xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision 9336e0699bda8a301cd2bfa37106b6ec5e32012e)
1 /*-
2  * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
3  *
4  * The soft updates code is derived from the appendix of a University
5  * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
6  * "Soft Updates: A Solution to the Metadata Update Problem in File
7  * Systems", CSE-TR-254-95, August 1995).
8  *
9  * Further information about soft updates can be obtained from:
10  *
11  *	Marshall Kirk McKusick		http://www.mckusick.com/softdep/
12  *	1614 Oxford Street		mckusick@mckusick.com
13  *	Berkeley, CA 94709-1608		+1-510-843-9542
14  *	USA
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  *
20  * 1. Redistributions of source code must retain the above copyright
21  *    notice, this list of conditions and the following disclaimer.
22  * 2. Redistributions in binary form must reproduce the above copyright
23  *    notice, this list of conditions and the following disclaimer in the
24  *    documentation and/or other materials provided with the distribution.
25  *
26  * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
27  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
28  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
29  * DISCLAIMED.  IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
30  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)ffs_softdep.c	9.59 (McKusick) 6/21/00
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 /*
45  * For now we want the safety net that the DEBUG flag provides.
46  */
47 #ifndef DEBUG
48 #define DEBUG
49 #endif
50 
51 #include <sys/param.h>
52 #include <sys/kernel.h>
53 #include <sys/systm.h>
54 #include <sys/bio.h>
55 #include <sys/buf.h>
56 #include <sys/kdb.h>
57 #include <sys/kthread.h>
58 #include <sys/lock.h>
59 #include <sys/malloc.h>
60 #include <sys/mount.h>
61 #include <sys/mutex.h>
62 #include <sys/proc.h>
63 #include <sys/stat.h>
64 #include <sys/sysctl.h>
65 #include <sys/syslog.h>
66 #include <sys/vnode.h>
67 #include <sys/conf.h>
68 #include <ufs/ufs/dir.h>
69 #include <ufs/ufs/extattr.h>
70 #include <ufs/ufs/quota.h>
71 #include <ufs/ufs/inode.h>
72 #include <ufs/ufs/ufsmount.h>
73 #include <ufs/ffs/fs.h>
74 #include <ufs/ffs/softdep.h>
75 #include <ufs/ffs/ffs_extern.h>
76 #include <ufs/ufs/ufs_extern.h>
77 
78 #include <vm/vm.h>
79 
80 #include "opt_ffs.h"
81 #include "opt_quota.h"
82 
83 #ifndef SOFTUPDATES
84 
85 int
86 softdep_flushfiles(oldmnt, flags, td)
87 	struct mount *oldmnt;
88 	int flags;
89 	struct thread *td;
90 {
91 
92 	panic("softdep_flushfiles called");
93 }
94 
95 int
96 softdep_mount(devvp, mp, fs, cred)
97 	struct vnode *devvp;
98 	struct mount *mp;
99 	struct fs *fs;
100 	struct ucred *cred;
101 {
102 
103 	return (0);
104 }
105 
106 void
107 softdep_initialize()
108 {
109 
110 	return;
111 }
112 
113 void
114 softdep_uninitialize()
115 {
116 
117 	return;
118 }
119 
120 void
121 softdep_setup_inomapdep(bp, ip, newinum)
122 	struct buf *bp;
123 	struct inode *ip;
124 	ino_t newinum;
125 {
126 
127 	panic("softdep_setup_inomapdep called");
128 }
129 
130 void
131 softdep_setup_blkmapdep(bp, mp, newblkno)
132 	struct buf *bp;
133 	struct mount *mp;
134 	ufs2_daddr_t newblkno;
135 {
136 
137 	panic("softdep_setup_blkmapdep called");
138 }
139 
140 void
141 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
142 	struct inode *ip;
143 	ufs_lbn_t lbn;
144 	ufs2_daddr_t newblkno;
145 	ufs2_daddr_t oldblkno;
146 	long newsize;
147 	long oldsize;
148 	struct buf *bp;
149 {
150 
151 	panic("softdep_setup_allocdirect called");
152 }
153 
154 void
155 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
156 	struct inode *ip;
157 	ufs_lbn_t lbn;
158 	ufs2_daddr_t newblkno;
159 	ufs2_daddr_t oldblkno;
160 	long newsize;
161 	long oldsize;
162 	struct buf *bp;
163 {
164 
165 	panic("softdep_setup_allocext called");
166 }
167 
168 void
169 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
170 	struct inode *ip;
171 	ufs_lbn_t lbn;
172 	struct buf *bp;
173 	int ptrno;
174 	ufs2_daddr_t newblkno;
175 	ufs2_daddr_t oldblkno;
176 	struct buf *nbp;
177 {
178 
179 	panic("softdep_setup_allocindir_page called");
180 }
181 
182 void
183 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
184 	struct buf *nbp;
185 	struct inode *ip;
186 	struct buf *bp;
187 	int ptrno;
188 	ufs2_daddr_t newblkno;
189 {
190 
191 	panic("softdep_setup_allocindir_meta called");
192 }
193 
194 void
195 softdep_setup_freeblocks(ip, length, flags)
196 	struct inode *ip;
197 	off_t length;
198 	int flags;
199 {
200 
201 	panic("softdep_setup_freeblocks called");
202 }
203 
204 void
205 softdep_freefile(pvp, ino, mode)
206 		struct vnode *pvp;
207 		ino_t ino;
208 		int mode;
209 {
210 
211 	panic("softdep_freefile called");
212 }
213 
214 int
215 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
216 	struct buf *bp;
217 	struct inode *dp;
218 	off_t diroffset;
219 	ino_t newinum;
220 	struct buf *newdirbp;
221 	int isnewblk;
222 {
223 
224 	panic("softdep_setup_directory_add called");
225 }
226 
227 void
228 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
229 	struct inode *dp;
230 	caddr_t base;
231 	caddr_t oldloc;
232 	caddr_t newloc;
233 	int entrysize;
234 {
235 
236 	panic("softdep_change_directoryentry_offset called");
237 }
238 
239 void
240 softdep_setup_remove(bp, dp, ip, isrmdir)
241 	struct buf *bp;
242 	struct inode *dp;
243 	struct inode *ip;
244 	int isrmdir;
245 {
246 
247 	panic("softdep_setup_remove called");
248 }
249 
250 void
251 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
252 	struct buf *bp;
253 	struct inode *dp;
254 	struct inode *ip;
255 	ino_t newinum;
256 	int isrmdir;
257 {
258 
259 	panic("softdep_setup_directory_change called");
260 }
261 
262 void
263 softdep_change_linkcnt(ip)
264 	struct inode *ip;
265 {
266 
267 	panic("softdep_change_linkcnt called");
268 }
269 
270 void
271 softdep_load_inodeblock(ip)
272 	struct inode *ip;
273 {
274 
275 	panic("softdep_load_inodeblock called");
276 }
277 
278 void
279 softdep_update_inodeblock(ip, bp, waitfor)
280 	struct inode *ip;
281 	struct buf *bp;
282 	int waitfor;
283 {
284 
285 	panic("softdep_update_inodeblock called");
286 }
287 
288 int
289 softdep_fsync(vp)
290 	struct vnode *vp;	/* the "in_core" copy of the inode */
291 {
292 
293 	return (0);
294 }
295 
296 void
297 softdep_fsync_mountdev(vp)
298 	struct vnode *vp;
299 {
300 
301 	return;
302 }
303 
304 int
305 softdep_flushworklist(oldmnt, countp, td)
306 	struct mount *oldmnt;
307 	int *countp;
308 	struct thread *td;
309 {
310 
311 	*countp = 0;
312 	return (0);
313 }
314 
315 int
316 softdep_sync_metadata(struct vnode *vp)
317 {
318 
319 	return (0);
320 }
321 
322 int
323 softdep_slowdown(vp)
324 	struct vnode *vp;
325 {
326 
327 	panic("softdep_slowdown called");
328 }
329 
330 void
331 softdep_releasefile(ip)
332 	struct inode *ip;	/* inode with the zero effective link count */
333 {
334 
335 	panic("softdep_releasefile called");
336 }
337 
338 int
339 softdep_request_cleanup(fs, vp)
340 	struct fs *fs;
341 	struct vnode *vp;
342 {
343 
344 	return (0);
345 }
346 
347 int
348 softdep_check_suspend(struct mount *mp,
349 		      struct vnode *devvp,
350 		      int softdep_deps,
351 		      int softdep_accdeps,
352 		      int secondary_writes,
353 		      int secondary_accwrites)
354 {
355 	struct bufobj *bo;
356 	int error;
357 
358 	(void) softdep_deps,
359 	(void) softdep_accdeps;
360 
361 	ASSERT_VI_LOCKED(devvp, "softdep_check_suspend");
362 	bo = &devvp->v_bufobj;
363 
364 	for (;;) {
365 		if (!MNT_ITRYLOCK(mp)) {
366 			VI_UNLOCK(devvp);
367 			MNT_ILOCK(mp);
368 			MNT_IUNLOCK(mp);
369 			VI_LOCK(devvp);
370 			continue;
371 		}
372 		if (mp->mnt_secondary_writes != 0) {
373 			VI_UNLOCK(devvp);
374 			msleep(&mp->mnt_secondary_writes,
375 			       MNT_MTX(mp),
376 			       (PUSER - 1) | PDROP, "secwr", 0);
377 			VI_LOCK(devvp);
378 			continue;
379 		}
380 		break;
381 	}
382 
383 	/*
384 	 * Reasons for needing more work before suspend:
385 	 * - Dirty buffers on devvp.
386 	 * - Secondary writes occurred after start of vnode sync loop
387 	 */
388 	error = 0;
389 	if (bo->bo_numoutput > 0 ||
390 	    bo->bo_dirty.bv_cnt > 0 ||
391 	    secondary_writes != 0 ||
392 	    mp->mnt_secondary_writes != 0 ||
393 	    secondary_accwrites != mp->mnt_secondary_accwrites)
394 		error = EAGAIN;
395 	VI_UNLOCK(devvp);
396 	return (error);
397 }
398 
399 void
400 softdep_get_depcounts(struct mount *mp,
401 		      int *softdepactivep,
402 		      int *softdepactiveaccp)
403 {
404 	(void) mp;
405 	*softdepactivep = 0;
406 	*softdepactiveaccp = 0;
407 }
408 
409 #else
410 /*
411  * These definitions need to be adapted to the system to which
412  * this file is being ported.
413  */
414 /*
415  * malloc types defined for the softdep system.
416  */
417 static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
418 static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
419 static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
420 static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
421 static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
422 static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
423 static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
424 static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
425 static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
426 static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
427 static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
428 static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
429 static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
430 static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block");
431 static MALLOC_DEFINE(M_SAVEDINO, "savedino","Saved inodes");
432 
433 #define M_SOFTDEP_FLAGS	(M_WAITOK | M_USE_RESERVE)
434 
435 #define	D_PAGEDEP	0
436 #define	D_INODEDEP	1
437 #define	D_NEWBLK	2
438 #define	D_BMSAFEMAP	3
439 #define	D_ALLOCDIRECT	4
440 #define	D_INDIRDEP	5
441 #define	D_ALLOCINDIR	6
442 #define	D_FREEFRAG	7
443 #define	D_FREEBLKS	8
444 #define	D_FREEFILE	9
445 #define	D_DIRADD	10
446 #define	D_MKDIR		11
447 #define	D_DIRREM	12
448 #define	D_NEWDIRBLK	13
449 #define	D_LAST		D_NEWDIRBLK
450 
451 /*
452  * translate from workitem type to memory type
453  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
454  */
455 static struct malloc_type *memtype[] = {
456 	M_PAGEDEP,
457 	M_INODEDEP,
458 	M_NEWBLK,
459 	M_BMSAFEMAP,
460 	M_ALLOCDIRECT,
461 	M_INDIRDEP,
462 	M_ALLOCINDIR,
463 	M_FREEFRAG,
464 	M_FREEBLKS,
465 	M_FREEFILE,
466 	M_DIRADD,
467 	M_MKDIR,
468 	M_DIRREM,
469 	M_NEWDIRBLK
470 };
471 
472 #define DtoM(type) (memtype[type])
473 
474 /*
475  * Names of malloc types.
476  */
477 #define TYPENAME(type)  \
478 	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
479 /*
480  * End system adaptation definitions.
481  */
482 
483 /*
484  * Forward declarations.
485  */
486 struct inodedep_hashhead;
487 struct newblk_hashhead;
488 struct pagedep_hashhead;
489 
490 /*
491  * Internal function prototypes.
492  */
493 static	void softdep_error(char *, int);
494 static	void drain_output(struct vnode *);
495 static	struct buf *getdirtybuf(struct buf *, struct mtx *, int);
496 static	void clear_remove(struct thread *);
497 static	void clear_inodedeps(struct thread *);
498 static	int flush_pagedep_deps(struct vnode *, struct mount *,
499 	    struct diraddhd *);
500 static	int flush_inodedep_deps(struct mount *, ino_t);
501 static	int flush_deplist(struct allocdirectlst *, int, int *);
502 static	int handle_written_filepage(struct pagedep *, struct buf *);
503 static  void diradd_inode_written(struct diradd *, struct inodedep *);
504 static	int handle_written_inodeblock(struct inodedep *, struct buf *);
505 static	void handle_allocdirect_partdone(struct allocdirect *);
506 static	void handle_allocindir_partdone(struct allocindir *);
507 static	void initiate_write_filepage(struct pagedep *, struct buf *);
508 static	void handle_written_mkdir(struct mkdir *, int);
509 static	void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
510 static	void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
511 static	void handle_workitem_freefile(struct freefile *);
512 static	void handle_workitem_remove(struct dirrem *, struct vnode *);
513 static	struct dirrem *newdirrem(struct buf *, struct inode *,
514 	    struct inode *, int, struct dirrem **);
515 static	void free_diradd(struct diradd *);
516 static	void free_allocindir(struct allocindir *, struct inodedep *);
517 static	void free_newdirblk(struct newdirblk *);
518 static	int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t,
519 	    ufs2_daddr_t *);
520 static	void deallocate_dependencies(struct buf *, struct inodedep *);
521 static	void free_allocdirect(struct allocdirectlst *,
522 	    struct allocdirect *, int);
523 static	int check_inode_unwritten(struct inodedep *);
524 static	int free_inodedep(struct inodedep *);
525 static	void handle_workitem_freeblocks(struct freeblks *, int);
526 static	void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
527 static	void setup_allocindir_phase2(struct buf *, struct inode *,
528 	    struct allocindir *);
529 static	struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
530 	    ufs2_daddr_t);
531 static	void handle_workitem_freefrag(struct freefrag *);
532 static	struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long);
533 static	void allocdirect_merge(struct allocdirectlst *,
534 	    struct allocdirect *, struct allocdirect *);
535 static	struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *);
536 static	int newblk_find(struct newblk_hashhead *, struct fs *, ufs2_daddr_t,
537 	    struct newblk **);
538 static	int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **);
539 static	int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t,
540 	    struct inodedep **);
541 static	int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **);
542 static	int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **);
543 static	int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t,
544 	    struct mount *mp, int, struct pagedep **);
545 static	void pause_timer(void *);
546 static	int request_cleanup(struct mount *, int);
547 static	int process_worklist_item(struct mount *, int);
548 static	void add_to_worklist(struct worklist *);
549 static	void softdep_flush(void);
550 static	int softdep_speedup(void);
551 
552 /*
553  * Exported softdep operations.
554  */
555 static	void softdep_disk_io_initiation(struct buf *);
556 static	void softdep_disk_write_complete(struct buf *);
557 static	void softdep_deallocate_dependencies(struct buf *);
558 static	int softdep_count_dependencies(struct buf *bp, int);
559 
560 static struct mtx lk;
561 MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF);
562 
563 #define TRY_ACQUIRE_LOCK(lk)		mtx_trylock(lk)
564 #define ACQUIRE_LOCK(lk)		mtx_lock(lk)
565 #define FREE_LOCK(lk)			mtx_unlock(lk)
566 
567 /*
568  * Worklist queue management.
569  * These routines require that the lock be held.
570  */
571 #ifndef /* NOT */ DEBUG
572 #define WORKLIST_INSERT(head, item) do {	\
573 	(item)->wk_state |= ONWORKLIST;		\
574 	LIST_INSERT_HEAD(head, item, wk_list);	\
575 } while (0)
576 #define WORKLIST_REMOVE(item) do {		\
577 	(item)->wk_state &= ~ONWORKLIST;	\
578 	LIST_REMOVE(item, wk_list);		\
579 } while (0)
580 #else /* DEBUG */
581 static	void worklist_insert(struct workhead *, struct worklist *);
582 static	void worklist_remove(struct worklist *);
583 
584 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
585 #define WORKLIST_REMOVE(item) worklist_remove(item)
586 
587 static void
588 worklist_insert(head, item)
589 	struct workhead *head;
590 	struct worklist *item;
591 {
592 
593 	mtx_assert(&lk, MA_OWNED);
594 	if (item->wk_state & ONWORKLIST)
595 		panic("worklist_insert: already on list");
596 	item->wk_state |= ONWORKLIST;
597 	LIST_INSERT_HEAD(head, item, wk_list);
598 }
599 
600 static void
601 worklist_remove(item)
602 	struct worklist *item;
603 {
604 
605 	mtx_assert(&lk, MA_OWNED);
606 	if ((item->wk_state & ONWORKLIST) == 0)
607 		panic("worklist_remove: not on list");
608 	item->wk_state &= ~ONWORKLIST;
609 	LIST_REMOVE(item, wk_list);
610 }
611 #endif /* DEBUG */
612 
613 /*
614  * Routines for tracking and managing workitems.
615  */
616 static	void workitem_free(struct worklist *, int);
617 static	void workitem_alloc(struct worklist *, int, struct mount *);
618 
619 #define	WORKITEM_FREE(item, type) workitem_free((struct worklist *)(item), (type))
620 
621 static void
622 workitem_free(item, type)
623 	struct worklist *item;
624 	int type;
625 {
626 	struct ufsmount *ump;
627 	mtx_assert(&lk, MA_OWNED);
628 
629 #ifdef DEBUG
630 	if (item->wk_state & ONWORKLIST)
631 		panic("workitem_free: still on list");
632 	if (item->wk_type != type)
633 		panic("workitem_free: type mismatch");
634 #endif
635 	ump = VFSTOUFS(item->wk_mp);
636 	if (--ump->softdep_deps == 0 && ump->softdep_req)
637 		wakeup(&ump->softdep_deps);
638 	FREE(item, DtoM(type));
639 }
640 
641 static void
642 workitem_alloc(item, type, mp)
643 	struct worklist *item;
644 	int type;
645 	struct mount *mp;
646 {
647 	item->wk_type = type;
648 	item->wk_mp = mp;
649 	item->wk_state = 0;
650 	ACQUIRE_LOCK(&lk);
651 	VFSTOUFS(mp)->softdep_deps++;
652 	VFSTOUFS(mp)->softdep_accdeps++;
653 	FREE_LOCK(&lk);
654 }
655 
656 /*
657  * Workitem queue management
658  */
659 static int max_softdeps;	/* maximum number of structs before slowdown */
660 static int maxindirdeps = 50;	/* max number of indirdeps before slowdown */
661 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
662 static int proc_waiting;	/* tracks whether we have a timeout posted */
663 static int *stat_countp;	/* statistic to count in proc_waiting timeout */
664 static struct callout_handle handle; /* handle on posted proc_waiting timeout */
665 static int req_pending;
666 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
667 #define FLUSH_INODES		1
668 static int req_clear_remove;	/* syncer process flush some freeblks */
669 #define FLUSH_REMOVE		2
670 #define FLUSH_REMOVE_WAIT	3
671 /*
672  * runtime statistics
673  */
674 static int stat_worklist_push;	/* number of worklist cleanups */
675 static int stat_blk_limit_push;	/* number of times block limit neared */
676 static int stat_ino_limit_push;	/* number of times inode limit neared */
677 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
678 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
679 static int stat_sync_limit_hit;	/* number of synchronous slowdowns imposed */
680 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
681 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
682 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
683 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
684 
685 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
686 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
687 SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, "");
688 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
689 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
690 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
691 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
692 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
693 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
694 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
695 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
696 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
697 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
698 /* SYSCTL_INT(_debug, OID_AUTO, worklist_num, CTLFLAG_RD, &softdep_on_worklist, 0, ""); */
699 
700 SYSCTL_DECL(_vfs_ffs);
701 
702 static int compute_summary_at_mount = 0;	/* Whether to recompute the summary at mount time */
703 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW,
704 	   &compute_summary_at_mount, 0, "Recompute summary at mount");
705 
706 static struct proc *softdepproc;
707 static struct kproc_desc softdep_kp = {
708 	"softdepflush",
709 	softdep_flush,
710 	&softdepproc
711 };
712 SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start, &softdep_kp)
713 
714 static void
715 softdep_flush(void)
716 {
717 	struct mount *nmp;
718 	struct mount *mp;
719 	struct ufsmount *ump;
720 	struct thread *td;
721 	int remaining;
722 	int vfslocked;
723 
724 	td = curthread;
725 	td->td_pflags |= TDP_NORUNNINGBUF;
726 
727 	for (;;) {
728 		kproc_suspend_check(softdepproc);
729 		vfslocked = VFS_LOCK_GIANT((struct mount *)NULL);
730 		ACQUIRE_LOCK(&lk);
731 		/*
732 		 * If requested, try removing inode or removal dependencies.
733 		 */
734 		if (req_clear_inodedeps) {
735 			clear_inodedeps(td);
736 			req_clear_inodedeps -= 1;
737 			wakeup_one(&proc_waiting);
738 		}
739 		if (req_clear_remove) {
740 			clear_remove(td);
741 			req_clear_remove -= 1;
742 			wakeup_one(&proc_waiting);
743 		}
744 		FREE_LOCK(&lk);
745 		VFS_UNLOCK_GIANT(vfslocked);
746 		remaining = 0;
747 		mtx_lock(&mountlist_mtx);
748 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp)  {
749 			nmp = TAILQ_NEXT(mp, mnt_list);
750 			if ((mp->mnt_flag & MNT_SOFTDEP) == 0)
751 				continue;
752 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
753 				continue;
754 			vfslocked = VFS_LOCK_GIANT(mp);
755 			softdep_process_worklist(mp, 0);
756 			ump = VFSTOUFS(mp);
757 			remaining += ump->softdep_on_worklist -
758 				ump->softdep_on_worklist_inprogress;
759 			VFS_UNLOCK_GIANT(vfslocked);
760 			mtx_lock(&mountlist_mtx);
761 			nmp = TAILQ_NEXT(mp, mnt_list);
762 			vfs_unbusy(mp, td);
763 		}
764 		mtx_unlock(&mountlist_mtx);
765 		if (remaining)
766 			continue;
767 		ACQUIRE_LOCK(&lk);
768 		if (!req_pending)
769 			msleep(&req_pending, &lk, PVM, "sdflush", hz);
770 		req_pending = 0;
771 		FREE_LOCK(&lk);
772 	}
773 }
774 
775 static int
776 softdep_speedup(void)
777 {
778 
779 	mtx_assert(&lk, MA_OWNED);
780 	if (req_pending == 0) {
781 		req_pending = 1;
782 		wakeup(&req_pending);
783 	}
784 
785 	return speedup_syncer();
786 }
787 
788 /*
789  * Add an item to the end of the work queue.
790  * This routine requires that the lock be held.
791  * This is the only routine that adds items to the list.
792  * The following routine is the only one that removes items
793  * and does so in order from first to last.
794  */
795 static void
796 add_to_worklist(wk)
797 	struct worklist *wk;
798 {
799 	struct ufsmount *ump;
800 
801 	mtx_assert(&lk, MA_OWNED);
802 	ump = VFSTOUFS(wk->wk_mp);
803 	if (wk->wk_state & ONWORKLIST)
804 		panic("add_to_worklist: already on list");
805 	wk->wk_state |= ONWORKLIST;
806 	if (LIST_EMPTY(&ump->softdep_workitem_pending))
807 		LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list);
808 	else
809 		LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list);
810 	ump->softdep_worklist_tail = wk;
811 	ump->softdep_on_worklist += 1;
812 }
813 
814 /*
815  * Process that runs once per second to handle items in the background queue.
816  *
817  * Note that we ensure that everything is done in the order in which they
818  * appear in the queue. The code below depends on this property to ensure
819  * that blocks of a file are freed before the inode itself is freed. This
820  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
821  * until all the old ones have been purged from the dependency lists.
822  */
823 int
824 softdep_process_worklist(mp, full)
825 	struct mount *mp;
826 	int full;
827 {
828 	struct thread *td = curthread;
829 	int cnt, matchcnt, loopcount;
830 	struct ufsmount *ump;
831 	long starttime;
832 
833 	KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp"));
834 	/*
835 	 * Record the process identifier of our caller so that we can give
836 	 * this process preferential treatment in request_cleanup below.
837 	 */
838 	matchcnt = 0;
839 	ump = VFSTOUFS(mp);
840 	ACQUIRE_LOCK(&lk);
841 	loopcount = 1;
842 	starttime = time_second;
843 	while (ump->softdep_on_worklist > 0) {
844 		if ((cnt = process_worklist_item(mp, 0)) == -1)
845 			break;
846 		else
847 			matchcnt += cnt;
848 		/*
849 		 * If requested, try removing inode or removal dependencies.
850 		 */
851 		if (req_clear_inodedeps) {
852 			clear_inodedeps(td);
853 			req_clear_inodedeps -= 1;
854 			wakeup_one(&proc_waiting);
855 		}
856 		if (req_clear_remove) {
857 			clear_remove(td);
858 			req_clear_remove -= 1;
859 			wakeup_one(&proc_waiting);
860 		}
861 		/*
862 		 * We do not generally want to stop for buffer space, but if
863 		 * we are really being a buffer hog, we will stop and wait.
864 		 */
865 		if (loopcount++ % 128 == 0) {
866 			FREE_LOCK(&lk);
867 			bwillwrite();
868 			ACQUIRE_LOCK(&lk);
869 		}
870 		/*
871 		 * Never allow processing to run for more than one
872 		 * second. Otherwise the other mountpoints may get
873 		 * excessively backlogged.
874 		 */
875 		if (!full && starttime != time_second) {
876 			matchcnt = -1;
877 			break;
878 		}
879 	}
880 	FREE_LOCK(&lk);
881 	return (matchcnt);
882 }
883 
884 /*
885  * Process one item on the worklist.
886  */
887 static int
888 process_worklist_item(mp, flags)
889 	struct mount *mp;
890 	int flags;
891 {
892 	struct worklist *wk, *wkend;
893 	struct ufsmount *ump;
894 	struct vnode *vp;
895 	int matchcnt = 0;
896 
897 	mtx_assert(&lk, MA_OWNED);
898 	KASSERT(mp != NULL, ("process_worklist_item: NULL mp"));
899 	/*
900 	 * If we are being called because of a process doing a
901 	 * copy-on-write, then it is not safe to write as we may
902 	 * recurse into the copy-on-write routine.
903 	 */
904 	if (curthread->td_pflags & TDP_COWINPROGRESS)
905 		return (-1);
906 	/*
907 	 * Normally we just process each item on the worklist in order.
908 	 * However, if we are in a situation where we cannot lock any
909 	 * inodes, we have to skip over any dirrem requests whose
910 	 * vnodes are resident and locked.
911 	 */
912 	ump = VFSTOUFS(mp);
913 	vp = NULL;
914 	LIST_FOREACH(wk, &ump->softdep_workitem_pending, wk_list) {
915 		if (wk->wk_state & INPROGRESS)
916 			continue;
917 		if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
918 			break;
919 		wk->wk_state |= INPROGRESS;
920 		ump->softdep_on_worklist_inprogress++;
921 		FREE_LOCK(&lk);
922 		ffs_vget(mp, WK_DIRREM(wk)->dm_oldinum,
923 		    LK_NOWAIT | LK_EXCLUSIVE, &vp);
924 		ACQUIRE_LOCK(&lk);
925 		wk->wk_state &= ~INPROGRESS;
926 		ump->softdep_on_worklist_inprogress--;
927 		if (vp != NULL)
928 			break;
929 	}
930 	if (wk == 0)
931 		return (-1);
932 	/*
933 	 * Remove the item to be processed. If we are removing the last
934 	 * item on the list, we need to recalculate the tail pointer.
935 	 * As this happens rarely and usually when the list is short,
936 	 * we just run down the list to find it rather than tracking it
937 	 * in the above loop.
938 	 */
939 	WORKLIST_REMOVE(wk);
940 	if (wk == ump->softdep_worklist_tail) {
941 		LIST_FOREACH(wkend, &ump->softdep_workitem_pending, wk_list)
942 			if (LIST_NEXT(wkend, wk_list) == NULL)
943 				break;
944 		ump->softdep_worklist_tail = wkend;
945 	}
946 	ump->softdep_on_worklist -= 1;
947 	FREE_LOCK(&lk);
948 	if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
949 		panic("process_worklist_item: suspended filesystem");
950 	matchcnt++;
951 	switch (wk->wk_type) {
952 
953 	case D_DIRREM:
954 		/* removal of a directory entry */
955 		handle_workitem_remove(WK_DIRREM(wk), vp);
956 		break;
957 
958 	case D_FREEBLKS:
959 		/* releasing blocks and/or fragments from a file */
960 		handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT);
961 		break;
962 
963 	case D_FREEFRAG:
964 		/* releasing a fragment when replaced as a file grows */
965 		handle_workitem_freefrag(WK_FREEFRAG(wk));
966 		break;
967 
968 	case D_FREEFILE:
969 		/* releasing an inode when its link count drops to 0 */
970 		handle_workitem_freefile(WK_FREEFILE(wk));
971 		break;
972 
973 	default:
974 		panic("%s_process_worklist: Unknown type %s",
975 		    "softdep", TYPENAME(wk->wk_type));
976 		/* NOTREACHED */
977 	}
978 	vn_finished_secondary_write(mp);
979 	ACQUIRE_LOCK(&lk);
980 	return (matchcnt);
981 }
982 
983 /*
984  * Move dependencies from one buffer to another.
985  */
986 void
987 softdep_move_dependencies(oldbp, newbp)
988 	struct buf *oldbp;
989 	struct buf *newbp;
990 {
991 	struct worklist *wk, *wktail;
992 
993 	if (!LIST_EMPTY(&newbp->b_dep))
994 		panic("softdep_move_dependencies: need merge code");
995 	wktail = 0;
996 	ACQUIRE_LOCK(&lk);
997 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
998 		LIST_REMOVE(wk, wk_list);
999 		if (wktail == 0)
1000 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
1001 		else
1002 			LIST_INSERT_AFTER(wktail, wk, wk_list);
1003 		wktail = wk;
1004 	}
1005 	FREE_LOCK(&lk);
1006 }
1007 
1008 /*
1009  * Purge the work list of all items associated with a particular mount point.
1010  */
1011 int
1012 softdep_flushworklist(oldmnt, countp, td)
1013 	struct mount *oldmnt;
1014 	int *countp;
1015 	struct thread *td;
1016 {
1017 	struct vnode *devvp;
1018 	int count, error = 0;
1019 	struct ufsmount *ump;
1020 
1021 	/*
1022 	 * Alternately flush the block device associated with the mount
1023 	 * point and process any dependencies that the flushing
1024 	 * creates. We continue until no more worklist dependencies
1025 	 * are found.
1026 	 */
1027 	*countp = 0;
1028 	ump = VFSTOUFS(oldmnt);
1029 	devvp = ump->um_devvp;
1030 	while ((count = softdep_process_worklist(oldmnt, 1)) > 0) {
1031 		*countp += count;
1032 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td);
1033 		error = VOP_FSYNC(devvp, MNT_WAIT, td);
1034 		VOP_UNLOCK(devvp, 0, td);
1035 		if (error)
1036 			break;
1037 	}
1038 	return (error);
1039 }
1040 
1041 int
1042 softdep_waitidle(struct mount *mp)
1043 {
1044 	struct ufsmount *ump;
1045 	int error;
1046 	int i;
1047 
1048 	ump = VFSTOUFS(mp);
1049 	ACQUIRE_LOCK(&lk);
1050 	for (i = 0; i < 10 && ump->softdep_deps; i++) {
1051 		ump->softdep_req = 1;
1052 		if (ump->softdep_on_worklist)
1053 			panic("softdep_waitidle: work added after flush.");
1054 		msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1);
1055 	}
1056 	ump->softdep_req = 0;
1057 	FREE_LOCK(&lk);
1058 	error = 0;
1059 	if (i == 10) {
1060 		error = EBUSY;
1061 		printf("softdep_waitidle: Failed to flush worklist for %p\n",
1062 		    mp);
1063 	}
1064 
1065 	return (error);
1066 }
1067 
1068 /*
1069  * Flush all vnodes and worklist items associated with a specified mount point.
1070  */
1071 int
1072 softdep_flushfiles(oldmnt, flags, td)
1073 	struct mount *oldmnt;
1074 	int flags;
1075 	struct thread *td;
1076 {
1077 	int error, count, loopcnt;
1078 
1079 	error = 0;
1080 
1081 	/*
1082 	 * Alternately flush the vnodes associated with the mount
1083 	 * point and process any dependencies that the flushing
1084 	 * creates. In theory, this loop can happen at most twice,
1085 	 * but we give it a few extra just to be sure.
1086 	 */
1087 	for (loopcnt = 10; loopcnt > 0; loopcnt--) {
1088 		/*
1089 		 * Do another flush in case any vnodes were brought in
1090 		 * as part of the cleanup operations.
1091 		 */
1092 		if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
1093 			break;
1094 		if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 ||
1095 		    count == 0)
1096 			break;
1097 	}
1098 	/*
1099 	 * If we are unmounting then it is an error to fail. If we
1100 	 * are simply trying to downgrade to read-only, then filesystem
1101 	 * activity can keep us busy forever, so we just fail with EBUSY.
1102 	 */
1103 	if (loopcnt == 0) {
1104 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
1105 			panic("softdep_flushfiles: looping");
1106 		error = EBUSY;
1107 	}
1108 	if (!error)
1109 		error = softdep_waitidle(oldmnt);
1110 	return (error);
1111 }
1112 
1113 /*
1114  * Structure hashing.
1115  *
1116  * There are three types of structures that can be looked up:
1117  *	1) pagedep structures identified by mount point, inode number,
1118  *	   and logical block.
1119  *	2) inodedep structures identified by mount point and inode number.
1120  *	3) newblk structures identified by mount point and
1121  *	   physical block number.
1122  *
1123  * The "pagedep" and "inodedep" dependency structures are hashed
1124  * separately from the file blocks and inodes to which they correspond.
1125  * This separation helps when the in-memory copy of an inode or
1126  * file block must be replaced. It also obviates the need to access
1127  * an inode or file page when simply updating (or de-allocating)
1128  * dependency structures. Lookup of newblk structures is needed to
1129  * find newly allocated blocks when trying to associate them with
1130  * their allocdirect or allocindir structure.
1131  *
1132  * The lookup routines optionally create and hash a new instance when
1133  * an existing entry is not found.
1134  */
1135 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
1136 #define NODELAY		0x0002	/* cannot do background work */
1137 
1138 /*
1139  * Structures and routines associated with pagedep caching.
1140  */
1141 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
1142 u_long	pagedep_hash;		/* size of hash table - 1 */
1143 #define	PAGEDEP_HASH(mp, inum, lbn) \
1144 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
1145 	    pagedep_hash])
1146 
1147 static int
1148 pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp)
1149 	struct pagedep_hashhead *pagedephd;
1150 	ino_t ino;
1151 	ufs_lbn_t lbn;
1152 	struct mount *mp;
1153 	int flags;
1154 	struct pagedep **pagedeppp;
1155 {
1156 	struct pagedep *pagedep;
1157 
1158 	LIST_FOREACH(pagedep, pagedephd, pd_hash)
1159 		if (ino == pagedep->pd_ino &&
1160 		    lbn == pagedep->pd_lbn &&
1161 		    mp == pagedep->pd_list.wk_mp)
1162 			break;
1163 	if (pagedep) {
1164 		*pagedeppp = pagedep;
1165 		if ((flags & DEPALLOC) != 0 &&
1166 		    (pagedep->pd_state & ONWORKLIST) == 0)
1167 			return (0);
1168 		return (1);
1169 	}
1170 	*pagedeppp = NULL;
1171 	return (0);
1172 }
1173 /*
1174  * Look up a pagedep. Return 1 if found, 0 if not found or found
1175  * when asked to allocate but not associated with any buffer.
1176  * If not found, allocate if DEPALLOC flag is passed.
1177  * Found or allocated entry is returned in pagedeppp.
1178  * This routine must be called with splbio interrupts blocked.
1179  */
1180 static int
1181 pagedep_lookup(ip, lbn, flags, pagedeppp)
1182 	struct inode *ip;
1183 	ufs_lbn_t lbn;
1184 	int flags;
1185 	struct pagedep **pagedeppp;
1186 {
1187 	struct pagedep *pagedep;
1188 	struct pagedep_hashhead *pagedephd;
1189 	struct mount *mp;
1190 	int ret;
1191 	int i;
1192 
1193 	mtx_assert(&lk, MA_OWNED);
1194 	mp = ITOV(ip)->v_mount;
1195 	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
1196 
1197 	ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp);
1198 	if (*pagedeppp || (flags & DEPALLOC) == 0)
1199 		return (ret);
1200 	FREE_LOCK(&lk);
1201 	MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep),
1202 	    M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO);
1203 	workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp);
1204 	ACQUIRE_LOCK(&lk);
1205 	ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp);
1206 	if (*pagedeppp) {
1207 		WORKITEM_FREE(pagedep, D_PAGEDEP);
1208 		return (ret);
1209 	}
1210 	pagedep->pd_ino = ip->i_number;
1211 	pagedep->pd_lbn = lbn;
1212 	LIST_INIT(&pagedep->pd_dirremhd);
1213 	LIST_INIT(&pagedep->pd_pendinghd);
1214 	for (i = 0; i < DAHASHSZ; i++)
1215 		LIST_INIT(&pagedep->pd_diraddhd[i]);
1216 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
1217 	*pagedeppp = pagedep;
1218 	return (0);
1219 }
1220 
1221 /*
1222  * Structures and routines associated with inodedep caching.
1223  */
1224 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
1225 static u_long	inodedep_hash;	/* size of hash table - 1 */
1226 static long	num_inodedep;	/* number of inodedep allocated */
1227 #define	INODEDEP_HASH(fs, inum) \
1228       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
1229 
1230 static int
1231 inodedep_find(inodedephd, fs, inum, inodedeppp)
1232 	struct inodedep_hashhead *inodedephd;
1233 	struct fs *fs;
1234 	ino_t inum;
1235 	struct inodedep **inodedeppp;
1236 {
1237 	struct inodedep *inodedep;
1238 
1239 	LIST_FOREACH(inodedep, inodedephd, id_hash)
1240 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
1241 			break;
1242 	if (inodedep) {
1243 		*inodedeppp = inodedep;
1244 		return (1);
1245 	}
1246 	*inodedeppp = NULL;
1247 
1248 	return (0);
1249 }
1250 /*
1251  * Look up an inodedep. Return 1 if found, 0 if not found.
1252  * If not found, allocate if DEPALLOC flag is passed.
1253  * Found or allocated entry is returned in inodedeppp.
1254  * This routine must be called with splbio interrupts blocked.
1255  */
1256 static int
1257 inodedep_lookup(mp, inum, flags, inodedeppp)
1258 	struct mount *mp;
1259 	ino_t inum;
1260 	int flags;
1261 	struct inodedep **inodedeppp;
1262 {
1263 	struct inodedep *inodedep;
1264 	struct inodedep_hashhead *inodedephd;
1265 	struct fs *fs;
1266 
1267 	mtx_assert(&lk, MA_OWNED);
1268 	fs = VFSTOUFS(mp)->um_fs;
1269 	inodedephd = INODEDEP_HASH(fs, inum);
1270 
1271 	if (inodedep_find(inodedephd, fs, inum, inodedeppp))
1272 		return (1);
1273 	if ((flags & DEPALLOC) == 0)
1274 		return (0);
1275 	/*
1276 	 * If we are over our limit, try to improve the situation.
1277 	 */
1278 	if (num_inodedep > max_softdeps && (flags & NODELAY) == 0)
1279 		request_cleanup(mp, FLUSH_INODES);
1280 	FREE_LOCK(&lk);
1281 	MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1282 		M_INODEDEP, M_SOFTDEP_FLAGS);
1283 	workitem_alloc(&inodedep->id_list, D_INODEDEP, mp);
1284 	ACQUIRE_LOCK(&lk);
1285 	if (inodedep_find(inodedephd, fs, inum, inodedeppp)) {
1286 		WORKITEM_FREE(inodedep, D_INODEDEP);
1287 		return (1);
1288 	}
1289 	num_inodedep += 1;
1290 	inodedep->id_fs = fs;
1291 	inodedep->id_ino = inum;
1292 	inodedep->id_state = ALLCOMPLETE;
1293 	inodedep->id_nlinkdelta = 0;
1294 	inodedep->id_savedino1 = NULL;
1295 	inodedep->id_savedsize = -1;
1296 	inodedep->id_savedextsize = -1;
1297 	inodedep->id_buf = NULL;
1298 	LIST_INIT(&inodedep->id_pendinghd);
1299 	LIST_INIT(&inodedep->id_inowait);
1300 	LIST_INIT(&inodedep->id_bufwait);
1301 	TAILQ_INIT(&inodedep->id_inoupdt);
1302 	TAILQ_INIT(&inodedep->id_newinoupdt);
1303 	TAILQ_INIT(&inodedep->id_extupdt);
1304 	TAILQ_INIT(&inodedep->id_newextupdt);
1305 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1306 	*inodedeppp = inodedep;
1307 	return (0);
1308 }
1309 
1310 /*
1311  * Structures and routines associated with newblk caching.
1312  */
1313 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1314 u_long	newblk_hash;		/* size of hash table - 1 */
1315 #define	NEWBLK_HASH(fs, inum) \
1316 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1317 
1318 static int
1319 newblk_find(newblkhd, fs, newblkno, newblkpp)
1320 	struct newblk_hashhead *newblkhd;
1321 	struct fs *fs;
1322 	ufs2_daddr_t newblkno;
1323 	struct newblk **newblkpp;
1324 {
1325 	struct newblk *newblk;
1326 
1327 	LIST_FOREACH(newblk, newblkhd, nb_hash)
1328 		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1329 			break;
1330 	if (newblk) {
1331 		*newblkpp = newblk;
1332 		return (1);
1333 	}
1334 	*newblkpp = NULL;
1335 	return (0);
1336 }
1337 
1338 /*
1339  * Look up a newblk. Return 1 if found, 0 if not found.
1340  * If not found, allocate if DEPALLOC flag is passed.
1341  * Found or allocated entry is returned in newblkpp.
1342  */
1343 static int
1344 newblk_lookup(fs, newblkno, flags, newblkpp)
1345 	struct fs *fs;
1346 	ufs2_daddr_t newblkno;
1347 	int flags;
1348 	struct newblk **newblkpp;
1349 {
1350 	struct newblk *newblk;
1351 	struct newblk_hashhead *newblkhd;
1352 
1353 	newblkhd = NEWBLK_HASH(fs, newblkno);
1354 	if (newblk_find(newblkhd, fs, newblkno, newblkpp))
1355 		return (1);
1356 	if ((flags & DEPALLOC) == 0)
1357 		return (0);
1358 	FREE_LOCK(&lk);
1359 	MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1360 		M_NEWBLK, M_SOFTDEP_FLAGS);
1361 	ACQUIRE_LOCK(&lk);
1362 	if (newblk_find(newblkhd, fs, newblkno, newblkpp)) {
1363 		FREE(newblk, M_NEWBLK);
1364 		return (1);
1365 	}
1366 	newblk->nb_state = 0;
1367 	newblk->nb_fs = fs;
1368 	newblk->nb_newblkno = newblkno;
1369 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1370 	*newblkpp = newblk;
1371 	return (0);
1372 }
1373 
1374 /*
1375  * Executed during filesystem system initialization before
1376  * mounting any filesystems.
1377  */
1378 void
1379 softdep_initialize()
1380 {
1381 
1382 	LIST_INIT(&mkdirlisthd);
1383 	max_softdeps = desiredvnodes * 4;
1384 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1385 	    &pagedep_hash);
1386 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1387 	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1388 
1389 	/* initialise bioops hack */
1390 	bioops.io_start = softdep_disk_io_initiation;
1391 	bioops.io_complete = softdep_disk_write_complete;
1392 	bioops.io_deallocate = softdep_deallocate_dependencies;
1393 	bioops.io_countdeps = softdep_count_dependencies;
1394 }
1395 
1396 /*
1397  * Executed after all filesystems have been unmounted during
1398  * filesystem module unload.
1399  */
1400 void
1401 softdep_uninitialize()
1402 {
1403 
1404 	hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash);
1405 	hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash);
1406 	hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash);
1407 }
1408 
1409 /*
1410  * Called at mount time to notify the dependency code that a
1411  * filesystem wishes to use it.
1412  */
1413 int
1414 softdep_mount(devvp, mp, fs, cred)
1415 	struct vnode *devvp;
1416 	struct mount *mp;
1417 	struct fs *fs;
1418 	struct ucred *cred;
1419 {
1420 	struct csum_total cstotal;
1421 	struct ufsmount *ump;
1422 	struct cg *cgp;
1423 	struct buf *bp;
1424 	int error, cyl;
1425 
1426 	MNT_ILOCK(mp);
1427 	mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP;
1428 	if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) {
1429 		mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) |
1430 			MNTK_SOFTDEP;
1431 		mp->mnt_noasync++;
1432 	}
1433 	MNT_IUNLOCK(mp);
1434 	ump = VFSTOUFS(mp);
1435 	LIST_INIT(&ump->softdep_workitem_pending);
1436 	ump->softdep_worklist_tail = NULL;
1437 	ump->softdep_on_worklist = 0;
1438 	ump->softdep_deps = 0;
1439 	/*
1440 	 * When doing soft updates, the counters in the
1441 	 * superblock may have gotten out of sync. Recomputation
1442 	 * can take a long time and can be deferred for background
1443 	 * fsck.  However, the old behavior of scanning the cylinder
1444 	 * groups and recalculating them at mount time is available
1445 	 * by setting vfs.ffs.compute_summary_at_mount to one.
1446 	 */
1447 	if (compute_summary_at_mount == 0 || fs->fs_clean != 0)
1448 		return (0);
1449 	bzero(&cstotal, sizeof cstotal);
1450 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1451 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1452 		    fs->fs_cgsize, cred, &bp)) != 0) {
1453 			brelse(bp);
1454 			return (error);
1455 		}
1456 		cgp = (struct cg *)bp->b_data;
1457 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1458 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1459 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1460 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1461 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
1462 		brelse(bp);
1463 	}
1464 #ifdef DEBUG
1465 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1466 		printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
1467 #endif
1468 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1469 	return (0);
1470 }
1471 
1472 /*
1473  * Protecting the freemaps (or bitmaps).
1474  *
1475  * To eliminate the need to execute fsck before mounting a filesystem
1476  * after a power failure, one must (conservatively) guarantee that the
1477  * on-disk copy of the bitmaps never indicate that a live inode or block is
1478  * free.  So, when a block or inode is allocated, the bitmap should be
1479  * updated (on disk) before any new pointers.  When a block or inode is
1480  * freed, the bitmap should not be updated until all pointers have been
1481  * reset.  The latter dependency is handled by the delayed de-allocation
1482  * approach described below for block and inode de-allocation.  The former
1483  * dependency is handled by calling the following procedure when a block or
1484  * inode is allocated. When an inode is allocated an "inodedep" is created
1485  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1486  * Each "inodedep" is also inserted into the hash indexing structure so
1487  * that any additional link additions can be made dependent on the inode
1488  * allocation.
1489  *
1490  * The ufs filesystem maintains a number of free block counts (e.g., per
1491  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1492  * in addition to the bitmaps.  These counts are used to improve efficiency
1493  * during allocation and therefore must be consistent with the bitmaps.
1494  * There is no convenient way to guarantee post-crash consistency of these
1495  * counts with simple update ordering, for two main reasons: (1) The counts
1496  * and bitmaps for a single cylinder group block are not in the same disk
1497  * sector.  If a disk write is interrupted (e.g., by power failure), one may
1498  * be written and the other not.  (2) Some of the counts are located in the
1499  * superblock rather than the cylinder group block. So, we focus our soft
1500  * updates implementation on protecting the bitmaps. When mounting a
1501  * filesystem, we recompute the auxiliary counts from the bitmaps.
1502  */
1503 
1504 /*
1505  * Called just after updating the cylinder group block to allocate an inode.
1506  */
1507 void
1508 softdep_setup_inomapdep(bp, ip, newinum)
1509 	struct buf *bp;		/* buffer for cylgroup block with inode map */
1510 	struct inode *ip;	/* inode related to allocation */
1511 	ino_t newinum;		/* new inode number being allocated */
1512 {
1513 	struct inodedep *inodedep;
1514 	struct bmsafemap *bmsafemap;
1515 
1516 	/*
1517 	 * Create a dependency for the newly allocated inode.
1518 	 * Panic if it already exists as something is seriously wrong.
1519 	 * Otherwise add it to the dependency list for the buffer holding
1520 	 * the cylinder group map from which it was allocated.
1521 	 */
1522 	ACQUIRE_LOCK(&lk);
1523 	if ((inodedep_lookup(UFSTOVFS(ip->i_ump), newinum, DEPALLOC|NODELAY,
1524 	    &inodedep)))
1525 		panic("softdep_setup_inomapdep: dependency for new inode "
1526 		    "already exists");
1527 	inodedep->id_buf = bp;
1528 	inodedep->id_state &= ~DEPCOMPLETE;
1529 	bmsafemap = bmsafemap_lookup(inodedep->id_list.wk_mp, bp);
1530 	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1531 	FREE_LOCK(&lk);
1532 }
1533 
1534 /*
1535  * Called just after updating the cylinder group block to
1536  * allocate block or fragment.
1537  */
1538 void
1539 softdep_setup_blkmapdep(bp, mp, newblkno)
1540 	struct buf *bp;		/* buffer for cylgroup block with block map */
1541 	struct mount *mp;	/* filesystem doing allocation */
1542 	ufs2_daddr_t newblkno;	/* number of newly allocated block */
1543 {
1544 	struct newblk *newblk;
1545 	struct bmsafemap *bmsafemap;
1546 	struct fs *fs;
1547 
1548 	fs = VFSTOUFS(mp)->um_fs;
1549 	/*
1550 	 * Create a dependency for the newly allocated block.
1551 	 * Add it to the dependency list for the buffer holding
1552 	 * the cylinder group map from which it was allocated.
1553 	 */
1554 	ACQUIRE_LOCK(&lk);
1555 	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1556 		panic("softdep_setup_blkmapdep: found block");
1557 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp);
1558 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1559 	FREE_LOCK(&lk);
1560 }
1561 
1562 /*
1563  * Find the bmsafemap associated with a cylinder group buffer.
1564  * If none exists, create one. The buffer must be locked when
1565  * this routine is called and this routine must be called with
1566  * splbio interrupts blocked.
1567  */
1568 static struct bmsafemap *
1569 bmsafemap_lookup(mp, bp)
1570 	struct mount *mp;
1571 	struct buf *bp;
1572 {
1573 	struct bmsafemap *bmsafemap;
1574 	struct worklist *wk;
1575 
1576 	mtx_assert(&lk, MA_OWNED);
1577 	LIST_FOREACH(wk, &bp->b_dep, wk_list)
1578 		if (wk->wk_type == D_BMSAFEMAP)
1579 			return (WK_BMSAFEMAP(wk));
1580 	FREE_LOCK(&lk);
1581 	MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1582 		M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1583 	workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
1584 	bmsafemap->sm_buf = bp;
1585 	LIST_INIT(&bmsafemap->sm_allocdirecthd);
1586 	LIST_INIT(&bmsafemap->sm_allocindirhd);
1587 	LIST_INIT(&bmsafemap->sm_inodedephd);
1588 	LIST_INIT(&bmsafemap->sm_newblkhd);
1589 	ACQUIRE_LOCK(&lk);
1590 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1591 	return (bmsafemap);
1592 }
1593 
1594 /*
1595  * Direct block allocation dependencies.
1596  *
1597  * When a new block is allocated, the corresponding disk locations must be
1598  * initialized (with zeros or new data) before the on-disk inode points to
1599  * them.  Also, the freemap from which the block was allocated must be
1600  * updated (on disk) before the inode's pointer. These two dependencies are
1601  * independent of each other and are needed for all file blocks and indirect
1602  * blocks that are pointed to directly by the inode.  Just before the
1603  * "in-core" version of the inode is updated with a newly allocated block
1604  * number, a procedure (below) is called to setup allocation dependency
1605  * structures.  These structures are removed when the corresponding
1606  * dependencies are satisfied or when the block allocation becomes obsolete
1607  * (i.e., the file is deleted, the block is de-allocated, or the block is a
1608  * fragment that gets upgraded).  All of these cases are handled in
1609  * procedures described later.
1610  *
1611  * When a file extension causes a fragment to be upgraded, either to a larger
1612  * fragment or to a full block, the on-disk location may change (if the
1613  * previous fragment could not simply be extended). In this case, the old
1614  * fragment must be de-allocated, but not until after the inode's pointer has
1615  * been updated. In most cases, this is handled by later procedures, which
1616  * will construct a "freefrag" structure to be added to the workitem queue
1617  * when the inode update is complete (or obsolete).  The main exception to
1618  * this is when an allocation occurs while a pending allocation dependency
1619  * (for the same block pointer) remains.  This case is handled in the main
1620  * allocation dependency setup procedure by immediately freeing the
1621  * unreferenced fragments.
1622  */
1623 void
1624 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1625 	struct inode *ip;	/* inode to which block is being added */
1626 	ufs_lbn_t lbn;		/* block pointer within inode */
1627 	ufs2_daddr_t newblkno;	/* disk block number being added */
1628 	ufs2_daddr_t oldblkno;	/* previous block number, 0 unless frag */
1629 	long newsize;		/* size of new block */
1630 	long oldsize;		/* size of new block */
1631 	struct buf *bp;		/* bp for allocated block */
1632 {
1633 	struct allocdirect *adp, *oldadp;
1634 	struct allocdirectlst *adphead;
1635 	struct bmsafemap *bmsafemap;
1636 	struct inodedep *inodedep;
1637 	struct pagedep *pagedep;
1638 	struct newblk *newblk;
1639 	struct mount *mp;
1640 
1641 	mp = UFSTOVFS(ip->i_ump);
1642 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1643 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1644 	workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp);
1645 	adp->ad_lbn = lbn;
1646 	adp->ad_newblkno = newblkno;
1647 	adp->ad_oldblkno = oldblkno;
1648 	adp->ad_newsize = newsize;
1649 	adp->ad_oldsize = oldsize;
1650 	adp->ad_state = ATTACHED;
1651 	LIST_INIT(&adp->ad_newdirblk);
1652 	if (newblkno == oldblkno)
1653 		adp->ad_freefrag = NULL;
1654 	else
1655 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1656 
1657 	ACQUIRE_LOCK(&lk);
1658 	if (lbn >= NDADDR) {
1659 		/* allocating an indirect block */
1660 		if (oldblkno != 0)
1661 			panic("softdep_setup_allocdirect: non-zero indir");
1662 	} else {
1663 		/*
1664 		 * Allocating a direct block.
1665 		 *
1666 		 * If we are allocating a directory block, then we must
1667 		 * allocate an associated pagedep to track additions and
1668 		 * deletions.
1669 		 */
1670 		if ((ip->i_mode & IFMT) == IFDIR &&
1671 		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1672 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1673 	}
1674 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1675 		panic("softdep_setup_allocdirect: lost block");
1676 	if (newblk->nb_state == DEPCOMPLETE) {
1677 		adp->ad_state |= DEPCOMPLETE;
1678 		adp->ad_buf = NULL;
1679 	} else {
1680 		bmsafemap = newblk->nb_bmsafemap;
1681 		adp->ad_buf = bmsafemap->sm_buf;
1682 		LIST_REMOVE(newblk, nb_deps);
1683 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1684 	}
1685 	LIST_REMOVE(newblk, nb_hash);
1686 	FREE(newblk, M_NEWBLK);
1687 
1688 	inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1689 	adp->ad_inodedep = inodedep;
1690 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1691 	/*
1692 	 * The list of allocdirects must be kept in sorted and ascending
1693 	 * order so that the rollback routines can quickly determine the
1694 	 * first uncommitted block (the size of the file stored on disk
1695 	 * ends at the end of the lowest committed fragment, or if there
1696 	 * are no fragments, at the end of the highest committed block).
1697 	 * Since files generally grow, the typical case is that the new
1698 	 * block is to be added at the end of the list. We speed this
1699 	 * special case by checking against the last allocdirect in the
1700 	 * list before laboriously traversing the list looking for the
1701 	 * insertion point.
1702 	 */
1703 	adphead = &inodedep->id_newinoupdt;
1704 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1705 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1706 		/* insert at end of list */
1707 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1708 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1709 			allocdirect_merge(adphead, adp, oldadp);
1710 		FREE_LOCK(&lk);
1711 		return;
1712 	}
1713 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1714 		if (oldadp->ad_lbn >= lbn)
1715 			break;
1716 	}
1717 	if (oldadp == NULL)
1718 		panic("softdep_setup_allocdirect: lost entry");
1719 	/* insert in middle of list */
1720 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1721 	if (oldadp->ad_lbn == lbn)
1722 		allocdirect_merge(adphead, adp, oldadp);
1723 	FREE_LOCK(&lk);
1724 }
1725 
1726 /*
1727  * Replace an old allocdirect dependency with a newer one.
1728  * This routine must be called with splbio interrupts blocked.
1729  */
1730 static void
1731 allocdirect_merge(adphead, newadp, oldadp)
1732 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
1733 	struct allocdirect *newadp;	/* allocdirect being added */
1734 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
1735 {
1736 	struct worklist *wk;
1737 	struct freefrag *freefrag;
1738 	struct newdirblk *newdirblk;
1739 
1740 	mtx_assert(&lk, MA_OWNED);
1741 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1742 	    newadp->ad_oldsize != oldadp->ad_newsize ||
1743 	    newadp->ad_lbn >= NDADDR)
1744 		panic("%s %jd != new %jd || old size %ld != new %ld",
1745 		    "allocdirect_merge: old blkno",
1746 		    (intmax_t)newadp->ad_oldblkno,
1747 		    (intmax_t)oldadp->ad_newblkno,
1748 		    newadp->ad_oldsize, oldadp->ad_newsize);
1749 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
1750 	newadp->ad_oldsize = oldadp->ad_oldsize;
1751 	/*
1752 	 * If the old dependency had a fragment to free or had never
1753 	 * previously had a block allocated, then the new dependency
1754 	 * can immediately post its freefrag and adopt the old freefrag.
1755 	 * This action is done by swapping the freefrag dependencies.
1756 	 * The new dependency gains the old one's freefrag, and the
1757 	 * old one gets the new one and then immediately puts it on
1758 	 * the worklist when it is freed by free_allocdirect. It is
1759 	 * not possible to do this swap when the old dependency had a
1760 	 * non-zero size but no previous fragment to free. This condition
1761 	 * arises when the new block is an extension of the old block.
1762 	 * Here, the first part of the fragment allocated to the new
1763 	 * dependency is part of the block currently claimed on disk by
1764 	 * the old dependency, so cannot legitimately be freed until the
1765 	 * conditions for the new dependency are fulfilled.
1766 	 */
1767 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1768 		freefrag = newadp->ad_freefrag;
1769 		newadp->ad_freefrag = oldadp->ad_freefrag;
1770 		oldadp->ad_freefrag = freefrag;
1771 	}
1772 	/*
1773 	 * If we are tracking a new directory-block allocation,
1774 	 * move it from the old allocdirect to the new allocdirect.
1775 	 */
1776 	if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
1777 		newdirblk = WK_NEWDIRBLK(wk);
1778 		WORKLIST_REMOVE(&newdirblk->db_list);
1779 		if (!LIST_EMPTY(&oldadp->ad_newdirblk))
1780 			panic("allocdirect_merge: extra newdirblk");
1781 		WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list);
1782 	}
1783 	free_allocdirect(adphead, oldadp, 0);
1784 }
1785 
1786 /*
1787  * Allocate a new freefrag structure if needed.
1788  */
1789 static struct freefrag *
1790 newfreefrag(ip, blkno, size)
1791 	struct inode *ip;
1792 	ufs2_daddr_t blkno;
1793 	long size;
1794 {
1795 	struct freefrag *freefrag;
1796 	struct fs *fs;
1797 
1798 	if (blkno == 0)
1799 		return (NULL);
1800 	fs = ip->i_fs;
1801 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1802 		panic("newfreefrag: frag size");
1803 	MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1804 		M_FREEFRAG, M_SOFTDEP_FLAGS);
1805 	workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump));
1806 	freefrag->ff_inum = ip->i_number;
1807 	freefrag->ff_blkno = blkno;
1808 	freefrag->ff_fragsize = size;
1809 	return (freefrag);
1810 }
1811 
1812 /*
1813  * This workitem de-allocates fragments that were replaced during
1814  * file block allocation.
1815  */
1816 static void
1817 handle_workitem_freefrag(freefrag)
1818 	struct freefrag *freefrag;
1819 {
1820 	struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp);
1821 
1822 	ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
1823 	    freefrag->ff_fragsize, freefrag->ff_inum);
1824 	ACQUIRE_LOCK(&lk);
1825 	WORKITEM_FREE(freefrag, D_FREEFRAG);
1826 	FREE_LOCK(&lk);
1827 }
1828 
1829 /*
1830  * Set up a dependency structure for an external attributes data block.
1831  * This routine follows much of the structure of softdep_setup_allocdirect.
1832  * See the description of softdep_setup_allocdirect above for details.
1833  */
1834 void
1835 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1836 	struct inode *ip;
1837 	ufs_lbn_t lbn;
1838 	ufs2_daddr_t newblkno;
1839 	ufs2_daddr_t oldblkno;
1840 	long newsize;
1841 	long oldsize;
1842 	struct buf *bp;
1843 {
1844 	struct allocdirect *adp, *oldadp;
1845 	struct allocdirectlst *adphead;
1846 	struct bmsafemap *bmsafemap;
1847 	struct inodedep *inodedep;
1848 	struct newblk *newblk;
1849 	struct mount *mp;
1850 
1851 	mp = UFSTOVFS(ip->i_ump);
1852 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1853 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1854 	workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp);
1855 	adp->ad_lbn = lbn;
1856 	adp->ad_newblkno = newblkno;
1857 	adp->ad_oldblkno = oldblkno;
1858 	adp->ad_newsize = newsize;
1859 	adp->ad_oldsize = oldsize;
1860 	adp->ad_state = ATTACHED | EXTDATA;
1861 	LIST_INIT(&adp->ad_newdirblk);
1862 	if (newblkno == oldblkno)
1863 		adp->ad_freefrag = NULL;
1864 	else
1865 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1866 
1867 	ACQUIRE_LOCK(&lk);
1868 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1869 		panic("softdep_setup_allocext: lost block");
1870 
1871 	inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1872 	adp->ad_inodedep = inodedep;
1873 
1874 	if (newblk->nb_state == DEPCOMPLETE) {
1875 		adp->ad_state |= DEPCOMPLETE;
1876 		adp->ad_buf = NULL;
1877 	} else {
1878 		bmsafemap = newblk->nb_bmsafemap;
1879 		adp->ad_buf = bmsafemap->sm_buf;
1880 		LIST_REMOVE(newblk, nb_deps);
1881 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1882 	}
1883 	LIST_REMOVE(newblk, nb_hash);
1884 	FREE(newblk, M_NEWBLK);
1885 
1886 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1887 	if (lbn >= NXADDR)
1888 		panic("softdep_setup_allocext: lbn %lld > NXADDR",
1889 		    (long long)lbn);
1890 	/*
1891 	 * The list of allocdirects must be kept in sorted and ascending
1892 	 * order so that the rollback routines can quickly determine the
1893 	 * first uncommitted block (the size of the file stored on disk
1894 	 * ends at the end of the lowest committed fragment, or if there
1895 	 * are no fragments, at the end of the highest committed block).
1896 	 * Since files generally grow, the typical case is that the new
1897 	 * block is to be added at the end of the list. We speed this
1898 	 * special case by checking against the last allocdirect in the
1899 	 * list before laboriously traversing the list looking for the
1900 	 * insertion point.
1901 	 */
1902 	adphead = &inodedep->id_newextupdt;
1903 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1904 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1905 		/* insert at end of list */
1906 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1907 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1908 			allocdirect_merge(adphead, adp, oldadp);
1909 		FREE_LOCK(&lk);
1910 		return;
1911 	}
1912 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1913 		if (oldadp->ad_lbn >= lbn)
1914 			break;
1915 	}
1916 	if (oldadp == NULL)
1917 		panic("softdep_setup_allocext: lost entry");
1918 	/* insert in middle of list */
1919 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1920 	if (oldadp->ad_lbn == lbn)
1921 		allocdirect_merge(adphead, adp, oldadp);
1922 	FREE_LOCK(&lk);
1923 }
1924 
1925 /*
1926  * Indirect block allocation dependencies.
1927  *
1928  * The same dependencies that exist for a direct block also exist when
1929  * a new block is allocated and pointed to by an entry in a block of
1930  * indirect pointers. The undo/redo states described above are also
1931  * used here. Because an indirect block contains many pointers that
1932  * may have dependencies, a second copy of the entire in-memory indirect
1933  * block is kept. The buffer cache copy is always completely up-to-date.
1934  * The second copy, which is used only as a source for disk writes,
1935  * contains only the safe pointers (i.e., those that have no remaining
1936  * update dependencies). The second copy is freed when all pointers
1937  * are safe. The cache is not allowed to replace indirect blocks with
1938  * pending update dependencies. If a buffer containing an indirect
1939  * block with dependencies is written, these routines will mark it
1940  * dirty again. It can only be successfully written once all the
1941  * dependencies are removed. The ffs_fsync routine in conjunction with
1942  * softdep_sync_metadata work together to get all the dependencies
1943  * removed so that a file can be successfully written to disk. Three
1944  * procedures are used when setting up indirect block pointer
1945  * dependencies. The division is necessary because of the organization
1946  * of the "balloc" routine and because of the distinction between file
1947  * pages and file metadata blocks.
1948  */
1949 
1950 /*
1951  * Allocate a new allocindir structure.
1952  */
1953 static struct allocindir *
1954 newallocindir(ip, ptrno, newblkno, oldblkno)
1955 	struct inode *ip;	/* inode for file being extended */
1956 	int ptrno;		/* offset of pointer in indirect block */
1957 	ufs2_daddr_t newblkno;	/* disk block number being added */
1958 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1959 {
1960 	struct allocindir *aip;
1961 
1962 	MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1963 		M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO);
1964 	workitem_alloc(&aip->ai_list, D_ALLOCINDIR, UFSTOVFS(ip->i_ump));
1965 	aip->ai_state = ATTACHED;
1966 	aip->ai_offset = ptrno;
1967 	aip->ai_newblkno = newblkno;
1968 	aip->ai_oldblkno = oldblkno;
1969 	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1970 	return (aip);
1971 }
1972 
1973 /*
1974  * Called just before setting an indirect block pointer
1975  * to a newly allocated file page.
1976  */
1977 void
1978 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1979 	struct inode *ip;	/* inode for file being extended */
1980 	ufs_lbn_t lbn;		/* allocated block number within file */
1981 	struct buf *bp;		/* buffer with indirect blk referencing page */
1982 	int ptrno;		/* offset of pointer in indirect block */
1983 	ufs2_daddr_t newblkno;	/* disk block number being added */
1984 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1985 	struct buf *nbp;	/* buffer holding allocated page */
1986 {
1987 	struct allocindir *aip;
1988 	struct pagedep *pagedep;
1989 
1990 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page");
1991 	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1992 	ACQUIRE_LOCK(&lk);
1993 	/*
1994 	 * If we are allocating a directory page, then we must
1995 	 * allocate an associated pagedep to track additions and
1996 	 * deletions.
1997 	 */
1998 	if ((ip->i_mode & IFMT) == IFDIR &&
1999 	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
2000 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
2001 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
2002 	setup_allocindir_phase2(bp, ip, aip);
2003 	FREE_LOCK(&lk);
2004 }
2005 
2006 /*
2007  * Called just before setting an indirect block pointer to a
2008  * newly allocated indirect block.
2009  */
2010 void
2011 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
2012 	struct buf *nbp;	/* newly allocated indirect block */
2013 	struct inode *ip;	/* inode for file being extended */
2014 	struct buf *bp;		/* indirect block referencing allocated block */
2015 	int ptrno;		/* offset of pointer in indirect block */
2016 	ufs2_daddr_t newblkno;	/* disk block number being added */
2017 {
2018 	struct allocindir *aip;
2019 
2020 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta");
2021 	aip = newallocindir(ip, ptrno, newblkno, 0);
2022 	ACQUIRE_LOCK(&lk);
2023 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
2024 	setup_allocindir_phase2(bp, ip, aip);
2025 	FREE_LOCK(&lk);
2026 }
2027 
2028 /*
2029  * Called to finish the allocation of the "aip" allocated
2030  * by one of the two routines above.
2031  */
2032 static void
2033 setup_allocindir_phase2(bp, ip, aip)
2034 	struct buf *bp;		/* in-memory copy of the indirect block */
2035 	struct inode *ip;	/* inode for file being extended */
2036 	struct allocindir *aip;	/* allocindir allocated by the above routines */
2037 {
2038 	struct worklist *wk;
2039 	struct indirdep *indirdep, *newindirdep;
2040 	struct bmsafemap *bmsafemap;
2041 	struct allocindir *oldaip;
2042 	struct freefrag *freefrag;
2043 	struct newblk *newblk;
2044 	ufs2_daddr_t blkno;
2045 
2046 	mtx_assert(&lk, MA_OWNED);
2047 	if (bp->b_lblkno >= 0)
2048 		panic("setup_allocindir_phase2: not indir blk");
2049 	for (indirdep = NULL, newindirdep = NULL; ; ) {
2050 		LIST_FOREACH(wk, &bp->b_dep, wk_list) {
2051 			if (wk->wk_type != D_INDIRDEP)
2052 				continue;
2053 			indirdep = WK_INDIRDEP(wk);
2054 			break;
2055 		}
2056 		if (indirdep == NULL && newindirdep) {
2057 			indirdep = newindirdep;
2058 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
2059 			newindirdep = NULL;
2060 		}
2061 		if (indirdep) {
2062 			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
2063 			    &newblk) == 0)
2064 				panic("setup_allocindir: lost block");
2065 			if (newblk->nb_state == DEPCOMPLETE) {
2066 				aip->ai_state |= DEPCOMPLETE;
2067 				aip->ai_buf = NULL;
2068 			} else {
2069 				bmsafemap = newblk->nb_bmsafemap;
2070 				aip->ai_buf = bmsafemap->sm_buf;
2071 				LIST_REMOVE(newblk, nb_deps);
2072 				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
2073 				    aip, ai_deps);
2074 			}
2075 			LIST_REMOVE(newblk, nb_hash);
2076 			FREE(newblk, M_NEWBLK);
2077 			aip->ai_indirdep = indirdep;
2078 			/*
2079 			 * Check to see if there is an existing dependency
2080 			 * for this block. If there is, merge the old
2081 			 * dependency into the new one.
2082 			 */
2083 			if (aip->ai_oldblkno == 0)
2084 				oldaip = NULL;
2085 			else
2086 
2087 				LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
2088 					if (oldaip->ai_offset == aip->ai_offset)
2089 						break;
2090 			freefrag = NULL;
2091 			if (oldaip != NULL) {
2092 				if (oldaip->ai_newblkno != aip->ai_oldblkno)
2093 					panic("setup_allocindir_phase2: blkno");
2094 				aip->ai_oldblkno = oldaip->ai_oldblkno;
2095 				freefrag = aip->ai_freefrag;
2096 				aip->ai_freefrag = oldaip->ai_freefrag;
2097 				oldaip->ai_freefrag = NULL;
2098 				free_allocindir(oldaip, NULL);
2099 			}
2100 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
2101 			if (ip->i_ump->um_fstype == UFS1)
2102 				((ufs1_daddr_t *)indirdep->ir_savebp->b_data)
2103 				    [aip->ai_offset] = aip->ai_oldblkno;
2104 			else
2105 				((ufs2_daddr_t *)indirdep->ir_savebp->b_data)
2106 				    [aip->ai_offset] = aip->ai_oldblkno;
2107 			FREE_LOCK(&lk);
2108 			if (freefrag != NULL)
2109 				handle_workitem_freefrag(freefrag);
2110 		} else
2111 			FREE_LOCK(&lk);
2112 		if (newindirdep) {
2113 			newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
2114 			brelse(newindirdep->ir_savebp);
2115 			ACQUIRE_LOCK(&lk);
2116 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
2117 			if (indirdep)
2118 				break;
2119 			FREE_LOCK(&lk);
2120 		}
2121 		if (indirdep) {
2122 			ACQUIRE_LOCK(&lk);
2123 			break;
2124 		}
2125 		MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
2126 			M_INDIRDEP, M_SOFTDEP_FLAGS);
2127 		workitem_alloc(&newindirdep->ir_list, D_INDIRDEP,
2128 		    UFSTOVFS(ip->i_ump));
2129 		newindirdep->ir_state = ATTACHED;
2130 		if (ip->i_ump->um_fstype == UFS1)
2131 			newindirdep->ir_state |= UFS1FMT;
2132 		LIST_INIT(&newindirdep->ir_deplisthd);
2133 		LIST_INIT(&newindirdep->ir_donehd);
2134 		if (bp->b_blkno == bp->b_lblkno) {
2135 			ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
2136 			    NULL, NULL);
2137 			bp->b_blkno = blkno;
2138 		}
2139 		newindirdep->ir_savebp =
2140 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
2141 		BUF_KERNPROC(newindirdep->ir_savebp);
2142 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
2143 		ACQUIRE_LOCK(&lk);
2144 	}
2145 }
2146 
2147 /*
2148  * Block de-allocation dependencies.
2149  *
2150  * When blocks are de-allocated, the on-disk pointers must be nullified before
2151  * the blocks are made available for use by other files.  (The true
2152  * requirement is that old pointers must be nullified before new on-disk
2153  * pointers are set.  We chose this slightly more stringent requirement to
2154  * reduce complexity.) Our implementation handles this dependency by updating
2155  * the inode (or indirect block) appropriately but delaying the actual block
2156  * de-allocation (i.e., freemap and free space count manipulation) until
2157  * after the updated versions reach stable storage.  After the disk is
2158  * updated, the blocks can be safely de-allocated whenever it is convenient.
2159  * This implementation handles only the common case of reducing a file's
2160  * length to zero. Other cases are handled by the conventional synchronous
2161  * write approach.
2162  *
2163  * The ffs implementation with which we worked double-checks
2164  * the state of the block pointers and file size as it reduces
2165  * a file's length.  Some of this code is replicated here in our
2166  * soft updates implementation.  The freeblks->fb_chkcnt field is
2167  * used to transfer a part of this information to the procedure
2168  * that eventually de-allocates the blocks.
2169  *
2170  * This routine should be called from the routine that shortens
2171  * a file's length, before the inode's size or block pointers
2172  * are modified. It will save the block pointer information for
2173  * later release and zero the inode so that the calling routine
2174  * can release it.
2175  */
2176 void
2177 softdep_setup_freeblocks(ip, length, flags)
2178 	struct inode *ip;	/* The inode whose length is to be reduced */
2179 	off_t length;		/* The new length for the file */
2180 	int flags;		/* IO_EXT and/or IO_NORMAL */
2181 {
2182 	struct freeblks *freeblks;
2183 	struct inodedep *inodedep;
2184 	struct allocdirect *adp;
2185 	struct vnode *vp;
2186 	struct buf *bp;
2187 	struct fs *fs;
2188 	ufs2_daddr_t extblocks, datablocks;
2189 	struct mount *mp;
2190 	int i, delay, error;
2191 
2192 	fs = ip->i_fs;
2193 	mp = UFSTOVFS(ip->i_ump);
2194 	if (length != 0)
2195 		panic("softdep_setup_freeblocks: non-zero length");
2196 	MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
2197 		M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
2198 	workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp);
2199 	freeblks->fb_state = ATTACHED;
2200 	freeblks->fb_uid = ip->i_uid;
2201 	freeblks->fb_previousinum = ip->i_number;
2202 	freeblks->fb_devvp = ip->i_devvp;
2203 	extblocks = 0;
2204 	if (fs->fs_magic == FS_UFS2_MAGIC)
2205 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
2206 	datablocks = DIP(ip, i_blocks) - extblocks;
2207 	if ((flags & IO_NORMAL) == 0) {
2208 		freeblks->fb_oldsize = 0;
2209 		freeblks->fb_chkcnt = 0;
2210 	} else {
2211 		freeblks->fb_oldsize = ip->i_size;
2212 		ip->i_size = 0;
2213 		DIP_SET(ip, i_size, 0);
2214 		freeblks->fb_chkcnt = datablocks;
2215 		for (i = 0; i < NDADDR; i++) {
2216 			freeblks->fb_dblks[i] = DIP(ip, i_db[i]);
2217 			DIP_SET(ip, i_db[i], 0);
2218 		}
2219 		for (i = 0; i < NIADDR; i++) {
2220 			freeblks->fb_iblks[i] = DIP(ip, i_ib[i]);
2221 			DIP_SET(ip, i_ib[i], 0);
2222 		}
2223 		/*
2224 		 * If the file was removed, then the space being freed was
2225 		 * accounted for then (see softdep_releasefile()). If the
2226 		 * file is merely being truncated, then we account for it now.
2227 		 */
2228 		if ((ip->i_flag & IN_SPACECOUNTED) == 0) {
2229 			UFS_LOCK(ip->i_ump);
2230 			fs->fs_pendingblocks += datablocks;
2231 			UFS_UNLOCK(ip->i_ump);
2232 		}
2233 	}
2234 	if ((flags & IO_EXT) == 0) {
2235 		freeblks->fb_oldextsize = 0;
2236 	} else {
2237 		freeblks->fb_oldextsize = ip->i_din2->di_extsize;
2238 		ip->i_din2->di_extsize = 0;
2239 		freeblks->fb_chkcnt += extblocks;
2240 		for (i = 0; i < NXADDR; i++) {
2241 			freeblks->fb_eblks[i] = ip->i_din2->di_extb[i];
2242 			ip->i_din2->di_extb[i] = 0;
2243 		}
2244 	}
2245 	DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt);
2246 	/*
2247 	 * Push the zero'ed inode to to its disk buffer so that we are free
2248 	 * to delete its dependencies below. Once the dependencies are gone
2249 	 * the buffer can be safely released.
2250 	 */
2251 	if ((error = bread(ip->i_devvp,
2252 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
2253 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
2254 		brelse(bp);
2255 		softdep_error("softdep_setup_freeblocks", error);
2256 	}
2257 	if (ip->i_ump->um_fstype == UFS1)
2258 		*((struct ufs1_dinode *)bp->b_data +
2259 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
2260 	else
2261 		*((struct ufs2_dinode *)bp->b_data +
2262 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
2263 	/*
2264 	 * Find and eliminate any inode dependencies.
2265 	 */
2266 	ACQUIRE_LOCK(&lk);
2267 	(void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
2268 	if ((inodedep->id_state & IOSTARTED) != 0)
2269 		panic("softdep_setup_freeblocks: inode busy");
2270 	/*
2271 	 * Add the freeblks structure to the list of operations that
2272 	 * must await the zero'ed inode being written to disk. If we
2273 	 * still have a bitmap dependency (delay == 0), then the inode
2274 	 * has never been written to disk, so we can process the
2275 	 * freeblks below once we have deleted the dependencies.
2276 	 */
2277 	delay = (inodedep->id_state & DEPCOMPLETE);
2278 	if (delay)
2279 		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
2280 	/*
2281 	 * Because the file length has been truncated to zero, any
2282 	 * pending block allocation dependency structures associated
2283 	 * with this inode are obsolete and can simply be de-allocated.
2284 	 * We must first merge the two dependency lists to get rid of
2285 	 * any duplicate freefrag structures, then purge the merged list.
2286 	 * If we still have a bitmap dependency, then the inode has never
2287 	 * been written to disk, so we can free any fragments without delay.
2288 	 */
2289 	if (flags & IO_NORMAL) {
2290 		merge_inode_lists(&inodedep->id_newinoupdt,
2291 		    &inodedep->id_inoupdt);
2292 		while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
2293 			free_allocdirect(&inodedep->id_inoupdt, adp, delay);
2294 	}
2295 	if (flags & IO_EXT) {
2296 		merge_inode_lists(&inodedep->id_newextupdt,
2297 		    &inodedep->id_extupdt);
2298 		while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
2299 			free_allocdirect(&inodedep->id_extupdt, adp, delay);
2300 	}
2301 	FREE_LOCK(&lk);
2302 	bdwrite(bp);
2303 	/*
2304 	 * We must wait for any I/O in progress to finish so that
2305 	 * all potential buffers on the dirty list will be visible.
2306 	 * Once they are all there, walk the list and get rid of
2307 	 * any dependencies.
2308 	 */
2309 	vp = ITOV(ip);
2310 	VI_LOCK(vp);
2311 	drain_output(vp);
2312 restart:
2313 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
2314 		if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
2315 		    ((flags & IO_NORMAL) == 0 &&
2316 		      (bp->b_xflags & BX_ALTDATA) == 0))
2317 			continue;
2318 		if ((bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT)) == NULL)
2319 			goto restart;
2320 		VI_UNLOCK(vp);
2321 		ACQUIRE_LOCK(&lk);
2322 		(void) inodedep_lookup(mp, ip->i_number, 0, &inodedep);
2323 		deallocate_dependencies(bp, inodedep);
2324 		FREE_LOCK(&lk);
2325 		bp->b_flags |= B_INVAL | B_NOCACHE;
2326 		brelse(bp);
2327 		VI_LOCK(vp);
2328 		goto restart;
2329 	}
2330 	VI_UNLOCK(vp);
2331 	ACQUIRE_LOCK(&lk);
2332 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
2333 		(void) free_inodedep(inodedep);
2334 
2335 	if(delay) {
2336 		freeblks->fb_state |= DEPCOMPLETE;
2337 		/*
2338 		 * If the inode with zeroed block pointers is now on disk
2339 		 * we can start freeing blocks. Add freeblks to the worklist
2340 		 * instead of calling  handle_workitem_freeblocks directly as
2341 		 * it is more likely that additional IO is needed to complete
2342 		 * the request here than in the !delay case.
2343 		 */
2344 		if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
2345 			add_to_worklist(&freeblks->fb_list);
2346 	}
2347 
2348 	FREE_LOCK(&lk);
2349 	/*
2350 	 * If the inode has never been written to disk (delay == 0),
2351 	 * then we can process the freeblks now that we have deleted
2352 	 * the dependencies.
2353 	 */
2354 	if (!delay)
2355 		handle_workitem_freeblocks(freeblks, 0);
2356 }
2357 
2358 /*
2359  * Reclaim any dependency structures from a buffer that is about to
2360  * be reallocated to a new vnode. The buffer must be locked, thus,
2361  * no I/O completion operations can occur while we are manipulating
2362  * its associated dependencies. The mutex is held so that other I/O's
2363  * associated with related dependencies do not occur.
2364  */
2365 static void
2366 deallocate_dependencies(bp, inodedep)
2367 	struct buf *bp;
2368 	struct inodedep *inodedep;
2369 {
2370 	struct worklist *wk;
2371 	struct indirdep *indirdep;
2372 	struct allocindir *aip;
2373 	struct pagedep *pagedep;
2374 	struct dirrem *dirrem;
2375 	struct diradd *dap;
2376 	int i;
2377 
2378 	mtx_assert(&lk, MA_OWNED);
2379 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2380 		switch (wk->wk_type) {
2381 
2382 		case D_INDIRDEP:
2383 			indirdep = WK_INDIRDEP(wk);
2384 			/*
2385 			 * None of the indirect pointers will ever be visible,
2386 			 * so they can simply be tossed. GOINGAWAY ensures
2387 			 * that allocated pointers will be saved in the buffer
2388 			 * cache until they are freed. Note that they will
2389 			 * only be able to be found by their physical address
2390 			 * since the inode mapping the logical address will
2391 			 * be gone. The save buffer used for the safe copy
2392 			 * was allocated in setup_allocindir_phase2 using
2393 			 * the physical address so it could be used for this
2394 			 * purpose. Hence we swap the safe copy with the real
2395 			 * copy, allowing the safe copy to be freed and holding
2396 			 * on to the real copy for later use in indir_trunc.
2397 			 */
2398 			if (indirdep->ir_state & GOINGAWAY)
2399 				panic("deallocate_dependencies: already gone");
2400 			indirdep->ir_state |= GOINGAWAY;
2401 			VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1;
2402 			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
2403 				free_allocindir(aip, inodedep);
2404 			if (bp->b_lblkno >= 0 ||
2405 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno)
2406 				panic("deallocate_dependencies: not indir");
2407 			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2408 			    bp->b_bcount);
2409 			WORKLIST_REMOVE(wk);
2410 			WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
2411 			continue;
2412 
2413 		case D_PAGEDEP:
2414 			pagedep = WK_PAGEDEP(wk);
2415 			/*
2416 			 * None of the directory additions will ever be
2417 			 * visible, so they can simply be tossed.
2418 			 */
2419 			for (i = 0; i < DAHASHSZ; i++)
2420 				while ((dap =
2421 				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
2422 					free_diradd(dap);
2423 			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2424 				free_diradd(dap);
2425 			/*
2426 			 * Copy any directory remove dependencies to the list
2427 			 * to be processed after the zero'ed inode is written.
2428 			 * If the inode has already been written, then they
2429 			 * can be dumped directly onto the work list.
2430 			 */
2431 			LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2432 				LIST_REMOVE(dirrem, dm_next);
2433 				dirrem->dm_dirinum = pagedep->pd_ino;
2434 				if (inodedep == NULL ||
2435 				    (inodedep->id_state & ALLCOMPLETE) ==
2436 				     ALLCOMPLETE)
2437 					add_to_worklist(&dirrem->dm_list);
2438 				else
2439 					WORKLIST_INSERT(&inodedep->id_bufwait,
2440 					    &dirrem->dm_list);
2441 			}
2442 			if ((pagedep->pd_state & NEWBLOCK) != 0) {
2443 				LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list)
2444 					if (wk->wk_type == D_NEWDIRBLK &&
2445 					    WK_NEWDIRBLK(wk)->db_pagedep ==
2446 					      pagedep)
2447 						break;
2448 				if (wk != NULL) {
2449 					WORKLIST_REMOVE(wk);
2450 					free_newdirblk(WK_NEWDIRBLK(wk));
2451 				} else
2452 					panic("deallocate_dependencies: "
2453 					      "lost pagedep");
2454 			}
2455 			WORKLIST_REMOVE(&pagedep->pd_list);
2456 			LIST_REMOVE(pagedep, pd_hash);
2457 			WORKITEM_FREE(pagedep, D_PAGEDEP);
2458 			continue;
2459 
2460 		case D_ALLOCINDIR:
2461 			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2462 			continue;
2463 
2464 		case D_ALLOCDIRECT:
2465 		case D_INODEDEP:
2466 			panic("deallocate_dependencies: Unexpected type %s",
2467 			    TYPENAME(wk->wk_type));
2468 			/* NOTREACHED */
2469 
2470 		default:
2471 			panic("deallocate_dependencies: Unknown type %s",
2472 			    TYPENAME(wk->wk_type));
2473 			/* NOTREACHED */
2474 		}
2475 	}
2476 }
2477 
2478 /*
2479  * Free an allocdirect. Generate a new freefrag work request if appropriate.
2480  * This routine must be called with splbio interrupts blocked.
2481  */
2482 static void
2483 free_allocdirect(adphead, adp, delay)
2484 	struct allocdirectlst *adphead;
2485 	struct allocdirect *adp;
2486 	int delay;
2487 {
2488 	struct newdirblk *newdirblk;
2489 	struct worklist *wk;
2490 
2491 	mtx_assert(&lk, MA_OWNED);
2492 	if ((adp->ad_state & DEPCOMPLETE) == 0)
2493 		LIST_REMOVE(adp, ad_deps);
2494 	TAILQ_REMOVE(adphead, adp, ad_next);
2495 	if ((adp->ad_state & COMPLETE) == 0)
2496 		WORKLIST_REMOVE(&adp->ad_list);
2497 	if (adp->ad_freefrag != NULL) {
2498 		if (delay)
2499 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2500 			    &adp->ad_freefrag->ff_list);
2501 		else
2502 			add_to_worklist(&adp->ad_freefrag->ff_list);
2503 	}
2504 	if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) {
2505 		newdirblk = WK_NEWDIRBLK(wk);
2506 		WORKLIST_REMOVE(&newdirblk->db_list);
2507 		if (!LIST_EMPTY(&adp->ad_newdirblk))
2508 			panic("free_allocdirect: extra newdirblk");
2509 		if (delay)
2510 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2511 			    &newdirblk->db_list);
2512 		else
2513 			free_newdirblk(newdirblk);
2514 	}
2515 	WORKITEM_FREE(adp, D_ALLOCDIRECT);
2516 }
2517 
2518 /*
2519  * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
2520  * This routine must be called with splbio interrupts blocked.
2521  */
2522 static void
2523 free_newdirblk(newdirblk)
2524 	struct newdirblk *newdirblk;
2525 {
2526 	struct pagedep *pagedep;
2527 	struct diradd *dap;
2528 	int i;
2529 
2530 	mtx_assert(&lk, MA_OWNED);
2531 	/*
2532 	 * If the pagedep is still linked onto the directory buffer
2533 	 * dependency chain, then some of the entries on the
2534 	 * pd_pendinghd list may not be committed to disk yet. In
2535 	 * this case, we will simply clear the NEWBLOCK flag and
2536 	 * let the pd_pendinghd list be processed when the pagedep
2537 	 * is next written. If the pagedep is no longer on the buffer
2538 	 * dependency chain, then all the entries on the pd_pending
2539 	 * list are committed to disk and we can free them here.
2540 	 */
2541 	pagedep = newdirblk->db_pagedep;
2542 	pagedep->pd_state &= ~NEWBLOCK;
2543 	if ((pagedep->pd_state & ONWORKLIST) == 0)
2544 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
2545 			free_diradd(dap);
2546 	/*
2547 	 * If no dependencies remain, the pagedep will be freed.
2548 	 */
2549 	for (i = 0; i < DAHASHSZ; i++)
2550 		if (!LIST_EMPTY(&pagedep->pd_diraddhd[i]))
2551 			break;
2552 	if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) {
2553 		LIST_REMOVE(pagedep, pd_hash);
2554 		WORKITEM_FREE(pagedep, D_PAGEDEP);
2555 	}
2556 	WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2557 }
2558 
2559 /*
2560  * Prepare an inode to be freed. The actual free operation is not
2561  * done until the zero'ed inode has been written to disk.
2562  */
2563 void
2564 softdep_freefile(pvp, ino, mode)
2565 	struct vnode *pvp;
2566 	ino_t ino;
2567 	int mode;
2568 {
2569 	struct inode *ip = VTOI(pvp);
2570 	struct inodedep *inodedep;
2571 	struct freefile *freefile;
2572 
2573 	/*
2574 	 * This sets up the inode de-allocation dependency.
2575 	 */
2576 	MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2577 		M_FREEFILE, M_SOFTDEP_FLAGS);
2578 	workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount);
2579 	freefile->fx_mode = mode;
2580 	freefile->fx_oldinum = ino;
2581 	freefile->fx_devvp = ip->i_devvp;
2582 	if ((ip->i_flag & IN_SPACECOUNTED) == 0) {
2583 		UFS_LOCK(ip->i_ump);
2584 		ip->i_fs->fs_pendinginodes += 1;
2585 		UFS_UNLOCK(ip->i_ump);
2586 	}
2587 
2588 	/*
2589 	 * If the inodedep does not exist, then the zero'ed inode has
2590 	 * been written to disk. If the allocated inode has never been
2591 	 * written to disk, then the on-disk inode is zero'ed. In either
2592 	 * case we can free the file immediately.
2593 	 */
2594 	ACQUIRE_LOCK(&lk);
2595 	if (inodedep_lookup(pvp->v_mount, ino, 0, &inodedep) == 0 ||
2596 	    check_inode_unwritten(inodedep)) {
2597 		FREE_LOCK(&lk);
2598 		handle_workitem_freefile(freefile);
2599 		return;
2600 	}
2601 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2602 	FREE_LOCK(&lk);
2603 	ip->i_flag |= IN_MODIFIED;
2604 }
2605 
2606 /*
2607  * Check to see if an inode has never been written to disk. If
2608  * so free the inodedep and return success, otherwise return failure.
2609  * This routine must be called with splbio interrupts blocked.
2610  *
2611  * If we still have a bitmap dependency, then the inode has never
2612  * been written to disk. Drop the dependency as it is no longer
2613  * necessary since the inode is being deallocated. We set the
2614  * ALLCOMPLETE flags since the bitmap now properly shows that the
2615  * inode is not allocated. Even if the inode is actively being
2616  * written, it has been rolled back to its zero'ed state, so we
2617  * are ensured that a zero inode is what is on the disk. For short
2618  * lived files, this change will usually result in removing all the
2619  * dependencies from the inode so that it can be freed immediately.
2620  */
2621 static int
2622 check_inode_unwritten(inodedep)
2623 	struct inodedep *inodedep;
2624 {
2625 
2626 	mtx_assert(&lk, MA_OWNED);
2627 	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2628 	    !LIST_EMPTY(&inodedep->id_pendinghd) ||
2629 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
2630 	    !LIST_EMPTY(&inodedep->id_inowait) ||
2631 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
2632 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
2633 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
2634 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
2635 	    inodedep->id_nlinkdelta != 0)
2636 		return (0);
2637 
2638 	/*
2639 	 * Another process might be in initiate_write_inodeblock_ufs[12]
2640 	 * trying to allocate memory without holding "Softdep Lock".
2641 	 */
2642 	if ((inodedep->id_state & IOSTARTED) != 0 &&
2643 	    inodedep->id_savedino1 == NULL)
2644 		return (0);
2645 
2646 	inodedep->id_state |= ALLCOMPLETE;
2647 	LIST_REMOVE(inodedep, id_deps);
2648 	inodedep->id_buf = NULL;
2649 	if (inodedep->id_state & ONWORKLIST)
2650 		WORKLIST_REMOVE(&inodedep->id_list);
2651 	if (inodedep->id_savedino1 != NULL) {
2652 		FREE(inodedep->id_savedino1, M_SAVEDINO);
2653 		inodedep->id_savedino1 = NULL;
2654 	}
2655 	if (free_inodedep(inodedep) == 0)
2656 		panic("check_inode_unwritten: busy inode");
2657 	return (1);
2658 }
2659 
2660 /*
2661  * Try to free an inodedep structure. Return 1 if it could be freed.
2662  */
2663 static int
2664 free_inodedep(inodedep)
2665 	struct inodedep *inodedep;
2666 {
2667 
2668 	mtx_assert(&lk, MA_OWNED);
2669 	if ((inodedep->id_state & ONWORKLIST) != 0 ||
2670 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2671 	    !LIST_EMPTY(&inodedep->id_pendinghd) ||
2672 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
2673 	    !LIST_EMPTY(&inodedep->id_inowait) ||
2674 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
2675 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
2676 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
2677 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
2678 	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL)
2679 		return (0);
2680 	LIST_REMOVE(inodedep, id_hash);
2681 	WORKITEM_FREE(inodedep, D_INODEDEP);
2682 	num_inodedep -= 1;
2683 	return (1);
2684 }
2685 
2686 /*
2687  * This workitem routine performs the block de-allocation.
2688  * The workitem is added to the pending list after the updated
2689  * inode block has been written to disk.  As mentioned above,
2690  * checks regarding the number of blocks de-allocated (compared
2691  * to the number of blocks allocated for the file) are also
2692  * performed in this function.
2693  */
2694 static void
2695 handle_workitem_freeblocks(freeblks, flags)
2696 	struct freeblks *freeblks;
2697 	int flags;
2698 {
2699 	struct inode *ip;
2700 	struct vnode *vp;
2701 	struct fs *fs;
2702 	struct ufsmount *ump;
2703 	int i, nblocks, level, bsize;
2704 	ufs2_daddr_t bn, blocksreleased = 0;
2705 	int error, allerror = 0;
2706 	ufs_lbn_t baselbns[NIADDR], tmpval;
2707 	int fs_pendingblocks;
2708 
2709 	ump = VFSTOUFS(freeblks->fb_list.wk_mp);
2710 	fs = ump->um_fs;
2711 	fs_pendingblocks = 0;
2712 	tmpval = 1;
2713 	baselbns[0] = NDADDR;
2714 	for (i = 1; i < NIADDR; i++) {
2715 		tmpval *= NINDIR(fs);
2716 		baselbns[i] = baselbns[i - 1] + tmpval;
2717 	}
2718 	nblocks = btodb(fs->fs_bsize);
2719 	blocksreleased = 0;
2720 	/*
2721 	 * Release all extended attribute blocks or frags.
2722 	 */
2723 	if (freeblks->fb_oldextsize > 0) {
2724 		for (i = (NXADDR - 1); i >= 0; i--) {
2725 			if ((bn = freeblks->fb_eblks[i]) == 0)
2726 				continue;
2727 			bsize = sblksize(fs, freeblks->fb_oldextsize, i);
2728 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize,
2729 			    freeblks->fb_previousinum);
2730 			blocksreleased += btodb(bsize);
2731 		}
2732 	}
2733 	/*
2734 	 * Release all data blocks or frags.
2735 	 */
2736 	if (freeblks->fb_oldsize > 0) {
2737 		/*
2738 		 * Indirect blocks first.
2739 		 */
2740 		for (level = (NIADDR - 1); level >= 0; level--) {
2741 			if ((bn = freeblks->fb_iblks[level]) == 0)
2742 				continue;
2743 			if ((error = indir_trunc(freeblks, fsbtodb(fs, bn),
2744 			    level, baselbns[level], &blocksreleased)) != 0)
2745 				allerror = error;
2746 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn,
2747 			    fs->fs_bsize, freeblks->fb_previousinum);
2748 			fs_pendingblocks += nblocks;
2749 			blocksreleased += nblocks;
2750 		}
2751 		/*
2752 		 * All direct blocks or frags.
2753 		 */
2754 		for (i = (NDADDR - 1); i >= 0; i--) {
2755 			if ((bn = freeblks->fb_dblks[i]) == 0)
2756 				continue;
2757 			bsize = sblksize(fs, freeblks->fb_oldsize, i);
2758 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize,
2759 			    freeblks->fb_previousinum);
2760 			fs_pendingblocks += btodb(bsize);
2761 			blocksreleased += btodb(bsize);
2762 		}
2763 	}
2764 	UFS_LOCK(ump);
2765 	fs->fs_pendingblocks -= fs_pendingblocks;
2766 	UFS_UNLOCK(ump);
2767 	/*
2768 	 * If we still have not finished background cleanup, then check
2769 	 * to see if the block count needs to be adjusted.
2770 	 */
2771 	if (freeblks->fb_chkcnt != blocksreleased &&
2772 	    (fs->fs_flags & FS_UNCLEAN) != 0 &&
2773 	    ffs_vget(freeblks->fb_list.wk_mp, freeblks->fb_previousinum,
2774 	    (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) {
2775 		ip = VTOI(vp);
2776 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + \
2777 		    freeblks->fb_chkcnt - blocksreleased);
2778 		ip->i_flag |= IN_CHANGE;
2779 		vput(vp);
2780 	}
2781 
2782 #ifdef INVARIANTS
2783 	if (freeblks->fb_chkcnt != blocksreleased &&
2784 	    ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0))
2785 		printf("handle_workitem_freeblocks: block count\n");
2786 	if (allerror)
2787 		softdep_error("handle_workitem_freeblks", allerror);
2788 #endif /* INVARIANTS */
2789 
2790 	ACQUIRE_LOCK(&lk);
2791 	WORKITEM_FREE(freeblks, D_FREEBLKS);
2792 	FREE_LOCK(&lk);
2793 }
2794 
2795 /*
2796  * Release blocks associated with the inode ip and stored in the indirect
2797  * block dbn. If level is greater than SINGLE, the block is an indirect block
2798  * and recursive calls to indirtrunc must be used to cleanse other indirect
2799  * blocks.
2800  */
2801 static int
2802 indir_trunc(freeblks, dbn, level, lbn, countp)
2803 	struct freeblks *freeblks;
2804 	ufs2_daddr_t dbn;
2805 	int level;
2806 	ufs_lbn_t lbn;
2807 	ufs2_daddr_t *countp;
2808 {
2809 	struct buf *bp;
2810 	struct fs *fs;
2811 	struct worklist *wk;
2812 	struct indirdep *indirdep;
2813 	struct ufsmount *ump;
2814 	ufs1_daddr_t *bap1 = 0;
2815 	ufs2_daddr_t nb, *bap2 = 0;
2816 	ufs_lbn_t lbnadd;
2817 	int i, nblocks, ufs1fmt;
2818 	int error, allerror = 0;
2819 	int fs_pendingblocks;
2820 
2821 	ump = VFSTOUFS(freeblks->fb_list.wk_mp);
2822 	fs = ump->um_fs;
2823 	fs_pendingblocks = 0;
2824 	lbnadd = 1;
2825 	for (i = level; i > 0; i--)
2826 		lbnadd *= NINDIR(fs);
2827 	/*
2828 	 * Get buffer of block pointers to be freed. This routine is not
2829 	 * called until the zero'ed inode has been written, so it is safe
2830 	 * to free blocks as they are encountered. Because the inode has
2831 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
2832 	 * have to use the on-disk address and the block device for the
2833 	 * filesystem to look them up. If the file was deleted before its
2834 	 * indirect blocks were all written to disk, the routine that set
2835 	 * us up (deallocate_dependencies) will have arranged to leave
2836 	 * a complete copy of the indirect block in memory for our use.
2837 	 * Otherwise we have to read the blocks in from the disk.
2838 	 */
2839 #ifdef notyet
2840 	bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0,
2841 	    GB_NOCREAT);
2842 #else
2843 	bp = incore(&freeblks->fb_devvp->v_bufobj, dbn);
2844 #endif
2845 	ACQUIRE_LOCK(&lk);
2846 	if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2847 		if (wk->wk_type != D_INDIRDEP ||
2848 		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2849 		    (indirdep->ir_state & GOINGAWAY) == 0)
2850 			panic("indir_trunc: lost indirdep");
2851 		WORKLIST_REMOVE(wk);
2852 		WORKITEM_FREE(indirdep, D_INDIRDEP);
2853 		if (!LIST_EMPTY(&bp->b_dep))
2854 			panic("indir_trunc: dangling dep");
2855 		ump->um_numindirdeps -= 1;
2856 		FREE_LOCK(&lk);
2857 	} else {
2858 #ifdef notyet
2859 		if (bp)
2860 			brelse(bp);
2861 #endif
2862 		FREE_LOCK(&lk);
2863 		error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
2864 		    NOCRED, &bp);
2865 		if (error) {
2866 			brelse(bp);
2867 			return (error);
2868 		}
2869 	}
2870 	/*
2871 	 * Recursively free indirect blocks.
2872 	 */
2873 	if (ump->um_fstype == UFS1) {
2874 		ufs1fmt = 1;
2875 		bap1 = (ufs1_daddr_t *)bp->b_data;
2876 	} else {
2877 		ufs1fmt = 0;
2878 		bap2 = (ufs2_daddr_t *)bp->b_data;
2879 	}
2880 	nblocks = btodb(fs->fs_bsize);
2881 	for (i = NINDIR(fs) - 1; i >= 0; i--) {
2882 		if (ufs1fmt)
2883 			nb = bap1[i];
2884 		else
2885 			nb = bap2[i];
2886 		if (nb == 0)
2887 			continue;
2888 		if (level != 0) {
2889 			if ((error = indir_trunc(freeblks, fsbtodb(fs, nb),
2890 			     level - 1, lbn + (i * lbnadd), countp)) != 0)
2891 				allerror = error;
2892 		}
2893 		ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, fs->fs_bsize,
2894 		    freeblks->fb_previousinum);
2895 		fs_pendingblocks += nblocks;
2896 		*countp += nblocks;
2897 	}
2898 	UFS_LOCK(ump);
2899 	fs->fs_pendingblocks -= fs_pendingblocks;
2900 	UFS_UNLOCK(ump);
2901 	bp->b_flags |= B_INVAL | B_NOCACHE;
2902 	brelse(bp);
2903 	return (allerror);
2904 }
2905 
2906 /*
2907  * Free an allocindir.
2908  * This routine must be called with splbio interrupts blocked.
2909  */
2910 static void
2911 free_allocindir(aip, inodedep)
2912 	struct allocindir *aip;
2913 	struct inodedep *inodedep;
2914 {
2915 	struct freefrag *freefrag;
2916 
2917 	mtx_assert(&lk, MA_OWNED);
2918 	if ((aip->ai_state & DEPCOMPLETE) == 0)
2919 		LIST_REMOVE(aip, ai_deps);
2920 	if (aip->ai_state & ONWORKLIST)
2921 		WORKLIST_REMOVE(&aip->ai_list);
2922 	LIST_REMOVE(aip, ai_next);
2923 	if ((freefrag = aip->ai_freefrag) != NULL) {
2924 		if (inodedep == NULL)
2925 			add_to_worklist(&freefrag->ff_list);
2926 		else
2927 			WORKLIST_INSERT(&inodedep->id_bufwait,
2928 			    &freefrag->ff_list);
2929 	}
2930 	WORKITEM_FREE(aip, D_ALLOCINDIR);
2931 }
2932 
2933 /*
2934  * Directory entry addition dependencies.
2935  *
2936  * When adding a new directory entry, the inode (with its incremented link
2937  * count) must be written to disk before the directory entry's pointer to it.
2938  * Also, if the inode is newly allocated, the corresponding freemap must be
2939  * updated (on disk) before the directory entry's pointer. These requirements
2940  * are met via undo/redo on the directory entry's pointer, which consists
2941  * simply of the inode number.
2942  *
2943  * As directory entries are added and deleted, the free space within a
2944  * directory block can become fragmented.  The ufs filesystem will compact
2945  * a fragmented directory block to make space for a new entry. When this
2946  * occurs, the offsets of previously added entries change. Any "diradd"
2947  * dependency structures corresponding to these entries must be updated with
2948  * the new offsets.
2949  */
2950 
2951 /*
2952  * This routine is called after the in-memory inode's link
2953  * count has been incremented, but before the directory entry's
2954  * pointer to the inode has been set.
2955  */
2956 int
2957 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
2958 	struct buf *bp;		/* buffer containing directory block */
2959 	struct inode *dp;	/* inode for directory */
2960 	off_t diroffset;	/* offset of new entry in directory */
2961 	ino_t newinum;		/* inode referenced by new directory entry */
2962 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
2963 	int isnewblk;		/* entry is in a newly allocated block */
2964 {
2965 	int offset;		/* offset of new entry within directory block */
2966 	ufs_lbn_t lbn;		/* block in directory containing new entry */
2967 	struct fs *fs;
2968 	struct diradd *dap;
2969 	struct allocdirect *adp;
2970 	struct pagedep *pagedep;
2971 	struct inodedep *inodedep;
2972 	struct newdirblk *newdirblk = 0;
2973 	struct mkdir *mkdir1, *mkdir2;
2974 	struct mount *mp;
2975 
2976 	/*
2977 	 * Whiteouts have no dependencies.
2978 	 */
2979 	if (newinum == WINO) {
2980 		if (newdirbp != NULL)
2981 			bdwrite(newdirbp);
2982 		return (0);
2983 	}
2984 	mp = UFSTOVFS(dp->i_ump);
2985 	fs = dp->i_fs;
2986 	lbn = lblkno(fs, diroffset);
2987 	offset = blkoff(fs, diroffset);
2988 	MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2989 		M_SOFTDEP_FLAGS|M_ZERO);
2990 	workitem_alloc(&dap->da_list, D_DIRADD, mp);
2991 	dap->da_offset = offset;
2992 	dap->da_newinum = newinum;
2993 	dap->da_state = ATTACHED;
2994 	if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) {
2995 		MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk),
2996 		    M_NEWDIRBLK, M_SOFTDEP_FLAGS);
2997 		workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
2998 	}
2999 	if (newdirbp == NULL) {
3000 		dap->da_state |= DEPCOMPLETE;
3001 		ACQUIRE_LOCK(&lk);
3002 	} else {
3003 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
3004 		MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
3005 		    M_SOFTDEP_FLAGS);
3006 		workitem_alloc(&mkdir1->md_list, D_MKDIR, mp);
3007 		mkdir1->md_state = MKDIR_BODY;
3008 		mkdir1->md_diradd = dap;
3009 		MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
3010 		    M_SOFTDEP_FLAGS);
3011 		workitem_alloc(&mkdir2->md_list, D_MKDIR, mp);
3012 		mkdir2->md_state = MKDIR_PARENT;
3013 		mkdir2->md_diradd = dap;
3014 		/*
3015 		 * Dependency on "." and ".." being written to disk.
3016 		 */
3017 		mkdir1->md_buf = newdirbp;
3018 		ACQUIRE_LOCK(&lk);
3019 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
3020 		WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
3021 		FREE_LOCK(&lk);
3022 		bdwrite(newdirbp);
3023 		/*
3024 		 * Dependency on link count increase for parent directory
3025 		 */
3026 		ACQUIRE_LOCK(&lk);
3027 		if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0
3028 		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3029 			dap->da_state &= ~MKDIR_PARENT;
3030 			WORKITEM_FREE(mkdir2, D_MKDIR);
3031 		} else {
3032 			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
3033 			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
3034 		}
3035 	}
3036 	/*
3037 	 * Link into parent directory pagedep to await its being written.
3038 	 */
3039 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3040 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3041 	dap->da_pagedep = pagedep;
3042 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
3043 	    da_pdlist);
3044 	/*
3045 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3046 	 * is not yet written. If it is written, do the post-inode write
3047 	 * processing to put it on the id_pendinghd list.
3048 	 */
3049 	(void) inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
3050 	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
3051 		diradd_inode_written(dap, inodedep);
3052 	else
3053 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3054 	if (isnewblk) {
3055 		/*
3056 		 * Directories growing into indirect blocks are rare
3057 		 * enough and the frequency of new block allocation
3058 		 * in those cases even more rare, that we choose not
3059 		 * to bother tracking them. Rather we simply force the
3060 		 * new directory entry to disk.
3061 		 */
3062 		if (lbn >= NDADDR) {
3063 			FREE_LOCK(&lk);
3064 			/*
3065 			 * We only have a new allocation when at the
3066 			 * beginning of a new block, not when we are
3067 			 * expanding into an existing block.
3068 			 */
3069 			if (blkoff(fs, diroffset) == 0)
3070 				return (1);
3071 			return (0);
3072 		}
3073 		/*
3074 		 * We only have a new allocation when at the beginning
3075 		 * of a new fragment, not when we are expanding into an
3076 		 * existing fragment. Also, there is nothing to do if we
3077 		 * are already tracking this block.
3078 		 */
3079 		if (fragoff(fs, diroffset) != 0) {
3080 			FREE_LOCK(&lk);
3081 			return (0);
3082 		}
3083 		if ((pagedep->pd_state & NEWBLOCK) != 0) {
3084 			WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
3085 			FREE_LOCK(&lk);
3086 			return (0);
3087 		}
3088 		/*
3089 		 * Find our associated allocdirect and have it track us.
3090 		 */
3091 		if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0)
3092 			panic("softdep_setup_directory_add: lost inodedep");
3093 		adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst);
3094 		if (adp == NULL || adp->ad_lbn != lbn)
3095 			panic("softdep_setup_directory_add: lost entry");
3096 		pagedep->pd_state |= NEWBLOCK;
3097 		newdirblk->db_pagedep = pagedep;
3098 		WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list);
3099 	}
3100 	FREE_LOCK(&lk);
3101 	return (0);
3102 }
3103 
3104 /*
3105  * This procedure is called to change the offset of a directory
3106  * entry when compacting a directory block which must be owned
3107  * exclusively by the caller. Note that the actual entry movement
3108  * must be done in this procedure to ensure that no I/O completions
3109  * occur while the move is in progress.
3110  */
3111 void
3112 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
3113 	struct inode *dp;	/* inode for directory */
3114 	caddr_t base;		/* address of dp->i_offset */
3115 	caddr_t oldloc;		/* address of old directory location */
3116 	caddr_t newloc;		/* address of new directory location */
3117 	int entrysize;		/* size of directory entry */
3118 {
3119 	int offset, oldoffset, newoffset;
3120 	struct pagedep *pagedep;
3121 	struct diradd *dap;
3122 	ufs_lbn_t lbn;
3123 
3124 	ACQUIRE_LOCK(&lk);
3125 	lbn = lblkno(dp->i_fs, dp->i_offset);
3126 	offset = blkoff(dp->i_fs, dp->i_offset);
3127 	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
3128 		goto done;
3129 	oldoffset = offset + (oldloc - base);
3130 	newoffset = offset + (newloc - base);
3131 
3132 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
3133 		if (dap->da_offset != oldoffset)
3134 			continue;
3135 		dap->da_offset = newoffset;
3136 		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
3137 			break;
3138 		LIST_REMOVE(dap, da_pdlist);
3139 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
3140 		    dap, da_pdlist);
3141 		break;
3142 	}
3143 	if (dap == NULL) {
3144 
3145 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
3146 			if (dap->da_offset == oldoffset) {
3147 				dap->da_offset = newoffset;
3148 				break;
3149 			}
3150 		}
3151 	}
3152 done:
3153 	bcopy(oldloc, newloc, entrysize);
3154 	FREE_LOCK(&lk);
3155 }
3156 
3157 /*
3158  * Free a diradd dependency structure. This routine must be called
3159  * with splbio interrupts blocked.
3160  */
3161 static void
3162 free_diradd(dap)
3163 	struct diradd *dap;
3164 {
3165 	struct dirrem *dirrem;
3166 	struct pagedep *pagedep;
3167 	struct inodedep *inodedep;
3168 	struct mkdir *mkdir, *nextmd;
3169 
3170 	mtx_assert(&lk, MA_OWNED);
3171 	WORKLIST_REMOVE(&dap->da_list);
3172 	LIST_REMOVE(dap, da_pdlist);
3173 	if ((dap->da_state & DIRCHG) == 0) {
3174 		pagedep = dap->da_pagedep;
3175 	} else {
3176 		dirrem = dap->da_previous;
3177 		pagedep = dirrem->dm_pagedep;
3178 		dirrem->dm_dirinum = pagedep->pd_ino;
3179 		add_to_worklist(&dirrem->dm_list);
3180 	}
3181 	if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum,
3182 	    0, &inodedep) != 0)
3183 		(void) free_inodedep(inodedep);
3184 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
3185 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
3186 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
3187 			if (mkdir->md_diradd != dap)
3188 				continue;
3189 			dap->da_state &= ~mkdir->md_state;
3190 			WORKLIST_REMOVE(&mkdir->md_list);
3191 			LIST_REMOVE(mkdir, md_mkdirs);
3192 			WORKITEM_FREE(mkdir, D_MKDIR);
3193 		}
3194 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
3195 			panic("free_diradd: unfound ref");
3196 	}
3197 	WORKITEM_FREE(dap, D_DIRADD);
3198 }
3199 
3200 /*
3201  * Directory entry removal dependencies.
3202  *
3203  * When removing a directory entry, the entry's inode pointer must be
3204  * zero'ed on disk before the corresponding inode's link count is decremented
3205  * (possibly freeing the inode for re-use). This dependency is handled by
3206  * updating the directory entry but delaying the inode count reduction until
3207  * after the directory block has been written to disk. After this point, the
3208  * inode count can be decremented whenever it is convenient.
3209  */
3210 
3211 /*
3212  * This routine should be called immediately after removing
3213  * a directory entry.  The inode's link count should not be
3214  * decremented by the calling procedure -- the soft updates
3215  * code will do this task when it is safe.
3216  */
3217 void
3218 softdep_setup_remove(bp, dp, ip, isrmdir)
3219 	struct buf *bp;		/* buffer containing directory block */
3220 	struct inode *dp;	/* inode for the directory being modified */
3221 	struct inode *ip;	/* inode for directory entry being removed */
3222 	int isrmdir;		/* indicates if doing RMDIR */
3223 {
3224 	struct dirrem *dirrem, *prevdirrem;
3225 
3226 	/*
3227 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
3228 	 */
3229 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3230 
3231 	/*
3232 	 * If the COMPLETE flag is clear, then there were no active
3233 	 * entries and we want to roll back to a zeroed entry until
3234 	 * the new inode is committed to disk. If the COMPLETE flag is
3235 	 * set then we have deleted an entry that never made it to
3236 	 * disk. If the entry we deleted resulted from a name change,
3237 	 * then the old name still resides on disk. We cannot delete
3238 	 * its inode (returned to us in prevdirrem) until the zeroed
3239 	 * directory entry gets to disk. The new inode has never been
3240 	 * referenced on the disk, so can be deleted immediately.
3241 	 */
3242 	if ((dirrem->dm_state & COMPLETE) == 0) {
3243 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
3244 		    dm_next);
3245 		FREE_LOCK(&lk);
3246 	} else {
3247 		if (prevdirrem != NULL)
3248 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
3249 			    prevdirrem, dm_next);
3250 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
3251 		FREE_LOCK(&lk);
3252 		handle_workitem_remove(dirrem, NULL);
3253 	}
3254 }
3255 
3256 /*
3257  * Allocate a new dirrem if appropriate and return it along with
3258  * its associated pagedep. Called without a lock, returns with lock.
3259  */
3260 static long num_dirrem;		/* number of dirrem allocated */
3261 static struct dirrem *
3262 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
3263 	struct buf *bp;		/* buffer containing directory block */
3264 	struct inode *dp;	/* inode for the directory being modified */
3265 	struct inode *ip;	/* inode for directory entry being removed */
3266 	int isrmdir;		/* indicates if doing RMDIR */
3267 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
3268 {
3269 	int offset;
3270 	ufs_lbn_t lbn;
3271 	struct diradd *dap;
3272 	struct dirrem *dirrem;
3273 	struct pagedep *pagedep;
3274 
3275 	/*
3276 	 * Whiteouts have no deletion dependencies.
3277 	 */
3278 	if (ip == NULL)
3279 		panic("newdirrem: whiteout");
3280 	/*
3281 	 * If we are over our limit, try to improve the situation.
3282 	 * Limiting the number of dirrem structures will also limit
3283 	 * the number of freefile and freeblks structures.
3284 	 */
3285 	ACQUIRE_LOCK(&lk);
3286 	if (num_dirrem > max_softdeps / 2)
3287 		(void) request_cleanup(ITOV(dp)->v_mount, FLUSH_REMOVE);
3288 	num_dirrem += 1;
3289 	FREE_LOCK(&lk);
3290 	MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
3291 		M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO);
3292 	workitem_alloc(&dirrem->dm_list, D_DIRREM, ITOV(dp)->v_mount);
3293 	dirrem->dm_state = isrmdir ? RMDIR : 0;
3294 	dirrem->dm_oldinum = ip->i_number;
3295 	*prevdirremp = NULL;
3296 
3297 	ACQUIRE_LOCK(&lk);
3298 	lbn = lblkno(dp->i_fs, dp->i_offset);
3299 	offset = blkoff(dp->i_fs, dp->i_offset);
3300 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3301 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3302 	dirrem->dm_pagedep = pagedep;
3303 	/*
3304 	 * Check for a diradd dependency for the same directory entry.
3305 	 * If present, then both dependencies become obsolete and can
3306 	 * be de-allocated. Check for an entry on both the pd_dirraddhd
3307 	 * list and the pd_pendinghd list.
3308 	 */
3309 
3310 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
3311 		if (dap->da_offset == offset)
3312 			break;
3313 	if (dap == NULL) {
3314 
3315 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
3316 			if (dap->da_offset == offset)
3317 				break;
3318 		if (dap == NULL)
3319 			return (dirrem);
3320 	}
3321 	/*
3322 	 * Must be ATTACHED at this point.
3323 	 */
3324 	if ((dap->da_state & ATTACHED) == 0)
3325 		panic("newdirrem: not ATTACHED");
3326 	if (dap->da_newinum != ip->i_number)
3327 		panic("newdirrem: inum %d should be %d",
3328 		    ip->i_number, dap->da_newinum);
3329 	/*
3330 	 * If we are deleting a changed name that never made it to disk,
3331 	 * then return the dirrem describing the previous inode (which
3332 	 * represents the inode currently referenced from this entry on disk).
3333 	 */
3334 	if ((dap->da_state & DIRCHG) != 0) {
3335 		*prevdirremp = dap->da_previous;
3336 		dap->da_state &= ~DIRCHG;
3337 		dap->da_pagedep = pagedep;
3338 	}
3339 	/*
3340 	 * We are deleting an entry that never made it to disk.
3341 	 * Mark it COMPLETE so we can delete its inode immediately.
3342 	 */
3343 	dirrem->dm_state |= COMPLETE;
3344 	free_diradd(dap);
3345 	return (dirrem);
3346 }
3347 
3348 /*
3349  * Directory entry change dependencies.
3350  *
3351  * Changing an existing directory entry requires that an add operation
3352  * be completed first followed by a deletion. The semantics for the addition
3353  * are identical to the description of adding a new entry above except
3354  * that the rollback is to the old inode number rather than zero. Once
3355  * the addition dependency is completed, the removal is done as described
3356  * in the removal routine above.
3357  */
3358 
3359 /*
3360  * This routine should be called immediately after changing
3361  * a directory entry.  The inode's link count should not be
3362  * decremented by the calling procedure -- the soft updates
3363  * code will perform this task when it is safe.
3364  */
3365 void
3366 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
3367 	struct buf *bp;		/* buffer containing directory block */
3368 	struct inode *dp;	/* inode for the directory being modified */
3369 	struct inode *ip;	/* inode for directory entry being removed */
3370 	ino_t newinum;		/* new inode number for changed entry */
3371 	int isrmdir;		/* indicates if doing RMDIR */
3372 {
3373 	int offset;
3374 	struct diradd *dap = NULL;
3375 	struct dirrem *dirrem, *prevdirrem;
3376 	struct pagedep *pagedep;
3377 	struct inodedep *inodedep;
3378 	struct mount *mp;
3379 
3380 	offset = blkoff(dp->i_fs, dp->i_offset);
3381 	mp = UFSTOVFS(dp->i_ump);
3382 
3383 	/*
3384 	 * Whiteouts do not need diradd dependencies.
3385 	 */
3386 	if (newinum != WINO) {
3387 		MALLOC(dap, struct diradd *, sizeof(struct diradd),
3388 		    M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
3389 		workitem_alloc(&dap->da_list, D_DIRADD, mp);
3390 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
3391 		dap->da_offset = offset;
3392 		dap->da_newinum = newinum;
3393 	}
3394 
3395 	/*
3396 	 * Allocate a new dirrem and ACQUIRE_LOCK.
3397 	 */
3398 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3399 	pagedep = dirrem->dm_pagedep;
3400 	/*
3401 	 * The possible values for isrmdir:
3402 	 *	0 - non-directory file rename
3403 	 *	1 - directory rename within same directory
3404 	 *   inum - directory rename to new directory of given inode number
3405 	 * When renaming to a new directory, we are both deleting and
3406 	 * creating a new directory entry, so the link count on the new
3407 	 * directory should not change. Thus we do not need the followup
3408 	 * dirrem which is usually done in handle_workitem_remove. We set
3409 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
3410 	 * followup dirrem.
3411 	 */
3412 	if (isrmdir > 1)
3413 		dirrem->dm_state |= DIRCHG;
3414 
3415 	/*
3416 	 * Whiteouts have no additional dependencies,
3417 	 * so just put the dirrem on the correct list.
3418 	 */
3419 	if (newinum == WINO) {
3420 		if ((dirrem->dm_state & COMPLETE) == 0) {
3421 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
3422 			    dm_next);
3423 		} else {
3424 			dirrem->dm_dirinum = pagedep->pd_ino;
3425 			add_to_worklist(&dirrem->dm_list);
3426 		}
3427 		FREE_LOCK(&lk);
3428 		return;
3429 	}
3430 
3431 	/*
3432 	 * If the COMPLETE flag is clear, then there were no active
3433 	 * entries and we want to roll back to the previous inode until
3434 	 * the new inode is committed to disk. If the COMPLETE flag is
3435 	 * set, then we have deleted an entry that never made it to disk.
3436 	 * If the entry we deleted resulted from a name change, then the old
3437 	 * inode reference still resides on disk. Any rollback that we do
3438 	 * needs to be to that old inode (returned to us in prevdirrem). If
3439 	 * the entry we deleted resulted from a create, then there is
3440 	 * no entry on the disk, so we want to roll back to zero rather
3441 	 * than the uncommitted inode. In either of the COMPLETE cases we
3442 	 * want to immediately free the unwritten and unreferenced inode.
3443 	 */
3444 	if ((dirrem->dm_state & COMPLETE) == 0) {
3445 		dap->da_previous = dirrem;
3446 	} else {
3447 		if (prevdirrem != NULL) {
3448 			dap->da_previous = prevdirrem;
3449 		} else {
3450 			dap->da_state &= ~DIRCHG;
3451 			dap->da_pagedep = pagedep;
3452 		}
3453 		dirrem->dm_dirinum = pagedep->pd_ino;
3454 		add_to_worklist(&dirrem->dm_list);
3455 	}
3456 	/*
3457 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3458 	 * is not yet written. If it is written, do the post-inode write
3459 	 * processing to put it on the id_pendinghd list.
3460 	 */
3461 	if (inodedep_lookup(mp, newinum, DEPALLOC, &inodedep) == 0 ||
3462 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3463 		dap->da_state |= COMPLETE;
3464 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3465 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3466 	} else {
3467 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
3468 		    dap, da_pdlist);
3469 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3470 	}
3471 	FREE_LOCK(&lk);
3472 }
3473 
3474 /*
3475  * Called whenever the link count on an inode is changed.
3476  * It creates an inode dependency so that the new reference(s)
3477  * to the inode cannot be committed to disk until the updated
3478  * inode has been written.
3479  */
3480 void
3481 softdep_change_linkcnt(ip)
3482 	struct inode *ip;	/* the inode with the increased link count */
3483 {
3484 	struct inodedep *inodedep;
3485 
3486 	ACQUIRE_LOCK(&lk);
3487 	(void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number,
3488 	    DEPALLOC, &inodedep);
3489 	if (ip->i_nlink < ip->i_effnlink)
3490 		panic("softdep_change_linkcnt: bad delta");
3491 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3492 	FREE_LOCK(&lk);
3493 }
3494 
3495 /*
3496  * Called when the effective link count and the reference count
3497  * on an inode drops to zero. At this point there are no names
3498  * referencing the file in the filesystem and no active file
3499  * references. The space associated with the file will be freed
3500  * as soon as the necessary soft dependencies are cleared.
3501  */
3502 void
3503 softdep_releasefile(ip)
3504 	struct inode *ip;	/* inode with the zero effective link count */
3505 {
3506 	struct inodedep *inodedep;
3507 	struct fs *fs;
3508 	int extblocks;
3509 
3510 	if (ip->i_effnlink > 0)
3511 		panic("softdep_releasefile: file still referenced");
3512 	/*
3513 	 * We may be called several times as the on-disk link count
3514 	 * drops to zero. We only want to account for the space once.
3515 	 */
3516 	if (ip->i_flag & IN_SPACECOUNTED)
3517 		return;
3518 	/*
3519 	 * We have to deactivate a snapshot otherwise copyonwrites may
3520 	 * add blocks and the cleanup may remove blocks after we have
3521 	 * tried to account for them.
3522 	 */
3523 	if ((ip->i_flags & SF_SNAPSHOT) != 0)
3524 		ffs_snapremove(ITOV(ip));
3525 	/*
3526 	 * If we are tracking an nlinkdelta, we have to also remember
3527 	 * whether we accounted for the freed space yet.
3528 	 */
3529 	ACQUIRE_LOCK(&lk);
3530 	if ((inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, &inodedep)))
3531 		inodedep->id_state |= SPACECOUNTED;
3532 	FREE_LOCK(&lk);
3533 	fs = ip->i_fs;
3534 	extblocks = 0;
3535 	if (fs->fs_magic == FS_UFS2_MAGIC)
3536 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
3537 	UFS_LOCK(ip->i_ump);
3538 	ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks;
3539 	ip->i_fs->fs_pendinginodes += 1;
3540 	UFS_UNLOCK(ip->i_ump);
3541 	ip->i_flag |= IN_SPACECOUNTED;
3542 }
3543 
3544 /*
3545  * This workitem decrements the inode's link count.
3546  * If the link count reaches zero, the file is removed.
3547  */
3548 static void
3549 handle_workitem_remove(dirrem, xp)
3550 	struct dirrem *dirrem;
3551 	struct vnode *xp;
3552 {
3553 	struct thread *td = curthread;
3554 	struct inodedep *inodedep;
3555 	struct vnode *vp;
3556 	struct inode *ip;
3557 	ino_t oldinum;
3558 	int error;
3559 
3560 	if ((vp = xp) == NULL &&
3561 	    (error = ffs_vget(dirrem->dm_list.wk_mp,
3562 	    dirrem->dm_oldinum, LK_EXCLUSIVE, &vp)) != 0) {
3563 		softdep_error("handle_workitem_remove: vget", error);
3564 		return;
3565 	}
3566 	ip = VTOI(vp);
3567 	ACQUIRE_LOCK(&lk);
3568 	if ((inodedep_lookup(dirrem->dm_list.wk_mp,
3569 	    dirrem->dm_oldinum, 0, &inodedep)) == 0)
3570 		panic("handle_workitem_remove: lost inodedep");
3571 	/*
3572 	 * Normal file deletion.
3573 	 */
3574 	if ((dirrem->dm_state & RMDIR) == 0) {
3575 		ip->i_nlink--;
3576 		DIP_SET(ip, i_nlink, ip->i_nlink);
3577 		ip->i_flag |= IN_CHANGE;
3578 		if (ip->i_nlink < ip->i_effnlink)
3579 			panic("handle_workitem_remove: bad file delta");
3580 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3581 		num_dirrem -= 1;
3582 		WORKITEM_FREE(dirrem, D_DIRREM);
3583 		FREE_LOCK(&lk);
3584 		vput(vp);
3585 		return;
3586 	}
3587 	/*
3588 	 * Directory deletion. Decrement reference count for both the
3589 	 * just deleted parent directory entry and the reference for ".".
3590 	 * Next truncate the directory to length zero. When the
3591 	 * truncation completes, arrange to have the reference count on
3592 	 * the parent decremented to account for the loss of "..".
3593 	 */
3594 	ip->i_nlink -= 2;
3595 	DIP_SET(ip, i_nlink, ip->i_nlink);
3596 	ip->i_flag |= IN_CHANGE;
3597 	if (ip->i_nlink < ip->i_effnlink)
3598 		panic("handle_workitem_remove: bad dir delta");
3599 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3600 	FREE_LOCK(&lk);
3601 	if ((error = ffs_truncate(vp, (off_t)0, 0, td->td_ucred, td)) != 0)
3602 		softdep_error("handle_workitem_remove: truncate", error);
3603 	ACQUIRE_LOCK(&lk);
3604 	/*
3605 	 * Rename a directory to a new parent. Since, we are both deleting
3606 	 * and creating a new directory entry, the link count on the new
3607 	 * directory should not change. Thus we skip the followup dirrem.
3608 	 */
3609 	if (dirrem->dm_state & DIRCHG) {
3610 		num_dirrem -= 1;
3611 		WORKITEM_FREE(dirrem, D_DIRREM);
3612 		FREE_LOCK(&lk);
3613 		vput(vp);
3614 		return;
3615 	}
3616 	/*
3617 	 * If the inodedep does not exist, then the zero'ed inode has
3618 	 * been written to disk. If the allocated inode has never been
3619 	 * written to disk, then the on-disk inode is zero'ed. In either
3620 	 * case we can remove the file immediately.
3621 	 */
3622 	dirrem->dm_state = 0;
3623 	oldinum = dirrem->dm_oldinum;
3624 	dirrem->dm_oldinum = dirrem->dm_dirinum;
3625 	if (inodedep_lookup(dirrem->dm_list.wk_mp, oldinum,
3626 	    0, &inodedep) == 0 || check_inode_unwritten(inodedep)) {
3627 		if (xp != NULL)
3628 			add_to_worklist(&dirrem->dm_list);
3629 		FREE_LOCK(&lk);
3630 		vput(vp);
3631 		if (xp == NULL)
3632 			handle_workitem_remove(dirrem, NULL);
3633 		return;
3634 	}
3635 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3636 	FREE_LOCK(&lk);
3637 	ip->i_flag |= IN_CHANGE;
3638 	ffs_update(vp, 0);
3639 	vput(vp);
3640 }
3641 
3642 /*
3643  * Inode de-allocation dependencies.
3644  *
3645  * When an inode's link count is reduced to zero, it can be de-allocated. We
3646  * found it convenient to postpone de-allocation until after the inode is
3647  * written to disk with its new link count (zero).  At this point, all of the
3648  * on-disk inode's block pointers are nullified and, with careful dependency
3649  * list ordering, all dependencies related to the inode will be satisfied and
3650  * the corresponding dependency structures de-allocated.  So, if/when the
3651  * inode is reused, there will be no mixing of old dependencies with new
3652  * ones.  This artificial dependency is set up by the block de-allocation
3653  * procedure above (softdep_setup_freeblocks) and completed by the
3654  * following procedure.
3655  */
3656 static void
3657 handle_workitem_freefile(freefile)
3658 	struct freefile *freefile;
3659 {
3660 	struct fs *fs;
3661 	struct inodedep *idp;
3662 	struct ufsmount *ump;
3663 	int error;
3664 
3665 	ump = VFSTOUFS(freefile->fx_list.wk_mp);
3666 	fs = ump->um_fs;
3667 #ifdef DEBUG
3668 	ACQUIRE_LOCK(&lk);
3669 	error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp);
3670 	FREE_LOCK(&lk);
3671 	if (error)
3672 		panic("handle_workitem_freefile: inodedep survived");
3673 #endif
3674 	UFS_LOCK(ump);
3675 	fs->fs_pendinginodes -= 1;
3676 	UFS_UNLOCK(ump);
3677 	if ((error = ffs_freefile(ump, fs, freefile->fx_devvp,
3678 	    freefile->fx_oldinum, freefile->fx_mode)) != 0)
3679 		softdep_error("handle_workitem_freefile", error);
3680 	ACQUIRE_LOCK(&lk);
3681 	WORKITEM_FREE(freefile, D_FREEFILE);
3682 	FREE_LOCK(&lk);
3683 }
3684 
3685 
3686 /*
3687  * Helper function which unlinks marker element from work list and returns
3688  * the next element on the list.
3689  */
3690 static __inline struct worklist *
3691 markernext(struct worklist *marker)
3692 {
3693 	struct worklist *next;
3694 
3695 	next = LIST_NEXT(marker, wk_list);
3696 	LIST_REMOVE(marker, wk_list);
3697 	return next;
3698 }
3699 
3700 /*
3701  * Disk writes.
3702  *
3703  * The dependency structures constructed above are most actively used when file
3704  * system blocks are written to disk.  No constraints are placed on when a
3705  * block can be written, but unsatisfied update dependencies are made safe by
3706  * modifying (or replacing) the source memory for the duration of the disk
3707  * write.  When the disk write completes, the memory block is again brought
3708  * up-to-date.
3709  *
3710  * In-core inode structure reclamation.
3711  *
3712  * Because there are a finite number of "in-core" inode structures, they are
3713  * reused regularly.  By transferring all inode-related dependencies to the
3714  * in-memory inode block and indexing them separately (via "inodedep"s), we
3715  * can allow "in-core" inode structures to be reused at any time and avoid
3716  * any increase in contention.
3717  *
3718  * Called just before entering the device driver to initiate a new disk I/O.
3719  * The buffer must be locked, thus, no I/O completion operations can occur
3720  * while we are manipulating its associated dependencies.
3721  */
3722 static void
3723 softdep_disk_io_initiation(bp)
3724 	struct buf *bp;		/* structure describing disk write to occur */
3725 {
3726 	struct worklist *wk;
3727 	struct worklist marker;
3728 	struct indirdep *indirdep;
3729 	struct inodedep *inodedep;
3730 
3731 	/*
3732 	 * We only care about write operations. There should never
3733 	 * be dependencies for reads.
3734 	 */
3735 	if (bp->b_iocmd != BIO_WRITE)
3736 		panic("softdep_disk_io_initiation: not write");
3737 
3738 	marker.wk_type = D_LAST + 1;	/* Not a normal workitem */
3739 	PHOLD(curproc);			/* Don't swap out kernel stack */
3740 
3741 	ACQUIRE_LOCK(&lk);
3742 	/*
3743 	 * Do any necessary pre-I/O processing.
3744 	 */
3745 	for (wk = LIST_FIRST(&bp->b_dep); wk != NULL;
3746 	     wk = markernext(&marker)) {
3747 		LIST_INSERT_AFTER(wk, &marker, wk_list);
3748 		switch (wk->wk_type) {
3749 
3750 		case D_PAGEDEP:
3751 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
3752 			continue;
3753 
3754 		case D_INODEDEP:
3755 			inodedep = WK_INODEDEP(wk);
3756 			if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
3757 				initiate_write_inodeblock_ufs1(inodedep, bp);
3758 			else
3759 				initiate_write_inodeblock_ufs2(inodedep, bp);
3760 			continue;
3761 
3762 		case D_INDIRDEP:
3763 			indirdep = WK_INDIRDEP(wk);
3764 			if (indirdep->ir_state & GOINGAWAY)
3765 				panic("disk_io_initiation: indirdep gone");
3766 			/*
3767 			 * If there are no remaining dependencies, this
3768 			 * will be writing the real pointers, so the
3769 			 * dependency can be freed.
3770 			 */
3771 			if (LIST_EMPTY(&indirdep->ir_deplisthd)) {
3772 				struct buf *bp;
3773 
3774 				bp = indirdep->ir_savebp;
3775 				bp->b_flags |= B_INVAL | B_NOCACHE;
3776 				/* inline expand WORKLIST_REMOVE(wk); */
3777 				wk->wk_state &= ~ONWORKLIST;
3778 				LIST_REMOVE(wk, wk_list);
3779 				WORKITEM_FREE(indirdep, D_INDIRDEP);
3780 				FREE_LOCK(&lk);
3781 				brelse(bp);
3782 				ACQUIRE_LOCK(&lk);
3783 				continue;
3784 			}
3785 			/*
3786 			 * Replace up-to-date version with safe version.
3787 			 */
3788 			FREE_LOCK(&lk);
3789 			MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3790 			    M_INDIRDEP, M_SOFTDEP_FLAGS);
3791 			ACQUIRE_LOCK(&lk);
3792 			indirdep->ir_state &= ~ATTACHED;
3793 			indirdep->ir_state |= UNDONE;
3794 			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3795 			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3796 			    bp->b_bcount);
3797 			continue;
3798 
3799 		case D_MKDIR:
3800 		case D_BMSAFEMAP:
3801 		case D_ALLOCDIRECT:
3802 		case D_ALLOCINDIR:
3803 			continue;
3804 
3805 		default:
3806 			panic("handle_disk_io_initiation: Unexpected type %s",
3807 			    TYPENAME(wk->wk_type));
3808 			/* NOTREACHED */
3809 		}
3810 	}
3811 	FREE_LOCK(&lk);
3812 	PRELE(curproc);			/* Allow swapout of kernel stack */
3813 }
3814 
3815 /*
3816  * Called from within the procedure above to deal with unsatisfied
3817  * allocation dependencies in a directory. The buffer must be locked,
3818  * thus, no I/O completion operations can occur while we are
3819  * manipulating its associated dependencies.
3820  */
3821 static void
3822 initiate_write_filepage(pagedep, bp)
3823 	struct pagedep *pagedep;
3824 	struct buf *bp;
3825 {
3826 	struct diradd *dap;
3827 	struct direct *ep;
3828 	int i;
3829 
3830 	if (pagedep->pd_state & IOSTARTED) {
3831 		/*
3832 		 * This can only happen if there is a driver that does not
3833 		 * understand chaining. Here biodone will reissue the call
3834 		 * to strategy for the incomplete buffers.
3835 		 */
3836 		printf("initiate_write_filepage: already started\n");
3837 		return;
3838 	}
3839 	pagedep->pd_state |= IOSTARTED;
3840 	for (i = 0; i < DAHASHSZ; i++) {
3841 		LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3842 			ep = (struct direct *)
3843 			    ((char *)bp->b_data + dap->da_offset);
3844 			if (ep->d_ino != dap->da_newinum)
3845 				panic("%s: dir inum %d != new %d",
3846 				    "initiate_write_filepage",
3847 				    ep->d_ino, dap->da_newinum);
3848 			if (dap->da_state & DIRCHG)
3849 				ep->d_ino = dap->da_previous->dm_oldinum;
3850 			else
3851 				ep->d_ino = 0;
3852 			dap->da_state &= ~ATTACHED;
3853 			dap->da_state |= UNDONE;
3854 		}
3855 	}
3856 }
3857 
3858 /*
3859  * Version of initiate_write_inodeblock that handles UFS1 dinodes.
3860  * Note that any bug fixes made to this routine must be done in the
3861  * version found below.
3862  *
3863  * Called from within the procedure above to deal with unsatisfied
3864  * allocation dependencies in an inodeblock. The buffer must be
3865  * locked, thus, no I/O completion operations can occur while we
3866  * are manipulating its associated dependencies.
3867  */
3868 static void
3869 initiate_write_inodeblock_ufs1(inodedep, bp)
3870 	struct inodedep *inodedep;
3871 	struct buf *bp;			/* The inode block */
3872 {
3873 	struct allocdirect *adp, *lastadp;
3874 	struct ufs1_dinode *dp;
3875 	struct ufs1_dinode *sip;
3876 	struct fs *fs;
3877 	ufs_lbn_t i;
3878 #ifdef INVARIANTS
3879 	ufs_lbn_t prevlbn = 0;
3880 #endif
3881 	int deplist;
3882 
3883 	if (inodedep->id_state & IOSTARTED)
3884 		panic("initiate_write_inodeblock_ufs1: already started");
3885 	inodedep->id_state |= IOSTARTED;
3886 	fs = inodedep->id_fs;
3887 	dp = (struct ufs1_dinode *)bp->b_data +
3888 	    ino_to_fsbo(fs, inodedep->id_ino);
3889 	/*
3890 	 * If the bitmap is not yet written, then the allocated
3891 	 * inode cannot be written to disk.
3892 	 */
3893 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3894 		if (inodedep->id_savedino1 != NULL)
3895 			panic("initiate_write_inodeblock_ufs1: I/O underway");
3896 		FREE_LOCK(&lk);
3897 		MALLOC(sip, struct ufs1_dinode *,
3898 		    sizeof(struct ufs1_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS);
3899 		ACQUIRE_LOCK(&lk);
3900 		inodedep->id_savedino1 = sip;
3901 		*inodedep->id_savedino1 = *dp;
3902 		bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
3903 		dp->di_gen = inodedep->id_savedino1->di_gen;
3904 		return;
3905 	}
3906 	/*
3907 	 * If no dependencies, then there is nothing to roll back.
3908 	 */
3909 	inodedep->id_savedsize = dp->di_size;
3910 	inodedep->id_savedextsize = 0;
3911 	if (TAILQ_EMPTY(&inodedep->id_inoupdt))
3912 		return;
3913 	/*
3914 	 * Set the dependencies to busy.
3915 	 */
3916 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3917 	     adp = TAILQ_NEXT(adp, ad_next)) {
3918 #ifdef INVARIANTS
3919 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
3920 			panic("softdep_write_inodeblock: lbn order");
3921 		prevlbn = adp->ad_lbn;
3922 		if (adp->ad_lbn < NDADDR &&
3923 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno)
3924 			panic("%s: direct pointer #%jd mismatch %d != %jd",
3925 			    "softdep_write_inodeblock",
3926 			    (intmax_t)adp->ad_lbn,
3927 			    dp->di_db[adp->ad_lbn],
3928 			    (intmax_t)adp->ad_newblkno);
3929 		if (adp->ad_lbn >= NDADDR &&
3930 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno)
3931 			panic("%s: indirect pointer #%jd mismatch %d != %jd",
3932 			    "softdep_write_inodeblock",
3933 			    (intmax_t)adp->ad_lbn - NDADDR,
3934 			    dp->di_ib[adp->ad_lbn - NDADDR],
3935 			    (intmax_t)adp->ad_newblkno);
3936 		deplist |= 1 << adp->ad_lbn;
3937 		if ((adp->ad_state & ATTACHED) == 0)
3938 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3939 			    adp->ad_state);
3940 #endif /* INVARIANTS */
3941 		adp->ad_state &= ~ATTACHED;
3942 		adp->ad_state |= UNDONE;
3943 	}
3944 	/*
3945 	 * The on-disk inode cannot claim to be any larger than the last
3946 	 * fragment that has been written. Otherwise, the on-disk inode
3947 	 * might have fragments that were not the last block in the file
3948 	 * which would corrupt the filesystem.
3949 	 */
3950 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3951 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3952 		if (adp->ad_lbn >= NDADDR)
3953 			break;
3954 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3955 		/* keep going until hitting a rollback to a frag */
3956 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3957 			continue;
3958 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3959 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3960 #ifdef INVARIANTS
3961 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
3962 				panic("softdep_write_inodeblock: lost dep1");
3963 #endif /* INVARIANTS */
3964 			dp->di_db[i] = 0;
3965 		}
3966 		for (i = 0; i < NIADDR; i++) {
3967 #ifdef INVARIANTS
3968 			if (dp->di_ib[i] != 0 &&
3969 			    (deplist & ((1 << NDADDR) << i)) == 0)
3970 				panic("softdep_write_inodeblock: lost dep2");
3971 #endif /* INVARIANTS */
3972 			dp->di_ib[i] = 0;
3973 		}
3974 		return;
3975 	}
3976 	/*
3977 	 * If we have zero'ed out the last allocated block of the file,
3978 	 * roll back the size to the last currently allocated block.
3979 	 * We know that this last allocated block is a full-sized as
3980 	 * we already checked for fragments in the loop above.
3981 	 */
3982 	if (lastadp != NULL &&
3983 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3984 		for (i = lastadp->ad_lbn; i >= 0; i--)
3985 			if (dp->di_db[i] != 0)
3986 				break;
3987 		dp->di_size = (i + 1) * fs->fs_bsize;
3988 	}
3989 	/*
3990 	 * The only dependencies are for indirect blocks.
3991 	 *
3992 	 * The file size for indirect block additions is not guaranteed.
3993 	 * Such a guarantee would be non-trivial to achieve. The conventional
3994 	 * synchronous write implementation also does not make this guarantee.
3995 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3996 	 * can be over-estimated without destroying integrity when the file
3997 	 * moves into the indirect blocks (i.e., is large). If we want to
3998 	 * postpone fsck, we are stuck with this argument.
3999 	 */
4000 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
4001 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
4002 }
4003 
4004 /*
4005  * Version of initiate_write_inodeblock that handles UFS2 dinodes.
4006  * Note that any bug fixes made to this routine must be done in the
4007  * version found above.
4008  *
4009  * Called from within the procedure above to deal with unsatisfied
4010  * allocation dependencies in an inodeblock. The buffer must be
4011  * locked, thus, no I/O completion operations can occur while we
4012  * are manipulating its associated dependencies.
4013  */
4014 static void
4015 initiate_write_inodeblock_ufs2(inodedep, bp)
4016 	struct inodedep *inodedep;
4017 	struct buf *bp;			/* The inode block */
4018 {
4019 	struct allocdirect *adp, *lastadp;
4020 	struct ufs2_dinode *dp;
4021 	struct ufs2_dinode *sip;
4022 	struct fs *fs;
4023 	ufs_lbn_t i;
4024 #ifdef INVARIANTS
4025 	ufs_lbn_t prevlbn = 0;
4026 #endif
4027 	int deplist;
4028 
4029 	if (inodedep->id_state & IOSTARTED)
4030 		panic("initiate_write_inodeblock_ufs2: already started");
4031 	inodedep->id_state |= IOSTARTED;
4032 	fs = inodedep->id_fs;
4033 	dp = (struct ufs2_dinode *)bp->b_data +
4034 	    ino_to_fsbo(fs, inodedep->id_ino);
4035 	/*
4036 	 * If the bitmap is not yet written, then the allocated
4037 	 * inode cannot be written to disk.
4038 	 */
4039 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4040 		if (inodedep->id_savedino2 != NULL)
4041 			panic("initiate_write_inodeblock_ufs2: I/O underway");
4042 		FREE_LOCK(&lk);
4043 		MALLOC(sip, struct ufs2_dinode *,
4044 		    sizeof(struct ufs2_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS);
4045 		ACQUIRE_LOCK(&lk);
4046 		inodedep->id_savedino2 = sip;
4047 		*inodedep->id_savedino2 = *dp;
4048 		bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
4049 		dp->di_gen = inodedep->id_savedino2->di_gen;
4050 		return;
4051 	}
4052 	/*
4053 	 * If no dependencies, then there is nothing to roll back.
4054 	 */
4055 	inodedep->id_savedsize = dp->di_size;
4056 	inodedep->id_savedextsize = dp->di_extsize;
4057 	if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
4058 	    TAILQ_EMPTY(&inodedep->id_extupdt))
4059 		return;
4060 	/*
4061 	 * Set the ext data dependencies to busy.
4062 	 */
4063 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
4064 	     adp = TAILQ_NEXT(adp, ad_next)) {
4065 #ifdef INVARIANTS
4066 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
4067 			panic("softdep_write_inodeblock: lbn order");
4068 		prevlbn = adp->ad_lbn;
4069 		if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno)
4070 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
4071 			    "softdep_write_inodeblock",
4072 			    (intmax_t)adp->ad_lbn,
4073 			    (intmax_t)dp->di_extb[adp->ad_lbn],
4074 			    (intmax_t)adp->ad_newblkno);
4075 		deplist |= 1 << adp->ad_lbn;
4076 		if ((adp->ad_state & ATTACHED) == 0)
4077 			panic("softdep_write_inodeblock: Unknown state 0x%x",
4078 			    adp->ad_state);
4079 #endif /* INVARIANTS */
4080 		adp->ad_state &= ~ATTACHED;
4081 		adp->ad_state |= UNDONE;
4082 	}
4083 	/*
4084 	 * The on-disk inode cannot claim to be any larger than the last
4085 	 * fragment that has been written. Otherwise, the on-disk inode
4086 	 * might have fragments that were not the last block in the ext
4087 	 * data which would corrupt the filesystem.
4088 	 */
4089 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
4090 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
4091 		dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno;
4092 		/* keep going until hitting a rollback to a frag */
4093 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
4094 			continue;
4095 		dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
4096 		for (i = adp->ad_lbn + 1; i < NXADDR; i++) {
4097 #ifdef INVARIANTS
4098 			if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0)
4099 				panic("softdep_write_inodeblock: lost dep1");
4100 #endif /* INVARIANTS */
4101 			dp->di_extb[i] = 0;
4102 		}
4103 		lastadp = NULL;
4104 		break;
4105 	}
4106 	/*
4107 	 * If we have zero'ed out the last allocated block of the ext
4108 	 * data, roll back the size to the last currently allocated block.
4109 	 * We know that this last allocated block is a full-sized as
4110 	 * we already checked for fragments in the loop above.
4111 	 */
4112 	if (lastadp != NULL &&
4113 	    dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
4114 		for (i = lastadp->ad_lbn; i >= 0; i--)
4115 			if (dp->di_extb[i] != 0)
4116 				break;
4117 		dp->di_extsize = (i + 1) * fs->fs_bsize;
4118 	}
4119 	/*
4120 	 * Set the file data dependencies to busy.
4121 	 */
4122 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
4123 	     adp = TAILQ_NEXT(adp, ad_next)) {
4124 #ifdef INVARIANTS
4125 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
4126 			panic("softdep_write_inodeblock: lbn order");
4127 		prevlbn = adp->ad_lbn;
4128 		if (adp->ad_lbn < NDADDR &&
4129 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno)
4130 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
4131 			    "softdep_write_inodeblock",
4132 			    (intmax_t)adp->ad_lbn,
4133 			    (intmax_t)dp->di_db[adp->ad_lbn],
4134 			    (intmax_t)adp->ad_newblkno);
4135 		if (adp->ad_lbn >= NDADDR &&
4136 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno)
4137 			panic("%s indirect pointer #%jd mismatch %jd != %jd",
4138 			    "softdep_write_inodeblock:",
4139 			    (intmax_t)adp->ad_lbn - NDADDR,
4140 			    (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR],
4141 			    (intmax_t)adp->ad_newblkno);
4142 		deplist |= 1 << adp->ad_lbn;
4143 		if ((adp->ad_state & ATTACHED) == 0)
4144 			panic("softdep_write_inodeblock: Unknown state 0x%x",
4145 			    adp->ad_state);
4146 #endif /* INVARIANTS */
4147 		adp->ad_state &= ~ATTACHED;
4148 		adp->ad_state |= UNDONE;
4149 	}
4150 	/*
4151 	 * The on-disk inode cannot claim to be any larger than the last
4152 	 * fragment that has been written. Otherwise, the on-disk inode
4153 	 * might have fragments that were not the last block in the file
4154 	 * which would corrupt the filesystem.
4155 	 */
4156 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
4157 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
4158 		if (adp->ad_lbn >= NDADDR)
4159 			break;
4160 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
4161 		/* keep going until hitting a rollback to a frag */
4162 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
4163 			continue;
4164 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
4165 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
4166 #ifdef INVARIANTS
4167 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
4168 				panic("softdep_write_inodeblock: lost dep2");
4169 #endif /* INVARIANTS */
4170 			dp->di_db[i] = 0;
4171 		}
4172 		for (i = 0; i < NIADDR; i++) {
4173 #ifdef INVARIANTS
4174 			if (dp->di_ib[i] != 0 &&
4175 			    (deplist & ((1 << NDADDR) << i)) == 0)
4176 				panic("softdep_write_inodeblock: lost dep3");
4177 #endif /* INVARIANTS */
4178 			dp->di_ib[i] = 0;
4179 		}
4180 		return;
4181 	}
4182 	/*
4183 	 * If we have zero'ed out the last allocated block of the file,
4184 	 * roll back the size to the last currently allocated block.
4185 	 * We know that this last allocated block is a full-sized as
4186 	 * we already checked for fragments in the loop above.
4187 	 */
4188 	if (lastadp != NULL &&
4189 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
4190 		for (i = lastadp->ad_lbn; i >= 0; i--)
4191 			if (dp->di_db[i] != 0)
4192 				break;
4193 		dp->di_size = (i + 1) * fs->fs_bsize;
4194 	}
4195 	/*
4196 	 * The only dependencies are for indirect blocks.
4197 	 *
4198 	 * The file size for indirect block additions is not guaranteed.
4199 	 * Such a guarantee would be non-trivial to achieve. The conventional
4200 	 * synchronous write implementation also does not make this guarantee.
4201 	 * Fsck should catch and fix discrepancies. Arguably, the file size
4202 	 * can be over-estimated without destroying integrity when the file
4203 	 * moves into the indirect blocks (i.e., is large). If we want to
4204 	 * postpone fsck, we are stuck with this argument.
4205 	 */
4206 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
4207 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
4208 }
4209 
4210 /*
4211  * This routine is called during the completion interrupt
4212  * service routine for a disk write (from the procedure called
4213  * by the device driver to inform the filesystem caches of
4214  * a request completion).  It should be called early in this
4215  * procedure, before the block is made available to other
4216  * processes or other routines are called.
4217  */
4218 static void
4219 softdep_disk_write_complete(bp)
4220 	struct buf *bp;		/* describes the completed disk write */
4221 {
4222 	struct worklist *wk;
4223 	struct worklist *owk;
4224 	struct workhead reattach;
4225 	struct newblk *newblk;
4226 	struct allocindir *aip;
4227 	struct allocdirect *adp;
4228 	struct indirdep *indirdep;
4229 	struct inodedep *inodedep;
4230 	struct bmsafemap *bmsafemap;
4231 
4232 	/*
4233 	 * If an error occurred while doing the write, then the data
4234 	 * has not hit the disk and the dependencies cannot be unrolled.
4235 	 */
4236 	if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0)
4237 		return;
4238 	LIST_INIT(&reattach);
4239 	/*
4240 	 * This lock must not be released anywhere in this code segment.
4241 	 */
4242 	ACQUIRE_LOCK(&lk);
4243 	owk = NULL;
4244 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
4245 		WORKLIST_REMOVE(wk);
4246 		if (wk == owk)
4247 			panic("duplicate worklist: %p\n", wk);
4248 		owk = wk;
4249 		switch (wk->wk_type) {
4250 
4251 		case D_PAGEDEP:
4252 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
4253 				WORKLIST_INSERT(&reattach, wk);
4254 			continue;
4255 
4256 		case D_INODEDEP:
4257 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
4258 				WORKLIST_INSERT(&reattach, wk);
4259 			continue;
4260 
4261 		case D_BMSAFEMAP:
4262 			bmsafemap = WK_BMSAFEMAP(wk);
4263 			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
4264 				newblk->nb_state |= DEPCOMPLETE;
4265 				newblk->nb_bmsafemap = NULL;
4266 				LIST_REMOVE(newblk, nb_deps);
4267 			}
4268 			while ((adp =
4269 			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
4270 				adp->ad_state |= DEPCOMPLETE;
4271 				adp->ad_buf = NULL;
4272 				LIST_REMOVE(adp, ad_deps);
4273 				handle_allocdirect_partdone(adp);
4274 			}
4275 			while ((aip =
4276 			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
4277 				aip->ai_state |= DEPCOMPLETE;
4278 				aip->ai_buf = NULL;
4279 				LIST_REMOVE(aip, ai_deps);
4280 				handle_allocindir_partdone(aip);
4281 			}
4282 			while ((inodedep =
4283 			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
4284 				inodedep->id_state |= DEPCOMPLETE;
4285 				LIST_REMOVE(inodedep, id_deps);
4286 				inodedep->id_buf = NULL;
4287 			}
4288 			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
4289 			continue;
4290 
4291 		case D_MKDIR:
4292 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
4293 			continue;
4294 
4295 		case D_ALLOCDIRECT:
4296 			adp = WK_ALLOCDIRECT(wk);
4297 			adp->ad_state |= COMPLETE;
4298 			handle_allocdirect_partdone(adp);
4299 			continue;
4300 
4301 		case D_ALLOCINDIR:
4302 			aip = WK_ALLOCINDIR(wk);
4303 			aip->ai_state |= COMPLETE;
4304 			handle_allocindir_partdone(aip);
4305 			continue;
4306 
4307 		case D_INDIRDEP:
4308 			indirdep = WK_INDIRDEP(wk);
4309 			if (indirdep->ir_state & GOINGAWAY)
4310 				panic("disk_write_complete: indirdep gone");
4311 			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
4312 			FREE(indirdep->ir_saveddata, M_INDIRDEP);
4313 			indirdep->ir_saveddata = 0;
4314 			indirdep->ir_state &= ~UNDONE;
4315 			indirdep->ir_state |= ATTACHED;
4316 			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
4317 				handle_allocindir_partdone(aip);
4318 				if (aip == LIST_FIRST(&indirdep->ir_donehd))
4319 					panic("disk_write_complete: not gone");
4320 			}
4321 			WORKLIST_INSERT(&reattach, wk);
4322 			if ((bp->b_flags & B_DELWRI) == 0)
4323 				stat_indir_blk_ptrs++;
4324 			bdirty(bp);
4325 			continue;
4326 
4327 		default:
4328 			panic("handle_disk_write_complete: Unknown type %s",
4329 			    TYPENAME(wk->wk_type));
4330 			/* NOTREACHED */
4331 		}
4332 	}
4333 	/*
4334 	 * Reattach any requests that must be redone.
4335 	 */
4336 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
4337 		WORKLIST_REMOVE(wk);
4338 		WORKLIST_INSERT(&bp->b_dep, wk);
4339 	}
4340 	FREE_LOCK(&lk);
4341 }
4342 
4343 /*
4344  * Called from within softdep_disk_write_complete above. Note that
4345  * this routine is always called from interrupt level with further
4346  * splbio interrupts blocked.
4347  */
4348 static void
4349 handle_allocdirect_partdone(adp)
4350 	struct allocdirect *adp;	/* the completed allocdirect */
4351 {
4352 	struct allocdirectlst *listhead;
4353 	struct allocdirect *listadp;
4354 	struct inodedep *inodedep;
4355 	long bsize, delay;
4356 
4357 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4358 		return;
4359 	if (adp->ad_buf != NULL)
4360 		panic("handle_allocdirect_partdone: dangling dep");
4361 	/*
4362 	 * The on-disk inode cannot claim to be any larger than the last
4363 	 * fragment that has been written. Otherwise, the on-disk inode
4364 	 * might have fragments that were not the last block in the file
4365 	 * which would corrupt the filesystem. Thus, we cannot free any
4366 	 * allocdirects after one whose ad_oldblkno claims a fragment as
4367 	 * these blocks must be rolled back to zero before writing the inode.
4368 	 * We check the currently active set of allocdirects in id_inoupdt
4369 	 * or id_extupdt as appropriate.
4370 	 */
4371 	inodedep = adp->ad_inodedep;
4372 	bsize = inodedep->id_fs->fs_bsize;
4373 	if (adp->ad_state & EXTDATA)
4374 		listhead = &inodedep->id_extupdt;
4375 	else
4376 		listhead = &inodedep->id_inoupdt;
4377 	TAILQ_FOREACH(listadp, listhead, ad_next) {
4378 		/* found our block */
4379 		if (listadp == adp)
4380 			break;
4381 		/* continue if ad_oldlbn is not a fragment */
4382 		if (listadp->ad_oldsize == 0 ||
4383 		    listadp->ad_oldsize == bsize)
4384 			continue;
4385 		/* hit a fragment */
4386 		return;
4387 	}
4388 	/*
4389 	 * If we have reached the end of the current list without
4390 	 * finding the just finished dependency, then it must be
4391 	 * on the future dependency list. Future dependencies cannot
4392 	 * be freed until they are moved to the current list.
4393 	 */
4394 	if (listadp == NULL) {
4395 #ifdef DEBUG
4396 		if (adp->ad_state & EXTDATA)
4397 			listhead = &inodedep->id_newextupdt;
4398 		else
4399 			listhead = &inodedep->id_newinoupdt;
4400 		TAILQ_FOREACH(listadp, listhead, ad_next)
4401 			/* found our block */
4402 			if (listadp == adp)
4403 				break;
4404 		if (listadp == NULL)
4405 			panic("handle_allocdirect_partdone: lost dep");
4406 #endif /* DEBUG */
4407 		return;
4408 	}
4409 	/*
4410 	 * If we have found the just finished dependency, then free
4411 	 * it along with anything that follows it that is complete.
4412 	 * If the inode still has a bitmap dependency, then it has
4413 	 * never been written to disk, hence the on-disk inode cannot
4414 	 * reference the old fragment so we can free it without delay.
4415 	 */
4416 	delay = (inodedep->id_state & DEPCOMPLETE);
4417 	for (; adp; adp = listadp) {
4418 		listadp = TAILQ_NEXT(adp, ad_next);
4419 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4420 			return;
4421 		free_allocdirect(listhead, adp, delay);
4422 	}
4423 }
4424 
4425 /*
4426  * Called from within softdep_disk_write_complete above. Note that
4427  * this routine is always called from interrupt level with further
4428  * splbio interrupts blocked.
4429  */
4430 static void
4431 handle_allocindir_partdone(aip)
4432 	struct allocindir *aip;		/* the completed allocindir */
4433 {
4434 	struct indirdep *indirdep;
4435 
4436 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
4437 		return;
4438 	if (aip->ai_buf != NULL)
4439 		panic("handle_allocindir_partdone: dangling dependency");
4440 	indirdep = aip->ai_indirdep;
4441 	if (indirdep->ir_state & UNDONE) {
4442 		LIST_REMOVE(aip, ai_next);
4443 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
4444 		return;
4445 	}
4446 	if (indirdep->ir_state & UFS1FMT)
4447 		((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4448 		    aip->ai_newblkno;
4449 	else
4450 		((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4451 		    aip->ai_newblkno;
4452 	LIST_REMOVE(aip, ai_next);
4453 	if (aip->ai_freefrag != NULL)
4454 		add_to_worklist(&aip->ai_freefrag->ff_list);
4455 	WORKITEM_FREE(aip, D_ALLOCINDIR);
4456 }
4457 
4458 /*
4459  * Called from within softdep_disk_write_complete above to restore
4460  * in-memory inode block contents to their most up-to-date state. Note
4461  * that this routine is always called from interrupt level with further
4462  * splbio interrupts blocked.
4463  */
4464 static int
4465 handle_written_inodeblock(inodedep, bp)
4466 	struct inodedep *inodedep;
4467 	struct buf *bp;		/* buffer containing the inode block */
4468 {
4469 	struct worklist *wk, *filefree;
4470 	struct allocdirect *adp, *nextadp;
4471 	struct ufs1_dinode *dp1 = NULL;
4472 	struct ufs2_dinode *dp2 = NULL;
4473 	int hadchanges, fstype;
4474 
4475 	if ((inodedep->id_state & IOSTARTED) == 0)
4476 		panic("handle_written_inodeblock: not started");
4477 	inodedep->id_state &= ~IOSTARTED;
4478 	if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
4479 		fstype = UFS1;
4480 		dp1 = (struct ufs1_dinode *)bp->b_data +
4481 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4482 	} else {
4483 		fstype = UFS2;
4484 		dp2 = (struct ufs2_dinode *)bp->b_data +
4485 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4486 	}
4487 	/*
4488 	 * If we had to rollback the inode allocation because of
4489 	 * bitmaps being incomplete, then simply restore it.
4490 	 * Keep the block dirty so that it will not be reclaimed until
4491 	 * all associated dependencies have been cleared and the
4492 	 * corresponding updates written to disk.
4493 	 */
4494 	if (inodedep->id_savedino1 != NULL) {
4495 		if (fstype == UFS1)
4496 			*dp1 = *inodedep->id_savedino1;
4497 		else
4498 			*dp2 = *inodedep->id_savedino2;
4499 		FREE(inodedep->id_savedino1, M_SAVEDINO);
4500 		inodedep->id_savedino1 = NULL;
4501 		if ((bp->b_flags & B_DELWRI) == 0)
4502 			stat_inode_bitmap++;
4503 		bdirty(bp);
4504 		return (1);
4505 	}
4506 	inodedep->id_state |= COMPLETE;
4507 	/*
4508 	 * Roll forward anything that had to be rolled back before
4509 	 * the inode could be updated.
4510 	 */
4511 	hadchanges = 0;
4512 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
4513 		nextadp = TAILQ_NEXT(adp, ad_next);
4514 		if (adp->ad_state & ATTACHED)
4515 			panic("handle_written_inodeblock: new entry");
4516 		if (fstype == UFS1) {
4517 			if (adp->ad_lbn < NDADDR) {
4518 				if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno)
4519 					panic("%s %s #%jd mismatch %d != %jd",
4520 					    "handle_written_inodeblock:",
4521 					    "direct pointer",
4522 					    (intmax_t)adp->ad_lbn,
4523 					    dp1->di_db[adp->ad_lbn],
4524 					    (intmax_t)adp->ad_oldblkno);
4525 				dp1->di_db[adp->ad_lbn] = adp->ad_newblkno;
4526 			} else {
4527 				if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0)
4528 					panic("%s: %s #%jd allocated as %d",
4529 					    "handle_written_inodeblock",
4530 					    "indirect pointer",
4531 					    (intmax_t)adp->ad_lbn - NDADDR,
4532 					    dp1->di_ib[adp->ad_lbn - NDADDR]);
4533 				dp1->di_ib[adp->ad_lbn - NDADDR] =
4534 				    adp->ad_newblkno;
4535 			}
4536 		} else {
4537 			if (adp->ad_lbn < NDADDR) {
4538 				if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno)
4539 					panic("%s: %s #%jd %s %jd != %jd",
4540 					    "handle_written_inodeblock",
4541 					    "direct pointer",
4542 					    (intmax_t)adp->ad_lbn, "mismatch",
4543 					    (intmax_t)dp2->di_db[adp->ad_lbn],
4544 					    (intmax_t)adp->ad_oldblkno);
4545 				dp2->di_db[adp->ad_lbn] = adp->ad_newblkno;
4546 			} else {
4547 				if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0)
4548 					panic("%s: %s #%jd allocated as %jd",
4549 					    "handle_written_inodeblock",
4550 					    "indirect pointer",
4551 					    (intmax_t)adp->ad_lbn - NDADDR,
4552 					    (intmax_t)
4553 					    dp2->di_ib[adp->ad_lbn - NDADDR]);
4554 				dp2->di_ib[adp->ad_lbn - NDADDR] =
4555 				    adp->ad_newblkno;
4556 			}
4557 		}
4558 		adp->ad_state &= ~UNDONE;
4559 		adp->ad_state |= ATTACHED;
4560 		hadchanges = 1;
4561 	}
4562 	for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
4563 		nextadp = TAILQ_NEXT(adp, ad_next);
4564 		if (adp->ad_state & ATTACHED)
4565 			panic("handle_written_inodeblock: new entry");
4566 		if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno)
4567 			panic("%s: direct pointers #%jd %s %jd != %jd",
4568 			    "handle_written_inodeblock",
4569 			    (intmax_t)adp->ad_lbn, "mismatch",
4570 			    (intmax_t)dp2->di_extb[adp->ad_lbn],
4571 			    (intmax_t)adp->ad_oldblkno);
4572 		dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno;
4573 		adp->ad_state &= ~UNDONE;
4574 		adp->ad_state |= ATTACHED;
4575 		hadchanges = 1;
4576 	}
4577 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
4578 		stat_direct_blk_ptrs++;
4579 	/*
4580 	 * Reset the file size to its most up-to-date value.
4581 	 */
4582 	if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1)
4583 		panic("handle_written_inodeblock: bad size");
4584 	if (fstype == UFS1) {
4585 		if (dp1->di_size != inodedep->id_savedsize) {
4586 			dp1->di_size = inodedep->id_savedsize;
4587 			hadchanges = 1;
4588 		}
4589 	} else {
4590 		if (dp2->di_size != inodedep->id_savedsize) {
4591 			dp2->di_size = inodedep->id_savedsize;
4592 			hadchanges = 1;
4593 		}
4594 		if (dp2->di_extsize != inodedep->id_savedextsize) {
4595 			dp2->di_extsize = inodedep->id_savedextsize;
4596 			hadchanges = 1;
4597 		}
4598 	}
4599 	inodedep->id_savedsize = -1;
4600 	inodedep->id_savedextsize = -1;
4601 	/*
4602 	 * If there were any rollbacks in the inode block, then it must be
4603 	 * marked dirty so that its will eventually get written back in
4604 	 * its correct form.
4605 	 */
4606 	if (hadchanges)
4607 		bdirty(bp);
4608 	/*
4609 	 * Process any allocdirects that completed during the update.
4610 	 */
4611 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
4612 		handle_allocdirect_partdone(adp);
4613 	if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
4614 		handle_allocdirect_partdone(adp);
4615 	/*
4616 	 * Process deallocations that were held pending until the
4617 	 * inode had been written to disk. Freeing of the inode
4618 	 * is delayed until after all blocks have been freed to
4619 	 * avoid creation of new <vfsid, inum, lbn> triples
4620 	 * before the old ones have been deleted.
4621 	 */
4622 	filefree = NULL;
4623 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
4624 		WORKLIST_REMOVE(wk);
4625 		switch (wk->wk_type) {
4626 
4627 		case D_FREEFILE:
4628 			/*
4629 			 * We defer adding filefree to the worklist until
4630 			 * all other additions have been made to ensure
4631 			 * that it will be done after all the old blocks
4632 			 * have been freed.
4633 			 */
4634 			if (filefree != NULL)
4635 				panic("handle_written_inodeblock: filefree");
4636 			filefree = wk;
4637 			continue;
4638 
4639 		case D_MKDIR:
4640 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
4641 			continue;
4642 
4643 		case D_DIRADD:
4644 			diradd_inode_written(WK_DIRADD(wk), inodedep);
4645 			continue;
4646 
4647 		case D_FREEBLKS:
4648 			wk->wk_state |= COMPLETE;
4649 			if ((wk->wk_state  & ALLCOMPLETE) != ALLCOMPLETE)
4650 				continue;
4651 			 /* -- fall through -- */
4652 		case D_FREEFRAG:
4653 		case D_DIRREM:
4654 			add_to_worklist(wk);
4655 			continue;
4656 
4657 		case D_NEWDIRBLK:
4658 			free_newdirblk(WK_NEWDIRBLK(wk));
4659 			continue;
4660 
4661 		default:
4662 			panic("handle_written_inodeblock: Unknown type %s",
4663 			    TYPENAME(wk->wk_type));
4664 			/* NOTREACHED */
4665 		}
4666 	}
4667 	if (filefree != NULL) {
4668 		if (free_inodedep(inodedep) == 0)
4669 			panic("handle_written_inodeblock: live inodedep");
4670 		add_to_worklist(filefree);
4671 		return (0);
4672 	}
4673 
4674 	/*
4675 	 * If no outstanding dependencies, free it.
4676 	 */
4677 	if (free_inodedep(inodedep) ||
4678 	    (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
4679 	     TAILQ_FIRST(&inodedep->id_extupdt) == 0))
4680 		return (0);
4681 	return (hadchanges);
4682 }
4683 
4684 /*
4685  * Process a diradd entry after its dependent inode has been written.
4686  * This routine must be called with splbio interrupts blocked.
4687  */
4688 static void
4689 diradd_inode_written(dap, inodedep)
4690 	struct diradd *dap;
4691 	struct inodedep *inodedep;
4692 {
4693 	struct pagedep *pagedep;
4694 
4695 	dap->da_state |= COMPLETE;
4696 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4697 		if (dap->da_state & DIRCHG)
4698 			pagedep = dap->da_previous->dm_pagedep;
4699 		else
4700 			pagedep = dap->da_pagedep;
4701 		LIST_REMOVE(dap, da_pdlist);
4702 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4703 	}
4704 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
4705 }
4706 
4707 /*
4708  * Handle the completion of a mkdir dependency.
4709  */
4710 static void
4711 handle_written_mkdir(mkdir, type)
4712 	struct mkdir *mkdir;
4713 	int type;
4714 {
4715 	struct diradd *dap;
4716 	struct pagedep *pagedep;
4717 
4718 	if (mkdir->md_state != type)
4719 		panic("handle_written_mkdir: bad type");
4720 	dap = mkdir->md_diradd;
4721 	dap->da_state &= ~type;
4722 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
4723 		dap->da_state |= DEPCOMPLETE;
4724 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4725 		if (dap->da_state & DIRCHG)
4726 			pagedep = dap->da_previous->dm_pagedep;
4727 		else
4728 			pagedep = dap->da_pagedep;
4729 		LIST_REMOVE(dap, da_pdlist);
4730 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4731 	}
4732 	LIST_REMOVE(mkdir, md_mkdirs);
4733 	WORKITEM_FREE(mkdir, D_MKDIR);
4734 }
4735 
4736 /*
4737  * Called from within softdep_disk_write_complete above.
4738  * A write operation was just completed. Removed inodes can
4739  * now be freed and associated block pointers may be committed.
4740  * Note that this routine is always called from interrupt level
4741  * with further splbio interrupts blocked.
4742  */
4743 static int
4744 handle_written_filepage(pagedep, bp)
4745 	struct pagedep *pagedep;
4746 	struct buf *bp;		/* buffer containing the written page */
4747 {
4748 	struct dirrem *dirrem;
4749 	struct diradd *dap, *nextdap;
4750 	struct direct *ep;
4751 	int i, chgs;
4752 
4753 	if ((pagedep->pd_state & IOSTARTED) == 0)
4754 		panic("handle_written_filepage: not started");
4755 	pagedep->pd_state &= ~IOSTARTED;
4756 	/*
4757 	 * Process any directory removals that have been committed.
4758 	 */
4759 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
4760 		LIST_REMOVE(dirrem, dm_next);
4761 		dirrem->dm_dirinum = pagedep->pd_ino;
4762 		add_to_worklist(&dirrem->dm_list);
4763 	}
4764 	/*
4765 	 * Free any directory additions that have been committed.
4766 	 * If it is a newly allocated block, we have to wait until
4767 	 * the on-disk directory inode claims the new block.
4768 	 */
4769 	if ((pagedep->pd_state & NEWBLOCK) == 0)
4770 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
4771 			free_diradd(dap);
4772 	/*
4773 	 * Uncommitted directory entries must be restored.
4774 	 */
4775 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
4776 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
4777 		     dap = nextdap) {
4778 			nextdap = LIST_NEXT(dap, da_pdlist);
4779 			if (dap->da_state & ATTACHED)
4780 				panic("handle_written_filepage: attached");
4781 			ep = (struct direct *)
4782 			    ((char *)bp->b_data + dap->da_offset);
4783 			ep->d_ino = dap->da_newinum;
4784 			dap->da_state &= ~UNDONE;
4785 			dap->da_state |= ATTACHED;
4786 			chgs = 1;
4787 			/*
4788 			 * If the inode referenced by the directory has
4789 			 * been written out, then the dependency can be
4790 			 * moved to the pending list.
4791 			 */
4792 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4793 				LIST_REMOVE(dap, da_pdlist);
4794 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
4795 				    da_pdlist);
4796 			}
4797 		}
4798 	}
4799 	/*
4800 	 * If there were any rollbacks in the directory, then it must be
4801 	 * marked dirty so that its will eventually get written back in
4802 	 * its correct form.
4803 	 */
4804 	if (chgs) {
4805 		if ((bp->b_flags & B_DELWRI) == 0)
4806 			stat_dir_entry++;
4807 		bdirty(bp);
4808 		return (1);
4809 	}
4810 	/*
4811 	 * If we are not waiting for a new directory block to be
4812 	 * claimed by its inode, then the pagedep will be freed.
4813 	 * Otherwise it will remain to track any new entries on
4814 	 * the page in case they are fsync'ed.
4815 	 */
4816 	if ((pagedep->pd_state & NEWBLOCK) == 0) {
4817 		LIST_REMOVE(pagedep, pd_hash);
4818 		WORKITEM_FREE(pagedep, D_PAGEDEP);
4819 	}
4820 	return (0);
4821 }
4822 
4823 /*
4824  * Writing back in-core inode structures.
4825  *
4826  * The filesystem only accesses an inode's contents when it occupies an
4827  * "in-core" inode structure.  These "in-core" structures are separate from
4828  * the page frames used to cache inode blocks.  Only the latter are
4829  * transferred to/from the disk.  So, when the updated contents of the
4830  * "in-core" inode structure are copied to the corresponding in-memory inode
4831  * block, the dependencies are also transferred.  The following procedure is
4832  * called when copying a dirty "in-core" inode to a cached inode block.
4833  */
4834 
4835 /*
4836  * Called when an inode is loaded from disk. If the effective link count
4837  * differed from the actual link count when it was last flushed, then we
4838  * need to ensure that the correct effective link count is put back.
4839  */
4840 void
4841 softdep_load_inodeblock(ip)
4842 	struct inode *ip;	/* the "in_core" copy of the inode */
4843 {
4844 	struct inodedep *inodedep;
4845 
4846 	/*
4847 	 * Check for alternate nlink count.
4848 	 */
4849 	ip->i_effnlink = ip->i_nlink;
4850 	ACQUIRE_LOCK(&lk);
4851 	if (inodedep_lookup(UFSTOVFS(ip->i_ump),
4852 	    ip->i_number, 0, &inodedep) == 0) {
4853 		FREE_LOCK(&lk);
4854 		return;
4855 	}
4856 	ip->i_effnlink -= inodedep->id_nlinkdelta;
4857 	if (inodedep->id_state & SPACECOUNTED)
4858 		ip->i_flag |= IN_SPACECOUNTED;
4859 	FREE_LOCK(&lk);
4860 }
4861 
4862 /*
4863  * This routine is called just before the "in-core" inode
4864  * information is to be copied to the in-memory inode block.
4865  * Recall that an inode block contains several inodes. If
4866  * the force flag is set, then the dependencies will be
4867  * cleared so that the update can always be made. Note that
4868  * the buffer is locked when this routine is called, so we
4869  * will never be in the middle of writing the inode block
4870  * to disk.
4871  */
4872 void
4873 softdep_update_inodeblock(ip, bp, waitfor)
4874 	struct inode *ip;	/* the "in_core" copy of the inode */
4875 	struct buf *bp;		/* the buffer containing the inode block */
4876 	int waitfor;		/* nonzero => update must be allowed */
4877 {
4878 	struct inodedep *inodedep;
4879 	struct worklist *wk;
4880 	struct mount *mp;
4881 	struct buf *ibp;
4882 	int error;
4883 
4884 	/*
4885 	 * If the effective link count is not equal to the actual link
4886 	 * count, then we must track the difference in an inodedep while
4887 	 * the inode is (potentially) tossed out of the cache. Otherwise,
4888 	 * if there is no existing inodedep, then there are no dependencies
4889 	 * to track.
4890 	 */
4891 	mp = UFSTOVFS(ip->i_ump);
4892 	ACQUIRE_LOCK(&lk);
4893 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
4894 		FREE_LOCK(&lk);
4895 		if (ip->i_effnlink != ip->i_nlink)
4896 			panic("softdep_update_inodeblock: bad link count");
4897 		return;
4898 	}
4899 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
4900 		panic("softdep_update_inodeblock: bad delta");
4901 	/*
4902 	 * Changes have been initiated. Anything depending on these
4903 	 * changes cannot occur until this inode has been written.
4904 	 */
4905 	inodedep->id_state &= ~COMPLETE;
4906 	if ((inodedep->id_state & ONWORKLIST) == 0)
4907 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
4908 	/*
4909 	 * Any new dependencies associated with the incore inode must
4910 	 * now be moved to the list associated with the buffer holding
4911 	 * the in-memory copy of the inode. Once merged process any
4912 	 * allocdirects that are completed by the merger.
4913 	 */
4914 	merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
4915 	if (!TAILQ_EMPTY(&inodedep->id_inoupdt))
4916 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
4917 	merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
4918 	if (!TAILQ_EMPTY(&inodedep->id_extupdt))
4919 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt));
4920 	/*
4921 	 * Now that the inode has been pushed into the buffer, the
4922 	 * operations dependent on the inode being written to disk
4923 	 * can be moved to the id_bufwait so that they will be
4924 	 * processed when the buffer I/O completes.
4925 	 */
4926 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
4927 		WORKLIST_REMOVE(wk);
4928 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
4929 	}
4930 	/*
4931 	 * Newly allocated inodes cannot be written until the bitmap
4932 	 * that allocates them have been written (indicated by
4933 	 * DEPCOMPLETE being set in id_state). If we are doing a
4934 	 * forced sync (e.g., an fsync on a file), we force the bitmap
4935 	 * to be written so that the update can be done.
4936 	 */
4937 	if (waitfor == 0) {
4938 		FREE_LOCK(&lk);
4939 		return;
4940 	}
4941 retry:
4942 	if ((inodedep->id_state & DEPCOMPLETE) != 0) {
4943 		FREE_LOCK(&lk);
4944 		return;
4945 	}
4946 	ibp = inodedep->id_buf;
4947 	ibp = getdirtybuf(ibp, &lk, MNT_WAIT);
4948 	if (ibp == NULL) {
4949 		/*
4950 		 * If ibp came back as NULL, the dependency could have been
4951 		 * freed while we slept.  Look it up again, and check to see
4952 		 * that it has completed.
4953 		 */
4954 		if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
4955 			goto retry;
4956 		FREE_LOCK(&lk);
4957 		return;
4958 	}
4959 	FREE_LOCK(&lk);
4960 	if ((error = bwrite(ibp)) != 0)
4961 		softdep_error("softdep_update_inodeblock: bwrite", error);
4962 }
4963 
4964 /*
4965  * Merge the a new inode dependency list (such as id_newinoupdt) into an
4966  * old inode dependency list (such as id_inoupdt). This routine must be
4967  * called with splbio interrupts blocked.
4968  */
4969 static void
4970 merge_inode_lists(newlisthead, oldlisthead)
4971 	struct allocdirectlst *newlisthead;
4972 	struct allocdirectlst *oldlisthead;
4973 {
4974 	struct allocdirect *listadp, *newadp;
4975 
4976 	newadp = TAILQ_FIRST(newlisthead);
4977 	for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
4978 		if (listadp->ad_lbn < newadp->ad_lbn) {
4979 			listadp = TAILQ_NEXT(listadp, ad_next);
4980 			continue;
4981 		}
4982 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4983 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
4984 		if (listadp->ad_lbn == newadp->ad_lbn) {
4985 			allocdirect_merge(oldlisthead, newadp,
4986 			    listadp);
4987 			listadp = newadp;
4988 		}
4989 		newadp = TAILQ_FIRST(newlisthead);
4990 	}
4991 	while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
4992 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4993 		TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
4994 	}
4995 }
4996 
4997 /*
4998  * If we are doing an fsync, then we must ensure that any directory
4999  * entries for the inode have been written after the inode gets to disk.
5000  */
5001 int
5002 softdep_fsync(vp)
5003 	struct vnode *vp;	/* the "in_core" copy of the inode */
5004 {
5005 	struct inodedep *inodedep;
5006 	struct pagedep *pagedep;
5007 	struct worklist *wk;
5008 	struct diradd *dap;
5009 	struct mount *mp;
5010 	struct vnode *pvp;
5011 	struct inode *ip;
5012 	struct buf *bp;
5013 	struct fs *fs;
5014 	struct thread *td = curthread;
5015 	int error, flushparent, pagedep_new_block;
5016 	ino_t parentino;
5017 	ufs_lbn_t lbn;
5018 
5019 	ip = VTOI(vp);
5020 	fs = ip->i_fs;
5021 	mp = vp->v_mount;
5022 	ACQUIRE_LOCK(&lk);
5023 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
5024 		FREE_LOCK(&lk);
5025 		return (0);
5026 	}
5027 	if (!LIST_EMPTY(&inodedep->id_inowait) ||
5028 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
5029 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
5030 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
5031 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
5032 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt))
5033 		panic("softdep_fsync: pending ops");
5034 	for (error = 0, flushparent = 0; ; ) {
5035 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
5036 			break;
5037 		if (wk->wk_type != D_DIRADD)
5038 			panic("softdep_fsync: Unexpected type %s",
5039 			    TYPENAME(wk->wk_type));
5040 		dap = WK_DIRADD(wk);
5041 		/*
5042 		 * Flush our parent if this directory entry has a MKDIR_PARENT
5043 		 * dependency or is contained in a newly allocated block.
5044 		 */
5045 		if (dap->da_state & DIRCHG)
5046 			pagedep = dap->da_previous->dm_pagedep;
5047 		else
5048 			pagedep = dap->da_pagedep;
5049 		parentino = pagedep->pd_ino;
5050 		lbn = pagedep->pd_lbn;
5051 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
5052 			panic("softdep_fsync: dirty");
5053 		if ((dap->da_state & MKDIR_PARENT) ||
5054 		    (pagedep->pd_state & NEWBLOCK))
5055 			flushparent = 1;
5056 		else
5057 			flushparent = 0;
5058 		/*
5059 		 * If we are being fsync'ed as part of vgone'ing this vnode,
5060 		 * then we will not be able to release and recover the
5061 		 * vnode below, so we just have to give up on writing its
5062 		 * directory entry out. It will eventually be written, just
5063 		 * not now, but then the user was not asking to have it
5064 		 * written, so we are not breaking any promises.
5065 		 */
5066 		if (vp->v_iflag & VI_DOOMED)
5067 			break;
5068 		/*
5069 		 * We prevent deadlock by always fetching inodes from the
5070 		 * root, moving down the directory tree. Thus, when fetching
5071 		 * our parent directory, we first try to get the lock. If
5072 		 * that fails, we must unlock ourselves before requesting
5073 		 * the lock on our parent. See the comment in ufs_lookup
5074 		 * for details on possible races.
5075 		 */
5076 		FREE_LOCK(&lk);
5077 		if (ffs_vget(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) {
5078 			VOP_UNLOCK(vp, 0, td);
5079 			error = ffs_vget(mp, parentino, LK_EXCLUSIVE, &pvp);
5080 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
5081 			if (error != 0)
5082 				return (error);
5083 		}
5084 		/*
5085 		 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
5086 		 * that are contained in direct blocks will be resolved by
5087 		 * doing a ffs_update. Pagedeps contained in indirect blocks
5088 		 * may require a complete sync'ing of the directory. So, we
5089 		 * try the cheap and fast ffs_update first, and if that fails,
5090 		 * then we do the slower ffs_syncvnode of the directory.
5091 		 */
5092 		if (flushparent) {
5093 			int locked;
5094 
5095 			if ((error = ffs_update(pvp, 1)) != 0) {
5096 				vput(pvp);
5097 				return (error);
5098 			}
5099 			ACQUIRE_LOCK(&lk);
5100 			locked = 1;
5101 			if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) {
5102 				if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) {
5103 					if (wk->wk_type != D_DIRADD)
5104 						panic("softdep_fsync: Unexpected type %s",
5105 						      TYPENAME(wk->wk_type));
5106 					dap = WK_DIRADD(wk);
5107 					if (dap->da_state & DIRCHG)
5108 						pagedep = dap->da_previous->dm_pagedep;
5109 					else
5110 						pagedep = dap->da_pagedep;
5111 					pagedep_new_block = pagedep->pd_state & NEWBLOCK;
5112 					FREE_LOCK(&lk);
5113 					locked = 0;
5114 					if (pagedep_new_block &&
5115 					    (error = ffs_syncvnode(pvp, MNT_WAIT))) {
5116 						vput(pvp);
5117 						return (error);
5118 					}
5119 				}
5120 			}
5121 			if (locked)
5122 				FREE_LOCK(&lk);
5123 		}
5124 		/*
5125 		 * Flush directory page containing the inode's name.
5126 		 */
5127 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
5128 		    &bp);
5129 		if (error == 0)
5130 			error = bwrite(bp);
5131 		else
5132 			brelse(bp);
5133 		vput(pvp);
5134 		if (error != 0)
5135 			return (error);
5136 		ACQUIRE_LOCK(&lk);
5137 		if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
5138 			break;
5139 	}
5140 	FREE_LOCK(&lk);
5141 	return (0);
5142 }
5143 
5144 /*
5145  * Flush all the dirty bitmaps associated with the block device
5146  * before flushing the rest of the dirty blocks so as to reduce
5147  * the number of dependencies that will have to be rolled back.
5148  */
5149 void
5150 softdep_fsync_mountdev(vp)
5151 	struct vnode *vp;
5152 {
5153 	struct buf *bp, *nbp;
5154 	struct worklist *wk;
5155 
5156 	if (!vn_isdisk(vp, NULL))
5157 		panic("softdep_fsync_mountdev: vnode not a disk");
5158 restart:
5159 	ACQUIRE_LOCK(&lk);
5160 	VI_LOCK(vp);
5161 	TAILQ_FOREACH_SAFE(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs, nbp) {
5162 		/*
5163 		 * If it is already scheduled, skip to the next buffer.
5164 		 */
5165 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
5166 			continue;
5167 
5168 		if ((bp->b_flags & B_DELWRI) == 0)
5169 			panic("softdep_fsync_mountdev: not dirty");
5170 		/*
5171 		 * We are only interested in bitmaps with outstanding
5172 		 * dependencies.
5173 		 */
5174 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
5175 		    wk->wk_type != D_BMSAFEMAP ||
5176 		    (bp->b_vflags & BV_BKGRDINPROG)) {
5177 			BUF_UNLOCK(bp);
5178 			continue;
5179 		}
5180 		VI_UNLOCK(vp);
5181 		FREE_LOCK(&lk);
5182 		bremfree(bp);
5183 		(void) bawrite(bp);
5184 		goto restart;
5185 	}
5186 	FREE_LOCK(&lk);
5187 	drain_output(vp);
5188 	VI_UNLOCK(vp);
5189 }
5190 
5191 /*
5192  * This routine is called when we are trying to synchronously flush a
5193  * file. This routine must eliminate any filesystem metadata dependencies
5194  * so that the syncing routine can succeed by pushing the dirty blocks
5195  * associated with the file. If any I/O errors occur, they are returned.
5196  */
5197 int
5198 softdep_sync_metadata(struct vnode *vp)
5199 {
5200 	struct pagedep *pagedep;
5201 	struct allocdirect *adp;
5202 	struct allocindir *aip;
5203 	struct buf *bp, *nbp;
5204 	struct worklist *wk;
5205 	int i, error, waitfor;
5206 
5207 	if (!DOINGSOFTDEP(vp))
5208 		return (0);
5209 	/*
5210 	 * Ensure that any direct block dependencies have been cleared.
5211 	 */
5212 	ACQUIRE_LOCK(&lk);
5213 	if ((error = flush_inodedep_deps(vp->v_mount, VTOI(vp)->i_number))) {
5214 		FREE_LOCK(&lk);
5215 		return (error);
5216 	}
5217 	FREE_LOCK(&lk);
5218 	/*
5219 	 * For most files, the only metadata dependencies are the
5220 	 * cylinder group maps that allocate their inode or blocks.
5221 	 * The block allocation dependencies can be found by traversing
5222 	 * the dependency lists for any buffers that remain on their
5223 	 * dirty buffer list. The inode allocation dependency will
5224 	 * be resolved when the inode is updated with MNT_WAIT.
5225 	 * This work is done in two passes. The first pass grabs most
5226 	 * of the buffers and begins asynchronously writing them. The
5227 	 * only way to wait for these asynchronous writes is to sleep
5228 	 * on the filesystem vnode which may stay busy for a long time
5229 	 * if the filesystem is active. So, instead, we make a second
5230 	 * pass over the dependencies blocking on each write. In the
5231 	 * usual case we will be blocking against a write that we
5232 	 * initiated, so when it is done the dependency will have been
5233 	 * resolved. Thus the second pass is expected to end quickly.
5234 	 */
5235 	waitfor = MNT_NOWAIT;
5236 
5237 top:
5238 	/*
5239 	 * We must wait for any I/O in progress to finish so that
5240 	 * all potential buffers on the dirty list will be visible.
5241 	 */
5242 	VI_LOCK(vp);
5243 	drain_output(vp);
5244 	while ((bp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd)) != NULL) {
5245 		bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT);
5246 		if (bp)
5247 			break;
5248 	}
5249 	VI_UNLOCK(vp);
5250 	if (bp == NULL)
5251 		return (0);
5252 loop:
5253 	/* While syncing snapshots, we must allow recursive lookups */
5254 	bp->b_lock.lk_flags |= LK_CANRECURSE;
5255 	ACQUIRE_LOCK(&lk);
5256 	/*
5257 	 * As we hold the buffer locked, none of its dependencies
5258 	 * will disappear.
5259 	 */
5260 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5261 		switch (wk->wk_type) {
5262 
5263 		case D_ALLOCDIRECT:
5264 			adp = WK_ALLOCDIRECT(wk);
5265 			if (adp->ad_state & DEPCOMPLETE)
5266 				continue;
5267 			nbp = adp->ad_buf;
5268 			nbp = getdirtybuf(nbp, &lk, waitfor);
5269 			if (nbp == NULL)
5270 				continue;
5271 			FREE_LOCK(&lk);
5272 			if (waitfor == MNT_NOWAIT) {
5273 				bawrite(nbp);
5274 			} else if ((error = bwrite(nbp)) != 0) {
5275 				break;
5276 			}
5277 			ACQUIRE_LOCK(&lk);
5278 			continue;
5279 
5280 		case D_ALLOCINDIR:
5281 			aip = WK_ALLOCINDIR(wk);
5282 			if (aip->ai_state & DEPCOMPLETE)
5283 				continue;
5284 			nbp = aip->ai_buf;
5285 			nbp = getdirtybuf(nbp, &lk, waitfor);
5286 			if (nbp == NULL)
5287 				continue;
5288 			FREE_LOCK(&lk);
5289 			if (waitfor == MNT_NOWAIT) {
5290 				bawrite(nbp);
5291 			} else if ((error = bwrite(nbp)) != 0) {
5292 				break;
5293 			}
5294 			ACQUIRE_LOCK(&lk);
5295 			continue;
5296 
5297 		case D_INDIRDEP:
5298 		restart:
5299 
5300 			LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
5301 				if (aip->ai_state & DEPCOMPLETE)
5302 					continue;
5303 				nbp = aip->ai_buf;
5304 				nbp = getdirtybuf(nbp, &lk, MNT_WAIT);
5305 				if (nbp == NULL)
5306 					goto restart;
5307 				FREE_LOCK(&lk);
5308 				if ((error = bwrite(nbp)) != 0) {
5309 					goto loop_end;
5310 				}
5311 				ACQUIRE_LOCK(&lk);
5312 				goto restart;
5313 			}
5314 			continue;
5315 
5316 		case D_INODEDEP:
5317 			if ((error = flush_inodedep_deps(wk->wk_mp,
5318 			    WK_INODEDEP(wk)->id_ino)) != 0) {
5319 				FREE_LOCK(&lk);
5320 				break;
5321 			}
5322 			continue;
5323 
5324 		case D_PAGEDEP:
5325 			/*
5326 			 * We are trying to sync a directory that may
5327 			 * have dependencies on both its own metadata
5328 			 * and/or dependencies on the inodes of any
5329 			 * recently allocated files. We walk its diradd
5330 			 * lists pushing out the associated inode.
5331 			 */
5332 			pagedep = WK_PAGEDEP(wk);
5333 			for (i = 0; i < DAHASHSZ; i++) {
5334 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
5335 					continue;
5336 				if ((error =
5337 				    flush_pagedep_deps(vp, wk->wk_mp,
5338 						&pagedep->pd_diraddhd[i]))) {
5339 					FREE_LOCK(&lk);
5340 					goto loop_end;
5341 				}
5342 			}
5343 			continue;
5344 
5345 		case D_MKDIR:
5346 			/*
5347 			 * This case should never happen if the vnode has
5348 			 * been properly sync'ed. However, if this function
5349 			 * is used at a place where the vnode has not yet
5350 			 * been sync'ed, this dependency can show up. So,
5351 			 * rather than panic, just flush it.
5352 			 */
5353 			nbp = WK_MKDIR(wk)->md_buf;
5354 			nbp = getdirtybuf(nbp, &lk, waitfor);
5355 			if (nbp == NULL)
5356 				continue;
5357 			FREE_LOCK(&lk);
5358 			if (waitfor == MNT_NOWAIT) {
5359 				bawrite(nbp);
5360 			} else if ((error = bwrite(nbp)) != 0) {
5361 				break;
5362 			}
5363 			ACQUIRE_LOCK(&lk);
5364 			continue;
5365 
5366 		case D_BMSAFEMAP:
5367 			/*
5368 			 * This case should never happen if the vnode has
5369 			 * been properly sync'ed. However, if this function
5370 			 * is used at a place where the vnode has not yet
5371 			 * been sync'ed, this dependency can show up. So,
5372 			 * rather than panic, just flush it.
5373 			 */
5374 			nbp = WK_BMSAFEMAP(wk)->sm_buf;
5375 			nbp = getdirtybuf(nbp, &lk, waitfor);
5376 			if (nbp == NULL)
5377 				continue;
5378 			FREE_LOCK(&lk);
5379 			if (waitfor == MNT_NOWAIT) {
5380 				bawrite(nbp);
5381 			} else if ((error = bwrite(nbp)) != 0) {
5382 				break;
5383 			}
5384 			ACQUIRE_LOCK(&lk);
5385 			continue;
5386 
5387 		default:
5388 			panic("softdep_sync_metadata: Unknown type %s",
5389 			    TYPENAME(wk->wk_type));
5390 			/* NOTREACHED */
5391 		}
5392 	loop_end:
5393 		/* We reach here only in error and unlocked */
5394 		if (error == 0)
5395 			panic("softdep_sync_metadata: zero error");
5396 		bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5397 		bawrite(bp);
5398 		return (error);
5399 	}
5400 	FREE_LOCK(&lk);
5401 	VI_LOCK(vp);
5402 	while ((nbp = TAILQ_NEXT(bp, b_bobufs)) != NULL) {
5403 		nbp = getdirtybuf(nbp, VI_MTX(vp), MNT_WAIT);
5404 		if (nbp)
5405 			break;
5406 	}
5407 	VI_UNLOCK(vp);
5408 	bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5409 	bawrite(bp);
5410 	if (nbp != NULL) {
5411 		bp = nbp;
5412 		goto loop;
5413 	}
5414 	/*
5415 	 * The brief unlock is to allow any pent up dependency
5416 	 * processing to be done. Then proceed with the second pass.
5417 	 */
5418 	if (waitfor == MNT_NOWAIT) {
5419 		waitfor = MNT_WAIT;
5420 		goto top;
5421 	}
5422 
5423 	/*
5424 	 * If we have managed to get rid of all the dirty buffers,
5425 	 * then we are done. For certain directories and block
5426 	 * devices, we may need to do further work.
5427 	 *
5428 	 * We must wait for any I/O in progress to finish so that
5429 	 * all potential buffers on the dirty list will be visible.
5430 	 */
5431 	VI_LOCK(vp);
5432 	drain_output(vp);
5433 	VI_UNLOCK(vp);
5434 	return (0);
5435 }
5436 
5437 /*
5438  * Flush the dependencies associated with an inodedep.
5439  * Called with splbio blocked.
5440  */
5441 static int
5442 flush_inodedep_deps(mp, ino)
5443 	struct mount *mp;
5444 	ino_t ino;
5445 {
5446 	struct inodedep *inodedep;
5447 	int error, waitfor;
5448 
5449 	/*
5450 	 * This work is done in two passes. The first pass grabs most
5451 	 * of the buffers and begins asynchronously writing them. The
5452 	 * only way to wait for these asynchronous writes is to sleep
5453 	 * on the filesystem vnode which may stay busy for a long time
5454 	 * if the filesystem is active. So, instead, we make a second
5455 	 * pass over the dependencies blocking on each write. In the
5456 	 * usual case we will be blocking against a write that we
5457 	 * initiated, so when it is done the dependency will have been
5458 	 * resolved. Thus the second pass is expected to end quickly.
5459 	 * We give a brief window at the top of the loop to allow
5460 	 * any pending I/O to complete.
5461 	 */
5462 	for (error = 0, waitfor = MNT_NOWAIT; ; ) {
5463 		if (error)
5464 			return (error);
5465 		FREE_LOCK(&lk);
5466 		ACQUIRE_LOCK(&lk);
5467 		if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
5468 			return (0);
5469 		if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
5470 		    flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
5471 		    flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
5472 		    flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
5473 			continue;
5474 		/*
5475 		 * If pass2, we are done, otherwise do pass 2.
5476 		 */
5477 		if (waitfor == MNT_WAIT)
5478 			break;
5479 		waitfor = MNT_WAIT;
5480 	}
5481 	/*
5482 	 * Try freeing inodedep in case all dependencies have been removed.
5483 	 */
5484 	if (inodedep_lookup(mp, ino, 0, &inodedep) != 0)
5485 		(void) free_inodedep(inodedep);
5486 	return (0);
5487 }
5488 
5489 /*
5490  * Flush an inode dependency list.
5491  * Called with splbio blocked.
5492  */
5493 static int
5494 flush_deplist(listhead, waitfor, errorp)
5495 	struct allocdirectlst *listhead;
5496 	int waitfor;
5497 	int *errorp;
5498 {
5499 	struct allocdirect *adp;
5500 	struct buf *bp;
5501 
5502 	mtx_assert(&lk, MA_OWNED);
5503 	TAILQ_FOREACH(adp, listhead, ad_next) {
5504 		if (adp->ad_state & DEPCOMPLETE)
5505 			continue;
5506 		bp = adp->ad_buf;
5507 		bp = getdirtybuf(bp, &lk, waitfor);
5508 		if (bp == NULL) {
5509 			if (waitfor == MNT_NOWAIT)
5510 				continue;
5511 			return (1);
5512 		}
5513 		FREE_LOCK(&lk);
5514 		if (waitfor == MNT_NOWAIT) {
5515 			bawrite(bp);
5516 		} else if ((*errorp = bwrite(bp)) != 0) {
5517 			ACQUIRE_LOCK(&lk);
5518 			return (1);
5519 		}
5520 		ACQUIRE_LOCK(&lk);
5521 		return (1);
5522 	}
5523 	return (0);
5524 }
5525 
5526 /*
5527  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
5528  * Called with splbio blocked.
5529  */
5530 static int
5531 flush_pagedep_deps(pvp, mp, diraddhdp)
5532 	struct vnode *pvp;
5533 	struct mount *mp;
5534 	struct diraddhd *diraddhdp;
5535 {
5536 	struct inodedep *inodedep;
5537 	struct ufsmount *ump;
5538 	struct diradd *dap;
5539 	struct vnode *vp;
5540 	int error = 0;
5541 	struct buf *bp;
5542 	ino_t inum;
5543 	struct worklist *wk;
5544 
5545 	ump = VFSTOUFS(mp);
5546 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
5547 		/*
5548 		 * Flush ourselves if this directory entry
5549 		 * has a MKDIR_PARENT dependency.
5550 		 */
5551 		if (dap->da_state & MKDIR_PARENT) {
5552 			FREE_LOCK(&lk);
5553 			if ((error = ffs_update(pvp, 1)) != 0)
5554 				break;
5555 			ACQUIRE_LOCK(&lk);
5556 			/*
5557 			 * If that cleared dependencies, go on to next.
5558 			 */
5559 			if (dap != LIST_FIRST(diraddhdp))
5560 				continue;
5561 			if (dap->da_state & MKDIR_PARENT)
5562 				panic("flush_pagedep_deps: MKDIR_PARENT");
5563 		}
5564 		/*
5565 		 * A newly allocated directory must have its "." and
5566 		 * ".." entries written out before its name can be
5567 		 * committed in its parent. We do not want or need
5568 		 * the full semantics of a synchronous ffs_syncvnode as
5569 		 * that may end up here again, once for each directory
5570 		 * level in the filesystem. Instead, we push the blocks
5571 		 * and wait for them to clear. We have to fsync twice
5572 		 * because the first call may choose to defer blocks
5573 		 * that still have dependencies, but deferral will
5574 		 * happen at most once.
5575 		 */
5576 		inum = dap->da_newinum;
5577 		if (dap->da_state & MKDIR_BODY) {
5578 			FREE_LOCK(&lk);
5579 			if ((error = ffs_vget(mp, inum, LK_EXCLUSIVE, &vp)))
5580 				break;
5581 			if ((error=ffs_syncvnode(vp, MNT_NOWAIT)) ||
5582 			    (error=ffs_syncvnode(vp, MNT_NOWAIT))) {
5583 				vput(vp);
5584 				break;
5585 			}
5586 			VI_LOCK(vp);
5587 			drain_output(vp);
5588 			/*
5589 			 * If first block is still dirty with a D_MKDIR
5590 			 * dependency then it needs to be written now.
5591 			 */
5592 			for (;;) {
5593 				error = 0;
5594 				bp = gbincore(&vp->v_bufobj, 0);
5595 				if (bp == NULL)
5596 					break;	/* First block not present */
5597 				error = BUF_LOCK(bp,
5598 						 LK_EXCLUSIVE |
5599 						 LK_SLEEPFAIL |
5600 						 LK_INTERLOCK,
5601 						 VI_MTX(vp));
5602 				VI_LOCK(vp);
5603 				if (error == ENOLCK)
5604 					continue;	/* Slept, retry */
5605 				if (error != 0)
5606 					break;		/* Failed */
5607 				if ((bp->b_flags & B_DELWRI) == 0) {
5608 					BUF_UNLOCK(bp);
5609 					break;	/* Buffer not dirty */
5610 				}
5611 				for (wk = LIST_FIRST(&bp->b_dep);
5612 				     wk != NULL;
5613 				     wk = LIST_NEXT(wk, wk_list))
5614 					if (wk->wk_type == D_MKDIR)
5615 						break;
5616 				if (wk == NULL)
5617 					BUF_UNLOCK(bp);	/* Dependency gone */
5618 				else {
5619 					/*
5620 					 * D_MKDIR dependency remains,
5621 					 * must write buffer to stable
5622 					 * storage.
5623 					 */
5624 					VI_UNLOCK(vp);
5625 					bremfree(bp);
5626 					error = bwrite(bp);
5627 					VI_LOCK(vp);
5628 				}
5629 				break;
5630 			}
5631 			VI_UNLOCK(vp);
5632 			vput(vp);
5633 			if (error != 0)
5634 				break;	/* Flushing of first block failed */
5635 			ACQUIRE_LOCK(&lk);
5636 			/*
5637 			 * If that cleared dependencies, go on to next.
5638 			 */
5639 			if (dap != LIST_FIRST(diraddhdp))
5640 				continue;
5641 			if (dap->da_state & MKDIR_BODY)
5642 				panic("flush_pagedep_deps: MKDIR_BODY");
5643 		}
5644 		/*
5645 		 * Flush the inode on which the directory entry depends.
5646 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
5647 		 * the only remaining dependency is that the updated inode
5648 		 * count must get pushed to disk. The inode has already
5649 		 * been pushed into its inode buffer (via VOP_UPDATE) at
5650 		 * the time of the reference count change. So we need only
5651 		 * locate that buffer, ensure that there will be no rollback
5652 		 * caused by a bitmap dependency, then write the inode buffer.
5653 		 */
5654 retry:
5655 		if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
5656 			panic("flush_pagedep_deps: lost inode");
5657 		/*
5658 		 * If the inode still has bitmap dependencies,
5659 		 * push them to disk.
5660 		 */
5661 		if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5662 			bp = inodedep->id_buf;
5663 			bp = getdirtybuf(bp, &lk, MNT_WAIT);
5664 			if (bp == NULL)
5665 				goto retry;
5666 			FREE_LOCK(&lk);
5667 			if ((error = bwrite(bp)) != 0)
5668 				break;
5669 			ACQUIRE_LOCK(&lk);
5670 			if (dap != LIST_FIRST(diraddhdp))
5671 				continue;
5672 		}
5673 		/*
5674 		 * If the inode is still sitting in a buffer waiting
5675 		 * to be written, push it to disk.
5676 		 */
5677 		FREE_LOCK(&lk);
5678 		if ((error = bread(ump->um_devvp,
5679 		    fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
5680 		    (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) {
5681 			brelse(bp);
5682 			break;
5683 		}
5684 		if ((error = bwrite(bp)) != 0)
5685 			break;
5686 		ACQUIRE_LOCK(&lk);
5687 		/*
5688 		 * If we have failed to get rid of all the dependencies
5689 		 * then something is seriously wrong.
5690 		 */
5691 		if (dap == LIST_FIRST(diraddhdp))
5692 			panic("flush_pagedep_deps: flush failed");
5693 	}
5694 	if (error)
5695 		ACQUIRE_LOCK(&lk);
5696 	return (error);
5697 }
5698 
5699 /*
5700  * A large burst of file addition or deletion activity can drive the
5701  * memory load excessively high. First attempt to slow things down
5702  * using the techniques below. If that fails, this routine requests
5703  * the offending operations to fall back to running synchronously
5704  * until the memory load returns to a reasonable level.
5705  */
5706 int
5707 softdep_slowdown(vp)
5708 	struct vnode *vp;
5709 {
5710 	int max_softdeps_hard;
5711 
5712 	ACQUIRE_LOCK(&lk);
5713 	max_softdeps_hard = max_softdeps * 11 / 10;
5714 	if (num_dirrem < max_softdeps_hard / 2 &&
5715 	    num_inodedep < max_softdeps_hard &&
5716 	    VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps) {
5717 		FREE_LOCK(&lk);
5718   		return (0);
5719 	}
5720 	if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps)
5721 		softdep_speedup();
5722 	stat_sync_limit_hit += 1;
5723 	FREE_LOCK(&lk);
5724 	return (1);
5725 }
5726 
5727 /*
5728  * Called by the allocation routines when they are about to fail
5729  * in the hope that we can free up some disk space.
5730  *
5731  * First check to see if the work list has anything on it. If it has,
5732  * clean up entries until we successfully free some space. Because this
5733  * process holds inodes locked, we cannot handle any remove requests
5734  * that might block on a locked inode as that could lead to deadlock.
5735  * If the worklist yields no free space, encourage the syncer daemon
5736  * to help us. In no event will we try for longer than tickdelay seconds.
5737  */
5738 int
5739 softdep_request_cleanup(fs, vp)
5740 	struct fs *fs;
5741 	struct vnode *vp;
5742 {
5743 	struct ufsmount *ump;
5744 	long starttime;
5745 	ufs2_daddr_t needed;
5746 	int error;
5747 
5748 	ump = VTOI(vp)->i_ump;
5749 	mtx_assert(UFS_MTX(ump), MA_OWNED);
5750 	needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize;
5751 	starttime = time_second + tickdelay;
5752 	/*
5753 	 * If we are being called because of a process doing a
5754 	 * copy-on-write, then it is not safe to update the vnode
5755 	 * as we may recurse into the copy-on-write routine.
5756 	 */
5757 	if (!(curthread->td_pflags & TDP_COWINPROGRESS)) {
5758 		UFS_UNLOCK(ump);
5759 		error = ffs_update(vp, 1);
5760 		UFS_LOCK(ump);
5761 		if (error != 0)
5762 			return (0);
5763 	}
5764 	while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) {
5765 		if (time_second > starttime)
5766 			return (0);
5767 		UFS_UNLOCK(ump);
5768 		ACQUIRE_LOCK(&lk);
5769 		if (ump->softdep_on_worklist > 0 &&
5770 		    process_worklist_item(UFSTOVFS(ump), LK_NOWAIT) != -1) {
5771 			stat_worklist_push += 1;
5772 			FREE_LOCK(&lk);
5773 			UFS_LOCK(ump);
5774 			continue;
5775 		}
5776 		request_cleanup(UFSTOVFS(ump), FLUSH_REMOVE_WAIT);
5777 		FREE_LOCK(&lk);
5778 		UFS_LOCK(ump);
5779 	}
5780 	return (1);
5781 }
5782 
5783 /*
5784  * If memory utilization has gotten too high, deliberately slow things
5785  * down and speed up the I/O processing.
5786  */
5787 extern struct thread *syncertd;
5788 static int
5789 request_cleanup(mp, resource)
5790 	struct mount *mp;
5791 	int resource;
5792 {
5793 	struct thread *td = curthread;
5794 	struct ufsmount *ump;
5795 
5796 	mtx_assert(&lk, MA_OWNED);
5797 	/*
5798 	 * We never hold up the filesystem syncer or buf daemon.
5799 	 */
5800 	if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF))
5801 		return (0);
5802 	ump = VFSTOUFS(mp);
5803 	/*
5804 	 * First check to see if the work list has gotten backlogged.
5805 	 * If it has, co-opt this process to help clean up two entries.
5806 	 * Because this process may hold inodes locked, we cannot
5807 	 * handle any remove requests that might block on a locked
5808 	 * inode as that could lead to deadlock.  We set TDP_SOFTDEP
5809 	 * to avoid recursively processing the worklist.
5810 	 */
5811 	if (ump->softdep_on_worklist > max_softdeps / 10) {
5812 		td->td_pflags |= TDP_SOFTDEP;
5813 		process_worklist_item(mp, LK_NOWAIT);
5814 		process_worklist_item(mp, LK_NOWAIT);
5815 		td->td_pflags &= ~TDP_SOFTDEP;
5816 		stat_worklist_push += 2;
5817 		return(1);
5818 	}
5819 	/*
5820 	 * Next, we attempt to speed up the syncer process. If that
5821 	 * is successful, then we allow the process to continue.
5822 	 */
5823 	if (softdep_speedup() && resource != FLUSH_REMOVE_WAIT)
5824 		return(0);
5825 	/*
5826 	 * If we are resource constrained on inode dependencies, try
5827 	 * flushing some dirty inodes. Otherwise, we are constrained
5828 	 * by file deletions, so try accelerating flushes of directories
5829 	 * with removal dependencies. We would like to do the cleanup
5830 	 * here, but we probably hold an inode locked at this point and
5831 	 * that might deadlock against one that we try to clean. So,
5832 	 * the best that we can do is request the syncer daemon to do
5833 	 * the cleanup for us.
5834 	 */
5835 	switch (resource) {
5836 
5837 	case FLUSH_INODES:
5838 		stat_ino_limit_push += 1;
5839 		req_clear_inodedeps += 1;
5840 		stat_countp = &stat_ino_limit_hit;
5841 		break;
5842 
5843 	case FLUSH_REMOVE:
5844 	case FLUSH_REMOVE_WAIT:
5845 		stat_blk_limit_push += 1;
5846 		req_clear_remove += 1;
5847 		stat_countp = &stat_blk_limit_hit;
5848 		break;
5849 
5850 	default:
5851 		panic("request_cleanup: unknown type");
5852 	}
5853 	/*
5854 	 * Hopefully the syncer daemon will catch up and awaken us.
5855 	 * We wait at most tickdelay before proceeding in any case.
5856 	 */
5857 	proc_waiting += 1;
5858 	if (handle.callout == NULL)
5859 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5860 	msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0);
5861 	proc_waiting -= 1;
5862 	return (1);
5863 }
5864 
5865 /*
5866  * Awaken processes pausing in request_cleanup and clear proc_waiting
5867  * to indicate that there is no longer a timer running.
5868  */
5869 static void
5870 pause_timer(arg)
5871 	void *arg;
5872 {
5873 
5874 	ACQUIRE_LOCK(&lk);
5875 	*stat_countp += 1;
5876 	wakeup_one(&proc_waiting);
5877 	if (proc_waiting > 0)
5878 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5879 	else
5880 		handle.callout = NULL;
5881 	FREE_LOCK(&lk);
5882 }
5883 
5884 /*
5885  * Flush out a directory with at least one removal dependency in an effort to
5886  * reduce the number of dirrem, freefile, and freeblks dependency structures.
5887  */
5888 static void
5889 clear_remove(td)
5890 	struct thread *td;
5891 {
5892 	struct pagedep_hashhead *pagedephd;
5893 	struct pagedep *pagedep;
5894 	static int next = 0;
5895 	struct mount *mp;
5896 	struct vnode *vp;
5897 	int error, cnt;
5898 	ino_t ino;
5899 
5900 	mtx_assert(&lk, MA_OWNED);
5901 
5902 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
5903 		pagedephd = &pagedep_hashtbl[next++];
5904 		if (next >= pagedep_hash)
5905 			next = 0;
5906 		LIST_FOREACH(pagedep, pagedephd, pd_hash) {
5907 			if (LIST_EMPTY(&pagedep->pd_dirremhd))
5908 				continue;
5909 			mp = pagedep->pd_list.wk_mp;
5910 			ino = pagedep->pd_ino;
5911 			if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5912 				continue;
5913 			FREE_LOCK(&lk);
5914 			if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp))) {
5915 				softdep_error("clear_remove: vget", error);
5916 				vn_finished_write(mp);
5917 				ACQUIRE_LOCK(&lk);
5918 				return;
5919 			}
5920 			if ((error = ffs_syncvnode(vp, MNT_NOWAIT)))
5921 				softdep_error("clear_remove: fsync", error);
5922 			VI_LOCK(vp);
5923 			drain_output(vp);
5924 			VI_UNLOCK(vp);
5925 			vput(vp);
5926 			vn_finished_write(mp);
5927 			ACQUIRE_LOCK(&lk);
5928 			return;
5929 		}
5930 	}
5931 }
5932 
5933 /*
5934  * Clear out a block of dirty inodes in an effort to reduce
5935  * the number of inodedep dependency structures.
5936  */
5937 static void
5938 clear_inodedeps(td)
5939 	struct thread *td;
5940 {
5941 	struct inodedep_hashhead *inodedephd;
5942 	struct inodedep *inodedep;
5943 	static int next = 0;
5944 	struct mount *mp;
5945 	struct vnode *vp;
5946 	struct fs *fs;
5947 	int error, cnt;
5948 	ino_t firstino, lastino, ino;
5949 
5950 	mtx_assert(&lk, MA_OWNED);
5951 	/*
5952 	 * Pick a random inode dependency to be cleared.
5953 	 * We will then gather up all the inodes in its block
5954 	 * that have dependencies and flush them out.
5955 	 */
5956 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
5957 		inodedephd = &inodedep_hashtbl[next++];
5958 		if (next >= inodedep_hash)
5959 			next = 0;
5960 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
5961 			break;
5962 	}
5963 	if (inodedep == NULL)
5964 		return;
5965 	fs = inodedep->id_fs;
5966 	mp = inodedep->id_list.wk_mp;
5967 	/*
5968 	 * Find the last inode in the block with dependencies.
5969 	 */
5970 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
5971 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
5972 		if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0)
5973 			break;
5974 	/*
5975 	 * Asynchronously push all but the last inode with dependencies.
5976 	 * Synchronously push the last inode with dependencies to ensure
5977 	 * that the inode block gets written to free up the inodedeps.
5978 	 */
5979 	for (ino = firstino; ino <= lastino; ino++) {
5980 		if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
5981 			continue;
5982 		if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5983 			continue;
5984 		FREE_LOCK(&lk);
5985 		if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp)) != 0) {
5986 			softdep_error("clear_inodedeps: vget", error);
5987 			vn_finished_write(mp);
5988 			ACQUIRE_LOCK(&lk);
5989 			return;
5990 		}
5991 		if (ino == lastino) {
5992 			if ((error = ffs_syncvnode(vp, MNT_WAIT)))
5993 				softdep_error("clear_inodedeps: fsync1", error);
5994 		} else {
5995 			if ((error = ffs_syncvnode(vp, MNT_NOWAIT)))
5996 				softdep_error("clear_inodedeps: fsync2", error);
5997 			VI_LOCK(vp);
5998 			drain_output(vp);
5999 			VI_UNLOCK(vp);
6000 		}
6001 		vput(vp);
6002 		vn_finished_write(mp);
6003 		ACQUIRE_LOCK(&lk);
6004 	}
6005 }
6006 
6007 /*
6008  * Function to determine if the buffer has outstanding dependencies
6009  * that will cause a roll-back if the buffer is written. If wantcount
6010  * is set, return number of dependencies, otherwise just yes or no.
6011  */
6012 static int
6013 softdep_count_dependencies(bp, wantcount)
6014 	struct buf *bp;
6015 	int wantcount;
6016 {
6017 	struct worklist *wk;
6018 	struct inodedep *inodedep;
6019 	struct indirdep *indirdep;
6020 	struct allocindir *aip;
6021 	struct pagedep *pagedep;
6022 	struct diradd *dap;
6023 	int i, retval;
6024 
6025 	retval = 0;
6026 	ACQUIRE_LOCK(&lk);
6027 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
6028 		switch (wk->wk_type) {
6029 
6030 		case D_INODEDEP:
6031 			inodedep = WK_INODEDEP(wk);
6032 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
6033 				/* bitmap allocation dependency */
6034 				retval += 1;
6035 				if (!wantcount)
6036 					goto out;
6037 			}
6038 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
6039 				/* direct block pointer dependency */
6040 				retval += 1;
6041 				if (!wantcount)
6042 					goto out;
6043 			}
6044 			if (TAILQ_FIRST(&inodedep->id_extupdt)) {
6045 				/* direct block pointer dependency */
6046 				retval += 1;
6047 				if (!wantcount)
6048 					goto out;
6049 			}
6050 			continue;
6051 
6052 		case D_INDIRDEP:
6053 			indirdep = WK_INDIRDEP(wk);
6054 
6055 			LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
6056 				/* indirect block pointer dependency */
6057 				retval += 1;
6058 				if (!wantcount)
6059 					goto out;
6060 			}
6061 			continue;
6062 
6063 		case D_PAGEDEP:
6064 			pagedep = WK_PAGEDEP(wk);
6065 			for (i = 0; i < DAHASHSZ; i++) {
6066 
6067 				LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
6068 					/* directory entry dependency */
6069 					retval += 1;
6070 					if (!wantcount)
6071 						goto out;
6072 				}
6073 			}
6074 			continue;
6075 
6076 		case D_BMSAFEMAP:
6077 		case D_ALLOCDIRECT:
6078 		case D_ALLOCINDIR:
6079 		case D_MKDIR:
6080 			/* never a dependency on these blocks */
6081 			continue;
6082 
6083 		default:
6084 			panic("softdep_check_for_rollback: Unexpected type %s",
6085 			    TYPENAME(wk->wk_type));
6086 			/* NOTREACHED */
6087 		}
6088 	}
6089 out:
6090 	FREE_LOCK(&lk);
6091 	return retval;
6092 }
6093 
6094 /*
6095  * Acquire exclusive access to a buffer.
6096  * Must be called with a locked mtx parameter.
6097  * Return acquired buffer or NULL on failure.
6098  */
6099 static struct buf *
6100 getdirtybuf(bp, mtx, waitfor)
6101 	struct buf *bp;
6102 	struct mtx *mtx;
6103 	int waitfor;
6104 {
6105 	int error;
6106 
6107 	mtx_assert(mtx, MA_OWNED);
6108 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) {
6109 		if (waitfor != MNT_WAIT)
6110 			return (NULL);
6111 		error = BUF_LOCK(bp,
6112 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx);
6113 		/*
6114 		 * Even if we sucessfully acquire bp here, we have dropped
6115 		 * mtx, which may violates our guarantee.
6116 		 */
6117 		if (error == 0)
6118 			BUF_UNLOCK(bp);
6119 		else if (error != ENOLCK)
6120 			panic("getdirtybuf: inconsistent lock: %d", error);
6121 		mtx_lock(mtx);
6122 		return (NULL);
6123 	}
6124 	if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
6125 		if (mtx == &lk && waitfor == MNT_WAIT) {
6126 			mtx_unlock(mtx);
6127 			BO_LOCK(bp->b_bufobj);
6128 			BUF_UNLOCK(bp);
6129 			if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
6130 				bp->b_vflags |= BV_BKGRDWAIT;
6131 				msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj),
6132 				       PRIBIO | PDROP, "getbuf", 0);
6133 			} else
6134 				BO_UNLOCK(bp->b_bufobj);
6135 			mtx_lock(mtx);
6136 			return (NULL);
6137 		}
6138 		BUF_UNLOCK(bp);
6139 		if (waitfor != MNT_WAIT)
6140 			return (NULL);
6141 		/*
6142 		 * The mtx argument must be bp->b_vp's mutex in
6143 		 * this case.
6144 		 */
6145 #ifdef	DEBUG_VFS_LOCKS
6146 		if (bp->b_vp->v_type != VCHR)
6147 			ASSERT_VI_LOCKED(bp->b_vp, "getdirtybuf");
6148 #endif
6149 		bp->b_vflags |= BV_BKGRDWAIT;
6150 		msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0);
6151 		return (NULL);
6152 	}
6153 	if ((bp->b_flags & B_DELWRI) == 0) {
6154 		BUF_UNLOCK(bp);
6155 		return (NULL);
6156 	}
6157 	bremfree(bp);
6158 	return (bp);
6159 }
6160 
6161 
6162 /*
6163  * Check if it is safe to suspend the file system now.  On entry,
6164  * the vnode interlock for devvp should be held.  Return 0 with
6165  * the mount interlock held if the file system can be suspended now,
6166  * otherwise return EAGAIN with the mount interlock held.
6167  */
6168 int
6169 softdep_check_suspend(struct mount *mp,
6170 		      struct vnode *devvp,
6171 		      int softdep_deps,
6172 		      int softdep_accdeps,
6173 		      int secondary_writes,
6174 		      int secondary_accwrites)
6175 {
6176 	struct bufobj *bo;
6177 	struct ufsmount *ump;
6178 	int error;
6179 
6180 	ASSERT_VI_LOCKED(devvp, "softdep_check_suspend");
6181 	ump = VFSTOUFS(mp);
6182 	bo = &devvp->v_bufobj;
6183 
6184 	for (;;) {
6185 		if (!TRY_ACQUIRE_LOCK(&lk)) {
6186 			VI_UNLOCK(devvp);
6187 			ACQUIRE_LOCK(&lk);
6188 			FREE_LOCK(&lk);
6189 			VI_LOCK(devvp);
6190 			continue;
6191 		}
6192 		if (!MNT_ITRYLOCK(mp)) {
6193 			FREE_LOCK(&lk);
6194 			VI_UNLOCK(devvp);
6195 			MNT_ILOCK(mp);
6196 			MNT_IUNLOCK(mp);
6197 			VI_LOCK(devvp);
6198 			continue;
6199 		}
6200 		if (mp->mnt_secondary_writes != 0) {
6201 			FREE_LOCK(&lk);
6202 			VI_UNLOCK(devvp);
6203 			msleep(&mp->mnt_secondary_writes,
6204 			       MNT_MTX(mp),
6205 			       (PUSER - 1) | PDROP, "secwr", 0);
6206 			VI_LOCK(devvp);
6207 			continue;
6208 		}
6209 		break;
6210 	}
6211 
6212 	/*
6213 	 * Reasons for needing more work before suspend:
6214 	 * - Dirty buffers on devvp.
6215 	 * - Softdep activity occurred after start of vnode sync loop
6216 	 * - Secondary writes occurred after start of vnode sync loop
6217 	 */
6218 	error = 0;
6219 	if (bo->bo_numoutput > 0 ||
6220 	    bo->bo_dirty.bv_cnt > 0 ||
6221 	    softdep_deps != 0 ||
6222 	    ump->softdep_deps != 0 ||
6223 	    softdep_accdeps != ump->softdep_accdeps ||
6224 	    secondary_writes != 0 ||
6225 	    mp->mnt_secondary_writes != 0 ||
6226 	    secondary_accwrites != mp->mnt_secondary_accwrites)
6227 		error = EAGAIN;
6228 	FREE_LOCK(&lk);
6229 	VI_UNLOCK(devvp);
6230 	return (error);
6231 }
6232 
6233 
6234 /*
6235  * Get the number of dependency structures for the file system, both
6236  * the current number and the total number allocated.  These will
6237  * later be used to detect that softdep processing has occurred.
6238  */
6239 void
6240 softdep_get_depcounts(struct mount *mp,
6241 		      int *softdep_depsp,
6242 		      int *softdep_accdepsp)
6243 {
6244 	struct ufsmount *ump;
6245 
6246 	ump = VFSTOUFS(mp);
6247 	ACQUIRE_LOCK(&lk);
6248 	*softdep_depsp = ump->softdep_deps;
6249 	*softdep_accdepsp = ump->softdep_accdeps;
6250 	FREE_LOCK(&lk);
6251 }
6252 
6253 /*
6254  * Wait for pending output on a vnode to complete.
6255  * Must be called with vnode lock and interlock locked.
6256  *
6257  * XXX: Should just be a call to bufobj_wwait().
6258  */
6259 static void
6260 drain_output(vp)
6261 	struct vnode *vp;
6262 {
6263 	ASSERT_VOP_LOCKED(vp, "drain_output");
6264 	ASSERT_VI_LOCKED(vp, "drain_output");
6265 
6266 	while (vp->v_bufobj.bo_numoutput) {
6267 		vp->v_bufobj.bo_flag |= BO_WWAIT;
6268 		msleep((caddr_t)&vp->v_bufobj.bo_numoutput,
6269 		    VI_MTX(vp), PRIBIO + 1, "drainvp", 0);
6270 	}
6271 }
6272 
6273 /*
6274  * Called whenever a buffer that is being invalidated or reallocated
6275  * contains dependencies. This should only happen if an I/O error has
6276  * occurred. The routine is called with the buffer locked.
6277  */
6278 static void
6279 softdep_deallocate_dependencies(bp)
6280 	struct buf *bp;
6281 {
6282 
6283 	if ((bp->b_ioflags & BIO_ERROR) == 0)
6284 		panic("softdep_deallocate_dependencies: dangling deps");
6285 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
6286 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
6287 }
6288 
6289 /*
6290  * Function to handle asynchronous write errors in the filesystem.
6291  */
6292 static void
6293 softdep_error(func, error)
6294 	char *func;
6295 	int error;
6296 {
6297 
6298 	/* XXX should do something better! */
6299 	printf("%s: got error %d while accessing filesystem\n", func, error);
6300 }
6301 
6302 #endif /* SOFTUPDATES */
6303