A number of structures related to UFS and QUOTAS have changed name.
[dragonfly.git] / sys / vfs / ufs / ffs_softdep.c
... / ...
CommitLineData
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 * $FreeBSD: src/sys/ufs/ffs/ffs_softdep.c,v 1.57.2.11 2002/02/05 18:46:53 dillon Exp $
40 * $DragonFly: src/sys/vfs/ufs/ffs_softdep.c,v 1.36 2006/03/24 18:35:34 dillon Exp $
41 */
42
43/*
44 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
45 */
46#ifndef DIAGNOSTIC
47#define DIAGNOSTIC
48#endif
49#ifndef DEBUG
50#define DEBUG
51#endif
52
53#include <sys/param.h>
54#include <sys/kernel.h>
55#include <sys/systm.h>
56#include <sys/buf.h>
57#include <sys/malloc.h>
58#include <sys/mount.h>
59#include <sys/proc.h>
60#include <sys/syslog.h>
61#include <sys/vnode.h>
62#include <sys/conf.h>
63#include <sys/buf2.h>
64#include <machine/inttypes.h>
65#include "dir.h"
66#include "quota.h"
67#include "inode.h"
68#include "ufsmount.h"
69#include "fs.h"
70#include "softdep.h"
71#include "ffs_extern.h"
72#include "ufs_extern.h"
73
74#include <sys/thread2.h>
75
76/*
77 * These definitions need to be adapted to the system to which
78 * this file is being ported.
79 */
80/*
81 * malloc types defined for the softdep system.
82 */
83MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
84MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
85MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
86MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
87MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
88MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
89MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
90MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
91MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
92MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
93MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
94MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
95MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
96
97#define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE)
98
99#define D_PAGEDEP 0
100#define D_INODEDEP 1
101#define D_NEWBLK 2
102#define D_BMSAFEMAP 3
103#define D_ALLOCDIRECT 4
104#define D_INDIRDEP 5
105#define D_ALLOCINDIR 6
106#define D_FREEFRAG 7
107#define D_FREEBLKS 8
108#define D_FREEFILE 9
109#define D_DIRADD 10
110#define D_MKDIR 11
111#define D_DIRREM 12
112#define D_LAST D_DIRREM
113
114/*
115 * translate from workitem type to memory type
116 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
117 */
118static struct malloc_type *memtype[] = {
119 M_PAGEDEP,
120 M_INODEDEP,
121 M_NEWBLK,
122 M_BMSAFEMAP,
123 M_ALLOCDIRECT,
124 M_INDIRDEP,
125 M_ALLOCINDIR,
126 M_FREEFRAG,
127 M_FREEBLKS,
128 M_FREEFILE,
129 M_DIRADD,
130 M_MKDIR,
131 M_DIRREM
132};
133
134#define DtoM(type) (memtype[type])
135
136/*
137 * Names of malloc types.
138 */
139#define TYPENAME(type) \
140 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
141/*
142 * End system adaptaion definitions.
143 */
144
145/*
146 * Internal function prototypes.
147 */
148static void softdep_error(char *, int);
149static void drain_output(struct vnode *, int);
150static int getdirtybuf(struct buf **, int);
151static void clear_remove(struct thread *);
152static void clear_inodedeps(struct thread *);
153static int flush_pagedep_deps(struct vnode *, struct mount *,
154 struct diraddhd *);
155static int flush_inodedep_deps(struct fs *, ino_t);
156static int handle_written_filepage(struct pagedep *, struct buf *);
157static void diradd_inode_written(struct diradd *, struct inodedep *);
158static int handle_written_inodeblock(struct inodedep *, struct buf *);
159static void handle_allocdirect_partdone(struct allocdirect *);
160static void handle_allocindir_partdone(struct allocindir *);
161static void initiate_write_filepage(struct pagedep *, struct buf *);
162static void handle_written_mkdir(struct mkdir *, int);
163static void initiate_write_inodeblock(struct inodedep *, struct buf *);
164static void handle_workitem_freefile(struct freefile *);
165static void handle_workitem_remove(struct dirrem *);
166static struct dirrem *newdirrem(struct buf *, struct inode *,
167 struct inode *, int, struct dirrem **);
168static void free_diradd(struct diradd *);
169static void free_allocindir(struct allocindir *, struct inodedep *);
170static int indir_trunc (struct inode *, off_t, int, ufs_lbn_t, long *);
171static void deallocate_dependencies(struct buf *, struct inodedep *);
172static void free_allocdirect(struct allocdirectlst *,
173 struct allocdirect *, int);
174static int check_inode_unwritten(struct inodedep *);
175static int free_inodedep(struct inodedep *);
176static void handle_workitem_freeblocks(struct freeblks *);
177static void merge_inode_lists(struct inodedep *);
178static void setup_allocindir_phase2(struct buf *, struct inode *,
179 struct allocindir *);
180static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t,
181 ufs_daddr_t);
182static void handle_workitem_freefrag(struct freefrag *);
183static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long);
184static void allocdirect_merge(struct allocdirectlst *,
185 struct allocdirect *, struct allocdirect *);
186static struct bmsafemap *bmsafemap_lookup(struct buf *);
187static int newblk_lookup(struct fs *, ufs_daddr_t, int,
188 struct newblk **);
189static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
190static int pagedep_lookup(struct inode *, ufs_lbn_t, int,
191 struct pagedep **);
192static void pause_timer(void *);
193static int request_cleanup(int, int);
194static int process_worklist_item(struct mount *, int);
195static void add_to_worklist(struct worklist *);
196
197/*
198 * Exported softdep operations.
199 */
200static void softdep_disk_io_initiation(struct buf *);
201static void softdep_disk_write_complete(struct buf *);
202static void softdep_deallocate_dependencies(struct buf *);
203static int softdep_fsync(struct vnode *);
204static int softdep_process_worklist(struct mount *);
205static void softdep_move_dependencies(struct buf *, struct buf *);
206static int softdep_count_dependencies(struct buf *bp, int);
207
208static struct bio_ops softdep_bioops = {
209 softdep_disk_io_initiation, /* io_start */
210 softdep_disk_write_complete, /* io_complete */
211 softdep_deallocate_dependencies, /* io_deallocate */
212 softdep_fsync, /* io_fsync */
213 softdep_process_worklist, /* io_sync */
214 softdep_move_dependencies, /* io_movedeps */
215 softdep_count_dependencies, /* io_countdeps */
216};
217
218/*
219 * Locking primitives.
220 *
221 * For a uniprocessor, all we need to do is protect against disk
222 * interrupts. For a multiprocessor, this lock would have to be
223 * a mutex. A single mutex is used throughout this file, though
224 * finer grain locking could be used if contention warranted it.
225 *
226 * For a multiprocessor, the sleep call would accept a lock and
227 * release it after the sleep processing was complete. In a uniprocessor
228 * implementation there is no such interlock, so we simple mark
229 * the places where it needs to be done with the `interlocked' form
230 * of the lock calls. Since the uniprocessor sleep already interlocks
231 * the spl, there is nothing that really needs to be done.
232 */
233#ifndef /* NOT */ DEBUG
234static struct lockit {
235} lk = { 0 };
236#define ACQUIRE_LOCK(lk) crit_enter_id("softupdates");
237#define FREE_LOCK(lk) crit_exit_id("softupdates");
238
239#else /* DEBUG */
240#define NOHOLDER ((struct thread *)-1)
241#define SPECIAL_FLAG ((struct thread *)-2)
242static struct lockit {
243 int lkt_spl;
244 struct thread *lkt_held;
245} lk = { 0, NOHOLDER };
246static int lockcnt;
247
248static void acquire_lock(struct lockit *);
249static void free_lock(struct lockit *);
250void softdep_panic(char *);
251
252#define ACQUIRE_LOCK(lk) acquire_lock(lk)
253#define FREE_LOCK(lk) free_lock(lk)
254
255static void
256acquire_lock(lk)
257 struct lockit *lk;
258{
259 thread_t holder;
260
261 if (lk->lkt_held != NOHOLDER) {
262 holder = lk->lkt_held;
263 FREE_LOCK(lk);
264 if (holder == curthread)
265 panic("softdep_lock: locking against myself");
266 else
267 panic("softdep_lock: lock held by %p", holder);
268 }
269 crit_enter_id("softupdates");
270 lk->lkt_held = curthread;
271 lockcnt++;
272}
273
274static void
275free_lock(lk)
276 struct lockit *lk;
277{
278
279 if (lk->lkt_held == NOHOLDER)
280 panic("softdep_unlock: lock not held");
281 lk->lkt_held = NOHOLDER;
282 crit_exit_id("softupdates");
283}
284
285/*
286 * Function to release soft updates lock and panic.
287 */
288void
289softdep_panic(msg)
290 char *msg;
291{
292
293 if (lk.lkt_held != NOHOLDER)
294 FREE_LOCK(&lk);
295 panic(msg);
296}
297#endif /* DEBUG */
298
299static int interlocked_sleep(struct lockit *, int, void *, int,
300 const char *, int);
301
302/*
303 * When going to sleep, we must save our SPL so that it does
304 * not get lost if some other process uses the lock while we
305 * are sleeping. We restore it after we have slept. This routine
306 * wraps the interlocking with functions that sleep. The list
307 * below enumerates the available set of operations.
308 */
309#define UNKNOWN 0
310#define SLEEP 1
311#define LOCKBUF 2
312
313static int
314interlocked_sleep(lk, op, ident, flags, wmesg, timo)
315 struct lockit *lk;
316 int op;
317 void *ident;
318 int flags;
319 const char *wmesg;
320 int timo;
321{
322 thread_t holder;
323 int s, retval;
324
325 s = lk->lkt_spl;
326# ifdef DEBUG
327 if (lk->lkt_held == NOHOLDER)
328 panic("interlocked_sleep: lock not held");
329 lk->lkt_held = NOHOLDER;
330# endif /* DEBUG */
331 switch (op) {
332 case SLEEP:
333 retval = tsleep(ident, flags, wmesg, timo);
334 break;
335 case LOCKBUF:
336 retval = BUF_LOCK((struct buf *)ident, flags);
337 break;
338 default:
339 panic("interlocked_sleep: unknown operation");
340 }
341# ifdef DEBUG
342 if (lk->lkt_held != NOHOLDER) {
343 holder = lk->lkt_held;
344 FREE_LOCK(lk);
345 if (holder == curthread)
346 panic("interlocked_sleep: locking against self");
347 else
348 panic("interlocked_sleep: lock held by %p", holder);
349 }
350 lk->lkt_held = curthread;
351 lockcnt++;
352# endif /* DEBUG */
353 lk->lkt_spl = s;
354 return (retval);
355}
356
357/*
358 * Place holder for real semaphores.
359 */
360struct sema {
361 int value;
362 thread_t holder;
363 char *name;
364 int prio;
365 int timo;
366};
367static void sema_init(struct sema *, char *, int, int);
368static int sema_get(struct sema *, struct lockit *);
369static void sema_release(struct sema *);
370
371static void
372sema_init(semap, name, prio, timo)
373 struct sema *semap;
374 char *name;
375 int prio, timo;
376{
377
378 semap->holder = NOHOLDER;
379 semap->value = 0;
380 semap->name = name;
381 semap->prio = prio;
382 semap->timo = timo;
383}
384
385static int
386sema_get(semap, interlock)
387 struct sema *semap;
388 struct lockit *interlock;
389{
390
391 if (semap->value++ > 0) {
392 if (interlock != NULL) {
393 interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
394 semap->prio, semap->name, semap->timo);
395 FREE_LOCK(interlock);
396 } else {
397 tsleep((caddr_t)semap, semap->prio, semap->name,
398 semap->timo);
399 }
400 return (0);
401 }
402 semap->holder = curthread;
403 if (interlock != NULL)
404 FREE_LOCK(interlock);
405 return (1);
406}
407
408static void
409sema_release(semap)
410 struct sema *semap;
411{
412
413 if (semap->value <= 0 || semap->holder != curthread) {
414 if (lk.lkt_held != NOHOLDER)
415 FREE_LOCK(&lk);
416 panic("sema_release: not held");
417 }
418 if (--semap->value > 0) {
419 semap->value = 0;
420 wakeup(semap);
421 }
422 semap->holder = NOHOLDER;
423}
424
425/*
426 * Worklist queue management.
427 * These routines require that the lock be held.
428 */
429#ifndef /* NOT */ DEBUG
430#define WORKLIST_INSERT(head, item) do { \
431 (item)->wk_state |= ONWORKLIST; \
432 LIST_INSERT_HEAD(head, item, wk_list); \
433} while (0)
434#define WORKLIST_REMOVE(item) do { \
435 (item)->wk_state &= ~ONWORKLIST; \
436 LIST_REMOVE(item, wk_list); \
437} while (0)
438#define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
439
440#else /* DEBUG */
441static void worklist_insert(struct workhead *, struct worklist *);
442static void worklist_remove(struct worklist *);
443static void workitem_free(struct worklist *, int);
444
445#define WORKLIST_INSERT(head, item) worklist_insert(head, item)
446#define WORKLIST_REMOVE(item) worklist_remove(item)
447#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
448
449static void
450worklist_insert(head, item)
451 struct workhead *head;
452 struct worklist *item;
453{
454
455 if (lk.lkt_held == NOHOLDER)
456 panic("worklist_insert: lock not held");
457 if (item->wk_state & ONWORKLIST) {
458 FREE_LOCK(&lk);
459 panic("worklist_insert: already on list");
460 }
461 item->wk_state |= ONWORKLIST;
462 LIST_INSERT_HEAD(head, item, wk_list);
463}
464
465static void
466worklist_remove(item)
467 struct worklist *item;
468{
469
470 if (lk.lkt_held == NOHOLDER)
471 panic("worklist_remove: lock not held");
472 if ((item->wk_state & ONWORKLIST) == 0) {
473 FREE_LOCK(&lk);
474 panic("worklist_remove: not on list");
475 }
476 item->wk_state &= ~ONWORKLIST;
477 LIST_REMOVE(item, wk_list);
478}
479
480static void
481workitem_free(item, type)
482 struct worklist *item;
483 int type;
484{
485
486 if (item->wk_state & ONWORKLIST) {
487 if (lk.lkt_held != NOHOLDER)
488 FREE_LOCK(&lk);
489 panic("workitem_free: still on list");
490 }
491 if (item->wk_type != type) {
492 if (lk.lkt_held != NOHOLDER)
493 FREE_LOCK(&lk);
494 panic("workitem_free: type mismatch");
495 }
496 FREE(item, DtoM(type));
497}
498#endif /* DEBUG */
499
500/*
501 * Workitem queue management
502 */
503static struct workhead softdep_workitem_pending;
504static int num_on_worklist; /* number of worklist items to be processed */
505static int softdep_worklist_busy; /* 1 => trying to do unmount */
506static int softdep_worklist_req; /* serialized waiters */
507static int max_softdeps; /* maximum number of structs before slowdown */
508static int tickdelay = 2; /* number of ticks to pause during slowdown */
509static int *stat_countp; /* statistic to count in proc_waiting timeout */
510static int proc_waiting; /* tracks whether we have a timeout posted */
511static struct callout handle; /* handle on posted proc_waiting timeout */
512static struct thread *filesys_syncer; /* proc of filesystem syncer process */
513static int req_clear_inodedeps; /* syncer process flush some inodedeps */
514#define FLUSH_INODES 1
515static int req_clear_remove; /* syncer process flush some freeblks */
516#define FLUSH_REMOVE 2
517/*
518 * runtime statistics
519 */
520static int stat_worklist_push; /* number of worklist cleanups */
521static int stat_blk_limit_push; /* number of times block limit neared */
522static int stat_ino_limit_push; /* number of times inode limit neared */
523static int stat_blk_limit_hit; /* number of times block slowdown imposed */
524static int stat_ino_limit_hit; /* number of times inode slowdown imposed */
525static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */
526static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */
527static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */
528static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
529static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */
530#ifdef DEBUG
531#include <vm/vm.h>
532#include <sys/sysctl.h>
533SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
534SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
535SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
536SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
537SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
538SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
539SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
540SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
541SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
542SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
543SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
544SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
545#endif /* DEBUG */
546
547/*
548 * Add an item to the end of the work queue.
549 * This routine requires that the lock be held.
550 * This is the only routine that adds items to the list.
551 * The following routine is the only one that removes items
552 * and does so in order from first to last.
553 */
554static void
555add_to_worklist(wk)
556 struct worklist *wk;
557{
558 static struct worklist *worklist_tail;
559
560 if (wk->wk_state & ONWORKLIST) {
561 if (lk.lkt_held != NOHOLDER)
562 FREE_LOCK(&lk);
563 panic("add_to_worklist: already on list");
564 }
565 wk->wk_state |= ONWORKLIST;
566 if (LIST_FIRST(&softdep_workitem_pending) == NULL)
567 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
568 else
569 LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
570 worklist_tail = wk;
571 num_on_worklist += 1;
572}
573
574/*
575 * Process that runs once per second to handle items in the background queue.
576 *
577 * Note that we ensure that everything is done in the order in which they
578 * appear in the queue. The code below depends on this property to ensure
579 * that blocks of a file are freed before the inode itself is freed. This
580 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
581 * until all the old ones have been purged from the dependency lists.
582 */
583static int
584softdep_process_worklist(matchmnt)
585 struct mount *matchmnt;
586{
587 thread_t td = curthread;
588 int matchcnt, loopcount;
589 long starttime;
590
591 /*
592 * Record the process identifier of our caller so that we can give
593 * this process preferential treatment in request_cleanup below.
594 */
595 filesys_syncer = td;
596 matchcnt = 0;
597
598 /*
599 * There is no danger of having multiple processes run this
600 * code, but we have to single-thread it when softdep_flushfiles()
601 * is in operation to get an accurate count of the number of items
602 * related to its mount point that are in the list.
603 */
604 if (matchmnt == NULL) {
605 if (softdep_worklist_busy < 0)
606 return(-1);
607 softdep_worklist_busy += 1;
608 }
609
610 /*
611 * If requested, try removing inode or removal dependencies.
612 */
613 if (req_clear_inodedeps) {
614 clear_inodedeps(td);
615 req_clear_inodedeps -= 1;
616 wakeup_one(&proc_waiting);
617 }
618 if (req_clear_remove) {
619 clear_remove(td);
620 req_clear_remove -= 1;
621 wakeup_one(&proc_waiting);
622 }
623 loopcount = 1;
624 starttime = time_second;
625 while (num_on_worklist > 0) {
626 matchcnt += process_worklist_item(matchmnt, 0);
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;
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 */
677static int
678process_worklist_item(matchmnt, flags)
679 struct mount *matchmnt;
680 int flags;
681{
682 struct worklist *wk;
683 struct dirrem *dirrem;
684 struct fs *matchfs;
685 struct vnode *vp;
686 int matchcnt = 0;
687
688 matchfs = NULL;
689 if (matchmnt != NULL)
690 matchfs = VFSTOUFS(matchmnt)->um_fs;
691 ACQUIRE_LOCK(&lk);
692 /*
693 * Normally we just process each item on the worklist in order.
694 * However, if we are in a situation where we cannot lock any
695 * inodes, we have to skip over any dirrem requests whose
696 * vnodes are resident and locked.
697 */
698 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
699 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
700 break;
701 dirrem = WK_DIRREM(wk);
702 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev,
703 dirrem->dm_oldinum);
704 if (vp == NULL || !VOP_ISLOCKED(vp, curthread))
705 break;
706 }
707 if (wk == 0) {
708 FREE_LOCK(&lk);
709 return (0);
710 }
711 WORKLIST_REMOVE(wk);
712 num_on_worklist -= 1;
713 FREE_LOCK(&lk);
714 switch (wk->wk_type) {
715
716 case D_DIRREM:
717 /* removal of a directory entry */
718 if (WK_DIRREM(wk)->dm_mnt == matchmnt)
719 matchcnt += 1;
720 handle_workitem_remove(WK_DIRREM(wk));
721 break;
722
723 case D_FREEBLKS:
724 /* releasing blocks and/or fragments from a file */
725 if (WK_FREEBLKS(wk)->fb_fs == matchfs)
726 matchcnt += 1;
727 handle_workitem_freeblocks(WK_FREEBLKS(wk));
728 break;
729
730 case D_FREEFRAG:
731 /* releasing a fragment when replaced as a file grows */
732 if (WK_FREEFRAG(wk)->ff_fs == matchfs)
733 matchcnt += 1;
734 handle_workitem_freefrag(WK_FREEFRAG(wk));
735 break;
736
737 case D_FREEFILE:
738 /* releasing an inode when its link count drops to 0 */
739 if (WK_FREEFILE(wk)->fx_fs == matchfs)
740 matchcnt += 1;
741 handle_workitem_freefile(WK_FREEFILE(wk));
742 break;
743
744 default:
745 panic("%s_process_worklist: Unknown type %s",
746 "softdep", TYPENAME(wk->wk_type));
747 /* NOTREACHED */
748 }
749 return (matchcnt);
750}
751
752/*
753 * Move dependencies from one buffer to another.
754 */
755static void
756softdep_move_dependencies(oldbp, newbp)
757 struct buf *oldbp;
758 struct buf *newbp;
759{
760 struct worklist *wk, *wktail;
761
762 if (LIST_FIRST(&newbp->b_dep) != NULL)
763 panic("softdep_move_dependencies: need merge code");
764 wktail = 0;
765 ACQUIRE_LOCK(&lk);
766 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
767 LIST_REMOVE(wk, wk_list);
768 if (wktail == 0)
769 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
770 else
771 LIST_INSERT_AFTER(wktail, wk, wk_list);
772 wktail = wk;
773 }
774 FREE_LOCK(&lk);
775}
776
777/*
778 * Purge the work list of all items associated with a particular mount point.
779 */
780int
781softdep_flushfiles(struct mount *oldmnt, int flags, struct thread *td)
782{
783 struct vnode *devvp;
784 int error, loopcnt;
785
786 /*
787 * Await our turn to clear out the queue, then serialize access.
788 */
789 while (softdep_worklist_busy != 0) {
790 softdep_worklist_req += 1;
791 tsleep(&softdep_worklist_req, 0, "softflush", 0);
792 softdep_worklist_req -= 1;
793 }
794 softdep_worklist_busy = -1;
795
796 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) {
797 softdep_worklist_busy = 0;
798 if (softdep_worklist_req)
799 wakeup(&softdep_worklist_req);
800 return (error);
801 }
802 /*
803 * Alternately flush the block device associated with the mount
804 * point and process any dependencies that the flushing
805 * creates. In theory, this loop can happen at most twice,
806 * but we give it a few extra just to be sure.
807 */
808 devvp = VFSTOUFS(oldmnt)->um_devvp;
809 for (loopcnt = 10; loopcnt > 0; ) {
810 if (softdep_process_worklist(oldmnt) == 0) {
811 loopcnt--;
812 /*
813 * Do another flush in case any vnodes were brought in
814 * as part of the cleanup operations.
815 */
816 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
817 break;
818 /*
819 * If we still found nothing to do, we are really done.
820 */
821 if (softdep_process_worklist(oldmnt) == 0)
822 break;
823 }
824 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td);
825 error = VOP_FSYNC(devvp, MNT_WAIT, td);
826 VOP_UNLOCK(devvp, 0, td);
827 if (error)
828 break;
829 }
830 softdep_worklist_busy = 0;
831 if (softdep_worklist_req)
832 wakeup(&softdep_worklist_req);
833
834 /*
835 * If we are unmounting then it is an error to fail. If we
836 * are simply trying to downgrade to read-only, then filesystem
837 * activity can keep us busy forever, so we just fail with EBUSY.
838 */
839 if (loopcnt == 0) {
840 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
841 panic("softdep_flushfiles: looping");
842 error = EBUSY;
843 }
844 return (error);
845}
846
847/*
848 * Structure hashing.
849 *
850 * There are three types of structures that can be looked up:
851 * 1) pagedep structures identified by mount point, inode number,
852 * and logical block.
853 * 2) inodedep structures identified by mount point and inode number.
854 * 3) newblk structures identified by mount point and
855 * physical block number.
856 *
857 * The "pagedep" and "inodedep" dependency structures are hashed
858 * separately from the file blocks and inodes to which they correspond.
859 * This separation helps when the in-memory copy of an inode or
860 * file block must be replaced. It also obviates the need to access
861 * an inode or file page when simply updating (or de-allocating)
862 * dependency structures. Lookup of newblk structures is needed to
863 * find newly allocated blocks when trying to associate them with
864 * their allocdirect or allocindir structure.
865 *
866 * The lookup routines optionally create and hash a new instance when
867 * an existing entry is not found.
868 */
869#define DEPALLOC 0x0001 /* allocate structure if lookup fails */
870#define NODELAY 0x0002 /* cannot do background work */
871
872/*
873 * Structures and routines associated with pagedep caching.
874 */
875LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
876u_long pagedep_hash; /* size of hash table - 1 */
877#define PAGEDEP_HASH(mp, inum, lbn) \
878 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
879 pagedep_hash])
880static struct sema pagedep_in_progress;
881
882/*
883 * Look up a pagedep. Return 1 if found, 0 if not found.
884 * If not found, allocate if DEPALLOC flag is passed.
885 * Found or allocated entry is returned in pagedeppp.
886 * This routine must be called with splbio interrupts blocked.
887 */
888static int
889pagedep_lookup(ip, lbn, flags, pagedeppp)
890 struct inode *ip;
891 ufs_lbn_t lbn;
892 int flags;
893 struct pagedep **pagedeppp;
894{
895 struct pagedep *pagedep;
896 struct pagedep_hashhead *pagedephd;
897 struct mount *mp;
898 int i;
899
900#ifdef DEBUG
901 if (lk.lkt_held == NOHOLDER)
902 panic("pagedep_lookup: lock not held");
903#endif
904 mp = ITOV(ip)->v_mount;
905 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
906top:
907 LIST_FOREACH(pagedep, pagedephd, pd_hash)
908 if (ip->i_number == pagedep->pd_ino &&
909 lbn == pagedep->pd_lbn &&
910 mp == pagedep->pd_mnt)
911 break;
912 if (pagedep) {
913 *pagedeppp = pagedep;
914 return (1);
915 }
916 if ((flags & DEPALLOC) == 0) {
917 *pagedeppp = NULL;
918 return (0);
919 }
920 if (sema_get(&pagedep_in_progress, &lk) == 0) {
921 ACQUIRE_LOCK(&lk);
922 goto top;
923 }
924 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
925 M_SOFTDEP_FLAGS);
926 bzero(pagedep, sizeof(struct pagedep));
927 pagedep->pd_list.wk_type = D_PAGEDEP;
928 pagedep->pd_mnt = mp;
929 pagedep->pd_ino = ip->i_number;
930 pagedep->pd_lbn = lbn;
931 LIST_INIT(&pagedep->pd_dirremhd);
932 LIST_INIT(&pagedep->pd_pendinghd);
933 for (i = 0; i < DAHASHSZ; i++)
934 LIST_INIT(&pagedep->pd_diraddhd[i]);
935 ACQUIRE_LOCK(&lk);
936 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
937 sema_release(&pagedep_in_progress);
938 *pagedeppp = pagedep;
939 return (0);
940}
941
942/*
943 * Structures and routines associated with inodedep caching.
944 */
945LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
946static u_long inodedep_hash; /* size of hash table - 1 */
947static long num_inodedep; /* number of inodedep allocated */
948#define INODEDEP_HASH(fs, inum) \
949 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
950static struct sema inodedep_in_progress;
951
952/*
953 * Look up a inodedep. Return 1 if found, 0 if not found.
954 * If not found, allocate if DEPALLOC flag is passed.
955 * Found or allocated entry is returned in inodedeppp.
956 * This routine must be called with splbio interrupts blocked.
957 */
958static int
959inodedep_lookup(fs, inum, flags, inodedeppp)
960 struct fs *fs;
961 ino_t inum;
962 int flags;
963 struct inodedep **inodedeppp;
964{
965 struct inodedep *inodedep;
966 struct inodedep_hashhead *inodedephd;
967 int firsttry;
968
969#ifdef DEBUG
970 if (lk.lkt_held == NOHOLDER)
971 panic("inodedep_lookup: lock not held");
972#endif
973 firsttry = 1;
974 inodedephd = INODEDEP_HASH(fs, inum);
975top:
976 LIST_FOREACH(inodedep, inodedephd, id_hash)
977 if (inum == inodedep->id_ino && fs == inodedep->id_fs)
978 break;
979 if (inodedep) {
980 *inodedeppp = inodedep;
981 return (1);
982 }
983 if ((flags & DEPALLOC) == 0) {
984 *inodedeppp = NULL;
985 return (0);
986 }
987 /*
988 * If we are over our limit, try to improve the situation.
989 */
990 if (num_inodedep > max_softdeps && firsttry &&
991 speedup_syncer() == 0 && (flags & NODELAY) == 0 &&
992 request_cleanup(FLUSH_INODES, 1)) {
993 firsttry = 0;
994 goto top;
995 }
996 if (sema_get(&inodedep_in_progress, &lk) == 0) {
997 ACQUIRE_LOCK(&lk);
998 goto top;
999 }
1000 num_inodedep += 1;
1001 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1002 M_INODEDEP, M_SOFTDEP_FLAGS);
1003 inodedep->id_list.wk_type = D_INODEDEP;
1004 inodedep->id_fs = fs;
1005 inodedep->id_ino = inum;
1006 inodedep->id_state = ALLCOMPLETE;
1007 inodedep->id_nlinkdelta = 0;
1008 inodedep->id_savedino = NULL;
1009 inodedep->id_savedsize = -1;
1010 inodedep->id_buf = NULL;
1011 LIST_INIT(&inodedep->id_pendinghd);
1012 LIST_INIT(&inodedep->id_inowait);
1013 LIST_INIT(&inodedep->id_bufwait);
1014 TAILQ_INIT(&inodedep->id_inoupdt);
1015 TAILQ_INIT(&inodedep->id_newinoupdt);
1016 ACQUIRE_LOCK(&lk);
1017 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1018 sema_release(&inodedep_in_progress);
1019 *inodedeppp = inodedep;
1020 return (0);
1021}
1022
1023/*
1024 * Structures and routines associated with newblk caching.
1025 */
1026LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1027u_long newblk_hash; /* size of hash table - 1 */
1028#define NEWBLK_HASH(fs, inum) \
1029 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1030static struct sema newblk_in_progress;
1031
1032/*
1033 * Look up a newblk. Return 1 if found, 0 if not found.
1034 * If not found, allocate if DEPALLOC flag is passed.
1035 * Found or allocated entry is returned in newblkpp.
1036 */
1037static int
1038newblk_lookup(fs, newblkno, flags, newblkpp)
1039 struct fs *fs;
1040 ufs_daddr_t newblkno;
1041 int flags;
1042 struct newblk **newblkpp;
1043{
1044 struct newblk *newblk;
1045 struct newblk_hashhead *newblkhd;
1046
1047 newblkhd = NEWBLK_HASH(fs, newblkno);
1048top:
1049 LIST_FOREACH(newblk, newblkhd, nb_hash)
1050 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1051 break;
1052 if (newblk) {
1053 *newblkpp = newblk;
1054 return (1);
1055 }
1056 if ((flags & DEPALLOC) == 0) {
1057 *newblkpp = NULL;
1058 return (0);
1059 }
1060 if (sema_get(&newblk_in_progress, 0) == 0)
1061 goto top;
1062 MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1063 M_NEWBLK, M_SOFTDEP_FLAGS);
1064 newblk->nb_state = 0;
1065 newblk->nb_fs = fs;
1066 newblk->nb_newblkno = newblkno;
1067 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1068 sema_release(&newblk_in_progress);
1069 *newblkpp = newblk;
1070 return (0);
1071}
1072
1073/*
1074 * Executed during filesystem system initialization before
1075 * mounting any filesystems.
1076 */
1077void
1078softdep_initialize()
1079{
1080 callout_init(&handle);
1081 bioops = softdep_bioops; /* XXX hack */
1082
1083 LIST_INIT(&mkdirlisthd);
1084 LIST_INIT(&softdep_workitem_pending);
1085 max_softdeps = min(desiredvnodes * 8,
1086 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep)));
1087 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1088 &pagedep_hash);
1089 sema_init(&pagedep_in_progress, "pagedep", 0, 0);
1090 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1091 sema_init(&inodedep_in_progress, "inodedep", 0, 0);
1092 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1093 sema_init(&newblk_in_progress, "newblk", 0, 0);
1094}
1095
1096/*
1097 * Called at mount time to notify the dependency code that a
1098 * filesystem wishes to use it.
1099 */
1100int
1101softdep_mount(devvp, mp, fs)
1102 struct vnode *devvp;
1103 struct mount *mp;
1104 struct fs *fs;
1105{
1106 struct csum cstotal;
1107 struct cg *cgp;
1108 struct buf *bp;
1109 int error, cyl;
1110
1111 mp->mnt_flag &= ~MNT_ASYNC;
1112 mp->mnt_flag |= MNT_SOFTDEP;
1113 /*
1114 * When doing soft updates, the counters in the
1115 * superblock may have gotten out of sync, so we have
1116 * to scan the cylinder groups and recalculate them.
1117 */
1118 if (fs->fs_clean != 0)
1119 return (0);
1120 bzero(&cstotal, sizeof cstotal);
1121 for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1122 if ((error = bread(devvp, fsbtodoff(fs, cgtod(fs, cyl)),
1123 fs->fs_cgsize, &bp)) != 0) {
1124 brelse(bp);
1125 return (error);
1126 }
1127 cgp = (struct cg *)bp->b_data;
1128 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1129 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1130 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1131 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1132 fs->fs_cs(fs, cyl) = cgp->cg_cs;
1133 brelse(bp);
1134 }
1135#ifdef DEBUG
1136 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1137 printf("ffs_mountfs: superblock updated for soft updates\n");
1138#endif
1139 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1140 return (0);
1141}
1142
1143/*
1144 * Protecting the freemaps (or bitmaps).
1145 *
1146 * To eliminate the need to execute fsck before mounting a filesystem
1147 * after a power failure, one must (conservatively) guarantee that the
1148 * on-disk copy of the bitmaps never indicate that a live inode or block is
1149 * free. So, when a block or inode is allocated, the bitmap should be
1150 * updated (on disk) before any new pointers. When a block or inode is
1151 * freed, the bitmap should not be updated until all pointers have been
1152 * reset. The latter dependency is handled by the delayed de-allocation
1153 * approach described below for block and inode de-allocation. The former
1154 * dependency is handled by calling the following procedure when a block or
1155 * inode is allocated. When an inode is allocated an "inodedep" is created
1156 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1157 * Each "inodedep" is also inserted into the hash indexing structure so
1158 * that any additional link additions can be made dependent on the inode
1159 * allocation.
1160 *
1161 * The ufs filesystem maintains a number of free block counts (e.g., per
1162 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1163 * in addition to the bitmaps. These counts are used to improve efficiency
1164 * during allocation and therefore must be consistent with the bitmaps.
1165 * There is no convenient way to guarantee post-crash consistency of these
1166 * counts with simple update ordering, for two main reasons: (1) The counts
1167 * and bitmaps for a single cylinder group block are not in the same disk
1168 * sector. If a disk write is interrupted (e.g., by power failure), one may
1169 * be written and the other not. (2) Some of the counts are located in the
1170 * superblock rather than the cylinder group block. So, we focus our soft
1171 * updates implementation on protecting the bitmaps. When mounting a
1172 * filesystem, we recompute the auxiliary counts from the bitmaps.
1173 */
1174
1175/*
1176 * Called just after updating the cylinder group block to allocate an inode.
1177 */
1178void
1179softdep_setup_inomapdep(bp, ip, newinum)
1180 struct buf *bp; /* buffer for cylgroup block with inode map */
1181 struct inode *ip; /* inode related to allocation */
1182 ino_t newinum; /* new inode number being allocated */
1183{
1184 struct inodedep *inodedep;
1185 struct bmsafemap *bmsafemap;
1186
1187 /*
1188 * Create a dependency for the newly allocated inode.
1189 * Panic if it already exists as something is seriously wrong.
1190 * Otherwise add it to the dependency list for the buffer holding
1191 * the cylinder group map from which it was allocated.
1192 */
1193 ACQUIRE_LOCK(&lk);
1194 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1195 FREE_LOCK(&lk);
1196 panic("softdep_setup_inomapdep: found inode");
1197 }
1198 inodedep->id_buf = bp;
1199 inodedep->id_state &= ~DEPCOMPLETE;
1200 bmsafemap = bmsafemap_lookup(bp);
1201 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1202 FREE_LOCK(&lk);
1203}
1204
1205/*
1206 * Called just after updating the cylinder group block to
1207 * allocate block or fragment.
1208 */
1209void
1210softdep_setup_blkmapdep(bp, fs, newblkno)
1211 struct buf *bp; /* buffer for cylgroup block with block map */
1212 struct fs *fs; /* filesystem doing allocation */
1213 ufs_daddr_t newblkno; /* number of newly allocated block */
1214{
1215 struct newblk *newblk;
1216 struct bmsafemap *bmsafemap;
1217
1218 /*
1219 * Create a dependency for the newly allocated block.
1220 * Add it to the dependency list for the buffer holding
1221 * the cylinder group map from which it was allocated.
1222 */
1223 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1224 panic("softdep_setup_blkmapdep: found block");
1225 ACQUIRE_LOCK(&lk);
1226 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1227 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1228 FREE_LOCK(&lk);
1229}
1230
1231/*
1232 * Find the bmsafemap associated with a cylinder group buffer.
1233 * If none exists, create one. The buffer must be locked when
1234 * this routine is called and this routine must be called with
1235 * splbio interrupts blocked.
1236 */
1237static struct bmsafemap *
1238bmsafemap_lookup(bp)
1239 struct buf *bp;
1240{
1241 struct bmsafemap *bmsafemap;
1242 struct worklist *wk;
1243
1244#ifdef DEBUG
1245 if (lk.lkt_held == NOHOLDER)
1246 panic("bmsafemap_lookup: lock not held");
1247#endif
1248 LIST_FOREACH(wk, &bp->b_dep, wk_list)
1249 if (wk->wk_type == D_BMSAFEMAP)
1250 return (WK_BMSAFEMAP(wk));
1251 FREE_LOCK(&lk);
1252 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1253 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1254 bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1255 bmsafemap->sm_list.wk_state = 0;
1256 bmsafemap->sm_buf = bp;
1257 LIST_INIT(&bmsafemap->sm_allocdirecthd);
1258 LIST_INIT(&bmsafemap->sm_allocindirhd);
1259 LIST_INIT(&bmsafemap->sm_inodedephd);
1260 LIST_INIT(&bmsafemap->sm_newblkhd);
1261 ACQUIRE_LOCK(&lk);
1262 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1263 return (bmsafemap);
1264}
1265
1266/*
1267 * Direct block allocation dependencies.
1268 *
1269 * When a new block is allocated, the corresponding disk locations must be
1270 * initialized (with zeros or new data) before the on-disk inode points to
1271 * them. Also, the freemap from which the block was allocated must be
1272 * updated (on disk) before the inode's pointer. These two dependencies are
1273 * independent of each other and are needed for all file blocks and indirect
1274 * blocks that are pointed to directly by the inode. Just before the
1275 * "in-core" version of the inode is updated with a newly allocated block
1276 * number, a procedure (below) is called to setup allocation dependency
1277 * structures. These structures are removed when the corresponding
1278 * dependencies are satisfied or when the block allocation becomes obsolete
1279 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1280 * fragment that gets upgraded). All of these cases are handled in
1281 * procedures described later.
1282 *
1283 * When a file extension causes a fragment to be upgraded, either to a larger
1284 * fragment or to a full block, the on-disk location may change (if the
1285 * previous fragment could not simply be extended). In this case, the old
1286 * fragment must be de-allocated, but not until after the inode's pointer has
1287 * been updated. In most cases, this is handled by later procedures, which
1288 * will construct a "freefrag" structure to be added to the workitem queue
1289 * when the inode update is complete (or obsolete). The main exception to
1290 * this is when an allocation occurs while a pending allocation dependency
1291 * (for the same block pointer) remains. This case is handled in the main
1292 * allocation dependency setup procedure by immediately freeing the
1293 * unreferenced fragments.
1294 */
1295void
1296softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1297 struct inode *ip; /* inode to which block is being added */
1298 ufs_lbn_t lbn; /* block pointer within inode */
1299 ufs_daddr_t newblkno; /* disk block number being added */
1300 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */
1301 long newsize; /* size of new block */
1302 long oldsize; /* size of new block */
1303 struct buf *bp; /* bp for allocated block */
1304{
1305 struct allocdirect *adp, *oldadp;
1306 struct allocdirectlst *adphead;
1307 struct bmsafemap *bmsafemap;
1308 struct inodedep *inodedep;
1309 struct pagedep *pagedep;
1310 struct newblk *newblk;
1311
1312 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1313 M_ALLOCDIRECT, M_SOFTDEP_FLAGS);
1314 bzero(adp, sizeof(struct allocdirect));
1315 adp->ad_list.wk_type = D_ALLOCDIRECT;
1316 adp->ad_lbn = lbn;
1317 adp->ad_newblkno = newblkno;
1318 adp->ad_oldblkno = oldblkno;
1319 adp->ad_newsize = newsize;
1320 adp->ad_oldsize = oldsize;
1321 adp->ad_state = ATTACHED;
1322 if (newblkno == oldblkno)
1323 adp->ad_freefrag = NULL;
1324 else
1325 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1326
1327 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1328 panic("softdep_setup_allocdirect: lost block");
1329
1330 ACQUIRE_LOCK(&lk);
1331 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1332 adp->ad_inodedep = inodedep;
1333
1334 if (newblk->nb_state == DEPCOMPLETE) {
1335 adp->ad_state |= DEPCOMPLETE;
1336 adp->ad_buf = NULL;
1337 } else {
1338 bmsafemap = newblk->nb_bmsafemap;
1339 adp->ad_buf = bmsafemap->sm_buf;
1340 LIST_REMOVE(newblk, nb_deps);
1341 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1342 }
1343 LIST_REMOVE(newblk, nb_hash);
1344 FREE(newblk, M_NEWBLK);
1345
1346 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1347 if (lbn >= NDADDR) {
1348 /* allocating an indirect block */
1349 if (oldblkno != 0) {
1350 FREE_LOCK(&lk);
1351 panic("softdep_setup_allocdirect: non-zero indir");
1352 }
1353 } else {
1354 /*
1355 * Allocating a direct block.
1356 *
1357 * If we are allocating a directory block, then we must
1358 * allocate an associated pagedep to track additions and
1359 * deletions.
1360 */
1361 if ((ip->i_mode & IFMT) == IFDIR &&
1362 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1363 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1364 }
1365 /*
1366 * The list of allocdirects must be kept in sorted and ascending
1367 * order so that the rollback routines can quickly determine the
1368 * first uncommitted block (the size of the file stored on disk
1369 * ends at the end of the lowest committed fragment, or if there
1370 * are no fragments, at the end of the highest committed block).
1371 * Since files generally grow, the typical case is that the new
1372 * block is to be added at the end of the list. We speed this
1373 * special case by checking against the last allocdirect in the
1374 * list before laboriously traversing the list looking for the
1375 * insertion point.
1376 */
1377 adphead = &inodedep->id_newinoupdt;
1378 oldadp = TAILQ_LAST(adphead, allocdirectlst);
1379 if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1380 /* insert at end of list */
1381 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1382 if (oldadp != NULL && oldadp->ad_lbn == lbn)
1383 allocdirect_merge(adphead, adp, oldadp);
1384 FREE_LOCK(&lk);
1385 return;
1386 }
1387 TAILQ_FOREACH(oldadp, adphead, ad_next) {
1388 if (oldadp->ad_lbn >= lbn)
1389 break;
1390 }
1391 if (oldadp == NULL) {
1392 FREE_LOCK(&lk);
1393 panic("softdep_setup_allocdirect: lost entry");
1394 }
1395 /* insert in middle of list */
1396 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1397 if (oldadp->ad_lbn == lbn)
1398 allocdirect_merge(adphead, adp, oldadp);
1399 FREE_LOCK(&lk);
1400}
1401
1402/*
1403 * Replace an old allocdirect dependency with a newer one.
1404 * This routine must be called with splbio interrupts blocked.
1405 */
1406static void
1407allocdirect_merge(adphead, newadp, oldadp)
1408 struct allocdirectlst *adphead; /* head of list holding allocdirects */
1409 struct allocdirect *newadp; /* allocdirect being added */
1410 struct allocdirect *oldadp; /* existing allocdirect being checked */
1411{
1412 struct freefrag *freefrag;
1413
1414#ifdef DEBUG
1415 if (lk.lkt_held == NOHOLDER)
1416 panic("allocdirect_merge: lock not held");
1417#endif
1418 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1419 newadp->ad_oldsize != oldadp->ad_newsize ||
1420 newadp->ad_lbn >= NDADDR) {
1421 FREE_LOCK(&lk);
1422 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1423 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1424 NDADDR);
1425 }
1426 newadp->ad_oldblkno = oldadp->ad_oldblkno;
1427 newadp->ad_oldsize = oldadp->ad_oldsize;
1428 /*
1429 * If the old dependency had a fragment to free or had never
1430 * previously had a block allocated, then the new dependency
1431 * can immediately post its freefrag and adopt the old freefrag.
1432 * This action is done by swapping the freefrag dependencies.
1433 * The new dependency gains the old one's freefrag, and the
1434 * old one gets the new one and then immediately puts it on
1435 * the worklist when it is freed by free_allocdirect. It is
1436 * not possible to do this swap when the old dependency had a
1437 * non-zero size but no previous fragment to free. This condition
1438 * arises when the new block is an extension of the old block.
1439 * Here, the first part of the fragment allocated to the new
1440 * dependency is part of the block currently claimed on disk by
1441 * the old dependency, so cannot legitimately be freed until the
1442 * conditions for the new dependency are fulfilled.
1443 */
1444 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1445 freefrag = newadp->ad_freefrag;
1446 newadp->ad_freefrag = oldadp->ad_freefrag;
1447 oldadp->ad_freefrag = freefrag;
1448 }
1449 free_allocdirect(adphead, oldadp, 0);
1450}
1451
1452/*
1453 * Allocate a new freefrag structure if needed.
1454 */
1455static struct freefrag *
1456newfreefrag(ip, blkno, size)
1457 struct inode *ip;
1458 ufs_daddr_t blkno;
1459 long size;
1460{
1461 struct freefrag *freefrag;
1462 struct fs *fs;
1463
1464 if (blkno == 0)
1465 return (NULL);
1466 fs = ip->i_fs;
1467 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1468 panic("newfreefrag: frag size");
1469 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1470 M_FREEFRAG, M_SOFTDEP_FLAGS);
1471 freefrag->ff_list.wk_type = D_FREEFRAG;
1472 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */
1473 freefrag->ff_inum = ip->i_number;
1474 freefrag->ff_fs = fs;
1475 freefrag->ff_devvp = ip->i_devvp;
1476 freefrag->ff_blkno = blkno;
1477 freefrag->ff_fragsize = size;
1478 return (freefrag);
1479}
1480
1481/*
1482 * This workitem de-allocates fragments that were replaced during
1483 * file block allocation.
1484 */
1485static void
1486handle_workitem_freefrag(freefrag)
1487 struct freefrag *freefrag;
1488{
1489 struct inode tip;
1490
1491 tip.i_fs = freefrag->ff_fs;
1492 tip.i_devvp = freefrag->ff_devvp;
1493 tip.i_dev = freefrag->ff_devvp->v_rdev;
1494 tip.i_number = freefrag->ff_inum;
1495 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */
1496 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1497 FREE(freefrag, M_FREEFRAG);
1498}
1499
1500/*
1501 * Indirect block allocation dependencies.
1502 *
1503 * The same dependencies that exist for a direct block also exist when
1504 * a new block is allocated and pointed to by an entry in a block of
1505 * indirect pointers. The undo/redo states described above are also
1506 * used here. Because an indirect block contains many pointers that
1507 * may have dependencies, a second copy of the entire in-memory indirect
1508 * block is kept. The buffer cache copy is always completely up-to-date.
1509 * The second copy, which is used only as a source for disk writes,
1510 * contains only the safe pointers (i.e., those that have no remaining
1511 * update dependencies). The second copy is freed when all pointers
1512 * are safe. The cache is not allowed to replace indirect blocks with
1513 * pending update dependencies. If a buffer containing an indirect
1514 * block with dependencies is written, these routines will mark it
1515 * dirty again. It can only be successfully written once all the
1516 * dependencies are removed. The ffs_fsync routine in conjunction with
1517 * softdep_sync_metadata work together to get all the dependencies
1518 * removed so that a file can be successfully written to disk. Three
1519 * procedures are used when setting up indirect block pointer
1520 * dependencies. The division is necessary because of the organization
1521 * of the "balloc" routine and because of the distinction between file
1522 * pages and file metadata blocks.
1523 */
1524
1525/*
1526 * Allocate a new allocindir structure.
1527 */
1528static struct allocindir *
1529newallocindir(ip, ptrno, newblkno, oldblkno)
1530 struct inode *ip; /* inode for file being extended */
1531 int ptrno; /* offset of pointer in indirect block */
1532 ufs_daddr_t newblkno; /* disk block number being added */
1533 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1534{
1535 struct allocindir *aip;
1536
1537 MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1538 M_ALLOCINDIR, M_SOFTDEP_FLAGS);
1539 bzero(aip, sizeof(struct allocindir));
1540 aip->ai_list.wk_type = D_ALLOCINDIR;
1541 aip->ai_state = ATTACHED;
1542 aip->ai_offset = ptrno;
1543 aip->ai_newblkno = newblkno;
1544 aip->ai_oldblkno = oldblkno;
1545 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1546 return (aip);
1547}
1548
1549/*
1550 * Called just before setting an indirect block pointer
1551 * to a newly allocated file page.
1552 */
1553void
1554softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1555 struct inode *ip; /* inode for file being extended */
1556 ufs_lbn_t lbn; /* allocated block number within file */
1557 struct buf *bp; /* buffer with indirect blk referencing page */
1558 int ptrno; /* offset of pointer in indirect block */
1559 ufs_daddr_t newblkno; /* disk block number being added */
1560 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1561 struct buf *nbp; /* buffer holding allocated page */
1562{
1563 struct allocindir *aip;
1564 struct pagedep *pagedep;
1565
1566 aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1567 ACQUIRE_LOCK(&lk);
1568 /*
1569 * If we are allocating a directory page, then we must
1570 * allocate an associated pagedep to track additions and
1571 * deletions.
1572 */
1573 if ((ip->i_mode & IFMT) == IFDIR &&
1574 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1575 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1576 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1577 FREE_LOCK(&lk);
1578 setup_allocindir_phase2(bp, ip, aip);
1579}
1580
1581/*
1582 * Called just before setting an indirect block pointer to a
1583 * newly allocated indirect block.
1584 */
1585void
1586softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1587 struct buf *nbp; /* newly allocated indirect block */
1588 struct inode *ip; /* inode for file being extended */
1589 struct buf *bp; /* indirect block referencing allocated block */
1590 int ptrno; /* offset of pointer in indirect block */
1591 ufs_daddr_t newblkno; /* disk block number being added */
1592{
1593 struct allocindir *aip;
1594
1595 aip = newallocindir(ip, ptrno, newblkno, 0);
1596 ACQUIRE_LOCK(&lk);
1597 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1598 FREE_LOCK(&lk);
1599 setup_allocindir_phase2(bp, ip, aip);
1600}
1601
1602/*
1603 * Called to finish the allocation of the "aip" allocated
1604 * by one of the two routines above.
1605 */
1606static void
1607setup_allocindir_phase2(bp, ip, aip)
1608 struct buf *bp; /* in-memory copy of the indirect block */
1609 struct inode *ip; /* inode for file being extended */
1610 struct allocindir *aip; /* allocindir allocated by the above routines */
1611{
1612 struct worklist *wk;
1613 struct indirdep *indirdep, *newindirdep;
1614 struct bmsafemap *bmsafemap;
1615 struct allocindir *oldaip;
1616 struct freefrag *freefrag;
1617 struct newblk *newblk;
1618
1619 if (bp->b_loffset >= 0)
1620 panic("setup_allocindir_phase2: not indir blk");
1621 for (indirdep = NULL, newindirdep = NULL; ; ) {
1622 ACQUIRE_LOCK(&lk);
1623 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1624 if (wk->wk_type != D_INDIRDEP)
1625 continue;
1626 indirdep = WK_INDIRDEP(wk);
1627 break;
1628 }
1629 if (indirdep == NULL && newindirdep) {
1630 indirdep = newindirdep;
1631 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1632 newindirdep = NULL;
1633 }
1634 FREE_LOCK(&lk);
1635 if (indirdep) {
1636 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1637 &newblk) == 0)
1638 panic("setup_allocindir: lost block");
1639 ACQUIRE_LOCK(&lk);
1640 if (newblk->nb_state == DEPCOMPLETE) {
1641 aip->ai_state |= DEPCOMPLETE;
1642 aip->ai_buf = NULL;
1643 } else {
1644 bmsafemap = newblk->nb_bmsafemap;
1645 aip->ai_buf = bmsafemap->sm_buf;
1646 LIST_REMOVE(newblk, nb_deps);
1647 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1648 aip, ai_deps);
1649 }
1650 LIST_REMOVE(newblk, nb_hash);
1651 FREE(newblk, M_NEWBLK);
1652 aip->ai_indirdep = indirdep;
1653 /*
1654 * Check to see if there is an existing dependency
1655 * for this block. If there is, merge the old
1656 * dependency into the new one.
1657 */
1658 if (aip->ai_oldblkno == 0)
1659 oldaip = NULL;
1660 else
1661
1662 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1663 if (oldaip->ai_offset == aip->ai_offset)
1664 break;
1665 if (oldaip != NULL) {
1666 if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1667 FREE_LOCK(&lk);
1668 panic("setup_allocindir_phase2: blkno");
1669 }
1670 aip->ai_oldblkno = oldaip->ai_oldblkno;
1671 freefrag = oldaip->ai_freefrag;
1672 oldaip->ai_freefrag = aip->ai_freefrag;
1673 aip->ai_freefrag = freefrag;
1674 free_allocindir(oldaip, NULL);
1675 }
1676 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1677 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1678 [aip->ai_offset] = aip->ai_oldblkno;
1679 FREE_LOCK(&lk);
1680 }
1681 if (newindirdep) {
1682 /*
1683 * Avoid any possibility of data corruption by
1684 * ensuring that our old version is thrown away.
1685 */
1686 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
1687 brelse(newindirdep->ir_savebp);
1688 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1689 }
1690 if (indirdep)
1691 break;
1692 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1693 M_INDIRDEP, M_SOFTDEP_FLAGS);
1694 newindirdep->ir_list.wk_type = D_INDIRDEP;
1695 newindirdep->ir_state = ATTACHED;
1696 LIST_INIT(&newindirdep->ir_deplisthd);
1697 LIST_INIT(&newindirdep->ir_donehd);
1698 if (bp->b_bio2.bio_offset == NOOFFSET) {
1699 VOP_BMAP(bp->b_vp, bp->b_bio1.bio_offset,
1700 NULL, &bp->b_bio2.bio_offset,
1701 NULL, NULL);
1702 }
1703 KKASSERT(bp->b_bio2.bio_offset != NOOFFSET);
1704 newindirdep->ir_savebp = getblk(ip->i_devvp,
1705 bp->b_bio2.bio_offset,
1706 bp->b_bcount, 0, 0);
1707 BUF_KERNPROC(newindirdep->ir_savebp);
1708 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1709 }
1710}
1711
1712/*
1713 * Block de-allocation dependencies.
1714 *
1715 * When blocks are de-allocated, the on-disk pointers must be nullified before
1716 * the blocks are made available for use by other files. (The true
1717 * requirement is that old pointers must be nullified before new on-disk
1718 * pointers are set. We chose this slightly more stringent requirement to
1719 * reduce complexity.) Our implementation handles this dependency by updating
1720 * the inode (or indirect block) appropriately but delaying the actual block
1721 * de-allocation (i.e., freemap and free space count manipulation) until
1722 * after the updated versions reach stable storage. After the disk is
1723 * updated, the blocks can be safely de-allocated whenever it is convenient.
1724 * This implementation handles only the common case of reducing a file's
1725 * length to zero. Other cases are handled by the conventional synchronous
1726 * write approach.
1727 *
1728 * The ffs implementation with which we worked double-checks
1729 * the state of the block pointers and file size as it reduces
1730 * a file's length. Some of this code is replicated here in our
1731 * soft updates implementation. The freeblks->fb_chkcnt field is
1732 * used to transfer a part of this information to the procedure
1733 * that eventually de-allocates the blocks.
1734 *
1735 * This routine should be called from the routine that shortens
1736 * a file's length, before the inode's size or block pointers
1737 * are modified. It will save the block pointer information for
1738 * later release and zero the inode so that the calling routine
1739 * can release it.
1740 */
1741struct softdep_setup_freeblocks_info {
1742 struct fs *fs;
1743 struct inode *ip;
1744};
1745
1746static int softdep_setup_freeblocks_bp(struct buf *bp, void *data);
1747
1748void
1749softdep_setup_freeblocks(ip, length)
1750 struct inode *ip; /* The inode whose length is to be reduced */
1751 off_t length; /* The new length for the file */
1752{
1753 struct softdep_setup_freeblocks_info info;
1754 struct freeblks *freeblks;
1755 struct inodedep *inodedep;
1756 struct allocdirect *adp;
1757 struct vnode *vp;
1758 struct buf *bp;
1759 struct fs *fs;
1760 int i, error, delay;
1761 int count;
1762
1763 fs = ip->i_fs;
1764 if (length != 0)
1765 panic("softde_setup_freeblocks: non-zero length");
1766 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1767 M_FREEBLKS, M_SOFTDEP_FLAGS);
1768 bzero(freeblks, sizeof(struct freeblks));
1769 freeblks->fb_list.wk_type = D_FREEBLKS;
1770 freeblks->fb_state = ATTACHED;
1771 freeblks->fb_uid = ip->i_uid;
1772 freeblks->fb_previousinum = ip->i_number;
1773 freeblks->fb_devvp = ip->i_devvp;
1774 freeblks->fb_fs = fs;
1775 freeblks->fb_oldsize = ip->i_size;
1776 freeblks->fb_newsize = length;
1777 freeblks->fb_chkcnt = ip->i_blocks;
1778 for (i = 0; i < NDADDR; i++) {
1779 freeblks->fb_dblks[i] = ip->i_db[i];
1780 ip->i_db[i] = 0;
1781 }
1782 for (i = 0; i < NIADDR; i++) {
1783 freeblks->fb_iblks[i] = ip->i_ib[i];
1784 ip->i_ib[i] = 0;
1785 }
1786 ip->i_blocks = 0;
1787 ip->i_size = 0;
1788 /*
1789 * Push the zero'ed inode to to its disk buffer so that we are free
1790 * to delete its dependencies below. Once the dependencies are gone
1791 * the buffer can be safely released.
1792 */
1793 if ((error = bread(ip->i_devvp,
1794 fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)),
1795 (int)fs->fs_bsize, &bp)) != 0)
1796 softdep_error("softdep_setup_freeblocks", error);
1797 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
1798 ip->i_din;
1799 /*
1800 * Find and eliminate any inode dependencies.
1801 */
1802 ACQUIRE_LOCK(&lk);
1803 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1804 if ((inodedep->id_state & IOSTARTED) != 0) {
1805 FREE_LOCK(&lk);
1806 panic("softdep_setup_freeblocks: inode busy");
1807 }
1808 /*
1809 * Add the freeblks structure to the list of operations that
1810 * must await the zero'ed inode being written to disk. If we
1811 * still have a bitmap dependency (delay == 0), then the inode
1812 * has never been written to disk, so we can process the
1813 * freeblks below once we have deleted the dependencies.
1814 */
1815 delay = (inodedep->id_state & DEPCOMPLETE);
1816 if (delay)
1817 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1818 /*
1819 * Because the file length has been truncated to zero, any
1820 * pending block allocation dependency structures associated
1821 * with this inode are obsolete and can simply be de-allocated.
1822 * We must first merge the two dependency lists to get rid of
1823 * any duplicate freefrag structures, then purge the merged list.
1824 */
1825 merge_inode_lists(inodedep);
1826 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1827 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
1828 FREE_LOCK(&lk);
1829 bdwrite(bp);
1830 /*
1831 * We must wait for any I/O in progress to finish so that
1832 * all potential buffers on the dirty list will be visible.
1833 * Once they are all there, walk the list and get rid of
1834 * any dependencies.
1835 */
1836 vp = ITOV(ip);
1837 ACQUIRE_LOCK(&lk);
1838 drain_output(vp, 1);
1839
1840 info.fs = fs;
1841 info.ip = ip;
1842 crit_enter();
1843 do {
1844 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
1845 softdep_setup_freeblocks_bp, &info);
1846 } while (count != 0);
1847 crit_exit();
1848 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
1849 (void)free_inodedep(inodedep);
1850
1851 if (delay) {
1852 freeblks->fb_state |= DEPCOMPLETE;
1853 /*
1854 * If the inode with zeroed block pointers is now on disk
1855 * we can start freeing blocks. Add freeblks to the worklist
1856 * instead of calling handle_workitem_freeblocks directly as
1857 * it is more likely that additional IO is needed to complete
1858 * the request here than in the !delay case.
1859 */
1860 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
1861 add_to_worklist(&freeblks->fb_list);
1862 }
1863
1864 FREE_LOCK(&lk);
1865 /*
1866 * If the inode has never been written to disk (delay == 0),
1867 * then we can process the freeblks now that we have deleted
1868 * the dependencies.
1869 */
1870 if (!delay)
1871 handle_workitem_freeblocks(freeblks);
1872}
1873
1874static int
1875softdep_setup_freeblocks_bp(struct buf *bp, void *data)
1876{
1877 struct softdep_setup_freeblocks_info *info = data;
1878 struct inodedep *inodedep;
1879
1880 if (getdirtybuf(&bp, MNT_WAIT) == 0) {
1881 printf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp);
1882 return(-1);
1883 }
1884 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) {
1885 printf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp);
1886 BUF_UNLOCK(bp);
1887 return(-1);
1888 }
1889 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep);
1890 deallocate_dependencies(bp, inodedep);
1891 bp->b_flags |= B_INVAL | B_NOCACHE;
1892 FREE_LOCK(&lk);
1893 brelse(bp);
1894 ACQUIRE_LOCK(&lk);
1895 return(1);
1896}
1897
1898/*
1899 * Reclaim any dependency structures from a buffer that is about to
1900 * be reallocated to a new vnode. The buffer must be locked, thus,
1901 * no I/O completion operations can occur while we are manipulating
1902 * its associated dependencies. The mutex is held so that other I/O's
1903 * associated with related dependencies do not occur.
1904 */
1905static void
1906deallocate_dependencies(bp, inodedep)
1907 struct buf *bp;
1908 struct inodedep *inodedep;
1909{
1910 struct worklist *wk;
1911 struct indirdep *indirdep;
1912 struct allocindir *aip;
1913 struct pagedep *pagedep;
1914 struct dirrem *dirrem;
1915 struct diradd *dap;
1916 int i;
1917
1918 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1919 switch (wk->wk_type) {
1920
1921 case D_INDIRDEP:
1922 indirdep = WK_INDIRDEP(wk);
1923 /*
1924 * None of the indirect pointers will ever be visible,
1925 * so they can simply be tossed. GOINGAWAY ensures
1926 * that allocated pointers will be saved in the buffer
1927 * cache until they are freed. Note that they will
1928 * only be able to be found by their physical address
1929 * since the inode mapping the logical address will
1930 * be gone. The save buffer used for the safe copy
1931 * was allocated in setup_allocindir_phase2 using
1932 * the physical address so it could be used for this
1933 * purpose. Hence we swap the safe copy with the real
1934 * copy, allowing the safe copy to be freed and holding
1935 * on to the real copy for later use in indir_trunc.
1936 *
1937 * NOTE: ir_savebp is relative to the block device
1938 * so b_bio1 contains the device block number.
1939 */
1940 if (indirdep->ir_state & GOINGAWAY) {
1941 FREE_LOCK(&lk);
1942 panic("deallocate_dependencies: already gone");
1943 }
1944 indirdep->ir_state |= GOINGAWAY;
1945 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
1946 free_allocindir(aip, inodedep);
1947 if (bp->b_bio1.bio_offset >= 0 ||
1948 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) {
1949 FREE_LOCK(&lk);
1950 panic("deallocate_dependencies: not indir");
1951 }
1952 bcopy(bp->b_data, indirdep->ir_savebp->b_data,
1953 bp->b_bcount);
1954 WORKLIST_REMOVE(wk);
1955 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
1956 continue;
1957
1958 case D_PAGEDEP:
1959 pagedep = WK_PAGEDEP(wk);
1960 /*
1961 * None of the directory additions will ever be
1962 * visible, so they can simply be tossed.
1963 */
1964 for (i = 0; i < DAHASHSZ; i++)
1965 while ((dap =
1966 LIST_FIRST(&pagedep->pd_diraddhd[i])))
1967 free_diradd(dap);
1968 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
1969 free_diradd(dap);
1970 /*
1971 * Copy any directory remove dependencies to the list
1972 * to be processed after the zero'ed inode is written.
1973 * If the inode has already been written, then they
1974 * can be dumped directly onto the work list.
1975 */
1976 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
1977 LIST_REMOVE(dirrem, dm_next);
1978 dirrem->dm_dirinum = pagedep->pd_ino;
1979 if (inodedep == NULL ||
1980 (inodedep->id_state & ALLCOMPLETE) ==
1981 ALLCOMPLETE)
1982 add_to_worklist(&dirrem->dm_list);
1983 else
1984 WORKLIST_INSERT(&inodedep->id_bufwait,
1985 &dirrem->dm_list);
1986 }
1987 WORKLIST_REMOVE(&pagedep->pd_list);
1988 LIST_REMOVE(pagedep, pd_hash);
1989 WORKITEM_FREE(pagedep, D_PAGEDEP);
1990 continue;
1991
1992 case D_ALLOCINDIR:
1993 free_allocindir(WK_ALLOCINDIR(wk), inodedep);
1994 continue;
1995
1996 case D_ALLOCDIRECT:
1997 case D_INODEDEP:
1998 FREE_LOCK(&lk);
1999 panic("deallocate_dependencies: Unexpected type %s",
2000 TYPENAME(wk->wk_type));
2001 /* NOTREACHED */
2002
2003 default:
2004 FREE_LOCK(&lk);
2005 panic("deallocate_dependencies: Unknown type %s",
2006 TYPENAME(wk->wk_type));
2007 /* NOTREACHED */
2008 }
2009 }
2010}
2011
2012/*
2013 * Free an allocdirect. Generate a new freefrag work request if appropriate.
2014 * This routine must be called with splbio interrupts blocked.
2015 */
2016static void
2017free_allocdirect(adphead, adp, delay)
2018 struct allocdirectlst *adphead;
2019 struct allocdirect *adp;
2020 int delay;
2021{
2022
2023#ifdef DEBUG
2024 if (lk.lkt_held == NOHOLDER)
2025 panic("free_allocdirect: lock not held");
2026#endif
2027 if ((adp->ad_state & DEPCOMPLETE) == 0)
2028 LIST_REMOVE(adp, ad_deps);
2029 TAILQ_REMOVE(adphead, adp, ad_next);
2030 if ((adp->ad_state & COMPLETE) == 0)
2031 WORKLIST_REMOVE(&adp->ad_list);
2032 if (adp->ad_freefrag != NULL) {
2033 if (delay)
2034 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2035 &adp->ad_freefrag->ff_list);
2036 else
2037 add_to_worklist(&adp->ad_freefrag->ff_list);
2038 }
2039 WORKITEM_FREE(adp, D_ALLOCDIRECT);
2040}
2041
2042/*
2043 * Prepare an inode to be freed. The actual free operation is not
2044 * done until the zero'ed inode has been written to disk.
2045 */
2046void
2047softdep_freefile(pvp, ino, mode)
2048 struct vnode *pvp;
2049 ino_t ino;
2050 int mode;
2051{
2052 struct inode *ip = VTOI(pvp);
2053 struct inodedep *inodedep;
2054 struct freefile *freefile;
2055
2056 /*
2057 * This sets up the inode de-allocation dependency.
2058 */
2059 MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2060 M_FREEFILE, M_SOFTDEP_FLAGS);
2061 freefile->fx_list.wk_type = D_FREEFILE;
2062 freefile->fx_list.wk_state = 0;
2063 freefile->fx_mode = mode;
2064 freefile->fx_oldinum = ino;
2065 freefile->fx_devvp = ip->i_devvp;
2066 freefile->fx_fs = ip->i_fs;
2067
2068 /*
2069 * If the inodedep does not exist, then the zero'ed inode has
2070 * been written to disk. If the allocated inode has never been
2071 * written to disk, then the on-disk inode is zero'ed. In either
2072 * case we can free the file immediately.
2073 */
2074 ACQUIRE_LOCK(&lk);
2075 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2076 check_inode_unwritten(inodedep)) {
2077 FREE_LOCK(&lk);
2078 handle_workitem_freefile(freefile);
2079 return;
2080 }
2081 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2082 FREE_LOCK(&lk);
2083}
2084
2085/*
2086 * Check to see if an inode has never been written to disk. If
2087 * so free the inodedep and return success, otherwise return failure.
2088 * This routine must be called with splbio interrupts blocked.
2089 *
2090 * If we still have a bitmap dependency, then the inode has never
2091 * been written to disk. Drop the dependency as it is no longer
2092 * necessary since the inode is being deallocated. We set the
2093 * ALLCOMPLETE flags since the bitmap now properly shows that the
2094 * inode is not allocated. Even if the inode is actively being
2095 * written, it has been rolled back to its zero'ed state, so we
2096 * are ensured that a zero inode is what is on the disk. For short
2097 * lived files, this change will usually result in removing all the
2098 * dependencies from the inode so that it can be freed immediately.
2099 */
2100static int
2101check_inode_unwritten(inodedep)
2102 struct inodedep *inodedep;
2103{
2104
2105 if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2106 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2107 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2108 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2109 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2110 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2111 inodedep->id_nlinkdelta != 0)
2112 return (0);
2113
2114 /*
2115 * Another process might be in initiate_write_inodeblock
2116 * trying to allocate memory without holding "Softdep Lock".
2117 */
2118 if ((inodedep->id_state & IOSTARTED) != 0 &&
2119 inodedep->id_savedino == NULL)
2120 return(0);
2121
2122 inodedep->id_state |= ALLCOMPLETE;
2123 LIST_REMOVE(inodedep, id_deps);
2124 inodedep->id_buf = NULL;
2125 if (inodedep->id_state & ONWORKLIST)
2126 WORKLIST_REMOVE(&inodedep->id_list);
2127 if (inodedep->id_savedino != NULL) {
2128 FREE(inodedep->id_savedino, M_INODEDEP);
2129 inodedep->id_savedino = NULL;
2130 }
2131 if (free_inodedep(inodedep) == 0) {
2132 FREE_LOCK(&lk);
2133 panic("check_inode_unwritten: busy inode");
2134 }
2135 return (1);
2136}
2137
2138/*
2139 * Try to free an inodedep structure. Return 1 if it could be freed.
2140 */
2141static int
2142free_inodedep(inodedep)
2143 struct inodedep *inodedep;
2144{
2145
2146 if ((inodedep->id_state & ONWORKLIST) != 0 ||
2147 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2148 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2149 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2150 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2151 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2152 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2153 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
2154 return (0);
2155 LIST_REMOVE(inodedep, id_hash);
2156 WORKITEM_FREE(inodedep, D_INODEDEP);
2157 num_inodedep -= 1;
2158 return (1);
2159}
2160
2161/*
2162 * This workitem routine performs the block de-allocation.
2163 * The workitem is added to the pending list after the updated
2164 * inode block has been written to disk. As mentioned above,
2165 * checks regarding the number of blocks de-allocated (compared
2166 * to the number of blocks allocated for the file) are also
2167 * performed in this function.
2168 */
2169static void
2170handle_workitem_freeblocks(freeblks)
2171 struct freeblks *freeblks;
2172{
2173 struct inode tip;
2174 ufs_daddr_t bn;
2175 struct fs *fs;
2176 int i, level, bsize;
2177 long nblocks, blocksreleased = 0;
2178 int error, allerror = 0;
2179 ufs_lbn_t baselbns[NIADDR], tmpval;
2180
2181 tip.i_number = freeblks->fb_previousinum;
2182 tip.i_devvp = freeblks->fb_devvp;
2183 tip.i_dev = freeblks->fb_devvp->v_rdev;
2184 tip.i_fs = freeblks->fb_fs;
2185 tip.i_size = freeblks->fb_oldsize;
2186 tip.i_uid = freeblks->fb_uid;
2187 fs = freeblks->fb_fs;
2188 tmpval = 1;
2189 baselbns[0] = NDADDR;
2190 for (i = 1; i < NIADDR; i++) {
2191 tmpval *= NINDIR(fs);
2192 baselbns[i] = baselbns[i - 1] + tmpval;
2193 }
2194 nblocks = btodb(fs->fs_bsize);
2195 blocksreleased = 0;
2196 /*
2197 * Indirect blocks first.
2198 */
2199 for (level = (NIADDR - 1); level >= 0; level--) {
2200 if ((bn = freeblks->fb_iblks[level]) == 0)
2201 continue;
2202 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level,
2203 baselbns[level], &blocksreleased)) == 0)
2204 allerror = error;
2205 ffs_blkfree(&tip, bn, fs->fs_bsize);
2206 blocksreleased += nblocks;
2207 }
2208 /*
2209 * All direct blocks or frags.
2210 */
2211 for (i = (NDADDR - 1); i >= 0; i--) {
2212 if ((bn = freeblks->fb_dblks[i]) == 0)
2213 continue;
2214 bsize = blksize(fs, &tip, i);
2215 ffs_blkfree(&tip, bn, bsize);
2216 blocksreleased += btodb(bsize);
2217 }
2218
2219#ifdef DIAGNOSTIC
2220 if (freeblks->fb_chkcnt != blocksreleased)
2221 printf("handle_workitem_freeblocks: block count\n");
2222 if (allerror)
2223 softdep_error("handle_workitem_freeblks", allerror);
2224#endif /* DIAGNOSTIC */
2225 WORKITEM_FREE(freeblks, D_FREEBLKS);
2226}
2227
2228/*
2229 * Release blocks associated with the inode ip and stored in the indirect
2230 * block at doffset. If level is greater than SINGLE, the block is an
2231 * indirect block and recursive calls to indirtrunc must be used to
2232 * cleanse other indirect blocks.
2233 */
2234static int
2235indir_trunc(ip, doffset, level, lbn, countp)
2236 struct inode *ip;
2237 off_t doffset;
2238 int level;
2239 ufs_lbn_t lbn;
2240 long *countp;
2241{
2242 struct buf *bp;
2243 ufs_daddr_t *bap;
2244 ufs_daddr_t nb;
2245 struct fs *fs;
2246 struct worklist *wk;
2247 struct indirdep *indirdep;
2248 int i, lbnadd, nblocks;
2249 int error, allerror = 0;
2250
2251 fs = ip->i_fs;
2252 lbnadd = 1;
2253 for (i = level; i > 0; i--)
2254 lbnadd *= NINDIR(fs);
2255 /*
2256 * Get buffer of block pointers to be freed. This routine is not
2257 * called until the zero'ed inode has been written, so it is safe
2258 * to free blocks as they are encountered. Because the inode has
2259 * been zero'ed, calls to bmap on these blocks will fail. So, we
2260 * have to use the on-disk address and the block device for the
2261 * filesystem to look them up. If the file was deleted before its
2262 * indirect blocks were all written to disk, the routine that set
2263 * us up (deallocate_dependencies) will have arranged to leave
2264 * a complete copy of the indirect block in memory for our use.
2265 * Otherwise we have to read the blocks in from the disk.
2266 */
2267 ACQUIRE_LOCK(&lk);
2268 if ((bp = findblk(ip->i_devvp, doffset)) != NULL &&
2269 (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2270 /*
2271 * bp must be ir_savebp, which is held locked for our use.
2272 */
2273 if (wk->wk_type != D_INDIRDEP ||
2274 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2275 (indirdep->ir_state & GOINGAWAY) == 0) {
2276 FREE_LOCK(&lk);
2277 panic("indir_trunc: lost indirdep");
2278 }
2279 WORKLIST_REMOVE(wk);
2280 WORKITEM_FREE(indirdep, D_INDIRDEP);
2281 if (LIST_FIRST(&bp->b_dep) != NULL) {
2282 FREE_LOCK(&lk);
2283 panic("indir_trunc: dangling dep");
2284 }
2285 FREE_LOCK(&lk);
2286 } else {
2287 FREE_LOCK(&lk);
2288 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp);
2289 if (error)
2290 return (error);
2291 }
2292 /*
2293 * Recursively free indirect blocks.
2294 */
2295 bap = (ufs_daddr_t *)bp->b_data;
2296 nblocks = btodb(fs->fs_bsize);
2297 for (i = NINDIR(fs) - 1; i >= 0; i--) {
2298 if ((nb = bap[i]) == 0)
2299 continue;
2300 if (level != 0) {
2301 if ((error = indir_trunc(ip, fsbtodoff(fs, nb),
2302 level - 1, lbn + (i * lbnadd), countp)) != 0)
2303 allerror = error;
2304 }
2305 ffs_blkfree(ip, nb, fs->fs_bsize);
2306 *countp += nblocks;
2307 }
2308 bp->b_flags |= B_INVAL | B_NOCACHE;
2309 brelse(bp);
2310 return (allerror);
2311}
2312
2313/*
2314 * Free an allocindir.
2315 * This routine must be called with splbio interrupts blocked.
2316 */
2317static void
2318free_allocindir(aip, inodedep)
2319 struct allocindir *aip;
2320 struct inodedep *inodedep;
2321{
2322 struct freefrag *freefrag;
2323
2324#ifdef DEBUG
2325 if (lk.lkt_held == NOHOLDER)
2326 panic("free_allocindir: lock not held");
2327#endif
2328 if ((aip->ai_state & DEPCOMPLETE) == 0)
2329 LIST_REMOVE(aip, ai_deps);
2330 if (aip->ai_state & ONWORKLIST)
2331 WORKLIST_REMOVE(&aip->ai_list);
2332 LIST_REMOVE(aip, ai_next);
2333 if ((freefrag = aip->ai_freefrag) != NULL) {
2334 if (inodedep == NULL)
2335 add_to_worklist(&freefrag->ff_list);
2336 else
2337 WORKLIST_INSERT(&inodedep->id_bufwait,
2338 &freefrag->ff_list);
2339 }
2340 WORKITEM_FREE(aip, D_ALLOCINDIR);
2341}
2342
2343/*
2344 * Directory entry addition dependencies.
2345 *
2346 * When adding a new directory entry, the inode (with its incremented link
2347 * count) must be written to disk before the directory entry's pointer to it.
2348 * Also, if the inode is newly allocated, the corresponding freemap must be
2349 * updated (on disk) before the directory entry's pointer. These requirements
2350 * are met via undo/redo on the directory entry's pointer, which consists
2351 * simply of the inode number.
2352 *
2353 * As directory entries are added and deleted, the free space within a
2354 * directory block can become fragmented. The ufs filesystem will compact
2355 * a fragmented directory block to make space for a new entry. When this
2356 * occurs, the offsets of previously added entries change. Any "diradd"
2357 * dependency structures corresponding to these entries must be updated with
2358 * the new offsets.
2359 */
2360
2361/*
2362 * This routine is called after the in-memory inode's link
2363 * count has been incremented, but before the directory entry's
2364 * pointer to the inode has been set.
2365 */
2366void
2367softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp)
2368 struct buf *bp; /* buffer containing directory block */
2369 struct inode *dp; /* inode for directory */
2370 off_t diroffset; /* offset of new entry in directory */
2371 ino_t newinum; /* inode referenced by new directory entry */
2372 struct buf *newdirbp; /* non-NULL => contents of new mkdir */
2373{
2374 int offset; /* offset of new entry within directory block */
2375 ufs_lbn_t lbn; /* block in directory containing new entry */
2376 struct fs *fs;
2377 struct diradd *dap;
2378 struct pagedep *pagedep;
2379 struct inodedep *inodedep;
2380 struct mkdir *mkdir1, *mkdir2;
2381
2382 /*
2383 * Whiteouts have no dependencies.
2384 */
2385 if (newinum == WINO) {
2386 if (newdirbp != NULL)
2387 bdwrite(newdirbp);
2388 return;
2389 }
2390
2391 fs = dp->i_fs;
2392 lbn = lblkno(fs, diroffset);
2393 offset = blkoff(fs, diroffset);
2394 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2395 M_SOFTDEP_FLAGS);
2396 bzero(dap, sizeof(struct diradd));
2397 dap->da_list.wk_type = D_DIRADD;
2398 dap->da_offset = offset;
2399 dap->da_newinum = newinum;
2400 dap->da_state = ATTACHED;
2401 if (newdirbp == NULL) {
2402 dap->da_state |= DEPCOMPLETE;
2403 ACQUIRE_LOCK(&lk);
2404 } else {
2405 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2406 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2407 M_SOFTDEP_FLAGS);
2408 mkdir1->md_list.wk_type = D_MKDIR;
2409 mkdir1->md_state = MKDIR_BODY;
2410 mkdir1->md_diradd = dap;
2411 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2412 M_SOFTDEP_FLAGS);
2413 mkdir2->md_list.wk_type = D_MKDIR;
2414 mkdir2->md_state = MKDIR_PARENT;
2415 mkdir2->md_diradd = dap;
2416 /*
2417 * Dependency on "." and ".." being written to disk.
2418 */
2419 mkdir1->md_buf = newdirbp;
2420 ACQUIRE_LOCK(&lk);
2421 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2422 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2423 FREE_LOCK(&lk);
2424 bdwrite(newdirbp);
2425 /*
2426 * Dependency on link count increase for parent directory
2427 */
2428 ACQUIRE_LOCK(&lk);
2429 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2430 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2431 dap->da_state &= ~MKDIR_PARENT;
2432 WORKITEM_FREE(mkdir2, D_MKDIR);
2433 } else {
2434 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2435 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2436 }
2437 }
2438 /*
2439 * Link into parent directory pagedep to await its being written.
2440 */
2441 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2442 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2443 dap->da_pagedep = pagedep;
2444 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2445 da_pdlist);
2446 /*
2447 * Link into its inodedep. Put it on the id_bufwait list if the inode
2448 * is not yet written. If it is written, do the post-inode write
2449 * processing to put it on the id_pendinghd list.
2450 */
2451 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2452 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2453 diradd_inode_written(dap, inodedep);
2454 else
2455 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2456 FREE_LOCK(&lk);
2457}
2458
2459/*
2460 * This procedure is called to change the offset of a directory
2461 * entry when compacting a directory block which must be owned
2462 * exclusively by the caller. Note that the actual entry movement
2463 * must be done in this procedure to ensure that no I/O completions
2464 * occur while the move is in progress.
2465 */
2466void
2467softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2468 struct inode *dp; /* inode for directory */
2469 caddr_t base; /* address of dp->i_offset */
2470 caddr_t oldloc; /* address of old directory location */
2471 caddr_t newloc; /* address of new directory location */
2472 int entrysize; /* size of directory entry */
2473{
2474 int offset, oldoffset, newoffset;
2475 struct pagedep *pagedep;
2476 struct diradd *dap;
2477 ufs_lbn_t lbn;
2478
2479 ACQUIRE_LOCK(&lk);
2480 lbn = lblkno(dp->i_fs, dp->i_offset);
2481 offset = blkoff(dp->i_fs, dp->i_offset);
2482 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2483 goto done;
2484 oldoffset = offset + (oldloc - base);
2485 newoffset = offset + (newloc - base);
2486
2487 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2488 if (dap->da_offset != oldoffset)
2489 continue;
2490 dap->da_offset = newoffset;
2491 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2492 break;
2493 LIST_REMOVE(dap, da_pdlist);
2494 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2495 dap, da_pdlist);
2496 break;
2497 }
2498 if (dap == NULL) {
2499
2500 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2501 if (dap->da_offset == oldoffset) {
2502 dap->da_offset = newoffset;
2503 break;
2504 }
2505 }
2506 }
2507done:
2508 bcopy(oldloc, newloc, entrysize);
2509 FREE_LOCK(&lk);
2510}
2511
2512/*
2513 * Free a diradd dependency structure. This routine must be called
2514 * with splbio interrupts blocked.
2515 */
2516static void
2517free_diradd(dap)
2518 struct diradd *dap;
2519{
2520 struct dirrem *dirrem;
2521 struct pagedep *pagedep;
2522 struct inodedep *inodedep;
2523 struct mkdir *mkdir, *nextmd;
2524
2525#ifdef DEBUG
2526 if (lk.lkt_held == NOHOLDER)
2527 panic("free_diradd: lock not held");
2528#endif
2529 WORKLIST_REMOVE(&dap->da_list);
2530 LIST_REMOVE(dap, da_pdlist);
2531 if ((dap->da_state & DIRCHG) == 0) {
2532 pagedep = dap->da_pagedep;
2533 } else {
2534 dirrem = dap->da_previous;
2535 pagedep = dirrem->dm_pagedep;
2536 dirrem->dm_dirinum = pagedep->pd_ino;
2537 add_to_worklist(&dirrem->dm_list);
2538 }
2539 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2540 0, &inodedep) != 0)
2541 (void) free_inodedep(inodedep);
2542 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2543 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2544 nextmd = LIST_NEXT(mkdir, md_mkdirs);
2545 if (mkdir->md_diradd != dap)
2546 continue;
2547 dap->da_state &= ~mkdir->md_state;
2548 WORKLIST_REMOVE(&mkdir->md_list);
2549 LIST_REMOVE(mkdir, md_mkdirs);
2550 WORKITEM_FREE(mkdir, D_MKDIR);
2551 }
2552 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2553 FREE_LOCK(&lk);
2554 panic("free_diradd: unfound ref");
2555 }
2556 }
2557 WORKITEM_FREE(dap, D_DIRADD);
2558}
2559
2560/*
2561 * Directory entry removal dependencies.
2562 *
2563 * When removing a directory entry, the entry's inode pointer must be
2564 * zero'ed on disk before the corresponding inode's link count is decremented
2565 * (possibly freeing the inode for re-use). This dependency is handled by
2566 * updating the directory entry but delaying the inode count reduction until
2567 * after the directory block has been written to disk. After this point, the
2568 * inode count can be decremented whenever it is convenient.
2569 */
2570
2571/*
2572 * This routine should be called immediately after removing
2573 * a directory entry. The inode's link count should not be
2574 * decremented by the calling procedure -- the soft updates
2575 * code will do this task when it is safe.
2576 */
2577void
2578softdep_setup_remove(bp, dp, ip, isrmdir)
2579 struct buf *bp; /* buffer containing directory block */
2580 struct inode *dp; /* inode for the directory being modified */
2581 struct inode *ip; /* inode for directory entry being removed */
2582 int isrmdir; /* indicates if doing RMDIR */
2583{
2584 struct dirrem *dirrem, *prevdirrem;
2585
2586 /*
2587 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2588 */
2589 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2590
2591 /*
2592 * If the COMPLETE flag is clear, then there were no active
2593 * entries and we want to roll back to a zeroed entry until
2594 * the new inode is committed to disk. If the COMPLETE flag is
2595 * set then we have deleted an entry that never made it to
2596 * disk. If the entry we deleted resulted from a name change,
2597 * then the old name still resides on disk. We cannot delete
2598 * its inode (returned to us in prevdirrem) until the zeroed
2599 * directory entry gets to disk. The new inode has never been
2600 * referenced on the disk, so can be deleted immediately.
2601 */
2602 if ((dirrem->dm_state & COMPLETE) == 0) {
2603 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2604 dm_next);
2605 FREE_LOCK(&lk);
2606 } else {
2607 if (prevdirrem != NULL)
2608 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2609 prevdirrem, dm_next);
2610 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2611 FREE_LOCK(&lk);
2612 handle_workitem_remove(dirrem);
2613 }
2614}
2615
2616/*
2617 * Allocate a new dirrem if appropriate and return it along with
2618 * its associated pagedep. Called without a lock, returns with lock.
2619 */
2620static long num_dirrem; /* number of dirrem allocated */
2621static struct dirrem *
2622newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2623 struct buf *bp; /* buffer containing directory block */
2624 struct inode *dp; /* inode for the directory being modified */
2625 struct inode *ip; /* inode for directory entry being removed */
2626 int isrmdir; /* indicates if doing RMDIR */
2627 struct dirrem **prevdirremp; /* previously referenced inode, if any */
2628{
2629 int offset;
2630 ufs_lbn_t lbn;
2631 struct diradd *dap;
2632 struct dirrem *dirrem;
2633 struct pagedep *pagedep;
2634
2635 /*
2636 * Whiteouts have no deletion dependencies.
2637 */
2638 if (ip == NULL)
2639 panic("newdirrem: whiteout");
2640 /*
2641 * If we are over our limit, try to improve the situation.
2642 * Limiting the number of dirrem structures will also limit
2643 * the number of freefile and freeblks structures.
2644 */
2645 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2646 (void) request_cleanup(FLUSH_REMOVE, 0);
2647 num_dirrem += 1;
2648 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
2649 M_DIRREM, M_SOFTDEP_FLAGS);
2650 bzero(dirrem, sizeof(struct dirrem));
2651 dirrem->dm_list.wk_type = D_DIRREM;
2652 dirrem->dm_state = isrmdir ? RMDIR : 0;
2653 dirrem->dm_mnt = ITOV(ip)->v_mount;
2654 dirrem->dm_oldinum = ip->i_number;
2655 *prevdirremp = NULL;
2656
2657 ACQUIRE_LOCK(&lk);
2658 lbn = lblkno(dp->i_fs, dp->i_offset);
2659 offset = blkoff(dp->i_fs, dp->i_offset);
2660 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2661 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2662 dirrem->dm_pagedep = pagedep;
2663 /*
2664 * Check for a diradd dependency for the same directory entry.
2665 * If present, then both dependencies become obsolete and can
2666 * be de-allocated. Check for an entry on both the pd_dirraddhd
2667 * list and the pd_pendinghd list.
2668 */
2669
2670 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
2671 if (dap->da_offset == offset)
2672 break;
2673 if (dap == NULL) {
2674
2675 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
2676 if (dap->da_offset == offset)
2677 break;
2678 if (dap == NULL)
2679 return (dirrem);
2680 }
2681 /*
2682 * Must be ATTACHED at this point.
2683 */
2684 if ((dap->da_state & ATTACHED) == 0) {
2685 FREE_LOCK(&lk);
2686 panic("newdirrem: not ATTACHED");
2687 }
2688 if (dap->da_newinum != ip->i_number) {
2689 FREE_LOCK(&lk);
2690 panic("newdirrem: inum %"PRId64" should be %"PRId64,
2691 ip->i_number, dap->da_newinum);
2692 }
2693 /*
2694 * If we are deleting a changed name that never made it to disk,
2695 * then return the dirrem describing the previous inode (which
2696 * represents the inode currently referenced from this entry on disk).
2697 */
2698 if ((dap->da_state & DIRCHG) != 0) {
2699 *prevdirremp = dap->da_previous;
2700 dap->da_state &= ~DIRCHG;
2701 dap->da_pagedep = pagedep;
2702 }
2703 /*
2704 * We are deleting an entry that never made it to disk.
2705 * Mark it COMPLETE so we can delete its inode immediately.
2706 */
2707 dirrem->dm_state |= COMPLETE;
2708 free_diradd(dap);
2709 return (dirrem);
2710}
2711
2712/*
2713 * Directory entry change dependencies.
2714 *
2715 * Changing an existing directory entry requires that an add operation
2716 * be completed first followed by a deletion. The semantics for the addition
2717 * are identical to the description of adding a new entry above except
2718 * that the rollback is to the old inode number rather than zero. Once
2719 * the addition dependency is completed, the removal is done as described
2720 * in the removal routine above.
2721 */
2722
2723/*
2724 * This routine should be called immediately after changing
2725 * a directory entry. The inode's link count should not be
2726 * decremented by the calling procedure -- the soft updates
2727 * code will perform this task when it is safe.
2728 */
2729void
2730softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
2731 struct buf *bp; /* buffer containing directory block */
2732 struct inode *dp; /* inode for the directory being modified */
2733 struct inode *ip; /* inode for directory entry being removed */
2734 ino_t newinum; /* new inode number for changed entry */
2735 int isrmdir; /* indicates if doing RMDIR */
2736{
2737 int offset;
2738 struct diradd *dap = NULL;
2739 struct dirrem *dirrem, *prevdirrem;
2740 struct pagedep *pagedep;
2741 struct inodedep *inodedep;
2742
2743 offset = blkoff(dp->i_fs, dp->i_offset);
2744
2745 /*
2746 * Whiteouts do not need diradd dependencies.
2747 */
2748 if (newinum != WINO) {
2749 MALLOC(dap, struct diradd *, sizeof(struct diradd),
2750 M_DIRADD, M_SOFTDEP_FLAGS);
2751 bzero(dap, sizeof(struct diradd));
2752 dap->da_list.wk_type = D_DIRADD;
2753 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2754 dap->da_offset = offset;
2755 dap->da_newinum = newinum;
2756 }
2757
2758 /*
2759 * Allocate a new dirrem and ACQUIRE_LOCK.
2760 */
2761 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2762 pagedep = dirrem->dm_pagedep;
2763 /*
2764 * The possible values for isrmdir:
2765 * 0 - non-directory file rename
2766 * 1 - directory rename within same directory
2767 * inum - directory rename to new directory of given inode number
2768 * When renaming to a new directory, we are both deleting and
2769 * creating a new directory entry, so the link count on the new
2770 * directory should not change. Thus we do not need the followup
2771 * dirrem which is usually done in handle_workitem_remove. We set
2772 * the DIRCHG flag to tell handle_workitem_remove to skip the
2773 * followup dirrem.
2774 */
2775 if (isrmdir > 1)
2776 dirrem->dm_state |= DIRCHG;
2777
2778 /*
2779 * Whiteouts have no additional dependencies,
2780 * so just put the dirrem on the correct list.
2781 */
2782 if (newinum == WINO) {
2783 if ((dirrem->dm_state & COMPLETE) == 0) {
2784 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2785 dm_next);
2786 } else {
2787 dirrem->dm_dirinum = pagedep->pd_ino;
2788 add_to_worklist(&dirrem->dm_list);
2789 }
2790 FREE_LOCK(&lk);
2791 return;
2792 }
2793
2794 /*
2795 * If the COMPLETE flag is clear, then there were no active
2796 * entries and we want to roll back to the previous inode until
2797 * the new inode is committed to disk. If the COMPLETE flag is
2798 * set, then we have deleted an entry that never made it to disk.
2799 * If the entry we deleted resulted from a name change, then the old
2800 * inode reference still resides on disk. Any rollback that we do
2801 * needs to be to that old inode (returned to us in prevdirrem). If
2802 * the entry we deleted resulted from a create, then there is
2803 * no entry on the disk, so we want to roll back to zero rather
2804 * than the uncommitted inode. In either of the COMPLETE cases we
2805 * want to immediately free the unwritten and unreferenced inode.
2806 */
2807 if ((dirrem->dm_state & COMPLETE) == 0) {
2808 dap->da_previous = dirrem;
2809 } else {
2810 if (prevdirrem != NULL) {
2811 dap->da_previous = prevdirrem;
2812 } else {
2813 dap->da_state &= ~DIRCHG;
2814 dap->da_pagedep = pagedep;
2815 }
2816 dirrem->dm_dirinum = pagedep->pd_ino;
2817 add_to_worklist(&dirrem->dm_list);
2818 }
2819 /*
2820 * Link into its inodedep. Put it on the id_bufwait list if the inode
2821 * is not yet written. If it is written, do the post-inode write
2822 * processing to put it on the id_pendinghd list.
2823 */
2824 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2825 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2826 dap->da_state |= COMPLETE;
2827 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2828 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2829 } else {
2830 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2831 dap, da_pdlist);
2832 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2833 }
2834 FREE_LOCK(&lk);
2835}
2836
2837/*
2838 * Called whenever the link count on an inode is changed.
2839 * It creates an inode dependency so that the new reference(s)
2840 * to the inode cannot be committed to disk until the updated
2841 * inode has been written.
2842 */
2843void
2844softdep_change_linkcnt(ip)
2845 struct inode *ip; /* the inode with the increased link count */
2846{
2847 struct inodedep *inodedep;
2848
2849 ACQUIRE_LOCK(&lk);
2850 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2851 if (ip->i_nlink < ip->i_effnlink) {
2852 FREE_LOCK(&lk);
2853 panic("softdep_change_linkcnt: bad delta");
2854 }
2855 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2856 FREE_LOCK(&lk);
2857}
2858
2859/*
2860 * This workitem decrements the inode's link count.
2861 * If the link count reaches zero, the file is removed.
2862 */
2863static void
2864handle_workitem_remove(dirrem)
2865 struct dirrem *dirrem;
2866{
2867 struct thread *td = curthread; /* XXX */
2868 struct inodedep *inodedep;
2869 struct vnode *vp;
2870 struct inode *ip;
2871 ino_t oldinum;
2872 int error;
2873
2874 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) {
2875 softdep_error("handle_workitem_remove: vget", error);
2876 return;
2877 }
2878 ip = VTOI(vp);
2879 ACQUIRE_LOCK(&lk);
2880 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
2881 FREE_LOCK(&lk);
2882 panic("handle_workitem_remove: lost inodedep");
2883 }
2884 /*
2885 * Normal file deletion.
2886 */
2887 if ((dirrem->dm_state & RMDIR) == 0) {
2888 ip->i_nlink--;
2889 ip->i_flag |= IN_CHANGE;
2890 if (ip->i_nlink < ip->i_effnlink) {
2891 FREE_LOCK(&lk);
2892 panic("handle_workitem_remove: bad file delta");
2893 }
2894 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2895 FREE_LOCK(&lk);
2896 vput(vp);
2897 num_dirrem -= 1;
2898 WORKITEM_FREE(dirrem, D_DIRREM);
2899 return;
2900 }
2901 /*
2902 * Directory deletion. Decrement reference count for both the
2903 * just deleted parent directory entry and the reference for ".".
2904 * Next truncate the directory to length zero. When the
2905 * truncation completes, arrange to have the reference count on
2906 * the parent decremented to account for the loss of "..".
2907 */
2908 ip->i_nlink -= 2;
2909 ip->i_flag |= IN_CHANGE;
2910 if (ip->i_nlink < ip->i_effnlink) {
2911 FREE_LOCK(&lk);
2912 panic("handle_workitem_remove: bad dir delta");
2913 }
2914 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2915 FREE_LOCK(&lk);
2916 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, proc0.p_ucred, td)) != 0)
2917 softdep_error("handle_workitem_remove: truncate", error);
2918 /*
2919 * Rename a directory to a new parent. Since, we are both deleting
2920 * and creating a new directory entry, the link count on the new
2921 * directory should not change. Thus we skip the followup dirrem.
2922 */
2923 if (dirrem->dm_state & DIRCHG) {
2924 vput(vp);
2925 num_dirrem -= 1;
2926 WORKITEM_FREE(dirrem, D_DIRREM);
2927 return;
2928 }
2929 /*
2930 * If the inodedep does not exist, then the zero'ed inode has
2931 * been written to disk. If the allocated inode has never been
2932 * written to disk, then the on-disk inode is zero'ed. In either
2933 * case we can remove the file immediately.
2934 */
2935 ACQUIRE_LOCK(&lk);
2936 dirrem->dm_state = 0;
2937 oldinum = dirrem->dm_oldinum;
2938 dirrem->dm_oldinum = dirrem->dm_dirinum;
2939 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
2940 check_inode_unwritten(inodedep)) {
2941 FREE_LOCK(&lk);
2942 vput(vp);
2943 handle_workitem_remove(dirrem);
2944 return;
2945 }
2946 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
2947 FREE_LOCK(&lk);
2948 ip->i_flag |= IN_CHANGE;
2949 ffs_update(vp, 0);
2950 vput(vp);
2951}
2952
2953/*
2954 * Inode de-allocation dependencies.
2955 *
2956 * When an inode's link count is reduced to zero, it can be de-allocated. We
2957 * found it convenient to postpone de-allocation until after the inode is
2958 * written to disk with its new link count (zero). At this point, all of the
2959 * on-disk inode's block pointers are nullified and, with careful dependency
2960 * list ordering, all dependencies related to the inode will be satisfied and
2961 * the corresponding dependency structures de-allocated. So, if/when the
2962 * inode is reused, there will be no mixing of old dependencies with new
2963 * ones. This artificial dependency is set up by the block de-allocation
2964 * procedure above (softdep_setup_freeblocks) and completed by the
2965 * following procedure.
2966 */
2967static void
2968handle_workitem_freefile(freefile)
2969 struct freefile *freefile;
2970{
2971 struct vnode vp;
2972 struct inode tip;
2973 struct inodedep *idp;
2974 int error;
2975
2976#ifdef DEBUG
2977 ACQUIRE_LOCK(&lk);
2978 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp);
2979 FREE_LOCK(&lk);
2980 if (error)
2981 panic("handle_workitem_freefile: inodedep survived");
2982#endif
2983 tip.i_devvp = freefile->fx_devvp;
2984 tip.i_dev = freefile->fx_devvp->v_rdev;
2985 tip.i_fs = freefile->fx_fs;
2986 vp.v_data = &tip;
2987 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
2988 softdep_error("handle_workitem_freefile", error);
2989 WORKITEM_FREE(freefile, D_FREEFILE);
2990}
2991
2992/*
2993 * Helper function which unlinks marker element from work list and returns
2994 * the next element on the list.
2995 */
2996static __inline struct worklist *
2997markernext(struct worklist *marker)
2998{
2999 struct worklist *next;
3000
3001 next = LIST_NEXT(marker, wk_list);
3002 LIST_REMOVE(marker, wk_list);
3003 return next;
3004}
3005
3006/*
3007 * Disk writes.
3008 *
3009 * The dependency structures constructed above are most actively used when file
3010 * system blocks are written to disk. No constraints are placed on when a
3011 * block can be written, but unsatisfied update dependencies are made safe by
3012 * modifying (or replacing) the source memory for the duration of the disk
3013 * write. When the disk write completes, the memory block is again brought
3014 * up-to-date.
3015 *
3016 * In-core inode structure reclamation.
3017 *
3018 * Because there are a finite number of "in-core" inode structures, they are
3019 * reused regularly. By transferring all inode-related dependencies to the
3020 * in-memory inode block and indexing them separately (via "inodedep"s), we
3021 * can allow "in-core" inode structures to be reused at any time and avoid
3022 * any increase in contention.
3023 *
3024 * Called just before entering the device driver to initiate a new disk I/O.
3025 * The buffer must be locked, thus, no I/O completion operations can occur
3026 * while we are manipulating its associated dependencies.
3027 */
3028static void
3029softdep_disk_io_initiation(bp)
3030 struct buf *bp; /* structure describing disk write to occur */
3031{
3032 struct worklist *wk;
3033 struct worklist marker;
3034 struct indirdep *indirdep;
3035
3036 /*
3037 * We only care about write operations. There should never
3038 * be dependencies for reads.
3039 */
3040 if (bp->b_flags & B_READ)
3041 panic("softdep_disk_io_initiation: read");
3042
3043 marker.wk_type = D_LAST + 1; /* Not a normal workitem */
3044
3045 /*
3046 * Do any necessary pre-I/O processing.
3047 */
3048 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) {
3049 LIST_INSERT_AFTER(wk, &marker, wk_list);
3050
3051 switch (wk->wk_type) {
3052
3053 case D_PAGEDEP:
3054 initiate_write_filepage(WK_PAGEDEP(wk), bp);
3055 continue;
3056
3057 case D_INODEDEP:
3058 initiate_write_inodeblock(WK_INODEDEP(wk), bp);
3059 continue;
3060
3061 case D_INDIRDEP:
3062 indirdep = WK_INDIRDEP(wk);
3063 if (indirdep->ir_state & GOINGAWAY)
3064 panic("disk_io_initiation: indirdep gone");
3065 /*
3066 * If there are no remaining dependencies, this
3067 * will be writing the real pointers, so the
3068 * dependency can be freed.
3069 */
3070 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
3071 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
3072 brelse(indirdep->ir_savebp);
3073 /* inline expand WORKLIST_REMOVE(wk); */
3074 wk->wk_state &= ~ONWORKLIST;
3075 LIST_REMOVE(wk, wk_list);
3076 WORKITEM_FREE(indirdep, D_INDIRDEP);
3077 continue;
3078 }
3079 /*
3080 * Replace up-to-date version with safe version.
3081 */
3082 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3083 M_INDIRDEP, M_SOFTDEP_FLAGS);
3084 ACQUIRE_LOCK(&lk);
3085 indirdep->ir_state &= ~ATTACHED;
3086 indirdep->ir_state |= UNDONE;
3087 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3088 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3089 bp->b_bcount);
3090 FREE_LOCK(&lk);
3091 continue;
3092
3093 case D_MKDIR:
3094 case D_BMSAFEMAP:
3095 case D_ALLOCDIRECT:
3096 case D_ALLOCINDIR:
3097 continue;
3098
3099 default:
3100 panic("handle_disk_io_initiation: Unexpected type %s",
3101 TYPENAME(wk->wk_type));
3102 /* NOTREACHED */
3103 }
3104 }
3105}
3106
3107/*
3108 * Called from within the procedure above to deal with unsatisfied
3109 * allocation dependencies in a directory. The buffer must be locked,
3110 * thus, no I/O completion operations can occur while we are
3111 * manipulating its associated dependencies.
3112 */
3113static void
3114initiate_write_filepage(pagedep, bp)
3115 struct pagedep *pagedep;
3116 struct buf *bp;
3117{
3118 struct diradd *dap;
3119 struct direct *ep;
3120 int i;
3121
3122 if (pagedep->pd_state & IOSTARTED) {
3123 /*
3124 * This can only happen if there is a driver that does not
3125 * understand chaining. Here biodone will reissue the call
3126 * to strategy for the incomplete buffers.
3127 */
3128 printf("initiate_write_filepage: already started\n");
3129 return;
3130 }
3131 pagedep->pd_state |= IOSTARTED;
3132 ACQUIRE_LOCK(&lk);
3133 for (i = 0; i < DAHASHSZ; i++) {
3134 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3135 ep = (struct direct *)
3136 ((char *)bp->b_data + dap->da_offset);
3137 if (ep->d_ino != dap->da_newinum) {
3138 FREE_LOCK(&lk);
3139 panic("%s: dir inum %d != new %"PRId64,
3140 "initiate_write_filepage",
3141 ep->d_ino, dap->da_newinum);
3142 }
3143 if (dap->da_state & DIRCHG)
3144 ep->d_ino = dap->da_previous->dm_oldinum;
3145 else
3146 ep->d_ino = 0;
3147 dap->da_state &= ~ATTACHED;
3148 dap->da_state |= UNDONE;
3149 }
3150 }
3151 FREE_LOCK(&lk);
3152}
3153
3154/*
3155 * Called from within the procedure above to deal with unsatisfied
3156 * allocation dependencies in an inodeblock. The buffer must be
3157 * locked, thus, no I/O completion operations can occur while we
3158 * are manipulating its associated dependencies.
3159 */
3160static void
3161initiate_write_inodeblock(inodedep, bp)
3162 struct inodedep *inodedep;
3163 struct buf *bp; /* The inode block */
3164{
3165 struct allocdirect *adp, *lastadp;
3166 struct dinode *dp;
3167 struct dinode *sip;
3168 struct fs *fs;
3169 ufs_lbn_t prevlbn = 0;
3170 int i, deplist;
3171
3172 if (inodedep->id_state & IOSTARTED)
3173 panic("initiate_write_inodeblock: already started");
3174 inodedep->id_state |= IOSTARTED;
3175 fs = inodedep->id_fs;
3176 dp = (struct dinode *)bp->b_data +
3177 ino_to_fsbo(fs, inodedep->id_ino);
3178 /*
3179 * If the bitmap is not yet written, then the allocated
3180 * inode cannot be written to disk.
3181 */
3182 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3183 if (inodedep->id_savedino != NULL)
3184 panic("initiate_write_inodeblock: already doing I/O");
3185 MALLOC(sip, struct dinode *,
3186 sizeof(struct dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3187 inodedep->id_savedino = sip;
3188 *inodedep->id_savedino = *dp;
3189 bzero((caddr_t)dp, sizeof(struct dinode));
3190 dp->di_gen = inodedep->id_savedino->di_gen;
3191 return;
3192 }
3193 /*
3194 * If no dependencies, then there is nothing to roll back.
3195 */
3196 inodedep->id_savedsize = dp->di_size;
3197 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3198 return;
3199 /*
3200 * Set the dependencies to busy.
3201 */
3202 ACQUIRE_LOCK(&lk);
3203 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3204 adp = TAILQ_NEXT(adp, ad_next)) {
3205#ifdef DIAGNOSTIC
3206 if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3207 FREE_LOCK(&lk);
3208 panic("softdep_write_inodeblock: lbn order");
3209 }
3210 prevlbn = adp->ad_lbn;
3211 if (adp->ad_lbn < NDADDR &&
3212 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3213 FREE_LOCK(&lk);
3214 panic("%s: direct pointer #%ld mismatch %d != %d",
3215 "softdep_write_inodeblock", adp->ad_lbn,
3216 dp->di_db[adp->ad_lbn], adp->ad_newblkno);
3217 }
3218 if (adp->ad_lbn >= NDADDR &&
3219 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3220 FREE_LOCK(&lk);
3221 panic("%s: indirect pointer #%ld mismatch %d != %d",
3222 "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
3223 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
3224 }
3225 deplist |= 1 << adp->ad_lbn;
3226 if ((adp->ad_state & ATTACHED) == 0) {
3227 FREE_LOCK(&lk);
3228 panic("softdep_write_inodeblock: Unknown state 0x%x",
3229 adp->ad_state);
3230 }
3231#endif /* DIAGNOSTIC */
3232 adp->ad_state &= ~ATTACHED;
3233 adp->ad_state |= UNDONE;
3234 }
3235 /*
3236 * The on-disk inode cannot claim to be any larger than the last
3237 * fragment that has been written. Otherwise, the on-disk inode
3238 * might have fragments that were not the last block in the file
3239 * which would corrupt the filesystem.
3240 */
3241 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3242 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3243 if (adp->ad_lbn >= NDADDR)
3244 break;
3245 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3246 /* keep going until hitting a rollback to a frag */
3247 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3248 continue;
3249 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3250 for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3251#ifdef DIAGNOSTIC
3252 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3253 FREE_LOCK(&lk);
3254 panic("softdep_write_inodeblock: lost dep1");
3255 }
3256#endif /* DIAGNOSTIC */
3257 dp->di_db[i] = 0;
3258 }
3259 for (i = 0; i < NIADDR; i++) {
3260#ifdef DIAGNOSTIC
3261 if (dp->di_ib[i] != 0 &&
3262 (deplist & ((1 << NDADDR) << i)) == 0) {
3263 FREE_LOCK(&lk);
3264 panic("softdep_write_inodeblock: lost dep2");
3265 }
3266#endif /* DIAGNOSTIC */
3267 dp->di_ib[i] = 0;
3268 }
3269 FREE_LOCK(&lk);
3270 return;
3271 }
3272 /*
3273 * If we have zero'ed out the last allocated block of the file,
3274 * roll back the size to the last currently allocated block.
3275 * We know that this last allocated block is a full-sized as
3276 * we already checked for fragments in the loop above.
3277 */
3278 if (lastadp != NULL &&
3279 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3280 for (i = lastadp->ad_lbn; i >= 0; i--)
3281 if (dp->di_db[i] != 0)
3282 break;
3283 dp->di_size = (i + 1) * fs->fs_bsize;
3284 }
3285 /*
3286 * The only dependencies are for indirect blocks.
3287 *
3288 * The file size for indirect block additions is not guaranteed.
3289 * Such a guarantee would be non-trivial to achieve. The conventional
3290 * synchronous write implementation also does not make this guarantee.
3291 * Fsck should catch and fix discrepancies. Arguably, the file size
3292 * can be over-estimated without destroying integrity when the file
3293 * moves into the indirect blocks (i.e., is large). If we want to
3294 * postpone fsck, we are stuck with this argument.
3295 */
3296 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3297 dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3298 FREE_LOCK(&lk);
3299}
3300
3301/*
3302 * This routine is called during the completion interrupt
3303 * service routine for a disk write (from the procedure called
3304 * by the device driver to inform the filesystem caches of
3305 * a request completion). It should be called early in this
3306 * procedure, before the block is made available to other
3307 * processes or other routines are called.
3308 */
3309static void
3310softdep_disk_write_complete(bp)
3311 struct buf *bp; /* describes the completed disk write */
3312{
3313 struct worklist *wk;
3314 struct workhead reattach;
3315 struct newblk *newblk;
3316 struct allocindir *aip;
3317 struct allocdirect *adp;
3318 struct indirdep *indirdep;
3319 struct inodedep *inodedep;
3320 struct bmsafemap *bmsafemap;
3321
3322#ifdef DEBUG
3323 if (lk.lkt_held != NOHOLDER)
3324 panic("softdep_disk_write_complete: lock is held");
3325 lk.lkt_held = SPECIAL_FLAG;
3326#endif
3327 LIST_INIT(&reattach);
3328 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3329 WORKLIST_REMOVE(wk);
3330 switch (wk->wk_type) {
3331
3332 case D_PAGEDEP:
3333 if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3334 WORKLIST_INSERT(&reattach, wk);
3335 continue;
3336
3337 case D_INODEDEP:
3338 if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3339 WORKLIST_INSERT(&reattach, wk);
3340 continue;
3341
3342 case D_BMSAFEMAP:
3343 bmsafemap = WK_BMSAFEMAP(wk);
3344 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3345 newblk->nb_state |= DEPCOMPLETE;
3346 newblk->nb_bmsafemap = NULL;
3347 LIST_REMOVE(newblk, nb_deps);
3348 }
3349 while ((adp =
3350 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3351 adp->ad_state |= DEPCOMPLETE;
3352 adp->ad_buf = NULL;
3353 LIST_REMOVE(adp, ad_deps);
3354 handle_allocdirect_partdone(adp);
3355 }
3356 while ((aip =
3357 LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3358 aip->ai_state |= DEPCOMPLETE;
3359 aip->ai_buf = NULL;
3360 LIST_REMOVE(aip, ai_deps);
3361 handle_allocindir_partdone(aip);
3362 }
3363 while ((inodedep =
3364 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3365 inodedep->id_state |= DEPCOMPLETE;
3366 LIST_REMOVE(inodedep, id_deps);
3367 inodedep->id_buf = NULL;
3368 }
3369 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3370 continue;
3371
3372 case D_MKDIR:
3373 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3374 continue;
3375
3376 case D_ALLOCDIRECT:
3377 adp = WK_ALLOCDIRECT(wk);
3378 adp->ad_state |= COMPLETE;
3379 handle_allocdirect_partdone(adp);
3380 continue;
3381
3382 case D_ALLOCINDIR:
3383 aip = WK_ALLOCINDIR(wk);
3384 aip->ai_state |= COMPLETE;
3385 handle_allocindir_partdone(aip);
3386 continue;
3387
3388 case D_INDIRDEP:
3389 indirdep = WK_INDIRDEP(wk);
3390 if (indirdep->ir_state & GOINGAWAY) {
3391 lk.lkt_held = NOHOLDER;
3392 panic("disk_write_complete: indirdep gone");
3393 }
3394 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3395 FREE(indirdep->ir_saveddata, M_INDIRDEP);
3396 indirdep->ir_saveddata = 0;
3397 indirdep->ir_state &= ~UNDONE;
3398 indirdep->ir_state |= ATTACHED;
3399 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3400 handle_allocindir_partdone(aip);
3401 if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
3402 lk.lkt_held = NOHOLDER;
3403 panic("disk_write_complete: not gone");
3404 }
3405 }
3406 WORKLIST_INSERT(&reattach, wk);
3407 if ((bp->b_flags & B_DELWRI) == 0)
3408 stat_indir_blk_ptrs++;
3409 bdirty(bp);
3410 continue;
3411
3412 default:
3413 lk.lkt_held = NOHOLDER;
3414 panic("handle_disk_write_complete: Unknown type %s",
3415 TYPENAME(wk->wk_type));
3416 /* NOTREACHED */
3417 }
3418 }
3419 /*
3420 * Reattach any requests that must be redone.
3421 */
3422 while ((wk = LIST_FIRST(&reattach)) != NULL) {
3423 WORKLIST_REMOVE(wk);
3424 WORKLIST_INSERT(&bp->b_dep, wk);
3425 }
3426#ifdef DEBUG
3427 if (lk.lkt_held != SPECIAL_FLAG)
3428 panic("softdep_disk_write_complete: lock lost");
3429 lk.lkt_held = NOHOLDER;
3430#endif
3431}
3432
3433/*
3434 * Called from within softdep_disk_write_complete above. Note that
3435 * this routine is always called from interrupt level with further
3436 * splbio interrupts blocked.
3437 */
3438static void
3439handle_allocdirect_partdone(adp)
3440 struct allocdirect *adp; /* the completed allocdirect */
3441{
3442 struct allocdirect *listadp;
3443 struct inodedep *inodedep;
3444 long bsize;
3445
3446 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3447 return;
3448 if (adp->ad_buf != NULL) {
3449 lk.lkt_held = NOHOLDER;
3450 panic("handle_allocdirect_partdone: dangling dep");
3451 }
3452 /*
3453 * The on-disk inode cannot claim to be any larger than the last
3454 * fragment that has been written. Otherwise, the on-disk inode
3455 * might have fragments that were not the last block in the file
3456 * which would corrupt the filesystem. Thus, we cannot free any
3457 * allocdirects after one whose ad_oldblkno claims a fragment as
3458 * these blocks must be rolled back to zero before writing the inode.
3459 * We check the currently active set of allocdirects in id_inoupdt.
3460 */
3461 inodedep = adp->ad_inodedep;
3462 bsize = inodedep->id_fs->fs_bsize;
3463 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) {
3464 /* found our block */
3465 if (listadp == adp)
3466 break;
3467 /* continue if ad_oldlbn is not a fragment */
3468 if (listadp->ad_oldsize == 0 ||
3469 listadp->ad_oldsize == bsize)
3470 continue;
3471 /* hit a fragment */
3472 return;
3473 }
3474 /*
3475 * If we have reached the end of the current list without
3476 * finding the just finished dependency, then it must be
3477 * on the future dependency list. Future dependencies cannot
3478 * be freed until they are moved to the current list.
3479 */
3480 if (listadp == NULL) {
3481#ifdef DEBUG
3482 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next)
3483 /* found our block */
3484 if (listadp == adp)
3485 break;
3486 if (listadp == NULL) {
3487 lk.lkt_held = NOHOLDER;
3488 panic("handle_allocdirect_partdone: lost dep");
3489 }
3490#endif /* DEBUG */
3491 return;
3492 }
3493 /*
3494 * If we have found the just finished dependency, then free
3495 * it along with anything that follows it that is complete.
3496 */
3497 for (; adp; adp = listadp) {
3498 listadp = TAILQ_NEXT(adp, ad_next);
3499 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3500 return;
3501 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
3502 }
3503}
3504
3505/*
3506 * Called from within softdep_disk_write_complete above. Note that
3507 * this routine is always called from interrupt level with further
3508 * splbio interrupts blocked.
3509 */
3510static void
3511handle_allocindir_partdone(aip)
3512 struct allocindir *aip; /* the completed allocindir */
3513{
3514 struct indirdep *indirdep;
3515
3516 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
3517 return;
3518 if (aip->ai_buf != NULL) {
3519 lk.lkt_held = NOHOLDER;
3520 panic("handle_allocindir_partdone: dangling dependency");
3521 }
3522 indirdep = aip->ai_indirdep;
3523 if (indirdep->ir_state & UNDONE) {
3524 LIST_REMOVE(aip, ai_next);
3525 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
3526 return;
3527 }
3528 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
3529 aip->ai_newblkno;
3530 LIST_REMOVE(aip, ai_next);
3531 if (aip->ai_freefrag != NULL)
3532 add_to_worklist(&aip->ai_freefrag->ff_list);
3533 WORKITEM_FREE(aip, D_ALLOCINDIR);
3534}
3535
3536/*
3537 * Called from within softdep_disk_write_complete above to restore
3538 * in-memory inode block contents to their most up-to-date state. Note
3539 * that this routine is always called from interrupt level with further
3540 * splbio interrupts blocked.
3541 */
3542static int
3543handle_written_inodeblock(inodedep, bp)
3544 struct inodedep *inodedep;
3545 struct buf *bp; /* buffer containing the inode block */
3546{
3547 struct worklist *wk, *filefree;
3548 struct allocdirect *adp, *nextadp;
3549 struct dinode *dp;
3550 int hadchanges;
3551
3552 if ((inodedep->id_state & IOSTARTED) == 0) {
3553 lk.lkt_held = NOHOLDER;
3554 panic("handle_written_inodeblock: not started");
3555 }
3556 inodedep->id_state &= ~IOSTARTED;
3557 dp = (struct dinode *)bp->b_data +
3558 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
3559 /*
3560 * If we had to rollback the inode allocation because of
3561 * bitmaps being incomplete, then simply restore it.
3562 * Keep the block dirty so that it will not be reclaimed until
3563 * all associated dependencies have been cleared and the
3564 * corresponding updates written to disk.
3565 */
3566 if (inodedep->id_savedino != NULL) {
3567 *dp = *inodedep->id_savedino;
3568 FREE(inodedep->id_savedino, M_INODEDEP);
3569 inodedep->id_savedino = NULL;
3570 if ((bp->b_flags & B_DELWRI) == 0)
3571 stat_inode_bitmap++;
3572 bdirty(bp);
3573 return (1);
3574 }
3575 inodedep->id_state |= COMPLETE;
3576 /*
3577 * Roll forward anything that had to be rolled back before
3578 * the inode could be updated.
3579 */
3580 hadchanges = 0;
3581 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
3582 nextadp = TAILQ_NEXT(adp, ad_next);
3583 if (adp->ad_state & ATTACHED) {
3584 lk.lkt_held = NOHOLDER;
3585 panic("handle_written_inodeblock: new entry");
3586 }
3587 if (adp->ad_lbn < NDADDR) {
3588 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) {
3589 lk.lkt_held = NOHOLDER;
3590 panic("%s: %s #%ld mismatch %d != %d",
3591 "handle_written_inodeblock",
3592 "direct pointer", adp->ad_lbn,
3593 dp->di_db[adp->ad_lbn], adp->ad_oldblkno);
3594 }
3595 dp->di_db[adp->ad_lbn] = adp->ad_newblkno;
3596 } else {
3597 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) {
3598 lk.lkt_held = NOHOLDER;
3599 panic("%s: %s #%ld allocated as %d",
3600 "handle_written_inodeblock",
3601 "indirect pointer", adp->ad_lbn - NDADDR,
3602 dp->di_ib[adp->ad_lbn - NDADDR]);
3603 }
3604 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno;
3605 }
3606 adp->ad_state &= ~UNDONE;
3607 adp->ad_state |= ATTACHED;
3608 hadchanges = 1;
3609 }
3610 if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
3611 stat_direct_blk_ptrs++;
3612 /*
3613 * Reset the file size to its most up-to-date value.
3614 */
3615 if (inodedep->id_savedsize == -1) {
3616 lk.lkt_held = NOHOLDER;
3617 panic("handle_written_inodeblock: bad size");
3618 }
3619 if (dp->di_size != inodedep->id_savedsize) {
3620 dp->di_size = inodedep->id_savedsize;
3621 hadchanges = 1;
3622 }
3623 inodedep->id_savedsize = -1;
3624 /*
3625 * If there were any rollbacks in the inode block, then it must be
3626 * marked dirty so that its will eventually get written back in
3627 * its correct form.
3628 */
3629 if (hadchanges)
3630 bdirty(bp);
3631 /*
3632 * Process any allocdirects that completed during the update.
3633 */
3634 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
3635 handle_allocdirect_partdone(adp);
3636 /*
3637 * Process deallocations that were held pending until the
3638 * inode had been written to disk. Freeing of the inode
3639 * is delayed until after all blocks have been freed to
3640 * avoid creation of new <vfsid, inum, lbn> triples
3641 * before the old ones have been deleted.
3642 */
3643 filefree = NULL;
3644 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
3645 WORKLIST_REMOVE(wk);
3646 switch (wk->wk_type) {
3647
3648 case D_FREEFILE:
3649 /*
3650 * We defer adding filefree to the worklist until
3651 * all other additions have been made to ensure
3652 * that it will be done after all the old blocks
3653 * have been freed.
3654 */
3655 if (filefree != NULL) {
3656 lk.lkt_held = NOHOLDER;
3657 panic("handle_written_inodeblock: filefree");
3658 }
3659 filefree = wk;
3660 continue;
3661
3662 case D_MKDIR:
3663 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
3664 continue;
3665
3666 case D_DIRADD:
3667 diradd_inode_written(WK_DIRADD(wk), inodedep);
3668 continue;
3669
3670 case D_FREEBLKS:
3671 wk->wk_state |= COMPLETE;
3672 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE)
3673 continue;
3674 /* -- fall through -- */
3675 case D_FREEFRAG:
3676 case D_DIRREM:
3677 add_to_worklist(wk);
3678 continue;
3679
3680 default:
3681 lk.lkt_held = NOHOLDER;
3682 panic("handle_written_inodeblock: Unknown type %s",
3683 TYPENAME(wk->wk_type));
3684 /* NOTREACHED */
3685 }
3686 }
3687 if (filefree != NULL) {
3688 if (free_inodedep(inodedep) == 0) {
3689 lk.lkt_held = NOHOLDER;
3690 panic("handle_written_inodeblock: live inodedep");
3691 }
3692 add_to_worklist(filefree);
3693 return (0);
3694 }
3695
3696 /*
3697 * If no outstanding dependencies, free it.
3698 */
3699 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0)
3700 return (0);
3701 return (hadchanges);
3702}
3703
3704/*
3705 * Process a diradd entry after its dependent inode has been written.
3706 * This routine must be called with splbio interrupts blocked.
3707 */
3708static void
3709diradd_inode_written(dap, inodedep)
3710 struct diradd *dap;
3711 struct inodedep *inodedep;
3712{
3713 struct pagedep *pagedep;
3714
3715 dap->da_state |= COMPLETE;
3716 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3717 if (dap->da_state & DIRCHG)
3718 pagedep = dap->da_previous->dm_pagedep;
3719 else
3720 pagedep = dap->da_pagedep;
3721 LIST_REMOVE(dap, da_pdlist);
3722 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3723 }
3724 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3725}
3726
3727/*
3728 * Handle the completion of a mkdir dependency.
3729 */
3730static void
3731handle_written_mkdir(mkdir, type)
3732 struct mkdir *mkdir;
3733 int type;
3734{
3735 struct diradd *dap;
3736 struct pagedep *pagedep;
3737
3738 if (mkdir->md_state != type) {
3739 lk.lkt_held = NOHOLDER;
3740 panic("handle_written_mkdir: bad type");
3741 }
3742 dap = mkdir->md_diradd;
3743 dap->da_state &= ~type;
3744 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
3745 dap->da_state |= DEPCOMPLETE;
3746 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3747 if (dap->da_state & DIRCHG)
3748 pagedep = dap->da_previous->dm_pagedep;
3749 else
3750 pagedep = dap->da_pagedep;
3751 LIST_REMOVE(dap, da_pdlist);
3752 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3753 }
3754 LIST_REMOVE(mkdir, md_mkdirs);
3755 WORKITEM_FREE(mkdir, D_MKDIR);
3756}
3757
3758/*
3759 * Called from within softdep_disk_write_complete above.
3760 * A write operation was just completed. Removed inodes can
3761 * now be freed and associated block pointers may be committed.
3762 * Note that this routine is always called from interrupt level
3763 * with further splbio interrupts blocked.
3764 */
3765static int
3766handle_written_filepage(pagedep, bp)
3767 struct pagedep *pagedep;
3768 struct buf *bp; /* buffer containing the written page */
3769{
3770 struct dirrem *dirrem;
3771 struct diradd *dap, *nextdap;
3772 struct direct *ep;
3773 int i, chgs;
3774
3775 if ((pagedep->pd_state & IOSTARTED) == 0) {
3776 lk.lkt_held = NOHOLDER;
3777 panic("handle_written_filepage: not started");
3778 }
3779 pagedep->pd_state &= ~IOSTARTED;
3780 /*
3781 * Process any directory removals that have been committed.
3782 */
3783 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
3784 LIST_REMOVE(dirrem, dm_next);
3785 dirrem->dm_dirinum = pagedep->pd_ino;
3786 add_to_worklist(&dirrem->dm_list);
3787 }
3788 /*
3789 * Free any directory additions that have been committed.
3790 */
3791 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
3792 free_diradd(dap);
3793 /*
3794 * Uncommitted directory entries must be restored.
3795 */
3796 for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
3797 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
3798 dap = nextdap) {
3799 nextdap = LIST_NEXT(dap, da_pdlist);
3800 if (dap->da_state & ATTACHED) {
3801 lk.lkt_held = NOHOLDER;
3802 panic("handle_written_filepage: attached");
3803 }
3804 ep = (struct direct *)
3805 ((char *)bp->b_data + dap->da_offset);
3806 ep->d_ino = dap->da_newinum;
3807 dap->da_state &= ~UNDONE;
3808 dap->da_state |= ATTACHED;
3809 chgs = 1;
3810 /*
3811 * If the inode referenced by the directory has
3812 * been written out, then the dependency can be
3813 * moved to the pending list.
3814 */
3815 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3816 LIST_REMOVE(dap, da_pdlist);
3817 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
3818 da_pdlist);
3819 }
3820 }
3821 }
3822 /*
3823 * If there were any rollbacks in the directory, then it must be
3824 * marked dirty so that its will eventually get written back in
3825 * its correct form.
3826 */
3827 if (chgs) {
3828 if ((bp->b_flags & B_DELWRI) == 0)
3829 stat_dir_entry++;
3830 bdirty(bp);
3831 }
3832 /*
3833 * If no dependencies remain, the pagedep will be freed.
3834 * Otherwise it will remain to update the page before it
3835 * is written back to disk.
3836 */
3837 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) {
3838 for (i = 0; i < DAHASHSZ; i++)
3839 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
3840 break;
3841 if (i == DAHASHSZ) {
3842 LIST_REMOVE(pagedep, pd_hash);
3843 WORKITEM_FREE(pagedep, D_PAGEDEP);
3844 return (0);
3845 }
3846 }
3847 return (1);
3848}
3849
3850/*
3851 * Writing back in-core inode structures.
3852 *
3853 * The filesystem only accesses an inode's contents when it occupies an
3854 * "in-core" inode structure. These "in-core" structures are separate from
3855 * the page frames used to cache inode blocks. Only the latter are
3856 * transferred to/from the disk. So, when the updated contents of the
3857 * "in-core" inode structure are copied to the corresponding in-memory inode
3858 * block, the dependencies are also transferred. The following procedure is
3859 * called when copying a dirty "in-core" inode to a cached inode block.
3860 */
3861
3862/*
3863 * Called when an inode is loaded from disk. If the effective link count
3864 * differed from the actual link count when it was last flushed, then we
3865 * need to ensure that the correct effective link count is put back.
3866 */
3867void
3868softdep_load_inodeblock(ip)
3869 struct inode *ip; /* the "in_core" copy of the inode */
3870{
3871 struct inodedep *inodedep;
3872
3873 /*
3874 * Check for alternate nlink count.
3875 */
3876 ip->i_effnlink = ip->i_nlink;
3877 ACQUIRE_LOCK(&lk);
3878 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3879 FREE_LOCK(&lk);
3880 return;
3881 }
3882 ip->i_effnlink -= inodedep->id_nlinkdelta;
3883 FREE_LOCK(&lk);
3884}
3885
3886/*
3887 * This routine is called just before the "in-core" inode
3888 * information is to be copied to the in-memory inode block.
3889 * Recall that an inode block contains several inodes. If
3890 * the force flag is set, then the dependencies will be
3891 * cleared so that the update can always be made. Note that
3892 * the buffer is locked when this routine is called, so we
3893 * will never be in the middle of writing the inode block
3894 * to disk.
3895 */
3896void
3897softdep_update_inodeblock(ip, bp, waitfor)
3898 struct inode *ip; /* the "in_core" copy of the inode */
3899 struct buf *bp; /* the buffer containing the inode block */
3900 int waitfor; /* nonzero => update must be allowed */
3901{
3902 struct inodedep *inodedep;
3903 struct worklist *wk;
3904 int error, gotit;
3905
3906 /*
3907 * If the effective link count is not equal to the actual link
3908 * count, then we must track the difference in an inodedep while
3909 * the inode is (potentially) tossed out of the cache. Otherwise,
3910 * if there is no existing inodedep, then there are no dependencies
3911 * to track.
3912 */
3913 ACQUIRE_LOCK(&lk);
3914 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3915 FREE_LOCK(&lk);
3916 if (ip->i_effnlink != ip->i_nlink)
3917 panic("softdep_update_inodeblock: bad link count");
3918 return;
3919 }
3920 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
3921 FREE_LOCK(&lk);
3922 panic("softdep_update_inodeblock: bad delta");
3923 }
3924 /*
3925 * Changes have been initiated. Anything depending on these
3926 * changes cannot occur until this inode has been written.
3927 */
3928 inodedep->id_state &= ~COMPLETE;
3929 if ((inodedep->id_state & ONWORKLIST) == 0)
3930 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
3931 /*
3932 * Any new dependencies associated with the incore inode must
3933 * now be moved to the list associated with the buffer holding
3934 * the in-memory copy of the inode. Once merged process any
3935 * allocdirects that are completed by the merger.
3936 */
3937 merge_inode_lists(inodedep);
3938 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
3939 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
3940 /*
3941 * Now that the inode has been pushed into the buffer, the
3942 * operations dependent on the inode being written to disk
3943 * can be moved to the id_bufwait so that they will be
3944 * processed when the buffer I/O completes.
3945 */
3946 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
3947 WORKLIST_REMOVE(wk);
3948 WORKLIST_INSERT(&inodedep->id_bufwait, wk);
3949 }
3950 /*
3951 * Newly allocated inodes cannot be written until the bitmap
3952 * that allocates them have been written (indicated by
3953 * DEPCOMPLETE being set in id_state). If we are doing a
3954 * forced sync (e.g., an fsync on a file), we force the bitmap
3955 * to be written so that the update can be done.
3956 */
3957 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
3958 FREE_LOCK(&lk);
3959 return;
3960 }
3961 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
3962 FREE_LOCK(&lk);
3963 if (gotit &&
3964 (error = VOP_BWRITE(inodedep->id_buf->b_vp, inodedep->id_buf)) != 0)
3965 softdep_error("softdep_update_inodeblock: bwrite", error);
3966 if ((inodedep->id_state & DEPCOMPLETE) == 0)
3967 panic("softdep_update_inodeblock: update failed");
3968}
3969
3970/*
3971 * Merge the new inode dependency list (id_newinoupdt) into the old
3972 * inode dependency list (id_inoupdt). This routine must be called
3973 * with splbio interrupts blocked.
3974 */
3975static void
3976merge_inode_lists(inodedep)
3977 struct inodedep *inodedep;
3978{
3979 struct allocdirect *listadp, *newadp;
3980
3981 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3982 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) {
3983 if (listadp->ad_lbn < newadp->ad_lbn) {
3984 listadp = TAILQ_NEXT(listadp, ad_next);
3985 continue;
3986 }
3987 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3988 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
3989 if (listadp->ad_lbn == newadp->ad_lbn) {
3990 allocdirect_merge(&inodedep->id_inoupdt, newadp,
3991 listadp);
3992 listadp = newadp;
3993 }
3994 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3995 }
3996 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) {
3997 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3998 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next);
3999 }
4000}
4001
4002/*
4003 * If we are doing an fsync, then we must ensure that any directory
4004 * entries for the inode have been written after the inode gets to disk.
4005 */
4006static int
4007softdep_fsync(vp)
4008 struct vnode *vp; /* the "in_core" copy of the inode */
4009{
4010 struct inodedep *inodedep;
4011 struct pagedep *pagedep;
4012 struct worklist *wk;
4013 struct diradd *dap;
4014 struct mount *mnt;
4015 struct vnode *pvp;
4016 struct inode *ip;
4017 struct buf *bp;
4018 struct fs *fs;
4019 struct thread *td = curthread; /* XXX */
4020 int error, flushparent;
4021 ino_t parentino;
4022 ufs_lbn_t lbn;
4023
4024 ip = VTOI(vp);
4025 fs = ip->i_fs;
4026 ACQUIRE_LOCK(&lk);
4027 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
4028 FREE_LOCK(&lk);
4029 return (0);
4030 }
4031 if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
4032 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
4033 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
4034 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) {
4035 FREE_LOCK(&lk);
4036 panic("softdep_fsync: pending ops");
4037 }
4038 for (error = 0, flushparent = 0; ; ) {
4039 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
4040 break;
4041 if (wk->wk_type != D_DIRADD) {
4042 FREE_LOCK(&lk);
4043 panic("softdep_fsync: Unexpected type %s",
4044 TYPENAME(wk->wk_type));
4045 }
4046 dap = WK_DIRADD(wk);
4047 /*
4048 * Flush our parent if this directory entry
4049 * has a MKDIR_PARENT dependency.
4050 */
4051 if (dap->da_state & DIRCHG)
4052 pagedep = dap->da_previous->dm_pagedep;
4053 else
4054 pagedep = dap->da_pagedep;
4055 mnt = pagedep->pd_mnt;
4056 parentino = pagedep->pd_ino;
4057 lbn = pagedep->pd_lbn;
4058 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) {
4059 FREE_LOCK(&lk);
4060 panic("softdep_fsync: dirty");
4061 }
4062 flushparent = dap->da_state & MKDIR_PARENT;
4063 /*
4064 * If we are being fsync'ed as part of vgone'ing this vnode,
4065 * then we will not be able to release and recover the
4066 * vnode below, so we just have to give up on writing its
4067 * directory entry out. It will eventually be written, just
4068 * not now, but then the user was not asking to have it
4069 * written, so we are not breaking any promises.
4070 */
4071 if (vp->v_flag & VRECLAIMED)
4072 break;
4073 /*
4074 * We prevent deadlock by always fetching inodes from the
4075 * root, moving down the directory tree. Thus, when fetching
4076 * our parent directory, we must unlock ourselves before
4077 * requesting the lock on our parent. See the comment in
4078 * ufs_lookup for details on possible races.
4079 */
4080 FREE_LOCK(&lk);
4081 VOP_UNLOCK(vp, 0, td);
4082 error = VFS_VGET(mnt, parentino, &pvp);
4083 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
4084 if (error != 0)
4085 return (error);
4086 if (flushparent) {
4087 if ((error = UFS_UPDATE(pvp, 1)) != 0) {
4088 vput(pvp);
4089 return (error);
4090 }
4091 }
4092 /*
4093 * Flush directory page containing the inode's name.
4094 */
4095 error = bread(pvp, lblktodoff(fs, lbn), blksize(fs, VTOI(pvp), lbn), &bp);
4096 if (error == 0)
4097 error = VOP_BWRITE(bp->b_vp, bp);
4098 vput(pvp);
4099 if (error != 0)
4100 return (error);
4101 ACQUIRE_LOCK(&lk);
4102 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
4103 break;
4104 }
4105 FREE_LOCK(&lk);
4106 return (0);
4107}
4108
4109/*
4110 * Flush all the dirty bitmaps associated with the block device
4111 * before flushing the rest of the dirty blocks so as to reduce
4112 * the number of dependencies that will have to be rolled back.
4113 */
4114static int softdep_fsync_mountdev_bp(struct buf *bp, void *data);
4115
4116void
4117softdep_fsync_mountdev(vp)
4118 struct vnode *vp;
4119{
4120 if (!vn_isdisk(vp, NULL))
4121 panic("softdep_fsync_mountdev: vnode not a disk");
4122 ACQUIRE_LOCK(&lk);
4123 crit_enter();
4124 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
4125 softdep_fsync_mountdev_bp, vp);
4126 crit_exit();
4127 drain_output(vp, 1);
4128 FREE_LOCK(&lk);
4129}
4130
4131static int
4132softdep_fsync_mountdev_bp(struct buf *bp, void *data)
4133{
4134 struct worklist *wk;
4135 struct vnode *vp = data;
4136
4137 /*
4138 * If it is already scheduled, skip to the next buffer.
4139 */
4140 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
4141 return(0);
4142 if (bp->b_vp != vp || (bp->b_flags & B_DELWRI) == 0) {
4143 BUF_UNLOCK(bp);
4144 printf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp);
4145 return(0);
4146 }
4147 /*
4148 * We are only interested in bitmaps with outstanding
4149 * dependencies.
4150 */
4151 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
4152 wk->wk_type != D_BMSAFEMAP ||
4153 (bp->b_xflags & BX_BKGRDINPROG)) {
4154 BUF_UNLOCK(bp);
4155 return(0);
4156 }
4157 bremfree(bp);
4158 FREE_LOCK(&lk);
4159 (void) bawrite(bp);
4160 ACQUIRE_LOCK(&lk);
4161 return(0);
4162}
4163
4164/*
4165 * This routine is called when we are trying to synchronously flush a
4166 * file. This routine must eliminate any filesystem metadata dependencies
4167 * so that the syncing routine can succeed by pushing the dirty blocks
4168 * associated with the file. If any I/O errors occur, they are returned.
4169 */
4170struct softdep_sync_metadata_info {
4171 struct vnode *vp;
4172 int waitfor;
4173};
4174
4175static int softdep_sync_metadata_bp(struct buf *bp, void *data);
4176
4177int
4178softdep_sync_metadata(struct vnode *vp, struct thread *td)
4179{
4180 struct softdep_sync_metadata_info info;
4181 int error, waitfor;
4182
4183 /*
4184 * Check whether this vnode is involved in a filesystem
4185 * that is doing soft dependency processing.
4186 */
4187 if (!vn_isdisk(vp, NULL)) {
4188 if (!DOINGSOFTDEP(vp))
4189 return (0);
4190 } else
4191 if (vp->v_rdev->si_mountpoint == NULL ||
4192 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0)
4193 return (0);
4194 /*
4195 * Ensure that any direct block dependencies have been cleared.
4196 */
4197 ACQUIRE_LOCK(&lk);
4198 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
4199 FREE_LOCK(&lk);
4200 return (error);
4201 }
4202 /*
4203 * For most files, the only metadata dependencies are the
4204 * cylinder group maps that allocate their inode or blocks.
4205 * The block allocation dependencies can be found by traversing
4206 * the dependency lists for any buffers that remain on their
4207 * dirty buffer list. The inode allocation dependency will
4208 * be resolved when the inode is updated with MNT_WAIT.
4209 * This work is done in two passes. The first pass grabs most
4210 * of the buffers and begins asynchronously writing them. The
4211 * only way to wait for these asynchronous writes is to sleep
4212 * on the filesystem vnode which may stay busy for a long time
4213 * if the filesystem is active. So, instead, we make a second
4214 * pass over the dependencies blocking on each write. In the
4215 * usual case we will be blocking against a write that we
4216 * initiated, so when it is done the dependency will have been
4217 * resolved. Thus the second pass is expected to end quickly.
4218 */
4219 waitfor = MNT_NOWAIT;
4220top:
4221 /*
4222 * We must wait for any I/O in progress to finish so that
4223 * all potential buffers on the dirty list will be visible.
4224 */
4225 drain_output(vp, 1);
4226 info.vp = vp;
4227 info.waitfor = waitfor;
4228 crit_enter();
4229 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
4230 softdep_sync_metadata_bp, &info);
4231 crit_exit();
4232 if (error < 0) {
4233 FREE_LOCK(&lk);
4234 return(-error); /* error code */
4235 }
4236
4237 /*
4238 * The brief unlock is to allow any pent up dependency
4239 * processing to be done. Then proceed with the second pass.
4240 */
4241 if (waitfor == MNT_NOWAIT) {
4242 waitfor = MNT_WAIT;
4243 FREE_LOCK(&lk);
4244 ACQUIRE_LOCK(&lk);
4245 goto top;
4246 }
4247
4248 /*
4249 * If we have managed to get rid of all the dirty buffers,
4250 * then we are done. For certain directories and block
4251 * devices, we may need to do further work.
4252 *
4253 * We must wait for any I/O in progress to finish so that
4254 * all potential buffers on the dirty list will be visible.
4255 */
4256 drain_output(vp, 1);
4257 if (RB_EMPTY(&vp->v_rbdirty_tree)) {
4258 FREE_LOCK(&lk);
4259 return (0);
4260 }
4261
4262 FREE_LOCK(&lk);
4263 /*
4264 * If we are trying to sync a block device, some of its buffers may
4265 * contain metadata that cannot be written until the contents of some
4266 * partially written files have been written to disk. The only easy
4267 * way to accomplish this is to sync the entire filesystem (luckily
4268 * this happens rarely).
4269 */
4270 if (vn_isdisk(vp, NULL) &&
4271 vp->v_rdev &&
4272 vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) &&
4273 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, td)) != 0)
4274 return (error);
4275 return (0);
4276}
4277
4278static int
4279softdep_sync_metadata_bp(struct buf *bp, void *data)
4280{
4281 struct softdep_sync_metadata_info *info = data;
4282 struct pagedep *pagedep;
4283 struct allocdirect *adp;
4284 struct allocindir *aip;
4285 struct worklist *wk;
4286 struct buf *nbp;
4287 int error;
4288 int i;
4289
4290 if (getdirtybuf(&bp, MNT_WAIT) == 0) {
4291 printf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp);
4292 return (1);
4293 }
4294 if (bp->b_vp != info->vp || (bp->b_flags & B_DELWRI) == 0) {
4295 printf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp, info->vp);
4296 BUF_UNLOCK(bp);
4297 return(1);
4298 }
4299
4300 /*
4301 * As we hold the buffer locked, none of its dependencies
4302 * will disappear.
4303 */
4304 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4305 switch (wk->wk_type) {
4306
4307 case D_ALLOCDIRECT:
4308 adp = WK_ALLOCDIRECT(wk);
4309 if (adp->ad_state & DEPCOMPLETE)
4310 break;
4311 nbp = adp->ad_buf;
4312 if (getdirtybuf(&nbp, info->waitfor) == 0)
4313 break;
4314 FREE_LOCK(&lk);
4315 if (info->waitfor == MNT_NOWAIT) {
4316 bawrite(nbp);
4317 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4318 bawrite(bp);
4319 ACQUIRE_LOCK(&lk);
4320 return (-error);
4321 }
4322 ACQUIRE_LOCK(&lk);
4323 break;
4324
4325 case D_ALLOCINDIR:
4326 aip = WK_ALLOCINDIR(wk);
4327 if (aip->ai_state & DEPCOMPLETE)
4328 break;
4329 nbp = aip->ai_buf;
4330 if (getdirtybuf(&nbp, info->waitfor) == 0)
4331 break;
4332 FREE_LOCK(&lk);
4333 if (info->waitfor == MNT_NOWAIT) {
4334 bawrite(nbp);
4335 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4336 bawrite(bp);
4337 ACQUIRE_LOCK(&lk);
4338 return (-error);
4339 }
4340 ACQUIRE_LOCK(&lk);
4341 break;
4342
4343 case D_INDIRDEP:
4344 restart:
4345
4346 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
4347 if (aip->ai_state & DEPCOMPLETE)
4348 continue;
4349 nbp = aip->ai_buf;
4350 if (getdirtybuf(&nbp, MNT_WAIT) == 0)
4351 goto restart;
4352 FREE_LOCK(&lk);
4353 if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4354 bawrite(bp);
4355 ACQUIRE_LOCK(&lk);
4356 return (-error);
4357 }
4358 ACQUIRE_LOCK(&lk);
4359 goto restart;
4360 }
4361 break;
4362
4363 case D_INODEDEP:
4364 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
4365 WK_INODEDEP(wk)->id_ino)) != 0) {
4366 FREE_LOCK(&lk);
4367 bawrite(bp);
4368 ACQUIRE_LOCK(&lk);
4369 return (-error);
4370 }
4371 break;
4372
4373 case D_PAGEDEP:
4374 /*
4375 * We are trying to sync a directory that may
4376 * have dependencies on both its own metadata
4377 * and/or dependencies on the inodes of any
4378 * recently allocated files. We walk its diradd
4379 * lists pushing out the associated inode.
4380 */
4381 pagedep = WK_PAGEDEP(wk);
4382 for (i = 0; i < DAHASHSZ; i++) {
4383 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
4384 continue;
4385 if ((error =
4386 flush_pagedep_deps(info->vp,
4387 pagedep->pd_mnt,
4388 &pagedep->pd_diraddhd[i]))) {
4389 FREE_LOCK(&lk);
4390 bawrite(bp);
4391 ACQUIRE_LOCK(&lk);
4392 return (-error);
4393 }
4394 }
4395 break;
4396
4397 case D_MKDIR:
4398 /*
4399 * This case should never happen if the vnode has
4400 * been properly sync'ed. However, if this function
4401 * is used at a place where the vnode has not yet
4402 * been sync'ed, this dependency can show up. So,
4403 * rather than panic, just flush it.
4404 */
4405 nbp = WK_MKDIR(wk)->md_buf;
4406 if (getdirtybuf(&nbp, info->waitfor) == 0)
4407 break;
4408 FREE_LOCK(&lk);
4409 if (info->waitfor == MNT_NOWAIT) {
4410 bawrite(nbp);
4411 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4412 bawrite(bp);
4413 ACQUIRE_LOCK(&lk);
4414 return (-error);
4415 }
4416 ACQUIRE_LOCK(&lk);
4417 break;
4418
4419 case D_BMSAFEMAP:
4420 /*
4421 * This case should never happen if the vnode has
4422 * been properly sync'ed. However, if this function
4423 * is used at a place where the vnode has not yet
4424 * been sync'ed, this dependency can show up. So,
4425 * rather than panic, just flush it.
4426 *
4427 * nbp can wind up == bp if a device node for the
4428 * same filesystem is being fsynced at the same time,
4429 * leading to a panic if we don't catch the case.
4430 */
4431 nbp = WK_BMSAFEMAP(wk)->sm_buf;
4432 if (nbp == bp)
4433 break;
4434 if (getdirtybuf(&nbp, info->waitfor) == 0)
4435 break;
4436 FREE_LOCK(&lk);
4437 if (info->waitfor == MNT_NOWAIT) {
4438 bawrite(nbp);
4439 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4440 bawrite(bp);
4441 ACQUIRE_LOCK(&lk);
4442 return (-error);
4443 }
4444 ACQUIRE_LOCK(&lk);
4445 break;
4446
4447 default:
4448 FREE_LOCK(&lk);
4449 panic("softdep_sync_metadata: Unknown type %s",
4450 TYPENAME(wk->wk_type));
4451 /* NOTREACHED */
4452 }
4453 }
4454 FREE_LOCK(&lk);
4455 bawrite(bp);
4456 ACQUIRE_LOCK(&lk);
4457 return(0);
4458}
4459
4460/*
4461 * Flush the dependencies associated with an inodedep.
4462 * Called with splbio blocked.
4463 */
4464static int
4465flush_inodedep_deps(fs, ino)
4466 struct fs *fs;
4467 ino_t ino;
4468{
4469 struct inodedep *inodedep;
4470 struct allocdirect *adp;
4471 int error, waitfor;
4472 struct buf *bp;
4473
4474 /*
4475 * This work is done in two passes. The first pass grabs most
4476 * of the buffers and begins asynchronously writing them. The
4477 * only way to wait for these asynchronous writes is to sleep
4478 * on the filesystem vnode which may stay busy for a long time
4479 * if the filesystem is active. So, instead, we make a second
4480 * pass over the dependencies blocking on each write. In the
4481 * usual case we will be blocking against a write that we
4482 * initiated, so when it is done the dependency will have been
4483 * resolved. Thus the second pass is expected to end quickly.
4484 * We give a brief window at the top of the loop to allow
4485 * any pending I/O to complete.
4486 */
4487 for (waitfor = MNT_NOWAIT; ; ) {
4488 FREE_LOCK(&lk);
4489 ACQUIRE_LOCK(&lk);
4490 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4491 return (0);
4492 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) {
4493 if (adp->ad_state & DEPCOMPLETE)
4494 continue;
4495 bp = adp->ad_buf;
4496 if (getdirtybuf(&bp, waitfor) == 0) {
4497 if (waitfor == MNT_NOWAIT)
4498 continue;
4499 break;
4500 }
4501 FREE_LOCK(&lk);
4502 if (waitfor == MNT_NOWAIT) {
4503 bawrite(bp);
4504 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) {
4505 ACQUIRE_LOCK(&lk);
4506 return (error);
4507 }
4508 ACQUIRE_LOCK(&lk);
4509 break;
4510 }
4511 if (adp != NULL)
4512 continue;
4513 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) {
4514 if (adp->ad_state & DEPCOMPLETE)
4515 continue;
4516 bp = adp->ad_buf;
4517 if (getdirtybuf(&bp, waitfor) == 0) {
4518 if (waitfor == MNT_NOWAIT)
4519 continue;
4520 break;
4521 }
4522 FREE_LOCK(&lk);
4523 if (waitfor == MNT_NOWAIT) {
4524 bawrite(bp);
4525 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) {
4526 ACQUIRE_LOCK(&lk);
4527 return (error);
4528 }
4529 ACQUIRE_LOCK(&lk);
4530 break;
4531 }
4532 if (adp != NULL)
4533 continue;
4534 /*
4535 * If pass2, we are done, otherwise do pass 2.
4536 */
4537 if (waitfor == MNT_WAIT)
4538 break;
4539 waitfor = MNT_WAIT;
4540 }
4541 /*
4542 * Try freeing inodedep in case all dependencies have been removed.
4543 */
4544 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
4545 (void) free_inodedep(inodedep);
4546 return (0);
4547}
4548
4549/*
4550 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
4551 * Called with splbio blocked.
4552 */
4553static int
4554flush_pagedep_deps(pvp, mp, diraddhdp)
4555 struct vnode *pvp;
4556 struct mount *mp;
4557 struct diraddhd *diraddhdp;
4558{
4559 struct thread *td = curthread; /* XXX */
4560 struct inodedep *inodedep;
4561 struct ufsmount *ump;
4562 struct diradd *dap;
4563 struct vnode *vp;
4564 int gotit, error = 0;
4565 struct buf *bp;
4566 ino_t inum;
4567
4568 ump = VFSTOUFS(mp);
4569 while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
4570 /*
4571 * Flush ourselves if this directory entry
4572 * has a MKDIR_PARENT dependency.
4573 */
4574 if (dap->da_state & MKDIR_PARENT) {
4575 FREE_LOCK(&lk);
4576 if ((error = UFS_UPDATE(pvp, 1)) != 0)
4577 break;
4578 ACQUIRE_LOCK(&lk);
4579 /*
4580 * If that cleared dependencies, go on to next.
4581 */
4582 if (dap != LIST_FIRST(diraddhdp))
4583 continue;
4584 if (dap->da_state & MKDIR_PARENT) {
4585 FREE_LOCK(&lk);
4586 panic("flush_pagedep_deps: MKDIR_PARENT");
4587 }
4588 }
4589 /*
4590 * A newly allocated directory must have its "." and
4591 * ".." entries written out before its name can be
4592 * committed in its parent. We do not want or need
4593 * the full semantics of a synchronous VOP_FSYNC as
4594 * that may end up here again, once for each directory
4595 * level in the filesystem. Instead, we push the blocks
4596 * and wait for them to clear. We have to fsync twice
4597 * because the first call may choose to defer blocks
4598 * that still have dependencies, but deferral will
4599 * happen at most once.
4600 */
4601 inum = dap->da_newinum;
4602 if (dap->da_state & MKDIR_BODY) {
4603 FREE_LOCK(&lk);
4604 if ((error = VFS_VGET(mp, inum, &vp)) != 0)
4605 break;
4606 if ((error=VOP_FSYNC(vp, MNT_NOWAIT, td)) ||
4607 (error=VOP_FSYNC(vp, MNT_NOWAIT, td))) {
4608 vput(vp);
4609 break;
4610 }
4611 drain_output(vp, 0);
4612 vput(vp);
4613 ACQUIRE_LOCK(&lk);
4614 /*
4615 * If that cleared dependencies, go on to next.
4616 */
4617 if (dap != LIST_FIRST(diraddhdp))
4618 continue;
4619 if (dap->da_state & MKDIR_BODY) {
4620 FREE_LOCK(&lk);
4621 panic("flush_pagedep_deps: MKDIR_BODY");
4622 }
4623 }
4624 /*
4625 * Flush the inode on which the directory entry depends.
4626 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
4627 * the only remaining dependency is that the updated inode
4628 * count must get pushed to disk. The inode has already
4629 * been pushed into its inode buffer (via VOP_UPDATE) at
4630 * the time of the reference count change. So we need only
4631 * locate that buffer, ensure that there will be no rollback
4632 * caused by a bitmap dependency, then write the inode buffer.
4633 */
4634 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) {
4635 FREE_LOCK(&lk);
4636 panic("flush_pagedep_deps: lost inode");
4637 }
4638 /*
4639 * If the inode still has bitmap dependencies,
4640 * push them to disk.
4641 */
4642 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4643 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4644 FREE_LOCK(&lk);
4645 if (gotit &&
4646 (error = VOP_BWRITE(inodedep->id_buf->b_vp,
4647 inodedep->id_buf)) != 0)
4648 break;
4649 ACQUIRE_LOCK(&lk);
4650 if (dap != LIST_FIRST(diraddhdp))
4651 continue;
4652 }
4653 /*
4654 * If the inode is still sitting in a buffer waiting
4655 * to be written, push it to disk.
4656 */
4657 FREE_LOCK(&lk);
4658 if ((error = bread(ump->um_devvp,
4659 fsbtodoff(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
4660 (int)ump->um_fs->fs_bsize, &bp)) != 0)
4661 break;
4662 if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0)
4663 break;
4664 ACQUIRE_LOCK(&lk);
4665 /*
4666 * If we have failed to get rid of all the dependencies
4667 * then something is seriously wrong.
4668 */
4669 if (dap == LIST_FIRST(diraddhdp)) {
4670 FREE_LOCK(&lk);
4671 panic("flush_pagedep_deps: flush failed");
4672 }
4673 }
4674 if (error)
4675 ACQUIRE_LOCK(&lk);
4676 return (error);
4677}
4678
4679/*
4680 * A large burst of file addition or deletion activity can drive the
4681 * memory load excessively high. First attempt to slow things down
4682 * using the techniques below. If that fails, this routine requests
4683 * the offending operations to fall back to running synchronously
4684 * until the memory load returns to a reasonable level.
4685 */
4686int
4687softdep_slowdown(vp)
4688 struct vnode *vp;
4689{
4690 int max_softdeps_hard;
4691
4692 max_softdeps_hard = max_softdeps * 11 / 10;
4693 if (num_dirrem < max_softdeps_hard / 2 &&
4694 num_inodedep < max_softdeps_hard)
4695 return (0);
4696 stat_sync_limit_hit += 1;
4697 return (1);
4698}
4699
4700/*
4701 * If memory utilization has gotten too high, deliberately slow things
4702 * down and speed up the I/O processing.
4703 */
4704static int
4705request_cleanup(resource, islocked)
4706 int resource;
4707 int islocked;
4708{
4709 struct thread *td = curthread; /* XXX */
4710
4711 /*
4712 * We never hold up the filesystem syncer process.
4713 */
4714 if (td == filesys_syncer)
4715 return (0);
4716 /*
4717 * First check to see if the work list has gotten backlogged.
4718 * If it has, co-opt this process to help clean up two entries.
4719 * Because this process may hold inodes locked, we cannot
4720 * handle any remove requests that might block on a locked
4721 * inode as that could lead to deadlock.
4722 */
4723 if (num_on_worklist > max_softdeps / 10) {
4724 if (islocked)
4725 FREE_LOCK(&lk);
4726 process_worklist_item(NULL, LK_NOWAIT);
4727 process_worklist_item(NULL, LK_NOWAIT);
4728 stat_worklist_push += 2;
4729 if (islocked)
4730 ACQUIRE_LOCK(&lk);
4731 return(1);
4732 }
4733
4734 /*
4735 * If we are resource constrained on inode dependencies, try
4736 * flushing some dirty inodes. Otherwise, we are constrained
4737 * by file deletions, so try accelerating flushes of directories
4738 * with removal dependencies. We would like to do the cleanup
4739 * here, but we probably hold an inode locked at this point and
4740 * that might deadlock against one that we try to clean. So,
4741 * the best that we can do is request the syncer daemon to do
4742 * the cleanup for us.
4743 */
4744 switch (resource) {
4745
4746 case FLUSH_INODES:
4747 stat_ino_limit_push += 1;
4748 req_clear_inodedeps += 1;
4749 stat_countp = &stat_ino_limit_hit;
4750 break;
4751
4752 case FLUSH_REMOVE:
4753 stat_blk_limit_push += 1;
4754 req_clear_remove += 1;
4755 stat_countp = &stat_blk_limit_hit;
4756 break;
4757
4758 default:
4759 if (islocked)
4760 FREE_LOCK(&lk);
4761 panic("request_cleanup: unknown type");
4762 }
4763 /*
4764 * Hopefully the syncer daemon will catch up and awaken us.
4765 * We wait at most tickdelay before proceeding in any case.
4766 */
4767 if (islocked == 0)
4768 ACQUIRE_LOCK(&lk);
4769 crit_enter();
4770 proc_waiting += 1;
4771 if (!callout_active(&handle))
4772 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2,
4773 pause_timer, NULL);
4774 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0,
4775 "softupdate", 0);
4776 proc_waiting -= 1;
4777 crit_exit();
4778 if (islocked == 0)
4779 FREE_LOCK(&lk);
4780 return (1);
4781}
4782
4783/*
4784 * Awaken processes pausing in request_cleanup and clear proc_waiting
4785 * to indicate that there is no longer a timer running.
4786 */
4787void
4788pause_timer(arg)
4789 void *arg;
4790{
4791 *stat_countp += 1;
4792 wakeup_one(&proc_waiting);
4793 if (proc_waiting > 0)
4794 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2,
4795 pause_timer, NULL);
4796 else
4797 callout_deactivate(&handle);
4798}
4799
4800/*
4801 * Flush out a directory with at least one removal dependency in an effort to
4802 * reduce the number of dirrem, freefile, and freeblks dependency structures.
4803 */
4804static void
4805clear_remove(struct thread *td)
4806{
4807 struct pagedep_hashhead *pagedephd;
4808 struct pagedep *pagedep;
4809 static int next = 0;
4810 struct mount *mp;
4811 struct vnode *vp;
4812 int error, cnt;
4813 ino_t ino;
4814
4815 ACQUIRE_LOCK(&lk);
4816 for (cnt = 0; cnt < pagedep_hash; cnt++) {
4817 pagedephd = &pagedep_hashtbl[next++];
4818 if (next >= pagedep_hash)
4819 next = 0;
4820 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
4821 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
4822 continue;
4823 mp = pagedep->pd_mnt;
4824 ino = pagedep->pd_ino;
4825 FREE_LOCK(&lk);
4826 if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4827 softdep_error("clear_remove: vget", error);
4828 return;
4829 }
4830 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, td)))
4831 softdep_error("clear_remove: fsync", error);
4832 drain_output(vp, 0);
4833 vput(vp);
4834 return;
4835 }
4836 }
4837 FREE_LOCK(&lk);
4838}
4839
4840/*
4841 * Clear out a block of dirty inodes in an effort to reduce
4842 * the number of inodedep dependency structures.
4843 */
4844struct clear_inodedeps_info {
4845 struct fs *fs;
4846 struct mount *mp;
4847};
4848
4849static int
4850clear_inodedeps_mountlist_callback(struct mount *mp, void *data)
4851{
4852 struct clear_inodedeps_info *info = data;
4853
4854 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) {
4855 info->mp = mp;
4856 return(-1);
4857 }
4858 return(0);
4859}
4860
4861static void
4862clear_inodedeps(struct thread *td)
4863{
4864 struct clear_inodedeps_info info;
4865 struct inodedep_hashhead *inodedephd;
4866 struct inodedep *inodedep;
4867 static int next = 0;
4868 struct vnode *vp;
4869 struct fs *fs;
4870 int error, cnt;
4871 ino_t firstino, lastino, ino;
4872
4873 ACQUIRE_LOCK(&lk);
4874 /*
4875 * Pick a random inode dependency to be cleared.
4876 * We will then gather up all the inodes in its block
4877 * that have dependencies and flush them out.
4878 */
4879 for (cnt = 0; cnt < inodedep_hash; cnt++) {
4880 inodedephd = &inodedep_hashtbl[next++];
4881 if (next >= inodedep_hash)
4882 next = 0;
4883 if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
4884 break;
4885 }
4886 if (inodedep == NULL) {
4887 FREE_LOCK(&lk);
4888 return;
4889 }
4890 /*
4891 * Ugly code to find mount point given pointer to superblock.
4892 */
4893 fs = inodedep->id_fs;
4894 info.mp = NULL;
4895 info.fs = fs;
4896 mountlist_scan(clear_inodedeps_mountlist_callback,
4897 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
4898 /*
4899 * Find the last inode in the block with dependencies.
4900 */
4901 firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
4902 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
4903 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
4904 break;
4905 /*
4906 * Asynchronously push all but the last inode with dependencies.
4907 * Synchronously push the last inode with dependencies to ensure
4908 * that the inode block gets written to free up the inodedeps.
4909 */
4910 for (ino = firstino; ino <= lastino; ino++) {
4911 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4912 continue;
4913 FREE_LOCK(&lk);
4914 if ((error = VFS_VGET(info.mp, ino, &vp)) != 0) {
4915 softdep_error("clear_inodedeps: vget", error);
4916 return;
4917 }
4918 if (ino == lastino) {
4919 if ((error = VOP_FSYNC(vp, MNT_WAIT, td)))
4920 softdep_error("clear_inodedeps: fsync1", error);
4921 } else {
4922 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, td)))
4923 softdep_error("clear_inodedeps: fsync2", error);
4924 drain_output(vp, 0);
4925 }
4926 vput(vp);
4927 ACQUIRE_LOCK(&lk);
4928 }
4929 FREE_LOCK(&lk);
4930}
4931
4932/*
4933 * Function to determine if the buffer has outstanding dependencies
4934 * that will cause a roll-back if the buffer is written. If wantcount
4935 * is set, return number of dependencies, otherwise just yes or no.
4936 */
4937static int
4938softdep_count_dependencies(bp, wantcount)
4939 struct buf *bp;
4940 int wantcount;
4941{
4942 struct worklist *wk;
4943 struct inodedep *inodedep;
4944 struct indirdep *indirdep;
4945 struct allocindir *aip;
4946 struct pagedep *pagedep;
4947 struct diradd *dap;
4948 int i, retval;
4949
4950 retval = 0;
4951 ACQUIRE_LOCK(&lk);
4952 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4953 switch (wk->wk_type) {
4954
4955 case D_INODEDEP:
4956 inodedep = WK_INODEDEP(wk);
4957 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4958 /* bitmap allocation dependency */
4959 retval += 1;
4960 if (!wantcount)
4961 goto out;
4962 }
4963 if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
4964 /* direct block pointer dependency */
4965 retval += 1;
4966 if (!wantcount)
4967 goto out;
4968 }
4969 continue;
4970
4971 case D_INDIRDEP:
4972 indirdep = WK_INDIRDEP(wk);
4973
4974 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
4975 /* indirect block pointer dependency */
4976 retval += 1;
4977 if (!wantcount)
4978 goto out;
4979 }
4980 continue;
4981
4982 case D_PAGEDEP:
4983 pagedep = WK_PAGEDEP(wk);
4984 for (i = 0; i < DAHASHSZ; i++) {
4985
4986 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
4987 /* directory entry dependency */
4988 retval += 1;
4989 if (!wantcount)
4990 goto out;
4991 }
4992 }
4993 continue;
4994
4995 case D_BMSAFEMAP:
4996 case D_ALLOCDIRECT:
4997 case D_ALLOCINDIR:
4998 case D_MKDIR:
4999 /* never a dependency on these blocks */
5000 continue;
5001
5002 default:
5003 FREE_LOCK(&lk);
5004 panic("softdep_check_for_rollback: Unexpected type %s",
5005 TYPENAME(wk->wk_type));
5006 /* NOTREACHED */
5007 }
5008 }
5009out:
5010 FREE_LOCK(&lk);
5011 return retval;
5012}
5013
5014/*
5015 * Acquire exclusive access to a buffer.
5016 * Must be called with splbio blocked.
5017 * Return 1 if buffer was acquired.
5018 */
5019static int
5020getdirtybuf(bpp, waitfor)
5021 struct buf **bpp;
5022 int waitfor;
5023{
5024 struct buf *bp;
5025 int error;
5026
5027 for (;;) {
5028 if ((bp = *bpp) == NULL)
5029 return (0);
5030 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
5031 if ((bp->b_xflags & BX_BKGRDINPROG) == 0)
5032 break;
5033 BUF_UNLOCK(bp);
5034 if (waitfor != MNT_WAIT)
5035 return (0);
5036 bp->b_xflags |= BX_BKGRDWAIT;
5037 interlocked_sleep(&lk, SLEEP, &bp->b_xflags, 0,
5038 "getbuf", 0);
5039 continue;
5040 }
5041 if (waitfor != MNT_WAIT)
5042 return (0);
5043 error = interlocked_sleep(&lk, LOCKBUF, bp,
5044 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0);
5045 if (error != ENOLCK) {
5046 FREE_LOCK(&lk);
5047 panic("getdirtybuf: inconsistent lock");
5048 }
5049 }
5050 if ((bp->b_flags & B_DELWRI) == 0) {
5051 BUF_UNLOCK(bp);
5052 return (0);
5053 }
5054 bremfree(bp);
5055 return (1);
5056}
5057
5058/*
5059 * Wait for pending output on a vnode to complete.
5060 * Must be called with vnode locked.
5061 */
5062static void
5063drain_output(vp, islocked)
5064 struct vnode *vp;
5065 int islocked;
5066{
5067
5068 if (!islocked)
5069 ACQUIRE_LOCK(&lk);
5070 while (vp->v_track_write.bk_active) {
5071 vp->v_track_write.bk_waitflag = 1;
5072 interlocked_sleep(&lk, SLEEP, &vp->v_track_write,
5073 0, "drainvp", 0);
5074 }
5075 if (!islocked)
5076 FREE_LOCK(&lk);
5077}
5078
5079/*
5080 * Called whenever a buffer that is being invalidated or reallocated
5081 * contains dependencies. This should only happen if an I/O error has
5082 * occurred. The routine is called with the buffer locked.
5083 */
5084static void
5085softdep_deallocate_dependencies(bp)
5086 struct buf *bp;
5087{
5088
5089 if ((bp->b_flags & B_ERROR) == 0)
5090 panic("softdep_deallocate_dependencies: dangling deps");
5091 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error);
5092 panic("softdep_deallocate_dependencies: unrecovered I/O error");
5093}
5094
5095/*
5096 * Function to handle asynchronous write errors in the filesystem.
5097 */
5098void
5099softdep_error(func, error)
5100 char *func;
5101 int error;
5102{
5103
5104 /* XXX should do something better! */
5105 printf("%s: got error %d while accessing filesystem\n", func, error);
5106}