xref: /freebsd/sys/ufs/ffs/ffs_softdep.c (revision 4b2eaea43fec8e8792be611dea204071a10b655a)
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/stdint.h>
58 #include <sys/bio.h>
59 #include <sys/buf.h>
60 #include <sys/malloc.h>
61 #include <sys/mount.h>
62 #include <sys/proc.h>
63 #include <sys/stat.h>
64 #include <sys/syslog.h>
65 #include <sys/vnode.h>
66 #include <sys/conf.h>
67 #include <ufs/ufs/dir.h>
68 #include <ufs/ufs/extattr.h>
69 #include <ufs/ufs/quota.h>
70 #include <ufs/ufs/inode.h>
71 #include <ufs/ufs/ufsmount.h>
72 #include <ufs/ffs/fs.h>
73 #include <ufs/ffs/softdep.h>
74 #include <ufs/ffs/ffs_extern.h>
75 #include <ufs/ufs/ufs_extern.h>
76 
77 /*
78  * These definitions need to be adapted to the system to which
79  * this file is being ported.
80  */
81 /*
82  * malloc types defined for the softdep system.
83  */
84 static MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
85 static MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
86 static MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
87 static MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
88 static MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
89 static MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
90 static MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
91 static MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
92 static MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
93 static MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
94 static MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
95 static MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
96 static MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
97 static MALLOC_DEFINE(M_NEWDIRBLK, "newdirblk","Unclaimed new directory block");
98 
99 #define M_SOFTDEP_FLAGS	(M_USE_RESERVE)
100 
101 #define	D_PAGEDEP	0
102 #define	D_INODEDEP	1
103 #define	D_NEWBLK	2
104 #define	D_BMSAFEMAP	3
105 #define	D_ALLOCDIRECT	4
106 #define	D_INDIRDEP	5
107 #define	D_ALLOCINDIR	6
108 #define	D_FREEFRAG	7
109 #define	D_FREEBLKS	8
110 #define	D_FREEFILE	9
111 #define	D_DIRADD	10
112 #define	D_MKDIR		11
113 #define	D_DIRREM	12
114 #define	D_NEWDIRBLK	13
115 #define	D_LAST		D_NEWDIRBLK
116 
117 /*
118  * translate from workitem type to memory type
119  * MUST match the defines above, such that memtype[D_XXX] == M_XXX
120  */
121 static struct malloc_type *memtype[] = {
122 	M_PAGEDEP,
123 	M_INODEDEP,
124 	M_NEWBLK,
125 	M_BMSAFEMAP,
126 	M_ALLOCDIRECT,
127 	M_INDIRDEP,
128 	M_ALLOCINDIR,
129 	M_FREEFRAG,
130 	M_FREEBLKS,
131 	M_FREEFILE,
132 	M_DIRADD,
133 	M_MKDIR,
134 	M_DIRREM,
135 	M_NEWDIRBLK
136 };
137 
138 #define DtoM(type) (memtype[type])
139 
140 /*
141  * Names of malloc types.
142  */
143 #define TYPENAME(type)  \
144 	((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
145 /*
146  * End system adaptaion definitions.
147  */
148 
149 /*
150  * Internal function prototypes.
151  */
152 static	void softdep_error(char *, int);
153 static	void drain_output(struct vnode *, int);
154 static	int getdirtybuf(struct buf **, int);
155 static	void clear_remove(struct thread *);
156 static	void clear_inodedeps(struct thread *);
157 static	int flush_pagedep_deps(struct vnode *, struct mount *,
158 	    struct diraddhd *);
159 static	int flush_inodedep_deps(struct fs *, ino_t);
160 static	int flush_deplist(struct allocdirectlst *, int, int *);
161 static	int handle_written_filepage(struct pagedep *, struct buf *);
162 static  void diradd_inode_written(struct diradd *, struct inodedep *);
163 static	int handle_written_inodeblock(struct inodedep *, struct buf *);
164 static	void handle_allocdirect_partdone(struct allocdirect *);
165 static	void handle_allocindir_partdone(struct allocindir *);
166 static	void initiate_write_filepage(struct pagedep *, struct buf *);
167 static	void handle_written_mkdir(struct mkdir *, int);
168 static	void initiate_write_inodeblock_ufs1(struct inodedep *, struct buf *);
169 static	void initiate_write_inodeblock_ufs2(struct inodedep *, struct buf *);
170 static	void handle_workitem_freefile(struct freefile *);
171 static	void handle_workitem_remove(struct dirrem *, struct vnode *);
172 static	struct dirrem *newdirrem(struct buf *, struct inode *,
173 	    struct inode *, int, struct dirrem **);
174 static	void free_diradd(struct diradd *);
175 static	void free_allocindir(struct allocindir *, struct inodedep *);
176 static	void free_newdirblk(struct newdirblk *);
177 static	int indir_trunc(struct freeblks *, ufs2_daddr_t, int, ufs_lbn_t,
178 	    ufs2_daddr_t *);
179 static	void deallocate_dependencies(struct buf *, struct inodedep *);
180 static	void free_allocdirect(struct allocdirectlst *,
181 	    struct allocdirect *, int);
182 static	int check_inode_unwritten(struct inodedep *);
183 static	int free_inodedep(struct inodedep *);
184 static	void handle_workitem_freeblocks(struct freeblks *, int);
185 static	void merge_inode_lists(struct allocdirectlst *,struct allocdirectlst *);
186 static	void setup_allocindir_phase2(struct buf *, struct inode *,
187 	    struct allocindir *);
188 static	struct allocindir *newallocindir(struct inode *, int, ufs2_daddr_t,
189 	    ufs2_daddr_t);
190 static	void handle_workitem_freefrag(struct freefrag *);
191 static	struct freefrag *newfreefrag(struct inode *, ufs2_daddr_t, long);
192 static	void allocdirect_merge(struct allocdirectlst *,
193 	    struct allocdirect *, struct allocdirect *);
194 static	struct bmsafemap *bmsafemap_lookup(struct buf *);
195 static	int newblk_lookup(struct fs *, ufs2_daddr_t, int, struct newblk **);
196 static	int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
197 static	int pagedep_lookup(struct inode *, ufs_lbn_t, int, struct pagedep **);
198 static	void pause_timer(void *);
199 static	int request_cleanup(int, int);
200 static	int process_worklist_item(struct mount *, int);
201 static	void add_to_worklist(struct worklist *);
202 
203 /*
204  * Exported softdep operations.
205  */
206 static	void softdep_disk_io_initiation(struct buf *);
207 static	void softdep_disk_write_complete(struct buf *);
208 static	void softdep_deallocate_dependencies(struct buf *);
209 static	void softdep_move_dependencies(struct buf *, struct buf *);
210 static	int softdep_count_dependencies(struct buf *bp, int);
211 
212 /*
213  * Locking primitives.
214  *
215  * For a uniprocessor, all we need to do is protect against disk
216  * interrupts. For a multiprocessor, this lock would have to be
217  * a mutex. A single mutex is used throughout this file, though
218  * finer grain locking could be used if contention warranted it.
219  *
220  * For a multiprocessor, the sleep call would accept a lock and
221  * release it after the sleep processing was complete. In a uniprocessor
222  * implementation there is no such interlock, so we simple mark
223  * the places where it needs to be done with the `interlocked' form
224  * of the lock calls. Since the uniprocessor sleep already interlocks
225  * the spl, there is nothing that really needs to be done.
226  */
227 #ifndef /* NOT */ DEBUG
228 static struct lockit {
229 	int	lkt_spl;
230 } lk = { 0 };
231 #define ACQUIRE_LOCK(lk)		(lk)->lkt_spl = splbio()
232 #define FREE_LOCK(lk)			splx((lk)->lkt_spl)
233 
234 #else /* DEBUG */
235 #define NOHOLDER	((struct thread *)-1)
236 #define SPECIAL_FLAG	((struct thread *)-2)
237 static struct lockit {
238 	int	lkt_spl;
239 	struct	thread *lkt_held;
240 } lk = { 0, NOHOLDER };
241 
242 static	void acquire_lock(struct lockit *);
243 static	void free_lock(struct lockit *);
244 void	softdep_panic(char *);
245 
246 #define ACQUIRE_LOCK(lk)		acquire_lock(lk)
247 #define FREE_LOCK(lk)			free_lock(lk)
248 
249 static void
250 acquire_lock(lk)
251 	struct lockit *lk;
252 {
253 	struct thread *holder;
254 
255 	if (lk->lkt_held != NOHOLDER) {
256 		holder = lk->lkt_held;
257 		FREE_LOCK(lk);
258 		if (holder == curthread)
259 			panic("softdep_lock: locking against myself");
260 		else
261 			panic("softdep_lock: lock held by %p", holder);
262 	}
263 	lk->lkt_spl = splbio();
264 	lk->lkt_held = curthread;
265 }
266 
267 static void
268 free_lock(lk)
269 	struct lockit *lk;
270 {
271 
272 	if (lk->lkt_held == NOHOLDER)
273 		panic("softdep_unlock: lock not held");
274 	lk->lkt_held = NOHOLDER;
275 	splx(lk->lkt_spl);
276 }
277 
278 /*
279  * Function to release soft updates lock and panic.
280  */
281 void
282 softdep_panic(msg)
283 	char *msg;
284 {
285 
286 	if (lk.lkt_held != NOHOLDER)
287 		FREE_LOCK(&lk);
288 	panic(msg);
289 }
290 #endif /* DEBUG */
291 
292 static	int interlocked_sleep(struct lockit *, int, void *, struct mtx *, int,
293 	    const char *, int);
294 
295 /*
296  * When going to sleep, we must save our SPL so that it does
297  * not get lost if some other process uses the lock while we
298  * are sleeping. We restore it after we have slept. This routine
299  * wraps the interlocking with functions that sleep. The list
300  * below enumerates the available set of operations.
301  */
302 #define	UNKNOWN		0
303 #define	SLEEP		1
304 #define	LOCKBUF		2
305 
306 static int
307 interlocked_sleep(lk, op, ident, mtx, flags, wmesg, timo)
308 	struct lockit *lk;
309 	int op;
310 	void *ident;
311 	struct mtx *mtx;
312 	int flags;
313 	const char *wmesg;
314 	int timo;
315 {
316 	struct thread *holder;
317 	int s, retval;
318 
319 	s = lk->lkt_spl;
320 #	ifdef DEBUG
321 	if (lk->lkt_held == NOHOLDER)
322 		panic("interlocked_sleep: lock not held");
323 	lk->lkt_held = NOHOLDER;
324 #	endif /* DEBUG */
325 	switch (op) {
326 	case SLEEP:
327 		retval = msleep(ident, mtx, flags, wmesg, timo);
328 		break;
329 	case LOCKBUF:
330 		retval = BUF_LOCK((struct buf *)ident, flags);
331 		break;
332 	default:
333 		panic("interlocked_sleep: unknown operation");
334 	}
335 #	ifdef DEBUG
336 	if (lk->lkt_held != NOHOLDER) {
337 		holder = lk->lkt_held;
338 		FREE_LOCK(lk);
339 		if (holder == curthread)
340 			panic("interlocked_sleep: locking against self");
341 		else
342 			panic("interlocked_sleep: lock held by %p", holder);
343 	}
344 	lk->lkt_held = curthread;
345 #	endif /* DEBUG */
346 	lk->lkt_spl = s;
347 	return (retval);
348 }
349 
350 /*
351  * Place holder for real semaphores.
352  */
353 struct sema {
354 	int	value;
355 	struct	thread *holder;
356 	char	*name;
357 	int	prio;
358 	int	timo;
359 };
360 static	void sema_init(struct sema *, char *, int, int);
361 static	int sema_get(struct sema *, struct lockit *);
362 static	void sema_release(struct sema *);
363 
364 static void
365 sema_init(semap, name, prio, timo)
366 	struct sema *semap;
367 	char *name;
368 	int prio, timo;
369 {
370 
371 	semap->holder = NOHOLDER;
372 	semap->value = 0;
373 	semap->name = name;
374 	semap->prio = prio;
375 	semap->timo = timo;
376 }
377 
378 static int
379 sema_get(semap, interlock)
380 	struct sema *semap;
381 	struct lockit *interlock;
382 {
383 
384 	if (semap->value++ > 0) {
385 		if (interlock != NULL) {
386 			interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
387 			    NULL, semap->prio, semap->name,
388 			    semap->timo);
389 			FREE_LOCK(interlock);
390 		} else {
391 			tsleep((caddr_t)semap, semap->prio, semap->name,
392 			    semap->timo);
393 		}
394 		return (0);
395 	}
396 	semap->holder = curthread;
397 	if (interlock != NULL)
398 		FREE_LOCK(interlock);
399 	return (1);
400 }
401 
402 static void
403 sema_release(semap)
404 	struct sema *semap;
405 {
406 
407 	if (semap->value <= 0 || semap->holder != curthread) {
408 		if (lk.lkt_held != NOHOLDER)
409 			FREE_LOCK(&lk);
410 		panic("sema_release: not held");
411 	}
412 	if (--semap->value > 0) {
413 		semap->value = 0;
414 		wakeup(semap);
415 	}
416 	semap->holder = NOHOLDER;
417 }
418 
419 /*
420  * Worklist queue management.
421  * These routines require that the lock be held.
422  */
423 #ifndef /* NOT */ DEBUG
424 #define WORKLIST_INSERT(head, item) do {	\
425 	(item)->wk_state |= ONWORKLIST;		\
426 	LIST_INSERT_HEAD(head, item, wk_list);	\
427 } while (0)
428 #define WORKLIST_REMOVE(item) do {		\
429 	(item)->wk_state &= ~ONWORKLIST;	\
430 	LIST_REMOVE(item, wk_list);		\
431 } while (0)
432 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
433 
434 #else /* DEBUG */
435 static	void worklist_insert(struct workhead *, struct worklist *);
436 static	void worklist_remove(struct worklist *);
437 static	void workitem_free(struct worklist *, int);
438 
439 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
440 #define WORKLIST_REMOVE(item) worklist_remove(item)
441 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
442 
443 static void
444 worklist_insert(head, item)
445 	struct workhead *head;
446 	struct worklist *item;
447 {
448 
449 	if (lk.lkt_held == NOHOLDER)
450 		panic("worklist_insert: lock not held");
451 	if (item->wk_state & ONWORKLIST) {
452 		FREE_LOCK(&lk);
453 		panic("worklist_insert: already on list");
454 	}
455 	item->wk_state |= ONWORKLIST;
456 	LIST_INSERT_HEAD(head, item, wk_list);
457 }
458 
459 static void
460 worklist_remove(item)
461 	struct worklist *item;
462 {
463 
464 	if (lk.lkt_held == NOHOLDER)
465 		panic("worklist_remove: lock not held");
466 	if ((item->wk_state & ONWORKLIST) == 0) {
467 		FREE_LOCK(&lk);
468 		panic("worklist_remove: not on list");
469 	}
470 	item->wk_state &= ~ONWORKLIST;
471 	LIST_REMOVE(item, wk_list);
472 }
473 
474 static void
475 workitem_free(item, type)
476 	struct worklist *item;
477 	int type;
478 {
479 
480 	if (item->wk_state & ONWORKLIST) {
481 		if (lk.lkt_held != NOHOLDER)
482 			FREE_LOCK(&lk);
483 		panic("workitem_free: still on list");
484 	}
485 	if (item->wk_type != type) {
486 		if (lk.lkt_held != NOHOLDER)
487 			FREE_LOCK(&lk);
488 		panic("workitem_free: type mismatch");
489 	}
490 	FREE(item, DtoM(type));
491 }
492 #endif /* DEBUG */
493 
494 /*
495  * Workitem queue management
496  */
497 static struct workhead softdep_workitem_pending;
498 static int num_on_worklist;	/* number of worklist items to be processed */
499 static int softdep_worklist_busy; /* 1 => trying to do unmount */
500 static int softdep_worklist_req; /* serialized waiters */
501 static int max_softdeps;	/* maximum number of structs before slowdown */
502 static int maxindirdeps = 50;	/* max number of indirdeps before slowdown */
503 static int tickdelay = 2;	/* number of ticks to pause during slowdown */
504 static int proc_waiting;	/* tracks whether we have a timeout posted */
505 static int *stat_countp;	/* statistic to count in proc_waiting timeout */
506 static struct callout_handle handle; /* handle on posted proc_waiting timeout */
507 static struct thread *filesys_syncer; /* proc of filesystem syncer process */
508 static int req_clear_inodedeps;	/* syncer process flush some inodedeps */
509 #define FLUSH_INODES		1
510 static int req_clear_remove;	/* syncer process flush some freeblks */
511 #define FLUSH_REMOVE		2
512 #define FLUSH_REMOVE_WAIT	3
513 /*
514  * runtime statistics
515  */
516 static int stat_worklist_push;	/* number of worklist cleanups */
517 static int stat_blk_limit_push;	/* number of times block limit neared */
518 static int stat_ino_limit_push;	/* number of times inode limit neared */
519 static int stat_blk_limit_hit;	/* number of times block slowdown imposed */
520 static int stat_ino_limit_hit;	/* number of times inode slowdown imposed */
521 static int stat_sync_limit_hit;	/* number of synchronous slowdowns imposed */
522 static int stat_indir_blk_ptrs;	/* bufs redirtied as indir ptrs not written */
523 static int stat_inode_bitmap;	/* bufs redirtied as inode bitmap not written */
524 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
525 static int stat_dir_entry;	/* bufs redirtied as dir entry cannot write */
526 #ifdef DEBUG
527 #include <vm/vm.h>
528 #include <sys/sysctl.h>
529 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
530 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
531 SYSCTL_INT(_debug, OID_AUTO, maxindirdeps, CTLFLAG_RW, &maxindirdeps, 0, "");
532 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
533 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
534 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
535 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
536 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
537 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
538 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
539 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
540 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
541 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
542 #endif /* DEBUG */
543 
544 /*
545  * Add an item to the end of the work queue.
546  * This routine requires that the lock be held.
547  * This is the only routine that adds items to the list.
548  * The following routine is the only one that removes items
549  * and does so in order from first to last.
550  */
551 static void
552 add_to_worklist(wk)
553 	struct worklist *wk;
554 {
555 	static struct worklist *worklist_tail;
556 
557 	if (wk->wk_state & ONWORKLIST) {
558 		if (lk.lkt_held != NOHOLDER)
559 			FREE_LOCK(&lk);
560 		panic("add_to_worklist: already on list");
561 	}
562 	wk->wk_state |= ONWORKLIST;
563 	if (LIST_FIRST(&softdep_workitem_pending) == NULL)
564 		LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
565 	else
566 		LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
567 	worklist_tail = wk;
568 	num_on_worklist += 1;
569 }
570 
571 /*
572  * Process that runs once per second to handle items in the background queue.
573  *
574  * Note that we ensure that everything is done in the order in which they
575  * appear in the queue. The code below depends on this property to ensure
576  * that blocks of a file are freed before the inode itself is freed. This
577  * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
578  * until all the old ones have been purged from the dependency lists.
579  */
580 int
581 softdep_process_worklist(matchmnt)
582 	struct mount *matchmnt;
583 {
584 	struct thread *td = curthread;
585 	int cnt, matchcnt, loopcount;
586 	long starttime;
587 
588 	/*
589 	 * Record the process identifier of our caller so that we can give
590 	 * this process preferential treatment in request_cleanup below.
591 	 */
592 	filesys_syncer = td;
593 	matchcnt = 0;
594 
595 	/*
596 	 * There is no danger of having multiple processes run this
597 	 * code, but we have to single-thread it when softdep_flushfiles()
598 	 * is in operation to get an accurate count of the number of items
599 	 * related to its mount point that are in the list.
600 	 */
601 	if (matchmnt == NULL) {
602 		if (softdep_worklist_busy < 0)
603 			return(-1);
604 		softdep_worklist_busy += 1;
605 	}
606 
607 	/*
608 	 * If requested, try removing inode or removal dependencies.
609 	 */
610 	if (req_clear_inodedeps) {
611 		clear_inodedeps(td);
612 		req_clear_inodedeps -= 1;
613 		wakeup_one(&proc_waiting);
614 	}
615 	if (req_clear_remove) {
616 		clear_remove(td);
617 		req_clear_remove -= 1;
618 		wakeup_one(&proc_waiting);
619 	}
620 	loopcount = 1;
621 	starttime = time_second;
622 	while (num_on_worklist > 0) {
623 		if ((cnt = process_worklist_item(matchmnt, 0)) == -1)
624 			break;
625 		else
626 			matchcnt += cnt;
627 
628 		/*
629 		 * If a umount operation wants to run the worklist
630 		 * accurately, abort.
631 		 */
632 		if (softdep_worklist_req && matchmnt == NULL) {
633 			matchcnt = -1;
634 			break;
635 		}
636 
637 		/*
638 		 * If requested, try removing inode or removal dependencies.
639 		 */
640 		if (req_clear_inodedeps) {
641 			clear_inodedeps(td);
642 			req_clear_inodedeps -= 1;
643 			wakeup_one(&proc_waiting);
644 		}
645 		if (req_clear_remove) {
646 			clear_remove(td);
647 			req_clear_remove -= 1;
648 			wakeup_one(&proc_waiting);
649 		}
650 		/*
651 		 * We do not generally want to stop for buffer space, but if
652 		 * we are really being a buffer hog, we will stop and wait.
653 		 */
654 		if (loopcount++ % 128 == 0)
655 			bwillwrite();
656 		/*
657 		 * Never allow processing to run for more than one
658 		 * second. Otherwise the other syncer tasks may get
659 		 * excessively backlogged.
660 		 */
661 		if (starttime != time_second && matchmnt == NULL) {
662 			matchcnt = -1;
663 			break;
664 		}
665 	}
666 	if (matchmnt == NULL) {
667 		softdep_worklist_busy -= 1;
668 		if (softdep_worklist_req && softdep_worklist_busy == 0)
669 			wakeup(&softdep_worklist_req);
670 	}
671 	return (matchcnt);
672 }
673 
674 /*
675  * Process one item on the worklist.
676  */
677 static int
678 process_worklist_item(matchmnt, flags)
679 	struct mount *matchmnt;
680 	int flags;
681 {
682 	struct worklist *wk;
683 	struct mount *mp;
684 	struct vnode *vp;
685 	int matchcnt = 0;
686 
687 	/*
688 	 * If we are being called because of a process doing a
689 	 * copy-on-write, then it is not safe to write as we may
690 	 * recurse into the copy-on-write routine.
691 	 */
692 	if (curthread->td_proc->p_flag & P_COWINPROGRESS)
693 		return (-1);
694 	ACQUIRE_LOCK(&lk);
695 	/*
696 	 * Normally we just process each item on the worklist in order.
697 	 * However, if we are in a situation where we cannot lock any
698 	 * inodes, we have to skip over any dirrem requests whose
699 	 * vnodes are resident and locked.
700 	 */
701 	vp = NULL;
702 	LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
703 		if (wk->wk_state & INPROGRESS)
704 			continue;
705 		if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
706 			break;
707 		wk->wk_state |= INPROGRESS;
708 		FREE_LOCK(&lk);
709 		VFS_VGET(WK_DIRREM(wk)->dm_mnt, WK_DIRREM(wk)->dm_oldinum,
710 		    LK_NOWAIT | LK_EXCLUSIVE, &vp);
711 		ACQUIRE_LOCK(&lk);
712 		wk->wk_state &= ~INPROGRESS;
713 		if (vp != NULL)
714 			break;
715 	}
716 	if (wk == 0) {
717 		FREE_LOCK(&lk);
718 		return (-1);
719 	}
720 	WORKLIST_REMOVE(wk);
721 	num_on_worklist -= 1;
722 	FREE_LOCK(&lk);
723 	switch (wk->wk_type) {
724 
725 	case D_DIRREM:
726 		/* removal of a directory entry */
727 		mp = WK_DIRREM(wk)->dm_mnt;
728 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
729 			panic("%s: dirrem on suspended filesystem",
730 				"process_worklist_item");
731 		if (mp == matchmnt)
732 			matchcnt += 1;
733 		handle_workitem_remove(WK_DIRREM(wk), vp);
734 		break;
735 
736 	case D_FREEBLKS:
737 		/* releasing blocks and/or fragments from a file */
738 		mp = WK_FREEBLKS(wk)->fb_mnt;
739 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
740 			panic("%s: freeblks on suspended filesystem",
741 				"process_worklist_item");
742 		if (mp == matchmnt)
743 			matchcnt += 1;
744 		handle_workitem_freeblocks(WK_FREEBLKS(wk), flags & LK_NOWAIT);
745 		break;
746 
747 	case D_FREEFRAG:
748 		/* releasing a fragment when replaced as a file grows */
749 		mp = WK_FREEFRAG(wk)->ff_mnt;
750 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
751 			panic("%s: freefrag on suspended filesystem",
752 				"process_worklist_item");
753 		if (mp == matchmnt)
754 			matchcnt += 1;
755 		handle_workitem_freefrag(WK_FREEFRAG(wk));
756 		break;
757 
758 	case D_FREEFILE:
759 		/* releasing an inode when its link count drops to 0 */
760 		mp = WK_FREEFILE(wk)->fx_mnt;
761 		if (vn_write_suspend_wait(NULL, mp, V_NOWAIT))
762 			panic("%s: freefile on suspended filesystem",
763 				"process_worklist_item");
764 		if (mp == matchmnt)
765 			matchcnt += 1;
766 		handle_workitem_freefile(WK_FREEFILE(wk));
767 		break;
768 
769 	default:
770 		panic("%s_process_worklist: Unknown type %s",
771 		    "softdep", TYPENAME(wk->wk_type));
772 		/* NOTREACHED */
773 	}
774 	return (matchcnt);
775 }
776 
777 /*
778  * Move dependencies from one buffer to another.
779  */
780 static void
781 softdep_move_dependencies(oldbp, newbp)
782 	struct buf *oldbp;
783 	struct buf *newbp;
784 {
785 	struct worklist *wk, *wktail;
786 
787 	if (LIST_FIRST(&newbp->b_dep) != NULL)
788 		panic("softdep_move_dependencies: need merge code");
789 	wktail = 0;
790 	ACQUIRE_LOCK(&lk);
791 	while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
792 		LIST_REMOVE(wk, wk_list);
793 		if (wktail == 0)
794 			LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
795 		else
796 			LIST_INSERT_AFTER(wktail, wk, wk_list);
797 		wktail = wk;
798 	}
799 	FREE_LOCK(&lk);
800 }
801 
802 /*
803  * Purge the work list of all items associated with a particular mount point.
804  */
805 int
806 softdep_flushworklist(oldmnt, countp, td)
807 	struct mount *oldmnt;
808 	int *countp;
809 	struct thread *td;
810 {
811 	struct vnode *devvp;
812 	int count, error = 0;
813 
814 	/*
815 	 * Await our turn to clear out the queue, then serialize access.
816 	 */
817 	while (softdep_worklist_busy) {
818 		softdep_worklist_req += 1;
819 		tsleep(&softdep_worklist_req, PRIBIO, "softflush", 0);
820 		softdep_worklist_req -= 1;
821 	}
822 	softdep_worklist_busy = -1;
823 	/*
824 	 * Alternately flush the block device associated with the mount
825 	 * point and process any dependencies that the flushing
826 	 * creates. We continue until no more worklist dependencies
827 	 * are found.
828 	 */
829 	*countp = 0;
830 	devvp = VFSTOUFS(oldmnt)->um_devvp;
831 	while ((count = softdep_process_worklist(oldmnt)) > 0) {
832 		*countp += count;
833 		vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td);
834 		error = VOP_FSYNC(devvp, td->td_ucred, MNT_WAIT, td);
835 		VOP_UNLOCK(devvp, 0, td);
836 		if (error)
837 			break;
838 	}
839 	softdep_worklist_busy = 0;
840 	if (softdep_worklist_req)
841 		wakeup(&softdep_worklist_req);
842 	return (error);
843 }
844 
845 /*
846  * Flush all vnodes and worklist items associated with a specified mount point.
847  */
848 int
849 softdep_flushfiles(oldmnt, flags, td)
850 	struct mount *oldmnt;
851 	int flags;
852 	struct thread *td;
853 {
854 	int error, count, loopcnt;
855 
856 	error = 0;
857 
858 	/*
859 	 * Alternately flush the vnodes associated with the mount
860 	 * point and process any dependencies that the flushing
861 	 * creates. In theory, this loop can happen at most twice,
862 	 * but we give it a few extra just to be sure.
863 	 */
864 	for (loopcnt = 10; loopcnt > 0; loopcnt--) {
865 		/*
866 		 * Do another flush in case any vnodes were brought in
867 		 * as part of the cleanup operations.
868 		 */
869 		if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
870 			break;
871 		if ((error = softdep_flushworklist(oldmnt, &count, td)) != 0 ||
872 		    count == 0)
873 			break;
874 	}
875 	/*
876 	 * If we are unmounting then it is an error to fail. If we
877 	 * are simply trying to downgrade to read-only, then filesystem
878 	 * activity can keep us busy forever, so we just fail with EBUSY.
879 	 */
880 	if (loopcnt == 0) {
881 		if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
882 			panic("softdep_flushfiles: looping");
883 		error = EBUSY;
884 	}
885 	return (error);
886 }
887 
888 /*
889  * Structure hashing.
890  *
891  * There are three types of structures that can be looked up:
892  *	1) pagedep structures identified by mount point, inode number,
893  *	   and logical block.
894  *	2) inodedep structures identified by mount point and inode number.
895  *	3) newblk structures identified by mount point and
896  *	   physical block number.
897  *
898  * The "pagedep" and "inodedep" dependency structures are hashed
899  * separately from the file blocks and inodes to which they correspond.
900  * This separation helps when the in-memory copy of an inode or
901  * file block must be replaced. It also obviates the need to access
902  * an inode or file page when simply updating (or de-allocating)
903  * dependency structures. Lookup of newblk structures is needed to
904  * find newly allocated blocks when trying to associate them with
905  * their allocdirect or allocindir structure.
906  *
907  * The lookup routines optionally create and hash a new instance when
908  * an existing entry is not found.
909  */
910 #define DEPALLOC	0x0001	/* allocate structure if lookup fails */
911 #define NODELAY		0x0002	/* cannot do background work */
912 
913 /*
914  * Structures and routines associated with pagedep caching.
915  */
916 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
917 u_long	pagedep_hash;		/* size of hash table - 1 */
918 #define	PAGEDEP_HASH(mp, inum, lbn) \
919 	(&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
920 	    pagedep_hash])
921 static struct sema pagedep_in_progress;
922 
923 /*
924  * Look up a pagedep. Return 1 if found, 0 if not found or found
925  * when asked to allocate but not associated with any buffer.
926  * If not found, allocate if DEPALLOC flag is passed.
927  * Found or allocated entry is returned in pagedeppp.
928  * This routine must be called with splbio interrupts blocked.
929  */
930 static int
931 pagedep_lookup(ip, lbn, flags, pagedeppp)
932 	struct inode *ip;
933 	ufs_lbn_t lbn;
934 	int flags;
935 	struct pagedep **pagedeppp;
936 {
937 	struct pagedep *pagedep;
938 	struct pagedep_hashhead *pagedephd;
939 	struct mount *mp;
940 	int i;
941 
942 #ifdef DEBUG
943 	if (lk.lkt_held == NOHOLDER)
944 		panic("pagedep_lookup: lock not held");
945 #endif
946 	mp = ITOV(ip)->v_mount;
947 	pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
948 top:
949 	LIST_FOREACH(pagedep, pagedephd, pd_hash)
950 		if (ip->i_number == pagedep->pd_ino &&
951 		    lbn == pagedep->pd_lbn &&
952 		    mp == pagedep->pd_mnt)
953 			break;
954 	if (pagedep) {
955 		*pagedeppp = pagedep;
956 		if ((flags & DEPALLOC) != 0 &&
957 		    (pagedep->pd_state & ONWORKLIST) == 0)
958 			return (0);
959 		return (1);
960 	}
961 	if ((flags & DEPALLOC) == 0) {
962 		*pagedeppp = NULL;
963 		return (0);
964 	}
965 	if (sema_get(&pagedep_in_progress, &lk) == 0) {
966 		ACQUIRE_LOCK(&lk);
967 		goto top;
968 	}
969 	MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
970 		M_SOFTDEP_FLAGS|M_ZERO);
971 	pagedep->pd_list.wk_type = D_PAGEDEP;
972 	pagedep->pd_mnt = mp;
973 	pagedep->pd_ino = ip->i_number;
974 	pagedep->pd_lbn = lbn;
975 	LIST_INIT(&pagedep->pd_dirremhd);
976 	LIST_INIT(&pagedep->pd_pendinghd);
977 	for (i = 0; i < DAHASHSZ; i++)
978 		LIST_INIT(&pagedep->pd_diraddhd[i]);
979 	ACQUIRE_LOCK(&lk);
980 	LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
981 	sema_release(&pagedep_in_progress);
982 	*pagedeppp = pagedep;
983 	return (0);
984 }
985 
986 /*
987  * Structures and routines associated with inodedep caching.
988  */
989 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
990 static u_long	inodedep_hash;	/* size of hash table - 1 */
991 static long	num_inodedep;	/* number of inodedep allocated */
992 #define	INODEDEP_HASH(fs, inum) \
993       (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
994 static struct sema inodedep_in_progress;
995 
996 /*
997  * Look up an inodedep. Return 1 if found, 0 if not found.
998  * If not found, allocate if DEPALLOC flag is passed.
999  * Found or allocated entry is returned in inodedeppp.
1000  * This routine must be called with splbio interrupts blocked.
1001  */
1002 static int
1003 inodedep_lookup(fs, inum, flags, inodedeppp)
1004 	struct fs *fs;
1005 	ino_t inum;
1006 	int flags;
1007 	struct inodedep **inodedeppp;
1008 {
1009 	struct inodedep *inodedep;
1010 	struct inodedep_hashhead *inodedephd;
1011 	int firsttry;
1012 
1013 #ifdef DEBUG
1014 	if (lk.lkt_held == NOHOLDER)
1015 		panic("inodedep_lookup: lock not held");
1016 #endif
1017 	firsttry = 1;
1018 	inodedephd = INODEDEP_HASH(fs, inum);
1019 top:
1020 	LIST_FOREACH(inodedep, inodedephd, id_hash)
1021 		if (inum == inodedep->id_ino && fs == inodedep->id_fs)
1022 			break;
1023 	if (inodedep) {
1024 		*inodedeppp = inodedep;
1025 		return (1);
1026 	}
1027 	if ((flags & DEPALLOC) == 0) {
1028 		*inodedeppp = NULL;
1029 		return (0);
1030 	}
1031 	/*
1032 	 * If we are over our limit, try to improve the situation.
1033 	 */
1034 	if (num_inodedep > max_softdeps && firsttry && (flags & NODELAY) == 0 &&
1035 	    request_cleanup(FLUSH_INODES, 1)) {
1036 		firsttry = 0;
1037 		goto top;
1038 	}
1039 	if (sema_get(&inodedep_in_progress, &lk) == 0) {
1040 		ACQUIRE_LOCK(&lk);
1041 		goto top;
1042 	}
1043 	num_inodedep += 1;
1044 	MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1045 		M_INODEDEP, M_SOFTDEP_FLAGS);
1046 	inodedep->id_list.wk_type = D_INODEDEP;
1047 	inodedep->id_fs = fs;
1048 	inodedep->id_ino = inum;
1049 	inodedep->id_state = ALLCOMPLETE;
1050 	inodedep->id_nlinkdelta = 0;
1051 	inodedep->id_savedino1 = NULL;
1052 	inodedep->id_savedsize = -1;
1053 	inodedep->id_savedextsize = -1;
1054 	inodedep->id_buf = NULL;
1055 	LIST_INIT(&inodedep->id_pendinghd);
1056 	LIST_INIT(&inodedep->id_inowait);
1057 	LIST_INIT(&inodedep->id_bufwait);
1058 	TAILQ_INIT(&inodedep->id_inoupdt);
1059 	TAILQ_INIT(&inodedep->id_newinoupdt);
1060 	TAILQ_INIT(&inodedep->id_extupdt);
1061 	TAILQ_INIT(&inodedep->id_newextupdt);
1062 	ACQUIRE_LOCK(&lk);
1063 	LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1064 	sema_release(&inodedep_in_progress);
1065 	*inodedeppp = inodedep;
1066 	return (0);
1067 }
1068 
1069 /*
1070  * Structures and routines associated with newblk caching.
1071  */
1072 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1073 u_long	newblk_hash;		/* size of hash table - 1 */
1074 #define	NEWBLK_HASH(fs, inum) \
1075 	(&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1076 static struct sema newblk_in_progress;
1077 
1078 /*
1079  * Look up a newblk. Return 1 if found, 0 if not found.
1080  * If not found, allocate if DEPALLOC flag is passed.
1081  * Found or allocated entry is returned in newblkpp.
1082  */
1083 static int
1084 newblk_lookup(fs, newblkno, flags, newblkpp)
1085 	struct fs *fs;
1086 	ufs2_daddr_t newblkno;
1087 	int flags;
1088 	struct newblk **newblkpp;
1089 {
1090 	struct newblk *newblk;
1091 	struct newblk_hashhead *newblkhd;
1092 
1093 	newblkhd = NEWBLK_HASH(fs, newblkno);
1094 top:
1095 	LIST_FOREACH(newblk, newblkhd, nb_hash)
1096 		if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1097 			break;
1098 	if (newblk) {
1099 		*newblkpp = newblk;
1100 		return (1);
1101 	}
1102 	if ((flags & DEPALLOC) == 0) {
1103 		*newblkpp = NULL;
1104 		return (0);
1105 	}
1106 	if (sema_get(&newblk_in_progress, 0) == 0)
1107 		goto top;
1108 	MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1109 		M_NEWBLK, M_SOFTDEP_FLAGS);
1110 	newblk->nb_state = 0;
1111 	newblk->nb_fs = fs;
1112 	newblk->nb_newblkno = newblkno;
1113 	LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1114 	sema_release(&newblk_in_progress);
1115 	*newblkpp = newblk;
1116 	return (0);
1117 }
1118 
1119 /*
1120  * Executed during filesystem system initialization before
1121  * mounting any filesystems.
1122  */
1123 void
1124 softdep_initialize()
1125 {
1126 
1127 	LIST_INIT(&mkdirlisthd);
1128 	LIST_INIT(&softdep_workitem_pending);
1129 	max_softdeps = desiredvnodes * 4;
1130 	pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1131 	    &pagedep_hash);
1132 	sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0);
1133 	inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1134 	sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0);
1135 	newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1136 	sema_init(&newblk_in_progress, "newblk", PRIBIO, 0);
1137 
1138 	/* hooks through which the main kernel code calls us */
1139 	softdep_process_worklist_hook = softdep_process_worklist;
1140 	softdep_fsync_hook = softdep_fsync;
1141 
1142 	/* initialise bioops hack */
1143 	bioops.io_start = softdep_disk_io_initiation;
1144 	bioops.io_complete = softdep_disk_write_complete;
1145 	bioops.io_deallocate = softdep_deallocate_dependencies;
1146 	bioops.io_movedeps = softdep_move_dependencies;
1147 	bioops.io_countdeps = softdep_count_dependencies;
1148 }
1149 
1150 /*
1151  * Executed after all filesystems have been unmounted during
1152  * filesystem module unload.
1153  */
1154 void
1155 softdep_uninitialize()
1156 {
1157 
1158 	softdep_process_worklist_hook = NULL;
1159 	softdep_fsync_hook = NULL;
1160 	hashdestroy(pagedep_hashtbl, M_PAGEDEP, pagedep_hash);
1161 	hashdestroy(inodedep_hashtbl, M_INODEDEP, inodedep_hash);
1162 	hashdestroy(newblk_hashtbl, M_NEWBLK, newblk_hash);
1163 }
1164 
1165 /*
1166  * Called at mount time to notify the dependency code that a
1167  * filesystem wishes to use it.
1168  */
1169 int
1170 softdep_mount(devvp, mp, fs, cred)
1171 	struct vnode *devvp;
1172 	struct mount *mp;
1173 	struct fs *fs;
1174 	struct ucred *cred;
1175 {
1176 	struct csum_total cstotal;
1177 	struct cg *cgp;
1178 	struct buf *bp;
1179 	int error, cyl;
1180 
1181 	mp->mnt_flag &= ~MNT_ASYNC;
1182 	mp->mnt_flag |= MNT_SOFTDEP;
1183 	/*
1184 	 * When doing soft updates, the counters in the
1185 	 * superblock may have gotten out of sync, so we have
1186 	 * to scan the cylinder groups and recalculate them.
1187 	 */
1188 	if (fs->fs_clean != 0)
1189 		return (0);
1190 	bzero(&cstotal, sizeof cstotal);
1191 	for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1192 		if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1193 		    fs->fs_cgsize, cred, &bp)) != 0) {
1194 			brelse(bp);
1195 			return (error);
1196 		}
1197 		cgp = (struct cg *)bp->b_data;
1198 		cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1199 		cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1200 		cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1201 		cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1202 		fs->fs_cs(fs, cyl) = cgp->cg_cs;
1203 		brelse(bp);
1204 	}
1205 #ifdef DEBUG
1206 	if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1207 		printf("%s: superblock summary recomputed\n", fs->fs_fsmnt);
1208 #endif
1209 	bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1210 	return (0);
1211 }
1212 
1213 /*
1214  * Protecting the freemaps (or bitmaps).
1215  *
1216  * To eliminate the need to execute fsck before mounting a filesystem
1217  * after a power failure, one must (conservatively) guarantee that the
1218  * on-disk copy of the bitmaps never indicate that a live inode or block is
1219  * free.  So, when a block or inode is allocated, the bitmap should be
1220  * updated (on disk) before any new pointers.  When a block or inode is
1221  * freed, the bitmap should not be updated until all pointers have been
1222  * reset.  The latter dependency is handled by the delayed de-allocation
1223  * approach described below for block and inode de-allocation.  The former
1224  * dependency is handled by calling the following procedure when a block or
1225  * inode is allocated. When an inode is allocated an "inodedep" is created
1226  * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1227  * Each "inodedep" is also inserted into the hash indexing structure so
1228  * that any additional link additions can be made dependent on the inode
1229  * allocation.
1230  *
1231  * The ufs filesystem maintains a number of free block counts (e.g., per
1232  * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1233  * in addition to the bitmaps.  These counts are used to improve efficiency
1234  * during allocation and therefore must be consistent with the bitmaps.
1235  * There is no convenient way to guarantee post-crash consistency of these
1236  * counts with simple update ordering, for two main reasons: (1) The counts
1237  * and bitmaps for a single cylinder group block are not in the same disk
1238  * sector.  If a disk write is interrupted (e.g., by power failure), one may
1239  * be written and the other not.  (2) Some of the counts are located in the
1240  * superblock rather than the cylinder group block. So, we focus our soft
1241  * updates implementation on protecting the bitmaps. When mounting a
1242  * filesystem, we recompute the auxiliary counts from the bitmaps.
1243  */
1244 
1245 /*
1246  * Called just after updating the cylinder group block to allocate an inode.
1247  */
1248 void
1249 softdep_setup_inomapdep(bp, ip, newinum)
1250 	struct buf *bp;		/* buffer for cylgroup block with inode map */
1251 	struct inode *ip;	/* inode related to allocation */
1252 	ino_t newinum;		/* new inode number being allocated */
1253 {
1254 	struct inodedep *inodedep;
1255 	struct bmsafemap *bmsafemap;
1256 
1257 	/*
1258 	 * Create a dependency for the newly allocated inode.
1259 	 * Panic if it already exists as something is seriously wrong.
1260 	 * Otherwise add it to the dependency list for the buffer holding
1261 	 * the cylinder group map from which it was allocated.
1262 	 */
1263 	ACQUIRE_LOCK(&lk);
1264 	if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1265 		FREE_LOCK(&lk);
1266 		panic("softdep_setup_inomapdep: found inode");
1267 	}
1268 	inodedep->id_buf = bp;
1269 	inodedep->id_state &= ~DEPCOMPLETE;
1270 	bmsafemap = bmsafemap_lookup(bp);
1271 	LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1272 	FREE_LOCK(&lk);
1273 }
1274 
1275 /*
1276  * Called just after updating the cylinder group block to
1277  * allocate block or fragment.
1278  */
1279 void
1280 softdep_setup_blkmapdep(bp, fs, newblkno)
1281 	struct buf *bp;		/* buffer for cylgroup block with block map */
1282 	struct fs *fs;		/* filesystem doing allocation */
1283 	ufs2_daddr_t newblkno;	/* number of newly allocated block */
1284 {
1285 	struct newblk *newblk;
1286 	struct bmsafemap *bmsafemap;
1287 
1288 	/*
1289 	 * Create a dependency for the newly allocated block.
1290 	 * Add it to the dependency list for the buffer holding
1291 	 * the cylinder group map from which it was allocated.
1292 	 */
1293 	if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1294 		panic("softdep_setup_blkmapdep: found block");
1295 	ACQUIRE_LOCK(&lk);
1296 	newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1297 	LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1298 	FREE_LOCK(&lk);
1299 }
1300 
1301 /*
1302  * Find the bmsafemap associated with a cylinder group buffer.
1303  * If none exists, create one. The buffer must be locked when
1304  * this routine is called and this routine must be called with
1305  * splbio interrupts blocked.
1306  */
1307 static struct bmsafemap *
1308 bmsafemap_lookup(bp)
1309 	struct buf *bp;
1310 {
1311 	struct bmsafemap *bmsafemap;
1312 	struct worklist *wk;
1313 
1314 #ifdef DEBUG
1315 	if (lk.lkt_held == NOHOLDER)
1316 		panic("bmsafemap_lookup: lock not held");
1317 #endif
1318 	LIST_FOREACH(wk, &bp->b_dep, wk_list)
1319 		if (wk->wk_type == D_BMSAFEMAP)
1320 			return (WK_BMSAFEMAP(wk));
1321 	FREE_LOCK(&lk);
1322 	MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1323 		M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1324 	bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1325 	bmsafemap->sm_list.wk_state = 0;
1326 	bmsafemap->sm_buf = bp;
1327 	LIST_INIT(&bmsafemap->sm_allocdirecthd);
1328 	LIST_INIT(&bmsafemap->sm_allocindirhd);
1329 	LIST_INIT(&bmsafemap->sm_inodedephd);
1330 	LIST_INIT(&bmsafemap->sm_newblkhd);
1331 	ACQUIRE_LOCK(&lk);
1332 	WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1333 	return (bmsafemap);
1334 }
1335 
1336 /*
1337  * Direct block allocation dependencies.
1338  *
1339  * When a new block is allocated, the corresponding disk locations must be
1340  * initialized (with zeros or new data) before the on-disk inode points to
1341  * them.  Also, the freemap from which the block was allocated must be
1342  * updated (on disk) before the inode's pointer. These two dependencies are
1343  * independent of each other and are needed for all file blocks and indirect
1344  * blocks that are pointed to directly by the inode.  Just before the
1345  * "in-core" version of the inode is updated with a newly allocated block
1346  * number, a procedure (below) is called to setup allocation dependency
1347  * structures.  These structures are removed when the corresponding
1348  * dependencies are satisfied or when the block allocation becomes obsolete
1349  * (i.e., the file is deleted, the block is de-allocated, or the block is a
1350  * fragment that gets upgraded).  All of these cases are handled in
1351  * procedures described later.
1352  *
1353  * When a file extension causes a fragment to be upgraded, either to a larger
1354  * fragment or to a full block, the on-disk location may change (if the
1355  * previous fragment could not simply be extended). In this case, the old
1356  * fragment must be de-allocated, but not until after the inode's pointer has
1357  * been updated. In most cases, this is handled by later procedures, which
1358  * will construct a "freefrag" structure to be added to the workitem queue
1359  * when the inode update is complete (or obsolete).  The main exception to
1360  * this is when an allocation occurs while a pending allocation dependency
1361  * (for the same block pointer) remains.  This case is handled in the main
1362  * allocation dependency setup procedure by immediately freeing the
1363  * unreferenced fragments.
1364  */
1365 void
1366 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1367 	struct inode *ip;	/* inode to which block is being added */
1368 	ufs_lbn_t lbn;		/* block pointer within inode */
1369 	ufs2_daddr_t newblkno;	/* disk block number being added */
1370 	ufs2_daddr_t oldblkno;	/* previous block number, 0 unless frag */
1371 	long newsize;		/* size of new block */
1372 	long oldsize;		/* size of new block */
1373 	struct buf *bp;		/* bp for allocated block */
1374 {
1375 	struct allocdirect *adp, *oldadp;
1376 	struct allocdirectlst *adphead;
1377 	struct bmsafemap *bmsafemap;
1378 	struct inodedep *inodedep;
1379 	struct pagedep *pagedep;
1380 	struct newblk *newblk;
1381 
1382 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1383 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1384 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1385 	adp->ad_lbn = lbn;
1386 	adp->ad_newblkno = newblkno;
1387 	adp->ad_oldblkno = oldblkno;
1388 	adp->ad_newsize = newsize;
1389 	adp->ad_oldsize = oldsize;
1390 	adp->ad_state = ATTACHED;
1391 	LIST_INIT(&adp->ad_newdirblk);
1392 	if (newblkno == oldblkno)
1393 		adp->ad_freefrag = NULL;
1394 	else
1395 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1396 
1397 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1398 		panic("softdep_setup_allocdirect: lost block");
1399 
1400 	ACQUIRE_LOCK(&lk);
1401 	inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1402 	adp->ad_inodedep = inodedep;
1403 
1404 	if (newblk->nb_state == DEPCOMPLETE) {
1405 		adp->ad_state |= DEPCOMPLETE;
1406 		adp->ad_buf = NULL;
1407 	} else {
1408 		bmsafemap = newblk->nb_bmsafemap;
1409 		adp->ad_buf = bmsafemap->sm_buf;
1410 		LIST_REMOVE(newblk, nb_deps);
1411 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1412 	}
1413 	LIST_REMOVE(newblk, nb_hash);
1414 	FREE(newblk, M_NEWBLK);
1415 
1416 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1417 	if (lbn >= NDADDR) {
1418 		/* allocating an indirect block */
1419 		if (oldblkno != 0) {
1420 			FREE_LOCK(&lk);
1421 			panic("softdep_setup_allocdirect: non-zero indir");
1422 		}
1423 	} else {
1424 		/*
1425 		 * Allocating a direct block.
1426 		 *
1427 		 * If we are allocating a directory block, then we must
1428 		 * allocate an associated pagedep to track additions and
1429 		 * deletions.
1430 		 */
1431 		if ((ip->i_mode & IFMT) == IFDIR &&
1432 		    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1433 			WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1434 	}
1435 	/*
1436 	 * The list of allocdirects must be kept in sorted and ascending
1437 	 * order so that the rollback routines can quickly determine the
1438 	 * first uncommitted block (the size of the file stored on disk
1439 	 * ends at the end of the lowest committed fragment, or if there
1440 	 * are no fragments, at the end of the highest committed block).
1441 	 * Since files generally grow, the typical case is that the new
1442 	 * block is to be added at the end of the list. We speed this
1443 	 * special case by checking against the last allocdirect in the
1444 	 * list before laboriously traversing the list looking for the
1445 	 * insertion point.
1446 	 */
1447 	adphead = &inodedep->id_newinoupdt;
1448 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1449 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1450 		/* insert at end of list */
1451 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1452 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1453 			allocdirect_merge(adphead, adp, oldadp);
1454 		FREE_LOCK(&lk);
1455 		return;
1456 	}
1457 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1458 		if (oldadp->ad_lbn >= lbn)
1459 			break;
1460 	}
1461 	if (oldadp == NULL) {
1462 		FREE_LOCK(&lk);
1463 		panic("softdep_setup_allocdirect: lost entry");
1464 	}
1465 	/* insert in middle of list */
1466 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1467 	if (oldadp->ad_lbn == lbn)
1468 		allocdirect_merge(adphead, adp, oldadp);
1469 	FREE_LOCK(&lk);
1470 }
1471 
1472 /*
1473  * Replace an old allocdirect dependency with a newer one.
1474  * This routine must be called with splbio interrupts blocked.
1475  */
1476 static void
1477 allocdirect_merge(adphead, newadp, oldadp)
1478 	struct allocdirectlst *adphead;	/* head of list holding allocdirects */
1479 	struct allocdirect *newadp;	/* allocdirect being added */
1480 	struct allocdirect *oldadp;	/* existing allocdirect being checked */
1481 {
1482 	struct worklist *wk;
1483 	struct freefrag *freefrag;
1484 	struct newdirblk *newdirblk;
1485 
1486 #ifdef DEBUG
1487 	if (lk.lkt_held == NOHOLDER)
1488 		panic("allocdirect_merge: lock not held");
1489 #endif
1490 	if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1491 	    newadp->ad_oldsize != oldadp->ad_newsize ||
1492 	    newadp->ad_lbn >= NDADDR) {
1493 		FREE_LOCK(&lk);
1494 		panic("%s %jd != new %jd || old size %ld != new %ld",
1495 		    "allocdirect_merge: old blkno",
1496 		    (intmax_t)newadp->ad_oldblkno,
1497 		    (intmax_t)oldadp->ad_newblkno,
1498 		    newadp->ad_oldsize, oldadp->ad_newsize);
1499 	}
1500 	newadp->ad_oldblkno = oldadp->ad_oldblkno;
1501 	newadp->ad_oldsize = oldadp->ad_oldsize;
1502 	/*
1503 	 * If the old dependency had a fragment to free or had never
1504 	 * previously had a block allocated, then the new dependency
1505 	 * can immediately post its freefrag and adopt the old freefrag.
1506 	 * This action is done by swapping the freefrag dependencies.
1507 	 * The new dependency gains the old one's freefrag, and the
1508 	 * old one gets the new one and then immediately puts it on
1509 	 * the worklist when it is freed by free_allocdirect. It is
1510 	 * not possible to do this swap when the old dependency had a
1511 	 * non-zero size but no previous fragment to free. This condition
1512 	 * arises when the new block is an extension of the old block.
1513 	 * Here, the first part of the fragment allocated to the new
1514 	 * dependency is part of the block currently claimed on disk by
1515 	 * the old dependency, so cannot legitimately be freed until the
1516 	 * conditions for the new dependency are fulfilled.
1517 	 */
1518 	if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1519 		freefrag = newadp->ad_freefrag;
1520 		newadp->ad_freefrag = oldadp->ad_freefrag;
1521 		oldadp->ad_freefrag = freefrag;
1522 	}
1523 	/*
1524 	 * If we are tracking a new directory-block allocation,
1525 	 * move it from the old allocdirect to the new allocdirect.
1526 	 */
1527 	if ((wk = LIST_FIRST(&oldadp->ad_newdirblk)) != NULL) {
1528 		newdirblk = WK_NEWDIRBLK(wk);
1529 		WORKLIST_REMOVE(&newdirblk->db_list);
1530 		if (LIST_FIRST(&oldadp->ad_newdirblk) != NULL)
1531 			panic("allocdirect_merge: extra newdirblk");
1532 		WORKLIST_INSERT(&newadp->ad_newdirblk, &newdirblk->db_list);
1533 	}
1534 	free_allocdirect(adphead, oldadp, 0);
1535 }
1536 
1537 /*
1538  * Allocate a new freefrag structure if needed.
1539  */
1540 static struct freefrag *
1541 newfreefrag(ip, blkno, size)
1542 	struct inode *ip;
1543 	ufs2_daddr_t blkno;
1544 	long size;
1545 {
1546 	struct freefrag *freefrag;
1547 	struct fs *fs;
1548 
1549 	if (blkno == 0)
1550 		return (NULL);
1551 	fs = ip->i_fs;
1552 	if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1553 		panic("newfreefrag: frag size");
1554 	MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1555 		M_FREEFRAG, M_SOFTDEP_FLAGS);
1556 	freefrag->ff_list.wk_type = D_FREEFRAG;
1557 	freefrag->ff_state = 0;
1558 	freefrag->ff_inum = ip->i_number;
1559 	freefrag->ff_mnt = ITOV(ip)->v_mount;
1560 	freefrag->ff_blkno = blkno;
1561 	freefrag->ff_fragsize = size;
1562 	return (freefrag);
1563 }
1564 
1565 /*
1566  * This workitem de-allocates fragments that were replaced during
1567  * file block allocation.
1568  */
1569 static void
1570 handle_workitem_freefrag(freefrag)
1571 	struct freefrag *freefrag;
1572 {
1573 	struct ufsmount *ump = VFSTOUFS(freefrag->ff_mnt);
1574 
1575 	ffs_blkfree(ump->um_fs, ump->um_devvp, freefrag->ff_blkno,
1576 	    freefrag->ff_fragsize, freefrag->ff_inum);
1577 	FREE(freefrag, M_FREEFRAG);
1578 }
1579 
1580 /*
1581  * Set up a dependency structure for an external attributes data block.
1582  * This routine follows much of the structure of softdep_setup_allocdirect.
1583  * See the description of softdep_setup_allocdirect above for details.
1584  */
1585 void
1586 softdep_setup_allocext(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1587 	struct inode *ip;
1588 	ufs_lbn_t lbn;
1589 	ufs2_daddr_t newblkno;
1590 	ufs2_daddr_t oldblkno;
1591 	long newsize;
1592 	long oldsize;
1593 	struct buf *bp;
1594 {
1595 	struct allocdirect *adp, *oldadp;
1596 	struct allocdirectlst *adphead;
1597 	struct bmsafemap *bmsafemap;
1598 	struct inodedep *inodedep;
1599 	struct newblk *newblk;
1600 
1601 	MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1602 		M_ALLOCDIRECT, M_SOFTDEP_FLAGS|M_ZERO);
1603 	adp->ad_list.wk_type = D_ALLOCDIRECT;
1604 	adp->ad_lbn = lbn;
1605 	adp->ad_newblkno = newblkno;
1606 	adp->ad_oldblkno = oldblkno;
1607 	adp->ad_newsize = newsize;
1608 	adp->ad_oldsize = oldsize;
1609 	adp->ad_state = ATTACHED | EXTDATA;
1610 	LIST_INIT(&adp->ad_newdirblk);
1611 	if (newblkno == oldblkno)
1612 		adp->ad_freefrag = NULL;
1613 	else
1614 		adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1615 
1616 	if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1617 		panic("softdep_setup_allocext: lost block");
1618 
1619 	ACQUIRE_LOCK(&lk);
1620 	inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1621 	adp->ad_inodedep = inodedep;
1622 
1623 	if (newblk->nb_state == DEPCOMPLETE) {
1624 		adp->ad_state |= DEPCOMPLETE;
1625 		adp->ad_buf = NULL;
1626 	} else {
1627 		bmsafemap = newblk->nb_bmsafemap;
1628 		adp->ad_buf = bmsafemap->sm_buf;
1629 		LIST_REMOVE(newblk, nb_deps);
1630 		LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1631 	}
1632 	LIST_REMOVE(newblk, nb_hash);
1633 	FREE(newblk, M_NEWBLK);
1634 
1635 	WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1636 	if (lbn >= NXADDR) {
1637 		FREE_LOCK(&lk);
1638 		panic("softdep_setup_allocext: lbn %lld > NXADDR",
1639 		    (long long)lbn);
1640 	}
1641 	/*
1642 	 * The list of allocdirects must be kept in sorted and ascending
1643 	 * order so that the rollback routines can quickly determine the
1644 	 * first uncommitted block (the size of the file stored on disk
1645 	 * ends at the end of the lowest committed fragment, or if there
1646 	 * are no fragments, at the end of the highest committed block).
1647 	 * Since files generally grow, the typical case is that the new
1648 	 * block is to be added at the end of the list. We speed this
1649 	 * special case by checking against the last allocdirect in the
1650 	 * list before laboriously traversing the list looking for the
1651 	 * insertion point.
1652 	 */
1653 	adphead = &inodedep->id_newextupdt;
1654 	oldadp = TAILQ_LAST(adphead, allocdirectlst);
1655 	if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1656 		/* insert at end of list */
1657 		TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1658 		if (oldadp != NULL && oldadp->ad_lbn == lbn)
1659 			allocdirect_merge(adphead, adp, oldadp);
1660 		FREE_LOCK(&lk);
1661 		return;
1662 	}
1663 	TAILQ_FOREACH(oldadp, adphead, ad_next) {
1664 		if (oldadp->ad_lbn >= lbn)
1665 			break;
1666 	}
1667 	if (oldadp == NULL) {
1668 		FREE_LOCK(&lk);
1669 		panic("softdep_setup_allocext: lost entry");
1670 	}
1671 	/* insert in middle of list */
1672 	TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1673 	if (oldadp->ad_lbn == lbn)
1674 		allocdirect_merge(adphead, adp, oldadp);
1675 	FREE_LOCK(&lk);
1676 }
1677 
1678 /*
1679  * Indirect block allocation dependencies.
1680  *
1681  * The same dependencies that exist for a direct block also exist when
1682  * a new block is allocated and pointed to by an entry in a block of
1683  * indirect pointers. The undo/redo states described above are also
1684  * used here. Because an indirect block contains many pointers that
1685  * may have dependencies, a second copy of the entire in-memory indirect
1686  * block is kept. The buffer cache copy is always completely up-to-date.
1687  * The second copy, which is used only as a source for disk writes,
1688  * contains only the safe pointers (i.e., those that have no remaining
1689  * update dependencies). The second copy is freed when all pointers
1690  * are safe. The cache is not allowed to replace indirect blocks with
1691  * pending update dependencies. If a buffer containing an indirect
1692  * block with dependencies is written, these routines will mark it
1693  * dirty again. It can only be successfully written once all the
1694  * dependencies are removed. The ffs_fsync routine in conjunction with
1695  * softdep_sync_metadata work together to get all the dependencies
1696  * removed so that a file can be successfully written to disk. Three
1697  * procedures are used when setting up indirect block pointer
1698  * dependencies. The division is necessary because of the organization
1699  * of the "balloc" routine and because of the distinction between file
1700  * pages and file metadata blocks.
1701  */
1702 
1703 /*
1704  * Allocate a new allocindir structure.
1705  */
1706 static struct allocindir *
1707 newallocindir(ip, ptrno, newblkno, oldblkno)
1708 	struct inode *ip;	/* inode for file being extended */
1709 	int ptrno;		/* offset of pointer in indirect block */
1710 	ufs2_daddr_t newblkno;	/* disk block number being added */
1711 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1712 {
1713 	struct allocindir *aip;
1714 
1715 	MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1716 		M_ALLOCINDIR, M_SOFTDEP_FLAGS|M_ZERO);
1717 	aip->ai_list.wk_type = D_ALLOCINDIR;
1718 	aip->ai_state = ATTACHED;
1719 	aip->ai_offset = ptrno;
1720 	aip->ai_newblkno = newblkno;
1721 	aip->ai_oldblkno = oldblkno;
1722 	aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1723 	return (aip);
1724 }
1725 
1726 /*
1727  * Called just before setting an indirect block pointer
1728  * to a newly allocated file page.
1729  */
1730 void
1731 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1732 	struct inode *ip;	/* inode for file being extended */
1733 	ufs_lbn_t lbn;		/* allocated block number within file */
1734 	struct buf *bp;		/* buffer with indirect blk referencing page */
1735 	int ptrno;		/* offset of pointer in indirect block */
1736 	ufs2_daddr_t newblkno;	/* disk block number being added */
1737 	ufs2_daddr_t oldblkno;	/* previous block number, 0 if none */
1738 	struct buf *nbp;	/* buffer holding allocated page */
1739 {
1740 	struct allocindir *aip;
1741 	struct pagedep *pagedep;
1742 
1743 	aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1744 	ACQUIRE_LOCK(&lk);
1745 	/*
1746 	 * If we are allocating a directory page, then we must
1747 	 * allocate an associated pagedep to track additions and
1748 	 * deletions.
1749 	 */
1750 	if ((ip->i_mode & IFMT) == IFDIR &&
1751 	    pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1752 		WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1753 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1754 	FREE_LOCK(&lk);
1755 	setup_allocindir_phase2(bp, ip, aip);
1756 }
1757 
1758 /*
1759  * Called just before setting an indirect block pointer to a
1760  * newly allocated indirect block.
1761  */
1762 void
1763 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1764 	struct buf *nbp;	/* newly allocated indirect block */
1765 	struct inode *ip;	/* inode for file being extended */
1766 	struct buf *bp;		/* indirect block referencing allocated block */
1767 	int ptrno;		/* offset of pointer in indirect block */
1768 	ufs2_daddr_t newblkno;	/* disk block number being added */
1769 {
1770 	struct allocindir *aip;
1771 
1772 	aip = newallocindir(ip, ptrno, newblkno, 0);
1773 	ACQUIRE_LOCK(&lk);
1774 	WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1775 	FREE_LOCK(&lk);
1776 	setup_allocindir_phase2(bp, ip, aip);
1777 }
1778 
1779 /*
1780  * Called to finish the allocation of the "aip" allocated
1781  * by one of the two routines above.
1782  */
1783 static void
1784 setup_allocindir_phase2(bp, ip, aip)
1785 	struct buf *bp;		/* in-memory copy of the indirect block */
1786 	struct inode *ip;	/* inode for file being extended */
1787 	struct allocindir *aip;	/* allocindir allocated by the above routines */
1788 {
1789 	struct worklist *wk;
1790 	struct indirdep *indirdep, *newindirdep;
1791 	struct bmsafemap *bmsafemap;
1792 	struct allocindir *oldaip;
1793 	struct freefrag *freefrag;
1794 	struct newblk *newblk;
1795 	ufs2_daddr_t blkno;
1796 
1797 	if (bp->b_lblkno >= 0)
1798 		panic("setup_allocindir_phase2: not indir blk");
1799 	for (indirdep = NULL, newindirdep = NULL; ; ) {
1800 		ACQUIRE_LOCK(&lk);
1801 		LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1802 			if (wk->wk_type != D_INDIRDEP)
1803 				continue;
1804 			indirdep = WK_INDIRDEP(wk);
1805 			break;
1806 		}
1807 		if (indirdep == NULL && newindirdep) {
1808 			indirdep = newindirdep;
1809 			WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1810 			newindirdep = NULL;
1811 		}
1812 		FREE_LOCK(&lk);
1813 		if (indirdep) {
1814 			if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1815 			    &newblk) == 0)
1816 				panic("setup_allocindir: lost block");
1817 			ACQUIRE_LOCK(&lk);
1818 			if (newblk->nb_state == DEPCOMPLETE) {
1819 				aip->ai_state |= DEPCOMPLETE;
1820 				aip->ai_buf = NULL;
1821 			} else {
1822 				bmsafemap = newblk->nb_bmsafemap;
1823 				aip->ai_buf = bmsafemap->sm_buf;
1824 				LIST_REMOVE(newblk, nb_deps);
1825 				LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1826 				    aip, ai_deps);
1827 			}
1828 			LIST_REMOVE(newblk, nb_hash);
1829 			FREE(newblk, M_NEWBLK);
1830 			aip->ai_indirdep = indirdep;
1831 			/*
1832 			 * Check to see if there is an existing dependency
1833 			 * for this block. If there is, merge the old
1834 			 * dependency into the new one.
1835 			 */
1836 			if (aip->ai_oldblkno == 0)
1837 				oldaip = NULL;
1838 			else
1839 
1840 				LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1841 					if (oldaip->ai_offset == aip->ai_offset)
1842 						break;
1843 			freefrag = NULL;
1844 			if (oldaip != NULL) {
1845 				if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1846 					FREE_LOCK(&lk);
1847 					panic("setup_allocindir_phase2: blkno");
1848 				}
1849 				aip->ai_oldblkno = oldaip->ai_oldblkno;
1850 				freefrag = aip->ai_freefrag;
1851 				aip->ai_freefrag = oldaip->ai_freefrag;
1852 				oldaip->ai_freefrag = NULL;
1853 				free_allocindir(oldaip, NULL);
1854 			}
1855 			LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1856 			if (ip->i_ump->um_fstype == UFS1)
1857 				((ufs1_daddr_t *)indirdep->ir_savebp->b_data)
1858 				    [aip->ai_offset] = aip->ai_oldblkno;
1859 			else
1860 				((ufs2_daddr_t *)indirdep->ir_savebp->b_data)
1861 				    [aip->ai_offset] = aip->ai_oldblkno;
1862 			FREE_LOCK(&lk);
1863 			if (freefrag != NULL)
1864 				handle_workitem_freefrag(freefrag);
1865 		}
1866 		if (newindirdep) {
1867 			brelse(newindirdep->ir_savebp);
1868 			WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1869 		}
1870 		if (indirdep)
1871 			break;
1872 		MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1873 			M_INDIRDEP, M_SOFTDEP_FLAGS);
1874 		newindirdep->ir_list.wk_type = D_INDIRDEP;
1875 		newindirdep->ir_state = ATTACHED;
1876 		if (ip->i_ump->um_fstype == UFS1)
1877 			newindirdep->ir_state |= UFS1FMT;
1878 		LIST_INIT(&newindirdep->ir_deplisthd);
1879 		LIST_INIT(&newindirdep->ir_donehd);
1880 		if (bp->b_blkno == bp->b_lblkno) {
1881 			ufs_bmaparray(bp->b_vp, bp->b_lblkno, &blkno, bp,
1882 			    NULL, NULL);
1883 			bp->b_blkno = blkno;
1884 		}
1885 		newindirdep->ir_savebp =
1886 		    getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0);
1887 		BUF_KERNPROC(newindirdep->ir_savebp);
1888 		bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1889 	}
1890 }
1891 
1892 /*
1893  * Block de-allocation dependencies.
1894  *
1895  * When blocks are de-allocated, the on-disk pointers must be nullified before
1896  * the blocks are made available for use by other files.  (The true
1897  * requirement is that old pointers must be nullified before new on-disk
1898  * pointers are set.  We chose this slightly more stringent requirement to
1899  * reduce complexity.) Our implementation handles this dependency by updating
1900  * the inode (or indirect block) appropriately but delaying the actual block
1901  * de-allocation (i.e., freemap and free space count manipulation) until
1902  * after the updated versions reach stable storage.  After the disk is
1903  * updated, the blocks can be safely de-allocated whenever it is convenient.
1904  * This implementation handles only the common case of reducing a file's
1905  * length to zero. Other cases are handled by the conventional synchronous
1906  * write approach.
1907  *
1908  * The ffs implementation with which we worked double-checks
1909  * the state of the block pointers and file size as it reduces
1910  * a file's length.  Some of this code is replicated here in our
1911  * soft updates implementation.  The freeblks->fb_chkcnt field is
1912  * used to transfer a part of this information to the procedure
1913  * that eventually de-allocates the blocks.
1914  *
1915  * This routine should be called from the routine that shortens
1916  * a file's length, before the inode's size or block pointers
1917  * are modified. It will save the block pointer information for
1918  * later release and zero the inode so that the calling routine
1919  * can release it.
1920  */
1921 void
1922 softdep_setup_freeblocks(ip, length, flags)
1923 	struct inode *ip;	/* The inode whose length is to be reduced */
1924 	off_t length;		/* The new length for the file */
1925 	int flags;		/* IO_EXT and/or IO_NORMAL */
1926 {
1927 	struct freeblks *freeblks;
1928 	struct inodedep *inodedep;
1929 	struct allocdirect *adp;
1930 	struct vnode *vp;
1931 	struct buf *bp;
1932 	struct fs *fs;
1933 	ufs2_daddr_t extblocks, datablocks;
1934 	int i, delay, error;
1935 
1936 	fs = ip->i_fs;
1937 	if (length != 0)
1938 		panic("softdep_setup_freeblocks: non-zero length");
1939 	MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1940 		M_FREEBLKS, M_SOFTDEP_FLAGS|M_ZERO);
1941 	freeblks->fb_list.wk_type = D_FREEBLKS;
1942 	freeblks->fb_uid = ip->i_uid;
1943 	freeblks->fb_previousinum = ip->i_number;
1944 	freeblks->fb_devvp = ip->i_devvp;
1945 	freeblks->fb_mnt = ITOV(ip)->v_mount;
1946 	extblocks = 0;
1947 	if (fs->fs_magic == FS_UFS2_MAGIC)
1948 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
1949 	datablocks = DIP(ip, i_blocks) - extblocks;
1950 	if ((flags & IO_NORMAL) == 0) {
1951 		freeblks->fb_oldsize = 0;
1952 		freeblks->fb_chkcnt = 0;
1953 	} else {
1954 		freeblks->fb_oldsize = ip->i_size;
1955 		ip->i_size = 0;
1956 		DIP(ip, i_size) = 0;
1957 		freeblks->fb_chkcnt = datablocks;
1958 		for (i = 0; i < NDADDR; i++) {
1959 			freeblks->fb_dblks[i] = DIP(ip, i_db[i]);
1960 			DIP(ip, i_db[i]) = 0;
1961 		}
1962 		for (i = 0; i < NIADDR; i++) {
1963 			freeblks->fb_iblks[i] = DIP(ip, i_ib[i]);
1964 			DIP(ip, i_ib[i]) = 0;
1965 		}
1966 		/*
1967 		 * If the file was removed, then the space being freed was
1968 		 * accounted for then (see softdep_filereleased()). If the
1969 		 * file is merely being truncated, then we account for it now.
1970 		 */
1971 		if ((ip->i_flag & IN_SPACECOUNTED) == 0)
1972 			fs->fs_pendingblocks += datablocks;
1973 	}
1974 	if ((flags & IO_EXT) == 0) {
1975 		freeblks->fb_oldextsize = 0;
1976 	} else {
1977 		freeblks->fb_oldextsize = ip->i_din2->di_extsize;
1978 		ip->i_din2->di_extsize = 0;
1979 		freeblks->fb_chkcnt += extblocks;
1980 		for (i = 0; i < NXADDR; i++) {
1981 			freeblks->fb_eblks[i] = ip->i_din2->di_extb[i];
1982 			ip->i_din2->di_extb[i] = 0;
1983 		}
1984 	}
1985 	DIP(ip, i_blocks) -= freeblks->fb_chkcnt;
1986 	/*
1987 	 * Push the zero'ed inode to to its disk buffer so that we are free
1988 	 * to delete its dependencies below. Once the dependencies are gone
1989 	 * the buffer can be safely released.
1990 	 */
1991 	if ((error = bread(ip->i_devvp,
1992 	    fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
1993 	    (int)fs->fs_bsize, NOCRED, &bp)) != 0) {
1994 		brelse(bp);
1995 		softdep_error("softdep_setup_freeblocks", error);
1996 	}
1997 	if (ip->i_ump->um_fstype == UFS1)
1998 		*((struct ufs1_dinode *)bp->b_data +
1999 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din1;
2000 	else
2001 		*((struct ufs2_dinode *)bp->b_data +
2002 		    ino_to_fsbo(fs, ip->i_number)) = *ip->i_din2;
2003 	/*
2004 	 * Find and eliminate any inode dependencies.
2005 	 */
2006 	ACQUIRE_LOCK(&lk);
2007 	(void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
2008 	if ((inodedep->id_state & IOSTARTED) != 0) {
2009 		FREE_LOCK(&lk);
2010 		panic("softdep_setup_freeblocks: inode busy");
2011 	}
2012 	/*
2013 	 * Add the freeblks structure to the list of operations that
2014 	 * must await the zero'ed inode being written to disk. If we
2015 	 * still have a bitmap dependency (delay == 0), then the inode
2016 	 * has never been written to disk, so we can process the
2017 	 * freeblks below once we have deleted the dependencies.
2018 	 */
2019 	delay = (inodedep->id_state & DEPCOMPLETE);
2020 	if (delay)
2021 		WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
2022 	/*
2023 	 * Because the file length has been truncated to zero, any
2024 	 * pending block allocation dependency structures associated
2025 	 * with this inode are obsolete and can simply be de-allocated.
2026 	 * We must first merge the two dependency lists to get rid of
2027 	 * any duplicate freefrag structures, then purge the merged list.
2028 	 * If we still have a bitmap dependency, then the inode has never
2029 	 * been written to disk, so we can free any fragments without delay.
2030 	 */
2031 	if (flags & IO_NORMAL) {
2032 		merge_inode_lists(&inodedep->id_newinoupdt,
2033 		    &inodedep->id_inoupdt);
2034 		while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
2035 			free_allocdirect(&inodedep->id_inoupdt, adp, delay);
2036 	}
2037 	if (flags & IO_EXT) {
2038 		merge_inode_lists(&inodedep->id_newextupdt,
2039 		    &inodedep->id_extupdt);
2040 		while ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != 0)
2041 			free_allocdirect(&inodedep->id_extupdt, adp, delay);
2042 	}
2043 	FREE_LOCK(&lk);
2044 	bdwrite(bp);
2045 	/*
2046 	 * We must wait for any I/O in progress to finish so that
2047 	 * all potential buffers on the dirty list will be visible.
2048 	 * Once they are all there, walk the list and get rid of
2049 	 * any dependencies.
2050 	 */
2051 	vp = ITOV(ip);
2052 	ACQUIRE_LOCK(&lk);
2053 	drain_output(vp, 1);
2054 restart:
2055 	VI_LOCK(vp);
2056 	TAILQ_FOREACH(bp, &vp->v_dirtyblkhd, b_vnbufs) {
2057 		if (((flags & IO_EXT) == 0 && (bp->b_xflags & BX_ALTDATA)) ||
2058 		    ((flags & IO_NORMAL) == 0 &&
2059 		      (bp->b_xflags & BX_ALTDATA) == 0))
2060 			continue;
2061 		VI_UNLOCK(vp);
2062 		if (getdirtybuf(&bp, MNT_WAIT) == 0)
2063 			goto restart;
2064 		(void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
2065 		deallocate_dependencies(bp, inodedep);
2066 		bp->b_flags |= B_INVAL | B_NOCACHE;
2067 		FREE_LOCK(&lk);
2068 		brelse(bp);
2069 		ACQUIRE_LOCK(&lk);
2070 		goto restart;
2071 	}
2072 	VI_UNLOCK(vp);
2073 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
2074 		(void) free_inodedep(inodedep);
2075 	FREE_LOCK(&lk);
2076 	/*
2077 	 * If the inode has never been written to disk (delay == 0),
2078 	 * then we can process the freeblks now that we have deleted
2079 	 * the dependencies.
2080 	 */
2081 	if (!delay)
2082 		handle_workitem_freeblocks(freeblks, 0);
2083 }
2084 
2085 /*
2086  * Reclaim any dependency structures from a buffer that is about to
2087  * be reallocated to a new vnode. The buffer must be locked, thus,
2088  * no I/O completion operations can occur while we are manipulating
2089  * its associated dependencies. The mutex is held so that other I/O's
2090  * associated with related dependencies do not occur.
2091  */
2092 static void
2093 deallocate_dependencies(bp, inodedep)
2094 	struct buf *bp;
2095 	struct inodedep *inodedep;
2096 {
2097 	struct worklist *wk;
2098 	struct indirdep *indirdep;
2099 	struct allocindir *aip;
2100 	struct pagedep *pagedep;
2101 	struct dirrem *dirrem;
2102 	struct diradd *dap;
2103 	int i;
2104 
2105 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2106 		switch (wk->wk_type) {
2107 
2108 		case D_INDIRDEP:
2109 			indirdep = WK_INDIRDEP(wk);
2110 			/*
2111 			 * None of the indirect pointers will ever be visible,
2112 			 * so they can simply be tossed. GOINGAWAY ensures
2113 			 * that allocated pointers will be saved in the buffer
2114 			 * cache until they are freed. Note that they will
2115 			 * only be able to be found by their physical address
2116 			 * since the inode mapping the logical address will
2117 			 * be gone. The save buffer used for the safe copy
2118 			 * was allocated in setup_allocindir_phase2 using
2119 			 * the physical address so it could be used for this
2120 			 * purpose. Hence we swap the safe copy with the real
2121 			 * copy, allowing the safe copy to be freed and holding
2122 			 * on to the real copy for later use in indir_trunc.
2123 			 */
2124 			if (indirdep->ir_state & GOINGAWAY) {
2125 				FREE_LOCK(&lk);
2126 				panic("deallocate_dependencies: already gone");
2127 			}
2128 			indirdep->ir_state |= GOINGAWAY;
2129 			VFSTOUFS(bp->b_vp->v_mount)->um_numindirdeps += 1;
2130 			while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
2131 				free_allocindir(aip, inodedep);
2132 			if (bp->b_lblkno >= 0 ||
2133 			    bp->b_blkno != indirdep->ir_savebp->b_lblkno) {
2134 				FREE_LOCK(&lk);
2135 				panic("deallocate_dependencies: not indir");
2136 			}
2137 			bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2138 			    bp->b_bcount);
2139 			WORKLIST_REMOVE(wk);
2140 			WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
2141 			continue;
2142 
2143 		case D_PAGEDEP:
2144 			pagedep = WK_PAGEDEP(wk);
2145 			/*
2146 			 * None of the directory additions will ever be
2147 			 * visible, so they can simply be tossed.
2148 			 */
2149 			for (i = 0; i < DAHASHSZ; i++)
2150 				while ((dap =
2151 				    LIST_FIRST(&pagedep->pd_diraddhd[i])))
2152 					free_diradd(dap);
2153 			while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2154 				free_diradd(dap);
2155 			/*
2156 			 * Copy any directory remove dependencies to the list
2157 			 * to be processed after the zero'ed inode is written.
2158 			 * If the inode has already been written, then they
2159 			 * can be dumped directly onto the work list.
2160 			 */
2161 			LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2162 				LIST_REMOVE(dirrem, dm_next);
2163 				dirrem->dm_dirinum = pagedep->pd_ino;
2164 				if (inodedep == NULL ||
2165 				    (inodedep->id_state & ALLCOMPLETE) ==
2166 				     ALLCOMPLETE)
2167 					add_to_worklist(&dirrem->dm_list);
2168 				else
2169 					WORKLIST_INSERT(&inodedep->id_bufwait,
2170 					    &dirrem->dm_list);
2171 			}
2172 			if ((pagedep->pd_state & NEWBLOCK) != 0) {
2173 				LIST_FOREACH(wk, &inodedep->id_bufwait, wk_list)
2174 					if (wk->wk_type == D_NEWDIRBLK &&
2175 					    WK_NEWDIRBLK(wk)->db_pagedep ==
2176 					      pagedep)
2177 						break;
2178 				if (wk != NULL) {
2179 					WORKLIST_REMOVE(wk);
2180 					free_newdirblk(WK_NEWDIRBLK(wk));
2181 				} else {
2182 					FREE_LOCK(&lk);
2183 					panic("deallocate_dependencies: "
2184 					      "lost pagedep");
2185 				}
2186 			}
2187 			WORKLIST_REMOVE(&pagedep->pd_list);
2188 			LIST_REMOVE(pagedep, pd_hash);
2189 			WORKITEM_FREE(pagedep, D_PAGEDEP);
2190 			continue;
2191 
2192 		case D_ALLOCINDIR:
2193 			free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2194 			continue;
2195 
2196 		case D_ALLOCDIRECT:
2197 		case D_INODEDEP:
2198 			FREE_LOCK(&lk);
2199 			panic("deallocate_dependencies: Unexpected type %s",
2200 			    TYPENAME(wk->wk_type));
2201 			/* NOTREACHED */
2202 
2203 		default:
2204 			FREE_LOCK(&lk);
2205 			panic("deallocate_dependencies: Unknown type %s",
2206 			    TYPENAME(wk->wk_type));
2207 			/* NOTREACHED */
2208 		}
2209 	}
2210 }
2211 
2212 /*
2213  * Free an allocdirect. Generate a new freefrag work request if appropriate.
2214  * This routine must be called with splbio interrupts blocked.
2215  */
2216 static void
2217 free_allocdirect(adphead, adp, delay)
2218 	struct allocdirectlst *adphead;
2219 	struct allocdirect *adp;
2220 	int delay;
2221 {
2222 	struct newdirblk *newdirblk;
2223 	struct worklist *wk;
2224 
2225 #ifdef DEBUG
2226 	if (lk.lkt_held == NOHOLDER)
2227 		panic("free_allocdirect: lock not held");
2228 #endif
2229 	if ((adp->ad_state & DEPCOMPLETE) == 0)
2230 		LIST_REMOVE(adp, ad_deps);
2231 	TAILQ_REMOVE(adphead, adp, ad_next);
2232 	if ((adp->ad_state & COMPLETE) == 0)
2233 		WORKLIST_REMOVE(&adp->ad_list);
2234 	if (adp->ad_freefrag != NULL) {
2235 		if (delay)
2236 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2237 			    &adp->ad_freefrag->ff_list);
2238 		else
2239 			add_to_worklist(&adp->ad_freefrag->ff_list);
2240 	}
2241 	if ((wk = LIST_FIRST(&adp->ad_newdirblk)) != NULL) {
2242 		newdirblk = WK_NEWDIRBLK(wk);
2243 		WORKLIST_REMOVE(&newdirblk->db_list);
2244 		if (LIST_FIRST(&adp->ad_newdirblk) != NULL)
2245 			panic("free_allocdirect: extra newdirblk");
2246 		if (delay)
2247 			WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2248 			    &newdirblk->db_list);
2249 		else
2250 			free_newdirblk(newdirblk);
2251 	}
2252 	WORKITEM_FREE(adp, D_ALLOCDIRECT);
2253 }
2254 
2255 /*
2256  * Free a newdirblk. Clear the NEWBLOCK flag on its associated pagedep.
2257  * This routine must be called with splbio interrupts blocked.
2258  */
2259 static void
2260 free_newdirblk(newdirblk)
2261 	struct newdirblk *newdirblk;
2262 {
2263 	struct pagedep *pagedep;
2264 	struct diradd *dap;
2265 	int i;
2266 
2267 #ifdef DEBUG
2268 	if (lk.lkt_held == NOHOLDER)
2269 		panic("free_newdirblk: lock not held");
2270 #endif
2271 	/*
2272 	 * If the pagedep is still linked onto the directory buffer
2273 	 * dependency chain, then some of the entries on the
2274 	 * pd_pendinghd list may not be committed to disk yet. In
2275 	 * this case, we will simply clear the NEWBLOCK flag and
2276 	 * let the pd_pendinghd list be processed when the pagedep
2277 	 * is next written. If the pagedep is no longer on the buffer
2278 	 * dependency chain, then all the entries on the pd_pending
2279 	 * list are committed to disk and we can free them here.
2280 	 */
2281 	pagedep = newdirblk->db_pagedep;
2282 	pagedep->pd_state &= ~NEWBLOCK;
2283 	if ((pagedep->pd_state & ONWORKLIST) == 0)
2284 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
2285 			free_diradd(dap);
2286 	/*
2287 	 * If no dependencies remain, the pagedep will be freed.
2288 	 */
2289 	for (i = 0; i < DAHASHSZ; i++)
2290 		if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
2291 			break;
2292 	if (i == DAHASHSZ && (pagedep->pd_state & ONWORKLIST) == 0) {
2293 		LIST_REMOVE(pagedep, pd_hash);
2294 		WORKITEM_FREE(pagedep, D_PAGEDEP);
2295 	}
2296 	WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2297 }
2298 
2299 /*
2300  * Prepare an inode to be freed. The actual free operation is not
2301  * done until the zero'ed inode has been written to disk.
2302  */
2303 void
2304 softdep_freefile(pvp, ino, mode)
2305 	struct vnode *pvp;
2306 	ino_t ino;
2307 	int mode;
2308 {
2309 	struct inode *ip = VTOI(pvp);
2310 	struct inodedep *inodedep;
2311 	struct freefile *freefile;
2312 
2313 	/*
2314 	 * This sets up the inode de-allocation dependency.
2315 	 */
2316 	MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2317 		M_FREEFILE, M_SOFTDEP_FLAGS);
2318 	freefile->fx_list.wk_type = D_FREEFILE;
2319 	freefile->fx_list.wk_state = 0;
2320 	freefile->fx_mode = mode;
2321 	freefile->fx_oldinum = ino;
2322 	freefile->fx_devvp = ip->i_devvp;
2323 	freefile->fx_mnt = ITOV(ip)->v_mount;
2324 	if ((ip->i_flag & IN_SPACECOUNTED) == 0)
2325 		ip->i_fs->fs_pendinginodes += 1;
2326 
2327 	/*
2328 	 * If the inodedep does not exist, then the zero'ed inode has
2329 	 * been written to disk. If the allocated inode has never been
2330 	 * written to disk, then the on-disk inode is zero'ed. In either
2331 	 * case we can free the file immediately.
2332 	 */
2333 	ACQUIRE_LOCK(&lk);
2334 	if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2335 	    check_inode_unwritten(inodedep)) {
2336 		FREE_LOCK(&lk);
2337 		handle_workitem_freefile(freefile);
2338 		return;
2339 	}
2340 	WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2341 	FREE_LOCK(&lk);
2342 }
2343 
2344 /*
2345  * Check to see if an inode has never been written to disk. If
2346  * so free the inodedep and return success, otherwise return failure.
2347  * This routine must be called with splbio interrupts blocked.
2348  *
2349  * If we still have a bitmap dependency, then the inode has never
2350  * been written to disk. Drop the dependency as it is no longer
2351  * necessary since the inode is being deallocated. We set the
2352  * ALLCOMPLETE flags since the bitmap now properly shows that the
2353  * inode is not allocated. Even if the inode is actively being
2354  * written, it has been rolled back to its zero'ed state, so we
2355  * are ensured that a zero inode is what is on the disk. For short
2356  * lived files, this change will usually result in removing all the
2357  * dependencies from the inode so that it can be freed immediately.
2358  */
2359 static int
2360 check_inode_unwritten(inodedep)
2361 	struct inodedep *inodedep;
2362 {
2363 
2364 	if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2365 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2366 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2367 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
2368 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2369 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2370 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
2371 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
2372 	    inodedep->id_nlinkdelta != 0)
2373 		return (0);
2374 	inodedep->id_state |= ALLCOMPLETE;
2375 	LIST_REMOVE(inodedep, id_deps);
2376 	inodedep->id_buf = NULL;
2377 	if (inodedep->id_state & ONWORKLIST)
2378 		WORKLIST_REMOVE(&inodedep->id_list);
2379 	if (inodedep->id_savedino1 != NULL) {
2380 		FREE(inodedep->id_savedino1, M_INODEDEP);
2381 		inodedep->id_savedino1 = NULL;
2382 	}
2383 	if (free_inodedep(inodedep) == 0) {
2384 		FREE_LOCK(&lk);
2385 		panic("check_inode_unwritten: busy inode");
2386 	}
2387 	return (1);
2388 }
2389 
2390 /*
2391  * Try to free an inodedep structure. Return 1 if it could be freed.
2392  */
2393 static int
2394 free_inodedep(inodedep)
2395 	struct inodedep *inodedep;
2396 {
2397 
2398 	if ((inodedep->id_state & ONWORKLIST) != 0 ||
2399 	    (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2400 	    LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2401 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2402 	    LIST_FIRST(&inodedep->id_inowait) != NULL ||
2403 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2404 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2405 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
2406 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
2407 	    inodedep->id_nlinkdelta != 0 || inodedep->id_savedino1 != NULL)
2408 		return (0);
2409 	LIST_REMOVE(inodedep, id_hash);
2410 	WORKITEM_FREE(inodedep, D_INODEDEP);
2411 	num_inodedep -= 1;
2412 	return (1);
2413 }
2414 
2415 /*
2416  * This workitem routine performs the block de-allocation.
2417  * The workitem is added to the pending list after the updated
2418  * inode block has been written to disk.  As mentioned above,
2419  * checks regarding the number of blocks de-allocated (compared
2420  * to the number of blocks allocated for the file) are also
2421  * performed in this function.
2422  */
2423 static void
2424 handle_workitem_freeblocks(freeblks, flags)
2425 	struct freeblks *freeblks;
2426 	int flags;
2427 {
2428 	struct inode *ip;
2429 	struct vnode *vp;
2430 	struct fs *fs;
2431 	int i, nblocks, level, bsize;
2432 	ufs2_daddr_t bn, blocksreleased = 0;
2433 	int error, allerror = 0;
2434 	ufs_lbn_t baselbns[NIADDR], tmpval;
2435 
2436 	fs = VFSTOUFS(freeblks->fb_mnt)->um_fs;
2437 	tmpval = 1;
2438 	baselbns[0] = NDADDR;
2439 	for (i = 1; i < NIADDR; i++) {
2440 		tmpval *= NINDIR(fs);
2441 		baselbns[i] = baselbns[i - 1] + tmpval;
2442 	}
2443 	nblocks = btodb(fs->fs_bsize);
2444 	blocksreleased = 0;
2445 	/*
2446 	 * Release all extended attribute blocks or frags.
2447 	 */
2448 	if (freeblks->fb_oldextsize > 0) {
2449 		for (i = (NXADDR - 1); i >= 0; i--) {
2450 			if ((bn = freeblks->fb_eblks[i]) == 0)
2451 				continue;
2452 			bsize = sblksize(fs, freeblks->fb_oldextsize, i);
2453 			ffs_blkfree(fs, freeblks->fb_devvp, bn, bsize,
2454 			    freeblks->fb_previousinum);
2455 			blocksreleased += btodb(bsize);
2456 		}
2457 	}
2458 	/*
2459 	 * Release all data blocks or frags.
2460 	 */
2461 	if (freeblks->fb_oldsize > 0) {
2462 		/*
2463 		 * Indirect blocks first.
2464 		 */
2465 		for (level = (NIADDR - 1); level >= 0; level--) {
2466 			if ((bn = freeblks->fb_iblks[level]) == 0)
2467 				continue;
2468 			if ((error = indir_trunc(freeblks, fsbtodb(fs, bn),
2469 			    level, baselbns[level], &blocksreleased)) == 0)
2470 				allerror = error;
2471 			ffs_blkfree(fs, freeblks->fb_devvp, bn, fs->fs_bsize,
2472 			    freeblks->fb_previousinum);
2473 			fs->fs_pendingblocks -= nblocks;
2474 			blocksreleased += nblocks;
2475 		}
2476 		/*
2477 		 * All direct blocks or frags.
2478 		 */
2479 		for (i = (NDADDR - 1); i >= 0; i--) {
2480 			if ((bn = freeblks->fb_dblks[i]) == 0)
2481 				continue;
2482 			bsize = sblksize(fs, freeblks->fb_oldsize, i);
2483 			ffs_blkfree(fs, freeblks->fb_devvp, bn, bsize,
2484 			    freeblks->fb_previousinum);
2485 			fs->fs_pendingblocks -= btodb(bsize);
2486 			blocksreleased += btodb(bsize);
2487 		}
2488 	}
2489 	/*
2490 	 * If we still have not finished background cleanup, then check
2491 	 * to see if the block count needs to be adjusted.
2492 	 */
2493 	if (freeblks->fb_chkcnt != blocksreleased &&
2494 	    (fs->fs_flags & FS_UNCLEAN) != 0 &&
2495 	    VFS_VGET(freeblks->fb_mnt, freeblks->fb_previousinum,
2496 	    (flags & LK_NOWAIT) | LK_EXCLUSIVE, &vp) == 0) {
2497 		ip = VTOI(vp);
2498 		DIP(ip, i_blocks) += freeblks->fb_chkcnt - blocksreleased;
2499 		ip->i_flag |= IN_CHANGE;
2500 		vput(vp);
2501 	}
2502 
2503 #ifdef DIAGNOSTIC
2504 	if (freeblks->fb_chkcnt != blocksreleased &&
2505 	    ((fs->fs_flags & FS_UNCLEAN) == 0 || (flags & LK_NOWAIT) != 0))
2506 		printf("handle_workitem_freeblocks: block count\n");
2507 	if (allerror)
2508 		softdep_error("handle_workitem_freeblks", allerror);
2509 #endif /* DIAGNOSTIC */
2510 
2511 	WORKITEM_FREE(freeblks, D_FREEBLKS);
2512 }
2513 
2514 /*
2515  * Release blocks associated with the inode ip and stored in the indirect
2516  * block dbn. If level is greater than SINGLE, the block is an indirect block
2517  * and recursive calls to indirtrunc must be used to cleanse other indirect
2518  * blocks.
2519  */
2520 static int
2521 indir_trunc(freeblks, dbn, level, lbn, countp)
2522 	struct freeblks *freeblks;
2523 	ufs2_daddr_t dbn;
2524 	int level;
2525 	ufs_lbn_t lbn;
2526 	ufs2_daddr_t *countp;
2527 {
2528 	struct buf *bp;
2529 	struct fs *fs;
2530 	struct worklist *wk;
2531 	struct indirdep *indirdep;
2532 	ufs1_daddr_t *bap1 = 0;
2533 	ufs2_daddr_t nb, *bap2 = 0;
2534 	ufs_lbn_t lbnadd;
2535 	int i, nblocks, ufs1fmt;
2536 	int error, allerror = 0;
2537 
2538 	fs = VFSTOUFS(freeblks->fb_mnt)->um_fs;
2539 	lbnadd = 1;
2540 	for (i = level; i > 0; i--)
2541 		lbnadd *= NINDIR(fs);
2542 	/*
2543 	 * Get buffer of block pointers to be freed. This routine is not
2544 	 * called until the zero'ed inode has been written, so it is safe
2545 	 * to free blocks as they are encountered. Because the inode has
2546 	 * been zero'ed, calls to bmap on these blocks will fail. So, we
2547 	 * have to use the on-disk address and the block device for the
2548 	 * filesystem to look them up. If the file was deleted before its
2549 	 * indirect blocks were all written to disk, the routine that set
2550 	 * us up (deallocate_dependencies) will have arranged to leave
2551 	 * a complete copy of the indirect block in memory for our use.
2552 	 * Otherwise we have to read the blocks in from the disk.
2553 	 */
2554 	ACQUIRE_LOCK(&lk);
2555 	if ((bp = incore(freeblks->fb_devvp, dbn)) != NULL &&
2556 	    (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2557 		if (wk->wk_type != D_INDIRDEP ||
2558 		    (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2559 		    (indirdep->ir_state & GOINGAWAY) == 0) {
2560 			FREE_LOCK(&lk);
2561 			panic("indir_trunc: lost indirdep");
2562 		}
2563 		WORKLIST_REMOVE(wk);
2564 		WORKITEM_FREE(indirdep, D_INDIRDEP);
2565 		if (LIST_FIRST(&bp->b_dep) != NULL) {
2566 			FREE_LOCK(&lk);
2567 			panic("indir_trunc: dangling dep");
2568 		}
2569 		VFSTOUFS(freeblks->fb_mnt)->um_numindirdeps -= 1;
2570 		FREE_LOCK(&lk);
2571 	} else {
2572 		FREE_LOCK(&lk);
2573 		error = bread(freeblks->fb_devvp, dbn, (int)fs->fs_bsize,
2574 		    NOCRED, &bp);
2575 		if (error) {
2576 			brelse(bp);
2577 			return (error);
2578 		}
2579 	}
2580 	/*
2581 	 * Recursively free indirect blocks.
2582 	 */
2583 	if (VFSTOUFS(freeblks->fb_mnt)->um_fstype == UFS1) {
2584 		ufs1fmt = 1;
2585 		bap1 = (ufs1_daddr_t *)bp->b_data;
2586 	} else {
2587 		ufs1fmt = 0;
2588 		bap2 = (ufs2_daddr_t *)bp->b_data;
2589 	}
2590 	nblocks = btodb(fs->fs_bsize);
2591 	for (i = NINDIR(fs) - 1; i >= 0; i--) {
2592 		if (ufs1fmt)
2593 			nb = bap1[i];
2594 		else
2595 			nb = bap2[i];
2596 		if (nb == 0)
2597 			continue;
2598 		if (level != 0) {
2599 			if ((error = indir_trunc(freeblks, fsbtodb(fs, nb),
2600 			     level - 1, lbn + (i * lbnadd), countp)) != 0)
2601 				allerror = error;
2602 		}
2603 		ffs_blkfree(fs, freeblks->fb_devvp, nb, fs->fs_bsize,
2604 		    freeblks->fb_previousinum);
2605 		fs->fs_pendingblocks -= nblocks;
2606 		*countp += nblocks;
2607 	}
2608 	bp->b_flags |= B_INVAL | B_NOCACHE;
2609 	brelse(bp);
2610 	return (allerror);
2611 }
2612 
2613 /*
2614  * Free an allocindir.
2615  * This routine must be called with splbio interrupts blocked.
2616  */
2617 static void
2618 free_allocindir(aip, inodedep)
2619 	struct allocindir *aip;
2620 	struct inodedep *inodedep;
2621 {
2622 	struct freefrag *freefrag;
2623 
2624 #ifdef DEBUG
2625 	if (lk.lkt_held == NOHOLDER)
2626 		panic("free_allocindir: lock not held");
2627 #endif
2628 	if ((aip->ai_state & DEPCOMPLETE) == 0)
2629 		LIST_REMOVE(aip, ai_deps);
2630 	if (aip->ai_state & ONWORKLIST)
2631 		WORKLIST_REMOVE(&aip->ai_list);
2632 	LIST_REMOVE(aip, ai_next);
2633 	if ((freefrag = aip->ai_freefrag) != NULL) {
2634 		if (inodedep == NULL)
2635 			add_to_worklist(&freefrag->ff_list);
2636 		else
2637 			WORKLIST_INSERT(&inodedep->id_bufwait,
2638 			    &freefrag->ff_list);
2639 	}
2640 	WORKITEM_FREE(aip, D_ALLOCINDIR);
2641 }
2642 
2643 /*
2644  * Directory entry addition dependencies.
2645  *
2646  * When adding a new directory entry, the inode (with its incremented link
2647  * count) must be written to disk before the directory entry's pointer to it.
2648  * Also, if the inode is newly allocated, the corresponding freemap must be
2649  * updated (on disk) before the directory entry's pointer. These requirements
2650  * are met via undo/redo on the directory entry's pointer, which consists
2651  * simply of the inode number.
2652  *
2653  * As directory entries are added and deleted, the free space within a
2654  * directory block can become fragmented.  The ufs filesystem will compact
2655  * a fragmented directory block to make space for a new entry. When this
2656  * occurs, the offsets of previously added entries change. Any "diradd"
2657  * dependency structures corresponding to these entries must be updated with
2658  * the new offsets.
2659  */
2660 
2661 /*
2662  * This routine is called after the in-memory inode's link
2663  * count has been incremented, but before the directory entry's
2664  * pointer to the inode has been set.
2665  */
2666 int
2667 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp, isnewblk)
2668 	struct buf *bp;		/* buffer containing directory block */
2669 	struct inode *dp;	/* inode for directory */
2670 	off_t diroffset;	/* offset of new entry in directory */
2671 	ino_t newinum;		/* inode referenced by new directory entry */
2672 	struct buf *newdirbp;	/* non-NULL => contents of new mkdir */
2673 	int isnewblk;		/* entry is in a newly allocated block */
2674 {
2675 	int offset;		/* offset of new entry within directory block */
2676 	ufs_lbn_t lbn;		/* block in directory containing new entry */
2677 	struct fs *fs;
2678 	struct diradd *dap;
2679 	struct allocdirect *adp;
2680 	struct pagedep *pagedep;
2681 	struct inodedep *inodedep;
2682 	struct newdirblk *newdirblk = 0;
2683 	struct mkdir *mkdir1, *mkdir2;
2684 
2685 	/*
2686 	 * Whiteouts have no dependencies.
2687 	 */
2688 	if (newinum == WINO) {
2689 		if (newdirbp != NULL)
2690 			bdwrite(newdirbp);
2691 		return (0);
2692 	}
2693 
2694 	fs = dp->i_fs;
2695 	lbn = lblkno(fs, diroffset);
2696 	offset = blkoff(fs, diroffset);
2697 	MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2698 		M_SOFTDEP_FLAGS|M_ZERO);
2699 	dap->da_list.wk_type = D_DIRADD;
2700 	dap->da_offset = offset;
2701 	dap->da_newinum = newinum;
2702 	dap->da_state = ATTACHED;
2703 	if (isnewblk && lbn < NDADDR && fragoff(fs, diroffset) == 0) {
2704 		MALLOC(newdirblk, struct newdirblk *, sizeof(struct newdirblk),
2705 		    M_NEWDIRBLK, M_SOFTDEP_FLAGS);
2706 		newdirblk->db_list.wk_type = D_NEWDIRBLK;
2707 		newdirblk->db_state = 0;
2708 	}
2709 	if (newdirbp == NULL) {
2710 		dap->da_state |= DEPCOMPLETE;
2711 		ACQUIRE_LOCK(&lk);
2712 	} else {
2713 		dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2714 		MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2715 		    M_SOFTDEP_FLAGS);
2716 		mkdir1->md_list.wk_type = D_MKDIR;
2717 		mkdir1->md_state = MKDIR_BODY;
2718 		mkdir1->md_diradd = dap;
2719 		MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2720 		    M_SOFTDEP_FLAGS);
2721 		mkdir2->md_list.wk_type = D_MKDIR;
2722 		mkdir2->md_state = MKDIR_PARENT;
2723 		mkdir2->md_diradd = dap;
2724 		/*
2725 		 * Dependency on "." and ".." being written to disk.
2726 		 */
2727 		mkdir1->md_buf = newdirbp;
2728 		ACQUIRE_LOCK(&lk);
2729 		LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2730 		WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2731 		FREE_LOCK(&lk);
2732 		bdwrite(newdirbp);
2733 		/*
2734 		 * Dependency on link count increase for parent directory
2735 		 */
2736 		ACQUIRE_LOCK(&lk);
2737 		if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0
2738 		    || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2739 			dap->da_state &= ~MKDIR_PARENT;
2740 			WORKITEM_FREE(mkdir2, D_MKDIR);
2741 		} else {
2742 			LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2743 			WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2744 		}
2745 	}
2746 	/*
2747 	 * Link into parent directory pagedep to await its being written.
2748 	 */
2749 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2750 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2751 	dap->da_pagedep = pagedep;
2752 	LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2753 	    da_pdlist);
2754 	/*
2755 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
2756 	 * is not yet written. If it is written, do the post-inode write
2757 	 * processing to put it on the id_pendinghd list.
2758 	 */
2759 	(void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2760 	if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2761 		diradd_inode_written(dap, inodedep);
2762 	else
2763 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2764 	if (isnewblk) {
2765 		/*
2766 		 * Directories growing into indirect blocks are rare
2767 		 * enough and the frequency of new block allocation
2768 		 * in those cases even more rare, that we choose not
2769 		 * to bother tracking them. Rather we simply force the
2770 		 * new directory entry to disk.
2771 		 */
2772 		if (lbn >= NDADDR) {
2773 			FREE_LOCK(&lk);
2774 			/*
2775 			 * We only have a new allocation when at the
2776 			 * beginning of a new block, not when we are
2777 			 * expanding into an existing block.
2778 			 */
2779 			if (blkoff(fs, diroffset) == 0)
2780 				return (1);
2781 			return (0);
2782 		}
2783 		/*
2784 		 * We only have a new allocation when at the beginning
2785 		 * of a new fragment, not when we are expanding into an
2786 		 * existing fragment. Also, there is nothing to do if we
2787 		 * are already tracking this block.
2788 		 */
2789 		if (fragoff(fs, diroffset) != 0) {
2790 			FREE_LOCK(&lk);
2791 			return (0);
2792 		}
2793 		if ((pagedep->pd_state & NEWBLOCK) != 0) {
2794 			WORKITEM_FREE(newdirblk, D_NEWDIRBLK);
2795 			FREE_LOCK(&lk);
2796 			return (0);
2797 		}
2798 		/*
2799 		 * Find our associated allocdirect and have it track us.
2800 		 */
2801 		if (inodedep_lookup(fs, dp->i_number, 0, &inodedep) == 0)
2802 			panic("softdep_setup_directory_add: lost inodedep");
2803 		adp = TAILQ_LAST(&inodedep->id_newinoupdt, allocdirectlst);
2804 		if (adp == NULL || adp->ad_lbn != lbn) {
2805 			FREE_LOCK(&lk);
2806 			panic("softdep_setup_directory_add: lost entry");
2807 		}
2808 		pagedep->pd_state |= NEWBLOCK;
2809 		newdirblk->db_pagedep = pagedep;
2810 		WORKLIST_INSERT(&adp->ad_newdirblk, &newdirblk->db_list);
2811 	}
2812 	FREE_LOCK(&lk);
2813 	return (0);
2814 }
2815 
2816 /*
2817  * This procedure is called to change the offset of a directory
2818  * entry when compacting a directory block which must be owned
2819  * exclusively by the caller. Note that the actual entry movement
2820  * must be done in this procedure to ensure that no I/O completions
2821  * occur while the move is in progress.
2822  */
2823 void
2824 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2825 	struct inode *dp;	/* inode for directory */
2826 	caddr_t base;		/* address of dp->i_offset */
2827 	caddr_t oldloc;		/* address of old directory location */
2828 	caddr_t newloc;		/* address of new directory location */
2829 	int entrysize;		/* size of directory entry */
2830 {
2831 	int offset, oldoffset, newoffset;
2832 	struct pagedep *pagedep;
2833 	struct diradd *dap;
2834 	ufs_lbn_t lbn;
2835 
2836 	ACQUIRE_LOCK(&lk);
2837 	lbn = lblkno(dp->i_fs, dp->i_offset);
2838 	offset = blkoff(dp->i_fs, dp->i_offset);
2839 	if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2840 		goto done;
2841 	oldoffset = offset + (oldloc - base);
2842 	newoffset = offset + (newloc - base);
2843 
2844 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2845 		if (dap->da_offset != oldoffset)
2846 			continue;
2847 		dap->da_offset = newoffset;
2848 		if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2849 			break;
2850 		LIST_REMOVE(dap, da_pdlist);
2851 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2852 		    dap, da_pdlist);
2853 		break;
2854 	}
2855 	if (dap == NULL) {
2856 
2857 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2858 			if (dap->da_offset == oldoffset) {
2859 				dap->da_offset = newoffset;
2860 				break;
2861 			}
2862 		}
2863 	}
2864 done:
2865 	bcopy(oldloc, newloc, entrysize);
2866 	FREE_LOCK(&lk);
2867 }
2868 
2869 /*
2870  * Free a diradd dependency structure. This routine must be called
2871  * with splbio interrupts blocked.
2872  */
2873 static void
2874 free_diradd(dap)
2875 	struct diradd *dap;
2876 {
2877 	struct dirrem *dirrem;
2878 	struct pagedep *pagedep;
2879 	struct inodedep *inodedep;
2880 	struct mkdir *mkdir, *nextmd;
2881 
2882 #ifdef DEBUG
2883 	if (lk.lkt_held == NOHOLDER)
2884 		panic("free_diradd: lock not held");
2885 #endif
2886 	WORKLIST_REMOVE(&dap->da_list);
2887 	LIST_REMOVE(dap, da_pdlist);
2888 	if ((dap->da_state & DIRCHG) == 0) {
2889 		pagedep = dap->da_pagedep;
2890 	} else {
2891 		dirrem = dap->da_previous;
2892 		pagedep = dirrem->dm_pagedep;
2893 		dirrem->dm_dirinum = pagedep->pd_ino;
2894 		add_to_worklist(&dirrem->dm_list);
2895 	}
2896 	if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2897 	    0, &inodedep) != 0)
2898 		(void) free_inodedep(inodedep);
2899 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2900 		for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2901 			nextmd = LIST_NEXT(mkdir, md_mkdirs);
2902 			if (mkdir->md_diradd != dap)
2903 				continue;
2904 			dap->da_state &= ~mkdir->md_state;
2905 			WORKLIST_REMOVE(&mkdir->md_list);
2906 			LIST_REMOVE(mkdir, md_mkdirs);
2907 			WORKITEM_FREE(mkdir, D_MKDIR);
2908 		}
2909 		if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2910 			FREE_LOCK(&lk);
2911 			panic("free_diradd: unfound ref");
2912 		}
2913 	}
2914 	WORKITEM_FREE(dap, D_DIRADD);
2915 }
2916 
2917 /*
2918  * Directory entry removal dependencies.
2919  *
2920  * When removing a directory entry, the entry's inode pointer must be
2921  * zero'ed on disk before the corresponding inode's link count is decremented
2922  * (possibly freeing the inode for re-use). This dependency is handled by
2923  * updating the directory entry but delaying the inode count reduction until
2924  * after the directory block has been written to disk. After this point, the
2925  * inode count can be decremented whenever it is convenient.
2926  */
2927 
2928 /*
2929  * This routine should be called immediately after removing
2930  * a directory entry.  The inode's link count should not be
2931  * decremented by the calling procedure -- the soft updates
2932  * code will do this task when it is safe.
2933  */
2934 void
2935 softdep_setup_remove(bp, dp, ip, isrmdir)
2936 	struct buf *bp;		/* buffer containing directory block */
2937 	struct inode *dp;	/* inode for the directory being modified */
2938 	struct inode *ip;	/* inode for directory entry being removed */
2939 	int isrmdir;		/* indicates if doing RMDIR */
2940 {
2941 	struct dirrem *dirrem, *prevdirrem;
2942 
2943 	/*
2944 	 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2945 	 */
2946 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2947 
2948 	/*
2949 	 * If the COMPLETE flag is clear, then there were no active
2950 	 * entries and we want to roll back to a zeroed entry until
2951 	 * the new inode is committed to disk. If the COMPLETE flag is
2952 	 * set then we have deleted an entry that never made it to
2953 	 * disk. If the entry we deleted resulted from a name change,
2954 	 * then the old name still resides on disk. We cannot delete
2955 	 * its inode (returned to us in prevdirrem) until the zeroed
2956 	 * directory entry gets to disk. The new inode has never been
2957 	 * referenced on the disk, so can be deleted immediately.
2958 	 */
2959 	if ((dirrem->dm_state & COMPLETE) == 0) {
2960 		LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2961 		    dm_next);
2962 		FREE_LOCK(&lk);
2963 	} else {
2964 		if (prevdirrem != NULL)
2965 			LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2966 			    prevdirrem, dm_next);
2967 		dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2968 		FREE_LOCK(&lk);
2969 		handle_workitem_remove(dirrem, NULL);
2970 	}
2971 }
2972 
2973 /*
2974  * Allocate a new dirrem if appropriate and return it along with
2975  * its associated pagedep. Called without a lock, returns with lock.
2976  */
2977 static long num_dirrem;		/* number of dirrem allocated */
2978 static struct dirrem *
2979 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2980 	struct buf *bp;		/* buffer containing directory block */
2981 	struct inode *dp;	/* inode for the directory being modified */
2982 	struct inode *ip;	/* inode for directory entry being removed */
2983 	int isrmdir;		/* indicates if doing RMDIR */
2984 	struct dirrem **prevdirremp; /* previously referenced inode, if any */
2985 {
2986 	int offset;
2987 	ufs_lbn_t lbn;
2988 	struct diradd *dap;
2989 	struct dirrem *dirrem;
2990 	struct pagedep *pagedep;
2991 
2992 	/*
2993 	 * Whiteouts have no deletion dependencies.
2994 	 */
2995 	if (ip == NULL)
2996 		panic("newdirrem: whiteout");
2997 	/*
2998 	 * If we are over our limit, try to improve the situation.
2999 	 * Limiting the number of dirrem structures will also limit
3000 	 * the number of freefile and freeblks structures.
3001 	 */
3002 	if (num_dirrem > max_softdeps / 2)
3003 		(void) request_cleanup(FLUSH_REMOVE, 0);
3004 	num_dirrem += 1;
3005 	MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
3006 		M_DIRREM, M_SOFTDEP_FLAGS|M_ZERO);
3007 	dirrem->dm_list.wk_type = D_DIRREM;
3008 	dirrem->dm_state = isrmdir ? RMDIR : 0;
3009 	dirrem->dm_mnt = ITOV(ip)->v_mount;
3010 	dirrem->dm_oldinum = ip->i_number;
3011 	*prevdirremp = NULL;
3012 
3013 	ACQUIRE_LOCK(&lk);
3014 	lbn = lblkno(dp->i_fs, dp->i_offset);
3015 	offset = blkoff(dp->i_fs, dp->i_offset);
3016 	if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
3017 		WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
3018 	dirrem->dm_pagedep = pagedep;
3019 	/*
3020 	 * Check for a diradd dependency for the same directory entry.
3021 	 * If present, then both dependencies become obsolete and can
3022 	 * be de-allocated. Check for an entry on both the pd_dirraddhd
3023 	 * list and the pd_pendinghd list.
3024 	 */
3025 
3026 	LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
3027 		if (dap->da_offset == offset)
3028 			break;
3029 	if (dap == NULL) {
3030 
3031 		LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
3032 			if (dap->da_offset == offset)
3033 				break;
3034 		if (dap == NULL)
3035 			return (dirrem);
3036 	}
3037 	/*
3038 	 * Must be ATTACHED at this point.
3039 	 */
3040 	if ((dap->da_state & ATTACHED) == 0) {
3041 		FREE_LOCK(&lk);
3042 		panic("newdirrem: not ATTACHED");
3043 	}
3044 	if (dap->da_newinum != ip->i_number) {
3045 		FREE_LOCK(&lk);
3046 		panic("newdirrem: inum %d should be %d",
3047 		    ip->i_number, dap->da_newinum);
3048 	}
3049 	/*
3050 	 * If we are deleting a changed name that never made it to disk,
3051 	 * then return the dirrem describing the previous inode (which
3052 	 * represents the inode currently referenced from this entry on disk).
3053 	 */
3054 	if ((dap->da_state & DIRCHG) != 0) {
3055 		*prevdirremp = dap->da_previous;
3056 		dap->da_state &= ~DIRCHG;
3057 		dap->da_pagedep = pagedep;
3058 	}
3059 	/*
3060 	 * We are deleting an entry that never made it to disk.
3061 	 * Mark it COMPLETE so we can delete its inode immediately.
3062 	 */
3063 	dirrem->dm_state |= COMPLETE;
3064 	free_diradd(dap);
3065 	return (dirrem);
3066 }
3067 
3068 /*
3069  * Directory entry change dependencies.
3070  *
3071  * Changing an existing directory entry requires that an add operation
3072  * be completed first followed by a deletion. The semantics for the addition
3073  * are identical to the description of adding a new entry above except
3074  * that the rollback is to the old inode number rather than zero. Once
3075  * the addition dependency is completed, the removal is done as described
3076  * in the removal routine above.
3077  */
3078 
3079 /*
3080  * This routine should be called immediately after changing
3081  * a directory entry.  The inode's link count should not be
3082  * decremented by the calling procedure -- the soft updates
3083  * code will perform this task when it is safe.
3084  */
3085 void
3086 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
3087 	struct buf *bp;		/* buffer containing directory block */
3088 	struct inode *dp;	/* inode for the directory being modified */
3089 	struct inode *ip;	/* inode for directory entry being removed */
3090 	ino_t newinum;		/* new inode number for changed entry */
3091 	int isrmdir;		/* indicates if doing RMDIR */
3092 {
3093 	int offset;
3094 	struct diradd *dap = NULL;
3095 	struct dirrem *dirrem, *prevdirrem;
3096 	struct pagedep *pagedep;
3097 	struct inodedep *inodedep;
3098 
3099 	offset = blkoff(dp->i_fs, dp->i_offset);
3100 
3101 	/*
3102 	 * Whiteouts do not need diradd dependencies.
3103 	 */
3104 	if (newinum != WINO) {
3105 		MALLOC(dap, struct diradd *, sizeof(struct diradd),
3106 		    M_DIRADD, M_SOFTDEP_FLAGS|M_ZERO);
3107 		dap->da_list.wk_type = D_DIRADD;
3108 		dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
3109 		dap->da_offset = offset;
3110 		dap->da_newinum = newinum;
3111 	}
3112 
3113 	/*
3114 	 * Allocate a new dirrem and ACQUIRE_LOCK.
3115 	 */
3116 	dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
3117 	pagedep = dirrem->dm_pagedep;
3118 	/*
3119 	 * The possible values for isrmdir:
3120 	 *	0 - non-directory file rename
3121 	 *	1 - directory rename within same directory
3122 	 *   inum - directory rename to new directory of given inode number
3123 	 * When renaming to a new directory, we are both deleting and
3124 	 * creating a new directory entry, so the link count on the new
3125 	 * directory should not change. Thus we do not need the followup
3126 	 * dirrem which is usually done in handle_workitem_remove. We set
3127 	 * the DIRCHG flag to tell handle_workitem_remove to skip the
3128 	 * followup dirrem.
3129 	 */
3130 	if (isrmdir > 1)
3131 		dirrem->dm_state |= DIRCHG;
3132 
3133 	/*
3134 	 * Whiteouts have no additional dependencies,
3135 	 * so just put the dirrem on the correct list.
3136 	 */
3137 	if (newinum == WINO) {
3138 		if ((dirrem->dm_state & COMPLETE) == 0) {
3139 			LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
3140 			    dm_next);
3141 		} else {
3142 			dirrem->dm_dirinum = pagedep->pd_ino;
3143 			add_to_worklist(&dirrem->dm_list);
3144 		}
3145 		FREE_LOCK(&lk);
3146 		return;
3147 	}
3148 
3149 	/*
3150 	 * If the COMPLETE flag is clear, then there were no active
3151 	 * entries and we want to roll back to the previous inode until
3152 	 * the new inode is committed to disk. If the COMPLETE flag is
3153 	 * set, then we have deleted an entry that never made it to disk.
3154 	 * If the entry we deleted resulted from a name change, then the old
3155 	 * inode reference still resides on disk. Any rollback that we do
3156 	 * needs to be to that old inode (returned to us in prevdirrem). If
3157 	 * the entry we deleted resulted from a create, then there is
3158 	 * no entry on the disk, so we want to roll back to zero rather
3159 	 * than the uncommitted inode. In either of the COMPLETE cases we
3160 	 * want to immediately free the unwritten and unreferenced inode.
3161 	 */
3162 	if ((dirrem->dm_state & COMPLETE) == 0) {
3163 		dap->da_previous = dirrem;
3164 	} else {
3165 		if (prevdirrem != NULL) {
3166 			dap->da_previous = prevdirrem;
3167 		} else {
3168 			dap->da_state &= ~DIRCHG;
3169 			dap->da_pagedep = pagedep;
3170 		}
3171 		dirrem->dm_dirinum = pagedep->pd_ino;
3172 		add_to_worklist(&dirrem->dm_list);
3173 	}
3174 	/*
3175 	 * Link into its inodedep. Put it on the id_bufwait list if the inode
3176 	 * is not yet written. If it is written, do the post-inode write
3177 	 * processing to put it on the id_pendinghd list.
3178 	 */
3179 	if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
3180 	    (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
3181 		dap->da_state |= COMPLETE;
3182 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3183 		WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3184 	} else {
3185 		LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
3186 		    dap, da_pdlist);
3187 		WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
3188 	}
3189 	FREE_LOCK(&lk);
3190 }
3191 
3192 /*
3193  * Called whenever the link count on an inode is changed.
3194  * It creates an inode dependency so that the new reference(s)
3195  * to the inode cannot be committed to disk until the updated
3196  * inode has been written.
3197  */
3198 void
3199 softdep_change_linkcnt(ip)
3200 	struct inode *ip;	/* the inode with the increased link count */
3201 {
3202 	struct inodedep *inodedep;
3203 
3204 	ACQUIRE_LOCK(&lk);
3205 	(void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
3206 	if (ip->i_nlink < ip->i_effnlink) {
3207 		FREE_LOCK(&lk);
3208 		panic("softdep_change_linkcnt: bad delta");
3209 	}
3210 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3211 	FREE_LOCK(&lk);
3212 }
3213 
3214 /*
3215  * Called when the effective link count and the reference count
3216  * on an inode drops to zero. At this point there are no names
3217  * referencing the file in the filesystem and no active file
3218  * references. The space associated with the file will be freed
3219  * as soon as the necessary soft dependencies are cleared.
3220  */
3221 void
3222 softdep_releasefile(ip)
3223 	struct inode *ip;	/* inode with the zero effective link count */
3224 {
3225 	struct inodedep *inodedep;
3226 	struct fs *fs;
3227 	int extblocks;
3228 
3229 	if (ip->i_effnlink > 0)
3230 		panic("softdep_filerelease: file still referenced");
3231 	/*
3232 	 * We may be called several times as the real reference count
3233 	 * drops to zero. We only want to account for the space once.
3234 	 */
3235 	if (ip->i_flag & IN_SPACECOUNTED)
3236 		return;
3237 	/*
3238 	 * We have to deactivate a snapshot otherwise copyonwrites may
3239 	 * add blocks and the cleanup may remove blocks after we have
3240 	 * tried to account for them.
3241 	 */
3242 	if ((ip->i_flags & SF_SNAPSHOT) != 0)
3243 		ffs_snapremove(ITOV(ip));
3244 	/*
3245 	 * If we are tracking an nlinkdelta, we have to also remember
3246 	 * whether we accounted for the freed space yet.
3247 	 */
3248 	ACQUIRE_LOCK(&lk);
3249 	if ((inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep)))
3250 		inodedep->id_state |= SPACECOUNTED;
3251 	FREE_LOCK(&lk);
3252 	fs = ip->i_fs;
3253 	extblocks = 0;
3254 	if (fs->fs_magic == FS_UFS2_MAGIC)
3255 		extblocks = btodb(fragroundup(fs, ip->i_din2->di_extsize));
3256 	ip->i_fs->fs_pendingblocks += DIP(ip, i_blocks) - extblocks;
3257 	ip->i_fs->fs_pendinginodes += 1;
3258 	ip->i_flag |= IN_SPACECOUNTED;
3259 }
3260 
3261 /*
3262  * This workitem decrements the inode's link count.
3263  * If the link count reaches zero, the file is removed.
3264  */
3265 static void
3266 handle_workitem_remove(dirrem, xp)
3267 	struct dirrem *dirrem;
3268 	struct vnode *xp;
3269 {
3270 	struct thread *td = curthread;
3271 	struct inodedep *inodedep;
3272 	struct vnode *vp;
3273 	struct inode *ip;
3274 	ino_t oldinum;
3275 	int error;
3276 
3277 	if ((vp = xp) == NULL &&
3278 	    (error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, LK_EXCLUSIVE,
3279 	     &vp)) != 0) {
3280 		softdep_error("handle_workitem_remove: vget", error);
3281 		return;
3282 	}
3283 	ip = VTOI(vp);
3284 	ACQUIRE_LOCK(&lk);
3285 	if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
3286 		FREE_LOCK(&lk);
3287 		panic("handle_workitem_remove: lost inodedep");
3288 	}
3289 	/*
3290 	 * Normal file deletion.
3291 	 */
3292 	if ((dirrem->dm_state & RMDIR) == 0) {
3293 		ip->i_nlink--;
3294 		DIP(ip, i_nlink) = ip->i_nlink;
3295 		ip->i_flag |= IN_CHANGE;
3296 		if (ip->i_nlink < ip->i_effnlink) {
3297 			FREE_LOCK(&lk);
3298 			panic("handle_workitem_remove: bad file delta");
3299 		}
3300 		inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3301 		FREE_LOCK(&lk);
3302 		vput(vp);
3303 		num_dirrem -= 1;
3304 		WORKITEM_FREE(dirrem, D_DIRREM);
3305 		return;
3306 	}
3307 	/*
3308 	 * Directory deletion. Decrement reference count for both the
3309 	 * just deleted parent directory entry and the reference for ".".
3310 	 * Next truncate the directory to length zero. When the
3311 	 * truncation completes, arrange to have the reference count on
3312 	 * the parent decremented to account for the loss of "..".
3313 	 */
3314 	ip->i_nlink -= 2;
3315 	DIP(ip, i_nlink) = ip->i_nlink;
3316 	ip->i_flag |= IN_CHANGE;
3317 	if (ip->i_nlink < ip->i_effnlink) {
3318 		FREE_LOCK(&lk);
3319 		panic("handle_workitem_remove: bad dir delta");
3320 	}
3321 	inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3322 	FREE_LOCK(&lk);
3323 	if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, td->td_ucred, td)) != 0)
3324 		softdep_error("handle_workitem_remove: truncate", error);
3325 	/*
3326 	 * Rename a directory to a new parent. Since, we are both deleting
3327 	 * and creating a new directory entry, the link count on the new
3328 	 * directory should not change. Thus we skip the followup dirrem.
3329 	 */
3330 	if (dirrem->dm_state & DIRCHG) {
3331 		vput(vp);
3332 		num_dirrem -= 1;
3333 		WORKITEM_FREE(dirrem, D_DIRREM);
3334 		return;
3335 	}
3336 	/*
3337 	 * If the inodedep does not exist, then the zero'ed inode has
3338 	 * been written to disk. If the allocated inode has never been
3339 	 * written to disk, then the on-disk inode is zero'ed. In either
3340 	 * case we can remove the file immediately.
3341 	 */
3342 	ACQUIRE_LOCK(&lk);
3343 	dirrem->dm_state = 0;
3344 	oldinum = dirrem->dm_oldinum;
3345 	dirrem->dm_oldinum = dirrem->dm_dirinum;
3346 	if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
3347 	    check_inode_unwritten(inodedep)) {
3348 		FREE_LOCK(&lk);
3349 		vput(vp);
3350 		handle_workitem_remove(dirrem, NULL);
3351 		return;
3352 	}
3353 	WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3354 	FREE_LOCK(&lk);
3355 	vput(vp);
3356 }
3357 
3358 /*
3359  * Inode de-allocation dependencies.
3360  *
3361  * When an inode's link count is reduced to zero, it can be de-allocated. We
3362  * found it convenient to postpone de-allocation until after the inode is
3363  * written to disk with its new link count (zero).  At this point, all of the
3364  * on-disk inode's block pointers are nullified and, with careful dependency
3365  * list ordering, all dependencies related to the inode will be satisfied and
3366  * the corresponding dependency structures de-allocated.  So, if/when the
3367  * inode is reused, there will be no mixing of old dependencies with new
3368  * ones.  This artificial dependency is set up by the block de-allocation
3369  * procedure above (softdep_setup_freeblocks) and completed by the
3370  * following procedure.
3371  */
3372 static void
3373 handle_workitem_freefile(freefile)
3374 	struct freefile *freefile;
3375 {
3376 	struct fs *fs;
3377 	struct inodedep *idp;
3378 	int error;
3379 
3380 	fs = VFSTOUFS(freefile->fx_mnt)->um_fs;
3381 #ifdef DEBUG
3382 	ACQUIRE_LOCK(&lk);
3383 	error = inodedep_lookup(fs, freefile->fx_oldinum, 0, &idp);
3384 	FREE_LOCK(&lk);
3385 	if (error)
3386 		panic("handle_workitem_freefile: inodedep survived");
3387 #endif
3388 	fs->fs_pendinginodes -= 1;
3389 	if ((error = ffs_freefile(fs, freefile->fx_devvp, freefile->fx_oldinum,
3390 	     freefile->fx_mode)) != 0)
3391 		softdep_error("handle_workitem_freefile", error);
3392 	WORKITEM_FREE(freefile, D_FREEFILE);
3393 }
3394 
3395 /*
3396  * Disk writes.
3397  *
3398  * The dependency structures constructed above are most actively used when file
3399  * system blocks are written to disk.  No constraints are placed on when a
3400  * block can be written, but unsatisfied update dependencies are made safe by
3401  * modifying (or replacing) the source memory for the duration of the disk
3402  * write.  When the disk write completes, the memory block is again brought
3403  * up-to-date.
3404  *
3405  * In-core inode structure reclamation.
3406  *
3407  * Because there are a finite number of "in-core" inode structures, they are
3408  * reused regularly.  By transferring all inode-related dependencies to the
3409  * in-memory inode block and indexing them separately (via "inodedep"s), we
3410  * can allow "in-core" inode structures to be reused at any time and avoid
3411  * any increase in contention.
3412  *
3413  * Called just before entering the device driver to initiate a new disk I/O.
3414  * The buffer must be locked, thus, no I/O completion operations can occur
3415  * while we are manipulating its associated dependencies.
3416  */
3417 static void
3418 softdep_disk_io_initiation(bp)
3419 	struct buf *bp;		/* structure describing disk write to occur */
3420 {
3421 	struct worklist *wk, *nextwk;
3422 	struct indirdep *indirdep;
3423 	struct inodedep *inodedep;
3424 
3425 	/*
3426 	 * We only care about write operations. There should never
3427 	 * be dependencies for reads.
3428 	 */
3429 	if (bp->b_iocmd == BIO_READ)
3430 		panic("softdep_disk_io_initiation: read");
3431 	/*
3432 	 * Do any necessary pre-I/O processing.
3433 	 */
3434 	for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
3435 		nextwk = LIST_NEXT(wk, wk_list);
3436 		switch (wk->wk_type) {
3437 
3438 		case D_PAGEDEP:
3439 			initiate_write_filepage(WK_PAGEDEP(wk), bp);
3440 			continue;
3441 
3442 		case D_INODEDEP:
3443 			inodedep = WK_INODEDEP(wk);
3444 			if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC)
3445 				initiate_write_inodeblock_ufs1(inodedep, bp);
3446 			else
3447 				initiate_write_inodeblock_ufs2(inodedep, bp);
3448 			continue;
3449 
3450 		case D_INDIRDEP:
3451 			indirdep = WK_INDIRDEP(wk);
3452 			if (indirdep->ir_state & GOINGAWAY)
3453 				panic("disk_io_initiation: indirdep gone");
3454 			/*
3455 			 * If there are no remaining dependencies, this
3456 			 * will be writing the real pointers, so the
3457 			 * dependency can be freed.
3458 			 */
3459 			if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
3460 				indirdep->ir_savebp->b_flags |=
3461 				    B_INVAL | B_NOCACHE;
3462 				brelse(indirdep->ir_savebp);
3463 				/* inline expand WORKLIST_REMOVE(wk); */
3464 				wk->wk_state &= ~ONWORKLIST;
3465 				LIST_REMOVE(wk, wk_list);
3466 				WORKITEM_FREE(indirdep, D_INDIRDEP);
3467 				continue;
3468 			}
3469 			/*
3470 			 * Replace up-to-date version with safe version.
3471 			 */
3472 			MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3473 			    M_INDIRDEP, M_SOFTDEP_FLAGS);
3474 			ACQUIRE_LOCK(&lk);
3475 			indirdep->ir_state &= ~ATTACHED;
3476 			indirdep->ir_state |= UNDONE;
3477 			bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3478 			bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3479 			    bp->b_bcount);
3480 			FREE_LOCK(&lk);
3481 			continue;
3482 
3483 		case D_MKDIR:
3484 		case D_BMSAFEMAP:
3485 		case D_ALLOCDIRECT:
3486 		case D_ALLOCINDIR:
3487 			continue;
3488 
3489 		default:
3490 			panic("handle_disk_io_initiation: Unexpected type %s",
3491 			    TYPENAME(wk->wk_type));
3492 			/* NOTREACHED */
3493 		}
3494 	}
3495 }
3496 
3497 /*
3498  * Called from within the procedure above to deal with unsatisfied
3499  * allocation dependencies in a directory. The buffer must be locked,
3500  * thus, no I/O completion operations can occur while we are
3501  * manipulating its associated dependencies.
3502  */
3503 static void
3504 initiate_write_filepage(pagedep, bp)
3505 	struct pagedep *pagedep;
3506 	struct buf *bp;
3507 {
3508 	struct diradd *dap;
3509 	struct direct *ep;
3510 	int i;
3511 
3512 	if (pagedep->pd_state & IOSTARTED) {
3513 		/*
3514 		 * This can only happen if there is a driver that does not
3515 		 * understand chaining. Here biodone will reissue the call
3516 		 * to strategy for the incomplete buffers.
3517 		 */
3518 		printf("initiate_write_filepage: already started\n");
3519 		return;
3520 	}
3521 	pagedep->pd_state |= IOSTARTED;
3522 	ACQUIRE_LOCK(&lk);
3523 	for (i = 0; i < DAHASHSZ; i++) {
3524 		LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3525 			ep = (struct direct *)
3526 			    ((char *)bp->b_data + dap->da_offset);
3527 			if (ep->d_ino != dap->da_newinum) {
3528 				FREE_LOCK(&lk);
3529 				panic("%s: dir inum %d != new %d",
3530 				    "initiate_write_filepage",
3531 				    ep->d_ino, dap->da_newinum);
3532 			}
3533 			if (dap->da_state & DIRCHG)
3534 				ep->d_ino = dap->da_previous->dm_oldinum;
3535 			else
3536 				ep->d_ino = 0;
3537 			dap->da_state &= ~ATTACHED;
3538 			dap->da_state |= UNDONE;
3539 		}
3540 	}
3541 	FREE_LOCK(&lk);
3542 }
3543 
3544 /*
3545  * Version of initiate_write_inodeblock that handles UFS1 dinodes.
3546  * Note that any bug fixes made to this routine must be done in the
3547  * version found below.
3548  *
3549  * Called from within the procedure above to deal with unsatisfied
3550  * allocation dependencies in an inodeblock. The buffer must be
3551  * locked, thus, no I/O completion operations can occur while we
3552  * are manipulating its associated dependencies.
3553  */
3554 static void
3555 initiate_write_inodeblock_ufs1(inodedep, bp)
3556 	struct inodedep *inodedep;
3557 	struct buf *bp;			/* The inode block */
3558 {
3559 	struct allocdirect *adp, *lastadp;
3560 	struct ufs1_dinode *dp;
3561 	struct fs *fs;
3562 	ufs_lbn_t i, prevlbn = 0;
3563 	int deplist;
3564 
3565 	if (inodedep->id_state & IOSTARTED)
3566 		panic("initiate_write_inodeblock_ufs1: already started");
3567 	inodedep->id_state |= IOSTARTED;
3568 	fs = inodedep->id_fs;
3569 	dp = (struct ufs1_dinode *)bp->b_data +
3570 	    ino_to_fsbo(fs, inodedep->id_ino);
3571 	/*
3572 	 * If the bitmap is not yet written, then the allocated
3573 	 * inode cannot be written to disk.
3574 	 */
3575 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3576 		if (inodedep->id_savedino1 != NULL)
3577 			panic("initiate_write_inodeblock_ufs1: I/O underway");
3578 		MALLOC(inodedep->id_savedino1, struct ufs1_dinode *,
3579 		    sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3580 		*inodedep->id_savedino1 = *dp;
3581 		bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
3582 		return;
3583 	}
3584 	/*
3585 	 * If no dependencies, then there is nothing to roll back.
3586 	 */
3587 	inodedep->id_savedsize = dp->di_size;
3588 	inodedep->id_savedextsize = 0;
3589 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3590 		return;
3591 	/*
3592 	 * Set the dependencies to busy.
3593 	 */
3594 	ACQUIRE_LOCK(&lk);
3595 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3596 	     adp = TAILQ_NEXT(adp, ad_next)) {
3597 #ifdef DIAGNOSTIC
3598 		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3599 			FREE_LOCK(&lk);
3600 			panic("softdep_write_inodeblock: lbn order");
3601 		}
3602 		prevlbn = adp->ad_lbn;
3603 		if (adp->ad_lbn < NDADDR &&
3604 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3605 			FREE_LOCK(&lk);
3606 			panic("%s: direct pointer #%jd mismatch %d != %jd",
3607 			    "softdep_write_inodeblock",
3608 			    (intmax_t)adp->ad_lbn,
3609 			    dp->di_db[adp->ad_lbn],
3610 			    (intmax_t)adp->ad_newblkno);
3611 		}
3612 		if (adp->ad_lbn >= NDADDR &&
3613 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3614 			FREE_LOCK(&lk);
3615 			panic("%s: indirect pointer #%jd mismatch %d != %jd",
3616 			    "softdep_write_inodeblock",
3617 			    (intmax_t)adp->ad_lbn - NDADDR,
3618 			    dp->di_ib[adp->ad_lbn - NDADDR],
3619 			    (intmax_t)adp->ad_newblkno);
3620 		}
3621 		deplist |= 1 << adp->ad_lbn;
3622 		if ((adp->ad_state & ATTACHED) == 0) {
3623 			FREE_LOCK(&lk);
3624 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3625 			    adp->ad_state);
3626 		}
3627 #endif /* DIAGNOSTIC */
3628 		adp->ad_state &= ~ATTACHED;
3629 		adp->ad_state |= UNDONE;
3630 	}
3631 	/*
3632 	 * The on-disk inode cannot claim to be any larger than the last
3633 	 * fragment that has been written. Otherwise, the on-disk inode
3634 	 * might have fragments that were not the last block in the file
3635 	 * which would corrupt the filesystem.
3636 	 */
3637 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3638 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3639 		if (adp->ad_lbn >= NDADDR)
3640 			break;
3641 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3642 		/* keep going until hitting a rollback to a frag */
3643 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3644 			continue;
3645 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3646 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3647 #ifdef DIAGNOSTIC
3648 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3649 				FREE_LOCK(&lk);
3650 				panic("softdep_write_inodeblock: lost dep1");
3651 			}
3652 #endif /* DIAGNOSTIC */
3653 			dp->di_db[i] = 0;
3654 		}
3655 		for (i = 0; i < NIADDR; i++) {
3656 #ifdef DIAGNOSTIC
3657 			if (dp->di_ib[i] != 0 &&
3658 			    (deplist & ((1 << NDADDR) << i)) == 0) {
3659 				FREE_LOCK(&lk);
3660 				panic("softdep_write_inodeblock: lost dep2");
3661 			}
3662 #endif /* DIAGNOSTIC */
3663 			dp->di_ib[i] = 0;
3664 		}
3665 		FREE_LOCK(&lk);
3666 		return;
3667 	}
3668 	/*
3669 	 * If we have zero'ed out the last allocated block of the file,
3670 	 * roll back the size to the last currently allocated block.
3671 	 * We know that this last allocated block is a full-sized as
3672 	 * we already checked for fragments in the loop above.
3673 	 */
3674 	if (lastadp != NULL &&
3675 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3676 		for (i = lastadp->ad_lbn; i >= 0; i--)
3677 			if (dp->di_db[i] != 0)
3678 				break;
3679 		dp->di_size = (i + 1) * fs->fs_bsize;
3680 	}
3681 	/*
3682 	 * The only dependencies are for indirect blocks.
3683 	 *
3684 	 * The file size for indirect block additions is not guaranteed.
3685 	 * Such a guarantee would be non-trivial to achieve. The conventional
3686 	 * synchronous write implementation also does not make this guarantee.
3687 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3688 	 * can be over-estimated without destroying integrity when the file
3689 	 * moves into the indirect blocks (i.e., is large). If we want to
3690 	 * postpone fsck, we are stuck with this argument.
3691 	 */
3692 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3693 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3694 	FREE_LOCK(&lk);
3695 }
3696 
3697 /*
3698  * Version of initiate_write_inodeblock that handles UFS2 dinodes.
3699  * Note that any bug fixes made to this routine must be done in the
3700  * version found above.
3701  *
3702  * Called from within the procedure above to deal with unsatisfied
3703  * allocation dependencies in an inodeblock. The buffer must be
3704  * locked, thus, no I/O completion operations can occur while we
3705  * are manipulating its associated dependencies.
3706  */
3707 static void
3708 initiate_write_inodeblock_ufs2(inodedep, bp)
3709 	struct inodedep *inodedep;
3710 	struct buf *bp;			/* The inode block */
3711 {
3712 	struct allocdirect *adp, *lastadp;
3713 	struct ufs2_dinode *dp;
3714 	struct fs *fs;
3715 	ufs_lbn_t i, prevlbn = 0;
3716 	int deplist;
3717 
3718 	if (inodedep->id_state & IOSTARTED)
3719 		panic("initiate_write_inodeblock_ufs2: already started");
3720 	inodedep->id_state |= IOSTARTED;
3721 	fs = inodedep->id_fs;
3722 	dp = (struct ufs2_dinode *)bp->b_data +
3723 	    ino_to_fsbo(fs, inodedep->id_ino);
3724 	/*
3725 	 * If the bitmap is not yet written, then the allocated
3726 	 * inode cannot be written to disk.
3727 	 */
3728 	if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3729 		if (inodedep->id_savedino2 != NULL)
3730 			panic("initiate_write_inodeblock_ufs2: I/O underway");
3731 		MALLOC(inodedep->id_savedino2, struct ufs2_dinode *,
3732 		    sizeof(struct ufs2_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3733 		*inodedep->id_savedino2 = *dp;
3734 		bzero((caddr_t)dp, sizeof(struct ufs2_dinode));
3735 		return;
3736 	}
3737 	/*
3738 	 * If no dependencies, then there is nothing to roll back.
3739 	 */
3740 	inodedep->id_savedsize = dp->di_size;
3741 	inodedep->id_savedextsize = dp->di_extsize;
3742 	if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL &&
3743 	    TAILQ_FIRST(&inodedep->id_extupdt) == NULL)
3744 		return;
3745 	/*
3746 	 * Set the ext data dependencies to busy.
3747 	 */
3748 	ACQUIRE_LOCK(&lk);
3749 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
3750 	     adp = TAILQ_NEXT(adp, ad_next)) {
3751 #ifdef DIAGNOSTIC
3752 		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3753 			FREE_LOCK(&lk);
3754 			panic("softdep_write_inodeblock: lbn order");
3755 		}
3756 		prevlbn = adp->ad_lbn;
3757 		if (dp->di_extb[adp->ad_lbn] != adp->ad_newblkno) {
3758 			FREE_LOCK(&lk);
3759 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
3760 			    "softdep_write_inodeblock",
3761 			    (intmax_t)adp->ad_lbn,
3762 			    (intmax_t)dp->di_extb[adp->ad_lbn],
3763 			    (intmax_t)adp->ad_newblkno);
3764 		}
3765 		deplist |= 1 << adp->ad_lbn;
3766 		if ((adp->ad_state & ATTACHED) == 0) {
3767 			FREE_LOCK(&lk);
3768 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3769 			    adp->ad_state);
3770 		}
3771 #endif /* DIAGNOSTIC */
3772 		adp->ad_state &= ~ATTACHED;
3773 		adp->ad_state |= UNDONE;
3774 	}
3775 	/*
3776 	 * The on-disk inode cannot claim to be any larger than the last
3777 	 * fragment that has been written. Otherwise, the on-disk inode
3778 	 * might have fragments that were not the last block in the ext
3779 	 * data which would corrupt the filesystem.
3780 	 */
3781 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_extupdt); adp;
3782 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3783 		dp->di_extb[adp->ad_lbn] = adp->ad_oldblkno;
3784 		/* keep going until hitting a rollback to a frag */
3785 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3786 			continue;
3787 		dp->di_extsize = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3788 		for (i = adp->ad_lbn + 1; i < NXADDR; i++) {
3789 #ifdef DIAGNOSTIC
3790 			if (dp->di_extb[i] != 0 && (deplist & (1 << i)) == 0) {
3791 				FREE_LOCK(&lk);
3792 				panic("softdep_write_inodeblock: lost dep1");
3793 			}
3794 #endif /* DIAGNOSTIC */
3795 			dp->di_extb[i] = 0;
3796 		}
3797 		lastadp = NULL;
3798 		break;
3799 	}
3800 	/*
3801 	 * If we have zero'ed out the last allocated block of the ext
3802 	 * data, roll back the size to the last currently allocated block.
3803 	 * We know that this last allocated block is a full-sized as
3804 	 * we already checked for fragments in the loop above.
3805 	 */
3806 	if (lastadp != NULL &&
3807 	    dp->di_extsize <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3808 		for (i = lastadp->ad_lbn; i >= 0; i--)
3809 			if (dp->di_extb[i] != 0)
3810 				break;
3811 		dp->di_extsize = (i + 1) * fs->fs_bsize;
3812 	}
3813 	/*
3814 	 * Set the file data dependencies to busy.
3815 	 */
3816 	for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3817 	     adp = TAILQ_NEXT(adp, ad_next)) {
3818 #ifdef DIAGNOSTIC
3819 		if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3820 			FREE_LOCK(&lk);
3821 			panic("softdep_write_inodeblock: lbn order");
3822 		}
3823 		prevlbn = adp->ad_lbn;
3824 		if (adp->ad_lbn < NDADDR &&
3825 		    dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3826 			FREE_LOCK(&lk);
3827 			panic("%s: direct pointer #%jd mismatch %jd != %jd",
3828 			    "softdep_write_inodeblock",
3829 			    (intmax_t)adp->ad_lbn,
3830 			    (intmax_t)dp->di_db[adp->ad_lbn],
3831 			    (intmax_t)adp->ad_newblkno);
3832 		}
3833 		if (adp->ad_lbn >= NDADDR &&
3834 		    dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3835 			FREE_LOCK(&lk);
3836 			panic("%s indirect pointer #%jd mismatch %jd != %jd",
3837 			    "softdep_write_inodeblock:",
3838 			    (intmax_t)adp->ad_lbn - NDADDR,
3839 			    (intmax_t)dp->di_ib[adp->ad_lbn - NDADDR],
3840 			    (intmax_t)adp->ad_newblkno);
3841 		}
3842 		deplist |= 1 << adp->ad_lbn;
3843 		if ((adp->ad_state & ATTACHED) == 0) {
3844 			FREE_LOCK(&lk);
3845 			panic("softdep_write_inodeblock: Unknown state 0x%x",
3846 			    adp->ad_state);
3847 		}
3848 #endif /* DIAGNOSTIC */
3849 		adp->ad_state &= ~ATTACHED;
3850 		adp->ad_state |= UNDONE;
3851 	}
3852 	/*
3853 	 * The on-disk inode cannot claim to be any larger than the last
3854 	 * fragment that has been written. Otherwise, the on-disk inode
3855 	 * might have fragments that were not the last block in the file
3856 	 * which would corrupt the filesystem.
3857 	 */
3858 	for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3859 	     lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3860 		if (adp->ad_lbn >= NDADDR)
3861 			break;
3862 		dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3863 		/* keep going until hitting a rollback to a frag */
3864 		if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3865 			continue;
3866 		dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3867 		for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3868 #ifdef DIAGNOSTIC
3869 			if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3870 				FREE_LOCK(&lk);
3871 				panic("softdep_write_inodeblock: lost dep2");
3872 			}
3873 #endif /* DIAGNOSTIC */
3874 			dp->di_db[i] = 0;
3875 		}
3876 		for (i = 0; i < NIADDR; i++) {
3877 #ifdef DIAGNOSTIC
3878 			if (dp->di_ib[i] != 0 &&
3879 			    (deplist & ((1 << NDADDR) << i)) == 0) {
3880 				FREE_LOCK(&lk);
3881 				panic("softdep_write_inodeblock: lost dep3");
3882 			}
3883 #endif /* DIAGNOSTIC */
3884 			dp->di_ib[i] = 0;
3885 		}
3886 		FREE_LOCK(&lk);
3887 		return;
3888 	}
3889 	/*
3890 	 * If we have zero'ed out the last allocated block of the file,
3891 	 * roll back the size to the last currently allocated block.
3892 	 * We know that this last allocated block is a full-sized as
3893 	 * we already checked for fragments in the loop above.
3894 	 */
3895 	if (lastadp != NULL &&
3896 	    dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3897 		for (i = lastadp->ad_lbn; i >= 0; i--)
3898 			if (dp->di_db[i] != 0)
3899 				break;
3900 		dp->di_size = (i + 1) * fs->fs_bsize;
3901 	}
3902 	/*
3903 	 * The only dependencies are for indirect blocks.
3904 	 *
3905 	 * The file size for indirect block additions is not guaranteed.
3906 	 * Such a guarantee would be non-trivial to achieve. The conventional
3907 	 * synchronous write implementation also does not make this guarantee.
3908 	 * Fsck should catch and fix discrepancies. Arguably, the file size
3909 	 * can be over-estimated without destroying integrity when the file
3910 	 * moves into the indirect blocks (i.e., is large). If we want to
3911 	 * postpone fsck, we are stuck with this argument.
3912 	 */
3913 	for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3914 		dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3915 	FREE_LOCK(&lk);
3916 }
3917 
3918 /*
3919  * This routine is called during the completion interrupt
3920  * service routine for a disk write (from the procedure called
3921  * by the device driver to inform the filesystem caches of
3922  * a request completion).  It should be called early in this
3923  * procedure, before the block is made available to other
3924  * processes or other routines are called.
3925  */
3926 static void
3927 softdep_disk_write_complete(bp)
3928 	struct buf *bp;		/* describes the completed disk write */
3929 {
3930 	struct worklist *wk;
3931 	struct workhead reattach;
3932 	struct newblk *newblk;
3933 	struct allocindir *aip;
3934 	struct allocdirect *adp;
3935 	struct indirdep *indirdep;
3936 	struct inodedep *inodedep;
3937 	struct bmsafemap *bmsafemap;
3938 
3939 	/*
3940 	 * If an error occurred while doing the write, then the data
3941 	 * has not hit the disk and the dependencies cannot be unrolled.
3942 	 */
3943 	if ((bp->b_ioflags & BIO_ERROR) != 0 && (bp->b_flags & B_INVAL) == 0)
3944 		return;
3945 #ifdef DEBUG
3946 	if (lk.lkt_held != NOHOLDER)
3947 		panic("softdep_disk_write_complete: lock is held");
3948 	lk.lkt_held = SPECIAL_FLAG;
3949 #endif
3950 	LIST_INIT(&reattach);
3951 	while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3952 		WORKLIST_REMOVE(wk);
3953 		switch (wk->wk_type) {
3954 
3955 		case D_PAGEDEP:
3956 			if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3957 				WORKLIST_INSERT(&reattach, wk);
3958 			continue;
3959 
3960 		case D_INODEDEP:
3961 			if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3962 				WORKLIST_INSERT(&reattach, wk);
3963 			continue;
3964 
3965 		case D_BMSAFEMAP:
3966 			bmsafemap = WK_BMSAFEMAP(wk);
3967 			while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3968 				newblk->nb_state |= DEPCOMPLETE;
3969 				newblk->nb_bmsafemap = NULL;
3970 				LIST_REMOVE(newblk, nb_deps);
3971 			}
3972 			while ((adp =
3973 			   LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3974 				adp->ad_state |= DEPCOMPLETE;
3975 				adp->ad_buf = NULL;
3976 				LIST_REMOVE(adp, ad_deps);
3977 				handle_allocdirect_partdone(adp);
3978 			}
3979 			while ((aip =
3980 			    LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3981 				aip->ai_state |= DEPCOMPLETE;
3982 				aip->ai_buf = NULL;
3983 				LIST_REMOVE(aip, ai_deps);
3984 				handle_allocindir_partdone(aip);
3985 			}
3986 			while ((inodedep =
3987 			     LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3988 				inodedep->id_state |= DEPCOMPLETE;
3989 				LIST_REMOVE(inodedep, id_deps);
3990 				inodedep->id_buf = NULL;
3991 			}
3992 			WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3993 			continue;
3994 
3995 		case D_MKDIR:
3996 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3997 			continue;
3998 
3999 		case D_ALLOCDIRECT:
4000 			adp = WK_ALLOCDIRECT(wk);
4001 			adp->ad_state |= COMPLETE;
4002 			handle_allocdirect_partdone(adp);
4003 			continue;
4004 
4005 		case D_ALLOCINDIR:
4006 			aip = WK_ALLOCINDIR(wk);
4007 			aip->ai_state |= COMPLETE;
4008 			handle_allocindir_partdone(aip);
4009 			continue;
4010 
4011 		case D_INDIRDEP:
4012 			indirdep = WK_INDIRDEP(wk);
4013 			if (indirdep->ir_state & GOINGAWAY) {
4014 				lk.lkt_held = NOHOLDER;
4015 				panic("disk_write_complete: indirdep gone");
4016 			}
4017 			bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
4018 			FREE(indirdep->ir_saveddata, M_INDIRDEP);
4019 			indirdep->ir_saveddata = 0;
4020 			indirdep->ir_state &= ~UNDONE;
4021 			indirdep->ir_state |= ATTACHED;
4022 			while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
4023 				handle_allocindir_partdone(aip);
4024 				if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
4025 					lk.lkt_held = NOHOLDER;
4026 					panic("disk_write_complete: not gone");
4027 				}
4028 			}
4029 			WORKLIST_INSERT(&reattach, wk);
4030 			if ((bp->b_flags & B_DELWRI) == 0)
4031 				stat_indir_blk_ptrs++;
4032 			bdirty(bp);
4033 			continue;
4034 
4035 		default:
4036 			lk.lkt_held = NOHOLDER;
4037 			panic("handle_disk_write_complete: Unknown type %s",
4038 			    TYPENAME(wk->wk_type));
4039 			/* NOTREACHED */
4040 		}
4041 	}
4042 	/*
4043 	 * Reattach any requests that must be redone.
4044 	 */
4045 	while ((wk = LIST_FIRST(&reattach)) != NULL) {
4046 		WORKLIST_REMOVE(wk);
4047 		WORKLIST_INSERT(&bp->b_dep, wk);
4048 	}
4049 #ifdef DEBUG
4050 	if (lk.lkt_held != SPECIAL_FLAG)
4051 		panic("softdep_disk_write_complete: lock lost");
4052 	lk.lkt_held = NOHOLDER;
4053 #endif
4054 }
4055 
4056 /*
4057  * Called from within softdep_disk_write_complete above. Note that
4058  * this routine is always called from interrupt level with further
4059  * splbio interrupts blocked.
4060  */
4061 static void
4062 handle_allocdirect_partdone(adp)
4063 	struct allocdirect *adp;	/* the completed allocdirect */
4064 {
4065 	struct allocdirectlst *listhead;
4066 	struct allocdirect *listadp;
4067 	struct inodedep *inodedep;
4068 	long bsize, delay;
4069 
4070 	if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4071 		return;
4072 	if (adp->ad_buf != NULL) {
4073 		lk.lkt_held = NOHOLDER;
4074 		panic("handle_allocdirect_partdone: dangling dep");
4075 	}
4076 	/*
4077 	 * The on-disk inode cannot claim to be any larger than the last
4078 	 * fragment that has been written. Otherwise, the on-disk inode
4079 	 * might have fragments that were not the last block in the file
4080 	 * which would corrupt the filesystem. Thus, we cannot free any
4081 	 * allocdirects after one whose ad_oldblkno claims a fragment as
4082 	 * these blocks must be rolled back to zero before writing the inode.
4083 	 * We check the currently active set of allocdirects in id_inoupdt
4084 	 * or id_extupdt as appropriate.
4085 	 */
4086 	inodedep = adp->ad_inodedep;
4087 	bsize = inodedep->id_fs->fs_bsize;
4088 	if (adp->ad_state & EXTDATA)
4089 		listhead = &inodedep->id_extupdt;
4090 	else
4091 		listhead = &inodedep->id_inoupdt;
4092 	TAILQ_FOREACH(listadp, listhead, ad_next) {
4093 		/* found our block */
4094 		if (listadp == adp)
4095 			break;
4096 		/* continue if ad_oldlbn is not a fragment */
4097 		if (listadp->ad_oldsize == 0 ||
4098 		    listadp->ad_oldsize == bsize)
4099 			continue;
4100 		/* hit a fragment */
4101 		return;
4102 	}
4103 	/*
4104 	 * If we have reached the end of the current list without
4105 	 * finding the just finished dependency, then it must be
4106 	 * on the future dependency list. Future dependencies cannot
4107 	 * be freed until they are moved to the current list.
4108 	 */
4109 	if (listadp == NULL) {
4110 #ifdef DEBUG
4111 		if (adp->ad_state & EXTDATA)
4112 			listhead = &inodedep->id_newextupdt;
4113 		else
4114 			listhead = &inodedep->id_newinoupdt;
4115 		TAILQ_FOREACH(listadp, listhead, ad_next)
4116 			/* found our block */
4117 			if (listadp == adp)
4118 				break;
4119 		if (listadp == NULL) {
4120 			lk.lkt_held = NOHOLDER;
4121 			panic("handle_allocdirect_partdone: lost dep");
4122 		}
4123 #endif /* DEBUG */
4124 		return;
4125 	}
4126 	/*
4127 	 * If we have found the just finished dependency, then free
4128 	 * it along with anything that follows it that is complete.
4129 	 * If the inode still has a bitmap dependency, then it has
4130 	 * never been written to disk, hence the on-disk inode cannot
4131 	 * reference the old fragment so we can free it without delay.
4132 	 */
4133 	delay = (inodedep->id_state & DEPCOMPLETE);
4134 	for (; adp; adp = listadp) {
4135 		listadp = TAILQ_NEXT(adp, ad_next);
4136 		if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
4137 			return;
4138 		free_allocdirect(listhead, adp, delay);
4139 	}
4140 }
4141 
4142 /*
4143  * Called from within softdep_disk_write_complete above. Note that
4144  * this routine is always called from interrupt level with further
4145  * splbio interrupts blocked.
4146  */
4147 static void
4148 handle_allocindir_partdone(aip)
4149 	struct allocindir *aip;		/* the completed allocindir */
4150 {
4151 	struct indirdep *indirdep;
4152 
4153 	if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
4154 		return;
4155 	if (aip->ai_buf != NULL) {
4156 		lk.lkt_held = NOHOLDER;
4157 		panic("handle_allocindir_partdone: dangling dependency");
4158 	}
4159 	indirdep = aip->ai_indirdep;
4160 	if (indirdep->ir_state & UNDONE) {
4161 		LIST_REMOVE(aip, ai_next);
4162 		LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
4163 		return;
4164 	}
4165 	if (indirdep->ir_state & UFS1FMT)
4166 		((ufs1_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4167 		    aip->ai_newblkno;
4168 	else
4169 		((ufs2_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
4170 		    aip->ai_newblkno;
4171 	LIST_REMOVE(aip, ai_next);
4172 	if (aip->ai_freefrag != NULL)
4173 		add_to_worklist(&aip->ai_freefrag->ff_list);
4174 	WORKITEM_FREE(aip, D_ALLOCINDIR);
4175 }
4176 
4177 /*
4178  * Called from within softdep_disk_write_complete above to restore
4179  * in-memory inode block contents to their most up-to-date state. Note
4180  * that this routine is always called from interrupt level with further
4181  * splbio interrupts blocked.
4182  */
4183 static int
4184 handle_written_inodeblock(inodedep, bp)
4185 	struct inodedep *inodedep;
4186 	struct buf *bp;		/* buffer containing the inode block */
4187 {
4188 	struct worklist *wk, *filefree;
4189 	struct allocdirect *adp, *nextadp;
4190 	struct ufs1_dinode *dp1 = NULL;
4191 	struct ufs2_dinode *dp2 = NULL;
4192 	int hadchanges, fstype;
4193 
4194 	if ((inodedep->id_state & IOSTARTED) == 0) {
4195 		lk.lkt_held = NOHOLDER;
4196 		panic("handle_written_inodeblock: not started");
4197 	}
4198 	inodedep->id_state &= ~IOSTARTED;
4199 	inodedep->id_state |= COMPLETE;
4200 	if (inodedep->id_fs->fs_magic == FS_UFS1_MAGIC) {
4201 		fstype = UFS1;
4202 		dp1 = (struct ufs1_dinode *)bp->b_data +
4203 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4204 	} else {
4205 		fstype = UFS2;
4206 		dp2 = (struct ufs2_dinode *)bp->b_data +
4207 		    ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
4208 	}
4209 	/*
4210 	 * If we had to rollback the inode allocation because of
4211 	 * bitmaps being incomplete, then simply restore it.
4212 	 * Keep the block dirty so that it will not be reclaimed until
4213 	 * all associated dependencies have been cleared and the
4214 	 * corresponding updates written to disk.
4215 	 */
4216 	if (inodedep->id_savedino1 != NULL) {
4217 		if (fstype == UFS1)
4218 			*dp1 = *inodedep->id_savedino1;
4219 		else
4220 			*dp2 = *inodedep->id_savedino2;
4221 		FREE(inodedep->id_savedino1, M_INODEDEP);
4222 		inodedep->id_savedino1 = NULL;
4223 		if ((bp->b_flags & B_DELWRI) == 0)
4224 			stat_inode_bitmap++;
4225 		bdirty(bp);
4226 		return (1);
4227 	}
4228 	/*
4229 	 * Roll forward anything that had to be rolled back before
4230 	 * the inode could be updated.
4231 	 */
4232 	hadchanges = 0;
4233 	for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
4234 		nextadp = TAILQ_NEXT(adp, ad_next);
4235 		if (adp->ad_state & ATTACHED) {
4236 			lk.lkt_held = NOHOLDER;
4237 			panic("handle_written_inodeblock: new entry");
4238 		}
4239 		if (fstype == UFS1) {
4240 			if (adp->ad_lbn < NDADDR) {
4241 				if (dp1->di_db[adp->ad_lbn]!=adp->ad_oldblkno) {
4242 					lk.lkt_held = NOHOLDER;
4243 					panic("%s %s #%jd mismatch %d != %jd",
4244 					    "handle_written_inodeblock:",
4245 					    "direct pointer",
4246 					    (intmax_t)adp->ad_lbn,
4247 					    dp1->di_db[adp->ad_lbn],
4248 					    (intmax_t)adp->ad_oldblkno);
4249 				}
4250 				dp1->di_db[adp->ad_lbn] = adp->ad_newblkno;
4251 			} else {
4252 				if (dp1->di_ib[adp->ad_lbn - NDADDR] != 0) {
4253 					lk.lkt_held = NOHOLDER;
4254 					panic("%s: %s #%jd allocated as %d",
4255 					    "handle_written_inodeblock",
4256 					    "indirect pointer",
4257 					    (intmax_t)adp->ad_lbn - NDADDR,
4258 					    dp1->di_ib[adp->ad_lbn - NDADDR]);
4259 				}
4260 				dp1->di_ib[adp->ad_lbn - NDADDR] =
4261 				    adp->ad_newblkno;
4262 			}
4263 		} else {
4264 			if (adp->ad_lbn < NDADDR) {
4265 				if (dp2->di_db[adp->ad_lbn]!=adp->ad_oldblkno) {
4266 					lk.lkt_held = NOHOLDER;
4267 					panic("%s: %s #%jd %s %jd != %jd",
4268 					    "handle_written_inodeblock",
4269 					    "direct pointer",
4270 					    (intmax_t)adp->ad_lbn, "mismatch",
4271 					    (intmax_t)dp2->di_db[adp->ad_lbn],
4272 					    (intmax_t)adp->ad_oldblkno);
4273 				}
4274 				dp2->di_db[adp->ad_lbn] = adp->ad_newblkno;
4275 			} else {
4276 				if (dp2->di_ib[adp->ad_lbn - NDADDR] != 0) {
4277 					lk.lkt_held = NOHOLDER;
4278 					panic("%s: %s #%jd allocated as %jd",
4279 					    "handle_written_inodeblock",
4280 					    "indirect pointer",
4281 					    (intmax_t)adp->ad_lbn - NDADDR,
4282 					    (intmax_t)
4283 					    dp2->di_ib[adp->ad_lbn - NDADDR]);
4284 				}
4285 				dp2->di_ib[adp->ad_lbn - NDADDR] =
4286 				    adp->ad_newblkno;
4287 			}
4288 		}
4289 		adp->ad_state &= ~UNDONE;
4290 		adp->ad_state |= ATTACHED;
4291 		hadchanges = 1;
4292 	}
4293 	for (adp = TAILQ_FIRST(&inodedep->id_extupdt); adp; adp = nextadp) {
4294 		nextadp = TAILQ_NEXT(adp, ad_next);
4295 		if (adp->ad_state & ATTACHED) {
4296 			lk.lkt_held = NOHOLDER;
4297 			panic("handle_written_inodeblock: new entry");
4298 		}
4299 		if (dp2->di_extb[adp->ad_lbn] != adp->ad_oldblkno) {
4300 			lk.lkt_held = NOHOLDER;
4301 			panic("%s: direct pointers #%jd %s %jd != %jd",
4302 			    "handle_written_inodeblock",
4303 			    (intmax_t)adp->ad_lbn, "mismatch",
4304 			    (intmax_t)dp2->di_extb[adp->ad_lbn],
4305 			    (intmax_t)adp->ad_oldblkno);
4306 		}
4307 		dp2->di_extb[adp->ad_lbn] = adp->ad_newblkno;
4308 		adp->ad_state &= ~UNDONE;
4309 		adp->ad_state |= ATTACHED;
4310 		hadchanges = 1;
4311 	}
4312 	if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
4313 		stat_direct_blk_ptrs++;
4314 	/*
4315 	 * Reset the file size to its most up-to-date value.
4316 	 */
4317 	if (inodedep->id_savedsize == -1 || inodedep->id_savedextsize == -1) {
4318 		lk.lkt_held = NOHOLDER;
4319 		panic("handle_written_inodeblock: bad size");
4320 	}
4321 	if (fstype == UFS1) {
4322 		if (dp1->di_size != inodedep->id_savedsize) {
4323 			dp1->di_size = inodedep->id_savedsize;
4324 			hadchanges = 1;
4325 		}
4326 	} else {
4327 		if (dp2->di_size != inodedep->id_savedsize) {
4328 			dp2->di_size = inodedep->id_savedsize;
4329 			hadchanges = 1;
4330 		}
4331 		if (dp2->di_extsize != inodedep->id_savedextsize) {
4332 			dp2->di_extsize = inodedep->id_savedextsize;
4333 			hadchanges = 1;
4334 		}
4335 	}
4336 	inodedep->id_savedsize = -1;
4337 	inodedep->id_savedextsize = -1;
4338 	/*
4339 	 * If there were any rollbacks in the inode block, then it must be
4340 	 * marked dirty so that its will eventually get written back in
4341 	 * its correct form.
4342 	 */
4343 	if (hadchanges)
4344 		bdirty(bp);
4345 	/*
4346 	 * Process any allocdirects that completed during the update.
4347 	 */
4348 	if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
4349 		handle_allocdirect_partdone(adp);
4350 	if ((adp = TAILQ_FIRST(&inodedep->id_extupdt)) != NULL)
4351 		handle_allocdirect_partdone(adp);
4352 	/*
4353 	 * Process deallocations that were held pending until the
4354 	 * inode had been written to disk. Freeing of the inode
4355 	 * is delayed until after all blocks have been freed to
4356 	 * avoid creation of new <vfsid, inum, lbn> triples
4357 	 * before the old ones have been deleted.
4358 	 */
4359 	filefree = NULL;
4360 	while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
4361 		WORKLIST_REMOVE(wk);
4362 		switch (wk->wk_type) {
4363 
4364 		case D_FREEFILE:
4365 			/*
4366 			 * We defer adding filefree to the worklist until
4367 			 * all other additions have been made to ensure
4368 			 * that it will be done after all the old blocks
4369 			 * have been freed.
4370 			 */
4371 			if (filefree != NULL) {
4372 				lk.lkt_held = NOHOLDER;
4373 				panic("handle_written_inodeblock: filefree");
4374 			}
4375 			filefree = wk;
4376 			continue;
4377 
4378 		case D_MKDIR:
4379 			handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
4380 			continue;
4381 
4382 		case D_DIRADD:
4383 			diradd_inode_written(WK_DIRADD(wk), inodedep);
4384 			continue;
4385 
4386 		case D_FREEBLKS:
4387 		case D_FREEFRAG:
4388 		case D_DIRREM:
4389 			add_to_worklist(wk);
4390 			continue;
4391 
4392 		case D_NEWDIRBLK:
4393 			free_newdirblk(WK_NEWDIRBLK(wk));
4394 			continue;
4395 
4396 		default:
4397 			lk.lkt_held = NOHOLDER;
4398 			panic("handle_written_inodeblock: Unknown type %s",
4399 			    TYPENAME(wk->wk_type));
4400 			/* NOTREACHED */
4401 		}
4402 	}
4403 	if (filefree != NULL) {
4404 		if (free_inodedep(inodedep) == 0) {
4405 			lk.lkt_held = NOHOLDER;
4406 			panic("handle_written_inodeblock: live inodedep");
4407 		}
4408 		add_to_worklist(filefree);
4409 		return (0);
4410 	}
4411 
4412 	/*
4413 	 * If no outstanding dependencies, free it.
4414 	 */
4415 	if (free_inodedep(inodedep) ||
4416 	    (TAILQ_FIRST(&inodedep->id_inoupdt) == 0 &&
4417 	     TAILQ_FIRST(&inodedep->id_extupdt) == 0))
4418 		return (0);
4419 	return (hadchanges);
4420 }
4421 
4422 /*
4423  * Process a diradd entry after its dependent inode has been written.
4424  * This routine must be called with splbio interrupts blocked.
4425  */
4426 static void
4427 diradd_inode_written(dap, inodedep)
4428 	struct diradd *dap;
4429 	struct inodedep *inodedep;
4430 {
4431 	struct pagedep *pagedep;
4432 
4433 	dap->da_state |= COMPLETE;
4434 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4435 		if (dap->da_state & DIRCHG)
4436 			pagedep = dap->da_previous->dm_pagedep;
4437 		else
4438 			pagedep = dap->da_pagedep;
4439 		LIST_REMOVE(dap, da_pdlist);
4440 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4441 	}
4442 	WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
4443 }
4444 
4445 /*
4446  * Handle the completion of a mkdir dependency.
4447  */
4448 static void
4449 handle_written_mkdir(mkdir, type)
4450 	struct mkdir *mkdir;
4451 	int type;
4452 {
4453 	struct diradd *dap;
4454 	struct pagedep *pagedep;
4455 
4456 	if (mkdir->md_state != type) {
4457 		lk.lkt_held = NOHOLDER;
4458 		panic("handle_written_mkdir: bad type");
4459 	}
4460 	dap = mkdir->md_diradd;
4461 	dap->da_state &= ~type;
4462 	if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
4463 		dap->da_state |= DEPCOMPLETE;
4464 	if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4465 		if (dap->da_state & DIRCHG)
4466 			pagedep = dap->da_previous->dm_pagedep;
4467 		else
4468 			pagedep = dap->da_pagedep;
4469 		LIST_REMOVE(dap, da_pdlist);
4470 		LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
4471 	}
4472 	LIST_REMOVE(mkdir, md_mkdirs);
4473 	WORKITEM_FREE(mkdir, D_MKDIR);
4474 }
4475 
4476 /*
4477  * Called from within softdep_disk_write_complete above.
4478  * A write operation was just completed. Removed inodes can
4479  * now be freed and associated block pointers may be committed.
4480  * Note that this routine is always called from interrupt level
4481  * with further splbio interrupts blocked.
4482  */
4483 static int
4484 handle_written_filepage(pagedep, bp)
4485 	struct pagedep *pagedep;
4486 	struct buf *bp;		/* buffer containing the written page */
4487 {
4488 	struct dirrem *dirrem;
4489 	struct diradd *dap, *nextdap;
4490 	struct direct *ep;
4491 	int i, chgs;
4492 
4493 	if ((pagedep->pd_state & IOSTARTED) == 0) {
4494 		lk.lkt_held = NOHOLDER;
4495 		panic("handle_written_filepage: not started");
4496 	}
4497 	pagedep->pd_state &= ~IOSTARTED;
4498 	/*
4499 	 * Process any directory removals that have been committed.
4500 	 */
4501 	while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
4502 		LIST_REMOVE(dirrem, dm_next);
4503 		dirrem->dm_dirinum = pagedep->pd_ino;
4504 		add_to_worklist(&dirrem->dm_list);
4505 	}
4506 	/*
4507 	 * Free any directory additions that have been committed.
4508 	 * If it is a newly allocated block, we have to wait until
4509 	 * the on-disk directory inode claims the new block.
4510 	 */
4511 	if ((pagedep->pd_state & NEWBLOCK) == 0)
4512 		while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
4513 			free_diradd(dap);
4514 	/*
4515 	 * Uncommitted directory entries must be restored.
4516 	 */
4517 	for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
4518 		for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
4519 		     dap = nextdap) {
4520 			nextdap = LIST_NEXT(dap, da_pdlist);
4521 			if (dap->da_state & ATTACHED) {
4522 				lk.lkt_held = NOHOLDER;
4523 				panic("handle_written_filepage: attached");
4524 			}
4525 			ep = (struct direct *)
4526 			    ((char *)bp->b_data + dap->da_offset);
4527 			ep->d_ino = dap->da_newinum;
4528 			dap->da_state &= ~UNDONE;
4529 			dap->da_state |= ATTACHED;
4530 			chgs = 1;
4531 			/*
4532 			 * If the inode referenced by the directory has
4533 			 * been written out, then the dependency can be
4534 			 * moved to the pending list.
4535 			 */
4536 			if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
4537 				LIST_REMOVE(dap, da_pdlist);
4538 				LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
4539 				    da_pdlist);
4540 			}
4541 		}
4542 	}
4543 	/*
4544 	 * If there were any rollbacks in the directory, then it must be
4545 	 * marked dirty so that its will eventually get written back in
4546 	 * its correct form.
4547 	 */
4548 	if (chgs) {
4549 		if ((bp->b_flags & B_DELWRI) == 0)
4550 			stat_dir_entry++;
4551 		bdirty(bp);
4552 		return (1);
4553 	}
4554 	/*
4555 	 * If we are not waiting for a new directory block to be
4556 	 * claimed by its inode, then the pagedep will be freed.
4557 	 * Otherwise it will remain to track any new entries on
4558 	 * the page in case they are fsync'ed.
4559 	 */
4560 	if ((pagedep->pd_state & NEWBLOCK) == 0) {
4561 		LIST_REMOVE(pagedep, pd_hash);
4562 		WORKITEM_FREE(pagedep, D_PAGEDEP);
4563 	}
4564 	return (0);
4565 }
4566 
4567 /*
4568  * Writing back in-core inode structures.
4569  *
4570  * The filesystem only accesses an inode's contents when it occupies an
4571  * "in-core" inode structure.  These "in-core" structures are separate from
4572  * the page frames used to cache inode blocks.  Only the latter are
4573  * transferred to/from the disk.  So, when the updated contents of the
4574  * "in-core" inode structure are copied to the corresponding in-memory inode
4575  * block, the dependencies are also transferred.  The following procedure is
4576  * called when copying a dirty "in-core" inode to a cached inode block.
4577  */
4578 
4579 /*
4580  * Called when an inode is loaded from disk. If the effective link count
4581  * differed from the actual link count when it was last flushed, then we
4582  * need to ensure that the correct effective link count is put back.
4583  */
4584 void
4585 softdep_load_inodeblock(ip)
4586 	struct inode *ip;	/* the "in_core" copy of the inode */
4587 {
4588 	struct inodedep *inodedep;
4589 
4590 	/*
4591 	 * Check for alternate nlink count.
4592 	 */
4593 	ip->i_effnlink = ip->i_nlink;
4594 	ACQUIRE_LOCK(&lk);
4595 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
4596 		FREE_LOCK(&lk);
4597 		return;
4598 	}
4599 	ip->i_effnlink -= inodedep->id_nlinkdelta;
4600 	if (inodedep->id_state & SPACECOUNTED)
4601 		ip->i_flag |= IN_SPACECOUNTED;
4602 	FREE_LOCK(&lk);
4603 }
4604 
4605 /*
4606  * This routine is called just before the "in-core" inode
4607  * information is to be copied to the in-memory inode block.
4608  * Recall that an inode block contains several inodes. If
4609  * the force flag is set, then the dependencies will be
4610  * cleared so that the update can always be made. Note that
4611  * the buffer is locked when this routine is called, so we
4612  * will never be in the middle of writing the inode block
4613  * to disk.
4614  */
4615 void
4616 softdep_update_inodeblock(ip, bp, waitfor)
4617 	struct inode *ip;	/* the "in_core" copy of the inode */
4618 	struct buf *bp;		/* the buffer containing the inode block */
4619 	int waitfor;		/* nonzero => update must be allowed */
4620 {
4621 	struct inodedep *inodedep;
4622 	struct worklist *wk;
4623 	int error, gotit;
4624 
4625 	/*
4626 	 * If the effective link count is not equal to the actual link
4627 	 * count, then we must track the difference in an inodedep while
4628 	 * the inode is (potentially) tossed out of the cache. Otherwise,
4629 	 * if there is no existing inodedep, then there are no dependencies
4630 	 * to track.
4631 	 */
4632 	ACQUIRE_LOCK(&lk);
4633 	if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
4634 		FREE_LOCK(&lk);
4635 		if (ip->i_effnlink != ip->i_nlink)
4636 			panic("softdep_update_inodeblock: bad link count");
4637 		return;
4638 	}
4639 	if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
4640 		FREE_LOCK(&lk);
4641 		panic("softdep_update_inodeblock: bad delta");
4642 	}
4643 	/*
4644 	 * Changes have been initiated. Anything depending on these
4645 	 * changes cannot occur until this inode has been written.
4646 	 */
4647 	inodedep->id_state &= ~COMPLETE;
4648 	if ((inodedep->id_state & ONWORKLIST) == 0)
4649 		WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
4650 	/*
4651 	 * Any new dependencies associated with the incore inode must
4652 	 * now be moved to the list associated with the buffer holding
4653 	 * the in-memory copy of the inode. Once merged process any
4654 	 * allocdirects that are completed by the merger.
4655 	 */
4656 	merge_inode_lists(&inodedep->id_newinoupdt, &inodedep->id_inoupdt);
4657 	if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
4658 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
4659 	merge_inode_lists(&inodedep->id_newextupdt, &inodedep->id_extupdt);
4660 	if (TAILQ_FIRST(&inodedep->id_extupdt) != NULL)
4661 		handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_extupdt));
4662 	/*
4663 	 * Now that the inode has been pushed into the buffer, the
4664 	 * operations dependent on the inode being written to disk
4665 	 * can be moved to the id_bufwait so that they will be
4666 	 * processed when the buffer I/O completes.
4667 	 */
4668 	while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
4669 		WORKLIST_REMOVE(wk);
4670 		WORKLIST_INSERT(&inodedep->id_bufwait, wk);
4671 	}
4672 	/*
4673 	 * Newly allocated inodes cannot be written until the bitmap
4674 	 * that allocates them have been written (indicated by
4675 	 * DEPCOMPLETE being set in id_state). If we are doing a
4676 	 * forced sync (e.g., an fsync on a file), we force the bitmap
4677 	 * to be written so that the update can be done.
4678 	 */
4679 	if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
4680 		FREE_LOCK(&lk);
4681 		return;
4682 	}
4683 	gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4684 	FREE_LOCK(&lk);
4685 	if (gotit &&
4686 	    (error = BUF_WRITE(inodedep->id_buf)) != 0)
4687 		softdep_error("softdep_update_inodeblock: bwrite", error);
4688 	if ((inodedep->id_state & DEPCOMPLETE) == 0)
4689 		panic("softdep_update_inodeblock: update failed");
4690 }
4691 
4692 /*
4693  * Merge the a new inode dependency list (such as id_newinoupdt) into an
4694  * old inode dependency list (such as id_inoupdt). This routine must be
4695  * called with splbio interrupts blocked.
4696  */
4697 static void
4698 merge_inode_lists(newlisthead, oldlisthead)
4699 	struct allocdirectlst *newlisthead;
4700 	struct allocdirectlst *oldlisthead;
4701 {
4702 	struct allocdirect *listadp, *newadp;
4703 
4704 	newadp = TAILQ_FIRST(newlisthead);
4705 	for (listadp = TAILQ_FIRST(oldlisthead); listadp && newadp;) {
4706 		if (listadp->ad_lbn < newadp->ad_lbn) {
4707 			listadp = TAILQ_NEXT(listadp, ad_next);
4708 			continue;
4709 		}
4710 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4711 		TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
4712 		if (listadp->ad_lbn == newadp->ad_lbn) {
4713 			allocdirect_merge(oldlisthead, newadp,
4714 			    listadp);
4715 			listadp = newadp;
4716 		}
4717 		newadp = TAILQ_FIRST(newlisthead);
4718 	}
4719 	while ((newadp = TAILQ_FIRST(newlisthead)) != NULL) {
4720 		TAILQ_REMOVE(newlisthead, newadp, ad_next);
4721 		TAILQ_INSERT_TAIL(oldlisthead, newadp, ad_next);
4722 	}
4723 }
4724 
4725 /*
4726  * If we are doing an fsync, then we must ensure that any directory
4727  * entries for the inode have been written after the inode gets to disk.
4728  */
4729 int
4730 softdep_fsync(vp)
4731 	struct vnode *vp;	/* the "in_core" copy of the inode */
4732 {
4733 	struct inodedep *inodedep;
4734 	struct pagedep *pagedep;
4735 	struct worklist *wk;
4736 	struct diradd *dap;
4737 	struct mount *mnt;
4738 	struct vnode *pvp;
4739 	struct inode *ip;
4740 	struct buf *bp;
4741 	struct fs *fs;
4742 	struct thread *td = curthread;
4743 	int error, flushparent;
4744 	ino_t parentino;
4745 	ufs_lbn_t lbn;
4746 
4747 	ip = VTOI(vp);
4748 	fs = ip->i_fs;
4749 	ACQUIRE_LOCK(&lk);
4750 	if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
4751 		FREE_LOCK(&lk);
4752 		return (0);
4753 	}
4754 	if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
4755 	    LIST_FIRST(&inodedep->id_bufwait) != NULL ||
4756 	    TAILQ_FIRST(&inodedep->id_extupdt) != NULL ||
4757 	    TAILQ_FIRST(&inodedep->id_newextupdt) != NULL ||
4758 	    TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
4759 	    TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) {
4760 		FREE_LOCK(&lk);
4761 		panic("softdep_fsync: pending ops");
4762 	}
4763 	for (error = 0, flushparent = 0; ; ) {
4764 		if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
4765 			break;
4766 		if (wk->wk_type != D_DIRADD) {
4767 			FREE_LOCK(&lk);
4768 			panic("softdep_fsync: Unexpected type %s",
4769 			    TYPENAME(wk->wk_type));
4770 		}
4771 		dap = WK_DIRADD(wk);
4772 		/*
4773 		 * Flush our parent if this directory entry has a MKDIR_PARENT
4774 		 * dependency or is contained in a newly allocated block.
4775 		 */
4776 		if (dap->da_state & DIRCHG)
4777 			pagedep = dap->da_previous->dm_pagedep;
4778 		else
4779 			pagedep = dap->da_pagedep;
4780 		mnt = pagedep->pd_mnt;
4781 		parentino = pagedep->pd_ino;
4782 		lbn = pagedep->pd_lbn;
4783 		if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) {
4784 			FREE_LOCK(&lk);
4785 			panic("softdep_fsync: dirty");
4786 		}
4787 		if ((dap->da_state & MKDIR_PARENT) ||
4788 		    (pagedep->pd_state & NEWBLOCK))
4789 			flushparent = 1;
4790 		else
4791 			flushparent = 0;
4792 		/*
4793 		 * If we are being fsync'ed as part of vgone'ing this vnode,
4794 		 * then we will not be able to release and recover the
4795 		 * vnode below, so we just have to give up on writing its
4796 		 * directory entry out. It will eventually be written, just
4797 		 * not now, but then the user was not asking to have it
4798 		 * written, so we are not breaking any promises.
4799 		 */
4800 		mp_fixme("This operation is not atomic wrt the rest of the code");
4801 		VI_LOCK(vp);
4802 		if (vp->v_iflag & VI_XLOCK) {
4803 			VI_UNLOCK(vp);
4804 			break;
4805 		} else
4806 			VI_UNLOCK(vp);
4807 		/*
4808 		 * We prevent deadlock by always fetching inodes from the
4809 		 * root, moving down the directory tree. Thus, when fetching
4810 		 * our parent directory, we first try to get the lock. If
4811 		 * that fails, we must unlock ourselves before requesting
4812 		 * the lock on our parent. See the comment in ufs_lookup
4813 		 * for details on possible races.
4814 		 */
4815 		FREE_LOCK(&lk);
4816 		if (VFS_VGET(mnt, parentino, LK_NOWAIT | LK_EXCLUSIVE, &pvp)) {
4817 			VOP_UNLOCK(vp, 0, td);
4818 			error = VFS_VGET(mnt, parentino, LK_EXCLUSIVE, &pvp);
4819 			vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
4820 			if (error != 0)
4821 				return (error);
4822 		}
4823 		/*
4824 		 * All MKDIR_PARENT dependencies and all the NEWBLOCK pagedeps
4825 		 * that are contained in direct blocks will be resolved by
4826 		 * doing a UFS_UPDATE. Pagedeps contained in indirect blocks
4827 		 * may require a complete sync'ing of the directory. So, we
4828 		 * try the cheap and fast UFS_UPDATE first, and if that fails,
4829 		 * then we do the slower VOP_FSYNC of the directory.
4830 		 */
4831 		if (flushparent) {
4832 			if ((error = UFS_UPDATE(pvp, 1)) != 0) {
4833 				vput(pvp);
4834 				return (error);
4835 			}
4836 			if ((pagedep->pd_state & NEWBLOCK) &&
4837 			    (error = VOP_FSYNC(pvp, td->td_ucred, MNT_WAIT, td))) {
4838 				vput(pvp);
4839 				return (error);
4840 			}
4841 		}
4842 		/*
4843 		 * Flush directory page containing the inode's name.
4844 		 */
4845 		error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), td->td_ucred,
4846 		    &bp);
4847 		if (error == 0)
4848 			error = BUF_WRITE(bp);
4849 		else
4850 			brelse(bp);
4851 		vput(pvp);
4852 		if (error != 0)
4853 			return (error);
4854 		ACQUIRE_LOCK(&lk);
4855 		if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
4856 			break;
4857 	}
4858 	FREE_LOCK(&lk);
4859 	return (0);
4860 }
4861 
4862 /*
4863  * Flush all the dirty bitmaps associated with the block device
4864  * before flushing the rest of the dirty blocks so as to reduce
4865  * the number of dependencies that will have to be rolled back.
4866  */
4867 void
4868 softdep_fsync_mountdev(vp)
4869 	struct vnode *vp;
4870 {
4871 	struct buf *bp, *nbp;
4872 	struct worklist *wk;
4873 
4874 	if (!vn_isdisk(vp, NULL))
4875 		panic("softdep_fsync_mountdev: vnode not a disk");
4876 	ACQUIRE_LOCK(&lk);
4877 	VI_LOCK(vp);
4878 	for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
4879 		nbp = TAILQ_NEXT(bp, b_vnbufs);
4880 		VI_UNLOCK(vp);
4881 		/*
4882 		 * If it is already scheduled, skip to the next buffer.
4883 		 */
4884 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
4885 			VI_LOCK(vp);
4886 			continue;
4887 		}
4888 		if ((bp->b_flags & B_DELWRI) == 0) {
4889 			FREE_LOCK(&lk);
4890 			panic("softdep_fsync_mountdev: not dirty");
4891 		}
4892 		/*
4893 		 * We are only interested in bitmaps with outstanding
4894 		 * dependencies.
4895 		 */
4896 		if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
4897 		    wk->wk_type != D_BMSAFEMAP ||
4898 		    (bp->b_xflags & BX_BKGRDINPROG)) {
4899 			BUF_UNLOCK(bp);
4900 			VI_LOCK(vp);
4901 			continue;
4902 		}
4903 		bremfree(bp);
4904 		FREE_LOCK(&lk);
4905 		(void) bawrite(bp);
4906 		ACQUIRE_LOCK(&lk);
4907 		/*
4908 		 * Since we may have slept during the I/O, we need
4909 		 * to start from a known point.
4910 		 */
4911 		VI_LOCK(vp);
4912 		nbp = TAILQ_FIRST(&vp->v_dirtyblkhd);
4913 	}
4914 	VI_UNLOCK(vp);
4915 	drain_output(vp, 1);
4916 	FREE_LOCK(&lk);
4917 }
4918 
4919 /*
4920  * This routine is called when we are trying to synchronously flush a
4921  * file. This routine must eliminate any filesystem metadata dependencies
4922  * so that the syncing routine can succeed by pushing the dirty blocks
4923  * associated with the file. If any I/O errors occur, they are returned.
4924  */
4925 int
4926 softdep_sync_metadata(ap)
4927 	struct vop_fsync_args /* {
4928 		struct vnode *a_vp;
4929 		struct ucred *a_cred;
4930 		int a_waitfor;
4931 		struct thread *a_td;
4932 	} */ *ap;
4933 {
4934 	struct vnode *vp = ap->a_vp;
4935 	struct pagedep *pagedep;
4936 	struct allocdirect *adp;
4937 	struct allocindir *aip;
4938 	struct buf *bp, *nbp;
4939 	struct worklist *wk;
4940 	int i, error, waitfor;
4941 
4942 	/*
4943 	 * Check whether this vnode is involved in a filesystem
4944 	 * that is doing soft dependency processing.
4945 	 */
4946 	if (!vn_isdisk(vp, NULL)) {
4947 		if (!DOINGSOFTDEP(vp))
4948 			return (0);
4949 	} else
4950 		if (vp->v_rdev->si_mountpoint == NULL ||
4951 		    (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0)
4952 			return (0);
4953 	/*
4954 	 * Ensure that any direct block dependencies have been cleared.
4955 	 */
4956 	ACQUIRE_LOCK(&lk);
4957 	if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
4958 		FREE_LOCK(&lk);
4959 		return (error);
4960 	}
4961 	/*
4962 	 * For most files, the only metadata dependencies are the
4963 	 * cylinder group maps that allocate their inode or blocks.
4964 	 * The block allocation dependencies can be found by traversing
4965 	 * the dependency lists for any buffers that remain on their
4966 	 * dirty buffer list. The inode allocation dependency will
4967 	 * be resolved when the inode is updated with MNT_WAIT.
4968 	 * This work is done in two passes. The first pass grabs most
4969 	 * of the buffers and begins asynchronously writing them. The
4970 	 * only way to wait for these asynchronous writes is to sleep
4971 	 * on the filesystem vnode which may stay busy for a long time
4972 	 * if the filesystem is active. So, instead, we make a second
4973 	 * pass over the dependencies blocking on each write. In the
4974 	 * usual case we will be blocking against a write that we
4975 	 * initiated, so when it is done the dependency will have been
4976 	 * resolved. Thus the second pass is expected to end quickly.
4977 	 */
4978 	waitfor = MNT_NOWAIT;
4979 top:
4980 	/*
4981 	 * We must wait for any I/O in progress to finish so that
4982 	 * all potential buffers on the dirty list will be visible.
4983 	 */
4984 	drain_output(vp, 1);
4985 	if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) {
4986 		FREE_LOCK(&lk);
4987 		return (0);
4988 	}
4989 	mp_fixme("The locking is somewhat complicated nonexistant here.");
4990 	bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
4991 	/* While syncing snapshots, we must allow recursive lookups */
4992 	bp->b_lock.lk_flags |= LK_CANRECURSE;
4993 loop:
4994 	/*
4995 	 * As we hold the buffer locked, none of its dependencies
4996 	 * will disappear.
4997 	 */
4998 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4999 		switch (wk->wk_type) {
5000 
5001 		case D_ALLOCDIRECT:
5002 			adp = WK_ALLOCDIRECT(wk);
5003 			if (adp->ad_state & DEPCOMPLETE)
5004 				continue;
5005 			nbp = adp->ad_buf;
5006 			if (getdirtybuf(&nbp, waitfor) == 0)
5007 				continue;
5008 			FREE_LOCK(&lk);
5009 			if (waitfor == MNT_NOWAIT) {
5010 				bawrite(nbp);
5011 			} else if ((error = BUF_WRITE(nbp)) != 0) {
5012 				break;
5013 			}
5014 			ACQUIRE_LOCK(&lk);
5015 			continue;
5016 
5017 		case D_ALLOCINDIR:
5018 			aip = WK_ALLOCINDIR(wk);
5019 			if (aip->ai_state & DEPCOMPLETE)
5020 				continue;
5021 			nbp = aip->ai_buf;
5022 			if (getdirtybuf(&nbp, waitfor) == 0)
5023 				continue;
5024 			FREE_LOCK(&lk);
5025 			if (waitfor == MNT_NOWAIT) {
5026 				bawrite(nbp);
5027 			} else if ((error = BUF_WRITE(nbp)) != 0) {
5028 				break;
5029 			}
5030 			ACQUIRE_LOCK(&lk);
5031 			continue;
5032 
5033 		case D_INDIRDEP:
5034 		restart:
5035 
5036 			LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
5037 				if (aip->ai_state & DEPCOMPLETE)
5038 					continue;
5039 				nbp = aip->ai_buf;
5040 				if (getdirtybuf(&nbp, MNT_WAIT) == 0)
5041 					goto restart;
5042 				FREE_LOCK(&lk);
5043 				if ((error = BUF_WRITE(nbp)) != 0) {
5044 					break;
5045 				}
5046 				ACQUIRE_LOCK(&lk);
5047 				goto restart;
5048 			}
5049 			continue;
5050 
5051 		case D_INODEDEP:
5052 			if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
5053 			    WK_INODEDEP(wk)->id_ino)) != 0) {
5054 				FREE_LOCK(&lk);
5055 				break;
5056 			}
5057 			continue;
5058 
5059 		case D_PAGEDEP:
5060 			/*
5061 			 * We are trying to sync a directory that may
5062 			 * have dependencies on both its own metadata
5063 			 * and/or dependencies on the inodes of any
5064 			 * recently allocated files. We walk its diradd
5065 			 * lists pushing out the associated inode.
5066 			 */
5067 			pagedep = WK_PAGEDEP(wk);
5068 			for (i = 0; i < DAHASHSZ; i++) {
5069 				if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
5070 					continue;
5071 				if ((error =
5072 				    flush_pagedep_deps(vp, pagedep->pd_mnt,
5073 						&pagedep->pd_diraddhd[i]))) {
5074 					FREE_LOCK(&lk);
5075 					break;
5076 				}
5077 			}
5078 			continue;
5079 
5080 		case D_MKDIR:
5081 			/*
5082 			 * This case should never happen if the vnode has
5083 			 * been properly sync'ed. However, if this function
5084 			 * is used at a place where the vnode has not yet
5085 			 * been sync'ed, this dependency can show up. So,
5086 			 * rather than panic, just flush it.
5087 			 */
5088 			nbp = WK_MKDIR(wk)->md_buf;
5089 			if (getdirtybuf(&nbp, waitfor) == 0)
5090 				continue;
5091 			FREE_LOCK(&lk);
5092 			if (waitfor == MNT_NOWAIT) {
5093 				bawrite(nbp);
5094 			} else if ((error = BUF_WRITE(nbp)) != 0) {
5095 				break;
5096 			}
5097 			ACQUIRE_LOCK(&lk);
5098 			continue;
5099 
5100 		case D_BMSAFEMAP:
5101 			/*
5102 			 * This case should never happen if the vnode has
5103 			 * been properly sync'ed. However, if this function
5104 			 * is used at a place where the vnode has not yet
5105 			 * been sync'ed, this dependency can show up. So,
5106 			 * rather than panic, just flush it.
5107 			 */
5108 			nbp = WK_BMSAFEMAP(wk)->sm_buf;
5109 			if (getdirtybuf(&nbp, waitfor) == 0)
5110 				continue;
5111 			FREE_LOCK(&lk);
5112 			if (waitfor == MNT_NOWAIT) {
5113 				bawrite(nbp);
5114 			} else if ((error = BUF_WRITE(nbp)) != 0) {
5115 				break;
5116 			}
5117 			ACQUIRE_LOCK(&lk);
5118 			continue;
5119 
5120 		default:
5121 			FREE_LOCK(&lk);
5122 			panic("softdep_sync_metadata: Unknown type %s",
5123 			    TYPENAME(wk->wk_type));
5124 			/* NOTREACHED */
5125 		}
5126 		/* We reach here only in error and unlocked */
5127 		if (error == 0)
5128 			panic("softdep_sync_metadata: zero error");
5129 		bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5130 		bawrite(bp);
5131 		return (error);
5132 	}
5133 	(void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT);
5134 	nbp = TAILQ_NEXT(bp, b_vnbufs);
5135 	FREE_LOCK(&lk);
5136 	bp->b_lock.lk_flags &= ~LK_CANRECURSE;
5137 	bawrite(bp);
5138 	ACQUIRE_LOCK(&lk);
5139 	if (nbp != NULL) {
5140 		bp = nbp;
5141 		goto loop;
5142 	}
5143 	/*
5144 	 * The brief unlock is to allow any pent up dependency
5145 	 * processing to be done. Then proceed with the second pass.
5146 	 */
5147 	if (waitfor == MNT_NOWAIT) {
5148 		waitfor = MNT_WAIT;
5149 		FREE_LOCK(&lk);
5150 		ACQUIRE_LOCK(&lk);
5151 		goto top;
5152 	}
5153 
5154 	/*
5155 	 * If we have managed to get rid of all the dirty buffers,
5156 	 * then we are done. For certain directories and block
5157 	 * devices, we may need to do further work.
5158 	 *
5159 	 * We must wait for any I/O in progress to finish so that
5160 	 * all potential buffers on the dirty list will be visible.
5161 	 */
5162 	drain_output(vp, 1);
5163 	if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) {
5164 		FREE_LOCK(&lk);
5165 		return (0);
5166 	}
5167 
5168 	FREE_LOCK(&lk);
5169 	/*
5170 	 * If we are trying to sync a block device, some of its buffers may
5171 	 * contain metadata that cannot be written until the contents of some
5172 	 * partially written files have been written to disk. The only easy
5173 	 * way to accomplish this is to sync the entire filesystem (luckily
5174 	 * this happens rarely).
5175 	 */
5176 	if (vn_isdisk(vp, NULL) &&
5177 	    vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) &&
5178 	    (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, ap->a_cred,
5179 	     ap->a_td)) != 0)
5180 		return (error);
5181 	return (0);
5182 }
5183 
5184 /*
5185  * Flush the dependencies associated with an inodedep.
5186  * Called with splbio blocked.
5187  */
5188 static int
5189 flush_inodedep_deps(fs, ino)
5190 	struct fs *fs;
5191 	ino_t ino;
5192 {
5193 	struct inodedep *inodedep;
5194 	int error, waitfor;
5195 
5196 	/*
5197 	 * This work is done in two passes. The first pass grabs most
5198 	 * of the buffers and begins asynchronously writing them. The
5199 	 * only way to wait for these asynchronous writes is to sleep
5200 	 * on the filesystem vnode which may stay busy for a long time
5201 	 * if the filesystem is active. So, instead, we make a second
5202 	 * pass over the dependencies blocking on each write. In the
5203 	 * usual case we will be blocking against a write that we
5204 	 * initiated, so when it is done the dependency will have been
5205 	 * resolved. Thus the second pass is expected to end quickly.
5206 	 * We give a brief window at the top of the loop to allow
5207 	 * any pending I/O to complete.
5208 	 */
5209 	for (error = 0, waitfor = MNT_NOWAIT; ; ) {
5210 		if (error)
5211 			return (error);
5212 		FREE_LOCK(&lk);
5213 		ACQUIRE_LOCK(&lk);
5214 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
5215 			return (0);
5216 		if (flush_deplist(&inodedep->id_inoupdt, waitfor, &error) ||
5217 		    flush_deplist(&inodedep->id_newinoupdt, waitfor, &error) ||
5218 		    flush_deplist(&inodedep->id_extupdt, waitfor, &error) ||
5219 		    flush_deplist(&inodedep->id_newextupdt, waitfor, &error))
5220 			continue;
5221 		/*
5222 		 * If pass2, we are done, otherwise do pass 2.
5223 		 */
5224 		if (waitfor == MNT_WAIT)
5225 			break;
5226 		waitfor = MNT_WAIT;
5227 	}
5228 	/*
5229 	 * Try freeing inodedep in case all dependencies have been removed.
5230 	 */
5231 	if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
5232 		(void) free_inodedep(inodedep);
5233 	return (0);
5234 }
5235 
5236 /*
5237  * Flush an inode dependency list.
5238  * Called with splbio blocked.
5239  */
5240 static int
5241 flush_deplist(listhead, waitfor, errorp)
5242 	struct allocdirectlst *listhead;
5243 	int waitfor;
5244 	int *errorp;
5245 {
5246 	struct allocdirect *adp;
5247 	struct buf *bp;
5248 
5249 	TAILQ_FOREACH(adp, listhead, ad_next) {
5250 		if (adp->ad_state & DEPCOMPLETE)
5251 			continue;
5252 		bp = adp->ad_buf;
5253 		if (getdirtybuf(&bp, waitfor) == 0) {
5254 			if (waitfor == MNT_NOWAIT)
5255 				continue;
5256 			return (1);
5257 		}
5258 		FREE_LOCK(&lk);
5259 		if (waitfor == MNT_NOWAIT) {
5260 			bawrite(bp);
5261 		} else if ((*errorp = BUF_WRITE(bp)) != 0) {
5262 			ACQUIRE_LOCK(&lk);
5263 			return (1);
5264 		}
5265 		ACQUIRE_LOCK(&lk);
5266 		return (1);
5267 	}
5268 	return (0);
5269 }
5270 
5271 /*
5272  * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
5273  * Called with splbio blocked.
5274  */
5275 static int
5276 flush_pagedep_deps(pvp, mp, diraddhdp)
5277 	struct vnode *pvp;
5278 	struct mount *mp;
5279 	struct diraddhd *diraddhdp;
5280 {
5281 	struct thread *td = curthread;
5282 	struct inodedep *inodedep;
5283 	struct ufsmount *ump;
5284 	struct diradd *dap;
5285 	struct vnode *vp;
5286 	int gotit, error = 0;
5287 	struct buf *bp;
5288 	ino_t inum;
5289 
5290 	ump = VFSTOUFS(mp);
5291 	while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
5292 		/*
5293 		 * Flush ourselves if this directory entry
5294 		 * has a MKDIR_PARENT dependency.
5295 		 */
5296 		if (dap->da_state & MKDIR_PARENT) {
5297 			FREE_LOCK(&lk);
5298 			if ((error = UFS_UPDATE(pvp, 1)) != 0)
5299 				break;
5300 			ACQUIRE_LOCK(&lk);
5301 			/*
5302 			 * If that cleared dependencies, go on to next.
5303 			 */
5304 			if (dap != LIST_FIRST(diraddhdp))
5305 				continue;
5306 			if (dap->da_state & MKDIR_PARENT) {
5307 				FREE_LOCK(&lk);
5308 				panic("flush_pagedep_deps: MKDIR_PARENT");
5309 			}
5310 		}
5311 		/*
5312 		 * A newly allocated directory must have its "." and
5313 		 * ".." entries written out before its name can be
5314 		 * committed in its parent. We do not want or need
5315 		 * the full semantics of a synchronous VOP_FSYNC as
5316 		 * that may end up here again, once for each directory
5317 		 * level in the filesystem. Instead, we push the blocks
5318 		 * and wait for them to clear. We have to fsync twice
5319 		 * because the first call may choose to defer blocks
5320 		 * that still have dependencies, but deferral will
5321 		 * happen at most once.
5322 		 */
5323 		inum = dap->da_newinum;
5324 		if (dap->da_state & MKDIR_BODY) {
5325 			FREE_LOCK(&lk);
5326 			if ((error = VFS_VGET(mp, inum, LK_EXCLUSIVE, &vp)))
5327 				break;
5328 			if ((error=VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)) ||
5329 			    (error=VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td))) {
5330 				vput(vp);
5331 				break;
5332 			}
5333 			drain_output(vp, 0);
5334 			vput(vp);
5335 			ACQUIRE_LOCK(&lk);
5336 			/*
5337 			 * If that cleared dependencies, go on to next.
5338 			 */
5339 			if (dap != LIST_FIRST(diraddhdp))
5340 				continue;
5341 			if (dap->da_state & MKDIR_BODY) {
5342 				FREE_LOCK(&lk);
5343 				panic("flush_pagedep_deps: MKDIR_BODY");
5344 			}
5345 		}
5346 		/*
5347 		 * Flush the inode on which the directory entry depends.
5348 		 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
5349 		 * the only remaining dependency is that the updated inode
5350 		 * count must get pushed to disk. The inode has already
5351 		 * been pushed into its inode buffer (via VOP_UPDATE) at
5352 		 * the time of the reference count change. So we need only
5353 		 * locate that buffer, ensure that there will be no rollback
5354 		 * caused by a bitmap dependency, then write the inode buffer.
5355 		 */
5356 		if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) {
5357 			FREE_LOCK(&lk);
5358 			panic("flush_pagedep_deps: lost inode");
5359 		}
5360 		/*
5361 		 * If the inode still has bitmap dependencies,
5362 		 * push them to disk.
5363 		 */
5364 		if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5365 			gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
5366 			FREE_LOCK(&lk);
5367 			if (gotit &&
5368 			    (error = BUF_WRITE(inodedep->id_buf)) != 0)
5369 				break;
5370 			ACQUIRE_LOCK(&lk);
5371 			if (dap != LIST_FIRST(diraddhdp))
5372 				continue;
5373 		}
5374 		/*
5375 		 * If the inode is still sitting in a buffer waiting
5376 		 * to be written, push it to disk.
5377 		 */
5378 		FREE_LOCK(&lk);
5379 		if ((error = bread(ump->um_devvp,
5380 		    fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
5381 		    (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0) {
5382 			brelse(bp);
5383 			break;
5384 		}
5385 		if ((error = BUF_WRITE(bp)) != 0)
5386 			break;
5387 		ACQUIRE_LOCK(&lk);
5388 		/*
5389 		 * If we have failed to get rid of all the dependencies
5390 		 * then something is seriously wrong.
5391 		 */
5392 		if (dap == LIST_FIRST(diraddhdp)) {
5393 			FREE_LOCK(&lk);
5394 			panic("flush_pagedep_deps: flush failed");
5395 		}
5396 	}
5397 	if (error)
5398 		ACQUIRE_LOCK(&lk);
5399 	return (error);
5400 }
5401 
5402 /*
5403  * A large burst of file addition or deletion activity can drive the
5404  * memory load excessively high. First attempt to slow things down
5405  * using the techniques below. If that fails, this routine requests
5406  * the offending operations to fall back to running synchronously
5407  * until the memory load returns to a reasonable level.
5408  */
5409 int
5410 softdep_slowdown(vp)
5411 	struct vnode *vp;
5412 {
5413 	int max_softdeps_hard;
5414 
5415 	max_softdeps_hard = max_softdeps * 11 / 10;
5416 	if (num_dirrem < max_softdeps_hard / 2 &&
5417 	    num_inodedep < max_softdeps_hard &&
5418 	    VFSTOUFS(vp->v_mount)->um_numindirdeps < maxindirdeps)
5419   		return (0);
5420 	if (VFSTOUFS(vp->v_mount)->um_numindirdeps >= maxindirdeps)
5421 		speedup_syncer();
5422 	stat_sync_limit_hit += 1;
5423 	return (1);
5424 }
5425 
5426 /*
5427  * Called by the allocation routines when they are about to fail
5428  * in the hope that we can free up some disk space.
5429  *
5430  * First check to see if the work list has anything on it. If it has,
5431  * clean up entries until we successfully free some space. Because this
5432  * process holds inodes locked, we cannot handle any remove requests
5433  * that might block on a locked inode as that could lead to deadlock.
5434  * If the worklist yields no free space, encourage the syncer daemon
5435  * to help us. In no event will we try for longer than tickdelay seconds.
5436  */
5437 int
5438 softdep_request_cleanup(fs, vp)
5439 	struct fs *fs;
5440 	struct vnode *vp;
5441 {
5442 	long starttime;
5443 	ufs2_daddr_t needed;
5444 
5445 	needed = fs->fs_cstotal.cs_nbfree + fs->fs_contigsumsize;
5446 	starttime = time_second + tickdelay;
5447 	/*
5448 	 * If we are being called because of a process doing a
5449 	 * copy-on-write, then it is not safe to update the vnode
5450 	 * as we may recurse into the copy-on-write routine.
5451 	 */
5452 	if ((curthread->td_proc->p_flag & P_COWINPROGRESS) == 0 &&
5453 	    UFS_UPDATE(vp, 1) != 0)
5454 		return (0);
5455 	while (fs->fs_pendingblocks > 0 && fs->fs_cstotal.cs_nbfree <= needed) {
5456 		if (time_second > starttime)
5457 			return (0);
5458 		if (num_on_worklist > 0 &&
5459 		    process_worklist_item(NULL, LK_NOWAIT) != -1) {
5460 			stat_worklist_push += 1;
5461 			continue;
5462 		}
5463 		request_cleanup(FLUSH_REMOVE_WAIT, 0);
5464 	}
5465 	return (1);
5466 }
5467 
5468 /*
5469  * If memory utilization has gotten too high, deliberately slow things
5470  * down and speed up the I/O processing.
5471  */
5472 static int
5473 request_cleanup(resource, islocked)
5474 	int resource;
5475 	int islocked;
5476 {
5477 	struct thread *td = curthread;
5478 
5479 	/*
5480 	 * We never hold up the filesystem syncer process.
5481 	 */
5482 	if (td == filesys_syncer)
5483 		return (0);
5484 	/*
5485 	 * First check to see if the work list has gotten backlogged.
5486 	 * If it has, co-opt this process to help clean up two entries.
5487 	 * Because this process may hold inodes locked, we cannot
5488 	 * handle any remove requests that might block on a locked
5489 	 * inode as that could lead to deadlock.
5490 	 */
5491 	if (num_on_worklist > max_softdeps / 10) {
5492 		if (islocked)
5493 			FREE_LOCK(&lk);
5494 		process_worklist_item(NULL, LK_NOWAIT);
5495 		process_worklist_item(NULL, LK_NOWAIT);
5496 		stat_worklist_push += 2;
5497 		if (islocked)
5498 			ACQUIRE_LOCK(&lk);
5499 		return(1);
5500 	}
5501 	/*
5502 	 * Next, we attempt to speed up the syncer process. If that
5503 	 * is successful, then we allow the process to continue.
5504 	 */
5505 	if (speedup_syncer() && resource != FLUSH_REMOVE_WAIT)
5506 		return(0);
5507 	/*
5508 	 * If we are resource constrained on inode dependencies, try
5509 	 * flushing some dirty inodes. Otherwise, we are constrained
5510 	 * by file deletions, so try accelerating flushes of directories
5511 	 * with removal dependencies. We would like to do the cleanup
5512 	 * here, but we probably hold an inode locked at this point and
5513 	 * that might deadlock against one that we try to clean. So,
5514 	 * the best that we can do is request the syncer daemon to do
5515 	 * the cleanup for us.
5516 	 */
5517 	switch (resource) {
5518 
5519 	case FLUSH_INODES:
5520 		stat_ino_limit_push += 1;
5521 		req_clear_inodedeps += 1;
5522 		stat_countp = &stat_ino_limit_hit;
5523 		break;
5524 
5525 	case FLUSH_REMOVE:
5526 	case FLUSH_REMOVE_WAIT:
5527 		stat_blk_limit_push += 1;
5528 		req_clear_remove += 1;
5529 		stat_countp = &stat_blk_limit_hit;
5530 		break;
5531 
5532 	default:
5533 		if (islocked)
5534 			FREE_LOCK(&lk);
5535 		panic("request_cleanup: unknown type");
5536 	}
5537 	/*
5538 	 * Hopefully the syncer daemon will catch up and awaken us.
5539 	 * We wait at most tickdelay before proceeding in any case.
5540 	 */
5541 	if (islocked == 0)
5542 		ACQUIRE_LOCK(&lk);
5543 	proc_waiting += 1;
5544 	if (handle.callout == NULL)
5545 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5546 	interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, NULL, PPAUSE,
5547 	    "softupdate", 0);
5548 	proc_waiting -= 1;
5549 	if (islocked == 0)
5550 		FREE_LOCK(&lk);
5551 	return (1);
5552 }
5553 
5554 /*
5555  * Awaken processes pausing in request_cleanup and clear proc_waiting
5556  * to indicate that there is no longer a timer running.
5557  */
5558 static void
5559 pause_timer(arg)
5560 	void *arg;
5561 {
5562 
5563 	*stat_countp += 1;
5564 	wakeup_one(&proc_waiting);
5565 	if (proc_waiting > 0)
5566 		handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
5567 	else
5568 		handle.callout = NULL;
5569 }
5570 
5571 /*
5572  * Flush out a directory with at least one removal dependency in an effort to
5573  * reduce the number of dirrem, freefile, and freeblks dependency structures.
5574  */
5575 static void
5576 clear_remove(td)
5577 	struct thread *td;
5578 {
5579 	struct pagedep_hashhead *pagedephd;
5580 	struct pagedep *pagedep;
5581 	static int next = 0;
5582 	struct mount *mp;
5583 	struct vnode *vp;
5584 	int error, cnt;
5585 	ino_t ino;
5586 
5587 	ACQUIRE_LOCK(&lk);
5588 	for (cnt = 0; cnt < pagedep_hash; cnt++) {
5589 		pagedephd = &pagedep_hashtbl[next++];
5590 		if (next >= pagedep_hash)
5591 			next = 0;
5592 		LIST_FOREACH(pagedep, pagedephd, pd_hash) {
5593 			if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
5594 				continue;
5595 			mp = pagedep->pd_mnt;
5596 			ino = pagedep->pd_ino;
5597 			if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5598 				continue;
5599 			FREE_LOCK(&lk);
5600 			if ((error = VFS_VGET(mp, ino, LK_EXCLUSIVE, &vp))) {
5601 				softdep_error("clear_remove: vget", error);
5602 				vn_finished_write(mp);
5603 				return;
5604 			}
5605 			if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)))
5606 				softdep_error("clear_remove: fsync", error);
5607 			drain_output(vp, 0);
5608 			vput(vp);
5609 			vn_finished_write(mp);
5610 			return;
5611 		}
5612 	}
5613 	FREE_LOCK(&lk);
5614 }
5615 
5616 /*
5617  * Clear out a block of dirty inodes in an effort to reduce
5618  * the number of inodedep dependency structures.
5619  */
5620 static void
5621 clear_inodedeps(td)
5622 	struct thread *td;
5623 {
5624 	struct inodedep_hashhead *inodedephd;
5625 	struct inodedep *inodedep;
5626 	static int next = 0;
5627 	struct mount *mp;
5628 	struct vnode *vp;
5629 	struct fs *fs;
5630 	int error, cnt;
5631 	ino_t firstino, lastino, ino;
5632 
5633 	ACQUIRE_LOCK(&lk);
5634 	/*
5635 	 * Pick a random inode dependency to be cleared.
5636 	 * We will then gather up all the inodes in its block
5637 	 * that have dependencies and flush them out.
5638 	 */
5639 	for (cnt = 0; cnt < inodedep_hash; cnt++) {
5640 		inodedephd = &inodedep_hashtbl[next++];
5641 		if (next >= inodedep_hash)
5642 			next = 0;
5643 		if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
5644 			break;
5645 	}
5646 	if (inodedep == NULL)
5647 		return;
5648 	/*
5649 	 * Ugly code to find mount point given pointer to superblock.
5650 	 */
5651 	fs = inodedep->id_fs;
5652 	TAILQ_FOREACH(mp, &mountlist, mnt_list)
5653 		if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs)
5654 			break;
5655 	/*
5656 	 * Find the last inode in the block with dependencies.
5657 	 */
5658 	firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
5659 	for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
5660 		if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
5661 			break;
5662 	/*
5663 	 * Asynchronously push all but the last inode with dependencies.
5664 	 * Synchronously push the last inode with dependencies to ensure
5665 	 * that the inode block gets written to free up the inodedeps.
5666 	 */
5667 	for (ino = firstino; ino <= lastino; ino++) {
5668 		if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
5669 			continue;
5670 		FREE_LOCK(&lk);
5671 		if (vn_start_write(NULL, &mp, V_NOWAIT) != 0)
5672 			continue;
5673 		if ((error = VFS_VGET(mp, ino, LK_EXCLUSIVE, &vp)) != 0) {
5674 			softdep_error("clear_inodedeps: vget", error);
5675 			vn_finished_write(mp);
5676 			return;
5677 		}
5678 		if (ino == lastino) {
5679 			if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_WAIT, td)))
5680 				softdep_error("clear_inodedeps: fsync1", error);
5681 		} else {
5682 			if ((error = VOP_FSYNC(vp, td->td_ucred, MNT_NOWAIT, td)))
5683 				softdep_error("clear_inodedeps: fsync2", error);
5684 			drain_output(vp, 0);
5685 		}
5686 		vput(vp);
5687 		vn_finished_write(mp);
5688 		ACQUIRE_LOCK(&lk);
5689 	}
5690 	FREE_LOCK(&lk);
5691 }
5692 
5693 /*
5694  * Function to determine if the buffer has outstanding dependencies
5695  * that will cause a roll-back if the buffer is written. If wantcount
5696  * is set, return number of dependencies, otherwise just yes or no.
5697  */
5698 static int
5699 softdep_count_dependencies(bp, wantcount)
5700 	struct buf *bp;
5701 	int wantcount;
5702 {
5703 	struct worklist *wk;
5704 	struct inodedep *inodedep;
5705 	struct indirdep *indirdep;
5706 	struct allocindir *aip;
5707 	struct pagedep *pagedep;
5708 	struct diradd *dap;
5709 	int i, retval;
5710 
5711 	retval = 0;
5712 	ACQUIRE_LOCK(&lk);
5713 	LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5714 		switch (wk->wk_type) {
5715 
5716 		case D_INODEDEP:
5717 			inodedep = WK_INODEDEP(wk);
5718 			if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5719 				/* bitmap allocation dependency */
5720 				retval += 1;
5721 				if (!wantcount)
5722 					goto out;
5723 			}
5724 			if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
5725 				/* direct block pointer dependency */
5726 				retval += 1;
5727 				if (!wantcount)
5728 					goto out;
5729 			}
5730 			if (TAILQ_FIRST(&inodedep->id_extupdt)) {
5731 				/* direct block pointer dependency */
5732 				retval += 1;
5733 				if (!wantcount)
5734 					goto out;
5735 			}
5736 			continue;
5737 
5738 		case D_INDIRDEP:
5739 			indirdep = WK_INDIRDEP(wk);
5740 
5741 			LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
5742 				/* indirect block pointer dependency */
5743 				retval += 1;
5744 				if (!wantcount)
5745 					goto out;
5746 			}
5747 			continue;
5748 
5749 		case D_PAGEDEP:
5750 			pagedep = WK_PAGEDEP(wk);
5751 			for (i = 0; i < DAHASHSZ; i++) {
5752 
5753 				LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
5754 					/* directory entry dependency */
5755 					retval += 1;
5756 					if (!wantcount)
5757 						goto out;
5758 				}
5759 			}
5760 			continue;
5761 
5762 		case D_BMSAFEMAP:
5763 		case D_ALLOCDIRECT:
5764 		case D_ALLOCINDIR:
5765 		case D_MKDIR:
5766 			/* never a dependency on these blocks */
5767 			continue;
5768 
5769 		default:
5770 			FREE_LOCK(&lk);
5771 			panic("softdep_check_for_rollback: Unexpected type %s",
5772 			    TYPENAME(wk->wk_type));
5773 			/* NOTREACHED */
5774 		}
5775 	}
5776 out:
5777 	FREE_LOCK(&lk);
5778 	return retval;
5779 }
5780 
5781 /*
5782  * Acquire exclusive access to a buffer.
5783  * Must be called with splbio blocked.
5784  * Return 1 if buffer was acquired.
5785  */
5786 static int
5787 getdirtybuf(bpp, waitfor)
5788 	struct buf **bpp;
5789 	int waitfor;
5790 {
5791 	struct buf *bp;
5792 	int error;
5793 
5794 	for (;;) {
5795 		if ((bp = *bpp) == NULL)
5796 			return (0);
5797 		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
5798 			if ((bp->b_xflags & BX_BKGRDINPROG) == 0)
5799 				break;
5800 			BUF_UNLOCK(bp);
5801 			if (waitfor != MNT_WAIT)
5802 				return (0);
5803 			bp->b_xflags |= BX_BKGRDWAIT;
5804 			interlocked_sleep(&lk, SLEEP, &bp->b_xflags, NULL,
5805 			    PRIBIO, "getbuf", 0);
5806 			continue;
5807 		}
5808 		if (waitfor != MNT_WAIT)
5809 			return (0);
5810 		error = interlocked_sleep(&lk, LOCKBUF, bp, NULL,
5811 		    LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0);
5812 		if (error != ENOLCK) {
5813 			FREE_LOCK(&lk);
5814 			panic("getdirtybuf: inconsistent lock");
5815 		}
5816 	}
5817 	if ((bp->b_flags & B_DELWRI) == 0) {
5818 		BUF_UNLOCK(bp);
5819 		return (0);
5820 	}
5821 	bremfree(bp);
5822 	return (1);
5823 }
5824 
5825 /*
5826  * Wait for pending output on a vnode to complete.
5827  * Must be called with vnode locked.
5828  */
5829 static void
5830 drain_output(vp, islocked)
5831 	struct vnode *vp;
5832 	int islocked;
5833 {
5834 
5835 	if (!islocked)
5836 		ACQUIRE_LOCK(&lk);
5837 	VI_LOCK(vp);
5838 	while (vp->v_numoutput) {
5839 		vp->v_iflag |= VI_BWAIT;
5840 		interlocked_sleep(&lk, SLEEP, (caddr_t)&vp->v_numoutput,
5841 		    VI_MTX(vp), PRIBIO + 1, "drainvp", 0);
5842 	}
5843 	VI_UNLOCK(vp);
5844 	if (!islocked)
5845 		FREE_LOCK(&lk);
5846 }
5847 
5848 /*
5849  * Called whenever a buffer that is being invalidated or reallocated
5850  * contains dependencies. This should only happen if an I/O error has
5851  * occurred. The routine is called with the buffer locked.
5852  */
5853 static void
5854 softdep_deallocate_dependencies(bp)
5855 	struct buf *bp;
5856 {
5857 
5858 	if ((bp->b_ioflags & BIO_ERROR) == 0)
5859 		panic("softdep_deallocate_dependencies: dangling deps");
5860 	softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
5861 	panic("softdep_deallocate_dependencies: unrecovered I/O error");
5862 }
5863 
5864 /*
5865  * Function to handle asynchronous write errors in the filesystem.
5866  */
5867 static void
5868 softdep_error(func, error)
5869 	char *func;
5870 	int error;
5871 {
5872 
5873 	/* XXX should do something better! */
5874 	printf("%s: got error %d while accessing filesystem\n", func, error);
5875 }
5876