xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision b28624fde638caadd4a89f50c9b7e7da0f98c4d2)
1 /*-
2  * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
3  *
4  * The soft updates code is derived from the appendix of a University
5  * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
6  * "Soft Updates: A Solution to the Metadata Update Problem in File
7  * Systems", CSE-TR-254-95, August 1995).
8  *
9  * Further information about soft updates can be obtained from:
10  *
11  *	Marshall Kirk McKusick		http://www.mckusick.com/softdep/
12  *	1614 Oxford Street		mckusick@mckusick.com
13  *	Berkeley, CA 94709-1608		+1-510-843-9542
14  *	USA
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  *
20  * 1. Redistributions of source code must retain the above copyright
21  *    notice, this list of conditions and the following disclaimer.
22  * 2. Redistributions in binary form must reproduce the above copyright
23  *    notice, this list of conditions and the following disclaimer in the
24  *    documentation and/or other materials provided with the distribution.
25  *
26  * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
27  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
28  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
29  * DISCLAIMED.  IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
30  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)ffs_softdep.c	9.59 (McKusick) 6/21/00
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 /*
45  * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
46  */
47 #ifndef DIAGNOSTIC
48 #define DIAGNOSTIC
49 #endif
50 #ifndef DEBUG
51 #define DEBUG
52 #endif
53 
54 #include <sys/param.h>
55 #include <sys/kernel.h>
56 #include <sys/systm.h>
57 #include <sys/bio.h>
58 #include <sys/buf.h>
59 #include <sys/kdb.h>
60 #include <sys/kthread.h>
61 #include <sys/lock.h>
62 #include <sys/malloc.h>
63 #include <sys/mount.h>
64 #include <sys/mutex.h>
65 #include <sys/proc.h>
66 #include <sys/stat.h>
67 #include <sys/sysctl.h>
68 #include <sys/syslog.h>
69 #include <sys/vnode.h>
70 #include <sys/conf.h>
71 #include <ufs/ufs/dir.h>
72 #include <ufs/ufs/extattr.h>
73 #include <ufs/ufs/quota.h>
74 #include <ufs/ufs/inode.h>
75 #include <ufs/ufs/ufsmount.h>
76 #include <ufs/ffs/fs.h>
77 #include <ufs/ffs/softdep.h>
78 #include <ufs/ffs/ffs_extern.h>
79 #include <ufs/ufs/ufs_extern.h>
80 
81 #include <vm/vm.h>
82 
83 #include "opt_ffs.h"
84 #include "opt_quota.h"
85 
86 #ifndef SOFTUPDATES
87 
88 int
89 softdep_flushfiles(oldmnt, flags, td)
90 	struct mount *oldmnt;
91 	int flags;
92 	struct thread *td;
93 {
94 
95 	panic("softdep_flushfiles called");
96 }
97 
98 int
99 softdep_mount(devvp, mp, fs, cred)
100 	struct vnode *devvp;
101 	struct mount *mp;
102 	struct fs *fs;
103 	struct ucred *cred;
104 {
105 
106 	return (0);
107 }
108 
109 void
110 softdep_initialize()
111 {
112 
113 	return;
114 }
115 
116 void
117 softdep_uninitialize()
118 {
119 
120 	return;
121 }
122 
123 void
124 softdep_setup_inomapdep(bp, ip, newinum)
125 	struct buf *bp;
126 	struct inode *ip;
127 	ino_t newinum;
128 {
129 
130 	panic("softdep_setup_inomapdep called");
131 }
132 
133 void
134 softdep_setup_blkmapdep(bp, mp, newblkno)
135 	struct buf *bp;
136 	struct mount *mp;
137 	ufs2_daddr_t newblkno;
138 {
139 
140 	panic("softdep_setup_blkmapdep called");
141 }
142 
143 void
144 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
145 	struct inode *ip;
146 	ufs_lbn_t lbn;
147 	ufs2_daddr_t newblkno;
148 	ufs2_daddr_t oldblkno;
149 	long newsize;
150 	long oldsize;
151 	struct buf *bp;
152 {
153 
154 	panic("softdep_setup_allocdirect called");
155 }
156 
157 void
158 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
159 	struct inode *ip;
160 	ufs_lbn_t lbn;
161 	ufs2_daddr_t newblkno;
162 	ufs2_daddr_t oldblkno;
163 	long newsize;
164 	long oldsize;
165 	struct buf *bp;
166 {
167 
168 	panic("softdep_setup_allocext called");
169 }
170 
171 void
172 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
173 	struct inode *ip;
174 	ufs_lbn_t lbn;
175 	struct buf *bp;
176 	int ptrno;
177 	ufs2_daddr_t newblkno;
178 	ufs2_daddr_t oldblkno;
179 	struct buf *nbp;
180 {
181 
182 	panic("softdep_setup_allocindir_page called");
183 }
184 
185 void
186 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
187 	struct buf *nbp;
188 	struct inode *ip;
189 	struct buf *bp;
190 	int ptrno;
191 	ufs2_daddr_t newblkno;
192 {
193 
194 	panic("softdep_setup_allocindir_meta called");
195 }
196 
197 void
198 softdep_setup_freeblocks(ip, length, flags)
199 	struct inode *ip;
200 	off_t length;
201 	int flags;
202 {
203 
204 	panic("softdep_setup_freeblocks called");
205 }
206 
207 void
208 softdep_freefile(pvp, ino, mode)
209 		struct vnode *pvp;
210 		ino_t ino;
211 		int mode;
212 {
213 
214 	panic("softdep_freefile called");
215 }
216 
217 int
218 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
219 	struct buf *bp;
220 	struct inode *dp;
221 	off_t diroffset;
222 	ino_t newinum;
223 	struct buf *newdirbp;
224 	int isnewblk;
225 {
226 
227 	panic("softdep_setup_directory_add called");
228 }
229 
230 void
231 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
232 	struct inode *dp;
233 	caddr_t base;
234 	caddr_t oldloc;
235 	caddr_t newloc;
236 	int entrysize;
237 {
238 
239 	panic("softdep_change_directoryentry_offset called");
240 }
241 
242 void
243 softdep_setup_remove(bp, dp, ip, isrmdir)
244 	struct buf *bp;
245 	struct inode *dp;
246 	struct inode *ip;
247 	int isrmdir;
248 {
249 
250 	panic("softdep_setup_remove called");
251 }
252 
253 void
254 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
255 	struct buf *bp;
256 	struct inode *dp;
257 	struct inode *ip;
258 	ino_t newinum;
259 	int isrmdir;
260 {
261 
262 	panic("softdep_setup_directory_change called");
263 }
264 
265 void
266 softdep_change_linkcnt(ip)
267 	struct inode *ip;
268 {
269 
270 	panic("softdep_change_linkcnt called");
271 }
272 
273 void
274 softdep_load_inodeblock(ip)
275 	struct inode *ip;
276 {
277 
278 	panic("softdep_load_inodeblock called");
279 }
280 
281 void
282 softdep_update_inodeblock(ip, bp, waitfor)
283 	struct inode *ip;
284 	struct buf *bp;
285 	int waitfor;
286 {
287 
288 	panic("softdep_update_inodeblock called");
289 }
290 
291 int
292 softdep_fsync(vp)
293 	struct vnode *vp;	/* the "in_core" copy of the inode */
294 {
295 
296 	return (0);
297 }
298 
299 void
300 softdep_fsync_mountdev(vp)
301 	struct vnode *vp;
302 {
303 
304 	return;
305 }
306 
307 int
308 softdep_flushworklist(oldmnt, countp, td)
309 	struct mount *oldmnt;
310 	int *countp;
311 	struct thread *td;
312 {
313 
314 	*countp = 0;
315 	return (0);
316 }
317 
318 int
319 softdep_sync_metadata(struct vnode *vp)
320 {
321 
322 	return (0);
323 }
324 
325 int
326 softdep_slowdown(vp)
327 	struct vnode *vp;
328 {
329 
330 	panic("softdep_slowdown called");
331 }
332 
333 void
334 softdep_releasefile(ip)
335 	struct inode *ip;	/* inode with the zero effective link count */
336 {
337 
338 	panic("softdep_releasefile called");
339 }
340 
341 int
342 softdep_request_cleanup(fs, vp)
343 	struct fs *fs;
344 	struct vnode *vp;
345 {
346 
347 	return (0);
348 }
349 
350 int
351 softdep_check_suspend(struct mount *mp,
352 		      struct vnode *devvp,
353 		      int softdep_deps,
354 		      int softdep_accdeps,
355 		      int secondary_writes,
356 		      int secondary_accwrites)
357 {
358 	struct bufobj *bo;
359 	int error;
360 
361 	(void) softdep_deps,
362 	(void) softdep_accdeps;
363 
364 	ASSERT_VI_LOCKED(devvp, "softdep_check_suspend");
365 	bo = &devvp->v_bufobj;
366 
367 	for (;;) {
368 		if (!MNT_ITRYLOCK(mp)) {
369 			VI_UNLOCK(devvp);
370 			MNT_ILOCK(mp);
371 			MNT_IUNLOCK(mp);
372 			VI_LOCK(devvp);
373 			continue;
374 		}
375 		if (mp->mnt_secondary_writes != 0) {
376 			VI_UNLOCK(devvp);
377 			msleep(&mp->mnt_secondary_writes,
378 			       MNT_MTX(mp),
379 			       (PUSER - 1) | PDROP, "secwr", 0);
380 			VI_LOCK(devvp);
381 			continue;
382 		}
383 		break;
384 	}
385 
386 	/*
387 	 * Reasons for needing more work before suspend:
388 	 * - Dirty buffers on devvp.
389 	 * - Secondary writes occurred after start of vnode sync loop
390 	 */
391 	error = 0;
392 	if (bo->bo_numoutput > 0 ||
393 	    bo->bo_dirty.bv_cnt > 0 ||
394 	    secondary_writes != 0 ||
395 	    mp->mnt_secondary_writes != 0 ||
396 	    secondary_accwrites != mp->mnt_secondary_accwrites)
397 		error = EAGAIN;
398 	VI_UNLOCK(devvp);
399 	return (error);
400 }
401 
402 void
403 softdep_get_depcounts(struct mount *mp,
404 		      int *softdepactivep,
405 		      int *softdepactiveaccp)
406 {
407 	(void) mp;
408 	*softdepactivep = 0;
409 	*softdepactiveaccp = 0;
410 }
411 
412 #else
413 /*
414  * These definitions need to be adapted to the system to which
415  * this file is being ported.
416  */
417 /*
418  * malloc types defined for the softdep system.
419  */
420 static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
421 static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
422 static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
423 static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
424 static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
425 static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
426 static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
427 static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
428 static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
429 static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
430 static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
431 static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
432 static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
433 static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block");
434 static MALLOC_DEFINE(M_SAVEDINO, "savedino","Saved inodes");
435 
436 #define M_SOFTDEP_FLAGS	(M_WAITOK | M_USE_RESERVE)
437 
438 #define	D_PAGEDEP	0
439 #define	D_INODEDEP	1
440 #define	D_NEWBLK	2
441 #define	D_BMSAFEMAP	3
442 #define	D_ALLOCDIRECT	4
443 #define	D_INDIRDEP	5
444 #define	D_ALLOCINDIR	6
445 #define	D_FREEFRAG	7
446 #define	D_FREEBLKS	8
447 #define	D_FREEFILE	9
448 #define	D_DIRADD	10
449 #define	D_MKDIR		11
450 #define	D_DIRREM	12
451 #define	D_NEWDIRBLK	13
452 #define	D_LAST		D_NEWDIRBLK
453 
454 /*
455  * translate from workitem type to memory type
456  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
457  */
458 static struct malloc_type *memtype[] = {
459 	M_PAGEDEP,
460 	M_INODEDEP,
461 	M_NEWBLK,
462 	M_BMSAFEMAP,
463 	M_ALLOCDIRECT,
464 	M_INDIRDEP,
465 	M_ALLOCINDIR,
466 	M_FREEFRAG,
467 	M_FREEBLKS,
468 	M_FREEFILE,
469 	M_DIRADD,
470 	M_MKDIR,
471 	M_DIRREM,
472 	M_NEWDIRBLK
473 };
474 
475 #define DtoM(type) (memtype[type])
476 
477 /*
478  * Names of malloc types.
479  */
480 #define TYPENAME(type)  \
481 	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
482 /*
483  * End system adaptation definitions.
484  */
485 
486 /*
487  * Forward declarations.
488  */
489 struct inodedep_hashhead;
490 struct newblk_hashhead;
491 struct pagedep_hashhead;
492 
493 /*
494  * Internal function prototypes.
495  */
496 static	void softdep_error(char *, int);
497 static	void drain_output(struct vnode *);
498 static	struct buf *getdirtybuf(struct buf *, struct mtx *, int);
499 static	void clear_remove(struct thread *);
500 static	void clear_inodedeps(struct thread *);
501 static	int flush_pagedep_deps(struct vnode *, struct mount *,
502 	    struct diraddhd *);
503 static	int flush_inodedep_deps(struct mount *, ino_t);
504 static	int flush_deplist(struct allocdirectlst *, int, int *);
505 static	int handle_written_filepage(struct pagedep *, struct buf *);
506 static  void diradd_inode_written(struct diradd *, struct inodedep *);
507 static	int handle_written_inodeblock(struct inodedep *, struct buf *);
508 static	void handle_allocdirect_partdone(struct allocdirect *);
509 static	void handle_allocindir_partdone(struct allocindir *);
510 static	void initiate_write_filepage(struct pagedep *, struct buf *);
511 static	void handle_written_mkdir(struct mkdir *, int);
512 static	void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
513 static	void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
514 static	void handle_workitem_freefile(struct freefile *);
515 static	void handle_workitem_remove(struct dirrem *, struct vnode *);
516 static	struct dirrem *newdirrem(struct buf *, struct inode *,
517 	    struct inode *, int, struct dirrem **);
518 static	void free_diradd(struct diradd *);
519 static	void free_allocindir(struct allocindir *, struct inodedep *);
520 static	void free_newdirblk(struct newdirblk *);
521 static	int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t,
522 	    ufs2_daddr_t *);
523 static	void deallocate_dependencies(struct buf *, struct inodedep *);
524 static	void free_allocdirect(struct allocdirectlst *,
525 	    struct allocdirect *, int);
526 static	int check_inode_unwritten(struct inodedep *);
527 static	int free_inodedep(struct inodedep *);
528 static	void handle_workitem_freeblocks(struct freeblks *, int);
529 static	void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
530 static	void setup_allocindir_phase2(struct buf *, struct inode *,
531 	    struct allocindir *);
532 static	struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
533 	    ufs2_daddr_t);
534 static	void handle_workitem_freefrag(struct freefrag *);
535 static	struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long);
536 static	void allocdirect_merge(struct allocdirectlst *,
537 	    struct allocdirect *, struct allocdirect *);
538 static	struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *);
539 static	int newblk_find(struct newblk_hashhead *, struct fs *, ufs2_daddr_t,
540 	    struct newblk **);
541 static	int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **);
542 static	int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t,
543 	    struct inodedep **);
544 static	int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **);
545 static	int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **);
546 static	int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t,
547 	    struct mount *mp, int, struct pagedep **);
548 static	void pause_timer(void *);
549 static	int request_cleanup(struct mount *, int);
550 static	int process_worklist_item(struct mount *, int);
551 static	void add_to_worklist(struct worklist *);
552 static	void softdep_flush(void);
553 static	int softdep_speedup(void);
554 
555 /*
556  * Exported softdep operations.
557  */
558 static	void softdep_disk_io_initiation(struct buf *);
559 static	void softdep_disk_write_complete(struct buf *);
560 static	void softdep_deallocate_dependencies(struct buf *);
561 static	int softdep_count_dependencies(struct buf *bp, int);
562 
563 static struct mtx lk;
564 MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF);
565 
566 #define TRY_ACQUIRE_LOCK(lk)		mtx_trylock(lk)
567 #define ACQUIRE_LOCK(lk)		mtx_lock(lk)
568 #define FREE_LOCK(lk)			mtx_unlock(lk)
569 
570 /*
571  * Worklist queue management.
572  * These routines require that the lock be held.
573  */
574 #ifndef /* NOT */ DEBUG
575 #define WORKLIST_INSERT(head, item) do {	\
576 	(item)->wk_state |= ONWORKLIST;		\
577 	LIST_INSERT_HEAD(head, item, wk_list);	\
578 } while (0)
579 #define WORKLIST_REMOVE(item) do {		\
580 	(item)->wk_state &= ~ONWORKLIST;	\
581 	LIST_REMOVE(item, wk_list);		\
582 } while (0)
583 #else /* DEBUG */
584 static	void worklist_insert(struct workhead *, struct worklist *);
585 static	void worklist_remove(struct worklist *);
586 
587 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
588 #define WORKLIST_REMOVE(item) worklist_remove(item)
589 
590 static void
591 worklist_insert(head, item)
592 	struct workhead *head;
593 	struct worklist *item;
594 {
595 
596 	mtx_assert(&lk, MA_OWNED);
597 	if (item->wk_state & ONWORKLIST)
598 		panic("worklist_insert: already on list");
599 	item->wk_state |= ONWORKLIST;
600 	LIST_INSERT_HEAD(head, item, wk_list);
601 }
602 
603 static void
604 worklist_remove(item)
605 	struct worklist *item;
606 {
607 
608 	mtx_assert(&lk, MA_OWNED);
609 	if ((item->wk_state & ONWORKLIST) == 0)
610 		panic("worklist_remove: not on list");
611 	item->wk_state &= ~ONWORKLIST;
612 	LIST_REMOVE(item, wk_list);
613 }
614 #endif /* DEBUG */
615 
616 /*
617  * Routines for tracking and managing workitems.
618  */
619 static	void workitem_free(struct worklist *, int);
620 static	void workitem_alloc(struct worklist *, int, struct mount *);
621 
622 #define	WORKITEM_FREE(item, type) workitem_free((struct worklist *)(item), (type))
623 
624 static void
625 workitem_free(item, type)
626 	struct worklist *item;
627 	int type;
628 {
629 	struct ufsmount *ump;
630 	mtx_assert(&lk, MA_OWNED);
631 
632 #ifdef DEBUG
633 	if (item->wk_state & ONWORKLIST)
634 		panic("workitem_free: still on list");
635 	if (item->wk_type != type)
636 		panic("workitem_free: type mismatch");
637 #endif
638 	ump = VFSTOUFS(item->wk_mp);
639 	if (--ump->softdep_deps == 0 && ump->softdep_req)
640 		wakeup(&ump->softdep_deps);
641 	FREE(item, DtoM(type));
642 }
643 
644 static void
645 workitem_alloc(item, type, mp)
646 	struct worklist *item;
647 	int type;
648 	struct mount *mp;
649 {
650 	item->wk_type = type;
651 	item->wk_mp = mp;
652 	item->wk_state = 0;
653 	ACQUIRE_LOCK(&lk);
654 	VFSTOUFS(mp)->softdep_deps++;
655 	VFSTOUFS(mp)->softdep_accdeps++;
656 	FREE_LOCK(&lk);
657 }
658 
659 /*
660  * Workitem queue management
661  */
662 static int max_softdeps;	/* maximum number of structs before slowdown */
663 static int maxindirdeps = 50;	/* max number of indirdeps before slowdown */
664 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
665 static int proc_waiting;	/* tracks whether we have a timeout posted */
666 static int *stat_countp;	/* statistic to count in proc_waiting timeout */
667 static struct callout_handle handle; /* handle on posted proc_waiting timeout */
668 static int req_pending;
669 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
670 #define FLUSH_INODES		1
671 static int req_clear_remove;	/* syncer process flush some freeblks */
672 #define FLUSH_REMOVE		2
673 #define FLUSH_REMOVE_WAIT	3
674 /*
675  * runtime statistics
676  */
677 static int stat_worklist_push;	/* number of worklist cleanups */
678 static int stat_blk_limit_push;	/* number of times block limit neared */
679 static int stat_ino_limit_push;	/* number of times inode limit neared */
680 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
681 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
682 static int stat_sync_limit_hit;	/* number of synchronous slowdowns imposed */
683 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
684 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
685 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
686 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
687 
688 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
689 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
690 SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, "");
691 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
692 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
693 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
694 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
695 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
696 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
697 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
698 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
699 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
700 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
701 /* SYSCTL_INT(_debug, OID_AUTO, worklist_num, CTLFLAG_RD, &softdep_on_worklist, 0, ""); */
702 
703 SYSCTL_DECL(_vfs_ffs);
704 
705 static int compute_summary_at_mount = 0;	/* Whether to recompute the summary at mount time */
706 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW,
707 	   &compute_summary_at_mount, 0, "Recompute summary at mount");
708 
709 static struct proc *softdepproc;
710 static struct kproc_desc softdep_kp = {
711 	"softdepflush",
712 	softdep_flush,
713 	&softdepproc
714 };
715 SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start, &softdep_kp)
716 
717 static void
718 softdep_flush(void)
719 {
720 	struct mount *nmp;
721 	struct mount *mp;
722 	struct ufsmount *ump;
723 	struct thread *td;
724 	int remaining;
725 	int vfslocked;
726 
727 	td = curthread;
728 	td->td_pflags |= TDP_NORUNNINGBUF;
729 
730 	for (;;) {
731 		kthread_suspend_check(softdepproc);
732 		vfslocked = VFS_LOCK_GIANT((struct mount *)NULL);
733 		ACQUIRE_LOCK(&lk);
734 		/*
735 		 * If requested, try removing inode or removal dependencies.
736 		 */
737 		if (req_clear_inodedeps) {
738 			clear_inodedeps(td);
739 			req_clear_inodedeps -= 1;
740 			wakeup_one(&proc_waiting);
741 		}
742 		if (req_clear_remove) {
743 			clear_remove(td);
744 			req_clear_remove -= 1;
745 			wakeup_one(&proc_waiting);
746 		}
747 		FREE_LOCK(&lk);
748 		VFS_UNLOCK_GIANT(vfslocked);
749 		remaining = 0;
750 		mtx_lock(&mountlist_mtx);
751 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp)  {
752 			nmp = TAILQ_NEXT(mp, mnt_list);
753 			if ((mp->mnt_flag & MNT_SOFTDEP) == 0)
754 				continue;
755 			if (vfs_busy(mp, LK_NOWAIT, &mountlist_mtx, td))
756 				continue;
757 			vfslocked = VFS_LOCK_GIANT(mp);
758 			softdep_process_worklist(mp, 0);
759 			ump = VFSTOUFS(mp);
760 			remaining += ump->softdep_on_worklist -
761 				ump->softdep_on_worklist_inprogress;
762 			VFS_UNLOCK_GIANT(vfslocked);
763 			mtx_lock(&mountlist_mtx);
764 			nmp = TAILQ_NEXT(mp, mnt_list);
765 			vfs_unbusy(mp, td);
766 		}
767 		mtx_unlock(&mountlist_mtx);
768 		if (remaining)
769 			continue;
770 		ACQUIRE_LOCK(&lk);
771 		if (!req_pending)
772 			msleep(&req_pending, &lk, PVM, "sdflush", hz);
773 		req_pending = 0;
774 		FREE_LOCK(&lk);
775 	}
776 }
777 
778 static int
779 softdep_speedup(void)
780 {
781 
782 	mtx_assert(&lk, MA_OWNED);
783 	if (req_pending == 0) {
784 		req_pending = 1;
785 		wakeup(&req_pending);
786 	}
787 
788 	return speedup_syncer();
789 }
790 
791 /*
792  * Add an item to the end of the work queue.
793  * This routine requires that the lock be held.
794  * This is the only routine that adds items to the list.
795  * The following routine is the only one that removes items
796  * and does so in order from first to last.
797  */
798 static void
799 add_to_worklist(wk)
800 	struct worklist *wk;
801 {
802 	struct ufsmount *ump;
803 
804 	mtx_assert(&lk, MA_OWNED);
805 	ump = VFSTOUFS(wk->wk_mp);
806 	if (wk->wk_state & ONWORKLIST)
807 		panic("add_to_worklist: already on list");
808 	wk->wk_state |= ONWORKLIST;
809 	if (LIST_EMPTY(&ump->softdep_workitem_pending))
810 		LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list);
811 	else
812 		LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list);
813 	ump->softdep_worklist_tail = wk;
814 	ump->softdep_on_worklist += 1;
815 }
816 
817 /*
818  * Process that runs once per second to handle items in the background queue.
819  *
820  * Note that we ensure that everything is done in the order in which they
821  * appear in the queue. The code below depends on this property to ensure
822  * that blocks of a file are freed before the inode itself is freed. This
823  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
824  * until all the old ones have been purged from the dependency lists.
825  */
826 int
827 softdep_process_worklist(mp, full)
828 	struct mount *mp;
829 	int full;
830 {
831 	struct thread *td = curthread;
832 	int cnt, matchcnt, loopcount;
833 	struct ufsmount *ump;
834 	long starttime;
835 
836 	KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp"));
837 	/*
838 	 * Record the process identifier of our caller so that we can give
839 	 * this process preferential treatment in request_cleanup below.
840 	 */
841 	matchcnt = 0;
842 	ump = VFSTOUFS(mp);
843 	ACQUIRE_LOCK(&lk);
844 	loopcount = 1;
845 	starttime = time_second;
846 	while (ump->softdep_on_worklist > 0) {
847 		if ((cnt = process_worklist_item(mp, 0)) == -1)
848 			break;
849 		else
850 			matchcnt += cnt;
851 		/*
852 		 * If requested, try removing inode or removal dependencies.
853 		 */
854 		if (req_clear_inodedeps) {
855 			clear_inodedeps(td);
856 			req_clear_inodedeps -= 1;
857 			wakeup_one(&proc_waiting);
858 		}
859 		if (req_clear_remove) {
860 			clear_remove(td);
861 			req_clear_remove -= 1;
862 			wakeup_one(&proc_waiting);
863 		}
864 		/*
865 		 * We do not generally want to stop for buffer space, but if
866 		 * we are really being a buffer hog, we will stop and wait.
867 		 */
868 		if (loopcount++ % 128 == 0) {
869 			FREE_LOCK(&lk);
870 			bwillwrite();
871 			ACQUIRE_LOCK(&lk);
872 		}
873 		/*
874 		 * Never allow processing to run for more than one
875 		 * second. Otherwise the other mountpoints may get
876 		 * excessively backlogged.
877 		 */
878 		if (!full && starttime != time_second) {
879 			matchcnt = -1;
880 			break;
881 		}
882 	}
883 	FREE_LOCK(&lk);
884 	return (matchcnt);
885 }
886 
887 /*
888  * Process one item on the worklist.
889  */
890 static int
891 process_worklist_item(mp, flags)
892 	struct mount *mp;
893 	int flags;
894 {
895 	struct worklist *wk, *wkend;
896 	struct ufsmount *ump;
897 	struct vnode *vp;
898 	int matchcnt = 0;
899 
900 	mtx_assert(&lk, MA_OWNED);
901 	KASSERT(mp != NULL, ("process_worklist_item: NULL mp"));
902 	/*
903 	 * If we are being called because of a process doing a
904 	 * copy-on-write, then it is not safe to write as we may
905 	 * recurse into the copy-on-write routine.
906 	 */
907 	if (curthread->td_pflags & TDP_COWINPROGRESS)
908 		return (-1);
909 	/*
910 	 * Normally we just process each item on the worklist in order.
911 	 * However, if we are in a situation where we cannot lock any
912 	 * inodes, we have to skip over any dirrem requests whose
913 	 * vnodes are resident and locked.
914 	 */
915 	ump = VFSTOUFS(mp);
916 	vp = NULL;
917 	LIST_FOREACH(wk, &ump->softdep_workitem_pending, wk_list) {
918 		if (wk->wk_state & INPROGRESS)
919 			continue;
920 		if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
921 			break;
922 		wk->wk_state |= INPROGRESS;
923 		ump->softdep_on_worklist_inprogress++;
924 		FREE_LOCK(&lk);
925 		ffs_vget(mp, WK_DIRREM(wk)->dm_oldinum,
926 		    LK_NOWAIT | LK_EXCLUSIVE, &vp);
927 		ACQUIRE_LOCK(&lk);
928 		wk->wk_state &= ~INPROGRESS;
929 		ump->softdep_on_worklist_inprogress--;
930 		if (vp != NULL)
931 			break;
932 	}
933 	if (wk == 0)
934 		return (-1);
935 	/*
936 	 * Remove the item to be processed. If we are removing the last
937 	 * item on the list, we need to recalculate the tail pointer.
938 	 * As this happens rarely and usually when the list is short,
939 	 * we just run down the list to find it rather than tracking it
940 	 * in the above loop.
941 	 */
942 	WORKLIST_REMOVE(wk);
943 	if (wk == ump->softdep_worklist_tail) {
944 		LIST_FOREACH(wkend, &ump->softdep_workitem_pending, wk_list)
945 			if (LIST_NEXT(wkend, wk_list) == NULL)
946 				break;
947 		ump->softdep_worklist_tail = wkend;
948 	}
949 	ump->softdep_on_worklist -= 1;
950 	FREE_LOCK(&lk);
951 	if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
952 		panic("process_worklist_item: suspended filesystem");
953 	matchcnt++;
954 	switch (wk->wk_type) {
955 
956 	case D_DIRREM:
957 		/* removal of a directory entry */
958 		handle_workitem_remove(WK_DIRREM(wk), vp);
959 		break;
960 
961 	case D_FREEBLKS:
962 		/* releasing blocks and/or fragments from a file */
963 		handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT);
964 		break;
965 
966 	case D_FREEFRAG:
967 		/* releasing a fragment when replaced as a file grows */
968 		handle_workitem_freefrag(WK_FREEFRAG(wk));
969 		break;
970 
971 	case D_FREEFILE:
972 		/* releasing an inode when its link count drops to 0 */
973 		handle_workitem_freefile(WK_FREEFILE(wk));
974 		break;
975 
976 	default:
977 		panic("%s_process_worklist: Unknown type %s",
978 		    "softdep", TYPENAME(wk->wk_type));
979 		/* NOTREACHED */
980 	}
981 	vn_finished_secondary_write(mp);
982 	ACQUIRE_LOCK(&lk);
983 	return (matchcnt);
984 }
985 
986 /*
987  * Move dependencies from one buffer to another.
988  */
989 void
990 softdep_move_dependencies(oldbp, newbp)
991 	struct buf *oldbp;
992 	struct buf *newbp;
993 {
994 	struct worklist *wk, *wktail;
995 
996 	if (!LIST_EMPTY(&newbp->b_dep))
997 		panic("softdep_move_dependencies: need merge code");
998 	wktail = 0;
999 	ACQUIRE_LOCK(&lk);
1000 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
1001 		LIST_REMOVE(wk, wk_list);
1002 		if (wktail == 0)
1003 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
1004 		else
1005 			LIST_INSERT_AFTER(wktail, wk, wk_list);
1006 		wktail = wk;
1007 	}
1008 	FREE_LOCK(&lk);
1009 }
1010 
1011 /*
1012  * Purge the work list of all items associated with a particular mount point.
1013  */
1014 int
1015 softdep_flushworklist(oldmnt, countp, td)
1016 	struct mount *oldmnt;
1017 	int *countp;
1018 	struct thread *td;
1019 {
1020 	struct vnode *devvp;
1021 	int count, error = 0;
1022 	struct ufsmount *ump;
1023 
1024 	/*
1025 	 * Alternately flush the block device associated with the mount
1026 	 * point and process any dependencies that the flushing
1027 	 * creates. We continue until no more worklist dependencies
1028 	 * are found.
1029 	 */
1030 	*countp = 0;
1031 	ump = VFSTOUFS(oldmnt);
1032 	devvp = ump->um_devvp;
1033 	while ((count = softdep_process_worklist(oldmnt, 1)) > 0) {
1034 		*countp += count;
1035 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td);
1036 		error = VOP_FSYNC(devvp, MNT_WAIT, td);
1037 		VOP_UNLOCK(devvp, 0, td);
1038 		if (error)
1039 			break;
1040 	}
1041 	return (error);
1042 }
1043 
1044 int
1045 softdep_waitidle(struct mount *mp)
1046 {
1047 	struct ufsmount *ump;
1048 	int error;
1049 	int i;
1050 
1051 	ump = VFSTOUFS(mp);
1052 	ACQUIRE_LOCK(&lk);
1053 	for (i = 0; i < 10 && ump->softdep_deps; i++) {
1054 		ump->softdep_req = 1;
1055 		if (ump->softdep_on_worklist)
1056 			panic("softdep_waitidle: work added after flush.");
1057 		msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1);
1058 	}
1059 	ump->softdep_req = 0;
1060 	FREE_LOCK(&lk);
1061 	error = 0;
1062 	if (i == 10) {
1063 		error = EBUSY;
1064 		printf("softdep_waitidle: Failed to flush worklist for %p\n",
1065 		    mp);
1066 	}
1067 
1068 	return (error);
1069 }
1070 
1071 /*
1072  * Flush all vnodes and worklist items associated with a specified mount point.
1073  */
1074 int
1075 softdep_flushfiles(oldmnt, flags, td)
1076 	struct mount *oldmnt;
1077 	int flags;
1078 	struct thread *td;
1079 {
1080 	int error, count, loopcnt;
1081 
1082 	error = 0;
1083 
1084 	/*
1085 	 * Alternately flush the vnodes associated with the mount
1086 	 * point and process any dependencies that the flushing
1087 	 * creates. In theory, this loop can happen at most twice,
1088 	 * but we give it a few extra just to be sure.
1089 	 */
1090 	for (loopcnt = 10; loopcnt > 0; loopcnt--) {
1091 		/*
1092 		 * Do another flush in case any vnodes were brought in
1093 		 * as part of the cleanup operations.
1094 		 */
1095 		if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
1096 			break;
1097 		if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 ||
1098 		    count == 0)
1099 			break;
1100 	}
1101 	/*
1102 	 * If we are unmounting then it is an error to fail. If we
1103 	 * are simply trying to downgrade to read-only, then filesystem
1104 	 * activity can keep us busy forever, so we just fail with EBUSY.
1105 	 */
1106 	if (loopcnt == 0) {
1107 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
1108 			panic("softdep_flushfiles: looping");
1109 		error = EBUSY;
1110 	}
1111 	if (!error)
1112 		error = softdep_waitidle(oldmnt);
1113 	return (error);
1114 }
1115 
1116 /*
1117  * Structure hashing.
1118  *
1119  * There are three types of structures that can be looked up:
1120  *	1) pagedep structures identified by mount point, inode number,
1121  *	   and logical block.
1122  *	2) inodedep structures identified by mount point and inode number.
1123  *	3) newblk structures identified by mount point and
1124  *	   physical block number.
1125  *
1126  * The "pagedep" and "inodedep" dependency structures are hashed
1127  * separately from the file blocks and inodes to which they correspond.
1128  * This separation helps when the in-memory copy of an inode or
1129  * file block must be replaced. It also obviates the need to access
1130  * an inode or file page when simply updating (or de-allocating)
1131  * dependency structures. Lookup of newblk structures is needed to
1132  * find newly allocated blocks when trying to associate them with
1133  * their allocdirect or allocindir structure.
1134  *
1135  * The lookup routines optionally create and hash a new instance when
1136  * an existing entry is not found.
1137  */
1138 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
1139 #define NODELAY		0x0002	/* cannot do background work */
1140 
1141 /*
1142  * Structures and routines associated with pagedep caching.
1143  */
1144 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
1145 u_long	pagedep_hash;		/* size of hash table - 1 */
1146 #define	PAGEDEP_HASH(mp, inum, lbn) \
1147 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
1148 	    pagedep_hash])
1149 
1150 static int
1151 pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp)
1152 	struct pagedep_hashhead *pagedephd;
1153 	ino_t ino;
1154 	ufs_lbn_t lbn;
1155 	struct mount *mp;
1156 	int flags;
1157 	struct pagedep **pagedeppp;
1158 {
1159 	struct pagedep *pagedep;
1160 
1161 	LIST_FOREACH(pagedep, pagedephd, pd_hash)
1162 		if (ino == pagedep->pd_ino &&
1163 		    lbn == pagedep->pd_lbn &&
1164 		    mp == pagedep->pd_list.wk_mp)
1165 			break;
1166 	if (pagedep) {
1167 		*pagedeppp = pagedep;
1168 		if ((flags & DEPALLOC) != 0 &&
1169 		    (pagedep->pd_state & ONWORKLIST) == 0)
1170 			return (0);
1171 		return (1);
1172 	}
1173 	*pagedeppp = NULL;
1174 	return (0);
1175 }
1176 /*
1177  * Look up a pagedep. Return 1 if found, 0 if not found or found
1178  * when asked to allocate but not associated with any buffer.
1179  * If not found, allocate if DEPALLOC flag is passed.
1180  * Found or allocated entry is returned in pagedeppp.
1181  * This routine must be called with splbio interrupts blocked.
1182  */
1183 static int
1184 pagedep_lookup(ip, lbn, flags, pagedeppp)
1185 	struct inode *ip;
1186 	ufs_lbn_t lbn;
1187 	int flags;
1188 	struct pagedep **pagedeppp;
1189 {
1190 	struct pagedep *pagedep;
1191 	struct pagedep_hashhead *pagedephd;
1192 	struct mount *mp;
1193 	int ret;
1194 	int i;
1195 
1196 	mtx_assert(&lk, MA_OWNED);
1197 	mp = ITOV(ip)->v_mount;
1198 	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
1199 
1200 	ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp);
1201 	if (*pagedeppp || (flags & DEPALLOC) == 0)
1202 		return (ret);
1203 	FREE_LOCK(&lk);
1204 	MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep),
1205 	    M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO);
1206 	workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp);
1207 	ACQUIRE_LOCK(&lk);
1208 	ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp);
1209 	if (*pagedeppp) {
1210 		WORKITEM_FREE(pagedep, D_PAGEDEP);
1211 		return (ret);
1212 	}
1213 	pagedep->pd_ino = ip->i_number;
1214 	pagedep->pd_lbn = lbn;
1215 	LIST_INIT(&pagedep->pd_dirremhd);
1216 	LIST_INIT(&pagedep->pd_pendinghd);
1217 	for (i = 0; i < DAHASHSZ; i++)
1218 		LIST_INIT(&pagedep->pd_diraddhd[i]);
1219 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
1220 	*pagedeppp = pagedep;
1221 	return (0);
1222 }
1223 
1224 /*
1225  * Structures and routines associated with inodedep caching.
1226  */
1227 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
1228 static u_long	inodedep_hash;	/* size of hash table - 1 */
1229 static long	num_inodedep;	/* number of inodedep allocated */
1230 #define	INODEDEP_HASH(fs, inum) \
1231       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
1232 
1233 static int
1234 inodedep_find(inodedephd, fs, inum, inodedeppp)
1235 	struct inodedep_hashhead *inodedephd;
1236 	struct fs *fs;
1237 	ino_t inum;
1238 	struct inodedep **inodedeppp;
1239 {
1240 	struct inodedep *inodedep;
1241 
1242 	LIST_FOREACH(inodedep, inodedephd, id_hash)
1243 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
1244 			break;
1245 	if (inodedep) {
1246 		*inodedeppp = inodedep;
1247 		return (1);
1248 	}
1249 	*inodedeppp = NULL;
1250 
1251 	return (0);
1252 }
1253 /*
1254  * Look up an inodedep. Return 1 if found, 0 if not found.
1255  * If not found, allocate if DEPALLOC flag is passed.
1256  * Found or allocated entry is returned in inodedeppp.
1257  * This routine must be called with splbio interrupts blocked.
1258  */
1259 static int
1260 inodedep_lookup(mp, inum, flags, inodedeppp)
1261 	struct mount *mp;
1262 	ino_t inum;
1263 	int flags;
1264 	struct inodedep **inodedeppp;
1265 {
1266 	struct inodedep *inodedep;
1267 	struct inodedep_hashhead *inodedephd;
1268 	struct fs *fs;
1269 
1270 	mtx_assert(&lk, MA_OWNED);
1271 	fs = VFSTOUFS(mp)->um_fs;
1272 	inodedephd = INODEDEP_HASH(fs, inum);
1273 
1274 	if (inodedep_find(inodedephd, fs, inum, inodedeppp))
1275 		return (1);
1276 	if ((flags & DEPALLOC) == 0)
1277 		return (0);
1278 	/*
1279 	 * If we are over our limit, try to improve the situation.
1280 	 */
1281 	if (num_inodedep > max_softdeps && (flags & NODELAY) == 0)
1282 		request_cleanup(mp, FLUSH_INODES);
1283 	FREE_LOCK(&lk);
1284 	MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1285 		M_INODEDEP, M_SOFTDEP_FLAGS);
1286 	workitem_alloc(&inodedep->id_list, D_INODEDEP, mp);
1287 	ACQUIRE_LOCK(&lk);
1288 	if (inodedep_find(inodedephd, fs, inum, inodedeppp)) {
1289 		WORKITEM_FREE(inodedep, D_INODEDEP);
1290 		return (1);
1291 	}
1292 	num_inodedep += 1;
1293 	inodedep->id_fs = fs;
1294 	inodedep->id_ino = inum;
1295 	inodedep->id_state = ALLCOMPLETE;
1296 	inodedep->id_nlinkdelta = 0;
1297 	inodedep->id_savedino1 = NULL;
1298 	inodedep->id_savedsize = -1;
1299 	inodedep->id_savedextsize = -1;
1300 	inodedep->id_buf = NULL;
1301 	LIST_INIT(&inodedep->id_pendinghd);
1302 	LIST_INIT(&inodedep->id_inowait);
1303 	LIST_INIT(&inodedep->id_bufwait);
1304 	TAILQ_INIT(&inodedep->id_inoupdt);
1305 	TAILQ_INIT(&inodedep->id_newinoupdt);
1306 	TAILQ_INIT(&inodedep->id_extupdt);
1307 	TAILQ_INIT(&inodedep->id_newextupdt);
1308 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1309 	*inodedeppp = inodedep;
1310 	return (0);
1311 }
1312 
1313 /*
1314  * Structures and routines associated with newblk caching.
1315  */
1316 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1317 u_long	newblk_hash;		/* size of hash table - 1 */
1318 #define	NEWBLK_HASH(fs, inum) \
1319 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1320 
1321 static int
1322 newblk_find(newblkhd, fs, newblkno, newblkpp)
1323 	struct newblk_hashhead *newblkhd;
1324 	struct fs *fs;
1325 	ufs2_daddr_t newblkno;
1326 	struct newblk **newblkpp;
1327 {
1328 	struct newblk *newblk;
1329 
1330 	LIST_FOREACH(newblk, newblkhd, nb_hash)
1331 		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1332 			break;
1333 	if (newblk) {
1334 		*newblkpp = newblk;
1335 		return (1);
1336 	}
1337 	*newblkpp = NULL;
1338 	return (0);
1339 }
1340 
1341 /*
1342  * Look up a newblk. Return 1 if found, 0 if not found.
1343  * If not found, allocate if DEPALLOC flag is passed.
1344  * Found or allocated entry is returned in newblkpp.
1345  */
1346 static int
1347 newblk_lookup(fs, newblkno, flags, newblkpp)
1348 	struct fs *fs;
1349 	ufs2_daddr_t newblkno;
1350 	int flags;
1351 	struct newblk **newblkpp;
1352 {
1353 	struct newblk *newblk;
1354 	struct newblk_hashhead *newblkhd;
1355 
1356 	newblkhd = NEWBLK_HASH(fs, newblkno);
1357 	if (newblk_find(newblkhd, fs, newblkno, newblkpp))
1358 		return (1);
1359 	if ((flags & DEPALLOC) == 0)
1360 		return (0);
1361 	FREE_LOCK(&lk);
1362 	MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1363 		M_NEWBLK, M_SOFTDEP_FLAGS);
1364 	ACQUIRE_LOCK(&lk);
1365 	if (newblk_find(newblkhd, fs, newblkno, newblkpp)) {
1366 		FREE(newblk, M_NEWBLK);
1367 		return (1);
1368 	}
1369 	newblk->nb_state = 0;
1370 	newblk->nb_fs = fs;
1371 	newblk->nb_newblkno = newblkno;
1372 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1373 	*newblkpp = newblk;
1374 	return (0);
1375 }
1376 
1377 /*
1378  * Executed during filesystem system initialization before
1379  * mounting any filesystems.
1380  */
1381 void
1382 softdep_initialize()
1383 {
1384 
1385 	LIST_INIT(&mkdirlisthd);
1386 	max_softdeps = desiredvnodes * 4;
1387 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1388 	    &pagedep_hash);
1389 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1390 	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1391 
1392 	/* initialise bioops hack */
1393 	bioops.io_start = softdep_disk_io_initiation;
1394 	bioops.io_complete = softdep_disk_write_complete;
1395 	bioops.io_deallocate = softdep_deallocate_dependencies;
1396 	bioops.io_countdeps = softdep_count_dependencies;
1397 }
1398 
1399 /*
1400  * Executed after all filesystems have been unmounted during
1401  * filesystem module unload.
1402  */
1403 void
1404 softdep_uninitialize()
1405 {
1406 
1407 	hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash);
1408 	hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash);
1409 	hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash);
1410 }
1411 
1412 /*
1413  * Called at mount time to notify the dependency code that a
1414  * filesystem wishes to use it.
1415  */
1416 int
1417 softdep_mount(devvp, mp, fs, cred)
1418 	struct vnode *devvp;
1419 	struct mount *mp;
1420 	struct fs *fs;
1421 	struct ucred *cred;
1422 {
1423 	struct csum_total cstotal;
1424 	struct ufsmount *ump;
1425 	struct cg *cgp;
1426 	struct buf *bp;
1427 	int error, cyl;
1428 
1429 	MNT_ILOCK(mp);
1430 	mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP;
1431 	if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) {
1432 		mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) |
1433 			MNTK_SOFTDEP;
1434 		mp->mnt_noasync++;
1435 	}
1436 	MNT_IUNLOCK(mp);
1437 	ump = VFSTOUFS(mp);
1438 	LIST_INIT(&ump->softdep_workitem_pending);
1439 	ump->softdep_worklist_tail = NULL;
1440 	ump->softdep_on_worklist = 0;
1441 	ump->softdep_deps = 0;
1442 	/*
1443 	 * When doing soft updates, the counters in the
1444 	 * superblock may have gotten out of sync. Recomputation
1445 	 * can take a long time and can be deferred for background
1446 	 * fsck.  However, the old behavior of scanning the cylinder
1447 	 * groups and recalculating them at mount time is available
1448 	 * by setting vfs.ffs.compute_summary_at_mount to one.
1449 	 */
1450 	if (compute_summary_at_mount == 0 || fs->fs_clean != 0)
1451 		return (0);
1452 	bzero(&cstotal, sizeof cstotal);
1453 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1454 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1455 		    fs->fs_cgsize, cred, &bp)) != 0) {
1456 			brelse(bp);
1457 			return (error);
1458 		}
1459 		cgp = (struct cg *)bp->b_data;
1460 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1461 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1462 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1463 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1464 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
1465 		brelse(bp);
1466 	}
1467 #ifdef DEBUG
1468 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1469 		printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
1470 #endif
1471 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1472 	return (0);
1473 }
1474 
1475 /*
1476  * Protecting the freemaps (or bitmaps).
1477  *
1478  * To eliminate the need to execute fsck before mounting a filesystem
1479  * after a power failure, one must (conservatively) guarantee that the
1480  * on-disk copy of the bitmaps never indicate that a live inode or block is
1481  * free.  So, when a block or inode is allocated, the bitmap should be
1482  * updated (on disk) before any new pointers.  When a block or inode is
1483  * freed, the bitmap should not be updated until all pointers have been
1484  * reset.  The latter dependency is handled by the delayed de-allocation
1485  * approach described below for block and inode de-allocation.  The former
1486  * dependency is handled by calling the following procedure when a block or
1487  * inode is allocated. When an inode is allocated an "inodedep" is created
1488  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1489  * Each "inodedep" is also inserted into the hash indexing structure so
1490  * that any additional link additions can be made dependent on the inode
1491  * allocation.
1492  *
1493  * The ufs filesystem maintains a number of free block counts (e.g., per
1494  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1495  * in addition to the bitmaps.  These counts are used to improve efficiency
1496  * during allocation and therefore must be consistent with the bitmaps.
1497  * There is no convenient way to guarantee post-crash consistency of these
1498  * counts with simple update ordering, for two main reasons: (1) The counts
1499  * and bitmaps for a single cylinder group block are not in the same disk
1500  * sector.  If a disk write is interrupted (e.g., by power failure), one may
1501  * be written and the other not.  (2) Some of the counts are located in the
1502  * superblock rather than the cylinder group block. So, we focus our soft
1503  * updates implementation on protecting the bitmaps. When mounting a
1504  * filesystem, we recompute the auxiliary counts from the bitmaps.
1505  */
1506 
1507 /*
1508  * Called just after updating the cylinder group block to allocate an inode.
1509  */
1510 void
1511 softdep_setup_inomapdep(bp, ip, newinum)
1512 	struct buf *bp;		/* buffer for cylgroup block with inode map */
1513 	struct inode *ip;	/* inode related to allocation */
1514 	ino_t newinum;		/* new inode number being allocated */
1515 {
1516 	struct inodedep *inodedep;
1517 	struct bmsafemap *bmsafemap;
1518 
1519 	/*
1520 	 * Create a dependency for the newly allocated inode.
1521 	 * Panic if it already exists as something is seriously wrong.
1522 	 * Otherwise add it to the dependency list for the buffer holding
1523 	 * the cylinder group map from which it was allocated.
1524 	 */
1525 	ACQUIRE_LOCK(&lk);
1526 	if ((inodedep_lookup(UFSTOVFS(ip->i_ump), newinum, DEPALLOC|NODELAY,
1527 	    &inodedep)))
1528 		panic("softdep_setup_inomapdep: dependency for new inode "
1529 		    "already exists");
1530 	inodedep->id_buf = bp;
1531 	inodedep->id_state &= ~DEPCOMPLETE;
1532 	bmsafemap = bmsafemap_lookup(inodedep->id_list.wk_mp, bp);
1533 	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1534 	FREE_LOCK(&lk);
1535 }
1536 
1537 /*
1538  * Called just after updating the cylinder group block to
1539  * allocate block or fragment.
1540  */
1541 void
1542 softdep_setup_blkmapdep(bp, mp, newblkno)
1543 	struct buf *bp;		/* buffer for cylgroup block with block map */
1544 	struct mount *mp;	/* filesystem doing allocation */
1545 	ufs2_daddr_t newblkno;	/* number of newly allocated block */
1546 {
1547 	struct newblk *newblk;
1548 	struct bmsafemap *bmsafemap;
1549 	struct fs *fs;
1550 
1551 	fs = VFSTOUFS(mp)->um_fs;
1552 	/*
1553 	 * Create a dependency for the newly allocated block.
1554 	 * Add it to the dependency list for the buffer holding
1555 	 * the cylinder group map from which it was allocated.
1556 	 */
1557 	ACQUIRE_LOCK(&lk);
1558 	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1559 		panic("softdep_setup_blkmapdep: found block");
1560 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp);
1561 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1562 	FREE_LOCK(&lk);
1563 }
1564 
1565 /*
1566  * Find the bmsafemap associated with a cylinder group buffer.
1567  * If none exists, create one. The buffer must be locked when
1568  * this routine is called and this routine must be called with
1569  * splbio interrupts blocked.
1570  */
1571 static struct bmsafemap *
1572 bmsafemap_lookup(mp, bp)
1573 	struct mount *mp;
1574 	struct buf *bp;
1575 {
1576 	struct bmsafemap *bmsafemap;
1577 	struct worklist *wk;
1578 
1579 	mtx_assert(&lk, MA_OWNED);
1580 	LIST_FOREACH(wk, &bp->b_dep, wk_list)
1581 		if (wk->wk_type == D_BMSAFEMAP)
1582 			return (WK_BMSAFEMAP(wk));
1583 	FREE_LOCK(&lk);
1584 	MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1585 		M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1586 	workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
1587 	bmsafemap->sm_buf = bp;
1588 	LIST_INIT(&bmsafemap->sm_allocdirecthd);
1589 	LIST_INIT(&bmsafemap->sm_allocindirhd);
1590 	LIST_INIT(&bmsafemap->sm_inodedephd);
1591 	LIST_INIT(&bmsafemap->sm_newblkhd);
1592 	ACQUIRE_LOCK(&lk);
1593 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1594 	return (bmsafemap);
1595 }
1596 
1597 /*
1598  * Direct block allocation dependencies.
1599  *
1600  * When a new block is allocated, the corresponding disk locations must be
1601  * initialized (with zeros or new data) before the on-disk inode points to
1602  * them.  Also, the freemap from which the block was allocated must be
1603  * updated (on disk) before the inode's pointer. These two dependencies are
1604  * independent of each other and are needed for all file blocks and indirect
1605  * blocks that are pointed to directly by the inode.  Just before the
1606  * "in-core" version of the inode is updated with a newly allocated block
1607  * number, a procedure (below) is called to setup allocation dependency
1608  * structures.  These structures are removed when the corresponding
1609  * dependencies are satisfied or when the block allocation becomes obsolete
1610  * (i.e., the file is deleted, the block is de-allocated, or the block is a
1611  * fragment that gets upgraded).  All of these cases are handled in
1612  * procedures described later.
1613  *
1614  * When a file extension causes a fragment to be upgraded, either to a larger
1615  * fragment or to a full block, the on-disk location may change (if the
1616  * previous fragment could not simply be extended). In this case, the old
1617  * fragment must be de-allocated, but not until after the inode's pointer has
1618  * been updated. In most cases, this is handled by later procedures, which
1619  * will construct a "freefrag" structure to be added to the workitem queue
1620  * when the inode update is complete (or obsolete).  The main exception to
1621  * this is when an allocation occurs while a pending allocation dependency
1622  * (for the same block pointer) remains.  This case is handled in the main
1623  * allocation dependency setup procedure by immediately freeing the
1624  * unreferenced fragments.
1625  */
1626 void
1627 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1628 	struct inode *ip;	/* inode to which block is being added */
1629 	ufs_lbn_t lbn;		/* block pointer within inode */
1630 	ufs2_daddr_t newblkno;	/* disk block number being added */
1631 	ufs2_daddr_t oldblkno;	/* previous block number, 0 unless frag */
1632 	long newsize;		/* size of new block */
1633 	long oldsize;		/* size of new block */
1634 	struct buf *bp;		/* bp for allocated block */
1635 {
1636 	struct allocdirect *adp, *oldadp;
1637 	struct allocdirectlst *adphead;
1638 	struct bmsafemap *bmsafemap;
1639 	struct inodedep *inodedep;
1640 	struct pagedep *pagedep;
1641 	struct newblk *newblk;
1642 	struct mount *mp;
1643 
1644 	mp = UFSTOVFS(ip->i_ump);
1645 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1646 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1647 	workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp);
1648 	adp->ad_lbn = lbn;
1649 	adp->ad_newblkno = newblkno;
1650 	adp->ad_oldblkno = oldblkno;
1651 	adp->ad_newsize = newsize;
1652 	adp->ad_oldsize = oldsize;
1653 	adp->ad_state = ATTACHED;
1654 	LIST_INIT(&adp->ad_newdirblk);
1655 	if (newblkno == oldblkno)
1656 		adp->ad_freefrag = NULL;
1657 	else
1658 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1659 
1660 	ACQUIRE_LOCK(&lk);
1661 	if (lbn >= NDADDR) {
1662 		/* allocating an indirect block */
1663 		if (oldblkno != 0)
1664 			panic("softdep_setup_allocdirect: non-zero indir");
1665 	} else {
1666 		/*
1667 		 * Allocating a direct block.
1668 		 *
1669 		 * If we are allocating a directory block, then we must
1670 		 * allocate an associated pagedep to track additions and
1671 		 * deletions.
1672 		 */
1673 		if ((ip->i_mode & IFMT) == IFDIR &&
1674 		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1675 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1676 	}
1677 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1678 		panic("softdep_setup_allocdirect: lost block");
1679 	if (newblk->nb_state == DEPCOMPLETE) {
1680 		adp->ad_state |= DEPCOMPLETE;
1681 		adp->ad_buf = NULL;
1682 	} else {
1683 		bmsafemap = newblk->nb_bmsafemap;
1684 		adp->ad_buf = bmsafemap->sm_buf;
1685 		LIST_REMOVE(newblk, nb_deps);
1686 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1687 	}
1688 	LIST_REMOVE(newblk, nb_hash);
1689 	FREE(newblk, M_NEWBLK);
1690 
1691 	inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1692 	adp->ad_inodedep = inodedep;
1693 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1694 	/*
1695 	 * The list of allocdirects must be kept in sorted and ascending
1696 	 * order so that the rollback routines can quickly determine the
1697 	 * first uncommitted block (the size of the file stored on disk
1698 	 * ends at the end of the lowest committed fragment, or if there
1699 	 * are no fragments, at the end of the highest committed block).
1700 	 * Since files generally grow, the typical case is that the new
1701 	 * block is to be added at the end of the list. We speed this
1702 	 * special case by checking against the last allocdirect in the
1703 	 * list before laboriously traversing the list looking for the
1704 	 * insertion point.
1705 	 */
1706 	adphead = &inodedep->id_newinoupdt;
1707 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1708 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1709 		/* insert at end of list */
1710 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1711 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1712 			allocdirect_merge(adphead, adp, oldadp);
1713 		FREE_LOCK(&lk);
1714 		return;
1715 	}
1716 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1717 		if (oldadp->ad_lbn >= lbn)
1718 			break;
1719 	}
1720 	if (oldadp == NULL)
1721 		panic("softdep_setup_allocdirect: lost entry");
1722 	/* insert in middle of list */
1723 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1724 	if (oldadp->ad_lbn == lbn)
1725 		allocdirect_merge(adphead, adp, oldadp);
1726 	FREE_LOCK(&lk);
1727 }
1728 
1729 /*
1730  * Replace an old allocdirect dependency with a newer one.
1731  * This routine must be called with splbio interrupts blocked.
1732  */
1733 static void
1734 allocdirect_merge(adphead, newadp, oldadp)
1735 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
1736 	struct allocdirect *newadp;	/* allocdirect being added */
1737 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
1738 {
1739 	struct worklist *wk;
1740 	struct freefrag *freefrag;
1741 	struct newdirblk *newdirblk;
1742 
1743 	mtx_assert(&lk, MA_OWNED);
1744 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1745 	    newadp->ad_oldsize != oldadp->ad_newsize ||
1746 	    newadp->ad_lbn >= NDADDR)
1747 		panic("%s %jd != new %jd || old size %ld != new %ld",
1748 		    "allocdirect_merge: old blkno",
1749 		    (intmax_t)newadp->ad_oldblkno,
1750 		    (intmax_t)oldadp->ad_newblkno,
1751 		    newadp->ad_oldsize, oldadp->ad_newsize);
1752 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
1753 	newadp->ad_oldsize = oldadp->ad_oldsize;
1754 	/*
1755 	 * If the old dependency had a fragment to free or had never
1756 	 * previously had a block allocated, then the new dependency
1757 	 * can immediately post its freefrag and adopt the old freefrag.
1758 	 * This action is done by swapping the freefrag dependencies.
1759 	 * The new dependency gains the old one's freefrag, and the
1760 	 * old one gets the new one and then immediately puts it on
1761 	 * the worklist when it is freed by free_allocdirect. It is
1762 	 * not possible to do this swap when the old dependency had a
1763 	 * non-zero size but no previous fragment to free. This condition
1764 	 * arises when the new block is an extension of the old block.
1765 	 * Here, the first part of the fragment allocated to the new
1766 	 * dependency is part of the block currently claimed on disk by
1767 	 * the old dependency, so cannot legitimately be freed until the
1768 	 * conditions for the new dependency are fulfilled.
1769 	 */
1770 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1771 		freefrag = newadp->ad_freefrag;
1772 		newadp->ad_freefrag = oldadp->ad_freefrag;
1773 		oldadp->ad_freefrag = freefrag;
1774 	}
1775 	/*
1776 	 * If we are tracking a new directory-block allocation,
1777 	 * move it from the old allocdirect to the new allocdirect.
1778 	 */
1779 	if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
1780 		newdirblk = WK_NEWDIRBLK(wk);
1781 		WORKLIST_REMOVE(&newdirblk->db_list);
1782 		if (!LIST_EMPTY(&oldadp->ad_newdirblk))
1783 			panic("allocdirect_merge: extra newdirblk");
1784 		WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list);
1785 	}
1786 	free_allocdirect(adphead, oldadp, 0);
1787 }
1788 
1789 /*
1790  * Allocate a new freefrag structure if needed.
1791  */
1792 static struct freefrag *
1793 newfreefrag(ip, blkno, size)
1794 	struct inode *ip;
1795 	ufs2_daddr_t blkno;
1796 	long size;
1797 {
1798 	struct freefrag *freefrag;
1799 	struct fs *fs;
1800 
1801 	if (blkno == 0)
1802 		return (NULL);
1803 	fs = ip->i_fs;
1804 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1805 		panic("newfreefrag: frag size");
1806 	MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1807 		M_FREEFRAG, M_SOFTDEP_FLAGS);
1808 	workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump));
1809 	freefrag->ff_inum = ip->i_number;
1810 	freefrag->ff_blkno = blkno;
1811 	freefrag->ff_fragsize = size;
1812 	return (freefrag);
1813 }
1814 
1815 /*
1816  * This workitem de-allocates fragments that were replaced during
1817  * file block allocation.
1818  */
1819 static void
1820 handle_workitem_freefrag(freefrag)
1821 	struct freefrag *freefrag;
1822 {
1823 	struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp);
1824 
1825 	ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
1826 	    freefrag->ff_fragsize, freefrag->ff_inum);
1827 	ACQUIRE_LOCK(&lk);
1828 	WORKITEM_FREE(freefrag, D_FREEFRAG);
1829 	FREE_LOCK(&lk);
1830 }
1831 
1832 /*
1833  * Set up a dependency structure for an external attributes data block.
1834  * This routine follows much of the structure of softdep_setup_allocdirect.
1835  * See the description of softdep_setup_allocdirect above for details.
1836  */
1837 void
1838 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1839 	struct inode *ip;
1840 	ufs_lbn_t lbn;
1841 	ufs2_daddr_t newblkno;
1842 	ufs2_daddr_t oldblkno;
1843 	long newsize;
1844 	long oldsize;
1845 	struct buf *bp;
1846 {
1847 	struct allocdirect *adp, *oldadp;
1848 	struct allocdirectlst *adphead;
1849 	struct bmsafemap *bmsafemap;
1850 	struct inodedep *inodedep;
1851 	struct newblk *newblk;
1852 	struct mount *mp;
1853 
1854 	mp = UFSTOVFS(ip->i_ump);
1855 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1856 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1857 	workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp);
1858 	adp->ad_lbn = lbn;
1859 	adp->ad_newblkno = newblkno;
1860 	adp->ad_oldblkno = oldblkno;
1861 	adp->ad_newsize = newsize;
1862 	adp->ad_oldsize = oldsize;
1863 	adp->ad_state = ATTACHED | EXTDATA;
1864 	LIST_INIT(&adp->ad_newdirblk);
1865 	if (newblkno == oldblkno)
1866 		adp->ad_freefrag = NULL;
1867 	else
1868 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1869 
1870 	ACQUIRE_LOCK(&lk);
1871 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1872 		panic("softdep_setup_allocext: lost block");
1873 
1874 	inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1875 	adp->ad_inodedep = inodedep;
1876 
1877 	if (newblk->nb_state == DEPCOMPLETE) {
1878 		adp->ad_state |= DEPCOMPLETE;
1879 		adp->ad_buf = NULL;
1880 	} else {
1881 		bmsafemap = newblk->nb_bmsafemap;
1882 		adp->ad_buf = bmsafemap->sm_buf;
1883 		LIST_REMOVE(newblk, nb_deps);
1884 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1885 	}
1886 	LIST_REMOVE(newblk, nb_hash);
1887 	FREE(newblk, M_NEWBLK);
1888 
1889 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1890 	if (lbn >= NXADDR)
1891 		panic("softdep_setup_allocext: lbn %lld > NXADDR",
1892 		    (long long)lbn);
1893 	/*
1894 	 * The list of allocdirects must be kept in sorted and ascending
1895 	 * order so that the rollback routines can quickly determine the
1896 	 * first uncommitted block (the size of the file stored on disk
1897 	 * ends at the end of the lowest committed fragment, or if there
1898 	 * are no fragments, at the end of the highest committed block).
1899 	 * Since files generally grow, the typical case is that the new
1900 	 * block is to be added at the end of the list. We speed this
1901 	 * special case by checking against the last allocdirect in the
1902 	 * list before laboriously traversing the list looking for the
1903 	 * insertion point.
1904 	 */
1905 	adphead = &inodedep->id_newextupdt;
1906 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1907 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1908 		/* insert at end of list */
1909 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1910 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1911 			allocdirect_merge(adphead, adp, oldadp);
1912 		FREE_LOCK(&lk);
1913 		return;
1914 	}
1915 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1916 		if (oldadp->ad_lbn >= lbn)
1917 			break;
1918 	}
1919 	if (oldadp == NULL)
1920 		panic("softdep_setup_allocext: lost entry");
1921 	/* insert in middle of list */
1922 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1923 	if (oldadp->ad_lbn == lbn)
1924 		allocdirect_merge(adphead, adp, oldadp);
1925 	FREE_LOCK(&lk);
1926 }
1927 
1928 /*
1929  * Indirect block allocation dependencies.
1930  *
1931  * The same dependencies that exist for a direct block also exist when
1932  * a new block is allocated and pointed to by an entry in a block of
1933  * indirect pointers. The undo/redo states described above are also
1934  * used here. Because an indirect block contains many pointers that
1935  * may have dependencies, a second copy of the entire in-memory indirect
1936  * block is kept. The buffer cache copy is always completely up-to-date.
1937  * The second copy, which is used only as a source for disk writes,
1938  * contains only the safe pointers (i.e., those that have no remaining
1939  * update dependencies). The second copy is freed when all pointers
1940  * are safe. The cache is not allowed to replace indirect blocks with
1941  * pending update dependencies. If a buffer containing an indirect
1942  * block with dependencies is written, these routines will mark it
1943  * dirty again. It can only be successfully written once all the
1944  * dependencies are removed. The ffs_fsync routine in conjunction with
1945  * softdep_sync_metadata work together to get all the dependencies
1946  * removed so that a file can be successfully written to disk. Three
1947  * procedures are used when setting up indirect block pointer
1948  * dependencies. The division is necessary because of the organization
1949  * of the "balloc" routine and because of the distinction between file
1950  * pages and file metadata blocks.
1951  */
1952 
1953 /*
1954  * Allocate a new allocindir structure.
1955  */
1956 static struct allocindir *
1957 newallocindir(ip, ptrno, newblkno, oldblkno)
1958 	struct inode *ip;	/* inode for file being extended */
1959 	int ptrno;		/* offset of pointer in indirect block */
1960 	ufs2_daddr_t newblkno;	/* disk block number being added */
1961 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1962 {
1963 	struct allocindir *aip;
1964 
1965 	MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1966 		M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO);
1967 	workitem_alloc(&aip->ai_list, D_ALLOCINDIR, UFSTOVFS(ip->i_ump));
1968 	aip->ai_state = ATTACHED;
1969 	aip->ai_offset = ptrno;
1970 	aip->ai_newblkno = newblkno;
1971 	aip->ai_oldblkno = oldblkno;
1972 	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1973 	return (aip);
1974 }
1975 
1976 /*
1977  * Called just before setting an indirect block pointer
1978  * to a newly allocated file page.
1979  */
1980 void
1981 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1982 	struct inode *ip;	/* inode for file being extended */
1983 	ufs_lbn_t lbn;		/* allocated block number within file */
1984 	struct buf *bp;		/* buffer with indirect blk referencing page */
1985 	int ptrno;		/* offset of pointer in indirect block */
1986 	ufs2_daddr_t newblkno;	/* disk block number being added */
1987 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1988 	struct buf *nbp;	/* buffer holding allocated page */
1989 {
1990 	struct allocindir *aip;
1991 	struct pagedep *pagedep;
1992 
1993 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page");
1994 	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1995 	ACQUIRE_LOCK(&lk);
1996 	/*
1997 	 * If we are allocating a directory page, then we must
1998 	 * allocate an associated pagedep to track additions and
1999 	 * deletions.
2000 	 */
2001 	if ((ip->i_mode & IFMT) == IFDIR &&
2002 	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
2003 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
2004 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
2005 	setup_allocindir_phase2(bp, ip, aip);
2006 	FREE_LOCK(&lk);
2007 }
2008 
2009 /*
2010  * Called just before setting an indirect block pointer to a
2011  * newly allocated indirect block.
2012  */
2013 void
2014 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
2015 	struct buf *nbp;	/* newly allocated indirect block */
2016 	struct inode *ip;	/* inode for file being extended */
2017 	struct buf *bp;		/* indirect block referencing allocated block */
2018 	int ptrno;		/* offset of pointer in indirect block */
2019 	ufs2_daddr_t newblkno;	/* disk block number being added */
2020 {
2021 	struct allocindir *aip;
2022 
2023 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta");
2024 	aip = newallocindir(ip, ptrno, newblkno, 0);
2025 	ACQUIRE_LOCK(&lk);
2026 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
2027 	setup_allocindir_phase2(bp, ip, aip);
2028 	FREE_LOCK(&lk);
2029 }
2030 
2031 /*
2032  * Called to finish the allocation of the "aip" allocated
2033  * by one of the two routines above.
2034  */
2035 static void
2036 setup_allocindir_phase2(bp, ip, aip)
2037 	struct buf *bp;		/* in-memory copy of the indirect block */
2038 	struct inode *ip;	/* inode for file being extended */
2039 	struct allocindir *aip;	/* allocindir allocated by the above routines */
2040 {
2041 	struct worklist *wk;
2042 	struct indirdep *indirdep, *newindirdep;
2043 	struct bmsafemap *bmsafemap;
2044 	struct allocindir *oldaip;
2045 	struct freefrag *freefrag;
2046 	struct newblk *newblk;
2047 	ufs2_daddr_t blkno;
2048 
2049 	mtx_assert(&lk, MA_OWNED);
2050 	if (bp->b_lblkno >= 0)
2051 		panic("setup_allocindir_phase2: not indir blk");
2052 	for (indirdep = NULL, newindirdep = NULL; ; ) {
2053 		LIST_FOREACH(wk, &bp->b_dep, wk_list) {
2054 			if (wk->wk_type != D_INDIRDEP)
2055 				continue;
2056 			indirdep = WK_INDIRDEP(wk);
2057 			break;
2058 		}
2059 		if (indirdep == NULL && newindirdep) {
2060 			indirdep = newindirdep;
2061 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
2062 			newindirdep = NULL;
2063 		}
2064 		if (indirdep) {
2065 			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
2066 			    &newblk) == 0)
2067 				panic("setup_allocindir: lost block");
2068 			if (newblk->nb_state == DEPCOMPLETE) {
2069 				aip->ai_state |= DEPCOMPLETE;
2070 				aip->ai_buf = NULL;
2071 			} else {
2072 				bmsafemap = newblk->nb_bmsafemap;
2073 				aip->ai_buf = bmsafemap->sm_buf;
2074 				LIST_REMOVE(newblk, nb_deps);
2075 				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
2076 				    aip, ai_deps);
2077 			}
2078 			LIST_REMOVE(newblk, nb_hash);
2079 			FREE(newblk, M_NEWBLK);
2080 			aip->ai_indirdep = indirdep;
2081 			/*
2082 			 * Check to see if there is an existing dependency
2083 			 * for this block. If there is, merge the old
2084 			 * dependency into the new one.
2085 			 */
2086 			if (aip->ai_oldblkno == 0)
2087 				oldaip = NULL;
2088 			else
2089 
2090 				LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
2091 					if (oldaip->ai_offset == aip->ai_offset)
2092 						break;
2093 			freefrag = NULL;
2094 			if (oldaip != NULL) {
2095 				if (oldaip->ai_newblkno != aip->ai_oldblkno)
2096 					panic("setup_allocindir_phase2: blkno");
2097 				aip->ai_oldblkno = oldaip->ai_oldblkno;
2098 				freefrag = aip->ai_freefrag;
2099 				aip->ai_freefrag = oldaip->ai_freefrag;
2100 				oldaip->ai_freefrag = NULL;
2101 				free_allocindir(oldaip, NULL);
2102 			}
2103 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
2104 			if (ip->i_ump->um_fstype == UFS1)
2105 				((ufs1_daddr_t *)indirdep->ir_savebp->b_data)
2106 				    [aip->ai_offset] = aip->ai_oldblkno;
2107 			else
2108 				((ufs2_daddr_t *)indirdep->ir_savebp->b_data)
2109 				    [aip->ai_offset] = aip->ai_oldblkno;
2110 			FREE_LOCK(&lk);
2111 			if (freefrag != NULL)
2112 				handle_workitem_freefrag(freefrag);
2113 		} else
2114 			FREE_LOCK(&lk);
2115 		if (newindirdep) {
2116 			newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
2117 			brelse(newindirdep->ir_savebp);
2118 			ACQUIRE_LOCK(&lk);
2119 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
2120 			if (indirdep)
2121 				break;
2122 			FREE_LOCK(&lk);
2123 		}
2124 		if (indirdep) {
2125 			ACQUIRE_LOCK(&lk);
2126 			break;
2127 		}
2128 		MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
2129 			M_INDIRDEP, M_SOFTDEP_FLAGS);
2130 		workitem_alloc(&newindirdep->ir_list, D_INDIRDEP,
2131 		    UFSTOVFS(ip->i_ump));
2132 		newindirdep->ir_state = ATTACHED;
2133 		if (ip->i_ump->um_fstype == UFS1)
2134 			newindirdep->ir_state |= UFS1FMT;
2135 		LIST_INIT(&newindirdep->ir_deplisthd);
2136 		LIST_INIT(&newindirdep->ir_donehd);
2137 		if (bp->b_blkno == bp->b_lblkno) {
2138 			ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
2139 			    NULL, NULL);
2140 			bp->b_blkno = blkno;
2141 		}
2142 		newindirdep->ir_savebp =
2143 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
2144 		BUF_KERNPROC(newindirdep->ir_savebp);
2145 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
2146 		ACQUIRE_LOCK(&lk);
2147 	}
2148 }
2149 
2150 /*
2151  * Block de-allocation dependencies.
2152  *
2153  * When blocks are de-allocated, the on-disk pointers must be nullified before
2154  * the blocks are made available for use by other files.  (The true
2155  * requirement is that old pointers must be nullified before new on-disk
2156  * pointers are set.  We chose this slightly more stringent requirement to
2157  * reduce complexity.) Our implementation handles this dependency by updating
2158  * the inode (or indirect block) appropriately but delaying the actual block
2159  * de-allocation (i.e., freemap and free space count manipulation) until
2160  * after the updated versions reach stable storage.  After the disk is
2161  * updated, the blocks can be safely de-allocated whenever it is convenient.
2162  * This implementation handles only the common case of reducing a file's
2163  * length to zero. Other cases are handled by the conventional synchronous
2164  * write approach.
2165  *
2166  * The ffs implementation with which we worked double-checks
2167  * the state of the block pointers and file size as it reduces
2168  * a file's length.  Some of this code is replicated here in our
2169  * soft updates implementation.  The freeblks->fb_chkcnt field is
2170  * used to transfer a part of this information to the procedure
2171  * that eventually de-allocates the blocks.
2172  *
2173  * This routine should be called from the routine that shortens
2174  * a file's length, before the inode's size or block pointers
2175  * are modified. It will save the block pointer information for
2176  * later release and zero the inode so that the calling routine
2177  * can release it.
2178  */
2179 void
2180 softdep_setup_freeblocks(ip, length, flags)
2181 	struct inode *ip;	/* The inode whose length is to be reduced */
2182 	off_t length;		/* The new length for the file */
2183 	int flags;		/* IO_EXT and/or IO_NORMAL */
2184 {
2185 	struct freeblks *freeblks;
2186 	struct inodedep *inodedep;
2187 	struct allocdirect *adp;
2188 	struct vnode *vp;
2189 	struct buf *bp;
2190 	struct fs *fs;
2191 	ufs2_daddr_t extblocks, datablocks;
2192 	struct mount *mp;
2193 	int i, delay, error;
2194 
2195 	fs = ip->i_fs;
2196 	mp = UFSTOVFS(ip->i_ump);
2197 	if (length != 0)
2198 		panic("softdep_setup_freeblocks: non-zero length");
2199 	MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
2200 		M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
2201 	workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp);
2202 	freeblks->fb_state = ATTACHED;
2203 	freeblks->fb_uid = ip->i_uid;
2204 	freeblks->fb_previousinum = ip->i_number;
2205 	freeblks->fb_devvp = ip->i_devvp;
2206 	extblocks = 0;
2207 	if (fs->fs_magic == FS_UFS2_MAGIC)
2208 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
2209 	datablocks = DIP(ip, i_blocks) - extblocks;
2210 	if ((flags & IO_NORMAL) == 0) {
2211 		freeblks->fb_oldsize = 0;
2212 		freeblks->fb_chkcnt = 0;
2213 	} else {
2214 		freeblks->fb_oldsize = ip->i_size;
2215 		ip->i_size = 0;
2216 		DIP_SET(ip, i_size, 0);
2217 		freeblks->fb_chkcnt = datablocks;
2218 		for (i = 0; i < NDADDR; i++) {
2219 			freeblks->fb_dblks[i] = DIP(ip, i_db[i]);
2220 			DIP_SET(ip, i_db[i], 0);
2221 		}
2222 		for (i = 0; i < NIADDR; i++) {
2223 			freeblks->fb_iblks[i] = DIP(ip, i_ib[i]);
2224 			DIP_SET(ip, i_ib[i], 0);
2225 		}
2226 		/*
2227 		 * If the file was removed, then the space being freed was
2228 		 * accounted for then (see softdep_releasefile()). If the
2229 		 * file is merely being truncated, then we account for it now.
2230 		 */
2231 		if ((ip->i_flag & IN_SPACECOUNTED) == 0) {
2232 			UFS_LOCK(ip->i_ump);
2233 			fs->fs_pendingblocks += datablocks;
2234 			UFS_UNLOCK(ip->i_ump);
2235 		}
2236 	}
2237 	if ((flags & IO_EXT) == 0) {
2238 		freeblks->fb_oldextsize = 0;
2239 	} else {
2240 		freeblks->fb_oldextsize = ip->i_din2->di_extsize;
2241 		ip->i_din2->di_extsize = 0;
2242 		freeblks->fb_chkcnt += extblocks;
2243 		for (i = 0; i < NXADDR; i++) {
2244 			freeblks->fb_eblks[i] = ip->i_din2->di_extb[i];
2245 			ip->i_din2->di_extb[i] = 0;
2246 		}
2247 	}
2248 	DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt);
2249 	/*
2250 	 * Push the zero'ed inode to to its disk buffer so that we are free
2251 	 * to delete its dependencies below. Once the dependencies are gone
2252 	 * the buffer can be safely released.
2253 	 */
2254 	if ((error = bread(ip->i_devvp,
2255 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
2256 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
2257 		brelse(bp);
2258 		softdep_error("softdep_setup_freeblocks", error);
2259 	}
2260 	if (ip->i_ump->um_fstype == UFS1)
2261 		*((struct ufs1_dinode *)bp->b_data +
2262 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
2263 	else
2264 		*((struct ufs2_dinode *)bp->b_data +
2265 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
2266 	/*
2267 	 * Find and eliminate any inode dependencies.
2268 	 */
2269 	ACQUIRE_LOCK(&lk);
2270 	(void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
2271 	if ((inodedep->id_state & IOSTARTED) != 0)
2272 		panic("softdep_setup_freeblocks: inode busy");
2273 	/*
2274 	 * Add the freeblks structure to the list of operations that
2275 	 * must await the zero'ed inode being written to disk. If we
2276 	 * still have a bitmap dependency (delay == 0), then the inode
2277 	 * has never been written to disk, so we can process the
2278 	 * freeblks below once we have deleted the dependencies.
2279 	 */
2280 	delay = (inodedep->id_state & DEPCOMPLETE);
2281 	if (delay)
2282 		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
2283 	/*
2284 	 * Because the file length has been truncated to zero, any
2285 	 * pending block allocation dependency structures associated
2286 	 * with this inode are obsolete and can simply be de-allocated.
2287 	 * We must first merge the two dependency lists to get rid of
2288 	 * any duplicate freefrag structures, then purge the merged list.
2289 	 * If we still have a bitmap dependency, then the inode has never
2290 	 * been written to disk, so we can free any fragments without delay.
2291 	 */
2292 	if (flags & IO_NORMAL) {
2293 		merge_inode_lists(&inodedep->id_newinoupdt,
2294 		    &inodedep->id_inoupdt);
2295 		while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
2296 			free_allocdirect(&inodedep->id_inoupdt, adp, delay);
2297 	}
2298 	if (flags & IO_EXT) {
2299 		merge_inode_lists(&inodedep->id_newextupdt,
2300 		    &inodedep->id_extupdt);
2301 		while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
2302 			free_allocdirect(&inodedep->id_extupdt, adp, delay);
2303 	}
2304 	FREE_LOCK(&lk);
2305 	bdwrite(bp);
2306 	/*
2307 	 * We must wait for any I/O in progress to finish so that
2308 	 * all potential buffers on the dirty list will be visible.
2309 	 * Once they are all there, walk the list and get rid of
2310 	 * any dependencies.
2311 	 */
2312 	vp = ITOV(ip);
2313 	VI_LOCK(vp);
2314 	drain_output(vp);
2315 restart:
2316 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
2317 		if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
2318 		    ((flags & IO_NORMAL) == 0 &&
2319 		      (bp->b_xflags & BX_ALTDATA) == 0))
2320 			continue;
2321 		if ((bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT)) == NULL)
2322 			goto restart;
2323 		VI_UNLOCK(vp);
2324 		ACQUIRE_LOCK(&lk);
2325 		(void) inodedep_lookup(mp, ip->i_number, 0, &inodedep);
2326 		deallocate_dependencies(bp, inodedep);
2327 		FREE_LOCK(&lk);
2328 		bp->b_flags |= B_INVAL | B_NOCACHE;
2329 		brelse(bp);
2330 		VI_LOCK(vp);
2331 		goto restart;
2332 	}
2333 	VI_UNLOCK(vp);
2334 	ACQUIRE_LOCK(&lk);
2335 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
2336 		(void) free_inodedep(inodedep);
2337 
2338 	if(delay) {
2339 		freeblks->fb_state |= DEPCOMPLETE;
2340 		/*
2341 		 * If the inode with zeroed block pointers is now on disk
2342 		 * we can start freeing blocks. Add freeblks to the worklist
2343 		 * instead of calling  handle_workitem_freeblocks directly as
2344 		 * it is more likely that additional IO is needed to complete
2345 		 * the request here than in the !delay case.
2346 		 */
2347 		if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
2348 			add_to_worklist(&freeblks->fb_list);
2349 	}
2350 
2351 	FREE_LOCK(&lk);
2352 	/*
2353 	 * If the inode has never been written to disk (delay == 0),
2354 	 * then we can process the freeblks now that we have deleted
2355 	 * the dependencies.
2356 	 */
2357 	if (!delay)
2358 		handle_workitem_freeblocks(freeblks, 0);
2359 }
2360 
2361 /*
2362  * Reclaim any dependency structures from a buffer that is about to
2363  * be reallocated to a new vnode. The buffer must be locked, thus,
2364  * no I/O completion operations can occur while we are manipulating
2365  * its associated dependencies. The mutex is held so that other I/O's
2366  * associated with related dependencies do not occur.
2367  */
2368 static void
2369 deallocate_dependencies(bp, inodedep)
2370 	struct buf *bp;
2371 	struct inodedep *inodedep;
2372 {
2373 	struct worklist *wk;
2374 	struct indirdep *indirdep;
2375 	struct allocindir *aip;
2376 	struct pagedep *pagedep;
2377 	struct dirrem *dirrem;
2378 	struct diradd *dap;
2379 	int i;
2380 
2381 	mtx_assert(&lk, MA_OWNED);
2382 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2383 		switch (wk->wk_type) {
2384 
2385 		case D_INDIRDEP:
2386 			indirdep = WK_INDIRDEP(wk);
2387 			/*
2388 			 * None of the indirect pointers will ever be visible,
2389 			 * so they can simply be tossed. GOINGAWAY ensures
2390 			 * that allocated pointers will be saved in the buffer
2391 			 * cache until they are freed. Note that they will
2392 			 * only be able to be found by their physical address
2393 			 * since the inode mapping the logical address will
2394 			 * be gone. The save buffer used for the safe copy
2395 			 * was allocated in setup_allocindir_phase2 using
2396 			 * the physical address so it could be used for this
2397 			 * purpose. Hence we swap the safe copy with the real
2398 			 * copy, allowing the safe copy to be freed and holding
2399 			 * on to the real copy for later use in indir_trunc.
2400 			 */
2401 			if (indirdep->ir_state & GOINGAWAY)
2402 				panic("deallocate_dependencies: already gone");
2403 			indirdep->ir_state |= GOINGAWAY;
2404 			VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1;
2405 			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
2406 				free_allocindir(aip, inodedep);
2407 			if (bp->b_lblkno >= 0 ||
2408 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno)
2409 				panic("deallocate_dependencies: not indir");
2410 			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2411 			    bp->b_bcount);
2412 			WORKLIST_REMOVE(wk);
2413 			WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
2414 			continue;
2415 
2416 		case D_PAGEDEP:
2417 			pagedep = WK_PAGEDEP(wk);
2418 			/*
2419 			 * None of the directory additions will ever be
2420 			 * visible, so they can simply be tossed.
2421 			 */
2422 			for (i = 0; i < DAHASHSZ; i++)
2423 				while ((dap =
2424 				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
2425 					free_diradd(dap);
2426 			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2427 				free_diradd(dap);
2428 			/*
2429 			 * Copy any directory remove dependencies to the list
2430 			 * to be processed after the zero'ed inode is written.
2431 			 * If the inode has already been written, then they
2432 			 * can be dumped directly onto the work list.
2433 			 */
2434 			LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2435 				LIST_REMOVE(dirrem, dm_next);
2436 				dirrem->dm_dirinum = pagedep->pd_ino;
2437 				if (inodedep == NULL ||
2438 				    (inodedep->id_state & ALLCOMPLETE) ==
2439 				     ALLCOMPLETE)
2440 					add_to_worklist(&dirrem->dm_list);
2441 				else
2442 					WORKLIST_INSERT(&inodedep->id_bufwait,
2443 					    &dirrem->dm_list);
2444 			}
2445 			if ((pagedep->pd_state & NEWBLOCK) != 0) {
2446 				LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list)
2447 					if (wk->wk_type == D_NEWDIRBLK &&
2448 					    WK_NEWDIRBLK(wk)->db_pagedep ==
2449 					      pagedep)
2450 						break;
2451 				if (wk != NULL) {
2452 					WORKLIST_REMOVE(wk);
2453 					free_newdirblk(WK_NEWDIRBLK(wk));
2454 				} else
2455 					panic("deallocate_dependencies: "
2456 					      "lost pagedep");
2457 			}
2458 			WORKLIST_REMOVE(&pagedep->pd_list);
2459 			LIST_REMOVE(pagedep, pd_hash);
2460 			WORKITEM_FREE(pagedep, D_PAGEDEP);
2461 			continue;
2462 
2463 		case D_ALLOCINDIR:
2464 			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2465 			continue;
2466 
2467 		case D_ALLOCDIRECT:
2468 		case D_INODEDEP:
2469 			panic("deallocate_dependencies: Unexpected type %s",
2470 			    TYPENAME(wk->wk_type));
2471 			/* NOTREACHED */
2472 
2473 		default:
2474 			panic("deallocate_dependencies: Unknown type %s",
2475 			    TYPENAME(wk->wk_type));
2476 			/* NOTREACHED */
2477 		}
2478 	}
2479 }
2480 
2481 /*
2482  * Free an allocdirect. Generate a new freefrag work request if appropriate.
2483  * This routine must be called with splbio interrupts blocked.
2484  */
2485 static void
2486 free_allocdirect(adphead, adp, delay)
2487 	struct allocdirectlst *adphead;
2488 	struct allocdirect *adp;
2489 	int delay;
2490 {
2491 	struct newdirblk *newdirblk;
2492 	struct worklist *wk;
2493 
2494 	mtx_assert(&lk, MA_OWNED);
2495 	if ((adp->ad_state & DEPCOMPLETE) == 0)
2496 		LIST_REMOVE(adp, ad_deps);
2497 	TAILQ_REMOVE(adphead, adp, ad_next);
2498 	if ((adp->ad_state & COMPLETE) == 0)
2499 		WORKLIST_REMOVE(&adp->ad_list);
2500 	if (adp->ad_freefrag != NULL) {
2501 		if (delay)
2502 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2503 			    &adp->ad_freefrag->ff_list);
2504 		else
2505 			add_to_worklist(&adp->ad_freefrag->ff_list);
2506 	}
2507 	if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) {
2508 		newdirblk = WK_NEWDIRBLK(wk);
2509 		WORKLIST_REMOVE(&newdirblk->db_list);
2510 		if (!LIST_EMPTY(&adp->ad_newdirblk))
2511 			panic("free_allocdirect: extra newdirblk");
2512 		if (delay)
2513 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2514 			    &newdirblk->db_list);
2515 		else
2516 			free_newdirblk(newdirblk);
2517 	}
2518 	WORKITEM_FREE(adp, D_ALLOCDIRECT);
2519 }
2520 
2521 /*
2522  * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
2523  * This routine must be called with splbio interrupts blocked.
2524  */
2525 static void
2526 free_newdirblk(newdirblk)
2527 	struct newdirblk *newdirblk;
2528 {
2529 	struct pagedep *pagedep;
2530 	struct diradd *dap;
2531 	int i;
2532 
2533 	mtx_assert(&lk, MA_OWNED);
2534 	/*
2535 	 * If the pagedep is still linked onto the directory buffer
2536 	 * dependency chain, then some of the entries on the
2537 	 * pd_pendinghd list may not be committed to disk yet. In
2538 	 * this case, we will simply clear the NEWBLOCK flag and
2539 	 * let the pd_pendinghd list be processed when the pagedep
2540 	 * is next written. If the pagedep is no longer on the buffer
2541 	 * dependency chain, then all the entries on the pd_pending
2542 	 * list are committed to disk and we can free them here.
2543 	 */
2544 	pagedep = newdirblk->db_pagedep;
2545 	pagedep->pd_state &= ~NEWBLOCK;
2546 	if ((pagedep->pd_state & ONWORKLIST) == 0)
2547 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
2548 			free_diradd(dap);
2549 	/*
2550 	 * If no dependencies remain, the pagedep will be freed.
2551 	 */
2552 	for (i = 0; i < DAHASHSZ; i++)
2553 		if (!LIST_EMPTY(&pagedep->pd_diraddhd[i]))
2554 			break;
2555 	if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) {
2556 		LIST_REMOVE(pagedep, pd_hash);
2557 		WORKITEM_FREE(pagedep, D_PAGEDEP);
2558 	}
2559 	WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2560 }
2561 
2562 /*
2563  * Prepare an inode to be freed. The actual free operation is not
2564  * done until the zero'ed inode has been written to disk.
2565  */
2566 void
2567 softdep_freefile(pvp, ino, mode)
2568 	struct vnode *pvp;
2569 	ino_t ino;
2570 	int mode;
2571 {
2572 	struct inode *ip = VTOI(pvp);
2573 	struct inodedep *inodedep;
2574 	struct freefile *freefile;
2575 
2576 	/*
2577 	 * This sets up the inode de-allocation dependency.
2578 	 */
2579 	MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2580 		M_FREEFILE, M_SOFTDEP_FLAGS);
2581 	workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount);
2582 	freefile->fx_mode = mode;
2583 	freefile->fx_oldinum = ino;
2584 	freefile->fx_devvp = ip->i_devvp;
2585 	if ((ip->i_flag & IN_SPACECOUNTED) == 0) {
2586 		UFS_LOCK(ip->i_ump);
2587 		ip->i_fs->fs_pendinginodes += 1;
2588 		UFS_UNLOCK(ip->i_ump);
2589 	}
2590 
2591 	/*
2592 	 * If the inodedep does not exist, then the zero'ed inode has
2593 	 * been written to disk. If the allocated inode has never been
2594 	 * written to disk, then the on-disk inode is zero'ed. In either
2595 	 * case we can free the file immediately.
2596 	 */
2597 	ACQUIRE_LOCK(&lk);
2598 	if (inodedep_lookup(pvp->v_mount, ino, 0, &inodedep) == 0 ||
2599 	    check_inode_unwritten(inodedep)) {
2600 		FREE_LOCK(&lk);
2601 		handle_workitem_freefile(freefile);
2602 		return;
2603 	}
2604 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2605 	FREE_LOCK(&lk);
2606 	ip->i_flag |= IN_MODIFIED;
2607 }
2608 
2609 /*
2610  * Check to see if an inode has never been written to disk. If
2611  * so free the inodedep and return success, otherwise return failure.
2612  * This routine must be called with splbio interrupts blocked.
2613  *
2614  * If we still have a bitmap dependency, then the inode has never
2615  * been written to disk. Drop the dependency as it is no longer
2616  * necessary since the inode is being deallocated. We set the
2617  * ALLCOMPLETE flags since the bitmap now properly shows that the
2618  * inode is not allocated. Even if the inode is actively being
2619  * written, it has been rolled back to its zero'ed state, so we
2620  * are ensured that a zero inode is what is on the disk. For short
2621  * lived files, this change will usually result in removing all the
2622  * dependencies from the inode so that it can be freed immediately.
2623  */
2624 static int
2625 check_inode_unwritten(inodedep)
2626 	struct inodedep *inodedep;
2627 {
2628 
2629 	mtx_assert(&lk, MA_OWNED);
2630 	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2631 	    !LIST_EMPTY(&inodedep->id_pendinghd) ||
2632 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
2633 	    !LIST_EMPTY(&inodedep->id_inowait) ||
2634 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
2635 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
2636 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
2637 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
2638 	    inodedep->id_nlinkdelta != 0)
2639 		return (0);
2640 
2641 	/*
2642 	 * Another process might be in initiate_write_inodeblock_ufs[12]
2643 	 * trying to allocate memory without holding "Softdep Lock".
2644 	 */
2645 	if ((inodedep->id_state & IOSTARTED) != 0 &&
2646 	    inodedep->id_savedino1 == NULL)
2647 		return (0);
2648 
2649 	inodedep->id_state |= ALLCOMPLETE;
2650 	LIST_REMOVE(inodedep, id_deps);
2651 	inodedep->id_buf = NULL;
2652 	if (inodedep->id_state & ONWORKLIST)
2653 		WORKLIST_REMOVE(&inodedep->id_list);
2654 	if (inodedep->id_savedino1 != NULL) {
2655 		FREE(inodedep->id_savedino1, M_SAVEDINO);
2656 		inodedep->id_savedino1 = NULL;
2657 	}
2658 	if (free_inodedep(inodedep) == 0)
2659 		panic("check_inode_unwritten: busy inode");
2660 	return (1);
2661 }
2662 
2663 /*
2664  * Try to free an inodedep structure. Return 1 if it could be freed.
2665  */
2666 static int
2667 free_inodedep(inodedep)
2668 	struct inodedep *inodedep;
2669 {
2670 
2671 	mtx_assert(&lk, MA_OWNED);
2672 	if ((inodedep->id_state & ONWORKLIST) != 0 ||
2673 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2674 	    !LIST_EMPTY(&inodedep->id_pendinghd) ||
2675 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
2676 	    !LIST_EMPTY(&inodedep->id_inowait) ||
2677 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
2678 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
2679 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
2680 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
2681 	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL)
2682 		return (0);
2683 	LIST_REMOVE(inodedep, id_hash);
2684 	WORKITEM_FREE(inodedep, D_INODEDEP);
2685 	num_inodedep -= 1;
2686 	return (1);
2687 }
2688 
2689 /*
2690  * This workitem routine performs the block de-allocation.
2691  * The workitem is added to the pending list after the updated
2692  * inode block has been written to disk.  As mentioned above,
2693  * checks regarding the number of blocks de-allocated (compared
2694  * to the number of blocks allocated for the file) are also
2695  * performed in this function.
2696  */
2697 static void
2698 handle_workitem_freeblocks(freeblks, flags)
2699 	struct freeblks *freeblks;
2700 	int flags;
2701 {
2702 	struct inode *ip;
2703 	struct vnode *vp;
2704 	struct fs *fs;
2705 	struct ufsmount *ump;
2706 	int i, nblocks, level, bsize;
2707 	ufs2_daddr_t bn, blocksreleased = 0;
2708 	int error, allerror = 0;
2709 	ufs_lbn_t baselbns[NIADDR], tmpval;
2710 	int fs_pendingblocks;
2711 
2712 	ump = VFSTOUFS(freeblks->fb_list.wk_mp);
2713 	fs = ump->um_fs;
2714 	fs_pendingblocks = 0;
2715 	tmpval = 1;
2716 	baselbns[0] = NDADDR;
2717 	for (i = 1; i < NIADDR; i++) {
2718 		tmpval *= NINDIR(fs);
2719 		baselbns[i] = baselbns[i - 1] + tmpval;
2720 	}
2721 	nblocks = btodb(fs->fs_bsize);
2722 	blocksreleased = 0;
2723 	/*
2724 	 * Release all extended attribute blocks or frags.
2725 	 */
2726 	if (freeblks->fb_oldextsize > 0) {
2727 		for (i = (NXADDR - 1); i >= 0; i--) {
2728 			if ((bn = freeblks->fb_eblks[i]) == 0)
2729 				continue;
2730 			bsize = sblksize(fs, freeblks->fb_oldextsize, i);
2731 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize,
2732 			    freeblks->fb_previousinum);
2733 			blocksreleased += btodb(bsize);
2734 		}
2735 	}
2736 	/*
2737 	 * Release all data blocks or frags.
2738 	 */
2739 	if (freeblks->fb_oldsize > 0) {
2740 		/*
2741 		 * Indirect blocks first.
2742 		 */
2743 		for (level = (NIADDR - 1); level >= 0; level--) {
2744 			if ((bn = freeblks->fb_iblks[level]) == 0)
2745 				continue;
2746 			if ((error = indir_trunc(freeblks, fsbtodb(fs, bn),
2747 			    level, baselbns[level], &blocksreleased)) != 0)
2748 				allerror = error;
2749 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn,
2750 			    fs->fs_bsize, freeblks->fb_previousinum);
2751 			fs_pendingblocks += nblocks;
2752 			blocksreleased += nblocks;
2753 		}
2754 		/*
2755 		 * All direct blocks or frags.
2756 		 */
2757 		for (i = (NDADDR - 1); i >= 0; i--) {
2758 			if ((bn = freeblks->fb_dblks[i]) == 0)
2759 				continue;
2760 			bsize = sblksize(fs, freeblks->fb_oldsize, i);
2761 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize,
2762 			    freeblks->fb_previousinum);
2763 			fs_pendingblocks += btodb(bsize);
2764 			blocksreleased += btodb(bsize);
2765 		}
2766 	}
2767 	UFS_LOCK(ump);
2768 	fs->fs_pendingblocks -= fs_pendingblocks;
2769 	UFS_UNLOCK(ump);
2770 	/*
2771 	 * If we still have not finished background cleanup, then check
2772 	 * to see if the block count needs to be adjusted.
2773 	 */
2774 	if (freeblks->fb_chkcnt != blocksreleased &&
2775 	    (fs->fs_flags & FS_UNCLEAN) != 0 &&
2776 	    ffs_vget(freeblks->fb_list.wk_mp, freeblks->fb_previousinum,
2777 	    (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) {
2778 		ip = VTOI(vp);
2779 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + \
2780 		    freeblks->fb_chkcnt - blocksreleased);
2781 		ip->i_flag |= IN_CHANGE;
2782 		vput(vp);
2783 	}
2784 
2785 #ifdef DIAGNOSTIC
2786 	if (freeblks->fb_chkcnt != blocksreleased &&
2787 	    ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0))
2788 		printf("handle_workitem_freeblocks: block count\n");
2789 	if (allerror)
2790 		softdep_error("handle_workitem_freeblks", allerror);
2791 #endif /* DIAGNOSTIC */
2792 
2793 	ACQUIRE_LOCK(&lk);
2794 	WORKITEM_FREE(freeblks, D_FREEBLKS);
2795 	FREE_LOCK(&lk);
2796 }
2797 
2798 /*
2799  * Release blocks associated with the inode ip and stored in the indirect
2800  * block dbn. If level is greater than SINGLE, the block is an indirect block
2801  * and recursive calls to indirtrunc must be used to cleanse other indirect
2802  * blocks.
2803  */
2804 static int
2805 indir_trunc(freeblks, dbn, level, lbn, countp)
2806 	struct freeblks *freeblks;
2807 	ufs2_daddr_t dbn;
2808 	int level;
2809 	ufs_lbn_t lbn;
2810 	ufs2_daddr_t *countp;
2811 {
2812 	struct buf *bp;
2813 	struct fs *fs;
2814 	struct worklist *wk;
2815 	struct indirdep *indirdep;
2816 	struct ufsmount *ump;
2817 	ufs1_daddr_t *bap1 = 0;
2818 	ufs2_daddr_t nb, *bap2 = 0;
2819 	ufs_lbn_t lbnadd;
2820 	int i, nblocks, ufs1fmt;
2821 	int error, allerror = 0;
2822 	int fs_pendingblocks;
2823 
2824 	ump = VFSTOUFS(freeblks->fb_list.wk_mp);
2825 	fs = ump->um_fs;
2826 	fs_pendingblocks = 0;
2827 	lbnadd = 1;
2828 	for (i = level; i > 0; i--)
2829 		lbnadd *= NINDIR(fs);
2830 	/*
2831 	 * Get buffer of block pointers to be freed. This routine is not
2832 	 * called until the zero'ed inode has been written, so it is safe
2833 	 * to free blocks as they are encountered. Because the inode has
2834 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
2835 	 * have to use the on-disk address and the block device for the
2836 	 * filesystem to look them up. If the file was deleted before its
2837 	 * indirect blocks were all written to disk, the routine that set
2838 	 * us up (deallocate_dependencies) will have arranged to leave
2839 	 * a complete copy of the indirect block in memory for our use.
2840 	 * Otherwise we have to read the blocks in from the disk.
2841 	 */
2842 #ifdef notyet
2843 	bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0,
2844 	    GB_NOCREAT);
2845 #else
2846 	bp = incore(&freeblks->fb_devvp->v_bufobj, dbn);
2847 #endif
2848 	ACQUIRE_LOCK(&lk);
2849 	if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2850 		if (wk->wk_type != D_INDIRDEP ||
2851 		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2852 		    (indirdep->ir_state & GOINGAWAY) == 0)
2853 			panic("indir_trunc: lost indirdep");
2854 		WORKLIST_REMOVE(wk);
2855 		WORKITEM_FREE(indirdep, D_INDIRDEP);
2856 		if (!LIST_EMPTY(&bp->b_dep))
2857 			panic("indir_trunc: dangling dep");
2858 		ump->um_numindirdeps -= 1;
2859 		FREE_LOCK(&lk);
2860 	} else {
2861 #ifdef notyet
2862 		if (bp)
2863 			brelse(bp);
2864 #endif
2865 		FREE_LOCK(&lk);
2866 		error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
2867 		    NOCRED, &bp);
2868 		if (error) {
2869 			brelse(bp);
2870 			return (error);
2871 		}
2872 	}
2873 	/*
2874 	 * Recursively free indirect blocks.
2875 	 */
2876 	if (ump->um_fstype == UFS1) {
2877 		ufs1fmt = 1;
2878 		bap1 = (ufs1_daddr_t *)bp->b_data;
2879 	} else {
2880 		ufs1fmt = 0;
2881 		bap2 = (ufs2_daddr_t *)bp->b_data;
2882 	}
2883 	nblocks = btodb(fs->fs_bsize);
2884 	for (i = NINDIR(fs) - 1; i >= 0; i--) {
2885 		if (ufs1fmt)
2886 			nb = bap1[i];
2887 		else
2888 			nb = bap2[i];
2889 		if (nb == 0)
2890 			continue;
2891 		if (level != 0) {
2892 			if ((error = indir_trunc(freeblks, fsbtodb(fs, nb),
2893 			     level - 1, lbn + (i * lbnadd), countp)) != 0)
2894 				allerror = error;
2895 		}
2896 		ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, fs->fs_bsize,
2897 		    freeblks->fb_previousinum);
2898 		fs_pendingblocks += nblocks;
2899 		*countp += nblocks;
2900 	}
2901 	UFS_LOCK(ump);
2902 	fs->fs_pendingblocks -= fs_pendingblocks;
2903 	UFS_UNLOCK(ump);
2904 	bp->b_flags |= B_INVAL | B_NOCACHE;
2905 	brelse(bp);
2906 	return (allerror);
2907 }
2908 
2909 /*
2910  * Free an allocindir.
2911  * This routine must be called with splbio interrupts blocked.
2912  */
2913 static void
2914 free_allocindir(aip, inodedep)
2915 	struct allocindir *aip;
2916 	struct inodedep *inodedep;
2917 {
2918 	struct freefrag *freefrag;
2919 
2920 	mtx_assert(&lk, MA_OWNED);
2921 	if ((aip->ai_state & DEPCOMPLETE) == 0)
2922 		LIST_REMOVE(aip, ai_deps);
2923 	if (aip->ai_state & ONWORKLIST)
2924 		WORKLIST_REMOVE(&aip->ai_list);
2925 	LIST_REMOVE(aip, ai_next);
2926 	if ((freefrag = aip->ai_freefrag) != NULL) {
2927 		if (inodedep == NULL)
2928 			add_to_worklist(&freefrag->ff_list);
2929 		else
2930 			WORKLIST_INSERT(&inodedep->id_bufwait,
2931 			    &freefrag->ff_list);
2932 	}
2933 	WORKITEM_FREE(aip, D_ALLOCINDIR);
2934 }
2935 
2936 /*
2937  * Directory entry addition dependencies.
2938  *
2939  * When adding a new directory entry, the inode (with its incremented link
2940  * count) must be written to disk before the directory entry's pointer to it.
2941  * Also, if the inode is newly allocated, the corresponding freemap must be
2942  * updated (on disk) before the directory entry's pointer. These requirements
2943  * are met via undo/redo on the directory entry's pointer, which consists
2944  * simply of the inode number.
2945  *
2946  * As directory entries are added and deleted, the free space within a
2947  * directory block can become fragmented.  The ufs filesystem will compact
2948  * a fragmented directory block to make space for a new entry. When this
2949  * occurs, the offsets of previously added entries change. Any "diradd"
2950  * dependency structures corresponding to these entries must be updated with
2951  * the new offsets.
2952  */
2953 
2954 /*
2955  * This routine is called after the in-memory inode's link
2956  * count has been incremented, but before the directory entry's
2957  * pointer to the inode has been set.
2958  */
2959 int
2960 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
2961 	struct buf *bp;		/* buffer containing directory block */
2962 	struct inode *dp;	/* inode for directory */
2963 	off_t diroffset;	/* offset of new entry in directory */
2964 	ino_t newinum;		/* inode referenced by new directory entry */
2965 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
2966 	int isnewblk;		/* entry is in a newly allocated block */
2967 {
2968 	int offset;		/* offset of new entry within directory block */
2969 	ufs_lbn_t lbn;		/* block in directory containing new entry */
2970 	struct fs *fs;
2971 	struct diradd *dap;
2972 	struct allocdirect *adp;
2973 	struct pagedep *pagedep;
2974 	struct inodedep *inodedep;
2975 	struct newdirblk *newdirblk = 0;
2976 	struct mkdir *mkdir1, *mkdir2;
2977 	struct mount *mp;
2978 
2979 	/*
2980 	 * Whiteouts have no dependencies.
2981 	 */
2982 	if (newinum == WINO) {
2983 		if (newdirbp != NULL)
2984 			bdwrite(newdirbp);
2985 		return (0);
2986 	}
2987 	mp = UFSTOVFS(dp->i_ump);
2988 	fs = dp->i_fs;
2989 	lbn = lblkno(fs, diroffset);
2990 	offset = blkoff(fs, diroffset);
2991 	MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2992 		M_SOFTDEP_FLAGS|M_ZERO);
2993 	workitem_alloc(&dap->da_list, D_DIRADD, mp);
2994 	dap->da_offset = offset;
2995 	dap->da_newinum = newinum;
2996 	dap->da_state = ATTACHED;
2997 	if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) {
2998 		MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk),
2999 		    M_NEWDIRBLK, M_SOFTDEP_FLAGS);
3000 		workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
3001 	}
3002 	if (newdirbp == NULL) {
3003 		dap->da_state |= DEPCOMPLETE;
3004 		ACQUIRE_LOCK(&lk);
3005 	} else {
3006 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
3007 		MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
3008 		    M_SOFTDEP_FLAGS);
3009 		workitem_alloc(&mkdir1->md_list, D_MKDIR, mp);
3010 		mkdir1->md_state = MKDIR_BODY;
3011 		mkdir1->md_diradd = dap;
3012 		MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
3013 		    M_SOFTDEP_FLAGS);
3014 		workitem_alloc(&mkdir2->md_list, D_MKDIR, mp);
3015 		mkdir2->md_state = MKDIR_PARENT;
3016 		mkdir2->md_diradd = dap;
3017 		/*
3018 		 * Dependency on "." and ".." being written to disk.
3019 		 */
3020 		mkdir1->md_buf = newdirbp;
3021 		ACQUIRE_LOCK(&lk);
3022 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
3023 		WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
3024 		FREE_LOCK(&lk);
3025 		bdwrite(newdirbp);
3026 		/*
3027 		 * Dependency on link count increase for parent directory
3028 		 */
3029 		ACQUIRE_LOCK(&lk);
3030 		if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0
3031 		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3032 			dap->da_state &= ~MKDIR_PARENT;
3033 			WORKITEM_FREE(mkdir2, D_MKDIR);
3034 		} else {
3035 			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
3036 			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
3037 		}
3038 	}
3039 	/*
3040 	 * Link into parent directory pagedep to await its being written.
3041 	 */
3042 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3043 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3044 	dap->da_pagedep = pagedep;
3045 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
3046 	    da_pdlist);
3047 	/*
3048 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3049 	 * is not yet written. If it is written, do the post-inode write
3050 	 * processing to put it on the id_pendinghd list.
3051 	 */
3052 	(void) inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
3053 	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
3054 		diradd_inode_written(dap, inodedep);
3055 	else
3056 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3057 	if (isnewblk) {
3058 		/*
3059 		 * Directories growing into indirect blocks are rare
3060 		 * enough and the frequency of new block allocation
3061 		 * in those cases even more rare, that we choose not
3062 		 * to bother tracking them. Rather we simply force the
3063 		 * new directory entry to disk.
3064 		 */
3065 		if (lbn >= NDADDR) {
3066 			FREE_LOCK(&lk);
3067 			/*
3068 			 * We only have a new allocation when at the
3069 			 * beginning of a new block, not when we are
3070 			 * expanding into an existing block.
3071 			 */
3072 			if (blkoff(fs, diroffset) == 0)
3073 				return (1);
3074 			return (0);
3075 		}
3076 		/*
3077 		 * We only have a new allocation when at the beginning
3078 		 * of a new fragment, not when we are expanding into an
3079 		 * existing fragment. Also, there is nothing to do if we
3080 		 * are already tracking this block.
3081 		 */
3082 		if (fragoff(fs, diroffset) != 0) {
3083 			FREE_LOCK(&lk);
3084 			return (0);
3085 		}
3086 		if ((pagedep->pd_state & NEWBLOCK) != 0) {
3087 			WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
3088 			FREE_LOCK(&lk);
3089 			return (0);
3090 		}
3091 		/*
3092 		 * Find our associated allocdirect and have it track us.
3093 		 */
3094 		if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0)
3095 			panic("softdep_setup_directory_add: lost inodedep");
3096 		adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst);
3097 		if (adp == NULL || adp->ad_lbn != lbn)
3098 			panic("softdep_setup_directory_add: lost entry");
3099 		pagedep->pd_state |= NEWBLOCK;
3100 		newdirblk->db_pagedep = pagedep;
3101 		WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list);
3102 	}
3103 	FREE_LOCK(&lk);
3104 	return (0);
3105 }
3106 
3107 /*
3108  * This procedure is called to change the offset of a directory
3109  * entry when compacting a directory block which must be owned
3110  * exclusively by the caller. Note that the actual entry movement
3111  * must be done in this procedure to ensure that no I/O completions
3112  * occur while the move is in progress.
3113  */
3114 void
3115 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
3116 	struct inode *dp;	/* inode for directory */
3117 	caddr_t base;		/* address of dp->i_offset */
3118 	caddr_t oldloc;		/* address of old directory location */
3119 	caddr_t newloc;		/* address of new directory location */
3120 	int entrysize;		/* size of directory entry */
3121 {
3122 	int offset, oldoffset, newoffset;
3123 	struct pagedep *pagedep;
3124 	struct diradd *dap;
3125 	ufs_lbn_t lbn;
3126 
3127 	ACQUIRE_LOCK(&lk);
3128 	lbn = lblkno(dp->i_fs, dp->i_offset);
3129 	offset = blkoff(dp->i_fs, dp->i_offset);
3130 	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
3131 		goto done;
3132 	oldoffset = offset + (oldloc - base);
3133 	newoffset = offset + (newloc - base);
3134 
3135 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
3136 		if (dap->da_offset != oldoffset)
3137 			continue;
3138 		dap->da_offset = newoffset;
3139 		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
3140 			break;
3141 		LIST_REMOVE(dap, da_pdlist);
3142 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
3143 		    dap, da_pdlist);
3144 		break;
3145 	}
3146 	if (dap == NULL) {
3147 
3148 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
3149 			if (dap->da_offset == oldoffset) {
3150 				dap->da_offset = newoffset;
3151 				break;
3152 			}
3153 		}
3154 	}
3155 done:
3156 	bcopy(oldloc, newloc, entrysize);
3157 	FREE_LOCK(&lk);
3158 }
3159 
3160 /*
3161  * Free a diradd dependency structure. This routine must be called
3162  * with splbio interrupts blocked.
3163  */
3164 static void
3165 free_diradd(dap)
3166 	struct diradd *dap;
3167 {
3168 	struct dirrem *dirrem;
3169 	struct pagedep *pagedep;
3170 	struct inodedep *inodedep;
3171 	struct mkdir *mkdir, *nextmd;
3172 
3173 	mtx_assert(&lk, MA_OWNED);
3174 	WORKLIST_REMOVE(&dap->da_list);
3175 	LIST_REMOVE(dap, da_pdlist);
3176 	if ((dap->da_state & DIRCHG) == 0) {
3177 		pagedep = dap->da_pagedep;
3178 	} else {
3179 		dirrem = dap->da_previous;
3180 		pagedep = dirrem->dm_pagedep;
3181 		dirrem->dm_dirinum = pagedep->pd_ino;
3182 		add_to_worklist(&dirrem->dm_list);
3183 	}
3184 	if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum,
3185 	    0, &inodedep) != 0)
3186 		(void) free_inodedep(inodedep);
3187 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
3188 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
3189 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
3190 			if (mkdir->md_diradd != dap)
3191 				continue;
3192 			dap->da_state &= ~mkdir->md_state;
3193 			WORKLIST_REMOVE(&mkdir->md_list);
3194 			LIST_REMOVE(mkdir, md_mkdirs);
3195 			WORKITEM_FREE(mkdir, D_MKDIR);
3196 		}
3197 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
3198 			panic("free_diradd: unfound ref");
3199 	}
3200 	WORKITEM_FREE(dap, D_DIRADD);
3201 }
3202 
3203 /*
3204  * Directory entry removal dependencies.
3205  *
3206  * When removing a directory entry, the entry's inode pointer must be
3207  * zero'ed on disk before the corresponding inode's link count is decremented
3208  * (possibly freeing the inode for re-use). This dependency is handled by
3209  * updating the directory entry but delaying the inode count reduction until
3210  * after the directory block has been written to disk. After this point, the
3211  * inode count can be decremented whenever it is convenient.
3212  */
3213 
3214 /*
3215  * This routine should be called immediately after removing
3216  * a directory entry.  The inode's link count should not be
3217  * decremented by the calling procedure -- the soft updates
3218  * code will do this task when it is safe.
3219  */
3220 void
3221 softdep_setup_remove(bp, dp, ip, isrmdir)
3222 	struct buf *bp;		/* buffer containing directory block */
3223 	struct inode *dp;	/* inode for the directory being modified */
3224 	struct inode *ip;	/* inode for directory entry being removed */
3225 	int isrmdir;		/* indicates if doing RMDIR */
3226 {
3227 	struct dirrem *dirrem, *prevdirrem;
3228 
3229 	/*
3230 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
3231 	 */
3232 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3233 
3234 	/*
3235 	 * If the COMPLETE flag is clear, then there were no active
3236 	 * entries and we want to roll back to a zeroed entry until
3237 	 * the new inode is committed to disk. If the COMPLETE flag is
3238 	 * set then we have deleted an entry that never made it to
3239 	 * disk. If the entry we deleted resulted from a name change,
3240 	 * then the old name still resides on disk. We cannot delete
3241 	 * its inode (returned to us in prevdirrem) until the zeroed
3242 	 * directory entry gets to disk. The new inode has never been
3243 	 * referenced on the disk, so can be deleted immediately.
3244 	 */
3245 	if ((dirrem->dm_state & COMPLETE) == 0) {
3246 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
3247 		    dm_next);
3248 		FREE_LOCK(&lk);
3249 	} else {
3250 		if (prevdirrem != NULL)
3251 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
3252 			    prevdirrem, dm_next);
3253 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
3254 		FREE_LOCK(&lk);
3255 		handle_workitem_remove(dirrem, NULL);
3256 	}
3257 }
3258 
3259 /*
3260  * Allocate a new dirrem if appropriate and return it along with
3261  * its associated pagedep. Called without a lock, returns with lock.
3262  */
3263 static long num_dirrem;		/* number of dirrem allocated */
3264 static struct dirrem *
3265 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
3266 	struct buf *bp;		/* buffer containing directory block */
3267 	struct inode *dp;	/* inode for the directory being modified */
3268 	struct inode *ip;	/* inode for directory entry being removed */
3269 	int isrmdir;		/* indicates if doing RMDIR */
3270 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
3271 {
3272 	int offset;
3273 	ufs_lbn_t lbn;
3274 	struct diradd *dap;
3275 	struct dirrem *dirrem;
3276 	struct pagedep *pagedep;
3277 
3278 	/*
3279 	 * Whiteouts have no deletion dependencies.
3280 	 */
3281 	if (ip == NULL)
3282 		panic("newdirrem: whiteout");
3283 	/*
3284 	 * If we are over our limit, try to improve the situation.
3285 	 * Limiting the number of dirrem structures will also limit
3286 	 * the number of freefile and freeblks structures.
3287 	 */
3288 	ACQUIRE_LOCK(&lk);
3289 	if (num_dirrem > max_softdeps / 2)
3290 		(void) request_cleanup(ITOV(dp)->v_mount, FLUSH_REMOVE);
3291 	num_dirrem += 1;
3292 	FREE_LOCK(&lk);
3293 	MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
3294 		M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO);
3295 	workitem_alloc(&dirrem->dm_list, D_DIRREM, ITOV(dp)->v_mount);
3296 	dirrem->dm_state = isrmdir ? RMDIR : 0;
3297 	dirrem->dm_oldinum = ip->i_number;
3298 	*prevdirremp = NULL;
3299 
3300 	ACQUIRE_LOCK(&lk);
3301 	lbn = lblkno(dp->i_fs, dp->i_offset);
3302 	offset = blkoff(dp->i_fs, dp->i_offset);
3303 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3304 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3305 	dirrem->dm_pagedep = pagedep;
3306 	/*
3307 	 * Check for a diradd dependency for the same directory entry.
3308 	 * If present, then both dependencies become obsolete and can
3309 	 * be de-allocated. Check for an entry on both the pd_dirraddhd
3310 	 * list and the pd_pendinghd list.
3311 	 */
3312 
3313 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
3314 		if (dap->da_offset == offset)
3315 			break;
3316 	if (dap == NULL) {
3317 
3318 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
3319 			if (dap->da_offset == offset)
3320 				break;
3321 		if (dap == NULL)
3322 			return (dirrem);
3323 	}
3324 	/*
3325 	 * Must be ATTACHED at this point.
3326 	 */
3327 	if ((dap->da_state & ATTACHED) == 0)
3328 		panic("newdirrem: not ATTACHED");
3329 	if (dap->da_newinum != ip->i_number)
3330 		panic("newdirrem: inum %d should be %d",
3331 		    ip->i_number, dap->da_newinum);
3332 	/*
3333 	 * If we are deleting a changed name that never made it to disk,
3334 	 * then return the dirrem describing the previous inode (which
3335 	 * represents the inode currently referenced from this entry on disk).
3336 	 */
3337 	if ((dap->da_state & DIRCHG) != 0) {
3338 		*prevdirremp = dap->da_previous;
3339 		dap->da_state &= ~DIRCHG;
3340 		dap->da_pagedep = pagedep;
3341 	}
3342 	/*
3343 	 * We are deleting an entry that never made it to disk.
3344 	 * Mark it COMPLETE so we can delete its inode immediately.
3345 	 */
3346 	dirrem->dm_state |= COMPLETE;
3347 	free_diradd(dap);
3348 	return (dirrem);
3349 }
3350 
3351 /*
3352  * Directory entry change dependencies.
3353  *
3354  * Changing an existing directory entry requires that an add operation
3355  * be completed first followed by a deletion. The semantics for the addition
3356  * are identical to the description of adding a new entry above except
3357  * that the rollback is to the old inode number rather than zero. Once
3358  * the addition dependency is completed, the removal is done as described
3359  * in the removal routine above.
3360  */
3361 
3362 /*
3363  * This routine should be called immediately after changing
3364  * a directory entry.  The inode's link count should not be
3365  * decremented by the calling procedure -- the soft updates
3366  * code will perform this task when it is safe.
3367  */
3368 void
3369 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
3370 	struct buf *bp;		/* buffer containing directory block */
3371 	struct inode *dp;	/* inode for the directory being modified */
3372 	struct inode *ip;	/* inode for directory entry being removed */
3373 	ino_t newinum;		/* new inode number for changed entry */
3374 	int isrmdir;		/* indicates if doing RMDIR */
3375 {
3376 	int offset;
3377 	struct diradd *dap = NULL;
3378 	struct dirrem *dirrem, *prevdirrem;
3379 	struct pagedep *pagedep;
3380 	struct inodedep *inodedep;
3381 	struct mount *mp;
3382 
3383 	offset = blkoff(dp->i_fs, dp->i_offset);
3384 	mp = UFSTOVFS(dp->i_ump);
3385 
3386 	/*
3387 	 * Whiteouts do not need diradd dependencies.
3388 	 */
3389 	if (newinum != WINO) {
3390 		MALLOC(dap, struct diradd *, sizeof(struct diradd),
3391 		    M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
3392 		workitem_alloc(&dap->da_list, D_DIRADD, mp);
3393 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
3394 		dap->da_offset = offset;
3395 		dap->da_newinum = newinum;
3396 	}
3397 
3398 	/*
3399 	 * Allocate a new dirrem and ACQUIRE_LOCK.
3400 	 */
3401 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3402 	pagedep = dirrem->dm_pagedep;
3403 	/*
3404 	 * The possible values for isrmdir:
3405 	 *	0 - non-directory file rename
3406 	 *	1 - directory rename within same directory
3407 	 *   inum - directory rename to new directory of given inode number
3408 	 * When renaming to a new directory, we are both deleting and
3409 	 * creating a new directory entry, so the link count on the new
3410 	 * directory should not change. Thus we do not need the followup
3411 	 * dirrem which is usually done in handle_workitem_remove. We set
3412 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
3413 	 * followup dirrem.
3414 	 */
3415 	if (isrmdir > 1)
3416 		dirrem->dm_state |= DIRCHG;
3417 
3418 	/*
3419 	 * Whiteouts have no additional dependencies,
3420 	 * so just put the dirrem on the correct list.
3421 	 */
3422 	if (newinum == WINO) {
3423 		if ((dirrem->dm_state & COMPLETE) == 0) {
3424 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
3425 			    dm_next);
3426 		} else {
3427 			dirrem->dm_dirinum = pagedep->pd_ino;
3428 			add_to_worklist(&dirrem->dm_list);
3429 		}
3430 		FREE_LOCK(&lk);
3431 		return;
3432 	}
3433 
3434 	/*
3435 	 * If the COMPLETE flag is clear, then there were no active
3436 	 * entries and we want to roll back to the previous inode until
3437 	 * the new inode is committed to disk. If the COMPLETE flag is
3438 	 * set, then we have deleted an entry that never made it to disk.
3439 	 * If the entry we deleted resulted from a name change, then the old
3440 	 * inode reference still resides on disk. Any rollback that we do
3441 	 * needs to be to that old inode (returned to us in prevdirrem). If
3442 	 * the entry we deleted resulted from a create, then there is
3443 	 * no entry on the disk, so we want to roll back to zero rather
3444 	 * than the uncommitted inode. In either of the COMPLETE cases we
3445 	 * want to immediately free the unwritten and unreferenced inode.
3446 	 */
3447 	if ((dirrem->dm_state & COMPLETE) == 0) {
3448 		dap->da_previous = dirrem;
3449 	} else {
3450 		if (prevdirrem != NULL) {
3451 			dap->da_previous = prevdirrem;
3452 		} else {
3453 			dap->da_state &= ~DIRCHG;
3454 			dap->da_pagedep = pagedep;
3455 		}
3456 		dirrem->dm_dirinum = pagedep->pd_ino;
3457 		add_to_worklist(&dirrem->dm_list);
3458 	}
3459 	/*
3460 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3461 	 * is not yet written. If it is written, do the post-inode write
3462 	 * processing to put it on the id_pendinghd list.
3463 	 */
3464 	if (inodedep_lookup(mp, newinum, DEPALLOC, &inodedep) == 0 ||
3465 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3466 		dap->da_state |= COMPLETE;
3467 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3468 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3469 	} else {
3470 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
3471 		    dap, da_pdlist);
3472 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3473 	}
3474 	FREE_LOCK(&lk);
3475 }
3476 
3477 /*
3478  * Called whenever the link count on an inode is changed.
3479  * It creates an inode dependency so that the new reference(s)
3480  * to the inode cannot be committed to disk until the updated
3481  * inode has been written.
3482  */
3483 void
3484 softdep_change_linkcnt(ip)
3485 	struct inode *ip;	/* the inode with the increased link count */
3486 {
3487 	struct inodedep *inodedep;
3488 
3489 	ACQUIRE_LOCK(&lk);
3490 	(void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number,
3491 	    DEPALLOC, &inodedep);
3492 	if (ip->i_nlink < ip->i_effnlink)
3493 		panic("softdep_change_linkcnt: bad delta");
3494 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3495 	FREE_LOCK(&lk);
3496 }
3497 
3498 /*
3499  * Called when the effective link count and the reference count
3500  * on an inode drops to zero. At this point there are no names
3501  * referencing the file in the filesystem and no active file
3502  * references. The space associated with the file will be freed
3503  * as soon as the necessary soft dependencies are cleared.
3504  */
3505 void
3506 softdep_releasefile(ip)
3507 	struct inode *ip;	/* inode with the zero effective link count */
3508 {
3509 	struct inodedep *inodedep;
3510 	struct fs *fs;
3511 	int extblocks;
3512 
3513 	if (ip->i_effnlink > 0)
3514 		panic("softdep_releasefile: file still referenced");
3515 	/*
3516 	 * We may be called several times as the on-disk link count
3517 	 * drops to zero. We only want to account for the space once.
3518 	 */
3519 	if (ip->i_flag & IN_SPACECOUNTED)
3520 		return;
3521 	/*
3522 	 * We have to deactivate a snapshot otherwise copyonwrites may
3523 	 * add blocks and the cleanup may remove blocks after we have
3524 	 * tried to account for them.
3525 	 */
3526 	if ((ip->i_flags & SF_SNAPSHOT) != 0)
3527 		ffs_snapremove(ITOV(ip));
3528 	/*
3529 	 * If we are tracking an nlinkdelta, we have to also remember
3530 	 * whether we accounted for the freed space yet.
3531 	 */
3532 	ACQUIRE_LOCK(&lk);
3533 	if ((inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, &inodedep)))
3534 		inodedep->id_state |= SPACECOUNTED;
3535 	FREE_LOCK(&lk);
3536 	fs = ip->i_fs;
3537 	extblocks = 0;
3538 	if (fs->fs_magic == FS_UFS2_MAGIC)
3539 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
3540 	UFS_LOCK(ip->i_ump);
3541 	ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks;
3542 	ip->i_fs->fs_pendinginodes += 1;
3543 	UFS_UNLOCK(ip->i_ump);
3544 	ip->i_flag |= IN_SPACECOUNTED;
3545 }
3546 
3547 /*
3548  * This workitem decrements the inode's link count.
3549  * If the link count reaches zero, the file is removed.
3550  */
3551 static void
3552 handle_workitem_remove(dirrem, xp)
3553 	struct dirrem *dirrem;
3554 	struct vnode *xp;
3555 {
3556 	struct thread *td = curthread;
3557 	struct inodedep *inodedep;
3558 	struct vnode *vp;
3559 	struct inode *ip;
3560 	ino_t oldinum;
3561 	int error;
3562 
3563 	if ((vp = xp) == NULL &&
3564 	    (error = ffs_vget(dirrem->dm_list.wk_mp,
3565 	    dirrem->dm_oldinum, LK_EXCLUSIVE, &vp)) != 0) {
3566 		softdep_error("handle_workitem_remove: vget", error);
3567 		return;
3568 	}
3569 	ip = VTOI(vp);
3570 	ACQUIRE_LOCK(&lk);
3571 	if ((inodedep_lookup(dirrem->dm_list.wk_mp,
3572 	    dirrem->dm_oldinum, 0, &inodedep)) == 0)
3573 		panic("handle_workitem_remove: lost inodedep");
3574 	/*
3575 	 * Normal file deletion.
3576 	 */
3577 	if ((dirrem->dm_state & RMDIR) == 0) {
3578 		ip->i_nlink--;
3579 		DIP_SET(ip, i_nlink, ip->i_nlink);
3580 		ip->i_flag |= IN_CHANGE;
3581 		if (ip->i_nlink < ip->i_effnlink)
3582 			panic("handle_workitem_remove: bad file delta");
3583 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3584 		num_dirrem -= 1;
3585 		WORKITEM_FREE(dirrem, D_DIRREM);
3586 		FREE_LOCK(&lk);
3587 		vput(vp);
3588 		return;
3589 	}
3590 	/*
3591 	 * Directory deletion. Decrement reference count for both the
3592 	 * just deleted parent directory entry and the reference for ".".
3593 	 * Next truncate the directory to length zero. When the
3594 	 * truncation completes, arrange to have the reference count on
3595 	 * the parent decremented to account for the loss of "..".
3596 	 */
3597 	ip->i_nlink -= 2;
3598 	DIP_SET(ip, i_nlink, ip->i_nlink);
3599 	ip->i_flag |= IN_CHANGE;
3600 	if (ip->i_nlink < ip->i_effnlink)
3601 		panic("handle_workitem_remove: bad dir delta");
3602 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3603 	FREE_LOCK(&lk);
3604 	if ((error = ffs_truncate(vp, (off_t)0, 0, td->td_ucred, td)) != 0)
3605 		softdep_error("handle_workitem_remove: truncate", error);
3606 	ACQUIRE_LOCK(&lk);
3607 	/*
3608 	 * Rename a directory to a new parent. Since, we are both deleting
3609 	 * and creating a new directory entry, the link count on the new
3610 	 * directory should not change. Thus we skip the followup dirrem.
3611 	 */
3612 	if (dirrem->dm_state & DIRCHG) {
3613 		num_dirrem -= 1;
3614 		WORKITEM_FREE(dirrem, D_DIRREM);
3615 		FREE_LOCK(&lk);
3616 		vput(vp);
3617 		return;
3618 	}
3619 	/*
3620 	 * If the inodedep does not exist, then the zero'ed inode has
3621 	 * been written to disk. If the allocated inode has never been
3622 	 * written to disk, then the on-disk inode is zero'ed. In either
3623 	 * case we can remove the file immediately.
3624 	 */
3625 	dirrem->dm_state = 0;
3626 	oldinum = dirrem->dm_oldinum;
3627 	dirrem->dm_oldinum = dirrem->dm_dirinum;
3628 	if (inodedep_lookup(dirrem->dm_list.wk_mp, oldinum,
3629 	    0, &inodedep) == 0 || check_inode_unwritten(inodedep)) {
3630 		if (xp != NULL)
3631 			add_to_worklist(&dirrem->dm_list);
3632 		FREE_LOCK(&lk);
3633 		vput(vp);
3634 		if (xp == NULL)
3635 			handle_workitem_remove(dirrem, NULL);
3636 		return;
3637 	}
3638 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3639 	FREE_LOCK(&lk);
3640 	ip->i_flag |= IN_CHANGE;
3641 	ffs_update(vp, 0);
3642 	vput(vp);
3643 }
3644 
3645 /*
3646  * Inode de-allocation dependencies.
3647  *
3648  * When an inode's link count is reduced to zero, it can be de-allocated. We
3649  * found it convenient to postpone de-allocation until after the inode is
3650  * written to disk with its new link count (zero).  At this point, all of the
3651  * on-disk inode's block pointers are nullified and, with careful dependency
3652  * list ordering, all dependencies related to the inode will be satisfied and
3653  * the corresponding dependency structures de-allocated.  So, if/when the
3654  * inode is reused, there will be no mixing of old dependencies with new
3655  * ones.  This artificial dependency is set up by the block de-allocation
3656  * procedure above (softdep_setup_freeblocks) and completed by the
3657  * following procedure.
3658  */
3659 static void
3660 handle_workitem_freefile(freefile)
3661 	struct freefile *freefile;
3662 {
3663 	struct fs *fs;
3664 	struct inodedep *idp;
3665 	struct ufsmount *ump;
3666 	int error;
3667 
3668 	ump = VFSTOUFS(freefile->fx_list.wk_mp);
3669 	fs = ump->um_fs;
3670 #ifdef DEBUG
3671 	ACQUIRE_LOCK(&lk);
3672 	error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp);
3673 	FREE_LOCK(&lk);
3674 	if (error)
3675 		panic("handle_workitem_freefile: inodedep survived");
3676 #endif
3677 	UFS_LOCK(ump);
3678 	fs->fs_pendinginodes -= 1;
3679 	UFS_UNLOCK(ump);
3680 	if ((error = ffs_freefile(ump, fs, freefile->fx_devvp,
3681 	    freefile->fx_oldinum, freefile->fx_mode)) != 0)
3682 		softdep_error("handle_workitem_freefile", error);
3683 	ACQUIRE_LOCK(&lk);
3684 	WORKITEM_FREE(freefile, D_FREEFILE);
3685 	FREE_LOCK(&lk);
3686 }
3687 
3688 
3689 /*
3690  * Helper function which unlinks marker element from work list and returns
3691  * the next element on the list.
3692  */
3693 static __inline struct worklist *
3694 markernext(struct worklist *marker)
3695 {
3696 	struct worklist *next;
3697 
3698 	next = LIST_NEXT(marker, wk_list);
3699 	LIST_REMOVE(marker, wk_list);
3700 	return next;
3701 }
3702 
3703 /*
3704  * Disk writes.
3705  *
3706  * The dependency structures constructed above are most actively used when file
3707  * system blocks are written to disk.  No constraints are placed on when a
3708  * block can be written, but unsatisfied update dependencies are made safe by
3709  * modifying (or replacing) the source memory for the duration of the disk
3710  * write.  When the disk write completes, the memory block is again brought
3711  * up-to-date.
3712  *
3713  * In-core inode structure reclamation.
3714  *
3715  * Because there are a finite number of "in-core" inode structures, they are
3716  * reused regularly.  By transferring all inode-related dependencies to the
3717  * in-memory inode block and indexing them separately (via "inodedep"s), we
3718  * can allow "in-core" inode structures to be reused at any time and avoid
3719  * any increase in contention.
3720  *
3721  * Called just before entering the device driver to initiate a new disk I/O.
3722  * The buffer must be locked, thus, no I/O completion operations can occur
3723  * while we are manipulating its associated dependencies.
3724  */
3725 static void
3726 softdep_disk_io_initiation(bp)
3727 	struct buf *bp;		/* structure describing disk write to occur */
3728 {
3729 	struct worklist *wk;
3730 	struct worklist marker;
3731 	struct indirdep *indirdep;
3732 	struct inodedep *inodedep;
3733 
3734 	/*
3735 	 * We only care about write operations. There should never
3736 	 * be dependencies for reads.
3737 	 */
3738 	if (bp->b_iocmd != BIO_WRITE)
3739 		panic("softdep_disk_io_initiation: not write");
3740 
3741 	marker.wk_type = D_LAST + 1;	/* Not a normal workitem */
3742 	PHOLD(curproc);			/* Don't swap out kernel stack */
3743 
3744 	ACQUIRE_LOCK(&lk);
3745 	/*
3746 	 * Do any necessary pre-I/O processing.
3747 	 */
3748 	for (wk = LIST_FIRST(&bp->b_dep); wk != NULL;
3749 	     wk = markernext(&marker)) {
3750 		LIST_INSERT_AFTER(wk, &marker, wk_list);
3751 		switch (wk->wk_type) {
3752 
3753 		case D_PAGEDEP:
3754 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
3755 			continue;
3756 
3757 		case D_INODEDEP:
3758 			inodedep = WK_INODEDEP(wk);
3759 			if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
3760 				initiate_write_inodeblock_ufs1(inodedep, bp);
3761 			else
3762 				initiate_write_inodeblock_ufs2(inodedep, bp);
3763 			continue;
3764 
3765 		case D_INDIRDEP:
3766 			indirdep = WK_INDIRDEP(wk);
3767 			if (indirdep->ir_state & GOINGAWAY)
3768 				panic("disk_io_initiation: indirdep gone");
3769 			/*
3770 			 * If there are no remaining dependencies, this
3771 			 * will be writing the real pointers, so the
3772 			 * dependency can be freed.
3773 			 */
3774 			if (LIST_EMPTY(&indirdep->ir_deplisthd)) {
3775 				struct buf *bp;
3776 
3777 				bp = indirdep->ir_savebp;
3778 				bp->b_flags |= B_INVAL | B_NOCACHE;
3779 				/* inline expand WORKLIST_REMOVE(wk); */
3780 				wk->wk_state &= ~ONWORKLIST;
3781 				LIST_REMOVE(wk, wk_list);
3782 				WORKITEM_FREE(indirdep, D_INDIRDEP);
3783 				FREE_LOCK(&lk);
3784 				brelse(bp);
3785 				ACQUIRE_LOCK(&lk);
3786 				continue;
3787 			}
3788 			/*
3789 			 * Replace up-to-date version with safe version.
3790 			 */
3791 			FREE_LOCK(&lk);
3792 			MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3793 			    M_INDIRDEP, M_SOFTDEP_FLAGS);
3794 			ACQUIRE_LOCK(&lk);
3795 			indirdep->ir_state &= ~ATTACHED;
3796 			indirdep->ir_state |= UNDONE;
3797 			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3798 			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3799 			    bp->b_bcount);
3800 			continue;
3801 
3802 		case D_MKDIR:
3803 		case D_BMSAFEMAP:
3804 		case D_ALLOCDIRECT:
3805 		case D_ALLOCINDIR:
3806 			continue;
3807 
3808 		default:
3809 			panic("handle_disk_io_initiation: Unexpected type %s",
3810 			    TYPENAME(wk->wk_type));
3811 			/* NOTREACHED */
3812 		}
3813 	}
3814 	FREE_LOCK(&lk);
3815 	PRELE(curproc);			/* Allow swapout of kernel stack */
3816 }
3817 
3818 /*
3819  * Called from within the procedure above to deal with unsatisfied
3820  * allocation dependencies in a directory. The buffer must be locked,
3821  * thus, no I/O completion operations can occur while we are
3822  * manipulating its associated dependencies.
3823  */
3824 static void
3825 initiate_write_filepage(pagedep, bp)
3826 	struct pagedep *pagedep;
3827 	struct buf *bp;
3828 {
3829 	struct diradd *dap;
3830 	struct direct *ep;
3831 	int i;
3832 
3833 	if (pagedep->pd_state & IOSTARTED) {
3834 		/*
3835 		 * This can only happen if there is a driver that does not
3836 		 * understand chaining. Here biodone will reissue the call
3837 		 * to strategy for the incomplete buffers.
3838 		 */
3839 		printf("initiate_write_filepage: already started\n");
3840 		return;
3841 	}
3842 	pagedep->pd_state |= IOSTARTED;
3843 	for (i = 0; i < DAHASHSZ; i++) {
3844 		LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3845 			ep = (struct direct *)
3846 			    ((char *)bp->b_data + dap->da_offset);
3847 			if (ep->d_ino != dap->da_newinum)
3848 				panic("%s: dir inum %d != new %d",
3849 				    "initiate_write_filepage",
3850 				    ep->d_ino, dap->da_newinum);
3851 			if (dap->da_state & DIRCHG)
3852 				ep->d_ino = dap->da_previous->dm_oldinum;
3853 			else
3854 				ep->d_ino = 0;
3855 			dap->da_state &= ~ATTACHED;
3856 			dap->da_state |= UNDONE;
3857 		}
3858 	}
3859 }
3860 
3861 /*
3862  * Version of initiate_write_inodeblock that handles UFS1 dinodes.
3863  * Note that any bug fixes made to this routine must be done in the
3864  * version found below.
3865  *
3866  * Called from within the procedure above to deal with unsatisfied
3867  * allocation dependencies in an inodeblock. The buffer must be
3868  * locked, thus, no I/O completion operations can occur while we
3869  * are manipulating its associated dependencies.
3870  */
3871 static void
3872 initiate_write_inodeblock_ufs1(inodedep, bp)
3873 	struct inodedep *inodedep;
3874 	struct buf *bp;			/* The inode block */
3875 {
3876 	struct allocdirect *adp, *lastadp;
3877 	struct ufs1_dinode *dp;
3878 	struct ufs1_dinode *sip;
3879 	struct fs *fs;
3880 	ufs_lbn_t i, prevlbn = 0;
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 DIAGNOSTIC
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 /* DIAGNOSTIC */
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 DIAGNOSTIC
3961 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
3962 				panic("softdep_write_inodeblock: lost dep1");
3963 #endif /* DIAGNOSTIC */
3964 			dp->di_db[i] = 0;
3965 		}
3966 		for (i = 0; i < NIADDR; i++) {
3967 #ifdef DIAGNOSTIC
3968 			if (dp->di_ib[i] != 0 &&
3969 			    (deplist & ((1 << NDADDR) << i)) == 0)
3970 				panic("softdep_write_inodeblock: lost dep2");
3971 #endif /* DIAGNOSTIC */
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, prevlbn = 0;
4024 	int deplist;
4025 
4026 	if (inodedep->id_state & IOSTARTED)
4027 		panic("initiate_write_inodeblock_ufs2: already started");
4028 	inodedep->id_state |= IOSTARTED;
4029 	fs = inodedep->id_fs;
4030 	dp = (struct ufs2_dinode *)bp->b_data +
4031 	    ino_to_fsbo(fs, inodedep->id_ino);
4032 	/*
4033 	 * If the bitmap is not yet written, then the allocated
4034 	 * inode cannot be written to disk.
4035 	 */
4036 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4037 		if (inodedep->id_savedino2 != NULL)
4038 			panic("initiate_write_inodeblock_ufs2: I/O underway");
4039 		FREE_LOCK(&lk);
4040 		MALLOC(sip, struct ufs2_dinode *,
4041 		    sizeof(struct ufs2_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS);
4042 		ACQUIRE_LOCK(&lk);
4043 		inodedep->id_savedino2 = sip;
4044 		*inodedep->id_savedino2 = *dp;
4045 		bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
4046 		dp->di_gen = inodedep->id_savedino2->di_gen;
4047 		return;
4048 	}
4049 	/*
4050 	 * If no dependencies, then there is nothing to roll back.
4051 	 */
4052 	inodedep->id_savedsize = dp->di_size;
4053 	inodedep->id_savedextsize = dp->di_extsize;
4054 	if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
4055 	    TAILQ_EMPTY(&inodedep->id_extupdt))
4056 		return;
4057 	/*
4058 	 * Set the ext data dependencies to busy.
4059 	 */
4060 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
4061 	     adp = TAILQ_NEXT(adp, ad_next)) {
4062 #ifdef DIAGNOSTIC
4063 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
4064 			panic("softdep_write_inodeblock: lbn order");
4065 		prevlbn = adp->ad_lbn;
4066 		if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno)
4067 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
4068 			    "softdep_write_inodeblock",
4069 			    (intmax_t)adp->ad_lbn,
4070 			    (intmax_t)dp->di_extb[adp->ad_lbn],
4071 			    (intmax_t)adp->ad_newblkno);
4072 		deplist |= 1 << adp->ad_lbn;
4073 		if ((adp->ad_state & ATTACHED) == 0)
4074 			panic("softdep_write_inodeblock: Unknown state 0x%x",
4075 			    adp->ad_state);
4076 #endif /* DIAGNOSTIC */
4077 		adp->ad_state &= ~ATTACHED;
4078 		adp->ad_state |= UNDONE;
4079 	}
4080 	/*
4081 	 * The on-disk inode cannot claim to be any larger than the last
4082 	 * fragment that has been written. Otherwise, the on-disk inode
4083 	 * might have fragments that were not the last block in the ext
4084 	 * data which would corrupt the filesystem.
4085 	 */
4086 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
4087 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
4088 		dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno;
4089 		/* keep going until hitting a rollback to a frag */
4090 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
4091 			continue;
4092 		dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
4093 		for (i = adp->ad_lbn + 1; i < NXADDR; i++) {
4094 #ifdef DIAGNOSTIC
4095 			if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0)
4096 				panic("softdep_write_inodeblock: lost dep1");
4097 #endif /* DIAGNOSTIC */
4098 			dp->di_extb[i] = 0;
4099 		}
4100 		lastadp = NULL;
4101 		break;
4102 	}
4103 	/*
4104 	 * If we have zero'ed out the last allocated block of the ext
4105 	 * data, roll back the size to the last currently allocated block.
4106 	 * We know that this last allocated block is a full-sized as
4107 	 * we already checked for fragments in the loop above.
4108 	 */
4109 	if (lastadp != NULL &&
4110 	    dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
4111 		for (i = lastadp->ad_lbn; i >= 0; i--)
4112 			if (dp->di_extb[i] != 0)
4113 				break;
4114 		dp->di_extsize = (i + 1) * fs->fs_bsize;
4115 	}
4116 	/*
4117 	 * Set the file data dependencies to busy.
4118 	 */
4119 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
4120 	     adp = TAILQ_NEXT(adp, ad_next)) {
4121 #ifdef DIAGNOSTIC
4122 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
4123 			panic("softdep_write_inodeblock: lbn order");
4124 		prevlbn = adp->ad_lbn;
4125 		if (adp->ad_lbn < NDADDR &&
4126 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno)
4127 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
4128 			    "softdep_write_inodeblock",
4129 			    (intmax_t)adp->ad_lbn,
4130 			    (intmax_t)dp->di_db[adp->ad_lbn],
4131 			    (intmax_t)adp->ad_newblkno);
4132 		if (adp->ad_lbn >= NDADDR &&
4133 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno)
4134 			panic("%s indirect pointer #%jd mismatch %jd != %jd",
4135 			    "softdep_write_inodeblock:",
4136 			    (intmax_t)adp->ad_lbn - NDADDR,
4137 			    (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR],
4138 			    (intmax_t)adp->ad_newblkno);
4139 		deplist |= 1 << adp->ad_lbn;
4140 		if ((adp->ad_state & ATTACHED) == 0)
4141 			panic("softdep_write_inodeblock: Unknown state 0x%x",
4142 			    adp->ad_state);
4143 #endif /* DIAGNOSTIC */
4144 		adp->ad_state &= ~ATTACHED;
4145 		adp->ad_state |= UNDONE;
4146 	}
4147 	/*
4148 	 * The on-disk inode cannot claim to be any larger than the last
4149 	 * fragment that has been written. Otherwise, the on-disk inode
4150 	 * might have fragments that were not the last block in the file
4151 	 * which would corrupt the filesystem.
4152 	 */
4153 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
4154 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
4155 		if (adp->ad_lbn >= NDADDR)
4156 			break;
4157 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
4158 		/* keep going until hitting a rollback to a frag */
4159 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
4160 			continue;
4161 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
4162 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
4163 #ifdef DIAGNOSTIC
4164 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
4165 				panic("softdep_write_inodeblock: lost dep2");
4166 #endif /* DIAGNOSTIC */
4167 			dp->di_db[i] = 0;
4168 		}
4169 		for (i = 0; i < NIADDR; i++) {
4170 #ifdef DIAGNOSTIC
4171 			if (dp->di_ib[i] != 0 &&
4172 			    (deplist & ((1 << NDADDR) << i)) == 0)
4173 				panic("softdep_write_inodeblock: lost dep3");
4174 #endif /* DIAGNOSTIC */
4175 			dp->di_ib[i] = 0;
4176 		}
4177 		return;
4178 	}
4179 	/*
4180 	 * If we have zero'ed out the last allocated block of the file,
4181 	 * roll back the size to the last currently allocated block.
4182 	 * We know that this last allocated block is a full-sized as
4183 	 * we already checked for fragments in the loop above.
4184 	 */
4185 	if (lastadp != NULL &&
4186 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
4187 		for (i = lastadp->ad_lbn; i >= 0; i--)
4188 			if (dp->di_db[i] != 0)
4189 				break;
4190 		dp->di_size = (i + 1) * fs->fs_bsize;
4191 	}
4192 	/*
4193 	 * The only dependencies are for indirect blocks.
4194 	 *
4195 	 * The file size for indirect block additions is not guaranteed.
4196 	 * Such a guarantee would be non-trivial to achieve. The conventional
4197 	 * synchronous write implementation also does not make this guarantee.
4198 	 * Fsck should catch and fix discrepancies. Arguably, the file size
4199 	 * can be over-estimated without destroying integrity when the file
4200 	 * moves into the indirect blocks (i.e., is large). If we want to
4201 	 * postpone fsck, we are stuck with this argument.
4202 	 */
4203 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
4204 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
4205 }
4206 
4207 /*
4208  * This routine is called during the completion interrupt
4209  * service routine for a disk write (from the procedure called
4210  * by the device driver to inform the filesystem caches of
4211  * a request completion).  It should be called early in this
4212  * procedure, before the block is made available to other
4213  * processes or other routines are called.
4214  */
4215 static void
4216 softdep_disk_write_complete(bp)
4217 	struct buf *bp;		/* describes the completed disk write */
4218 {
4219 	struct worklist *wk;
4220 	struct worklist *owk;
4221 	struct workhead reattach;
4222 	struct newblk *newblk;
4223 	struct allocindir *aip;
4224 	struct allocdirect *adp;
4225 	struct indirdep *indirdep;
4226 	struct inodedep *inodedep;
4227 	struct bmsafemap *bmsafemap;
4228 
4229 	/*
4230 	 * If an error occurred while doing the write, then the data
4231 	 * has not hit the disk and the dependencies cannot be unrolled.
4232 	 */
4233 	if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0)
4234 		return;
4235 	LIST_INIT(&reattach);
4236 	/*
4237 	 * This lock must not be released anywhere in this code segment.
4238 	 */
4239 	ACQUIRE_LOCK(&lk);
4240 	owk = NULL;
4241 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
4242 		WORKLIST_REMOVE(wk);
4243 		if (wk == owk)
4244 			panic("duplicate worklist: %p\n", wk);
4245 		owk = wk;
4246 		switch (wk->wk_type) {
4247 
4248 		case D_PAGEDEP:
4249 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
4250 				WORKLIST_INSERT(&reattach, wk);
4251 			continue;
4252 
4253 		case D_INODEDEP:
4254 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
4255 				WORKLIST_INSERT(&reattach, wk);
4256 			continue;
4257 
4258 		case D_BMSAFEMAP:
4259 			bmsafemap = WK_BMSAFEMAP(wk);
4260 			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
4261 				newblk->nb_state |= DEPCOMPLETE;
4262 				newblk->nb_bmsafemap = NULL;
4263 				LIST_REMOVE(newblk, nb_deps);
4264 			}
4265 			while ((adp =
4266 			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
4267 				adp->ad_state |= DEPCOMPLETE;
4268 				adp->ad_buf = NULL;
4269 				LIST_REMOVE(adp, ad_deps);
4270 				handle_allocdirect_partdone(adp);
4271 			}
4272 			while ((aip =
4273 			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
4274 				aip->ai_state |= DEPCOMPLETE;
4275 				aip->ai_buf = NULL;
4276 				LIST_REMOVE(aip, ai_deps);
4277 				handle_allocindir_partdone(aip);
4278 			}
4279 			while ((inodedep =
4280 			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
4281 				inodedep->id_state |= DEPCOMPLETE;
4282 				LIST_REMOVE(inodedep, id_deps);
4283 				inodedep->id_buf = NULL;
4284 			}
4285 			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
4286 			continue;
4287 
4288 		case D_MKDIR:
4289 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
4290 			continue;
4291 
4292 		case D_ALLOCDIRECT:
4293 			adp = WK_ALLOCDIRECT(wk);
4294 			adp->ad_state |= COMPLETE;
4295 			handle_allocdirect_partdone(adp);
4296 			continue;
4297 
4298 		case D_ALLOCINDIR:
4299 			aip = WK_ALLOCINDIR(wk);
4300 			aip->ai_state |= COMPLETE;
4301 			handle_allocindir_partdone(aip);
4302 			continue;
4303 
4304 		case D_INDIRDEP:
4305 			indirdep = WK_INDIRDEP(wk);
4306 			if (indirdep->ir_state & GOINGAWAY)
4307 				panic("disk_write_complete: indirdep gone");
4308 			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
4309 			FREE(indirdep->ir_saveddata, M_INDIRDEP);
4310 			indirdep->ir_saveddata = 0;
4311 			indirdep->ir_state &= ~UNDONE;
4312 			indirdep->ir_state |= ATTACHED;
4313 			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
4314 				handle_allocindir_partdone(aip);
4315 				if (aip == LIST_FIRST(&indirdep->ir_donehd))
4316 					panic("disk_write_complete: not gone");
4317 			}
4318 			WORKLIST_INSERT(&reattach, wk);
4319 			if ((bp->b_flags & B_DELWRI) == 0)
4320 				stat_indir_blk_ptrs++;
4321 			bdirty(bp);
4322 			continue;
4323 
4324 		default:
4325 			panic("handle_disk_write_complete: Unknown type %s",
4326 			    TYPENAME(wk->wk_type));
4327 			/* NOTREACHED */
4328 		}
4329 	}
4330 	/*
4331 	 * Reattach any requests that must be redone.
4332 	 */
4333 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
4334 		WORKLIST_REMOVE(wk);
4335 		WORKLIST_INSERT(&bp->b_dep, wk);
4336 	}
4337 	FREE_LOCK(&lk);
4338 }
4339 
4340 /*
4341  * Called from within softdep_disk_write_complete above. Note that
4342  * this routine is always called from interrupt level with further
4343  * splbio interrupts blocked.
4344  */
4345 static void
4346 handle_allocdirect_partdone(adp)
4347 	struct allocdirect *adp;	/* the completed allocdirect */
4348 {
4349 	struct allocdirectlst *listhead;
4350 	struct allocdirect *listadp;
4351 	struct inodedep *inodedep;
4352 	long bsize, delay;
4353 
4354 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4355 		return;
4356 	if (adp->ad_buf != NULL)
4357 		panic("handle_allocdirect_partdone: dangling dep");
4358 	/*
4359 	 * The on-disk inode cannot claim to be any larger than the last
4360 	 * fragment that has been written. Otherwise, the on-disk inode
4361 	 * might have fragments that were not the last block in the file
4362 	 * which would corrupt the filesystem. Thus, we cannot free any
4363 	 * allocdirects after one whose ad_oldblkno claims a fragment as
4364 	 * these blocks must be rolled back to zero before writing the inode.
4365 	 * We check the currently active set of allocdirects in id_inoupdt
4366 	 * or id_extupdt as appropriate.
4367 	 */
4368 	inodedep = adp->ad_inodedep;
4369 	bsize = inodedep->id_fs->fs_bsize;
4370 	if (adp->ad_state & EXTDATA)
4371 		listhead = &inodedep->id_extupdt;
4372 	else
4373 		listhead = &inodedep->id_inoupdt;
4374 	TAILQ_FOREACH(listadp, listhead, ad_next) {
4375 		/* found our block */
4376 		if (listadp == adp)
4377 			break;
4378 		/* continue if ad_oldlbn is not a fragment */
4379 		if (listadp->ad_oldsize == 0 ||
4380 		    listadp->ad_oldsize == bsize)
4381 			continue;
4382 		/* hit a fragment */
4383 		return;
4384 	}
4385 	/*
4386 	 * If we have reached the end of the current list without
4387 	 * finding the just finished dependency, then it must be
4388 	 * on the future dependency list. Future dependencies cannot
4389 	 * be freed until they are moved to the current list.
4390 	 */
4391 	if (listadp == NULL) {
4392 #ifdef DEBUG
4393 		if (adp->ad_state & EXTDATA)
4394 			listhead = &inodedep->id_newextupdt;
4395 		else
4396 			listhead = &inodedep->id_newinoupdt;
4397 		TAILQ_FOREACH(listadp, listhead, ad_next)
4398 			/* found our block */
4399 			if (listadp == adp)
4400 				break;
4401 		if (listadp == NULL)
4402 			panic("handle_allocdirect_partdone: lost dep");
4403 #endif /* DEBUG */
4404 		return;
4405 	}
4406 	/*
4407 	 * If we have found the just finished dependency, then free
4408 	 * it along with anything that follows it that is complete.
4409 	 * If the inode still has a bitmap dependency, then it has
4410 	 * never been written to disk, hence the on-disk inode cannot
4411 	 * reference the old fragment so we can free it without delay.
4412 	 */
4413 	delay = (inodedep->id_state & DEPCOMPLETE);
4414 	for (; adp; adp = listadp) {
4415 		listadp = TAILQ_NEXT(adp, ad_next);
4416 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4417 			return;
4418 		free_allocdirect(listhead, adp, delay);
4419 	}
4420 }
4421 
4422 /*
4423  * Called from within softdep_disk_write_complete above. Note that
4424  * this routine is always called from interrupt level with further
4425  * splbio interrupts blocked.
4426  */
4427 static void
4428 handle_allocindir_partdone(aip)
4429 	struct allocindir *aip;		/* the completed allocindir */
4430 {
4431 	struct indirdep *indirdep;
4432 
4433 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
4434 		return;
4435 	if (aip->ai_buf != NULL)
4436 		panic("handle_allocindir_partdone: dangling dependency");
4437 	indirdep = aip->ai_indirdep;
4438 	if (indirdep->ir_state & UNDONE) {
4439 		LIST_REMOVE(aip, ai_next);
4440 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
4441 		return;
4442 	}
4443 	if (indirdep->ir_state & UFS1FMT)
4444 		((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4445 		    aip->ai_newblkno;
4446 	else
4447 		((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4448 		    aip->ai_newblkno;
4449 	LIST_REMOVE(aip, ai_next);
4450 	if (aip->ai_freefrag != NULL)
4451 		add_to_worklist(&aip->ai_freefrag->ff_list);
4452 	WORKITEM_FREE(aip, D_ALLOCINDIR);
4453 }
4454 
4455 /*
4456  * Called from within softdep_disk_write_complete above to restore
4457  * in-memory inode block contents to their most up-to-date state. Note
4458  * that this routine is always called from interrupt level with further
4459  * splbio interrupts blocked.
4460  */
4461 static int
4462 handle_written_inodeblock(inodedep, bp)
4463 	struct inodedep *inodedep;
4464 	struct buf *bp;		/* buffer containing the inode block */
4465 {
4466 	struct worklist *wk, *filefree;
4467 	struct allocdirect *adp, *nextadp;
4468 	struct ufs1_dinode *dp1 = NULL;
4469 	struct ufs2_dinode *dp2 = NULL;
4470 	int hadchanges, fstype;
4471 
4472 	if ((inodedep->id_state & IOSTARTED) == 0)
4473 		panic("handle_written_inodeblock: not started");
4474 	inodedep->id_state &= ~IOSTARTED;
4475 	if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
4476 		fstype = UFS1;
4477 		dp1 = (struct ufs1_dinode *)bp->b_data +
4478 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4479 	} else {
4480 		fstype = UFS2;
4481 		dp2 = (struct ufs2_dinode *)bp->b_data +
4482 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4483 	}
4484 	/*
4485 	 * If we had to rollback the inode allocation because of
4486 	 * bitmaps being incomplete, then simply restore it.
4487 	 * Keep the block dirty so that it will not be reclaimed until
4488 	 * all associated dependencies have been cleared and the
4489 	 * corresponding updates written to disk.
4490 	 */
4491 	if (inodedep->id_savedino1 != NULL) {
4492 		if (fstype == UFS1)
4493 			*dp1 = *inodedep->id_savedino1;
4494 		else
4495 			*dp2 = *inodedep->id_savedino2;
4496 		FREE(inodedep->id_savedino1, M_SAVEDINO);
4497 		inodedep->id_savedino1 = NULL;
4498 		if ((bp->b_flags & B_DELWRI) == 0)
4499 			stat_inode_bitmap++;
4500 		bdirty(bp);
4501 		return (1);
4502 	}
4503 	inodedep->id_state |= COMPLETE;
4504 	/*
4505 	 * Roll forward anything that had to be rolled back before
4506 	 * the inode could be updated.
4507 	 */
4508 	hadchanges = 0;
4509 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
4510 		nextadp = TAILQ_NEXT(adp, ad_next);
4511 		if (adp->ad_state & ATTACHED)
4512 			panic("handle_written_inodeblock: new entry");
4513 		if (fstype == UFS1) {
4514 			if (adp->ad_lbn < NDADDR) {
4515 				if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno)
4516 					panic("%s %s #%jd mismatch %d != %jd",
4517 					    "handle_written_inodeblock:",
4518 					    "direct pointer",
4519 					    (intmax_t)adp->ad_lbn,
4520 					    dp1->di_db[adp->ad_lbn],
4521 					    (intmax_t)adp->ad_oldblkno);
4522 				dp1->di_db[adp->ad_lbn] = adp->ad_newblkno;
4523 			} else {
4524 				if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0)
4525 					panic("%s: %s #%jd allocated as %d",
4526 					    "handle_written_inodeblock",
4527 					    "indirect pointer",
4528 					    (intmax_t)adp->ad_lbn - NDADDR,
4529 					    dp1->di_ib[adp->ad_lbn - NDADDR]);
4530 				dp1->di_ib[adp->ad_lbn - NDADDR] =
4531 				    adp->ad_newblkno;
4532 			}
4533 		} else {
4534 			if (adp->ad_lbn < NDADDR) {
4535 				if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno)
4536 					panic("%s: %s #%jd %s %jd != %jd",
4537 					    "handle_written_inodeblock",
4538 					    "direct pointer",
4539 					    (intmax_t)adp->ad_lbn, "mismatch",
4540 					    (intmax_t)dp2->di_db[adp->ad_lbn],
4541 					    (intmax_t)adp->ad_oldblkno);
4542 				dp2->di_db[adp->ad_lbn] = adp->ad_newblkno;
4543 			} else {
4544 				if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0)
4545 					panic("%s: %s #%jd allocated as %jd",
4546 					    "handle_written_inodeblock",
4547 					    "indirect pointer",
4548 					    (intmax_t)adp->ad_lbn - NDADDR,
4549 					    (intmax_t)
4550 					    dp2->di_ib[adp->ad_lbn - NDADDR]);
4551 				dp2->di_ib[adp->ad_lbn - NDADDR] =
4552 				    adp->ad_newblkno;
4553 			}
4554 		}
4555 		adp->ad_state &= ~UNDONE;
4556 		adp->ad_state |= ATTACHED;
4557 		hadchanges = 1;
4558 	}
4559 	for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
4560 		nextadp = TAILQ_NEXT(adp, ad_next);
4561 		if (adp->ad_state & ATTACHED)
4562 			panic("handle_written_inodeblock: new entry");
4563 		if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno)
4564 			panic("%s: direct pointers #%jd %s %jd != %jd",
4565 			    "handle_written_inodeblock",
4566 			    (intmax_t)adp->ad_lbn, "mismatch",
4567 			    (intmax_t)dp2->di_extb[adp->ad_lbn],
4568 			    (intmax_t)adp->ad_oldblkno);
4569 		dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno;
4570 		adp->ad_state &= ~UNDONE;
4571 		adp->ad_state |= ATTACHED;
4572 		hadchanges = 1;
4573 	}
4574 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
4575 		stat_direct_blk_ptrs++;
4576 	/*
4577 	 * Reset the file size to its most up-to-date value.
4578 	 */
4579 	if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1)
4580 		panic("handle_written_inodeblock: bad size");
4581 	if (fstype == UFS1) {
4582 		if (dp1->di_size != inodedep->id_savedsize) {
4583 			dp1->di_size = inodedep->id_savedsize;
4584 			hadchanges = 1;
4585 		}
4586 	} else {
4587 		if (dp2->di_size != inodedep->id_savedsize) {
4588 			dp2->di_size = inodedep->id_savedsize;
4589 			hadchanges = 1;
4590 		}
4591 		if (dp2->di_extsize != inodedep->id_savedextsize) {
4592 			dp2->di_extsize = inodedep->id_savedextsize;
4593 			hadchanges = 1;
4594 		}
4595 	}
4596 	inodedep->id_savedsize = -1;
4597 	inodedep->id_savedextsize = -1;
4598 	/*
4599 	 * If there were any rollbacks in the inode block, then it must be
4600 	 * marked dirty so that its will eventually get written back in
4601 	 * its correct form.
4602 	 */
4603 	if (hadchanges)
4604 		bdirty(bp);
4605 	/*
4606 	 * Process any allocdirects that completed during the update.
4607 	 */
4608 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
4609 		handle_allocdirect_partdone(adp);
4610 	if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
4611 		handle_allocdirect_partdone(adp);
4612 	/*
4613 	 * Process deallocations that were held pending until the
4614 	 * inode had been written to disk. Freeing of the inode
4615 	 * is delayed until after all blocks have been freed to
4616 	 * avoid creation of new <vfsid, inum, lbn> triples
4617 	 * before the old ones have been deleted.
4618 	 */
4619 	filefree = NULL;
4620 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
4621 		WORKLIST_REMOVE(wk);
4622 		switch (wk->wk_type) {
4623 
4624 		case D_FREEFILE:
4625 			/*
4626 			 * We defer adding filefree to the worklist until
4627 			 * all other additions have been made to ensure
4628 			 * that it will be done after all the old blocks
4629 			 * have been freed.
4630 			 */
4631 			if (filefree != NULL)
4632 				panic("handle_written_inodeblock: filefree");
4633 			filefree = wk;
4634 			continue;
4635 
4636 		case D_MKDIR:
4637 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
4638 			continue;
4639 
4640 		case D_DIRADD:
4641 			diradd_inode_written(WK_DIRADD(wk), inodedep);
4642 			continue;
4643 
4644 		case D_FREEBLKS:
4645 			wk->wk_state |= COMPLETE;
4646 			if ((wk->wk_state  & ALLCOMPLETE) != ALLCOMPLETE)
4647 				continue;
4648 			 /* -- fall through -- */
4649 		case D_FREEFRAG:
4650 		case D_DIRREM:
4651 			add_to_worklist(wk);
4652 			continue;
4653 
4654 		case D_NEWDIRBLK:
4655 			free_newdirblk(WK_NEWDIRBLK(wk));
4656 			continue;
4657 
4658 		default:
4659 			panic("handle_written_inodeblock: Unknown type %s",
4660 			    TYPENAME(wk->wk_type));
4661 			/* NOTREACHED */
4662 		}
4663 	}
4664 	if (filefree != NULL) {
4665 		if (free_inodedep(inodedep) == 0)
4666 			panic("handle_written_inodeblock: live inodedep");
4667 		add_to_worklist(filefree);
4668 		return (0);
4669 	}
4670 
4671 	/*
4672 	 * If no outstanding dependencies, free it.
4673 	 */
4674 	if (free_inodedep(inodedep) ||
4675 	    (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
4676 	     TAILQ_FIRST(&inodedep->id_extupdt) == 0))
4677 		return (0);
4678 	return (hadchanges);
4679 }
4680 
4681 /*
4682  * Process a diradd entry after its dependent inode has been written.
4683  * This routine must be called with splbio interrupts blocked.
4684  */
4685 static void
4686 diradd_inode_written(dap, inodedep)
4687 	struct diradd *dap;
4688 	struct inodedep *inodedep;
4689 {
4690 	struct pagedep *pagedep;
4691 
4692 	dap->da_state |= COMPLETE;
4693 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4694 		if (dap->da_state & DIRCHG)
4695 			pagedep = dap->da_previous->dm_pagedep;
4696 		else
4697 			pagedep = dap->da_pagedep;
4698 		LIST_REMOVE(dap, da_pdlist);
4699 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4700 	}
4701 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
4702 }
4703 
4704 /*
4705  * Handle the completion of a mkdir dependency.
4706  */
4707 static void
4708 handle_written_mkdir(mkdir, type)
4709 	struct mkdir *mkdir;
4710 	int type;
4711 {
4712 	struct diradd *dap;
4713 	struct pagedep *pagedep;
4714 
4715 	if (mkdir->md_state != type)
4716 		panic("handle_written_mkdir: bad type");
4717 	dap = mkdir->md_diradd;
4718 	dap->da_state &= ~type;
4719 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
4720 		dap->da_state |= DEPCOMPLETE;
4721 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4722 		if (dap->da_state & DIRCHG)
4723 			pagedep = dap->da_previous->dm_pagedep;
4724 		else
4725 			pagedep = dap->da_pagedep;
4726 		LIST_REMOVE(dap, da_pdlist);
4727 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4728 	}
4729 	LIST_REMOVE(mkdir, md_mkdirs);
4730 	WORKITEM_FREE(mkdir, D_MKDIR);
4731 }
4732 
4733 /*
4734  * Called from within softdep_disk_write_complete above.
4735  * A write operation was just completed. Removed inodes can
4736  * now be freed and associated block pointers may be committed.
4737  * Note that this routine is always called from interrupt level
4738  * with further splbio interrupts blocked.
4739  */
4740 static int
4741 handle_written_filepage(pagedep, bp)
4742 	struct pagedep *pagedep;
4743 	struct buf *bp;		/* buffer containing the written page */
4744 {
4745 	struct dirrem *dirrem;
4746 	struct diradd *dap, *nextdap;
4747 	struct direct *ep;
4748 	int i, chgs;
4749 
4750 	if ((pagedep->pd_state & IOSTARTED) == 0)
4751 		panic("handle_written_filepage: not started");
4752 	pagedep->pd_state &= ~IOSTARTED;
4753 	/*
4754 	 * Process any directory removals that have been committed.
4755 	 */
4756 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
4757 		LIST_REMOVE(dirrem, dm_next);
4758 		dirrem->dm_dirinum = pagedep->pd_ino;
4759 		add_to_worklist(&dirrem->dm_list);
4760 	}
4761 	/*
4762 	 * Free any directory additions that have been committed.
4763 	 * If it is a newly allocated block, we have to wait until
4764 	 * the on-disk directory inode claims the new block.
4765 	 */
4766 	if ((pagedep->pd_state & NEWBLOCK) == 0)
4767 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
4768 			free_diradd(dap);
4769 	/*
4770 	 * Uncommitted directory entries must be restored.
4771 	 */
4772 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
4773 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
4774 		     dap = nextdap) {
4775 			nextdap = LIST_NEXT(dap, da_pdlist);
4776 			if (dap->da_state & ATTACHED)
4777 				panic("handle_written_filepage: attached");
4778 			ep = (struct direct *)
4779 			    ((char *)bp->b_data + dap->da_offset);
4780 			ep->d_ino = dap->da_newinum;
4781 			dap->da_state &= ~UNDONE;
4782 			dap->da_state |= ATTACHED;
4783 			chgs = 1;
4784 			/*
4785 			 * If the inode referenced by the directory has
4786 			 * been written out, then the dependency can be
4787 			 * moved to the pending list.
4788 			 */
4789 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4790 				LIST_REMOVE(dap, da_pdlist);
4791 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
4792 				    da_pdlist);
4793 			}
4794 		}
4795 	}
4796 	/*
4797 	 * If there were any rollbacks in the directory, then it must be
4798 	 * marked dirty so that its will eventually get written back in
4799 	 * its correct form.
4800 	 */
4801 	if (chgs) {
4802 		if ((bp->b_flags & B_DELWRI) == 0)
4803 			stat_dir_entry++;
4804 		bdirty(bp);
4805 		return (1);
4806 	}
4807 	/*
4808 	 * If we are not waiting for a new directory block to be
4809 	 * claimed by its inode, then the pagedep will be freed.
4810 	 * Otherwise it will remain to track any new entries on
4811 	 * the page in case they are fsync'ed.
4812 	 */
4813 	if ((pagedep->pd_state & NEWBLOCK) == 0) {
4814 		LIST_REMOVE(pagedep, pd_hash);
4815 		WORKITEM_FREE(pagedep, D_PAGEDEP);
4816 	}
4817 	return (0);
4818 }
4819 
4820 /*
4821  * Writing back in-core inode structures.
4822  *
4823  * The filesystem only accesses an inode's contents when it occupies an
4824  * "in-core" inode structure.  These "in-core" structures are separate from
4825  * the page frames used to cache inode blocks.  Only the latter are
4826  * transferred to/from the disk.  So, when the updated contents of the
4827  * "in-core" inode structure are copied to the corresponding in-memory inode
4828  * block, the dependencies are also transferred.  The following procedure is
4829  * called when copying a dirty "in-core" inode to a cached inode block.
4830  */
4831 
4832 /*
4833  * Called when an inode is loaded from disk. If the effective link count
4834  * differed from the actual link count when it was last flushed, then we
4835  * need to ensure that the correct effective link count is put back.
4836  */
4837 void
4838 softdep_load_inodeblock(ip)
4839 	struct inode *ip;	/* the "in_core" copy of the inode */
4840 {
4841 	struct inodedep *inodedep;
4842 
4843 	/*
4844 	 * Check for alternate nlink count.
4845 	 */
4846 	ip->i_effnlink = ip->i_nlink;
4847 	ACQUIRE_LOCK(&lk);
4848 	if (inodedep_lookup(UFSTOVFS(ip->i_ump),
4849 	    ip->i_number, 0, &inodedep) == 0) {
4850 		FREE_LOCK(&lk);
4851 		return;
4852 	}
4853 	ip->i_effnlink -= inodedep->id_nlinkdelta;
4854 	if (inodedep->id_state & SPACECOUNTED)
4855 		ip->i_flag |= IN_SPACECOUNTED;
4856 	FREE_LOCK(&lk);
4857 }
4858 
4859 /*
4860  * This routine is called just before the "in-core" inode
4861  * information is to be copied to the in-memory inode block.
4862  * Recall that an inode block contains several inodes. If
4863  * the force flag is set, then the dependencies will be
4864  * cleared so that the update can always be made. Note that
4865  * the buffer is locked when this routine is called, so we
4866  * will never be in the middle of writing the inode block
4867  * to disk.
4868  */
4869 void
4870 softdep_update_inodeblock(ip, bp, waitfor)
4871 	struct inode *ip;	/* the "in_core" copy of the inode */
4872 	struct buf *bp;		/* the buffer containing the inode block */
4873 	int waitfor;		/* nonzero => update must be allowed */
4874 {
4875 	struct inodedep *inodedep;
4876 	struct worklist *wk;
4877 	struct mount *mp;
4878 	struct buf *ibp;
4879 	int error;
4880 
4881 	/*
4882 	 * If the effective link count is not equal to the actual link
4883 	 * count, then we must track the difference in an inodedep while
4884 	 * the inode is (potentially) tossed out of the cache. Otherwise,
4885 	 * if there is no existing inodedep, then there are no dependencies
4886 	 * to track.
4887 	 */
4888 	mp = UFSTOVFS(ip->i_ump);
4889 	ACQUIRE_LOCK(&lk);
4890 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
4891 		FREE_LOCK(&lk);
4892 		if (ip->i_effnlink != ip->i_nlink)
4893 			panic("softdep_update_inodeblock: bad link count");
4894 		return;
4895 	}
4896 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
4897 		panic("softdep_update_inodeblock: bad delta");
4898 	/*
4899 	 * Changes have been initiated. Anything depending on these
4900 	 * changes cannot occur until this inode has been written.
4901 	 */
4902 	inodedep->id_state &= ~COMPLETE;
4903 	if ((inodedep->id_state & ONWORKLIST) == 0)
4904 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
4905 	/*
4906 	 * Any new dependencies associated with the incore inode must
4907 	 * now be moved to the list associated with the buffer holding
4908 	 * the in-memory copy of the inode. Once merged process any
4909 	 * allocdirects that are completed by the merger.
4910 	 */
4911 	merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
4912 	if (!TAILQ_EMPTY(&inodedep->id_inoupdt))
4913 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
4914 	merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
4915 	if (!TAILQ_EMPTY(&inodedep->id_extupdt))
4916 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt));
4917 	/*
4918 	 * Now that the inode has been pushed into the buffer, the
4919 	 * operations dependent on the inode being written to disk
4920 	 * can be moved to the id_bufwait so that they will be
4921 	 * processed when the buffer I/O completes.
4922 	 */
4923 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
4924 		WORKLIST_REMOVE(wk);
4925 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
4926 	}
4927 	/*
4928 	 * Newly allocated inodes cannot be written until the bitmap
4929 	 * that allocates them have been written (indicated by
4930 	 * DEPCOMPLETE being set in id_state). If we are doing a
4931 	 * forced sync (e.g., an fsync on a file), we force the bitmap
4932 	 * to be written so that the update can be done.
4933 	 */
4934 	if (waitfor == 0) {
4935 		FREE_LOCK(&lk);
4936 		return;
4937 	}
4938 retry:
4939 	if ((inodedep->id_state & DEPCOMPLETE) != 0) {
4940 		FREE_LOCK(&lk);
4941 		return;
4942 	}
4943 	ibp = inodedep->id_buf;
4944 	ibp = getdirtybuf(ibp, &lk, MNT_WAIT);
4945 	if (ibp == NULL) {
4946 		/*
4947 		 * If ibp came back as NULL, the dependency could have been
4948 		 * freed while we slept.  Look it up again, and check to see
4949 		 * that it has completed.
4950 		 */
4951 		if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
4952 			goto retry;
4953 		FREE_LOCK(&lk);
4954 		return;
4955 	}
4956 	FREE_LOCK(&lk);
4957 	if ((error = bwrite(ibp)) != 0)
4958 		softdep_error("softdep_update_inodeblock: bwrite", error);
4959 }
4960 
4961 /*
4962  * Merge the a new inode dependency list (such as id_newinoupdt) into an
4963  * old inode dependency list (such as id_inoupdt). This routine must be
4964  * called with splbio interrupts blocked.
4965  */
4966 static void
4967 merge_inode_lists(newlisthead, oldlisthead)
4968 	struct allocdirectlst *newlisthead;
4969 	struct allocdirectlst *oldlisthead;
4970 {
4971 	struct allocdirect *listadp, *newadp;
4972 
4973 	newadp = TAILQ_FIRST(newlisthead);
4974 	for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
4975 		if (listadp->ad_lbn < newadp->ad_lbn) {
4976 			listadp = TAILQ_NEXT(listadp, ad_next);
4977 			continue;
4978 		}
4979 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4980 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
4981 		if (listadp->ad_lbn == newadp->ad_lbn) {
4982 			allocdirect_merge(oldlisthead, newadp,
4983 			    listadp);
4984 			listadp = newadp;
4985 		}
4986 		newadp = TAILQ_FIRST(newlisthead);
4987 	}
4988 	while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
4989 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4990 		TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
4991 	}
4992 }
4993 
4994 /*
4995  * If we are doing an fsync, then we must ensure that any directory
4996  * entries for the inode have been written after the inode gets to disk.
4997  */
4998 int
4999 softdep_fsync(vp)
5000 	struct vnode *vp;	/* the "in_core" copy of the inode */
5001 {
5002 	struct inodedep *inodedep;
5003 	struct pagedep *pagedep;
5004 	struct worklist *wk;
5005 	struct diradd *dap;
5006 	struct mount *mp;
5007 	struct vnode *pvp;
5008 	struct inode *ip;
5009 	struct buf *bp;
5010 	struct fs *fs;
5011 	struct thread *td = curthread;
5012 	int error, flushparent, pagedep_new_block;
5013 	ino_t parentino;
5014 	ufs_lbn_t lbn;
5015 
5016 	ip = VTOI(vp);
5017 	fs = ip->i_fs;
5018 	mp = vp->v_mount;
5019 	ACQUIRE_LOCK(&lk);
5020 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
5021 		FREE_LOCK(&lk);
5022 		return (0);
5023 	}
5024 	if (!LIST_EMPTY(&inodedep->id_inowait) ||
5025 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
5026 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
5027 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
5028 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
5029 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt))
5030 		panic("softdep_fsync: pending ops");
5031 	for (error = 0, flushparent = 0; ; ) {
5032 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
5033 			break;
5034 		if (wk->wk_type != D_DIRADD)
5035 			panic("softdep_fsync: Unexpected type %s",
5036 			    TYPENAME(wk->wk_type));
5037 		dap = WK_DIRADD(wk);
5038 		/*
5039 		 * Flush our parent if this directory entry has a MKDIR_PARENT
5040 		 * dependency or is contained in a newly allocated block.
5041 		 */
5042 		if (dap->da_state & DIRCHG)
5043 			pagedep = dap->da_previous->dm_pagedep;
5044 		else
5045 			pagedep = dap->da_pagedep;
5046 		parentino = pagedep->pd_ino;
5047 		lbn = pagedep->pd_lbn;
5048 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
5049 			panic("softdep_fsync: dirty");
5050 		if ((dap->da_state & MKDIR_PARENT) ||
5051 		    (pagedep->pd_state & NEWBLOCK))
5052 			flushparent = 1;
5053 		else
5054 			flushparent = 0;
5055 		/*
5056 		 * If we are being fsync'ed as part of vgone'ing this vnode,
5057 		 * then we will not be able to release and recover the
5058 		 * vnode below, so we just have to give up on writing its
5059 		 * directory entry out. It will eventually be written, just
5060 		 * not now, but then the user was not asking to have it
5061 		 * written, so we are not breaking any promises.
5062 		 */
5063 		if (vp->v_iflag & VI_DOOMED)
5064 			break;
5065 		/*
5066 		 * We prevent deadlock by always fetching inodes from the
5067 		 * root, moving down the directory tree. Thus, when fetching
5068 		 * our parent directory, we first try to get the lock. If
5069 		 * that fails, we must unlock ourselves before requesting
5070 		 * the lock on our parent. See the comment in ufs_lookup
5071 		 * for details on possible races.
5072 		 */
5073 		FREE_LOCK(&lk);
5074 		if (ffs_vget(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) {
5075 			VOP_UNLOCK(vp, 0, td);
5076 			error = ffs_vget(mp, parentino, LK_EXCLUSIVE, &pvp);
5077 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
5078 			if (error != 0)
5079 				return (error);
5080 		}
5081 		/*
5082 		 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
5083 		 * that are contained in direct blocks will be resolved by
5084 		 * doing a ffs_update. Pagedeps contained in indirect blocks
5085 		 * may require a complete sync'ing of the directory. So, we
5086 		 * try the cheap and fast ffs_update first, and if that fails,
5087 		 * then we do the slower ffs_syncvnode of the directory.
5088 		 */
5089 		if (flushparent) {
5090 			int locked;
5091 
5092 			if ((error = ffs_update(pvp, 1)) != 0) {
5093 				vput(pvp);
5094 				return (error);
5095 			}
5096 			ACQUIRE_LOCK(&lk);
5097 			locked = 1;
5098 			if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) {
5099 				if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) {
5100 					if (wk->wk_type != D_DIRADD)
5101 						panic("softdep_fsync: Unexpected type %s",
5102 						      TYPENAME(wk->wk_type));
5103 					dap = WK_DIRADD(wk);
5104 					if (dap->da_state & DIRCHG)
5105 						pagedep = dap->da_previous->dm_pagedep;
5106 					else
5107 						pagedep = dap->da_pagedep;
5108 					pagedep_new_block = pagedep->pd_state & NEWBLOCK;
5109 					FREE_LOCK(&lk);
5110 					locked = 0;
5111 					if (pagedep_new_block &&
5112 					    (error = ffs_syncvnode(pvp, MNT_WAIT))) {
5113 						vput(pvp);
5114 						return (error);
5115 					}
5116 				}
5117 			}
5118 			if (locked)
5119 				FREE_LOCK(&lk);
5120 		}
5121 		/*
5122 		 * Flush directory page containing the inode's name.
5123 		 */
5124 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
5125 		    &bp);
5126 		if (error == 0)
5127 			error = bwrite(bp);
5128 		else
5129 			brelse(bp);
5130 		vput(pvp);
5131 		if (error != 0)
5132 			return (error);
5133 		ACQUIRE_LOCK(&lk);
5134 		if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
5135 			break;
5136 	}
5137 	FREE_LOCK(&lk);
5138 	return (0);
5139 }
5140 
5141 /*
5142  * Flush all the dirty bitmaps associated with the block device
5143  * before flushing the rest of the dirty blocks so as to reduce
5144  * the number of dependencies that will have to be rolled back.
5145  */
5146 void
5147 softdep_fsync_mountdev(vp)
5148 	struct vnode *vp;
5149 {
5150 	struct buf *bp, *nbp;
5151 	struct worklist *wk;
5152 
5153 	if (!vn_isdisk(vp, NULL))
5154 		panic("softdep_fsync_mountdev: vnode not a disk");
5155 restart:
5156 	ACQUIRE_LOCK(&lk);
5157 	VI_LOCK(vp);
5158 	TAILQ_FOREACH_SAFE(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs, nbp) {
5159 		/*
5160 		 * If it is already scheduled, skip to the next buffer.
5161 		 */
5162 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
5163 			continue;
5164 
5165 		if ((bp->b_flags & B_DELWRI) == 0)
5166 			panic("softdep_fsync_mountdev: not dirty");
5167 		/*
5168 		 * We are only interested in bitmaps with outstanding
5169 		 * dependencies.
5170 		 */
5171 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
5172 		    wk->wk_type != D_BMSAFEMAP ||
5173 		    (bp->b_vflags & BV_BKGRDINPROG)) {
5174 			BUF_UNLOCK(bp);
5175 			continue;
5176 		}
5177 		VI_UNLOCK(vp);
5178 		FREE_LOCK(&lk);
5179 		bremfree(bp);
5180 		(void) bawrite(bp);
5181 		goto restart;
5182 	}
5183 	FREE_LOCK(&lk);
5184 	drain_output(vp);
5185 	VI_UNLOCK(vp);
5186 }
5187 
5188 /*
5189  * This routine is called when we are trying to synchronously flush a
5190  * file. This routine must eliminate any filesystem metadata dependencies
5191  * so that the syncing routine can succeed by pushing the dirty blocks
5192  * associated with the file. If any I/O errors occur, they are returned.
5193  */
5194 int
5195 softdep_sync_metadata(struct vnode *vp)
5196 {
5197 	struct pagedep *pagedep;
5198 	struct allocdirect *adp;
5199 	struct allocindir *aip;
5200 	struct buf *bp, *nbp;
5201 	struct worklist *wk;
5202 	int i, error, waitfor;
5203 
5204 	if (!DOINGSOFTDEP(vp))
5205 		return (0);
5206 	/*
5207 	 * Ensure that any direct block dependencies have been cleared.
5208 	 */
5209 	ACQUIRE_LOCK(&lk);
5210 	if ((error = flush_inodedep_deps(vp->v_mount, VTOI(vp)->i_number))) {
5211 		FREE_LOCK(&lk);
5212 		return (error);
5213 	}
5214 	FREE_LOCK(&lk);
5215 	/*
5216 	 * For most files, the only metadata dependencies are the
5217 	 * cylinder group maps that allocate their inode or blocks.
5218 	 * The block allocation dependencies can be found by traversing
5219 	 * the dependency lists for any buffers that remain on their
5220 	 * dirty buffer list. The inode allocation dependency will
5221 	 * be resolved when the inode is updated with MNT_WAIT.
5222 	 * This work is done in two passes. The first pass grabs most
5223 	 * of the buffers and begins asynchronously writing them. The
5224 	 * only way to wait for these asynchronous writes is to sleep
5225 	 * on the filesystem vnode which may stay busy for a long time
5226 	 * if the filesystem is active. So, instead, we make a second
5227 	 * pass over the dependencies blocking on each write. In the
5228 	 * usual case we will be blocking against a write that we
5229 	 * initiated, so when it is done the dependency will have been
5230 	 * resolved. Thus the second pass is expected to end quickly.
5231 	 */
5232 	waitfor = MNT_NOWAIT;
5233 
5234 top:
5235 	/*
5236 	 * We must wait for any I/O in progress to finish so that
5237 	 * all potential buffers on the dirty list will be visible.
5238 	 */
5239 	VI_LOCK(vp);
5240 	drain_output(vp);
5241 	while ((bp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd)) != NULL) {
5242 		bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT);
5243 		if (bp)
5244 			break;
5245 	}
5246 	VI_UNLOCK(vp);
5247 	if (bp == NULL)
5248 		return (0);
5249 loop:
5250 	/* While syncing snapshots, we must allow recursive lookups */
5251 	bp->b_lock.lk_flags |= LK_CANRECURSE;
5252 	ACQUIRE_LOCK(&lk);
5253 	/*
5254 	 * As we hold the buffer locked, none of its dependencies
5255 	 * will disappear.
5256 	 */
5257 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5258 		switch (wk->wk_type) {
5259 
5260 		case D_ALLOCDIRECT:
5261 			adp = WK_ALLOCDIRECT(wk);
5262 			if (adp->ad_state & DEPCOMPLETE)
5263 				continue;
5264 			nbp = adp->ad_buf;
5265 			nbp = getdirtybuf(nbp, &lk, waitfor);
5266 			if (nbp == NULL)
5267 				continue;
5268 			FREE_LOCK(&lk);
5269 			if (waitfor == MNT_NOWAIT) {
5270 				bawrite(nbp);
5271 			} else if ((error = bwrite(nbp)) != 0) {
5272 				break;
5273 			}
5274 			ACQUIRE_LOCK(&lk);
5275 			continue;
5276 
5277 		case D_ALLOCINDIR:
5278 			aip = WK_ALLOCINDIR(wk);
5279 			if (aip->ai_state & DEPCOMPLETE)
5280 				continue;
5281 			nbp = aip->ai_buf;
5282 			nbp = getdirtybuf(nbp, &lk, waitfor);
5283 			if (nbp == NULL)
5284 				continue;
5285 			FREE_LOCK(&lk);
5286 			if (waitfor == MNT_NOWAIT) {
5287 				bawrite(nbp);
5288 			} else if ((error = bwrite(nbp)) != 0) {
5289 				break;
5290 			}
5291 			ACQUIRE_LOCK(&lk);
5292 			continue;
5293 
5294 		case D_INDIRDEP:
5295 		restart:
5296 
5297 			LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
5298 				if (aip->ai_state & DEPCOMPLETE)
5299 					continue;
5300 				nbp = aip->ai_buf;
5301 				nbp = getdirtybuf(nbp, &lk, MNT_WAIT);
5302 				if (nbp == NULL)
5303 					goto restart;
5304 				FREE_LOCK(&lk);
5305 				if ((error = bwrite(nbp)) != 0) {
5306 					goto loop_end;
5307 				}
5308 				ACQUIRE_LOCK(&lk);
5309 				goto restart;
5310 			}
5311 			continue;
5312 
5313 		case D_INODEDEP:
5314 			if ((error = flush_inodedep_deps(wk->wk_mp,
5315 			    WK_INODEDEP(wk)->id_ino)) != 0) {
5316 				FREE_LOCK(&lk);
5317 				break;
5318 			}
5319 			continue;
5320 
5321 		case D_PAGEDEP:
5322 			/*
5323 			 * We are trying to sync a directory that may
5324 			 * have dependencies on both its own metadata
5325 			 * and/or dependencies on the inodes of any
5326 			 * recently allocated files. We walk its diradd
5327 			 * lists pushing out the associated inode.
5328 			 */
5329 			pagedep = WK_PAGEDEP(wk);
5330 			for (i = 0; i < DAHASHSZ; i++) {
5331 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
5332 					continue;
5333 				if ((error =
5334 				    flush_pagedep_deps(vp, wk->wk_mp,
5335 						&pagedep->pd_diraddhd[i]))) {
5336 					FREE_LOCK(&lk);
5337 					goto loop_end;
5338 				}
5339 			}
5340 			continue;
5341 
5342 		case D_MKDIR:
5343 			/*
5344 			 * This case should never happen if the vnode has
5345 			 * been properly sync'ed. However, if this function
5346 			 * is used at a place where the vnode has not yet
5347 			 * been sync'ed, this dependency can show up. So,
5348 			 * rather than panic, just flush it.
5349 			 */
5350 			nbp = WK_MKDIR(wk)->md_buf;
5351 			nbp = getdirtybuf(nbp, &lk, waitfor);
5352 			if (nbp == NULL)
5353 				continue;
5354 			FREE_LOCK(&lk);
5355 			if (waitfor == MNT_NOWAIT) {
5356 				bawrite(nbp);
5357 			} else if ((error = bwrite(nbp)) != 0) {
5358 				break;
5359 			}
5360 			ACQUIRE_LOCK(&lk);
5361 			continue;
5362 
5363 		case D_BMSAFEMAP:
5364 			/*
5365 			 * This case should never happen if the vnode has
5366 			 * been properly sync'ed. However, if this function
5367 			 * is used at a place where the vnode has not yet
5368 			 * been sync'ed, this dependency can show up. So,
5369 			 * rather than panic, just flush it.
5370 			 */
5371 			nbp = WK_BMSAFEMAP(wk)->sm_buf;
5372 			nbp = getdirtybuf(nbp, &lk, waitfor);
5373 			if (nbp == NULL)
5374 				continue;
5375 			FREE_LOCK(&lk);
5376 			if (waitfor == MNT_NOWAIT) {
5377 				bawrite(nbp);
5378 			} else if ((error = bwrite(nbp)) != 0) {
5379 				break;
5380 			}
5381 			ACQUIRE_LOCK(&lk);
5382 			continue;
5383 
5384 		default:
5385 			panic("softdep_sync_metadata: Unknown type %s",
5386 			    TYPENAME(wk->wk_type));
5387 			/* NOTREACHED */
5388 		}
5389 	loop_end:
5390 		/* We reach here only in error and unlocked */
5391 		if (error == 0)
5392 			panic("softdep_sync_metadata: zero error");
5393 		bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5394 		bawrite(bp);
5395 		return (error);
5396 	}
5397 	FREE_LOCK(&lk);
5398 	VI_LOCK(vp);
5399 	while ((nbp = TAILQ_NEXT(bp, b_bobufs)) != NULL) {
5400 		nbp = getdirtybuf(nbp, VI_MTX(vp), MNT_WAIT);
5401 		if (nbp)
5402 			break;
5403 	}
5404 	VI_UNLOCK(vp);
5405 	bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5406 	bawrite(bp);
5407 	if (nbp != NULL) {
5408 		bp = nbp;
5409 		goto loop;
5410 	}
5411 	/*
5412 	 * The brief unlock is to allow any pent up dependency
5413 	 * processing to be done. Then proceed with the second pass.
5414 	 */
5415 	if (waitfor == MNT_NOWAIT) {
5416 		waitfor = MNT_WAIT;
5417 		goto top;
5418 	}
5419 
5420 	/*
5421 	 * If we have managed to get rid of all the dirty buffers,
5422 	 * then we are done. For certain directories and block
5423 	 * devices, we may need to do further work.
5424 	 *
5425 	 * We must wait for any I/O in progress to finish so that
5426 	 * all potential buffers on the dirty list will be visible.
5427 	 */
5428 	VI_LOCK(vp);
5429 	drain_output(vp);
5430 	VI_UNLOCK(vp);
5431 	return (0);
5432 }
5433 
5434 /*
5435  * Flush the dependencies associated with an inodedep.
5436  * Called with splbio blocked.
5437  */
5438 static int
5439 flush_inodedep_deps(mp, ino)
5440 	struct mount *mp;
5441 	ino_t ino;
5442 {
5443 	struct inodedep *inodedep;
5444 	int error, waitfor;
5445 
5446 	/*
5447 	 * This work is done in two passes. The first pass grabs most
5448 	 * of the buffers and begins asynchronously writing them. The
5449 	 * only way to wait for these asynchronous writes is to sleep
5450 	 * on the filesystem vnode which may stay busy for a long time
5451 	 * if the filesystem is active. So, instead, we make a second
5452 	 * pass over the dependencies blocking on each write. In the
5453 	 * usual case we will be blocking against a write that we
5454 	 * initiated, so when it is done the dependency will have been
5455 	 * resolved. Thus the second pass is expected to end quickly.
5456 	 * We give a brief window at the top of the loop to allow
5457 	 * any pending I/O to complete.
5458 	 */
5459 	for (error = 0, waitfor = MNT_NOWAIT; ; ) {
5460 		if (error)
5461 			return (error);
5462 		FREE_LOCK(&lk);
5463 		ACQUIRE_LOCK(&lk);
5464 		if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
5465 			return (0);
5466 		if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
5467 		    flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
5468 		    flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
5469 		    flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
5470 			continue;
5471 		/*
5472 		 * If pass2, we are done, otherwise do pass 2.
5473 		 */
5474 		if (waitfor == MNT_WAIT)
5475 			break;
5476 		waitfor = MNT_WAIT;
5477 	}
5478 	/*
5479 	 * Try freeing inodedep in case all dependencies have been removed.
5480 	 */
5481 	if (inodedep_lookup(mp, ino, 0, &inodedep) != 0)
5482 		(void) free_inodedep(inodedep);
5483 	return (0);
5484 }
5485 
5486 /*
5487  * Flush an inode dependency list.
5488  * Called with splbio blocked.
5489  */
5490 static int
5491 flush_deplist(listhead, waitfor, errorp)
5492 	struct allocdirectlst *listhead;
5493 	int waitfor;
5494 	int *errorp;
5495 {
5496 	struct allocdirect *adp;
5497 	struct buf *bp;
5498 
5499 	mtx_assert(&lk, MA_OWNED);
5500 	TAILQ_FOREACH(adp, listhead, ad_next) {
5501 		if (adp->ad_state & DEPCOMPLETE)
5502 			continue;
5503 		bp = adp->ad_buf;
5504 		bp = getdirtybuf(bp, &lk, waitfor);
5505 		if (bp == NULL) {
5506 			if (waitfor == MNT_NOWAIT)
5507 				continue;
5508 			return (1);
5509 		}
5510 		FREE_LOCK(&lk);
5511 		if (waitfor == MNT_NOWAIT) {
5512 			bawrite(bp);
5513 		} else if ((*errorp = bwrite(bp)) != 0) {
5514 			ACQUIRE_LOCK(&lk);
5515 			return (1);
5516 		}
5517 		ACQUIRE_LOCK(&lk);
5518 		return (1);
5519 	}
5520 	return (0);
5521 }
5522 
5523 /*
5524  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
5525  * Called with splbio blocked.
5526  */
5527 static int
5528 flush_pagedep_deps(pvp, mp, diraddhdp)
5529 	struct vnode *pvp;
5530 	struct mount *mp;
5531 	struct diraddhd *diraddhdp;
5532 {
5533 	struct inodedep *inodedep;
5534 	struct ufsmount *ump;
5535 	struct diradd *dap;
5536 	struct vnode *vp;
5537 	int error = 0;
5538 	struct buf *bp;
5539 	ino_t inum;
5540 	struct worklist *wk;
5541 
5542 	ump = VFSTOUFS(mp);
5543 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
5544 		/*
5545 		 * Flush ourselves if this directory entry
5546 		 * has a MKDIR_PARENT dependency.
5547 		 */
5548 		if (dap->da_state & MKDIR_PARENT) {
5549 			FREE_LOCK(&lk);
5550 			if ((error = ffs_update(pvp, 1)) != 0)
5551 				break;
5552 			ACQUIRE_LOCK(&lk);
5553 			/*
5554 			 * If that cleared dependencies, go on to next.
5555 			 */
5556 			if (dap != LIST_FIRST(diraddhdp))
5557 				continue;
5558 			if (dap->da_state & MKDIR_PARENT)
5559 				panic("flush_pagedep_deps: MKDIR_PARENT");
5560 		}
5561 		/*
5562 		 * A newly allocated directory must have its "." and
5563 		 * ".." entries written out before its name can be
5564 		 * committed in its parent. We do not want or need
5565 		 * the full semantics of a synchronous ffs_syncvnode as
5566 		 * that may end up here again, once for each directory
5567 		 * level in the filesystem. Instead, we push the blocks
5568 		 * and wait for them to clear. We have to fsync twice
5569 		 * because the first call may choose to defer blocks
5570 		 * that still have dependencies, but deferral will
5571 		 * happen at most once.
5572 		 */
5573 		inum = dap->da_newinum;
5574 		if (dap->da_state & MKDIR_BODY) {
5575 			FREE_LOCK(&lk);
5576 			if ((error = ffs_vget(mp, inum, LK_EXCLUSIVE, &vp)))
5577 				break;
5578 			if ((error=ffs_syncvnode(vp, MNT_NOWAIT)) ||
5579 			    (error=ffs_syncvnode(vp, MNT_NOWAIT))) {
5580 				vput(vp);
5581 				break;
5582 			}
5583 			VI_LOCK(vp);
5584 			drain_output(vp);
5585 			/*
5586 			 * If first block is still dirty with a D_MKDIR
5587 			 * dependency then it needs to be written now.
5588 			 */
5589 			for (;;) {
5590 				error = 0;
5591 				bp = gbincore(&vp->v_bufobj, 0);
5592 				if (bp == NULL)
5593 					break;	/* First block not present */
5594 				error = BUF_LOCK(bp,
5595 						 LK_EXCLUSIVE |
5596 						 LK_SLEEPFAIL |
5597 						 LK_INTERLOCK,
5598 						 VI_MTX(vp));
5599 				VI_LOCK(vp);
5600 				if (error == ENOLCK)
5601 					continue;	/* Slept, retry */
5602 				if (error != 0)
5603 					break;		/* Failed */
5604 				if ((bp->b_flags & B_DELWRI) == 0) {
5605 					BUF_UNLOCK(bp);
5606 					break;	/* Buffer not dirty */
5607 				}
5608 				for (wk = LIST_FIRST(&bp->b_dep);
5609 				     wk != NULL;
5610 				     wk = LIST_NEXT(wk, wk_list))
5611 					if (wk->wk_type == D_MKDIR)
5612 						break;
5613 				if (wk == NULL)
5614 					BUF_UNLOCK(bp);	/* Dependency gone */
5615 				else {
5616 					/*
5617 					 * D_MKDIR dependency remains,
5618 					 * must write buffer to stable
5619 					 * storage.
5620 					 */
5621 					VI_UNLOCK(vp);
5622 					bremfree(bp);
5623 					error = bwrite(bp);
5624 					VI_LOCK(vp);
5625 				}
5626 				break;
5627 			}
5628 			VI_UNLOCK(vp);
5629 			vput(vp);
5630 			if (error != 0)
5631 				break;	/* Flushing of first block failed */
5632 			ACQUIRE_LOCK(&lk);
5633 			/*
5634 			 * If that cleared dependencies, go on to next.
5635 			 */
5636 			if (dap != LIST_FIRST(diraddhdp))
5637 				continue;
5638 			if (dap->da_state & MKDIR_BODY)
5639 				panic("flush_pagedep_deps: MKDIR_BODY");
5640 		}
5641 		/*
5642 		 * Flush the inode on which the directory entry depends.
5643 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
5644 		 * the only remaining dependency is that the updated inode
5645 		 * count must get pushed to disk. The inode has already
5646 		 * been pushed into its inode buffer (via VOP_UPDATE) at
5647 		 * the time of the reference count change. So we need only
5648 		 * locate that buffer, ensure that there will be no rollback
5649 		 * caused by a bitmap dependency, then write the inode buffer.
5650 		 */
5651 retry:
5652 		if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
5653 			panic("flush_pagedep_deps: lost inode");
5654 		/*
5655 		 * If the inode still has bitmap dependencies,
5656 		 * push them to disk.
5657 		 */
5658 		if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5659 			bp = inodedep->id_buf;
5660 			bp = getdirtybuf(bp, &lk, MNT_WAIT);
5661 			if (bp == NULL)
5662 				goto retry;
5663 			FREE_LOCK(&lk);
5664 			if ((error = bwrite(bp)) != 0)
5665 				break;
5666 			ACQUIRE_LOCK(&lk);
5667 			if (dap != LIST_FIRST(diraddhdp))
5668 				continue;
5669 		}
5670 		/*
5671 		 * If the inode is still sitting in a buffer waiting
5672 		 * to be written, push it to disk.
5673 		 */
5674 		FREE_LOCK(&lk);
5675 		if ((error = bread(ump->um_devvp,
5676 		    fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
5677 		    (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) {
5678 			brelse(bp);
5679 			break;
5680 		}
5681 		if ((error = bwrite(bp)) != 0)
5682 			break;
5683 		ACQUIRE_LOCK(&lk);
5684 		/*
5685 		 * If we have failed to get rid of all the dependencies
5686 		 * then something is seriously wrong.
5687 		 */
5688 		if (dap == LIST_FIRST(diraddhdp))
5689 			panic("flush_pagedep_deps: flush failed");
5690 	}
5691 	if (error)
5692 		ACQUIRE_LOCK(&lk);
5693 	return (error);
5694 }
5695 
5696 /*
5697  * A large burst of file addition or deletion activity can drive the
5698  * memory load excessively high. First attempt to slow things down
5699  * using the techniques below. If that fails, this routine requests
5700  * the offending operations to fall back to running synchronously
5701  * until the memory load returns to a reasonable level.
5702  */
5703 int
5704 softdep_slowdown(vp)
5705 	struct vnode *vp;
5706 {
5707 	int max_softdeps_hard;
5708 
5709 	ACQUIRE_LOCK(&lk);
5710 	max_softdeps_hard = max_softdeps * 11 / 10;
5711 	if (num_dirrem < max_softdeps_hard / 2 &&
5712 	    num_inodedep < max_softdeps_hard &&
5713 	    VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps) {
5714 		FREE_LOCK(&lk);
5715   		return (0);
5716 	}
5717 	if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps)
5718 		softdep_speedup();
5719 	stat_sync_limit_hit += 1;
5720 	FREE_LOCK(&lk);
5721 	return (1);
5722 }
5723 
5724 /*
5725  * Called by the allocation routines when they are about to fail
5726  * in the hope that we can free up some disk space.
5727  *
5728  * First check to see if the work list has anything on it. If it has,
5729  * clean up entries until we successfully free some space. Because this
5730  * process holds inodes locked, we cannot handle any remove requests
5731  * that might block on a locked inode as that could lead to deadlock.
5732  * If the worklist yields no free space, encourage the syncer daemon
5733  * to help us. In no event will we try for longer than tickdelay seconds.
5734  */
5735 int
5736 softdep_request_cleanup(fs, vp)
5737 	struct fs *fs;
5738 	struct vnode *vp;
5739 {
5740 	struct ufsmount *ump;
5741 	long starttime;
5742 	ufs2_daddr_t needed;
5743 	int error;
5744 
5745 	ump = VTOI(vp)->i_ump;
5746 	mtx_assert(UFS_MTX(ump), MA_OWNED);
5747 	needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize;
5748 	starttime = time_second + tickdelay;
5749 	/*
5750 	 * If we are being called because of a process doing a
5751 	 * copy-on-write, then it is not safe to update the vnode
5752 	 * as we may recurse into the copy-on-write routine.
5753 	 */
5754 	if (!(curthread->td_pflags & TDP_COWINPROGRESS)) {
5755 		UFS_UNLOCK(ump);
5756 		error = ffs_update(vp, 1);
5757 		UFS_LOCK(ump);
5758 		if (error != 0)
5759 			return (0);
5760 	}
5761 	while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) {
5762 		if (time_second > starttime)
5763 			return (0);
5764 		UFS_UNLOCK(ump);
5765 		ACQUIRE_LOCK(&lk);
5766 		if (ump->softdep_on_worklist > 0 &&
5767 		    process_worklist_item(UFSTOVFS(ump), LK_NOWAIT) != -1) {
5768 			stat_worklist_push += 1;
5769 			FREE_LOCK(&lk);
5770 			UFS_LOCK(ump);
5771 			continue;
5772 		}
5773 		request_cleanup(UFSTOVFS(ump), FLUSH_REMOVE_WAIT);
5774 		FREE_LOCK(&lk);
5775 		UFS_LOCK(ump);
5776 	}
5777 	return (1);
5778 }
5779 
5780 /*
5781  * If memory utilization has gotten too high, deliberately slow things
5782  * down and speed up the I/O processing.
5783  */
5784 extern struct thread *syncertd;
5785 static int
5786 request_cleanup(mp, resource)
5787 	struct mount *mp;
5788 	int resource;
5789 {
5790 	struct thread *td = curthread;
5791 	struct ufsmount *ump;
5792 
5793 	mtx_assert(&lk, MA_OWNED);
5794 	/*
5795 	 * We never hold up the filesystem syncer or buf daemon.
5796 	 */
5797 	if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF))
5798 		return (0);
5799 	ump = VFSTOUFS(mp);
5800 	/*
5801 	 * First check to see if the work list has gotten backlogged.
5802 	 * If it has, co-opt this process to help clean up two entries.
5803 	 * Because this process may hold inodes locked, we cannot
5804 	 * handle any remove requests that might block on a locked
5805 	 * inode as that could lead to deadlock.  We set TDP_SOFTDEP
5806 	 * to avoid recursively processing the worklist.
5807 	 */
5808 	if (ump->softdep_on_worklist > max_softdeps / 10) {
5809 		td->td_pflags |= TDP_SOFTDEP;
5810 		process_worklist_item(mp, LK_NOWAIT);
5811 		process_worklist_item(mp, LK_NOWAIT);
5812 		td->td_pflags &= ~TDP_SOFTDEP;
5813 		stat_worklist_push += 2;
5814 		return(1);
5815 	}
5816 	/*
5817 	 * Next, we attempt to speed up the syncer process. If that
5818 	 * is successful, then we allow the process to continue.
5819 	 */
5820 	if (softdep_speedup() && resource != FLUSH_REMOVE_WAIT)
5821 		return(0);
5822 	/*
5823 	 * If we are resource constrained on inode dependencies, try
5824 	 * flushing some dirty inodes. Otherwise, we are constrained
5825 	 * by file deletions, so try accelerating flushes of directories
5826 	 * with removal dependencies. We would like to do the cleanup
5827 	 * here, but we probably hold an inode locked at this point and
5828 	 * that might deadlock against one that we try to clean. So,
5829 	 * the best that we can do is request the syncer daemon to do
5830 	 * the cleanup for us.
5831 	 */
5832 	switch (resource) {
5833 
5834 	case FLUSH_INODES:
5835 		stat_ino_limit_push += 1;
5836 		req_clear_inodedeps += 1;
5837 		stat_countp = &stat_ino_limit_hit;
5838 		break;
5839 
5840 	case FLUSH_REMOVE:
5841 	case FLUSH_REMOVE_WAIT:
5842 		stat_blk_limit_push += 1;
5843 		req_clear_remove += 1;
5844 		stat_countp = &stat_blk_limit_hit;
5845 		break;
5846 
5847 	default:
5848 		panic("request_cleanup: unknown type");
5849 	}
5850 	/*
5851 	 * Hopefully the syncer daemon will catch up and awaken us.
5852 	 * We wait at most tickdelay before proceeding in any case.
5853 	 */
5854 	proc_waiting += 1;
5855 	if (handle.callout == NULL)
5856 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5857 	msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0);
5858 	proc_waiting -= 1;
5859 	return (1);
5860 }
5861 
5862 /*
5863  * Awaken processes pausing in request_cleanup and clear proc_waiting
5864  * to indicate that there is no longer a timer running.
5865  */
5866 static void
5867 pause_timer(arg)
5868 	void *arg;
5869 {
5870 
5871 	ACQUIRE_LOCK(&lk);
5872 	*stat_countp += 1;
5873 	wakeup_one(&proc_waiting);
5874 	if (proc_waiting > 0)
5875 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5876 	else
5877 		handle.callout = NULL;
5878 	FREE_LOCK(&lk);
5879 }
5880 
5881 /*
5882  * Flush out a directory with at least one removal dependency in an effort to
5883  * reduce the number of dirrem, freefile, and freeblks dependency structures.
5884  */
5885 static void
5886 clear_remove(td)
5887 	struct thread *td;
5888 {
5889 	struct pagedep_hashhead *pagedephd;
5890 	struct pagedep *pagedep;
5891 	static int next = 0;
5892 	struct mount *mp;
5893 	struct vnode *vp;
5894 	int error, cnt;
5895 	ino_t ino;
5896 
5897 	mtx_assert(&lk, MA_OWNED);
5898 
5899 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
5900 		pagedephd = &pagedep_hashtbl[next++];
5901 		if (next >= pagedep_hash)
5902 			next = 0;
5903 		LIST_FOREACH(pagedep, pagedephd, pd_hash) {
5904 			if (LIST_EMPTY(&pagedep->pd_dirremhd))
5905 				continue;
5906 			mp = pagedep->pd_list.wk_mp;
5907 			ino = pagedep->pd_ino;
5908 			if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5909 				continue;
5910 			FREE_LOCK(&lk);
5911 			if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp))) {
5912 				softdep_error("clear_remove: vget", error);
5913 				vn_finished_write(mp);
5914 				ACQUIRE_LOCK(&lk);
5915 				return;
5916 			}
5917 			if ((error = ffs_syncvnode(vp, MNT_NOWAIT)))
5918 				softdep_error("clear_remove: fsync", error);
5919 			VI_LOCK(vp);
5920 			drain_output(vp);
5921 			VI_UNLOCK(vp);
5922 			vput(vp);
5923 			vn_finished_write(mp);
5924 			ACQUIRE_LOCK(&lk);
5925 			return;
5926 		}
5927 	}
5928 }
5929 
5930 /*
5931  * Clear out a block of dirty inodes in an effort to reduce
5932  * the number of inodedep dependency structures.
5933  */
5934 static void
5935 clear_inodedeps(td)
5936 	struct thread *td;
5937 {
5938 	struct inodedep_hashhead *inodedephd;
5939 	struct inodedep *inodedep;
5940 	static int next = 0;
5941 	struct mount *mp;
5942 	struct vnode *vp;
5943 	struct fs *fs;
5944 	int error, cnt;
5945 	ino_t firstino, lastino, ino;
5946 
5947 	mtx_assert(&lk, MA_OWNED);
5948 	/*
5949 	 * Pick a random inode dependency to be cleared.
5950 	 * We will then gather up all the inodes in its block
5951 	 * that have dependencies and flush them out.
5952 	 */
5953 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
5954 		inodedephd = &inodedep_hashtbl[next++];
5955 		if (next >= inodedep_hash)
5956 			next = 0;
5957 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
5958 			break;
5959 	}
5960 	if (inodedep == NULL)
5961 		return;
5962 	fs = inodedep->id_fs;
5963 	mp = inodedep->id_list.wk_mp;
5964 	/*
5965 	 * Find the last inode in the block with dependencies.
5966 	 */
5967 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
5968 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
5969 		if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0)
5970 			break;
5971 	/*
5972 	 * Asynchronously push all but the last inode with dependencies.
5973 	 * Synchronously push the last inode with dependencies to ensure
5974 	 * that the inode block gets written to free up the inodedeps.
5975 	 */
5976 	for (ino = firstino; ino <= lastino; ino++) {
5977 		if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
5978 			continue;
5979 		if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5980 			continue;
5981 		FREE_LOCK(&lk);
5982 		if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp)) != 0) {
5983 			softdep_error("clear_inodedeps: vget", error);
5984 			vn_finished_write(mp);
5985 			ACQUIRE_LOCK(&lk);
5986 			return;
5987 		}
5988 		if (ino == lastino) {
5989 			if ((error = ffs_syncvnode(vp, MNT_WAIT)))
5990 				softdep_error("clear_inodedeps: fsync1", error);
5991 		} else {
5992 			if ((error = ffs_syncvnode(vp, MNT_NOWAIT)))
5993 				softdep_error("clear_inodedeps: fsync2", error);
5994 			VI_LOCK(vp);
5995 			drain_output(vp);
5996 			VI_UNLOCK(vp);
5997 		}
5998 		vput(vp);
5999 		vn_finished_write(mp);
6000 		ACQUIRE_LOCK(&lk);
6001 	}
6002 }
6003 
6004 /*
6005  * Function to determine if the buffer has outstanding dependencies
6006  * that will cause a roll-back if the buffer is written. If wantcount
6007  * is set, return number of dependencies, otherwise just yes or no.
6008  */
6009 static int
6010 softdep_count_dependencies(bp, wantcount)
6011 	struct buf *bp;
6012 	int wantcount;
6013 {
6014 	struct worklist *wk;
6015 	struct inodedep *inodedep;
6016 	struct indirdep *indirdep;
6017 	struct allocindir *aip;
6018 	struct pagedep *pagedep;
6019 	struct diradd *dap;
6020 	int i, retval;
6021 
6022 	retval = 0;
6023 	ACQUIRE_LOCK(&lk);
6024 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
6025 		switch (wk->wk_type) {
6026 
6027 		case D_INODEDEP:
6028 			inodedep = WK_INODEDEP(wk);
6029 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
6030 				/* bitmap allocation dependency */
6031 				retval += 1;
6032 				if (!wantcount)
6033 					goto out;
6034 			}
6035 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
6036 				/* direct block pointer dependency */
6037 				retval += 1;
6038 				if (!wantcount)
6039 					goto out;
6040 			}
6041 			if (TAILQ_FIRST(&inodedep->id_extupdt)) {
6042 				/* direct block pointer dependency */
6043 				retval += 1;
6044 				if (!wantcount)
6045 					goto out;
6046 			}
6047 			continue;
6048 
6049 		case D_INDIRDEP:
6050 			indirdep = WK_INDIRDEP(wk);
6051 
6052 			LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
6053 				/* indirect block pointer dependency */
6054 				retval += 1;
6055 				if (!wantcount)
6056 					goto out;
6057 			}
6058 			continue;
6059 
6060 		case D_PAGEDEP:
6061 			pagedep = WK_PAGEDEP(wk);
6062 			for (i = 0; i < DAHASHSZ; i++) {
6063 
6064 				LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
6065 					/* directory entry dependency */
6066 					retval += 1;
6067 					if (!wantcount)
6068 						goto out;
6069 				}
6070 			}
6071 			continue;
6072 
6073 		case D_BMSAFEMAP:
6074 		case D_ALLOCDIRECT:
6075 		case D_ALLOCINDIR:
6076 		case D_MKDIR:
6077 			/* never a dependency on these blocks */
6078 			continue;
6079 
6080 		default:
6081 			panic("softdep_check_for_rollback: Unexpected type %s",
6082 			    TYPENAME(wk->wk_type));
6083 			/* NOTREACHED */
6084 		}
6085 	}
6086 out:
6087 	FREE_LOCK(&lk);
6088 	return retval;
6089 }
6090 
6091 /*
6092  * Acquire exclusive access to a buffer.
6093  * Must be called with a locked mtx parameter.
6094  * Return acquired buffer or NULL on failure.
6095  */
6096 static struct buf *
6097 getdirtybuf(bp, mtx, waitfor)
6098 	struct buf *bp;
6099 	struct mtx *mtx;
6100 	int waitfor;
6101 {
6102 	int error;
6103 
6104 	mtx_assert(mtx, MA_OWNED);
6105 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) {
6106 		if (waitfor != MNT_WAIT)
6107 			return (NULL);
6108 		error = BUF_LOCK(bp,
6109 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx);
6110 		/*
6111 		 * Even if we sucessfully acquire bp here, we have dropped
6112 		 * mtx, which may violates our guarantee.
6113 		 */
6114 		if (error == 0)
6115 			BUF_UNLOCK(bp);
6116 		else if (error != ENOLCK)
6117 			panic("getdirtybuf: inconsistent lock: %d", error);
6118 		mtx_lock(mtx);
6119 		return (NULL);
6120 	}
6121 	if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
6122 		if (mtx == &lk && waitfor == MNT_WAIT) {
6123 			mtx_unlock(mtx);
6124 			BO_LOCK(bp->b_bufobj);
6125 			BUF_UNLOCK(bp);
6126 			if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
6127 				bp->b_vflags |= BV_BKGRDWAIT;
6128 				msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj),
6129 				       PRIBIO | PDROP, "getbuf", 0);
6130 			} else
6131 				BO_UNLOCK(bp->b_bufobj);
6132 			mtx_lock(mtx);
6133 			return (NULL);
6134 		}
6135 		BUF_UNLOCK(bp);
6136 		if (waitfor != MNT_WAIT)
6137 			return (NULL);
6138 		/*
6139 		 * The mtx argument must be bp->b_vp's mutex in
6140 		 * this case.
6141 		 */
6142 #ifdef	DEBUG_VFS_LOCKS
6143 		if (bp->b_vp->v_type != VCHR)
6144 			ASSERT_VI_LOCKED(bp->b_vp, "getdirtybuf");
6145 #endif
6146 		bp->b_vflags |= BV_BKGRDWAIT;
6147 		msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0);
6148 		return (NULL);
6149 	}
6150 	if ((bp->b_flags & B_DELWRI) == 0) {
6151 		BUF_UNLOCK(bp);
6152 		return (NULL);
6153 	}
6154 	bremfree(bp);
6155 	return (bp);
6156 }
6157 
6158 
6159 /*
6160  * Check if it is safe to suspend the file system now.  On entry,
6161  * the vnode interlock for devvp should be held.  Return 0 with
6162  * the mount interlock held if the file system can be suspended now,
6163  * otherwise return EAGAIN with the mount interlock held.
6164  */
6165 int
6166 softdep_check_suspend(struct mount *mp,
6167 		      struct vnode *devvp,
6168 		      int softdep_deps,
6169 		      int softdep_accdeps,
6170 		      int secondary_writes,
6171 		      int secondary_accwrites)
6172 {
6173 	struct bufobj *bo;
6174 	struct ufsmount *ump;
6175 	int error;
6176 
6177 	ASSERT_VI_LOCKED(devvp, "softdep_check_suspend");
6178 	ump = VFSTOUFS(mp);
6179 	bo = &devvp->v_bufobj;
6180 
6181 	for (;;) {
6182 		if (!TRY_ACQUIRE_LOCK(&lk)) {
6183 			VI_UNLOCK(devvp);
6184 			ACQUIRE_LOCK(&lk);
6185 			FREE_LOCK(&lk);
6186 			VI_LOCK(devvp);
6187 			continue;
6188 		}
6189 		if (!MNT_ITRYLOCK(mp)) {
6190 			FREE_LOCK(&lk);
6191 			VI_UNLOCK(devvp);
6192 			MNT_ILOCK(mp);
6193 			MNT_IUNLOCK(mp);
6194 			VI_LOCK(devvp);
6195 			continue;
6196 		}
6197 		if (mp->mnt_secondary_writes != 0) {
6198 			FREE_LOCK(&lk);
6199 			VI_UNLOCK(devvp);
6200 			msleep(&mp->mnt_secondary_writes,
6201 			       MNT_MTX(mp),
6202 			       (PUSER - 1) | PDROP, "secwr", 0);
6203 			VI_LOCK(devvp);
6204 			continue;
6205 		}
6206 		break;
6207 	}
6208 
6209 	/*
6210 	 * Reasons for needing more work before suspend:
6211 	 * - Dirty buffers on devvp.
6212 	 * - Softdep activity occurred after start of vnode sync loop
6213 	 * - Secondary writes occurred after start of vnode sync loop
6214 	 */
6215 	error = 0;
6216 	if (bo->bo_numoutput > 0 ||
6217 	    bo->bo_dirty.bv_cnt > 0 ||
6218 	    softdep_deps != 0 ||
6219 	    ump->softdep_deps != 0 ||
6220 	    softdep_accdeps != ump->softdep_accdeps ||
6221 	    secondary_writes != 0 ||
6222 	    mp->mnt_secondary_writes != 0 ||
6223 	    secondary_accwrites != mp->mnt_secondary_accwrites)
6224 		error = EAGAIN;
6225 	FREE_LOCK(&lk);
6226 	VI_UNLOCK(devvp);
6227 	return (error);
6228 }
6229 
6230 
6231 /*
6232  * Get the number of dependency structures for the file system, both
6233  * the current number and the total number allocated.  These will
6234  * later be used to detect that softdep processing has occurred.
6235  */
6236 void
6237 softdep_get_depcounts(struct mount *mp,
6238 		      int *softdep_depsp,
6239 		      int *softdep_accdepsp)
6240 {
6241 	struct ufsmount *ump;
6242 
6243 	ump = VFSTOUFS(mp);
6244 	ACQUIRE_LOCK(&lk);
6245 	*softdep_depsp = ump->softdep_deps;
6246 	*softdep_accdepsp = ump->softdep_accdeps;
6247 	FREE_LOCK(&lk);
6248 }
6249 
6250 /*
6251  * Wait for pending output on a vnode to complete.
6252  * Must be called with vnode lock and interlock locked.
6253  *
6254  * XXX: Should just be a call to bufobj_wwait().
6255  */
6256 static void
6257 drain_output(vp)
6258 	struct vnode *vp;
6259 {
6260 	ASSERT_VOP_LOCKED(vp, "drain_output");
6261 	ASSERT_VI_LOCKED(vp, "drain_output");
6262 
6263 	while (vp->v_bufobj.bo_numoutput) {
6264 		vp->v_bufobj.bo_flag |= BO_WWAIT;
6265 		msleep((caddr_t)&vp->v_bufobj.bo_numoutput,
6266 		    VI_MTX(vp), PRIBIO + 1, "drainvp", 0);
6267 	}
6268 }
6269 
6270 /*
6271  * Called whenever a buffer that is being invalidated or reallocated
6272  * contains dependencies. This should only happen if an I/O error has
6273  * occurred. The routine is called with the buffer locked.
6274  */
6275 static void
6276 softdep_deallocate_dependencies(bp)
6277 	struct buf *bp;
6278 {
6279 
6280 	if ((bp->b_ioflags & BIO_ERROR) == 0)
6281 		panic("softdep_deallocate_dependencies: dangling deps");
6282 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
6283 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
6284 }
6285 
6286 /*
6287  * Function to handle asynchronous write errors in the filesystem.
6288  */
6289 static void
6290 softdep_error(func, error)
6291 	char *func;
6292 	int error;
6293 {
6294 
6295 	/* XXX should do something better! */
6296 	printf("%s: got error %d while accessing filesystem\n", func, error);
6297 }
6298 
6299 #endif /* SOFTUPDATES */
6300