xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision 830940567b49bb0c08dfaed40418999e76616909)
1 /*-
2  * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
3  *
4  * The soft updates code is derived from the appendix of a University
5  * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
6  * "Soft Updates: A Solution to the Metadata Update Problem in File
7  * Systems", CSE-TR-254-95, August 1995).
8  *
9  * Further information about soft updates can be obtained from:
10  *
11  *	Marshall Kirk McKusick		http://www.mckusick.com/softdep/
12  *	1614 Oxford Street		mckusick@mckusick.com
13  *	Berkeley, CA 94709-1608		+1-510-843-9542
14  *	USA
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  *
20  * 1. Redistributions of source code must retain the above copyright
21  *    notice, this list of conditions and the following disclaimer.
22  * 2. Redistributions in binary form must reproduce the above copyright
23  *    notice, this list of conditions and the following disclaimer in the
24  *    documentation and/or other materials provided with the distribution.
25  *
26  * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
27  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
28  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
29  * DISCLAIMED.  IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
30  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)ffs_softdep.c	9.59 (McKusick) 6/21/00
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_ffs.h"
45 #include "opt_ddb.h"
46 
47 /*
48  * For now we want the safety net that the DEBUG flag provides.
49  */
50 #ifndef DEBUG
51 #define DEBUG
52 #endif
53 
54 #include <sys/param.h>
55 #include <sys/kernel.h>
56 #include <sys/systm.h>
57 #include <sys/bio.h>
58 #include <sys/buf.h>
59 #include <sys/kdb.h>
60 #include <sys/kthread.h>
61 #include <sys/lock.h>
62 #include <sys/malloc.h>
63 #include <sys/mount.h>
64 #include <sys/mutex.h>
65 #include <sys/proc.h>
66 #include <sys/stat.h>
67 #include <sys/sysctl.h>
68 #include <sys/syslog.h>
69 #include <sys/vnode.h>
70 #include <sys/conf.h>
71 #include <ufs/ufs/dir.h>
72 #include <ufs/ufs/extattr.h>
73 #include <ufs/ufs/quota.h>
74 #include <ufs/ufs/inode.h>
75 #include <ufs/ufs/ufsmount.h>
76 #include <ufs/ffs/fs.h>
77 #include <ufs/ffs/softdep.h>
78 #include <ufs/ffs/ffs_extern.h>
79 #include <ufs/ufs/ufs_extern.h>
80 
81 #include <vm/vm.h>
82 
83 #include <ddb/ddb.h>
84 
85 #ifndef SOFTUPDATES
86 
87 int
88 softdep_flushfiles(oldmnt, flags, td)
89 	struct mount *oldmnt;
90 	int flags;
91 	struct thread *td;
92 {
93 
94 	panic("softdep_flushfiles called");
95 }
96 
97 int
98 softdep_mount(devvp, mp, fs, cred)
99 	struct vnode *devvp;
100 	struct mount *mp;
101 	struct fs *fs;
102 	struct ucred *cred;
103 {
104 
105 	return (0);
106 }
107 
108 void
109 softdep_initialize()
110 {
111 
112 	return;
113 }
114 
115 void
116 softdep_uninitialize()
117 {
118 
119 	return;
120 }
121 
122 void
123 softdep_setup_inomapdep(bp, ip, newinum)
124 	struct buf *bp;
125 	struct inode *ip;
126 	ino_t newinum;
127 {
128 
129 	panic("softdep_setup_inomapdep called");
130 }
131 
132 void
133 softdep_setup_blkmapdep(bp, mp, newblkno)
134 	struct buf *bp;
135 	struct mount *mp;
136 	ufs2_daddr_t newblkno;
137 {
138 
139 	panic("softdep_setup_blkmapdep called");
140 }
141 
142 void
143 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
144 	struct inode *ip;
145 	ufs_lbn_t lbn;
146 	ufs2_daddr_t newblkno;
147 	ufs2_daddr_t oldblkno;
148 	long newsize;
149 	long oldsize;
150 	struct buf *bp;
151 {
152 
153 	panic("softdep_setup_allocdirect called");
154 }
155 
156 void
157 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
158 	struct inode *ip;
159 	ufs_lbn_t lbn;
160 	ufs2_daddr_t newblkno;
161 	ufs2_daddr_t oldblkno;
162 	long newsize;
163 	long oldsize;
164 	struct buf *bp;
165 {
166 
167 	panic("softdep_setup_allocext called");
168 }
169 
170 void
171 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
172 	struct inode *ip;
173 	ufs_lbn_t lbn;
174 	struct buf *bp;
175 	int ptrno;
176 	ufs2_daddr_t newblkno;
177 	ufs2_daddr_t oldblkno;
178 	struct buf *nbp;
179 {
180 
181 	panic("softdep_setup_allocindir_page called");
182 }
183 
184 void
185 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
186 	struct buf *nbp;
187 	struct inode *ip;
188 	struct buf *bp;
189 	int ptrno;
190 	ufs2_daddr_t newblkno;
191 {
192 
193 	panic("softdep_setup_allocindir_meta called");
194 }
195 
196 void
197 softdep_setup_freeblocks(ip, length, flags)
198 	struct inode *ip;
199 	off_t length;
200 	int flags;
201 {
202 
203 	panic("softdep_setup_freeblocks called");
204 }
205 
206 void
207 softdep_freefile(pvp, ino, mode)
208 		struct vnode *pvp;
209 		ino_t ino;
210 		int mode;
211 {
212 
213 	panic("softdep_freefile called");
214 }
215 
216 int
217 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
218 	struct buf *bp;
219 	struct inode *dp;
220 	off_t diroffset;
221 	ino_t newinum;
222 	struct buf *newdirbp;
223 	int isnewblk;
224 {
225 
226 	panic("softdep_setup_directory_add called");
227 }
228 
229 void
230 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
231 	struct inode *dp;
232 	caddr_t base;
233 	caddr_t oldloc;
234 	caddr_t newloc;
235 	int entrysize;
236 {
237 
238 	panic("softdep_change_directoryentry_offset called");
239 }
240 
241 void
242 softdep_setup_remove(bp, dp, ip, isrmdir)
243 	struct buf *bp;
244 	struct inode *dp;
245 	struct inode *ip;
246 	int isrmdir;
247 {
248 
249 	panic("softdep_setup_remove called");
250 }
251 
252 void
253 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
254 	struct buf *bp;
255 	struct inode *dp;
256 	struct inode *ip;
257 	ino_t newinum;
258 	int isrmdir;
259 {
260 
261 	panic("softdep_setup_directory_change called");
262 }
263 
264 void
265 softdep_change_linkcnt(ip)
266 	struct inode *ip;
267 {
268 
269 	panic("softdep_change_linkcnt called");
270 }
271 
272 void
273 softdep_load_inodeblock(ip)
274 	struct inode *ip;
275 {
276 
277 	panic("softdep_load_inodeblock called");
278 }
279 
280 void
281 softdep_update_inodeblock(ip, bp, waitfor)
282 	struct inode *ip;
283 	struct buf *bp;
284 	int waitfor;
285 {
286 
287 	panic("softdep_update_inodeblock called");
288 }
289 
290 int
291 softdep_fsync(vp)
292 	struct vnode *vp;	/* the "in_core" copy of the inode */
293 {
294 
295 	return (0);
296 }
297 
298 void
299 softdep_fsync_mountdev(vp)
300 	struct vnode *vp;
301 {
302 
303 	return;
304 }
305 
306 int
307 softdep_flushworklist(oldmnt, countp, td)
308 	struct mount *oldmnt;
309 	int *countp;
310 	struct thread *td;
311 {
312 
313 	*countp = 0;
314 	return (0);
315 }
316 
317 int
318 softdep_sync_metadata(struct vnode *vp)
319 {
320 
321 	return (0);
322 }
323 
324 int
325 softdep_slowdown(vp)
326 	struct vnode *vp;
327 {
328 
329 	panic("softdep_slowdown called");
330 }
331 
332 void
333 softdep_releasefile(ip)
334 	struct inode *ip;	/* inode with the zero effective link count */
335 {
336 
337 	panic("softdep_releasefile called");
338 }
339 
340 int
341 softdep_request_cleanup(fs, vp)
342 	struct fs *fs;
343 	struct vnode *vp;
344 {
345 
346 	return (0);
347 }
348 
349 int
350 softdep_check_suspend(struct mount *mp,
351 		      struct vnode *devvp,
352 		      int softdep_deps,
353 		      int softdep_accdeps,
354 		      int secondary_writes,
355 		      int secondary_accwrites)
356 {
357 	struct bufobj *bo;
358 	int error;
359 
360 	(void) softdep_deps,
361 	(void) softdep_accdeps;
362 
363 	bo = &devvp->v_bufobj;
364 	ASSERT_BO_LOCKED(bo);
365 
366 	MNT_ILOCK(mp);
367 	while (mp->mnt_secondary_writes != 0) {
368 		BO_UNLOCK(bo);
369 		msleep(&mp->mnt_secondary_writes, MNT_MTX(mp),
370 		    (PUSER - 1) | PDROP, "secwr", 0);
371 		BO_LOCK(bo);
372 		MNT_ILOCK(mp);
373 	}
374 
375 	/*
376 	 * Reasons for needing more work before suspend:
377 	 * - Dirty buffers on devvp.
378 	 * - Secondary writes occurred after start of vnode sync loop
379 	 */
380 	error = 0;
381 	if (bo->bo_numoutput > 0 ||
382 	    bo->bo_dirty.bv_cnt > 0 ||
383 	    secondary_writes != 0 ||
384 	    mp->mnt_secondary_writes != 0 ||
385 	    secondary_accwrites != mp->mnt_secondary_accwrites)
386 		error = EAGAIN;
387 	BO_UNLOCK(bo);
388 	return (error);
389 }
390 
391 void
392 softdep_get_depcounts(struct mount *mp,
393 		      int *softdepactivep,
394 		      int *softdepactiveaccp)
395 {
396 	(void) mp;
397 	*softdepactivep = 0;
398 	*softdepactiveaccp = 0;
399 }
400 
401 #else
402 /*
403  * These definitions need to be adapted to the system to which
404  * this file is being ported.
405  */
406 /*
407  * malloc types defined for the softdep system.
408  */
409 static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
410 static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
411 static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
412 static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
413 static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
414 static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
415 static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
416 static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
417 static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
418 static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
419 static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
420 static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
421 static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
422 static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block");
423 static MALLOC_DEFINE(M_SAVEDINO, "savedino","Saved inodes");
424 
425 #define M_SOFTDEP_FLAGS	(M_WAITOK | M_USE_RESERVE)
426 
427 #define	D_PAGEDEP	0
428 #define	D_INODEDEP	1
429 #define	D_NEWBLK	2
430 #define	D_BMSAFEMAP	3
431 #define	D_ALLOCDIRECT	4
432 #define	D_INDIRDEP	5
433 #define	D_ALLOCINDIR	6
434 #define	D_FREEFRAG	7
435 #define	D_FREEBLKS	8
436 #define	D_FREEFILE	9
437 #define	D_DIRADD	10
438 #define	D_MKDIR		11
439 #define	D_DIRREM	12
440 #define	D_NEWDIRBLK	13
441 #define	D_LAST		D_NEWDIRBLK
442 
443 /*
444  * translate from workitem type to memory type
445  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
446  */
447 static struct malloc_type *memtype[] = {
448 	M_PAGEDEP,
449 	M_INODEDEP,
450 	M_NEWBLK,
451 	M_BMSAFEMAP,
452 	M_ALLOCDIRECT,
453 	M_INDIRDEP,
454 	M_ALLOCINDIR,
455 	M_FREEFRAG,
456 	M_FREEBLKS,
457 	M_FREEFILE,
458 	M_DIRADD,
459 	M_MKDIR,
460 	M_DIRREM,
461 	M_NEWDIRBLK
462 };
463 
464 #define DtoM(type) (memtype[type])
465 
466 /*
467  * Names of malloc types.
468  */
469 #define TYPENAME(type)  \
470 	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
471 /*
472  * End system adaptation definitions.
473  */
474 
475 /*
476  * Forward declarations.
477  */
478 struct inodedep_hashhead;
479 struct newblk_hashhead;
480 struct pagedep_hashhead;
481 
482 /*
483  * Internal function prototypes.
484  */
485 static	void softdep_error(char *, int);
486 static	void drain_output(struct vnode *);
487 static	struct buf *getdirtybuf(struct buf *, struct mtx *, int);
488 static	void clear_remove(struct thread *);
489 static	void clear_inodedeps(struct thread *);
490 static	int flush_pagedep_deps(struct vnode *, struct mount *,
491 	    struct diraddhd *);
492 static	int flush_inodedep_deps(struct mount *, ino_t);
493 static	int flush_deplist(struct allocdirectlst *, int, int *);
494 static	int handle_written_filepage(struct pagedep *, struct buf *);
495 static  void diradd_inode_written(struct diradd *, struct inodedep *);
496 static	int handle_written_inodeblock(struct inodedep *, struct buf *);
497 static	void handle_allocdirect_partdone(struct allocdirect *);
498 static	void handle_allocindir_partdone(struct allocindir *);
499 static	void initiate_write_filepage(struct pagedep *, struct buf *);
500 static	void handle_written_mkdir(struct mkdir *, int);
501 static	void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
502 static	void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
503 static	void handle_workitem_freefile(struct freefile *);
504 static	void handle_workitem_remove(struct dirrem *, struct vnode *);
505 static	struct dirrem *newdirrem(struct buf *, struct inode *,
506 	    struct inode *, int, struct dirrem **);
507 static	void free_diradd(struct diradd *);
508 static	void free_allocindir(struct allocindir *, struct inodedep *);
509 static	void free_newdirblk(struct newdirblk *);
510 static	int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t,
511 	    ufs2_daddr_t *);
512 static	void deallocate_dependencies(struct buf *, struct inodedep *);
513 static	void free_allocdirect(struct allocdirectlst *,
514 	    struct allocdirect *, int);
515 static	int check_inode_unwritten(struct inodedep *);
516 static	int free_inodedep(struct inodedep *);
517 static	void handle_workitem_freeblocks(struct freeblks *, int);
518 static	void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
519 static	void setup_allocindir_phase2(struct buf *, struct inode *,
520 	    struct allocindir *);
521 static	struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
522 	    ufs2_daddr_t);
523 static	void handle_workitem_freefrag(struct freefrag *);
524 static	struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long);
525 static	void allocdirect_merge(struct allocdirectlst *,
526 	    struct allocdirect *, struct allocdirect *);
527 static	struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *);
528 static	int newblk_find(struct newblk_hashhead *, struct fs *, ufs2_daddr_t,
529 	    struct newblk **);
530 static	int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **);
531 static	int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t,
532 	    struct inodedep **);
533 static	int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **);
534 static	int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **);
535 static	int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t,
536 	    struct mount *mp, int, struct pagedep **);
537 static	void pause_timer(void *);
538 static	int request_cleanup(struct mount *, int);
539 static	int process_worklist_item(struct mount *, int);
540 static	void add_to_worklist(struct worklist *);
541 static	void softdep_flush(void);
542 static	int softdep_speedup(void);
543 
544 /*
545  * Exported softdep operations.
546  */
547 static	void softdep_disk_io_initiation(struct buf *);
548 static	void softdep_disk_write_complete(struct buf *);
549 static	void softdep_deallocate_dependencies(struct buf *);
550 static	int softdep_count_dependencies(struct buf *bp, int);
551 
552 static struct mtx lk;
553 MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF);
554 
555 #define TRY_ACQUIRE_LOCK(lk)		mtx_trylock(lk)
556 #define ACQUIRE_LOCK(lk)		mtx_lock(lk)
557 #define FREE_LOCK(lk)			mtx_unlock(lk)
558 
559 #define	BUF_AREC(bp)	((bp)->b_lock.lock_object.lo_flags |= LO_RECURSABLE)
560 #define	BUF_NOREC(bp)	((bp)->b_lock.lock_object.lo_flags &= ~LO_RECURSABLE)
561 
562 /*
563  * Worklist queue management.
564  * These routines require that the lock be held.
565  */
566 #ifndef /* NOT */ DEBUG
567 #define WORKLIST_INSERT(head, item) do {	\
568 	(item)->wk_state |= ONWORKLIST;		\
569 	LIST_INSERT_HEAD(head, item, wk_list);	\
570 } while (0)
571 #define WORKLIST_REMOVE(item) do {		\
572 	(item)->wk_state &= ~ONWORKLIST;	\
573 	LIST_REMOVE(item, wk_list);		\
574 } while (0)
575 #else /* DEBUG */
576 static	void worklist_insert(struct workhead *, struct worklist *);
577 static	void worklist_remove(struct worklist *);
578 
579 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
580 #define WORKLIST_REMOVE(item) worklist_remove(item)
581 
582 static void
583 worklist_insert(head, item)
584 	struct workhead *head;
585 	struct worklist *item;
586 {
587 
588 	mtx_assert(&lk, MA_OWNED);
589 	if (item->wk_state & ONWORKLIST)
590 		panic("worklist_insert: already on list");
591 	item->wk_state |= ONWORKLIST;
592 	LIST_INSERT_HEAD(head, item, wk_list);
593 }
594 
595 static void
596 worklist_remove(item)
597 	struct worklist *item;
598 {
599 
600 	mtx_assert(&lk, MA_OWNED);
601 	if ((item->wk_state & ONWORKLIST) == 0)
602 		panic("worklist_remove: not on list");
603 	item->wk_state &= ~ONWORKLIST;
604 	LIST_REMOVE(item, wk_list);
605 }
606 #endif /* DEBUG */
607 
608 /*
609  * Routines for tracking and managing workitems.
610  */
611 static	void workitem_free(struct worklist *, int);
612 static	void workitem_alloc(struct worklist *, int, struct mount *);
613 
614 #define	WORKITEM_FREE(item, type) workitem_free((struct worklist *)(item), (type))
615 
616 static void
617 workitem_free(item, type)
618 	struct worklist *item;
619 	int type;
620 {
621 	struct ufsmount *ump;
622 	mtx_assert(&lk, MA_OWNED);
623 
624 #ifdef DEBUG
625 	if (item->wk_state & ONWORKLIST)
626 		panic("workitem_free: still on list");
627 	if (item->wk_type != type)
628 		panic("workitem_free: type mismatch");
629 #endif
630 	ump = VFSTOUFS(item->wk_mp);
631 	if (--ump->softdep_deps == 0 && ump->softdep_req)
632 		wakeup(&ump->softdep_deps);
633 	free(item, DtoM(type));
634 }
635 
636 static void
637 workitem_alloc(item, type, mp)
638 	struct worklist *item;
639 	int type;
640 	struct mount *mp;
641 {
642 	item->wk_type = type;
643 	item->wk_mp = mp;
644 	item->wk_state = 0;
645 	ACQUIRE_LOCK(&lk);
646 	VFSTOUFS(mp)->softdep_deps++;
647 	VFSTOUFS(mp)->softdep_accdeps++;
648 	FREE_LOCK(&lk);
649 }
650 
651 /*
652  * Workitem queue management
653  */
654 static int max_softdeps;	/* maximum number of structs before slowdown */
655 static int maxindirdeps = 50;	/* max number of indirdeps before slowdown */
656 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
657 static int proc_waiting;	/* tracks whether we have a timeout posted */
658 static int *stat_countp;	/* statistic to count in proc_waiting timeout */
659 static struct callout softdep_callout;
660 static int req_pending;
661 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
662 #define FLUSH_INODES		1
663 static int req_clear_remove;	/* syncer process flush some freeblks */
664 #define FLUSH_REMOVE		2
665 #define FLUSH_REMOVE_WAIT	3
666 static long num_freeblkdep;	/* number of freeblks workitems allocated */
667 
668 /*
669  * runtime statistics
670  */
671 static int stat_worklist_push;	/* number of worklist cleanups */
672 static int stat_blk_limit_push;	/* number of times block limit neared */
673 static int stat_ino_limit_push;	/* number of times inode limit neared */
674 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
675 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
676 static int stat_sync_limit_hit;	/* number of synchronous slowdowns imposed */
677 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
678 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
679 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
680 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
681 
682 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
683 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
684 SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, "");
685 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
686 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
687 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
688 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
689 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
690 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
691 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
692 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
693 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
694 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
695 /* SYSCTL_INT(_debug, OID_AUTO, worklist_num, CTLFLAG_RD, &softdep_on_worklist, 0, ""); */
696 
697 SYSCTL_DECL(_vfs_ffs);
698 
699 static int compute_summary_at_mount = 0;	/* Whether to recompute the summary at mount time */
700 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW,
701 	   &compute_summary_at_mount, 0, "Recompute summary at mount");
702 
703 static struct proc *softdepproc;
704 static struct kproc_desc softdep_kp = {
705 	"softdepflush",
706 	softdep_flush,
707 	&softdepproc
708 };
709 SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start,
710     &softdep_kp);
711 
712 static void
713 softdep_flush(void)
714 {
715 	struct mount *nmp;
716 	struct mount *mp;
717 	struct ufsmount *ump;
718 	struct thread *td;
719 	int remaining;
720 	int vfslocked;
721 
722 	td = curthread;
723 	td->td_pflags |= TDP_NORUNNINGBUF;
724 
725 	for (;;) {
726 		kproc_suspend_check(softdepproc);
727 		vfslocked = VFS_LOCK_GIANT((struct mount *)NULL);
728 		ACQUIRE_LOCK(&lk);
729 		/*
730 		 * If requested, try removing inode or removal dependencies.
731 		 */
732 		if (req_clear_inodedeps) {
733 			clear_inodedeps(td);
734 			req_clear_inodedeps -= 1;
735 			wakeup_one(&proc_waiting);
736 		}
737 		if (req_clear_remove) {
738 			clear_remove(td);
739 			req_clear_remove -= 1;
740 			wakeup_one(&proc_waiting);
741 		}
742 		FREE_LOCK(&lk);
743 		VFS_UNLOCK_GIANT(vfslocked);
744 		remaining = 0;
745 		mtx_lock(&mountlist_mtx);
746 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp)  {
747 			nmp = TAILQ_NEXT(mp, mnt_list);
748 			if ((mp->mnt_flag & MNT_SOFTDEP) == 0)
749 				continue;
750 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
751 				continue;
752 			vfslocked = VFS_LOCK_GIANT(mp);
753 			softdep_process_worklist(mp, 0);
754 			ump = VFSTOUFS(mp);
755 			remaining += ump->softdep_on_worklist -
756 				ump->softdep_on_worklist_inprogress;
757 			VFS_UNLOCK_GIANT(vfslocked);
758 			mtx_lock(&mountlist_mtx);
759 			nmp = TAILQ_NEXT(mp, mnt_list);
760 			vfs_unbusy(mp);
761 		}
762 		mtx_unlock(&mountlist_mtx);
763 		if (remaining)
764 			continue;
765 		ACQUIRE_LOCK(&lk);
766 		if (!req_pending)
767 			msleep(&req_pending, &lk, PVM, "sdflush", hz);
768 		req_pending = 0;
769 		FREE_LOCK(&lk);
770 	}
771 }
772 
773 static int
774 softdep_speedup(void)
775 {
776 
777 	mtx_assert(&lk, MA_OWNED);
778 	if (req_pending == 0) {
779 		req_pending = 1;
780 		wakeup(&req_pending);
781 	}
782 
783 	return speedup_syncer();
784 }
785 
786 /*
787  * Add an item to the end of the work queue.
788  * This routine requires that the lock be held.
789  * This is the only routine that adds items to the list.
790  * The following routine is the only one that removes items
791  * and does so in order from first to last.
792  */
793 static void
794 add_to_worklist(wk)
795 	struct worklist *wk;
796 {
797 	struct ufsmount *ump;
798 
799 	mtx_assert(&lk, MA_OWNED);
800 	ump = VFSTOUFS(wk->wk_mp);
801 	if (wk->wk_state & ONWORKLIST)
802 		panic("add_to_worklist: already on list");
803 	wk->wk_state |= ONWORKLIST;
804 	if (LIST_EMPTY(&ump->softdep_workitem_pending))
805 		LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list);
806 	else
807 		LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list);
808 	ump->softdep_worklist_tail = wk;
809 	ump->softdep_on_worklist += 1;
810 }
811 
812 /*
813  * Process that runs once per second to handle items in the background queue.
814  *
815  * Note that we ensure that everything is done in the order in which they
816  * appear in the queue. The code below depends on this property to ensure
817  * that blocks of a file are freed before the inode itself is freed. This
818  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
819  * until all the old ones have been purged from the dependency lists.
820  */
821 int
822 softdep_process_worklist(mp, full)
823 	struct mount *mp;
824 	int full;
825 {
826 	struct thread *td = curthread;
827 	int cnt, matchcnt, loopcount;
828 	struct ufsmount *ump;
829 	long starttime;
830 
831 	KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp"));
832 	/*
833 	 * Record the process identifier of our caller so that we can give
834 	 * this process preferential treatment in request_cleanup below.
835 	 */
836 	matchcnt = 0;
837 	ump = VFSTOUFS(mp);
838 	ACQUIRE_LOCK(&lk);
839 	loopcount = 1;
840 	starttime = time_second;
841 	while (ump->softdep_on_worklist > 0) {
842 		if ((cnt = process_worklist_item(mp, 0)) == -1)
843 			break;
844 		else
845 			matchcnt += cnt;
846 		/*
847 		 * If requested, try removing inode or removal dependencies.
848 		 */
849 		if (req_clear_inodedeps) {
850 			clear_inodedeps(td);
851 			req_clear_inodedeps -= 1;
852 			wakeup_one(&proc_waiting);
853 		}
854 		if (req_clear_remove) {
855 			clear_remove(td);
856 			req_clear_remove -= 1;
857 			wakeup_one(&proc_waiting);
858 		}
859 		/*
860 		 * We do not generally want to stop for buffer space, but if
861 		 * we are really being a buffer hog, we will stop and wait.
862 		 */
863 		if (loopcount++ % 128 == 0) {
864 			FREE_LOCK(&lk);
865 			uio_yield();
866 			bwillwrite();
867 			ACQUIRE_LOCK(&lk);
868 		}
869 		/*
870 		 * Never allow processing to run for more than one
871 		 * second. Otherwise the other mountpoints may get
872 		 * excessively backlogged.
873 		 */
874 		if (!full && starttime != time_second) {
875 			matchcnt = -1;
876 			break;
877 		}
878 	}
879 	FREE_LOCK(&lk);
880 	return (matchcnt);
881 }
882 
883 /*
884  * Process one item on the worklist.
885  */
886 static int
887 process_worklist_item(mp, flags)
888 	struct mount *mp;
889 	int flags;
890 {
891 	struct worklist *wk, *wkend;
892 	struct ufsmount *ump;
893 	struct vnode *vp;
894 	int matchcnt = 0;
895 
896 	mtx_assert(&lk, MA_OWNED);
897 	KASSERT(mp != NULL, ("process_worklist_item: NULL mp"));
898 	/*
899 	 * If we are being called because of a process doing a
900 	 * copy-on-write, then it is not safe to write as we may
901 	 * recurse into the copy-on-write routine.
902 	 */
903 	if (curthread->td_pflags & TDP_COWINPROGRESS)
904 		return (-1);
905 	/*
906 	 * Normally we just process each item on the worklist in order.
907 	 * However, if we are in a situation where we cannot lock any
908 	 * inodes, we have to skip over any dirrem requests whose
909 	 * vnodes are resident and locked.
910 	 */
911 	ump = VFSTOUFS(mp);
912 	vp = NULL;
913 	LIST_FOREACH(wk, &ump->softdep_workitem_pending, wk_list) {
914 		if (wk->wk_state & INPROGRESS)
915 			continue;
916 		if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
917 			break;
918 		wk->wk_state |= INPROGRESS;
919 		ump->softdep_on_worklist_inprogress++;
920 		FREE_LOCK(&lk);
921 		ffs_vgetf(mp, WK_DIRREM(wk)->dm_oldinum,
922 		    LK_NOWAIT | LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ);
923 		ACQUIRE_LOCK(&lk);
924 		wk->wk_state &= ~INPROGRESS;
925 		ump->softdep_on_worklist_inprogress--;
926 		if (vp != NULL)
927 			break;
928 	}
929 	if (wk == 0)
930 		return (-1);
931 	/*
932 	 * Remove the item to be processed. If we are removing the last
933 	 * item on the list, we need to recalculate the tail pointer.
934 	 * As this happens rarely and usually when the list is short,
935 	 * we just run down the list to find it rather than tracking it
936 	 * in the above loop.
937 	 */
938 	WORKLIST_REMOVE(wk);
939 	if (wk == ump->softdep_worklist_tail) {
940 		LIST_FOREACH(wkend, &ump->softdep_workitem_pending, wk_list)
941 			if (LIST_NEXT(wkend, wk_list) == NULL)
942 				break;
943 		ump->softdep_worklist_tail = wkend;
944 	}
945 	ump->softdep_on_worklist -= 1;
946 	FREE_LOCK(&lk);
947 	if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
948 		panic("process_worklist_item: suspended filesystem");
949 	matchcnt++;
950 	switch (wk->wk_type) {
951 
952 	case D_DIRREM:
953 		/* removal of a directory entry */
954 		handle_workitem_remove(WK_DIRREM(wk), vp);
955 		break;
956 
957 	case D_FREEBLKS:
958 		/* releasing blocks and/or fragments from a file */
959 		handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT);
960 		break;
961 
962 	case D_FREEFRAG:
963 		/* releasing a fragment when replaced as a file grows */
964 		handle_workitem_freefrag(WK_FREEFRAG(wk));
965 		break;
966 
967 	case D_FREEFILE:
968 		/* releasing an inode when its link count drops to 0 */
969 		handle_workitem_freefile(WK_FREEFILE(wk));
970 		break;
971 
972 	default:
973 		panic("%s_process_worklist: Unknown type %s",
974 		    "softdep", TYPENAME(wk->wk_type));
975 		/* NOTREACHED */
976 	}
977 	vn_finished_secondary_write(mp);
978 	ACQUIRE_LOCK(&lk);
979 	return (matchcnt);
980 }
981 
982 /*
983  * Move dependencies from one buffer to another.
984  */
985 void
986 softdep_move_dependencies(oldbp, newbp)
987 	struct buf *oldbp;
988 	struct buf *newbp;
989 {
990 	struct worklist *wk, *wktail;
991 
992 	if (!LIST_EMPTY(&newbp->b_dep))
993 		panic("softdep_move_dependencies: need merge code");
994 	wktail = 0;
995 	ACQUIRE_LOCK(&lk);
996 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
997 		LIST_REMOVE(wk, wk_list);
998 		if (wktail == 0)
999 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
1000 		else
1001 			LIST_INSERT_AFTER(wktail, wk, wk_list);
1002 		wktail = wk;
1003 	}
1004 	FREE_LOCK(&lk);
1005 }
1006 
1007 /*
1008  * Purge the work list of all items associated with a particular mount point.
1009  */
1010 int
1011 softdep_flushworklist(oldmnt, countp, td)
1012 	struct mount *oldmnt;
1013 	int *countp;
1014 	struct thread *td;
1015 {
1016 	struct vnode *devvp;
1017 	int count, error = 0;
1018 	struct ufsmount *ump;
1019 
1020 	/*
1021 	 * Alternately flush the block device associated with the mount
1022 	 * point and process any dependencies that the flushing
1023 	 * creates. We continue until no more worklist dependencies
1024 	 * are found.
1025 	 */
1026 	*countp = 0;
1027 	ump = VFSTOUFS(oldmnt);
1028 	devvp = ump->um_devvp;
1029 	while ((count = softdep_process_worklist(oldmnt, 1)) > 0) {
1030 		*countp += count;
1031 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1032 		error = VOP_FSYNC(devvp, MNT_WAIT, td);
1033 		VOP_UNLOCK(devvp, 0);
1034 		if (error)
1035 			break;
1036 	}
1037 	return (error);
1038 }
1039 
1040 int
1041 softdep_waitidle(struct mount *mp)
1042 {
1043 	struct ufsmount *ump;
1044 	int error;
1045 	int i;
1046 
1047 	ump = VFSTOUFS(mp);
1048 	ACQUIRE_LOCK(&lk);
1049 	for (i = 0; i < 10 && ump->softdep_deps; i++) {
1050 		ump->softdep_req = 1;
1051 		if (ump->softdep_on_worklist)
1052 			panic("softdep_waitidle: work added after flush.");
1053 		msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1);
1054 	}
1055 	ump->softdep_req = 0;
1056 	FREE_LOCK(&lk);
1057 	error = 0;
1058 	if (i == 10) {
1059 		error = EBUSY;
1060 		printf("softdep_waitidle: Failed to flush worklist for %p\n",
1061 		    mp);
1062 	}
1063 
1064 	return (error);
1065 }
1066 
1067 /*
1068  * Flush all vnodes and worklist items associated with a specified mount point.
1069  */
1070 int
1071 softdep_flushfiles(oldmnt, flags, td)
1072 	struct mount *oldmnt;
1073 	int flags;
1074 	struct thread *td;
1075 {
1076 	int error, depcount, loopcnt, retry_flush_count, retry;
1077 
1078 	loopcnt = 10;
1079 	retry_flush_count = 3;
1080 retry_flush:
1081 	error = 0;
1082 
1083 	/*
1084 	 * Alternately flush the vnodes associated with the mount
1085 	 * point and process any dependencies that the flushing
1086 	 * creates. In theory, this loop can happen at most twice,
1087 	 * but we give it a few extra just to be sure.
1088 	 */
1089 	for (; loopcnt > 0; loopcnt--) {
1090 		/*
1091 		 * Do another flush in case any vnodes were brought in
1092 		 * as part of the cleanup operations.
1093 		 */
1094 		if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
1095 			break;
1096 		if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 ||
1097 		    depcount == 0)
1098 			break;
1099 	}
1100 	/*
1101 	 * If we are unmounting then it is an error to fail. If we
1102 	 * are simply trying to downgrade to read-only, then filesystem
1103 	 * activity can keep us busy forever, so we just fail with EBUSY.
1104 	 */
1105 	if (loopcnt == 0) {
1106 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
1107 			panic("softdep_flushfiles: looping");
1108 		error = EBUSY;
1109 	}
1110 	if (!error)
1111 		error = softdep_waitidle(oldmnt);
1112 	if (!error) {
1113 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) {
1114 			retry = 0;
1115 			MNT_ILOCK(oldmnt);
1116 			KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0,
1117 			    ("softdep_flushfiles: !MNTK_NOINSMNTQ"));
1118 			if (oldmnt->mnt_nvnodelistsize > 0) {
1119 				if (--retry_flush_count > 0) {
1120 					retry = 1;
1121 					loopcnt = 3;
1122 				} else
1123 					error = EBUSY;
1124 			}
1125 			MNT_IUNLOCK(oldmnt);
1126 			if (retry)
1127 				goto retry_flush;
1128 		}
1129 	}
1130 	return (error);
1131 }
1132 
1133 /*
1134  * Structure hashing.
1135  *
1136  * There are three types of structures that can be looked up:
1137  *	1) pagedep structures identified by mount point, inode number,
1138  *	   and logical block.
1139  *	2) inodedep structures identified by mount point and inode number.
1140  *	3) newblk structures identified by mount point and
1141  *	   physical block number.
1142  *
1143  * The "pagedep" and "inodedep" dependency structures are hashed
1144  * separately from the file blocks and inodes to which they correspond.
1145  * This separation helps when the in-memory copy of an inode or
1146  * file block must be replaced. It also obviates the need to access
1147  * an inode or file page when simply updating (or de-allocating)
1148  * dependency structures. Lookup of newblk structures is needed to
1149  * find newly allocated blocks when trying to associate them with
1150  * their allocdirect or allocindir structure.
1151  *
1152  * The lookup routines optionally create and hash a new instance when
1153  * an existing entry is not found.
1154  */
1155 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
1156 #define NODELAY		0x0002	/* cannot do background work */
1157 
1158 /*
1159  * Structures and routines associated with pagedep caching.
1160  */
1161 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
1162 u_long	pagedep_hash;		/* size of hash table - 1 */
1163 #define	PAGEDEP_HASH(mp, inum, lbn) \
1164 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
1165 	    pagedep_hash])
1166 
1167 static int
1168 pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp)
1169 	struct pagedep_hashhead *pagedephd;
1170 	ino_t ino;
1171 	ufs_lbn_t lbn;
1172 	struct mount *mp;
1173 	int flags;
1174 	struct pagedep **pagedeppp;
1175 {
1176 	struct pagedep *pagedep;
1177 
1178 	LIST_FOREACH(pagedep, pagedephd, pd_hash)
1179 		if (ino == pagedep->pd_ino &&
1180 		    lbn == pagedep->pd_lbn &&
1181 		    mp == pagedep->pd_list.wk_mp)
1182 			break;
1183 	if (pagedep) {
1184 		*pagedeppp = pagedep;
1185 		if ((flags & DEPALLOC) != 0 &&
1186 		    (pagedep->pd_state & ONWORKLIST) == 0)
1187 			return (0);
1188 		return (1);
1189 	}
1190 	*pagedeppp = NULL;
1191 	return (0);
1192 }
1193 /*
1194  * Look up a pagedep. Return 1 if found, 0 if not found or found
1195  * when asked to allocate but not associated with any buffer.
1196  * If not found, allocate if DEPALLOC flag is passed.
1197  * Found or allocated entry is returned in pagedeppp.
1198  * This routine must be called with splbio interrupts blocked.
1199  */
1200 static int
1201 pagedep_lookup(ip, lbn, flags, pagedeppp)
1202 	struct inode *ip;
1203 	ufs_lbn_t lbn;
1204 	int flags;
1205 	struct pagedep **pagedeppp;
1206 {
1207 	struct pagedep *pagedep;
1208 	struct pagedep_hashhead *pagedephd;
1209 	struct mount *mp;
1210 	int ret;
1211 	int i;
1212 
1213 	mtx_assert(&lk, MA_OWNED);
1214 	mp = ITOV(ip)->v_mount;
1215 	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
1216 
1217 	ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp);
1218 	if (*pagedeppp || (flags & DEPALLOC) == 0)
1219 		return (ret);
1220 	FREE_LOCK(&lk);
1221 	pagedep = malloc(sizeof(struct pagedep),
1222 	    M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO);
1223 	workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp);
1224 	ACQUIRE_LOCK(&lk);
1225 	ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp);
1226 	if (*pagedeppp) {
1227 		WORKITEM_FREE(pagedep, D_PAGEDEP);
1228 		return (ret);
1229 	}
1230 	pagedep->pd_ino = ip->i_number;
1231 	pagedep->pd_lbn = lbn;
1232 	LIST_INIT(&pagedep->pd_dirremhd);
1233 	LIST_INIT(&pagedep->pd_pendinghd);
1234 	for (i = 0; i < DAHASHSZ; i++)
1235 		LIST_INIT(&pagedep->pd_diraddhd[i]);
1236 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
1237 	*pagedeppp = pagedep;
1238 	return (0);
1239 }
1240 
1241 /*
1242  * Structures and routines associated with inodedep caching.
1243  */
1244 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
1245 static u_long	inodedep_hash;	/* size of hash table - 1 */
1246 static long	num_inodedep;	/* number of inodedep allocated */
1247 #define	INODEDEP_HASH(fs, inum) \
1248       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
1249 
1250 static int
1251 inodedep_find(inodedephd, fs, inum, inodedeppp)
1252 	struct inodedep_hashhead *inodedephd;
1253 	struct fs *fs;
1254 	ino_t inum;
1255 	struct inodedep **inodedeppp;
1256 {
1257 	struct inodedep *inodedep;
1258 
1259 	LIST_FOREACH(inodedep, inodedephd, id_hash)
1260 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
1261 			break;
1262 	if (inodedep) {
1263 		*inodedeppp = inodedep;
1264 		return (1);
1265 	}
1266 	*inodedeppp = NULL;
1267 
1268 	return (0);
1269 }
1270 /*
1271  * Look up an inodedep. Return 1 if found, 0 if not found.
1272  * If not found, allocate if DEPALLOC flag is passed.
1273  * Found or allocated entry is returned in inodedeppp.
1274  * This routine must be called with splbio interrupts blocked.
1275  */
1276 static int
1277 inodedep_lookup(mp, inum, flags, inodedeppp)
1278 	struct mount *mp;
1279 	ino_t inum;
1280 	int flags;
1281 	struct inodedep **inodedeppp;
1282 {
1283 	struct inodedep *inodedep;
1284 	struct inodedep_hashhead *inodedephd;
1285 	struct fs *fs;
1286 
1287 	mtx_assert(&lk, MA_OWNED);
1288 	fs = VFSTOUFS(mp)->um_fs;
1289 	inodedephd = INODEDEP_HASH(fs, inum);
1290 
1291 	if (inodedep_find(inodedephd, fs, inum, inodedeppp))
1292 		return (1);
1293 	if ((flags & DEPALLOC) == 0)
1294 		return (0);
1295 	/*
1296 	 * If we are over our limit, try to improve the situation.
1297 	 */
1298 	if (num_inodedep > max_softdeps && (flags & NODELAY) == 0)
1299 		request_cleanup(mp, FLUSH_INODES);
1300 	FREE_LOCK(&lk);
1301 	inodedep = malloc(sizeof(struct inodedep),
1302 		M_INODEDEP, M_SOFTDEP_FLAGS);
1303 	workitem_alloc(&inodedep->id_list, D_INODEDEP, mp);
1304 	ACQUIRE_LOCK(&lk);
1305 	if (inodedep_find(inodedephd, fs, inum, inodedeppp)) {
1306 		WORKITEM_FREE(inodedep, D_INODEDEP);
1307 		return (1);
1308 	}
1309 	num_inodedep += 1;
1310 	inodedep->id_fs = fs;
1311 	inodedep->id_ino = inum;
1312 	inodedep->id_state = ALLCOMPLETE;
1313 	inodedep->id_nlinkdelta = 0;
1314 	inodedep->id_savedino1 = NULL;
1315 	inodedep->id_savedsize = -1;
1316 	inodedep->id_savedextsize = -1;
1317 	inodedep->id_buf = NULL;
1318 	LIST_INIT(&inodedep->id_pendinghd);
1319 	LIST_INIT(&inodedep->id_inowait);
1320 	LIST_INIT(&inodedep->id_bufwait);
1321 	TAILQ_INIT(&inodedep->id_inoupdt);
1322 	TAILQ_INIT(&inodedep->id_newinoupdt);
1323 	TAILQ_INIT(&inodedep->id_extupdt);
1324 	TAILQ_INIT(&inodedep->id_newextupdt);
1325 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1326 	*inodedeppp = inodedep;
1327 	return (0);
1328 }
1329 
1330 /*
1331  * Structures and routines associated with newblk caching.
1332  */
1333 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1334 u_long	newblk_hash;		/* size of hash table - 1 */
1335 #define	NEWBLK_HASH(fs, inum) \
1336 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1337 
1338 static int
1339 newblk_find(newblkhd, fs, newblkno, newblkpp)
1340 	struct newblk_hashhead *newblkhd;
1341 	struct fs *fs;
1342 	ufs2_daddr_t newblkno;
1343 	struct newblk **newblkpp;
1344 {
1345 	struct newblk *newblk;
1346 
1347 	LIST_FOREACH(newblk, newblkhd, nb_hash)
1348 		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1349 			break;
1350 	if (newblk) {
1351 		*newblkpp = newblk;
1352 		return (1);
1353 	}
1354 	*newblkpp = NULL;
1355 	return (0);
1356 }
1357 
1358 /*
1359  * Look up a newblk. Return 1 if found, 0 if not found.
1360  * If not found, allocate if DEPALLOC flag is passed.
1361  * Found or allocated entry is returned in newblkpp.
1362  */
1363 static int
1364 newblk_lookup(fs, newblkno, flags, newblkpp)
1365 	struct fs *fs;
1366 	ufs2_daddr_t newblkno;
1367 	int flags;
1368 	struct newblk **newblkpp;
1369 {
1370 	struct newblk *newblk;
1371 	struct newblk_hashhead *newblkhd;
1372 
1373 	newblkhd = NEWBLK_HASH(fs, newblkno);
1374 	if (newblk_find(newblkhd, fs, newblkno, newblkpp))
1375 		return (1);
1376 	if ((flags & DEPALLOC) == 0)
1377 		return (0);
1378 	FREE_LOCK(&lk);
1379 	newblk = malloc(sizeof(struct newblk),
1380 		M_NEWBLK, M_SOFTDEP_FLAGS);
1381 	ACQUIRE_LOCK(&lk);
1382 	if (newblk_find(newblkhd, fs, newblkno, newblkpp)) {
1383 		free(newblk, M_NEWBLK);
1384 		return (1);
1385 	}
1386 	newblk->nb_state = 0;
1387 	newblk->nb_fs = fs;
1388 	newblk->nb_newblkno = newblkno;
1389 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1390 	*newblkpp = newblk;
1391 	return (0);
1392 }
1393 
1394 /*
1395  * Executed during filesystem system initialization before
1396  * mounting any filesystems.
1397  */
1398 void
1399 softdep_initialize()
1400 {
1401 
1402 	LIST_INIT(&mkdirlisthd);
1403 	max_softdeps = desiredvnodes * 4;
1404 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1405 	    &pagedep_hash);
1406 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1407 	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1408 
1409 	/* initialise bioops hack */
1410 	bioops.io_start = softdep_disk_io_initiation;
1411 	bioops.io_complete = softdep_disk_write_complete;
1412 	bioops.io_deallocate = softdep_deallocate_dependencies;
1413 	bioops.io_countdeps = softdep_count_dependencies;
1414 
1415 	/* Initialize the callout with an mtx. */
1416 	callout_init_mtx(&softdep_callout, &lk, 0);
1417 }
1418 
1419 /*
1420  * Executed after all filesystems have been unmounted during
1421  * filesystem module unload.
1422  */
1423 void
1424 softdep_uninitialize()
1425 {
1426 
1427 	callout_drain(&softdep_callout);
1428 	hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash);
1429 	hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash);
1430 	hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash);
1431 }
1432 
1433 /*
1434  * Called at mount time to notify the dependency code that a
1435  * filesystem wishes to use it.
1436  */
1437 int
1438 softdep_mount(devvp, mp, fs, cred)
1439 	struct vnode *devvp;
1440 	struct mount *mp;
1441 	struct fs *fs;
1442 	struct ucred *cred;
1443 {
1444 	struct csum_total cstotal;
1445 	struct ufsmount *ump;
1446 	struct cg *cgp;
1447 	struct buf *bp;
1448 	int error, cyl;
1449 
1450 	MNT_ILOCK(mp);
1451 	mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP;
1452 	if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) {
1453 		mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) |
1454 			MNTK_SOFTDEP;
1455 		mp->mnt_noasync++;
1456 	}
1457 	MNT_IUNLOCK(mp);
1458 	ump = VFSTOUFS(mp);
1459 	LIST_INIT(&ump->softdep_workitem_pending);
1460 	ump->softdep_worklist_tail = NULL;
1461 	ump->softdep_on_worklist = 0;
1462 	ump->softdep_deps = 0;
1463 	/*
1464 	 * When doing soft updates, the counters in the
1465 	 * superblock may have gotten out of sync. Recomputation
1466 	 * can take a long time and can be deferred for background
1467 	 * fsck.  However, the old behavior of scanning the cylinder
1468 	 * groups and recalculating them at mount time is available
1469 	 * by setting vfs.ffs.compute_summary_at_mount to one.
1470 	 */
1471 	if (compute_summary_at_mount == 0 || fs->fs_clean != 0)
1472 		return (0);
1473 	bzero(&cstotal, sizeof cstotal);
1474 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1475 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1476 		    fs->fs_cgsize, cred, &bp)) != 0) {
1477 			brelse(bp);
1478 			return (error);
1479 		}
1480 		cgp = (struct cg *)bp->b_data;
1481 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1482 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1483 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1484 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1485 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
1486 		brelse(bp);
1487 	}
1488 #ifdef DEBUG
1489 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1490 		printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
1491 #endif
1492 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1493 	return (0);
1494 }
1495 
1496 /*
1497  * Protecting the freemaps (or bitmaps).
1498  *
1499  * To eliminate the need to execute fsck before mounting a filesystem
1500  * after a power failure, one must (conservatively) guarantee that the
1501  * on-disk copy of the bitmaps never indicate that a live inode or block is
1502  * free.  So, when a block or inode is allocated, the bitmap should be
1503  * updated (on disk) before any new pointers.  When a block or inode is
1504  * freed, the bitmap should not be updated until all pointers have been
1505  * reset.  The latter dependency is handled by the delayed de-allocation
1506  * approach described below for block and inode de-allocation.  The former
1507  * dependency is handled by calling the following procedure when a block or
1508  * inode is allocated. When an inode is allocated an "inodedep" is created
1509  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1510  * Each "inodedep" is also inserted into the hash indexing structure so
1511  * that any additional link additions can be made dependent on the inode
1512  * allocation.
1513  *
1514  * The ufs filesystem maintains a number of free block counts (e.g., per
1515  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1516  * in addition to the bitmaps.  These counts are used to improve efficiency
1517  * during allocation and therefore must be consistent with the bitmaps.
1518  * There is no convenient way to guarantee post-crash consistency of these
1519  * counts with simple update ordering, for two main reasons: (1) The counts
1520  * and bitmaps for a single cylinder group block are not in the same disk
1521  * sector.  If a disk write is interrupted (e.g., by power failure), one may
1522  * be written and the other not.  (2) Some of the counts are located in the
1523  * superblock rather than the cylinder group block. So, we focus our soft
1524  * updates implementation on protecting the bitmaps. When mounting a
1525  * filesystem, we recompute the auxiliary counts from the bitmaps.
1526  */
1527 
1528 /*
1529  * Called just after updating the cylinder group block to allocate an inode.
1530  */
1531 void
1532 softdep_setup_inomapdep(bp, ip, newinum)
1533 	struct buf *bp;		/* buffer for cylgroup block with inode map */
1534 	struct inode *ip;	/* inode related to allocation */
1535 	ino_t newinum;		/* new inode number being allocated */
1536 {
1537 	struct inodedep *inodedep;
1538 	struct bmsafemap *bmsafemap;
1539 
1540 	/*
1541 	 * Create a dependency for the newly allocated inode.
1542 	 * Panic if it already exists as something is seriously wrong.
1543 	 * Otherwise add it to the dependency list for the buffer holding
1544 	 * the cylinder group map from which it was allocated.
1545 	 */
1546 	ACQUIRE_LOCK(&lk);
1547 	if ((inodedep_lookup(UFSTOVFS(ip->i_ump), newinum, DEPALLOC|NODELAY,
1548 	    &inodedep)))
1549 		panic("softdep_setup_inomapdep: dependency for new inode "
1550 		    "already exists");
1551 	inodedep->id_buf = bp;
1552 	inodedep->id_state &= ~DEPCOMPLETE;
1553 	bmsafemap = bmsafemap_lookup(inodedep->id_list.wk_mp, bp);
1554 	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1555 	FREE_LOCK(&lk);
1556 }
1557 
1558 /*
1559  * Called just after updating the cylinder group block to
1560  * allocate block or fragment.
1561  */
1562 void
1563 softdep_setup_blkmapdep(bp, mp, newblkno)
1564 	struct buf *bp;		/* buffer for cylgroup block with block map */
1565 	struct mount *mp;	/* filesystem doing allocation */
1566 	ufs2_daddr_t newblkno;	/* number of newly allocated block */
1567 {
1568 	struct newblk *newblk;
1569 	struct bmsafemap *bmsafemap;
1570 	struct fs *fs;
1571 
1572 	fs = VFSTOUFS(mp)->um_fs;
1573 	/*
1574 	 * Create a dependency for the newly allocated block.
1575 	 * Add it to the dependency list for the buffer holding
1576 	 * the cylinder group map from which it was allocated.
1577 	 */
1578 	ACQUIRE_LOCK(&lk);
1579 	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1580 		panic("softdep_setup_blkmapdep: found block");
1581 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp);
1582 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1583 	FREE_LOCK(&lk);
1584 }
1585 
1586 /*
1587  * Find the bmsafemap associated with a cylinder group buffer.
1588  * If none exists, create one. The buffer must be locked when
1589  * this routine is called and this routine must be called with
1590  * splbio interrupts blocked.
1591  */
1592 static struct bmsafemap *
1593 bmsafemap_lookup(mp, bp)
1594 	struct mount *mp;
1595 	struct buf *bp;
1596 {
1597 	struct bmsafemap *bmsafemap;
1598 	struct worklist *wk;
1599 
1600 	mtx_assert(&lk, MA_OWNED);
1601 	LIST_FOREACH(wk, &bp->b_dep, wk_list)
1602 		if (wk->wk_type == D_BMSAFEMAP)
1603 			return (WK_BMSAFEMAP(wk));
1604 	FREE_LOCK(&lk);
1605 	bmsafemap = malloc(sizeof(struct bmsafemap),
1606 		M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1607 	workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
1608 	bmsafemap->sm_buf = bp;
1609 	LIST_INIT(&bmsafemap->sm_allocdirecthd);
1610 	LIST_INIT(&bmsafemap->sm_allocindirhd);
1611 	LIST_INIT(&bmsafemap->sm_inodedephd);
1612 	LIST_INIT(&bmsafemap->sm_newblkhd);
1613 	ACQUIRE_LOCK(&lk);
1614 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1615 	return (bmsafemap);
1616 }
1617 
1618 /*
1619  * Direct block allocation dependencies.
1620  *
1621  * When a new block is allocated, the corresponding disk locations must be
1622  * initialized (with zeros or new data) before the on-disk inode points to
1623  * them.  Also, the freemap from which the block was allocated must be
1624  * updated (on disk) before the inode's pointer. These two dependencies are
1625  * independent of each other and are needed for all file blocks and indirect
1626  * blocks that are pointed to directly by the inode.  Just before the
1627  * "in-core" version of the inode is updated with a newly allocated block
1628  * number, a procedure (below) is called to setup allocation dependency
1629  * structures.  These structures are removed when the corresponding
1630  * dependencies are satisfied or when the block allocation becomes obsolete
1631  * (i.e., the file is deleted, the block is de-allocated, or the block is a
1632  * fragment that gets upgraded).  All of these cases are handled in
1633  * procedures described later.
1634  *
1635  * When a file extension causes a fragment to be upgraded, either to a larger
1636  * fragment or to a full block, the on-disk location may change (if the
1637  * previous fragment could not simply be extended). In this case, the old
1638  * fragment must be de-allocated, but not until after the inode's pointer has
1639  * been updated. In most cases, this is handled by later procedures, which
1640  * will construct a "freefrag" structure to be added to the workitem queue
1641  * when the inode update is complete (or obsolete).  The main exception to
1642  * this is when an allocation occurs while a pending allocation dependency
1643  * (for the same block pointer) remains.  This case is handled in the main
1644  * allocation dependency setup procedure by immediately freeing the
1645  * unreferenced fragments.
1646  */
1647 void
1648 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1649 	struct inode *ip;	/* inode to which block is being added */
1650 	ufs_lbn_t lbn;		/* block pointer within inode */
1651 	ufs2_daddr_t newblkno;	/* disk block number being added */
1652 	ufs2_daddr_t oldblkno;	/* previous block number, 0 unless frag */
1653 	long newsize;		/* size of new block */
1654 	long oldsize;		/* size of new block */
1655 	struct buf *bp;		/* bp for allocated block */
1656 {
1657 	struct allocdirect *adp, *oldadp;
1658 	struct allocdirectlst *adphead;
1659 	struct bmsafemap *bmsafemap;
1660 	struct inodedep *inodedep;
1661 	struct pagedep *pagedep;
1662 	struct newblk *newblk;
1663 	struct mount *mp;
1664 
1665 	mp = UFSTOVFS(ip->i_ump);
1666 	adp = malloc(sizeof(struct allocdirect),
1667 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1668 	workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp);
1669 	adp->ad_lbn = lbn;
1670 	adp->ad_newblkno = newblkno;
1671 	adp->ad_oldblkno = oldblkno;
1672 	adp->ad_newsize = newsize;
1673 	adp->ad_oldsize = oldsize;
1674 	adp->ad_state = ATTACHED;
1675 	LIST_INIT(&adp->ad_newdirblk);
1676 	if (newblkno == oldblkno)
1677 		adp->ad_freefrag = NULL;
1678 	else
1679 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1680 
1681 	ACQUIRE_LOCK(&lk);
1682 	if (lbn >= NDADDR) {
1683 		/* allocating an indirect block */
1684 		if (oldblkno != 0)
1685 			panic("softdep_setup_allocdirect: non-zero indir");
1686 	} else {
1687 		/*
1688 		 * Allocating a direct block.
1689 		 *
1690 		 * If we are allocating a directory block, then we must
1691 		 * allocate an associated pagedep to track additions and
1692 		 * deletions.
1693 		 */
1694 		if ((ip->i_mode & IFMT) == IFDIR &&
1695 		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1696 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1697 	}
1698 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1699 		panic("softdep_setup_allocdirect: lost block");
1700 	if (newblk->nb_state == DEPCOMPLETE) {
1701 		adp->ad_state |= DEPCOMPLETE;
1702 		adp->ad_buf = NULL;
1703 	} else {
1704 		bmsafemap = newblk->nb_bmsafemap;
1705 		adp->ad_buf = bmsafemap->sm_buf;
1706 		LIST_REMOVE(newblk, nb_deps);
1707 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1708 	}
1709 	LIST_REMOVE(newblk, nb_hash);
1710 	free(newblk, M_NEWBLK);
1711 
1712 	inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1713 	adp->ad_inodedep = inodedep;
1714 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1715 	/*
1716 	 * The list of allocdirects must be kept in sorted and ascending
1717 	 * order so that the rollback routines can quickly determine the
1718 	 * first uncommitted block (the size of the file stored on disk
1719 	 * ends at the end of the lowest committed fragment, or if there
1720 	 * are no fragments, at the end of the highest committed block).
1721 	 * Since files generally grow, the typical case is that the new
1722 	 * block is to be added at the end of the list. We speed this
1723 	 * special case by checking against the last allocdirect in the
1724 	 * list before laboriously traversing the list looking for the
1725 	 * insertion point.
1726 	 */
1727 	adphead = &inodedep->id_newinoupdt;
1728 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1729 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1730 		/* insert at end of list */
1731 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1732 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1733 			allocdirect_merge(adphead, adp, oldadp);
1734 		FREE_LOCK(&lk);
1735 		return;
1736 	}
1737 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1738 		if (oldadp->ad_lbn >= lbn)
1739 			break;
1740 	}
1741 	if (oldadp == NULL)
1742 		panic("softdep_setup_allocdirect: lost entry");
1743 	/* insert in middle of list */
1744 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1745 	if (oldadp->ad_lbn == lbn)
1746 		allocdirect_merge(adphead, adp, oldadp);
1747 	FREE_LOCK(&lk);
1748 }
1749 
1750 /*
1751  * Replace an old allocdirect dependency with a newer one.
1752  * This routine must be called with splbio interrupts blocked.
1753  */
1754 static void
1755 allocdirect_merge(adphead, newadp, oldadp)
1756 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
1757 	struct allocdirect *newadp;	/* allocdirect being added */
1758 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
1759 {
1760 	struct worklist *wk;
1761 	struct freefrag *freefrag;
1762 	struct newdirblk *newdirblk;
1763 
1764 	mtx_assert(&lk, MA_OWNED);
1765 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1766 	    newadp->ad_oldsize != oldadp->ad_newsize ||
1767 	    newadp->ad_lbn >= NDADDR)
1768 		panic("%s %jd != new %jd || old size %ld != new %ld",
1769 		    "allocdirect_merge: old blkno",
1770 		    (intmax_t)newadp->ad_oldblkno,
1771 		    (intmax_t)oldadp->ad_newblkno,
1772 		    newadp->ad_oldsize, oldadp->ad_newsize);
1773 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
1774 	newadp->ad_oldsize = oldadp->ad_oldsize;
1775 	/*
1776 	 * If the old dependency had a fragment to free or had never
1777 	 * previously had a block allocated, then the new dependency
1778 	 * can immediately post its freefrag and adopt the old freefrag.
1779 	 * This action is done by swapping the freefrag dependencies.
1780 	 * The new dependency gains the old one's freefrag, and the
1781 	 * old one gets the new one and then immediately puts it on
1782 	 * the worklist when it is freed by free_allocdirect. It is
1783 	 * not possible to do this swap when the old dependency had a
1784 	 * non-zero size but no previous fragment to free. This condition
1785 	 * arises when the new block is an extension of the old block.
1786 	 * Here, the first part of the fragment allocated to the new
1787 	 * dependency is part of the block currently claimed on disk by
1788 	 * the old dependency, so cannot legitimately be freed until the
1789 	 * conditions for the new dependency are fulfilled.
1790 	 */
1791 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1792 		freefrag = newadp->ad_freefrag;
1793 		newadp->ad_freefrag = oldadp->ad_freefrag;
1794 		oldadp->ad_freefrag = freefrag;
1795 	}
1796 	/*
1797 	 * If we are tracking a new directory-block allocation,
1798 	 * move it from the old allocdirect to the new allocdirect.
1799 	 */
1800 	if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
1801 		newdirblk = WK_NEWDIRBLK(wk);
1802 		WORKLIST_REMOVE(&newdirblk->db_list);
1803 		if (!LIST_EMPTY(&oldadp->ad_newdirblk))
1804 			panic("allocdirect_merge: extra newdirblk");
1805 		WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list);
1806 	}
1807 	free_allocdirect(adphead, oldadp, 0);
1808 }
1809 
1810 /*
1811  * Allocate a new freefrag structure if needed.
1812  */
1813 static struct freefrag *
1814 newfreefrag(ip, blkno, size)
1815 	struct inode *ip;
1816 	ufs2_daddr_t blkno;
1817 	long size;
1818 {
1819 	struct freefrag *freefrag;
1820 	struct fs *fs;
1821 
1822 	if (blkno == 0)
1823 		return (NULL);
1824 	fs = ip->i_fs;
1825 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1826 		panic("newfreefrag: frag size");
1827 	freefrag = malloc(sizeof(struct freefrag),
1828 		M_FREEFRAG, M_SOFTDEP_FLAGS);
1829 	workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump));
1830 	freefrag->ff_inum = ip->i_number;
1831 	freefrag->ff_blkno = blkno;
1832 	freefrag->ff_fragsize = size;
1833 	return (freefrag);
1834 }
1835 
1836 /*
1837  * This workitem de-allocates fragments that were replaced during
1838  * file block allocation.
1839  */
1840 static void
1841 handle_workitem_freefrag(freefrag)
1842 	struct freefrag *freefrag;
1843 {
1844 	struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp);
1845 
1846 	ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
1847 	    freefrag->ff_fragsize, freefrag->ff_inum);
1848 	ACQUIRE_LOCK(&lk);
1849 	WORKITEM_FREE(freefrag, D_FREEFRAG);
1850 	FREE_LOCK(&lk);
1851 }
1852 
1853 /*
1854  * Set up a dependency structure for an external attributes data block.
1855  * This routine follows much of the structure of softdep_setup_allocdirect.
1856  * See the description of softdep_setup_allocdirect above for details.
1857  */
1858 void
1859 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1860 	struct inode *ip;
1861 	ufs_lbn_t lbn;
1862 	ufs2_daddr_t newblkno;
1863 	ufs2_daddr_t oldblkno;
1864 	long newsize;
1865 	long oldsize;
1866 	struct buf *bp;
1867 {
1868 	struct allocdirect *adp, *oldadp;
1869 	struct allocdirectlst *adphead;
1870 	struct bmsafemap *bmsafemap;
1871 	struct inodedep *inodedep;
1872 	struct newblk *newblk;
1873 	struct mount *mp;
1874 
1875 	mp = UFSTOVFS(ip->i_ump);
1876 	adp = malloc(sizeof(struct allocdirect),
1877 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1878 	workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp);
1879 	adp->ad_lbn = lbn;
1880 	adp->ad_newblkno = newblkno;
1881 	adp->ad_oldblkno = oldblkno;
1882 	adp->ad_newsize = newsize;
1883 	adp->ad_oldsize = oldsize;
1884 	adp->ad_state = ATTACHED | EXTDATA;
1885 	LIST_INIT(&adp->ad_newdirblk);
1886 	if (newblkno == oldblkno)
1887 		adp->ad_freefrag = NULL;
1888 	else
1889 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1890 
1891 	ACQUIRE_LOCK(&lk);
1892 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1893 		panic("softdep_setup_allocext: lost block");
1894 
1895 	inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1896 	adp->ad_inodedep = inodedep;
1897 
1898 	if (newblk->nb_state == DEPCOMPLETE) {
1899 		adp->ad_state |= DEPCOMPLETE;
1900 		adp->ad_buf = NULL;
1901 	} else {
1902 		bmsafemap = newblk->nb_bmsafemap;
1903 		adp->ad_buf = bmsafemap->sm_buf;
1904 		LIST_REMOVE(newblk, nb_deps);
1905 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1906 	}
1907 	LIST_REMOVE(newblk, nb_hash);
1908 	free(newblk, M_NEWBLK);
1909 
1910 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1911 	if (lbn >= NXADDR)
1912 		panic("softdep_setup_allocext: lbn %lld > NXADDR",
1913 		    (long long)lbn);
1914 	/*
1915 	 * The list of allocdirects must be kept in sorted and ascending
1916 	 * order so that the rollback routines can quickly determine the
1917 	 * first uncommitted block (the size of the file stored on disk
1918 	 * ends at the end of the lowest committed fragment, or if there
1919 	 * are no fragments, at the end of the highest committed block).
1920 	 * Since files generally grow, the typical case is that the new
1921 	 * block is to be added at the end of the list. We speed this
1922 	 * special case by checking against the last allocdirect in the
1923 	 * list before laboriously traversing the list looking for the
1924 	 * insertion point.
1925 	 */
1926 	adphead = &inodedep->id_newextupdt;
1927 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1928 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1929 		/* insert at end of list */
1930 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1931 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1932 			allocdirect_merge(adphead, adp, oldadp);
1933 		FREE_LOCK(&lk);
1934 		return;
1935 	}
1936 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1937 		if (oldadp->ad_lbn >= lbn)
1938 			break;
1939 	}
1940 	if (oldadp == NULL)
1941 		panic("softdep_setup_allocext: lost entry");
1942 	/* insert in middle of list */
1943 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1944 	if (oldadp->ad_lbn == lbn)
1945 		allocdirect_merge(adphead, adp, oldadp);
1946 	FREE_LOCK(&lk);
1947 }
1948 
1949 /*
1950  * Indirect block allocation dependencies.
1951  *
1952  * The same dependencies that exist for a direct block also exist when
1953  * a new block is allocated and pointed to by an entry in a block of
1954  * indirect pointers. The undo/redo states described above are also
1955  * used here. Because an indirect block contains many pointers that
1956  * may have dependencies, a second copy of the entire in-memory indirect
1957  * block is kept. The buffer cache copy is always completely up-to-date.
1958  * The second copy, which is used only as a source for disk writes,
1959  * contains only the safe pointers (i.e., those that have no remaining
1960  * update dependencies). The second copy is freed when all pointers
1961  * are safe. The cache is not allowed to replace indirect blocks with
1962  * pending update dependencies. If a buffer containing an indirect
1963  * block with dependencies is written, these routines will mark it
1964  * dirty again. It can only be successfully written once all the
1965  * dependencies are removed. The ffs_fsync routine in conjunction with
1966  * softdep_sync_metadata work together to get all the dependencies
1967  * removed so that a file can be successfully written to disk. Three
1968  * procedures are used when setting up indirect block pointer
1969  * dependencies. The division is necessary because of the organization
1970  * of the "balloc" routine and because of the distinction between file
1971  * pages and file metadata blocks.
1972  */
1973 
1974 /*
1975  * Allocate a new allocindir structure.
1976  */
1977 static struct allocindir *
1978 newallocindir(ip, ptrno, newblkno, oldblkno)
1979 	struct inode *ip;	/* inode for file being extended */
1980 	int ptrno;		/* offset of pointer in indirect block */
1981 	ufs2_daddr_t newblkno;	/* disk block number being added */
1982 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1983 {
1984 	struct allocindir *aip;
1985 
1986 	aip = malloc(sizeof(struct allocindir),
1987 		M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO);
1988 	workitem_alloc(&aip->ai_list, D_ALLOCINDIR, UFSTOVFS(ip->i_ump));
1989 	aip->ai_state = ATTACHED;
1990 	aip->ai_offset = ptrno;
1991 	aip->ai_newblkno = newblkno;
1992 	aip->ai_oldblkno = oldblkno;
1993 	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1994 	return (aip);
1995 }
1996 
1997 /*
1998  * Called just before setting an indirect block pointer
1999  * to a newly allocated file page.
2000  */
2001 void
2002 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
2003 	struct inode *ip;	/* inode for file being extended */
2004 	ufs_lbn_t lbn;		/* allocated block number within file */
2005 	struct buf *bp;		/* buffer with indirect blk referencing page */
2006 	int ptrno;		/* offset of pointer in indirect block */
2007 	ufs2_daddr_t newblkno;	/* disk block number being added */
2008 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
2009 	struct buf *nbp;	/* buffer holding allocated page */
2010 {
2011 	struct allocindir *aip;
2012 	struct pagedep *pagedep;
2013 
2014 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page");
2015 	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
2016 	ACQUIRE_LOCK(&lk);
2017 	/*
2018 	 * If we are allocating a directory page, then we must
2019 	 * allocate an associated pagedep to track additions and
2020 	 * deletions.
2021 	 */
2022 	if ((ip->i_mode & IFMT) == IFDIR &&
2023 	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
2024 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
2025 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
2026 	setup_allocindir_phase2(bp, ip, aip);
2027 	FREE_LOCK(&lk);
2028 }
2029 
2030 /*
2031  * Called just before setting an indirect block pointer to a
2032  * newly allocated indirect block.
2033  */
2034 void
2035 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
2036 	struct buf *nbp;	/* newly allocated indirect block */
2037 	struct inode *ip;	/* inode for file being extended */
2038 	struct buf *bp;		/* indirect block referencing allocated block */
2039 	int ptrno;		/* offset of pointer in indirect block */
2040 	ufs2_daddr_t newblkno;	/* disk block number being added */
2041 {
2042 	struct allocindir *aip;
2043 
2044 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta");
2045 	aip = newallocindir(ip, ptrno, newblkno, 0);
2046 	ACQUIRE_LOCK(&lk);
2047 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
2048 	setup_allocindir_phase2(bp, ip, aip);
2049 	FREE_LOCK(&lk);
2050 }
2051 
2052 /*
2053  * Called to finish the allocation of the "aip" allocated
2054  * by one of the two routines above.
2055  */
2056 static void
2057 setup_allocindir_phase2(bp, ip, aip)
2058 	struct buf *bp;		/* in-memory copy of the indirect block */
2059 	struct inode *ip;	/* inode for file being extended */
2060 	struct allocindir *aip;	/* allocindir allocated by the above routines */
2061 {
2062 	struct worklist *wk;
2063 	struct indirdep *indirdep, *newindirdep;
2064 	struct bmsafemap *bmsafemap;
2065 	struct allocindir *oldaip;
2066 	struct freefrag *freefrag;
2067 	struct newblk *newblk;
2068 	ufs2_daddr_t blkno;
2069 
2070 	mtx_assert(&lk, MA_OWNED);
2071 	if (bp->b_lblkno >= 0)
2072 		panic("setup_allocindir_phase2: not indir blk");
2073 	for (indirdep = NULL, newindirdep = NULL; ; ) {
2074 		LIST_FOREACH(wk, &bp->b_dep, wk_list) {
2075 			if (wk->wk_type != D_INDIRDEP)
2076 				continue;
2077 			indirdep = WK_INDIRDEP(wk);
2078 			break;
2079 		}
2080 		if (indirdep == NULL && newindirdep) {
2081 			indirdep = newindirdep;
2082 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
2083 			newindirdep = NULL;
2084 		}
2085 		if (indirdep) {
2086 			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
2087 			    &newblk) == 0)
2088 				panic("setup_allocindir: lost block");
2089 			if (newblk->nb_state == DEPCOMPLETE) {
2090 				aip->ai_state |= DEPCOMPLETE;
2091 				aip->ai_buf = NULL;
2092 			} else {
2093 				bmsafemap = newblk->nb_bmsafemap;
2094 				aip->ai_buf = bmsafemap->sm_buf;
2095 				LIST_REMOVE(newblk, nb_deps);
2096 				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
2097 				    aip, ai_deps);
2098 			}
2099 			LIST_REMOVE(newblk, nb_hash);
2100 			free(newblk, M_NEWBLK);
2101 			aip->ai_indirdep = indirdep;
2102 			/*
2103 			 * Check to see if there is an existing dependency
2104 			 * for this block. If there is, merge the old
2105 			 * dependency into the new one.
2106 			 */
2107 			if (aip->ai_oldblkno == 0)
2108 				oldaip = NULL;
2109 			else
2110 
2111 				LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
2112 					if (oldaip->ai_offset == aip->ai_offset)
2113 						break;
2114 			freefrag = NULL;
2115 			if (oldaip != NULL) {
2116 				if (oldaip->ai_newblkno != aip->ai_oldblkno)
2117 					panic("setup_allocindir_phase2: blkno");
2118 				aip->ai_oldblkno = oldaip->ai_oldblkno;
2119 				freefrag = aip->ai_freefrag;
2120 				aip->ai_freefrag = oldaip->ai_freefrag;
2121 				oldaip->ai_freefrag = NULL;
2122 				free_allocindir(oldaip, NULL);
2123 			}
2124 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
2125 			if (ip->i_ump->um_fstype == UFS1)
2126 				((ufs1_daddr_t *)indirdep->ir_savebp->b_data)
2127 				    [aip->ai_offset] = aip->ai_oldblkno;
2128 			else
2129 				((ufs2_daddr_t *)indirdep->ir_savebp->b_data)
2130 				    [aip->ai_offset] = aip->ai_oldblkno;
2131 			FREE_LOCK(&lk);
2132 			if (freefrag != NULL)
2133 				handle_workitem_freefrag(freefrag);
2134 		} else
2135 			FREE_LOCK(&lk);
2136 		if (newindirdep) {
2137 			newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
2138 			brelse(newindirdep->ir_savebp);
2139 			ACQUIRE_LOCK(&lk);
2140 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
2141 			if (indirdep)
2142 				break;
2143 			FREE_LOCK(&lk);
2144 		}
2145 		if (indirdep) {
2146 			ACQUIRE_LOCK(&lk);
2147 			break;
2148 		}
2149 		newindirdep = malloc(sizeof(struct indirdep),
2150 			M_INDIRDEP, M_SOFTDEP_FLAGS);
2151 		workitem_alloc(&newindirdep->ir_list, D_INDIRDEP,
2152 		    UFSTOVFS(ip->i_ump));
2153 		newindirdep->ir_state = ATTACHED;
2154 		if (ip->i_ump->um_fstype == UFS1)
2155 			newindirdep->ir_state |= UFS1FMT;
2156 		LIST_INIT(&newindirdep->ir_deplisthd);
2157 		LIST_INIT(&newindirdep->ir_donehd);
2158 		if (bp->b_blkno == bp->b_lblkno) {
2159 			ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
2160 			    NULL, NULL);
2161 			bp->b_blkno = blkno;
2162 		}
2163 		newindirdep->ir_savebp =
2164 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
2165 		BUF_KERNPROC(newindirdep->ir_savebp);
2166 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
2167 		ACQUIRE_LOCK(&lk);
2168 	}
2169 }
2170 
2171 /*
2172  * Block de-allocation dependencies.
2173  *
2174  * When blocks are de-allocated, the on-disk pointers must be nullified before
2175  * the blocks are made available for use by other files.  (The true
2176  * requirement is that old pointers must be nullified before new on-disk
2177  * pointers are set.  We chose this slightly more stringent requirement to
2178  * reduce complexity.) Our implementation handles this dependency by updating
2179  * the inode (or indirect block) appropriately but delaying the actual block
2180  * de-allocation (i.e., freemap and free space count manipulation) until
2181  * after the updated versions reach stable storage.  After the disk is
2182  * updated, the blocks can be safely de-allocated whenever it is convenient.
2183  * This implementation handles only the common case of reducing a file's
2184  * length to zero. Other cases are handled by the conventional synchronous
2185  * write approach.
2186  *
2187  * The ffs implementation with which we worked double-checks
2188  * the state of the block pointers and file size as it reduces
2189  * a file's length.  Some of this code is replicated here in our
2190  * soft updates implementation.  The freeblks->fb_chkcnt field is
2191  * used to transfer a part of this information to the procedure
2192  * that eventually de-allocates the blocks.
2193  *
2194  * This routine should be called from the routine that shortens
2195  * a file's length, before the inode's size or block pointers
2196  * are modified. It will save the block pointer information for
2197  * later release and zero the inode so that the calling routine
2198  * can release it.
2199  */
2200 void
2201 softdep_setup_freeblocks(ip, length, flags)
2202 	struct inode *ip;	/* The inode whose length is to be reduced */
2203 	off_t length;		/* The new length for the file */
2204 	int flags;		/* IO_EXT and/or IO_NORMAL */
2205 {
2206 	struct freeblks *freeblks;
2207 	struct inodedep *inodedep;
2208 	struct allocdirect *adp;
2209 	struct bufobj *bo;
2210 	struct vnode *vp;
2211 	struct buf *bp;
2212 	struct fs *fs;
2213 	ufs2_daddr_t extblocks, datablocks;
2214 	struct mount *mp;
2215 	int i, delay, error;
2216 
2217 	fs = ip->i_fs;
2218 	mp = UFSTOVFS(ip->i_ump);
2219 	if (length != 0)
2220 		panic("softdep_setup_freeblocks: non-zero length");
2221 	freeblks = malloc(sizeof(struct freeblks),
2222 		M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
2223 	workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp);
2224 	freeblks->fb_state = ATTACHED;
2225 	freeblks->fb_uid = ip->i_uid;
2226 	freeblks->fb_previousinum = ip->i_number;
2227 	freeblks->fb_devvp = ip->i_devvp;
2228 	ACQUIRE_LOCK(&lk);
2229 	num_freeblkdep++;
2230 	FREE_LOCK(&lk);
2231 	extblocks = 0;
2232 	if (fs->fs_magic == FS_UFS2_MAGIC)
2233 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
2234 	datablocks = DIP(ip, i_blocks) - extblocks;
2235 	if ((flags & IO_NORMAL) == 0) {
2236 		freeblks->fb_oldsize = 0;
2237 		freeblks->fb_chkcnt = 0;
2238 	} else {
2239 		freeblks->fb_oldsize = ip->i_size;
2240 		ip->i_size = 0;
2241 		DIP_SET(ip, i_size, 0);
2242 		freeblks->fb_chkcnt = datablocks;
2243 		for (i = 0; i < NDADDR; i++) {
2244 			freeblks->fb_dblks[i] = DIP(ip, i_db[i]);
2245 			DIP_SET(ip, i_db[i], 0);
2246 		}
2247 		for (i = 0; i < NIADDR; i++) {
2248 			freeblks->fb_iblks[i] = DIP(ip, i_ib[i]);
2249 			DIP_SET(ip, i_ib[i], 0);
2250 		}
2251 		/*
2252 		 * If the file was removed, then the space being freed was
2253 		 * accounted for then (see softdep_releasefile()). If the
2254 		 * file is merely being truncated, then we account for it now.
2255 		 */
2256 		if ((ip->i_flag & IN_SPACECOUNTED) == 0) {
2257 			UFS_LOCK(ip->i_ump);
2258 			fs->fs_pendingblocks += datablocks;
2259 			UFS_UNLOCK(ip->i_ump);
2260 		}
2261 	}
2262 	if ((flags & IO_EXT) == 0) {
2263 		freeblks->fb_oldextsize = 0;
2264 	} else {
2265 		freeblks->fb_oldextsize = ip->i_din2->di_extsize;
2266 		ip->i_din2->di_extsize = 0;
2267 		freeblks->fb_chkcnt += extblocks;
2268 		for (i = 0; i < NXADDR; i++) {
2269 			freeblks->fb_eblks[i] = ip->i_din2->di_extb[i];
2270 			ip->i_din2->di_extb[i] = 0;
2271 		}
2272 	}
2273 	DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt);
2274 	/*
2275 	 * Push the zero'ed inode to to its disk buffer so that we are free
2276 	 * to delete its dependencies below. Once the dependencies are gone
2277 	 * the buffer can be safely released.
2278 	 */
2279 	if ((error = bread(ip->i_devvp,
2280 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
2281 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
2282 		brelse(bp);
2283 		softdep_error("softdep_setup_freeblocks", error);
2284 	}
2285 	if (ip->i_ump->um_fstype == UFS1)
2286 		*((struct ufs1_dinode *)bp->b_data +
2287 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
2288 	else
2289 		*((struct ufs2_dinode *)bp->b_data +
2290 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
2291 	/*
2292 	 * Find and eliminate any inode dependencies.
2293 	 */
2294 	ACQUIRE_LOCK(&lk);
2295 	(void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
2296 	if ((inodedep->id_state & IOSTARTED) != 0)
2297 		panic("softdep_setup_freeblocks: inode busy");
2298 	/*
2299 	 * Add the freeblks structure to the list of operations that
2300 	 * must await the zero'ed inode being written to disk. If we
2301 	 * still have a bitmap dependency (delay == 0), then the inode
2302 	 * has never been written to disk, so we can process the
2303 	 * freeblks below once we have deleted the dependencies.
2304 	 */
2305 	delay = (inodedep->id_state & DEPCOMPLETE);
2306 	if (delay)
2307 		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
2308 	/*
2309 	 * Because the file length has been truncated to zero, any
2310 	 * pending block allocation dependency structures associated
2311 	 * with this inode are obsolete and can simply be de-allocated.
2312 	 * We must first merge the two dependency lists to get rid of
2313 	 * any duplicate freefrag structures, then purge the merged list.
2314 	 * If we still have a bitmap dependency, then the inode has never
2315 	 * been written to disk, so we can free any fragments without delay.
2316 	 */
2317 	if (flags & IO_NORMAL) {
2318 		merge_inode_lists(&inodedep->id_newinoupdt,
2319 		    &inodedep->id_inoupdt);
2320 		while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
2321 			free_allocdirect(&inodedep->id_inoupdt, adp, delay);
2322 	}
2323 	if (flags & IO_EXT) {
2324 		merge_inode_lists(&inodedep->id_newextupdt,
2325 		    &inodedep->id_extupdt);
2326 		while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
2327 			free_allocdirect(&inodedep->id_extupdt, adp, delay);
2328 	}
2329 	FREE_LOCK(&lk);
2330 	bdwrite(bp);
2331 	/*
2332 	 * We must wait for any I/O in progress to finish so that
2333 	 * all potential buffers on the dirty list will be visible.
2334 	 * Once they are all there, walk the list and get rid of
2335 	 * any dependencies.
2336 	 */
2337 	vp = ITOV(ip);
2338 	bo = &vp->v_bufobj;
2339 	BO_LOCK(bo);
2340 	drain_output(vp);
2341 restart:
2342 	TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2343 		if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
2344 		    ((flags & IO_NORMAL) == 0 &&
2345 		      (bp->b_xflags & BX_ALTDATA) == 0))
2346 			continue;
2347 		if ((bp = getdirtybuf(bp, BO_MTX(bo), MNT_WAIT)) == NULL)
2348 			goto restart;
2349 		BO_UNLOCK(bo);
2350 		ACQUIRE_LOCK(&lk);
2351 		(void) inodedep_lookup(mp, ip->i_number, 0, &inodedep);
2352 		deallocate_dependencies(bp, inodedep);
2353 		FREE_LOCK(&lk);
2354 		bp->b_flags |= B_INVAL | B_NOCACHE;
2355 		brelse(bp);
2356 		BO_LOCK(bo);
2357 		goto restart;
2358 	}
2359 	BO_UNLOCK(bo);
2360 	ACQUIRE_LOCK(&lk);
2361 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
2362 		(void) free_inodedep(inodedep);
2363 
2364 	if(delay) {
2365 		freeblks->fb_state |= DEPCOMPLETE;
2366 		/*
2367 		 * If the inode with zeroed block pointers is now on disk
2368 		 * we can start freeing blocks. Add freeblks to the worklist
2369 		 * instead of calling  handle_workitem_freeblocks directly as
2370 		 * it is more likely that additional IO is needed to complete
2371 		 * the request here than in the !delay case.
2372 		 */
2373 		if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
2374 			add_to_worklist(&freeblks->fb_list);
2375 	}
2376 
2377 	FREE_LOCK(&lk);
2378 	/*
2379 	 * If the inode has never been written to disk (delay == 0),
2380 	 * then we can process the freeblks now that we have deleted
2381 	 * the dependencies.
2382 	 */
2383 	if (!delay)
2384 		handle_workitem_freeblocks(freeblks, 0);
2385 }
2386 
2387 /*
2388  * Reclaim any dependency structures from a buffer that is about to
2389  * be reallocated to a new vnode. The buffer must be locked, thus,
2390  * no I/O completion operations can occur while we are manipulating
2391  * its associated dependencies. The mutex is held so that other I/O's
2392  * associated with related dependencies do not occur.
2393  */
2394 static void
2395 deallocate_dependencies(bp, inodedep)
2396 	struct buf *bp;
2397 	struct inodedep *inodedep;
2398 {
2399 	struct worklist *wk;
2400 	struct indirdep *indirdep;
2401 	struct allocindir *aip;
2402 	struct pagedep *pagedep;
2403 	struct dirrem *dirrem;
2404 	struct diradd *dap;
2405 	int i;
2406 
2407 	mtx_assert(&lk, MA_OWNED);
2408 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2409 		switch (wk->wk_type) {
2410 
2411 		case D_INDIRDEP:
2412 			indirdep = WK_INDIRDEP(wk);
2413 			/*
2414 			 * None of the indirect pointers will ever be visible,
2415 			 * so they can simply be tossed. GOINGAWAY ensures
2416 			 * that allocated pointers will be saved in the buffer
2417 			 * cache until they are freed. Note that they will
2418 			 * only be able to be found by their physical address
2419 			 * since the inode mapping the logical address will
2420 			 * be gone. The save buffer used for the safe copy
2421 			 * was allocated in setup_allocindir_phase2 using
2422 			 * the physical address so it could be used for this
2423 			 * purpose. Hence we swap the safe copy with the real
2424 			 * copy, allowing the safe copy to be freed and holding
2425 			 * on to the real copy for later use in indir_trunc.
2426 			 */
2427 			if (indirdep->ir_state & GOINGAWAY)
2428 				panic("deallocate_dependencies: already gone");
2429 			indirdep->ir_state |= GOINGAWAY;
2430 			VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1;
2431 			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
2432 				free_allocindir(aip, inodedep);
2433 			if (bp->b_lblkno >= 0 ||
2434 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno)
2435 				panic("deallocate_dependencies: not indir");
2436 			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2437 			    bp->b_bcount);
2438 			WORKLIST_REMOVE(wk);
2439 			WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
2440 			continue;
2441 
2442 		case D_PAGEDEP:
2443 			pagedep = WK_PAGEDEP(wk);
2444 			/*
2445 			 * None of the directory additions will ever be
2446 			 * visible, so they can simply be tossed.
2447 			 */
2448 			for (i = 0; i < DAHASHSZ; i++)
2449 				while ((dap =
2450 				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
2451 					free_diradd(dap);
2452 			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2453 				free_diradd(dap);
2454 			/*
2455 			 * Copy any directory remove dependencies to the list
2456 			 * to be processed after the zero'ed inode is written.
2457 			 * If the inode has already been written, then they
2458 			 * can be dumped directly onto the work list.
2459 			 */
2460 			LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2461 				LIST_REMOVE(dirrem, dm_next);
2462 				dirrem->dm_dirinum = pagedep->pd_ino;
2463 				if (inodedep == NULL ||
2464 				    (inodedep->id_state & ALLCOMPLETE) ==
2465 				     ALLCOMPLETE)
2466 					add_to_worklist(&dirrem->dm_list);
2467 				else
2468 					WORKLIST_INSERT(&inodedep->id_bufwait,
2469 					    &dirrem->dm_list);
2470 			}
2471 			if ((pagedep->pd_state & NEWBLOCK) != 0) {
2472 				LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list)
2473 					if (wk->wk_type == D_NEWDIRBLK &&
2474 					    WK_NEWDIRBLK(wk)->db_pagedep ==
2475 					      pagedep)
2476 						break;
2477 				if (wk != NULL) {
2478 					WORKLIST_REMOVE(wk);
2479 					free_newdirblk(WK_NEWDIRBLK(wk));
2480 				} else
2481 					panic("deallocate_dependencies: "
2482 					      "lost pagedep");
2483 			}
2484 			WORKLIST_REMOVE(&pagedep->pd_list);
2485 			LIST_REMOVE(pagedep, pd_hash);
2486 			WORKITEM_FREE(pagedep, D_PAGEDEP);
2487 			continue;
2488 
2489 		case D_ALLOCINDIR:
2490 			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2491 			continue;
2492 
2493 		case D_ALLOCDIRECT:
2494 		case D_INODEDEP:
2495 			panic("deallocate_dependencies: Unexpected type %s",
2496 			    TYPENAME(wk->wk_type));
2497 			/* NOTREACHED */
2498 
2499 		default:
2500 			panic("deallocate_dependencies: Unknown type %s",
2501 			    TYPENAME(wk->wk_type));
2502 			/* NOTREACHED */
2503 		}
2504 	}
2505 }
2506 
2507 /*
2508  * Free an allocdirect. Generate a new freefrag work request if appropriate.
2509  * This routine must be called with splbio interrupts blocked.
2510  */
2511 static void
2512 free_allocdirect(adphead, adp, delay)
2513 	struct allocdirectlst *adphead;
2514 	struct allocdirect *adp;
2515 	int delay;
2516 {
2517 	struct newdirblk *newdirblk;
2518 	struct worklist *wk;
2519 
2520 	mtx_assert(&lk, MA_OWNED);
2521 	if ((adp->ad_state & DEPCOMPLETE) == 0)
2522 		LIST_REMOVE(adp, ad_deps);
2523 	TAILQ_REMOVE(adphead, adp, ad_next);
2524 	if ((adp->ad_state & COMPLETE) == 0)
2525 		WORKLIST_REMOVE(&adp->ad_list);
2526 	if (adp->ad_freefrag != NULL) {
2527 		if (delay)
2528 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2529 			    &adp->ad_freefrag->ff_list);
2530 		else
2531 			add_to_worklist(&adp->ad_freefrag->ff_list);
2532 	}
2533 	if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) {
2534 		newdirblk = WK_NEWDIRBLK(wk);
2535 		WORKLIST_REMOVE(&newdirblk->db_list);
2536 		if (!LIST_EMPTY(&adp->ad_newdirblk))
2537 			panic("free_allocdirect: extra newdirblk");
2538 		if (delay)
2539 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2540 			    &newdirblk->db_list);
2541 		else
2542 			free_newdirblk(newdirblk);
2543 	}
2544 	WORKITEM_FREE(adp, D_ALLOCDIRECT);
2545 }
2546 
2547 /*
2548  * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
2549  * This routine must be called with splbio interrupts blocked.
2550  */
2551 static void
2552 free_newdirblk(newdirblk)
2553 	struct newdirblk *newdirblk;
2554 {
2555 	struct pagedep *pagedep;
2556 	struct diradd *dap;
2557 	int i;
2558 
2559 	mtx_assert(&lk, MA_OWNED);
2560 	/*
2561 	 * If the pagedep is still linked onto the directory buffer
2562 	 * dependency chain, then some of the entries on the
2563 	 * pd_pendinghd list may not be committed to disk yet. In
2564 	 * this case, we will simply clear the NEWBLOCK flag and
2565 	 * let the pd_pendinghd list be processed when the pagedep
2566 	 * is next written. If the pagedep is no longer on the buffer
2567 	 * dependency chain, then all the entries on the pd_pending
2568 	 * list are committed to disk and we can free them here.
2569 	 */
2570 	pagedep = newdirblk->db_pagedep;
2571 	pagedep->pd_state &= ~NEWBLOCK;
2572 	if ((pagedep->pd_state & ONWORKLIST) == 0)
2573 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
2574 			free_diradd(dap);
2575 	/*
2576 	 * If no dependencies remain, the pagedep will be freed.
2577 	 */
2578 	for (i = 0; i < DAHASHSZ; i++)
2579 		if (!LIST_EMPTY(&pagedep->pd_diraddhd[i]))
2580 			break;
2581 	if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) {
2582 		LIST_REMOVE(pagedep, pd_hash);
2583 		WORKITEM_FREE(pagedep, D_PAGEDEP);
2584 	}
2585 	WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2586 }
2587 
2588 /*
2589  * Prepare an inode to be freed. The actual free operation is not
2590  * done until the zero'ed inode has been written to disk.
2591  */
2592 void
2593 softdep_freefile(pvp, ino, mode)
2594 	struct vnode *pvp;
2595 	ino_t ino;
2596 	int mode;
2597 {
2598 	struct inode *ip = VTOI(pvp);
2599 	struct inodedep *inodedep;
2600 	struct freefile *freefile;
2601 
2602 	/*
2603 	 * This sets up the inode de-allocation dependency.
2604 	 */
2605 	freefile = malloc(sizeof(struct freefile),
2606 		M_FREEFILE, M_SOFTDEP_FLAGS);
2607 	workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount);
2608 	freefile->fx_mode = mode;
2609 	freefile->fx_oldinum = ino;
2610 	freefile->fx_devvp = ip->i_devvp;
2611 	if ((ip->i_flag & IN_SPACECOUNTED) == 0) {
2612 		UFS_LOCK(ip->i_ump);
2613 		ip->i_fs->fs_pendinginodes += 1;
2614 		UFS_UNLOCK(ip->i_ump);
2615 	}
2616 
2617 	/*
2618 	 * If the inodedep does not exist, then the zero'ed inode has
2619 	 * been written to disk. If the allocated inode has never been
2620 	 * written to disk, then the on-disk inode is zero'ed. In either
2621 	 * case we can free the file immediately.
2622 	 */
2623 	ACQUIRE_LOCK(&lk);
2624 	if (inodedep_lookup(pvp->v_mount, ino, 0, &inodedep) == 0 ||
2625 	    check_inode_unwritten(inodedep)) {
2626 		FREE_LOCK(&lk);
2627 		handle_workitem_freefile(freefile);
2628 		return;
2629 	}
2630 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2631 	FREE_LOCK(&lk);
2632 	if (ip->i_number == ino)
2633 		ip->i_flag |= IN_MODIFIED;
2634 }
2635 
2636 /*
2637  * Check to see if an inode has never been written to disk. If
2638  * so free the inodedep and return success, otherwise return failure.
2639  * This routine must be called with splbio interrupts blocked.
2640  *
2641  * If we still have a bitmap dependency, then the inode has never
2642  * been written to disk. Drop the dependency as it is no longer
2643  * necessary since the inode is being deallocated. We set the
2644  * ALLCOMPLETE flags since the bitmap now properly shows that the
2645  * inode is not allocated. Even if the inode is actively being
2646  * written, it has been rolled back to its zero'ed state, so we
2647  * are ensured that a zero inode is what is on the disk. For short
2648  * lived files, this change will usually result in removing all the
2649  * dependencies from the inode so that it can be freed immediately.
2650  */
2651 static int
2652 check_inode_unwritten(inodedep)
2653 	struct inodedep *inodedep;
2654 {
2655 
2656 	mtx_assert(&lk, MA_OWNED);
2657 	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2658 	    !LIST_EMPTY(&inodedep->id_pendinghd) ||
2659 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
2660 	    !LIST_EMPTY(&inodedep->id_inowait) ||
2661 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
2662 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
2663 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
2664 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
2665 	    inodedep->id_nlinkdelta != 0)
2666 		return (0);
2667 
2668 	/*
2669 	 * Another process might be in initiate_write_inodeblock_ufs[12]
2670 	 * trying to allocate memory without holding "Softdep Lock".
2671 	 */
2672 	if ((inodedep->id_state & IOSTARTED) != 0 &&
2673 	    inodedep->id_savedino1 == NULL)
2674 		return (0);
2675 
2676 	inodedep->id_state |= ALLCOMPLETE;
2677 	LIST_REMOVE(inodedep, id_deps);
2678 	inodedep->id_buf = NULL;
2679 	if (inodedep->id_state & ONWORKLIST)
2680 		WORKLIST_REMOVE(&inodedep->id_list);
2681 	if (inodedep->id_savedino1 != NULL) {
2682 		free(inodedep->id_savedino1, M_SAVEDINO);
2683 		inodedep->id_savedino1 = NULL;
2684 	}
2685 	if (free_inodedep(inodedep) == 0)
2686 		panic("check_inode_unwritten: busy inode");
2687 	return (1);
2688 }
2689 
2690 /*
2691  * Try to free an inodedep structure. Return 1 if it could be freed.
2692  */
2693 static int
2694 free_inodedep(inodedep)
2695 	struct inodedep *inodedep;
2696 {
2697 
2698 	mtx_assert(&lk, MA_OWNED);
2699 	if ((inodedep->id_state & ONWORKLIST) != 0 ||
2700 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2701 	    !LIST_EMPTY(&inodedep->id_pendinghd) ||
2702 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
2703 	    !LIST_EMPTY(&inodedep->id_inowait) ||
2704 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
2705 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
2706 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
2707 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
2708 	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL)
2709 		return (0);
2710 	LIST_REMOVE(inodedep, id_hash);
2711 	WORKITEM_FREE(inodedep, D_INODEDEP);
2712 	num_inodedep -= 1;
2713 	return (1);
2714 }
2715 
2716 /*
2717  * This workitem routine performs the block de-allocation.
2718  * The workitem is added to the pending list after the updated
2719  * inode block has been written to disk.  As mentioned above,
2720  * checks regarding the number of blocks de-allocated (compared
2721  * to the number of blocks allocated for the file) are also
2722  * performed in this function.
2723  */
2724 static void
2725 handle_workitem_freeblocks(freeblks, flags)
2726 	struct freeblks *freeblks;
2727 	int flags;
2728 {
2729 	struct inode *ip;
2730 	struct vnode *vp;
2731 	struct fs *fs;
2732 	struct ufsmount *ump;
2733 	int i, nblocks, level, bsize;
2734 	ufs2_daddr_t bn, blocksreleased = 0;
2735 	int error, allerror = 0;
2736 	ufs_lbn_t baselbns[NIADDR], tmpval;
2737 	int fs_pendingblocks;
2738 
2739 	ump = VFSTOUFS(freeblks->fb_list.wk_mp);
2740 	fs = ump->um_fs;
2741 	fs_pendingblocks = 0;
2742 	tmpval = 1;
2743 	baselbns[0] = NDADDR;
2744 	for (i = 1; i < NIADDR; i++) {
2745 		tmpval *= NINDIR(fs);
2746 		baselbns[i] = baselbns[i - 1] + tmpval;
2747 	}
2748 	nblocks = btodb(fs->fs_bsize);
2749 	blocksreleased = 0;
2750 	/*
2751 	 * Release all extended attribute blocks or frags.
2752 	 */
2753 	if (freeblks->fb_oldextsize > 0) {
2754 		for (i = (NXADDR - 1); i >= 0; i--) {
2755 			if ((bn = freeblks->fb_eblks[i]) == 0)
2756 				continue;
2757 			bsize = sblksize(fs, freeblks->fb_oldextsize, i);
2758 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize,
2759 			    freeblks->fb_previousinum);
2760 			blocksreleased += btodb(bsize);
2761 		}
2762 	}
2763 	/*
2764 	 * Release all data blocks or frags.
2765 	 */
2766 	if (freeblks->fb_oldsize > 0) {
2767 		/*
2768 		 * Indirect blocks first.
2769 		 */
2770 		for (level = (NIADDR - 1); level >= 0; level--) {
2771 			if ((bn = freeblks->fb_iblks[level]) == 0)
2772 				continue;
2773 			if ((error = indir_trunc(freeblks, fsbtodb(fs, bn),
2774 			    level, baselbns[level], &blocksreleased)) != 0)
2775 				allerror = error;
2776 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn,
2777 			    fs->fs_bsize, freeblks->fb_previousinum);
2778 			fs_pendingblocks += nblocks;
2779 			blocksreleased += nblocks;
2780 		}
2781 		/*
2782 		 * All direct blocks or frags.
2783 		 */
2784 		for (i = (NDADDR - 1); i >= 0; i--) {
2785 			if ((bn = freeblks->fb_dblks[i]) == 0)
2786 				continue;
2787 			bsize = sblksize(fs, freeblks->fb_oldsize, i);
2788 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize,
2789 			    freeblks->fb_previousinum);
2790 			fs_pendingblocks += btodb(bsize);
2791 			blocksreleased += btodb(bsize);
2792 		}
2793 	}
2794 	UFS_LOCK(ump);
2795 	fs->fs_pendingblocks -= fs_pendingblocks;
2796 	UFS_UNLOCK(ump);
2797 	/*
2798 	 * If we still have not finished background cleanup, then check
2799 	 * to see if the block count needs to be adjusted.
2800 	 */
2801 	if (freeblks->fb_chkcnt != blocksreleased &&
2802 	    (fs->fs_flags & FS_UNCLEAN) != 0 &&
2803 	    ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_previousinum,
2804 		(flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ)
2805 	    == 0) {
2806 		ip = VTOI(vp);
2807 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + \
2808 		    freeblks->fb_chkcnt - blocksreleased);
2809 		ip->i_flag |= IN_CHANGE;
2810 		vput(vp);
2811 	}
2812 
2813 #ifdef INVARIANTS
2814 	if (freeblks->fb_chkcnt != blocksreleased &&
2815 	    ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0))
2816 		printf("handle_workitem_freeblocks: block count\n");
2817 	if (allerror)
2818 		softdep_error("handle_workitem_freeblks", allerror);
2819 #endif /* INVARIANTS */
2820 
2821 	ACQUIRE_LOCK(&lk);
2822 	WORKITEM_FREE(freeblks, D_FREEBLKS);
2823 	num_freeblkdep--;
2824 	FREE_LOCK(&lk);
2825 }
2826 
2827 /*
2828  * Release blocks associated with the inode ip and stored in the indirect
2829  * block dbn. If level is greater than SINGLE, the block is an indirect block
2830  * and recursive calls to indirtrunc must be used to cleanse other indirect
2831  * blocks.
2832  */
2833 static int
2834 indir_trunc(freeblks, dbn, level, lbn, countp)
2835 	struct freeblks *freeblks;
2836 	ufs2_daddr_t dbn;
2837 	int level;
2838 	ufs_lbn_t lbn;
2839 	ufs2_daddr_t *countp;
2840 {
2841 	struct buf *bp;
2842 	struct fs *fs;
2843 	struct worklist *wk;
2844 	struct indirdep *indirdep;
2845 	struct ufsmount *ump;
2846 	ufs1_daddr_t *bap1 = 0;
2847 	ufs2_daddr_t nb, *bap2 = 0;
2848 	ufs_lbn_t lbnadd;
2849 	int i, nblocks, ufs1fmt;
2850 	int error, allerror = 0;
2851 	int fs_pendingblocks;
2852 
2853 	ump = VFSTOUFS(freeblks->fb_list.wk_mp);
2854 	fs = ump->um_fs;
2855 	fs_pendingblocks = 0;
2856 	lbnadd = 1;
2857 	for (i = level; i > 0; i--)
2858 		lbnadd *= NINDIR(fs);
2859 	/*
2860 	 * Get buffer of block pointers to be freed. This routine is not
2861 	 * called until the zero'ed inode has been written, so it is safe
2862 	 * to free blocks as they are encountered. Because the inode has
2863 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
2864 	 * have to use the on-disk address and the block device for the
2865 	 * filesystem to look them up. If the file was deleted before its
2866 	 * indirect blocks were all written to disk, the routine that set
2867 	 * us up (deallocate_dependencies) will have arranged to leave
2868 	 * a complete copy of the indirect block in memory for our use.
2869 	 * Otherwise we have to read the blocks in from the disk.
2870 	 */
2871 #ifdef notyet
2872 	bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0,
2873 	    GB_NOCREAT);
2874 #else
2875 	bp = incore(&freeblks->fb_devvp->v_bufobj, dbn);
2876 #endif
2877 	ACQUIRE_LOCK(&lk);
2878 	if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2879 		if (wk->wk_type != D_INDIRDEP ||
2880 		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2881 		    (indirdep->ir_state & GOINGAWAY) == 0)
2882 			panic("indir_trunc: lost indirdep");
2883 		WORKLIST_REMOVE(wk);
2884 		WORKITEM_FREE(indirdep, D_INDIRDEP);
2885 		if (!LIST_EMPTY(&bp->b_dep))
2886 			panic("indir_trunc: dangling dep");
2887 		ump->um_numindirdeps -= 1;
2888 		FREE_LOCK(&lk);
2889 	} else {
2890 #ifdef notyet
2891 		if (bp)
2892 			brelse(bp);
2893 #endif
2894 		FREE_LOCK(&lk);
2895 		error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
2896 		    NOCRED, &bp);
2897 		if (error) {
2898 			brelse(bp);
2899 			return (error);
2900 		}
2901 	}
2902 	/*
2903 	 * Recursively free indirect blocks.
2904 	 */
2905 	if (ump->um_fstype == UFS1) {
2906 		ufs1fmt = 1;
2907 		bap1 = (ufs1_daddr_t *)bp->b_data;
2908 	} else {
2909 		ufs1fmt = 0;
2910 		bap2 = (ufs2_daddr_t *)bp->b_data;
2911 	}
2912 	nblocks = btodb(fs->fs_bsize);
2913 	for (i = NINDIR(fs) - 1; i >= 0; i--) {
2914 		if (ufs1fmt)
2915 			nb = bap1[i];
2916 		else
2917 			nb = bap2[i];
2918 		if (nb == 0)
2919 			continue;
2920 		if (level != 0) {
2921 			if ((error = indir_trunc(freeblks, fsbtodb(fs, nb),
2922 			     level - 1, lbn + (i * lbnadd), countp)) != 0)
2923 				allerror = error;
2924 		}
2925 		ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, fs->fs_bsize,
2926 		    freeblks->fb_previousinum);
2927 		fs_pendingblocks += nblocks;
2928 		*countp += nblocks;
2929 	}
2930 	UFS_LOCK(ump);
2931 	fs->fs_pendingblocks -= fs_pendingblocks;
2932 	UFS_UNLOCK(ump);
2933 	bp->b_flags |= B_INVAL | B_NOCACHE;
2934 	brelse(bp);
2935 	return (allerror);
2936 }
2937 
2938 /*
2939  * Free an allocindir.
2940  * This routine must be called with splbio interrupts blocked.
2941  */
2942 static void
2943 free_allocindir(aip, inodedep)
2944 	struct allocindir *aip;
2945 	struct inodedep *inodedep;
2946 {
2947 	struct freefrag *freefrag;
2948 
2949 	mtx_assert(&lk, MA_OWNED);
2950 	if ((aip->ai_state & DEPCOMPLETE) == 0)
2951 		LIST_REMOVE(aip, ai_deps);
2952 	if (aip->ai_state & ONWORKLIST)
2953 		WORKLIST_REMOVE(&aip->ai_list);
2954 	LIST_REMOVE(aip, ai_next);
2955 	if ((freefrag = aip->ai_freefrag) != NULL) {
2956 		if (inodedep == NULL)
2957 			add_to_worklist(&freefrag->ff_list);
2958 		else
2959 			WORKLIST_INSERT(&inodedep->id_bufwait,
2960 			    &freefrag->ff_list);
2961 	}
2962 	WORKITEM_FREE(aip, D_ALLOCINDIR);
2963 }
2964 
2965 /*
2966  * Directory entry addition dependencies.
2967  *
2968  * When adding a new directory entry, the inode (with its incremented link
2969  * count) must be written to disk before the directory entry's pointer to it.
2970  * Also, if the inode is newly allocated, the corresponding freemap must be
2971  * updated (on disk) before the directory entry's pointer. These requirements
2972  * are met via undo/redo on the directory entry's pointer, which consists
2973  * simply of the inode number.
2974  *
2975  * As directory entries are added and deleted, the free space within a
2976  * directory block can become fragmented.  The ufs filesystem will compact
2977  * a fragmented directory block to make space for a new entry. When this
2978  * occurs, the offsets of previously added entries change. Any "diradd"
2979  * dependency structures corresponding to these entries must be updated with
2980  * the new offsets.
2981  */
2982 
2983 /*
2984  * This routine is called after the in-memory inode's link
2985  * count has been incremented, but before the directory entry's
2986  * pointer to the inode has been set.
2987  */
2988 int
2989 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
2990 	struct buf *bp;		/* buffer containing directory block */
2991 	struct inode *dp;	/* inode for directory */
2992 	off_t diroffset;	/* offset of new entry in directory */
2993 	ino_t newinum;		/* inode referenced by new directory entry */
2994 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
2995 	int isnewblk;		/* entry is in a newly allocated block */
2996 {
2997 	int offset;		/* offset of new entry within directory block */
2998 	ufs_lbn_t lbn;		/* block in directory containing new entry */
2999 	struct fs *fs;
3000 	struct diradd *dap;
3001 	struct allocdirect *adp;
3002 	struct pagedep *pagedep;
3003 	struct inodedep *inodedep;
3004 	struct newdirblk *newdirblk = 0;
3005 	struct mkdir *mkdir1, *mkdir2;
3006 	struct mount *mp;
3007 
3008 	/*
3009 	 * Whiteouts have no dependencies.
3010 	 */
3011 	if (newinum == WINO) {
3012 		if (newdirbp != NULL)
3013 			bdwrite(newdirbp);
3014 		return (0);
3015 	}
3016 	mp = UFSTOVFS(dp->i_ump);
3017 	fs = dp->i_fs;
3018 	lbn = lblkno(fs, diroffset);
3019 	offset = blkoff(fs, diroffset);
3020 	dap = malloc(sizeof(struct diradd), M_DIRADD,
3021 		M_SOFTDEP_FLAGS|M_ZERO);
3022 	workitem_alloc(&dap->da_list, D_DIRADD, mp);
3023 	dap->da_offset = offset;
3024 	dap->da_newinum = newinum;
3025 	dap->da_state = ATTACHED;
3026 	if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) {
3027 		newdirblk = malloc(sizeof(struct newdirblk),
3028 		    M_NEWDIRBLK, M_SOFTDEP_FLAGS);
3029 		workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
3030 	}
3031 	if (newdirbp == NULL) {
3032 		dap->da_state |= DEPCOMPLETE;
3033 		ACQUIRE_LOCK(&lk);
3034 	} else {
3035 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
3036 		mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR,
3037 		    M_SOFTDEP_FLAGS);
3038 		workitem_alloc(&mkdir1->md_list, D_MKDIR, mp);
3039 		mkdir1->md_state = MKDIR_BODY;
3040 		mkdir1->md_diradd = dap;
3041 		mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR,
3042 		    M_SOFTDEP_FLAGS);
3043 		workitem_alloc(&mkdir2->md_list, D_MKDIR, mp);
3044 		mkdir2->md_state = MKDIR_PARENT;
3045 		mkdir2->md_diradd = dap;
3046 		/*
3047 		 * Dependency on "." and ".." being written to disk.
3048 		 */
3049 		mkdir1->md_buf = newdirbp;
3050 		ACQUIRE_LOCK(&lk);
3051 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
3052 		WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
3053 		FREE_LOCK(&lk);
3054 		bdwrite(newdirbp);
3055 		/*
3056 		 * Dependency on link count increase for parent directory
3057 		 */
3058 		ACQUIRE_LOCK(&lk);
3059 		if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0
3060 		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3061 			dap->da_state &= ~MKDIR_PARENT;
3062 			WORKITEM_FREE(mkdir2, D_MKDIR);
3063 		} else {
3064 			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
3065 			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
3066 		}
3067 	}
3068 	/*
3069 	 * Link into parent directory pagedep to await its being written.
3070 	 */
3071 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3072 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3073 	dap->da_pagedep = pagedep;
3074 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
3075 	    da_pdlist);
3076 	/*
3077 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3078 	 * is not yet written. If it is written, do the post-inode write
3079 	 * processing to put it on the id_pendinghd list.
3080 	 */
3081 	(void) inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
3082 	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
3083 		diradd_inode_written(dap, inodedep);
3084 	else
3085 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3086 	if (isnewblk) {
3087 		/*
3088 		 * Directories growing into indirect blocks are rare
3089 		 * enough and the frequency of new block allocation
3090 		 * in those cases even more rare, that we choose not
3091 		 * to bother tracking them. Rather we simply force the
3092 		 * new directory entry to disk.
3093 		 */
3094 		if (lbn >= NDADDR) {
3095 			FREE_LOCK(&lk);
3096 			/*
3097 			 * We only have a new allocation when at the
3098 			 * beginning of a new block, not when we are
3099 			 * expanding into an existing block.
3100 			 */
3101 			if (blkoff(fs, diroffset) == 0)
3102 				return (1);
3103 			return (0);
3104 		}
3105 		/*
3106 		 * We only have a new allocation when at the beginning
3107 		 * of a new fragment, not when we are expanding into an
3108 		 * existing fragment. Also, there is nothing to do if we
3109 		 * are already tracking this block.
3110 		 */
3111 		if (fragoff(fs, diroffset) != 0) {
3112 			FREE_LOCK(&lk);
3113 			return (0);
3114 		}
3115 		if ((pagedep->pd_state & NEWBLOCK) != 0) {
3116 			WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
3117 			FREE_LOCK(&lk);
3118 			return (0);
3119 		}
3120 		/*
3121 		 * Find our associated allocdirect and have it track us.
3122 		 */
3123 		if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0)
3124 			panic("softdep_setup_directory_add: lost inodedep");
3125 		adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst);
3126 		if (adp == NULL || adp->ad_lbn != lbn)
3127 			panic("softdep_setup_directory_add: lost entry");
3128 		pagedep->pd_state |= NEWBLOCK;
3129 		newdirblk->db_pagedep = pagedep;
3130 		WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list);
3131 	}
3132 	FREE_LOCK(&lk);
3133 	return (0);
3134 }
3135 
3136 /*
3137  * This procedure is called to change the offset of a directory
3138  * entry when compacting a directory block which must be owned
3139  * exclusively by the caller. Note that the actual entry movement
3140  * must be done in this procedure to ensure that no I/O completions
3141  * occur while the move is in progress.
3142  */
3143 void
3144 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
3145 	struct inode *dp;	/* inode for directory */
3146 	caddr_t base;		/* address of dp->i_offset */
3147 	caddr_t oldloc;		/* address of old directory location */
3148 	caddr_t newloc;		/* address of new directory location */
3149 	int entrysize;		/* size of directory entry */
3150 {
3151 	int offset, oldoffset, newoffset;
3152 	struct pagedep *pagedep;
3153 	struct diradd *dap;
3154 	ufs_lbn_t lbn;
3155 
3156 	ACQUIRE_LOCK(&lk);
3157 	lbn = lblkno(dp->i_fs, dp->i_offset);
3158 	offset = blkoff(dp->i_fs, dp->i_offset);
3159 	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
3160 		goto done;
3161 	oldoffset = offset + (oldloc - base);
3162 	newoffset = offset + (newloc - base);
3163 
3164 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
3165 		if (dap->da_offset != oldoffset)
3166 			continue;
3167 		dap->da_offset = newoffset;
3168 		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
3169 			break;
3170 		LIST_REMOVE(dap, da_pdlist);
3171 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
3172 		    dap, da_pdlist);
3173 		break;
3174 	}
3175 	if (dap == NULL) {
3176 
3177 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
3178 			if (dap->da_offset == oldoffset) {
3179 				dap->da_offset = newoffset;
3180 				break;
3181 			}
3182 		}
3183 	}
3184 done:
3185 	bcopy(oldloc, newloc, entrysize);
3186 	FREE_LOCK(&lk);
3187 }
3188 
3189 /*
3190  * Free a diradd dependency structure. This routine must be called
3191  * with splbio interrupts blocked.
3192  */
3193 static void
3194 free_diradd(dap)
3195 	struct diradd *dap;
3196 {
3197 	struct dirrem *dirrem;
3198 	struct pagedep *pagedep;
3199 	struct inodedep *inodedep;
3200 	struct mkdir *mkdir, *nextmd;
3201 
3202 	mtx_assert(&lk, MA_OWNED);
3203 	WORKLIST_REMOVE(&dap->da_list);
3204 	LIST_REMOVE(dap, da_pdlist);
3205 	if ((dap->da_state & DIRCHG) == 0) {
3206 		pagedep = dap->da_pagedep;
3207 	} else {
3208 		dirrem = dap->da_previous;
3209 		pagedep = dirrem->dm_pagedep;
3210 		dirrem->dm_dirinum = pagedep->pd_ino;
3211 		add_to_worklist(&dirrem->dm_list);
3212 	}
3213 	if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum,
3214 	    0, &inodedep) != 0)
3215 		(void) free_inodedep(inodedep);
3216 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
3217 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
3218 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
3219 			if (mkdir->md_diradd != dap)
3220 				continue;
3221 			dap->da_state &= ~mkdir->md_state;
3222 			WORKLIST_REMOVE(&mkdir->md_list);
3223 			LIST_REMOVE(mkdir, md_mkdirs);
3224 			WORKITEM_FREE(mkdir, D_MKDIR);
3225 		}
3226 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
3227 			panic("free_diradd: unfound ref");
3228 	}
3229 	WORKITEM_FREE(dap, D_DIRADD);
3230 }
3231 
3232 /*
3233  * Directory entry removal dependencies.
3234  *
3235  * When removing a directory entry, the entry's inode pointer must be
3236  * zero'ed on disk before the corresponding inode's link count is decremented
3237  * (possibly freeing the inode for re-use). This dependency is handled by
3238  * updating the directory entry but delaying the inode count reduction until
3239  * after the directory block has been written to disk. After this point, the
3240  * inode count can be decremented whenever it is convenient.
3241  */
3242 
3243 /*
3244  * This routine should be called immediately after removing
3245  * a directory entry.  The inode's link count should not be
3246  * decremented by the calling procedure -- the soft updates
3247  * code will do this task when it is safe.
3248  */
3249 void
3250 softdep_setup_remove(bp, dp, ip, isrmdir)
3251 	struct buf *bp;		/* buffer containing directory block */
3252 	struct inode *dp;	/* inode for the directory being modified */
3253 	struct inode *ip;	/* inode for directory entry being removed */
3254 	int isrmdir;		/* indicates if doing RMDIR */
3255 {
3256 	struct dirrem *dirrem, *prevdirrem;
3257 
3258 	/*
3259 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
3260 	 */
3261 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3262 
3263 	/*
3264 	 * If the COMPLETE flag is clear, then there were no active
3265 	 * entries and we want to roll back to a zeroed entry until
3266 	 * the new inode is committed to disk. If the COMPLETE flag is
3267 	 * set then we have deleted an entry that never made it to
3268 	 * disk. If the entry we deleted resulted from a name change,
3269 	 * then the old name still resides on disk. We cannot delete
3270 	 * its inode (returned to us in prevdirrem) until the zeroed
3271 	 * directory entry gets to disk. The new inode has never been
3272 	 * referenced on the disk, so can be deleted immediately.
3273 	 */
3274 	if ((dirrem->dm_state & COMPLETE) == 0) {
3275 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
3276 		    dm_next);
3277 		FREE_LOCK(&lk);
3278 	} else {
3279 		if (prevdirrem != NULL)
3280 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
3281 			    prevdirrem, dm_next);
3282 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
3283 		FREE_LOCK(&lk);
3284 		handle_workitem_remove(dirrem, NULL);
3285 	}
3286 }
3287 
3288 /*
3289  * Allocate a new dirrem if appropriate and return it along with
3290  * its associated pagedep. Called without a lock, returns with lock.
3291  */
3292 static long num_dirrem;		/* number of dirrem allocated */
3293 static struct dirrem *
3294 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
3295 	struct buf *bp;		/* buffer containing directory block */
3296 	struct inode *dp;	/* inode for the directory being modified */
3297 	struct inode *ip;	/* inode for directory entry being removed */
3298 	int isrmdir;		/* indicates if doing RMDIR */
3299 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
3300 {
3301 	int offset;
3302 	ufs_lbn_t lbn;
3303 	struct diradd *dap;
3304 	struct dirrem *dirrem;
3305 	struct pagedep *pagedep;
3306 
3307 	/*
3308 	 * Whiteouts have no deletion dependencies.
3309 	 */
3310 	if (ip == NULL)
3311 		panic("newdirrem: whiteout");
3312 	/*
3313 	 * If we are over our limit, try to improve the situation.
3314 	 * Limiting the number of dirrem structures will also limit
3315 	 * the number of freefile and freeblks structures.
3316 	 */
3317 	ACQUIRE_LOCK(&lk);
3318 	if (!(ip->i_flags & SF_SNAPSHOT) && num_dirrem > max_softdeps / 2)
3319 		(void) request_cleanup(ITOV(dp)->v_mount, FLUSH_REMOVE);
3320 	num_dirrem += 1;
3321 	FREE_LOCK(&lk);
3322 	dirrem = malloc(sizeof(struct dirrem),
3323 		M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO);
3324 	workitem_alloc(&dirrem->dm_list, D_DIRREM, ITOV(dp)->v_mount);
3325 	dirrem->dm_state = isrmdir ? RMDIR : 0;
3326 	dirrem->dm_oldinum = ip->i_number;
3327 	*prevdirremp = NULL;
3328 
3329 	ACQUIRE_LOCK(&lk);
3330 	lbn = lblkno(dp->i_fs, dp->i_offset);
3331 	offset = blkoff(dp->i_fs, dp->i_offset);
3332 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3333 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3334 	dirrem->dm_pagedep = pagedep;
3335 	/*
3336 	 * Check for a diradd dependency for the same directory entry.
3337 	 * If present, then both dependencies become obsolete and can
3338 	 * be de-allocated. Check for an entry on both the pd_dirraddhd
3339 	 * list and the pd_pendinghd list.
3340 	 */
3341 
3342 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
3343 		if (dap->da_offset == offset)
3344 			break;
3345 	if (dap == NULL) {
3346 
3347 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
3348 			if (dap->da_offset == offset)
3349 				break;
3350 		if (dap == NULL)
3351 			return (dirrem);
3352 	}
3353 	/*
3354 	 * Must be ATTACHED at this point.
3355 	 */
3356 	if ((dap->da_state & ATTACHED) == 0)
3357 		panic("newdirrem: not ATTACHED");
3358 	if (dap->da_newinum != ip->i_number)
3359 		panic("newdirrem: inum %d should be %d",
3360 		    ip->i_number, dap->da_newinum);
3361 	/*
3362 	 * If we are deleting a changed name that never made it to disk,
3363 	 * then return the dirrem describing the previous inode (which
3364 	 * represents the inode currently referenced from this entry on disk).
3365 	 */
3366 	if ((dap->da_state & DIRCHG) != 0) {
3367 		*prevdirremp = dap->da_previous;
3368 		dap->da_state &= ~DIRCHG;
3369 		dap->da_pagedep = pagedep;
3370 	}
3371 	/*
3372 	 * We are deleting an entry that never made it to disk.
3373 	 * Mark it COMPLETE so we can delete its inode immediately.
3374 	 */
3375 	dirrem->dm_state |= COMPLETE;
3376 	free_diradd(dap);
3377 	return (dirrem);
3378 }
3379 
3380 /*
3381  * Directory entry change dependencies.
3382  *
3383  * Changing an existing directory entry requires that an add operation
3384  * be completed first followed by a deletion. The semantics for the addition
3385  * are identical to the description of adding a new entry above except
3386  * that the rollback is to the old inode number rather than zero. Once
3387  * the addition dependency is completed, the removal is done as described
3388  * in the removal routine above.
3389  */
3390 
3391 /*
3392  * This routine should be called immediately after changing
3393  * a directory entry.  The inode's link count should not be
3394  * decremented by the calling procedure -- the soft updates
3395  * code will perform this task when it is safe.
3396  */
3397 void
3398 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
3399 	struct buf *bp;		/* buffer containing directory block */
3400 	struct inode *dp;	/* inode for the directory being modified */
3401 	struct inode *ip;	/* inode for directory entry being removed */
3402 	ino_t newinum;		/* new inode number for changed entry */
3403 	int isrmdir;		/* indicates if doing RMDIR */
3404 {
3405 	int offset;
3406 	struct diradd *dap = NULL;
3407 	struct dirrem *dirrem, *prevdirrem;
3408 	struct pagedep *pagedep;
3409 	struct inodedep *inodedep;
3410 	struct mount *mp;
3411 
3412 	offset = blkoff(dp->i_fs, dp->i_offset);
3413 	mp = UFSTOVFS(dp->i_ump);
3414 
3415 	/*
3416 	 * Whiteouts do not need diradd dependencies.
3417 	 */
3418 	if (newinum != WINO) {
3419 		dap = malloc(sizeof(struct diradd),
3420 		    M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
3421 		workitem_alloc(&dap->da_list, D_DIRADD, mp);
3422 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
3423 		dap->da_offset = offset;
3424 		dap->da_newinum = newinum;
3425 	}
3426 
3427 	/*
3428 	 * Allocate a new dirrem and ACQUIRE_LOCK.
3429 	 */
3430 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3431 	pagedep = dirrem->dm_pagedep;
3432 	/*
3433 	 * The possible values for isrmdir:
3434 	 *	0 - non-directory file rename
3435 	 *	1 - directory rename within same directory
3436 	 *   inum - directory rename to new directory of given inode number
3437 	 * When renaming to a new directory, we are both deleting and
3438 	 * creating a new directory entry, so the link count on the new
3439 	 * directory should not change. Thus we do not need the followup
3440 	 * dirrem which is usually done in handle_workitem_remove. We set
3441 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
3442 	 * followup dirrem.
3443 	 */
3444 	if (isrmdir > 1)
3445 		dirrem->dm_state |= DIRCHG;
3446 
3447 	/*
3448 	 * Whiteouts have no additional dependencies,
3449 	 * so just put the dirrem on the correct list.
3450 	 */
3451 	if (newinum == WINO) {
3452 		if ((dirrem->dm_state & COMPLETE) == 0) {
3453 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
3454 			    dm_next);
3455 		} else {
3456 			dirrem->dm_dirinum = pagedep->pd_ino;
3457 			add_to_worklist(&dirrem->dm_list);
3458 		}
3459 		FREE_LOCK(&lk);
3460 		return;
3461 	}
3462 
3463 	/*
3464 	 * If the COMPLETE flag is clear, then there were no active
3465 	 * entries and we want to roll back to the previous inode until
3466 	 * the new inode is committed to disk. If the COMPLETE flag is
3467 	 * set, then we have deleted an entry that never made it to disk.
3468 	 * If the entry we deleted resulted from a name change, then the old
3469 	 * inode reference still resides on disk. Any rollback that we do
3470 	 * needs to be to that old inode (returned to us in prevdirrem). If
3471 	 * the entry we deleted resulted from a create, then there is
3472 	 * no entry on the disk, so we want to roll back to zero rather
3473 	 * than the uncommitted inode. In either of the COMPLETE cases we
3474 	 * want to immediately free the unwritten and unreferenced inode.
3475 	 */
3476 	if ((dirrem->dm_state & COMPLETE) == 0) {
3477 		dap->da_previous = dirrem;
3478 	} else {
3479 		if (prevdirrem != NULL) {
3480 			dap->da_previous = prevdirrem;
3481 		} else {
3482 			dap->da_state &= ~DIRCHG;
3483 			dap->da_pagedep = pagedep;
3484 		}
3485 		dirrem->dm_dirinum = pagedep->pd_ino;
3486 		add_to_worklist(&dirrem->dm_list);
3487 	}
3488 	/*
3489 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3490 	 * is not yet written. If it is written, do the post-inode write
3491 	 * processing to put it on the id_pendinghd list.
3492 	 */
3493 	if (inodedep_lookup(mp, newinum, DEPALLOC, &inodedep) == 0 ||
3494 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3495 		dap->da_state |= COMPLETE;
3496 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3497 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3498 	} else {
3499 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
3500 		    dap, da_pdlist);
3501 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3502 	}
3503 	FREE_LOCK(&lk);
3504 }
3505 
3506 /*
3507  * Called whenever the link count on an inode is changed.
3508  * It creates an inode dependency so that the new reference(s)
3509  * to the inode cannot be committed to disk until the updated
3510  * inode has been written.
3511  */
3512 void
3513 softdep_change_linkcnt(ip)
3514 	struct inode *ip;	/* the inode with the increased link count */
3515 {
3516 	struct inodedep *inodedep;
3517 
3518 	ACQUIRE_LOCK(&lk);
3519 	(void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number,
3520 	    DEPALLOC, &inodedep);
3521 	if (ip->i_nlink < ip->i_effnlink)
3522 		panic("softdep_change_linkcnt: bad delta");
3523 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3524 	FREE_LOCK(&lk);
3525 }
3526 
3527 /*
3528  * Called when the effective link count and the reference count
3529  * on an inode drops to zero. At this point there are no names
3530  * referencing the file in the filesystem and no active file
3531  * references. The space associated with the file will be freed
3532  * as soon as the necessary soft dependencies are cleared.
3533  */
3534 void
3535 softdep_releasefile(ip)
3536 	struct inode *ip;	/* inode with the zero effective link count */
3537 {
3538 	struct inodedep *inodedep;
3539 	struct fs *fs;
3540 	int extblocks;
3541 
3542 	if (ip->i_effnlink > 0)
3543 		panic("softdep_releasefile: file still referenced");
3544 	/*
3545 	 * We may be called several times as the on-disk link count
3546 	 * drops to zero. We only want to account for the space once.
3547 	 */
3548 	if (ip->i_flag & IN_SPACECOUNTED)
3549 		return;
3550 	/*
3551 	 * We have to deactivate a snapshot otherwise copyonwrites may
3552 	 * add blocks and the cleanup may remove blocks after we have
3553 	 * tried to account for them.
3554 	 */
3555 	if ((ip->i_flags & SF_SNAPSHOT) != 0)
3556 		ffs_snapremove(ITOV(ip));
3557 	/*
3558 	 * If we are tracking an nlinkdelta, we have to also remember
3559 	 * whether we accounted for the freed space yet.
3560 	 */
3561 	ACQUIRE_LOCK(&lk);
3562 	if ((inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, &inodedep)))
3563 		inodedep->id_state |= SPACECOUNTED;
3564 	FREE_LOCK(&lk);
3565 	fs = ip->i_fs;
3566 	extblocks = 0;
3567 	if (fs->fs_magic == FS_UFS2_MAGIC)
3568 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
3569 	UFS_LOCK(ip->i_ump);
3570 	ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks;
3571 	ip->i_fs->fs_pendinginodes += 1;
3572 	UFS_UNLOCK(ip->i_ump);
3573 	ip->i_flag |= IN_SPACECOUNTED;
3574 }
3575 
3576 /*
3577  * This workitem decrements the inode's link count.
3578  * If the link count reaches zero, the file is removed.
3579  */
3580 static void
3581 handle_workitem_remove(dirrem, xp)
3582 	struct dirrem *dirrem;
3583 	struct vnode *xp;
3584 {
3585 	struct thread *td = curthread;
3586 	struct inodedep *inodedep;
3587 	struct vnode *vp;
3588 	struct inode *ip;
3589 	ino_t oldinum;
3590 	int error;
3591 
3592 	if ((vp = xp) == NULL &&
3593 	    (error = ffs_vgetf(dirrem->dm_list.wk_mp,
3594 		    dirrem->dm_oldinum, LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ)) != 0) {
3595 		softdep_error("handle_workitem_remove: vget", error);
3596 		return;
3597 	}
3598 	ip = VTOI(vp);
3599 	ACQUIRE_LOCK(&lk);
3600 	if ((inodedep_lookup(dirrem->dm_list.wk_mp,
3601 	    dirrem->dm_oldinum, 0, &inodedep)) == 0)
3602 		panic("handle_workitem_remove: lost inodedep");
3603 	/*
3604 	 * Normal file deletion.
3605 	 */
3606 	if ((dirrem->dm_state & RMDIR) == 0) {
3607 		ip->i_nlink--;
3608 		DIP_SET(ip, i_nlink, ip->i_nlink);
3609 		ip->i_flag |= IN_CHANGE;
3610 		if (ip->i_nlink < ip->i_effnlink)
3611 			panic("handle_workitem_remove: bad file delta");
3612 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3613 		num_dirrem -= 1;
3614 		WORKITEM_FREE(dirrem, D_DIRREM);
3615 		FREE_LOCK(&lk);
3616 		vput(vp);
3617 		return;
3618 	}
3619 	/*
3620 	 * Directory deletion. Decrement reference count for both the
3621 	 * just deleted parent directory entry and the reference for ".".
3622 	 * Next truncate the directory to length zero. When the
3623 	 * truncation completes, arrange to have the reference count on
3624 	 * the parent decremented to account for the loss of "..".
3625 	 */
3626 	ip->i_nlink -= 2;
3627 	DIP_SET(ip, i_nlink, ip->i_nlink);
3628 	ip->i_flag |= IN_CHANGE;
3629 	if (ip->i_nlink < ip->i_effnlink)
3630 		panic("handle_workitem_remove: bad dir delta");
3631 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3632 	FREE_LOCK(&lk);
3633 	if ((error = ffs_truncate(vp, (off_t)0, 0, td->td_ucred, td)) != 0)
3634 		softdep_error("handle_workitem_remove: truncate", error);
3635 	ACQUIRE_LOCK(&lk);
3636 	/*
3637 	 * Rename a directory to a new parent. Since, we are both deleting
3638 	 * and creating a new directory entry, the link count on the new
3639 	 * directory should not change. Thus we skip the followup dirrem.
3640 	 */
3641 	if (dirrem->dm_state & DIRCHG) {
3642 		num_dirrem -= 1;
3643 		WORKITEM_FREE(dirrem, D_DIRREM);
3644 		FREE_LOCK(&lk);
3645 		vput(vp);
3646 		return;
3647 	}
3648 	/*
3649 	 * If the inodedep does not exist, then the zero'ed inode has
3650 	 * been written to disk. If the allocated inode has never been
3651 	 * written to disk, then the on-disk inode is zero'ed. In either
3652 	 * case we can remove the file immediately.
3653 	 */
3654 	dirrem->dm_state = 0;
3655 	oldinum = dirrem->dm_oldinum;
3656 	dirrem->dm_oldinum = dirrem->dm_dirinum;
3657 	if (inodedep_lookup(dirrem->dm_list.wk_mp, oldinum,
3658 	    0, &inodedep) == 0 || check_inode_unwritten(inodedep)) {
3659 		if (xp != NULL)
3660 			add_to_worklist(&dirrem->dm_list);
3661 		FREE_LOCK(&lk);
3662 		vput(vp);
3663 		if (xp == NULL)
3664 			handle_workitem_remove(dirrem, NULL);
3665 		return;
3666 	}
3667 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3668 	FREE_LOCK(&lk);
3669 	ip->i_flag |= IN_CHANGE;
3670 	ffs_update(vp, 0);
3671 	vput(vp);
3672 }
3673 
3674 /*
3675  * Inode de-allocation dependencies.
3676  *
3677  * When an inode's link count is reduced to zero, it can be de-allocated. We
3678  * found it convenient to postpone de-allocation until after the inode is
3679  * written to disk with its new link count (zero).  At this point, all of the
3680  * on-disk inode's block pointers are nullified and, with careful dependency
3681  * list ordering, all dependencies related to the inode will be satisfied and
3682  * the corresponding dependency structures de-allocated.  So, if/when the
3683  * inode is reused, there will be no mixing of old dependencies with new
3684  * ones.  This artificial dependency is set up by the block de-allocation
3685  * procedure above (softdep_setup_freeblocks) and completed by the
3686  * following procedure.
3687  */
3688 static void
3689 handle_workitem_freefile(freefile)
3690 	struct freefile *freefile;
3691 {
3692 	struct fs *fs;
3693 	struct inodedep *idp;
3694 	struct ufsmount *ump;
3695 	int error;
3696 
3697 	ump = VFSTOUFS(freefile->fx_list.wk_mp);
3698 	fs = ump->um_fs;
3699 #ifdef DEBUG
3700 	ACQUIRE_LOCK(&lk);
3701 	error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp);
3702 	FREE_LOCK(&lk);
3703 	if (error)
3704 		panic("handle_workitem_freefile: inodedep survived");
3705 #endif
3706 	UFS_LOCK(ump);
3707 	fs->fs_pendinginodes -= 1;
3708 	UFS_UNLOCK(ump);
3709 	if ((error = ffs_freefile(ump, fs, freefile->fx_devvp,
3710 	    freefile->fx_oldinum, freefile->fx_mode)) != 0)
3711 		softdep_error("handle_workitem_freefile", error);
3712 	ACQUIRE_LOCK(&lk);
3713 	WORKITEM_FREE(freefile, D_FREEFILE);
3714 	FREE_LOCK(&lk);
3715 }
3716 
3717 
3718 /*
3719  * Helper function which unlinks marker element from work list and returns
3720  * the next element on the list.
3721  */
3722 static __inline struct worklist *
3723 markernext(struct worklist *marker)
3724 {
3725 	struct worklist *next;
3726 
3727 	next = LIST_NEXT(marker, wk_list);
3728 	LIST_REMOVE(marker, wk_list);
3729 	return next;
3730 }
3731 
3732 /*
3733  * Disk writes.
3734  *
3735  * The dependency structures constructed above are most actively used when file
3736  * system blocks are written to disk.  No constraints are placed on when a
3737  * block can be written, but unsatisfied update dependencies are made safe by
3738  * modifying (or replacing) the source memory for the duration of the disk
3739  * write.  When the disk write completes, the memory block is again brought
3740  * up-to-date.
3741  *
3742  * In-core inode structure reclamation.
3743  *
3744  * Because there are a finite number of "in-core" inode structures, they are
3745  * reused regularly.  By transferring all inode-related dependencies to the
3746  * in-memory inode block and indexing them separately (via "inodedep"s), we
3747  * can allow "in-core" inode structures to be reused at any time and avoid
3748  * any increase in contention.
3749  *
3750  * Called just before entering the device driver to initiate a new disk I/O.
3751  * The buffer must be locked, thus, no I/O completion operations can occur
3752  * while we are manipulating its associated dependencies.
3753  */
3754 static void
3755 softdep_disk_io_initiation(bp)
3756 	struct buf *bp;		/* structure describing disk write to occur */
3757 {
3758 	struct worklist *wk;
3759 	struct worklist marker;
3760 	struct indirdep *indirdep;
3761 	struct inodedep *inodedep;
3762 
3763 	/*
3764 	 * We only care about write operations. There should never
3765 	 * be dependencies for reads.
3766 	 */
3767 	if (bp->b_iocmd != BIO_WRITE)
3768 		panic("softdep_disk_io_initiation: not write");
3769 
3770 	marker.wk_type = D_LAST + 1;	/* Not a normal workitem */
3771 	PHOLD(curproc);			/* Don't swap out kernel stack */
3772 
3773 	ACQUIRE_LOCK(&lk);
3774 	/*
3775 	 * Do any necessary pre-I/O processing.
3776 	 */
3777 	for (wk = LIST_FIRST(&bp->b_dep); wk != NULL;
3778 	     wk = markernext(&marker)) {
3779 		LIST_INSERT_AFTER(wk, &marker, wk_list);
3780 		switch (wk->wk_type) {
3781 
3782 		case D_PAGEDEP:
3783 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
3784 			continue;
3785 
3786 		case D_INODEDEP:
3787 			inodedep = WK_INODEDEP(wk);
3788 			if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
3789 				initiate_write_inodeblock_ufs1(inodedep, bp);
3790 			else
3791 				initiate_write_inodeblock_ufs2(inodedep, bp);
3792 			continue;
3793 
3794 		case D_INDIRDEP:
3795 			indirdep = WK_INDIRDEP(wk);
3796 			if (indirdep->ir_state & GOINGAWAY)
3797 				panic("disk_io_initiation: indirdep gone");
3798 			/*
3799 			 * If there are no remaining dependencies, this
3800 			 * will be writing the real pointers, so the
3801 			 * dependency can be freed.
3802 			 */
3803 			if (LIST_EMPTY(&indirdep->ir_deplisthd)) {
3804 				struct buf *bp;
3805 
3806 				bp = indirdep->ir_savebp;
3807 				bp->b_flags |= B_INVAL | B_NOCACHE;
3808 				/* inline expand WORKLIST_REMOVE(wk); */
3809 				wk->wk_state &= ~ONWORKLIST;
3810 				LIST_REMOVE(wk, wk_list);
3811 				WORKITEM_FREE(indirdep, D_INDIRDEP);
3812 				FREE_LOCK(&lk);
3813 				brelse(bp);
3814 				ACQUIRE_LOCK(&lk);
3815 				continue;
3816 			}
3817 			/*
3818 			 * Replace up-to-date version with safe version.
3819 			 */
3820 			FREE_LOCK(&lk);
3821 			indirdep->ir_saveddata = malloc(bp->b_bcount,
3822 			    M_INDIRDEP, M_SOFTDEP_FLAGS);
3823 			ACQUIRE_LOCK(&lk);
3824 			indirdep->ir_state &= ~ATTACHED;
3825 			indirdep->ir_state |= UNDONE;
3826 			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3827 			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3828 			    bp->b_bcount);
3829 			continue;
3830 
3831 		case D_MKDIR:
3832 		case D_BMSAFEMAP:
3833 		case D_ALLOCDIRECT:
3834 		case D_ALLOCINDIR:
3835 			continue;
3836 
3837 		default:
3838 			panic("handle_disk_io_initiation: Unexpected type %s",
3839 			    TYPENAME(wk->wk_type));
3840 			/* NOTREACHED */
3841 		}
3842 	}
3843 	FREE_LOCK(&lk);
3844 	PRELE(curproc);			/* Allow swapout of kernel stack */
3845 }
3846 
3847 /*
3848  * Called from within the procedure above to deal with unsatisfied
3849  * allocation dependencies in a directory. The buffer must be locked,
3850  * thus, no I/O completion operations can occur while we are
3851  * manipulating its associated dependencies.
3852  */
3853 static void
3854 initiate_write_filepage(pagedep, bp)
3855 	struct pagedep *pagedep;
3856 	struct buf *bp;
3857 {
3858 	struct diradd *dap;
3859 	struct direct *ep;
3860 	int i;
3861 
3862 	if (pagedep->pd_state & IOSTARTED) {
3863 		/*
3864 		 * This can only happen if there is a driver that does not
3865 		 * understand chaining. Here biodone will reissue the call
3866 		 * to strategy for the incomplete buffers.
3867 		 */
3868 		printf("initiate_write_filepage: already started\n");
3869 		return;
3870 	}
3871 	pagedep->pd_state |= IOSTARTED;
3872 	for (i = 0; i < DAHASHSZ; i++) {
3873 		LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3874 			ep = (struct direct *)
3875 			    ((char *)bp->b_data + dap->da_offset);
3876 			if (ep->d_ino != dap->da_newinum)
3877 				panic("%s: dir inum %d != new %d",
3878 				    "initiate_write_filepage",
3879 				    ep->d_ino, dap->da_newinum);
3880 			if (dap->da_state & DIRCHG)
3881 				ep->d_ino = dap->da_previous->dm_oldinum;
3882 			else
3883 				ep->d_ino = 0;
3884 			dap->da_state &= ~ATTACHED;
3885 			dap->da_state |= UNDONE;
3886 		}
3887 	}
3888 }
3889 
3890 /*
3891  * Version of initiate_write_inodeblock that handles UFS1 dinodes.
3892  * Note that any bug fixes made to this routine must be done in the
3893  * version found below.
3894  *
3895  * Called from within the procedure above to deal with unsatisfied
3896  * allocation dependencies in an inodeblock. The buffer must be
3897  * locked, thus, no I/O completion operations can occur while we
3898  * are manipulating its associated dependencies.
3899  */
3900 static void
3901 initiate_write_inodeblock_ufs1(inodedep, bp)
3902 	struct inodedep *inodedep;
3903 	struct buf *bp;			/* The inode block */
3904 {
3905 	struct allocdirect *adp, *lastadp;
3906 	struct ufs1_dinode *dp;
3907 	struct ufs1_dinode *sip;
3908 	struct fs *fs;
3909 	ufs_lbn_t i;
3910 #ifdef INVARIANTS
3911 	ufs_lbn_t prevlbn = 0;
3912 #endif
3913 	int deplist;
3914 
3915 	if (inodedep->id_state & IOSTARTED)
3916 		panic("initiate_write_inodeblock_ufs1: already started");
3917 	inodedep->id_state |= IOSTARTED;
3918 	fs = inodedep->id_fs;
3919 	dp = (struct ufs1_dinode *)bp->b_data +
3920 	    ino_to_fsbo(fs, inodedep->id_ino);
3921 	/*
3922 	 * If the bitmap is not yet written, then the allocated
3923 	 * inode cannot be written to disk.
3924 	 */
3925 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3926 		if (inodedep->id_savedino1 != NULL)
3927 			panic("initiate_write_inodeblock_ufs1: I/O underway");
3928 		FREE_LOCK(&lk);
3929 		sip = malloc(sizeof(struct ufs1_dinode),
3930 		    M_SAVEDINO, M_SOFTDEP_FLAGS);
3931 		ACQUIRE_LOCK(&lk);
3932 		inodedep->id_savedino1 = sip;
3933 		*inodedep->id_savedino1 = *dp;
3934 		bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
3935 		dp->di_gen = inodedep->id_savedino1->di_gen;
3936 		return;
3937 	}
3938 	/*
3939 	 * If no dependencies, then there is nothing to roll back.
3940 	 */
3941 	inodedep->id_savedsize = dp->di_size;
3942 	inodedep->id_savedextsize = 0;
3943 	if (TAILQ_EMPTY(&inodedep->id_inoupdt))
3944 		return;
3945 	/*
3946 	 * Set the dependencies to busy.
3947 	 */
3948 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3949 	     adp = TAILQ_NEXT(adp, ad_next)) {
3950 #ifdef INVARIANTS
3951 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
3952 			panic("softdep_write_inodeblock: lbn order");
3953 		prevlbn = adp->ad_lbn;
3954 		if (adp->ad_lbn < NDADDR &&
3955 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno)
3956 			panic("%s: direct pointer #%jd mismatch %d != %jd",
3957 			    "softdep_write_inodeblock",
3958 			    (intmax_t)adp->ad_lbn,
3959 			    dp->di_db[adp->ad_lbn],
3960 			    (intmax_t)adp->ad_newblkno);
3961 		if (adp->ad_lbn >= NDADDR &&
3962 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno)
3963 			panic("%s: indirect pointer #%jd mismatch %d != %jd",
3964 			    "softdep_write_inodeblock",
3965 			    (intmax_t)adp->ad_lbn - NDADDR,
3966 			    dp->di_ib[adp->ad_lbn - NDADDR],
3967 			    (intmax_t)adp->ad_newblkno);
3968 		deplist |= 1 << adp->ad_lbn;
3969 		if ((adp->ad_state & ATTACHED) == 0)
3970 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3971 			    adp->ad_state);
3972 #endif /* INVARIANTS */
3973 		adp->ad_state &= ~ATTACHED;
3974 		adp->ad_state |= UNDONE;
3975 	}
3976 	/*
3977 	 * The on-disk inode cannot claim to be any larger than the last
3978 	 * fragment that has been written. Otherwise, the on-disk inode
3979 	 * might have fragments that were not the last block in the file
3980 	 * which would corrupt the filesystem.
3981 	 */
3982 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3983 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3984 		if (adp->ad_lbn >= NDADDR)
3985 			break;
3986 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3987 		/* keep going until hitting a rollback to a frag */
3988 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3989 			continue;
3990 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3991 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3992 #ifdef INVARIANTS
3993 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
3994 				panic("softdep_write_inodeblock: lost dep1");
3995 #endif /* INVARIANTS */
3996 			dp->di_db[i] = 0;
3997 		}
3998 		for (i = 0; i < NIADDR; i++) {
3999 #ifdef INVARIANTS
4000 			if (dp->di_ib[i] != 0 &&
4001 			    (deplist & ((1 << NDADDR) << i)) == 0)
4002 				panic("softdep_write_inodeblock: lost dep2");
4003 #endif /* INVARIANTS */
4004 			dp->di_ib[i] = 0;
4005 		}
4006 		return;
4007 	}
4008 	/*
4009 	 * If we have zero'ed out the last allocated block of the file,
4010 	 * roll back the size to the last currently allocated block.
4011 	 * We know that this last allocated block is a full-sized as
4012 	 * we already checked for fragments in the loop above.
4013 	 */
4014 	if (lastadp != NULL &&
4015 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
4016 		for (i = lastadp->ad_lbn; i >= 0; i--)
4017 			if (dp->di_db[i] != 0)
4018 				break;
4019 		dp->di_size = (i + 1) * fs->fs_bsize;
4020 	}
4021 	/*
4022 	 * The only dependencies are for indirect blocks.
4023 	 *
4024 	 * The file size for indirect block additions is not guaranteed.
4025 	 * Such a guarantee would be non-trivial to achieve. The conventional
4026 	 * synchronous write implementation also does not make this guarantee.
4027 	 * Fsck should catch and fix discrepancies. Arguably, the file size
4028 	 * can be over-estimated without destroying integrity when the file
4029 	 * moves into the indirect blocks (i.e., is large). If we want to
4030 	 * postpone fsck, we are stuck with this argument.
4031 	 */
4032 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
4033 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
4034 }
4035 
4036 /*
4037  * Version of initiate_write_inodeblock that handles UFS2 dinodes.
4038  * Note that any bug fixes made to this routine must be done in the
4039  * version found above.
4040  *
4041  * Called from within the procedure above to deal with unsatisfied
4042  * allocation dependencies in an inodeblock. The buffer must be
4043  * locked, thus, no I/O completion operations can occur while we
4044  * are manipulating its associated dependencies.
4045  */
4046 static void
4047 initiate_write_inodeblock_ufs2(inodedep, bp)
4048 	struct inodedep *inodedep;
4049 	struct buf *bp;			/* The inode block */
4050 {
4051 	struct allocdirect *adp, *lastadp;
4052 	struct ufs2_dinode *dp;
4053 	struct ufs2_dinode *sip;
4054 	struct fs *fs;
4055 	ufs_lbn_t i;
4056 #ifdef INVARIANTS
4057 	ufs_lbn_t prevlbn = 0;
4058 #endif
4059 	int deplist;
4060 
4061 	if (inodedep->id_state & IOSTARTED)
4062 		panic("initiate_write_inodeblock_ufs2: already started");
4063 	inodedep->id_state |= IOSTARTED;
4064 	fs = inodedep->id_fs;
4065 	dp = (struct ufs2_dinode *)bp->b_data +
4066 	    ino_to_fsbo(fs, inodedep->id_ino);
4067 	/*
4068 	 * If the bitmap is not yet written, then the allocated
4069 	 * inode cannot be written to disk.
4070 	 */
4071 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4072 		if (inodedep->id_savedino2 != NULL)
4073 			panic("initiate_write_inodeblock_ufs2: I/O underway");
4074 		FREE_LOCK(&lk);
4075 		sip = malloc(sizeof(struct ufs2_dinode),
4076 		    M_SAVEDINO, M_SOFTDEP_FLAGS);
4077 		ACQUIRE_LOCK(&lk);
4078 		inodedep->id_savedino2 = sip;
4079 		*inodedep->id_savedino2 = *dp;
4080 		bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
4081 		dp->di_gen = inodedep->id_savedino2->di_gen;
4082 		return;
4083 	}
4084 	/*
4085 	 * If no dependencies, then there is nothing to roll back.
4086 	 */
4087 	inodedep->id_savedsize = dp->di_size;
4088 	inodedep->id_savedextsize = dp->di_extsize;
4089 	if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
4090 	    TAILQ_EMPTY(&inodedep->id_extupdt))
4091 		return;
4092 	/*
4093 	 * Set the ext data dependencies to busy.
4094 	 */
4095 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
4096 	     adp = TAILQ_NEXT(adp, ad_next)) {
4097 #ifdef INVARIANTS
4098 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
4099 			panic("softdep_write_inodeblock: lbn order");
4100 		prevlbn = adp->ad_lbn;
4101 		if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno)
4102 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
4103 			    "softdep_write_inodeblock",
4104 			    (intmax_t)adp->ad_lbn,
4105 			    (intmax_t)dp->di_extb[adp->ad_lbn],
4106 			    (intmax_t)adp->ad_newblkno);
4107 		deplist |= 1 << adp->ad_lbn;
4108 		if ((adp->ad_state & ATTACHED) == 0)
4109 			panic("softdep_write_inodeblock: Unknown state 0x%x",
4110 			    adp->ad_state);
4111 #endif /* INVARIANTS */
4112 		adp->ad_state &= ~ATTACHED;
4113 		adp->ad_state |= UNDONE;
4114 	}
4115 	/*
4116 	 * The on-disk inode cannot claim to be any larger than the last
4117 	 * fragment that has been written. Otherwise, the on-disk inode
4118 	 * might have fragments that were not the last block in the ext
4119 	 * data which would corrupt the filesystem.
4120 	 */
4121 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
4122 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
4123 		dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno;
4124 		/* keep going until hitting a rollback to a frag */
4125 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
4126 			continue;
4127 		dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
4128 		for (i = adp->ad_lbn + 1; i < NXADDR; i++) {
4129 #ifdef INVARIANTS
4130 			if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0)
4131 				panic("softdep_write_inodeblock: lost dep1");
4132 #endif /* INVARIANTS */
4133 			dp->di_extb[i] = 0;
4134 		}
4135 		lastadp = NULL;
4136 		break;
4137 	}
4138 	/*
4139 	 * If we have zero'ed out the last allocated block of the ext
4140 	 * data, roll back the size to the last currently allocated block.
4141 	 * We know that this last allocated block is a full-sized as
4142 	 * we already checked for fragments in the loop above.
4143 	 */
4144 	if (lastadp != NULL &&
4145 	    dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
4146 		for (i = lastadp->ad_lbn; i >= 0; i--)
4147 			if (dp->di_extb[i] != 0)
4148 				break;
4149 		dp->di_extsize = (i + 1) * fs->fs_bsize;
4150 	}
4151 	/*
4152 	 * Set the file data dependencies to busy.
4153 	 */
4154 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
4155 	     adp = TAILQ_NEXT(adp, ad_next)) {
4156 #ifdef INVARIANTS
4157 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
4158 			panic("softdep_write_inodeblock: lbn order");
4159 		prevlbn = adp->ad_lbn;
4160 		if (adp->ad_lbn < NDADDR &&
4161 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno)
4162 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
4163 			    "softdep_write_inodeblock",
4164 			    (intmax_t)adp->ad_lbn,
4165 			    (intmax_t)dp->di_db[adp->ad_lbn],
4166 			    (intmax_t)adp->ad_newblkno);
4167 		if (adp->ad_lbn >= NDADDR &&
4168 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno)
4169 			panic("%s indirect pointer #%jd mismatch %jd != %jd",
4170 			    "softdep_write_inodeblock:",
4171 			    (intmax_t)adp->ad_lbn - NDADDR,
4172 			    (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR],
4173 			    (intmax_t)adp->ad_newblkno);
4174 		deplist |= 1 << adp->ad_lbn;
4175 		if ((adp->ad_state & ATTACHED) == 0)
4176 			panic("softdep_write_inodeblock: Unknown state 0x%x",
4177 			    adp->ad_state);
4178 #endif /* INVARIANTS */
4179 		adp->ad_state &= ~ATTACHED;
4180 		adp->ad_state |= UNDONE;
4181 	}
4182 	/*
4183 	 * The on-disk inode cannot claim to be any larger than the last
4184 	 * fragment that has been written. Otherwise, the on-disk inode
4185 	 * might have fragments that were not the last block in the file
4186 	 * which would corrupt the filesystem.
4187 	 */
4188 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
4189 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
4190 		if (adp->ad_lbn >= NDADDR)
4191 			break;
4192 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
4193 		/* keep going until hitting a rollback to a frag */
4194 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
4195 			continue;
4196 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
4197 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
4198 #ifdef INVARIANTS
4199 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
4200 				panic("softdep_write_inodeblock: lost dep2");
4201 #endif /* INVARIANTS */
4202 			dp->di_db[i] = 0;
4203 		}
4204 		for (i = 0; i < NIADDR; i++) {
4205 #ifdef INVARIANTS
4206 			if (dp->di_ib[i] != 0 &&
4207 			    (deplist & ((1 << NDADDR) << i)) == 0)
4208 				panic("softdep_write_inodeblock: lost dep3");
4209 #endif /* INVARIANTS */
4210 			dp->di_ib[i] = 0;
4211 		}
4212 		return;
4213 	}
4214 	/*
4215 	 * If we have zero'ed out the last allocated block of the file,
4216 	 * roll back the size to the last currently allocated block.
4217 	 * We know that this last allocated block is a full-sized as
4218 	 * we already checked for fragments in the loop above.
4219 	 */
4220 	if (lastadp != NULL &&
4221 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
4222 		for (i = lastadp->ad_lbn; i >= 0; i--)
4223 			if (dp->di_db[i] != 0)
4224 				break;
4225 		dp->di_size = (i + 1) * fs->fs_bsize;
4226 	}
4227 	/*
4228 	 * The only dependencies are for indirect blocks.
4229 	 *
4230 	 * The file size for indirect block additions is not guaranteed.
4231 	 * Such a guarantee would be non-trivial to achieve. The conventional
4232 	 * synchronous write implementation also does not make this guarantee.
4233 	 * Fsck should catch and fix discrepancies. Arguably, the file size
4234 	 * can be over-estimated without destroying integrity when the file
4235 	 * moves into the indirect blocks (i.e., is large). If we want to
4236 	 * postpone fsck, we are stuck with this argument.
4237 	 */
4238 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
4239 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
4240 }
4241 
4242 /*
4243  * This routine is called during the completion interrupt
4244  * service routine for a disk write (from the procedure called
4245  * by the device driver to inform the filesystem caches of
4246  * a request completion).  It should be called early in this
4247  * procedure, before the block is made available to other
4248  * processes or other routines are called.
4249  */
4250 static void
4251 softdep_disk_write_complete(bp)
4252 	struct buf *bp;		/* describes the completed disk write */
4253 {
4254 	struct worklist *wk;
4255 	struct worklist *owk;
4256 	struct workhead reattach;
4257 	struct newblk *newblk;
4258 	struct allocindir *aip;
4259 	struct allocdirect *adp;
4260 	struct indirdep *indirdep;
4261 	struct inodedep *inodedep;
4262 	struct bmsafemap *bmsafemap;
4263 
4264 	/*
4265 	 * If an error occurred while doing the write, then the data
4266 	 * has not hit the disk and the dependencies cannot be unrolled.
4267 	 */
4268 	if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0)
4269 		return;
4270 	LIST_INIT(&reattach);
4271 	/*
4272 	 * This lock must not be released anywhere in this code segment.
4273 	 */
4274 	ACQUIRE_LOCK(&lk);
4275 	owk = NULL;
4276 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
4277 		WORKLIST_REMOVE(wk);
4278 		if (wk == owk)
4279 			panic("duplicate worklist: %p\n", wk);
4280 		owk = wk;
4281 		switch (wk->wk_type) {
4282 
4283 		case D_PAGEDEP:
4284 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
4285 				WORKLIST_INSERT(&reattach, wk);
4286 			continue;
4287 
4288 		case D_INODEDEP:
4289 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
4290 				WORKLIST_INSERT(&reattach, wk);
4291 			continue;
4292 
4293 		case D_BMSAFEMAP:
4294 			bmsafemap = WK_BMSAFEMAP(wk);
4295 			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
4296 				newblk->nb_state |= DEPCOMPLETE;
4297 				newblk->nb_bmsafemap = NULL;
4298 				LIST_REMOVE(newblk, nb_deps);
4299 			}
4300 			while ((adp =
4301 			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
4302 				adp->ad_state |= DEPCOMPLETE;
4303 				adp->ad_buf = NULL;
4304 				LIST_REMOVE(adp, ad_deps);
4305 				handle_allocdirect_partdone(adp);
4306 			}
4307 			while ((aip =
4308 			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
4309 				aip->ai_state |= DEPCOMPLETE;
4310 				aip->ai_buf = NULL;
4311 				LIST_REMOVE(aip, ai_deps);
4312 				handle_allocindir_partdone(aip);
4313 			}
4314 			while ((inodedep =
4315 			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
4316 				inodedep->id_state |= DEPCOMPLETE;
4317 				LIST_REMOVE(inodedep, id_deps);
4318 				inodedep->id_buf = NULL;
4319 			}
4320 			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
4321 			continue;
4322 
4323 		case D_MKDIR:
4324 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
4325 			continue;
4326 
4327 		case D_ALLOCDIRECT:
4328 			adp = WK_ALLOCDIRECT(wk);
4329 			adp->ad_state |= COMPLETE;
4330 			handle_allocdirect_partdone(adp);
4331 			continue;
4332 
4333 		case D_ALLOCINDIR:
4334 			aip = WK_ALLOCINDIR(wk);
4335 			aip->ai_state |= COMPLETE;
4336 			handle_allocindir_partdone(aip);
4337 			continue;
4338 
4339 		case D_INDIRDEP:
4340 			indirdep = WK_INDIRDEP(wk);
4341 			if (indirdep->ir_state & GOINGAWAY)
4342 				panic("disk_write_complete: indirdep gone");
4343 			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
4344 			free(indirdep->ir_saveddata, M_INDIRDEP);
4345 			indirdep->ir_saveddata = 0;
4346 			indirdep->ir_state &= ~UNDONE;
4347 			indirdep->ir_state |= ATTACHED;
4348 			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
4349 				handle_allocindir_partdone(aip);
4350 				if (aip == LIST_FIRST(&indirdep->ir_donehd))
4351 					panic("disk_write_complete: not gone");
4352 			}
4353 			WORKLIST_INSERT(&reattach, wk);
4354 			if ((bp->b_flags & B_DELWRI) == 0)
4355 				stat_indir_blk_ptrs++;
4356 			bdirty(bp);
4357 			continue;
4358 
4359 		default:
4360 			panic("handle_disk_write_complete: Unknown type %s",
4361 			    TYPENAME(wk->wk_type));
4362 			/* NOTREACHED */
4363 		}
4364 	}
4365 	/*
4366 	 * Reattach any requests that must be redone.
4367 	 */
4368 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
4369 		WORKLIST_REMOVE(wk);
4370 		WORKLIST_INSERT(&bp->b_dep, wk);
4371 	}
4372 	FREE_LOCK(&lk);
4373 }
4374 
4375 /*
4376  * Called from within softdep_disk_write_complete above. Note that
4377  * this routine is always called from interrupt level with further
4378  * splbio interrupts blocked.
4379  */
4380 static void
4381 handle_allocdirect_partdone(adp)
4382 	struct allocdirect *adp;	/* the completed allocdirect */
4383 {
4384 	struct allocdirectlst *listhead;
4385 	struct allocdirect *listadp;
4386 	struct inodedep *inodedep;
4387 	long bsize, delay;
4388 
4389 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4390 		return;
4391 	if (adp->ad_buf != NULL)
4392 		panic("handle_allocdirect_partdone: dangling dep");
4393 	/*
4394 	 * The on-disk inode cannot claim to be any larger than the last
4395 	 * fragment that has been written. Otherwise, the on-disk inode
4396 	 * might have fragments that were not the last block in the file
4397 	 * which would corrupt the filesystem. Thus, we cannot free any
4398 	 * allocdirects after one whose ad_oldblkno claims a fragment as
4399 	 * these blocks must be rolled back to zero before writing the inode.
4400 	 * We check the currently active set of allocdirects in id_inoupdt
4401 	 * or id_extupdt as appropriate.
4402 	 */
4403 	inodedep = adp->ad_inodedep;
4404 	bsize = inodedep->id_fs->fs_bsize;
4405 	if (adp->ad_state & EXTDATA)
4406 		listhead = &inodedep->id_extupdt;
4407 	else
4408 		listhead = &inodedep->id_inoupdt;
4409 	TAILQ_FOREACH(listadp, listhead, ad_next) {
4410 		/* found our block */
4411 		if (listadp == adp)
4412 			break;
4413 		/* continue if ad_oldlbn is not a fragment */
4414 		if (listadp->ad_oldsize == 0 ||
4415 		    listadp->ad_oldsize == bsize)
4416 			continue;
4417 		/* hit a fragment */
4418 		return;
4419 	}
4420 	/*
4421 	 * If we have reached the end of the current list without
4422 	 * finding the just finished dependency, then it must be
4423 	 * on the future dependency list. Future dependencies cannot
4424 	 * be freed until they are moved to the current list.
4425 	 */
4426 	if (listadp == NULL) {
4427 #ifdef DEBUG
4428 		if (adp->ad_state & EXTDATA)
4429 			listhead = &inodedep->id_newextupdt;
4430 		else
4431 			listhead = &inodedep->id_newinoupdt;
4432 		TAILQ_FOREACH(listadp, listhead, ad_next)
4433 			/* found our block */
4434 			if (listadp == adp)
4435 				break;
4436 		if (listadp == NULL)
4437 			panic("handle_allocdirect_partdone: lost dep");
4438 #endif /* DEBUG */
4439 		return;
4440 	}
4441 	/*
4442 	 * If we have found the just finished dependency, then free
4443 	 * it along with anything that follows it that is complete.
4444 	 * If the inode still has a bitmap dependency, then it has
4445 	 * never been written to disk, hence the on-disk inode cannot
4446 	 * reference the old fragment so we can free it without delay.
4447 	 */
4448 	delay = (inodedep->id_state & DEPCOMPLETE);
4449 	for (; adp; adp = listadp) {
4450 		listadp = TAILQ_NEXT(adp, ad_next);
4451 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4452 			return;
4453 		free_allocdirect(listhead, adp, delay);
4454 	}
4455 }
4456 
4457 /*
4458  * Called from within softdep_disk_write_complete above. Note that
4459  * this routine is always called from interrupt level with further
4460  * splbio interrupts blocked.
4461  */
4462 static void
4463 handle_allocindir_partdone(aip)
4464 	struct allocindir *aip;		/* the completed allocindir */
4465 {
4466 	struct indirdep *indirdep;
4467 
4468 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
4469 		return;
4470 	if (aip->ai_buf != NULL)
4471 		panic("handle_allocindir_partdone: dangling dependency");
4472 	indirdep = aip->ai_indirdep;
4473 	if (indirdep->ir_state & UNDONE) {
4474 		LIST_REMOVE(aip, ai_next);
4475 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
4476 		return;
4477 	}
4478 	if (indirdep->ir_state & UFS1FMT)
4479 		((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4480 		    aip->ai_newblkno;
4481 	else
4482 		((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4483 		    aip->ai_newblkno;
4484 	LIST_REMOVE(aip, ai_next);
4485 	if (aip->ai_freefrag != NULL)
4486 		add_to_worklist(&aip->ai_freefrag->ff_list);
4487 	WORKITEM_FREE(aip, D_ALLOCINDIR);
4488 }
4489 
4490 /*
4491  * Called from within softdep_disk_write_complete above to restore
4492  * in-memory inode block contents to their most up-to-date state. Note
4493  * that this routine is always called from interrupt level with further
4494  * splbio interrupts blocked.
4495  */
4496 static int
4497 handle_written_inodeblock(inodedep, bp)
4498 	struct inodedep *inodedep;
4499 	struct buf *bp;		/* buffer containing the inode block */
4500 {
4501 	struct worklist *wk, *filefree;
4502 	struct allocdirect *adp, *nextadp;
4503 	struct ufs1_dinode *dp1 = NULL;
4504 	struct ufs2_dinode *dp2 = NULL;
4505 	int hadchanges, fstype;
4506 
4507 	if ((inodedep->id_state & IOSTARTED) == 0)
4508 		panic("handle_written_inodeblock: not started");
4509 	inodedep->id_state &= ~IOSTARTED;
4510 	if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
4511 		fstype = UFS1;
4512 		dp1 = (struct ufs1_dinode *)bp->b_data +
4513 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4514 	} else {
4515 		fstype = UFS2;
4516 		dp2 = (struct ufs2_dinode *)bp->b_data +
4517 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4518 	}
4519 	/*
4520 	 * If we had to rollback the inode allocation because of
4521 	 * bitmaps being incomplete, then simply restore it.
4522 	 * Keep the block dirty so that it will not be reclaimed until
4523 	 * all associated dependencies have been cleared and the
4524 	 * corresponding updates written to disk.
4525 	 */
4526 	if (inodedep->id_savedino1 != NULL) {
4527 		if (fstype == UFS1)
4528 			*dp1 = *inodedep->id_savedino1;
4529 		else
4530 			*dp2 = *inodedep->id_savedino2;
4531 		free(inodedep->id_savedino1, M_SAVEDINO);
4532 		inodedep->id_savedino1 = NULL;
4533 		if ((bp->b_flags & B_DELWRI) == 0)
4534 			stat_inode_bitmap++;
4535 		bdirty(bp);
4536 		return (1);
4537 	}
4538 	inodedep->id_state |= COMPLETE;
4539 	/*
4540 	 * Roll forward anything that had to be rolled back before
4541 	 * the inode could be updated.
4542 	 */
4543 	hadchanges = 0;
4544 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
4545 		nextadp = TAILQ_NEXT(adp, ad_next);
4546 		if (adp->ad_state & ATTACHED)
4547 			panic("handle_written_inodeblock: new entry");
4548 		if (fstype == UFS1) {
4549 			if (adp->ad_lbn < NDADDR) {
4550 				if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno)
4551 					panic("%s %s #%jd mismatch %d != %jd",
4552 					    "handle_written_inodeblock:",
4553 					    "direct pointer",
4554 					    (intmax_t)adp->ad_lbn,
4555 					    dp1->di_db[adp->ad_lbn],
4556 					    (intmax_t)adp->ad_oldblkno);
4557 				dp1->di_db[adp->ad_lbn] = adp->ad_newblkno;
4558 			} else {
4559 				if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0)
4560 					panic("%s: %s #%jd allocated as %d",
4561 					    "handle_written_inodeblock",
4562 					    "indirect pointer",
4563 					    (intmax_t)adp->ad_lbn - NDADDR,
4564 					    dp1->di_ib[adp->ad_lbn - NDADDR]);
4565 				dp1->di_ib[adp->ad_lbn - NDADDR] =
4566 				    adp->ad_newblkno;
4567 			}
4568 		} else {
4569 			if (adp->ad_lbn < NDADDR) {
4570 				if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno)
4571 					panic("%s: %s #%jd %s %jd != %jd",
4572 					    "handle_written_inodeblock",
4573 					    "direct pointer",
4574 					    (intmax_t)adp->ad_lbn, "mismatch",
4575 					    (intmax_t)dp2->di_db[adp->ad_lbn],
4576 					    (intmax_t)adp->ad_oldblkno);
4577 				dp2->di_db[adp->ad_lbn] = adp->ad_newblkno;
4578 			} else {
4579 				if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0)
4580 					panic("%s: %s #%jd allocated as %jd",
4581 					    "handle_written_inodeblock",
4582 					    "indirect pointer",
4583 					    (intmax_t)adp->ad_lbn - NDADDR,
4584 					    (intmax_t)
4585 					    dp2->di_ib[adp->ad_lbn - NDADDR]);
4586 				dp2->di_ib[adp->ad_lbn - NDADDR] =
4587 				    adp->ad_newblkno;
4588 			}
4589 		}
4590 		adp->ad_state &= ~UNDONE;
4591 		adp->ad_state |= ATTACHED;
4592 		hadchanges = 1;
4593 	}
4594 	for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
4595 		nextadp = TAILQ_NEXT(adp, ad_next);
4596 		if (adp->ad_state & ATTACHED)
4597 			panic("handle_written_inodeblock: new entry");
4598 		if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno)
4599 			panic("%s: direct pointers #%jd %s %jd != %jd",
4600 			    "handle_written_inodeblock",
4601 			    (intmax_t)adp->ad_lbn, "mismatch",
4602 			    (intmax_t)dp2->di_extb[adp->ad_lbn],
4603 			    (intmax_t)adp->ad_oldblkno);
4604 		dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno;
4605 		adp->ad_state &= ~UNDONE;
4606 		adp->ad_state |= ATTACHED;
4607 		hadchanges = 1;
4608 	}
4609 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
4610 		stat_direct_blk_ptrs++;
4611 	/*
4612 	 * Reset the file size to its most up-to-date value.
4613 	 */
4614 	if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1)
4615 		panic("handle_written_inodeblock: bad size");
4616 	if (fstype == UFS1) {
4617 		if (dp1->di_size != inodedep->id_savedsize) {
4618 			dp1->di_size = inodedep->id_savedsize;
4619 			hadchanges = 1;
4620 		}
4621 	} else {
4622 		if (dp2->di_size != inodedep->id_savedsize) {
4623 			dp2->di_size = inodedep->id_savedsize;
4624 			hadchanges = 1;
4625 		}
4626 		if (dp2->di_extsize != inodedep->id_savedextsize) {
4627 			dp2->di_extsize = inodedep->id_savedextsize;
4628 			hadchanges = 1;
4629 		}
4630 	}
4631 	inodedep->id_savedsize = -1;
4632 	inodedep->id_savedextsize = -1;
4633 	/*
4634 	 * If there were any rollbacks in the inode block, then it must be
4635 	 * marked dirty so that its will eventually get written back in
4636 	 * its correct form.
4637 	 */
4638 	if (hadchanges)
4639 		bdirty(bp);
4640 	/*
4641 	 * Process any allocdirects that completed during the update.
4642 	 */
4643 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
4644 		handle_allocdirect_partdone(adp);
4645 	if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
4646 		handle_allocdirect_partdone(adp);
4647 	/*
4648 	 * Process deallocations that were held pending until the
4649 	 * inode had been written to disk. Freeing of the inode
4650 	 * is delayed until after all blocks have been freed to
4651 	 * avoid creation of new <vfsid, inum, lbn> triples
4652 	 * before the old ones have been deleted.
4653 	 */
4654 	filefree = NULL;
4655 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
4656 		WORKLIST_REMOVE(wk);
4657 		switch (wk->wk_type) {
4658 
4659 		case D_FREEFILE:
4660 			/*
4661 			 * We defer adding filefree to the worklist until
4662 			 * all other additions have been made to ensure
4663 			 * that it will be done after all the old blocks
4664 			 * have been freed.
4665 			 */
4666 			if (filefree != NULL)
4667 				panic("handle_written_inodeblock: filefree");
4668 			filefree = wk;
4669 			continue;
4670 
4671 		case D_MKDIR:
4672 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
4673 			continue;
4674 
4675 		case D_DIRADD:
4676 			diradd_inode_written(WK_DIRADD(wk), inodedep);
4677 			continue;
4678 
4679 		case D_FREEBLKS:
4680 			wk->wk_state |= COMPLETE;
4681 			if ((wk->wk_state  & ALLCOMPLETE) != ALLCOMPLETE)
4682 				continue;
4683 			 /* -- fall through -- */
4684 		case D_FREEFRAG:
4685 		case D_DIRREM:
4686 			add_to_worklist(wk);
4687 			continue;
4688 
4689 		case D_NEWDIRBLK:
4690 			free_newdirblk(WK_NEWDIRBLK(wk));
4691 			continue;
4692 
4693 		default:
4694 			panic("handle_written_inodeblock: Unknown type %s",
4695 			    TYPENAME(wk->wk_type));
4696 			/* NOTREACHED */
4697 		}
4698 	}
4699 	if (filefree != NULL) {
4700 		if (free_inodedep(inodedep) == 0)
4701 			panic("handle_written_inodeblock: live inodedep");
4702 		add_to_worklist(filefree);
4703 		return (0);
4704 	}
4705 
4706 	/*
4707 	 * If no outstanding dependencies, free it.
4708 	 */
4709 	if (free_inodedep(inodedep) ||
4710 	    (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
4711 	     TAILQ_FIRST(&inodedep->id_extupdt) == 0))
4712 		return (0);
4713 	return (hadchanges);
4714 }
4715 
4716 /*
4717  * Process a diradd entry after its dependent inode has been written.
4718  * This routine must be called with splbio interrupts blocked.
4719  */
4720 static void
4721 diradd_inode_written(dap, inodedep)
4722 	struct diradd *dap;
4723 	struct inodedep *inodedep;
4724 {
4725 	struct pagedep *pagedep;
4726 
4727 	dap->da_state |= COMPLETE;
4728 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4729 		if (dap->da_state & DIRCHG)
4730 			pagedep = dap->da_previous->dm_pagedep;
4731 		else
4732 			pagedep = dap->da_pagedep;
4733 		LIST_REMOVE(dap, da_pdlist);
4734 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4735 	}
4736 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
4737 }
4738 
4739 /*
4740  * Handle the completion of a mkdir dependency.
4741  */
4742 static void
4743 handle_written_mkdir(mkdir, type)
4744 	struct mkdir *mkdir;
4745 	int type;
4746 {
4747 	struct diradd *dap;
4748 	struct pagedep *pagedep;
4749 
4750 	if (mkdir->md_state != type)
4751 		panic("handle_written_mkdir: bad type");
4752 	dap = mkdir->md_diradd;
4753 	dap->da_state &= ~type;
4754 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
4755 		dap->da_state |= DEPCOMPLETE;
4756 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4757 		if (dap->da_state & DIRCHG)
4758 			pagedep = dap->da_previous->dm_pagedep;
4759 		else
4760 			pagedep = dap->da_pagedep;
4761 		LIST_REMOVE(dap, da_pdlist);
4762 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4763 	}
4764 	LIST_REMOVE(mkdir, md_mkdirs);
4765 	WORKITEM_FREE(mkdir, D_MKDIR);
4766 }
4767 
4768 /*
4769  * Called from within softdep_disk_write_complete above.
4770  * A write operation was just completed. Removed inodes can
4771  * now be freed and associated block pointers may be committed.
4772  * Note that this routine is always called from interrupt level
4773  * with further splbio interrupts blocked.
4774  */
4775 static int
4776 handle_written_filepage(pagedep, bp)
4777 	struct pagedep *pagedep;
4778 	struct buf *bp;		/* buffer containing the written page */
4779 {
4780 	struct dirrem *dirrem;
4781 	struct diradd *dap, *nextdap;
4782 	struct direct *ep;
4783 	int i, chgs;
4784 
4785 	if ((pagedep->pd_state & IOSTARTED) == 0)
4786 		panic("handle_written_filepage: not started");
4787 	pagedep->pd_state &= ~IOSTARTED;
4788 	/*
4789 	 * Process any directory removals that have been committed.
4790 	 */
4791 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
4792 		LIST_REMOVE(dirrem, dm_next);
4793 		dirrem->dm_dirinum = pagedep->pd_ino;
4794 		add_to_worklist(&dirrem->dm_list);
4795 	}
4796 	/*
4797 	 * Free any directory additions that have been committed.
4798 	 * If it is a newly allocated block, we have to wait until
4799 	 * the on-disk directory inode claims the new block.
4800 	 */
4801 	if ((pagedep->pd_state & NEWBLOCK) == 0)
4802 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
4803 			free_diradd(dap);
4804 	/*
4805 	 * Uncommitted directory entries must be restored.
4806 	 */
4807 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
4808 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
4809 		     dap = nextdap) {
4810 			nextdap = LIST_NEXT(dap, da_pdlist);
4811 			if (dap->da_state & ATTACHED)
4812 				panic("handle_written_filepage: attached");
4813 			ep = (struct direct *)
4814 			    ((char *)bp->b_data + dap->da_offset);
4815 			ep->d_ino = dap->da_newinum;
4816 			dap->da_state &= ~UNDONE;
4817 			dap->da_state |= ATTACHED;
4818 			chgs = 1;
4819 			/*
4820 			 * If the inode referenced by the directory has
4821 			 * been written out, then the dependency can be
4822 			 * moved to the pending list.
4823 			 */
4824 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4825 				LIST_REMOVE(dap, da_pdlist);
4826 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
4827 				    da_pdlist);
4828 			}
4829 		}
4830 	}
4831 	/*
4832 	 * If there were any rollbacks in the directory, then it must be
4833 	 * marked dirty so that its will eventually get written back in
4834 	 * its correct form.
4835 	 */
4836 	if (chgs) {
4837 		if ((bp->b_flags & B_DELWRI) == 0)
4838 			stat_dir_entry++;
4839 		bdirty(bp);
4840 		return (1);
4841 	}
4842 	/*
4843 	 * If we are not waiting for a new directory block to be
4844 	 * claimed by its inode, then the pagedep will be freed.
4845 	 * Otherwise it will remain to track any new entries on
4846 	 * the page in case they are fsync'ed.
4847 	 */
4848 	if ((pagedep->pd_state & NEWBLOCK) == 0) {
4849 		LIST_REMOVE(pagedep, pd_hash);
4850 		WORKITEM_FREE(pagedep, D_PAGEDEP);
4851 	}
4852 	return (0);
4853 }
4854 
4855 /*
4856  * Writing back in-core inode structures.
4857  *
4858  * The filesystem only accesses an inode's contents when it occupies an
4859  * "in-core" inode structure.  These "in-core" structures are separate from
4860  * the page frames used to cache inode blocks.  Only the latter are
4861  * transferred to/from the disk.  So, when the updated contents of the
4862  * "in-core" inode structure are copied to the corresponding in-memory inode
4863  * block, the dependencies are also transferred.  The following procedure is
4864  * called when copying a dirty "in-core" inode to a cached inode block.
4865  */
4866 
4867 /*
4868  * Called when an inode is loaded from disk. If the effective link count
4869  * differed from the actual link count when it was last flushed, then we
4870  * need to ensure that the correct effective link count is put back.
4871  */
4872 void
4873 softdep_load_inodeblock(ip)
4874 	struct inode *ip;	/* the "in_core" copy of the inode */
4875 {
4876 	struct inodedep *inodedep;
4877 
4878 	/*
4879 	 * Check for alternate nlink count.
4880 	 */
4881 	ip->i_effnlink = ip->i_nlink;
4882 	ACQUIRE_LOCK(&lk);
4883 	if (inodedep_lookup(UFSTOVFS(ip->i_ump),
4884 	    ip->i_number, 0, &inodedep) == 0) {
4885 		FREE_LOCK(&lk);
4886 		return;
4887 	}
4888 	ip->i_effnlink -= inodedep->id_nlinkdelta;
4889 	if (inodedep->id_state & SPACECOUNTED)
4890 		ip->i_flag |= IN_SPACECOUNTED;
4891 	FREE_LOCK(&lk);
4892 }
4893 
4894 /*
4895  * This routine is called just before the "in-core" inode
4896  * information is to be copied to the in-memory inode block.
4897  * Recall that an inode block contains several inodes. If
4898  * the force flag is set, then the dependencies will be
4899  * cleared so that the update can always be made. Note that
4900  * the buffer is locked when this routine is called, so we
4901  * will never be in the middle of writing the inode block
4902  * to disk.
4903  */
4904 void
4905 softdep_update_inodeblock(ip, bp, waitfor)
4906 	struct inode *ip;	/* the "in_core" copy of the inode */
4907 	struct buf *bp;		/* the buffer containing the inode block */
4908 	int waitfor;		/* nonzero => update must be allowed */
4909 {
4910 	struct inodedep *inodedep;
4911 	struct worklist *wk;
4912 	struct mount *mp;
4913 	struct buf *ibp;
4914 	int error;
4915 
4916 	/*
4917 	 * If the effective link count is not equal to the actual link
4918 	 * count, then we must track the difference in an inodedep while
4919 	 * the inode is (potentially) tossed out of the cache. Otherwise,
4920 	 * if there is no existing inodedep, then there are no dependencies
4921 	 * to track.
4922 	 */
4923 	mp = UFSTOVFS(ip->i_ump);
4924 	ACQUIRE_LOCK(&lk);
4925 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
4926 		FREE_LOCK(&lk);
4927 		if (ip->i_effnlink != ip->i_nlink)
4928 			panic("softdep_update_inodeblock: bad link count");
4929 		return;
4930 	}
4931 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
4932 		panic("softdep_update_inodeblock: bad delta");
4933 	/*
4934 	 * Changes have been initiated. Anything depending on these
4935 	 * changes cannot occur until this inode has been written.
4936 	 */
4937 	inodedep->id_state &= ~COMPLETE;
4938 	if ((inodedep->id_state & ONWORKLIST) == 0)
4939 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
4940 	/*
4941 	 * Any new dependencies associated with the incore inode must
4942 	 * now be moved to the list associated with the buffer holding
4943 	 * the in-memory copy of the inode. Once merged process any
4944 	 * allocdirects that are completed by the merger.
4945 	 */
4946 	merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
4947 	if (!TAILQ_EMPTY(&inodedep->id_inoupdt))
4948 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
4949 	merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
4950 	if (!TAILQ_EMPTY(&inodedep->id_extupdt))
4951 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt));
4952 	/*
4953 	 * Now that the inode has been pushed into the buffer, the
4954 	 * operations dependent on the inode being written to disk
4955 	 * can be moved to the id_bufwait so that they will be
4956 	 * processed when the buffer I/O completes.
4957 	 */
4958 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
4959 		WORKLIST_REMOVE(wk);
4960 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
4961 	}
4962 	/*
4963 	 * Newly allocated inodes cannot be written until the bitmap
4964 	 * that allocates them have been written (indicated by
4965 	 * DEPCOMPLETE being set in id_state). If we are doing a
4966 	 * forced sync (e.g., an fsync on a file), we force the bitmap
4967 	 * to be written so that the update can be done.
4968 	 */
4969 	if (waitfor == 0) {
4970 		FREE_LOCK(&lk);
4971 		return;
4972 	}
4973 retry:
4974 	if ((inodedep->id_state & DEPCOMPLETE) != 0) {
4975 		FREE_LOCK(&lk);
4976 		return;
4977 	}
4978 	ibp = inodedep->id_buf;
4979 	ibp = getdirtybuf(ibp, &lk, MNT_WAIT);
4980 	if (ibp == NULL) {
4981 		/*
4982 		 * If ibp came back as NULL, the dependency could have been
4983 		 * freed while we slept.  Look it up again, and check to see
4984 		 * that it has completed.
4985 		 */
4986 		if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
4987 			goto retry;
4988 		FREE_LOCK(&lk);
4989 		return;
4990 	}
4991 	FREE_LOCK(&lk);
4992 	if ((error = bwrite(ibp)) != 0)
4993 		softdep_error("softdep_update_inodeblock: bwrite", error);
4994 }
4995 
4996 /*
4997  * Merge the a new inode dependency list (such as id_newinoupdt) into an
4998  * old inode dependency list (such as id_inoupdt). This routine must be
4999  * called with splbio interrupts blocked.
5000  */
5001 static void
5002 merge_inode_lists(newlisthead, oldlisthead)
5003 	struct allocdirectlst *newlisthead;
5004 	struct allocdirectlst *oldlisthead;
5005 {
5006 	struct allocdirect *listadp, *newadp;
5007 
5008 	newadp = TAILQ_FIRST(newlisthead);
5009 	for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
5010 		if (listadp->ad_lbn < newadp->ad_lbn) {
5011 			listadp = TAILQ_NEXT(listadp, ad_next);
5012 			continue;
5013 		}
5014 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
5015 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
5016 		if (listadp->ad_lbn == newadp->ad_lbn) {
5017 			allocdirect_merge(oldlisthead, newadp,
5018 			    listadp);
5019 			listadp = newadp;
5020 		}
5021 		newadp = TAILQ_FIRST(newlisthead);
5022 	}
5023 	while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
5024 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
5025 		TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
5026 	}
5027 }
5028 
5029 /*
5030  * If we are doing an fsync, then we must ensure that any directory
5031  * entries for the inode have been written after the inode gets to disk.
5032  */
5033 int
5034 softdep_fsync(vp)
5035 	struct vnode *vp;	/* the "in_core" copy of the inode */
5036 {
5037 	struct inodedep *inodedep;
5038 	struct pagedep *pagedep;
5039 	struct worklist *wk;
5040 	struct diradd *dap;
5041 	struct mount *mp;
5042 	struct vnode *pvp;
5043 	struct inode *ip;
5044 	struct buf *bp;
5045 	struct fs *fs;
5046 	struct thread *td = curthread;
5047 	int error, flushparent, pagedep_new_block;
5048 	ino_t parentino;
5049 	ufs_lbn_t lbn;
5050 
5051 	ip = VTOI(vp);
5052 	fs = ip->i_fs;
5053 	mp = vp->v_mount;
5054 	ACQUIRE_LOCK(&lk);
5055 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
5056 		FREE_LOCK(&lk);
5057 		return (0);
5058 	}
5059 	if (!LIST_EMPTY(&inodedep->id_inowait) ||
5060 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
5061 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
5062 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
5063 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
5064 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt))
5065 		panic("softdep_fsync: pending ops");
5066 	for (error = 0, flushparent = 0; ; ) {
5067 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
5068 			break;
5069 		if (wk->wk_type != D_DIRADD)
5070 			panic("softdep_fsync: Unexpected type %s",
5071 			    TYPENAME(wk->wk_type));
5072 		dap = WK_DIRADD(wk);
5073 		/*
5074 		 * Flush our parent if this directory entry has a MKDIR_PARENT
5075 		 * dependency or is contained in a newly allocated block.
5076 		 */
5077 		if (dap->da_state & DIRCHG)
5078 			pagedep = dap->da_previous->dm_pagedep;
5079 		else
5080 			pagedep = dap->da_pagedep;
5081 		parentino = pagedep->pd_ino;
5082 		lbn = pagedep->pd_lbn;
5083 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
5084 			panic("softdep_fsync: dirty");
5085 		if ((dap->da_state & MKDIR_PARENT) ||
5086 		    (pagedep->pd_state & NEWBLOCK))
5087 			flushparent = 1;
5088 		else
5089 			flushparent = 0;
5090 		/*
5091 		 * If we are being fsync'ed as part of vgone'ing this vnode,
5092 		 * then we will not be able to release and recover the
5093 		 * vnode below, so we just have to give up on writing its
5094 		 * directory entry out. It will eventually be written, just
5095 		 * not now, but then the user was not asking to have it
5096 		 * written, so we are not breaking any promises.
5097 		 */
5098 		if (vp->v_iflag & VI_DOOMED)
5099 			break;
5100 		/*
5101 		 * We prevent deadlock by always fetching inodes from the
5102 		 * root, moving down the directory tree. Thus, when fetching
5103 		 * our parent directory, we first try to get the lock. If
5104 		 * that fails, we must unlock ourselves before requesting
5105 		 * the lock on our parent. See the comment in ufs_lookup
5106 		 * for details on possible races.
5107 		 */
5108 		FREE_LOCK(&lk);
5109 		if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp,
5110 		    FFSV_FORCEINSMQ)) {
5111 			error = vfs_busy(mp, MBF_NOWAIT);
5112 			if (error != 0) {
5113 				vfs_ref(mp);
5114 				VOP_UNLOCK(vp, 0);
5115 				error = vfs_busy(mp, 0);
5116 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5117 				vfs_rel(mp);
5118 				if (error != 0)
5119 					return (ENOENT);
5120 				if (vp->v_iflag & VI_DOOMED) {
5121 					vfs_unbusy(mp);
5122 					return (ENOENT);
5123 				}
5124 			}
5125 			VOP_UNLOCK(vp, 0);
5126 			error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE,
5127 			    &pvp, FFSV_FORCEINSMQ);
5128 			vfs_unbusy(mp);
5129 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5130 			if (vp->v_iflag & VI_DOOMED) {
5131 				if (error == 0)
5132 					vput(pvp);
5133 				error = ENOENT;
5134 			}
5135 			if (error != 0)
5136 				return (error);
5137 		}
5138 		/*
5139 		 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
5140 		 * that are contained in direct blocks will be resolved by
5141 		 * doing a ffs_update. Pagedeps contained in indirect blocks
5142 		 * may require a complete sync'ing of the directory. So, we
5143 		 * try the cheap and fast ffs_update first, and if that fails,
5144 		 * then we do the slower ffs_syncvnode of the directory.
5145 		 */
5146 		if (flushparent) {
5147 			int locked;
5148 
5149 			if ((error = ffs_update(pvp, 1)) != 0) {
5150 				vput(pvp);
5151 				return (error);
5152 			}
5153 			ACQUIRE_LOCK(&lk);
5154 			locked = 1;
5155 			if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) {
5156 				if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) {
5157 					if (wk->wk_type != D_DIRADD)
5158 						panic("softdep_fsync: Unexpected type %s",
5159 						      TYPENAME(wk->wk_type));
5160 					dap = WK_DIRADD(wk);
5161 					if (dap->da_state & DIRCHG)
5162 						pagedep = dap->da_previous->dm_pagedep;
5163 					else
5164 						pagedep = dap->da_pagedep;
5165 					pagedep_new_block = pagedep->pd_state & NEWBLOCK;
5166 					FREE_LOCK(&lk);
5167 					locked = 0;
5168 					if (pagedep_new_block &&
5169 					    (error = ffs_syncvnode(pvp, MNT_WAIT))) {
5170 						vput(pvp);
5171 						return (error);
5172 					}
5173 				}
5174 			}
5175 			if (locked)
5176 				FREE_LOCK(&lk);
5177 		}
5178 		/*
5179 		 * Flush directory page containing the inode's name.
5180 		 */
5181 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
5182 		    &bp);
5183 		if (error == 0)
5184 			error = bwrite(bp);
5185 		else
5186 			brelse(bp);
5187 		vput(pvp);
5188 		if (error != 0)
5189 			return (error);
5190 		ACQUIRE_LOCK(&lk);
5191 		if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
5192 			break;
5193 	}
5194 	FREE_LOCK(&lk);
5195 	return (0);
5196 }
5197 
5198 /*
5199  * Flush all the dirty bitmaps associated with the block device
5200  * before flushing the rest of the dirty blocks so as to reduce
5201  * the number of dependencies that will have to be rolled back.
5202  */
5203 void
5204 softdep_fsync_mountdev(vp)
5205 	struct vnode *vp;
5206 {
5207 	struct buf *bp, *nbp;
5208 	struct worklist *wk;
5209 	struct bufobj *bo;
5210 
5211 	if (!vn_isdisk(vp, NULL))
5212 		panic("softdep_fsync_mountdev: vnode not a disk");
5213 	bo = &vp->v_bufobj;
5214 restart:
5215 	BO_LOCK(bo);
5216 	ACQUIRE_LOCK(&lk);
5217 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
5218 		/*
5219 		 * If it is already scheduled, skip to the next buffer.
5220 		 */
5221 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
5222 			continue;
5223 
5224 		if ((bp->b_flags & B_DELWRI) == 0)
5225 			panic("softdep_fsync_mountdev: not dirty");
5226 		/*
5227 		 * We are only interested in bitmaps with outstanding
5228 		 * dependencies.
5229 		 */
5230 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
5231 		    wk->wk_type != D_BMSAFEMAP ||
5232 		    (bp->b_vflags & BV_BKGRDINPROG)) {
5233 			BUF_UNLOCK(bp);
5234 			continue;
5235 		}
5236 		FREE_LOCK(&lk);
5237 		BO_UNLOCK(bo);
5238 		bremfree(bp);
5239 		(void) bawrite(bp);
5240 		goto restart;
5241 	}
5242 	FREE_LOCK(&lk);
5243 	drain_output(vp);
5244 	BO_UNLOCK(bo);
5245 }
5246 
5247 /*
5248  * This routine is called when we are trying to synchronously flush a
5249  * file. This routine must eliminate any filesystem metadata dependencies
5250  * so that the syncing routine can succeed by pushing the dirty blocks
5251  * associated with the file. If any I/O errors occur, they are returned.
5252  */
5253 int
5254 softdep_sync_metadata(struct vnode *vp)
5255 {
5256 	struct pagedep *pagedep;
5257 	struct allocdirect *adp;
5258 	struct allocindir *aip;
5259 	struct buf *bp, *nbp;
5260 	struct worklist *wk;
5261 	struct bufobj *bo;
5262 	int i, error, waitfor;
5263 
5264 	if (!DOINGSOFTDEP(vp))
5265 		return (0);
5266 	/*
5267 	 * Ensure that any direct block dependencies have been cleared.
5268 	 */
5269 	ACQUIRE_LOCK(&lk);
5270 	if ((error = flush_inodedep_deps(vp->v_mount, VTOI(vp)->i_number))) {
5271 		FREE_LOCK(&lk);
5272 		return (error);
5273 	}
5274 	FREE_LOCK(&lk);
5275 	/*
5276 	 * For most files, the only metadata dependencies are the
5277 	 * cylinder group maps that allocate their inode or blocks.
5278 	 * The block allocation dependencies can be found by traversing
5279 	 * the dependency lists for any buffers that remain on their
5280 	 * dirty buffer list. The inode allocation dependency will
5281 	 * be resolved when the inode is updated with MNT_WAIT.
5282 	 * This work is done in two passes. The first pass grabs most
5283 	 * of the buffers and begins asynchronously writing them. The
5284 	 * only way to wait for these asynchronous writes is to sleep
5285 	 * on the filesystem vnode which may stay busy for a long time
5286 	 * if the filesystem is active. So, instead, we make a second
5287 	 * pass over the dependencies blocking on each write. In the
5288 	 * usual case we will be blocking against a write that we
5289 	 * initiated, so when it is done the dependency will have been
5290 	 * resolved. Thus the second pass is expected to end quickly.
5291 	 */
5292 	waitfor = MNT_NOWAIT;
5293 	bo = &vp->v_bufobj;
5294 
5295 top:
5296 	/*
5297 	 * We must wait for any I/O in progress to finish so that
5298 	 * all potential buffers on the dirty list will be visible.
5299 	 */
5300 	BO_LOCK(bo);
5301 	drain_output(vp);
5302 	while ((bp = TAILQ_FIRST(&bo->bo_dirty.bv_hd)) != NULL) {
5303 		bp = getdirtybuf(bp, BO_MTX(bo), MNT_WAIT);
5304 		if (bp)
5305 			break;
5306 	}
5307 	BO_UNLOCK(bo);
5308 	if (bp == NULL)
5309 		return (0);
5310 loop:
5311 	/* While syncing snapshots, we must allow recursive lookups */
5312 	BUF_AREC(bp);
5313 	ACQUIRE_LOCK(&lk);
5314 	/*
5315 	 * As we hold the buffer locked, none of its dependencies
5316 	 * will disappear.
5317 	 */
5318 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5319 		switch (wk->wk_type) {
5320 
5321 		case D_ALLOCDIRECT:
5322 			adp = WK_ALLOCDIRECT(wk);
5323 			if (adp->ad_state & DEPCOMPLETE)
5324 				continue;
5325 			nbp = adp->ad_buf;
5326 			nbp = getdirtybuf(nbp, &lk, waitfor);
5327 			if (nbp == NULL)
5328 				continue;
5329 			FREE_LOCK(&lk);
5330 			if (waitfor == MNT_NOWAIT) {
5331 				bawrite(nbp);
5332 			} else if ((error = bwrite(nbp)) != 0) {
5333 				break;
5334 			}
5335 			ACQUIRE_LOCK(&lk);
5336 			continue;
5337 
5338 		case D_ALLOCINDIR:
5339 			aip = WK_ALLOCINDIR(wk);
5340 			if (aip->ai_state & DEPCOMPLETE)
5341 				continue;
5342 			nbp = aip->ai_buf;
5343 			nbp = getdirtybuf(nbp, &lk, waitfor);
5344 			if (nbp == NULL)
5345 				continue;
5346 			FREE_LOCK(&lk);
5347 			if (waitfor == MNT_NOWAIT) {
5348 				bawrite(nbp);
5349 			} else if ((error = bwrite(nbp)) != 0) {
5350 				break;
5351 			}
5352 			ACQUIRE_LOCK(&lk);
5353 			continue;
5354 
5355 		case D_INDIRDEP:
5356 		restart:
5357 
5358 			LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
5359 				if (aip->ai_state & DEPCOMPLETE)
5360 					continue;
5361 				nbp = aip->ai_buf;
5362 				nbp = getdirtybuf(nbp, &lk, MNT_WAIT);
5363 				if (nbp == NULL)
5364 					goto restart;
5365 				FREE_LOCK(&lk);
5366 				if ((error = bwrite(nbp)) != 0) {
5367 					goto loop_end;
5368 				}
5369 				ACQUIRE_LOCK(&lk);
5370 				goto restart;
5371 			}
5372 			continue;
5373 
5374 		case D_INODEDEP:
5375 			if ((error = flush_inodedep_deps(wk->wk_mp,
5376 			    WK_INODEDEP(wk)->id_ino)) != 0) {
5377 				FREE_LOCK(&lk);
5378 				break;
5379 			}
5380 			continue;
5381 
5382 		case D_PAGEDEP:
5383 			/*
5384 			 * We are trying to sync a directory that may
5385 			 * have dependencies on both its own metadata
5386 			 * and/or dependencies on the inodes of any
5387 			 * recently allocated files. We walk its diradd
5388 			 * lists pushing out the associated inode.
5389 			 */
5390 			pagedep = WK_PAGEDEP(wk);
5391 			for (i = 0; i < DAHASHSZ; i++) {
5392 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
5393 					continue;
5394 				if ((error =
5395 				    flush_pagedep_deps(vp, wk->wk_mp,
5396 						&pagedep->pd_diraddhd[i]))) {
5397 					FREE_LOCK(&lk);
5398 					goto loop_end;
5399 				}
5400 			}
5401 			continue;
5402 
5403 		case D_MKDIR:
5404 			/*
5405 			 * This case should never happen if the vnode has
5406 			 * been properly sync'ed. However, if this function
5407 			 * is used at a place where the vnode has not yet
5408 			 * been sync'ed, this dependency can show up. So,
5409 			 * rather than panic, just flush it.
5410 			 */
5411 			nbp = WK_MKDIR(wk)->md_buf;
5412 			nbp = getdirtybuf(nbp, &lk, waitfor);
5413 			if (nbp == NULL)
5414 				continue;
5415 			FREE_LOCK(&lk);
5416 			if (waitfor == MNT_NOWAIT) {
5417 				bawrite(nbp);
5418 			} else if ((error = bwrite(nbp)) != 0) {
5419 				break;
5420 			}
5421 			ACQUIRE_LOCK(&lk);
5422 			continue;
5423 
5424 		case D_BMSAFEMAP:
5425 			/*
5426 			 * This case should never happen if the vnode has
5427 			 * been properly sync'ed. However, if this function
5428 			 * is used at a place where the vnode has not yet
5429 			 * been sync'ed, this dependency can show up. So,
5430 			 * rather than panic, just flush it.
5431 			 */
5432 			nbp = WK_BMSAFEMAP(wk)->sm_buf;
5433 			nbp = getdirtybuf(nbp, &lk, waitfor);
5434 			if (nbp == NULL)
5435 				continue;
5436 			FREE_LOCK(&lk);
5437 			if (waitfor == MNT_NOWAIT) {
5438 				bawrite(nbp);
5439 			} else if ((error = bwrite(nbp)) != 0) {
5440 				break;
5441 			}
5442 			ACQUIRE_LOCK(&lk);
5443 			continue;
5444 
5445 		default:
5446 			panic("softdep_sync_metadata: Unknown type %s",
5447 			    TYPENAME(wk->wk_type));
5448 			/* NOTREACHED */
5449 		}
5450 	loop_end:
5451 		/* We reach here only in error and unlocked */
5452 		if (error == 0)
5453 			panic("softdep_sync_metadata: zero error");
5454 		BUF_NOREC(bp);
5455 		bawrite(bp);
5456 		return (error);
5457 	}
5458 	FREE_LOCK(&lk);
5459 	BO_LOCK(bo);
5460 	while ((nbp = TAILQ_NEXT(bp, b_bobufs)) != NULL) {
5461 		nbp = getdirtybuf(nbp, BO_MTX(bo), MNT_WAIT);
5462 		if (nbp)
5463 			break;
5464 	}
5465 	BO_UNLOCK(bo);
5466 	BUF_NOREC(bp);
5467 	bawrite(bp);
5468 	if (nbp != NULL) {
5469 		bp = nbp;
5470 		goto loop;
5471 	}
5472 	/*
5473 	 * The brief unlock is to allow any pent up dependency
5474 	 * processing to be done. Then proceed with the second pass.
5475 	 */
5476 	if (waitfor == MNT_NOWAIT) {
5477 		waitfor = MNT_WAIT;
5478 		goto top;
5479 	}
5480 
5481 	/*
5482 	 * If we have managed to get rid of all the dirty buffers,
5483 	 * then we are done. For certain directories and block
5484 	 * devices, we may need to do further work.
5485 	 *
5486 	 * We must wait for any I/O in progress to finish so that
5487 	 * all potential buffers on the dirty list will be visible.
5488 	 */
5489 	BO_LOCK(bo);
5490 	drain_output(vp);
5491 	BO_UNLOCK(bo);
5492 	return (0);
5493 }
5494 
5495 /*
5496  * Flush the dependencies associated with an inodedep.
5497  * Called with splbio blocked.
5498  */
5499 static int
5500 flush_inodedep_deps(mp, ino)
5501 	struct mount *mp;
5502 	ino_t ino;
5503 {
5504 	struct inodedep *inodedep;
5505 	int error, waitfor;
5506 
5507 	/*
5508 	 * This work is done in two passes. The first pass grabs most
5509 	 * of the buffers and begins asynchronously writing them. The
5510 	 * only way to wait for these asynchronous writes is to sleep
5511 	 * on the filesystem vnode which may stay busy for a long time
5512 	 * if the filesystem is active. So, instead, we make a second
5513 	 * pass over the dependencies blocking on each write. In the
5514 	 * usual case we will be blocking against a write that we
5515 	 * initiated, so when it is done the dependency will have been
5516 	 * resolved. Thus the second pass is expected to end quickly.
5517 	 * We give a brief window at the top of the loop to allow
5518 	 * any pending I/O to complete.
5519 	 */
5520 	for (error = 0, waitfor = MNT_NOWAIT; ; ) {
5521 		if (error)
5522 			return (error);
5523 		FREE_LOCK(&lk);
5524 		ACQUIRE_LOCK(&lk);
5525 		if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
5526 			return (0);
5527 		if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
5528 		    flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
5529 		    flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
5530 		    flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
5531 			continue;
5532 		/*
5533 		 * If pass2, we are done, otherwise do pass 2.
5534 		 */
5535 		if (waitfor == MNT_WAIT)
5536 			break;
5537 		waitfor = MNT_WAIT;
5538 	}
5539 	/*
5540 	 * Try freeing inodedep in case all dependencies have been removed.
5541 	 */
5542 	if (inodedep_lookup(mp, ino, 0, &inodedep) != 0)
5543 		(void) free_inodedep(inodedep);
5544 	return (0);
5545 }
5546 
5547 /*
5548  * Flush an inode dependency list.
5549  * Called with splbio blocked.
5550  */
5551 static int
5552 flush_deplist(listhead, waitfor, errorp)
5553 	struct allocdirectlst *listhead;
5554 	int waitfor;
5555 	int *errorp;
5556 {
5557 	struct allocdirect *adp;
5558 	struct buf *bp;
5559 
5560 	mtx_assert(&lk, MA_OWNED);
5561 	TAILQ_FOREACH(adp, listhead, ad_next) {
5562 		if (adp->ad_state & DEPCOMPLETE)
5563 			continue;
5564 		bp = adp->ad_buf;
5565 		bp = getdirtybuf(bp, &lk, waitfor);
5566 		if (bp == NULL) {
5567 			if (waitfor == MNT_NOWAIT)
5568 				continue;
5569 			return (1);
5570 		}
5571 		FREE_LOCK(&lk);
5572 		if (waitfor == MNT_NOWAIT) {
5573 			bawrite(bp);
5574 		} else if ((*errorp = bwrite(bp)) != 0) {
5575 			ACQUIRE_LOCK(&lk);
5576 			return (1);
5577 		}
5578 		ACQUIRE_LOCK(&lk);
5579 		return (1);
5580 	}
5581 	return (0);
5582 }
5583 
5584 /*
5585  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
5586  * Called with splbio blocked.
5587  */
5588 static int
5589 flush_pagedep_deps(pvp, mp, diraddhdp)
5590 	struct vnode *pvp;
5591 	struct mount *mp;
5592 	struct diraddhd *diraddhdp;
5593 {
5594 	struct inodedep *inodedep;
5595 	struct ufsmount *ump;
5596 	struct diradd *dap;
5597 	struct vnode *vp;
5598 	struct bufobj *bo;
5599 	int error = 0;
5600 	struct buf *bp;
5601 	ino_t inum;
5602 	struct worklist *wk;
5603 
5604 	ump = VFSTOUFS(mp);
5605 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
5606 		/*
5607 		 * Flush ourselves if this directory entry
5608 		 * has a MKDIR_PARENT dependency.
5609 		 */
5610 		if (dap->da_state & MKDIR_PARENT) {
5611 			FREE_LOCK(&lk);
5612 			if ((error = ffs_update(pvp, 1)) != 0)
5613 				break;
5614 			ACQUIRE_LOCK(&lk);
5615 			/*
5616 			 * If that cleared dependencies, go on to next.
5617 			 */
5618 			if (dap != LIST_FIRST(diraddhdp))
5619 				continue;
5620 			if (dap->da_state & MKDIR_PARENT)
5621 				panic("flush_pagedep_deps: MKDIR_PARENT");
5622 		}
5623 		/*
5624 		 * A newly allocated directory must have its "." and
5625 		 * ".." entries written out before its name can be
5626 		 * committed in its parent. We do not want or need
5627 		 * the full semantics of a synchronous ffs_syncvnode as
5628 		 * that may end up here again, once for each directory
5629 		 * level in the filesystem. Instead, we push the blocks
5630 		 * and wait for them to clear. We have to fsync twice
5631 		 * because the first call may choose to defer blocks
5632 		 * that still have dependencies, but deferral will
5633 		 * happen at most once.
5634 		 */
5635 		inum = dap->da_newinum;
5636 		if (dap->da_state & MKDIR_BODY) {
5637 			FREE_LOCK(&lk);
5638 			if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
5639 			    FFSV_FORCEINSMQ)))
5640 				break;
5641 			if ((error=ffs_syncvnode(vp, MNT_NOWAIT)) ||
5642 			    (error=ffs_syncvnode(vp, MNT_NOWAIT))) {
5643 				vput(vp);
5644 				break;
5645 			}
5646 			bo = &vp->v_bufobj;
5647 			BO_LOCK(bo);
5648 			drain_output(vp);
5649 			/*
5650 			 * If first block is still dirty with a D_MKDIR
5651 			 * dependency then it needs to be written now.
5652 			 */
5653 			for (;;) {
5654 				error = 0;
5655 				bp = gbincore(bo, 0);
5656 				if (bp == NULL)
5657 					break;	/* First block not present */
5658 				error = BUF_LOCK(bp,
5659 						 LK_EXCLUSIVE |
5660 						 LK_SLEEPFAIL |
5661 						 LK_INTERLOCK,
5662 						 BO_MTX(bo));
5663 				BO_LOCK(bo);
5664 				if (error == ENOLCK)
5665 					continue;	/* Slept, retry */
5666 				if (error != 0)
5667 					break;		/* Failed */
5668 				if ((bp->b_flags & B_DELWRI) == 0) {
5669 					BUF_UNLOCK(bp);
5670 					break;	/* Buffer not dirty */
5671 				}
5672 				for (wk = LIST_FIRST(&bp->b_dep);
5673 				     wk != NULL;
5674 				     wk = LIST_NEXT(wk, wk_list))
5675 					if (wk->wk_type == D_MKDIR)
5676 						break;
5677 				if (wk == NULL)
5678 					BUF_UNLOCK(bp);	/* Dependency gone */
5679 				else {
5680 					/*
5681 					 * D_MKDIR dependency remains,
5682 					 * must write buffer to stable
5683 					 * storage.
5684 					 */
5685 					BO_UNLOCK(bo);
5686 					bremfree(bp);
5687 					error = bwrite(bp);
5688 					BO_LOCK(bo);
5689 				}
5690 				break;
5691 			}
5692 			BO_UNLOCK(bo);
5693 			vput(vp);
5694 			if (error != 0)
5695 				break;	/* Flushing of first block failed */
5696 			ACQUIRE_LOCK(&lk);
5697 			/*
5698 			 * If that cleared dependencies, go on to next.
5699 			 */
5700 			if (dap != LIST_FIRST(diraddhdp))
5701 				continue;
5702 			if (dap->da_state & MKDIR_BODY)
5703 				panic("flush_pagedep_deps: MKDIR_BODY");
5704 		}
5705 		/*
5706 		 * Flush the inode on which the directory entry depends.
5707 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
5708 		 * the only remaining dependency is that the updated inode
5709 		 * count must get pushed to disk. The inode has already
5710 		 * been pushed into its inode buffer (via VOP_UPDATE) at
5711 		 * the time of the reference count change. So we need only
5712 		 * locate that buffer, ensure that there will be no rollback
5713 		 * caused by a bitmap dependency, then write the inode buffer.
5714 		 */
5715 retry:
5716 		if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
5717 			panic("flush_pagedep_deps: lost inode");
5718 		/*
5719 		 * If the inode still has bitmap dependencies,
5720 		 * push them to disk.
5721 		 */
5722 		if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5723 			bp = inodedep->id_buf;
5724 			bp = getdirtybuf(bp, &lk, MNT_WAIT);
5725 			if (bp == NULL)
5726 				goto retry;
5727 			FREE_LOCK(&lk);
5728 			if ((error = bwrite(bp)) != 0)
5729 				break;
5730 			ACQUIRE_LOCK(&lk);
5731 			if (dap != LIST_FIRST(diraddhdp))
5732 				continue;
5733 		}
5734 		/*
5735 		 * If the inode is still sitting in a buffer waiting
5736 		 * to be written, push it to disk.
5737 		 */
5738 		FREE_LOCK(&lk);
5739 		if ((error = bread(ump->um_devvp,
5740 		    fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
5741 		    (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) {
5742 			brelse(bp);
5743 			break;
5744 		}
5745 		if ((error = bwrite(bp)) != 0)
5746 			break;
5747 		ACQUIRE_LOCK(&lk);
5748 		/*
5749 		 * If we have failed to get rid of all the dependencies
5750 		 * then something is seriously wrong.
5751 		 */
5752 		if (dap == LIST_FIRST(diraddhdp))
5753 			panic("flush_pagedep_deps: flush failed");
5754 	}
5755 	if (error)
5756 		ACQUIRE_LOCK(&lk);
5757 	return (error);
5758 }
5759 
5760 /*
5761  * A large burst of file addition or deletion activity can drive the
5762  * memory load excessively high. First attempt to slow things down
5763  * using the techniques below. If that fails, this routine requests
5764  * the offending operations to fall back to running synchronously
5765  * until the memory load returns to a reasonable level.
5766  */
5767 int
5768 softdep_slowdown(vp)
5769 	struct vnode *vp;
5770 {
5771 	int max_softdeps_hard;
5772 
5773 	ACQUIRE_LOCK(&lk);
5774 	max_softdeps_hard = max_softdeps * 11 / 10;
5775 	if (num_dirrem < max_softdeps_hard / 2 &&
5776 	    num_inodedep < max_softdeps_hard &&
5777 	    VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps &&
5778 	    num_freeblkdep < max_softdeps_hard) {
5779 		FREE_LOCK(&lk);
5780   		return (0);
5781 	}
5782 	if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps)
5783 		softdep_speedup();
5784 	stat_sync_limit_hit += 1;
5785 	FREE_LOCK(&lk);
5786 	return (1);
5787 }
5788 
5789 /*
5790  * Called by the allocation routines when they are about to fail
5791  * in the hope that we can free up some disk space.
5792  *
5793  * First check to see if the work list has anything on it. If it has,
5794  * clean up entries until we successfully free some space. Because this
5795  * process holds inodes locked, we cannot handle any remove requests
5796  * that might block on a locked inode as that could lead to deadlock.
5797  * If the worklist yields no free space, encourage the syncer daemon
5798  * to help us. In no event will we try for longer than tickdelay seconds.
5799  */
5800 int
5801 softdep_request_cleanup(fs, vp)
5802 	struct fs *fs;
5803 	struct vnode *vp;
5804 {
5805 	struct ufsmount *ump;
5806 	long starttime;
5807 	ufs2_daddr_t needed;
5808 	int error;
5809 
5810 	ump = VTOI(vp)->i_ump;
5811 	mtx_assert(UFS_MTX(ump), MA_OWNED);
5812 	needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize;
5813 	starttime = time_second + tickdelay;
5814 	/*
5815 	 * If we are being called because of a process doing a
5816 	 * copy-on-write, then it is not safe to update the vnode
5817 	 * as we may recurse into the copy-on-write routine.
5818 	 */
5819 	if (!(curthread->td_pflags & TDP_COWINPROGRESS)) {
5820 		UFS_UNLOCK(ump);
5821 		error = ffs_update(vp, 1);
5822 		UFS_LOCK(ump);
5823 		if (error != 0)
5824 			return (0);
5825 	}
5826 	while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) {
5827 		if (time_second > starttime)
5828 			return (0);
5829 		UFS_UNLOCK(ump);
5830 		ACQUIRE_LOCK(&lk);
5831 		if (ump->softdep_on_worklist > 0 &&
5832 		    process_worklist_item(UFSTOVFS(ump), LK_NOWAIT) != -1) {
5833 			stat_worklist_push += 1;
5834 			FREE_LOCK(&lk);
5835 			UFS_LOCK(ump);
5836 			continue;
5837 		}
5838 		request_cleanup(UFSTOVFS(ump), FLUSH_REMOVE_WAIT);
5839 		FREE_LOCK(&lk);
5840 		UFS_LOCK(ump);
5841 	}
5842 	return (1);
5843 }
5844 
5845 /*
5846  * If memory utilization has gotten too high, deliberately slow things
5847  * down and speed up the I/O processing.
5848  */
5849 extern struct thread *syncertd;
5850 static int
5851 request_cleanup(mp, resource)
5852 	struct mount *mp;
5853 	int resource;
5854 {
5855 	struct thread *td = curthread;
5856 	struct ufsmount *ump;
5857 
5858 	mtx_assert(&lk, MA_OWNED);
5859 	/*
5860 	 * We never hold up the filesystem syncer or buf daemon.
5861 	 */
5862 	if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF))
5863 		return (0);
5864 	ump = VFSTOUFS(mp);
5865 	/*
5866 	 * First check to see if the work list has gotten backlogged.
5867 	 * If it has, co-opt this process to help clean up two entries.
5868 	 * Because this process may hold inodes locked, we cannot
5869 	 * handle any remove requests that might block on a locked
5870 	 * inode as that could lead to deadlock.  We set TDP_SOFTDEP
5871 	 * to avoid recursively processing the worklist.
5872 	 */
5873 	if (ump->softdep_on_worklist > max_softdeps / 10) {
5874 		td->td_pflags |= TDP_SOFTDEP;
5875 		process_worklist_item(mp, LK_NOWAIT);
5876 		process_worklist_item(mp, LK_NOWAIT);
5877 		td->td_pflags &= ~TDP_SOFTDEP;
5878 		stat_worklist_push += 2;
5879 		return(1);
5880 	}
5881 	/*
5882 	 * Next, we attempt to speed up the syncer process. If that
5883 	 * is successful, then we allow the process to continue.
5884 	 */
5885 	if (softdep_speedup() && resource != FLUSH_REMOVE_WAIT)
5886 		return(0);
5887 	/*
5888 	 * If we are resource constrained on inode dependencies, try
5889 	 * flushing some dirty inodes. Otherwise, we are constrained
5890 	 * by file deletions, so try accelerating flushes of directories
5891 	 * with removal dependencies. We would like to do the cleanup
5892 	 * here, but we probably hold an inode locked at this point and
5893 	 * that might deadlock against one that we try to clean. So,
5894 	 * the best that we can do is request the syncer daemon to do
5895 	 * the cleanup for us.
5896 	 */
5897 	switch (resource) {
5898 
5899 	case FLUSH_INODES:
5900 		stat_ino_limit_push += 1;
5901 		req_clear_inodedeps += 1;
5902 		stat_countp = &stat_ino_limit_hit;
5903 		break;
5904 
5905 	case FLUSH_REMOVE:
5906 	case FLUSH_REMOVE_WAIT:
5907 		stat_blk_limit_push += 1;
5908 		req_clear_remove += 1;
5909 		stat_countp = &stat_blk_limit_hit;
5910 		break;
5911 
5912 	default:
5913 		panic("request_cleanup: unknown type");
5914 	}
5915 	/*
5916 	 * Hopefully the syncer daemon will catch up and awaken us.
5917 	 * We wait at most tickdelay before proceeding in any case.
5918 	 */
5919 	proc_waiting += 1;
5920 	if (callout_pending(&softdep_callout) == FALSE)
5921 		callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2,
5922 		    pause_timer, 0);
5923 
5924 	msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0);
5925 	proc_waiting -= 1;
5926 	return (1);
5927 }
5928 
5929 /*
5930  * Awaken processes pausing in request_cleanup and clear proc_waiting
5931  * to indicate that there is no longer a timer running.
5932  */
5933 static void
5934 pause_timer(arg)
5935 	void *arg;
5936 {
5937 
5938 	/*
5939 	 * The callout_ API has acquired mtx and will hold it around this
5940 	 * function call.
5941 	 */
5942 	*stat_countp += 1;
5943 	wakeup_one(&proc_waiting);
5944 	if (proc_waiting > 0)
5945 		callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2,
5946 		    pause_timer, 0);
5947 }
5948 
5949 /*
5950  * Flush out a directory with at least one removal dependency in an effort to
5951  * reduce the number of dirrem, freefile, and freeblks dependency structures.
5952  */
5953 static void
5954 clear_remove(td)
5955 	struct thread *td;
5956 {
5957 	struct pagedep_hashhead *pagedephd;
5958 	struct pagedep *pagedep;
5959 	static int next = 0;
5960 	struct mount *mp;
5961 	struct vnode *vp;
5962 	struct bufobj *bo;
5963 	int error, cnt;
5964 	ino_t ino;
5965 
5966 	mtx_assert(&lk, MA_OWNED);
5967 
5968 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
5969 		pagedephd = &pagedep_hashtbl[next++];
5970 		if (next >= pagedep_hash)
5971 			next = 0;
5972 		LIST_FOREACH(pagedep, pagedephd, pd_hash) {
5973 			if (LIST_EMPTY(&pagedep->pd_dirremhd))
5974 				continue;
5975 			mp = pagedep->pd_list.wk_mp;
5976 			ino = pagedep->pd_ino;
5977 			if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5978 				continue;
5979 			FREE_LOCK(&lk);
5980 			if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
5981 			     FFSV_FORCEINSMQ))) {
5982 				softdep_error("clear_remove: vget", error);
5983 				vn_finished_write(mp);
5984 				ACQUIRE_LOCK(&lk);
5985 				return;
5986 			}
5987 			if ((error = ffs_syncvnode(vp, MNT_NOWAIT)))
5988 				softdep_error("clear_remove: fsync", error);
5989 			bo = &vp->v_bufobj;
5990 			BO_LOCK(bo);
5991 			drain_output(vp);
5992 			BO_UNLOCK(bo);
5993 			vput(vp);
5994 			vn_finished_write(mp);
5995 			ACQUIRE_LOCK(&lk);
5996 			return;
5997 		}
5998 	}
5999 }
6000 
6001 /*
6002  * Clear out a block of dirty inodes in an effort to reduce
6003  * the number of inodedep dependency structures.
6004  */
6005 static void
6006 clear_inodedeps(td)
6007 	struct thread *td;
6008 {
6009 	struct inodedep_hashhead *inodedephd;
6010 	struct inodedep *inodedep;
6011 	static int next = 0;
6012 	struct mount *mp;
6013 	struct vnode *vp;
6014 	struct fs *fs;
6015 	int error, cnt;
6016 	ino_t firstino, lastino, ino;
6017 
6018 	mtx_assert(&lk, MA_OWNED);
6019 	/*
6020 	 * Pick a random inode dependency to be cleared.
6021 	 * We will then gather up all the inodes in its block
6022 	 * that have dependencies and flush them out.
6023 	 */
6024 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
6025 		inodedephd = &inodedep_hashtbl[next++];
6026 		if (next >= inodedep_hash)
6027 			next = 0;
6028 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
6029 			break;
6030 	}
6031 	if (inodedep == NULL)
6032 		return;
6033 	fs = inodedep->id_fs;
6034 	mp = inodedep->id_list.wk_mp;
6035 	/*
6036 	 * Find the last inode in the block with dependencies.
6037 	 */
6038 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
6039 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
6040 		if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0)
6041 			break;
6042 	/*
6043 	 * Asynchronously push all but the last inode with dependencies.
6044 	 * Synchronously push the last inode with dependencies to ensure
6045 	 * that the inode block gets written to free up the inodedeps.
6046 	 */
6047 	for (ino = firstino; ino <= lastino; ino++) {
6048 		if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
6049 			continue;
6050 		if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
6051 			continue;
6052 		FREE_LOCK(&lk);
6053 		if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
6054 		    FFSV_FORCEINSMQ)) != 0) {
6055 			softdep_error("clear_inodedeps: vget", error);
6056 			vn_finished_write(mp);
6057 			ACQUIRE_LOCK(&lk);
6058 			return;
6059 		}
6060 		if (ino == lastino) {
6061 			if ((error = ffs_syncvnode(vp, MNT_WAIT)))
6062 				softdep_error("clear_inodedeps: fsync1", error);
6063 		} else {
6064 			if ((error = ffs_syncvnode(vp, MNT_NOWAIT)))
6065 				softdep_error("clear_inodedeps: fsync2", error);
6066 			BO_LOCK(&vp->v_bufobj);
6067 			drain_output(vp);
6068 			BO_UNLOCK(&vp->v_bufobj);
6069 		}
6070 		vput(vp);
6071 		vn_finished_write(mp);
6072 		ACQUIRE_LOCK(&lk);
6073 	}
6074 }
6075 
6076 /*
6077  * Function to determine if the buffer has outstanding dependencies
6078  * that will cause a roll-back if the buffer is written. If wantcount
6079  * is set, return number of dependencies, otherwise just yes or no.
6080  */
6081 static int
6082 softdep_count_dependencies(bp, wantcount)
6083 	struct buf *bp;
6084 	int wantcount;
6085 {
6086 	struct worklist *wk;
6087 	struct inodedep *inodedep;
6088 	struct indirdep *indirdep;
6089 	struct allocindir *aip;
6090 	struct pagedep *pagedep;
6091 	struct diradd *dap;
6092 	int i, retval;
6093 
6094 	retval = 0;
6095 	ACQUIRE_LOCK(&lk);
6096 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
6097 		switch (wk->wk_type) {
6098 
6099 		case D_INODEDEP:
6100 			inodedep = WK_INODEDEP(wk);
6101 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
6102 				/* bitmap allocation dependency */
6103 				retval += 1;
6104 				if (!wantcount)
6105 					goto out;
6106 			}
6107 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
6108 				/* direct block pointer dependency */
6109 				retval += 1;
6110 				if (!wantcount)
6111 					goto out;
6112 			}
6113 			if (TAILQ_FIRST(&inodedep->id_extupdt)) {
6114 				/* direct block pointer dependency */
6115 				retval += 1;
6116 				if (!wantcount)
6117 					goto out;
6118 			}
6119 			continue;
6120 
6121 		case D_INDIRDEP:
6122 			indirdep = WK_INDIRDEP(wk);
6123 
6124 			LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
6125 				/* indirect block pointer dependency */
6126 				retval += 1;
6127 				if (!wantcount)
6128 					goto out;
6129 			}
6130 			continue;
6131 
6132 		case D_PAGEDEP:
6133 			pagedep = WK_PAGEDEP(wk);
6134 			for (i = 0; i < DAHASHSZ; i++) {
6135 
6136 				LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
6137 					/* directory entry dependency */
6138 					retval += 1;
6139 					if (!wantcount)
6140 						goto out;
6141 				}
6142 			}
6143 			continue;
6144 
6145 		case D_BMSAFEMAP:
6146 		case D_ALLOCDIRECT:
6147 		case D_ALLOCINDIR:
6148 		case D_MKDIR:
6149 			/* never a dependency on these blocks */
6150 			continue;
6151 
6152 		default:
6153 			panic("softdep_check_for_rollback: Unexpected type %s",
6154 			    TYPENAME(wk->wk_type));
6155 			/* NOTREACHED */
6156 		}
6157 	}
6158 out:
6159 	FREE_LOCK(&lk);
6160 	return retval;
6161 }
6162 
6163 /*
6164  * Acquire exclusive access to a buffer.
6165  * Must be called with a locked mtx parameter.
6166  * Return acquired buffer or NULL on failure.
6167  */
6168 static struct buf *
6169 getdirtybuf(bp, mtx, waitfor)
6170 	struct buf *bp;
6171 	struct mtx *mtx;
6172 	int waitfor;
6173 {
6174 	int error;
6175 
6176 	mtx_assert(mtx, MA_OWNED);
6177 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) {
6178 		if (waitfor != MNT_WAIT)
6179 			return (NULL);
6180 		error = BUF_LOCK(bp,
6181 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx);
6182 		/*
6183 		 * Even if we sucessfully acquire bp here, we have dropped
6184 		 * mtx, which may violates our guarantee.
6185 		 */
6186 		if (error == 0)
6187 			BUF_UNLOCK(bp);
6188 		else if (error != ENOLCK)
6189 			panic("getdirtybuf: inconsistent lock: %d", error);
6190 		mtx_lock(mtx);
6191 		return (NULL);
6192 	}
6193 	if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
6194 		if (mtx == &lk && waitfor == MNT_WAIT) {
6195 			mtx_unlock(mtx);
6196 			BO_LOCK(bp->b_bufobj);
6197 			BUF_UNLOCK(bp);
6198 			if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
6199 				bp->b_vflags |= BV_BKGRDWAIT;
6200 				msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj),
6201 				       PRIBIO | PDROP, "getbuf", 0);
6202 			} else
6203 				BO_UNLOCK(bp->b_bufobj);
6204 			mtx_lock(mtx);
6205 			return (NULL);
6206 		}
6207 		BUF_UNLOCK(bp);
6208 		if (waitfor != MNT_WAIT)
6209 			return (NULL);
6210 		/*
6211 		 * The mtx argument must be bp->b_vp's mutex in
6212 		 * this case.
6213 		 */
6214 #ifdef	DEBUG_VFS_LOCKS
6215 		if (bp->b_vp->v_type != VCHR)
6216 			ASSERT_BO_LOCKED(bp->b_bufobj);
6217 #endif
6218 		bp->b_vflags |= BV_BKGRDWAIT;
6219 		msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0);
6220 		return (NULL);
6221 	}
6222 	if ((bp->b_flags & B_DELWRI) == 0) {
6223 		BUF_UNLOCK(bp);
6224 		return (NULL);
6225 	}
6226 	bremfree(bp);
6227 	return (bp);
6228 }
6229 
6230 
6231 /*
6232  * Check if it is safe to suspend the file system now.  On entry,
6233  * the vnode interlock for devvp should be held.  Return 0 with
6234  * the mount interlock held if the file system can be suspended now,
6235  * otherwise return EAGAIN with the mount interlock held.
6236  */
6237 int
6238 softdep_check_suspend(struct mount *mp,
6239 		      struct vnode *devvp,
6240 		      int softdep_deps,
6241 		      int softdep_accdeps,
6242 		      int secondary_writes,
6243 		      int secondary_accwrites)
6244 {
6245 	struct bufobj *bo;
6246 	struct ufsmount *ump;
6247 	int error;
6248 
6249 	ump = VFSTOUFS(mp);
6250 	bo = &devvp->v_bufobj;
6251 	ASSERT_BO_LOCKED(bo);
6252 
6253 	for (;;) {
6254 		if (!TRY_ACQUIRE_LOCK(&lk)) {
6255 			BO_UNLOCK(bo);
6256 			ACQUIRE_LOCK(&lk);
6257 			FREE_LOCK(&lk);
6258 			BO_LOCK(bo);
6259 			continue;
6260 		}
6261 		MNT_ILOCK(mp);
6262 		if (mp->mnt_secondary_writes != 0) {
6263 			FREE_LOCK(&lk);
6264 			BO_UNLOCK(bo);
6265 			msleep(&mp->mnt_secondary_writes,
6266 			       MNT_MTX(mp),
6267 			       (PUSER - 1) | PDROP, "secwr", 0);
6268 			BO_LOCK(bo);
6269 			continue;
6270 		}
6271 		break;
6272 	}
6273 
6274 	/*
6275 	 * Reasons for needing more work before suspend:
6276 	 * - Dirty buffers on devvp.
6277 	 * - Softdep activity occurred after start of vnode sync loop
6278 	 * - Secondary writes occurred after start of vnode sync loop
6279 	 */
6280 	error = 0;
6281 	if (bo->bo_numoutput > 0 ||
6282 	    bo->bo_dirty.bv_cnt > 0 ||
6283 	    softdep_deps != 0 ||
6284 	    ump->softdep_deps != 0 ||
6285 	    softdep_accdeps != ump->softdep_accdeps ||
6286 	    secondary_writes != 0 ||
6287 	    mp->mnt_secondary_writes != 0 ||
6288 	    secondary_accwrites != mp->mnt_secondary_accwrites)
6289 		error = EAGAIN;
6290 	FREE_LOCK(&lk);
6291 	BO_UNLOCK(bo);
6292 	return (error);
6293 }
6294 
6295 
6296 /*
6297  * Get the number of dependency structures for the file system, both
6298  * the current number and the total number allocated.  These will
6299  * later be used to detect that softdep processing has occurred.
6300  */
6301 void
6302 softdep_get_depcounts(struct mount *mp,
6303 		      int *softdep_depsp,
6304 		      int *softdep_accdepsp)
6305 {
6306 	struct ufsmount *ump;
6307 
6308 	ump = VFSTOUFS(mp);
6309 	ACQUIRE_LOCK(&lk);
6310 	*softdep_depsp = ump->softdep_deps;
6311 	*softdep_accdepsp = ump->softdep_accdeps;
6312 	FREE_LOCK(&lk);
6313 }
6314 
6315 /*
6316  * Wait for pending output on a vnode to complete.
6317  * Must be called with vnode lock and interlock locked.
6318  *
6319  * XXX: Should just be a call to bufobj_wwait().
6320  */
6321 static void
6322 drain_output(vp)
6323 	struct vnode *vp;
6324 {
6325 	struct bufobj *bo;
6326 
6327 	bo = &vp->v_bufobj;
6328 	ASSERT_VOP_LOCKED(vp, "drain_output");
6329 	ASSERT_BO_LOCKED(bo);
6330 
6331 	while (bo->bo_numoutput) {
6332 		bo->bo_flag |= BO_WWAIT;
6333 		msleep((caddr_t)&bo->bo_numoutput,
6334 		    BO_MTX(bo), PRIBIO + 1, "drainvp", 0);
6335 	}
6336 }
6337 
6338 /*
6339  * Called whenever a buffer that is being invalidated or reallocated
6340  * contains dependencies. This should only happen if an I/O error has
6341  * occurred. The routine is called with the buffer locked.
6342  */
6343 static void
6344 softdep_deallocate_dependencies(bp)
6345 	struct buf *bp;
6346 {
6347 
6348 	if ((bp->b_ioflags & BIO_ERROR) == 0)
6349 		panic("softdep_deallocate_dependencies: dangling deps");
6350 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
6351 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
6352 }
6353 
6354 /*
6355  * Function to handle asynchronous write errors in the filesystem.
6356  */
6357 static void
6358 softdep_error(func, error)
6359 	char *func;
6360 	int error;
6361 {
6362 
6363 	/* XXX should do something better! */
6364 	printf("%s: got error %d while accessing filesystem\n", func, error);
6365 }
6366 
6367 #ifdef DDB
6368 
6369 DB_SHOW_COMMAND(inodedeps, db_show_inodedeps)
6370 {
6371 	struct inodedep_hashhead *inodedephd;
6372 	struct inodedep *inodedep;
6373 	struct fs *fs;
6374 	int cnt;
6375 
6376 	fs = have_addr ? (struct fs *)addr : NULL;
6377 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
6378 		inodedephd = &inodedep_hashtbl[cnt];
6379 		LIST_FOREACH(inodedep, inodedephd, id_hash) {
6380 			if (fs != NULL && fs != inodedep->id_fs)
6381 				continue;
6382 			db_printf("%p fs %p st %x ino %jd inoblk %jd\n",
6383 			    inodedep, inodedep->id_fs, inodedep->id_state,
6384 			    (intmax_t)inodedep->id_ino,
6385 			    (intmax_t)fsbtodb(inodedep->id_fs,
6386 			    ino_to_fsba(inodedep->id_fs, inodedep->id_ino)));
6387 		}
6388 	}
6389 }
6390 
6391 #endif /* DDB */
6392 
6393 #endif /* SOFTUPDATES */
6394