xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision a3e8fd0b7f663db7eafff527d5c3ca3bcfa8a537)
1 /*
2  * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
3  *
4  * The soft updates code is derived from the appendix of a University
5  * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
6  * "Soft Updates: A Solution to the Metadata Update Problem in File
7  * Systems", CSE-TR-254-95, August 1995).
8  *
9  * Further information about soft updates can be obtained from:
10  *
11  *	Marshall Kirk McKusick		http://www.mckusick.com/softdep/
12  *	1614 Oxford Street		mckusick@mckusick.com
13  *	Berkeley, CA 94709-1608		+1-510-843-9542
14  *	USA
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  *
20  * 1. Redistributions of source code must retain the above copyright
21  *    notice, this list of conditions and the following disclaimer.
22  * 2. Redistributions in binary form must reproduce the above copyright
23  *    notice, this list of conditions and the following disclaimer in the
24  *    documentation and/or other materials provided with the distribution.
25  *
26  * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
27  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
28  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
29  * DISCLAIMED.  IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
30  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)ffs_softdep.c	9.59 (McKusick) 6/21/00
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 /*
45  * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
46  */
47 #ifndef DIAGNOSTIC
48 #define DIAGNOSTIC
49 #endif
50 #ifndef DEBUG
51 #define DEBUG
52 #endif
53 
54 #include <sys/param.h>
55 #include <sys/kernel.h>
56 #include <sys/systm.h>
57 #include <sys/stdint.h>
58 #include <sys/bio.h>
59 #include <sys/buf.h>
60 #include <sys/malloc.h>
61 #include <sys/mount.h>
62 #include <sys/proc.h>
63 #include <sys/stat.h>
64 #include <sys/syslog.h>
65 #include <sys/vnode.h>
66 #include <sys/conf.h>
67 #include <ufs/ufs/dir.h>
68 #include <ufs/ufs/extattr.h>
69 #include <ufs/ufs/quota.h>
70 #include <ufs/ufs/inode.h>
71 #include <ufs/ufs/ufsmount.h>
72 #include <ufs/ffs/fs.h>
73 #include <ufs/ffs/softdep.h>
74 #include <ufs/ffs/ffs_extern.h>
75 #include <ufs/ufs/ufs_extern.h>
76 
77 /*
78  * These definitions need to be adapted to the system to which
79  * this file is being ported.
80  */
81 /*
82  * malloc types defined for the softdep system.
83  */
84 static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
85 static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
86 static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
87 static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
88 static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
89 static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
90 static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
91 static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
92 static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
93 static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
94 static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
95 static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
96 static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
97 static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block");
98 
99 #define M_SOFTDEP_FLAGS	(M_WAITOK | M_USE_RESERVE)
100 
101 #define	D_PAGEDEP	0
102 #define	D_INODEDEP	1
103 #define	D_NEWBLK	2
104 #define	D_BMSAFEMAP	3
105 #define	D_ALLOCDIRECT	4
106 #define	D_INDIRDEP	5
107 #define	D_ALLOCINDIR	6
108 #define	D_FREEFRAG	7
109 #define	D_FREEBLKS	8
110 #define	D_FREEFILE	9
111 #define	D_DIRADD	10
112 #define	D_MKDIR		11
113 #define	D_DIRREM	12
114 #define	D_NEWDIRBLK	13
115 #define	D_LAST		D_NEWDIRBLK
116 
117 /*
118  * translate from workitem type to memory type
119  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
120  */
121 static struct malloc_type *memtype[] = {
122 	M_PAGEDEP,
123 	M_INODEDEP,
124 	M_NEWBLK,
125 	M_BMSAFEMAP,
126 	M_ALLOCDIRECT,
127 	M_INDIRDEP,
128 	M_ALLOCINDIR,
129 	M_FREEFRAG,
130 	M_FREEBLKS,
131 	M_FREEFILE,
132 	M_DIRADD,
133 	M_MKDIR,
134 	M_DIRREM,
135 	M_NEWDIRBLK
136 };
137 
138 #define DtoM(type) (memtype[type])
139 
140 /*
141  * Names of malloc types.
142  */
143 #define TYPENAME(type)  \
144 	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
145 /*
146  * End system adaptaion definitions.
147  */
148 
149 /*
150  * Internal function prototypes.
151  */
152 static	void softdep_error(char *, int);
153 static	void drain_output(struct vnode *, int);
154 static	int getdirtybuf(struct buf **, int);
155 static	void clear_remove(struct thread *);
156 static	void clear_inodedeps(struct thread *);
157 static	int flush_pagedep_deps(struct vnode *, struct mount *,
158 	    struct diraddhd *);
159 static	int flush_inodedep_deps(struct fs *, ino_t);
160 static	int flush_deplist(struct allocdirectlst *, int, int *);
161 static	int handle_written_filepage(struct pagedep *, struct buf *);
162 static  void diradd_inode_written(struct diradd *, struct inodedep *);
163 static	int handle_written_inodeblock(struct inodedep *, struct buf *);
164 static	void handle_allocdirect_partdone(struct allocdirect *);
165 static	void handle_allocindir_partdone(struct allocindir *);
166 static	void initiate_write_filepage(struct pagedep *, struct buf *);
167 static	void handle_written_mkdir(struct mkdir *, int);
168 static	void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
169 static	void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
170 static	void handle_workitem_freefile(struct freefile *);
171 static	void handle_workitem_remove(struct dirrem *, struct vnode *);
172 static	struct dirrem *newdirrem(struct buf *, struct inode *,
173 	    struct inode *, int, struct dirrem **);
174 static	void free_diradd(struct diradd *);
175 static	void free_allocindir(struct allocindir *, struct inodedep *);
176 static	void free_newdirblk(struct newdirblk *);
177 static	int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t,
178 	    ufs2_daddr_t *);
179 static	void deallocate_dependencies(struct buf *, struct inodedep *);
180 static	void free_allocdirect(struct allocdirectlst *,
181 	    struct allocdirect *, int);
182 static	int check_inode_unwritten(struct inodedep *);
183 static	int free_inodedep(struct inodedep *);
184 static	void handle_workitem_freeblocks(struct freeblks *, int);
185 static	void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
186 static	void setup_allocindir_phase2(struct buf *, struct inode *,
187 	    struct allocindir *);
188 static	struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
189 	    ufs2_daddr_t);
190 static	void handle_workitem_freefrag(struct freefrag *);
191 static	struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long);
192 static	void allocdirect_merge(struct allocdirectlst *,
193 	    struct allocdirect *, struct allocdirect *);
194 static	struct bmsafemap *bmsafemap_lookup(struct buf *);
195 static	int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **);
196 static	int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
197 static	int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **);
198 static	void pause_timer(void *);
199 static	int request_cleanup(int, int);
200 static	int process_worklist_item(struct mount *, int);
201 static	void add_to_worklist(struct worklist *);
202 
203 /*
204  * Exported softdep operations.
205  */
206 static	void softdep_disk_io_initiation(struct buf *);
207 static	void softdep_disk_write_complete(struct buf *);
208 static	void softdep_deallocate_dependencies(struct buf *);
209 static	void softdep_move_dependencies(struct buf *, struct buf *);
210 static	int softdep_count_dependencies(struct buf *bp, int);
211 
212 /*
213  * Locking primitives.
214  *
215  * For a uniprocessor, all we need to do is protect against disk
216  * interrupts. For a multiprocessor, this lock would have to be
217  * a mutex. A single mutex is used throughout this file, though
218  * finer grain locking could be used if contention warranted it.
219  *
220  * For a multiprocessor, the sleep call would accept a lock and
221  * release it after the sleep processing was complete. In a uniprocessor
222  * implementation there is no such interlock, so we simple mark
223  * the places where it needs to be done with the `interlocked' form
224  * of the lock calls. Since the uniprocessor sleep already interlocks
225  * the spl, there is nothing that really needs to be done.
226  */
227 #ifndef /* NOT */ DEBUG
228 static struct lockit {
229 	int	lkt_spl;
230 } lk = { 0 };
231 #define ACQUIRE_LOCK(lk)		(lk)->lkt_spl = splbio()
232 #define FREE_LOCK(lk)			splx((lk)->lkt_spl)
233 
234 #else /* DEBUG */
235 #define NOHOLDER	((struct thread *)-1)
236 #define SPECIAL_FLAG	((struct thread *)-2)
237 static struct lockit {
238 	int	lkt_spl;
239 	struct	thread *lkt_held;
240 } lk = { 0, NOHOLDER };
241 static int lockcnt;
242 
243 static	void acquire_lock(struct lockit *);
244 static	void free_lock(struct lockit *);
245 void	softdep_panic(char *);
246 
247 #define ACQUIRE_LOCK(lk)		acquire_lock(lk)
248 #define FREE_LOCK(lk)			free_lock(lk)
249 
250 static void
251 acquire_lock(lk)
252 	struct lockit *lk;
253 {
254 	struct thread *holder;
255 
256 	if (lk->lkt_held != NOHOLDER) {
257 		holder = lk->lkt_held;
258 		FREE_LOCK(lk);
259 		if (holder == curthread)
260 			panic("softdep_lock: locking against myself");
261 		else
262 			panic("softdep_lock: lock held by %p", holder);
263 	}
264 	lk->lkt_spl = splbio();
265 	lk->lkt_held = curthread;
266 	lockcnt++;
267 }
268 
269 static void
270 free_lock(lk)
271 	struct lockit *lk;
272 {
273 
274 	if (lk->lkt_held == NOHOLDER)
275 		panic("softdep_unlock: lock not held");
276 	lk->lkt_held = NOHOLDER;
277 	splx(lk->lkt_spl);
278 }
279 
280 /*
281  * Function to release soft updates lock and panic.
282  */
283 void
284 softdep_panic(msg)
285 	char *msg;
286 {
287 
288 	if (lk.lkt_held != NOHOLDER)
289 		FREE_LOCK(&lk);
290 	panic(msg);
291 }
292 #endif /* DEBUG */
293 
294 static	int interlocked_sleep(struct lockit *, int, void *, struct mtx *, int,
295 	    const char *, int);
296 
297 /*
298  * When going to sleep, we must save our SPL so that it does
299  * not get lost if some other process uses the lock while we
300  * are sleeping. We restore it after we have slept. This routine
301  * wraps the interlocking with functions that sleep. The list
302  * below enumerates the available set of operations.
303  */
304 #define	UNKNOWN		0
305 #define	SLEEP		1
306 #define	LOCKBUF		2
307 
308 static int
309 interlocked_sleep(lk, op, ident, mtx, flags, wmesg, timo)
310 	struct lockit *lk;
311 	int op;
312 	void *ident;
313 	struct mtx *mtx;
314 	int flags;
315 	const char *wmesg;
316 	int timo;
317 {
318 	struct thread *holder;
319 	int s, retval;
320 
321 	s = lk->lkt_spl;
322 #	ifdef DEBUG
323 	if (lk->lkt_held == NOHOLDER)
324 		panic("interlocked_sleep: lock not held");
325 	lk->lkt_held = NOHOLDER;
326 #	endif /* DEBUG */
327 	switch (op) {
328 	case SLEEP:
329 		retval = msleep(ident, mtx, flags, wmesg, timo);
330 		break;
331 	case LOCKBUF:
332 		retval = BUF_LOCK((struct buf *)ident, flags);
333 		break;
334 	default:
335 		panic("interlocked_sleep: unknown operation");
336 	}
337 #	ifdef DEBUG
338 	if (lk->lkt_held != NOHOLDER) {
339 		holder = lk->lkt_held;
340 		FREE_LOCK(lk);
341 		if (holder == curthread)
342 			panic("interlocked_sleep: locking against self");
343 		else
344 			panic("interlocked_sleep: lock held by %p", holder);
345 	}
346 	lk->lkt_held = curthread;
347 	lockcnt++;
348 #	endif /* DEBUG */
349 	lk->lkt_spl = s;
350 	return (retval);
351 }
352 
353 /*
354  * Place holder for real semaphores.
355  */
356 struct sema {
357 	int	value;
358 	struct	thread *holder;
359 	char	*name;
360 	int	prio;
361 	int	timo;
362 };
363 static	void sema_init(struct sema *, char *, int, int);
364 static	int sema_get(struct sema *, struct lockit *);
365 static	void sema_release(struct sema *);
366 
367 static void
368 sema_init(semap, name, prio, timo)
369 	struct sema *semap;
370 	char *name;
371 	int prio, timo;
372 {
373 
374 	semap->holder = NOHOLDER;
375 	semap->value = 0;
376 	semap->name = name;
377 	semap->prio = prio;
378 	semap->timo = timo;
379 }
380 
381 static int
382 sema_get(semap, interlock)
383 	struct sema *semap;
384 	struct lockit *interlock;
385 {
386 
387 	if (semap->value++ > 0) {
388 		if (interlock != NULL) {
389 			interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
390 			    NULL, semap->prio, semap->name,
391 			    semap->timo);
392 			FREE_LOCK(interlock);
393 		} else {
394 			tsleep((caddr_t)semap, semap->prio, semap->name,
395 			    semap->timo);
396 		}
397 		return (0);
398 	}
399 	semap->holder = curthread;
400 	if (interlock != NULL)
401 		FREE_LOCK(interlock);
402 	return (1);
403 }
404 
405 static void
406 sema_release(semap)
407 	struct sema *semap;
408 {
409 
410 	if (semap->value <= 0 || semap->holder != curthread) {
411 		if (lk.lkt_held != NOHOLDER)
412 			FREE_LOCK(&lk);
413 		panic("sema_release: not held");
414 	}
415 	if (--semap->value > 0) {
416 		semap->value = 0;
417 		wakeup(semap);
418 	}
419 	semap->holder = NOHOLDER;
420 }
421 
422 /*
423  * Worklist queue management.
424  * These routines require that the lock be held.
425  */
426 #ifndef /* NOT */ DEBUG
427 #define WORKLIST_INSERT(head, item) do {	\
428 	(item)->wk_state |= ONWORKLIST;		\
429 	LIST_INSERT_HEAD(head, item, wk_list);	\
430 } while (0)
431 #define WORKLIST_REMOVE(item) do {		\
432 	(item)->wk_state &= ~ONWORKLIST;	\
433 	LIST_REMOVE(item, wk_list);		\
434 } while (0)
435 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
436 
437 #else /* DEBUG */
438 static	void worklist_insert(struct workhead *, struct worklist *);
439 static	void worklist_remove(struct worklist *);
440 static	void workitem_free(struct worklist *, int);
441 
442 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
443 #define WORKLIST_REMOVE(item) worklist_remove(item)
444 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
445 
446 static void
447 worklist_insert(head, item)
448 	struct workhead *head;
449 	struct worklist *item;
450 {
451 
452 	if (lk.lkt_held == NOHOLDER)
453 		panic("worklist_insert: lock not held");
454 	if (item->wk_state & ONWORKLIST) {
455 		FREE_LOCK(&lk);
456 		panic("worklist_insert: already on list");
457 	}
458 	item->wk_state |= ONWORKLIST;
459 	LIST_INSERT_HEAD(head, item, wk_list);
460 }
461 
462 static void
463 worklist_remove(item)
464 	struct worklist *item;
465 {
466 
467 	if (lk.lkt_held == NOHOLDER)
468 		panic("worklist_remove: lock not held");
469 	if ((item->wk_state & ONWORKLIST) == 0) {
470 		FREE_LOCK(&lk);
471 		panic("worklist_remove: not on list");
472 	}
473 	item->wk_state &= ~ONWORKLIST;
474 	LIST_REMOVE(item, wk_list);
475 }
476 
477 static void
478 workitem_free(item, type)
479 	struct worklist *item;
480 	int type;
481 {
482 
483 	if (item->wk_state & ONWORKLIST) {
484 		if (lk.lkt_held != NOHOLDER)
485 			FREE_LOCK(&lk);
486 		panic("workitem_free: still on list");
487 	}
488 	if (item->wk_type != type) {
489 		if (lk.lkt_held != NOHOLDER)
490 			FREE_LOCK(&lk);
491 		panic("workitem_free: type mismatch");
492 	}
493 	FREE(item, DtoM(type));
494 }
495 #endif /* DEBUG */
496 
497 /*
498  * Workitem queue management
499  */
500 static struct workhead softdep_workitem_pending;
501 static int num_on_worklist;	/* number of worklist items to be processed */
502 static int softdep_worklist_busy; /* 1 => trying to do unmount */
503 static int softdep_worklist_req; /* serialized waiters */
504 static int max_softdeps;	/* maximum number of structs before slowdown */
505 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
506 static int proc_waiting;	/* tracks whether we have a timeout posted */
507 static int *stat_countp;	/* statistic to count in proc_waiting timeout */
508 static struct callout_handle handle; /* handle on posted proc_waiting timeout */
509 static struct thread *filesys_syncer; /* proc of filesystem syncer process */
510 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
511 #define FLUSH_INODES		1
512 static int req_clear_remove;	/* syncer process flush some freeblks */
513 #define FLUSH_REMOVE		2
514 #define FLUSH_REMOVE_WAIT	3
515 /*
516  * runtime statistics
517  */
518 static int stat_worklist_push;	/* number of worklist cleanups */
519 static int stat_blk_limit_push;	/* number of times block limit neared */
520 static int stat_ino_limit_push;	/* number of times inode limit neared */
521 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
522 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
523 static int stat_sync_limit_hit;	/* number of synchronous slowdowns imposed */
524 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
525 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
526 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
527 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
528 #ifdef DEBUG
529 #include <vm/vm.h>
530 #include <sys/sysctl.h>
531 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
532 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
533 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
534 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
535 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
536 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
537 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
538 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
539 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
540 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
541 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
542 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
543 #endif /* DEBUG */
544 
545 /*
546  * Add an item to the end of the work queue.
547  * This routine requires that the lock be held.
548  * This is the only routine that adds items to the list.
549  * The following routine is the only one that removes items
550  * and does so in order from first to last.
551  */
552 static void
553 add_to_worklist(wk)
554 	struct worklist *wk;
555 {
556 	static struct worklist *worklist_tail;
557 
558 	if (wk->wk_state & ONWORKLIST) {
559 		if (lk.lkt_held != NOHOLDER)
560 			FREE_LOCK(&lk);
561 		panic("add_to_worklist: already on list");
562 	}
563 	wk->wk_state |= ONWORKLIST;
564 	if (LIST_FIRST(&softdep_workitem_pending) == NULL)
565 		LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
566 	else
567 		LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
568 	worklist_tail = wk;
569 	num_on_worklist += 1;
570 }
571 
572 /*
573  * Process that runs once per second to handle items in the background queue.
574  *
575  * Note that we ensure that everything is done in the order in which they
576  * appear in the queue. The code below depends on this property to ensure
577  * that blocks of a file are freed before the inode itself is freed. This
578  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
579  * until all the old ones have been purged from the dependency lists.
580  */
581 int
582 softdep_process_worklist(matchmnt)
583 	struct mount *matchmnt;
584 {
585 	struct thread *td = curthread;
586 	int cnt, matchcnt, loopcount;
587 	long starttime;
588 
589 	/*
590 	 * Record the process identifier of our caller so that we can give
591 	 * this process preferential treatment in request_cleanup below.
592 	 */
593 	filesys_syncer = td;
594 	matchcnt = 0;
595 
596 	/*
597 	 * There is no danger of having multiple processes run this
598 	 * code, but we have to single-thread it when softdep_flushfiles()
599 	 * is in operation to get an accurate count of the number of items
600 	 * related to its mount point that are in the list.
601 	 */
602 	if (matchmnt == NULL) {
603 		if (softdep_worklist_busy < 0)
604 			return(-1);
605 		softdep_worklist_busy += 1;
606 	}
607 
608 	/*
609 	 * If requested, try removing inode or removal dependencies.
610 	 */
611 	if (req_clear_inodedeps) {
612 		clear_inodedeps(td);
613 		req_clear_inodedeps -= 1;
614 		wakeup_one(&proc_waiting);
615 	}
616 	if (req_clear_remove) {
617 		clear_remove(td);
618 		req_clear_remove -= 1;
619 		wakeup_one(&proc_waiting);
620 	}
621 	loopcount = 1;
622 	starttime = time_second;
623 	while (num_on_worklist > 0) {
624 		if ((cnt = process_worklist_item(matchmnt, 0)) == -1)
625 			break;
626 		else
627 			matchcnt += cnt;
628 
629 		/*
630 		 * If a umount operation wants to run the worklist
631 		 * accurately, abort.
632 		 */
633 		if (softdep_worklist_req && matchmnt == NULL) {
634 			matchcnt = -1;
635 			break;
636 		}
637 
638 		/*
639 		 * If requested, try removing inode or removal dependencies.
640 		 */
641 		if (req_clear_inodedeps) {
642 			clear_inodedeps(td);
643 			req_clear_inodedeps -= 1;
644 			wakeup_one(&proc_waiting);
645 		}
646 		if (req_clear_remove) {
647 			clear_remove(td);
648 			req_clear_remove -= 1;
649 			wakeup_one(&proc_waiting);
650 		}
651 		/*
652 		 * We do not generally want to stop for buffer space, but if
653 		 * we are really being a buffer hog, we will stop and wait.
654 		 */
655 		if (loopcount++ % 128 == 0)
656 			bwillwrite();
657 		/*
658 		 * Never allow processing to run for more than one
659 		 * second. Otherwise the other syncer tasks may get
660 		 * excessively backlogged.
661 		 */
662 		if (starttime != time_second && matchmnt == NULL) {
663 			matchcnt = -1;
664 			break;
665 		}
666 	}
667 	if (matchmnt == NULL) {
668 		softdep_worklist_busy -= 1;
669 		if (softdep_worklist_req && softdep_worklist_busy == 0)
670 			wakeup(&softdep_worklist_req);
671 	}
672 	return (matchcnt);
673 }
674 
675 /*
676  * Process one item on the worklist.
677  */
678 static int
679 process_worklist_item(matchmnt, flags)
680 	struct mount *matchmnt;
681 	int flags;
682 {
683 	struct worklist *wk;
684 	struct mount *mp;
685 	struct vnode *vp;
686 	int matchcnt = 0;
687 
688 	/*
689 	 * If we are being called because of a process doing a
690 	 * copy-on-write, then it is not safe to write as we may
691 	 * recurse into the copy-on-write routine.
692 	 */
693 	if (curthread->td_proc->p_flag & P_COWINPROGRESS)
694 		return (-1);
695 	ACQUIRE_LOCK(&lk);
696 	/*
697 	 * Normally we just process each item on the worklist in order.
698 	 * However, if we are in a situation where we cannot lock any
699 	 * inodes, we have to skip over any dirrem requests whose
700 	 * vnodes are resident and locked.
701 	 */
702 	vp = NULL;
703 	LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
704 		if (wk->wk_state & INPROGRESS)
705 			continue;
706 		if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
707 			break;
708 		wk->wk_state |= INPROGRESS;
709 		FREE_LOCK(&lk);
710 		VFS_VGET(WK_DIRREM(wk)->dm_mnt, WK_DIRREM(wk)->dm_oldinum,
711 		    LK_NOWAIT | LK_EXCLUSIVE, &vp);
712 		ACQUIRE_LOCK(&lk);
713 		wk->wk_state &= ~INPROGRESS;
714 		if (vp != NULL)
715 			break;
716 	}
717 	if (wk == 0) {
718 		FREE_LOCK(&lk);
719 		return (-1);
720 	}
721 	WORKLIST_REMOVE(wk);
722 	num_on_worklist -= 1;
723 	FREE_LOCK(&lk);
724 	switch (wk->wk_type) {
725 
726 	case D_DIRREM:
727 		/* removal of a directory entry */
728 		mp = WK_DIRREM(wk)->dm_mnt;
729 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
730 			panic("%s: dirrem on suspended filesystem",
731 				"process_worklist_item");
732 		if (mp == matchmnt)
733 			matchcnt += 1;
734 		handle_workitem_remove(WK_DIRREM(wk), vp);
735 		break;
736 
737 	case D_FREEBLKS:
738 		/* releasing blocks and/or fragments from a file */
739 		mp = WK_FREEBLKS(wk)->fb_mnt;
740 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
741 			panic("%s: freeblks on suspended filesystem",
742 				"process_worklist_item");
743 		if (mp == matchmnt)
744 			matchcnt += 1;
745 		handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT);
746 		break;
747 
748 	case D_FREEFRAG:
749 		/* releasing a fragment when replaced as a file grows */
750 		mp = WK_FREEFRAG(wk)->ff_mnt;
751 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
752 			panic("%s: freefrag on suspended filesystem",
753 				"process_worklist_item");
754 		if (mp == matchmnt)
755 			matchcnt += 1;
756 		handle_workitem_freefrag(WK_FREEFRAG(wk));
757 		break;
758 
759 	case D_FREEFILE:
760 		/* releasing an inode when its link count drops to 0 */
761 		mp = WK_FREEFILE(wk)->fx_mnt;
762 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
763 			panic("%s: freefile on suspended filesystem",
764 				"process_worklist_item");
765 		if (mp == matchmnt)
766 			matchcnt += 1;
767 		handle_workitem_freefile(WK_FREEFILE(wk));
768 		break;
769 
770 	default:
771 		panic("%s_process_worklist: Unknown type %s",
772 		    "softdep", TYPENAME(wk->wk_type));
773 		/* NOTREACHED */
774 	}
775 	return (matchcnt);
776 }
777 
778 /*
779  * Move dependencies from one buffer to another.
780  */
781 static void
782 softdep_move_dependencies(oldbp, newbp)
783 	struct buf *oldbp;
784 	struct buf *newbp;
785 {
786 	struct worklist *wk, *wktail;
787 
788 	if (LIST_FIRST(&newbp->b_dep) != NULL)
789 		panic("softdep_move_dependencies: need merge code");
790 	wktail = 0;
791 	ACQUIRE_LOCK(&lk);
792 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
793 		LIST_REMOVE(wk, wk_list);
794 		if (wktail == 0)
795 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
796 		else
797 			LIST_INSERT_AFTER(wktail, wk, wk_list);
798 		wktail = wk;
799 	}
800 	FREE_LOCK(&lk);
801 }
802 
803 /*
804  * Purge the work list of all items associated with a particular mount point.
805  */
806 int
807 softdep_flushworklist(oldmnt, countp, td)
808 	struct mount *oldmnt;
809 	int *countp;
810 	struct thread *td;
811 {
812 	struct vnode *devvp;
813 	int count, error = 0;
814 
815 	/*
816 	 * Await our turn to clear out the queue, then serialize access.
817 	 */
818 	while (softdep_worklist_busy) {
819 		softdep_worklist_req += 1;
820 		tsleep(&softdep_worklist_req, PRIBIO, "softflush", 0);
821 		softdep_worklist_req -= 1;
822 	}
823 	softdep_worklist_busy = -1;
824 	/*
825 	 * Alternately flush the block device associated with the mount
826 	 * point and process any dependencies that the flushing
827 	 * creates. We continue until no more worklist dependencies
828 	 * are found.
829 	 */
830 	*countp = 0;
831 	devvp = VFSTOUFS(oldmnt)->um_devvp;
832 	while ((count = softdep_process_worklist(oldmnt)) > 0) {
833 		*countp += count;
834 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td);
835 		error = VOP_FSYNC(devvp, td->td_ucred, MNT_WAIT, td);
836 		VOP_UNLOCK(devvp, 0, td);
837 		if (error)
838 			break;
839 	}
840 	softdep_worklist_busy = 0;
841 	if (softdep_worklist_req)
842 		wakeup(&softdep_worklist_req);
843 	return (error);
844 }
845 
846 /*
847  * Flush all vnodes and worklist items associated with a specified mount point.
848  */
849 int
850 softdep_flushfiles(oldmnt, flags, td)
851 	struct mount *oldmnt;
852 	int flags;
853 	struct thread *td;
854 {
855 	int error, count, loopcnt;
856 
857 	error = 0;
858 
859 	/*
860 	 * Alternately flush the vnodes associated with the mount
861 	 * point and process any dependencies that the flushing
862 	 * creates. In theory, this loop can happen at most twice,
863 	 * but we give it a few extra just to be sure.
864 	 */
865 	for (loopcnt = 10; loopcnt > 0; loopcnt--) {
866 		/*
867 		 * Do another flush in case any vnodes were brought in
868 		 * as part of the cleanup operations.
869 		 */
870 		if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
871 			break;
872 		if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 ||
873 		    count == 0)
874 			break;
875 	}
876 	/*
877 	 * If we are unmounting then it is an error to fail. If we
878 	 * are simply trying to downgrade to read-only, then filesystem
879 	 * activity can keep us busy forever, so we just fail with EBUSY.
880 	 */
881 	if (loopcnt == 0) {
882 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
883 			panic("softdep_flushfiles: looping");
884 		error = EBUSY;
885 	}
886 	return (error);
887 }
888 
889 /*
890  * Structure hashing.
891  *
892  * There are three types of structures that can be looked up:
893  *	1) pagedep structures identified by mount point, inode number,
894  *	   and logical block.
895  *	2) inodedep structures identified by mount point and inode number.
896  *	3) newblk structures identified by mount point and
897  *	   physical block number.
898  *
899  * The "pagedep" and "inodedep" dependency structures are hashed
900  * separately from the file blocks and inodes to which they correspond.
901  * This separation helps when the in-memory copy of an inode or
902  * file block must be replaced. It also obviates the need to access
903  * an inode or file page when simply updating (or de-allocating)
904  * dependency structures. Lookup of newblk structures is needed to
905  * find newly allocated blocks when trying to associate them with
906  * their allocdirect or allocindir structure.
907  *
908  * The lookup routines optionally create and hash a new instance when
909  * an existing entry is not found.
910  */
911 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
912 #define NODELAY		0x0002	/* cannot do background work */
913 
914 /*
915  * Structures and routines associated with pagedep caching.
916  */
917 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
918 u_long	pagedep_hash;		/* size of hash table - 1 */
919 #define	PAGEDEP_HASH(mp, inum, lbn) \
920 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
921 	    pagedep_hash])
922 static struct sema pagedep_in_progress;
923 
924 /*
925  * Look up a pagedep. Return 1 if found, 0 if not found or found
926  * when asked to allocate but not associated with any buffer.
927  * If not found, allocate if DEPALLOC flag is passed.
928  * Found or allocated entry is returned in pagedeppp.
929  * This routine must be called with splbio interrupts blocked.
930  */
931 static int
932 pagedep_lookup(ip, lbn, flags, pagedeppp)
933 	struct inode *ip;
934 	ufs_lbn_t lbn;
935 	int flags;
936 	struct pagedep **pagedeppp;
937 {
938 	struct pagedep *pagedep;
939 	struct pagedep_hashhead *pagedephd;
940 	struct mount *mp;
941 	int i;
942 
943 #ifdef DEBUG
944 	if (lk.lkt_held == NOHOLDER)
945 		panic("pagedep_lookup: lock not held");
946 #endif
947 	mp = ITOV(ip)->v_mount;
948 	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
949 top:
950 	LIST_FOREACH(pagedep, pagedephd, pd_hash)
951 		if (ip->i_number == pagedep->pd_ino &&
952 		    lbn == pagedep->pd_lbn &&
953 		    mp == pagedep->pd_mnt)
954 			break;
955 	if (pagedep) {
956 		*pagedeppp = pagedep;
957 		if ((flags & DEPALLOC) != 0 &&
958 		    (pagedep->pd_state & ONWORKLIST) == 0)
959 			return (0);
960 		return (1);
961 	}
962 	if ((flags & DEPALLOC) == 0) {
963 		*pagedeppp = NULL;
964 		return (0);
965 	}
966 	if (sema_get(&pagedep_in_progress, &lk) == 0) {
967 		ACQUIRE_LOCK(&lk);
968 		goto top;
969 	}
970 	MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
971 		M_SOFTDEP_FLAGS|M_ZERO);
972 	pagedep->pd_list.wk_type = D_PAGEDEP;
973 	pagedep->pd_mnt = mp;
974 	pagedep->pd_ino = ip->i_number;
975 	pagedep->pd_lbn = lbn;
976 	LIST_INIT(&pagedep->pd_dirremhd);
977 	LIST_INIT(&pagedep->pd_pendinghd);
978 	for (i = 0; i < DAHASHSZ; i++)
979 		LIST_INIT(&pagedep->pd_diraddhd[i]);
980 	ACQUIRE_LOCK(&lk);
981 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
982 	sema_release(&pagedep_in_progress);
983 	*pagedeppp = pagedep;
984 	return (0);
985 }
986 
987 /*
988  * Structures and routines associated with inodedep caching.
989  */
990 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
991 static u_long	inodedep_hash;	/* size of hash table - 1 */
992 static long	num_inodedep;	/* number of inodedep allocated */
993 #define	INODEDEP_HASH(fs, inum) \
994       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
995 static struct sema inodedep_in_progress;
996 
997 /*
998  * Look up a inodedep. Return 1 if found, 0 if not found.
999  * If not found, allocate if DEPALLOC flag is passed.
1000  * Found or allocated entry is returned in inodedeppp.
1001  * This routine must be called with splbio interrupts blocked.
1002  */
1003 static int
1004 inodedep_lookup(fs, inum, flags, inodedeppp)
1005 	struct fs *fs;
1006 	ino_t inum;
1007 	int flags;
1008 	struct inodedep **inodedeppp;
1009 {
1010 	struct inodedep *inodedep;
1011 	struct inodedep_hashhead *inodedephd;
1012 	int firsttry;
1013 
1014 #ifdef DEBUG
1015 	if (lk.lkt_held == NOHOLDER)
1016 		panic("inodedep_lookup: lock not held");
1017 #endif
1018 	firsttry = 1;
1019 	inodedephd = INODEDEP_HASH(fs, inum);
1020 top:
1021 	LIST_FOREACH(inodedep, inodedephd, id_hash)
1022 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
1023 			break;
1024 	if (inodedep) {
1025 		*inodedeppp = inodedep;
1026 		return (1);
1027 	}
1028 	if ((flags & DEPALLOC) == 0) {
1029 		*inodedeppp = NULL;
1030 		return (0);
1031 	}
1032 	/*
1033 	 * If we are over our limit, try to improve the situation.
1034 	 */
1035 	if (num_inodedep > max_softdeps && firsttry && (flags & NODELAY) == 0 &&
1036 	    request_cleanup(FLUSH_INODES, 1)) {
1037 		firsttry = 0;
1038 		goto top;
1039 	}
1040 	if (sema_get(&inodedep_in_progress, &lk) == 0) {
1041 		ACQUIRE_LOCK(&lk);
1042 		goto top;
1043 	}
1044 	num_inodedep += 1;
1045 	MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1046 		M_INODEDEP, M_SOFTDEP_FLAGS);
1047 	inodedep->id_list.wk_type = D_INODEDEP;
1048 	inodedep->id_fs = fs;
1049 	inodedep->id_ino = inum;
1050 	inodedep->id_state = ALLCOMPLETE;
1051 	inodedep->id_nlinkdelta = 0;
1052 	inodedep->id_savedino1 = NULL;
1053 	inodedep->id_savedsize = -1;
1054 	inodedep->id_savedextsize = -1;
1055 	inodedep->id_buf = NULL;
1056 	LIST_INIT(&inodedep->id_pendinghd);
1057 	LIST_INIT(&inodedep->id_inowait);
1058 	LIST_INIT(&inodedep->id_bufwait);
1059 	TAILQ_INIT(&inodedep->id_inoupdt);
1060 	TAILQ_INIT(&inodedep->id_newinoupdt);
1061 	TAILQ_INIT(&inodedep->id_extupdt);
1062 	TAILQ_INIT(&inodedep->id_newextupdt);
1063 	ACQUIRE_LOCK(&lk);
1064 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1065 	sema_release(&inodedep_in_progress);
1066 	*inodedeppp = inodedep;
1067 	return (0);
1068 }
1069 
1070 /*
1071  * Structures and routines associated with newblk caching.
1072  */
1073 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1074 u_long	newblk_hash;		/* size of hash table - 1 */
1075 #define	NEWBLK_HASH(fs, inum) \
1076 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1077 static struct sema newblk_in_progress;
1078 
1079 /*
1080  * Look up a newblk. Return 1 if found, 0 if not found.
1081  * If not found, allocate if DEPALLOC flag is passed.
1082  * Found or allocated entry is returned in newblkpp.
1083  */
1084 static int
1085 newblk_lookup(fs, newblkno, flags, newblkpp)
1086 	struct fs *fs;
1087 	ufs2_daddr_t newblkno;
1088 	int flags;
1089 	struct newblk **newblkpp;
1090 {
1091 	struct newblk *newblk;
1092 	struct newblk_hashhead *newblkhd;
1093 
1094 	newblkhd = NEWBLK_HASH(fs, newblkno);
1095 top:
1096 	LIST_FOREACH(newblk, newblkhd, nb_hash)
1097 		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1098 			break;
1099 	if (newblk) {
1100 		*newblkpp = newblk;
1101 		return (1);
1102 	}
1103 	if ((flags & DEPALLOC) == 0) {
1104 		*newblkpp = NULL;
1105 		return (0);
1106 	}
1107 	if (sema_get(&newblk_in_progress, 0) == 0)
1108 		goto top;
1109 	MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1110 		M_NEWBLK, M_SOFTDEP_FLAGS);
1111 	newblk->nb_state = 0;
1112 	newblk->nb_fs = fs;
1113 	newblk->nb_newblkno = newblkno;
1114 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1115 	sema_release(&newblk_in_progress);
1116 	*newblkpp = newblk;
1117 	return (0);
1118 }
1119 
1120 /*
1121  * Executed during filesystem system initialization before
1122  * mounting any filesystems.
1123  */
1124 void
1125 softdep_initialize()
1126 {
1127 
1128 	LIST_INIT(&mkdirlisthd);
1129 	LIST_INIT(&softdep_workitem_pending);
1130 	max_softdeps = desiredvnodes * 8;
1131 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1132 	    &pagedep_hash);
1133 	sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0);
1134 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1135 	sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0);
1136 	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1137 	sema_init(&newblk_in_progress, "newblk", PRIBIO, 0);
1138 
1139 	/* hooks through which the main kernel code calls us */
1140 	softdep_process_worklist_hook = softdep_process_worklist;
1141 	softdep_fsync_hook = softdep_fsync;
1142 
1143 	/* initialise bioops hack */
1144 	bioops.io_start = softdep_disk_io_initiation;
1145 	bioops.io_complete = softdep_disk_write_complete;
1146 	bioops.io_deallocate = softdep_deallocate_dependencies;
1147 	bioops.io_movedeps = softdep_move_dependencies;
1148 	bioops.io_countdeps = softdep_count_dependencies;
1149 }
1150 
1151 /*
1152  * Executed after all filesystems have been unmounted during
1153  * filesystem module unload.
1154  */
1155 void
1156 softdep_uninitialize()
1157 {
1158 
1159 	softdep_process_worklist_hook = NULL;
1160 	softdep_fsync_hook = NULL;
1161 	hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash);
1162 	hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash);
1163 	hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash);
1164 }
1165 
1166 /*
1167  * Called at mount time to notify the dependency code that a
1168  * filesystem wishes to use it.
1169  */
1170 int
1171 softdep_mount(devvp, mp, fs, cred)
1172 	struct vnode *devvp;
1173 	struct mount *mp;
1174 	struct fs *fs;
1175 	struct ucred *cred;
1176 {
1177 	struct csum_total cstotal;
1178 	struct cg *cgp;
1179 	struct buf *bp;
1180 	int error, cyl;
1181 
1182 	mp->mnt_flag &= ~MNT_ASYNC;
1183 	mp->mnt_flag |= MNT_SOFTDEP;
1184 	/*
1185 	 * When doing soft updates, the counters in the
1186 	 * superblock may have gotten out of sync, so we have
1187 	 * to scan the cylinder groups and recalculate them.
1188 	 */
1189 	if (fs->fs_clean != 0)
1190 		return (0);
1191 	bzero(&cstotal, sizeof cstotal);
1192 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1193 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1194 		    fs->fs_cgsize, cred, &bp)) != 0) {
1195 			brelse(bp);
1196 			return (error);
1197 		}
1198 		cgp = (struct cg *)bp->b_data;
1199 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1200 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1201 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1202 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1203 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
1204 		brelse(bp);
1205 	}
1206 #ifdef DEBUG
1207 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1208 		printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
1209 #endif
1210 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1211 	return (0);
1212 }
1213 
1214 /*
1215  * Protecting the freemaps (or bitmaps).
1216  *
1217  * To eliminate the need to execute fsck before mounting a filesystem
1218  * after a power failure, one must (conservatively) guarantee that the
1219  * on-disk copy of the bitmaps never indicate that a live inode or block is
1220  * free.  So, when a block or inode is allocated, the bitmap should be
1221  * updated (on disk) before any new pointers.  When a block or inode is
1222  * freed, the bitmap should not be updated until all pointers have been
1223  * reset.  The latter dependency is handled by the delayed de-allocation
1224  * approach described below for block and inode de-allocation.  The former
1225  * dependency is handled by calling the following procedure when a block or
1226  * inode is allocated. When an inode is allocated an "inodedep" is created
1227  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1228  * Each "inodedep" is also inserted into the hash indexing structure so
1229  * that any additional link additions can be made dependent on the inode
1230  * allocation.
1231  *
1232  * The ufs filesystem maintains a number of free block counts (e.g., per
1233  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1234  * in addition to the bitmaps.  These counts are used to improve efficiency
1235  * during allocation and therefore must be consistent with the bitmaps.
1236  * There is no convenient way to guarantee post-crash consistency of these
1237  * counts with simple update ordering, for two main reasons: (1) The counts
1238  * and bitmaps for a single cylinder group block are not in the same disk
1239  * sector.  If a disk write is interrupted (e.g., by power failure), one may
1240  * be written and the other not.  (2) Some of the counts are located in the
1241  * superblock rather than the cylinder group block. So, we focus our soft
1242  * updates implementation on protecting the bitmaps. When mounting a
1243  * filesystem, we recompute the auxiliary counts from the bitmaps.
1244  */
1245 
1246 /*
1247  * Called just after updating the cylinder group block to allocate an inode.
1248  */
1249 void
1250 softdep_setup_inomapdep(bp, ip, newinum)
1251 	struct buf *bp;		/* buffer for cylgroup block with inode map */
1252 	struct inode *ip;	/* inode related to allocation */
1253 	ino_t newinum;		/* new inode number being allocated */
1254 {
1255 	struct inodedep *inodedep;
1256 	struct bmsafemap *bmsafemap;
1257 
1258 	/*
1259 	 * Create a dependency for the newly allocated inode.
1260 	 * Panic if it already exists as something is seriously wrong.
1261 	 * Otherwise add it to the dependency list for the buffer holding
1262 	 * the cylinder group map from which it was allocated.
1263 	 */
1264 	ACQUIRE_LOCK(&lk);
1265 	if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1266 		FREE_LOCK(&lk);
1267 		panic("softdep_setup_inomapdep: found inode");
1268 	}
1269 	inodedep->id_buf = bp;
1270 	inodedep->id_state &= ~DEPCOMPLETE;
1271 	bmsafemap = bmsafemap_lookup(bp);
1272 	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1273 	FREE_LOCK(&lk);
1274 }
1275 
1276 /*
1277  * Called just after updating the cylinder group block to
1278  * allocate block or fragment.
1279  */
1280 void
1281 softdep_setup_blkmapdep(bp, fs, newblkno)
1282 	struct buf *bp;		/* buffer for cylgroup block with block map */
1283 	struct fs *fs;		/* filesystem doing allocation */
1284 	ufs2_daddr_t newblkno;	/* number of newly allocated block */
1285 {
1286 	struct newblk *newblk;
1287 	struct bmsafemap *bmsafemap;
1288 
1289 	/*
1290 	 * Create a dependency for the newly allocated block.
1291 	 * Add it to the dependency list for the buffer holding
1292 	 * the cylinder group map from which it was allocated.
1293 	 */
1294 	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1295 		panic("softdep_setup_blkmapdep: found block");
1296 	ACQUIRE_LOCK(&lk);
1297 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1298 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1299 	FREE_LOCK(&lk);
1300 }
1301 
1302 /*
1303  * Find the bmsafemap associated with a cylinder group buffer.
1304  * If none exists, create one. The buffer must be locked when
1305  * this routine is called and this routine must be called with
1306  * splbio interrupts blocked.
1307  */
1308 static struct bmsafemap *
1309 bmsafemap_lookup(bp)
1310 	struct buf *bp;
1311 {
1312 	struct bmsafemap *bmsafemap;
1313 	struct worklist *wk;
1314 
1315 #ifdef DEBUG
1316 	if (lk.lkt_held == NOHOLDER)
1317 		panic("bmsafemap_lookup: lock not held");
1318 #endif
1319 	LIST_FOREACH(wk, &bp->b_dep, wk_list)
1320 		if (wk->wk_type == D_BMSAFEMAP)
1321 			return (WK_BMSAFEMAP(wk));
1322 	FREE_LOCK(&lk);
1323 	MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1324 		M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1325 	bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1326 	bmsafemap->sm_list.wk_state = 0;
1327 	bmsafemap->sm_buf = bp;
1328 	LIST_INIT(&bmsafemap->sm_allocdirecthd);
1329 	LIST_INIT(&bmsafemap->sm_allocindirhd);
1330 	LIST_INIT(&bmsafemap->sm_inodedephd);
1331 	LIST_INIT(&bmsafemap->sm_newblkhd);
1332 	ACQUIRE_LOCK(&lk);
1333 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1334 	return (bmsafemap);
1335 }
1336 
1337 /*
1338  * Direct block allocation dependencies.
1339  *
1340  * When a new block is allocated, the corresponding disk locations must be
1341  * initialized (with zeros or new data) before the on-disk inode points to
1342  * them.  Also, the freemap from which the block was allocated must be
1343  * updated (on disk) before the inode's pointer. These two dependencies are
1344  * independent of each other and are needed for all file blocks and indirect
1345  * blocks that are pointed to directly by the inode.  Just before the
1346  * "in-core" version of the inode is updated with a newly allocated block
1347  * number, a procedure (below) is called to setup allocation dependency
1348  * structures.  These structures are removed when the corresponding
1349  * dependencies are satisfied or when the block allocation becomes obsolete
1350  * (i.e., the file is deleted, the block is de-allocated, or the block is a
1351  * fragment that gets upgraded).  All of these cases are handled in
1352  * procedures described later.
1353  *
1354  * When a file extension causes a fragment to be upgraded, either to a larger
1355  * fragment or to a full block, the on-disk location may change (if the
1356  * previous fragment could not simply be extended). In this case, the old
1357  * fragment must be de-allocated, but not until after the inode's pointer has
1358  * been updated. In most cases, this is handled by later procedures, which
1359  * will construct a "freefrag" structure to be added to the workitem queue
1360  * when the inode update is complete (or obsolete).  The main exception to
1361  * this is when an allocation occurs while a pending allocation dependency
1362  * (for the same block pointer) remains.  This case is handled in the main
1363  * allocation dependency setup procedure by immediately freeing the
1364  * unreferenced fragments.
1365  */
1366 void
1367 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1368 	struct inode *ip;	/* inode to which block is being added */
1369 	ufs_lbn_t lbn;		/* block pointer within inode */
1370 	ufs2_daddr_t newblkno;	/* disk block number being added */
1371 	ufs2_daddr_t oldblkno;	/* previous block number, 0 unless frag */
1372 	long newsize;		/* size of new block */
1373 	long oldsize;		/* size of new block */
1374 	struct buf *bp;		/* bp for allocated block */
1375 {
1376 	struct allocdirect *adp, *oldadp;
1377 	struct allocdirectlst *adphead;
1378 	struct bmsafemap *bmsafemap;
1379 	struct inodedep *inodedep;
1380 	struct pagedep *pagedep;
1381 	struct newblk *newblk;
1382 
1383 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1384 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1385 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1386 	adp->ad_lbn = lbn;
1387 	adp->ad_newblkno = newblkno;
1388 	adp->ad_oldblkno = oldblkno;
1389 	adp->ad_newsize = newsize;
1390 	adp->ad_oldsize = oldsize;
1391 	adp->ad_state = ATTACHED;
1392 	LIST_INIT(&adp->ad_newdirblk);
1393 	if (newblkno == oldblkno)
1394 		adp->ad_freefrag = NULL;
1395 	else
1396 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1397 
1398 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1399 		panic("softdep_setup_allocdirect: lost block");
1400 
1401 	ACQUIRE_LOCK(&lk);
1402 	inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1403 	adp->ad_inodedep = inodedep;
1404 
1405 	if (newblk->nb_state == DEPCOMPLETE) {
1406 		adp->ad_state |= DEPCOMPLETE;
1407 		adp->ad_buf = NULL;
1408 	} else {
1409 		bmsafemap = newblk->nb_bmsafemap;
1410 		adp->ad_buf = bmsafemap->sm_buf;
1411 		LIST_REMOVE(newblk, nb_deps);
1412 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1413 	}
1414 	LIST_REMOVE(newblk, nb_hash);
1415 	FREE(newblk, M_NEWBLK);
1416 
1417 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1418 	if (lbn >= NDADDR) {
1419 		/* allocating an indirect block */
1420 		if (oldblkno != 0) {
1421 			FREE_LOCK(&lk);
1422 			panic("softdep_setup_allocdirect: non-zero indir");
1423 		}
1424 	} else {
1425 		/*
1426 		 * Allocating a direct block.
1427 		 *
1428 		 * If we are allocating a directory block, then we must
1429 		 * allocate an associated pagedep to track additions and
1430 		 * deletions.
1431 		 */
1432 		if ((ip->i_mode & IFMT) == IFDIR &&
1433 		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1434 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1435 	}
1436 	/*
1437 	 * The list of allocdirects must be kept in sorted and ascending
1438 	 * order so that the rollback routines can quickly determine the
1439 	 * first uncommitted block (the size of the file stored on disk
1440 	 * ends at the end of the lowest committed fragment, or if there
1441 	 * are no fragments, at the end of the highest committed block).
1442 	 * Since files generally grow, the typical case is that the new
1443 	 * block is to be added at the end of the list. We speed this
1444 	 * special case by checking against the last allocdirect in the
1445 	 * list before laboriously traversing the list looking for the
1446 	 * insertion point.
1447 	 */
1448 	adphead = &inodedep->id_newinoupdt;
1449 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1450 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1451 		/* insert at end of list */
1452 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1453 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1454 			allocdirect_merge(adphead, adp, oldadp);
1455 		FREE_LOCK(&lk);
1456 		return;
1457 	}
1458 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1459 		if (oldadp->ad_lbn >= lbn)
1460 			break;
1461 	}
1462 	if (oldadp == NULL) {
1463 		FREE_LOCK(&lk);
1464 		panic("softdep_setup_allocdirect: lost entry");
1465 	}
1466 	/* insert in middle of list */
1467 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1468 	if (oldadp->ad_lbn == lbn)
1469 		allocdirect_merge(adphead, adp, oldadp);
1470 	FREE_LOCK(&lk);
1471 }
1472 
1473 /*
1474  * Replace an old allocdirect dependency with a newer one.
1475  * This routine must be called with splbio interrupts blocked.
1476  */
1477 static void
1478 allocdirect_merge(adphead, newadp, oldadp)
1479 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
1480 	struct allocdirect *newadp;	/* allocdirect being added */
1481 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
1482 {
1483 	struct worklist *wk;
1484 	struct freefrag *freefrag;
1485 	struct newdirblk *newdirblk;
1486 
1487 #ifdef DEBUG
1488 	if (lk.lkt_held == NOHOLDER)
1489 		panic("allocdirect_merge: lock not held");
1490 #endif
1491 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1492 	    newadp->ad_oldsize != oldadp->ad_newsize ||
1493 	    newadp->ad_lbn >= NDADDR) {
1494 		FREE_LOCK(&lk);
1495 		panic("%s %jd != new %jd || old size %ld != new %ld",
1496 		    "allocdirect_merge: old blkno",
1497 		    (intmax_t)newadp->ad_oldblkno,
1498 		    (intmax_t)oldadp->ad_newblkno,
1499 		    newadp->ad_oldsize, oldadp->ad_newsize);
1500 	}
1501 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
1502 	newadp->ad_oldsize = oldadp->ad_oldsize;
1503 	/*
1504 	 * If the old dependency had a fragment to free or had never
1505 	 * previously had a block allocated, then the new dependency
1506 	 * can immediately post its freefrag and adopt the old freefrag.
1507 	 * This action is done by swapping the freefrag dependencies.
1508 	 * The new dependency gains the old one's freefrag, and the
1509 	 * old one gets the new one and then immediately puts it on
1510 	 * the worklist when it is freed by free_allocdirect. It is
1511 	 * not possible to do this swap when the old dependency had a
1512 	 * non-zero size but no previous fragment to free. This condition
1513 	 * arises when the new block is an extension of the old block.
1514 	 * Here, the first part of the fragment allocated to the new
1515 	 * dependency is part of the block currently claimed on disk by
1516 	 * the old dependency, so cannot legitimately be freed until the
1517 	 * conditions for the new dependency are fulfilled.
1518 	 */
1519 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1520 		freefrag = newadp->ad_freefrag;
1521 		newadp->ad_freefrag = oldadp->ad_freefrag;
1522 		oldadp->ad_freefrag = freefrag;
1523 	}
1524 	/*
1525 	 * If we are tracking a new directory-block allocation,
1526 	 * move it from the old allocdirect to the new allocdirect.
1527 	 */
1528 	if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
1529 		newdirblk = WK_NEWDIRBLK(wk);
1530 		WORKLIST_REMOVE(&newdirblk->db_list);
1531 		if (LIST_FIRST(&oldadp->ad_newdirblk) != NULL)
1532 			panic("allocdirect_merge: extra newdirblk");
1533 		WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list);
1534 	}
1535 	free_allocdirect(adphead, oldadp, 0);
1536 }
1537 
1538 /*
1539  * Allocate a new freefrag structure if needed.
1540  */
1541 static struct freefrag *
1542 newfreefrag(ip, blkno, size)
1543 	struct inode *ip;
1544 	ufs2_daddr_t blkno;
1545 	long size;
1546 {
1547 	struct freefrag *freefrag;
1548 	struct fs *fs;
1549 
1550 	if (blkno == 0)
1551 		return (NULL);
1552 	fs = ip->i_fs;
1553 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1554 		panic("newfreefrag: frag size");
1555 	MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1556 		M_FREEFRAG, M_SOFTDEP_FLAGS);
1557 	freefrag->ff_list.wk_type = D_FREEFRAG;
1558 	freefrag->ff_state = 0;
1559 	freefrag->ff_inum = ip->i_number;
1560 	freefrag->ff_mnt = ITOV(ip)->v_mount;
1561 	freefrag->ff_blkno = blkno;
1562 	freefrag->ff_fragsize = size;
1563 	return (freefrag);
1564 }
1565 
1566 /*
1567  * This workitem de-allocates fragments that were replaced during
1568  * file block allocation.
1569  */
1570 static void
1571 handle_workitem_freefrag(freefrag)
1572 	struct freefrag *freefrag;
1573 {
1574 	struct ufsmount *ump = VFSTOUFS(freefrag->ff_mnt);
1575 
1576 	ffs_blkfree(ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
1577 	    freefrag->ff_fragsize, freefrag->ff_inum);
1578 	FREE(freefrag, M_FREEFRAG);
1579 }
1580 
1581 /*
1582  * Set up a dependency structure for an external attributes data block.
1583  * This routine follows much of the structure of softdep_setup_allocdirect.
1584  * See the description of softdep_setup_allocdirect above for details.
1585  */
1586 void
1587 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1588 	struct inode *ip;
1589 	ufs_lbn_t lbn;
1590 	ufs2_daddr_t newblkno;
1591 	ufs2_daddr_t oldblkno;
1592 	long newsize;
1593 	long oldsize;
1594 	struct buf *bp;
1595 {
1596 	struct allocdirect *adp, *oldadp;
1597 	struct allocdirectlst *adphead;
1598 	struct bmsafemap *bmsafemap;
1599 	struct inodedep *inodedep;
1600 	struct newblk *newblk;
1601 
1602 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1603 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1604 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1605 	adp->ad_lbn = lbn;
1606 	adp->ad_newblkno = newblkno;
1607 	adp->ad_oldblkno = oldblkno;
1608 	adp->ad_newsize = newsize;
1609 	adp->ad_oldsize = oldsize;
1610 	adp->ad_state = ATTACHED | EXTDATA;
1611 	LIST_INIT(&adp->ad_newdirblk);
1612 	if (newblkno == oldblkno)
1613 		adp->ad_freefrag = NULL;
1614 	else
1615 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1616 
1617 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1618 		panic("softdep_setup_allocext: lost block");
1619 
1620 	ACQUIRE_LOCK(&lk);
1621 	inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1622 	adp->ad_inodedep = inodedep;
1623 
1624 	if (newblk->nb_state == DEPCOMPLETE) {
1625 		adp->ad_state |= DEPCOMPLETE;
1626 		adp->ad_buf = NULL;
1627 	} else {
1628 		bmsafemap = newblk->nb_bmsafemap;
1629 		adp->ad_buf = bmsafemap->sm_buf;
1630 		LIST_REMOVE(newblk, nb_deps);
1631 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1632 	}
1633 	LIST_REMOVE(newblk, nb_hash);
1634 	FREE(newblk, M_NEWBLK);
1635 
1636 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1637 	if (lbn >= NXADDR) {
1638 		FREE_LOCK(&lk);
1639 		panic("softdep_setup_allocext: lbn %lld > NXADDR",
1640 		    (long long)lbn);
1641 	}
1642 	/*
1643 	 * The list of allocdirects must be kept in sorted and ascending
1644 	 * order so that the rollback routines can quickly determine the
1645 	 * first uncommitted block (the size of the file stored on disk
1646 	 * ends at the end of the lowest committed fragment, or if there
1647 	 * are no fragments, at the end of the highest committed block).
1648 	 * Since files generally grow, the typical case is that the new
1649 	 * block is to be added at the end of the list. We speed this
1650 	 * special case by checking against the last allocdirect in the
1651 	 * list before laboriously traversing the list looking for the
1652 	 * insertion point.
1653 	 */
1654 	adphead = &inodedep->id_newextupdt;
1655 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1656 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1657 		/* insert at end of list */
1658 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1659 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1660 			allocdirect_merge(adphead, adp, oldadp);
1661 		FREE_LOCK(&lk);
1662 		return;
1663 	}
1664 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1665 		if (oldadp->ad_lbn >= lbn)
1666 			break;
1667 	}
1668 	if (oldadp == NULL) {
1669 		FREE_LOCK(&lk);
1670 		panic("softdep_setup_allocext: lost entry");
1671 	}
1672 	/* insert in middle of list */
1673 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1674 	if (oldadp->ad_lbn == lbn)
1675 		allocdirect_merge(adphead, adp, oldadp);
1676 	FREE_LOCK(&lk);
1677 }
1678 
1679 /*
1680  * Indirect block allocation dependencies.
1681  *
1682  * The same dependencies that exist for a direct block also exist when
1683  * a new block is allocated and pointed to by an entry in a block of
1684  * indirect pointers. The undo/redo states described above are also
1685  * used here. Because an indirect block contains many pointers that
1686  * may have dependencies, a second copy of the entire in-memory indirect
1687  * block is kept. The buffer cache copy is always completely up-to-date.
1688  * The second copy, which is used only as a source for disk writes,
1689  * contains only the safe pointers (i.e., those that have no remaining
1690  * update dependencies). The second copy is freed when all pointers
1691  * are safe. The cache is not allowed to replace indirect blocks with
1692  * pending update dependencies. If a buffer containing an indirect
1693  * block with dependencies is written, these routines will mark it
1694  * dirty again. It can only be successfully written once all the
1695  * dependencies are removed. The ffs_fsync routine in conjunction with
1696  * softdep_sync_metadata work together to get all the dependencies
1697  * removed so that a file can be successfully written to disk. Three
1698  * procedures are used when setting up indirect block pointer
1699  * dependencies. The division is necessary because of the organization
1700  * of the "balloc" routine and because of the distinction between file
1701  * pages and file metadata blocks.
1702  */
1703 
1704 /*
1705  * Allocate a new allocindir structure.
1706  */
1707 static struct allocindir *
1708 newallocindir(ip, ptrno, newblkno, oldblkno)
1709 	struct inode *ip;	/* inode for file being extended */
1710 	int ptrno;		/* offset of pointer in indirect block */
1711 	ufs2_daddr_t newblkno;	/* disk block number being added */
1712 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1713 {
1714 	struct allocindir *aip;
1715 
1716 	MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1717 		M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO);
1718 	aip->ai_list.wk_type = D_ALLOCINDIR;
1719 	aip->ai_state = ATTACHED;
1720 	aip->ai_offset = ptrno;
1721 	aip->ai_newblkno = newblkno;
1722 	aip->ai_oldblkno = oldblkno;
1723 	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1724 	return (aip);
1725 }
1726 
1727 /*
1728  * Called just before setting an indirect block pointer
1729  * to a newly allocated file page.
1730  */
1731 void
1732 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1733 	struct inode *ip;	/* inode for file being extended */
1734 	ufs_lbn_t lbn;		/* allocated block number within file */
1735 	struct buf *bp;		/* buffer with indirect blk referencing page */
1736 	int ptrno;		/* offset of pointer in indirect block */
1737 	ufs2_daddr_t newblkno;	/* disk block number being added */
1738 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1739 	struct buf *nbp;	/* buffer holding allocated page */
1740 {
1741 	struct allocindir *aip;
1742 	struct pagedep *pagedep;
1743 
1744 	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1745 	ACQUIRE_LOCK(&lk);
1746 	/*
1747 	 * If we are allocating a directory page, then we must
1748 	 * allocate an associated pagedep to track additions and
1749 	 * deletions.
1750 	 */
1751 	if ((ip->i_mode & IFMT) == IFDIR &&
1752 	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1753 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1754 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1755 	FREE_LOCK(&lk);
1756 	setup_allocindir_phase2(bp, ip, aip);
1757 }
1758 
1759 /*
1760  * Called just before setting an indirect block pointer to a
1761  * newly allocated indirect block.
1762  */
1763 void
1764 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1765 	struct buf *nbp;	/* newly allocated indirect block */
1766 	struct inode *ip;	/* inode for file being extended */
1767 	struct buf *bp;		/* indirect block referencing allocated block */
1768 	int ptrno;		/* offset of pointer in indirect block */
1769 	ufs2_daddr_t newblkno;	/* disk block number being added */
1770 {
1771 	struct allocindir *aip;
1772 
1773 	aip = newallocindir(ip, ptrno, newblkno, 0);
1774 	ACQUIRE_LOCK(&lk);
1775 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1776 	FREE_LOCK(&lk);
1777 	setup_allocindir_phase2(bp, ip, aip);
1778 }
1779 
1780 /*
1781  * Called to finish the allocation of the "aip" allocated
1782  * by one of the two routines above.
1783  */
1784 static void
1785 setup_allocindir_phase2(bp, ip, aip)
1786 	struct buf *bp;		/* in-memory copy of the indirect block */
1787 	struct inode *ip;	/* inode for file being extended */
1788 	struct allocindir *aip;	/* allocindir allocated by the above routines */
1789 {
1790 	struct worklist *wk;
1791 	struct indirdep *indirdep, *newindirdep;
1792 	struct bmsafemap *bmsafemap;
1793 	struct allocindir *oldaip;
1794 	struct freefrag *freefrag;
1795 	struct newblk *newblk;
1796 	ufs2_daddr_t blkno;
1797 
1798 	if (bp->b_lblkno >= 0)
1799 		panic("setup_allocindir_phase2: not indir blk");
1800 	for (indirdep = NULL, newindirdep = NULL; ; ) {
1801 		ACQUIRE_LOCK(&lk);
1802 		LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1803 			if (wk->wk_type != D_INDIRDEP)
1804 				continue;
1805 			indirdep = WK_INDIRDEP(wk);
1806 			break;
1807 		}
1808 		if (indirdep == NULL && newindirdep) {
1809 			indirdep = newindirdep;
1810 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1811 			newindirdep = NULL;
1812 		}
1813 		FREE_LOCK(&lk);
1814 		if (indirdep) {
1815 			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1816 			    &newblk) == 0)
1817 				panic("setup_allocindir: lost block");
1818 			ACQUIRE_LOCK(&lk);
1819 			if (newblk->nb_state == DEPCOMPLETE) {
1820 				aip->ai_state |= DEPCOMPLETE;
1821 				aip->ai_buf = NULL;
1822 			} else {
1823 				bmsafemap = newblk->nb_bmsafemap;
1824 				aip->ai_buf = bmsafemap->sm_buf;
1825 				LIST_REMOVE(newblk, nb_deps);
1826 				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1827 				    aip, ai_deps);
1828 			}
1829 			LIST_REMOVE(newblk, nb_hash);
1830 			FREE(newblk, M_NEWBLK);
1831 			aip->ai_indirdep = indirdep;
1832 			/*
1833 			 * Check to see if there is an existing dependency
1834 			 * for this block. If there is, merge the old
1835 			 * dependency into the new one.
1836 			 */
1837 			if (aip->ai_oldblkno == 0)
1838 				oldaip = NULL;
1839 			else
1840 
1841 				LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1842 					if (oldaip->ai_offset == aip->ai_offset)
1843 						break;
1844 			freefrag = NULL;
1845 			if (oldaip != NULL) {
1846 				if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1847 					FREE_LOCK(&lk);
1848 					panic("setup_allocindir_phase2: blkno");
1849 				}
1850 				aip->ai_oldblkno = oldaip->ai_oldblkno;
1851 				freefrag = aip->ai_freefrag;
1852 				aip->ai_freefrag = oldaip->ai_freefrag;
1853 				oldaip->ai_freefrag = NULL;
1854 				free_allocindir(oldaip, NULL);
1855 			}
1856 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1857 			if (ip->i_ump->um_fstype == UFS1)
1858 				((ufs1_daddr_t *)indirdep->ir_savebp->b_data)
1859 				    [aip->ai_offset] = aip->ai_oldblkno;
1860 			else
1861 				((ufs2_daddr_t *)indirdep->ir_savebp->b_data)
1862 				    [aip->ai_offset] = aip->ai_oldblkno;
1863 			FREE_LOCK(&lk);
1864 			if (freefrag != NULL)
1865 				handle_workitem_freefrag(freefrag);
1866 		}
1867 		if (newindirdep) {
1868 			if (indirdep->ir_savebp != NULL)
1869 				brelse(newindirdep->ir_savebp);
1870 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1871 		}
1872 		if (indirdep)
1873 			break;
1874 		MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1875 			M_INDIRDEP, M_SOFTDEP_FLAGS);
1876 		newindirdep->ir_list.wk_type = D_INDIRDEP;
1877 		newindirdep->ir_state = ATTACHED;
1878 		if (ip->i_ump->um_fstype == UFS1)
1879 			newindirdep->ir_state |= UFS1FMT;
1880 		LIST_INIT(&newindirdep->ir_deplisthd);
1881 		LIST_INIT(&newindirdep->ir_donehd);
1882 		if (bp->b_blkno == bp->b_lblkno) {
1883 			ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
1884 			    NULL, NULL);
1885 			bp->b_blkno = blkno;
1886 		}
1887 		newindirdep->ir_savebp =
1888 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0);
1889 		BUF_KERNPROC(newindirdep->ir_savebp);
1890 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1891 	}
1892 }
1893 
1894 /*
1895  * Block de-allocation dependencies.
1896  *
1897  * When blocks are de-allocated, the on-disk pointers must be nullified before
1898  * the blocks are made available for use by other files.  (The true
1899  * requirement is that old pointers must be nullified before new on-disk
1900  * pointers are set.  We chose this slightly more stringent requirement to
1901  * reduce complexity.) Our implementation handles this dependency by updating
1902  * the inode (or indirect block) appropriately but delaying the actual block
1903  * de-allocation (i.e., freemap and free space count manipulation) until
1904  * after the updated versions reach stable storage.  After the disk is
1905  * updated, the blocks can be safely de-allocated whenever it is convenient.
1906  * This implementation handles only the common case of reducing a file's
1907  * length to zero. Other cases are handled by the conventional synchronous
1908  * write approach.
1909  *
1910  * The ffs implementation with which we worked double-checks
1911  * the state of the block pointers and file size as it reduces
1912  * a file's length.  Some of this code is replicated here in our
1913  * soft updates implementation.  The freeblks->fb_chkcnt field is
1914  * used to transfer a part of this information to the procedure
1915  * that eventually de-allocates the blocks.
1916  *
1917  * This routine should be called from the routine that shortens
1918  * a file's length, before the inode's size or block pointers
1919  * are modified. It will save the block pointer information for
1920  * later release and zero the inode so that the calling routine
1921  * can release it.
1922  */
1923 void
1924 softdep_setup_freeblocks(ip, length, flags)
1925 	struct inode *ip;	/* The inode whose length is to be reduced */
1926 	off_t length;		/* The new length for the file */
1927 	int flags;		/* IO_EXT and/or IO_NORMAL */
1928 {
1929 	struct freeblks *freeblks;
1930 	struct inodedep *inodedep;
1931 	struct allocdirect *adp;
1932 	struct vnode *vp;
1933 	struct buf *bp;
1934 	struct fs *fs;
1935 	ufs2_daddr_t extblocks, datablocks;
1936 	int i, delay, error;
1937 
1938 	fs = ip->i_fs;
1939 	if (length != 0)
1940 		panic("softdep_setup_freeblocks: non-zero length");
1941 	MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1942 		M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
1943 	freeblks->fb_list.wk_type = D_FREEBLKS;
1944 	freeblks->fb_uid = ip->i_uid;
1945 	freeblks->fb_previousinum = ip->i_number;
1946 	freeblks->fb_devvp = ip->i_devvp;
1947 	freeblks->fb_mnt = ITOV(ip)->v_mount;
1948 	extblocks = 0;
1949 	if (fs->fs_magic == FS_UFS2_MAGIC)
1950 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
1951 	datablocks = DIP(ip, i_blocks) - extblocks;
1952 	if ((flags & IO_NORMAL) == 0) {
1953 		freeblks->fb_oldsize = 0;
1954 		freeblks->fb_chkcnt = 0;
1955 	} else {
1956 		freeblks->fb_oldsize = ip->i_size;
1957 		ip->i_size = 0;
1958 		DIP(ip, i_size) = 0;
1959 		freeblks->fb_chkcnt = datablocks;
1960 		for (i = 0; i < NDADDR; i++) {
1961 			freeblks->fb_dblks[i] = DIP(ip, i_db[i]);
1962 			DIP(ip, i_db[i]) = 0;
1963 		}
1964 		for (i = 0; i < NIADDR; i++) {
1965 			freeblks->fb_iblks[i] = DIP(ip, i_ib[i]);
1966 			DIP(ip, i_ib[i]) = 0;
1967 		}
1968 		/*
1969 		 * If the file was removed, then the space being freed was
1970 		 * accounted for then (see softdep_filereleased()). If the
1971 		 * file is merely being truncated, then we account for it now.
1972 		 */
1973 		if ((ip->i_flag & IN_SPACECOUNTED) == 0)
1974 			fs->fs_pendingblocks += datablocks;
1975 	}
1976 	if ((flags & IO_EXT) == 0) {
1977 		freeblks->fb_oldextsize = 0;
1978 	} else {
1979 		freeblks->fb_oldextsize = ip->i_din2->di_extsize;
1980 		ip->i_din2->di_extsize = 0;
1981 		freeblks->fb_chkcnt += extblocks;
1982 		for (i = 0; i < NXADDR; i++) {
1983 			freeblks->fb_eblks[i] = ip->i_din2->di_extb[i];
1984 			ip->i_din2->di_extb[i] = 0;
1985 		}
1986 	}
1987 	DIP(ip, i_blocks) -= freeblks->fb_chkcnt;
1988 	/*
1989 	 * Push the zero'ed inode to to its disk buffer so that we are free
1990 	 * to delete its dependencies below. Once the dependencies are gone
1991 	 * the buffer can be safely released.
1992 	 */
1993 	if ((error = bread(ip->i_devvp,
1994 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
1995 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
1996 		brelse(bp);
1997 		softdep_error("softdep_setup_freeblocks", error);
1998 	}
1999 	if (ip->i_ump->um_fstype == UFS1)
2000 		*((struct ufs1_dinode *)bp->b_data +
2001 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
2002 	else
2003 		*((struct ufs2_dinode *)bp->b_data +
2004 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
2005 	/*
2006 	 * Find and eliminate any inode dependencies.
2007 	 */
2008 	ACQUIRE_LOCK(&lk);
2009 	(void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
2010 	if ((inodedep->id_state & IOSTARTED) != 0) {
2011 		FREE_LOCK(&lk);
2012 		panic("softdep_setup_freeblocks: inode busy");
2013 	}
2014 	/*
2015 	 * Add the freeblks structure to the list of operations that
2016 	 * must await the zero'ed inode being written to disk. If we
2017 	 * still have a bitmap dependency (delay == 0), then the inode
2018 	 * has never been written to disk, so we can process the
2019 	 * freeblks below once we have deleted the dependencies.
2020 	 */
2021 	delay = (inodedep->id_state & DEPCOMPLETE);
2022 	if (delay)
2023 		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
2024 	/*
2025 	 * Because the file length has been truncated to zero, any
2026 	 * pending block allocation dependency structures associated
2027 	 * with this inode are obsolete and can simply be de-allocated.
2028 	 * We must first merge the two dependency lists to get rid of
2029 	 * any duplicate freefrag structures, then purge the merged list.
2030 	 * If we still have a bitmap dependency, then the inode has never
2031 	 * been written to disk, so we can free any fragments without delay.
2032 	 */
2033 	if (flags & IO_NORMAL) {
2034 		merge_inode_lists(&inodedep->id_newinoupdt,
2035 		    &inodedep->id_inoupdt);
2036 		while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
2037 			free_allocdirect(&inodedep->id_inoupdt, adp, delay);
2038 	}
2039 	if (flags & IO_EXT) {
2040 		merge_inode_lists(&inodedep->id_newextupdt,
2041 		    &inodedep->id_extupdt);
2042 		while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
2043 			free_allocdirect(&inodedep->id_extupdt, adp, delay);
2044 	}
2045 	FREE_LOCK(&lk);
2046 	bdwrite(bp);
2047 	/*
2048 	 * We must wait for any I/O in progress to finish so that
2049 	 * all potential buffers on the dirty list will be visible.
2050 	 * Once they are all there, walk the list and get rid of
2051 	 * any dependencies.
2052 	 */
2053 	vp = ITOV(ip);
2054 	ACQUIRE_LOCK(&lk);
2055 	drain_output(vp, 1);
2056 restart:
2057 	VI_LOCK(vp);
2058 	TAILQ_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
2059 		if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
2060 		    ((flags & IO_NORMAL) == 0 &&
2061 		      (bp->b_xflags & BX_ALTDATA) == 0))
2062 			continue;
2063 		VI_UNLOCK(vp);
2064 		if (getdirtybuf(&bp, MNT_WAIT) == 0)
2065 			goto restart;
2066 		(void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
2067 		deallocate_dependencies(bp, inodedep);
2068 		bp->b_flags |= B_INVAL | B_NOCACHE;
2069 		FREE_LOCK(&lk);
2070 		brelse(bp);
2071 		ACQUIRE_LOCK(&lk);
2072 		goto restart;
2073 	}
2074 	VI_UNLOCK(vp);
2075 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
2076 		(void) free_inodedep(inodedep);
2077 	FREE_LOCK(&lk);
2078 	/*
2079 	 * If the inode has never been written to disk (delay == 0),
2080 	 * then we can process the freeblks now that we have deleted
2081 	 * the dependencies.
2082 	 */
2083 	if (!delay)
2084 		handle_workitem_freeblocks(freeblks, 0);
2085 }
2086 
2087 /*
2088  * Reclaim any dependency structures from a buffer that is about to
2089  * be reallocated to a new vnode. The buffer must be locked, thus,
2090  * no I/O completion operations can occur while we are manipulating
2091  * its associated dependencies. The mutex is held so that other I/O's
2092  * associated with related dependencies do not occur.
2093  */
2094 static void
2095 deallocate_dependencies(bp, inodedep)
2096 	struct buf *bp;
2097 	struct inodedep *inodedep;
2098 {
2099 	struct worklist *wk;
2100 	struct indirdep *indirdep;
2101 	struct allocindir *aip;
2102 	struct pagedep *pagedep;
2103 	struct dirrem *dirrem;
2104 	struct diradd *dap;
2105 	int i;
2106 
2107 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2108 		switch (wk->wk_type) {
2109 
2110 		case D_INDIRDEP:
2111 			indirdep = WK_INDIRDEP(wk);
2112 			/*
2113 			 * None of the indirect pointers will ever be visible,
2114 			 * so they can simply be tossed. GOINGAWAY ensures
2115 			 * that allocated pointers will be saved in the buffer
2116 			 * cache until they are freed. Note that they will
2117 			 * only be able to be found by their physical address
2118 			 * since the inode mapping the logical address will
2119 			 * be gone. The save buffer used for the safe copy
2120 			 * was allocated in setup_allocindir_phase2 using
2121 			 * the physical address so it could be used for this
2122 			 * purpose. Hence we swap the safe copy with the real
2123 			 * copy, allowing the safe copy to be freed and holding
2124 			 * on to the real copy for later use in indir_trunc.
2125 			 */
2126 			if (indirdep->ir_state & GOINGAWAY) {
2127 				FREE_LOCK(&lk);
2128 				panic("deallocate_dependencies: already gone");
2129 			}
2130 			indirdep->ir_state |= GOINGAWAY;
2131 			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
2132 				free_allocindir(aip, inodedep);
2133 			if (bp->b_lblkno >= 0 ||
2134 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno) {
2135 				FREE_LOCK(&lk);
2136 				panic("deallocate_dependencies: not indir");
2137 			}
2138 			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2139 			    bp->b_bcount);
2140 			WORKLIST_REMOVE(wk);
2141 			WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
2142 			continue;
2143 
2144 		case D_PAGEDEP:
2145 			pagedep = WK_PAGEDEP(wk);
2146 			/*
2147 			 * None of the directory additions will ever be
2148 			 * visible, so they can simply be tossed.
2149 			 */
2150 			for (i = 0; i < DAHASHSZ; i++)
2151 				while ((dap =
2152 				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
2153 					free_diradd(dap);
2154 			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2155 				free_diradd(dap);
2156 			/*
2157 			 * Copy any directory remove dependencies to the list
2158 			 * to be processed after the zero'ed inode is written.
2159 			 * If the inode has already been written, then they
2160 			 * can be dumped directly onto the work list.
2161 			 */
2162 			LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2163 				LIST_REMOVE(dirrem, dm_next);
2164 				dirrem->dm_dirinum = pagedep->pd_ino;
2165 				if (inodedep == NULL ||
2166 				    (inodedep->id_state & ALLCOMPLETE) ==
2167 				     ALLCOMPLETE)
2168 					add_to_worklist(&dirrem->dm_list);
2169 				else
2170 					WORKLIST_INSERT(&inodedep->id_bufwait,
2171 					    &dirrem->dm_list);
2172 			}
2173 			if ((pagedep->pd_state & NEWBLOCK) != 0) {
2174 				LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list)
2175 					if (wk->wk_type == D_NEWDIRBLK &&
2176 					    WK_NEWDIRBLK(wk)->db_pagedep ==
2177 					      pagedep)
2178 						break;
2179 				if (wk != NULL) {
2180 					WORKLIST_REMOVE(wk);
2181 					free_newdirblk(WK_NEWDIRBLK(wk));
2182 				} else {
2183 					FREE_LOCK(&lk);
2184 					panic("deallocate_dependencies: "
2185 					      "lost pagedep");
2186 				}
2187 			}
2188 			WORKLIST_REMOVE(&pagedep->pd_list);
2189 			LIST_REMOVE(pagedep, pd_hash);
2190 			WORKITEM_FREE(pagedep, D_PAGEDEP);
2191 			continue;
2192 
2193 		case D_ALLOCINDIR:
2194 			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2195 			continue;
2196 
2197 		case D_ALLOCDIRECT:
2198 		case D_INODEDEP:
2199 			FREE_LOCK(&lk);
2200 			panic("deallocate_dependencies: Unexpected type %s",
2201 			    TYPENAME(wk->wk_type));
2202 			/* NOTREACHED */
2203 
2204 		default:
2205 			FREE_LOCK(&lk);
2206 			panic("deallocate_dependencies: Unknown type %s",
2207 			    TYPENAME(wk->wk_type));
2208 			/* NOTREACHED */
2209 		}
2210 	}
2211 }
2212 
2213 /*
2214  * Free an allocdirect. Generate a new freefrag work request if appropriate.
2215  * This routine must be called with splbio interrupts blocked.
2216  */
2217 static void
2218 free_allocdirect(adphead, adp, delay)
2219 	struct allocdirectlst *adphead;
2220 	struct allocdirect *adp;
2221 	int delay;
2222 {
2223 	struct newdirblk *newdirblk;
2224 	struct worklist *wk;
2225 
2226 #ifdef DEBUG
2227 	if (lk.lkt_held == NOHOLDER)
2228 		panic("free_allocdirect: lock not held");
2229 #endif
2230 	if ((adp->ad_state & DEPCOMPLETE) == 0)
2231 		LIST_REMOVE(adp, ad_deps);
2232 	TAILQ_REMOVE(adphead, adp, ad_next);
2233 	if ((adp->ad_state & COMPLETE) == 0)
2234 		WORKLIST_REMOVE(&adp->ad_list);
2235 	if (adp->ad_freefrag != NULL) {
2236 		if (delay)
2237 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2238 			    &adp->ad_freefrag->ff_list);
2239 		else
2240 			add_to_worklist(&adp->ad_freefrag->ff_list);
2241 	}
2242 	if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) {
2243 		newdirblk = WK_NEWDIRBLK(wk);
2244 		WORKLIST_REMOVE(&newdirblk->db_list);
2245 		if (LIST_FIRST(&adp->ad_newdirblk) != NULL)
2246 			panic("free_allocdirect: extra newdirblk");
2247 		if (delay)
2248 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2249 			    &newdirblk->db_list);
2250 		else
2251 			free_newdirblk(newdirblk);
2252 	}
2253 	WORKITEM_FREE(adp, D_ALLOCDIRECT);
2254 }
2255 
2256 /*
2257  * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
2258  * This routine must be called with splbio interrupts blocked.
2259  */
2260 static void
2261 free_newdirblk(newdirblk)
2262 	struct newdirblk *newdirblk;
2263 {
2264 	struct pagedep *pagedep;
2265 	struct diradd *dap;
2266 	int i;
2267 
2268 #ifdef DEBUG
2269 	if (lk.lkt_held == NOHOLDER)
2270 		panic("free_newdirblk: lock not held");
2271 #endif
2272 	/*
2273 	 * If the pagedep is still linked onto the directory buffer
2274 	 * dependency chain, then some of the entries on the
2275 	 * pd_pendinghd list may not be committed to disk yet. In
2276 	 * this case, we will simply clear the NEWBLOCK flag and
2277 	 * let the pd_pendinghd list be processed when the pagedep
2278 	 * is next written. If the pagedep is no longer on the buffer
2279 	 * dependency chain, then all the entries on the pd_pending
2280 	 * list are committed to disk and we can free them here.
2281 	 */
2282 	pagedep = newdirblk->db_pagedep;
2283 	pagedep->pd_state &= ~NEWBLOCK;
2284 	if ((pagedep->pd_state & ONWORKLIST) == 0)
2285 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
2286 			free_diradd(dap);
2287 	/*
2288 	 * If no dependencies remain, the pagedep will be freed.
2289 	 */
2290 	for (i = 0; i < DAHASHSZ; i++)
2291 		if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
2292 			break;
2293 	if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) {
2294 		LIST_REMOVE(pagedep, pd_hash);
2295 		WORKITEM_FREE(pagedep, D_PAGEDEP);
2296 	}
2297 	WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2298 }
2299 
2300 /*
2301  * Prepare an inode to be freed. The actual free operation is not
2302  * done until the zero'ed inode has been written to disk.
2303  */
2304 void
2305 softdep_freefile(pvp, ino, mode)
2306 	struct vnode *pvp;
2307 	ino_t ino;
2308 	int mode;
2309 {
2310 	struct inode *ip = VTOI(pvp);
2311 	struct inodedep *inodedep;
2312 	struct freefile *freefile;
2313 
2314 	/*
2315 	 * This sets up the inode de-allocation dependency.
2316 	 */
2317 	MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2318 		M_FREEFILE, M_SOFTDEP_FLAGS);
2319 	freefile->fx_list.wk_type = D_FREEFILE;
2320 	freefile->fx_list.wk_state = 0;
2321 	freefile->fx_mode = mode;
2322 	freefile->fx_oldinum = ino;
2323 	freefile->fx_devvp = ip->i_devvp;
2324 	freefile->fx_mnt = ITOV(ip)->v_mount;
2325 	if ((ip->i_flag & IN_SPACECOUNTED) == 0)
2326 		ip->i_fs->fs_pendinginodes += 1;
2327 
2328 	/*
2329 	 * If the inodedep does not exist, then the zero'ed inode has
2330 	 * been written to disk. If the allocated inode has never been
2331 	 * written to disk, then the on-disk inode is zero'ed. In either
2332 	 * case we can free the file immediately.
2333 	 */
2334 	ACQUIRE_LOCK(&lk);
2335 	if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2336 	    check_inode_unwritten(inodedep)) {
2337 		FREE_LOCK(&lk);
2338 		handle_workitem_freefile(freefile);
2339 		return;
2340 	}
2341 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2342 	FREE_LOCK(&lk);
2343 }
2344 
2345 /*
2346  * Check to see if an inode has never been written to disk. If
2347  * so free the inodedep and return success, otherwise return failure.
2348  * This routine must be called with splbio interrupts blocked.
2349  *
2350  * If we still have a bitmap dependency, then the inode has never
2351  * been written to disk. Drop the dependency as it is no longer
2352  * necessary since the inode is being deallocated. We set the
2353  * ALLCOMPLETE flags since the bitmap now properly shows that the
2354  * inode is not allocated. Even if the inode is actively being
2355  * written, it has been rolled back to its zero'ed state, so we
2356  * are ensured that a zero inode is what is on the disk. For short
2357  * lived files, this change will usually result in removing all the
2358  * dependencies from the inode so that it can be freed immediately.
2359  */
2360 static int
2361 check_inode_unwritten(inodedep)
2362 	struct inodedep *inodedep;
2363 {
2364 
2365 	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2366 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2367 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2368 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
2369 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2370 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2371 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
2372 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
2373 	    inodedep->id_nlinkdelta != 0)
2374 		return (0);
2375 	inodedep->id_state |= ALLCOMPLETE;
2376 	LIST_REMOVE(inodedep, id_deps);
2377 	inodedep->id_buf = NULL;
2378 	if (inodedep->id_state & ONWORKLIST)
2379 		WORKLIST_REMOVE(&inodedep->id_list);
2380 	if (inodedep->id_savedino1 != NULL) {
2381 		FREE(inodedep->id_savedino1, M_INODEDEP);
2382 		inodedep->id_savedino1 = NULL;
2383 	}
2384 	if (free_inodedep(inodedep) == 0) {
2385 		FREE_LOCK(&lk);
2386 		panic("check_inode_unwritten: busy inode");
2387 	}
2388 	return (1);
2389 }
2390 
2391 /*
2392  * Try to free an inodedep structure. Return 1 if it could be freed.
2393  */
2394 static int
2395 free_inodedep(inodedep)
2396 	struct inodedep *inodedep;
2397 {
2398 
2399 	if ((inodedep->id_state & ONWORKLIST) != 0 ||
2400 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2401 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2402 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2403 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
2404 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2405 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2406 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
2407 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
2408 	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL)
2409 		return (0);
2410 	LIST_REMOVE(inodedep, id_hash);
2411 	WORKITEM_FREE(inodedep, D_INODEDEP);
2412 	num_inodedep -= 1;
2413 	return (1);
2414 }
2415 
2416 /*
2417  * This workitem routine performs the block de-allocation.
2418  * The workitem is added to the pending list after the updated
2419  * inode block has been written to disk.  As mentioned above,
2420  * checks regarding the number of blocks de-allocated (compared
2421  * to the number of blocks allocated for the file) are also
2422  * performed in this function.
2423  */
2424 static void
2425 handle_workitem_freeblocks(freeblks, flags)
2426 	struct freeblks *freeblks;
2427 	int flags;
2428 {
2429 	struct inode *ip;
2430 	struct vnode *vp;
2431 	struct fs *fs;
2432 	int i, nblocks, level, bsize;
2433 	ufs2_daddr_t bn, blocksreleased = 0;
2434 	int error, allerror = 0;
2435 	ufs_lbn_t baselbns[NIADDR], tmpval;
2436 
2437 	fs = VFSTOUFS(freeblks->fb_mnt)->um_fs;
2438 	tmpval = 1;
2439 	baselbns[0] = NDADDR;
2440 	for (i = 1; i < NIADDR; i++) {
2441 		tmpval *= NINDIR(fs);
2442 		baselbns[i] = baselbns[i - 1] + tmpval;
2443 	}
2444 	nblocks = btodb(fs->fs_bsize);
2445 	blocksreleased = 0;
2446 	/*
2447 	 * Release all extended attribute blocks or frags.
2448 	 */
2449 	if (freeblks->fb_oldextsize > 0) {
2450 		for (i = (NXADDR - 1); i >= 0; i--) {
2451 			if ((bn = freeblks->fb_eblks[i]) == 0)
2452 				continue;
2453 			bsize = sblksize(fs, freeblks->fb_oldextsize, i);
2454 			ffs_blkfree(fs, freeblks->fb_devvp, bn, bsize,
2455 			    freeblks->fb_previousinum);
2456 			blocksreleased += btodb(bsize);
2457 		}
2458 	}
2459 	/*
2460 	 * Release all data blocks or frags.
2461 	 */
2462 	if (freeblks->fb_oldsize > 0) {
2463 		/*
2464 		 * Indirect blocks first.
2465 		 */
2466 		for (level = (NIADDR - 1); level >= 0; level--) {
2467 			if ((bn = freeblks->fb_iblks[level]) == 0)
2468 				continue;
2469 			if ((error = indir_trunc(freeblks, fsbtodb(fs, bn),
2470 			    level, baselbns[level], &blocksreleased)) == 0)
2471 				allerror = error;
2472 			ffs_blkfree(fs, freeblks->fb_devvp, bn, fs->fs_bsize,
2473 			    freeblks->fb_previousinum);
2474 			fs->fs_pendingblocks -= nblocks;
2475 			blocksreleased += nblocks;
2476 		}
2477 		/*
2478 		 * All direct blocks or frags.
2479 		 */
2480 		for (i = (NDADDR - 1); i >= 0; i--) {
2481 			if ((bn = freeblks->fb_dblks[i]) == 0)
2482 				continue;
2483 			bsize = sblksize(fs, freeblks->fb_oldsize, i);
2484 			ffs_blkfree(fs, freeblks->fb_devvp, bn, bsize,
2485 			    freeblks->fb_previousinum);
2486 			fs->fs_pendingblocks -= btodb(bsize);
2487 			blocksreleased += btodb(bsize);
2488 		}
2489 	}
2490 	/*
2491 	 * If we still have not finished background cleanup, then check
2492 	 * to see if the block count needs to be adjusted.
2493 	 */
2494 	if (freeblks->fb_chkcnt != blocksreleased &&
2495 	    (fs->fs_flags & FS_UNCLEAN) != 0 &&
2496 	    VFS_VGET(freeblks->fb_mnt, freeblks->fb_previousinum,
2497 	    (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) {
2498 		ip = VTOI(vp);
2499 		DIP(ip, i_blocks) += freeblks->fb_chkcnt - blocksreleased;
2500 		ip->i_flag |= IN_CHANGE;
2501 		vput(vp);
2502 	}
2503 
2504 #ifdef DIAGNOSTIC
2505 	if (freeblks->fb_chkcnt != blocksreleased &&
2506 	    ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0))
2507 		printf("handle_workitem_freeblocks: block count\n");
2508 	if (allerror)
2509 		softdep_error("handle_workitem_freeblks", allerror);
2510 #endif /* DIAGNOSTIC */
2511 
2512 	WORKITEM_FREE(freeblks, D_FREEBLKS);
2513 }
2514 
2515 /*
2516  * Release blocks associated with the inode ip and stored in the indirect
2517  * block dbn. If level is greater than SINGLE, the block is an indirect block
2518  * and recursive calls to indirtrunc must be used to cleanse other indirect
2519  * blocks.
2520  */
2521 static int
2522 indir_trunc(freeblks, dbn, level, lbn, countp)
2523 	struct freeblks *freeblks;
2524 	ufs2_daddr_t dbn;
2525 	int level;
2526 	ufs_lbn_t lbn;
2527 	ufs2_daddr_t *countp;
2528 {
2529 	struct buf *bp;
2530 	struct fs *fs;
2531 	struct worklist *wk;
2532 	struct indirdep *indirdep;
2533 	ufs1_daddr_t *bap1 = 0;
2534 	ufs2_daddr_t nb, *bap2 = 0;
2535 	ufs_lbn_t lbnadd;
2536 	int i, nblocks, ufs1fmt;
2537 	int error, allerror = 0;
2538 
2539 	fs = VFSTOUFS(freeblks->fb_mnt)->um_fs;
2540 	lbnadd = 1;
2541 	for (i = level; i > 0; i--)
2542 		lbnadd *= NINDIR(fs);
2543 	/*
2544 	 * Get buffer of block pointers to be freed. This routine is not
2545 	 * called until the zero'ed inode has been written, so it is safe
2546 	 * to free blocks as they are encountered. Because the inode has
2547 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
2548 	 * have to use the on-disk address and the block device for the
2549 	 * filesystem to look them up. If the file was deleted before its
2550 	 * indirect blocks were all written to disk, the routine that set
2551 	 * us up (deallocate_dependencies) will have arranged to leave
2552 	 * a complete copy of the indirect block in memory for our use.
2553 	 * Otherwise we have to read the blocks in from the disk.
2554 	 */
2555 	ACQUIRE_LOCK(&lk);
2556 	if ((bp = incore(freeblks->fb_devvp, dbn)) != NULL &&
2557 	    (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2558 		if (wk->wk_type != D_INDIRDEP ||
2559 		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2560 		    (indirdep->ir_state & GOINGAWAY) == 0) {
2561 			FREE_LOCK(&lk);
2562 			panic("indir_trunc: lost indirdep");
2563 		}
2564 		WORKLIST_REMOVE(wk);
2565 		WORKITEM_FREE(indirdep, D_INDIRDEP);
2566 		if (LIST_FIRST(&bp->b_dep) != NULL) {
2567 			FREE_LOCK(&lk);
2568 			panic("indir_trunc: dangling dep");
2569 		}
2570 		FREE_LOCK(&lk);
2571 	} else {
2572 		FREE_LOCK(&lk);
2573 		error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
2574 		    NOCRED, &bp);
2575 		if (error) {
2576 			brelse(bp);
2577 			return (error);
2578 		}
2579 	}
2580 	/*
2581 	 * Recursively free indirect blocks.
2582 	 */
2583 	if (VFSTOUFS(freeblks->fb_mnt)->um_fstype == UFS1) {
2584 		ufs1fmt = 1;
2585 		bap1 = (ufs1_daddr_t *)bp->b_data;
2586 	} else {
2587 		ufs1fmt = 0;
2588 		bap2 = (ufs2_daddr_t *)bp->b_data;
2589 	}
2590 	nblocks = btodb(fs->fs_bsize);
2591 	for (i = NINDIR(fs) - 1; i >= 0; i--) {
2592 		if (ufs1fmt)
2593 			nb = bap1[i];
2594 		else
2595 			nb = bap2[i];
2596 		if (nb == 0)
2597 			continue;
2598 		if (level != 0) {
2599 			if ((error = indir_trunc(freeblks, fsbtodb(fs, nb),
2600 			     level - 1, lbn + (i * lbnadd), countp)) != 0)
2601 				allerror = error;
2602 		}
2603 		ffs_blkfree(fs, freeblks->fb_devvp, nb, fs->fs_bsize,
2604 		    freeblks->fb_previousinum);
2605 		fs->fs_pendingblocks -= nblocks;
2606 		*countp += nblocks;
2607 	}
2608 	bp->b_flags |= B_INVAL | B_NOCACHE;
2609 	brelse(bp);
2610 	return (allerror);
2611 }
2612 
2613 /*
2614  * Free an allocindir.
2615  * This routine must be called with splbio interrupts blocked.
2616  */
2617 static void
2618 free_allocindir(aip, inodedep)
2619 	struct allocindir *aip;
2620 	struct inodedep *inodedep;
2621 {
2622 	struct freefrag *freefrag;
2623 
2624 #ifdef DEBUG
2625 	if (lk.lkt_held == NOHOLDER)
2626 		panic("free_allocindir: lock not held");
2627 #endif
2628 	if ((aip->ai_state & DEPCOMPLETE) == 0)
2629 		LIST_REMOVE(aip, ai_deps);
2630 	if (aip->ai_state & ONWORKLIST)
2631 		WORKLIST_REMOVE(&aip->ai_list);
2632 	LIST_REMOVE(aip, ai_next);
2633 	if ((freefrag = aip->ai_freefrag) != NULL) {
2634 		if (inodedep == NULL)
2635 			add_to_worklist(&freefrag->ff_list);
2636 		else
2637 			WORKLIST_INSERT(&inodedep->id_bufwait,
2638 			    &freefrag->ff_list);
2639 	}
2640 	WORKITEM_FREE(aip, D_ALLOCINDIR);
2641 }
2642 
2643 /*
2644  * Directory entry addition dependencies.
2645  *
2646  * When adding a new directory entry, the inode (with its incremented link
2647  * count) must be written to disk before the directory entry's pointer to it.
2648  * Also, if the inode is newly allocated, the corresponding freemap must be
2649  * updated (on disk) before the directory entry's pointer. These requirements
2650  * are met via undo/redo on the directory entry's pointer, which consists
2651  * simply of the inode number.
2652  *
2653  * As directory entries are added and deleted, the free space within a
2654  * directory block can become fragmented.  The ufs filesystem will compact
2655  * a fragmented directory block to make space for a new entry. When this
2656  * occurs, the offsets of previously added entries change. Any "diradd"
2657  * dependency structures corresponding to these entries must be updated with
2658  * the new offsets.
2659  */
2660 
2661 /*
2662  * This routine is called after the in-memory inode's link
2663  * count has been incremented, but before the directory entry's
2664  * pointer to the inode has been set.
2665  */
2666 int
2667 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
2668 	struct buf *bp;		/* buffer containing directory block */
2669 	struct inode *dp;	/* inode for directory */
2670 	off_t diroffset;	/* offset of new entry in directory */
2671 	ino_t newinum;		/* inode referenced by new directory entry */
2672 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
2673 	int isnewblk;		/* entry is in a newly allocated block */
2674 {
2675 	int offset;		/* offset of new entry within directory block */
2676 	ufs_lbn_t lbn;		/* block in directory containing new entry */
2677 	struct fs *fs;
2678 	struct diradd *dap;
2679 	struct allocdirect *adp;
2680 	struct pagedep *pagedep;
2681 	struct inodedep *inodedep;
2682 	struct newdirblk *newdirblk = 0;
2683 	struct mkdir *mkdir1, *mkdir2;
2684 
2685 	/*
2686 	 * Whiteouts have no dependencies.
2687 	 */
2688 	if (newinum == WINO) {
2689 		if (newdirbp != NULL)
2690 			bdwrite(newdirbp);
2691 		return (0);
2692 	}
2693 
2694 	fs = dp->i_fs;
2695 	lbn = lblkno(fs, diroffset);
2696 	offset = blkoff(fs, diroffset);
2697 	MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2698 		M_SOFTDEP_FLAGS|M_ZERO);
2699 	dap->da_list.wk_type = D_DIRADD;
2700 	dap->da_offset = offset;
2701 	dap->da_newinum = newinum;
2702 	dap->da_state = ATTACHED;
2703 	if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) {
2704 		MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk),
2705 		    M_NEWDIRBLK, M_SOFTDEP_FLAGS);
2706 		newdirblk->db_list.wk_type = D_NEWDIRBLK;
2707 		newdirblk->db_state = 0;
2708 	}
2709 	if (newdirbp == NULL) {
2710 		dap->da_state |= DEPCOMPLETE;
2711 		ACQUIRE_LOCK(&lk);
2712 	} else {
2713 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2714 		MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2715 		    M_SOFTDEP_FLAGS);
2716 		mkdir1->md_list.wk_type = D_MKDIR;
2717 		mkdir1->md_state = MKDIR_BODY;
2718 		mkdir1->md_diradd = dap;
2719 		MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2720 		    M_SOFTDEP_FLAGS);
2721 		mkdir2->md_list.wk_type = D_MKDIR;
2722 		mkdir2->md_state = MKDIR_PARENT;
2723 		mkdir2->md_diradd = dap;
2724 		/*
2725 		 * Dependency on "." and ".." being written to disk.
2726 		 */
2727 		mkdir1->md_buf = newdirbp;
2728 		ACQUIRE_LOCK(&lk);
2729 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2730 		WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2731 		FREE_LOCK(&lk);
2732 		bdwrite(newdirbp);
2733 		/*
2734 		 * Dependency on link count increase for parent directory
2735 		 */
2736 		ACQUIRE_LOCK(&lk);
2737 		if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0
2738 		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2739 			dap->da_state &= ~MKDIR_PARENT;
2740 			WORKITEM_FREE(mkdir2, D_MKDIR);
2741 		} else {
2742 			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2743 			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2744 		}
2745 	}
2746 	/*
2747 	 * Link into parent directory pagedep to await its being written.
2748 	 */
2749 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2750 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2751 	dap->da_pagedep = pagedep;
2752 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2753 	    da_pdlist);
2754 	/*
2755 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
2756 	 * is not yet written. If it is written, do the post-inode write
2757 	 * processing to put it on the id_pendinghd list.
2758 	 */
2759 	(void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2760 	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2761 		diradd_inode_written(dap, inodedep);
2762 	else
2763 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2764 	if (isnewblk) {
2765 		/*
2766 		 * Directories growing into indirect blocks are rare
2767 		 * enough and the frequency of new block allocation
2768 		 * in those cases even more rare, that we choose not
2769 		 * to bother tracking them. Rather we simply force the
2770 		 * new directory entry to disk.
2771 		 */
2772 		if (lbn >= NDADDR) {
2773 			FREE_LOCK(&lk);
2774 			/*
2775 			 * We only have a new allocation when at the
2776 			 * beginning of a new block, not when we are
2777 			 * expanding into an existing block.
2778 			 */
2779 			if (blkoff(fs, diroffset) == 0)
2780 				return (1);
2781 			return (0);
2782 		}
2783 		/*
2784 		 * We only have a new allocation when at the beginning
2785 		 * of a new fragment, not when we are expanding into an
2786 		 * existing fragment. Also, there is nothing to do if we
2787 		 * are already tracking this block.
2788 		 */
2789 		if (fragoff(fs, diroffset) != 0) {
2790 			FREE_LOCK(&lk);
2791 			return (0);
2792 		}
2793 		if ((pagedep->pd_state & NEWBLOCK) != 0) {
2794 			WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2795 			FREE_LOCK(&lk);
2796 			return (0);
2797 		}
2798 		/*
2799 		 * Find our associated allocdirect and have it track us.
2800 		 */
2801 		if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0)
2802 			panic("softdep_setup_directory_add: lost inodedep");
2803 		adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst);
2804 		if (adp == NULL || adp->ad_lbn != lbn) {
2805 			FREE_LOCK(&lk);
2806 			panic("softdep_setup_directory_add: lost entry");
2807 		}
2808 		pagedep->pd_state |= NEWBLOCK;
2809 		newdirblk->db_pagedep = pagedep;
2810 		WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list);
2811 	}
2812 	FREE_LOCK(&lk);
2813 	return (0);
2814 }
2815 
2816 /*
2817  * This procedure is called to change the offset of a directory
2818  * entry when compacting a directory block which must be owned
2819  * exclusively by the caller. Note that the actual entry movement
2820  * must be done in this procedure to ensure that no I/O completions
2821  * occur while the move is in progress.
2822  */
2823 void
2824 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2825 	struct inode *dp;	/* inode for directory */
2826 	caddr_t base;		/* address of dp->i_offset */
2827 	caddr_t oldloc;		/* address of old directory location */
2828 	caddr_t newloc;		/* address of new directory location */
2829 	int entrysize;		/* size of directory entry */
2830 {
2831 	int offset, oldoffset, newoffset;
2832 	struct pagedep *pagedep;
2833 	struct diradd *dap;
2834 	ufs_lbn_t lbn;
2835 
2836 	ACQUIRE_LOCK(&lk);
2837 	lbn = lblkno(dp->i_fs, dp->i_offset);
2838 	offset = blkoff(dp->i_fs, dp->i_offset);
2839 	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2840 		goto done;
2841 	oldoffset = offset + (oldloc - base);
2842 	newoffset = offset + (newloc - base);
2843 
2844 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2845 		if (dap->da_offset != oldoffset)
2846 			continue;
2847 		dap->da_offset = newoffset;
2848 		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2849 			break;
2850 		LIST_REMOVE(dap, da_pdlist);
2851 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2852 		    dap, da_pdlist);
2853 		break;
2854 	}
2855 	if (dap == NULL) {
2856 
2857 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2858 			if (dap->da_offset == oldoffset) {
2859 				dap->da_offset = newoffset;
2860 				break;
2861 			}
2862 		}
2863 	}
2864 done:
2865 	bcopy(oldloc, newloc, entrysize);
2866 	FREE_LOCK(&lk);
2867 }
2868 
2869 /*
2870  * Free a diradd dependency structure. This routine must be called
2871  * with splbio interrupts blocked.
2872  */
2873 static void
2874 free_diradd(dap)
2875 	struct diradd *dap;
2876 {
2877 	struct dirrem *dirrem;
2878 	struct pagedep *pagedep;
2879 	struct inodedep *inodedep;
2880 	struct mkdir *mkdir, *nextmd;
2881 
2882 #ifdef DEBUG
2883 	if (lk.lkt_held == NOHOLDER)
2884 		panic("free_diradd: lock not held");
2885 #endif
2886 	WORKLIST_REMOVE(&dap->da_list);
2887 	LIST_REMOVE(dap, da_pdlist);
2888 	if ((dap->da_state & DIRCHG) == 0) {
2889 		pagedep = dap->da_pagedep;
2890 	} else {
2891 		dirrem = dap->da_previous;
2892 		pagedep = dirrem->dm_pagedep;
2893 		dirrem->dm_dirinum = pagedep->pd_ino;
2894 		add_to_worklist(&dirrem->dm_list);
2895 	}
2896 	if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2897 	    0, &inodedep) != 0)
2898 		(void) free_inodedep(inodedep);
2899 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2900 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2901 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
2902 			if (mkdir->md_diradd != dap)
2903 				continue;
2904 			dap->da_state &= ~mkdir->md_state;
2905 			WORKLIST_REMOVE(&mkdir->md_list);
2906 			LIST_REMOVE(mkdir, md_mkdirs);
2907 			WORKITEM_FREE(mkdir, D_MKDIR);
2908 		}
2909 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2910 			FREE_LOCK(&lk);
2911 			panic("free_diradd: unfound ref");
2912 		}
2913 	}
2914 	WORKITEM_FREE(dap, D_DIRADD);
2915 }
2916 
2917 /*
2918  * Directory entry removal dependencies.
2919  *
2920  * When removing a directory entry, the entry's inode pointer must be
2921  * zero'ed on disk before the corresponding inode's link count is decremented
2922  * (possibly freeing the inode for re-use). This dependency is handled by
2923  * updating the directory entry but delaying the inode count reduction until
2924  * after the directory block has been written to disk. After this point, the
2925  * inode count can be decremented whenever it is convenient.
2926  */
2927 
2928 /*
2929  * This routine should be called immediately after removing
2930  * a directory entry.  The inode's link count should not be
2931  * decremented by the calling procedure -- the soft updates
2932  * code will do this task when it is safe.
2933  */
2934 void
2935 softdep_setup_remove(bp, dp, ip, isrmdir)
2936 	struct buf *bp;		/* buffer containing directory block */
2937 	struct inode *dp;	/* inode for the directory being modified */
2938 	struct inode *ip;	/* inode for directory entry being removed */
2939 	int isrmdir;		/* indicates if doing RMDIR */
2940 {
2941 	struct dirrem *dirrem, *prevdirrem;
2942 
2943 	/*
2944 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2945 	 */
2946 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2947 
2948 	/*
2949 	 * If the COMPLETE flag is clear, then there were no active
2950 	 * entries and we want to roll back to a zeroed entry until
2951 	 * the new inode is committed to disk. If the COMPLETE flag is
2952 	 * set then we have deleted an entry that never made it to
2953 	 * disk. If the entry we deleted resulted from a name change,
2954 	 * then the old name still resides on disk. We cannot delete
2955 	 * its inode (returned to us in prevdirrem) until the zeroed
2956 	 * directory entry gets to disk. The new inode has never been
2957 	 * referenced on the disk, so can be deleted immediately.
2958 	 */
2959 	if ((dirrem->dm_state & COMPLETE) == 0) {
2960 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2961 		    dm_next);
2962 		FREE_LOCK(&lk);
2963 	} else {
2964 		if (prevdirrem != NULL)
2965 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2966 			    prevdirrem, dm_next);
2967 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2968 		FREE_LOCK(&lk);
2969 		handle_workitem_remove(dirrem, NULL);
2970 	}
2971 }
2972 
2973 /*
2974  * Allocate a new dirrem if appropriate and return it along with
2975  * its associated pagedep. Called without a lock, returns with lock.
2976  */
2977 static long num_dirrem;		/* number of dirrem allocated */
2978 static struct dirrem *
2979 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2980 	struct buf *bp;		/* buffer containing directory block */
2981 	struct inode *dp;	/* inode for the directory being modified */
2982 	struct inode *ip;	/* inode for directory entry being removed */
2983 	int isrmdir;		/* indicates if doing RMDIR */
2984 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
2985 {
2986 	int offset;
2987 	ufs_lbn_t lbn;
2988 	struct diradd *dap;
2989 	struct dirrem *dirrem;
2990 	struct pagedep *pagedep;
2991 
2992 	/*
2993 	 * Whiteouts have no deletion dependencies.
2994 	 */
2995 	if (ip == NULL)
2996 		panic("newdirrem: whiteout");
2997 	/*
2998 	 * If we are over our limit, try to improve the situation.
2999 	 * Limiting the number of dirrem structures will also limit
3000 	 * the number of freefile and freeblks structures.
3001 	 */
3002 	if (num_dirrem > max_softdeps / 2)
3003 		(void) request_cleanup(FLUSH_REMOVE, 0);
3004 	num_dirrem += 1;
3005 	MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
3006 		M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO);
3007 	dirrem->dm_list.wk_type = D_DIRREM;
3008 	dirrem->dm_state = isrmdir ? RMDIR : 0;
3009 	dirrem->dm_mnt = ITOV(ip)->v_mount;
3010 	dirrem->dm_oldinum = ip->i_number;
3011 	*prevdirremp = NULL;
3012 
3013 	ACQUIRE_LOCK(&lk);
3014 	lbn = lblkno(dp->i_fs, dp->i_offset);
3015 	offset = blkoff(dp->i_fs, dp->i_offset);
3016 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3017 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3018 	dirrem->dm_pagedep = pagedep;
3019 	/*
3020 	 * Check for a diradd dependency for the same directory entry.
3021 	 * If present, then both dependencies become obsolete and can
3022 	 * be de-allocated. Check for an entry on both the pd_dirraddhd
3023 	 * list and the pd_pendinghd list.
3024 	 */
3025 
3026 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
3027 		if (dap->da_offset == offset)
3028 			break;
3029 	if (dap == NULL) {
3030 
3031 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
3032 			if (dap->da_offset == offset)
3033 				break;
3034 		if (dap == NULL)
3035 			return (dirrem);
3036 	}
3037 	/*
3038 	 * Must be ATTACHED at this point.
3039 	 */
3040 	if ((dap->da_state & ATTACHED) == 0) {
3041 		FREE_LOCK(&lk);
3042 		panic("newdirrem: not ATTACHED");
3043 	}
3044 	if (dap->da_newinum != ip->i_number) {
3045 		FREE_LOCK(&lk);
3046 		panic("newdirrem: inum %d should be %d",
3047 		    ip->i_number, dap->da_newinum);
3048 	}
3049 	/*
3050 	 * If we are deleting a changed name that never made it to disk,
3051 	 * then return the dirrem describing the previous inode (which
3052 	 * represents the inode currently referenced from this entry on disk).
3053 	 */
3054 	if ((dap->da_state & DIRCHG) != 0) {
3055 		*prevdirremp = dap->da_previous;
3056 		dap->da_state &= ~DIRCHG;
3057 		dap->da_pagedep = pagedep;
3058 	}
3059 	/*
3060 	 * We are deleting an entry that never made it to disk.
3061 	 * Mark it COMPLETE so we can delete its inode immediately.
3062 	 */
3063 	dirrem->dm_state |= COMPLETE;
3064 	free_diradd(dap);
3065 	return (dirrem);
3066 }
3067 
3068 /*
3069  * Directory entry change dependencies.
3070  *
3071  * Changing an existing directory entry requires that an add operation
3072  * be completed first followed by a deletion. The semantics for the addition
3073  * are identical to the description of adding a new entry above except
3074  * that the rollback is to the old inode number rather than zero. Once
3075  * the addition dependency is completed, the removal is done as described
3076  * in the removal routine above.
3077  */
3078 
3079 /*
3080  * This routine should be called immediately after changing
3081  * a directory entry.  The inode's link count should not be
3082  * decremented by the calling procedure -- the soft updates
3083  * code will perform this task when it is safe.
3084  */
3085 void
3086 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
3087 	struct buf *bp;		/* buffer containing directory block */
3088 	struct inode *dp;	/* inode for the directory being modified */
3089 	struct inode *ip;	/* inode for directory entry being removed */
3090 	ino_t newinum;		/* new inode number for changed entry */
3091 	int isrmdir;		/* indicates if doing RMDIR */
3092 {
3093 	int offset;
3094 	struct diradd *dap = NULL;
3095 	struct dirrem *dirrem, *prevdirrem;
3096 	struct pagedep *pagedep;
3097 	struct inodedep *inodedep;
3098 
3099 	offset = blkoff(dp->i_fs, dp->i_offset);
3100 
3101 	/*
3102 	 * Whiteouts do not need diradd dependencies.
3103 	 */
3104 	if (newinum != WINO) {
3105 		MALLOC(dap, struct diradd *, sizeof(struct diradd),
3106 		    M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
3107 		dap->da_list.wk_type = D_DIRADD;
3108 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
3109 		dap->da_offset = offset;
3110 		dap->da_newinum = newinum;
3111 	}
3112 
3113 	/*
3114 	 * Allocate a new dirrem and ACQUIRE_LOCK.
3115 	 */
3116 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3117 	pagedep = dirrem->dm_pagedep;
3118 	/*
3119 	 * The possible values for isrmdir:
3120 	 *	0 - non-directory file rename
3121 	 *	1 - directory rename within same directory
3122 	 *   inum - directory rename to new directory of given inode number
3123 	 * When renaming to a new directory, we are both deleting and
3124 	 * creating a new directory entry, so the link count on the new
3125 	 * directory should not change. Thus we do not need the followup
3126 	 * dirrem which is usually done in handle_workitem_remove. We set
3127 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
3128 	 * followup dirrem.
3129 	 */
3130 	if (isrmdir > 1)
3131 		dirrem->dm_state |= DIRCHG;
3132 
3133 	/*
3134 	 * Whiteouts have no additional dependencies,
3135 	 * so just put the dirrem on the correct list.
3136 	 */
3137 	if (newinum == WINO) {
3138 		if ((dirrem->dm_state & COMPLETE) == 0) {
3139 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
3140 			    dm_next);
3141 		} else {
3142 			dirrem->dm_dirinum = pagedep->pd_ino;
3143 			add_to_worklist(&dirrem->dm_list);
3144 		}
3145 		FREE_LOCK(&lk);
3146 		return;
3147 	}
3148 
3149 	/*
3150 	 * If the COMPLETE flag is clear, then there were no active
3151 	 * entries and we want to roll back to the previous inode until
3152 	 * the new inode is committed to disk. If the COMPLETE flag is
3153 	 * set, then we have deleted an entry that never made it to disk.
3154 	 * If the entry we deleted resulted from a name change, then the old
3155 	 * inode reference still resides on disk. Any rollback that we do
3156 	 * needs to be to that old inode (returned to us in prevdirrem). If
3157 	 * the entry we deleted resulted from a create, then there is
3158 	 * no entry on the disk, so we want to roll back to zero rather
3159 	 * than the uncommitted inode. In either of the COMPLETE cases we
3160 	 * want to immediately free the unwritten and unreferenced inode.
3161 	 */
3162 	if ((dirrem->dm_state & COMPLETE) == 0) {
3163 		dap->da_previous = dirrem;
3164 	} else {
3165 		if (prevdirrem != NULL) {
3166 			dap->da_previous = prevdirrem;
3167 		} else {
3168 			dap->da_state &= ~DIRCHG;
3169 			dap->da_pagedep = pagedep;
3170 		}
3171 		dirrem->dm_dirinum = pagedep->pd_ino;
3172 		add_to_worklist(&dirrem->dm_list);
3173 	}
3174 	/*
3175 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3176 	 * is not yet written. If it is written, do the post-inode write
3177 	 * processing to put it on the id_pendinghd list.
3178 	 */
3179 	if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
3180 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3181 		dap->da_state |= COMPLETE;
3182 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3183 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3184 	} else {
3185 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
3186 		    dap, da_pdlist);
3187 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3188 	}
3189 	FREE_LOCK(&lk);
3190 }
3191 
3192 /*
3193  * Called whenever the link count on an inode is changed.
3194  * It creates an inode dependency so that the new reference(s)
3195  * to the inode cannot be committed to disk until the updated
3196  * inode has been written.
3197  */
3198 void
3199 softdep_change_linkcnt(ip)
3200 	struct inode *ip;	/* the inode with the increased link count */
3201 {
3202 	struct inodedep *inodedep;
3203 
3204 	ACQUIRE_LOCK(&lk);
3205 	(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
3206 	if (ip->i_nlink < ip->i_effnlink) {
3207 		FREE_LOCK(&lk);
3208 		panic("softdep_change_linkcnt: bad delta");
3209 	}
3210 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3211 	FREE_LOCK(&lk);
3212 }
3213 
3214 /*
3215  * Called when the effective link count and the reference count
3216  * on an inode drops to zero. At this point there are no names
3217  * referencing the file in the filesystem and no active file
3218  * references. The space associated with the file will be freed
3219  * as soon as the necessary soft dependencies are cleared.
3220  */
3221 void
3222 softdep_releasefile(ip)
3223 	struct inode *ip;	/* inode with the zero effective link count */
3224 {
3225 	struct inodedep *inodedep;
3226 	struct fs *fs;
3227 	int extblocks;
3228 
3229 	if (ip->i_effnlink > 0)
3230 		panic("softdep_filerelease: file still referenced");
3231 	/*
3232 	 * We may be called several times as the real reference count
3233 	 * drops to zero. We only want to account for the space once.
3234 	 */
3235 	if (ip->i_flag & IN_SPACECOUNTED)
3236 		return;
3237 	/*
3238 	 * We have to deactivate a snapshot otherwise copyonwrites may
3239 	 * add blocks and the cleanup may remove blocks after we have
3240 	 * tried to account for them.
3241 	 */
3242 	if ((ip->i_flags & SF_SNAPSHOT) != 0)
3243 		ffs_snapremove(ITOV(ip));
3244 	/*
3245 	 * If we are tracking an nlinkdelta, we have to also remember
3246 	 * whether we accounted for the freed space yet.
3247 	 */
3248 	ACQUIRE_LOCK(&lk);
3249 	if ((inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep)))
3250 		inodedep->id_state |= SPACECOUNTED;
3251 	FREE_LOCK(&lk);
3252 	fs = ip->i_fs;
3253 	extblocks = 0;
3254 	if (fs->fs_magic == FS_UFS2_MAGIC)
3255 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
3256 	ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks;
3257 	ip->i_fs->fs_pendinginodes += 1;
3258 	ip->i_flag |= IN_SPACECOUNTED;
3259 }
3260 
3261 /*
3262  * This workitem decrements the inode's link count.
3263  * If the link count reaches zero, the file is removed.
3264  */
3265 static void
3266 handle_workitem_remove(dirrem, xp)
3267 	struct dirrem *dirrem;
3268 	struct vnode *xp;
3269 {
3270 	struct thread *td = curthread;
3271 	struct inodedep *inodedep;
3272 	struct vnode *vp;
3273 	struct inode *ip;
3274 	ino_t oldinum;
3275 	int error;
3276 
3277 	if ((vp = xp) == NULL &&
3278 	    (error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, LK_EXCLUSIVE,
3279 	     &vp)) != 0) {
3280 		softdep_error("handle_workitem_remove: vget", error);
3281 		return;
3282 	}
3283 	ip = VTOI(vp);
3284 	ACQUIRE_LOCK(&lk);
3285 	if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
3286 		FREE_LOCK(&lk);
3287 		panic("handle_workitem_remove: lost inodedep");
3288 	}
3289 	/*
3290 	 * Normal file deletion.
3291 	 */
3292 	if ((dirrem->dm_state & RMDIR) == 0) {
3293 		ip->i_nlink--;
3294 		DIP(ip, i_nlink) = ip->i_nlink;
3295 		ip->i_flag |= IN_CHANGE;
3296 		if (ip->i_nlink < ip->i_effnlink) {
3297 			FREE_LOCK(&lk);
3298 			panic("handle_workitem_remove: bad file delta");
3299 		}
3300 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3301 		FREE_LOCK(&lk);
3302 		vput(vp);
3303 		num_dirrem -= 1;
3304 		WORKITEM_FREE(dirrem, D_DIRREM);
3305 		return;
3306 	}
3307 	/*
3308 	 * Directory deletion. Decrement reference count for both the
3309 	 * just deleted parent directory entry and the reference for ".".
3310 	 * Next truncate the directory to length zero. When the
3311 	 * truncation completes, arrange to have the reference count on
3312 	 * the parent decremented to account for the loss of "..".
3313 	 */
3314 	ip->i_nlink -= 2;
3315 	DIP(ip, i_nlink) = ip->i_nlink;
3316 	ip->i_flag |= IN_CHANGE;
3317 	if (ip->i_nlink < ip->i_effnlink) {
3318 		FREE_LOCK(&lk);
3319 		panic("handle_workitem_remove: bad dir delta");
3320 	}
3321 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3322 	FREE_LOCK(&lk);
3323 	if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, td->td_ucred, td)) != 0)
3324 		softdep_error("handle_workitem_remove: truncate", error);
3325 	/*
3326 	 * Rename a directory to a new parent. Since, we are both deleting
3327 	 * and creating a new directory entry, the link count on the new
3328 	 * directory should not change. Thus we skip the followup dirrem.
3329 	 */
3330 	if (dirrem->dm_state & DIRCHG) {
3331 		vput(vp);
3332 		num_dirrem -= 1;
3333 		WORKITEM_FREE(dirrem, D_DIRREM);
3334 		return;
3335 	}
3336 	/*
3337 	 * If the inodedep does not exist, then the zero'ed inode has
3338 	 * been written to disk. If the allocated inode has never been
3339 	 * written to disk, then the on-disk inode is zero'ed. In either
3340 	 * case we can remove the file immediately.
3341 	 */
3342 	ACQUIRE_LOCK(&lk);
3343 	dirrem->dm_state = 0;
3344 	oldinum = dirrem->dm_oldinum;
3345 	dirrem->dm_oldinum = dirrem->dm_dirinum;
3346 	if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
3347 	    check_inode_unwritten(inodedep)) {
3348 		FREE_LOCK(&lk);
3349 		vput(vp);
3350 		handle_workitem_remove(dirrem, NULL);
3351 		return;
3352 	}
3353 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3354 	FREE_LOCK(&lk);
3355 	vput(vp);
3356 }
3357 
3358 /*
3359  * Inode de-allocation dependencies.
3360  *
3361  * When an inode's link count is reduced to zero, it can be de-allocated. We
3362  * found it convenient to postpone de-allocation until after the inode is
3363  * written to disk with its new link count (zero).  At this point, all of the
3364  * on-disk inode's block pointers are nullified and, with careful dependency
3365  * list ordering, all dependencies related to the inode will be satisfied and
3366  * the corresponding dependency structures de-allocated.  So, if/when the
3367  * inode is reused, there will be no mixing of old dependencies with new
3368  * ones.  This artificial dependency is set up by the block de-allocation
3369  * procedure above (softdep_setup_freeblocks) and completed by the
3370  * following procedure.
3371  */
3372 static void
3373 handle_workitem_freefile(freefile)
3374 	struct freefile *freefile;
3375 {
3376 	struct fs *fs;
3377 	struct inodedep *idp;
3378 	int error;
3379 
3380 	fs = VFSTOUFS(freefile->fx_mnt)->um_fs;
3381 #ifdef DEBUG
3382 	ACQUIRE_LOCK(&lk);
3383 	error = inodedep_lookup(fs, freefile->fx_oldinum, 0, &idp);
3384 	FREE_LOCK(&lk);
3385 	if (error)
3386 		panic("handle_workitem_freefile: inodedep survived");
3387 #endif
3388 	fs->fs_pendinginodes -= 1;
3389 	if ((error = ffs_freefile(fs, freefile->fx_devvp, freefile->fx_oldinum,
3390 	     freefile->fx_mode)) != 0)
3391 		softdep_error("handle_workitem_freefile", error);
3392 	WORKITEM_FREE(freefile, D_FREEFILE);
3393 }
3394 
3395 /*
3396  * Disk writes.
3397  *
3398  * The dependency structures constructed above are most actively used when file
3399  * system blocks are written to disk.  No constraints are placed on when a
3400  * block can be written, but unsatisfied update dependencies are made safe by
3401  * modifying (or replacing) the source memory for the duration of the disk
3402  * write.  When the disk write completes, the memory block is again brought
3403  * up-to-date.
3404  *
3405  * In-core inode structure reclamation.
3406  *
3407  * Because there are a finite number of "in-core" inode structures, they are
3408  * reused regularly.  By transferring all inode-related dependencies to the
3409  * in-memory inode block and indexing them separately (via "inodedep"s), we
3410  * can allow "in-core" inode structures to be reused at any time and avoid
3411  * any increase in contention.
3412  *
3413  * Called just before entering the device driver to initiate a new disk I/O.
3414  * The buffer must be locked, thus, no I/O completion operations can occur
3415  * while we are manipulating its associated dependencies.
3416  */
3417 static void
3418 softdep_disk_io_initiation(bp)
3419 	struct buf *bp;		/* structure describing disk write to occur */
3420 {
3421 	struct worklist *wk, *nextwk;
3422 	struct indirdep *indirdep;
3423 	struct inodedep *inodedep;
3424 
3425 	/*
3426 	 * We only care about write operations. There should never
3427 	 * be dependencies for reads.
3428 	 */
3429 	if (bp->b_iocmd == BIO_READ)
3430 		panic("softdep_disk_io_initiation: read");
3431 	/*
3432 	 * Do any necessary pre-I/O processing.
3433 	 */
3434 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
3435 		nextwk = LIST_NEXT(wk, wk_list);
3436 		switch (wk->wk_type) {
3437 
3438 		case D_PAGEDEP:
3439 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
3440 			continue;
3441 
3442 		case D_INODEDEP:
3443 			inodedep = WK_INODEDEP(wk);
3444 			if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
3445 				initiate_write_inodeblock_ufs1(inodedep, bp);
3446 			else
3447 				initiate_write_inodeblock_ufs2(inodedep, bp);
3448 			continue;
3449 
3450 		case D_INDIRDEP:
3451 			indirdep = WK_INDIRDEP(wk);
3452 			if (indirdep->ir_state & GOINGAWAY)
3453 				panic("disk_io_initiation: indirdep gone");
3454 			/*
3455 			 * If there are no remaining dependencies, this
3456 			 * will be writing the real pointers, so the
3457 			 * dependency can be freed.
3458 			 */
3459 			if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
3460 				indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
3461 				brelse(indirdep->ir_savebp);
3462 				/* inline expand WORKLIST_REMOVE(wk); */
3463 				wk->wk_state &= ~ONWORKLIST;
3464 				LIST_REMOVE(wk, wk_list);
3465 				WORKITEM_FREE(indirdep, D_INDIRDEP);
3466 				continue;
3467 			}
3468 			/*
3469 			 * Replace up-to-date version with safe version.
3470 			 */
3471 			MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3472 			    M_INDIRDEP, M_SOFTDEP_FLAGS);
3473 			ACQUIRE_LOCK(&lk);
3474 			indirdep->ir_state &= ~ATTACHED;
3475 			indirdep->ir_state |= UNDONE;
3476 			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3477 			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3478 			    bp->b_bcount);
3479 			FREE_LOCK(&lk);
3480 			continue;
3481 
3482 		case D_MKDIR:
3483 		case D_BMSAFEMAP:
3484 		case D_ALLOCDIRECT:
3485 		case D_ALLOCINDIR:
3486 			continue;
3487 
3488 		default:
3489 			panic("handle_disk_io_initiation: Unexpected type %s",
3490 			    TYPENAME(wk->wk_type));
3491 			/* NOTREACHED */
3492 		}
3493 	}
3494 }
3495 
3496 /*
3497  * Called from within the procedure above to deal with unsatisfied
3498  * allocation dependencies in a directory. The buffer must be locked,
3499  * thus, no I/O completion operations can occur while we are
3500  * manipulating its associated dependencies.
3501  */
3502 static void
3503 initiate_write_filepage(pagedep, bp)
3504 	struct pagedep *pagedep;
3505 	struct buf *bp;
3506 {
3507 	struct diradd *dap;
3508 	struct direct *ep;
3509 	int i;
3510 
3511 	if (pagedep->pd_state & IOSTARTED) {
3512 		/*
3513 		 * This can only happen if there is a driver that does not
3514 		 * understand chaining. Here biodone will reissue the call
3515 		 * to strategy for the incomplete buffers.
3516 		 */
3517 		printf("initiate_write_filepage: already started\n");
3518 		return;
3519 	}
3520 	pagedep->pd_state |= IOSTARTED;
3521 	ACQUIRE_LOCK(&lk);
3522 	for (i = 0; i < DAHASHSZ; i++) {
3523 		LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3524 			ep = (struct direct *)
3525 			    ((char *)bp->b_data + dap->da_offset);
3526 			if (ep->d_ino != dap->da_newinum) {
3527 				FREE_LOCK(&lk);
3528 				panic("%s: dir inum %d != new %d",
3529 				    "initiate_write_filepage",
3530 				    ep->d_ino, dap->da_newinum);
3531 			}
3532 			if (dap->da_state & DIRCHG)
3533 				ep->d_ino = dap->da_previous->dm_oldinum;
3534 			else
3535 				ep->d_ino = 0;
3536 			dap->da_state &= ~ATTACHED;
3537 			dap->da_state |= UNDONE;
3538 		}
3539 	}
3540 	FREE_LOCK(&lk);
3541 }
3542 
3543 /*
3544  * Version of initiate_write_inodeblock that handles UFS1 dinodes.
3545  * Note that any bug fixes made to this routine must be done in the
3546  * version found below.
3547  *
3548  * Called from within the procedure above to deal with unsatisfied
3549  * allocation dependencies in an inodeblock. The buffer must be
3550  * locked, thus, no I/O completion operations can occur while we
3551  * are manipulating its associated dependencies.
3552  */
3553 static void
3554 initiate_write_inodeblock_ufs1(inodedep, bp)
3555 	struct inodedep *inodedep;
3556 	struct buf *bp;			/* The inode block */
3557 {
3558 	struct allocdirect *adp, *lastadp;
3559 	struct ufs1_dinode *dp;
3560 	struct fs *fs;
3561 	ufs_lbn_t i, prevlbn = 0;
3562 	int deplist;
3563 
3564 	if (inodedep->id_state & IOSTARTED)
3565 		panic("initiate_write_inodeblock_ufs1: already started");
3566 	inodedep->id_state |= IOSTARTED;
3567 	fs = inodedep->id_fs;
3568 	dp = (struct ufs1_dinode *)bp->b_data +
3569 	    ino_to_fsbo(fs, inodedep->id_ino);
3570 	/*
3571 	 * If the bitmap is not yet written, then the allocated
3572 	 * inode cannot be written to disk.
3573 	 */
3574 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3575 		if (inodedep->id_savedino1 != NULL)
3576 			panic("initiate_write_inodeblock_ufs1: I/O underway");
3577 		MALLOC(inodedep->id_savedino1, struct ufs1_dinode *,
3578 		    sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3579 		*inodedep->id_savedino1 = *dp;
3580 		bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
3581 		return;
3582 	}
3583 	/*
3584 	 * If no dependencies, then there is nothing to roll back.
3585 	 */
3586 	inodedep->id_savedsize = dp->di_size;
3587 	inodedep->id_savedextsize = 0;
3588 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3589 		return;
3590 	/*
3591 	 * Set the dependencies to busy.
3592 	 */
3593 	ACQUIRE_LOCK(&lk);
3594 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3595 	     adp = TAILQ_NEXT(adp, ad_next)) {
3596 #ifdef DIAGNOSTIC
3597 		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3598 			FREE_LOCK(&lk);
3599 			panic("softdep_write_inodeblock: lbn order");
3600 		}
3601 		prevlbn = adp->ad_lbn;
3602 		if (adp->ad_lbn < NDADDR &&
3603 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3604 			FREE_LOCK(&lk);
3605 			panic("%s: direct pointer #%jd mismatch %d != %jd",
3606 			    "softdep_write_inodeblock",
3607 			    (intmax_t)adp->ad_lbn,
3608 			    dp->di_db[adp->ad_lbn],
3609 			    (intmax_t)adp->ad_newblkno);
3610 		}
3611 		if (adp->ad_lbn >= NDADDR &&
3612 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3613 			FREE_LOCK(&lk);
3614 			panic("%s: indirect pointer #%jd mismatch %d != %jd",
3615 			    "softdep_write_inodeblock",
3616 			    (intmax_t)adp->ad_lbn - NDADDR,
3617 			    dp->di_ib[adp->ad_lbn - NDADDR],
3618 			    (intmax_t)adp->ad_newblkno);
3619 		}
3620 		deplist |= 1 << adp->ad_lbn;
3621 		if ((adp->ad_state & ATTACHED) == 0) {
3622 			FREE_LOCK(&lk);
3623 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3624 			    adp->ad_state);
3625 		}
3626 #endif /* DIAGNOSTIC */
3627 		adp->ad_state &= ~ATTACHED;
3628 		adp->ad_state |= UNDONE;
3629 	}
3630 	/*
3631 	 * The on-disk inode cannot claim to be any larger than the last
3632 	 * fragment that has been written. Otherwise, the on-disk inode
3633 	 * might have fragments that were not the last block in the file
3634 	 * which would corrupt the filesystem.
3635 	 */
3636 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3637 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3638 		if (adp->ad_lbn >= NDADDR)
3639 			break;
3640 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3641 		/* keep going until hitting a rollback to a frag */
3642 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3643 			continue;
3644 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3645 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3646 #ifdef DIAGNOSTIC
3647 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3648 				FREE_LOCK(&lk);
3649 				panic("softdep_write_inodeblock: lost dep1");
3650 			}
3651 #endif /* DIAGNOSTIC */
3652 			dp->di_db[i] = 0;
3653 		}
3654 		for (i = 0; i < NIADDR; i++) {
3655 #ifdef DIAGNOSTIC
3656 			if (dp->di_ib[i] != 0 &&
3657 			    (deplist & ((1 << NDADDR) << i)) == 0) {
3658 				FREE_LOCK(&lk);
3659 				panic("softdep_write_inodeblock: lost dep2");
3660 			}
3661 #endif /* DIAGNOSTIC */
3662 			dp->di_ib[i] = 0;
3663 		}
3664 		FREE_LOCK(&lk);
3665 		return;
3666 	}
3667 	/*
3668 	 * If we have zero'ed out the last allocated block of the file,
3669 	 * roll back the size to the last currently allocated block.
3670 	 * We know that this last allocated block is a full-sized as
3671 	 * we already checked for fragments in the loop above.
3672 	 */
3673 	if (lastadp != NULL &&
3674 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3675 		for (i = lastadp->ad_lbn; i >= 0; i--)
3676 			if (dp->di_db[i] != 0)
3677 				break;
3678 		dp->di_size = (i + 1) * fs->fs_bsize;
3679 	}
3680 	/*
3681 	 * The only dependencies are for indirect blocks.
3682 	 *
3683 	 * The file size for indirect block additions is not guaranteed.
3684 	 * Such a guarantee would be non-trivial to achieve. The conventional
3685 	 * synchronous write implementation also does not make this guarantee.
3686 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3687 	 * can be over-estimated without destroying integrity when the file
3688 	 * moves into the indirect blocks (i.e., is large). If we want to
3689 	 * postpone fsck, we are stuck with this argument.
3690 	 */
3691 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3692 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3693 	FREE_LOCK(&lk);
3694 }
3695 
3696 /*
3697  * Version of initiate_write_inodeblock that handles UFS2 dinodes.
3698  * Note that any bug fixes made to this routine must be done in the
3699  * version found above.
3700  *
3701  * Called from within the procedure above to deal with unsatisfied
3702  * allocation dependencies in an inodeblock. The buffer must be
3703  * locked, thus, no I/O completion operations can occur while we
3704  * are manipulating its associated dependencies.
3705  */
3706 static void
3707 initiate_write_inodeblock_ufs2(inodedep, bp)
3708 	struct inodedep *inodedep;
3709 	struct buf *bp;			/* The inode block */
3710 {
3711 	struct allocdirect *adp, *lastadp;
3712 	struct ufs2_dinode *dp;
3713 	struct fs *fs;
3714 	ufs_lbn_t i, prevlbn = 0;
3715 	int deplist;
3716 
3717 	if (inodedep->id_state & IOSTARTED)
3718 		panic("initiate_write_inodeblock_ufs2: already started");
3719 	inodedep->id_state |= IOSTARTED;
3720 	fs = inodedep->id_fs;
3721 	dp = (struct ufs2_dinode *)bp->b_data +
3722 	    ino_to_fsbo(fs, inodedep->id_ino);
3723 	/*
3724 	 * If the bitmap is not yet written, then the allocated
3725 	 * inode cannot be written to disk.
3726 	 */
3727 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3728 		if (inodedep->id_savedino2 != NULL)
3729 			panic("initiate_write_inodeblock_ufs2: I/O underway");
3730 		MALLOC(inodedep->id_savedino2, struct ufs2_dinode *,
3731 		    sizeof(struct ufs2_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3732 		*inodedep->id_savedino2 = *dp;
3733 		bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
3734 		return;
3735 	}
3736 	/*
3737 	 * If no dependencies, then there is nothing to roll back.
3738 	 */
3739 	inodedep->id_savedsize = dp->di_size;
3740 	inodedep->id_savedextsize = dp->di_extsize;
3741 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL &&
3742 	    TAILQ_FIRST(&inodedep->id_extupdt) == NULL)
3743 		return;
3744 	/*
3745 	 * Set the ext data dependencies to busy.
3746 	 */
3747 	ACQUIRE_LOCK(&lk);
3748 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
3749 	     adp = TAILQ_NEXT(adp, ad_next)) {
3750 #ifdef DIAGNOSTIC
3751 		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3752 			FREE_LOCK(&lk);
3753 			panic("softdep_write_inodeblock: lbn order");
3754 		}
3755 		prevlbn = adp->ad_lbn;
3756 		if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno) {
3757 			FREE_LOCK(&lk);
3758 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
3759 			    "softdep_write_inodeblock",
3760 			    (intmax_t)adp->ad_lbn,
3761 			    (intmax_t)dp->di_extb[adp->ad_lbn],
3762 			    (intmax_t)adp->ad_newblkno);
3763 		}
3764 		deplist |= 1 << adp->ad_lbn;
3765 		if ((adp->ad_state & ATTACHED) == 0) {
3766 			FREE_LOCK(&lk);
3767 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3768 			    adp->ad_state);
3769 		}
3770 #endif /* DIAGNOSTIC */
3771 		adp->ad_state &= ~ATTACHED;
3772 		adp->ad_state |= UNDONE;
3773 	}
3774 	/*
3775 	 * The on-disk inode cannot claim to be any larger than the last
3776 	 * fragment that has been written. Otherwise, the on-disk inode
3777 	 * might have fragments that were not the last block in the ext
3778 	 * data which would corrupt the filesystem.
3779 	 */
3780 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
3781 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3782 		dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno;
3783 		/* keep going until hitting a rollback to a frag */
3784 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3785 			continue;
3786 		dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3787 		for (i = adp->ad_lbn + 1; i < NXADDR; i++) {
3788 #ifdef DIAGNOSTIC
3789 			if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) {
3790 				FREE_LOCK(&lk);
3791 				panic("softdep_write_inodeblock: lost dep1");
3792 			}
3793 #endif /* DIAGNOSTIC */
3794 			dp->di_extb[i] = 0;
3795 		}
3796 		lastadp = NULL;
3797 		break;
3798 	}
3799 	/*
3800 	 * If we have zero'ed out the last allocated block of the ext
3801 	 * data, roll back the size to the last currently allocated block.
3802 	 * We know that this last allocated block is a full-sized as
3803 	 * we already checked for fragments in the loop above.
3804 	 */
3805 	if (lastadp != NULL &&
3806 	    dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3807 		for (i = lastadp->ad_lbn; i >= 0; i--)
3808 			if (dp->di_extb[i] != 0)
3809 				break;
3810 		dp->di_extsize = (i + 1) * fs->fs_bsize;
3811 	}
3812 	/*
3813 	 * Set the file data dependencies to busy.
3814 	 */
3815 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3816 	     adp = TAILQ_NEXT(adp, ad_next)) {
3817 #ifdef DIAGNOSTIC
3818 		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3819 			FREE_LOCK(&lk);
3820 			panic("softdep_write_inodeblock: lbn order");
3821 		}
3822 		prevlbn = adp->ad_lbn;
3823 		if (adp->ad_lbn < NDADDR &&
3824 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3825 			FREE_LOCK(&lk);
3826 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
3827 			    "softdep_write_inodeblock",
3828 			    (intmax_t)adp->ad_lbn,
3829 			    (intmax_t)dp->di_db[adp->ad_lbn],
3830 			    (intmax_t)adp->ad_newblkno);
3831 		}
3832 		if (adp->ad_lbn >= NDADDR &&
3833 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3834 			FREE_LOCK(&lk);
3835 			panic("%s indirect pointer #%jd mismatch %jd != %jd",
3836 			    "softdep_write_inodeblock:",
3837 			    (intmax_t)adp->ad_lbn - NDADDR,
3838 			    (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR],
3839 			    (intmax_t)adp->ad_newblkno);
3840 		}
3841 		deplist |= 1 << adp->ad_lbn;
3842 		if ((adp->ad_state & ATTACHED) == 0) {
3843 			FREE_LOCK(&lk);
3844 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3845 			    adp->ad_state);
3846 		}
3847 #endif /* DIAGNOSTIC */
3848 		adp->ad_state &= ~ATTACHED;
3849 		adp->ad_state |= UNDONE;
3850 	}
3851 	/*
3852 	 * The on-disk inode cannot claim to be any larger than the last
3853 	 * fragment that has been written. Otherwise, the on-disk inode
3854 	 * might have fragments that were not the last block in the file
3855 	 * which would corrupt the filesystem.
3856 	 */
3857 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3858 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3859 		if (adp->ad_lbn >= NDADDR)
3860 			break;
3861 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3862 		/* keep going until hitting a rollback to a frag */
3863 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3864 			continue;
3865 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3866 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3867 #ifdef DIAGNOSTIC
3868 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3869 				FREE_LOCK(&lk);
3870 				panic("softdep_write_inodeblock: lost dep2");
3871 			}
3872 #endif /* DIAGNOSTIC */
3873 			dp->di_db[i] = 0;
3874 		}
3875 		for (i = 0; i < NIADDR; i++) {
3876 #ifdef DIAGNOSTIC
3877 			if (dp->di_ib[i] != 0 &&
3878 			    (deplist & ((1 << NDADDR) << i)) == 0) {
3879 				FREE_LOCK(&lk);
3880 				panic("softdep_write_inodeblock: lost dep3");
3881 			}
3882 #endif /* DIAGNOSTIC */
3883 			dp->di_ib[i] = 0;
3884 		}
3885 		FREE_LOCK(&lk);
3886 		return;
3887 	}
3888 	/*
3889 	 * If we have zero'ed out the last allocated block of the file,
3890 	 * roll back the size to the last currently allocated block.
3891 	 * We know that this last allocated block is a full-sized as
3892 	 * we already checked for fragments in the loop above.
3893 	 */
3894 	if (lastadp != NULL &&
3895 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3896 		for (i = lastadp->ad_lbn; i >= 0; i--)
3897 			if (dp->di_db[i] != 0)
3898 				break;
3899 		dp->di_size = (i + 1) * fs->fs_bsize;
3900 	}
3901 	/*
3902 	 * The only dependencies are for indirect blocks.
3903 	 *
3904 	 * The file size for indirect block additions is not guaranteed.
3905 	 * Such a guarantee would be non-trivial to achieve. The conventional
3906 	 * synchronous write implementation also does not make this guarantee.
3907 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3908 	 * can be over-estimated without destroying integrity when the file
3909 	 * moves into the indirect blocks (i.e., is large). If we want to
3910 	 * postpone fsck, we are stuck with this argument.
3911 	 */
3912 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3913 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3914 	FREE_LOCK(&lk);
3915 }
3916 
3917 /*
3918  * This routine is called during the completion interrupt
3919  * service routine for a disk write (from the procedure called
3920  * by the device driver to inform the filesystem caches of
3921  * a request completion).  It should be called early in this
3922  * procedure, before the block is made available to other
3923  * processes or other routines are called.
3924  */
3925 static void
3926 softdep_disk_write_complete(bp)
3927 	struct buf *bp;		/* describes the completed disk write */
3928 {
3929 	struct worklist *wk;
3930 	struct workhead reattach;
3931 	struct newblk *newblk;
3932 	struct allocindir *aip;
3933 	struct allocdirect *adp;
3934 	struct indirdep *indirdep;
3935 	struct inodedep *inodedep;
3936 	struct bmsafemap *bmsafemap;
3937 
3938 #ifdef DEBUG
3939 	if (lk.lkt_held != NOHOLDER)
3940 		panic("softdep_disk_write_complete: lock is held");
3941 	lk.lkt_held = SPECIAL_FLAG;
3942 #endif
3943 	LIST_INIT(&reattach);
3944 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3945 		WORKLIST_REMOVE(wk);
3946 		switch (wk->wk_type) {
3947 
3948 		case D_PAGEDEP:
3949 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3950 				WORKLIST_INSERT(&reattach, wk);
3951 			continue;
3952 
3953 		case D_INODEDEP:
3954 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3955 				WORKLIST_INSERT(&reattach, wk);
3956 			continue;
3957 
3958 		case D_BMSAFEMAP:
3959 			bmsafemap = WK_BMSAFEMAP(wk);
3960 			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3961 				newblk->nb_state |= DEPCOMPLETE;
3962 				newblk->nb_bmsafemap = NULL;
3963 				LIST_REMOVE(newblk, nb_deps);
3964 			}
3965 			while ((adp =
3966 			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3967 				adp->ad_state |= DEPCOMPLETE;
3968 				adp->ad_buf = NULL;
3969 				LIST_REMOVE(adp, ad_deps);
3970 				handle_allocdirect_partdone(adp);
3971 			}
3972 			while ((aip =
3973 			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3974 				aip->ai_state |= DEPCOMPLETE;
3975 				aip->ai_buf = NULL;
3976 				LIST_REMOVE(aip, ai_deps);
3977 				handle_allocindir_partdone(aip);
3978 			}
3979 			while ((inodedep =
3980 			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3981 				inodedep->id_state |= DEPCOMPLETE;
3982 				LIST_REMOVE(inodedep, id_deps);
3983 				inodedep->id_buf = NULL;
3984 			}
3985 			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3986 			continue;
3987 
3988 		case D_MKDIR:
3989 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3990 			continue;
3991 
3992 		case D_ALLOCDIRECT:
3993 			adp = WK_ALLOCDIRECT(wk);
3994 			adp->ad_state |= COMPLETE;
3995 			handle_allocdirect_partdone(adp);
3996 			continue;
3997 
3998 		case D_ALLOCINDIR:
3999 			aip = WK_ALLOCINDIR(wk);
4000 			aip->ai_state |= COMPLETE;
4001 			handle_allocindir_partdone(aip);
4002 			continue;
4003 
4004 		case D_INDIRDEP:
4005 			indirdep = WK_INDIRDEP(wk);
4006 			if (indirdep->ir_state & GOINGAWAY) {
4007 				lk.lkt_held = NOHOLDER;
4008 				panic("disk_write_complete: indirdep gone");
4009 			}
4010 			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
4011 			FREE(indirdep->ir_saveddata, M_INDIRDEP);
4012 			indirdep->ir_saveddata = 0;
4013 			indirdep->ir_state &= ~UNDONE;
4014 			indirdep->ir_state |= ATTACHED;
4015 			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
4016 				handle_allocindir_partdone(aip);
4017 				if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
4018 					lk.lkt_held = NOHOLDER;
4019 					panic("disk_write_complete: not gone");
4020 				}
4021 			}
4022 			WORKLIST_INSERT(&reattach, wk);
4023 			if ((bp->b_flags & B_DELWRI) == 0)
4024 				stat_indir_blk_ptrs++;
4025 			bdirty(bp);
4026 			continue;
4027 
4028 		default:
4029 			lk.lkt_held = NOHOLDER;
4030 			panic("handle_disk_write_complete: Unknown type %s",
4031 			    TYPENAME(wk->wk_type));
4032 			/* NOTREACHED */
4033 		}
4034 	}
4035 	/*
4036 	 * Reattach any requests that must be redone.
4037 	 */
4038 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
4039 		WORKLIST_REMOVE(wk);
4040 		WORKLIST_INSERT(&bp->b_dep, wk);
4041 	}
4042 #ifdef DEBUG
4043 	if (lk.lkt_held != SPECIAL_FLAG)
4044 		panic("softdep_disk_write_complete: lock lost");
4045 	lk.lkt_held = NOHOLDER;
4046 #endif
4047 }
4048 
4049 /*
4050  * Called from within softdep_disk_write_complete above. Note that
4051  * this routine is always called from interrupt level with further
4052  * splbio interrupts blocked.
4053  */
4054 static void
4055 handle_allocdirect_partdone(adp)
4056 	struct allocdirect *adp;	/* the completed allocdirect */
4057 {
4058 	struct allocdirectlst *listhead;
4059 	struct allocdirect *listadp;
4060 	struct inodedep *inodedep;
4061 	long bsize, delay;
4062 
4063 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4064 		return;
4065 	if (adp->ad_buf != NULL) {
4066 		lk.lkt_held = NOHOLDER;
4067 		panic("handle_allocdirect_partdone: dangling dep");
4068 	}
4069 	/*
4070 	 * The on-disk inode cannot claim to be any larger than the last
4071 	 * fragment that has been written. Otherwise, the on-disk inode
4072 	 * might have fragments that were not the last block in the file
4073 	 * which would corrupt the filesystem. Thus, we cannot free any
4074 	 * allocdirects after one whose ad_oldblkno claims a fragment as
4075 	 * these blocks must be rolled back to zero before writing the inode.
4076 	 * We check the currently active set of allocdirects in id_inoupdt
4077 	 * or id_extupdt as appropriate.
4078 	 */
4079 	inodedep = adp->ad_inodedep;
4080 	bsize = inodedep->id_fs->fs_bsize;
4081 	if (adp->ad_state & EXTDATA)
4082 		listhead = &inodedep->id_extupdt;
4083 	else
4084 		listhead = &inodedep->id_inoupdt;
4085 	TAILQ_FOREACH(listadp, listhead, ad_next) {
4086 		/* found our block */
4087 		if (listadp == adp)
4088 			break;
4089 		/* continue if ad_oldlbn is not a fragment */
4090 		if (listadp->ad_oldsize == 0 ||
4091 		    listadp->ad_oldsize == bsize)
4092 			continue;
4093 		/* hit a fragment */
4094 		return;
4095 	}
4096 	/*
4097 	 * If we have reached the end of the current list without
4098 	 * finding the just finished dependency, then it must be
4099 	 * on the future dependency list. Future dependencies cannot
4100 	 * be freed until they are moved to the current list.
4101 	 */
4102 	if (listadp == NULL) {
4103 #ifdef DEBUG
4104 		if (adp->ad_state & EXTDATA)
4105 			listhead = &inodedep->id_newextupdt;
4106 		else
4107 			listhead = &inodedep->id_newinoupdt;
4108 		TAILQ_FOREACH(listadp, listhead, ad_next)
4109 			/* found our block */
4110 			if (listadp == adp)
4111 				break;
4112 		if (listadp == NULL) {
4113 			lk.lkt_held = NOHOLDER;
4114 			panic("handle_allocdirect_partdone: lost dep");
4115 		}
4116 #endif /* DEBUG */
4117 		return;
4118 	}
4119 	/*
4120 	 * If we have found the just finished dependency, then free
4121 	 * it along with anything that follows it that is complete.
4122 	 * If the inode still has a bitmap dependency, then it has
4123 	 * never been written to disk, hence the on-disk inode cannot
4124 	 * reference the old fragment so we can free it without delay.
4125 	 */
4126 	delay = (inodedep->id_state & DEPCOMPLETE);
4127 	for (; adp; adp = listadp) {
4128 		listadp = TAILQ_NEXT(adp, ad_next);
4129 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4130 			return;
4131 		free_allocdirect(listhead, adp, delay);
4132 	}
4133 }
4134 
4135 /*
4136  * Called from within softdep_disk_write_complete above. Note that
4137  * this routine is always called from interrupt level with further
4138  * splbio interrupts blocked.
4139  */
4140 static void
4141 handle_allocindir_partdone(aip)
4142 	struct allocindir *aip;		/* the completed allocindir */
4143 {
4144 	struct indirdep *indirdep;
4145 
4146 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
4147 		return;
4148 	if (aip->ai_buf != NULL) {
4149 		lk.lkt_held = NOHOLDER;
4150 		panic("handle_allocindir_partdone: dangling dependency");
4151 	}
4152 	indirdep = aip->ai_indirdep;
4153 	if (indirdep->ir_state & UNDONE) {
4154 		LIST_REMOVE(aip, ai_next);
4155 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
4156 		return;
4157 	}
4158 	if (indirdep->ir_state & UFS1FMT)
4159 		((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4160 		    aip->ai_newblkno;
4161 	else
4162 		((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4163 		    aip->ai_newblkno;
4164 	LIST_REMOVE(aip, ai_next);
4165 	if (aip->ai_freefrag != NULL)
4166 		add_to_worklist(&aip->ai_freefrag->ff_list);
4167 	WORKITEM_FREE(aip, D_ALLOCINDIR);
4168 }
4169 
4170 /*
4171  * Called from within softdep_disk_write_complete above to restore
4172  * in-memory inode block contents to their most up-to-date state. Note
4173  * that this routine is always called from interrupt level with further
4174  * splbio interrupts blocked.
4175  */
4176 static int
4177 handle_written_inodeblock(inodedep, bp)
4178 	struct inodedep *inodedep;
4179 	struct buf *bp;		/* buffer containing the inode block */
4180 {
4181 	struct worklist *wk, *filefree;
4182 	struct allocdirect *adp, *nextadp;
4183 	struct ufs1_dinode *dp1 = NULL;
4184 	struct ufs2_dinode *dp2 = NULL;
4185 	int hadchanges, fstype;
4186 
4187 	if ((inodedep->id_state & IOSTARTED) == 0) {
4188 		lk.lkt_held = NOHOLDER;
4189 		panic("handle_written_inodeblock: not started");
4190 	}
4191 	inodedep->id_state &= ~IOSTARTED;
4192 	inodedep->id_state |= COMPLETE;
4193 	if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
4194 		fstype = UFS1;
4195 		dp1 = (struct ufs1_dinode *)bp->b_data +
4196 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4197 	} else {
4198 		fstype = UFS2;
4199 		dp2 = (struct ufs2_dinode *)bp->b_data +
4200 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4201 	}
4202 	/*
4203 	 * If we had to rollback the inode allocation because of
4204 	 * bitmaps being incomplete, then simply restore it.
4205 	 * Keep the block dirty so that it will not be reclaimed until
4206 	 * all associated dependencies have been cleared and the
4207 	 * corresponding updates written to disk.
4208 	 */
4209 	if (inodedep->id_savedino1 != NULL) {
4210 		if (fstype == UFS1)
4211 			*dp1 = *inodedep->id_savedino1;
4212 		else
4213 			*dp2 = *inodedep->id_savedino2;
4214 		FREE(inodedep->id_savedino1, M_INODEDEP);
4215 		inodedep->id_savedino1 = NULL;
4216 		if ((bp->b_flags & B_DELWRI) == 0)
4217 			stat_inode_bitmap++;
4218 		bdirty(bp);
4219 		return (1);
4220 	}
4221 	/*
4222 	 * Roll forward anything that had to be rolled back before
4223 	 * the inode could be updated.
4224 	 */
4225 	hadchanges = 0;
4226 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
4227 		nextadp = TAILQ_NEXT(adp, ad_next);
4228 		if (adp->ad_state & ATTACHED) {
4229 			lk.lkt_held = NOHOLDER;
4230 			panic("handle_written_inodeblock: new entry");
4231 		}
4232 		if (fstype == UFS1) {
4233 			if (adp->ad_lbn < NDADDR) {
4234 				if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno) {
4235 					lk.lkt_held = NOHOLDER;
4236 					panic("%s %s #%jd mismatch %d != %jd",
4237 					    "handle_written_inodeblock:",
4238 					    "direct pointer",
4239 					    (intmax_t)adp->ad_lbn,
4240 					    dp1->di_db[adp->ad_lbn],
4241 					    (intmax_t)adp->ad_oldblkno);
4242 				}
4243 				dp1->di_db[adp->ad_lbn] = adp->ad_newblkno;
4244 			} else {
4245 				if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0) {
4246 					lk.lkt_held = NOHOLDER;
4247 					panic("%s: %s #%jd allocated as %d",
4248 					    "handle_written_inodeblock",
4249 					    "indirect pointer",
4250 					    (intmax_t)adp->ad_lbn - NDADDR,
4251 					    dp1->di_ib[adp->ad_lbn - NDADDR]);
4252 				}
4253 				dp1->di_ib[adp->ad_lbn - NDADDR] =
4254 				    adp->ad_newblkno;
4255 			}
4256 		} else {
4257 			if (adp->ad_lbn < NDADDR) {
4258 				if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno) {
4259 					lk.lkt_held = NOHOLDER;
4260 					panic("%s: %s #%jd %s %jd != %jd",
4261 					    "handle_written_inodeblock",
4262 					    "direct pointer",
4263 					    (intmax_t)adp->ad_lbn, "mismatch",
4264 					    (intmax_t)dp2->di_db[adp->ad_lbn],
4265 					    (intmax_t)adp->ad_oldblkno);
4266 				}
4267 				dp2->di_db[adp->ad_lbn] = adp->ad_newblkno;
4268 			} else {
4269 				if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0) {
4270 					lk.lkt_held = NOHOLDER;
4271 					panic("%s: %s #%jd allocated as %jd",
4272 					    "handle_written_inodeblock",
4273 					    "indirect pointer",
4274 					    (intmax_t)adp->ad_lbn - NDADDR,
4275 					    (intmax_t)
4276 					    dp2->di_ib[adp->ad_lbn - NDADDR]);
4277 				}
4278 				dp2->di_ib[adp->ad_lbn - NDADDR] =
4279 				    adp->ad_newblkno;
4280 			}
4281 		}
4282 		adp->ad_state &= ~UNDONE;
4283 		adp->ad_state |= ATTACHED;
4284 		hadchanges = 1;
4285 	}
4286 	for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
4287 		nextadp = TAILQ_NEXT(adp, ad_next);
4288 		if (adp->ad_state & ATTACHED) {
4289 			lk.lkt_held = NOHOLDER;
4290 			panic("handle_written_inodeblock: new entry");
4291 		}
4292 		if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno) {
4293 			lk.lkt_held = NOHOLDER;
4294 			panic("%s: direct pointers #%jd %s %jd != %jd",
4295 			    "handle_written_inodeblock",
4296 			    (intmax_t)adp->ad_lbn, "mismatch",
4297 			    (intmax_t)dp2->di_extb[adp->ad_lbn],
4298 			    (intmax_t)adp->ad_oldblkno);
4299 		}
4300 		dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno;
4301 		adp->ad_state &= ~UNDONE;
4302 		adp->ad_state |= ATTACHED;
4303 		hadchanges = 1;
4304 	}
4305 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
4306 		stat_direct_blk_ptrs++;
4307 	/*
4308 	 * Reset the file size to its most up-to-date value.
4309 	 */
4310 	if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) {
4311 		lk.lkt_held = NOHOLDER;
4312 		panic("handle_written_inodeblock: bad size");
4313 	}
4314 	if (fstype == UFS1) {
4315 		if (dp1->di_size != inodedep->id_savedsize) {
4316 			dp1->di_size = inodedep->id_savedsize;
4317 			hadchanges = 1;
4318 		}
4319 	} else {
4320 		if (dp2->di_size != inodedep->id_savedsize) {
4321 			dp2->di_size = inodedep->id_savedsize;
4322 			hadchanges = 1;
4323 		}
4324 		if (dp2->di_extsize != inodedep->id_savedextsize) {
4325 			dp2->di_extsize = inodedep->id_savedextsize;
4326 			hadchanges = 1;
4327 		}
4328 	}
4329 	inodedep->id_savedsize = -1;
4330 	inodedep->id_savedextsize = -1;
4331 	/*
4332 	 * If there were any rollbacks in the inode block, then it must be
4333 	 * marked dirty so that its will eventually get written back in
4334 	 * its correct form.
4335 	 */
4336 	if (hadchanges)
4337 		bdirty(bp);
4338 	/*
4339 	 * Process any allocdirects that completed during the update.
4340 	 */
4341 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
4342 		handle_allocdirect_partdone(adp);
4343 	if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
4344 		handle_allocdirect_partdone(adp);
4345 	/*
4346 	 * Process deallocations that were held pending until the
4347 	 * inode had been written to disk. Freeing of the inode
4348 	 * is delayed until after all blocks have been freed to
4349 	 * avoid creation of new <vfsid, inum, lbn> triples
4350 	 * before the old ones have been deleted.
4351 	 */
4352 	filefree = NULL;
4353 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
4354 		WORKLIST_REMOVE(wk);
4355 		switch (wk->wk_type) {
4356 
4357 		case D_FREEFILE:
4358 			/*
4359 			 * We defer adding filefree to the worklist until
4360 			 * all other additions have been made to ensure
4361 			 * that it will be done after all the old blocks
4362 			 * have been freed.
4363 			 */
4364 			if (filefree != NULL) {
4365 				lk.lkt_held = NOHOLDER;
4366 				panic("handle_written_inodeblock: filefree");
4367 			}
4368 			filefree = wk;
4369 			continue;
4370 
4371 		case D_MKDIR:
4372 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
4373 			continue;
4374 
4375 		case D_DIRADD:
4376 			diradd_inode_written(WK_DIRADD(wk), inodedep);
4377 			continue;
4378 
4379 		case D_FREEBLKS:
4380 		case D_FREEFRAG:
4381 		case D_DIRREM:
4382 			add_to_worklist(wk);
4383 			continue;
4384 
4385 		case D_NEWDIRBLK:
4386 			free_newdirblk(WK_NEWDIRBLK(wk));
4387 			continue;
4388 
4389 		default:
4390 			lk.lkt_held = NOHOLDER;
4391 			panic("handle_written_inodeblock: Unknown type %s",
4392 			    TYPENAME(wk->wk_type));
4393 			/* NOTREACHED */
4394 		}
4395 	}
4396 	if (filefree != NULL) {
4397 		if (free_inodedep(inodedep) == 0) {
4398 			lk.lkt_held = NOHOLDER;
4399 			panic("handle_written_inodeblock: live inodedep");
4400 		}
4401 		add_to_worklist(filefree);
4402 		return (0);
4403 	}
4404 
4405 	/*
4406 	 * If no outstanding dependencies, free it.
4407 	 */
4408 	if (free_inodedep(inodedep) ||
4409 	    (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
4410 	     TAILQ_FIRST(&inodedep->id_extupdt) == 0))
4411 		return (0);
4412 	return (hadchanges);
4413 }
4414 
4415 /*
4416  * Process a diradd entry after its dependent inode has been written.
4417  * This routine must be called with splbio interrupts blocked.
4418  */
4419 static void
4420 diradd_inode_written(dap, inodedep)
4421 	struct diradd *dap;
4422 	struct inodedep *inodedep;
4423 {
4424 	struct pagedep *pagedep;
4425 
4426 	dap->da_state |= COMPLETE;
4427 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4428 		if (dap->da_state & DIRCHG)
4429 			pagedep = dap->da_previous->dm_pagedep;
4430 		else
4431 			pagedep = dap->da_pagedep;
4432 		LIST_REMOVE(dap, da_pdlist);
4433 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4434 	}
4435 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
4436 }
4437 
4438 /*
4439  * Handle the completion of a mkdir dependency.
4440  */
4441 static void
4442 handle_written_mkdir(mkdir, type)
4443 	struct mkdir *mkdir;
4444 	int type;
4445 {
4446 	struct diradd *dap;
4447 	struct pagedep *pagedep;
4448 
4449 	if (mkdir->md_state != type) {
4450 		lk.lkt_held = NOHOLDER;
4451 		panic("handle_written_mkdir: bad type");
4452 	}
4453 	dap = mkdir->md_diradd;
4454 	dap->da_state &= ~type;
4455 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
4456 		dap->da_state |= DEPCOMPLETE;
4457 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4458 		if (dap->da_state & DIRCHG)
4459 			pagedep = dap->da_previous->dm_pagedep;
4460 		else
4461 			pagedep = dap->da_pagedep;
4462 		LIST_REMOVE(dap, da_pdlist);
4463 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4464 	}
4465 	LIST_REMOVE(mkdir, md_mkdirs);
4466 	WORKITEM_FREE(mkdir, D_MKDIR);
4467 }
4468 
4469 /*
4470  * Called from within softdep_disk_write_complete above.
4471  * A write operation was just completed. Removed inodes can
4472  * now be freed and associated block pointers may be committed.
4473  * Note that this routine is always called from interrupt level
4474  * with further splbio interrupts blocked.
4475  */
4476 static int
4477 handle_written_filepage(pagedep, bp)
4478 	struct pagedep *pagedep;
4479 	struct buf *bp;		/* buffer containing the written page */
4480 {
4481 	struct dirrem *dirrem;
4482 	struct diradd *dap, *nextdap;
4483 	struct direct *ep;
4484 	int i, chgs;
4485 
4486 	if ((pagedep->pd_state & IOSTARTED) == 0) {
4487 		lk.lkt_held = NOHOLDER;
4488 		panic("handle_written_filepage: not started");
4489 	}
4490 	pagedep->pd_state &= ~IOSTARTED;
4491 	/*
4492 	 * Process any directory removals that have been committed.
4493 	 */
4494 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
4495 		LIST_REMOVE(dirrem, dm_next);
4496 		dirrem->dm_dirinum = pagedep->pd_ino;
4497 		add_to_worklist(&dirrem->dm_list);
4498 	}
4499 	/*
4500 	 * Free any directory additions that have been committed.
4501 	 * If it is a newly allocated block, we have to wait until
4502 	 * the on-disk directory inode claims the new block.
4503 	 */
4504 	if ((pagedep->pd_state & NEWBLOCK) == 0)
4505 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
4506 			free_diradd(dap);
4507 	/*
4508 	 * Uncommitted directory entries must be restored.
4509 	 */
4510 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
4511 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
4512 		     dap = nextdap) {
4513 			nextdap = LIST_NEXT(dap, da_pdlist);
4514 			if (dap->da_state & ATTACHED) {
4515 				lk.lkt_held = NOHOLDER;
4516 				panic("handle_written_filepage: attached");
4517 			}
4518 			ep = (struct direct *)
4519 			    ((char *)bp->b_data + dap->da_offset);
4520 			ep->d_ino = dap->da_newinum;
4521 			dap->da_state &= ~UNDONE;
4522 			dap->da_state |= ATTACHED;
4523 			chgs = 1;
4524 			/*
4525 			 * If the inode referenced by the directory has
4526 			 * been written out, then the dependency can be
4527 			 * moved to the pending list.
4528 			 */
4529 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4530 				LIST_REMOVE(dap, da_pdlist);
4531 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
4532 				    da_pdlist);
4533 			}
4534 		}
4535 	}
4536 	/*
4537 	 * If there were any rollbacks in the directory, then it must be
4538 	 * marked dirty so that its will eventually get written back in
4539 	 * its correct form.
4540 	 */
4541 	if (chgs) {
4542 		if ((bp->b_flags & B_DELWRI) == 0)
4543 			stat_dir_entry++;
4544 		bdirty(bp);
4545 		return (1);
4546 	}
4547 	/*
4548 	 * If we are not waiting for a new directory block to be
4549 	 * claimed by its inode, then the pagedep will be freed.
4550 	 * Otherwise it will remain to track any new entries on
4551 	 * the page in case they are fsync'ed.
4552 	 */
4553 	if ((pagedep->pd_state & NEWBLOCK) == 0) {
4554 		LIST_REMOVE(pagedep, pd_hash);
4555 		WORKITEM_FREE(pagedep, D_PAGEDEP);
4556 	}
4557 	return (0);
4558 }
4559 
4560 /*
4561  * Writing back in-core inode structures.
4562  *
4563  * The filesystem only accesses an inode's contents when it occupies an
4564  * "in-core" inode structure.  These "in-core" structures are separate from
4565  * the page frames used to cache inode blocks.  Only the latter are
4566  * transferred to/from the disk.  So, when the updated contents of the
4567  * "in-core" inode structure are copied to the corresponding in-memory inode
4568  * block, the dependencies are also transferred.  The following procedure is
4569  * called when copying a dirty "in-core" inode to a cached inode block.
4570  */
4571 
4572 /*
4573  * Called when an inode is loaded from disk. If the effective link count
4574  * differed from the actual link count when it was last flushed, then we
4575  * need to ensure that the correct effective link count is put back.
4576  */
4577 void
4578 softdep_load_inodeblock(ip)
4579 	struct inode *ip;	/* the "in_core" copy of the inode */
4580 {
4581 	struct inodedep *inodedep;
4582 
4583 	/*
4584 	 * Check for alternate nlink count.
4585 	 */
4586 	ip->i_effnlink = ip->i_nlink;
4587 	ACQUIRE_LOCK(&lk);
4588 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
4589 		FREE_LOCK(&lk);
4590 		return;
4591 	}
4592 	ip->i_effnlink -= inodedep->id_nlinkdelta;
4593 	if (inodedep->id_state & SPACECOUNTED)
4594 		ip->i_flag |= IN_SPACECOUNTED;
4595 	FREE_LOCK(&lk);
4596 }
4597 
4598 /*
4599  * This routine is called just before the "in-core" inode
4600  * information is to be copied to the in-memory inode block.
4601  * Recall that an inode block contains several inodes. If
4602  * the force flag is set, then the dependencies will be
4603  * cleared so that the update can always be made. Note that
4604  * the buffer is locked when this routine is called, so we
4605  * will never be in the middle of writing the inode block
4606  * to disk.
4607  */
4608 void
4609 softdep_update_inodeblock(ip, bp, waitfor)
4610 	struct inode *ip;	/* the "in_core" copy of the inode */
4611 	struct buf *bp;		/* the buffer containing the inode block */
4612 	int waitfor;		/* nonzero => update must be allowed */
4613 {
4614 	struct inodedep *inodedep;
4615 	struct worklist *wk;
4616 	int error, gotit;
4617 
4618 	/*
4619 	 * If the effective link count is not equal to the actual link
4620 	 * count, then we must track the difference in an inodedep while
4621 	 * the inode is (potentially) tossed out of the cache. Otherwise,
4622 	 * if there is no existing inodedep, then there are no dependencies
4623 	 * to track.
4624 	 */
4625 	ACQUIRE_LOCK(&lk);
4626 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
4627 		FREE_LOCK(&lk);
4628 		if (ip->i_effnlink != ip->i_nlink)
4629 			panic("softdep_update_inodeblock: bad link count");
4630 		return;
4631 	}
4632 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
4633 		FREE_LOCK(&lk);
4634 		panic("softdep_update_inodeblock: bad delta");
4635 	}
4636 	/*
4637 	 * Changes have been initiated. Anything depending on these
4638 	 * changes cannot occur until this inode has been written.
4639 	 */
4640 	inodedep->id_state &= ~COMPLETE;
4641 	if ((inodedep->id_state & ONWORKLIST) == 0)
4642 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
4643 	/*
4644 	 * Any new dependencies associated with the incore inode must
4645 	 * now be moved to the list associated with the buffer holding
4646 	 * the in-memory copy of the inode. Once merged process any
4647 	 * allocdirects that are completed by the merger.
4648 	 */
4649 	merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
4650 	if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
4651 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
4652 	merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
4653 	if (TAILQ_FIRST(&inodedep->id_extupdt) != NULL)
4654 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt));
4655 	/*
4656 	 * Now that the inode has been pushed into the buffer, the
4657 	 * operations dependent on the inode being written to disk
4658 	 * can be moved to the id_bufwait so that they will be
4659 	 * processed when the buffer I/O completes.
4660 	 */
4661 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
4662 		WORKLIST_REMOVE(wk);
4663 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
4664 	}
4665 	/*
4666 	 * Newly allocated inodes cannot be written until the bitmap
4667 	 * that allocates them have been written (indicated by
4668 	 * DEPCOMPLETE being set in id_state). If we are doing a
4669 	 * forced sync (e.g., an fsync on a file), we force the bitmap
4670 	 * to be written so that the update can be done.
4671 	 */
4672 	if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
4673 		FREE_LOCK(&lk);
4674 		return;
4675 	}
4676 	gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4677 	FREE_LOCK(&lk);
4678 	if (gotit &&
4679 	    (error = BUF_WRITE(inodedep->id_buf)) != 0)
4680 		softdep_error("softdep_update_inodeblock: bwrite", error);
4681 	if ((inodedep->id_state & DEPCOMPLETE) == 0)
4682 		panic("softdep_update_inodeblock: update failed");
4683 }
4684 
4685 /*
4686  * Merge the a new inode dependency list (such as id_newinoupdt) into an
4687  * old inode dependency list (such as id_inoupdt). This routine must be
4688  * called with splbio interrupts blocked.
4689  */
4690 static void
4691 merge_inode_lists(newlisthead, oldlisthead)
4692 	struct allocdirectlst *newlisthead;
4693 	struct allocdirectlst *oldlisthead;
4694 {
4695 	struct allocdirect *listadp, *newadp;
4696 
4697 	newadp = TAILQ_FIRST(newlisthead);
4698 	for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
4699 		if (listadp->ad_lbn < newadp->ad_lbn) {
4700 			listadp = TAILQ_NEXT(listadp, ad_next);
4701 			continue;
4702 		}
4703 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4704 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
4705 		if (listadp->ad_lbn == newadp->ad_lbn) {
4706 			allocdirect_merge(oldlisthead, newadp,
4707 			    listadp);
4708 			listadp = newadp;
4709 		}
4710 		newadp = TAILQ_FIRST(newlisthead);
4711 	}
4712 	while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
4713 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4714 		TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
4715 	}
4716 }
4717 
4718 /*
4719  * If we are doing an fsync, then we must ensure that any directory
4720  * entries for the inode have been written after the inode gets to disk.
4721  */
4722 int
4723 softdep_fsync(vp)
4724 	struct vnode *vp;	/* the "in_core" copy of the inode */
4725 {
4726 	struct inodedep *inodedep;
4727 	struct pagedep *pagedep;
4728 	struct worklist *wk;
4729 	struct diradd *dap;
4730 	struct mount *mnt;
4731 	struct vnode *pvp;
4732 	struct inode *ip;
4733 	struct buf *bp;
4734 	struct fs *fs;
4735 	struct thread *td = curthread;
4736 	int error, flushparent;
4737 	ino_t parentino;
4738 	ufs_lbn_t lbn;
4739 
4740 	ip = VTOI(vp);
4741 	fs = ip->i_fs;
4742 	ACQUIRE_LOCK(&lk);
4743 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
4744 		FREE_LOCK(&lk);
4745 		return (0);
4746 	}
4747 	if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
4748 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
4749 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
4750 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
4751 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
4752 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) {
4753 		FREE_LOCK(&lk);
4754 		panic("softdep_fsync: pending ops");
4755 	}
4756 	for (error = 0, flushparent = 0; ; ) {
4757 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
4758 			break;
4759 		if (wk->wk_type != D_DIRADD) {
4760 			FREE_LOCK(&lk);
4761 			panic("softdep_fsync: Unexpected type %s",
4762 			    TYPENAME(wk->wk_type));
4763 		}
4764 		dap = WK_DIRADD(wk);
4765 		/*
4766 		 * Flush our parent if this directory entry has a MKDIR_PARENT
4767 		 * dependency or is contained in a newly allocated block.
4768 		 */
4769 		if (dap->da_state & DIRCHG)
4770 			pagedep = dap->da_previous->dm_pagedep;
4771 		else
4772 			pagedep = dap->da_pagedep;
4773 		mnt = pagedep->pd_mnt;
4774 		parentino = pagedep->pd_ino;
4775 		lbn = pagedep->pd_lbn;
4776 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) {
4777 			FREE_LOCK(&lk);
4778 			panic("softdep_fsync: dirty");
4779 		}
4780 		if ((dap->da_state & MKDIR_PARENT) ||
4781 		    (pagedep->pd_state & NEWBLOCK))
4782 			flushparent = 1;
4783 		else
4784 			flushparent = 0;
4785 		/*
4786 		 * If we are being fsync'ed as part of vgone'ing this vnode,
4787 		 * then we will not be able to release and recover the
4788 		 * vnode below, so we just have to give up on writing its
4789 		 * directory entry out. It will eventually be written, just
4790 		 * not now, but then the user was not asking to have it
4791 		 * written, so we are not breaking any promises.
4792 		 */
4793 		mp_fixme("This operation is not atomic wrt the rest of the code");
4794 		VI_LOCK(vp);
4795 		if (vp->v_iflag & VI_XLOCK) {
4796 			VI_UNLOCK(vp);
4797 			break;
4798 		} else
4799 			VI_UNLOCK(vp);
4800 		/*
4801 		 * We prevent deadlock by always fetching inodes from the
4802 		 * root, moving down the directory tree. Thus, when fetching
4803 		 * our parent directory, we first try to get the lock. If
4804 		 * that fails, we must unlock ourselves before requesting
4805 		 * the lock on our parent. See the comment in ufs_lookup
4806 		 * for details on possible races.
4807 		 */
4808 		FREE_LOCK(&lk);
4809 		if (VFS_VGET(mnt, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) {
4810 			VOP_UNLOCK(vp, 0, td);
4811 			error = VFS_VGET(mnt, parentino, LK_EXCLUSIVE, &pvp);
4812 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
4813 			if (error != 0)
4814 				return (error);
4815 		}
4816 		/*
4817 		 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
4818 		 * that are contained in direct blocks will be resolved by
4819 		 * doing a UFS_UPDATE. Pagedeps contained in indirect blocks
4820 		 * may require a complete sync'ing of the directory. So, we
4821 		 * try the cheap and fast UFS_UPDATE first, and if that fails,
4822 		 * then we do the slower VOP_FSYNC of the directory.
4823 		 */
4824 		if (flushparent) {
4825 			if ((error = UFS_UPDATE(pvp, 1)) != 0) {
4826 				vput(pvp);
4827 				return (error);
4828 			}
4829 			if ((pagedep->pd_state & NEWBLOCK) &&
4830 			    (error = VOP_FSYNC(pvp, td->td_ucred, MNT_WAIT, td))) {
4831 				vput(pvp);
4832 				return (error);
4833 			}
4834 		}
4835 		/*
4836 		 * Flush directory page containing the inode's name.
4837 		 */
4838 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
4839 		    &bp);
4840 		if (error == 0)
4841 			error = BUF_WRITE(bp);
4842 		else
4843 			brelse(bp);
4844 		vput(pvp);
4845 		if (error != 0)
4846 			return (error);
4847 		ACQUIRE_LOCK(&lk);
4848 		if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
4849 			break;
4850 	}
4851 	FREE_LOCK(&lk);
4852 	return (0);
4853 }
4854 
4855 /*
4856  * Flush all the dirty bitmaps associated with the block device
4857  * before flushing the rest of the dirty blocks so as to reduce
4858  * the number of dependencies that will have to be rolled back.
4859  */
4860 void
4861 softdep_fsync_mountdev(vp)
4862 	struct vnode *vp;
4863 {
4864 	struct buf *bp, *nbp;
4865 	struct worklist *wk;
4866 
4867 	if (!vn_isdisk(vp, NULL))
4868 		panic("softdep_fsync_mountdev: vnode not a disk");
4869 	ACQUIRE_LOCK(&lk);
4870 	VI_LOCK(vp);
4871 	for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
4872 		nbp = TAILQ_NEXT(bp, b_vnbufs);
4873 		VI_UNLOCK(vp);
4874 		/*
4875 		 * If it is already scheduled, skip to the next buffer.
4876 		 */
4877 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
4878 			VI_LOCK(vp);
4879 			continue;
4880 		}
4881 		if ((bp->b_flags & B_DELWRI) == 0) {
4882 			FREE_LOCK(&lk);
4883 			panic("softdep_fsync_mountdev: not dirty");
4884 		}
4885 		/*
4886 		 * We are only interested in bitmaps with outstanding
4887 		 * dependencies.
4888 		 */
4889 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
4890 		    wk->wk_type != D_BMSAFEMAP ||
4891 		    (bp->b_xflags & BX_BKGRDINPROG)) {
4892 			BUF_UNLOCK(bp);
4893 			VI_LOCK(vp);
4894 			continue;
4895 		}
4896 		bremfree(bp);
4897 		FREE_LOCK(&lk);
4898 		(void) bawrite(bp);
4899 		ACQUIRE_LOCK(&lk);
4900 		/*
4901 		 * Since we may have slept during the I/O, we need
4902 		 * to start from a known point.
4903 		 */
4904 		VI_LOCK(vp);
4905 		nbp = TAILQ_FIRST(&vp->v_dirtyblkhd);
4906 	}
4907 	VI_UNLOCK(vp);
4908 	drain_output(vp, 1);
4909 	FREE_LOCK(&lk);
4910 }
4911 
4912 /*
4913  * This routine is called when we are trying to synchronously flush a
4914  * file. This routine must eliminate any filesystem metadata dependencies
4915  * so that the syncing routine can succeed by pushing the dirty blocks
4916  * associated with the file. If any I/O errors occur, they are returned.
4917  */
4918 int
4919 softdep_sync_metadata(ap)
4920 	struct vop_fsync_args /* {
4921 		struct vnode *a_vp;
4922 		struct ucred *a_cred;
4923 		int a_waitfor;
4924 		struct thread *a_td;
4925 	} */ *ap;
4926 {
4927 	struct vnode *vp = ap->a_vp;
4928 	struct pagedep *pagedep;
4929 	struct allocdirect *adp;
4930 	struct allocindir *aip;
4931 	struct buf *bp, *nbp;
4932 	struct worklist *wk;
4933 	int i, error, waitfor;
4934 
4935 	/*
4936 	 * Check whether this vnode is involved in a filesystem
4937 	 * that is doing soft dependency processing.
4938 	 */
4939 	if (!vn_isdisk(vp, NULL)) {
4940 		if (!DOINGSOFTDEP(vp))
4941 			return (0);
4942 	} else
4943 		if (vp->v_rdev->si_mountpoint == NULL ||
4944 		    (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0)
4945 			return (0);
4946 	/*
4947 	 * Ensure that any direct block dependencies have been cleared.
4948 	 */
4949 	ACQUIRE_LOCK(&lk);
4950 	if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
4951 		FREE_LOCK(&lk);
4952 		return (error);
4953 	}
4954 	/*
4955 	 * For most files, the only metadata dependencies are the
4956 	 * cylinder group maps that allocate their inode or blocks.
4957 	 * The block allocation dependencies can be found by traversing
4958 	 * the dependency lists for any buffers that remain on their
4959 	 * dirty buffer list. The inode allocation dependency will
4960 	 * be resolved when the inode is updated with MNT_WAIT.
4961 	 * This work is done in two passes. The first pass grabs most
4962 	 * of the buffers and begins asynchronously writing them. The
4963 	 * only way to wait for these asynchronous writes is to sleep
4964 	 * on the filesystem vnode which may stay busy for a long time
4965 	 * if the filesystem is active. So, instead, we make a second
4966 	 * pass over the dependencies blocking on each write. In the
4967 	 * usual case we will be blocking against a write that we
4968 	 * initiated, so when it is done the dependency will have been
4969 	 * resolved. Thus the second pass is expected to end quickly.
4970 	 */
4971 	waitfor = MNT_NOWAIT;
4972 top:
4973 	/*
4974 	 * We must wait for any I/O in progress to finish so that
4975 	 * all potential buffers on the dirty list will be visible.
4976 	 */
4977 	drain_output(vp, 1);
4978 	if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) {
4979 		FREE_LOCK(&lk);
4980 		return (0);
4981 	}
4982 	mp_fixme("The locking is somewhat complicated nonexistant here.");
4983 	bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
4984 	/* While syncing snapshots, we must allow recursive lookups */
4985 	bp->b_lock.lk_flags |= LK_CANRECURSE;
4986 loop:
4987 	/*
4988 	 * As we hold the buffer locked, none of its dependencies
4989 	 * will disappear.
4990 	 */
4991 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4992 		switch (wk->wk_type) {
4993 
4994 		case D_ALLOCDIRECT:
4995 			adp = WK_ALLOCDIRECT(wk);
4996 			if (adp->ad_state & DEPCOMPLETE)
4997 				continue;
4998 			nbp = adp->ad_buf;
4999 			if (getdirtybuf(&nbp, waitfor) == 0)
5000 				continue;
5001 			FREE_LOCK(&lk);
5002 			if (waitfor == MNT_NOWAIT) {
5003 				bawrite(nbp);
5004 			} else if ((error = BUF_WRITE(nbp)) != 0) {
5005 				break;
5006 			}
5007 			ACQUIRE_LOCK(&lk);
5008 			continue;
5009 
5010 		case D_ALLOCINDIR:
5011 			aip = WK_ALLOCINDIR(wk);
5012 			if (aip->ai_state & DEPCOMPLETE)
5013 				continue;
5014 			nbp = aip->ai_buf;
5015 			if (getdirtybuf(&nbp, waitfor) == 0)
5016 				continue;
5017 			FREE_LOCK(&lk);
5018 			if (waitfor == MNT_NOWAIT) {
5019 				bawrite(nbp);
5020 			} else if ((error = BUF_WRITE(nbp)) != 0) {
5021 				break;
5022 			}
5023 			ACQUIRE_LOCK(&lk);
5024 			continue;
5025 
5026 		case D_INDIRDEP:
5027 		restart:
5028 
5029 			LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
5030 				if (aip->ai_state & DEPCOMPLETE)
5031 					continue;
5032 				nbp = aip->ai_buf;
5033 				if (getdirtybuf(&nbp, MNT_WAIT) == 0)
5034 					goto restart;
5035 				FREE_LOCK(&lk);
5036 				if ((error = BUF_WRITE(nbp)) != 0) {
5037 					break;
5038 				}
5039 				ACQUIRE_LOCK(&lk);
5040 				goto restart;
5041 			}
5042 			continue;
5043 
5044 		case D_INODEDEP:
5045 			if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
5046 			    WK_INODEDEP(wk)->id_ino)) != 0) {
5047 				FREE_LOCK(&lk);
5048 				break;
5049 			}
5050 			continue;
5051 
5052 		case D_PAGEDEP:
5053 			/*
5054 			 * We are trying to sync a directory that may
5055 			 * have dependencies on both its own metadata
5056 			 * and/or dependencies on the inodes of any
5057 			 * recently allocated files. We walk its diradd
5058 			 * lists pushing out the associated inode.
5059 			 */
5060 			pagedep = WK_PAGEDEP(wk);
5061 			for (i = 0; i < DAHASHSZ; i++) {
5062 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
5063 					continue;
5064 				if ((error =
5065 				    flush_pagedep_deps(vp, pagedep->pd_mnt,
5066 						&pagedep->pd_diraddhd[i]))) {
5067 					FREE_LOCK(&lk);
5068 					break;
5069 				}
5070 			}
5071 			continue;
5072 
5073 		case D_MKDIR:
5074 			/*
5075 			 * This case should never happen if the vnode has
5076 			 * been properly sync'ed. However, if this function
5077 			 * is used at a place where the vnode has not yet
5078 			 * been sync'ed, this dependency can show up. So,
5079 			 * rather than panic, just flush it.
5080 			 */
5081 			nbp = WK_MKDIR(wk)->md_buf;
5082 			if (getdirtybuf(&nbp, waitfor) == 0)
5083 				continue;
5084 			FREE_LOCK(&lk);
5085 			if (waitfor == MNT_NOWAIT) {
5086 				bawrite(nbp);
5087 			} else if ((error = BUF_WRITE(nbp)) != 0) {
5088 				break;
5089 			}
5090 			ACQUIRE_LOCK(&lk);
5091 			continue;
5092 
5093 		case D_BMSAFEMAP:
5094 			/*
5095 			 * This case should never happen if the vnode has
5096 			 * been properly sync'ed. However, if this function
5097 			 * is used at a place where the vnode has not yet
5098 			 * been sync'ed, this dependency can show up. So,
5099 			 * rather than panic, just flush it.
5100 			 */
5101 			nbp = WK_BMSAFEMAP(wk)->sm_buf;
5102 			if (getdirtybuf(&nbp, waitfor) == 0)
5103 				continue;
5104 			FREE_LOCK(&lk);
5105 			if (waitfor == MNT_NOWAIT) {
5106 				bawrite(nbp);
5107 			} else if ((error = BUF_WRITE(nbp)) != 0) {
5108 				break;
5109 			}
5110 			ACQUIRE_LOCK(&lk);
5111 			continue;
5112 
5113 		default:
5114 			FREE_LOCK(&lk);
5115 			panic("softdep_sync_metadata: Unknown type %s",
5116 			    TYPENAME(wk->wk_type));
5117 			/* NOTREACHED */
5118 		}
5119 		/* We reach here only in error and unlocked */
5120 		if (error == 0)
5121 			panic("softdep_sync_metadata: zero error");
5122 		bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5123 		bawrite(bp);
5124 		return (error);
5125 	}
5126 	(void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT);
5127 	nbp = TAILQ_NEXT(bp, b_vnbufs);
5128 	FREE_LOCK(&lk);
5129 	bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5130 	bawrite(bp);
5131 	ACQUIRE_LOCK(&lk);
5132 	if (nbp != NULL) {
5133 		bp = nbp;
5134 		goto loop;
5135 	}
5136 	/*
5137 	 * The brief unlock is to allow any pent up dependency
5138 	 * processing to be done. Then proceed with the second pass.
5139 	 */
5140 	if (waitfor == MNT_NOWAIT) {
5141 		waitfor = MNT_WAIT;
5142 		FREE_LOCK(&lk);
5143 		ACQUIRE_LOCK(&lk);
5144 		goto top;
5145 	}
5146 
5147 	/*
5148 	 * If we have managed to get rid of all the dirty buffers,
5149 	 * then we are done. For certain directories and block
5150 	 * devices, we may need to do further work.
5151 	 *
5152 	 * We must wait for any I/O in progress to finish so that
5153 	 * all potential buffers on the dirty list will be visible.
5154 	 */
5155 	drain_output(vp, 1);
5156 	if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) {
5157 		FREE_LOCK(&lk);
5158 		return (0);
5159 	}
5160 
5161 	FREE_LOCK(&lk);
5162 	/*
5163 	 * If we are trying to sync a block device, some of its buffers may
5164 	 * contain metadata that cannot be written until the contents of some
5165 	 * partially written files have been written to disk. The only easy
5166 	 * way to accomplish this is to sync the entire filesystem (luckily
5167 	 * this happens rarely).
5168 	 */
5169 	if (vn_isdisk(vp, NULL) &&
5170 	    vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) &&
5171 	    (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, ap->a_cred,
5172 	     ap->a_td)) != 0)
5173 		return (error);
5174 	return (0);
5175 }
5176 
5177 /*
5178  * Flush the dependencies associated with an inodedep.
5179  * Called with splbio blocked.
5180  */
5181 static int
5182 flush_inodedep_deps(fs, ino)
5183 	struct fs *fs;
5184 	ino_t ino;
5185 {
5186 	struct inodedep *inodedep;
5187 	int error, waitfor;
5188 
5189 	/*
5190 	 * This work is done in two passes. The first pass grabs most
5191 	 * of the buffers and begins asynchronously writing them. The
5192 	 * only way to wait for these asynchronous writes is to sleep
5193 	 * on the filesystem vnode which may stay busy for a long time
5194 	 * if the filesystem is active. So, instead, we make a second
5195 	 * pass over the dependencies blocking on each write. In the
5196 	 * usual case we will be blocking against a write that we
5197 	 * initiated, so when it is done the dependency will have been
5198 	 * resolved. Thus the second pass is expected to end quickly.
5199 	 * We give a brief window at the top of the loop to allow
5200 	 * any pending I/O to complete.
5201 	 */
5202 	for (error = 0, waitfor = MNT_NOWAIT; ; ) {
5203 		if (error)
5204 			return (error);
5205 		FREE_LOCK(&lk);
5206 		ACQUIRE_LOCK(&lk);
5207 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
5208 			return (0);
5209 		if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
5210 		    flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
5211 		    flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
5212 		    flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
5213 			continue;
5214 		/*
5215 		 * If pass2, we are done, otherwise do pass 2.
5216 		 */
5217 		if (waitfor == MNT_WAIT)
5218 			break;
5219 		waitfor = MNT_WAIT;
5220 	}
5221 	/*
5222 	 * Try freeing inodedep in case all dependencies have been removed.
5223 	 */
5224 	if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
5225 		(void) free_inodedep(inodedep);
5226 	return (0);
5227 }
5228 
5229 /*
5230  * Flush an inode dependency list.
5231  * Called with splbio blocked.
5232  */
5233 static int
5234 flush_deplist(listhead, waitfor, errorp)
5235 	struct allocdirectlst *listhead;
5236 	int waitfor;
5237 	int *errorp;
5238 {
5239 	struct allocdirect *adp;
5240 	struct buf *bp;
5241 
5242 	TAILQ_FOREACH(adp, listhead, ad_next) {
5243 		if (adp->ad_state & DEPCOMPLETE)
5244 			continue;
5245 		bp = adp->ad_buf;
5246 		if (getdirtybuf(&bp, waitfor) == 0) {
5247 			if (waitfor == MNT_NOWAIT)
5248 				continue;
5249 			return (1);
5250 		}
5251 		FREE_LOCK(&lk);
5252 		if (waitfor == MNT_NOWAIT) {
5253 			bawrite(bp);
5254 		} else if ((*errorp = BUF_WRITE(bp)) != 0) {
5255 			ACQUIRE_LOCK(&lk);
5256 			return (1);
5257 		}
5258 		ACQUIRE_LOCK(&lk);
5259 		return (1);
5260 	}
5261 	return (0);
5262 }
5263 
5264 /*
5265  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
5266  * Called with splbio blocked.
5267  */
5268 static int
5269 flush_pagedep_deps(pvp, mp, diraddhdp)
5270 	struct vnode *pvp;
5271 	struct mount *mp;
5272 	struct diraddhd *diraddhdp;
5273 {
5274 	struct thread *td = curthread;
5275 	struct inodedep *inodedep;
5276 	struct ufsmount *ump;
5277 	struct diradd *dap;
5278 	struct vnode *vp;
5279 	int gotit, error = 0;
5280 	struct buf *bp;
5281 	ino_t inum;
5282 
5283 	ump = VFSTOUFS(mp);
5284 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
5285 		/*
5286 		 * Flush ourselves if this directory entry
5287 		 * has a MKDIR_PARENT dependency.
5288 		 */
5289 		if (dap->da_state & MKDIR_PARENT) {
5290 			FREE_LOCK(&lk);
5291 			if ((error = UFS_UPDATE(pvp, 1)) != 0)
5292 				break;
5293 			ACQUIRE_LOCK(&lk);
5294 			/*
5295 			 * If that cleared dependencies, go on to next.
5296 			 */
5297 			if (dap != LIST_FIRST(diraddhdp))
5298 				continue;
5299 			if (dap->da_state & MKDIR_PARENT) {
5300 				FREE_LOCK(&lk);
5301 				panic("flush_pagedep_deps: MKDIR_PARENT");
5302 			}
5303 		}
5304 		/*
5305 		 * A newly allocated directory must have its "." and
5306 		 * ".." entries written out before its name can be
5307 		 * committed in its parent. We do not want or need
5308 		 * the full semantics of a synchronous VOP_FSYNC as
5309 		 * that may end up here again, once for each directory
5310 		 * level in the filesystem. Instead, we push the blocks
5311 		 * and wait for them to clear. We have to fsync twice
5312 		 * because the first call may choose to defer blocks
5313 		 * that still have dependencies, but deferral will
5314 		 * happen at most once.
5315 		 */
5316 		inum = dap->da_newinum;
5317 		if (dap->da_state & MKDIR_BODY) {
5318 			FREE_LOCK(&lk);
5319 			if ((error = VFS_VGET(mp, inum, LK_EXCLUSIVE, &vp)))
5320 				break;
5321 			if ((error=VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)) ||
5322 			    (error=VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td))) {
5323 				vput(vp);
5324 				break;
5325 			}
5326 			drain_output(vp, 0);
5327 			vput(vp);
5328 			ACQUIRE_LOCK(&lk);
5329 			/*
5330 			 * If that cleared dependencies, go on to next.
5331 			 */
5332 			if (dap != LIST_FIRST(diraddhdp))
5333 				continue;
5334 			if (dap->da_state & MKDIR_BODY) {
5335 				FREE_LOCK(&lk);
5336 				panic("flush_pagedep_deps: MKDIR_BODY");
5337 			}
5338 		}
5339 		/*
5340 		 * Flush the inode on which the directory entry depends.
5341 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
5342 		 * the only remaining dependency is that the updated inode
5343 		 * count must get pushed to disk. The inode has already
5344 		 * been pushed into its inode buffer (via VOP_UPDATE) at
5345 		 * the time of the reference count change. So we need only
5346 		 * locate that buffer, ensure that there will be no rollback
5347 		 * caused by a bitmap dependency, then write the inode buffer.
5348 		 */
5349 		if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) {
5350 			FREE_LOCK(&lk);
5351 			panic("flush_pagedep_deps: lost inode");
5352 		}
5353 		/*
5354 		 * If the inode still has bitmap dependencies,
5355 		 * push them to disk.
5356 		 */
5357 		if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5358 			gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
5359 			FREE_LOCK(&lk);
5360 			if (gotit &&
5361 			    (error = BUF_WRITE(inodedep->id_buf)) != 0)
5362 				break;
5363 			ACQUIRE_LOCK(&lk);
5364 			if (dap != LIST_FIRST(diraddhdp))
5365 				continue;
5366 		}
5367 		/*
5368 		 * If the inode is still sitting in a buffer waiting
5369 		 * to be written, push it to disk.
5370 		 */
5371 		FREE_LOCK(&lk);
5372 		if ((error = bread(ump->um_devvp,
5373 		    fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
5374 		    (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) {
5375 			brelse(bp);
5376 			break;
5377 		}
5378 		if ((error = BUF_WRITE(bp)) != 0)
5379 			break;
5380 		ACQUIRE_LOCK(&lk);
5381 		/*
5382 		 * If we have failed to get rid of all the dependencies
5383 		 * then something is seriously wrong.
5384 		 */
5385 		if (dap == LIST_FIRST(diraddhdp)) {
5386 			FREE_LOCK(&lk);
5387 			panic("flush_pagedep_deps: flush failed");
5388 		}
5389 	}
5390 	if (error)
5391 		ACQUIRE_LOCK(&lk);
5392 	return (error);
5393 }
5394 
5395 /*
5396  * A large burst of file addition or deletion activity can drive the
5397  * memory load excessively high. First attempt to slow things down
5398  * using the techniques below. If that fails, this routine requests
5399  * the offending operations to fall back to running synchronously
5400  * until the memory load returns to a reasonable level.
5401  */
5402 int
5403 softdep_slowdown(vp)
5404 	struct vnode *vp;
5405 {
5406 	int max_softdeps_hard;
5407 
5408 	max_softdeps_hard = max_softdeps * 11 / 10;
5409 	if (num_dirrem < max_softdeps_hard / 2 &&
5410 	    num_inodedep < max_softdeps_hard)
5411 		return (0);
5412 	stat_sync_limit_hit += 1;
5413 	return (1);
5414 }
5415 
5416 /*
5417  * Called by the allocation routines when they are about to fail
5418  * in the hope that we can free up some disk space.
5419  *
5420  * First check to see if the work list has anything on it. If it has,
5421  * clean up entries until we successfully free some space. Because this
5422  * process holds inodes locked, we cannot handle any remove requests
5423  * that might block on a locked inode as that could lead to deadlock.
5424  * If the worklist yields no free space, encourage the syncer daemon
5425  * to help us. In no event will we try for longer than tickdelay seconds.
5426  */
5427 int
5428 softdep_request_cleanup(fs, vp)
5429 	struct fs *fs;
5430 	struct vnode *vp;
5431 {
5432 	long starttime;
5433 	ufs2_daddr_t needed;
5434 
5435 	needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize;
5436 	starttime = time_second + tickdelay;
5437 	/*
5438 	 * If we are being called because of a process doing a
5439 	 * copy-on-write, then it is not safe to update the vnode
5440 	 * as we may recurse into the copy-on-write routine.
5441 	 */
5442 	if ((curthread->td_proc->p_flag & P_COWINPROGRESS) == 0 &&
5443 	    UFS_UPDATE(vp, 1) != 0)
5444 		return (0);
5445 	while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) {
5446 		if (time_second > starttime)
5447 			return (0);
5448 		if (num_on_worklist > 0 &&
5449 		    process_worklist_item(NULL, LK_NOWAIT) != -1) {
5450 			stat_worklist_push += 1;
5451 			continue;
5452 		}
5453 		request_cleanup(FLUSH_REMOVE_WAIT, 0);
5454 	}
5455 	return (1);
5456 }
5457 
5458 /*
5459  * If memory utilization has gotten too high, deliberately slow things
5460  * down and speed up the I/O processing.
5461  */
5462 static int
5463 request_cleanup(resource, islocked)
5464 	int resource;
5465 	int islocked;
5466 {
5467 	struct thread *td = curthread;
5468 
5469 	/*
5470 	 * We never hold up the filesystem syncer process.
5471 	 */
5472 	if (td == filesys_syncer)
5473 		return (0);
5474 	/*
5475 	 * First check to see if the work list has gotten backlogged.
5476 	 * If it has, co-opt this process to help clean up two entries.
5477 	 * Because this process may hold inodes locked, we cannot
5478 	 * handle any remove requests that might block on a locked
5479 	 * inode as that could lead to deadlock.
5480 	 */
5481 	if (num_on_worklist > max_softdeps / 10) {
5482 		if (islocked)
5483 			FREE_LOCK(&lk);
5484 		process_worklist_item(NULL, LK_NOWAIT);
5485 		process_worklist_item(NULL, LK_NOWAIT);
5486 		stat_worklist_push += 2;
5487 		if (islocked)
5488 			ACQUIRE_LOCK(&lk);
5489 		return(1);
5490 	}
5491 	/*
5492 	 * Next, we attempt to speed up the syncer process. If that
5493 	 * is successful, then we allow the process to continue.
5494 	 */
5495 	if (speedup_syncer() && resource != FLUSH_REMOVE_WAIT)
5496 		return(0);
5497 	/*
5498 	 * If we are resource constrained on inode dependencies, try
5499 	 * flushing some dirty inodes. Otherwise, we are constrained
5500 	 * by file deletions, so try accelerating flushes of directories
5501 	 * with removal dependencies. We would like to do the cleanup
5502 	 * here, but we probably hold an inode locked at this point and
5503 	 * that might deadlock against one that we try to clean. So,
5504 	 * the best that we can do is request the syncer daemon to do
5505 	 * the cleanup for us.
5506 	 */
5507 	switch (resource) {
5508 
5509 	case FLUSH_INODES:
5510 		stat_ino_limit_push += 1;
5511 		req_clear_inodedeps += 1;
5512 		stat_countp = &stat_ino_limit_hit;
5513 		break;
5514 
5515 	case FLUSH_REMOVE:
5516 	case FLUSH_REMOVE_WAIT:
5517 		stat_blk_limit_push += 1;
5518 		req_clear_remove += 1;
5519 		stat_countp = &stat_blk_limit_hit;
5520 		break;
5521 
5522 	default:
5523 		if (islocked)
5524 			FREE_LOCK(&lk);
5525 		panic("request_cleanup: unknown type");
5526 	}
5527 	/*
5528 	 * Hopefully the syncer daemon will catch up and awaken us.
5529 	 * We wait at most tickdelay before proceeding in any case.
5530 	 */
5531 	if (islocked == 0)
5532 		ACQUIRE_LOCK(&lk);
5533 	proc_waiting += 1;
5534 	if (handle.callout == NULL)
5535 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5536 	interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, NULL, PPAUSE,
5537 	    "softupdate", 0);
5538 	proc_waiting -= 1;
5539 	if (islocked == 0)
5540 		FREE_LOCK(&lk);
5541 	return (1);
5542 }
5543 
5544 /*
5545  * Awaken processes pausing in request_cleanup and clear proc_waiting
5546  * to indicate that there is no longer a timer running.
5547  */
5548 static void
5549 pause_timer(arg)
5550 	void *arg;
5551 {
5552 
5553 	*stat_countp += 1;
5554 	wakeup_one(&proc_waiting);
5555 	if (proc_waiting > 0)
5556 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5557 	else
5558 		handle.callout = NULL;
5559 }
5560 
5561 /*
5562  * Flush out a directory with at least one removal dependency in an effort to
5563  * reduce the number of dirrem, freefile, and freeblks dependency structures.
5564  */
5565 static void
5566 clear_remove(td)
5567 	struct thread *td;
5568 {
5569 	struct pagedep_hashhead *pagedephd;
5570 	struct pagedep *pagedep;
5571 	static int next = 0;
5572 	struct mount *mp;
5573 	struct vnode *vp;
5574 	int error, cnt;
5575 	ino_t ino;
5576 
5577 	ACQUIRE_LOCK(&lk);
5578 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
5579 		pagedephd = &pagedep_hashtbl[next++];
5580 		if (next >= pagedep_hash)
5581 			next = 0;
5582 		LIST_FOREACH(pagedep, pagedephd, pd_hash) {
5583 			if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
5584 				continue;
5585 			mp = pagedep->pd_mnt;
5586 			ino = pagedep->pd_ino;
5587 			if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5588 				continue;
5589 			FREE_LOCK(&lk);
5590 			if ((error = VFS_VGET(mp, ino, LK_EXCLUSIVE, &vp))) {
5591 				softdep_error("clear_remove: vget", error);
5592 				vn_finished_write(mp);
5593 				return;
5594 			}
5595 			if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)))
5596 				softdep_error("clear_remove: fsync", error);
5597 			drain_output(vp, 0);
5598 			vput(vp);
5599 			vn_finished_write(mp);
5600 			return;
5601 		}
5602 	}
5603 	FREE_LOCK(&lk);
5604 }
5605 
5606 /*
5607  * Clear out a block of dirty inodes in an effort to reduce
5608  * the number of inodedep dependency structures.
5609  */
5610 static void
5611 clear_inodedeps(td)
5612 	struct thread *td;
5613 {
5614 	struct inodedep_hashhead *inodedephd;
5615 	struct inodedep *inodedep;
5616 	static int next = 0;
5617 	struct mount *mp;
5618 	struct vnode *vp;
5619 	struct fs *fs;
5620 	int error, cnt;
5621 	ino_t firstino, lastino, ino;
5622 
5623 	ACQUIRE_LOCK(&lk);
5624 	/*
5625 	 * Pick a random inode dependency to be cleared.
5626 	 * We will then gather up all the inodes in its block
5627 	 * that have dependencies and flush them out.
5628 	 */
5629 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
5630 		inodedephd = &inodedep_hashtbl[next++];
5631 		if (next >= inodedep_hash)
5632 			next = 0;
5633 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
5634 			break;
5635 	}
5636 	if (inodedep == NULL)
5637 		return;
5638 	/*
5639 	 * Ugly code to find mount point given pointer to superblock.
5640 	 */
5641 	fs = inodedep->id_fs;
5642 	TAILQ_FOREACH(mp, &mountlist, mnt_list)
5643 		if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs)
5644 			break;
5645 	/*
5646 	 * Find the last inode in the block with dependencies.
5647 	 */
5648 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
5649 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
5650 		if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
5651 			break;
5652 	/*
5653 	 * Asynchronously push all but the last inode with dependencies.
5654 	 * Synchronously push the last inode with dependencies to ensure
5655 	 * that the inode block gets written to free up the inodedeps.
5656 	 */
5657 	for (ino = firstino; ino <= lastino; ino++) {
5658 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
5659 			continue;
5660 		FREE_LOCK(&lk);
5661 		if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5662 			continue;
5663 		if ((error = VFS_VGET(mp, ino, LK_EXCLUSIVE, &vp)) != 0) {
5664 			softdep_error("clear_inodedeps: vget", error);
5665 			vn_finished_write(mp);
5666 			return;
5667 		}
5668 		if (ino == lastino) {
5669 			if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_WAIT, td)))
5670 				softdep_error("clear_inodedeps: fsync1", error);
5671 		} else {
5672 			if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)))
5673 				softdep_error("clear_inodedeps: fsync2", error);
5674 			drain_output(vp, 0);
5675 		}
5676 		vput(vp);
5677 		vn_finished_write(mp);
5678 		ACQUIRE_LOCK(&lk);
5679 	}
5680 	FREE_LOCK(&lk);
5681 }
5682 
5683 /*
5684  * Function to determine if the buffer has outstanding dependencies
5685  * that will cause a roll-back if the buffer is written. If wantcount
5686  * is set, return number of dependencies, otherwise just yes or no.
5687  */
5688 static int
5689 softdep_count_dependencies(bp, wantcount)
5690 	struct buf *bp;
5691 	int wantcount;
5692 {
5693 	struct worklist *wk;
5694 	struct inodedep *inodedep;
5695 	struct indirdep *indirdep;
5696 	struct allocindir *aip;
5697 	struct pagedep *pagedep;
5698 	struct diradd *dap;
5699 	int i, retval;
5700 
5701 	retval = 0;
5702 	ACQUIRE_LOCK(&lk);
5703 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5704 		switch (wk->wk_type) {
5705 
5706 		case D_INODEDEP:
5707 			inodedep = WK_INODEDEP(wk);
5708 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5709 				/* bitmap allocation dependency */
5710 				retval += 1;
5711 				if (!wantcount)
5712 					goto out;
5713 			}
5714 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
5715 				/* direct block pointer dependency */
5716 				retval += 1;
5717 				if (!wantcount)
5718 					goto out;
5719 			}
5720 			if (TAILQ_FIRST(&inodedep->id_extupdt)) {
5721 				/* direct block pointer dependency */
5722 				retval += 1;
5723 				if (!wantcount)
5724 					goto out;
5725 			}
5726 			continue;
5727 
5728 		case D_INDIRDEP:
5729 			indirdep = WK_INDIRDEP(wk);
5730 
5731 			LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
5732 				/* indirect block pointer dependency */
5733 				retval += 1;
5734 				if (!wantcount)
5735 					goto out;
5736 			}
5737 			continue;
5738 
5739 		case D_PAGEDEP:
5740 			pagedep = WK_PAGEDEP(wk);
5741 			for (i = 0; i < DAHASHSZ; i++) {
5742 
5743 				LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
5744 					/* directory entry dependency */
5745 					retval += 1;
5746 					if (!wantcount)
5747 						goto out;
5748 				}
5749 			}
5750 			continue;
5751 
5752 		case D_BMSAFEMAP:
5753 		case D_ALLOCDIRECT:
5754 		case D_ALLOCINDIR:
5755 		case D_MKDIR:
5756 			/* never a dependency on these blocks */
5757 			continue;
5758 
5759 		default:
5760 			FREE_LOCK(&lk);
5761 			panic("softdep_check_for_rollback: Unexpected type %s",
5762 			    TYPENAME(wk->wk_type));
5763 			/* NOTREACHED */
5764 		}
5765 	}
5766 out:
5767 	FREE_LOCK(&lk);
5768 	return retval;
5769 }
5770 
5771 /*
5772  * Acquire exclusive access to a buffer.
5773  * Must be called with splbio blocked.
5774  * Return 1 if buffer was acquired.
5775  */
5776 static int
5777 getdirtybuf(bpp, waitfor)
5778 	struct buf **bpp;
5779 	int waitfor;
5780 {
5781 	struct buf *bp;
5782 	int error;
5783 
5784 	for (;;) {
5785 		if ((bp = *bpp) == NULL)
5786 			return (0);
5787 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
5788 			if ((bp->b_xflags & BX_BKGRDINPROG) == 0)
5789 				break;
5790 			BUF_UNLOCK(bp);
5791 			if (waitfor != MNT_WAIT)
5792 				return (0);
5793 			bp->b_xflags |= BX_BKGRDWAIT;
5794 			interlocked_sleep(&lk, SLEEP, &bp->b_xflags, NULL,
5795 			    PRIBIO, "getbuf", 0);
5796 			continue;
5797 		}
5798 		if (waitfor != MNT_WAIT)
5799 			return (0);
5800 		error = interlocked_sleep(&lk, LOCKBUF, bp, NULL,
5801 		    LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0);
5802 		if (error != ENOLCK) {
5803 			FREE_LOCK(&lk);
5804 			panic("getdirtybuf: inconsistent lock");
5805 		}
5806 	}
5807 	if ((bp->b_flags & B_DELWRI) == 0) {
5808 		BUF_UNLOCK(bp);
5809 		return (0);
5810 	}
5811 	bremfree(bp);
5812 	return (1);
5813 }
5814 
5815 /*
5816  * Wait for pending output on a vnode to complete.
5817  * Must be called with vnode locked.
5818  */
5819 static void
5820 drain_output(vp, islocked)
5821 	struct vnode *vp;
5822 	int islocked;
5823 {
5824 
5825 	if (!islocked)
5826 		ACQUIRE_LOCK(&lk);
5827 	VI_LOCK(vp);
5828 	while (vp->v_numoutput) {
5829 		vp->v_iflag |= VI_BWAIT;
5830 		interlocked_sleep(&lk, SLEEP, (caddr_t)&vp->v_numoutput,
5831 		    VI_MTX(vp), PRIBIO + 1, "drainvp", 0);
5832 	}
5833 	VI_UNLOCK(vp);
5834 	if (!islocked)
5835 		FREE_LOCK(&lk);
5836 }
5837 
5838 /*
5839  * Called whenever a buffer that is being invalidated or reallocated
5840  * contains dependencies. This should only happen if an I/O error has
5841  * occurred. The routine is called with the buffer locked.
5842  */
5843 static void
5844 softdep_deallocate_dependencies(bp)
5845 	struct buf *bp;
5846 {
5847 
5848 	if ((bp->b_ioflags & BIO_ERROR) == 0)
5849 		panic("softdep_deallocate_dependencies: dangling deps");
5850 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
5851 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
5852 }
5853 
5854 /*
5855  * Function to handle asynchronous write errors in the filesystem.
5856  */
5857 static void
5858 softdep_error(func, error)
5859 	char *func;
5860 	int error;
5861 {
5862 
5863 	/* XXX should do something better! */
5864 	printf("%s: got error %d while accessing filesystem\n", func, error);
5865 }
5866