xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision d37ea99837e6ad50837fd9fe1771ddf1c3ba6002)
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/bio.h>
58 #include <sys/buf.h>
59 #include <sys/malloc.h>
60 #include <sys/mount.h>
61 #include <sys/proc.h>
62 #include <sys/stat.h>
63 #include <sys/syslog.h>
64 #include <sys/vnode.h>
65 #include <sys/conf.h>
66 #include <ufs/ufs/dir.h>
67 #include <ufs/ufs/extattr.h>
68 #include <ufs/ufs/quota.h>
69 #include <ufs/ufs/inode.h>
70 #include <ufs/ufs/ufsmount.h>
71 #include <ufs/ffs/fs.h>
72 #include <ufs/ffs/softdep.h>
73 #include <ufs/ffs/ffs_extern.h>
74 #include <ufs/ufs/ufs_extern.h>
75 
76 /*
77  * These definitions need to be adapted to the system to which
78  * this file is being ported.
79  */
80 /*
81  * malloc types defined for the softdep system.
82  */
83 static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
84 static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
85 static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
86 static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
87 static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
88 static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
89 static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
90 static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
91 static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
92 static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
93 static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
94 static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
95 static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
96 static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block");
97 
98 #define M_SOFTDEP_FLAGS	(M_WAITOK | M_USE_RESERVE)
99 
100 #define	D_PAGEDEP	0
101 #define	D_INODEDEP	1
102 #define	D_NEWBLK	2
103 #define	D_BMSAFEMAP	3
104 #define	D_ALLOCDIRECT	4
105 #define	D_INDIRDEP	5
106 #define	D_ALLOCINDIR	6
107 #define	D_FREEFRAG	7
108 #define	D_FREEBLKS	8
109 #define	D_FREEFILE	9
110 #define	D_DIRADD	10
111 #define	D_MKDIR		11
112 #define	D_DIRREM	12
113 #define	D_NEWDIRBLK	13
114 #define	D_LAST		D_NEWDIRBLK
115 
116 /*
117  * translate from workitem type to memory type
118  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
119  */
120 static struct malloc_type *memtype[] = {
121 	M_PAGEDEP,
122 	M_INODEDEP,
123 	M_NEWBLK,
124 	M_BMSAFEMAP,
125 	M_ALLOCDIRECT,
126 	M_INDIRDEP,
127 	M_ALLOCINDIR,
128 	M_FREEFRAG,
129 	M_FREEBLKS,
130 	M_FREEFILE,
131 	M_DIRADD,
132 	M_MKDIR,
133 	M_DIRREM,
134 	M_NEWDIRBLK
135 };
136 
137 #define DtoM(type) (memtype[type])
138 
139 /*
140  * Names of malloc types.
141  */
142 #define TYPENAME(type)  \
143 	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
144 /*
145  * End system adaptaion definitions.
146  */
147 
148 /*
149  * Internal function prototypes.
150  */
151 static	void softdep_error(char *, int);
152 static	void drain_output(struct vnode *, int);
153 static	struct buf *getdirtybuf(struct buf **, struct mtx *, int);
154 static	void clear_remove(struct thread *);
155 static	void clear_inodedeps(struct thread *);
156 static	int flush_pagedep_deps(struct vnode *, struct mount *,
157 	    struct diraddhd *);
158 static	int flush_inodedep_deps(struct fs *, ino_t);
159 static	int flush_deplist(struct allocdirectlst *, int, int *);
160 static	int handle_written_filepage(struct pagedep *, struct buf *);
161 static  void diradd_inode_written(struct diradd *, struct inodedep *);
162 static	int handle_written_inodeblock(struct inodedep *, struct buf *);
163 static	void handle_allocdirect_partdone(struct allocdirect *);
164 static	void handle_allocindir_partdone(struct allocindir *);
165 static	void initiate_write_filepage(struct pagedep *, struct buf *);
166 static	void handle_written_mkdir(struct mkdir *, int);
167 static	void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
168 static	void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
169 static	void handle_workitem_freefile(struct freefile *);
170 static	void handle_workitem_remove(struct dirrem *, struct vnode *);
171 static	struct dirrem *newdirrem(struct buf *, struct inode *,
172 	    struct inode *, int, struct dirrem **);
173 static	void free_diradd(struct diradd *);
174 static	void free_allocindir(struct allocindir *, struct inodedep *);
175 static	void free_newdirblk(struct newdirblk *);
176 static	int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t,
177 	    ufs2_daddr_t *);
178 static	void deallocate_dependencies(struct buf *, struct inodedep *);
179 static	void free_allocdirect(struct allocdirectlst *,
180 	    struct allocdirect *, int);
181 static	int check_inode_unwritten(struct inodedep *);
182 static	int free_inodedep(struct inodedep *);
183 static	void handle_workitem_freeblocks(struct freeblks *, int);
184 static	void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
185 static	void setup_allocindir_phase2(struct buf *, struct inode *,
186 	    struct allocindir *);
187 static	struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
188 	    ufs2_daddr_t);
189 static	void handle_workitem_freefrag(struct freefrag *);
190 static	struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long);
191 static	void allocdirect_merge(struct allocdirectlst *,
192 	    struct allocdirect *, struct allocdirect *);
193 static	struct bmsafemap *bmsafemap_lookup(struct buf *);
194 static	int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **);
195 static	int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
196 static	int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **);
197 static	void pause_timer(void *);
198 static	int request_cleanup(int, int);
199 static	int process_worklist_item(struct mount *, int);
200 static	void add_to_worklist(struct worklist *);
201 
202 /*
203  * Exported softdep operations.
204  */
205 static	int softdep_disk_prewrite(struct vnode *vp, struct buf *bp);
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 
242 static	void acquire_lock(struct lockit *);
243 static	void free_lock(struct lockit *);
244 void	softdep_panic(char *);
245 
246 #define ACQUIRE_LOCK(lk)		acquire_lock(lk)
247 #define FREE_LOCK(lk)			free_lock(lk)
248 
249 static void
250 acquire_lock(lk)
251 	struct lockit *lk;
252 {
253 	struct thread *holder;
254 
255 	if (lk->lkt_held != NOHOLDER) {
256 		holder = lk->lkt_held;
257 		FREE_LOCK(lk);
258 		if (holder == curthread)
259 			panic("softdep_lock: locking against myself");
260 		else
261 			panic("softdep_lock: lock held by %p", holder);
262 	}
263 	lk->lkt_spl = splbio();
264 	lk->lkt_held = curthread;
265 }
266 
267 static void
268 free_lock(lk)
269 	struct lockit *lk;
270 {
271 
272 	if (lk->lkt_held == NOHOLDER)
273 		panic("softdep_unlock: lock not held");
274 	lk->lkt_held = NOHOLDER;
275 	splx(lk->lkt_spl);
276 }
277 
278 /*
279  * Function to release soft updates lock and panic.
280  */
281 void
282 softdep_panic(msg)
283 	char *msg;
284 {
285 
286 	if (lk.lkt_held != NOHOLDER)
287 		FREE_LOCK(&lk);
288 	panic(msg);
289 }
290 #endif /* DEBUG */
291 
292 static	int interlocked_sleep(struct lockit *, int, void *, struct mtx *, int,
293 	    const char *, int);
294 
295 /*
296  * When going to sleep, we must save our SPL so that it does
297  * not get lost if some other process uses the lock while we
298  * are sleeping. We restore it after we have slept. This routine
299  * wraps the interlocking with functions that sleep. The list
300  * below enumerates the available set of operations.
301  */
302 #define	UNKNOWN		0
303 #define	SLEEP		1
304 #define	LOCKBUF		2
305 
306 static int
307 interlocked_sleep(lk, op, ident, mtx, flags, wmesg, timo)
308 	struct lockit *lk;
309 	int op;
310 	void *ident;
311 	struct mtx *mtx;
312 	int flags;
313 	const char *wmesg;
314 	int timo;
315 {
316 	struct thread *holder;
317 	int s, retval;
318 
319 	s = lk->lkt_spl;
320 #	ifdef DEBUG
321 	if (lk->lkt_held == NOHOLDER)
322 		panic("interlocked_sleep: lock not held");
323 	lk->lkt_held = NOHOLDER;
324 #	endif /* DEBUG */
325 	switch (op) {
326 	case SLEEP:
327 		retval = msleep(ident, mtx, flags, wmesg, timo);
328 		break;
329 	case LOCKBUF:
330 		retval = BUF_LOCK((struct buf *)ident, flags, mtx);
331 		break;
332 	default:
333 		panic("interlocked_sleep: unknown operation");
334 	}
335 #	ifdef DEBUG
336 	if (lk->lkt_held != NOHOLDER) {
337 		holder = lk->lkt_held;
338 		FREE_LOCK(lk);
339 		if (holder == curthread)
340 			panic("interlocked_sleep: locking against self");
341 		else
342 			panic("interlocked_sleep: lock held by %p", holder);
343 	}
344 	lk->lkt_held = curthread;
345 #	endif /* DEBUG */
346 	lk->lkt_spl = s;
347 	return (retval);
348 }
349 
350 /*
351  * Place holder for real semaphores.
352  */
353 struct sema {
354 	int	value;
355 	struct	thread *holder;
356 	char	*name;
357 	int	prio;
358 	int	timo;
359 };
360 static	void sema_init(struct sema *, char *, int, int);
361 static	int sema_get(struct sema *, struct lockit *);
362 static	void sema_release(struct sema *);
363 
364 static void
365 sema_init(semap, name, prio, timo)
366 	struct sema *semap;
367 	char *name;
368 	int prio, timo;
369 {
370 
371 	semap->holder = NOHOLDER;
372 	semap->value = 0;
373 	semap->name = name;
374 	semap->prio = prio;
375 	semap->timo = timo;
376 }
377 
378 static int
379 sema_get(semap, interlock)
380 	struct sema *semap;
381 	struct lockit *interlock;
382 {
383 
384 	if (semap->value++ > 0) {
385 		if (interlock != NULL) {
386 			interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
387 			    NULL, semap->prio, semap->name,
388 			    semap->timo);
389 			FREE_LOCK(interlock);
390 		} else {
391 			tsleep(semap, semap->prio, semap->name,
392 			    semap->timo);
393 		}
394 		return (0);
395 	}
396 	semap->holder = curthread;
397 	if (interlock != NULL)
398 		FREE_LOCK(interlock);
399 	return (1);
400 }
401 
402 static void
403 sema_release(semap)
404 	struct sema *semap;
405 {
406 
407 	if (semap->value <= 0 || semap->holder != curthread) {
408 		if (lk.lkt_held != NOHOLDER)
409 			FREE_LOCK(&lk);
410 		panic("sema_release: not held");
411 	}
412 	if (--semap->value > 0) {
413 		semap->value = 0;
414 		wakeup(semap);
415 	}
416 	semap->holder = NOHOLDER;
417 }
418 
419 /*
420  * Worklist queue management.
421  * These routines require that the lock be held.
422  */
423 #ifndef /* NOT */ DEBUG
424 #define WORKLIST_INSERT(head, item) do {	\
425 	(item)->wk_state |= ONWORKLIST;		\
426 	LIST_INSERT_HEAD(head, item, wk_list);	\
427 } while (0)
428 #define WORKLIST_REMOVE(item) do {		\
429 	(item)->wk_state &= ~ONWORKLIST;	\
430 	LIST_REMOVE(item, wk_list);		\
431 } while (0)
432 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
433 
434 #else /* DEBUG */
435 static	void worklist_insert(struct workhead *, struct worklist *);
436 static	void worklist_remove(struct worklist *);
437 static	void workitem_free(struct worklist *, int);
438 
439 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
440 #define WORKLIST_REMOVE(item) worklist_remove(item)
441 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
442 
443 static void
444 worklist_insert(head, item)
445 	struct workhead *head;
446 	struct worklist *item;
447 {
448 
449 	if (lk.lkt_held == NOHOLDER)
450 		panic("worklist_insert: lock not held");
451 	if (item->wk_state & ONWORKLIST) {
452 		FREE_LOCK(&lk);
453 		panic("worklist_insert: already on list");
454 	}
455 	item->wk_state |= ONWORKLIST;
456 	LIST_INSERT_HEAD(head, item, wk_list);
457 }
458 
459 static void
460 worklist_remove(item)
461 	struct worklist *item;
462 {
463 
464 	if (lk.lkt_held == NOHOLDER)
465 		panic("worklist_remove: lock not held");
466 	if ((item->wk_state & ONWORKLIST) == 0) {
467 		FREE_LOCK(&lk);
468 		panic("worklist_remove: not on list");
469 	}
470 	item->wk_state &= ~ONWORKLIST;
471 	LIST_REMOVE(item, wk_list);
472 }
473 
474 static void
475 workitem_free(item, type)
476 	struct worklist *item;
477 	int type;
478 {
479 
480 	if (item->wk_state & ONWORKLIST) {
481 		if (lk.lkt_held != NOHOLDER)
482 			FREE_LOCK(&lk);
483 		panic("workitem_free: still on list");
484 	}
485 	if (item->wk_type != type) {
486 		if (lk.lkt_held != NOHOLDER)
487 			FREE_LOCK(&lk);
488 		panic("workitem_free: type mismatch");
489 	}
490 	FREE(item, DtoM(type));
491 }
492 #endif /* DEBUG */
493 
494 /*
495  * Workitem queue management
496  */
497 static struct workhead softdep_workitem_pending;
498 static struct worklist *worklist_tail;
499 static int num_on_worklist;	/* number of worklist items to be processed */
500 static int softdep_worklist_busy; /* 1 => trying to do unmount */
501 static int softdep_worklist_req; /* serialized waiters */
502 static int max_softdeps;	/* maximum number of structs before slowdown */
503 static int maxindirdeps = 50;	/* max number of indirdeps before slowdown */
504 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
505 static int proc_waiting;	/* tracks whether we have a timeout posted */
506 static int *stat_countp;	/* statistic to count in proc_waiting timeout */
507 static struct callout_handle handle; /* handle on posted proc_waiting timeout */
508 static struct thread *filesys_syncer; /* proc of filesystem syncer process */
509 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
510 #define FLUSH_INODES		1
511 static int req_clear_remove;	/* syncer process flush some freeblks */
512 #define FLUSH_REMOVE		2
513 #define FLUSH_REMOVE_WAIT	3
514 /*
515  * runtime statistics
516  */
517 static int stat_worklist_push;	/* number of worklist cleanups */
518 static int stat_blk_limit_push;	/* number of times block limit neared */
519 static int stat_ino_limit_push;	/* number of times inode limit neared */
520 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
521 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
522 static int stat_sync_limit_hit;	/* number of synchronous slowdowns imposed */
523 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
524 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
525 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
526 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
527 #ifdef DEBUG
528 #include <vm/vm.h>
529 #include <sys/sysctl.h>
530 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
531 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
532 SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 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 
557 	if (wk->wk_state & ONWORKLIST) {
558 		if (lk.lkt_held != NOHOLDER)
559 			FREE_LOCK(&lk);
560 		panic("add_to_worklist: already on list");
561 	}
562 	wk->wk_state |= ONWORKLIST;
563 	if (LIST_FIRST(&softdep_workitem_pending) == NULL)
564 		LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
565 	else
566 		LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
567 	worklist_tail = wk;
568 	num_on_worklist += 1;
569 }
570 
571 /*
572  * Process that runs once per second to handle items in the background queue.
573  *
574  * Note that we ensure that everything is done in the order in which they
575  * appear in the queue. The code below depends on this property to ensure
576  * that blocks of a file are freed before the inode itself is freed. This
577  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
578  * until all the old ones have been purged from the dependency lists.
579  */
580 int
581 softdep_process_worklist(matchmnt)
582 	struct mount *matchmnt;
583 {
584 	struct thread *td = curthread;
585 	int cnt, matchcnt, loopcount;
586 	long starttime;
587 
588 	/*
589 	 * Record the process identifier of our caller so that we can give
590 	 * this process preferential treatment in request_cleanup below.
591 	 */
592 	filesys_syncer = td;
593 	matchcnt = 0;
594 
595 	/*
596 	 * There is no danger of having multiple processes run this
597 	 * code, but we have to single-thread it when softdep_flushfiles()
598 	 * is in operation to get an accurate count of the number of items
599 	 * related to its mount point that are in the list.
600 	 */
601 	if (matchmnt == NULL) {
602 		if (softdep_worklist_busy < 0)
603 			return(-1);
604 		softdep_worklist_busy += 1;
605 	}
606 
607 	/*
608 	 * If requested, try removing inode or removal dependencies.
609 	 */
610 	if (req_clear_inodedeps) {
611 		clear_inodedeps(td);
612 		req_clear_inodedeps -= 1;
613 		wakeup_one(&proc_waiting);
614 	}
615 	if (req_clear_remove) {
616 		clear_remove(td);
617 		req_clear_remove -= 1;
618 		wakeup_one(&proc_waiting);
619 	}
620 	loopcount = 1;
621 	starttime = time_second;
622 	while (num_on_worklist > 0) {
623 		if ((cnt = process_worklist_item(matchmnt, 0)) == -1)
624 			break;
625 		else
626 			matchcnt += cnt;
627 
628 		/*
629 		 * If a umount operation wants to run the worklist
630 		 * accurately, abort.
631 		 */
632 		if (softdep_worklist_req && matchmnt == NULL) {
633 			matchcnt = -1;
634 			break;
635 		}
636 
637 		/*
638 		 * If requested, try removing inode or removal dependencies.
639 		 */
640 		if (req_clear_inodedeps) {
641 			clear_inodedeps(td);
642 			req_clear_inodedeps -= 1;
643 			wakeup_one(&proc_waiting);
644 		}
645 		if (req_clear_remove) {
646 			clear_remove(td);
647 			req_clear_remove -= 1;
648 			wakeup_one(&proc_waiting);
649 		}
650 		/*
651 		 * We do not generally want to stop for buffer space, but if
652 		 * we are really being a buffer hog, we will stop and wait.
653 		 */
654 		if (loopcount++ % 128 == 0)
655 			bwillwrite();
656 		/*
657 		 * Never allow processing to run for more than one
658 		 * second. Otherwise the other syncer tasks may get
659 		 * excessively backlogged.
660 		 */
661 		if (starttime != time_second && matchmnt == NULL) {
662 			matchcnt = -1;
663 			break;
664 		}
665 	}
666 	if (matchmnt == NULL) {
667 		softdep_worklist_busy -= 1;
668 		if (softdep_worklist_req && softdep_worklist_busy == 0)
669 			wakeup(&softdep_worklist_req);
670 	}
671 	return (matchcnt);
672 }
673 
674 /*
675  * Process one item on the worklist.
676  */
677 static int
678 process_worklist_item(matchmnt, flags)
679 	struct mount *matchmnt;
680 	int flags;
681 {
682 	struct worklist *wk, *wkend;
683 	struct mount *mp;
684 	struct vnode *vp;
685 	int matchcnt = 0;
686 
687 	/*
688 	 * If we are being called because of a process doing a
689 	 * copy-on-write, then it is not safe to write as we may
690 	 * recurse into the copy-on-write routine.
691 	 */
692 	if (curthread->td_pflags & TDP_COWINPROGRESS)
693 		return (-1);
694 	ACQUIRE_LOCK(&lk);
695 	/*
696 	 * Normally we just process each item on the worklist in order.
697 	 * However, if we are in a situation where we cannot lock any
698 	 * inodes, we have to skip over any dirrem requests whose
699 	 * vnodes are resident and locked.
700 	 */
701 	vp = NULL;
702 	LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
703 		if (wk->wk_state & INPROGRESS)
704 			continue;
705 		if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
706 			break;
707 		wk->wk_state |= INPROGRESS;
708 		FREE_LOCK(&lk);
709 		VFS_VGET(WK_DIRREM(wk)->dm_mnt, WK_DIRREM(wk)->dm_oldinum,
710 		    LK_NOWAIT | LK_EXCLUSIVE, &vp);
711 		ACQUIRE_LOCK(&lk);
712 		wk->wk_state &= ~INPROGRESS;
713 		if (vp != NULL)
714 			break;
715 	}
716 	if (wk == 0) {
717 		FREE_LOCK(&lk);
718 		return (-1);
719 	}
720 	/*
721 	 * Remove the item to be processed. If we are removing the last
722 	 * item on the list, we need to recalculate the tail pointer.
723 	 * As this happens rarely and usually when the list is short,
724 	 * we just run down the list to find it rather than tracking it
725 	 * in the above loop.
726 	 */
727 	WORKLIST_REMOVE(wk);
728 	if (wk == worklist_tail) {
729 		LIST_FOREACH(wkend, &softdep_workitem_pending, wk_list)
730 			if (LIST_NEXT(wkend, wk_list) == NULL)
731 				break;
732 		worklist_tail = wkend;
733 	}
734 	num_on_worklist -= 1;
735 	FREE_LOCK(&lk);
736 	switch (wk->wk_type) {
737 
738 	case D_DIRREM:
739 		/* removal of a directory entry */
740 		mp = WK_DIRREM(wk)->dm_mnt;
741 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
742 			panic("%s: dirrem on suspended filesystem",
743 				"process_worklist_item");
744 		if (mp == matchmnt)
745 			matchcnt += 1;
746 		handle_workitem_remove(WK_DIRREM(wk), vp);
747 		break;
748 
749 	case D_FREEBLKS:
750 		/* releasing blocks and/or fragments from a file */
751 		mp = WK_FREEBLKS(wk)->fb_mnt;
752 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
753 			panic("%s: freeblks on suspended filesystem",
754 				"process_worklist_item");
755 		if (mp == matchmnt)
756 			matchcnt += 1;
757 		handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT);
758 		break;
759 
760 	case D_FREEFRAG:
761 		/* releasing a fragment when replaced as a file grows */
762 		mp = WK_FREEFRAG(wk)->ff_mnt;
763 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
764 			panic("%s: freefrag on suspended filesystem",
765 				"process_worklist_item");
766 		if (mp == matchmnt)
767 			matchcnt += 1;
768 		handle_workitem_freefrag(WK_FREEFRAG(wk));
769 		break;
770 
771 	case D_FREEFILE:
772 		/* releasing an inode when its link count drops to 0 */
773 		mp = WK_FREEFILE(wk)->fx_mnt;
774 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
775 			panic("%s: freefile on suspended filesystem",
776 				"process_worklist_item");
777 		if (mp == matchmnt)
778 			matchcnt += 1;
779 		handle_workitem_freefile(WK_FREEFILE(wk));
780 		break;
781 
782 	default:
783 		panic("%s_process_worklist: Unknown type %s",
784 		    "softdep", TYPENAME(wk->wk_type));
785 		/* NOTREACHED */
786 	}
787 	return (matchcnt);
788 }
789 
790 /*
791  * Move dependencies from one buffer to another.
792  */
793 static void
794 softdep_move_dependencies(oldbp, newbp)
795 	struct buf *oldbp;
796 	struct buf *newbp;
797 {
798 	struct worklist *wk, *wktail;
799 
800 	if (LIST_FIRST(&newbp->b_dep) != NULL)
801 		panic("softdep_move_dependencies: need merge code");
802 	wktail = 0;
803 	ACQUIRE_LOCK(&lk);
804 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
805 		LIST_REMOVE(wk, wk_list);
806 		if (wktail == 0)
807 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
808 		else
809 			LIST_INSERT_AFTER(wktail, wk, wk_list);
810 		wktail = wk;
811 	}
812 	FREE_LOCK(&lk);
813 }
814 
815 /*
816  * Purge the work list of all items associated with a particular mount point.
817  */
818 int
819 softdep_flushworklist(oldmnt, countp, td)
820 	struct mount *oldmnt;
821 	int *countp;
822 	struct thread *td;
823 {
824 	struct vnode *devvp;
825 	int count, error = 0;
826 
827 	/*
828 	 * Await our turn to clear out the queue, then serialize access.
829 	 */
830 	while (softdep_worklist_busy) {
831 		softdep_worklist_req += 1;
832 		tsleep(&softdep_worklist_req, PRIBIO, "softflush", 0);
833 		softdep_worklist_req -= 1;
834 	}
835 	softdep_worklist_busy = -1;
836 	/*
837 	 * Alternately flush the block device associated with the mount
838 	 * point and process any dependencies that the flushing
839 	 * creates. We continue until no more worklist dependencies
840 	 * are found.
841 	 */
842 	*countp = 0;
843 	devvp = VFSTOUFS(oldmnt)->um_devvp;
844 	while ((count = softdep_process_worklist(oldmnt)) > 0) {
845 		*countp += count;
846 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td);
847 		error = VOP_FSYNC(devvp, td->td_ucred, MNT_WAIT, td);
848 		VOP_UNLOCK(devvp, 0, td);
849 		if (error)
850 			break;
851 	}
852 	softdep_worklist_busy = 0;
853 	if (softdep_worklist_req)
854 		wakeup(&softdep_worklist_req);
855 	return (error);
856 }
857 
858 /*
859  * Flush all vnodes and worklist items associated with a specified mount point.
860  */
861 int
862 softdep_flushfiles(oldmnt, flags, td)
863 	struct mount *oldmnt;
864 	int flags;
865 	struct thread *td;
866 {
867 	int error, count, loopcnt;
868 
869 	error = 0;
870 
871 	/*
872 	 * Alternately flush the vnodes associated with the mount
873 	 * point and process any dependencies that the flushing
874 	 * creates. In theory, this loop can happen at most twice,
875 	 * but we give it a few extra just to be sure.
876 	 */
877 	for (loopcnt = 10; loopcnt > 0; loopcnt--) {
878 		/*
879 		 * Do another flush in case any vnodes were brought in
880 		 * as part of the cleanup operations.
881 		 */
882 		if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
883 			break;
884 		if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 ||
885 		    count == 0)
886 			break;
887 	}
888 	/*
889 	 * If we are unmounting then it is an error to fail. If we
890 	 * are simply trying to downgrade to read-only, then filesystem
891 	 * activity can keep us busy forever, so we just fail with EBUSY.
892 	 */
893 	if (loopcnt == 0) {
894 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
895 			panic("softdep_flushfiles: looping");
896 		error = EBUSY;
897 	}
898 	return (error);
899 }
900 
901 /*
902  * Structure hashing.
903  *
904  * There are three types of structures that can be looked up:
905  *	1) pagedep structures identified by mount point, inode number,
906  *	   and logical block.
907  *	2) inodedep structures identified by mount point and inode number.
908  *	3) newblk structures identified by mount point and
909  *	   physical block number.
910  *
911  * The "pagedep" and "inodedep" dependency structures are hashed
912  * separately from the file blocks and inodes to which they correspond.
913  * This separation helps when the in-memory copy of an inode or
914  * file block must be replaced. It also obviates the need to access
915  * an inode or file page when simply updating (or de-allocating)
916  * dependency structures. Lookup of newblk structures is needed to
917  * find newly allocated blocks when trying to associate them with
918  * their allocdirect or allocindir structure.
919  *
920  * The lookup routines optionally create and hash a new instance when
921  * an existing entry is not found.
922  */
923 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
924 #define NODELAY		0x0002	/* cannot do background work */
925 
926 /*
927  * Structures and routines associated with pagedep caching.
928  */
929 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
930 u_long	pagedep_hash;		/* size of hash table - 1 */
931 #define	PAGEDEP_HASH(mp, inum, lbn) \
932 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
933 	    pagedep_hash])
934 static struct sema pagedep_in_progress;
935 
936 /*
937  * Look up a pagedep. Return 1 if found, 0 if not found or found
938  * when asked to allocate but not associated with any buffer.
939  * If not found, allocate if DEPALLOC flag is passed.
940  * Found or allocated entry is returned in pagedeppp.
941  * This routine must be called with splbio interrupts blocked.
942  */
943 static int
944 pagedep_lookup(ip, lbn, flags, pagedeppp)
945 	struct inode *ip;
946 	ufs_lbn_t lbn;
947 	int flags;
948 	struct pagedep **pagedeppp;
949 {
950 	struct pagedep *pagedep;
951 	struct pagedep_hashhead *pagedephd;
952 	struct mount *mp;
953 	int i;
954 
955 #ifdef DEBUG
956 	if (lk.lkt_held == NOHOLDER)
957 		panic("pagedep_lookup: lock not held");
958 #endif
959 	mp = ITOV(ip)->v_mount;
960 	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
961 top:
962 	LIST_FOREACH(pagedep, pagedephd, pd_hash)
963 		if (ip->i_number == pagedep->pd_ino &&
964 		    lbn == pagedep->pd_lbn &&
965 		    mp == pagedep->pd_mnt)
966 			break;
967 	if (pagedep) {
968 		*pagedeppp = pagedep;
969 		if ((flags & DEPALLOC) != 0 &&
970 		    (pagedep->pd_state & ONWORKLIST) == 0)
971 			return (0);
972 		return (1);
973 	}
974 	if ((flags & DEPALLOC) == 0) {
975 		*pagedeppp = NULL;
976 		return (0);
977 	}
978 	if (sema_get(&pagedep_in_progress, &lk) == 0) {
979 		ACQUIRE_LOCK(&lk);
980 		goto top;
981 	}
982 	MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
983 		M_SOFTDEP_FLAGS|M_ZERO);
984 	pagedep->pd_list.wk_type = D_PAGEDEP;
985 	pagedep->pd_mnt = mp;
986 	pagedep->pd_ino = ip->i_number;
987 	pagedep->pd_lbn = lbn;
988 	LIST_INIT(&pagedep->pd_dirremhd);
989 	LIST_INIT(&pagedep->pd_pendinghd);
990 	for (i = 0; i < DAHASHSZ; i++)
991 		LIST_INIT(&pagedep->pd_diraddhd[i]);
992 	ACQUIRE_LOCK(&lk);
993 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
994 	sema_release(&pagedep_in_progress);
995 	*pagedeppp = pagedep;
996 	return (0);
997 }
998 
999 /*
1000  * Structures and routines associated with inodedep caching.
1001  */
1002 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
1003 static u_long	inodedep_hash;	/* size of hash table - 1 */
1004 static long	num_inodedep;	/* number of inodedep allocated */
1005 #define	INODEDEP_HASH(fs, inum) \
1006       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
1007 static struct sema inodedep_in_progress;
1008 
1009 /*
1010  * Look up an inodedep. Return 1 if found, 0 if not found.
1011  * If not found, allocate if DEPALLOC flag is passed.
1012  * Found or allocated entry is returned in inodedeppp.
1013  * This routine must be called with splbio interrupts blocked.
1014  */
1015 static int
1016 inodedep_lookup(fs, inum, flags, inodedeppp)
1017 	struct fs *fs;
1018 	ino_t inum;
1019 	int flags;
1020 	struct inodedep **inodedeppp;
1021 {
1022 	struct inodedep *inodedep;
1023 	struct inodedep_hashhead *inodedephd;
1024 	int firsttry;
1025 
1026 #ifdef DEBUG
1027 	if (lk.lkt_held == NOHOLDER)
1028 		panic("inodedep_lookup: lock not held");
1029 #endif
1030 	firsttry = 1;
1031 	inodedephd = INODEDEP_HASH(fs, inum);
1032 top:
1033 	LIST_FOREACH(inodedep, inodedephd, id_hash)
1034 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
1035 			break;
1036 	if (inodedep) {
1037 		*inodedeppp = inodedep;
1038 		return (1);
1039 	}
1040 	if ((flags & DEPALLOC) == 0) {
1041 		*inodedeppp = NULL;
1042 		return (0);
1043 	}
1044 	/*
1045 	 * If we are over our limit, try to improve the situation.
1046 	 */
1047 	if (num_inodedep > max_softdeps && firsttry && (flags & NODELAY) == 0 &&
1048 	    request_cleanup(FLUSH_INODES, 1)) {
1049 		firsttry = 0;
1050 		goto top;
1051 	}
1052 	if (sema_get(&inodedep_in_progress, &lk) == 0) {
1053 		ACQUIRE_LOCK(&lk);
1054 		goto top;
1055 	}
1056 	num_inodedep += 1;
1057 	MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1058 		M_INODEDEP, M_SOFTDEP_FLAGS);
1059 	inodedep->id_list.wk_type = D_INODEDEP;
1060 	inodedep->id_fs = fs;
1061 	inodedep->id_ino = inum;
1062 	inodedep->id_state = ALLCOMPLETE;
1063 	inodedep->id_nlinkdelta = 0;
1064 	inodedep->id_savedino1 = NULL;
1065 	inodedep->id_savedsize = -1;
1066 	inodedep->id_savedextsize = -1;
1067 	inodedep->id_buf = NULL;
1068 	LIST_INIT(&inodedep->id_pendinghd);
1069 	LIST_INIT(&inodedep->id_inowait);
1070 	LIST_INIT(&inodedep->id_bufwait);
1071 	TAILQ_INIT(&inodedep->id_inoupdt);
1072 	TAILQ_INIT(&inodedep->id_newinoupdt);
1073 	TAILQ_INIT(&inodedep->id_extupdt);
1074 	TAILQ_INIT(&inodedep->id_newextupdt);
1075 	ACQUIRE_LOCK(&lk);
1076 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1077 	sema_release(&inodedep_in_progress);
1078 	*inodedeppp = inodedep;
1079 	return (0);
1080 }
1081 
1082 /*
1083  * Structures and routines associated with newblk caching.
1084  */
1085 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1086 u_long	newblk_hash;		/* size of hash table - 1 */
1087 #define	NEWBLK_HASH(fs, inum) \
1088 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1089 static struct sema newblk_in_progress;
1090 
1091 /*
1092  * Look up a newblk. Return 1 if found, 0 if not found.
1093  * If not found, allocate if DEPALLOC flag is passed.
1094  * Found or allocated entry is returned in newblkpp.
1095  */
1096 static int
1097 newblk_lookup(fs, newblkno, flags, newblkpp)
1098 	struct fs *fs;
1099 	ufs2_daddr_t newblkno;
1100 	int flags;
1101 	struct newblk **newblkpp;
1102 {
1103 	struct newblk *newblk;
1104 	struct newblk_hashhead *newblkhd;
1105 
1106 	newblkhd = NEWBLK_HASH(fs, newblkno);
1107 top:
1108 	LIST_FOREACH(newblk, newblkhd, nb_hash)
1109 		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1110 			break;
1111 	if (newblk) {
1112 		*newblkpp = newblk;
1113 		return (1);
1114 	}
1115 	if ((flags & DEPALLOC) == 0) {
1116 		*newblkpp = NULL;
1117 		return (0);
1118 	}
1119 	if (sema_get(&newblk_in_progress, 0) == 0)
1120 		goto top;
1121 	MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1122 		M_NEWBLK, M_SOFTDEP_FLAGS);
1123 	newblk->nb_state = 0;
1124 	newblk->nb_fs = fs;
1125 	newblk->nb_newblkno = newblkno;
1126 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1127 	sema_release(&newblk_in_progress);
1128 	*newblkpp = newblk;
1129 	return (0);
1130 }
1131 
1132 /*
1133  * Executed during filesystem system initialization before
1134  * mounting any filesystems.
1135  */
1136 void
1137 softdep_initialize()
1138 {
1139 
1140 	LIST_INIT(&mkdirlisthd);
1141 	LIST_INIT(&softdep_workitem_pending);
1142 	max_softdeps = desiredvnodes * 4;
1143 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1144 	    &pagedep_hash);
1145 	sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0);
1146 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1147 	sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0);
1148 	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1149 	sema_init(&newblk_in_progress, "newblk", PRIBIO, 0);
1150 
1151 	/* hooks through which the main kernel code calls us */
1152 	softdep_process_worklist_hook = softdep_process_worklist;
1153 	softdep_fsync_hook = softdep_fsync;
1154 
1155 	/* initialise bioops hack */
1156 	bioops.io_prewrite = softdep_disk_prewrite;
1157 	bioops.io_start = softdep_disk_io_initiation;
1158 	bioops.io_complete = softdep_disk_write_complete;
1159 	bioops.io_deallocate = softdep_deallocate_dependencies;
1160 	bioops.io_movedeps = softdep_move_dependencies;
1161 	bioops.io_countdeps = softdep_count_dependencies;
1162 }
1163 
1164 /*
1165  * Executed after all filesystems have been unmounted during
1166  * filesystem module unload.
1167  */
1168 void
1169 softdep_uninitialize()
1170 {
1171 
1172 	softdep_process_worklist_hook = NULL;
1173 	softdep_fsync_hook = NULL;
1174 	hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash);
1175 	hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash);
1176 	hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash);
1177 }
1178 
1179 /*
1180  * Called at mount time to notify the dependency code that a
1181  * filesystem wishes to use it.
1182  */
1183 int
1184 softdep_mount(devvp, mp, fs, cred)
1185 	struct vnode *devvp;
1186 	struct mount *mp;
1187 	struct fs *fs;
1188 	struct ucred *cred;
1189 {
1190 	struct csum_total cstotal;
1191 	struct cg *cgp;
1192 	struct buf *bp;
1193 	int error, cyl;
1194 
1195 	mp->mnt_flag &= ~MNT_ASYNC;
1196 	mp->mnt_flag |= MNT_SOFTDEP;
1197 	/*
1198 	 * When doing soft updates, the counters in the
1199 	 * superblock may have gotten out of sync, so we have
1200 	 * to scan the cylinder groups and recalculate them.
1201 	 */
1202 	if (fs->fs_clean != 0)
1203 		return (0);
1204 	bzero(&cstotal, sizeof cstotal);
1205 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1206 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1207 		    fs->fs_cgsize, cred, &bp)) != 0) {
1208 			brelse(bp);
1209 			return (error);
1210 		}
1211 		cgp = (struct cg *)bp->b_data;
1212 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1213 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1214 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1215 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1216 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
1217 		brelse(bp);
1218 	}
1219 #ifdef DEBUG
1220 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1221 		printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
1222 #endif
1223 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1224 	return (0);
1225 }
1226 
1227 /*
1228  * Protecting the freemaps (or bitmaps).
1229  *
1230  * To eliminate the need to execute fsck before mounting a filesystem
1231  * after a power failure, one must (conservatively) guarantee that the
1232  * on-disk copy of the bitmaps never indicate that a live inode or block is
1233  * free.  So, when a block or inode is allocated, the bitmap should be
1234  * updated (on disk) before any new pointers.  When a block or inode is
1235  * freed, the bitmap should not be updated until all pointers have been
1236  * reset.  The latter dependency is handled by the delayed de-allocation
1237  * approach described below for block and inode de-allocation.  The former
1238  * dependency is handled by calling the following procedure when a block or
1239  * inode is allocated. When an inode is allocated an "inodedep" is created
1240  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1241  * Each "inodedep" is also inserted into the hash indexing structure so
1242  * that any additional link additions can be made dependent on the inode
1243  * allocation.
1244  *
1245  * The ufs filesystem maintains a number of free block counts (e.g., per
1246  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1247  * in addition to the bitmaps.  These counts are used to improve efficiency
1248  * during allocation and therefore must be consistent with the bitmaps.
1249  * There is no convenient way to guarantee post-crash consistency of these
1250  * counts with simple update ordering, for two main reasons: (1) The counts
1251  * and bitmaps for a single cylinder group block are not in the same disk
1252  * sector.  If a disk write is interrupted (e.g., by power failure), one may
1253  * be written and the other not.  (2) Some of the counts are located in the
1254  * superblock rather than the cylinder group block. So, we focus our soft
1255  * updates implementation on protecting the bitmaps. When mounting a
1256  * filesystem, we recompute the auxiliary counts from the bitmaps.
1257  */
1258 
1259 /*
1260  * Called just after updating the cylinder group block to allocate an inode.
1261  */
1262 void
1263 softdep_setup_inomapdep(bp, ip, newinum)
1264 	struct buf *bp;		/* buffer for cylgroup block with inode map */
1265 	struct inode *ip;	/* inode related to allocation */
1266 	ino_t newinum;		/* new inode number being allocated */
1267 {
1268 	struct inodedep *inodedep;
1269 	struct bmsafemap *bmsafemap;
1270 
1271 	/*
1272 	 * Create a dependency for the newly allocated inode.
1273 	 * Panic if it already exists as something is seriously wrong.
1274 	 * Otherwise add it to the dependency list for the buffer holding
1275 	 * the cylinder group map from which it was allocated.
1276 	 */
1277 	ACQUIRE_LOCK(&lk);
1278 	if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1279 		FREE_LOCK(&lk);
1280 		panic("softdep_setup_inomapdep: found inode");
1281 	}
1282 	inodedep->id_buf = bp;
1283 	inodedep->id_state &= ~DEPCOMPLETE;
1284 	bmsafemap = bmsafemap_lookup(bp);
1285 	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1286 	FREE_LOCK(&lk);
1287 }
1288 
1289 /*
1290  * Called just after updating the cylinder group block to
1291  * allocate block or fragment.
1292  */
1293 void
1294 softdep_setup_blkmapdep(bp, fs, newblkno)
1295 	struct buf *bp;		/* buffer for cylgroup block with block map */
1296 	struct fs *fs;		/* filesystem doing allocation */
1297 	ufs2_daddr_t newblkno;	/* number of newly allocated block */
1298 {
1299 	struct newblk *newblk;
1300 	struct bmsafemap *bmsafemap;
1301 
1302 	/*
1303 	 * Create a dependency for the newly allocated block.
1304 	 * Add it to the dependency list for the buffer holding
1305 	 * the cylinder group map from which it was allocated.
1306 	 */
1307 	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1308 		panic("softdep_setup_blkmapdep: found block");
1309 	ACQUIRE_LOCK(&lk);
1310 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1311 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1312 	FREE_LOCK(&lk);
1313 }
1314 
1315 /*
1316  * Find the bmsafemap associated with a cylinder group buffer.
1317  * If none exists, create one. The buffer must be locked when
1318  * this routine is called and this routine must be called with
1319  * splbio interrupts blocked.
1320  */
1321 static struct bmsafemap *
1322 bmsafemap_lookup(bp)
1323 	struct buf *bp;
1324 {
1325 	struct bmsafemap *bmsafemap;
1326 	struct worklist *wk;
1327 
1328 #ifdef DEBUG
1329 	if (lk.lkt_held == NOHOLDER)
1330 		panic("bmsafemap_lookup: lock not held");
1331 #endif
1332 	LIST_FOREACH(wk, &bp->b_dep, wk_list)
1333 		if (wk->wk_type == D_BMSAFEMAP)
1334 			return (WK_BMSAFEMAP(wk));
1335 	FREE_LOCK(&lk);
1336 	MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1337 		M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1338 	bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1339 	bmsafemap->sm_list.wk_state = 0;
1340 	bmsafemap->sm_buf = bp;
1341 	LIST_INIT(&bmsafemap->sm_allocdirecthd);
1342 	LIST_INIT(&bmsafemap->sm_allocindirhd);
1343 	LIST_INIT(&bmsafemap->sm_inodedephd);
1344 	LIST_INIT(&bmsafemap->sm_newblkhd);
1345 	ACQUIRE_LOCK(&lk);
1346 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1347 	return (bmsafemap);
1348 }
1349 
1350 /*
1351  * Direct block allocation dependencies.
1352  *
1353  * When a new block is allocated, the corresponding disk locations must be
1354  * initialized (with zeros or new data) before the on-disk inode points to
1355  * them.  Also, the freemap from which the block was allocated must be
1356  * updated (on disk) before the inode's pointer. These two dependencies are
1357  * independent of each other and are needed for all file blocks and indirect
1358  * blocks that are pointed to directly by the inode.  Just before the
1359  * "in-core" version of the inode is updated with a newly allocated block
1360  * number, a procedure (below) is called to setup allocation dependency
1361  * structures.  These structures are removed when the corresponding
1362  * dependencies are satisfied or when the block allocation becomes obsolete
1363  * (i.e., the file is deleted, the block is de-allocated, or the block is a
1364  * fragment that gets upgraded).  All of these cases are handled in
1365  * procedures described later.
1366  *
1367  * When a file extension causes a fragment to be upgraded, either to a larger
1368  * fragment or to a full block, the on-disk location may change (if the
1369  * previous fragment could not simply be extended). In this case, the old
1370  * fragment must be de-allocated, but not until after the inode's pointer has
1371  * been updated. In most cases, this is handled by later procedures, which
1372  * will construct a "freefrag" structure to be added to the workitem queue
1373  * when the inode update is complete (or obsolete).  The main exception to
1374  * this is when an allocation occurs while a pending allocation dependency
1375  * (for the same block pointer) remains.  This case is handled in the main
1376  * allocation dependency setup procedure by immediately freeing the
1377  * unreferenced fragments.
1378  */
1379 void
1380 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1381 	struct inode *ip;	/* inode to which block is being added */
1382 	ufs_lbn_t lbn;		/* block pointer within inode */
1383 	ufs2_daddr_t newblkno;	/* disk block number being added */
1384 	ufs2_daddr_t oldblkno;	/* previous block number, 0 unless frag */
1385 	long newsize;		/* size of new block */
1386 	long oldsize;		/* size of new block */
1387 	struct buf *bp;		/* bp for allocated block */
1388 {
1389 	struct allocdirect *adp, *oldadp;
1390 	struct allocdirectlst *adphead;
1391 	struct bmsafemap *bmsafemap;
1392 	struct inodedep *inodedep;
1393 	struct pagedep *pagedep;
1394 	struct newblk *newblk;
1395 
1396 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1397 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1398 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1399 	adp->ad_lbn = lbn;
1400 	adp->ad_newblkno = newblkno;
1401 	adp->ad_oldblkno = oldblkno;
1402 	adp->ad_newsize = newsize;
1403 	adp->ad_oldsize = oldsize;
1404 	adp->ad_state = ATTACHED;
1405 	LIST_INIT(&adp->ad_newdirblk);
1406 	if (newblkno == oldblkno)
1407 		adp->ad_freefrag = NULL;
1408 	else
1409 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1410 
1411 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1412 		panic("softdep_setup_allocdirect: lost block");
1413 
1414 	ACQUIRE_LOCK(&lk);
1415 	inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1416 	adp->ad_inodedep = inodedep;
1417 
1418 	if (newblk->nb_state == DEPCOMPLETE) {
1419 		adp->ad_state |= DEPCOMPLETE;
1420 		adp->ad_buf = NULL;
1421 	} else {
1422 		bmsafemap = newblk->nb_bmsafemap;
1423 		adp->ad_buf = bmsafemap->sm_buf;
1424 		LIST_REMOVE(newblk, nb_deps);
1425 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1426 	}
1427 	LIST_REMOVE(newblk, nb_hash);
1428 	FREE(newblk, M_NEWBLK);
1429 
1430 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1431 	if (lbn >= NDADDR) {
1432 		/* allocating an indirect block */
1433 		if (oldblkno != 0) {
1434 			FREE_LOCK(&lk);
1435 			panic("softdep_setup_allocdirect: non-zero indir");
1436 		}
1437 	} else {
1438 		/*
1439 		 * Allocating a direct block.
1440 		 *
1441 		 * If we are allocating a directory block, then we must
1442 		 * allocate an associated pagedep to track additions and
1443 		 * deletions.
1444 		 */
1445 		if ((ip->i_mode & IFMT) == IFDIR &&
1446 		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1447 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1448 	}
1449 	/*
1450 	 * The list of allocdirects must be kept in sorted and ascending
1451 	 * order so that the rollback routines can quickly determine the
1452 	 * first uncommitted block (the size of the file stored on disk
1453 	 * ends at the end of the lowest committed fragment, or if there
1454 	 * are no fragments, at the end of the highest committed block).
1455 	 * Since files generally grow, the typical case is that the new
1456 	 * block is to be added at the end of the list. We speed this
1457 	 * special case by checking against the last allocdirect in the
1458 	 * list before laboriously traversing the list looking for the
1459 	 * insertion point.
1460 	 */
1461 	adphead = &inodedep->id_newinoupdt;
1462 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1463 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1464 		/* insert at end of list */
1465 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1466 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1467 			allocdirect_merge(adphead, adp, oldadp);
1468 		FREE_LOCK(&lk);
1469 		return;
1470 	}
1471 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1472 		if (oldadp->ad_lbn >= lbn)
1473 			break;
1474 	}
1475 	if (oldadp == NULL) {
1476 		FREE_LOCK(&lk);
1477 		panic("softdep_setup_allocdirect: lost entry");
1478 	}
1479 	/* insert in middle of list */
1480 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1481 	if (oldadp->ad_lbn == lbn)
1482 		allocdirect_merge(adphead, adp, oldadp);
1483 	FREE_LOCK(&lk);
1484 }
1485 
1486 /*
1487  * Replace an old allocdirect dependency with a newer one.
1488  * This routine must be called with splbio interrupts blocked.
1489  */
1490 static void
1491 allocdirect_merge(adphead, newadp, oldadp)
1492 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
1493 	struct allocdirect *newadp;	/* allocdirect being added */
1494 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
1495 {
1496 	struct worklist *wk;
1497 	struct freefrag *freefrag;
1498 	struct newdirblk *newdirblk;
1499 
1500 #ifdef DEBUG
1501 	if (lk.lkt_held == NOHOLDER)
1502 		panic("allocdirect_merge: lock not held");
1503 #endif
1504 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1505 	    newadp->ad_oldsize != oldadp->ad_newsize ||
1506 	    newadp->ad_lbn >= NDADDR) {
1507 		FREE_LOCK(&lk);
1508 		panic("%s %jd != new %jd || old size %ld != new %ld",
1509 		    "allocdirect_merge: old blkno",
1510 		    (intmax_t)newadp->ad_oldblkno,
1511 		    (intmax_t)oldadp->ad_newblkno,
1512 		    newadp->ad_oldsize, oldadp->ad_newsize);
1513 	}
1514 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
1515 	newadp->ad_oldsize = oldadp->ad_oldsize;
1516 	/*
1517 	 * If the old dependency had a fragment to free or had never
1518 	 * previously had a block allocated, then the new dependency
1519 	 * can immediately post its freefrag and adopt the old freefrag.
1520 	 * This action is done by swapping the freefrag dependencies.
1521 	 * The new dependency gains the old one's freefrag, and the
1522 	 * old one gets the new one and then immediately puts it on
1523 	 * the worklist when it is freed by free_allocdirect. It is
1524 	 * not possible to do this swap when the old dependency had a
1525 	 * non-zero size but no previous fragment to free. This condition
1526 	 * arises when the new block is an extension of the old block.
1527 	 * Here, the first part of the fragment allocated to the new
1528 	 * dependency is part of the block currently claimed on disk by
1529 	 * the old dependency, so cannot legitimately be freed until the
1530 	 * conditions for the new dependency are fulfilled.
1531 	 */
1532 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1533 		freefrag = newadp->ad_freefrag;
1534 		newadp->ad_freefrag = oldadp->ad_freefrag;
1535 		oldadp->ad_freefrag = freefrag;
1536 	}
1537 	/*
1538 	 * If we are tracking a new directory-block allocation,
1539 	 * move it from the old allocdirect to the new allocdirect.
1540 	 */
1541 	if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
1542 		newdirblk = WK_NEWDIRBLK(wk);
1543 		WORKLIST_REMOVE(&newdirblk->db_list);
1544 		if (LIST_FIRST(&oldadp->ad_newdirblk) != NULL)
1545 			panic("allocdirect_merge: extra newdirblk");
1546 		WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list);
1547 	}
1548 	free_allocdirect(adphead, oldadp, 0);
1549 }
1550 
1551 /*
1552  * Allocate a new freefrag structure if needed.
1553  */
1554 static struct freefrag *
1555 newfreefrag(ip, blkno, size)
1556 	struct inode *ip;
1557 	ufs2_daddr_t blkno;
1558 	long size;
1559 {
1560 	struct freefrag *freefrag;
1561 	struct fs *fs;
1562 
1563 	if (blkno == 0)
1564 		return (NULL);
1565 	fs = ip->i_fs;
1566 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1567 		panic("newfreefrag: frag size");
1568 	MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1569 		M_FREEFRAG, M_SOFTDEP_FLAGS);
1570 	freefrag->ff_list.wk_type = D_FREEFRAG;
1571 	freefrag->ff_state = 0;
1572 	freefrag->ff_inum = ip->i_number;
1573 	freefrag->ff_mnt = ITOV(ip)->v_mount;
1574 	freefrag->ff_blkno = blkno;
1575 	freefrag->ff_fragsize = size;
1576 	return (freefrag);
1577 }
1578 
1579 /*
1580  * This workitem de-allocates fragments that were replaced during
1581  * file block allocation.
1582  */
1583 static void
1584 handle_workitem_freefrag(freefrag)
1585 	struct freefrag *freefrag;
1586 {
1587 	struct ufsmount *ump = VFSTOUFS(freefrag->ff_mnt);
1588 
1589 	ffs_blkfree(ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
1590 	    freefrag->ff_fragsize, freefrag->ff_inum);
1591 	FREE(freefrag, M_FREEFRAG);
1592 }
1593 
1594 /*
1595  * Set up a dependency structure for an external attributes data block.
1596  * This routine follows much of the structure of softdep_setup_allocdirect.
1597  * See the description of softdep_setup_allocdirect above for details.
1598  */
1599 void
1600 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1601 	struct inode *ip;
1602 	ufs_lbn_t lbn;
1603 	ufs2_daddr_t newblkno;
1604 	ufs2_daddr_t oldblkno;
1605 	long newsize;
1606 	long oldsize;
1607 	struct buf *bp;
1608 {
1609 	struct allocdirect *adp, *oldadp;
1610 	struct allocdirectlst *adphead;
1611 	struct bmsafemap *bmsafemap;
1612 	struct inodedep *inodedep;
1613 	struct newblk *newblk;
1614 
1615 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1616 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1617 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1618 	adp->ad_lbn = lbn;
1619 	adp->ad_newblkno = newblkno;
1620 	adp->ad_oldblkno = oldblkno;
1621 	adp->ad_newsize = newsize;
1622 	adp->ad_oldsize = oldsize;
1623 	adp->ad_state = ATTACHED | EXTDATA;
1624 	LIST_INIT(&adp->ad_newdirblk);
1625 	if (newblkno == oldblkno)
1626 		adp->ad_freefrag = NULL;
1627 	else
1628 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1629 
1630 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1631 		panic("softdep_setup_allocext: lost block");
1632 
1633 	ACQUIRE_LOCK(&lk);
1634 	inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1635 	adp->ad_inodedep = inodedep;
1636 
1637 	if (newblk->nb_state == DEPCOMPLETE) {
1638 		adp->ad_state |= DEPCOMPLETE;
1639 		adp->ad_buf = NULL;
1640 	} else {
1641 		bmsafemap = newblk->nb_bmsafemap;
1642 		adp->ad_buf = bmsafemap->sm_buf;
1643 		LIST_REMOVE(newblk, nb_deps);
1644 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1645 	}
1646 	LIST_REMOVE(newblk, nb_hash);
1647 	FREE(newblk, M_NEWBLK);
1648 
1649 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1650 	if (lbn >= NXADDR) {
1651 		FREE_LOCK(&lk);
1652 		panic("softdep_setup_allocext: lbn %lld > NXADDR",
1653 		    (long long)lbn);
1654 	}
1655 	/*
1656 	 * The list of allocdirects must be kept in sorted and ascending
1657 	 * order so that the rollback routines can quickly determine the
1658 	 * first uncommitted block (the size of the file stored on disk
1659 	 * ends at the end of the lowest committed fragment, or if there
1660 	 * are no fragments, at the end of the highest committed block).
1661 	 * Since files generally grow, the typical case is that the new
1662 	 * block is to be added at the end of the list. We speed this
1663 	 * special case by checking against the last allocdirect in the
1664 	 * list before laboriously traversing the list looking for the
1665 	 * insertion point.
1666 	 */
1667 	adphead = &inodedep->id_newextupdt;
1668 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1669 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1670 		/* insert at end of list */
1671 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1672 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1673 			allocdirect_merge(adphead, adp, oldadp);
1674 		FREE_LOCK(&lk);
1675 		return;
1676 	}
1677 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1678 		if (oldadp->ad_lbn >= lbn)
1679 			break;
1680 	}
1681 	if (oldadp == NULL) {
1682 		FREE_LOCK(&lk);
1683 		panic("softdep_setup_allocext: lost entry");
1684 	}
1685 	/* insert in middle of list */
1686 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1687 	if (oldadp->ad_lbn == lbn)
1688 		allocdirect_merge(adphead, adp, oldadp);
1689 	FREE_LOCK(&lk);
1690 }
1691 
1692 /*
1693  * Indirect block allocation dependencies.
1694  *
1695  * The same dependencies that exist for a direct block also exist when
1696  * a new block is allocated and pointed to by an entry in a block of
1697  * indirect pointers. The undo/redo states described above are also
1698  * used here. Because an indirect block contains many pointers that
1699  * may have dependencies, a second copy of the entire in-memory indirect
1700  * block is kept. The buffer cache copy is always completely up-to-date.
1701  * The second copy, which is used only as a source for disk writes,
1702  * contains only the safe pointers (i.e., those that have no remaining
1703  * update dependencies). The second copy is freed when all pointers
1704  * are safe. The cache is not allowed to replace indirect blocks with
1705  * pending update dependencies. If a buffer containing an indirect
1706  * block with dependencies is written, these routines will mark it
1707  * dirty again. It can only be successfully written once all the
1708  * dependencies are removed. The ffs_fsync routine in conjunction with
1709  * softdep_sync_metadata work together to get all the dependencies
1710  * removed so that a file can be successfully written to disk. Three
1711  * procedures are used when setting up indirect block pointer
1712  * dependencies. The division is necessary because of the organization
1713  * of the "balloc" routine and because of the distinction between file
1714  * pages and file metadata blocks.
1715  */
1716 
1717 /*
1718  * Allocate a new allocindir structure.
1719  */
1720 static struct allocindir *
1721 newallocindir(ip, ptrno, newblkno, oldblkno)
1722 	struct inode *ip;	/* inode for file being extended */
1723 	int ptrno;		/* offset of pointer in indirect block */
1724 	ufs2_daddr_t newblkno;	/* disk block number being added */
1725 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1726 {
1727 	struct allocindir *aip;
1728 
1729 	MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1730 		M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO);
1731 	aip->ai_list.wk_type = D_ALLOCINDIR;
1732 	aip->ai_state = ATTACHED;
1733 	aip->ai_offset = ptrno;
1734 	aip->ai_newblkno = newblkno;
1735 	aip->ai_oldblkno = oldblkno;
1736 	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1737 	return (aip);
1738 }
1739 
1740 /*
1741  * Called just before setting an indirect block pointer
1742  * to a newly allocated file page.
1743  */
1744 void
1745 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1746 	struct inode *ip;	/* inode for file being extended */
1747 	ufs_lbn_t lbn;		/* allocated block number within file */
1748 	struct buf *bp;		/* buffer with indirect blk referencing page */
1749 	int ptrno;		/* offset of pointer in indirect block */
1750 	ufs2_daddr_t newblkno;	/* disk block number being added */
1751 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1752 	struct buf *nbp;	/* buffer holding allocated page */
1753 {
1754 	struct allocindir *aip;
1755 	struct pagedep *pagedep;
1756 
1757 	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1758 	ACQUIRE_LOCK(&lk);
1759 	/*
1760 	 * If we are allocating a directory page, then we must
1761 	 * allocate an associated pagedep to track additions and
1762 	 * deletions.
1763 	 */
1764 	if ((ip->i_mode & IFMT) == IFDIR &&
1765 	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1766 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1767 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1768 	FREE_LOCK(&lk);
1769 	setup_allocindir_phase2(bp, ip, aip);
1770 }
1771 
1772 /*
1773  * Called just before setting an indirect block pointer to a
1774  * newly allocated indirect block.
1775  */
1776 void
1777 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1778 	struct buf *nbp;	/* newly allocated indirect block */
1779 	struct inode *ip;	/* inode for file being extended */
1780 	struct buf *bp;		/* indirect block referencing allocated block */
1781 	int ptrno;		/* offset of pointer in indirect block */
1782 	ufs2_daddr_t newblkno;	/* disk block number being added */
1783 {
1784 	struct allocindir *aip;
1785 
1786 	aip = newallocindir(ip, ptrno, newblkno, 0);
1787 	ACQUIRE_LOCK(&lk);
1788 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1789 	FREE_LOCK(&lk);
1790 	setup_allocindir_phase2(bp, ip, aip);
1791 }
1792 
1793 /*
1794  * Called to finish the allocation of the "aip" allocated
1795  * by one of the two routines above.
1796  */
1797 static void
1798 setup_allocindir_phase2(bp, ip, aip)
1799 	struct buf *bp;		/* in-memory copy of the indirect block */
1800 	struct inode *ip;	/* inode for file being extended */
1801 	struct allocindir *aip;	/* allocindir allocated by the above routines */
1802 {
1803 	struct worklist *wk;
1804 	struct indirdep *indirdep, *newindirdep;
1805 	struct bmsafemap *bmsafemap;
1806 	struct allocindir *oldaip;
1807 	struct freefrag *freefrag;
1808 	struct newblk *newblk;
1809 	ufs2_daddr_t blkno;
1810 
1811 	if (bp->b_lblkno >= 0)
1812 		panic("setup_allocindir_phase2: not indir blk");
1813 	for (indirdep = NULL, newindirdep = NULL; ; ) {
1814 		ACQUIRE_LOCK(&lk);
1815 		LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1816 			if (wk->wk_type != D_INDIRDEP)
1817 				continue;
1818 			indirdep = WK_INDIRDEP(wk);
1819 			break;
1820 		}
1821 		if (indirdep == NULL && newindirdep) {
1822 			indirdep = newindirdep;
1823 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1824 			newindirdep = NULL;
1825 		}
1826 		FREE_LOCK(&lk);
1827 		if (indirdep) {
1828 			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1829 			    &newblk) == 0)
1830 				panic("setup_allocindir: lost block");
1831 			ACQUIRE_LOCK(&lk);
1832 			if (newblk->nb_state == DEPCOMPLETE) {
1833 				aip->ai_state |= DEPCOMPLETE;
1834 				aip->ai_buf = NULL;
1835 			} else {
1836 				bmsafemap = newblk->nb_bmsafemap;
1837 				aip->ai_buf = bmsafemap->sm_buf;
1838 				LIST_REMOVE(newblk, nb_deps);
1839 				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1840 				    aip, ai_deps);
1841 			}
1842 			LIST_REMOVE(newblk, nb_hash);
1843 			FREE(newblk, M_NEWBLK);
1844 			aip->ai_indirdep = indirdep;
1845 			/*
1846 			 * Check to see if there is an existing dependency
1847 			 * for this block. If there is, merge the old
1848 			 * dependency into the new one.
1849 			 */
1850 			if (aip->ai_oldblkno == 0)
1851 				oldaip = NULL;
1852 			else
1853 
1854 				LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1855 					if (oldaip->ai_offset == aip->ai_offset)
1856 						break;
1857 			freefrag = NULL;
1858 			if (oldaip != NULL) {
1859 				if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1860 					FREE_LOCK(&lk);
1861 					panic("setup_allocindir_phase2: blkno");
1862 				}
1863 				aip->ai_oldblkno = oldaip->ai_oldblkno;
1864 				freefrag = aip->ai_freefrag;
1865 				aip->ai_freefrag = oldaip->ai_freefrag;
1866 				oldaip->ai_freefrag = NULL;
1867 				free_allocindir(oldaip, NULL);
1868 			}
1869 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1870 			if (ip->i_ump->um_fstype == UFS1)
1871 				((ufs1_daddr_t *)indirdep->ir_savebp->b_data)
1872 				    [aip->ai_offset] = aip->ai_oldblkno;
1873 			else
1874 				((ufs2_daddr_t *)indirdep->ir_savebp->b_data)
1875 				    [aip->ai_offset] = aip->ai_oldblkno;
1876 			FREE_LOCK(&lk);
1877 			if (freefrag != NULL)
1878 				handle_workitem_freefrag(freefrag);
1879 		}
1880 		if (newindirdep) {
1881 			brelse(newindirdep->ir_savebp);
1882 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1883 		}
1884 		if (indirdep)
1885 			break;
1886 		MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1887 			M_INDIRDEP, M_SOFTDEP_FLAGS);
1888 		newindirdep->ir_list.wk_type = D_INDIRDEP;
1889 		newindirdep->ir_state = ATTACHED;
1890 		if (ip->i_ump->um_fstype == UFS1)
1891 			newindirdep->ir_state |= UFS1FMT;
1892 		LIST_INIT(&newindirdep->ir_deplisthd);
1893 		LIST_INIT(&newindirdep->ir_donehd);
1894 		if (bp->b_blkno == bp->b_lblkno) {
1895 			ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
1896 			    NULL, NULL);
1897 			bp->b_blkno = blkno;
1898 		}
1899 		newindirdep->ir_savebp =
1900 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
1901 		BUF_KERNPROC(newindirdep->ir_savebp);
1902 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1903 	}
1904 }
1905 
1906 /*
1907  * Block de-allocation dependencies.
1908  *
1909  * When blocks are de-allocated, the on-disk pointers must be nullified before
1910  * the blocks are made available for use by other files.  (The true
1911  * requirement is that old pointers must be nullified before new on-disk
1912  * pointers are set.  We chose this slightly more stringent requirement to
1913  * reduce complexity.) Our implementation handles this dependency by updating
1914  * the inode (or indirect block) appropriately but delaying the actual block
1915  * de-allocation (i.e., freemap and free space count manipulation) until
1916  * after the updated versions reach stable storage.  After the disk is
1917  * updated, the blocks can be safely de-allocated whenever it is convenient.
1918  * This implementation handles only the common case of reducing a file's
1919  * length to zero. Other cases are handled by the conventional synchronous
1920  * write approach.
1921  *
1922  * The ffs implementation with which we worked double-checks
1923  * the state of the block pointers and file size as it reduces
1924  * a file's length.  Some of this code is replicated here in our
1925  * soft updates implementation.  The freeblks->fb_chkcnt field is
1926  * used to transfer a part of this information to the procedure
1927  * that eventually de-allocates the blocks.
1928  *
1929  * This routine should be called from the routine that shortens
1930  * a file's length, before the inode's size or block pointers
1931  * are modified. It will save the block pointer information for
1932  * later release and zero the inode so that the calling routine
1933  * can release it.
1934  */
1935 void
1936 softdep_setup_freeblocks(ip, length, flags)
1937 	struct inode *ip;	/* The inode whose length is to be reduced */
1938 	off_t length;		/* The new length for the file */
1939 	int flags;		/* IO_EXT and/or IO_NORMAL */
1940 {
1941 	struct freeblks *freeblks;
1942 	struct inodedep *inodedep;
1943 	struct allocdirect *adp;
1944 	struct vnode *vp;
1945 	struct buf *bp;
1946 	struct fs *fs;
1947 	ufs2_daddr_t extblocks, datablocks;
1948 	int i, delay, error;
1949 
1950 	fs = ip->i_fs;
1951 	if (length != 0)
1952 		panic("softdep_setup_freeblocks: non-zero length");
1953 	MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1954 		M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
1955 	freeblks->fb_list.wk_type = D_FREEBLKS;
1956 	freeblks->fb_uid = ip->i_uid;
1957 	freeblks->fb_previousinum = ip->i_number;
1958 	freeblks->fb_devvp = ip->i_devvp;
1959 	freeblks->fb_mnt = ITOV(ip)->v_mount;
1960 	extblocks = 0;
1961 	if (fs->fs_magic == FS_UFS2_MAGIC)
1962 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
1963 	datablocks = DIP(ip, i_blocks) - extblocks;
1964 	if ((flags & IO_NORMAL) == 0) {
1965 		freeblks->fb_oldsize = 0;
1966 		freeblks->fb_chkcnt = 0;
1967 	} else {
1968 		freeblks->fb_oldsize = ip->i_size;
1969 		ip->i_size = 0;
1970 		DIP(ip, i_size) = 0;
1971 		freeblks->fb_chkcnt = datablocks;
1972 		for (i = 0; i < NDADDR; i++) {
1973 			freeblks->fb_dblks[i] = DIP(ip, i_db[i]);
1974 			DIP(ip, i_db[i]) = 0;
1975 		}
1976 		for (i = 0; i < NIADDR; i++) {
1977 			freeblks->fb_iblks[i] = DIP(ip, i_ib[i]);
1978 			DIP(ip, i_ib[i]) = 0;
1979 		}
1980 		/*
1981 		 * If the file was removed, then the space being freed was
1982 		 * accounted for then (see softdep_filereleased()). If the
1983 		 * file is merely being truncated, then we account for it now.
1984 		 */
1985 		if ((ip->i_flag & IN_SPACECOUNTED) == 0)
1986 			fs->fs_pendingblocks += datablocks;
1987 	}
1988 	if ((flags & IO_EXT) == 0) {
1989 		freeblks->fb_oldextsize = 0;
1990 	} else {
1991 		freeblks->fb_oldextsize = ip->i_din2->di_extsize;
1992 		ip->i_din2->di_extsize = 0;
1993 		freeblks->fb_chkcnt += extblocks;
1994 		for (i = 0; i < NXADDR; i++) {
1995 			freeblks->fb_eblks[i] = ip->i_din2->di_extb[i];
1996 			ip->i_din2->di_extb[i] = 0;
1997 		}
1998 	}
1999 	DIP(ip, i_blocks) -= freeblks->fb_chkcnt;
2000 	/*
2001 	 * Push the zero'ed inode to to its disk buffer so that we are free
2002 	 * to delete its dependencies below. Once the dependencies are gone
2003 	 * the buffer can be safely released.
2004 	 */
2005 	if ((error = bread(ip->i_devvp,
2006 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
2007 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
2008 		brelse(bp);
2009 		softdep_error("softdep_setup_freeblocks", error);
2010 	}
2011 	if (ip->i_ump->um_fstype == UFS1)
2012 		*((struct ufs1_dinode *)bp->b_data +
2013 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
2014 	else
2015 		*((struct ufs2_dinode *)bp->b_data +
2016 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
2017 	/*
2018 	 * Find and eliminate any inode dependencies.
2019 	 */
2020 	ACQUIRE_LOCK(&lk);
2021 	(void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
2022 	if ((inodedep->id_state & IOSTARTED) != 0) {
2023 		FREE_LOCK(&lk);
2024 		panic("softdep_setup_freeblocks: inode busy");
2025 	}
2026 	/*
2027 	 * Add the freeblks structure to the list of operations that
2028 	 * must await the zero'ed inode being written to disk. If we
2029 	 * still have a bitmap dependency (delay == 0), then the inode
2030 	 * has never been written to disk, so we can process the
2031 	 * freeblks below once we have deleted the dependencies.
2032 	 */
2033 	delay = (inodedep->id_state & DEPCOMPLETE);
2034 	if (delay)
2035 		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
2036 	/*
2037 	 * Because the file length has been truncated to zero, any
2038 	 * pending block allocation dependency structures associated
2039 	 * with this inode are obsolete and can simply be de-allocated.
2040 	 * We must first merge the two dependency lists to get rid of
2041 	 * any duplicate freefrag structures, then purge the merged list.
2042 	 * If we still have a bitmap dependency, then the inode has never
2043 	 * been written to disk, so we can free any fragments without delay.
2044 	 */
2045 	if (flags & IO_NORMAL) {
2046 		merge_inode_lists(&inodedep->id_newinoupdt,
2047 		    &inodedep->id_inoupdt);
2048 		while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
2049 			free_allocdirect(&inodedep->id_inoupdt, adp, delay);
2050 	}
2051 	if (flags & IO_EXT) {
2052 		merge_inode_lists(&inodedep->id_newextupdt,
2053 		    &inodedep->id_extupdt);
2054 		while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
2055 			free_allocdirect(&inodedep->id_extupdt, adp, delay);
2056 	}
2057 	FREE_LOCK(&lk);
2058 	bdwrite(bp);
2059 	/*
2060 	 * We must wait for any I/O in progress to finish so that
2061 	 * all potential buffers on the dirty list will be visible.
2062 	 * Once they are all there, walk the list and get rid of
2063 	 * any dependencies.
2064 	 */
2065 	vp = ITOV(ip);
2066 	ACQUIRE_LOCK(&lk);
2067 	VI_LOCK(vp);
2068 	drain_output(vp, 1);
2069 restart:
2070 	TAILQ_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
2071 		if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
2072 		    ((flags & IO_NORMAL) == 0 &&
2073 		      (bp->b_xflags & BX_ALTDATA) == 0))
2074 			continue;
2075 		if ((bp = getdirtybuf(&bp, VI_MTX(vp), MNT_WAIT)) == NULL)
2076 			goto restart;
2077 		(void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
2078 		deallocate_dependencies(bp, inodedep);
2079 		bp->b_flags |= B_INVAL | B_NOCACHE;
2080 		FREE_LOCK(&lk);
2081 		brelse(bp);
2082 		ACQUIRE_LOCK(&lk);
2083 		VI_LOCK(vp);
2084 		goto restart;
2085 	}
2086 	VI_UNLOCK(vp);
2087 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
2088 		(void) free_inodedep(inodedep);
2089 	FREE_LOCK(&lk);
2090 	/*
2091 	 * If the inode has never been written to disk (delay == 0),
2092 	 * then we can process the freeblks now that we have deleted
2093 	 * the dependencies.
2094 	 */
2095 	if (!delay)
2096 		handle_workitem_freeblocks(freeblks, 0);
2097 }
2098 
2099 /*
2100  * Reclaim any dependency structures from a buffer that is about to
2101  * be reallocated to a new vnode. The buffer must be locked, thus,
2102  * no I/O completion operations can occur while we are manipulating
2103  * its associated dependencies. The mutex is held so that other I/O's
2104  * associated with related dependencies do not occur.
2105  */
2106 static void
2107 deallocate_dependencies(bp, inodedep)
2108 	struct buf *bp;
2109 	struct inodedep *inodedep;
2110 {
2111 	struct worklist *wk;
2112 	struct indirdep *indirdep;
2113 	struct allocindir *aip;
2114 	struct pagedep *pagedep;
2115 	struct dirrem *dirrem;
2116 	struct diradd *dap;
2117 	int i;
2118 
2119 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2120 		switch (wk->wk_type) {
2121 
2122 		case D_INDIRDEP:
2123 			indirdep = WK_INDIRDEP(wk);
2124 			/*
2125 			 * None of the indirect pointers will ever be visible,
2126 			 * so they can simply be tossed. GOINGAWAY ensures
2127 			 * that allocated pointers will be saved in the buffer
2128 			 * cache until they are freed. Note that they will
2129 			 * only be able to be found by their physical address
2130 			 * since the inode mapping the logical address will
2131 			 * be gone. The save buffer used for the safe copy
2132 			 * was allocated in setup_allocindir_phase2 using
2133 			 * the physical address so it could be used for this
2134 			 * purpose. Hence we swap the safe copy with the real
2135 			 * copy, allowing the safe copy to be freed and holding
2136 			 * on to the real copy for later use in indir_trunc.
2137 			 */
2138 			if (indirdep->ir_state & GOINGAWAY) {
2139 				FREE_LOCK(&lk);
2140 				panic("deallocate_dependencies: already gone");
2141 			}
2142 			indirdep->ir_state |= GOINGAWAY;
2143 			VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1;
2144 			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
2145 				free_allocindir(aip, inodedep);
2146 			if (bp->b_lblkno >= 0 ||
2147 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno) {
2148 				FREE_LOCK(&lk);
2149 				panic("deallocate_dependencies: not indir");
2150 			}
2151 			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2152 			    bp->b_bcount);
2153 			WORKLIST_REMOVE(wk);
2154 			WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
2155 			continue;
2156 
2157 		case D_PAGEDEP:
2158 			pagedep = WK_PAGEDEP(wk);
2159 			/*
2160 			 * None of the directory additions will ever be
2161 			 * visible, so they can simply be tossed.
2162 			 */
2163 			for (i = 0; i < DAHASHSZ; i++)
2164 				while ((dap =
2165 				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
2166 					free_diradd(dap);
2167 			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2168 				free_diradd(dap);
2169 			/*
2170 			 * Copy any directory remove dependencies to the list
2171 			 * to be processed after the zero'ed inode is written.
2172 			 * If the inode has already been written, then they
2173 			 * can be dumped directly onto the work list.
2174 			 */
2175 			LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2176 				LIST_REMOVE(dirrem, dm_next);
2177 				dirrem->dm_dirinum = pagedep->pd_ino;
2178 				if (inodedep == NULL ||
2179 				    (inodedep->id_state & ALLCOMPLETE) ==
2180 				     ALLCOMPLETE)
2181 					add_to_worklist(&dirrem->dm_list);
2182 				else
2183 					WORKLIST_INSERT(&inodedep->id_bufwait,
2184 					    &dirrem->dm_list);
2185 			}
2186 			if ((pagedep->pd_state & NEWBLOCK) != 0) {
2187 				LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list)
2188 					if (wk->wk_type == D_NEWDIRBLK &&
2189 					    WK_NEWDIRBLK(wk)->db_pagedep ==
2190 					      pagedep)
2191 						break;
2192 				if (wk != NULL) {
2193 					WORKLIST_REMOVE(wk);
2194 					free_newdirblk(WK_NEWDIRBLK(wk));
2195 				} else {
2196 					FREE_LOCK(&lk);
2197 					panic("deallocate_dependencies: "
2198 					      "lost pagedep");
2199 				}
2200 			}
2201 			WORKLIST_REMOVE(&pagedep->pd_list);
2202 			LIST_REMOVE(pagedep, pd_hash);
2203 			WORKITEM_FREE(pagedep, D_PAGEDEP);
2204 			continue;
2205 
2206 		case D_ALLOCINDIR:
2207 			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2208 			continue;
2209 
2210 		case D_ALLOCDIRECT:
2211 		case D_INODEDEP:
2212 			FREE_LOCK(&lk);
2213 			panic("deallocate_dependencies: Unexpected type %s",
2214 			    TYPENAME(wk->wk_type));
2215 			/* NOTREACHED */
2216 
2217 		default:
2218 			FREE_LOCK(&lk);
2219 			panic("deallocate_dependencies: Unknown type %s",
2220 			    TYPENAME(wk->wk_type));
2221 			/* NOTREACHED */
2222 		}
2223 	}
2224 }
2225 
2226 /*
2227  * Free an allocdirect. Generate a new freefrag work request if appropriate.
2228  * This routine must be called with splbio interrupts blocked.
2229  */
2230 static void
2231 free_allocdirect(adphead, adp, delay)
2232 	struct allocdirectlst *adphead;
2233 	struct allocdirect *adp;
2234 	int delay;
2235 {
2236 	struct newdirblk *newdirblk;
2237 	struct worklist *wk;
2238 
2239 #ifdef DEBUG
2240 	if (lk.lkt_held == NOHOLDER)
2241 		panic("free_allocdirect: lock not held");
2242 #endif
2243 	if ((adp->ad_state & DEPCOMPLETE) == 0)
2244 		LIST_REMOVE(adp, ad_deps);
2245 	TAILQ_REMOVE(adphead, adp, ad_next);
2246 	if ((adp->ad_state & COMPLETE) == 0)
2247 		WORKLIST_REMOVE(&adp->ad_list);
2248 	if (adp->ad_freefrag != NULL) {
2249 		if (delay)
2250 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2251 			    &adp->ad_freefrag->ff_list);
2252 		else
2253 			add_to_worklist(&adp->ad_freefrag->ff_list);
2254 	}
2255 	if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) {
2256 		newdirblk = WK_NEWDIRBLK(wk);
2257 		WORKLIST_REMOVE(&newdirblk->db_list);
2258 		if (LIST_FIRST(&adp->ad_newdirblk) != NULL)
2259 			panic("free_allocdirect: extra newdirblk");
2260 		if (delay)
2261 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2262 			    &newdirblk->db_list);
2263 		else
2264 			free_newdirblk(newdirblk);
2265 	}
2266 	WORKITEM_FREE(adp, D_ALLOCDIRECT);
2267 }
2268 
2269 /*
2270  * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
2271  * This routine must be called with splbio interrupts blocked.
2272  */
2273 static void
2274 free_newdirblk(newdirblk)
2275 	struct newdirblk *newdirblk;
2276 {
2277 	struct pagedep *pagedep;
2278 	struct diradd *dap;
2279 	int i;
2280 
2281 #ifdef DEBUG
2282 	if (lk.lkt_held == NOHOLDER)
2283 		panic("free_newdirblk: lock not held");
2284 #endif
2285 	/*
2286 	 * If the pagedep is still linked onto the directory buffer
2287 	 * dependency chain, then some of the entries on the
2288 	 * pd_pendinghd list may not be committed to disk yet. In
2289 	 * this case, we will simply clear the NEWBLOCK flag and
2290 	 * let the pd_pendinghd list be processed when the pagedep
2291 	 * is next written. If the pagedep is no longer on the buffer
2292 	 * dependency chain, then all the entries on the pd_pending
2293 	 * list are committed to disk and we can free them here.
2294 	 */
2295 	pagedep = newdirblk->db_pagedep;
2296 	pagedep->pd_state &= ~NEWBLOCK;
2297 	if ((pagedep->pd_state & ONWORKLIST) == 0)
2298 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
2299 			free_diradd(dap);
2300 	/*
2301 	 * If no dependencies remain, the pagedep will be freed.
2302 	 */
2303 	for (i = 0; i < DAHASHSZ; i++)
2304 		if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
2305 			break;
2306 	if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) {
2307 		LIST_REMOVE(pagedep, pd_hash);
2308 		WORKITEM_FREE(pagedep, D_PAGEDEP);
2309 	}
2310 	WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2311 }
2312 
2313 /*
2314  * Prepare an inode to be freed. The actual free operation is not
2315  * done until the zero'ed inode has been written to disk.
2316  */
2317 void
2318 softdep_freefile(pvp, ino, mode)
2319 	struct vnode *pvp;
2320 	ino_t ino;
2321 	int mode;
2322 {
2323 	struct inode *ip = VTOI(pvp);
2324 	struct inodedep *inodedep;
2325 	struct freefile *freefile;
2326 
2327 	/*
2328 	 * This sets up the inode de-allocation dependency.
2329 	 */
2330 	MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2331 		M_FREEFILE, M_SOFTDEP_FLAGS);
2332 	freefile->fx_list.wk_type = D_FREEFILE;
2333 	freefile->fx_list.wk_state = 0;
2334 	freefile->fx_mode = mode;
2335 	freefile->fx_oldinum = ino;
2336 	freefile->fx_devvp = ip->i_devvp;
2337 	freefile->fx_mnt = ITOV(ip)->v_mount;
2338 	if ((ip->i_flag & IN_SPACECOUNTED) == 0)
2339 		ip->i_fs->fs_pendinginodes += 1;
2340 
2341 	/*
2342 	 * If the inodedep does not exist, then the zero'ed inode has
2343 	 * been written to disk. If the allocated inode has never been
2344 	 * written to disk, then the on-disk inode is zero'ed. In either
2345 	 * case we can free the file immediately.
2346 	 */
2347 	ACQUIRE_LOCK(&lk);
2348 	if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2349 	    check_inode_unwritten(inodedep)) {
2350 		FREE_LOCK(&lk);
2351 		handle_workitem_freefile(freefile);
2352 		return;
2353 	}
2354 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2355 	FREE_LOCK(&lk);
2356 }
2357 
2358 /*
2359  * Check to see if an inode has never been written to disk. If
2360  * so free the inodedep and return success, otherwise return failure.
2361  * This routine must be called with splbio interrupts blocked.
2362  *
2363  * If we still have a bitmap dependency, then the inode has never
2364  * been written to disk. Drop the dependency as it is no longer
2365  * necessary since the inode is being deallocated. We set the
2366  * ALLCOMPLETE flags since the bitmap now properly shows that the
2367  * inode is not allocated. Even if the inode is actively being
2368  * written, it has been rolled back to its zero'ed state, so we
2369  * are ensured that a zero inode is what is on the disk. For short
2370  * lived files, this change will usually result in removing all the
2371  * dependencies from the inode so that it can be freed immediately.
2372  */
2373 static int
2374 check_inode_unwritten(inodedep)
2375 	struct inodedep *inodedep;
2376 {
2377 
2378 	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2379 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2380 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2381 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
2382 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2383 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2384 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
2385 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
2386 	    inodedep->id_nlinkdelta != 0)
2387 		return (0);
2388 	inodedep->id_state |= ALLCOMPLETE;
2389 	LIST_REMOVE(inodedep, id_deps);
2390 	inodedep->id_buf = NULL;
2391 	if (inodedep->id_state & ONWORKLIST)
2392 		WORKLIST_REMOVE(&inodedep->id_list);
2393 	if (inodedep->id_savedino1 != NULL) {
2394 		FREE(inodedep->id_savedino1, M_INODEDEP);
2395 		inodedep->id_savedino1 = NULL;
2396 	}
2397 	if (free_inodedep(inodedep) == 0) {
2398 		FREE_LOCK(&lk);
2399 		panic("check_inode_unwritten: busy inode");
2400 	}
2401 	return (1);
2402 }
2403 
2404 /*
2405  * Try to free an inodedep structure. Return 1 if it could be freed.
2406  */
2407 static int
2408 free_inodedep(inodedep)
2409 	struct inodedep *inodedep;
2410 {
2411 
2412 	if ((inodedep->id_state & ONWORKLIST) != 0 ||
2413 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2414 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2415 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2416 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
2417 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2418 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2419 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
2420 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
2421 	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL)
2422 		return (0);
2423 	LIST_REMOVE(inodedep, id_hash);
2424 	WORKITEM_FREE(inodedep, D_INODEDEP);
2425 	num_inodedep -= 1;
2426 	return (1);
2427 }
2428 
2429 /*
2430  * This workitem routine performs the block de-allocation.
2431  * The workitem is added to the pending list after the updated
2432  * inode block has been written to disk.  As mentioned above,
2433  * checks regarding the number of blocks de-allocated (compared
2434  * to the number of blocks allocated for the file) are also
2435  * performed in this function.
2436  */
2437 static void
2438 handle_workitem_freeblocks(freeblks, flags)
2439 	struct freeblks *freeblks;
2440 	int flags;
2441 {
2442 	struct inode *ip;
2443 	struct vnode *vp;
2444 	struct fs *fs;
2445 	int i, nblocks, level, bsize;
2446 	ufs2_daddr_t bn, blocksreleased = 0;
2447 	int error, allerror = 0;
2448 	ufs_lbn_t baselbns[NIADDR], tmpval;
2449 
2450 	fs = VFSTOUFS(freeblks->fb_mnt)->um_fs;
2451 	tmpval = 1;
2452 	baselbns[0] = NDADDR;
2453 	for (i = 1; i < NIADDR; i++) {
2454 		tmpval *= NINDIR(fs);
2455 		baselbns[i] = baselbns[i - 1] + tmpval;
2456 	}
2457 	nblocks = btodb(fs->fs_bsize);
2458 	blocksreleased = 0;
2459 	/*
2460 	 * Release all extended attribute blocks or frags.
2461 	 */
2462 	if (freeblks->fb_oldextsize > 0) {
2463 		for (i = (NXADDR - 1); i >= 0; i--) {
2464 			if ((bn = freeblks->fb_eblks[i]) == 0)
2465 				continue;
2466 			bsize = sblksize(fs, freeblks->fb_oldextsize, i);
2467 			ffs_blkfree(fs, freeblks->fb_devvp, bn, bsize,
2468 			    freeblks->fb_previousinum);
2469 			blocksreleased += btodb(bsize);
2470 		}
2471 	}
2472 	/*
2473 	 * Release all data blocks or frags.
2474 	 */
2475 	if (freeblks->fb_oldsize > 0) {
2476 		/*
2477 		 * Indirect blocks first.
2478 		 */
2479 		for (level = (NIADDR - 1); level >= 0; level--) {
2480 			if ((bn = freeblks->fb_iblks[level]) == 0)
2481 				continue;
2482 			if ((error = indir_trunc(freeblks, fsbtodb(fs, bn),
2483 			    level, baselbns[level], &blocksreleased)) == 0)
2484 				allerror = error;
2485 			ffs_blkfree(fs, freeblks->fb_devvp, bn, fs->fs_bsize,
2486 			    freeblks->fb_previousinum);
2487 			fs->fs_pendingblocks -= nblocks;
2488 			blocksreleased += nblocks;
2489 		}
2490 		/*
2491 		 * All direct blocks or frags.
2492 		 */
2493 		for (i = (NDADDR - 1); i >= 0; i--) {
2494 			if ((bn = freeblks->fb_dblks[i]) == 0)
2495 				continue;
2496 			bsize = sblksize(fs, freeblks->fb_oldsize, i);
2497 			ffs_blkfree(fs, freeblks->fb_devvp, bn, bsize,
2498 			    freeblks->fb_previousinum);
2499 			fs->fs_pendingblocks -= btodb(bsize);
2500 			blocksreleased += btodb(bsize);
2501 		}
2502 	}
2503 	/*
2504 	 * If we still have not finished background cleanup, then check
2505 	 * to see if the block count needs to be adjusted.
2506 	 */
2507 	if (freeblks->fb_chkcnt != blocksreleased &&
2508 	    (fs->fs_flags & FS_UNCLEAN) != 0 &&
2509 	    VFS_VGET(freeblks->fb_mnt, freeblks->fb_previousinum,
2510 	    (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) {
2511 		ip = VTOI(vp);
2512 		DIP(ip, i_blocks) += freeblks->fb_chkcnt - blocksreleased;
2513 		ip->i_flag |= IN_CHANGE;
2514 		vput(vp);
2515 	}
2516 
2517 #ifdef DIAGNOSTIC
2518 	if (freeblks->fb_chkcnt != blocksreleased &&
2519 	    ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0))
2520 		printf("handle_workitem_freeblocks: block count\n");
2521 	if (allerror)
2522 		softdep_error("handle_workitem_freeblks", allerror);
2523 #endif /* DIAGNOSTIC */
2524 
2525 	WORKITEM_FREE(freeblks, D_FREEBLKS);
2526 }
2527 
2528 /*
2529  * Release blocks associated with the inode ip and stored in the indirect
2530  * block dbn. If level is greater than SINGLE, the block is an indirect block
2531  * and recursive calls to indirtrunc must be used to cleanse other indirect
2532  * blocks.
2533  */
2534 static int
2535 indir_trunc(freeblks, dbn, level, lbn, countp)
2536 	struct freeblks *freeblks;
2537 	ufs2_daddr_t dbn;
2538 	int level;
2539 	ufs_lbn_t lbn;
2540 	ufs2_daddr_t *countp;
2541 {
2542 	struct buf *bp;
2543 	struct fs *fs;
2544 	struct worklist *wk;
2545 	struct indirdep *indirdep;
2546 	ufs1_daddr_t *bap1 = 0;
2547 	ufs2_daddr_t nb, *bap2 = 0;
2548 	ufs_lbn_t lbnadd;
2549 	int i, nblocks, ufs1fmt;
2550 	int error, allerror = 0;
2551 
2552 	fs = VFSTOUFS(freeblks->fb_mnt)->um_fs;
2553 	lbnadd = 1;
2554 	for (i = level; i > 0; i--)
2555 		lbnadd *= NINDIR(fs);
2556 	/*
2557 	 * Get buffer of block pointers to be freed. This routine is not
2558 	 * called until the zero'ed inode has been written, so it is safe
2559 	 * to free blocks as they are encountered. Because the inode has
2560 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
2561 	 * have to use the on-disk address and the block device for the
2562 	 * filesystem to look them up. If the file was deleted before its
2563 	 * indirect blocks were all written to disk, the routine that set
2564 	 * us up (deallocate_dependencies) will have arranged to leave
2565 	 * a complete copy of the indirect block in memory for our use.
2566 	 * Otherwise we have to read the blocks in from the disk.
2567 	 */
2568 #ifdef notyet
2569 	bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0,
2570 	    GB_NOCREAT);
2571 #else
2572 	bp = incore(freeblks->fb_devvp, dbn);
2573 #endif
2574 	ACQUIRE_LOCK(&lk);
2575 	if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2576 		if (wk->wk_type != D_INDIRDEP ||
2577 		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2578 		    (indirdep->ir_state & GOINGAWAY) == 0) {
2579 			FREE_LOCK(&lk);
2580 			panic("indir_trunc: lost indirdep");
2581 		}
2582 		WORKLIST_REMOVE(wk);
2583 		WORKITEM_FREE(indirdep, D_INDIRDEP);
2584 		if (LIST_FIRST(&bp->b_dep) != NULL) {
2585 			FREE_LOCK(&lk);
2586 			panic("indir_trunc: dangling dep");
2587 		}
2588 		VFSTOUFS(freeblks->fb_mnt)->um_numindirdeps -= 1;
2589 		FREE_LOCK(&lk);
2590 	} else {
2591 #ifdef notyet
2592 		if (bp)
2593 			brelse(bp);
2594 #endif
2595 		FREE_LOCK(&lk);
2596 		error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
2597 		    NOCRED, &bp);
2598 		if (error) {
2599 			brelse(bp);
2600 			return (error);
2601 		}
2602 	}
2603 	/*
2604 	 * Recursively free indirect blocks.
2605 	 */
2606 	if (VFSTOUFS(freeblks->fb_mnt)->um_fstype == UFS1) {
2607 		ufs1fmt = 1;
2608 		bap1 = (ufs1_daddr_t *)bp->b_data;
2609 	} else {
2610 		ufs1fmt = 0;
2611 		bap2 = (ufs2_daddr_t *)bp->b_data;
2612 	}
2613 	nblocks = btodb(fs->fs_bsize);
2614 	for (i = NINDIR(fs) - 1; i >= 0; i--) {
2615 		if (ufs1fmt)
2616 			nb = bap1[i];
2617 		else
2618 			nb = bap2[i];
2619 		if (nb == 0)
2620 			continue;
2621 		if (level != 0) {
2622 			if ((error = indir_trunc(freeblks, fsbtodb(fs, nb),
2623 			     level - 1, lbn + (i * lbnadd), countp)) != 0)
2624 				allerror = error;
2625 		}
2626 		ffs_blkfree(fs, freeblks->fb_devvp, nb, fs->fs_bsize,
2627 		    freeblks->fb_previousinum);
2628 		fs->fs_pendingblocks -= nblocks;
2629 		*countp += nblocks;
2630 	}
2631 	bp->b_flags |= B_INVAL | B_NOCACHE;
2632 	brelse(bp);
2633 	return (allerror);
2634 }
2635 
2636 /*
2637  * Free an allocindir.
2638  * This routine must be called with splbio interrupts blocked.
2639  */
2640 static void
2641 free_allocindir(aip, inodedep)
2642 	struct allocindir *aip;
2643 	struct inodedep *inodedep;
2644 {
2645 	struct freefrag *freefrag;
2646 
2647 #ifdef DEBUG
2648 	if (lk.lkt_held == NOHOLDER)
2649 		panic("free_allocindir: lock not held");
2650 #endif
2651 	if ((aip->ai_state & DEPCOMPLETE) == 0)
2652 		LIST_REMOVE(aip, ai_deps);
2653 	if (aip->ai_state & ONWORKLIST)
2654 		WORKLIST_REMOVE(&aip->ai_list);
2655 	LIST_REMOVE(aip, ai_next);
2656 	if ((freefrag = aip->ai_freefrag) != NULL) {
2657 		if (inodedep == NULL)
2658 			add_to_worklist(&freefrag->ff_list);
2659 		else
2660 			WORKLIST_INSERT(&inodedep->id_bufwait,
2661 			    &freefrag->ff_list);
2662 	}
2663 	WORKITEM_FREE(aip, D_ALLOCINDIR);
2664 }
2665 
2666 /*
2667  * Directory entry addition dependencies.
2668  *
2669  * When adding a new directory entry, the inode (with its incremented link
2670  * count) must be written to disk before the directory entry's pointer to it.
2671  * Also, if the inode is newly allocated, the corresponding freemap must be
2672  * updated (on disk) before the directory entry's pointer. These requirements
2673  * are met via undo/redo on the directory entry's pointer, which consists
2674  * simply of the inode number.
2675  *
2676  * As directory entries are added and deleted, the free space within a
2677  * directory block can become fragmented.  The ufs filesystem will compact
2678  * a fragmented directory block to make space for a new entry. When this
2679  * occurs, the offsets of previously added entries change. Any "diradd"
2680  * dependency structures corresponding to these entries must be updated with
2681  * the new offsets.
2682  */
2683 
2684 /*
2685  * This routine is called after the in-memory inode's link
2686  * count has been incremented, but before the directory entry's
2687  * pointer to the inode has been set.
2688  */
2689 int
2690 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
2691 	struct buf *bp;		/* buffer containing directory block */
2692 	struct inode *dp;	/* inode for directory */
2693 	off_t diroffset;	/* offset of new entry in directory */
2694 	ino_t newinum;		/* inode referenced by new directory entry */
2695 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
2696 	int isnewblk;		/* entry is in a newly allocated block */
2697 {
2698 	int offset;		/* offset of new entry within directory block */
2699 	ufs_lbn_t lbn;		/* block in directory containing new entry */
2700 	struct fs *fs;
2701 	struct diradd *dap;
2702 	struct allocdirect *adp;
2703 	struct pagedep *pagedep;
2704 	struct inodedep *inodedep;
2705 	struct newdirblk *newdirblk = 0;
2706 	struct mkdir *mkdir1, *mkdir2;
2707 
2708 	/*
2709 	 * Whiteouts have no dependencies.
2710 	 */
2711 	if (newinum == WINO) {
2712 		if (newdirbp != NULL)
2713 			bdwrite(newdirbp);
2714 		return (0);
2715 	}
2716 
2717 	fs = dp->i_fs;
2718 	lbn = lblkno(fs, diroffset);
2719 	offset = blkoff(fs, diroffset);
2720 	MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2721 		M_SOFTDEP_FLAGS|M_ZERO);
2722 	dap->da_list.wk_type = D_DIRADD;
2723 	dap->da_offset = offset;
2724 	dap->da_newinum = newinum;
2725 	dap->da_state = ATTACHED;
2726 	if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) {
2727 		MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk),
2728 		    M_NEWDIRBLK, M_SOFTDEP_FLAGS);
2729 		newdirblk->db_list.wk_type = D_NEWDIRBLK;
2730 		newdirblk->db_state = 0;
2731 	}
2732 	if (newdirbp == NULL) {
2733 		dap->da_state |= DEPCOMPLETE;
2734 		ACQUIRE_LOCK(&lk);
2735 	} else {
2736 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2737 		MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2738 		    M_SOFTDEP_FLAGS);
2739 		mkdir1->md_list.wk_type = D_MKDIR;
2740 		mkdir1->md_state = MKDIR_BODY;
2741 		mkdir1->md_diradd = dap;
2742 		MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2743 		    M_SOFTDEP_FLAGS);
2744 		mkdir2->md_list.wk_type = D_MKDIR;
2745 		mkdir2->md_state = MKDIR_PARENT;
2746 		mkdir2->md_diradd = dap;
2747 		/*
2748 		 * Dependency on "." and ".." being written to disk.
2749 		 */
2750 		mkdir1->md_buf = newdirbp;
2751 		ACQUIRE_LOCK(&lk);
2752 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2753 		WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2754 		FREE_LOCK(&lk);
2755 		bdwrite(newdirbp);
2756 		/*
2757 		 * Dependency on link count increase for parent directory
2758 		 */
2759 		ACQUIRE_LOCK(&lk);
2760 		if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0
2761 		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2762 			dap->da_state &= ~MKDIR_PARENT;
2763 			WORKITEM_FREE(mkdir2, D_MKDIR);
2764 		} else {
2765 			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2766 			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2767 		}
2768 	}
2769 	/*
2770 	 * Link into parent directory pagedep to await its being written.
2771 	 */
2772 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2773 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2774 	dap->da_pagedep = pagedep;
2775 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2776 	    da_pdlist);
2777 	/*
2778 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
2779 	 * is not yet written. If it is written, do the post-inode write
2780 	 * processing to put it on the id_pendinghd list.
2781 	 */
2782 	(void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2783 	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2784 		diradd_inode_written(dap, inodedep);
2785 	else
2786 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2787 	if (isnewblk) {
2788 		/*
2789 		 * Directories growing into indirect blocks are rare
2790 		 * enough and the frequency of new block allocation
2791 		 * in those cases even more rare, that we choose not
2792 		 * to bother tracking them. Rather we simply force the
2793 		 * new directory entry to disk.
2794 		 */
2795 		if (lbn >= NDADDR) {
2796 			FREE_LOCK(&lk);
2797 			/*
2798 			 * We only have a new allocation when at the
2799 			 * beginning of a new block, not when we are
2800 			 * expanding into an existing block.
2801 			 */
2802 			if (blkoff(fs, diroffset) == 0)
2803 				return (1);
2804 			return (0);
2805 		}
2806 		/*
2807 		 * We only have a new allocation when at the beginning
2808 		 * of a new fragment, not when we are expanding into an
2809 		 * existing fragment. Also, there is nothing to do if we
2810 		 * are already tracking this block.
2811 		 */
2812 		if (fragoff(fs, diroffset) != 0) {
2813 			FREE_LOCK(&lk);
2814 			return (0);
2815 		}
2816 		if ((pagedep->pd_state & NEWBLOCK) != 0) {
2817 			WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2818 			FREE_LOCK(&lk);
2819 			return (0);
2820 		}
2821 		/*
2822 		 * Find our associated allocdirect and have it track us.
2823 		 */
2824 		if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0)
2825 			panic("softdep_setup_directory_add: lost inodedep");
2826 		adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst);
2827 		if (adp == NULL || adp->ad_lbn != lbn) {
2828 			FREE_LOCK(&lk);
2829 			panic("softdep_setup_directory_add: lost entry");
2830 		}
2831 		pagedep->pd_state |= NEWBLOCK;
2832 		newdirblk->db_pagedep = pagedep;
2833 		WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list);
2834 	}
2835 	FREE_LOCK(&lk);
2836 	return (0);
2837 }
2838 
2839 /*
2840  * This procedure is called to change the offset of a directory
2841  * entry when compacting a directory block which must be owned
2842  * exclusively by the caller. Note that the actual entry movement
2843  * must be done in this procedure to ensure that no I/O completions
2844  * occur while the move is in progress.
2845  */
2846 void
2847 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2848 	struct inode *dp;	/* inode for directory */
2849 	caddr_t base;		/* address of dp->i_offset */
2850 	caddr_t oldloc;		/* address of old directory location */
2851 	caddr_t newloc;		/* address of new directory location */
2852 	int entrysize;		/* size of directory entry */
2853 {
2854 	int offset, oldoffset, newoffset;
2855 	struct pagedep *pagedep;
2856 	struct diradd *dap;
2857 	ufs_lbn_t lbn;
2858 
2859 	ACQUIRE_LOCK(&lk);
2860 	lbn = lblkno(dp->i_fs, dp->i_offset);
2861 	offset = blkoff(dp->i_fs, dp->i_offset);
2862 	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2863 		goto done;
2864 	oldoffset = offset + (oldloc - base);
2865 	newoffset = offset + (newloc - base);
2866 
2867 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2868 		if (dap->da_offset != oldoffset)
2869 			continue;
2870 		dap->da_offset = newoffset;
2871 		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2872 			break;
2873 		LIST_REMOVE(dap, da_pdlist);
2874 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2875 		    dap, da_pdlist);
2876 		break;
2877 	}
2878 	if (dap == NULL) {
2879 
2880 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2881 			if (dap->da_offset == oldoffset) {
2882 				dap->da_offset = newoffset;
2883 				break;
2884 			}
2885 		}
2886 	}
2887 done:
2888 	bcopy(oldloc, newloc, entrysize);
2889 	FREE_LOCK(&lk);
2890 }
2891 
2892 /*
2893  * Free a diradd dependency structure. This routine must be called
2894  * with splbio interrupts blocked.
2895  */
2896 static void
2897 free_diradd(dap)
2898 	struct diradd *dap;
2899 {
2900 	struct dirrem *dirrem;
2901 	struct pagedep *pagedep;
2902 	struct inodedep *inodedep;
2903 	struct mkdir *mkdir, *nextmd;
2904 
2905 #ifdef DEBUG
2906 	if (lk.lkt_held == NOHOLDER)
2907 		panic("free_diradd: lock not held");
2908 #endif
2909 	WORKLIST_REMOVE(&dap->da_list);
2910 	LIST_REMOVE(dap, da_pdlist);
2911 	if ((dap->da_state & DIRCHG) == 0) {
2912 		pagedep = dap->da_pagedep;
2913 	} else {
2914 		dirrem = dap->da_previous;
2915 		pagedep = dirrem->dm_pagedep;
2916 		dirrem->dm_dirinum = pagedep->pd_ino;
2917 		add_to_worklist(&dirrem->dm_list);
2918 	}
2919 	if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2920 	    0, &inodedep) != 0)
2921 		(void) free_inodedep(inodedep);
2922 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2923 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2924 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
2925 			if (mkdir->md_diradd != dap)
2926 				continue;
2927 			dap->da_state &= ~mkdir->md_state;
2928 			WORKLIST_REMOVE(&mkdir->md_list);
2929 			LIST_REMOVE(mkdir, md_mkdirs);
2930 			WORKITEM_FREE(mkdir, D_MKDIR);
2931 		}
2932 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2933 			FREE_LOCK(&lk);
2934 			panic("free_diradd: unfound ref");
2935 		}
2936 	}
2937 	WORKITEM_FREE(dap, D_DIRADD);
2938 }
2939 
2940 /*
2941  * Directory entry removal dependencies.
2942  *
2943  * When removing a directory entry, the entry's inode pointer must be
2944  * zero'ed on disk before the corresponding inode's link count is decremented
2945  * (possibly freeing the inode for re-use). This dependency is handled by
2946  * updating the directory entry but delaying the inode count reduction until
2947  * after the directory block has been written to disk. After this point, the
2948  * inode count can be decremented whenever it is convenient.
2949  */
2950 
2951 /*
2952  * This routine should be called immediately after removing
2953  * a directory entry.  The inode's link count should not be
2954  * decremented by the calling procedure -- the soft updates
2955  * code will do this task when it is safe.
2956  */
2957 void
2958 softdep_setup_remove(bp, dp, ip, isrmdir)
2959 	struct buf *bp;		/* buffer containing directory block */
2960 	struct inode *dp;	/* inode for the directory being modified */
2961 	struct inode *ip;	/* inode for directory entry being removed */
2962 	int isrmdir;		/* indicates if doing RMDIR */
2963 {
2964 	struct dirrem *dirrem, *prevdirrem;
2965 
2966 	/*
2967 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2968 	 */
2969 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2970 
2971 	/*
2972 	 * If the COMPLETE flag is clear, then there were no active
2973 	 * entries and we want to roll back to a zeroed entry until
2974 	 * the new inode is committed to disk. If the COMPLETE flag is
2975 	 * set then we have deleted an entry that never made it to
2976 	 * disk. If the entry we deleted resulted from a name change,
2977 	 * then the old name still resides on disk. We cannot delete
2978 	 * its inode (returned to us in prevdirrem) until the zeroed
2979 	 * directory entry gets to disk. The new inode has never been
2980 	 * referenced on the disk, so can be deleted immediately.
2981 	 */
2982 	if ((dirrem->dm_state & COMPLETE) == 0) {
2983 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2984 		    dm_next);
2985 		FREE_LOCK(&lk);
2986 	} else {
2987 		if (prevdirrem != NULL)
2988 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2989 			    prevdirrem, dm_next);
2990 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2991 		FREE_LOCK(&lk);
2992 		handle_workitem_remove(dirrem, NULL);
2993 	}
2994 }
2995 
2996 /*
2997  * Allocate a new dirrem if appropriate and return it along with
2998  * its associated pagedep. Called without a lock, returns with lock.
2999  */
3000 static long num_dirrem;		/* number of dirrem allocated */
3001 static struct dirrem *
3002 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
3003 	struct buf *bp;		/* buffer containing directory block */
3004 	struct inode *dp;	/* inode for the directory being modified */
3005 	struct inode *ip;	/* inode for directory entry being removed */
3006 	int isrmdir;		/* indicates if doing RMDIR */
3007 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
3008 {
3009 	int offset;
3010 	ufs_lbn_t lbn;
3011 	struct diradd *dap;
3012 	struct dirrem *dirrem;
3013 	struct pagedep *pagedep;
3014 
3015 	/*
3016 	 * Whiteouts have no deletion dependencies.
3017 	 */
3018 	if (ip == NULL)
3019 		panic("newdirrem: whiteout");
3020 	/*
3021 	 * If we are over our limit, try to improve the situation.
3022 	 * Limiting the number of dirrem structures will also limit
3023 	 * the number of freefile and freeblks structures.
3024 	 */
3025 	if (num_dirrem > max_softdeps / 2)
3026 		(void) request_cleanup(FLUSH_REMOVE, 0);
3027 	num_dirrem += 1;
3028 	MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
3029 		M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO);
3030 	dirrem->dm_list.wk_type = D_DIRREM;
3031 	dirrem->dm_state = isrmdir ? RMDIR : 0;
3032 	dirrem->dm_mnt = ITOV(ip)->v_mount;
3033 	dirrem->dm_oldinum = ip->i_number;
3034 	*prevdirremp = NULL;
3035 
3036 	ACQUIRE_LOCK(&lk);
3037 	lbn = lblkno(dp->i_fs, dp->i_offset);
3038 	offset = blkoff(dp->i_fs, dp->i_offset);
3039 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3040 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3041 	dirrem->dm_pagedep = pagedep;
3042 	/*
3043 	 * Check for a diradd dependency for the same directory entry.
3044 	 * If present, then both dependencies become obsolete and can
3045 	 * be de-allocated. Check for an entry on both the pd_dirraddhd
3046 	 * list and the pd_pendinghd list.
3047 	 */
3048 
3049 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
3050 		if (dap->da_offset == offset)
3051 			break;
3052 	if (dap == NULL) {
3053 
3054 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
3055 			if (dap->da_offset == offset)
3056 				break;
3057 		if (dap == NULL)
3058 			return (dirrem);
3059 	}
3060 	/*
3061 	 * Must be ATTACHED at this point.
3062 	 */
3063 	if ((dap->da_state & ATTACHED) == 0) {
3064 		FREE_LOCK(&lk);
3065 		panic("newdirrem: not ATTACHED");
3066 	}
3067 	if (dap->da_newinum != ip->i_number) {
3068 		FREE_LOCK(&lk);
3069 		panic("newdirrem: inum %d should be %d",
3070 		    ip->i_number, dap->da_newinum);
3071 	}
3072 	/*
3073 	 * If we are deleting a changed name that never made it to disk,
3074 	 * then return the dirrem describing the previous inode (which
3075 	 * represents the inode currently referenced from this entry on disk).
3076 	 */
3077 	if ((dap->da_state & DIRCHG) != 0) {
3078 		*prevdirremp = dap->da_previous;
3079 		dap->da_state &= ~DIRCHG;
3080 		dap->da_pagedep = pagedep;
3081 	}
3082 	/*
3083 	 * We are deleting an entry that never made it to disk.
3084 	 * Mark it COMPLETE so we can delete its inode immediately.
3085 	 */
3086 	dirrem->dm_state |= COMPLETE;
3087 	free_diradd(dap);
3088 	return (dirrem);
3089 }
3090 
3091 /*
3092  * Directory entry change dependencies.
3093  *
3094  * Changing an existing directory entry requires that an add operation
3095  * be completed first followed by a deletion. The semantics for the addition
3096  * are identical to the description of adding a new entry above except
3097  * that the rollback is to the old inode number rather than zero. Once
3098  * the addition dependency is completed, the removal is done as described
3099  * in the removal routine above.
3100  */
3101 
3102 /*
3103  * This routine should be called immediately after changing
3104  * a directory entry.  The inode's link count should not be
3105  * decremented by the calling procedure -- the soft updates
3106  * code will perform this task when it is safe.
3107  */
3108 void
3109 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
3110 	struct buf *bp;		/* buffer containing directory block */
3111 	struct inode *dp;	/* inode for the directory being modified */
3112 	struct inode *ip;	/* inode for directory entry being removed */
3113 	ino_t newinum;		/* new inode number for changed entry */
3114 	int isrmdir;		/* indicates if doing RMDIR */
3115 {
3116 	int offset;
3117 	struct diradd *dap = NULL;
3118 	struct dirrem *dirrem, *prevdirrem;
3119 	struct pagedep *pagedep;
3120 	struct inodedep *inodedep;
3121 
3122 	offset = blkoff(dp->i_fs, dp->i_offset);
3123 
3124 	/*
3125 	 * Whiteouts do not need diradd dependencies.
3126 	 */
3127 	if (newinum != WINO) {
3128 		MALLOC(dap, struct diradd *, sizeof(struct diradd),
3129 		    M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
3130 		dap->da_list.wk_type = D_DIRADD;
3131 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
3132 		dap->da_offset = offset;
3133 		dap->da_newinum = newinum;
3134 	}
3135 
3136 	/*
3137 	 * Allocate a new dirrem and ACQUIRE_LOCK.
3138 	 */
3139 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3140 	pagedep = dirrem->dm_pagedep;
3141 	/*
3142 	 * The possible values for isrmdir:
3143 	 *	0 - non-directory file rename
3144 	 *	1 - directory rename within same directory
3145 	 *   inum - directory rename to new directory of given inode number
3146 	 * When renaming to a new directory, we are both deleting and
3147 	 * creating a new directory entry, so the link count on the new
3148 	 * directory should not change. Thus we do not need the followup
3149 	 * dirrem which is usually done in handle_workitem_remove. We set
3150 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
3151 	 * followup dirrem.
3152 	 */
3153 	if (isrmdir > 1)
3154 		dirrem->dm_state |= DIRCHG;
3155 
3156 	/*
3157 	 * Whiteouts have no additional dependencies,
3158 	 * so just put the dirrem on the correct list.
3159 	 */
3160 	if (newinum == WINO) {
3161 		if ((dirrem->dm_state & COMPLETE) == 0) {
3162 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
3163 			    dm_next);
3164 		} else {
3165 			dirrem->dm_dirinum = pagedep->pd_ino;
3166 			add_to_worklist(&dirrem->dm_list);
3167 		}
3168 		FREE_LOCK(&lk);
3169 		return;
3170 	}
3171 
3172 	/*
3173 	 * If the COMPLETE flag is clear, then there were no active
3174 	 * entries and we want to roll back to the previous inode until
3175 	 * the new inode is committed to disk. If the COMPLETE flag is
3176 	 * set, then we have deleted an entry that never made it to disk.
3177 	 * If the entry we deleted resulted from a name change, then the old
3178 	 * inode reference still resides on disk. Any rollback that we do
3179 	 * needs to be to that old inode (returned to us in prevdirrem). If
3180 	 * the entry we deleted resulted from a create, then there is
3181 	 * no entry on the disk, so we want to roll back to zero rather
3182 	 * than the uncommitted inode. In either of the COMPLETE cases we
3183 	 * want to immediately free the unwritten and unreferenced inode.
3184 	 */
3185 	if ((dirrem->dm_state & COMPLETE) == 0) {
3186 		dap->da_previous = dirrem;
3187 	} else {
3188 		if (prevdirrem != NULL) {
3189 			dap->da_previous = prevdirrem;
3190 		} else {
3191 			dap->da_state &= ~DIRCHG;
3192 			dap->da_pagedep = pagedep;
3193 		}
3194 		dirrem->dm_dirinum = pagedep->pd_ino;
3195 		add_to_worklist(&dirrem->dm_list);
3196 	}
3197 	/*
3198 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3199 	 * is not yet written. If it is written, do the post-inode write
3200 	 * processing to put it on the id_pendinghd list.
3201 	 */
3202 	if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
3203 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3204 		dap->da_state |= COMPLETE;
3205 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3206 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3207 	} else {
3208 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
3209 		    dap, da_pdlist);
3210 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3211 	}
3212 	FREE_LOCK(&lk);
3213 }
3214 
3215 /*
3216  * Called whenever the link count on an inode is changed.
3217  * It creates an inode dependency so that the new reference(s)
3218  * to the inode cannot be committed to disk until the updated
3219  * inode has been written.
3220  */
3221 void
3222 softdep_change_linkcnt(ip)
3223 	struct inode *ip;	/* the inode with the increased link count */
3224 {
3225 	struct inodedep *inodedep;
3226 
3227 	ACQUIRE_LOCK(&lk);
3228 	(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
3229 	if (ip->i_nlink < ip->i_effnlink) {
3230 		FREE_LOCK(&lk);
3231 		panic("softdep_change_linkcnt: bad delta");
3232 	}
3233 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3234 	FREE_LOCK(&lk);
3235 }
3236 
3237 /*
3238  * Called when the effective link count and the reference count
3239  * on an inode drops to zero. At this point there are no names
3240  * referencing the file in the filesystem and no active file
3241  * references. The space associated with the file will be freed
3242  * as soon as the necessary soft dependencies are cleared.
3243  */
3244 void
3245 softdep_releasefile(ip)
3246 	struct inode *ip;	/* inode with the zero effective link count */
3247 {
3248 	struct inodedep *inodedep;
3249 	struct fs *fs;
3250 	int extblocks;
3251 
3252 	if (ip->i_effnlink > 0)
3253 		panic("softdep_filerelease: file still referenced");
3254 	/*
3255 	 * We may be called several times as the real reference count
3256 	 * drops to zero. We only want to account for the space once.
3257 	 */
3258 	if (ip->i_flag & IN_SPACECOUNTED)
3259 		return;
3260 	/*
3261 	 * We have to deactivate a snapshot otherwise copyonwrites may
3262 	 * add blocks and the cleanup may remove blocks after we have
3263 	 * tried to account for them.
3264 	 */
3265 	if ((ip->i_flags & SF_SNAPSHOT) != 0)
3266 		ffs_snapremove(ITOV(ip));
3267 	/*
3268 	 * If we are tracking an nlinkdelta, we have to also remember
3269 	 * whether we accounted for the freed space yet.
3270 	 */
3271 	ACQUIRE_LOCK(&lk);
3272 	if ((inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep)))
3273 		inodedep->id_state |= SPACECOUNTED;
3274 	FREE_LOCK(&lk);
3275 	fs = ip->i_fs;
3276 	extblocks = 0;
3277 	if (fs->fs_magic == FS_UFS2_MAGIC)
3278 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
3279 	ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks;
3280 	ip->i_fs->fs_pendinginodes += 1;
3281 	ip->i_flag |= IN_SPACECOUNTED;
3282 }
3283 
3284 /*
3285  * This workitem decrements the inode's link count.
3286  * If the link count reaches zero, the file is removed.
3287  */
3288 static void
3289 handle_workitem_remove(dirrem, xp)
3290 	struct dirrem *dirrem;
3291 	struct vnode *xp;
3292 {
3293 	struct thread *td = curthread;
3294 	struct inodedep *inodedep;
3295 	struct vnode *vp;
3296 	struct inode *ip;
3297 	ino_t oldinum;
3298 	int error;
3299 
3300 	if ((vp = xp) == NULL &&
3301 	    (error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, LK_EXCLUSIVE,
3302 	     &vp)) != 0) {
3303 		softdep_error("handle_workitem_remove: vget", error);
3304 		return;
3305 	}
3306 	ip = VTOI(vp);
3307 	ACQUIRE_LOCK(&lk);
3308 	if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
3309 		FREE_LOCK(&lk);
3310 		panic("handle_workitem_remove: lost inodedep");
3311 	}
3312 	/*
3313 	 * Normal file deletion.
3314 	 */
3315 	if ((dirrem->dm_state & RMDIR) == 0) {
3316 		ip->i_nlink--;
3317 		DIP(ip, i_nlink) = ip->i_nlink;
3318 		ip->i_flag |= IN_CHANGE;
3319 		if (ip->i_nlink < ip->i_effnlink) {
3320 			FREE_LOCK(&lk);
3321 			panic("handle_workitem_remove: bad file delta");
3322 		}
3323 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3324 		FREE_LOCK(&lk);
3325 		vput(vp);
3326 		num_dirrem -= 1;
3327 		WORKITEM_FREE(dirrem, D_DIRREM);
3328 		return;
3329 	}
3330 	/*
3331 	 * Directory deletion. Decrement reference count for both the
3332 	 * just deleted parent directory entry and the reference for ".".
3333 	 * Next truncate the directory to length zero. When the
3334 	 * truncation completes, arrange to have the reference count on
3335 	 * the parent decremented to account for the loss of "..".
3336 	 */
3337 	ip->i_nlink -= 2;
3338 	DIP(ip, i_nlink) = ip->i_nlink;
3339 	ip->i_flag |= IN_CHANGE;
3340 	if (ip->i_nlink < ip->i_effnlink) {
3341 		FREE_LOCK(&lk);
3342 		panic("handle_workitem_remove: bad dir delta");
3343 	}
3344 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3345 	FREE_LOCK(&lk);
3346 	if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, td->td_ucred, td)) != 0)
3347 		softdep_error("handle_workitem_remove: truncate", error);
3348 	/*
3349 	 * Rename a directory to a new parent. Since, we are both deleting
3350 	 * and creating a new directory entry, the link count on the new
3351 	 * directory should not change. Thus we skip the followup dirrem.
3352 	 */
3353 	if (dirrem->dm_state & DIRCHG) {
3354 		vput(vp);
3355 		num_dirrem -= 1;
3356 		WORKITEM_FREE(dirrem, D_DIRREM);
3357 		return;
3358 	}
3359 	/*
3360 	 * If the inodedep does not exist, then the zero'ed inode has
3361 	 * been written to disk. If the allocated inode has never been
3362 	 * written to disk, then the on-disk inode is zero'ed. In either
3363 	 * case we can remove the file immediately.
3364 	 */
3365 	ACQUIRE_LOCK(&lk);
3366 	dirrem->dm_state = 0;
3367 	oldinum = dirrem->dm_oldinum;
3368 	dirrem->dm_oldinum = dirrem->dm_dirinum;
3369 	if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
3370 	    check_inode_unwritten(inodedep)) {
3371 		FREE_LOCK(&lk);
3372 		vput(vp);
3373 		handle_workitem_remove(dirrem, NULL);
3374 		return;
3375 	}
3376 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3377 	FREE_LOCK(&lk);
3378 	vput(vp);
3379 }
3380 
3381 /*
3382  * Inode de-allocation dependencies.
3383  *
3384  * When an inode's link count is reduced to zero, it can be de-allocated. We
3385  * found it convenient to postpone de-allocation until after the inode is
3386  * written to disk with its new link count (zero).  At this point, all of the
3387  * on-disk inode's block pointers are nullified and, with careful dependency
3388  * list ordering, all dependencies related to the inode will be satisfied and
3389  * the corresponding dependency structures de-allocated.  So, if/when the
3390  * inode is reused, there will be no mixing of old dependencies with new
3391  * ones.  This artificial dependency is set up by the block de-allocation
3392  * procedure above (softdep_setup_freeblocks) and completed by the
3393  * following procedure.
3394  */
3395 static void
3396 handle_workitem_freefile(freefile)
3397 	struct freefile *freefile;
3398 {
3399 	struct fs *fs;
3400 	struct inodedep *idp;
3401 	int error;
3402 
3403 	fs = VFSTOUFS(freefile->fx_mnt)->um_fs;
3404 #ifdef DEBUG
3405 	ACQUIRE_LOCK(&lk);
3406 	error = inodedep_lookup(fs, freefile->fx_oldinum, 0, &idp);
3407 	FREE_LOCK(&lk);
3408 	if (error)
3409 		panic("handle_workitem_freefile: inodedep survived");
3410 #endif
3411 	fs->fs_pendinginodes -= 1;
3412 	if ((error = ffs_freefile(fs, freefile->fx_devvp, freefile->fx_oldinum,
3413 	     freefile->fx_mode)) != 0)
3414 		softdep_error("handle_workitem_freefile", error);
3415 	WORKITEM_FREE(freefile, D_FREEFILE);
3416 }
3417 
3418 static int
3419 softdep_disk_prewrite(struct vnode *vp, struct buf *bp)
3420 {
3421 	int error;
3422 
3423 	if (bp->b_iocmd != BIO_WRITE)
3424 		return (0);
3425 	if ((bp->b_flags & B_VALIDSUSPWRT) == 0 &&
3426 	    bp->b_vp != NULL && bp->b_vp->v_mount != NULL &&
3427 	    (bp->b_vp->v_mount->mnt_kern_flag & MNTK_SUSPENDED) != 0)
3428 		panic("softdep_disk_prewrite: bad I/O");
3429 	bp->b_flags &= ~B_VALIDSUSPWRT;
3430 	if (LIST_FIRST(&bp->b_dep) != NULL)
3431 		buf_start(bp);
3432 	mp_fixme("This should require the vnode lock.");
3433 	if ((vp->v_vflag & VV_COPYONWRITE) &&
3434 	    vp->v_rdev->si_copyonwrite &&
3435 	    (error = (*vp->v_rdev->si_copyonwrite)(vp, bp)) != 0 &&
3436 	    error != EOPNOTSUPP) {
3437 		bp->b_io.bio_error = error;
3438 		bp->b_io.bio_flags |= BIO_ERROR;
3439 		biodone(&bp->b_io);
3440 		return (1);
3441 	}
3442 	return (0);
3443 }
3444 
3445 
3446 /*
3447  * Disk writes.
3448  *
3449  * The dependency structures constructed above are most actively used when file
3450  * system blocks are written to disk.  No constraints are placed on when a
3451  * block can be written, but unsatisfied update dependencies are made safe by
3452  * modifying (or replacing) the source memory for the duration of the disk
3453  * write.  When the disk write completes, the memory block is again brought
3454  * up-to-date.
3455  *
3456  * In-core inode structure reclamation.
3457  *
3458  * Because there are a finite number of "in-core" inode structures, they are
3459  * reused regularly.  By transferring all inode-related dependencies to the
3460  * in-memory inode block and indexing them separately (via "inodedep"s), we
3461  * can allow "in-core" inode structures to be reused at any time and avoid
3462  * any increase in contention.
3463  *
3464  * Called just before entering the device driver to initiate a new disk I/O.
3465  * The buffer must be locked, thus, no I/O completion operations can occur
3466  * while we are manipulating its associated dependencies.
3467  */
3468 static void
3469 softdep_disk_io_initiation(bp)
3470 	struct buf *bp;		/* structure describing disk write to occur */
3471 {
3472 	struct worklist *wk, *nextwk;
3473 	struct indirdep *indirdep;
3474 	struct inodedep *inodedep;
3475 
3476 	/*
3477 	 * We only care about write operations. There should never
3478 	 * be dependencies for reads.
3479 	 */
3480 	if (bp->b_iocmd == BIO_READ)
3481 		panic("softdep_disk_io_initiation: read");
3482 	/*
3483 	 * Do any necessary pre-I/O processing.
3484 	 */
3485 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
3486 		nextwk = LIST_NEXT(wk, wk_list);
3487 		switch (wk->wk_type) {
3488 
3489 		case D_PAGEDEP:
3490 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
3491 			continue;
3492 
3493 		case D_INODEDEP:
3494 			inodedep = WK_INODEDEP(wk);
3495 			if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
3496 				initiate_write_inodeblock_ufs1(inodedep, bp);
3497 			else
3498 				initiate_write_inodeblock_ufs2(inodedep, bp);
3499 			continue;
3500 
3501 		case D_INDIRDEP:
3502 			indirdep = WK_INDIRDEP(wk);
3503 			if (indirdep->ir_state & GOINGAWAY)
3504 				panic("disk_io_initiation: indirdep gone");
3505 			/*
3506 			 * If there are no remaining dependencies, this
3507 			 * will be writing the real pointers, so the
3508 			 * dependency can be freed.
3509 			 */
3510 			if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
3511 				indirdep->ir_savebp->b_flags |=
3512 				    B_INVAL | B_NOCACHE;
3513 				brelse(indirdep->ir_savebp);
3514 				/* inline expand WORKLIST_REMOVE(wk); */
3515 				wk->wk_state &= ~ONWORKLIST;
3516 				LIST_REMOVE(wk, wk_list);
3517 				WORKITEM_FREE(indirdep, D_INDIRDEP);
3518 				continue;
3519 			}
3520 			/*
3521 			 * Replace up-to-date version with safe version.
3522 			 */
3523 			MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3524 			    M_INDIRDEP, M_SOFTDEP_FLAGS);
3525 			ACQUIRE_LOCK(&lk);
3526 			indirdep->ir_state &= ~ATTACHED;
3527 			indirdep->ir_state |= UNDONE;
3528 			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3529 			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3530 			    bp->b_bcount);
3531 			FREE_LOCK(&lk);
3532 			continue;
3533 
3534 		case D_MKDIR:
3535 		case D_BMSAFEMAP:
3536 		case D_ALLOCDIRECT:
3537 		case D_ALLOCINDIR:
3538 			continue;
3539 
3540 		default:
3541 			panic("handle_disk_io_initiation: Unexpected type %s",
3542 			    TYPENAME(wk->wk_type));
3543 			/* NOTREACHED */
3544 		}
3545 	}
3546 }
3547 
3548 /*
3549  * Called from within the procedure above to deal with unsatisfied
3550  * allocation dependencies in a directory. The buffer must be locked,
3551  * thus, no I/O completion operations can occur while we are
3552  * manipulating its associated dependencies.
3553  */
3554 static void
3555 initiate_write_filepage(pagedep, bp)
3556 	struct pagedep *pagedep;
3557 	struct buf *bp;
3558 {
3559 	struct diradd *dap;
3560 	struct direct *ep;
3561 	int i;
3562 
3563 	if (pagedep->pd_state & IOSTARTED) {
3564 		/*
3565 		 * This can only happen if there is a driver that does not
3566 		 * understand chaining. Here biodone will reissue the call
3567 		 * to strategy for the incomplete buffers.
3568 		 */
3569 		printf("initiate_write_filepage: already started\n");
3570 		return;
3571 	}
3572 	pagedep->pd_state |= IOSTARTED;
3573 	ACQUIRE_LOCK(&lk);
3574 	for (i = 0; i < DAHASHSZ; i++) {
3575 		LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3576 			ep = (struct direct *)
3577 			    ((char *)bp->b_data + dap->da_offset);
3578 			if (ep->d_ino != dap->da_newinum) {
3579 				FREE_LOCK(&lk);
3580 				panic("%s: dir inum %d != new %d",
3581 				    "initiate_write_filepage",
3582 				    ep->d_ino, dap->da_newinum);
3583 			}
3584 			if (dap->da_state & DIRCHG)
3585 				ep->d_ino = dap->da_previous->dm_oldinum;
3586 			else
3587 				ep->d_ino = 0;
3588 			dap->da_state &= ~ATTACHED;
3589 			dap->da_state |= UNDONE;
3590 		}
3591 	}
3592 	FREE_LOCK(&lk);
3593 }
3594 
3595 /*
3596  * Version of initiate_write_inodeblock that handles UFS1 dinodes.
3597  * Note that any bug fixes made to this routine must be done in the
3598  * version found below.
3599  *
3600  * Called from within the procedure above to deal with unsatisfied
3601  * allocation dependencies in an inodeblock. The buffer must be
3602  * locked, thus, no I/O completion operations can occur while we
3603  * are manipulating its associated dependencies.
3604  */
3605 static void
3606 initiate_write_inodeblock_ufs1(inodedep, bp)
3607 	struct inodedep *inodedep;
3608 	struct buf *bp;			/* The inode block */
3609 {
3610 	struct allocdirect *adp, *lastadp;
3611 	struct ufs1_dinode *dp;
3612 	struct fs *fs;
3613 	ufs_lbn_t i, prevlbn = 0;
3614 	int deplist;
3615 
3616 	if (inodedep->id_state & IOSTARTED)
3617 		panic("initiate_write_inodeblock_ufs1: already started");
3618 	inodedep->id_state |= IOSTARTED;
3619 	fs = inodedep->id_fs;
3620 	dp = (struct ufs1_dinode *)bp->b_data +
3621 	    ino_to_fsbo(fs, inodedep->id_ino);
3622 	/*
3623 	 * If the bitmap is not yet written, then the allocated
3624 	 * inode cannot be written to disk.
3625 	 */
3626 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3627 		if (inodedep->id_savedino1 != NULL)
3628 			panic("initiate_write_inodeblock_ufs1: I/O underway");
3629 		MALLOC(inodedep->id_savedino1, struct ufs1_dinode *,
3630 		    sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3631 		*inodedep->id_savedino1 = *dp;
3632 		bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
3633 		return;
3634 	}
3635 	/*
3636 	 * If no dependencies, then there is nothing to roll back.
3637 	 */
3638 	inodedep->id_savedsize = dp->di_size;
3639 	inodedep->id_savedextsize = 0;
3640 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3641 		return;
3642 	/*
3643 	 * Set the dependencies to busy.
3644 	 */
3645 	ACQUIRE_LOCK(&lk);
3646 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3647 	     adp = TAILQ_NEXT(adp, ad_next)) {
3648 #ifdef DIAGNOSTIC
3649 		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3650 			FREE_LOCK(&lk);
3651 			panic("softdep_write_inodeblock: lbn order");
3652 		}
3653 		prevlbn = adp->ad_lbn;
3654 		if (adp->ad_lbn < NDADDR &&
3655 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3656 			FREE_LOCK(&lk);
3657 			panic("%s: direct pointer #%jd mismatch %d != %jd",
3658 			    "softdep_write_inodeblock",
3659 			    (intmax_t)adp->ad_lbn,
3660 			    dp->di_db[adp->ad_lbn],
3661 			    (intmax_t)adp->ad_newblkno);
3662 		}
3663 		if (adp->ad_lbn >= NDADDR &&
3664 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3665 			FREE_LOCK(&lk);
3666 			panic("%s: indirect pointer #%jd mismatch %d != %jd",
3667 			    "softdep_write_inodeblock",
3668 			    (intmax_t)adp->ad_lbn - NDADDR,
3669 			    dp->di_ib[adp->ad_lbn - NDADDR],
3670 			    (intmax_t)adp->ad_newblkno);
3671 		}
3672 		deplist |= 1 << adp->ad_lbn;
3673 		if ((adp->ad_state & ATTACHED) == 0) {
3674 			FREE_LOCK(&lk);
3675 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3676 			    adp->ad_state);
3677 		}
3678 #endif /* DIAGNOSTIC */
3679 		adp->ad_state &= ~ATTACHED;
3680 		adp->ad_state |= UNDONE;
3681 	}
3682 	/*
3683 	 * The on-disk inode cannot claim to be any larger than the last
3684 	 * fragment that has been written. Otherwise, the on-disk inode
3685 	 * might have fragments that were not the last block in the file
3686 	 * which would corrupt the filesystem.
3687 	 */
3688 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3689 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3690 		if (adp->ad_lbn >= NDADDR)
3691 			break;
3692 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3693 		/* keep going until hitting a rollback to a frag */
3694 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3695 			continue;
3696 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3697 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3698 #ifdef DIAGNOSTIC
3699 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3700 				FREE_LOCK(&lk);
3701 				panic("softdep_write_inodeblock: lost dep1");
3702 			}
3703 #endif /* DIAGNOSTIC */
3704 			dp->di_db[i] = 0;
3705 		}
3706 		for (i = 0; i < NIADDR; i++) {
3707 #ifdef DIAGNOSTIC
3708 			if (dp->di_ib[i] != 0 &&
3709 			    (deplist & ((1 << NDADDR) << i)) == 0) {
3710 				FREE_LOCK(&lk);
3711 				panic("softdep_write_inodeblock: lost dep2");
3712 			}
3713 #endif /* DIAGNOSTIC */
3714 			dp->di_ib[i] = 0;
3715 		}
3716 		FREE_LOCK(&lk);
3717 		return;
3718 	}
3719 	/*
3720 	 * If we have zero'ed out the last allocated block of the file,
3721 	 * roll back the size to the last currently allocated block.
3722 	 * We know that this last allocated block is a full-sized as
3723 	 * we already checked for fragments in the loop above.
3724 	 */
3725 	if (lastadp != NULL &&
3726 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3727 		for (i = lastadp->ad_lbn; i >= 0; i--)
3728 			if (dp->di_db[i] != 0)
3729 				break;
3730 		dp->di_size = (i + 1) * fs->fs_bsize;
3731 	}
3732 	/*
3733 	 * The only dependencies are for indirect blocks.
3734 	 *
3735 	 * The file size for indirect block additions is not guaranteed.
3736 	 * Such a guarantee would be non-trivial to achieve. The conventional
3737 	 * synchronous write implementation also does not make this guarantee.
3738 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3739 	 * can be over-estimated without destroying integrity when the file
3740 	 * moves into the indirect blocks (i.e., is large). If we want to
3741 	 * postpone fsck, we are stuck with this argument.
3742 	 */
3743 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3744 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3745 	FREE_LOCK(&lk);
3746 }
3747 
3748 /*
3749  * Version of initiate_write_inodeblock that handles UFS2 dinodes.
3750  * Note that any bug fixes made to this routine must be done in the
3751  * version found above.
3752  *
3753  * Called from within the procedure above to deal with unsatisfied
3754  * allocation dependencies in an inodeblock. The buffer must be
3755  * locked, thus, no I/O completion operations can occur while we
3756  * are manipulating its associated dependencies.
3757  */
3758 static void
3759 initiate_write_inodeblock_ufs2(inodedep, bp)
3760 	struct inodedep *inodedep;
3761 	struct buf *bp;			/* The inode block */
3762 {
3763 	struct allocdirect *adp, *lastadp;
3764 	struct ufs2_dinode *dp;
3765 	struct fs *fs;
3766 	ufs_lbn_t i, prevlbn = 0;
3767 	int deplist;
3768 
3769 	if (inodedep->id_state & IOSTARTED)
3770 		panic("initiate_write_inodeblock_ufs2: already started");
3771 	inodedep->id_state |= IOSTARTED;
3772 	fs = inodedep->id_fs;
3773 	dp = (struct ufs2_dinode *)bp->b_data +
3774 	    ino_to_fsbo(fs, inodedep->id_ino);
3775 	/*
3776 	 * If the bitmap is not yet written, then the allocated
3777 	 * inode cannot be written to disk.
3778 	 */
3779 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3780 		if (inodedep->id_savedino2 != NULL)
3781 			panic("initiate_write_inodeblock_ufs2: I/O underway");
3782 		MALLOC(inodedep->id_savedino2, struct ufs2_dinode *,
3783 		    sizeof(struct ufs2_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3784 		*inodedep->id_savedino2 = *dp;
3785 		bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
3786 		return;
3787 	}
3788 	/*
3789 	 * If no dependencies, then there is nothing to roll back.
3790 	 */
3791 	inodedep->id_savedsize = dp->di_size;
3792 	inodedep->id_savedextsize = dp->di_extsize;
3793 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL &&
3794 	    TAILQ_FIRST(&inodedep->id_extupdt) == NULL)
3795 		return;
3796 	/*
3797 	 * Set the ext data dependencies to busy.
3798 	 */
3799 	ACQUIRE_LOCK(&lk);
3800 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
3801 	     adp = TAILQ_NEXT(adp, ad_next)) {
3802 #ifdef DIAGNOSTIC
3803 		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3804 			FREE_LOCK(&lk);
3805 			panic("softdep_write_inodeblock: lbn order");
3806 		}
3807 		prevlbn = adp->ad_lbn;
3808 		if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno) {
3809 			FREE_LOCK(&lk);
3810 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
3811 			    "softdep_write_inodeblock",
3812 			    (intmax_t)adp->ad_lbn,
3813 			    (intmax_t)dp->di_extb[adp->ad_lbn],
3814 			    (intmax_t)adp->ad_newblkno);
3815 		}
3816 		deplist |= 1 << adp->ad_lbn;
3817 		if ((adp->ad_state & ATTACHED) == 0) {
3818 			FREE_LOCK(&lk);
3819 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3820 			    adp->ad_state);
3821 		}
3822 #endif /* DIAGNOSTIC */
3823 		adp->ad_state &= ~ATTACHED;
3824 		adp->ad_state |= UNDONE;
3825 	}
3826 	/*
3827 	 * The on-disk inode cannot claim to be any larger than the last
3828 	 * fragment that has been written. Otherwise, the on-disk inode
3829 	 * might have fragments that were not the last block in the ext
3830 	 * data which would corrupt the filesystem.
3831 	 */
3832 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
3833 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3834 		dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno;
3835 		/* keep going until hitting a rollback to a frag */
3836 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3837 			continue;
3838 		dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3839 		for (i = adp->ad_lbn + 1; i < NXADDR; i++) {
3840 #ifdef DIAGNOSTIC
3841 			if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) {
3842 				FREE_LOCK(&lk);
3843 				panic("softdep_write_inodeblock: lost dep1");
3844 			}
3845 #endif /* DIAGNOSTIC */
3846 			dp->di_extb[i] = 0;
3847 		}
3848 		lastadp = NULL;
3849 		break;
3850 	}
3851 	/*
3852 	 * If we have zero'ed out the last allocated block of the ext
3853 	 * data, roll back the size to the last currently allocated block.
3854 	 * We know that this last allocated block is a full-sized as
3855 	 * we already checked for fragments in the loop above.
3856 	 */
3857 	if (lastadp != NULL &&
3858 	    dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3859 		for (i = lastadp->ad_lbn; i >= 0; i--)
3860 			if (dp->di_extb[i] != 0)
3861 				break;
3862 		dp->di_extsize = (i + 1) * fs->fs_bsize;
3863 	}
3864 	/*
3865 	 * Set the file data dependencies to busy.
3866 	 */
3867 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3868 	     adp = TAILQ_NEXT(adp, ad_next)) {
3869 #ifdef DIAGNOSTIC
3870 		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3871 			FREE_LOCK(&lk);
3872 			panic("softdep_write_inodeblock: lbn order");
3873 		}
3874 		prevlbn = adp->ad_lbn;
3875 		if (adp->ad_lbn < NDADDR &&
3876 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3877 			FREE_LOCK(&lk);
3878 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
3879 			    "softdep_write_inodeblock",
3880 			    (intmax_t)adp->ad_lbn,
3881 			    (intmax_t)dp->di_db[adp->ad_lbn],
3882 			    (intmax_t)adp->ad_newblkno);
3883 		}
3884 		if (adp->ad_lbn >= NDADDR &&
3885 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3886 			FREE_LOCK(&lk);
3887 			panic("%s indirect pointer #%jd mismatch %jd != %jd",
3888 			    "softdep_write_inodeblock:",
3889 			    (intmax_t)adp->ad_lbn - NDADDR,
3890 			    (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR],
3891 			    (intmax_t)adp->ad_newblkno);
3892 		}
3893 		deplist |= 1 << adp->ad_lbn;
3894 		if ((adp->ad_state & ATTACHED) == 0) {
3895 			FREE_LOCK(&lk);
3896 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3897 			    adp->ad_state);
3898 		}
3899 #endif /* DIAGNOSTIC */
3900 		adp->ad_state &= ~ATTACHED;
3901 		adp->ad_state |= UNDONE;
3902 	}
3903 	/*
3904 	 * The on-disk inode cannot claim to be any larger than the last
3905 	 * fragment that has been written. Otherwise, the on-disk inode
3906 	 * might have fragments that were not the last block in the file
3907 	 * which would corrupt the filesystem.
3908 	 */
3909 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3910 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3911 		if (adp->ad_lbn >= NDADDR)
3912 			break;
3913 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3914 		/* keep going until hitting a rollback to a frag */
3915 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3916 			continue;
3917 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3918 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3919 #ifdef DIAGNOSTIC
3920 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3921 				FREE_LOCK(&lk);
3922 				panic("softdep_write_inodeblock: lost dep2");
3923 			}
3924 #endif /* DIAGNOSTIC */
3925 			dp->di_db[i] = 0;
3926 		}
3927 		for (i = 0; i < NIADDR; i++) {
3928 #ifdef DIAGNOSTIC
3929 			if (dp->di_ib[i] != 0 &&
3930 			    (deplist & ((1 << NDADDR) << i)) == 0) {
3931 				FREE_LOCK(&lk);
3932 				panic("softdep_write_inodeblock: lost dep3");
3933 			}
3934 #endif /* DIAGNOSTIC */
3935 			dp->di_ib[i] = 0;
3936 		}
3937 		FREE_LOCK(&lk);
3938 		return;
3939 	}
3940 	/*
3941 	 * If we have zero'ed out the last allocated block of the file,
3942 	 * roll back the size to the last currently allocated block.
3943 	 * We know that this last allocated block is a full-sized as
3944 	 * we already checked for fragments in the loop above.
3945 	 */
3946 	if (lastadp != NULL &&
3947 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3948 		for (i = lastadp->ad_lbn; i >= 0; i--)
3949 			if (dp->di_db[i] != 0)
3950 				break;
3951 		dp->di_size = (i + 1) * fs->fs_bsize;
3952 	}
3953 	/*
3954 	 * The only dependencies are for indirect blocks.
3955 	 *
3956 	 * The file size for indirect block additions is not guaranteed.
3957 	 * Such a guarantee would be non-trivial to achieve. The conventional
3958 	 * synchronous write implementation also does not make this guarantee.
3959 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3960 	 * can be over-estimated without destroying integrity when the file
3961 	 * moves into the indirect blocks (i.e., is large). If we want to
3962 	 * postpone fsck, we are stuck with this argument.
3963 	 */
3964 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3965 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3966 	FREE_LOCK(&lk);
3967 }
3968 
3969 /*
3970  * This routine is called during the completion interrupt
3971  * service routine for a disk write (from the procedure called
3972  * by the device driver to inform the filesystem caches of
3973  * a request completion).  It should be called early in this
3974  * procedure, before the block is made available to other
3975  * processes or other routines are called.
3976  */
3977 static void
3978 softdep_disk_write_complete(bp)
3979 	struct buf *bp;		/* describes the completed disk write */
3980 {
3981 	struct worklist *wk;
3982 	struct workhead reattach;
3983 	struct newblk *newblk;
3984 	struct allocindir *aip;
3985 	struct allocdirect *adp;
3986 	struct indirdep *indirdep;
3987 	struct inodedep *inodedep;
3988 	struct bmsafemap *bmsafemap;
3989 
3990 	/*
3991 	 * If an error occurred while doing the write, then the data
3992 	 * has not hit the disk and the dependencies cannot be unrolled.
3993 	 */
3994 	if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0)
3995 		return;
3996 #ifdef DEBUG
3997 	if (lk.lkt_held != NOHOLDER)
3998 		panic("softdep_disk_write_complete: lock is held");
3999 	lk.lkt_held = SPECIAL_FLAG;
4000 #endif
4001 	LIST_INIT(&reattach);
4002 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
4003 		WORKLIST_REMOVE(wk);
4004 		switch (wk->wk_type) {
4005 
4006 		case D_PAGEDEP:
4007 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
4008 				WORKLIST_INSERT(&reattach, wk);
4009 			continue;
4010 
4011 		case D_INODEDEP:
4012 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
4013 				WORKLIST_INSERT(&reattach, wk);
4014 			continue;
4015 
4016 		case D_BMSAFEMAP:
4017 			bmsafemap = WK_BMSAFEMAP(wk);
4018 			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
4019 				newblk->nb_state |= DEPCOMPLETE;
4020 				newblk->nb_bmsafemap = NULL;
4021 				LIST_REMOVE(newblk, nb_deps);
4022 			}
4023 			while ((adp =
4024 			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
4025 				adp->ad_state |= DEPCOMPLETE;
4026 				adp->ad_buf = NULL;
4027 				LIST_REMOVE(adp, ad_deps);
4028 				handle_allocdirect_partdone(adp);
4029 			}
4030 			while ((aip =
4031 			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
4032 				aip->ai_state |= DEPCOMPLETE;
4033 				aip->ai_buf = NULL;
4034 				LIST_REMOVE(aip, ai_deps);
4035 				handle_allocindir_partdone(aip);
4036 			}
4037 			while ((inodedep =
4038 			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
4039 				inodedep->id_state |= DEPCOMPLETE;
4040 				LIST_REMOVE(inodedep, id_deps);
4041 				inodedep->id_buf = NULL;
4042 			}
4043 			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
4044 			continue;
4045 
4046 		case D_MKDIR:
4047 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
4048 			continue;
4049 
4050 		case D_ALLOCDIRECT:
4051 			adp = WK_ALLOCDIRECT(wk);
4052 			adp->ad_state |= COMPLETE;
4053 			handle_allocdirect_partdone(adp);
4054 			continue;
4055 
4056 		case D_ALLOCINDIR:
4057 			aip = WK_ALLOCINDIR(wk);
4058 			aip->ai_state |= COMPLETE;
4059 			handle_allocindir_partdone(aip);
4060 			continue;
4061 
4062 		case D_INDIRDEP:
4063 			indirdep = WK_INDIRDEP(wk);
4064 			if (indirdep->ir_state & GOINGAWAY) {
4065 				lk.lkt_held = NOHOLDER;
4066 				panic("disk_write_complete: indirdep gone");
4067 			}
4068 			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
4069 			FREE(indirdep->ir_saveddata, M_INDIRDEP);
4070 			indirdep->ir_saveddata = 0;
4071 			indirdep->ir_state &= ~UNDONE;
4072 			indirdep->ir_state |= ATTACHED;
4073 			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
4074 				handle_allocindir_partdone(aip);
4075 				if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
4076 					lk.lkt_held = NOHOLDER;
4077 					panic("disk_write_complete: not gone");
4078 				}
4079 			}
4080 			WORKLIST_INSERT(&reattach, wk);
4081 			if ((bp->b_flags & B_DELWRI) == 0)
4082 				stat_indir_blk_ptrs++;
4083 			bdirty(bp);
4084 			continue;
4085 
4086 		default:
4087 			lk.lkt_held = NOHOLDER;
4088 			panic("handle_disk_write_complete: Unknown type %s",
4089 			    TYPENAME(wk->wk_type));
4090 			/* NOTREACHED */
4091 		}
4092 	}
4093 	/*
4094 	 * Reattach any requests that must be redone.
4095 	 */
4096 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
4097 		WORKLIST_REMOVE(wk);
4098 		WORKLIST_INSERT(&bp->b_dep, wk);
4099 	}
4100 #ifdef DEBUG
4101 	if (lk.lkt_held != SPECIAL_FLAG)
4102 		panic("softdep_disk_write_complete: lock lost");
4103 	lk.lkt_held = NOHOLDER;
4104 #endif
4105 }
4106 
4107 /*
4108  * Called from within softdep_disk_write_complete above. Note that
4109  * this routine is always called from interrupt level with further
4110  * splbio interrupts blocked.
4111  */
4112 static void
4113 handle_allocdirect_partdone(adp)
4114 	struct allocdirect *adp;	/* the completed allocdirect */
4115 {
4116 	struct allocdirectlst *listhead;
4117 	struct allocdirect *listadp;
4118 	struct inodedep *inodedep;
4119 	long bsize, delay;
4120 
4121 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4122 		return;
4123 	if (adp->ad_buf != NULL) {
4124 		lk.lkt_held = NOHOLDER;
4125 		panic("handle_allocdirect_partdone: dangling dep");
4126 	}
4127 	/*
4128 	 * The on-disk inode cannot claim to be any larger than the last
4129 	 * fragment that has been written. Otherwise, the on-disk inode
4130 	 * might have fragments that were not the last block in the file
4131 	 * which would corrupt the filesystem. Thus, we cannot free any
4132 	 * allocdirects after one whose ad_oldblkno claims a fragment as
4133 	 * these blocks must be rolled back to zero before writing the inode.
4134 	 * We check the currently active set of allocdirects in id_inoupdt
4135 	 * or id_extupdt as appropriate.
4136 	 */
4137 	inodedep = adp->ad_inodedep;
4138 	bsize = inodedep->id_fs->fs_bsize;
4139 	if (adp->ad_state & EXTDATA)
4140 		listhead = &inodedep->id_extupdt;
4141 	else
4142 		listhead = &inodedep->id_inoupdt;
4143 	TAILQ_FOREACH(listadp, listhead, ad_next) {
4144 		/* found our block */
4145 		if (listadp == adp)
4146 			break;
4147 		/* continue if ad_oldlbn is not a fragment */
4148 		if (listadp->ad_oldsize == 0 ||
4149 		    listadp->ad_oldsize == bsize)
4150 			continue;
4151 		/* hit a fragment */
4152 		return;
4153 	}
4154 	/*
4155 	 * If we have reached the end of the current list without
4156 	 * finding the just finished dependency, then it must be
4157 	 * on the future dependency list. Future dependencies cannot
4158 	 * be freed until they are moved to the current list.
4159 	 */
4160 	if (listadp == NULL) {
4161 #ifdef DEBUG
4162 		if (adp->ad_state & EXTDATA)
4163 			listhead = &inodedep->id_newextupdt;
4164 		else
4165 			listhead = &inodedep->id_newinoupdt;
4166 		TAILQ_FOREACH(listadp, listhead, ad_next)
4167 			/* found our block */
4168 			if (listadp == adp)
4169 				break;
4170 		if (listadp == NULL) {
4171 			lk.lkt_held = NOHOLDER;
4172 			panic("handle_allocdirect_partdone: lost dep");
4173 		}
4174 #endif /* DEBUG */
4175 		return;
4176 	}
4177 	/*
4178 	 * If we have found the just finished dependency, then free
4179 	 * it along with anything that follows it that is complete.
4180 	 * If the inode still has a bitmap dependency, then it has
4181 	 * never been written to disk, hence the on-disk inode cannot
4182 	 * reference the old fragment so we can free it without delay.
4183 	 */
4184 	delay = (inodedep->id_state & DEPCOMPLETE);
4185 	for (; adp; adp = listadp) {
4186 		listadp = TAILQ_NEXT(adp, ad_next);
4187 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4188 			return;
4189 		free_allocdirect(listhead, adp, delay);
4190 	}
4191 }
4192 
4193 /*
4194  * Called from within softdep_disk_write_complete above. Note that
4195  * this routine is always called from interrupt level with further
4196  * splbio interrupts blocked.
4197  */
4198 static void
4199 handle_allocindir_partdone(aip)
4200 	struct allocindir *aip;		/* the completed allocindir */
4201 {
4202 	struct indirdep *indirdep;
4203 
4204 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
4205 		return;
4206 	if (aip->ai_buf != NULL) {
4207 		lk.lkt_held = NOHOLDER;
4208 		panic("handle_allocindir_partdone: dangling dependency");
4209 	}
4210 	indirdep = aip->ai_indirdep;
4211 	if (indirdep->ir_state & UNDONE) {
4212 		LIST_REMOVE(aip, ai_next);
4213 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
4214 		return;
4215 	}
4216 	if (indirdep->ir_state & UFS1FMT)
4217 		((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4218 		    aip->ai_newblkno;
4219 	else
4220 		((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4221 		    aip->ai_newblkno;
4222 	LIST_REMOVE(aip, ai_next);
4223 	if (aip->ai_freefrag != NULL)
4224 		add_to_worklist(&aip->ai_freefrag->ff_list);
4225 	WORKITEM_FREE(aip, D_ALLOCINDIR);
4226 }
4227 
4228 /*
4229  * Called from within softdep_disk_write_complete above to restore
4230  * in-memory inode block contents to their most up-to-date state. Note
4231  * that this routine is always called from interrupt level with further
4232  * splbio interrupts blocked.
4233  */
4234 static int
4235 handle_written_inodeblock(inodedep, bp)
4236 	struct inodedep *inodedep;
4237 	struct buf *bp;		/* buffer containing the inode block */
4238 {
4239 	struct worklist *wk, *filefree;
4240 	struct allocdirect *adp, *nextadp;
4241 	struct ufs1_dinode *dp1 = NULL;
4242 	struct ufs2_dinode *dp2 = NULL;
4243 	int hadchanges, fstype;
4244 
4245 	if ((inodedep->id_state & IOSTARTED) == 0) {
4246 		lk.lkt_held = NOHOLDER;
4247 		panic("handle_written_inodeblock: not started");
4248 	}
4249 	inodedep->id_state &= ~IOSTARTED;
4250 	inodedep->id_state |= COMPLETE;
4251 	if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
4252 		fstype = UFS1;
4253 		dp1 = (struct ufs1_dinode *)bp->b_data +
4254 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4255 	} else {
4256 		fstype = UFS2;
4257 		dp2 = (struct ufs2_dinode *)bp->b_data +
4258 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4259 	}
4260 	/*
4261 	 * If we had to rollback the inode allocation because of
4262 	 * bitmaps being incomplete, then simply restore it.
4263 	 * Keep the block dirty so that it will not be reclaimed until
4264 	 * all associated dependencies have been cleared and the
4265 	 * corresponding updates written to disk.
4266 	 */
4267 	if (inodedep->id_savedino1 != NULL) {
4268 		if (fstype == UFS1)
4269 			*dp1 = *inodedep->id_savedino1;
4270 		else
4271 			*dp2 = *inodedep->id_savedino2;
4272 		FREE(inodedep->id_savedino1, M_INODEDEP);
4273 		inodedep->id_savedino1 = NULL;
4274 		if ((bp->b_flags & B_DELWRI) == 0)
4275 			stat_inode_bitmap++;
4276 		bdirty(bp);
4277 		return (1);
4278 	}
4279 	/*
4280 	 * Roll forward anything that had to be rolled back before
4281 	 * the inode could be updated.
4282 	 */
4283 	hadchanges = 0;
4284 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
4285 		nextadp = TAILQ_NEXT(adp, ad_next);
4286 		if (adp->ad_state & ATTACHED) {
4287 			lk.lkt_held = NOHOLDER;
4288 			panic("handle_written_inodeblock: new entry");
4289 		}
4290 		if (fstype == UFS1) {
4291 			if (adp->ad_lbn < NDADDR) {
4292 				if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno) {
4293 					lk.lkt_held = NOHOLDER;
4294 					panic("%s %s #%jd mismatch %d != %jd",
4295 					    "handle_written_inodeblock:",
4296 					    "direct pointer",
4297 					    (intmax_t)adp->ad_lbn,
4298 					    dp1->di_db[adp->ad_lbn],
4299 					    (intmax_t)adp->ad_oldblkno);
4300 				}
4301 				dp1->di_db[adp->ad_lbn] = adp->ad_newblkno;
4302 			} else {
4303 				if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0) {
4304 					lk.lkt_held = NOHOLDER;
4305 					panic("%s: %s #%jd allocated as %d",
4306 					    "handle_written_inodeblock",
4307 					    "indirect pointer",
4308 					    (intmax_t)adp->ad_lbn - NDADDR,
4309 					    dp1->di_ib[adp->ad_lbn - NDADDR]);
4310 				}
4311 				dp1->di_ib[adp->ad_lbn - NDADDR] =
4312 				    adp->ad_newblkno;
4313 			}
4314 		} else {
4315 			if (adp->ad_lbn < NDADDR) {
4316 				if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno) {
4317 					lk.lkt_held = NOHOLDER;
4318 					panic("%s: %s #%jd %s %jd != %jd",
4319 					    "handle_written_inodeblock",
4320 					    "direct pointer",
4321 					    (intmax_t)adp->ad_lbn, "mismatch",
4322 					    (intmax_t)dp2->di_db[adp->ad_lbn],
4323 					    (intmax_t)adp->ad_oldblkno);
4324 				}
4325 				dp2->di_db[adp->ad_lbn] = adp->ad_newblkno;
4326 			} else {
4327 				if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0) {
4328 					lk.lkt_held = NOHOLDER;
4329 					panic("%s: %s #%jd allocated as %jd",
4330 					    "handle_written_inodeblock",
4331 					    "indirect pointer",
4332 					    (intmax_t)adp->ad_lbn - NDADDR,
4333 					    (intmax_t)
4334 					    dp2->di_ib[adp->ad_lbn - NDADDR]);
4335 				}
4336 				dp2->di_ib[adp->ad_lbn - NDADDR] =
4337 				    adp->ad_newblkno;
4338 			}
4339 		}
4340 		adp->ad_state &= ~UNDONE;
4341 		adp->ad_state |= ATTACHED;
4342 		hadchanges = 1;
4343 	}
4344 	for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
4345 		nextadp = TAILQ_NEXT(adp, ad_next);
4346 		if (adp->ad_state & ATTACHED) {
4347 			lk.lkt_held = NOHOLDER;
4348 			panic("handle_written_inodeblock: new entry");
4349 		}
4350 		if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno) {
4351 			lk.lkt_held = NOHOLDER;
4352 			panic("%s: direct pointers #%jd %s %jd != %jd",
4353 			    "handle_written_inodeblock",
4354 			    (intmax_t)adp->ad_lbn, "mismatch",
4355 			    (intmax_t)dp2->di_extb[adp->ad_lbn],
4356 			    (intmax_t)adp->ad_oldblkno);
4357 		}
4358 		dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno;
4359 		adp->ad_state &= ~UNDONE;
4360 		adp->ad_state |= ATTACHED;
4361 		hadchanges = 1;
4362 	}
4363 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
4364 		stat_direct_blk_ptrs++;
4365 	/*
4366 	 * Reset the file size to its most up-to-date value.
4367 	 */
4368 	if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) {
4369 		lk.lkt_held = NOHOLDER;
4370 		panic("handle_written_inodeblock: bad size");
4371 	}
4372 	if (fstype == UFS1) {
4373 		if (dp1->di_size != inodedep->id_savedsize) {
4374 			dp1->di_size = inodedep->id_savedsize;
4375 			hadchanges = 1;
4376 		}
4377 	} else {
4378 		if (dp2->di_size != inodedep->id_savedsize) {
4379 			dp2->di_size = inodedep->id_savedsize;
4380 			hadchanges = 1;
4381 		}
4382 		if (dp2->di_extsize != inodedep->id_savedextsize) {
4383 			dp2->di_extsize = inodedep->id_savedextsize;
4384 			hadchanges = 1;
4385 		}
4386 	}
4387 	inodedep->id_savedsize = -1;
4388 	inodedep->id_savedextsize = -1;
4389 	/*
4390 	 * If there were any rollbacks in the inode block, then it must be
4391 	 * marked dirty so that its will eventually get written back in
4392 	 * its correct form.
4393 	 */
4394 	if (hadchanges)
4395 		bdirty(bp);
4396 	/*
4397 	 * Process any allocdirects that completed during the update.
4398 	 */
4399 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
4400 		handle_allocdirect_partdone(adp);
4401 	if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
4402 		handle_allocdirect_partdone(adp);
4403 	/*
4404 	 * Process deallocations that were held pending until the
4405 	 * inode had been written to disk. Freeing of the inode
4406 	 * is delayed until after all blocks have been freed to
4407 	 * avoid creation of new <vfsid, inum, lbn> triples
4408 	 * before the old ones have been deleted.
4409 	 */
4410 	filefree = NULL;
4411 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
4412 		WORKLIST_REMOVE(wk);
4413 		switch (wk->wk_type) {
4414 
4415 		case D_FREEFILE:
4416 			/*
4417 			 * We defer adding filefree to the worklist until
4418 			 * all other additions have been made to ensure
4419 			 * that it will be done after all the old blocks
4420 			 * have been freed.
4421 			 */
4422 			if (filefree != NULL) {
4423 				lk.lkt_held = NOHOLDER;
4424 				panic("handle_written_inodeblock: filefree");
4425 			}
4426 			filefree = wk;
4427 			continue;
4428 
4429 		case D_MKDIR:
4430 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
4431 			continue;
4432 
4433 		case D_DIRADD:
4434 			diradd_inode_written(WK_DIRADD(wk), inodedep);
4435 			continue;
4436 
4437 		case D_FREEBLKS:
4438 		case D_FREEFRAG:
4439 		case D_DIRREM:
4440 			add_to_worklist(wk);
4441 			continue;
4442 
4443 		case D_NEWDIRBLK:
4444 			free_newdirblk(WK_NEWDIRBLK(wk));
4445 			continue;
4446 
4447 		default:
4448 			lk.lkt_held = NOHOLDER;
4449 			panic("handle_written_inodeblock: Unknown type %s",
4450 			    TYPENAME(wk->wk_type));
4451 			/* NOTREACHED */
4452 		}
4453 	}
4454 	if (filefree != NULL) {
4455 		if (free_inodedep(inodedep) == 0) {
4456 			lk.lkt_held = NOHOLDER;
4457 			panic("handle_written_inodeblock: live inodedep");
4458 		}
4459 		add_to_worklist(filefree);
4460 		return (0);
4461 	}
4462 
4463 	/*
4464 	 * If no outstanding dependencies, free it.
4465 	 */
4466 	if (free_inodedep(inodedep) ||
4467 	    (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
4468 	     TAILQ_FIRST(&inodedep->id_extupdt) == 0))
4469 		return (0);
4470 	return (hadchanges);
4471 }
4472 
4473 /*
4474  * Process a diradd entry after its dependent inode has been written.
4475  * This routine must be called with splbio interrupts blocked.
4476  */
4477 static void
4478 diradd_inode_written(dap, inodedep)
4479 	struct diradd *dap;
4480 	struct inodedep *inodedep;
4481 {
4482 	struct pagedep *pagedep;
4483 
4484 	dap->da_state |= COMPLETE;
4485 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4486 		if (dap->da_state & DIRCHG)
4487 			pagedep = dap->da_previous->dm_pagedep;
4488 		else
4489 			pagedep = dap->da_pagedep;
4490 		LIST_REMOVE(dap, da_pdlist);
4491 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4492 	}
4493 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
4494 }
4495 
4496 /*
4497  * Handle the completion of a mkdir dependency.
4498  */
4499 static void
4500 handle_written_mkdir(mkdir, type)
4501 	struct mkdir *mkdir;
4502 	int type;
4503 {
4504 	struct diradd *dap;
4505 	struct pagedep *pagedep;
4506 
4507 	if (mkdir->md_state != type) {
4508 		lk.lkt_held = NOHOLDER;
4509 		panic("handle_written_mkdir: bad type");
4510 	}
4511 	dap = mkdir->md_diradd;
4512 	dap->da_state &= ~type;
4513 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
4514 		dap->da_state |= DEPCOMPLETE;
4515 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4516 		if (dap->da_state & DIRCHG)
4517 			pagedep = dap->da_previous->dm_pagedep;
4518 		else
4519 			pagedep = dap->da_pagedep;
4520 		LIST_REMOVE(dap, da_pdlist);
4521 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4522 	}
4523 	LIST_REMOVE(mkdir, md_mkdirs);
4524 	WORKITEM_FREE(mkdir, D_MKDIR);
4525 }
4526 
4527 /*
4528  * Called from within softdep_disk_write_complete above.
4529  * A write operation was just completed. Removed inodes can
4530  * now be freed and associated block pointers may be committed.
4531  * Note that this routine is always called from interrupt level
4532  * with further splbio interrupts blocked.
4533  */
4534 static int
4535 handle_written_filepage(pagedep, bp)
4536 	struct pagedep *pagedep;
4537 	struct buf *bp;		/* buffer containing the written page */
4538 {
4539 	struct dirrem *dirrem;
4540 	struct diradd *dap, *nextdap;
4541 	struct direct *ep;
4542 	int i, chgs;
4543 
4544 	if ((pagedep->pd_state & IOSTARTED) == 0) {
4545 		lk.lkt_held = NOHOLDER;
4546 		panic("handle_written_filepage: not started");
4547 	}
4548 	pagedep->pd_state &= ~IOSTARTED;
4549 	/*
4550 	 * Process any directory removals that have been committed.
4551 	 */
4552 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
4553 		LIST_REMOVE(dirrem, dm_next);
4554 		dirrem->dm_dirinum = pagedep->pd_ino;
4555 		add_to_worklist(&dirrem->dm_list);
4556 	}
4557 	/*
4558 	 * Free any directory additions that have been committed.
4559 	 * If it is a newly allocated block, we have to wait until
4560 	 * the on-disk directory inode claims the new block.
4561 	 */
4562 	if ((pagedep->pd_state & NEWBLOCK) == 0)
4563 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
4564 			free_diradd(dap);
4565 	/*
4566 	 * Uncommitted directory entries must be restored.
4567 	 */
4568 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
4569 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
4570 		     dap = nextdap) {
4571 			nextdap = LIST_NEXT(dap, da_pdlist);
4572 			if (dap->da_state & ATTACHED) {
4573 				lk.lkt_held = NOHOLDER;
4574 				panic("handle_written_filepage: attached");
4575 			}
4576 			ep = (struct direct *)
4577 			    ((char *)bp->b_data + dap->da_offset);
4578 			ep->d_ino = dap->da_newinum;
4579 			dap->da_state &= ~UNDONE;
4580 			dap->da_state |= ATTACHED;
4581 			chgs = 1;
4582 			/*
4583 			 * If the inode referenced by the directory has
4584 			 * been written out, then the dependency can be
4585 			 * moved to the pending list.
4586 			 */
4587 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4588 				LIST_REMOVE(dap, da_pdlist);
4589 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
4590 				    da_pdlist);
4591 			}
4592 		}
4593 	}
4594 	/*
4595 	 * If there were any rollbacks in the directory, then it must be
4596 	 * marked dirty so that its will eventually get written back in
4597 	 * its correct form.
4598 	 */
4599 	if (chgs) {
4600 		if ((bp->b_flags & B_DELWRI) == 0)
4601 			stat_dir_entry++;
4602 		bdirty(bp);
4603 		return (1);
4604 	}
4605 	/*
4606 	 * If we are not waiting for a new directory block to be
4607 	 * claimed by its inode, then the pagedep will be freed.
4608 	 * Otherwise it will remain to track any new entries on
4609 	 * the page in case they are fsync'ed.
4610 	 */
4611 	if ((pagedep->pd_state & NEWBLOCK) == 0) {
4612 		LIST_REMOVE(pagedep, pd_hash);
4613 		WORKITEM_FREE(pagedep, D_PAGEDEP);
4614 	}
4615 	return (0);
4616 }
4617 
4618 /*
4619  * Writing back in-core inode structures.
4620  *
4621  * The filesystem only accesses an inode's contents when it occupies an
4622  * "in-core" inode structure.  These "in-core" structures are separate from
4623  * the page frames used to cache inode blocks.  Only the latter are
4624  * transferred to/from the disk.  So, when the updated contents of the
4625  * "in-core" inode structure are copied to the corresponding in-memory inode
4626  * block, the dependencies are also transferred.  The following procedure is
4627  * called when copying a dirty "in-core" inode to a cached inode block.
4628  */
4629 
4630 /*
4631  * Called when an inode is loaded from disk. If the effective link count
4632  * differed from the actual link count when it was last flushed, then we
4633  * need to ensure that the correct effective link count is put back.
4634  */
4635 void
4636 softdep_load_inodeblock(ip)
4637 	struct inode *ip;	/* the "in_core" copy of the inode */
4638 {
4639 	struct inodedep *inodedep;
4640 
4641 	/*
4642 	 * Check for alternate nlink count.
4643 	 */
4644 	ip->i_effnlink = ip->i_nlink;
4645 	ACQUIRE_LOCK(&lk);
4646 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
4647 		FREE_LOCK(&lk);
4648 		return;
4649 	}
4650 	ip->i_effnlink -= inodedep->id_nlinkdelta;
4651 	if (inodedep->id_state & SPACECOUNTED)
4652 		ip->i_flag |= IN_SPACECOUNTED;
4653 	FREE_LOCK(&lk);
4654 }
4655 
4656 /*
4657  * This routine is called just before the "in-core" inode
4658  * information is to be copied to the in-memory inode block.
4659  * Recall that an inode block contains several inodes. If
4660  * the force flag is set, then the dependencies will be
4661  * cleared so that the update can always be made. Note that
4662  * the buffer is locked when this routine is called, so we
4663  * will never be in the middle of writing the inode block
4664  * to disk.
4665  */
4666 void
4667 softdep_update_inodeblock(ip, bp, waitfor)
4668 	struct inode *ip;	/* the "in_core" copy of the inode */
4669 	struct buf *bp;		/* the buffer containing the inode block */
4670 	int waitfor;		/* nonzero => update must be allowed */
4671 {
4672 	struct inodedep *inodedep;
4673 	struct worklist *wk;
4674 	struct buf *ibp;
4675 	int error;
4676 
4677 	/*
4678 	 * If the effective link count is not equal to the actual link
4679 	 * count, then we must track the difference in an inodedep while
4680 	 * the inode is (potentially) tossed out of the cache. Otherwise,
4681 	 * if there is no existing inodedep, then there are no dependencies
4682 	 * to track.
4683 	 */
4684 	ACQUIRE_LOCK(&lk);
4685 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
4686 		FREE_LOCK(&lk);
4687 		if (ip->i_effnlink != ip->i_nlink)
4688 			panic("softdep_update_inodeblock: bad link count");
4689 		return;
4690 	}
4691 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
4692 		FREE_LOCK(&lk);
4693 		panic("softdep_update_inodeblock: bad delta");
4694 	}
4695 	/*
4696 	 * Changes have been initiated. Anything depending on these
4697 	 * changes cannot occur until this inode has been written.
4698 	 */
4699 	inodedep->id_state &= ~COMPLETE;
4700 	if ((inodedep->id_state & ONWORKLIST) == 0)
4701 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
4702 	/*
4703 	 * Any new dependencies associated with the incore inode must
4704 	 * now be moved to the list associated with the buffer holding
4705 	 * the in-memory copy of the inode. Once merged process any
4706 	 * allocdirects that are completed by the merger.
4707 	 */
4708 	merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
4709 	if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
4710 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
4711 	merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
4712 	if (TAILQ_FIRST(&inodedep->id_extupdt) != NULL)
4713 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt));
4714 	/*
4715 	 * Now that the inode has been pushed into the buffer, the
4716 	 * operations dependent on the inode being written to disk
4717 	 * can be moved to the id_bufwait so that they will be
4718 	 * processed when the buffer I/O completes.
4719 	 */
4720 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
4721 		WORKLIST_REMOVE(wk);
4722 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
4723 	}
4724 	/*
4725 	 * Newly allocated inodes cannot be written until the bitmap
4726 	 * that allocates them have been written (indicated by
4727 	 * DEPCOMPLETE being set in id_state). If we are doing a
4728 	 * forced sync (e.g., an fsync on a file), we force the bitmap
4729 	 * to be written so that the update can be done.
4730 	 */
4731 	if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
4732 		FREE_LOCK(&lk);
4733 		return;
4734 	}
4735 	ibp = inodedep->id_buf;
4736 	ibp = getdirtybuf(&ibp, NULL, MNT_WAIT);
4737 	FREE_LOCK(&lk);
4738 	if (ibp && (error = bwrite(ibp)) != 0)
4739 		softdep_error("softdep_update_inodeblock: bwrite", error);
4740 	if ((inodedep->id_state & DEPCOMPLETE) == 0)
4741 		panic("softdep_update_inodeblock: update failed");
4742 }
4743 
4744 /*
4745  * Merge the a new inode dependency list (such as id_newinoupdt) into an
4746  * old inode dependency list (such as id_inoupdt). This routine must be
4747  * called with splbio interrupts blocked.
4748  */
4749 static void
4750 merge_inode_lists(newlisthead, oldlisthead)
4751 	struct allocdirectlst *newlisthead;
4752 	struct allocdirectlst *oldlisthead;
4753 {
4754 	struct allocdirect *listadp, *newadp;
4755 
4756 	newadp = TAILQ_FIRST(newlisthead);
4757 	for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
4758 		if (listadp->ad_lbn < newadp->ad_lbn) {
4759 			listadp = TAILQ_NEXT(listadp, ad_next);
4760 			continue;
4761 		}
4762 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4763 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
4764 		if (listadp->ad_lbn == newadp->ad_lbn) {
4765 			allocdirect_merge(oldlisthead, newadp,
4766 			    listadp);
4767 			listadp = newadp;
4768 		}
4769 		newadp = TAILQ_FIRST(newlisthead);
4770 	}
4771 	while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
4772 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4773 		TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
4774 	}
4775 }
4776 
4777 /*
4778  * If we are doing an fsync, then we must ensure that any directory
4779  * entries for the inode have been written after the inode gets to disk.
4780  */
4781 int
4782 softdep_fsync(vp)
4783 	struct vnode *vp;	/* the "in_core" copy of the inode */
4784 {
4785 	struct inodedep *inodedep;
4786 	struct pagedep *pagedep;
4787 	struct worklist *wk;
4788 	struct diradd *dap;
4789 	struct mount *mnt;
4790 	struct vnode *pvp;
4791 	struct inode *ip;
4792 	struct buf *bp;
4793 	struct fs *fs;
4794 	struct thread *td = curthread;
4795 	int error, flushparent;
4796 	ino_t parentino;
4797 	ufs_lbn_t lbn;
4798 
4799 	ip = VTOI(vp);
4800 	fs = ip->i_fs;
4801 	ACQUIRE_LOCK(&lk);
4802 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
4803 		FREE_LOCK(&lk);
4804 		return (0);
4805 	}
4806 	if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
4807 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
4808 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
4809 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
4810 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
4811 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) {
4812 		FREE_LOCK(&lk);
4813 		panic("softdep_fsync: pending ops");
4814 	}
4815 	for (error = 0, flushparent = 0; ; ) {
4816 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
4817 			break;
4818 		if (wk->wk_type != D_DIRADD) {
4819 			FREE_LOCK(&lk);
4820 			panic("softdep_fsync: Unexpected type %s",
4821 			    TYPENAME(wk->wk_type));
4822 		}
4823 		dap = WK_DIRADD(wk);
4824 		/*
4825 		 * Flush our parent if this directory entry has a MKDIR_PARENT
4826 		 * dependency or is contained in a newly allocated block.
4827 		 */
4828 		if (dap->da_state & DIRCHG)
4829 			pagedep = dap->da_previous->dm_pagedep;
4830 		else
4831 			pagedep = dap->da_pagedep;
4832 		mnt = pagedep->pd_mnt;
4833 		parentino = pagedep->pd_ino;
4834 		lbn = pagedep->pd_lbn;
4835 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) {
4836 			FREE_LOCK(&lk);
4837 			panic("softdep_fsync: dirty");
4838 		}
4839 		if ((dap->da_state & MKDIR_PARENT) ||
4840 		    (pagedep->pd_state & NEWBLOCK))
4841 			flushparent = 1;
4842 		else
4843 			flushparent = 0;
4844 		/*
4845 		 * If we are being fsync'ed as part of vgone'ing this vnode,
4846 		 * then we will not be able to release and recover the
4847 		 * vnode below, so we just have to give up on writing its
4848 		 * directory entry out. It will eventually be written, just
4849 		 * not now, but then the user was not asking to have it
4850 		 * written, so we are not breaking any promises.
4851 		 */
4852 		if (vp->v_iflag & VI_XLOCK)
4853 			break;
4854 		/*
4855 		 * We prevent deadlock by always fetching inodes from the
4856 		 * root, moving down the directory tree. Thus, when fetching
4857 		 * our parent directory, we first try to get the lock. If
4858 		 * that fails, we must unlock ourselves before requesting
4859 		 * the lock on our parent. See the comment in ufs_lookup
4860 		 * for details on possible races.
4861 		 */
4862 		FREE_LOCK(&lk);
4863 		if (VFS_VGET(mnt, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) {
4864 			VOP_UNLOCK(vp, 0, td);
4865 			error = VFS_VGET(mnt, parentino, LK_EXCLUSIVE, &pvp);
4866 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
4867 			if (error != 0)
4868 				return (error);
4869 		}
4870 		/*
4871 		 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
4872 		 * that are contained in direct blocks will be resolved by
4873 		 * doing a UFS_UPDATE. Pagedeps contained in indirect blocks
4874 		 * may require a complete sync'ing of the directory. So, we
4875 		 * try the cheap and fast UFS_UPDATE first, and if that fails,
4876 		 * then we do the slower VOP_FSYNC of the directory.
4877 		 */
4878 		if (flushparent) {
4879 			if ((error = UFS_UPDATE(pvp, 1)) != 0) {
4880 				vput(pvp);
4881 				return (error);
4882 			}
4883 			if ((pagedep->pd_state & NEWBLOCK) &&
4884 			    (error = VOP_FSYNC(pvp, td->td_ucred, MNT_WAIT, td))) {
4885 				vput(pvp);
4886 				return (error);
4887 			}
4888 		}
4889 		/*
4890 		 * Flush directory page containing the inode's name.
4891 		 */
4892 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
4893 		    &bp);
4894 		if (error == 0)
4895 			error = bwrite(bp);
4896 		else
4897 			brelse(bp);
4898 		vput(pvp);
4899 		if (error != 0)
4900 			return (error);
4901 		ACQUIRE_LOCK(&lk);
4902 		if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
4903 			break;
4904 	}
4905 	FREE_LOCK(&lk);
4906 	return (0);
4907 }
4908 
4909 /*
4910  * Flush all the dirty bitmaps associated with the block device
4911  * before flushing the rest of the dirty blocks so as to reduce
4912  * the number of dependencies that will have to be rolled back.
4913  */
4914 void
4915 softdep_fsync_mountdev(vp)
4916 	struct vnode *vp;
4917 {
4918 	struct buf *bp, *nbp;
4919 	struct worklist *wk;
4920 
4921 	if (!vn_isdisk(vp, NULL))
4922 		panic("softdep_fsync_mountdev: vnode not a disk");
4923 	ACQUIRE_LOCK(&lk);
4924 	VI_LOCK(vp);
4925 	for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
4926 		nbp = TAILQ_NEXT(bp, b_vnbufs);
4927 		/*
4928 		 * If it is already scheduled, skip to the next buffer.
4929 		 */
4930 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
4931 			continue;
4932 
4933 		if ((bp->b_flags & B_DELWRI) == 0) {
4934 			FREE_LOCK(&lk);
4935 			panic("softdep_fsync_mountdev: not dirty");
4936 		}
4937 		/*
4938 		 * We are only interested in bitmaps with outstanding
4939 		 * dependencies.
4940 		 */
4941 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
4942 		    wk->wk_type != D_BMSAFEMAP ||
4943 		    (bp->b_vflags & BV_BKGRDINPROG)) {
4944 			BUF_UNLOCK(bp);
4945 			continue;
4946 		}
4947 		VI_UNLOCK(vp);
4948 		bremfree(bp);
4949 		FREE_LOCK(&lk);
4950 		(void) bawrite(bp);
4951 		ACQUIRE_LOCK(&lk);
4952 		/*
4953 		 * Since we may have slept during the I/O, we need
4954 		 * to start from a known point.
4955 		 */
4956 		VI_LOCK(vp);
4957 		nbp = TAILQ_FIRST(&vp->v_dirtyblkhd);
4958 	}
4959 	drain_output(vp, 1);
4960 	VI_UNLOCK(vp);
4961 	FREE_LOCK(&lk);
4962 }
4963 
4964 /*
4965  * This routine is called when we are trying to synchronously flush a
4966  * file. This routine must eliminate any filesystem metadata dependencies
4967  * so that the syncing routine can succeed by pushing the dirty blocks
4968  * associated with the file. If any I/O errors occur, they are returned.
4969  */
4970 int
4971 softdep_sync_metadata(ap)
4972 	struct vop_fsync_args /* {
4973 		struct vnode *a_vp;
4974 		struct ucred *a_cred;
4975 		int a_waitfor;
4976 		struct thread *a_td;
4977 	} */ *ap;
4978 {
4979 	struct vnode *vp = ap->a_vp;
4980 	struct pagedep *pagedep;
4981 	struct allocdirect *adp;
4982 	struct allocindir *aip;
4983 	struct buf *bp, *nbp;
4984 	struct worklist *wk;
4985 	int i, error, waitfor;
4986 
4987 	/*
4988 	 * Check whether this vnode is involved in a filesystem
4989 	 * that is doing soft dependency processing.
4990 	 */
4991 	if (!vn_isdisk(vp, NULL)) {
4992 		if (!DOINGSOFTDEP(vp))
4993 			return (0);
4994 	} else
4995 		if (vp->v_rdev->si_mountpoint == NULL ||
4996 		    (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0)
4997 			return (0);
4998 	/*
4999 	 * Ensure that any direct block dependencies have been cleared.
5000 	 */
5001 	ACQUIRE_LOCK(&lk);
5002 	if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
5003 		FREE_LOCK(&lk);
5004 		return (error);
5005 	}
5006 	/*
5007 	 * For most files, the only metadata dependencies are the
5008 	 * cylinder group maps that allocate their inode or blocks.
5009 	 * The block allocation dependencies can be found by traversing
5010 	 * the dependency lists for any buffers that remain on their
5011 	 * dirty buffer list. The inode allocation dependency will
5012 	 * be resolved when the inode is updated with MNT_WAIT.
5013 	 * This work is done in two passes. The first pass grabs most
5014 	 * of the buffers and begins asynchronously writing them. The
5015 	 * only way to wait for these asynchronous writes is to sleep
5016 	 * on the filesystem vnode which may stay busy for a long time
5017 	 * if the filesystem is active. So, instead, we make a second
5018 	 * pass over the dependencies blocking on each write. In the
5019 	 * usual case we will be blocking against a write that we
5020 	 * initiated, so when it is done the dependency will have been
5021 	 * resolved. Thus the second pass is expected to end quickly.
5022 	 */
5023 	waitfor = MNT_NOWAIT;
5024 top:
5025 	/*
5026 	 * We must wait for any I/O in progress to finish so that
5027 	 * all potential buffers on the dirty list will be visible.
5028 	 */
5029 	VI_LOCK(vp);
5030 	drain_output(vp, 1);
5031 	bp = getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd),
5032 	    VI_MTX(vp), MNT_WAIT);
5033 	if (bp == NULL) {
5034 		VI_UNLOCK(vp);
5035 		FREE_LOCK(&lk);
5036 		return (0);
5037 	}
5038 	/* While syncing snapshots, we must allow recursive lookups */
5039 	bp->b_lock.lk_flags |= LK_CANRECURSE;
5040 loop:
5041 	/*
5042 	 * As we hold the buffer locked, none of its dependencies
5043 	 * will disappear.
5044 	 */
5045 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5046 		switch (wk->wk_type) {
5047 
5048 		case D_ALLOCDIRECT:
5049 			adp = WK_ALLOCDIRECT(wk);
5050 			if (adp->ad_state & DEPCOMPLETE)
5051 				continue;
5052 			nbp = adp->ad_buf;
5053 			nbp = getdirtybuf(&nbp, NULL, waitfor);
5054 			if (nbp == NULL)
5055 				continue;
5056 			FREE_LOCK(&lk);
5057 			if (waitfor == MNT_NOWAIT) {
5058 				bawrite(nbp);
5059 			} else if ((error = bwrite(nbp)) != 0) {
5060 				break;
5061 			}
5062 			ACQUIRE_LOCK(&lk);
5063 			continue;
5064 
5065 		case D_ALLOCINDIR:
5066 			aip = WK_ALLOCINDIR(wk);
5067 			if (aip->ai_state & DEPCOMPLETE)
5068 				continue;
5069 			nbp = aip->ai_buf;
5070 			nbp = getdirtybuf(&nbp, NULL, waitfor);
5071 			if (nbp == NULL)
5072 				continue;
5073 			FREE_LOCK(&lk);
5074 			if (waitfor == MNT_NOWAIT) {
5075 				bawrite(nbp);
5076 			} else if ((error = bwrite(nbp)) != 0) {
5077 				break;
5078 			}
5079 			ACQUIRE_LOCK(&lk);
5080 			continue;
5081 
5082 		case D_INDIRDEP:
5083 		restart:
5084 
5085 			LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
5086 				if (aip->ai_state & DEPCOMPLETE)
5087 					continue;
5088 				nbp = aip->ai_buf;
5089 				nbp = getdirtybuf(&nbp, NULL, MNT_WAIT);
5090 				if (nbp == NULL)
5091 					goto restart;
5092 				FREE_LOCK(&lk);
5093 				if ((error = bwrite(nbp)) != 0) {
5094 					break;
5095 				}
5096 				ACQUIRE_LOCK(&lk);
5097 				goto restart;
5098 			}
5099 			continue;
5100 
5101 		case D_INODEDEP:
5102 			if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
5103 			    WK_INODEDEP(wk)->id_ino)) != 0) {
5104 				FREE_LOCK(&lk);
5105 				break;
5106 			}
5107 			continue;
5108 
5109 		case D_PAGEDEP:
5110 			/*
5111 			 * We are trying to sync a directory that may
5112 			 * have dependencies on both its own metadata
5113 			 * and/or dependencies on the inodes of any
5114 			 * recently allocated files. We walk its diradd
5115 			 * lists pushing out the associated inode.
5116 			 */
5117 			pagedep = WK_PAGEDEP(wk);
5118 			for (i = 0; i < DAHASHSZ; i++) {
5119 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
5120 					continue;
5121 				if ((error =
5122 				    flush_pagedep_deps(vp, pagedep->pd_mnt,
5123 						&pagedep->pd_diraddhd[i]))) {
5124 					FREE_LOCK(&lk);
5125 					break;
5126 				}
5127 			}
5128 			continue;
5129 
5130 		case D_MKDIR:
5131 			/*
5132 			 * This case should never happen if the vnode has
5133 			 * been properly sync'ed. However, if this function
5134 			 * is used at a place where the vnode has not yet
5135 			 * been sync'ed, this dependency can show up. So,
5136 			 * rather than panic, just flush it.
5137 			 */
5138 			nbp = WK_MKDIR(wk)->md_buf;
5139 			nbp = getdirtybuf(&nbp, NULL, waitfor);
5140 			if (nbp == NULL)
5141 				continue;
5142 			FREE_LOCK(&lk);
5143 			if (waitfor == MNT_NOWAIT) {
5144 				bawrite(nbp);
5145 			} else if ((error = bwrite(nbp)) != 0) {
5146 				break;
5147 			}
5148 			ACQUIRE_LOCK(&lk);
5149 			continue;
5150 
5151 		case D_BMSAFEMAP:
5152 			/*
5153 			 * This case should never happen if the vnode has
5154 			 * been properly sync'ed. However, if this function
5155 			 * is used at a place where the vnode has not yet
5156 			 * been sync'ed, this dependency can show up. So,
5157 			 * rather than panic, just flush it.
5158 			 */
5159 			nbp = WK_BMSAFEMAP(wk)->sm_buf;
5160 			nbp = getdirtybuf(&nbp, NULL, waitfor);
5161 			if (nbp == NULL)
5162 				continue;
5163 			FREE_LOCK(&lk);
5164 			if (waitfor == MNT_NOWAIT) {
5165 				bawrite(nbp);
5166 			} else if ((error = bwrite(nbp)) != 0) {
5167 				break;
5168 			}
5169 			ACQUIRE_LOCK(&lk);
5170 			continue;
5171 
5172 		default:
5173 			FREE_LOCK(&lk);
5174 			panic("softdep_sync_metadata: Unknown type %s",
5175 			    TYPENAME(wk->wk_type));
5176 			/* NOTREACHED */
5177 		}
5178 		/* We reach here only in error and unlocked */
5179 		if (error == 0)
5180 			panic("softdep_sync_metadata: zero error");
5181 		bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5182 		bawrite(bp);
5183 		return (error);
5184 	}
5185 	VI_LOCK(vp);
5186 	nbp = getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), VI_MTX(vp), MNT_WAIT);
5187 	if (nbp == NULL)
5188 		VI_UNLOCK(vp);
5189 	FREE_LOCK(&lk);
5190 	bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5191 	bawrite(bp);
5192 	ACQUIRE_LOCK(&lk);
5193 	if (nbp != NULL) {
5194 		bp = nbp;
5195 		goto loop;
5196 	}
5197 	/*
5198 	 * The brief unlock is to allow any pent up dependency
5199 	 * processing to be done. Then proceed with the second pass.
5200 	 */
5201 	if (waitfor == MNT_NOWAIT) {
5202 		waitfor = MNT_WAIT;
5203 		FREE_LOCK(&lk);
5204 		ACQUIRE_LOCK(&lk);
5205 		goto top;
5206 	}
5207 
5208 	/*
5209 	 * If we have managed to get rid of all the dirty buffers,
5210 	 * then we are done. For certain directories and block
5211 	 * devices, we may need to do further work.
5212 	 *
5213 	 * We must wait for any I/O in progress to finish so that
5214 	 * all potential buffers on the dirty list will be visible.
5215 	 */
5216 	VI_LOCK(vp);
5217 	drain_output(vp, 1);
5218 	if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) {
5219 		VI_UNLOCK(vp);
5220 		FREE_LOCK(&lk);
5221 		return (0);
5222 	}
5223 	VI_UNLOCK(vp);
5224 
5225 	FREE_LOCK(&lk);
5226 	/*
5227 	 * If we are trying to sync a block device, some of its buffers may
5228 	 * contain metadata that cannot be written until the contents of some
5229 	 * partially written files have been written to disk. The only easy
5230 	 * way to accomplish this is to sync the entire filesystem (luckily
5231 	 * this happens rarely).
5232 	 */
5233 	if (vn_isdisk(vp, NULL) &&
5234 	    vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) &&
5235 	    (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, ap->a_cred,
5236 	     ap->a_td)) != 0)
5237 		return (error);
5238 	return (0);
5239 }
5240 
5241 /*
5242  * Flush the dependencies associated with an inodedep.
5243  * Called with splbio blocked.
5244  */
5245 static int
5246 flush_inodedep_deps(fs, ino)
5247 	struct fs *fs;
5248 	ino_t ino;
5249 {
5250 	struct inodedep *inodedep;
5251 	int error, waitfor;
5252 
5253 	/*
5254 	 * This work is done in two passes. The first pass grabs most
5255 	 * of the buffers and begins asynchronously writing them. The
5256 	 * only way to wait for these asynchronous writes is to sleep
5257 	 * on the filesystem vnode which may stay busy for a long time
5258 	 * if the filesystem is active. So, instead, we make a second
5259 	 * pass over the dependencies blocking on each write. In the
5260 	 * usual case we will be blocking against a write that we
5261 	 * initiated, so when it is done the dependency will have been
5262 	 * resolved. Thus the second pass is expected to end quickly.
5263 	 * We give a brief window at the top of the loop to allow
5264 	 * any pending I/O to complete.
5265 	 */
5266 	for (error = 0, waitfor = MNT_NOWAIT; ; ) {
5267 		if (error)
5268 			return (error);
5269 		FREE_LOCK(&lk);
5270 		ACQUIRE_LOCK(&lk);
5271 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
5272 			return (0);
5273 		if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
5274 		    flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
5275 		    flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
5276 		    flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
5277 			continue;
5278 		/*
5279 		 * If pass2, we are done, otherwise do pass 2.
5280 		 */
5281 		if (waitfor == MNT_WAIT)
5282 			break;
5283 		waitfor = MNT_WAIT;
5284 	}
5285 	/*
5286 	 * Try freeing inodedep in case all dependencies have been removed.
5287 	 */
5288 	if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
5289 		(void) free_inodedep(inodedep);
5290 	return (0);
5291 }
5292 
5293 /*
5294  * Flush an inode dependency list.
5295  * Called with splbio blocked.
5296  */
5297 static int
5298 flush_deplist(listhead, waitfor, errorp)
5299 	struct allocdirectlst *listhead;
5300 	int waitfor;
5301 	int *errorp;
5302 {
5303 	struct allocdirect *adp;
5304 	struct buf *bp;
5305 
5306 	TAILQ_FOREACH(adp, listhead, ad_next) {
5307 		if (adp->ad_state & DEPCOMPLETE)
5308 			continue;
5309 		bp = adp->ad_buf;
5310 		bp = getdirtybuf(&bp, NULL, waitfor);
5311 		if (bp == NULL) {
5312 			if (waitfor == MNT_NOWAIT)
5313 				continue;
5314 			return (1);
5315 		}
5316 		FREE_LOCK(&lk);
5317 		if (waitfor == MNT_NOWAIT) {
5318 			bawrite(bp);
5319 		} else if ((*errorp = bwrite(bp)) != 0) {
5320 			ACQUIRE_LOCK(&lk);
5321 			return (1);
5322 		}
5323 		ACQUIRE_LOCK(&lk);
5324 		return (1);
5325 	}
5326 	return (0);
5327 }
5328 
5329 /*
5330  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
5331  * Called with splbio blocked.
5332  */
5333 static int
5334 flush_pagedep_deps(pvp, mp, diraddhdp)
5335 	struct vnode *pvp;
5336 	struct mount *mp;
5337 	struct diraddhd *diraddhdp;
5338 {
5339 	struct thread *td = curthread;
5340 	struct inodedep *inodedep;
5341 	struct ufsmount *ump;
5342 	struct diradd *dap;
5343 	struct vnode *vp;
5344 	int error = 0;
5345 	struct buf *bp;
5346 	ino_t inum;
5347 
5348 	ump = VFSTOUFS(mp);
5349 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
5350 		/*
5351 		 * Flush ourselves if this directory entry
5352 		 * has a MKDIR_PARENT dependency.
5353 		 */
5354 		if (dap->da_state & MKDIR_PARENT) {
5355 			FREE_LOCK(&lk);
5356 			if ((error = UFS_UPDATE(pvp, 1)) != 0)
5357 				break;
5358 			ACQUIRE_LOCK(&lk);
5359 			/*
5360 			 * If that cleared dependencies, go on to next.
5361 			 */
5362 			if (dap != LIST_FIRST(diraddhdp))
5363 				continue;
5364 			if (dap->da_state & MKDIR_PARENT) {
5365 				FREE_LOCK(&lk);
5366 				panic("flush_pagedep_deps: MKDIR_PARENT");
5367 			}
5368 		}
5369 		/*
5370 		 * A newly allocated directory must have its "." and
5371 		 * ".." entries written out before its name can be
5372 		 * committed in its parent. We do not want or need
5373 		 * the full semantics of a synchronous VOP_FSYNC as
5374 		 * that may end up here again, once for each directory
5375 		 * level in the filesystem. Instead, we push the blocks
5376 		 * and wait for them to clear. We have to fsync twice
5377 		 * because the first call may choose to defer blocks
5378 		 * that still have dependencies, but deferral will
5379 		 * happen at most once.
5380 		 */
5381 		inum = dap->da_newinum;
5382 		if (dap->da_state & MKDIR_BODY) {
5383 			FREE_LOCK(&lk);
5384 			if ((error = VFS_VGET(mp, inum, LK_EXCLUSIVE, &vp)))
5385 				break;
5386 			if ((error=VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)) ||
5387 			    (error=VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td))) {
5388 				vput(vp);
5389 				break;
5390 			}
5391 			VI_LOCK(vp);
5392 			drain_output(vp, 0);
5393 			VI_UNLOCK(vp);
5394 			vput(vp);
5395 			ACQUIRE_LOCK(&lk);
5396 			/*
5397 			 * If that cleared dependencies, go on to next.
5398 			 */
5399 			if (dap != LIST_FIRST(diraddhdp))
5400 				continue;
5401 			if (dap->da_state & MKDIR_BODY) {
5402 				FREE_LOCK(&lk);
5403 				panic("flush_pagedep_deps: MKDIR_BODY");
5404 			}
5405 		}
5406 		/*
5407 		 * Flush the inode on which the directory entry depends.
5408 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
5409 		 * the only remaining dependency is that the updated inode
5410 		 * count must get pushed to disk. The inode has already
5411 		 * been pushed into its inode buffer (via VOP_UPDATE) at
5412 		 * the time of the reference count change. So we need only
5413 		 * locate that buffer, ensure that there will be no rollback
5414 		 * caused by a bitmap dependency, then write the inode buffer.
5415 		 */
5416 		if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) {
5417 			FREE_LOCK(&lk);
5418 			panic("flush_pagedep_deps: lost inode");
5419 		}
5420 		/*
5421 		 * If the inode still has bitmap dependencies,
5422 		 * push them to disk.
5423 		 */
5424 		if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5425 			bp = inodedep->id_buf;
5426 			bp = getdirtybuf(&bp, NULL, MNT_WAIT);
5427 			FREE_LOCK(&lk);
5428 			if (bp && (error = bwrite(bp)) != 0)
5429 				break;
5430 			ACQUIRE_LOCK(&lk);
5431 			if (dap != LIST_FIRST(diraddhdp))
5432 				continue;
5433 		}
5434 		/*
5435 		 * If the inode is still sitting in a buffer waiting
5436 		 * to be written, push it to disk.
5437 		 */
5438 		FREE_LOCK(&lk);
5439 		if ((error = bread(ump->um_devvp,
5440 		    fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
5441 		    (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) {
5442 			brelse(bp);
5443 			break;
5444 		}
5445 		if ((error = bwrite(bp)) != 0)
5446 			break;
5447 		ACQUIRE_LOCK(&lk);
5448 		/*
5449 		 * If we have failed to get rid of all the dependencies
5450 		 * then something is seriously wrong.
5451 		 */
5452 		if (dap == LIST_FIRST(diraddhdp)) {
5453 			FREE_LOCK(&lk);
5454 			panic("flush_pagedep_deps: flush failed");
5455 		}
5456 	}
5457 	if (error)
5458 		ACQUIRE_LOCK(&lk);
5459 	return (error);
5460 }
5461 
5462 /*
5463  * A large burst of file addition or deletion activity can drive the
5464  * memory load excessively high. First attempt to slow things down
5465  * using the techniques below. If that fails, this routine requests
5466  * the offending operations to fall back to running synchronously
5467  * until the memory load returns to a reasonable level.
5468  */
5469 int
5470 softdep_slowdown(vp)
5471 	struct vnode *vp;
5472 {
5473 	int max_softdeps_hard;
5474 
5475 	max_softdeps_hard = max_softdeps * 11 / 10;
5476 	if (num_dirrem < max_softdeps_hard / 2 &&
5477 	    num_inodedep < max_softdeps_hard &&
5478 	    VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps)
5479   		return (0);
5480 	if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps)
5481 		speedup_syncer();
5482 	stat_sync_limit_hit += 1;
5483 	return (1);
5484 }
5485 
5486 /*
5487  * Called by the allocation routines when they are about to fail
5488  * in the hope that we can free up some disk space.
5489  *
5490  * First check to see if the work list has anything on it. If it has,
5491  * clean up entries until we successfully free some space. Because this
5492  * process holds inodes locked, we cannot handle any remove requests
5493  * that might block on a locked inode as that could lead to deadlock.
5494  * If the worklist yields no free space, encourage the syncer daemon
5495  * to help us. In no event will we try for longer than tickdelay seconds.
5496  */
5497 int
5498 softdep_request_cleanup(fs, vp)
5499 	struct fs *fs;
5500 	struct vnode *vp;
5501 {
5502 	long starttime;
5503 	ufs2_daddr_t needed;
5504 
5505 	needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize;
5506 	starttime = time_second + tickdelay;
5507 	/*
5508 	 * If we are being called because of a process doing a
5509 	 * copy-on-write, then it is not safe to update the vnode
5510 	 * as we may recurse into the copy-on-write routine.
5511 	 */
5512 	if (!(curthread->td_pflags & TDP_COWINPROGRESS) &&
5513 	    UFS_UPDATE(vp, 1) != 0)
5514 		return (0);
5515 	while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) {
5516 		if (time_second > starttime)
5517 			return (0);
5518 		if (num_on_worklist > 0 &&
5519 		    process_worklist_item(NULL, LK_NOWAIT) != -1) {
5520 			stat_worklist_push += 1;
5521 			continue;
5522 		}
5523 		request_cleanup(FLUSH_REMOVE_WAIT, 0);
5524 	}
5525 	return (1);
5526 }
5527 
5528 /*
5529  * If memory utilization has gotten too high, deliberately slow things
5530  * down and speed up the I/O processing.
5531  */
5532 static int
5533 request_cleanup(resource, islocked)
5534 	int resource;
5535 	int islocked;
5536 {
5537 	struct thread *td = curthread;
5538 
5539 	/*
5540 	 * We never hold up the filesystem syncer process.
5541 	 */
5542 	if (td == filesys_syncer)
5543 		return (0);
5544 	/*
5545 	 * First check to see if the work list has gotten backlogged.
5546 	 * If it has, co-opt this process to help clean up two entries.
5547 	 * Because this process may hold inodes locked, we cannot
5548 	 * handle any remove requests that might block on a locked
5549 	 * inode as that could lead to deadlock.
5550 	 */
5551 	if (num_on_worklist > max_softdeps / 10) {
5552 		if (islocked)
5553 			FREE_LOCK(&lk);
5554 		process_worklist_item(NULL, LK_NOWAIT);
5555 		process_worklist_item(NULL, LK_NOWAIT);
5556 		stat_worklist_push += 2;
5557 		if (islocked)
5558 			ACQUIRE_LOCK(&lk);
5559 		return(1);
5560 	}
5561 	/*
5562 	 * Next, we attempt to speed up the syncer process. If that
5563 	 * is successful, then we allow the process to continue.
5564 	 */
5565 	if (speedup_syncer() && resource != FLUSH_REMOVE_WAIT)
5566 		return(0);
5567 	/*
5568 	 * If we are resource constrained on inode dependencies, try
5569 	 * flushing some dirty inodes. Otherwise, we are constrained
5570 	 * by file deletions, so try accelerating flushes of directories
5571 	 * with removal dependencies. We would like to do the cleanup
5572 	 * here, but we probably hold an inode locked at this point and
5573 	 * that might deadlock against one that we try to clean. So,
5574 	 * the best that we can do is request the syncer daemon to do
5575 	 * the cleanup for us.
5576 	 */
5577 	switch (resource) {
5578 
5579 	case FLUSH_INODES:
5580 		stat_ino_limit_push += 1;
5581 		req_clear_inodedeps += 1;
5582 		stat_countp = &stat_ino_limit_hit;
5583 		break;
5584 
5585 	case FLUSH_REMOVE:
5586 	case FLUSH_REMOVE_WAIT:
5587 		stat_blk_limit_push += 1;
5588 		req_clear_remove += 1;
5589 		stat_countp = &stat_blk_limit_hit;
5590 		break;
5591 
5592 	default:
5593 		if (islocked)
5594 			FREE_LOCK(&lk);
5595 		panic("request_cleanup: unknown type");
5596 	}
5597 	/*
5598 	 * Hopefully the syncer daemon will catch up and awaken us.
5599 	 * We wait at most tickdelay before proceeding in any case.
5600 	 */
5601 	if (islocked == 0)
5602 		ACQUIRE_LOCK(&lk);
5603 	proc_waiting += 1;
5604 	if (handle.callout == NULL)
5605 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5606 	interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, NULL, PPAUSE,
5607 	    "softupdate", 0);
5608 	proc_waiting -= 1;
5609 	if (islocked == 0)
5610 		FREE_LOCK(&lk);
5611 	return (1);
5612 }
5613 
5614 /*
5615  * Awaken processes pausing in request_cleanup and clear proc_waiting
5616  * to indicate that there is no longer a timer running.
5617  */
5618 static void
5619 pause_timer(arg)
5620 	void *arg;
5621 {
5622 
5623 	*stat_countp += 1;
5624 	wakeup_one(&proc_waiting);
5625 	if (proc_waiting > 0)
5626 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5627 	else
5628 		handle.callout = NULL;
5629 }
5630 
5631 /*
5632  * Flush out a directory with at least one removal dependency in an effort to
5633  * reduce the number of dirrem, freefile, and freeblks dependency structures.
5634  */
5635 static void
5636 clear_remove(td)
5637 	struct thread *td;
5638 {
5639 	struct pagedep_hashhead *pagedephd;
5640 	struct pagedep *pagedep;
5641 	static int next = 0;
5642 	struct mount *mp;
5643 	struct vnode *vp;
5644 	int error, cnt;
5645 	ino_t ino;
5646 
5647 	ACQUIRE_LOCK(&lk);
5648 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
5649 		pagedephd = &pagedep_hashtbl[next++];
5650 		if (next >= pagedep_hash)
5651 			next = 0;
5652 		LIST_FOREACH(pagedep, pagedephd, pd_hash) {
5653 			if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
5654 				continue;
5655 			mp = pagedep->pd_mnt;
5656 			ino = pagedep->pd_ino;
5657 			if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5658 				continue;
5659 			FREE_LOCK(&lk);
5660 			if ((error = VFS_VGET(mp, ino, LK_EXCLUSIVE, &vp))) {
5661 				softdep_error("clear_remove: vget", error);
5662 				vn_finished_write(mp);
5663 				return;
5664 			}
5665 			if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)))
5666 				softdep_error("clear_remove: fsync", error);
5667 			VI_LOCK(vp);
5668 			drain_output(vp, 0);
5669 			VI_UNLOCK(vp);
5670 			vput(vp);
5671 			vn_finished_write(mp);
5672 			return;
5673 		}
5674 	}
5675 	FREE_LOCK(&lk);
5676 }
5677 
5678 /*
5679  * Clear out a block of dirty inodes in an effort to reduce
5680  * the number of inodedep dependency structures.
5681  */
5682 static void
5683 clear_inodedeps(td)
5684 	struct thread *td;
5685 {
5686 	struct inodedep_hashhead *inodedephd;
5687 	struct inodedep *inodedep;
5688 	static int next = 0;
5689 	struct mount *mp;
5690 	struct vnode *vp;
5691 	struct fs *fs;
5692 	int error, cnt;
5693 	ino_t firstino, lastino, ino;
5694 
5695 	ACQUIRE_LOCK(&lk);
5696 	/*
5697 	 * Pick a random inode dependency to be cleared.
5698 	 * We will then gather up all the inodes in its block
5699 	 * that have dependencies and flush them out.
5700 	 */
5701 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
5702 		inodedephd = &inodedep_hashtbl[next++];
5703 		if (next >= inodedep_hash)
5704 			next = 0;
5705 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
5706 			break;
5707 	}
5708 	if (inodedep == NULL)
5709 		return;
5710 	/*
5711 	 * Ugly code to find mount point given pointer to superblock.
5712 	 */
5713 	fs = inodedep->id_fs;
5714 	TAILQ_FOREACH(mp, &mountlist, mnt_list)
5715 		if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs)
5716 			break;
5717 	/*
5718 	 * Find the last inode in the block with dependencies.
5719 	 */
5720 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
5721 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
5722 		if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
5723 			break;
5724 	/*
5725 	 * Asynchronously push all but the last inode with dependencies.
5726 	 * Synchronously push the last inode with dependencies to ensure
5727 	 * that the inode block gets written to free up the inodedeps.
5728 	 */
5729 	for (ino = firstino; ino <= lastino; ino++) {
5730 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
5731 			continue;
5732 		if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5733 			continue;
5734 		FREE_LOCK(&lk);
5735 		if ((error = VFS_VGET(mp, ino, LK_EXCLUSIVE, &vp)) != 0) {
5736 			softdep_error("clear_inodedeps: vget", error);
5737 			vn_finished_write(mp);
5738 			return;
5739 		}
5740 		if (ino == lastino) {
5741 			if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_WAIT, td)))
5742 				softdep_error("clear_inodedeps: fsync1", error);
5743 		} else {
5744 			if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)))
5745 				softdep_error("clear_inodedeps: fsync2", error);
5746 			VI_LOCK(vp);
5747 			drain_output(vp, 0);
5748 			VI_UNLOCK(vp);
5749 		}
5750 		vput(vp);
5751 		vn_finished_write(mp);
5752 		ACQUIRE_LOCK(&lk);
5753 	}
5754 	FREE_LOCK(&lk);
5755 }
5756 
5757 /*
5758  * Function to determine if the buffer has outstanding dependencies
5759  * that will cause a roll-back if the buffer is written. If wantcount
5760  * is set, return number of dependencies, otherwise just yes or no.
5761  */
5762 static int
5763 softdep_count_dependencies(bp, wantcount)
5764 	struct buf *bp;
5765 	int wantcount;
5766 {
5767 	struct worklist *wk;
5768 	struct inodedep *inodedep;
5769 	struct indirdep *indirdep;
5770 	struct allocindir *aip;
5771 	struct pagedep *pagedep;
5772 	struct diradd *dap;
5773 	int i, retval;
5774 
5775 	retval = 0;
5776 	ACQUIRE_LOCK(&lk);
5777 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5778 		switch (wk->wk_type) {
5779 
5780 		case D_INODEDEP:
5781 			inodedep = WK_INODEDEP(wk);
5782 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5783 				/* bitmap allocation dependency */
5784 				retval += 1;
5785 				if (!wantcount)
5786 					goto out;
5787 			}
5788 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
5789 				/* direct block pointer dependency */
5790 				retval += 1;
5791 				if (!wantcount)
5792 					goto out;
5793 			}
5794 			if (TAILQ_FIRST(&inodedep->id_extupdt)) {
5795 				/* direct block pointer dependency */
5796 				retval += 1;
5797 				if (!wantcount)
5798 					goto out;
5799 			}
5800 			continue;
5801 
5802 		case D_INDIRDEP:
5803 			indirdep = WK_INDIRDEP(wk);
5804 
5805 			LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
5806 				/* indirect block pointer dependency */
5807 				retval += 1;
5808 				if (!wantcount)
5809 					goto out;
5810 			}
5811 			continue;
5812 
5813 		case D_PAGEDEP:
5814 			pagedep = WK_PAGEDEP(wk);
5815 			for (i = 0; i < DAHASHSZ; i++) {
5816 
5817 				LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
5818 					/* directory entry dependency */
5819 					retval += 1;
5820 					if (!wantcount)
5821 						goto out;
5822 				}
5823 			}
5824 			continue;
5825 
5826 		case D_BMSAFEMAP:
5827 		case D_ALLOCDIRECT:
5828 		case D_ALLOCINDIR:
5829 		case D_MKDIR:
5830 			/* never a dependency on these blocks */
5831 			continue;
5832 
5833 		default:
5834 			FREE_LOCK(&lk);
5835 			panic("softdep_check_for_rollback: Unexpected type %s",
5836 			    TYPENAME(wk->wk_type));
5837 			/* NOTREACHED */
5838 		}
5839 	}
5840 out:
5841 	FREE_LOCK(&lk);
5842 	return retval;
5843 }
5844 
5845 /*
5846  * Acquire exclusive access to a buffer.
5847  * Must be called with splbio blocked.
5848  * Return acquired buffer or NULL on failure.  mtx, if provided, will be
5849  * released on success but held on failure.
5850  */
5851 static struct buf *
5852 getdirtybuf(bpp, mtx, waitfor)
5853 	struct buf **bpp;
5854 	struct mtx *mtx;
5855 	int waitfor;
5856 {
5857 	struct buf *bp;
5858 	int error;
5859 
5860 	/*
5861 	 * XXX This code and the code that calls it need to be reviewed to
5862 	 * verify its use of the vnode interlock.
5863 	 */
5864 
5865 	for (;;) {
5866 		if ((bp = *bpp) == NULL)
5867 			return (0);
5868 		if (bp->b_vp == NULL)
5869 			backtrace();
5870 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) == 0) {
5871 			if ((bp->b_vflags & BV_BKGRDINPROG) == 0)
5872 				break;
5873 			BUF_UNLOCK(bp);
5874 			if (waitfor != MNT_WAIT)
5875 				return (NULL);
5876 			/*
5877 			 * The mtx argument must be bp->b_vp's mutex in
5878 			 * this case.
5879 			 */
5880 #ifdef	DEBUG_VFS_LOCKS
5881 			if (bp->b_vp->v_type != VCHR)
5882 				ASSERT_VI_LOCKED(bp->b_vp, "getdirtybuf");
5883 #endif
5884 			bp->b_vflags |= BV_BKGRDWAIT;
5885 			interlocked_sleep(&lk, SLEEP, &bp->b_xflags, mtx,
5886 			    PRIBIO, "getbuf", 0);
5887 			continue;
5888 		}
5889 		if (waitfor != MNT_WAIT)
5890 			return (NULL);
5891 		if (mtx) {
5892 			error = interlocked_sleep(&lk, LOCKBUF, bp, mtx,
5893 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, 0, 0);
5894 			mtx_lock(mtx);
5895 		} else
5896 			error = interlocked_sleep(&lk, LOCKBUF, bp, NULL,
5897 			    LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0);
5898 		if (error != ENOLCK) {
5899 			FREE_LOCK(&lk);
5900 			panic("getdirtybuf: inconsistent lock");
5901 		}
5902 	}
5903 	if ((bp->b_flags & B_DELWRI) == 0) {
5904 		BUF_UNLOCK(bp);
5905 		return (NULL);
5906 	}
5907 	if (mtx)
5908 		mtx_unlock(mtx);
5909 	bremfree(bp);
5910 	return (bp);
5911 }
5912 
5913 /*
5914  * Wait for pending output on a vnode to complete.
5915  * Must be called with vnode lock and interlock locked.
5916  */
5917 static void
5918 drain_output(vp, islocked)
5919 	struct vnode *vp;
5920 	int islocked;
5921 {
5922 	ASSERT_VOP_LOCKED(vp, "drain_output");
5923 	ASSERT_VI_LOCKED(vp, "drain_output");
5924 
5925 	if (!islocked)
5926 		ACQUIRE_LOCK(&lk);
5927 	while (vp->v_numoutput) {
5928 		vp->v_iflag |= VI_BWAIT;
5929 		interlocked_sleep(&lk, SLEEP, (caddr_t)&vp->v_numoutput,
5930 		    VI_MTX(vp), PRIBIO + 1, "drainvp", 0);
5931 	}
5932 	if (!islocked)
5933 		FREE_LOCK(&lk);
5934 }
5935 
5936 /*
5937  * Called whenever a buffer that is being invalidated or reallocated
5938  * contains dependencies. This should only happen if an I/O error has
5939  * occurred. The routine is called with the buffer locked.
5940  */
5941 static void
5942 softdep_deallocate_dependencies(bp)
5943 	struct buf *bp;
5944 {
5945 
5946 	if ((bp->b_ioflags & BIO_ERROR) == 0)
5947 		panic("softdep_deallocate_dependencies: dangling deps");
5948 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
5949 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
5950 }
5951 
5952 /*
5953  * Function to handle asynchronous write errors in the filesystem.
5954  */
5955 static void
5956 softdep_error(func, error)
5957 	char *func;
5958 	int error;
5959 {
5960 
5961 	/* XXX should do something better! */
5962 	printf("%s: got error %d while accessing filesystem\n", func, error);
5963 }
5964