xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision a3cf0ef5a295c885c895fabfd56470c0d1db322d)
1 /*-
2  * Copyright 1998, 2000 Marshall Kirk McKusick.
3  * Copyright 2009, 2010 Jeffrey W. Roberson <jeff@FreeBSD.org>
4  * All rights reserved.
5  *
6  * The soft updates code is derived from the appendix of a University
7  * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
8  * "Soft Updates: A Solution to the Metadata Update Problem in File
9  * Systems", CSE-TR-254-95, August 1995).
10  *
11  * Further information about soft updates can be obtained from:
12  *
13  *	Marshall Kirk McKusick		http://www.mckusick.com/softdep/
14  *	1614 Oxford Street		mckusick@mckusick.com
15  *	Berkeley, CA 94709-1608		+1-510-843-9542
16  *	USA
17  *
18  * Redistribution and use in source and binary forms, with or without
19  * modification, are permitted provided that the following conditions
20  * are met:
21  *
22  * 1. Redistributions of source code must retain the above copyright
23  *    notice, this list of conditions and the following disclaimer.
24  * 2. Redistributions in binary form must reproduce the above copyright
25  *    notice, this list of conditions and the following disclaimer in the
26  *    documentation and/or other materials provided with the distribution.
27  *
28  * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR
29  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
30  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
31  * IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY DIRECT, INDIRECT,
32  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
33  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
34  * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
35  * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR
36  * TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
37  * USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
38  *
39  *	from: @(#)ffs_softdep.c	9.59 (McKusick) 6/21/00
40  */
41 
42 #include <sys/cdefs.h>
43 __FBSDID("$FreeBSD$");
44 
45 #include "opt_ffs.h"
46 #include "opt_ddb.h"
47 
48 /*
49  * For now we want the safety net that the DEBUG flag provides.
50  */
51 #ifndef DEBUG
52 #define DEBUG
53 #endif
54 
55 #include <sys/param.h>
56 #include <sys/kernel.h>
57 #include <sys/systm.h>
58 #include <sys/bio.h>
59 #include <sys/buf.h>
60 #include <sys/kdb.h>
61 #include <sys/kthread.h>
62 #include <sys/lock.h>
63 #include <sys/malloc.h>
64 #include <sys/mount.h>
65 #include <sys/mutex.h>
66 #include <sys/namei.h>
67 #include <sys/proc.h>
68 #include <sys/stat.h>
69 #include <sys/sysctl.h>
70 #include <sys/syslog.h>
71 #include <sys/vnode.h>
72 #include <sys/conf.h>
73 #include <ufs/ufs/dir.h>
74 #include <ufs/ufs/extattr.h>
75 #include <ufs/ufs/quota.h>
76 #include <ufs/ufs/inode.h>
77 #include <ufs/ufs/ufsmount.h>
78 #include <ufs/ffs/fs.h>
79 #include <ufs/ffs/softdep.h>
80 #include <ufs/ffs/ffs_extern.h>
81 #include <ufs/ufs/ufs_extern.h>
82 
83 #include <vm/vm.h>
84 
85 #include <ddb/ddb.h>
86 
87 #ifndef SOFTUPDATES
88 
89 int
90 softdep_flushfiles(oldmnt, flags, td)
91 	struct mount *oldmnt;
92 	int flags;
93 	struct thread *td;
94 {
95 
96 	panic("softdep_flushfiles called");
97 }
98 
99 int
100 softdep_mount(devvp, mp, fs, cred)
101 	struct vnode *devvp;
102 	struct mount *mp;
103 	struct fs *fs;
104 	struct ucred *cred;
105 {
106 
107 	return (0);
108 }
109 
110 void
111 softdep_initialize()
112 {
113 
114 	return;
115 }
116 
117 void
118 softdep_uninitialize()
119 {
120 
121 	return;
122 }
123 
124 void
125 softdep_unmount(mp)
126 	struct mount *mp;
127 {
128 
129 }
130 
131 void
132 softdep_setup_sbupdate(ump, fs, bp)
133 	struct ufsmount *ump;
134 	struct fs *fs;
135 	struct buf *bp;
136 {
137 }
138 
139 void
140 softdep_setup_inomapdep(bp, ip, newinum)
141 	struct buf *bp;
142 	struct inode *ip;
143 	ino_t newinum;
144 {
145 
146 	panic("softdep_setup_inomapdep called");
147 }
148 
149 void
150 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags)
151 	struct buf *bp;
152 	struct mount *mp;
153 	ufs2_daddr_t newblkno;
154 	int frags;
155 	int oldfrags;
156 {
157 
158 	panic("softdep_setup_blkmapdep called");
159 }
160 
161 void
162 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
163 	struct inode *ip;
164 	ufs_lbn_t lbn;
165 	ufs2_daddr_t newblkno;
166 	ufs2_daddr_t oldblkno;
167 	long newsize;
168 	long oldsize;
169 	struct buf *bp;
170 {
171 
172 	panic("softdep_setup_allocdirect called");
173 }
174 
175 void
176 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
177 	struct inode *ip;
178 	ufs_lbn_t lbn;
179 	ufs2_daddr_t newblkno;
180 	ufs2_daddr_t oldblkno;
181 	long newsize;
182 	long oldsize;
183 	struct buf *bp;
184 {
185 
186 	panic("softdep_setup_allocext called");
187 }
188 
189 void
190 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
191 	struct inode *ip;
192 	ufs_lbn_t lbn;
193 	struct buf *bp;
194 	int ptrno;
195 	ufs2_daddr_t newblkno;
196 	ufs2_daddr_t oldblkno;
197 	struct buf *nbp;
198 {
199 
200 	panic("softdep_setup_allocindir_page called");
201 }
202 
203 void
204 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
205 	struct buf *nbp;
206 	struct inode *ip;
207 	struct buf *bp;
208 	int ptrno;
209 	ufs2_daddr_t newblkno;
210 {
211 
212 	panic("softdep_setup_allocindir_meta called");
213 }
214 
215 void
216 softdep_setup_freeblocks(ip, length, flags)
217 	struct inode *ip;
218 	off_t length;
219 	int flags;
220 {
221 
222 	panic("softdep_setup_freeblocks called");
223 }
224 
225 void
226 softdep_freefile(pvp, ino, mode)
227 		struct vnode *pvp;
228 		ino_t ino;
229 		int mode;
230 {
231 
232 	panic("softdep_freefile called");
233 }
234 
235 int
236 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
237 	struct buf *bp;
238 	struct inode *dp;
239 	off_t diroffset;
240 	ino_t newinum;
241 	struct buf *newdirbp;
242 	int isnewblk;
243 {
244 
245 	panic("softdep_setup_directory_add called");
246 }
247 
248 void
249 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize)
250 	struct buf *bp;
251 	struct inode *dp;
252 	caddr_t base;
253 	caddr_t oldloc;
254 	caddr_t newloc;
255 	int entrysize;
256 {
257 
258 	panic("softdep_change_directoryentry_offset called");
259 }
260 
261 void
262 softdep_setup_remove(bp, dp, ip, isrmdir)
263 	struct buf *bp;
264 	struct inode *dp;
265 	struct inode *ip;
266 	int isrmdir;
267 {
268 
269 	panic("softdep_setup_remove called");
270 }
271 
272 void
273 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
274 	struct buf *bp;
275 	struct inode *dp;
276 	struct inode *ip;
277 	ino_t newinum;
278 	int isrmdir;
279 {
280 
281 	panic("softdep_setup_directory_change called");
282 }
283 
284 void *
285 softdep_setup_trunc(vp, length, flags)
286 	struct vnode *vp;
287 	off_t length;
288 	int flags;
289 {
290 
291 	panic("%s called", __FUNCTION__);
292 
293 	return (NULL);
294 }
295 
296 int
297 softdep_complete_trunc(vp, cookie)
298 	struct vnode *vp;
299 	void *cookie;
300 {
301 
302 	panic("%s called", __FUNCTION__);
303 
304 	return (0);
305 }
306 
307 void
308 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd)
309 	struct mount *mp;
310 	struct buf *bp;
311 	ufs2_daddr_t blkno;
312 	int frags;
313 	struct workhead *wkhd;
314 {
315 
316 	panic("%s called", __FUNCTION__);
317 }
318 
319 void
320 softdep_setup_inofree(mp, bp, ino, wkhd)
321 	struct mount *mp;
322 	struct buf *bp;
323 	ino_t ino;
324 	struct workhead *wkhd;
325 {
326 
327 	panic("%s called", __FUNCTION__);
328 }
329 
330 void
331 softdep_setup_unlink(dp, ip)
332 	struct inode *dp;
333 	struct inode *ip;
334 {
335 
336 	panic("%s called", __FUNCTION__);
337 }
338 
339 void
340 softdep_setup_link(dp, ip)
341 	struct inode *dp;
342 	struct inode *ip;
343 {
344 
345 	panic("%s called", __FUNCTION__);
346 }
347 
348 void
349 softdep_revert_link(dp, ip)
350 	struct inode *dp;
351 	struct inode *ip;
352 {
353 
354 	panic("%s called", __FUNCTION__);
355 }
356 
357 void
358 softdep_setup_rmdir(dp, ip)
359 	struct inode *dp;
360 	struct inode *ip;
361 {
362 
363 	panic("%s called", __FUNCTION__);
364 }
365 
366 void
367 softdep_revert_rmdir(dp, ip)
368 	struct inode *dp;
369 	struct inode *ip;
370 {
371 
372 	panic("%s called", __FUNCTION__);
373 }
374 
375 void
376 softdep_setup_create(dp, ip)
377 	struct inode *dp;
378 	struct inode *ip;
379 {
380 
381 	panic("%s called", __FUNCTION__);
382 }
383 
384 void
385 softdep_revert_create(dp, ip)
386 	struct inode *dp;
387 	struct inode *ip;
388 {
389 
390 	panic("%s called", __FUNCTION__);
391 }
392 
393 void
394 softdep_setup_mkdir(dp, ip)
395 	struct inode *dp;
396 	struct inode *ip;
397 {
398 
399 	panic("%s called", __FUNCTION__);
400 }
401 
402 void
403 softdep_revert_mkdir(dp, ip)
404 	struct inode *dp;
405 	struct inode *ip;
406 {
407 
408 	panic("%s called", __FUNCTION__);
409 }
410 
411 void
412 softdep_setup_dotdot_link(dp, ip)
413 	struct inode *dp;
414 	struct inode *ip;
415 {
416 
417 	panic("%s called", __FUNCTION__);
418 }
419 
420 int
421 softdep_prealloc(vp, waitok)
422 	struct vnode *vp;
423 	int waitok;
424 {
425 
426 	panic("%s called", __FUNCTION__);
427 
428 	return (0);
429 }
430 
431 int
432 softdep_journal_lookup(mp, vpp)
433 	struct mount *mp;
434 	struct vnode **vpp;
435 {
436 
437 	return (ENOENT);
438 }
439 
440 void
441 softdep_change_linkcnt(ip)
442 	struct inode *ip;
443 {
444 
445 	panic("softdep_change_linkcnt called");
446 }
447 
448 void
449 softdep_load_inodeblock(ip)
450 	struct inode *ip;
451 {
452 
453 	panic("softdep_load_inodeblock called");
454 }
455 
456 void
457 softdep_update_inodeblock(ip, bp, waitfor)
458 	struct inode *ip;
459 	struct buf *bp;
460 	int waitfor;
461 {
462 
463 	panic("softdep_update_inodeblock called");
464 }
465 
466 int
467 softdep_fsync(vp)
468 	struct vnode *vp;	/* the "in_core" copy of the inode */
469 {
470 
471 	return (0);
472 }
473 
474 void
475 softdep_fsync_mountdev(vp)
476 	struct vnode *vp;
477 {
478 
479 	return;
480 }
481 
482 int
483 softdep_flushworklist(oldmnt, countp, td)
484 	struct mount *oldmnt;
485 	int *countp;
486 	struct thread *td;
487 {
488 
489 	*countp = 0;
490 	return (0);
491 }
492 
493 int
494 softdep_sync_metadata(struct vnode *vp)
495 {
496 
497 	return (0);
498 }
499 
500 int
501 softdep_slowdown(vp)
502 	struct vnode *vp;
503 {
504 
505 	panic("softdep_slowdown called");
506 }
507 
508 void
509 softdep_releasefile(ip)
510 	struct inode *ip;	/* inode with the zero effective link count */
511 {
512 
513 	panic("softdep_releasefile called");
514 }
515 
516 int
517 softdep_request_cleanup(fs, vp)
518 	struct fs *fs;
519 	struct vnode *vp;
520 {
521 
522 	return (0);
523 }
524 
525 int
526 softdep_check_suspend(struct mount *mp,
527 		      struct vnode *devvp,
528 		      int softdep_deps,
529 		      int softdep_accdeps,
530 		      int secondary_writes,
531 		      int secondary_accwrites)
532 {
533 	struct bufobj *bo;
534 	int error;
535 
536 	(void) softdep_deps,
537 	(void) softdep_accdeps;
538 
539 	bo = &devvp->v_bufobj;
540 	ASSERT_BO_LOCKED(bo);
541 
542 	MNT_ILOCK(mp);
543 	while (mp->mnt_secondary_writes != 0) {
544 		BO_UNLOCK(bo);
545 		msleep(&mp->mnt_secondary_writes, MNT_MTX(mp),
546 		    (PUSER - 1) | PDROP, "secwr", 0);
547 		BO_LOCK(bo);
548 		MNT_ILOCK(mp);
549 	}
550 
551 	/*
552 	 * Reasons for needing more work before suspend:
553 	 * - Dirty buffers on devvp.
554 	 * - Secondary writes occurred after start of vnode sync loop
555 	 */
556 	error = 0;
557 	if (bo->bo_numoutput > 0 ||
558 	    bo->bo_dirty.bv_cnt > 0 ||
559 	    secondary_writes != 0 ||
560 	    mp->mnt_secondary_writes != 0 ||
561 	    secondary_accwrites != mp->mnt_secondary_accwrites)
562 		error = EAGAIN;
563 	BO_UNLOCK(bo);
564 	return (error);
565 }
566 
567 void
568 softdep_get_depcounts(struct mount *mp,
569 		      int *softdepactivep,
570 		      int *softdepactiveaccp)
571 {
572 	(void) mp;
573 	*softdepactivep = 0;
574 	*softdepactiveaccp = 0;
575 }
576 
577 #else
578 /*
579  * These definitions need to be adapted to the system to which
580  * this file is being ported.
581  */
582 
583 #define M_SOFTDEP_FLAGS	(M_WAITOK)
584 
585 #define	D_PAGEDEP	0
586 #define	D_INODEDEP	1
587 #define	D_BMSAFEMAP	2
588 #define	D_NEWBLK	3
589 #define	D_ALLOCDIRECT	4
590 #define	D_INDIRDEP	5
591 #define	D_ALLOCINDIR	6
592 #define	D_FREEFRAG	7
593 #define	D_FREEBLKS	8
594 #define	D_FREEFILE	9
595 #define	D_DIRADD	10
596 #define	D_MKDIR		11
597 #define	D_DIRREM	12
598 #define	D_NEWDIRBLK	13
599 #define	D_FREEWORK	14
600 #define	D_FREEDEP	15
601 #define	D_JADDREF	16
602 #define	D_JREMREF	17
603 #define	D_JMVREF	18
604 #define	D_JNEWBLK	19
605 #define	D_JFREEBLK	20
606 #define	D_JFREEFRAG	21
607 #define	D_JSEG		22
608 #define	D_JSEGDEP	23
609 #define	D_SBDEP		24
610 #define	D_JTRUNC	25
611 #define	D_LAST		D_JTRUNC
612 
613 unsigned long dep_current[D_LAST + 1];
614 unsigned long dep_total[D_LAST + 1];
615 
616 
617 SYSCTL_NODE(_debug, OID_AUTO, softdep, CTLFLAG_RW, 0, "soft updates stats");
618 SYSCTL_NODE(_debug_softdep, OID_AUTO, total, CTLFLAG_RW, 0,
619     "total dependencies allocated");
620 SYSCTL_NODE(_debug_softdep, OID_AUTO, current, CTLFLAG_RW, 0,
621     "current dependencies allocated");
622 
623 #define	SOFTDEP_TYPE(type, str, long)					\
624     static MALLOC_DEFINE(M_ ## type, #str, long);			\
625     SYSCTL_LONG(_debug_softdep_total, OID_AUTO, str, CTLFLAG_RD,	\
626 	&dep_total[D_ ## type], 0, "");					\
627     SYSCTL_LONG(_debug_softdep_current, OID_AUTO, str, CTLFLAG_RD, 	\
628 	&dep_current[D_ ## type], 0, "");
629 
630 SOFTDEP_TYPE(PAGEDEP, pagedep, "File page dependencies");
631 SOFTDEP_TYPE(INODEDEP, inodedep, "Inode dependencies");
632 SOFTDEP_TYPE(BMSAFEMAP, bmsafemap,
633     "Block or frag allocated from cyl group map");
634 SOFTDEP_TYPE(NEWBLK, newblk, "New block or frag allocation dependency");
635 SOFTDEP_TYPE(ALLOCDIRECT, allocdirect, "Block or frag dependency for an inode");
636 SOFTDEP_TYPE(INDIRDEP, indirdep, "Indirect block dependencies");
637 SOFTDEP_TYPE(ALLOCINDIR, allocindir, "Block dependency for an indirect block");
638 SOFTDEP_TYPE(FREEFRAG, freefrag, "Previously used frag for an inode");
639 SOFTDEP_TYPE(FREEBLKS, freeblks, "Blocks freed from an inode");
640 SOFTDEP_TYPE(FREEFILE, freefile, "Inode deallocated");
641 SOFTDEP_TYPE(DIRADD, diradd, "New directory entry");
642 SOFTDEP_TYPE(MKDIR, mkdir, "New directory");
643 SOFTDEP_TYPE(DIRREM, dirrem, "Directory entry deleted");
644 SOFTDEP_TYPE(NEWDIRBLK, newdirblk, "Unclaimed new directory block");
645 SOFTDEP_TYPE(FREEWORK, freework, "free an inode block");
646 SOFTDEP_TYPE(FREEDEP, freedep, "track a block free");
647 SOFTDEP_TYPE(JADDREF, jaddref, "Journal inode ref add");
648 SOFTDEP_TYPE(JREMREF, jremref, "Journal inode ref remove");
649 SOFTDEP_TYPE(JMVREF, jmvref, "Journal inode ref move");
650 SOFTDEP_TYPE(JNEWBLK, jnewblk, "Journal new block");
651 SOFTDEP_TYPE(JFREEBLK, jfreeblk, "Journal free block");
652 SOFTDEP_TYPE(JFREEFRAG, jfreefrag, "Journal free frag");
653 SOFTDEP_TYPE(JSEG, jseg, "Journal segment");
654 SOFTDEP_TYPE(JSEGDEP, jsegdep, "Journal segment complete");
655 SOFTDEP_TYPE(SBDEP, sbdep, "Superblock write dependency");
656 SOFTDEP_TYPE(JTRUNC, jtrunc, "Journal inode truncation");
657 
658 static MALLOC_DEFINE(M_SAVEDINO, "savedino", "Saved inodes");
659 static MALLOC_DEFINE(M_JBLOCKS, "jblocks", "Journal block locations");
660 
661 /*
662  * translate from workitem type to memory type
663  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
664  */
665 static struct malloc_type *memtype[] = {
666 	M_PAGEDEP,
667 	M_INODEDEP,
668 	M_BMSAFEMAP,
669 	M_NEWBLK,
670 	M_ALLOCDIRECT,
671 	M_INDIRDEP,
672 	M_ALLOCINDIR,
673 	M_FREEFRAG,
674 	M_FREEBLKS,
675 	M_FREEFILE,
676 	M_DIRADD,
677 	M_MKDIR,
678 	M_DIRREM,
679 	M_NEWDIRBLK,
680 	M_FREEWORK,
681 	M_FREEDEP,
682 	M_JADDREF,
683 	M_JREMREF,
684 	M_JMVREF,
685 	M_JNEWBLK,
686 	M_JFREEBLK,
687 	M_JFREEFRAG,
688 	M_JSEG,
689 	M_JSEGDEP,
690 	M_SBDEP,
691 	M_JTRUNC
692 };
693 
694 #define DtoM(type) (memtype[type])
695 
696 /*
697  * Names of malloc types.
698  */
699 #define TYPENAME(type)  \
700 	((unsigned)(type) <= D_LAST ? memtype[type]->ks_shortdesc : "???")
701 /*
702  * End system adaptation definitions.
703  */
704 
705 #define	DOTDOT_OFFSET	offsetof(struct dirtemplate, dotdot_ino)
706 #define	DOT_OFFSET	offsetof(struct dirtemplate, dot_ino)
707 
708 /*
709  * Forward declarations.
710  */
711 struct inodedep_hashhead;
712 struct newblk_hashhead;
713 struct pagedep_hashhead;
714 struct bmsafemap_hashhead;
715 
716 /*
717  * Internal function prototypes.
718  */
719 static	void softdep_error(char *, int);
720 static	void drain_output(struct vnode *);
721 static	struct buf *getdirtybuf(struct buf *, struct mtx *, int);
722 static	void clear_remove(struct thread *);
723 static	void clear_inodedeps(struct thread *);
724 static	void unlinked_inodedep(struct mount *, struct inodedep *);
725 static	void clear_unlinked_inodedep(struct inodedep *);
726 static	struct inodedep *first_unlinked_inodedep(struct ufsmount *);
727 static	int flush_pagedep_deps(struct vnode *, struct mount *,
728 	    struct diraddhd *);
729 static	void free_pagedep(struct pagedep *);
730 static	int flush_newblk_dep(struct vnode *, struct mount *, ufs_lbn_t);
731 static	int flush_inodedep_deps(struct mount *, ino_t);
732 static	int flush_deplist(struct allocdirectlst *, int, int *);
733 static	int handle_written_filepage(struct pagedep *, struct buf *);
734 static	int handle_written_sbdep(struct sbdep *, struct buf *);
735 static	void initiate_write_sbdep(struct sbdep *);
736 static  void diradd_inode_written(struct diradd *, struct inodedep *);
737 static	int handle_written_indirdep(struct indirdep *, struct buf *,
738 	    struct buf**);
739 static	int handle_written_inodeblock(struct inodedep *, struct buf *);
740 static	int handle_written_bmsafemap(struct bmsafemap *, struct buf *);
741 static	void handle_written_jaddref(struct jaddref *);
742 static	void handle_written_jremref(struct jremref *);
743 static	void handle_written_jseg(struct jseg *, struct buf *);
744 static	void handle_written_jnewblk(struct jnewblk *);
745 static	void handle_written_jfreeblk(struct jfreeblk *);
746 static	void handle_written_jfreefrag(struct jfreefrag *);
747 static	void complete_jseg(struct jseg *);
748 static	void jseg_write(struct fs *, struct jblocks *, struct jseg *,
749 	    uint8_t *);
750 static	void jaddref_write(struct jaddref *, struct jseg *, uint8_t *);
751 static	void jremref_write(struct jremref *, struct jseg *, uint8_t *);
752 static	void jmvref_write(struct jmvref *, struct jseg *, uint8_t *);
753 static	void jtrunc_write(struct jtrunc *, struct jseg *, uint8_t *);
754 static	void jnewblk_write(struct jnewblk *, struct jseg *, uint8_t *);
755 static	void jfreeblk_write(struct jfreeblk *, struct jseg *, uint8_t *);
756 static	void jfreefrag_write(struct jfreefrag *, struct jseg *, uint8_t *);
757 static	inline void inoref_write(struct inoref *, struct jseg *,
758 	    struct jrefrec *);
759 static	void handle_allocdirect_partdone(struct allocdirect *,
760 	    struct workhead *);
761 static	void cancel_newblk(struct newblk *, struct workhead *);
762 static	void indirdep_complete(struct indirdep *);
763 static	void handle_allocindir_partdone(struct allocindir *);
764 static	void initiate_write_filepage(struct pagedep *, struct buf *);
765 static	void initiate_write_indirdep(struct indirdep*, struct buf *);
766 static	void handle_written_mkdir(struct mkdir *, int);
767 static	void initiate_write_bmsafemap(struct bmsafemap *, struct buf *);
768 static	void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
769 static	void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
770 static	void handle_workitem_freefile(struct freefile *);
771 static	void handle_workitem_remove(struct dirrem *, struct vnode *);
772 static	struct dirrem *newdirrem(struct buf *, struct inode *,
773 	    struct inode *, int, struct dirrem **);
774 static	void cancel_indirdep(struct indirdep *, struct buf *, struct inodedep *,
775 	    struct freeblks *);
776 static	void free_indirdep(struct indirdep *);
777 static	void free_diradd(struct diradd *, struct workhead *);
778 static	void merge_diradd(struct inodedep *, struct diradd *);
779 static	void complete_diradd(struct diradd *);
780 static	struct diradd *diradd_lookup(struct pagedep *, int);
781 static	struct jremref *cancel_diradd_dotdot(struct inode *, struct dirrem *,
782 	    struct jremref *);
783 static	struct jremref *cancel_mkdir_dotdot(struct inode *, struct dirrem *,
784 	    struct jremref *);
785 static	void cancel_diradd(struct diradd *, struct dirrem *, struct jremref *,
786 	    struct jremref *, struct jremref *);
787 static	void dirrem_journal(struct dirrem *, struct jremref *, struct jremref *,
788 	    struct jremref *);
789 static	void cancel_allocindir(struct allocindir *, struct inodedep *,
790 	    struct freeblks *);
791 static	void complete_mkdir(struct mkdir *);
792 static	void free_newdirblk(struct newdirblk *);
793 static	void free_jremref(struct jremref *);
794 static	void free_jaddref(struct jaddref *);
795 static	void free_jsegdep(struct jsegdep *);
796 static	void free_jseg(struct jseg *);
797 static	void free_jnewblk(struct jnewblk *);
798 static	void free_jfreeblk(struct jfreeblk *);
799 static	void free_jfreefrag(struct jfreefrag *);
800 static	void free_freedep(struct freedep *);
801 static	void journal_jremref(struct dirrem *, struct jremref *,
802 	    struct inodedep *);
803 static	void cancel_jnewblk(struct jnewblk *, struct workhead *);
804 static	int cancel_jaddref(struct jaddref *, struct inodedep *,
805 	    struct workhead *);
806 static	void cancel_jfreefrag(struct jfreefrag *);
807 static	void indir_trunc(struct freework *, ufs2_daddr_t, ufs_lbn_t);
808 static	int deallocate_dependencies(struct buf *, struct inodedep *,
809 	    struct freeblks *);
810 static	void free_newblk(struct newblk *);
811 static	void cancel_allocdirect(struct allocdirectlst *,
812 	    struct allocdirect *, struct freeblks *, int);
813 static	int check_inode_unwritten(struct inodedep *);
814 static	int free_inodedep(struct inodedep *);
815 static	void freework_freeblock(struct freework *);
816 static	void handle_workitem_freeblocks(struct freeblks *, int);
817 static	void handle_complete_freeblocks(struct freeblks *);
818 static	void handle_workitem_indirblk(struct freework *);
819 static	void handle_written_freework(struct freework *);
820 static	void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
821 static	void setup_allocindir_phase2(struct buf *, struct inode *,
822 	    struct inodedep *, struct allocindir *, ufs_lbn_t);
823 static	struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
824 	    ufs2_daddr_t, ufs_lbn_t);
825 static	void handle_workitem_freefrag(struct freefrag *);
826 static	struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long,
827 	    ufs_lbn_t);
828 static	void allocdirect_merge(struct allocdirectlst *,
829 	    struct allocdirect *, struct allocdirect *);
830 static	struct freefrag *allocindir_merge(struct allocindir *,
831 	    struct allocindir *);
832 static	int bmsafemap_find(struct bmsafemap_hashhead *, struct mount *, int,
833 	    struct bmsafemap **);
834 static	struct bmsafemap *bmsafemap_lookup(struct mount *, struct buf *,
835 	    int cg);
836 static	int newblk_find(struct newblk_hashhead *, struct mount *, ufs2_daddr_t,
837 	    int, struct newblk **);
838 static	int newblk_lookup(struct mount *, ufs2_daddr_t, int, struct newblk **);
839 static	int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t,
840 	    struct inodedep **);
841 static	int inodedep_lookup(struct mount *, ino_t, int, struct inodedep **);
842 static	int pagedep_lookup(struct mount *, ino_t, ufs_lbn_t, int,
843 	    struct pagedep **);
844 static	int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t,
845 	    struct mount *mp, int, struct pagedep **);
846 static	void pause_timer(void *);
847 static	int request_cleanup(struct mount *, int);
848 static	int process_worklist_item(struct mount *, int);
849 static	void process_removes(struct vnode *);
850 static	void jwork_move(struct workhead *, struct workhead *);
851 static	void add_to_worklist(struct worklist *, int);
852 static	void remove_from_worklist(struct worklist *);
853 static	void softdep_flush(void);
854 static	int softdep_speedup(void);
855 static	void worklist_speedup(void);
856 static	int journal_mount(struct mount *, struct fs *, struct ucred *);
857 static	void journal_unmount(struct mount *);
858 static	int journal_space(struct ufsmount *, int);
859 static	void journal_suspend(struct ufsmount *);
860 static	int journal_unsuspend(struct ufsmount *ump);
861 static	void softdep_prelink(struct vnode *, struct vnode *);
862 static	void add_to_journal(struct worklist *);
863 static	void remove_from_journal(struct worklist *);
864 static	void softdep_process_journal(struct mount *, int);
865 static	struct jremref *newjremref(struct dirrem *, struct inode *,
866 	    struct inode *ip, off_t, nlink_t);
867 static	struct jaddref *newjaddref(struct inode *, ino_t, off_t, int16_t,
868 	    uint16_t);
869 static inline void newinoref(struct inoref *, ino_t, ino_t, off_t, nlink_t,
870 	    uint16_t);
871 static inline struct jsegdep *inoref_jseg(struct inoref *);
872 static	struct jmvref *newjmvref(struct inode *, ino_t, off_t, off_t);
873 static	struct jfreeblk *newjfreeblk(struct freeblks *, ufs_lbn_t,
874 	    ufs2_daddr_t, int);
875 static	struct jfreefrag *newjfreefrag(struct freefrag *, struct inode *,
876 	    ufs2_daddr_t, long, ufs_lbn_t);
877 static	struct freework *newfreework(struct freeblks *, struct freework *,
878 	    ufs_lbn_t, ufs2_daddr_t, int, int);
879 static	void jwait(struct worklist *wk);
880 static	struct inodedep *inodedep_lookup_ip(struct inode *);
881 static	int bmsafemap_rollbacks(struct bmsafemap *);
882 static	struct freefile *handle_bufwait(struct inodedep *, struct workhead *);
883 static	void handle_jwork(struct workhead *);
884 static	struct mkdir *setup_newdir(struct diradd *, ino_t, ino_t, struct buf *,
885 	    struct mkdir **);
886 static	struct jblocks *jblocks_create(void);
887 static	ufs2_daddr_t jblocks_alloc(struct jblocks *, int, int *);
888 static	void jblocks_free(struct jblocks *, struct mount *, int);
889 static	void jblocks_destroy(struct jblocks *);
890 static	void jblocks_add(struct jblocks *, ufs2_daddr_t, int);
891 
892 /*
893  * Exported softdep operations.
894  */
895 static	void softdep_disk_io_initiation(struct buf *);
896 static	void softdep_disk_write_complete(struct buf *);
897 static	void softdep_deallocate_dependencies(struct buf *);
898 static	int softdep_count_dependencies(struct buf *bp, int);
899 
900 static struct mtx lk;
901 MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF);
902 
903 #define TRY_ACQUIRE_LOCK(lk)		mtx_trylock(lk)
904 #define ACQUIRE_LOCK(lk)		mtx_lock(lk)
905 #define FREE_LOCK(lk)			mtx_unlock(lk)
906 
907 #define	BUF_AREC(bp)			lockallowrecurse(&(bp)->b_lock)
908 #define	BUF_NOREC(bp)			lockdisablerecurse(&(bp)->b_lock)
909 
910 /*
911  * Worklist queue management.
912  * These routines require that the lock be held.
913  */
914 #ifndef /* NOT */ DEBUG
915 #define WORKLIST_INSERT(head, item) do {	\
916 	(item)->wk_state |= ONWORKLIST;		\
917 	LIST_INSERT_HEAD(head, item, wk_list);	\
918 } while (0)
919 #define WORKLIST_REMOVE(item) do {		\
920 	(item)->wk_state &= ~ONWORKLIST;	\
921 	LIST_REMOVE(item, wk_list);		\
922 } while (0)
923 #define WORKLIST_INSERT_UNLOCKED	WORKLIST_INSERT
924 #define WORKLIST_REMOVE_UNLOCKED	WORKLIST_REMOVE
925 
926 #else /* DEBUG */
927 static	void worklist_insert(struct workhead *, struct worklist *, int);
928 static	void worklist_remove(struct worklist *, int);
929 
930 #define WORKLIST_INSERT(head, item) worklist_insert(head, item, 1)
931 #define WORKLIST_INSERT_UNLOCKED(head, item) worklist_insert(head, item, 0)
932 #define WORKLIST_REMOVE(item) worklist_remove(item, 1)
933 #define WORKLIST_REMOVE_UNLOCKED(item) worklist_remove(item, 0)
934 
935 static void
936 worklist_insert(head, item, locked)
937 	struct workhead *head;
938 	struct worklist *item;
939 	int locked;
940 {
941 
942 	if (locked)
943 		mtx_assert(&lk, MA_OWNED);
944 	if (item->wk_state & ONWORKLIST)
945 		panic("worklist_insert: %p %s(0x%X) already on list",
946 		    item, TYPENAME(item->wk_type), item->wk_state);
947 	item->wk_state |= ONWORKLIST;
948 	LIST_INSERT_HEAD(head, item, wk_list);
949 }
950 
951 static void
952 worklist_remove(item, locked)
953 	struct worklist *item;
954 	int locked;
955 {
956 
957 	if (locked)
958 		mtx_assert(&lk, MA_OWNED);
959 	if ((item->wk_state & ONWORKLIST) == 0)
960 		panic("worklist_remove: %p %s(0x%X) not on list",
961 		    item, TYPENAME(item->wk_type), item->wk_state);
962 	item->wk_state &= ~ONWORKLIST;
963 	LIST_REMOVE(item, wk_list);
964 }
965 #endif /* DEBUG */
966 
967 /*
968  * Merge two jsegdeps keeping only the oldest one as newer references
969  * can't be discarded until after older references.
970  */
971 static inline struct jsegdep *
972 jsegdep_merge(struct jsegdep *one, struct jsegdep *two)
973 {
974 	struct jsegdep *swp;
975 
976 	if (two == NULL)
977 		return (one);
978 
979 	if (one->jd_seg->js_seq > two->jd_seg->js_seq) {
980 		swp = one;
981 		one = two;
982 		two = swp;
983 	}
984 	WORKLIST_REMOVE(&two->jd_list);
985 	free_jsegdep(two);
986 
987 	return (one);
988 }
989 
990 /*
991  * If two freedeps are compatible free one to reduce list size.
992  */
993 static inline struct freedep *
994 freedep_merge(struct freedep *one, struct freedep *two)
995 {
996 	if (two == NULL)
997 		return (one);
998 
999 	if (one->fd_freework == two->fd_freework) {
1000 		WORKLIST_REMOVE(&two->fd_list);
1001 		free_freedep(two);
1002 	}
1003 	return (one);
1004 }
1005 
1006 /*
1007  * Move journal work from one list to another.  Duplicate freedeps and
1008  * jsegdeps are coalesced to keep the lists as small as possible.
1009  */
1010 static void
1011 jwork_move(dst, src)
1012 	struct workhead *dst;
1013 	struct workhead *src;
1014 {
1015 	struct freedep *freedep;
1016 	struct jsegdep *jsegdep;
1017 	struct worklist *wkn;
1018 	struct worklist *wk;
1019 
1020 	KASSERT(dst != src,
1021 	    ("jwork_move: dst == src"));
1022 	freedep = NULL;
1023 	jsegdep = NULL;
1024 	LIST_FOREACH_SAFE(wk, dst, wk_list, wkn) {
1025 		if (wk->wk_type == D_JSEGDEP)
1026 			jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep);
1027 		if (wk->wk_type == D_FREEDEP)
1028 			freedep = freedep_merge(WK_FREEDEP(wk), freedep);
1029 	}
1030 
1031 	mtx_assert(&lk, MA_OWNED);
1032 	while ((wk = LIST_FIRST(src)) != NULL) {
1033 		WORKLIST_REMOVE(wk);
1034 		WORKLIST_INSERT(dst, wk);
1035 		if (wk->wk_type == D_JSEGDEP) {
1036 			jsegdep = jsegdep_merge(WK_JSEGDEP(wk), jsegdep);
1037 			continue;
1038 		}
1039 		if (wk->wk_type == D_FREEDEP)
1040 			freedep = freedep_merge(WK_FREEDEP(wk), freedep);
1041 	}
1042 }
1043 
1044 /*
1045  * Routines for tracking and managing workitems.
1046  */
1047 static	void workitem_free(struct worklist *, int);
1048 static	void workitem_alloc(struct worklist *, int, struct mount *);
1049 
1050 #define	WORKITEM_FREE(item, type) workitem_free((struct worklist *)(item), (type))
1051 
1052 static void
1053 workitem_free(item, type)
1054 	struct worklist *item;
1055 	int type;
1056 {
1057 	struct ufsmount *ump;
1058 	mtx_assert(&lk, MA_OWNED);
1059 
1060 #ifdef DEBUG
1061 	if (item->wk_state & ONWORKLIST)
1062 		panic("workitem_free: %s(0x%X) still on list",
1063 		    TYPENAME(item->wk_type), item->wk_state);
1064 	if (item->wk_type != type)
1065 		panic("workitem_free: type mismatch %s != %s",
1066 		    TYPENAME(item->wk_type), TYPENAME(type));
1067 #endif
1068 	ump = VFSTOUFS(item->wk_mp);
1069 	if (--ump->softdep_deps == 0 && ump->softdep_req)
1070 		wakeup(&ump->softdep_deps);
1071 	dep_current[type]--;
1072 	free(item, DtoM(type));
1073 }
1074 
1075 static void
1076 workitem_alloc(item, type, mp)
1077 	struct worklist *item;
1078 	int type;
1079 	struct mount *mp;
1080 {
1081 	item->wk_type = type;
1082 	item->wk_mp = mp;
1083 	item->wk_state = 0;
1084 	ACQUIRE_LOCK(&lk);
1085 	dep_current[type]++;
1086 	dep_total[type]++;
1087 	VFSTOUFS(mp)->softdep_deps++;
1088 	VFSTOUFS(mp)->softdep_accdeps++;
1089 	FREE_LOCK(&lk);
1090 }
1091 
1092 /*
1093  * Workitem queue management
1094  */
1095 static int max_softdeps;	/* maximum number of structs before slowdown */
1096 static int maxindirdeps = 50;	/* max number of indirdeps before slowdown */
1097 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
1098 static int proc_waiting;	/* tracks whether we have a timeout posted */
1099 static int *stat_countp;	/* statistic to count in proc_waiting timeout */
1100 static struct callout softdep_callout;
1101 static int req_pending;
1102 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
1103 #define FLUSH_INODES		1
1104 static int req_clear_remove;	/* syncer process flush some freeblks */
1105 #define FLUSH_REMOVE		2
1106 #define FLUSH_REMOVE_WAIT	3
1107 static long num_freeblkdep;	/* number of freeblks workitems allocated */
1108 
1109 /*
1110  * runtime statistics
1111  */
1112 static int stat_worklist_push;	/* number of worklist cleanups */
1113 static int stat_blk_limit_push;	/* number of times block limit neared */
1114 static int stat_ino_limit_push;	/* number of times inode limit neared */
1115 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
1116 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
1117 static int stat_sync_limit_hit;	/* number of synchronous slowdowns imposed */
1118 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
1119 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
1120 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
1121 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
1122 static int stat_jaddref;	/* bufs redirtied as ino bitmap can not write */
1123 static int stat_jnewblk;	/* bufs redirtied as blk bitmap can not write */
1124 static int stat_journal_min;	/* Times hit journal min threshold */
1125 static int stat_journal_low;	/* Times hit journal low threshold */
1126 static int stat_journal_wait;	/* Times blocked in jwait(). */
1127 static int stat_jwait_filepage;	/* Times blocked in jwait() for filepage. */
1128 static int stat_jwait_freeblks;	/* Times blocked in jwait() for freeblks. */
1129 static int stat_jwait_inode;	/* Times blocked in jwait() for inodes. */
1130 static int stat_jwait_newblk;	/* Times blocked in jwait() for newblks. */
1131 
1132 SYSCTL_INT(_debug_softdep, OID_AUTO, max_softdeps, CTLFLAG_RW,
1133     &max_softdeps, 0, "");
1134 SYSCTL_INT(_debug_softdep, OID_AUTO, tickdelay, CTLFLAG_RW,
1135     &tickdelay, 0, "");
1136 SYSCTL_INT(_debug_softdep, OID_AUTO, maxindirdeps, CTLFLAG_RW,
1137     &maxindirdeps, 0, "");
1138 SYSCTL_INT(_debug_softdep, OID_AUTO, worklist_push, CTLFLAG_RW,
1139     &stat_worklist_push, 0,"");
1140 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_push, CTLFLAG_RW,
1141     &stat_blk_limit_push, 0,"");
1142 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_push, CTLFLAG_RW,
1143     &stat_ino_limit_push, 0,"");
1144 SYSCTL_INT(_debug_softdep, OID_AUTO, blk_limit_hit, CTLFLAG_RW,
1145     &stat_blk_limit_hit, 0, "");
1146 SYSCTL_INT(_debug_softdep, OID_AUTO, ino_limit_hit, CTLFLAG_RW,
1147     &stat_ino_limit_hit, 0, "");
1148 SYSCTL_INT(_debug_softdep, OID_AUTO, sync_limit_hit, CTLFLAG_RW,
1149     &stat_sync_limit_hit, 0, "");
1150 SYSCTL_INT(_debug_softdep, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW,
1151     &stat_indir_blk_ptrs, 0, "");
1152 SYSCTL_INT(_debug_softdep, OID_AUTO, inode_bitmap, CTLFLAG_RW,
1153     &stat_inode_bitmap, 0, "");
1154 SYSCTL_INT(_debug_softdep, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW,
1155     &stat_direct_blk_ptrs, 0, "");
1156 SYSCTL_INT(_debug_softdep, OID_AUTO, dir_entry, CTLFLAG_RW,
1157     &stat_dir_entry, 0, "");
1158 SYSCTL_INT(_debug_softdep, OID_AUTO, jaddref_rollback, CTLFLAG_RW,
1159     &stat_jaddref, 0, "");
1160 SYSCTL_INT(_debug_softdep, OID_AUTO, jnewblk_rollback, CTLFLAG_RW,
1161     &stat_jnewblk, 0, "");
1162 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_low, CTLFLAG_RW,
1163     &stat_journal_low, 0, "");
1164 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_min, CTLFLAG_RW,
1165     &stat_journal_min, 0, "");
1166 SYSCTL_INT(_debug_softdep, OID_AUTO, journal_wait, CTLFLAG_RW,
1167     &stat_journal_wait, 0, "");
1168 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_filepage, CTLFLAG_RW,
1169     &stat_jwait_filepage, 0, "");
1170 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_freeblks, CTLFLAG_RW,
1171     &stat_jwait_freeblks, 0, "");
1172 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_inode, CTLFLAG_RW,
1173     &stat_jwait_inode, 0, "");
1174 SYSCTL_INT(_debug_softdep, OID_AUTO, jwait_newblk, CTLFLAG_RW,
1175     &stat_jwait_newblk, 0, "");
1176 
1177 SYSCTL_DECL(_vfs_ffs);
1178 
1179 LIST_HEAD(bmsafemap_hashhead, bmsafemap) *bmsafemap_hashtbl;
1180 static u_long	bmsafemap_hash;	/* size of hash table - 1 */
1181 
1182 static int compute_summary_at_mount = 0;	/* Whether to recompute the summary at mount time */
1183 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW,
1184 	   &compute_summary_at_mount, 0, "Recompute summary at mount");
1185 
1186 static struct proc *softdepproc;
1187 static struct kproc_desc softdep_kp = {
1188 	"softdepflush",
1189 	softdep_flush,
1190 	&softdepproc
1191 };
1192 SYSINIT(sdproc, SI_SUB_KTHREAD_UPDATE, SI_ORDER_ANY, kproc_start,
1193     &softdep_kp);
1194 
1195 static void
1196 softdep_flush(void)
1197 {
1198 	struct mount *nmp;
1199 	struct mount *mp;
1200 	struct ufsmount *ump;
1201 	struct thread *td;
1202 	int remaining;
1203 	int progress;
1204 	int vfslocked;
1205 
1206 	td = curthread;
1207 	td->td_pflags |= TDP_NORUNNINGBUF;
1208 
1209 	for (;;) {
1210 		kproc_suspend_check(softdepproc);
1211 		vfslocked = VFS_LOCK_GIANT((struct mount *)NULL);
1212 		ACQUIRE_LOCK(&lk);
1213 		/*
1214 		 * If requested, try removing inode or removal dependencies.
1215 		 */
1216 		if (req_clear_inodedeps) {
1217 			clear_inodedeps(td);
1218 			req_clear_inodedeps -= 1;
1219 			wakeup_one(&proc_waiting);
1220 		}
1221 		if (req_clear_remove) {
1222 			clear_remove(td);
1223 			req_clear_remove -= 1;
1224 			wakeup_one(&proc_waiting);
1225 		}
1226 		FREE_LOCK(&lk);
1227 		VFS_UNLOCK_GIANT(vfslocked);
1228 		remaining = progress = 0;
1229 		mtx_lock(&mountlist_mtx);
1230 		for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp)  {
1231 			nmp = TAILQ_NEXT(mp, mnt_list);
1232 			if ((mp->mnt_flag & MNT_SOFTDEP) == 0)
1233 				continue;
1234 			if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
1235 				continue;
1236 			vfslocked = VFS_LOCK_GIANT(mp);
1237 			progress += softdep_process_worklist(mp, 0);
1238 			ump = VFSTOUFS(mp);
1239 			remaining += ump->softdep_on_worklist -
1240 				ump->softdep_on_worklist_inprogress;
1241 			VFS_UNLOCK_GIANT(vfslocked);
1242 			mtx_lock(&mountlist_mtx);
1243 			nmp = TAILQ_NEXT(mp, mnt_list);
1244 			vfs_unbusy(mp);
1245 		}
1246 		mtx_unlock(&mountlist_mtx);
1247 		if (remaining && progress)
1248 			continue;
1249 		ACQUIRE_LOCK(&lk);
1250 		if (!req_pending)
1251 			msleep(&req_pending, &lk, PVM, "sdflush", hz);
1252 		req_pending = 0;
1253 		FREE_LOCK(&lk);
1254 	}
1255 }
1256 
1257 static void
1258 worklist_speedup(void)
1259 {
1260 	mtx_assert(&lk, MA_OWNED);
1261 	if (req_pending == 0) {
1262 		req_pending = 1;
1263 		wakeup(&req_pending);
1264 	}
1265 }
1266 
1267 static int
1268 softdep_speedup(void)
1269 {
1270 
1271 	worklist_speedup();
1272 	bd_speedup();
1273 	return speedup_syncer();
1274 }
1275 
1276 /*
1277  * Add an item to the end of the work queue.
1278  * This routine requires that the lock be held.
1279  * This is the only routine that adds items to the list.
1280  * The following routine is the only one that removes items
1281  * and does so in order from first to last.
1282  */
1283 static void
1284 add_to_worklist(wk, nodelay)
1285 	struct worklist *wk;
1286 	int nodelay;
1287 {
1288 	struct ufsmount *ump;
1289 
1290 	mtx_assert(&lk, MA_OWNED);
1291 	ump = VFSTOUFS(wk->wk_mp);
1292 	if (wk->wk_state & ONWORKLIST)
1293 		panic("add_to_worklist: %s(0x%X) already on list",
1294 		    TYPENAME(wk->wk_type), wk->wk_state);
1295 	wk->wk_state |= ONWORKLIST;
1296 	if (LIST_EMPTY(&ump->softdep_workitem_pending))
1297 		LIST_INSERT_HEAD(&ump->softdep_workitem_pending, wk, wk_list);
1298 	else
1299 		LIST_INSERT_AFTER(ump->softdep_worklist_tail, wk, wk_list);
1300 	ump->softdep_worklist_tail = wk;
1301 	ump->softdep_on_worklist += 1;
1302 	if (nodelay)
1303 		worklist_speedup();
1304 }
1305 
1306 /*
1307  * Remove the item to be processed. If we are removing the last
1308  * item on the list, we need to recalculate the tail pointer.
1309  */
1310 static void
1311 remove_from_worklist(wk)
1312 	struct worklist *wk;
1313 {
1314 	struct ufsmount *ump;
1315 	struct worklist *wkend;
1316 
1317 	ump = VFSTOUFS(wk->wk_mp);
1318 	WORKLIST_REMOVE(wk);
1319 	if (wk == ump->softdep_worklist_tail) {
1320 		LIST_FOREACH(wkend, &ump->softdep_workitem_pending, wk_list)
1321 			if (LIST_NEXT(wkend, wk_list) == NULL)
1322 				break;
1323 		ump->softdep_worklist_tail = wkend;
1324 	}
1325 	ump->softdep_on_worklist -= 1;
1326 }
1327 
1328 /*
1329  * Process that runs once per second to handle items in the background queue.
1330  *
1331  * Note that we ensure that everything is done in the order in which they
1332  * appear in the queue. The code below depends on this property to ensure
1333  * that blocks of a file are freed before the inode itself is freed. This
1334  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
1335  * until all the old ones have been purged from the dependency lists.
1336  */
1337 int
1338 softdep_process_worklist(mp, full)
1339 	struct mount *mp;
1340 	int full;
1341 {
1342 	struct thread *td = curthread;
1343 	int cnt, matchcnt, loopcount;
1344 	struct ufsmount *ump;
1345 	long starttime;
1346 
1347 	KASSERT(mp != NULL, ("softdep_process_worklist: NULL mp"));
1348 	/*
1349 	 * Record the process identifier of our caller so that we can give
1350 	 * this process preferential treatment in request_cleanup below.
1351 	 */
1352 	matchcnt = 0;
1353 	ump = VFSTOUFS(mp);
1354 	ACQUIRE_LOCK(&lk);
1355 	loopcount = 1;
1356 	starttime = time_second;
1357 	softdep_process_journal(mp, full?MNT_WAIT:0);
1358 	while (ump->softdep_on_worklist > 0) {
1359 		if ((cnt = process_worklist_item(mp, LK_NOWAIT)) == -1)
1360 			break;
1361 		else
1362 			matchcnt += cnt;
1363 		/*
1364 		 * If requested, try removing inode or removal dependencies.
1365 		 */
1366 		if (req_clear_inodedeps) {
1367 			clear_inodedeps(td);
1368 			req_clear_inodedeps -= 1;
1369 			wakeup_one(&proc_waiting);
1370 		}
1371 		if (req_clear_remove) {
1372 			clear_remove(td);
1373 			req_clear_remove -= 1;
1374 			wakeup_one(&proc_waiting);
1375 		}
1376 		/*
1377 		 * We do not generally want to stop for buffer space, but if
1378 		 * we are really being a buffer hog, we will stop and wait.
1379 		 */
1380 		if (loopcount++ % 128 == 0) {
1381 			FREE_LOCK(&lk);
1382 			uio_yield();
1383 			bwillwrite();
1384 			ACQUIRE_LOCK(&lk);
1385 		}
1386 		/*
1387 		 * Never allow processing to run for more than one
1388 		 * second. Otherwise the other mountpoints may get
1389 		 * excessively backlogged.
1390 		 */
1391 		if (!full && starttime != time_second)
1392 			break;
1393 	}
1394 	if (full == 0)
1395 		journal_unsuspend(ump);
1396 	FREE_LOCK(&lk);
1397 	return (matchcnt);
1398 }
1399 
1400 /*
1401  * Process all removes associated with a vnode if we are running out of
1402  * journal space.  Any other process which attempts to flush these will
1403  * be unable as we have the vnodes locked.
1404  */
1405 static void
1406 process_removes(vp)
1407 	struct vnode *vp;
1408 {
1409 	struct inodedep *inodedep;
1410 	struct dirrem *dirrem;
1411 	struct mount *mp;
1412 	ino_t inum;
1413 
1414 	mtx_assert(&lk, MA_OWNED);
1415 
1416 	mp = vp->v_mount;
1417 	inum = VTOI(vp)->i_number;
1418 	for (;;) {
1419 		if (inodedep_lookup(mp, inum, 0, &inodedep) == 0)
1420 			return;
1421 		LIST_FOREACH(dirrem, &inodedep->id_dirremhd, dm_inonext)
1422 			if ((dirrem->dm_state & (COMPLETE | ONWORKLIST)) ==
1423 			    (COMPLETE | ONWORKLIST))
1424 				break;
1425 		if (dirrem == NULL)
1426 			return;
1427 		/*
1428 		 * If another thread is trying to lock this vnode it will
1429 		 * fail but we must wait for it to do so before we can
1430 		 * proceed.
1431 		 */
1432 		if (dirrem->dm_state & INPROGRESS) {
1433 			dirrem->dm_state |= IOWAITING;
1434 			msleep(&dirrem->dm_list, &lk, PVM, "pwrwait", 0);
1435 			continue;
1436 		}
1437 		remove_from_worklist(&dirrem->dm_list);
1438 		FREE_LOCK(&lk);
1439 		if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
1440 			panic("process_removes: suspended filesystem");
1441 		handle_workitem_remove(dirrem, vp);
1442 		vn_finished_secondary_write(mp);
1443 		ACQUIRE_LOCK(&lk);
1444 	}
1445 }
1446 
1447 /*
1448  * Process one item on the worklist.
1449  */
1450 static int
1451 process_worklist_item(mp, flags)
1452 	struct mount *mp;
1453 	int flags;
1454 {
1455 	struct worklist *wk;
1456 	struct ufsmount *ump;
1457 	struct vnode *vp;
1458 	int matchcnt = 0;
1459 
1460 	mtx_assert(&lk, MA_OWNED);
1461 	KASSERT(mp != NULL, ("process_worklist_item: NULL mp"));
1462 	/*
1463 	 * If we are being called because of a process doing a
1464 	 * copy-on-write, then it is not safe to write as we may
1465 	 * recurse into the copy-on-write routine.
1466 	 */
1467 	if (curthread->td_pflags & TDP_COWINPROGRESS)
1468 		return (-1);
1469 	/*
1470 	 * Normally we just process each item on the worklist in order.
1471 	 * However, if we are in a situation where we cannot lock any
1472 	 * inodes, we have to skip over any dirrem requests whose
1473 	 * vnodes are resident and locked.
1474 	 */
1475 	vp = NULL;
1476 	ump = VFSTOUFS(mp);
1477 	LIST_FOREACH(wk, &ump->softdep_workitem_pending, wk_list) {
1478 		if (wk->wk_state & INPROGRESS)
1479 			continue;
1480 		if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
1481 			break;
1482 		wk->wk_state |= INPROGRESS;
1483 		ump->softdep_on_worklist_inprogress++;
1484 		FREE_LOCK(&lk);
1485 		ffs_vgetf(mp, WK_DIRREM(wk)->dm_oldinum,
1486 		    LK_NOWAIT | LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ);
1487 		ACQUIRE_LOCK(&lk);
1488 		if (wk->wk_state & IOWAITING) {
1489 			wk->wk_state &= ~IOWAITING;
1490 			wakeup(wk);
1491 		}
1492 		wk->wk_state &= ~INPROGRESS;
1493 		ump->softdep_on_worklist_inprogress--;
1494 		if (vp != NULL)
1495 			break;
1496 	}
1497 	if (wk == 0)
1498 		return (-1);
1499 	remove_from_worklist(wk);
1500 	FREE_LOCK(&lk);
1501 	if (vn_start_secondary_write(NULL, &mp, V_NOWAIT))
1502 		panic("process_worklist_item: suspended filesystem");
1503 	matchcnt++;
1504 	switch (wk->wk_type) {
1505 
1506 	case D_DIRREM:
1507 		/* removal of a directory entry */
1508 		handle_workitem_remove(WK_DIRREM(wk), vp);
1509 		if (vp)
1510 			vput(vp);
1511 		break;
1512 
1513 	case D_FREEBLKS:
1514 		/* releasing blocks and/or fragments from a file */
1515 		handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT);
1516 		break;
1517 
1518 	case D_FREEFRAG:
1519 		/* releasing a fragment when replaced as a file grows */
1520 		handle_workitem_freefrag(WK_FREEFRAG(wk));
1521 		break;
1522 
1523 	case D_FREEFILE:
1524 		/* releasing an inode when its link count drops to 0 */
1525 		handle_workitem_freefile(WK_FREEFILE(wk));
1526 		break;
1527 
1528 	case D_FREEWORK:
1529 		/* Final block in an indirect was freed. */
1530 		handle_workitem_indirblk(WK_FREEWORK(wk));
1531 		break;
1532 
1533 	default:
1534 		panic("%s_process_worklist: Unknown type %s",
1535 		    "softdep", TYPENAME(wk->wk_type));
1536 		/* NOTREACHED */
1537 	}
1538 	vn_finished_secondary_write(mp);
1539 	ACQUIRE_LOCK(&lk);
1540 	return (matchcnt);
1541 }
1542 
1543 /*
1544  * Move dependencies from one buffer to another.
1545  */
1546 int
1547 softdep_move_dependencies(oldbp, newbp)
1548 	struct buf *oldbp;
1549 	struct buf *newbp;
1550 {
1551 	struct worklist *wk, *wktail;
1552 	int dirty;
1553 
1554 	dirty = 0;
1555 	wktail = NULL;
1556 	ACQUIRE_LOCK(&lk);
1557 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
1558 		LIST_REMOVE(wk, wk_list);
1559 		if (wk->wk_type == D_BMSAFEMAP &&
1560 		    bmsafemap_rollbacks(WK_BMSAFEMAP(wk)))
1561 			dirty = 1;
1562 		if (wktail == 0)
1563 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
1564 		else
1565 			LIST_INSERT_AFTER(wktail, wk, wk_list);
1566 		wktail = wk;
1567 	}
1568 	FREE_LOCK(&lk);
1569 
1570 	return (dirty);
1571 }
1572 
1573 /*
1574  * Purge the work list of all items associated with a particular mount point.
1575  */
1576 int
1577 softdep_flushworklist(oldmnt, countp, td)
1578 	struct mount *oldmnt;
1579 	int *countp;
1580 	struct thread *td;
1581 {
1582 	struct vnode *devvp;
1583 	int count, error = 0;
1584 	struct ufsmount *ump;
1585 
1586 	/*
1587 	 * Alternately flush the block device associated with the mount
1588 	 * point and process any dependencies that the flushing
1589 	 * creates. We continue until no more worklist dependencies
1590 	 * are found.
1591 	 */
1592 	*countp = 0;
1593 	ump = VFSTOUFS(oldmnt);
1594 	devvp = ump->um_devvp;
1595 	while ((count = softdep_process_worklist(oldmnt, 1)) > 0) {
1596 		*countp += count;
1597 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1598 		error = VOP_FSYNC(devvp, MNT_WAIT, td);
1599 		VOP_UNLOCK(devvp, 0);
1600 		if (error)
1601 			break;
1602 	}
1603 	return (error);
1604 }
1605 
1606 int
1607 softdep_waitidle(struct mount *mp)
1608 {
1609 	struct ufsmount *ump;
1610 	int error;
1611 	int i;
1612 
1613 	ump = VFSTOUFS(mp);
1614 	ACQUIRE_LOCK(&lk);
1615 	for (i = 0; i < 10 && ump->softdep_deps; i++) {
1616 		ump->softdep_req = 1;
1617 		if (ump->softdep_on_worklist)
1618 			panic("softdep_waitidle: work added after flush.");
1619 		msleep(&ump->softdep_deps, &lk, PVM, "softdeps", 1);
1620 	}
1621 	ump->softdep_req = 0;
1622 	FREE_LOCK(&lk);
1623 	error = 0;
1624 	if (i == 10) {
1625 		error = EBUSY;
1626 		printf("softdep_waitidle: Failed to flush worklist for %p\n",
1627 		    mp);
1628 	}
1629 
1630 	return (error);
1631 }
1632 
1633 /*
1634  * Flush all vnodes and worklist items associated with a specified mount point.
1635  */
1636 int
1637 softdep_flushfiles(oldmnt, flags, td)
1638 	struct mount *oldmnt;
1639 	int flags;
1640 	struct thread *td;
1641 {
1642 	int error, depcount, loopcnt, retry_flush_count, retry;
1643 
1644 	loopcnt = 10;
1645 	retry_flush_count = 3;
1646 retry_flush:
1647 	error = 0;
1648 
1649 	/*
1650 	 * Alternately flush the vnodes associated with the mount
1651 	 * point and process any dependencies that the flushing
1652 	 * creates. In theory, this loop can happen at most twice,
1653 	 * but we give it a few extra just to be sure.
1654 	 */
1655 	for (; loopcnt > 0; loopcnt--) {
1656 		/*
1657 		 * Do another flush in case any vnodes were brought in
1658 		 * as part of the cleanup operations.
1659 		 */
1660 		if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
1661 			break;
1662 		if ((error = softdep_flushworklist(oldmnt, &depcount, td)) != 0 ||
1663 		    depcount == 0)
1664 			break;
1665 	}
1666 	/*
1667 	 * If we are unmounting then it is an error to fail. If we
1668 	 * are simply trying to downgrade to read-only, then filesystem
1669 	 * activity can keep us busy forever, so we just fail with EBUSY.
1670 	 */
1671 	if (loopcnt == 0) {
1672 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
1673 			panic("softdep_flushfiles: looping");
1674 		error = EBUSY;
1675 	}
1676 	if (!error)
1677 		error = softdep_waitidle(oldmnt);
1678 	if (!error) {
1679 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT) {
1680 			retry = 0;
1681 			MNT_ILOCK(oldmnt);
1682 			KASSERT((oldmnt->mnt_kern_flag & MNTK_NOINSMNTQ) != 0,
1683 			    ("softdep_flushfiles: !MNTK_NOINSMNTQ"));
1684 			if (oldmnt->mnt_nvnodelistsize > 0) {
1685 				if (--retry_flush_count > 0) {
1686 					retry = 1;
1687 					loopcnt = 3;
1688 				} else
1689 					error = EBUSY;
1690 			}
1691 			MNT_IUNLOCK(oldmnt);
1692 			if (retry)
1693 				goto retry_flush;
1694 		}
1695 	}
1696 	return (error);
1697 }
1698 
1699 /*
1700  * Structure hashing.
1701  *
1702  * There are three types of structures that can be looked up:
1703  *	1) pagedep structures identified by mount point, inode number,
1704  *	   and logical block.
1705  *	2) inodedep structures identified by mount point and inode number.
1706  *	3) newblk structures identified by mount point and
1707  *	   physical block number.
1708  *
1709  * The "pagedep" and "inodedep" dependency structures are hashed
1710  * separately from the file blocks and inodes to which they correspond.
1711  * This separation helps when the in-memory copy of an inode or
1712  * file block must be replaced. It also obviates the need to access
1713  * an inode or file page when simply updating (or de-allocating)
1714  * dependency structures. Lookup of newblk structures is needed to
1715  * find newly allocated blocks when trying to associate them with
1716  * their allocdirect or allocindir structure.
1717  *
1718  * The lookup routines optionally create and hash a new instance when
1719  * an existing entry is not found.
1720  */
1721 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
1722 #define NODELAY		0x0002	/* cannot do background work */
1723 
1724 /*
1725  * Structures and routines associated with pagedep caching.
1726  */
1727 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
1728 u_long	pagedep_hash;		/* size of hash table - 1 */
1729 #define	PAGEDEP_HASH(mp, inum, lbn) \
1730 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
1731 	    pagedep_hash])
1732 
1733 static int
1734 pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp)
1735 	struct pagedep_hashhead *pagedephd;
1736 	ino_t ino;
1737 	ufs_lbn_t lbn;
1738 	struct mount *mp;
1739 	int flags;
1740 	struct pagedep **pagedeppp;
1741 {
1742 	struct pagedep *pagedep;
1743 
1744 	LIST_FOREACH(pagedep, pagedephd, pd_hash)
1745 		if (ino == pagedep->pd_ino &&
1746 		    lbn == pagedep->pd_lbn &&
1747 		    mp == pagedep->pd_list.wk_mp)
1748 			break;
1749 	if (pagedep) {
1750 		*pagedeppp = pagedep;
1751 		if ((flags & DEPALLOC) != 0 &&
1752 		    (pagedep->pd_state & ONWORKLIST) == 0)
1753 			return (0);
1754 		return (1);
1755 	}
1756 	*pagedeppp = NULL;
1757 	return (0);
1758 }
1759 /*
1760  * Look up a pagedep. Return 1 if found, 0 if not found or found
1761  * when asked to allocate but not associated with any buffer.
1762  * If not found, allocate if DEPALLOC flag is passed.
1763  * Found or allocated entry is returned in pagedeppp.
1764  * This routine must be called with splbio interrupts blocked.
1765  */
1766 static int
1767 pagedep_lookup(mp, ino, lbn, flags, pagedeppp)
1768 	struct mount *mp;
1769 	ino_t ino;
1770 	ufs_lbn_t lbn;
1771 	int flags;
1772 	struct pagedep **pagedeppp;
1773 {
1774 	struct pagedep *pagedep;
1775 	struct pagedep_hashhead *pagedephd;
1776 	int ret;
1777 	int i;
1778 
1779 	mtx_assert(&lk, MA_OWNED);
1780 	pagedephd = PAGEDEP_HASH(mp, ino, lbn);
1781 
1782 	ret = pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp);
1783 	if (*pagedeppp || (flags & DEPALLOC) == 0)
1784 		return (ret);
1785 	FREE_LOCK(&lk);
1786 	pagedep = malloc(sizeof(struct pagedep),
1787 	    M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO);
1788 	workitem_alloc(&pagedep->pd_list, D_PAGEDEP, mp);
1789 	ACQUIRE_LOCK(&lk);
1790 	ret = pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp);
1791 	if (*pagedeppp) {
1792 		WORKITEM_FREE(pagedep, D_PAGEDEP);
1793 		return (ret);
1794 	}
1795 	pagedep->pd_ino = ino;
1796 	pagedep->pd_lbn = lbn;
1797 	LIST_INIT(&pagedep->pd_dirremhd);
1798 	LIST_INIT(&pagedep->pd_pendinghd);
1799 	for (i = 0; i < DAHASHSZ; i++)
1800 		LIST_INIT(&pagedep->pd_diraddhd[i]);
1801 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
1802 	*pagedeppp = pagedep;
1803 	return (0);
1804 }
1805 
1806 /*
1807  * Structures and routines associated with inodedep caching.
1808  */
1809 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
1810 static u_long	inodedep_hash;	/* size of hash table - 1 */
1811 static long	num_inodedep;	/* number of inodedep allocated */
1812 #define	INODEDEP_HASH(fs, inum) \
1813       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
1814 
1815 static int
1816 inodedep_find(inodedephd, fs, inum, inodedeppp)
1817 	struct inodedep_hashhead *inodedephd;
1818 	struct fs *fs;
1819 	ino_t inum;
1820 	struct inodedep **inodedeppp;
1821 {
1822 	struct inodedep *inodedep;
1823 
1824 	LIST_FOREACH(inodedep, inodedephd, id_hash)
1825 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
1826 			break;
1827 	if (inodedep) {
1828 		*inodedeppp = inodedep;
1829 		return (1);
1830 	}
1831 	*inodedeppp = NULL;
1832 
1833 	return (0);
1834 }
1835 /*
1836  * Look up an inodedep. Return 1 if found, 0 if not found.
1837  * If not found, allocate if DEPALLOC flag is passed.
1838  * Found or allocated entry is returned in inodedeppp.
1839  * This routine must be called with splbio interrupts blocked.
1840  */
1841 static int
1842 inodedep_lookup(mp, inum, flags, inodedeppp)
1843 	struct mount *mp;
1844 	ino_t inum;
1845 	int flags;
1846 	struct inodedep **inodedeppp;
1847 {
1848 	struct inodedep *inodedep;
1849 	struct inodedep_hashhead *inodedephd;
1850 	struct fs *fs;
1851 
1852 	mtx_assert(&lk, MA_OWNED);
1853 	fs = VFSTOUFS(mp)->um_fs;
1854 	inodedephd = INODEDEP_HASH(fs, inum);
1855 
1856 	if (inodedep_find(inodedephd, fs, inum, inodedeppp))
1857 		return (1);
1858 	if ((flags & DEPALLOC) == 0)
1859 		return (0);
1860 	/*
1861 	 * If we are over our limit, try to improve the situation.
1862 	 */
1863 	if (num_inodedep > max_softdeps && (flags & NODELAY) == 0)
1864 		request_cleanup(mp, FLUSH_INODES);
1865 	FREE_LOCK(&lk);
1866 	inodedep = malloc(sizeof(struct inodedep),
1867 		M_INODEDEP, M_SOFTDEP_FLAGS);
1868 	workitem_alloc(&inodedep->id_list, D_INODEDEP, mp);
1869 	ACQUIRE_LOCK(&lk);
1870 	if (inodedep_find(inodedephd, fs, inum, inodedeppp)) {
1871 		WORKITEM_FREE(inodedep, D_INODEDEP);
1872 		return (1);
1873 	}
1874 	num_inodedep += 1;
1875 	inodedep->id_fs = fs;
1876 	inodedep->id_ino = inum;
1877 	inodedep->id_state = ALLCOMPLETE;
1878 	inodedep->id_nlinkdelta = 0;
1879 	inodedep->id_savedino1 = NULL;
1880 	inodedep->id_savedsize = -1;
1881 	inodedep->id_savedextsize = -1;
1882 	inodedep->id_savednlink = -1;
1883 	inodedep->id_bmsafemap = NULL;
1884 	inodedep->id_mkdiradd = NULL;
1885 	LIST_INIT(&inodedep->id_dirremhd);
1886 	LIST_INIT(&inodedep->id_pendinghd);
1887 	LIST_INIT(&inodedep->id_inowait);
1888 	LIST_INIT(&inodedep->id_bufwait);
1889 	TAILQ_INIT(&inodedep->id_inoreflst);
1890 	TAILQ_INIT(&inodedep->id_inoupdt);
1891 	TAILQ_INIT(&inodedep->id_newinoupdt);
1892 	TAILQ_INIT(&inodedep->id_extupdt);
1893 	TAILQ_INIT(&inodedep->id_newextupdt);
1894 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1895 	*inodedeppp = inodedep;
1896 	return (0);
1897 }
1898 
1899 /*
1900  * Structures and routines associated with newblk caching.
1901  */
1902 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1903 u_long	newblk_hash;		/* size of hash table - 1 */
1904 #define	NEWBLK_HASH(fs, inum) \
1905 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1906 
1907 static int
1908 newblk_find(newblkhd, mp, newblkno, flags, newblkpp)
1909 	struct newblk_hashhead *newblkhd;
1910 	struct mount *mp;
1911 	ufs2_daddr_t newblkno;
1912 	int flags;
1913 	struct newblk **newblkpp;
1914 {
1915 	struct newblk *newblk;
1916 
1917 	LIST_FOREACH(newblk, newblkhd, nb_hash) {
1918 		if (newblkno != newblk->nb_newblkno)
1919 			continue;
1920 		if (mp != newblk->nb_list.wk_mp)
1921 			continue;
1922 		/*
1923 		 * If we're creating a new dependency don't match those that
1924 		 * have already been converted to allocdirects.  This is for
1925 		 * a frag extend.
1926 		 */
1927 		if ((flags & DEPALLOC) && newblk->nb_list.wk_type != D_NEWBLK)
1928 			continue;
1929 		break;
1930 	}
1931 	if (newblk) {
1932 		*newblkpp = newblk;
1933 		return (1);
1934 	}
1935 	*newblkpp = NULL;
1936 	return (0);
1937 }
1938 
1939 /*
1940  * Look up a newblk. Return 1 if found, 0 if not found.
1941  * If not found, allocate if DEPALLOC flag is passed.
1942  * Found or allocated entry is returned in newblkpp.
1943  */
1944 static int
1945 newblk_lookup(mp, newblkno, flags, newblkpp)
1946 	struct mount *mp;
1947 	ufs2_daddr_t newblkno;
1948 	int flags;
1949 	struct newblk **newblkpp;
1950 {
1951 	struct newblk *newblk;
1952 	struct newblk_hashhead *newblkhd;
1953 
1954 	newblkhd = NEWBLK_HASH(VFSTOUFS(mp)->um_fs, newblkno);
1955 	if (newblk_find(newblkhd, mp, newblkno, flags, newblkpp))
1956 		return (1);
1957 	if ((flags & DEPALLOC) == 0)
1958 		return (0);
1959 	FREE_LOCK(&lk);
1960 	newblk = malloc(sizeof(union allblk), M_NEWBLK,
1961 	    M_SOFTDEP_FLAGS | M_ZERO);
1962 	workitem_alloc(&newblk->nb_list, D_NEWBLK, mp);
1963 	ACQUIRE_LOCK(&lk);
1964 	if (newblk_find(newblkhd, mp, newblkno, flags, newblkpp)) {
1965 		WORKITEM_FREE(newblk, D_NEWBLK);
1966 		return (1);
1967 	}
1968 	newblk->nb_freefrag = NULL;
1969 	LIST_INIT(&newblk->nb_indirdeps);
1970 	LIST_INIT(&newblk->nb_newdirblk);
1971 	LIST_INIT(&newblk->nb_jwork);
1972 	newblk->nb_state = ATTACHED;
1973 	newblk->nb_newblkno = newblkno;
1974 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1975 	*newblkpp = newblk;
1976 	return (0);
1977 }
1978 
1979 /*
1980  * Executed during filesystem system initialization before
1981  * mounting any filesystems.
1982  */
1983 void
1984 softdep_initialize()
1985 {
1986 
1987 	LIST_INIT(&mkdirlisthd);
1988 	max_softdeps = desiredvnodes * 4;
1989 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP, &pagedep_hash);
1990 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1991 	newblk_hashtbl = hashinit(desiredvnodes / 5,  M_NEWBLK, &newblk_hash);
1992 	bmsafemap_hashtbl = hashinit(1024, M_BMSAFEMAP, &bmsafemap_hash);
1993 
1994 	/* initialise bioops hack */
1995 	bioops.io_start = softdep_disk_io_initiation;
1996 	bioops.io_complete = softdep_disk_write_complete;
1997 	bioops.io_deallocate = softdep_deallocate_dependencies;
1998 	bioops.io_countdeps = softdep_count_dependencies;
1999 
2000 	/* Initialize the callout with an mtx. */
2001 	callout_init_mtx(&softdep_callout, &lk, 0);
2002 }
2003 
2004 /*
2005  * Executed after all filesystems have been unmounted during
2006  * filesystem module unload.
2007  */
2008 void
2009 softdep_uninitialize()
2010 {
2011 
2012 	callout_drain(&softdep_callout);
2013 	hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash);
2014 	hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash);
2015 	hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash);
2016 	hashdestroy(bmsafemap_hashtbl, M_BMSAFEMAP, bmsafemap_hash);
2017 }
2018 
2019 /*
2020  * Called at mount time to notify the dependency code that a
2021  * filesystem wishes to use it.
2022  */
2023 int
2024 softdep_mount(devvp, mp, fs, cred)
2025 	struct vnode *devvp;
2026 	struct mount *mp;
2027 	struct fs *fs;
2028 	struct ucred *cred;
2029 {
2030 	struct csum_total cstotal;
2031 	struct ufsmount *ump;
2032 	struct cg *cgp;
2033 	struct buf *bp;
2034 	int error, cyl;
2035 
2036 	MNT_ILOCK(mp);
2037 	mp->mnt_flag = (mp->mnt_flag & ~MNT_ASYNC) | MNT_SOFTDEP;
2038 	if ((mp->mnt_kern_flag & MNTK_SOFTDEP) == 0) {
2039 		mp->mnt_kern_flag = (mp->mnt_kern_flag & ~MNTK_ASYNC) |
2040 			MNTK_SOFTDEP;
2041 		mp->mnt_noasync++;
2042 	}
2043 	MNT_IUNLOCK(mp);
2044 	ump = VFSTOUFS(mp);
2045 	LIST_INIT(&ump->softdep_workitem_pending);
2046 	LIST_INIT(&ump->softdep_journal_pending);
2047 	TAILQ_INIT(&ump->softdep_unlinked);
2048 	ump->softdep_worklist_tail = NULL;
2049 	ump->softdep_on_worklist = 0;
2050 	ump->softdep_deps = 0;
2051 	if ((fs->fs_flags & FS_SUJ) &&
2052 	    (error = journal_mount(mp, fs, cred)) != 0) {
2053 		printf("Failed to start journal: %d\n", error);
2054 		return (error);
2055 	}
2056 	/*
2057 	 * When doing soft updates, the counters in the
2058 	 * superblock may have gotten out of sync. Recomputation
2059 	 * can take a long time and can be deferred for background
2060 	 * fsck.  However, the old behavior of scanning the cylinder
2061 	 * groups and recalculating them at mount time is available
2062 	 * by setting vfs.ffs.compute_summary_at_mount to one.
2063 	 */
2064 	if (compute_summary_at_mount == 0 || fs->fs_clean != 0)
2065 		return (0);
2066 	bzero(&cstotal, sizeof cstotal);
2067 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
2068 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
2069 		    fs->fs_cgsize, cred, &bp)) != 0) {
2070 			brelse(bp);
2071 			return (error);
2072 		}
2073 		cgp = (struct cg *)bp->b_data;
2074 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
2075 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
2076 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
2077 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
2078 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
2079 		brelse(bp);
2080 	}
2081 #ifdef DEBUG
2082 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
2083 		printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
2084 #endif
2085 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
2086 	return (0);
2087 }
2088 
2089 void
2090 softdep_unmount(mp)
2091 	struct mount *mp;
2092 {
2093 
2094 	if (mp->mnt_kern_flag & MNTK_SUJ)
2095 		journal_unmount(mp);
2096 }
2097 
2098 struct jblocks {
2099 	struct jseglst	jb_segs;	/* TAILQ of current segments. */
2100 	struct jseg	*jb_writeseg;	/* Next write to complete. */
2101 	struct jextent	*jb_extent;	/* Extent array. */
2102 	uint64_t	jb_nextseq;	/* Next sequence number. */
2103 	uint64_t	jb_oldestseq;	/* Oldest active sequence number. */
2104 	int		jb_avail;	/* Available extents. */
2105 	int		jb_used;	/* Last used extent. */
2106 	int		jb_head;	/* Allocator head. */
2107 	int		jb_off;		/* Allocator extent offset. */
2108 	int		jb_blocks;	/* Total disk blocks covered. */
2109 	int		jb_free;	/* Total disk blocks free. */
2110 	int		jb_min;		/* Minimum free space. */
2111 	int		jb_low;		/* Low on space. */
2112 	int		jb_age;		/* Insertion time of oldest rec. */
2113 	int		jb_suspended;	/* Did journal suspend writes? */
2114 };
2115 
2116 struct jextent {
2117 	ufs2_daddr_t	je_daddr;	/* Disk block address. */
2118 	int		je_blocks;	/* Disk block count. */
2119 };
2120 
2121 static struct jblocks *
2122 jblocks_create(void)
2123 {
2124 	struct jblocks *jblocks;
2125 
2126 	jblocks = malloc(sizeof(*jblocks), M_JBLOCKS, M_WAITOK | M_ZERO);
2127 	TAILQ_INIT(&jblocks->jb_segs);
2128 	jblocks->jb_avail = 10;
2129 	jblocks->jb_extent = malloc(sizeof(struct jextent) * jblocks->jb_avail,
2130 	    M_JBLOCKS, M_WAITOK | M_ZERO);
2131 
2132 	return (jblocks);
2133 }
2134 
2135 static ufs2_daddr_t
2136 jblocks_alloc(jblocks, bytes, actual)
2137 	struct jblocks *jblocks;
2138 	int bytes;
2139 	int *actual;
2140 {
2141 	ufs2_daddr_t daddr;
2142 	struct jextent *jext;
2143 	int freecnt;
2144 	int blocks;
2145 
2146 	blocks = bytes / DEV_BSIZE;
2147 	jext = &jblocks->jb_extent[jblocks->jb_head];
2148 	freecnt = jext->je_blocks - jblocks->jb_off;
2149 	if (freecnt == 0) {
2150 		jblocks->jb_off = 0;
2151 		if (++jblocks->jb_head > jblocks->jb_used)
2152 			jblocks->jb_head = 0;
2153 		jext = &jblocks->jb_extent[jblocks->jb_head];
2154 		freecnt = jext->je_blocks;
2155 	}
2156 	if (freecnt > blocks)
2157 		freecnt = blocks;
2158 	*actual = freecnt * DEV_BSIZE;
2159 	daddr = jext->je_daddr + jblocks->jb_off;
2160 	jblocks->jb_off += freecnt;
2161 	jblocks->jb_free -= freecnt;
2162 
2163 	return (daddr);
2164 }
2165 
2166 static void
2167 jblocks_free(jblocks, mp, bytes)
2168 	struct jblocks *jblocks;
2169 	struct mount *mp;
2170 	int bytes;
2171 {
2172 
2173 	jblocks->jb_free += bytes / DEV_BSIZE;
2174 	if (jblocks->jb_suspended)
2175 		worklist_speedup();
2176 	wakeup(jblocks);
2177 }
2178 
2179 static void
2180 jblocks_destroy(jblocks)
2181 	struct jblocks *jblocks;
2182 {
2183 
2184 	if (jblocks->jb_extent)
2185 		free(jblocks->jb_extent, M_JBLOCKS);
2186 	free(jblocks, M_JBLOCKS);
2187 }
2188 
2189 static void
2190 jblocks_add(jblocks, daddr, blocks)
2191 	struct jblocks *jblocks;
2192 	ufs2_daddr_t daddr;
2193 	int blocks;
2194 {
2195 	struct jextent *jext;
2196 
2197 	jblocks->jb_blocks += blocks;
2198 	jblocks->jb_free += blocks;
2199 	jext = &jblocks->jb_extent[jblocks->jb_used];
2200 	/* Adding the first block. */
2201 	if (jext->je_daddr == 0) {
2202 		jext->je_daddr = daddr;
2203 		jext->je_blocks = blocks;
2204 		return;
2205 	}
2206 	/* Extending the last extent. */
2207 	if (jext->je_daddr + jext->je_blocks == daddr) {
2208 		jext->je_blocks += blocks;
2209 		return;
2210 	}
2211 	/* Adding a new extent. */
2212 	if (++jblocks->jb_used == jblocks->jb_avail) {
2213 		jblocks->jb_avail *= 2;
2214 		jext = malloc(sizeof(struct jextent) * jblocks->jb_avail,
2215 		    M_JBLOCKS, M_WAITOK | M_ZERO);
2216 		memcpy(jext, jblocks->jb_extent,
2217 		    sizeof(struct jextent) * jblocks->jb_used);
2218 		free(jblocks->jb_extent, M_JBLOCKS);
2219 		jblocks->jb_extent = jext;
2220 	}
2221 	jext = &jblocks->jb_extent[jblocks->jb_used];
2222 	jext->je_daddr = daddr;
2223 	jext->je_blocks = blocks;
2224 	return;
2225 }
2226 
2227 int
2228 softdep_journal_lookup(mp, vpp)
2229 	struct mount *mp;
2230 	struct vnode **vpp;
2231 {
2232 	struct componentname cnp;
2233 	struct vnode *dvp;
2234 	ino_t sujournal;
2235 	int error;
2236 
2237 	error = VFS_VGET(mp, ROOTINO, LK_EXCLUSIVE, &dvp);
2238 	if (error)
2239 		return (error);
2240 	bzero(&cnp, sizeof(cnp));
2241 	cnp.cn_nameiop = LOOKUP;
2242 	cnp.cn_flags = ISLASTCN;
2243 	cnp.cn_thread = curthread;
2244 	cnp.cn_cred = curthread->td_ucred;
2245 	cnp.cn_pnbuf = SUJ_FILE;
2246 	cnp.cn_nameptr = SUJ_FILE;
2247 	cnp.cn_namelen = strlen(SUJ_FILE);
2248 	error = ufs_lookup_ino(dvp, NULL, &cnp, &sujournal);
2249 	vput(dvp);
2250 	if (error != 0)
2251 		return (error);
2252 	error = VFS_VGET(mp, sujournal, LK_EXCLUSIVE, vpp);
2253 	return (error);
2254 }
2255 
2256 /*
2257  * Open and verify the journal file.
2258  */
2259 static int
2260 journal_mount(mp, fs, cred)
2261 	struct mount *mp;
2262 	struct fs *fs;
2263 	struct ucred *cred;
2264 {
2265 	struct jblocks *jblocks;
2266 	struct vnode *vp;
2267 	struct inode *ip;
2268 	ufs2_daddr_t blkno;
2269 	int bcount;
2270 	int error;
2271 	int i;
2272 
2273 	error = softdep_journal_lookup(mp, &vp);
2274 	if (error != 0) {
2275 		printf("Failed to find journal.  Use tunefs to create one\n");
2276 		return (error);
2277 	}
2278 	ip = VTOI(vp);
2279 	if (ip->i_size < SUJ_MIN) {
2280 		error = ENOSPC;
2281 		goto out;
2282 	}
2283 	bcount = lblkno(fs, ip->i_size);	/* Only use whole blocks. */
2284 	jblocks = jblocks_create();
2285 	for (i = 0; i < bcount; i++) {
2286 		error = ufs_bmaparray(vp, i, &blkno, NULL, NULL, NULL);
2287 		if (error)
2288 			break;
2289 		jblocks_add(jblocks, blkno, fsbtodb(fs, fs->fs_frag));
2290 	}
2291 	if (error) {
2292 		jblocks_destroy(jblocks);
2293 		goto out;
2294 	}
2295 	jblocks->jb_low = jblocks->jb_free / 3;	/* Reserve 33%. */
2296 	jblocks->jb_min = jblocks->jb_free / 10; /* Suspend at 10%. */
2297 	VFSTOUFS(mp)->softdep_jblocks = jblocks;
2298 out:
2299 	if (error == 0) {
2300 		MNT_ILOCK(mp);
2301 		mp->mnt_kern_flag |= MNTK_SUJ;
2302 		MNT_IUNLOCK(mp);
2303 		/*
2304 		 * Only validate the journal contents if the
2305 		 * filesystem is clean, otherwise we write the logs
2306 		 * but they'll never be used.  If the filesystem was
2307 		 * still dirty when we mounted it the journal is
2308 		 * invalid and a new journal can only be valid if it
2309 		 * starts from a clean mount.
2310 		 */
2311 		if (fs->fs_clean) {
2312 			DIP_SET(ip, i_modrev, fs->fs_mtime);
2313 			ip->i_flags |= IN_MODIFIED;
2314 			ffs_update(vp, 1);
2315 		}
2316 	}
2317 	vput(vp);
2318 	return (error);
2319 }
2320 
2321 static void
2322 journal_unmount(mp)
2323 	struct mount *mp;
2324 {
2325 	struct ufsmount *ump;
2326 
2327 	ump = VFSTOUFS(mp);
2328 	if (ump->softdep_jblocks)
2329 		jblocks_destroy(ump->softdep_jblocks);
2330 	ump->softdep_jblocks = NULL;
2331 }
2332 
2333 /*
2334  * Called when a journal record is ready to be written.  Space is allocated
2335  * and the journal entry is created when the journal is flushed to stable
2336  * store.
2337  */
2338 static void
2339 add_to_journal(wk)
2340 	struct worklist *wk;
2341 {
2342 	struct ufsmount *ump;
2343 
2344 	mtx_assert(&lk, MA_OWNED);
2345 	ump = VFSTOUFS(wk->wk_mp);
2346 	if (wk->wk_state & ONWORKLIST)
2347 		panic("add_to_journal: %s(0x%X) already on list",
2348 		    TYPENAME(wk->wk_type), wk->wk_state);
2349 	wk->wk_state |= ONWORKLIST | DEPCOMPLETE;
2350 	if (LIST_EMPTY(&ump->softdep_journal_pending)) {
2351 		ump->softdep_jblocks->jb_age = ticks;
2352 		LIST_INSERT_HEAD(&ump->softdep_journal_pending, wk, wk_list);
2353 	} else
2354 		LIST_INSERT_AFTER(ump->softdep_journal_tail, wk, wk_list);
2355 	ump->softdep_journal_tail = wk;
2356 	ump->softdep_on_journal += 1;
2357 }
2358 
2359 /*
2360  * Remove an arbitrary item for the journal worklist maintain the tail
2361  * pointer.  This happens when a new operation obviates the need to
2362  * journal an old operation.
2363  */
2364 static void
2365 remove_from_journal(wk)
2366 	struct worklist *wk;
2367 {
2368 	struct ufsmount *ump;
2369 
2370 	mtx_assert(&lk, MA_OWNED);
2371 	ump = VFSTOUFS(wk->wk_mp);
2372 #ifdef SUJ_DEBUG
2373 	{
2374 		struct worklist *wkn;
2375 
2376 		LIST_FOREACH(wkn, &ump->softdep_journal_pending, wk_list)
2377 			if (wkn == wk)
2378 				break;
2379 		if (wkn == NULL)
2380 			panic("remove_from_journal: %p is not in journal", wk);
2381 	}
2382 #endif
2383 	/*
2384 	 * We emulate a TAILQ to save space in most structures which do not
2385 	 * require TAILQ semantics.  Here we must update the tail position
2386 	 * when removing the tail which is not the final entry. This works
2387 	 * only if the worklist linkage are at the beginning of the structure.
2388 	 */
2389 	if (ump->softdep_journal_tail == wk)
2390 		ump->softdep_journal_tail =
2391 		    (struct worklist *)wk->wk_list.le_prev;
2392 
2393 	WORKLIST_REMOVE(wk);
2394 	ump->softdep_on_journal -= 1;
2395 }
2396 
2397 /*
2398  * Check for journal space as well as dependency limits so the prelink
2399  * code can throttle both journaled and non-journaled filesystems.
2400  * Threshold is 0 for low and 1 for min.
2401  */
2402 static int
2403 journal_space(ump, thresh)
2404 	struct ufsmount *ump;
2405 	int thresh;
2406 {
2407 	struct jblocks *jblocks;
2408 	int avail;
2409 
2410 	jblocks = ump->softdep_jblocks;
2411 	if (jblocks == NULL)
2412 		return (1);
2413 	/*
2414 	 * We use a tighter restriction here to prevent request_cleanup()
2415 	 * running in threads from running into locks we currently hold.
2416 	 */
2417 	if (num_inodedep > (max_softdeps / 10) * 9)
2418 		return (0);
2419 	if (thresh)
2420 		thresh = jblocks->jb_min;
2421 	else
2422 		thresh = jblocks->jb_low;
2423 	avail = (ump->softdep_on_journal * JREC_SIZE) / DEV_BSIZE;
2424 	avail = jblocks->jb_free - avail;
2425 
2426 	return (avail > thresh);
2427 }
2428 
2429 static void
2430 journal_suspend(ump)
2431 	struct ufsmount *ump;
2432 {
2433 	struct jblocks *jblocks;
2434 	struct mount *mp;
2435 
2436 	mp = UFSTOVFS(ump);
2437 	jblocks = ump->softdep_jblocks;
2438 	MNT_ILOCK(mp);
2439 	if ((mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
2440 		stat_journal_min++;
2441 		mp->mnt_kern_flag |= MNTK_SUSPEND;
2442 		mp->mnt_susp_owner = FIRST_THREAD_IN_PROC(softdepproc);
2443 	}
2444 	jblocks->jb_suspended = 1;
2445 	MNT_IUNLOCK(mp);
2446 }
2447 
2448 static int
2449 journal_unsuspend(struct ufsmount *ump)
2450 {
2451 	struct jblocks *jblocks;
2452 	struct mount *mp;
2453 
2454 	mp = UFSTOVFS(ump);
2455 	jblocks = ump->softdep_jblocks;
2456 
2457 	if (jblocks != NULL && jblocks->jb_suspended &&
2458 	    journal_space(ump, jblocks->jb_min)) {
2459 		jblocks->jb_suspended = 0;
2460 		FREE_LOCK(&lk);
2461 		mp->mnt_susp_owner = curthread;
2462 		vfs_write_resume(mp);
2463 		ACQUIRE_LOCK(&lk);
2464 		return (1);
2465 	}
2466 	return (0);
2467 }
2468 
2469 /*
2470  * Called before any allocation function to be certain that there is
2471  * sufficient space in the journal prior to creating any new records.
2472  * Since in the case of block allocation we may have multiple locked
2473  * buffers at the time of the actual allocation we can not block
2474  * when the journal records are created.  Doing so would create a deadlock
2475  * if any of these buffers needed to be flushed to reclaim space.  Instead
2476  * we require a sufficiently large amount of available space such that
2477  * each thread in the system could have passed this allocation check and
2478  * still have sufficient free space.  With 20% of a minimum journal size
2479  * of 1MB we have 6553 records available.
2480  */
2481 int
2482 softdep_prealloc(vp, waitok)
2483 	struct vnode *vp;
2484 	int waitok;
2485 {
2486 	struct ufsmount *ump;
2487 
2488 	if (DOINGSUJ(vp) == 0)
2489 		return (0);
2490 	ump = VFSTOUFS(vp->v_mount);
2491 	ACQUIRE_LOCK(&lk);
2492 	if (journal_space(ump, 0)) {
2493 		FREE_LOCK(&lk);
2494 		return (0);
2495 	}
2496 	stat_journal_low++;
2497 	FREE_LOCK(&lk);
2498 	if (waitok == MNT_NOWAIT)
2499 		return (ENOSPC);
2500 	/*
2501 	 * Attempt to sync this vnode once to flush any journal
2502 	 * work attached to it.
2503 	 */
2504 	if ((curthread->td_pflags & TDP_COWINPROGRESS) == 0)
2505 		ffs_syncvnode(vp, waitok);
2506 	ACQUIRE_LOCK(&lk);
2507 	process_removes(vp);
2508 	if (journal_space(ump, 0) == 0) {
2509 		softdep_speedup();
2510 		if (journal_space(ump, 1) == 0)
2511 			journal_suspend(ump);
2512 	}
2513 	FREE_LOCK(&lk);
2514 
2515 	return (0);
2516 }
2517 
2518 /*
2519  * Before adjusting a link count on a vnode verify that we have sufficient
2520  * journal space.  If not, process operations that depend on the currently
2521  * locked pair of vnodes to try to flush space as the syncer, buf daemon,
2522  * and softdep flush threads can not acquire these locks to reclaim space.
2523  */
2524 static void
2525 softdep_prelink(dvp, vp)
2526 	struct vnode *dvp;
2527 	struct vnode *vp;
2528 {
2529 	struct ufsmount *ump;
2530 
2531 	ump = VFSTOUFS(dvp->v_mount);
2532 	mtx_assert(&lk, MA_OWNED);
2533 	if (journal_space(ump, 0))
2534 		return;
2535 	stat_journal_low++;
2536 	FREE_LOCK(&lk);
2537 	if (vp)
2538 		ffs_syncvnode(vp, MNT_NOWAIT);
2539 	ffs_syncvnode(dvp, MNT_WAIT);
2540 	ACQUIRE_LOCK(&lk);
2541 	/* Process vp before dvp as it may create .. removes. */
2542 	if (vp)
2543 		process_removes(vp);
2544 	process_removes(dvp);
2545 	softdep_speedup();
2546 	process_worklist_item(UFSTOVFS(ump), LK_NOWAIT);
2547 	process_worklist_item(UFSTOVFS(ump), LK_NOWAIT);
2548 	if (journal_space(ump, 0) == 0) {
2549 		softdep_speedup();
2550 		if (journal_space(ump, 1) == 0)
2551 			journal_suspend(ump);
2552 	}
2553 }
2554 
2555 static void
2556 jseg_write(fs, jblocks, jseg, data)
2557 	struct fs *fs;
2558 	struct jblocks *jblocks;
2559 	struct jseg *jseg;
2560 	uint8_t *data;
2561 {
2562 	struct jsegrec *rec;
2563 
2564 	rec = (struct jsegrec *)data;
2565 	rec->jsr_seq = jseg->js_seq;
2566 	rec->jsr_oldest = jblocks->jb_oldestseq;
2567 	rec->jsr_cnt = jseg->js_cnt;
2568 	rec->jsr_blocks = jseg->js_size / DEV_BSIZE;
2569 	rec->jsr_crc = 0;
2570 	rec->jsr_time = fs->fs_mtime;
2571 }
2572 
2573 static inline void
2574 inoref_write(inoref, jseg, rec)
2575 	struct inoref *inoref;
2576 	struct jseg *jseg;
2577 	struct jrefrec *rec;
2578 {
2579 
2580 	inoref->if_jsegdep->jd_seg = jseg;
2581 	rec->jr_ino = inoref->if_ino;
2582 	rec->jr_parent = inoref->if_parent;
2583 	rec->jr_nlink = inoref->if_nlink;
2584 	rec->jr_mode = inoref->if_mode;
2585 	rec->jr_diroff = inoref->if_diroff;
2586 }
2587 
2588 static void
2589 jaddref_write(jaddref, jseg, data)
2590 	struct jaddref *jaddref;
2591 	struct jseg *jseg;
2592 	uint8_t *data;
2593 {
2594 	struct jrefrec *rec;
2595 
2596 	rec = (struct jrefrec *)data;
2597 	rec->jr_op = JOP_ADDREF;
2598 	inoref_write(&jaddref->ja_ref, jseg, rec);
2599 }
2600 
2601 static void
2602 jremref_write(jremref, jseg, data)
2603 	struct jremref *jremref;
2604 	struct jseg *jseg;
2605 	uint8_t *data;
2606 {
2607 	struct jrefrec *rec;
2608 
2609 	rec = (struct jrefrec *)data;
2610 	rec->jr_op = JOP_REMREF;
2611 	inoref_write(&jremref->jr_ref, jseg, rec);
2612 }
2613 
2614 static void
2615 jmvref_write(jmvref, jseg, data)
2616 	struct jmvref *jmvref;
2617 	struct jseg *jseg;
2618 	uint8_t *data;
2619 {
2620 	struct jmvrec *rec;
2621 
2622 	rec = (struct jmvrec *)data;
2623 	rec->jm_op = JOP_MVREF;
2624 	rec->jm_ino = jmvref->jm_ino;
2625 	rec->jm_parent = jmvref->jm_parent;
2626 	rec->jm_oldoff = jmvref->jm_oldoff;
2627 	rec->jm_newoff = jmvref->jm_newoff;
2628 }
2629 
2630 static void
2631 jnewblk_write(jnewblk, jseg, data)
2632 	struct jnewblk *jnewblk;
2633 	struct jseg *jseg;
2634 	uint8_t *data;
2635 {
2636 	struct jblkrec *rec;
2637 
2638 	jnewblk->jn_jsegdep->jd_seg = jseg;
2639 	rec = (struct jblkrec *)data;
2640 	rec->jb_op = JOP_NEWBLK;
2641 	rec->jb_ino = jnewblk->jn_ino;
2642 	rec->jb_blkno = jnewblk->jn_blkno;
2643 	rec->jb_lbn = jnewblk->jn_lbn;
2644 	rec->jb_frags = jnewblk->jn_frags;
2645 	rec->jb_oldfrags = jnewblk->jn_oldfrags;
2646 }
2647 
2648 static void
2649 jfreeblk_write(jfreeblk, jseg, data)
2650 	struct jfreeblk *jfreeblk;
2651 	struct jseg *jseg;
2652 	uint8_t *data;
2653 {
2654 	struct jblkrec *rec;
2655 
2656 	jfreeblk->jf_jsegdep->jd_seg = jseg;
2657 	rec = (struct jblkrec *)data;
2658 	rec->jb_op = JOP_FREEBLK;
2659 	rec->jb_ino = jfreeblk->jf_ino;
2660 	rec->jb_blkno = jfreeblk->jf_blkno;
2661 	rec->jb_lbn = jfreeblk->jf_lbn;
2662 	rec->jb_frags = jfreeblk->jf_frags;
2663 	rec->jb_oldfrags = 0;
2664 }
2665 
2666 static void
2667 jfreefrag_write(jfreefrag, jseg, data)
2668 	struct jfreefrag *jfreefrag;
2669 	struct jseg *jseg;
2670 	uint8_t *data;
2671 {
2672 	struct jblkrec *rec;
2673 
2674 	jfreefrag->fr_jsegdep->jd_seg = jseg;
2675 	rec = (struct jblkrec *)data;
2676 	rec->jb_op = JOP_FREEBLK;
2677 	rec->jb_ino = jfreefrag->fr_ino;
2678 	rec->jb_blkno = jfreefrag->fr_blkno;
2679 	rec->jb_lbn = jfreefrag->fr_lbn;
2680 	rec->jb_frags = jfreefrag->fr_frags;
2681 	rec->jb_oldfrags = 0;
2682 }
2683 
2684 static void
2685 jtrunc_write(jtrunc, jseg, data)
2686 	struct jtrunc *jtrunc;
2687 	struct jseg *jseg;
2688 	uint8_t *data;
2689 {
2690 	struct jtrncrec *rec;
2691 
2692 	rec = (struct jtrncrec *)data;
2693 	rec->jt_op = JOP_TRUNC;
2694 	rec->jt_ino = jtrunc->jt_ino;
2695 	rec->jt_size = jtrunc->jt_size;
2696 	rec->jt_extsize = jtrunc->jt_extsize;
2697 }
2698 
2699 /*
2700  * Flush some journal records to disk.
2701  */
2702 static void
2703 softdep_process_journal(mp, flags)
2704 	struct mount *mp;
2705 	int flags;
2706 {
2707 	struct jblocks *jblocks;
2708 	struct ufsmount *ump;
2709 	struct worklist *wk;
2710 	struct jseg *jseg;
2711 	struct buf *bp;
2712 	uint8_t *data;
2713 	struct fs *fs;
2714 	int segwritten;
2715 	int jrecmin;	/* Minimum records per block. */
2716 	int jrecmax;	/* Maximum records per block. */
2717 	int size;
2718 	int cnt;
2719 	int off;
2720 
2721 	if ((mp->mnt_kern_flag & MNTK_SUJ) == 0)
2722 		return;
2723 	ump = VFSTOUFS(mp);
2724 	fs = ump->um_fs;
2725 	jblocks = ump->softdep_jblocks;
2726 	/*
2727 	 * We write anywhere between a disk block and fs block.  The upper
2728 	 * bound is picked to prevent buffer cache fragmentation and limit
2729 	 * processing time per I/O.
2730 	 */
2731 	jrecmin = (DEV_BSIZE / JREC_SIZE) - 1; /* -1 for seg header */
2732 	jrecmax = (fs->fs_bsize / DEV_BSIZE) * jrecmin;
2733 	segwritten = 0;
2734 	while ((cnt = ump->softdep_on_journal) != 0) {
2735 		/*
2736 		 * Create a new segment to hold as many as 'cnt' journal
2737 		 * entries and add them to the segment.  Notice cnt is
2738 		 * off by one to account for the space required by the
2739 		 * jsegrec.  If we don't have a full block to log skip it
2740 		 * unless we haven't written anything.
2741 		 */
2742 		cnt++;
2743 		if (cnt < jrecmax && segwritten)
2744 			break;
2745 		/*
2746 		 * Verify some free journal space.  softdep_prealloc() should
2747 	 	 * guarantee that we don't run out so this is indicative of
2748 		 * a problem with the flow control.  Try to recover
2749 		 * gracefully in any event.
2750 		 */
2751 		while (jblocks->jb_free == 0) {
2752 			if (flags != MNT_WAIT)
2753 				break;
2754 			printf("softdep: Out of journal space!\n");
2755 			softdep_speedup();
2756 			msleep(jblocks, &lk, PRIBIO, "jblocks", hz);
2757 		}
2758 		FREE_LOCK(&lk);
2759 		jseg = malloc(sizeof(*jseg), M_JSEG, M_SOFTDEP_FLAGS);
2760 		workitem_alloc(&jseg->js_list, D_JSEG, mp);
2761 		LIST_INIT(&jseg->js_entries);
2762 		jseg->js_state = ATTACHED;
2763 		jseg->js_jblocks = jblocks;
2764 		bp = geteblk(fs->fs_bsize, 0);
2765 		ACQUIRE_LOCK(&lk);
2766 		/*
2767 		 * If there was a race while we were allocating the block
2768 		 * and jseg the entry we care about was likely written.
2769 		 * We bail out in both the WAIT and NOWAIT case and assume
2770 		 * the caller will loop if the entry it cares about is
2771 		 * not written.
2772 		 */
2773 		if (ump->softdep_on_journal == 0 || jblocks->jb_free == 0) {
2774 			bp->b_flags |= B_INVAL | B_NOCACHE;
2775 			WORKITEM_FREE(jseg, D_JSEG);
2776 			FREE_LOCK(&lk);
2777 			brelse(bp);
2778 			ACQUIRE_LOCK(&lk);
2779 			break;
2780 		}
2781 		/*
2782 		 * Calculate the disk block size required for the available
2783 		 * records rounded to the min size.
2784 		 */
2785 		cnt = ump->softdep_on_journal;
2786 		if (cnt < jrecmax)
2787 			size = howmany(cnt, jrecmin) * DEV_BSIZE;
2788 		else
2789 			size = fs->fs_bsize;
2790 		/*
2791 		 * Allocate a disk block for this journal data and account
2792 		 * for truncation of the requested size if enough contiguous
2793 		 * space was not available.
2794 		 */
2795 		bp->b_blkno = jblocks_alloc(jblocks, size, &size);
2796 		bp->b_lblkno = bp->b_blkno;
2797 		bp->b_offset = bp->b_blkno * DEV_BSIZE;
2798 		bp->b_bcount = size;
2799 		bp->b_bufobj = &ump->um_devvp->v_bufobj;
2800 		bp->b_flags &= ~B_INVAL;
2801 		bp->b_flags |= B_VALIDSUSPWRT | B_NOCOPY;
2802 		/*
2803 		 * Initialize our jseg with cnt records.  Assign the next
2804 		 * sequence number to it and link it in-order.
2805 		 */
2806 		cnt = MIN(ump->softdep_on_journal,
2807 		    (size / DEV_BSIZE) * jrecmin);
2808 		jseg->js_buf = bp;
2809 		jseg->js_cnt = cnt;
2810 		jseg->js_refs = cnt + 1;	/* Self ref. */
2811 		jseg->js_size = size;
2812 		jseg->js_seq = jblocks->jb_nextseq++;
2813 		if (TAILQ_EMPTY(&jblocks->jb_segs))
2814 			jblocks->jb_oldestseq = jseg->js_seq;
2815 		TAILQ_INSERT_TAIL(&jblocks->jb_segs, jseg, js_next);
2816 		if (jblocks->jb_writeseg == NULL)
2817 			jblocks->jb_writeseg = jseg;
2818 		/*
2819 		 * Start filling in records from the pending list.
2820 		 */
2821 		data = bp->b_data;
2822 		off = 0;
2823 		while ((wk = LIST_FIRST(&ump->softdep_journal_pending))
2824 		    != NULL) {
2825 			/* Place a segment header on every device block. */
2826 			if ((off % DEV_BSIZE) == 0) {
2827 				jseg_write(fs, jblocks, jseg, data);
2828 				off += JREC_SIZE;
2829 				data = bp->b_data + off;
2830 			}
2831 			remove_from_journal(wk);
2832 			wk->wk_state |= IOSTARTED;
2833 			WORKLIST_INSERT(&jseg->js_entries, wk);
2834 			switch (wk->wk_type) {
2835 			case D_JADDREF:
2836 				jaddref_write(WK_JADDREF(wk), jseg, data);
2837 				break;
2838 			case D_JREMREF:
2839 				jremref_write(WK_JREMREF(wk), jseg, data);
2840 				break;
2841 			case D_JMVREF:
2842 				jmvref_write(WK_JMVREF(wk), jseg, data);
2843 				break;
2844 			case D_JNEWBLK:
2845 				jnewblk_write(WK_JNEWBLK(wk), jseg, data);
2846 				break;
2847 			case D_JFREEBLK:
2848 				jfreeblk_write(WK_JFREEBLK(wk), jseg, data);
2849 				break;
2850 			case D_JFREEFRAG:
2851 				jfreefrag_write(WK_JFREEFRAG(wk), jseg, data);
2852 				break;
2853 			case D_JTRUNC:
2854 				jtrunc_write(WK_JTRUNC(wk), jseg, data);
2855 				break;
2856 			default:
2857 				panic("process_journal: Unknown type %s",
2858 				    TYPENAME(wk->wk_type));
2859 				/* NOTREACHED */
2860 			}
2861 			if (--cnt == 0)
2862 				break;
2863 			off += JREC_SIZE;
2864 			data = bp->b_data + off;
2865 		}
2866 		/*
2867 		 * Write this one buffer and continue.
2868 		 */
2869 		WORKLIST_INSERT(&bp->b_dep, &jseg->js_list);
2870 		FREE_LOCK(&lk);
2871 		BO_LOCK(bp->b_bufobj);
2872 		bgetvp(ump->um_devvp, bp);
2873 		BO_UNLOCK(bp->b_bufobj);
2874 		if (flags == MNT_NOWAIT)
2875 			bawrite(bp);
2876 		else
2877 			bwrite(bp);
2878 		ACQUIRE_LOCK(&lk);
2879 	}
2880 	/*
2881 	 * If we've suspended the filesystem because we ran out of journal
2882 	 * space either try to sync it here to make some progress or
2883 	 * unsuspend it if we already have.
2884 	 */
2885 	if (flags == 0 && jblocks->jb_suspended) {
2886 		if (journal_unsuspend(ump))
2887 			return;
2888 		FREE_LOCK(&lk);
2889 		VFS_SYNC(mp, MNT_NOWAIT);
2890 		ffs_sbupdate(ump, MNT_WAIT, 0);
2891 		ACQUIRE_LOCK(&lk);
2892 	}
2893 }
2894 
2895 /*
2896  * Complete a jseg, allowing all dependencies awaiting journal writes
2897  * to proceed.  Each journal dependency also attaches a jsegdep to dependent
2898  * structures so that the journal segment can be freed to reclaim space.
2899  */
2900 static void
2901 complete_jseg(jseg)
2902 	struct jseg *jseg;
2903 {
2904 	struct worklist *wk;
2905 	struct jmvref *jmvref;
2906 	int waiting;
2907 #ifdef INVARIANTS
2908 	int i = 0;
2909 #endif
2910 
2911 	while ((wk = LIST_FIRST(&jseg->js_entries)) != NULL) {
2912 		WORKLIST_REMOVE(wk);
2913 		waiting = wk->wk_state & IOWAITING;
2914 		wk->wk_state &= ~(IOSTARTED | IOWAITING);
2915 		wk->wk_state |= COMPLETE;
2916 		KASSERT(i++ < jseg->js_cnt,
2917 		    ("handle_written_jseg: overflow %d >= %d",
2918 		    i - 1, jseg->js_cnt));
2919 		switch (wk->wk_type) {
2920 		case D_JADDREF:
2921 			handle_written_jaddref(WK_JADDREF(wk));
2922 			break;
2923 		case D_JREMREF:
2924 			handle_written_jremref(WK_JREMREF(wk));
2925 			break;
2926 		case D_JMVREF:
2927 			/* No jsegdep here. */
2928 			free_jseg(jseg);
2929 			jmvref = WK_JMVREF(wk);
2930 			LIST_REMOVE(jmvref, jm_deps);
2931 			free_pagedep(jmvref->jm_pagedep);
2932 			WORKITEM_FREE(jmvref, D_JMVREF);
2933 			break;
2934 		case D_JNEWBLK:
2935 			handle_written_jnewblk(WK_JNEWBLK(wk));
2936 			break;
2937 		case D_JFREEBLK:
2938 			handle_written_jfreeblk(WK_JFREEBLK(wk));
2939 			break;
2940 		case D_JFREEFRAG:
2941 			handle_written_jfreefrag(WK_JFREEFRAG(wk));
2942 			break;
2943 		case D_JTRUNC:
2944 			WK_JTRUNC(wk)->jt_jsegdep->jd_seg = jseg;
2945 			WORKITEM_FREE(wk, D_JTRUNC);
2946 			break;
2947 		default:
2948 			panic("handle_written_jseg: Unknown type %s",
2949 			    TYPENAME(wk->wk_type));
2950 			/* NOTREACHED */
2951 		}
2952 		if (waiting)
2953 			wakeup(wk);
2954 	}
2955 	/* Release the self reference so the structure may be freed. */
2956 	free_jseg(jseg);
2957 }
2958 
2959 /*
2960  * Mark a jseg as DEPCOMPLETE and throw away the buffer.  Handle jseg
2961  * completions in order only.
2962  */
2963 static void
2964 handle_written_jseg(jseg, bp)
2965 	struct jseg *jseg;
2966 	struct buf *bp;
2967 {
2968 	struct jblocks *jblocks;
2969 	struct jseg *jsegn;
2970 
2971 	if (jseg->js_refs == 0)
2972 		panic("handle_written_jseg: No self-reference on %p", jseg);
2973 	jseg->js_state |= DEPCOMPLETE;
2974 	/*
2975 	 * We'll never need this buffer again, set flags so it will be
2976 	 * discarded.
2977 	 */
2978 	bp->b_flags |= B_INVAL | B_NOCACHE;
2979 	jblocks = jseg->js_jblocks;
2980 	/*
2981 	 * Don't allow out of order completions.  If this isn't the first
2982 	 * block wait for it to write before we're done.
2983 	 */
2984 	if (jseg != jblocks->jb_writeseg)
2985 		return;
2986 	/* Iterate through available jsegs processing their entries. */
2987 	do {
2988 		jsegn = TAILQ_NEXT(jseg, js_next);
2989 		complete_jseg(jseg);
2990 		jseg = jsegn;
2991 	} while (jseg && jseg->js_state & DEPCOMPLETE);
2992 	jblocks->jb_writeseg = jseg;
2993 }
2994 
2995 static inline struct jsegdep *
2996 inoref_jseg(inoref)
2997 	struct inoref *inoref;
2998 {
2999 	struct jsegdep *jsegdep;
3000 
3001 	jsegdep = inoref->if_jsegdep;
3002 	inoref->if_jsegdep = NULL;
3003 
3004 	return (jsegdep);
3005 }
3006 
3007 /*
3008  * Called once a jremref has made it to stable store.  The jremref is marked
3009  * complete and we attempt to free it.  Any pagedeps writes sleeping waiting
3010  * for the jremref to complete will be awoken by free_jremref.
3011  */
3012 static void
3013 handle_written_jremref(jremref)
3014 	struct jremref *jremref;
3015 {
3016 	struct inodedep *inodedep;
3017 	struct jsegdep *jsegdep;
3018 	struct dirrem *dirrem;
3019 
3020 	/* Grab the jsegdep. */
3021 	jsegdep = inoref_jseg(&jremref->jr_ref);
3022 	/*
3023 	 * Remove us from the inoref list.
3024 	 */
3025 	if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino,
3026 	    0, &inodedep) == 0)
3027 		panic("handle_written_jremref: Lost inodedep");
3028 	TAILQ_REMOVE(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps);
3029 	/*
3030 	 * Complete the dirrem.
3031 	 */
3032 	dirrem = jremref->jr_dirrem;
3033 	jremref->jr_dirrem = NULL;
3034 	LIST_REMOVE(jremref, jr_deps);
3035 	jsegdep->jd_state |= jremref->jr_state & MKDIR_PARENT;
3036 	WORKLIST_INSERT(&dirrem->dm_jwork, &jsegdep->jd_list);
3037 	if (LIST_EMPTY(&dirrem->dm_jremrefhd) &&
3038 	    (dirrem->dm_state & COMPLETE) != 0)
3039 		add_to_worklist(&dirrem->dm_list, 0);
3040 	free_jremref(jremref);
3041 }
3042 
3043 /*
3044  * Called once a jaddref has made it to stable store.  The dependency is
3045  * marked complete and any dependent structures are added to the inode
3046  * bufwait list to be completed as soon as it is written.  If a bitmap write
3047  * depends on this entry we move the inode into the inodedephd of the
3048  * bmsafemap dependency and attempt to remove the jaddref from the bmsafemap.
3049  */
3050 static void
3051 handle_written_jaddref(jaddref)
3052 	struct jaddref *jaddref;
3053 {
3054 	struct jsegdep *jsegdep;
3055 	struct inodedep *inodedep;
3056 	struct diradd *diradd;
3057 	struct mkdir *mkdir;
3058 
3059 	/* Grab the jsegdep. */
3060 	jsegdep = inoref_jseg(&jaddref->ja_ref);
3061 	mkdir = NULL;
3062 	diradd = NULL;
3063 	if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino,
3064 	    0, &inodedep) == 0)
3065 		panic("handle_written_jaddref: Lost inodedep.");
3066 	if (jaddref->ja_diradd == NULL)
3067 		panic("handle_written_jaddref: No dependency");
3068 	if (jaddref->ja_diradd->da_list.wk_type == D_DIRADD) {
3069 		diradd = jaddref->ja_diradd;
3070 		WORKLIST_INSERT(&inodedep->id_bufwait, &diradd->da_list);
3071 	} else if (jaddref->ja_state & MKDIR_PARENT) {
3072 		mkdir = jaddref->ja_mkdir;
3073 		WORKLIST_INSERT(&inodedep->id_bufwait, &mkdir->md_list);
3074 	} else if (jaddref->ja_state & MKDIR_BODY)
3075 		mkdir = jaddref->ja_mkdir;
3076 	else
3077 		panic("handle_written_jaddref: Unknown dependency %p",
3078 		    jaddref->ja_diradd);
3079 	jaddref->ja_diradd = NULL;	/* also clears ja_mkdir */
3080 	/*
3081 	 * Remove us from the inode list.
3082 	 */
3083 	TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref, if_deps);
3084 	/*
3085 	 * The mkdir may be waiting on the jaddref to clear before freeing.
3086 	 */
3087 	if (mkdir) {
3088 		KASSERT(mkdir->md_list.wk_type == D_MKDIR,
3089 		    ("handle_written_jaddref: Incorrect type for mkdir %s",
3090 		    TYPENAME(mkdir->md_list.wk_type)));
3091 		mkdir->md_jaddref = NULL;
3092 		diradd = mkdir->md_diradd;
3093 		mkdir->md_state |= DEPCOMPLETE;
3094 		complete_mkdir(mkdir);
3095 	}
3096 	WORKLIST_INSERT(&diradd->da_jwork, &jsegdep->jd_list);
3097 	if (jaddref->ja_state & NEWBLOCK) {
3098 		inodedep->id_state |= ONDEPLIST;
3099 		LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_inodedephd,
3100 		    inodedep, id_deps);
3101 	}
3102 	free_jaddref(jaddref);
3103 }
3104 
3105 /*
3106  * Called once a jnewblk journal is written.  The allocdirect or allocindir
3107  * is placed in the bmsafemap to await notification of a written bitmap.
3108  */
3109 static void
3110 handle_written_jnewblk(jnewblk)
3111 	struct jnewblk *jnewblk;
3112 {
3113 	struct bmsafemap *bmsafemap;
3114 	struct jsegdep *jsegdep;
3115 	struct newblk *newblk;
3116 
3117 	/* Grab the jsegdep. */
3118 	jsegdep = jnewblk->jn_jsegdep;
3119 	jnewblk->jn_jsegdep = NULL;
3120 	/*
3121 	 * Add the written block to the bmsafemap so it can be notified when
3122 	 * the bitmap is on disk.
3123 	 */
3124 	newblk = jnewblk->jn_newblk;
3125 	jnewblk->jn_newblk = NULL;
3126 	if (newblk == NULL)
3127 		panic("handle_written_jnewblk: No dependency for the segdep.");
3128 
3129 	newblk->nb_jnewblk = NULL;
3130 	bmsafemap = newblk->nb_bmsafemap;
3131 	WORKLIST_INSERT(&newblk->nb_jwork, &jsegdep->jd_list);
3132 	newblk->nb_state |= ONDEPLIST;
3133 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
3134 	free_jnewblk(jnewblk);
3135 }
3136 
3137 /*
3138  * Cancel a jfreefrag that won't be needed, probably due to colliding with
3139  * an in-flight allocation that has not yet been committed.  Divorce us
3140  * from the freefrag and mark it DEPCOMPLETE so that it may be added
3141  * to the worklist.
3142  */
3143 static void
3144 cancel_jfreefrag(jfreefrag)
3145 	struct jfreefrag *jfreefrag;
3146 {
3147 	struct freefrag *freefrag;
3148 
3149 	if (jfreefrag->fr_jsegdep) {
3150 		free_jsegdep(jfreefrag->fr_jsegdep);
3151 		jfreefrag->fr_jsegdep = NULL;
3152 	}
3153 	freefrag = jfreefrag->fr_freefrag;
3154 	jfreefrag->fr_freefrag = NULL;
3155 	freefrag->ff_jfreefrag = NULL;
3156 	free_jfreefrag(jfreefrag);
3157 	freefrag->ff_state |= DEPCOMPLETE;
3158 }
3159 
3160 /*
3161  * Free a jfreefrag when the parent freefrag is rendered obsolete.
3162  */
3163 static void
3164 free_jfreefrag(jfreefrag)
3165 	struct jfreefrag *jfreefrag;
3166 {
3167 
3168 	if (jfreefrag->fr_state & IOSTARTED)
3169 		WORKLIST_REMOVE(&jfreefrag->fr_list);
3170 	else if (jfreefrag->fr_state & ONWORKLIST)
3171 		remove_from_journal(&jfreefrag->fr_list);
3172 	if (jfreefrag->fr_freefrag != NULL)
3173 		panic("free_jfreefrag:  Still attached to a freefrag.");
3174 	WORKITEM_FREE(jfreefrag, D_JFREEFRAG);
3175 }
3176 
3177 /*
3178  * Called when the journal write for a jfreefrag completes.  The parent
3179  * freefrag is added to the worklist if this completes its dependencies.
3180  */
3181 static void
3182 handle_written_jfreefrag(jfreefrag)
3183 	struct jfreefrag *jfreefrag;
3184 {
3185 	struct jsegdep *jsegdep;
3186 	struct freefrag *freefrag;
3187 
3188 	/* Grab the jsegdep. */
3189 	jsegdep = jfreefrag->fr_jsegdep;
3190 	jfreefrag->fr_jsegdep = NULL;
3191 	freefrag = jfreefrag->fr_freefrag;
3192 	if (freefrag == NULL)
3193 		panic("handle_written_jfreefrag: No freefrag.");
3194 	freefrag->ff_state |= DEPCOMPLETE;
3195 	freefrag->ff_jfreefrag = NULL;
3196 	WORKLIST_INSERT(&freefrag->ff_jwork, &jsegdep->jd_list);
3197 	if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE)
3198 		add_to_worklist(&freefrag->ff_list, 0);
3199 	jfreefrag->fr_freefrag = NULL;
3200 	free_jfreefrag(jfreefrag);
3201 }
3202 
3203 /*
3204  * Called when the journal write for a jfreeblk completes.  The jfreeblk
3205  * is removed from the freeblks list of pending journal writes and the
3206  * jsegdep is moved to the freeblks jwork to be completed when all blocks
3207  * have been reclaimed.
3208  */
3209 static void
3210 handle_written_jfreeblk(jfreeblk)
3211 	struct jfreeblk *jfreeblk;
3212 {
3213 	struct freeblks *freeblks;
3214 	struct jsegdep *jsegdep;
3215 
3216 	/* Grab the jsegdep. */
3217 	jsegdep = jfreeblk->jf_jsegdep;
3218 	jfreeblk->jf_jsegdep = NULL;
3219 	freeblks = jfreeblk->jf_freeblks;
3220 	LIST_REMOVE(jfreeblk, jf_deps);
3221 	WORKLIST_INSERT(&freeblks->fb_jwork, &jsegdep->jd_list);
3222 	/*
3223 	 * If the freeblks is all journaled, we can add it to the worklist.
3224 	 */
3225 	if (LIST_EMPTY(&freeblks->fb_jfreeblkhd) &&
3226 	    (freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE) {
3227 		/* Remove from the b_dep that is waiting on this write. */
3228 		if (freeblks->fb_state & ONWORKLIST)
3229 			WORKLIST_REMOVE(&freeblks->fb_list);
3230 		add_to_worklist(&freeblks->fb_list, 1);
3231 	}
3232 
3233 	free_jfreeblk(jfreeblk);
3234 }
3235 
3236 static struct jsegdep *
3237 newjsegdep(struct worklist *wk)
3238 {
3239 	struct jsegdep *jsegdep;
3240 
3241 	jsegdep = malloc(sizeof(*jsegdep), M_JSEGDEP, M_SOFTDEP_FLAGS);
3242 	workitem_alloc(&jsegdep->jd_list, D_JSEGDEP, wk->wk_mp);
3243 	jsegdep->jd_seg = NULL;
3244 
3245 	return (jsegdep);
3246 }
3247 
3248 static struct jmvref *
3249 newjmvref(dp, ino, oldoff, newoff)
3250 	struct inode *dp;
3251 	ino_t ino;
3252 	off_t oldoff;
3253 	off_t newoff;
3254 {
3255 	struct jmvref *jmvref;
3256 
3257 	jmvref = malloc(sizeof(*jmvref), M_JMVREF, M_SOFTDEP_FLAGS);
3258 	workitem_alloc(&jmvref->jm_list, D_JMVREF, UFSTOVFS(dp->i_ump));
3259 	jmvref->jm_list.wk_state = ATTACHED | DEPCOMPLETE;
3260 	jmvref->jm_parent = dp->i_number;
3261 	jmvref->jm_ino = ino;
3262 	jmvref->jm_oldoff = oldoff;
3263 	jmvref->jm_newoff = newoff;
3264 
3265 	return (jmvref);
3266 }
3267 
3268 /*
3269  * Allocate a new jremref that tracks the removal of ip from dp with the
3270  * directory entry offset of diroff.  Mark the entry as ATTACHED and
3271  * DEPCOMPLETE as we have all the information required for the journal write
3272  * and the directory has already been removed from the buffer.  The caller
3273  * is responsible for linking the jremref into the pagedep and adding it
3274  * to the journal to write.  The MKDIR_PARENT flag is set if we're doing
3275  * a DOTDOT addition so handle_workitem_remove() can properly assign
3276  * the jsegdep when we're done.
3277  */
3278 static struct jremref *
3279 newjremref(struct dirrem *dirrem, struct inode *dp, struct inode *ip,
3280     off_t diroff, nlink_t nlink)
3281 {
3282 	struct jremref *jremref;
3283 
3284 	jremref = malloc(sizeof(*jremref), M_JREMREF, M_SOFTDEP_FLAGS);
3285 	workitem_alloc(&jremref->jr_list, D_JREMREF, UFSTOVFS(dp->i_ump));
3286 	jremref->jr_state = ATTACHED;
3287 	newinoref(&jremref->jr_ref, ip->i_number, dp->i_number, diroff,
3288 	   nlink, ip->i_mode);
3289 	jremref->jr_dirrem = dirrem;
3290 
3291 	return (jremref);
3292 }
3293 
3294 static inline void
3295 newinoref(struct inoref *inoref, ino_t ino, ino_t parent, off_t diroff,
3296     nlink_t nlink, uint16_t mode)
3297 {
3298 
3299 	inoref->if_jsegdep = newjsegdep(&inoref->if_list);
3300 	inoref->if_diroff = diroff;
3301 	inoref->if_ino = ino;
3302 	inoref->if_parent = parent;
3303 	inoref->if_nlink = nlink;
3304 	inoref->if_mode = mode;
3305 }
3306 
3307 /*
3308  * Allocate a new jaddref to track the addition of ino to dp at diroff.  The
3309  * directory offset may not be known until later.  The caller is responsible
3310  * adding the entry to the journal when this information is available.  nlink
3311  * should be the link count prior to the addition and mode is only required
3312  * to have the correct FMT.
3313  */
3314 static struct jaddref *
3315 newjaddref(struct inode *dp, ino_t ino, off_t diroff, int16_t nlink,
3316     uint16_t mode)
3317 {
3318 	struct jaddref *jaddref;
3319 
3320 	jaddref = malloc(sizeof(*jaddref), M_JADDREF, M_SOFTDEP_FLAGS);
3321 	workitem_alloc(&jaddref->ja_list, D_JADDREF, UFSTOVFS(dp->i_ump));
3322 	jaddref->ja_state = ATTACHED;
3323 	jaddref->ja_mkdir = NULL;
3324 	newinoref(&jaddref->ja_ref, ino, dp->i_number, diroff, nlink, mode);
3325 
3326 	return (jaddref);
3327 }
3328 
3329 /*
3330  * Create a new free dependency for a freework.  The caller is responsible
3331  * for adjusting the reference count when it has the lock held.  The freedep
3332  * will track an outstanding bitmap write that will ultimately clear the
3333  * freework to continue.
3334  */
3335 static struct freedep *
3336 newfreedep(struct freework *freework)
3337 {
3338 	struct freedep *freedep;
3339 
3340 	freedep = malloc(sizeof(*freedep), M_FREEDEP, M_SOFTDEP_FLAGS);
3341 	workitem_alloc(&freedep->fd_list, D_FREEDEP, freework->fw_list.wk_mp);
3342 	freedep->fd_freework = freework;
3343 
3344 	return (freedep);
3345 }
3346 
3347 /*
3348  * Free a freedep structure once the buffer it is linked to is written.  If
3349  * this is the last reference to the freework schedule it for completion.
3350  */
3351 static void
3352 free_freedep(freedep)
3353 	struct freedep *freedep;
3354 {
3355 
3356 	if (--freedep->fd_freework->fw_ref == 0)
3357 		add_to_worklist(&freedep->fd_freework->fw_list, 1);
3358 	WORKITEM_FREE(freedep, D_FREEDEP);
3359 }
3360 
3361 /*
3362  * Allocate a new freework structure that may be a level in an indirect
3363  * when parent is not NULL or a top level block when it is.  The top level
3364  * freework structures are allocated without lk held and before the freeblks
3365  * is visible outside of softdep_setup_freeblocks().
3366  */
3367 static struct freework *
3368 newfreework(freeblks, parent, lbn, nb, frags, journal)
3369 	struct freeblks *freeblks;
3370 	struct freework *parent;
3371 	ufs_lbn_t lbn;
3372 	ufs2_daddr_t nb;
3373 	int frags;
3374 	int journal;
3375 {
3376 	struct freework *freework;
3377 
3378 	freework = malloc(sizeof(*freework), M_FREEWORK, M_SOFTDEP_FLAGS);
3379 	workitem_alloc(&freework->fw_list, D_FREEWORK, freeblks->fb_list.wk_mp);
3380 	freework->fw_freeblks = freeblks;
3381 	freework->fw_parent = parent;
3382 	freework->fw_lbn = lbn;
3383 	freework->fw_blkno = nb;
3384 	freework->fw_frags = frags;
3385 	freework->fw_ref = 0;
3386 	freework->fw_off = 0;
3387 	LIST_INIT(&freework->fw_jwork);
3388 
3389 	if (parent == NULL) {
3390 		WORKLIST_INSERT_UNLOCKED(&freeblks->fb_freeworkhd,
3391 		    &freework->fw_list);
3392 		freeblks->fb_ref++;
3393 	}
3394 	if (journal)
3395 		newjfreeblk(freeblks, lbn, nb, frags);
3396 
3397 	return (freework);
3398 }
3399 
3400 /*
3401  * Allocate a new jfreeblk to journal top level block pointer when truncating
3402  * a file.  The caller must add this to the worklist when lk is held.
3403  */
3404 static struct jfreeblk *
3405 newjfreeblk(freeblks, lbn, blkno, frags)
3406 	struct freeblks *freeblks;
3407 	ufs_lbn_t lbn;
3408 	ufs2_daddr_t blkno;
3409 	int frags;
3410 {
3411 	struct jfreeblk *jfreeblk;
3412 
3413 	jfreeblk = malloc(sizeof(*jfreeblk), M_JFREEBLK, M_SOFTDEP_FLAGS);
3414 	workitem_alloc(&jfreeblk->jf_list, D_JFREEBLK, freeblks->fb_list.wk_mp);
3415 	jfreeblk->jf_jsegdep = newjsegdep(&jfreeblk->jf_list);
3416 	jfreeblk->jf_state = ATTACHED | DEPCOMPLETE;
3417 	jfreeblk->jf_ino = freeblks->fb_previousinum;
3418 	jfreeblk->jf_lbn = lbn;
3419 	jfreeblk->jf_blkno = blkno;
3420 	jfreeblk->jf_frags = frags;
3421 	jfreeblk->jf_freeblks = freeblks;
3422 	LIST_INSERT_HEAD(&freeblks->fb_jfreeblkhd, jfreeblk, jf_deps);
3423 
3424 	return (jfreeblk);
3425 }
3426 
3427 static void move_newblock_dep(struct jaddref *, struct inodedep *);
3428 /*
3429  * If we're canceling a new bitmap we have to search for another ref
3430  * to move into the bmsafemap dep.  This might be better expressed
3431  * with another structure.
3432  */
3433 static void
3434 move_newblock_dep(jaddref, inodedep)
3435 	struct jaddref *jaddref;
3436 	struct inodedep *inodedep;
3437 {
3438 	struct inoref *inoref;
3439 	struct jaddref *jaddrefn;
3440 
3441 	jaddrefn = NULL;
3442 	for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref;
3443 	    inoref = TAILQ_NEXT(inoref, if_deps)) {
3444 		if ((jaddref->ja_state & NEWBLOCK) &&
3445 		    inoref->if_list.wk_type == D_JADDREF) {
3446 			jaddrefn = (struct jaddref *)inoref;
3447 			break;
3448 		}
3449 	}
3450 	if (jaddrefn == NULL)
3451 		return;
3452 	jaddrefn->ja_state &= ~(ATTACHED | UNDONE);
3453 	jaddrefn->ja_state |= jaddref->ja_state &
3454 	    (ATTACHED | UNDONE | NEWBLOCK);
3455 	jaddref->ja_state &= ~(ATTACHED | UNDONE | NEWBLOCK);
3456 	jaddref->ja_state |= ATTACHED;
3457 	LIST_REMOVE(jaddref, ja_bmdeps);
3458 	LIST_INSERT_HEAD(&inodedep->id_bmsafemap->sm_jaddrefhd, jaddrefn,
3459 	    ja_bmdeps);
3460 }
3461 
3462 /*
3463  * Cancel a jaddref either before it has been written or while it is being
3464  * written.  This happens when a link is removed before the add reaches
3465  * the disk.  The jaddref dependency is kept linked into the bmsafemap
3466  * and inode to prevent the link count or bitmap from reaching the disk
3467  * until handle_workitem_remove() re-adjusts the counts and bitmaps as
3468  * required.
3469  *
3470  * Returns 1 if the canceled addref requires journaling of the remove and
3471  * 0 otherwise.
3472  */
3473 static int
3474 cancel_jaddref(jaddref, inodedep, wkhd)
3475 	struct jaddref *jaddref;
3476 	struct inodedep *inodedep;
3477 	struct workhead *wkhd;
3478 {
3479 	struct inoref *inoref;
3480 	struct jsegdep *jsegdep;
3481 	int needsj;
3482 
3483 	KASSERT((jaddref->ja_state & COMPLETE) == 0,
3484 	    ("cancel_jaddref: Canceling complete jaddref"));
3485 	if (jaddref->ja_state & (IOSTARTED | COMPLETE))
3486 		needsj = 1;
3487 	else
3488 		needsj = 0;
3489 	if (inodedep == NULL)
3490 		if (inodedep_lookup(jaddref->ja_list.wk_mp, jaddref->ja_ino,
3491 		    0, &inodedep) == 0)
3492 			panic("cancel_jaddref: Lost inodedep");
3493 	/*
3494 	 * We must adjust the nlink of any reference operation that follows
3495 	 * us so that it is consistent with the in-memory reference.  This
3496 	 * ensures that inode nlink rollbacks always have the correct link.
3497 	 */
3498 	if (needsj == 0)
3499 		for (inoref = TAILQ_NEXT(&jaddref->ja_ref, if_deps); inoref;
3500 		    inoref = TAILQ_NEXT(inoref, if_deps))
3501 			inoref->if_nlink--;
3502 	jsegdep = inoref_jseg(&jaddref->ja_ref);
3503 	if (jaddref->ja_state & NEWBLOCK)
3504 		move_newblock_dep(jaddref, inodedep);
3505 	if (jaddref->ja_state & IOWAITING) {
3506 		jaddref->ja_state &= ~IOWAITING;
3507 		wakeup(&jaddref->ja_list);
3508 	}
3509 	jaddref->ja_mkdir = NULL;
3510 	if (jaddref->ja_state & IOSTARTED) {
3511 		jaddref->ja_state &= ~IOSTARTED;
3512 		WORKLIST_REMOVE(&jaddref->ja_list);
3513 		WORKLIST_INSERT(wkhd, &jsegdep->jd_list);
3514 	} else {
3515 		free_jsegdep(jsegdep);
3516 		if (jaddref->ja_state & DEPCOMPLETE)
3517 			remove_from_journal(&jaddref->ja_list);
3518 	}
3519 	/*
3520 	 * Leave NEWBLOCK jaddrefs on the inodedep so handle_workitem_remove
3521 	 * can arrange for them to be freed with the bitmap.  Otherwise we
3522 	 * no longer need this addref attached to the inoreflst and it
3523 	 * will incorrectly adjust nlink if we leave it.
3524 	 */
3525 	if ((jaddref->ja_state & NEWBLOCK) == 0) {
3526 		TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref,
3527 		    if_deps);
3528 		jaddref->ja_state |= COMPLETE;
3529 		free_jaddref(jaddref);
3530 		return (needsj);
3531 	}
3532 	jaddref->ja_state |= GOINGAWAY;
3533 	/*
3534 	 * Leave the head of the list for jsegdeps for fast merging.
3535 	 */
3536 	if (LIST_FIRST(wkhd) != NULL) {
3537 		jaddref->ja_state |= ONWORKLIST;
3538 		LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jaddref->ja_list, wk_list);
3539 	} else
3540 		WORKLIST_INSERT(wkhd, &jaddref->ja_list);
3541 
3542 	return (needsj);
3543 }
3544 
3545 /*
3546  * Attempt to free a jaddref structure when some work completes.  This
3547  * should only succeed once the entry is written and all dependencies have
3548  * been notified.
3549  */
3550 static void
3551 free_jaddref(jaddref)
3552 	struct jaddref *jaddref;
3553 {
3554 
3555 	if ((jaddref->ja_state & ALLCOMPLETE) != ALLCOMPLETE)
3556 		return;
3557 	if (jaddref->ja_ref.if_jsegdep)
3558 		panic("free_jaddref: segdep attached to jaddref %p(0x%X)\n",
3559 		    jaddref, jaddref->ja_state);
3560 	if (jaddref->ja_state & NEWBLOCK)
3561 		LIST_REMOVE(jaddref, ja_bmdeps);
3562 	if (jaddref->ja_state & (IOSTARTED | ONWORKLIST))
3563 		panic("free_jaddref: Bad state %p(0x%X)",
3564 		    jaddref, jaddref->ja_state);
3565 	if (jaddref->ja_mkdir != NULL)
3566 		panic("free_jaddref: Work pending, 0x%X\n", jaddref->ja_state);
3567 	WORKITEM_FREE(jaddref, D_JADDREF);
3568 }
3569 
3570 /*
3571  * Free a jremref structure once it has been written or discarded.
3572  */
3573 static void
3574 free_jremref(jremref)
3575 	struct jremref *jremref;
3576 {
3577 
3578 	if (jremref->jr_ref.if_jsegdep)
3579 		free_jsegdep(jremref->jr_ref.if_jsegdep);
3580 	if (jremref->jr_state & IOSTARTED)
3581 		panic("free_jremref: IO still pending");
3582 	WORKITEM_FREE(jremref, D_JREMREF);
3583 }
3584 
3585 /*
3586  * Free a jnewblk structure.
3587  */
3588 static void
3589 free_jnewblk(jnewblk)
3590 	struct jnewblk *jnewblk;
3591 {
3592 
3593 	if ((jnewblk->jn_state & ALLCOMPLETE) != ALLCOMPLETE)
3594 		return;
3595 	LIST_REMOVE(jnewblk, jn_deps);
3596 	if (jnewblk->jn_newblk != NULL)
3597 		panic("free_jnewblk: Dependency still attached.");
3598 	WORKITEM_FREE(jnewblk, D_JNEWBLK);
3599 }
3600 
3601 /*
3602  * Cancel a jnewblk which has been superseded by a freeblk.  The jnewblk
3603  * is kept linked into the bmsafemap until the free completes, thus
3604  * preventing the modified state from ever reaching disk.  The free
3605  * routine must pass this structure via ffs_blkfree() to
3606  * softdep_setup_freeblks() so there is no race in releasing the space.
3607  */
3608 static void
3609 cancel_jnewblk(jnewblk, wkhd)
3610 	struct jnewblk *jnewblk;
3611 	struct workhead *wkhd;
3612 {
3613 	struct jsegdep *jsegdep;
3614 
3615 	jsegdep = jnewblk->jn_jsegdep;
3616 	jnewblk->jn_jsegdep  = NULL;
3617 	free_jsegdep(jsegdep);
3618 	jnewblk->jn_newblk = NULL;
3619 	jnewblk->jn_state |= GOINGAWAY;
3620 	if (jnewblk->jn_state & IOSTARTED) {
3621 		jnewblk->jn_state &= ~IOSTARTED;
3622 		WORKLIST_REMOVE(&jnewblk->jn_list);
3623 	} else
3624 		remove_from_journal(&jnewblk->jn_list);
3625 	/*
3626 	 * Leave the head of the list for jsegdeps for fast merging.
3627 	 */
3628 	if (LIST_FIRST(wkhd) != NULL) {
3629 		jnewblk->jn_state |= ONWORKLIST;
3630 		LIST_INSERT_AFTER(LIST_FIRST(wkhd), &jnewblk->jn_list, wk_list);
3631 	} else
3632 		WORKLIST_INSERT(wkhd, &jnewblk->jn_list);
3633 	if (jnewblk->jn_state & IOWAITING) {
3634 		jnewblk->jn_state &= ~IOWAITING;
3635 		wakeup(&jnewblk->jn_list);
3636 	}
3637 }
3638 
3639 static void
3640 free_jfreeblk(jfreeblk)
3641 	struct jfreeblk *jfreeblk;
3642 {
3643 
3644 	WORKITEM_FREE(jfreeblk, D_JFREEBLK);
3645 }
3646 
3647 /*
3648  * Release one reference to a jseg and free it if the count reaches 0.  This
3649  * should eventually reclaim journal space as well.
3650  */
3651 static void
3652 free_jseg(jseg)
3653 	struct jseg *jseg;
3654 {
3655 	struct jblocks *jblocks;
3656 
3657 	KASSERT(jseg->js_refs > 0,
3658 	    ("free_jseg: Invalid refcnt %d", jseg->js_refs));
3659 	if (--jseg->js_refs != 0)
3660 		return;
3661 	/*
3662 	 * Free only those jsegs which have none allocated before them to
3663 	 * preserve the journal space ordering.
3664 	 */
3665 	jblocks = jseg->js_jblocks;
3666 	while ((jseg = TAILQ_FIRST(&jblocks->jb_segs)) != NULL) {
3667 		jblocks->jb_oldestseq = jseg->js_seq;
3668 		if (jseg->js_refs != 0)
3669 			break;
3670 		TAILQ_REMOVE(&jblocks->jb_segs, jseg, js_next);
3671 		jblocks_free(jblocks, jseg->js_list.wk_mp, jseg->js_size);
3672 		KASSERT(LIST_EMPTY(&jseg->js_entries),
3673 		    ("free_jseg: Freed jseg has valid entries."));
3674 		WORKITEM_FREE(jseg, D_JSEG);
3675 	}
3676 }
3677 
3678 /*
3679  * Release a jsegdep and decrement the jseg count.
3680  */
3681 static void
3682 free_jsegdep(jsegdep)
3683 	struct jsegdep *jsegdep;
3684 {
3685 
3686 	if (jsegdep->jd_seg)
3687 		free_jseg(jsegdep->jd_seg);
3688 	WORKITEM_FREE(jsegdep, D_JSEGDEP);
3689 }
3690 
3691 /*
3692  * Wait for a journal item to make it to disk.  Initiate journal processing
3693  * if required.
3694  */
3695 static void
3696 jwait(wk)
3697 	struct worklist *wk;
3698 {
3699 
3700 	stat_journal_wait++;
3701 	/*
3702 	 * If IO has not started we process the journal.  We can't mark the
3703 	 * worklist item as IOWAITING because we drop the lock while
3704 	 * processing the journal and the worklist entry may be freed after
3705 	 * this point.  The caller may call back in and re-issue the request.
3706 	 */
3707 	if ((wk->wk_state & IOSTARTED) == 0) {
3708 		softdep_process_journal(wk->wk_mp, MNT_WAIT);
3709 		return;
3710 	}
3711 	wk->wk_state |= IOWAITING;
3712 	msleep(wk, &lk, PRIBIO, "jwait", 0);
3713 }
3714 
3715 /*
3716  * Lookup an inodedep based on an inode pointer and set the nlinkdelta as
3717  * appropriate.  This is a convenience function to reduce duplicate code
3718  * for the setup and revert functions below.
3719  */
3720 static struct inodedep *
3721 inodedep_lookup_ip(ip)
3722 	struct inode *ip;
3723 {
3724 	struct inodedep *inodedep;
3725 
3726 	KASSERT(ip->i_nlink >= ip->i_effnlink,
3727 	    ("inodedep_lookup_ip: bad delta"));
3728 	(void) inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number,
3729 	    DEPALLOC, &inodedep);
3730 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3731 
3732 	return (inodedep);
3733 }
3734 
3735 /*
3736  * Create a journal entry that describes a truncate that we're about to
3737  * perform.  The inode allocations and frees between here and the completion
3738  * of the operation are done asynchronously and without journaling.  At
3739  * the end of the operation the vnode is sync'd and the journal space
3740  * is released.  Recovery will discover the partially completed truncate
3741  * and complete it.
3742  */
3743 void *
3744 softdep_setup_trunc(vp, length, flags)
3745 	struct vnode *vp;
3746 	off_t length;
3747 	int flags;
3748 {
3749 	struct jsegdep *jsegdep;
3750 	struct jtrunc *jtrunc;
3751 	struct ufsmount *ump;
3752 	struct inode *ip;
3753 
3754 	softdep_prealloc(vp, MNT_WAIT);
3755 	ip = VTOI(vp);
3756 	ump = VFSTOUFS(vp->v_mount);
3757 	jtrunc = malloc(sizeof(*jtrunc), M_JTRUNC, M_SOFTDEP_FLAGS);
3758 	workitem_alloc(&jtrunc->jt_list, D_JTRUNC, vp->v_mount);
3759 	jsegdep = jtrunc->jt_jsegdep = newjsegdep(&jtrunc->jt_list);
3760 	jtrunc->jt_ino = ip->i_number;
3761 	jtrunc->jt_extsize = 0;
3762 	jtrunc->jt_size = length;
3763 	if ((flags & IO_EXT) == 0 && ump->um_fstype == UFS2)
3764 		jtrunc->jt_extsize = ip->i_din2->di_extsize;
3765 	if ((flags & IO_NORMAL) == 0)
3766 		jtrunc->jt_size = DIP(ip, i_size);
3767 	ACQUIRE_LOCK(&lk);
3768 	add_to_journal(&jtrunc->jt_list);
3769 	while (jsegdep->jd_seg == NULL) {
3770 		stat_jwait_freeblks++;
3771 		jwait(&jtrunc->jt_list);
3772 	}
3773 	FREE_LOCK(&lk);
3774 
3775 	return (jsegdep);
3776 }
3777 
3778 /*
3779  * After synchronous truncation is complete we free sync the vnode and
3780  * release the jsegdep so the journal space can be freed.
3781  */
3782 int
3783 softdep_complete_trunc(vp, cookie)
3784 	struct vnode *vp;
3785 	void *cookie;
3786 {
3787 	int error;
3788 
3789 	error = ffs_syncvnode(vp, MNT_WAIT);
3790 	ACQUIRE_LOCK(&lk);
3791 	free_jsegdep((struct jsegdep *)cookie);
3792 	FREE_LOCK(&lk);
3793 
3794 	return (error);
3795 }
3796 
3797 /*
3798  * Called prior to creating a new inode and linking it to a directory.  The
3799  * jaddref structure must already be allocated by softdep_setup_inomapdep
3800  * and it is discovered here so we can initialize the mode and update
3801  * nlinkdelta.
3802  */
3803 void
3804 softdep_setup_create(dp, ip)
3805 	struct inode *dp;
3806 	struct inode *ip;
3807 {
3808 	struct inodedep *inodedep;
3809 	struct jaddref *jaddref;
3810 	struct vnode *dvp;
3811 
3812 	KASSERT(ip->i_nlink == 1,
3813 	    ("softdep_setup_create: Invalid link count."));
3814 	dvp = ITOV(dp);
3815 	ACQUIRE_LOCK(&lk);
3816 	inodedep = inodedep_lookup_ip(ip);
3817 	if (DOINGSUJ(dvp)) {
3818 		jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
3819 		    inoreflst);
3820 		KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
3821 		    ("softdep_setup_create: No addref structure present."));
3822 		jaddref->ja_mode = ip->i_mode;
3823 	}
3824 	softdep_prelink(dvp, NULL);
3825 	FREE_LOCK(&lk);
3826 }
3827 
3828 /*
3829  * Create a jaddref structure to track the addition of a DOTDOT link when
3830  * we are reparenting an inode as part of a rename.  This jaddref will be
3831  * found by softdep_setup_directory_change.  Adjusts nlinkdelta for
3832  * non-journaling softdep.
3833  */
3834 void
3835 softdep_setup_dotdot_link(dp, ip)
3836 	struct inode *dp;
3837 	struct inode *ip;
3838 {
3839 	struct inodedep *inodedep;
3840 	struct jaddref *jaddref;
3841 	struct vnode *dvp;
3842 	struct vnode *vp;
3843 
3844 	dvp = ITOV(dp);
3845 	vp = ITOV(ip);
3846 	jaddref = NULL;
3847 	/*
3848 	 * We don't set MKDIR_PARENT as this is not tied to a mkdir and
3849 	 * is used as a normal link would be.
3850 	 */
3851 	if (DOINGSUJ(dvp))
3852 		jaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET,
3853 		    dp->i_effnlink - 1, dp->i_mode);
3854 	ACQUIRE_LOCK(&lk);
3855 	inodedep = inodedep_lookup_ip(dp);
3856 	if (jaddref)
3857 		TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
3858 		    if_deps);
3859 	softdep_prelink(dvp, ITOV(ip));
3860 	FREE_LOCK(&lk);
3861 }
3862 
3863 /*
3864  * Create a jaddref structure to track a new link to an inode.  The directory
3865  * offset is not known until softdep_setup_directory_add or
3866  * softdep_setup_directory_change.  Adjusts nlinkdelta for non-journaling
3867  * softdep.
3868  */
3869 void
3870 softdep_setup_link(dp, ip)
3871 	struct inode *dp;
3872 	struct inode *ip;
3873 {
3874 	struct inodedep *inodedep;
3875 	struct jaddref *jaddref;
3876 	struct vnode *dvp;
3877 
3878 	dvp = ITOV(dp);
3879 	jaddref = NULL;
3880 	if (DOINGSUJ(dvp))
3881 		jaddref = newjaddref(dp, ip->i_number, 0, ip->i_effnlink - 1,
3882 		    ip->i_mode);
3883 	ACQUIRE_LOCK(&lk);
3884 	inodedep = inodedep_lookup_ip(ip);
3885 	if (jaddref)
3886 		TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
3887 		    if_deps);
3888 	softdep_prelink(dvp, ITOV(ip));
3889 	FREE_LOCK(&lk);
3890 }
3891 
3892 /*
3893  * Called to create the jaddref structures to track . and .. references as
3894  * well as lookup and further initialize the incomplete jaddref created
3895  * by softdep_setup_inomapdep when the inode was allocated.  Adjusts
3896  * nlinkdelta for non-journaling softdep.
3897  */
3898 void
3899 softdep_setup_mkdir(dp, ip)
3900 	struct inode *dp;
3901 	struct inode *ip;
3902 {
3903 	struct inodedep *inodedep;
3904 	struct jaddref *dotdotaddref;
3905 	struct jaddref *dotaddref;
3906 	struct jaddref *jaddref;
3907 	struct vnode *dvp;
3908 
3909 	dvp = ITOV(dp);
3910 	dotaddref = dotdotaddref = NULL;
3911 	if (DOINGSUJ(dvp)) {
3912 		dotaddref = newjaddref(ip, ip->i_number, DOT_OFFSET, 1,
3913 		    ip->i_mode);
3914 		dotaddref->ja_state |= MKDIR_BODY;
3915 		dotdotaddref = newjaddref(ip, dp->i_number, DOTDOT_OFFSET,
3916 		    dp->i_effnlink - 1, dp->i_mode);
3917 		dotdotaddref->ja_state |= MKDIR_PARENT;
3918 	}
3919 	ACQUIRE_LOCK(&lk);
3920 	inodedep = inodedep_lookup_ip(ip);
3921 	if (DOINGSUJ(dvp)) {
3922 		jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
3923 		    inoreflst);
3924 		KASSERT(jaddref != NULL,
3925 		    ("softdep_setup_mkdir: No addref structure present."));
3926 		KASSERT(jaddref->ja_parent == dp->i_number,
3927 		    ("softdep_setup_mkdir: bad parent %d",
3928 		    jaddref->ja_parent));
3929 		jaddref->ja_mode = ip->i_mode;
3930 		TAILQ_INSERT_BEFORE(&jaddref->ja_ref, &dotaddref->ja_ref,
3931 		    if_deps);
3932 	}
3933 	inodedep = inodedep_lookup_ip(dp);
3934 	if (DOINGSUJ(dvp))
3935 		TAILQ_INSERT_TAIL(&inodedep->id_inoreflst,
3936 		    &dotdotaddref->ja_ref, if_deps);
3937 	softdep_prelink(ITOV(dp), NULL);
3938 	FREE_LOCK(&lk);
3939 }
3940 
3941 /*
3942  * Called to track nlinkdelta of the inode and parent directories prior to
3943  * unlinking a directory.
3944  */
3945 void
3946 softdep_setup_rmdir(dp, ip)
3947 	struct inode *dp;
3948 	struct inode *ip;
3949 {
3950 	struct vnode *dvp;
3951 
3952 	dvp = ITOV(dp);
3953 	ACQUIRE_LOCK(&lk);
3954 	(void) inodedep_lookup_ip(ip);
3955 	(void) inodedep_lookup_ip(dp);
3956 	softdep_prelink(dvp, ITOV(ip));
3957 	FREE_LOCK(&lk);
3958 }
3959 
3960 /*
3961  * Called to track nlinkdelta of the inode and parent directories prior to
3962  * unlink.
3963  */
3964 void
3965 softdep_setup_unlink(dp, ip)
3966 	struct inode *dp;
3967 	struct inode *ip;
3968 {
3969 	struct vnode *dvp;
3970 
3971 	dvp = ITOV(dp);
3972 	ACQUIRE_LOCK(&lk);
3973 	(void) inodedep_lookup_ip(ip);
3974 	(void) inodedep_lookup_ip(dp);
3975 	softdep_prelink(dvp, ITOV(ip));
3976 	FREE_LOCK(&lk);
3977 }
3978 
3979 /*
3980  * Called to release the journal structures created by a failed non-directory
3981  * creation.  Adjusts nlinkdelta for non-journaling softdep.
3982  */
3983 void
3984 softdep_revert_create(dp, ip)
3985 	struct inode *dp;
3986 	struct inode *ip;
3987 {
3988 	struct inodedep *inodedep;
3989 	struct jaddref *jaddref;
3990 	struct vnode *dvp;
3991 
3992 	dvp = ITOV(dp);
3993 	ACQUIRE_LOCK(&lk);
3994 	inodedep = inodedep_lookup_ip(ip);
3995 	if (DOINGSUJ(dvp)) {
3996 		jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
3997 		    inoreflst);
3998 		KASSERT(jaddref->ja_parent == dp->i_number,
3999 		    ("softdep_revert_create: addref parent mismatch"));
4000 		cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4001 	}
4002 	FREE_LOCK(&lk);
4003 }
4004 
4005 /*
4006  * Called to release the journal structures created by a failed dotdot link
4007  * creation.  Adjusts nlinkdelta for non-journaling softdep.
4008  */
4009 void
4010 softdep_revert_dotdot_link(dp, ip)
4011 	struct inode *dp;
4012 	struct inode *ip;
4013 {
4014 	struct inodedep *inodedep;
4015 	struct jaddref *jaddref;
4016 	struct vnode *dvp;
4017 
4018 	dvp = ITOV(dp);
4019 	ACQUIRE_LOCK(&lk);
4020 	inodedep = inodedep_lookup_ip(dp);
4021 	if (DOINGSUJ(dvp)) {
4022 		jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4023 		    inoreflst);
4024 		KASSERT(jaddref->ja_parent == ip->i_number,
4025 		    ("softdep_revert_dotdot_link: addref parent mismatch"));
4026 		cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4027 	}
4028 	FREE_LOCK(&lk);
4029 }
4030 
4031 /*
4032  * Called to release the journal structures created by a failed link
4033  * addition.  Adjusts nlinkdelta for non-journaling softdep.
4034  */
4035 void
4036 softdep_revert_link(dp, ip)
4037 	struct inode *dp;
4038 	struct inode *ip;
4039 {
4040 	struct inodedep *inodedep;
4041 	struct jaddref *jaddref;
4042 	struct vnode *dvp;
4043 
4044 	dvp = ITOV(dp);
4045 	ACQUIRE_LOCK(&lk);
4046 	inodedep = inodedep_lookup_ip(ip);
4047 	if (DOINGSUJ(dvp)) {
4048 		jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4049 		    inoreflst);
4050 		KASSERT(jaddref->ja_parent == dp->i_number,
4051 		    ("softdep_revert_link: addref parent mismatch"));
4052 		cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4053 	}
4054 	FREE_LOCK(&lk);
4055 }
4056 
4057 /*
4058  * Called to release the journal structures created by a failed mkdir
4059  * attempt.  Adjusts nlinkdelta for non-journaling softdep.
4060  */
4061 void
4062 softdep_revert_mkdir(dp, ip)
4063 	struct inode *dp;
4064 	struct inode *ip;
4065 {
4066 	struct inodedep *inodedep;
4067 	struct jaddref *jaddref;
4068 	struct vnode *dvp;
4069 
4070 	dvp = ITOV(dp);
4071 
4072 	ACQUIRE_LOCK(&lk);
4073 	inodedep = inodedep_lookup_ip(dp);
4074 	if (DOINGSUJ(dvp)) {
4075 		jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4076 		    inoreflst);
4077 		KASSERT(jaddref->ja_parent == ip->i_number,
4078 		    ("softdep_revert_mkdir: dotdot addref parent mismatch"));
4079 		cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4080 	}
4081 	inodedep = inodedep_lookup_ip(ip);
4082 	if (DOINGSUJ(dvp)) {
4083 		jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4084 		    inoreflst);
4085 		KASSERT(jaddref->ja_parent == dp->i_number,
4086 		    ("softdep_revert_mkdir: addref parent mismatch"));
4087 		cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4088 		jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
4089 		    inoreflst);
4090 		KASSERT(jaddref->ja_parent == ip->i_number,
4091 		    ("softdep_revert_mkdir: dot addref parent mismatch"));
4092 		cancel_jaddref(jaddref, inodedep, &inodedep->id_inowait);
4093 	}
4094 	FREE_LOCK(&lk);
4095 }
4096 
4097 /*
4098  * Called to correct nlinkdelta after a failed rmdir.
4099  */
4100 void
4101 softdep_revert_rmdir(dp, ip)
4102 	struct inode *dp;
4103 	struct inode *ip;
4104 {
4105 
4106 	ACQUIRE_LOCK(&lk);
4107 	(void) inodedep_lookup_ip(ip);
4108 	(void) inodedep_lookup_ip(dp);
4109 	FREE_LOCK(&lk);
4110 }
4111 
4112 /*
4113  * Protecting the freemaps (or bitmaps).
4114  *
4115  * To eliminate the need to execute fsck before mounting a filesystem
4116  * after a power failure, one must (conservatively) guarantee that the
4117  * on-disk copy of the bitmaps never indicate that a live inode or block is
4118  * free.  So, when a block or inode is allocated, the bitmap should be
4119  * updated (on disk) before any new pointers.  When a block or inode is
4120  * freed, the bitmap should not be updated until all pointers have been
4121  * reset.  The latter dependency is handled by the delayed de-allocation
4122  * approach described below for block and inode de-allocation.  The former
4123  * dependency is handled by calling the following procedure when a block or
4124  * inode is allocated. When an inode is allocated an "inodedep" is created
4125  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
4126  * Each "inodedep" is also inserted into the hash indexing structure so
4127  * that any additional link additions can be made dependent on the inode
4128  * allocation.
4129  *
4130  * The ufs filesystem maintains a number of free block counts (e.g., per
4131  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
4132  * in addition to the bitmaps.  These counts are used to improve efficiency
4133  * during allocation and therefore must be consistent with the bitmaps.
4134  * There is no convenient way to guarantee post-crash consistency of these
4135  * counts with simple update ordering, for two main reasons: (1) The counts
4136  * and bitmaps for a single cylinder group block are not in the same disk
4137  * sector.  If a disk write is interrupted (e.g., by power failure), one may
4138  * be written and the other not.  (2) Some of the counts are located in the
4139  * superblock rather than the cylinder group block. So, we focus our soft
4140  * updates implementation on protecting the bitmaps. When mounting a
4141  * filesystem, we recompute the auxiliary counts from the bitmaps.
4142  */
4143 
4144 /*
4145  * Called just after updating the cylinder group block to allocate an inode.
4146  */
4147 void
4148 softdep_setup_inomapdep(bp, ip, newinum)
4149 	struct buf *bp;		/* buffer for cylgroup block with inode map */
4150 	struct inode *ip;	/* inode related to allocation */
4151 	ino_t newinum;		/* new inode number being allocated */
4152 {
4153 	struct inodedep *inodedep;
4154 	struct bmsafemap *bmsafemap;
4155 	struct jaddref *jaddref;
4156 	struct mount *mp;
4157 	struct fs *fs;
4158 
4159 	mp = UFSTOVFS(ip->i_ump);
4160 	fs = ip->i_ump->um_fs;
4161 	jaddref = NULL;
4162 
4163 	/*
4164 	 * Allocate the journal reference add structure so that the bitmap
4165 	 * can be dependent on it.
4166 	 */
4167 	if (mp->mnt_kern_flag & MNTK_SUJ) {
4168 		jaddref = newjaddref(ip, newinum, 0, 0, 0);
4169 		jaddref->ja_state |= NEWBLOCK;
4170 	}
4171 
4172 	/*
4173 	 * Create a dependency for the newly allocated inode.
4174 	 * Panic if it already exists as something is seriously wrong.
4175 	 * Otherwise add it to the dependency list for the buffer holding
4176 	 * the cylinder group map from which it was allocated.
4177 	 */
4178 	ACQUIRE_LOCK(&lk);
4179 	if ((inodedep_lookup(mp, newinum, DEPALLOC|NODELAY, &inodedep)))
4180 		panic("softdep_setup_inomapdep: dependency %p for new"
4181 		    "inode already exists", inodedep);
4182 	bmsafemap = bmsafemap_lookup(mp, bp, ino_to_cg(fs, newinum));
4183 	if (jaddref) {
4184 		LIST_INSERT_HEAD(&bmsafemap->sm_jaddrefhd, jaddref, ja_bmdeps);
4185 		TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jaddref->ja_ref,
4186 		    if_deps);
4187 	} else {
4188 		inodedep->id_state |= ONDEPLIST;
4189 		LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
4190 	}
4191 	inodedep->id_bmsafemap = bmsafemap;
4192 	inodedep->id_state &= ~DEPCOMPLETE;
4193 	FREE_LOCK(&lk);
4194 }
4195 
4196 /*
4197  * Called just after updating the cylinder group block to
4198  * allocate block or fragment.
4199  */
4200 void
4201 softdep_setup_blkmapdep(bp, mp, newblkno, frags, oldfrags)
4202 	struct buf *bp;		/* buffer for cylgroup block with block map */
4203 	struct mount *mp;	/* filesystem doing allocation */
4204 	ufs2_daddr_t newblkno;	/* number of newly allocated block */
4205 	int frags;		/* Number of fragments. */
4206 	int oldfrags;		/* Previous number of fragments for extend. */
4207 {
4208 	struct newblk *newblk;
4209 	struct bmsafemap *bmsafemap;
4210 	struct jnewblk *jnewblk;
4211 	struct fs *fs;
4212 
4213 	fs = VFSTOUFS(mp)->um_fs;
4214 	jnewblk = NULL;
4215 	/*
4216 	 * Create a dependency for the newly allocated block.
4217 	 * Add it to the dependency list for the buffer holding
4218 	 * the cylinder group map from which it was allocated.
4219 	 */
4220 	if (mp->mnt_kern_flag & MNTK_SUJ) {
4221 		jnewblk = malloc(sizeof(*jnewblk), M_JNEWBLK, M_SOFTDEP_FLAGS);
4222 		workitem_alloc(&jnewblk->jn_list, D_JNEWBLK, mp);
4223 		jnewblk->jn_jsegdep = newjsegdep(&jnewblk->jn_list);
4224 		jnewblk->jn_state = ATTACHED;
4225 		jnewblk->jn_blkno = newblkno;
4226 		jnewblk->jn_frags = frags;
4227 		jnewblk->jn_oldfrags = oldfrags;
4228 #ifdef SUJ_DEBUG
4229 		{
4230 			struct cg *cgp;
4231 			uint8_t *blksfree;
4232 			long bno;
4233 			int i;
4234 
4235 			cgp = (struct cg *)bp->b_data;
4236 			blksfree = cg_blksfree(cgp);
4237 			bno = dtogd(fs, jnewblk->jn_blkno);
4238 			for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags;
4239 			    i++) {
4240 				if (isset(blksfree, bno + i))
4241 					panic("softdep_setup_blkmapdep: "
4242 					    "free fragment %d from %d-%d "
4243 					    "state 0x%X dep %p", i,
4244 					    jnewblk->jn_oldfrags,
4245 					    jnewblk->jn_frags,
4246 					    jnewblk->jn_state,
4247 					    jnewblk->jn_newblk);
4248 			}
4249 		}
4250 #endif
4251 	}
4252 	ACQUIRE_LOCK(&lk);
4253 	if (newblk_lookup(mp, newblkno, DEPALLOC, &newblk) != 0)
4254 		panic("softdep_setup_blkmapdep: found block");
4255 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(mp, bp,
4256 	    dtog(fs, newblkno));
4257 	if (jnewblk) {
4258 		jnewblk->jn_newblk = newblk;
4259 		LIST_INSERT_HEAD(&bmsafemap->sm_jnewblkhd, jnewblk, jn_deps);
4260 	} else {
4261 		newblk->nb_state |= ONDEPLIST;
4262 		LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
4263 	}
4264 	newblk->nb_bmsafemap = bmsafemap;
4265 	newblk->nb_jnewblk = jnewblk;
4266 	FREE_LOCK(&lk);
4267 }
4268 
4269 #define	BMSAFEMAP_HASH(fs, cg) \
4270       (&bmsafemap_hashtbl[((((register_t)(fs)) >> 13) + (cg)) & bmsafemap_hash])
4271 
4272 static int
4273 bmsafemap_find(bmsafemaphd, mp, cg, bmsafemapp)
4274 	struct bmsafemap_hashhead *bmsafemaphd;
4275 	struct mount *mp;
4276 	int cg;
4277 	struct bmsafemap **bmsafemapp;
4278 {
4279 	struct bmsafemap *bmsafemap;
4280 
4281 	LIST_FOREACH(bmsafemap, bmsafemaphd, sm_hash)
4282 		if (bmsafemap->sm_list.wk_mp == mp && bmsafemap->sm_cg == cg)
4283 			break;
4284 	if (bmsafemap) {
4285 		*bmsafemapp = bmsafemap;
4286 		return (1);
4287 	}
4288 	*bmsafemapp = NULL;
4289 
4290 	return (0);
4291 }
4292 
4293 /*
4294  * Find the bmsafemap associated with a cylinder group buffer.
4295  * If none exists, create one. The buffer must be locked when
4296  * this routine is called and this routine must be called with
4297  * splbio interrupts blocked.
4298  */
4299 static struct bmsafemap *
4300 bmsafemap_lookup(mp, bp, cg)
4301 	struct mount *mp;
4302 	struct buf *bp;
4303 	int cg;
4304 {
4305 	struct bmsafemap_hashhead *bmsafemaphd;
4306 	struct bmsafemap *bmsafemap, *collision;
4307 	struct worklist *wk;
4308 	struct fs *fs;
4309 
4310 	mtx_assert(&lk, MA_OWNED);
4311 	if (bp)
4312 		LIST_FOREACH(wk, &bp->b_dep, wk_list)
4313 			if (wk->wk_type == D_BMSAFEMAP)
4314 				return (WK_BMSAFEMAP(wk));
4315 	fs = VFSTOUFS(mp)->um_fs;
4316 	bmsafemaphd = BMSAFEMAP_HASH(fs, cg);
4317 	if (bmsafemap_find(bmsafemaphd, mp, cg, &bmsafemap) == 1)
4318 		return (bmsafemap);
4319 	FREE_LOCK(&lk);
4320 	bmsafemap = malloc(sizeof(struct bmsafemap),
4321 		M_BMSAFEMAP, M_SOFTDEP_FLAGS);
4322 	workitem_alloc(&bmsafemap->sm_list, D_BMSAFEMAP, mp);
4323 	bmsafemap->sm_buf = bp;
4324 	LIST_INIT(&bmsafemap->sm_inodedephd);
4325 	LIST_INIT(&bmsafemap->sm_inodedepwr);
4326 	LIST_INIT(&bmsafemap->sm_newblkhd);
4327 	LIST_INIT(&bmsafemap->sm_newblkwr);
4328 	LIST_INIT(&bmsafemap->sm_jaddrefhd);
4329 	LIST_INIT(&bmsafemap->sm_jnewblkhd);
4330 	ACQUIRE_LOCK(&lk);
4331 	if (bmsafemap_find(bmsafemaphd, mp, cg, &collision) == 1) {
4332 		WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
4333 		return (collision);
4334 	}
4335 	bmsafemap->sm_cg = cg;
4336 	LIST_INSERT_HEAD(bmsafemaphd, bmsafemap, sm_hash);
4337 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
4338 	return (bmsafemap);
4339 }
4340 
4341 /*
4342  * Direct block allocation dependencies.
4343  *
4344  * When a new block is allocated, the corresponding disk locations must be
4345  * initialized (with zeros or new data) before the on-disk inode points to
4346  * them.  Also, the freemap from which the block was allocated must be
4347  * updated (on disk) before the inode's pointer. These two dependencies are
4348  * independent of each other and are needed for all file blocks and indirect
4349  * blocks that are pointed to directly by the inode.  Just before the
4350  * "in-core" version of the inode is updated with a newly allocated block
4351  * number, a procedure (below) is called to setup allocation dependency
4352  * structures.  These structures are removed when the corresponding
4353  * dependencies are satisfied or when the block allocation becomes obsolete
4354  * (i.e., the file is deleted, the block is de-allocated, or the block is a
4355  * fragment that gets upgraded).  All of these cases are handled in
4356  * procedures described later.
4357  *
4358  * When a file extension causes a fragment to be upgraded, either to a larger
4359  * fragment or to a full block, the on-disk location may change (if the
4360  * previous fragment could not simply be extended). In this case, the old
4361  * fragment must be de-allocated, but not until after the inode's pointer has
4362  * been updated. In most cases, this is handled by later procedures, which
4363  * will construct a "freefrag" structure to be added to the workitem queue
4364  * when the inode update is complete (or obsolete).  The main exception to
4365  * this is when an allocation occurs while a pending allocation dependency
4366  * (for the same block pointer) remains.  This case is handled in the main
4367  * allocation dependency setup procedure by immediately freeing the
4368  * unreferenced fragments.
4369  */
4370 void
4371 softdep_setup_allocdirect(ip, off, newblkno, oldblkno, newsize, oldsize, bp)
4372 	struct inode *ip;	/* inode to which block is being added */
4373 	ufs_lbn_t off;		/* block pointer within inode */
4374 	ufs2_daddr_t newblkno;	/* disk block number being added */
4375 	ufs2_daddr_t oldblkno;	/* previous block number, 0 unless frag */
4376 	long newsize;		/* size of new block */
4377 	long oldsize;		/* size of new block */
4378 	struct buf *bp;		/* bp for allocated block */
4379 {
4380 	struct allocdirect *adp, *oldadp;
4381 	struct allocdirectlst *adphead;
4382 	struct freefrag *freefrag;
4383 	struct inodedep *inodedep;
4384 	struct pagedep *pagedep;
4385 	struct jnewblk *jnewblk;
4386 	struct newblk *newblk;
4387 	struct mount *mp;
4388 	ufs_lbn_t lbn;
4389 
4390 	lbn = bp->b_lblkno;
4391 	mp = UFSTOVFS(ip->i_ump);
4392 	if (oldblkno && oldblkno != newblkno)
4393 		freefrag = newfreefrag(ip, oldblkno, oldsize, lbn);
4394 	else
4395 		freefrag = NULL;
4396 
4397 	ACQUIRE_LOCK(&lk);
4398 	if (off >= NDADDR) {
4399 		if (lbn > 0)
4400 			panic("softdep_setup_allocdirect: bad lbn %jd, off %jd",
4401 			    lbn, off);
4402 		/* allocating an indirect block */
4403 		if (oldblkno != 0)
4404 			panic("softdep_setup_allocdirect: non-zero indir");
4405 	} else {
4406 		if (off != lbn)
4407 			panic("softdep_setup_allocdirect: lbn %jd != off %jd",
4408 			    lbn, off);
4409 		/*
4410 		 * Allocating a direct block.
4411 		 *
4412 		 * If we are allocating a directory block, then we must
4413 		 * allocate an associated pagedep to track additions and
4414 		 * deletions.
4415 		 */
4416 		if ((ip->i_mode & IFMT) == IFDIR &&
4417 		    pagedep_lookup(mp, ip->i_number, off, DEPALLOC,
4418 		    &pagedep) == 0)
4419 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
4420 	}
4421 	if (newblk_lookup(mp, newblkno, 0, &newblk) == 0)
4422 		panic("softdep_setup_allocdirect: lost block");
4423 	KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
4424 	    ("softdep_setup_allocdirect: newblk already initialized"));
4425 	/*
4426 	 * Convert the newblk to an allocdirect.
4427 	 */
4428 	newblk->nb_list.wk_type = D_ALLOCDIRECT;
4429 	adp = (struct allocdirect *)newblk;
4430 	newblk->nb_freefrag = freefrag;
4431 	adp->ad_offset = off;
4432 	adp->ad_oldblkno = oldblkno;
4433 	adp->ad_newsize = newsize;
4434 	adp->ad_oldsize = oldsize;
4435 
4436 	/*
4437 	 * Finish initializing the journal.
4438 	 */
4439 	if ((jnewblk = newblk->nb_jnewblk) != NULL) {
4440 		jnewblk->jn_ino = ip->i_number;
4441 		jnewblk->jn_lbn = lbn;
4442 		add_to_journal(&jnewblk->jn_list);
4443 	}
4444 	if (freefrag && freefrag->ff_jfreefrag != NULL)
4445 		add_to_journal(&freefrag->ff_jfreefrag->fr_list);
4446 	inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep);
4447 	adp->ad_inodedep = inodedep;
4448 
4449 	WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list);
4450 	/*
4451 	 * The list of allocdirects must be kept in sorted and ascending
4452 	 * order so that the rollback routines can quickly determine the
4453 	 * first uncommitted block (the size of the file stored on disk
4454 	 * ends at the end of the lowest committed fragment, or if there
4455 	 * are no fragments, at the end of the highest committed block).
4456 	 * Since files generally grow, the typical case is that the new
4457 	 * block is to be added at the end of the list. We speed this
4458 	 * special case by checking against the last allocdirect in the
4459 	 * list before laboriously traversing the list looking for the
4460 	 * insertion point.
4461 	 */
4462 	adphead = &inodedep->id_newinoupdt;
4463 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
4464 	if (oldadp == NULL || oldadp->ad_offset <= off) {
4465 		/* insert at end of list */
4466 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
4467 		if (oldadp != NULL && oldadp->ad_offset == off)
4468 			allocdirect_merge(adphead, adp, oldadp);
4469 		FREE_LOCK(&lk);
4470 		return;
4471 	}
4472 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
4473 		if (oldadp->ad_offset >= off)
4474 			break;
4475 	}
4476 	if (oldadp == NULL)
4477 		panic("softdep_setup_allocdirect: lost entry");
4478 	/* insert in middle of list */
4479 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
4480 	if (oldadp->ad_offset == off)
4481 		allocdirect_merge(adphead, adp, oldadp);
4482 
4483 	FREE_LOCK(&lk);
4484 }
4485 
4486 /*
4487  * Replace an old allocdirect dependency with a newer one.
4488  * This routine must be called with splbio interrupts blocked.
4489  */
4490 static void
4491 allocdirect_merge(adphead, newadp, oldadp)
4492 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
4493 	struct allocdirect *newadp;	/* allocdirect being added */
4494 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
4495 {
4496 	struct worklist *wk;
4497 	struct freefrag *freefrag;
4498 	struct newdirblk *newdirblk;
4499 
4500 	freefrag = NULL;
4501 	mtx_assert(&lk, MA_OWNED);
4502 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
4503 	    newadp->ad_oldsize != oldadp->ad_newsize ||
4504 	    newadp->ad_offset >= NDADDR)
4505 		panic("%s %jd != new %jd || old size %ld != new %ld",
4506 		    "allocdirect_merge: old blkno",
4507 		    (intmax_t)newadp->ad_oldblkno,
4508 		    (intmax_t)oldadp->ad_newblkno,
4509 		    newadp->ad_oldsize, oldadp->ad_newsize);
4510 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
4511 	newadp->ad_oldsize = oldadp->ad_oldsize;
4512 	/*
4513 	 * If the old dependency had a fragment to free or had never
4514 	 * previously had a block allocated, then the new dependency
4515 	 * can immediately post its freefrag and adopt the old freefrag.
4516 	 * This action is done by swapping the freefrag dependencies.
4517 	 * The new dependency gains the old one's freefrag, and the
4518 	 * old one gets the new one and then immediately puts it on
4519 	 * the worklist when it is freed by free_newblk. It is
4520 	 * not possible to do this swap when the old dependency had a
4521 	 * non-zero size but no previous fragment to free. This condition
4522 	 * arises when the new block is an extension of the old block.
4523 	 * Here, the first part of the fragment allocated to the new
4524 	 * dependency is part of the block currently claimed on disk by
4525 	 * the old dependency, so cannot legitimately be freed until the
4526 	 * conditions for the new dependency are fulfilled.
4527 	 */
4528 	freefrag = newadp->ad_freefrag;
4529 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
4530 		newadp->ad_freefrag = oldadp->ad_freefrag;
4531 		oldadp->ad_freefrag = freefrag;
4532 	}
4533 	/*
4534 	 * If we are tracking a new directory-block allocation,
4535 	 * move it from the old allocdirect to the new allocdirect.
4536 	 */
4537 	if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
4538 		newdirblk = WK_NEWDIRBLK(wk);
4539 		WORKLIST_REMOVE(&newdirblk->db_list);
4540 		if (!LIST_EMPTY(&oldadp->ad_newdirblk))
4541 			panic("allocdirect_merge: extra newdirblk");
4542 		WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list);
4543 	}
4544 	TAILQ_REMOVE(adphead, oldadp, ad_next);
4545 	/*
4546 	 * We need to move any journal dependencies over to the freefrag
4547 	 * that releases this block if it exists.  Otherwise we are
4548 	 * extending an existing block and we'll wait until that is
4549 	 * complete to release the journal space and extend the
4550 	 * new journal to cover this old space as well.
4551 	 */
4552 	if (freefrag == NULL) {
4553 		struct jnewblk *jnewblk;
4554 		struct jnewblk *njnewblk;
4555 
4556 		if (oldadp->ad_newblkno != newadp->ad_newblkno)
4557 			panic("allocdirect_merge: %jd != %jd",
4558 			    oldadp->ad_newblkno, newadp->ad_newblkno);
4559 		jnewblk = oldadp->ad_block.nb_jnewblk;
4560 		cancel_newblk(&oldadp->ad_block, &newadp->ad_block.nb_jwork);
4561 		/*
4562 		 * We have an unwritten jnewblk, we need to merge the
4563 		 * frag bits with our own.  The newer adp's journal can not
4564 		 * be written prior to the old one so no need to check for
4565 		 * it here.
4566 		 */
4567 		if (jnewblk) {
4568 			njnewblk = newadp->ad_block.nb_jnewblk;
4569 			if (njnewblk == NULL)
4570 				panic("allocdirect_merge: No jnewblk");
4571 			if (jnewblk->jn_state & UNDONE) {
4572 				njnewblk->jn_state |= UNDONE | NEWBLOCK;
4573 				njnewblk->jn_state &= ~ATTACHED;
4574 				jnewblk->jn_state &= ~UNDONE;
4575 			}
4576 			njnewblk->jn_oldfrags = jnewblk->jn_oldfrags;
4577 			WORKLIST_REMOVE(&jnewblk->jn_list);
4578 			jnewblk->jn_state |= ATTACHED | COMPLETE;
4579 			free_jnewblk(jnewblk);
4580 		}
4581 	} else {
4582 		/*
4583 		 * We can skip journaling for this freefrag and just complete
4584 		 * any pending journal work for the allocdirect that is being
4585 		 * removed after the freefrag completes.
4586 		 */
4587 		if (freefrag->ff_jfreefrag)
4588 			cancel_jfreefrag(freefrag->ff_jfreefrag);
4589 		cancel_newblk(&oldadp->ad_block, &freefrag->ff_jwork);
4590 	}
4591 	free_newblk(&oldadp->ad_block);
4592 }
4593 
4594 /*
4595  * Allocate a jfreefrag structure to journal a single block free.
4596  */
4597 static struct jfreefrag *
4598 newjfreefrag(freefrag, ip, blkno, size, lbn)
4599 	struct freefrag *freefrag;
4600 	struct inode *ip;
4601 	ufs2_daddr_t blkno;
4602 	long size;
4603 	ufs_lbn_t lbn;
4604 {
4605 	struct jfreefrag *jfreefrag;
4606 	struct fs *fs;
4607 
4608 	fs = ip->i_fs;
4609 	jfreefrag = malloc(sizeof(struct jfreefrag), M_JFREEFRAG,
4610 	    M_SOFTDEP_FLAGS);
4611 	workitem_alloc(&jfreefrag->fr_list, D_JFREEFRAG, UFSTOVFS(ip->i_ump));
4612 	jfreefrag->fr_jsegdep = newjsegdep(&jfreefrag->fr_list);
4613 	jfreefrag->fr_state = ATTACHED | DEPCOMPLETE;
4614 	jfreefrag->fr_ino = ip->i_number;
4615 	jfreefrag->fr_lbn = lbn;
4616 	jfreefrag->fr_blkno = blkno;
4617 	jfreefrag->fr_frags = numfrags(fs, size);
4618 	jfreefrag->fr_freefrag = freefrag;
4619 
4620 	return (jfreefrag);
4621 }
4622 
4623 /*
4624  * Allocate a new freefrag structure.
4625  */
4626 static struct freefrag *
4627 newfreefrag(ip, blkno, size, lbn)
4628 	struct inode *ip;
4629 	ufs2_daddr_t blkno;
4630 	long size;
4631 	ufs_lbn_t lbn;
4632 {
4633 	struct freefrag *freefrag;
4634 	struct fs *fs;
4635 
4636 	fs = ip->i_fs;
4637 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
4638 		panic("newfreefrag: frag size");
4639 	freefrag = malloc(sizeof(struct freefrag),
4640 	    M_FREEFRAG, M_SOFTDEP_FLAGS);
4641 	workitem_alloc(&freefrag->ff_list, D_FREEFRAG, UFSTOVFS(ip->i_ump));
4642 	freefrag->ff_state = ATTACHED;
4643 	LIST_INIT(&freefrag->ff_jwork);
4644 	freefrag->ff_inum = ip->i_number;
4645 	freefrag->ff_blkno = blkno;
4646 	freefrag->ff_fragsize = size;
4647 
4648 	if (fs->fs_flags & FS_SUJ) {
4649 		freefrag->ff_jfreefrag =
4650 		    newjfreefrag(freefrag, ip, blkno, size, lbn);
4651 	} else {
4652 		freefrag->ff_state |= DEPCOMPLETE;
4653 		freefrag->ff_jfreefrag = NULL;
4654 	}
4655 
4656 	return (freefrag);
4657 }
4658 
4659 /*
4660  * This workitem de-allocates fragments that were replaced during
4661  * file block allocation.
4662  */
4663 static void
4664 handle_workitem_freefrag(freefrag)
4665 	struct freefrag *freefrag;
4666 {
4667 	struct ufsmount *ump = VFSTOUFS(freefrag->ff_list.wk_mp);
4668 	struct workhead wkhd;
4669 
4670 	/*
4671 	 * It would be illegal to add new completion items to the
4672 	 * freefrag after it was schedule to be done so it must be
4673 	 * safe to modify the list head here.
4674 	 */
4675 	LIST_INIT(&wkhd);
4676 	LIST_SWAP(&freefrag->ff_jwork, &wkhd, worklist, wk_list);
4677 	ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
4678 	    freefrag->ff_fragsize, freefrag->ff_inum, &wkhd);
4679 	ACQUIRE_LOCK(&lk);
4680 	WORKITEM_FREE(freefrag, D_FREEFRAG);
4681 	FREE_LOCK(&lk);
4682 }
4683 
4684 /*
4685  * Set up a dependency structure for an external attributes data block.
4686  * This routine follows much of the structure of softdep_setup_allocdirect.
4687  * See the description of softdep_setup_allocdirect above for details.
4688  */
4689 void
4690 softdep_setup_allocext(ip, off, newblkno, oldblkno, newsize, oldsize, bp)
4691 	struct inode *ip;
4692 	ufs_lbn_t off;
4693 	ufs2_daddr_t newblkno;
4694 	ufs2_daddr_t oldblkno;
4695 	long newsize;
4696 	long oldsize;
4697 	struct buf *bp;
4698 {
4699 	struct allocdirect *adp, *oldadp;
4700 	struct allocdirectlst *adphead;
4701 	struct freefrag *freefrag;
4702 	struct inodedep *inodedep;
4703 	struct jnewblk *jnewblk;
4704 	struct newblk *newblk;
4705 	struct mount *mp;
4706 	ufs_lbn_t lbn;
4707 
4708 	if (off >= NXADDR)
4709 		panic("softdep_setup_allocext: lbn %lld > NXADDR",
4710 		    (long long)off);
4711 
4712 	lbn = bp->b_lblkno;
4713 	mp = UFSTOVFS(ip->i_ump);
4714 	if (oldblkno && oldblkno != newblkno)
4715 		freefrag = newfreefrag(ip, oldblkno, oldsize, lbn);
4716 	else
4717 		freefrag = NULL;
4718 
4719 	ACQUIRE_LOCK(&lk);
4720 	if (newblk_lookup(mp, newblkno, 0, &newblk) == 0)
4721 		panic("softdep_setup_allocext: lost block");
4722 	KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
4723 	    ("softdep_setup_allocext: newblk already initialized"));
4724 	/*
4725 	 * Convert the newblk to an allocdirect.
4726 	 */
4727 	newblk->nb_list.wk_type = D_ALLOCDIRECT;
4728 	adp = (struct allocdirect *)newblk;
4729 	newblk->nb_freefrag = freefrag;
4730 	adp->ad_offset = off;
4731 	adp->ad_oldblkno = oldblkno;
4732 	adp->ad_newsize = newsize;
4733 	adp->ad_oldsize = oldsize;
4734 	adp->ad_state |=  EXTDATA;
4735 
4736 	/*
4737 	 * Finish initializing the journal.
4738 	 */
4739 	if ((jnewblk = newblk->nb_jnewblk) != NULL) {
4740 		jnewblk->jn_ino = ip->i_number;
4741 		jnewblk->jn_lbn = lbn;
4742 		add_to_journal(&jnewblk->jn_list);
4743 	}
4744 	if (freefrag && freefrag->ff_jfreefrag != NULL)
4745 		add_to_journal(&freefrag->ff_jfreefrag->fr_list);
4746 	inodedep_lookup(mp, ip->i_number, DEPALLOC | NODELAY, &inodedep);
4747 	adp->ad_inodedep = inodedep;
4748 
4749 	WORKLIST_INSERT(&bp->b_dep, &newblk->nb_list);
4750 	/*
4751 	 * The list of allocdirects must be kept in sorted and ascending
4752 	 * order so that the rollback routines can quickly determine the
4753 	 * first uncommitted block (the size of the file stored on disk
4754 	 * ends at the end of the lowest committed fragment, or if there
4755 	 * are no fragments, at the end of the highest committed block).
4756 	 * Since files generally grow, the typical case is that the new
4757 	 * block is to be added at the end of the list. We speed this
4758 	 * special case by checking against the last allocdirect in the
4759 	 * list before laboriously traversing the list looking for the
4760 	 * insertion point.
4761 	 */
4762 	adphead = &inodedep->id_newextupdt;
4763 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
4764 	if (oldadp == NULL || oldadp->ad_offset <= off) {
4765 		/* insert at end of list */
4766 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
4767 		if (oldadp != NULL && oldadp->ad_offset == off)
4768 			allocdirect_merge(adphead, adp, oldadp);
4769 		FREE_LOCK(&lk);
4770 		return;
4771 	}
4772 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
4773 		if (oldadp->ad_offset >= off)
4774 			break;
4775 	}
4776 	if (oldadp == NULL)
4777 		panic("softdep_setup_allocext: lost entry");
4778 	/* insert in middle of list */
4779 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
4780 	if (oldadp->ad_offset == off)
4781 		allocdirect_merge(adphead, adp, oldadp);
4782 	FREE_LOCK(&lk);
4783 }
4784 
4785 /*
4786  * Indirect block allocation dependencies.
4787  *
4788  * The same dependencies that exist for a direct block also exist when
4789  * a new block is allocated and pointed to by an entry in a block of
4790  * indirect pointers. The undo/redo states described above are also
4791  * used here. Because an indirect block contains many pointers that
4792  * may have dependencies, a second copy of the entire in-memory indirect
4793  * block is kept. The buffer cache copy is always completely up-to-date.
4794  * The second copy, which is used only as a source for disk writes,
4795  * contains only the safe pointers (i.e., those that have no remaining
4796  * update dependencies). The second copy is freed when all pointers
4797  * are safe. The cache is not allowed to replace indirect blocks with
4798  * pending update dependencies. If a buffer containing an indirect
4799  * block with dependencies is written, these routines will mark it
4800  * dirty again. It can only be successfully written once all the
4801  * dependencies are removed. The ffs_fsync routine in conjunction with
4802  * softdep_sync_metadata work together to get all the dependencies
4803  * removed so that a file can be successfully written to disk. Three
4804  * procedures are used when setting up indirect block pointer
4805  * dependencies. The division is necessary because of the organization
4806  * of the "balloc" routine and because of the distinction between file
4807  * pages and file metadata blocks.
4808  */
4809 
4810 /*
4811  * Allocate a new allocindir structure.
4812  */
4813 static struct allocindir *
4814 newallocindir(ip, ptrno, newblkno, oldblkno, lbn)
4815 	struct inode *ip;	/* inode for file being extended */
4816 	int ptrno;		/* offset of pointer in indirect block */
4817 	ufs2_daddr_t newblkno;	/* disk block number being added */
4818 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
4819 	ufs_lbn_t lbn;
4820 {
4821 	struct newblk *newblk;
4822 	struct allocindir *aip;
4823 	struct freefrag *freefrag;
4824 	struct jnewblk *jnewblk;
4825 
4826 	if (oldblkno)
4827 		freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize, lbn);
4828 	else
4829 		freefrag = NULL;
4830 	ACQUIRE_LOCK(&lk);
4831 	if (newblk_lookup(UFSTOVFS(ip->i_ump), newblkno, 0, &newblk) == 0)
4832 		panic("new_allocindir: lost block");
4833 	KASSERT(newblk->nb_list.wk_type == D_NEWBLK,
4834 	    ("newallocindir: newblk already initialized"));
4835 	newblk->nb_list.wk_type = D_ALLOCINDIR;
4836 	newblk->nb_freefrag = freefrag;
4837 	aip = (struct allocindir *)newblk;
4838 	aip->ai_offset = ptrno;
4839 	aip->ai_oldblkno = oldblkno;
4840 	if ((jnewblk = newblk->nb_jnewblk) != NULL) {
4841 		jnewblk->jn_ino = ip->i_number;
4842 		jnewblk->jn_lbn = lbn;
4843 		add_to_journal(&jnewblk->jn_list);
4844 	}
4845 	if (freefrag && freefrag->ff_jfreefrag != NULL)
4846 		add_to_journal(&freefrag->ff_jfreefrag->fr_list);
4847 	return (aip);
4848 }
4849 
4850 /*
4851  * Called just before setting an indirect block pointer
4852  * to a newly allocated file page.
4853  */
4854 void
4855 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
4856 	struct inode *ip;	/* inode for file being extended */
4857 	ufs_lbn_t lbn;		/* allocated block number within file */
4858 	struct buf *bp;		/* buffer with indirect blk referencing page */
4859 	int ptrno;		/* offset of pointer in indirect block */
4860 	ufs2_daddr_t newblkno;	/* disk block number being added */
4861 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
4862 	struct buf *nbp;	/* buffer holding allocated page */
4863 {
4864 	struct inodedep *inodedep;
4865 	struct allocindir *aip;
4866 	struct pagedep *pagedep;
4867 	struct mount *mp;
4868 
4869 	if (lbn != nbp->b_lblkno)
4870 		panic("softdep_setup_allocindir_page: lbn %jd != lblkno %jd",
4871 		    lbn, bp->b_lblkno);
4872 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page");
4873 	mp = UFSTOVFS(ip->i_ump);
4874 	aip = newallocindir(ip, ptrno, newblkno, oldblkno, lbn);
4875 	(void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
4876 	/*
4877 	 * If we are allocating a directory page, then we must
4878 	 * allocate an associated pagedep to track additions and
4879 	 * deletions.
4880 	 */
4881 	if ((ip->i_mode & IFMT) == IFDIR &&
4882 	    pagedep_lookup(mp, ip->i_number, lbn, DEPALLOC, &pagedep) == 0)
4883 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
4884 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
4885 	setup_allocindir_phase2(bp, ip, inodedep, aip, lbn);
4886 	FREE_LOCK(&lk);
4887 }
4888 
4889 /*
4890  * Called just before setting an indirect block pointer to a
4891  * newly allocated indirect block.
4892  */
4893 void
4894 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
4895 	struct buf *nbp;	/* newly allocated indirect block */
4896 	struct inode *ip;	/* inode for file being extended */
4897 	struct buf *bp;		/* indirect block referencing allocated block */
4898 	int ptrno;		/* offset of pointer in indirect block */
4899 	ufs2_daddr_t newblkno;	/* disk block number being added */
4900 {
4901 	struct inodedep *inodedep;
4902 	struct allocindir *aip;
4903 	ufs_lbn_t lbn;
4904 
4905 	lbn = nbp->b_lblkno;
4906 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta");
4907 	aip = newallocindir(ip, ptrno, newblkno, 0, lbn);
4908 	inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC, &inodedep);
4909 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_block.nb_list);
4910 	setup_allocindir_phase2(bp, ip, inodedep, aip, lbn);
4911 	FREE_LOCK(&lk);
4912 }
4913 
4914 static void
4915 indirdep_complete(indirdep)
4916 	struct indirdep *indirdep;
4917 {
4918 	struct allocindir *aip;
4919 
4920 	LIST_REMOVE(indirdep, ir_next);
4921 	indirdep->ir_state &= ~ONDEPLIST;
4922 
4923 	while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) {
4924 		LIST_REMOVE(aip, ai_next);
4925 		free_newblk(&aip->ai_block);
4926 	}
4927 	/*
4928 	 * If this indirdep is not attached to a buf it was simply waiting
4929 	 * on completion to clear completehd.  free_indirdep() asserts
4930 	 * that nothing is dangling.
4931 	 */
4932 	if ((indirdep->ir_state & ONWORKLIST) == 0)
4933 		free_indirdep(indirdep);
4934 }
4935 
4936 /*
4937  * Called to finish the allocation of the "aip" allocated
4938  * by one of the two routines above.
4939  */
4940 static void
4941 setup_allocindir_phase2(bp, ip, inodedep, aip, lbn)
4942 	struct buf *bp;		/* in-memory copy of the indirect block */
4943 	struct inode *ip;	/* inode for file being extended */
4944 	struct inodedep *inodedep; /* Inodedep for ip */
4945 	struct allocindir *aip;	/* allocindir allocated by the above routines */
4946 	ufs_lbn_t lbn;		/* Logical block number for this block. */
4947 {
4948 	struct worklist *wk;
4949 	struct fs *fs;
4950 	struct newblk *newblk;
4951 	struct indirdep *indirdep, *newindirdep;
4952 	struct allocindir *oldaip;
4953 	struct freefrag *freefrag;
4954 	struct mount *mp;
4955 	ufs2_daddr_t blkno;
4956 
4957 	mp = UFSTOVFS(ip->i_ump);
4958 	fs = ip->i_fs;
4959 	mtx_assert(&lk, MA_OWNED);
4960 	if (bp->b_lblkno >= 0)
4961 		panic("setup_allocindir_phase2: not indir blk");
4962 	for (freefrag = NULL, indirdep = NULL, newindirdep = NULL; ; ) {
4963 		LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4964 			if (wk->wk_type != D_INDIRDEP)
4965 				continue;
4966 			indirdep = WK_INDIRDEP(wk);
4967 			break;
4968 		}
4969 		if (indirdep == NULL && newindirdep) {
4970 			indirdep = newindirdep;
4971 			newindirdep = NULL;
4972 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
4973 			if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0,
4974 			    &newblk)) {
4975 				indirdep->ir_state |= ONDEPLIST;
4976 				LIST_INSERT_HEAD(&newblk->nb_indirdeps,
4977 				    indirdep, ir_next);
4978 			} else
4979 				indirdep->ir_state |= DEPCOMPLETE;
4980 		}
4981 		if (indirdep) {
4982 			aip->ai_indirdep = indirdep;
4983 			/*
4984 			 * Check to see if there is an existing dependency
4985 			 * for this block. If there is, merge the old
4986 			 * dependency into the new one.  This happens
4987 			 * as a result of reallocblk only.
4988 			 */
4989 			if (aip->ai_oldblkno == 0)
4990 				oldaip = NULL;
4991 			else
4992 
4993 				LIST_FOREACH(oldaip, &indirdep->ir_deplisthd,
4994 				    ai_next)
4995 					if (oldaip->ai_offset == aip->ai_offset)
4996 						break;
4997 			if (oldaip != NULL)
4998 				freefrag = allocindir_merge(aip, oldaip);
4999 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
5000 			KASSERT(aip->ai_offset >= 0 &&
5001 			    aip->ai_offset < NINDIR(ip->i_ump->um_fs),
5002 			    ("setup_allocindir_phase2: Bad offset %d",
5003 			    aip->ai_offset));
5004 			KASSERT(indirdep->ir_savebp != NULL,
5005 			    ("setup_allocindir_phase2 NULL ir_savebp"));
5006 			if (ip->i_ump->um_fstype == UFS1)
5007 				((ufs1_daddr_t *)indirdep->ir_savebp->b_data)
5008 				    [aip->ai_offset] = aip->ai_oldblkno;
5009 			else
5010 				((ufs2_daddr_t *)indirdep->ir_savebp->b_data)
5011 				    [aip->ai_offset] = aip->ai_oldblkno;
5012 			FREE_LOCK(&lk);
5013 			if (freefrag != NULL)
5014 				handle_workitem_freefrag(freefrag);
5015 		} else
5016 			FREE_LOCK(&lk);
5017 		if (newindirdep) {
5018 			newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
5019 			brelse(newindirdep->ir_savebp);
5020 			ACQUIRE_LOCK(&lk);
5021 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
5022 			if (indirdep)
5023 				break;
5024 			FREE_LOCK(&lk);
5025 		}
5026 		if (indirdep) {
5027 			ACQUIRE_LOCK(&lk);
5028 			break;
5029 		}
5030 		newindirdep = malloc(sizeof(struct indirdep),
5031 			M_INDIRDEP, M_SOFTDEP_FLAGS);
5032 		workitem_alloc(&newindirdep->ir_list, D_INDIRDEP, mp);
5033 		newindirdep->ir_state = ATTACHED;
5034 		if (ip->i_ump->um_fstype == UFS1)
5035 			newindirdep->ir_state |= UFS1FMT;
5036 		newindirdep->ir_saveddata = NULL;
5037 		LIST_INIT(&newindirdep->ir_deplisthd);
5038 		LIST_INIT(&newindirdep->ir_donehd);
5039 		LIST_INIT(&newindirdep->ir_writehd);
5040 		LIST_INIT(&newindirdep->ir_completehd);
5041 		LIST_INIT(&newindirdep->ir_jwork);
5042 		if (bp->b_blkno == bp->b_lblkno) {
5043 			ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
5044 			    NULL, NULL);
5045 			bp->b_blkno = blkno;
5046 		}
5047 		newindirdep->ir_savebp =
5048 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
5049 		BUF_KERNPROC(newindirdep->ir_savebp);
5050 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
5051 		ACQUIRE_LOCK(&lk);
5052 	}
5053 }
5054 
5055 /*
5056  * Merge two allocindirs which refer to the same block.  Move newblock
5057  * dependencies and setup the freefrags appropriately.
5058  */
5059 static struct freefrag *
5060 allocindir_merge(aip, oldaip)
5061 	struct allocindir *aip;
5062 	struct allocindir *oldaip;
5063 {
5064 	struct newdirblk *newdirblk;
5065 	struct freefrag *freefrag;
5066 	struct worklist *wk;
5067 
5068 	if (oldaip->ai_newblkno != aip->ai_oldblkno)
5069 		panic("allocindir_merge: blkno");
5070 	aip->ai_oldblkno = oldaip->ai_oldblkno;
5071 	freefrag = aip->ai_freefrag;
5072 	aip->ai_freefrag = oldaip->ai_freefrag;
5073 	oldaip->ai_freefrag = NULL;
5074 	KASSERT(freefrag != NULL, ("setup_allocindir_phase2: No freefrag"));
5075 	/*
5076 	 * If we are tracking a new directory-block allocation,
5077 	 * move it from the old allocindir to the new allocindir.
5078 	 */
5079 	if ((wk = LIST_FIRST(&oldaip->ai_newdirblk)) != NULL) {
5080 		newdirblk = WK_NEWDIRBLK(wk);
5081 		WORKLIST_REMOVE(&newdirblk->db_list);
5082 		if (!LIST_EMPTY(&oldaip->ai_newdirblk))
5083 			panic("allocindir_merge: extra newdirblk");
5084 		WORKLIST_INSERT(&aip->ai_newdirblk, &newdirblk->db_list);
5085 	}
5086 	/*
5087 	 * We can skip journaling for this freefrag and just complete
5088 	 * any pending journal work for the allocindir that is being
5089 	 * removed after the freefrag completes.
5090 	 */
5091 	if (freefrag->ff_jfreefrag)
5092 		cancel_jfreefrag(freefrag->ff_jfreefrag);
5093 	LIST_REMOVE(oldaip, ai_next);
5094 	cancel_newblk(&oldaip->ai_block, &freefrag->ff_jwork);
5095 	free_newblk(&oldaip->ai_block);
5096 
5097 	return (freefrag);
5098 }
5099 
5100 /*
5101  * Block de-allocation dependencies.
5102  *
5103  * When blocks are de-allocated, the on-disk pointers must be nullified before
5104  * the blocks are made available for use by other files.  (The true
5105  * requirement is that old pointers must be nullified before new on-disk
5106  * pointers are set.  We chose this slightly more stringent requirement to
5107  * reduce complexity.) Our implementation handles this dependency by updating
5108  * the inode (or indirect block) appropriately but delaying the actual block
5109  * de-allocation (i.e., freemap and free space count manipulation) until
5110  * after the updated versions reach stable storage.  After the disk is
5111  * updated, the blocks can be safely de-allocated whenever it is convenient.
5112  * This implementation handles only the common case of reducing a file's
5113  * length to zero. Other cases are handled by the conventional synchronous
5114  * write approach.
5115  *
5116  * The ffs implementation with which we worked double-checks
5117  * the state of the block pointers and file size as it reduces
5118  * a file's length.  Some of this code is replicated here in our
5119  * soft updates implementation.  The freeblks->fb_chkcnt field is
5120  * used to transfer a part of this information to the procedure
5121  * that eventually de-allocates the blocks.
5122  *
5123  * This routine should be called from the routine that shortens
5124  * a file's length, before the inode's size or block pointers
5125  * are modified. It will save the block pointer information for
5126  * later release and zero the inode so that the calling routine
5127  * can release it.
5128  */
5129 void
5130 softdep_setup_freeblocks(ip, length, flags)
5131 	struct inode *ip;	/* The inode whose length is to be reduced */
5132 	off_t length;		/* The new length for the file */
5133 	int flags;		/* IO_EXT and/or IO_NORMAL */
5134 {
5135 	struct ufs1_dinode *dp1;
5136 	struct ufs2_dinode *dp2;
5137 	struct freeblks *freeblks;
5138 	struct inodedep *inodedep;
5139 	struct allocdirect *adp;
5140 	struct jfreeblk *jfreeblk;
5141 	struct bufobj *bo;
5142 	struct vnode *vp;
5143 	struct buf *bp;
5144 	struct fs *fs;
5145 	ufs2_daddr_t extblocks, datablocks;
5146 	struct mount *mp;
5147 	int i, delay, error;
5148 	ufs2_daddr_t blkno;
5149 	ufs_lbn_t tmpval;
5150 	ufs_lbn_t lbn;
5151 	long oldextsize;
5152 	long oldsize;
5153 	int frags;
5154 	int needj;
5155 
5156 	fs = ip->i_fs;
5157 	mp = UFSTOVFS(ip->i_ump);
5158 	if (length != 0)
5159 		panic("softdep_setup_freeblocks: non-zero length");
5160 	freeblks = malloc(sizeof(struct freeblks),
5161 		M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
5162 	workitem_alloc(&freeblks->fb_list, D_FREEBLKS, mp);
5163 	LIST_INIT(&freeblks->fb_jfreeblkhd);
5164 	LIST_INIT(&freeblks->fb_jwork);
5165 	freeblks->fb_state = ATTACHED;
5166 	freeblks->fb_uid = ip->i_uid;
5167 	freeblks->fb_previousinum = ip->i_number;
5168 	freeblks->fb_devvp = ip->i_devvp;
5169 	freeblks->fb_chkcnt = 0;
5170 	ACQUIRE_LOCK(&lk);
5171 	/*
5172 	 * If we're truncating a removed file that will never be written
5173 	 * we don't need to journal the block frees.  The canceled journals
5174 	 * for the allocations will suffice.
5175 	 */
5176 	inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5177 	if ((inodedep->id_state & (UNLINKED | DEPCOMPLETE)) == UNLINKED ||
5178 	    (fs->fs_flags & FS_SUJ) == 0)
5179 		needj = 0;
5180 	else
5181 		needj = 1;
5182 	num_freeblkdep++;
5183 	FREE_LOCK(&lk);
5184 	extblocks = 0;
5185 	if (fs->fs_magic == FS_UFS2_MAGIC)
5186 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
5187 	datablocks = DIP(ip, i_blocks) - extblocks;
5188 	if ((flags & IO_NORMAL) != 0) {
5189 		oldsize = ip->i_size;
5190 		ip->i_size = 0;
5191 		DIP_SET(ip, i_size, 0);
5192 		freeblks->fb_chkcnt = datablocks;
5193 		for (i = 0; i < NDADDR; i++) {
5194 			blkno = DIP(ip, i_db[i]);
5195 			DIP_SET(ip, i_db[i], 0);
5196 			if (blkno == 0)
5197 				continue;
5198 			frags = sblksize(fs, oldsize, i);
5199 			frags = numfrags(fs, frags);
5200 			newfreework(freeblks, NULL, i, blkno, frags, needj);
5201 		}
5202 		for (i = 0, tmpval = NINDIR(fs), lbn = NDADDR; i < NIADDR;
5203 		    i++, tmpval *= NINDIR(fs)) {
5204 			blkno = DIP(ip, i_ib[i]);
5205 			DIP_SET(ip, i_ib[i], 0);
5206 			if (blkno)
5207 				newfreework(freeblks, NULL, -lbn - i, blkno,
5208 				    fs->fs_frag, needj);
5209 			lbn += tmpval;
5210 		}
5211 		UFS_LOCK(ip->i_ump);
5212 		fs->fs_pendingblocks += datablocks;
5213 		UFS_UNLOCK(ip->i_ump);
5214 	}
5215 	if ((flags & IO_EXT) != 0) {
5216 		oldextsize = ip->i_din2->di_extsize;
5217 		ip->i_din2->di_extsize = 0;
5218 		freeblks->fb_chkcnt += extblocks;
5219 		for (i = 0; i < NXADDR; i++) {
5220 			blkno = ip->i_din2->di_extb[i];
5221 			ip->i_din2->di_extb[i] = 0;
5222 			if (blkno == 0)
5223 				continue;
5224 			frags = sblksize(fs, oldextsize, i);
5225 			frags = numfrags(fs, frags);
5226 			newfreework(freeblks, NULL, -1 - i, blkno, frags,
5227 			    needj);
5228 		}
5229 	}
5230 	if (LIST_EMPTY(&freeblks->fb_jfreeblkhd))
5231 		needj = 0;
5232 	DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt);
5233 	/*
5234 	 * Push the zero'ed inode to to its disk buffer so that we are free
5235 	 * to delete its dependencies below. Once the dependencies are gone
5236 	 * the buffer can be safely released.
5237 	 */
5238 	if ((error = bread(ip->i_devvp,
5239 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
5240 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
5241 		brelse(bp);
5242 		softdep_error("softdep_setup_freeblocks", error);
5243 	}
5244 	if (ip->i_ump->um_fstype == UFS1) {
5245 		dp1 = ((struct ufs1_dinode *)bp->b_data +
5246 		    ino_to_fsbo(fs, ip->i_number));
5247 		ip->i_din1->di_freelink = dp1->di_freelink;
5248 		*dp1 = *ip->i_din1;
5249 	} else {
5250 		dp2 = ((struct ufs2_dinode *)bp->b_data +
5251 		    ino_to_fsbo(fs, ip->i_number));
5252 		ip->i_din2->di_freelink = dp2->di_freelink;
5253 		*dp2 = *ip->i_din2;
5254 	}
5255 	/*
5256 	 * Find and eliminate any inode dependencies.
5257 	 */
5258 	ACQUIRE_LOCK(&lk);
5259 	(void) inodedep_lookup(mp, ip->i_number, DEPALLOC, &inodedep);
5260 	if ((inodedep->id_state & IOSTARTED) != 0)
5261 		panic("softdep_setup_freeblocks: inode busy");
5262 	/*
5263 	 * Add the freeblks structure to the list of operations that
5264 	 * must await the zero'ed inode being written to disk. If we
5265 	 * still have a bitmap dependency (delay == 0), then the inode
5266 	 * has never been written to disk, so we can process the
5267 	 * freeblks below once we have deleted the dependencies.
5268 	 */
5269 	delay = (inodedep->id_state & DEPCOMPLETE);
5270 	if (delay)
5271 		WORKLIST_INSERT(&bp->b_dep, &freeblks->fb_list);
5272 	else if (needj)
5273 		freeblks->fb_state |= COMPLETE;
5274 	/*
5275 	 * Because the file length has been truncated to zero, any
5276 	 * pending block allocation dependency structures associated
5277 	 * with this inode are obsolete and can simply be de-allocated.
5278 	 * We must first merge the two dependency lists to get rid of
5279 	 * any duplicate freefrag structures, then purge the merged list.
5280 	 * If we still have a bitmap dependency, then the inode has never
5281 	 * been written to disk, so we can free any fragments without delay.
5282 	 */
5283 	if (flags & IO_NORMAL) {
5284 		merge_inode_lists(&inodedep->id_newinoupdt,
5285 		    &inodedep->id_inoupdt);
5286 		while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
5287 			cancel_allocdirect(&inodedep->id_inoupdt, adp,
5288 			    freeblks, delay);
5289 	}
5290 	if (flags & IO_EXT) {
5291 		merge_inode_lists(&inodedep->id_newextupdt,
5292 		    &inodedep->id_extupdt);
5293 		while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
5294 			cancel_allocdirect(&inodedep->id_extupdt, adp,
5295 			    freeblks, delay);
5296 	}
5297 	LIST_FOREACH(jfreeblk, &freeblks->fb_jfreeblkhd, jf_deps)
5298 		add_to_journal(&jfreeblk->jf_list);
5299 
5300 	FREE_LOCK(&lk);
5301 	bdwrite(bp);
5302 	/*
5303 	 * We must wait for any I/O in progress to finish so that
5304 	 * all potential buffers on the dirty list will be visible.
5305 	 * Once they are all there, walk the list and get rid of
5306 	 * any dependencies.
5307 	 */
5308 	vp = ITOV(ip);
5309 	bo = &vp->v_bufobj;
5310 	BO_LOCK(bo);
5311 	drain_output(vp);
5312 restart:
5313 	TAILQ_FOREACH(bp, &bo->bo_dirty.bv_hd, b_bobufs) {
5314 		if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
5315 		    ((flags & IO_NORMAL) == 0 &&
5316 		      (bp->b_xflags & BX_ALTDATA) == 0))
5317 			continue;
5318 		if ((bp = getdirtybuf(bp, BO_MTX(bo), MNT_WAIT)) == NULL)
5319 			goto restart;
5320 		BO_UNLOCK(bo);
5321 		ACQUIRE_LOCK(&lk);
5322 		(void) inodedep_lookup(mp, ip->i_number, 0, &inodedep);
5323 		if (deallocate_dependencies(bp, inodedep, freeblks))
5324 			bp->b_flags |= B_INVAL | B_NOCACHE;
5325 		FREE_LOCK(&lk);
5326 		brelse(bp);
5327 		BO_LOCK(bo);
5328 		goto restart;
5329 	}
5330 	BO_UNLOCK(bo);
5331 	ACQUIRE_LOCK(&lk);
5332 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
5333 		(void) free_inodedep(inodedep);
5334 
5335 	if (delay || needj)
5336 		freeblks->fb_state |= DEPCOMPLETE;
5337 	if (delay) {
5338 		/*
5339 		 * If the inode with zeroed block pointers is now on disk
5340 		 * we can start freeing blocks. Add freeblks to the worklist
5341 		 * instead of calling  handle_workitem_freeblocks directly as
5342 		 * it is more likely that additional IO is needed to complete
5343 		 * the request here than in the !delay case.
5344 		 */
5345 		if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
5346 			add_to_worklist(&freeblks->fb_list, 1);
5347 	}
5348 	if (needj && LIST_EMPTY(&freeblks->fb_jfreeblkhd))
5349 		needj = 0;
5350 
5351 	FREE_LOCK(&lk);
5352 	/*
5353 	 * If the inode has never been written to disk (delay == 0) and
5354 	 * we're not waiting on any journal writes, then we can process the
5355 	 * freeblks now that we have deleted the dependencies.
5356 	 */
5357 	if (!delay && !needj)
5358 		handle_workitem_freeblocks(freeblks, 0);
5359 }
5360 
5361 /*
5362  * Reclaim any dependency structures from a buffer that is about to
5363  * be reallocated to a new vnode. The buffer must be locked, thus,
5364  * no I/O completion operations can occur while we are manipulating
5365  * its associated dependencies. The mutex is held so that other I/O's
5366  * associated with related dependencies do not occur.  Returns 1 if
5367  * all dependencies were cleared, 0 otherwise.
5368  */
5369 static int
5370 deallocate_dependencies(bp, inodedep, freeblks)
5371 	struct buf *bp;
5372 	struct inodedep *inodedep;
5373 	struct freeblks *freeblks;
5374 {
5375 	struct worklist *wk;
5376 	struct indirdep *indirdep;
5377 	struct newdirblk *newdirblk;
5378 	struct allocindir *aip;
5379 	struct pagedep *pagedep;
5380 	struct jremref *jremref;
5381 	struct jmvref *jmvref;
5382 	struct dirrem *dirrem;
5383 	int i;
5384 
5385 	mtx_assert(&lk, MA_OWNED);
5386 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
5387 		switch (wk->wk_type) {
5388 
5389 		case D_INDIRDEP:
5390 			indirdep = WK_INDIRDEP(wk);
5391 			if (bp->b_lblkno >= 0 ||
5392 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno)
5393 				panic("deallocate_dependencies: not indir");
5394 			cancel_indirdep(indirdep, bp, inodedep, freeblks);
5395 			continue;
5396 
5397 		case D_PAGEDEP:
5398 			pagedep = WK_PAGEDEP(wk);
5399 			/*
5400 			 * There should be no directory add dependencies present
5401 			 * as the directory could not be truncated until all
5402 			 * children were removed.
5403 			 */
5404 			KASSERT(LIST_FIRST(&pagedep->pd_pendinghd) == NULL,
5405 			    ("deallocate_dependencies: pendinghd != NULL"));
5406 			for (i = 0; i < DAHASHSZ; i++)
5407 				KASSERT(LIST_FIRST(&pagedep->pd_diraddhd[i]) == NULL,
5408 				    ("deallocate_dependencies: diraddhd != NULL"));
5409 			/*
5410 			 * Copy any directory remove dependencies to the list
5411 			 * to be processed after the zero'ed inode is written.
5412 			 * If the inode has already been written, then they
5413 			 * can be dumped directly onto the work list.
5414 			 */
5415 			LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
5416 				/*
5417 				 * If there are any dirrems we wait for
5418 				 * the journal write to complete and
5419 				 * then restart the buf scan as the lock
5420 				 * has been dropped.
5421 				 */
5422 				while ((jremref =
5423 				    LIST_FIRST(&dirrem->dm_jremrefhd))
5424 				    != NULL) {
5425 					stat_jwait_filepage++;
5426 					jwait(&jremref->jr_list);
5427 					return (0);
5428 				}
5429 				LIST_REMOVE(dirrem, dm_next);
5430 				dirrem->dm_dirinum = pagedep->pd_ino;
5431 				if (inodedep == NULL ||
5432 				    (inodedep->id_state & ALLCOMPLETE) ==
5433 				     ALLCOMPLETE) {
5434 					dirrem->dm_state |= COMPLETE;
5435 					add_to_worklist(&dirrem->dm_list, 0);
5436 				} else
5437 					WORKLIST_INSERT(&inodedep->id_bufwait,
5438 					    &dirrem->dm_list);
5439 			}
5440 			if ((pagedep->pd_state & NEWBLOCK) != 0) {
5441 				newdirblk = pagedep->pd_newdirblk;
5442 				WORKLIST_REMOVE(&newdirblk->db_list);
5443 				free_newdirblk(newdirblk);
5444 			}
5445 			while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd))
5446 			    != NULL) {
5447 				stat_jwait_filepage++;
5448 				jwait(&jmvref->jm_list);
5449 				return (0);
5450 			}
5451 			WORKLIST_REMOVE(&pagedep->pd_list);
5452 			LIST_REMOVE(pagedep, pd_hash);
5453 			WORKITEM_FREE(pagedep, D_PAGEDEP);
5454 			continue;
5455 
5456 		case D_ALLOCINDIR:
5457 			aip = WK_ALLOCINDIR(wk);
5458 			cancel_allocindir(aip, inodedep, freeblks);
5459 			continue;
5460 
5461 		case D_ALLOCDIRECT:
5462 		case D_INODEDEP:
5463 			panic("deallocate_dependencies: Unexpected type %s",
5464 			    TYPENAME(wk->wk_type));
5465 			/* NOTREACHED */
5466 
5467 		default:
5468 			panic("deallocate_dependencies: Unknown type %s",
5469 			    TYPENAME(wk->wk_type));
5470 			/* NOTREACHED */
5471 		}
5472 	}
5473 
5474 	return (1);
5475 }
5476 
5477 /*
5478  * An allocdirect is being canceled due to a truncate.  We must make sure
5479  * the journal entry is released in concert with the blkfree that releases
5480  * the storage.  Completed journal entries must not be released until the
5481  * space is no longer pointed to by the inode or in the bitmap.
5482  */
5483 static void
5484 cancel_allocdirect(adphead, adp, freeblks, delay)
5485 	struct allocdirectlst *adphead;
5486 	struct allocdirect *adp;
5487 	struct freeblks *freeblks;
5488 	int delay;
5489 {
5490 	struct freework *freework;
5491 	struct newblk *newblk;
5492 	struct worklist *wk;
5493 	ufs_lbn_t lbn;
5494 
5495 	TAILQ_REMOVE(adphead, adp, ad_next);
5496 	newblk = (struct newblk *)adp;
5497 	/*
5498 	 * If the journal hasn't been written the jnewblk must be passed
5499 	 * to the call to ffs_blkfree that reclaims the space.  We accomplish
5500 	 * this by linking the journal dependency into the freework to be
5501 	 * freed when freework_freeblock() is called.  If the journal has
5502 	 * been written we can simply reclaim the journal space when the
5503 	 * freeblks work is complete.
5504 	 */
5505 	if (newblk->nb_jnewblk == NULL) {
5506 		cancel_newblk(newblk, &freeblks->fb_jwork);
5507 		goto found;
5508 	}
5509 	lbn = newblk->nb_jnewblk->jn_lbn;
5510 	/*
5511 	 * Find the correct freework structure so it releases the canceled
5512 	 * journal when the bitmap is cleared.  This preserves rollback
5513 	 * until the allocation is reverted.
5514 	 */
5515 	LIST_FOREACH(wk, &freeblks->fb_freeworkhd, wk_list) {
5516 		freework = WK_FREEWORK(wk);
5517 		if (freework->fw_lbn != lbn)
5518 			continue;
5519 		cancel_newblk(newblk, &freework->fw_jwork);
5520 		goto found;
5521 	}
5522 	panic("cancel_allocdirect: Freework not found for lbn %jd\n", lbn);
5523 found:
5524 	if (delay)
5525 		WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
5526 		    &newblk->nb_list);
5527 	else
5528 		free_newblk(newblk);
5529 	return;
5530 }
5531 
5532 
5533 static void
5534 cancel_newblk(newblk, wkhd)
5535 	struct newblk *newblk;
5536 	struct workhead *wkhd;
5537 {
5538 	struct indirdep *indirdep;
5539 	struct allocindir *aip;
5540 
5541 	while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) {
5542 		indirdep->ir_state &= ~ONDEPLIST;
5543 		LIST_REMOVE(indirdep, ir_next);
5544 		/*
5545 		 * If an indirdep is not on the buf worklist we need to
5546 		 * free it here as deallocate_dependencies() will never
5547 		 * find it.  These pointers were never visible on disk and
5548 		 * can be discarded immediately.
5549 		 */
5550 		while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != NULL) {
5551 			LIST_REMOVE(aip, ai_next);
5552 			cancel_newblk(&aip->ai_block, wkhd);
5553 			free_newblk(&aip->ai_block);
5554 		}
5555 		/*
5556 		 * If this indirdep is not attached to a buf it was simply
5557 		 * waiting on completion to clear completehd.  free_indirdep()
5558 		 * asserts that nothing is dangling.
5559 		 */
5560 		if ((indirdep->ir_state & ONWORKLIST) == 0)
5561 			free_indirdep(indirdep);
5562 	}
5563 	if (newblk->nb_state & ONDEPLIST) {
5564 		newblk->nb_state &= ~ONDEPLIST;
5565 		LIST_REMOVE(newblk, nb_deps);
5566 	}
5567 	if (newblk->nb_state & ONWORKLIST)
5568 		WORKLIST_REMOVE(&newblk->nb_list);
5569 	/*
5570 	 * If the journal entry hasn't been written we hold onto the dep
5571 	 * until it is safe to free along with the other journal work.
5572 	 */
5573 	if (newblk->nb_jnewblk != NULL) {
5574 		cancel_jnewblk(newblk->nb_jnewblk, wkhd);
5575 		newblk->nb_jnewblk = NULL;
5576 	}
5577 	if (!LIST_EMPTY(&newblk->nb_jwork))
5578 		jwork_move(wkhd, &newblk->nb_jwork);
5579 }
5580 
5581 /*
5582  * Free a newblk. Generate a new freefrag work request if appropriate.
5583  * This must be called after the inode pointer and any direct block pointers
5584  * are valid or fully removed via truncate or frag extension.
5585  */
5586 static void
5587 free_newblk(newblk)
5588 	struct newblk *newblk;
5589 {
5590 	struct indirdep *indirdep;
5591 	struct newdirblk *newdirblk;
5592 	struct freefrag *freefrag;
5593 	struct worklist *wk;
5594 
5595 	mtx_assert(&lk, MA_OWNED);
5596 	if (newblk->nb_state & ONDEPLIST)
5597 		LIST_REMOVE(newblk, nb_deps);
5598 	if (newblk->nb_state & ONWORKLIST)
5599 		WORKLIST_REMOVE(&newblk->nb_list);
5600 	LIST_REMOVE(newblk, nb_hash);
5601 	if ((freefrag = newblk->nb_freefrag) != NULL) {
5602 		freefrag->ff_state |= COMPLETE;
5603 		if ((freefrag->ff_state & ALLCOMPLETE) == ALLCOMPLETE)
5604 			add_to_worklist(&freefrag->ff_list, 0);
5605 	}
5606 	if ((wk = LIST_FIRST(&newblk->nb_newdirblk)) != NULL) {
5607 		newdirblk = WK_NEWDIRBLK(wk);
5608 		WORKLIST_REMOVE(&newdirblk->db_list);
5609 		if (!LIST_EMPTY(&newblk->nb_newdirblk))
5610 			panic("free_newblk: extra newdirblk");
5611 		free_newdirblk(newdirblk);
5612 	}
5613 	while ((indirdep = LIST_FIRST(&newblk->nb_indirdeps)) != NULL) {
5614 		indirdep->ir_state |= DEPCOMPLETE;
5615 		indirdep_complete(indirdep);
5616 	}
5617 	KASSERT(newblk->nb_jnewblk == NULL,
5618 	    ("free_newblk; jnewblk %p still attached", newblk->nb_jnewblk));
5619 	handle_jwork(&newblk->nb_jwork);
5620 	newblk->nb_list.wk_type = D_NEWBLK;
5621 	WORKITEM_FREE(newblk, D_NEWBLK);
5622 }
5623 
5624 /*
5625  * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
5626  * This routine must be called with splbio interrupts blocked.
5627  */
5628 static void
5629 free_newdirblk(newdirblk)
5630 	struct newdirblk *newdirblk;
5631 {
5632 	struct pagedep *pagedep;
5633 	struct diradd *dap;
5634 	struct worklist *wk;
5635 	int i;
5636 
5637 	mtx_assert(&lk, MA_OWNED);
5638 	/*
5639 	 * If the pagedep is still linked onto the directory buffer
5640 	 * dependency chain, then some of the entries on the
5641 	 * pd_pendinghd list may not be committed to disk yet. In
5642 	 * this case, we will simply clear the NEWBLOCK flag and
5643 	 * let the pd_pendinghd list be processed when the pagedep
5644 	 * is next written. If the pagedep is no longer on the buffer
5645 	 * dependency chain, then all the entries on the pd_pending
5646 	 * list are committed to disk and we can free them here.
5647 	 */
5648 	pagedep = newdirblk->db_pagedep;
5649 	pagedep->pd_state &= ~NEWBLOCK;
5650 	if ((pagedep->pd_state & ONWORKLIST) == 0)
5651 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
5652 			free_diradd(dap, NULL);
5653 	/*
5654 	 * If no dependencies remain, the pagedep will be freed.
5655 	 */
5656 	for (i = 0; i < DAHASHSZ; i++)
5657 		if (!LIST_EMPTY(&pagedep->pd_diraddhd[i]))
5658 			break;
5659 	if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0 &&
5660 	    LIST_EMPTY(&pagedep->pd_jmvrefhd)) {
5661 		KASSERT(LIST_FIRST(&pagedep->pd_dirremhd) == NULL,
5662 		    ("free_newdirblk: Freeing non-free pagedep %p", pagedep));
5663 		LIST_REMOVE(pagedep, pd_hash);
5664 		WORKITEM_FREE(pagedep, D_PAGEDEP);
5665 	}
5666 	/* Should only ever be one item in the list. */
5667 	while ((wk = LIST_FIRST(&newdirblk->db_mkdir)) != NULL) {
5668 		WORKLIST_REMOVE(wk);
5669 		handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
5670 	}
5671 	WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
5672 }
5673 
5674 /*
5675  * Prepare an inode to be freed. The actual free operation is not
5676  * done until the zero'ed inode has been written to disk.
5677  */
5678 void
5679 softdep_freefile(pvp, ino, mode)
5680 	struct vnode *pvp;
5681 	ino_t ino;
5682 	int mode;
5683 {
5684 	struct inode *ip = VTOI(pvp);
5685 	struct inodedep *inodedep;
5686 	struct freefile *freefile;
5687 
5688 	/*
5689 	 * This sets up the inode de-allocation dependency.
5690 	 */
5691 	freefile = malloc(sizeof(struct freefile),
5692 		M_FREEFILE, M_SOFTDEP_FLAGS);
5693 	workitem_alloc(&freefile->fx_list, D_FREEFILE, pvp->v_mount);
5694 	freefile->fx_mode = mode;
5695 	freefile->fx_oldinum = ino;
5696 	freefile->fx_devvp = ip->i_devvp;
5697 	LIST_INIT(&freefile->fx_jwork);
5698 	UFS_LOCK(ip->i_ump);
5699 	ip->i_fs->fs_pendinginodes += 1;
5700 	UFS_UNLOCK(ip->i_ump);
5701 
5702 	/*
5703 	 * If the inodedep does not exist, then the zero'ed inode has
5704 	 * been written to disk. If the allocated inode has never been
5705 	 * written to disk, then the on-disk inode is zero'ed. In either
5706 	 * case we can free the file immediately.  If the journal was
5707 	 * canceled before being written the inode will never make it to
5708 	 * disk and we must send the canceled journal entrys to
5709 	 * ffs_freefile() to be cleared in conjunction with the bitmap.
5710 	 * Any blocks waiting on the inode to write can be safely freed
5711 	 * here as it will never been written.
5712 	 */
5713 	ACQUIRE_LOCK(&lk);
5714 	inodedep_lookup(pvp->v_mount, ino, 0, &inodedep);
5715 	/*
5716 	 * Remove this inode from the unlinked list and set
5717 	 * GOINGAWAY as appropriate to indicate that this inode
5718 	 * will never be written.
5719 	 */
5720 	if (inodedep && inodedep->id_state & UNLINKED) {
5721 		/*
5722 		 * Save the journal work to be freed with the bitmap
5723 		 * before we clear UNLINKED.  Otherwise it can be lost
5724 		 * if the inode block is written.
5725 		 */
5726 		handle_bufwait(inodedep, &freefile->fx_jwork);
5727 		clear_unlinked_inodedep(inodedep);
5728 		/* Re-acquire inodedep as we've dropped lk. */
5729 		inodedep_lookup(pvp->v_mount, ino, 0, &inodedep);
5730 		if (inodedep && (inodedep->id_state & DEPCOMPLETE) == 0)
5731 			inodedep->id_state |= GOINGAWAY;
5732 	}
5733 	if (inodedep == NULL || check_inode_unwritten(inodedep)) {
5734 		FREE_LOCK(&lk);
5735 		handle_workitem_freefile(freefile);
5736 		return;
5737 	}
5738 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
5739 	FREE_LOCK(&lk);
5740 	if (ip->i_number == ino)
5741 		ip->i_flag |= IN_MODIFIED;
5742 }
5743 
5744 /*
5745  * Check to see if an inode has never been written to disk. If
5746  * so free the inodedep and return success, otherwise return failure.
5747  * This routine must be called with splbio interrupts blocked.
5748  *
5749  * If we still have a bitmap dependency, then the inode has never
5750  * been written to disk. Drop the dependency as it is no longer
5751  * necessary since the inode is being deallocated. We set the
5752  * ALLCOMPLETE flags since the bitmap now properly shows that the
5753  * inode is not allocated. Even if the inode is actively being
5754  * written, it has been rolled back to its zero'ed state, so we
5755  * are ensured that a zero inode is what is on the disk. For short
5756  * lived files, this change will usually result in removing all the
5757  * dependencies from the inode so that it can be freed immediately.
5758  */
5759 static int
5760 check_inode_unwritten(inodedep)
5761 	struct inodedep *inodedep;
5762 {
5763 
5764 	mtx_assert(&lk, MA_OWNED);
5765 
5766 	if ((inodedep->id_state & (DEPCOMPLETE | UNLINKED)) != 0 ||
5767 	    !LIST_EMPTY(&inodedep->id_pendinghd) ||
5768 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
5769 	    !LIST_EMPTY(&inodedep->id_inowait) ||
5770 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
5771 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
5772 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
5773 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
5774 	    inodedep->id_mkdiradd != NULL ||
5775 	    inodedep->id_nlinkdelta != 0)
5776 		return (0);
5777 	/*
5778 	 * Another process might be in initiate_write_inodeblock_ufs[12]
5779 	 * trying to allocate memory without holding "Softdep Lock".
5780 	 */
5781 	if ((inodedep->id_state & IOSTARTED) != 0 &&
5782 	    inodedep->id_savedino1 == NULL)
5783 		return (0);
5784 
5785 	if (inodedep->id_state & ONDEPLIST)
5786 		LIST_REMOVE(inodedep, id_deps);
5787 	inodedep->id_state &= ~ONDEPLIST;
5788 	inodedep->id_state |= ALLCOMPLETE;
5789 	inodedep->id_bmsafemap = NULL;
5790 	if (inodedep->id_state & ONWORKLIST)
5791 		WORKLIST_REMOVE(&inodedep->id_list);
5792 	if (inodedep->id_savedino1 != NULL) {
5793 		free(inodedep->id_savedino1, M_SAVEDINO);
5794 		inodedep->id_savedino1 = NULL;
5795 	}
5796 	if (free_inodedep(inodedep) == 0)
5797 		panic("check_inode_unwritten: busy inode");
5798 	return (1);
5799 }
5800 
5801 /*
5802  * Try to free an inodedep structure. Return 1 if it could be freed.
5803  */
5804 static int
5805 free_inodedep(inodedep)
5806 	struct inodedep *inodedep;
5807 {
5808 
5809 	mtx_assert(&lk, MA_OWNED);
5810 	if ((inodedep->id_state & (ONWORKLIST | UNLINKED)) != 0 ||
5811 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
5812 	    !LIST_EMPTY(&inodedep->id_dirremhd) ||
5813 	    !LIST_EMPTY(&inodedep->id_pendinghd) ||
5814 	    !LIST_EMPTY(&inodedep->id_bufwait) ||
5815 	    !LIST_EMPTY(&inodedep->id_inowait) ||
5816 	    !TAILQ_EMPTY(&inodedep->id_inoreflst) ||
5817 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
5818 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt) ||
5819 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
5820 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
5821 	    inodedep->id_mkdiradd != NULL ||
5822 	    inodedep->id_nlinkdelta != 0 ||
5823 	    inodedep->id_savedino1 != NULL)
5824 		return (0);
5825 	if (inodedep->id_state & ONDEPLIST)
5826 		LIST_REMOVE(inodedep, id_deps);
5827 	LIST_REMOVE(inodedep, id_hash);
5828 	WORKITEM_FREE(inodedep, D_INODEDEP);
5829 	num_inodedep -= 1;
5830 	return (1);
5831 }
5832 
5833 /*
5834  * Free the block referenced by a freework structure.  The parent freeblks
5835  * structure is released and completed when the final cg bitmap reaches
5836  * the disk.  This routine may be freeing a jnewblk which never made it to
5837  * disk in which case we do not have to wait as the operation is undone
5838  * in memory immediately.
5839  */
5840 static void
5841 freework_freeblock(freework)
5842 	struct freework *freework;
5843 {
5844 	struct freeblks *freeblks;
5845 	struct ufsmount *ump;
5846 	struct workhead wkhd;
5847 	struct fs *fs;
5848 	int complete;
5849 	int pending;
5850 	int bsize;
5851 	int needj;
5852 
5853 	freeblks = freework->fw_freeblks;
5854 	ump = VFSTOUFS(freeblks->fb_list.wk_mp);
5855 	fs = ump->um_fs;
5856 	needj = freeblks->fb_list.wk_mp->mnt_kern_flag & MNTK_SUJ;
5857 	complete = 0;
5858 	LIST_INIT(&wkhd);
5859 	/*
5860 	 * If we are canceling an existing jnewblk pass it to the free
5861 	 * routine, otherwise pass the freeblk which will ultimately
5862 	 * release the freeblks.  If we're not journaling, we can just
5863 	 * free the freeblks immediately.
5864 	 */
5865 	if (!LIST_EMPTY(&freework->fw_jwork)) {
5866 		LIST_SWAP(&wkhd, &freework->fw_jwork, worklist, wk_list);
5867 		complete = 1;
5868 	} else if (needj)
5869 		WORKLIST_INSERT_UNLOCKED(&wkhd, &freework->fw_list);
5870 	bsize = lfragtosize(fs, freework->fw_frags);
5871 	pending = btodb(bsize);
5872 	ACQUIRE_LOCK(&lk);
5873 	freeblks->fb_chkcnt -= pending;
5874 	FREE_LOCK(&lk);
5875 	/*
5876 	 * extattr blocks don't show up in pending blocks.  XXX why?
5877 	 */
5878 	if (freework->fw_lbn >= 0 || freework->fw_lbn <= -NDADDR) {
5879 		UFS_LOCK(ump);
5880 		fs->fs_pendingblocks -= pending;
5881 		UFS_UNLOCK(ump);
5882 	}
5883 	ffs_blkfree(ump, fs, freeblks->fb_devvp, freework->fw_blkno,
5884 	    bsize, freeblks->fb_previousinum, &wkhd);
5885 	if (complete == 0 && needj)
5886 		return;
5887 	/*
5888 	 * The jnewblk will be discarded and the bits in the map never
5889 	 * made it to disk.  We can immediately free the freeblk.
5890 	 */
5891 	ACQUIRE_LOCK(&lk);
5892 	handle_written_freework(freework);
5893 	FREE_LOCK(&lk);
5894 }
5895 
5896 /*
5897  * Start, continue, or finish the process of freeing an indirect block tree.
5898  * The free operation may be paused at any point with fw_off containing the
5899  * offset to restart from.  This enables us to implement some flow control
5900  * for large truncates which may fan out and generate a huge number of
5901  * dependencies.
5902  */
5903 static void
5904 handle_workitem_indirblk(freework)
5905 	struct freework *freework;
5906 {
5907 	struct freeblks *freeblks;
5908 	struct ufsmount *ump;
5909 	struct fs *fs;
5910 
5911 
5912 	freeblks = freework->fw_freeblks;
5913 	ump = VFSTOUFS(freeblks->fb_list.wk_mp);
5914 	fs = ump->um_fs;
5915 	if (freework->fw_off == NINDIR(fs))
5916 		freework_freeblock(freework);
5917 	else
5918 		indir_trunc(freework, fsbtodb(fs, freework->fw_blkno),
5919 		    freework->fw_lbn);
5920 }
5921 
5922 /*
5923  * Called when a freework structure attached to a cg buf is written.  The
5924  * ref on either the parent or the freeblks structure is released and
5925  * either may be added to the worklist if it is the final ref.
5926  */
5927 static void
5928 handle_written_freework(freework)
5929 	struct freework *freework;
5930 {
5931 	struct freeblks *freeblks;
5932 	struct freework *parent;
5933 
5934 	freeblks = freework->fw_freeblks;
5935 	parent = freework->fw_parent;
5936 	if (parent) {
5937 		if (--parent->fw_ref != 0)
5938 			parent = NULL;
5939 		freeblks = NULL;
5940 	} else if (--freeblks->fb_ref != 0)
5941 		freeblks = NULL;
5942 	WORKITEM_FREE(freework, D_FREEWORK);
5943 	/*
5944 	 * Don't delay these block frees or it takes an intolerable amount
5945 	 * of time to process truncates and free their journal entries.
5946 	 */
5947 	if (freeblks)
5948 		add_to_worklist(&freeblks->fb_list, 1);
5949 	if (parent)
5950 		add_to_worklist(&parent->fw_list, 1);
5951 }
5952 
5953 /*
5954  * This workitem routine performs the block de-allocation.
5955  * The workitem is added to the pending list after the updated
5956  * inode block has been written to disk.  As mentioned above,
5957  * checks regarding the number of blocks de-allocated (compared
5958  * to the number of blocks allocated for the file) are also
5959  * performed in this function.
5960  */
5961 static void
5962 handle_workitem_freeblocks(freeblks, flags)
5963 	struct freeblks *freeblks;
5964 	int flags;
5965 {
5966 	struct freework *freework;
5967 	struct worklist *wk;
5968 
5969 	KASSERT(LIST_EMPTY(&freeblks->fb_jfreeblkhd),
5970 	    ("handle_workitem_freeblocks: Journal entries not written."));
5971 	if (LIST_EMPTY(&freeblks->fb_freeworkhd)) {
5972 		handle_complete_freeblocks(freeblks);
5973 		return;
5974 	}
5975 	freeblks->fb_ref++;
5976 	while ((wk = LIST_FIRST(&freeblks->fb_freeworkhd)) != NULL) {
5977 		KASSERT(wk->wk_type == D_FREEWORK,
5978 		    ("handle_workitem_freeblocks: Unknown type %s",
5979 		    TYPENAME(wk->wk_type)));
5980 		WORKLIST_REMOVE_UNLOCKED(wk);
5981 		freework = WK_FREEWORK(wk);
5982 		if (freework->fw_lbn <= -NDADDR)
5983 			handle_workitem_indirblk(freework);
5984 		else
5985 			freework_freeblock(freework);
5986 	}
5987 	ACQUIRE_LOCK(&lk);
5988 	if (--freeblks->fb_ref != 0)
5989 		freeblks = NULL;
5990 	FREE_LOCK(&lk);
5991 	if (freeblks)
5992 		handle_complete_freeblocks(freeblks);
5993 }
5994 
5995 /*
5996  * Once all of the freework workitems are complete we can retire the
5997  * freeblocks dependency and any journal work awaiting completion.  This
5998  * can not be called until all other dependencies are stable on disk.
5999  */
6000 static void
6001 handle_complete_freeblocks(freeblks)
6002 	struct freeblks *freeblks;
6003 {
6004 	struct inode *ip;
6005 	struct vnode *vp;
6006 	struct fs *fs;
6007 	struct ufsmount *ump;
6008 	int flags;
6009 
6010 	ump = VFSTOUFS(freeblks->fb_list.wk_mp);
6011 	fs = ump->um_fs;
6012 	flags = LK_NOWAIT;
6013 
6014 	/*
6015 	 * If we still have not finished background cleanup, then check
6016 	 * to see if the block count needs to be adjusted.
6017 	 */
6018 	if (freeblks->fb_chkcnt != 0 && (fs->fs_flags & FS_UNCLEAN) != 0 &&
6019 	    ffs_vgetf(freeblks->fb_list.wk_mp, freeblks->fb_previousinum,
6020 	    (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp, FFSV_FORCEINSMQ) == 0) {
6021 		ip = VTOI(vp);
6022 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + freeblks->fb_chkcnt);
6023 		ip->i_flag |= IN_CHANGE;
6024 		vput(vp);
6025 	}
6026 
6027 	KASSERT(freeblks->fb_chkcnt == 0 ||
6028 	    ((fs->fs_flags & FS_UNCLEAN) != 0 && (flags & LK_NOWAIT) == 0),
6029 	    ("handle_workitem_freeblocks: inode %ju block count %jd\n",
6030 	    (uintmax_t)freeblks->fb_previousinum,
6031 	    (intmax_t)freeblks->fb_chkcnt));
6032 
6033 	ACQUIRE_LOCK(&lk);
6034 	/*
6035 	 * All of the freeblock deps must be complete prior to this call
6036 	 * so it's now safe to complete earlier outstanding journal entries.
6037 	 */
6038 	handle_jwork(&freeblks->fb_jwork);
6039 	WORKITEM_FREE(freeblks, D_FREEBLKS);
6040 	num_freeblkdep--;
6041 	FREE_LOCK(&lk);
6042 }
6043 
6044 /*
6045  * Release blocks associated with the inode ip and stored in the indirect
6046  * block dbn. If level is greater than SINGLE, the block is an indirect block
6047  * and recursive calls to indirtrunc must be used to cleanse other indirect
6048  * blocks.
6049  */
6050 static void
6051 indir_trunc(freework, dbn, lbn)
6052 	struct freework *freework;
6053 	ufs2_daddr_t dbn;
6054 	ufs_lbn_t lbn;
6055 {
6056 	struct freework *nfreework;
6057 	struct workhead wkhd;
6058 	struct jnewblk *jnewblk;
6059 	struct freeblks *freeblks;
6060 	struct buf *bp;
6061 	struct fs *fs;
6062 	struct worklist *wkn;
6063 	struct worklist *wk;
6064 	struct indirdep *indirdep;
6065 	struct ufsmount *ump;
6066 	ufs1_daddr_t *bap1 = 0;
6067 	ufs2_daddr_t nb, nnb, *bap2 = 0;
6068 	ufs_lbn_t lbnadd;
6069 	int i, nblocks, ufs1fmt;
6070 	int fs_pendingblocks;
6071 	int freedeps;
6072 	int needj;
6073 	int level;
6074 	int cnt;
6075 
6076 	LIST_INIT(&wkhd);
6077 	level = lbn_level(lbn);
6078 	if (level == -1)
6079 		panic("indir_trunc: Invalid lbn %jd\n", lbn);
6080 	freeblks = freework->fw_freeblks;
6081 	ump = VFSTOUFS(freeblks->fb_list.wk_mp);
6082 	fs = ump->um_fs;
6083 	fs_pendingblocks = 0;
6084 	freedeps = 0;
6085 	needj = UFSTOVFS(ump)->mnt_kern_flag & MNTK_SUJ;
6086 	lbnadd = lbn_offset(fs, level);
6087 	/*
6088 	 * Get buffer of block pointers to be freed. This routine is not
6089 	 * called until the zero'ed inode has been written, so it is safe
6090 	 * to free blocks as they are encountered. Because the inode has
6091 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
6092 	 * have to use the on-disk address and the block device for the
6093 	 * filesystem to look them up. If the file was deleted before its
6094 	 * indirect blocks were all written to disk, the routine that set
6095 	 * us up (deallocate_dependencies) will have arranged to leave
6096 	 * a complete copy of the indirect block in memory for our use.
6097 	 * Otherwise we have to read the blocks in from the disk.
6098 	 */
6099 #ifdef notyet
6100 	bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0,
6101 	    GB_NOCREAT);
6102 #else
6103 	bp = incore(&freeblks->fb_devvp->v_bufobj, dbn);
6104 #endif
6105 	ACQUIRE_LOCK(&lk);
6106 	if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
6107 		if (wk->wk_type != D_INDIRDEP ||
6108 		    (wk->wk_state & GOINGAWAY) == 0)
6109 			panic("indir_trunc: lost indirdep %p", wk);
6110 		indirdep = WK_INDIRDEP(wk);
6111 		LIST_SWAP(&wkhd, &indirdep->ir_jwork, worklist, wk_list);
6112 		free_indirdep(indirdep);
6113 		if (!LIST_EMPTY(&bp->b_dep))
6114 			panic("indir_trunc: dangling dep %p",
6115 			    LIST_FIRST(&bp->b_dep));
6116 		ump->um_numindirdeps -= 1;
6117 		FREE_LOCK(&lk);
6118 	} else {
6119 #ifdef notyet
6120 		if (bp)
6121 			brelse(bp);
6122 #endif
6123 		FREE_LOCK(&lk);
6124 		if (bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
6125 		    NOCRED, &bp) != 0) {
6126 			brelse(bp);
6127 			return;
6128 		}
6129 	}
6130 	/*
6131 	 * Recursively free indirect blocks.
6132 	 */
6133 	if (ump->um_fstype == UFS1) {
6134 		ufs1fmt = 1;
6135 		bap1 = (ufs1_daddr_t *)bp->b_data;
6136 	} else {
6137 		ufs1fmt = 0;
6138 		bap2 = (ufs2_daddr_t *)bp->b_data;
6139 	}
6140 	/*
6141 	 * Reclaim indirect blocks which never made it to disk.
6142 	 */
6143 	cnt = 0;
6144 	LIST_FOREACH_SAFE(wk, &wkhd, wk_list, wkn) {
6145 		struct workhead freewk;
6146 		if (wk->wk_type != D_JNEWBLK)
6147 			continue;
6148 		WORKLIST_REMOVE_UNLOCKED(wk);
6149 		LIST_INIT(&freewk);
6150 		WORKLIST_INSERT_UNLOCKED(&freewk, wk);
6151 		jnewblk = WK_JNEWBLK(wk);
6152 		if (jnewblk->jn_lbn > 0)
6153 			i = (jnewblk->jn_lbn - -lbn) / lbnadd;
6154 		else
6155 			i = (jnewblk->jn_lbn - (lbn + 1)) / lbnadd;
6156 		KASSERT(i >= 0 && i < NINDIR(fs),
6157 		    ("indir_trunc: Index out of range %d parent %jd lbn %jd",
6158 		    i, lbn, jnewblk->jn_lbn));
6159 		/* Clear the pointer so it isn't found below. */
6160 		if (ufs1fmt) {
6161 			nb = bap1[i];
6162 			bap1[i] = 0;
6163 		} else {
6164 			nb = bap2[i];
6165 			bap2[i] = 0;
6166 		}
6167 		KASSERT(nb == jnewblk->jn_blkno,
6168 		    ("indir_trunc: Block mismatch %jd != %jd",
6169 		    nb, jnewblk->jn_blkno));
6170 		ffs_blkfree(ump, fs, freeblks->fb_devvp, jnewblk->jn_blkno,
6171 		    fs->fs_bsize, freeblks->fb_previousinum, &freewk);
6172 		cnt++;
6173 	}
6174 	ACQUIRE_LOCK(&lk);
6175 	if (needj)
6176 		freework->fw_ref += NINDIR(fs) + 1;
6177 	/* Any remaining journal work can be completed with freeblks. */
6178 	jwork_move(&freeblks->fb_jwork, &wkhd);
6179 	FREE_LOCK(&lk);
6180 	nblocks = btodb(fs->fs_bsize);
6181 	if (ufs1fmt)
6182 		nb = bap1[0];
6183 	else
6184 		nb = bap2[0];
6185 	nfreework = freework;
6186 	/*
6187 	 * Reclaim on disk blocks.
6188 	 */
6189 	for (i = freework->fw_off; i < NINDIR(fs); i++, nb = nnb) {
6190 		if (i != NINDIR(fs) - 1) {
6191 			if (ufs1fmt)
6192 				nnb = bap1[i+1];
6193 			else
6194 				nnb = bap2[i+1];
6195 		} else
6196 			nnb = 0;
6197 		if (nb == 0)
6198 			continue;
6199 		cnt++;
6200 		if (level != 0) {
6201 			ufs_lbn_t nlbn;
6202 
6203 			nlbn = (lbn + 1) - (i * lbnadd);
6204 			if (needj != 0) {
6205 				nfreework = newfreework(freeblks, freework,
6206 				    nlbn, nb, fs->fs_frag, 0);
6207 				freedeps++;
6208 			}
6209 			indir_trunc(nfreework, fsbtodb(fs, nb), nlbn);
6210 		} else {
6211 			struct freedep *freedep;
6212 
6213 			/*
6214 			 * Attempt to aggregate freedep dependencies for
6215 			 * all blocks being released to the same CG.
6216 			 */
6217 			LIST_INIT(&wkhd);
6218 			if (needj != 0 &&
6219 			    (nnb == 0 || (dtog(fs, nb) != dtog(fs, nnb)))) {
6220 				freedep = newfreedep(freework);
6221 				WORKLIST_INSERT_UNLOCKED(&wkhd,
6222 				    &freedep->fd_list);
6223 				freedeps++;
6224 			}
6225 			ffs_blkfree(ump, fs, freeblks->fb_devvp, nb,
6226 			    fs->fs_bsize, freeblks->fb_previousinum, &wkhd);
6227 		}
6228 	}
6229 	if (level == 0)
6230 		fs_pendingblocks = (nblocks * cnt);
6231 	/*
6232 	 * If we're not journaling we can free the indirect now.  Otherwise
6233 	 * setup the ref counts and offset so this indirect can be completed
6234 	 * when its children are free.
6235 	 */
6236 	if (needj == 0) {
6237 		fs_pendingblocks += nblocks;
6238 		dbn = dbtofsb(fs, dbn);
6239 		ffs_blkfree(ump, fs, freeblks->fb_devvp, dbn, fs->fs_bsize,
6240 		    freeblks->fb_previousinum, NULL);
6241 		ACQUIRE_LOCK(&lk);
6242 		freeblks->fb_chkcnt -= fs_pendingblocks;
6243 		if (freework->fw_blkno == dbn)
6244 			handle_written_freework(freework);
6245 		FREE_LOCK(&lk);
6246 		freework = NULL;
6247 	} else {
6248 		ACQUIRE_LOCK(&lk);
6249 		freework->fw_off = i;
6250 		freework->fw_ref += freedeps;
6251 		freework->fw_ref -= NINDIR(fs) + 1;
6252 		if (freework->fw_ref != 0)
6253 			freework = NULL;
6254 		freeblks->fb_chkcnt -= fs_pendingblocks;
6255 		FREE_LOCK(&lk);
6256 	}
6257 	if (fs_pendingblocks) {
6258 		UFS_LOCK(ump);
6259 		fs->fs_pendingblocks -= fs_pendingblocks;
6260 		UFS_UNLOCK(ump);
6261 	}
6262 	bp->b_flags |= B_INVAL | B_NOCACHE;
6263 	brelse(bp);
6264 	if (freework)
6265 		handle_workitem_indirblk(freework);
6266 	return;
6267 }
6268 
6269 /*
6270  * Cancel an allocindir when it is removed via truncation.
6271  */
6272 static void
6273 cancel_allocindir(aip, inodedep, freeblks)
6274 	struct allocindir *aip;
6275 	struct inodedep *inodedep;
6276 	struct freeblks *freeblks;
6277 {
6278 	struct newblk *newblk;
6279 
6280 	/*
6281 	 * If the journal hasn't been written the jnewblk must be passed
6282 	 * to the call to ffs_blkfree that reclaims the space.  We accomplish
6283 	 * this by linking the journal dependency into the indirdep to be
6284 	 * freed when indir_trunc() is called.  If the journal has already
6285 	 * been written we can simply reclaim the journal space when the
6286 	 * freeblks work is complete.
6287 	 */
6288 	LIST_REMOVE(aip, ai_next);
6289 	newblk = (struct newblk *)aip;
6290 	if (newblk->nb_jnewblk == NULL)
6291 		cancel_newblk(newblk, &freeblks->fb_jwork);
6292 	else
6293 		cancel_newblk(newblk, &aip->ai_indirdep->ir_jwork);
6294 	if (inodedep && inodedep->id_state & DEPCOMPLETE)
6295 		WORKLIST_INSERT(&inodedep->id_bufwait, &newblk->nb_list);
6296 	else
6297 		free_newblk(newblk);
6298 }
6299 
6300 /*
6301  * Create the mkdir dependencies for . and .. in a new directory.  Link them
6302  * in to a newdirblk so any subsequent additions are tracked properly.  The
6303  * caller is responsible for adding the mkdir1 dependency to the journal
6304  * and updating id_mkdiradd.  This function returns with lk held.
6305  */
6306 static struct mkdir *
6307 setup_newdir(dap, newinum, dinum, newdirbp, mkdirp)
6308 	struct diradd *dap;
6309 	ino_t newinum;
6310 	ino_t dinum;
6311 	struct buf *newdirbp;
6312 	struct mkdir **mkdirp;
6313 {
6314 	struct newblk *newblk;
6315 	struct pagedep *pagedep;
6316 	struct inodedep *inodedep;
6317 	struct newdirblk *newdirblk = 0;
6318 	struct mkdir *mkdir1, *mkdir2;
6319 	struct worklist *wk;
6320 	struct jaddref *jaddref;
6321 	struct mount *mp;
6322 
6323 	mp = dap->da_list.wk_mp;
6324 	newdirblk = malloc(sizeof(struct newdirblk), M_NEWDIRBLK,
6325 	    M_SOFTDEP_FLAGS);
6326 	workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
6327 	LIST_INIT(&newdirblk->db_mkdir);
6328 	mkdir1 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS);
6329 	workitem_alloc(&mkdir1->md_list, D_MKDIR, mp);
6330 	mkdir1->md_state = ATTACHED | MKDIR_BODY;
6331 	mkdir1->md_diradd = dap;
6332 	mkdir1->md_jaddref = NULL;
6333 	mkdir2 = malloc(sizeof(struct mkdir), M_MKDIR, M_SOFTDEP_FLAGS);
6334 	workitem_alloc(&mkdir2->md_list, D_MKDIR, mp);
6335 	mkdir2->md_state = ATTACHED | MKDIR_PARENT;
6336 	mkdir2->md_diradd = dap;
6337 	mkdir2->md_jaddref = NULL;
6338 	if ((mp->mnt_kern_flag & MNTK_SUJ) == 0) {
6339 		mkdir1->md_state |= DEPCOMPLETE;
6340 		mkdir2->md_state |= DEPCOMPLETE;
6341 	}
6342 	/*
6343 	 * Dependency on "." and ".." being written to disk.
6344 	 */
6345 	mkdir1->md_buf = newdirbp;
6346 	ACQUIRE_LOCK(&lk);
6347 	LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
6348 	/*
6349 	 * We must link the pagedep, allocdirect, and newdirblk for
6350 	 * the initial file page so the pointer to the new directory
6351 	 * is not written until the directory contents are live and
6352 	 * any subsequent additions are not marked live until the
6353 	 * block is reachable via the inode.
6354 	 */
6355 	if (pagedep_lookup(mp, newinum, 0, 0, &pagedep) == 0)
6356 		panic("setup_newdir: lost pagedep");
6357 	LIST_FOREACH(wk, &newdirbp->b_dep, wk_list)
6358 		if (wk->wk_type == D_ALLOCDIRECT)
6359 			break;
6360 	if (wk == NULL)
6361 		panic("setup_newdir: lost allocdirect");
6362 	newblk = WK_NEWBLK(wk);
6363 	pagedep->pd_state |= NEWBLOCK;
6364 	pagedep->pd_newdirblk = newdirblk;
6365 	newdirblk->db_pagedep = pagedep;
6366 	WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list);
6367 	WORKLIST_INSERT(&newdirblk->db_mkdir, &mkdir1->md_list);
6368 	/*
6369 	 * Look up the inodedep for the parent directory so that we
6370 	 * can link mkdir2 into the pending dotdot jaddref or
6371 	 * the inode write if there is none.  If the inode is
6372 	 * ALLCOMPLETE and no jaddref is present all dependencies have
6373 	 * been satisfied and mkdir2 can be freed.
6374 	 */
6375 	inodedep_lookup(mp, dinum, 0, &inodedep);
6376 	if (mp->mnt_kern_flag & MNTK_SUJ) {
6377 		if (inodedep == NULL)
6378 			panic("setup_newdir: Lost parent.");
6379 		jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
6380 		    inoreflst);
6381 		KASSERT(jaddref != NULL && jaddref->ja_parent == newinum &&
6382 		    (jaddref->ja_state & MKDIR_PARENT),
6383 		    ("setup_newdir: bad dotdot jaddref %p", jaddref));
6384 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
6385 		mkdir2->md_jaddref = jaddref;
6386 		jaddref->ja_mkdir = mkdir2;
6387 	} else if (inodedep == NULL ||
6388 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
6389 		dap->da_state &= ~MKDIR_PARENT;
6390 		WORKITEM_FREE(mkdir2, D_MKDIR);
6391 	} else {
6392 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
6393 		WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
6394 	}
6395 	*mkdirp = mkdir2;
6396 
6397 	return (mkdir1);
6398 }
6399 
6400 /*
6401  * Directory entry addition dependencies.
6402  *
6403  * When adding a new directory entry, the inode (with its incremented link
6404  * count) must be written to disk before the directory entry's pointer to it.
6405  * Also, if the inode is newly allocated, the corresponding freemap must be
6406  * updated (on disk) before the directory entry's pointer. These requirements
6407  * are met via undo/redo on the directory entry's pointer, which consists
6408  * simply of the inode number.
6409  *
6410  * As directory entries are added and deleted, the free space within a
6411  * directory block can become fragmented.  The ufs filesystem will compact
6412  * a fragmented directory block to make space for a new entry. When this
6413  * occurs, the offsets of previously added entries change. Any "diradd"
6414  * dependency structures corresponding to these entries must be updated with
6415  * the new offsets.
6416  */
6417 
6418 /*
6419  * This routine is called after the in-memory inode's link
6420  * count has been incremented, but before the directory entry's
6421  * pointer to the inode has been set.
6422  */
6423 int
6424 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
6425 	struct buf *bp;		/* buffer containing directory block */
6426 	struct inode *dp;	/* inode for directory */
6427 	off_t diroffset;	/* offset of new entry in directory */
6428 	ino_t newinum;		/* inode referenced by new directory entry */
6429 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
6430 	int isnewblk;		/* entry is in a newly allocated block */
6431 {
6432 	int offset;		/* offset of new entry within directory block */
6433 	ufs_lbn_t lbn;		/* block in directory containing new entry */
6434 	struct fs *fs;
6435 	struct diradd *dap;
6436 	struct newblk *newblk;
6437 	struct pagedep *pagedep;
6438 	struct inodedep *inodedep;
6439 	struct newdirblk *newdirblk = 0;
6440 	struct mkdir *mkdir1, *mkdir2;
6441 	struct jaddref *jaddref;
6442 	struct mount *mp;
6443 	int isindir;
6444 
6445 	/*
6446 	 * Whiteouts have no dependencies.
6447 	 */
6448 	if (newinum == WINO) {
6449 		if (newdirbp != NULL)
6450 			bdwrite(newdirbp);
6451 		return (0);
6452 	}
6453 	jaddref = NULL;
6454 	mkdir1 = mkdir2 = NULL;
6455 	mp = UFSTOVFS(dp->i_ump);
6456 	fs = dp->i_fs;
6457 	lbn = lblkno(fs, diroffset);
6458 	offset = blkoff(fs, diroffset);
6459 	dap = malloc(sizeof(struct diradd), M_DIRADD,
6460 		M_SOFTDEP_FLAGS|M_ZERO);
6461 	workitem_alloc(&dap->da_list, D_DIRADD, mp);
6462 	dap->da_offset = offset;
6463 	dap->da_newinum = newinum;
6464 	dap->da_state = ATTACHED;
6465 	LIST_INIT(&dap->da_jwork);
6466 	isindir = bp->b_lblkno >= NDADDR;
6467 	if (isnewblk &&
6468 	    (isindir ? blkoff(fs, diroffset) : fragoff(fs, diroffset)) == 0) {
6469 		newdirblk = malloc(sizeof(struct newdirblk),
6470 		    M_NEWDIRBLK, M_SOFTDEP_FLAGS);
6471 		workitem_alloc(&newdirblk->db_list, D_NEWDIRBLK, mp);
6472 		LIST_INIT(&newdirblk->db_mkdir);
6473 	}
6474 	/*
6475 	 * If we're creating a new directory setup the dependencies and set
6476 	 * the dap state to wait for them.  Otherwise it's COMPLETE and
6477 	 * we can move on.
6478 	 */
6479 	if (newdirbp == NULL) {
6480 		dap->da_state |= DEPCOMPLETE;
6481 		ACQUIRE_LOCK(&lk);
6482 	} else {
6483 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
6484 		mkdir1 = setup_newdir(dap, newinum, dp->i_number, newdirbp,
6485 		    &mkdir2);
6486 	}
6487 	/*
6488 	 * Link into parent directory pagedep to await its being written.
6489 	 */
6490 	if (pagedep_lookup(mp, dp->i_number, lbn, DEPALLOC, &pagedep) == 0)
6491 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
6492 #ifdef DEBUG
6493 	if (diradd_lookup(pagedep, offset) != NULL)
6494 		panic("softdep_setup_directory_add: %p already at off %d\n",
6495 		    diradd_lookup(pagedep, offset), offset);
6496 #endif
6497 	dap->da_pagedep = pagedep;
6498 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
6499 	    da_pdlist);
6500 	inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
6501 	/*
6502 	 * If we're journaling, link the diradd into the jaddref so it
6503 	 * may be completed after the journal entry is written.  Otherwise,
6504 	 * link the diradd into its inodedep.  If the inode is not yet
6505 	 * written place it on the bufwait list, otherwise do the post-inode
6506 	 * write processing to put it on the id_pendinghd list.
6507 	 */
6508 	if (mp->mnt_kern_flag & MNTK_SUJ) {
6509 		jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
6510 		    inoreflst);
6511 		KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
6512 		    ("softdep_setup_directory_add: bad jaddref %p", jaddref));
6513 		jaddref->ja_diroff = diroffset;
6514 		jaddref->ja_diradd = dap;
6515 		add_to_journal(&jaddref->ja_list);
6516 	} else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
6517 		diradd_inode_written(dap, inodedep);
6518 	else
6519 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
6520 	/*
6521 	 * Add the journal entries for . and .. links now that the primary
6522 	 * link is written.
6523 	 */
6524 	if (mkdir1 != NULL && mp->mnt_kern_flag & MNTK_SUJ) {
6525 		jaddref = (struct jaddref *)TAILQ_PREV(&jaddref->ja_ref,
6526 		    inoreflst, if_deps);
6527 		KASSERT(jaddref != NULL &&
6528 		    jaddref->ja_ino == jaddref->ja_parent &&
6529 		    (jaddref->ja_state & MKDIR_BODY),
6530 		    ("softdep_setup_directory_add: bad dot jaddref %p",
6531 		    jaddref));
6532 		mkdir1->md_jaddref = jaddref;
6533 		jaddref->ja_mkdir = mkdir1;
6534 		/*
6535 		 * It is important that the dotdot journal entry
6536 		 * is added prior to the dot entry since dot writes
6537 		 * both the dot and dotdot links.  These both must
6538 		 * be added after the primary link for the journal
6539 		 * to remain consistent.
6540 		 */
6541 		add_to_journal(&mkdir2->md_jaddref->ja_list);
6542 		add_to_journal(&jaddref->ja_list);
6543 	}
6544 	/*
6545 	 * If we are adding a new directory remember this diradd so that if
6546 	 * we rename it we can keep the dot and dotdot dependencies.  If
6547 	 * we are adding a new name for an inode that has a mkdiradd we
6548 	 * must be in rename and we have to move the dot and dotdot
6549 	 * dependencies to this new name.  The old name is being orphaned
6550 	 * soon.
6551 	 */
6552 	if (mkdir1 != NULL) {
6553 		if (inodedep->id_mkdiradd != NULL)
6554 			panic("softdep_setup_directory_add: Existing mkdir");
6555 		inodedep->id_mkdiradd = dap;
6556 	} else if (inodedep->id_mkdiradd)
6557 		merge_diradd(inodedep, dap);
6558 	if (newdirblk) {
6559 		/*
6560 		 * There is nothing to do if we are already tracking
6561 		 * this block.
6562 		 */
6563 		if ((pagedep->pd_state & NEWBLOCK) != 0) {
6564 			WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
6565 			FREE_LOCK(&lk);
6566 			return (0);
6567 		}
6568 		if (newblk_lookup(mp, dbtofsb(fs, bp->b_blkno), 0, &newblk)
6569 		    == 0)
6570 			panic("softdep_setup_directory_add: lost entry");
6571 		WORKLIST_INSERT(&newblk->nb_newdirblk, &newdirblk->db_list);
6572 		pagedep->pd_state |= NEWBLOCK;
6573 		pagedep->pd_newdirblk = newdirblk;
6574 		newdirblk->db_pagedep = pagedep;
6575 		FREE_LOCK(&lk);
6576 		/*
6577 		 * If we extended into an indirect signal direnter to sync.
6578 		 */
6579 		if (isindir)
6580 			return (1);
6581 		return (0);
6582 	}
6583 	FREE_LOCK(&lk);
6584 	return (0);
6585 }
6586 
6587 /*
6588  * This procedure is called to change the offset of a directory
6589  * entry when compacting a directory block which must be owned
6590  * exclusively by the caller. Note that the actual entry movement
6591  * must be done in this procedure to ensure that no I/O completions
6592  * occur while the move is in progress.
6593  */
6594 void
6595 softdep_change_directoryentry_offset(bp, dp, base, oldloc, newloc, entrysize)
6596 	struct buf *bp;		/* Buffer holding directory block. */
6597 	struct inode *dp;	/* inode for directory */
6598 	caddr_t base;		/* address of dp->i_offset */
6599 	caddr_t oldloc;		/* address of old directory location */
6600 	caddr_t newloc;		/* address of new directory location */
6601 	int entrysize;		/* size of directory entry */
6602 {
6603 	int offset, oldoffset, newoffset;
6604 	struct pagedep *pagedep;
6605 	struct jmvref *jmvref;
6606 	struct diradd *dap;
6607 	struct direct *de;
6608 	struct mount *mp;
6609 	ufs_lbn_t lbn;
6610 	int flags;
6611 
6612 	mp = UFSTOVFS(dp->i_ump);
6613 	de = (struct direct *)oldloc;
6614 	jmvref = NULL;
6615 	flags = 0;
6616 	/*
6617 	 * Moves are always journaled as it would be too complex to
6618 	 * determine if any affected adds or removes are present in the
6619 	 * journal.
6620 	 */
6621 	if (mp->mnt_kern_flag & MNTK_SUJ)  {
6622 		flags = DEPALLOC;
6623 		jmvref = newjmvref(dp, de->d_ino,
6624 		    dp->i_offset + (oldloc - base),
6625 		    dp->i_offset + (newloc - base));
6626 	}
6627 	lbn = lblkno(dp->i_fs, dp->i_offset);
6628 	offset = blkoff(dp->i_fs, dp->i_offset);
6629 	oldoffset = offset + (oldloc - base);
6630 	newoffset = offset + (newloc - base);
6631 	ACQUIRE_LOCK(&lk);
6632 	if (pagedep_lookup(mp, dp->i_number, lbn, flags, &pagedep) == 0) {
6633 		if (pagedep)
6634 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
6635 		goto done;
6636 	}
6637 	dap = diradd_lookup(pagedep, oldoffset);
6638 	if (dap) {
6639 		dap->da_offset = newoffset;
6640 		newoffset = DIRADDHASH(newoffset);
6641 		oldoffset = DIRADDHASH(oldoffset);
6642 		if ((dap->da_state & ALLCOMPLETE) != ALLCOMPLETE &&
6643 		    newoffset != oldoffset) {
6644 			LIST_REMOVE(dap, da_pdlist);
6645 			LIST_INSERT_HEAD(&pagedep->pd_diraddhd[newoffset],
6646 			    dap, da_pdlist);
6647 		}
6648 	}
6649 done:
6650 	if (jmvref) {
6651 		jmvref->jm_pagedep = pagedep;
6652 		LIST_INSERT_HEAD(&pagedep->pd_jmvrefhd, jmvref, jm_deps);
6653 		add_to_journal(&jmvref->jm_list);
6654 	}
6655 	bcopy(oldloc, newloc, entrysize);
6656 	FREE_LOCK(&lk);
6657 }
6658 
6659 /*
6660  * Move the mkdir dependencies and journal work from one diradd to another
6661  * when renaming a directory.  The new name must depend on the mkdir deps
6662  * completing as the old name did.  Directories can only have one valid link
6663  * at a time so one must be canonical.
6664  */
6665 static void
6666 merge_diradd(inodedep, newdap)
6667 	struct inodedep *inodedep;
6668 	struct diradd *newdap;
6669 {
6670 	struct diradd *olddap;
6671 	struct mkdir *mkdir, *nextmd;
6672 	short state;
6673 
6674 	olddap = inodedep->id_mkdiradd;
6675 	inodedep->id_mkdiradd = newdap;
6676 	if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
6677 		newdap->da_state &= ~DEPCOMPLETE;
6678 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
6679 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
6680 			if (mkdir->md_diradd != olddap)
6681 				continue;
6682 			mkdir->md_diradd = newdap;
6683 			state = mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY);
6684 			newdap->da_state |= state;
6685 			olddap->da_state &= ~state;
6686 			if ((olddap->da_state &
6687 			    (MKDIR_PARENT | MKDIR_BODY)) == 0)
6688 				break;
6689 		}
6690 		if ((olddap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
6691 			panic("merge_diradd: unfound ref");
6692 	}
6693 	/*
6694 	 * Any mkdir related journal items are not safe to be freed until
6695 	 * the new name is stable.
6696 	 */
6697 	jwork_move(&newdap->da_jwork, &olddap->da_jwork);
6698 	olddap->da_state |= DEPCOMPLETE;
6699 	complete_diradd(olddap);
6700 }
6701 
6702 /*
6703  * Move the diradd to the pending list when all diradd dependencies are
6704  * complete.
6705  */
6706 static void
6707 complete_diradd(dap)
6708 	struct diradd *dap;
6709 {
6710 	struct pagedep *pagedep;
6711 
6712 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
6713 		if (dap->da_state & DIRCHG)
6714 			pagedep = dap->da_previous->dm_pagedep;
6715 		else
6716 			pagedep = dap->da_pagedep;
6717 		LIST_REMOVE(dap, da_pdlist);
6718 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
6719 	}
6720 }
6721 
6722 /*
6723  * Cancel a diradd when a dirrem overlaps with it.  We must cancel the journal
6724  * add entries and conditonally journal the remove.
6725  */
6726 static void
6727 cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref)
6728 	struct diradd *dap;
6729 	struct dirrem *dirrem;
6730 	struct jremref *jremref;
6731 	struct jremref *dotremref;
6732 	struct jremref *dotdotremref;
6733 {
6734 	struct inodedep *inodedep;
6735 	struct jaddref *jaddref;
6736 	struct inoref *inoref;
6737 	struct mkdir *mkdir;
6738 
6739 	/*
6740 	 * If no remove references were allocated we're on a non-journaled
6741 	 * filesystem and can skip the cancel step.
6742 	 */
6743 	if (jremref == NULL) {
6744 		free_diradd(dap, NULL);
6745 		return;
6746 	}
6747 	/*
6748 	 * Cancel the primary name an free it if it does not require
6749 	 * journaling.
6750 	 */
6751 	if (inodedep_lookup(dap->da_list.wk_mp, dap->da_newinum,
6752 	    0, &inodedep) != 0) {
6753 		/* Abort the addref that reference this diradd.  */
6754 		TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
6755 			if (inoref->if_list.wk_type != D_JADDREF)
6756 				continue;
6757 			jaddref = (struct jaddref *)inoref;
6758 			if (jaddref->ja_diradd != dap)
6759 				continue;
6760 			if (cancel_jaddref(jaddref, inodedep,
6761 			    &dirrem->dm_jwork) == 0) {
6762 				free_jremref(jremref);
6763 				jremref = NULL;
6764 			}
6765 			break;
6766 		}
6767 	}
6768 	/*
6769 	 * Cancel subordinate names and free them if they do not require
6770 	 * journaling.
6771 	 */
6772 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
6773 		LIST_FOREACH(mkdir, &mkdirlisthd, md_mkdirs) {
6774 			if (mkdir->md_diradd != dap)
6775 				continue;
6776 			if ((jaddref = mkdir->md_jaddref) == NULL)
6777 				continue;
6778 			mkdir->md_jaddref = NULL;
6779 			if (mkdir->md_state & MKDIR_PARENT) {
6780 				if (cancel_jaddref(jaddref, NULL,
6781 				    &dirrem->dm_jwork) == 0) {
6782 					free_jremref(dotdotremref);
6783 					dotdotremref = NULL;
6784 				}
6785 			} else {
6786 				if (cancel_jaddref(jaddref, inodedep,
6787 				    &dirrem->dm_jwork) == 0) {
6788 					free_jremref(dotremref);
6789 					dotremref = NULL;
6790 				}
6791 			}
6792 		}
6793 	}
6794 
6795 	if (jremref)
6796 		journal_jremref(dirrem, jremref, inodedep);
6797 	if (dotremref)
6798 		journal_jremref(dirrem, dotremref, inodedep);
6799 	if (dotdotremref)
6800 		journal_jremref(dirrem, dotdotremref, NULL);
6801 	jwork_move(&dirrem->dm_jwork, &dap->da_jwork);
6802 	free_diradd(dap, &dirrem->dm_jwork);
6803 }
6804 
6805 /*
6806  * Free a diradd dependency structure. This routine must be called
6807  * with splbio interrupts blocked.
6808  */
6809 static void
6810 free_diradd(dap, wkhd)
6811 	struct diradd *dap;
6812 	struct workhead *wkhd;
6813 {
6814 	struct dirrem *dirrem;
6815 	struct pagedep *pagedep;
6816 	struct inodedep *inodedep;
6817 	struct mkdir *mkdir, *nextmd;
6818 
6819 	mtx_assert(&lk, MA_OWNED);
6820 	LIST_REMOVE(dap, da_pdlist);
6821 	if (dap->da_state & ONWORKLIST)
6822 		WORKLIST_REMOVE(&dap->da_list);
6823 	if ((dap->da_state & DIRCHG) == 0) {
6824 		pagedep = dap->da_pagedep;
6825 	} else {
6826 		dirrem = dap->da_previous;
6827 		pagedep = dirrem->dm_pagedep;
6828 		dirrem->dm_dirinum = pagedep->pd_ino;
6829 		dirrem->dm_state |= COMPLETE;
6830 		if (LIST_EMPTY(&dirrem->dm_jremrefhd))
6831 			add_to_worklist(&dirrem->dm_list, 0);
6832 	}
6833 	if (inodedep_lookup(pagedep->pd_list.wk_mp, dap->da_newinum,
6834 	    0, &inodedep) != 0)
6835 		if (inodedep->id_mkdiradd == dap)
6836 			inodedep->id_mkdiradd = NULL;
6837 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
6838 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
6839 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
6840 			if (mkdir->md_diradd != dap)
6841 				continue;
6842 			dap->da_state &=
6843 			    ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY));
6844 			LIST_REMOVE(mkdir, md_mkdirs);
6845 			if (mkdir->md_state & ONWORKLIST)
6846 				WORKLIST_REMOVE(&mkdir->md_list);
6847 			if (mkdir->md_jaddref != NULL)
6848 				panic("free_diradd: Unexpected jaddref");
6849 			WORKITEM_FREE(mkdir, D_MKDIR);
6850 			if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
6851 				break;
6852 		}
6853 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
6854 			panic("free_diradd: unfound ref");
6855 	}
6856 	if (inodedep)
6857 		free_inodedep(inodedep);
6858 	/*
6859 	 * Free any journal segments waiting for the directory write.
6860 	 */
6861 	handle_jwork(&dap->da_jwork);
6862 	WORKITEM_FREE(dap, D_DIRADD);
6863 }
6864 
6865 /*
6866  * Directory entry removal dependencies.
6867  *
6868  * When removing a directory entry, the entry's inode pointer must be
6869  * zero'ed on disk before the corresponding inode's link count is decremented
6870  * (possibly freeing the inode for re-use). This dependency is handled by
6871  * updating the directory entry but delaying the inode count reduction until
6872  * after the directory block has been written to disk. After this point, the
6873  * inode count can be decremented whenever it is convenient.
6874  */
6875 
6876 /*
6877  * This routine should be called immediately after removing
6878  * a directory entry.  The inode's link count should not be
6879  * decremented by the calling procedure -- the soft updates
6880  * code will do this task when it is safe.
6881  */
6882 void
6883 softdep_setup_remove(bp, dp, ip, isrmdir)
6884 	struct buf *bp;		/* buffer containing directory block */
6885 	struct inode *dp;	/* inode for the directory being modified */
6886 	struct inode *ip;	/* inode for directory entry being removed */
6887 	int isrmdir;		/* indicates if doing RMDIR */
6888 {
6889 	struct dirrem *dirrem, *prevdirrem;
6890 	struct inodedep *inodedep;
6891 	int direct;
6892 
6893 	/*
6894 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.  We want
6895 	 * newdirrem() to setup the full directory remove which requires
6896 	 * isrmdir > 1.
6897 	 */
6898 	dirrem = newdirrem(bp, dp, ip, isrmdir?2:0, &prevdirrem);
6899 	/*
6900 	 * Add the dirrem to the inodedep's pending remove list for quick
6901 	 * discovery later.
6902 	 */
6903 	if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
6904 	    &inodedep) == 0)
6905 		panic("softdep_setup_remove: Lost inodedep.");
6906 	dirrem->dm_state |= ONDEPLIST;
6907 	LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
6908 
6909 	/*
6910 	 * If the COMPLETE flag is clear, then there were no active
6911 	 * entries and we want to roll back to a zeroed entry until
6912 	 * the new inode is committed to disk. If the COMPLETE flag is
6913 	 * set then we have deleted an entry that never made it to
6914 	 * disk. If the entry we deleted resulted from a name change,
6915 	 * then the old name still resides on disk. We cannot delete
6916 	 * its inode (returned to us in prevdirrem) until the zeroed
6917 	 * directory entry gets to disk. The new inode has never been
6918 	 * referenced on the disk, so can be deleted immediately.
6919 	 */
6920 	if ((dirrem->dm_state & COMPLETE) == 0) {
6921 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
6922 		    dm_next);
6923 		FREE_LOCK(&lk);
6924 	} else {
6925 		if (prevdirrem != NULL)
6926 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
6927 			    prevdirrem, dm_next);
6928 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
6929 		direct = LIST_EMPTY(&dirrem->dm_jremrefhd);
6930 		FREE_LOCK(&lk);
6931 		if (direct)
6932 			handle_workitem_remove(dirrem, NULL);
6933 	}
6934 }
6935 
6936 /*
6937  * Check for an entry matching 'offset' on both the pd_dirraddhd list and the
6938  * pd_pendinghd list of a pagedep.
6939  */
6940 static struct diradd *
6941 diradd_lookup(pagedep, offset)
6942 	struct pagedep *pagedep;
6943 	int offset;
6944 {
6945 	struct diradd *dap;
6946 
6947 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
6948 		if (dap->da_offset == offset)
6949 			return (dap);
6950 	LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
6951 		if (dap->da_offset == offset)
6952 			return (dap);
6953 	return (NULL);
6954 }
6955 
6956 /*
6957  * Search for a .. diradd dependency in a directory that is being removed.
6958  * If the directory was renamed to a new parent we have a diradd rather
6959  * than a mkdir for the .. entry.  We need to cancel it now before
6960  * it is found in truncate().
6961  */
6962 static struct jremref *
6963 cancel_diradd_dotdot(ip, dirrem, jremref)
6964 	struct inode *ip;
6965 	struct dirrem *dirrem;
6966 	struct jremref *jremref;
6967 {
6968 	struct pagedep *pagedep;
6969 	struct diradd *dap;
6970 	struct worklist *wk;
6971 
6972 	if (pagedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0, 0,
6973 	    &pagedep) == 0)
6974 		return (jremref);
6975 	dap = diradd_lookup(pagedep, DOTDOT_OFFSET);
6976 	if (dap == NULL)
6977 		return (jremref);
6978 	cancel_diradd(dap, dirrem, jremref, NULL, NULL);
6979 	/*
6980 	 * Mark any journal work as belonging to the parent so it is freed
6981 	 * with the .. reference.
6982 	 */
6983 	LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list)
6984 		wk->wk_state |= MKDIR_PARENT;
6985 	return (NULL);
6986 }
6987 
6988 /*
6989  * Cancel the MKDIR_PARENT mkdir component of a diradd when we're going to
6990  * replace it with a dirrem/diradd pair as a result of re-parenting a
6991  * directory.  This ensures that we don't simultaneously have a mkdir and
6992  * a diradd for the same .. entry.
6993  */
6994 static struct jremref *
6995 cancel_mkdir_dotdot(ip, dirrem, jremref)
6996 	struct inode *ip;
6997 	struct dirrem *dirrem;
6998 	struct jremref *jremref;
6999 {
7000 	struct inodedep *inodedep;
7001 	struct jaddref *jaddref;
7002 	struct mkdir *mkdir;
7003 	struct diradd *dap;
7004 
7005 	if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
7006 	    &inodedep) == 0)
7007 		panic("cancel_mkdir_dotdot: Lost inodedep");
7008 	dap = inodedep->id_mkdiradd;
7009 	if (dap == NULL || (dap->da_state & MKDIR_PARENT) == 0)
7010 		return (jremref);
7011 	for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir;
7012 	    mkdir = LIST_NEXT(mkdir, md_mkdirs))
7013 		if (mkdir->md_diradd == dap && mkdir->md_state & MKDIR_PARENT)
7014 			break;
7015 	if (mkdir == NULL)
7016 		panic("cancel_mkdir_dotdot: Unable to find mkdir\n");
7017 	if ((jaddref = mkdir->md_jaddref) != NULL) {
7018 		mkdir->md_jaddref = NULL;
7019 		jaddref->ja_state &= ~MKDIR_PARENT;
7020 		if (inodedep_lookup(UFSTOVFS(ip->i_ump), jaddref->ja_ino, 0,
7021 		    &inodedep) == 0)
7022 			panic("cancel_mkdir_dotdot: Lost parent inodedep");
7023 		if (cancel_jaddref(jaddref, inodedep, &dirrem->dm_jwork)) {
7024 			journal_jremref(dirrem, jremref, inodedep);
7025 			jremref = NULL;
7026 		}
7027 	}
7028 	if (mkdir->md_state & ONWORKLIST)
7029 		WORKLIST_REMOVE(&mkdir->md_list);
7030 	mkdir->md_state |= ALLCOMPLETE;
7031 	complete_mkdir(mkdir);
7032 	return (jremref);
7033 }
7034 
7035 static void
7036 journal_jremref(dirrem, jremref, inodedep)
7037 	struct dirrem *dirrem;
7038 	struct jremref *jremref;
7039 	struct inodedep *inodedep;
7040 {
7041 
7042 	if (inodedep == NULL)
7043 		if (inodedep_lookup(jremref->jr_list.wk_mp,
7044 		    jremref->jr_ref.if_ino, 0, &inodedep) == 0)
7045 			panic("journal_jremref: Lost inodedep");
7046 	LIST_INSERT_HEAD(&dirrem->dm_jremrefhd, jremref, jr_deps);
7047 	TAILQ_INSERT_TAIL(&inodedep->id_inoreflst, &jremref->jr_ref, if_deps);
7048 	add_to_journal(&jremref->jr_list);
7049 }
7050 
7051 static void
7052 dirrem_journal(dirrem, jremref, dotremref, dotdotremref)
7053 	struct dirrem *dirrem;
7054 	struct jremref *jremref;
7055 	struct jremref *dotremref;
7056 	struct jremref *dotdotremref;
7057 {
7058 	struct inodedep *inodedep;
7059 
7060 
7061 	if (inodedep_lookup(jremref->jr_list.wk_mp, jremref->jr_ref.if_ino, 0,
7062 	    &inodedep) == 0)
7063 		panic("dirrem_journal: Lost inodedep");
7064 	journal_jremref(dirrem, jremref, inodedep);
7065 	if (dotremref)
7066 		journal_jremref(dirrem, dotremref, inodedep);
7067 	if (dotdotremref)
7068 		journal_jremref(dirrem, dotdotremref, NULL);
7069 }
7070 
7071 /*
7072  * Allocate a new dirrem if appropriate and return it along with
7073  * its associated pagedep. Called without a lock, returns with lock.
7074  */
7075 static long num_dirrem;		/* number of dirrem allocated */
7076 static struct dirrem *
7077 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
7078 	struct buf *bp;		/* buffer containing directory block */
7079 	struct inode *dp;	/* inode for the directory being modified */
7080 	struct inode *ip;	/* inode for directory entry being removed */
7081 	int isrmdir;		/* indicates if doing RMDIR */
7082 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
7083 {
7084 	int offset;
7085 	ufs_lbn_t lbn;
7086 	struct diradd *dap;
7087 	struct dirrem *dirrem;
7088 	struct pagedep *pagedep;
7089 	struct jremref *jremref;
7090 	struct jremref *dotremref;
7091 	struct jremref *dotdotremref;
7092 	struct vnode *dvp;
7093 
7094 	/*
7095 	 * Whiteouts have no deletion dependencies.
7096 	 */
7097 	if (ip == NULL)
7098 		panic("newdirrem: whiteout");
7099 	dvp = ITOV(dp);
7100 	/*
7101 	 * If we are over our limit, try to improve the situation.
7102 	 * Limiting the number of dirrem structures will also limit
7103 	 * the number of freefile and freeblks structures.
7104 	 */
7105 	ACQUIRE_LOCK(&lk);
7106 	if (!(ip->i_flags & SF_SNAPSHOT) && num_dirrem > max_softdeps / 2)
7107 		(void) request_cleanup(ITOV(dp)->v_mount, FLUSH_REMOVE);
7108 	num_dirrem += 1;
7109 	FREE_LOCK(&lk);
7110 	dirrem = malloc(sizeof(struct dirrem),
7111 		M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO);
7112 	workitem_alloc(&dirrem->dm_list, D_DIRREM, dvp->v_mount);
7113 	LIST_INIT(&dirrem->dm_jremrefhd);
7114 	LIST_INIT(&dirrem->dm_jwork);
7115 	dirrem->dm_state = isrmdir ? RMDIR : 0;
7116 	dirrem->dm_oldinum = ip->i_number;
7117 	*prevdirremp = NULL;
7118 	/*
7119 	 * Allocate remove reference structures to track journal write
7120 	 * dependencies.  We will always have one for the link and
7121 	 * when doing directories we will always have one more for dot.
7122 	 * When renaming a directory we skip the dotdot link change so
7123 	 * this is not needed.
7124 	 */
7125 	jremref = dotremref = dotdotremref = NULL;
7126 	if (DOINGSUJ(dvp)) {
7127 		if (isrmdir) {
7128 			jremref = newjremref(dirrem, dp, ip, dp->i_offset,
7129 			    ip->i_effnlink + 2);
7130 			dotremref = newjremref(dirrem, ip, ip, DOT_OFFSET,
7131 			    ip->i_effnlink + 1);
7132 		} else
7133 			jremref = newjremref(dirrem, dp, ip, dp->i_offset,
7134 			    ip->i_effnlink + 1);
7135 		if (isrmdir > 1) {
7136 			dotdotremref = newjremref(dirrem, ip, dp, DOTDOT_OFFSET,
7137 			    dp->i_effnlink + 1);
7138 			dotdotremref->jr_state |= MKDIR_PARENT;
7139 		}
7140 	}
7141 	ACQUIRE_LOCK(&lk);
7142 	lbn = lblkno(dp->i_fs, dp->i_offset);
7143 	offset = blkoff(dp->i_fs, dp->i_offset);
7144 	if (pagedep_lookup(UFSTOVFS(dp->i_ump), dp->i_number, lbn, DEPALLOC,
7145 	    &pagedep) == 0)
7146 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
7147 	dirrem->dm_pagedep = pagedep;
7148 	/*
7149 	 * If we're renaming a .. link to a new directory, cancel any
7150 	 * existing MKDIR_PARENT mkdir.  If it has already been canceled
7151 	 * the jremref is preserved for any potential diradd in this
7152 	 * location.  This can not coincide with a rmdir.
7153 	 */
7154 	if (dp->i_offset == DOTDOT_OFFSET) {
7155 		if (isrmdir)
7156 			panic("newdirrem: .. directory change during remove?");
7157 		jremref = cancel_mkdir_dotdot(dp, dirrem, jremref);
7158 	}
7159 	/*
7160 	 * If we're removing a directory search for the .. dependency now and
7161 	 * cancel it.  Any pending journal work will be added to the dirrem
7162 	 * to be completed when the workitem remove completes.
7163 	 */
7164 	if (isrmdir > 1)
7165 		dotdotremref = cancel_diradd_dotdot(ip, dirrem, dotdotremref);
7166 	/*
7167 	 * Check for a diradd dependency for the same directory entry.
7168 	 * If present, then both dependencies become obsolete and can
7169 	 * be de-allocated.
7170 	 */
7171 	dap = diradd_lookup(pagedep, offset);
7172 	if (dap == NULL) {
7173 		/*
7174 		 * Link the jremref structures into the dirrem so they are
7175 		 * written prior to the pagedep.
7176 		 */
7177 		if (jremref)
7178 			dirrem_journal(dirrem, jremref, dotremref,
7179 			    dotdotremref);
7180 		return (dirrem);
7181 	}
7182 	/*
7183 	 * Must be ATTACHED at this point.
7184 	 */
7185 	if ((dap->da_state & ATTACHED) == 0)
7186 		panic("newdirrem: not ATTACHED");
7187 	if (dap->da_newinum != ip->i_number)
7188 		panic("newdirrem: inum %d should be %d",
7189 		    ip->i_number, dap->da_newinum);
7190 	/*
7191 	 * If we are deleting a changed name that never made it to disk,
7192 	 * then return the dirrem describing the previous inode (which
7193 	 * represents the inode currently referenced from this entry on disk).
7194 	 */
7195 	if ((dap->da_state & DIRCHG) != 0) {
7196 		*prevdirremp = dap->da_previous;
7197 		dap->da_state &= ~DIRCHG;
7198 		dap->da_pagedep = pagedep;
7199 	}
7200 	/*
7201 	 * We are deleting an entry that never made it to disk.
7202 	 * Mark it COMPLETE so we can delete its inode immediately.
7203 	 */
7204 	dirrem->dm_state |= COMPLETE;
7205 	cancel_diradd(dap, dirrem, jremref, dotremref, dotdotremref);
7206 #ifdef SUJ_DEBUG
7207 	if (isrmdir == 0) {
7208 		struct worklist *wk;
7209 
7210 		LIST_FOREACH(wk, &dirrem->dm_jwork, wk_list)
7211 			if (wk->wk_state & (MKDIR_BODY | MKDIR_PARENT))
7212 				panic("bad wk %p (0x%X)\n", wk, wk->wk_state);
7213 	}
7214 #endif
7215 
7216 	return (dirrem);
7217 }
7218 
7219 /*
7220  * Directory entry change dependencies.
7221  *
7222  * Changing an existing directory entry requires that an add operation
7223  * be completed first followed by a deletion. The semantics for the addition
7224  * are identical to the description of adding a new entry above except
7225  * that the rollback is to the old inode number rather than zero. Once
7226  * the addition dependency is completed, the removal is done as described
7227  * in the removal routine above.
7228  */
7229 
7230 /*
7231  * This routine should be called immediately after changing
7232  * a directory entry.  The inode's link count should not be
7233  * decremented by the calling procedure -- the soft updates
7234  * code will perform this task when it is safe.
7235  */
7236 void
7237 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
7238 	struct buf *bp;		/* buffer containing directory block */
7239 	struct inode *dp;	/* inode for the directory being modified */
7240 	struct inode *ip;	/* inode for directory entry being removed */
7241 	ino_t newinum;		/* new inode number for changed entry */
7242 	int isrmdir;		/* indicates if doing RMDIR */
7243 {
7244 	int offset;
7245 	struct diradd *dap = NULL;
7246 	struct dirrem *dirrem, *prevdirrem;
7247 	struct pagedep *pagedep;
7248 	struct inodedep *inodedep;
7249 	struct jaddref *jaddref;
7250 	struct mount *mp;
7251 
7252 	offset = blkoff(dp->i_fs, dp->i_offset);
7253 	mp = UFSTOVFS(dp->i_ump);
7254 
7255 	/*
7256 	 * Whiteouts do not need diradd dependencies.
7257 	 */
7258 	if (newinum != WINO) {
7259 		dap = malloc(sizeof(struct diradd),
7260 		    M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
7261 		workitem_alloc(&dap->da_list, D_DIRADD, mp);
7262 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
7263 		dap->da_offset = offset;
7264 		dap->da_newinum = newinum;
7265 		LIST_INIT(&dap->da_jwork);
7266 	}
7267 
7268 	/*
7269 	 * Allocate a new dirrem and ACQUIRE_LOCK.
7270 	 */
7271 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
7272 	pagedep = dirrem->dm_pagedep;
7273 	/*
7274 	 * The possible values for isrmdir:
7275 	 *	0 - non-directory file rename
7276 	 *	1 - directory rename within same directory
7277 	 *   inum - directory rename to new directory of given inode number
7278 	 * When renaming to a new directory, we are both deleting and
7279 	 * creating a new directory entry, so the link count on the new
7280 	 * directory should not change. Thus we do not need the followup
7281 	 * dirrem which is usually done in handle_workitem_remove. We set
7282 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
7283 	 * followup dirrem.
7284 	 */
7285 	if (isrmdir > 1)
7286 		dirrem->dm_state |= DIRCHG;
7287 
7288 	/*
7289 	 * Whiteouts have no additional dependencies,
7290 	 * so just put the dirrem on the correct list.
7291 	 */
7292 	if (newinum == WINO) {
7293 		if ((dirrem->dm_state & COMPLETE) == 0) {
7294 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
7295 			    dm_next);
7296 		} else {
7297 			dirrem->dm_dirinum = pagedep->pd_ino;
7298 			if (LIST_EMPTY(&dirrem->dm_jremrefhd))
7299 				add_to_worklist(&dirrem->dm_list, 0);
7300 		}
7301 		FREE_LOCK(&lk);
7302 		return;
7303 	}
7304 	/*
7305 	 * Add the dirrem to the inodedep's pending remove list for quick
7306 	 * discovery later.  A valid nlinkdelta ensures that this lookup
7307 	 * will not fail.
7308 	 */
7309 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
7310 		panic("softdep_setup_directory_change: Lost inodedep.");
7311 	dirrem->dm_state |= ONDEPLIST;
7312 	LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
7313 
7314 	/*
7315 	 * If the COMPLETE flag is clear, then there were no active
7316 	 * entries and we want to roll back to the previous inode until
7317 	 * the new inode is committed to disk. If the COMPLETE flag is
7318 	 * set, then we have deleted an entry that never made it to disk.
7319 	 * If the entry we deleted resulted from a name change, then the old
7320 	 * inode reference still resides on disk. Any rollback that we do
7321 	 * needs to be to that old inode (returned to us in prevdirrem). If
7322 	 * the entry we deleted resulted from a create, then there is
7323 	 * no entry on the disk, so we want to roll back to zero rather
7324 	 * than the uncommitted inode. In either of the COMPLETE cases we
7325 	 * want to immediately free the unwritten and unreferenced inode.
7326 	 */
7327 	if ((dirrem->dm_state & COMPLETE) == 0) {
7328 		dap->da_previous = dirrem;
7329 	} else {
7330 		if (prevdirrem != NULL) {
7331 			dap->da_previous = prevdirrem;
7332 		} else {
7333 			dap->da_state &= ~DIRCHG;
7334 			dap->da_pagedep = pagedep;
7335 		}
7336 		dirrem->dm_dirinum = pagedep->pd_ino;
7337 		if (LIST_EMPTY(&dirrem->dm_jremrefhd))
7338 			add_to_worklist(&dirrem->dm_list, 0);
7339 	}
7340 	/*
7341 	 * Lookup the jaddref for this journal entry.  We must finish
7342 	 * initializing it and make the diradd write dependent on it.
7343 	 * If we're not journaling Put it on the id_bufwait list if the inode
7344 	 * is not yet written. If it is written, do the post-inode write
7345 	 * processing to put it on the id_pendinghd list.
7346 	 */
7347 	inodedep_lookup(mp, newinum, DEPALLOC, &inodedep);
7348 	if (mp->mnt_kern_flag & MNTK_SUJ) {
7349 		jaddref = (struct jaddref *)TAILQ_LAST(&inodedep->id_inoreflst,
7350 		    inoreflst);
7351 		KASSERT(jaddref != NULL && jaddref->ja_parent == dp->i_number,
7352 		    ("softdep_setup_directory_change: bad jaddref %p",
7353 		    jaddref));
7354 		jaddref->ja_diroff = dp->i_offset;
7355 		jaddref->ja_diradd = dap;
7356 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
7357 		    dap, da_pdlist);
7358 		add_to_journal(&jaddref->ja_list);
7359 	} else if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
7360 		dap->da_state |= COMPLETE;
7361 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
7362 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
7363 	} else {
7364 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
7365 		    dap, da_pdlist);
7366 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
7367 	}
7368 	/*
7369 	 * If we're making a new name for a directory that has not been
7370 	 * committed when need to move the dot and dotdot references to
7371 	 * this new name.
7372 	 */
7373 	if (inodedep->id_mkdiradd && dp->i_offset != DOTDOT_OFFSET)
7374 		merge_diradd(inodedep, dap);
7375 	FREE_LOCK(&lk);
7376 }
7377 
7378 /*
7379  * Called whenever the link count on an inode is changed.
7380  * It creates an inode dependency so that the new reference(s)
7381  * to the inode cannot be committed to disk until the updated
7382  * inode has been written.
7383  */
7384 void
7385 softdep_change_linkcnt(ip)
7386 	struct inode *ip;	/* the inode with the increased link count */
7387 {
7388 	struct inodedep *inodedep;
7389 
7390 	ACQUIRE_LOCK(&lk);
7391 	inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, DEPALLOC, &inodedep);
7392 	if (ip->i_nlink < ip->i_effnlink)
7393 		panic("softdep_change_linkcnt: bad delta");
7394 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
7395 	FREE_LOCK(&lk);
7396 }
7397 
7398 /*
7399  * Attach a sbdep dependency to the superblock buf so that we can keep
7400  * track of the head of the linked list of referenced but unlinked inodes.
7401  */
7402 void
7403 softdep_setup_sbupdate(ump, fs, bp)
7404 	struct ufsmount *ump;
7405 	struct fs *fs;
7406 	struct buf *bp;
7407 {
7408 	struct sbdep *sbdep;
7409 	struct worklist *wk;
7410 
7411 	if ((fs->fs_flags & FS_SUJ) == 0)
7412 		return;
7413 	LIST_FOREACH(wk, &bp->b_dep, wk_list)
7414 		if (wk->wk_type == D_SBDEP)
7415 			break;
7416 	if (wk != NULL)
7417 		return;
7418 	sbdep = malloc(sizeof(struct sbdep), M_SBDEP, M_SOFTDEP_FLAGS);
7419 	workitem_alloc(&sbdep->sb_list, D_SBDEP, UFSTOVFS(ump));
7420 	sbdep->sb_fs = fs;
7421 	sbdep->sb_ump = ump;
7422 	ACQUIRE_LOCK(&lk);
7423 	WORKLIST_INSERT(&bp->b_dep, &sbdep->sb_list);
7424 	FREE_LOCK(&lk);
7425 }
7426 
7427 /*
7428  * Return the first unlinked inodedep which is ready to be the head of the
7429  * list.  The inodedep and all those after it must have valid next pointers.
7430  */
7431 static struct inodedep *
7432 first_unlinked_inodedep(ump)
7433 	struct ufsmount *ump;
7434 {
7435 	struct inodedep *inodedep;
7436 	struct inodedep *idp;
7437 
7438 	for (inodedep = TAILQ_LAST(&ump->softdep_unlinked, inodedeplst);
7439 	    inodedep; inodedep = idp) {
7440 		if ((inodedep->id_state & UNLINKNEXT) == 0)
7441 			return (NULL);
7442 		idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
7443 		if (idp == NULL || (idp->id_state & UNLINKNEXT) == 0)
7444 			break;
7445 		if ((inodedep->id_state & UNLINKPREV) == 0)
7446 			panic("first_unlinked_inodedep: prev != next");
7447 	}
7448 	if (inodedep == NULL)
7449 		return (NULL);
7450 
7451 	return (inodedep);
7452 }
7453 
7454 /*
7455  * Set the sujfree unlinked head pointer prior to writing a superblock.
7456  */
7457 static void
7458 initiate_write_sbdep(sbdep)
7459 	struct sbdep *sbdep;
7460 {
7461 	struct inodedep *inodedep;
7462 	struct fs *bpfs;
7463 	struct fs *fs;
7464 
7465 	bpfs = sbdep->sb_fs;
7466 	fs = sbdep->sb_ump->um_fs;
7467 	inodedep = first_unlinked_inodedep(sbdep->sb_ump);
7468 	if (inodedep) {
7469 		fs->fs_sujfree = inodedep->id_ino;
7470 		inodedep->id_state |= UNLINKPREV;
7471 	} else
7472 		fs->fs_sujfree = 0;
7473 	bpfs->fs_sujfree = fs->fs_sujfree;
7474 }
7475 
7476 /*
7477  * After a superblock is written determine whether it must be written again
7478  * due to a changing unlinked list head.
7479  */
7480 static int
7481 handle_written_sbdep(sbdep, bp)
7482 	struct sbdep *sbdep;
7483 	struct buf *bp;
7484 {
7485 	struct inodedep *inodedep;
7486 	struct mount *mp;
7487 	struct fs *fs;
7488 
7489 	fs = sbdep->sb_fs;
7490 	mp = UFSTOVFS(sbdep->sb_ump);
7491 	inodedep = first_unlinked_inodedep(sbdep->sb_ump);
7492 	if ((inodedep && fs->fs_sujfree != inodedep->id_ino) ||
7493 	    (inodedep == NULL && fs->fs_sujfree != 0)) {
7494 		bdirty(bp);
7495 		return (1);
7496 	}
7497 	WORKITEM_FREE(sbdep, D_SBDEP);
7498 	if (fs->fs_sujfree == 0)
7499 		return (0);
7500 	if (inodedep_lookup(mp, fs->fs_sujfree, 0, &inodedep) == 0)
7501 		panic("handle_written_sbdep: lost inodedep");
7502 	/*
7503 	 * Now that we have a record of this inode in stable store allow it
7504 	 * to be written to free up pending work.  Inodes may see a lot of
7505 	 * write activity after they are unlinked which we must not hold up.
7506 	 */
7507 	for (; inodedep != NULL; inodedep = TAILQ_NEXT(inodedep, id_unlinked)) {
7508 		if ((inodedep->id_state & UNLINKLINKS) != UNLINKLINKS)
7509 			panic("handle_written_sbdep: Bad inodedep %p (0x%X)",
7510 			    inodedep, inodedep->id_state);
7511 		if (inodedep->id_state & UNLINKONLIST)
7512 			break;
7513 		inodedep->id_state |= DEPCOMPLETE | UNLINKONLIST;
7514 	}
7515 
7516 	return (0);
7517 }
7518 
7519 /*
7520  * Mark an inodedep as unlinked and insert it into the in-memory unlinked list.
7521  */
7522 static void
7523 unlinked_inodedep(mp, inodedep)
7524 	struct mount *mp;
7525 	struct inodedep *inodedep;
7526 {
7527 	struct ufsmount *ump;
7528 
7529 	if ((mp->mnt_kern_flag & MNTK_SUJ) == 0)
7530 		return;
7531 	ump = VFSTOUFS(mp);
7532 	ump->um_fs->fs_fmod = 1;
7533 	inodedep->id_state |= UNLINKED;
7534 	TAILQ_INSERT_HEAD(&ump->softdep_unlinked, inodedep, id_unlinked);
7535 }
7536 
7537 /*
7538  * Remove an inodedep from the unlinked inodedep list.  This may require
7539  * disk writes if the inode has made it that far.
7540  */
7541 static void
7542 clear_unlinked_inodedep(inodedep)
7543 	struct inodedep *inodedep;
7544 {
7545 	struct ufsmount *ump;
7546 	struct inodedep *idp;
7547 	struct inodedep *idn;
7548 	struct fs *fs;
7549 	struct buf *bp;
7550 	ino_t ino;
7551 	ino_t nino;
7552 	ino_t pino;
7553 	int error;
7554 
7555 	ump = VFSTOUFS(inodedep->id_list.wk_mp);
7556 	fs = ump->um_fs;
7557 	ino = inodedep->id_ino;
7558 	error = 0;
7559 	for (;;) {
7560 		/*
7561 		 * If nothing has yet been written simply remove us from
7562 		 * the in memory list and return.  This is the most common
7563 		 * case where handle_workitem_remove() loses the final
7564 		 * reference.
7565 		 */
7566 		if ((inodedep->id_state & UNLINKLINKS) == 0)
7567 			break;
7568 		/*
7569 		 * If we have a NEXT pointer and no PREV pointer we can simply
7570 		 * clear NEXT's PREV and remove ourselves from the list.  Be
7571 		 * careful not to clear PREV if the superblock points at
7572 		 * next as well.
7573 		 */
7574 		idn = TAILQ_NEXT(inodedep, id_unlinked);
7575 		if ((inodedep->id_state & UNLINKLINKS) == UNLINKNEXT) {
7576 			if (idn && fs->fs_sujfree != idn->id_ino)
7577 				idn->id_state &= ~UNLINKPREV;
7578 			break;
7579 		}
7580 		/*
7581 		 * Here we have an inodedep which is actually linked into
7582 		 * the list.  We must remove it by forcing a write to the
7583 		 * link before us, whether it be the superblock or an inode.
7584 		 * Unfortunately the list may change while we're waiting
7585 		 * on the buf lock for either resource so we must loop until
7586 		 * we lock the right one.  If both the superblock and an
7587 		 * inode point to this inode we must clear the inode first
7588 		 * followed by the superblock.
7589 		 */
7590 		idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
7591 		pino = 0;
7592 		if (idp && (idp->id_state & UNLINKNEXT))
7593 			pino = idp->id_ino;
7594 		FREE_LOCK(&lk);
7595 		if (pino == 0)
7596 			bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc),
7597 			    (int)fs->fs_sbsize, 0, 0, 0);
7598 		else
7599 			error = bread(ump->um_devvp,
7600 			    fsbtodb(fs, ino_to_fsba(fs, pino)),
7601 			    (int)fs->fs_bsize, NOCRED, &bp);
7602 		ACQUIRE_LOCK(&lk);
7603 		if (error)
7604 			break;
7605 		/* If the list has changed restart the loop. */
7606 		idp = TAILQ_PREV(inodedep, inodedeplst, id_unlinked);
7607 		nino = 0;
7608 		if (idp && (idp->id_state & UNLINKNEXT))
7609 			nino = idp->id_ino;
7610 		if (nino != pino ||
7611 		    (inodedep->id_state & UNLINKPREV) != UNLINKPREV) {
7612 			FREE_LOCK(&lk);
7613 			brelse(bp);
7614 			ACQUIRE_LOCK(&lk);
7615 			continue;
7616 		}
7617 		/*
7618 		 * Remove us from the in memory list.  After this we cannot
7619 		 * access the inodedep.
7620 		 */
7621 		idn = TAILQ_NEXT(inodedep, id_unlinked);
7622 		inodedep->id_state &= ~(UNLINKED | UNLINKLINKS);
7623 		TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked);
7624 		/*
7625 		 * Determine the next inode number.
7626 		 */
7627 		nino = 0;
7628 		if (idn) {
7629 			/*
7630 			 * If next isn't on the list we can just clear prev's
7631 			 * state and schedule it to be fixed later.  No need
7632 			 * to synchronously write if we're not in the real
7633 			 * list.
7634 			 */
7635 			if ((idn->id_state & UNLINKPREV) == 0 && pino != 0) {
7636 				idp->id_state &= ~UNLINKNEXT;
7637 				if ((idp->id_state & ONWORKLIST) == 0)
7638 					WORKLIST_INSERT(&bp->b_dep,
7639 					    &idp->id_list);
7640 				FREE_LOCK(&lk);
7641 				bawrite(bp);
7642 				ACQUIRE_LOCK(&lk);
7643 				return;
7644 			}
7645 			nino = idn->id_ino;
7646 		}
7647 		FREE_LOCK(&lk);
7648 		/*
7649 		 * The predecessor's next pointer is manually updated here
7650 		 * so that the NEXT flag is never cleared for an element
7651 		 * that is in the list.
7652 		 */
7653 		if (pino == 0) {
7654 			bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
7655 			ffs_oldfscompat_write((struct fs *)bp->b_data, ump);
7656 			softdep_setup_sbupdate(ump, (struct fs *)bp->b_data,
7657 			    bp);
7658 		} else if (fs->fs_magic == FS_UFS1_MAGIC)
7659 			((struct ufs1_dinode *)bp->b_data +
7660 			    ino_to_fsbo(fs, pino))->di_freelink = nino;
7661 		else
7662 			((struct ufs2_dinode *)bp->b_data +
7663 			    ino_to_fsbo(fs, pino))->di_freelink = nino;
7664 		/*
7665 		 * If the bwrite fails we have no recourse to recover.  The
7666 		 * filesystem is corrupted already.
7667 		 */
7668 		bwrite(bp);
7669 		ACQUIRE_LOCK(&lk);
7670 		/*
7671 		 * If the superblock pointer still needs to be cleared force
7672 		 * a write here.
7673 		 */
7674 		if (fs->fs_sujfree == ino) {
7675 			FREE_LOCK(&lk);
7676 			bp = getblk(ump->um_devvp, btodb(fs->fs_sblockloc),
7677 			    (int)fs->fs_sbsize, 0, 0, 0);
7678 			bcopy((caddr_t)fs, bp->b_data, (u_int)fs->fs_sbsize);
7679 			ffs_oldfscompat_write((struct fs *)bp->b_data, ump);
7680 			softdep_setup_sbupdate(ump, (struct fs *)bp->b_data,
7681 			    bp);
7682 			bwrite(bp);
7683 			ACQUIRE_LOCK(&lk);
7684 		}
7685 		if (fs->fs_sujfree != ino)
7686 			return;
7687 		panic("clear_unlinked_inodedep: Failed to clear free head");
7688 	}
7689 	if (inodedep->id_ino == fs->fs_sujfree)
7690 		panic("clear_unlinked_inodedep: Freeing head of free list");
7691 	inodedep->id_state &= ~(UNLINKED | UNLINKLINKS);
7692 	TAILQ_REMOVE(&ump->softdep_unlinked, inodedep, id_unlinked);
7693 	return;
7694 }
7695 
7696 /*
7697  * This workitem decrements the inode's link count.
7698  * If the link count reaches zero, the file is removed.
7699  */
7700 static void
7701 handle_workitem_remove(dirrem, xp)
7702 	struct dirrem *dirrem;
7703 	struct vnode *xp;
7704 {
7705 	struct inodedep *inodedep;
7706 	struct workhead dotdotwk;
7707 	struct worklist *wk;
7708 	struct ufsmount *ump;
7709 	struct mount *mp;
7710 	struct vnode *vp;
7711 	struct inode *ip;
7712 	ino_t oldinum;
7713 	int error;
7714 
7715 	if (dirrem->dm_state & ONWORKLIST)
7716 		panic("handle_workitem_remove: dirrem %p still on worklist",
7717 		    dirrem);
7718 	oldinum = dirrem->dm_oldinum;
7719 	mp = dirrem->dm_list.wk_mp;
7720 	ump = VFSTOUFS(mp);
7721 	if ((vp = xp) == NULL &&
7722 	    (error = ffs_vgetf(mp, oldinum, LK_EXCLUSIVE, &vp,
7723 	    FFSV_FORCEINSMQ)) != 0) {
7724 		softdep_error("handle_workitem_remove: vget", error);
7725 		return;
7726 	}
7727 	ip = VTOI(vp);
7728 	ACQUIRE_LOCK(&lk);
7729 	if ((inodedep_lookup(mp, oldinum, 0, &inodedep)) == 0)
7730 		panic("handle_workitem_remove: lost inodedep");
7731 	if (dirrem->dm_state & ONDEPLIST)
7732 		LIST_REMOVE(dirrem, dm_inonext);
7733 	KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd),
7734 	    ("handle_workitem_remove:  Journal entries not written."));
7735 
7736 	/*
7737 	 * Move all dependencies waiting on the remove to complete
7738 	 * from the dirrem to the inode inowait list to be completed
7739 	 * after the inode has been updated and written to disk.  Any
7740 	 * marked MKDIR_PARENT are saved to be completed when the .. ref
7741 	 * is removed.
7742 	 */
7743 	LIST_INIT(&dotdotwk);
7744 	while ((wk = LIST_FIRST(&dirrem->dm_jwork)) != NULL) {
7745 		WORKLIST_REMOVE(wk);
7746 		if (wk->wk_state & MKDIR_PARENT) {
7747 			wk->wk_state &= ~MKDIR_PARENT;
7748 			WORKLIST_INSERT(&dotdotwk, wk);
7749 			continue;
7750 		}
7751 		WORKLIST_INSERT(&inodedep->id_inowait, wk);
7752 	}
7753 	LIST_SWAP(&dirrem->dm_jwork, &dotdotwk, worklist, wk_list);
7754 	/*
7755 	 * Normal file deletion.
7756 	 */
7757 	if ((dirrem->dm_state & RMDIR) == 0) {
7758 		ip->i_nlink--;
7759 		DIP_SET(ip, i_nlink, ip->i_nlink);
7760 		ip->i_flag |= IN_CHANGE;
7761 		if (ip->i_nlink < ip->i_effnlink)
7762 			panic("handle_workitem_remove: bad file delta");
7763 		if (ip->i_nlink == 0)
7764 			unlinked_inodedep(mp, inodedep);
7765 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
7766 		num_dirrem -= 1;
7767 		KASSERT(LIST_EMPTY(&dirrem->dm_jwork),
7768 		    ("handle_workitem_remove: worklist not empty. %s",
7769 		    TYPENAME(LIST_FIRST(&dirrem->dm_jwork)->wk_type)));
7770 		WORKITEM_FREE(dirrem, D_DIRREM);
7771 		FREE_LOCK(&lk);
7772 		goto out;
7773 	}
7774 	/*
7775 	 * Directory deletion. Decrement reference count for both the
7776 	 * just deleted parent directory entry and the reference for ".".
7777 	 * Arrange to have the reference count on the parent decremented
7778 	 * to account for the loss of "..".
7779 	 */
7780 	ip->i_nlink -= 2;
7781 	DIP_SET(ip, i_nlink, ip->i_nlink);
7782 	ip->i_flag |= IN_CHANGE;
7783 	if (ip->i_nlink < ip->i_effnlink)
7784 		panic("handle_workitem_remove: bad dir delta");
7785 	if (ip->i_nlink == 0)
7786 		unlinked_inodedep(mp, inodedep);
7787 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
7788 	/*
7789 	 * Rename a directory to a new parent. Since, we are both deleting
7790 	 * and creating a new directory entry, the link count on the new
7791 	 * directory should not change. Thus we skip the followup dirrem.
7792 	 */
7793 	if (dirrem->dm_state & DIRCHG) {
7794 		KASSERT(LIST_EMPTY(&dirrem->dm_jwork),
7795 		    ("handle_workitem_remove: DIRCHG and worklist not empty."));
7796 		num_dirrem -= 1;
7797 		WORKITEM_FREE(dirrem, D_DIRREM);
7798 		FREE_LOCK(&lk);
7799 		goto out;
7800 	}
7801 	dirrem->dm_state = ONDEPLIST;
7802 	dirrem->dm_oldinum = dirrem->dm_dirinum;
7803 	/*
7804 	 * Place the dirrem on the parent's diremhd list.
7805 	 */
7806 	if (inodedep_lookup(mp, dirrem->dm_oldinum, 0, &inodedep) == 0)
7807 		panic("handle_workitem_remove: lost dir inodedep");
7808 	LIST_INSERT_HEAD(&inodedep->id_dirremhd, dirrem, dm_inonext);
7809 	/*
7810 	 * If the allocated inode has never been written to disk, then
7811 	 * the on-disk inode is zero'ed and we can remove the file
7812 	 * immediately.  When journaling if the inode has been marked
7813 	 * unlinked and not DEPCOMPLETE we know it can never be written.
7814 	 */
7815 	inodedep_lookup(mp, oldinum, 0, &inodedep);
7816 	if (inodedep == NULL ||
7817 	    (inodedep->id_state & (DEPCOMPLETE | UNLINKED)) == UNLINKED ||
7818 	    check_inode_unwritten(inodedep)) {
7819 		if (xp != NULL)
7820 			add_to_worklist(&dirrem->dm_list, 0);
7821 		FREE_LOCK(&lk);
7822 		if (xp == NULL) {
7823 			vput(vp);
7824 			handle_workitem_remove(dirrem, NULL);
7825 		}
7826 		return;
7827 	}
7828 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
7829 	FREE_LOCK(&lk);
7830 	ip->i_flag |= IN_CHANGE;
7831 out:
7832 	ffs_update(vp, 0);
7833 	if (xp == NULL)
7834 		vput(vp);
7835 }
7836 
7837 /*
7838  * Inode de-allocation dependencies.
7839  *
7840  * When an inode's link count is reduced to zero, it can be de-allocated. We
7841  * found it convenient to postpone de-allocation until after the inode is
7842  * written to disk with its new link count (zero).  At this point, all of the
7843  * on-disk inode's block pointers are nullified and, with careful dependency
7844  * list ordering, all dependencies related to the inode will be satisfied and
7845  * the corresponding dependency structures de-allocated.  So, if/when the
7846  * inode is reused, there will be no mixing of old dependencies with new
7847  * ones.  This artificial dependency is set up by the block de-allocation
7848  * procedure above (softdep_setup_freeblocks) and completed by the
7849  * following procedure.
7850  */
7851 static void
7852 handle_workitem_freefile(freefile)
7853 	struct freefile *freefile;
7854 {
7855 	struct workhead wkhd;
7856 	struct fs *fs;
7857 	struct inodedep *idp;
7858 	struct ufsmount *ump;
7859 	int error;
7860 
7861 	ump = VFSTOUFS(freefile->fx_list.wk_mp);
7862 	fs = ump->um_fs;
7863 #ifdef DEBUG
7864 	ACQUIRE_LOCK(&lk);
7865 	error = inodedep_lookup(UFSTOVFS(ump), freefile->fx_oldinum, 0, &idp);
7866 	FREE_LOCK(&lk);
7867 	if (error)
7868 		panic("handle_workitem_freefile: inodedep %p survived", idp);
7869 #endif
7870 	UFS_LOCK(ump);
7871 	fs->fs_pendinginodes -= 1;
7872 	UFS_UNLOCK(ump);
7873 	LIST_INIT(&wkhd);
7874 	LIST_SWAP(&freefile->fx_jwork, &wkhd, worklist, wk_list);
7875 	if ((error = ffs_freefile(ump, fs, freefile->fx_devvp,
7876 	    freefile->fx_oldinum, freefile->fx_mode, &wkhd)) != 0)
7877 		softdep_error("handle_workitem_freefile", error);
7878 	ACQUIRE_LOCK(&lk);
7879 	WORKITEM_FREE(freefile, D_FREEFILE);
7880 	FREE_LOCK(&lk);
7881 }
7882 
7883 
7884 /*
7885  * Helper function which unlinks marker element from work list and returns
7886  * the next element on the list.
7887  */
7888 static __inline struct worklist *
7889 markernext(struct worklist *marker)
7890 {
7891 	struct worklist *next;
7892 
7893 	next = LIST_NEXT(marker, wk_list);
7894 	LIST_REMOVE(marker, wk_list);
7895 	return next;
7896 }
7897 
7898 /*
7899  * Disk writes.
7900  *
7901  * The dependency structures constructed above are most actively used when file
7902  * system blocks are written to disk.  No constraints are placed on when a
7903  * block can be written, but unsatisfied update dependencies are made safe by
7904  * modifying (or replacing) the source memory for the duration of the disk
7905  * write.  When the disk write completes, the memory block is again brought
7906  * up-to-date.
7907  *
7908  * In-core inode structure reclamation.
7909  *
7910  * Because there are a finite number of "in-core" inode structures, they are
7911  * reused regularly.  By transferring all inode-related dependencies to the
7912  * in-memory inode block and indexing them separately (via "inodedep"s), we
7913  * can allow "in-core" inode structures to be reused at any time and avoid
7914  * any increase in contention.
7915  *
7916  * Called just before entering the device driver to initiate a new disk I/O.
7917  * The buffer must be locked, thus, no I/O completion operations can occur
7918  * while we are manipulating its associated dependencies.
7919  */
7920 static void
7921 softdep_disk_io_initiation(bp)
7922 	struct buf *bp;		/* structure describing disk write to occur */
7923 {
7924 	struct worklist *wk;
7925 	struct worklist marker;
7926 	struct inodedep *inodedep;
7927 	struct freeblks *freeblks;
7928 	struct jfreeblk *jfreeblk;
7929 	struct newblk *newblk;
7930 
7931 	/*
7932 	 * We only care about write operations. There should never
7933 	 * be dependencies for reads.
7934 	 */
7935 	if (bp->b_iocmd != BIO_WRITE)
7936 		panic("softdep_disk_io_initiation: not write");
7937 
7938 	if (bp->b_vflags & BV_BKGRDINPROG)
7939 		panic("softdep_disk_io_initiation: Writing buffer with "
7940 		    "background write in progress: %p", bp);
7941 
7942 	marker.wk_type = D_LAST + 1;	/* Not a normal workitem */
7943 	PHOLD(curproc);			/* Don't swap out kernel stack */
7944 
7945 	ACQUIRE_LOCK(&lk);
7946 	/*
7947 	 * Do any necessary pre-I/O processing.
7948 	 */
7949 	for (wk = LIST_FIRST(&bp->b_dep); wk != NULL;
7950 	     wk = markernext(&marker)) {
7951 		LIST_INSERT_AFTER(wk, &marker, wk_list);
7952 		switch (wk->wk_type) {
7953 
7954 		case D_PAGEDEP:
7955 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
7956 			continue;
7957 
7958 		case D_INODEDEP:
7959 			inodedep = WK_INODEDEP(wk);
7960 			if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
7961 				initiate_write_inodeblock_ufs1(inodedep, bp);
7962 			else
7963 				initiate_write_inodeblock_ufs2(inodedep, bp);
7964 			continue;
7965 
7966 		case D_INDIRDEP:
7967 			initiate_write_indirdep(WK_INDIRDEP(wk), bp);
7968 			continue;
7969 
7970 		case D_BMSAFEMAP:
7971 			initiate_write_bmsafemap(WK_BMSAFEMAP(wk), bp);
7972 			continue;
7973 
7974 		case D_JSEG:
7975 			WK_JSEG(wk)->js_buf = NULL;
7976 			continue;
7977 
7978 		case D_FREEBLKS:
7979 			freeblks = WK_FREEBLKS(wk);
7980 			jfreeblk = LIST_FIRST(&freeblks->fb_jfreeblkhd);
7981 			/*
7982 			 * We have to wait for the jfreeblks to be journaled
7983 			 * before we can write an inodeblock with updated
7984 			 * pointers.  Be careful to arrange the marker so
7985 			 * we revisit the jfreeblk if it's not removed by
7986 			 * the first jwait().
7987 			 */
7988 			if (jfreeblk != NULL) {
7989 				LIST_REMOVE(&marker, wk_list);
7990 				LIST_INSERT_BEFORE(wk, &marker, wk_list);
7991 				jwait(&jfreeblk->jf_list);
7992 			}
7993 			continue;
7994 		case D_ALLOCDIRECT:
7995 		case D_ALLOCINDIR:
7996 			/*
7997 			 * We have to wait for the jnewblk to be journaled
7998 			 * before we can write to a block otherwise the
7999 			 * contents may be confused with an earlier file
8000 			 * at recovery time.  Handle the marker as described
8001 			 * above.
8002 			 */
8003 			newblk = WK_NEWBLK(wk);
8004 			if (newblk->nb_jnewblk != NULL) {
8005 				LIST_REMOVE(&marker, wk_list);
8006 				LIST_INSERT_BEFORE(wk, &marker, wk_list);
8007 				jwait(&newblk->nb_jnewblk->jn_list);
8008 			}
8009 			continue;
8010 
8011 		case D_SBDEP:
8012 			initiate_write_sbdep(WK_SBDEP(wk));
8013 			continue;
8014 
8015 		case D_MKDIR:
8016 		case D_FREEWORK:
8017 		case D_FREEDEP:
8018 		case D_JSEGDEP:
8019 			continue;
8020 
8021 		default:
8022 			panic("handle_disk_io_initiation: Unexpected type %s",
8023 			    TYPENAME(wk->wk_type));
8024 			/* NOTREACHED */
8025 		}
8026 	}
8027 	FREE_LOCK(&lk);
8028 	PRELE(curproc);			/* Allow swapout of kernel stack */
8029 }
8030 
8031 /*
8032  * Called from within the procedure above to deal with unsatisfied
8033  * allocation dependencies in a directory. The buffer must be locked,
8034  * thus, no I/O completion operations can occur while we are
8035  * manipulating its associated dependencies.
8036  */
8037 static void
8038 initiate_write_filepage(pagedep, bp)
8039 	struct pagedep *pagedep;
8040 	struct buf *bp;
8041 {
8042 	struct jremref *jremref;
8043 	struct jmvref *jmvref;
8044 	struct dirrem *dirrem;
8045 	struct diradd *dap;
8046 	struct direct *ep;
8047 	int i;
8048 
8049 	if (pagedep->pd_state & IOSTARTED) {
8050 		/*
8051 		 * This can only happen if there is a driver that does not
8052 		 * understand chaining. Here biodone will reissue the call
8053 		 * to strategy for the incomplete buffers.
8054 		 */
8055 		printf("initiate_write_filepage: already started\n");
8056 		return;
8057 	}
8058 	pagedep->pd_state |= IOSTARTED;
8059 	/*
8060 	 * Wait for all journal remove dependencies to hit the disk.
8061 	 * We can not allow any potentially conflicting directory adds
8062 	 * to be visible before removes and rollback is too difficult.
8063 	 * lk may be dropped and re-acquired, however we hold the buf
8064 	 * locked so the dependency can not go away.
8065 	 */
8066 	LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next)
8067 		while ((jremref = LIST_FIRST(&dirrem->dm_jremrefhd)) != NULL) {
8068 			stat_jwait_filepage++;
8069 			jwait(&jremref->jr_list);
8070 		}
8071 	while ((jmvref = LIST_FIRST(&pagedep->pd_jmvrefhd)) != NULL) {
8072 		stat_jwait_filepage++;
8073 		jwait(&jmvref->jm_list);
8074 	}
8075 	for (i = 0; i < DAHASHSZ; i++) {
8076 		LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
8077 			ep = (struct direct *)
8078 			    ((char *)bp->b_data + dap->da_offset);
8079 			if (ep->d_ino != dap->da_newinum)
8080 				panic("%s: dir inum %d != new %d",
8081 				    "initiate_write_filepage",
8082 				    ep->d_ino, dap->da_newinum);
8083 			if (dap->da_state & DIRCHG)
8084 				ep->d_ino = dap->da_previous->dm_oldinum;
8085 			else
8086 				ep->d_ino = 0;
8087 			dap->da_state &= ~ATTACHED;
8088 			dap->da_state |= UNDONE;
8089 		}
8090 	}
8091 }
8092 
8093 /*
8094  * Version of initiate_write_inodeblock that handles UFS1 dinodes.
8095  * Note that any bug fixes made to this routine must be done in the
8096  * version found below.
8097  *
8098  * Called from within the procedure above to deal with unsatisfied
8099  * allocation dependencies in an inodeblock. The buffer must be
8100  * locked, thus, no I/O completion operations can occur while we
8101  * are manipulating its associated dependencies.
8102  */
8103 static void
8104 initiate_write_inodeblock_ufs1(inodedep, bp)
8105 	struct inodedep *inodedep;
8106 	struct buf *bp;			/* The inode block */
8107 {
8108 	struct allocdirect *adp, *lastadp;
8109 	struct ufs1_dinode *dp;
8110 	struct ufs1_dinode *sip;
8111 	struct inoref *inoref;
8112 	struct fs *fs;
8113 	ufs_lbn_t i;
8114 #ifdef INVARIANTS
8115 	ufs_lbn_t prevlbn = 0;
8116 #endif
8117 	int deplist;
8118 
8119 	if (inodedep->id_state & IOSTARTED)
8120 		panic("initiate_write_inodeblock_ufs1: already started");
8121 	inodedep->id_state |= IOSTARTED;
8122 	fs = inodedep->id_fs;
8123 	dp = (struct ufs1_dinode *)bp->b_data +
8124 	    ino_to_fsbo(fs, inodedep->id_ino);
8125 
8126 	/*
8127 	 * If we're on the unlinked list but have not yet written our
8128 	 * next pointer initialize it here.
8129 	 */
8130 	if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
8131 		struct inodedep *inon;
8132 
8133 		inon = TAILQ_NEXT(inodedep, id_unlinked);
8134 		dp->di_freelink = inon ? inon->id_ino : 0;
8135 	}
8136 	/*
8137 	 * If the bitmap is not yet written, then the allocated
8138 	 * inode cannot be written to disk.
8139 	 */
8140 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
8141 		if (inodedep->id_savedino1 != NULL)
8142 			panic("initiate_write_inodeblock_ufs1: I/O underway");
8143 		FREE_LOCK(&lk);
8144 		sip = malloc(sizeof(struct ufs1_dinode),
8145 		    M_SAVEDINO, M_SOFTDEP_FLAGS);
8146 		ACQUIRE_LOCK(&lk);
8147 		inodedep->id_savedino1 = sip;
8148 		*inodedep->id_savedino1 = *dp;
8149 		bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
8150 		dp->di_gen = inodedep->id_savedino1->di_gen;
8151 		dp->di_freelink = inodedep->id_savedino1->di_freelink;
8152 		return;
8153 	}
8154 	/*
8155 	 * If no dependencies, then there is nothing to roll back.
8156 	 */
8157 	inodedep->id_savedsize = dp->di_size;
8158 	inodedep->id_savedextsize = 0;
8159 	inodedep->id_savednlink = dp->di_nlink;
8160 	if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
8161 	    TAILQ_EMPTY(&inodedep->id_inoreflst))
8162 		return;
8163 	/*
8164 	 * Revert the link count to that of the first unwritten journal entry.
8165 	 */
8166 	inoref = TAILQ_FIRST(&inodedep->id_inoreflst);
8167 	if (inoref)
8168 		dp->di_nlink = inoref->if_nlink;
8169 	/*
8170 	 * Set the dependencies to busy.
8171 	 */
8172 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
8173 	     adp = TAILQ_NEXT(adp, ad_next)) {
8174 #ifdef INVARIANTS
8175 		if (deplist != 0 && prevlbn >= adp->ad_offset)
8176 			panic("softdep_write_inodeblock: lbn order");
8177 		prevlbn = adp->ad_offset;
8178 		if (adp->ad_offset < NDADDR &&
8179 		    dp->di_db[adp->ad_offset] != adp->ad_newblkno)
8180 			panic("%s: direct pointer #%jd mismatch %d != %jd",
8181 			    "softdep_write_inodeblock",
8182 			    (intmax_t)adp->ad_offset,
8183 			    dp->di_db[adp->ad_offset],
8184 			    (intmax_t)adp->ad_newblkno);
8185 		if (adp->ad_offset >= NDADDR &&
8186 		    dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno)
8187 			panic("%s: indirect pointer #%jd mismatch %d != %jd",
8188 			    "softdep_write_inodeblock",
8189 			    (intmax_t)adp->ad_offset - NDADDR,
8190 			    dp->di_ib[adp->ad_offset - NDADDR],
8191 			    (intmax_t)adp->ad_newblkno);
8192 		deplist |= 1 << adp->ad_offset;
8193 		if ((adp->ad_state & ATTACHED) == 0)
8194 			panic("softdep_write_inodeblock: Unknown state 0x%x",
8195 			    adp->ad_state);
8196 #endif /* INVARIANTS */
8197 		adp->ad_state &= ~ATTACHED;
8198 		adp->ad_state |= UNDONE;
8199 	}
8200 	/*
8201 	 * The on-disk inode cannot claim to be any larger than the last
8202 	 * fragment that has been written. Otherwise, the on-disk inode
8203 	 * might have fragments that were not the last block in the file
8204 	 * which would corrupt the filesystem.
8205 	 */
8206 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
8207 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
8208 		if (adp->ad_offset >= NDADDR)
8209 			break;
8210 		dp->di_db[adp->ad_offset] = adp->ad_oldblkno;
8211 		/* keep going until hitting a rollback to a frag */
8212 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
8213 			continue;
8214 		dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
8215 		for (i = adp->ad_offset + 1; i < NDADDR; i++) {
8216 #ifdef INVARIANTS
8217 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
8218 				panic("softdep_write_inodeblock: lost dep1");
8219 #endif /* INVARIANTS */
8220 			dp->di_db[i] = 0;
8221 		}
8222 		for (i = 0; i < NIADDR; i++) {
8223 #ifdef INVARIANTS
8224 			if (dp->di_ib[i] != 0 &&
8225 			    (deplist & ((1 << NDADDR) << i)) == 0)
8226 				panic("softdep_write_inodeblock: lost dep2");
8227 #endif /* INVARIANTS */
8228 			dp->di_ib[i] = 0;
8229 		}
8230 		return;
8231 	}
8232 	/*
8233 	 * If we have zero'ed out the last allocated block of the file,
8234 	 * roll back the size to the last currently allocated block.
8235 	 * We know that this last allocated block is a full-sized as
8236 	 * we already checked for fragments in the loop above.
8237 	 */
8238 	if (lastadp != NULL &&
8239 	    dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
8240 		for (i = lastadp->ad_offset; i >= 0; i--)
8241 			if (dp->di_db[i] != 0)
8242 				break;
8243 		dp->di_size = (i + 1) * fs->fs_bsize;
8244 	}
8245 	/*
8246 	 * The only dependencies are for indirect blocks.
8247 	 *
8248 	 * The file size for indirect block additions is not guaranteed.
8249 	 * Such a guarantee would be non-trivial to achieve. The conventional
8250 	 * synchronous write implementation also does not make this guarantee.
8251 	 * Fsck should catch and fix discrepancies. Arguably, the file size
8252 	 * can be over-estimated without destroying integrity when the file
8253 	 * moves into the indirect blocks (i.e., is large). If we want to
8254 	 * postpone fsck, we are stuck with this argument.
8255 	 */
8256 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
8257 		dp->di_ib[adp->ad_offset - NDADDR] = 0;
8258 }
8259 
8260 /*
8261  * Version of initiate_write_inodeblock that handles UFS2 dinodes.
8262  * Note that any bug fixes made to this routine must be done in the
8263  * version found above.
8264  *
8265  * Called from within the procedure above to deal with unsatisfied
8266  * allocation dependencies in an inodeblock. The buffer must be
8267  * locked, thus, no I/O completion operations can occur while we
8268  * are manipulating its associated dependencies.
8269  */
8270 static void
8271 initiate_write_inodeblock_ufs2(inodedep, bp)
8272 	struct inodedep *inodedep;
8273 	struct buf *bp;			/* The inode block */
8274 {
8275 	struct allocdirect *adp, *lastadp;
8276 	struct ufs2_dinode *dp;
8277 	struct ufs2_dinode *sip;
8278 	struct inoref *inoref;
8279 	struct fs *fs;
8280 	ufs_lbn_t i;
8281 #ifdef INVARIANTS
8282 	ufs_lbn_t prevlbn = 0;
8283 #endif
8284 	int deplist;
8285 
8286 	if (inodedep->id_state & IOSTARTED)
8287 		panic("initiate_write_inodeblock_ufs2: already started");
8288 	inodedep->id_state |= IOSTARTED;
8289 	fs = inodedep->id_fs;
8290 	dp = (struct ufs2_dinode *)bp->b_data +
8291 	    ino_to_fsbo(fs, inodedep->id_ino);
8292 
8293 	/*
8294 	 * If we're on the unlinked list but have not yet written our
8295 	 * next pointer initialize it here.
8296 	 */
8297 	if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
8298 		struct inodedep *inon;
8299 
8300 		inon = TAILQ_NEXT(inodedep, id_unlinked);
8301 		dp->di_freelink = inon ? inon->id_ino : 0;
8302 	}
8303 	if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) ==
8304 	    (UNLINKED | UNLINKNEXT)) {
8305 		struct inodedep *inon;
8306 		ino_t freelink;
8307 
8308 		inon = TAILQ_NEXT(inodedep, id_unlinked);
8309 		freelink = inon ? inon->id_ino : 0;
8310 		if (freelink != dp->di_freelink)
8311 			panic("ino %p(0x%X) %d, %d != %d",
8312 			    inodedep, inodedep->id_state, inodedep->id_ino,
8313 			    freelink, dp->di_freelink);
8314 	}
8315 	/*
8316 	 * If the bitmap is not yet written, then the allocated
8317 	 * inode cannot be written to disk.
8318 	 */
8319 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
8320 		if (inodedep->id_savedino2 != NULL)
8321 			panic("initiate_write_inodeblock_ufs2: I/O underway");
8322 		FREE_LOCK(&lk);
8323 		sip = malloc(sizeof(struct ufs2_dinode),
8324 		    M_SAVEDINO, M_SOFTDEP_FLAGS);
8325 		ACQUIRE_LOCK(&lk);
8326 		inodedep->id_savedino2 = sip;
8327 		*inodedep->id_savedino2 = *dp;
8328 		bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
8329 		dp->di_gen = inodedep->id_savedino2->di_gen;
8330 		dp->di_freelink = inodedep->id_savedino2->di_freelink;
8331 		return;
8332 	}
8333 	/*
8334 	 * If no dependencies, then there is nothing to roll back.
8335 	 */
8336 	inodedep->id_savedsize = dp->di_size;
8337 	inodedep->id_savedextsize = dp->di_extsize;
8338 	inodedep->id_savednlink = dp->di_nlink;
8339 	if (TAILQ_EMPTY(&inodedep->id_inoupdt) &&
8340 	    TAILQ_EMPTY(&inodedep->id_extupdt) &&
8341 	    TAILQ_EMPTY(&inodedep->id_inoreflst))
8342 		return;
8343 	/*
8344 	 * Revert the link count to that of the first unwritten journal entry.
8345 	 */
8346 	inoref = TAILQ_FIRST(&inodedep->id_inoreflst);
8347 	if (inoref)
8348 		dp->di_nlink = inoref->if_nlink;
8349 
8350 	/*
8351 	 * Set the ext data dependencies to busy.
8352 	 */
8353 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
8354 	     adp = TAILQ_NEXT(adp, ad_next)) {
8355 #ifdef INVARIANTS
8356 		if (deplist != 0 && prevlbn >= adp->ad_offset)
8357 			panic("softdep_write_inodeblock: lbn order");
8358 		prevlbn = adp->ad_offset;
8359 		if (dp->di_extb[adp->ad_offset] != adp->ad_newblkno)
8360 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
8361 			    "softdep_write_inodeblock",
8362 			    (intmax_t)adp->ad_offset,
8363 			    (intmax_t)dp->di_extb[adp->ad_offset],
8364 			    (intmax_t)adp->ad_newblkno);
8365 		deplist |= 1 << adp->ad_offset;
8366 		if ((adp->ad_state & ATTACHED) == 0)
8367 			panic("softdep_write_inodeblock: Unknown state 0x%x",
8368 			    adp->ad_state);
8369 #endif /* INVARIANTS */
8370 		adp->ad_state &= ~ATTACHED;
8371 		adp->ad_state |= UNDONE;
8372 	}
8373 	/*
8374 	 * The on-disk inode cannot claim to be any larger than the last
8375 	 * fragment that has been written. Otherwise, the on-disk inode
8376 	 * might have fragments that were not the last block in the ext
8377 	 * data which would corrupt the filesystem.
8378 	 */
8379 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
8380 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
8381 		dp->di_extb[adp->ad_offset] = adp->ad_oldblkno;
8382 		/* keep going until hitting a rollback to a frag */
8383 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
8384 			continue;
8385 		dp->di_extsize = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
8386 		for (i = adp->ad_offset + 1; i < NXADDR; i++) {
8387 #ifdef INVARIANTS
8388 			if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0)
8389 				panic("softdep_write_inodeblock: lost dep1");
8390 #endif /* INVARIANTS */
8391 			dp->di_extb[i] = 0;
8392 		}
8393 		lastadp = NULL;
8394 		break;
8395 	}
8396 	/*
8397 	 * If we have zero'ed out the last allocated block of the ext
8398 	 * data, roll back the size to the last currently allocated block.
8399 	 * We know that this last allocated block is a full-sized as
8400 	 * we already checked for fragments in the loop above.
8401 	 */
8402 	if (lastadp != NULL &&
8403 	    dp->di_extsize <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
8404 		for (i = lastadp->ad_offset; i >= 0; i--)
8405 			if (dp->di_extb[i] != 0)
8406 				break;
8407 		dp->di_extsize = (i + 1) * fs->fs_bsize;
8408 	}
8409 	/*
8410 	 * Set the file data dependencies to busy.
8411 	 */
8412 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
8413 	     adp = TAILQ_NEXT(adp, ad_next)) {
8414 #ifdef INVARIANTS
8415 		if (deplist != 0 && prevlbn >= adp->ad_offset)
8416 			panic("softdep_write_inodeblock: lbn order");
8417 		prevlbn = adp->ad_offset;
8418 		if (adp->ad_offset < NDADDR &&
8419 		    dp->di_db[adp->ad_offset] != adp->ad_newblkno)
8420 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
8421 			    "softdep_write_inodeblock",
8422 			    (intmax_t)adp->ad_offset,
8423 			    (intmax_t)dp->di_db[adp->ad_offset],
8424 			    (intmax_t)adp->ad_newblkno);
8425 		if (adp->ad_offset >= NDADDR &&
8426 		    dp->di_ib[adp->ad_offset - NDADDR] != adp->ad_newblkno)
8427 			panic("%s indirect pointer #%jd mismatch %jd != %jd",
8428 			    "softdep_write_inodeblock:",
8429 			    (intmax_t)adp->ad_offset - NDADDR,
8430 			    (intmax_t)dp->di_ib[adp->ad_offset - NDADDR],
8431 			    (intmax_t)adp->ad_newblkno);
8432 		deplist |= 1 << adp->ad_offset;
8433 		if ((adp->ad_state & ATTACHED) == 0)
8434 			panic("softdep_write_inodeblock: Unknown state 0x%x",
8435 			    adp->ad_state);
8436 #endif /* INVARIANTS */
8437 		adp->ad_state &= ~ATTACHED;
8438 		adp->ad_state |= UNDONE;
8439 	}
8440 	/*
8441 	 * The on-disk inode cannot claim to be any larger than the last
8442 	 * fragment that has been written. Otherwise, the on-disk inode
8443 	 * might have fragments that were not the last block in the file
8444 	 * which would corrupt the filesystem.
8445 	 */
8446 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
8447 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
8448 		if (adp->ad_offset >= NDADDR)
8449 			break;
8450 		dp->di_db[adp->ad_offset] = adp->ad_oldblkno;
8451 		/* keep going until hitting a rollback to a frag */
8452 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
8453 			continue;
8454 		dp->di_size = fs->fs_bsize * adp->ad_offset + adp->ad_oldsize;
8455 		for (i = adp->ad_offset + 1; i < NDADDR; i++) {
8456 #ifdef INVARIANTS
8457 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
8458 				panic("softdep_write_inodeblock: lost dep2");
8459 #endif /* INVARIANTS */
8460 			dp->di_db[i] = 0;
8461 		}
8462 		for (i = 0; i < NIADDR; i++) {
8463 #ifdef INVARIANTS
8464 			if (dp->di_ib[i] != 0 &&
8465 			    (deplist & ((1 << NDADDR) << i)) == 0)
8466 				panic("softdep_write_inodeblock: lost dep3");
8467 #endif /* INVARIANTS */
8468 			dp->di_ib[i] = 0;
8469 		}
8470 		return;
8471 	}
8472 	/*
8473 	 * If we have zero'ed out the last allocated block of the file,
8474 	 * roll back the size to the last currently allocated block.
8475 	 * We know that this last allocated block is a full-sized as
8476 	 * we already checked for fragments in the loop above.
8477 	 */
8478 	if (lastadp != NULL &&
8479 	    dp->di_size <= (lastadp->ad_offset + 1) * fs->fs_bsize) {
8480 		for (i = lastadp->ad_offset; i >= 0; i--)
8481 			if (dp->di_db[i] != 0)
8482 				break;
8483 		dp->di_size = (i + 1) * fs->fs_bsize;
8484 	}
8485 	/*
8486 	 * The only dependencies are for indirect blocks.
8487 	 *
8488 	 * The file size for indirect block additions is not guaranteed.
8489 	 * Such a guarantee would be non-trivial to achieve. The conventional
8490 	 * synchronous write implementation also does not make this guarantee.
8491 	 * Fsck should catch and fix discrepancies. Arguably, the file size
8492 	 * can be over-estimated without destroying integrity when the file
8493 	 * moves into the indirect blocks (i.e., is large). If we want to
8494 	 * postpone fsck, we are stuck with this argument.
8495 	 */
8496 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
8497 		dp->di_ib[adp->ad_offset - NDADDR] = 0;
8498 }
8499 
8500 /*
8501  * Cancel an indirdep as a result of truncation.  Release all of the
8502  * children allocindirs and place their journal work on the appropriate
8503  * list.
8504  */
8505 static void
8506 cancel_indirdep(indirdep, bp, inodedep, freeblks)
8507 	struct indirdep *indirdep;
8508 	struct buf *bp;
8509 	struct inodedep *inodedep;
8510 	struct freeblks *freeblks;
8511 {
8512 	struct allocindir *aip;
8513 
8514 	/*
8515 	 * None of the indirect pointers will ever be visible,
8516 	 * so they can simply be tossed. GOINGAWAY ensures
8517 	 * that allocated pointers will be saved in the buffer
8518 	 * cache until they are freed. Note that they will
8519 	 * only be able to be found by their physical address
8520 	 * since the inode mapping the logical address will
8521 	 * be gone. The save buffer used for the safe copy
8522 	 * was allocated in setup_allocindir_phase2 using
8523 	 * the physical address so it could be used for this
8524 	 * purpose. Hence we swap the safe copy with the real
8525 	 * copy, allowing the safe copy to be freed and holding
8526 	 * on to the real copy for later use in indir_trunc.
8527 	 */
8528 	if (indirdep->ir_state & GOINGAWAY)
8529 		panic("cancel_indirdep: already gone");
8530 	if (indirdep->ir_state & ONDEPLIST) {
8531 		indirdep->ir_state &= ~ONDEPLIST;
8532 		LIST_REMOVE(indirdep, ir_next);
8533 	}
8534 	indirdep->ir_state |= GOINGAWAY;
8535 	VFSTOUFS(indirdep->ir_list.wk_mp)->um_numindirdeps += 1;
8536 	while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
8537 		cancel_allocindir(aip, inodedep, freeblks);
8538 	while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0)
8539 		cancel_allocindir(aip, inodedep, freeblks);
8540 	while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0)
8541 		cancel_allocindir(aip, inodedep, freeblks);
8542 	while ((aip = LIST_FIRST(&indirdep->ir_completehd)) != 0)
8543 		cancel_allocindir(aip, inodedep, freeblks);
8544 	bcopy(bp->b_data, indirdep->ir_savebp->b_data, bp->b_bcount);
8545 	WORKLIST_REMOVE(&indirdep->ir_list);
8546 	WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, &indirdep->ir_list);
8547 	indirdep->ir_savebp = NULL;
8548 }
8549 
8550 /*
8551  * Free an indirdep once it no longer has new pointers to track.
8552  */
8553 static void
8554 free_indirdep(indirdep)
8555 	struct indirdep *indirdep;
8556 {
8557 
8558 	KASSERT(LIST_EMPTY(&indirdep->ir_jwork),
8559 	    ("free_indirdep: Journal work not empty."));
8560 	KASSERT(LIST_EMPTY(&indirdep->ir_completehd),
8561 	    ("free_indirdep: Complete head not empty."));
8562 	KASSERT(LIST_EMPTY(&indirdep->ir_writehd),
8563 	    ("free_indirdep: write head not empty."));
8564 	KASSERT(LIST_EMPTY(&indirdep->ir_donehd),
8565 	    ("free_indirdep: done head not empty."));
8566 	KASSERT(LIST_EMPTY(&indirdep->ir_deplisthd),
8567 	    ("free_indirdep: deplist head not empty."));
8568 	KASSERT(indirdep->ir_savebp == NULL,
8569 	    ("free_indirdep: %p ir_savebp != NULL", indirdep));
8570 	KASSERT((indirdep->ir_state & ONDEPLIST) == 0,
8571 	    ("free_indirdep: %p still on deplist.", indirdep));
8572 	if (indirdep->ir_state & ONWORKLIST)
8573 		WORKLIST_REMOVE(&indirdep->ir_list);
8574 	WORKITEM_FREE(indirdep, D_INDIRDEP);
8575 }
8576 
8577 /*
8578  * Called before a write to an indirdep.  This routine is responsible for
8579  * rolling back pointers to a safe state which includes only those
8580  * allocindirs which have been completed.
8581  */
8582 static void
8583 initiate_write_indirdep(indirdep, bp)
8584 	struct indirdep *indirdep;
8585 	struct buf *bp;
8586 {
8587 
8588 	if (indirdep->ir_state & GOINGAWAY)
8589 		panic("disk_io_initiation: indirdep gone");
8590 
8591 	/*
8592 	 * If there are no remaining dependencies, this will be writing
8593 	 * the real pointers.
8594 	 */
8595 	if (LIST_EMPTY(&indirdep->ir_deplisthd))
8596 		return;
8597 	/*
8598 	 * Replace up-to-date version with safe version.
8599 	 */
8600 	FREE_LOCK(&lk);
8601 	indirdep->ir_saveddata = malloc(bp->b_bcount, M_INDIRDEP,
8602 	    M_SOFTDEP_FLAGS);
8603 	ACQUIRE_LOCK(&lk);
8604 	indirdep->ir_state &= ~ATTACHED;
8605 	indirdep->ir_state |= UNDONE;
8606 	bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
8607 	bcopy(indirdep->ir_savebp->b_data, bp->b_data,
8608 	    bp->b_bcount);
8609 }
8610 
8611 /*
8612  * Called when an inode has been cleared in a cg bitmap.  This finally
8613  * eliminates any canceled jaddrefs
8614  */
8615 void
8616 softdep_setup_inofree(mp, bp, ino, wkhd)
8617 	struct mount *mp;
8618 	struct buf *bp;
8619 	ino_t ino;
8620 	struct workhead *wkhd;
8621 {
8622 	struct worklist *wk, *wkn;
8623 	struct inodedep *inodedep;
8624 	uint8_t *inosused;
8625 	struct cg *cgp;
8626 	struct fs *fs;
8627 
8628 	ACQUIRE_LOCK(&lk);
8629 	fs = VFSTOUFS(mp)->um_fs;
8630 	cgp = (struct cg *)bp->b_data;
8631 	inosused = cg_inosused(cgp);
8632 	if (isset(inosused, ino % fs->fs_ipg))
8633 		panic("softdep_setup_inofree: inode %d not freed.", ino);
8634 	if (inodedep_lookup(mp, ino, 0, &inodedep))
8635 		panic("softdep_setup_inofree: ino %d has existing inodedep %p",
8636 		    ino, inodedep);
8637 	if (wkhd) {
8638 		LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) {
8639 			if (wk->wk_type != D_JADDREF)
8640 				continue;
8641 			WORKLIST_REMOVE(wk);
8642 			/*
8643 			 * We can free immediately even if the jaddref
8644 			 * isn't attached in a background write as now
8645 			 * the bitmaps are reconciled.
8646 		 	 */
8647 			wk->wk_state |= COMPLETE | ATTACHED;
8648 			free_jaddref(WK_JADDREF(wk));
8649 		}
8650 		jwork_move(&bp->b_dep, wkhd);
8651 	}
8652 	FREE_LOCK(&lk);
8653 }
8654 
8655 
8656 /*
8657  * Called via ffs_blkfree() after a set of frags has been cleared from a cg
8658  * map.  Any dependencies waiting for the write to clear are added to the
8659  * buf's list and any jnewblks that are being canceled are discarded
8660  * immediately.
8661  */
8662 void
8663 softdep_setup_blkfree(mp, bp, blkno, frags, wkhd)
8664 	struct mount *mp;
8665 	struct buf *bp;
8666 	ufs2_daddr_t blkno;
8667 	int frags;
8668 	struct workhead *wkhd;
8669 {
8670 	struct jnewblk *jnewblk;
8671 	struct worklist *wk, *wkn;
8672 #ifdef SUJ_DEBUG
8673 	struct bmsafemap *bmsafemap;
8674 	struct fs *fs;
8675 	uint8_t *blksfree;
8676 	struct cg *cgp;
8677 	ufs2_daddr_t jstart;
8678 	ufs2_daddr_t jend;
8679 	ufs2_daddr_t end;
8680 	long bno;
8681 	int i;
8682 #endif
8683 
8684 	ACQUIRE_LOCK(&lk);
8685 	/*
8686 	 * Detach any jnewblks which have been canceled.  They must linger
8687 	 * until the bitmap is cleared again by ffs_blkfree() to prevent
8688 	 * an unjournaled allocation from hitting the disk.
8689 	 */
8690 	if (wkhd) {
8691 		LIST_FOREACH_SAFE(wk, wkhd, wk_list, wkn) {
8692 			if (wk->wk_type != D_JNEWBLK)
8693 				continue;
8694 			jnewblk = WK_JNEWBLK(wk);
8695 			KASSERT(jnewblk->jn_state & GOINGAWAY,
8696 			    ("softdep_setup_blkfree: jnewblk not canceled."));
8697 			WORKLIST_REMOVE(wk);
8698 #ifdef SUJ_DEBUG
8699 			/*
8700 			 * Assert that this block is free in the bitmap
8701 			 * before we discard the jnewblk.
8702 			 */
8703 			fs = VFSTOUFS(mp)->um_fs;
8704 			cgp = (struct cg *)bp->b_data;
8705 			blksfree = cg_blksfree(cgp);
8706 			bno = dtogd(fs, jnewblk->jn_blkno);
8707 			for (i = jnewblk->jn_oldfrags;
8708 			    i < jnewblk->jn_frags; i++) {
8709 				if (isset(blksfree, bno + i))
8710 					continue;
8711 				panic("softdep_setup_blkfree: not free");
8712 			}
8713 #endif
8714 			/*
8715 			 * Even if it's not attached we can free immediately
8716 			 * as the new bitmap is correct.
8717 			 */
8718 			wk->wk_state |= COMPLETE | ATTACHED;
8719 			free_jnewblk(jnewblk);
8720 		}
8721 		/*
8722 		 * The buf must be locked by the caller otherwise these could
8723 		 * be added while it's being written and the write would
8724 		 * complete them before they made it to disk.
8725 		 */
8726 		jwork_move(&bp->b_dep, wkhd);
8727 	}
8728 
8729 #ifdef SUJ_DEBUG
8730 	/*
8731 	 * Assert that we are not freeing a block which has an outstanding
8732 	 * allocation dependency.
8733 	 */
8734 	fs = VFSTOUFS(mp)->um_fs;
8735 	bmsafemap = bmsafemap_lookup(mp, bp, dtog(fs, blkno));
8736 	end = blkno + frags;
8737 	LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) {
8738 		/*
8739 		 * Don't match against blocks that will be freed when the
8740 		 * background write is done.
8741 		 */
8742 		if ((jnewblk->jn_state & (ATTACHED | COMPLETE | DEPCOMPLETE)) ==
8743 		    (COMPLETE | DEPCOMPLETE))
8744 			continue;
8745 		jstart = jnewblk->jn_blkno + jnewblk->jn_oldfrags;
8746 		jend = jnewblk->jn_blkno + jnewblk->jn_frags;
8747 		if ((blkno >= jstart && blkno < jend) ||
8748 		    (end > jstart && end <= jend)) {
8749 			printf("state 0x%X %jd - %d %d dep %p\n",
8750 			    jnewblk->jn_state, jnewblk->jn_blkno,
8751 			    jnewblk->jn_oldfrags, jnewblk->jn_frags,
8752 			    jnewblk->jn_newblk);
8753 			panic("softdep_setup_blkfree: "
8754 			    "%jd-%jd(%d) overlaps with %jd-%jd",
8755 			    blkno, end, frags, jstart, jend);
8756 		}
8757 	}
8758 #endif
8759 	FREE_LOCK(&lk);
8760 }
8761 
8762 static void
8763 initiate_write_bmsafemap(bmsafemap, bp)
8764 	struct bmsafemap *bmsafemap;
8765 	struct buf *bp;			/* The cg block. */
8766 {
8767 	struct jaddref *jaddref;
8768 	struct jnewblk *jnewblk;
8769 	uint8_t *inosused;
8770 	uint8_t *blksfree;
8771 	struct cg *cgp;
8772 	struct fs *fs;
8773 	int cleared;
8774 	ino_t ino;
8775 	long bno;
8776 	int i;
8777 
8778 	if (bmsafemap->sm_state & IOSTARTED)
8779 		panic("initiate_write_bmsafemap: Already started\n");
8780 	bmsafemap->sm_state |= IOSTARTED;
8781 	/*
8782 	 * Clear any inode allocations which are pending journal writes.
8783 	 */
8784 	if (LIST_FIRST(&bmsafemap->sm_jaddrefhd) != NULL) {
8785 		cgp = (struct cg *)bp->b_data;
8786 		fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
8787 		inosused = cg_inosused(cgp);
8788 		LIST_FOREACH(jaddref, &bmsafemap->sm_jaddrefhd, ja_bmdeps) {
8789 			ino = jaddref->ja_ino % fs->fs_ipg;
8790 			/*
8791 			 * If this is a background copy the inode may not
8792 			 * be marked used yet.
8793 			 */
8794 			if (isset(inosused, ino)) {
8795 				if ((jaddref->ja_mode & IFMT) == IFDIR)
8796 					cgp->cg_cs.cs_ndir--;
8797 				cgp->cg_cs.cs_nifree++;
8798 				clrbit(inosused, ino);
8799 				jaddref->ja_state &= ~ATTACHED;
8800 				jaddref->ja_state |= UNDONE;
8801 				stat_jaddref++;
8802 			} else if ((bp->b_xflags & BX_BKGRDMARKER) == 0)
8803 				panic("initiate_write_bmsafemap: inode %d "
8804 				    "marked free", jaddref->ja_ino);
8805 		}
8806 	}
8807 	/*
8808 	 * Clear any block allocations which are pending journal writes.
8809 	 */
8810 	if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) {
8811 		cgp = (struct cg *)bp->b_data;
8812 		fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
8813 		blksfree = cg_blksfree(cgp);
8814 		LIST_FOREACH(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps) {
8815 			bno = dtogd(fs, jnewblk->jn_blkno);
8816 			cleared = 0;
8817 			for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags;
8818 			    i++) {
8819 				if (isclr(blksfree, bno + i)) {
8820 					cleared = 1;
8821 					setbit(blksfree, bno + i);
8822 				}
8823 			}
8824 			/*
8825 			 * We may not clear the block if it's a background
8826 			 * copy.  In that case there is no reason to detach
8827 			 * it.
8828 			 */
8829 			if (cleared) {
8830 				stat_jnewblk++;
8831 				jnewblk->jn_state &= ~ATTACHED;
8832 				jnewblk->jn_state |= UNDONE;
8833 			} else if ((bp->b_xflags & BX_BKGRDMARKER) == 0)
8834 				panic("initiate_write_bmsafemap: block %jd "
8835 				    "marked free", jnewblk->jn_blkno);
8836 		}
8837 	}
8838 	/*
8839 	 * Move allocation lists to the written lists so they can be
8840 	 * cleared once the block write is complete.
8841 	 */
8842 	LIST_SWAP(&bmsafemap->sm_inodedephd, &bmsafemap->sm_inodedepwr,
8843 	    inodedep, id_deps);
8844 	LIST_SWAP(&bmsafemap->sm_newblkhd, &bmsafemap->sm_newblkwr,
8845 	    newblk, nb_deps);
8846 }
8847 
8848 /*
8849  * This routine is called during the completion interrupt
8850  * service routine for a disk write (from the procedure called
8851  * by the device driver to inform the filesystem caches of
8852  * a request completion).  It should be called early in this
8853  * procedure, before the block is made available to other
8854  * processes or other routines are called.
8855  *
8856  */
8857 static void
8858 softdep_disk_write_complete(bp)
8859 	struct buf *bp;		/* describes the completed disk write */
8860 {
8861 	struct worklist *wk;
8862 	struct worklist *owk;
8863 	struct workhead reattach;
8864 	struct buf *sbp;
8865 
8866 	/*
8867 	 * If an error occurred while doing the write, then the data
8868 	 * has not hit the disk and the dependencies cannot be unrolled.
8869 	 */
8870 	if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0)
8871 		return;
8872 	LIST_INIT(&reattach);
8873 	/*
8874 	 * This lock must not be released anywhere in this code segment.
8875 	 */
8876 	sbp = NULL;
8877 	owk = NULL;
8878 	ACQUIRE_LOCK(&lk);
8879 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
8880 		WORKLIST_REMOVE(wk);
8881 		if (wk == owk)
8882 			panic("duplicate worklist: %p\n", wk);
8883 		owk = wk;
8884 		switch (wk->wk_type) {
8885 
8886 		case D_PAGEDEP:
8887 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
8888 				WORKLIST_INSERT(&reattach, wk);
8889 			continue;
8890 
8891 		case D_INODEDEP:
8892 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
8893 				WORKLIST_INSERT(&reattach, wk);
8894 			continue;
8895 
8896 		case D_BMSAFEMAP:
8897 			if (handle_written_bmsafemap(WK_BMSAFEMAP(wk), bp))
8898 				WORKLIST_INSERT(&reattach, wk);
8899 			continue;
8900 
8901 		case D_MKDIR:
8902 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
8903 			continue;
8904 
8905 		case D_ALLOCDIRECT:
8906 			wk->wk_state |= COMPLETE;
8907 			handle_allocdirect_partdone(WK_ALLOCDIRECT(wk), NULL);
8908 			continue;
8909 
8910 		case D_ALLOCINDIR:
8911 			wk->wk_state |= COMPLETE;
8912 			handle_allocindir_partdone(WK_ALLOCINDIR(wk));
8913 			continue;
8914 
8915 		case D_INDIRDEP:
8916 			if (handle_written_indirdep(WK_INDIRDEP(wk), bp, &sbp))
8917 				WORKLIST_INSERT(&reattach, wk);
8918 			continue;
8919 
8920 		case D_FREEBLKS:
8921 			wk->wk_state |= COMPLETE;
8922 			if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE)
8923 				add_to_worklist(wk, 1);
8924 			continue;
8925 
8926 		case D_FREEWORK:
8927 			handle_written_freework(WK_FREEWORK(wk));
8928 			break;
8929 
8930 		case D_FREEDEP:
8931 			free_freedep(WK_FREEDEP(wk));
8932 			continue;
8933 
8934 		case D_JSEGDEP:
8935 			free_jsegdep(WK_JSEGDEP(wk));
8936 			continue;
8937 
8938 		case D_JSEG:
8939 			handle_written_jseg(WK_JSEG(wk), bp);
8940 			continue;
8941 
8942 		case D_SBDEP:
8943 			if (handle_written_sbdep(WK_SBDEP(wk), bp))
8944 				WORKLIST_INSERT(&reattach, wk);
8945 			continue;
8946 
8947 		default:
8948 			panic("handle_disk_write_complete: Unknown type %s",
8949 			    TYPENAME(wk->wk_type));
8950 			/* NOTREACHED */
8951 		}
8952 	}
8953 	/*
8954 	 * Reattach any requests that must be redone.
8955 	 */
8956 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
8957 		WORKLIST_REMOVE(wk);
8958 		WORKLIST_INSERT(&bp->b_dep, wk);
8959 	}
8960 	FREE_LOCK(&lk);
8961 	if (sbp)
8962 		brelse(sbp);
8963 }
8964 
8965 /*
8966  * Called from within softdep_disk_write_complete above. Note that
8967  * this routine is always called from interrupt level with further
8968  * splbio interrupts blocked.
8969  */
8970 static void
8971 handle_allocdirect_partdone(adp, wkhd)
8972 	struct allocdirect *adp;	/* the completed allocdirect */
8973 	struct workhead *wkhd;		/* Work to do when inode is writtne. */
8974 {
8975 	struct allocdirectlst *listhead;
8976 	struct allocdirect *listadp;
8977 	struct inodedep *inodedep;
8978 	long bsize;
8979 
8980 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
8981 		return;
8982 	/*
8983 	 * The on-disk inode cannot claim to be any larger than the last
8984 	 * fragment that has been written. Otherwise, the on-disk inode
8985 	 * might have fragments that were not the last block in the file
8986 	 * which would corrupt the filesystem. Thus, we cannot free any
8987 	 * allocdirects after one whose ad_oldblkno claims a fragment as
8988 	 * these blocks must be rolled back to zero before writing the inode.
8989 	 * We check the currently active set of allocdirects in id_inoupdt
8990 	 * or id_extupdt as appropriate.
8991 	 */
8992 	inodedep = adp->ad_inodedep;
8993 	bsize = inodedep->id_fs->fs_bsize;
8994 	if (adp->ad_state & EXTDATA)
8995 		listhead = &inodedep->id_extupdt;
8996 	else
8997 		listhead = &inodedep->id_inoupdt;
8998 	TAILQ_FOREACH(listadp, listhead, ad_next) {
8999 		/* found our block */
9000 		if (listadp == adp)
9001 			break;
9002 		/* continue if ad_oldlbn is not a fragment */
9003 		if (listadp->ad_oldsize == 0 ||
9004 		    listadp->ad_oldsize == bsize)
9005 			continue;
9006 		/* hit a fragment */
9007 		return;
9008 	}
9009 	/*
9010 	 * If we have reached the end of the current list without
9011 	 * finding the just finished dependency, then it must be
9012 	 * on the future dependency list. Future dependencies cannot
9013 	 * be freed until they are moved to the current list.
9014 	 */
9015 	if (listadp == NULL) {
9016 #ifdef DEBUG
9017 		if (adp->ad_state & EXTDATA)
9018 			listhead = &inodedep->id_newextupdt;
9019 		else
9020 			listhead = &inodedep->id_newinoupdt;
9021 		TAILQ_FOREACH(listadp, listhead, ad_next)
9022 			/* found our block */
9023 			if (listadp == adp)
9024 				break;
9025 		if (listadp == NULL)
9026 			panic("handle_allocdirect_partdone: lost dep");
9027 #endif /* DEBUG */
9028 		return;
9029 	}
9030 	/*
9031 	 * If we have found the just finished dependency, then queue
9032 	 * it along with anything that follows it that is complete.
9033 	 * Since the pointer has not yet been written in the inode
9034 	 * as the dependency prevents it, place the allocdirect on the
9035 	 * bufwait list where it will be freed once the pointer is
9036 	 * valid.
9037 	 */
9038 	if (wkhd == NULL)
9039 		wkhd = &inodedep->id_bufwait;
9040 	for (; adp; adp = listadp) {
9041 		listadp = TAILQ_NEXT(adp, ad_next);
9042 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
9043 			return;
9044 		TAILQ_REMOVE(listhead, adp, ad_next);
9045 		WORKLIST_INSERT(wkhd, &adp->ad_block.nb_list);
9046 	}
9047 }
9048 
9049 /*
9050  * Called from within softdep_disk_write_complete above.  This routine
9051  * completes successfully written allocindirs.
9052  */
9053 static void
9054 handle_allocindir_partdone(aip)
9055 	struct allocindir *aip;		/* the completed allocindir */
9056 {
9057 	struct indirdep *indirdep;
9058 
9059 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
9060 		return;
9061 	indirdep = aip->ai_indirdep;
9062 	LIST_REMOVE(aip, ai_next);
9063 	if (indirdep->ir_state & UNDONE) {
9064 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
9065 		return;
9066 	}
9067 	if (indirdep->ir_state & UFS1FMT)
9068 		((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
9069 		    aip->ai_newblkno;
9070 	else
9071 		((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
9072 		    aip->ai_newblkno;
9073 	/*
9074 	 * Await the pointer write before freeing the allocindir.
9075 	 */
9076 	LIST_INSERT_HEAD(&indirdep->ir_writehd, aip, ai_next);
9077 }
9078 
9079 /*
9080  * Release segments held on a jwork list.
9081  */
9082 static void
9083 handle_jwork(wkhd)
9084 	struct workhead *wkhd;
9085 {
9086 	struct worklist *wk;
9087 
9088 	while ((wk = LIST_FIRST(wkhd)) != NULL) {
9089 		WORKLIST_REMOVE(wk);
9090 		switch (wk->wk_type) {
9091 		case D_JSEGDEP:
9092 			free_jsegdep(WK_JSEGDEP(wk));
9093 			continue;
9094 		default:
9095 			panic("handle_jwork: Unknown type %s\n",
9096 			    TYPENAME(wk->wk_type));
9097 		}
9098 	}
9099 }
9100 
9101 /*
9102  * Handle the bufwait list on an inode when it is safe to release items
9103  * held there.  This normally happens after an inode block is written but
9104  * may be delayed and handled later if there are pending journal items that
9105  * are not yet safe to be released.
9106  */
9107 static struct freefile *
9108 handle_bufwait(inodedep, refhd)
9109 	struct inodedep *inodedep;
9110 	struct workhead *refhd;
9111 {
9112 	struct jaddref *jaddref;
9113 	struct freefile *freefile;
9114 	struct worklist *wk;
9115 
9116 	freefile = NULL;
9117 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
9118 		WORKLIST_REMOVE(wk);
9119 		switch (wk->wk_type) {
9120 		case D_FREEFILE:
9121 			/*
9122 			 * We defer adding freefile to the worklist
9123 			 * until all other additions have been made to
9124 			 * ensure that it will be done after all the
9125 			 * old blocks have been freed.
9126 			 */
9127 			if (freefile != NULL)
9128 				panic("handle_bufwait: freefile");
9129 			freefile = WK_FREEFILE(wk);
9130 			continue;
9131 
9132 		case D_MKDIR:
9133 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
9134 			continue;
9135 
9136 		case D_DIRADD:
9137 			diradd_inode_written(WK_DIRADD(wk), inodedep);
9138 			continue;
9139 
9140 		case D_FREEFRAG:
9141 			wk->wk_state |= COMPLETE;
9142 			if ((wk->wk_state & ALLCOMPLETE) == ALLCOMPLETE)
9143 				add_to_worklist(wk, 0);
9144 			continue;
9145 
9146 		case D_DIRREM:
9147 			wk->wk_state |= COMPLETE;
9148 			add_to_worklist(wk, 0);
9149 			continue;
9150 
9151 		case D_ALLOCDIRECT:
9152 		case D_ALLOCINDIR:
9153 			free_newblk(WK_NEWBLK(wk));
9154 			continue;
9155 
9156 		case D_JNEWBLK:
9157 			wk->wk_state |= COMPLETE;
9158 			free_jnewblk(WK_JNEWBLK(wk));
9159 			continue;
9160 
9161 		/*
9162 		 * Save freed journal segments and add references on
9163 		 * the supplied list which will delay their release
9164 		 * until the cg bitmap is cleared on disk.
9165 		 */
9166 		case D_JSEGDEP:
9167 			if (refhd == NULL)
9168 				free_jsegdep(WK_JSEGDEP(wk));
9169 			else
9170 				WORKLIST_INSERT(refhd, wk);
9171 			continue;
9172 
9173 		case D_JADDREF:
9174 			jaddref = WK_JADDREF(wk);
9175 			TAILQ_REMOVE(&inodedep->id_inoreflst, &jaddref->ja_ref,
9176 			    if_deps);
9177 			/*
9178 			 * Transfer any jaddrefs to the list to be freed with
9179 			 * the bitmap if we're handling a removed file.
9180 			 */
9181 			if (refhd == NULL) {
9182 				wk->wk_state |= COMPLETE;
9183 				free_jaddref(jaddref);
9184 			} else
9185 				WORKLIST_INSERT(refhd, wk);
9186 			continue;
9187 
9188 		default:
9189 			panic("handle_bufwait: Unknown type %p(%s)",
9190 			    wk, TYPENAME(wk->wk_type));
9191 			/* NOTREACHED */
9192 		}
9193 	}
9194 	return (freefile);
9195 }
9196 /*
9197  * Called from within softdep_disk_write_complete above to restore
9198  * in-memory inode block contents to their most up-to-date state. Note
9199  * that this routine is always called from interrupt level with further
9200  * splbio interrupts blocked.
9201  */
9202 static int
9203 handle_written_inodeblock(inodedep, bp)
9204 	struct inodedep *inodedep;
9205 	struct buf *bp;		/* buffer containing the inode block */
9206 {
9207 	struct freefile *freefile;
9208 	struct allocdirect *adp, *nextadp;
9209 	struct ufs1_dinode *dp1 = NULL;
9210 	struct ufs2_dinode *dp2 = NULL;
9211 	struct workhead wkhd;
9212 	int hadchanges, fstype;
9213 	ino_t freelink;
9214 
9215 	LIST_INIT(&wkhd);
9216 	hadchanges = 0;
9217 	freefile = NULL;
9218 	if ((inodedep->id_state & IOSTARTED) == 0)
9219 		panic("handle_written_inodeblock: not started");
9220 	inodedep->id_state &= ~IOSTARTED;
9221 	if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
9222 		fstype = UFS1;
9223 		dp1 = (struct ufs1_dinode *)bp->b_data +
9224 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
9225 		freelink = dp1->di_freelink;
9226 	} else {
9227 		fstype = UFS2;
9228 		dp2 = (struct ufs2_dinode *)bp->b_data +
9229 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
9230 		freelink = dp2->di_freelink;
9231 	}
9232 	/*
9233 	 * If we wrote a valid freelink pointer during the last write
9234 	 * record it here.
9235 	 */
9236 	if ((inodedep->id_state & (UNLINKED | UNLINKNEXT)) == UNLINKED) {
9237 		struct inodedep *inon;
9238 
9239 		inon = TAILQ_NEXT(inodedep, id_unlinked);
9240 		if ((inon == NULL && freelink == 0) ||
9241 		    (inon && inon->id_ino == freelink)) {
9242 			if (inon)
9243 				inon->id_state |= UNLINKPREV;
9244 			inodedep->id_state |= UNLINKNEXT;
9245 		} else
9246 			hadchanges = 1;
9247 	}
9248 	/* Leave this inodeblock dirty until it's in the list. */
9249 	if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) == UNLINKED)
9250 		hadchanges = 1;
9251 	/*
9252 	 * If we had to rollback the inode allocation because of
9253 	 * bitmaps being incomplete, then simply restore it.
9254 	 * Keep the block dirty so that it will not be reclaimed until
9255 	 * all associated dependencies have been cleared and the
9256 	 * corresponding updates written to disk.
9257 	 */
9258 	if (inodedep->id_savedino1 != NULL) {
9259 		hadchanges = 1;
9260 		if (fstype == UFS1)
9261 			*dp1 = *inodedep->id_savedino1;
9262 		else
9263 			*dp2 = *inodedep->id_savedino2;
9264 		free(inodedep->id_savedino1, M_SAVEDINO);
9265 		inodedep->id_savedino1 = NULL;
9266 		if ((bp->b_flags & B_DELWRI) == 0)
9267 			stat_inode_bitmap++;
9268 		bdirty(bp);
9269 		/*
9270 		 * If the inode is clear here and GOINGAWAY it will never
9271 		 * be written.  Process the bufwait and clear any pending
9272 		 * work which may include the freefile.
9273 		 */
9274 		if (inodedep->id_state & GOINGAWAY)
9275 			goto bufwait;
9276 		return (1);
9277 	}
9278 	inodedep->id_state |= COMPLETE;
9279 	/*
9280 	 * Roll forward anything that had to be rolled back before
9281 	 * the inode could be updated.
9282 	 */
9283 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
9284 		nextadp = TAILQ_NEXT(adp, ad_next);
9285 		if (adp->ad_state & ATTACHED)
9286 			panic("handle_written_inodeblock: new entry");
9287 		if (fstype == UFS1) {
9288 			if (adp->ad_offset < NDADDR) {
9289 				if (dp1->di_db[adp->ad_offset]!=adp->ad_oldblkno)
9290 					panic("%s %s #%jd mismatch %d != %jd",
9291 					    "handle_written_inodeblock:",
9292 					    "direct pointer",
9293 					    (intmax_t)adp->ad_offset,
9294 					    dp1->di_db[adp->ad_offset],
9295 					    (intmax_t)adp->ad_oldblkno);
9296 				dp1->di_db[adp->ad_offset] = adp->ad_newblkno;
9297 			} else {
9298 				if (dp1->di_ib[adp->ad_offset - NDADDR] != 0)
9299 					panic("%s: %s #%jd allocated as %d",
9300 					    "handle_written_inodeblock",
9301 					    "indirect pointer",
9302 					    (intmax_t)adp->ad_offset - NDADDR,
9303 					    dp1->di_ib[adp->ad_offset - NDADDR]);
9304 				dp1->di_ib[adp->ad_offset - NDADDR] =
9305 				    adp->ad_newblkno;
9306 			}
9307 		} else {
9308 			if (adp->ad_offset < NDADDR) {
9309 				if (dp2->di_db[adp->ad_offset]!=adp->ad_oldblkno)
9310 					panic("%s: %s #%jd %s %jd != %jd",
9311 					    "handle_written_inodeblock",
9312 					    "direct pointer",
9313 					    (intmax_t)adp->ad_offset, "mismatch",
9314 					    (intmax_t)dp2->di_db[adp->ad_offset],
9315 					    (intmax_t)adp->ad_oldblkno);
9316 				dp2->di_db[adp->ad_offset] = adp->ad_newblkno;
9317 			} else {
9318 				if (dp2->di_ib[adp->ad_offset - NDADDR] != 0)
9319 					panic("%s: %s #%jd allocated as %jd",
9320 					    "handle_written_inodeblock",
9321 					    "indirect pointer",
9322 					    (intmax_t)adp->ad_offset - NDADDR,
9323 					    (intmax_t)
9324 					    dp2->di_ib[adp->ad_offset - NDADDR]);
9325 				dp2->di_ib[adp->ad_offset - NDADDR] =
9326 				    adp->ad_newblkno;
9327 			}
9328 		}
9329 		adp->ad_state &= ~UNDONE;
9330 		adp->ad_state |= ATTACHED;
9331 		hadchanges = 1;
9332 	}
9333 	for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
9334 		nextadp = TAILQ_NEXT(adp, ad_next);
9335 		if (adp->ad_state & ATTACHED)
9336 			panic("handle_written_inodeblock: new entry");
9337 		if (dp2->di_extb[adp->ad_offset] != adp->ad_oldblkno)
9338 			panic("%s: direct pointers #%jd %s %jd != %jd",
9339 			    "handle_written_inodeblock",
9340 			    (intmax_t)adp->ad_offset, "mismatch",
9341 			    (intmax_t)dp2->di_extb[adp->ad_offset],
9342 			    (intmax_t)adp->ad_oldblkno);
9343 		dp2->di_extb[adp->ad_offset] = adp->ad_newblkno;
9344 		adp->ad_state &= ~UNDONE;
9345 		adp->ad_state |= ATTACHED;
9346 		hadchanges = 1;
9347 	}
9348 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
9349 		stat_direct_blk_ptrs++;
9350 	/*
9351 	 * Reset the file size to its most up-to-date value.
9352 	 */
9353 	if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1)
9354 		panic("handle_written_inodeblock: bad size");
9355 	if (inodedep->id_savednlink > LINK_MAX)
9356 		panic("handle_written_inodeblock: Invalid link count "
9357 		    "%d for inodedep %p", inodedep->id_savednlink, inodedep);
9358 	if (fstype == UFS1) {
9359 		if (dp1->di_nlink != inodedep->id_savednlink) {
9360 			dp1->di_nlink = inodedep->id_savednlink;
9361 			hadchanges = 1;
9362 		}
9363 		if (dp1->di_size != inodedep->id_savedsize) {
9364 			dp1->di_size = inodedep->id_savedsize;
9365 			hadchanges = 1;
9366 		}
9367 	} else {
9368 		if (dp2->di_nlink != inodedep->id_savednlink) {
9369 			dp2->di_nlink = inodedep->id_savednlink;
9370 			hadchanges = 1;
9371 		}
9372 		if (dp2->di_size != inodedep->id_savedsize) {
9373 			dp2->di_size = inodedep->id_savedsize;
9374 			hadchanges = 1;
9375 		}
9376 		if (dp2->di_extsize != inodedep->id_savedextsize) {
9377 			dp2->di_extsize = inodedep->id_savedextsize;
9378 			hadchanges = 1;
9379 		}
9380 	}
9381 	inodedep->id_savedsize = -1;
9382 	inodedep->id_savedextsize = -1;
9383 	inodedep->id_savednlink = -1;
9384 	/*
9385 	 * If there were any rollbacks in the inode block, then it must be
9386 	 * marked dirty so that its will eventually get written back in
9387 	 * its correct form.
9388 	 */
9389 	if (hadchanges)
9390 		bdirty(bp);
9391 bufwait:
9392 	/*
9393 	 * Process any allocdirects that completed during the update.
9394 	 */
9395 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
9396 		handle_allocdirect_partdone(adp, &wkhd);
9397 	if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
9398 		handle_allocdirect_partdone(adp, &wkhd);
9399 	/*
9400 	 * Process deallocations that were held pending until the
9401 	 * inode had been written to disk. Freeing of the inode
9402 	 * is delayed until after all blocks have been freed to
9403 	 * avoid creation of new <vfsid, inum, lbn> triples
9404 	 * before the old ones have been deleted.  Completely
9405 	 * unlinked inodes are not processed until the unlinked
9406 	 * inode list is written or the last reference is removed.
9407 	 */
9408 	if ((inodedep->id_state & (UNLINKED | UNLINKONLIST)) != UNLINKED) {
9409 		freefile = handle_bufwait(inodedep, NULL);
9410 		if (freefile && !LIST_EMPTY(&wkhd)) {
9411 			WORKLIST_INSERT(&wkhd, &freefile->fx_list);
9412 			freefile = NULL;
9413 		}
9414 	}
9415 	/*
9416 	 * Move rolled forward dependency completions to the bufwait list
9417 	 * now that those that were already written have been processed.
9418 	 */
9419 	if (!LIST_EMPTY(&wkhd) && hadchanges == 0)
9420 		panic("handle_written_inodeblock: bufwait but no changes");
9421 	jwork_move(&inodedep->id_bufwait, &wkhd);
9422 
9423 	if (freefile != NULL) {
9424 		/*
9425 		 * If the inode is goingaway it was never written.  Fake up
9426 		 * the state here so free_inodedep() can succeed.
9427 		 */
9428 		if (inodedep->id_state & GOINGAWAY)
9429 			inodedep->id_state |= COMPLETE | DEPCOMPLETE;
9430 		if (free_inodedep(inodedep) == 0)
9431 			panic("handle_written_inodeblock: live inodedep %p",
9432 			    inodedep);
9433 		add_to_worklist(&freefile->fx_list, 0);
9434 		return (0);
9435 	}
9436 
9437 	/*
9438 	 * If no outstanding dependencies, free it.
9439 	 */
9440 	if (free_inodedep(inodedep) ||
9441 	    (TAILQ_FIRST(&inodedep->id_inoreflst) == 0 &&
9442 	     TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
9443 	     TAILQ_FIRST(&inodedep->id_extupdt) == 0 &&
9444 	     LIST_FIRST(&inodedep->id_bufwait) == 0))
9445 		return (0);
9446 	return (hadchanges);
9447 }
9448 
9449 static int
9450 handle_written_indirdep(indirdep, bp, bpp)
9451 	struct indirdep *indirdep;
9452 	struct buf *bp;
9453 	struct buf **bpp;
9454 {
9455 	struct allocindir *aip;
9456 	int chgs;
9457 
9458 	if (indirdep->ir_state & GOINGAWAY)
9459 		panic("disk_write_complete: indirdep gone");
9460 	chgs = 0;
9461 	/*
9462 	 * If there were rollbacks revert them here.
9463 	 */
9464 	if (indirdep->ir_saveddata) {
9465 		bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
9466 		free(indirdep->ir_saveddata, M_INDIRDEP);
9467 		indirdep->ir_saveddata = 0;
9468 		chgs = 1;
9469 	}
9470 	indirdep->ir_state &= ~UNDONE;
9471 	indirdep->ir_state |= ATTACHED;
9472 	/*
9473 	 * Move allocindirs with written pointers to the completehd if
9474 	 * the indirdep's pointer is not yet written.  Otherwise
9475 	 * free them here.
9476 	 */
9477 	while ((aip = LIST_FIRST(&indirdep->ir_writehd)) != 0) {
9478 		LIST_REMOVE(aip, ai_next);
9479 		if ((indirdep->ir_state & DEPCOMPLETE) == 0) {
9480 			LIST_INSERT_HEAD(&indirdep->ir_completehd, aip,
9481 			    ai_next);
9482 			continue;
9483 		}
9484 		free_newblk(&aip->ai_block);
9485 	}
9486 	/*
9487 	 * Move allocindirs that have finished dependency processing from
9488 	 * the done list to the write list after updating the pointers.
9489 	 */
9490 	while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
9491 		handle_allocindir_partdone(aip);
9492 		if (aip == LIST_FIRST(&indirdep->ir_donehd))
9493 			panic("disk_write_complete: not gone");
9494 		chgs = 1;
9495 	}
9496 	/*
9497 	 * If this indirdep has been detached from its newblk during
9498 	 * I/O we need to keep this dep attached to the buffer so
9499 	 * deallocate_dependencies can find it and properly resolve
9500 	 * any outstanding dependencies.
9501 	 */
9502 	if ((indirdep->ir_state & (ONDEPLIST | DEPCOMPLETE)) == 0)
9503 		chgs = 1;
9504 	if ((bp->b_flags & B_DELWRI) == 0)
9505 		stat_indir_blk_ptrs++;
9506 	/*
9507 	 * If there were no changes we can discard the savedbp and detach
9508 	 * ourselves from the buf.  We are only carrying completed pointers
9509 	 * in this case.
9510 	 */
9511 	if (chgs == 0) {
9512 		struct buf *sbp;
9513 
9514 		sbp = indirdep->ir_savebp;
9515 		sbp->b_flags |= B_INVAL | B_NOCACHE;
9516 		indirdep->ir_savebp = NULL;
9517 		if (*bpp != NULL)
9518 			panic("handle_written_indirdep: bp already exists.");
9519 		*bpp = sbp;
9520 	} else
9521 		bdirty(bp);
9522 	/*
9523 	 * If there are no fresh dependencies and none waiting on writes
9524 	 * we can free the indirdep.
9525 	 */
9526 	if ((indirdep->ir_state & DEPCOMPLETE) && chgs == 0) {
9527 		if (indirdep->ir_state & ONDEPLIST)
9528 			LIST_REMOVE(indirdep, ir_next);
9529 		free_indirdep(indirdep);
9530 		return (0);
9531 	}
9532 
9533 	return (chgs);
9534 }
9535 
9536 /*
9537  * Process a diradd entry after its dependent inode has been written.
9538  * This routine must be called with splbio interrupts blocked.
9539  */
9540 static void
9541 diradd_inode_written(dap, inodedep)
9542 	struct diradd *dap;
9543 	struct inodedep *inodedep;
9544 {
9545 
9546 	dap->da_state |= COMPLETE;
9547 	complete_diradd(dap);
9548 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
9549 }
9550 
9551 /*
9552  * Returns true if the bmsafemap will have rollbacks when written.  Must
9553  * only be called with lk and the buf lock on the cg held.
9554  */
9555 static int
9556 bmsafemap_rollbacks(bmsafemap)
9557 	struct bmsafemap *bmsafemap;
9558 {
9559 
9560 	return (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd) |
9561 	    !LIST_EMPTY(&bmsafemap->sm_jnewblkhd));
9562 }
9563 
9564 /*
9565  * Complete a write to a bmsafemap structure.  Roll forward any bitmap
9566  * changes if it's not a background write.  Set all written dependencies
9567  * to DEPCOMPLETE and free the structure if possible.
9568  */
9569 static int
9570 handle_written_bmsafemap(bmsafemap, bp)
9571 	struct bmsafemap *bmsafemap;
9572 	struct buf *bp;
9573 {
9574 	struct newblk *newblk;
9575 	struct inodedep *inodedep;
9576 	struct jaddref *jaddref, *jatmp;
9577 	struct jnewblk *jnewblk, *jntmp;
9578 	uint8_t *inosused;
9579 	uint8_t *blksfree;
9580 	struct cg *cgp;
9581 	struct fs *fs;
9582 	ino_t ino;
9583 	long bno;
9584 	int chgs;
9585 	int i;
9586 
9587 	if ((bmsafemap->sm_state & IOSTARTED) == 0)
9588 		panic("initiate_write_bmsafemap: Not started\n");
9589 	chgs = 0;
9590 	bmsafemap->sm_state &= ~IOSTARTED;
9591 	/*
9592 	 * Restore unwritten inode allocation pending jaddref writes.
9593 	 */
9594 	if (!LIST_EMPTY(&bmsafemap->sm_jaddrefhd)) {
9595 		cgp = (struct cg *)bp->b_data;
9596 		fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
9597 		inosused = cg_inosused(cgp);
9598 		LIST_FOREACH_SAFE(jaddref, &bmsafemap->sm_jaddrefhd,
9599 		    ja_bmdeps, jatmp) {
9600 			if ((jaddref->ja_state & UNDONE) == 0)
9601 				continue;
9602 			ino = jaddref->ja_ino % fs->fs_ipg;
9603 			if (isset(inosused, ino))
9604 				panic("handle_written_bmsafemap: "
9605 				    "re-allocated inode");
9606 			if ((bp->b_xflags & BX_BKGRDMARKER) == 0) {
9607 				if ((jaddref->ja_mode & IFMT) == IFDIR)
9608 					cgp->cg_cs.cs_ndir++;
9609 				cgp->cg_cs.cs_nifree--;
9610 				setbit(inosused, ino);
9611 				chgs = 1;
9612 			}
9613 			jaddref->ja_state &= ~UNDONE;
9614 			jaddref->ja_state |= ATTACHED;
9615 			free_jaddref(jaddref);
9616 		}
9617 	}
9618 	/*
9619 	 * Restore any block allocations which are pending journal writes.
9620 	 */
9621 	if (LIST_FIRST(&bmsafemap->sm_jnewblkhd) != NULL) {
9622 		cgp = (struct cg *)bp->b_data;
9623 		fs = VFSTOUFS(bmsafemap->sm_list.wk_mp)->um_fs;
9624 		blksfree = cg_blksfree(cgp);
9625 		LIST_FOREACH_SAFE(jnewblk, &bmsafemap->sm_jnewblkhd, jn_deps,
9626 		    jntmp) {
9627 			if ((jnewblk->jn_state & UNDONE) == 0)
9628 				continue;
9629 			bno = dtogd(fs, jnewblk->jn_blkno);
9630 			for (i = jnewblk->jn_oldfrags; i < jnewblk->jn_frags;
9631 			    i++) {
9632 				if (bp->b_xflags & BX_BKGRDMARKER)
9633 					break;
9634 				if ((jnewblk->jn_state & NEWBLOCK) == 0 &&
9635 				    isclr(blksfree, bno + i))
9636 					panic("handle_written_bmsafemap: "
9637 					    "re-allocated fragment");
9638 				clrbit(blksfree, bno + i);
9639 				chgs = 1;
9640 			}
9641 			jnewblk->jn_state &= ~(UNDONE | NEWBLOCK);
9642 			jnewblk->jn_state |= ATTACHED;
9643 			free_jnewblk(jnewblk);
9644 		}
9645 	}
9646 	while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkwr))) {
9647 		newblk->nb_state |= DEPCOMPLETE;
9648 		newblk->nb_state &= ~ONDEPLIST;
9649 		newblk->nb_bmsafemap = NULL;
9650 		LIST_REMOVE(newblk, nb_deps);
9651 		if (newblk->nb_list.wk_type == D_ALLOCDIRECT)
9652 			handle_allocdirect_partdone(
9653 			    WK_ALLOCDIRECT(&newblk->nb_list), NULL);
9654 		else if (newblk->nb_list.wk_type == D_ALLOCINDIR)
9655 			handle_allocindir_partdone(
9656 			    WK_ALLOCINDIR(&newblk->nb_list));
9657 		else if (newblk->nb_list.wk_type != D_NEWBLK)
9658 			panic("handle_written_bmsafemap: Unexpected type: %s",
9659 			    TYPENAME(newblk->nb_list.wk_type));
9660 	}
9661 	while ((inodedep = LIST_FIRST(&bmsafemap->sm_inodedepwr)) != NULL) {
9662 		inodedep->id_state |= DEPCOMPLETE;
9663 		inodedep->id_state &= ~ONDEPLIST;
9664 		LIST_REMOVE(inodedep, id_deps);
9665 		inodedep->id_bmsafemap = NULL;
9666 	}
9667 	if (LIST_EMPTY(&bmsafemap->sm_jaddrefhd) &&
9668 	    LIST_EMPTY(&bmsafemap->sm_jnewblkhd) &&
9669 	    LIST_EMPTY(&bmsafemap->sm_newblkhd) &&
9670 	    LIST_EMPTY(&bmsafemap->sm_inodedephd)) {
9671 		if (chgs)
9672 			bdirty(bp);
9673 		LIST_REMOVE(bmsafemap, sm_hash);
9674 		WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
9675 		return (0);
9676 	}
9677 	bdirty(bp);
9678 	return (1);
9679 }
9680 
9681 /*
9682  * Try to free a mkdir dependency.
9683  */
9684 static void
9685 complete_mkdir(mkdir)
9686 	struct mkdir *mkdir;
9687 {
9688 	struct diradd *dap;
9689 
9690 	if ((mkdir->md_state & ALLCOMPLETE) != ALLCOMPLETE)
9691 		return;
9692 	LIST_REMOVE(mkdir, md_mkdirs);
9693 	dap = mkdir->md_diradd;
9694 	dap->da_state &= ~(mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY));
9695 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0) {
9696 		dap->da_state |= DEPCOMPLETE;
9697 		complete_diradd(dap);
9698 	}
9699 	WORKITEM_FREE(mkdir, D_MKDIR);
9700 }
9701 
9702 /*
9703  * Handle the completion of a mkdir dependency.
9704  */
9705 static void
9706 handle_written_mkdir(mkdir, type)
9707 	struct mkdir *mkdir;
9708 	int type;
9709 {
9710 
9711 	if ((mkdir->md_state & (MKDIR_PARENT | MKDIR_BODY)) != type)
9712 		panic("handle_written_mkdir: bad type");
9713 	mkdir->md_state |= COMPLETE;
9714 	complete_mkdir(mkdir);
9715 }
9716 
9717 static void
9718 free_pagedep(pagedep)
9719 	struct pagedep *pagedep;
9720 {
9721 	int i;
9722 
9723 	if (pagedep->pd_state & (NEWBLOCK | ONWORKLIST))
9724 		return;
9725 	for (i = 0; i < DAHASHSZ; i++)
9726 		if (!LIST_EMPTY(&pagedep->pd_diraddhd[i]))
9727 			return;
9728 	if (!LIST_EMPTY(&pagedep->pd_jmvrefhd))
9729 		return;
9730 	if (!LIST_EMPTY(&pagedep->pd_dirremhd))
9731 		return;
9732 	if (!LIST_EMPTY(&pagedep->pd_pendinghd))
9733 		return;
9734 	LIST_REMOVE(pagedep, pd_hash);
9735 	WORKITEM_FREE(pagedep, D_PAGEDEP);
9736 }
9737 
9738 /*
9739  * Called from within softdep_disk_write_complete above.
9740  * A write operation was just completed. Removed inodes can
9741  * now be freed and associated block pointers may be committed.
9742  * Note that this routine is always called from interrupt level
9743  * with further splbio interrupts blocked.
9744  */
9745 static int
9746 handle_written_filepage(pagedep, bp)
9747 	struct pagedep *pagedep;
9748 	struct buf *bp;		/* buffer containing the written page */
9749 {
9750 	struct dirrem *dirrem;
9751 	struct diradd *dap, *nextdap;
9752 	struct direct *ep;
9753 	int i, chgs;
9754 
9755 	if ((pagedep->pd_state & IOSTARTED) == 0)
9756 		panic("handle_written_filepage: not started");
9757 	pagedep->pd_state &= ~IOSTARTED;
9758 	/*
9759 	 * Process any directory removals that have been committed.
9760 	 */
9761 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
9762 		LIST_REMOVE(dirrem, dm_next);
9763 		dirrem->dm_state |= COMPLETE;
9764 		dirrem->dm_dirinum = pagedep->pd_ino;
9765 		KASSERT(LIST_EMPTY(&dirrem->dm_jremrefhd),
9766 		    ("handle_written_filepage: Journal entries not written."));
9767 		add_to_worklist(&dirrem->dm_list, 0);
9768 	}
9769 	/*
9770 	 * Free any directory additions that have been committed.
9771 	 * If it is a newly allocated block, we have to wait until
9772 	 * the on-disk directory inode claims the new block.
9773 	 */
9774 	if ((pagedep->pd_state & NEWBLOCK) == 0)
9775 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
9776 			free_diradd(dap, NULL);
9777 	/*
9778 	 * Uncommitted directory entries must be restored.
9779 	 */
9780 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
9781 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
9782 		     dap = nextdap) {
9783 			nextdap = LIST_NEXT(dap, da_pdlist);
9784 			if (dap->da_state & ATTACHED)
9785 				panic("handle_written_filepage: attached");
9786 			ep = (struct direct *)
9787 			    ((char *)bp->b_data + dap->da_offset);
9788 			ep->d_ino = dap->da_newinum;
9789 			dap->da_state &= ~UNDONE;
9790 			dap->da_state |= ATTACHED;
9791 			chgs = 1;
9792 			/*
9793 			 * If the inode referenced by the directory has
9794 			 * been written out, then the dependency can be
9795 			 * moved to the pending list.
9796 			 */
9797 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
9798 				LIST_REMOVE(dap, da_pdlist);
9799 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
9800 				    da_pdlist);
9801 			}
9802 		}
9803 	}
9804 	/*
9805 	 * If there were any rollbacks in the directory, then it must be
9806 	 * marked dirty so that its will eventually get written back in
9807 	 * its correct form.
9808 	 */
9809 	if (chgs) {
9810 		if ((bp->b_flags & B_DELWRI) == 0)
9811 			stat_dir_entry++;
9812 		bdirty(bp);
9813 		return (1);
9814 	}
9815 	/*
9816 	 * If we are not waiting for a new directory block to be
9817 	 * claimed by its inode, then the pagedep will be freed.
9818 	 * Otherwise it will remain to track any new entries on
9819 	 * the page in case they are fsync'ed.
9820 	 */
9821 	if ((pagedep->pd_state & NEWBLOCK) == 0 &&
9822 	    LIST_EMPTY(&pagedep->pd_jmvrefhd)) {
9823 		LIST_REMOVE(pagedep, pd_hash);
9824 		WORKITEM_FREE(pagedep, D_PAGEDEP);
9825 	}
9826 	return (0);
9827 }
9828 
9829 /*
9830  * Writing back in-core inode structures.
9831  *
9832  * The filesystem only accesses an inode's contents when it occupies an
9833  * "in-core" inode structure.  These "in-core" structures are separate from
9834  * the page frames used to cache inode blocks.  Only the latter are
9835  * transferred to/from the disk.  So, when the updated contents of the
9836  * "in-core" inode structure are copied to the corresponding in-memory inode
9837  * block, the dependencies are also transferred.  The following procedure is
9838  * called when copying a dirty "in-core" inode to a cached inode block.
9839  */
9840 
9841 /*
9842  * Called when an inode is loaded from disk. If the effective link count
9843  * differed from the actual link count when it was last flushed, then we
9844  * need to ensure that the correct effective link count is put back.
9845  */
9846 void
9847 softdep_load_inodeblock(ip)
9848 	struct inode *ip;	/* the "in_core" copy of the inode */
9849 {
9850 	struct inodedep *inodedep;
9851 
9852 	/*
9853 	 * Check for alternate nlink count.
9854 	 */
9855 	ip->i_effnlink = ip->i_nlink;
9856 	ACQUIRE_LOCK(&lk);
9857 	if (inodedep_lookup(UFSTOVFS(ip->i_ump), ip->i_number, 0,
9858 	    &inodedep) == 0) {
9859 		FREE_LOCK(&lk);
9860 		return;
9861 	}
9862 	ip->i_effnlink -= inodedep->id_nlinkdelta;
9863 	FREE_LOCK(&lk);
9864 }
9865 
9866 /*
9867  * This routine is called just before the "in-core" inode
9868  * information is to be copied to the in-memory inode block.
9869  * Recall that an inode block contains several inodes. If
9870  * the force flag is set, then the dependencies will be
9871  * cleared so that the update can always be made. Note that
9872  * the buffer is locked when this routine is called, so we
9873  * will never be in the middle of writing the inode block
9874  * to disk.
9875  */
9876 void
9877 softdep_update_inodeblock(ip, bp, waitfor)
9878 	struct inode *ip;	/* the "in_core" copy of the inode */
9879 	struct buf *bp;		/* the buffer containing the inode block */
9880 	int waitfor;		/* nonzero => update must be allowed */
9881 {
9882 	struct inodedep *inodedep;
9883 	struct inoref *inoref;
9884 	struct worklist *wk;
9885 	struct mount *mp;
9886 	struct buf *ibp;
9887 	struct fs *fs;
9888 	int error;
9889 
9890 	mp = UFSTOVFS(ip->i_ump);
9891 	fs = ip->i_fs;
9892 	/*
9893 	 * Preserve the freelink that is on disk.  clear_unlinked_inodedep()
9894 	 * does not have access to the in-core ip so must write directly into
9895 	 * the inode block buffer when setting freelink.
9896 	 */
9897 	if (fs->fs_magic == FS_UFS1_MAGIC)
9898 		DIP_SET(ip, i_freelink, ((struct ufs1_dinode *)bp->b_data +
9899 		    ino_to_fsbo(fs, ip->i_number))->di_freelink);
9900 	else
9901 		DIP_SET(ip, i_freelink, ((struct ufs2_dinode *)bp->b_data +
9902 		    ino_to_fsbo(fs, ip->i_number))->di_freelink);
9903 	/*
9904 	 * If the effective link count is not equal to the actual link
9905 	 * count, then we must track the difference in an inodedep while
9906 	 * the inode is (potentially) tossed out of the cache. Otherwise,
9907 	 * if there is no existing inodedep, then there are no dependencies
9908 	 * to track.
9909 	 */
9910 	ACQUIRE_LOCK(&lk);
9911 again:
9912 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
9913 		FREE_LOCK(&lk);
9914 		if (ip->i_effnlink != ip->i_nlink)
9915 			panic("softdep_update_inodeblock: bad link count");
9916 		return;
9917 	}
9918 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
9919 		panic("softdep_update_inodeblock: bad delta");
9920 	/*
9921 	 * If we're flushing all dependencies we must also move any waiting
9922 	 * for journal writes onto the bufwait list prior to I/O.
9923 	 */
9924 	if (waitfor) {
9925 		TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
9926 			if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
9927 			    == DEPCOMPLETE) {
9928 				stat_jwait_inode++;
9929 				jwait(&inoref->if_list);
9930 				goto again;
9931 			}
9932 		}
9933 	}
9934 	/*
9935 	 * Changes have been initiated. Anything depending on these
9936 	 * changes cannot occur until this inode has been written.
9937 	 */
9938 	inodedep->id_state &= ~COMPLETE;
9939 	if ((inodedep->id_state & ONWORKLIST) == 0)
9940 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
9941 	/*
9942 	 * Any new dependencies associated with the incore inode must
9943 	 * now be moved to the list associated with the buffer holding
9944 	 * the in-memory copy of the inode. Once merged process any
9945 	 * allocdirects that are completed by the merger.
9946 	 */
9947 	merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
9948 	if (!TAILQ_EMPTY(&inodedep->id_inoupdt))
9949 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt),
9950 		    NULL);
9951 	merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
9952 	if (!TAILQ_EMPTY(&inodedep->id_extupdt))
9953 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt),
9954 		    NULL);
9955 	/*
9956 	 * Now that the inode has been pushed into the buffer, the
9957 	 * operations dependent on the inode being written to disk
9958 	 * can be moved to the id_bufwait so that they will be
9959 	 * processed when the buffer I/O completes.
9960 	 */
9961 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
9962 		WORKLIST_REMOVE(wk);
9963 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
9964 	}
9965 	/*
9966 	 * Newly allocated inodes cannot be written until the bitmap
9967 	 * that allocates them have been written (indicated by
9968 	 * DEPCOMPLETE being set in id_state). If we are doing a
9969 	 * forced sync (e.g., an fsync on a file), we force the bitmap
9970 	 * to be written so that the update can be done.
9971 	 */
9972 	if (waitfor == 0) {
9973 		FREE_LOCK(&lk);
9974 		return;
9975 	}
9976 retry:
9977 	if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) != 0) {
9978 		FREE_LOCK(&lk);
9979 		return;
9980 	}
9981 	ibp = inodedep->id_bmsafemap->sm_buf;
9982 	ibp = getdirtybuf(ibp, &lk, MNT_WAIT);
9983 	if (ibp == NULL) {
9984 		/*
9985 		 * If ibp came back as NULL, the dependency could have been
9986 		 * freed while we slept.  Look it up again, and check to see
9987 		 * that it has completed.
9988 		 */
9989 		if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0)
9990 			goto retry;
9991 		FREE_LOCK(&lk);
9992 		return;
9993 	}
9994 	FREE_LOCK(&lk);
9995 	if ((error = bwrite(ibp)) != 0)
9996 		softdep_error("softdep_update_inodeblock: bwrite", error);
9997 }
9998 
9999 /*
10000  * Merge the a new inode dependency list (such as id_newinoupdt) into an
10001  * old inode dependency list (such as id_inoupdt). This routine must be
10002  * called with splbio interrupts blocked.
10003  */
10004 static void
10005 merge_inode_lists(newlisthead, oldlisthead)
10006 	struct allocdirectlst *newlisthead;
10007 	struct allocdirectlst *oldlisthead;
10008 {
10009 	struct allocdirect *listadp, *newadp;
10010 
10011 	newadp = TAILQ_FIRST(newlisthead);
10012 	for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
10013 		if (listadp->ad_offset < newadp->ad_offset) {
10014 			listadp = TAILQ_NEXT(listadp, ad_next);
10015 			continue;
10016 		}
10017 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
10018 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
10019 		if (listadp->ad_offset == newadp->ad_offset) {
10020 			allocdirect_merge(oldlisthead, newadp,
10021 			    listadp);
10022 			listadp = newadp;
10023 		}
10024 		newadp = TAILQ_FIRST(newlisthead);
10025 	}
10026 	while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
10027 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
10028 		TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
10029 	}
10030 }
10031 
10032 /*
10033  * If we are doing an fsync, then we must ensure that any directory
10034  * entries for the inode have been written after the inode gets to disk.
10035  */
10036 int
10037 softdep_fsync(vp)
10038 	struct vnode *vp;	/* the "in_core" copy of the inode */
10039 {
10040 	struct inodedep *inodedep;
10041 	struct pagedep *pagedep;
10042 	struct inoref *inoref;
10043 	struct worklist *wk;
10044 	struct diradd *dap;
10045 	struct mount *mp;
10046 	struct vnode *pvp;
10047 	struct inode *ip;
10048 	struct buf *bp;
10049 	struct fs *fs;
10050 	struct thread *td = curthread;
10051 	int error, flushparent, pagedep_new_block;
10052 	ino_t parentino;
10053 	ufs_lbn_t lbn;
10054 
10055 	ip = VTOI(vp);
10056 	fs = ip->i_fs;
10057 	mp = vp->v_mount;
10058 	ACQUIRE_LOCK(&lk);
10059 restart:
10060 	if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0) {
10061 		FREE_LOCK(&lk);
10062 		return (0);
10063 	}
10064 	TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
10065 		if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
10066 		    == DEPCOMPLETE) {
10067 			stat_jwait_inode++;
10068 			jwait(&inoref->if_list);
10069 			goto restart;
10070 		}
10071 	}
10072 	if (!LIST_EMPTY(&inodedep->id_inowait) ||
10073 	    !TAILQ_EMPTY(&inodedep->id_extupdt) ||
10074 	    !TAILQ_EMPTY(&inodedep->id_newextupdt) ||
10075 	    !TAILQ_EMPTY(&inodedep->id_inoupdt) ||
10076 	    !TAILQ_EMPTY(&inodedep->id_newinoupdt))
10077 		panic("softdep_fsync: pending ops %p", inodedep);
10078 	for (error = 0, flushparent = 0; ; ) {
10079 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
10080 			break;
10081 		if (wk->wk_type != D_DIRADD)
10082 			panic("softdep_fsync: Unexpected type %s",
10083 			    TYPENAME(wk->wk_type));
10084 		dap = WK_DIRADD(wk);
10085 		/*
10086 		 * Flush our parent if this directory entry has a MKDIR_PARENT
10087 		 * dependency or is contained in a newly allocated block.
10088 		 */
10089 		if (dap->da_state & DIRCHG)
10090 			pagedep = dap->da_previous->dm_pagedep;
10091 		else
10092 			pagedep = dap->da_pagedep;
10093 		parentino = pagedep->pd_ino;
10094 		lbn = pagedep->pd_lbn;
10095 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
10096 			panic("softdep_fsync: dirty");
10097 		if ((dap->da_state & MKDIR_PARENT) ||
10098 		    (pagedep->pd_state & NEWBLOCK))
10099 			flushparent = 1;
10100 		else
10101 			flushparent = 0;
10102 		/*
10103 		 * If we are being fsync'ed as part of vgone'ing this vnode,
10104 		 * then we will not be able to release and recover the
10105 		 * vnode below, so we just have to give up on writing its
10106 		 * directory entry out. It will eventually be written, just
10107 		 * not now, but then the user was not asking to have it
10108 		 * written, so we are not breaking any promises.
10109 		 */
10110 		if (vp->v_iflag & VI_DOOMED)
10111 			break;
10112 		/*
10113 		 * We prevent deadlock by always fetching inodes from the
10114 		 * root, moving down the directory tree. Thus, when fetching
10115 		 * our parent directory, we first try to get the lock. If
10116 		 * that fails, we must unlock ourselves before requesting
10117 		 * the lock on our parent. See the comment in ufs_lookup
10118 		 * for details on possible races.
10119 		 */
10120 		FREE_LOCK(&lk);
10121 		if (ffs_vgetf(mp, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp,
10122 		    FFSV_FORCEINSMQ)) {
10123 			error = vfs_busy(mp, MBF_NOWAIT);
10124 			if (error != 0) {
10125 				vfs_ref(mp);
10126 				VOP_UNLOCK(vp, 0);
10127 				error = vfs_busy(mp, 0);
10128 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
10129 				vfs_rel(mp);
10130 				if (error != 0)
10131 					return (ENOENT);
10132 				if (vp->v_iflag & VI_DOOMED) {
10133 					vfs_unbusy(mp);
10134 					return (ENOENT);
10135 				}
10136 			}
10137 			VOP_UNLOCK(vp, 0);
10138 			error = ffs_vgetf(mp, parentino, LK_EXCLUSIVE,
10139 			    &pvp, FFSV_FORCEINSMQ);
10140 			vfs_unbusy(mp);
10141 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
10142 			if (vp->v_iflag & VI_DOOMED) {
10143 				if (error == 0)
10144 					vput(pvp);
10145 				error = ENOENT;
10146 			}
10147 			if (error != 0)
10148 				return (error);
10149 		}
10150 		/*
10151 		 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
10152 		 * that are contained in direct blocks will be resolved by
10153 		 * doing a ffs_update. Pagedeps contained in indirect blocks
10154 		 * may require a complete sync'ing of the directory. So, we
10155 		 * try the cheap and fast ffs_update first, and if that fails,
10156 		 * then we do the slower ffs_syncvnode of the directory.
10157 		 */
10158 		if (flushparent) {
10159 			int locked;
10160 
10161 			if ((error = ffs_update(pvp, 1)) != 0) {
10162 				vput(pvp);
10163 				return (error);
10164 			}
10165 			ACQUIRE_LOCK(&lk);
10166 			locked = 1;
10167 			if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) != 0) {
10168 				if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) != NULL) {
10169 					if (wk->wk_type != D_DIRADD)
10170 						panic("softdep_fsync: Unexpected type %s",
10171 						      TYPENAME(wk->wk_type));
10172 					dap = WK_DIRADD(wk);
10173 					if (dap->da_state & DIRCHG)
10174 						pagedep = dap->da_previous->dm_pagedep;
10175 					else
10176 						pagedep = dap->da_pagedep;
10177 					pagedep_new_block = pagedep->pd_state & NEWBLOCK;
10178 					FREE_LOCK(&lk);
10179 					locked = 0;
10180 					if (pagedep_new_block &&
10181 					    (error = ffs_syncvnode(pvp, MNT_WAIT))) {
10182 						vput(pvp);
10183 						return (error);
10184 					}
10185 				}
10186 			}
10187 			if (locked)
10188 				FREE_LOCK(&lk);
10189 		}
10190 		/*
10191 		 * Flush directory page containing the inode's name.
10192 		 */
10193 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
10194 		    &bp);
10195 		if (error == 0)
10196 			error = bwrite(bp);
10197 		else
10198 			brelse(bp);
10199 		vput(pvp);
10200 		if (error != 0)
10201 			return (error);
10202 		ACQUIRE_LOCK(&lk);
10203 		if (inodedep_lookup(mp, ip->i_number, 0, &inodedep) == 0)
10204 			break;
10205 	}
10206 	FREE_LOCK(&lk);
10207 	return (0);
10208 }
10209 
10210 /*
10211  * Flush all the dirty bitmaps associated with the block device
10212  * before flushing the rest of the dirty blocks so as to reduce
10213  * the number of dependencies that will have to be rolled back.
10214  */
10215 void
10216 softdep_fsync_mountdev(vp)
10217 	struct vnode *vp;
10218 {
10219 	struct buf *bp, *nbp;
10220 	struct worklist *wk;
10221 	struct bufobj *bo;
10222 
10223 	if (!vn_isdisk(vp, NULL))
10224 		panic("softdep_fsync_mountdev: vnode not a disk");
10225 	bo = &vp->v_bufobj;
10226 restart:
10227 	BO_LOCK(bo);
10228 	ACQUIRE_LOCK(&lk);
10229 	TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
10230 		/*
10231 		 * If it is already scheduled, skip to the next buffer.
10232 		 */
10233 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
10234 			continue;
10235 
10236 		if ((bp->b_flags & B_DELWRI) == 0)
10237 			panic("softdep_fsync_mountdev: not dirty");
10238 		/*
10239 		 * We are only interested in bitmaps with outstanding
10240 		 * dependencies.
10241 		 */
10242 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
10243 		    wk->wk_type != D_BMSAFEMAP ||
10244 		    (bp->b_vflags & BV_BKGRDINPROG)) {
10245 			BUF_UNLOCK(bp);
10246 			continue;
10247 		}
10248 		FREE_LOCK(&lk);
10249 		BO_UNLOCK(bo);
10250 		bremfree(bp);
10251 		(void) bawrite(bp);
10252 		goto restart;
10253 	}
10254 	FREE_LOCK(&lk);
10255 	drain_output(vp);
10256 	BO_UNLOCK(bo);
10257 }
10258 
10259 /*
10260  * This routine is called when we are trying to synchronously flush a
10261  * file. This routine must eliminate any filesystem metadata dependencies
10262  * so that the syncing routine can succeed by pushing the dirty blocks
10263  * associated with the file. If any I/O errors occur, they are returned.
10264  */
10265 int
10266 softdep_sync_metadata(struct vnode *vp)
10267 {
10268 	struct pagedep *pagedep;
10269 	struct allocindir *aip;
10270 	struct newblk *newblk;
10271 	struct buf *bp, *nbp;
10272 	struct worklist *wk;
10273 	struct bufobj *bo;
10274 	int i, error, waitfor;
10275 
10276 	if (!DOINGSOFTDEP(vp))
10277 		return (0);
10278 	/*
10279 	 * Ensure that any direct block dependencies have been cleared.
10280 	 */
10281 	ACQUIRE_LOCK(&lk);
10282 	if ((error = flush_inodedep_deps(vp->v_mount, VTOI(vp)->i_number))) {
10283 		FREE_LOCK(&lk);
10284 		return (error);
10285 	}
10286 	FREE_LOCK(&lk);
10287 	/*
10288 	 * For most files, the only metadata dependencies are the
10289 	 * cylinder group maps that allocate their inode or blocks.
10290 	 * The block allocation dependencies can be found by traversing
10291 	 * the dependency lists for any buffers that remain on their
10292 	 * dirty buffer list. The inode allocation dependency will
10293 	 * be resolved when the inode is updated with MNT_WAIT.
10294 	 * This work is done in two passes. The first pass grabs most
10295 	 * of the buffers and begins asynchronously writing them. The
10296 	 * only way to wait for these asynchronous writes is to sleep
10297 	 * on the filesystem vnode which may stay busy for a long time
10298 	 * if the filesystem is active. So, instead, we make a second
10299 	 * pass over the dependencies blocking on each write. In the
10300 	 * usual case we will be blocking against a write that we
10301 	 * initiated, so when it is done the dependency will have been
10302 	 * resolved. Thus the second pass is expected to end quickly.
10303 	 */
10304 	waitfor = MNT_NOWAIT;
10305 	bo = &vp->v_bufobj;
10306 
10307 top:
10308 	/*
10309 	 * We must wait for any I/O in progress to finish so that
10310 	 * all potential buffers on the dirty list will be visible.
10311 	 */
10312 	BO_LOCK(bo);
10313 	drain_output(vp);
10314 	while ((bp = TAILQ_FIRST(&bo->bo_dirty.bv_hd)) != NULL) {
10315 		bp = getdirtybuf(bp, BO_MTX(bo), MNT_WAIT);
10316 		if (bp)
10317 			break;
10318 	}
10319 	BO_UNLOCK(bo);
10320 	if (bp == NULL)
10321 		return (0);
10322 loop:
10323 	/* While syncing snapshots, we must allow recursive lookups */
10324 	BUF_AREC(bp);
10325 	ACQUIRE_LOCK(&lk);
10326 	/*
10327 	 * As we hold the buffer locked, none of its dependencies
10328 	 * will disappear.
10329 	 */
10330 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
10331 		switch (wk->wk_type) {
10332 
10333 		case D_ALLOCDIRECT:
10334 		case D_ALLOCINDIR:
10335 			newblk = WK_NEWBLK(wk);
10336 			if (newblk->nb_jnewblk != NULL) {
10337 				stat_jwait_newblk++;
10338 				jwait(&newblk->nb_jnewblk->jn_list);
10339 				goto restart;
10340 			}
10341 			if (newblk->nb_state & DEPCOMPLETE)
10342 				continue;
10343 			nbp = newblk->nb_bmsafemap->sm_buf;
10344 			nbp = getdirtybuf(nbp, &lk, waitfor);
10345 			if (nbp == NULL)
10346 				continue;
10347 			FREE_LOCK(&lk);
10348 			if (waitfor == MNT_NOWAIT) {
10349 				bawrite(nbp);
10350 			} else if ((error = bwrite(nbp)) != 0) {
10351 				break;
10352 			}
10353 			ACQUIRE_LOCK(&lk);
10354 			continue;
10355 
10356 		case D_INDIRDEP:
10357 		restart:
10358 
10359 			LIST_FOREACH(aip,
10360 			    &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
10361 				newblk = (struct newblk *)aip;
10362 				if (newblk->nb_jnewblk != NULL) {
10363 					stat_jwait_newblk++;
10364 					jwait(&newblk->nb_jnewblk->jn_list);
10365 					goto restart;
10366 				}
10367 				if (newblk->nb_state & DEPCOMPLETE)
10368 					continue;
10369 				nbp = newblk->nb_bmsafemap->sm_buf;
10370 				nbp = getdirtybuf(nbp, &lk, MNT_WAIT);
10371 				if (nbp == NULL)
10372 					goto restart;
10373 				FREE_LOCK(&lk);
10374 				if ((error = bwrite(nbp)) != 0) {
10375 					goto loop_end;
10376 				}
10377 				ACQUIRE_LOCK(&lk);
10378 				goto restart;
10379 			}
10380 			continue;
10381 
10382 		case D_PAGEDEP:
10383 			/*
10384 			 * We are trying to sync a directory that may
10385 			 * have dependencies on both its own metadata
10386 			 * and/or dependencies on the inodes of any
10387 			 * recently allocated files. We walk its diradd
10388 			 * lists pushing out the associated inode.
10389 			 */
10390 			pagedep = WK_PAGEDEP(wk);
10391 			for (i = 0; i < DAHASHSZ; i++) {
10392 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
10393 					continue;
10394 				if ((error =
10395 				    flush_pagedep_deps(vp, wk->wk_mp,
10396 						&pagedep->pd_diraddhd[i]))) {
10397 					FREE_LOCK(&lk);
10398 					goto loop_end;
10399 				}
10400 			}
10401 			continue;
10402 
10403 		default:
10404 			panic("softdep_sync_metadata: Unknown type %s",
10405 			    TYPENAME(wk->wk_type));
10406 			/* NOTREACHED */
10407 		}
10408 	loop_end:
10409 		/* We reach here only in error and unlocked */
10410 		if (error == 0)
10411 			panic("softdep_sync_metadata: zero error");
10412 		BUF_NOREC(bp);
10413 		bawrite(bp);
10414 		return (error);
10415 	}
10416 	FREE_LOCK(&lk);
10417 	BO_LOCK(bo);
10418 	while ((nbp = TAILQ_NEXT(bp, b_bobufs)) != NULL) {
10419 		nbp = getdirtybuf(nbp, BO_MTX(bo), MNT_WAIT);
10420 		if (nbp)
10421 			break;
10422 	}
10423 	BO_UNLOCK(bo);
10424 	BUF_NOREC(bp);
10425 	bawrite(bp);
10426 	if (nbp != NULL) {
10427 		bp = nbp;
10428 		goto loop;
10429 	}
10430 	/*
10431 	 * The brief unlock is to allow any pent up dependency
10432 	 * processing to be done. Then proceed with the second pass.
10433 	 */
10434 	if (waitfor == MNT_NOWAIT) {
10435 		waitfor = MNT_WAIT;
10436 		goto top;
10437 	}
10438 
10439 	/*
10440 	 * If we have managed to get rid of all the dirty buffers,
10441 	 * then we are done. For certain directories and block
10442 	 * devices, we may need to do further work.
10443 	 *
10444 	 * We must wait for any I/O in progress to finish so that
10445 	 * all potential buffers on the dirty list will be visible.
10446 	 */
10447 	BO_LOCK(bo);
10448 	drain_output(vp);
10449 	BO_UNLOCK(bo);
10450 	return ffs_update(vp, 1);
10451 	/* return (0); */
10452 }
10453 
10454 /*
10455  * Flush the dependencies associated with an inodedep.
10456  * Called with splbio blocked.
10457  */
10458 static int
10459 flush_inodedep_deps(mp, ino)
10460 	struct mount *mp;
10461 	ino_t ino;
10462 {
10463 	struct inodedep *inodedep;
10464 	struct inoref *inoref;
10465 	int error, waitfor;
10466 
10467 	/*
10468 	 * This work is done in two passes. The first pass grabs most
10469 	 * of the buffers and begins asynchronously writing them. The
10470 	 * only way to wait for these asynchronous writes is to sleep
10471 	 * on the filesystem vnode which may stay busy for a long time
10472 	 * if the filesystem is active. So, instead, we make a second
10473 	 * pass over the dependencies blocking on each write. In the
10474 	 * usual case we will be blocking against a write that we
10475 	 * initiated, so when it is done the dependency will have been
10476 	 * resolved. Thus the second pass is expected to end quickly.
10477 	 * We give a brief window at the top of the loop to allow
10478 	 * any pending I/O to complete.
10479 	 */
10480 	for (error = 0, waitfor = MNT_NOWAIT; ; ) {
10481 		if (error)
10482 			return (error);
10483 		FREE_LOCK(&lk);
10484 		ACQUIRE_LOCK(&lk);
10485 restart:
10486 		if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
10487 			return (0);
10488 		TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
10489 			if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
10490 			    == DEPCOMPLETE) {
10491 				stat_jwait_inode++;
10492 				jwait(&inoref->if_list);
10493 				goto restart;
10494 			}
10495 		}
10496 		if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
10497 		    flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
10498 		    flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
10499 		    flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
10500 			continue;
10501 		/*
10502 		 * If pass2, we are done, otherwise do pass 2.
10503 		 */
10504 		if (waitfor == MNT_WAIT)
10505 			break;
10506 		waitfor = MNT_WAIT;
10507 	}
10508 	/*
10509 	 * Try freeing inodedep in case all dependencies have been removed.
10510 	 */
10511 	if (inodedep_lookup(mp, ino, 0, &inodedep) != 0)
10512 		(void) free_inodedep(inodedep);
10513 	return (0);
10514 }
10515 
10516 /*
10517  * Flush an inode dependency list.
10518  * Called with splbio blocked.
10519  */
10520 static int
10521 flush_deplist(listhead, waitfor, errorp)
10522 	struct allocdirectlst *listhead;
10523 	int waitfor;
10524 	int *errorp;
10525 {
10526 	struct allocdirect *adp;
10527 	struct newblk *newblk;
10528 	struct buf *bp;
10529 
10530 	mtx_assert(&lk, MA_OWNED);
10531 	TAILQ_FOREACH(adp, listhead, ad_next) {
10532 		newblk = (struct newblk *)adp;
10533 		if (newblk->nb_jnewblk != NULL) {
10534 			stat_jwait_newblk++;
10535 			jwait(&newblk->nb_jnewblk->jn_list);
10536 			return (1);
10537 		}
10538 		if (newblk->nb_state & DEPCOMPLETE)
10539 			continue;
10540 		bp = newblk->nb_bmsafemap->sm_buf;
10541 		bp = getdirtybuf(bp, &lk, waitfor);
10542 		if (bp == NULL) {
10543 			if (waitfor == MNT_NOWAIT)
10544 				continue;
10545 			return (1);
10546 		}
10547 		FREE_LOCK(&lk);
10548 		if (waitfor == MNT_NOWAIT) {
10549 			bawrite(bp);
10550 		} else if ((*errorp = bwrite(bp)) != 0) {
10551 			ACQUIRE_LOCK(&lk);
10552 			return (1);
10553 		}
10554 		ACQUIRE_LOCK(&lk);
10555 		return (1);
10556 	}
10557 	return (0);
10558 }
10559 
10560 /*
10561  * Flush dependencies associated with an allocdirect block.
10562  */
10563 static int
10564 flush_newblk_dep(vp, mp, lbn)
10565 	struct vnode *vp;
10566 	struct mount *mp;
10567 	ufs_lbn_t lbn;
10568 {
10569 	struct newblk *newblk;
10570 	struct bufobj *bo;
10571 	struct inode *ip;
10572 	struct buf *bp;
10573 	ufs2_daddr_t blkno;
10574 	int error;
10575 
10576 	error = 0;
10577 	bo = &vp->v_bufobj;
10578 	ip = VTOI(vp);
10579 	blkno = DIP(ip, i_db[lbn]);
10580 	if (blkno == 0)
10581 		panic("flush_newblk_dep: Missing block");
10582 	ACQUIRE_LOCK(&lk);
10583 	/*
10584 	 * Loop until all dependencies related to this block are satisfied.
10585 	 * We must be careful to restart after each sleep in case a write
10586 	 * completes some part of this process for us.
10587 	 */
10588 	for (;;) {
10589 		if (newblk_lookup(mp, blkno, 0, &newblk) == 0) {
10590 			FREE_LOCK(&lk);
10591 			break;
10592 		}
10593 		if (newblk->nb_list.wk_type != D_ALLOCDIRECT)
10594 			panic("flush_newblk_deps: Bad newblk %p", newblk);
10595 		/*
10596 		 * Flush the journal.
10597 		 */
10598 		if (newblk->nb_jnewblk != NULL) {
10599 			stat_jwait_newblk++;
10600 			jwait(&newblk->nb_jnewblk->jn_list);
10601 			continue;
10602 		}
10603 		/*
10604 		 * Write the bitmap dependency.
10605 		 */
10606 		if ((newblk->nb_state & DEPCOMPLETE) == 0) {
10607 			bp = newblk->nb_bmsafemap->sm_buf;
10608 			bp = getdirtybuf(bp, &lk, MNT_WAIT);
10609 			if (bp == NULL)
10610 				continue;
10611 			FREE_LOCK(&lk);
10612 			error = bwrite(bp);
10613 			if (error)
10614 				break;
10615 			ACQUIRE_LOCK(&lk);
10616 			continue;
10617 		}
10618 		/*
10619 		 * Write the buffer.
10620 		 */
10621 		FREE_LOCK(&lk);
10622 		BO_LOCK(bo);
10623 		bp = gbincore(bo, lbn);
10624 		if (bp != NULL) {
10625 			error = BUF_LOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
10626 			    LK_INTERLOCK, BO_MTX(bo));
10627 			if (error == ENOLCK) {
10628 				ACQUIRE_LOCK(&lk);
10629 				continue; /* Slept, retry */
10630 			}
10631 			if (error != 0)
10632 				break;	/* Failed */
10633 			if (bp->b_flags & B_DELWRI) {
10634 				bremfree(bp);
10635 				error = bwrite(bp);
10636 				if (error)
10637 					break;
10638 			} else
10639 				BUF_UNLOCK(bp);
10640 		} else
10641 			BO_UNLOCK(bo);
10642 		/*
10643 		 * We have to wait for the direct pointers to
10644 		 * point at the newdirblk before the dependency
10645 		 * will go away.
10646 		 */
10647 		error = ffs_update(vp, MNT_WAIT);
10648 		if (error)
10649 			break;
10650 		ACQUIRE_LOCK(&lk);
10651 	}
10652 	return (error);
10653 }
10654 
10655 /*
10656  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
10657  * Called with splbio blocked.
10658  */
10659 static int
10660 flush_pagedep_deps(pvp, mp, diraddhdp)
10661 	struct vnode *pvp;
10662 	struct mount *mp;
10663 	struct diraddhd *diraddhdp;
10664 {
10665 	struct inodedep *inodedep;
10666 	struct inoref *inoref;
10667 	struct ufsmount *ump;
10668 	struct diradd *dap;
10669 	struct vnode *vp;
10670 	int error = 0;
10671 	struct buf *bp;
10672 	ino_t inum;
10673 
10674 	ump = VFSTOUFS(mp);
10675 restart:
10676 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
10677 		/*
10678 		 * Flush ourselves if this directory entry
10679 		 * has a MKDIR_PARENT dependency.
10680 		 */
10681 		if (dap->da_state & MKDIR_PARENT) {
10682 			FREE_LOCK(&lk);
10683 			if ((error = ffs_update(pvp, MNT_WAIT)) != 0)
10684 				break;
10685 			ACQUIRE_LOCK(&lk);
10686 			/*
10687 			 * If that cleared dependencies, go on to next.
10688 			 */
10689 			if (dap != LIST_FIRST(diraddhdp))
10690 				continue;
10691 			if (dap->da_state & MKDIR_PARENT)
10692 				panic("flush_pagedep_deps: MKDIR_PARENT");
10693 		}
10694 		/*
10695 		 * A newly allocated directory must have its "." and
10696 		 * ".." entries written out before its name can be
10697 		 * committed in its parent.
10698 		 */
10699 		inum = dap->da_newinum;
10700 		if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
10701 			panic("flush_pagedep_deps: lost inode1");
10702 		/*
10703 		 * Wait for any pending journal adds to complete so we don't
10704 		 * cause rollbacks while syncing.
10705 		 */
10706 		TAILQ_FOREACH(inoref, &inodedep->id_inoreflst, if_deps) {
10707 			if ((inoref->if_state & (DEPCOMPLETE | GOINGAWAY))
10708 			    == DEPCOMPLETE) {
10709 				stat_jwait_inode++;
10710 				jwait(&inoref->if_list);
10711 				goto restart;
10712 			}
10713 		}
10714 		if (dap->da_state & MKDIR_BODY) {
10715 			FREE_LOCK(&lk);
10716 			if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
10717 			    FFSV_FORCEINSMQ)))
10718 				break;
10719 			error = flush_newblk_dep(vp, mp, 0);
10720 			/*
10721 			 * If we still have the dependency we might need to
10722 			 * update the vnode to sync the new link count to
10723 			 * disk.
10724 			 */
10725 			if (error == 0 && dap == LIST_FIRST(diraddhdp))
10726 				error = ffs_update(vp, MNT_WAIT);
10727 			vput(vp);
10728 			if (error != 0)
10729 				break;
10730 			ACQUIRE_LOCK(&lk);
10731 			/*
10732 			 * If that cleared dependencies, go on to next.
10733 			 */
10734 			if (dap != LIST_FIRST(diraddhdp))
10735 				continue;
10736 			if (dap->da_state & MKDIR_BODY) {
10737 				inodedep_lookup(UFSTOVFS(ump), inum, 0,
10738 				    &inodedep);
10739 				panic("flush_pagedep_deps: MKDIR_BODY "
10740 				    "inodedep %p dap %p vp %p",
10741 				    inodedep, dap, vp);
10742 			}
10743 		}
10744 		/*
10745 		 * Flush the inode on which the directory entry depends.
10746 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
10747 		 * the only remaining dependency is that the updated inode
10748 		 * count must get pushed to disk. The inode has already
10749 		 * been pushed into its inode buffer (via VOP_UPDATE) at
10750 		 * the time of the reference count change. So we need only
10751 		 * locate that buffer, ensure that there will be no rollback
10752 		 * caused by a bitmap dependency, then write the inode buffer.
10753 		 */
10754 retry:
10755 		if (inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep) == 0)
10756 			panic("flush_pagedep_deps: lost inode");
10757 		/*
10758 		 * If the inode still has bitmap dependencies,
10759 		 * push them to disk.
10760 		 */
10761 		if ((inodedep->id_state & (DEPCOMPLETE | GOINGAWAY)) == 0) {
10762 			bp = inodedep->id_bmsafemap->sm_buf;
10763 			bp = getdirtybuf(bp, &lk, MNT_WAIT);
10764 			if (bp == NULL)
10765 				goto retry;
10766 			FREE_LOCK(&lk);
10767 			if ((error = bwrite(bp)) != 0)
10768 				break;
10769 			ACQUIRE_LOCK(&lk);
10770 			if (dap != LIST_FIRST(diraddhdp))
10771 				continue;
10772 		}
10773 		/*
10774 		 * If the inode is still sitting in a buffer waiting
10775 		 * to be written or waiting for the link count to be
10776 		 * adjusted update it here to flush it to disk.
10777 		 */
10778 		if (dap == LIST_FIRST(diraddhdp)) {
10779 			FREE_LOCK(&lk);
10780 			if ((error = ffs_vgetf(mp, inum, LK_EXCLUSIVE, &vp,
10781 			    FFSV_FORCEINSMQ)))
10782 				break;
10783 			error = ffs_update(vp, MNT_WAIT);
10784 			vput(vp);
10785 			if (error)
10786 				break;
10787 			ACQUIRE_LOCK(&lk);
10788 		}
10789 		/*
10790 		 * If we have failed to get rid of all the dependencies
10791 		 * then something is seriously wrong.
10792 		 */
10793 		if (dap == LIST_FIRST(diraddhdp)) {
10794 			inodedep_lookup(UFSTOVFS(ump), inum, 0, &inodedep);
10795 			panic("flush_pagedep_deps: failed to flush "
10796 			    "inodedep %p ino %d dap %p", inodedep, inum, dap);
10797 		}
10798 	}
10799 	if (error)
10800 		ACQUIRE_LOCK(&lk);
10801 	return (error);
10802 }
10803 
10804 /*
10805  * A large burst of file addition or deletion activity can drive the
10806  * memory load excessively high. First attempt to slow things down
10807  * using the techniques below. If that fails, this routine requests
10808  * the offending operations to fall back to running synchronously
10809  * until the memory load returns to a reasonable level.
10810  */
10811 int
10812 softdep_slowdown(vp)
10813 	struct vnode *vp;
10814 {
10815 	struct ufsmount *ump;
10816 	int jlow;
10817 	int max_softdeps_hard;
10818 
10819 	ACQUIRE_LOCK(&lk);
10820 	jlow = 0;
10821 	/*
10822 	 * Check for journal space if needed.
10823 	 */
10824 	if (DOINGSUJ(vp)) {
10825 		ump = VFSTOUFS(vp->v_mount);
10826 		if (journal_space(ump, 0) == 0)
10827 			jlow = 1;
10828 	}
10829 	max_softdeps_hard = max_softdeps * 11 / 10;
10830 	if (num_dirrem < max_softdeps_hard / 2 &&
10831 	    num_inodedep < max_softdeps_hard &&
10832 	    VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps &&
10833 	    num_freeblkdep < max_softdeps_hard && jlow == 0) {
10834 		FREE_LOCK(&lk);
10835   		return (0);
10836 	}
10837 	if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps || jlow)
10838 		softdep_speedup();
10839 	stat_sync_limit_hit += 1;
10840 	FREE_LOCK(&lk);
10841 	return (1);
10842 }
10843 
10844 /*
10845  * Called by the allocation routines when they are about to fail
10846  * in the hope that we can free up some disk space.
10847  *
10848  * First check to see if the work list has anything on it. If it has,
10849  * clean up entries until we successfully free some space. Because this
10850  * process holds inodes locked, we cannot handle any remove requests
10851  * that might block on a locked inode as that could lead to deadlock.
10852  * If the worklist yields no free space, encourage the syncer daemon
10853  * to help us. In no event will we try for longer than tickdelay seconds.
10854  */
10855 int
10856 softdep_request_cleanup(fs, vp)
10857 	struct fs *fs;
10858 	struct vnode *vp;
10859 {
10860 	struct ufsmount *ump;
10861 	long starttime;
10862 	ufs2_daddr_t needed;
10863 	int error;
10864 
10865 	ump = VTOI(vp)->i_ump;
10866 	mtx_assert(UFS_MTX(ump), MA_OWNED);
10867 	needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize;
10868 	starttime = time_second + tickdelay;
10869 	/*
10870 	 * If we are being called because of a process doing a
10871 	 * copy-on-write, then it is not safe to update the vnode
10872 	 * as we may recurse into the copy-on-write routine.
10873 	 */
10874 	if (!(curthread->td_pflags & TDP_COWINPROGRESS)) {
10875 		UFS_UNLOCK(ump);
10876 		error = ffs_update(vp, 1);
10877 		UFS_LOCK(ump);
10878 		if (error != 0)
10879 			return (0);
10880 	}
10881 	while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) {
10882 		if (time_second > starttime)
10883 			return (0);
10884 		UFS_UNLOCK(ump);
10885 		ACQUIRE_LOCK(&lk);
10886 		process_removes(vp);
10887 		if (ump->softdep_on_worklist > 0 &&
10888 		    process_worklist_item(UFSTOVFS(ump), LK_NOWAIT) != -1) {
10889 			stat_worklist_push += 1;
10890 			FREE_LOCK(&lk);
10891 			UFS_LOCK(ump);
10892 			continue;
10893 		}
10894 		request_cleanup(UFSTOVFS(ump), FLUSH_REMOVE_WAIT);
10895 		FREE_LOCK(&lk);
10896 		UFS_LOCK(ump);
10897 	}
10898 	return (1);
10899 }
10900 
10901 /*
10902  * If memory utilization has gotten too high, deliberately slow things
10903  * down and speed up the I/O processing.
10904  */
10905 extern struct thread *syncertd;
10906 static int
10907 request_cleanup(mp, resource)
10908 	struct mount *mp;
10909 	int resource;
10910 {
10911 	struct thread *td = curthread;
10912 	struct ufsmount *ump;
10913 
10914 	mtx_assert(&lk, MA_OWNED);
10915 	/*
10916 	 * We never hold up the filesystem syncer or buf daemon.
10917 	 */
10918 	if (td->td_pflags & (TDP_SOFTDEP|TDP_NORUNNINGBUF))
10919 		return (0);
10920 	ump = VFSTOUFS(mp);
10921 	/*
10922 	 * First check to see if the work list has gotten backlogged.
10923 	 * If it has, co-opt this process to help clean up two entries.
10924 	 * Because this process may hold inodes locked, we cannot
10925 	 * handle any remove requests that might block on a locked
10926 	 * inode as that could lead to deadlock.  We set TDP_SOFTDEP
10927 	 * to avoid recursively processing the worklist.
10928 	 */
10929 	if (ump->softdep_on_worklist > max_softdeps / 10) {
10930 		td->td_pflags |= TDP_SOFTDEP;
10931 		process_worklist_item(mp, LK_NOWAIT);
10932 		process_worklist_item(mp, LK_NOWAIT);
10933 		td->td_pflags &= ~TDP_SOFTDEP;
10934 		stat_worklist_push += 2;
10935 		return(1);
10936 	}
10937 	/*
10938 	 * Next, we attempt to speed up the syncer process. If that
10939 	 * is successful, then we allow the process to continue.
10940 	 */
10941 	if (softdep_speedup() && resource != FLUSH_REMOVE_WAIT)
10942 		return(0);
10943 	/*
10944 	 * If we are resource constrained on inode dependencies, try
10945 	 * flushing some dirty inodes. Otherwise, we are constrained
10946 	 * by file deletions, so try accelerating flushes of directories
10947 	 * with removal dependencies. We would like to do the cleanup
10948 	 * here, but we probably hold an inode locked at this point and
10949 	 * that might deadlock against one that we try to clean. So,
10950 	 * the best that we can do is request the syncer daemon to do
10951 	 * the cleanup for us.
10952 	 */
10953 	switch (resource) {
10954 
10955 	case FLUSH_INODES:
10956 		stat_ino_limit_push += 1;
10957 		req_clear_inodedeps += 1;
10958 		stat_countp = &stat_ino_limit_hit;
10959 		break;
10960 
10961 	case FLUSH_REMOVE:
10962 	case FLUSH_REMOVE_WAIT:
10963 		stat_blk_limit_push += 1;
10964 		req_clear_remove += 1;
10965 		stat_countp = &stat_blk_limit_hit;
10966 		break;
10967 
10968 	default:
10969 		panic("request_cleanup: unknown type");
10970 	}
10971 	/*
10972 	 * Hopefully the syncer daemon will catch up and awaken us.
10973 	 * We wait at most tickdelay before proceeding in any case.
10974 	 */
10975 	proc_waiting += 1;
10976 	if (callout_pending(&softdep_callout) == FALSE)
10977 		callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2,
10978 		    pause_timer, 0);
10979 
10980 	msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0);
10981 	proc_waiting -= 1;
10982 	return (1);
10983 }
10984 
10985 /*
10986  * Awaken processes pausing in request_cleanup and clear proc_waiting
10987  * to indicate that there is no longer a timer running.
10988  */
10989 static void
10990 pause_timer(arg)
10991 	void *arg;
10992 {
10993 
10994 	/*
10995 	 * The callout_ API has acquired mtx and will hold it around this
10996 	 * function call.
10997 	 */
10998 	*stat_countp += 1;
10999 	wakeup_one(&proc_waiting);
11000 	if (proc_waiting > 0)
11001 		callout_reset(&softdep_callout, tickdelay > 2 ? tickdelay : 2,
11002 		    pause_timer, 0);
11003 }
11004 
11005 /*
11006  * Flush out a directory with at least one removal dependency in an effort to
11007  * reduce the number of dirrem, freefile, and freeblks dependency structures.
11008  */
11009 static void
11010 clear_remove(td)
11011 	struct thread *td;
11012 {
11013 	struct pagedep_hashhead *pagedephd;
11014 	struct pagedep *pagedep;
11015 	static int next = 0;
11016 	struct mount *mp;
11017 	struct vnode *vp;
11018 	struct bufobj *bo;
11019 	int error, cnt;
11020 	ino_t ino;
11021 
11022 	mtx_assert(&lk, MA_OWNED);
11023 
11024 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
11025 		pagedephd = &pagedep_hashtbl[next++];
11026 		if (next >= pagedep_hash)
11027 			next = 0;
11028 		LIST_FOREACH(pagedep, pagedephd, pd_hash) {
11029 			if (LIST_EMPTY(&pagedep->pd_dirremhd))
11030 				continue;
11031 			mp = pagedep->pd_list.wk_mp;
11032 			ino = pagedep->pd_ino;
11033 			if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
11034 				continue;
11035 			FREE_LOCK(&lk);
11036 
11037 			/*
11038 			 * Let unmount clear deps
11039 			 */
11040 			error = vfs_busy(mp, MBF_NOWAIT);
11041 			if (error != 0)
11042 				goto finish_write;
11043 			error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
11044 			     FFSV_FORCEINSMQ);
11045 			vfs_unbusy(mp);
11046 			if (error != 0) {
11047 				softdep_error("clear_remove: vget", error);
11048 				goto finish_write;
11049 			}
11050 			if ((error = ffs_syncvnode(vp, MNT_NOWAIT)))
11051 				softdep_error("clear_remove: fsync", error);
11052 			bo = &vp->v_bufobj;
11053 			BO_LOCK(bo);
11054 			drain_output(vp);
11055 			BO_UNLOCK(bo);
11056 			vput(vp);
11057 		finish_write:
11058 			vn_finished_write(mp);
11059 			ACQUIRE_LOCK(&lk);
11060 			return;
11061 		}
11062 	}
11063 }
11064 
11065 /*
11066  * Clear out a block of dirty inodes in an effort to reduce
11067  * the number of inodedep dependency structures.
11068  */
11069 static void
11070 clear_inodedeps(td)
11071 	struct thread *td;
11072 {
11073 	struct inodedep_hashhead *inodedephd;
11074 	struct inodedep *inodedep;
11075 	static int next = 0;
11076 	struct mount *mp;
11077 	struct vnode *vp;
11078 	struct fs *fs;
11079 	int error, cnt;
11080 	ino_t firstino, lastino, ino;
11081 
11082 	mtx_assert(&lk, MA_OWNED);
11083 	/*
11084 	 * Pick a random inode dependency to be cleared.
11085 	 * We will then gather up all the inodes in its block
11086 	 * that have dependencies and flush them out.
11087 	 */
11088 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
11089 		inodedephd = &inodedep_hashtbl[next++];
11090 		if (next >= inodedep_hash)
11091 			next = 0;
11092 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
11093 			break;
11094 	}
11095 	if (inodedep == NULL)
11096 		return;
11097 	fs = inodedep->id_fs;
11098 	mp = inodedep->id_list.wk_mp;
11099 	/*
11100 	 * Find the last inode in the block with dependencies.
11101 	 */
11102 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
11103 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
11104 		if (inodedep_lookup(mp, lastino, 0, &inodedep) != 0)
11105 			break;
11106 	/*
11107 	 * Asynchronously push all but the last inode with dependencies.
11108 	 * Synchronously push the last inode with dependencies to ensure
11109 	 * that the inode block gets written to free up the inodedeps.
11110 	 */
11111 	for (ino = firstino; ino <= lastino; ino++) {
11112 		if (inodedep_lookup(mp, ino, 0, &inodedep) == 0)
11113 			continue;
11114 		if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
11115 			continue;
11116 		FREE_LOCK(&lk);
11117 		error = vfs_busy(mp, MBF_NOWAIT); /* Let unmount clear deps */
11118 		if (error != 0) {
11119 			vn_finished_write(mp);
11120 			ACQUIRE_LOCK(&lk);
11121 			return;
11122 		}
11123 		if ((error = ffs_vgetf(mp, ino, LK_EXCLUSIVE, &vp,
11124 		    FFSV_FORCEINSMQ)) != 0) {
11125 			softdep_error("clear_inodedeps: vget", error);
11126 			vfs_unbusy(mp);
11127 			vn_finished_write(mp);
11128 			ACQUIRE_LOCK(&lk);
11129 			return;
11130 		}
11131 		vfs_unbusy(mp);
11132 		if (ino == lastino) {
11133 			if ((error = ffs_syncvnode(vp, MNT_WAIT)))
11134 				softdep_error("clear_inodedeps: fsync1", error);
11135 		} else {
11136 			if ((error = ffs_syncvnode(vp, MNT_NOWAIT)))
11137 				softdep_error("clear_inodedeps: fsync2", error);
11138 			BO_LOCK(&vp->v_bufobj);
11139 			drain_output(vp);
11140 			BO_UNLOCK(&vp->v_bufobj);
11141 		}
11142 		vput(vp);
11143 		vn_finished_write(mp);
11144 		ACQUIRE_LOCK(&lk);
11145 	}
11146 }
11147 
11148 /*
11149  * Function to determine if the buffer has outstanding dependencies
11150  * that will cause a roll-back if the buffer is written. If wantcount
11151  * is set, return number of dependencies, otherwise just yes or no.
11152  */
11153 static int
11154 softdep_count_dependencies(bp, wantcount)
11155 	struct buf *bp;
11156 	int wantcount;
11157 {
11158 	struct worklist *wk;
11159 	struct bmsafemap *bmsafemap;
11160 	struct inodedep *inodedep;
11161 	struct indirdep *indirdep;
11162 	struct freeblks *freeblks;
11163 	struct allocindir *aip;
11164 	struct pagedep *pagedep;
11165 	struct dirrem *dirrem;
11166 	struct newblk *newblk;
11167 	struct mkdir *mkdir;
11168 	struct diradd *dap;
11169 	int i, retval;
11170 
11171 	retval = 0;
11172 	ACQUIRE_LOCK(&lk);
11173 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
11174 		switch (wk->wk_type) {
11175 
11176 		case D_INODEDEP:
11177 			inodedep = WK_INODEDEP(wk);
11178 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
11179 				/* bitmap allocation dependency */
11180 				retval += 1;
11181 				if (!wantcount)
11182 					goto out;
11183 			}
11184 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
11185 				/* direct block pointer dependency */
11186 				retval += 1;
11187 				if (!wantcount)
11188 					goto out;
11189 			}
11190 			if (TAILQ_FIRST(&inodedep->id_extupdt)) {
11191 				/* direct block pointer dependency */
11192 				retval += 1;
11193 				if (!wantcount)
11194 					goto out;
11195 			}
11196 			if (TAILQ_FIRST(&inodedep->id_inoreflst)) {
11197 				/* Add reference dependency. */
11198 				retval += 1;
11199 				if (!wantcount)
11200 					goto out;
11201 			}
11202 			continue;
11203 
11204 		case D_INDIRDEP:
11205 			indirdep = WK_INDIRDEP(wk);
11206 
11207 			LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
11208 				/* indirect block pointer dependency */
11209 				retval += 1;
11210 				if (!wantcount)
11211 					goto out;
11212 			}
11213 			continue;
11214 
11215 		case D_PAGEDEP:
11216 			pagedep = WK_PAGEDEP(wk);
11217 			LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
11218 				if (LIST_FIRST(&dirrem->dm_jremrefhd)) {
11219 					/* Journal remove ref dependency. */
11220 					retval += 1;
11221 					if (!wantcount)
11222 						goto out;
11223 				}
11224 			}
11225 			for (i = 0; i < DAHASHSZ; i++) {
11226 
11227 				LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
11228 					/* directory entry dependency */
11229 					retval += 1;
11230 					if (!wantcount)
11231 						goto out;
11232 				}
11233 			}
11234 			continue;
11235 
11236 		case D_BMSAFEMAP:
11237 			bmsafemap = WK_BMSAFEMAP(wk);
11238 			if (LIST_FIRST(&bmsafemap->sm_jaddrefhd)) {
11239 				/* Add reference dependency. */
11240 				retval += 1;
11241 				if (!wantcount)
11242 					goto out;
11243 			}
11244 			if (LIST_FIRST(&bmsafemap->sm_jnewblkhd)) {
11245 				/* Allocate block dependency. */
11246 				retval += 1;
11247 				if (!wantcount)
11248 					goto out;
11249 			}
11250 			continue;
11251 
11252 		case D_FREEBLKS:
11253 			freeblks = WK_FREEBLKS(wk);
11254 			if (LIST_FIRST(&freeblks->fb_jfreeblkhd)) {
11255 				/* Freeblk journal dependency. */
11256 				retval += 1;
11257 				if (!wantcount)
11258 					goto out;
11259 			}
11260 			continue;
11261 
11262 		case D_ALLOCDIRECT:
11263 		case D_ALLOCINDIR:
11264 			newblk = WK_NEWBLK(wk);
11265 			if (newblk->nb_jnewblk) {
11266 				/* Journal allocate dependency. */
11267 				retval += 1;
11268 				if (!wantcount)
11269 					goto out;
11270 			}
11271 			continue;
11272 
11273 		case D_MKDIR:
11274 			mkdir = WK_MKDIR(wk);
11275 			if (mkdir->md_jaddref) {
11276 				/* Journal reference dependency. */
11277 				retval += 1;
11278 				if (!wantcount)
11279 					goto out;
11280 			}
11281 			continue;
11282 
11283 		case D_FREEWORK:
11284 		case D_FREEDEP:
11285 		case D_JSEGDEP:
11286 		case D_JSEG:
11287 		case D_SBDEP:
11288 			/* never a dependency on these blocks */
11289 			continue;
11290 
11291 		default:
11292 			panic("softdep_count_dependencies: Unexpected type %s",
11293 			    TYPENAME(wk->wk_type));
11294 			/* NOTREACHED */
11295 		}
11296 	}
11297 out:
11298 	FREE_LOCK(&lk);
11299 	return retval;
11300 }
11301 
11302 /*
11303  * Acquire exclusive access to a buffer.
11304  * Must be called with a locked mtx parameter.
11305  * Return acquired buffer or NULL on failure.
11306  */
11307 static struct buf *
11308 getdirtybuf(bp, mtx, waitfor)
11309 	struct buf *bp;
11310 	struct mtx *mtx;
11311 	int waitfor;
11312 {
11313 	int error;
11314 
11315 	mtx_assert(mtx, MA_OWNED);
11316 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) {
11317 		if (waitfor != MNT_WAIT)
11318 			return (NULL);
11319 		error = BUF_LOCK(bp,
11320 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx);
11321 		/*
11322 		 * Even if we sucessfully acquire bp here, we have dropped
11323 		 * mtx, which may violates our guarantee.
11324 		 */
11325 		if (error == 0)
11326 			BUF_UNLOCK(bp);
11327 		else if (error != ENOLCK)
11328 			panic("getdirtybuf: inconsistent lock: %d", error);
11329 		mtx_lock(mtx);
11330 		return (NULL);
11331 	}
11332 	if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
11333 		if (mtx == &lk && waitfor == MNT_WAIT) {
11334 			mtx_unlock(mtx);
11335 			BO_LOCK(bp->b_bufobj);
11336 			BUF_UNLOCK(bp);
11337 			if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
11338 				bp->b_vflags |= BV_BKGRDWAIT;
11339 				msleep(&bp->b_xflags, BO_MTX(bp->b_bufobj),
11340 				       PRIBIO | PDROP, "getbuf", 0);
11341 			} else
11342 				BO_UNLOCK(bp->b_bufobj);
11343 			mtx_lock(mtx);
11344 			return (NULL);
11345 		}
11346 		BUF_UNLOCK(bp);
11347 		if (waitfor != MNT_WAIT)
11348 			return (NULL);
11349 		/*
11350 		 * The mtx argument must be bp->b_vp's mutex in
11351 		 * this case.
11352 		 */
11353 #ifdef	DEBUG_VFS_LOCKS
11354 		if (bp->b_vp->v_type != VCHR)
11355 			ASSERT_BO_LOCKED(bp->b_bufobj);
11356 #endif
11357 		bp->b_vflags |= BV_BKGRDWAIT;
11358 		msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0);
11359 		return (NULL);
11360 	}
11361 	if ((bp->b_flags & B_DELWRI) == 0) {
11362 		BUF_UNLOCK(bp);
11363 		return (NULL);
11364 	}
11365 	bremfree(bp);
11366 	return (bp);
11367 }
11368 
11369 
11370 /*
11371  * Check if it is safe to suspend the file system now.  On entry,
11372  * the vnode interlock for devvp should be held.  Return 0 with
11373  * the mount interlock held if the file system can be suspended now,
11374  * otherwise return EAGAIN with the mount interlock held.
11375  */
11376 int
11377 softdep_check_suspend(struct mount *mp,
11378 		      struct vnode *devvp,
11379 		      int softdep_deps,
11380 		      int softdep_accdeps,
11381 		      int secondary_writes,
11382 		      int secondary_accwrites)
11383 {
11384 	struct bufobj *bo;
11385 	struct ufsmount *ump;
11386 	int error;
11387 
11388 	ump = VFSTOUFS(mp);
11389 	bo = &devvp->v_bufobj;
11390 	ASSERT_BO_LOCKED(bo);
11391 
11392 	for (;;) {
11393 		if (!TRY_ACQUIRE_LOCK(&lk)) {
11394 			BO_UNLOCK(bo);
11395 			ACQUIRE_LOCK(&lk);
11396 			FREE_LOCK(&lk);
11397 			BO_LOCK(bo);
11398 			continue;
11399 		}
11400 		MNT_ILOCK(mp);
11401 		if (mp->mnt_secondary_writes != 0) {
11402 			FREE_LOCK(&lk);
11403 			BO_UNLOCK(bo);
11404 			msleep(&mp->mnt_secondary_writes,
11405 			       MNT_MTX(mp),
11406 			       (PUSER - 1) | PDROP, "secwr", 0);
11407 			BO_LOCK(bo);
11408 			continue;
11409 		}
11410 		break;
11411 	}
11412 
11413 	/*
11414 	 * Reasons for needing more work before suspend:
11415 	 * - Dirty buffers on devvp.
11416 	 * - Softdep activity occurred after start of vnode sync loop
11417 	 * - Secondary writes occurred after start of vnode sync loop
11418 	 */
11419 	error = 0;
11420 	if (bo->bo_numoutput > 0 ||
11421 	    bo->bo_dirty.bv_cnt > 0 ||
11422 	    softdep_deps != 0 ||
11423 	    ump->softdep_deps != 0 ||
11424 	    softdep_accdeps != ump->softdep_accdeps ||
11425 	    secondary_writes != 0 ||
11426 	    mp->mnt_secondary_writes != 0 ||
11427 	    secondary_accwrites != mp->mnt_secondary_accwrites)
11428 		error = EAGAIN;
11429 	FREE_LOCK(&lk);
11430 	BO_UNLOCK(bo);
11431 	return (error);
11432 }
11433 
11434 
11435 /*
11436  * Get the number of dependency structures for the file system, both
11437  * the current number and the total number allocated.  These will
11438  * later be used to detect that softdep processing has occurred.
11439  */
11440 void
11441 softdep_get_depcounts(struct mount *mp,
11442 		      int *softdep_depsp,
11443 		      int *softdep_accdepsp)
11444 {
11445 	struct ufsmount *ump;
11446 
11447 	ump = VFSTOUFS(mp);
11448 	ACQUIRE_LOCK(&lk);
11449 	*softdep_depsp = ump->softdep_deps;
11450 	*softdep_accdepsp = ump->softdep_accdeps;
11451 	FREE_LOCK(&lk);
11452 }
11453 
11454 /*
11455  * Wait for pending output on a vnode to complete.
11456  * Must be called with vnode lock and interlock locked.
11457  *
11458  * XXX: Should just be a call to bufobj_wwait().
11459  */
11460 static void
11461 drain_output(vp)
11462 	struct vnode *vp;
11463 {
11464 	struct bufobj *bo;
11465 
11466 	bo = &vp->v_bufobj;
11467 	ASSERT_VOP_LOCKED(vp, "drain_output");
11468 	ASSERT_BO_LOCKED(bo);
11469 
11470 	while (bo->bo_numoutput) {
11471 		bo->bo_flag |= BO_WWAIT;
11472 		msleep((caddr_t)&bo->bo_numoutput,
11473 		    BO_MTX(bo), PRIBIO + 1, "drainvp", 0);
11474 	}
11475 }
11476 
11477 /*
11478  * Called whenever a buffer that is being invalidated or reallocated
11479  * contains dependencies. This should only happen if an I/O error has
11480  * occurred. The routine is called with the buffer locked.
11481  */
11482 static void
11483 softdep_deallocate_dependencies(bp)
11484 	struct buf *bp;
11485 {
11486 
11487 	if ((bp->b_ioflags & BIO_ERROR) == 0)
11488 		panic("softdep_deallocate_dependencies: dangling deps");
11489 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
11490 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
11491 }
11492 
11493 /*
11494  * Function to handle asynchronous write errors in the filesystem.
11495  */
11496 static void
11497 softdep_error(func, error)
11498 	char *func;
11499 	int error;
11500 {
11501 
11502 	/* XXX should do something better! */
11503 	printf("%s: got error %d while accessing filesystem\n", func, error);
11504 }
11505 
11506 #ifdef DDB
11507 
11508 static void
11509 inodedep_print(struct inodedep *inodedep, int verbose)
11510 {
11511 	db_printf("%p fs %p st %x ino %jd inoblk %jd delta %d nlink %d"
11512 	    " saveino %p\n",
11513 	    inodedep, inodedep->id_fs, inodedep->id_state,
11514 	    (intmax_t)inodedep->id_ino,
11515 	    (intmax_t)fsbtodb(inodedep->id_fs,
11516 	    ino_to_fsba(inodedep->id_fs, inodedep->id_ino)),
11517 	    inodedep->id_nlinkdelta, inodedep->id_savednlink,
11518 	    inodedep->id_savedino1);
11519 
11520 	if (verbose == 0)
11521 		return;
11522 
11523 	db_printf("\tpendinghd %p, bufwait %p, inowait %p, inoreflst %p, "
11524 	    "mkdiradd %p\n",
11525 	    LIST_FIRST(&inodedep->id_pendinghd),
11526 	    LIST_FIRST(&inodedep->id_bufwait),
11527 	    LIST_FIRST(&inodedep->id_inowait),
11528 	    TAILQ_FIRST(&inodedep->id_inoreflst),
11529 	    inodedep->id_mkdiradd);
11530 	db_printf("\tinoupdt %p, newinoupdt %p, extupdt %p, newextupdt %p\n",
11531 	    TAILQ_FIRST(&inodedep->id_inoupdt),
11532 	    TAILQ_FIRST(&inodedep->id_newinoupdt),
11533 	    TAILQ_FIRST(&inodedep->id_extupdt),
11534 	    TAILQ_FIRST(&inodedep->id_newextupdt));
11535 }
11536 
11537 DB_SHOW_COMMAND(inodedep, db_show_inodedep)
11538 {
11539 
11540 	if (have_addr == 0) {
11541 		db_printf("Address required\n");
11542 		return;
11543 	}
11544 	inodedep_print((struct inodedep*)addr, 1);
11545 }
11546 
11547 DB_SHOW_COMMAND(inodedeps, db_show_inodedeps)
11548 {
11549 	struct inodedep_hashhead *inodedephd;
11550 	struct inodedep *inodedep;
11551 	struct fs *fs;
11552 	int cnt;
11553 
11554 	fs = have_addr ? (struct fs *)addr : NULL;
11555 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
11556 		inodedephd = &inodedep_hashtbl[cnt];
11557 		LIST_FOREACH(inodedep, inodedephd, id_hash) {
11558 			if (fs != NULL && fs != inodedep->id_fs)
11559 				continue;
11560 			inodedep_print(inodedep, 0);
11561 		}
11562 	}
11563 }
11564 
11565 DB_SHOW_COMMAND(worklist, db_show_worklist)
11566 {
11567 	struct worklist *wk;
11568 
11569 	if (have_addr == 0) {
11570 		db_printf("Address required\n");
11571 		return;
11572 	}
11573 	wk = (struct worklist *)addr;
11574 	printf("worklist: %p type %s state 0x%X\n",
11575 	    wk, TYPENAME(wk->wk_type), wk->wk_state);
11576 }
11577 
11578 DB_SHOW_COMMAND(workhead, db_show_workhead)
11579 {
11580 	struct workhead *wkhd;
11581 	struct worklist *wk;
11582 	int i;
11583 
11584 	if (have_addr == 0) {
11585 		db_printf("Address required\n");
11586 		return;
11587 	}
11588 	wkhd = (struct workhead *)addr;
11589 	wk = LIST_FIRST(wkhd);
11590 	for (i = 0; i < 100 && wk != NULL; i++, wk = LIST_NEXT(wk, wk_list))
11591 		db_printf("worklist: %p type %s state 0x%X",
11592 		    wk, TYPENAME(wk->wk_type), wk->wk_state);
11593 	if (i == 100)
11594 		db_printf("workhead overflow");
11595 	printf("\n");
11596 }
11597 
11598 
11599 DB_SHOW_COMMAND(mkdirs, db_show_mkdirs)
11600 {
11601 	struct jaddref *jaddref;
11602 	struct diradd *diradd;
11603 	struct mkdir *mkdir;
11604 
11605 	LIST_FOREACH(mkdir, &mkdirlisthd, md_mkdirs) {
11606 		diradd = mkdir->md_diradd;
11607 		db_printf("mkdir: %p state 0x%X dap %p state 0x%X",
11608 		    mkdir, mkdir->md_state, diradd, diradd->da_state);
11609 		if ((jaddref = mkdir->md_jaddref) != NULL)
11610 			db_printf(" jaddref %p jaddref state 0x%X",
11611 			    jaddref, jaddref->ja_state);
11612 		db_printf("\n");
11613 	}
11614 }
11615 
11616 #endif /* DDB */
11617 
11618 #endif /* SOFTUPDATES */
11619