lock(lk)
248 #define FREE_LOCK(lk)			free_lock(lk)
249 
250 static void
251 acquire_lock(lk)
252 	struct lockit *lk;
253 {
254 	struct thread *holder;
255 
256 	if (lk->lkt_held != NOHOLDER) {
257 		holder = lk->lkt_held;
258 		FREE_LOCK(lk);
259 		if (holder == curthread)
260 			panic("softdep_lock: locking against myself");
261 		else
262 			panic("softdep_lock: lock held by %p", holder);
263 	}
264 	lk->lkt_spl = splbio();
265 	lk->lkt_held = curthread;
266 	lockcnt++;
267 }
268 
269 static void
270 free_lock(lk)
271 	struct lockit *lk;
272 {
273 
274 	if (lk->lkt_held == NOHOLDER)
275 		panic("softdep_unlock: lock not held");
276 	lk->lkt_held = NOHOLDER;
277 	splx(lk->lkt_spl);
278 }
279 
280 /*
281  * Function to release soft updates lock and panic.
282  */
283 void
284 softdep_panic(msg)
285 	char *msg;
286 {
287 
288 	if (lk.lkt_held != NOHOLDER)
289 		FREE_LOCK(&lk);
290 	panic(msg);
291 }
292 #endif /* DEBUG */
293 
294 static	int interlocked_sleep(struct lockit *, int, void *, struct mtx *, int,
295 	    const char *, int);
296 
297 /*
298  * When going to sleep, we must save our SPL so that it does
299  * not get lost if some other process uses the lock while we
300  * are sleeping. We restore it after we have slept. This routine
301  * wraps the interlocking with functions that sleep. The list
302  * below enumerates the available set of operations.
303  */
304 #define	UNKNOWN		0
305 #define	SLEEP		1
306 #define	LOCKBUF		2
307 
308 static int
309 interlocked_sleep(lk, op, ident, mtx, flags, wmesg, timo)
310 	struct lockit *lk;
311 	int op;
312 	void *ident;
313 	struct mtx *mtx;
314 	int flags;
315 	const char *wmesg;
316 	int timo;
317 {
318 	struct thread *holder;
319 	int s, retval;
320 
321 	s = lk->lkt_spl;
322 #	ifdef DEBUG
323 	if (lk->lkt_held == NOHOLDER)
324 		panic("interlocked_sleep: lock not held");
325 	lk->lkt_held = NOHOLDER;
326 #	endif /* DEBUG */
327 	switch (op) {
328 	case SLEEP:
329 		retval = msleep(ident, mtx, flags, wmesg, timo);
330 		break;
331 	case LOCKBUF:
332 		retval = BUF_LOCK((struct buf *)ident, flags);
333 		break;
334 	default:
335 		panic("interlocked_sleep: unknown operation");
336 	}
337 #	ifdef DEBUG
338 	if (lk->lkt_held != NOHOLDER) {
339 		holder = lk->lkt_held;
340 		FREE_LOCK(lk);
341 		if (holder == curthread)
342 			panic("interlocked_sleep: locking against self");
343 		else
344 			panic("interlocked_sleep: lock held by %p", holder);
345 	}
346 	lk->lkt_held = curthread;
347 	lockcnt++;
348 #	endif /* DEBUG */
349 	lk->lkt_spl = s;
350 	return (retval);
351 }
352 
353 /*
354  * Place holder for real semaphores.
355  */
356 struct sema {
357 	int	value;
358 	struct	thread *holder;
359 	char	*name;
360 	int	prio;
361 	int	timo;
362 };
363 static	void sema_init(struct sema *, char *, int, int);
364 static	int sema_get(struct sema *, struct lockit *);
365 static	void sema_release(struct sema *);
366 
367 static void
368 sema_init(semap, name, prio, timo)
369 	struct sema *semap;
370 	char *name;
371 	int prio, timo;
372 {
373 
374 	semap->holder = NOHOLDER;
375 	semap->value = 0;
376 	semap->name = name;
377 	semap->prio = prio;
378 	semap->timo = timo;
379 }
380 
381 static int
382 sema_get(semap, interlock)
383 	struct sema *semap;
384 	struct lockit *interlock;
385 {
386 
387 	if (semap->value++ > 0) {
388 		if (interlock != NULL) {
389 			interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
390 			    NULL, semap->prio, semap->name,
391 			    semap->timo);
392 			FREE_LOCK(interlock);
393 		} else {
394 			tsleep((caddr_t)semap, semap->prio, semap->name,
395 			    semap->timo);
396 		}
397 		return (0);
398 	}
399 	semap->holder = curthread;
400 	if (interlock != NULL)
401 		FREE_LOCK(interlock);
402 	return (1);
403 }
404 
405 static void
406 sema_release(semap)
407 	struct sema *semap;
408 {
409 
410 	if (semap->value <= 0 || semap->holder != curthread) {
411 		if (lk.lkt_held != NOHOLDER)
412 			FREE_LOCK(&lk);
413 		panic("sema_release: not held");
414 	}
415 	if (--semap->value > 0) {
416 		semap->value = 0;
417 		wakeup(semap);
418 	}
419 	semap->holder = NOHOLDER;
420 }
421 
422 /*
423  * Worklist queue management.
424  * These routines require that the lock be held.
425  */
426 #ifndef /* NOT */ DEBUG
427 #define WORKLIST_INSERT(head, item) do {	\
428 	(item)->wk_state |= ONWORKLIST;		\
429 	LIST_INSERT_HEAD(head, item, wk_list);	\
430 } while (0)
431 #define WORKLIST_REMOVE(item) do {		\
432 	(item)->wk_state &= ~ONWORKLIST;	\
433 	LIST_REMOVE(item, wk_list);		\
434 } while (0)
435 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
436 
437 #else /* DEBUG */
438 static	void worklist_insert(struct workhead *, struct worklist *);
439 static	void worklist_remove(struct worklist *);
440 static	void workitem_free(struct worklist *, int);
441 
442 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
443 #define WORKLIST_REMOVE(item) worklist_remove(item)
444 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
445 
446 static void
447 worklist_insert(head, item)
448 	struct workhead *head;
449 	struct worklist *item;
450 {
451 
452 	if (lk.lkt_held == NOHOLDER)
453 		panic("worklist_insert: lock not held");
454 	if (item->wk_state & ONWORKLIST) {
455 		FREE_LOCK(&lk);
456 		panic("worklist_insert: already on list");
457 	}
458 	item->wk_state |= ONWORKLIST;
459 	LIST_INSERT_HEAD(head, item, wk_list);
460 }
461 
462 static void
463 worklist_remove(item)
464 	struct worklist *item;
465 {
466 
467 	if (lk.lkt_held == NOHOLDER)
468 		panic("worklist_remove: lock not held");
469 	if ((item->wk_state & ONWORKLIST) == 0) {
470 		FREE_LOCK(&lk);
471 		panic("worklist_remove: not on list");
472 	}
473 	item->wk_state &= ~ONWORKLIST;
474 	LIST_REMOVE(item, wk_list);
475 }
476 
477 static void
478 workitem_free(item, type)
479 	struct worklist *item;
480 	int type;
481 {
482 
483 	if (item->wk_state & ONWORKLIST) {
484 		if (lk.lkt_held != NOHOLDER)
485 			FREE_LOCK(&lk);
486 		panic("workitem_free: still on list");
487 	}
488 	if (item->wk_type != type) {
489 		if (lk.lkt_held != NOHOLDER)
490 			FREE_LOCK(&lk);
491 		panic("workitem_free: type mismatch");
492 	}
493 	FREE(item, DtoM(type));
494 }
495 #endif /* DEBUG */
496 
497 /*
498  * Workitem queue management
499  */
500 static struct workhead softdep_workitem_pending;
501 static int num_on_worklist;	/* number of worklist items to be processed */
502 static int softdep_worklist_busy; /* 1 => trying to do unmount */
503 static int softdep_worklist_req; /* serialized waiters */
504 static int max_softdeps;	/* maximum number of structs before slowdown */
505 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
506 static int proc_waiting;	/* tracks whether we have a timeout posted */
507 static int *stat_countp;	/* statistic to count in proc_waiting timeout */
508 static struct callout_handle handle; /* handle on posted proc_waiting timeout */
509 static struct thread *filesys_syncer; /* proc of filesystem syncer process */
510 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
511 #define FLUSH_INODES		1
512 static int req_clear_remove;	/* syncer process flush some freeblks */
513 #define FLUSH_REMOVE		2
514 #define FLUSH_REMOVE_WAIT	3
515 /*
516  * runtime statistics
517  */
518 static int stat_worklist_push;	/* number of worklist cleanups */
519 static int stat_blk_limit_push;	/* number of times block limit neared */
520 static int stat_ino_limit_push;	/* number of times inode limit neared */
521 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
522 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
523 static int stat_sync_limit_hit;	/* number of synchronous slowdowns imposed */
524 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
525 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
526 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
527 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
528 #ifdef DEBUG
529 #include <vm/vm.h>
530 #include <sys/sysctl.h>
531 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
532 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
533 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
534 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
535 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
536 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
537 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
538 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
539 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
540 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
541 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
542 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
543 #endif /* DEBUG */
544 
545 /*
546  * Add an item to the end of the work queue.
547  * This routine requires that the lock be held.
548  * This is the only routine that adds items to the list.
549  * The following routine is the only one that removes items
550  * and does so in order from first to last.
551  */
552 static void
553 add_to_worklist(wk)
554 	struct worklist *wk;
555 {
556 	static struct worklist *worklist_tail;
557 
558 	if (wk->wk_state & ONWORKLIST) {
559 		if (lk.lkt_held != NOHOLDER)
560 			FREE_LOCK(&lk);
561 		panic("add_to_worklist: already on list");
562 	}
563 	wk->wk_state |= ONWORKLIST;
564 	if (LIST_FIRST(&softdep_workitem_pending) == NULL)
565 		LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
566 	else
567 		LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
568 	worklist_tail = wk;
569 	num_on_worklist += 1;
570 }
571 
572 /*
573  * Process that runs once per second to handle items in the background queue.
574  *
575  * Note that we ensure that everything is done in the order in which they
576  * appear in the queue. The code below depends on this property to ensure
577  * that blocks of a file are freed before the inode itself is freed. This
578  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
579  * until all the old ones have been purged from the dependency lists.
580  */
581 int
582 softdep_process_worklist(matchmnt)
583 	struct mount *matchmnt;
584 {
585 	struct thread *td = curthread;
586 	int cnt, matchcnt, loopcount;
587 	long starttime;
588 
589 	/*
590 	 * Record the process identifier of our caller so that we can give
591 	 * this process preferential treatment in request_cleanup below.
592 	 */
593 	filesys_syncer = td;
594 	matchcnt = 0;
595 
596 	/*
597 	 * There is no danger of having multiple processes run this
598 	 * code, but we have to single-thread it when softdep_flushfiles()
599 	 * is in operation to get an accurate count of the number of items
600 	 * related to its mount point that are in the list.
601 	 */
602 	if (matchmnt == NULL) {
603 		if (softdep_worklist_busy < 0)
604 			return(-1);
605 		softdep_worklist_busy += 1;
606 	}
607 
608 	/*
609 	 * If requested, try removing inode or removal dependencies.
610 	 */
611 	if (req_clear_inodedeps) {
612 		clear_inodedeps(td);
613 		req_clear_inodedeps -= 1;
614 		wakeup_one(&proc_waiting);
615 	}
616 	if (req_clear_remove) {
617 		clear_remove(td);
618 		req_clear_remove -= 1;
619 		wakeup_one(&proc_waiting);
620 	}
621 	loopcount = 1;
622 	starttime = time_second;
623 	while (num_on_worklist > 0) {
624 		if ((cnt = process_worklist_item(matchmnt, 0)) == -1)
625 			break;
626 		else
627 			matchcnt += cnt;
628 
629 		/*
630 		 * If a umount operation wants to run the worklist
631 		 * accurately, abort.
632 		 */
633 		if (softdep_worklist_req && matchmnt == NULL) {
634 			matchcnt = -1;
635 			break;
636 		}
637 
638 		/*
639 		 * If requested, try removing inode or removal dependencies.
640 		 */
641 		if (req_clear_inodedeps) {
642 			clear_inodedeps(td);
643 			req_clear_inodedeps -= 1;
644 			wakeup_one(&proc_waiting);
645 		}
646 		if (req_clear_remove) {
647 			clear_remove(td);
648 			req_clear_remove -= 1;
649 			wakeup_one(&proc_waiting);
650 		}
651 		/*
652 		 * We do not generally want to stop for buffer space, but if
653 		 * we are really being a buffer hog, we will stop and wait.
654 		 */
655 		if (loopcount++ % 128 == 0)
656 			bwillwrite();
657 		/*
658 		 * Never allow processing to run for more than one
659 		 * second. Otherwise the other syncer tasks may get
660 		 * excessively backlogged.
661 		 */
662 		if (starttime != time_second && matchmnt == NULL) {
663 			matchcnt = -1;
664 			break;
665 		}
666 	}
667 	if (matchmnt == NULL) {
668 		softdep_worklist_busy -= 1;
669 		if (softdep_worklist_req && softdep_worklist_busy == 0)
670 			wakeup(&softdep_worklist_req);
671 	}
672 	return (matchcnt);
673 }
674 
675 /*
676  * Process one item on the worklist.
677  */
678 static int
679 process_worklist_item(matchmnt, flags)
680 	struct mount *matchmnt;
681 	int flags;
682 {
683 	struct worklist *wk;
684 	struct mount *mp;
685 	struct vnode *vp;
686 	int matchcnt = 0;
687 
688 	ACQUIRE_LOCK(&lk);
689 	/*
690 	 * Normally we just process each item on the worklist in order.
691 	 * However, if we are in a situation where we cannot lock any
692 	 * inodes, we have to skip over any dirrem requests whose
693 	 * vnodes are resident and locked.
694 	 */
695 	vp = NULL;
696 	LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
697 		if (wk->wk_state & INPROGRESS)
698 			continue;
699 		if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
700 			break;
701 		wk->wk_state |= INPROGRESS;
702 		FREE_LOCK(&lk);
703 		VFS_VGET(WK_DIRREM(wk)->dm_mnt, WK_DIRREM(wk)->dm_oldinum,
704 		    LK_NOWAIT | LK_EXCLUSIVE, &vp);
705 		ACQUIRE_LOCK(&lk);
706 		wk->wk_state &= ~INPROGRESS;
707 		if (vp != NULL)
708 			break;
709 	}
710 	if (wk == 0) {
711 		FREE_LOCK(&lk);
712 		return (-1);
713 	}
714 	WORKLIST_REMOVE(wk);
715 	num_on_worklist -= 1;
716 	FREE_LOCK(&lk);
717 	switch (wk->wk_type) {
718 
719 	case D_DIRREM:
720 		/* removal of a directory entry */
721 		mp = WK_DIRREM(wk)->dm_mnt;
722 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
723 			panic("%s: dirrem on suspended filesystem",
724 				"process_worklist_item");
725 		if (mp == matchmnt)
726 			matchcnt += 1;
727 		handle_workitem_remove(WK_DIRREM(wk), vp);
728 		break;
729 
730 	case D_FREEBLKS:
731 		/* releasing blocks and/or fragments from a file */
732 		mp = WK_FREEBLKS(wk)->fb_mnt;
733 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
734 			panic("%s: freeblks on suspended filesystem",
735 				"process_worklist_item");
736 		if (mp == matchmnt)
737 			matchcnt += 1;
738 		handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT);
739 		break;
740 
741 	case D_FREEFRAG:
742 		/* releasing a fragment when replaced as a file grows */
743 		mp = WK_FREEFRAG(wk)->ff_mnt;
744 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
745 			panic("%s: freefrag on suspended filesystem",
746 				"process_worklist_item");
747 		if (mp == matchmnt)
748 			matchcnt += 1;
749 		handle_workitem_freefrag(WK_FREEFRAG(wk));
750 		break;
751 
752 	case D_FREEFILE:
753 		/* releasing an inode when its link count drops to 0 */
754 		mp = WK_FREEFILE(wk)->fx_mnt;
755 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
756 			panic("%s: freefile on suspended filesystem",
757 				"process_worklist_item");
758 		if (mp == matchmnt)
759 			matchcnt += 1;
760 		handle_workitem_freefile(WK_FREEFILE(wk));
761 		break;
762 
763 	default:
764 		panic("%s_process_worklist: Unknown type %s",
765 		    "softdep", TYPENAME(wk->wk_type));
766 		/* NOTREACHED */
767 	}
768 	return (matchcnt);
769 }
770 
771 /*
772  * Move dependencies from one buffer to another.
773  */
774 static void
775 softdep_move_dependencies(oldbp, newbp)
776 	struct buf *oldbp;
777 	struct buf *newbp;
778 {
779 	struct worklist *wk, *wktail;
780 
781 	if (LIST_FIRST(&newbp->b_dep) != NULL)
782 		panic("softdep_move_dependencies: need merge code");
783 	wktail = 0;
784 	ACQUIRE_LOCK(&lk);
785 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
786 		LIST_REMOVE(wk, wk_list);
787 		if (wktail == 0)
788 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
789 		else
790 			LIST_INSERT_AFTER(wktail, wk, wk_list);
791 		wktail = wk;
792 	}
793 	FREE_LOCK(&lk);
794 }
795 
796 /*
797  * Purge the work list of all items associated with a particular mount point.
798  */
799 int
800 softdep_flushworklist(oldmnt, countp, td)
801 	struct mount *oldmnt;
802 	int *countp;
803 	struct thread *td;
804 {
805 	struct vnode *devvp;
806 	int count, error = 0;
807 
808 	/*
809 	 * Await our turn to clear out the queue, then serialize access.
810 	 */
811 	while (softdep_worklist_busy) {
812 		softdep_worklist_req += 1;
813 		tsleep(&softdep_worklist_req, PRIBIO, "softflush", 0);
814 		softdep_worklist_req -= 1;
815 	}
816 	softdep_worklist_busy = -1;
817 	/*
818 	 * Alternately flush the block device associated with the mount
819 	 * point and process any dependencies that the flushing
820 	 * creates. We continue until no more worklist dependencies
821 	 * are found.
822 	 */
823 	*countp = 0;
824 	devvp = VFSTOUFS(oldmnt)->um_devvp;
825 	while ((count = softdep_process_worklist(oldmnt)) > 0) {
826 		*countp += count;
827 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td);
828 		error = VOP_FSYNC(devvp, td->td_ucred, MNT_WAIT, td);
829 		VOP_UNLOCK(devvp, 0, td);
830 		if (error)
831 			break;
832 	}
833 	softdep_worklist_busy = 0;
834 	if (softdep_worklist_req)
835 		wakeup(&softdep_worklist_req);
836 	return (error);
837 }
838 
839 /*
840  * Flush all vnodes and worklist items associated with a specified mount point.
841  */
842 int
843 softdep_flushfiles(oldmnt, flags, td)
844 	struct mount *oldmnt;
845 	int flags;
846 	struct thread *td;
847 {
848 	int error, count, loopcnt;
849 
850 	error = 0;
851 
852 	/*
853 	 * Alternately flush the vnodes associated with the mount
854 	 * point and process any dependencies that the flushing
855 	 * creates. In theory, this loop can happen at most twice,
856 	 * but we give it a few extra just to be sure.
857 	 */
858 	for (loopcnt = 10; loopcnt > 0; loopcnt--) {
859 		/*
860 		 * Do another flush in case any vnodes were brought in
861 		 * as part of the cleanup operations.
862 		 */
863 		if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
864 			break;
865 		if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 ||
866 		    count == 0)
867 			break;
868 	}
869 	/*
870 	 * If we are unmounting then it is an error to fail. If we
871 	 * are simply trying to downgrade to read-only, then filesystem
872 	 * activity can keep us busy forever, so we just fail with EBUSY.
873 	 */
874 	if (loopcnt == 0) {
875 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
876 			panic("softdep_flushfiles: looping");
877 		error = EBUSY;
878 	}
879 	return (error);
880 }
881 
882 /*
883  * Structure hashing.
884  *
885  * There are three types of structures that can be looked up:
886  *	1) pagedep structures identified by mount point, inode number,
887  *	   and logical block.
888  *	2) inodedep structures identified by mount point and inode number.
889  *	3) newblk structures identified by mount point and
890  *	   physical block number.
891  *
892  * The "pagedep" and "inodedep" dependency structures are hashed
893  * separately from the file blocks and inodes to which they correspond.
894  * This separation helps when the in-memory copy of an inode or
895  * file block must be replaced. It also obviates the need to access
896  * an inode or file page when simply updating (or de-allocating)
897  * dependency structures. Lookup of newblk structures is needed to
898  * find newly allocated blocks when trying to associate them with
899  * their allocdirect or allocindir structure.
900  *
901  * The lookup routines optionally create and hash a new instance when
902  * an existing entry is not found.
903  */
904 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
905 #define NODELAY		0x0002	/* cannot do background work */
906 
907 /*
908  * Structures and routines associated with pagedep caching.
909  */
910 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
911 u_long	pagedep_hash;		/* size of hash table - 1 */
912 #define	PAGEDEP_HASH(mp, inum, lbn) \
913 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
914 	    pagedep_hash])
915 static struct sema pagedep_in_progress;
916 
917 /*
918  * Look up a pagedep. Return 1 if found, 0 if not found or found
919  * when asked to allocate but not associated with any buffer.
920  * If not found, allocate if DEPALLOC flag is passed.
921  * Found or allocated entry is returned in pagedeppp.
922  * This routine must be called with splbio interrupts blocked.
923  */
924 static int
925 pagedep_lookup(ip, lbn, flags, pagedeppp)
926 	struct inode *ip;
927 	ufs_lbn_t lbn;
928 	int flags;
929 	struct pagedep **pagedeppp;
930 {
931 	struct pagedep *pagedep;
932 	struct pagedep_hashhead *pagedephd;
933 	struct mount *mp;
934 	int i;
935 
936 #ifdef DEBUG
937 	if (lk.lkt_held == NOHOLDER)
938 		panic("pagedep_lookup: lock not held");
939 #endif
940 	mp = ITOV(ip)->v_mount;
941 	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
942 top:
943 	LIST_FOREACH(pagedep, pagedephd, pd_hash)
944 		if (ip->i_number == pagedep->pd_ino &&
945 		    lbn == pagedep->pd_lbn &&
946 		    mp == pagedep->pd_mnt)
947 			break;
948 	if (pagedep) {
949 		*pagedeppp = pagedep;
950 		if ((flags & DEPALLOC) != 0 &&
951 		    (pagedep->pd_state & ONWORKLIST) == 0)
952 			return (0);
953 		return (1);
954 	}
955 	if ((flags & DEPALLOC) == 0) {
956 		*pagedeppp = NULL;
957 		return (0);
958 	}
959 	if (sema_get(&pagedep_in_progress, &lk) == 0) {
960 		ACQUIRE_LOCK(&lk);
961 		goto top;
962 	}
963 	MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
964 		M_SOFTDEP_FLAGS|M_ZERO);
965 	pagedep->pd_list.wk_type = D_PAGEDEP;
966 	pagedep->pd_mnt = mp;
967 	pagedep->pd_ino = ip->i_number;
968 	pagedep->pd_lbn = lbn;
969 	LIST_INIT(&pagedep->pd_dirremhd);
970 	LIST_INIT(&pagedep->pd_pendinghd);
971 	for (i = 0; i < DAHASHSZ; i++)
972 		LIST_INIT(&pagedep->pd_diraddhd[i]);
973 	ACQUIRE_LOCK(&lk);
974 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
975 	sema_release(&pagedep_in_progress);
976 	*pagedeppp = pagedep;
977 	return (0);
978 }
979 
980 /*
981  * Structures and routines associated with inodedep caching.
982  */
983 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
984 static u_long	inodedep_hash;	/* size of hash table - 1 */
985 static long	num_inodedep;	/* number of inodedep allocated */
986 #define	INODEDEP_HASH(fs, inum) \
987       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
988 static struct sema inodedep_in_progress;
989 
990 /*
991  * Look up a inodedep. Return 1 if found, 0 if not found.
992  * If not found, allocate if DEPALLOC flag is passed.
993  * Found or allocated entry is returned in inodedeppp.
994  * This routine must be called with splbio interrupts blocked.
995  */
996 static int
997 inodedep_lookup(fs, inum, flags, inodedeppp)
998 	struct fs *fs;
999 	ino_t inum;
1000 	int flags;
1001 	struct inodedep **inodedeppp;
1002 {
1003 	struct inodedep *inodedep;
1004 	struct inodedep_hashhead *inodedephd;
1005 	int firsttry;
1006 
1007 #ifdef DEBUG
1008 	if (lk.lkt_held == NOHOLDER)
1009 		panic("inodedep_lookup: lock not held");
1010 #endif
1011 	firsttry = 1;
1012 	inodedephd = INODEDEP_HASH(fs, inum);
1013 top:
1014 	LIST_FOREACH(inodedep, inodedephd, id_hash)
1015 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
1016 			break;
1017 	if (inodedep) {
1018 		*inodedeppp = inodedep;
1019 		return (1);
1020 	}
1021 	if ((flags & DEPALLOC) == 0) {
1022 		*inodedeppp = NULL;
1023 		return (0);
1024 	}
1025 	/*
1026 	 * If we are over our limit, try to improve the situation.
1027 	 */
1028 	if (num_inodedep > max_softdeps && firsttry && (flags & NODELAY) == 0 &&
1029 	    request_cleanup(FLUSH_INODES, 1)) {
1030 		firsttry = 0;
1031 		goto top;
1032 	}
1033 	if (sema_get(&inodedep_in_progress, &lk) == 0) {
1034 		ACQUIRE_LOCK(&lk);
1035 		goto top;
1036 	}
1037 	num_inodedep += 1;
1038 	MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1039 		M_INODEDEP, M_SOFTDEP_FLAGS);
1040 	inodedep->id_list.wk_type = D_INODEDEP;
1041 	inodedep->id_fs = fs;
1042 	inodedep->id_ino = inum;
1043 	inodedep->id_state = ALLCOMPLETE;
1044 	inodedep->id_nlinkdelta = 0;
1045 	inodedep->id_savedino1 = NULL;
1046 	inodedep->id_savedsize = -1;
1047 	inodedep->id_savedextsize = -1;
1048 	inodedep->id_buf = NULL;
1049 	LIST_INIT(&inodedep->id_pendinghd);
1050 	LIST_INIT(&inodedep->id_inowait);
1051 	LIST_INIT(&inodedep->id_bufwait);
1052 	TAILQ_INIT(&inodedep->id_inoupdt);
1053 	TAILQ_INIT(&inodedep->id_newinoupdt);
1054 	TAILQ_INIT(&inodedep->id_extupdt);
1055 	TAILQ_INIT(&inodedep->id_newextupdt);
1056 	ACQUIRE_LOCK(&lk);
1057 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1058 	sema_release(&inodedep_in_progress);
1059 	*inodedeppp = inodedep;
1060 	return (0);
1061 }
1062 
1063 /*
1064  * Structures and routines associated with newblk caching.
1065  */
1066 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1067 u_long	newblk_hash;		/* size of hash table - 1 */
1068 #define	NEWBLK_HASH(fs, inum) \
1069 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1070 static struct sema newblk_in_progress;
1071 
1072 /*
1073  * Look up a newblk. Return 1 if found, 0 if not found.
1074  * If not found, allocate if DEPALLOC flag is passed.
1075  * Found or allocated entry is returned in newblkpp.
1076  */
1077 static int
1078 newblk_lookup(fs, newblkno, flags, newblkpp)
1079 	struct fs *fs;
1080 	ufs2_daddr_t newblkno;
1081 	int flags;
1082 	struct newblk **newblkpp;
1083 {
1084 	struct newblk *newblk;
1085 	struct newblk_hashhead *newblkhd;
1086 
1087 	newblkhd = NEWBLK_HASH(fs, newblkno);
1088 top:
1089 	LIST_FOREACH(newblk, newblkhd, nb_hash)
1090 		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1091 			break;
1092 	if (newblk) {
1093 		*newblkpp = newblk;
1094 		return (1);
1095 	}
1096 	if ((flags & DEPALLOC) == 0) {
1097 		*newblkpp = NULL;
1098 		return (0);
1099 	}
1100 	if (sema_get(&newblk_in_progress, 0) == 0)
1101 		goto top;
1102 	MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1103 		M_NEWBLK, M_SOFTDEP_FLAGS);
1104 	newblk->nb_state = 0;
1105 	newblk->nb_fs = fs;
1106 	newblk->nb_newblkno = newblkno;
1107 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1108 	sema_release(&newblk_in_progress);
1109 	*newblkpp = newblk;
1110 	return (0);
1111 }
1112 
1113 /*
1114  * Executed during filesystem system initialization before
1115  * mounting any filesystems.
1116  */
1117 void
1118 softdep_initialize()
1119 {
1120 
1121 	LIST_INIT(&mkdirlisthd);
1122 	LIST_INIT(&softdep_workitem_pending);
1123 	max_softdeps = desiredvnodes * 8;
1124 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1125 	    &pagedep_hash);
1126 	sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0);
1127 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1128 	sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0);
1129 	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1130 	sema_init(&newblk_in_progress, "newblk", PRIBIO, 0);
1131 
1132 	/* hooks through which the main kernel code calls us */
1133 	softdep_process_worklist_hook = softdep_process_worklist;
1134 	softdep_fsync_hook = softdep_fsync;
1135 
1136 	/* initialise bioops hack */
1137 	bioops.io_start = softdep_disk_io_initiation;
1138 	bioops.io_complete = softdep_disk_write_complete;
1139 	bioops.io_deallocate = softdep_deallocate_dependencies;
1140 	bioops.io_movedeps = softdep_move_dependencies;
1141 	bioops.io_countdeps = softdep_count_dependencies;
1142 }
1143 
1144 /*
1145  * Executed after all filesystems have been unmounted during
1146  * filesystem module unload.
1147  */
1148 void
1149 softdep_uninitialize()
1150 {
1151 
1152 	softdep_process_worklist_hook = NULL;
1153 	softdep_fsync_hook = NULL;
1154 	hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash);
1155 	hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash);
1156 	hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash);
1157 }
1158 
1159 /*
1160  * Called at mount time to notify the dependency code that a
1161  * filesystem wishes to use it.
1162  */
1163 int
1164 softdep_mount(devvp, mp, fs, cred)
1165 	struct vnode *devvp;
1166 	struct mount *mp;
1167 	struct fs *fs;
1168 	struct ucred *cred;
1169 {
1170 	struct csum_total cstotal;
1171 	struct cg *cgp;
1172 	struct buf *bp;
1173 	int error, cyl;
1174 
1175 	mp->mnt_flag &= ~MNT_ASYNC;
1176 	mp->mnt_flag |= MNT_SOFTDEP;
1177 	/*
1178 	 * When doing soft updates, the counters in the
1179 	 * superblock may have gotten out of sync, so we have
1180 	 * to scan the cylinder groups and recalculate them.
1181 	 */
1182 	if (fs->fs_clean != 0)
1183 		return (0);
1184 	bzero(&cstotal, sizeof cstotal);
1185 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1186 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1187 		    fs->fs_cgsize, cred, &bp)) != 0) {
1188 			brelse(bp);
1189 			return (error);
1190 		}
1191 		cgp = (struct cg *)bp->b_data;
1192 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1193 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1194 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1195 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1196 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
1197 		brelse(bp);
1198 	}
1199 #ifdef DEBUG
1200 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1201 		printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
1202 #endif
1203 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1204 	return (0);
1205 }
1206 
1207 /*
1208  * Protecting the freemaps (or bitmaps).
1209  *
1210  * To eliminate the need to execute fsck before mounting a filesystem
1211  * after a power failure, one must (conservatively) guarantee that the
1212  * on-disk copy of the bitmaps never indicate that a live inode or block is
1213  * free.  So, when a block or inode is allocated, the bitmap should be
1214  * updated (on disk) before any new pointers.  When a block or inode is
1215  * freed, the bitmap should not be updated until all pointers have been
1216  * reset.  The latter dependency is handled by the delayed de-allocation
1217  * approach described below for block and inode de-allocation.  The former
1218  * dependency is handled by calling the following procedure when a block or
1219  * inode is allocated. When an inode is allocated an "inodedep" is created
1220  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1221  * Each "inodedep" is also inserted into the hash indexing structure so
1222  * that any additional link additions can be made dependent on the inode
1223  * allocation.
1224  *
1225  * The ufs filesystem maintains a number of free block counts (e.g., per
1226  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1227  * in addition to the bitmaps.  These counts are used to improve efficiency
1228  * during allocation and therefore must be consistent with the bitmaps.
1229  * There is no convenient way to guarantee post-crash consistency of these
1230  * counts with simple update ordering, for two main reasons: (1) The counts
1231  * and bitmaps for a single cylinder group block are not in the same disk
1232  * sector.  If a disk write is interrupted (e.g., by power failure), one may
1233  * be written and the other not.  (2) Some of the counts are located in the
1234  * superblock rather than the cylinder group block. So, we focus our soft
1235  * updates implementation on protecting the bitmaps. When mounting a
1236  * filesystem, we recompute the auxiliary counts from the bitmaps.
1237  */
1238 
1239 /*
1240  * Called just after updating the cylinder group block to allocate an inode.
1241  */
1242 void
1243 softdep_setup_inomapdep(bp, ip, newinum)
1244 	struct buf *bp;		/* buffer for cylgroup block with inode map */
1245 	struct inode *ip;	/* inode related to allocation */
1246 	ino_t newinum;		/* new inode number being allocated */
1247 {
1248 	struct inodedep *inodedep;
1249 	struct bmsafemap *bmsafemap;
1250 
1251 	/*
1252 	 * Create a dependency for the newly allocated inode.
1253 	 * Panic if it already exists as something is seriously wrong.
1254 	 * Otherwise add it to the dependency list for the buffer holding
1255 	 * the cylinder group map from which it was allocated.
1256 	 */
1257 	ACQUIRE_LOCK(&lk);
1258 	if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1259 		FREE_LOCK(&lk);
1260 		panic("softdep_setup_inomapdep: found inode");
1261 	}
1262 	inodedep->id_buf = bp;
1263 	inodedep->id_state &= ~DEPCOMPLETE;
1264 	bmsafemap = bmsafemap_lookup(bp);
1265 	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1266 	FREE_LOCK(&lk);
1267 }
1268 
1269 /*
1270  * Called just after updating the cylinder group block to
1271  * allocate block or fragment.
1272  */
1273 void
1274 softdep_setup_blkmapdep(bp, fs, newblkno)
1275 	struct buf *bp;		/* buffer for cylgroup block with block map */
1276 	struct fs *fs;		/* filesystem doing allocation */
1277 	ufs2_daddr_t newblkno;	/* number of newly allocated block */
1278 {
1279 	struct newblk *newblk;
1280 	struct bmsafemap *bmsafemap;
1281 
1282 	/*
1283 	 * Create a dependency for the newly allocated block.
1284 	 * Add it to the dependency list for the buffer holding
1285 	 * the cylinder group map from which it was allocated.
1286 	 */
1287 	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1288 		panic("softdep_setup_blkmapdep: found block");
1289 	ACQUIRE_LOCK(&lk);
1290 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1291 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1292 	FREE_LOCK(&lk);
1293 }
1294 
1295 /*
1296  * Find the bmsafemap associated with a cylinder group buffer.
1297  * If none exists, create one. The buffer must be locked when
1298  * this routine is called and this routine must be called with
1299  * splbio interrupts blocked.
1300  */
1301 static struct bmsafemap *
1302 bmsafemap_lookup(bp)
1303 	struct buf *bp;
1304 {
1305 	struct bmsafemap *bmsafemap;
1306 	struct worklist *wk;
1307 
1308 #ifdef DEBUG
1309 	if (lk.lkt_held == NOHOLDER)
1310 		panic("bmsafemap_lookup: lock not held");
1311 #endif
1312 	LIST_FOREACH(wk, &bp->b_dep, wk_list)
1313 		if (wk->wk_type == D_BMSAFEMAP)
1314 			return (WK_BMSAFEMAP(wk));
1315 	FREE_LOCK(&lk);
1316 	MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1317 		M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1318 	bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1319 	bmsafemap->sm_list.wk_state = 0;
1320 	bmsafemap->sm_buf = bp;
1321 	LIST_INIT(&bmsafemap->sm_allocdirecthd);
1322 	LIST_INIT(&bmsafemap->sm_allocindirhd);
1323 	LIST_INIT(&bmsafemap->sm_inodedephd);
1324 	LIST_INIT(&bmsafemap->sm_newblkhd);
1325 	ACQUIRE_LOCK(&lk);
1326 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1327 	return (bmsafemap);
1328 }
1329 
1330 /*
1331  * Direct block allocation dependencies.
1332  *
1333  * When a new block is allocated, the corresponding disk locations must be
1334  * initialized (with zeros or new data) before the on-disk inode points to
1335  * them.  Also, the freemap from which the block was allocated must be
1336  * updated (on disk) before the inode's pointer. These two dependencies are
1337  * independent of each other and are needed for all file blocks and indirect
1338  * blocks that are pointed to directly by the inode.  Just before the
1339  * "in-core" version of the inode is updated with a newly allocated block
1340  * number, a procedure (below) is called to setup allocation dependency
1341  * structures.  These structures are removed when the corresponding
1342  * dependencies are satisfied or when the block allocation becomes obsolete
1343  * (i.e., the file is deleted, the block is de-allocated, or the block is a
1344  * fragment that gets upgraded).  All of these cases are handled in
1345  * procedures described later.
1346  *
1347  * When a file extension causes a fragment to be upgraded, either to a larger
1348  * fragment or to a full block, the on-disk location may change (if the
1349  * previous fragment could not simply be extended). In this case, the old
1350  * fragment must be de-allocated, but not until after the inode's pointer has
1351  * been updated. In most cases, this is handled by later procedures, which
1352  * will construct a "freefrag" structure to be added to the workitem queue
1353  * when the inode update is complete (or obsolete).  The main exception to
1354  * this is when an allocation occurs while a pending allocation dependency
1355  * (for the same block pointer) remains.  This case is handled in the main
1356  * allocation dependency setup procedure by immediately freeing the
1357  * unreferenced fragments.
1358  */
1359 void
1360 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1361 	struct inode *ip;	/* inode to which block is being added */
1362 	ufs_lbn_t lbn;		/* block pointer within inode */
1363 	ufs2_daddr_t newblkno;	/* disk block number being added */
1364 	ufs2_daddr_t oldblkno;	/* previous block number, 0 unless frag */
1365 	long newsize;		/* size of new block */
1366 	long oldsize;		/* size of new block */
1367 	struct buf *bp;		/* bp for allocated block */
1368 {
1369 	struct allocdirect *adp, *oldadp;
1370 	struct allocdirectlst *adphead;
1371 	struct bmsafemap *bmsafemap;
1372 	struct inodedep *inodedep;
1373 	struct pagedep *pagedep;
1374 	struct newblk *newblk;
1375 
1376 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1377 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1378 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1379 	adp->ad_lbn = lbn;
1380 	adp->ad_newblkno = newblkno;
1381 	adp->ad_oldblkno = oldblkno;
1382 	adp->ad_newsize = newsize;
1383 	adp->ad_oldsize = oldsize;
1384 	adp->ad_state = ATTACHED;
1385 	LIST_INIT(&adp->ad_newdirblk);
1386 	if (newblkno == oldblkno)
1387 		adp->ad_freefrag = NULL;
1388 	else
1389 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1390 
1391 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1392 		panic("softdep_setup_allocdirect: lost block");
1393 
1394 	ACQUIRE_LOCK(&lk);
1395 	inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1396 	adp->ad_inodedep = inodedep;
1397 
1398 	if (newblk->nb_state == DEPCOMPLETE) {
1399 		adp->ad_state |= DEPCOMPLETE;
1400 		adp->ad_buf = NULL;
1401 	} else {
1402 		bmsafemap = newblk->nb_bmsafemap;
1403 		adp->ad_buf = bmsafemap->sm_buf;
1404 		LIST_REMOVE(newblk, nb_deps);
1405 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1406 	}
1407 	LIST_REMOVE(newblk, nb_hash);
1408 	FREE(newblk, M_NEWBLK);
1409 
1410 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1411 	if (lbn >= NDADDR) {
1412 		/* allocating an indirect block */
1413 		if (oldblkno != 0) {
1414 			FREE_LOCK(&lk);
1415 			panic("softdep_setup_allocdirect: non-zero indir");
1416 		}
1417 	} else {
1418 		/*
1419 		 * Allocating a direct block.
1420 		 *
1421 		 * If we are allocating a directory block, then we must
1422 		 * allocate an associated pagedep to track additions and
1423 		 * deletions.
1424 		 */
1425 		if ((ip->i_mode & IFMT) == IFDIR &&
1426 		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1427 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1428 	}
1429 	/*
1430 	 * The list of allocdirects must be kept in sorted and ascending
1431 	 * order so that the rollback routines can quickly determine the
1432 	 * first uncommitted block (the size of the file stored on disk
1433 	 * ends at the end of the lowest committed fragment, or if there
1434 	 * are no fragments, at the end of the highest committed block).
1435 	 * Since files generally grow, the typical case is that the new
1436 	 * block is to be added at the end of the list. We speed this
1437 	 * special case by checking against the last allocdirect in the
1438 	 * list before laboriously traversing the list looking for the
1439 	 * insertion point.
1440 	 */
1441 	adphead = &inodedep->id_newinoupdt;
1442 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1443 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1444 		/* insert at end of list */
1445 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1446 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1447 			allocdirect_merge(adphead, adp, oldadp);
1448 		FREE_LOCK(&lk);
1449 		return;
1450 	}
1451 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1452 		if (oldadp->ad_lbn >= lbn)
1453 			break;
1454 	}
1455 	if (oldadp == NULL) {
1456 		FREE_LOCK(&lk);
1457 		panic("softdep_setup_allocdirect: lost entry");
1458 	}
1459 	/* insert in middle of list */
1460 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1461 	if (oldadp->ad_lbn == lbn)
1462 		allocdirect_merge(adphead, adp, oldadp);
1463 	FREE_LOCK(&lk);
1464 }
1465 
1466 /*
1467  * Replace an old allocdirect dependency with a newer one.
1468  * This routine must be called with splbio interrupts blocked.
1469  */
1470 static void
1471 allocdirect_merge(adphead, newadp, oldadp)
1472 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
1473 	struct allocdirect *newadp;	/* allocdirect being added */
1474 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
1475 {
1476 	struct worklist *wk;
1477 	struct freefrag *freefrag;
1478 	struct newdirblk *newdirblk;
1479 
1480 #ifdef DEBUG
1481 	if (lk.lkt_held == NOHOLDER)
1482 		panic("allocdirect_merge: lock not held");
1483 #endif
1484 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1485 	    newadp->ad_oldsize != oldadp->ad_newsize ||
1486 	    newadp->ad_lbn >= NDADDR) {
1487 		FREE_LOCK(&lk);
1488 		panic("%s %jd != new %jd || old size %ld != new %ld",
1489 		    "allocdirect_merge: old blkno",
1490 		    (intmax_t)newadp->ad_oldblkno,
1491 		    (intmax_t)oldadp->ad_newblkno,
1492 		    newadp->ad_oldsize, oldadp->ad_newsize);
1493 	}
1494 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
1495 	newadp->ad_oldsize = oldadp->ad_oldsize;
1496 	/*
1497 	 * If the old dependency had a fragment to free or had never
1498 	 * previously had a block allocated, then the new dependency
1499 	 * can immediately post its freefrag and adopt the old freefrag.
1500 	 * This action is done by swapping the freefrag dependencies.
1501 	 * The new dependency gains the old one's freefrag, and the
1502 	 * old one gets the new one and then immediately puts it on
1503 	 * the worklist when it is freed by free_allocdirect. It is
1504 	 * not possible to do this swap when the old dependency had a
1505 	 * non-zero size but no previous fragment to free. This condition
1506 	 * arises when the new block is an extension of the old block.
1507 	 * Here, the first part of the fragment allocated to the new
1508 	 * dependency is part of the block currently claimed on disk by
1509 	 * the old dependency, so cannot legitimately be freed until the
1510 	 * conditions for the new dependency are fulfilled.
1511 	 */
1512 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1513 		freefrag = newadp->ad_freefrag;
1514 		newadp->ad_freefrag = oldadp->ad_freefrag;
1515 		oldadp->ad_freefrag = freefrag;
1516 	}
1517 	/*
1518 	 * If we are tracking a new directory-block allocation,
1519 	 * move it from the old allocdirect to the new allocdirect.
1520 	 */
1521 	if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
1522 		newdirblk = WK_NEWDIRBLK(wk);
1523 		WORKLIST_REMOVE(&newdirblk->db_list);
1524 		if (LIST_FIRST(&oldadp->ad_newdirblk) != NULL)
1525 			panic("allocdirect_merge: extra newdirblk");
1526 		WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list);
1527 	}
1528 	free_allocdirect(adphead, oldadp, 0);
1529 }
1530 
1531 /*
1532  * Allocate a new freefrag structure if needed.
1533  */
1534 static struct freefrag *
1535 newfreefrag(ip, blkno, size)
1536 	struct inode *ip;
1537 	ufs2_daddr_t blkno;
1538 	long size;
1539 {
1540 	struct freefrag *freefrag;
1541 	struct fs *fs;
1542 
1543 	if (blkno == 0)
1544 		return (NULL);
1545 	fs = ip->i_fs;
1546 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1547 		panic("newfreefrag: frag size");
1548 	MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1549 		M_FREEFRAG, M_SOFTDEP_FLAGS);
1550 	freefrag->ff_list.wk_type = D_FREEFRAG;
1551 	freefrag->ff_state = 0;
1552 	freefrag->ff_inum = ip->i_number;
1553 	freefrag->ff_mnt = ITOV(ip)->v_mount;
1554 	freefrag->ff_blkno = blkno;
1555 	freefrag->ff_fragsize = size;
1556 	return (freefrag);
1557 }
1558 
1559 /*
1560  * This workitem de-allocates fragments that were replaced during
1561  * file block allocation.
1562  */
1563 static void
1564 handle_workitem_freefrag(freefrag)
1565 	struct freefrag *freefrag;
1566 {
1567 	struct ufsmount *ump = VFSTOUFS(freefrag->ff_mnt);
1568 
1569 	ffs_blkfree(ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
1570 	    freefrag->ff_fragsize, freefrag->ff_inum);
1571 	FREE(freefrag, M_FREEFRAG);
1572 }
1573 
1574 /*
1575  * Set up a dependency structure for an external attributes data block.
1576  * This routine follows much of the structure of softdep_setup_allocdirect.
1577  * See the description of softdep_setup_allocdirect above for details.
1578  */
1579 void
1580 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1581 	struct inode *ip;
1582 	ufs_lbn_t lbn;
1583 	ufs2_daddr_t newblkno;
1584 	ufs2_daddr_t oldblkno;
1585 	long newsize;
1586 	long oldsize;
1587 	struct buf *bp;
1588 {
1589 	struct allocdirect *adp, *oldadp;
1590 	struct allocdirectlst *adphead;
1591 	struct bmsafemap *bmsafemap;
1592 	struct inodedep *inodedep;
1593 	struct newblk *newblk;
1594 
1595 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1596 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1597 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1598 	adp->ad_lbn = lbn;
1599 	adp->ad_newblkno = newblkno;
1600 	adp->ad_oldblkno = oldblkno;
1601 	adp->ad_newsize = newsize;
1602 	adp->ad_oldsize = oldsize;
1603 	adp->ad_state = ATTACHED | EXTDATA;
1604 	LIST_INIT(&adp->ad_newdirblk);
1605 	if (newblkno == oldblkno)
1606 		adp->ad_freefrag = NULL;
1607 	else
1608 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1609 
1610 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1611 		panic("softdep_setup_allocext: lost block");
1612 
1613 	ACQUIRE_LOCK(&lk);
1614 	inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1615 	adp->ad_inodedep = inodedep;
1616 
1617 	if (newblk->nb_state == DEPCOMPLETE) {
1618 		adp->ad_state |= DEPCOMPLETE;
1619 		adp->ad_buf = NULL;
1620 	} else {
1621 		bmsafemap = newblk->nb_bmsafemap;
1622 		adp->ad_buf = bmsafemap->sm_buf;
1623 		LIST_REMOVE(newblk, nb_deps);
1624 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1625 	}
1626 	LIST_REMOVE(newblk, nb_hash);
1627 	FREE(newblk, M_NEWBLK);
1628 
1629 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1630 	if (lbn >= NXADDR) {
1631 		FREE_LOCK(&lk);
1632 		panic("softdep_setup_allocext: lbn %lld > NXADDR",
1633 		    (long long)lbn);
1634 	}
1635 	/*
1636 	 * The list of allocdirects must be kept in sorted and ascending
1637 	 * order so that the rollback routines can quickly determine the
1638 	 * first uncommitted block (the size of the file stored on disk
1639 	 * ends at the end of the lowest committed fragment, or if there
1640 	 * are no fragments, at the end of the highest committed block).
1641 	 * Since files generally grow, the typical case is that the new
1642 	 * block is to be added at the end of the list. We speed this
1643 	 * special case by checking against the last allocdirect in the
1644 	 * list before laboriously traversing the list looking for the
1645 	 * insertion point.
1646 	 */
1647 	adphead = &inodedep->id_newextupdt;
1648 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1649 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1650 		/* insert at end of list */
1651 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1652 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1653 			allocdirect_merge(adphead, adp, oldadp);
1654 		FREE_LOCK(&lk);
1655 		return;
1656 	}
1657 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1658 		if (oldadp->ad_lbn >= lbn)
1659 			break;
1660 	}
1661 	if (oldadp == NULL) {
1662 		FREE_LOCK(&lk);
1663 		panic("softdep_setup_allocext: lost entry");
1664 	}
1665 	/* insert in middle of list */
1666 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1667 	if (oldadp->ad_lbn == lbn)
1668 		allocdirect_merge(adphead, adp, oldadp);
1669 	FREE_LOCK(&lk);
1670 }
1671 
1672 /*
1673  * Indirect block allocation dependencies.
1674  *
1675  * The same dependencies that exist for a direct block also exist when
1676  * a new block is allocated and pointed to by an entry in a block of
1677  * indirect pointers. The undo/redo states described above are also
1678  * used here. Because an indirect block contains many pointers that
1679  * may have dependencies, a second copy of the entire in-memory indirect
1680  * block is kept. The buffer cache copy is always completely up-to-date.
1681  * The second copy, which is used only as a source for disk writes,
1682  * contains only the safe pointers (i.e., those that have no remaining
1683  * update dependencies). The second copy is freed when all pointers
1684  * are safe. The cache is not allowed to replace indirect blocks with
1685  * pending update dependencies. If a buffer containing an indirect
1686  * block with dependencies is written, these routines will mark it
1687  * dirty again. It can only be successfully written once all the
1688  * dependencies are removed. The ffs_fsync routine in conjunction with
1689  * softdep_sync_metadata work together to get all the dependencies
1690  * removed so that a file can be successfully written to disk. Three
1691  * procedures are used when setting up indirect block pointer
1692  * dependencies. The division is necessary because of the organization
1693  * of the "balloc" routine and because of the distinction between file
1694  * pages and file metadata blocks.
1695  */
1696 
1697 /*
1698  * Allocate a new allocindir structure.
1699  */
1700 static struct allocindir *
1701 newallocindir(ip, ptrno, newblkno, oldblkno)
1702 	struct inode *ip;	/* inode for file being extended */
1703 	int ptrno;		/* offset of pointer in indirect block */
1704 	ufs2_daddr_t newblkno;	/* disk block number being added */
1705 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1706 {
1707 	struct allocindir *aip;
1708 
1709 	MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1710 		M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO);
1711 	aip->ai_list.wk_type = D_ALLOCINDIR;
1712 	aip->ai_state = ATTACHED;
1713 	aip->ai_offset = ptrno;
1714 	aip->ai_newblkno = newblkno;
1715 	aip->ai_oldblkno = oldblkno;
1716 	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1717 	return (aip);
1718 }
1719 
1720 /*
1721  * Called just before setting an indirect block pointer
1722  * to a newly allocated file page.
1723  */
1724 void
1725 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1726 	struct inode *ip;	/* inode for file being extended */
1727 	ufs_lbn_t lbn;		/* allocated block number within file */
1728 	struct buf *bp;		/* buffer with indirect blk referencing page */
1729 	int ptrno;		/* offset of pointer in indirect block */
1730 	ufs2_daddr_t newblkno;	/* disk block number being added */
1731 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1732 	struct buf *nbp;	/* buffer holding allocated page */
1733 {
1734 	struct allocindir *aip;
1735 	struct pagedep *pagedep;
1736 
1737 	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1738 	ACQUIRE_LOCK(&lk);
1739 	/*
1740 	 * If we are allocating a directory page, then we must
1741 	 * allocate an associated pagedep to track additions and
1742 	 * deletions.
1743 	 */
1744 	if ((ip->i_mode & IFMT) == IFDIR &&
1745 	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1746 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1747 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1748 	FREE_LOCK(&lk);
1749 	setup_allocindir_phase2(bp, ip, aip);
1750 }
1751 
1752 /*
1753  * Called just before setting an indirect block pointer to a
1754  * newly allocated indirect block.
1755  */
1756 void
1757 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1758 	struct buf *nbp;	/* newly allocated indirect block */
1759 	struct inode *ip;	/* inode for file being extended */
1760 	struct buf *bp;		/* indirect block referencing allocated block */
1761 	int ptrno;		/* offset of pointer in indirect block */
1762 	ufs2_daddr_t newblkno;	/* disk block number being added */
1763 {
1764 	struct allocindir *aip;
1765 
1766 	aip = newallocindir(ip, ptrno, newblkno, 0);
1767 	ACQUIRE_LOCK(&lk);
1768 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1769 	FREE_LOCK(&lk);
1770 	setup_allocindir_phase2(bp, ip, aip);
1771 }
1772 
1773 /*
1774  * Called to finish the allocation of the "aip" allocated
1775  * by one of the two routines above.
1776  */
1777 static void
1778 setup_allocindir_phase2(bp, ip, aip)
1779 	struct buf *bp;		/* in-memory copy of the indirect block */
1780 	struct inode *ip;	/* inode for file being extended */
1781 	struct allocindir *aip;	/* allocindir allocated by the above routines */
1782 {
1783 	struct worklist *wk;
1784 	struct indirdep *indirdep, *newindirdep;
1785 	struct bmsafemap *bmsafemap;
1786 	struct allocindir *oldaip;
1787 	struct freefrag *freefrag;
1788 	struct newblk *newblk;
1789 	ufs2_daddr_t blkno;
1790 
1791 	if (bp->b_lblkno >= 0)
1792 		panic("setup_allocindir_phase2: not indir blk");
1793 	for (indirdep = NULL, newindirdep = NULL; ; ) {
1794 		ACQUIRE_LOCK(&lk);
1795 		LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1796 			if (wk->wk_type != D_INDIRDEP)
1797 				continue;
1798 			indirdep = WK_INDIRDEP(wk);
1799 			break;
1800 		}
1801 		if (indirdep == NULL && newindirdep) {
1802 			indirdep = newindirdep;
1803 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1804 			newindirdep = NULL;
1805 		}
1806 		FREE_LOCK(&lk);
1807 		if (indirdep) {
1808 			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1809 			    &newblk) == 0)
1810 				panic("setup_allocindir: lost block");
1811 			ACQUIRE_LOCK(&lk);
1812 			if (newblk->nb_state == DEPCOMPLETE) {
1813 				aip->ai_state |= DEPCOMPLETE;
1814 				aip->ai_buf = NULL;
1815 			} else {
1816 				bmsafemap = newblk->nb_bmsafemap;
1817 				aip->ai_buf = bmsafemap->sm_buf;
1818 				LIST_REMOVE(newblk, nb_deps);
1819 				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1820 				    aip, ai_deps);
1821 			}
1822 			LIST_REMOVE(newblk, nb_hash);
1823 			FREE(newblk, M_NEWBLK);
1824 			aip->ai_indirdep = indirdep;
1825 			/*
1826 			 * Check to see if there is an existing dependency
1827 			 * for this block. If there is, merge the old
1828 			 * dependency into the new one.
1829 			 */
1830 			if (aip->ai_oldblkno == 0)
1831 				oldaip = NULL;
1832 			else
1833 
1834 				LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1835 					if (oldaip->ai_offset == aip->ai_offset)
1836 						break;
1837 			freefrag = NULL;
1838 			if (oldaip != NULL) {
1839 				if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1840 					FREE_LOCK(&lk);
1841 					panic("setup_allocindir_phase2: blkno");
1842 				}
1843 				aip->ai_oldblkno = oldaip->ai_oldblkno;
1844 				freefrag = aip->ai_freefrag;
1845 				aip->ai_freefrag = oldaip->ai_freefrag;
1846 				oldaip->ai_freefrag = NULL;
1847 				free_allocindir(oldaip, NULL);
1848 			}
1849 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1850 			if (ip->i_ump->um_fstype == UFS1)
1851 				((ufs1_daddr_t *)indirdep->ir_savebp->b_data)
1852 				    [aip->ai_offset] = aip->ai_oldblkno;
1853 			else
1854 				((ufs2_daddr_t *)indirdep->ir_savebp->b_data)
1855 				    [aip->ai_offset] = aip->ai_oldblkno;
1856 			FREE_LOCK(&lk);
1857 			if (freefrag != NULL)
1858 				handle_workitem_freefrag(freefrag);
1859 		}
1860 		if (newindirdep) {
1861 			if (indirdep->ir_savebp != NULL)
1862 				brelse(newindirdep->ir_savebp);
1863 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1864 		}
1865 		if (indirdep)
1866 			break;
1867 		MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1868 			M_INDIRDEP, M_SOFTDEP_FLAGS);
1869 		newindirdep->ir_list.wk_type = D_INDIRDEP;
1870 		newindirdep->ir_state = ATTACHED;
1871 		if (ip->i_ump->um_fstype == UFS1)
1872 			newindirdep->ir_state |= UFS1FMT;
1873 		LIST_INIT(&newindirdep->ir_deplisthd);
1874 		LIST_INIT(&newindirdep->ir_donehd);
1875 		if (bp->b_blkno == bp->b_lblkno) {
1876 			ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
1877 			    NULL, NULL);
1878 			bp->b_blkno = blkno;
1879 		}
1880 		newindirdep->ir_savebp =
1881 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0);
1882 		BUF_KERNPROC(newindirdep->ir_savebp);
1883 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1884 	}
1885 }
1886 
1887 /*
1888  * Block de-allocation dependencies.
1889  *
1890  * When blocks are de-allocated, the on-disk pointers must be nullified before
1891  * the blocks are made available for use by other files.  (The true
1892  * requirement is that old pointers must be nullified before new on-disk
1893  * pointers are set.  We chose this slightly more stringent requirement to
1894  * reduce complexity.) Our implementation handles this dependency by updating
1895  * the inode (or indirect block) appropriately but delaying the actual block
1896  * de-allocation (i.e., freemap and free space count manipulation) until
1897  * after the updated versions reach stable storage.  After the disk is
1898  * updated, the blocks can be safely de-allocated whenever it is convenient.
1899  * This implementation handles only the common case of reducing a file's
1900  * length to zero. Other cases are handled by the conventional synchronous
1901  * write approach.
1902  *
1903  * The ffs implementation with which we worked double-checks
1904  * the state of the block pointers and file size as it reduces
1905  * a file's length.  Some of this code is replicated here in our
1906  * soft updates implementation.  The freeblks->fb_chkcnt field is
1907  * used to transfer a part of this information to the procedure
1908  * that eventually de-allocates the blocks.
1909  *
1910  * This routine should be called from the routine that shortens
1911  * a file's length, before the inode's size or block pointers
1912  * are modified. It will save the block pointer information for
1913  * later release and zero the inode so that the calling routine
1914  * can release it.
1915  */
1916 void
1917 softdep_setup_freeblocks(ip, length, flags)
1918 	struct inode *ip;	/* The inode whose length is to be reduced */
1919 	off_t length;		/* The new length for the file */
1920 	int flags;		/* IO_EXT and/or IO_NORMAL */
1921 {
1922 	struct freeblks *freeblks;
1923 	struct inodedep *inodedep;
1924 	struct allocdirect *adp;
1925 	struct vnode *vp;
1926 	struct buf *bp;
1927 	struct fs *fs;
1928 	ufs2_daddr_t extblocks, datablocks;
1929 	int i, delay, error;
1930 
1931 	fs = ip->i_fs;
1932 	if (length != 0)
1933 		panic("softdep_setup_freeblocks: non-zero length");
1934 	MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1935 		M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
1936 	freeblks->fb_list.wk_type = D_FREEBLKS;
1937 	freeblks->fb_uid = ip->i_uid;
1938 	freeblks->fb_previousinum = ip->i_number;
1939 	freeblks->fb_devvp = ip->i_devvp;
1940 	freeblks->fb_mnt = ITOV(ip)->v_mount;
1941 	extblocks = 0;
1942 	if (fs->fs_magic == FS_UFS2_MAGIC)
1943 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
1944 	datablocks = DIP(ip, i_blocks) - extblocks;
1945 	if ((flags & IO_NORMAL) == 0) {
1946 		freeblks->fb_oldsize = 0;
1947 		freeblks->fb_chkcnt = 0;
1948 	} else {
1949 		freeblks->fb_oldsize = ip->i_size;
1950 		ip->i_size = 0;
1951 		DIP(ip, i_size) = 0;
1952 		freeblks->fb_chkcnt = datablocks;
1953 		for (i = 0; i < NDADDR; i++) {
1954 			freeblks->fb_dblks[i] = DIP(ip, i_db[i]);
1955 			DIP(ip, i_db[i]) = 0;
1956 		}
1957 		for (i = 0; i < NIADDR; i++) {
1958 			freeblks->fb_iblks[i] = DIP(ip, i_ib[i]);
1959 			DIP(ip, i_ib[i]) = 0;
1960 		}
1961 		/*
1962 		 * If the file was removed, then the space being freed was
1963 		 * accounted for then (see softdep_filereleased()). If the
1964 		 * file is merely being truncated, then we account for it now.
1965 		 */
1966 		if ((ip->i_flag & IN_SPACECOUNTED) == 0)
1967 			fs->fs_pendingblocks += datablocks;
1968 	}
1969 	if ((flags & IO_EXT) == 0) {
1970 		freeblks->fb_oldextsize = 0;
1971 	} else {
1972 		freeblks->fb_oldextsize = ip->i_din2->di_extsize;
1973 		ip->i_din2->di_extsize = 0;
1974 		freeblks->fb_chkcnt += extblocks;
1975 		for (i = 0; i < NXADDR; i++) {
1976 			freeblks->fb_eblks[i] = ip->i_din2->di_extb[i];
1977 			ip->i_din2->di_extb[i] = 0;
1978 		}
1979 	}
1980 	DIP(ip, i_blocks) -= freeblks->fb_chkcnt;
1981 	/*
1982 	 * Push the zero'ed inode to to its disk buffer so that we are free
1983 	 * to delete its dependencies below. Once the dependencies are gone
1984 	 * the buffer can be safely released.
1985 	 */
1986 	if ((error = bread(ip->i_devvp,
1987 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
1988 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
1989 		brelse(bp);
1990 		softdep_error("softdep_setup_freeblocks", error);
1991 	}
1992 	if (ip->i_ump->um_fstype == UFS1)
1993 		*((struct ufs1_dinode *)bp->b_data +
1994 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
1995 	else
1996 		*((struct ufs2_dinode *)bp->b_data +
1997 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
1998 	/*
1999 	 * Find and eliminate any inode dependencies.
2000 	 */
2001 	ACQUIRE_LOCK(&lk);
2002 	(void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
2003 	if ((inodedep->id_state & IOSTARTED) != 0) {
2004 		FREE_LOCK(&lk);
2005 		panic("softdep_setup_freeblocks: inode busy");
2006 	}
2007 	/*
2008 	 * Add the freeblks structure to the list of operations that
2009 	 * must await the zero'ed inode being written to disk. If we
2010 	 * still have a bitmap dependency (delay == 0), then the inode
2011 	 * has never been written to disk, so we can process the
2012 	 * freeblks below once we have deleted the dependencies.
2013 	 */
2014 	delay = (inodedep->id_state & DEPCOMPLETE);
2015 	if (delay)
2016 		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
2017 	/*
2018 	 * Because the file length has been truncated to zero, any
2019 	 * pending block allocation dependency structures associated
2020 	 * with this inode are obsolete and can simply be de-allocated.
2021 	 * We must first merge the two dependency lists to get rid of
2022 	 * any duplicate freefrag structures, then purge the merged list.
2023 	 * If we still have a bitmap dependency, then the inode has never
2024 	 * been written to disk, so we can free any fragments without delay.
2025 	 */
2026 	if (flags & IO_NORMAL) {
2027 		merge_inode_lists(&inodedep->id_newinoupdt,
2028 		    &inodedep->id_inoupdt);
2029 		while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
2030 			free_allocdirect(&inodedep->id_inoupdt, adp, delay);
2031 	}
2032 	if (flags & IO_EXT) {
2033 		merge_inode_lists(&inodedep->id_newextupdt,
2034 		    &inodedep->id_extupdt);
2035 		while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
2036 			free_allocdirect(&inodedep->id_extupdt, adp, delay);
2037 	}
2038 	FREE_LOCK(&lk);
2039 	bdwrite(bp);
2040 	/*
2041 	 * We must wait for any I/O in progress to finish so that
2042 	 * all potential buffers on the dirty list will be visible.
2043 	 * Once they are all there, walk the list and get rid of
2044 	 * any dependencies.
2045 	 */
2046 	vp = ITOV(ip);
2047 	ACQUIRE_LOCK(&lk);
2048 	drain_output(vp, 1);
2049 restart:
2050 	TAILQ_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
2051 		if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
2052 		    ((flags & IO_NORMAL) == 0 &&
2053 		      (bp->b_xflags & BX_ALTDATA) == 0))
2054 			continue;
2055 		if (getdirtybuf(&bp, MNT_WAIT) == 0)
2056 			goto restart;
2057 		(void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
2058 		deallocate_dependencies(bp, inodedep);
2059 		bp->b_flags |= B_INVAL | B_NOCACHE;
2060 		FREE_LOCK(&lk);
2061 		brelse(bp);
2062 		ACQUIRE_LOCK(&lk);
2063 		goto restart;
2064 	}
2065 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
2066 		(void) free_inodedep(inodedep);
2067 	FREE_LOCK(&lk);
2068 	/*
2069 	 * If the inode has never been written to disk (delay == 0),
2070 	 * then we can process the freeblks now that we have deleted
2071 	 * the dependencies.
2072 	 */
2073 	if (!delay)
2074 		handle_workitem_freeblocks(freeblks, 0);
2075 }
2076 
2077 /*
2078  * Reclaim any dependency structures from a buffer that is about to
2079  * be reallocated to a new vnode. The buffer must be locked, thus,
2080  * no I/O completion operations can occur while we are manipulating
2081  * its associated dependencies. The mutex is held so that other I/O's
2082  * associated with related dependencies do not occur.
2083  */
2084 static void
2085 deallocate_dependencies(bp, inodedep)
2086 	struct buf *bp;
2087 	struct inodedep *inodedep;
2088 {
2089 	struct worklist *wk;
2090 	struct indirdep *indirdep;
2091 	struct allocindir *aip;
2092 	struct pagedep *pagedep;
2093 	struct dirrem *dirrem;
2094 	struct diradd *dap;
2095 	int i;
2096 
2097 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2098 		switch (wk->wk_type) {
2099 
2100 		case D_INDIRDEP:
2101 			indirdep = WK_INDIRDEP(wk);
2102 			/*
2103 			 * None of the indirect pointers will ever be visible,
2104 			 * so they can simply be tossed. GOINGAWAY ensures
2105 			 * that allocated pointers will be saved in the buffer
2106 			 * cache until they are freed. Note that they will
2107 			 * only be able to be found by their physical address
2108 			 * since the inode mapping the logical address will
2109 			 * be gone. The save buffer used for the safe copy
2110 			 * was allocated in setup_allocindir_phase2 using
2111 			 * the physical address so it could be used for this
2112 			 * purpose. Hence we swap the safe copy with the real
2113 			 * copy, allowing the safe copy to be freed and holding
2114 			 * on to the real copy for later use in indir_trunc.
2115 			 */
2116 			if (indirdep->ir_state & GOINGAWAY) {
2117 				FREE_LOCK(&lk);
2118 				panic("deallocate_dependencies: already gone");
2119 			}
2120 			indirdep->ir_state |= GOINGAWAY;
2121 			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
2122 				free_allocindir(aip, inodedep);
2123 			if (bp->b_lblkno >= 0 ||
2124 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno) {
2125 				FREE_LOCK(&lk);
2126 				panic("deallocate_dependencies: not indir");
2127 			}
2128 			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2129 			    bp->b_bcount);
2130 			WORKLIST_REMOVE(wk);
2131 			WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
2132 			continue;
2133 
2134 		case D_PAGEDEP:
2135 			pagedep = WK_PAGEDEP(wk);
2136 			/*
2137 			 * None of the directory additions will ever be
2138 			 * visible, so they can simply be tossed.
2139 			 */
2140 			for (i = 0; i < DAHASHSZ; i++)
2141 				while ((dap =
2142 				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
2143 					free_diradd(dap);
2144 			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2145 				free_diradd(dap);
2146 			/*
2147 			 * Copy any directory remove dependencies to the list
2148 			 * to be processed after the zero'ed inode is written.
2149 			 * If the inode has already been written, then they
2150 			 * can be dumped directly onto the work list.
2151 			 */
2152 			LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2153 				LIST_REMOVE(dirrem, dm_next);
2154 				dirrem->dm_dirinum = pagedep->pd_ino;
2155 				if (inodedep == NULL ||
2156 				    (inodedep->id_state & ALLCOMPLETE) ==
2157 				     ALLCOMPLETE)
2158 					add_to_worklist(&dirrem->dm_list);
2159 				else
2160 					WORKLIST_INSERT(&inodedep->id_bufwait,
2161 					    &dirrem->dm_list);
2162 			}
2163 			if ((pagedep->pd_state & NEWBLOCK) != 0) {
2164 				LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list)
2165 					if (wk->wk_type == D_NEWDIRBLK &&
2166 					    WK_NEWDIRBLK(wk)->db_pagedep ==
2167 					      pagedep)
2168 						break;
2169 				if (wk != NULL) {
2170 					WORKLIST_REMOVE(wk);
2171 					free_newdirblk(WK_NEWDIRBLK(wk));
2172 				} else {
2173 					FREE_LOCK(&lk);
2174 					panic("deallocate_dependencies: "
2175 					      "lost pagedep");
2176 				}
2177 			}
2178 			WORKLIST_REMOVE(&pagedep->pd_list);
2179 			LIST_REMOVE(pagedep, pd_hash);
2180 			WORKITEM_FREE(pagedep, D_PAGEDEP);
2181 			continue;
2182 
2183 		case D_ALLOCINDIR:
2184 			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2185 			continue;
2186 
2187 		case D_ALLOCDIRECT:
2188 		case D_INODEDEP:
2189 			FREE_LOCK(&lk);
2190 			panic("deallocate_dependencies: Unexpected type %s",
2191 			    TYPENAME(wk->wk_type));
2192 			/* NOTREACHED */
2193 
2194 		default:
2195 			FREE_LOCK(&lk);
2196 			panic("deallocate_dependencies: Unknown type %s",
2197 			    TYPENAME(wk->wk_type));
2198 			/* NOTREACHED */
2199 		}
2200 	}
2201 }
2202 
2203 /*
2204  * Free an allocdirect. Generate a new freefrag work request if appropriate.
2205  * This routine must be called with splbio interrupts blocked.
2206  */
2207 static void
2208 free_allocdirect(adphead, adp, delay)
2209 	struct allocdirectlst *adphead;
2210 	struct allocdirect *adp;
2211 	int delay;
2212 {
2213 	struct newdirblk *newdirblk;
2214 	struct worklist *wk;
2215 
2216 #ifdef DEBUG
2217 	if (lk.lkt_held == NOHOLDER)
2218 		panic("free_allocdirect: lock not held");
2219 #endif
2220 	if ((adp->ad_state & DEPCOMPLETE) == 0)
2221 		LIST_REMOVE(adp, ad_deps);
2222 	TAILQ_REMOVE(adphead, adp, ad_next);
2223 	if ((adp->ad_state & COMPLETE) == 0)
2224 		WORKLIST_REMOVE(&adp->ad_list);
2225 	if (adp->ad_freefrag != NULL) {
2226 		if (delay)
2227 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2228 			    &adp->ad_freefrag->ff_list);
2229 		else
2230 			add_to_worklist(&adp->ad_freefrag->ff_list);
2231 	}
2232 	if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) {
2233 		newdirblk = WK_NEWDIRBLK(wk);
2234 		WORKLIST_REMOVE(&newdirblk->db_list);
2235 		if (LIST_FIRST(&adp->ad_newdirblk) != NULL)
2236 			panic("free_allocdirect: extra newdirblk");
2237 		if (delay)
2238 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2239 			    &newdirblk->db_list);
2240 		else
2241 			free_newdirblk(newdirblk);
2242 	}
2243 	WORKITEM_FREE(adp, D_ALLOCDIRECT);
2244 }
2245 
2246 /*
2247  * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
2248  * This routine must be called with splbio interrupts blocked.
2249  */
2250 static void
2251 free_newdirblk(newdirblk)
2252 	struct newdirblk *newdirblk;
2253 {
2254 	struct pagedep *pagedep;
2255 	struct diradd *dap;
2256 	int i;
2257 
2258 #ifdef DEBUG
2259 	if (lk.lkt_held == NOHOLDER)
2260 		panic("free_newdirblk: lock not held");
2261 #endif
2262 	/*
2263 	 * If the pagedep is still linked onto the directory buffer
2264 	 * dependency chain, then some of the entries on the
2265 	 * pd_pendinghd list may not be committed to disk yet. In
2266 	 * this case, we will simply clear the NEWBLOCK flag and
2267 	 * let the pd_pendinghd list be processed when the pagedep
2268 	 * is next written. If the pagedep is no longer on the buffer
2269 	 * dependency chain, then all the entries on the pd_pending
2270 	 * list are committed to disk and we can free them here.
2271 	 */
2272 	pagedep = newdirblk->db_pagedep;
2273 	pagedep->pd_state &= ~NEWBLOCK;
2274 	if ((pagedep->pd_state & ONWORKLIST) == 0)
2275 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
2276 			free_diradd(dap);
2277 	/*
2278 	 * If no dependencies remain, the pagedep will be freed.
2279 	 */
2280 	for (i = 0; i < DAHASHSZ; i++)
2281 		if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
2282 			break;
2283 	if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) {
2284 		LIST_REMOVE(pagedep, pd_hash);
2285 		WORKITEM_FREE(pagedep, D_PAGEDEP);
2286 	}
2287 	WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2288 }
2289 
2290 /*
2291  * Prepare an inode to be freed. The actual free operation is not
2292  * done until the zero'ed inode has been written to disk.
2293  */
2294 void
2295 softdep_freefile(pvp, ino, mode)
2296 	struct vnode *pvp;
2297 	ino_t ino;
2298 	int mode;
2299 {
2300 	struct inode *ip = VTOI(pvp);
2301 	struct inodedep *inodedep;
2302 	struct freefile *freefile;
2303 
2304 	/*
2305 	 * This sets up the inode de-allocation dependency.
2306 	 */
2307 	MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2308 		M_FREEFILE, M_SOFTDEP_FLAGS);
2309 	freefile->fx_list.wk_type = D_FREEFILE;
2310 	freefile->fx_list.wk_state = 0;
2311 	freefile->fx_mode = mode;
2312 	freefile->fx_oldinum = ino;
2313 	freefile->fx_devvp = ip->i_devvp;
2314 	freefile->fx_mnt = ITOV(ip)->v_mount;
2315 	if ((ip->i_flag & IN_SPACECOUNTED) == 0)
2316 		ip->i_fs->fs_pendinginodes += 1;
2317 
2318 	/*
2319 	 * If the inodedep does not exist, then the zero'ed inode has
2320 	 * been written to disk. If the allocated inode has never been
2321 	 * written to disk, then the on-disk inode is zero'ed. In either
2322 	 * case we can free the file immediately.
2323 	 */
2324 	ACQUIRE_LOCK(&lk);
2325 	if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2326 	    check_inode_unwritten(inodedep)) {
2327 		FREE_LOCK(&lk);
2328 		handle_workitem_freefile(freefile);
2329 		return;
2330 	}
2331 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2332 	FREE_LOCK(&lk);
2333 }
2334 
2335 /*
2336  * Check to see if an inode has never been written to disk. If
2337  * so free the inodedep and return success, otherwise return failure.
2338  * This routine must be called with splbio interrupts blocked.
2339  *
2340  * If we still have a bitmap dependency, then the inode has never
2341  * been written to disk. Drop the dependency as it is no longer
2342  * necessary since the inode is being deallocated. We set the
2343  * ALLCOMPLETE flags since the bitmap now properly shows that the
2344  * inode is not allocated. Even if the inode is actively being
2345  * written, it has been rolled back to its zero'ed state, so we
2346  * are ensured that a zero inode is what is on the disk. For short
2347  * lived files, this change will usually result in removing all the
2348  * dependencies from the inode so that it can be freed immediately.
2349  */
2350 static int
2351 check_inode_unwritten(inodedep)
2352 	struct inodedep *inodedep;
2353 {
2354 
2355 	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2356 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2357 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2358 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
2359 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2360 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2361 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
2362 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
2363 	    inodedep->id_nlinkdelta != 0)
2364 		return (0);
2365 	inodedep->id_state |= ALLCOMPLETE;
2366 	LIST_REMOVE(inodedep, id_deps);
2367 	inodedep->id_buf = NULL;
2368 	if (inodedep->id_state & ONWORKLIST)
2369 		WORKLIST_REMOVE(&inodedep->id_list);
2370 	if (inodedep->id_savedino1 != NULL) {
2371 		FREE(inodedep->id_savedino1, M_INODEDEP);
2372 		inodedep->id_savedino1 = NULL;
2373 	}
2374 	if (free_inodedep(inodedep) == 0) {
2375 		FREE_LOCK(&lk);
2376 		panic("check_inode_unwritten: busy inode");
2377 	}
2378 	return (1);
2379 }
2380 
2381 /*
2382  * Try to free an inodedep structure. Return 1 if it could be freed.
2383  */
2384 static int
2385 free_inodedep(inodedep)
2386 	struct inodedep *inodedep;
2387 {
2388 
2389 	if ((inodedep->id_state & ONWORKLIST) != 0 ||
2390 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2391 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2392 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2393 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
2394 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2395 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2396 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
2397 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
2398 	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL)
2399 		return (0);
2400 	LIST_REMOVE(inodedep, id_hash);
2401 	WORKITEM_FREE(inodedep, D_INODEDEP);
2402 	num_inodedep -= 1;
2403 	return (1);
2404 }
2405 
2406 /*
2407  * This workitem routine performs the block de-allocation.
2408  * The workitem is added to the pending list after the updated
2409  * inode block has been written to disk.  As mentioned above,
2410  * checks regarding the number of blocks de-allocated (compared
2411  * to the number of blocks allocated for the file) are also
2412  * performed in this function.
2413  */
2414 static void
2415 handle_workitem_freeblocks(freeblks, flags)
2416 	struct freeblks *freeblks;
2417 	int flags;
2418 {
2419 	struct inode *ip;
2420 	struct vnode *vp;
2421 	struct fs *fs;
2422 	int i, nblocks, level, bsize;
2423 	ufs2_daddr_t bn, blocksreleased = 0;
2424 	int error, allerror = 0;
2425 	ufs_lbn_t baselbns[NIADDR], tmpval;
2426 
2427 	fs = VFSTOUFS(freeblks->fb_mnt)->um_fs;
2428 	tmpval = 1;
2429 	baselbns[0] = NDADDR;
2430 	for (i = 1; i < NIADDR; i++) {
2431 		tmpval *= NINDIR(fs);
2432 		baselbns[i] = baselbns[i - 1] + tmpval;
2433 	}
2434 	nblocks = btodb(fs->fs_bsize);
2435 	blocksreleased = 0;
2436 	/*
2437 	 * Release all extended attribute blocks or frags.
2438 	 */
2439 	if (freeblks->fb_oldextsize > 0) {
2440 		for (i = (NXADDR - 1); i >= 0; i--) {
2441 			if ((bn = freeblks->fb_eblks[i]) == 0)
2442 				continue;
2443 			bsize = sblksize(fs, freeblks->fb_oldextsize, i);
2444 			ffs_blkfree(fs, freeblks->fb_devvp, bn, bsize,
2445 			    freeblks->fb_previousinum);
2446 			blocksreleased += btodb(bsize);
2447 		}
2448 	}
2449 	/*
2450 	 * Release all data blocks or frags.
2451 	 */
2452 	if (freeblks->fb_oldsize > 0) {
2453 		/*
2454 		 * Indirect blocks first.
2455 		 */
2456 		for (level = (NIADDR - 1); level >= 0; level--) {
2457 			if ((bn = freeblks->fb_iblks[level]) == 0)
2458 				continue;
2459 			if ((error = indir_trunc(freeblks, fsbtodb(fs, bn),
2460 			    level, baselbns[level], &blocksreleased)) == 0)
2461 				allerror = error;
2462 			ffs_blkfree(fs, freeblks->fb_devvp, bn, fs->fs_bsize,
2463 			    freeblks->fb_previousinum);
2464 			fs->fs_pendingblocks -= nblocks;
2465 			blocksreleased += nblocks;
2466 		}
2467 		/*
2468 		 * All direct blocks or frags.
2469 		 */
2470 		for (i = (NDADDR - 1); i >= 0; i--) {
2471 			if ((bn = freeblks->fb_dblks[i]) == 0)
2472 				continue;
2473 			bsize = sblksize(fs, freeblks->fb_oldsize, i);
2474 			ffs_blkfree(fs, freeblks->fb_devvp, bn, bsize,
2475 			    freeblks->fb_previousinum);
2476 			fs->fs_pendingblocks -= btodb(bsize);
2477 			blocksreleased += btodb(bsize);
2478 		}
2479 	}
2480 	/*
2481 	 * If we still have not finished background cleanup, then check
2482 	 * to see if the block count needs to be adjusted.
2483 	 */
2484 	if (freeblks->fb_chkcnt != blocksreleased &&
2485 	    (fs->fs_flags & FS_UNCLEAN) != 0 &&
2486 	    VFS_VGET(freeblks->fb_mnt, freeblks->fb_previousinum,
2487 	    (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) {
2488 		ip = VTOI(vp);
2489 		DIP(ip, i_blocks) += freeblks->fb_chkcnt - blocksreleased;
2490 		ip->i_flag |= IN_CHANGE;
2491 		vput(vp);
2492 	}
2493 
2494 #ifdef DIAGNOSTIC
2495 	if (freeblks->fb_chkcnt != blocksreleased &&
2496 	    ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0))
2497 		printf("handle_workitem_freeblocks: block count");
2498 	if (allerror)
2499 		softdep_error("handle_workitem_freeblks", allerror);
2500 #endif /* DIAGNOSTIC */
2501 
2502 	WORKITEM_FREE(freeblks, D_FREEBLKS);
2503 }
2504 
2505 /*
2506  * Release blocks associated with the inode ip and stored in the indirect
2507  * block dbn. If level is greater than SINGLE, the block is an indirect block
2508  * and recursive calls to indirtrunc must be used to cleanse other indirect
2509  * blocks.
2510  */
2511 static int
2512 indir_trunc(freeblks, dbn, level, lbn, countp)
2513 	struct freeblks *freeblks;
2514 	ufs2_daddr_t dbn;
2515 	int level;
2516 	ufs_lbn_t lbn;
2517 	ufs2_daddr_t *countp;
2518 {
2519 	struct buf *bp;
2520 	struct fs *fs;
2521 	struct worklist *wk;
2522 	struct indirdep *indirdep;
2523 	ufs1_daddr_t *bap1 = 0;
2524 	ufs2_daddr_t nb, *bap2 = 0;
2525 	ufs_lbn_t lbnadd;
2526 	int i, nblocks, ufs1fmt;
2527 	int error, allerror = 0;
2528 
2529 	fs = VFSTOUFS(freeblks->fb_mnt)->um_fs;
2530 	lbnadd = 1;
2531 	for (i = level; i > 0; i--)
2532 		lbnadd *= NINDIR(fs);
2533 	/*
2534 	 * Get buffer of block pointers to be freed. This routine is not
2535 	 * called until the zero'ed inode has been written, so it is safe
2536 	 * to free blocks as they are encountered. Because the inode has
2537 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
2538 	 * have to use the on-disk address and the block device for the
2539 	 * filesystem to look them up. If the file was deleted before its
2540 	 * indirect blocks were all written to disk, the routine that set
2541 	 * us up (deallocate_dependencies) will have arranged to leave
2542 	 * a complete copy of the indirect block in memory for our use.
2543 	 * Otherwise we have to read the blocks in from the disk.
2544 	 */
2545 	ACQUIRE_LOCK(&lk);
2546 	if ((bp = incore(freeblks->fb_devvp, dbn)) != NULL &&
2547 	    (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2548 		if (wk->wk_type != D_INDIRDEP ||
2549 		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2550 		    (indirdep->ir_state & GOINGAWAY) == 0) {
2551 			FREE_LOCK(&lk);
2552 			panic("indir_trunc: lost indirdep");
2553 		}
2554 		WORKLIST_REMOVE(wk);
2555 		WORKITEM_FREE(indirdep, D_INDIRDEP);
2556 		if (LIST_FIRST(&bp->b_dep) != NULL) {
2557 			FREE_LOCK(&lk);
2558 			panic("indir_trunc: dangling dep");
2559 		}
2560 		FREE_LOCK(&lk);
2561 	} else {
2562 		FREE_LOCK(&lk);
2563 		error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
2564 		    NOCRED, &bp);
2565 		if (error) {
2566 			brelse(bp);
2567 			return (error);
2568 		}
2569 	}
2570 	/*
2571 	 * Recursively free indirect blocks.
2572 	 */
2573 	if (VFSTOUFS(freeblks->fb_mnt)->um_fstype == UFS1) {
2574 		ufs1fmt = 1;
2575 		bap1 = (ufs1_daddr_t *)bp->b_data;
2576 	} else {
2577 		ufs1fmt = 0;
2578 		bap2 = (ufs2_daddr_t *)bp->b_data;
2579 	}
2580 	nblocks = btodb(fs->fs_bsize);
2581 	for (i = NINDIR(fs) - 1; i >= 0; i--) {
2582 		if (ufs1fmt)
2583 			nb = bap1[i];
2584 		else
2585 			nb = bap2[i];
2586 		if (nb == 0)
2587 			continue;
2588 		if (level != 0) {
2589 			if ((error = indir_trunc(freeblks, fsbtodb(fs, nb),
2590 			     level - 1, lbn + (i * lbnadd), countp)) != 0)
2591 				allerror = error;
2592 		}
2593 		ffs_blkfree(fs, freeblks->fb_devvp, nb, fs->fs_bsize,
2594 		    freeblks->fb_previousinum);
2595 		fs->fs_pendingblocks -= nblocks;
2596 		*countp += nblocks;
2597 	}
2598 	bp->b_flags |= B_INVAL | B_NOCACHE;
2599 	brelse(bp);
2600 	return (allerror);
2601 }
2602 
2603 /*
2604  * Free an allocindir.
2605  * This routine must be called with splbio interrupts blocked.
2606  */
2607 static void
2608 free_allocindir(aip, inodedep)
2609 	struct allocindir *aip;
2610 	struct inodedep *inodedep;
2611 {
2612 	struct freefrag *freefrag;
2613 
2614 #ifdef DEBUG
2615 	if (lk.lkt_held == NOHOLDER)
2616 		panic("free_allocindir: lock not held");
2617 #endif
2618 	if ((aip->ai_state & DEPCOMPLETE) == 0)
2619 		LIST_REMOVE(aip, ai_deps);
2620 	if (aip->ai_state & ONWORKLIST)
2621 		WORKLIST_REMOVE(&aip->ai_list);
2622 	LIST_REMOVE(aip, ai_next);
2623 	if ((freefrag = aip->ai_freefrag) != NULL) {
2624 		if (inodedep == NULL)
2625 			add_to_worklist(&freefrag->ff_list);
2626 		else
2627 			WORKLIST_INSERT(&inodedep->id_bufwait,
2628 			    &freefrag->ff_list);
2629 	}
2630 	WORKITEM_FREE(aip, D_ALLOCINDIR);
2631 }
2632 
2633 /*
2634  * Directory entry addition dependencies.
2635  *
2636  * When adding a new directory entry, the inode (with its incremented link
2637  * count) must be written to disk before the directory entry's pointer to it.
2638  * Also, if the inode is newly allocated, the corresponding freemap must be
2639  * updated (on disk) before the directory entry's pointer. These requirements
2640  * are met via undo/redo on the directory entry's pointer, which consists
2641  * simply of the inode number.
2642  *
2643  * As directory entries are added and deleted, the free space within a
2644  * directory block can become fragmented.  The ufs filesystem will compact
2645  * a fragmented directory block to make space for a new entry. When this
2646  * occurs, the offsets of previously added entries change. Any "diradd"
2647  * dependency structures corresponding to these entries must be updated with
2648  * the new offsets.
2649  */
2650 
2651 /*
2652  * This routine is called after the in-memory inode's link
2653  * count has been incremented, but before the directory entry's
2654  * pointer to the inode has been set.
2655  */
2656 int
2657 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
2658 	struct buf *bp;		/* buffer containing directory block */
2659 	struct inode *dp;	/* inode for directory */
2660 	off_t diroffset;	/* offset of new entry in directory */
2661 	ino_t newinum;		/* inode referenced by new directory entry */
2662 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
2663 	int isnewblk;		/* entry is in a newly allocated block */
2664 {
2665 	int offset;		/* offset of new entry within directory block */
2666 	ufs_lbn_t lbn;		/* block in directory containing new entry */
2667 	struct fs *fs;
2668 	struct diradd *dap;
2669 	struct allocdirect *adp;
2670 	struct pagedep *pagedep;
2671 	struct inodedep *inodedep;
2672 	struct newdirblk *newdirblk = 0;
2673 	struct mkdir *mkdir1, *mkdir2;
2674 
2675 	/*
2676 	 * Whiteouts have no dependencies.
2677 	 */
2678 	if (newinum == WINO) {
2679 		if (newdirbp != NULL)
2680 			bdwrite(newdirbp);
2681 		return (0);
2682 	}
2683 
2684 	fs = dp->i_fs;
2685 	lbn = lblkno(fs, diroffset);
2686 	offset = blkoff(fs, diroffset);
2687 	MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2688 		M_SOFTDEP_FLAGS|M_ZERO);
2689 	dap->da_list.wk_type = D_DIRADD;
2690 	dap->da_offset = offset;
2691 	dap->da_newinum = newinum;
2692 	dap->da_state = ATTACHED;
2693 	if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) {
2694 		MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk),
2695 		    M_NEWDIRBLK, M_SOFTDEP_FLAGS);
2696 		newdirblk->db_list.wk_type = D_NEWDIRBLK;
2697 		newdirblk->db_state = 0;
2698 	}
2699 	if (newdirbp == NULL) {
2700 		dap->da_state |= DEPCOMPLETE;
2701 		ACQUIRE_LOCK(&lk);
2702 	} else {
2703 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2704 		MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2705 		    M_SOFTDEP_FLAGS);
2706 		mkdir1->md_list.wk_type = D_MKDIR;
2707 		mkdir1->md_state = MKDIR_BODY;
2708 		mkdir1->md_diradd = dap;
2709 		MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2710 		    M_SOFTDEP_FLAGS);
2711 		mkdir2->md_list.wk_type = D_MKDIR;
2712 		mkdir2->md_state = MKDIR_PARENT;
2713 		mkdir2->md_diradd = dap;
2714 		/*
2715 		 * Dependency on "." and ".." being written to disk.
2716 		 */
2717 		mkdir1->md_buf = newdirbp;
2718 		ACQUIRE_LOCK(&lk);
2719 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2720 		WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2721 		FREE_LOCK(&lk);
2722 		bdwrite(newdirbp);
2723 		/*
2724 		 * Dependency on link count increase for parent directory
2725 		 */
2726 		ACQUIRE_LOCK(&lk);
2727 		if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0
2728 		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2729 			dap->da_state &= ~MKDIR_PARENT;
2730 			WORKITEM_FREE(mkdir2, D_MKDIR);
2731 		} else {
2732 			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2733 			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2734 		}
2735 	}
2736 	/*
2737 	 * Link into parent directory pagedep to await its being written.
2738 	 */
2739 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2740 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2741 	dap->da_pagedep = pagedep;
2742 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2743 	    da_pdlist);
2744 	/*
2745 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
2746 	 * is not yet written. If it is written, do the post-inode write
2747 	 * processing to put it on the id_pendinghd list.
2748 	 */
2749 	(void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2750 	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2751 		diradd_inode_written(dap, inodedep);
2752 	else
2753 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2754 	if (isnewblk) {
2755 		/*
2756 		 * Directories growing into indirect blocks are rare
2757 		 * enough and the frequency of new block allocation
2758 		 * in those cases even more rare, that we choose not
2759 		 * to bother tracking them. Rather we simply force the
2760 		 * new directory entry to disk.
2761 		 */
2762 		if (lbn >= NDADDR) {
2763 			FREE_LOCK(&lk);
2764 			/*
2765 			 * We only have a new allocation when at the
2766 			 * beginning of a new block, not when we are
2767 			 * expanding into an existing block.
2768 			 */
2769 			if (blkoff(fs, diroffset) == 0)
2770 				return (1);
2771 			return (0);
2772 		}
2773 		/*
2774 		 * We only have a new allocation when at the beginning
2775 		 * of a new fragment, not when we are expanding into an
2776 		 * existing fragment. Also, there is nothing to do if we
2777 		 * are already tracking this block.
2778 		 */
2779 		if (fragoff(fs, diroffset) != 0) {
2780 			FREE_LOCK(&lk);
2781 			return (0);
2782 		}
2783 		if ((pagedep->pd_state & NEWBLOCK) != 0) {
2784 			WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2785 			FREE_LOCK(&lk);
2786 			return (0);
2787 		}
2788 		/*
2789 		 * Find our associated allocdirect and have it track us.
2790 		 */
2791 		if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0)
2792 			panic("softdep_setup_directory_add: lost inodedep");
2793 		adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst);
2794 		if (adp == NULL || adp->ad_lbn != lbn) {
2795 			FREE_LOCK(&lk);
2796 			panic("softdep_setup_directory_add: lost entry");
2797 		}
2798 		pagedep->pd_state |= NEWBLOCK;
2799 		newdirblk->db_pagedep = pagedep;
2800 		WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list);
2801 	}
2802 	FREE_LOCK(&lk);
2803 	return (0);
2804 }
2805 
2806 /*
2807  * This procedure is called to change the offset of a directory
2808  * entry when compacting a directory block which must be owned
2809  * exclusively by the caller. Note that the actual entry movement
2810  * must be done in this procedure to ensure that no I/O completions
2811  * occur while the move is in progress.
2812  */
2813 void
2814 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2815 	struct inode *dp;	/* inode for directory */
2816 	caddr_t base;		/* address of dp->i_offset */
2817 	caddr_t oldloc;		/* address of old directory location */
2818 	caddr_t newloc;		/* address of new directory location */
2819 	int entrysize;		/* size of directory entry */
2820 {
2821 	int offset, oldoffset, newoffset;
2822 	struct pagedep *pagedep;
2823 	struct diradd *dap;
2824 	ufs_lbn_t lbn;
2825 
2826 	ACQUIRE_LOCK(&lk);
2827 	lbn = lblkno(dp->i_fs, dp->i_offset);
2828 	offset = blkoff(dp->i_fs, dp->i_offset);
2829 	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2830 		goto done;
2831 	oldoffset = offset + (oldloc - base);
2832 	newoffset = offset + (newloc - base);
2833 
2834 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2835 		if (dap->da_offset != oldoffset)
2836 			continue;
2837 		dap->da_offset = newoffset;
2838 		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2839 			break;
2840 		LIST_REMOVE(dap, da_pdlist);
2841 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2842 		    dap, da_pdlist);
2843 		break;
2844 	}
2845 	if (dap == NULL) {
2846 
2847 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2848 			if (dap->da_offset == oldoffset) {
2849 				dap->da_offset = newoffset;
2850 				break;
2851 			}
2852 		}
2853 	}
2854 done:
2855 	bcopy(oldloc, newloc, entrysize);
2856 	FREE_LOCK(&lk);
2857 }
2858 
2859 /*
2860  * Free a diradd dependency structure. This routine must be called
2861  * with splbio interrupts blocked.
2862  */
2863 static void
2864 free_diradd(dap)
2865 	struct diradd *dap;
2866 {
2867 	struct dirrem *dirrem;
2868 	struct pagedep *pagedep;
2869 	struct inodedep *inodedep;
2870 	struct mkdir *mkdir, *nextmd;
2871 
2872 #ifdef DEBUG
2873 	if (lk.lkt_held == NOHOLDER)
2874 		panic("free_diradd: lock not held");
2875 #endif
2876 	WORKLIST_REMOVE(&dap->da_list);
2877 	LIST_REMOVE(dap, da_pdlist);
2878 	if ((dap->da_state & DIRCHG) == 0) {
2879 		pagedep = dap->da_pagedep;
2880 	} else {
2881 		dirrem = dap->da_previous;
2882 		pagedep = dirrem->dm_pagedep;
2883 		dirrem->dm_dirinum = pagedep->pd_ino;
2884 		add_to_worklist(&dirrem->dm_list);
2885 	}
2886 	if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2887 	    0, &inodedep) != 0)
2888 		(void) free_inodedep(inodedep);
2889 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2890 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2891 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
2892 			if (mkdir->md_diradd != dap)
2893 				continue;
2894 			dap->da_state &= ~mkdir->md_state;
2895 			WORKLIST_REMOVE(&mkdir->md_list);
2896 			LIST_REMOVE(mkdir, md_mkdirs);
2897 			WORKITEM_FREE(mkdir, D_MKDIR);
2898 		}
2899 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2900 			FREE_LOCK(&lk);
2901 			panic("free_diradd: unfound ref");
2902 		}
2903 	}
2904 	WORKITEM_FREE(dap, D_DIRADD);
2905 }
2906 
2907 /*
2908  * Directory entry removal dependencies.
2909  *
2910  * When removing a directory entry, the entry's inode pointer must be
2911  * zero'ed on disk before the corresponding inode's link count is decremented
2912  * (possibly freeing the inode for re-use). This dependency is handled by
2913  * updating the directory entry but delaying the inode count reduction until
2914  * after the directory block has been written to disk. After this point, the
2915  * inode count can be decremented whenever it is convenient.
2916  */
2917 
2918 /*
2919  * This routine should be called immediately after removing
2920  * a directory entry.  The inode's link count should not be
2921  * decremented by the calling procedure -- the soft updates
2922  * code will do this task when it is safe.
2923  */
2924 void
2925 softdep_setup_remove(bp, dp, ip, isrmdir)
2926 	struct buf *bp;		/* buffer containing directory block */
2927 	struct inode *dp;	/* inode for the directory being modified */
2928 	struct inode *ip;	/* inode for directory entry being removed */
2929 	int isrmdir;		/* indicates if doing RMDIR */
2930 {
2931 	struct dirrem *dirrem, *prevdirrem;
2932 
2933 	/*
2934 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2935 	 */
2936 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2937 
2938 	/*
2939 	 * If the COMPLETE flag is clear, then there were no active
2940 	 * entries and we want to roll back to a zeroed entry until
2941 	 * the new inode is committed to disk. If the COMPLETE flag is
2942 	 * set then we have deleted an entry that never made it to
2943 	 * disk. If the entry we deleted resulted from a name change,
2944 	 * then the old name still resides on disk. We cannot delete
2945 	 * its inode (returned to us in prevdirrem) until the zeroed
2946 	 * directory entry gets to disk. The new inode has never been
2947 	 * referenced on the disk, so can be deleted immediately.
2948 	 */
2949 	if ((dirrem->dm_state & COMPLETE) == 0) {
2950 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2951 		    dm_next);
2952 		FREE_LOCK(&lk);
2953 	} else {
2954 		if (prevdirrem != NULL)
2955 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2956 			    prevdirrem, dm_next);
2957 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2958 		FREE_LOCK(&lk);
2959 		handle_workitem_remove(dirrem, NULL);
2960 	}
2961 }
2962 
2963 /*
2964  * Allocate a new dirrem if appropriate and return it along with
2965  * its associated pagedep. Called without a lock, returns with lock.
2966  */
2967 static long num_dirrem;		/* number of dirrem allocated */
2968 static struct dirrem *
2969 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2970 	struct buf *bp;		/* buffer containing directory block */
2971 	struct inode *dp;	/* inode for the directory being modified */
2972 	struct inode *ip;	/* inode for directory entry being removed */
2973 	int isrmdir;		/* indicates if doing RMDIR */
2974 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
2975 {
2976 	int offset;
2977 	ufs_lbn_t lbn;
2978 	struct diradd *dap;
2979 	struct dirrem *dirrem;
2980 	struct pagedep *pagedep;
2981 
2982 	/*
2983 	 * Whiteouts have no deletion dependencies.
2984 	 */
2985 	if (ip == NULL)
2986 		panic("newdirrem: whiteout");
2987 	/*
2988 	 * If we are over our limit, try to improve the situation.
2989 	 * Limiting the number of dirrem structures will also limit
2990 	 * the number of freefile and freeblks structures.
2991 	 */
2992 	if (num_dirrem > max_softdeps / 2)
2993 		(void) request_cleanup(FLUSH_REMOVE, 0);
2994 	num_dirrem += 1;
2995 	MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
2996 		M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO);
2997 	dirrem->dm_list.wk_type = D_DIRREM;
2998 	dirrem->dm_state = isrmdir ? RMDIR : 0;
2999 	dirrem->dm_mnt = ITOV(ip)->v_mount;
3000 	dirrem->dm_oldinum = ip->i_number;
3001 	*prevdirremp = NULL;
3002 
3003 	ACQUIRE_LOCK(&lk);
3004 	lbn = lblkno(dp->i_fs, dp->i_offset);
3005 	offset = blkoff(dp->i_fs, dp->i_offset);
3006 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3007 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3008 	dirrem->dm_pagedep = pagedep;
3009 	/*
3010 	 * Check for a diradd dependency for the same directory entry.
3011 	 * If present, then both dependencies become obsolete and can
3012 	 * be de-allocated. Check for an entry on both the pd_dirraddhd
3013 	 * list and the pd_pendinghd list.
3014 	 */
3015 
3016 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
3017 		if (dap->da_offset == offset)
3018 			break;
3019 	if (dap == NULL) {
3020 
3021 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
3022 			if (dap->da_offset == offset)
3023 				break;
3024 		if (dap == NULL)
3025 			return (dirrem);
3026 	}
3027 	/*
3028 	 * Must be ATTACHED at this point.
3029 	 */
3030 	if ((dap->da_state & ATTACHED) == 0) {
3031 		FREE_LOCK(&lk);
3032 		panic("newdirrem: not ATTACHED");
3033 	}
3034 	if (dap->da_newinum != ip->i_number) {
3035 		FREE_LOCK(&lk);
3036 		panic("newdirrem: inum %d should be %d",
3037 		    ip->i_number, dap->da_newinum);
3038 	}
3039 	/*
3040 	 * If we are deleting a changed name that never made it to disk,
3041 	 * then return the dirrem describing the previous inode (which
3042 	 * represents the inode currently referenced from this entry on disk).
3043 	 */
3044 	if ((dap->da_state & DIRCHG) != 0) {
3045 		*prevdirremp = dap->da_previous;
3046 		dap->da_state &= ~DIRCHG;
3047 		dap->da_pagedep = pagedep;
3048 	}
3049 	/*
3050 	 * We are deleting an entry that never made it to disk.
3051 	 * Mark it COMPLETE so we can delete its inode immediately.
3052 	 */
3053 	dirrem->dm_state |= COMPLETE;
3054 	free_diradd(dap);
3055 	return (dirrem);
3056 }
3057 
3058 /*
3059  * Directory entry change dependencies.
3060  *
3061  * Changing an existing directory entry requires that an add operation
3062  * be completed first followed by a deletion. The semantics for the addition
3063  * are identical to the description of adding a new entry above except
3064  * that the rollback is to the old inode number rather than zero. Once
3065  * the addition dependency is completed, the removal is done as described
3066  * in the removal routine above.
3067  */
3068 
3069 /*
3070  * This routine should be called immediately after changing
3071  * a directory entry.  The inode's link count should not be
3072  * decremented by the calling procedure -- the soft updates
3073  * code will perform this task when it is safe.
3074  */
3075 void
3076 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
3077 	struct buf *bp;		/* buffer containing directory block */
3078 	struct inode *dp;	/* inode for the directory being modified */
3079 	struct inode *ip;	/* inode for directory entry being removed */
3080 	ino_t newinum;		/* new inode number for changed entry */
3081 	int isrmdir;		/* indicates if doing RMDIR */
3082 {
3083 	int offset;
3084 	struct diradd *dap = NULL;
3085 	struct dirrem *dirrem, *prevdirrem;
3086 	struct pagedep *pagedep;
3087 	struct inodedep *inodedep;
3088 
3089 	offset = blkoff(dp->i_fs, dp->i_offset);
3090 
3091 	/*
3092 	 * Whiteouts do not need diradd dependencies.
3093 	 */
3094 	if (newinum != WINO) {
3095 		MALLOC(dap, struct diradd *, sizeof(struct diradd),
3096 		    M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
3097 		dap->da_list.wk_type = D_DIRADD;
3098 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
3099 		dap->da_offset = offset;
3100 		dap->da_newinum = newinum;
3101 	}
3102 
3103 	/*
3104 	 * Allocate a new dirrem and ACQUIRE_LOCK.
3105 	 */
3106 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3107 	pagedep = dirrem->dm_pagedep;
3108 	/*
3109 	 * The possible values for isrmdir:
3110 	 *	0 - non-directory file rename
3111 	 *	1 - directory rename within same directory
3112 	 *   inum - directory rename to new directory of given inode number
3113 	 * When renaming to a new directory, we are both deleting and
3114 	 * creating a new directory entry, so the link count on the new
3115 	 * directory should not change. Thus we do not need the followup
3116 	 * dirrem which is usually done in handle_workitem_remove. We set
3117 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
3118 	 * followup dirrem.
3119 	 */
3120 	if (isrmdir > 1)
3121 		dirrem->dm_state |= DIRCHG;
3122 
3123 	/*
3124 	 * Whiteouts have no additional dependencies,
3125 	 * so just put the dirrem on the correct list.
3126 	 */
3127 	if (newinum == WINO) {
3128 		if ((dirrem->dm_state & COMPLETE) == 0) {
3129 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
3130 			    dm_next);
3131 		} else {
3132 			dirrem->dm_dirinum = pagedep->pd_ino;
3133 			add_to_worklist(&dirrem->dm_list);
3134 		}
3135 		FREE_LOCK(&lk);
3136 		return;
3137 	}
3138 
3139 	/*
3140 	 * If the COMPLETE flag is clear, then there were no active
3141 	 * entries and we want to roll back to the previous inode until
3142 	 * the new inode is committed to disk. If the COMPLETE flag is
3143 	 * set, then we have deleted an entry that never made it to disk.
3144 	 * If the entry we deleted resulted from a name change, then the old
3145 	 * inode reference still resides on disk. Any rollback that we do
3146 	 * needs to be to that old inode (returned to us in prevdirrem). If
3147 	 * the entry we deleted resulted from a create, then there is
3148 	 * no entry on the disk, so we want to roll back to zero rather
3149 	 * than the uncommitted inode. In either of the COMPLETE cases we
3150 	 * want to immediately free the unwritten and unreferenced inode.
3151 	 */
3152 	if ((dirrem->dm_state & COMPLETE) == 0) {
3153 		dap->da_previous = dirrem;
3154 	} else {
3155 		if (prevdirrem != NULL) {
3156 			dap->da_previous = prevdirrem;
3157 		} else {
3158 			dap->da_state &= ~DIRCHG;
3159 			dap->da_pagedep = pagedep;
3160 		}
3161 		dirrem->dm_dirinum = pagedep->pd_ino;
3162 		add_to_worklist(&dirrem->dm_list);
3163 	}
3164 	/*
3165 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3166 	 * is not yet written. If it is written, do the post-inode write
3167 	 * processing to put it on the id_pendinghd list.
3168 	 */
3169 	if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
3170 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3171 		dap->da_state |= COMPLETE;
3172 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3173 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3174 	} else {
3175 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
3176 		    dap, da_pdlist);
3177 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3178 	}
3179 	FREE_LOCK(&lk);
3180 }
3181 
3182 /*
3183  * Called whenever the link count on an inode is changed.
3184  * It creates an inode dependency so that the new reference(s)
3185  * to the inode cannot be committed to disk until the updated
3186  * inode has been written.
3187  */
3188 void
3189 softdep_change_linkcnt(ip)
3190 	struct inode *ip;	/* the inode with the increased link count */
3191 {
3192 	struct inodedep *inodedep;
3193 
3194 	ACQUIRE_LOCK(&lk);
3195 	(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
3196 	if (ip->i_nlink < ip->i_effnlink) {
3197 		FREE_LOCK(&lk);
3198 		panic("softdep_change_linkcnt: bad delta");
3199 	}
3200 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3201 	FREE_LOCK(&lk);
3202 }
3203 
3204 /*
3205  * Called when the effective link count and the reference count
3206  * on an inode drops to zero. At this point there are no names
3207  * referencing the file in the filesystem and no active file
3208  * references. The space associated with the file will be freed
3209  * as soon as the necessary soft dependencies are cleared.
3210  */
3211 void
3212 softdep_releasefile(ip)
3213 	struct inode *ip;	/* inode with the zero effective link count */
3214 {
3215 	struct inodedep *inodedep;
3216 	struct fs *fs;
3217 	int extblocks;
3218 
3219 	if (ip->i_effnlink > 0)
3220 		panic("softdep_filerelease: file still referenced");
3221 	/*
3222 	 * We may be called several times as the real reference count
3223 	 * drops to zero. We only want to account for the space once.
3224 	 */
3225 	if (ip->i_flag & IN_SPACECOUNTED)
3226 		return;
3227 	/*
3228 	 * We have to deactivate a snapshot otherwise copyonwrites may
3229 	 * add blocks and the cleanup may remove blocks after we have
3230 	 * tried to account for them.
3231 	 */
3232 	if ((ip->i_flags & SF_SNAPSHOT) != 0)
3233 		ffs_snapremove(ITOV(ip));
3234 	/*
3235 	 * If we are tracking an nlinkdelta, we have to also remember
3236 	 * whether we accounted for the freed space yet.
3237 	 */
3238 	ACQUIRE_LOCK(&lk);
3239 	if ((inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep)))
3240 		inodedep->id_state |= SPACECOUNTED;
3241 	FREE_LOCK(&lk);
3242 	fs = ip->i_fs;
3243 	extblocks = 0;
3244 	if (fs->fs_magic == FS_UFS2_MAGIC)
3245 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
3246 	ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks;
3247 	ip->i_fs->fs_pendinginodes += 1;
3248 	ip->i_flag |= IN_SPACECOUNTED;
3249 }
3250 
3251 /*
3252  * This workitem decrements the inode's link count.
3253  * If the link count reaches zero, the file is removed.
3254  */
3255 static void
3256 handle_workitem_remove(dirrem, xp)
3257 	struct dirrem *dirrem;
3258 	struct vnode *xp;
3259 {
3260 	struct thread *td = curthread;
3261 	struct inodedep *inodedep;
3262 	struct vnode *vp;
3263 	struct inode *ip;
3264 	ino_t oldinum;
3265 	int error;
3266 
3267 	if ((vp = xp) == NULL &&
3268 	    (error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, LK_EXCLUSIVE,
3269 	     &vp)) != 0) {
3270 		softdep_error("handle_workitem_remove: vget", error);
3271 		return;
3272 	}
3273 	ip = VTOI(vp);
3274 	ACQUIRE_LOCK(&lk);
3275 	if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
3276 		FREE_LOCK(&lk);
3277 		panic("handle_workitem_remove: lost inodedep");
3278 	}
3279 	/*
3280 	 * Normal file deletion.
3281 	 */
3282 	if ((dirrem->dm_state & RMDIR) == 0) {
3283 		ip->i_nlink--;
3284 		DIP(ip, i_nlink) = ip->i_nlink;
3285 		ip->i_flag |= IN_CHANGE;
3286 		if (ip->i_nlink < ip->i_effnlink) {
3287 			FREE_LOCK(&lk);
3288 			panic("handle_workitem_remove: bad file delta");
3289 		}
3290 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3291 		FREE_LOCK(&lk);
3292 		vput(vp);
3293 		num_dirrem -= 1;
3294 		WORKITEM_FREE(dirrem, D_DIRREM);
3295 		return;
3296 	}
3297 	/*
3298 	 * Directory deletion. Decrement reference count for both the
3299 	 * just deleted parent directory entry and the reference for ".".
3300 	 * Next truncate the directory to length zero. When the
3301 	 * truncation completes, arrange to have the reference count on
3302 	 * the parent decremented to account for the loss of "..".
3303 	 */
3304 	ip->i_nlink -= 2;
3305 	DIP(ip, i_nlink) = ip->i_nlink;
3306 	ip->i_flag |= IN_CHANGE;
3307 	if (ip->i_nlink < ip->i_effnlink) {
3308 		FREE_LOCK(&lk);
3309 		panic("handle_workitem_remove: bad dir delta");
3310 	}
3311 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3312 	FREE_LOCK(&lk);
3313 	if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, td->td_ucred, td)) != 0)
3314 		softdep_error("handle_workitem_remove: truncate", error);
3315 	/*
3316 	 * Rename a directory to a new parent. Since, we are both deleting
3317 	 * and creating a new directory entry, the link count on the new
3318 	 * directory should not change. Thus we skip the followup dirrem.
3319 	 */
3320 	if (dirrem->dm_state & DIRCHG) {
3321 		vput(vp);
3322 		num_dirrem -= 1;
3323 		WORKITEM_FREE(dirrem, D_DIRREM);
3324 		return;
3325 	}
3326 	/*
3327 	 * If the inodedep does not exist, then the zero'ed inode has
3328 	 * been written to disk. If the allocated inode has never been
3329 	 * written to disk, then the on-disk inode is zero'ed. In either
3330 	 * case we can remove the file immediately.
3331 	 */
3332 	ACQUIRE_LOCK(&lk);
3333 	dirrem->dm_state = 0;
3334 	oldinum = dirrem->dm_oldinum;
3335 	dirrem->dm_oldinum = dirrem->dm_dirinum;
3336 	if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
3337 	    check_inode_unwritten(inodedep)) {
3338 		FREE_LOCK(&lk);
3339 		vput(vp);
3340 		handle_workitem_remove(dirrem, NULL);
3341 		return;
3342 	}
3343 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3344 	FREE_LOCK(&lk);
3345 	vput(vp);
3346 }
3347 
3348 /*
3349  * Inode de-allocation dependencies.
3350  *
3351  * When an inode's link count is reduced to zero, it can be de-allocated. We
3352  * found it convenient to postpone de-allocation until after the inode is
3353  * written to disk with its new link count (zero).  At this point, all of the
3354  * on-disk inode's block pointers are nullified and, with careful dependency
3355  * list ordering, all dependencies related to the inode will be satisfied and
3356  * the corresponding dependency structures de-allocated.  So, if/when the
3357  * inode is reused, there will be no mixing of old dependencies with new
3358  * ones.  This artificial dependency is set up by the block de-allocation
3359  * procedure above (softdep_setup_freeblocks) and completed by the
3360  * following procedure.
3361  */
3362 static void
3363 handle_workitem_freefile(freefile)
3364 	struct freefile *freefile;
3365 {
3366 	struct fs *fs;
3367 	struct inodedep *idp;
3368 	int error;
3369 
3370 	fs = VFSTOUFS(freefile->fx_mnt)->um_fs;
3371 #ifdef DEBUG
3372 	ACQUIRE_LOCK(&lk);
3373 	error = inodedep_lookup(fs, freefile->fx_oldinum, 0, &idp);
3374 	FREE_LOCK(&lk);
3375 	if (error)
3376 		panic("handle_workitem_freefile: inodedep survived");
3377 #endif
3378 	fs->fs_pendinginodes -= 1;
3379 	if ((error = ffs_freefile(fs, freefile->fx_devvp, freefile->fx_oldinum,
3380 	     freefile->fx_mode)) != 0)
3381 		softdep_error("handle_workitem_freefile", error);
3382 	WORKITEM_FREE(freefile, D_FREEFILE);
3383 }
3384 
3385 /*
3386  * Disk writes.
3387  *
3388  * The dependency structures constructed above are most actively used when file
3389  * system blocks are written to disk.  No constraints are placed on when a
3390  * block can be written, but unsatisfied update dependencies are made safe by
3391  * modifying (or replacing) the source memory for the duration of the disk
3392  * write.  When the disk write completes, the memory block is again brought
3393  * up-to-date.
3394  *
3395  * In-core inode structure reclamation.
3396  *
3397  * Because there are a finite number of "in-core" inode structures, they are
3398  * reused regularly.  By transferring all inode-related dependencies to the
3399  * in-memory inode block and indexing them separately (via "inodedep"s), we
3400  * can allow "in-core" inode structures to be reused at any time and avoid
3401  * any increase in contention.
3402  *
3403  * Called just before entering the device driver to initiate a new disk I/O.
3404  * The buffer must be locked, thus, no I/O completion operations can occur
3405  * while we are manipulating its associated dependencies.
3406  */
3407 static void
3408 softdep_disk_io_initiation(bp)
3409 	struct buf *bp;		/* structure describing disk write to occur */
3410 {
3411 	struct worklist *wk, *nextwk;
3412 	struct indirdep *indirdep;
3413 	struct inodedep *inodedep;
3414 
3415 	/*
3416 	 * We only care about write operations. There should never
3417 	 * be dependencies for reads.
3418 	 */
3419 	if (bp->b_iocmd == BIO_READ)
3420 		panic("softdep_disk_io_initiation: read");
3421 	/*
3422 	 * Do any necessary pre-I/O processing.
3423 	 */
3424 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
3425 		nextwk = LIST_NEXT(wk, wk_list);
3426 		switch (wk->wk_type) {
3427 
3428 		case D_PAGEDEP:
3429 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
3430 			continue;
3431 
3432 		case D_INODEDEP:
3433 			inodedep = WK_INODEDEP(wk);
3434 			if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
3435 				initiate_write_inodeblock_ufs1(inodedep, bp);
3436 			else
3437 				initiate_write_inodeblock_ufs2(inodedep, bp);
3438 			continue;
3439 
3440 		case D_INDIRDEP:
3441 			indirdep = WK_INDIRDEP(wk);
3442 			if (indirdep->ir_state & GOINGAWAY)
3443 				panic("disk_io_initiation: indirdep gone");
3444 			/*
3445 			 * If there are no remaining dependencies, this
3446 			 * will be writing the real pointers, so the
3447 			 * dependency can be freed.
3448 			 */
3449 			if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
3450 				indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
3451 				brelse(indirdep->ir_savebp);
3452 				/* inline expand WORKLIST_REMOVE(wk); */
3453 				wk->wk_state &= ~ONWORKLIST;
3454 				LIST_REMOVE(wk, wk_list);
3455 				WORKITEM_FREE(indirdep, D_INDIRDEP);
3456 				continue;
3457 			}
3458 			/*
3459 			 * Replace up-to-date version with safe version.
3460 			 */
3461 			MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3462 			    M_INDIRDEP, M_SOFTDEP_FLAGS);
3463 			ACQUIRE_LOCK(&lk);
3464 			indirdep->ir_state &= ~ATTACHED;
3465 			indirdep->ir_state |= UNDONE;
3466 			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3467 			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3468 			    bp->b_bcount);
3469 			FREE_LOCK(&lk);
3470 			continue;
3471 
3472 		case D_MKDIR:
3473 		case D_BMSAFEMAP:
3474 		case D_ALLOCDIRECT:
3475 		case D_ALLOCINDIR:
3476 			continue;
3477 
3478 		default:
3479 			panic("handle_disk_io_initiation: Unexpected type %s",
3480 			    TYPENAME(wk->wk_type));
3481 			/* NOTREACHED */
3482 		}
3483 	}
3484 }
3485 
3486 /*
3487  * Called from within the procedure above to deal with unsatisfied
3488  * allocation dependencies in a directory. The buffer must be locked,
3489  * thus, no I/O completion operations can occur while we are
3490  * manipulating its associated dependencies.
3491  */
3492 static void
3493 initiate_write_filepage(pagedep, bp)
3494 	struct pagedep *pagedep;
3495 	struct buf *bp;
3496 {
3497 	struct diradd *dap;
3498 	struct direct *ep;
3499 	int i;
3500 
3501 	if (pagedep->pd_state & IOSTARTED) {
3502 		/*
3503 		 * This can only happen if there is a driver that does not
3504 		 * understand chaining. Here biodone will reissue the call
3505 		 * to strategy for the incomplete buffers.
3506 		 */
3507 		printf("initiate_write_filepage: already started\n");
3508 		return;
3509 	}
3510 	pagedep->pd_state |= IOSTARTED;
3511 	ACQUIRE_LOCK(&lk);
3512 	for (i = 0; i < DAHASHSZ; i++) {
3513 		LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3514 			ep = (struct direct *)
3515 			    ((char *)bp->b_data + dap->da_offset);
3516 			if (ep->d_ino != dap->da_newinum) {
3517 				FREE_LOCK(&lk);
3518 				panic("%s: dir inum %d != new %d",
3519 				    "initiate_write_filepage",
3520 				    ep->d_ino, dap->da_newinum);
3521 			}
3522 			if (dap->da_state & DIRCHG)
3523 				ep->d_ino = dap->da_previous->dm_oldinum;
3524 			else
3525 				ep->d_ino = 0;
3526 			dap->da_state &= ~ATTACHED;
3527 			dap->da_state |= UNDONE;
3528 		}
3529 	}
3530 	FREE_LOCK(&lk);
3531 }
3532 
3533 /*
3534  * Version of initiate_write_inodeblock that handles UFS1 dinodes.
3535  * Note that any bug fixes made to this routine must be done in the
3536  * version found below.
3537  *
3538  * Called from within the procedure above to deal with unsatisfied
3539  * allocation dependencies in an inodeblock. The buffer must be
3540  * locked, thus, no I/O completion operations can occur while we
3541  * are manipulating its associated dependencies.
3542  */
3543 static void
3544 initiate_write_inodeblock_ufs1(inodedep, bp)
3545 	struct inodedep *inodedep;
3546 	struct buf *bp;			/* The inode block */
3547 {
3548 	struct allocdirect *adp, *lastadp;
3549 	struct ufs1_dinode *dp;
3550 	struct fs *fs;
3551 	ufs_lbn_t i, prevlbn = 0;
3552 	int deplist;
3553 
3554 	if (inodedep->id_state & IOSTARTED)
3555 		panic("initiate_write_inodeblock_ufs1: already started");
3556 	inodedep->id_state |= IOSTARTED;
3557 	fs = inodedep->id_fs;
3558 	dp = (struct ufs1_dinode *)bp->b_data +
3559 	    ino_to_fsbo(fs, inodedep->id_ino);
3560 	/*
3561 	 * If the bitmap is not yet written, then the allocated
3562 	 * inode cannot be written to disk.
3563 	 */
3564 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3565 		if (inodedep->id_savedino1 != NULL)
3566 			panic("initiate_write_inodeblock_ufs1: I/O underway");
3567 		MALLOC(inodedep->id_savedino1, struct ufs1_dinode *,
3568 		    sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3569 		*inodedep->id_savedino1 = *dp;
3570 		bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
3571 		return;
3572 	}
3573 	/*
3574 	 * If no dependencies, then there is nothing to roll back.
3575 	 */
3576 	inodedep->id_savedsize = dp->di_size;
3577 	inodedep->id_savedextsize = 0;
3578 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3579 		return;
3580 	/*
3581 	 * Set the dependencies to busy.
3582 	 */
3583 	ACQUIRE_LOCK(&lk);
3584 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3585 	     adp = TAILQ_NEXT(adp, ad_next)) {
3586 #ifdef DIAGNOSTIC
3587 		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3588 			FREE_LOCK(&lk);
3589 			panic("softdep_write_inodeblock: lbn order");
3590 		}
3591 		prevlbn = adp->ad_lbn;
3592 		if (adp->ad_lbn < NDADDR &&
3593 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3594 			FREE_LOCK(&lk);
3595 			panic("%s: direct pointer #%jd mismatch %d != %jd",
3596 			    "softdep_write_inodeblock",
3597 			    (intmax_t)adp->ad_lbn,
3598 			    dp->di_db[adp->ad_lbn],
3599 			    (intmax_t)adp->ad_newblkno);
3600 		}
3601 		if (adp->ad_lbn >= NDADDR &&
3602 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3603 			FREE_LOCK(&lk);
3604 			panic("%s: indirect pointer #%jd mismatch %d != %jd",
3605 			    "softdep_write_inodeblock",
3606 			    (intmax_t)adp->ad_lbn - NDADDR,
3607 			    dp->di_ib[adp->ad_lbn - NDADDR],
3608 			    (intmax_t)adp->ad_newblkno);
3609 		}
3610 		deplist |= 1 << adp->ad_lbn;
3611 		if ((adp->ad_state & ATTACHED) == 0) {
3612 			FREE_LOCK(&lk);
3613 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3614 			    adp->ad_state);
3615 		}
3616 #endif /* DIAGNOSTIC */
3617 		adp->ad_state &= ~ATTACHED;
3618 		adp->ad_state |= UNDONE;
3619 	}
3620 	/*
3621 	 * The on-disk inode cannot claim to be any larger than the last
3622 	 * fragment that has been written. Otherwise, the on-disk inode
3623 	 * might have fragments that were not the last block in the file
3624 	 * which would corrupt the filesystem.
3625 	 */
3626 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3627 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3628 		if (adp->ad_lbn >= NDADDR)
3629 			break;
3630 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3631 		/* keep going until hitting a rollback to a frag */
3632 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3633 			continue;
3634 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3635 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3636 #ifdef DIAGNOSTIC
3637 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3638 				FREE_LOCK(&lk);
3639 				panic("softdep_write_inodeblock: lost dep1");
3640 			}
3641 #endif /* DIAGNOSTIC */
3642 			dp->di_db[i] = 0;
3643 		}
3644 		for (i = 0; i < NIADDR; i++) {
3645 #ifdef DIAGNOSTIC
3646 			if (dp->di_ib[i] != 0 &&
3647 			    (deplist & ((1 << NDADDR) << i)) == 0) {
3648 				FREE_LOCK(&lk);
3649 				panic("softdep_write_inodeblock: lost dep2");
3650 			}
3651 #endif /* DIAGNOSTIC */
3652 			dp->di_ib[i] = 0;
3653 		}
3654 		FREE_LOCK(&lk);
3655 		return;
3656 	}
3657 	/*
3658 	 * If we have zero'ed out the last allocated block of the file,
3659 	 * roll back the size to the last currently allocated block.
3660 	 * We know that this last allocated block is a full-sized as
3661 	 * we already checked for fragments in the loop above.
3662 	 */
3663 	if (lastadp != NULL &&
3664 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3665 		for (i = lastadp->ad_lbn; i >= 0; i--)
3666 			if (dp->di_db[i] != 0)
3667 				break;
3668 		dp->di_size = (i + 1) * fs->fs_bsize;
3669 	}
3670 	/*
3671 	 * The only dependencies are for indirect blocks.
3672 	 *
3673 	 * The file size for indirect block additions is not guaranteed.
3674 	 * Such a guarantee would be non-trivial to achieve. The conventional
3675 	 * synchronous write implementation also does not make this guarantee.
3676 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3677 	 * can be over-estimated without destroying integrity when the file
3678 	 * moves into the indirect blocks (i.e., is large). If we want to
3679 	 * postpone fsck, we are stuck with this argument.
3680 	 */
3681 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3682 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3683 	FREE_LOCK(&lk);
3684 }
3685 
3686 /*
3687  * Version of initiate_write_inodeblock that handles UFS2 dinodes.
3688  * Note that any bug fixes made to this routine must be done in the
3689  * version found above.
3690  *
3691  * Called from within the procedure above to deal with unsatisfied
3692  * allocation dependencies in an inodeblock. The buffer must be
3693  * locked, thus, no I/O completion operations can occur while we
3694  * are manipulating its associated dependencies.
3695  */
3696 static void
3697 initiate_write_inodeblock_ufs2(inodedep, bp)
3698 	struct inodedep *inodedep;
3699 	struct buf *bp;			/* The inode block */
3700 {
3701 	struct allocdirect *adp, *lastadp;
3702 	struct ufs2_dinode *dp;
3703 	struct fs *fs;
3704 	ufs_lbn_t i, prevlbn = 0;
3705 	int deplist;
3706 
3707 	if (inodedep->id_state & IOSTARTED)
3708 		panic("initiate_write_inodeblock_ufs2: already started");
3709 	inodedep->id_state |= IOSTARTED;
3710 	fs = inodedep->id_fs;
3711 	dp = (struct ufs2_dinode *)bp->b_data +
3712 	    ino_to_fsbo(fs, inodedep->id_ino);
3713 	/*
3714 	 * If the bitmap is not yet written, then the allocated
3715 	 * inode cannot be written to disk.
3716 	 */
3717 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3718 		if (inodedep->id_savedino2 != NULL)
3719 			panic("initiate_write_inodeblock_ufs2: I/O underway");
3720 		MALLOC(inodedep->id_savedino2, struct ufs2_dinode *,
3721 		    sizeof(struct ufs2_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3722 		*inodedep->id_savedino2 = *dp;
3723 		bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
3724 		return;
3725 	}
3726 	/*
3727 	 * If no dependencies, then there is nothing to roll back.
3728 	 */
3729 	inodedep->id_savedsize = dp->di_size;
3730 	inodedep->id_savedextsize = dp->di_extsize;
3731 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL &&
3732 	    TAILQ_FIRST(&inodedep->id_extupdt) == NULL)
3733 		return;
3734 	/*
3735 	 * Set the ext data dependencies to busy.
3736 	 */
3737 	ACQUIRE_LOCK(&lk);
3738 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
3739 	     adp = TAILQ_NEXT(adp, ad_next)) {
3740 #ifdef DIAGNOSTIC
3741 		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3742 			FREE_LOCK(&lk);
3743 			panic("softdep_write_inodeblock: lbn order");
3744 		}
3745 		prevlbn = adp->ad_lbn;
3746 		if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno) {
3747 			FREE_LOCK(&lk);
3748 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
3749 			    "softdep_write_inodeblock",
3750 			    (intmax_t)adp->ad_lbn,
3751 			    (intmax_t)dp->di_extb[adp->ad_lbn],
3752 			    (intmax_t)adp->ad_newblkno);
3753 		}
3754 		deplist |= 1 << adp->ad_lbn;
3755 		if ((adp->ad_state & ATTACHED) == 0) {
3756 			FREE_LOCK(&lk);
3757 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3758 			    adp->ad_state);
3759 		}
3760 #endif /* DIAGNOSTIC */
3761 		adp->ad_state &= ~ATTACHED;
3762 		adp->ad_state |= UNDONE;
3763 	}
3764 	/*
3765 	 * The on-disk inode cannot claim to be any larger than the last
3766 	 * fragment that has been written. Otherwise, the on-disk inode
3767 	 * might have fragments that were not the last block in the ext
3768 	 * data which would corrupt the filesystem.
3769 	 */
3770 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
3771 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3772 		dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno;
3773 		/* keep going until hitting a rollback to a frag */
3774 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3775 			continue;
3776 		dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3777 		for (i = adp->ad_lbn + 1; i < NXADDR; i++) {
3778 #ifdef DIAGNOSTIC
3779 			if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) {
3780 				FREE_LOCK(&lk);
3781 				panic("softdep_write_inodeblock: lost dep1");
3782 			}
3783 #endif /* DIAGNOSTIC */
3784 			dp->di_extb[i] = 0;
3785 		}
3786 		lastadp = NULL;
3787 		break;
3788 	}
3789 	/*
3790 	 * If we have zero'ed out the last allocated block of the ext
3791 	 * data, roll back the size to the last currently allocated block.
3792 	 * We know that this last allocated block is a full-sized as
3793 	 * we already checked for fragments in the loop above.
3794 	 */
3795 	if (lastadp != NULL &&
3796 	    dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3797 		for (i = lastadp->ad_lbn; i >= 0; i--)
3798 			if (dp->di_extb[i] != 0)
3799 				break;
3800 		dp->di_extsize = (i + 1) * fs->fs_bsize;
3801 	}
3802 	/*
3803 	 * Set the file data dependencies to busy.
3804 	 */
3805 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3806 	     adp = TAILQ_NEXT(adp, ad_next)) {
3807 #ifdef DIAGNOSTIC
3808 		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3809 			FREE_LOCK(&lk);
3810 			panic("softdep_write_inodeblock: lbn order");
3811 		}
3812 		prevlbn = adp->ad_lbn;
3813 		if (adp->ad_lbn < NDADDR &&
3814 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3815 			FREE_LOCK(&lk);
3816 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
3817 			    "softdep_write_inodeblock",
3818 			    (intmax_t)adp->ad_lbn,
3819 			    (intmax_t)dp->di_db[adp->ad_lbn],
3820 			    (intmax_t)adp->ad_newblkno);
3821 		}
3822 		if (adp->ad_lbn >= NDADDR &&
3823 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3824 			FREE_LOCK(&lk);
3825 			panic("%s indirect pointer #%jd mismatch %jd != %jd",
3826 			    "softdep_write_inodeblock:",
3827 			    (intmax_t)adp->ad_lbn - NDADDR,
3828 			    (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR],
3829 			    (intmax_t)adp->ad_newblkno);
3830 		}
3831 		deplist |= 1 << adp->ad_lbn;
3832 		if ((adp->ad_state & ATTACHED) == 0) {
3833 			FREE_LOCK(&lk);
3834 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3835 			    adp->ad_state);
3836 		}
3837 #endif /* DIAGNOSTIC */
3838 		adp->ad_state &= ~ATTACHED;
3839 		adp->ad_state |= UNDONE;
3840 	}
3841 	/*
3842 	 * The on-disk inode cannot claim to be any larger than the last
3843 	 * fragment that has been written. Otherwise, the on-disk inode
3844 	 * might have fragments that were not the last block in the file
3845 	 * which would corrupt the filesystem.
3846 	 */
3847 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3848 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3849 		if (adp->ad_lbn >= NDADDR)
3850 			break;
3851 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3852 		/* keep going until hitting a rollback to a frag */
3853 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3854 			continue;
3855 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3856 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3857 #ifdef DIAGNOSTIC
3858 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3859 				FREE_LOCK(&lk);
3860 				panic("softdep_write_inodeblock: lost dep2");
3861 			}
3862 #endif /* DIAGNOSTIC */
3863 			dp->di_db[i] = 0;
3864 		}
3865 		for (i = 0; i < NIADDR; i++) {
3866 #ifdef DIAGNOSTIC
3867 			if (dp->di_ib[i] != 0 &&
3868 			    (deplist & ((1 << NDADDR) << i)) == 0) {
3869 				FREE_LOCK(&lk);
3870 				panic("softdep_write_inodeblock: lost dep3");
3871 			}
3872 #endif /* DIAGNOSTIC */
3873 			dp->di_ib[i] = 0;
3874 		}
3875 		FREE_LOCK(&lk);
3876 		return;
3877 	}
3878 	/*
3879 	 * If we have zero'ed out the last allocated block of the file,
3880 	 * roll back the size to the last currently allocated block.
3881 	 * We know that this last allocated block is a full-sized as
3882 	 * we already checked for fragments in the loop above.
3883 	 */
3884 	if (lastadp != NULL &&
3885 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3886 		for (i = lastadp->ad_lbn; i >= 0; i--)
3887 			if (dp->di_db[i] != 0)
3888 				break;
3889 		dp->di_size = (i + 1) * fs->fs_bsize;
3890 	}
3891 	/*
3892 	 * The only dependencies are for indirect blocks.
3893 	 *
3894 	 * The file size for indirect block additions is not guaranteed.
3895 	 * Such a guarantee would be non-trivial to achieve. The conventional
3896 	 * synchronous write implementation also does not make this guarantee.
3897 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3898 	 * can be over-estimated without destroying integrity when the file
3899 	 * moves into the indirect blocks (i.e., is large). If we want to
3900 	 * postpone fsck, we are stuck with this argument.
3901 	 */
3902 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3903 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3904 	FREE_LOCK(&lk);
3905 }
3906 
3907 /*
3908  * This routine is called during the completion interrupt
3909  * service routine for a disk write (from the procedure called
3910  * by the device driver to inform the filesystem caches of
3911  * a request completion).  It should be called early in this
3912  * procedure, before the block is made available to other
3913  * processes or other routines are called.
3914  */
3915 static void
3916 softdep_disk_write_complete(bp)
3917 	struct buf *bp;		/* describes the completed disk write */
3918 {
3919 	struct worklist *wk;
3920 	struct workhead reattach;
3921 	struct newblk *newblk;
3922 	struct allocindir *aip;
3923 	struct allocdirect *adp;
3924 	struct indirdep *indirdep;
3925 	struct inodedep *inodedep;
3926 	struct bmsafemap *bmsafemap;
3927 
3928 #ifdef DEBUG
3929 	if (lk.lkt_held != NOHOLDER)
3930 		panic("softdep_disk_write_complete: lock is held");
3931 	lk.lkt_held = SPECIAL_FLAG;
3932 #endif
3933 	LIST_INIT(&reattach);
3934 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3935 		WORKLIST_REMOVE(wk);
3936 		switch (wk->wk_type) {
3937 
3938 		case D_PAGEDEP:
3939 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3940 				WORKLIST_INSERT(&reattach, wk);
3941 			continue;
3942 
3943 		case D_INODEDEP:
3944 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3945 				WORKLIST_INSERT(&reattach, wk);
3946 			continue;
3947 
3948 		case D_BMSAFEMAP:
3949 			bmsafemap = WK_BMSAFEMAP(wk);
3950 			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3951 				newblk->nb_state |= DEPCOMPLETE;
3952 				newblk->nb_bmsafemap = NULL;
3953 				LIST_REMOVE(newblk, nb_deps);
3954 			}
3955 			while ((adp =
3956 			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3957 				adp->ad_state |= DEPCOMPLETE;
3958 				adp->ad_buf = NULL;
3959 				LIST_REMOVE(adp, ad_deps);
3960 				handle_allocdirect_partdone(adp);
3961 			}
3962 			while ((aip =
3963 			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3964 				aip->ai_state |= DEPCOMPLETE;
3965 				aip->ai_buf = NULL;
3966 				LIST_REMOVE(aip, ai_deps);
3967 				handle_allocindir_partdone(aip);
3968 			}
3969 			while ((inodedep =
3970 			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3971 				inodedep->id_state |= DEPCOMPLETE;
3972 				LIST_REMOVE(inodedep, id_deps);
3973 				inodedep->id_buf = NULL;
3974 			}
3975 			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3976 			continue;
3977 
3978 		case D_MKDIR:
3979 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3980 			continue;
3981 
3982 		case D_ALLOCDIRECT:
3983 			adp = WK_ALLOCDIRECT(wk);
3984 			adp->ad_state |= COMPLETE;
3985 			handle_allocdirect_partdone(adp);
3986 			continue;
3987 
3988 		case D_ALLOCINDIR:
3989 			aip = WK_ALLOCINDIR(wk);
3990 			aip->ai_state |= COMPLETE;
3991 			handle_allocindir_partdone(aip);
3992 			continue;
3993 
3994 		case D_INDIRDEP:
3995 			indirdep = WK_INDIRDEP(wk);
3996 			if (indirdep->ir_state & GOINGAWAY) {
3997 				lk.lkt_held = NOHOLDER;
3998 				panic("disk_write_complete: indirdep gone");
3999 			}
4000 			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
4001 			FREE(indirdep->ir_saveddata, M_INDIRDEP);
4002 			indirdep->ir_saveddata = 0;
4003 			indirdep->ir_state &= ~UNDONE;
4004 			indirdep->ir_state |= ATTACHED;
4005 			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
4006 				handle_allocindir_partdone(aip);
4007 				if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
4008 					lk.lkt_held = NOHOLDER;
4009 					panic("disk_write_complete: not gone");
4010 				}
4011 			}
4012 			WORKLIST_INSERT(&reattach, wk);
4013 			if ((bp->b_flags & B_DELWRI) == 0)
4014 				stat_indir_blk_ptrs++;
4015 			bdirty(bp);
4016 			continue;
4017 
4018 		default:
4019 			lk.lkt_held = NOHOLDER;
4020 			panic("handle_disk_write_complete: Unknown type %s",
4021 			    TYPENAME(wk->wk_type));
4022 			/* NOTREACHED */
4023 		}
4024 	}
4025 	/*
4026 	 * Reattach any requests that must be redone.
4027 	 */
4028 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
4029 		WORKLIST_REMOVE(wk);
4030 		WORKLIST_INSERT(&bp->b_dep, wk);
4031 	}
4032 #ifdef DEBUG
4033 	if (lk.lkt_held != SPECIAL_FLAG)
4034 		panic("softdep_disk_write_complete: lock lost");
4035 	lk.lkt_held = NOHOLDER;
4036 #endif
4037 }
4038 
4039 /*
4040  * Called from within softdep_disk_write_complete above. Note that
4041  * this routine is always called from interrupt level with further
4042  * splbio interrupts blocked.
4043  */
4044 static void
4045 handle_allocdirect_partdone(adp)
4046 	struct allocdirect *adp;	/* the completed allocdirect */
4047 {
4048 	struct allocdirectlst *listhead;
4049 	struct allocdirect *listadp;
4050 	struct inodedep *inodedep;
4051 	long bsize, delay;
4052 
4053 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4054 		return;
4055 	if (adp->ad_buf != NULL) {
4056 		lk.lkt_held = NOHOLDER;
4057 		panic("handle_allocdirect_partdone: dangling dep");
4058 	}
4059 	/*
4060 	 * The on-disk inode cannot claim to be any larger than the last
4061 	 * fragment that has been written. Otherwise, the on-disk inode
4062 	 * might have fragments that were not the last block in the file
4063 	 * which would corrupt the filesystem. Thus, we cannot free any
4064 	 * allocdirects after one whose ad_oldblkno claims a fragment as
4065 	 * these blocks must be rolled back to zero before writing the inode.
4066 	 * We check the currently active set of allocdirects in id_inoupdt
4067 	 * or id_extupdt as appropriate.
4068 	 */
4069 	inodedep = adp->ad_inodedep;
4070 	bsize = inodedep->id_fs->fs_bsize;
4071 	if (adp->ad_state & EXTDATA)
4072 		listhead = &inodedep->id_extupdt;
4073 	else
4074 		listhead = &inodedep->id_inoupdt;
4075 	TAILQ_FOREACH(listadp, listhead, ad_next) {
4076 		/* found our block */
4077 		if (listadp == adp)
4078 			break;
4079 		/* continue if ad_oldlbn is not a fragment */
4080 		if (listadp->ad_oldsize == 0 ||
4081 		    listadp->ad_oldsize == bsize)
4082 			continue;
4083 		/* hit a fragment */
4084 		return;
4085 	}
4086 	/*
4087 	 * If we have reached the end of the current list without
4088 	 * finding the just finished dependency, then it must be
4089 	 * on the future dependency list. Future dependencies cannot
4090 	 * be freed until they are moved to the current list.
4091 	 */
4092 	if (listadp == NULL) {
4093 #ifdef DEBUG
4094 		if (adp->ad_state & EXTDATA)
4095 			listhead = &inodedep->id_newextupdt;
4096 		else
4097 			listhead = &inodedep->id_newinoupdt;
4098 		TAILQ_FOREACH(listadp, listhead, ad_next)
4099 			/* found our block */
4100 			if (listadp == adp)
4101 				break;
4102 		if (listadp == NULL) {
4103 			lk.lkt_held = NOHOLDER;
4104 			panic("handle_allocdirect_partdone: lost dep");
4105 		}
4106 #endif /* DEBUG */
4107 		return;
4108 	}
4109 	/*
4110 	 * If we have found the just finished dependency, then free
4111 	 * it along with anything that follows it that is complete.
4112 	 * If the inode still has a bitmap dependency, then it has
4113 	 * never been written to disk, hence the on-disk inode cannot
4114 	 * reference the old fragment so we can free it without delay.
4115 	 */
4116 	delay = (inodedep->id_state & DEPCOMPLETE);
4117 	for (; adp; adp = listadp) {
4118 		listadp = TAILQ_NEXT(adp, ad_next);
4119 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4120 			return;
4121 		free_allocdirect(listhead, adp, delay);
4122 	}
4123 }
4124 
4125 /*
4126  * Called from within softdep_disk_write_complete above. Note that
4127  * this routine is always called from interrupt level with further
4128  * splbio interrupts blocked.
4129  */
4130 static void
4131 handle_allocindir_partdone(aip)
4132 	struct allocindir *aip;		/* the completed allocindir */
4133 {
4134 	struct indirdep *indirdep;
4135 
4136 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
4137 		return;
4138 	if (aip->ai_buf != NULL) {
4139 		lk.lkt_held = NOHOLDER;
4140 		panic("handle_allocindir_partdone: dangling dependency");
4141 	}
4142 	indirdep = aip->ai_indirdep;
4143 	if (indirdep->ir_state & UNDONE) {
4144 		LIST_REMOVE(aip, ai_next);
4145 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
4146 		return;
4147 	}
4148 	if (indirdep->ir_state & UFS1FMT)
4149 		((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4150 		    aip->ai_newblkno;
4151 	else
4152 		((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4153 		    aip->ai_newblkno;
4154 	LIST_REMOVE(aip, ai_next);
4155 	if (aip->ai_freefrag != NULL)
4156 		add_to_worklist(&aip->ai_freefrag->ff_list);
4157 	WORKITEM_FREE(aip, D_ALLOCINDIR);
4158 }
4159 
4160 /*
4161  * Called from within softdep_disk_write_complete above to restore
4162  * in-memory inode block contents to their most up-to-date state. Note
4163  * that this routine is always called from interrupt level with further
4164  * splbio interrupts blocked.
4165  */
4166 static int
4167 handle_written_inodeblock(inodedep, bp)
4168 	struct inodedep *inodedep;
4169 	struct buf *bp;		/* buffer containing the inode block */
4170 {
4171 	struct worklist *wk, *filefree;
4172 	struct allocdirect *adp, *nextadp;
4173 	struct ufs1_dinode *dp1 = NULL;
4174 	struct ufs2_dinode *dp2 = NULL;
4175 	int hadchanges, fstype;
4176 
4177 	if ((inodedep->id_state & IOSTARTED) == 0) {
4178 		lk.lkt_held = NOHOLDER;
4179 		panic("handle_written_inodeblock: not started");
4180 	}
4181 	inodedep->id_state &= ~IOSTARTED;
4182 	inodedep->id_state |= COMPLETE;
4183 	if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
4184 		fstype = UFS1;
4185 		dp1 = (struct ufs1_dinode *)bp->b_data +
4186 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4187 	} else {
4188 		fstype = UFS2;
4189 		dp2 = (struct ufs2_dinode *)bp->b_data +
4190 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4191 	}
4192 	/*
4193 	 * If we had to rollback the inode allocation because of
4194 	 * bitmaps being incomplete, then simply restore it.
4195 	 * Keep the block dirty so that it will not be reclaimed until
4196 	 * all associated dependencies have been cleared and the
4197 	 * corresponding updates written to disk.
4198 	 */
4199 	if (inodedep->id_savedino1 != NULL) {
4200 		if (fstype == UFS1)
4201 			*dp1 = *inodedep->id_savedino1;
4202 		else
4203 			*dp2 = *inodedep->id_savedino2;
4204 		FREE(inodedep->id_savedino1, M_INODEDEP);
4205 		inodedep->id_savedino1 = NULL;
4206 		if ((bp->b_flags & B_DELWRI) == 0)
4207 			stat_inode_bitmap++;
4208 		bdirty(bp);
4209 		return (1);
4210 	}
4211 	/*
4212 	 * Roll forward anything that had to be rolled back before
4213 	 * the inode could be updated.
4214 	 */
4215 	hadchanges = 0;
4216 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
4217 		nextadp = TAILQ_NEXT(adp, ad_next);
4218 		if (adp->ad_state & ATTACHED) {
4219 			lk.lkt_held = NOHOLDER;
4220 			panic("handle_written_inodeblock: new entry");
4221 		}
4222 		if (fstype == UFS1) {
4223 			if (adp->ad_lbn < NDADDR) {
4224 				if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno) {
4225 					lk.lkt_held = NOHOLDER;
4226 					panic("%s %s #%jd mismatch %d != %jd",
4227 					    "handle_written_inodeblock:",
4228 					    "direct pointer",
4229 					    (intmax_t)adp->ad_lbn,
4230 					    dp1->di_db[adp->ad_lbn],
4231 					    (intmax_t)adp->ad_oldblkno);
4232 				}
4233 				dp1->di_db[adp->ad_lbn] = adp->ad_newblkno;
4234 			} else {
4235 				if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0) {
4236 					lk.lkt_held = NOHOLDER;
4237 					panic("%s: %s #%jd allocated as %d",
4238 					    "handle_written_inodeblock",
4239 					    "indirect pointer",
4240 					    (intmax_t)adp->ad_lbn - NDADDR,
4241 					    dp1->di_ib[adp->ad_lbn - NDADDR]);
4242 				}
4243 				dp1->di_ib[adp->ad_lbn - NDADDR] =
4244 				    adp->ad_newblkno;
4245 			}
4246 		} else {
4247 			if (adp->ad_lbn < NDADDR) {
4248 				if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno) {
4249 					lk.lkt_held = NOHOLDER;
4250 					panic("%s: %s #%jd %s %jd != %jd",
4251 					    "handle_written_inodeblock",
4252 					    "direct pointer",
4253 					    (intmax_t)adp->ad_lbn, "mismatch",
4254 					    (intmax_t)dp2->di_db[adp->ad_lbn],
4255 					    (intmax_t)adp->ad_oldblkno);
4256 				}
4257 				dp2->di_db[adp->ad_lbn] = adp->ad_newblkno;
4258 			} else {
4259 				if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0) {
4260 					lk.lkt_held = NOHOLDER;
4261 					panic("%s: %s #%jd allocated as %jd",
4262 					    "handle_written_inodeblock",
4263 					    "indirect pointer",
4264 					    (intmax_t)adp->ad_lbn - NDADDR,
4265 					    (intmax_t)
4266 					    dp2->di_ib[adp->ad_lbn - NDADDR]);
4267 				}
4268 				dp2->di_ib[adp->ad_lbn - NDADDR] =
4269 				    adp->ad_newblkno;
4270 			}
4271 		}
4272 		adp->ad_state &= ~UNDONE;
4273 		adp->ad_state |= ATTACHED;
4274 		hadchanges = 1;
4275 	}
4276 	for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
4277 		nextadp = TAILQ_NEXT(adp, ad_next);
4278 		if (adp->ad_state & ATTACHED) {
4279 			lk.lkt_held = NOHOLDER;
4280 			panic("handle_written_inodeblock: new entry");
4281 		}
4282 		if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno) {
4283 			lk.lkt_held = NOHOLDER;
4284 			panic("%s: direct pointers #%jd %s %jd != %jd",
4285 			    "handle_written_inodeblock",
4286 			    (intmax_t)adp->ad_lbn, "mismatch",
4287 			    (intmax_t)dp2->di_extb[adp->ad_lbn],
4288 			    (intmax_t)adp->ad_oldblkno);
4289 		}
4290 		dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno;
4291 		adp->ad_state &= ~UNDONE;
4292 		adp->ad_state |= ATTACHED;
4293 		hadchanges = 1;
4294 	}
4295 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
4296 		stat_direct_blk_ptrs++;
4297 	/*
4298 	 * Reset the file size to its most up-to-date value.
4299 	 */
4300 	if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) {
4301 		lk.lkt_held = NOHOLDER;
4302 		panic("handle_written_inodeblock: bad size");
4303 	}
4304 	if (fstype == UFS1) {
4305 		if (dp1->di_size != inodedep->id_savedsize) {
4306 			dp1->di_size = inodedep->id_savedsize;
4307 			hadchanges = 1;
4308 		}
4309 	} else {
4310 		if (dp2->di_size != inodedep->id_savedsize) {
4311 			dp2->di_size = inodedep->id_savedsize;
4312 			hadchanges = 1;
4313 		}
4314 		if (dp2->di_extsize != inodedep->id_savedextsize) {
4315 			dp2->di_extsize = inodedep->id_savedextsize;
4316 			hadchanges = 1;
4317 		}
4318 	}
4319 	inodedep->id_savedsize = -1;
4320 	inodedep->id_savedextsize = -1;
4321 	/*
4322 	 * If there were any rollbacks in the inode block, then it must be
4323 	 * marked dirty so that its will eventually get written back in
4324 	 * its correct form.
4325 	 */
4326 	if (hadchanges)
4327 		bdirty(bp);
4328 	/*
4329 	 * Process any allocdirects that completed during the update.
4330 	 */
4331 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
4332 		handle_allocdirect_partdone(adp);
4333 	if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
4334 		handle_allocdirect_partdone(adp);
4335 	/*
4336 	 * Process deallocations that were held pending until the
4337 	 * inode had been written to disk. Freeing of the inode
4338 	 * is delayed until after all blocks have been freed to
4339 	 * avoid creation of new <vfsid, inum, lbn> triples
4340 	 * before the old ones have been deleted.
4341 	 */
4342 	filefree = NULL;
4343 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
4344 		WORKLIST_REMOVE(wk);
4345 		switch (wk->wk_type) {
4346 
4347 		case D_FREEFILE:
4348 			/*
4349 			 * We defer adding filefree to the worklist until
4350 			 * all other additions have been made to ensure
4351 			 * that it will be done after all the old blocks
4352 			 * have been freed.
4353 			 */
4354 			if (filefree != NULL) {
4355 				lk.lkt_held = NOHOLDER;
4356 				panic("handle_written_inodeblock: filefree");
4357 			}
4358 			filefree = wk;
4359 			continue;
4360 
4361 		case D_MKDIR:
4362 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
4363 			continue;
4364 
4365 		case D_DIRADD:
4366 			diradd_inode_written(WK_DIRADD(wk), inodedep);
4367 			continue;
4368 
4369 		case D_FREEBLKS:
4370 		case D_FREEFRAG:
4371 		case D_DIRREM:
4372 			add_to_worklist(wk);
4373 			continue;
4374 
4375 		case D_NEWDIRBLK:
4376 			free_newdirblk(WK_NEWDIRBLK(wk));
4377 			continue;
4378 
4379 		default:
4380 			lk.lkt_held = NOHOLDER;
4381 			panic("handle_written_inodeblock: Unknown type %s",
4382 			    TYPENAME(wk->wk_type));
4383 			/* NOTREACHED */
4384 		}
4385 	}
4386 	if (filefree != NULL) {
4387 		if (free_inodedep(inodedep) == 0) {
4388 			lk.lkt_held = NOHOLDER;
4389 			panic("handle_written_inodeblock: live inodedep");
4390 		}
4391 		add_to_worklist(filefree);
4392 		return (0);
4393 	}
4394 
4395 	/*
4396 	 * If no outstanding dependencies, free it.
4397 	 */
4398 	if (free_inodedep(inodedep) ||
4399 	    (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
4400 	     TAILQ_FIRST(&inodedep->id_extupdt) == 0))
4401 		return (0);
4402 	return (hadchanges);
4403 }
4404 
4405 /*
4406  * Process a diradd entry after its dependent inode has been written.
4407  * This routine must be called with splbio interrupts blocked.
4408  */
4409 static void
4410 diradd_inode_written(dap, inodedep)
4411 	struct diradd *dap;
4412 	struct inodedep *inodedep;
4413 {
4414 	struct pagedep *pagedep;
4415 
4416 	dap->da_state |= COMPLETE;
4417 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4418 		if (dap->da_state & DIRCHG)
4419 			pagedep = dap->da_previous->dm_pagedep;
4420 		else
4421 			pagedep = dap->da_pagedep;
4422 		LIST_REMOVE(dap, da_pdlist);
4423 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4424 	}
4425 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
4426 }
4427 
4428 /*
4429  * Handle the completion of a mkdir dependency.
4430  */
4431 static void
4432 handle_written_mkdir(mkdir, type)
4433 	struct mkdir *mkdir;
4434 	int type;
4435 {
4436 	struct diradd *dap;
4437 	struct pagedep *pagedep;
4438 
4439 	if (mkdir->md_state != type) {
4440 		lk.lkt_held = NOHOLDER;
4441 		panic("handle_written_mkdir: bad type");
4442 	}
4443 	dap = mkdir->md_diradd;
4444 	dap->da_state &= ~type;
4445 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
4446 		dap->da_state |= DEPCOMPLETE;
4447 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4448 		if (dap->da_state & DIRCHG)
4449 			pagedep = dap->da_previous->dm_pagedep;
4450 		else
4451 			pagedep = dap->da_pagedep;
4452 		LIST_REMOVE(dap, da_pdlist);
4453 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4454 	}
4455 	LIST_REMOVE(mkdir, md_mkdirs);
4456 	WORKITEM_FREE(mkdir, D_MKDIR);
4457 }
4458 
4459 /*
4460  * Called from within softdep_disk_write_complete above.
4461  * A write operation was just completed. Removed inodes can
4462  * now be freed and associated block pointers may be committed.
4463  * Note that this routine is always called from interrupt level
4464  * with further splbio interrupts blocked.
4465  */
4466 static int
4467 handle_written_filepage(pagedep, bp)
4468 	struct pagedep *pagedep;
4469 	struct buf *bp;		/* buffer containing the written page */
4470 {
4471 	struct dirrem *dirrem;
4472 	struct diradd *dap, *nextdap;
4473 	struct direct *ep;
4474 	int i, chgs;
4475 
4476 	if ((pagedep->pd_state & IOSTARTED) == 0) {
4477 		lk.lkt_held = NOHOLDER;
4478 		panic("handle_written_filepage: not started");
4479 	}
4480 	pagedep->pd_state &= ~IOSTARTED;
4481 	/*
4482 	 * Process any directory removals that have been committed.
4483 	 */
4484 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
4485 		LIST_REMOVE(dirrem, dm_next);
4486 		dirrem->dm_dirinum = pagedep->pd_ino;
4487 		add_to_worklist(&dirrem->dm_list);
4488 	}
4489 	/*
4490 	 * Free any directory additions that have been committed.
4491 	 * If it is a newly allocated block, we have to wait until
4492 	 * the on-disk directory inode claims the new block.
4493 	 */
4494 	if ((pagedep->pd_state & NEWBLOCK) == 0)
4495 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
4496 			free_diradd(dap);
4497 	/*
4498 	 * Uncommitted directory entries must be restored.
4499 	 */
4500 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
4501 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
4502 		     dap = nextdap) {
4503 			nextdap = LIST_NEXT(dap, da_pdlist);
4504 			if (dap->da_state & ATTACHED) {
4505 				lk.lkt_held = NOHOLDER;
4506 				panic("handle_written_filepage: attached");
4507 			}
4508 			ep = (struct direct *)
4509 			    ((char *)bp->b_data + dap->da_offset);
4510 			ep->d_ino = dap->da_newinum;
4511 			dap->da_state &= ~UNDONE;
4512 			dap->da_state |= ATTACHED;
4513 			chgs = 1;
4514 			/*
4515 			 * If the inode referenced by the directory has
4516 			 * been written out, then the dependency can be
4517 			 * moved to the pending list.
4518 			 */
4519 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4520 				LIST_REMOVE(dap, da_pdlist);
4521 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
4522 				    da_pdlist);
4523 			}
4524 		}
4525 	}
4526 	/*
4527 	 * If there were any rollbacks in the directory, then it must be
4528 	 * marked dirty so that its will eventually get written back in
4529 	 * its correct form.
4530 	 */
4531 	if (chgs) {
4532 		if ((bp->b_flags & B_DELWRI) == 0)
4533 			stat_dir_entry++;
4534 		bdirty(bp);
4535 		return (1);
4536 	}
4537 	/*
4538 	 * If we are not waiting for a new directory block to be
4539 	 * claimed by its inode, then the pagedep will be freed.
4540 	 * Otherwise it will remain to track any new entries on
4541 	 * the page in case they are fsync'ed.
4542 	 */
4543 	if ((pagedep->pd_state & NEWBLOCK) == 0) {
4544 		LIST_REMOVE(pagedep, pd_hash);
4545 		WORKITEM_FREE(pagedep, D_PAGEDEP);
4546 	}
4547 	return (0);
4548 }
4549 
4550 /*
4551  * Writing back in-core inode structures.
4552  *
4553  * The filesystem only accesses an inode's contents when it occupies an
4554  * "in-core" inode structure.  These "in-core" structures are separate from
4555  * the page frames used to cache inode blocks.  Only the latter are
4556  * transferred to/from the disk.  So, when the updated contents of the
4557  * "in-core" inode structure are copied to the corresponding in-memory inode
4558  * block, the dependencies are also transferred.  The following procedure is
4559  * called when copying a dirty "in-core" inode to a cached inode block.
4560  */
4561 
4562 /*
4563  * Called when an inode is loaded from disk. If the effective link count
4564  * differed from the actual link count when it was last flushed, then we
4565  * need to ensure that the correct effective link count is put back.
4566  */
4567 void
4568 softdep_load_inodeblock(ip)
4569 	struct inode *ip;	/* the "in_core" copy of the inode */
4570 {
4571 	struct inodedep *inodedep;
4572 
4573 	/*
4574 	 * Check for alternate nlink count.
4575 	 */
4576 	ip->i_effnlink = ip->i_nlink;
4577 	ACQUIRE_LOCK(&lk);
4578 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
4579 		FREE_LOCK(&lk);
4580 		return;
4581 	}
4582 	ip->i_effnlink -= inodedep->id_nlinkdelta;
4583 	if (inodedep->id_state & SPACECOUNTED)
4584 		ip->i_flag |= IN_SPACECOUNTED;
4585 	FREE_LOCK(&lk);
4586 }
4587 
4588 /*
4589  * This routine is called just before the "in-core" inode
4590  * information is to be copied to the in-memory inode block.
4591  * Recall that an inode block contains several inodes. If
4592  * the force flag is set, then the dependencies will be
4593  * cleared so that the update can always be made. Note that
4594  * the buffer is locked when this routine is called, so we
4595  * will never be in the middle of writing the inode block
4596  * to disk.
4597  */
4598 void
4599 softdep_update_inodeblock(ip, bp, waitfor)
4600 	struct inode *ip;	/* the "in_core" copy of the inode */
4601 	struct buf *bp;		/* the buffer containing the inode block */
4602 	int waitfor;		/* nonzero => update must be allowed */
4603 {
4604 	struct inodedep *inodedep;
4605 	struct worklist *wk;
4606 	int error, gotit;
4607 
4608 	/*
4609 	 * If the effective link count is not equal to the actual link
4610 	 * count, then we must track the difference in an inodedep while
4611 	 * the inode is (potentially) tossed out of the cache. Otherwise,
4612 	 * if there is no existing inodedep, then there are no dependencies
4613 	 * to track.
4614 	 */
4615 	ACQUIRE_LOCK(&lk);
4616 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
4617 		FREE_LOCK(&lk);
4618 		if (ip->i_effnlink != ip->i_nlink)
4619 			panic("softdep_update_inodeblock: bad link count");
4620 		return;
4621 	}
4622 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
4623 		FREE_LOCK(&lk);
4624 		panic("softdep_update_inodeblock: bad delta");
4625 	}
4626 	/*
4627 	 * Changes have been initiated. Anything depending on these
4628 	 * changes cannot occur until this inode has been written.
4629 	 */
4630 	inodedep->id_state &= ~COMPLETE;
4631 	if ((inodedep->id_state & ONWORKLIST) == 0)
4632 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
4633 	/*
4634 	 * Any new dependencies associated with the incore inode must
4635 	 * now be moved to the list associated with the buffer holding
4636 	 * the in-memory copy of the inode. Once merged process any
4637 	 * allocdirects that are completed by the merger.
4638 	 */
4639 	merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
4640 	if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
4641 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
4642 	merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
4643 	if (TAILQ_FIRST(&inodedep->id_extupdt) != NULL)
4644 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt));
4645 	/*
4646 	 * Now that the inode has been pushed into the buffer, the
4647 	 * operations dependent on the inode being written to disk
4648 	 * can be moved to the id_bufwait so that they will be
4649 	 * processed when the buffer I/O completes.
4650 	 */
4651 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
4652 		WORKLIST_REMOVE(wk);
4653 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
4654 	}
4655 	/*
4656 	 * Newly allocated inodes cannot be written until the bitmap
4657 	 * that allocates them have been written (indicated by
4658 	 * DEPCOMPLETE being set in id_state). If we are doing a
4659 	 * forced sync (e.g., an fsync on a file), we force the bitmap
4660 	 * to be written so that the update can be done.
4661 	 */
4662 	if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
4663 		FREE_LOCK(&lk);
4664 		return;
4665 	}
4666 	gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4667 	FREE_LOCK(&lk);
4668 	if (gotit &&
4669 	    (error = BUF_WRITE(inodedep->id_buf)) != 0)
4670 		softdep_error("softdep_update_inodeblock: bwrite", error);
4671 	if ((inodedep->id_state & DEPCOMPLETE) == 0)
4672 		panic("softdep_update_inodeblock: update failed");
4673 }
4674 
4675 /*
4676  * Merge the a new inode dependency list (such as id_newinoupdt) into an
4677  * old inode dependency list (such as id_inoupdt). This routine must be
4678  * called with splbio interrupts blocked.
4679  */
4680 static void
4681 merge_inode_lists(newlisthead, oldlisthead)
4682 	struct allocdirectlst *newlisthead;
4683 	struct allocdirectlst *oldlisthead;
4684 {
4685 	struct allocdirect *listadp, *newadp;
4686 
4687 	newadp = TAILQ_FIRST(newlisthead);
4688 	for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
4689 		if (listadp->ad_lbn < newadp->ad_lbn) {
4690 			listadp = TAILQ_NEXT(listadp, ad_next);
4691 			continue;
4692 		}
4693 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4694 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
4695 		if (listadp->ad_lbn == newadp->ad_lbn) {
4696 			allocdirect_merge(oldlisthead, newadp,
4697 			    listadp);
4698 			listadp = newadp;
4699 		}
4700 		newadp = TAILQ_FIRST(newlisthead);
4701 	}
4702 	while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
4703 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4704 		TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
4705 	}
4706 }
4707 
4708 /*
4709  * If we are doing an fsync, then we must ensure that any directory
4710  * entries for the inode have been written after the inode gets to disk.
4711  */
4712 int
4713 softdep_fsync(vp)
4714 	struct vnode *vp;	/* the "in_core" copy of the inode */
4715 {
4716 	struct inodedep *inodedep;
4717 	struct pagedep *pagedep;
4718 	struct worklist *wk;
4719 	struct diradd *dap;
4720 	struct mount *mnt;
4721 	struct vnode *pvp;
4722 	struct inode *ip;
4723 	struct buf *bp;
4724 	struct fs *fs;
4725 	struct thread *td = curthread;
4726 	int error, flushparent;
4727 	ino_t parentino;
4728 	ufs_lbn_t lbn;
4729 
4730 	ip = VTOI(vp);
4731 	fs = ip->i_fs;
4732 	ACQUIRE_LOCK(&lk);
4733 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
4734 		FREE_LOCK(&lk);
4735 		return (0);
4736 	}
4737 	if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
4738 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
4739 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
4740 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
4741 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
4742 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) {
4743 		FREE_LOCK(&lk);
4744 		panic("softdep_fsync: pending ops");
4745 	}
4746 	for (error = 0, flushparent = 0; ; ) {
4747 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
4748 			break;
4749 		if (wk->wk_type != D_DIRADD) {
4750 			FREE_LOCK(&lk);
4751 			panic("softdep_fsync: Unexpected type %s",
4752 			    TYPENAME(wk->wk_type));
4753 		}
4754 		dap = WK_DIRADD(wk);
4755 		/*
4756 		 * Flush our parent if this directory entry has a MKDIR_PARENT
4757 		 * dependency or is contained in a newly allocated block.
4758 		 */
4759 		if (dap->da_state & DIRCHG)
4760 			pagedep = dap->da_previous->dm_pagedep;
4761 		else
4762 			pagedep = dap->da_pagedep;
4763 		mnt = pagedep->pd_mnt;
4764 		parentino = pagedep->pd_ino;
4765 		lbn = pagedep->pd_lbn;
4766 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) {
4767 			FREE_LOCK(&lk);
4768 			panic("softdep_fsync: dirty");
4769 		}
4770 		if ((dap->da_state & MKDIR_PARENT) ||
4771 		    (pagedep->pd_state & NEWBLOCK))
4772 			flushparent = 1;
4773 		else
4774 			flushparent = 0;
4775 		/*
4776 		 * If we are being fsync'ed as part of vgone'ing this vnode,
4777 		 * then we will not be able to release and recover the
4778 		 * vnode below, so we just have to give up on writing its
4779 		 * directory entry out. It will eventually be written, just
4780 		 * not now, but then the user was not asking to have it
4781 		 * written, so we are not breaking any promises.
4782 		 */
4783 		mp_fixme("This operation is not atomic wrt the rest of the code");
4784 		VI_LOCK(vp);
4785 		if (vp->v_iflag & VI_XLOCK) {
4786 			VI_UNLOCK(vp);
4787 			break;
4788 		} else
4789 			VI_UNLOCK(vp);
4790 		/*
4791 		 * We prevent deadlock by always fetching inodes from the
4792 		 * root, moving down the directory tree. Thus, when fetching
4793 		 * our parent directory, we first try to get the lock. If
4794 		 * that fails, we must unlock ourselves before requesting
4795 		 * the lock on our parent. See the comment in ufs_lookup
4796 		 * for details on possible races.
4797 		 */
4798 		FREE_LOCK(&lk);
4799 		if (VFS_VGET(mnt, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) {
4800 			VOP_UNLOCK(vp, 0, td);
4801 			error = VFS_VGET(mnt, parentino, LK_EXCLUSIVE, &pvp);
4802 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
4803 			if (error != 0)
4804 				return (error);
4805 		}
4806 		/*
4807 		 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
4808 		 * that are contained in direct blocks will be resolved by
4809 		 * doing a UFS_UPDATE. Pagedeps contained in indirect blocks
4810 		 * may require a complete sync'ing of the directory. So, we
4811 		 * try the cheap and fast UFS_UPDATE first, and if that fails,
4812 		 * then we do the slower VOP_FSYNC of the directory.
4813 		 */
4814 		if (flushparent) {
4815 			if ((error = UFS_UPDATE(pvp, 1)) != 0) {
4816 				vput(pvp);
4817 				return (error);
4818 			}
4819 			if ((pagedep->pd_state & NEWBLOCK) &&
4820 			    (error = VOP_FSYNC(pvp, td->td_ucred, MNT_WAIT, td))) {
4821 				vput(pvp);
4822 				return (error);
4823 			}
4824 		}
4825 		/*
4826 		 * Flush directory page containing the inode's name.
4827 		 */
4828 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
4829 		    &bp);
4830 		if (error == 0)
4831 			error = BUF_WRITE(bp);
4832 		else
4833 			brelse(bp);
4834 		vput(pvp);
4835 		if (error != 0)
4836 			return (error);
4837 		ACQUIRE_LOCK(&lk);
4838 		if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
4839 			break;
4840 	}
4841 	FREE_LOCK(&lk);
4842 	return (0);
4843 }
4844 
4845 /*
4846  * Flush all the dirty bitmaps associated with the block device
4847  * before flushing the rest of the dirty blocks so as to reduce
4848  * the number of dependencies that will have to be rolled back.
4849  */
4850 void
4851 softdep_fsync_mountdev(vp)
4852 	struct vnode *vp;
4853 {
4854 	struct buf *bp, *nbp;
4855 	struct worklist *wk;
4856 
4857 	if (!vn_isdisk(vp, NULL))
4858 		panic("softdep_fsync_mountdev: vnode not a disk");
4859 	ACQUIRE_LOCK(&lk);
4860 	for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
4861 		nbp = TAILQ_NEXT(bp, b_vnbufs);
4862 		/*
4863 		 * If it is already scheduled, skip to the next buffer.
4864 		 */
4865 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
4866 			continue;
4867 		if ((bp->b_flags & B_DELWRI) == 0) {
4868 			FREE_LOCK(&lk);
4869 			panic("softdep_fsync_mountdev: not dirty");
4870 		}
4871 		/*
4872 		 * We are only interested in bitmaps with outstanding
4873 		 * dependencies.
4874 		 */
4875 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
4876 		    wk->wk_type != D_BMSAFEMAP ||
4877 		    (bp->b_xflags & BX_BKGRDINPROG)) {
4878 			BUF_UNLOCK(bp);
4879 			continue;
4880 		}
4881 		bremfree(bp);
4882 		FREE_LOCK(&lk);
4883 		(void) bawrite(bp);
4884 		ACQUIRE_LOCK(&lk);
4885 		/*
4886 		 * Since we may have slept during the I/O, we need
4887 		 * to start from a known point.
4888 		 */
4889 		nbp = TAILQ_FIRST(&vp->v_dirtyblkhd);
4890 	}
4891 	drain_output(vp, 1);
4892 	FREE_LOCK(&lk);
4893 }
4894 
4895 /*
4896  * This routine is called when we are trying to synchronously flush a
4897  * file. This routine must eliminate any filesystem metadata dependencies
4898  * so that the syncing routine can succeed by pushing the dirty blocks
4899  * associated with the file. If any I/O errors occur, they are returned.
4900  */
4901 int
4902 softdep_sync_metadata(ap)
4903 	struct vop_fsync_args /* {
4904 		struct vnode *a_vp;
4905 		struct ucred *a_cred;
4906 		int a_waitfor;
4907 		struct thread *a_td;
4908 	} */ *ap;
4909 {
4910 	struct vnode *vp = ap->a_vp;
4911 	struct pagedep *pagedep;
4912 	struct allocdirect *adp;
4913 	struct allocindir *aip;
4914 	struct buf *bp, *nbp;
4915 	struct worklist *wk;
4916 	int i, error, waitfor;
4917 
4918 	/*
4919 	 * Check whether this vnode is involved in a filesystem
4920 	 * that is doing soft dependency processing.
4921 	 */
4922 	if (!vn_isdisk(vp, NULL)) {
4923 		if (!DOINGSOFTDEP(vp))
4924 			return (0);
4925 	} else
4926 		if (vp->v_rdev->si_mountpoint == NULL ||
4927 		    (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0)
4928 			return (0);
4929 	/*
4930 	 * Ensure that any direct block dependencies have been cleared.
4931 	 */
4932 	ACQUIRE_LOCK(&lk);
4933 	if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
4934 		FREE_LOCK(&lk);
4935 		return (error);
4936 	}
4937 	/*
4938 	 * For most files, the only metadata dependencies are the
4939 	 * cylinder group maps that allocate their inode or blocks.
4940 	 * The block allocation dependencies can be found by traversing
4941 	 * the dependency lists for any buffers that remain on their
4942 	 * dirty buffer list. The inode allocation dependency will
4943 	 * be resolved when the inode is updated with MNT_WAIT.
4944 	 * This work is done in two passes. The first pass grabs most
4945 	 * of the buffers and begins asynchronously writing them. The
4946 	 * only way to wait for these asynchronous writes is to sleep
4947 	 * on the filesystem vnode which may stay busy for a long time
4948 	 * if the filesystem is active. So, instead, we make a second
4949 	 * pass over the dependencies blocking on each write. In the
4950 	 * usual case we will be blocking against a write that we
4951 	 * initiated, so when it is done the dependency will have been
4952 	 * resolved. Thus the second pass is expected to end quickly.
4953 	 */
4954 	waitfor = MNT_NOWAIT;
4955 top:
4956 	/*
4957 	 * We must wait for any I/O in progress to finish so that
4958 	 * all potential buffers on the dirty list will be visible.
4959 	 */
4960 	drain_output(vp, 1);
4961 	if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) {
4962 		FREE_LOCK(&lk);
4963 		return (0);
4964 	}
4965 	bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
4966 	/* While syncing snapshots, we must allow recursive lookups */
4967 	bp->b_lock.lk_flags |= LK_CANRECURSE;
4968 loop:
4969 	/*
4970 	 * As we hold the buffer locked, none of its dependencies
4971 	 * will disappear.
4972 	 */
4973 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4974 		switch (wk->wk_type) {
4975 
4976 		case D_ALLOCDIRECT:
4977 			adp = WK_ALLOCDIRECT(wk);
4978 			if (adp->ad_state & DEPCOMPLETE)
4979 				continue;
4980 			nbp = adp->ad_buf;
4981 			if (getdirtybuf(&nbp, waitfor) == 0)
4982 				continue;
4983 			FREE_LOCK(&lk);
4984 			if (waitfor == MNT_NOWAIT) {
4985 				bawrite(nbp);
4986 			} else if ((error = BUF_WRITE(nbp)) != 0) {
4987 				break;
4988 			}
4989 			ACQUIRE_LOCK(&lk);
4990 			continue;
4991 
4992 		case D_ALLOCINDIR:
4993 			aip = WK_ALLOCINDIR(wk);
4994 			if (aip->ai_state & DEPCOMPLETE)
4995 				continue;
4996 			nbp = aip->ai_buf;
4997 			if (getdirtybuf(&nbp, waitfor) == 0)
4998 				continue;
4999 			FREE_LOCK(&lk);
5000 			if (waitfor == MNT_NOWAIT) {
5001 				bawrite(nbp);
5002 			} else if ((error = BUF_WRITE(nbp)) != 0) {
5003 				break;
5004 			}
5005 			ACQUIRE_LOCK(&lk);
5006 			continue;
5007 
5008 		case D_INDIRDEP:
5009 		restart:
5010 
5011 			LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
5012 				if (aip->ai_state & DEPCOMPLETE)
5013 					continue;
5014 				nbp = aip->ai_buf;
5015 				if (getdirtybuf(&nbp, MNT_WAIT) == 0)
5016 					goto restart;
5017 				FREE_LOCK(&lk);
5018 				if ((error = BUF_WRITE(nbp)) != 0) {
5019 					break;
5020 				}
5021 				ACQUIRE_LOCK(&lk);
5022 				goto restart;
5023 			}
5024 			continue;
5025 
5026 		case D_INODEDEP:
5027 			if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
5028 			    WK_INODEDEP(wk)->id_ino)) != 0) {
5029 				FREE_LOCK(&lk);
5030 				break;
5031 			}
5032 			continue;
5033 
5034 		case D_PAGEDEP:
5035 			/*
5036 			 * We are trying to sync a directory that may
5037 			 * have dependencies on both its own metadata
5038 			 * and/or dependencies on the inodes of any
5039 			 * recently allocated files. We walk its diradd
5040 			 * lists pushing out the associated inode.
5041 			 */
5042 			pagedep = WK_PAGEDEP(wk);
5043 			for (i = 0; i < DAHASHSZ; i++) {
5044 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
5045 					continue;
5046 				if ((error =
5047 				    flush_pagedep_deps(vp, pagedep->pd_mnt,
5048 						&pagedep->pd_diraddhd[i]))) {
5049 					FREE_LOCK(&lk);
5050 					break;
5051 				}
5052 			}
5053 			continue;
5054 
5055 		case D_MKDIR:
5056 			/*
5057 			 * This case should never happen if the vnode has
5058 			 * been properly sync'ed. However, if this function
5059 			 * is used at a place where the vnode has not yet
5060 			 * been sync'ed, this dependency can show up. So,
5061 			 * rather than panic, just flush it.
5062 			 */
5063 			nbp = WK_MKDIR(wk)->md_buf;
5064 			if (getdirtybuf(&nbp, waitfor) == 0)
5065 				continue;
5066 			FREE_LOCK(&lk);
5067 			if (waitfor == MNT_NOWAIT) {
5068 				bawrite(nbp);
5069 			} else if ((error = BUF_WRITE(nbp)) != 0) {
5070 				break;
5071 			}
5072 			ACQUIRE_LOCK(&lk);
5073 			continue;
5074 
5075 		case D_BMSAFEMAP:
5076 			/*
5077 			 * This case should never happen if the vnode has
5078 			 * been properly sync'ed. However, if this function
5079 			 * is used at a place where the vnode has not yet
5080 			 * been sync'ed, this dependency can show up. So,
5081 			 * rather than panic, just flush it.
5082 			 */
5083 			nbp = WK_BMSAFEMAP(wk)->sm_buf;
5084 			if (getdirtybuf(&nbp, waitfor) == 0)
5085 				continue;
5086 			FREE_LOCK(&lk);
5087 			if (waitfor == MNT_NOWAIT) {
5088 				bawrite(nbp);
5089 			} else if ((error = BUF_WRITE(nbp)) != 0) {
5090 				break;
5091 			}
5092 			ACQUIRE_LOCK(&lk);
5093 			continue;
5094 
5095 		default:
5096 			FREE_LOCK(&lk);
5097 			panic("softdep_sync_metadata: Unknown type %s",
5098 			    TYPENAME(wk->wk_type));
5099 			/* NOTREACHED */
5100 		}
5101 		/* We reach here only in error and unlocked */
5102 		if (error == 0)
5103 			panic("softdep_sync_metadata: zero error");
5104 		bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5105 		bawrite(bp);
5106 		return (error);
5107 	}
5108 	(void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT);
5109 	nbp = TAILQ_NEXT(bp, b_vnbufs);
5110 	FREE_LOCK(&lk);
5111 	bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5112 	bawrite(bp);
5113 	ACQUIRE_LOCK(&lk);
5114 	if (nbp != NULL) {
5115 		bp = nbp;
5116 		goto loop;
5117 	}
5118 	/*
5119 	 * The brief unlock is to allow any pent up dependency
5120 	 * processing to be done. Then proceed with the second pass.
5121 	 */
5122 	if (waitfor == MNT_NOWAIT) {
5123 		waitfor = MNT_WAIT;
5124 		FREE_LOCK(&lk);
5125 		ACQUIRE_LOCK(&lk);
5126 		goto top;
5127 	}
5128 
5129 	/*
5130 	 * If we have managed to get rid of all the dirty buffers,
5131 	 * then we are done. For certain directories and block
5132 	 * devices, we may need to do further work.
5133 	 *
5134 	 * We must wait for any I/O in progress to finish so that
5135 	 * all potential buffers on the dirty list will be visible.
5136 	 */
5137 	drain_output(vp, 1);
5138 	if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) {
5139 		FREE_LOCK(&lk);
5140 		return (0);
5141 	}
5142 
5143 	FREE_LOCK(&lk);
5144 	/*
5145 	 * If we are trying to sync a block device, some of its buffers may
5146 	 * contain metadata that cannot be written until the contents of some
5147 	 * partially written files have been written to disk. The only easy
5148 	 * way to accomplish this is to sync the entire filesystem (luckily
5149 	 * this happens rarely).
5150 	 */
5151 	if (vn_isdisk(vp, NULL) &&
5152 	    vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) &&
5153 	    (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, ap->a_cred,
5154 	     ap->a_td)) != 0)
5155 		return (error);
5156 	return (0);
5157 }
5158 
5159 /*
5160  * Flush the dependencies associated with an inodedep.
5161  * Called with splbio blocked.
5162  */
5163 static int
5164 flush_inodedep_deps(fs, ino)
5165 	struct fs *fs;
5166 	ino_t ino;
5167 {
5168 	struct inodedep *inodedep;
5169 	int error, waitfor;
5170 
5171 	/*
5172 	 * This work is done in two passes. The first pass grabs most
5173 	 * of the buffers and begins asynchronously writing them. The
5174 	 * only way to wait for these asynchronous writes is to sleep
5175 	 * on the filesystem vnode which may stay busy for a long time
5176 	 * if the filesystem is active. So, instead, we make a second
5177 	 * pass over the dependencies blocking on each write. In the
5178 	 * usual case we will be blocking against a write that we
5179 	 * initiated, so when it is done the dependency will have been
5180 	 * resolved. Thus the second pass is expected to end quickly.
5181 	 * We give a brief window at the top of the loop to allow
5182 	 * any pending I/O to complete.
5183 	 */
5184 	for (error = 0, waitfor = MNT_NOWAIT; ; ) {
5185 		if (error)
5186 			return (error);
5187 		FREE_LOCK(&lk);
5188 		ACQUIRE_LOCK(&lk);
5189 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
5190 			return (0);
5191 		if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
5192 		    flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
5193 		    flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
5194 		    flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
5195 			continue;
5196 		/*
5197 		 * If pass2, we are done, otherwise do pass 2.
5198 		 */
5199 		if (waitfor == MNT_WAIT)
5200 			break;
5201 		waitfor = MNT_WAIT;
5202 	}
5203 	/*
5204 	 * Try freeing inodedep in case all dependencies have been removed.
5205 	 */
5206 	if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
5207 		(void) free_inodedep(inodedep);
5208 	return (0);
5209 }
5210 
5211 /*
5212  * Flush an inode dependency list.
5213  * Called with splbio blocked.
5214  */
5215 static int
5216 flush_deplist(listhead, waitfor, errorp)
5217 	struct allocdirectlst *listhead;
5218 	int waitfor;
5219 	int *errorp;
5220 {
5221 	struct allocdirect *adp;
5222 	struct buf *bp;
5223 
5224 	TAILQ_FOREACH(adp, listhead, ad_next) {
5225 		if (adp->ad_state & DEPCOMPLETE)
5226 			continue;
5227 		bp = adp->ad_buf;
5228 		if (getdirtybuf(&bp, waitfor) == 0) {
5229 			if (waitfor == MNT_NOWAIT)
5230 				continue;
5231 			return (1);
5232 		}
5233 		FREE_LOCK(&lk);
5234 		if (waitfor == MNT_NOWAIT) {
5235 			bawrite(bp);
5236 		} else if ((*errorp = BUF_WRITE(bp)) != 0) {
5237 			ACQUIRE_LOCK(&lk);
5238 			return (1);
5239 		}
5240 		ACQUIRE_LOCK(&lk);
5241 		return (1);
5242 	}
5243 	return (0);
5244 }
5245 
5246 /*
5247  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
5248  * Called with splbio blocked.
5249  */
5250 static int
5251 flush_pagedep_deps(pvp, mp, diraddhdp)
5252 	struct vnode *pvp;
5253 	struct mount *mp;
5254 	struct diraddhd *diraddhdp;
5255 {
5256 	struct thread *td = curthread;
5257 	struct inodedep *inodedep;
5258 	struct ufsmount *ump;
5259 	struct diradd *dap;
5260 	struct vnode *vp;
5261 	int gotit, error = 0;
5262 	struct buf *bp;
5263 	ino_t inum;
5264 
5265 	ump = VFSTOUFS(mp);
5266 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
5267 		/*
5268 		 * Flush ourselves if this directory entry
5269 		 * has a MKDIR_PARENT dependency.
5270 		 */
5271 		if (dap->da_state & MKDIR_PARENT) {
5272 			FREE_LOCK(&lk);
5273 			if ((error = UFS_UPDATE(pvp, 1)) != 0)
5274 				break;
5275 			ACQUIRE_LOCK(&lk);
5276 			/*
5277 			 * If that cleared dependencies, go on to next.
5278 			 */
5279 			if (dap != LIST_FIRST(diraddhdp))
5280 				continue;
5281 			if (dap->da_state & MKDIR_PARENT) {
5282 				FREE_LOCK(&lk);
5283 				panic("flush_pagedep_deps: MKDIR_PARENT");
5284 			}
5285 		}
5286 		/*
5287 		 * A newly allocated directory must have its "." and
5288 		 * ".." entries written out before its name can be
5289 		 * committed in its parent. We do not want or need
5290 		 * the full semantics of a synchronous VOP_FSYNC as
5291 		 * that may end up here again, once for each directory
5292 		 * level in the filesystem. Instead, we push the blocks
5293 		 * and wait for them to clear. We have to fsync twice
5294 		 * because the first call may choose to defer blocks
5295 		 * that still have dependencies, but deferral will
5296 		 * happen at most once.
5297 		 */
5298 		inum = dap->da_newinum;
5299 		if (dap->da_state & MKDIR_BODY) {
5300 			FREE_LOCK(&lk);
5301 			if ((error = VFS_VGET(mp, inum, LK_EXCLUSIVE, &vp)))
5302 				break;
5303 			if ((error=VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)) ||
5304 			    (error=VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td))) {
5305 				vput(vp);
5306 				break;
5307 			}
5308 			drain_output(vp, 0);
5309 			vput(vp);
5310 			ACQUIRE_LOCK(&lk);
5311 			/*
5312 			 * If that cleared dependencies, go on to next.
5313 			 */
5314 			if (dap != LIST_FIRST(diraddhdp))
5315 				continue;
5316 			if (dap->da_state & MKDIR_BODY) {
5317 				FREE_LOCK(&lk);
5318 				panic("flush_pagedep_deps: MKDIR_BODY");
5319 			}
5320 		}
5321 		/*
5322 		 * Flush the inode on which the directory entry depends.
5323 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
5324 		 * the only remaining dependency is that the updated inode
5325 		 * count must get pushed to disk. The inode has already
5326 		 * been pushed into its inode buffer (via VOP_UPDATE) at
5327 		 * the time of the reference count change. So we need only
5328 		 * locate that buffer, ensure that there will be no rollback
5329 		 * caused by a bitmap dependency, then write the inode buffer.
5330 		 */
5331 		if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) {
5332 			FREE_LOCK(&lk);
5333 			panic("flush_pagedep_deps: lost inode");
5334 		}
5335 		/*
5336 		 * If the inode still has bitmap dependencies,
5337 		 * push them to disk.
5338 		 */
5339 		if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5340 			gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
5341 			FREE_LOCK(&lk);
5342 			if (gotit &&
5343 			    (error = BUF_WRITE(inodedep->id_buf)) != 0)
5344 				break;
5345 			ACQUIRE_LOCK(&lk);
5346 			if (dap != LIST_FIRST(diraddhdp))
5347 				continue;
5348 		}
5349 		/*
5350 		 * If the inode is still sitting in a buffer waiting
5351 		 * to be written, push it to disk.
5352 		 */
5353 		FREE_LOCK(&lk);
5354 		if ((error = bread(ump->um_devvp,
5355 		    fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
5356 		    (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) {
5357 			brelse(bp);
5358 			break;
5359 		}
5360 		if ((error = BUF_WRITE(bp)) != 0)
5361 			break;
5362 		ACQUIRE_LOCK(&lk);
5363 		/*
5364 		 * If we have failed to get rid of all the dependencies
5365 		 * then something is seriously wrong.
5366 		 */
5367 		if (dap == LIST_FIRST(diraddhdp)) {
5368 			FREE_LOCK(&lk);
5369 			panic("flush_pagedep_deps: flush failed");
5370 		}
5371 	}
5372 	if (error)
5373 		ACQUIRE_LOCK(&lk);
5374 	return (error);
5375 }
5376 
5377 /*
5378  * A large burst of file addition or deletion activity can drive the
5379  * memory load excessively high. First attempt to slow things down
5380  * using the techniques below. If that fails, this routine requests
5381  * the offending operations to fall back to running synchronously
5382  * until the memory load returns to a reasonable level.
5383  */
5384 int
5385 softdep_slowdown(vp)
5386 	struct vnode *vp;
5387 {
5388 	int max_softdeps_hard;
5389 
5390 	max_softdeps_hard = max_softdeps * 11 / 10;
5391 	if (num_dirrem < max_softdeps_hard / 2 &&
5392 	    num_inodedep < max_softdeps_hard)
5393 		return (0);
5394 	stat_sync_limit_hit += 1;
5395 	return (1);
5396 }
5397 
5398 /*
5399  * Called by the allocation routines when they are about to fail
5400  * in the hope that we can free up some disk space.
5401  *
5402  * First check to see if the work list has anything on it. If it has,
5403  * clean up entries until we successfully free some space. Because this
5404  * process holds inodes locked, we cannot handle any remove requests
5405  * that might block on a locked inode as that could lead to deadlock.
5406  * If the worklist yields no free space, encourage the syncer daemon
5407  * to help us. In no event will we try for longer than tickdelay seconds.
5408  */
5409 int
5410 softdep_request_cleanup(fs, vp)
5411 	struct fs *fs;
5412 	struct vnode *vp;
5413 {
5414 	long starttime;
5415 	ufs2_daddr_t needed;
5416 
5417 	needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize;
5418 	starttime = time_second + tickdelay;
5419 	if (UFS_UPDATE(vp, 1) != 0)
5420 		return (0);
5421 	while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) {
5422 		if (time_second > starttime)
5423 			return (0);
5424 		if (num_on_worklist > 0 &&
5425 		    process_worklist_item(NULL, LK_NOWAIT) != -1) {
5426 			stat_worklist_push += 1;
5427 			continue;
5428 		}
5429 		request_cleanup(FLUSH_REMOVE_WAIT, 0);
5430 	}
5431 	return (1);
5432 }
5433 
5434 /*
5435  * If memory utilization has gotten too high, deliberately slow things
5436  * down and speed up the I/O processing.
5437  */
5438 static int
5439 request_cleanup(resource, islocked)
5440 	int resource;
5441 	int islocked;
5442 {
5443 	struct thread *td = curthread;
5444 
5445 	/*
5446 	 * We never hold up the filesystem syncer process.
5447 	 */
5448 	if (td == filesys_syncer)
5449 		return (0);
5450 	/*
5451 	 * First check to see if the work list has gotten backlogged.
5452 	 * If it has, co-opt this process to help clean up two entries.
5453 	 * Because this process may hold inodes locked, we cannot
5454 	 * handle any remove requests that might block on a locked
5455 	 * inode as that could lead to deadlock.
5456 	 */
5457 	if (num_on_worklist > max_softdeps / 10) {
5458 		if (islocked)
5459 			FREE_LOCK(&lk);
5460 		process_worklist_item(NULL, LK_NOWAIT);
5461 		process_worklist_item(NULL, LK_NOWAIT);
5462 		stat_worklist_push += 2;
5463 		if (islocked)
5464 			ACQUIRE_LOCK(&lk);
5465 		return(1);
5466 	}
5467 	/*
5468 	 * Next, we attempt to speed up the syncer process. If that
5469 	 * is successful, then we allow the process to continue.
5470 	 */
5471 	if (speedup_syncer() && resource != FLUSH_REMOVE_WAIT)
5472 		return(0);
5473 	/*
5474 	 * If we are resource constrained on inode dependencies, try
5475 	 * flushing some dirty inodes. Otherwise, we are constrained
5476 	 * by file deletions, so try accelerating flushes of directories
5477 	 * with removal dependencies. We would like to do the cleanup
5478 	 * here, but we probably hold an inode locked at this point and
5479 	 * that might deadlock against one that we try to clean. So,
5480 	 * the best that we can do is request the syncer daemon to do
5481 	 * the cleanup for us.
5482 	 */
5483 	switch (resource) {
5484 
5485 	case FLUSH_INODES:
5486 		stat_ino_limit_push += 1;
5487 		req_clear_inodedeps += 1;
5488 		stat_countp = &stat_ino_limit_hit;
5489 		break;
5490 
5491 	case FLUSH_REMOVE:
5492 	case FLUSH_REMOVE_WAIT:
5493 		stat_blk_limit_push += 1;
5494 		req_clear_remove += 1;
5495 		stat_countp = &stat_blk_limit_hit;
5496 		break;
5497 
5498 	default:
5499 		if (islocked)
5500 			FREE_LOCK(&lk);
5501 		panic("request_cleanup: unknown type");
5502 	}
5503 	/*
5504 	 * Hopefully the syncer daemon will catch up and awaken us.
5505 	 * We wait at most tickdelay before proceeding in any case.
5506 	 */
5507 	if (islocked == 0)
5508 		ACQUIRE_LOCK(&lk);
5509 	proc_waiting += 1;
5510 	if (handle.callout == NULL)
5511 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5512 	interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, NULL, PPAUSE,
5513 	    "softupdate", 0);
5514 	proc_waiting -= 1;
5515 	if (islocked == 0)
5516 		FREE_LOCK(&lk);
5517 	return (1);
5518 }
5519 
5520 /*
5521  * Awaken processes pausing in request_cleanup and clear proc_waiting
5522  * to indicate that there is no longer a timer running.
5523  */
5524 void
5525 pause_timer(arg)
5526 	void *arg;
5527 {
5528 
5529 	*stat_countp += 1;
5530 	wakeup_one(&proc_waiting);
5531 	if (proc_waiting > 0)
5532 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5533 	else
5534 		handle.callout = NULL;
5535 }
5536 
5537 /*
5538  * Flush out a directory with at least one removal dependency in an effort to
5539  * reduce the number of dirrem, freefile, and freeblks dependency structures.
5540  */
5541 static void
5542 clear_remove(td)
5543 	struct thread *td;
5544 {
5545 	struct pagedep_hashhead *pagedephd;
5546 	struct pagedep *pagedep;
5547 	static int next = 0;
5548 	struct mount *mp;
5549 	struct vnode *vp;
5550 	int error, cnt;
5551 	ino_t ino;
5552 
5553 	ACQUIRE_LOCK(&lk);
5554 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
5555 		pagedephd = &pagedep_hashtbl[next++];
5556 		if (next >= pagedep_hash)
5557 			next = 0;
5558 		LIST_FOREACH(pagedep, pagedephd, pd_hash) {
5559 			if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
5560 				continue;
5561 			mp = pagedep->pd_mnt;
5562 			ino = pagedep->pd_ino;
5563 			FREE_LOCK(&lk);
5564 			if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5565 				continue;
5566 			if ((error = VFS_VGET(mp, ino, LK_EXCLUSIVE, &vp))) {
5567 				softdep_error("clear_remove: vget", error);
5568 				vn_finished_write(mp);
5569 				return;
5570 			}
5571 			if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)))
5572 				softdep_error("clear_remove: fsync", error);
5573 			drain_output(vp, 0);
5574 			vput(vp);
5575 			vn_finished_write(mp);
5576 			return;
5577 		}
5578 	}
5579 	FREE_LOCK(&lk);
5580 }
5581 
5582 /*
5583  * Clear out a block of dirty inodes in an effort to reduce
5584  * the number of inodedep dependency structures.
5585  */
5586 static void
5587 clear_inodedeps(td)
5588 	struct thread *td;
5589 {
5590 	struct inodedep_hashhead *inodedephd;
5591 	struct inodedep *inodedep;
5592 	static int next = 0;
5593 	struct mount *mp;
5594 	struct vnode *vp;
5595 	struct fs *fs;
5596 	int error, cnt;
5597 	ino_t firstino, lastino, ino;
5598 
5599 	ACQUIRE_LOCK(&lk);
5600 	/*
5601 	 * Pick a random inode dependency to be cleared.
5602 	 * We will then gather up all the inodes in its block
5603 	 * that have dependencies and flush them out.
5604 	 */
5605 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
5606 		inodedephd = &inodedep_hashtbl[next++];
5607 		if (next >= inodedep_hash)
5608 			next = 0;
5609 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
5610 			break;
5611 	}
5612 	if (inodedep == NULL)
5613 		return;
5614 	/*
5615 	 * Ugly code to find mount point given pointer to superblock.
5616 	 */
5617 	fs = inodedep->id_fs;
5618 	TAILQ_FOREACH(mp, &mountlist, mnt_list)
5619 		if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs)
5620 			break;
5621 	/*
5622 	 * Find the last inode in the block with dependencies.
5623 	 */
5624 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
5625 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
5626 		if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
5627 			break;
5628 	/*
5629 	 * Asynchronously push all but the last inode with dependencies.
5630 	 * Synchronously push the last inode with dependencies to ensure
5631 	 * that the inode block gets written to free up the inodedeps.
5632 	 */
5633 	for (ino = firstino; ino <= lastino; ino++) {
5634 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
5635 			continue;
5636 		FREE_LOCK(&lk);
5637 		if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5638 			continue;
5639 		if ((error = VFS_VGET(mp, ino, LK_EXCLUSIVE, &vp)) != 0) {
5640 			softdep_error("clear_inodedeps: vget", error);
5641 			vn_finished_write(mp);
5642 			return;
5643 		}
5644 		if (ino == lastino) {
5645 			if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_WAIT, td)))
5646 				softdep_error("clear_inodedeps: fsync1", error);
5647 		} else {
5648 			if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)))
5649 				softdep_error("clear_inodedeps: fsync2", error);
5650 			drain_output(vp, 0);
5651 		}
5652 		vput(vp);
5653 		vn_finished_write(mp);
5654 		ACQUIRE_LOCK(&lk);
5655 	}
5656 	FREE_LOCK(&lk);
5657 }
5658 
5659 /*
5660  * Function to determine if the buffer has outstanding dependencies
5661  * that will cause a roll-back if the buffer is written. If wantcount
5662  * is set, return number of dependencies, otherwise just yes or no.
5663  */
5664 static int
5665 softdep_count_dependencies(bp, wantcount)
5666 	struct buf *bp;
5667 	int wantcount;
5668 {
5669 	struct worklist *wk;
5670 	struct inodedep *inodedep;
5671 	struct indirdep *indirdep;
5672 	struct allocindir *aip;
5673 	struct pagedep *pagedep;
5674 	struct diradd *dap;
5675 	int i, retval;
5676 
5677 	retval = 0;
5678 	ACQUIRE_LOCK(&lk);
5679 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5680 		switch (wk->wk_type) {
5681 
5682 		case D_INODEDEP:
5683 			inodedep = WK_INODEDEP(wk);
5684 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5685 				/* bitmap allocation dependency */
5686 				retval += 1;
5687 				if (!wantcount)
5688 					goto out;
5689 			}
5690 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
5691 				/* direct block pointer dependency */
5692 				retval += 1;
5693 				if (!wantcount)
5694 					goto out;
5695 			}
5696 			if (TAILQ_FIRST(&inodedep->id_extupdt)) {
5697 				/* direct block pointer dependency */
5698 				retval += 1;
5699 				if (!wantcount)
5700 					goto out;
5701 			}
5702 			continue;
5703 
5704 		case D_INDIRDEP:
5705 			indirdep = WK_INDIRDEP(wk);
5706 
5707 			LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
5708 				/* indirect block pointer dependency */
5709 				retval += 1;
5710 				if (!wantcount)
5711 					goto out;
5712 			}
5713 			continue;
5714 
5715 		case D_PAGEDEP:
5716 			pagedep = WK_PAGEDEP(wk);
5717 			for (i = 0; i < DAHASHSZ; i++) {
5718 
5719 				LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
5720 					/* directory entry dependency */
5721 					retval += 1;
5722 					if (!wantcount)
5723 						goto out;
5724 				}
5725 			}
5726 			continue;
5727 
5728 		case D_BMSAFEMAP:
5729 		case D_ALLOCDIRECT:
5730 		case D_ALLOCINDIR:
5731 		case D_MKDIR:
5732 			/* never a dependency on these blocks */
5733 			continue;
5734 
5735 		default:
5736 			FREE_LOCK(&lk);
5737 			panic("softdep_check_for_rollback: Unexpected type %s",
5738 			    TYPENAME(wk->wk_type));
5739 			/* NOTREACHED */
5740 		}
5741 	}
5742 out:
5743 	FREE_LOCK(&lk);
5744 	return retval;
5745 }
5746 
5747 /*
5748  * Acquire exclusive access to a buffer.
5749  * Must be called with splbio blocked.
5750  * Return 1 if buffer was acquired.
5751  */
5752 static int
5753 getdirtybuf(bpp, waitfor)
5754 	struct buf **bpp;
5755 	int waitfor;
5756 {
5757 	struct buf *bp;
5758 	int error;
5759 
5760 	for (;;) {
5761 		if ((bp = *bpp) == NULL)
5762 			return (0);
5763 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
5764 			if ((bp->b_xflags & BX_BKGRDINPROG) == 0)
5765 				break;
5766 			BUF_UNLOCK(bp);
5767 			if (waitfor != MNT_WAIT)
5768 				return (0);
5769 			bp->b_xflags |= BX_BKGRDWAIT;
5770 			interlocked_sleep(&lk, SLEEP, &bp->b_xflags, NULL,
5771 			    PRIBIO, "getbuf", 0);
5772 			continue;
5773 		}
5774 		if (waitfor != MNT_WAIT)
5775 			return (0);
5776 		error = interlocked_sleep(&lk, LOCKBUF, bp, NULL,
5777 		    LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0);
5778 		if (error != ENOLCK) {
5779 			FREE_LOCK(&lk);
5780 			panic("getdirtybuf: inconsistent lock");
5781 		}
5782 	}
5783 	if ((bp->b_flags & B_DELWRI) == 0) {
5784 		BUF_UNLOCK(bp);
5785 		return (0);
5786 	}
5787 	bremfree(bp);
5788 	return (1);
5789 }
5790 
5791 /*
5792  * Wait for pending output on a vnode to complete.
5793  * Must be called with vnode locked.
5794  */
5795 static void
5796 drain_output(vp, islocked)
5797 	struct vnode *vp;
5798 	int islocked;
5799 {
5800 
5801 	if (!islocked)
5802 		ACQUIRE_LOCK(&lk);
5803 	VI_LOCK(vp);
5804 	while (vp->v_numoutput) {
5805 		vp->v_iflag |= VI_BWAIT;
5806 		interlocked_sleep(&lk, SLEEP, (caddr_t)&vp->v_numoutput,
5807 		    VI_MTX(vp), PRIBIO + 1, "drainvp", 0);
5808 	}
5809 	VI_UNLOCK(vp);
5810 	if (!islocked)
5811 		FREE_LOCK(&lk);
5812 }
5813 
5814 /*
5815  * Called whenever a buffer that is being invalidated or reallocated
5816  * contains dependencies. This should only happen if an I/O error has
5817  * occurred. The routine is called with the buffer locked.
5818  */
5819 static void
5820 softdep_deallocate_dependencies(bp)
5821 	struct buf *bp;
5822 {
5823 
5824 	if ((bp->b_ioflags & BIO_ERROR) == 0)
5825 		panic("softdep_deallocate_dependencies: dangling deps");
5826 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
5827 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
5828 }
5829 
5830 /*
5831  * Function to handle asynchronous write errors in the filesystem.
5832  */
5833 void
5834 softdep_error(func, error)
5835 	char *func;
5836 	int error;
5837 {
5838 
5839 	/* XXX should do something better! */
5840 	printf("%s: got error %d while accessing filesystem\n", func, error);
5841 }
5842