xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision 00a5db46de56179184c0f000eaacad695e2b0859)
1 /*-
2  * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
3  *
4  * The soft updates code is derived from the appendix of a University
5  * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
6  * "Soft Updates: A Solution to the Metadata Update Problem in File
7  * Systems", CSE-TR-254-95, August 1995).
8  *
9  * Further information about soft updates can be obtained from:
10  *
11  *	Marshall Kirk McKusick		http://www.mckusick.com/softdep/
12  *	1614 Oxford Street		mckusick@mckusick.com
13  *	Berkeley, CA 94709-1608		+1-510-843-9542
14  *	USA
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  *
20  * 1. Redistributions of source code must retain the above copyright
21  *    notice, this list of conditions and the following disclaimer.
22  * 2. Redistributions in binary form must reproduce the above copyright
23  *    notice, this list of conditions and the following disclaimer in the
24  *    documentation and/or other materials provided with the distribution.
25  *
26  * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
27  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
28  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
29  * DISCLAIMED.  IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
30  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)ffs_softdep.c	9.59 (McKusick) 6/21/00
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_ffs.h"
45 #include "opt_ddb.h"
46 
47 /*
48  * For now we want the safety net that the DEBUG flag provides.
49  */
50 #ifndef DEBUG
51 #define DEBUG
52 #endif
53 
54 #include <sys/param.h>
55 #include <sys/kernel.h>
56 #include <sys/systm.h>
57 #include <sys/bio.h>
58 #include <sys/buf.h>
59 #include <sys/kdb.h>
60 #include <sys/kthread.h>
61 #include <sys/lock.h>
62 #include <sys/malloc.h>
63 #include <sys/mount.h>
64 #include <sys/mutex.h>
65 #include <sys/proc.h>
66 #include <sys/stat.h>
67 #include <sys/sysctl.h>
68 #include <sys/syslog.h>
69 #include <sys/vnode.h>
70 #include <sys/conf.h>
71 #include <ufs/ufs/dir.h>
72 #include <ufs/ufs/extattr.h>
73 #include <ufs/ufs/quota.h>
74 #include <ufs/ufs/inode.h>
75 #include <ufs/ufs/ufsmount.h>
76 #include <ufs/ffs/fs.h>
77 #include <ufs/ffs/softdep.h>
78 #include <ufs/ffs/ffs_extern.h>
79 #include <ufs/ufs/ufs_extern.h>
80 
81 #include <vm/vm.h>
82 
83 #include <ddb/ddb.h>
84 
85 #ifndef SOFTUPDATES
86 
87 int
88 softdep_flushfiles(oldmnt, flags, td)
89 	struct mount *oldmnt;
90 	int flags;
91 	struct thread *td;
92 {
93 
94 	panic("softdep_flushfiles called");
95 }
96 
97 int
98 softdep_mount(devvp, mp, fs, cred)
99 	struct vnode *devvp;
100 	struct mount *mp;
101 	struct fs *fs;
102 	struct ucred *cred;
103 {
104 
105 	return (0);
106 }
107 
108 void
109 softdep_initialize()
110 {
111 
112 	return;
113 }
114 
115 void
116 softdep_uninitialize()
117 {
118 
119 	return;
120 }
121 
122 void
123 softdep_setup_inomapdep(bp, ip, newinum)
124 	struct buf *bp;
125 	struct inode *ip;
126 	ino_t newinum;
127 {
128 
129 	panic("softdep_setup_inomapdep called");
130 }
131 
132 void
133 softdep_setup_blkmapdep(bp, mp, newblkno)
134 	struct buf *bp;
135 	struct mount *mp;
136 	ufs2_daddr_t newblkno;
137 {
138 
139 	panic("softdep_setup_blkmapdep called");
140 }
141 
142 void
143 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
144 	struct inode *ip;
145 	ufs_lbn_t lbn;
146 	ufs2_daddr_t newblkno;
147 	ufs2_daddr_t oldblkno;
148 	long newsize;
149 	long oldsize;
150 	struct buf *bp;
151 {
152 
153 	panic("softdep_setup_allocdirect called");
154 }
155 
156 void
157 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
158 	struct inode *ip;
159 	ufs_lbn_t lbn;
160 	ufs2_daddr_t newblkno;
161 	ufs2_daddr_t oldblkno;
162 	long newsize;
163 	long oldsize;
164 	struct buf *bp;
165 {
166 
167 	panic("softdep_setup_allocext called");
168 }
169 
170 void
171 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
172 	struct inode *ip;
173 	ufs_lbn_t lbn;
174 	struct buf *bp;
175 	int ptrno;
176 	ufs2_daddr_t newblkno;
177 	ufs2_daddr_t oldblkno;
178 	struct buf *nbp;
179 {
180 
181 	panic("softdep_setup_allocindir_page called");
182 }
183 
184 void
185 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
186 	struct buf *nbp;
187 	struct inode *ip;
188 	struct buf *bp;
189 	int ptrno;
190 	ufs2_daddr_t newblkno;
191 {
192 
193 	panic("softdep_setup_allocindir_meta called");
194 }
195 
196 void
197 softdep_setup_freeblocks(ip, length, flags)
198 	struct inode *ip;
199 	off_t length;
200 	int flags;
201 {
202 
203 	panic("softdep_setup_freeblocks called");
204 }
205 
206 void
207 softdep_freefile(pvp, ino, mode)
208 		struct vnode *pvp;
209 		ino_t ino;
210 		int mode;
211 {
212 
213 	panic("softdep_freefile called");
214 }
215 
216 int
217 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
218 	struct buf *bp;
219 	struct inode *dp;
220 	off_t diroffset;
221 	ino_t newinum;
222 	struct buf *newdirbp;
223 	int isnewblk;
224 {
225 
226 	panic("softdep_setup_directory_add called");
227 }
228 
229 void
230 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
231 	struct inode *dp;
232 	caddr_t base;
233 	caddr_t oldloc;
234 	caddr_t newloc;
235 	int entrysize;
236 {
237 
238 	panic("softdep_change_directoryentry_offset called");
239 }
240 
241 void
242 softdep_setup_remove(bp, dp, ip, isrmdir)
243 	struct buf *bp;
244 	struct inode *dp;
245 	struct inode *ip;
246 	int isrmdir;
247 {
248 
249 	panic("softdep_setup_remove called");
250 }
251 
252 void
253 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
254 	struct buf *bp;
255 	struct inode *dp;
256 	struct inode *ip;
257 	ino_t newinum;
258 	int isrmdir;
259 {
260 
261 	panic("softdep_setup_directory_change called");
262 }
263 
264 void
265 softdep_change_linkcnt(ip)
266 	struct inode *ip;
267 {
268 
269 	panic("softdep_change_linkcnt called");
270 }
271 
272 void
273 softdep_load_inodeblock(ip)
274 	struct inode *ip;
275 {
276 
277 	panic("softdep_load_inodeblock called");
278 }
279 
280 void
281 softdep_update_inodeblock(ip, bp, waitfor)
282 	struct inode *ip;
283 	struct buf *bp;
284 	int waitfor;
285 {
286 
287 	panic("softdep_update_inodeblock called");
288 }
289 
290 int
291 softdep_fsync(vp)
292 	struct vnode *vp;	/* the "in_core" copy of the inode */
293 {
294 
295 	return (0);
296 }
297 
298 void
299 softdep_fsync_mountdev(vp)
300 	struct vnode *vp;
301 {
302 
303 	return;
304 }
305 
306 int
307 softdep_flushworklist(oldmnt, countp, td)
308 	struct mount *oldmnt;
309 	int *countp;
310 	struct thread *td;
311 {
312 
313 	*countp = 0;
314 	return (0);
315 }
316 
317 int
318 softdep_sync_metadata(struct vnode *vp)
319 {
320 
321 	return (0);
322 }
323 
324 int
325 softdep_slowdown(vp)
326 	struct vnode *vp;
327 {
328 
329 	panic("softdep_slowdown called");
330 }
331 
332 void
333 softdep_releasefile(ip)
334 	struct inode *ip;	/* inode with the zero effective link count */
335 {
336 
337 	panic("softdep_releasefile called");
338 }
339 
340 int
341 softdep_request_cleanup(fs, vp)
342 	struct fs *fs;
343 	struct vnode *vp;
344 {
345 
346 	return (0);
347 }
348 
349 int
350 softdep_check_suspend(struct mount *mp,
351 		      struct vnode *devvp,
352 		      int softdep_deps,
353 		      int softdep_accdeps,
354 		      int secondary_writes,
355 		      int secondary_accwrites)
356 {
357 	struct bufobj *bo;
358 	int error;
359 
360 	(void) softdep_deps,
361 	(void) softdep_accdeps;
362 
363 	bo = &devvp->v_bufobj;
364 	ASSERT_BO_LOCKED(bo);
365 
366 	MNT_ILOCK(mp);
367 	while (mp->mnt_secondary_writes != 0) {
368 		BO_UNLOCK(bo);
369 		msleep(&mp->mnt_secondary_writes, MNT_MTX(mp),
370 		    (PUSER - 1) | PDROP, "secwr", 0);
371 		BO_LOCK(bo);
372 		MNT_ILOCK(mp);
373 	}
374 
375 	/*
376 	 * Reasons for needing more work before suspend:
377 	 * - Dirty buffers on devvp.
378 	 * - Secondary writes occurred after start of vnode sync loop
379 	 */
380 	error = 0;
381 	if (bo->bo_numoutput > 0 ||
382 	    bo->bo_dirty.bv_cnt > 0 ||
383 	    secondary_writes != 0 ||
384 	    mp->mnt_secondary_writes != 0 ||
385 	    secondary_accwrites != mp->mnt_secondary_accwrites)
386 		error = EAGAIN;
387 	BO_UNLOCK(bo);
388 	return (error);
389 }
390 
391 void
392 softdep_get_depcounts(struct mount *mp,
393 		      int *softdepactivep,
394 		      int *softdepactiveaccp)
395 {
396 	(void) mp;
397 	*softdepactivep = 0;
398 	*softdepactiveaccp = 0;
399 }
400 
401 #else
402 /*
403  * These definitions need to be adapted to the system to which
404  * this file is being ported.
405  */
406 /*
407  * malloc types defined for the softdep system.
408  */
409 static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
410 static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
411 static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
412 static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
413 static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
414 static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
415 static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
416 static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
417 static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
418 static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
419 static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
420 static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
421 static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
422 static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block");
423 static MALLOC_DEFINE(M_SAVEDINO, "savedino","Saved inodes");
424 
425 #define M_SOFTDEP_FLAGS	(M_WAITOK | M_USE_RESERVE)
426 
427 #define	D_PAGEDEP	0
428 #define	D_INODEDEP	1
429 #define	D_NEWBLK	2
430 #define	D_BMSAFEMAP	3
431 #define	D_ALLOCDIRECT	4
432 #define	D_INDIRDEP	5
433 #define	D_ALLOCINDIR	6
434 #define	D_FREEFRAG	7
435 #define	D_FREEBLKS	8
436 #define	D_FREEFILE	9
437 #define	D_DIRADD	10
438 #define	D_MKDIR		11
439 #define	D_DIRREM	12
440 #define	D_NEWDIRBLK	13
441 #define	D_LAST		D_NEWDIRBLK
442 
443 /*
444  * translate from workitem type to memory type
445  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
446  */
447 static struct malloc_type *memtype[] = {
448 	M_PAGEDEP,
449 	M_INODEDEP,
450 	M_NEWBLK,
451 	M_BMSAFEMAP,
452 	M_ALLOCDIRECT,
453 	M_INDIRDEP,
454 	M_ALLOCINDIR,
455 	M_FREEFRAG,
456 	M_FREEBLKS,
457 	M_FREEFILE,
458 	M_DIRADD,
459 	M_MKDIR,
460 	M_DIRREM,
461 	M_NEWDIRBLK
462 };
463 
464 #define DtoM(type) (memtype[type])
465 
466 /*
467  * Names of malloc types.
468  */
469 #define TYPENAME(type)  \
470 	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
471 /*
472  * End system adaptation definitions.
473  */
474 
475 /*
476  * Forward declarations.
477  */
478 struct inodedep_hashhead;
479 struct newblk_hashhead;
480 struct pagedep_hashhead;
481 
482 /*
483  * Internal function prototypes.
484  */
485 static	void softdep_error(char *, int);
486 static	void drain_output(struct vnode *);
487 static	struct buf *getdirtybuf(struct buf *, struct mtx *, int);
488 static	void clear_remove(struct thread *);
489 static	void clear_inodedeps(struct thread *);
490 static	int flush_pagedep_deps(struct vnode *, struct mount *,
491 	    struct diraddhd *);
492 static	int flush_inodedep_deps(struct mount *, ino_t);
493 static	int flush_deplist(struct allocdirectlst *, int, int *);
494 static	int handle_written_filepage(struct pagedep *, struct buf *);
495 static  void diradd_inode_written(struct diradd *, struct inodedep *);
496 static	int handle_written_inodeblock(struct inodedep *, struct buf *);
497 static	void handle_allocdirect_partdone(struct allocdirect *);
498 static	void handle_allocindir_partdone(struct allocindir *);
499 static	void initiate_write_filepage(struct pagedep *, struct buf *);
500 static	void handle_written_mkdir(struct mkdir *, int);
501 static	void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
502 static	void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
503 static	void handle_workitem_freefile(struct freefile *);
504 static	void handle_workitem_remove(struct dirrem *, struct vnode *);
505 static	struct dirrem *newdirrem(struct buf *, struct inode *,
506 	    struct inode *, int, struct dirrem **);
507 static	void free_diradd(struct diradd *);
508 static	void free_allocindir(struct allocindir *, struct inodedep *);
509 static	void free_newdirblk(struct newdirblk *);
510 static	int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t,
511 	    ufs2_daddr_t *);
512 static	void deallocate_dependencies(struct buf *, struct inodedep *);
513 static	void free_allocdirect(struct allocdirectlst *,
514 	    struct allocdirect *, int);
515 static	int check_inode_unwritten(struct inodedep *);
516 static	int free_inodedep(struct inodedep *);
517 static	void handle_workitem_freeblocks(struct freeblks *, int);
518 static	void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
519 static	void setup_allocindir_phase2(struct buf *, struct inode *,
520 	    struct allocindir *);
521 static	struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
522 	    ufs2_daddr_t);
523 static	void handle_workitem_freefrag(struct freefrag *);
524 static	struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long);
525 static	void allocdirect_merge(struct allocdirectlst *,
526 	    struct allocdirect *, struct allocdirect *);
527 static	struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *);
528 static	int newblk_find(struct newblk_hashhead *, struct fs *, ufs2_daddr_t,
529 	    struct newblk **);
530 static	int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **);
531 static	int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t,
532 	    struct inodedep **);
533 static	int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **);
534 static	int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **);
535 static	int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t,
536 	    struct mount *mp, int, struct pagedep **);
537 static	void pause_timer(void *);
538 static	int request_cleanup(struct mount *, int);
539 static	int process_worklist_item(struct mount *, int);
540 static	void add_to_worklist(struct worklist *);
541 static	void softdep_flush(void);
542 static	int softdep_speedup(void);
543 
544 /*
545  * Exported softdep operations.
546  */
547 static	void softdep_disk_io_initiation(struct buf *);
548 static	void softdep_disk_write_complete(struct buf *);
549 static	void softdep_deallocate_dependencies(struct buf *);
550 static	int softdep_count_dependencies(struct buf *bp, int);
551 
552 static struct mtx lk;
553 MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF);
554 
555 #define TRY_ACQUIRE_LOCK(lk)		mtx_trylock(lk)
556 #define ACQUIRE_LOCK(lk)		mtx_lock(lk)
557 #define FREE_LOCK(lk)			mtx_unlock(lk)
558 
559 #define	BUF_AREC(bp)	((bp)->b_lock.lock_object.lo_flags |= LK_CANRECURSE)
560 #define	BUF_NOREC(bp)	((bp)->b_lock.lock_object.lo_flags &= ~LK_CANRECURSE)
561 
562 /*
563  * Worklist queue management.
564  * These routines require that the lock be held.
565  */
566 #ifndef /* NOT */ DEBUG
567 #define WORKLIST_INSERT(head, item) do {	\
568 	(item)->wk_state |= ONWORKLIST;		\
569 	LIST_INSERT_HEAD(head, item, wk_list);	\
570 } while (0)
571 #define WORKLIST_REMOVE(item) do {		\
572 	(item)->wk_state &= ~ONWORKLIST;	\
573 	LIST_REMOVE(item, wk_list);		\
574 } while (0)
575 #else /* DEBUG */
576 static	void worklist_insert(struct workhead *, struct worklist *);
577 static	void worklist_remove(struct worklist *);
578 
579 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
580 #define WORKLIST_REMOVE(item) worklist_remove(item)
581 
582 static void
583 worklist_insert(head, item)
584 	struct workhead *head;
585 	struct worklist *item;
586 {
587 
588 	mtx_assert(&lk, MA_OWNED);
589 	if (item->wk_state & ONWORKLIST)
590 		panic("worklist_insert: already on list");
591 	item->wk_state |= ONWORKLIST;
592 	LIST_INSERT_HEAD(head, item, wk_list);
593 }
594 
595 static void
596 worklist_remove(item)
597 	struct worklist *item;
598 {
599 
600 	mtx_assert(&lk, MA_OWNED);
601 	if ((item->wk_state & ONWORKLIST) == 0)
602 		panic("worklist_remove: not on list");
603 	item->wk_state &= ~ONWORKLIST;
604 	LIST_REMOVE(item, wk_list);
605 }
606 #endif /* DEBUG */
607 
608 /*
609  * Routines for tracking and managing workitems.
610  */
611 static	void workitem_free(struct worklist *, int);
612 static	void workitem_alloc(struct worklist *, int, struct mount *);
613 
614 #define	WORKITEM_FREE(item, type) workitem_free((struct worklist *)(item), (type))
615 
616 static void
617 workitem_free(item, type)
618 	struct worklist *item;
619 	int type;
620 {
621 	struct ufsmount *ump;
622 	mtx_assert(&lk, MA_OWNED);
623 
624 #ifdef DEBUG
625 	if (item->wk_state & ONWORKLIST)
626 		panic("workitem_free: still on list");
627 	if (item->wk_type != type)
628 		panic("workitem_free: type mismatch");
629 #endif
630 	ump = VFSTOUFS(item->wk_mp);
631 	if (--ump->softdep_deps == 0 && ump->softdep_req)
632 		wakeup(&ump->softdep_deps);
633 	free(item, DtoM(type));
634 }
635 
636 static void
637 workitem_alloc(item, type, mp)
638 	struct worklist *item;
639 	int type;
640 	struct mount *mp;
641 {
642 	item->wk_type = type;
643 	item->wk_mp = mp;
644 	item->wk_state = 0;
645 	ACQUIRE_LOCK(&lk);
646 	VFSTOUFS(mp)->softdep_deps++;
647 	VFSTOUFS(mp)->softdep_accdeps++;
648 	FREE_LOCK(&lk);
649 }
650 
651 /*
652  * Workitem queue management
653  */
654 static int max_softdeps;	/* maximum number of structs before slowdown */
655 static int maxindirdeps = 50;	/* max number of indirdeps before slowdown */
656 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
657 static int proc_waiting;	/* tracks whether we have a timeout posted */
658 static int *stat_countp;	/* statistic to count in proc_waiting timeout */
659 static struct callout softdep_callout;
660 static int req_pending;
661 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
662 #define FLUSH_INODES		1
663 static int req_clear_remove;	/* syncer process flush some freeblks */
664 #define FLUSH_REMOVE		2
665 #define FLUSH_REMOVE_WAIT	3
666 /*
667  * runtime statistics
668  */
669 static int stat_worklist_push;	/* number of worklist cleanups */
670 static int stat_blk_limit_push;	/* number of times block limit neared */
671 static int stat_ino_limit_push;	/* number of times inode limit neared */
672 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
673 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
674 static int stat_sync_limit_hit;	/* number of synchronous slowdowns imposed */
675 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
676 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
677 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
678 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
679 
680 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
681 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
682 SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, "");
683 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
684 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
685 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
686 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
687 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
688 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
689 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
690 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
691 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
692 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
693 /* SYSCTL_INT(_debug, OID_AUTO, worklist_num, CTLFLAG_RD, &softdep_on_worklist, 0, ""); */
694 
695 SYSCTL_DECL(_vfs_ffs);
696 
697 static int compute_summary_at_mount = 0;	/* Whether to recompute the summary at mount time */
698 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW,
699 	   &compute_summary_at_mount, 0, "Recompute summary at mount");
700 
701 static struct proc *softdepproc;
702 static struct kproc_desc softdep_kp = {
703 	"softdepflush",
704 	softdep_flush,
705 	&softdepproc
706 };
707 SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start,
708     &softdep_kp);
709 
710 static void
711 softdep_flush(void)
712 {
713 	struct mount *nmp;
714 	struct mount *mp;
715 	struct ufsmount *ump;
716 	struct thread *td;
717 	int remaining;
718 	int vfslocked;
719 
720 	td = curthread;
721 	td->td_pflags |= TDP_NORUNNINGBUF;
722 
723 	for (;;) {
724 		kproc_suspend_check(softdepproc);
725 		vfslocked = VFS_LOCK_GIANT((struct mount *)NULL);
726 		ACQUIRE_LOCK(&lk);
727 		/*
728 		 * If requested, try removing inode or removal dependencies.
729 		 */
730 		if (req_clear_inodedeps) {
731 			clear_inodedeps(td);
732 			req_clear_inodedeps -= 1;
733 			wakeup_one(&proc_waiting);
734 		}
735 		if (req_clear_remove) {
736 			clear_remove(td);
737 			req_clear_remove -= 1;
738 			wakeup_one(&proc_waiting);
739 		}
740 		FREE_LOCK(&lk);
741 		VFS_UNLOCK_GIANT(vfslocked);
742 		remaining = 0;
743 		mtx_lock(&mountlist_mtx);
744 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp)  {
745 			nmp = TAILQ_NEXT(mp, mnt_list);
746 			if ((mp->mnt_flag & MNT_SOFTDEP) == 0)
747 				continue;
748 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
749 				continue;
750 			vfslocked = VFS_LOCK_GIANT(mp);
751 			softdep_process_worklist(mp, 0);
752 			ump = VFSTOUFS(mp);
753 			remaining += ump->softdep_on_worklist -
754 				ump->softdep_on_worklist_inprogress;
755 			VFS_UNLOCK_GIANT(vfslocked);
756 			mtx_lock(&mountlist_mtx);
757 			nmp = TAILQ_NEXT(mp, mnt_list);
758 			vfs_unbusy(mp);
759 		}
760 		mtx_unlock(&mountlist_mtx);
761 		if (remaining)
762 			continue;
763 		ACQUIRE_LOCK(&lk);
764 		if (!req_pending)
765 			msleep(&req_pending, &lk, PVM, "sdflush", hz);
766 		req_pending = 0;
767 		FREE_LOCK(&lk);
768 	}
769 }
770 
771 static int
772 softdep_speedup(void)
773 {
774 
775 	mtx_assert(&lk, MA_OWNED);
776 	if (req_pending == 0) {
777 		req_pending = 1;
778 		wakeup(&req_pending);
779 	}
780 
781 	return speedup_syncer();
782 }
783 
784 /*
785  * Add an item to the end of the work queue.
786  * This routine requires that the lock be held.
787  * This is the only routine that adds items to the list.
788  * The following routine is the only one that removes items
789  * and does so in order from first to last.
790  */
791 static void
792 add_to_worklist(wk)
793 	struct worklist *wk;
794 {
795 	struct ufsmount *ump;
796 
797 	mtx_assert(&lk, MA_OWNED);
798 	ump = VFSTOUFS(wk->wk_mp);
799 	if (wk->wk_state & ONWORKLIST)
800 		panic("add_to_worklist: already on list");
801 	wk->wk_state |= ONWORKLIST;
802 	if (LIST_EMPTY(&ump->softdep_workitem_pending))
803 		LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list);
804 	else
805 		LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list);
806 	ump->softdep_worklist_tail = wk;
807 	ump->softdep_on_worklist += 1;
808 }
809 
810 /*
811  * Process that runs once per second to handle items in the background queue.
812  *
813  * Note that we ensure that everything is done in the order in which they
814  * appear in the queue. The code below depends on this property to ensure
815  * that blocks of a file are freed before the inode itself is freed. This
816  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
817  * until all the old ones have been purged from the dependency lists.
818  */
819 int
820 softdep_process_worklist(mp, full)
821 	struct mount *mp;
822 	int full;
823 {
824 	struct thread *td = curthread;
825 	int cnt, matchcnt, loopcount;
826 	struct ufsmount *ump;
827 	long starttime;
828 
829 	KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp"));
830 	/*
831 	 * Record the process identifier of our caller so that we can give
832 	 * this process preferential treatment in request_cleanup below.
833 	 */
834 	matchcnt = 0;
835 	ump = VFSTOUFS(mp);
836 	ACQUIRE_LOCK(&lk);
837 	loopcount = 1;
838 	starttime = time_second;
839 	while (ump->softdep_on_worklist > 0) {
840 		if ((cnt = process_worklist_item(mp, 0)) == -1)
841 			break;
842 		else
843 			matchcnt += cnt;
844 		/*
845 		 * If requested, try removing inode or removal dependencies.
846 		 */
847 		if (req_clear_inodedeps) {
848 			clear_inodedeps(td);
849 			req_clear_inodedeps -= 1;
850 			wakeup_one(&proc_waiting);
851 		}
852 		if (req_clear_remove) {
853 			clear_remove(td);
854 			req_clear_remove -= 1;
855 			wakeup_one(&proc_waiting);
856 		}
857 		/*
858 		 * We do not generally want to stop for buffer space, but if
859 		 * we are really being a buffer hog, we will stop and wait.
860 		 */
861 		if (loopcount++ % 128 == 0) {
862 			FREE_LOCK(&lk);
863 			uio_yield();
864 			bwillwrite();
865 			ACQUIRE_LOCK(&lk);
866 		}
867 		/*
868 		 * Never allow processing to run for more than one
869 		 * second. Otherwise the other mountpoints may get
870 		 * excessively backlogged.
871 		 */
872 		if (!full && starttime != time_second) {
873 			matchcnt = -1;
874 			break;
875 		}
876 	}
877 	FREE_LOCK(&lk);
878 	return (matchcnt);
879 }
880 
881 /*
882  * Process one item on the worklist.
883  */
884 static int
885 process_worklist_item(mp, flags)
886 	struct mount *mp;
887 	int flags;
888 {
889 	struct worklist *wk, *wkend;
890 	struct ufsmount *ump;
891 	struct vnode *vp;
892 	int matchcnt = 0;
893 
894 	mtx_assert(&lk, MA_OWNED);
895 	KASSERT(mp != NULL, ("process_worklist_item: NULL mp"));
896 	/*
897 	 * If we are being called because of a process doing a
898 	 * copy-on-write, then it is not safe to write as we may
899 	 * recurse into the copy-on-write routine.
900 	 */
901 	if (curthread->td_pflags & TDP_COWINPROGRESS)
902 		return (-1);
903 	/*
904 	 * Normally we just process each item on the worklist in order.
905 	 * However, if we are in a situation where we cannot lock any
906 	 * inodes, we have to skip over any dirrem requests whose
907 	 * vnodes are resident and locked.
908 	 */
909 	ump = VFSTOUFS(mp);
910 	vp = NULL;
911 	LIST_FOREACH(wk, &ump->softdep_workitem_pending, wk_list) {
912 		if (wk->wk_state & INPROGRESS)
913 			continue;
914 		if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
915 			break;
916 		wk->wk_state |= INPROGRESS;
917 		ump->softdep_on_worklist_inprogress++;
918 		FREE_LOCK(&lk);
919 		ffs_vgetf(mp, WK_DIRREM(wk)->dm_oldinum,
920 		    LK_NOWAIT | LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ);
921 		ACQUIRE_LOCK(&lk);
922 		wk->wk_state &= ~INPROGRESS;
923 		ump->softdep_on_worklist_inprogress--;
924 		if (vp != NULL)
925 			break;
926 	}
927 	if (wk == 0)
928 		return (-1);
929 	/*
930 	 * Remove the item to be processed. If we are removing the last
931 	 * item on the list, we need to recalculate the tail pointer.
932 	 * As this happens rarely and usually when the list is short,
933 	 * we just run down the list to find it rather than tracking it
934 	 * in the above loop.
935 	 */
936 	WORKLIST_REMOVE(wk);
937 	if (wk == ump->softdep_worklist_tail) {
938 		LIST_FOREACH(wkend, &ump->softdep_workitem_pending, wk_list)
939 			if (LIST_NEXT(wkend, wk_list) == NULL)
940 				break;
941 		ump->softdep_worklist_tail = wkend;
942 	}
943 	ump->softdep_on_worklist -= 1;
944 	FREE_LOCK(&lk);
945 	if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
946 		panic("process_worklist_item: suspended filesystem");
947 	matchcnt++;
948 	switch (wk->wk_type) {
949 
950 	case D_DIRREM:
951 		/* removal of a directory entry */
952 		handle_workitem_remove(WK_DIRREM(wk), vp);
953 		break;
954 
955 	case D_FREEBLKS:
956 		/* releasing blocks and/or fragments from a file */
957 		handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT);
958 		break;
959 
960 	case D_FREEFRAG:
961 		/* releasing a fragment when replaced as a file grows */
962 		handle_workitem_freefrag(WK_FREEFRAG(wk));
963 		break;
964 
965 	case D_FREEFILE:
966 		/* releasing an inode when its link count drops to 0 */
967 		handle_workitem_freefile(WK_FREEFILE(wk));
968 		break;
969 
970 	default:
971 		panic("%s_process_worklist: Unknown type %s",
972 		    "softdep", TYPENAME(wk->wk_type));
973 		/* NOTREACHED */
974 	}
975 	vn_finished_secondary_write(mp);
976 	ACQUIRE_LOCK(&lk);
977 	return (matchcnt);
978 }
979 
980 /*
981  * Move dependencies from one buffer to another.
982  */
983 void
984 softdep_move_dependencies(oldbp, newbp)
985 	struct buf *oldbp;
986 	struct buf *newbp;
987 {
988 	struct worklist *wk, *wktail;
989 
990 	if (!LIST_EMPTY(&newbp->b_dep))
991 		panic("softdep_move_dependencies: need merge code");
992 	wktail = 0;
993 	ACQUIRE_LOCK(&lk);
994 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
995 		LIST_REMOVE(wk, wk_list);
996 		if (wktail == 0)
997 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
998 		else
999 			LIST_INSERT_AFTER(wktail, wk, wk_list);
1000 		wktail = wk;
1001 	}
1002 	FREE_LOCK(&lk);
1003 }
1004 
1005 /*
1006  * Purge the work list of all items associated with a particular mount point.
1007  */
1008 int
1009 softdep_flushworklist(oldmnt, countp, td)
1010 	struct mount *oldmnt;
1011 	int *countp;
1012 	struct thread *td;
1013 {
1014 	struct vnode *devvp;
1015 	int count, error = 0;
1016 	struct ufsmount *ump;
1017 
1018 	/*
1019 	 * Alternately flush the block device associated with the mount
1020 	 * point and process any dependencies that the flushing
1021 	 * creates. We continue until no more worklist dependencies
1022 	 * are found.
1023 	 */
1024 	*countp = 0;
1025 	ump = VFSTOUFS(oldmnt);
1026 	devvp = ump->um_devvp;
1027 	while ((count = softdep_process_worklist(oldmnt, 1)) > 0) {
1028 		*countp += count;
1029 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1030 		error = VOP_FSYNC(devvp, MNT_WAIT, td);
1031 		VOP_UNLOCK(devvp, 0);
1032 		if (error)
1033 			break;
1034 	}
1035 	return (error);
1036 }
1037 
1038 int
1039 softdep_waitidle(struct mount *mp)
1040 {
1041 	struct ufsmount *ump;
1042 	int error;
1043 	int i;
1044 
1045 	ump = VFSTOUFS(mp);
1046 	ACQUIRE_LOCK(&lk);
1047 	for (i = 0; i < 10 && ump->softdep_deps; i++) {
1048 		ump->softdep_req = 1;
1049 		if (ump->softdep_on_worklist)
1050 			panic("softdep_waitidle: work added after flush.");
1051 		msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1);
1052 	}
1053 	ump->softdep_req = 0;
1054 	FREE_LOCK(&lk);
1055 	error = 0;
1056 	if (i == 10) {
1057 		error = EBUSY;
1058 		printf("softdep_waitidle: Failed to flush worklist for %p\n",
1059 		    mp);
1060 	}
1061 
1062 	return (error);
1063 }
1064 
1065 /*
1066  * Flush all vnodes and worklist items associated with a specified mount point.
1067  */
1068 int
1069 softdep_flushfiles(oldmnt, flags, td)
1070 	struct mount *oldmnt;
1071 	int flags;
1072 	struct thread *td;
1073 {
1074 	int error, depcount, loopcnt, retry_flush_count, retry;
1075 
1076 	loopcnt = 10;
1077 	retry_flush_count = 3;
1078 retry_flush:
1079 	error = 0;
1080 
1081 	/*
1082 	 * Alternately flush the vnodes associated with the mount
1083 	 * point and process any dependencies that the flushing
1084 	 * creates. In theory, this loop can happen at most twice,
1085 	 * but we give it a few extra just to be sure.
1086 	 */
1087 	for (; loopcnt > 0; loopcnt--) {
1088 		/*
1089 		 * Do another flush in case any vnodes were brought in
1090 		 * as part of the cleanup operations.
1091 		 */
1092 		if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
1093 			break;
1094 		if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 ||
1095 		    depcount == 0)
1096 			break;
1097 	}
1098 	/*
1099 	 * If we are unmounting then it is an error to fail. If we
1100 	 * are simply trying to downgrade to read-only, then filesystem
1101 	 * activity can keep us busy forever, so we just fail with EBUSY.
1102 	 */
1103 	if (loopcnt == 0) {
1104 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
1105 			panic("softdep_flushfiles: looping");
1106 		error = EBUSY;
1107 	}
1108 	if (!error)
1109 		error = softdep_waitidle(oldmnt);
1110 	if (!error) {
1111 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) {
1112 			retry = 0;
1113 			MNT_ILOCK(oldmnt);
1114 			KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0,
1115 			    ("softdep_flushfiles: !MNTK_NOINSMNTQ"));
1116 			if (oldmnt->mnt_nvnodelistsize > 0) {
1117 				if (--retry_flush_count > 0) {
1118 					retry = 1;
1119 					loopcnt = 3;
1120 				} else
1121 					error = EBUSY;
1122 			}
1123 			MNT_IUNLOCK(oldmnt);
1124 			if (retry)
1125 				goto retry_flush;
1126 		}
1127 	}
1128 	return (error);
1129 }
1130 
1131 /*
1132  * Structure hashing.
1133  *
1134  * There are three types of structures that can be looked up:
1135  *	1) pagedep structures identified by mount point, inode number,
1136  *	   and logical block.
1137  *	2) inodedep structures identified by mount point and inode number.
1138  *	3) newblk structures identified by mount point and
1139  *	   physical block number.
1140  *
1141  * The "pagedep" and "inodedep" dependency structures are hashed
1142  * separately from the file blocks and inodes to which they correspond.
1143  * This separation helps when the in-memory copy of an inode or
1144  * file block must be replaced. It also obviates the need to access
1145  * an inode or file page when simply updating (or de-allocating)
1146  * dependency structures. Lookup of newblk structures is needed to
1147  * find newly allocated blocks when trying to associate them with
1148  * their allocdirect or allocindir structure.
1149  *
1150  * The lookup routines optionally create and hash a new instance when
1151  * an existing entry is not found.
1152  */
1153 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
1154 #define NODELAY		0x0002	/* cannot do background work */
1155 
1156 /*
1157  * Structures and routines associated with pagedep caching.
1158  */
1159 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
1160 u_long	pagedep_hash;		/* size of hash table - 1 */
1161 #define	PAGEDEP_HASH(mp, inum, lbn) \
1162 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
1163 	    pagedep_hash])
1164 
1165 static int
1166 pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp)
1167 	struct pagedep_hashhead *pagedephd;
1168 	ino_t ino;
1169 	ufs_lbn_t lbn;
1170 	struct mount *mp;
1171 	int flags;
1172 	struct pagedep **pagedeppp;
1173 {
1174 	struct pagedep *pagedep;
1175 
1176 	LIST_FOREACH(pagedep, pagedephd, pd_hash)
1177 		if (ino == pagedep->pd_ino &&
1178 		    lbn == pagedep->pd_lbn &&
1179 		    mp == pagedep->pd_list.wk_mp)
1180 			break;
1181 	if (pagedep) {
1182 		*pagedeppp = pagedep;
1183 		if ((flags & DEPALLOC) != 0 &&
1184 		    (pagedep->pd_state & ONWORKLIST) == 0)
1185 			return (0);
1186 		return (1);
1187 	}
1188 	*pagedeppp = NULL;
1189 	return (0);
1190 }
1191 /*
1192  * Look up a pagedep. Return 1 if found, 0 if not found or found
1193  * when asked to allocate but not associated with any buffer.
1194  * If not found, allocate if DEPALLOC flag is passed.
1195  * Found or allocated entry is returned in pagedeppp.
1196  * This routine must be called with splbio interrupts blocked.
1197  */
1198 static int
1199 pagedep_lookup(ip, lbn, flags, pagedeppp)
1200 	struct inode *ip;
1201 	ufs_lbn_t lbn;
1202 	int flags;
1203 	struct pagedep **pagedeppp;
1204 {
1205 	struct pagedep *pagedep;
1206 	struct pagedep_hashhead *pagedephd;
1207 	struct mount *mp;
1208 	int ret;
1209 	int i;
1210 
1211 	mtx_assert(&lk, MA_OWNED);
1212 	mp = ITOV(ip)->v_mount;
1213 	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
1214 
1215 	ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp);
1216 	if (*pagedeppp || (flags & DEPALLOC) == 0)
1217 		return (ret);
1218 	FREE_LOCK(&lk);
1219 	pagedep = malloc(sizeof(struct pagedep),
1220 	    M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO);
1221 	workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp);
1222 	ACQUIRE_LOCK(&lk);
1223 	ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp);
1224 	if (*pagedeppp) {
1225 		WORKITEM_FREE(pagedep, D_PAGEDEP);
1226 		return (ret);
1227 	}
1228 	pagedep->pd_ino = ip->i_number;
1229 	pagedep->pd_lbn = lbn;
1230 	LIST_INIT(&pagedep->pd_dirremhd);
1231 	LIST_INIT(&pagedep->pd_pendinghd);
1232 	for (i = 0; i < DAHASHSZ; i++)
1233 		LIST_INIT(&pagedep->pd_diraddhd[i]);
1234 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
1235 	*pagedeppp = pagedep;
1236 	return (0);
1237 }
1238 
1239 /*
1240  * Structures and routines associated with inodedep caching.
1241  */
1242 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
1243 static u_long	inodedep_hash;	/* size of hash table - 1 */
1244 static long	num_inodedep;	/* number of inodedep allocated */
1245 #define	INODEDEP_HASH(fs, inum) \
1246       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
1247 
1248 static int
1249 inodedep_find(inodedephd, fs, inum, inodedeppp)
1250 	struct inodedep_hashhead *inodedephd;
1251 	struct fs *fs;
1252 	ino_t inum;
1253 	struct inodedep **inodedeppp;
1254 {
1255 	struct inodedep *inodedep;
1256 
1257 	LIST_FOREACH(inodedep, inodedephd, id_hash)
1258 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
1259 			break;
1260 	if (inodedep) {
1261 		*inodedeppp = inodedep;
1262 		return (1);
1263 	}
1264 	*inodedeppp = NULL;
1265 
1266 	return (0);
1267 }
1268 /*
1269  * Look up an inodedep. Return 1 if found, 0 if not found.
1270  * If not found, allocate if DEPALLOC flag is passed.
1271  * Found or allocated entry is returned in inodedeppp.
1272  * This routine must be called with splbio interrupts blocked.
1273  */
1274 static int
1275 inodedep_lookup(mp, inum, flags, inodedeppp)
1276 	struct mount *mp;
1277 	ino_t inum;
1278 	int flags;
1279 	struct inodedep **inodedeppp;
1280 {
1281 	struct inodedep *inodedep;
1282 	struct inodedep_hashhead *inodedephd;
1283 	struct fs *fs;
1284 
1285 	mtx_assert(&lk, MA_OWNED);
1286 	fs = VFSTOUFS(mp)->um_fs;
1287 	inodedephd = INODEDEP_HASH(fs, inum);
1288 
1289 	if (inodedep_find(inodedephd, fs, inum, inodedeppp))
1290 		return (1);
1291 	if ((flags & DEPALLOC) == 0)
1292 		return (0);
1293 	/*
1294 	 * If we are over our limit, try to improve the situation.
1295 	 */
1296 	if (num_inodedep > max_softdeps && (flags & NODELAY) == 0)
1297 		request_cleanup(mp, FLUSH_INODES);
1298 	FREE_LOCK(&lk);
1299 	inodedep = malloc(sizeof(struct inodedep),
1300 		M_INODEDEP, M_SOFTDEP_FLAGS);
1301 	workitem_alloc(&inodedep->id_list, D_INODEDEP, mp);
1302 	ACQUIRE_LOCK(&lk);
1303 	if (inodedep_find(inodedephd, fs, inum, inodedeppp)) {
1304 		WORKITEM_FREE(inodedep, D_INODEDEP);
1305 		return (1);
1306 	}
1307 	num_inodedep += 1;
1308 	inodedep->id_fs = fs;
1309 	inodedep->id_ino = inum;
1310 	inodedep->id_state = ALLCOMPLETE;
1311 	inodedep->id_nlinkdelta = 0;
1312 	inodedep->id_savedino1 = NULL;
1313 	inodedep->id_savedsize = -1;
1314 	inodedep->id_savedextsize = -1;
1315 	inodedep->id_buf = NULL;
1316 	LIST_INIT(&inodedep->id_pendinghd);
1317 	LIST_INIT(&inodedep->id_inowait);
1318 	LIST_INIT(&inodedep->id_bufwait);
1319 	TAILQ_INIT(&inodedep->id_inoupdt);
1320 	TAILQ_INIT(&inodedep->id_newinoupdt);
1321 	TAILQ_INIT(&inodedep->id_extupdt);
1322 	TAILQ_INIT(&inodedep->id_newextupdt);
1323 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1324 	*inodedeppp = inodedep;
1325 	return (0);
1326 }
1327 
1328 /*
1329  * Structures and routines associated with newblk caching.
1330  */
1331 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1332 u_long	newblk_hash;		/* size of hash table - 1 */
1333 #define	NEWBLK_HASH(fs, inum) \
1334 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1335 
1336 static int
1337 newblk_find(newblkhd, fs, newblkno, newblkpp)
1338 	struct newblk_hashhead *newblkhd;
1339 	struct fs *fs;
1340 	ufs2_daddr_t newblkno;
1341 	struct newblk **newblkpp;
1342 {
1343 	struct newblk *newblk;
1344 
1345 	LIST_FOREACH(newblk, newblkhd, nb_hash)
1346 		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1347 			break;
1348 	if (newblk) {
1349 		*newblkpp = newblk;
1350 		return (1);
1351 	}
1352 	*newblkpp = NULL;
1353 	return (0);
1354 }
1355 
1356 /*
1357  * Look up a newblk. Return 1 if found, 0 if not found.
1358  * If not found, allocate if DEPALLOC flag is passed.
1359  * Found or allocated entry is returned in newblkpp.
1360  */
1361 static int
1362 newblk_lookup(fs, newblkno, flags, newblkpp)
1363 	struct fs *fs;
1364 	ufs2_daddr_t newblkno;
1365 	int flags;
1366 	struct newblk **newblkpp;
1367 {
1368 	struct newblk *newblk;
1369 	struct newblk_hashhead *newblkhd;
1370 
1371 	newblkhd = NEWBLK_HASH(fs, newblkno);
1372 	if (newblk_find(newblkhd, fs, newblkno, newblkpp))
1373 		return (1);
1374 	if ((flags & DEPALLOC) == 0)
1375 		return (0);
1376 	FREE_LOCK(&lk);
1377 	newblk = malloc(sizeof(struct newblk),
1378 		M_NEWBLK, M_SOFTDEP_FLAGS);
1379 	ACQUIRE_LOCK(&lk);
1380 	if (newblk_find(newblkhd, fs, newblkno, newblkpp)) {
1381 		free(newblk, M_NEWBLK);
1382 		return (1);
1383 	}
1384 	newblk->nb_state = 0;
1385 	newblk->nb_fs = fs;
1386 	newblk->nb_newblkno = newblkno;
1387 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1388 	*newblkpp = newblk;
1389 	return (0);
1390 }
1391 
1392 /*
1393  * Executed during filesystem system initialization before
1394  * mounting any filesystems.
1395  */
1396 void
1397 softdep_initialize()
1398 {
1399 
1400 	LIST_INIT(&mkdirlisthd);
1401 	max_softdeps = desiredvnodes * 4;
1402 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1403 	    &pagedep_hash);
1404 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1405 	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1406 
1407 	/* initialise bioops hack */
1408 	bioops.io_start = softdep_disk_io_initiation;
1409 	bioops.io_complete = softdep_disk_write_complete;
1410 	bioops.io_deallocate = softdep_deallocate_dependencies;
1411 	bioops.io_countdeps = softdep_count_dependencies;
1412 
1413 	/* Initialize the callout with an mtx. */
1414 	callout_init_mtx(&softdep_callout, &lk, 0);
1415 }
1416 
1417 /*
1418  * Executed after all filesystems have been unmounted during
1419  * filesystem module unload.
1420  */
1421 void
1422 softdep_uninitialize()
1423 {
1424 
1425 	callout_drain(&softdep_callout);
1426 	hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash);
1427 	hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash);
1428 	hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash);
1429 }
1430 
1431 /*
1432  * Called at mount time to notify the dependency code that a
1433  * filesystem wishes to use it.
1434  */
1435 int
1436 softdep_mount(devvp, mp, fs, cred)
1437 	struct vnode *devvp;
1438 	struct mount *mp;
1439 	struct fs *fs;
1440 	struct ucred *cred;
1441 {
1442 	struct csum_total cstotal;
1443 	struct ufsmount *ump;
1444 	struct cg *cgp;
1445 	struct buf *bp;
1446 	int error, cyl;
1447 
1448 	MNT_ILOCK(mp);
1449 	mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP;
1450 	if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) {
1451 		mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) |
1452 			MNTK_SOFTDEP;
1453 		mp->mnt_noasync++;
1454 	}
1455 	MNT_IUNLOCK(mp);
1456 	ump = VFSTOUFS(mp);
1457 	LIST_INIT(&ump->softdep_workitem_pending);
1458 	ump->softdep_worklist_tail = NULL;
1459 	ump->softdep_on_worklist = 0;
1460 	ump->softdep_deps = 0;
1461 	/*
1462 	 * When doing soft updates, the counters in the
1463 	 * superblock may have gotten out of sync. Recomputation
1464 	 * can take a long time and can be deferred for background
1465 	 * fsck.  However, the old behavior of scanning the cylinder
1466 	 * groups and recalculating them at mount time is available
1467 	 * by setting vfs.ffs.compute_summary_at_mount to one.
1468 	 */
1469 	if (compute_summary_at_mount == 0 || fs->fs_clean != 0)
1470 		return (0);
1471 	bzero(&cstotal, sizeof cstotal);
1472 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1473 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1474 		    fs->fs_cgsize, cred, &bp)) != 0) {
1475 			brelse(bp);
1476 			return (error);
1477 		}
1478 		cgp = (struct cg *)bp->b_data;
1479 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1480 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1481 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1482 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1483 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
1484 		brelse(bp);
1485 	}
1486 #ifdef DEBUG
1487 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1488 		printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
1489 #endif
1490 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1491 	return (0);
1492 }
1493 
1494 /*
1495  * Protecting the freemaps (or bitmaps).
1496  *
1497  * To eliminate the need to execute fsck before mounting a filesystem
1498  * after a power failure, one must (conservatively) guarantee that the
1499  * on-disk copy of the bitmaps never indicate that a live inode or block is
1500  * free.  So, when a block or inode is allocated, the bitmap should be
1501  * updated (on disk) before any new pointers.  When a block or inode is
1502  * freed, the bitmap should not be updated until all pointers have been
1503  * reset.  The latter dependency is handled by the delayed de-allocation
1504  * approach described below for block and inode de-allocation.  The former
1505  * dependency is handled by calling the following procedure when a block or
1506  * inode is allocated. When an inode is allocated an "inodedep" is created
1507  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1508  * Each "inodedep" is also inserted into the hash indexing structure so
1509  * that any additional link additions can be made dependent on the inode
1510  * allocation.
1511  *
1512  * The ufs filesystem maintains a number of free block counts (e.g., per
1513  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1514  * in addition to the bitmaps.  These counts are used to improve efficiency
1515  * during allocation and therefore must be consistent with the bitmaps.
1516  * There is no convenient way to guarantee post-crash consistency of these
1517  * counts with simple update ordering, for two main reasons: (1) The counts
1518  * and bitmaps for a single cylinder group block are not in the same disk
1519  * sector.  If a disk write is interrupted (e.g., by power failure), one may
1520  * be written and the other not.  (2) Some of the counts are located in the
1521  * superblock rather than the cylinder group block. So, we focus our soft
1522  * updates implementation on protecting the bitmaps. When mounting a
1523  * filesystem, we recompute the auxiliary counts from the bitmaps.
1524  */
1525 
1526 /*
1527  * Called just after updating the cylinder group block to allocate an inode.
1528  */
1529 void
1530 softdep_setup_inomapdep(bp, ip, newinum)
1531 	struct buf *bp;		/* buffer for cylgroup block with inode map */
1532 	struct inode *ip;	/* inode related to allocation */
1533 	ino_t newinum;		/* new inode number being allocated */
1534 {
1535 	struct inodedep *inodedep;
1536 	struct bmsafemap *bmsafemap;
1537 
1538 	/*
1539 	 * Create a dependency for the newly allocated inode.
1540 	 * Panic if it already exists as something is seriously wrong.
1541 	 * Otherwise add it to the dependency list for the buffer holding
1542 	 * the cylinder group map from which it was allocated.
1543 	 */
1544 	ACQUIRE_LOCK(&lk);
1545 	if ((inodedep_lookup(UFSTOVFS(ip->i_ump), newinum, DEPALLOC|NODELAY,
1546 	    &inodedep)))
1547 		panic("softdep_setup_inomapdep: dependency for new inode "
1548 		    "already exists");
1549 	inodedep->id_buf = bp;
1550 	inodedep->id_state &= ~DEPCOMPLETE;
1551 	bmsafemap = bmsafemap_lookup(inodedep->id_list.wk_mp, bp);
1552 	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1553 	FREE_LOCK(&lk);
1554 }
1555 
1556 /*
1557  * Called just after updating the cylinder group block to
1558  * allocate block or fragment.
1559  */
1560 void
1561 softdep_setup_blkmapdep(bp, mp, newblkno)
1562 	struct buf *bp;		/* buffer for cylgroup block with block map */
1563 	struct mount *mp;	/* filesystem doing allocation */
1564 	ufs2_daddr_t newblkno;	/* number of newly allocated block */
1565 {
1566 	struct newblk *newblk;
1567 	struct bmsafemap *bmsafemap;
1568 	struct fs *fs;
1569 
1570 	fs = VFSTOUFS(mp)->um_fs;
1571 	/*
1572 	 * Create a dependency for the newly allocated block.
1573 	 * Add it to the dependency list for the buffer holding
1574 	 * the cylinder group map from which it was allocated.
1575 	 */
1576 	ACQUIRE_LOCK(&lk);
1577 	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1578 		panic("softdep_setup_blkmapdep: found block");
1579 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp);
1580 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1581 	FREE_LOCK(&lk);
1582 }
1583 
1584 /*
1585  * Find the bmsafemap associated with a cylinder group buffer.
1586  * If none exists, create one. The buffer must be locked when
1587  * this routine is called and this routine must be called with
1588  * splbio interrupts blocked.
1589  */
1590 static struct bmsafemap *
1591 bmsafemap_lookup(mp, bp)
1592 	struct mount *mp;
1593 	struct buf *bp;
1594 {
1595 	struct bmsafemap *bmsafemap;
1596 	struct worklist *wk;
1597 
1598 	mtx_assert(&lk, MA_OWNED);
1599 	LIST_FOREACH(wk, &bp->b_dep, wk_list)
1600 		if (wk->wk_type == D_BMSAFEMAP)
1601 			return (WK_BMSAFEMAP(wk));
1602 	FREE_LOCK(&lk);
1603 	bmsafemap = malloc(sizeof(struct bmsafemap),
1604 		M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1605 	workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
1606 	bmsafemap->sm_buf = bp;
1607 	LIST_INIT(&bmsafemap->sm_allocdirecthd);
1608 	LIST_INIT(&bmsafemap->sm_allocindirhd);
1609 	LIST_INIT(&bmsafemap->sm_inodedephd);
1610 	LIST_INIT(&bmsafemap->sm_newblkhd);
1611 	ACQUIRE_LOCK(&lk);
1612 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1613 	return (bmsafemap);
1614 }
1615 
1616 /*
1617  * Direct block allocation dependencies.
1618  *
1619  * When a new block is allocated, the corresponding disk locations must be
1620  * initialized (with zeros or new data) before the on-disk inode points to
1621  * them.  Also, the freemap from which the block was allocated must be
1622  * updated (on disk) before the inode's pointer. These two dependencies are
1623  * independent of each other and are needed for all file blocks and indirect
1624  * blocks that are pointed to directly by the inode.  Just before the
1625  * "in-core" version of the inode is updated with a newly allocated block
1626  * number, a procedure (below) is called to setup allocation dependency
1627  * structures.  These structures are removed when the corresponding
1628  * dependencies are satisfied or when the block allocation becomes obsolete
1629  * (i.e., the file is deleted, the block is de-allocated, or the block is a
1630  * fragment that gets upgraded).  All of these cases are handled in
1631  * procedures described later.
1632  *
1633  * When a file extension causes a fragment to be upgraded, either to a larger
1634  * fragment or to a full block, the on-disk location may change (if the
1635  * previous fragment could not simply be extended). In this case, the old
1636  * fragment must be de-allocated, but not until after the inode's pointer has
1637  * been updated. In most cases, this is handled by later procedures, which
1638  * will construct a "freefrag" structure to be added to the workitem queue
1639  * when the inode update is complete (or obsolete).  The main exception to
1640  * this is when an allocation occurs while a pending allocation dependency
1641  * (for the same block pointer) remains.  This case is handled in the main
1642  * allocation dependency setup procedure by immediately freeing the
1643  * unreferenced fragments.
1644  */
1645 void
1646 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1647 	struct inode *ip;	/* inode to which block is being added */
1648 	ufs_lbn_t lbn;		/* block pointer within inode */
1649 	ufs2_daddr_t newblkno;	/* disk block number being added */
1650 	ufs2_daddr_t oldblkno;	/* previous block number, 0 unless frag */
1651 	long newsize;		/* size of new block */
1652 	long oldsize;		/* size of new block */
1653 	struct buf *bp;		/* bp for allocated block */
1654 {
1655 	struct allocdirect *adp, *oldadp;
1656 	struct allocdirectlst *adphead;
1657 	struct bmsafemap *bmsafemap;
1658 	struct inodedep *inodedep;
1659 	struct pagedep *pagedep;
1660 	struct newblk *newblk;
1661 	struct mount *mp;
1662 
1663 	mp = UFSTOVFS(ip->i_ump);
1664 	adp = malloc(sizeof(struct allocdirect),
1665 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1666 	workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp);
1667 	adp->ad_lbn = lbn;
1668 	adp->ad_newblkno = newblkno;
1669 	adp->ad_oldblkno = oldblkno;
1670 	adp->ad_newsize = newsize;
1671 	adp->ad_oldsize = oldsize;
1672 	adp->ad_state = ATTACHED;
1673 	LIST_INIT(&adp->ad_newdirblk);
1674 	if (newblkno == oldblkno)
1675 		adp->ad_freefrag = NULL;
1676 	else
1677 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1678 
1679 	ACQUIRE_LOCK(&lk);
1680 	if (lbn >= NDADDR) {
1681 		/* allocating an indirect block */
1682 		if (oldblkno != 0)
1683 			panic("softdep_setup_allocdirect: non-zero indir");
1684 	} else {
1685 		/*
1686 		 * Allocating a direct block.
1687 		 *
1688 		 * If we are allocating a directory block, then we must
1689 		 * allocate an associated pagedep to track additions and
1690 		 * deletions.
1691 		 */
1692 		if ((ip->i_mode & IFMT) == IFDIR &&
1693 		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1694 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1695 	}
1696 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1697 		panic("softdep_setup_allocdirect: lost block");
1698 	if (newblk->nb_state == DEPCOMPLETE) {
1699 		adp->ad_state |= DEPCOMPLETE;
1700 		adp->ad_buf = NULL;
1701 	} else {
1702 		bmsafemap = newblk->nb_bmsafemap;
1703 		adp->ad_buf = bmsafemap->sm_buf;
1704 		LIST_REMOVE(newblk, nb_deps);
1705 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1706 	}
1707 	LIST_REMOVE(newblk, nb_hash);
1708 	free(newblk, M_NEWBLK);
1709 
1710 	inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1711 	adp->ad_inodedep = inodedep;
1712 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1713 	/*
1714 	 * The list of allocdirects must be kept in sorted and ascending
1715 	 * order so that the rollback routines can quickly determine the
1716 	 * first uncommitted block (the size of the file stored on disk
1717 	 * ends at the end of the lowest committed fragment, or if there
1718 	 * are no fragments, at the end of the highest committed block).
1719 	 * Since files generally grow, the typical case is that the new
1720 	 * block is to be added at the end of the list. We speed this
1721 	 * special case by checking against the last allocdirect in the
1722 	 * list before laboriously traversing the list looking for the
1723 	 * insertion point.
1724 	 */
1725 	adphead = &inodedep->id_newinoupdt;
1726 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1727 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1728 		/* insert at end of list */
1729 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1730 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1731 			allocdirect_merge(adphead, adp, oldadp);
1732 		FREE_LOCK(&lk);
1733 		return;
1734 	}
1735 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1736 		if (oldadp->ad_lbn >= lbn)
1737 			break;
1738 	}
1739 	if (oldadp == NULL)
1740 		panic("softdep_setup_allocdirect: lost entry");
1741 	/* insert in middle of list */
1742 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1743 	if (oldadp->ad_lbn == lbn)
1744 		allocdirect_merge(adphead, adp, oldadp);
1745 	FREE_LOCK(&lk);
1746 }
1747 
1748 /*
1749  * Replace an old allocdirect dependency with a newer one.
1750  * This routine must be called with splbio interrupts blocked.
1751  */
1752 static void
1753 allocdirect_merge(adphead, newadp, oldadp)
1754 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
1755 	struct allocdirect *newadp;	/* allocdirect being added */
1756 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
1757 {
1758 	struct worklist *wk;
1759 	struct freefrag *freefrag;
1760 	struct newdirblk *newdirblk;
1761 
1762 	mtx_assert(&lk, MA_OWNED);
1763 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1764 	    newadp->ad_oldsize != oldadp->ad_newsize ||
1765 	    newadp->ad_lbn >= NDADDR)
1766 		panic("%s %jd != new %jd || old size %ld != new %ld",
1767 		    "allocdirect_merge: old blkno",
1768 		    (intmax_t)newadp->ad_oldblkno,
1769 		    (intmax_t)oldadp->ad_newblkno,
1770 		    newadp->ad_oldsize, oldadp->ad_newsize);
1771 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
1772 	newadp->ad_oldsize = oldadp->ad_oldsize;
1773 	/*
1774 	 * If the old dependency had a fragment to free or had never
1775 	 * previously had a block allocated, then the new dependency
1776 	 * can immediately post its freefrag and adopt the old freefrag.
1777 	 * This action is done by swapping the freefrag dependencies.
1778 	 * The new dependency gains the old one's freefrag, and the
1779 	 * old one gets the new one and then immediately puts it on
1780 	 * the worklist when it is freed by free_allocdirect. It is
1781 	 * not possible to do this swap when the old dependency had a
1782 	 * non-zero size but no previous fragment to free. This condition
1783 	 * arises when the new block is an extension of the old block.
1784 	 * Here, the first part of the fragment allocated to the new
1785 	 * dependency is part of the block currently claimed on disk by
1786 	 * the old dependency, so cannot legitimately be freed until the
1787 	 * conditions for the new dependency are fulfilled.
1788 	 */
1789 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1790 		freefrag = newadp->ad_freefrag;
1791 		newadp->ad_freefrag = oldadp->ad_freefrag;
1792 		oldadp->ad_freefrag = freefrag;
1793 	}
1794 	/*
1795 	 * If we are tracking a new directory-block allocation,
1796 	 * move it from the old allocdirect to the new allocdirect.
1797 	 */
1798 	if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
1799 		newdirblk = WK_NEWDIRBLK(wk);
1800 		WORKLIST_REMOVE(&newdirblk->db_list);
1801 		if (!LIST_EMPTY(&oldadp->ad_newdirblk))
1802 			panic("allocdirect_merge: extra newdirblk");
1803 		WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list);
1804 	}
1805 	free_allocdirect(adphead, oldadp, 0);
1806 }
1807 
1808 /*
1809  * Allocate a new freefrag structure if needed.
1810  */
1811 static struct freefrag *
1812 newfreefrag(ip, blkno, size)
1813 	struct inode *ip;
1814 	ufs2_daddr_t blkno;
1815 	long size;
1816 {
1817 	struct freefrag *freefrag;
1818 	struct fs *fs;
1819 
1820 	if (blkno == 0)
1821 		return (NULL);
1822 	fs = ip->i_fs;
1823 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1824 		panic("newfreefrag: frag size");
1825 	freefrag = malloc(sizeof(struct freefrag),
1826 		M_FREEFRAG, M_SOFTDEP_FLAGS);
1827 	workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump));
1828 	freefrag->ff_inum = ip->i_number;
1829 	freefrag->ff_blkno = blkno;
1830 	freefrag->ff_fragsize = size;
1831 	return (freefrag);
1832 }
1833 
1834 /*
1835  * This workitem de-allocates fragments that were replaced during
1836  * file block allocation.
1837  */
1838 static void
1839 handle_workitem_freefrag(freefrag)
1840 	struct freefrag *freefrag;
1841 {
1842 	struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp);
1843 
1844 	ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
1845 	    freefrag->ff_fragsize, freefrag->ff_inum);
1846 	ACQUIRE_LOCK(&lk);
1847 	WORKITEM_FREE(freefrag, D_FREEFRAG);
1848 	FREE_LOCK(&lk);
1849 }
1850 
1851 /*
1852  * Set up a dependency structure for an external attributes data block.
1853  * This routine follows much of the structure of softdep_setup_allocdirect.
1854  * See the description of softdep_setup_allocdirect above for details.
1855  */
1856 void
1857 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1858 	struct inode *ip;
1859 	ufs_lbn_t lbn;
1860 	ufs2_daddr_t newblkno;
1861 	ufs2_daddr_t oldblkno;
1862 	long newsize;
1863 	long oldsize;
1864 	struct buf *bp;
1865 {
1866 	struct allocdirect *adp, *oldadp;
1867 	struct allocdirectlst *adphead;
1868 	struct bmsafemap *bmsafemap;
1869 	struct inodedep *inodedep;
1870 	struct newblk *newblk;
1871 	struct mount *mp;
1872 
1873 	mp = UFSTOVFS(ip->i_ump);
1874 	adp = malloc(sizeof(struct allocdirect),
1875 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1876 	workitem_alloc(&adp->ad_list, D_ALLOCDIRECT, mp);
1877 	adp->ad_lbn = lbn;
1878 	adp->ad_newblkno = newblkno;
1879 	adp->ad_oldblkno = oldblkno;
1880 	adp->ad_newsize = newsize;
1881 	adp->ad_oldsize = oldsize;
1882 	adp->ad_state = ATTACHED | EXTDATA;
1883 	LIST_INIT(&adp->ad_newdirblk);
1884 	if (newblkno == oldblkno)
1885 		adp->ad_freefrag = NULL;
1886 	else
1887 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1888 
1889 	ACQUIRE_LOCK(&lk);
1890 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1891 		panic("softdep_setup_allocext: lost block");
1892 
1893 	inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1894 	adp->ad_inodedep = inodedep;
1895 
1896 	if (newblk->nb_state == DEPCOMPLETE) {
1897 		adp->ad_state |= DEPCOMPLETE;
1898 		adp->ad_buf = NULL;
1899 	} else {
1900 		bmsafemap = newblk->nb_bmsafemap;
1901 		adp->ad_buf = bmsafemap->sm_buf;
1902 		LIST_REMOVE(newblk, nb_deps);
1903 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1904 	}
1905 	LIST_REMOVE(newblk, nb_hash);
1906 	free(newblk, M_NEWBLK);
1907 
1908 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1909 	if (lbn >= NXADDR)
1910 		panic("softdep_setup_allocext: lbn %lld > NXADDR",
1911 		    (long long)lbn);
1912 	/*
1913 	 * The list of allocdirects must be kept in sorted and ascending
1914 	 * order so that the rollback routines can quickly determine the
1915 	 * first uncommitted block (the size of the file stored on disk
1916 	 * ends at the end of the lowest committed fragment, or if there
1917 	 * are no fragments, at the end of the highest committed block).
1918 	 * Since files generally grow, the typical case is that the new
1919 	 * block is to be added at the end of the list. We speed this
1920 	 * special case by checking against the last allocdirect in the
1921 	 * list before laboriously traversing the list looking for the
1922 	 * insertion point.
1923 	 */
1924 	adphead = &inodedep->id_newextupdt;
1925 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1926 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1927 		/* insert at end of list */
1928 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1929 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1930 			allocdirect_merge(adphead, adp, oldadp);
1931 		FREE_LOCK(&lk);
1932 		return;
1933 	}
1934 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1935 		if (oldadp->ad_lbn >= lbn)
1936 			break;
1937 	}
1938 	if (oldadp == NULL)
1939 		panic("softdep_setup_allocext: lost entry");
1940 	/* insert in middle of list */
1941 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1942 	if (oldadp->ad_lbn == lbn)
1943 		allocdirect_merge(adphead, adp, oldadp);
1944 	FREE_LOCK(&lk);
1945 }
1946 
1947 /*
1948  * Indirect block allocation dependencies.
1949  *
1950  * The same dependencies that exist for a direct block also exist when
1951  * a new block is allocated and pointed to by an entry in a block of
1952  * indirect pointers. The undo/redo states described above are also
1953  * used here. Because an indirect block contains many pointers that
1954  * may have dependencies, a second copy of the entire in-memory indirect
1955  * block is kept. The buffer cache copy is always completely up-to-date.
1956  * The second copy, which is used only as a source for disk writes,
1957  * contains only the safe pointers (i.e., those that have no remaining
1958  * update dependencies). The second copy is freed when all pointers
1959  * are safe. The cache is not allowed to replace indirect blocks with
1960  * pending update dependencies. If a buffer containing an indirect
1961  * block with dependencies is written, these routines will mark it
1962  * dirty again. It can only be successfully written once all the
1963  * dependencies are removed. The ffs_fsync routine in conjunction with
1964  * softdep_sync_metadata work together to get all the dependencies
1965  * removed so that a file can be successfully written to disk. Three
1966  * procedures are used when setting up indirect block pointer
1967  * dependencies. The division is necessary because of the organization
1968  * of the "balloc" routine and because of the distinction between file
1969  * pages and file metadata blocks.
1970  */
1971 
1972 /*
1973  * Allocate a new allocindir structure.
1974  */
1975 static struct allocindir *
1976 newallocindir(ip, ptrno, newblkno, oldblkno)
1977 	struct inode *ip;	/* inode for file being extended */
1978 	int ptrno;		/* offset of pointer in indirect block */
1979 	ufs2_daddr_t newblkno;	/* disk block number being added */
1980 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1981 {
1982 	struct allocindir *aip;
1983 
1984 	aip = malloc(sizeof(struct allocindir),
1985 		M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO);
1986 	workitem_alloc(&aip->ai_list, D_ALLOCINDIR, UFSTOVFS(ip->i_ump));
1987 	aip->ai_state = ATTACHED;
1988 	aip->ai_offset = ptrno;
1989 	aip->ai_newblkno = newblkno;
1990 	aip->ai_oldblkno = oldblkno;
1991 	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1992 	return (aip);
1993 }
1994 
1995 /*
1996  * Called just before setting an indirect block pointer
1997  * to a newly allocated file page.
1998  */
1999 void
2000 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
2001 	struct inode *ip;	/* inode for file being extended */
2002 	ufs_lbn_t lbn;		/* allocated block number within file */
2003 	struct buf *bp;		/* buffer with indirect blk referencing page */
2004 	int ptrno;		/* offset of pointer in indirect block */
2005 	ufs2_daddr_t newblkno;	/* disk block number being added */
2006 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
2007 	struct buf *nbp;	/* buffer holding allocated page */
2008 {
2009 	struct allocindir *aip;
2010 	struct pagedep *pagedep;
2011 
2012 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page");
2013 	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
2014 	ACQUIRE_LOCK(&lk);
2015 	/*
2016 	 * If we are allocating a directory page, then we must
2017 	 * allocate an associated pagedep to track additions and
2018 	 * deletions.
2019 	 */
2020 	if ((ip->i_mode & IFMT) == IFDIR &&
2021 	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
2022 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
2023 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
2024 	setup_allocindir_phase2(bp, ip, aip);
2025 	FREE_LOCK(&lk);
2026 }
2027 
2028 /*
2029  * Called just before setting an indirect block pointer to a
2030  * newly allocated indirect block.
2031  */
2032 void
2033 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
2034 	struct buf *nbp;	/* newly allocated indirect block */
2035 	struct inode *ip;	/* inode for file being extended */
2036 	struct buf *bp;		/* indirect block referencing allocated block */
2037 	int ptrno;		/* offset of pointer in indirect block */
2038 	ufs2_daddr_t newblkno;	/* disk block number being added */
2039 {
2040 	struct allocindir *aip;
2041 
2042 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta");
2043 	aip = newallocindir(ip, ptrno, newblkno, 0);
2044 	ACQUIRE_LOCK(&lk);
2045 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
2046 	setup_allocindir_phase2(bp, ip, aip);
2047 	FREE_LOCK(&lk);
2048 }
2049 
2050 /*
2051  * Called to finish the allocation of the "aip" allocated
2052  * by one of the two routines above.
2053  */
2054 static void
2055 setup_allocindir_phase2(bp, ip, aip)
2056 	struct buf *bp;		/* in-memory copy of the indirect block */
2057 	struct inode *ip;	/* inode for file being extended */
2058 	struct allocindir *aip;	/* allocindir allocated by the above routines */
2059 {
2060 	struct worklist *wk;
2061 	struct indirdep *indirdep, *newindirdep;
2062 	struct bmsafemap *bmsafemap;
2063 	struct allocindir *oldaip;
2064 	struct freefrag *freefrag;
2065 	struct newblk *newblk;
2066 	ufs2_daddr_t blkno;
2067 
2068 	mtx_assert(&lk, MA_OWNED);
2069 	if (bp->b_lblkno >= 0)
2070 		panic("setup_allocindir_phase2: not indir blk");
2071 	for (indirdep = NULL, newindirdep = NULL; ; ) {
2072 		LIST_FOREACH(wk, &bp->b_dep, wk_list) {
2073 			if (wk->wk_type != D_INDIRDEP)
2074 				continue;
2075 			indirdep = WK_INDIRDEP(wk);
2076 			break;
2077 		}
2078 		if (indirdep == NULL && newindirdep) {
2079 			indirdep = newindirdep;
2080 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
2081 			newindirdep = NULL;
2082 		}
2083 		if (indirdep) {
2084 			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
2085 			    &newblk) == 0)
2086 				panic("setup_allocindir: lost block");
2087 			if (newblk->nb_state == DEPCOMPLETE) {
2088 				aip->ai_state |= DEPCOMPLETE;
2089 				aip->ai_buf = NULL;
2090 			} else {
2091 				bmsafemap = newblk->nb_bmsafemap;
2092 				aip->ai_buf = bmsafemap->sm_buf;
2093 				LIST_REMOVE(newblk, nb_deps);
2094 				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
2095 				    aip, ai_deps);
2096 			}
2097 			LIST_REMOVE(newblk, nb_hash);
2098 			free(newblk, M_NEWBLK);
2099 			aip->ai_indirdep = indirdep;
2100 			/*
2101 			 * Check to see if there is an existing dependency
2102 			 * for this block. If there is, merge the old
2103 			 * dependency into the new one.
2104 			 */
2105 			if (aip->ai_oldblkno == 0)
2106 				oldaip = NULL;
2107 			else
2108 
2109 				LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
2110 					if (oldaip->ai_offset == aip->ai_offset)
2111 						break;
2112 			freefrag = NULL;
2113 			if (oldaip != NULL) {
2114 				if (oldaip->ai_newblkno != aip->ai_oldblkno)
2115 					panic("setup_allocindir_phase2: blkno");
2116 				aip->ai_oldblkno = oldaip->ai_oldblkno;
2117 				freefrag = aip->ai_freefrag;
2118 				aip->ai_freefrag = oldaip->ai_freefrag;
2119 				oldaip->ai_freefrag = NULL;
2120 				free_allocindir(oldaip, NULL);
2121 			}
2122 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
2123 			if (ip->i_ump->um_fstype == UFS1)
2124 				((ufs1_daddr_t *)indirdep->ir_savebp->b_data)
2125 				    [aip->ai_offset] = aip->ai_oldblkno;
2126 			else
2127 				((ufs2_daddr_t *)indirdep->ir_savebp->b_data)
2128 				    [aip->ai_offset] = aip->ai_oldblkno;
2129 			FREE_LOCK(&lk);
2130 			if (freefrag != NULL)
2131 				handle_workitem_freefrag(freefrag);
2132 		} else
2133 			FREE_LOCK(&lk);
2134 		if (newindirdep) {
2135 			newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
2136 			brelse(newindirdep->ir_savebp);
2137 			ACQUIRE_LOCK(&lk);
2138 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
2139 			if (indirdep)
2140 				break;
2141 			FREE_LOCK(&lk);
2142 		}
2143 		if (indirdep) {
2144 			ACQUIRE_LOCK(&lk);
2145 			break;
2146 		}
2147 		newindirdep = malloc(sizeof(struct indirdep),
2148 			M_INDIRDEP, M_SOFTDEP_FLAGS);
2149 		workitem_alloc(&newindirdep->ir_list, D_INDIRDEP,
2150 		    UFSTOVFS(ip->i_ump));
2151 		newindirdep->ir_state = ATTACHED;
2152 		if (ip->i_ump->um_fstype == UFS1)
2153 			newindirdep->ir_state |= UFS1FMT;
2154 		LIST_INIT(&newindirdep->ir_deplisthd);
2155 		LIST_INIT(&newindirdep->ir_donehd);
2156 		if (bp->b_blkno == bp->b_lblkno) {
2157 			ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
2158 			    NULL, NULL);
2159 			bp->b_blkno = blkno;
2160 		}
2161 		newindirdep->ir_savebp =
2162 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
2163 		BUF_KERNPROC(newindirdep->ir_savebp);
2164 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
2165 		ACQUIRE_LOCK(&lk);
2166 	}
2167 }
2168 
2169 /*
2170  * Block de-allocation dependencies.
2171  *
2172  * When blocks are de-allocated, the on-disk pointers must be nullified before
2173  * the blocks are made available for use by other files.  (The true
2174  * requirement is that old pointers must be nullified before new on-disk
2175  * pointers are set.  We chose this slightly more stringent requirement to
2176  * reduce complexity.) Our implementation handles this dependency by updating
2177  * the inode (or indirect block) appropriately but delaying the actual block
2178  * de-allocation (i.e., freemap and free space count manipulation) until
2179  * after the updated versions reach stable storage.  After the disk is
2180  * updated, the blocks can be safely de-allocated whenever it is convenient.
2181  * This implementation handles only the common case of reducing a file's
2182  * length to zero. Other cases are handled by the conventional synchronous
2183  * write approach.
2184  *
2185  * The ffs implementation with which we worked double-checks
2186  * the state of the block pointers and file size as it reduces
2187  * a file's length.  Some of this code is replicated here in our
2188  * soft updates implementation.  The freeblks->fb_chkcnt field is
2189  * used to transfer a part of this information to the procedure
2190  * that eventually de-allocates the blocks.
2191  *
2192  * This routine should be called from the routine that shortens
2193  * a file's length, before the inode's size or block pointers
2194  * are modified. It will save the block pointer information for
2195  * later release and zero the inode so that the calling routine
2196  * can release it.
2197  */
2198 void
2199 softdep_setup_freeblocks(ip, length, flags)
2200 	struct inode *ip;	/* The inode whose length is to be reduced */
2201 	off_t length;		/* The new length for the file */
2202 	int flags;		/* IO_EXT and/or IO_NORMAL */
2203 {
2204 	struct freeblks *freeblks;
2205 	struct inodedep *inodedep;
2206 	struct allocdirect *adp;
2207 	struct bufobj *bo;
2208 	struct vnode *vp;
2209 	struct buf *bp;
2210 	struct fs *fs;
2211 	ufs2_daddr_t extblocks, datablocks;
2212 	struct mount *mp;
2213 	int i, delay, error;
2214 
2215 	fs = ip->i_fs;
2216 	mp = UFSTOVFS(ip->i_ump);
2217 	if (length != 0)
2218 		panic("softdep_setup_freeblocks: non-zero length");
2219 	freeblks = malloc(sizeof(struct freeblks),
2220 		M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
2221 	workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp);
2222 	freeblks->fb_state = ATTACHED;
2223 	freeblks->fb_uid = ip->i_uid;
2224 	freeblks->fb_previousinum = ip->i_number;
2225 	freeblks->fb_devvp = ip->i_devvp;
2226 	extblocks = 0;
2227 	if (fs->fs_magic == FS_UFS2_MAGIC)
2228 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
2229 	datablocks = DIP(ip, i_blocks) - extblocks;
2230 	if ((flags & IO_NORMAL) == 0) {
2231 		freeblks->fb_oldsize = 0;
2232 		freeblks->fb_chkcnt = 0;
2233 	} else {
2234 		freeblks->fb_oldsize = ip->i_size;
2235 		ip->i_size = 0;
2236 		DIP_SET(ip, i_size, 0);
2237 		freeblks->fb_chkcnt = datablocks;
2238 		for (i = 0; i < NDADDR; i++) {
2239 			freeblks->fb_dblks[i] = DIP(ip, i_db[i]);
2240 			DIP_SET(ip, i_db[i], 0);
2241 		}
2242 		for (i = 0; i < NIADDR; i++) {
2243 			freeblks->fb_iblks[i] = DIP(ip, i_ib[i]);
2244 			DIP_SET(ip, i_ib[i], 0);
2245 		}
2246 		/*
2247 		 * If the file was removed, then the space being freed was
2248 		 * accounted for then (see softdep_releasefile()). If the
2249 		 * file is merely being truncated, then we account for it now.
2250 		 */
2251 		if ((ip->i_flag & IN_SPACECOUNTED) == 0) {
2252 			UFS_LOCK(ip->i_ump);
2253 			fs->fs_pendingblocks += datablocks;
2254 			UFS_UNLOCK(ip->i_ump);
2255 		}
2256 	}
2257 	if ((flags & IO_EXT) == 0) {
2258 		freeblks->fb_oldextsize = 0;
2259 	} else {
2260 		freeblks->fb_oldextsize = ip->i_din2->di_extsize;
2261 		ip->i_din2->di_extsize = 0;
2262 		freeblks->fb_chkcnt += extblocks;
2263 		for (i = 0; i < NXADDR; i++) {
2264 			freeblks->fb_eblks[i] = ip->i_din2->di_extb[i];
2265 			ip->i_din2->di_extb[i] = 0;
2266 		}
2267 	}
2268 	DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt);
2269 	/*
2270 	 * Push the zero'ed inode to to its disk buffer so that we are free
2271 	 * to delete its dependencies below. Once the dependencies are gone
2272 	 * the buffer can be safely released.
2273 	 */
2274 	if ((error = bread(ip->i_devvp,
2275 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
2276 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
2277 		brelse(bp);
2278 		softdep_error("softdep_setup_freeblocks", error);
2279 	}
2280 	if (ip->i_ump->um_fstype == UFS1)
2281 		*((struct ufs1_dinode *)bp->b_data +
2282 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
2283 	else
2284 		*((struct ufs2_dinode *)bp->b_data +
2285 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
2286 	/*
2287 	 * Find and eliminate any inode dependencies.
2288 	 */
2289 	ACQUIRE_LOCK(&lk);
2290 	(void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
2291 	if ((inodedep->id_state & IOSTARTED) != 0)
2292 		panic("softdep_setup_freeblocks: inode busy");
2293 	/*
2294 	 * Add the freeblks structure to the list of operations that
2295 	 * must await the zero'ed inode being written to disk. If we
2296 	 * still have a bitmap dependency (delay == 0), then the inode
2297 	 * has never been written to disk, so we can process the
2298 	 * freeblks below once we have deleted the dependencies.
2299 	 */
2300 	delay = (inodedep->id_state & DEPCOMPLETE);
2301 	if (delay)
2302 		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
2303 	/*
2304 	 * Because the file length has been truncated to zero, any
2305 	 * pending block allocation dependency structures associated
2306 	 * with this inode are obsolete and can simply be de-allocated.
2307 	 * We must first merge the two dependency lists to get rid of
2308 	 * any duplicate freefrag structures, then purge the merged list.
2309 	 * If we still have a bitmap dependency, then the inode has never
2310 	 * been written to disk, so we can free any fragments without delay.
2311 	 */
2312 	if (flags & IO_NORMAL) {
2313 		merge_inode_lists(&inodedep->id_newinoupdt,
2314 		    &inodedep->id_inoupdt);
2315 		while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
2316 			free_allocdirect(&inodedep->id_inoupdt, adp, delay);
2317 	}
2318 	if (flags & IO_EXT) {
2319 		merge_inode_lists(&inodedep->id_newextupdt,
2320 		    &inodedep->id_extupdt);
2321 		while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
2322 			free_allocdirect(&inodedep->id_extupdt, adp, delay);
2323 	}
2324 	FREE_LOCK(&lk);
2325 	bdwrite(bp);
2326 	/*
2327 	 * We must wait for any I/O in progress to finish so that
2328 	 * all potential buffers on the dirty list will be visible.
2329 	 * Once they are all there, walk the list and get rid of
2330 	 * any dependencies.
2331 	 */
2332 	vp = ITOV(ip);
2333 	bo = &vp->v_bufobj;
2334 	BO_LOCK(bo);
2335 	drain_output(vp);
2336 restart:
2337 	TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
2338 		if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
2339 		    ((flags & IO_NORMAL) == 0 &&
2340 		      (bp->b_xflags & BX_ALTDATA) == 0))
2341 			continue;
2342 		if ((bp = getdirtybuf(bp, BO_MTX(bo), MNT_WAIT)) == NULL)
2343 			goto restart;
2344 		BO_UNLOCK(bo);
2345 		ACQUIRE_LOCK(&lk);
2346 		(void) inodedep_lookup(mp, ip->i_number, 0, &inodedep);
2347 		deallocate_dependencies(bp, inodedep);
2348 		FREE_LOCK(&lk);
2349 		bp->b_flags |= B_INVAL | B_NOCACHE;
2350 		brelse(bp);
2351 		BO_LOCK(bo);
2352 		goto restart;
2353 	}
2354 	BO_UNLOCK(bo);
2355 	ACQUIRE_LOCK(&lk);
2356 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
2357 		(void) free_inodedep(inodedep);
2358 
2359 	if(delay) {
2360 		freeblks->fb_state |= DEPCOMPLETE;
2361 		/*
2362 		 * If the inode with zeroed block pointers is now on disk
2363 		 * we can start freeing blocks. Add freeblks to the worklist
2364 		 * instead of calling  handle_workitem_freeblocks directly as
2365 		 * it is more likely that additional IO is needed to complete
2366 		 * the request here than in the !delay case.
2367 		 */
2368 		if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
2369 			add_to_worklist(&freeblks->fb_list);
2370 	}
2371 
2372 	FREE_LOCK(&lk);
2373 	/*
2374 	 * If the inode has never been written to disk (delay == 0),
2375 	 * then we can process the freeblks now that we have deleted
2376 	 * the dependencies.
2377 	 */
2378 	if (!delay)
2379 		handle_workitem_freeblocks(freeblks, 0);
2380 }
2381 
2382 /*
2383  * Reclaim any dependency structures from a buffer that is about to
2384  * be reallocated to a new vnode. The buffer must be locked, thus,
2385  * no I/O completion operations can occur while we are manipulating
2386  * its associated dependencies. The mutex is held so that other I/O's
2387  * associated with related dependencies do not occur.
2388  */
2389 static void
2390 deallocate_dependencies(bp, inodedep)
2391 	struct buf *bp;
2392 	struct inodedep *inodedep;
2393 {
2394 	struct worklist *wk;
2395 	struct indirdep *indirdep;
2396 	struct allocindir *aip;
2397 	struct pagedep *pagedep;
2398 	struct dirrem *dirrem;
2399 	struct diradd *dap;
2400 	int i;
2401 
2402 	mtx_assert(&lk, MA_OWNED);
2403 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2404 		switch (wk->wk_type) {
2405 
2406 		case D_INDIRDEP:
2407 			indirdep = WK_INDIRDEP(wk);
2408 			/*
2409 			 * None of the indirect pointers will ever be visible,
2410 			 * so they can simply be tossed. GOINGAWAY ensures
2411 			 * that allocated pointers will be saved in the buffer
2412 			 * cache until they are freed. Note that they will
2413 			 * only be able to be found by their physical address
2414 			 * since the inode mapping the logical address will
2415 			 * be gone. The save buffer used for the safe copy
2416 			 * was allocated in setup_allocindir_phase2 using
2417 			 * the physical address so it could be used for this
2418 			 * purpose. Hence we swap the safe copy with the real
2419 			 * copy, allowing the safe copy to be freed and holding
2420 			 * on to the real copy for later use in indir_trunc.
2421 			 */
2422 			if (indirdep->ir_state & GOINGAWAY)
2423 				panic("deallocate_dependencies: already gone");
2424 			indirdep->ir_state |= GOINGAWAY;
2425 			VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1;
2426 			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
2427 				free_allocindir(aip, inodedep);
2428 			if (bp->b_lblkno >= 0 ||
2429 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno)
2430 				panic("deallocate_dependencies: not indir");
2431 			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2432 			    bp->b_bcount);
2433 			WORKLIST_REMOVE(wk);
2434 			WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
2435 			continue;
2436 
2437 		case D_PAGEDEP:
2438 			pagedep = WK_PAGEDEP(wk);
2439 			/*
2440 			 * None of the directory additions will ever be
2441 			 * visible, so they can simply be tossed.
2442 			 */
2443 			for (i = 0; i < DAHASHSZ; i++)
2444 				while ((dap =
2445 				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
2446 					free_diradd(dap);
2447 			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2448 				free_diradd(dap);
2449 			/*
2450 			 * Copy any directory remove dependencies to the list
2451 			 * to be processed after the zero'ed inode is written.
2452 			 * If the inode has already been written, then they
2453 			 * can be dumped directly onto the work list.
2454 			 */
2455 			LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2456 				LIST_REMOVE(dirrem, dm_next);
2457 				dirrem->dm_dirinum = pagedep->pd_ino;
2458 				if (inodedep == NULL ||
2459 				    (inodedep->id_state & ALLCOMPLETE) ==
2460 				     ALLCOMPLETE)
2461 					add_to_worklist(&dirrem->dm_list);
2462 				else
2463 					WORKLIST_INSERT(&inodedep->id_bufwait,
2464 					    &dirrem->dm_list);
2465 			}
2466 			if ((pagedep->pd_state & NEWBLOCK) != 0) {
2467 				LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list)
2468 					if (wk->wk_type == D_NEWDIRBLK &&
2469 					    WK_NEWDIRBLK(wk)->db_pagedep ==
2470 					      pagedep)
2471 						break;
2472 				if (wk != NULL) {
2473 					WORKLIST_REMOVE(wk);
2474 					free_newdirblk(WK_NEWDIRBLK(wk));
2475 				} else
2476 					panic("deallocate_dependencies: "
2477 					      "lost pagedep");
2478 			}
2479 			WORKLIST_REMOVE(&pagedep->pd_list);
2480 			LIST_REMOVE(pagedep, pd_hash);
2481 			WORKITEM_FREE(pagedep, D_PAGEDEP);
2482 			continue;
2483 
2484 		case D_ALLOCINDIR:
2485 			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2486 			continue;
2487 
2488 		case D_ALLOCDIRECT:
2489 		case D_INODEDEP:
2490 			panic("deallocate_dependencies: Unexpected type %s",
2491 			    TYPENAME(wk->wk_type));
2492 			/* NOTREACHED */
2493 
2494 		default:
2495 			panic("deallocate_dependencies: Unknown type %s",
2496 			    TYPENAME(wk->wk_type));
2497 			/* NOTREACHED */
2498 		}
2499 	}
2500 }
2501 
2502 /*
2503  * Free an allocdirect. Generate a new freefrag work request if appropriate.
2504  * This routine must be called with splbio interrupts blocked.
2505  */
2506 static void
2507 free_allocdirect(adphead, adp, delay)
2508 	struct allocdirectlst *adphead;
2509 	struct allocdirect *adp;
2510 	int delay;
2511 {
2512 	struct newdirblk *newdirblk;
2513 	struct worklist *wk;
2514 
2515 	mtx_assert(&lk, MA_OWNED);
2516 	if ((adp->ad_state & DEPCOMPLETE) == 0)
2517 		LIST_REMOVE(adp, ad_deps);
2518 	TAILQ_REMOVE(adphead, adp, ad_next);
2519 	if ((adp->ad_state & COMPLETE) == 0)
2520 		WORKLIST_REMOVE(&adp->ad_list);
2521 	if (adp->ad_freefrag != NULL) {
2522 		if (delay)
2523 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2524 			    &adp->ad_freefrag->ff_list);
2525 		else
2526 			add_to_worklist(&adp->ad_freefrag->ff_list);
2527 	}
2528 	if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) {
2529 		newdirblk = WK_NEWDIRBLK(wk);
2530 		WORKLIST_REMOVE(&newdirblk->db_list);
2531 		if (!LIST_EMPTY(&adp->ad_newdirblk))
2532 			panic("free_allocdirect: extra newdirblk");
2533 		if (delay)
2534 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2535 			    &newdirblk->db_list);
2536 		else
2537 			free_newdirblk(newdirblk);
2538 	}
2539 	WORKITEM_FREE(adp, D_ALLOCDIRECT);
2540 }
2541 
2542 /*
2543  * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
2544  * This routine must be called with splbio interrupts blocked.
2545  */
2546 static void
2547 free_newdirblk(newdirblk)
2548 	struct newdirblk *newdirblk;
2549 {
2550 	struct pagedep *pagedep;
2551 	struct diradd *dap;
2552 	int i;
2553 
2554 	mtx_assert(&lk, MA_OWNED);
2555 	/*
2556 	 * If the pagedep is still linked onto the directory buffer
2557 	 * dependency chain, then some of the entries on the
2558 	 * pd_pendinghd list may not be committed to disk yet. In
2559 	 * this case, we will simply clear the NEWBLOCK flag and
2560 	 * let the pd_pendinghd list be processed when the pagedep
2561 	 * is next written. If the pagedep is no longer on the buffer
2562 	 * dependency chain, then all the entries on the pd_pending
2563 	 * list are committed to disk and we can free them here.
2564 	 */
2565 	pagedep = newdirblk->db_pagedep;
2566 	pagedep->pd_state &= ~NEWBLOCK;
2567 	if ((pagedep->pd_state & ONWORKLIST) == 0)
2568 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
2569 			free_diradd(dap);
2570 	/*
2571 	 * If no dependencies remain, the pagedep will be freed.
2572 	 */
2573 	for (i = 0; i < DAHASHSZ; i++)
2574 		if (!LIST_EMPTY(&pagedep->pd_diraddhd[i]))
2575 			break;
2576 	if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) {
2577 		LIST_REMOVE(pagedep, pd_hash);
2578 		WORKITEM_FREE(pagedep, D_PAGEDEP);
2579 	}
2580 	WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2581 }
2582 
2583 /*
2584  * Prepare an inode to be freed. The actual free operation is not
2585  * done until the zero'ed inode has been written to disk.
2586  */
2587 void
2588 softdep_freefile(pvp, ino, mode)
2589 	struct vnode *pvp;
2590 	ino_t ino;
2591 	int mode;
2592 {
2593 	struct inode *ip = VTOI(pvp);
2594 	struct inodedep *inodedep;
2595 	struct freefile *freefile;
2596 
2597 	/*
2598 	 * This sets up the inode de-allocation dependency.
2599 	 */
2600 	freefile = malloc(sizeof(struct freefile),
2601 		M_FREEFILE, M_SOFTDEP_FLAGS);
2602 	workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount);
2603 	freefile->fx_mode = mode;
2604 	freefile->fx_oldinum = ino;
2605 	freefile->fx_devvp = ip->i_devvp;
2606 	if ((ip->i_flag & IN_SPACECOUNTED) == 0) {
2607 		UFS_LOCK(ip->i_ump);
2608 		ip->i_fs->fs_pendinginodes += 1;
2609 		UFS_UNLOCK(ip->i_ump);
2610 	}
2611 
2612 	/*
2613 	 * If the inodedep does not exist, then the zero'ed inode has
2614 	 * been written to disk. If the allocated inode has never been
2615 	 * written to disk, then the on-disk inode is zero'ed. In either
2616 	 * case we can free the file immediately.
2617 	 */
2618 	ACQUIRE_LOCK(&lk);
2619 	if (inodedep_lookup(pvp->v_mount, ino, 0, &inodedep) == 0 ||
2620 	    check_inode_unwritten(inodedep)) {
2621 		FREE_LOCK(&lk);
2622 		handle_workitem_freefile(freefile);
2623 		return;
2624 	}
2625 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2626 	FREE_LOCK(&lk);
2627 	if (ip->i_number == ino)
2628 		ip->i_flag |= IN_MODIFIED;
2629 }
2630 
2631 /*
2632  * Check to see if an inode has never been written to disk. If
2633  * so free the inodedep and return success, otherwise return failure.
2634  * This routine must be called with splbio interrupts blocked.
2635  *
2636  * If we still have a bitmap dependency, then the inode has never
2637  * been written to disk. Drop the dependency as it is no longer
2638  * necessary since the inode is being deallocated. We set the
2639  * ALLCOMPLETE flags since the bitmap now properly shows that the
2640  * inode is not allocated. Even if the inode is actively being
2641  * written, it has been rolled back to its zero'ed state, so we
2642  * are ensured that a zero inode is what is on the disk. For short
2643  * lived files, this change will usually result in removing all the
2644  * dependencies from the inode so that it can be freed immediately.
2645  */
2646 static int
2647 check_inode_unwritten(inodedep)
2648 	struct inodedep *inodedep;
2649 {
2650 
2651 	mtx_assert(&lk, MA_OWNED);
2652 	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2653 	    !LIST_EMPTY(&inodedep->id_pendinghd) ||
2654 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
2655 	    !LIST_EMPTY(&inodedep->id_inowait) ||
2656 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
2657 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
2658 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
2659 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
2660 	    inodedep->id_nlinkdelta != 0)
2661 		return (0);
2662 
2663 	/*
2664 	 * Another process might be in initiate_write_inodeblock_ufs[12]
2665 	 * trying to allocate memory without holding "Softdep Lock".
2666 	 */
2667 	if ((inodedep->id_state & IOSTARTED) != 0 &&
2668 	    inodedep->id_savedino1 == NULL)
2669 		return (0);
2670 
2671 	inodedep->id_state |= ALLCOMPLETE;
2672 	LIST_REMOVE(inodedep, id_deps);
2673 	inodedep->id_buf = NULL;
2674 	if (inodedep->id_state & ONWORKLIST)
2675 		WORKLIST_REMOVE(&inodedep->id_list);
2676 	if (inodedep->id_savedino1 != NULL) {
2677 		free(inodedep->id_savedino1, M_SAVEDINO);
2678 		inodedep->id_savedino1 = NULL;
2679 	}
2680 	if (free_inodedep(inodedep) == 0)
2681 		panic("check_inode_unwritten: busy inode");
2682 	return (1);
2683 }
2684 
2685 /*
2686  * Try to free an inodedep structure. Return 1 if it could be freed.
2687  */
2688 static int
2689 free_inodedep(inodedep)
2690 	struct inodedep *inodedep;
2691 {
2692 
2693 	mtx_assert(&lk, MA_OWNED);
2694 	if ((inodedep->id_state & ONWORKLIST) != 0 ||
2695 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2696 	    !LIST_EMPTY(&inodedep->id_pendinghd) ||
2697 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
2698 	    !LIST_EMPTY(&inodedep->id_inowait) ||
2699 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
2700 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
2701 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
2702 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
2703 	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL)
2704 		return (0);
2705 	LIST_REMOVE(inodedep, id_hash);
2706 	WORKITEM_FREE(inodedep, D_INODEDEP);
2707 	num_inodedep -= 1;
2708 	return (1);
2709 }
2710 
2711 /*
2712  * This workitem routine performs the block de-allocation.
2713  * The workitem is added to the pending list after the updated
2714  * inode block has been written to disk.  As mentioned above,
2715  * checks regarding the number of blocks de-allocated (compared
2716  * to the number of blocks allocated for the file) are also
2717  * performed in this function.
2718  */
2719 static void
2720 handle_workitem_freeblocks(freeblks, flags)
2721 	struct freeblks *freeblks;
2722 	int flags;
2723 {
2724 	struct inode *ip;
2725 	struct vnode *vp;
2726 	struct fs *fs;
2727 	struct ufsmount *ump;
2728 	int i, nblocks, level, bsize;
2729 	ufs2_daddr_t bn, blocksreleased = 0;
2730 	int error, allerror = 0;
2731 	ufs_lbn_t baselbns[NIADDR], tmpval;
2732 	int fs_pendingblocks;
2733 
2734 	ump = VFSTOUFS(freeblks->fb_list.wk_mp);
2735 	fs = ump->um_fs;
2736 	fs_pendingblocks = 0;
2737 	tmpval = 1;
2738 	baselbns[0] = NDADDR;
2739 	for (i = 1; i < NIADDR; i++) {
2740 		tmpval *= NINDIR(fs);
2741 		baselbns[i] = baselbns[i - 1] + tmpval;
2742 	}
2743 	nblocks = btodb(fs->fs_bsize);
2744 	blocksreleased = 0;
2745 	/*
2746 	 * Release all extended attribute blocks or frags.
2747 	 */
2748 	if (freeblks->fb_oldextsize > 0) {
2749 		for (i = (NXADDR - 1); i >= 0; i--) {
2750 			if ((bn = freeblks->fb_eblks[i]) == 0)
2751 				continue;
2752 			bsize = sblksize(fs, freeblks->fb_oldextsize, i);
2753 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize,
2754 			    freeblks->fb_previousinum);
2755 			blocksreleased += btodb(bsize);
2756 		}
2757 	}
2758 	/*
2759 	 * Release all data blocks or frags.
2760 	 */
2761 	if (freeblks->fb_oldsize > 0) {
2762 		/*
2763 		 * Indirect blocks first.
2764 		 */
2765 		for (level = (NIADDR - 1); level >= 0; level--) {
2766 			if ((bn = freeblks->fb_iblks[level]) == 0)
2767 				continue;
2768 			if ((error = indir_trunc(freeblks, fsbtodb(fs, bn),
2769 			    level, baselbns[level], &blocksreleased)) != 0)
2770 				allerror = error;
2771 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn,
2772 			    fs->fs_bsize, freeblks->fb_previousinum);
2773 			fs_pendingblocks += nblocks;
2774 			blocksreleased += nblocks;
2775 		}
2776 		/*
2777 		 * All direct blocks or frags.
2778 		 */
2779 		for (i = (NDADDR - 1); i >= 0; i--) {
2780 			if ((bn = freeblks->fb_dblks[i]) == 0)
2781 				continue;
2782 			bsize = sblksize(fs, freeblks->fb_oldsize, i);
2783 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize,
2784 			    freeblks->fb_previousinum);
2785 			fs_pendingblocks += btodb(bsize);
2786 			blocksreleased += btodb(bsize);
2787 		}
2788 	}
2789 	UFS_LOCK(ump);
2790 	fs->fs_pendingblocks -= fs_pendingblocks;
2791 	UFS_UNLOCK(ump);
2792 	/*
2793 	 * If we still have not finished background cleanup, then check
2794 	 * to see if the block count needs to be adjusted.
2795 	 */
2796 	if (freeblks->fb_chkcnt != blocksreleased &&
2797 	    (fs->fs_flags & FS_UNCLEAN) != 0 &&
2798 	    ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_previousinum,
2799 		(flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ)
2800 	    == 0) {
2801 		ip = VTOI(vp);
2802 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + \
2803 		    freeblks->fb_chkcnt - blocksreleased);
2804 		ip->i_flag |= IN_CHANGE;
2805 		vput(vp);
2806 	}
2807 
2808 #ifdef INVARIANTS
2809 	if (freeblks->fb_chkcnt != blocksreleased &&
2810 	    ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0))
2811 		printf("handle_workitem_freeblocks: block count\n");
2812 	if (allerror)
2813 		softdep_error("handle_workitem_freeblks", allerror);
2814 #endif /* INVARIANTS */
2815 
2816 	ACQUIRE_LOCK(&lk);
2817 	WORKITEM_FREE(freeblks, D_FREEBLKS);
2818 	FREE_LOCK(&lk);
2819 }
2820 
2821 /*
2822  * Release blocks associated with the inode ip and stored in the indirect
2823  * block dbn. If level is greater than SINGLE, the block is an indirect block
2824  * and recursive calls to indirtrunc must be used to cleanse other indirect
2825  * blocks.
2826  */
2827 static int
2828 indir_trunc(freeblks, dbn, level, lbn, countp)
2829 	struct freeblks *freeblks;
2830 	ufs2_daddr_t dbn;
2831 	int level;
2832 	ufs_lbn_t lbn;
2833 	ufs2_daddr_t *countp;
2834 {
2835 	struct buf *bp;
2836 	struct fs *fs;
2837 	struct worklist *wk;
2838 	struct indirdep *indirdep;
2839 	struct ufsmount *ump;
2840 	ufs1_daddr_t *bap1 = 0;
2841 	ufs2_daddr_t nb, *bap2 = 0;
2842 	ufs_lbn_t lbnadd;
2843 	int i, nblocks, ufs1fmt;
2844 	int error, allerror = 0;
2845 	int fs_pendingblocks;
2846 
2847 	ump = VFSTOUFS(freeblks->fb_list.wk_mp);
2848 	fs = ump->um_fs;
2849 	fs_pendingblocks = 0;
2850 	lbnadd = 1;
2851 	for (i = level; i > 0; i--)
2852 		lbnadd *= NINDIR(fs);
2853 	/*
2854 	 * Get buffer of block pointers to be freed. This routine is not
2855 	 * called until the zero'ed inode has been written, so it is safe
2856 	 * to free blocks as they are encountered. Because the inode has
2857 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
2858 	 * have to use the on-disk address and the block device for the
2859 	 * filesystem to look them up. If the file was deleted before its
2860 	 * indirect blocks were all written to disk, the routine that set
2861 	 * us up (deallocate_dependencies) will have arranged to leave
2862 	 * a complete copy of the indirect block in memory for our use.
2863 	 * Otherwise we have to read the blocks in from the disk.
2864 	 */
2865 #ifdef notyet
2866 	bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0,
2867 	    GB_NOCREAT);
2868 #else
2869 	bp = incore(&freeblks->fb_devvp->v_bufobj, dbn);
2870 #endif
2871 	ACQUIRE_LOCK(&lk);
2872 	if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2873 		if (wk->wk_type != D_INDIRDEP ||
2874 		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2875 		    (indirdep->ir_state & GOINGAWAY) == 0)
2876 			panic("indir_trunc: lost indirdep");
2877 		WORKLIST_REMOVE(wk);
2878 		WORKITEM_FREE(indirdep, D_INDIRDEP);
2879 		if (!LIST_EMPTY(&bp->b_dep))
2880 			panic("indir_trunc: dangling dep");
2881 		ump->um_numindirdeps -= 1;
2882 		FREE_LOCK(&lk);
2883 	} else {
2884 #ifdef notyet
2885 		if (bp)
2886 			brelse(bp);
2887 #endif
2888 		FREE_LOCK(&lk);
2889 		error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
2890 		    NOCRED, &bp);
2891 		if (error) {
2892 			brelse(bp);
2893 			return (error);
2894 		}
2895 	}
2896 	/*
2897 	 * Recursively free indirect blocks.
2898 	 */
2899 	if (ump->um_fstype == UFS1) {
2900 		ufs1fmt = 1;
2901 		bap1 = (ufs1_daddr_t *)bp->b_data;
2902 	} else {
2903 		ufs1fmt = 0;
2904 		bap2 = (ufs2_daddr_t *)bp->b_data;
2905 	}
2906 	nblocks = btodb(fs->fs_bsize);
2907 	for (i = NINDIR(fs) - 1; i >= 0; i--) {
2908 		if (ufs1fmt)
2909 			nb = bap1[i];
2910 		else
2911 			nb = bap2[i];
2912 		if (nb == 0)
2913 			continue;
2914 		if (level != 0) {
2915 			if ((error = indir_trunc(freeblks, fsbtodb(fs, nb),
2916 			     level - 1, lbn + (i * lbnadd), countp)) != 0)
2917 				allerror = error;
2918 		}
2919 		ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, fs->fs_bsize,
2920 		    freeblks->fb_previousinum);
2921 		fs_pendingblocks += nblocks;
2922 		*countp += nblocks;
2923 	}
2924 	UFS_LOCK(ump);
2925 	fs->fs_pendingblocks -= fs_pendingblocks;
2926 	UFS_UNLOCK(ump);
2927 	bp->b_flags |= B_INVAL | B_NOCACHE;
2928 	brelse(bp);
2929 	return (allerror);
2930 }
2931 
2932 /*
2933  * Free an allocindir.
2934  * This routine must be called with splbio interrupts blocked.
2935  */
2936 static void
2937 free_allocindir(aip, inodedep)
2938 	struct allocindir *aip;
2939 	struct inodedep *inodedep;
2940 {
2941 	struct freefrag *freefrag;
2942 
2943 	mtx_assert(&lk, MA_OWNED);
2944 	if ((aip->ai_state & DEPCOMPLETE) == 0)
2945 		LIST_REMOVE(aip, ai_deps);
2946 	if (aip->ai_state & ONWORKLIST)
2947 		WORKLIST_REMOVE(&aip->ai_list);
2948 	LIST_REMOVE(aip, ai_next);
2949 	if ((freefrag = aip->ai_freefrag) != NULL) {
2950 		if (inodedep == NULL)
2951 			add_to_worklist(&freefrag->ff_list);
2952 		else
2953 			WORKLIST_INSERT(&inodedep->id_bufwait,
2954 			    &freefrag->ff_list);
2955 	}
2956 	WORKITEM_FREE(aip, D_ALLOCINDIR);
2957 }
2958 
2959 /*
2960  * Directory entry addition dependencies.
2961  *
2962  * When adding a new directory entry, the inode (with its incremented link
2963  * count) must be written to disk before the directory entry's pointer to it.
2964  * Also, if the inode is newly allocated, the corresponding freemap must be
2965  * updated (on disk) before the directory entry's pointer. These requirements
2966  * are met via undo/redo on the directory entry's pointer, which consists
2967  * simply of the inode number.
2968  *
2969  * As directory entries are added and deleted, the free space within a
2970  * directory block can become fragmented.  The ufs filesystem will compact
2971  * a fragmented directory block to make space for a new entry. When this
2972  * occurs, the offsets of previously added entries change. Any "diradd"
2973  * dependency structures corresponding to these entries must be updated with
2974  * the new offsets.
2975  */
2976 
2977 /*
2978  * This routine is called after the in-memory inode's link
2979  * count has been incremented, but before the directory entry's
2980  * pointer to the inode has been set.
2981  */
2982 int
2983 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
2984 	struct buf *bp;		/* buffer containing directory block */
2985 	struct inode *dp;	/* inode for directory */
2986 	off_t diroffset;	/* offset of new entry in directory */
2987 	ino_t newinum;		/* inode referenced by new directory entry */
2988 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
2989 	int isnewblk;		/* entry is in a newly allocated block */
2990 {
2991 	int offset;		/* offset of new entry within directory block */
2992 	ufs_lbn_t lbn;		/* block in directory containing new entry */
2993 	struct fs *fs;
2994 	struct diradd *dap;
2995 	struct allocdirect *adp;
2996 	struct pagedep *pagedep;
2997 	struct inodedep *inodedep;
2998 	struct newdirblk *newdirblk = 0;
2999 	struct mkdir *mkdir1, *mkdir2;
3000 	struct mount *mp;
3001 
3002 	/*
3003 	 * Whiteouts have no dependencies.
3004 	 */
3005 	if (newinum == WINO) {
3006 		if (newdirbp != NULL)
3007 			bdwrite(newdirbp);
3008 		return (0);
3009 	}
3010 	mp = UFSTOVFS(dp->i_ump);
3011 	fs = dp->i_fs;
3012 	lbn = lblkno(fs, diroffset);
3013 	offset = blkoff(fs, diroffset);
3014 	dap = malloc(sizeof(struct diradd), M_DIRADD,
3015 		M_SOFTDEP_FLAGS|M_ZERO);
3016 	workitem_alloc(&dap->da_list, D_DIRADD, mp);
3017 	dap->da_offset = offset;
3018 	dap->da_newinum = newinum;
3019 	dap->da_state = ATTACHED;
3020 	if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) {
3021 		newdirblk = malloc(sizeof(struct newdirblk),
3022 		    M_NEWDIRBLK, M_SOFTDEP_FLAGS);
3023 		workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
3024 	}
3025 	if (newdirbp == NULL) {
3026 		dap->da_state |= DEPCOMPLETE;
3027 		ACQUIRE_LOCK(&lk);
3028 	} else {
3029 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
3030 		mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR,
3031 		    M_SOFTDEP_FLAGS);
3032 		workitem_alloc(&mkdir1->md_list, D_MKDIR, mp);
3033 		mkdir1->md_state = MKDIR_BODY;
3034 		mkdir1->md_diradd = dap;
3035 		mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR,
3036 		    M_SOFTDEP_FLAGS);
3037 		workitem_alloc(&mkdir2->md_list, D_MKDIR, mp);
3038 		mkdir2->md_state = MKDIR_PARENT;
3039 		mkdir2->md_diradd = dap;
3040 		/*
3041 		 * Dependency on "." and ".." being written to disk.
3042 		 */
3043 		mkdir1->md_buf = newdirbp;
3044 		ACQUIRE_LOCK(&lk);
3045 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
3046 		WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
3047 		FREE_LOCK(&lk);
3048 		bdwrite(newdirbp);
3049 		/*
3050 		 * Dependency on link count increase for parent directory
3051 		 */
3052 		ACQUIRE_LOCK(&lk);
3053 		if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0
3054 		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3055 			dap->da_state &= ~MKDIR_PARENT;
3056 			WORKITEM_FREE(mkdir2, D_MKDIR);
3057 		} else {
3058 			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
3059 			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
3060 		}
3061 	}
3062 	/*
3063 	 * Link into parent directory pagedep to await its being written.
3064 	 */
3065 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3066 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3067 	dap->da_pagedep = pagedep;
3068 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
3069 	    da_pdlist);
3070 	/*
3071 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3072 	 * is not yet written. If it is written, do the post-inode write
3073 	 * processing to put it on the id_pendinghd list.
3074 	 */
3075 	(void) inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
3076 	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
3077 		diradd_inode_written(dap, inodedep);
3078 	else
3079 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3080 	if (isnewblk) {
3081 		/*
3082 		 * Directories growing into indirect blocks are rare
3083 		 * enough and the frequency of new block allocation
3084 		 * in those cases even more rare, that we choose not
3085 		 * to bother tracking them. Rather we simply force the
3086 		 * new directory entry to disk.
3087 		 */
3088 		if (lbn >= NDADDR) {
3089 			FREE_LOCK(&lk);
3090 			/*
3091 			 * We only have a new allocation when at the
3092 			 * beginning of a new block, not when we are
3093 			 * expanding into an existing block.
3094 			 */
3095 			if (blkoff(fs, diroffset) == 0)
3096 				return (1);
3097 			return (0);
3098 		}
3099 		/*
3100 		 * We only have a new allocation when at the beginning
3101 		 * of a new fragment, not when we are expanding into an
3102 		 * existing fragment. Also, there is nothing to do if we
3103 		 * are already tracking this block.
3104 		 */
3105 		if (fragoff(fs, diroffset) != 0) {
3106 			FREE_LOCK(&lk);
3107 			return (0);
3108 		}
3109 		if ((pagedep->pd_state & NEWBLOCK) != 0) {
3110 			WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
3111 			FREE_LOCK(&lk);
3112 			return (0);
3113 		}
3114 		/*
3115 		 * Find our associated allocdirect and have it track us.
3116 		 */
3117 		if (inodedep_lookup(mp, dp->i_number, 0, &inodedep) == 0)
3118 			panic("softdep_setup_directory_add: lost inodedep");
3119 		adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst);
3120 		if (adp == NULL || adp->ad_lbn != lbn)
3121 			panic("softdep_setup_directory_add: lost entry");
3122 		pagedep->pd_state |= NEWBLOCK;
3123 		newdirblk->db_pagedep = pagedep;
3124 		WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list);
3125 	}
3126 	FREE_LOCK(&lk);
3127 	return (0);
3128 }
3129 
3130 /*
3131  * This procedure is called to change the offset of a directory
3132  * entry when compacting a directory block which must be owned
3133  * exclusively by the caller. Note that the actual entry movement
3134  * must be done in this procedure to ensure that no I/O completions
3135  * occur while the move is in progress.
3136  */
3137 void
3138 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
3139 	struct inode *dp;	/* inode for directory */
3140 	caddr_t base;		/* address of dp->i_offset */
3141 	caddr_t oldloc;		/* address of old directory location */
3142 	caddr_t newloc;		/* address of new directory location */
3143 	int entrysize;		/* size of directory entry */
3144 {
3145 	int offset, oldoffset, newoffset;
3146 	struct pagedep *pagedep;
3147 	struct diradd *dap;
3148 	ufs_lbn_t lbn;
3149 
3150 	ACQUIRE_LOCK(&lk);
3151 	lbn = lblkno(dp->i_fs, dp->i_offset);
3152 	offset = blkoff(dp->i_fs, dp->i_offset);
3153 	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
3154 		goto done;
3155 	oldoffset = offset + (oldloc - base);
3156 	newoffset = offset + (newloc - base);
3157 
3158 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
3159 		if (dap->da_offset != oldoffset)
3160 			continue;
3161 		dap->da_offset = newoffset;
3162 		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
3163 			break;
3164 		LIST_REMOVE(dap, da_pdlist);
3165 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
3166 		    dap, da_pdlist);
3167 		break;
3168 	}
3169 	if (dap == NULL) {
3170 
3171 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
3172 			if (dap->da_offset == oldoffset) {
3173 				dap->da_offset = newoffset;
3174 				break;
3175 			}
3176 		}
3177 	}
3178 done:
3179 	bcopy(oldloc, newloc, entrysize);
3180 	FREE_LOCK(&lk);
3181 }
3182 
3183 /*
3184  * Free a diradd dependency structure. This routine must be called
3185  * with splbio interrupts blocked.
3186  */
3187 static void
3188 free_diradd(dap)
3189 	struct diradd *dap;
3190 {
3191 	struct dirrem *dirrem;
3192 	struct pagedep *pagedep;
3193 	struct inodedep *inodedep;
3194 	struct mkdir *mkdir, *nextmd;
3195 
3196 	mtx_assert(&lk, MA_OWNED);
3197 	WORKLIST_REMOVE(&dap->da_list);
3198 	LIST_REMOVE(dap, da_pdlist);
3199 	if ((dap->da_state & DIRCHG) == 0) {
3200 		pagedep = dap->da_pagedep;
3201 	} else {
3202 		dirrem = dap->da_previous;
3203 		pagedep = dirrem->dm_pagedep;
3204 		dirrem->dm_dirinum = pagedep->pd_ino;
3205 		add_to_worklist(&dirrem->dm_list);
3206 	}
3207 	if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum,
3208 	    0, &inodedep) != 0)
3209 		(void) free_inodedep(inodedep);
3210 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
3211 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
3212 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
3213 			if (mkdir->md_diradd != dap)
3214 				continue;
3215 			dap->da_state &= ~mkdir->md_state;
3216 			WORKLIST_REMOVE(&mkdir->md_list);
3217 			LIST_REMOVE(mkdir, md_mkdirs);
3218 			WORKITEM_FREE(mkdir, D_MKDIR);
3219 		}
3220 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
3221 			panic("free_diradd: unfound ref");
3222 	}
3223 	WORKITEM_FREE(dap, D_DIRADD);
3224 }
3225 
3226 /*
3227  * Directory entry removal dependencies.
3228  *
3229  * When removing a directory entry, the entry's inode pointer must be
3230  * zero'ed on disk before the corresponding inode's link count is decremented
3231  * (possibly freeing the inode for re-use). This dependency is handled by
3232  * updating the directory entry but delaying the inode count reduction until
3233  * after the directory block has been written to disk. After this point, the
3234  * inode count can be decremented whenever it is convenient.
3235  */
3236 
3237 /*
3238  * This routine should be called immediately after removing
3239  * a directory entry.  The inode's link count should not be
3240  * decremented by the calling procedure -- the soft updates
3241  * code will do this task when it is safe.
3242  */
3243 void
3244 softdep_setup_remove(bp, dp, ip, isrmdir)
3245 	struct buf *bp;		/* buffer containing directory block */
3246 	struct inode *dp;	/* inode for the directory being modified */
3247 	struct inode *ip;	/* inode for directory entry being removed */
3248 	int isrmdir;		/* indicates if doing RMDIR */
3249 {
3250 	struct dirrem *dirrem, *prevdirrem;
3251 
3252 	/*
3253 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
3254 	 */
3255 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3256 
3257 	/*
3258 	 * If the COMPLETE flag is clear, then there were no active
3259 	 * entries and we want to roll back to a zeroed entry until
3260 	 * the new inode is committed to disk. If the COMPLETE flag is
3261 	 * set then we have deleted an entry that never made it to
3262 	 * disk. If the entry we deleted resulted from a name change,
3263 	 * then the old name still resides on disk. We cannot delete
3264 	 * its inode (returned to us in prevdirrem) until the zeroed
3265 	 * directory entry gets to disk. The new inode has never been
3266 	 * referenced on the disk, so can be deleted immediately.
3267 	 */
3268 	if ((dirrem->dm_state & COMPLETE) == 0) {
3269 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
3270 		    dm_next);
3271 		FREE_LOCK(&lk);
3272 	} else {
3273 		if (prevdirrem != NULL)
3274 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
3275 			    prevdirrem, dm_next);
3276 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
3277 		FREE_LOCK(&lk);
3278 		handle_workitem_remove(dirrem, NULL);
3279 	}
3280 }
3281 
3282 /*
3283  * Allocate a new dirrem if appropriate and return it along with
3284  * its associated pagedep. Called without a lock, returns with lock.
3285  */
3286 static long num_dirrem;		/* number of dirrem allocated */
3287 static struct dirrem *
3288 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
3289 	struct buf *bp;		/* buffer containing directory block */
3290 	struct inode *dp;	/* inode for the directory being modified */
3291 	struct inode *ip;	/* inode for directory entry being removed */
3292 	int isrmdir;		/* indicates if doing RMDIR */
3293 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
3294 {
3295 	int offset;
3296 	ufs_lbn_t lbn;
3297 	struct diradd *dap;
3298 	struct dirrem *dirrem;
3299 	struct pagedep *pagedep;
3300 
3301 	/*
3302 	 * Whiteouts have no deletion dependencies.
3303 	 */
3304 	if (ip == NULL)
3305 		panic("newdirrem: whiteout");
3306 	/*
3307 	 * If we are over our limit, try to improve the situation.
3308 	 * Limiting the number of dirrem structures will also limit
3309 	 * the number of freefile and freeblks structures.
3310 	 */
3311 	ACQUIRE_LOCK(&lk);
3312 	if (!(ip->i_flags & SF_SNAPSHOT) && num_dirrem > max_softdeps / 2)
3313 		(void) request_cleanup(ITOV(dp)->v_mount, FLUSH_REMOVE);
3314 	num_dirrem += 1;
3315 	FREE_LOCK(&lk);
3316 	dirrem = malloc(sizeof(struct dirrem),
3317 		M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO);
3318 	workitem_alloc(&dirrem->dm_list, D_DIRREM, ITOV(dp)->v_mount);
3319 	dirrem->dm_state = isrmdir ? RMDIR : 0;
3320 	dirrem->dm_oldinum = ip->i_number;
3321 	*prevdirremp = NULL;
3322 
3323 	ACQUIRE_LOCK(&lk);
3324 	lbn = lblkno(dp->i_fs, dp->i_offset);
3325 	offset = blkoff(dp->i_fs, dp->i_offset);
3326 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3327 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3328 	dirrem->dm_pagedep = pagedep;
3329 	/*
3330 	 * Check for a diradd dependency for the same directory entry.
3331 	 * If present, then both dependencies become obsolete and can
3332 	 * be de-allocated. Check for an entry on both the pd_dirraddhd
3333 	 * list and the pd_pendinghd list.
3334 	 */
3335 
3336 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
3337 		if (dap->da_offset == offset)
3338 			break;
3339 	if (dap == NULL) {
3340 
3341 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
3342 			if (dap->da_offset == offset)
3343 				break;
3344 		if (dap == NULL)
3345 			return (dirrem);
3346 	}
3347 	/*
3348 	 * Must be ATTACHED at this point.
3349 	 */
3350 	if ((dap->da_state & ATTACHED) == 0)
3351 		panic("newdirrem: not ATTACHED");
3352 	if (dap->da_newinum != ip->i_number)
3353 		panic("newdirrem: inum %d should be %d",
3354 		    ip->i_number, dap->da_newinum);
3355 	/*
3356 	 * If we are deleting a changed name that never made it to disk,
3357 	 * then return the dirrem describing the previous inode (which
3358 	 * represents the inode currently referenced from this entry on disk).
3359 	 */
3360 	if ((dap->da_state & DIRCHG) != 0) {
3361 		*prevdirremp = dap->da_previous;
3362 		dap->da_state &= ~DIRCHG;
3363 		dap->da_pagedep = pagedep;
3364 	}
3365 	/*
3366 	 * We are deleting an entry that never made it to disk.
3367 	 * Mark it COMPLETE so we can delete its inode immediately.
3368 	 */
3369 	dirrem->dm_state |= COMPLETE;
3370 	free_diradd(dap);
3371 	return (dirrem);
3372 }
3373 
3374 /*
3375  * Directory entry change dependencies.
3376  *
3377  * Changing an existing directory entry requires that an add operation
3378  * be completed first followed by a deletion. The semantics for the addition
3379  * are identical to the description of adding a new entry above except
3380  * that the rollback is to the old inode number rather than zero. Once
3381  * the addition dependency is completed, the removal is done as described
3382  * in the removal routine above.
3383  */
3384 
3385 /*
3386  * This routine should be called immediately after changing
3387  * a directory entry.  The inode's link count should not be
3388  * decremented by the calling procedure -- the soft updates
3389  * code will perform this task when it is safe.
3390  */
3391 void
3392 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
3393 	struct buf *bp;		/* buffer containing directory block */
3394 	struct inode *dp;	/* inode for the directory being modified */
3395 	struct inode *ip;	/* inode for directory entry being removed */
3396 	ino_t newinum;		/* new inode number for changed entry */
3397 	int isrmdir;		/* indicates if doing RMDIR */
3398 {
3399 	int offset;
3400 	struct diradd *dap = NULL;
3401 	struct dirrem *dirrem, *prevdirrem;
3402 	struct pagedep *pagedep;
3403 	struct inodedep *inodedep;
3404 	struct mount *mp;
3405 
3406 	offset = blkoff(dp->i_fs, dp->i_offset);
3407 	mp = UFSTOVFS(dp->i_ump);
3408 
3409 	/*
3410 	 * Whiteouts do not need diradd dependencies.
3411 	 */
3412 	if (newinum != WINO) {
3413 		dap = malloc(sizeof(struct diradd),
3414 		    M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
3415 		workitem_alloc(&dap->da_list, D_DIRADD, mp);
3416 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
3417 		dap->da_offset = offset;
3418 		dap->da_newinum = newinum;
3419 	}
3420 
3421 	/*
3422 	 * Allocate a new dirrem and ACQUIRE_LOCK.
3423 	 */
3424 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3425 	pagedep = dirrem->dm_pagedep;
3426 	/*
3427 	 * The possible values for isrmdir:
3428 	 *	0 - non-directory file rename
3429 	 *	1 - directory rename within same directory
3430 	 *   inum - directory rename to new directory of given inode number
3431 	 * When renaming to a new directory, we are both deleting and
3432 	 * creating a new directory entry, so the link count on the new
3433 	 * directory should not change. Thus we do not need the followup
3434 	 * dirrem which is usually done in handle_workitem_remove. We set
3435 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
3436 	 * followup dirrem.
3437 	 */
3438 	if (isrmdir > 1)
3439 		dirrem->dm_state |= DIRCHG;
3440 
3441 	/*
3442 	 * Whiteouts have no additional dependencies,
3443 	 * so just put the dirrem on the correct list.
3444 	 */
3445 	if (newinum == WINO) {
3446 		if ((dirrem->dm_state & COMPLETE) == 0) {
3447 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
3448 			    dm_next);
3449 		} else {
3450 			dirrem->dm_dirinum = pagedep->pd_ino;
3451 			add_to_worklist(&dirrem->dm_list);
3452 		}
3453 		FREE_LOCK(&lk);
3454 		return;
3455 	}
3456 
3457 	/*
3458 	 * If the COMPLETE flag is clear, then there were no active
3459 	 * entries and we want to roll back to the previous inode until
3460 	 * the new inode is committed to disk. If the COMPLETE flag is
3461 	 * set, then we have deleted an entry that never made it to disk.
3462 	 * If the entry we deleted resulted from a name change, then the old
3463 	 * inode reference still resides on disk. Any rollback that we do
3464 	 * needs to be to that old inode (returned to us in prevdirrem). If
3465 	 * the entry we deleted resulted from a create, then there is
3466 	 * no entry on the disk, so we want to roll back to zero rather
3467 	 * than the uncommitted inode. In either of the COMPLETE cases we
3468 	 * want to immediately free the unwritten and unreferenced inode.
3469 	 */
3470 	if ((dirrem->dm_state & COMPLETE) == 0) {
3471 		dap->da_previous = dirrem;
3472 	} else {
3473 		if (prevdirrem != NULL) {
3474 			dap->da_previous = prevdirrem;
3475 		} else {
3476 			dap->da_state &= ~DIRCHG;
3477 			dap->da_pagedep = pagedep;
3478 		}
3479 		dirrem->dm_dirinum = pagedep->pd_ino;
3480 		add_to_worklist(&dirrem->dm_list);
3481 	}
3482 	/*
3483 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3484 	 * is not yet written. If it is written, do the post-inode write
3485 	 * processing to put it on the id_pendinghd list.
3486 	 */
3487 	if (inodedep_lookup(mp, newinum, DEPALLOC, &inodedep) == 0 ||
3488 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3489 		dap->da_state |= COMPLETE;
3490 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3491 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3492 	} else {
3493 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
3494 		    dap, da_pdlist);
3495 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3496 	}
3497 	FREE_LOCK(&lk);
3498 }
3499 
3500 /*
3501  * Called whenever the link count on an inode is changed.
3502  * It creates an inode dependency so that the new reference(s)
3503  * to the inode cannot be committed to disk until the updated
3504  * inode has been written.
3505  */
3506 void
3507 softdep_change_linkcnt(ip)
3508 	struct inode *ip;	/* the inode with the increased link count */
3509 {
3510 	struct inodedep *inodedep;
3511 
3512 	ACQUIRE_LOCK(&lk);
3513 	(void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number,
3514 	    DEPALLOC, &inodedep);
3515 	if (ip->i_nlink < ip->i_effnlink)
3516 		panic("softdep_change_linkcnt: bad delta");
3517 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3518 	FREE_LOCK(&lk);
3519 }
3520 
3521 /*
3522  * Called when the effective link count and the reference count
3523  * on an inode drops to zero. At this point there are no names
3524  * referencing the file in the filesystem and no active file
3525  * references. The space associated with the file will be freed
3526  * as soon as the necessary soft dependencies are cleared.
3527  */
3528 void
3529 softdep_releasefile(ip)
3530 	struct inode *ip;	/* inode with the zero effective link count */
3531 {
3532 	struct inodedep *inodedep;
3533 	struct fs *fs;
3534 	int extblocks;
3535 
3536 	if (ip->i_effnlink > 0)
3537 		panic("softdep_releasefile: file still referenced");
3538 	/*
3539 	 * We may be called several times as the on-disk link count
3540 	 * drops to zero. We only want to account for the space once.
3541 	 */
3542 	if (ip->i_flag & IN_SPACECOUNTED)
3543 		return;
3544 	/*
3545 	 * We have to deactivate a snapshot otherwise copyonwrites may
3546 	 * add blocks and the cleanup may remove blocks after we have
3547 	 * tried to account for them.
3548 	 */
3549 	if ((ip->i_flags & SF_SNAPSHOT) != 0)
3550 		ffs_snapremove(ITOV(ip));
3551 	/*
3552 	 * If we are tracking an nlinkdelta, we have to also remember
3553 	 * whether we accounted for the freed space yet.
3554 	 */
3555 	ACQUIRE_LOCK(&lk);
3556 	if ((inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, &inodedep)))
3557 		inodedep->id_state |= SPACECOUNTED;
3558 	FREE_LOCK(&lk);
3559 	fs = ip->i_fs;
3560 	extblocks = 0;
3561 	if (fs->fs_magic == FS_UFS2_MAGIC)
3562 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
3563 	UFS_LOCK(ip->i_ump);
3564 	ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks;
3565 	ip->i_fs->fs_pendinginodes += 1;
3566 	UFS_UNLOCK(ip->i_ump);
3567 	ip->i_flag |= IN_SPACECOUNTED;
3568 }
3569 
3570 /*
3571  * This workitem decrements the inode's link count.
3572  * If the link count reaches zero, the file is removed.
3573  */
3574 static void
3575 handle_workitem_remove(dirrem, xp)
3576 	struct dirrem *dirrem;
3577 	struct vnode *xp;
3578 {
3579 	struct thread *td = curthread;
3580 	struct inodedep *inodedep;
3581 	struct vnode *vp;
3582 	struct inode *ip;
3583 	ino_t oldinum;
3584 	int error;
3585 
3586 	if ((vp = xp) == NULL &&
3587 	    (error = ffs_vgetf(dirrem->dm_list.wk_mp,
3588 		    dirrem->dm_oldinum, LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ)) != 0) {
3589 		softdep_error("handle_workitem_remove: vget", error);
3590 		return;
3591 	}
3592 	ip = VTOI(vp);
3593 	ACQUIRE_LOCK(&lk);
3594 	if ((inodedep_lookup(dirrem->dm_list.wk_mp,
3595 	    dirrem->dm_oldinum, 0, &inodedep)) == 0)
3596 		panic("handle_workitem_remove: lost inodedep");
3597 	/*
3598 	 * Normal file deletion.
3599 	 */
3600 	if ((dirrem->dm_state & RMDIR) == 0) {
3601 		ip->i_nlink--;
3602 		DIP_SET(ip, i_nlink, ip->i_nlink);
3603 		ip->i_flag |= IN_CHANGE;
3604 		if (ip->i_nlink < ip->i_effnlink)
3605 			panic("handle_workitem_remove: bad file delta");
3606 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3607 		num_dirrem -= 1;
3608 		WORKITEM_FREE(dirrem, D_DIRREM);
3609 		FREE_LOCK(&lk);
3610 		vput(vp);
3611 		return;
3612 	}
3613 	/*
3614 	 * Directory deletion. Decrement reference count for both the
3615 	 * just deleted parent directory entry and the reference for ".".
3616 	 * Next truncate the directory to length zero. When the
3617 	 * truncation completes, arrange to have the reference count on
3618 	 * the parent decremented to account for the loss of "..".
3619 	 */
3620 	ip->i_nlink -= 2;
3621 	DIP_SET(ip, i_nlink, ip->i_nlink);
3622 	ip->i_flag |= IN_CHANGE;
3623 	if (ip->i_nlink < ip->i_effnlink)
3624 		panic("handle_workitem_remove: bad dir delta");
3625 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3626 	FREE_LOCK(&lk);
3627 	if ((error = ffs_truncate(vp, (off_t)0, 0, td->td_ucred, td)) != 0)
3628 		softdep_error("handle_workitem_remove: truncate", error);
3629 	ACQUIRE_LOCK(&lk);
3630 	/*
3631 	 * Rename a directory to a new parent. Since, we are both deleting
3632 	 * and creating a new directory entry, the link count on the new
3633 	 * directory should not change. Thus we skip the followup dirrem.
3634 	 */
3635 	if (dirrem->dm_state & DIRCHG) {
3636 		num_dirrem -= 1;
3637 		WORKITEM_FREE(dirrem, D_DIRREM);
3638 		FREE_LOCK(&lk);
3639 		vput(vp);
3640 		return;
3641 	}
3642 	/*
3643 	 * If the inodedep does not exist, then the zero'ed inode has
3644 	 * been written to disk. If the allocated inode has never been
3645 	 * written to disk, then the on-disk inode is zero'ed. In either
3646 	 * case we can remove the file immediately.
3647 	 */
3648 	dirrem->dm_state = 0;
3649 	oldinum = dirrem->dm_oldinum;
3650 	dirrem->dm_oldinum = dirrem->dm_dirinum;
3651 	if (inodedep_lookup(dirrem->dm_list.wk_mp, oldinum,
3652 	    0, &inodedep) == 0 || check_inode_unwritten(inodedep)) {
3653 		if (xp != NULL)
3654 			add_to_worklist(&dirrem->dm_list);
3655 		FREE_LOCK(&lk);
3656 		vput(vp);
3657 		if (xp == NULL)
3658 			handle_workitem_remove(dirrem, NULL);
3659 		return;
3660 	}
3661 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3662 	FREE_LOCK(&lk);
3663 	ip->i_flag |= IN_CHANGE;
3664 	ffs_update(vp, 0);
3665 	vput(vp);
3666 }
3667 
3668 /*
3669  * Inode de-allocation dependencies.
3670  *
3671  * When an inode's link count is reduced to zero, it can be de-allocated. We
3672  * found it convenient to postpone de-allocation until after the inode is
3673  * written to disk with its new link count (zero).  At this point, all of the
3674  * on-disk inode's block pointers are nullified and, with careful dependency
3675  * list ordering, all dependencies related to the inode will be satisfied and
3676  * the corresponding dependency structures de-allocated.  So, if/when the
3677  * inode is reused, there will be no mixing of old dependencies with new
3678  * ones.  This artificial dependency is set up by the block de-allocation
3679  * procedure above (softdep_setup_freeblocks) and completed by the
3680  * following procedure.
3681  */
3682 static void
3683 handle_workitem_freefile(freefile)
3684 	struct freefile *freefile;
3685 {
3686 	struct fs *fs;
3687 	struct inodedep *idp;
3688 	struct ufsmount *ump;
3689 	int error;
3690 
3691 	ump = VFSTOUFS(freefile->fx_list.wk_mp);
3692 	fs = ump->um_fs;
3693 #ifdef DEBUG
3694 	ACQUIRE_LOCK(&lk);
3695 	error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp);
3696 	FREE_LOCK(&lk);
3697 	if (error)
3698 		panic("handle_workitem_freefile: inodedep survived");
3699 #endif
3700 	UFS_LOCK(ump);
3701 	fs->fs_pendinginodes -= 1;
3702 	UFS_UNLOCK(ump);
3703 	if ((error = ffs_freefile(ump, fs, freefile->fx_devvp,
3704 	    freefile->fx_oldinum, freefile->fx_mode)) != 0)
3705 		softdep_error("handle_workitem_freefile", error);
3706 	ACQUIRE_LOCK(&lk);
3707 	WORKITEM_FREE(freefile, D_FREEFILE);
3708 	FREE_LOCK(&lk);
3709 }
3710 
3711 
3712 /*
3713  * Helper function which unlinks marker element from work list and returns
3714  * the next element on the list.
3715  */
3716 static __inline struct worklist *
3717 markernext(struct worklist *marker)
3718 {
3719 	struct worklist *next;
3720 
3721 	next = LIST_NEXT(marker, wk_list);
3722 	LIST_REMOVE(marker, wk_list);
3723 	return next;
3724 }
3725 
3726 /*
3727  * Disk writes.
3728  *
3729  * The dependency structures constructed above are most actively used when file
3730  * system blocks are written to disk.  No constraints are placed on when a
3731  * block can be written, but unsatisfied update dependencies are made safe by
3732  * modifying (or replacing) the source memory for the duration of the disk
3733  * write.  When the disk write completes, the memory block is again brought
3734  * up-to-date.
3735  *
3736  * In-core inode structure reclamation.
3737  *
3738  * Because there are a finite number of "in-core" inode structures, they are
3739  * reused regularly.  By transferring all inode-related dependencies to the
3740  * in-memory inode block and indexing them separately (via "inodedep"s), we
3741  * can allow "in-core" inode structures to be reused at any time and avoid
3742  * any increase in contention.
3743  *
3744  * Called just before entering the device driver to initiate a new disk I/O.
3745  * The buffer must be locked, thus, no I/O completion operations can occur
3746  * while we are manipulating its associated dependencies.
3747  */
3748 static void
3749 softdep_disk_io_initiation(bp)
3750 	struct buf *bp;		/* structure describing disk write to occur */
3751 {
3752 	struct worklist *wk;
3753 	struct worklist marker;
3754 	struct indirdep *indirdep;
3755 	struct inodedep *inodedep;
3756 
3757 	/*
3758 	 * We only care about write operations. There should never
3759 	 * be dependencies for reads.
3760 	 */
3761 	if (bp->b_iocmd != BIO_WRITE)
3762 		panic("softdep_disk_io_initiation: not write");
3763 
3764 	marker.wk_type = D_LAST + 1;	/* Not a normal workitem */
3765 	PHOLD(curproc);			/* Don't swap out kernel stack */
3766 
3767 	ACQUIRE_LOCK(&lk);
3768 	/*
3769 	 * Do any necessary pre-I/O processing.
3770 	 */
3771 	for (wk = LIST_FIRST(&bp->b_dep); wk != NULL;
3772 	     wk = markernext(&marker)) {
3773 		LIST_INSERT_AFTER(wk, &marker, wk_list);
3774 		switch (wk->wk_type) {
3775 
3776 		case D_PAGEDEP:
3777 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
3778 			continue;
3779 
3780 		case D_INODEDEP:
3781 			inodedep = WK_INODEDEP(wk);
3782 			if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
3783 				initiate_write_inodeblock_ufs1(inodedep, bp);
3784 			else
3785 				initiate_write_inodeblock_ufs2(inodedep, bp);
3786 			continue;
3787 
3788 		case D_INDIRDEP:
3789 			indirdep = WK_INDIRDEP(wk);
3790 			if (indirdep->ir_state & GOINGAWAY)
3791 				panic("disk_io_initiation: indirdep gone");
3792 			/*
3793 			 * If there are no remaining dependencies, this
3794 			 * will be writing the real pointers, so the
3795 			 * dependency can be freed.
3796 			 */
3797 			if (LIST_EMPTY(&indirdep->ir_deplisthd)) {
3798 				struct buf *bp;
3799 
3800 				bp = indirdep->ir_savebp;
3801 				bp->b_flags |= B_INVAL | B_NOCACHE;
3802 				/* inline expand WORKLIST_REMOVE(wk); */
3803 				wk->wk_state &= ~ONWORKLIST;
3804 				LIST_REMOVE(wk, wk_list);
3805 				WORKITEM_FREE(indirdep, D_INDIRDEP);
3806 				FREE_LOCK(&lk);
3807 				brelse(bp);
3808 				ACQUIRE_LOCK(&lk);
3809 				continue;
3810 			}
3811 			/*
3812 			 * Replace up-to-date version with safe version.
3813 			 */
3814 			FREE_LOCK(&lk);
3815 			indirdep->ir_saveddata = malloc(bp->b_bcount,
3816 			    M_INDIRDEP, M_SOFTDEP_FLAGS);
3817 			ACQUIRE_LOCK(&lk);
3818 			indirdep->ir_state &= ~ATTACHED;
3819 			indirdep->ir_state |= UNDONE;
3820 			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3821 			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3822 			    bp->b_bcount);
3823 			continue;
3824 
3825 		case D_MKDIR:
3826 		case D_BMSAFEMAP:
3827 		case D_ALLOCDIRECT:
3828 		case D_ALLOCINDIR:
3829 			continue;
3830 
3831 		default:
3832 			panic("handle_disk_io_initiation: Unexpected type %s",
3833 			    TYPENAME(wk->wk_type));
3834 			/* NOTREACHED */
3835 		}
3836 	}
3837 	FREE_LOCK(&lk);
3838 	PRELE(curproc);			/* Allow swapout of kernel stack */
3839 }
3840 
3841 /*
3842  * Called from within the procedure above to deal with unsatisfied
3843  * allocation dependencies in a directory. The buffer must be locked,
3844  * thus, no I/O completion operations can occur while we are
3845  * manipulating its associated dependencies.
3846  */
3847 static void
3848 initiate_write_filepage(pagedep, bp)
3849 	struct pagedep *pagedep;
3850 	struct buf *bp;
3851 {
3852 	struct diradd *dap;
3853 	struct direct *ep;
3854 	int i;
3855 
3856 	if (pagedep->pd_state & IOSTARTED) {
3857 		/*
3858 		 * This can only happen if there is a driver that does not
3859 		 * understand chaining. Here biodone will reissue the call
3860 		 * to strategy for the incomplete buffers.
3861 		 */
3862 		printf("initiate_write_filepage: already started\n");
3863 		return;
3864 	}
3865 	pagedep->pd_state |= IOSTARTED;
3866 	for (i = 0; i < DAHASHSZ; i++) {
3867 		LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3868 			ep = (struct direct *)
3869 			    ((char *)bp->b_data + dap->da_offset);
3870 			if (ep->d_ino != dap->da_newinum)
3871 				panic("%s: dir inum %d != new %d",
3872 				    "initiate_write_filepage",
3873 				    ep->d_ino, dap->da_newinum);
3874 			if (dap->da_state & DIRCHG)
3875 				ep->d_ino = dap->da_previous->dm_oldinum;
3876 			else
3877 				ep->d_ino = 0;
3878 			dap->da_state &= ~ATTACHED;
3879 			dap->da_state |= UNDONE;
3880 		}
3881 	}
3882 }
3883 
3884 /*
3885  * Version of initiate_write_inodeblock that handles UFS1 dinodes.
3886  * Note that any bug fixes made to this routine must be done in the
3887  * version found below.
3888  *
3889  * Called from within the procedure above to deal with unsatisfied
3890  * allocation dependencies in an inodeblock. The buffer must be
3891  * locked, thus, no I/O completion operations can occur while we
3892  * are manipulating its associated dependencies.
3893  */
3894 static void
3895 initiate_write_inodeblock_ufs1(inodedep, bp)
3896 	struct inodedep *inodedep;
3897 	struct buf *bp;			/* The inode block */
3898 {
3899 	struct allocdirect *adp, *lastadp;
3900 	struct ufs1_dinode *dp;
3901 	struct ufs1_dinode *sip;
3902 	struct fs *fs;
3903 	ufs_lbn_t i;
3904 #ifdef INVARIANTS
3905 	ufs_lbn_t prevlbn = 0;
3906 #endif
3907 	int deplist;
3908 
3909 	if (inodedep->id_state & IOSTARTED)
3910 		panic("initiate_write_inodeblock_ufs1: already started");
3911 	inodedep->id_state |= IOSTARTED;
3912 	fs = inodedep->id_fs;
3913 	dp = (struct ufs1_dinode *)bp->b_data +
3914 	    ino_to_fsbo(fs, inodedep->id_ino);
3915 	/*
3916 	 * If the bitmap is not yet written, then the allocated
3917 	 * inode cannot be written to disk.
3918 	 */
3919 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3920 		if (inodedep->id_savedino1 != NULL)
3921 			panic("initiate_write_inodeblock_ufs1: I/O underway");
3922 		FREE_LOCK(&lk);
3923 		sip = malloc(sizeof(struct ufs1_dinode),
3924 		    M_SAVEDINO, M_SOFTDEP_FLAGS);
3925 		ACQUIRE_LOCK(&lk);
3926 		inodedep->id_savedino1 = sip;
3927 		*inodedep->id_savedino1 = *dp;
3928 		bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
3929 		dp->di_gen = inodedep->id_savedino1->di_gen;
3930 		return;
3931 	}
3932 	/*
3933 	 * If no dependencies, then there is nothing to roll back.
3934 	 */
3935 	inodedep->id_savedsize = dp->di_size;
3936 	inodedep->id_savedextsize = 0;
3937 	if (TAILQ_EMPTY(&inodedep->id_inoupdt))
3938 		return;
3939 	/*
3940 	 * Set the dependencies to busy.
3941 	 */
3942 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3943 	     adp = TAILQ_NEXT(adp, ad_next)) {
3944 #ifdef INVARIANTS
3945 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
3946 			panic("softdep_write_inodeblock: lbn order");
3947 		prevlbn = adp->ad_lbn;
3948 		if (adp->ad_lbn < NDADDR &&
3949 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno)
3950 			panic("%s: direct pointer #%jd mismatch %d != %jd",
3951 			    "softdep_write_inodeblock",
3952 			    (intmax_t)adp->ad_lbn,
3953 			    dp->di_db[adp->ad_lbn],
3954 			    (intmax_t)adp->ad_newblkno);
3955 		if (adp->ad_lbn >= NDADDR &&
3956 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno)
3957 			panic("%s: indirect pointer #%jd mismatch %d != %jd",
3958 			    "softdep_write_inodeblock",
3959 			    (intmax_t)adp->ad_lbn - NDADDR,
3960 			    dp->di_ib[adp->ad_lbn - NDADDR],
3961 			    (intmax_t)adp->ad_newblkno);
3962 		deplist |= 1 << adp->ad_lbn;
3963 		if ((adp->ad_state & ATTACHED) == 0)
3964 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3965 			    adp->ad_state);
3966 #endif /* INVARIANTS */
3967 		adp->ad_state &= ~ATTACHED;
3968 		adp->ad_state |= UNDONE;
3969 	}
3970 	/*
3971 	 * The on-disk inode cannot claim to be any larger than the last
3972 	 * fragment that has been written. Otherwise, the on-disk inode
3973 	 * might have fragments that were not the last block in the file
3974 	 * which would corrupt the filesystem.
3975 	 */
3976 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3977 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3978 		if (adp->ad_lbn >= NDADDR)
3979 			break;
3980 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3981 		/* keep going until hitting a rollback to a frag */
3982 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3983 			continue;
3984 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3985 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3986 #ifdef INVARIANTS
3987 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
3988 				panic("softdep_write_inodeblock: lost dep1");
3989 #endif /* INVARIANTS */
3990 			dp->di_db[i] = 0;
3991 		}
3992 		for (i = 0; i < NIADDR; i++) {
3993 #ifdef INVARIANTS
3994 			if (dp->di_ib[i] != 0 &&
3995 			    (deplist & ((1 << NDADDR) << i)) == 0)
3996 				panic("softdep_write_inodeblock: lost dep2");
3997 #endif /* INVARIANTS */
3998 			dp->di_ib[i] = 0;
3999 		}
4000 		return;
4001 	}
4002 	/*
4003 	 * If we have zero'ed out the last allocated block of the file,
4004 	 * roll back the size to the last currently allocated block.
4005 	 * We know that this last allocated block is a full-sized as
4006 	 * we already checked for fragments in the loop above.
4007 	 */
4008 	if (lastadp != NULL &&
4009 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
4010 		for (i = lastadp->ad_lbn; i >= 0; i--)
4011 			if (dp->di_db[i] != 0)
4012 				break;
4013 		dp->di_size = (i + 1) * fs->fs_bsize;
4014 	}
4015 	/*
4016 	 * The only dependencies are for indirect blocks.
4017 	 *
4018 	 * The file size for indirect block additions is not guaranteed.
4019 	 * Such a guarantee would be non-trivial to achieve. The conventional
4020 	 * synchronous write implementation also does not make this guarantee.
4021 	 * Fsck should catch and fix discrepancies. Arguably, the file size
4022 	 * can be over-estimated without destroying integrity when the file
4023 	 * moves into the indirect blocks (i.e., is large). If we want to
4024 	 * postpone fsck, we are stuck with this argument.
4025 	 */
4026 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
4027 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
4028 }
4029 
4030 /*
4031  * Version of initiate_write_inodeblock that handles UFS2 dinodes.
4032  * Note that any bug fixes made to this routine must be done in the
4033  * version found above.
4034  *
4035  * Called from within the procedure above to deal with unsatisfied
4036  * allocation dependencies in an inodeblock. The buffer must be
4037  * locked, thus, no I/O completion operations can occur while we
4038  * are manipulating its associated dependencies.
4039  */
4040 static void
4041 initiate_write_inodeblock_ufs2(inodedep, bp)
4042 	struct inodedep *inodedep;
4043 	struct buf *bp;			/* The inode block */
4044 {
4045 	struct allocdirect *adp, *lastadp;
4046 	struct ufs2_dinode *dp;
4047 	struct ufs2_dinode *sip;
4048 	struct fs *fs;
4049 	ufs_lbn_t i;
4050 #ifdef INVARIANTS
4051 	ufs_lbn_t prevlbn = 0;
4052 #endif
4053 	int deplist;
4054 
4055 	if (inodedep->id_state & IOSTARTED)
4056 		panic("initiate_write_inodeblock_ufs2: already started");
4057 	inodedep->id_state |= IOSTARTED;
4058 	fs = inodedep->id_fs;
4059 	dp = (struct ufs2_dinode *)bp->b_data +
4060 	    ino_to_fsbo(fs, inodedep->id_ino);
4061 	/*
4062 	 * If the bitmap is not yet written, then the allocated
4063 	 * inode cannot be written to disk.
4064 	 */
4065 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4066 		if (inodedep->id_savedino2 != NULL)
4067 			panic("initiate_write_inodeblock_ufs2: I/O underway");
4068 		FREE_LOCK(&lk);
4069 		sip = malloc(sizeof(struct ufs2_dinode),
4070 		    M_SAVEDINO, M_SOFTDEP_FLAGS);
4071 		ACQUIRE_LOCK(&lk);
4072 		inodedep->id_savedino2 = sip;
4073 		*inodedep->id_savedino2 = *dp;
4074 		bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
4075 		dp->di_gen = inodedep->id_savedino2->di_gen;
4076 		return;
4077 	}
4078 	/*
4079 	 * If no dependencies, then there is nothing to roll back.
4080 	 */
4081 	inodedep->id_savedsize = dp->di_size;
4082 	inodedep->id_savedextsize = dp->di_extsize;
4083 	if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
4084 	    TAILQ_EMPTY(&inodedep->id_extupdt))
4085 		return;
4086 	/*
4087 	 * Set the ext data dependencies to busy.
4088 	 */
4089 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
4090 	     adp = TAILQ_NEXT(adp, ad_next)) {
4091 #ifdef INVARIANTS
4092 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
4093 			panic("softdep_write_inodeblock: lbn order");
4094 		prevlbn = adp->ad_lbn;
4095 		if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno)
4096 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
4097 			    "softdep_write_inodeblock",
4098 			    (intmax_t)adp->ad_lbn,
4099 			    (intmax_t)dp->di_extb[adp->ad_lbn],
4100 			    (intmax_t)adp->ad_newblkno);
4101 		deplist |= 1 << adp->ad_lbn;
4102 		if ((adp->ad_state & ATTACHED) == 0)
4103 			panic("softdep_write_inodeblock: Unknown state 0x%x",
4104 			    adp->ad_state);
4105 #endif /* INVARIANTS */
4106 		adp->ad_state &= ~ATTACHED;
4107 		adp->ad_state |= UNDONE;
4108 	}
4109 	/*
4110 	 * The on-disk inode cannot claim to be any larger than the last
4111 	 * fragment that has been written. Otherwise, the on-disk inode
4112 	 * might have fragments that were not the last block in the ext
4113 	 * data which would corrupt the filesystem.
4114 	 */
4115 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
4116 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
4117 		dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno;
4118 		/* keep going until hitting a rollback to a frag */
4119 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
4120 			continue;
4121 		dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
4122 		for (i = adp->ad_lbn + 1; i < NXADDR; i++) {
4123 #ifdef INVARIANTS
4124 			if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0)
4125 				panic("softdep_write_inodeblock: lost dep1");
4126 #endif /* INVARIANTS */
4127 			dp->di_extb[i] = 0;
4128 		}
4129 		lastadp = NULL;
4130 		break;
4131 	}
4132 	/*
4133 	 * If we have zero'ed out the last allocated block of the ext
4134 	 * data, roll back the size to the last currently allocated block.
4135 	 * We know that this last allocated block is a full-sized as
4136 	 * we already checked for fragments in the loop above.
4137 	 */
4138 	if (lastadp != NULL &&
4139 	    dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
4140 		for (i = lastadp->ad_lbn; i >= 0; i--)
4141 			if (dp->di_extb[i] != 0)
4142 				break;
4143 		dp->di_extsize = (i + 1) * fs->fs_bsize;
4144 	}
4145 	/*
4146 	 * Set the file data dependencies to busy.
4147 	 */
4148 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
4149 	     adp = TAILQ_NEXT(adp, ad_next)) {
4150 #ifdef INVARIANTS
4151 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
4152 			panic("softdep_write_inodeblock: lbn order");
4153 		prevlbn = adp->ad_lbn;
4154 		if (adp->ad_lbn < NDADDR &&
4155 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno)
4156 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
4157 			    "softdep_write_inodeblock",
4158 			    (intmax_t)adp->ad_lbn,
4159 			    (intmax_t)dp->di_db[adp->ad_lbn],
4160 			    (intmax_t)adp->ad_newblkno);
4161 		if (adp->ad_lbn >= NDADDR &&
4162 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno)
4163 			panic("%s indirect pointer #%jd mismatch %jd != %jd",
4164 			    "softdep_write_inodeblock:",
4165 			    (intmax_t)adp->ad_lbn - NDADDR,
4166 			    (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR],
4167 			    (intmax_t)adp->ad_newblkno);
4168 		deplist |= 1 << adp->ad_lbn;
4169 		if ((adp->ad_state & ATTACHED) == 0)
4170 			panic("softdep_write_inodeblock: Unknown state 0x%x",
4171 			    adp->ad_state);
4172 #endif /* INVARIANTS */
4173 		adp->ad_state &= ~ATTACHED;
4174 		adp->ad_state |= UNDONE;
4175 	}
4176 	/*
4177 	 * The on-disk inode cannot claim to be any larger than the last
4178 	 * fragment that has been written. Otherwise, the on-disk inode
4179 	 * might have fragments that were not the last block in the file
4180 	 * which would corrupt the filesystem.
4181 	 */
4182 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
4183 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
4184 		if (adp->ad_lbn >= NDADDR)
4185 			break;
4186 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
4187 		/* keep going until hitting a rollback to a frag */
4188 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
4189 			continue;
4190 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
4191 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
4192 #ifdef INVARIANTS
4193 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
4194 				panic("softdep_write_inodeblock: lost dep2");
4195 #endif /* INVARIANTS */
4196 			dp->di_db[i] = 0;
4197 		}
4198 		for (i = 0; i < NIADDR; i++) {
4199 #ifdef INVARIANTS
4200 			if (dp->di_ib[i] != 0 &&
4201 			    (deplist & ((1 << NDADDR) << i)) == 0)
4202 				panic("softdep_write_inodeblock: lost dep3");
4203 #endif /* INVARIANTS */
4204 			dp->di_ib[i] = 0;
4205 		}
4206 		return;
4207 	}
4208 	/*
4209 	 * If we have zero'ed out the last allocated block of the file,
4210 	 * roll back the size to the last currently allocated block.
4211 	 * We know that this last allocated block is a full-sized as
4212 	 * we already checked for fragments in the loop above.
4213 	 */
4214 	if (lastadp != NULL &&
4215 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
4216 		for (i = lastadp->ad_lbn; i >= 0; i--)
4217 			if (dp->di_db[i] != 0)
4218 				break;
4219 		dp->di_size = (i + 1) * fs->fs_bsize;
4220 	}
4221 	/*
4222 	 * The only dependencies are for indirect blocks.
4223 	 *
4224 	 * The file size for indirect block additions is not guaranteed.
4225 	 * Such a guarantee would be non-trivial to achieve. The conventional
4226 	 * synchronous write implementation also does not make this guarantee.
4227 	 * Fsck should catch and fix discrepancies. Arguably, the file size
4228 	 * can be over-estimated without destroying integrity when the file
4229 	 * moves into the indirect blocks (i.e., is large). If we want to
4230 	 * postpone fsck, we are stuck with this argument.
4231 	 */
4232 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
4233 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
4234 }
4235 
4236 /*
4237  * This routine is called during the completion interrupt
4238  * service routine for a disk write (from the procedure called
4239  * by the device driver to inform the filesystem caches of
4240  * a request completion).  It should be called early in this
4241  * procedure, before the block is made available to other
4242  * processes or other routines are called.
4243  */
4244 static void
4245 softdep_disk_write_complete(bp)
4246 	struct buf *bp;		/* describes the completed disk write */
4247 {
4248 	struct worklist *wk;
4249 	struct worklist *owk;
4250 	struct workhead reattach;
4251 	struct newblk *newblk;
4252 	struct allocindir *aip;
4253 	struct allocdirect *adp;
4254 	struct indirdep *indirdep;
4255 	struct inodedep *inodedep;
4256 	struct bmsafemap *bmsafemap;
4257 
4258 	/*
4259 	 * If an error occurred while doing the write, then the data
4260 	 * has not hit the disk and the dependencies cannot be unrolled.
4261 	 */
4262 	if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0)
4263 		return;
4264 	LIST_INIT(&reattach);
4265 	/*
4266 	 * This lock must not be released anywhere in this code segment.
4267 	 */
4268 	ACQUIRE_LOCK(&lk);
4269 	owk = NULL;
4270 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
4271 		WORKLIST_REMOVE(wk);
4272 		if (wk == owk)
4273 			panic("duplicate worklist: %p\n", wk);
4274 		owk = wk;
4275 		switch (wk->wk_type) {
4276 
4277 		case D_PAGEDEP:
4278 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
4279 				WORKLIST_INSERT(&reattach, wk);
4280 			continue;
4281 
4282 		case D_INODEDEP:
4283 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
4284 				WORKLIST_INSERT(&reattach, wk);
4285 			continue;
4286 
4287 		case D_BMSAFEMAP:
4288 			bmsafemap = WK_BMSAFEMAP(wk);
4289 			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
4290 				newblk->nb_state |= DEPCOMPLETE;
4291 				newblk->nb_bmsafemap = NULL;
4292 				LIST_REMOVE(newblk, nb_deps);
4293 			}
4294 			while ((adp =
4295 			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
4296 				adp->ad_state |= DEPCOMPLETE;
4297 				adp->ad_buf = NULL;
4298 				LIST_REMOVE(adp, ad_deps);
4299 				handle_allocdirect_partdone(adp);
4300 			}
4301 			while ((aip =
4302 			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
4303 				aip->ai_state |= DEPCOMPLETE;
4304 				aip->ai_buf = NULL;
4305 				LIST_REMOVE(aip, ai_deps);
4306 				handle_allocindir_partdone(aip);
4307 			}
4308 			while ((inodedep =
4309 			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
4310 				inodedep->id_state |= DEPCOMPLETE;
4311 				LIST_REMOVE(inodedep, id_deps);
4312 				inodedep->id_buf = NULL;
4313 			}
4314 			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
4315 			continue;
4316 
4317 		case D_MKDIR:
4318 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
4319 			continue;
4320 
4321 		case D_ALLOCDIRECT:
4322 			adp = WK_ALLOCDIRECT(wk);
4323 			adp->ad_state |= COMPLETE;
4324 			handle_allocdirect_partdone(adp);
4325 			continue;
4326 
4327 		case D_ALLOCINDIR:
4328 			aip = WK_ALLOCINDIR(wk);
4329 			aip->ai_state |= COMPLETE;
4330 			handle_allocindir_partdone(aip);
4331 			continue;
4332 
4333 		case D_INDIRDEP:
4334 			indirdep = WK_INDIRDEP(wk);
4335 			if (indirdep->ir_state & GOINGAWAY)
4336 				panic("disk_write_complete: indirdep gone");
4337 			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
4338 			free(indirdep->ir_saveddata, M_INDIRDEP);
4339 			indirdep->ir_saveddata = 0;
4340 			indirdep->ir_state &= ~UNDONE;
4341 			indirdep->ir_state |= ATTACHED;
4342 			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
4343 				handle_allocindir_partdone(aip);
4344 				if (aip == LIST_FIRST(&indirdep->ir_donehd))
4345 					panic("disk_write_complete: not gone");
4346 			}
4347 			WORKLIST_INSERT(&reattach, wk);
4348 			if ((bp->b_flags & B_DELWRI) == 0)
4349 				stat_indir_blk_ptrs++;
4350 			bdirty(bp);
4351 			continue;
4352 
4353 		default:
4354 			panic("handle_disk_write_complete: Unknown type %s",
4355 			    TYPENAME(wk->wk_type));
4356 			/* NOTREACHED */
4357 		}
4358 	}
4359 	/*
4360 	 * Reattach any requests that must be redone.
4361 	 */
4362 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
4363 		WORKLIST_REMOVE(wk);
4364 		WORKLIST_INSERT(&bp->b_dep, wk);
4365 	}
4366 	FREE_LOCK(&lk);
4367 }
4368 
4369 /*
4370  * Called from within softdep_disk_write_complete above. Note that
4371  * this routine is always called from interrupt level with further
4372  * splbio interrupts blocked.
4373  */
4374 static void
4375 handle_allocdirect_partdone(adp)
4376 	struct allocdirect *adp;	/* the completed allocdirect */
4377 {
4378 	struct allocdirectlst *listhead;
4379 	struct allocdirect *listadp;
4380 	struct inodedep *inodedep;
4381 	long bsize, delay;
4382 
4383 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4384 		return;
4385 	if (adp->ad_buf != NULL)
4386 		panic("handle_allocdirect_partdone: dangling dep");
4387 	/*
4388 	 * The on-disk inode cannot claim to be any larger than the last
4389 	 * fragment that has been written. Otherwise, the on-disk inode
4390 	 * might have fragments that were not the last block in the file
4391 	 * which would corrupt the filesystem. Thus, we cannot free any
4392 	 * allocdirects after one whose ad_oldblkno claims a fragment as
4393 	 * these blocks must be rolled back to zero before writing the inode.
4394 	 * We check the currently active set of allocdirects in id_inoupdt
4395 	 * or id_extupdt as appropriate.
4396 	 */
4397 	inodedep = adp->ad_inodedep;
4398 	bsize = inodedep->id_fs->fs_bsize;
4399 	if (adp->ad_state & EXTDATA)
4400 		listhead = &inodedep->id_extupdt;
4401 	else
4402 		listhead = &inodedep->id_inoupdt;
4403 	TAILQ_FOREACH(listadp, listhead, ad_next) {
4404 		/* found our block */
4405 		if (listadp == adp)
4406 			break;
4407 		/* continue if ad_oldlbn is not a fragment */
4408 		if (listadp->ad_oldsize == 0 ||
4409 		    listadp->ad_oldsize == bsize)
4410 			continue;
4411 		/* hit a fragment */
4412 		return;
4413 	}
4414 	/*
4415 	 * If we have reached the end of the current list without
4416 	 * finding the just finished dependency, then it must be
4417 	 * on the future dependency list. Future dependencies cannot
4418 	 * be freed until they are moved to the current list.
4419 	 */
4420 	if (listadp == NULL) {
4421 #ifdef DEBUG
4422 		if (adp->ad_state & EXTDATA)
4423 			listhead = &inodedep->id_newextupdt;
4424 		else
4425 			listhead = &inodedep->id_newinoupdt;
4426 		TAILQ_FOREACH(listadp, listhead, ad_next)
4427 			/* found our block */
4428 			if (listadp == adp)
4429 				break;
4430 		if (listadp == NULL)
4431 			panic("handle_allocdirect_partdone: lost dep");
4432 #endif /* DEBUG */
4433 		return;
4434 	}
4435 	/*
4436 	 * If we have found the just finished dependency, then free
4437 	 * it along with anything that follows it that is complete.
4438 	 * If the inode still has a bitmap dependency, then it has
4439 	 * never been written to disk, hence the on-disk inode cannot
4440 	 * reference the old fragment so we can free it without delay.
4441 	 */
4442 	delay = (inodedep->id_state & DEPCOMPLETE);
4443 	for (; adp; adp = listadp) {
4444 		listadp = TAILQ_NEXT(adp, ad_next);
4445 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4446 			return;
4447 		free_allocdirect(listhead, adp, delay);
4448 	}
4449 }
4450 
4451 /*
4452  * Called from within softdep_disk_write_complete above. Note that
4453  * this routine is always called from interrupt level with further
4454  * splbio interrupts blocked.
4455  */
4456 static void
4457 handle_allocindir_partdone(aip)
4458 	struct allocindir *aip;		/* the completed allocindir */
4459 {
4460 	struct indirdep *indirdep;
4461 
4462 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
4463 		return;
4464 	if (aip->ai_buf != NULL)
4465 		panic("handle_allocindir_partdone: dangling dependency");
4466 	indirdep = aip->ai_indirdep;
4467 	if (indirdep->ir_state & UNDONE) {
4468 		LIST_REMOVE(aip, ai_next);
4469 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
4470 		return;
4471 	}
4472 	if (indirdep->ir_state & UFS1FMT)
4473 		((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4474 		    aip->ai_newblkno;
4475 	else
4476 		((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4477 		    aip->ai_newblkno;
4478 	LIST_REMOVE(aip, ai_next);
4479 	if (aip->ai_freefrag != NULL)
4480 		add_to_worklist(&aip->ai_freefrag->ff_list);
4481 	WORKITEM_FREE(aip, D_ALLOCINDIR);
4482 }
4483 
4484 /*
4485  * Called from within softdep_disk_write_complete above to restore
4486  * in-memory inode block contents to their most up-to-date state. Note
4487  * that this routine is always called from interrupt level with further
4488  * splbio interrupts blocked.
4489  */
4490 static int
4491 handle_written_inodeblock(inodedep, bp)
4492 	struct inodedep *inodedep;
4493 	struct buf *bp;		/* buffer containing the inode block */
4494 {
4495 	struct worklist *wk, *filefree;
4496 	struct allocdirect *adp, *nextadp;
4497 	struct ufs1_dinode *dp1 = NULL;
4498 	struct ufs2_dinode *dp2 = NULL;
4499 	int hadchanges, fstype;
4500 
4501 	if ((inodedep->id_state & IOSTARTED) == 0)
4502 		panic("handle_written_inodeblock: not started");
4503 	inodedep->id_state &= ~IOSTARTED;
4504 	if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
4505 		fstype = UFS1;
4506 		dp1 = (struct ufs1_dinode *)bp->b_data +
4507 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4508 	} else {
4509 		fstype = UFS2;
4510 		dp2 = (struct ufs2_dinode *)bp->b_data +
4511 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4512 	}
4513 	/*
4514 	 * If we had to rollback the inode allocation because of
4515 	 * bitmaps being incomplete, then simply restore it.
4516 	 * Keep the block dirty so that it will not be reclaimed until
4517 	 * all associated dependencies have been cleared and the
4518 	 * corresponding updates written to disk.
4519 	 */
4520 	if (inodedep->id_savedino1 != NULL) {
4521 		if (fstype == UFS1)
4522 			*dp1 = *inodedep->id_savedino1;
4523 		else
4524 			*dp2 = *inodedep->id_savedino2;
4525 		free(inodedep->id_savedino1, M_SAVEDINO);
4526 		inodedep->id_savedino1 = NULL;
4527 		if ((bp->b_flags & B_DELWRI) == 0)
4528 			stat_inode_bitmap++;
4529 		bdirty(bp);
4530 		return (1);
4531 	}
4532 	inodedep->id_state |= COMPLETE;
4533 	/*
4534 	 * Roll forward anything that had to be rolled back before
4535 	 * the inode could be updated.
4536 	 */
4537 	hadchanges = 0;
4538 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
4539 		nextadp = TAILQ_NEXT(adp, ad_next);
4540 		if (adp->ad_state & ATTACHED)
4541 			panic("handle_written_inodeblock: new entry");
4542 		if (fstype == UFS1) {
4543 			if (adp->ad_lbn < NDADDR) {
4544 				if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno)
4545 					panic("%s %s #%jd mismatch %d != %jd",
4546 					    "handle_written_inodeblock:",
4547 					    "direct pointer",
4548 					    (intmax_t)adp->ad_lbn,
4549 					    dp1->di_db[adp->ad_lbn],
4550 					    (intmax_t)adp->ad_oldblkno);
4551 				dp1->di_db[adp->ad_lbn] = adp->ad_newblkno;
4552 			} else {
4553 				if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0)
4554 					panic("%s: %s #%jd allocated as %d",
4555 					    "handle_written_inodeblock",
4556 					    "indirect pointer",
4557 					    (intmax_t)adp->ad_lbn - NDADDR,
4558 					    dp1->di_ib[adp->ad_lbn - NDADDR]);
4559 				dp1->di_ib[adp->ad_lbn - NDADDR] =
4560 				    adp->ad_newblkno;
4561 			}
4562 		} else {
4563 			if (adp->ad_lbn < NDADDR) {
4564 				if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno)
4565 					panic("%s: %s #%jd %s %jd != %jd",
4566 					    "handle_written_inodeblock",
4567 					    "direct pointer",
4568 					    (intmax_t)adp->ad_lbn, "mismatch",
4569 					    (intmax_t)dp2->di_db[adp->ad_lbn],
4570 					    (intmax_t)adp->ad_oldblkno);
4571 				dp2->di_db[adp->ad_lbn] = adp->ad_newblkno;
4572 			} else {
4573 				if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0)
4574 					panic("%s: %s #%jd allocated as %jd",
4575 					    "handle_written_inodeblock",
4576 					    "indirect pointer",
4577 					    (intmax_t)adp->ad_lbn - NDADDR,
4578 					    (intmax_t)
4579 					    dp2->di_ib[adp->ad_lbn - NDADDR]);
4580 				dp2->di_ib[adp->ad_lbn - NDADDR] =
4581 				    adp->ad_newblkno;
4582 			}
4583 		}
4584 		adp->ad_state &= ~UNDONE;
4585 		adp->ad_state |= ATTACHED;
4586 		hadchanges = 1;
4587 	}
4588 	for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
4589 		nextadp = TAILQ_NEXT(adp, ad_next);
4590 		if (adp->ad_state & ATTACHED)
4591 			panic("handle_written_inodeblock: new entry");
4592 		if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno)
4593 			panic("%s: direct pointers #%jd %s %jd != %jd",
4594 			    "handle_written_inodeblock",
4595 			    (intmax_t)adp->ad_lbn, "mismatch",
4596 			    (intmax_t)dp2->di_extb[adp->ad_lbn],
4597 			    (intmax_t)adp->ad_oldblkno);
4598 		dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno;
4599 		adp->ad_state &= ~UNDONE;
4600 		adp->ad_state |= ATTACHED;
4601 		hadchanges = 1;
4602 	}
4603 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
4604 		stat_direct_blk_ptrs++;
4605 	/*
4606 	 * Reset the file size to its most up-to-date value.
4607 	 */
4608 	if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1)
4609 		panic("handle_written_inodeblock: bad size");
4610 	if (fstype == UFS1) {
4611 		if (dp1->di_size != inodedep->id_savedsize) {
4612 			dp1->di_size = inodedep->id_savedsize;
4613 			hadchanges = 1;
4614 		}
4615 	} else {
4616 		if (dp2->di_size != inodedep->id_savedsize) {
4617 			dp2->di_size = inodedep->id_savedsize;
4618 			hadchanges = 1;
4619 		}
4620 		if (dp2->di_extsize != inodedep->id_savedextsize) {
4621 			dp2->di_extsize = inodedep->id_savedextsize;
4622 			hadchanges = 1;
4623 		}
4624 	}
4625 	inodedep->id_savedsize = -1;
4626 	inodedep->id_savedextsize = -1;
4627 	/*
4628 	 * If there were any rollbacks in the inode block, then it must be
4629 	 * marked dirty so that its will eventually get written back in
4630 	 * its correct form.
4631 	 */
4632 	if (hadchanges)
4633 		bdirty(bp);
4634 	/*
4635 	 * Process any allocdirects that completed during the update.
4636 	 */
4637 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
4638 		handle_allocdirect_partdone(adp);
4639 	if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
4640 		handle_allocdirect_partdone(adp);
4641 	/*
4642 	 * Process deallocations that were held pending until the
4643 	 * inode had been written to disk. Freeing of the inode
4644 	 * is delayed until after all blocks have been freed to
4645 	 * avoid creation of new <vfsid, inum, lbn> triples
4646 	 * before the old ones have been deleted.
4647 	 */
4648 	filefree = NULL;
4649 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
4650 		WORKLIST_REMOVE(wk);
4651 		switch (wk->wk_type) {
4652 
4653 		case D_FREEFILE:
4654 			/*
4655 			 * We defer adding filefree to the worklist until
4656 			 * all other additions have been made to ensure
4657 			 * that it will be done after all the old blocks
4658 			 * have been freed.
4659 			 */
4660 			if (filefree != NULL)
4661 				panic("handle_written_inodeblock: filefree");
4662 			filefree = wk;
4663 			continue;
4664 
4665 		case D_MKDIR:
4666 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
4667 			continue;
4668 
4669 		case D_DIRADD:
4670 			diradd_inode_written(WK_DIRADD(wk), inodedep);
4671 			continue;
4672 
4673 		case D_FREEBLKS:
4674 			wk->wk_state |= COMPLETE;
4675 			if ((wk->wk_state  & ALLCOMPLETE) != ALLCOMPLETE)
4676 				continue;
4677 			 /* -- fall through -- */
4678 		case D_FREEFRAG:
4679 		case D_DIRREM:
4680 			add_to_worklist(wk);
4681 			continue;
4682 
4683 		case D_NEWDIRBLK:
4684 			free_newdirblk(WK_NEWDIRBLK(wk));
4685 			continue;
4686 
4687 		default:
4688 			panic("handle_written_inodeblock: Unknown type %s",
4689 			    TYPENAME(wk->wk_type));
4690 			/* NOTREACHED */
4691 		}
4692 	}
4693 	if (filefree != NULL) {
4694 		if (free_inodedep(inodedep) == 0)
4695 			panic("handle_written_inodeblock: live inodedep");
4696 		add_to_worklist(filefree);
4697 		return (0);
4698 	}
4699 
4700 	/*
4701 	 * If no outstanding dependencies, free it.
4702 	 */
4703 	if (free_inodedep(inodedep) ||
4704 	    (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
4705 	     TAILQ_FIRST(&inodedep->id_extupdt) == 0))
4706 		return (0);
4707 	return (hadchanges);
4708 }
4709 
4710 /*
4711  * Process a diradd entry after its dependent inode has been written.
4712  * This routine must be called with splbio interrupts blocked.
4713  */
4714 static void
4715 diradd_inode_written(dap, inodedep)
4716 	struct diradd *dap;
4717 	struct inodedep *inodedep;
4718 {
4719 	struct pagedep *pagedep;
4720 
4721 	dap->da_state |= COMPLETE;
4722 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4723 		if (dap->da_state & DIRCHG)
4724 			pagedep = dap->da_previous->dm_pagedep;
4725 		else
4726 			pagedep = dap->da_pagedep;
4727 		LIST_REMOVE(dap, da_pdlist);
4728 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4729 	}
4730 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
4731 }
4732 
4733 /*
4734  * Handle the completion of a mkdir dependency.
4735  */
4736 static void
4737 handle_written_mkdir(mkdir, type)
4738 	struct mkdir *mkdir;
4739 	int type;
4740 {
4741 	struct diradd *dap;
4742 	struct pagedep *pagedep;
4743 
4744 	if (mkdir->md_state != type)
4745 		panic("handle_written_mkdir: bad type");
4746 	dap = mkdir->md_diradd;
4747 	dap->da_state &= ~type;
4748 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
4749 		dap->da_state |= DEPCOMPLETE;
4750 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4751 		if (dap->da_state & DIRCHG)
4752 			pagedep = dap->da_previous->dm_pagedep;
4753 		else
4754 			pagedep = dap->da_pagedep;
4755 		LIST_REMOVE(dap, da_pdlist);
4756 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4757 	}
4758 	LIST_REMOVE(mkdir, md_mkdirs);
4759 	WORKITEM_FREE(mkdir, D_MKDIR);
4760 }
4761 
4762 /*
4763  * Called from within softdep_disk_write_complete above.
4764  * A write operation was just completed. Removed inodes can
4765  * now be freed and associated block pointers may be committed.
4766  * Note that this routine is always called from interrupt level
4767  * with further splbio interrupts blocked.
4768  */
4769 static int
4770 handle_written_filepage(pagedep, bp)
4771 	struct pagedep *pagedep;
4772 	struct buf *bp;		/* buffer containing the written page */
4773 {
4774 	struct dirrem *dirrem;
4775 	struct diradd *dap, *nextdap;
4776 	struct direct *ep;
4777 	int i, chgs;
4778 
4779 	if ((pagedep->pd_state & IOSTARTED) == 0)
4780 		panic("handle_written_filepage: not started");
4781 	pagedep->pd_state &= ~IOSTARTED;
4782 	/*
4783 	 * Process any directory removals that have been committed.
4784 	 */
4785 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
4786 		LIST_REMOVE(dirrem, dm_next);
4787 		dirrem->dm_dirinum = pagedep->pd_ino;
4788 		add_to_worklist(&dirrem->dm_list);
4789 	}
4790 	/*
4791 	 * Free any directory additions that have been committed.
4792 	 * If it is a newly allocated block, we have to wait until
4793 	 * the on-disk directory inode claims the new block.
4794 	 */
4795 	if ((pagedep->pd_state & NEWBLOCK) == 0)
4796 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
4797 			free_diradd(dap);
4798 	/*
4799 	 * Uncommitted directory entries must be restored.
4800 	 */
4801 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
4802 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
4803 		     dap = nextdap) {
4804 			nextdap = LIST_NEXT(dap, da_pdlist);
4805 			if (dap->da_state & ATTACHED)
4806 				panic("handle_written_filepage: attached");
4807 			ep = (struct direct *)
4808 			    ((char *)bp->b_data + dap->da_offset);
4809 			ep->d_ino = dap->da_newinum;
4810 			dap->da_state &= ~UNDONE;
4811 			dap->da_state |= ATTACHED;
4812 			chgs = 1;
4813 			/*
4814 			 * If the inode referenced by the directory has
4815 			 * been written out, then the dependency can be
4816 			 * moved to the pending list.
4817 			 */
4818 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4819 				LIST_REMOVE(dap, da_pdlist);
4820 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
4821 				    da_pdlist);
4822 			}
4823 		}
4824 	}
4825 	/*
4826 	 * If there were any rollbacks in the directory, then it must be
4827 	 * marked dirty so that its will eventually get written back in
4828 	 * its correct form.
4829 	 */
4830 	if (chgs) {
4831 		if ((bp->b_flags & B_DELWRI) == 0)
4832 			stat_dir_entry++;
4833 		bdirty(bp);
4834 		return (1);
4835 	}
4836 	/*
4837 	 * If we are not waiting for a new directory block to be
4838 	 * claimed by its inode, then the pagedep will be freed.
4839 	 * Otherwise it will remain to track any new entries on
4840 	 * the page in case they are fsync'ed.
4841 	 */
4842 	if ((pagedep->pd_state & NEWBLOCK) == 0) {
4843 		LIST_REMOVE(pagedep, pd_hash);
4844 		WORKITEM_FREE(pagedep, D_PAGEDEP);
4845 	}
4846 	return (0);
4847 }
4848 
4849 /*
4850  * Writing back in-core inode structures.
4851  *
4852  * The filesystem only accesses an inode's contents when it occupies an
4853  * "in-core" inode structure.  These "in-core" structures are separate from
4854  * the page frames used to cache inode blocks.  Only the latter are
4855  * transferred to/from the disk.  So, when the updated contents of the
4856  * "in-core" inode structure are copied to the corresponding in-memory inode
4857  * block, the dependencies are also transferred.  The following procedure is
4858  * called when copying a dirty "in-core" inode to a cached inode block.
4859  */
4860 
4861 /*
4862  * Called when an inode is loaded from disk. If the effective link count
4863  * differed from the actual link count when it was last flushed, then we
4864  * need to ensure that the correct effective link count is put back.
4865  */
4866 void
4867 softdep_load_inodeblock(ip)
4868 	struct inode *ip;	/* the "in_core" copy of the inode */
4869 {
4870 	struct inodedep *inodedep;
4871 
4872 	/*
4873 	 * Check for alternate nlink count.
4874 	 */
4875 	ip->i_effnlink = ip->i_nlink;
4876 	ACQUIRE_LOCK(&lk);
4877 	if (inodedep_lookup(UFSTOVFS(ip->i_ump),
4878 	    ip->i_number, 0, &inodedep) == 0) {
4879 		FREE_LOCK(&lk);
4880 		return;
4881 	}
4882 	ip->i_effnlink -= inodedep->id_nlinkdelta;
4883 	if (inodedep->id_state & SPACECOUNTED)
4884 		ip->i_flag |= IN_SPACECOUNTED;
4885 	FREE_LOCK(&lk);
4886 }
4887 
4888 /*
4889  * This routine is called just before the "in-core" inode
4890  * information is to be copied to the in-memory inode block.
4891  * Recall that an inode block contains several inodes. If
4892  * the force flag is set, then the dependencies will be
4893  * cleared so that the update can always be made. Note that
4894  * the buffer is locked when this routine is called, so we
4895  * will never be in the middle of writing the inode block
4896  * to disk.
4897  */
4898 void
4899 softdep_update_inodeblock(ip, bp, waitfor)
4900 	struct inode *ip;	/* the "in_core" copy of the inode */
4901 	struct buf *bp;		/* the buffer containing the inode block */
4902 	int waitfor;		/* nonzero => update must be allowed */
4903 {
4904 	struct inodedep *inodedep;
4905 	struct worklist *wk;
4906 	struct mount *mp;
4907 	struct buf *ibp;
4908 	int error;
4909 
4910 	/*
4911 	 * If the effective link count is not equal to the actual link
4912 	 * count, then we must track the difference in an inodedep while
4913 	 * the inode is (potentially) tossed out of the cache. Otherwise,
4914 	 * if there is no existing inodedep, then there are no dependencies
4915 	 * to track.
4916 	 */
4917 	mp = UFSTOVFS(ip->i_ump);
4918 	ACQUIRE_LOCK(&lk);
4919 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
4920 		FREE_LOCK(&lk);
4921 		if (ip->i_effnlink != ip->i_nlink)
4922 			panic("softdep_update_inodeblock: bad link count");
4923 		return;
4924 	}
4925 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
4926 		panic("softdep_update_inodeblock: bad delta");
4927 	/*
4928 	 * Changes have been initiated. Anything depending on these
4929 	 * changes cannot occur until this inode has been written.
4930 	 */
4931 	inodedep->id_state &= ~COMPLETE;
4932 	if ((inodedep->id_state & ONWORKLIST) == 0)
4933 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
4934 	/*
4935 	 * Any new dependencies associated with the incore inode must
4936 	 * now be moved to the list associated with the buffer holding
4937 	 * the in-memory copy of the inode. Once merged process any
4938 	 * allocdirects that are completed by the merger.
4939 	 */
4940 	merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
4941 	if (!TAILQ_EMPTY(&inodedep->id_inoupdt))
4942 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
4943 	merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
4944 	if (!TAILQ_EMPTY(&inodedep->id_extupdt))
4945 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt));
4946 	/*
4947 	 * Now that the inode has been pushed into the buffer, the
4948 	 * operations dependent on the inode being written to disk
4949 	 * can be moved to the id_bufwait so that they will be
4950 	 * processed when the buffer I/O completes.
4951 	 */
4952 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
4953 		WORKLIST_REMOVE(wk);
4954 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
4955 	}
4956 	/*
4957 	 * Newly allocated inodes cannot be written until the bitmap
4958 	 * that allocates them have been written (indicated by
4959 	 * DEPCOMPLETE being set in id_state). If we are doing a
4960 	 * forced sync (e.g., an fsync on a file), we force the bitmap
4961 	 * to be written so that the update can be done.
4962 	 */
4963 	if (waitfor == 0) {
4964 		FREE_LOCK(&lk);
4965 		return;
4966 	}
4967 retry:
4968 	if ((inodedep->id_state & DEPCOMPLETE) != 0) {
4969 		FREE_LOCK(&lk);
4970 		return;
4971 	}
4972 	ibp = inodedep->id_buf;
4973 	ibp = getdirtybuf(ibp, &lk, MNT_WAIT);
4974 	if (ibp == NULL) {
4975 		/*
4976 		 * If ibp came back as NULL, the dependency could have been
4977 		 * freed while we slept.  Look it up again, and check to see
4978 		 * that it has completed.
4979 		 */
4980 		if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
4981 			goto retry;
4982 		FREE_LOCK(&lk);
4983 		return;
4984 	}
4985 	FREE_LOCK(&lk);
4986 	if ((error = bwrite(ibp)) != 0)
4987 		softdep_error("softdep_update_inodeblock: bwrite", error);
4988 }
4989 
4990 /*
4991  * Merge the a new inode dependency list (such as id_newinoupdt) into an
4992  * old inode dependency list (such as id_inoupdt). This routine must be
4993  * called with splbio interrupts blocked.
4994  */
4995 static void
4996 merge_inode_lists(newlisthead, oldlisthead)
4997 	struct allocdirectlst *newlisthead;
4998 	struct allocdirectlst *oldlisthead;
4999 {
5000 	struct allocdirect *listadp, *newadp;
5001 
5002 	newadp = TAILQ_FIRST(newlisthead);
5003 	for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
5004 		if (listadp->ad_lbn < newadp->ad_lbn) {
5005 			listadp = TAILQ_NEXT(listadp, ad_next);
5006 			continue;
5007 		}
5008 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
5009 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
5010 		if (listadp->ad_lbn == newadp->ad_lbn) {
5011 			allocdirect_merge(oldlisthead, newadp,
5012 			    listadp);
5013 			listadp = newadp;
5014 		}
5015 		newadp = TAILQ_FIRST(newlisthead);
5016 	}
5017 	while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
5018 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
5019 		TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
5020 	}
5021 }
5022 
5023 /*
5024  * If we are doing an fsync, then we must ensure that any directory
5025  * entries for the inode have been written after the inode gets to disk.
5026  */
5027 int
5028 softdep_fsync(vp)
5029 	struct vnode *vp;	/* the "in_core" copy of the inode */
5030 {
5031 	struct inodedep *inodedep;
5032 	struct pagedep *pagedep;
5033 	struct worklist *wk;
5034 	struct diradd *dap;
5035 	struct mount *mp;
5036 	struct vnode *pvp;
5037 	struct inode *ip;
5038 	struct buf *bp;
5039 	struct fs *fs;
5040 	struct thread *td = curthread;
5041 	int error, flushparent, pagedep_new_block;
5042 	ino_t parentino;
5043 	ufs_lbn_t lbn;
5044 
5045 	ip = VTOI(vp);
5046 	fs = ip->i_fs;
5047 	mp = vp->v_mount;
5048 	ACQUIRE_LOCK(&lk);
5049 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
5050 		FREE_LOCK(&lk);
5051 		return (0);
5052 	}
5053 	if (!LIST_EMPTY(&inodedep->id_inowait) ||
5054 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
5055 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
5056 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
5057 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
5058 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt))
5059 		panic("softdep_fsync: pending ops");
5060 	for (error = 0, flushparent = 0; ; ) {
5061 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
5062 			break;
5063 		if (wk->wk_type != D_DIRADD)
5064 			panic("softdep_fsync: Unexpected type %s",
5065 			    TYPENAME(wk->wk_type));
5066 		dap = WK_DIRADD(wk);
5067 		/*
5068 		 * Flush our parent if this directory entry has a MKDIR_PARENT
5069 		 * dependency or is contained in a newly allocated block.
5070 		 */
5071 		if (dap->da_state & DIRCHG)
5072 			pagedep = dap->da_previous->dm_pagedep;
5073 		else
5074 			pagedep = dap->da_pagedep;
5075 		parentino = pagedep->pd_ino;
5076 		lbn = pagedep->pd_lbn;
5077 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
5078 			panic("softdep_fsync: dirty");
5079 		if ((dap->da_state & MKDIR_PARENT) ||
5080 		    (pagedep->pd_state & NEWBLOCK))
5081 			flushparent = 1;
5082 		else
5083 			flushparent = 0;
5084 		/*
5085 		 * If we are being fsync'ed as part of vgone'ing this vnode,
5086 		 * then we will not be able to release and recover the
5087 		 * vnode below, so we just have to give up on writing its
5088 		 * directory entry out. It will eventually be written, just
5089 		 * not now, but then the user was not asking to have it
5090 		 * written, so we are not breaking any promises.
5091 		 */
5092 		if (vp->v_iflag & VI_DOOMED)
5093 			break;
5094 		/*
5095 		 * We prevent deadlock by always fetching inodes from the
5096 		 * root, moving down the directory tree. Thus, when fetching
5097 		 * our parent directory, we first try to get the lock. If
5098 		 * that fails, we must unlock ourselves before requesting
5099 		 * the lock on our parent. See the comment in ufs_lookup
5100 		 * for details on possible races.
5101 		 */
5102 		FREE_LOCK(&lk);
5103 		if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp,
5104 		    FFSV_FORCEINSMQ)) {
5105 			VOP_UNLOCK(vp, 0);
5106 			error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE,
5107 			    &pvp, FFSV_FORCEINSMQ);
5108 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5109 			if (error != 0)
5110 				return (error);
5111 		}
5112 		/*
5113 		 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
5114 		 * that are contained in direct blocks will be resolved by
5115 		 * doing a ffs_update. Pagedeps contained in indirect blocks
5116 		 * may require a complete sync'ing of the directory. So, we
5117 		 * try the cheap and fast ffs_update first, and if that fails,
5118 		 * then we do the slower ffs_syncvnode of the directory.
5119 		 */
5120 		if (flushparent) {
5121 			int locked;
5122 
5123 			if ((error = ffs_update(pvp, 1)) != 0) {
5124 				vput(pvp);
5125 				return (error);
5126 			}
5127 			ACQUIRE_LOCK(&lk);
5128 			locked = 1;
5129 			if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) {
5130 				if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) {
5131 					if (wk->wk_type != D_DIRADD)
5132 						panic("softdep_fsync: Unexpected type %s",
5133 						      TYPENAME(wk->wk_type));
5134 					dap = WK_DIRADD(wk);
5135 					if (dap->da_state & DIRCHG)
5136 						pagedep = dap->da_previous->dm_pagedep;
5137 					else
5138 						pagedep = dap->da_pagedep;
5139 					pagedep_new_block = pagedep->pd_state & NEWBLOCK;
5140 					FREE_LOCK(&lk);
5141 					locked = 0;
5142 					if (pagedep_new_block &&
5143 					    (error = ffs_syncvnode(pvp, MNT_WAIT))) {
5144 						vput(pvp);
5145 						return (error);
5146 					}
5147 				}
5148 			}
5149 			if (locked)
5150 				FREE_LOCK(&lk);
5151 		}
5152 		/*
5153 		 * Flush directory page containing the inode's name.
5154 		 */
5155 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
5156 		    &bp);
5157 		if (error == 0)
5158 			error = bwrite(bp);
5159 		else
5160 			brelse(bp);
5161 		vput(pvp);
5162 		if (error != 0)
5163 			return (error);
5164 		ACQUIRE_LOCK(&lk);
5165 		if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
5166 			break;
5167 	}
5168 	FREE_LOCK(&lk);
5169 	return (0);
5170 }
5171 
5172 /*
5173  * Flush all the dirty bitmaps associated with the block device
5174  * before flushing the rest of the dirty blocks so as to reduce
5175  * the number of dependencies that will have to be rolled back.
5176  */
5177 void
5178 softdep_fsync_mountdev(vp)
5179 	struct vnode *vp;
5180 {
5181 	struct buf *bp, *nbp;
5182 	struct worklist *wk;
5183 	struct bufobj *bo;
5184 
5185 	if (!vn_isdisk(vp, NULL))
5186 		panic("softdep_fsync_mountdev: vnode not a disk");
5187 	bo = &vp->v_bufobj;
5188 restart:
5189 	BO_LOCK(bo);
5190 	ACQUIRE_LOCK(&lk);
5191 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
5192 		/*
5193 		 * If it is already scheduled, skip to the next buffer.
5194 		 */
5195 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
5196 			continue;
5197 
5198 		if ((bp->b_flags & B_DELWRI) == 0)
5199 			panic("softdep_fsync_mountdev: not dirty");
5200 		/*
5201 		 * We are only interested in bitmaps with outstanding
5202 		 * dependencies.
5203 		 */
5204 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
5205 		    wk->wk_type != D_BMSAFEMAP ||
5206 		    (bp->b_vflags & BV_BKGRDINPROG)) {
5207 			BUF_UNLOCK(bp);
5208 			continue;
5209 		}
5210 		FREE_LOCK(&lk);
5211 		BO_UNLOCK(bo);
5212 		bremfree(bp);
5213 		(void) bawrite(bp);
5214 		goto restart;
5215 	}
5216 	FREE_LOCK(&lk);
5217 	drain_output(vp);
5218 	BO_UNLOCK(bo);
5219 }
5220 
5221 /*
5222  * This routine is called when we are trying to synchronously flush a
5223  * file. This routine must eliminate any filesystem metadata dependencies
5224  * so that the syncing routine can succeed by pushing the dirty blocks
5225  * associated with the file. If any I/O errors occur, they are returned.
5226  */
5227 int
5228 softdep_sync_metadata(struct vnode *vp)
5229 {
5230 	struct pagedep *pagedep;
5231 	struct allocdirect *adp;
5232 	struct allocindir *aip;
5233 	struct buf *bp, *nbp;
5234 	struct worklist *wk;
5235 	struct bufobj *bo;
5236 	int i, error, waitfor;
5237 
5238 	if (!DOINGSOFTDEP(vp))
5239 		return (0);
5240 	/*
5241 	 * Ensure that any direct block dependencies have been cleared.
5242 	 */
5243 	ACQUIRE_LOCK(&lk);
5244 	if ((error = flush_inodedep_deps(vp->v_mount, VTOI(vp)->i_number))) {
5245 		FREE_LOCK(&lk);
5246 		return (error);
5247 	}
5248 	FREE_LOCK(&lk);
5249 	/*
5250 	 * For most files, the only metadata dependencies are the
5251 	 * cylinder group maps that allocate their inode or blocks.
5252 	 * The block allocation dependencies can be found by traversing
5253 	 * the dependency lists for any buffers that remain on their
5254 	 * dirty buffer list. The inode allocation dependency will
5255 	 * be resolved when the inode is updated with MNT_WAIT.
5256 	 * This work is done in two passes. The first pass grabs most
5257 	 * of the buffers and begins asynchronously writing them. The
5258 	 * only way to wait for these asynchronous writes is to sleep
5259 	 * on the filesystem vnode which may stay busy for a long time
5260 	 * if the filesystem is active. So, instead, we make a second
5261 	 * pass over the dependencies blocking on each write. In the
5262 	 * usual case we will be blocking against a write that we
5263 	 * initiated, so when it is done the dependency will have been
5264 	 * resolved. Thus the second pass is expected to end quickly.
5265 	 */
5266 	waitfor = MNT_NOWAIT;
5267 	bo = &vp->v_bufobj;
5268 
5269 top:
5270 	/*
5271 	 * We must wait for any I/O in progress to finish so that
5272 	 * all potential buffers on the dirty list will be visible.
5273 	 */
5274 	BO_LOCK(bo);
5275 	drain_output(vp);
5276 	while ((bp = TAILQ_FIRST(&bo->bo_dirty.bv_hd)) != NULL) {
5277 		bp = getdirtybuf(bp, BO_MTX(bo), MNT_WAIT);
5278 		if (bp)
5279 			break;
5280 	}
5281 	BO_UNLOCK(bo);
5282 	if (bp == NULL)
5283 		return (0);
5284 loop:
5285 	/* While syncing snapshots, we must allow recursive lookups */
5286 	BUF_AREC(bp);
5287 	ACQUIRE_LOCK(&lk);
5288 	/*
5289 	 * As we hold the buffer locked, none of its dependencies
5290 	 * will disappear.
5291 	 */
5292 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5293 		switch (wk->wk_type) {
5294 
5295 		case D_ALLOCDIRECT:
5296 			adp = WK_ALLOCDIRECT(wk);
5297 			if (adp->ad_state & DEPCOMPLETE)
5298 				continue;
5299 			nbp = adp->ad_buf;
5300 			nbp = getdirtybuf(nbp, &lk, waitfor);
5301 			if (nbp == NULL)
5302 				continue;
5303 			FREE_LOCK(&lk);
5304 			if (waitfor == MNT_NOWAIT) {
5305 				bawrite(nbp);
5306 			} else if ((error = bwrite(nbp)) != 0) {
5307 				break;
5308 			}
5309 			ACQUIRE_LOCK(&lk);
5310 			continue;
5311 
5312 		case D_ALLOCINDIR:
5313 			aip = WK_ALLOCINDIR(wk);
5314 			if (aip->ai_state & DEPCOMPLETE)
5315 				continue;
5316 			nbp = aip->ai_buf;
5317 			nbp = getdirtybuf(nbp, &lk, waitfor);
5318 			if (nbp == NULL)
5319 				continue;
5320 			FREE_LOCK(&lk);
5321 			if (waitfor == MNT_NOWAIT) {
5322 				bawrite(nbp);
5323 			} else if ((error = bwrite(nbp)) != 0) {
5324 				break;
5325 			}
5326 			ACQUIRE_LOCK(&lk);
5327 			continue;
5328 
5329 		case D_INDIRDEP:
5330 		restart:
5331 
5332 			LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
5333 				if (aip->ai_state & DEPCOMPLETE)
5334 					continue;
5335 				nbp = aip->ai_buf;
5336 				nbp = getdirtybuf(nbp, &lk, MNT_WAIT);
5337 				if (nbp == NULL)
5338 					goto restart;
5339 				FREE_LOCK(&lk);
5340 				if ((error = bwrite(nbp)) != 0) {
5341 					goto loop_end;
5342 				}
5343 				ACQUIRE_LOCK(&lk);
5344 				goto restart;
5345 			}
5346 			continue;
5347 
5348 		case D_INODEDEP:
5349 			if ((error = flush_inodedep_deps(wk->wk_mp,
5350 			    WK_INODEDEP(wk)->id_ino)) != 0) {
5351 				FREE_LOCK(&lk);
5352 				break;
5353 			}
5354 			continue;
5355 
5356 		case D_PAGEDEP:
5357 			/*
5358 			 * We are trying to sync a directory that may
5359 			 * have dependencies on both its own metadata
5360 			 * and/or dependencies on the inodes of any
5361 			 * recently allocated files. We walk its diradd
5362 			 * lists pushing out the associated inode.
5363 			 */
5364 			pagedep = WK_PAGEDEP(wk);
5365 			for (i = 0; i < DAHASHSZ; i++) {
5366 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
5367 					continue;
5368 				if ((error =
5369 				    flush_pagedep_deps(vp, wk->wk_mp,
5370 						&pagedep->pd_diraddhd[i]))) {
5371 					FREE_LOCK(&lk);
5372 					goto loop_end;
5373 				}
5374 			}
5375 			continue;
5376 
5377 		case D_MKDIR:
5378 			/*
5379 			 * This case should never happen if the vnode has
5380 			 * been properly sync'ed. However, if this function
5381 			 * is used at a place where the vnode has not yet
5382 			 * been sync'ed, this dependency can show up. So,
5383 			 * rather than panic, just flush it.
5384 			 */
5385 			nbp = WK_MKDIR(wk)->md_buf;
5386 			nbp = getdirtybuf(nbp, &lk, waitfor);
5387 			if (nbp == NULL)
5388 				continue;
5389 			FREE_LOCK(&lk);
5390 			if (waitfor == MNT_NOWAIT) {
5391 				bawrite(nbp);
5392 			} else if ((error = bwrite(nbp)) != 0) {
5393 				break;
5394 			}
5395 			ACQUIRE_LOCK(&lk);
5396 			continue;
5397 
5398 		case D_BMSAFEMAP:
5399 			/*
5400 			 * This case should never happen if the vnode has
5401 			 * been properly sync'ed. However, if this function
5402 			 * is used at a place where the vnode has not yet
5403 			 * been sync'ed, this dependency can show up. So,
5404 			 * rather than panic, just flush it.
5405 			 */
5406 			nbp = WK_BMSAFEMAP(wk)->sm_buf;
5407 			nbp = getdirtybuf(nbp, &lk, waitfor);
5408 			if (nbp == NULL)
5409 				continue;
5410 			FREE_LOCK(&lk);
5411 			if (waitfor == MNT_NOWAIT) {
5412 				bawrite(nbp);
5413 			} else if ((error = bwrite(nbp)) != 0) {
5414 				break;
5415 			}
5416 			ACQUIRE_LOCK(&lk);
5417 			continue;
5418 
5419 		default:
5420 			panic("softdep_sync_metadata: Unknown type %s",
5421 			    TYPENAME(wk->wk_type));
5422 			/* NOTREACHED */
5423 		}
5424 	loop_end:
5425 		/* We reach here only in error and unlocked */
5426 		if (error == 0)
5427 			panic("softdep_sync_metadata: zero error");
5428 		BUF_NOREC(bp);
5429 		bawrite(bp);
5430 		return (error);
5431 	}
5432 	FREE_LOCK(&lk);
5433 	BO_LOCK(bo);
5434 	while ((nbp = TAILQ_NEXT(bp, b_bobufs)) != NULL) {
5435 		nbp = getdirtybuf(nbp, BO_MTX(bo), MNT_WAIT);
5436 		if (nbp)
5437 			break;
5438 	}
5439 	BO_UNLOCK(bo);
5440 	BUF_NOREC(bp);
5441 	bawrite(bp);
5442 	if (nbp != NULL) {
5443 		bp = nbp;
5444 		goto loop;
5445 	}
5446 	/*
5447 	 * The brief unlock is to allow any pent up dependency
5448 	 * processing to be done. Then proceed with the second pass.
5449 	 */
5450 	if (waitfor == MNT_NOWAIT) {
5451 		waitfor = MNT_WAIT;
5452 		goto top;
5453 	}
5454 
5455 	/*
5456 	 * If we have managed to get rid of all the dirty buffers,
5457 	 * then we are done. For certain directories and block
5458 	 * devices, we may need to do further work.
5459 	 *
5460 	 * We must wait for any I/O in progress to finish so that
5461 	 * all potential buffers on the dirty list will be visible.
5462 	 */
5463 	BO_LOCK(bo);
5464 	drain_output(vp);
5465 	BO_UNLOCK(bo);
5466 	return (0);
5467 }
5468 
5469 /*
5470  * Flush the dependencies associated with an inodedep.
5471  * Called with splbio blocked.
5472  */
5473 static int
5474 flush_inodedep_deps(mp, ino)
5475 	struct mount *mp;
5476 	ino_t ino;
5477 {
5478 	struct inodedep *inodedep;
5479 	int error, waitfor;
5480 
5481 	/*
5482 	 * This work is done in two passes. The first pass grabs most
5483 	 * of the buffers and begins asynchronously writing them. The
5484 	 * only way to wait for these asynchronous writes is to sleep
5485 	 * on the filesystem vnode which may stay busy for a long time
5486 	 * if the filesystem is active. So, instead, we make a second
5487 	 * pass over the dependencies blocking on each write. In the
5488 	 * usual case we will be blocking against a write that we
5489 	 * initiated, so when it is done the dependency will have been
5490 	 * resolved. Thus the second pass is expected to end quickly.
5491 	 * We give a brief window at the top of the loop to allow
5492 	 * any pending I/O to complete.
5493 	 */
5494 	for (error = 0, waitfor = MNT_NOWAIT; ; ) {
5495 		if (error)
5496 			return (error);
5497 		FREE_LOCK(&lk);
5498 		ACQUIRE_LOCK(&lk);
5499 		if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
5500 			return (0);
5501 		if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
5502 		    flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
5503 		    flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
5504 		    flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
5505 			continue;
5506 		/*
5507 		 * If pass2, we are done, otherwise do pass 2.
5508 		 */
5509 		if (waitfor == MNT_WAIT)
5510 			break;
5511 		waitfor = MNT_WAIT;
5512 	}
5513 	/*
5514 	 * Try freeing inodedep in case all dependencies have been removed.
5515 	 */
5516 	if (inodedep_lookup(mp, ino, 0, &inodedep) != 0)
5517 		(void) free_inodedep(inodedep);
5518 	return (0);
5519 }
5520 
5521 /*
5522  * Flush an inode dependency list.
5523  * Called with splbio blocked.
5524  */
5525 static int
5526 flush_deplist(listhead, waitfor, errorp)
5527 	struct allocdirectlst *listhead;
5528 	int waitfor;
5529 	int *errorp;
5530 {
5531 	struct allocdirect *adp;
5532 	struct buf *bp;
5533 
5534 	mtx_assert(&lk, MA_OWNED);
5535 	TAILQ_FOREACH(adp, listhead, ad_next) {
5536 		if (adp->ad_state & DEPCOMPLETE)
5537 			continue;
5538 		bp = adp->ad_buf;
5539 		bp = getdirtybuf(bp, &lk, waitfor);
5540 		if (bp == NULL) {
5541 			if (waitfor == MNT_NOWAIT)
5542 				continue;
5543 			return (1);
5544 		}
5545 		FREE_LOCK(&lk);
5546 		if (waitfor == MNT_NOWAIT) {
5547 			bawrite(bp);
5548 		} else if ((*errorp = bwrite(bp)) != 0) {
5549 			ACQUIRE_LOCK(&lk);
5550 			return (1);
5551 		}
5552 		ACQUIRE_LOCK(&lk);
5553 		return (1);
5554 	}
5555 	return (0);
5556 }
5557 
5558 /*
5559  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
5560  * Called with splbio blocked.
5561  */
5562 static int
5563 flush_pagedep_deps(pvp, mp, diraddhdp)
5564 	struct vnode *pvp;
5565 	struct mount *mp;
5566 	struct diraddhd *diraddhdp;
5567 {
5568 	struct inodedep *inodedep;
5569 	struct ufsmount *ump;
5570 	struct diradd *dap;
5571 	struct vnode *vp;
5572 	struct bufobj *bo;
5573 	int error = 0;
5574 	struct buf *bp;
5575 	ino_t inum;
5576 	struct worklist *wk;
5577 
5578 	ump = VFSTOUFS(mp);
5579 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
5580 		/*
5581 		 * Flush ourselves if this directory entry
5582 		 * has a MKDIR_PARENT dependency.
5583 		 */
5584 		if (dap->da_state & MKDIR_PARENT) {
5585 			FREE_LOCK(&lk);
5586 			if ((error = ffs_update(pvp, 1)) != 0)
5587 				break;
5588 			ACQUIRE_LOCK(&lk);
5589 			/*
5590 			 * If that cleared dependencies, go on to next.
5591 			 */
5592 			if (dap != LIST_FIRST(diraddhdp))
5593 				continue;
5594 			if (dap->da_state & MKDIR_PARENT)
5595 				panic("flush_pagedep_deps: MKDIR_PARENT");
5596 		}
5597 		/*
5598 		 * A newly allocated directory must have its "." and
5599 		 * ".." entries written out before its name can be
5600 		 * committed in its parent. We do not want or need
5601 		 * the full semantics of a synchronous ffs_syncvnode as
5602 		 * that may end up here again, once for each directory
5603 		 * level in the filesystem. Instead, we push the blocks
5604 		 * and wait for them to clear. We have to fsync twice
5605 		 * because the first call may choose to defer blocks
5606 		 * that still have dependencies, but deferral will
5607 		 * happen at most once.
5608 		 */
5609 		inum = dap->da_newinum;
5610 		if (dap->da_state & MKDIR_BODY) {
5611 			FREE_LOCK(&lk);
5612 			if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
5613 			    FFSV_FORCEINSMQ)))
5614 				break;
5615 			if ((error=ffs_syncvnode(vp, MNT_NOWAIT)) ||
5616 			    (error=ffs_syncvnode(vp, MNT_NOWAIT))) {
5617 				vput(vp);
5618 				break;
5619 			}
5620 			bo = &vp->v_bufobj;
5621 			BO_LOCK(bo);
5622 			drain_output(vp);
5623 			/*
5624 			 * If first block is still dirty with a D_MKDIR
5625 			 * dependency then it needs to be written now.
5626 			 */
5627 			for (;;) {
5628 				error = 0;
5629 				bp = gbincore(bo, 0);
5630 				if (bp == NULL)
5631 					break;	/* First block not present */
5632 				error = BUF_LOCK(bp,
5633 						 LK_EXCLUSIVE |
5634 						 LK_SLEEPFAIL |
5635 						 LK_INTERLOCK,
5636 						 BO_MTX(bo));
5637 				BO_LOCK(bo);
5638 				if (error == ENOLCK)
5639 					continue;	/* Slept, retry */
5640 				if (error != 0)
5641 					break;		/* Failed */
5642 				if ((bp->b_flags & B_DELWRI) == 0) {
5643 					BUF_UNLOCK(bp);
5644 					break;	/* Buffer not dirty */
5645 				}
5646 				for (wk = LIST_FIRST(&bp->b_dep);
5647 				     wk != NULL;
5648 				     wk = LIST_NEXT(wk, wk_list))
5649 					if (wk->wk_type == D_MKDIR)
5650 						break;
5651 				if (wk == NULL)
5652 					BUF_UNLOCK(bp);	/* Dependency gone */
5653 				else {
5654 					/*
5655 					 * D_MKDIR dependency remains,
5656 					 * must write buffer to stable
5657 					 * storage.
5658 					 */
5659 					BO_UNLOCK(bo);
5660 					bremfree(bp);
5661 					error = bwrite(bp);
5662 					BO_LOCK(bo);
5663 				}
5664 				break;
5665 			}
5666 			BO_UNLOCK(bo);
5667 			vput(vp);
5668 			if (error != 0)
5669 				break;	/* Flushing of first block failed */
5670 			ACQUIRE_LOCK(&lk);
5671 			/*
5672 			 * If that cleared dependencies, go on to next.
5673 			 */
5674 			if (dap != LIST_FIRST(diraddhdp))
5675 				continue;
5676 			if (dap->da_state & MKDIR_BODY)
5677 				panic("flush_pagedep_deps: MKDIR_BODY");
5678 		}
5679 		/*
5680 		 * Flush the inode on which the directory entry depends.
5681 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
5682 		 * the only remaining dependency is that the updated inode
5683 		 * count must get pushed to disk. The inode has already
5684 		 * been pushed into its inode buffer (via VOP_UPDATE) at
5685 		 * the time of the reference count change. So we need only
5686 		 * locate that buffer, ensure that there will be no rollback
5687 		 * caused by a bitmap dependency, then write the inode buffer.
5688 		 */
5689 retry:
5690 		if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
5691 			panic("flush_pagedep_deps: lost inode");
5692 		/*
5693 		 * If the inode still has bitmap dependencies,
5694 		 * push them to disk.
5695 		 */
5696 		if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5697 			bp = inodedep->id_buf;
5698 			bp = getdirtybuf(bp, &lk, MNT_WAIT);
5699 			if (bp == NULL)
5700 				goto retry;
5701 			FREE_LOCK(&lk);
5702 			if ((error = bwrite(bp)) != 0)
5703 				break;
5704 			ACQUIRE_LOCK(&lk);
5705 			if (dap != LIST_FIRST(diraddhdp))
5706 				continue;
5707 		}
5708 		/*
5709 		 * If the inode is still sitting in a buffer waiting
5710 		 * to be written, push it to disk.
5711 		 */
5712 		FREE_LOCK(&lk);
5713 		if ((error = bread(ump->um_devvp,
5714 		    fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
5715 		    (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) {
5716 			brelse(bp);
5717 			break;
5718 		}
5719 		if ((error = bwrite(bp)) != 0)
5720 			break;
5721 		ACQUIRE_LOCK(&lk);
5722 		/*
5723 		 * If we have failed to get rid of all the dependencies
5724 		 * then something is seriously wrong.
5725 		 */
5726 		if (dap == LIST_FIRST(diraddhdp))
5727 			panic("flush_pagedep_deps: flush failed");
5728 	}
5729 	if (error)
5730 		ACQUIRE_LOCK(&lk);
5731 	return (error);
5732 }
5733 
5734 /*
5735  * A large burst of file addition or deletion activity can drive the
5736  * memory load excessively high. First attempt to slow things down
5737  * using the techniques below. If that fails, this routine requests
5738  * the offending operations to fall back to running synchronously
5739  * until the memory load returns to a reasonable level.
5740  */
5741 int
5742 softdep_slowdown(vp)
5743 	struct vnode *vp;
5744 {
5745 	int max_softdeps_hard;
5746 
5747 	ACQUIRE_LOCK(&lk);
5748 	max_softdeps_hard = max_softdeps * 11 / 10;
5749 	if (num_dirrem < max_softdeps_hard / 2 &&
5750 	    num_inodedep < max_softdeps_hard &&
5751 	    VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps) {
5752 		FREE_LOCK(&lk);
5753   		return (0);
5754 	}
5755 	if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps)
5756 		softdep_speedup();
5757 	stat_sync_limit_hit += 1;
5758 	FREE_LOCK(&lk);
5759 	return (1);
5760 }
5761 
5762 /*
5763  * Called by the allocation routines when they are about to fail
5764  * in the hope that we can free up some disk space.
5765  *
5766  * First check to see if the work list has anything on it. If it has,
5767  * clean up entries until we successfully free some space. Because this
5768  * process holds inodes locked, we cannot handle any remove requests
5769  * that might block on a locked inode as that could lead to deadlock.
5770  * If the worklist yields no free space, encourage the syncer daemon
5771  * to help us. In no event will we try for longer than tickdelay seconds.
5772  */
5773 int
5774 softdep_request_cleanup(fs, vp)
5775 	struct fs *fs;
5776 	struct vnode *vp;
5777 {
5778 	struct ufsmount *ump;
5779 	long starttime;
5780 	ufs2_daddr_t needed;
5781 	int error;
5782 
5783 	ump = VTOI(vp)->i_ump;
5784 	mtx_assert(UFS_MTX(ump), MA_OWNED);
5785 	needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize;
5786 	starttime = time_second + tickdelay;
5787 	/*
5788 	 * If we are being called because of a process doing a
5789 	 * copy-on-write, then it is not safe to update the vnode
5790 	 * as we may recurse into the copy-on-write routine.
5791 	 */
5792 	if (!(curthread->td_pflags & TDP_COWINPROGRESS)) {
5793 		UFS_UNLOCK(ump);
5794 		error = ffs_update(vp, 1);
5795 		UFS_LOCK(ump);
5796 		if (error != 0)
5797 			return (0);
5798 	}
5799 	while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) {
5800 		if (time_second > starttime)
5801 			return (0);
5802 		UFS_UNLOCK(ump);
5803 		ACQUIRE_LOCK(&lk);
5804 		if (ump->softdep_on_worklist > 0 &&
5805 		    process_worklist_item(UFSTOVFS(ump), LK_NOWAIT) != -1) {
5806 			stat_worklist_push += 1;
5807 			FREE_LOCK(&lk);
5808 			UFS_LOCK(ump);
5809 			continue;
5810 		}
5811 		request_cleanup(UFSTOVFS(ump), FLUSH_REMOVE_WAIT);
5812 		FREE_LOCK(&lk);
5813 		UFS_LOCK(ump);
5814 	}
5815 	return (1);
5816 }
5817 
5818 /*
5819  * If memory utilization has gotten too high, deliberately slow things
5820  * down and speed up the I/O processing.
5821  */
5822 extern struct thread *syncertd;
5823 static int
5824 request_cleanup(mp, resource)
5825 	struct mount *mp;
5826 	int resource;
5827 {
5828 	struct thread *td = curthread;
5829 	struct ufsmount *ump;
5830 
5831 	mtx_assert(&lk, MA_OWNED);
5832 	/*
5833 	 * We never hold up the filesystem syncer or buf daemon.
5834 	 */
5835 	if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF))
5836 		return (0);
5837 	ump = VFSTOUFS(mp);
5838 	/*
5839 	 * First check to see if the work list has gotten backlogged.
5840 	 * If it has, co-opt this process to help clean up two entries.
5841 	 * Because this process may hold inodes locked, we cannot
5842 	 * handle any remove requests that might block on a locked
5843 	 * inode as that could lead to deadlock.  We set TDP_SOFTDEP
5844 	 * to avoid recursively processing the worklist.
5845 	 */
5846 	if (ump->softdep_on_worklist > max_softdeps / 10) {
5847 		td->td_pflags |= TDP_SOFTDEP;
5848 		process_worklist_item(mp, LK_NOWAIT);
5849 		process_worklist_item(mp, LK_NOWAIT);
5850 		td->td_pflags &= ~TDP_SOFTDEP;
5851 		stat_worklist_push += 2;
5852 		return(1);
5853 	}
5854 	/*
5855 	 * Next, we attempt to speed up the syncer process. If that
5856 	 * is successful, then we allow the process to continue.
5857 	 */
5858 	if (softdep_speedup() && resource != FLUSH_REMOVE_WAIT)
5859 		return(0);
5860 	/*
5861 	 * If we are resource constrained on inode dependencies, try
5862 	 * flushing some dirty inodes. Otherwise, we are constrained
5863 	 * by file deletions, so try accelerating flushes of directories
5864 	 * with removal dependencies. We would like to do the cleanup
5865 	 * here, but we probably hold an inode locked at this point and
5866 	 * that might deadlock against one that we try to clean. So,
5867 	 * the best that we can do is request the syncer daemon to do
5868 	 * the cleanup for us.
5869 	 */
5870 	switch (resource) {
5871 
5872 	case FLUSH_INODES:
5873 		stat_ino_limit_push += 1;
5874 		req_clear_inodedeps += 1;
5875 		stat_countp = &stat_ino_limit_hit;
5876 		break;
5877 
5878 	case FLUSH_REMOVE:
5879 	case FLUSH_REMOVE_WAIT:
5880 		stat_blk_limit_push += 1;
5881 		req_clear_remove += 1;
5882 		stat_countp = &stat_blk_limit_hit;
5883 		break;
5884 
5885 	default:
5886 		panic("request_cleanup: unknown type");
5887 	}
5888 	/*
5889 	 * Hopefully the syncer daemon will catch up and awaken us.
5890 	 * We wait at most tickdelay before proceeding in any case.
5891 	 */
5892 	proc_waiting += 1;
5893 	if (callout_pending(&softdep_callout) == FALSE)
5894 		callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2,
5895 		    pause_timer, 0);
5896 
5897 	msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0);
5898 	proc_waiting -= 1;
5899 	return (1);
5900 }
5901 
5902 /*
5903  * Awaken processes pausing in request_cleanup and clear proc_waiting
5904  * to indicate that there is no longer a timer running.
5905  */
5906 static void
5907 pause_timer(arg)
5908 	void *arg;
5909 {
5910 
5911 	/*
5912 	 * The callout_ API has acquired mtx and will hold it around this
5913 	 * function call.
5914 	 */
5915 	*stat_countp += 1;
5916 	wakeup_one(&proc_waiting);
5917 	if (proc_waiting > 0)
5918 		callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2,
5919 		    pause_timer, 0);
5920 }
5921 
5922 /*
5923  * Flush out a directory with at least one removal dependency in an effort to
5924  * reduce the number of dirrem, freefile, and freeblks dependency structures.
5925  */
5926 static void
5927 clear_remove(td)
5928 	struct thread *td;
5929 {
5930 	struct pagedep_hashhead *pagedephd;
5931 	struct pagedep *pagedep;
5932 	static int next = 0;
5933 	struct mount *mp;
5934 	struct vnode *vp;
5935 	struct bufobj *bo;
5936 	int error, cnt;
5937 	ino_t ino;
5938 
5939 	mtx_assert(&lk, MA_OWNED);
5940 
5941 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
5942 		pagedephd = &pagedep_hashtbl[next++];
5943 		if (next >= pagedep_hash)
5944 			next = 0;
5945 		LIST_FOREACH(pagedep, pagedephd, pd_hash) {
5946 			if (LIST_EMPTY(&pagedep->pd_dirremhd))
5947 				continue;
5948 			mp = pagedep->pd_list.wk_mp;
5949 			ino = pagedep->pd_ino;
5950 			if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5951 				continue;
5952 			FREE_LOCK(&lk);
5953 			if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
5954 			     FFSV_FORCEINSMQ))) {
5955 				softdep_error("clear_remove: vget", error);
5956 				vn_finished_write(mp);
5957 				ACQUIRE_LOCK(&lk);
5958 				return;
5959 			}
5960 			if ((error = ffs_syncvnode(vp, MNT_NOWAIT)))
5961 				softdep_error("clear_remove: fsync", error);
5962 			bo = &vp->v_bufobj;
5963 			BO_LOCK(bo);
5964 			drain_output(vp);
5965 			BO_UNLOCK(bo);
5966 			vput(vp);
5967 			vn_finished_write(mp);
5968 			ACQUIRE_LOCK(&lk);
5969 			return;
5970 		}
5971 	}
5972 }
5973 
5974 /*
5975  * Clear out a block of dirty inodes in an effort to reduce
5976  * the number of inodedep dependency structures.
5977  */
5978 static void
5979 clear_inodedeps(td)
5980 	struct thread *td;
5981 {
5982 	struct inodedep_hashhead *inodedephd;
5983 	struct inodedep *inodedep;
5984 	static int next = 0;
5985 	struct mount *mp;
5986 	struct vnode *vp;
5987 	struct fs *fs;
5988 	int error, cnt;
5989 	ino_t firstino, lastino, ino;
5990 
5991 	mtx_assert(&lk, MA_OWNED);
5992 	/*
5993 	 * Pick a random inode dependency to be cleared.
5994 	 * We will then gather up all the inodes in its block
5995 	 * that have dependencies and flush them out.
5996 	 */
5997 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
5998 		inodedephd = &inodedep_hashtbl[next++];
5999 		if (next >= inodedep_hash)
6000 			next = 0;
6001 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
6002 			break;
6003 	}
6004 	if (inodedep == NULL)
6005 		return;
6006 	fs = inodedep->id_fs;
6007 	mp = inodedep->id_list.wk_mp;
6008 	/*
6009 	 * Find the last inode in the block with dependencies.
6010 	 */
6011 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
6012 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
6013 		if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0)
6014 			break;
6015 	/*
6016 	 * Asynchronously push all but the last inode with dependencies.
6017 	 * Synchronously push the last inode with dependencies to ensure
6018 	 * that the inode block gets written to free up the inodedeps.
6019 	 */
6020 	for (ino = firstino; ino <= lastino; ino++) {
6021 		if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
6022 			continue;
6023 		if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
6024 			continue;
6025 		FREE_LOCK(&lk);
6026 		if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
6027 		    FFSV_FORCEINSMQ)) != 0) {
6028 			softdep_error("clear_inodedeps: vget", error);
6029 			vn_finished_write(mp);
6030 			ACQUIRE_LOCK(&lk);
6031 			return;
6032 		}
6033 		if (ino == lastino) {
6034 			if ((error = ffs_syncvnode(vp, MNT_WAIT)))
6035 				softdep_error("clear_inodedeps: fsync1", error);
6036 		} else {
6037 			if ((error = ffs_syncvnode(vp, MNT_NOWAIT)))
6038 				softdep_error("clear_inodedeps: fsync2", error);
6039 			BO_LOCK(&vp->v_bufobj);
6040 			drain_output(vp);
6041 			BO_UNLOCK(&vp->v_bufobj);
6042 		}
6043 		vput(vp);
6044 		vn_finished_write(mp);
6045 		ACQUIRE_LOCK(&lk);
6046 	}
6047 }
6048 
6049 /*
6050  * Function to determine if the buffer has outstanding dependencies
6051  * that will cause a roll-back if the buffer is written. If wantcount
6052  * is set, return number of dependencies, otherwise just yes or no.
6053  */
6054 static int
6055 softdep_count_dependencies(bp, wantcount)
6056 	struct buf *bp;
6057 	int wantcount;
6058 {
6059 	struct worklist *wk;
6060 	struct inodedep *inodedep;
6061 	struct indirdep *indirdep;
6062 	struct allocindir *aip;
6063 	struct pagedep *pagedep;
6064 	struct diradd *dap;
6065 	int i, retval;
6066 
6067 	retval = 0;
6068 	ACQUIRE_LOCK(&lk);
6069 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
6070 		switch (wk->wk_type) {
6071 
6072 		case D_INODEDEP:
6073 			inodedep = WK_INODEDEP(wk);
6074 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
6075 				/* bitmap allocation dependency */
6076 				retval += 1;
6077 				if (!wantcount)
6078 					goto out;
6079 			}
6080 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
6081 				/* direct block pointer dependency */
6082 				retval += 1;
6083 				if (!wantcount)
6084 					goto out;
6085 			}
6086 			if (TAILQ_FIRST(&inodedep->id_extupdt)) {
6087 				/* direct block pointer dependency */
6088 				retval += 1;
6089 				if (!wantcount)
6090 					goto out;
6091 			}
6092 			continue;
6093 
6094 		case D_INDIRDEP:
6095 			indirdep = WK_INDIRDEP(wk);
6096 
6097 			LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
6098 				/* indirect block pointer dependency */
6099 				retval += 1;
6100 				if (!wantcount)
6101 					goto out;
6102 			}
6103 			continue;
6104 
6105 		case D_PAGEDEP:
6106 			pagedep = WK_PAGEDEP(wk);
6107 			for (i = 0; i < DAHASHSZ; i++) {
6108 
6109 				LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
6110 					/* directory entry dependency */
6111 					retval += 1;
6112 					if (!wantcount)
6113 						goto out;
6114 				}
6115 			}
6116 			continue;
6117 
6118 		case D_BMSAFEMAP:
6119 		case D_ALLOCDIRECT:
6120 		case D_ALLOCINDIR:
6121 		case D_MKDIR:
6122 			/* never a dependency on these blocks */
6123 			continue;
6124 
6125 		default:
6126 			panic("softdep_check_for_rollback: Unexpected type %s",
6127 			    TYPENAME(wk->wk_type));
6128 			/* NOTREACHED */
6129 		}
6130 	}
6131 out:
6132 	FREE_LOCK(&lk);
6133 	return retval;
6134 }
6135 
6136 /*
6137  * Acquire exclusive access to a buffer.
6138  * Must be called with a locked mtx parameter.
6139  * Return acquired buffer or NULL on failure.
6140  */
6141 static struct buf *
6142 getdirtybuf(bp, mtx, waitfor)
6143 	struct buf *bp;
6144 	struct mtx *mtx;
6145 	int waitfor;
6146 {
6147 	int error;
6148 
6149 	mtx_assert(mtx, MA_OWNED);
6150 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) {
6151 		if (waitfor != MNT_WAIT)
6152 			return (NULL);
6153 		error = BUF_LOCK(bp,
6154 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx);
6155 		/*
6156 		 * Even if we sucessfully acquire bp here, we have dropped
6157 		 * mtx, which may violates our guarantee.
6158 		 */
6159 		if (error == 0)
6160 			BUF_UNLOCK(bp);
6161 		else if (error != ENOLCK)
6162 			panic("getdirtybuf: inconsistent lock: %d", error);
6163 		mtx_lock(mtx);
6164 		return (NULL);
6165 	}
6166 	if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
6167 		if (mtx == &lk && waitfor == MNT_WAIT) {
6168 			mtx_unlock(mtx);
6169 			BO_LOCK(bp->b_bufobj);
6170 			BUF_UNLOCK(bp);
6171 			if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
6172 				bp->b_vflags |= BV_BKGRDWAIT;
6173 				msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj),
6174 				       PRIBIO | PDROP, "getbuf", 0);
6175 			} else
6176 				BO_UNLOCK(bp->b_bufobj);
6177 			mtx_lock(mtx);
6178 			return (NULL);
6179 		}
6180 		BUF_UNLOCK(bp);
6181 		if (waitfor != MNT_WAIT)
6182 			return (NULL);
6183 		/*
6184 		 * The mtx argument must be bp->b_vp's mutex in
6185 		 * this case.
6186 		 */
6187 #ifdef	DEBUG_VFS_LOCKS
6188 		if (bp->b_vp->v_type != VCHR)
6189 			ASSERT_BO_LOCKED(bp->b_bufobj);
6190 #endif
6191 		bp->b_vflags |= BV_BKGRDWAIT;
6192 		msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0);
6193 		return (NULL);
6194 	}
6195 	if ((bp->b_flags & B_DELWRI) == 0) {
6196 		BUF_UNLOCK(bp);
6197 		return (NULL);
6198 	}
6199 	bremfree(bp);
6200 	return (bp);
6201 }
6202 
6203 
6204 /*
6205  * Check if it is safe to suspend the file system now.  On entry,
6206  * the vnode interlock for devvp should be held.  Return 0 with
6207  * the mount interlock held if the file system can be suspended now,
6208  * otherwise return EAGAIN with the mount interlock held.
6209  */
6210 int
6211 softdep_check_suspend(struct mount *mp,
6212 		      struct vnode *devvp,
6213 		      int softdep_deps,
6214 		      int softdep_accdeps,
6215 		      int secondary_writes,
6216 		      int secondary_accwrites)
6217 {
6218 	struct bufobj *bo;
6219 	struct ufsmount *ump;
6220 	int error;
6221 
6222 	ump = VFSTOUFS(mp);
6223 	bo = &devvp->v_bufobj;
6224 	ASSERT_BO_LOCKED(bo);
6225 
6226 	for (;;) {
6227 		if (!TRY_ACQUIRE_LOCK(&lk)) {
6228 			BO_UNLOCK(bo);
6229 			ACQUIRE_LOCK(&lk);
6230 			FREE_LOCK(&lk);
6231 			BO_LOCK(bo);
6232 			continue;
6233 		}
6234 		MNT_ILOCK(mp);
6235 		if (mp->mnt_secondary_writes != 0) {
6236 			FREE_LOCK(&lk);
6237 			BO_UNLOCK(bo);
6238 			msleep(&mp->mnt_secondary_writes,
6239 			       MNT_MTX(mp),
6240 			       (PUSER - 1) | PDROP, "secwr", 0);
6241 			BO_LOCK(bo);
6242 			continue;
6243 		}
6244 		break;
6245 	}
6246 
6247 	/*
6248 	 * Reasons for needing more work before suspend:
6249 	 * - Dirty buffers on devvp.
6250 	 * - Softdep activity occurred after start of vnode sync loop
6251 	 * - Secondary writes occurred after start of vnode sync loop
6252 	 */
6253 	error = 0;
6254 	if (bo->bo_numoutput > 0 ||
6255 	    bo->bo_dirty.bv_cnt > 0 ||
6256 	    softdep_deps != 0 ||
6257 	    ump->softdep_deps != 0 ||
6258 	    softdep_accdeps != ump->softdep_accdeps ||
6259 	    secondary_writes != 0 ||
6260 	    mp->mnt_secondary_writes != 0 ||
6261 	    secondary_accwrites != mp->mnt_secondary_accwrites)
6262 		error = EAGAIN;
6263 	FREE_LOCK(&lk);
6264 	BO_UNLOCK(bo);
6265 	return (error);
6266 }
6267 
6268 
6269 /*
6270  * Get the number of dependency structures for the file system, both
6271  * the current number and the total number allocated.  These will
6272  * later be used to detect that softdep processing has occurred.
6273  */
6274 void
6275 softdep_get_depcounts(struct mount *mp,
6276 		      int *softdep_depsp,
6277 		      int *softdep_accdepsp)
6278 {
6279 	struct ufsmount *ump;
6280 
6281 	ump = VFSTOUFS(mp);
6282 	ACQUIRE_LOCK(&lk);
6283 	*softdep_depsp = ump->softdep_deps;
6284 	*softdep_accdepsp = ump->softdep_accdeps;
6285 	FREE_LOCK(&lk);
6286 }
6287 
6288 /*
6289  * Wait for pending output on a vnode to complete.
6290  * Must be called with vnode lock and interlock locked.
6291  *
6292  * XXX: Should just be a call to bufobj_wwait().
6293  */
6294 static void
6295 drain_output(vp)
6296 	struct vnode *vp;
6297 {
6298 	struct bufobj *bo;
6299 
6300 	bo = &vp->v_bufobj;
6301 	ASSERT_VOP_LOCKED(vp, "drain_output");
6302 	ASSERT_BO_LOCKED(bo);
6303 
6304 	while (bo->bo_numoutput) {
6305 		bo->bo_flag |= BO_WWAIT;
6306 		msleep((caddr_t)&bo->bo_numoutput,
6307 		    BO_MTX(bo), PRIBIO + 1, "drainvp", 0);
6308 	}
6309 }
6310 
6311 /*
6312  * Called whenever a buffer that is being invalidated or reallocated
6313  * contains dependencies. This should only happen if an I/O error has
6314  * occurred. The routine is called with the buffer locked.
6315  */
6316 static void
6317 softdep_deallocate_dependencies(bp)
6318 	struct buf *bp;
6319 {
6320 
6321 	if ((bp->b_ioflags & BIO_ERROR) == 0)
6322 		panic("softdep_deallocate_dependencies: dangling deps");
6323 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
6324 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
6325 }
6326 
6327 /*
6328  * Function to handle asynchronous write errors in the filesystem.
6329  */
6330 static void
6331 softdep_error(func, error)
6332 	char *func;
6333 	int error;
6334 {
6335 
6336 	/* XXX should do something better! */
6337 	printf("%s: got error %d while accessing filesystem\n", func, error);
6338 }
6339 
6340 #ifdef DDB
6341 
6342 DB_SHOW_COMMAND(inodedeps, db_show_inodedeps)
6343 {
6344 	struct inodedep_hashhead *inodedephd;
6345 	struct inodedep *inodedep;
6346 	struct fs *fs;
6347 	int cnt;
6348 
6349 	fs = have_addr ? (struct fs *)addr : NULL;
6350 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
6351 		inodedephd = &inodedep_hashtbl[cnt];
6352 		LIST_FOREACH(inodedep, inodedephd, id_hash) {
6353 			if (fs != NULL && fs != inodedep->id_fs)
6354 				continue;
6355 			db_printf("%p fs %p st %x ino %jd inoblk %jd\n",
6356 			    inodedep, inodedep->id_fs, inodedep->id_state,
6357 			    (intmax_t)inodedep->id_ino,
6358 			    (intmax_t)fsbtodb(inodedep->id_fs,
6359 			    ino_to_fsba(inodedep->id_fs, inodedep->id_ino)));
6360 		}
6361 	}
6362 }
6363 
6364 #endif /* DDB */
6365 
6366 #endif /* SOFTUPDATES */
6367