xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision 77a0943ded95b9e6438f7db70c4a28e4d93946d4)
1 /*
2  * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
3  *
4  * The soft updates code is derived from the appendix of a University
5  * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
6  * "Soft Updates: A Solution to the Metadata Update Problem in File
7  * Systems", CSE-TR-254-95, August 1995).
8  *
9  * Further information about soft updates can be obtained from:
10  *
11  *	Marshall Kirk McKusick		http://www.mckusick.com/softdep/
12  *	1614 Oxford Street		mckusick@mckusick.com
13  *	Berkeley, CA 94709-1608		+1-510-843-9542
14  *	USA
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  *
20  * 1. Redistributions of source code must retain the above copyright
21  *    notice, this list of conditions and the following disclaimer.
22  * 2. Redistributions in binary form must reproduce the above copyright
23  *    notice, this list of conditions and the following disclaimer in the
24  *    documentation and/or other materials provided with the distribution.
25  *
26  * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
27  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
28  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
29  * DISCLAIMED.  IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
30  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)ffs_softdep.c	9.59 (McKusick) 6/21/00
39  * $FreeBSD$
40  */
41 
42 /*
43  * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
44  */
45 #ifndef DIAGNOSTIC
46 #define DIAGNOSTIC
47 #endif
48 #ifndef DEBUG
49 #define DEBUG
50 #endif
51 
52 #include <sys/param.h>
53 #include <sys/kernel.h>
54 #include <sys/systm.h>
55 #include <sys/bio.h>
56 #include <sys/buf.h>
57 #include <sys/malloc.h>
58 #include <sys/mount.h>
59 #include <sys/proc.h>
60 #include <sys/syslog.h>
61 #include <sys/vnode.h>
62 #include <sys/conf.h>
63 #include <ufs/ufs/dir.h>
64 #include <ufs/ufs/extattr.h>
65 #include <ufs/ufs/quota.h>
66 #include <ufs/ufs/inode.h>
67 #include <ufs/ufs/ufsmount.h>
68 #include <ufs/ffs/fs.h>
69 #include <ufs/ffs/softdep.h>
70 #include <ufs/ffs/ffs_extern.h>
71 #include <ufs/ufs/ufs_extern.h>
72 
73 /*
74  * These definitions need to be adapted to the system to which
75  * this file is being ported.
76  */
77 /*
78  * malloc types defined for the softdep system.
79  */
80 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
81 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
82 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
83 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
84 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
85 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
86 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
87 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
88 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
89 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
90 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
91 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
92 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
93 
94 #define M_SOFTDEP_FLAGS	(M_WAITOK | M_USE_RESERVE)
95 
96 #define	D_PAGEDEP	0
97 #define	D_INODEDEP	1
98 #define	D_NEWBLK	2
99 #define	D_BMSAFEMAP	3
100 #define	D_ALLOCDIRECT	4
101 #define	D_INDIRDEP	5
102 #define	D_ALLOCINDIR	6
103 #define	D_FREEFRAG	7
104 #define	D_FREEBLKS	8
105 #define	D_FREEFILE	9
106 #define	D_DIRADD	10
107 #define	D_MKDIR		11
108 #define	D_DIRREM	12
109 #define D_LAST		D_DIRREM
110 
111 /*
112  * translate from workitem type to memory type
113  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
114  */
115 static struct malloc_type *memtype[] = {
116 	M_PAGEDEP,
117 	M_INODEDEP,
118 	M_NEWBLK,
119 	M_BMSAFEMAP,
120 	M_ALLOCDIRECT,
121 	M_INDIRDEP,
122 	M_ALLOCINDIR,
123 	M_FREEFRAG,
124 	M_FREEBLKS,
125 	M_FREEFILE,
126 	M_DIRADD,
127 	M_MKDIR,
128 	M_DIRREM
129 };
130 
131 #define DtoM(type) (memtype[type])
132 
133 /*
134  * Names of malloc types.
135  */
136 #define TYPENAME(type)  \
137 	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
138 /*
139  * End system adaptaion definitions.
140  */
141 
142 /*
143  * Internal function prototypes.
144  */
145 static	void softdep_error __P((char *, int));
146 static	void drain_output __P((struct vnode *, int));
147 static	int getdirtybuf __P((struct buf **, int));
148 static	void clear_remove __P((struct proc *));
149 static	void clear_inodedeps __P((struct proc *));
150 static	int flush_pagedep_deps __P((struct vnode *, struct mount *,
151 	    struct diraddhd *));
152 static	int flush_inodedep_deps __P((struct fs *, ino_t));
153 static	int handle_written_filepage __P((struct pagedep *, struct buf *));
154 static  void diradd_inode_written __P((struct diradd *, struct inodedep *));
155 static	int handle_written_inodeblock __P((struct inodedep *, struct buf *));
156 static	void handle_allocdirect_partdone __P((struct allocdirect *));
157 static	void handle_allocindir_partdone __P((struct allocindir *));
158 static	void initiate_write_filepage __P((struct pagedep *, struct buf *));
159 static	void handle_written_mkdir __P((struct mkdir *, int));
160 static	void initiate_write_inodeblock __P((struct inodedep *, struct buf *));
161 static	void handle_workitem_freefile __P((struct freefile *));
162 static	void handle_workitem_remove __P((struct dirrem *));
163 static	struct dirrem *newdirrem __P((struct buf *, struct inode *,
164 	    struct inode *, int, struct dirrem **));
165 static	void free_diradd __P((struct diradd *));
166 static	void free_allocindir __P((struct allocindir *, struct inodedep *));
167 static	int indir_trunc __P((struct inode *, ufs_daddr_t, int, ufs_lbn_t,
168 	    long *));
169 static	void deallocate_dependencies __P((struct buf *, struct inodedep *));
170 static	void free_allocdirect __P((struct allocdirectlst *,
171 	    struct allocdirect *, int));
172 static	int check_inode_unwritten __P((struct inodedep *));
173 static	int free_inodedep __P((struct inodedep *));
174 static	void handle_workitem_freeblocks __P((struct freeblks *));
175 static	void merge_inode_lists __P((struct inodedep *));
176 static	void setup_allocindir_phase2 __P((struct buf *, struct inode *,
177 	    struct allocindir *));
178 static	struct allocindir *newallocindir __P((struct inode *, int, ufs_daddr_t,
179 	    ufs_daddr_t));
180 static	void handle_workitem_freefrag __P((struct freefrag *));
181 static	struct freefrag *newfreefrag __P((struct inode *, ufs_daddr_t, long));
182 static	void allocdirect_merge __P((struct allocdirectlst *,
183 	    struct allocdirect *, struct allocdirect *));
184 static	struct bmsafemap *bmsafemap_lookup __P((struct buf *));
185 static	int newblk_lookup __P((struct fs *, ufs_daddr_t, int,
186 	    struct newblk **));
187 static	int inodedep_lookup __P((struct fs *, ino_t, int, struct inodedep **));
188 static	int pagedep_lookup __P((struct inode *, ufs_lbn_t, int,
189 	    struct pagedep **));
190 static	void pause_timer __P((void *));
191 static	int request_cleanup __P((int, int));
192 static	void add_to_worklist __P((struct worklist *));
193 
194 /*
195  * Exported softdep operations.
196  */
197 static	void softdep_disk_io_initiation __P((struct buf *));
198 static	void softdep_disk_write_complete __P((struct buf *));
199 static	void softdep_deallocate_dependencies __P((struct buf *));
200 static	void softdep_move_dependencies __P((struct buf *, struct buf *));
201 static	int softdep_count_dependencies __P((struct buf *bp, int));
202 
203 struct bio_ops bioops = {
204 	softdep_disk_io_initiation,		/* io_start */
205 	softdep_disk_write_complete,		/* io_complete */
206 	softdep_deallocate_dependencies,	/* io_deallocate */
207 	softdep_move_dependencies,		/* io_movedeps */
208 	softdep_count_dependencies,		/* io_countdeps */
209 };
210 
211 /*
212  * Locking primitives.
213  *
214  * For a uniprocessor, all we need to do is protect against disk
215  * interrupts. For a multiprocessor, this lock would have to be
216  * a mutex. A single mutex is used throughout this file, though
217  * finer grain locking could be used if contention warranted it.
218  *
219  * For a multiprocessor, the sleep call would accept a lock and
220  * release it after the sleep processing was complete. In a uniprocessor
221  * implementation there is no such interlock, so we simple mark
222  * the places where it needs to be done with the `interlocked' form
223  * of the lock calls. Since the uniprocessor sleep already interlocks
224  * the spl, there is nothing that really needs to be done.
225  */
226 #ifndef /* NOT */ DEBUG
227 static struct lockit {
228 	int	lkt_spl;
229 } lk = { 0 };
230 #define ACQUIRE_LOCK(lk)		(lk)->lkt_spl = splbio()
231 #define FREE_LOCK(lk)			splx((lk)->lkt_spl)
232 #define ACQUIRE_LOCK_INTERLOCKED(lk)
233 #define FREE_LOCK_INTERLOCKED(lk)
234 
235 #else /* DEBUG */
236 static struct lockit {
237 	int	lkt_spl;
238 	pid_t	lkt_held;
239 } lk = { 0, -1 };
240 static int lockcnt;
241 
242 static	void acquire_lock __P((struct lockit *));
243 static	void free_lock __P((struct lockit *));
244 static	void acquire_lock_interlocked __P((struct lockit *));
245 static	void free_lock_interlocked __P((struct lockit *));
246 
247 #define ACQUIRE_LOCK(lk)		acquire_lock(lk)
248 #define FREE_LOCK(lk)			free_lock(lk)
249 #define ACQUIRE_LOCK_INTERLOCKED(lk)	acquire_lock_interlocked(lk)
250 #define FREE_LOCK_INTERLOCKED(lk)	free_lock_interlocked(lk)
251 
252 static void
253 acquire_lock(lk)
254 	struct lockit *lk;
255 {
256 
257 	if (lk->lkt_held != -1) {
258 		if (lk->lkt_held == CURPROC->p_pid)
259 			panic("softdep_lock: locking against myself");
260 		else
261 			panic("softdep_lock: lock held by %d", lk->lkt_held);
262 	}
263 	lk->lkt_spl = splbio();
264 	lk->lkt_held = CURPROC->p_pid;
265 	lockcnt++;
266 }
267 
268 static void
269 free_lock(lk)
270 	struct lockit *lk;
271 {
272 
273 	if (lk->lkt_held == -1)
274 		panic("softdep_unlock: lock not held");
275 	lk->lkt_held = -1;
276 	splx(lk->lkt_spl);
277 }
278 
279 static void
280 acquire_lock_interlocked(lk)
281 	struct lockit *lk;
282 {
283 
284 	if (lk->lkt_held != -1) {
285 		if (lk->lkt_held == CURPROC->p_pid)
286 			panic("softdep_lock_interlocked: locking against self");
287 		else
288 			panic("softdep_lock_interlocked: lock held by %d",
289 			    lk->lkt_held);
290 	}
291 	lk->lkt_held = CURPROC->p_pid;
292 	lockcnt++;
293 }
294 
295 static void
296 free_lock_interlocked(lk)
297 	struct lockit *lk;
298 {
299 
300 	if (lk->lkt_held == -1)
301 		panic("softdep_unlock_interlocked: lock not held");
302 	lk->lkt_held = -1;
303 }
304 #endif /* DEBUG */
305 
306 /*
307  * Place holder for real semaphores.
308  */
309 struct sema {
310 	int	value;
311 	pid_t	holder;
312 	char	*name;
313 	int	prio;
314 	int	timo;
315 };
316 static	void sema_init __P((struct sema *, char *, int, int));
317 static	int sema_get __P((struct sema *, struct lockit *));
318 static	void sema_release __P((struct sema *));
319 
320 static void
321 sema_init(semap, name, prio, timo)
322 	struct sema *semap;
323 	char *name;
324 	int prio, timo;
325 {
326 
327 	semap->holder = -1;
328 	semap->value = 0;
329 	semap->name = name;
330 	semap->prio = prio;
331 	semap->timo = timo;
332 }
333 
334 static int
335 sema_get(semap, interlock)
336 	struct sema *semap;
337 	struct lockit *interlock;
338 {
339 
340 	if (semap->value++ > 0) {
341 		if (interlock != NULL)
342 			FREE_LOCK_INTERLOCKED(interlock);
343 		tsleep((caddr_t)semap, semap->prio, semap->name, semap->timo);
344 		if (interlock != NULL) {
345 			ACQUIRE_LOCK_INTERLOCKED(interlock);
346 			FREE_LOCK(interlock);
347 		}
348 		return (0);
349 	}
350 	semap->holder = CURPROC->p_pid;
351 	if (interlock != NULL)
352 		FREE_LOCK(interlock);
353 	return (1);
354 }
355 
356 static void
357 sema_release(semap)
358 	struct sema *semap;
359 {
360 
361 	if (semap->value <= 0 || semap->holder != CURPROC->p_pid)
362 		panic("sema_release: not held");
363 	if (--semap->value > 0) {
364 		semap->value = 0;
365 		wakeup(semap);
366 	}
367 	semap->holder = -1;
368 }
369 
370 /*
371  * Worklist queue management.
372  * These routines require that the lock be held.
373  */
374 #ifndef /* NOT */ DEBUG
375 #define WORKLIST_INSERT(head, item) do {	\
376 	(item)->wk_state |= ONWORKLIST;		\
377 	LIST_INSERT_HEAD(head, item, wk_list);	\
378 } while (0)
379 #define WORKLIST_REMOVE(item) do {		\
380 	(item)->wk_state &= ~ONWORKLIST;	\
381 	LIST_REMOVE(item, wk_list);		\
382 } while (0)
383 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
384 
385 #else /* DEBUG */
386 static	void worklist_insert __P((struct workhead *, struct worklist *));
387 static	void worklist_remove __P((struct worklist *));
388 static	void workitem_free __P((struct worklist *, int));
389 
390 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
391 #define WORKLIST_REMOVE(item) worklist_remove(item)
392 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
393 
394 static void
395 worklist_insert(head, item)
396 	struct workhead *head;
397 	struct worklist *item;
398 {
399 
400 	if (lk.lkt_held == -1)
401 		panic("worklist_insert: lock not held");
402 	if (item->wk_state & ONWORKLIST)
403 		panic("worklist_insert: already on list");
404 	item->wk_state |= ONWORKLIST;
405 	LIST_INSERT_HEAD(head, item, wk_list);
406 }
407 
408 static void
409 worklist_remove(item)
410 	struct worklist *item;
411 {
412 
413 	if (lk.lkt_held == -1)
414 		panic("worklist_remove: lock not held");
415 	if ((item->wk_state & ONWORKLIST) == 0)
416 		panic("worklist_remove: not on list");
417 	item->wk_state &= ~ONWORKLIST;
418 	LIST_REMOVE(item, wk_list);
419 }
420 
421 static void
422 workitem_free(item, type)
423 	struct worklist *item;
424 	int type;
425 {
426 
427 	if (item->wk_state & ONWORKLIST)
428 		panic("workitem_free: still on list");
429 	if (item->wk_type != type)
430 		panic("workitem_free: type mismatch");
431 	FREE(item, DtoM(type));
432 }
433 #endif /* DEBUG */
434 
435 /*
436  * Workitem queue management
437  */
438 static struct workhead softdep_workitem_pending;
439 static int softdep_worklist_busy;
440 static int max_softdeps;	/* maximum number of structs before slowdown */
441 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
442 static int proc_waiting;	/* tracks whether we have a timeout posted */
443 static int *stat_countp;	/* statistic to count in proc_waiting timeout */
444 static struct callout_handle handle; /* handle on posted proc_waiting timeout */
445 static struct proc *filesys_syncer; /* proc of filesystem syncer process */
446 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
447 #define FLUSH_INODES	1
448 static int req_clear_remove;	/* syncer process flush some freeblks */
449 #define FLUSH_REMOVE	2
450 /*
451  * runtime statistics
452  */
453 static int stat_blk_limit_push;	/* number of times block limit neared */
454 static int stat_ino_limit_push;	/* number of times inode limit neared */
455 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
456 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
457 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
458 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
459 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
460 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
461 #ifdef DEBUG
462 #include <vm/vm.h>
463 #include <sys/sysctl.h>
464 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
465 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
466 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
467 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
468 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
469 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
470 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
471 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
472 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
473 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
474 #endif /* DEBUG */
475 
476 /*
477  * Add an item to the end of the work queue.
478  * This routine requires that the lock be held.
479  * This is the only routine that adds items to the list.
480  * The following routine is the only one that removes items
481  * and does so in order from first to last.
482  */
483 static void
484 add_to_worklist(wk)
485 	struct worklist *wk;
486 {
487 	static struct worklist *worklist_tail;
488 
489 	if (wk->wk_state & ONWORKLIST)
490 		panic("add_to_worklist: already on list");
491 	wk->wk_state |= ONWORKLIST;
492 	if (LIST_FIRST(&softdep_workitem_pending) == NULL)
493 		LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
494 	else
495 		LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
496 	worklist_tail = wk;
497 }
498 
499 /*
500  * Process that runs once per second to handle items in the background queue.
501  *
502  * Note that we ensure that everything is done in the order in which they
503  * appear in the queue. The code below depends on this property to ensure
504  * that blocks of a file are freed before the inode itself is freed. This
505  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
506  * until all the old ones have been purged from the dependency lists.
507  */
508 int
509 softdep_process_worklist(matchmnt)
510 	struct mount *matchmnt;
511 {
512 	struct proc *p = CURPROC;
513 	struct worklist *wk;
514 	struct mount *mp;
515 	int matchcnt, loopcount;
516 
517 	/*
518 	 * Record the process identifier of our caller so that we can give
519 	 * this process preferential treatment in request_cleanup below.
520 	 */
521 	filesys_syncer = p;
522 	matchcnt = 0;
523 	/*
524 	 * There is no danger of having multiple processes run this
525 	 * code. It is single threaded solely so that softdep_flushfiles
526 	 * (below) can get an accurate count of the number of items
527 	 * related to its mount point that are in the list.
528 	 */
529 	if (softdep_worklist_busy && matchmnt == NULL)
530 		return (-1);
531 	/*
532 	 * If requested, try removing inode or removal dependencies.
533 	 */
534 	if (req_clear_inodedeps) {
535 		clear_inodedeps(p);
536 		req_clear_inodedeps -= 1;
537 		wakeup_one(&proc_waiting);
538 	}
539 	if (req_clear_remove) {
540 		clear_remove(p);
541 		req_clear_remove -= 1;
542 		wakeup_one(&proc_waiting);
543 	}
544 	ACQUIRE_LOCK(&lk);
545 	loopcount = 1;
546 	while ((wk = LIST_FIRST(&softdep_workitem_pending)) != 0) {
547 		WORKLIST_REMOVE(wk);
548 		FREE_LOCK(&lk);
549 		switch (wk->wk_type) {
550 
551 		case D_DIRREM:
552 			/* removal of a directory entry */
553 			mp = WK_DIRREM(wk)->dm_mnt;
554 			if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
555 				panic("%s: dirrem on suspended filesystem",
556 					"softdep_process_worklist");
557 			if (mp == matchmnt)
558 				matchcnt += 1;
559 			handle_workitem_remove(WK_DIRREM(wk));
560 			break;
561 
562 		case D_FREEBLKS:
563 			/* releasing blocks and/or fragments from a file */
564 			mp = WK_FREEBLKS(wk)->fb_mnt;
565 			if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
566 				panic("%s: freeblks on suspended filesystem",
567 					"softdep_process_worklist");
568 			if (mp == matchmnt)
569 				matchcnt += 1;
570 			handle_workitem_freeblocks(WK_FREEBLKS(wk));
571 			break;
572 
573 		case D_FREEFRAG:
574 			/* releasing a fragment when replaced as a file grows */
575 			mp = WK_FREEFRAG(wk)->ff_mnt;
576 			if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
577 				panic("%s: freefrag on suspended filesystem",
578 					"softdep_process_worklist");
579 			if (mp == matchmnt)
580 				matchcnt += 1;
581 			handle_workitem_freefrag(WK_FREEFRAG(wk));
582 			break;
583 
584 		case D_FREEFILE:
585 			/* releasing an inode when its link count drops to 0 */
586 			mp = WK_FREEFILE(wk)->fx_mnt;
587 			if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
588 				panic("%s: freefile on suspended filesystem",
589 					"softdep_process_worklist");
590 			if (mp == matchmnt)
591 				matchcnt += 1;
592 			handle_workitem_freefile(WK_FREEFILE(wk));
593 			break;
594 
595 		default:
596 			panic("%s_process_worklist: Unknown type %s",
597 			    "softdep", TYPENAME(wk->wk_type));
598 			/* NOTREACHED */
599 		}
600 		if (softdep_worklist_busy && matchmnt == NULL)
601 			return (-1);
602 		/*
603 		 * If requested, try removing inode or removal dependencies.
604 		 */
605 		if (req_clear_inodedeps) {
606 			clear_inodedeps(p);
607 			req_clear_inodedeps -= 1;
608 			wakeup_one(&proc_waiting);
609 		}
610 		if (req_clear_remove) {
611 			clear_remove(p);
612 			req_clear_remove -= 1;
613 			wakeup_one(&proc_waiting);
614 		}
615 		/*
616 		 * We do not generally want to stop for buffer space, but if
617 		 * we are really being a buffer hog, we will stop and wait.
618 		 */
619 		if (loopcount++ % 128 == 0)
620 			bwillwrite();
621 		ACQUIRE_LOCK(&lk);
622 	}
623 	FREE_LOCK(&lk);
624 	return (matchcnt);
625 }
626 
627 /*
628  * Move dependencies from one buffer to another.
629  */
630 static void
631 softdep_move_dependencies(oldbp, newbp)
632 	struct buf *oldbp;
633 	struct buf *newbp;
634 {
635 	struct worklist *wk, *wktail;
636 
637 	if (LIST_FIRST(&newbp->b_dep) != NULL)
638 		panic("softdep_move_dependencies: need merge code");
639 	wktail = 0;
640 	ACQUIRE_LOCK(&lk);
641 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
642 		LIST_REMOVE(wk, wk_list);
643 		if (wktail == 0)
644 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
645 		else
646 			LIST_INSERT_AFTER(wktail, wk, wk_list);
647 		wktail = wk;
648 	}
649 	FREE_LOCK(&lk);
650 }
651 
652 /*
653  * Purge the work list of all items associated with a particular mount point.
654  */
655 int
656 softdep_flushworklist(oldmnt, countp, p)
657 	struct mount *oldmnt;
658 	int *countp;
659 	struct proc *p;
660 {
661 	struct vnode *devvp;
662 	int count, error = 0;
663 
664 	/*
665 	 * Await our turn to clear out the queue.
666 	 */
667 	while (softdep_worklist_busy)
668 		tsleep(&lbolt, PRIBIO, "softflush", 0);
669 	softdep_worklist_busy = 1;
670 	/*
671 	 * Alternately flush the block device associated with the mount
672 	 * point and process any dependencies that the flushing
673 	 * creates. We continue until no more worklist dependencies
674 	 * are found.
675 	 */
676 	*countp = 0;
677 	devvp = VFSTOUFS(oldmnt)->um_devvp;
678 	while ((count = softdep_process_worklist(oldmnt)) > 0) {
679 		*countp += count;
680 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, p);
681 		error = VOP_FSYNC(devvp, p->p_ucred, MNT_WAIT, p);
682 		VOP_UNLOCK(devvp, 0, p);
683 		if (error)
684 			break;
685 	}
686 	softdep_worklist_busy = 0;
687 	return (error);
688 }
689 
690 /*
691  * Flush all vnodes and worklist items associated with a specified mount point.
692  */
693 int
694 softdep_flushfiles(oldmnt, flags, p)
695 	struct mount *oldmnt;
696 	int flags;
697 	struct proc *p;
698 {
699 	int error, count, loopcnt;
700 
701 	/*
702 	 * Alternately flush the vnodes associated with the mount
703 	 * point and process any dependencies that the flushing
704 	 * creates. In theory, this loop can happen at most twice,
705 	 * but we give it a few extra just to be sure.
706 	 */
707 	for (loopcnt = 10; loopcnt > 0; loopcnt--) {
708 		/*
709 		 * Do another flush in case any vnodes were brought in
710 		 * as part of the cleanup operations.
711 		 */
712 		if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0)
713 			break;
714 		if ((error = softdep_flushworklist(oldmnt, &count, p)) != 0 ||
715 		    count == 0)
716 			break;
717 	}
718 	/*
719 	 * If we are unmounting then it is an error to fail. If we
720 	 * are simply trying to downgrade to read-only, then filesystem
721 	 * activity can keep us busy forever, so we just fail with EBUSY.
722 	 */
723 	if (loopcnt == 0) {
724 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
725 			panic("softdep_flushfiles: looping");
726 		error = EBUSY;
727 	}
728 	return (error);
729 }
730 
731 /*
732  * Structure hashing.
733  *
734  * There are three types of structures that can be looked up:
735  *	1) pagedep structures identified by mount point, inode number,
736  *	   and logical block.
737  *	2) inodedep structures identified by mount point and inode number.
738  *	3) newblk structures identified by mount point and
739  *	   physical block number.
740  *
741  * The "pagedep" and "inodedep" dependency structures are hashed
742  * separately from the file blocks and inodes to which they correspond.
743  * This separation helps when the in-memory copy of an inode or
744  * file block must be replaced. It also obviates the need to access
745  * an inode or file page when simply updating (or de-allocating)
746  * dependency structures. Lookup of newblk structures is needed to
747  * find newly allocated blocks when trying to associate them with
748  * their allocdirect or allocindir structure.
749  *
750  * The lookup routines optionally create and hash a new instance when
751  * an existing entry is not found.
752  */
753 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
754 
755 /*
756  * Structures and routines associated with pagedep caching.
757  */
758 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
759 u_long	pagedep_hash;		/* size of hash table - 1 */
760 #define	PAGEDEP_HASH(mp, inum, lbn) \
761 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
762 	    pagedep_hash])
763 static struct sema pagedep_in_progress;
764 
765 /*
766  * Look up a pagedep. Return 1 if found, 0 if not found.
767  * If not found, allocate if DEPALLOC flag is passed.
768  * Found or allocated entry is returned in pagedeppp.
769  * This routine must be called with splbio interrupts blocked.
770  */
771 static int
772 pagedep_lookup(ip, lbn, flags, pagedeppp)
773 	struct inode *ip;
774 	ufs_lbn_t lbn;
775 	int flags;
776 	struct pagedep **pagedeppp;
777 {
778 	struct pagedep *pagedep;
779 	struct pagedep_hashhead *pagedephd;
780 	struct mount *mp;
781 	int i;
782 
783 #ifdef DEBUG
784 	if (lk.lkt_held == -1)
785 		panic("pagedep_lookup: lock not held");
786 #endif
787 	mp = ITOV(ip)->v_mount;
788 	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
789 top:
790 	for (pagedep = LIST_FIRST(pagedephd); pagedep;
791 	     pagedep = LIST_NEXT(pagedep, pd_hash))
792 		if (ip->i_number == pagedep->pd_ino &&
793 		    lbn == pagedep->pd_lbn &&
794 		    mp == pagedep->pd_mnt)
795 			break;
796 	if (pagedep) {
797 		*pagedeppp = pagedep;
798 		return (1);
799 	}
800 	if ((flags & DEPALLOC) == 0) {
801 		*pagedeppp = NULL;
802 		return (0);
803 	}
804 	if (sema_get(&pagedep_in_progress, &lk) == 0) {
805 		ACQUIRE_LOCK(&lk);
806 		goto top;
807 	}
808 	MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
809 		M_SOFTDEP_FLAGS);
810 	bzero(pagedep, sizeof(struct pagedep));
811 	pagedep->pd_list.wk_type = D_PAGEDEP;
812 	pagedep->pd_mnt = mp;
813 	pagedep->pd_ino = ip->i_number;
814 	pagedep->pd_lbn = lbn;
815 	LIST_INIT(&pagedep->pd_dirremhd);
816 	LIST_INIT(&pagedep->pd_pendinghd);
817 	for (i = 0; i < DAHASHSZ; i++)
818 		LIST_INIT(&pagedep->pd_diraddhd[i]);
819 	ACQUIRE_LOCK(&lk);
820 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
821 	sema_release(&pagedep_in_progress);
822 	*pagedeppp = pagedep;
823 	return (0);
824 }
825 
826 /*
827  * Structures and routines associated with inodedep caching.
828  */
829 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
830 static u_long	inodedep_hash;	/* size of hash table - 1 */
831 static long	num_inodedep;	/* number of inodedep allocated */
832 #define	INODEDEP_HASH(fs, inum) \
833       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
834 static struct sema inodedep_in_progress;
835 
836 /*
837  * Look up a inodedep. Return 1 if found, 0 if not found.
838  * If not found, allocate if DEPALLOC flag is passed.
839  * Found or allocated entry is returned in inodedeppp.
840  * This routine must be called with splbio interrupts blocked.
841  */
842 static int
843 inodedep_lookup(fs, inum, flags, inodedeppp)
844 	struct fs *fs;
845 	ino_t inum;
846 	int flags;
847 	struct inodedep **inodedeppp;
848 {
849 	struct inodedep *inodedep;
850 	struct inodedep_hashhead *inodedephd;
851 	int firsttry;
852 
853 #ifdef DEBUG
854 	if (lk.lkt_held == -1)
855 		panic("inodedep_lookup: lock not held");
856 #endif
857 	firsttry = 1;
858 	inodedephd = INODEDEP_HASH(fs, inum);
859 top:
860 	for (inodedep = LIST_FIRST(inodedephd); inodedep;
861 	     inodedep = LIST_NEXT(inodedep, id_hash))
862 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
863 			break;
864 	if (inodedep) {
865 		*inodedeppp = inodedep;
866 		return (1);
867 	}
868 	if ((flags & DEPALLOC) == 0) {
869 		*inodedeppp = NULL;
870 		return (0);
871 	}
872 	/*
873 	 * If we are over our limit, try to improve the situation.
874 	 */
875 	if (num_inodedep > max_softdeps && firsttry && speedup_syncer() == 0 &&
876 	    request_cleanup(FLUSH_INODES, 1)) {
877 		firsttry = 0;
878 		goto top;
879 	}
880 	if (sema_get(&inodedep_in_progress, &lk) == 0) {
881 		ACQUIRE_LOCK(&lk);
882 		goto top;
883 	}
884 	num_inodedep += 1;
885 	MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
886 		M_INODEDEP, M_SOFTDEP_FLAGS);
887 	inodedep->id_list.wk_type = D_INODEDEP;
888 	inodedep->id_fs = fs;
889 	inodedep->id_ino = inum;
890 	inodedep->id_state = ALLCOMPLETE;
891 	inodedep->id_nlinkdelta = 0;
892 	inodedep->id_savedino = NULL;
893 	inodedep->id_savedsize = -1;
894 	inodedep->id_buf = NULL;
895 	LIST_INIT(&inodedep->id_pendinghd);
896 	LIST_INIT(&inodedep->id_inowait);
897 	LIST_INIT(&inodedep->id_bufwait);
898 	TAILQ_INIT(&inodedep->id_inoupdt);
899 	TAILQ_INIT(&inodedep->id_newinoupdt);
900 	ACQUIRE_LOCK(&lk);
901 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
902 	sema_release(&inodedep_in_progress);
903 	*inodedeppp = inodedep;
904 	return (0);
905 }
906 
907 /*
908  * Structures and routines associated with newblk caching.
909  */
910 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
911 u_long	newblk_hash;		/* size of hash table - 1 */
912 #define	NEWBLK_HASH(fs, inum) \
913 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
914 static struct sema newblk_in_progress;
915 
916 /*
917  * Look up a newblk. Return 1 if found, 0 if not found.
918  * If not found, allocate if DEPALLOC flag is passed.
919  * Found or allocated entry is returned in newblkpp.
920  */
921 static int
922 newblk_lookup(fs, newblkno, flags, newblkpp)
923 	struct fs *fs;
924 	ufs_daddr_t newblkno;
925 	int flags;
926 	struct newblk **newblkpp;
927 {
928 	struct newblk *newblk;
929 	struct newblk_hashhead *newblkhd;
930 
931 	newblkhd = NEWBLK_HASH(fs, newblkno);
932 top:
933 	for (newblk = LIST_FIRST(newblkhd); newblk;
934 	     newblk = LIST_NEXT(newblk, nb_hash))
935 		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
936 			break;
937 	if (newblk) {
938 		*newblkpp = newblk;
939 		return (1);
940 	}
941 	if ((flags & DEPALLOC) == 0) {
942 		*newblkpp = NULL;
943 		return (0);
944 	}
945 	if (sema_get(&newblk_in_progress, 0) == 0)
946 		goto top;
947 	MALLOC(newblk, struct newblk *, sizeof(struct newblk),
948 		M_NEWBLK, M_SOFTDEP_FLAGS);
949 	newblk->nb_state = 0;
950 	newblk->nb_fs = fs;
951 	newblk->nb_newblkno = newblkno;
952 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
953 	sema_release(&newblk_in_progress);
954 	*newblkpp = newblk;
955 	return (0);
956 }
957 
958 /*
959  * Executed during filesystem system initialization before
960  * mounting any file systems.
961  */
962 void
963 softdep_initialize()
964 {
965 
966 	LIST_INIT(&mkdirlisthd);
967 	LIST_INIT(&softdep_workitem_pending);
968 	max_softdeps = desiredvnodes * 8;
969 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
970 	    &pagedep_hash);
971 	sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0);
972 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
973 	sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0);
974 	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
975 	sema_init(&newblk_in_progress, "newblk", PRIBIO, 0);
976 }
977 
978 /*
979  * Called at mount time to notify the dependency code that a
980  * filesystem wishes to use it.
981  */
982 int
983 softdep_mount(devvp, mp, fs, cred)
984 	struct vnode *devvp;
985 	struct mount *mp;
986 	struct fs *fs;
987 	struct ucred *cred;
988 {
989 	struct csum cstotal;
990 	struct cg *cgp;
991 	struct buf *bp;
992 	int error, cyl;
993 
994 	mp->mnt_flag &= ~MNT_ASYNC;
995 	mp->mnt_flag |= MNT_SOFTDEP;
996 	/*
997 	 * When doing soft updates, the counters in the
998 	 * superblock may have gotten out of sync, so we have
999 	 * to scan the cylinder groups and recalculate them.
1000 	 */
1001 	if (fs->fs_clean != 0)
1002 		return (0);
1003 	bzero(&cstotal, sizeof cstotal);
1004 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1005 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1006 		    fs->fs_cgsize, cred, &bp)) != 0) {
1007 			brelse(bp);
1008 			return (error);
1009 		}
1010 		cgp = (struct cg *)bp->b_data;
1011 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1012 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1013 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1014 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1015 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
1016 		brelse(bp);
1017 	}
1018 #ifdef DEBUG
1019 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1020 		printf("ffs_mountfs: superblock updated for soft updates\n");
1021 #endif
1022 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1023 	return (0);
1024 }
1025 
1026 /*
1027  * Protecting the freemaps (or bitmaps).
1028  *
1029  * To eliminate the need to execute fsck before mounting a file system
1030  * after a power failure, one must (conservatively) guarantee that the
1031  * on-disk copy of the bitmaps never indicate that a live inode or block is
1032  * free.  So, when a block or inode is allocated, the bitmap should be
1033  * updated (on disk) before any new pointers.  When a block or inode is
1034  * freed, the bitmap should not be updated until all pointers have been
1035  * reset.  The latter dependency is handled by the delayed de-allocation
1036  * approach described below for block and inode de-allocation.  The former
1037  * dependency is handled by calling the following procedure when a block or
1038  * inode is allocated. When an inode is allocated an "inodedep" is created
1039  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1040  * Each "inodedep" is also inserted into the hash indexing structure so
1041  * that any additional link additions can be made dependent on the inode
1042  * allocation.
1043  *
1044  * The ufs file system maintains a number of free block counts (e.g., per
1045  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1046  * in addition to the bitmaps.  These counts are used to improve efficiency
1047  * during allocation and therefore must be consistent with the bitmaps.
1048  * There is no convenient way to guarantee post-crash consistency of these
1049  * counts with simple update ordering, for two main reasons: (1) The counts
1050  * and bitmaps for a single cylinder group block are not in the same disk
1051  * sector.  If a disk write is interrupted (e.g., by power failure), one may
1052  * be written and the other not.  (2) Some of the counts are located in the
1053  * superblock rather than the cylinder group block. So, we focus our soft
1054  * updates implementation on protecting the bitmaps. When mounting a
1055  * filesystem, we recompute the auxiliary counts from the bitmaps.
1056  */
1057 
1058 /*
1059  * Called just after updating the cylinder group block to allocate an inode.
1060  */
1061 void
1062 softdep_setup_inomapdep(bp, ip, newinum)
1063 	struct buf *bp;		/* buffer for cylgroup block with inode map */
1064 	struct inode *ip;	/* inode related to allocation */
1065 	ino_t newinum;		/* new inode number being allocated */
1066 {
1067 	struct inodedep *inodedep;
1068 	struct bmsafemap *bmsafemap;
1069 
1070 	/*
1071 	 * Create a dependency for the newly allocated inode.
1072 	 * Panic if it already exists as something is seriously wrong.
1073 	 * Otherwise add it to the dependency list for the buffer holding
1074 	 * the cylinder group map from which it was allocated.
1075 	 */
1076 	ACQUIRE_LOCK(&lk);
1077 	if (inodedep_lookup(ip->i_fs, newinum, DEPALLOC, &inodedep) != 0)
1078 		panic("softdep_setup_inomapdep: found inode");
1079 	inodedep->id_buf = bp;
1080 	inodedep->id_state &= ~DEPCOMPLETE;
1081 	bmsafemap = bmsafemap_lookup(bp);
1082 	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1083 	FREE_LOCK(&lk);
1084 }
1085 
1086 /*
1087  * Called just after updating the cylinder group block to
1088  * allocate block or fragment.
1089  */
1090 void
1091 softdep_setup_blkmapdep(bp, fs, newblkno)
1092 	struct buf *bp;		/* buffer for cylgroup block with block map */
1093 	struct fs *fs;		/* filesystem doing allocation */
1094 	ufs_daddr_t newblkno;	/* number of newly allocated block */
1095 {
1096 	struct newblk *newblk;
1097 	struct bmsafemap *bmsafemap;
1098 
1099 	/*
1100 	 * Create a dependency for the newly allocated block.
1101 	 * Add it to the dependency list for the buffer holding
1102 	 * the cylinder group map from which it was allocated.
1103 	 */
1104 	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1105 		panic("softdep_setup_blkmapdep: found block");
1106 	ACQUIRE_LOCK(&lk);
1107 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1108 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1109 	FREE_LOCK(&lk);
1110 }
1111 
1112 /*
1113  * Find the bmsafemap associated with a cylinder group buffer.
1114  * If none exists, create one. The buffer must be locked when
1115  * this routine is called and this routine must be called with
1116  * splbio interrupts blocked.
1117  */
1118 static struct bmsafemap *
1119 bmsafemap_lookup(bp)
1120 	struct buf *bp;
1121 {
1122 	struct bmsafemap *bmsafemap;
1123 	struct worklist *wk;
1124 
1125 #ifdef DEBUG
1126 	if (lk.lkt_held == -1)
1127 		panic("bmsafemap_lookup: lock not held");
1128 #endif
1129 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = LIST_NEXT(wk, wk_list))
1130 		if (wk->wk_type == D_BMSAFEMAP)
1131 			return (WK_BMSAFEMAP(wk));
1132 	FREE_LOCK(&lk);
1133 	MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1134 		M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1135 	bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1136 	bmsafemap->sm_list.wk_state = 0;
1137 	bmsafemap->sm_buf = bp;
1138 	LIST_INIT(&bmsafemap->sm_allocdirecthd);
1139 	LIST_INIT(&bmsafemap->sm_allocindirhd);
1140 	LIST_INIT(&bmsafemap->sm_inodedephd);
1141 	LIST_INIT(&bmsafemap->sm_newblkhd);
1142 	ACQUIRE_LOCK(&lk);
1143 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1144 	return (bmsafemap);
1145 }
1146 
1147 /*
1148  * Direct block allocation dependencies.
1149  *
1150  * When a new block is allocated, the corresponding disk locations must be
1151  * initialized (with zeros or new data) before the on-disk inode points to
1152  * them.  Also, the freemap from which the block was allocated must be
1153  * updated (on disk) before the inode's pointer. These two dependencies are
1154  * independent of each other and are needed for all file blocks and indirect
1155  * blocks that are pointed to directly by the inode.  Just before the
1156  * "in-core" version of the inode is updated with a newly allocated block
1157  * number, a procedure (below) is called to setup allocation dependency
1158  * structures.  These structures are removed when the corresponding
1159  * dependencies are satisfied or when the block allocation becomes obsolete
1160  * (i.e., the file is deleted, the block is de-allocated, or the block is a
1161  * fragment that gets upgraded).  All of these cases are handled in
1162  * procedures described later.
1163  *
1164  * When a file extension causes a fragment to be upgraded, either to a larger
1165  * fragment or to a full block, the on-disk location may change (if the
1166  * previous fragment could not simply be extended). In this case, the old
1167  * fragment must be de-allocated, but not until after the inode's pointer has
1168  * been updated. In most cases, this is handled by later procedures, which
1169  * will construct a "freefrag" structure to be added to the workitem queue
1170  * when the inode update is complete (or obsolete).  The main exception to
1171  * this is when an allocation occurs while a pending allocation dependency
1172  * (for the same block pointer) remains.  This case is handled in the main
1173  * allocation dependency setup procedure by immediately freeing the
1174  * unreferenced fragments.
1175  */
1176 void
1177 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1178 	struct inode *ip;	/* inode to which block is being added */
1179 	ufs_lbn_t lbn;		/* block pointer within inode */
1180 	ufs_daddr_t newblkno;	/* disk block number being added */
1181 	ufs_daddr_t oldblkno;	/* previous block number, 0 unless frag */
1182 	long newsize;		/* size of new block */
1183 	long oldsize;		/* size of new block */
1184 	struct buf *bp;		/* bp for allocated block */
1185 {
1186 	struct allocdirect *adp, *oldadp;
1187 	struct allocdirectlst *adphead;
1188 	struct bmsafemap *bmsafemap;
1189 	struct inodedep *inodedep;
1190 	struct pagedep *pagedep;
1191 	struct newblk *newblk;
1192 
1193 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1194 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS);
1195 	bzero(adp, sizeof(struct allocdirect));
1196 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1197 	adp->ad_lbn = lbn;
1198 	adp->ad_newblkno = newblkno;
1199 	adp->ad_oldblkno = oldblkno;
1200 	adp->ad_newsize = newsize;
1201 	adp->ad_oldsize = oldsize;
1202 	adp->ad_state = ATTACHED;
1203 	if (newblkno == oldblkno)
1204 		adp->ad_freefrag = NULL;
1205 	else
1206 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1207 
1208 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1209 		panic("softdep_setup_allocdirect: lost block");
1210 
1211 	ACQUIRE_LOCK(&lk);
1212 	(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
1213 	adp->ad_inodedep = inodedep;
1214 
1215 	if (newblk->nb_state == DEPCOMPLETE) {
1216 		adp->ad_state |= DEPCOMPLETE;
1217 		adp->ad_buf = NULL;
1218 	} else {
1219 		bmsafemap = newblk->nb_bmsafemap;
1220 		adp->ad_buf = bmsafemap->sm_buf;
1221 		LIST_REMOVE(newblk, nb_deps);
1222 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1223 	}
1224 	LIST_REMOVE(newblk, nb_hash);
1225 	FREE(newblk, M_NEWBLK);
1226 
1227 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1228 	if (lbn >= NDADDR) {
1229 		/* allocating an indirect block */
1230 		if (oldblkno != 0)
1231 			panic("softdep_setup_allocdirect: non-zero indir");
1232 	} else {
1233 		/*
1234 		 * Allocating a direct block.
1235 		 *
1236 		 * If we are allocating a directory block, then we must
1237 		 * allocate an associated pagedep to track additions and
1238 		 * deletions.
1239 		 */
1240 		if ((ip->i_mode & IFMT) == IFDIR &&
1241 		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1242 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1243 	}
1244 	/*
1245 	 * The list of allocdirects must be kept in sorted and ascending
1246 	 * order so that the rollback routines can quickly determine the
1247 	 * first uncommitted block (the size of the file stored on disk
1248 	 * ends at the end of the lowest committed fragment, or if there
1249 	 * are no fragments, at the end of the highest committed block).
1250 	 * Since files generally grow, the typical case is that the new
1251 	 * block is to be added at the end of the list. We speed this
1252 	 * special case by checking against the last allocdirect in the
1253 	 * list before laboriously traversing the list looking for the
1254 	 * insertion point.
1255 	 */
1256 	adphead = &inodedep->id_newinoupdt;
1257 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1258 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1259 		/* insert at end of list */
1260 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1261 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1262 			allocdirect_merge(adphead, adp, oldadp);
1263 		FREE_LOCK(&lk);
1264 		return;
1265 	}
1266 	for (oldadp = TAILQ_FIRST(adphead); oldadp;
1267 	     oldadp = TAILQ_NEXT(oldadp, ad_next)) {
1268 		if (oldadp->ad_lbn >= lbn)
1269 			break;
1270 	}
1271 	if (oldadp == NULL)
1272 		panic("softdep_setup_allocdirect: lost entry");
1273 	/* insert in middle of list */
1274 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1275 	if (oldadp->ad_lbn == lbn)
1276 		allocdirect_merge(adphead, adp, oldadp);
1277 	FREE_LOCK(&lk);
1278 }
1279 
1280 /*
1281  * Replace an old allocdirect dependency with a newer one.
1282  * This routine must be called with splbio interrupts blocked.
1283  */
1284 static void
1285 allocdirect_merge(adphead, newadp, oldadp)
1286 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
1287 	struct allocdirect *newadp;	/* allocdirect being added */
1288 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
1289 {
1290 	struct freefrag *freefrag;
1291 
1292 #ifdef DEBUG
1293 	if (lk.lkt_held == -1)
1294 		panic("allocdirect_merge: lock not held");
1295 #endif
1296 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1297 	    newadp->ad_oldsize != oldadp->ad_newsize ||
1298 	    newadp->ad_lbn >= NDADDR)
1299 		panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1300 		    newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1301 		    NDADDR);
1302 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
1303 	newadp->ad_oldsize = oldadp->ad_oldsize;
1304 	/*
1305 	 * If the old dependency had a fragment to free or had never
1306 	 * previously had a block allocated, then the new dependency
1307 	 * can immediately post its freefrag and adopt the old freefrag.
1308 	 * This action is done by swapping the freefrag dependencies.
1309 	 * The new dependency gains the old one's freefrag, and the
1310 	 * old one gets the new one and then immediately puts it on
1311 	 * the worklist when it is freed by free_allocdirect. It is
1312 	 * not possible to do this swap when the old dependency had a
1313 	 * non-zero size but no previous fragment to free. This condition
1314 	 * arises when the new block is an extension of the old block.
1315 	 * Here, the first part of the fragment allocated to the new
1316 	 * dependency is part of the block currently claimed on disk by
1317 	 * the old dependency, so cannot legitimately be freed until the
1318 	 * conditions for the new dependency are fulfilled.
1319 	 */
1320 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1321 		freefrag = newadp->ad_freefrag;
1322 		newadp->ad_freefrag = oldadp->ad_freefrag;
1323 		oldadp->ad_freefrag = freefrag;
1324 	}
1325 	free_allocdirect(adphead, oldadp, 0);
1326 }
1327 
1328 /*
1329  * Allocate a new freefrag structure if needed.
1330  */
1331 static struct freefrag *
1332 newfreefrag(ip, blkno, size)
1333 	struct inode *ip;
1334 	ufs_daddr_t blkno;
1335 	long size;
1336 {
1337 	struct freefrag *freefrag;
1338 	struct fs *fs;
1339 
1340 	if (blkno == 0)
1341 		return (NULL);
1342 	fs = ip->i_fs;
1343 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1344 		panic("newfreefrag: frag size");
1345 	MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1346 		M_FREEFRAG, M_SOFTDEP_FLAGS);
1347 	freefrag->ff_list.wk_type = D_FREEFRAG;
1348 	freefrag->ff_state = ip->i_uid & ~ONWORKLIST;	/* XXX - used below */
1349 	freefrag->ff_inum = ip->i_number;
1350 	freefrag->ff_mnt = ITOV(ip)->v_mount;
1351 	freefrag->ff_devvp = ip->i_devvp;
1352 	freefrag->ff_blkno = blkno;
1353 	freefrag->ff_fragsize = size;
1354 	return (freefrag);
1355 }
1356 
1357 /*
1358  * This workitem de-allocates fragments that were replaced during
1359  * file block allocation.
1360  */
1361 static void
1362 handle_workitem_freefrag(freefrag)
1363 	struct freefrag *freefrag;
1364 {
1365 	struct inode tip;
1366 
1367 	tip.i_vnode = NULL;
1368 	tip.i_fs = VFSTOUFS(freefrag->ff_mnt)->um_fs;
1369 	tip.i_devvp = freefrag->ff_devvp;
1370 	tip.i_dev = freefrag->ff_devvp->v_rdev;
1371 	tip.i_number = freefrag->ff_inum;
1372 	tip.i_uid = freefrag->ff_state & ~ONWORKLIST;	/* XXX - set above */
1373 	ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1374 	FREE(freefrag, M_FREEFRAG);
1375 }
1376 
1377 /*
1378  * Indirect block allocation dependencies.
1379  *
1380  * The same dependencies that exist for a direct block also exist when
1381  * a new block is allocated and pointed to by an entry in a block of
1382  * indirect pointers. The undo/redo states described above are also
1383  * used here. Because an indirect block contains many pointers that
1384  * may have dependencies, a second copy of the entire in-memory indirect
1385  * block is kept. The buffer cache copy is always completely up-to-date.
1386  * The second copy, which is used only as a source for disk writes,
1387  * contains only the safe pointers (i.e., those that have no remaining
1388  * update dependencies). The second copy is freed when all pointers
1389  * are safe. The cache is not allowed to replace indirect blocks with
1390  * pending update dependencies. If a buffer containing an indirect
1391  * block with dependencies is written, these routines will mark it
1392  * dirty again. It can only be successfully written once all the
1393  * dependencies are removed. The ffs_fsync routine in conjunction with
1394  * softdep_sync_metadata work together to get all the dependencies
1395  * removed so that a file can be successfully written to disk. Three
1396  * procedures are used when setting up indirect block pointer
1397  * dependencies. The division is necessary because of the organization
1398  * of the "balloc" routine and because of the distinction between file
1399  * pages and file metadata blocks.
1400  */
1401 
1402 /*
1403  * Allocate a new allocindir structure.
1404  */
1405 static struct allocindir *
1406 newallocindir(ip, ptrno, newblkno, oldblkno)
1407 	struct inode *ip;	/* inode for file being extended */
1408 	int ptrno;		/* offset of pointer in indirect block */
1409 	ufs_daddr_t newblkno;	/* disk block number being added */
1410 	ufs_daddr_t oldblkno;	/* previous block number, 0 if none */
1411 {
1412 	struct allocindir *aip;
1413 
1414 	MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1415 		M_ALLOCINDIR, M_SOFTDEP_FLAGS);
1416 	bzero(aip, sizeof(struct allocindir));
1417 	aip->ai_list.wk_type = D_ALLOCINDIR;
1418 	aip->ai_state = ATTACHED;
1419 	aip->ai_offset = ptrno;
1420 	aip->ai_newblkno = newblkno;
1421 	aip->ai_oldblkno = oldblkno;
1422 	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1423 	return (aip);
1424 }
1425 
1426 /*
1427  * Called just before setting an indirect block pointer
1428  * to a newly allocated file page.
1429  */
1430 void
1431 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1432 	struct inode *ip;	/* inode for file being extended */
1433 	ufs_lbn_t lbn;		/* allocated block number within file */
1434 	struct buf *bp;		/* buffer with indirect blk referencing page */
1435 	int ptrno;		/* offset of pointer in indirect block */
1436 	ufs_daddr_t newblkno;	/* disk block number being added */
1437 	ufs_daddr_t oldblkno;	/* previous block number, 0 if none */
1438 	struct buf *nbp;	/* buffer holding allocated page */
1439 {
1440 	struct allocindir *aip;
1441 	struct pagedep *pagedep;
1442 
1443 	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1444 	ACQUIRE_LOCK(&lk);
1445 	/*
1446 	 * If we are allocating a directory page, then we must
1447 	 * allocate an associated pagedep to track additions and
1448 	 * deletions.
1449 	 */
1450 	if ((ip->i_mode & IFMT) == IFDIR &&
1451 	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1452 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1453 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1454 	FREE_LOCK(&lk);
1455 	setup_allocindir_phase2(bp, ip, aip);
1456 }
1457 
1458 /*
1459  * Called just before setting an indirect block pointer to a
1460  * newly allocated indirect block.
1461  */
1462 void
1463 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1464 	struct buf *nbp;	/* newly allocated indirect block */
1465 	struct inode *ip;	/* inode for file being extended */
1466 	struct buf *bp;		/* indirect block referencing allocated block */
1467 	int ptrno;		/* offset of pointer in indirect block */
1468 	ufs_daddr_t newblkno;	/* disk block number being added */
1469 {
1470 	struct allocindir *aip;
1471 
1472 	aip = newallocindir(ip, ptrno, newblkno, 0);
1473 	ACQUIRE_LOCK(&lk);
1474 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1475 	FREE_LOCK(&lk);
1476 	setup_allocindir_phase2(bp, ip, aip);
1477 }
1478 
1479 /*
1480  * Called to finish the allocation of the "aip" allocated
1481  * by one of the two routines above.
1482  */
1483 static void
1484 setup_allocindir_phase2(bp, ip, aip)
1485 	struct buf *bp;		/* in-memory copy of the indirect block */
1486 	struct inode *ip;	/* inode for file being extended */
1487 	struct allocindir *aip;	/* allocindir allocated by the above routines */
1488 {
1489 	struct worklist *wk;
1490 	struct indirdep *indirdep, *newindirdep;
1491 	struct bmsafemap *bmsafemap;
1492 	struct allocindir *oldaip;
1493 	struct freefrag *freefrag;
1494 	struct newblk *newblk;
1495 
1496 	if (bp->b_lblkno >= 0)
1497 		panic("setup_allocindir_phase2: not indir blk");
1498 	for (indirdep = NULL, newindirdep = NULL; ; ) {
1499 		ACQUIRE_LOCK(&lk);
1500 		for (wk = LIST_FIRST(&bp->b_dep); wk;
1501 		     wk = LIST_NEXT(wk, wk_list)) {
1502 			if (wk->wk_type != D_INDIRDEP)
1503 				continue;
1504 			indirdep = WK_INDIRDEP(wk);
1505 			break;
1506 		}
1507 		if (indirdep == NULL && newindirdep) {
1508 			indirdep = newindirdep;
1509 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1510 			newindirdep = NULL;
1511 		}
1512 		FREE_LOCK(&lk);
1513 		if (indirdep) {
1514 			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1515 			    &newblk) == 0)
1516 				panic("setup_allocindir: lost block");
1517 			ACQUIRE_LOCK(&lk);
1518 			if (newblk->nb_state == DEPCOMPLETE) {
1519 				aip->ai_state |= DEPCOMPLETE;
1520 				aip->ai_buf = NULL;
1521 			} else {
1522 				bmsafemap = newblk->nb_bmsafemap;
1523 				aip->ai_buf = bmsafemap->sm_buf;
1524 				LIST_REMOVE(newblk, nb_deps);
1525 				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1526 				    aip, ai_deps);
1527 			}
1528 			LIST_REMOVE(newblk, nb_hash);
1529 			FREE(newblk, M_NEWBLK);
1530 			aip->ai_indirdep = indirdep;
1531 			/*
1532 			 * Check to see if there is an existing dependency
1533 			 * for this block. If there is, merge the old
1534 			 * dependency into the new one.
1535 			 */
1536 			if (aip->ai_oldblkno == 0)
1537 				oldaip = NULL;
1538 			else
1539 				for (oldaip=LIST_FIRST(&indirdep->ir_deplisthd);
1540 				    oldaip; oldaip = LIST_NEXT(oldaip, ai_next))
1541 					if (oldaip->ai_offset == aip->ai_offset)
1542 						break;
1543 			freefrag = NULL;
1544 			if (oldaip != NULL) {
1545 				if (oldaip->ai_newblkno != aip->ai_oldblkno)
1546 					panic("setup_allocindir_phase2: blkno");
1547 				aip->ai_oldblkno = oldaip->ai_oldblkno;
1548 				freefrag = aip->ai_freefrag;
1549 				aip->ai_freefrag = oldaip->ai_freefrag;
1550 				oldaip->ai_freefrag = NULL;
1551 				free_allocindir(oldaip, NULL);
1552 			}
1553 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1554 			((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1555 			    [aip->ai_offset] = aip->ai_oldblkno;
1556 			FREE_LOCK(&lk);
1557 			if (freefrag != NULL)
1558 				handle_workitem_freefrag(freefrag);
1559 		}
1560 		if (newindirdep) {
1561 			if (indirdep->ir_savebp != NULL)
1562 				brelse(newindirdep->ir_savebp);
1563 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1564 		}
1565 		if (indirdep)
1566 			break;
1567 		MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1568 			M_INDIRDEP, M_SOFTDEP_FLAGS);
1569 		newindirdep->ir_list.wk_type = D_INDIRDEP;
1570 		newindirdep->ir_state = ATTACHED;
1571 		LIST_INIT(&newindirdep->ir_deplisthd);
1572 		LIST_INIT(&newindirdep->ir_donehd);
1573 		if (bp->b_blkno == bp->b_lblkno) {
1574 			VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
1575 				NULL, NULL);
1576 		}
1577 		newindirdep->ir_savebp =
1578 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0);
1579 		BUF_KERNPROC(newindirdep->ir_savebp);
1580 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1581 	}
1582 }
1583 
1584 /*
1585  * Block de-allocation dependencies.
1586  *
1587  * When blocks are de-allocated, the on-disk pointers must be nullified before
1588  * the blocks are made available for use by other files.  (The true
1589  * requirement is that old pointers must be nullified before new on-disk
1590  * pointers are set.  We chose this slightly more stringent requirement to
1591  * reduce complexity.) Our implementation handles this dependency by updating
1592  * the inode (or indirect block) appropriately but delaying the actual block
1593  * de-allocation (i.e., freemap and free space count manipulation) until
1594  * after the updated versions reach stable storage.  After the disk is
1595  * updated, the blocks can be safely de-allocated whenever it is convenient.
1596  * This implementation handles only the common case of reducing a file's
1597  * length to zero. Other cases are handled by the conventional synchronous
1598  * write approach.
1599  *
1600  * The ffs implementation with which we worked double-checks
1601  * the state of the block pointers and file size as it reduces
1602  * a file's length.  Some of this code is replicated here in our
1603  * soft updates implementation.  The freeblks->fb_chkcnt field is
1604  * used to transfer a part of this information to the procedure
1605  * that eventually de-allocates the blocks.
1606  *
1607  * This routine should be called from the routine that shortens
1608  * a file's length, before the inode's size or block pointers
1609  * are modified. It will save the block pointer information for
1610  * later release and zero the inode so that the calling routine
1611  * can release it.
1612  */
1613 void
1614 softdep_setup_freeblocks(ip, length)
1615 	struct inode *ip;	/* The inode whose length is to be reduced */
1616 	off_t length;		/* The new length for the file */
1617 {
1618 	struct freeblks *freeblks;
1619 	struct inodedep *inodedep;
1620 	struct allocdirect *adp;
1621 	struct vnode *vp;
1622 	struct buf *bp;
1623 	struct fs *fs;
1624 	int i, delay, error;
1625 
1626 	fs = ip->i_fs;
1627 	if (length != 0)
1628 		panic("softde_setup_freeblocks: non-zero length");
1629 	MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1630 		M_FREEBLKS, M_SOFTDEP_FLAGS);
1631 	bzero(freeblks, sizeof(struct freeblks));
1632 	freeblks->fb_list.wk_type = D_FREEBLKS;
1633 	freeblks->fb_uid = ip->i_uid;
1634 	freeblks->fb_previousinum = ip->i_number;
1635 	freeblks->fb_devvp = ip->i_devvp;
1636 	freeblks->fb_mnt = ITOV(ip)->v_mount;
1637 	freeblks->fb_oldsize = ip->i_size;
1638 	freeblks->fb_newsize = length;
1639 	freeblks->fb_chkcnt = ip->i_blocks;
1640 	for (i = 0; i < NDADDR; i++) {
1641 		freeblks->fb_dblks[i] = ip->i_db[i];
1642 		ip->i_db[i] = 0;
1643 	}
1644 	for (i = 0; i < NIADDR; i++) {
1645 		freeblks->fb_iblks[i] = ip->i_ib[i];
1646 		ip->i_ib[i] = 0;
1647 	}
1648 	ip->i_blocks = 0;
1649 	ip->i_size = 0;
1650 	/*
1651 	 * Push the zero'ed inode to to its disk buffer so that we are free
1652 	 * to delete its dependencies below. Once the dependencies are gone
1653 	 * the buffer can be safely released.
1654 	 */
1655 	if ((error = bread(ip->i_devvp,
1656 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
1657 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0)
1658 		softdep_error("softdep_setup_freeblocks", error);
1659 	*((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
1660 	    ip->i_din;
1661 	/*
1662 	 * Find and eliminate any inode dependencies.
1663 	 */
1664 	ACQUIRE_LOCK(&lk);
1665 	(void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1666 	if ((inodedep->id_state & IOSTARTED) != 0)
1667 		panic("softdep_setup_freeblocks: inode busy");
1668 	/*
1669 	 * Add the freeblks structure to the list of operations that
1670 	 * must await the zero'ed inode being written to disk. If we
1671 	 * still have a bitmap dependency (delay == 0), then the inode
1672 	 * has never been written to disk, so we can process the
1673 	 * freeblks below once we have deleted the dependencies.
1674 	 */
1675 	delay = (inodedep->id_state & DEPCOMPLETE);
1676 	if (delay)
1677 		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1678 	/*
1679 	 * Because the file length has been truncated to zero, any
1680 	 * pending block allocation dependency structures associated
1681 	 * with this inode are obsolete and can simply be de-allocated.
1682 	 * We must first merge the two dependency lists to get rid of
1683 	 * any duplicate freefrag structures, then purge the merged list.
1684 	 * If we still have a bitmap dependency, then the inode has never
1685 	 * been written to disk, so we can free any fragments without delay.
1686 	 */
1687 	merge_inode_lists(inodedep);
1688 	while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1689 		free_allocdirect(&inodedep->id_inoupdt, adp, delay);
1690 	FREE_LOCK(&lk);
1691 	bdwrite(bp);
1692 	/*
1693 	 * We must wait for any I/O in progress to finish so that
1694 	 * all potential buffers on the dirty list will be visible.
1695 	 * Once they are all there, walk the list and get rid of
1696 	 * any dependencies.
1697 	 */
1698 	vp = ITOV(ip);
1699 	ACQUIRE_LOCK(&lk);
1700 	drain_output(vp, 1);
1701 	while (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) {
1702 		bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
1703 		(void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
1704 		deallocate_dependencies(bp, inodedep);
1705 		bp->b_flags |= B_INVAL | B_NOCACHE;
1706 		FREE_LOCK(&lk);
1707 		brelse(bp);
1708 		ACQUIRE_LOCK(&lk);
1709 	}
1710 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
1711 		(void) free_inodedep(inodedep);
1712 	FREE_LOCK(&lk);
1713 	/*
1714 	 * If the inode has never been written to disk (delay == 0),
1715 	 * then we can process the freeblks now that we have deleted
1716 	 * the dependencies.
1717 	 */
1718 	if (!delay)
1719 		handle_workitem_freeblocks(freeblks);
1720 }
1721 
1722 /*
1723  * Reclaim any dependency structures from a buffer that is about to
1724  * be reallocated to a new vnode. The buffer must be locked, thus,
1725  * no I/O completion operations can occur while we are manipulating
1726  * its associated dependencies. The mutex is held so that other I/O's
1727  * associated with related dependencies do not occur.
1728  */
1729 static void
1730 deallocate_dependencies(bp, inodedep)
1731 	struct buf *bp;
1732 	struct inodedep *inodedep;
1733 {
1734 	struct worklist *wk;
1735 	struct indirdep *indirdep;
1736 	struct allocindir *aip;
1737 	struct pagedep *pagedep;
1738 	struct dirrem *dirrem;
1739 	struct diradd *dap;
1740 	int i;
1741 
1742 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1743 		switch (wk->wk_type) {
1744 
1745 		case D_INDIRDEP:
1746 			indirdep = WK_INDIRDEP(wk);
1747 			/*
1748 			 * None of the indirect pointers will ever be visible,
1749 			 * so they can simply be tossed. GOINGAWAY ensures
1750 			 * that allocated pointers will be saved in the buffer
1751 			 * cache until they are freed. Note that they will
1752 			 * only be able to be found by their physical address
1753 			 * since the inode mapping the logical address will
1754 			 * be gone. The save buffer used for the safe copy
1755 			 * was allocated in setup_allocindir_phase2 using
1756 			 * the physical address so it could be used for this
1757 			 * purpose. Hence we swap the safe copy with the real
1758 			 * copy, allowing the safe copy to be freed and holding
1759 			 * on to the real copy for later use in indir_trunc.
1760 			 */
1761 			if (indirdep->ir_state & GOINGAWAY)
1762 				panic("deallocate_dependencies: already gone");
1763 			indirdep->ir_state |= GOINGAWAY;
1764 			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
1765 				free_allocindir(aip, inodedep);
1766 			if (bp->b_lblkno >= 0 ||
1767 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno)
1768 				panic("deallocate_dependencies: not indir");
1769 			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
1770 			    bp->b_bcount);
1771 			WORKLIST_REMOVE(wk);
1772 			WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
1773 			continue;
1774 
1775 		case D_PAGEDEP:
1776 			pagedep = WK_PAGEDEP(wk);
1777 			/*
1778 			 * None of the directory additions will ever be
1779 			 * visible, so they can simply be tossed.
1780 			 */
1781 			for (i = 0; i < DAHASHSZ; i++)
1782 				while ((dap =
1783 				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
1784 					free_diradd(dap);
1785 			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
1786 				free_diradd(dap);
1787 			/*
1788 			 * Copy any directory remove dependencies to the list
1789 			 * to be processed after the zero'ed inode is written.
1790 			 * If the inode has already been written, then they
1791 			 * can be dumped directly onto the work list.
1792 			 */
1793 			for (dirrem = LIST_FIRST(&pagedep->pd_dirremhd); dirrem;
1794 			     dirrem = LIST_NEXT(dirrem, dm_next)) {
1795 				LIST_REMOVE(dirrem, dm_next);
1796 				dirrem->dm_dirinum = pagedep->pd_ino;
1797 				if (inodedep == NULL ||
1798 				    (inodedep->id_state & ALLCOMPLETE) ==
1799 				     ALLCOMPLETE)
1800 					add_to_worklist(&dirrem->dm_list);
1801 				else
1802 					WORKLIST_INSERT(&inodedep->id_bufwait,
1803 					    &dirrem->dm_list);
1804 			}
1805 			WORKLIST_REMOVE(&pagedep->pd_list);
1806 			LIST_REMOVE(pagedep, pd_hash);
1807 			WORKITEM_FREE(pagedep, D_PAGEDEP);
1808 			continue;
1809 
1810 		case D_ALLOCINDIR:
1811 			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
1812 			continue;
1813 
1814 		case D_ALLOCDIRECT:
1815 		case D_INODEDEP:
1816 			panic("deallocate_dependencies: Unexpected type %s",
1817 			    TYPENAME(wk->wk_type));
1818 			/* NOTREACHED */
1819 
1820 		default:
1821 			panic("deallocate_dependencies: Unknown type %s",
1822 			    TYPENAME(wk->wk_type));
1823 			/* NOTREACHED */
1824 		}
1825 	}
1826 }
1827 
1828 /*
1829  * Free an allocdirect. Generate a new freefrag work request if appropriate.
1830  * This routine must be called with splbio interrupts blocked.
1831  */
1832 static void
1833 free_allocdirect(adphead, adp, delay)
1834 	struct allocdirectlst *adphead;
1835 	struct allocdirect *adp;
1836 	int delay;
1837 {
1838 
1839 #ifdef DEBUG
1840 	if (lk.lkt_held == -1)
1841 		panic("free_allocdirect: lock not held");
1842 #endif
1843 	if ((adp->ad_state & DEPCOMPLETE) == 0)
1844 		LIST_REMOVE(adp, ad_deps);
1845 	TAILQ_REMOVE(adphead, adp, ad_next);
1846 	if ((adp->ad_state & COMPLETE) == 0)
1847 		WORKLIST_REMOVE(&adp->ad_list);
1848 	if (adp->ad_freefrag != NULL) {
1849 		if (delay)
1850 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
1851 			    &adp->ad_freefrag->ff_list);
1852 		else
1853 			add_to_worklist(&adp->ad_freefrag->ff_list);
1854 	}
1855 	WORKITEM_FREE(adp, D_ALLOCDIRECT);
1856 }
1857 
1858 /*
1859  * Prepare an inode to be freed. The actual free operation is not
1860  * done until the zero'ed inode has been written to disk.
1861  */
1862 void
1863 softdep_freefile(pvp, ino, mode)
1864 		struct vnode *pvp;
1865 		ino_t ino;
1866 		int mode;
1867 {
1868 	struct inode *ip = VTOI(pvp);
1869 	struct inodedep *inodedep;
1870 	struct freefile *freefile;
1871 
1872 	/*
1873 	 * This sets up the inode de-allocation dependency.
1874 	 */
1875 	MALLOC(freefile, struct freefile *, sizeof(struct freefile),
1876 		M_FREEFILE, M_SOFTDEP_FLAGS);
1877 	freefile->fx_list.wk_type = D_FREEFILE;
1878 	freefile->fx_list.wk_state = 0;
1879 	freefile->fx_mode = mode;
1880 	freefile->fx_oldinum = ino;
1881 	freefile->fx_devvp = ip->i_devvp;
1882 	freefile->fx_mnt = ITOV(ip)->v_mount;
1883 
1884 	/*
1885 	 * If the inodedep does not exist, then the zero'ed inode has
1886 	 * been written to disk. If the allocated inode has never been
1887 	 * written to disk, then the on-disk inode is zero'ed. In either
1888 	 * case we can free the file immediately.
1889 	 */
1890 	ACQUIRE_LOCK(&lk);
1891 	if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
1892 	    check_inode_unwritten(inodedep)) {
1893 		FREE_LOCK(&lk);
1894 		handle_workitem_freefile(freefile);
1895 		return;
1896 	}
1897 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
1898 	FREE_LOCK(&lk);
1899 }
1900 
1901 /*
1902  * Check to see if an inode has never been written to disk. If
1903  * so free the inodedep and return success, otherwise return failure.
1904  * This routine must be called with splbio interrupts blocked.
1905  *
1906  * If we still have a bitmap dependency, then the inode has never
1907  * been written to disk. Drop the dependency as it is no longer
1908  * necessary since the inode is being deallocated. We set the
1909  * ALLCOMPLETE flags since the bitmap now properly shows that the
1910  * inode is not allocated. Even if the inode is actively being
1911  * written, it has been rolled back to its zero'ed state, so we
1912  * are ensured that a zero inode is what is on the disk. For short
1913  * lived files, this change will usually result in removing all the
1914  * dependencies from the inode so that it can be freed immediately.
1915  */
1916 static int
1917 check_inode_unwritten(inodedep)
1918 	struct inodedep *inodedep;
1919 {
1920 
1921 	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
1922 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
1923 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
1924 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
1925 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
1926 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
1927 	    inodedep->id_nlinkdelta != 0)
1928 		return (0);
1929 	inodedep->id_state |= ALLCOMPLETE;
1930 	LIST_REMOVE(inodedep, id_deps);
1931 	inodedep->id_buf = NULL;
1932 	if (inodedep->id_state & ONWORKLIST)
1933 		WORKLIST_REMOVE(&inodedep->id_list);
1934 	if (inodedep->id_savedino != NULL) {
1935 		FREE(inodedep->id_savedino, M_INODEDEP);
1936 		inodedep->id_savedino = NULL;
1937 	}
1938 	if (free_inodedep(inodedep) == 0)
1939 		panic("check_inode_unwritten: busy inode");
1940 	return (1);
1941 }
1942 
1943 /*
1944  * Try to free an inodedep structure. Return 1 if it could be freed.
1945  */
1946 static int
1947 free_inodedep(inodedep)
1948 	struct inodedep *inodedep;
1949 {
1950 
1951 	if ((inodedep->id_state & ONWORKLIST) != 0 ||
1952 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
1953 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
1954 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
1955 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
1956 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
1957 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
1958 	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
1959 		return (0);
1960 	LIST_REMOVE(inodedep, id_hash);
1961 	WORKITEM_FREE(inodedep, D_INODEDEP);
1962 	num_inodedep -= 1;
1963 	return (1);
1964 }
1965 
1966 /*
1967  * This workitem routine performs the block de-allocation.
1968  * The workitem is added to the pending list after the updated
1969  * inode block has been written to disk.  As mentioned above,
1970  * checks regarding the number of blocks de-allocated (compared
1971  * to the number of blocks allocated for the file) are also
1972  * performed in this function.
1973  */
1974 static void
1975 handle_workitem_freeblocks(freeblks)
1976 	struct freeblks *freeblks;
1977 {
1978 	struct inode tip;
1979 	ufs_daddr_t bn;
1980 	struct fs *fs;
1981 	int i, level, bsize;
1982 	long nblocks, blocksreleased = 0;
1983 	int error, allerror = 0;
1984 	ufs_lbn_t baselbns[NIADDR], tmpval;
1985 
1986 	tip.i_fs = fs = VFSTOUFS(freeblks->fb_mnt)->um_fs;
1987 	tip.i_number = freeblks->fb_previousinum;
1988 	tip.i_devvp = freeblks->fb_devvp;
1989 	tip.i_dev = freeblks->fb_devvp->v_rdev;
1990 	tip.i_size = freeblks->fb_oldsize;
1991 	tip.i_uid = freeblks->fb_uid;
1992 	tip.i_vnode = NULL;
1993 	tmpval = 1;
1994 	baselbns[0] = NDADDR;
1995 	for (i = 1; i < NIADDR; i++) {
1996 		tmpval *= NINDIR(fs);
1997 		baselbns[i] = baselbns[i - 1] + tmpval;
1998 	}
1999 	nblocks = btodb(fs->fs_bsize);
2000 	blocksreleased = 0;
2001 	/*
2002 	 * Indirect blocks first.
2003 	 */
2004 	for (level = (NIADDR - 1); level >= 0; level--) {
2005 		if ((bn = freeblks->fb_iblks[level]) == 0)
2006 			continue;
2007 		if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level,
2008 		    baselbns[level], &blocksreleased)) == 0)
2009 			allerror = error;
2010 		ffs_blkfree(&tip, bn, fs->fs_bsize);
2011 		blocksreleased += nblocks;
2012 	}
2013 	/*
2014 	 * All direct blocks or frags.
2015 	 */
2016 	for (i = (NDADDR - 1); i >= 0; i--) {
2017 		if ((bn = freeblks->fb_dblks[i]) == 0)
2018 			continue;
2019 		bsize = blksize(fs, &tip, i);
2020 		ffs_blkfree(&tip, bn, bsize);
2021 		blocksreleased += btodb(bsize);
2022 	}
2023 
2024 #ifdef DIAGNOSTIC
2025 	if (freeblks->fb_chkcnt != blocksreleased)
2026 		printf("handle_workitem_freeblocks: block count");
2027 	if (allerror)
2028 		softdep_error("handle_workitem_freeblks", allerror);
2029 #endif /* DIAGNOSTIC */
2030 	WORKITEM_FREE(freeblks, D_FREEBLKS);
2031 }
2032 
2033 /*
2034  * Release blocks associated with the inode ip and stored in the indirect
2035  * block dbn. If level is greater than SINGLE, the block is an indirect block
2036  * and recursive calls to indirtrunc must be used to cleanse other indirect
2037  * blocks.
2038  */
2039 static int
2040 indir_trunc(ip, dbn, level, lbn, countp)
2041 	struct inode *ip;
2042 	ufs_daddr_t dbn;
2043 	int level;
2044 	ufs_lbn_t lbn;
2045 	long *countp;
2046 {
2047 	struct buf *bp;
2048 	ufs_daddr_t *bap;
2049 	ufs_daddr_t nb;
2050 	struct fs *fs;
2051 	struct worklist *wk;
2052 	struct indirdep *indirdep;
2053 	int i, lbnadd, nblocks;
2054 	int error, allerror = 0;
2055 
2056 	fs = ip->i_fs;
2057 	lbnadd = 1;
2058 	for (i = level; i > 0; i--)
2059 		lbnadd *= NINDIR(fs);
2060 	/*
2061 	 * Get buffer of block pointers to be freed. This routine is not
2062 	 * called until the zero'ed inode has been written, so it is safe
2063 	 * to free blocks as they are encountered. Because the inode has
2064 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
2065 	 * have to use the on-disk address and the block device for the
2066 	 * filesystem to look them up. If the file was deleted before its
2067 	 * indirect blocks were all written to disk, the routine that set
2068 	 * us up (deallocate_dependencies) will have arranged to leave
2069 	 * a complete copy of the indirect block in memory for our use.
2070 	 * Otherwise we have to read the blocks in from the disk.
2071 	 */
2072 	ACQUIRE_LOCK(&lk);
2073 	if ((bp = incore(ip->i_devvp, dbn)) != NULL &&
2074 	    (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2075 		if (wk->wk_type != D_INDIRDEP ||
2076 		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2077 		    (indirdep->ir_state & GOINGAWAY) == 0)
2078 			panic("indir_trunc: lost indirdep");
2079 		WORKLIST_REMOVE(wk);
2080 		WORKITEM_FREE(indirdep, D_INDIRDEP);
2081 		if (LIST_FIRST(&bp->b_dep) != NULL)
2082 			panic("indir_trunc: dangling dep");
2083 		FREE_LOCK(&lk);
2084 	} else {
2085 		FREE_LOCK(&lk);
2086 		error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, NOCRED, &bp);
2087 		if (error)
2088 			return (error);
2089 	}
2090 	/*
2091 	 * Recursively free indirect blocks.
2092 	 */
2093 	bap = (ufs_daddr_t *)bp->b_data;
2094 	nblocks = btodb(fs->fs_bsize);
2095 	for (i = NINDIR(fs) - 1; i >= 0; i--) {
2096 		if ((nb = bap[i]) == 0)
2097 			continue;
2098 		if (level != 0) {
2099 			if ((error = indir_trunc(ip, fsbtodb(fs, nb),
2100 			     level - 1, lbn + (i * lbnadd), countp)) != 0)
2101 				allerror = error;
2102 		}
2103 		ffs_blkfree(ip, nb, fs->fs_bsize);
2104 		*countp += nblocks;
2105 	}
2106 	bp->b_flags |= B_INVAL | B_NOCACHE;
2107 	brelse(bp);
2108 	return (allerror);
2109 }
2110 
2111 /*
2112  * Free an allocindir.
2113  * This routine must be called with splbio interrupts blocked.
2114  */
2115 static void
2116 free_allocindir(aip, inodedep)
2117 	struct allocindir *aip;
2118 	struct inodedep *inodedep;
2119 {
2120 	struct freefrag *freefrag;
2121 
2122 #ifdef DEBUG
2123 	if (lk.lkt_held == -1)
2124 		panic("free_allocindir: lock not held");
2125 #endif
2126 	if ((aip->ai_state & DEPCOMPLETE) == 0)
2127 		LIST_REMOVE(aip, ai_deps);
2128 	if (aip->ai_state & ONWORKLIST)
2129 		WORKLIST_REMOVE(&aip->ai_list);
2130 	LIST_REMOVE(aip, ai_next);
2131 	if ((freefrag = aip->ai_freefrag) != NULL) {
2132 		if (inodedep == NULL)
2133 			add_to_worklist(&freefrag->ff_list);
2134 		else
2135 			WORKLIST_INSERT(&inodedep->id_bufwait,
2136 			    &freefrag->ff_list);
2137 	}
2138 	WORKITEM_FREE(aip, D_ALLOCINDIR);
2139 }
2140 
2141 /*
2142  * Directory entry addition dependencies.
2143  *
2144  * When adding a new directory entry, the inode (with its incremented link
2145  * count) must be written to disk before the directory entry's pointer to it.
2146  * Also, if the inode is newly allocated, the corresponding freemap must be
2147  * updated (on disk) before the directory entry's pointer. These requirements
2148  * are met via undo/redo on the directory entry's pointer, which consists
2149  * simply of the inode number.
2150  *
2151  * As directory entries are added and deleted, the free space within a
2152  * directory block can become fragmented.  The ufs file system will compact
2153  * a fragmented directory block to make space for a new entry. When this
2154  * occurs, the offsets of previously added entries change. Any "diradd"
2155  * dependency structures corresponding to these entries must be updated with
2156  * the new offsets.
2157  */
2158 
2159 /*
2160  * This routine is called after the in-memory inode's link
2161  * count has been incremented, but before the directory entry's
2162  * pointer to the inode has been set.
2163  */
2164 void
2165 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp)
2166 	struct buf *bp;		/* buffer containing directory block */
2167 	struct inode *dp;	/* inode for directory */
2168 	off_t diroffset;	/* offset of new entry in directory */
2169 	long newinum;		/* inode referenced by new directory entry */
2170 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
2171 {
2172 	int offset;		/* offset of new entry within directory block */
2173 	ufs_lbn_t lbn;		/* block in directory containing new entry */
2174 	struct fs *fs;
2175 	struct diradd *dap;
2176 	struct pagedep *pagedep;
2177 	struct inodedep *inodedep;
2178 	struct mkdir *mkdir1, *mkdir2;
2179 
2180 	/*
2181 	 * Whiteouts have no dependencies.
2182 	 */
2183 	if (newinum == WINO) {
2184 		if (newdirbp != NULL)
2185 			bdwrite(newdirbp);
2186 		return;
2187 	}
2188 
2189 	fs = dp->i_fs;
2190 	lbn = lblkno(fs, diroffset);
2191 	offset = blkoff(fs, diroffset);
2192 	MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD, M_SOFTDEP_FLAGS);
2193 	bzero(dap, sizeof(struct diradd));
2194 	dap->da_list.wk_type = D_DIRADD;
2195 	dap->da_offset = offset;
2196 	dap->da_newinum = newinum;
2197 	dap->da_state = ATTACHED;
2198 	if (newdirbp == NULL) {
2199 		dap->da_state |= DEPCOMPLETE;
2200 		ACQUIRE_LOCK(&lk);
2201 	} else {
2202 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2203 		MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2204 		    M_SOFTDEP_FLAGS);
2205 		mkdir1->md_list.wk_type = D_MKDIR;
2206 		mkdir1->md_state = MKDIR_BODY;
2207 		mkdir1->md_diradd = dap;
2208 		MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2209 		    M_SOFTDEP_FLAGS);
2210 		mkdir2->md_list.wk_type = D_MKDIR;
2211 		mkdir2->md_state = MKDIR_PARENT;
2212 		mkdir2->md_diradd = dap;
2213 		/*
2214 		 * Dependency on "." and ".." being written to disk.
2215 		 */
2216 		mkdir1->md_buf = newdirbp;
2217 		ACQUIRE_LOCK(&lk);
2218 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2219 		WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2220 		FREE_LOCK(&lk);
2221 		bdwrite(newdirbp);
2222 		/*
2223 		 * Dependency on link count increase for parent directory
2224 		 */
2225 		ACQUIRE_LOCK(&lk);
2226 		if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2227 		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2228 			dap->da_state &= ~MKDIR_PARENT;
2229 			WORKITEM_FREE(mkdir2, D_MKDIR);
2230 		} else {
2231 			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2232 			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2233 		}
2234 	}
2235 	/*
2236 	 * Link into parent directory pagedep to await its being written.
2237 	 */
2238 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2239 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2240 	dap->da_pagedep = pagedep;
2241 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2242 	    da_pdlist);
2243 	/*
2244 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
2245 	 * is not yet written. If it is written, do the post-inode write
2246 	 * processing to put it on the id_pendinghd list.
2247 	 */
2248 	(void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2249 	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2250 		diradd_inode_written(dap, inodedep);
2251 	else
2252 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2253 	FREE_LOCK(&lk);
2254 }
2255 
2256 /*
2257  * This procedure is called to change the offset of a directory
2258  * entry when compacting a directory block which must be owned
2259  * exclusively by the caller. Note that the actual entry movement
2260  * must be done in this procedure to ensure that no I/O completions
2261  * occur while the move is in progress.
2262  */
2263 void
2264 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2265 	struct inode *dp;	/* inode for directory */
2266 	caddr_t base;		/* address of dp->i_offset */
2267 	caddr_t oldloc;		/* address of old directory location */
2268 	caddr_t newloc;		/* address of new directory location */
2269 	int entrysize;		/* size of directory entry */
2270 {
2271 	int offset, oldoffset, newoffset;
2272 	struct pagedep *pagedep;
2273 	struct diradd *dap;
2274 	ufs_lbn_t lbn;
2275 
2276 	ACQUIRE_LOCK(&lk);
2277 	lbn = lblkno(dp->i_fs, dp->i_offset);
2278 	offset = blkoff(dp->i_fs, dp->i_offset);
2279 	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2280 		goto done;
2281 	oldoffset = offset + (oldloc - base);
2282 	newoffset = offset + (newloc - base);
2283 	for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(oldoffset)]);
2284 	     dap; dap = LIST_NEXT(dap, da_pdlist)) {
2285 		if (dap->da_offset != oldoffset)
2286 			continue;
2287 		dap->da_offset = newoffset;
2288 		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2289 			break;
2290 		LIST_REMOVE(dap, da_pdlist);
2291 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2292 		    dap, da_pdlist);
2293 		break;
2294 	}
2295 	if (dap == NULL) {
2296 		for (dap = LIST_FIRST(&pagedep->pd_pendinghd);
2297 		     dap; dap = LIST_NEXT(dap, da_pdlist)) {
2298 			if (dap->da_offset == oldoffset) {
2299 				dap->da_offset = newoffset;
2300 				break;
2301 			}
2302 		}
2303 	}
2304 done:
2305 	bcopy(oldloc, newloc, entrysize);
2306 	FREE_LOCK(&lk);
2307 }
2308 
2309 /*
2310  * Free a diradd dependency structure. This routine must be called
2311  * with splbio interrupts blocked.
2312  */
2313 static void
2314 free_diradd(dap)
2315 	struct diradd *dap;
2316 {
2317 	struct dirrem *dirrem;
2318 	struct pagedep *pagedep;
2319 	struct inodedep *inodedep;
2320 	struct mkdir *mkdir, *nextmd;
2321 
2322 #ifdef DEBUG
2323 	if (lk.lkt_held == -1)
2324 		panic("free_diradd: lock not held");
2325 #endif
2326 	WORKLIST_REMOVE(&dap->da_list);
2327 	LIST_REMOVE(dap, da_pdlist);
2328 	if ((dap->da_state & DIRCHG) == 0) {
2329 		pagedep = dap->da_pagedep;
2330 	} else {
2331 		dirrem = dap->da_previous;
2332 		pagedep = dirrem->dm_pagedep;
2333 		dirrem->dm_dirinum = pagedep->pd_ino;
2334 		add_to_worklist(&dirrem->dm_list);
2335 	}
2336 	if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2337 	    0, &inodedep) != 0)
2338 		(void) free_inodedep(inodedep);
2339 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2340 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2341 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
2342 			if (mkdir->md_diradd != dap)
2343 				continue;
2344 			dap->da_state &= ~mkdir->md_state;
2345 			WORKLIST_REMOVE(&mkdir->md_list);
2346 			LIST_REMOVE(mkdir, md_mkdirs);
2347 			WORKITEM_FREE(mkdir, D_MKDIR);
2348 		}
2349 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
2350 			panic("free_diradd: unfound ref");
2351 	}
2352 	WORKITEM_FREE(dap, D_DIRADD);
2353 }
2354 
2355 /*
2356  * Directory entry removal dependencies.
2357  *
2358  * When removing a directory entry, the entry's inode pointer must be
2359  * zero'ed on disk before the corresponding inode's link count is decremented
2360  * (possibly freeing the inode for re-use). This dependency is handled by
2361  * updating the directory entry but delaying the inode count reduction until
2362  * after the directory block has been written to disk. After this point, the
2363  * inode count can be decremented whenever it is convenient.
2364  */
2365 
2366 /*
2367  * This routine should be called immediately after removing
2368  * a directory entry.  The inode's link count should not be
2369  * decremented by the calling procedure -- the soft updates
2370  * code will do this task when it is safe.
2371  */
2372 void
2373 softdep_setup_remove(bp, dp, ip, isrmdir)
2374 	struct buf *bp;		/* buffer containing directory block */
2375 	struct inode *dp;	/* inode for the directory being modified */
2376 	struct inode *ip;	/* inode for directory entry being removed */
2377 	int isrmdir;		/* indicates if doing RMDIR */
2378 {
2379 	struct dirrem *dirrem, *prevdirrem;
2380 
2381 	/*
2382 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2383 	 */
2384 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2385 
2386 	/*
2387 	 * If the COMPLETE flag is clear, then there were no active
2388 	 * entries and we want to roll back to a zeroed entry until
2389 	 * the new inode is committed to disk. If the COMPLETE flag is
2390 	 * set then we have deleted an entry that never made it to
2391 	 * disk. If the entry we deleted resulted from a name change,
2392 	 * then the old name still resides on disk. We cannot delete
2393 	 * its inode (returned to us in prevdirrem) until the zeroed
2394 	 * directory entry gets to disk. The new inode has never been
2395 	 * referenced on the disk, so can be deleted immediately.
2396 	 */
2397 	if ((dirrem->dm_state & COMPLETE) == 0) {
2398 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2399 		    dm_next);
2400 		FREE_LOCK(&lk);
2401 	} else {
2402 		if (prevdirrem != NULL)
2403 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2404 			    prevdirrem, dm_next);
2405 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2406 		FREE_LOCK(&lk);
2407 		handle_workitem_remove(dirrem);
2408 	}
2409 }
2410 
2411 /*
2412  * Allocate a new dirrem if appropriate and return it along with
2413  * its associated pagedep. Called without a lock, returns with lock.
2414  */
2415 static long num_dirrem;		/* number of dirrem allocated */
2416 static struct dirrem *
2417 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2418 	struct buf *bp;		/* buffer containing directory block */
2419 	struct inode *dp;	/* inode for the directory being modified */
2420 	struct inode *ip;	/* inode for directory entry being removed */
2421 	int isrmdir;		/* indicates if doing RMDIR */
2422 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
2423 {
2424 	int offset;
2425 	ufs_lbn_t lbn;
2426 	struct diradd *dap;
2427 	struct dirrem *dirrem;
2428 	struct pagedep *pagedep;
2429 
2430 	/*
2431 	 * Whiteouts have no deletion dependencies.
2432 	 */
2433 	if (ip == NULL)
2434 		panic("newdirrem: whiteout");
2435 	/*
2436 	 * If we are over our limit, try to improve the situation.
2437 	 * Limiting the number of dirrem structures will also limit
2438 	 * the number of freefile and freeblks structures.
2439 	 */
2440 	if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2441 		(void) request_cleanup(FLUSH_REMOVE, 0);
2442 	num_dirrem += 1;
2443 	MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
2444 		M_DIRREM, M_SOFTDEP_FLAGS);
2445 	bzero(dirrem, sizeof(struct dirrem));
2446 	dirrem->dm_list.wk_type = D_DIRREM;
2447 	dirrem->dm_state = isrmdir ? RMDIR : 0;
2448 	dirrem->dm_mnt = ITOV(ip)->v_mount;
2449 	dirrem->dm_oldinum = ip->i_number;
2450 	*prevdirremp = NULL;
2451 
2452 	ACQUIRE_LOCK(&lk);
2453 	lbn = lblkno(dp->i_fs, dp->i_offset);
2454 	offset = blkoff(dp->i_fs, dp->i_offset);
2455 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2456 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2457 	dirrem->dm_pagedep = pagedep;
2458 	/*
2459 	 * Check for a diradd dependency for the same directory entry.
2460 	 * If present, then both dependencies become obsolete and can
2461 	 * be de-allocated. Check for an entry on both the pd_dirraddhd
2462 	 * list and the pd_pendinghd list.
2463 	 */
2464 	for (dap = LIST_FIRST(&pagedep->pd_diraddhd[DIRADDHASH(offset)]);
2465 	     dap; dap = LIST_NEXT(dap, da_pdlist))
2466 		if (dap->da_offset == offset)
2467 			break;
2468 	if (dap == NULL) {
2469 		for (dap = LIST_FIRST(&pagedep->pd_pendinghd);
2470 		     dap; dap = LIST_NEXT(dap, da_pdlist))
2471 			if (dap->da_offset == offset)
2472 				break;
2473 		if (dap == NULL)
2474 			return (dirrem);
2475 	}
2476 	/*
2477 	 * Must be ATTACHED at this point.
2478 	 */
2479 	if ((dap->da_state & ATTACHED) == 0)
2480 		panic("newdirrem: not ATTACHED");
2481 	if (dap->da_newinum != ip->i_number)
2482 		panic("newdirrem: inum %d should be %d",
2483 		    ip->i_number, dap->da_newinum);
2484 	/*
2485 	 * If we are deleting a changed name that never made it to disk,
2486 	 * then return the dirrem describing the previous inode (which
2487 	 * represents the inode currently referenced from this entry on disk).
2488 	 */
2489 	if ((dap->da_state & DIRCHG) != 0) {
2490 		*prevdirremp = dap->da_previous;
2491 		dap->da_state &= ~DIRCHG;
2492 		dap->da_pagedep = pagedep;
2493 	}
2494 	/*
2495 	 * We are deleting an entry that never made it to disk.
2496 	 * Mark it COMPLETE so we can delete its inode immediately.
2497 	 */
2498 	dirrem->dm_state |= COMPLETE;
2499 	free_diradd(dap);
2500 	return (dirrem);
2501 }
2502 
2503 /*
2504  * Directory entry change dependencies.
2505  *
2506  * Changing an existing directory entry requires that an add operation
2507  * be completed first followed by a deletion. The semantics for the addition
2508  * are identical to the description of adding a new entry above except
2509  * that the rollback is to the old inode number rather than zero. Once
2510  * the addition dependency is completed, the removal is done as described
2511  * in the removal routine above.
2512  */
2513 
2514 /*
2515  * This routine should be called immediately after changing
2516  * a directory entry.  The inode's link count should not be
2517  * decremented by the calling procedure -- the soft updates
2518  * code will perform this task when it is safe.
2519  */
2520 void
2521 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
2522 	struct buf *bp;		/* buffer containing directory block */
2523 	struct inode *dp;	/* inode for the directory being modified */
2524 	struct inode *ip;	/* inode for directory entry being removed */
2525 	long newinum;		/* new inode number for changed entry */
2526 	int isrmdir;		/* indicates if doing RMDIR */
2527 {
2528 	int offset;
2529 	struct diradd *dap = NULL;
2530 	struct dirrem *dirrem, *prevdirrem;
2531 	struct pagedep *pagedep;
2532 	struct inodedep *inodedep;
2533 
2534 	offset = blkoff(dp->i_fs, dp->i_offset);
2535 
2536 	/*
2537 	 * Whiteouts do not need diradd dependencies.
2538 	 */
2539 	if (newinum != WINO) {
2540 		MALLOC(dap, struct diradd *, sizeof(struct diradd),
2541 		    M_DIRADD, M_SOFTDEP_FLAGS);
2542 		bzero(dap, sizeof(struct diradd));
2543 		dap->da_list.wk_type = D_DIRADD;
2544 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2545 		dap->da_offset = offset;
2546 		dap->da_newinum = newinum;
2547 	}
2548 
2549 	/*
2550 	 * Allocate a new dirrem and ACQUIRE_LOCK.
2551 	 */
2552 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2553 	pagedep = dirrem->dm_pagedep;
2554 	/*
2555 	 * The possible values for isrmdir:
2556 	 *	0 - non-directory file rename
2557 	 *	1 - directory rename within same directory
2558 	 *   inum - directory rename to new directory of given inode number
2559 	 * When renaming to a new directory, we are both deleting and
2560 	 * creating a new directory entry, so the link count on the new
2561 	 * directory should not change. Thus we do not need the followup
2562 	 * dirrem which is usually done in handle_workitem_remove. We set
2563 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
2564 	 * followup dirrem.
2565 	 */
2566 	if (isrmdir > 1)
2567 		dirrem->dm_state |= DIRCHG;
2568 
2569 	/*
2570 	 * Whiteouts have no additional dependencies,
2571 	 * so just put the dirrem on the correct list.
2572 	 */
2573 	if (newinum == WINO) {
2574 		if ((dirrem->dm_state & COMPLETE) == 0) {
2575 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2576 			    dm_next);
2577 		} else {
2578 			dirrem->dm_dirinum = pagedep->pd_ino;
2579 			add_to_worklist(&dirrem->dm_list);
2580 		}
2581 		FREE_LOCK(&lk);
2582 		return;
2583 	}
2584 
2585 	/*
2586 	 * If the COMPLETE flag is clear, then there were no active
2587 	 * entries and we want to roll back to the previous inode until
2588 	 * the new inode is committed to disk. If the COMPLETE flag is
2589 	 * set, then we have deleted an entry that never made it to disk.
2590 	 * If the entry we deleted resulted from a name change, then the old
2591 	 * inode reference still resides on disk. Any rollback that we do
2592 	 * needs to be to that old inode (returned to us in prevdirrem). If
2593 	 * the entry we deleted resulted from a create, then there is
2594 	 * no entry on the disk, so we want to roll back to zero rather
2595 	 * than the uncommitted inode. In either of the COMPLETE cases we
2596 	 * want to immediately free the unwritten and unreferenced inode.
2597 	 */
2598 	if ((dirrem->dm_state & COMPLETE) == 0) {
2599 		dap->da_previous = dirrem;
2600 	} else {
2601 		if (prevdirrem != NULL) {
2602 			dap->da_previous = prevdirrem;
2603 		} else {
2604 			dap->da_state &= ~DIRCHG;
2605 			dap->da_pagedep = pagedep;
2606 		}
2607 		dirrem->dm_dirinum = pagedep->pd_ino;
2608 		add_to_worklist(&dirrem->dm_list);
2609 	}
2610 	/*
2611 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
2612 	 * is not yet written. If it is written, do the post-inode write
2613 	 * processing to put it on the id_pendinghd list.
2614 	 */
2615 	if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2616 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2617 		dap->da_state |= COMPLETE;
2618 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2619 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2620 	} else {
2621 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2622 		    dap, da_pdlist);
2623 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2624 	}
2625 	FREE_LOCK(&lk);
2626 }
2627 
2628 /*
2629  * Called whenever the link count on an inode is changed.
2630  * It creates an inode dependency so that the new reference(s)
2631  * to the inode cannot be committed to disk until the updated
2632  * inode has been written.
2633  */
2634 void
2635 softdep_change_linkcnt(ip)
2636 	struct inode *ip;	/* the inode with the increased link count */
2637 {
2638 	struct inodedep *inodedep;
2639 
2640 	ACQUIRE_LOCK(&lk);
2641 	(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2642 	if (ip->i_nlink < ip->i_effnlink)
2643 		panic("softdep_change_linkcnt: bad delta");
2644 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2645 	FREE_LOCK(&lk);
2646 }
2647 
2648 /*
2649  * This workitem decrements the inode's link count.
2650  * If the link count reaches zero, the file is removed.
2651  */
2652 static void
2653 handle_workitem_remove(dirrem)
2654 	struct dirrem *dirrem;
2655 {
2656 	struct proc *p = CURPROC;	/* XXX */
2657 	struct inodedep *inodedep;
2658 	struct vnode *vp;
2659 	struct inode *ip;
2660 	ino_t oldinum;
2661 	int error;
2662 
2663 	if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) {
2664 		softdep_error("handle_workitem_remove: vget", error);
2665 		return;
2666 	}
2667 	ip = VTOI(vp);
2668 	ACQUIRE_LOCK(&lk);
2669 	if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0)
2670 		panic("handle_workitem_remove: lost inodedep");
2671 	/*
2672 	 * Normal file deletion.
2673 	 */
2674 	if ((dirrem->dm_state & RMDIR) == 0) {
2675 		ip->i_nlink--;
2676 		ip->i_flag |= IN_CHANGE;
2677 		if (ip->i_nlink < ip->i_effnlink)
2678 			panic("handle_workitem_remove: bad file delta");
2679 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2680 		FREE_LOCK(&lk);
2681 		vput(vp);
2682 		num_dirrem -= 1;
2683 		WORKITEM_FREE(dirrem, D_DIRREM);
2684 		return;
2685 	}
2686 	/*
2687 	 * Directory deletion. Decrement reference count for both the
2688 	 * just deleted parent directory entry and the reference for ".".
2689 	 * Next truncate the directory to length zero. When the
2690 	 * truncation completes, arrange to have the reference count on
2691 	 * the parent decremented to account for the loss of "..".
2692 	 */
2693 	ip->i_nlink -= 2;
2694 	ip->i_flag |= IN_CHANGE;
2695 	if (ip->i_nlink < ip->i_effnlink)
2696 		panic("handle_workitem_remove: bad dir delta");
2697 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2698 	FREE_LOCK(&lk);
2699 	if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, p->p_ucred, p)) != 0)
2700 		softdep_error("handle_workitem_remove: truncate", error);
2701 	/*
2702 	 * Rename a directory to a new parent. Since, we are both deleting
2703 	 * and creating a new directory entry, the link count on the new
2704 	 * directory should not change. Thus we skip the followup dirrem.
2705 	 */
2706 	if (dirrem->dm_state & DIRCHG) {
2707 		vput(vp);
2708 		num_dirrem -= 1;
2709 		WORKITEM_FREE(dirrem, D_DIRREM);
2710 		return;
2711 	}
2712 	/*
2713 	 * If the inodedep does not exist, then the zero'ed inode has
2714 	 * been written to disk. If the allocated inode has never been
2715 	 * written to disk, then the on-disk inode is zero'ed. In either
2716 	 * case we can remove the file immediately.
2717 	 */
2718 	ACQUIRE_LOCK(&lk);
2719 	dirrem->dm_state = 0;
2720 	oldinum = dirrem->dm_oldinum;
2721 	dirrem->dm_oldinum = dirrem->dm_dirinum;
2722 	if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
2723 	    check_inode_unwritten(inodedep)) {
2724 		FREE_LOCK(&lk);
2725 		vput(vp);
2726 		handle_workitem_remove(dirrem);
2727 		return;
2728 	}
2729 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
2730 	FREE_LOCK(&lk);
2731 	vput(vp);
2732 }
2733 
2734 /*
2735  * Inode de-allocation dependencies.
2736  *
2737  * When an inode's link count is reduced to zero, it can be de-allocated. We
2738  * found it convenient to postpone de-allocation until after the inode is
2739  * written to disk with its new link count (zero).  At this point, all of the
2740  * on-disk inode's block pointers are nullified and, with careful dependency
2741  * list ordering, all dependencies related to the inode will be satisfied and
2742  * the corresponding dependency structures de-allocated.  So, if/when the
2743  * inode is reused, there will be no mixing of old dependencies with new
2744  * ones.  This artificial dependency is set up by the block de-allocation
2745  * procedure above (softdep_setup_freeblocks) and completed by the
2746  * following procedure.
2747  */
2748 static void
2749 handle_workitem_freefile(freefile)
2750 	struct freefile *freefile;
2751 {
2752 	struct fs *fs;
2753 	struct vnode vp;
2754 	struct inode tip;
2755 	struct inodedep *idp;
2756 	int error;
2757 
2758 	fs = VFSTOUFS(freefile->fx_mnt)->um_fs;
2759 #ifdef DEBUG
2760 	ACQUIRE_LOCK(&lk);
2761 	if (inodedep_lookup(fs, freefile->fx_oldinum, 0, &idp))
2762 		panic("handle_workitem_freefile: inodedep survived");
2763 	FREE_LOCK(&lk);
2764 #endif
2765 	tip.i_devvp = freefile->fx_devvp;
2766 	tip.i_dev = freefile->fx_devvp->v_rdev;
2767 	tip.i_fs = fs;
2768 	tip.i_vnode = &vp;
2769 	vp.v_data = &tip;
2770 	if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
2771 		softdep_error("handle_workitem_freefile", error);
2772 	WORKITEM_FREE(freefile, D_FREEFILE);
2773 }
2774 
2775 /*
2776  * Disk writes.
2777  *
2778  * The dependency structures constructed above are most actively used when file
2779  * system blocks are written to disk.  No constraints are placed on when a
2780  * block can be written, but unsatisfied update dependencies are made safe by
2781  * modifying (or replacing) the source memory for the duration of the disk
2782  * write.  When the disk write completes, the memory block is again brought
2783  * up-to-date.
2784  *
2785  * In-core inode structure reclamation.
2786  *
2787  * Because there are a finite number of "in-core" inode structures, they are
2788  * reused regularly.  By transferring all inode-related dependencies to the
2789  * in-memory inode block and indexing them separately (via "inodedep"s), we
2790  * can allow "in-core" inode structures to be reused at any time and avoid
2791  * any increase in contention.
2792  *
2793  * Called just before entering the device driver to initiate a new disk I/O.
2794  * The buffer must be locked, thus, no I/O completion operations can occur
2795  * while we are manipulating its associated dependencies.
2796  */
2797 static void
2798 softdep_disk_io_initiation(bp)
2799 	struct buf *bp;		/* structure describing disk write to occur */
2800 {
2801 	struct worklist *wk, *nextwk;
2802 	struct indirdep *indirdep;
2803 
2804 	/*
2805 	 * We only care about write operations. There should never
2806 	 * be dependencies for reads.
2807 	 */
2808 	if (bp->b_iocmd == BIO_READ)
2809 		panic("softdep_disk_io_initiation: read");
2810 	/*
2811 	 * Do any necessary pre-I/O processing.
2812 	 */
2813 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
2814 		nextwk = LIST_NEXT(wk, wk_list);
2815 		switch (wk->wk_type) {
2816 
2817 		case D_PAGEDEP:
2818 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
2819 			continue;
2820 
2821 		case D_INODEDEP:
2822 			initiate_write_inodeblock(WK_INODEDEP(wk), bp);
2823 			continue;
2824 
2825 		case D_INDIRDEP:
2826 			indirdep = WK_INDIRDEP(wk);
2827 			if (indirdep->ir_state & GOINGAWAY)
2828 				panic("disk_io_initiation: indirdep gone");
2829 			/*
2830 			 * If there are no remaining dependencies, this
2831 			 * will be writing the real pointers, so the
2832 			 * dependency can be freed.
2833 			 */
2834 			if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
2835 				indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
2836 				brelse(indirdep->ir_savebp);
2837 				/* inline expand WORKLIST_REMOVE(wk); */
2838 				wk->wk_state &= ~ONWORKLIST;
2839 				LIST_REMOVE(wk, wk_list);
2840 				WORKITEM_FREE(indirdep, D_INDIRDEP);
2841 				continue;
2842 			}
2843 			/*
2844 			 * Replace up-to-date version with safe version.
2845 			 */
2846 			MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
2847 			    M_INDIRDEP, M_SOFTDEP_FLAGS);
2848 			ACQUIRE_LOCK(&lk);
2849 			indirdep->ir_state &= ~ATTACHED;
2850 			indirdep->ir_state |= UNDONE;
2851 			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
2852 			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
2853 			    bp->b_bcount);
2854 			FREE_LOCK(&lk);
2855 			continue;
2856 
2857 		case D_MKDIR:
2858 		case D_BMSAFEMAP:
2859 		case D_ALLOCDIRECT:
2860 		case D_ALLOCINDIR:
2861 			continue;
2862 
2863 		default:
2864 			panic("handle_disk_io_initiation: Unexpected type %s",
2865 			    TYPENAME(wk->wk_type));
2866 			/* NOTREACHED */
2867 		}
2868 	}
2869 }
2870 
2871 /*
2872  * Called from within the procedure above to deal with unsatisfied
2873  * allocation dependencies in a directory. The buffer must be locked,
2874  * thus, no I/O completion operations can occur while we are
2875  * manipulating its associated dependencies.
2876  */
2877 static void
2878 initiate_write_filepage(pagedep, bp)
2879 	struct pagedep *pagedep;
2880 	struct buf *bp;
2881 {
2882 	struct diradd *dap;
2883 	struct direct *ep;
2884 	int i;
2885 
2886 	if (pagedep->pd_state & IOSTARTED) {
2887 		/*
2888 		 * This can only happen if there is a driver that does not
2889 		 * understand chaining. Here biodone will reissue the call
2890 		 * to strategy for the incomplete buffers.
2891 		 */
2892 		printf("initiate_write_filepage: already started\n");
2893 		return;
2894 	}
2895 	pagedep->pd_state |= IOSTARTED;
2896 	ACQUIRE_LOCK(&lk);
2897 	for (i = 0; i < DAHASHSZ; i++) {
2898 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
2899 		     dap = LIST_NEXT(dap, da_pdlist)) {
2900 			ep = (struct direct *)
2901 			    ((char *)bp->b_data + dap->da_offset);
2902 			if (ep->d_ino != dap->da_newinum)
2903 				panic("%s: dir inum %d != new %d",
2904 				    "initiate_write_filepage",
2905 				    ep->d_ino, dap->da_newinum);
2906 			if (dap->da_state & DIRCHG)
2907 				ep->d_ino = dap->da_previous->dm_oldinum;
2908 			else
2909 				ep->d_ino = 0;
2910 			dap->da_state &= ~ATTACHED;
2911 			dap->da_state |= UNDONE;
2912 		}
2913 	}
2914 	FREE_LOCK(&lk);
2915 }
2916 
2917 /*
2918  * Called from within the procedure above to deal with unsatisfied
2919  * allocation dependencies in an inodeblock. The buffer must be
2920  * locked, thus, no I/O completion operations can occur while we
2921  * are manipulating its associated dependencies.
2922  */
2923 static void
2924 initiate_write_inodeblock(inodedep, bp)
2925 	struct inodedep *inodedep;
2926 	struct buf *bp;			/* The inode block */
2927 {
2928 	struct allocdirect *adp, *lastadp;
2929 	struct dinode *dp;
2930 	struct fs *fs;
2931 	ufs_lbn_t prevlbn = 0;
2932 	int i, deplist;
2933 
2934 	if (inodedep->id_state & IOSTARTED)
2935 		panic("initiate_write_inodeblock: already started");
2936 	inodedep->id_state |= IOSTARTED;
2937 	fs = inodedep->id_fs;
2938 	dp = (struct dinode *)bp->b_data +
2939 	    ino_to_fsbo(fs, inodedep->id_ino);
2940 	/*
2941 	 * If the bitmap is not yet written, then the allocated
2942 	 * inode cannot be written to disk.
2943 	 */
2944 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
2945 		if (inodedep->id_savedino != NULL)
2946 			panic("initiate_write_inodeblock: already doing I/O");
2947 		MALLOC(inodedep->id_savedino, struct dinode *,
2948 		    sizeof(struct dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
2949 		*inodedep->id_savedino = *dp;
2950 		bzero((caddr_t)dp, sizeof(struct dinode));
2951 		return;
2952 	}
2953 	/*
2954 	 * If no dependencies, then there is nothing to roll back.
2955 	 */
2956 	inodedep->id_savedsize = dp->di_size;
2957 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
2958 		return;
2959 	/*
2960 	 * Set the dependencies to busy.
2961 	 */
2962 	ACQUIRE_LOCK(&lk);
2963 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
2964 	     adp = TAILQ_NEXT(adp, ad_next)) {
2965 #ifdef DIAGNOSTIC
2966 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
2967 			panic("softdep_write_inodeblock: lbn order");
2968 		prevlbn = adp->ad_lbn;
2969 		if (adp->ad_lbn < NDADDR &&
2970 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno)
2971 			panic("%s: direct pointer #%ld mismatch %d != %d",
2972 			    "softdep_write_inodeblock", adp->ad_lbn,
2973 			    dp->di_db[adp->ad_lbn], adp->ad_newblkno);
2974 		if (adp->ad_lbn >= NDADDR &&
2975 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno)
2976 			panic("%s: indirect pointer #%ld mismatch %d != %d",
2977 			    "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
2978 			    dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
2979 		deplist |= 1 << adp->ad_lbn;
2980 		if ((adp->ad_state & ATTACHED) == 0)
2981 			panic("softdep_write_inodeblock: Unknown state 0x%x",
2982 			    adp->ad_state);
2983 #endif /* DIAGNOSTIC */
2984 		adp->ad_state &= ~ATTACHED;
2985 		adp->ad_state |= UNDONE;
2986 	}
2987 	/*
2988 	 * The on-disk inode cannot claim to be any larger than the last
2989 	 * fragment that has been written. Otherwise, the on-disk inode
2990 	 * might have fragments that were not the last block in the file
2991 	 * which would corrupt the filesystem.
2992 	 */
2993 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
2994 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
2995 		if (adp->ad_lbn >= NDADDR)
2996 			break;
2997 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
2998 		/* keep going until hitting a rollback to a frag */
2999 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3000 			continue;
3001 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3002 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3003 #ifdef DIAGNOSTIC
3004 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
3005 				panic("softdep_write_inodeblock: lost dep1");
3006 #endif /* DIAGNOSTIC */
3007 			dp->di_db[i] = 0;
3008 		}
3009 		for (i = 0; i < NIADDR; i++) {
3010 #ifdef DIAGNOSTIC
3011 			if (dp->di_ib[i] != 0 &&
3012 			    (deplist & ((1 << NDADDR) << i)) == 0)
3013 				panic("softdep_write_inodeblock: lost dep2");
3014 #endif /* DIAGNOSTIC */
3015 			dp->di_ib[i] = 0;
3016 		}
3017 		FREE_LOCK(&lk);
3018 		return;
3019 	}
3020 	/*
3021 	 * If we have zero'ed out the last allocated block of the file,
3022 	 * roll back the size to the last currently allocated block.
3023 	 * We know that this last allocated block is a full-sized as
3024 	 * we already checked for fragments in the loop above.
3025 	 */
3026 	if (lastadp != NULL &&
3027 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3028 		for (i = lastadp->ad_lbn; i >= 0; i--)
3029 			if (dp->di_db[i] != 0)
3030 				break;
3031 		dp->di_size = (i + 1) * fs->fs_bsize;
3032 	}
3033 	/*
3034 	 * The only dependencies are for indirect blocks.
3035 	 *
3036 	 * The file size for indirect block additions is not guaranteed.
3037 	 * Such a guarantee would be non-trivial to achieve. The conventional
3038 	 * synchronous write implementation also does not make this guarantee.
3039 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3040 	 * can be over-estimated without destroying integrity when the file
3041 	 * moves into the indirect blocks (i.e., is large). If we want to
3042 	 * postpone fsck, we are stuck with this argument.
3043 	 */
3044 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3045 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3046 	FREE_LOCK(&lk);
3047 }
3048 
3049 /*
3050  * This routine is called during the completion interrupt
3051  * service routine for a disk write (from the procedure called
3052  * by the device driver to inform the file system caches of
3053  * a request completion).  It should be called early in this
3054  * procedure, before the block is made available to other
3055  * processes or other routines are called.
3056  */
3057 static void
3058 softdep_disk_write_complete(bp)
3059 	struct buf *bp;		/* describes the completed disk write */
3060 {
3061 	struct worklist *wk;
3062 	struct workhead reattach;
3063 	struct newblk *newblk;
3064 	struct allocindir *aip;
3065 	struct allocdirect *adp;
3066 	struct indirdep *indirdep;
3067 	struct inodedep *inodedep;
3068 	struct bmsafemap *bmsafemap;
3069 
3070 #ifdef DEBUG
3071 	if (lk.lkt_held != -1)
3072 		panic("softdep_disk_write_complete: lock is held");
3073 	lk.lkt_held = -2;
3074 #endif
3075 	LIST_INIT(&reattach);
3076 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3077 		WORKLIST_REMOVE(wk);
3078 		switch (wk->wk_type) {
3079 
3080 		case D_PAGEDEP:
3081 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3082 				WORKLIST_INSERT(&reattach, wk);
3083 			continue;
3084 
3085 		case D_INODEDEP:
3086 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3087 				WORKLIST_INSERT(&reattach, wk);
3088 			continue;
3089 
3090 		case D_BMSAFEMAP:
3091 			bmsafemap = WK_BMSAFEMAP(wk);
3092 			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3093 				newblk->nb_state |= DEPCOMPLETE;
3094 				newblk->nb_bmsafemap = NULL;
3095 				LIST_REMOVE(newblk, nb_deps);
3096 			}
3097 			while ((adp =
3098 			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3099 				adp->ad_state |= DEPCOMPLETE;
3100 				adp->ad_buf = NULL;
3101 				LIST_REMOVE(adp, ad_deps);
3102 				handle_allocdirect_partdone(adp);
3103 			}
3104 			while ((aip =
3105 			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3106 				aip->ai_state |= DEPCOMPLETE;
3107 				aip->ai_buf = NULL;
3108 				LIST_REMOVE(aip, ai_deps);
3109 				handle_allocindir_partdone(aip);
3110 			}
3111 			while ((inodedep =
3112 			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3113 				inodedep->id_state |= DEPCOMPLETE;
3114 				LIST_REMOVE(inodedep, id_deps);
3115 				inodedep->id_buf = NULL;
3116 			}
3117 			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3118 			continue;
3119 
3120 		case D_MKDIR:
3121 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3122 			continue;
3123 
3124 		case D_ALLOCDIRECT:
3125 			adp = WK_ALLOCDIRECT(wk);
3126 			adp->ad_state |= COMPLETE;
3127 			handle_allocdirect_partdone(adp);
3128 			continue;
3129 
3130 		case D_ALLOCINDIR:
3131 			aip = WK_ALLOCINDIR(wk);
3132 			aip->ai_state |= COMPLETE;
3133 			handle_allocindir_partdone(aip);
3134 			continue;
3135 
3136 		case D_INDIRDEP:
3137 			indirdep = WK_INDIRDEP(wk);
3138 			if (indirdep->ir_state & GOINGAWAY)
3139 				panic("disk_write_complete: indirdep gone");
3140 			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3141 			FREE(indirdep->ir_saveddata, M_INDIRDEP);
3142 			indirdep->ir_saveddata = 0;
3143 			indirdep->ir_state &= ~UNDONE;
3144 			indirdep->ir_state |= ATTACHED;
3145 			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3146 				handle_allocindir_partdone(aip);
3147 				if (aip == LIST_FIRST(&indirdep->ir_donehd))
3148 					panic("disk_write_complete: not gone");
3149 			}
3150 			WORKLIST_INSERT(&reattach, wk);
3151 			if ((bp->b_flags & B_DELWRI) == 0)
3152 				stat_indir_blk_ptrs++;
3153 			bdirty(bp);
3154 			continue;
3155 
3156 		default:
3157 			panic("handle_disk_write_complete: Unknown type %s",
3158 			    TYPENAME(wk->wk_type));
3159 			/* NOTREACHED */
3160 		}
3161 	}
3162 	/*
3163 	 * Reattach any requests that must be redone.
3164 	 */
3165 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
3166 		WORKLIST_REMOVE(wk);
3167 		WORKLIST_INSERT(&bp->b_dep, wk);
3168 	}
3169 #ifdef DEBUG
3170 	if (lk.lkt_held != -2)
3171 		panic("softdep_disk_write_complete: lock lost");
3172 	lk.lkt_held = -1;
3173 #endif
3174 }
3175 
3176 /*
3177  * Called from within softdep_disk_write_complete above. Note that
3178  * this routine is always called from interrupt level with further
3179  * splbio interrupts blocked.
3180  */
3181 static void
3182 handle_allocdirect_partdone(adp)
3183 	struct allocdirect *adp;	/* the completed allocdirect */
3184 {
3185 	struct allocdirect *listadp;
3186 	struct inodedep *inodedep;
3187 	long bsize, delay;
3188 
3189 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3190 		return;
3191 	if (adp->ad_buf != NULL)
3192 		panic("handle_allocdirect_partdone: dangling dep");
3193 	/*
3194 	 * The on-disk inode cannot claim to be any larger than the last
3195 	 * fragment that has been written. Otherwise, the on-disk inode
3196 	 * might have fragments that were not the last block in the file
3197 	 * which would corrupt the filesystem. Thus, we cannot free any
3198 	 * allocdirects after one whose ad_oldblkno claims a fragment as
3199 	 * these blocks must be rolled back to zero before writing the inode.
3200 	 * We check the currently active set of allocdirects in id_inoupdt.
3201 	 */
3202 	inodedep = adp->ad_inodedep;
3203 	bsize = inodedep->id_fs->fs_bsize;
3204 	for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp;
3205 	     listadp = TAILQ_NEXT(listadp, ad_next)) {
3206 		/* found our block */
3207 		if (listadp == adp)
3208 			break;
3209 		/* continue if ad_oldlbn is not a fragment */
3210 		if (listadp->ad_oldsize == 0 ||
3211 		    listadp->ad_oldsize == bsize)
3212 			continue;
3213 		/* hit a fragment */
3214 		return;
3215 	}
3216 	/*
3217 	 * If we have reached the end of the current list without
3218 	 * finding the just finished dependency, then it must be
3219 	 * on the future dependency list. Future dependencies cannot
3220 	 * be freed until they are moved to the current list.
3221 	 */
3222 	if (listadp == NULL) {
3223 #ifdef DEBUG
3224 		for (listadp = TAILQ_FIRST(&inodedep->id_newinoupdt); listadp;
3225 		     listadp = TAILQ_NEXT(listadp, ad_next))
3226 			/* found our block */
3227 			if (listadp == adp)
3228 				break;
3229 		if (listadp == NULL)
3230 			panic("handle_allocdirect_partdone: lost dep");
3231 #endif /* DEBUG */
3232 		return;
3233 	}
3234 	/*
3235 	 * If we have found the just finished dependency, then free
3236 	 * it along with anything that follows it that is complete.
3237 	 * If the inode still has a bitmap dependency, then it has
3238 	 * never been written to disk, hence the on-disk inode cannot
3239 	 * reference the old fragment so we can free it without delay.
3240 	 */
3241 	delay = (inodedep->id_state & DEPCOMPLETE);
3242 	for (; adp; adp = listadp) {
3243 		listadp = TAILQ_NEXT(adp, ad_next);
3244 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3245 			return;
3246 		free_allocdirect(&inodedep->id_inoupdt, adp, delay);
3247 	}
3248 }
3249 
3250 /*
3251  * Called from within softdep_disk_write_complete above. Note that
3252  * this routine is always called from interrupt level with further
3253  * splbio interrupts blocked.
3254  */
3255 static void
3256 handle_allocindir_partdone(aip)
3257 	struct allocindir *aip;		/* the completed allocindir */
3258 {
3259 	struct indirdep *indirdep;
3260 
3261 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
3262 		return;
3263 	if (aip->ai_buf != NULL)
3264 		panic("handle_allocindir_partdone: dangling dependency");
3265 	indirdep = aip->ai_indirdep;
3266 	if (indirdep->ir_state & UNDONE) {
3267 		LIST_REMOVE(aip, ai_next);
3268 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
3269 		return;
3270 	}
3271 	((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
3272 	    aip->ai_newblkno;
3273 	LIST_REMOVE(aip, ai_next);
3274 	if (aip->ai_freefrag != NULL)
3275 		add_to_worklist(&aip->ai_freefrag->ff_list);
3276 	WORKITEM_FREE(aip, D_ALLOCINDIR);
3277 }
3278 
3279 /*
3280  * Called from within softdep_disk_write_complete above to restore
3281  * in-memory inode block contents to their most up-to-date state. Note
3282  * that this routine is always called from interrupt level with further
3283  * splbio interrupts blocked.
3284  */
3285 static int
3286 handle_written_inodeblock(inodedep, bp)
3287 	struct inodedep *inodedep;
3288 	struct buf *bp;		/* buffer containing the inode block */
3289 {
3290 	struct worklist *wk, *filefree;
3291 	struct allocdirect *adp, *nextadp;
3292 	struct dinode *dp;
3293 	int hadchanges;
3294 
3295 	if ((inodedep->id_state & IOSTARTED) == 0)
3296 		panic("handle_written_inodeblock: not started");
3297 	inodedep->id_state &= ~IOSTARTED;
3298 	inodedep->id_state |= COMPLETE;
3299 	dp = (struct dinode *)bp->b_data +
3300 	    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
3301 	/*
3302 	 * If we had to rollback the inode allocation because of
3303 	 * bitmaps being incomplete, then simply restore it.
3304 	 * Keep the block dirty so that it will not be reclaimed until
3305 	 * all associated dependencies have been cleared and the
3306 	 * corresponding updates written to disk.
3307 	 */
3308 	if (inodedep->id_savedino != NULL) {
3309 		*dp = *inodedep->id_savedino;
3310 		FREE(inodedep->id_savedino, M_INODEDEP);
3311 		inodedep->id_savedino = NULL;
3312 		if ((bp->b_flags & B_DELWRI) == 0)
3313 			stat_inode_bitmap++;
3314 		bdirty(bp);
3315 		return (1);
3316 	}
3317 	/*
3318 	 * Roll forward anything that had to be rolled back before
3319 	 * the inode could be updated.
3320 	 */
3321 	hadchanges = 0;
3322 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
3323 		nextadp = TAILQ_NEXT(adp, ad_next);
3324 		if (adp->ad_state & ATTACHED)
3325 			panic("handle_written_inodeblock: new entry");
3326 		if (adp->ad_lbn < NDADDR) {
3327 			if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno)
3328 				panic("%s: %s #%ld mismatch %d != %d",
3329 				    "handle_written_inodeblock",
3330 				    "direct pointer", adp->ad_lbn,
3331 				    dp->di_db[adp->ad_lbn], adp->ad_oldblkno);
3332 			dp->di_db[adp->ad_lbn] = adp->ad_newblkno;
3333 		} else {
3334 			if (dp->di_ib[adp->ad_lbn - NDADDR] != 0)
3335 				panic("%s: %s #%ld allocated as %d",
3336 				    "handle_written_inodeblock",
3337 				    "indirect pointer", adp->ad_lbn - NDADDR,
3338 				    dp->di_ib[adp->ad_lbn - NDADDR]);
3339 			dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno;
3340 		}
3341 		adp->ad_state &= ~UNDONE;
3342 		adp->ad_state |= ATTACHED;
3343 		hadchanges = 1;
3344 	}
3345 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
3346 		stat_direct_blk_ptrs++;
3347 	/*
3348 	 * Reset the file size to its most up-to-date value.
3349 	 */
3350 	if (inodedep->id_savedsize == -1)
3351 		panic("handle_written_inodeblock: bad size");
3352 	if (dp->di_size != inodedep->id_savedsize) {
3353 		dp->di_size = inodedep->id_savedsize;
3354 		hadchanges = 1;
3355 	}
3356 	inodedep->id_savedsize = -1;
3357 	/*
3358 	 * If there were any rollbacks in the inode block, then it must be
3359 	 * marked dirty so that its will eventually get written back in
3360 	 * its correct form.
3361 	 */
3362 	if (hadchanges)
3363 		bdirty(bp);
3364 	/*
3365 	 * Process any allocdirects that completed during the update.
3366 	 */
3367 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
3368 		handle_allocdirect_partdone(adp);
3369 	/*
3370 	 * Process deallocations that were held pending until the
3371 	 * inode had been written to disk. Freeing of the inode
3372 	 * is delayed until after all blocks have been freed to
3373 	 * avoid creation of new <vfsid, inum, lbn> triples
3374 	 * before the old ones have been deleted.
3375 	 */
3376 	filefree = NULL;
3377 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
3378 		WORKLIST_REMOVE(wk);
3379 		switch (wk->wk_type) {
3380 
3381 		case D_FREEFILE:
3382 			/*
3383 			 * We defer adding filefree to the worklist until
3384 			 * all other additions have been made to ensure
3385 			 * that it will be done after all the old blocks
3386 			 * have been freed.
3387 			 */
3388 			if (filefree != NULL)
3389 				panic("handle_written_inodeblock: filefree");
3390 			filefree = wk;
3391 			continue;
3392 
3393 		case D_MKDIR:
3394 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
3395 			continue;
3396 
3397 		case D_DIRADD:
3398 			diradd_inode_written(WK_DIRADD(wk), inodedep);
3399 			continue;
3400 
3401 		case D_FREEBLKS:
3402 		case D_FREEFRAG:
3403 		case D_DIRREM:
3404 			add_to_worklist(wk);
3405 			continue;
3406 
3407 		default:
3408 			panic("handle_written_inodeblock: Unknown type %s",
3409 			    TYPENAME(wk->wk_type));
3410 			/* NOTREACHED */
3411 		}
3412 	}
3413 	if (filefree != NULL) {
3414 		if (free_inodedep(inodedep) == 0)
3415 			panic("handle_written_inodeblock: live inodedep");
3416 		add_to_worklist(filefree);
3417 		return (0);
3418 	}
3419 
3420 	/*
3421 	 * If no outstanding dependencies, free it.
3422 	 */
3423 	if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0)
3424 		return (0);
3425 	return (hadchanges);
3426 }
3427 
3428 /*
3429  * Process a diradd entry after its dependent inode has been written.
3430  * This routine must be called with splbio interrupts blocked.
3431  */
3432 static void
3433 diradd_inode_written(dap, inodedep)
3434 	struct diradd *dap;
3435 	struct inodedep *inodedep;
3436 {
3437 	struct pagedep *pagedep;
3438 
3439 	dap->da_state |= COMPLETE;
3440 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3441 		if (dap->da_state & DIRCHG)
3442 			pagedep = dap->da_previous->dm_pagedep;
3443 		else
3444 			pagedep = dap->da_pagedep;
3445 		LIST_REMOVE(dap, da_pdlist);
3446 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3447 	}
3448 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3449 }
3450 
3451 /*
3452  * Handle the completion of a mkdir dependency.
3453  */
3454 static void
3455 handle_written_mkdir(mkdir, type)
3456 	struct mkdir *mkdir;
3457 	int type;
3458 {
3459 	struct diradd *dap;
3460 	struct pagedep *pagedep;
3461 
3462 	if (mkdir->md_state != type)
3463 		panic("handle_written_mkdir: bad type");
3464 	dap = mkdir->md_diradd;
3465 	dap->da_state &= ~type;
3466 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
3467 		dap->da_state |= DEPCOMPLETE;
3468 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3469 		if (dap->da_state & DIRCHG)
3470 			pagedep = dap->da_previous->dm_pagedep;
3471 		else
3472 			pagedep = dap->da_pagedep;
3473 		LIST_REMOVE(dap, da_pdlist);
3474 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3475 	}
3476 	LIST_REMOVE(mkdir, md_mkdirs);
3477 	WORKITEM_FREE(mkdir, D_MKDIR);
3478 }
3479 
3480 /*
3481  * Called from within softdep_disk_write_complete above.
3482  * A write operation was just completed. Removed inodes can
3483  * now be freed and associated block pointers may be committed.
3484  * Note that this routine is always called from interrupt level
3485  * with further splbio interrupts blocked.
3486  */
3487 static int
3488 handle_written_filepage(pagedep, bp)
3489 	struct pagedep *pagedep;
3490 	struct buf *bp;		/* buffer containing the written page */
3491 {
3492 	struct dirrem *dirrem;
3493 	struct diradd *dap, *nextdap;
3494 	struct direct *ep;
3495 	int i, chgs;
3496 
3497 	if ((pagedep->pd_state & IOSTARTED) == 0)
3498 		panic("handle_written_filepage: not started");
3499 	pagedep->pd_state &= ~IOSTARTED;
3500 	/*
3501 	 * Process any directory removals that have been committed.
3502 	 */
3503 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
3504 		LIST_REMOVE(dirrem, dm_next);
3505 		dirrem->dm_dirinum = pagedep->pd_ino;
3506 		add_to_worklist(&dirrem->dm_list);
3507 	}
3508 	/*
3509 	 * Free any directory additions that have been committed.
3510 	 */
3511 	while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
3512 		free_diradd(dap);
3513 	/*
3514 	 * Uncommitted directory entries must be restored.
3515 	 */
3516 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
3517 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
3518 		     dap = nextdap) {
3519 			nextdap = LIST_NEXT(dap, da_pdlist);
3520 			if (dap->da_state & ATTACHED)
3521 				panic("handle_written_filepage: attached");
3522 			ep = (struct direct *)
3523 			    ((char *)bp->b_data + dap->da_offset);
3524 			ep->d_ino = dap->da_newinum;
3525 			dap->da_state &= ~UNDONE;
3526 			dap->da_state |= ATTACHED;
3527 			chgs = 1;
3528 			/*
3529 			 * If the inode referenced by the directory has
3530 			 * been written out, then the dependency can be
3531 			 * moved to the pending list.
3532 			 */
3533 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3534 				LIST_REMOVE(dap, da_pdlist);
3535 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
3536 				    da_pdlist);
3537 			}
3538 		}
3539 	}
3540 	/*
3541 	 * If there were any rollbacks in the directory, then it must be
3542 	 * marked dirty so that its will eventually get written back in
3543 	 * its correct form.
3544 	 */
3545 	if (chgs) {
3546 		if ((bp->b_flags & B_DELWRI) == 0)
3547 			stat_dir_entry++;
3548 		bdirty(bp);
3549 	}
3550 	/*
3551 	 * If no dependencies remain, the pagedep will be freed.
3552 	 * Otherwise it will remain to update the page before it
3553 	 * is written back to disk.
3554 	 */
3555 	if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) {
3556 		for (i = 0; i < DAHASHSZ; i++)
3557 			if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
3558 				break;
3559 		if (i == DAHASHSZ) {
3560 			LIST_REMOVE(pagedep, pd_hash);
3561 			WORKITEM_FREE(pagedep, D_PAGEDEP);
3562 			return (0);
3563 		}
3564 	}
3565 	return (1);
3566 }
3567 
3568 /*
3569  * Writing back in-core inode structures.
3570  *
3571  * The file system only accesses an inode's contents when it occupies an
3572  * "in-core" inode structure.  These "in-core" structures are separate from
3573  * the page frames used to cache inode blocks.  Only the latter are
3574  * transferred to/from the disk.  So, when the updated contents of the
3575  * "in-core" inode structure are copied to the corresponding in-memory inode
3576  * block, the dependencies are also transferred.  The following procedure is
3577  * called when copying a dirty "in-core" inode to a cached inode block.
3578  */
3579 
3580 /*
3581  * Called when an inode is loaded from disk. If the effective link count
3582  * differed from the actual link count when it was last flushed, then we
3583  * need to ensure that the correct effective link count is put back.
3584  */
3585 void
3586 softdep_load_inodeblock(ip)
3587 	struct inode *ip;	/* the "in_core" copy of the inode */
3588 {
3589 	struct inodedep *inodedep;
3590 
3591 	/*
3592 	 * Check for alternate nlink count.
3593 	 */
3594 	ip->i_effnlink = ip->i_nlink;
3595 	ACQUIRE_LOCK(&lk);
3596 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3597 		FREE_LOCK(&lk);
3598 		return;
3599 	}
3600 	ip->i_effnlink -= inodedep->id_nlinkdelta;
3601 	FREE_LOCK(&lk);
3602 }
3603 
3604 /*
3605  * This routine is called just before the "in-core" inode
3606  * information is to be copied to the in-memory inode block.
3607  * Recall that an inode block contains several inodes. If
3608  * the force flag is set, then the dependencies will be
3609  * cleared so that the update can always be made. Note that
3610  * the buffer is locked when this routine is called, so we
3611  * will never be in the middle of writing the inode block
3612  * to disk.
3613  */
3614 void
3615 softdep_update_inodeblock(ip, bp, waitfor)
3616 	struct inode *ip;	/* the "in_core" copy of the inode */
3617 	struct buf *bp;		/* the buffer containing the inode block */
3618 	int waitfor;		/* nonzero => update must be allowed */
3619 {
3620 	struct inodedep *inodedep;
3621 	struct worklist *wk;
3622 	int error, gotit;
3623 
3624 	/*
3625 	 * If the effective link count is not equal to the actual link
3626 	 * count, then we must track the difference in an inodedep while
3627 	 * the inode is (potentially) tossed out of the cache. Otherwise,
3628 	 * if there is no existing inodedep, then there are no dependencies
3629 	 * to track.
3630 	 */
3631 	ACQUIRE_LOCK(&lk);
3632 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3633 		if (ip->i_effnlink != ip->i_nlink)
3634 			panic("softdep_update_inodeblock: bad link count");
3635 		FREE_LOCK(&lk);
3636 		return;
3637 	}
3638 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
3639 		panic("softdep_update_inodeblock: bad delta");
3640 	/*
3641 	 * Changes have been initiated. Anything depending on these
3642 	 * changes cannot occur until this inode has been written.
3643 	 */
3644 	inodedep->id_state &= ~COMPLETE;
3645 	if ((inodedep->id_state & ONWORKLIST) == 0)
3646 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
3647 	/*
3648 	 * Any new dependencies associated with the incore inode must
3649 	 * now be moved to the list associated with the buffer holding
3650 	 * the in-memory copy of the inode. Once merged process any
3651 	 * allocdirects that are completed by the merger.
3652 	 */
3653 	merge_inode_lists(inodedep);
3654 	if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
3655 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
3656 	/*
3657 	 * Now that the inode has been pushed into the buffer, the
3658 	 * operations dependent on the inode being written to disk
3659 	 * can be moved to the id_bufwait so that they will be
3660 	 * processed when the buffer I/O completes.
3661 	 */
3662 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
3663 		WORKLIST_REMOVE(wk);
3664 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
3665 	}
3666 	/*
3667 	 * Newly allocated inodes cannot be written until the bitmap
3668 	 * that allocates them have been written (indicated by
3669 	 * DEPCOMPLETE being set in id_state). If we are doing a
3670 	 * forced sync (e.g., an fsync on a file), we force the bitmap
3671 	 * to be written so that the update can be done.
3672 	 */
3673 	if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
3674 		FREE_LOCK(&lk);
3675 		return;
3676 	}
3677 	gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
3678 	FREE_LOCK(&lk);
3679 	if (gotit &&
3680 	    (error = BUF_WRITE(inodedep->id_buf)) != 0)
3681 		softdep_error("softdep_update_inodeblock: bwrite", error);
3682 	if ((inodedep->id_state & DEPCOMPLETE) == 0)
3683 		panic("softdep_update_inodeblock: update failed");
3684 }
3685 
3686 /*
3687  * Merge the new inode dependency list (id_newinoupdt) into the old
3688  * inode dependency list (id_inoupdt). This routine must be called
3689  * with splbio interrupts blocked.
3690  */
3691 static void
3692 merge_inode_lists(inodedep)
3693 	struct inodedep *inodedep;
3694 {
3695 	struct allocdirect *listadp, *newadp;
3696 
3697 	newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3698 	for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) {
3699 		if (listadp->ad_lbn < newadp->ad_lbn) {
3700 			listadp = TAILQ_NEXT(listadp, ad_next);
3701 			continue;
3702 		}
3703 		TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3704 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
3705 		if (listadp->ad_lbn == newadp->ad_lbn) {
3706 			allocdirect_merge(&inodedep->id_inoupdt, newadp,
3707 			    listadp);
3708 			listadp = newadp;
3709 		}
3710 		newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3711 	}
3712 	while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) {
3713 		TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3714 		TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next);
3715 	}
3716 }
3717 
3718 /*
3719  * If we are doing an fsync, then we must ensure that any directory
3720  * entries for the inode have been written after the inode gets to disk.
3721  */
3722 int
3723 softdep_fsync(vp)
3724 	struct vnode *vp;	/* the "in_core" copy of the inode */
3725 {
3726 	struct inodedep *inodedep;
3727 	struct pagedep *pagedep;
3728 	struct worklist *wk;
3729 	struct diradd *dap;
3730 	struct mount *mnt;
3731 	struct vnode *pvp;
3732 	struct inode *ip;
3733 	struct buf *bp;
3734 	struct fs *fs;
3735 	struct proc *p = CURPROC;		/* XXX */
3736 	int error, flushparent;
3737 	ino_t parentino;
3738 	ufs_lbn_t lbn;
3739 
3740 	ip = VTOI(vp);
3741 	fs = ip->i_fs;
3742 	ACQUIRE_LOCK(&lk);
3743 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
3744 		FREE_LOCK(&lk);
3745 		return (0);
3746 	}
3747 	if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
3748 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
3749 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
3750 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL)
3751 		panic("softdep_fsync: pending ops");
3752 	for (error = 0, flushparent = 0; ; ) {
3753 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
3754 			break;
3755 		if (wk->wk_type != D_DIRADD)
3756 			panic("softdep_fsync: Unexpected type %s",
3757 			    TYPENAME(wk->wk_type));
3758 		dap = WK_DIRADD(wk);
3759 		/*
3760 		 * Flush our parent if this directory entry
3761 		 * has a MKDIR_PARENT dependency.
3762 		 */
3763 		if (dap->da_state & DIRCHG)
3764 			pagedep = dap->da_previous->dm_pagedep;
3765 		else
3766 			pagedep = dap->da_pagedep;
3767 		mnt = pagedep->pd_mnt;
3768 		parentino = pagedep->pd_ino;
3769 		lbn = pagedep->pd_lbn;
3770 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
3771 			panic("softdep_fsync: dirty");
3772 		flushparent = dap->da_state & MKDIR_PARENT;
3773 		/*
3774 		 * If we are being fsync'ed as part of vgone'ing this vnode,
3775 		 * then we will not be able to release and recover the
3776 		 * vnode below, so we just have to give up on writing its
3777 		 * directory entry out. It will eventually be written, just
3778 		 * not now, but then the user was not asking to have it
3779 		 * written, so we are not breaking any promises.
3780 		 */
3781 		if (vp->v_flag & VXLOCK)
3782 			break;
3783 		/*
3784 		 * We prevent deadlock by always fetching inodes from the
3785 		 * root, moving down the directory tree. Thus, when fetching
3786 		 * our parent directory, we must unlock ourselves before
3787 		 * requesting the lock on our parent. See the comment in
3788 		 * ufs_lookup for details on possible races.
3789 		 */
3790 		FREE_LOCK(&lk);
3791 		VOP_UNLOCK(vp, 0, p);
3792 		error = VFS_VGET(mnt, parentino, &pvp);
3793 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p);
3794 		if (error != 0)
3795 			return (error);
3796 		if (flushparent) {
3797 			if ((error = UFS_UPDATE(pvp, 1)) != 0) {
3798 				vput(pvp);
3799 				return (error);
3800 			}
3801 		}
3802 		/*
3803 		 * Flush directory page containing the inode's name.
3804 		 */
3805 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), p->p_ucred,
3806 		    &bp);
3807 		if (error == 0)
3808 			error = BUF_WRITE(bp);
3809 		vput(pvp);
3810 		if (error != 0)
3811 			return (error);
3812 		ACQUIRE_LOCK(&lk);
3813 		if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
3814 			break;
3815 	}
3816 	FREE_LOCK(&lk);
3817 	return (0);
3818 }
3819 
3820 /*
3821  * Flush all the dirty bitmaps associated with the block device
3822  * before flushing the rest of the dirty blocks so as to reduce
3823  * the number of dependencies that will have to be rolled back.
3824  */
3825 void
3826 softdep_fsync_mountdev(vp)
3827 	struct vnode *vp;
3828 {
3829 	struct buf *bp, *nbp;
3830 	struct worklist *wk;
3831 
3832 	if (!vn_isdisk(vp, NULL))
3833 		panic("softdep_fsync_mountdev: vnode not a disk");
3834 	ACQUIRE_LOCK(&lk);
3835 	for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
3836 		nbp = TAILQ_NEXT(bp, b_vnbufs);
3837 		/*
3838 		 * If it is already scheduled, skip to the next buffer.
3839 		 */
3840 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
3841 			continue;
3842 		if ((bp->b_flags & B_DELWRI) == 0)
3843 			panic("softdep_fsync_mountdev: not dirty");
3844 		/*
3845 		 * We are only interested in bitmaps with outstanding
3846 		 * dependencies.
3847 		 */
3848 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
3849 		    wk->wk_type != D_BMSAFEMAP ||
3850 		    (bp->b_xflags & BX_BKGRDINPROG)) {
3851 			BUF_UNLOCK(bp);
3852 			continue;
3853 		}
3854 		bremfree(bp);
3855 		FREE_LOCK(&lk);
3856 		(void) bawrite(bp);
3857 		ACQUIRE_LOCK(&lk);
3858 		/*
3859 		 * Since we may have slept during the I/O, we need
3860 		 * to start from a known point.
3861 		 */
3862 		nbp = TAILQ_FIRST(&vp->v_dirtyblkhd);
3863 	}
3864 	drain_output(vp, 1);
3865 	FREE_LOCK(&lk);
3866 }
3867 
3868 /*
3869  * This routine is called when we are trying to synchronously flush a
3870  * file. This routine must eliminate any filesystem metadata dependencies
3871  * so that the syncing routine can succeed by pushing the dirty blocks
3872  * associated with the file. If any I/O errors occur, they are returned.
3873  */
3874 int
3875 softdep_sync_metadata(ap)
3876 	struct vop_fsync_args /* {
3877 		struct vnode *a_vp;
3878 		struct ucred *a_cred;
3879 		int a_waitfor;
3880 		struct proc *a_p;
3881 	} */ *ap;
3882 {
3883 	struct vnode *vp = ap->a_vp;
3884 	struct pagedep *pagedep;
3885 	struct allocdirect *adp;
3886 	struct allocindir *aip;
3887 	struct buf *bp, *nbp;
3888 	struct worklist *wk;
3889 	int i, error, waitfor;
3890 
3891 	/*
3892 	 * Check whether this vnode is involved in a filesystem
3893 	 * that is doing soft dependency processing.
3894 	 */
3895 	if (!vn_isdisk(vp, NULL)) {
3896 		if (!DOINGSOFTDEP(vp))
3897 			return (0);
3898 	} else
3899 		if (vp->v_rdev->si_mountpoint == NULL ||
3900 		    (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0)
3901 			return (0);
3902 	/*
3903 	 * Ensure that any direct block dependencies have been cleared.
3904 	 */
3905 	ACQUIRE_LOCK(&lk);
3906 	if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
3907 		FREE_LOCK(&lk);
3908 		return (error);
3909 	}
3910 	/*
3911 	 * For most files, the only metadata dependencies are the
3912 	 * cylinder group maps that allocate their inode or blocks.
3913 	 * The block allocation dependencies can be found by traversing
3914 	 * the dependency lists for any buffers that remain on their
3915 	 * dirty buffer list. The inode allocation dependency will
3916 	 * be resolved when the inode is updated with MNT_WAIT.
3917 	 * This work is done in two passes. The first pass grabs most
3918 	 * of the buffers and begins asynchronously writing them. The
3919 	 * only way to wait for these asynchronous writes is to sleep
3920 	 * on the filesystem vnode which may stay busy for a long time
3921 	 * if the filesystem is active. So, instead, we make a second
3922 	 * pass over the dependencies blocking on each write. In the
3923 	 * usual case we will be blocking against a write that we
3924 	 * initiated, so when it is done the dependency will have been
3925 	 * resolved. Thus the second pass is expected to end quickly.
3926 	 */
3927 	waitfor = MNT_NOWAIT;
3928 top:
3929 	if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) {
3930 		FREE_LOCK(&lk);
3931 		return (0);
3932 	}
3933 	bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
3934 loop:
3935 	/*
3936 	 * As we hold the buffer locked, none of its dependencies
3937 	 * will disappear.
3938 	 */
3939 	for (wk = LIST_FIRST(&bp->b_dep); wk;
3940 	     wk = LIST_NEXT(wk, wk_list)) {
3941 		switch (wk->wk_type) {
3942 
3943 		case D_ALLOCDIRECT:
3944 			adp = WK_ALLOCDIRECT(wk);
3945 			if (adp->ad_state & DEPCOMPLETE)
3946 				break;
3947 			nbp = adp->ad_buf;
3948 			if (getdirtybuf(&nbp, waitfor) == 0)
3949 				break;
3950 			FREE_LOCK(&lk);
3951 			if (waitfor == MNT_NOWAIT) {
3952 				bawrite(nbp);
3953 			} else if ((error = BUF_WRITE(nbp)) != 0) {
3954 				bawrite(bp);
3955 				return (error);
3956 			}
3957 			ACQUIRE_LOCK(&lk);
3958 			break;
3959 
3960 		case D_ALLOCINDIR:
3961 			aip = WK_ALLOCINDIR(wk);
3962 			if (aip->ai_state & DEPCOMPLETE)
3963 				break;
3964 			nbp = aip->ai_buf;
3965 			if (getdirtybuf(&nbp, waitfor) == 0)
3966 				break;
3967 			FREE_LOCK(&lk);
3968 			if (waitfor == MNT_NOWAIT) {
3969 				bawrite(nbp);
3970 			} else if ((error = BUF_WRITE(nbp)) != 0) {
3971 				bawrite(bp);
3972 				return (error);
3973 			}
3974 			ACQUIRE_LOCK(&lk);
3975 			break;
3976 
3977 		case D_INDIRDEP:
3978 		restart:
3979 			for (aip = LIST_FIRST(&WK_INDIRDEP(wk)->ir_deplisthd);
3980 			     aip; aip = LIST_NEXT(aip, ai_next)) {
3981 				if (aip->ai_state & DEPCOMPLETE)
3982 					continue;
3983 				nbp = aip->ai_buf;
3984 				if (getdirtybuf(&nbp, MNT_WAIT) == 0)
3985 					goto restart;
3986 				FREE_LOCK(&lk);
3987 				if ((error = BUF_WRITE(nbp)) != 0) {
3988 					bawrite(bp);
3989 					return (error);
3990 				}
3991 				ACQUIRE_LOCK(&lk);
3992 				goto restart;
3993 			}
3994 			break;
3995 
3996 		case D_INODEDEP:
3997 			if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
3998 			    WK_INODEDEP(wk)->id_ino)) != 0) {
3999 				FREE_LOCK(&lk);
4000 				bawrite(bp);
4001 				return (error);
4002 			}
4003 			break;
4004 
4005 		case D_PAGEDEP:
4006 			/*
4007 			 * We are trying to sync a directory that may
4008 			 * have dependencies on both its own metadata
4009 			 * and/or dependencies on the inodes of any
4010 			 * recently allocated files. We walk its diradd
4011 			 * lists pushing out the associated inode.
4012 			 */
4013 			pagedep = WK_PAGEDEP(wk);
4014 			for (i = 0; i < DAHASHSZ; i++) {
4015 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
4016 					continue;
4017 				if ((error =
4018 				    flush_pagedep_deps(vp, pagedep->pd_mnt,
4019 						&pagedep->pd_diraddhd[i]))) {
4020 					FREE_LOCK(&lk);
4021 					bawrite(bp);
4022 					return (error);
4023 				}
4024 			}
4025 			break;
4026 
4027 		case D_MKDIR:
4028 			/*
4029 			 * This case should never happen if the vnode has
4030 			 * been properly sync'ed. However, if this function
4031 			 * is used at a place where the vnode has not yet
4032 			 * been sync'ed, this dependency can show up. So,
4033 			 * rather than panic, just flush it.
4034 			 */
4035 			nbp = WK_MKDIR(wk)->md_buf;
4036 			if (getdirtybuf(&nbp, waitfor) == 0)
4037 				break;
4038 			FREE_LOCK(&lk);
4039 			if (waitfor == MNT_NOWAIT) {
4040 				bawrite(nbp);
4041 			} else if ((error = BUF_WRITE(nbp)) != 0) {
4042 				bawrite(bp);
4043 				return (error);
4044 			}
4045 			ACQUIRE_LOCK(&lk);
4046 			break;
4047 
4048 		case D_BMSAFEMAP:
4049 			/*
4050 			 * This case should never happen if the vnode has
4051 			 * been properly sync'ed. However, if this function
4052 			 * is used at a place where the vnode has not yet
4053 			 * been sync'ed, this dependency can show up. So,
4054 			 * rather than panic, just flush it.
4055 			 */
4056 			nbp = WK_BMSAFEMAP(wk)->sm_buf;
4057 			if (getdirtybuf(&nbp, waitfor) == 0)
4058 				break;
4059 			FREE_LOCK(&lk);
4060 			if (waitfor == MNT_NOWAIT) {
4061 				bawrite(nbp);
4062 			} else if ((error = BUF_WRITE(nbp)) != 0) {
4063 				bawrite(bp);
4064 				return (error);
4065 			}
4066 			ACQUIRE_LOCK(&lk);
4067 			break;
4068 
4069 		default:
4070 			panic("softdep_sync_metadata: Unknown type %s",
4071 			    TYPENAME(wk->wk_type));
4072 			/* NOTREACHED */
4073 		}
4074 	}
4075 	(void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT);
4076 	nbp = TAILQ_NEXT(bp, b_vnbufs);
4077 	FREE_LOCK(&lk);
4078 	bawrite(bp);
4079 	ACQUIRE_LOCK(&lk);
4080 	if (nbp != NULL) {
4081 		bp = nbp;
4082 		goto loop;
4083 	}
4084 	/*
4085 	 * We must wait for any I/O in progress to finish so that
4086 	 * all potential buffers on the dirty list will be visible.
4087 	 * Once they are all there, proceed with the second pass
4088 	 * which will wait for the I/O as per above.
4089 	 */
4090 	drain_output(vp, 1);
4091 	/*
4092 	 * The brief unlock is to allow any pent up dependency
4093 	 * processing to be done.
4094 	 */
4095 	if (waitfor == MNT_NOWAIT) {
4096 		waitfor = MNT_WAIT;
4097 		FREE_LOCK(&lk);
4098 		ACQUIRE_LOCK(&lk);
4099 		goto top;
4100 	}
4101 
4102 	/*
4103 	 * If we have managed to get rid of all the dirty buffers,
4104 	 * then we are done. For certain directories and block
4105 	 * devices, we may need to do further work.
4106 	 */
4107 	if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) {
4108 		FREE_LOCK(&lk);
4109 		return (0);
4110 	}
4111 
4112 	FREE_LOCK(&lk);
4113 	/*
4114 	 * If we are trying to sync a block device, some of its buffers may
4115 	 * contain metadata that cannot be written until the contents of some
4116 	 * partially written files have been written to disk. The only easy
4117 	 * way to accomplish this is to sync the entire filesystem (luckily
4118 	 * this happens rarely).
4119 	 */
4120 	if (vn_isdisk(vp, NULL) &&
4121 	    vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) &&
4122 	    (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, ap->a_cred,
4123 	     ap->a_p)) != 0)
4124 		return (error);
4125 	return (0);
4126 }
4127 
4128 /*
4129  * Flush the dependencies associated with an inodedep.
4130  * Called with splbio blocked.
4131  */
4132 static int
4133 flush_inodedep_deps(fs, ino)
4134 	struct fs *fs;
4135 	ino_t ino;
4136 {
4137 	struct inodedep *inodedep;
4138 	struct allocdirect *adp;
4139 	int error, waitfor;
4140 	struct buf *bp;
4141 
4142 	/*
4143 	 * This work is done in two passes. The first pass grabs most
4144 	 * of the buffers and begins asynchronously writing them. The
4145 	 * only way to wait for these asynchronous writes is to sleep
4146 	 * on the filesystem vnode which may stay busy for a long time
4147 	 * if the filesystem is active. So, instead, we make a second
4148 	 * pass over the dependencies blocking on each write. In the
4149 	 * usual case we will be blocking against a write that we
4150 	 * initiated, so when it is done the dependency will have been
4151 	 * resolved. Thus the second pass is expected to end quickly.
4152 	 * We give a brief window at the top of the loop to allow
4153 	 * any pending I/O to complete.
4154 	 */
4155 	for (waitfor = MNT_NOWAIT; ; ) {
4156 		FREE_LOCK(&lk);
4157 		ACQUIRE_LOCK(&lk);
4158 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4159 			return (0);
4160 		for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
4161 		     adp = TAILQ_NEXT(adp, ad_next)) {
4162 			if (adp->ad_state & DEPCOMPLETE)
4163 				continue;
4164 			bp = adp->ad_buf;
4165 			if (getdirtybuf(&bp, waitfor) == 0) {
4166 				if (waitfor == MNT_NOWAIT)
4167 					continue;
4168 				break;
4169 			}
4170 			FREE_LOCK(&lk);
4171 			if (waitfor == MNT_NOWAIT) {
4172 				bawrite(bp);
4173 			} else if ((error = BUF_WRITE(bp)) != 0) {
4174 				ACQUIRE_LOCK(&lk);
4175 				return (error);
4176 			}
4177 			ACQUIRE_LOCK(&lk);
4178 			break;
4179 		}
4180 		if (adp != NULL)
4181 			continue;
4182 		for (adp = TAILQ_FIRST(&inodedep->id_newinoupdt); adp;
4183 		     adp = TAILQ_NEXT(adp, ad_next)) {
4184 			if (adp->ad_state & DEPCOMPLETE)
4185 				continue;
4186 			bp = adp->ad_buf;
4187 			if (getdirtybuf(&bp, waitfor) == 0) {
4188 				if (waitfor == MNT_NOWAIT)
4189 					continue;
4190 				break;
4191 			}
4192 			FREE_LOCK(&lk);
4193 			if (waitfor == MNT_NOWAIT) {
4194 				bawrite(bp);
4195 			} else if ((error = BUF_WRITE(bp)) != 0) {
4196 				ACQUIRE_LOCK(&lk);
4197 				return (error);
4198 			}
4199 			ACQUIRE_LOCK(&lk);
4200 			break;
4201 		}
4202 		if (adp != NULL)
4203 			continue;
4204 		/*
4205 		 * If pass2, we are done, otherwise do pass 2.
4206 		 */
4207 		if (waitfor == MNT_WAIT)
4208 			break;
4209 		waitfor = MNT_WAIT;
4210 	}
4211 	/*
4212 	 * Try freeing inodedep in case all dependencies have been removed.
4213 	 */
4214 	if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
4215 		(void) free_inodedep(inodedep);
4216 	return (0);
4217 }
4218 
4219 /*
4220  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
4221  * Called with splbio blocked.
4222  */
4223 static int
4224 flush_pagedep_deps(pvp, mp, diraddhdp)
4225 	struct vnode *pvp;
4226 	struct mount *mp;
4227 	struct diraddhd *diraddhdp;
4228 {
4229 	struct proc *p = CURPROC;	/* XXX */
4230 	struct inodedep *inodedep;
4231 	struct ufsmount *ump;
4232 	struct diradd *dap;
4233 	struct vnode *vp;
4234 	int gotit, error = 0;
4235 	struct buf *bp;
4236 	ino_t inum;
4237 
4238 	ump = VFSTOUFS(mp);
4239 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
4240 		/*
4241 		 * Flush ourselves if this directory entry
4242 		 * has a MKDIR_PARENT dependency.
4243 		 */
4244 		if (dap->da_state & MKDIR_PARENT) {
4245 			FREE_LOCK(&lk);
4246 			if ((error = UFS_UPDATE(pvp, 1)) != 0)
4247 				break;
4248 			ACQUIRE_LOCK(&lk);
4249 			/*
4250 			 * If that cleared dependencies, go on to next.
4251 			 */
4252 			if (dap != LIST_FIRST(diraddhdp))
4253 				continue;
4254 			if (dap->da_state & MKDIR_PARENT)
4255 				panic("flush_pagedep_deps: MKDIR_PARENT");
4256 		}
4257 		/*
4258 		 * A newly allocated directory must have its "." and
4259 		 * ".." entries written out before its name can be
4260 		 * committed in its parent. We do not want or need
4261 		 * the full semantics of a synchronous VOP_FSYNC as
4262 		 * that may end up here again, once for each directory
4263 		 * level in the filesystem. Instead, we push the blocks
4264 		 * and wait for them to clear. We have to fsync twice
4265 		 * because the first call may choose to defer blocks
4266 		 * that still have dependencies, but deferral will
4267 		 * happen at most once.
4268 		 */
4269 		inum = dap->da_newinum;
4270 		if (dap->da_state & MKDIR_BODY) {
4271 			FREE_LOCK(&lk);
4272 			if ((error = VFS_VGET(mp, inum, &vp)) != 0)
4273 				break;
4274 			if ((error=VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)) ||
4275 			    (error=VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) {
4276 				vput(vp);
4277 				break;
4278 			}
4279 			drain_output(vp, 0);
4280 			vput(vp);
4281 			ACQUIRE_LOCK(&lk);
4282 			/*
4283 			 * If that cleared dependencies, go on to next.
4284 			 */
4285 			if (dap != LIST_FIRST(diraddhdp))
4286 				continue;
4287 			if (dap->da_state & MKDIR_BODY)
4288 				panic("flush_pagedep_deps: MKDIR_BODY");
4289 		}
4290 		/*
4291 		 * Flush the inode on which the directory entry depends.
4292 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
4293 		 * the only remaining dependency is that the updated inode
4294 		 * count must get pushed to disk. The inode has already
4295 		 * been pushed into its inode buffer (via VOP_UPDATE) at
4296 		 * the time of the reference count change. So we need only
4297 		 * locate that buffer, ensure that there will be no rollback
4298 		 * caused by a bitmap dependency, then write the inode buffer.
4299 		 */
4300 		if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0)
4301 			panic("flush_pagedep_deps: lost inode");
4302 		/*
4303 		 * If the inode still has bitmap dependencies,
4304 		 * push them to disk.
4305 		 */
4306 		if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4307 			gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4308 			FREE_LOCK(&lk);
4309 			if (gotit &&
4310 			    (error = BUF_WRITE(inodedep->id_buf)) != 0)
4311 				break;
4312 			ACQUIRE_LOCK(&lk);
4313 			if (dap != LIST_FIRST(diraddhdp))
4314 				continue;
4315 		}
4316 		/*
4317 		 * If the inode is still sitting in a buffer waiting
4318 		 * to be written, push it to disk.
4319 		 */
4320 		FREE_LOCK(&lk);
4321 		if ((error = bread(ump->um_devvp,
4322 		    fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
4323 		    (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0)
4324 			break;
4325 		if ((error = BUF_WRITE(bp)) != 0)
4326 			break;
4327 		ACQUIRE_LOCK(&lk);
4328 		/*
4329 		 * If we have failed to get rid of all the dependencies
4330 		 * then something is seriously wrong.
4331 		 */
4332 		if (dap == LIST_FIRST(diraddhdp))
4333 			panic("flush_pagedep_deps: flush failed");
4334 	}
4335 	if (error)
4336 		ACQUIRE_LOCK(&lk);
4337 	return (error);
4338 }
4339 
4340 /*
4341  * A large burst of file addition or deletion activity can drive the
4342  * memory load excessively high. Therefore we deliberately slow things
4343  * down and speed up the I/O processing if we find ourselves with too
4344  * many dependencies in progress.
4345  */
4346 static int
4347 request_cleanup(resource, islocked)
4348 	int resource;
4349 	int islocked;
4350 {
4351 	struct proc *p = CURPROC;
4352 
4353 	/*
4354 	 * We never hold up the filesystem syncer process.
4355 	 */
4356 	if (p == filesys_syncer)
4357 		return (0);
4358 	/*
4359 	 * If we are resource constrained on inode dependencies, try
4360 	 * flushing some dirty inodes. Otherwise, we are constrained
4361 	 * by file deletions, so try accelerating flushes of directories
4362 	 * with removal dependencies. We would like to do the cleanup
4363 	 * here, but we probably hold an inode locked at this point and
4364 	 * that might deadlock against one that we try to clean. So,
4365 	 * the best that we can do is request the syncer daemon to do
4366 	 * the cleanup for us.
4367 	 */
4368 	switch (resource) {
4369 
4370 	case FLUSH_INODES:
4371 		stat_ino_limit_push += 1;
4372 		req_clear_inodedeps += 1;
4373 		stat_countp = &stat_ino_limit_hit;
4374 		break;
4375 
4376 	case FLUSH_REMOVE:
4377 		stat_blk_limit_push += 1;
4378 		req_clear_remove += 1;
4379 		stat_countp = &stat_blk_limit_hit;
4380 		break;
4381 
4382 	default:
4383 		panic("request_cleanup: unknown type");
4384 	}
4385 	/*
4386 	 * Hopefully the syncer daemon will catch up and awaken us.
4387 	 * We wait at most tickdelay before proceeding in any case.
4388 	 */
4389 	if (islocked == 0)
4390 		ACQUIRE_LOCK(&lk);
4391 	if (proc_waiting++ == 0) {
4392 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
4393 	}
4394 	FREE_LOCK_INTERLOCKED(&lk);
4395 	(void) tsleep((caddr_t)&proc_waiting, PPAUSE, "softupdate", 0);
4396 	ACQUIRE_LOCK_INTERLOCKED(&lk);
4397 	if (--proc_waiting == 0)
4398 		untimeout(pause_timer, 0, handle);
4399 	if (islocked == 0)
4400 		FREE_LOCK(&lk);
4401 	return (1);
4402 }
4403 
4404 /*
4405  * Awaken processes pausing in request_cleanup and clear proc_waiting
4406  * to indicate that there is no longer a timer running.
4407  */
4408 void
4409 pause_timer(arg)
4410 	void *arg;
4411 {
4412 
4413 	*stat_countp += 1;
4414 	handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
4415 	wakeup_one(&proc_waiting);
4416 }
4417 
4418 /*
4419  * Flush out a directory with at least one removal dependency in an effort to
4420  * reduce the number of dirrem, freefile, and freeblks dependency structures.
4421  */
4422 static void
4423 clear_remove(p)
4424 	struct proc *p;
4425 {
4426 	struct pagedep_hashhead *pagedephd;
4427 	struct pagedep *pagedep;
4428 	static int next = 0;
4429 	struct mount *mp;
4430 	struct vnode *vp;
4431 	int error, cnt;
4432 	ino_t ino;
4433 
4434 	ACQUIRE_LOCK(&lk);
4435 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
4436 		pagedephd = &pagedep_hashtbl[next++];
4437 		if (next >= pagedep_hash)
4438 			next = 0;
4439 		for (pagedep = LIST_FIRST(pagedephd); pagedep;
4440 		     pagedep = LIST_NEXT(pagedep, pd_hash)) {
4441 			if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
4442 				continue;
4443 			mp = pagedep->pd_mnt;
4444 			ino = pagedep->pd_ino;
4445 			FREE_LOCK(&lk);
4446 			if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
4447 				continue;
4448 			if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4449 				softdep_error("clear_remove: vget", error);
4450 				vn_finished_write(mp);
4451 				return;
4452 			}
4453 			if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)))
4454 				softdep_error("clear_remove: fsync", error);
4455 			drain_output(vp, 0);
4456 			vput(vp);
4457 			vn_finished_write(mp);
4458 			return;
4459 		}
4460 	}
4461 	FREE_LOCK(&lk);
4462 }
4463 
4464 /*
4465  * Clear out a block of dirty inodes in an effort to reduce
4466  * the number of inodedep dependency structures.
4467  */
4468 static void
4469 clear_inodedeps(p)
4470 	struct proc *p;
4471 {
4472 	struct inodedep_hashhead *inodedephd;
4473 	struct inodedep *inodedep;
4474 	static int next = 0;
4475 	struct mount *mp;
4476 	struct vnode *vp;
4477 	struct fs *fs;
4478 	int error, cnt;
4479 	ino_t firstino, lastino, ino;
4480 
4481 	ACQUIRE_LOCK(&lk);
4482 	/*
4483 	 * Pick a random inode dependency to be cleared.
4484 	 * We will then gather up all the inodes in its block
4485 	 * that have dependencies and flush them out.
4486 	 */
4487 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
4488 		inodedephd = &inodedep_hashtbl[next++];
4489 		if (next >= inodedep_hash)
4490 			next = 0;
4491 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
4492 			break;
4493 	}
4494 	/*
4495 	 * Ugly code to find mount point given pointer to superblock.
4496 	 */
4497 	fs = inodedep->id_fs;
4498 	TAILQ_FOREACH(mp, &mountlist, mnt_list)
4499 		if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs)
4500 			break;
4501 	/*
4502 	 * Find the last inode in the block with dependencies.
4503 	 */
4504 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
4505 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
4506 		if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
4507 			break;
4508 	/*
4509 	 * Asynchronously push all but the last inode with dependencies.
4510 	 * Synchronously push the last inode with dependencies to ensure
4511 	 * that the inode block gets written to free up the inodedeps.
4512 	 */
4513 	for (ino = firstino; ino <= lastino; ino++) {
4514 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4515 			continue;
4516 		FREE_LOCK(&lk);
4517 		if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
4518 			continue;
4519 		if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4520 			softdep_error("clear_inodedeps: vget", error);
4521 			vn_finished_write(mp);
4522 			return;
4523 		}
4524 		if (ino == lastino) {
4525 			if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_WAIT, p)))
4526 				softdep_error("clear_inodedeps: fsync1", error);
4527 		} else {
4528 			if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)))
4529 				softdep_error("clear_inodedeps: fsync2", error);
4530 			drain_output(vp, 0);
4531 		}
4532 		vput(vp);
4533 		vn_finished_write(mp);
4534 		ACQUIRE_LOCK(&lk);
4535 	}
4536 	FREE_LOCK(&lk);
4537 }
4538 
4539 /*
4540  * Function to determine if the buffer has outstanding dependencies
4541  * that will cause a roll-back if the buffer is written. If wantcount
4542  * is set, return number of dependencies, otherwise just yes or no.
4543  */
4544 static int
4545 softdep_count_dependencies(bp, wantcount)
4546 	struct buf *bp;
4547 	int wantcount;
4548 {
4549 	struct worklist *wk;
4550 	struct inodedep *inodedep;
4551 	struct indirdep *indirdep;
4552 	struct allocindir *aip;
4553 	struct pagedep *pagedep;
4554 	struct diradd *dap;
4555 	int i, retval;
4556 
4557 	retval = 0;
4558 	ACQUIRE_LOCK(&lk);
4559 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = LIST_NEXT(wk, wk_list)) {
4560 		switch (wk->wk_type) {
4561 
4562 		case D_INODEDEP:
4563 			inodedep = WK_INODEDEP(wk);
4564 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4565 				/* bitmap allocation dependency */
4566 				retval += 1;
4567 				if (!wantcount)
4568 					goto out;
4569 			}
4570 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
4571 				/* direct block pointer dependency */
4572 				retval += 1;
4573 				if (!wantcount)
4574 					goto out;
4575 			}
4576 			continue;
4577 
4578 		case D_INDIRDEP:
4579 			indirdep = WK_INDIRDEP(wk);
4580 			for (aip = LIST_FIRST(&indirdep->ir_deplisthd);
4581 			     aip; aip = LIST_NEXT(aip, ai_next)) {
4582 				/* indirect block pointer dependency */
4583 				retval += 1;
4584 				if (!wantcount)
4585 					goto out;
4586 			}
4587 			continue;
4588 
4589 		case D_PAGEDEP:
4590 			pagedep = WK_PAGEDEP(wk);
4591 			for (i = 0; i < DAHASHSZ; i++) {
4592 				for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]);
4593 				     dap; dap = LIST_NEXT(dap, da_pdlist)) {
4594 					/* directory entry dependency */
4595 					retval += 1;
4596 					if (!wantcount)
4597 						goto out;
4598 				}
4599 			}
4600 			continue;
4601 
4602 		case D_BMSAFEMAP:
4603 		case D_ALLOCDIRECT:
4604 		case D_ALLOCINDIR:
4605 		case D_MKDIR:
4606 			/* never a dependency on these blocks */
4607 			continue;
4608 
4609 		default:
4610 			panic("softdep_check_for_rollback: Unexpected type %s",
4611 			    TYPENAME(wk->wk_type));
4612 			/* NOTREACHED */
4613 		}
4614 	}
4615 out:
4616 	FREE_LOCK(&lk);
4617 	return retval;
4618 }
4619 
4620 /*
4621  * Acquire exclusive access to a buffer.
4622  * Must be called with splbio blocked.
4623  * Return 1 if buffer was acquired.
4624  */
4625 static int
4626 getdirtybuf(bpp, waitfor)
4627 	struct buf **bpp;
4628 	int waitfor;
4629 {
4630 	struct buf *bp;
4631 
4632 	for (;;) {
4633 		if ((bp = *bpp) == NULL)
4634 			return (0);
4635 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
4636 			if ((bp->b_xflags & BX_BKGRDINPROG) == 0)
4637 				break;
4638 			BUF_UNLOCK(bp);
4639 			if (waitfor != MNT_WAIT)
4640 				return (0);
4641 			bp->b_xflags |= BX_BKGRDWAIT;
4642 			FREE_LOCK_INTERLOCKED(&lk);
4643 			tsleep(&bp->b_xflags, PRIBIO, "getbuf", 0);
4644 			ACQUIRE_LOCK_INTERLOCKED(&lk);
4645 			continue;
4646 		}
4647 		if (waitfor != MNT_WAIT)
4648 			return (0);
4649 		FREE_LOCK_INTERLOCKED(&lk);
4650 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL) != ENOLCK)
4651 			panic("getdirtybuf: inconsistent lock");
4652 		ACQUIRE_LOCK_INTERLOCKED(&lk);
4653 	}
4654 	if ((bp->b_flags & B_DELWRI) == 0) {
4655 		BUF_UNLOCK(bp);
4656 		return (0);
4657 	}
4658 	bremfree(bp);
4659 	return (1);
4660 }
4661 
4662 /*
4663  * Wait for pending output on a vnode to complete.
4664  * Must be called with vnode locked.
4665  */
4666 static void
4667 drain_output(vp, islocked)
4668 	struct vnode *vp;
4669 	int islocked;
4670 {
4671 
4672 	if (!islocked)
4673 		ACQUIRE_LOCK(&lk);
4674 	while (vp->v_numoutput) {
4675 		vp->v_flag |= VBWAIT;
4676 		FREE_LOCK_INTERLOCKED(&lk);
4677 		tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, "drainvp", 0);
4678 		ACQUIRE_LOCK_INTERLOCKED(&lk);
4679 	}
4680 	if (!islocked)
4681 		FREE_LOCK(&lk);
4682 }
4683 
4684 /*
4685  * Called whenever a buffer that is being invalidated or reallocated
4686  * contains dependencies. This should only happen if an I/O error has
4687  * occurred. The routine is called with the buffer locked.
4688  */
4689 static void
4690 softdep_deallocate_dependencies(bp)
4691 	struct buf *bp;
4692 {
4693 
4694 	if ((bp->b_ioflags & BIO_ERROR) == 0)
4695 		panic("softdep_deallocate_dependencies: dangling deps");
4696 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
4697 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
4698 }
4699 
4700 /*
4701  * Function to handle asynchronous write errors in the filesystem.
4702  */
4703 void
4704 softdep_error(func, error)
4705 	char *func;
4706 	int error;
4707 {
4708 
4709 	/* XXX should do something better! */
4710 	printf("%s: got error %d while accessing filesystem\n", func, error);
4711 }
4712