xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision 262e143bd46171a6415a5b28af260a5efa2a3db8)
1 /*-
2  * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
3  *
4  * The soft updates code is derived from the appendix of a University
5  * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
6  * "Soft Updates: A Solution to the Metadata Update Problem in File
7  * Systems", CSE-TR-254-95, August 1995).
8  *
9  * Further information about soft updates can be obtained from:
10  *
11  *	Marshall Kirk McKusick		http://www.mckusick.com/softdep/
12  *	1614 Oxford Street		mckusick@mckusick.com
13  *	Berkeley, CA 94709-1608		+1-510-843-9542
14  *	USA
15  *
16  * Redistribution and use in source and binary forms, with or without
17  * modification, are permitted provided that the following conditions
18  * are met:
19  *
20  * 1. Redistributions of source code must retain the above copyright
21  *    notice, this list of conditions and the following disclaimer.
22  * 2. Redistributions in binary form must reproduce the above copyright
23  *    notice, this list of conditions and the following disclaimer in the
24  *    documentation and/or other materials provided with the distribution.
25  *
26  * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
27  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
28  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
29  * DISCLAIMED.  IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
30  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)ffs_softdep.c	9.59 (McKusick) 6/21/00
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 /*
45  * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
46  */
47 #ifndef DIAGNOSTIC
48 #define DIAGNOSTIC
49 #endif
50 #ifndef DEBUG
51 #define DEBUG
52 #endif
53 
54 #include <sys/param.h>
55 #include <sys/kernel.h>
56 #include <sys/systm.h>
57 #include <sys/bio.h>
58 #include <sys/buf.h>
59 #include <sys/kdb.h>
60 #include <sys/lock.h>
61 #include <sys/malloc.h>
62 #include <sys/mount.h>
63 #include <sys/mutex.h>
64 #include <sys/proc.h>
65 #include <sys/stat.h>
66 #include <sys/sysctl.h>
67 #include <sys/syslog.h>
68 #include <sys/vnode.h>
69 #include <sys/conf.h>
70 #include <ufs/ufs/dir.h>
71 #include <ufs/ufs/extattr.h>
72 #include <ufs/ufs/quota.h>
73 #include <ufs/ufs/inode.h>
74 #include <ufs/ufs/ufsmount.h>
75 #include <ufs/ffs/fs.h>
76 #include <ufs/ffs/softdep.h>
77 #include <ufs/ffs/ffs_extern.h>
78 #include <ufs/ufs/ufs_extern.h>
79 
80 #include <vm/vm.h>
81 
82 #include "opt_ffs.h"
83 
84 #ifndef SOFTUPDATES
85 
86 int
87 softdep_flushfiles(oldmnt, flags, td)
88 	struct mount *oldmnt;
89 	int flags;
90 	struct thread *td;
91 {
92 
93 	panic("softdep_flushfiles called");
94 }
95 
96 int
97 softdep_mount(devvp, mp, fs, cred)
98 	struct vnode *devvp;
99 	struct mount *mp;
100 	struct fs *fs;
101 	struct ucred *cred;
102 {
103 
104 	return (0);
105 }
106 
107 void
108 softdep_initialize()
109 {
110 
111 	return;
112 }
113 
114 void
115 softdep_uninitialize()
116 {
117 
118 	return;
119 }
120 
121 void
122 softdep_setup_inomapdep(bp, ip, newinum)
123 	struct buf *bp;
124 	struct inode *ip;
125 	ino_t newinum;
126 {
127 
128 	panic("softdep_setup_inomapdep called");
129 }
130 
131 void
132 softdep_setup_blkmapdep(bp, fs, newblkno)
133 	struct buf *bp;
134 	struct fs *fs;
135 	ufs2_daddr_t newblkno;
136 {
137 
138 	panic("softdep_setup_blkmapdep called");
139 }
140 
141 void
142 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
143 	struct inode *ip;
144 	ufs_lbn_t lbn;
145 	ufs2_daddr_t newblkno;
146 	ufs2_daddr_t oldblkno;
147 	long newsize;
148 	long oldsize;
149 	struct buf *bp;
150 {
151 
152 	panic("softdep_setup_allocdirect called");
153 }
154 
155 void
156 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
157 	struct inode *ip;
158 	ufs_lbn_t lbn;
159 	ufs2_daddr_t newblkno;
160 	ufs2_daddr_t oldblkno;
161 	long newsize;
162 	long oldsize;
163 	struct buf *bp;
164 {
165 
166 	panic("softdep_setup_allocdirect called");
167 }
168 
169 void
170 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
171 	struct inode *ip;
172 	ufs_lbn_t lbn;
173 	struct buf *bp;
174 	int ptrno;
175 	ufs2_daddr_t newblkno;
176 	ufs2_daddr_t oldblkno;
177 	struct buf *nbp;
178 {
179 
180 	panic("softdep_setup_allocindir_page called");
181 }
182 
183 void
184 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
185 	struct buf *nbp;
186 	struct inode *ip;
187 	struct buf *bp;
188 	int ptrno;
189 	ufs2_daddr_t newblkno;
190 {
191 
192 	panic("softdep_setup_allocindir_meta called");
193 }
194 
195 void
196 softdep_setup_freeblocks(ip, length, flags)
197 	struct inode *ip;
198 	off_t length;
199 	int flags;
200 {
201 
202 	panic("softdep_setup_freeblocks called");
203 }
204 
205 void
206 softdep_freefile(pvp, ino, mode)
207 		struct vnode *pvp;
208 		ino_t ino;
209 		int mode;
210 {
211 
212 	panic("softdep_freefile called");
213 }
214 
215 int
216 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
217 	struct buf *bp;
218 	struct inode *dp;
219 	off_t diroffset;
220 	ino_t newinum;
221 	struct buf *newdirbp;
222 	int isnewblk;
223 {
224 
225 	panic("softdep_setup_directory_add called");
226 }
227 
228 void
229 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
230 	struct inode *dp;
231 	caddr_t base;
232 	caddr_t oldloc;
233 	caddr_t newloc;
234 	int entrysize;
235 {
236 
237 	panic("softdep_change_directoryentry_offset called");
238 }
239 
240 void
241 softdep_setup_remove(bp, dp, ip, isrmdir)
242 	struct buf *bp;
243 	struct inode *dp;
244 	struct inode *ip;
245 	int isrmdir;
246 {
247 
248 	panic("softdep_setup_remove called");
249 }
250 
251 void
252 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
253 	struct buf *bp;
254 	struct inode *dp;
255 	struct inode *ip;
256 	ino_t newinum;
257 	int isrmdir;
258 {
259 
260 	panic("softdep_setup_directory_change called");
261 }
262 
263 void
264 softdep_change_linkcnt(ip)
265 	struct inode *ip;
266 {
267 
268 	panic("softdep_change_linkcnt called");
269 }
270 
271 void
272 softdep_load_inodeblock(ip)
273 	struct inode *ip;
274 {
275 
276 	panic("softdep_load_inodeblock called");
277 }
278 
279 void
280 softdep_update_inodeblock(ip, bp, waitfor)
281 	struct inode *ip;
282 	struct buf *bp;
283 	int waitfor;
284 {
285 
286 	panic("softdep_update_inodeblock called");
287 }
288 
289 int
290 softdep_fsync(vp)
291 	struct vnode *vp;	/* the "in_core" copy of the inode */
292 {
293 
294 	return (0);
295 }
296 
297 void
298 softdep_fsync_mountdev(vp)
299 	struct vnode *vp;
300 {
301 
302 	return;
303 }
304 
305 int
306 softdep_flushworklist(oldmnt, countp, td)
307 	struct mount *oldmnt;
308 	int *countp;
309 	struct thread *td;
310 {
311 
312 	*countp = 0;
313 	return (0);
314 }
315 
316 int
317 softdep_sync_metadata(struct vnode *vp)
318 {
319 
320 	return (0);
321 }
322 
323 int
324 softdep_slowdown(vp)
325 	struct vnode *vp;
326 {
327 
328 	panic("softdep_slowdown called");
329 }
330 
331 void
332 softdep_releasefile(ip)
333 	struct inode *ip;	/* inode with the zero effective link count */
334 {
335 
336 	panic("softdep_releasefile called");
337 }
338 
339 int
340 softdep_request_cleanup(fs, vp)
341 	struct fs *fs;
342 	struct vnode *vp;
343 {
344 
345 	return (0);
346 }
347 
348 #else
349 /*
350  * These definitions need to be adapted to the system to which
351  * this file is being ported.
352  */
353 /*
354  * malloc types defined for the softdep system.
355  */
356 static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
357 static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
358 static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
359 static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
360 static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
361 static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
362 static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
363 static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
364 static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
365 static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
366 static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
367 static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
368 static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
369 static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block");
370 static MALLOC_DEFINE(M_SAVEDINO, "savedino","Saved inodes");
371 
372 #define M_SOFTDEP_FLAGS	(M_WAITOK | M_USE_RESERVE)
373 
374 #define	D_PAGEDEP	0
375 #define	D_INODEDEP	1
376 #define	D_NEWBLK	2
377 #define	D_BMSAFEMAP	3
378 #define	D_ALLOCDIRECT	4
379 #define	D_INDIRDEP	5
380 #define	D_ALLOCINDIR	6
381 #define	D_FREEFRAG	7
382 #define	D_FREEBLKS	8
383 #define	D_FREEFILE	9
384 #define	D_DIRADD	10
385 #define	D_MKDIR		11
386 #define	D_DIRREM	12
387 #define	D_NEWDIRBLK	13
388 #define	D_LAST		D_NEWDIRBLK
389 
390 /*
391  * translate from workitem type to memory type
392  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
393  */
394 static struct malloc_type *memtype[] = {
395 	M_PAGEDEP,
396 	M_INODEDEP,
397 	M_NEWBLK,
398 	M_BMSAFEMAP,
399 	M_ALLOCDIRECT,
400 	M_INDIRDEP,
401 	M_ALLOCINDIR,
402 	M_FREEFRAG,
403 	M_FREEBLKS,
404 	M_FREEFILE,
405 	M_DIRADD,
406 	M_MKDIR,
407 	M_DIRREM,
408 	M_NEWDIRBLK
409 };
410 
411 #define DtoM(type) (memtype[type])
412 
413 /*
414  * Names of malloc types.
415  */
416 #define TYPENAME(type)  \
417 	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
418 /*
419  * End system adaptaion definitions.
420  */
421 
422 /*
423  * Forward declarations.
424  */
425 struct inodedep_hashhead;
426 struct newblk_hashhead;
427 struct pagedep_hashhead;
428 
429 /*
430  * Internal function prototypes.
431  */
432 static	void softdep_error(char *, int);
433 static	void drain_output(struct vnode *);
434 static	struct buf *getdirtybuf(struct buf *, struct mtx *, int);
435 static	void clear_remove(struct thread *);
436 static	void clear_inodedeps(struct thread *);
437 static	int flush_pagedep_deps(struct vnode *, struct mount *,
438 	    struct diraddhd *);
439 static	int flush_inodedep_deps(struct fs *, ino_t);
440 static	int flush_deplist(struct allocdirectlst *, int, int *);
441 static	int handle_written_filepage(struct pagedep *, struct buf *);
442 static  void diradd_inode_written(struct diradd *, struct inodedep *);
443 static	int handle_written_inodeblock(struct inodedep *, struct buf *);
444 static	void handle_allocdirect_partdone(struct allocdirect *);
445 static	void handle_allocindir_partdone(struct allocindir *);
446 static	void initiate_write_filepage(struct pagedep *, struct buf *);
447 static	void handle_written_mkdir(struct mkdir *, int);
448 static	void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
449 static	void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
450 static	void handle_workitem_freefile(struct freefile *);
451 static	void handle_workitem_remove(struct dirrem *, struct vnode *);
452 static	struct dirrem *newdirrem(struct buf *, struct inode *,
453 	    struct inode *, int, struct dirrem **);
454 static	void free_diradd(struct diradd *);
455 static	void free_allocindir(struct allocindir *, struct inodedep *);
456 static	void free_newdirblk(struct newdirblk *);
457 static	int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t,
458 	    ufs2_daddr_t *);
459 static	void deallocate_dependencies(struct buf *, struct inodedep *);
460 static	void free_allocdirect(struct allocdirectlst *,
461 	    struct allocdirect *, int);
462 static	int check_inode_unwritten(struct inodedep *);
463 static	int free_inodedep(struct inodedep *);
464 static	void handle_workitem_freeblocks(struct freeblks *, int);
465 static	void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
466 static	void setup_allocindir_phase2(struct buf *, struct inode *,
467 	    struct allocindir *);
468 static	struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
469 	    ufs2_daddr_t);
470 static	void handle_workitem_freefrag(struct freefrag *);
471 static	struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long);
472 static	void allocdirect_merge(struct allocdirectlst *,
473 	    struct allocdirect *, struct allocdirect *);
474 static	struct bmsafemap *bmsafemap_lookup(struct buf *);
475 static	int newblk_find(struct newblk_hashhead *, struct fs *, ufs2_daddr_t,
476 	    struct newblk **);
477 static	int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **);
478 static	int inodedep_find(struct inodedep_hashhead *, struct fs *, ino_t,
479 	    struct inodedep **);
480 static	int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
481 static	int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **);
482 static	int pagedep_find(struct pagedep_hashhead *, ino_t, ufs_lbn_t,
483 	    struct mount *mp, int, struct pagedep **);
484 static	void pause_timer(void *);
485 static	int request_cleanup(int);
486 static	int process_worklist_item(struct mount *, int);
487 static	void add_to_worklist(struct worklist *);
488 
489 /*
490  * Exported softdep operations.
491  */
492 static	void softdep_disk_io_initiation(struct buf *);
493 static	void softdep_disk_write_complete(struct buf *);
494 static	void softdep_deallocate_dependencies(struct buf *);
495 static	int softdep_count_dependencies(struct buf *bp, int);
496 
497 static struct mtx lk;
498 MTX_SYSINIT(softdep_lock, &lk, "Softdep Lock", MTX_DEF);
499 
500 #define ACQUIRE_LOCK(lk)		mtx_lock(lk)
501 #define FREE_LOCK(lk)			mtx_unlock(lk)
502 
503 /*
504  * Worklist queue management.
505  * These routines require that the lock be held.
506  */
507 #ifndef /* NOT */ DEBUG
508 #define WORKLIST_INSERT(head, item) do {	\
509 	(item)->wk_state |= ONWORKLIST;		\
510 	LIST_INSERT_HEAD(head, item, wk_list);	\
511 } while (0)
512 #define WORKLIST_REMOVE(item) do {		\
513 	(item)->wk_state &= ~ONWORKLIST;	\
514 	LIST_REMOVE(item, wk_list);		\
515 } while (0)
516 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
517 
518 #else /* DEBUG */
519 static	void worklist_insert(struct workhead *, struct worklist *);
520 static	void worklist_remove(struct worklist *);
521 static	void workitem_free(struct worklist *, int);
522 
523 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
524 #define WORKLIST_REMOVE(item) worklist_remove(item)
525 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
526 
527 static void
528 worklist_insert(head, item)
529 	struct workhead *head;
530 	struct worklist *item;
531 {
532 
533 	mtx_assert(&lk, MA_OWNED);
534 	if (item->wk_state & ONWORKLIST)
535 		panic("worklist_insert: already on list");
536 	item->wk_state |= ONWORKLIST;
537 	LIST_INSERT_HEAD(head, item, wk_list);
538 }
539 
540 static void
541 worklist_remove(item)
542 	struct worklist *item;
543 {
544 
545 	mtx_assert(&lk, MA_OWNED);
546 	if ((item->wk_state & ONWORKLIST) == 0)
547 		panic("worklist_remove: not on list");
548 	item->wk_state &= ~ONWORKLIST;
549 	LIST_REMOVE(item, wk_list);
550 }
551 
552 static void
553 workitem_free(item, type)
554 	struct worklist *item;
555 	int type;
556 {
557 
558 	if (item->wk_state & ONWORKLIST)
559 		panic("workitem_free: still on list");
560 	if (item->wk_type != type)
561 		panic("workitem_free: type mismatch");
562 	FREE(item, DtoM(type));
563 }
564 #endif /* DEBUG */
565 
566 /*
567  * Workitem queue management
568  */
569 static struct workhead softdep_workitem_pending;
570 static struct worklist *worklist_tail;
571 static int num_on_worklist;	/* number of worklist items to be processed */
572 static int softdep_worklist_busy; /* 1 => trying to do unmount */
573 static int softdep_worklist_req; /* serialized waiters */
574 static int max_softdeps;	/* maximum number of structs before slowdown */
575 static int maxindirdeps = 50;	/* max number of indirdeps before slowdown */
576 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
577 static int proc_waiting;	/* tracks whether we have a timeout posted */
578 static int *stat_countp;	/* statistic to count in proc_waiting timeout */
579 static struct callout_handle handle; /* handle on posted proc_waiting timeout */
580 static struct thread *filesys_syncer; /* proc of filesystem syncer process */
581 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
582 #define FLUSH_INODES		1
583 static int req_clear_remove;	/* syncer process flush some freeblks */
584 #define FLUSH_REMOVE		2
585 #define FLUSH_REMOVE_WAIT	3
586 /*
587  * runtime statistics
588  */
589 static int stat_worklist_push;	/* number of worklist cleanups */
590 static int stat_blk_limit_push;	/* number of times block limit neared */
591 static int stat_ino_limit_push;	/* number of times inode limit neared */
592 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
593 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
594 static int stat_sync_limit_hit;	/* number of synchronous slowdowns imposed */
595 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
596 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
597 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
598 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
599 
600 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
601 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
602 SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, "");
603 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
604 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
605 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
606 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
607 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
608 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
609 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
610 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
611 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
612 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
613 SYSCTL_INT(_debug, OID_AUTO, worklist_num, CTLFLAG_RD, &num_on_worklist, 0, "");
614 
615 SYSCTL_DECL(_vfs_ffs);
616 
617 static int compute_summary_at_mount = 0;	/* Whether to recompute the summary at mount time */
618 SYSCTL_INT(_vfs_ffs, OID_AUTO, compute_summary_at_mount, CTLFLAG_RW,
619 	   &compute_summary_at_mount, 0, "Recompute summary at mount");
620 
621 /*
622  * Add an item to the end of the work queue.
623  * This routine requires that the lock be held.
624  * This is the only routine that adds items to the list.
625  * The following routine is the only one that removes items
626  * and does so in order from first to last.
627  */
628 static void
629 add_to_worklist(wk)
630 	struct worklist *wk;
631 {
632 
633 	mtx_assert(&lk, MA_OWNED);
634 	if (wk->wk_state & ONWORKLIST)
635 		panic("add_to_worklist: already on list");
636 	wk->wk_state |= ONWORKLIST;
637 	if (LIST_FIRST(&softdep_workitem_pending) == NULL)
638 		LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
639 	else
640 		LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
641 	worklist_tail = wk;
642 	num_on_worklist += 1;
643 }
644 
645 /*
646  * Process that runs once per second to handle items in the background queue.
647  *
648  * Note that we ensure that everything is done in the order in which they
649  * appear in the queue. The code below depends on this property to ensure
650  * that blocks of a file are freed before the inode itself is freed. This
651  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
652  * until all the old ones have been purged from the dependency lists.
653  */
654 int
655 softdep_process_worklist(matchmnt)
656 	struct mount *matchmnt;
657 {
658 	struct thread *td = curthread;
659 	int cnt, matchcnt, loopcount;
660 	long starttime;
661 
662 	/*
663 	 * Record the process identifier of our caller so that we can give
664 	 * this process preferential treatment in request_cleanup below.
665 	 */
666 	filesys_syncer = td;
667 	matchcnt = 0;
668 
669 	/*
670 	 * There is no danger of having multiple processes run this
671 	 * code, but we have to single-thread it when softdep_flushfiles()
672 	 * is in operation to get an accurate count of the number of items
673 	 * related to its mount point that are in the list.
674 	 */
675 	ACQUIRE_LOCK(&lk);
676 	if (matchmnt == NULL) {
677 		if (softdep_worklist_busy < 0) {
678 			FREE_LOCK(&lk);
679 			return(-1);
680 		}
681 		softdep_worklist_busy += 1;
682 	}
683 
684 	/*
685 	 * If requested, try removing inode or removal dependencies.
686 	 */
687 	if (req_clear_inodedeps) {
688 		clear_inodedeps(td);
689 		req_clear_inodedeps -= 1;
690 		wakeup_one(&proc_waiting);
691 	}
692 	if (req_clear_remove) {
693 		clear_remove(td);
694 		req_clear_remove -= 1;
695 		wakeup_one(&proc_waiting);
696 	}
697 	loopcount = 1;
698 	starttime = time_second;
699 	while (num_on_worklist > 0) {
700 		if ((cnt = process_worklist_item(matchmnt, 0)) == -1)
701 			break;
702 		else
703 			matchcnt += cnt;
704 
705 		/*
706 		 * If a umount operation wants to run the worklist
707 		 * accurately, abort.
708 		 */
709 		if (softdep_worklist_req && matchmnt == NULL) {
710 			matchcnt = -1;
711 			break;
712 		}
713 
714 		/*
715 		 * If requested, try removing inode or removal dependencies.
716 		 */
717 		if (req_clear_inodedeps) {
718 			clear_inodedeps(td);
719 			req_clear_inodedeps -= 1;
720 			wakeup_one(&proc_waiting);
721 		}
722 		if (req_clear_remove) {
723 			clear_remove(td);
724 			req_clear_remove -= 1;
725 			wakeup_one(&proc_waiting);
726 		}
727 		/*
728 		 * We do not generally want to stop for buffer space, but if
729 		 * we are really being a buffer hog, we will stop and wait.
730 		 */
731 		if (loopcount++ % 128 == 0) {
732 			FREE_LOCK(&lk);
733 			bwillwrite();
734 			ACQUIRE_LOCK(&lk);
735 		}
736 		/*
737 		 * Never allow processing to run for more than one
738 		 * second. Otherwise the other syncer tasks may get
739 		 * excessively backlogged.
740 		 */
741 		if (starttime != time_second && matchmnt == NULL) {
742 			matchcnt = -1;
743 			break;
744 		}
745 	}
746 	if (matchmnt == NULL) {
747 		softdep_worklist_busy -= 1;
748 		if (softdep_worklist_req && softdep_worklist_busy == 0)
749 			wakeup(&softdep_worklist_req);
750 	}
751 	FREE_LOCK(&lk);
752 	return (matchcnt);
753 }
754 
755 /*
756  * Process one item on the worklist.
757  */
758 static int
759 process_worklist_item(matchmnt, flags)
760 	struct mount *matchmnt;
761 	int flags;
762 {
763 	struct worklist *wk, *wkend;
764 	struct mount *mp;
765 	struct vnode *vp;
766 	int matchcnt = 0;
767 
768 	mtx_assert(&lk, MA_OWNED);
769 	/*
770 	 * If we are being called because of a process doing a
771 	 * copy-on-write, then it is not safe to write as we may
772 	 * recurse into the copy-on-write routine.
773 	 */
774 	if (curthread->td_pflags & TDP_COWINPROGRESS)
775 		return (-1);
776 	/*
777 	 * Normally we just process each item on the worklist in order.
778 	 * However, if we are in a situation where we cannot lock any
779 	 * inodes, we have to skip over any dirrem requests whose
780 	 * vnodes are resident and locked.
781 	 */
782 	vp = NULL;
783 	LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
784 		if (wk->wk_state & INPROGRESS)
785 			continue;
786 		if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
787 			break;
788 		wk->wk_state |= INPROGRESS;
789 		FREE_LOCK(&lk);
790 		ffs_vget(WK_DIRREM(wk)->dm_mnt, WK_DIRREM(wk)->dm_oldinum,
791 		    LK_NOWAIT | LK_EXCLUSIVE, &vp);
792 		ACQUIRE_LOCK(&lk);
793 		wk->wk_state &= ~INPROGRESS;
794 		if (vp != NULL)
795 			break;
796 	}
797 	if (wk == 0)
798 		return (-1);
799 	/*
800 	 * Remove the item to be processed. If we are removing the last
801 	 * item on the list, we need to recalculate the tail pointer.
802 	 * As this happens rarely and usually when the list is short,
803 	 * we just run down the list to find it rather than tracking it
804 	 * in the above loop.
805 	 */
806 	WORKLIST_REMOVE(wk);
807 	if (wk == worklist_tail) {
808 		LIST_FOREACH(wkend, &softdep_workitem_pending, wk_list)
809 			if (LIST_NEXT(wkend, wk_list) == NULL)
810 				break;
811 		worklist_tail = wkend;
812 	}
813 	num_on_worklist -= 1;
814 	FREE_LOCK(&lk);
815 	switch (wk->wk_type) {
816 
817 	case D_DIRREM:
818 		/* removal of a directory entry */
819 		mp = WK_DIRREM(wk)->dm_mnt;
820 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
821 			panic("%s: dirrem on suspended filesystem",
822 				"process_worklist_item");
823 		if (mp == matchmnt)
824 			matchcnt += 1;
825 		handle_workitem_remove(WK_DIRREM(wk), vp);
826 		break;
827 
828 	case D_FREEBLKS:
829 		/* releasing blocks and/or fragments from a file */
830 		mp = WK_FREEBLKS(wk)->fb_mnt;
831 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
832 			panic("%s: freeblks on suspended filesystem",
833 				"process_worklist_item");
834 		if (mp == matchmnt)
835 			matchcnt += 1;
836 		handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT);
837 		break;
838 
839 	case D_FREEFRAG:
840 		/* releasing a fragment when replaced as a file grows */
841 		mp = WK_FREEFRAG(wk)->ff_mnt;
842 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
843 			panic("%s: freefrag on suspended filesystem",
844 				"process_worklist_item");
845 		if (mp == matchmnt)
846 			matchcnt += 1;
847 		handle_workitem_freefrag(WK_FREEFRAG(wk));
848 		break;
849 
850 	case D_FREEFILE:
851 		/* releasing an inode when its link count drops to 0 */
852 		mp = WK_FREEFILE(wk)->fx_mnt;
853 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
854 			panic("%s: freefile on suspended filesystem",
855 				"process_worklist_item");
856 		if (mp == matchmnt)
857 			matchcnt += 1;
858 		handle_workitem_freefile(WK_FREEFILE(wk));
859 		break;
860 
861 	default:
862 		panic("%s_process_worklist: Unknown type %s",
863 		    "softdep", TYPENAME(wk->wk_type));
864 		/* NOTREACHED */
865 	}
866 	ACQUIRE_LOCK(&lk);
867 	return (matchcnt);
868 }
869 
870 /*
871  * Move dependencies from one buffer to another.
872  */
873 void
874 softdep_move_dependencies(oldbp, newbp)
875 	struct buf *oldbp;
876 	struct buf *newbp;
877 {
878 	struct worklist *wk, *wktail;
879 
880 	if (LIST_FIRST(&newbp->b_dep) != NULL)
881 		panic("softdep_move_dependencies: need merge code");
882 	wktail = 0;
883 	ACQUIRE_LOCK(&lk);
884 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
885 		LIST_REMOVE(wk, wk_list);
886 		if (wktail == 0)
887 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
888 		else
889 			LIST_INSERT_AFTER(wktail, wk, wk_list);
890 		wktail = wk;
891 	}
892 	FREE_LOCK(&lk);
893 }
894 
895 /*
896  * Purge the work list of all items associated with a particular mount point.
897  */
898 int
899 softdep_flushworklist(oldmnt, countp, td)
900 	struct mount *oldmnt;
901 	int *countp;
902 	struct thread *td;
903 {
904 	struct vnode *devvp;
905 	int count, error = 0;
906 
907 	/*
908 	 * Await our turn to clear out the queue, then serialize access.
909 	 */
910 	ACQUIRE_LOCK(&lk);
911 	while (softdep_worklist_busy) {
912 		softdep_worklist_req += 1;
913 		msleep(&softdep_worklist_req, &lk, PRIBIO, "softflush", 0);
914 		softdep_worklist_req -= 1;
915 	}
916 	softdep_worklist_busy = -1;
917 	FREE_LOCK(&lk);
918 	/*
919 	 * Alternately flush the block device associated with the mount
920 	 * point and process any dependencies that the flushing
921 	 * creates. We continue until no more worklist dependencies
922 	 * are found.
923 	 */
924 	*countp = 0;
925 	devvp = VFSTOUFS(oldmnt)->um_devvp;
926 	while ((count = softdep_process_worklist(oldmnt)) > 0) {
927 		*countp += count;
928 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td);
929 		error = VOP_FSYNC(devvp, MNT_WAIT, td);
930 		VOP_UNLOCK(devvp, 0, td);
931 		if (error)
932 			break;
933 	}
934 	ACQUIRE_LOCK(&lk);
935 	softdep_worklist_busy = 0;
936 	if (softdep_worklist_req)
937 		wakeup(&softdep_worklist_req);
938 	FREE_LOCK(&lk);
939 	return (error);
940 }
941 
942 /*
943  * Flush all vnodes and worklist items associated with a specified mount point.
944  */
945 int
946 softdep_flushfiles(oldmnt, flags, td)
947 	struct mount *oldmnt;
948 	int flags;
949 	struct thread *td;
950 {
951 	int error, count, loopcnt;
952 
953 	error = 0;
954 
955 	/*
956 	 * Alternately flush the vnodes associated with the mount
957 	 * point and process any dependencies that the flushing
958 	 * creates. In theory, this loop can happen at most twice,
959 	 * but we give it a few extra just to be sure.
960 	 */
961 	for (loopcnt = 10; loopcnt > 0; loopcnt--) {
962 		/*
963 		 * Do another flush in case any vnodes were brought in
964 		 * as part of the cleanup operations.
965 		 */
966 		if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
967 			break;
968 		if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 ||
969 		    count == 0)
970 			break;
971 	}
972 	/*
973 	 * If we are unmounting then it is an error to fail. If we
974 	 * are simply trying to downgrade to read-only, then filesystem
975 	 * activity can keep us busy forever, so we just fail with EBUSY.
976 	 */
977 	if (loopcnt == 0) {
978 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
979 			panic("softdep_flushfiles: looping");
980 		error = EBUSY;
981 	}
982 	return (error);
983 }
984 
985 /*
986  * Structure hashing.
987  *
988  * There are three types of structures that can be looked up:
989  *	1) pagedep structures identified by mount point, inode number,
990  *	   and logical block.
991  *	2) inodedep structures identified by mount point and inode number.
992  *	3) newblk structures identified by mount point and
993  *	   physical block number.
994  *
995  * The "pagedep" and "inodedep" dependency structures are hashed
996  * separately from the file blocks and inodes to which they correspond.
997  * This separation helps when the in-memory copy of an inode or
998  * file block must be replaced. It also obviates the need to access
999  * an inode or file page when simply updating (or de-allocating)
1000  * dependency structures. Lookup of newblk structures is needed to
1001  * find newly allocated blocks when trying to associate them with
1002  * their allocdirect or allocindir structure.
1003  *
1004  * The lookup routines optionally create and hash a new instance when
1005  * an existing entry is not found.
1006  */
1007 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
1008 #define NODELAY		0x0002	/* cannot do background work */
1009 
1010 /*
1011  * Structures and routines associated with pagedep caching.
1012  */
1013 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
1014 u_long	pagedep_hash;		/* size of hash table - 1 */
1015 #define	PAGEDEP_HASH(mp, inum, lbn) \
1016 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
1017 	    pagedep_hash])
1018 
1019 static int
1020 pagedep_find(pagedephd, ino, lbn, mp, flags, pagedeppp)
1021 	struct pagedep_hashhead *pagedephd;
1022 	ino_t ino;
1023 	ufs_lbn_t lbn;
1024 	struct mount *mp;
1025 	int flags;
1026 	struct pagedep **pagedeppp;
1027 {
1028 	struct pagedep *pagedep;
1029 
1030 	LIST_FOREACH(pagedep, pagedephd, pd_hash)
1031 		if (ino == pagedep->pd_ino &&
1032 		    lbn == pagedep->pd_lbn &&
1033 		    mp == pagedep->pd_mnt)
1034 			break;
1035 	if (pagedep) {
1036 		*pagedeppp = pagedep;
1037 		if ((flags & DEPALLOC) != 0 &&
1038 		    (pagedep->pd_state & ONWORKLIST) == 0)
1039 			return (0);
1040 		return (1);
1041 	}
1042 	*pagedeppp = NULL;
1043 	return (0);
1044 }
1045 /*
1046  * Look up a pagedep. Return 1 if found, 0 if not found or found
1047  * when asked to allocate but not associated with any buffer.
1048  * If not found, allocate if DEPALLOC flag is passed.
1049  * Found or allocated entry is returned in pagedeppp.
1050  * This routine must be called with splbio interrupts blocked.
1051  */
1052 static int
1053 pagedep_lookup(ip, lbn, flags, pagedeppp)
1054 	struct inode *ip;
1055 	ufs_lbn_t lbn;
1056 	int flags;
1057 	struct pagedep **pagedeppp;
1058 {
1059 	struct pagedep *pagedep;
1060 	struct pagedep_hashhead *pagedephd;
1061 	struct mount *mp;
1062 	int ret;
1063 	int i;
1064 
1065 	mtx_assert(&lk, MA_OWNED);
1066 	mp = ITOV(ip)->v_mount;
1067 	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
1068 
1069 	ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp);
1070 	if (*pagedeppp || (flags & DEPALLOC) == 0)
1071 		return (ret);
1072 	FREE_LOCK(&lk);
1073 	MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep),
1074 	    M_PAGEDEP, M_SOFTDEP_FLAGS|M_ZERO);
1075 	ACQUIRE_LOCK(&lk);
1076 	ret = pagedep_find(pagedephd, ip->i_number, lbn, mp, flags, pagedeppp);
1077 	if (*pagedeppp) {
1078 		FREE(pagedep, M_PAGEDEP);
1079 		return (ret);
1080 	}
1081 	pagedep->pd_list.wk_type = D_PAGEDEP;
1082 	pagedep->pd_mnt = mp;
1083 	pagedep->pd_ino = ip->i_number;
1084 	pagedep->pd_lbn = lbn;
1085 	LIST_INIT(&pagedep->pd_dirremhd);
1086 	LIST_INIT(&pagedep->pd_pendinghd);
1087 	for (i = 0; i < DAHASHSZ; i++)
1088 		LIST_INIT(&pagedep->pd_diraddhd[i]);
1089 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
1090 	*pagedeppp = pagedep;
1091 	return (0);
1092 }
1093 
1094 /*
1095  * Structures and routines associated with inodedep caching.
1096  */
1097 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
1098 static u_long	inodedep_hash;	/* size of hash table - 1 */
1099 static long	num_inodedep;	/* number of inodedep allocated */
1100 #define	INODEDEP_HASH(fs, inum) \
1101       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
1102 
1103 static int
1104 inodedep_find(inodedephd, fs, inum, inodedeppp)
1105 	struct inodedep_hashhead *inodedephd;
1106 	struct fs *fs;
1107 	ino_t inum;
1108 	struct inodedep **inodedeppp;
1109 {
1110 	struct inodedep *inodedep;
1111 
1112 	LIST_FOREACH(inodedep, inodedephd, id_hash)
1113 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
1114 			break;
1115 	if (inodedep) {
1116 		*inodedeppp = inodedep;
1117 		return (1);
1118 	}
1119 	*inodedeppp = NULL;
1120 
1121 	return (0);
1122 }
1123 /*
1124  * Look up an inodedep. Return 1 if found, 0 if not found.
1125  * If not found, allocate if DEPALLOC flag is passed.
1126  * Found or allocated entry is returned in inodedeppp.
1127  * This routine must be called with splbio interrupts blocked.
1128  */
1129 static int
1130 inodedep_lookup(fs, inum, flags, inodedeppp)
1131 	struct fs *fs;
1132 	ino_t inum;
1133 	int flags;
1134 	struct inodedep **inodedeppp;
1135 {
1136 	struct inodedep *inodedep;
1137 	struct inodedep_hashhead *inodedephd;
1138 
1139 	mtx_assert(&lk, MA_OWNED);
1140 	inodedephd = INODEDEP_HASH(fs, inum);
1141 
1142 	if (inodedep_find(inodedephd, fs, inum, inodedeppp))
1143 		return (1);
1144 	if ((flags & DEPALLOC) == 0)
1145 		return (0);
1146 	/*
1147 	 * If we are over our limit, try to improve the situation.
1148 	 */
1149 	if (num_inodedep > max_softdeps  && (flags & NODELAY) == 0)
1150 		request_cleanup(FLUSH_INODES);
1151 	FREE_LOCK(&lk);
1152 	MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1153 		M_INODEDEP, M_SOFTDEP_FLAGS);
1154 	ACQUIRE_LOCK(&lk);
1155 	if (inodedep_find(inodedephd, fs, inum, inodedeppp)) {
1156 		FREE(inodedep, M_INODEDEP);
1157 		return (1);
1158 	}
1159 	num_inodedep += 1;
1160 	inodedep->id_list.wk_type = D_INODEDEP;
1161 	inodedep->id_fs = fs;
1162 	inodedep->id_ino = inum;
1163 	inodedep->id_state = ALLCOMPLETE;
1164 	inodedep->id_nlinkdelta = 0;
1165 	inodedep->id_savedino1 = NULL;
1166 	inodedep->id_savedsize = -1;
1167 	inodedep->id_savedextsize = -1;
1168 	inodedep->id_buf = NULL;
1169 	LIST_INIT(&inodedep->id_pendinghd);
1170 	LIST_INIT(&inodedep->id_inowait);
1171 	LIST_INIT(&inodedep->id_bufwait);
1172 	TAILQ_INIT(&inodedep->id_inoupdt);
1173 	TAILQ_INIT(&inodedep->id_newinoupdt);
1174 	TAILQ_INIT(&inodedep->id_extupdt);
1175 	TAILQ_INIT(&inodedep->id_newextupdt);
1176 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1177 	*inodedeppp = inodedep;
1178 	return (0);
1179 }
1180 
1181 /*
1182  * Structures and routines associated with newblk caching.
1183  */
1184 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1185 u_long	newblk_hash;		/* size of hash table - 1 */
1186 #define	NEWBLK_HASH(fs, inum) \
1187 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1188 
1189 static int
1190 newblk_find(newblkhd, fs, newblkno, newblkpp)
1191 	struct newblk_hashhead *newblkhd;
1192 	struct fs *fs;
1193 	ufs2_daddr_t newblkno;
1194 	struct newblk **newblkpp;
1195 {
1196 	struct newblk *newblk;
1197 
1198 	LIST_FOREACH(newblk, newblkhd, nb_hash)
1199 		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1200 			break;
1201 	if (newblk) {
1202 		*newblkpp = newblk;
1203 		return (1);
1204 	}
1205 	*newblkpp = NULL;
1206 	return (0);
1207 }
1208 
1209 /*
1210  * Look up a newblk. Return 1 if found, 0 if not found.
1211  * If not found, allocate if DEPALLOC flag is passed.
1212  * Found or allocated entry is returned in newblkpp.
1213  */
1214 static int
1215 newblk_lookup(fs, newblkno, flags, newblkpp)
1216 	struct fs *fs;
1217 	ufs2_daddr_t newblkno;
1218 	int flags;
1219 	struct newblk **newblkpp;
1220 {
1221 	struct newblk *newblk;
1222 	struct newblk_hashhead *newblkhd;
1223 
1224 	newblkhd = NEWBLK_HASH(fs, newblkno);
1225 	if (newblk_find(newblkhd, fs, newblkno, newblkpp))
1226 		return (1);
1227 	if ((flags & DEPALLOC) == 0)
1228 		return (0);
1229 	FREE_LOCK(&lk);
1230 	MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1231 		M_NEWBLK, M_SOFTDEP_FLAGS);
1232 	ACQUIRE_LOCK(&lk);
1233 	if (newblk_find(newblkhd, fs, newblkno, newblkpp)) {
1234 		FREE(newblk, M_NEWBLK);
1235 		return (1);
1236 	}
1237 	newblk->nb_state = 0;
1238 	newblk->nb_fs = fs;
1239 	newblk->nb_newblkno = newblkno;
1240 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1241 	*newblkpp = newblk;
1242 	return (0);
1243 }
1244 
1245 /*
1246  * Executed during filesystem system initialization before
1247  * mounting any filesystems.
1248  */
1249 void
1250 softdep_initialize()
1251 {
1252 
1253 	LIST_INIT(&mkdirlisthd);
1254 	LIST_INIT(&softdep_workitem_pending);
1255 	max_softdeps = desiredvnodes * 4;
1256 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1257 	    &pagedep_hash);
1258 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1259 	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1260 
1261 	/* hooks through which the main kernel code calls us */
1262 	softdep_process_worklist_hook = softdep_process_worklist;
1263 
1264 	/* initialise bioops hack */
1265 	bioops.io_start = softdep_disk_io_initiation;
1266 	bioops.io_complete = softdep_disk_write_complete;
1267 	bioops.io_deallocate = softdep_deallocate_dependencies;
1268 	bioops.io_countdeps = softdep_count_dependencies;
1269 }
1270 
1271 /*
1272  * Executed after all filesystems have been unmounted during
1273  * filesystem module unload.
1274  */
1275 void
1276 softdep_uninitialize()
1277 {
1278 
1279 	softdep_process_worklist_hook = NULL;
1280 	hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash);
1281 	hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash);
1282 	hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash);
1283 }
1284 
1285 /*
1286  * Called at mount time to notify the dependency code that a
1287  * filesystem wishes to use it.
1288  */
1289 int
1290 softdep_mount(devvp, mp, fs, cred)
1291 	struct vnode *devvp;
1292 	struct mount *mp;
1293 	struct fs *fs;
1294 	struct ucred *cred;
1295 {
1296 	struct csum_total cstotal;
1297 	struct cg *cgp;
1298 	struct buf *bp;
1299 	int error, cyl;
1300 
1301 	mp->mnt_flag &= ~MNT_ASYNC;
1302 	mp->mnt_flag |= MNT_SOFTDEP;
1303 	/*
1304 	 * When doing soft updates, the counters in the
1305 	 * superblock may have gotten out of sync. Recomputation
1306 	 * can take a long time and can be deferred for background
1307 	 * fsck.  However, the old behavior of scanning the cylinder
1308 	 * groups and recalculating them at mount time is available
1309 	 * by setting vfs.ffs.compute_summary_at_mount to one.
1310 	 */
1311 	if (compute_summary_at_mount == 0 || fs->fs_clean != 0)
1312 		return (0);
1313 	bzero(&cstotal, sizeof cstotal);
1314 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1315 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1316 		    fs->fs_cgsize, cred, &bp)) != 0) {
1317 			brelse(bp);
1318 			return (error);
1319 		}
1320 		cgp = (struct cg *)bp->b_data;
1321 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1322 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1323 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1324 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1325 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
1326 		brelse(bp);
1327 	}
1328 #ifdef DEBUG
1329 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1330 		printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
1331 #endif
1332 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1333 	return (0);
1334 }
1335 
1336 /*
1337  * Protecting the freemaps (or bitmaps).
1338  *
1339  * To eliminate the need to execute fsck before mounting a filesystem
1340  * after a power failure, one must (conservatively) guarantee that the
1341  * on-disk copy of the bitmaps never indicate that a live inode or block is
1342  * free.  So, when a block or inode is allocated, the bitmap should be
1343  * updated (on disk) before any new pointers.  When a block or inode is
1344  * freed, the bitmap should not be updated until all pointers have been
1345  * reset.  The latter dependency is handled by the delayed de-allocation
1346  * approach described below for block and inode de-allocation.  The former
1347  * dependency is handled by calling the following procedure when a block or
1348  * inode is allocated. When an inode is allocated an "inodedep" is created
1349  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1350  * Each "inodedep" is also inserted into the hash indexing structure so
1351  * that any additional link additions can be made dependent on the inode
1352  * allocation.
1353  *
1354  * The ufs filesystem maintains a number of free block counts (e.g., per
1355  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1356  * in addition to the bitmaps.  These counts are used to improve efficiency
1357  * during allocation and therefore must be consistent with the bitmaps.
1358  * There is no convenient way to guarantee post-crash consistency of these
1359  * counts with simple update ordering, for two main reasons: (1) The counts
1360  * and bitmaps for a single cylinder group block are not in the same disk
1361  * sector.  If a disk write is interrupted (e.g., by power failure), one may
1362  * be written and the other not.  (2) Some of the counts are located in the
1363  * superblock rather than the cylinder group block. So, we focus our soft
1364  * updates implementation on protecting the bitmaps. When mounting a
1365  * filesystem, we recompute the auxiliary counts from the bitmaps.
1366  */
1367 
1368 /*
1369  * Called just after updating the cylinder group block to allocate an inode.
1370  */
1371 void
1372 softdep_setup_inomapdep(bp, ip, newinum)
1373 	struct buf *bp;		/* buffer for cylgroup block with inode map */
1374 	struct inode *ip;	/* inode related to allocation */
1375 	ino_t newinum;		/* new inode number being allocated */
1376 {
1377 	struct inodedep *inodedep;
1378 	struct bmsafemap *bmsafemap;
1379 
1380 	/*
1381 	 * Create a dependency for the newly allocated inode.
1382 	 * Panic if it already exists as something is seriously wrong.
1383 	 * Otherwise add it to the dependency list for the buffer holding
1384 	 * the cylinder group map from which it was allocated.
1385 	 */
1386 	ACQUIRE_LOCK(&lk);
1387 	if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep)))
1388 		panic("softdep_setup_inomapdep: found inode");
1389 	inodedep->id_buf = bp;
1390 	inodedep->id_state &= ~DEPCOMPLETE;
1391 	bmsafemap = bmsafemap_lookup(bp);
1392 	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1393 	FREE_LOCK(&lk);
1394 }
1395 
1396 /*
1397  * Called just after updating the cylinder group block to
1398  * allocate block or fragment.
1399  */
1400 void
1401 softdep_setup_blkmapdep(bp, fs, newblkno)
1402 	struct buf *bp;		/* buffer for cylgroup block with block map */
1403 	struct fs *fs;		/* filesystem doing allocation */
1404 	ufs2_daddr_t newblkno;	/* number of newly allocated block */
1405 {
1406 	struct newblk *newblk;
1407 	struct bmsafemap *bmsafemap;
1408 
1409 	/*
1410 	 * Create a dependency for the newly allocated block.
1411 	 * Add it to the dependency list for the buffer holding
1412 	 * the cylinder group map from which it was allocated.
1413 	 */
1414 	ACQUIRE_LOCK(&lk);
1415 	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1416 		panic("softdep_setup_blkmapdep: found block");
1417 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1418 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1419 	FREE_LOCK(&lk);
1420 }
1421 
1422 /*
1423  * Find the bmsafemap associated with a cylinder group buffer.
1424  * If none exists, create one. The buffer must be locked when
1425  * this routine is called and this routine must be called with
1426  * splbio interrupts blocked.
1427  */
1428 static struct bmsafemap *
1429 bmsafemap_lookup(bp)
1430 	struct buf *bp;
1431 {
1432 	struct bmsafemap *bmsafemap;
1433 	struct worklist *wk;
1434 
1435 	mtx_assert(&lk, MA_OWNED);
1436 	LIST_FOREACH(wk, &bp->b_dep, wk_list)
1437 		if (wk->wk_type == D_BMSAFEMAP)
1438 			return (WK_BMSAFEMAP(wk));
1439 	FREE_LOCK(&lk);
1440 	MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1441 		M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1442 	bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1443 	bmsafemap->sm_list.wk_state = 0;
1444 	bmsafemap->sm_buf = bp;
1445 	LIST_INIT(&bmsafemap->sm_allocdirecthd);
1446 	LIST_INIT(&bmsafemap->sm_allocindirhd);
1447 	LIST_INIT(&bmsafemap->sm_inodedephd);
1448 	LIST_INIT(&bmsafemap->sm_newblkhd);
1449 	ACQUIRE_LOCK(&lk);
1450 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1451 	return (bmsafemap);
1452 }
1453 
1454 /*
1455  * Direct block allocation dependencies.
1456  *
1457  * When a new block is allocated, the corresponding disk locations must be
1458  * initialized (with zeros or new data) before the on-disk inode points to
1459  * them.  Also, the freemap from which the block was allocated must be
1460  * updated (on disk) before the inode's pointer. These two dependencies are
1461  * independent of each other and are needed for all file blocks and indirect
1462  * blocks that are pointed to directly by the inode.  Just before the
1463  * "in-core" version of the inode is updated with a newly allocated block
1464  * number, a procedure (below) is called to setup allocation dependency
1465  * structures.  These structures are removed when the corresponding
1466  * dependencies are satisfied or when the block allocation becomes obsolete
1467  * (i.e., the file is deleted, the block is de-allocated, or the block is a
1468  * fragment that gets upgraded).  All of these cases are handled in
1469  * procedures described later.
1470  *
1471  * When a file extension causes a fragment to be upgraded, either to a larger
1472  * fragment or to a full block, the on-disk location may change (if the
1473  * previous fragment could not simply be extended). In this case, the old
1474  * fragment must be de-allocated, but not until after the inode's pointer has
1475  * been updated. In most cases, this is handled by later procedures, which
1476  * will construct a "freefrag" structure to be added to the workitem queue
1477  * when the inode update is complete (or obsolete).  The main exception to
1478  * this is when an allocation occurs while a pending allocation dependency
1479  * (for the same block pointer) remains.  This case is handled in the main
1480  * allocation dependency setup procedure by immediately freeing the
1481  * unreferenced fragments.
1482  */
1483 void
1484 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1485 	struct inode *ip;	/* inode to which block is being added */
1486 	ufs_lbn_t lbn;		/* block pointer within inode */
1487 	ufs2_daddr_t newblkno;	/* disk block number being added */
1488 	ufs2_daddr_t oldblkno;	/* previous block number, 0 unless frag */
1489 	long newsize;		/* size of new block */
1490 	long oldsize;		/* size of new block */
1491 	struct buf *bp;		/* bp for allocated block */
1492 {
1493 	struct allocdirect *adp, *oldadp;
1494 	struct allocdirectlst *adphead;
1495 	struct bmsafemap *bmsafemap;
1496 	struct inodedep *inodedep;
1497 	struct pagedep *pagedep;
1498 	struct newblk *newblk;
1499 
1500 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1501 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1502 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1503 	adp->ad_lbn = lbn;
1504 	adp->ad_newblkno = newblkno;
1505 	adp->ad_oldblkno = oldblkno;
1506 	adp->ad_newsize = newsize;
1507 	adp->ad_oldsize = oldsize;
1508 	adp->ad_state = ATTACHED;
1509 	LIST_INIT(&adp->ad_newdirblk);
1510 	if (newblkno == oldblkno)
1511 		adp->ad_freefrag = NULL;
1512 	else
1513 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1514 
1515 	ACQUIRE_LOCK(&lk);
1516 	if (lbn >= NDADDR) {
1517 		/* allocating an indirect block */
1518 		if (oldblkno != 0)
1519 			panic("softdep_setup_allocdirect: non-zero indir");
1520 	} else {
1521 		/*
1522 		 * Allocating a direct block.
1523 		 *
1524 		 * If we are allocating a directory block, then we must
1525 		 * allocate an associated pagedep to track additions and
1526 		 * deletions.
1527 		 */
1528 		if ((ip->i_mode & IFMT) == IFDIR &&
1529 		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1530 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1531 	}
1532 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1533 		panic("softdep_setup_allocdirect: lost block");
1534 	if (newblk->nb_state == DEPCOMPLETE) {
1535 		adp->ad_state |= DEPCOMPLETE;
1536 		adp->ad_buf = NULL;
1537 	} else {
1538 		bmsafemap = newblk->nb_bmsafemap;
1539 		adp->ad_buf = bmsafemap->sm_buf;
1540 		LIST_REMOVE(newblk, nb_deps);
1541 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1542 	}
1543 	LIST_REMOVE(newblk, nb_hash);
1544 	FREE(newblk, M_NEWBLK);
1545 
1546 	inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1547 	adp->ad_inodedep = inodedep;
1548 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1549 	/*
1550 	 * The list of allocdirects must be kept in sorted and ascending
1551 	 * order so that the rollback routines can quickly determine the
1552 	 * first uncommitted block (the size of the file stored on disk
1553 	 * ends at the end of the lowest committed fragment, or if there
1554 	 * are no fragments, at the end of the highest committed block).
1555 	 * Since files generally grow, the typical case is that the new
1556 	 * block is to be added at the end of the list. We speed this
1557 	 * special case by checking against the last allocdirect in the
1558 	 * list before laboriously traversing the list looking for the
1559 	 * insertion point.
1560 	 */
1561 	adphead = &inodedep->id_newinoupdt;
1562 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1563 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1564 		/* insert at end of list */
1565 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1566 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1567 			allocdirect_merge(adphead, adp, oldadp);
1568 		FREE_LOCK(&lk);
1569 		return;
1570 	}
1571 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1572 		if (oldadp->ad_lbn >= lbn)
1573 			break;
1574 	}
1575 	if (oldadp == NULL)
1576 		panic("softdep_setup_allocdirect: lost entry");
1577 	/* insert in middle of list */
1578 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1579 	if (oldadp->ad_lbn == lbn)
1580 		allocdirect_merge(adphead, adp, oldadp);
1581 	FREE_LOCK(&lk);
1582 }
1583 
1584 /*
1585  * Replace an old allocdirect dependency with a newer one.
1586  * This routine must be called with splbio interrupts blocked.
1587  */
1588 static void
1589 allocdirect_merge(adphead, newadp, oldadp)
1590 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
1591 	struct allocdirect *newadp;	/* allocdirect being added */
1592 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
1593 {
1594 	struct worklist *wk;
1595 	struct freefrag *freefrag;
1596 	struct newdirblk *newdirblk;
1597 
1598 	mtx_assert(&lk, MA_OWNED);
1599 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1600 	    newadp->ad_oldsize != oldadp->ad_newsize ||
1601 	    newadp->ad_lbn >= NDADDR)
1602 		panic("%s %jd != new %jd || old size %ld != new %ld",
1603 		    "allocdirect_merge: old blkno",
1604 		    (intmax_t)newadp->ad_oldblkno,
1605 		    (intmax_t)oldadp->ad_newblkno,
1606 		    newadp->ad_oldsize, oldadp->ad_newsize);
1607 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
1608 	newadp->ad_oldsize = oldadp->ad_oldsize;
1609 	/*
1610 	 * If the old dependency had a fragment to free or had never
1611 	 * previously had a block allocated, then the new dependency
1612 	 * can immediately post its freefrag and adopt the old freefrag.
1613 	 * This action is done by swapping the freefrag dependencies.
1614 	 * The new dependency gains the old one's freefrag, and the
1615 	 * old one gets the new one and then immediately puts it on
1616 	 * the worklist when it is freed by free_allocdirect. It is
1617 	 * not possible to do this swap when the old dependency had a
1618 	 * non-zero size but no previous fragment to free. This condition
1619 	 * arises when the new block is an extension of the old block.
1620 	 * Here, the first part of the fragment allocated to the new
1621 	 * dependency is part of the block currently claimed on disk by
1622 	 * the old dependency, so cannot legitimately be freed until the
1623 	 * conditions for the new dependency are fulfilled.
1624 	 */
1625 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1626 		freefrag = newadp->ad_freefrag;
1627 		newadp->ad_freefrag = oldadp->ad_freefrag;
1628 		oldadp->ad_freefrag = freefrag;
1629 	}
1630 	/*
1631 	 * If we are tracking a new directory-block allocation,
1632 	 * move it from the old allocdirect to the new allocdirect.
1633 	 */
1634 	if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
1635 		newdirblk = WK_NEWDIRBLK(wk);
1636 		WORKLIST_REMOVE(&newdirblk->db_list);
1637 		if (LIST_FIRST(&oldadp->ad_newdirblk) != NULL)
1638 			panic("allocdirect_merge: extra newdirblk");
1639 		WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list);
1640 	}
1641 	free_allocdirect(adphead, oldadp, 0);
1642 }
1643 
1644 /*
1645  * Allocate a new freefrag structure if needed.
1646  */
1647 static struct freefrag *
1648 newfreefrag(ip, blkno, size)
1649 	struct inode *ip;
1650 	ufs2_daddr_t blkno;
1651 	long size;
1652 {
1653 	struct freefrag *freefrag;
1654 	struct fs *fs;
1655 
1656 	if (blkno == 0)
1657 		return (NULL);
1658 	fs = ip->i_fs;
1659 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1660 		panic("newfreefrag: frag size");
1661 	MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1662 		M_FREEFRAG, M_SOFTDEP_FLAGS);
1663 	freefrag->ff_list.wk_type = D_FREEFRAG;
1664 	freefrag->ff_state = 0;
1665 	freefrag->ff_inum = ip->i_number;
1666 	freefrag->ff_mnt = ITOV(ip)->v_mount;
1667 	freefrag->ff_blkno = blkno;
1668 	freefrag->ff_fragsize = size;
1669 	return (freefrag);
1670 }
1671 
1672 /*
1673  * This workitem de-allocates fragments that were replaced during
1674  * file block allocation.
1675  */
1676 static void
1677 handle_workitem_freefrag(freefrag)
1678 	struct freefrag *freefrag;
1679 {
1680 	struct ufsmount *ump = VFSTOUFS(freefrag->ff_mnt);
1681 
1682 	ffs_blkfree(ump, ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
1683 	    freefrag->ff_fragsize, freefrag->ff_inum);
1684 	FREE(freefrag, M_FREEFRAG);
1685 }
1686 
1687 /*
1688  * Set up a dependency structure for an external attributes data block.
1689  * This routine follows much of the structure of softdep_setup_allocdirect.
1690  * See the description of softdep_setup_allocdirect above for details.
1691  */
1692 void
1693 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1694 	struct inode *ip;
1695 	ufs_lbn_t lbn;
1696 	ufs2_daddr_t newblkno;
1697 	ufs2_daddr_t oldblkno;
1698 	long newsize;
1699 	long oldsize;
1700 	struct buf *bp;
1701 {
1702 	struct allocdirect *adp, *oldadp;
1703 	struct allocdirectlst *adphead;
1704 	struct bmsafemap *bmsafemap;
1705 	struct inodedep *inodedep;
1706 	struct newblk *newblk;
1707 
1708 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1709 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1710 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1711 	adp->ad_lbn = lbn;
1712 	adp->ad_newblkno = newblkno;
1713 	adp->ad_oldblkno = oldblkno;
1714 	adp->ad_newsize = newsize;
1715 	adp->ad_oldsize = oldsize;
1716 	adp->ad_state = ATTACHED | EXTDATA;
1717 	LIST_INIT(&adp->ad_newdirblk);
1718 	if (newblkno == oldblkno)
1719 		adp->ad_freefrag = NULL;
1720 	else
1721 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1722 
1723 	ACQUIRE_LOCK(&lk);
1724 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1725 		panic("softdep_setup_allocext: lost block");
1726 
1727 	inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1728 	adp->ad_inodedep = inodedep;
1729 
1730 	if (newblk->nb_state == DEPCOMPLETE) {
1731 		adp->ad_state |= DEPCOMPLETE;
1732 		adp->ad_buf = NULL;
1733 	} else {
1734 		bmsafemap = newblk->nb_bmsafemap;
1735 		adp->ad_buf = bmsafemap->sm_buf;
1736 		LIST_REMOVE(newblk, nb_deps);
1737 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1738 	}
1739 	LIST_REMOVE(newblk, nb_hash);
1740 	FREE(newblk, M_NEWBLK);
1741 
1742 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1743 	if (lbn >= NXADDR)
1744 		panic("softdep_setup_allocext: lbn %lld > NXADDR",
1745 		    (long long)lbn);
1746 	/*
1747 	 * The list of allocdirects must be kept in sorted and ascending
1748 	 * order so that the rollback routines can quickly determine the
1749 	 * first uncommitted block (the size of the file stored on disk
1750 	 * ends at the end of the lowest committed fragment, or if there
1751 	 * are no fragments, at the end of the highest committed block).
1752 	 * Since files generally grow, the typical case is that the new
1753 	 * block is to be added at the end of the list. We speed this
1754 	 * special case by checking against the last allocdirect in the
1755 	 * list before laboriously traversing the list looking for the
1756 	 * insertion point.
1757 	 */
1758 	adphead = &inodedep->id_newextupdt;
1759 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1760 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1761 		/* insert at end of list */
1762 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1763 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1764 			allocdirect_merge(adphead, adp, oldadp);
1765 		FREE_LOCK(&lk);
1766 		return;
1767 	}
1768 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1769 		if (oldadp->ad_lbn >= lbn)
1770 			break;
1771 	}
1772 	if (oldadp == NULL)
1773 		panic("softdep_setup_allocext: lost entry");
1774 	/* insert in middle of list */
1775 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1776 	if (oldadp->ad_lbn == lbn)
1777 		allocdirect_merge(adphead, adp, oldadp);
1778 	FREE_LOCK(&lk);
1779 }
1780 
1781 /*
1782  * Indirect block allocation dependencies.
1783  *
1784  * The same dependencies that exist for a direct block also exist when
1785  * a new block is allocated and pointed to by an entry in a block of
1786  * indirect pointers. The undo/redo states described above are also
1787  * used here. Because an indirect block contains many pointers that
1788  * may have dependencies, a second copy of the entire in-memory indirect
1789  * block is kept. The buffer cache copy is always completely up-to-date.
1790  * The second copy, which is used only as a source for disk writes,
1791  * contains only the safe pointers (i.e., those that have no remaining
1792  * update dependencies). The second copy is freed when all pointers
1793  * are safe. The cache is not allowed to replace indirect blocks with
1794  * pending update dependencies. If a buffer containing an indirect
1795  * block with dependencies is written, these routines will mark it
1796  * dirty again. It can only be successfully written once all the
1797  * dependencies are removed. The ffs_fsync routine in conjunction with
1798  * softdep_sync_metadata work together to get all the dependencies
1799  * removed so that a file can be successfully written to disk. Three
1800  * procedures are used when setting up indirect block pointer
1801  * dependencies. The division is necessary because of the organization
1802  * of the "balloc" routine and because of the distinction between file
1803  * pages and file metadata blocks.
1804  */
1805 
1806 /*
1807  * Allocate a new allocindir structure.
1808  */
1809 static struct allocindir *
1810 newallocindir(ip, ptrno, newblkno, oldblkno)
1811 	struct inode *ip;	/* inode for file being extended */
1812 	int ptrno;		/* offset of pointer in indirect block */
1813 	ufs2_daddr_t newblkno;	/* disk block number being added */
1814 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1815 {
1816 	struct allocindir *aip;
1817 
1818 	MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1819 		M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO);
1820 	aip->ai_list.wk_type = D_ALLOCINDIR;
1821 	aip->ai_state = ATTACHED;
1822 	aip->ai_offset = ptrno;
1823 	aip->ai_newblkno = newblkno;
1824 	aip->ai_oldblkno = oldblkno;
1825 	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1826 	return (aip);
1827 }
1828 
1829 /*
1830  * Called just before setting an indirect block pointer
1831  * to a newly allocated file page.
1832  */
1833 void
1834 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1835 	struct inode *ip;	/* inode for file being extended */
1836 	ufs_lbn_t lbn;		/* allocated block number within file */
1837 	struct buf *bp;		/* buffer with indirect blk referencing page */
1838 	int ptrno;		/* offset of pointer in indirect block */
1839 	ufs2_daddr_t newblkno;	/* disk block number being added */
1840 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1841 	struct buf *nbp;	/* buffer holding allocated page */
1842 {
1843 	struct allocindir *aip;
1844 	struct pagedep *pagedep;
1845 
1846 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_page");
1847 	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1848 	ACQUIRE_LOCK(&lk);
1849 	/*
1850 	 * If we are allocating a directory page, then we must
1851 	 * allocate an associated pagedep to track additions and
1852 	 * deletions.
1853 	 */
1854 	if ((ip->i_mode & IFMT) == IFDIR &&
1855 	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1856 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1857 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1858 	setup_allocindir_phase2(bp, ip, aip);
1859 	FREE_LOCK(&lk);
1860 }
1861 
1862 /*
1863  * Called just before setting an indirect block pointer to a
1864  * newly allocated indirect block.
1865  */
1866 void
1867 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1868 	struct buf *nbp;	/* newly allocated indirect block */
1869 	struct inode *ip;	/* inode for file being extended */
1870 	struct buf *bp;		/* indirect block referencing allocated block */
1871 	int ptrno;		/* offset of pointer in indirect block */
1872 	ufs2_daddr_t newblkno;	/* disk block number being added */
1873 {
1874 	struct allocindir *aip;
1875 
1876 	ASSERT_VOP_LOCKED(ITOV(ip), "softdep_setup_allocindir_meta");
1877 	aip = newallocindir(ip, ptrno, newblkno, 0);
1878 	ACQUIRE_LOCK(&lk);
1879 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1880 	setup_allocindir_phase2(bp, ip, aip);
1881 	FREE_LOCK(&lk);
1882 }
1883 
1884 /*
1885  * Called to finish the allocation of the "aip" allocated
1886  * by one of the two routines above.
1887  */
1888 static void
1889 setup_allocindir_phase2(bp, ip, aip)
1890 	struct buf *bp;		/* in-memory copy of the indirect block */
1891 	struct inode *ip;	/* inode for file being extended */
1892 	struct allocindir *aip;	/* allocindir allocated by the above routines */
1893 {
1894 	struct worklist *wk;
1895 	struct indirdep *indirdep, *newindirdep;
1896 	struct bmsafemap *bmsafemap;
1897 	struct allocindir *oldaip;
1898 	struct freefrag *freefrag;
1899 	struct newblk *newblk;
1900 	ufs2_daddr_t blkno;
1901 
1902 	mtx_assert(&lk, MA_OWNED);
1903 	if (bp->b_lblkno >= 0)
1904 		panic("setup_allocindir_phase2: not indir blk");
1905 	for (indirdep = NULL, newindirdep = NULL; ; ) {
1906 		LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1907 			if (wk->wk_type != D_INDIRDEP)
1908 				continue;
1909 			indirdep = WK_INDIRDEP(wk);
1910 			break;
1911 		}
1912 		if (indirdep == NULL && newindirdep) {
1913 			indirdep = newindirdep;
1914 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1915 			newindirdep = NULL;
1916 		}
1917 		if (indirdep) {
1918 			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1919 			    &newblk) == 0)
1920 				panic("setup_allocindir: lost block");
1921 			if (newblk->nb_state == DEPCOMPLETE) {
1922 				aip->ai_state |= DEPCOMPLETE;
1923 				aip->ai_buf = NULL;
1924 			} else {
1925 				bmsafemap = newblk->nb_bmsafemap;
1926 				aip->ai_buf = bmsafemap->sm_buf;
1927 				LIST_REMOVE(newblk, nb_deps);
1928 				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1929 				    aip, ai_deps);
1930 			}
1931 			LIST_REMOVE(newblk, nb_hash);
1932 			FREE(newblk, M_NEWBLK);
1933 			aip->ai_indirdep = indirdep;
1934 			/*
1935 			 * Check to see if there is an existing dependency
1936 			 * for this block. If there is, merge the old
1937 			 * dependency into the new one.
1938 			 */
1939 			if (aip->ai_oldblkno == 0)
1940 				oldaip = NULL;
1941 			else
1942 
1943 				LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1944 					if (oldaip->ai_offset == aip->ai_offset)
1945 						break;
1946 			freefrag = NULL;
1947 			if (oldaip != NULL) {
1948 				if (oldaip->ai_newblkno != aip->ai_oldblkno)
1949 					panic("setup_allocindir_phase2: blkno");
1950 				aip->ai_oldblkno = oldaip->ai_oldblkno;
1951 				freefrag = aip->ai_freefrag;
1952 				aip->ai_freefrag = oldaip->ai_freefrag;
1953 				oldaip->ai_freefrag = NULL;
1954 				free_allocindir(oldaip, NULL);
1955 			}
1956 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1957 			if (ip->i_ump->um_fstype == UFS1)
1958 				((ufs1_daddr_t *)indirdep->ir_savebp->b_data)
1959 				    [aip->ai_offset] = aip->ai_oldblkno;
1960 			else
1961 				((ufs2_daddr_t *)indirdep->ir_savebp->b_data)
1962 				    [aip->ai_offset] = aip->ai_oldblkno;
1963 			FREE_LOCK(&lk);
1964 			if (freefrag != NULL)
1965 				handle_workitem_freefrag(freefrag);
1966 		} else
1967 			FREE_LOCK(&lk);
1968 		if (newindirdep) {
1969 			newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
1970 			brelse(newindirdep->ir_savebp);
1971 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1972 		}
1973 		if (indirdep) {
1974 			ACQUIRE_LOCK(&lk);
1975 			break;
1976 		}
1977 		MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1978 			M_INDIRDEP, M_SOFTDEP_FLAGS);
1979 		newindirdep->ir_list.wk_type = D_INDIRDEP;
1980 		newindirdep->ir_state = ATTACHED;
1981 		if (ip->i_ump->um_fstype == UFS1)
1982 			newindirdep->ir_state |= UFS1FMT;
1983 		LIST_INIT(&newindirdep->ir_deplisthd);
1984 		LIST_INIT(&newindirdep->ir_donehd);
1985 		if (bp->b_blkno == bp->b_lblkno) {
1986 			ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
1987 			    NULL, NULL);
1988 			bp->b_blkno = blkno;
1989 		}
1990 		newindirdep->ir_savebp =
1991 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0, 0);
1992 		BUF_KERNPROC(newindirdep->ir_savebp);
1993 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1994 		ACQUIRE_LOCK(&lk);
1995 	}
1996 }
1997 
1998 /*
1999  * Block de-allocation dependencies.
2000  *
2001  * When blocks are de-allocated, the on-disk pointers must be nullified before
2002  * the blocks are made available for use by other files.  (The true
2003  * requirement is that old pointers must be nullified before new on-disk
2004  * pointers are set.  We chose this slightly more stringent requirement to
2005  * reduce complexity.) Our implementation handles this dependency by updating
2006  * the inode (or indirect block) appropriately but delaying the actual block
2007  * de-allocation (i.e., freemap and free space count manipulation) until
2008  * after the updated versions reach stable storage.  After the disk is
2009  * updated, the blocks can be safely de-allocated whenever it is convenient.
2010  * This implementation handles only the common case of reducing a file's
2011  * length to zero. Other cases are handled by the conventional synchronous
2012  * write approach.
2013  *
2014  * The ffs implementation with which we worked double-checks
2015  * the state of the block pointers and file size as it reduces
2016  * a file's length.  Some of this code is replicated here in our
2017  * soft updates implementation.  The freeblks->fb_chkcnt field is
2018  * used to transfer a part of this information to the procedure
2019  * that eventually de-allocates the blocks.
2020  *
2021  * This routine should be called from the routine that shortens
2022  * a file's length, before the inode's size or block pointers
2023  * are modified. It will save the block pointer information for
2024  * later release and zero the inode so that the calling routine
2025  * can release it.
2026  */
2027 void
2028 softdep_setup_freeblocks(ip, length, flags)
2029 	struct inode *ip;	/* The inode whose length is to be reduced */
2030 	off_t length;		/* The new length for the file */
2031 	int flags;		/* IO_EXT and/or IO_NORMAL */
2032 {
2033 	struct freeblks *freeblks;
2034 	struct inodedep *inodedep;
2035 	struct allocdirect *adp;
2036 	struct vnode *vp;
2037 	struct buf *bp;
2038 	struct fs *fs;
2039 	ufs2_daddr_t extblocks, datablocks;
2040 	int i, delay, error;
2041 
2042 	fs = ip->i_fs;
2043 	if (length != 0)
2044 		panic("softdep_setup_freeblocks: non-zero length");
2045 	MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
2046 		M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
2047 	freeblks->fb_list.wk_type = D_FREEBLKS;
2048 	freeblks->fb_state = ATTACHED;
2049 	freeblks->fb_uid = ip->i_uid;
2050 	freeblks->fb_previousinum = ip->i_number;
2051 	freeblks->fb_devvp = ip->i_devvp;
2052 	freeblks->fb_mnt = ITOV(ip)->v_mount;
2053 	extblocks = 0;
2054 	if (fs->fs_magic == FS_UFS2_MAGIC)
2055 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
2056 	datablocks = DIP(ip, i_blocks) - extblocks;
2057 	if ((flags & IO_NORMAL) == 0) {
2058 		freeblks->fb_oldsize = 0;
2059 		freeblks->fb_chkcnt = 0;
2060 	} else {
2061 		freeblks->fb_oldsize = ip->i_size;
2062 		ip->i_size = 0;
2063 		DIP_SET(ip, i_size, 0);
2064 		freeblks->fb_chkcnt = datablocks;
2065 		for (i = 0; i < NDADDR; i++) {
2066 			freeblks->fb_dblks[i] = DIP(ip, i_db[i]);
2067 			DIP_SET(ip, i_db[i], 0);
2068 		}
2069 		for (i = 0; i < NIADDR; i++) {
2070 			freeblks->fb_iblks[i] = DIP(ip, i_ib[i]);
2071 			DIP_SET(ip, i_ib[i], 0);
2072 		}
2073 		/*
2074 		 * If the file was removed, then the space being freed was
2075 		 * accounted for then (see softdep_filereleased()). If the
2076 		 * file is merely being truncated, then we account for it now.
2077 		 */
2078 		if ((ip->i_flag & IN_SPACECOUNTED) == 0) {
2079 			UFS_LOCK(ip->i_ump);
2080 			fs->fs_pendingblocks += datablocks;
2081 			UFS_UNLOCK(ip->i_ump);
2082 		}
2083 	}
2084 	if ((flags & IO_EXT) == 0) {
2085 		freeblks->fb_oldextsize = 0;
2086 	} else {
2087 		freeblks->fb_oldextsize = ip->i_din2->di_extsize;
2088 		ip->i_din2->di_extsize = 0;
2089 		freeblks->fb_chkcnt += extblocks;
2090 		for (i = 0; i < NXADDR; i++) {
2091 			freeblks->fb_eblks[i] = ip->i_din2->di_extb[i];
2092 			ip->i_din2->di_extb[i] = 0;
2093 		}
2094 	}
2095 	DIP_SET(ip, i_blocks, DIP(ip, i_blocks) - freeblks->fb_chkcnt);
2096 	/*
2097 	 * Push the zero'ed inode to to its disk buffer so that we are free
2098 	 * to delete its dependencies below. Once the dependencies are gone
2099 	 * the buffer can be safely released.
2100 	 */
2101 	if ((error = bread(ip->i_devvp,
2102 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
2103 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
2104 		brelse(bp);
2105 		softdep_error("softdep_setup_freeblocks", error);
2106 	}
2107 	if (ip->i_ump->um_fstype == UFS1)
2108 		*((struct ufs1_dinode *)bp->b_data +
2109 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
2110 	else
2111 		*((struct ufs2_dinode *)bp->b_data +
2112 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
2113 	/*
2114 	 * Find and eliminate any inode dependencies.
2115 	 */
2116 	ACQUIRE_LOCK(&lk);
2117 	(void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
2118 	if ((inodedep->id_state & IOSTARTED) != 0)
2119 		panic("softdep_setup_freeblocks: inode busy");
2120 	/*
2121 	 * Add the freeblks structure to the list of operations that
2122 	 * must await the zero'ed inode being written to disk. If we
2123 	 * still have a bitmap dependency (delay == 0), then the inode
2124 	 * has never been written to disk, so we can process the
2125 	 * freeblks below once we have deleted the dependencies.
2126 	 */
2127 	delay = (inodedep->id_state & DEPCOMPLETE);
2128 	if (delay)
2129 		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
2130 	/*
2131 	 * Because the file length has been truncated to zero, any
2132 	 * pending block allocation dependency structures associated
2133 	 * with this inode are obsolete and can simply be de-allocated.
2134 	 * We must first merge the two dependency lists to get rid of
2135 	 * any duplicate freefrag structures, then purge the merged list.
2136 	 * If we still have a bitmap dependency, then the inode has never
2137 	 * been written to disk, so we can free any fragments without delay.
2138 	 */
2139 	if (flags & IO_NORMAL) {
2140 		merge_inode_lists(&inodedep->id_newinoupdt,
2141 		    &inodedep->id_inoupdt);
2142 		while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
2143 			free_allocdirect(&inodedep->id_inoupdt, adp, delay);
2144 	}
2145 	if (flags & IO_EXT) {
2146 		merge_inode_lists(&inodedep->id_newextupdt,
2147 		    &inodedep->id_extupdt);
2148 		while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
2149 			free_allocdirect(&inodedep->id_extupdt, adp, delay);
2150 	}
2151 	FREE_LOCK(&lk);
2152 	bdwrite(bp);
2153 	/*
2154 	 * We must wait for any I/O in progress to finish so that
2155 	 * all potential buffers on the dirty list will be visible.
2156 	 * Once they are all there, walk the list and get rid of
2157 	 * any dependencies.
2158 	 */
2159 	vp = ITOV(ip);
2160 	VI_LOCK(vp);
2161 	drain_output(vp);
2162 restart:
2163 	TAILQ_FOREACH(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs) {
2164 		if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
2165 		    ((flags & IO_NORMAL) == 0 &&
2166 		      (bp->b_xflags & BX_ALTDATA) == 0))
2167 			continue;
2168 		if ((bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT)) == NULL)
2169 			goto restart;
2170 		VI_UNLOCK(vp);
2171 		ACQUIRE_LOCK(&lk);
2172 		(void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
2173 		deallocate_dependencies(bp, inodedep);
2174 		FREE_LOCK(&lk);
2175 		bp->b_flags |= B_INVAL | B_NOCACHE;
2176 		brelse(bp);
2177 		VI_LOCK(vp);
2178 		goto restart;
2179 	}
2180 	VI_UNLOCK(vp);
2181 	ACQUIRE_LOCK(&lk);
2182 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
2183 		(void) free_inodedep(inodedep);
2184 
2185 	if(delay) {
2186 		freeblks->fb_state |= DEPCOMPLETE;
2187 		/*
2188 		 * If the inode with zeroed block pointers is now on disk
2189 		 * we can start freeing blocks. Add freeblks to the worklist
2190 		 * instead of calling  handle_workitem_freeblocks directly as
2191 		 * it is more likely that additional IO is needed to complete
2192 		 * the request here than in the !delay case.
2193 		 */
2194 		if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
2195 			add_to_worklist(&freeblks->fb_list);
2196 	}
2197 
2198 	FREE_LOCK(&lk);
2199 	/*
2200 	 * If the inode has never been written to disk (delay == 0),
2201 	 * then we can process the freeblks now that we have deleted
2202 	 * the dependencies.
2203 	 */
2204 	if (!delay)
2205 		handle_workitem_freeblocks(freeblks, 0);
2206 }
2207 
2208 /*
2209  * Reclaim any dependency structures from a buffer that is about to
2210  * be reallocated to a new vnode. The buffer must be locked, thus,
2211  * no I/O completion operations can occur while we are manipulating
2212  * its associated dependencies. The mutex is held so that other I/O's
2213  * associated with related dependencies do not occur.
2214  */
2215 static void
2216 deallocate_dependencies(bp, inodedep)
2217 	struct buf *bp;
2218 	struct inodedep *inodedep;
2219 {
2220 	struct worklist *wk;
2221 	struct indirdep *indirdep;
2222 	struct allocindir *aip;
2223 	struct pagedep *pagedep;
2224 	struct dirrem *dirrem;
2225 	struct diradd *dap;
2226 	int i;
2227 
2228 	mtx_assert(&lk, MA_OWNED);
2229 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2230 		switch (wk->wk_type) {
2231 
2232 		case D_INDIRDEP:
2233 			indirdep = WK_INDIRDEP(wk);
2234 			/*
2235 			 * None of the indirect pointers will ever be visible,
2236 			 * so they can simply be tossed. GOINGAWAY ensures
2237 			 * that allocated pointers will be saved in the buffer
2238 			 * cache until they are freed. Note that they will
2239 			 * only be able to be found by their physical address
2240 			 * since the inode mapping the logical address will
2241 			 * be gone. The save buffer used for the safe copy
2242 			 * was allocated in setup_allocindir_phase2 using
2243 			 * the physical address so it could be used for this
2244 			 * purpose. Hence we swap the safe copy with the real
2245 			 * copy, allowing the safe copy to be freed and holding
2246 			 * on to the real copy for later use in indir_trunc.
2247 			 */
2248 			if (indirdep->ir_state & GOINGAWAY)
2249 				panic("deallocate_dependencies: already gone");
2250 			indirdep->ir_state |= GOINGAWAY;
2251 			VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1;
2252 			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
2253 				free_allocindir(aip, inodedep);
2254 			if (bp->b_lblkno >= 0 ||
2255 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno)
2256 				panic("deallocate_dependencies: not indir");
2257 			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2258 			    bp->b_bcount);
2259 			WORKLIST_REMOVE(wk);
2260 			WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
2261 			continue;
2262 
2263 		case D_PAGEDEP:
2264 			pagedep = WK_PAGEDEP(wk);
2265 			/*
2266 			 * None of the directory additions will ever be
2267 			 * visible, so they can simply be tossed.
2268 			 */
2269 			for (i = 0; i < DAHASHSZ; i++)
2270 				while ((dap =
2271 				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
2272 					free_diradd(dap);
2273 			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2274 				free_diradd(dap);
2275 			/*
2276 			 * Copy any directory remove dependencies to the list
2277 			 * to be processed after the zero'ed inode is written.
2278 			 * If the inode has already been written, then they
2279 			 * can be dumped directly onto the work list.
2280 			 */
2281 			LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2282 				LIST_REMOVE(dirrem, dm_next);
2283 				dirrem->dm_dirinum = pagedep->pd_ino;
2284 				if (inodedep == NULL ||
2285 				    (inodedep->id_state & ALLCOMPLETE) ==
2286 				     ALLCOMPLETE)
2287 					add_to_worklist(&dirrem->dm_list);
2288 				else
2289 					WORKLIST_INSERT(&inodedep->id_bufwait,
2290 					    &dirrem->dm_list);
2291 			}
2292 			if ((pagedep->pd_state & NEWBLOCK) != 0) {
2293 				LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list)
2294 					if (wk->wk_type == D_NEWDIRBLK &&
2295 					    WK_NEWDIRBLK(wk)->db_pagedep ==
2296 					      pagedep)
2297 						break;
2298 				if (wk != NULL) {
2299 					WORKLIST_REMOVE(wk);
2300 					free_newdirblk(WK_NEWDIRBLK(wk));
2301 				} else
2302 					panic("deallocate_dependencies: "
2303 					      "lost pagedep");
2304 			}
2305 			WORKLIST_REMOVE(&pagedep->pd_list);
2306 			LIST_REMOVE(pagedep, pd_hash);
2307 			WORKITEM_FREE(pagedep, D_PAGEDEP);
2308 			continue;
2309 
2310 		case D_ALLOCINDIR:
2311 			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2312 			continue;
2313 
2314 		case D_ALLOCDIRECT:
2315 		case D_INODEDEP:
2316 			panic("deallocate_dependencies: Unexpected type %s",
2317 			    TYPENAME(wk->wk_type));
2318 			/* NOTREACHED */
2319 
2320 		default:
2321 			panic("deallocate_dependencies: Unknown type %s",
2322 			    TYPENAME(wk->wk_type));
2323 			/* NOTREACHED */
2324 		}
2325 	}
2326 }
2327 
2328 /*
2329  * Free an allocdirect. Generate a new freefrag work request if appropriate.
2330  * This routine must be called with splbio interrupts blocked.
2331  */
2332 static void
2333 free_allocdirect(adphead, adp, delay)
2334 	struct allocdirectlst *adphead;
2335 	struct allocdirect *adp;
2336 	int delay;
2337 {
2338 	struct newdirblk *newdirblk;
2339 	struct worklist *wk;
2340 
2341 	mtx_assert(&lk, MA_OWNED);
2342 	if ((adp->ad_state & DEPCOMPLETE) == 0)
2343 		LIST_REMOVE(adp, ad_deps);
2344 	TAILQ_REMOVE(adphead, adp, ad_next);
2345 	if ((adp->ad_state & COMPLETE) == 0)
2346 		WORKLIST_REMOVE(&adp->ad_list);
2347 	if (adp->ad_freefrag != NULL) {
2348 		if (delay)
2349 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2350 			    &adp->ad_freefrag->ff_list);
2351 		else
2352 			add_to_worklist(&adp->ad_freefrag->ff_list);
2353 	}
2354 	if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) {
2355 		newdirblk = WK_NEWDIRBLK(wk);
2356 		WORKLIST_REMOVE(&newdirblk->db_list);
2357 		if (LIST_FIRST(&adp->ad_newdirblk) != NULL)
2358 			panic("free_allocdirect: extra newdirblk");
2359 		if (delay)
2360 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2361 			    &newdirblk->db_list);
2362 		else
2363 			free_newdirblk(newdirblk);
2364 	}
2365 	WORKITEM_FREE(adp, D_ALLOCDIRECT);
2366 }
2367 
2368 /*
2369  * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
2370  * This routine must be called with splbio interrupts blocked.
2371  */
2372 static void
2373 free_newdirblk(newdirblk)
2374 	struct newdirblk *newdirblk;
2375 {
2376 	struct pagedep *pagedep;
2377 	struct diradd *dap;
2378 	int i;
2379 
2380 	mtx_assert(&lk, MA_OWNED);
2381 	/*
2382 	 * If the pagedep is still linked onto the directory buffer
2383 	 * dependency chain, then some of the entries on the
2384 	 * pd_pendinghd list may not be committed to disk yet. In
2385 	 * this case, we will simply clear the NEWBLOCK flag and
2386 	 * let the pd_pendinghd list be processed when the pagedep
2387 	 * is next written. If the pagedep is no longer on the buffer
2388 	 * dependency chain, then all the entries on the pd_pending
2389 	 * list are committed to disk and we can free them here.
2390 	 */
2391 	pagedep = newdirblk->db_pagedep;
2392 	pagedep->pd_state &= ~NEWBLOCK;
2393 	if ((pagedep->pd_state & ONWORKLIST) == 0)
2394 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
2395 			free_diradd(dap);
2396 	/*
2397 	 * If no dependencies remain, the pagedep will be freed.
2398 	 */
2399 	for (i = 0; i < DAHASHSZ; i++)
2400 		if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
2401 			break;
2402 	if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) {
2403 		LIST_REMOVE(pagedep, pd_hash);
2404 		WORKITEM_FREE(pagedep, D_PAGEDEP);
2405 	}
2406 	WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2407 }
2408 
2409 /*
2410  * Prepare an inode to be freed. The actual free operation is not
2411  * done until the zero'ed inode has been written to disk.
2412  */
2413 void
2414 softdep_freefile(pvp, ino, mode)
2415 	struct vnode *pvp;
2416 	ino_t ino;
2417 	int mode;
2418 {
2419 	struct inode *ip = VTOI(pvp);
2420 	struct inodedep *inodedep;
2421 	struct freefile *freefile;
2422 
2423 	/*
2424 	 * This sets up the inode de-allocation dependency.
2425 	 */
2426 	MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2427 		M_FREEFILE, M_SOFTDEP_FLAGS);
2428 	freefile->fx_list.wk_type = D_FREEFILE;
2429 	freefile->fx_list.wk_state = 0;
2430 	freefile->fx_mode = mode;
2431 	freefile->fx_oldinum = ino;
2432 	freefile->fx_devvp = ip->i_devvp;
2433 	freefile->fx_mnt = ITOV(ip)->v_mount;
2434 	if ((ip->i_flag & IN_SPACECOUNTED) == 0) {
2435 		UFS_LOCK(ip->i_ump);
2436 		ip->i_fs->fs_pendinginodes += 1;
2437 		UFS_UNLOCK(ip->i_ump);
2438 	}
2439 
2440 	/*
2441 	 * If the inodedep does not exist, then the zero'ed inode has
2442 	 * been written to disk. If the allocated inode has never been
2443 	 * written to disk, then the on-disk inode is zero'ed. In either
2444 	 * case we can free the file immediately.
2445 	 */
2446 	ACQUIRE_LOCK(&lk);
2447 	if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2448 	    check_inode_unwritten(inodedep)) {
2449 		FREE_LOCK(&lk);
2450 		handle_workitem_freefile(freefile);
2451 		return;
2452 	}
2453 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2454 	FREE_LOCK(&lk);
2455 }
2456 
2457 /*
2458  * Check to see if an inode has never been written to disk. If
2459  * so free the inodedep and return success, otherwise return failure.
2460  * This routine must be called with splbio interrupts blocked.
2461  *
2462  * If we still have a bitmap dependency, then the inode has never
2463  * been written to disk. Drop the dependency as it is no longer
2464  * necessary since the inode is being deallocated. We set the
2465  * ALLCOMPLETE flags since the bitmap now properly shows that the
2466  * inode is not allocated. Even if the inode is actively being
2467  * written, it has been rolled back to its zero'ed state, so we
2468  * are ensured that a zero inode is what is on the disk. For short
2469  * lived files, this change will usually result in removing all the
2470  * dependencies from the inode so that it can be freed immediately.
2471  */
2472 static int
2473 check_inode_unwritten(inodedep)
2474 	struct inodedep *inodedep;
2475 {
2476 
2477 	mtx_assert(&lk, MA_OWNED);
2478 	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2479 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2480 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2481 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
2482 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2483 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2484 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
2485 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
2486 	    inodedep->id_nlinkdelta != 0)
2487 		return (0);
2488 
2489 	/*
2490 	 * Another process might be in initiate_write_inodeblock_ufs[12]
2491 	 * trying to allocate memory without holding "Softdep Lock".
2492 	 */
2493 	if ((inodedep->id_state & IOSTARTED) != 0 &&
2494 	    inodedep->id_savedino1 == NULL)
2495 		return (0);
2496 
2497 	inodedep->id_state |= ALLCOMPLETE;
2498 	LIST_REMOVE(inodedep, id_deps);
2499 	inodedep->id_buf = NULL;
2500 	if (inodedep->id_state & ONWORKLIST)
2501 		WORKLIST_REMOVE(&inodedep->id_list);
2502 	if (inodedep->id_savedino1 != NULL) {
2503 		FREE(inodedep->id_savedino1, M_SAVEDINO);
2504 		inodedep->id_savedino1 = NULL;
2505 	}
2506 	if (free_inodedep(inodedep) == 0)
2507 		panic("check_inode_unwritten: busy inode");
2508 	return (1);
2509 }
2510 
2511 /*
2512  * Try to free an inodedep structure. Return 1 if it could be freed.
2513  */
2514 static int
2515 free_inodedep(inodedep)
2516 	struct inodedep *inodedep;
2517 {
2518 
2519 	mtx_assert(&lk, MA_OWNED);
2520 	if ((inodedep->id_state & ONWORKLIST) != 0 ||
2521 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2522 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2523 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2524 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
2525 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2526 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2527 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
2528 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
2529 	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL)
2530 		return (0);
2531 	LIST_REMOVE(inodedep, id_hash);
2532 	WORKITEM_FREE(inodedep, D_INODEDEP);
2533 	num_inodedep -= 1;
2534 	return (1);
2535 }
2536 
2537 /*
2538  * This workitem routine performs the block de-allocation.
2539  * The workitem is added to the pending list after the updated
2540  * inode block has been written to disk.  As mentioned above,
2541  * checks regarding the number of blocks de-allocated (compared
2542  * to the number of blocks allocated for the file) are also
2543  * performed in this function.
2544  */
2545 static void
2546 handle_workitem_freeblocks(freeblks, flags)
2547 	struct freeblks *freeblks;
2548 	int flags;
2549 {
2550 	struct inode *ip;
2551 	struct vnode *vp;
2552 	struct fs *fs;
2553 	struct ufsmount *ump;
2554 	int i, nblocks, level, bsize;
2555 	ufs2_daddr_t bn, blocksreleased = 0;
2556 	int error, allerror = 0;
2557 	ufs_lbn_t baselbns[NIADDR], tmpval;
2558 	int fs_pendingblocks;
2559 
2560 	ump = VFSTOUFS(freeblks->fb_mnt);
2561 	fs = ump->um_fs;
2562 	fs_pendingblocks = 0;
2563 	tmpval = 1;
2564 	baselbns[0] = NDADDR;
2565 	for (i = 1; i < NIADDR; i++) {
2566 		tmpval *= NINDIR(fs);
2567 		baselbns[i] = baselbns[i - 1] + tmpval;
2568 	}
2569 	nblocks = btodb(fs->fs_bsize);
2570 	blocksreleased = 0;
2571 	/*
2572 	 * Release all extended attribute blocks or frags.
2573 	 */
2574 	if (freeblks->fb_oldextsize > 0) {
2575 		for (i = (NXADDR - 1); i >= 0; i--) {
2576 			if ((bn = freeblks->fb_eblks[i]) == 0)
2577 				continue;
2578 			bsize = sblksize(fs, freeblks->fb_oldextsize, i);
2579 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize,
2580 			    freeblks->fb_previousinum);
2581 			blocksreleased += btodb(bsize);
2582 		}
2583 	}
2584 	/*
2585 	 * Release all data blocks or frags.
2586 	 */
2587 	if (freeblks->fb_oldsize > 0) {
2588 		/*
2589 		 * Indirect blocks first.
2590 		 */
2591 		for (level = (NIADDR - 1); level >= 0; level--) {
2592 			if ((bn = freeblks->fb_iblks[level]) == 0)
2593 				continue;
2594 			if ((error = indir_trunc(freeblks, fsbtodb(fs, bn),
2595 			    level, baselbns[level], &blocksreleased)) == 0)
2596 				allerror = error;
2597 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn,
2598 			    fs->fs_bsize, freeblks->fb_previousinum);
2599 			fs_pendingblocks += nblocks;
2600 			blocksreleased += nblocks;
2601 		}
2602 		/*
2603 		 * All direct blocks or frags.
2604 		 */
2605 		for (i = (NDADDR - 1); i >= 0; i--) {
2606 			if ((bn = freeblks->fb_dblks[i]) == 0)
2607 				continue;
2608 			bsize = sblksize(fs, freeblks->fb_oldsize, i);
2609 			ffs_blkfree(ump, fs, freeblks->fb_devvp, bn, bsize,
2610 			    freeblks->fb_previousinum);
2611 			fs_pendingblocks += btodb(bsize);
2612 			blocksreleased += btodb(bsize);
2613 		}
2614 	}
2615 	UFS_LOCK(ump);
2616 	fs->fs_pendingblocks -= fs_pendingblocks;
2617 	UFS_UNLOCK(ump);
2618 	/*
2619 	 * If we still have not finished background cleanup, then check
2620 	 * to see if the block count needs to be adjusted.
2621 	 */
2622 	if (freeblks->fb_chkcnt != blocksreleased &&
2623 	    (fs->fs_flags & FS_UNCLEAN) != 0 &&
2624 	    ffs_vget(freeblks->fb_mnt, freeblks->fb_previousinum,
2625 	    (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) {
2626 		ip = VTOI(vp);
2627 		DIP_SET(ip, i_blocks, DIP(ip, i_blocks) + \
2628 		    freeblks->fb_chkcnt - blocksreleased);
2629 		ip->i_flag |= IN_CHANGE;
2630 		vput(vp);
2631 	}
2632 
2633 #ifdef DIAGNOSTIC
2634 	if (freeblks->fb_chkcnt != blocksreleased &&
2635 	    ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0))
2636 		printf("handle_workitem_freeblocks: block count\n");
2637 	if (allerror)
2638 		softdep_error("handle_workitem_freeblks", allerror);
2639 #endif /* DIAGNOSTIC */
2640 
2641 	WORKITEM_FREE(freeblks, D_FREEBLKS);
2642 }
2643 
2644 /*
2645  * Release blocks associated with the inode ip and stored in the indirect
2646  * block dbn. If level is greater than SINGLE, the block is an indirect block
2647  * and recursive calls to indirtrunc must be used to cleanse other indirect
2648  * blocks.
2649  */
2650 static int
2651 indir_trunc(freeblks, dbn, level, lbn, countp)
2652 	struct freeblks *freeblks;
2653 	ufs2_daddr_t dbn;
2654 	int level;
2655 	ufs_lbn_t lbn;
2656 	ufs2_daddr_t *countp;
2657 {
2658 	struct buf *bp;
2659 	struct fs *fs;
2660 	struct worklist *wk;
2661 	struct indirdep *indirdep;
2662 	struct ufsmount *ump;
2663 	ufs1_daddr_t *bap1 = 0;
2664 	ufs2_daddr_t nb, *bap2 = 0;
2665 	ufs_lbn_t lbnadd;
2666 	int i, nblocks, ufs1fmt;
2667 	int error, allerror = 0;
2668 	int fs_pendingblocks;
2669 
2670 	ump = VFSTOUFS(freeblks->fb_mnt);
2671 	fs = ump->um_fs;
2672 	fs_pendingblocks = 0;
2673 	lbnadd = 1;
2674 	for (i = level; i > 0; i--)
2675 		lbnadd *= NINDIR(fs);
2676 	/*
2677 	 * Get buffer of block pointers to be freed. This routine is not
2678 	 * called until the zero'ed inode has been written, so it is safe
2679 	 * to free blocks as they are encountered. Because the inode has
2680 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
2681 	 * have to use the on-disk address and the block device for the
2682 	 * filesystem to look them up. If the file was deleted before its
2683 	 * indirect blocks were all written to disk, the routine that set
2684 	 * us up (deallocate_dependencies) will have arranged to leave
2685 	 * a complete copy of the indirect block in memory for our use.
2686 	 * Otherwise we have to read the blocks in from the disk.
2687 	 */
2688 #ifdef notyet
2689 	bp = getblk(freeblks->fb_devvp, dbn, (int)fs->fs_bsize, 0, 0,
2690 	    GB_NOCREAT);
2691 #else
2692 	bp = incore(&freeblks->fb_devvp->v_bufobj, dbn);
2693 #endif
2694 	ACQUIRE_LOCK(&lk);
2695 	if (bp != NULL && (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2696 		if (wk->wk_type != D_INDIRDEP ||
2697 		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2698 		    (indirdep->ir_state & GOINGAWAY) == 0)
2699 			panic("indir_trunc: lost indirdep");
2700 		WORKLIST_REMOVE(wk);
2701 		WORKITEM_FREE(indirdep, D_INDIRDEP);
2702 		if (LIST_FIRST(&bp->b_dep) != NULL)
2703 			panic("indir_trunc: dangling dep");
2704 		VFSTOUFS(freeblks->fb_mnt)->um_numindirdeps -= 1;
2705 		FREE_LOCK(&lk);
2706 	} else {
2707 #ifdef notyet
2708 		if (bp)
2709 			brelse(bp);
2710 #endif
2711 		FREE_LOCK(&lk);
2712 		error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
2713 		    NOCRED, &bp);
2714 		if (error) {
2715 			brelse(bp);
2716 			return (error);
2717 		}
2718 	}
2719 	/*
2720 	 * Recursively free indirect blocks.
2721 	 */
2722 	if (VFSTOUFS(freeblks->fb_mnt)->um_fstype == UFS1) {
2723 		ufs1fmt = 1;
2724 		bap1 = (ufs1_daddr_t *)bp->b_data;
2725 	} else {
2726 		ufs1fmt = 0;
2727 		bap2 = (ufs2_daddr_t *)bp->b_data;
2728 	}
2729 	nblocks = btodb(fs->fs_bsize);
2730 	for (i = NINDIR(fs) - 1; i >= 0; i--) {
2731 		if (ufs1fmt)
2732 			nb = bap1[i];
2733 		else
2734 			nb = bap2[i];
2735 		if (nb == 0)
2736 			continue;
2737 		if (level != 0) {
2738 			if ((error = indir_trunc(freeblks, fsbtodb(fs, nb),
2739 			     level - 1, lbn + (i * lbnadd), countp)) != 0)
2740 				allerror = error;
2741 		}
2742 		ffs_blkfree(ump, fs, freeblks->fb_devvp, nb, fs->fs_bsize,
2743 		    freeblks->fb_previousinum);
2744 		fs_pendingblocks += nblocks;
2745 		*countp += nblocks;
2746 	}
2747 	UFS_LOCK(ump);
2748 	fs->fs_pendingblocks -= fs_pendingblocks;
2749 	UFS_UNLOCK(ump);
2750 	bp->b_flags |= B_INVAL | B_NOCACHE;
2751 	brelse(bp);
2752 	return (allerror);
2753 }
2754 
2755 /*
2756  * Free an allocindir.
2757  * This routine must be called with splbio interrupts blocked.
2758  */
2759 static void
2760 free_allocindir(aip, inodedep)
2761 	struct allocindir *aip;
2762 	struct inodedep *inodedep;
2763 {
2764 	struct freefrag *freefrag;
2765 
2766 	mtx_assert(&lk, MA_OWNED);
2767 	if ((aip->ai_state & DEPCOMPLETE) == 0)
2768 		LIST_REMOVE(aip, ai_deps);
2769 	if (aip->ai_state & ONWORKLIST)
2770 		WORKLIST_REMOVE(&aip->ai_list);
2771 	LIST_REMOVE(aip, ai_next);
2772 	if ((freefrag = aip->ai_freefrag) != NULL) {
2773 		if (inodedep == NULL)
2774 			add_to_worklist(&freefrag->ff_list);
2775 		else
2776 			WORKLIST_INSERT(&inodedep->id_bufwait,
2777 			    &freefrag->ff_list);
2778 	}
2779 	WORKITEM_FREE(aip, D_ALLOCINDIR);
2780 }
2781 
2782 /*
2783  * Directory entry addition dependencies.
2784  *
2785  * When adding a new directory entry, the inode (with its incremented link
2786  * count) must be written to disk before the directory entry's pointer to it.
2787  * Also, if the inode is newly allocated, the corresponding freemap must be
2788  * updated (on disk) before the directory entry's pointer. These requirements
2789  * are met via undo/redo on the directory entry's pointer, which consists
2790  * simply of the inode number.
2791  *
2792  * As directory entries are added and deleted, the free space within a
2793  * directory block can become fragmented.  The ufs filesystem will compact
2794  * a fragmented directory block to make space for a new entry. When this
2795  * occurs, the offsets of previously added entries change. Any "diradd"
2796  * dependency structures corresponding to these entries must be updated with
2797  * the new offsets.
2798  */
2799 
2800 /*
2801  * This routine is called after the in-memory inode's link
2802  * count has been incremented, but before the directory entry's
2803  * pointer to the inode has been set.
2804  */
2805 int
2806 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
2807 	struct buf *bp;		/* buffer containing directory block */
2808 	struct inode *dp;	/* inode for directory */
2809 	off_t diroffset;	/* offset of new entry in directory */
2810 	ino_t newinum;		/* inode referenced by new directory entry */
2811 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
2812 	int isnewblk;		/* entry is in a newly allocated block */
2813 {
2814 	int offset;		/* offset of new entry within directory block */
2815 	ufs_lbn_t lbn;		/* block in directory containing new entry */
2816 	struct fs *fs;
2817 	struct diradd *dap;
2818 	struct allocdirect *adp;
2819 	struct pagedep *pagedep;
2820 	struct inodedep *inodedep;
2821 	struct newdirblk *newdirblk = 0;
2822 	struct mkdir *mkdir1, *mkdir2;
2823 
2824 	/*
2825 	 * Whiteouts have no dependencies.
2826 	 */
2827 	if (newinum == WINO) {
2828 		if (newdirbp != NULL)
2829 			bdwrite(newdirbp);
2830 		return (0);
2831 	}
2832 
2833 	fs = dp->i_fs;
2834 	lbn = lblkno(fs, diroffset);
2835 	offset = blkoff(fs, diroffset);
2836 	MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2837 		M_SOFTDEP_FLAGS|M_ZERO);
2838 	dap->da_list.wk_type = D_DIRADD;
2839 	dap->da_offset = offset;
2840 	dap->da_newinum = newinum;
2841 	dap->da_state = ATTACHED;
2842 	if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) {
2843 		MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk),
2844 		    M_NEWDIRBLK, M_SOFTDEP_FLAGS);
2845 		newdirblk->db_list.wk_type = D_NEWDIRBLK;
2846 		newdirblk->db_state = 0;
2847 	}
2848 	if (newdirbp == NULL) {
2849 		dap->da_state |= DEPCOMPLETE;
2850 		ACQUIRE_LOCK(&lk);
2851 	} else {
2852 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2853 		MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2854 		    M_SOFTDEP_FLAGS);
2855 		mkdir1->md_list.wk_type = D_MKDIR;
2856 		mkdir1->md_state = MKDIR_BODY;
2857 		mkdir1->md_diradd = dap;
2858 		MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2859 		    M_SOFTDEP_FLAGS);
2860 		mkdir2->md_list.wk_type = D_MKDIR;
2861 		mkdir2->md_state = MKDIR_PARENT;
2862 		mkdir2->md_diradd = dap;
2863 		/*
2864 		 * Dependency on "." and ".." being written to disk.
2865 		 */
2866 		mkdir1->md_buf = newdirbp;
2867 		ACQUIRE_LOCK(&lk);
2868 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2869 		WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2870 		FREE_LOCK(&lk);
2871 		bdwrite(newdirbp);
2872 		/*
2873 		 * Dependency on link count increase for parent directory
2874 		 */
2875 		ACQUIRE_LOCK(&lk);
2876 		if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0
2877 		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2878 			dap->da_state &= ~MKDIR_PARENT;
2879 			WORKITEM_FREE(mkdir2, D_MKDIR);
2880 		} else {
2881 			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2882 			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2883 		}
2884 	}
2885 	/*
2886 	 * Link into parent directory pagedep to await its being written.
2887 	 */
2888 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2889 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2890 	dap->da_pagedep = pagedep;
2891 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2892 	    da_pdlist);
2893 	/*
2894 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
2895 	 * is not yet written. If it is written, do the post-inode write
2896 	 * processing to put it on the id_pendinghd list.
2897 	 */
2898 	(void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2899 	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2900 		diradd_inode_written(dap, inodedep);
2901 	else
2902 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2903 	if (isnewblk) {
2904 		/*
2905 		 * Directories growing into indirect blocks are rare
2906 		 * enough and the frequency of new block allocation
2907 		 * in those cases even more rare, that we choose not
2908 		 * to bother tracking them. Rather we simply force the
2909 		 * new directory entry to disk.
2910 		 */
2911 		if (lbn >= NDADDR) {
2912 			FREE_LOCK(&lk);
2913 			/*
2914 			 * We only have a new allocation when at the
2915 			 * beginning of a new block, not when we are
2916 			 * expanding into an existing block.
2917 			 */
2918 			if (blkoff(fs, diroffset) == 0)
2919 				return (1);
2920 			return (0);
2921 		}
2922 		/*
2923 		 * We only have a new allocation when at the beginning
2924 		 * of a new fragment, not when we are expanding into an
2925 		 * existing fragment. Also, there is nothing to do if we
2926 		 * are already tracking this block.
2927 		 */
2928 		if (fragoff(fs, diroffset) != 0) {
2929 			FREE_LOCK(&lk);
2930 			return (0);
2931 		}
2932 		if ((pagedep->pd_state & NEWBLOCK) != 0) {
2933 			FREE_LOCK(&lk);
2934 			WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2935 			return (0);
2936 		}
2937 		/*
2938 		 * Find our associated allocdirect and have it track us.
2939 		 */
2940 		if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0)
2941 			panic("softdep_setup_directory_add: lost inodedep");
2942 		adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst);
2943 		if (adp == NULL || adp->ad_lbn != lbn)
2944 			panic("softdep_setup_directory_add: lost entry");
2945 		pagedep->pd_state |= NEWBLOCK;
2946 		newdirblk->db_pagedep = pagedep;
2947 		WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list);
2948 	}
2949 	FREE_LOCK(&lk);
2950 	return (0);
2951 }
2952 
2953 /*
2954  * This procedure is called to change the offset of a directory
2955  * entry when compacting a directory block which must be owned
2956  * exclusively by the caller. Note that the actual entry movement
2957  * must be done in this procedure to ensure that no I/O completions
2958  * occur while the move is in progress.
2959  */
2960 void
2961 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2962 	struct inode *dp;	/* inode for directory */
2963 	caddr_t base;		/* address of dp->i_offset */
2964 	caddr_t oldloc;		/* address of old directory location */
2965 	caddr_t newloc;		/* address of new directory location */
2966 	int entrysize;		/* size of directory entry */
2967 {
2968 	int offset, oldoffset, newoffset;
2969 	struct pagedep *pagedep;
2970 	struct diradd *dap;
2971 	ufs_lbn_t lbn;
2972 
2973 	ACQUIRE_LOCK(&lk);
2974 	lbn = lblkno(dp->i_fs, dp->i_offset);
2975 	offset = blkoff(dp->i_fs, dp->i_offset);
2976 	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2977 		goto done;
2978 	oldoffset = offset + (oldloc - base);
2979 	newoffset = offset + (newloc - base);
2980 
2981 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2982 		if (dap->da_offset != oldoffset)
2983 			continue;
2984 		dap->da_offset = newoffset;
2985 		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2986 			break;
2987 		LIST_REMOVE(dap, da_pdlist);
2988 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2989 		    dap, da_pdlist);
2990 		break;
2991 	}
2992 	if (dap == NULL) {
2993 
2994 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2995 			if (dap->da_offset == oldoffset) {
2996 				dap->da_offset = newoffset;
2997 				break;
2998 			}
2999 		}
3000 	}
3001 done:
3002 	bcopy(oldloc, newloc, entrysize);
3003 	FREE_LOCK(&lk);
3004 }
3005 
3006 /*
3007  * Free a diradd dependency structure. This routine must be called
3008  * with splbio interrupts blocked.
3009  */
3010 static void
3011 free_diradd(dap)
3012 	struct diradd *dap;
3013 {
3014 	struct dirrem *dirrem;
3015 	struct pagedep *pagedep;
3016 	struct inodedep *inodedep;
3017 	struct mkdir *mkdir, *nextmd;
3018 
3019 	mtx_assert(&lk, MA_OWNED);
3020 	WORKLIST_REMOVE(&dap->da_list);
3021 	LIST_REMOVE(dap, da_pdlist);
3022 	if ((dap->da_state & DIRCHG) == 0) {
3023 		pagedep = dap->da_pagedep;
3024 	} else {
3025 		dirrem = dap->da_previous;
3026 		pagedep = dirrem->dm_pagedep;
3027 		dirrem->dm_dirinum = pagedep->pd_ino;
3028 		add_to_worklist(&dirrem->dm_list);
3029 	}
3030 	if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
3031 	    0, &inodedep) != 0)
3032 		(void) free_inodedep(inodedep);
3033 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
3034 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
3035 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
3036 			if (mkdir->md_diradd != dap)
3037 				continue;
3038 			dap->da_state &= ~mkdir->md_state;
3039 			WORKLIST_REMOVE(&mkdir->md_list);
3040 			LIST_REMOVE(mkdir, md_mkdirs);
3041 			WORKITEM_FREE(mkdir, D_MKDIR);
3042 		}
3043 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0)
3044 			panic("free_diradd: unfound ref");
3045 	}
3046 	WORKITEM_FREE(dap, D_DIRADD);
3047 }
3048 
3049 /*
3050  * Directory entry removal dependencies.
3051  *
3052  * When removing a directory entry, the entry's inode pointer must be
3053  * zero'ed on disk before the corresponding inode's link count is decremented
3054  * (possibly freeing the inode for re-use). This dependency is handled by
3055  * updating the directory entry but delaying the inode count reduction until
3056  * after the directory block has been written to disk. After this point, the
3057  * inode count can be decremented whenever it is convenient.
3058  */
3059 
3060 /*
3061  * This routine should be called immediately after removing
3062  * a directory entry.  The inode's link count should not be
3063  * decremented by the calling procedure -- the soft updates
3064  * code will do this task when it is safe.
3065  */
3066 void
3067 softdep_setup_remove(bp, dp, ip, isrmdir)
3068 	struct buf *bp;		/* buffer containing directory block */
3069 	struct inode *dp;	/* inode for the directory being modified */
3070 	struct inode *ip;	/* inode for directory entry being removed */
3071 	int isrmdir;		/* indicates if doing RMDIR */
3072 {
3073 	struct dirrem *dirrem, *prevdirrem;
3074 
3075 	/*
3076 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
3077 	 */
3078 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3079 
3080 	/*
3081 	 * If the COMPLETE flag is clear, then there were no active
3082 	 * entries and we want to roll back to a zeroed entry until
3083 	 * the new inode is committed to disk. If the COMPLETE flag is
3084 	 * set then we have deleted an entry that never made it to
3085 	 * disk. If the entry we deleted resulted from a name change,
3086 	 * then the old name still resides on disk. We cannot delete
3087 	 * its inode (returned to us in prevdirrem) until the zeroed
3088 	 * directory entry gets to disk. The new inode has never been
3089 	 * referenced on the disk, so can be deleted immediately.
3090 	 */
3091 	if ((dirrem->dm_state & COMPLETE) == 0) {
3092 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
3093 		    dm_next);
3094 		FREE_LOCK(&lk);
3095 	} else {
3096 		if (prevdirrem != NULL)
3097 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
3098 			    prevdirrem, dm_next);
3099 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
3100 		FREE_LOCK(&lk);
3101 		handle_workitem_remove(dirrem, NULL);
3102 	}
3103 }
3104 
3105 /*
3106  * Allocate a new dirrem if appropriate and return it along with
3107  * its associated pagedep. Called without a lock, returns with lock.
3108  */
3109 static long num_dirrem;		/* number of dirrem allocated */
3110 static struct dirrem *
3111 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
3112 	struct buf *bp;		/* buffer containing directory block */
3113 	struct inode *dp;	/* inode for the directory being modified */
3114 	struct inode *ip;	/* inode for directory entry being removed */
3115 	int isrmdir;		/* indicates if doing RMDIR */
3116 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
3117 {
3118 	int offset;
3119 	ufs_lbn_t lbn;
3120 	struct diradd *dap;
3121 	struct dirrem *dirrem;
3122 	struct pagedep *pagedep;
3123 
3124 	/*
3125 	 * Whiteouts have no deletion dependencies.
3126 	 */
3127 	if (ip == NULL)
3128 		panic("newdirrem: whiteout");
3129 	/*
3130 	 * If we are over our limit, try to improve the situation.
3131 	 * Limiting the number of dirrem structures will also limit
3132 	 * the number of freefile and freeblks structures.
3133 	 */
3134 	ACQUIRE_LOCK(&lk);
3135 	if (num_dirrem > max_softdeps / 2)
3136 		(void) request_cleanup(FLUSH_REMOVE);
3137 	num_dirrem += 1;
3138 	FREE_LOCK(&lk);
3139 	MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
3140 		M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO);
3141 	dirrem->dm_list.wk_type = D_DIRREM;
3142 	dirrem->dm_state = isrmdir ? RMDIR : 0;
3143 	dirrem->dm_mnt = ITOV(ip)->v_mount;
3144 	dirrem->dm_oldinum = ip->i_number;
3145 	*prevdirremp = NULL;
3146 
3147 	ACQUIRE_LOCK(&lk);
3148 	lbn = lblkno(dp->i_fs, dp->i_offset);
3149 	offset = blkoff(dp->i_fs, dp->i_offset);
3150 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3151 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3152 	dirrem->dm_pagedep = pagedep;
3153 	/*
3154 	 * Check for a diradd dependency for the same directory entry.
3155 	 * If present, then both dependencies become obsolete and can
3156 	 * be de-allocated. Check for an entry on both the pd_dirraddhd
3157 	 * list and the pd_pendinghd list.
3158 	 */
3159 
3160 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
3161 		if (dap->da_offset == offset)
3162 			break;
3163 	if (dap == NULL) {
3164 
3165 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
3166 			if (dap->da_offset == offset)
3167 				break;
3168 		if (dap == NULL)
3169 			return (dirrem);
3170 	}
3171 	/*
3172 	 * Must be ATTACHED at this point.
3173 	 */
3174 	if ((dap->da_state & ATTACHED) == 0)
3175 		panic("newdirrem: not ATTACHED");
3176 	if (dap->da_newinum != ip->i_number)
3177 		panic("newdirrem: inum %d should be %d",
3178 		    ip->i_number, dap->da_newinum);
3179 	/*
3180 	 * If we are deleting a changed name that never made it to disk,
3181 	 * then return the dirrem describing the previous inode (which
3182 	 * represents the inode currently referenced from this entry on disk).
3183 	 */
3184 	if ((dap->da_state & DIRCHG) != 0) {
3185 		*prevdirremp = dap->da_previous;
3186 		dap->da_state &= ~DIRCHG;
3187 		dap->da_pagedep = pagedep;
3188 	}
3189 	/*
3190 	 * We are deleting an entry that never made it to disk.
3191 	 * Mark it COMPLETE so we can delete its inode immediately.
3192 	 */
3193 	dirrem->dm_state |= COMPLETE;
3194 	free_diradd(dap);
3195 	return (dirrem);
3196 }
3197 
3198 /*
3199  * Directory entry change dependencies.
3200  *
3201  * Changing an existing directory entry requires that an add operation
3202  * be completed first followed by a deletion. The semantics for the addition
3203  * are identical to the description of adding a new entry above except
3204  * that the rollback is to the old inode number rather than zero. Once
3205  * the addition dependency is completed, the removal is done as described
3206  * in the removal routine above.
3207  */
3208 
3209 /*
3210  * This routine should be called immediately after changing
3211  * a directory entry.  The inode's link count should not be
3212  * decremented by the calling procedure -- the soft updates
3213  * code will perform this task when it is safe.
3214  */
3215 void
3216 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
3217 	struct buf *bp;		/* buffer containing directory block */
3218 	struct inode *dp;	/* inode for the directory being modified */
3219 	struct inode *ip;	/* inode for directory entry being removed */
3220 	ino_t newinum;		/* new inode number for changed entry */
3221 	int isrmdir;		/* indicates if doing RMDIR */
3222 {
3223 	int offset;
3224 	struct diradd *dap = NULL;
3225 	struct dirrem *dirrem, *prevdirrem;
3226 	struct pagedep *pagedep;
3227 	struct inodedep *inodedep;
3228 
3229 	offset = blkoff(dp->i_fs, dp->i_offset);
3230 
3231 	/*
3232 	 * Whiteouts do not need diradd dependencies.
3233 	 */
3234 	if (newinum != WINO) {
3235 		MALLOC(dap, struct diradd *, sizeof(struct diradd),
3236 		    M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
3237 		dap->da_list.wk_type = D_DIRADD;
3238 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
3239 		dap->da_offset = offset;
3240 		dap->da_newinum = newinum;
3241 	}
3242 
3243 	/*
3244 	 * Allocate a new dirrem and ACQUIRE_LOCK.
3245 	 */
3246 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3247 	pagedep = dirrem->dm_pagedep;
3248 	/*
3249 	 * The possible values for isrmdir:
3250 	 *	0 - non-directory file rename
3251 	 *	1 - directory rename within same directory
3252 	 *   inum - directory rename to new directory of given inode number
3253 	 * When renaming to a new directory, we are both deleting and
3254 	 * creating a new directory entry, so the link count on the new
3255 	 * directory should not change. Thus we do not need the followup
3256 	 * dirrem which is usually done in handle_workitem_remove. We set
3257 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
3258 	 * followup dirrem.
3259 	 */
3260 	if (isrmdir > 1)
3261 		dirrem->dm_state |= DIRCHG;
3262 
3263 	/*
3264 	 * Whiteouts have no additional dependencies,
3265 	 * so just put the dirrem on the correct list.
3266 	 */
3267 	if (newinum == WINO) {
3268 		if ((dirrem->dm_state & COMPLETE) == 0) {
3269 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
3270 			    dm_next);
3271 		} else {
3272 			dirrem->dm_dirinum = pagedep->pd_ino;
3273 			add_to_worklist(&dirrem->dm_list);
3274 		}
3275 		FREE_LOCK(&lk);
3276 		return;
3277 	}
3278 
3279 	/*
3280 	 * If the COMPLETE flag is clear, then there were no active
3281 	 * entries and we want to roll back to the previous inode until
3282 	 * the new inode is committed to disk. If the COMPLETE flag is
3283 	 * set, then we have deleted an entry that never made it to disk.
3284 	 * If the entry we deleted resulted from a name change, then the old
3285 	 * inode reference still resides on disk. Any rollback that we do
3286 	 * needs to be to that old inode (returned to us in prevdirrem). If
3287 	 * the entry we deleted resulted from a create, then there is
3288 	 * no entry on the disk, so we want to roll back to zero rather
3289 	 * than the uncommitted inode. In either of the COMPLETE cases we
3290 	 * want to immediately free the unwritten and unreferenced inode.
3291 	 */
3292 	if ((dirrem->dm_state & COMPLETE) == 0) {
3293 		dap->da_previous = dirrem;
3294 	} else {
3295 		if (prevdirrem != NULL) {
3296 			dap->da_previous = prevdirrem;
3297 		} else {
3298 			dap->da_state &= ~DIRCHG;
3299 			dap->da_pagedep = pagedep;
3300 		}
3301 		dirrem->dm_dirinum = pagedep->pd_ino;
3302 		add_to_worklist(&dirrem->dm_list);
3303 	}
3304 	/*
3305 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3306 	 * is not yet written. If it is written, do the post-inode write
3307 	 * processing to put it on the id_pendinghd list.
3308 	 */
3309 	if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
3310 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3311 		dap->da_state |= COMPLETE;
3312 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3313 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3314 	} else {
3315 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
3316 		    dap, da_pdlist);
3317 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3318 	}
3319 	FREE_LOCK(&lk);
3320 }
3321 
3322 /*
3323  * Called whenever the link count on an inode is changed.
3324  * It creates an inode dependency so that the new reference(s)
3325  * to the inode cannot be committed to disk until the updated
3326  * inode has been written.
3327  */
3328 void
3329 softdep_change_linkcnt(ip)
3330 	struct inode *ip;	/* the inode with the increased link count */
3331 {
3332 	struct inodedep *inodedep;
3333 
3334 	ACQUIRE_LOCK(&lk);
3335 	(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
3336 	if (ip->i_nlink < ip->i_effnlink)
3337 		panic("softdep_change_linkcnt: bad delta");
3338 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3339 	FREE_LOCK(&lk);
3340 }
3341 
3342 /*
3343  * Called when the effective link count and the reference count
3344  * on an inode drops to zero. At this point there are no names
3345  * referencing the file in the filesystem and no active file
3346  * references. The space associated with the file will be freed
3347  * as soon as the necessary soft dependencies are cleared.
3348  */
3349 void
3350 softdep_releasefile(ip)
3351 	struct inode *ip;	/* inode with the zero effective link count */
3352 {
3353 	struct inodedep *inodedep;
3354 	struct fs *fs;
3355 	int extblocks;
3356 
3357 	if (ip->i_effnlink > 0)
3358 		panic("softdep_filerelease: file still referenced");
3359 	/*
3360 	 * We may be called several times as the real reference count
3361 	 * drops to zero. We only want to account for the space once.
3362 	 */
3363 	if (ip->i_flag & IN_SPACECOUNTED)
3364 		return;
3365 	/*
3366 	 * We have to deactivate a snapshot otherwise copyonwrites may
3367 	 * add blocks and the cleanup may remove blocks after we have
3368 	 * tried to account for them.
3369 	 */
3370 	if ((ip->i_flags & SF_SNAPSHOT) != 0)
3371 		ffs_snapremove(ITOV(ip));
3372 	/*
3373 	 * If we are tracking an nlinkdelta, we have to also remember
3374 	 * whether we accounted for the freed space yet.
3375 	 */
3376 	ACQUIRE_LOCK(&lk);
3377 	if ((inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep)))
3378 		inodedep->id_state |= SPACECOUNTED;
3379 	FREE_LOCK(&lk);
3380 	fs = ip->i_fs;
3381 	extblocks = 0;
3382 	if (fs->fs_magic == FS_UFS2_MAGIC)
3383 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
3384 	UFS_LOCK(ip->i_ump);
3385 	ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks;
3386 	ip->i_fs->fs_pendinginodes += 1;
3387 	UFS_UNLOCK(ip->i_ump);
3388 	ip->i_flag |= IN_SPACECOUNTED;
3389 }
3390 
3391 /*
3392  * This workitem decrements the inode's link count.
3393  * If the link count reaches zero, the file is removed.
3394  */
3395 static void
3396 handle_workitem_remove(dirrem, xp)
3397 	struct dirrem *dirrem;
3398 	struct vnode *xp;
3399 {
3400 	struct thread *td = curthread;
3401 	struct inodedep *inodedep;
3402 	struct vnode *vp;
3403 	struct inode *ip;
3404 	ino_t oldinum;
3405 	int error;
3406 
3407 	if ((vp = xp) == NULL &&
3408 	    (error = ffs_vget(dirrem->dm_mnt, dirrem->dm_oldinum, LK_EXCLUSIVE,
3409 	     &vp)) != 0) {
3410 		softdep_error("handle_workitem_remove: vget", error);
3411 		return;
3412 	}
3413 	ip = VTOI(vp);
3414 	ACQUIRE_LOCK(&lk);
3415 	if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0)
3416 		panic("handle_workitem_remove: lost inodedep");
3417 	/*
3418 	 * Normal file deletion.
3419 	 */
3420 	if ((dirrem->dm_state & RMDIR) == 0) {
3421 		ip->i_nlink--;
3422 		DIP_SET(ip, i_nlink, ip->i_nlink);
3423 		ip->i_flag |= IN_CHANGE;
3424 		if (ip->i_nlink < ip->i_effnlink)
3425 			panic("handle_workitem_remove: bad file delta");
3426 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3427 		num_dirrem -= 1;
3428 		FREE_LOCK(&lk);
3429 		vput(vp);
3430 		WORKITEM_FREE(dirrem, D_DIRREM);
3431 		return;
3432 	}
3433 	/*
3434 	 * Directory deletion. Decrement reference count for both the
3435 	 * just deleted parent directory entry and the reference for ".".
3436 	 * Next truncate the directory to length zero. When the
3437 	 * truncation completes, arrange to have the reference count on
3438 	 * the parent decremented to account for the loss of "..".
3439 	 */
3440 	ip->i_nlink -= 2;
3441 	DIP_SET(ip, i_nlink, ip->i_nlink);
3442 	ip->i_flag |= IN_CHANGE;
3443 	if (ip->i_nlink < ip->i_effnlink)
3444 		panic("handle_workitem_remove: bad dir delta");
3445 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3446 	FREE_LOCK(&lk);
3447 	if ((error = ffs_truncate(vp, (off_t)0, 0, td->td_ucred, td)) != 0)
3448 		softdep_error("handle_workitem_remove: truncate", error);
3449 	ACQUIRE_LOCK(&lk);
3450 	/*
3451 	 * Rename a directory to a new parent. Since, we are both deleting
3452 	 * and creating a new directory entry, the link count on the new
3453 	 * directory should not change. Thus we skip the followup dirrem.
3454 	 */
3455 	if (dirrem->dm_state & DIRCHG) {
3456 		num_dirrem -= 1;
3457 		FREE_LOCK(&lk);
3458 		vput(vp);
3459 		WORKITEM_FREE(dirrem, D_DIRREM);
3460 		return;
3461 	}
3462 	/*
3463 	 * If the inodedep does not exist, then the zero'ed inode has
3464 	 * been written to disk. If the allocated inode has never been
3465 	 * written to disk, then the on-disk inode is zero'ed. In either
3466 	 * case we can remove the file immediately.
3467 	 */
3468 	dirrem->dm_state = 0;
3469 	oldinum = dirrem->dm_oldinum;
3470 	dirrem->dm_oldinum = dirrem->dm_dirinum;
3471 	if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
3472 	    check_inode_unwritten(inodedep)) {
3473 		FREE_LOCK(&lk);
3474 		vput(vp);
3475 		handle_workitem_remove(dirrem, NULL);
3476 		return;
3477 	}
3478 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3479 	FREE_LOCK(&lk);
3480 	ip->i_flag |= IN_CHANGE;
3481 	ffs_update(vp, 0);
3482 	vput(vp);
3483 }
3484 
3485 /*
3486  * Inode de-allocation dependencies.
3487  *
3488  * When an inode's link count is reduced to zero, it can be de-allocated. We
3489  * found it convenient to postpone de-allocation until after the inode is
3490  * written to disk with its new link count (zero).  At this point, all of the
3491  * on-disk inode's block pointers are nullified and, with careful dependency
3492  * list ordering, all dependencies related to the inode will be satisfied and
3493  * the corresponding dependency structures de-allocated.  So, if/when the
3494  * inode is reused, there will be no mixing of old dependencies with new
3495  * ones.  This artificial dependency is set up by the block de-allocation
3496  * procedure above (softdep_setup_freeblocks) and completed by the
3497  * following procedure.
3498  */
3499 static void
3500 handle_workitem_freefile(freefile)
3501 	struct freefile *freefile;
3502 {
3503 	struct fs *fs;
3504 	struct inodedep *idp;
3505 	struct ufsmount *ump;
3506 	int error;
3507 
3508 	ump = VFSTOUFS(freefile->fx_mnt);
3509 	fs = ump->um_fs;
3510 #ifdef DEBUG
3511 	ACQUIRE_LOCK(&lk);
3512 	error = inodedep_lookup(fs, freefile->fx_oldinum, 0, &idp);
3513 	FREE_LOCK(&lk);
3514 	if (error)
3515 		panic("handle_workitem_freefile: inodedep survived");
3516 #endif
3517 	UFS_LOCK(ump);
3518 	fs->fs_pendinginodes -= 1;
3519 	UFS_UNLOCK(ump);
3520 	if ((error = ffs_freefile(VFSTOUFS(freefile->fx_mnt), fs,
3521 	    freefile->fx_devvp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
3522 		softdep_error("handle_workitem_freefile", error);
3523 	WORKITEM_FREE(freefile, D_FREEFILE);
3524 }
3525 
3526 
3527 /*
3528  * Helper function which unlinks marker element from work list and returns
3529  * the next element on the list.
3530  */
3531 static __inline struct worklist *
3532 markernext(struct worklist *marker)
3533 {
3534 	struct worklist *next;
3535 
3536 	next = LIST_NEXT(marker, wk_list);
3537 	LIST_REMOVE(marker, wk_list);
3538 	return next;
3539 }
3540 
3541 /*
3542  * Disk writes.
3543  *
3544  * The dependency structures constructed above are most actively used when file
3545  * system blocks are written to disk.  No constraints are placed on when a
3546  * block can be written, but unsatisfied update dependencies are made safe by
3547  * modifying (or replacing) the source memory for the duration of the disk
3548  * write.  When the disk write completes, the memory block is again brought
3549  * up-to-date.
3550  *
3551  * In-core inode structure reclamation.
3552  *
3553  * Because there are a finite number of "in-core" inode structures, they are
3554  * reused regularly.  By transferring all inode-related dependencies to the
3555  * in-memory inode block and indexing them separately (via "inodedep"s), we
3556  * can allow "in-core" inode structures to be reused at any time and avoid
3557  * any increase in contention.
3558  *
3559  * Called just before entering the device driver to initiate a new disk I/O.
3560  * The buffer must be locked, thus, no I/O completion operations can occur
3561  * while we are manipulating its associated dependencies.
3562  */
3563 static void
3564 softdep_disk_io_initiation(bp)
3565 	struct buf *bp;		/* structure describing disk write to occur */
3566 {
3567 	struct worklist *wk;
3568 	struct worklist marker;
3569 	struct indirdep *indirdep;
3570 	struct inodedep *inodedep;
3571 
3572 	/*
3573 	 * We only care about write operations. There should never
3574 	 * be dependencies for reads.
3575 	 */
3576 	if (bp->b_iocmd != BIO_WRITE)
3577 		panic("softdep_disk_io_initiation: not write");
3578 
3579 	marker.wk_type = D_LAST + 1;	/* Not a normal workitem */
3580 	PHOLD(curproc);			/* Don't swap out kernel stack */
3581 
3582 	ACQUIRE_LOCK(&lk);
3583 	/*
3584 	 * Do any necessary pre-I/O processing.
3585 	 */
3586 	for (wk = LIST_FIRST(&bp->b_dep); wk != NULL;
3587 	     wk = markernext(&marker)) {
3588 		LIST_INSERT_AFTER(wk, &marker, wk_list);
3589 		switch (wk->wk_type) {
3590 
3591 		case D_PAGEDEP:
3592 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
3593 			continue;
3594 
3595 		case D_INODEDEP:
3596 			inodedep = WK_INODEDEP(wk);
3597 			if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
3598 				initiate_write_inodeblock_ufs1(inodedep, bp);
3599 			else
3600 				initiate_write_inodeblock_ufs2(inodedep, bp);
3601 			continue;
3602 
3603 		case D_INDIRDEP:
3604 			indirdep = WK_INDIRDEP(wk);
3605 			if (indirdep->ir_state & GOINGAWAY)
3606 				panic("disk_io_initiation: indirdep gone");
3607 			/*
3608 			 * If there are no remaining dependencies, this
3609 			 * will be writing the real pointers, so the
3610 			 * dependency can be freed.
3611 			 */
3612 			if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
3613 				struct buf *bp;
3614 
3615 				bp = indirdep->ir_savebp;
3616 				bp->b_flags |= B_INVAL | B_NOCACHE;
3617 				/* inline expand WORKLIST_REMOVE(wk); */
3618 				wk->wk_state &= ~ONWORKLIST;
3619 				LIST_REMOVE(wk, wk_list);
3620 				WORKITEM_FREE(indirdep, D_INDIRDEP);
3621 				FREE_LOCK(&lk);
3622 				brelse(bp);
3623 				ACQUIRE_LOCK(&lk);
3624 				continue;
3625 			}
3626 			/*
3627 			 * Replace up-to-date version with safe version.
3628 			 */
3629 			FREE_LOCK(&lk);
3630 			MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3631 			    M_INDIRDEP, M_SOFTDEP_FLAGS);
3632 			ACQUIRE_LOCK(&lk);
3633 			indirdep->ir_state &= ~ATTACHED;
3634 			indirdep->ir_state |= UNDONE;
3635 			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3636 			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3637 			    bp->b_bcount);
3638 			continue;
3639 
3640 		case D_MKDIR:
3641 		case D_BMSAFEMAP:
3642 		case D_ALLOCDIRECT:
3643 		case D_ALLOCINDIR:
3644 			continue;
3645 
3646 		default:
3647 			panic("handle_disk_io_initiation: Unexpected type %s",
3648 			    TYPENAME(wk->wk_type));
3649 			/* NOTREACHED */
3650 		}
3651 	}
3652 	FREE_LOCK(&lk);
3653 	PRELE(curproc);			/* Allow swapout of kernel stack */
3654 }
3655 
3656 /*
3657  * Called from within the procedure above to deal with unsatisfied
3658  * allocation dependencies in a directory. The buffer must be locked,
3659  * thus, no I/O completion operations can occur while we are
3660  * manipulating its associated dependencies.
3661  */
3662 static void
3663 initiate_write_filepage(pagedep, bp)
3664 	struct pagedep *pagedep;
3665 	struct buf *bp;
3666 {
3667 	struct diradd *dap;
3668 	struct direct *ep;
3669 	int i;
3670 
3671 	if (pagedep->pd_state & IOSTARTED) {
3672 		/*
3673 		 * This can only happen if there is a driver that does not
3674 		 * understand chaining. Here biodone will reissue the call
3675 		 * to strategy for the incomplete buffers.
3676 		 */
3677 		printf("initiate_write_filepage: already started\n");
3678 		return;
3679 	}
3680 	pagedep->pd_state |= IOSTARTED;
3681 	for (i = 0; i < DAHASHSZ; i++) {
3682 		LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3683 			ep = (struct direct *)
3684 			    ((char *)bp->b_data + dap->da_offset);
3685 			if (ep->d_ino != dap->da_newinum)
3686 				panic("%s: dir inum %d != new %d",
3687 				    "initiate_write_filepage",
3688 				    ep->d_ino, dap->da_newinum);
3689 			if (dap->da_state & DIRCHG)
3690 				ep->d_ino = dap->da_previous->dm_oldinum;
3691 			else
3692 				ep->d_ino = 0;
3693 			dap->da_state &= ~ATTACHED;
3694 			dap->da_state |= UNDONE;
3695 		}
3696 	}
3697 }
3698 
3699 /*
3700  * Version of initiate_write_inodeblock that handles UFS1 dinodes.
3701  * Note that any bug fixes made to this routine must be done in the
3702  * version found below.
3703  *
3704  * Called from within the procedure above to deal with unsatisfied
3705  * allocation dependencies in an inodeblock. The buffer must be
3706  * locked, thus, no I/O completion operations can occur while we
3707  * are manipulating its associated dependencies.
3708  */
3709 static void
3710 initiate_write_inodeblock_ufs1(inodedep, bp)
3711 	struct inodedep *inodedep;
3712 	struct buf *bp;			/* The inode block */
3713 {
3714 	struct allocdirect *adp, *lastadp;
3715 	struct ufs1_dinode *dp;
3716 	struct ufs1_dinode *sip;
3717 	struct fs *fs;
3718 	ufs_lbn_t i, prevlbn = 0;
3719 	int deplist;
3720 
3721 	if (inodedep->id_state & IOSTARTED)
3722 		panic("initiate_write_inodeblock_ufs1: already started");
3723 	inodedep->id_state |= IOSTARTED;
3724 	fs = inodedep->id_fs;
3725 	dp = (struct ufs1_dinode *)bp->b_data +
3726 	    ino_to_fsbo(fs, inodedep->id_ino);
3727 	/*
3728 	 * If the bitmap is not yet written, then the allocated
3729 	 * inode cannot be written to disk.
3730 	 */
3731 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3732 		if (inodedep->id_savedino1 != NULL)
3733 			panic("initiate_write_inodeblock_ufs1: I/O underway");
3734 		FREE_LOCK(&lk);
3735 		MALLOC(sip, struct ufs1_dinode *,
3736 		    sizeof(struct ufs1_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS);
3737 		ACQUIRE_LOCK(&lk);
3738 		inodedep->id_savedino1 = sip;
3739 		*inodedep->id_savedino1 = *dp;
3740 		bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
3741 		dp->di_gen = inodedep->id_savedino1->di_gen;
3742 		return;
3743 	}
3744 	/*
3745 	 * If no dependencies, then there is nothing to roll back.
3746 	 */
3747 	inodedep->id_savedsize = dp->di_size;
3748 	inodedep->id_savedextsize = 0;
3749 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3750 		return;
3751 	/*
3752 	 * Set the dependencies to busy.
3753 	 */
3754 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3755 	     adp = TAILQ_NEXT(adp, ad_next)) {
3756 #ifdef DIAGNOSTIC
3757 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
3758 			panic("softdep_write_inodeblock: lbn order");
3759 		prevlbn = adp->ad_lbn;
3760 		if (adp->ad_lbn < NDADDR &&
3761 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno)
3762 			panic("%s: direct pointer #%jd mismatch %d != %jd",
3763 			    "softdep_write_inodeblock",
3764 			    (intmax_t)adp->ad_lbn,
3765 			    dp->di_db[adp->ad_lbn],
3766 			    (intmax_t)adp->ad_newblkno);
3767 		if (adp->ad_lbn >= NDADDR &&
3768 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno)
3769 			panic("%s: indirect pointer #%jd mismatch %d != %jd",
3770 			    "softdep_write_inodeblock",
3771 			    (intmax_t)adp->ad_lbn - NDADDR,
3772 			    dp->di_ib[adp->ad_lbn - NDADDR],
3773 			    (intmax_t)adp->ad_newblkno);
3774 		deplist |= 1 << adp->ad_lbn;
3775 		if ((adp->ad_state & ATTACHED) == 0)
3776 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3777 			    adp->ad_state);
3778 #endif /* DIAGNOSTIC */
3779 		adp->ad_state &= ~ATTACHED;
3780 		adp->ad_state |= UNDONE;
3781 	}
3782 	/*
3783 	 * The on-disk inode cannot claim to be any larger than the last
3784 	 * fragment that has been written. Otherwise, the on-disk inode
3785 	 * might have fragments that were not the last block in the file
3786 	 * which would corrupt the filesystem.
3787 	 */
3788 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3789 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3790 		if (adp->ad_lbn >= NDADDR)
3791 			break;
3792 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3793 		/* keep going until hitting a rollback to a frag */
3794 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3795 			continue;
3796 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3797 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3798 #ifdef DIAGNOSTIC
3799 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
3800 				panic("softdep_write_inodeblock: lost dep1");
3801 #endif /* DIAGNOSTIC */
3802 			dp->di_db[i] = 0;
3803 		}
3804 		for (i = 0; i < NIADDR; i++) {
3805 #ifdef DIAGNOSTIC
3806 			if (dp->di_ib[i] != 0 &&
3807 			    (deplist & ((1 << NDADDR) << i)) == 0)
3808 				panic("softdep_write_inodeblock: lost dep2");
3809 #endif /* DIAGNOSTIC */
3810 			dp->di_ib[i] = 0;
3811 		}
3812 		return;
3813 	}
3814 	/*
3815 	 * If we have zero'ed out the last allocated block of the file,
3816 	 * roll back the size to the last currently allocated block.
3817 	 * We know that this last allocated block is a full-sized as
3818 	 * we already checked for fragments in the loop above.
3819 	 */
3820 	if (lastadp != NULL &&
3821 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3822 		for (i = lastadp->ad_lbn; i >= 0; i--)
3823 			if (dp->di_db[i] != 0)
3824 				break;
3825 		dp->di_size = (i + 1) * fs->fs_bsize;
3826 	}
3827 	/*
3828 	 * The only dependencies are for indirect blocks.
3829 	 *
3830 	 * The file size for indirect block additions is not guaranteed.
3831 	 * Such a guarantee would be non-trivial to achieve. The conventional
3832 	 * synchronous write implementation also does not make this guarantee.
3833 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3834 	 * can be over-estimated without destroying integrity when the file
3835 	 * moves into the indirect blocks (i.e., is large). If we want to
3836 	 * postpone fsck, we are stuck with this argument.
3837 	 */
3838 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3839 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3840 }
3841 
3842 /*
3843  * Version of initiate_write_inodeblock that handles UFS2 dinodes.
3844  * Note that any bug fixes made to this routine must be done in the
3845  * version found above.
3846  *
3847  * Called from within the procedure above to deal with unsatisfied
3848  * allocation dependencies in an inodeblock. The buffer must be
3849  * locked, thus, no I/O completion operations can occur while we
3850  * are manipulating its associated dependencies.
3851  */
3852 static void
3853 initiate_write_inodeblock_ufs2(inodedep, bp)
3854 	struct inodedep *inodedep;
3855 	struct buf *bp;			/* The inode block */
3856 {
3857 	struct allocdirect *adp, *lastadp;
3858 	struct ufs2_dinode *dp;
3859 	struct ufs2_dinode *sip;
3860 	struct fs *fs;
3861 	ufs_lbn_t i, prevlbn = 0;
3862 	int deplist;
3863 
3864 	if (inodedep->id_state & IOSTARTED)
3865 		panic("initiate_write_inodeblock_ufs2: already started");
3866 	inodedep->id_state |= IOSTARTED;
3867 	fs = inodedep->id_fs;
3868 	dp = (struct ufs2_dinode *)bp->b_data +
3869 	    ino_to_fsbo(fs, inodedep->id_ino);
3870 	/*
3871 	 * If the bitmap is not yet written, then the allocated
3872 	 * inode cannot be written to disk.
3873 	 */
3874 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3875 		if (inodedep->id_savedino2 != NULL)
3876 			panic("initiate_write_inodeblock_ufs2: I/O underway");
3877 		FREE_LOCK(&lk);
3878 		MALLOC(sip, struct ufs2_dinode *,
3879 		    sizeof(struct ufs2_dinode), M_SAVEDINO, M_SOFTDEP_FLAGS);
3880 		ACQUIRE_LOCK(&lk);
3881 		inodedep->id_savedino2 = sip;
3882 		*inodedep->id_savedino2 = *dp;
3883 		bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
3884 		dp->di_gen = inodedep->id_savedino2->di_gen;
3885 		return;
3886 	}
3887 	/*
3888 	 * If no dependencies, then there is nothing to roll back.
3889 	 */
3890 	inodedep->id_savedsize = dp->di_size;
3891 	inodedep->id_savedextsize = dp->di_extsize;
3892 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL &&
3893 	    TAILQ_FIRST(&inodedep->id_extupdt) == NULL)
3894 		return;
3895 	/*
3896 	 * Set the ext data dependencies to busy.
3897 	 */
3898 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
3899 	     adp = TAILQ_NEXT(adp, ad_next)) {
3900 #ifdef DIAGNOSTIC
3901 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
3902 			panic("softdep_write_inodeblock: lbn order");
3903 		prevlbn = adp->ad_lbn;
3904 		if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno)
3905 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
3906 			    "softdep_write_inodeblock",
3907 			    (intmax_t)adp->ad_lbn,
3908 			    (intmax_t)dp->di_extb[adp->ad_lbn],
3909 			    (intmax_t)adp->ad_newblkno);
3910 		deplist |= 1 << adp->ad_lbn;
3911 		if ((adp->ad_state & ATTACHED) == 0)
3912 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3913 			    adp->ad_state);
3914 #endif /* DIAGNOSTIC */
3915 		adp->ad_state &= ~ATTACHED;
3916 		adp->ad_state |= UNDONE;
3917 	}
3918 	/*
3919 	 * The on-disk inode cannot claim to be any larger than the last
3920 	 * fragment that has been written. Otherwise, the on-disk inode
3921 	 * might have fragments that were not the last block in the ext
3922 	 * data which would corrupt the filesystem.
3923 	 */
3924 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
3925 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3926 		dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno;
3927 		/* keep going until hitting a rollback to a frag */
3928 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3929 			continue;
3930 		dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3931 		for (i = adp->ad_lbn + 1; i < NXADDR; i++) {
3932 #ifdef DIAGNOSTIC
3933 			if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0)
3934 				panic("softdep_write_inodeblock: lost dep1");
3935 #endif /* DIAGNOSTIC */
3936 			dp->di_extb[i] = 0;
3937 		}
3938 		lastadp = NULL;
3939 		break;
3940 	}
3941 	/*
3942 	 * If we have zero'ed out the last allocated block of the ext
3943 	 * data, roll back the size to the last currently allocated block.
3944 	 * We know that this last allocated block is a full-sized as
3945 	 * we already checked for fragments in the loop above.
3946 	 */
3947 	if (lastadp != NULL &&
3948 	    dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3949 		for (i = lastadp->ad_lbn; i >= 0; i--)
3950 			if (dp->di_extb[i] != 0)
3951 				break;
3952 		dp->di_extsize = (i + 1) * fs->fs_bsize;
3953 	}
3954 	/*
3955 	 * Set the file data dependencies to busy.
3956 	 */
3957 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3958 	     adp = TAILQ_NEXT(adp, ad_next)) {
3959 #ifdef DIAGNOSTIC
3960 		if (deplist != 0 && prevlbn >= adp->ad_lbn)
3961 			panic("softdep_write_inodeblock: lbn order");
3962 		prevlbn = adp->ad_lbn;
3963 		if (adp->ad_lbn < NDADDR &&
3964 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno)
3965 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
3966 			    "softdep_write_inodeblock",
3967 			    (intmax_t)adp->ad_lbn,
3968 			    (intmax_t)dp->di_db[adp->ad_lbn],
3969 			    (intmax_t)adp->ad_newblkno);
3970 		if (adp->ad_lbn >= NDADDR &&
3971 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno)
3972 			panic("%s indirect pointer #%jd mismatch %jd != %jd",
3973 			    "softdep_write_inodeblock:",
3974 			    (intmax_t)adp->ad_lbn - NDADDR,
3975 			    (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR],
3976 			    (intmax_t)adp->ad_newblkno);
3977 		deplist |= 1 << adp->ad_lbn;
3978 		if ((adp->ad_state & ATTACHED) == 0)
3979 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3980 			    adp->ad_state);
3981 #endif /* DIAGNOSTIC */
3982 		adp->ad_state &= ~ATTACHED;
3983 		adp->ad_state |= UNDONE;
3984 	}
3985 	/*
3986 	 * The on-disk inode cannot claim to be any larger than the last
3987 	 * fragment that has been written. Otherwise, the on-disk inode
3988 	 * might have fragments that were not the last block in the file
3989 	 * which would corrupt the filesystem.
3990 	 */
3991 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3992 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3993 		if (adp->ad_lbn >= NDADDR)
3994 			break;
3995 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3996 		/* keep going until hitting a rollback to a frag */
3997 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3998 			continue;
3999 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
4000 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
4001 #ifdef DIAGNOSTIC
4002 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0)
4003 				panic("softdep_write_inodeblock: lost dep2");
4004 #endif /* DIAGNOSTIC */
4005 			dp->di_db[i] = 0;
4006 		}
4007 		for (i = 0; i < NIADDR; i++) {
4008 #ifdef DIAGNOSTIC
4009 			if (dp->di_ib[i] != 0 &&
4010 			    (deplist & ((1 << NDADDR) << i)) == 0)
4011 				panic("softdep_write_inodeblock: lost dep3");
4012 #endif /* DIAGNOSTIC */
4013 			dp->di_ib[i] = 0;
4014 		}
4015 		return;
4016 	}
4017 	/*
4018 	 * If we have zero'ed out the last allocated block of the file,
4019 	 * roll back the size to the last currently allocated block.
4020 	 * We know that this last allocated block is a full-sized as
4021 	 * we already checked for fragments in the loop above.
4022 	 */
4023 	if (lastadp != NULL &&
4024 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
4025 		for (i = lastadp->ad_lbn; i >= 0; i--)
4026 			if (dp->di_db[i] != 0)
4027 				break;
4028 		dp->di_size = (i + 1) * fs->fs_bsize;
4029 	}
4030 	/*
4031 	 * The only dependencies are for indirect blocks.
4032 	 *
4033 	 * The file size for indirect block additions is not guaranteed.
4034 	 * Such a guarantee would be non-trivial to achieve. The conventional
4035 	 * synchronous write implementation also does not make this guarantee.
4036 	 * Fsck should catch and fix discrepancies. Arguably, the file size
4037 	 * can be over-estimated without destroying integrity when the file
4038 	 * moves into the indirect blocks (i.e., is large). If we want to
4039 	 * postpone fsck, we are stuck with this argument.
4040 	 */
4041 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
4042 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
4043 }
4044 
4045 /*
4046  * This routine is called during the completion interrupt
4047  * service routine for a disk write (from the procedure called
4048  * by the device driver to inform the filesystem caches of
4049  * a request completion).  It should be called early in this
4050  * procedure, before the block is made available to other
4051  * processes or other routines are called.
4052  */
4053 static void
4054 softdep_disk_write_complete(bp)
4055 	struct buf *bp;		/* describes the completed disk write */
4056 {
4057 	struct worklist *wk;
4058 	struct worklist *owk;
4059 	struct workhead reattach;
4060 	struct newblk *newblk;
4061 	struct allocindir *aip;
4062 	struct allocdirect *adp;
4063 	struct indirdep *indirdep;
4064 	struct inodedep *inodedep;
4065 	struct bmsafemap *bmsafemap;
4066 
4067 	/*
4068 	 * If an error occurred while doing the write, then the data
4069 	 * has not hit the disk and the dependencies cannot be unrolled.
4070 	 */
4071 	if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0)
4072 		return;
4073 	LIST_INIT(&reattach);
4074 	/*
4075 	 * This lock must not be released anywhere in this code segment.
4076 	 */
4077 	ACQUIRE_LOCK(&lk);
4078 	owk = NULL;
4079 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
4080 		WORKLIST_REMOVE(wk);
4081 		if (wk == owk)
4082 			panic("duplicate worklist: %p\n", wk);
4083 		owk = wk;
4084 		switch (wk->wk_type) {
4085 
4086 		case D_PAGEDEP:
4087 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
4088 				WORKLIST_INSERT(&reattach, wk);
4089 			continue;
4090 
4091 		case D_INODEDEP:
4092 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
4093 				WORKLIST_INSERT(&reattach, wk);
4094 			continue;
4095 
4096 		case D_BMSAFEMAP:
4097 			bmsafemap = WK_BMSAFEMAP(wk);
4098 			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
4099 				newblk->nb_state |= DEPCOMPLETE;
4100 				newblk->nb_bmsafemap = NULL;
4101 				LIST_REMOVE(newblk, nb_deps);
4102 			}
4103 			while ((adp =
4104 			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
4105 				adp->ad_state |= DEPCOMPLETE;
4106 				adp->ad_buf = NULL;
4107 				LIST_REMOVE(adp, ad_deps);
4108 				handle_allocdirect_partdone(adp);
4109 			}
4110 			while ((aip =
4111 			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
4112 				aip->ai_state |= DEPCOMPLETE;
4113 				aip->ai_buf = NULL;
4114 				LIST_REMOVE(aip, ai_deps);
4115 				handle_allocindir_partdone(aip);
4116 			}
4117 			while ((inodedep =
4118 			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
4119 				inodedep->id_state |= DEPCOMPLETE;
4120 				LIST_REMOVE(inodedep, id_deps);
4121 				inodedep->id_buf = NULL;
4122 			}
4123 			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
4124 			continue;
4125 
4126 		case D_MKDIR:
4127 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
4128 			continue;
4129 
4130 		case D_ALLOCDIRECT:
4131 			adp = WK_ALLOCDIRECT(wk);
4132 			adp->ad_state |= COMPLETE;
4133 			handle_allocdirect_partdone(adp);
4134 			continue;
4135 
4136 		case D_ALLOCINDIR:
4137 			aip = WK_ALLOCINDIR(wk);
4138 			aip->ai_state |= COMPLETE;
4139 			handle_allocindir_partdone(aip);
4140 			continue;
4141 
4142 		case D_INDIRDEP:
4143 			indirdep = WK_INDIRDEP(wk);
4144 			if (indirdep->ir_state & GOINGAWAY)
4145 				panic("disk_write_complete: indirdep gone");
4146 			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
4147 			FREE(indirdep->ir_saveddata, M_INDIRDEP);
4148 			indirdep->ir_saveddata = 0;
4149 			indirdep->ir_state &= ~UNDONE;
4150 			indirdep->ir_state |= ATTACHED;
4151 			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
4152 				handle_allocindir_partdone(aip);
4153 				if (aip == LIST_FIRST(&indirdep->ir_donehd))
4154 					panic("disk_write_complete: not gone");
4155 			}
4156 			WORKLIST_INSERT(&reattach, wk);
4157 			if ((bp->b_flags & B_DELWRI) == 0)
4158 				stat_indir_blk_ptrs++;
4159 			bdirty(bp);
4160 			continue;
4161 
4162 		default:
4163 			panic("handle_disk_write_complete: Unknown type %s",
4164 			    TYPENAME(wk->wk_type));
4165 			/* NOTREACHED */
4166 		}
4167 	}
4168 	/*
4169 	 * Reattach any requests that must be redone.
4170 	 */
4171 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
4172 		WORKLIST_REMOVE(wk);
4173 		WORKLIST_INSERT(&bp->b_dep, wk);
4174 	}
4175 	FREE_LOCK(&lk);
4176 }
4177 
4178 /*
4179  * Called from within softdep_disk_write_complete above. Note that
4180  * this routine is always called from interrupt level with further
4181  * splbio interrupts blocked.
4182  */
4183 static void
4184 handle_allocdirect_partdone(adp)
4185 	struct allocdirect *adp;	/* the completed allocdirect */
4186 {
4187 	struct allocdirectlst *listhead;
4188 	struct allocdirect *listadp;
4189 	struct inodedep *inodedep;
4190 	long bsize, delay;
4191 
4192 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4193 		return;
4194 	if (adp->ad_buf != NULL)
4195 		panic("handle_allocdirect_partdone: dangling dep");
4196 	/*
4197 	 * The on-disk inode cannot claim to be any larger than the last
4198 	 * fragment that has been written. Otherwise, the on-disk inode
4199 	 * might have fragments that were not the last block in the file
4200 	 * which would corrupt the filesystem. Thus, we cannot free any
4201 	 * allocdirects after one whose ad_oldblkno claims a fragment as
4202 	 * these blocks must be rolled back to zero before writing the inode.
4203 	 * We check the currently active set of allocdirects in id_inoupdt
4204 	 * or id_extupdt as appropriate.
4205 	 */
4206 	inodedep = adp->ad_inodedep;
4207 	bsize = inodedep->id_fs->fs_bsize;
4208 	if (adp->ad_state & EXTDATA)
4209 		listhead = &inodedep->id_extupdt;
4210 	else
4211 		listhead = &inodedep->id_inoupdt;
4212 	TAILQ_FOREACH(listadp, listhead, ad_next) {
4213 		/* found our block */
4214 		if (listadp == adp)
4215 			break;
4216 		/* continue if ad_oldlbn is not a fragment */
4217 		if (listadp->ad_oldsize == 0 ||
4218 		    listadp->ad_oldsize == bsize)
4219 			continue;
4220 		/* hit a fragment */
4221 		return;
4222 	}
4223 	/*
4224 	 * If we have reached the end of the current list without
4225 	 * finding the just finished dependency, then it must be
4226 	 * on the future dependency list. Future dependencies cannot
4227 	 * be freed until they are moved to the current list.
4228 	 */
4229 	if (listadp == NULL) {
4230 #ifdef DEBUG
4231 		if (adp->ad_state & EXTDATA)
4232 			listhead = &inodedep->id_newextupdt;
4233 		else
4234 			listhead = &inodedep->id_newinoupdt;
4235 		TAILQ_FOREACH(listadp, listhead, ad_next)
4236 			/* found our block */
4237 			if (listadp == adp)
4238 				break;
4239 		if (listadp == NULL)
4240 			panic("handle_allocdirect_partdone: lost dep");
4241 #endif /* DEBUG */
4242 		return;
4243 	}
4244 	/*
4245 	 * If we have found the just finished dependency, then free
4246 	 * it along with anything that follows it that is complete.
4247 	 * If the inode still has a bitmap dependency, then it has
4248 	 * never been written to disk, hence the on-disk inode cannot
4249 	 * reference the old fragment so we can free it without delay.
4250 	 */
4251 	delay = (inodedep->id_state & DEPCOMPLETE);
4252 	for (; adp; adp = listadp) {
4253 		listadp = TAILQ_NEXT(adp, ad_next);
4254 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4255 			return;
4256 		free_allocdirect(listhead, adp, delay);
4257 	}
4258 }
4259 
4260 /*
4261  * Called from within softdep_disk_write_complete above. Note that
4262  * this routine is always called from interrupt level with further
4263  * splbio interrupts blocked.
4264  */
4265 static void
4266 handle_allocindir_partdone(aip)
4267 	struct allocindir *aip;		/* the completed allocindir */
4268 {
4269 	struct indirdep *indirdep;
4270 
4271 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
4272 		return;
4273 	if (aip->ai_buf != NULL)
4274 		panic("handle_allocindir_partdone: dangling dependency");
4275 	indirdep = aip->ai_indirdep;
4276 	if (indirdep->ir_state & UNDONE) {
4277 		LIST_REMOVE(aip, ai_next);
4278 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
4279 		return;
4280 	}
4281 	if (indirdep->ir_state & UFS1FMT)
4282 		((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4283 		    aip->ai_newblkno;
4284 	else
4285 		((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4286 		    aip->ai_newblkno;
4287 	LIST_REMOVE(aip, ai_next);
4288 	if (aip->ai_freefrag != NULL)
4289 		add_to_worklist(&aip->ai_freefrag->ff_list);
4290 	WORKITEM_FREE(aip, D_ALLOCINDIR);
4291 }
4292 
4293 /*
4294  * Called from within softdep_disk_write_complete above to restore
4295  * in-memory inode block contents to their most up-to-date state. Note
4296  * that this routine is always called from interrupt level with further
4297  * splbio interrupts blocked.
4298  */
4299 static int
4300 handle_written_inodeblock(inodedep, bp)
4301 	struct inodedep *inodedep;
4302 	struct buf *bp;		/* buffer containing the inode block */
4303 {
4304 	struct worklist *wk, *filefree;
4305 	struct allocdirect *adp, *nextadp;
4306 	struct ufs1_dinode *dp1 = NULL;
4307 	struct ufs2_dinode *dp2 = NULL;
4308 	int hadchanges, fstype;
4309 
4310 	if ((inodedep->id_state & IOSTARTED) == 0)
4311 		panic("handle_written_inodeblock: not started");
4312 	inodedep->id_state &= ~IOSTARTED;
4313 	if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
4314 		fstype = UFS1;
4315 		dp1 = (struct ufs1_dinode *)bp->b_data +
4316 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4317 	} else {
4318 		fstype = UFS2;
4319 		dp2 = (struct ufs2_dinode *)bp->b_data +
4320 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4321 	}
4322 	/*
4323 	 * If we had to rollback the inode allocation because of
4324 	 * bitmaps being incomplete, then simply restore it.
4325 	 * Keep the block dirty so that it will not be reclaimed until
4326 	 * all associated dependencies have been cleared and the
4327 	 * corresponding updates written to disk.
4328 	 */
4329 	if (inodedep->id_savedino1 != NULL) {
4330 		if (fstype == UFS1)
4331 			*dp1 = *inodedep->id_savedino1;
4332 		else
4333 			*dp2 = *inodedep->id_savedino2;
4334 		FREE(inodedep->id_savedino1, M_SAVEDINO);
4335 		inodedep->id_savedino1 = NULL;
4336 		if ((bp->b_flags & B_DELWRI) == 0)
4337 			stat_inode_bitmap++;
4338 		bdirty(bp);
4339 		return (1);
4340 	}
4341 	inodedep->id_state |= COMPLETE;
4342 	/*
4343 	 * Roll forward anything that had to be rolled back before
4344 	 * the inode could be updated.
4345 	 */
4346 	hadchanges = 0;
4347 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
4348 		nextadp = TAILQ_NEXT(adp, ad_next);
4349 		if (adp->ad_state & ATTACHED)
4350 			panic("handle_written_inodeblock: new entry");
4351 		if (fstype == UFS1) {
4352 			if (adp->ad_lbn < NDADDR) {
4353 				if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno)
4354 					panic("%s %s #%jd mismatch %d != %jd",
4355 					    "handle_written_inodeblock:",
4356 					    "direct pointer",
4357 					    (intmax_t)adp->ad_lbn,
4358 					    dp1->di_db[adp->ad_lbn],
4359 					    (intmax_t)adp->ad_oldblkno);
4360 				dp1->di_db[adp->ad_lbn] = adp->ad_newblkno;
4361 			} else {
4362 				if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0)
4363 					panic("%s: %s #%jd allocated as %d",
4364 					    "handle_written_inodeblock",
4365 					    "indirect pointer",
4366 					    (intmax_t)adp->ad_lbn - NDADDR,
4367 					    dp1->di_ib[adp->ad_lbn - NDADDR]);
4368 				dp1->di_ib[adp->ad_lbn - NDADDR] =
4369 				    adp->ad_newblkno;
4370 			}
4371 		} else {
4372 			if (adp->ad_lbn < NDADDR) {
4373 				if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno)
4374 					panic("%s: %s #%jd %s %jd != %jd",
4375 					    "handle_written_inodeblock",
4376 					    "direct pointer",
4377 					    (intmax_t)adp->ad_lbn, "mismatch",
4378 					    (intmax_t)dp2->di_db[adp->ad_lbn],
4379 					    (intmax_t)adp->ad_oldblkno);
4380 				dp2->di_db[adp->ad_lbn] = adp->ad_newblkno;
4381 			} else {
4382 				if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0)
4383 					panic("%s: %s #%jd allocated as %jd",
4384 					    "handle_written_inodeblock",
4385 					    "indirect pointer",
4386 					    (intmax_t)adp->ad_lbn - NDADDR,
4387 					    (intmax_t)
4388 					    dp2->di_ib[adp->ad_lbn - NDADDR]);
4389 				dp2->di_ib[adp->ad_lbn - NDADDR] =
4390 				    adp->ad_newblkno;
4391 			}
4392 		}
4393 		adp->ad_state &= ~UNDONE;
4394 		adp->ad_state |= ATTACHED;
4395 		hadchanges = 1;
4396 	}
4397 	for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
4398 		nextadp = TAILQ_NEXT(adp, ad_next);
4399 		if (adp->ad_state & ATTACHED)
4400 			panic("handle_written_inodeblock: new entry");
4401 		if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno)
4402 			panic("%s: direct pointers #%jd %s %jd != %jd",
4403 			    "handle_written_inodeblock",
4404 			    (intmax_t)adp->ad_lbn, "mismatch",
4405 			    (intmax_t)dp2->di_extb[adp->ad_lbn],
4406 			    (intmax_t)adp->ad_oldblkno);
4407 		dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno;
4408 		adp->ad_state &= ~UNDONE;
4409 		adp->ad_state |= ATTACHED;
4410 		hadchanges = 1;
4411 	}
4412 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
4413 		stat_direct_blk_ptrs++;
4414 	/*
4415 	 * Reset the file size to its most up-to-date value.
4416 	 */
4417 	if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1)
4418 		panic("handle_written_inodeblock: bad size");
4419 	if (fstype == UFS1) {
4420 		if (dp1->di_size != inodedep->id_savedsize) {
4421 			dp1->di_size = inodedep->id_savedsize;
4422 			hadchanges = 1;
4423 		}
4424 	} else {
4425 		if (dp2->di_size != inodedep->id_savedsize) {
4426 			dp2->di_size = inodedep->id_savedsize;
4427 			hadchanges = 1;
4428 		}
4429 		if (dp2->di_extsize != inodedep->id_savedextsize) {
4430 			dp2->di_extsize = inodedep->id_savedextsize;
4431 			hadchanges = 1;
4432 		}
4433 	}
4434 	inodedep->id_savedsize = -1;
4435 	inodedep->id_savedextsize = -1;
4436 	/*
4437 	 * If there were any rollbacks in the inode block, then it must be
4438 	 * marked dirty so that its will eventually get written back in
4439 	 * its correct form.
4440 	 */
4441 	if (hadchanges)
4442 		bdirty(bp);
4443 	/*
4444 	 * Process any allocdirects that completed during the update.
4445 	 */
4446 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
4447 		handle_allocdirect_partdone(adp);
4448 	if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
4449 		handle_allocdirect_partdone(adp);
4450 	/*
4451 	 * Process deallocations that were held pending until the
4452 	 * inode had been written to disk. Freeing of the inode
4453 	 * is delayed until after all blocks have been freed to
4454 	 * avoid creation of new <vfsid, inum, lbn> triples
4455 	 * before the old ones have been deleted.
4456 	 */
4457 	filefree = NULL;
4458 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
4459 		WORKLIST_REMOVE(wk);
4460 		switch (wk->wk_type) {
4461 
4462 		case D_FREEFILE:
4463 			/*
4464 			 * We defer adding filefree to the worklist until
4465 			 * all other additions have been made to ensure
4466 			 * that it will be done after all the old blocks
4467 			 * have been freed.
4468 			 */
4469 			if (filefree != NULL)
4470 				panic("handle_written_inodeblock: filefree");
4471 			filefree = wk;
4472 			continue;
4473 
4474 		case D_MKDIR:
4475 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
4476 			continue;
4477 
4478 		case D_DIRADD:
4479 			diradd_inode_written(WK_DIRADD(wk), inodedep);
4480 			continue;
4481 
4482 		case D_FREEBLKS:
4483 			wk->wk_state |= COMPLETE;
4484 			if ((wk->wk_state  & ALLCOMPLETE) != ALLCOMPLETE)
4485 				continue;
4486 			 /* -- fall through -- */
4487 		case D_FREEFRAG:
4488 		case D_DIRREM:
4489 			add_to_worklist(wk);
4490 			continue;
4491 
4492 		case D_NEWDIRBLK:
4493 			free_newdirblk(WK_NEWDIRBLK(wk));
4494 			continue;
4495 
4496 		default:
4497 			panic("handle_written_inodeblock: Unknown type %s",
4498 			    TYPENAME(wk->wk_type));
4499 			/* NOTREACHED */
4500 		}
4501 	}
4502 	if (filefree != NULL) {
4503 		if (free_inodedep(inodedep) == 0)
4504 			panic("handle_written_inodeblock: live inodedep");
4505 		add_to_worklist(filefree);
4506 		return (0);
4507 	}
4508 
4509 	/*
4510 	 * If no outstanding dependencies, free it.
4511 	 */
4512 	if (free_inodedep(inodedep) ||
4513 	    (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
4514 	     TAILQ_FIRST(&inodedep->id_extupdt) == 0))
4515 		return (0);
4516 	return (hadchanges);
4517 }
4518 
4519 /*
4520  * Process a diradd entry after its dependent inode has been written.
4521  * This routine must be called with splbio interrupts blocked.
4522  */
4523 static void
4524 diradd_inode_written(dap, inodedep)
4525 	struct diradd *dap;
4526 	struct inodedep *inodedep;
4527 {
4528 	struct pagedep *pagedep;
4529 
4530 	dap->da_state |= COMPLETE;
4531 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4532 		if (dap->da_state & DIRCHG)
4533 			pagedep = dap->da_previous->dm_pagedep;
4534 		else
4535 			pagedep = dap->da_pagedep;
4536 		LIST_REMOVE(dap, da_pdlist);
4537 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4538 	}
4539 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
4540 }
4541 
4542 /*
4543  * Handle the completion of a mkdir dependency.
4544  */
4545 static void
4546 handle_written_mkdir(mkdir, type)
4547 	struct mkdir *mkdir;
4548 	int type;
4549 {
4550 	struct diradd *dap;
4551 	struct pagedep *pagedep;
4552 
4553 	if (mkdir->md_state != type)
4554 		panic("handle_written_mkdir: bad type");
4555 	dap = mkdir->md_diradd;
4556 	dap->da_state &= ~type;
4557 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
4558 		dap->da_state |= DEPCOMPLETE;
4559 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4560 		if (dap->da_state & DIRCHG)
4561 			pagedep = dap->da_previous->dm_pagedep;
4562 		else
4563 			pagedep = dap->da_pagedep;
4564 		LIST_REMOVE(dap, da_pdlist);
4565 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4566 	}
4567 	LIST_REMOVE(mkdir, md_mkdirs);
4568 	WORKITEM_FREE(mkdir, D_MKDIR);
4569 }
4570 
4571 /*
4572  * Called from within softdep_disk_write_complete above.
4573  * A write operation was just completed. Removed inodes can
4574  * now be freed and associated block pointers may be committed.
4575  * Note that this routine is always called from interrupt level
4576  * with further splbio interrupts blocked.
4577  */
4578 static int
4579 handle_written_filepage(pagedep, bp)
4580 	struct pagedep *pagedep;
4581 	struct buf *bp;		/* buffer containing the written page */
4582 {
4583 	struct dirrem *dirrem;
4584 	struct diradd *dap, *nextdap;
4585 	struct direct *ep;
4586 	int i, chgs;
4587 
4588 	if ((pagedep->pd_state & IOSTARTED) == 0)
4589 		panic("handle_written_filepage: not started");
4590 	pagedep->pd_state &= ~IOSTARTED;
4591 	/*
4592 	 * Process any directory removals that have been committed.
4593 	 */
4594 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
4595 		LIST_REMOVE(dirrem, dm_next);
4596 		dirrem->dm_dirinum = pagedep->pd_ino;
4597 		add_to_worklist(&dirrem->dm_list);
4598 	}
4599 	/*
4600 	 * Free any directory additions that have been committed.
4601 	 * If it is a newly allocated block, we have to wait until
4602 	 * the on-disk directory inode claims the new block.
4603 	 */
4604 	if ((pagedep->pd_state & NEWBLOCK) == 0)
4605 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
4606 			free_diradd(dap);
4607 	/*
4608 	 * Uncommitted directory entries must be restored.
4609 	 */
4610 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
4611 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
4612 		     dap = nextdap) {
4613 			nextdap = LIST_NEXT(dap, da_pdlist);
4614 			if (dap->da_state & ATTACHED)
4615 				panic("handle_written_filepage: attached");
4616 			ep = (struct direct *)
4617 			    ((char *)bp->b_data + dap->da_offset);
4618 			ep->d_ino = dap->da_newinum;
4619 			dap->da_state &= ~UNDONE;
4620 			dap->da_state |= ATTACHED;
4621 			chgs = 1;
4622 			/*
4623 			 * If the inode referenced by the directory has
4624 			 * been written out, then the dependency can be
4625 			 * moved to the pending list.
4626 			 */
4627 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4628 				LIST_REMOVE(dap, da_pdlist);
4629 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
4630 				    da_pdlist);
4631 			}
4632 		}
4633 	}
4634 	/*
4635 	 * If there were any rollbacks in the directory, then it must be
4636 	 * marked dirty so that its will eventually get written back in
4637 	 * its correct form.
4638 	 */
4639 	if (chgs) {
4640 		if ((bp->b_flags & B_DELWRI) == 0)
4641 			stat_dir_entry++;
4642 		bdirty(bp);
4643 		return (1);
4644 	}
4645 	/*
4646 	 * If we are not waiting for a new directory block to be
4647 	 * claimed by its inode, then the pagedep will be freed.
4648 	 * Otherwise it will remain to track any new entries on
4649 	 * the page in case they are fsync'ed.
4650 	 */
4651 	if ((pagedep->pd_state & NEWBLOCK) == 0) {
4652 		LIST_REMOVE(pagedep, pd_hash);
4653 		WORKITEM_FREE(pagedep, D_PAGEDEP);
4654 	}
4655 	return (0);
4656 }
4657 
4658 /*
4659  * Writing back in-core inode structures.
4660  *
4661  * The filesystem only accesses an inode's contents when it occupies an
4662  * "in-core" inode structure.  These "in-core" structures are separate from
4663  * the page frames used to cache inode blocks.  Only the latter are
4664  * transferred to/from the disk.  So, when the updated contents of the
4665  * "in-core" inode structure are copied to the corresponding in-memory inode
4666  * block, the dependencies are also transferred.  The following procedure is
4667  * called when copying a dirty "in-core" inode to a cached inode block.
4668  */
4669 
4670 /*
4671  * Called when an inode is loaded from disk. If the effective link count
4672  * differed from the actual link count when it was last flushed, then we
4673  * need to ensure that the correct effective link count is put back.
4674  */
4675 void
4676 softdep_load_inodeblock(ip)
4677 	struct inode *ip;	/* the "in_core" copy of the inode */
4678 {
4679 	struct inodedep *inodedep;
4680 
4681 	/*
4682 	 * Check for alternate nlink count.
4683 	 */
4684 	ip->i_effnlink = ip->i_nlink;
4685 	ACQUIRE_LOCK(&lk);
4686 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
4687 		FREE_LOCK(&lk);
4688 		return;
4689 	}
4690 	ip->i_effnlink -= inodedep->id_nlinkdelta;
4691 	if (inodedep->id_state & SPACECOUNTED)
4692 		ip->i_flag |= IN_SPACECOUNTED;
4693 	FREE_LOCK(&lk);
4694 }
4695 
4696 /*
4697  * This routine is called just before the "in-core" inode
4698  * information is to be copied to the in-memory inode block.
4699  * Recall that an inode block contains several inodes. If
4700  * the force flag is set, then the dependencies will be
4701  * cleared so that the update can always be made. Note that
4702  * the buffer is locked when this routine is called, so we
4703  * will never be in the middle of writing the inode block
4704  * to disk.
4705  */
4706 void
4707 softdep_update_inodeblock(ip, bp, waitfor)
4708 	struct inode *ip;	/* the "in_core" copy of the inode */
4709 	struct buf *bp;		/* the buffer containing the inode block */
4710 	int waitfor;		/* nonzero => update must be allowed */
4711 {
4712 	struct inodedep *inodedep;
4713 	struct worklist *wk;
4714 	struct buf *ibp;
4715 	int error;
4716 
4717 	/*
4718 	 * If the effective link count is not equal to the actual link
4719 	 * count, then we must track the difference in an inodedep while
4720 	 * the inode is (potentially) tossed out of the cache. Otherwise,
4721 	 * if there is no existing inodedep, then there are no dependencies
4722 	 * to track.
4723 	 */
4724 	ACQUIRE_LOCK(&lk);
4725 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
4726 		FREE_LOCK(&lk);
4727 		if (ip->i_effnlink != ip->i_nlink)
4728 			panic("softdep_update_inodeblock: bad link count");
4729 		return;
4730 	}
4731 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink)
4732 		panic("softdep_update_inodeblock: bad delta");
4733 	/*
4734 	 * Changes have been initiated. Anything depending on these
4735 	 * changes cannot occur until this inode has been written.
4736 	 */
4737 	inodedep->id_state &= ~COMPLETE;
4738 	if ((inodedep->id_state & ONWORKLIST) == 0)
4739 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
4740 	/*
4741 	 * Any new dependencies associated with the incore inode must
4742 	 * now be moved to the list associated with the buffer holding
4743 	 * the in-memory copy of the inode. Once merged process any
4744 	 * allocdirects that are completed by the merger.
4745 	 */
4746 	merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
4747 	if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
4748 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
4749 	merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
4750 	if (TAILQ_FIRST(&inodedep->id_extupdt) != NULL)
4751 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt));
4752 	/*
4753 	 * Now that the inode has been pushed into the buffer, the
4754 	 * operations dependent on the inode being written to disk
4755 	 * can be moved to the id_bufwait so that they will be
4756 	 * processed when the buffer I/O completes.
4757 	 */
4758 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
4759 		WORKLIST_REMOVE(wk);
4760 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
4761 	}
4762 	/*
4763 	 * Newly allocated inodes cannot be written until the bitmap
4764 	 * that allocates them have been written (indicated by
4765 	 * DEPCOMPLETE being set in id_state). If we are doing a
4766 	 * forced sync (e.g., an fsync on a file), we force the bitmap
4767 	 * to be written so that the update can be done.
4768 	 */
4769 	if (waitfor == 0) {
4770 		FREE_LOCK(&lk);
4771 		return;
4772 	}
4773 retry:
4774 	if ((inodedep->id_state & DEPCOMPLETE) != 0) {
4775 		FREE_LOCK(&lk);
4776 		return;
4777 	}
4778 	ibp = inodedep->id_buf;
4779 	ibp = getdirtybuf(ibp, &lk, MNT_WAIT);
4780 	if (ibp == NULL) {
4781 		/*
4782 		 * If ibp came back as NULL, the dependency could have been
4783 		 * freed while we slept.  Look it up again, and check to see
4784 		 * that it has completed.
4785 		 */
4786 		if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) != 0)
4787 			goto retry;
4788 		FREE_LOCK(&lk);
4789 		return;
4790 	}
4791 	FREE_LOCK(&lk);
4792 	if ((error = bwrite(ibp)) != 0)
4793 		softdep_error("softdep_update_inodeblock: bwrite", error);
4794 }
4795 
4796 /*
4797  * Merge the a new inode dependency list (such as id_newinoupdt) into an
4798  * old inode dependency list (such as id_inoupdt). This routine must be
4799  * called with splbio interrupts blocked.
4800  */
4801 static void
4802 merge_inode_lists(newlisthead, oldlisthead)
4803 	struct allocdirectlst *newlisthead;
4804 	struct allocdirectlst *oldlisthead;
4805 {
4806 	struct allocdirect *listadp, *newadp;
4807 
4808 	newadp = TAILQ_FIRST(newlisthead);
4809 	for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
4810 		if (listadp->ad_lbn < newadp->ad_lbn) {
4811 			listadp = TAILQ_NEXT(listadp, ad_next);
4812 			continue;
4813 		}
4814 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4815 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
4816 		if (listadp->ad_lbn == newadp->ad_lbn) {
4817 			allocdirect_merge(oldlisthead, newadp,
4818 			    listadp);
4819 			listadp = newadp;
4820 		}
4821 		newadp = TAILQ_FIRST(newlisthead);
4822 	}
4823 	while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
4824 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4825 		TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
4826 	}
4827 }
4828 
4829 /*
4830  * If we are doing an fsync, then we must ensure that any directory
4831  * entries for the inode have been written after the inode gets to disk.
4832  */
4833 int
4834 softdep_fsync(vp)
4835 	struct vnode *vp;	/* the "in_core" copy of the inode */
4836 {
4837 	struct inodedep *inodedep;
4838 	struct pagedep *pagedep;
4839 	struct worklist *wk;
4840 	struct diradd *dap;
4841 	struct mount *mnt;
4842 	struct vnode *pvp;
4843 	struct inode *ip;
4844 	struct buf *bp;
4845 	struct fs *fs;
4846 	struct thread *td = curthread;
4847 	int error, flushparent;
4848 	ino_t parentino;
4849 	ufs_lbn_t lbn;
4850 
4851 	ip = VTOI(vp);
4852 	fs = ip->i_fs;
4853 	ACQUIRE_LOCK(&lk);
4854 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
4855 		FREE_LOCK(&lk);
4856 		return (0);
4857 	}
4858 	if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
4859 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
4860 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
4861 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
4862 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
4863 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL)
4864 		panic("softdep_fsync: pending ops");
4865 	for (error = 0, flushparent = 0; ; ) {
4866 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
4867 			break;
4868 		if (wk->wk_type != D_DIRADD)
4869 			panic("softdep_fsync: Unexpected type %s",
4870 			    TYPENAME(wk->wk_type));
4871 		dap = WK_DIRADD(wk);
4872 		/*
4873 		 * Flush our parent if this directory entry has a MKDIR_PARENT
4874 		 * dependency or is contained in a newly allocated block.
4875 		 */
4876 		if (dap->da_state & DIRCHG)
4877 			pagedep = dap->da_previous->dm_pagedep;
4878 		else
4879 			pagedep = dap->da_pagedep;
4880 		mnt = pagedep->pd_mnt;
4881 		parentino = pagedep->pd_ino;
4882 		lbn = pagedep->pd_lbn;
4883 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE)
4884 			panic("softdep_fsync: dirty");
4885 		if ((dap->da_state & MKDIR_PARENT) ||
4886 		    (pagedep->pd_state & NEWBLOCK))
4887 			flushparent = 1;
4888 		else
4889 			flushparent = 0;
4890 		/*
4891 		 * If we are being fsync'ed as part of vgone'ing this vnode,
4892 		 * then we will not be able to release and recover the
4893 		 * vnode below, so we just have to give up on writing its
4894 		 * directory entry out. It will eventually be written, just
4895 		 * not now, but then the user was not asking to have it
4896 		 * written, so we are not breaking any promises.
4897 		 */
4898 		if (vp->v_iflag & VI_DOOMED)
4899 			break;
4900 		/*
4901 		 * We prevent deadlock by always fetching inodes from the
4902 		 * root, moving down the directory tree. Thus, when fetching
4903 		 * our parent directory, we first try to get the lock. If
4904 		 * that fails, we must unlock ourselves before requesting
4905 		 * the lock on our parent. See the comment in ufs_lookup
4906 		 * for details on possible races.
4907 		 */
4908 		FREE_LOCK(&lk);
4909 		if (ffs_vget(mnt, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) {
4910 			VOP_UNLOCK(vp, 0, td);
4911 			error = ffs_vget(mnt, parentino, LK_EXCLUSIVE, &pvp);
4912 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
4913 			if (error != 0)
4914 				return (error);
4915 		}
4916 		/*
4917 		 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
4918 		 * that are contained in direct blocks will be resolved by
4919 		 * doing a ffs_update. Pagedeps contained in indirect blocks
4920 		 * may require a complete sync'ing of the directory. So, we
4921 		 * try the cheap and fast ffs_update first, and if that fails,
4922 		 * then we do the slower ffs_syncvnode of the directory.
4923 		 */
4924 		if (flushparent) {
4925 			if ((error = ffs_update(pvp, 1)) != 0) {
4926 				vput(pvp);
4927 				return (error);
4928 			}
4929 			if ((pagedep->pd_state & NEWBLOCK) &&
4930 			    (error = ffs_syncvnode(pvp, MNT_WAIT))) {
4931 				vput(pvp);
4932 				return (error);
4933 			}
4934 		}
4935 		/*
4936 		 * Flush directory page containing the inode's name.
4937 		 */
4938 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
4939 		    &bp);
4940 		if (error == 0)
4941 			error = bwrite(bp);
4942 		else
4943 			brelse(bp);
4944 		vput(pvp);
4945 		if (error != 0)
4946 			return (error);
4947 		ACQUIRE_LOCK(&lk);
4948 		if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
4949 			break;
4950 	}
4951 	FREE_LOCK(&lk);
4952 	return (0);
4953 }
4954 
4955 /*
4956  * Flush all the dirty bitmaps associated with the block device
4957  * before flushing the rest of the dirty blocks so as to reduce
4958  * the number of dependencies that will have to be rolled back.
4959  */
4960 void
4961 softdep_fsync_mountdev(vp)
4962 	struct vnode *vp;
4963 {
4964 	struct buf *bp, *nbp;
4965 	struct worklist *wk;
4966 
4967 	if (!vn_isdisk(vp, NULL))
4968 		panic("softdep_fsync_mountdev: vnode not a disk");
4969 restart:
4970 	ACQUIRE_LOCK(&lk);
4971 	VI_LOCK(vp);
4972 	TAILQ_FOREACH_SAFE(bp, &vp->v_bufobj.bo_dirty.bv_hd, b_bobufs, nbp) {
4973 		/*
4974 		 * If it is already scheduled, skip to the next buffer.
4975 		 */
4976 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL))
4977 			continue;
4978 
4979 		if ((bp->b_flags & B_DELWRI) == 0)
4980 			panic("softdep_fsync_mountdev: not dirty");
4981 		/*
4982 		 * We are only interested in bitmaps with outstanding
4983 		 * dependencies.
4984 		 */
4985 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
4986 		    wk->wk_type != D_BMSAFEMAP ||
4987 		    (bp->b_vflags & BV_BKGRDINPROG)) {
4988 			BUF_UNLOCK(bp);
4989 			continue;
4990 		}
4991 		VI_UNLOCK(vp);
4992 		FREE_LOCK(&lk);
4993 		bremfree(bp);
4994 		(void) bawrite(bp);
4995 		goto restart;
4996 	}
4997 	FREE_LOCK(&lk);
4998 	drain_output(vp);
4999 	VI_UNLOCK(vp);
5000 }
5001 
5002 /*
5003  * This routine is called when we are trying to synchronously flush a
5004  * file. This routine must eliminate any filesystem metadata dependencies
5005  * so that the syncing routine can succeed by pushing the dirty blocks
5006  * associated with the file. If any I/O errors occur, they are returned.
5007  */
5008 int
5009 softdep_sync_metadata(struct vnode *vp)
5010 {
5011 	struct pagedep *pagedep;
5012 	struct allocdirect *adp;
5013 	struct allocindir *aip;
5014 	struct buf *bp, *nbp;
5015 	struct worklist *wk;
5016 	int i, error, waitfor;
5017 
5018 	if (!DOINGSOFTDEP(vp))
5019 		return (0);
5020 	/*
5021 	 * Ensure that any direct block dependencies have been cleared.
5022 	 */
5023 	ACQUIRE_LOCK(&lk);
5024 	if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
5025 		FREE_LOCK(&lk);
5026 		return (error);
5027 	}
5028 	FREE_LOCK(&lk);
5029 	/*
5030 	 * For most files, the only metadata dependencies are the
5031 	 * cylinder group maps that allocate their inode or blocks.
5032 	 * The block allocation dependencies can be found by traversing
5033 	 * the dependency lists for any buffers that remain on their
5034 	 * dirty buffer list. The inode allocation dependency will
5035 	 * be resolved when the inode is updated with MNT_WAIT.
5036 	 * This work is done in two passes. The first pass grabs most
5037 	 * of the buffers and begins asynchronously writing them. The
5038 	 * only way to wait for these asynchronous writes is to sleep
5039 	 * on the filesystem vnode which may stay busy for a long time
5040 	 * if the filesystem is active. So, instead, we make a second
5041 	 * pass over the dependencies blocking on each write. In the
5042 	 * usual case we will be blocking against a write that we
5043 	 * initiated, so when it is done the dependency will have been
5044 	 * resolved. Thus the second pass is expected to end quickly.
5045 	 */
5046 	waitfor = MNT_NOWAIT;
5047 
5048 top:
5049 	/*
5050 	 * We must wait for any I/O in progress to finish so that
5051 	 * all potential buffers on the dirty list will be visible.
5052 	 */
5053 	VI_LOCK(vp);
5054 	drain_output(vp);
5055 	while ((bp = TAILQ_FIRST(&vp->v_bufobj.bo_dirty.bv_hd)) != NULL) {
5056 		bp = getdirtybuf(bp, VI_MTX(vp), MNT_WAIT);
5057 		if (bp)
5058 			break;
5059 	}
5060 	VI_UNLOCK(vp);
5061 	if (bp == NULL)
5062 		return (0);
5063 loop:
5064 	/* While syncing snapshots, we must allow recursive lookups */
5065 	bp->b_lock.lk_flags |= LK_CANRECURSE;
5066 	ACQUIRE_LOCK(&lk);
5067 	/*
5068 	 * As we hold the buffer locked, none of its dependencies
5069 	 * will disappear.
5070 	 */
5071 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5072 		switch (wk->wk_type) {
5073 
5074 		case D_ALLOCDIRECT:
5075 			adp = WK_ALLOCDIRECT(wk);
5076 			if (adp->ad_state & DEPCOMPLETE)
5077 				continue;
5078 			nbp = adp->ad_buf;
5079 			nbp = getdirtybuf(nbp, &lk, waitfor);
5080 			if (nbp == NULL)
5081 				continue;
5082 			FREE_LOCK(&lk);
5083 			if (waitfor == MNT_NOWAIT) {
5084 				bawrite(nbp);
5085 			} else if ((error = bwrite(nbp)) != 0) {
5086 				break;
5087 			}
5088 			ACQUIRE_LOCK(&lk);
5089 			continue;
5090 
5091 		case D_ALLOCINDIR:
5092 			aip = WK_ALLOCINDIR(wk);
5093 			if (aip->ai_state & DEPCOMPLETE)
5094 				continue;
5095 			nbp = aip->ai_buf;
5096 			nbp = getdirtybuf(nbp, &lk, waitfor);
5097 			if (nbp == NULL)
5098 				continue;
5099 			FREE_LOCK(&lk);
5100 			if (waitfor == MNT_NOWAIT) {
5101 				bawrite(nbp);
5102 			} else if ((error = bwrite(nbp)) != 0) {
5103 				break;
5104 			}
5105 			ACQUIRE_LOCK(&lk);
5106 			continue;
5107 
5108 		case D_INDIRDEP:
5109 		restart:
5110 
5111 			LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
5112 				if (aip->ai_state & DEPCOMPLETE)
5113 					continue;
5114 				nbp = aip->ai_buf;
5115 				nbp = getdirtybuf(nbp, &lk, MNT_WAIT);
5116 				if (nbp == NULL)
5117 					goto restart;
5118 				FREE_LOCK(&lk);
5119 				if ((error = bwrite(nbp)) != 0) {
5120 					break;
5121 				}
5122 				ACQUIRE_LOCK(&lk);
5123 				goto restart;
5124 			}
5125 			continue;
5126 
5127 		case D_INODEDEP:
5128 			if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
5129 			    WK_INODEDEP(wk)->id_ino)) != 0) {
5130 				FREE_LOCK(&lk);
5131 				break;
5132 			}
5133 			continue;
5134 
5135 		case D_PAGEDEP:
5136 			/*
5137 			 * We are trying to sync a directory that may
5138 			 * have dependencies on both its own metadata
5139 			 * and/or dependencies on the inodes of any
5140 			 * recently allocated files. We walk its diradd
5141 			 * lists pushing out the associated inode.
5142 			 */
5143 			pagedep = WK_PAGEDEP(wk);
5144 			for (i = 0; i < DAHASHSZ; i++) {
5145 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
5146 					continue;
5147 				if ((error =
5148 				    flush_pagedep_deps(vp, pagedep->pd_mnt,
5149 						&pagedep->pd_diraddhd[i]))) {
5150 					FREE_LOCK(&lk);
5151 					break;
5152 				}
5153 			}
5154 			continue;
5155 
5156 		case D_MKDIR:
5157 			/*
5158 			 * This case should never happen if the vnode has
5159 			 * been properly sync'ed. However, if this function
5160 			 * is used at a place where the vnode has not yet
5161 			 * been sync'ed, this dependency can show up. So,
5162 			 * rather than panic, just flush it.
5163 			 */
5164 			nbp = WK_MKDIR(wk)->md_buf;
5165 			nbp = getdirtybuf(nbp, &lk, waitfor);
5166 			if (nbp == NULL)
5167 				continue;
5168 			FREE_LOCK(&lk);
5169 			if (waitfor == MNT_NOWAIT) {
5170 				bawrite(nbp);
5171 			} else if ((error = bwrite(nbp)) != 0) {
5172 				break;
5173 			}
5174 			ACQUIRE_LOCK(&lk);
5175 			continue;
5176 
5177 		case D_BMSAFEMAP:
5178 			/*
5179 			 * This case should never happen if the vnode has
5180 			 * been properly sync'ed. However, if this function
5181 			 * is used at a place where the vnode has not yet
5182 			 * been sync'ed, this dependency can show up. So,
5183 			 * rather than panic, just flush it.
5184 			 */
5185 			nbp = WK_BMSAFEMAP(wk)->sm_buf;
5186 			nbp = getdirtybuf(nbp, &lk, waitfor);
5187 			if (nbp == NULL)
5188 				continue;
5189 			FREE_LOCK(&lk);
5190 			if (waitfor == MNT_NOWAIT) {
5191 				bawrite(nbp);
5192 			} else if ((error = bwrite(nbp)) != 0) {
5193 				break;
5194 			}
5195 			ACQUIRE_LOCK(&lk);
5196 			continue;
5197 
5198 		default:
5199 			panic("softdep_sync_metadata: Unknown type %s",
5200 			    TYPENAME(wk->wk_type));
5201 			/* NOTREACHED */
5202 		}
5203 		/* We reach here only in error and unlocked */
5204 		if (error == 0)
5205 			panic("softdep_sync_metadata: zero error");
5206 		bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5207 		bawrite(bp);
5208 		return (error);
5209 	}
5210 	FREE_LOCK(&lk);
5211 	VI_LOCK(vp);
5212 	while ((nbp = TAILQ_NEXT(bp, b_bobufs)) != NULL) {
5213 		nbp = getdirtybuf(nbp, VI_MTX(vp), MNT_WAIT);
5214 		if (nbp)
5215 			break;
5216 	}
5217 	VI_UNLOCK(vp);
5218 	bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5219 	bawrite(bp);
5220 	if (nbp != NULL) {
5221 		bp = nbp;
5222 		goto loop;
5223 	}
5224 	/*
5225 	 * The brief unlock is to allow any pent up dependency
5226 	 * processing to be done. Then proceed with the second pass.
5227 	 */
5228 	if (waitfor == MNT_NOWAIT) {
5229 		waitfor = MNT_WAIT;
5230 		goto top;
5231 	}
5232 
5233 	/*
5234 	 * If we have managed to get rid of all the dirty buffers,
5235 	 * then we are done. For certain directories and block
5236 	 * devices, we may need to do further work.
5237 	 *
5238 	 * We must wait for any I/O in progress to finish so that
5239 	 * all potential buffers on the dirty list will be visible.
5240 	 */
5241 	VI_LOCK(vp);
5242 	drain_output(vp);
5243 	VI_UNLOCK(vp);
5244 	return (0);
5245 }
5246 
5247 /*
5248  * Flush the dependencies associated with an inodedep.
5249  * Called with splbio blocked.
5250  */
5251 static int
5252 flush_inodedep_deps(fs, ino)
5253 	struct fs *fs;
5254 	ino_t ino;
5255 {
5256 	struct inodedep *inodedep;
5257 	int error, waitfor;
5258 
5259 	/*
5260 	 * This work is done in two passes. The first pass grabs most
5261 	 * of the buffers and begins asynchronously writing them. The
5262 	 * only way to wait for these asynchronous writes is to sleep
5263 	 * on the filesystem vnode which may stay busy for a long time
5264 	 * if the filesystem is active. So, instead, we make a second
5265 	 * pass over the dependencies blocking on each write. In the
5266 	 * usual case we will be blocking against a write that we
5267 	 * initiated, so when it is done the dependency will have been
5268 	 * resolved. Thus the second pass is expected to end quickly.
5269 	 * We give a brief window at the top of the loop to allow
5270 	 * any pending I/O to complete.
5271 	 */
5272 	for (error = 0, waitfor = MNT_NOWAIT; ; ) {
5273 		if (error)
5274 			return (error);
5275 		FREE_LOCK(&lk);
5276 		ACQUIRE_LOCK(&lk);
5277 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
5278 			return (0);
5279 		if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
5280 		    flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
5281 		    flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
5282 		    flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
5283 			continue;
5284 		/*
5285 		 * If pass2, we are done, otherwise do pass 2.
5286 		 */
5287 		if (waitfor == MNT_WAIT)
5288 			break;
5289 		waitfor = MNT_WAIT;
5290 	}
5291 	/*
5292 	 * Try freeing inodedep in case all dependencies have been removed.
5293 	 */
5294 	if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
5295 		(void) free_inodedep(inodedep);
5296 	return (0);
5297 }
5298 
5299 /*
5300  * Flush an inode dependency list.
5301  * Called with splbio blocked.
5302  */
5303 static int
5304 flush_deplist(listhead, waitfor, errorp)
5305 	struct allocdirectlst *listhead;
5306 	int waitfor;
5307 	int *errorp;
5308 {
5309 	struct allocdirect *adp;
5310 	struct buf *bp;
5311 
5312 	mtx_assert(&lk, MA_OWNED);
5313 	TAILQ_FOREACH(adp, listhead, ad_next) {
5314 		if (adp->ad_state & DEPCOMPLETE)
5315 			continue;
5316 		bp = adp->ad_buf;
5317 		bp = getdirtybuf(bp, &lk, waitfor);
5318 		if (bp == NULL) {
5319 			if (waitfor == MNT_NOWAIT)
5320 				continue;
5321 			return (1);
5322 		}
5323 		FREE_LOCK(&lk);
5324 		if (waitfor == MNT_NOWAIT) {
5325 			bawrite(bp);
5326 		} else if ((*errorp = bwrite(bp)) != 0) {
5327 			ACQUIRE_LOCK(&lk);
5328 			return (1);
5329 		}
5330 		ACQUIRE_LOCK(&lk);
5331 		return (1);
5332 	}
5333 	return (0);
5334 }
5335 
5336 /*
5337  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
5338  * Called with splbio blocked.
5339  */
5340 static int
5341 flush_pagedep_deps(pvp, mp, diraddhdp)
5342 	struct vnode *pvp;
5343 	struct mount *mp;
5344 	struct diraddhd *diraddhdp;
5345 {
5346 	struct inodedep *inodedep;
5347 	struct ufsmount *ump;
5348 	struct diradd *dap;
5349 	struct vnode *vp;
5350 	int error = 0;
5351 	struct buf *bp;
5352 	ino_t inum;
5353 
5354 	ump = VFSTOUFS(mp);
5355 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
5356 		/*
5357 		 * Flush ourselves if this directory entry
5358 		 * has a MKDIR_PARENT dependency.
5359 		 */
5360 		if (dap->da_state & MKDIR_PARENT) {
5361 			FREE_LOCK(&lk);
5362 			if ((error = ffs_update(pvp, 1)) != 0)
5363 				break;
5364 			ACQUIRE_LOCK(&lk);
5365 			/*
5366 			 * If that cleared dependencies, go on to next.
5367 			 */
5368 			if (dap != LIST_FIRST(diraddhdp))
5369 				continue;
5370 			if (dap->da_state & MKDIR_PARENT)
5371 				panic("flush_pagedep_deps: MKDIR_PARENT");
5372 		}
5373 		/*
5374 		 * A newly allocated directory must have its "." and
5375 		 * ".." entries written out before its name can be
5376 		 * committed in its parent. We do not want or need
5377 		 * the full semantics of a synchronous ffs_syncvnode as
5378 		 * that may end up here again, once for each directory
5379 		 * level in the filesystem. Instead, we push the blocks
5380 		 * and wait for them to clear. We have to fsync twice
5381 		 * because the first call may choose to defer blocks
5382 		 * that still have dependencies, but deferral will
5383 		 * happen at most once.
5384 		 */
5385 		inum = dap->da_newinum;
5386 		if (dap->da_state & MKDIR_BODY) {
5387 			FREE_LOCK(&lk);
5388 			if ((error = ffs_vget(mp, inum, LK_EXCLUSIVE, &vp)))
5389 				break;
5390 			if ((error=ffs_syncvnode(vp, MNT_NOWAIT)) ||
5391 			    (error=ffs_syncvnode(vp, MNT_NOWAIT))) {
5392 				vput(vp);
5393 				break;
5394 			}
5395 			VI_LOCK(vp);
5396 			drain_output(vp);
5397 			VI_UNLOCK(vp);
5398 			vput(vp);
5399 			ACQUIRE_LOCK(&lk);
5400 			/*
5401 			 * If that cleared dependencies, go on to next.
5402 			 */
5403 			if (dap != LIST_FIRST(diraddhdp))
5404 				continue;
5405 			if (dap->da_state & MKDIR_BODY)
5406 				panic("flush_pagedep_deps: MKDIR_BODY");
5407 		}
5408 		/*
5409 		 * Flush the inode on which the directory entry depends.
5410 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
5411 		 * the only remaining dependency is that the updated inode
5412 		 * count must get pushed to disk. The inode has already
5413 		 * been pushed into its inode buffer (via VOP_UPDATE) at
5414 		 * the time of the reference count change. So we need only
5415 		 * locate that buffer, ensure that there will be no rollback
5416 		 * caused by a bitmap dependency, then write the inode buffer.
5417 		 */
5418 retry:
5419 		if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0)
5420 			panic("flush_pagedep_deps: lost inode");
5421 		/*
5422 		 * If the inode still has bitmap dependencies,
5423 		 * push them to disk.
5424 		 */
5425 		if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5426 			bp = inodedep->id_buf;
5427 			bp = getdirtybuf(bp, &lk, MNT_WAIT);
5428 			if (bp == NULL)
5429 				goto retry;
5430 			FREE_LOCK(&lk);
5431 			if ((error = bwrite(bp)) != 0)
5432 				break;
5433 			ACQUIRE_LOCK(&lk);
5434 			if (dap != LIST_FIRST(diraddhdp))
5435 				continue;
5436 		}
5437 		/*
5438 		 * If the inode is still sitting in a buffer waiting
5439 		 * to be written, push it to disk.
5440 		 */
5441 		FREE_LOCK(&lk);
5442 		if ((error = bread(ump->um_devvp,
5443 		    fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
5444 		    (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) {
5445 			brelse(bp);
5446 			break;
5447 		}
5448 		if ((error = bwrite(bp)) != 0)
5449 			break;
5450 		ACQUIRE_LOCK(&lk);
5451 		/*
5452 		 * If we have failed to get rid of all the dependencies
5453 		 * then something is seriously wrong.
5454 		 */
5455 		if (dap == LIST_FIRST(diraddhdp))
5456 			panic("flush_pagedep_deps: flush failed");
5457 	}
5458 	if (error)
5459 		ACQUIRE_LOCK(&lk);
5460 	return (error);
5461 }
5462 
5463 /*
5464  * A large burst of file addition or deletion activity can drive the
5465  * memory load excessively high. First attempt to slow things down
5466  * using the techniques below. If that fails, this routine requests
5467  * the offending operations to fall back to running synchronously
5468  * until the memory load returns to a reasonable level.
5469  */
5470 int
5471 softdep_slowdown(vp)
5472 	struct vnode *vp;
5473 {
5474 	int max_softdeps_hard;
5475 
5476 	max_softdeps_hard = max_softdeps * 11 / 10;
5477 	if (num_dirrem < max_softdeps_hard / 2 &&
5478 	    num_inodedep < max_softdeps_hard &&
5479 	    VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps)
5480   		return (0);
5481 	if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps)
5482 		speedup_syncer();
5483 	stat_sync_limit_hit += 1;
5484 	return (1);
5485 }
5486 
5487 /*
5488  * Called by the allocation routines when they are about to fail
5489  * in the hope that we can free up some disk space.
5490  *
5491  * First check to see if the work list has anything on it. If it has,
5492  * clean up entries until we successfully free some space. Because this
5493  * process holds inodes locked, we cannot handle any remove requests
5494  * that might block on a locked inode as that could lead to deadlock.
5495  * If the worklist yields no free space, encourage the syncer daemon
5496  * to help us. In no event will we try for longer than tickdelay seconds.
5497  */
5498 int
5499 softdep_request_cleanup(fs, vp)
5500 	struct fs *fs;
5501 	struct vnode *vp;
5502 {
5503 	struct ufsmount *ump;
5504 	long starttime;
5505 	ufs2_daddr_t needed;
5506 	int error;
5507 
5508 	ump = VTOI(vp)->i_ump;
5509 	mtx_assert(UFS_MTX(ump), MA_OWNED);
5510 	needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize;
5511 	starttime = time_second + tickdelay;
5512 	/*
5513 	 * If we are being called because of a process doing a
5514 	 * copy-on-write, then it is not safe to update the vnode
5515 	 * as we may recurse into the copy-on-write routine.
5516 	 */
5517 	if (!(curthread->td_pflags & TDP_COWINPROGRESS)) {
5518 		UFS_UNLOCK(ump);
5519 		error = ffs_update(vp, 1);
5520 		UFS_LOCK(ump);
5521 		if (error != 0)
5522 			return (0);
5523 	}
5524 	while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) {
5525 		if (time_second > starttime)
5526 			return (0);
5527 		UFS_UNLOCK(ump);
5528 		ACQUIRE_LOCK(&lk);
5529 		if (num_on_worklist > 0 &&
5530 		    process_worklist_item(NULL, LK_NOWAIT) != -1) {
5531 			stat_worklist_push += 1;
5532 			FREE_LOCK(&lk);
5533 			UFS_LOCK(ump);
5534 			continue;
5535 		}
5536 		request_cleanup(FLUSH_REMOVE_WAIT);
5537 		FREE_LOCK(&lk);
5538 		UFS_LOCK(ump);
5539 	}
5540 	return (1);
5541 }
5542 
5543 /*
5544  * If memory utilization has gotten too high, deliberately slow things
5545  * down and speed up the I/O processing.
5546  */
5547 static int
5548 request_cleanup(resource)
5549 	int resource;
5550 {
5551 	struct thread *td = curthread;
5552 
5553 	mtx_assert(&lk, MA_OWNED);
5554 	/*
5555 	 * We never hold up the filesystem syncer process.
5556 	 */
5557 	if (td == filesys_syncer || (td->td_pflags & TDP_SOFTDEP))
5558 		return (0);
5559 	/*
5560 	 * First check to see if the work list has gotten backlogged.
5561 	 * If it has, co-opt this process to help clean up two entries.
5562 	 * Because this process may hold inodes locked, we cannot
5563 	 * handle any remove requests that might block on a locked
5564 	 * inode as that could lead to deadlock.  We set TDP_SOFTDEP
5565 	 * to avoid recursively processing the worklist.
5566 	 */
5567 	if (num_on_worklist > max_softdeps / 10) {
5568 		td->td_pflags |= TDP_SOFTDEP;
5569 		process_worklist_item(NULL, LK_NOWAIT);
5570 		process_worklist_item(NULL, LK_NOWAIT);
5571 		td->td_pflags &= ~TDP_SOFTDEP;
5572 		stat_worklist_push += 2;
5573 		return(1);
5574 	}
5575 	/*
5576 	 * Next, we attempt to speed up the syncer process. If that
5577 	 * is successful, then we allow the process to continue.
5578 	 */
5579 	if (speedup_syncer() && resource != FLUSH_REMOVE_WAIT)
5580 		return(0);
5581 	/*
5582 	 * If we are resource constrained on inode dependencies, try
5583 	 * flushing some dirty inodes. Otherwise, we are constrained
5584 	 * by file deletions, so try accelerating flushes of directories
5585 	 * with removal dependencies. We would like to do the cleanup
5586 	 * here, but we probably hold an inode locked at this point and
5587 	 * that might deadlock against one that we try to clean. So,
5588 	 * the best that we can do is request the syncer daemon to do
5589 	 * the cleanup for us.
5590 	 */
5591 	switch (resource) {
5592 
5593 	case FLUSH_INODES:
5594 		stat_ino_limit_push += 1;
5595 		req_clear_inodedeps += 1;
5596 		stat_countp = &stat_ino_limit_hit;
5597 		break;
5598 
5599 	case FLUSH_REMOVE:
5600 	case FLUSH_REMOVE_WAIT:
5601 		stat_blk_limit_push += 1;
5602 		req_clear_remove += 1;
5603 		stat_countp = &stat_blk_limit_hit;
5604 		break;
5605 
5606 	default:
5607 		panic("request_cleanup: unknown type");
5608 	}
5609 	/*
5610 	 * Hopefully the syncer daemon will catch up and awaken us.
5611 	 * We wait at most tickdelay before proceeding in any case.
5612 	 */
5613 	proc_waiting += 1;
5614 	if (handle.callout == NULL)
5615 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5616 	msleep((caddr_t)&proc_waiting, &lk, PPAUSE, "softupdate", 0);
5617 	proc_waiting -= 1;
5618 	return (1);
5619 }
5620 
5621 /*
5622  * Awaken processes pausing in request_cleanup and clear proc_waiting
5623  * to indicate that there is no longer a timer running.
5624  */
5625 static void
5626 pause_timer(arg)
5627 	void *arg;
5628 {
5629 
5630 	ACQUIRE_LOCK(&lk);
5631 	*stat_countp += 1;
5632 	wakeup_one(&proc_waiting);
5633 	if (proc_waiting > 0)
5634 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5635 	else
5636 		handle.callout = NULL;
5637 	FREE_LOCK(&lk);
5638 }
5639 
5640 /*
5641  * Flush out a directory with at least one removal dependency in an effort to
5642  * reduce the number of dirrem, freefile, and freeblks dependency structures.
5643  */
5644 static void
5645 clear_remove(td)
5646 	struct thread *td;
5647 {
5648 	struct pagedep_hashhead *pagedephd;
5649 	struct pagedep *pagedep;
5650 	static int next = 0;
5651 	struct mount *mp;
5652 	struct vnode *vp;
5653 	int error, cnt;
5654 	ino_t ino;
5655 
5656 	mtx_assert(&lk, MA_OWNED);
5657 
5658 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
5659 		pagedephd = &pagedep_hashtbl[next++];
5660 		if (next >= pagedep_hash)
5661 			next = 0;
5662 		LIST_FOREACH(pagedep, pagedephd, pd_hash) {
5663 			if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
5664 				continue;
5665 			mp = pagedep->pd_mnt;
5666 			ino = pagedep->pd_ino;
5667 			if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5668 				continue;
5669 			FREE_LOCK(&lk);
5670 			if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp))) {
5671 				softdep_error("clear_remove: vget", error);
5672 				vn_finished_write(mp);
5673 				ACQUIRE_LOCK(&lk);
5674 				return;
5675 			}
5676 			if ((error = ffs_syncvnode(vp, MNT_NOWAIT)))
5677 				softdep_error("clear_remove: fsync", error);
5678 			VI_LOCK(vp);
5679 			drain_output(vp);
5680 			VI_UNLOCK(vp);
5681 			vput(vp);
5682 			vn_finished_write(mp);
5683 			ACQUIRE_LOCK(&lk);
5684 			return;
5685 		}
5686 	}
5687 }
5688 
5689 /*
5690  * Clear out a block of dirty inodes in an effort to reduce
5691  * the number of inodedep dependency structures.
5692  */
5693 static void
5694 clear_inodedeps(td)
5695 	struct thread *td;
5696 {
5697 	struct inodedep_hashhead *inodedephd;
5698 	struct inodedep *inodedep;
5699 	static int next = 0;
5700 	struct mount *mp;
5701 	struct vnode *vp;
5702 	struct fs *fs;
5703 	int error, cnt;
5704 	ino_t firstino, lastino, ino;
5705 
5706 	mtx_assert(&lk, MA_OWNED);
5707 	/*
5708 	 * Pick a random inode dependency to be cleared.
5709 	 * We will then gather up all the inodes in its block
5710 	 * that have dependencies and flush them out.
5711 	 */
5712 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
5713 		inodedephd = &inodedep_hashtbl[next++];
5714 		if (next >= inodedep_hash)
5715 			next = 0;
5716 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
5717 			break;
5718 	}
5719 	if (inodedep == NULL)
5720 		return;
5721 	/*
5722 	 * Ugly code to find mount point given pointer to superblock.
5723 	 */
5724 	fs = inodedep->id_fs;
5725 	TAILQ_FOREACH(mp, &mountlist, mnt_list)
5726 		if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs)
5727 			break;
5728 	/*
5729 	 * Find the last inode in the block with dependencies.
5730 	 */
5731 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
5732 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
5733 		if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
5734 			break;
5735 	/*
5736 	 * Asynchronously push all but the last inode with dependencies.
5737 	 * Synchronously push the last inode with dependencies to ensure
5738 	 * that the inode block gets written to free up the inodedeps.
5739 	 */
5740 	for (ino = firstino; ino <= lastino; ino++) {
5741 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
5742 			continue;
5743 		if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5744 			continue;
5745 		FREE_LOCK(&lk);
5746 		if ((error = ffs_vget(mp, ino, LK_EXCLUSIVE, &vp)) != 0) {
5747 			softdep_error("clear_inodedeps: vget", error);
5748 			vn_finished_write(mp);
5749 			ACQUIRE_LOCK(&lk);
5750 			return;
5751 		}
5752 		if (ino == lastino) {
5753 			if ((error = ffs_syncvnode(vp, MNT_WAIT)))
5754 				softdep_error("clear_inodedeps: fsync1", error);
5755 		} else {
5756 			if ((error = ffs_syncvnode(vp, MNT_NOWAIT)))
5757 				softdep_error("clear_inodedeps: fsync2", error);
5758 			VI_LOCK(vp);
5759 			drain_output(vp);
5760 			VI_UNLOCK(vp);
5761 		}
5762 		vput(vp);
5763 		vn_finished_write(mp);
5764 		ACQUIRE_LOCK(&lk);
5765 	}
5766 }
5767 
5768 /*
5769  * Function to determine if the buffer has outstanding dependencies
5770  * that will cause a roll-back if the buffer is written. If wantcount
5771  * is set, return number of dependencies, otherwise just yes or no.
5772  */
5773 static int
5774 softdep_count_dependencies(bp, wantcount)
5775 	struct buf *bp;
5776 	int wantcount;
5777 {
5778 	struct worklist *wk;
5779 	struct inodedep *inodedep;
5780 	struct indirdep *indirdep;
5781 	struct allocindir *aip;
5782 	struct pagedep *pagedep;
5783 	struct diradd *dap;
5784 	int i, retval;
5785 
5786 	retval = 0;
5787 	ACQUIRE_LOCK(&lk);
5788 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5789 		switch (wk->wk_type) {
5790 
5791 		case D_INODEDEP:
5792 			inodedep = WK_INODEDEP(wk);
5793 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5794 				/* bitmap allocation dependency */
5795 				retval += 1;
5796 				if (!wantcount)
5797 					goto out;
5798 			}
5799 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
5800 				/* direct block pointer dependency */
5801 				retval += 1;
5802 				if (!wantcount)
5803 					goto out;
5804 			}
5805 			if (TAILQ_FIRST(&inodedep->id_extupdt)) {
5806 				/* direct block pointer dependency */
5807 				retval += 1;
5808 				if (!wantcount)
5809 					goto out;
5810 			}
5811 			continue;
5812 
5813 		case D_INDIRDEP:
5814 			indirdep = WK_INDIRDEP(wk);
5815 
5816 			LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
5817 				/* indirect block pointer dependency */
5818 				retval += 1;
5819 				if (!wantcount)
5820 					goto out;
5821 			}
5822 			continue;
5823 
5824 		case D_PAGEDEP:
5825 			pagedep = WK_PAGEDEP(wk);
5826 			for (i = 0; i < DAHASHSZ; i++) {
5827 
5828 				LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
5829 					/* directory entry dependency */
5830 					retval += 1;
5831 					if (!wantcount)
5832 						goto out;
5833 				}
5834 			}
5835 			continue;
5836 
5837 		case D_BMSAFEMAP:
5838 		case D_ALLOCDIRECT:
5839 		case D_ALLOCINDIR:
5840 		case D_MKDIR:
5841 			/* never a dependency on these blocks */
5842 			continue;
5843 
5844 		default:
5845 			panic("softdep_check_for_rollback: Unexpected type %s",
5846 			    TYPENAME(wk->wk_type));
5847 			/* NOTREACHED */
5848 		}
5849 	}
5850 out:
5851 	FREE_LOCK(&lk);
5852 	return retval;
5853 }
5854 
5855 /*
5856  * Acquire exclusive access to a buffer.
5857  * Must be called with a locked mtx parameter.
5858  * Return acquired buffer or NULL on failure.
5859  */
5860 static struct buf *
5861 getdirtybuf(bp, mtx, waitfor)
5862 	struct buf *bp;
5863 	struct mtx *mtx;
5864 	int waitfor;
5865 {
5866 	int error;
5867 
5868 	mtx_assert(mtx, MA_OWNED);
5869 	if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT, NULL) != 0) {
5870 		if (waitfor != MNT_WAIT)
5871 			return (NULL);
5872 		error = BUF_LOCK(bp,
5873 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, mtx);
5874 		/*
5875 		 * Even if we sucessfully acquire bp here, we have dropped
5876 		 * mtx, which may violates our guarantee.
5877 		 */
5878 		if (error == 0)
5879 			BUF_UNLOCK(bp);
5880 		else if (error != ENOLCK)
5881 			panic("getdirtybuf: inconsistent lock: %d", error);
5882 		mtx_lock(mtx);
5883 		return (NULL);
5884 	}
5885 	if ((bp->b_vflags & BV_BKGRDINPROG) != 0) {
5886 		BUF_UNLOCK(bp);
5887 		if (waitfor != MNT_WAIT)
5888 			return (NULL);
5889 		/*
5890 		 * The mtx argument must be bp->b_vp's mutex in
5891 		 * this case.
5892 		 */
5893 #ifdef	DEBUG_VFS_LOCKS
5894 		if (bp->b_vp->v_type != VCHR)
5895 			ASSERT_VI_LOCKED(bp->b_vp, "getdirtybuf");
5896 #endif
5897 		bp->b_vflags |= BV_BKGRDWAIT;
5898 		msleep(&bp->b_xflags, mtx, PRIBIO, "getbuf", 0);
5899 		return (NULL);
5900 	}
5901 	if ((bp->b_flags & B_DELWRI) == 0) {
5902 		BUF_UNLOCK(bp);
5903 		return (NULL);
5904 	}
5905 	bremfree(bp);
5906 	return (bp);
5907 }
5908 
5909 /*
5910  * Wait for pending output on a vnode to complete.
5911  * Must be called with vnode lock and interlock locked.
5912  *
5913  * XXX: Should just be a call to bufobj_wwait().
5914  */
5915 static void
5916 drain_output(vp)
5917 	struct vnode *vp;
5918 {
5919 	ASSERT_VOP_LOCKED(vp, "drain_output");
5920 	ASSERT_VI_LOCKED(vp, "drain_output");
5921 
5922 	while (vp->v_bufobj.bo_numoutput) {
5923 		vp->v_bufobj.bo_flag |= BO_WWAIT;
5924 		msleep((caddr_t)&vp->v_bufobj.bo_numoutput,
5925 		    VI_MTX(vp), PRIBIO + 1, "drainvp", 0);
5926 	}
5927 }
5928 
5929 /*
5930  * Called whenever a buffer that is being invalidated or reallocated
5931  * contains dependencies. This should only happen if an I/O error has
5932  * occurred. The routine is called with the buffer locked.
5933  */
5934 static void
5935 softdep_deallocate_dependencies(bp)
5936 	struct buf *bp;
5937 {
5938 
5939 	if ((bp->b_ioflags & BIO_ERROR) == 0)
5940 		panic("softdep_deallocate_dependencies: dangling deps");
5941 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
5942 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
5943 }
5944 
5945 /*
5946  * Function to handle asynchronous write errors in the filesystem.
5947  */
5948 static void
5949 softdep_error(func, error)
5950 	char *func;
5951 	int error;
5952 {
5953 
5954 	/* XXX should do something better! */
5955 	printf("%s: got error %d while accessing filesystem\n", func, error);
5956 }
5957 
5958 #endif /* SOFTUPDATES */
5959