2 * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
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).
9 * Further information about soft updates can be obtained from:
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
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
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.
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
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 $
43 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
52 #include <sys/param.h>
53 #include <sys/kernel.h>
54 #include <sys/systm.h>
56 #include <sys/malloc.h>
57 #include <sys/mount.h>
59 #include <sys/syslog.h>
60 #include <sys/vnode.h>
62 #include <machine/inttypes.h>
69 #include "ffs_extern.h"
70 #include "ufs_extern.h"
73 #include <sys/thread2.h>
77 * These definitions need to be adapted to the system to which
78 * this file is being ported.
81 * malloc types defined for the softdep system.
83 MALLOC_DEFINE(M_PAGEDEP
, "pagedep","File page dependencies");
84 MALLOC_DEFINE(M_INODEDEP
, "inodedep","Inode dependencies");
85 MALLOC_DEFINE(M_NEWBLK
, "newblk","New block allocation");
86 MALLOC_DEFINE(M_BMSAFEMAP
, "bmsafemap","Block or frag allocated from cyl group map");
87 MALLOC_DEFINE(M_ALLOCDIRECT
, "allocdirect","Block or frag dependency for an inode");
88 MALLOC_DEFINE(M_INDIRDEP
, "indirdep","Indirect block dependencies");
89 MALLOC_DEFINE(M_ALLOCINDIR
, "allocindir","Block dependency for an indirect block");
90 MALLOC_DEFINE(M_FREEFRAG
, "freefrag","Previously used frag for an inode");
91 MALLOC_DEFINE(M_FREEBLKS
, "freeblks","Blocks freed from an inode");
92 MALLOC_DEFINE(M_FREEFILE
, "freefile","Inode deallocated");
93 MALLOC_DEFINE(M_DIRADD
, "diradd","New directory entry");
94 MALLOC_DEFINE(M_MKDIR
, "mkdir","New directory");
95 MALLOC_DEFINE(M_DIRREM
, "dirrem","Directory entry deleted");
97 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE)
102 #define D_BMSAFEMAP 3
103 #define D_ALLOCDIRECT 4
105 #define D_ALLOCINDIR 6
112 #define D_LAST D_DIRREM
115 * translate from workitem type to memory type
116 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
118 static struct malloc_type
*memtype
[] = {
134 #define DtoM(type) (memtype[type])
137 * Names of malloc types.
139 #define TYPENAME(type) \
140 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
142 * End system adaptaion definitions.
146 * Internal function prototypes.
148 static void softdep_error(char *, int);
149 static void drain_output(struct vnode
*, int);
150 static int getdirtybuf(struct buf
**, int);
151 static void clear_remove(struct thread
*);
152 static void clear_inodedeps(struct thread
*);
153 static int flush_pagedep_deps(struct vnode
*, struct mount
*,
155 static int flush_inodedep_deps(struct fs
*, ino_t
);
156 static int handle_written_filepage(struct pagedep
*, struct buf
*);
157 static void diradd_inode_written(struct diradd
*, struct inodedep
*);
158 static int handle_written_inodeblock(struct inodedep
*, struct buf
*);
159 static void handle_allocdirect_partdone(struct allocdirect
*);
160 static void handle_allocindir_partdone(struct allocindir
*);
161 static void initiate_write_filepage(struct pagedep
*, struct buf
*);
162 static void handle_written_mkdir(struct mkdir
*, int);
163 static void initiate_write_inodeblock(struct inodedep
*, struct buf
*);
164 static void handle_workitem_freefile(struct freefile
*);
165 static void handle_workitem_remove(struct dirrem
*);
166 static struct dirrem
*newdirrem(struct buf
*, struct inode
*,
167 struct inode
*, int, struct dirrem
**);
168 static void free_diradd(struct diradd
*);
169 static void free_allocindir(struct allocindir
*, struct inodedep
*);
170 static int indir_trunc (struct inode
*, off_t
, int, ufs_lbn_t
, long *);
171 static void deallocate_dependencies(struct buf
*, struct inodedep
*);
172 static void free_allocdirect(struct allocdirectlst
*,
173 struct allocdirect
*, int);
174 static int check_inode_unwritten(struct inodedep
*);
175 static int free_inodedep(struct inodedep
*);
176 static void handle_workitem_freeblocks(struct freeblks
*);
177 static void merge_inode_lists(struct inodedep
*);
178 static void setup_allocindir_phase2(struct buf
*, struct inode
*,
179 struct allocindir
*);
180 static struct allocindir
*newallocindir(struct inode
*, int, ufs_daddr_t
,
182 static void handle_workitem_freefrag(struct freefrag
*);
183 static struct freefrag
*newfreefrag(struct inode
*, ufs_daddr_t
, long);
184 static void allocdirect_merge(struct allocdirectlst
*,
185 struct allocdirect
*, struct allocdirect
*);
186 static struct bmsafemap
*bmsafemap_lookup(struct buf
*);
187 static int newblk_lookup(struct fs
*, ufs_daddr_t
, int,
189 static int inodedep_lookup(struct fs
*, ino_t
, int, struct inodedep
**);
190 static int pagedep_lookup(struct inode
*, ufs_lbn_t
, int,
192 static int request_cleanup(int);
193 static int process_worklist_item(struct mount
*, int);
194 static void add_to_worklist(struct worklist
*);
197 * Exported softdep operations.
199 static void softdep_disk_io_initiation(struct buf
*);
200 static void softdep_disk_write_complete(struct buf
*);
201 static void softdep_deallocate_dependencies(struct buf
*);
202 static int softdep_fsync(struct vnode
*);
203 static int softdep_process_worklist(struct mount
*);
204 static void softdep_move_dependencies(struct buf
*, struct buf
*);
205 static int softdep_count_dependencies(struct buf
*bp
, int);
206 static int softdep_checkread(struct buf
*bp
);
207 static int softdep_checkwrite(struct buf
*bp
);
209 static struct bio_ops softdep_bioops
= {
210 .io_start
= softdep_disk_io_initiation
,
211 .io_complete
= softdep_disk_write_complete
,
212 .io_deallocate
= softdep_deallocate_dependencies
,
213 .io_fsync
= softdep_fsync
,
214 .io_sync
= softdep_process_worklist
,
215 .io_movedeps
= softdep_move_dependencies
,
216 .io_countdeps
= softdep_count_dependencies
,
217 .io_checkread
= softdep_checkread
,
218 .io_checkwrite
= softdep_checkwrite
222 * Locking primitives.
224 static void acquire_lock(struct lock
*);
225 static void free_lock(struct lock
*);
227 static int lock_held(struct lock
*);
230 static struct lock lk
;
232 #define ACQUIRE_LOCK(lkp) acquire_lock(lkp)
233 #define FREE_LOCK(lkp) free_lock(lkp)
236 acquire_lock(struct lock
*lkp
)
238 lockmgr(lkp
, LK_EXCLUSIVE
);
242 free_lock(struct lock
*lkp
)
244 lockmgr(lkp
, LK_RELEASE
);
249 lock_held(struct lock
*lkp
)
251 return lockcountnb(lkp
);
256 * Place holder for real semaphores.
263 struct spinlock spin
;
265 static void sema_init(struct sema
*, char *, int);
266 static int sema_get(struct sema
*, struct lock
*);
267 static void sema_release(struct sema
*, struct lock
*);
269 #define NOHOLDER ((struct thread *) -1)
272 sema_init(struct sema
*semap
, char *name
, int timo
)
274 semap
->holder
= NOHOLDER
;
278 spin_init(&semap
->spin
, "ufssema");
282 * Obtain exclusive access, semaphore is protected by the interlock.
283 * If interlock is NULL we must protect the semaphore ourselves.
286 sema_get(struct sema
*semap
, struct lock
*interlock
)
291 if (semap
->value
> 0) {
292 ++semap
->value
; /* serves as wakeup flag */
293 lksleep(semap
, interlock
, 0,
294 semap
->name
, semap
->timo
);
297 semap
->value
= 1; /* serves as owned flag */
298 semap
->holder
= curthread
;
302 spin_lock(&semap
->spin
);
303 if (semap
->value
> 0) {
304 ++semap
->value
; /* serves as wakeup flag */
305 ssleep(semap
, &semap
->spin
, 0,
306 semap
->name
, semap
->timo
);
307 spin_unlock(&semap
->spin
);
310 semap
->value
= 1; /* serves as owned flag */
311 semap
->holder
= curthread
;
312 spin_unlock(&semap
->spin
);
320 sema_release(struct sema
*semap
, struct lock
*lk
)
322 if (semap
->value
<= 0 || semap
->holder
!= curthread
)
323 panic("sema_release: not held");
325 semap
->holder
= NOHOLDER
;
326 if (--semap
->value
> 0) {
331 spin_lock(&semap
->spin
);
332 semap
->holder
= NOHOLDER
;
333 if (--semap
->value
> 0) {
335 spin_unlock(&semap
->spin
);
338 spin_unlock(&semap
->spin
);
344 * Worklist queue management.
345 * These routines require that the lock be held.
347 static void worklist_insert(struct workhead
*, struct worklist
*);
348 static void worklist_remove(struct worklist
*);
349 static void workitem_free(struct worklist
*, int);
351 #define WORKLIST_INSERT_BP(bp, item) do { \
352 (bp)->b_ops = &softdep_bioops; \
353 worklist_insert(&(bp)->b_dep, item); \
356 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
357 #define WORKLIST_REMOVE(item) worklist_remove(item)
358 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
361 worklist_insert(struct workhead
*head
, struct worklist
*item
)
363 KKASSERT(lock_held(&lk
) > 0);
365 if (item
->wk_state
& ONWORKLIST
) {
366 panic("worklist_insert: already on list");
368 item
->wk_state
|= ONWORKLIST
;
369 LIST_INSERT_HEAD(head
, item
, wk_list
);
373 worklist_remove(struct worklist
*item
)
376 KKASSERT(lock_held(&lk
));
377 if ((item
->wk_state
& ONWORKLIST
) == 0)
378 panic("worklist_remove: not on list");
380 item
->wk_state
&= ~ONWORKLIST
;
381 LIST_REMOVE(item
, wk_list
);
385 workitem_free(struct worklist
*item
, int type
)
388 if (item
->wk_state
& ONWORKLIST
)
389 panic("workitem_free: still on list");
390 if (item
->wk_type
!= type
)
391 panic("workitem_free: type mismatch");
393 kfree(item
, DtoM(type
));
397 * Workitem queue management
399 static struct workhead softdep_workitem_pending
;
400 static int num_on_worklist
; /* number of worklist items to be processed */
401 static int softdep_worklist_busy
; /* 1 => trying to do unmount */
402 static int softdep_worklist_req
; /* serialized waiters */
403 static int max_softdeps
; /* maximum number of structs before slowdown */
404 static int tickdelay
= 2; /* number of ticks to pause during slowdown */
405 static int *stat_countp
; /* statistic to count in proc_waiting timeout */
406 static int proc_waiting
; /* tracks whether we have a timeout posted */
407 static struct thread
*filesys_syncer
; /* proc of filesystem syncer process */
408 static int req_clear_inodedeps
; /* syncer process flush some inodedeps */
409 #define FLUSH_INODES 1
410 static int req_clear_remove
; /* syncer process flush some freeblks */
411 #define FLUSH_REMOVE 2
415 static int stat_worklist_push
; /* number of worklist cleanups */
416 static int stat_blk_limit_push
; /* number of times block limit neared */
417 static int stat_ino_limit_push
; /* number of times inode limit neared */
418 static int stat_blk_limit_hit
; /* number of times block slowdown imposed */
419 static int stat_ino_limit_hit
; /* number of times inode slowdown imposed */
420 static int stat_sync_limit_hit
; /* number of synchronous slowdowns imposed */
421 static int stat_indir_blk_ptrs
; /* bufs redirtied as indir ptrs not written */
422 static int stat_inode_bitmap
; /* bufs redirtied as inode bitmap not written */
423 static int stat_direct_blk_ptrs
;/* bufs redirtied as direct ptrs not written */
424 static int stat_dir_entry
; /* bufs redirtied as dir entry cannot write */
427 #include <sys/sysctl.h>
428 SYSCTL_INT(_debug
, OID_AUTO
, max_softdeps
, CTLFLAG_RW
, &max_softdeps
, 0,
429 "Maximum soft dependencies before slowdown occurs");
430 SYSCTL_INT(_debug
, OID_AUTO
, tickdelay
, CTLFLAG_RW
, &tickdelay
, 0,
431 "Ticks to delay before allocating during slowdown");
432 SYSCTL_INT(_debug
, OID_AUTO
, worklist_push
, CTLFLAG_RW
, &stat_worklist_push
, 0,
433 "Number of worklist cleanups");
434 SYSCTL_INT(_debug
, OID_AUTO
, blk_limit_push
, CTLFLAG_RW
, &stat_blk_limit_push
, 0,
435 "Number of times block limit neared");
436 SYSCTL_INT(_debug
, OID_AUTO
, ino_limit_push
, CTLFLAG_RW
, &stat_ino_limit_push
, 0,
437 "Number of times inode limit neared");
438 SYSCTL_INT(_debug
, OID_AUTO
, blk_limit_hit
, CTLFLAG_RW
, &stat_blk_limit_hit
, 0,
439 "Number of times block slowdown imposed");
440 SYSCTL_INT(_debug
, OID_AUTO
, ino_limit_hit
, CTLFLAG_RW
, &stat_ino_limit_hit
, 0,
441 "Number of times inode slowdown imposed ");
442 SYSCTL_INT(_debug
, OID_AUTO
, sync_limit_hit
, CTLFLAG_RW
, &stat_sync_limit_hit
, 0,
443 "Number of synchronous slowdowns imposed");
444 SYSCTL_INT(_debug
, OID_AUTO
, indir_blk_ptrs
, CTLFLAG_RW
, &stat_indir_blk_ptrs
, 0,
445 "Bufs redirtied as indir ptrs not written");
446 SYSCTL_INT(_debug
, OID_AUTO
, inode_bitmap
, CTLFLAG_RW
, &stat_inode_bitmap
, 0,
447 "Bufs redirtied as inode bitmap not written");
448 SYSCTL_INT(_debug
, OID_AUTO
, direct_blk_ptrs
, CTLFLAG_RW
, &stat_direct_blk_ptrs
, 0,
449 "Bufs redirtied as direct ptrs not written");
450 SYSCTL_INT(_debug
, OID_AUTO
, dir_entry
, CTLFLAG_RW
, &stat_dir_entry
, 0,
451 "Bufs redirtied as dir entry cannot write");
455 * Add an item to the end of the work queue.
456 * This routine requires that the lock be held.
457 * This is the only routine that adds items to the list.
458 * The following routine is the only one that removes items
459 * and does so in order from first to last.
462 add_to_worklist(struct worklist
*wk
)
464 static struct worklist
*worklist_tail
;
466 if (wk
->wk_state
& ONWORKLIST
) {
467 panic("add_to_worklist: already on list");
469 wk
->wk_state
|= ONWORKLIST
;
470 if (LIST_FIRST(&softdep_workitem_pending
) == NULL
)
471 LIST_INSERT_HEAD(&softdep_workitem_pending
, wk
, wk_list
);
473 LIST_INSERT_AFTER(worklist_tail
, wk
, wk_list
);
475 num_on_worklist
+= 1;
479 * Process that runs once per second to handle items in the background queue.
481 * Note that we ensure that everything is done in the order in which they
482 * appear in the queue. The code below depends on this property to ensure
483 * that blocks of a file are freed before the inode itself is freed. This
484 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
485 * until all the old ones have been purged from the dependency lists.
487 * bioops callback - hold io_token
490 softdep_process_worklist(struct mount
*matchmnt
)
492 thread_t td
= curthread
;
493 int matchcnt
, loopcount
;
499 * Record the process identifier of our caller so that we can give
500 * this process preferential treatment in request_cleanup below.
506 * There is no danger of having multiple processes run this
507 * code, but we have to single-thread it when softdep_flushfiles()
508 * is in operation to get an accurate count of the number of items
509 * related to its mount point that are in the list.
511 if (matchmnt
== NULL
) {
512 if (softdep_worklist_busy
< 0) {
516 softdep_worklist_busy
+= 1;
520 * If requested, try removing inode or removal dependencies.
522 if (req_clear_inodedeps
) {
524 req_clear_inodedeps
-= 1;
525 wakeup_one(&proc_waiting
);
527 if (req_clear_remove
) {
529 req_clear_remove
-= 1;
530 wakeup_one(&proc_waiting
);
534 while (num_on_worklist
> 0) {
535 matchcnt
+= process_worklist_item(matchmnt
, 0);
538 * If a umount operation wants to run the worklist
541 if (softdep_worklist_req
&& matchmnt
== NULL
) {
547 * If requested, try removing inode or removal dependencies.
549 if (req_clear_inodedeps
) {
551 req_clear_inodedeps
-= 1;
552 wakeup_one(&proc_waiting
);
554 if (req_clear_remove
) {
556 req_clear_remove
-= 1;
557 wakeup_one(&proc_waiting
);
560 * We do not generally want to stop for buffer space, but if
561 * we are really being a buffer hog, we will stop and wait.
563 if (loopcount
++ % 128 == 0) {
570 * Never allow processing to run for more than one
571 * second. Otherwise the other syncer tasks may get
572 * excessively backlogged.
574 * Use ticks to avoid boundary condition w/time_second or
577 if ((ticks
- starttime
) > hz
&& matchmnt
== NULL
) {
582 if (matchmnt
== NULL
) {
583 --softdep_worklist_busy
;
584 if (softdep_worklist_req
&& softdep_worklist_busy
== 0)
585 wakeup(&softdep_worklist_req
);
593 * Process one item on the worklist.
596 process_worklist_item(struct mount
*matchmnt
, int flags
)
598 struct ufsmount
*ump
;
600 struct dirrem
*dirrem
;
605 KKASSERT(lock_held(&lk
) > 0);
608 if (matchmnt
!= NULL
)
609 matchfs
= VFSTOUFS(matchmnt
)->um_fs
;
612 * Normally we just process each item on the worklist in order.
613 * However, if we are in a situation where we cannot lock any
614 * inodes, we have to skip over any dirrem requests whose
615 * vnodes are resident and locked.
617 LIST_FOREACH(wk
, &softdep_workitem_pending
, wk_list
) {
618 if ((flags
& LK_NOWAIT
) == 0 || wk
->wk_type
!= D_DIRREM
)
620 dirrem
= WK_DIRREM(wk
);
621 ump
= VFSTOUFS(dirrem
->dm_mnt
);
622 lwkt_gettoken(&ump
->um_mountp
->mnt_token
);
623 vp
= ufs_ihashlookup(ump
, ump
->um_dev
, dirrem
->dm_oldinum
);
624 lwkt_reltoken(&ump
->um_mountp
->mnt_token
);
625 if (vp
== NULL
|| !vn_islocked(vp
))
632 num_on_worklist
-= 1;
634 switch (wk
->wk_type
) {
636 /* removal of a directory entry */
637 if (WK_DIRREM(wk
)->dm_mnt
== matchmnt
)
639 handle_workitem_remove(WK_DIRREM(wk
));
643 /* releasing blocks and/or fragments from a file */
644 if (WK_FREEBLKS(wk
)->fb_fs
== matchfs
)
646 handle_workitem_freeblocks(WK_FREEBLKS(wk
));
650 /* releasing a fragment when replaced as a file grows */
651 if (WK_FREEFRAG(wk
)->ff_fs
== matchfs
)
653 handle_workitem_freefrag(WK_FREEFRAG(wk
));
657 /* releasing an inode when its link count drops to 0 */
658 if (WK_FREEFILE(wk
)->fx_fs
== matchfs
)
660 handle_workitem_freefile(WK_FREEFILE(wk
));
664 panic("%s_process_worklist: Unknown type %s",
665 "softdep", TYPENAME(wk
->wk_type
));
673 * Move dependencies from one buffer to another.
675 * bioops callback - hold io_token
678 softdep_move_dependencies(struct buf
*oldbp
, struct buf
*newbp
)
680 struct worklist
*wk
, *wktail
;
682 if (LIST_FIRST(&newbp
->b_dep
) != NULL
)
683 panic("softdep_move_dependencies: need merge code");
686 while ((wk
= LIST_FIRST(&oldbp
->b_dep
)) != NULL
) {
687 LIST_REMOVE(wk
, wk_list
);
689 LIST_INSERT_HEAD(&newbp
->b_dep
, wk
, wk_list
);
691 LIST_INSERT_AFTER(wktail
, wk
, wk_list
);
693 newbp
->b_ops
= &softdep_bioops
;
699 * Purge the work list of all items associated with a particular mount point.
702 softdep_flushfiles(struct mount
*oldmnt
, int flags
)
708 * Await our turn to clear out the queue, then serialize access.
711 while (softdep_worklist_busy
!= 0) {
712 softdep_worklist_req
+= 1;
713 lksleep(&softdep_worklist_req
, &lk
, 0, "softflush", 0);
714 softdep_worklist_req
-= 1;
716 softdep_worklist_busy
= -1;
719 if ((error
= ffs_flushfiles(oldmnt
, flags
)) != 0) {
720 softdep_worklist_busy
= 0;
721 if (softdep_worklist_req
)
722 wakeup(&softdep_worklist_req
);
726 * Alternately flush the block device associated with the mount
727 * point and process any dependencies that the flushing
728 * creates. In theory, this loop can happen at most twice,
729 * but we give it a few extra just to be sure.
731 devvp
= VFSTOUFS(oldmnt
)->um_devvp
;
732 for (loopcnt
= 10; loopcnt
> 0; ) {
733 if (softdep_process_worklist(oldmnt
) == 0) {
736 * Do another flush in case any vnodes were brought in
737 * as part of the cleanup operations.
739 if ((error
= ffs_flushfiles(oldmnt
, flags
)) != 0)
742 * If we still found nothing to do, we are really done.
744 if (softdep_process_worklist(oldmnt
) == 0)
747 vn_lock(devvp
, LK_EXCLUSIVE
| LK_RETRY
);
748 error
= VOP_FSYNC(devvp
, MNT_WAIT
, 0);
754 softdep_worklist_busy
= 0;
755 if (softdep_worklist_req
)
756 wakeup(&softdep_worklist_req
);
760 * If we are unmounting then it is an error to fail. If we
761 * are simply trying to downgrade to read-only, then filesystem
762 * activity can keep us busy forever, so we just fail with EBUSY.
765 if (oldmnt
->mnt_kern_flag
& MNTK_UNMOUNT
)
766 panic("softdep_flushfiles: looping");
775 * There are three types of structures that can be looked up:
776 * 1) pagedep structures identified by mount point, inode number,
778 * 2) inodedep structures identified by mount point and inode number.
779 * 3) newblk structures identified by mount point and
780 * physical block number.
782 * The "pagedep" and "inodedep" dependency structures are hashed
783 * separately from the file blocks and inodes to which they correspond.
784 * This separation helps when the in-memory copy of an inode or
785 * file block must be replaced. It also obviates the need to access
786 * an inode or file page when simply updating (or de-allocating)
787 * dependency structures. Lookup of newblk structures is needed to
788 * find newly allocated blocks when trying to associate them with
789 * their allocdirect or allocindir structure.
791 * The lookup routines optionally create and hash a new instance when
792 * an existing entry is not found.
794 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */
795 #define NODELAY 0x0002 /* cannot do background work */
798 * Structures and routines associated with pagedep caching.
800 LIST_HEAD(pagedep_hashhead
, pagedep
) *pagedep_hashtbl
;
801 u_long pagedep_hash
; /* size of hash table - 1 */
802 #define PAGEDEP_HASH(mp, inum, lbn) \
803 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
805 static struct sema pagedep_in_progress
;
808 * Helper routine for pagedep_lookup()
812 pagedep_find(struct pagedep_hashhead
*pagedephd
, ino_t ino
, ufs_lbn_t lbn
,
815 struct pagedep
*pagedep
;
817 LIST_FOREACH(pagedep
, pagedephd
, pd_hash
) {
818 if (ino
== pagedep
->pd_ino
&&
819 lbn
== pagedep
->pd_lbn
&&
820 mp
== pagedep
->pd_mnt
) {
828 * Look up a pagedep. Return 1 if found, 0 if not found.
829 * If not found, allocate if DEPALLOC flag is passed.
830 * Found or allocated entry is returned in pagedeppp.
831 * This routine must be called with splbio interrupts blocked.
834 pagedep_lookup(struct inode
*ip
, ufs_lbn_t lbn
, int flags
,
835 struct pagedep
**pagedeppp
)
837 struct pagedep
*pagedep
;
838 struct pagedep_hashhead
*pagedephd
;
842 KKASSERT(lock_held(&lk
) > 0);
844 mp
= ITOV(ip
)->v_mount
;
845 pagedephd
= PAGEDEP_HASH(mp
, ip
->i_number
, lbn
);
847 *pagedeppp
= pagedep_find(pagedephd
, ip
->i_number
, lbn
, mp
);
850 if ((flags
& DEPALLOC
) == 0)
852 if (sema_get(&pagedep_in_progress
, &lk
) == 0)
856 pagedep
= kmalloc(sizeof(struct pagedep
), M_PAGEDEP
,
857 M_SOFTDEP_FLAGS
| M_ZERO
);
859 if (pagedep_find(pagedephd
, ip
->i_number
, lbn
, mp
)) {
860 kprintf("pagedep_lookup: blocking race avoided\n");
861 sema_release(&pagedep_in_progress
, &lk
);
862 kfree(pagedep
, M_PAGEDEP
);
866 pagedep
->pd_list
.wk_type
= D_PAGEDEP
;
867 pagedep
->pd_mnt
= mp
;
868 pagedep
->pd_ino
= ip
->i_number
;
869 pagedep
->pd_lbn
= lbn
;
870 LIST_INIT(&pagedep
->pd_dirremhd
);
871 LIST_INIT(&pagedep
->pd_pendinghd
);
872 for (i
= 0; i
< DAHASHSZ
; i
++)
873 LIST_INIT(&pagedep
->pd_diraddhd
[i
]);
874 LIST_INSERT_HEAD(pagedephd
, pagedep
, pd_hash
);
875 sema_release(&pagedep_in_progress
, &lk
);
876 *pagedeppp
= pagedep
;
881 * Structures and routines associated with inodedep caching.
883 LIST_HEAD(inodedep_hashhead
, inodedep
) *inodedep_hashtbl
;
884 static u_long inodedep_hash
; /* size of hash table - 1 */
885 static long num_inodedep
; /* number of inodedep allocated */
886 #define INODEDEP_HASH(fs, inum) \
887 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
888 static struct sema inodedep_in_progress
;
891 * Helper routine for inodedep_lookup()
895 inodedep_find(struct inodedep_hashhead
*inodedephd
, struct fs
*fs
, ino_t inum
)
897 struct inodedep
*inodedep
;
899 LIST_FOREACH(inodedep
, inodedephd
, id_hash
) {
900 if (inum
== inodedep
->id_ino
&& fs
== inodedep
->id_fs
)
907 * Look up a inodedep. Return 1 if found, 0 if not found.
908 * If not found, allocate if DEPALLOC flag is passed.
909 * Found or allocated entry is returned in inodedeppp.
910 * This routine must be called with splbio interrupts blocked.
913 inodedep_lookup(struct fs
*fs
, ino_t inum
, int flags
,
914 struct inodedep
**inodedeppp
)
916 struct inodedep
*inodedep
;
917 struct inodedep_hashhead
*inodedephd
;
919 KKASSERT(lock_held(&lk
) > 0);
921 inodedephd
= INODEDEP_HASH(fs
, inum
);
923 *inodedeppp
= inodedep_find(inodedephd
, fs
, inum
);
926 if ((flags
& DEPALLOC
) == 0)
930 * If we are over our limit, try to improve the situation.
932 if (num_inodedep
> max_softdeps
/ 2)
933 speedup_syncer(NULL
);
934 if (num_inodedep
> max_softdeps
&&
935 (flags
& NODELAY
) == 0 &&
936 request_cleanup(FLUSH_INODES
)) {
939 if (sema_get(&inodedep_in_progress
, &lk
) == 0)
943 inodedep
= kmalloc(sizeof(struct inodedep
), M_INODEDEP
,
944 M_SOFTDEP_FLAGS
| M_ZERO
);
946 if (inodedep_find(inodedephd
, fs
, inum
)) {
947 kprintf("inodedep_lookup: blocking race avoided\n");
948 sema_release(&inodedep_in_progress
, &lk
);
949 kfree(inodedep
, M_INODEDEP
);
952 inodedep
->id_list
.wk_type
= D_INODEDEP
;
953 inodedep
->id_fs
= fs
;
954 inodedep
->id_ino
= inum
;
955 inodedep
->id_state
= ALLCOMPLETE
;
956 inodedep
->id_nlinkdelta
= 0;
957 inodedep
->id_savedino
= NULL
;
958 inodedep
->id_savedsize
= -1;
959 inodedep
->id_buf
= NULL
;
960 LIST_INIT(&inodedep
->id_pendinghd
);
961 LIST_INIT(&inodedep
->id_inowait
);
962 LIST_INIT(&inodedep
->id_bufwait
);
963 TAILQ_INIT(&inodedep
->id_inoupdt
);
964 TAILQ_INIT(&inodedep
->id_newinoupdt
);
966 LIST_INSERT_HEAD(inodedephd
, inodedep
, id_hash
);
967 sema_release(&inodedep_in_progress
, &lk
);
968 *inodedeppp
= inodedep
;
973 * Structures and routines associated with newblk caching.
975 LIST_HEAD(newblk_hashhead
, newblk
) *newblk_hashtbl
;
976 u_long newblk_hash
; /* size of hash table - 1 */
977 #define NEWBLK_HASH(fs, inum) \
978 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
979 static struct sema newblk_in_progress
;
982 * Helper routine for newblk_lookup()
986 newblk_find(struct newblk_hashhead
*newblkhd
, struct fs
*fs
,
987 ufs_daddr_t newblkno
)
989 struct newblk
*newblk
;
991 LIST_FOREACH(newblk
, newblkhd
, nb_hash
) {
992 if (newblkno
== newblk
->nb_newblkno
&& fs
== newblk
->nb_fs
)
999 * Look up a newblk. Return 1 if found, 0 if not found.
1000 * If not found, allocate if DEPALLOC flag is passed.
1001 * Found or allocated entry is returned in newblkpp.
1004 newblk_lookup(struct fs
*fs
, ufs_daddr_t newblkno
, int flags
,
1005 struct newblk
**newblkpp
)
1007 struct newblk
*newblk
;
1008 struct newblk_hashhead
*newblkhd
;
1010 newblkhd
= NEWBLK_HASH(fs
, newblkno
);
1012 *newblkpp
= newblk_find(newblkhd
, fs
, newblkno
);
1015 if ((flags
& DEPALLOC
) == 0)
1017 if (sema_get(&newblk_in_progress
, NULL
) == 0)
1020 newblk
= kmalloc(sizeof(struct newblk
), M_NEWBLK
,
1021 M_SOFTDEP_FLAGS
| M_ZERO
);
1023 if (newblk_find(newblkhd
, fs
, newblkno
)) {
1024 kprintf("newblk_lookup: blocking race avoided\n");
1025 sema_release(&pagedep_in_progress
, NULL
);
1026 kfree(newblk
, M_NEWBLK
);
1029 newblk
->nb_state
= 0;
1031 newblk
->nb_newblkno
= newblkno
;
1032 LIST_INSERT_HEAD(newblkhd
, newblk
, nb_hash
);
1033 sema_release(&newblk_in_progress
, NULL
);
1039 * Executed during filesystem system initialization before
1040 * mounting any filesystems.
1043 softdep_initialize(void)
1045 LIST_INIT(&mkdirlisthd
);
1046 LIST_INIT(&softdep_workitem_pending
);
1047 max_softdeps
= min(maxvnodes
* 8,
1048 M_INODEDEP
->ks_limit
/ (2 * sizeof(struct inodedep
)));
1049 pagedep_hashtbl
= hashinit(maxvnodes
/ 5, M_PAGEDEP
, &pagedep_hash
);
1050 lockinit(&lk
, "ffs_softdep", 0, LK_CANRECURSE
);
1051 sema_init(&pagedep_in_progress
, "pagedep", 0);
1052 inodedep_hashtbl
= hashinit(maxvnodes
, M_INODEDEP
, &inodedep_hash
);
1053 sema_init(&inodedep_in_progress
, "inodedep", 0);
1054 newblk_hashtbl
= hashinit(64, M_NEWBLK
, &newblk_hash
);
1055 sema_init(&newblk_in_progress
, "newblk", 0);
1056 add_bio_ops(&softdep_bioops
);
1060 * Called at mount time to notify the dependency code that a
1061 * filesystem wishes to use it.
1064 softdep_mount(struct vnode
*devvp
, struct mount
*mp
, struct fs
*fs
)
1066 struct csum cstotal
;
1071 mp
->mnt_flag
&= ~MNT_ASYNC
;
1072 mp
->mnt_flag
|= MNT_SOFTDEP
;
1073 mp
->mnt_bioops
= &softdep_bioops
;
1075 * When doing soft updates, the counters in the
1076 * superblock may have gotten out of sync, so we have
1077 * to scan the cylinder groups and recalculate them.
1079 if (fs
->fs_clean
!= 0)
1081 bzero(&cstotal
, sizeof cstotal
);
1082 for (cyl
= 0; cyl
< fs
->fs_ncg
; cyl
++) {
1083 if ((error
= bread(devvp
, fsbtodoff(fs
, cgtod(fs
, cyl
)),
1084 fs
->fs_cgsize
, &bp
)) != 0) {
1088 cgp
= (struct cg
*)bp
->b_data
;
1089 cstotal
.cs_nffree
+= cgp
->cg_cs
.cs_nffree
;
1090 cstotal
.cs_nbfree
+= cgp
->cg_cs
.cs_nbfree
;
1091 cstotal
.cs_nifree
+= cgp
->cg_cs
.cs_nifree
;
1092 cstotal
.cs_ndir
+= cgp
->cg_cs
.cs_ndir
;
1093 fs
->fs_cs(fs
, cyl
) = cgp
->cg_cs
;
1097 if (bcmp(&cstotal
, &fs
->fs_cstotal
, sizeof cstotal
))
1098 kprintf("ffs_mountfs: superblock updated for soft updates\n");
1100 bcopy(&cstotal
, &fs
->fs_cstotal
, sizeof cstotal
);
1105 * Protecting the freemaps (or bitmaps).
1107 * To eliminate the need to execute fsck before mounting a filesystem
1108 * after a power failure, one must (conservatively) guarantee that the
1109 * on-disk copy of the bitmaps never indicate that a live inode or block is
1110 * free. So, when a block or inode is allocated, the bitmap should be
1111 * updated (on disk) before any new pointers. When a block or inode is
1112 * freed, the bitmap should not be updated until all pointers have been
1113 * reset. The latter dependency is handled by the delayed de-allocation
1114 * approach described below for block and inode de-allocation. The former
1115 * dependency is handled by calling the following procedure when a block or
1116 * inode is allocated. When an inode is allocated an "inodedep" is created
1117 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1118 * Each "inodedep" is also inserted into the hash indexing structure so
1119 * that any additional link additions can be made dependent on the inode
1122 * The ufs filesystem maintains a number of free block counts (e.g., per
1123 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1124 * in addition to the bitmaps. These counts are used to improve efficiency
1125 * during allocation and therefore must be consistent with the bitmaps.
1126 * There is no convenient way to guarantee post-crash consistency of these
1127 * counts with simple update ordering, for two main reasons: (1) The counts
1128 * and bitmaps for a single cylinder group block are not in the same disk
1129 * sector. If a disk write is interrupted (e.g., by power failure), one may
1130 * be written and the other not. (2) Some of the counts are located in the
1131 * superblock rather than the cylinder group block. So, we focus our soft
1132 * updates implementation on protecting the bitmaps. When mounting a
1133 * filesystem, we recompute the auxiliary counts from the bitmaps.
1137 * Called just after updating the cylinder group block to allocate an inode.
1140 * bp: buffer for cylgroup block with inode map
1141 * ip: inode related to allocation
1142 * newinum: new inode number being allocated
1145 softdep_setup_inomapdep(struct buf
*bp
, struct inode
*ip
, ino_t newinum
)
1147 struct inodedep
*inodedep
;
1148 struct bmsafemap
*bmsafemap
;
1151 * Create a dependency for the newly allocated inode.
1152 * Panic if it already exists as something is seriously wrong.
1153 * Otherwise add it to the dependency list for the buffer holding
1154 * the cylinder group map from which it was allocated.
1157 if ((inodedep_lookup(ip
->i_fs
, newinum
, DEPALLOC
|NODELAY
, &inodedep
))) {
1158 panic("softdep_setup_inomapdep: found inode");
1160 inodedep
->id_buf
= bp
;
1161 inodedep
->id_state
&= ~DEPCOMPLETE
;
1162 bmsafemap
= bmsafemap_lookup(bp
);
1163 LIST_INSERT_HEAD(&bmsafemap
->sm_inodedephd
, inodedep
, id_deps
);
1168 * Called just after updating the cylinder group block to
1169 * allocate block or fragment.
1172 * bp: buffer for cylgroup block with block map
1173 * fs: filesystem doing allocation
1174 * newblkno: number of newly allocated block
1177 softdep_setup_blkmapdep(struct buf
*bp
, struct fs
*fs
,
1178 ufs_daddr_t newblkno
)
1180 struct newblk
*newblk
;
1181 struct bmsafemap
*bmsafemap
;
1184 * Create a dependency for the newly allocated block.
1185 * Add it to the dependency list for the buffer holding
1186 * the cylinder group map from which it was allocated.
1188 if (newblk_lookup(fs
, newblkno
, DEPALLOC
, &newblk
) != 0)
1189 panic("softdep_setup_blkmapdep: found block");
1191 newblk
->nb_bmsafemap
= bmsafemap
= bmsafemap_lookup(bp
);
1192 LIST_INSERT_HEAD(&bmsafemap
->sm_newblkhd
, newblk
, nb_deps
);
1197 * Find the bmsafemap associated with a cylinder group buffer.
1198 * If none exists, create one. The buffer must be locked when
1199 * this routine is called and this routine must be called with
1200 * splbio interrupts blocked.
1202 static struct bmsafemap
*
1203 bmsafemap_lookup(struct buf
*bp
)
1205 struct bmsafemap
*bmsafemap
;
1206 struct worklist
*wk
;
1208 KKASSERT(lock_held(&lk
) > 0);
1210 LIST_FOREACH(wk
, &bp
->b_dep
, wk_list
) {
1211 if (wk
->wk_type
== D_BMSAFEMAP
)
1212 return (WK_BMSAFEMAP(wk
));
1215 bmsafemap
= kmalloc(sizeof(struct bmsafemap
), M_BMSAFEMAP
,
1217 bmsafemap
->sm_list
.wk_type
= D_BMSAFEMAP
;
1218 bmsafemap
->sm_list
.wk_state
= 0;
1219 bmsafemap
->sm_buf
= bp
;
1220 LIST_INIT(&bmsafemap
->sm_allocdirecthd
);
1221 LIST_INIT(&bmsafemap
->sm_allocindirhd
);
1222 LIST_INIT(&bmsafemap
->sm_inodedephd
);
1223 LIST_INIT(&bmsafemap
->sm_newblkhd
);
1225 WORKLIST_INSERT_BP(bp
, &bmsafemap
->sm_list
);
1230 * Direct block allocation dependencies.
1232 * When a new block is allocated, the corresponding disk locations must be
1233 * initialized (with zeros or new data) before the on-disk inode points to
1234 * them. Also, the freemap from which the block was allocated must be
1235 * updated (on disk) before the inode's pointer. These two dependencies are
1236 * independent of each other and are needed for all file blocks and indirect
1237 * blocks that are pointed to directly by the inode. Just before the
1238 * "in-core" version of the inode is updated with a newly allocated block
1239 * number, a procedure (below) is called to setup allocation dependency
1240 * structures. These structures are removed when the corresponding
1241 * dependencies are satisfied or when the block allocation becomes obsolete
1242 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1243 * fragment that gets upgraded). All of these cases are handled in
1244 * procedures described later.
1246 * When a file extension causes a fragment to be upgraded, either to a larger
1247 * fragment or to a full block, the on-disk location may change (if the
1248 * previous fragment could not simply be extended). In this case, the old
1249 * fragment must be de-allocated, but not until after the inode's pointer has
1250 * been updated. In most cases, this is handled by later procedures, which
1251 * will construct a "freefrag" structure to be added to the workitem queue
1252 * when the inode update is complete (or obsolete). The main exception to
1253 * this is when an allocation occurs while a pending allocation dependency
1254 * (for the same block pointer) remains. This case is handled in the main
1255 * allocation dependency setup procedure by immediately freeing the
1256 * unreferenced fragments.
1259 * ip: inode to which block is being added
1260 * lbn: block pointer within inode
1261 * newblkno: disk block number being added
1262 * oldblkno: previous block number, 0 unless frag
1263 * newsize: size of new block
1264 * oldsize: size of new block
1265 * bp: bp for allocated block
1268 softdep_setup_allocdirect(struct inode
*ip
, ufs_lbn_t lbn
, ufs_daddr_t newblkno
,
1269 ufs_daddr_t oldblkno
, long newsize
, long oldsize
,
1272 struct allocdirect
*adp
, *oldadp
;
1273 struct allocdirectlst
*adphead
;
1274 struct bmsafemap
*bmsafemap
;
1275 struct inodedep
*inodedep
;
1276 struct pagedep
*pagedep
;
1277 struct newblk
*newblk
;
1279 adp
= kmalloc(sizeof(struct allocdirect
), M_ALLOCDIRECT
,
1280 M_SOFTDEP_FLAGS
| M_ZERO
);
1281 adp
->ad_list
.wk_type
= D_ALLOCDIRECT
;
1283 adp
->ad_newblkno
= newblkno
;
1284 adp
->ad_oldblkno
= oldblkno
;
1285 adp
->ad_newsize
= newsize
;
1286 adp
->ad_oldsize
= oldsize
;
1287 adp
->ad_state
= ATTACHED
;
1288 if (newblkno
== oldblkno
)
1289 adp
->ad_freefrag
= NULL
;
1291 adp
->ad_freefrag
= newfreefrag(ip
, oldblkno
, oldsize
);
1293 if (newblk_lookup(ip
->i_fs
, newblkno
, 0, &newblk
) == 0)
1294 panic("softdep_setup_allocdirect: lost block");
1297 inodedep_lookup(ip
->i_fs
, ip
->i_number
, DEPALLOC
| NODELAY
, &inodedep
);
1298 adp
->ad_inodedep
= inodedep
;
1300 if (newblk
->nb_state
== DEPCOMPLETE
) {
1301 adp
->ad_state
|= DEPCOMPLETE
;
1304 bmsafemap
= newblk
->nb_bmsafemap
;
1305 adp
->ad_buf
= bmsafemap
->sm_buf
;
1306 LIST_REMOVE(newblk
, nb_deps
);
1307 LIST_INSERT_HEAD(&bmsafemap
->sm_allocdirecthd
, adp
, ad_deps
);
1309 LIST_REMOVE(newblk
, nb_hash
);
1310 kfree(newblk
, M_NEWBLK
);
1312 WORKLIST_INSERT_BP(bp
, &adp
->ad_list
);
1313 if (lbn
>= NDADDR
) {
1314 /* allocating an indirect block */
1315 if (oldblkno
!= 0) {
1316 panic("softdep_setup_allocdirect: non-zero indir");
1320 * Allocating a direct block.
1322 * If we are allocating a directory block, then we must
1323 * allocate an associated pagedep to track additions and
1326 if ((ip
->i_mode
& IFMT
) == IFDIR
&&
1327 pagedep_lookup(ip
, lbn
, DEPALLOC
, &pagedep
) == 0) {
1328 WORKLIST_INSERT_BP(bp
, &pagedep
->pd_list
);
1332 * The list of allocdirects must be kept in sorted and ascending
1333 * order so that the rollback routines can quickly determine the
1334 * first uncommitted block (the size of the file stored on disk
1335 * ends at the end of the lowest committed fragment, or if there
1336 * are no fragments, at the end of the highest committed block).
1337 * Since files generally grow, the typical case is that the new
1338 * block is to be added at the end of the list. We speed this
1339 * special case by checking against the last allocdirect in the
1340 * list before laboriously traversing the list looking for the
1343 adphead
= &inodedep
->id_newinoupdt
;
1344 oldadp
= TAILQ_LAST(adphead
, allocdirectlst
);
1345 if (oldadp
== NULL
|| oldadp
->ad_lbn
<= lbn
) {
1346 /* insert at end of list */
1347 TAILQ_INSERT_TAIL(adphead
, adp
, ad_next
);
1348 if (oldadp
!= NULL
&& oldadp
->ad_lbn
== lbn
)
1349 allocdirect_merge(adphead
, adp
, oldadp
);
1353 TAILQ_FOREACH(oldadp
, adphead
, ad_next
) {
1354 if (oldadp
->ad_lbn
>= lbn
)
1357 if (oldadp
== NULL
) {
1358 panic("softdep_setup_allocdirect: lost entry");
1360 /* insert in middle of list */
1361 TAILQ_INSERT_BEFORE(oldadp
, adp
, ad_next
);
1362 if (oldadp
->ad_lbn
== lbn
)
1363 allocdirect_merge(adphead
, adp
, oldadp
);
1368 * Replace an old allocdirect dependency with a newer one.
1369 * This routine must be called with splbio interrupts blocked.
1372 * adphead: head of list holding allocdirects
1373 * newadp: allocdirect being added
1374 * oldadp: existing allocdirect being checked
1377 allocdirect_merge(struct allocdirectlst
*adphead
,
1378 struct allocdirect
*newadp
,
1379 struct allocdirect
*oldadp
)
1381 struct freefrag
*freefrag
;
1383 KKASSERT(lock_held(&lk
) > 0);
1385 if (newadp
->ad_oldblkno
!= oldadp
->ad_newblkno
||
1386 newadp
->ad_oldsize
!= oldadp
->ad_newsize
||
1387 newadp
->ad_lbn
>= NDADDR
) {
1388 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1389 newadp
->ad_oldblkno
, oldadp
->ad_newblkno
, newadp
->ad_lbn
,
1392 newadp
->ad_oldblkno
= oldadp
->ad_oldblkno
;
1393 newadp
->ad_oldsize
= oldadp
->ad_oldsize
;
1395 * If the old dependency had a fragment to free or had never
1396 * previously had a block allocated, then the new dependency
1397 * can immediately post its freefrag and adopt the old freefrag.
1398 * This action is done by swapping the freefrag dependencies.
1399 * The new dependency gains the old one's freefrag, and the
1400 * old one gets the new one and then immediately puts it on
1401 * the worklist when it is freed by free_allocdirect. It is
1402 * not possible to do this swap when the old dependency had a
1403 * non-zero size but no previous fragment to free. This condition
1404 * arises when the new block is an extension of the old block.
1405 * Here, the first part of the fragment allocated to the new
1406 * dependency is part of the block currently claimed on disk by
1407 * the old dependency, so cannot legitimately be freed until the
1408 * conditions for the new dependency are fulfilled.
1410 if (oldadp
->ad_freefrag
!= NULL
|| oldadp
->ad_oldblkno
== 0) {
1411 freefrag
= newadp
->ad_freefrag
;
1412 newadp
->ad_freefrag
= oldadp
->ad_freefrag
;
1413 oldadp
->ad_freefrag
= freefrag
;
1415 free_allocdirect(adphead
, oldadp
, 0);
1419 * Allocate a new freefrag structure if needed.
1421 static struct freefrag
*
1422 newfreefrag(struct inode
*ip
, ufs_daddr_t blkno
, long size
)
1424 struct freefrag
*freefrag
;
1430 if (fragnum(fs
, blkno
) + numfrags(fs
, size
) > fs
->fs_frag
)
1431 panic("newfreefrag: frag size");
1432 freefrag
= kmalloc(sizeof(struct freefrag
), M_FREEFRAG
,
1434 freefrag
->ff_list
.wk_type
= D_FREEFRAG
;
1435 freefrag
->ff_state
= ip
->i_uid
& ~ONWORKLIST
; /* XXX - used below */
1436 freefrag
->ff_inum
= ip
->i_number
;
1437 freefrag
->ff_fs
= fs
;
1438 freefrag
->ff_devvp
= ip
->i_devvp
;
1439 freefrag
->ff_blkno
= blkno
;
1440 freefrag
->ff_fragsize
= size
;
1445 * This workitem de-allocates fragments that were replaced during
1446 * file block allocation.
1449 handle_workitem_freefrag(struct freefrag
*freefrag
)
1453 tip
.i_fs
= freefrag
->ff_fs
;
1454 tip
.i_devvp
= freefrag
->ff_devvp
;
1455 tip
.i_dev
= freefrag
->ff_devvp
->v_rdev
;
1456 tip
.i_number
= freefrag
->ff_inum
;
1457 tip
.i_uid
= freefrag
->ff_state
& ~ONWORKLIST
; /* XXX - set above */
1458 ffs_blkfree(&tip
, freefrag
->ff_blkno
, freefrag
->ff_fragsize
);
1459 kfree(freefrag
, M_FREEFRAG
);
1463 * Indirect block allocation dependencies.
1465 * The same dependencies that exist for a direct block also exist when
1466 * a new block is allocated and pointed to by an entry in a block of
1467 * indirect pointers. The undo/redo states described above are also
1468 * used here. Because an indirect block contains many pointers that
1469 * may have dependencies, a second copy of the entire in-memory indirect
1470 * block is kept. The buffer cache copy is always completely up-to-date.
1471 * The second copy, which is used only as a source for disk writes,
1472 * contains only the safe pointers (i.e., those that have no remaining
1473 * update dependencies). The second copy is freed when all pointers
1474 * are safe. The cache is not allowed to replace indirect blocks with
1475 * pending update dependencies. If a buffer containing an indirect
1476 * block with dependencies is written, these routines will mark it
1477 * dirty again. It can only be successfully written once all the
1478 * dependencies are removed. The ffs_fsync routine in conjunction with
1479 * softdep_sync_metadata work together to get all the dependencies
1480 * removed so that a file can be successfully written to disk. Three
1481 * procedures are used when setting up indirect block pointer
1482 * dependencies. The division is necessary because of the organization
1483 * of the "balloc" routine and because of the distinction between file
1484 * pages and file metadata blocks.
1488 * Allocate a new allocindir structure.
1491 * ip: inode for file being extended
1492 * ptrno: offset of pointer in indirect block
1493 * newblkno: disk block number being added
1494 * oldblkno: previous block number, 0 if none
1496 static struct allocindir
*
1497 newallocindir(struct inode
*ip
, int ptrno
, ufs_daddr_t newblkno
,
1498 ufs_daddr_t oldblkno
)
1500 struct allocindir
*aip
;
1502 aip
= kmalloc(sizeof(struct allocindir
), M_ALLOCINDIR
,
1503 M_SOFTDEP_FLAGS
| M_ZERO
);
1504 aip
->ai_list
.wk_type
= D_ALLOCINDIR
;
1505 aip
->ai_state
= ATTACHED
;
1506 aip
->ai_offset
= ptrno
;
1507 aip
->ai_newblkno
= newblkno
;
1508 aip
->ai_oldblkno
= oldblkno
;
1509 aip
->ai_freefrag
= newfreefrag(ip
, oldblkno
, ip
->i_fs
->fs_bsize
);
1514 * Called just before setting an indirect block pointer
1515 * to a newly allocated file page.
1518 * ip: inode for file being extended
1519 * lbn: allocated block number within file
1520 * bp: buffer with indirect blk referencing page
1521 * ptrno: offset of pointer in indirect block
1522 * newblkno: disk block number being added
1523 * oldblkno: previous block number, 0 if none
1524 * nbp: buffer holding allocated page
1527 softdep_setup_allocindir_page(struct inode
*ip
, ufs_lbn_t lbn
,
1528 struct buf
*bp
, int ptrno
,
1529 ufs_daddr_t newblkno
, ufs_daddr_t oldblkno
,
1532 struct allocindir
*aip
;
1533 struct pagedep
*pagedep
;
1535 aip
= newallocindir(ip
, ptrno
, newblkno
, oldblkno
);
1538 * If we are allocating a directory page, then we must
1539 * allocate an associated pagedep to track additions and
1542 if ((ip
->i_mode
& IFMT
) == IFDIR
&&
1543 pagedep_lookup(ip
, lbn
, DEPALLOC
, &pagedep
) == 0)
1544 WORKLIST_INSERT_BP(nbp
, &pagedep
->pd_list
);
1545 WORKLIST_INSERT_BP(nbp
, &aip
->ai_list
);
1547 setup_allocindir_phase2(bp
, ip
, aip
);
1551 * Called just before setting an indirect block pointer to a
1552 * newly allocated indirect block.
1554 * nbp: newly allocated indirect block
1555 * ip: inode for file being extended
1556 * bp: indirect block referencing allocated block
1557 * ptrno: offset of pointer in indirect block
1558 * newblkno: disk block number being added
1561 softdep_setup_allocindir_meta(struct buf
*nbp
, struct inode
*ip
,
1562 struct buf
*bp
, int ptrno
,
1563 ufs_daddr_t newblkno
)
1565 struct allocindir
*aip
;
1567 aip
= newallocindir(ip
, ptrno
, newblkno
, 0);
1569 WORKLIST_INSERT_BP(nbp
, &aip
->ai_list
);
1571 setup_allocindir_phase2(bp
, ip
, aip
);
1575 * Called to finish the allocation of the "aip" allocated
1576 * by one of the two routines above.
1579 * bp: in-memory copy of the indirect block
1580 * ip: inode for file being extended
1581 * aip: allocindir allocated by the above routines
1584 setup_allocindir_phase2(struct buf
*bp
, struct inode
*ip
,
1585 struct allocindir
*aip
)
1587 struct worklist
*wk
;
1588 struct indirdep
*indirdep
, *newindirdep
;
1589 struct bmsafemap
*bmsafemap
;
1590 struct allocindir
*oldaip
;
1591 struct freefrag
*freefrag
;
1592 struct newblk
*newblk
;
1594 if (bp
->b_loffset
>= 0)
1595 panic("setup_allocindir_phase2: not indir blk");
1596 for (indirdep
= NULL
, newindirdep
= NULL
; ; ) {
1598 LIST_FOREACH(wk
, &bp
->b_dep
, wk_list
) {
1599 if (wk
->wk_type
!= D_INDIRDEP
)
1601 indirdep
= WK_INDIRDEP(wk
);
1604 if (indirdep
== NULL
&& newindirdep
) {
1605 indirdep
= newindirdep
;
1606 WORKLIST_INSERT_BP(bp
, &indirdep
->ir_list
);
1611 if (newblk_lookup(ip
->i_fs
, aip
->ai_newblkno
, 0,
1613 panic("setup_allocindir: lost block");
1615 if (newblk
->nb_state
== DEPCOMPLETE
) {
1616 aip
->ai_state
|= DEPCOMPLETE
;
1619 bmsafemap
= newblk
->nb_bmsafemap
;
1620 aip
->ai_buf
= bmsafemap
->sm_buf
;
1621 LIST_REMOVE(newblk
, nb_deps
);
1622 LIST_INSERT_HEAD(&bmsafemap
->sm_allocindirhd
,
1625 LIST_REMOVE(newblk
, nb_hash
);
1626 kfree(newblk
, M_NEWBLK
);
1627 aip
->ai_indirdep
= indirdep
;
1629 * Check to see if there is an existing dependency
1630 * for this block. If there is, merge the old
1631 * dependency into the new one.
1633 if (aip
->ai_oldblkno
== 0)
1637 LIST_FOREACH(oldaip
, &indirdep
->ir_deplisthd
, ai_next
)
1638 if (oldaip
->ai_offset
== aip
->ai_offset
)
1640 if (oldaip
!= NULL
) {
1641 if (oldaip
->ai_newblkno
!= aip
->ai_oldblkno
) {
1642 panic("setup_allocindir_phase2: blkno");
1644 aip
->ai_oldblkno
= oldaip
->ai_oldblkno
;
1645 freefrag
= oldaip
->ai_freefrag
;
1646 oldaip
->ai_freefrag
= aip
->ai_freefrag
;
1647 aip
->ai_freefrag
= freefrag
;
1648 free_allocindir(oldaip
, NULL
);
1650 LIST_INSERT_HEAD(&indirdep
->ir_deplisthd
, aip
, ai_next
);
1651 ((ufs_daddr_t
*)indirdep
->ir_savebp
->b_data
)
1652 [aip
->ai_offset
] = aip
->ai_oldblkno
;
1657 * Avoid any possibility of data corruption by
1658 * ensuring that our old version is thrown away.
1660 newindirdep
->ir_savebp
->b_flags
|= B_INVAL
| B_NOCACHE
;
1661 brelse(newindirdep
->ir_savebp
);
1662 WORKITEM_FREE((caddr_t
)newindirdep
, D_INDIRDEP
);
1666 newindirdep
= kmalloc(sizeof(struct indirdep
), M_INDIRDEP
,
1668 newindirdep
->ir_list
.wk_type
= D_INDIRDEP
;
1669 newindirdep
->ir_state
= ATTACHED
;
1670 LIST_INIT(&newindirdep
->ir_deplisthd
);
1671 LIST_INIT(&newindirdep
->ir_donehd
);
1672 if (bp
->b_bio2
.bio_offset
== NOOFFSET
) {
1673 VOP_BMAP(bp
->b_vp
, bp
->b_bio1
.bio_offset
,
1674 &bp
->b_bio2
.bio_offset
, NULL
, NULL
,
1677 KKASSERT(bp
->b_bio2
.bio_offset
!= NOOFFSET
);
1678 newindirdep
->ir_savebp
= getblk(ip
->i_devvp
,
1679 bp
->b_bio2
.bio_offset
,
1680 bp
->b_bcount
, 0, 0);
1681 BUF_KERNPROC(newindirdep
->ir_savebp
);
1682 bcopy(bp
->b_data
, newindirdep
->ir_savebp
->b_data
, bp
->b_bcount
);
1687 * Block de-allocation dependencies.
1689 * When blocks are de-allocated, the on-disk pointers must be nullified before
1690 * the blocks are made available for use by other files. (The true
1691 * requirement is that old pointers must be nullified before new on-disk
1692 * pointers are set. We chose this slightly more stringent requirement to
1693 * reduce complexity.) Our implementation handles this dependency by updating
1694 * the inode (or indirect block) appropriately but delaying the actual block
1695 * de-allocation (i.e., freemap and free space count manipulation) until
1696 * after the updated versions reach stable storage. After the disk is
1697 * updated, the blocks can be safely de-allocated whenever it is convenient.
1698 * This implementation handles only the common case of reducing a file's
1699 * length to zero. Other cases are handled by the conventional synchronous
1702 * The ffs implementation with which we worked double-checks
1703 * the state of the block pointers and file size as it reduces
1704 * a file's length. Some of this code is replicated here in our
1705 * soft updates implementation. The freeblks->fb_chkcnt field is
1706 * used to transfer a part of this information to the procedure
1707 * that eventually de-allocates the blocks.
1709 * This routine should be called from the routine that shortens
1710 * a file's length, before the inode's size or block pointers
1711 * are modified. It will save the block pointer information for
1712 * later release and zero the inode so that the calling routine
1715 struct softdep_setup_freeblocks_info
{
1720 static int softdep_setup_freeblocks_bp(struct buf
*bp
, void *data
);
1724 * ip: The inode whose length is to be reduced
1725 * length: The new length for the file
1728 softdep_setup_freeblocks(struct inode
*ip
, off_t length
)
1730 struct softdep_setup_freeblocks_info info
;
1731 struct freeblks
*freeblks
;
1732 struct inodedep
*inodedep
;
1733 struct allocdirect
*adp
;
1737 int i
, error
, delay
;
1742 panic("softde_setup_freeblocks: non-zero length");
1743 freeblks
= kmalloc(sizeof(struct freeblks
), M_FREEBLKS
,
1744 M_SOFTDEP_FLAGS
| M_ZERO
);
1745 freeblks
->fb_list
.wk_type
= D_FREEBLKS
;
1746 freeblks
->fb_state
= ATTACHED
;
1747 freeblks
->fb_uid
= ip
->i_uid
;
1748 freeblks
->fb_previousinum
= ip
->i_number
;
1749 freeblks
->fb_devvp
= ip
->i_devvp
;
1750 freeblks
->fb_fs
= fs
;
1751 freeblks
->fb_oldsize
= ip
->i_size
;
1752 freeblks
->fb_newsize
= length
;
1753 freeblks
->fb_chkcnt
= ip
->i_blocks
;
1754 for (i
= 0; i
< NDADDR
; i
++) {
1755 freeblks
->fb_dblks
[i
] = ip
->i_db
[i
];
1758 for (i
= 0; i
< NIADDR
; i
++) {
1759 freeblks
->fb_iblks
[i
] = ip
->i_ib
[i
];
1765 * Push the zero'ed inode to to its disk buffer so that we are free
1766 * to delete its dependencies below. Once the dependencies are gone
1767 * the buffer can be safely released.
1769 if ((error
= bread(ip
->i_devvp
,
1770 fsbtodoff(fs
, ino_to_fsba(fs
, ip
->i_number
)),
1771 (int)fs
->fs_bsize
, &bp
)) != 0)
1772 softdep_error("softdep_setup_freeblocks", error
);
1773 *((struct ufs1_dinode
*)bp
->b_data
+ ino_to_fsbo(fs
, ip
->i_number
)) =
1776 * Find and eliminate any inode dependencies.
1779 (void) inodedep_lookup(fs
, ip
->i_number
, DEPALLOC
, &inodedep
);
1780 if ((inodedep
->id_state
& IOSTARTED
) != 0) {
1781 panic("softdep_setup_freeblocks: inode busy");
1784 * Add the freeblks structure to the list of operations that
1785 * must await the zero'ed inode being written to disk. If we
1786 * still have a bitmap dependency (delay == 0), then the inode
1787 * has never been written to disk, so we can process the
1788 * freeblks below once we have deleted the dependencies.
1790 delay
= (inodedep
->id_state
& DEPCOMPLETE
);
1792 WORKLIST_INSERT(&inodedep
->id_bufwait
, &freeblks
->fb_list
);
1794 * Because the file length has been truncated to zero, any
1795 * pending block allocation dependency structures associated
1796 * with this inode are obsolete and can simply be de-allocated.
1797 * We must first merge the two dependency lists to get rid of
1798 * any duplicate freefrag structures, then purge the merged list.
1800 merge_inode_lists(inodedep
);
1801 while ((adp
= TAILQ_FIRST(&inodedep
->id_inoupdt
)) != NULL
)
1802 free_allocdirect(&inodedep
->id_inoupdt
, adp
, 1);
1806 * We must wait for any I/O in progress to finish so that
1807 * all potential buffers on the dirty list will be visible.
1808 * Once they are all there, walk the list and get rid of
1813 drain_output(vp
, 1);
1817 lwkt_gettoken(&vp
->v_token
);
1819 count
= RB_SCAN(buf_rb_tree
, &vp
->v_rbdirty_tree
, NULL
,
1820 softdep_setup_freeblocks_bp
, &info
);
1821 } while (count
!= 0);
1822 lwkt_reltoken(&vp
->v_token
);
1824 if (inodedep_lookup(fs
, ip
->i_number
, 0, &inodedep
) != 0)
1825 (void)free_inodedep(inodedep
);
1828 freeblks
->fb_state
|= DEPCOMPLETE
;
1830 * If the inode with zeroed block pointers is now on disk
1831 * we can start freeing blocks. Add freeblks to the worklist
1832 * instead of calling handle_workitem_freeblocks directly as
1833 * it is more likely that additional IO is needed to complete
1834 * the request here than in the !delay case.
1836 if ((freeblks
->fb_state
& ALLCOMPLETE
) == ALLCOMPLETE
)
1837 add_to_worklist(&freeblks
->fb_list
);
1842 * If the inode has never been written to disk (delay == 0),
1843 * then we can process the freeblks now that we have deleted
1847 handle_workitem_freeblocks(freeblks
);
1851 softdep_setup_freeblocks_bp(struct buf
*bp
, void *data
)
1853 struct softdep_setup_freeblocks_info
*info
= data
;
1854 struct inodedep
*inodedep
;
1856 if (getdirtybuf(&bp
, MNT_WAIT
) == 0) {
1857 kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp
);
1860 if (bp
->b_vp
!= ITOV(info
->ip
) || (bp
->b_flags
& B_DELWRI
) == 0) {
1861 kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp
);
1865 (void) inodedep_lookup(info
->fs
, info
->ip
->i_number
, 0, &inodedep
);
1866 deallocate_dependencies(bp
, inodedep
);
1867 bp
->b_flags
|= B_INVAL
| B_NOCACHE
;
1875 * Reclaim any dependency structures from a buffer that is about to
1876 * be reallocated to a new vnode. The buffer must be locked, thus,
1877 * no I/O completion operations can occur while we are manipulating
1878 * its associated dependencies. The mutex is held so that other I/O's
1879 * associated with related dependencies do not occur.
1882 deallocate_dependencies(struct buf
*bp
, struct inodedep
*inodedep
)
1884 struct worklist
*wk
;
1885 struct indirdep
*indirdep
;
1886 struct allocindir
*aip
;
1887 struct pagedep
*pagedep
;
1888 struct dirrem
*dirrem
;
1892 while ((wk
= LIST_FIRST(&bp
->b_dep
)) != NULL
) {
1893 switch (wk
->wk_type
) {
1896 indirdep
= WK_INDIRDEP(wk
);
1898 * None of the indirect pointers will ever be visible,
1899 * so they can simply be tossed. GOINGAWAY ensures
1900 * that allocated pointers will be saved in the buffer
1901 * cache until they are freed. Note that they will
1902 * only be able to be found by their physical address
1903 * since the inode mapping the logical address will
1904 * be gone. The save buffer used for the safe copy
1905 * was allocated in setup_allocindir_phase2 using
1906 * the physical address so it could be used for this
1907 * purpose. Hence we swap the safe copy with the real
1908 * copy, allowing the safe copy to be freed and holding
1909 * on to the real copy for later use in indir_trunc.
1911 * NOTE: ir_savebp is relative to the block device
1912 * so b_bio1 contains the device block number.
1914 if (indirdep
->ir_state
& GOINGAWAY
) {
1915 panic("deallocate_dependencies: already gone");
1917 indirdep
->ir_state
|= GOINGAWAY
;
1918 while ((aip
= LIST_FIRST(&indirdep
->ir_deplisthd
)) != NULL
)
1919 free_allocindir(aip
, inodedep
);
1920 if (bp
->b_bio1
.bio_offset
>= 0 ||
1921 bp
->b_bio2
.bio_offset
!= indirdep
->ir_savebp
->b_bio1
.bio_offset
) {
1922 panic("deallocate_dependencies: not indir");
1924 bcopy(bp
->b_data
, indirdep
->ir_savebp
->b_data
,
1926 WORKLIST_REMOVE(wk
);
1927 WORKLIST_INSERT_BP(indirdep
->ir_savebp
, wk
);
1931 pagedep
= WK_PAGEDEP(wk
);
1933 * None of the directory additions will ever be
1934 * visible, so they can simply be tossed.
1936 for (i
= 0; i
< DAHASHSZ
; i
++)
1938 LIST_FIRST(&pagedep
->pd_diraddhd
[i
])))
1940 while ((dap
= LIST_FIRST(&pagedep
->pd_pendinghd
)) != NULL
)
1943 * Copy any directory remove dependencies to the list
1944 * to be processed after the zero'ed inode is written.
1945 * If the inode has already been written, then they
1946 * can be dumped directly onto the work list.
1948 LIST_FOREACH(dirrem
, &pagedep
->pd_dirremhd
, dm_next
) {
1949 LIST_REMOVE(dirrem
, dm_next
);
1950 dirrem
->dm_dirinum
= pagedep
->pd_ino
;
1951 if (inodedep
== NULL
||
1952 (inodedep
->id_state
& ALLCOMPLETE
) ==
1954 add_to_worklist(&dirrem
->dm_list
);
1956 WORKLIST_INSERT(&inodedep
->id_bufwait
,
1959 WORKLIST_REMOVE(&pagedep
->pd_list
);
1960 LIST_REMOVE(pagedep
, pd_hash
);
1961 WORKITEM_FREE(pagedep
, D_PAGEDEP
);
1965 free_allocindir(WK_ALLOCINDIR(wk
), inodedep
);
1970 panic("deallocate_dependencies: Unexpected type %s",
1971 TYPENAME(wk
->wk_type
));
1975 panic("deallocate_dependencies: Unknown type %s",
1976 TYPENAME(wk
->wk_type
));
1983 * Free an allocdirect. Generate a new freefrag work request if appropriate.
1984 * This routine must be called with splbio interrupts blocked.
1987 free_allocdirect(struct allocdirectlst
*adphead
,
1988 struct allocdirect
*adp
, int delay
)
1990 KKASSERT(lock_held(&lk
) > 0);
1992 if ((adp
->ad_state
& DEPCOMPLETE
) == 0)
1993 LIST_REMOVE(adp
, ad_deps
);
1994 TAILQ_REMOVE(adphead
, adp
, ad_next
);
1995 if ((adp
->ad_state
& COMPLETE
) == 0)
1996 WORKLIST_REMOVE(&adp
->ad_list
);
1997 if (adp
->ad_freefrag
!= NULL
) {
1999 WORKLIST_INSERT(&adp
->ad_inodedep
->id_bufwait
,
2000 &adp
->ad_freefrag
->ff_list
);
2002 add_to_worklist(&adp
->ad_freefrag
->ff_list
);
2004 WORKITEM_FREE(adp
, D_ALLOCDIRECT
);
2008 * Prepare an inode to be freed. The actual free operation is not
2009 * done until the zero'ed inode has been written to disk.
2012 softdep_freefile(struct vnode
*pvp
, ino_t ino
, int mode
)
2014 struct inode
*ip
= VTOI(pvp
);
2015 struct inodedep
*inodedep
;
2016 struct freefile
*freefile
;
2019 * This sets up the inode de-allocation dependency.
2021 freefile
= kmalloc(sizeof(struct freefile
), M_FREEFILE
,
2023 freefile
->fx_list
.wk_type
= D_FREEFILE
;
2024 freefile
->fx_list
.wk_state
= 0;
2025 freefile
->fx_mode
= mode
;
2026 freefile
->fx_oldinum
= ino
;
2027 freefile
->fx_devvp
= ip
->i_devvp
;
2028 freefile
->fx_fs
= ip
->i_fs
;
2031 * If the inodedep does not exist, then the zero'ed inode has
2032 * been written to disk. If the allocated inode has never been
2033 * written to disk, then the on-disk inode is zero'ed. In either
2034 * case we can free the file immediately.
2037 if (inodedep_lookup(ip
->i_fs
, ino
, 0, &inodedep
) == 0 ||
2038 check_inode_unwritten(inodedep
)) {
2040 handle_workitem_freefile(freefile
);
2043 WORKLIST_INSERT(&inodedep
->id_inowait
, &freefile
->fx_list
);
2048 * Check to see if an inode has never been written to disk. If
2049 * so free the inodedep and return success, otherwise return failure.
2050 * This routine must be called with splbio interrupts blocked.
2052 * If we still have a bitmap dependency, then the inode has never
2053 * been written to disk. Drop the dependency as it is no longer
2054 * necessary since the inode is being deallocated. We set the
2055 * ALLCOMPLETE flags since the bitmap now properly shows that the
2056 * inode is not allocated. Even if the inode is actively being
2057 * written, it has been rolled back to its zero'ed state, so we
2058 * are ensured that a zero inode is what is on the disk. For short
2059 * lived files, this change will usually result in removing all the
2060 * dependencies from the inode so that it can be freed immediately.
2063 check_inode_unwritten(struct inodedep
*inodedep
)
2066 if ((inodedep
->id_state
& DEPCOMPLETE
) != 0 ||
2067 LIST_FIRST(&inodedep
->id_pendinghd
) != NULL
||
2068 LIST_FIRST(&inodedep
->id_bufwait
) != NULL
||
2069 LIST_FIRST(&inodedep
->id_inowait
) != NULL
||
2070 TAILQ_FIRST(&inodedep
->id_inoupdt
) != NULL
||
2071 TAILQ_FIRST(&inodedep
->id_newinoupdt
) != NULL
||
2072 inodedep
->id_nlinkdelta
!= 0)
2076 * Another process might be in initiate_write_inodeblock
2077 * trying to allocate memory without holding "Softdep Lock".
2079 if ((inodedep
->id_state
& IOSTARTED
) != 0 &&
2080 inodedep
->id_savedino
== NULL
)
2083 inodedep
->id_state
|= ALLCOMPLETE
;
2084 LIST_REMOVE(inodedep
, id_deps
);
2085 inodedep
->id_buf
= NULL
;
2086 if (inodedep
->id_state
& ONWORKLIST
)
2087 WORKLIST_REMOVE(&inodedep
->id_list
);
2088 if (inodedep
->id_savedino
!= NULL
) {
2089 kfree(inodedep
->id_savedino
, M_INODEDEP
);
2090 inodedep
->id_savedino
= NULL
;
2092 if (free_inodedep(inodedep
) == 0) {
2093 panic("check_inode_unwritten: busy inode");
2099 * Try to free an inodedep structure. Return 1 if it could be freed.
2102 free_inodedep(struct inodedep
*inodedep
)
2105 if ((inodedep
->id_state
& ONWORKLIST
) != 0 ||
2106 (inodedep
->id_state
& ALLCOMPLETE
) != ALLCOMPLETE
||
2107 LIST_FIRST(&inodedep
->id_pendinghd
) != NULL
||
2108 LIST_FIRST(&inodedep
->id_bufwait
) != NULL
||
2109 LIST_FIRST(&inodedep
->id_inowait
) != NULL
||
2110 TAILQ_FIRST(&inodedep
->id_inoupdt
) != NULL
||
2111 TAILQ_FIRST(&inodedep
->id_newinoupdt
) != NULL
||
2112 inodedep
->id_nlinkdelta
!= 0 || inodedep
->id_savedino
!= NULL
)
2114 LIST_REMOVE(inodedep
, id_hash
);
2115 WORKITEM_FREE(inodedep
, D_INODEDEP
);
2121 * This workitem routine performs the block de-allocation.
2122 * The workitem is added to the pending list after the updated
2123 * inode block has been written to disk. As mentioned above,
2124 * checks regarding the number of blocks de-allocated (compared
2125 * to the number of blocks allocated for the file) are also
2126 * performed in this function.
2129 handle_workitem_freeblocks(struct freeblks
*freeblks
)
2134 int i
, level
, bsize
;
2135 long nblocks
, blocksreleased
= 0;
2136 int error
, allerror
= 0;
2137 ufs_lbn_t baselbns
[NIADDR
], tmpval
;
2139 tip
.i_number
= freeblks
->fb_previousinum
;
2140 tip
.i_devvp
= freeblks
->fb_devvp
;
2141 tip
.i_dev
= freeblks
->fb_devvp
->v_rdev
;
2142 tip
.i_fs
= freeblks
->fb_fs
;
2143 tip
.i_size
= freeblks
->fb_oldsize
;
2144 tip
.i_uid
= freeblks
->fb_uid
;
2145 fs
= freeblks
->fb_fs
;
2147 baselbns
[0] = NDADDR
;
2148 for (i
= 1; i
< NIADDR
; i
++) {
2149 tmpval
*= NINDIR(fs
);
2150 baselbns
[i
] = baselbns
[i
- 1] + tmpval
;
2152 nblocks
= btodb(fs
->fs_bsize
);
2155 * Indirect blocks first.
2157 for (level
= (NIADDR
- 1); level
>= 0; level
--) {
2158 if ((bn
= freeblks
->fb_iblks
[level
]) == 0)
2160 if ((error
= indir_trunc(&tip
, fsbtodoff(fs
, bn
), level
,
2161 baselbns
[level
], &blocksreleased
)) == 0)
2163 ffs_blkfree(&tip
, bn
, fs
->fs_bsize
);
2164 blocksreleased
+= nblocks
;
2167 * All direct blocks or frags.
2169 for (i
= (NDADDR
- 1); i
>= 0; i
--) {
2170 if ((bn
= freeblks
->fb_dblks
[i
]) == 0)
2172 bsize
= blksize(fs
, &tip
, i
);
2173 ffs_blkfree(&tip
, bn
, bsize
);
2174 blocksreleased
+= btodb(bsize
);
2178 if (freeblks
->fb_chkcnt
!= blocksreleased
)
2179 kprintf("handle_workitem_freeblocks: block count\n");
2181 softdep_error("handle_workitem_freeblks", allerror
);
2182 #endif /* DIAGNOSTIC */
2183 WORKITEM_FREE(freeblks
, D_FREEBLKS
);
2187 * Release blocks associated with the inode ip and stored in the indirect
2188 * block at doffset. If level is greater than SINGLE, the block is an
2189 * indirect block and recursive calls to indirtrunc must be used to
2190 * cleanse other indirect blocks.
2193 indir_trunc(struct inode
*ip
, off_t doffset
, int level
, ufs_lbn_t lbn
,
2200 struct worklist
*wk
;
2201 struct indirdep
*indirdep
;
2202 int i
, lbnadd
, nblocks
;
2203 int error
, allerror
= 0;
2207 for (i
= level
; i
> 0; i
--)
2208 lbnadd
*= NINDIR(fs
);
2210 * Get buffer of block pointers to be freed. This routine is not
2211 * called until the zero'ed inode has been written, so it is safe
2212 * to free blocks as they are encountered. Because the inode has
2213 * been zero'ed, calls to bmap on these blocks will fail. So, we
2214 * have to use the on-disk address and the block device for the
2215 * filesystem to look them up. If the file was deleted before its
2216 * indirect blocks were all written to disk, the routine that set
2217 * us up (deallocate_dependencies) will have arranged to leave
2218 * a complete copy of the indirect block in memory for our use.
2219 * Otherwise we have to read the blocks in from the disk.
2222 if ((bp
= findblk(ip
->i_devvp
, doffset
, FINDBLK_TEST
)) != NULL
&&
2223 (wk
= LIST_FIRST(&bp
->b_dep
)) != NULL
) {
2225 * bp must be ir_savebp, which is held locked for our use.
2227 if (wk
->wk_type
!= D_INDIRDEP
||
2228 (indirdep
= WK_INDIRDEP(wk
))->ir_savebp
!= bp
||
2229 (indirdep
->ir_state
& GOINGAWAY
) == 0) {
2230 panic("indir_trunc: lost indirdep");
2232 WORKLIST_REMOVE(wk
);
2233 WORKITEM_FREE(indirdep
, D_INDIRDEP
);
2234 if (LIST_FIRST(&bp
->b_dep
) != NULL
) {
2235 panic("indir_trunc: dangling dep");
2240 error
= bread(ip
->i_devvp
, doffset
, (int)fs
->fs_bsize
, &bp
);
2245 * Recursively free indirect blocks.
2247 bap
= (ufs_daddr_t
*)bp
->b_data
;
2248 nblocks
= btodb(fs
->fs_bsize
);
2249 for (i
= NINDIR(fs
) - 1; i
>= 0; i
--) {
2250 if ((nb
= bap
[i
]) == 0)
2253 if ((error
= indir_trunc(ip
, fsbtodoff(fs
, nb
),
2254 level
- 1, lbn
+ (i
* lbnadd
), countp
)) != 0)
2257 ffs_blkfree(ip
, nb
, fs
->fs_bsize
);
2260 bp
->b_flags
|= B_INVAL
| B_NOCACHE
;
2266 * Free an allocindir.
2267 * This routine must be called with splbio interrupts blocked.
2270 free_allocindir(struct allocindir
*aip
, struct inodedep
*inodedep
)
2272 struct freefrag
*freefrag
;
2274 KKASSERT(lock_held(&lk
) > 0);
2276 if ((aip
->ai_state
& DEPCOMPLETE
) == 0)
2277 LIST_REMOVE(aip
, ai_deps
);
2278 if (aip
->ai_state
& ONWORKLIST
)
2279 WORKLIST_REMOVE(&aip
->ai_list
);
2280 LIST_REMOVE(aip
, ai_next
);
2281 if ((freefrag
= aip
->ai_freefrag
) != NULL
) {
2282 if (inodedep
== NULL
)
2283 add_to_worklist(&freefrag
->ff_list
);
2285 WORKLIST_INSERT(&inodedep
->id_bufwait
,
2286 &freefrag
->ff_list
);
2288 WORKITEM_FREE(aip
, D_ALLOCINDIR
);
2292 * Directory entry addition dependencies.
2294 * When adding a new directory entry, the inode (with its incremented link
2295 * count) must be written to disk before the directory entry's pointer to it.
2296 * Also, if the inode is newly allocated, the corresponding freemap must be
2297 * updated (on disk) before the directory entry's pointer. These requirements
2298 * are met via undo/redo on the directory entry's pointer, which consists
2299 * simply of the inode number.
2301 * As directory entries are added and deleted, the free space within a
2302 * directory block can become fragmented. The ufs filesystem will compact
2303 * a fragmented directory block to make space for a new entry. When this
2304 * occurs, the offsets of previously added entries change. Any "diradd"
2305 * dependency structures corresponding to these entries must be updated with
2310 * This routine is called after the in-memory inode's link
2311 * count has been incremented, but before the directory entry's
2312 * pointer to the inode has been set.
2315 * bp: buffer containing directory block
2316 * dp: inode for directory
2317 * diroffset: offset of new entry in directory
2318 * newinum: inode referenced by new directory entry
2319 * newdirbp: non-NULL => contents of new mkdir
2322 softdep_setup_directory_add(struct buf
*bp
, struct inode
*dp
, off_t diroffset
,
2323 ino_t newinum
, struct buf
*newdirbp
)
2325 int offset
; /* offset of new entry within directory block */
2326 ufs_lbn_t lbn
; /* block in directory containing new entry */
2329 struct pagedep
*pagedep
;
2330 struct inodedep
*inodedep
;
2331 struct mkdir
*mkdir1
, *mkdir2
;
2334 * Whiteouts have no dependencies.
2336 if (newinum
== WINO
) {
2337 if (newdirbp
!= NULL
)
2343 lbn
= lblkno(fs
, diroffset
);
2344 offset
= blkoff(fs
, diroffset
);
2345 dap
= kmalloc(sizeof(struct diradd
), M_DIRADD
,
2346 M_SOFTDEP_FLAGS
| M_ZERO
);
2347 dap
->da_list
.wk_type
= D_DIRADD
;
2348 dap
->da_offset
= offset
;
2349 dap
->da_newinum
= newinum
;
2350 dap
->da_state
= ATTACHED
;
2351 if (newdirbp
== NULL
) {
2352 dap
->da_state
|= DEPCOMPLETE
;
2355 dap
->da_state
|= MKDIR_BODY
| MKDIR_PARENT
;
2356 mkdir1
= kmalloc(sizeof(struct mkdir
), M_MKDIR
,
2358 mkdir1
->md_list
.wk_type
= D_MKDIR
;
2359 mkdir1
->md_state
= MKDIR_BODY
;
2360 mkdir1
->md_diradd
= dap
;
2361 mkdir2
= kmalloc(sizeof(struct mkdir
), M_MKDIR
,
2363 mkdir2
->md_list
.wk_type
= D_MKDIR
;
2364 mkdir2
->md_state
= MKDIR_PARENT
;
2365 mkdir2
->md_diradd
= dap
;
2367 * Dependency on "." and ".." being written to disk.
2369 mkdir1
->md_buf
= newdirbp
;
2371 LIST_INSERT_HEAD(&mkdirlisthd
, mkdir1
, md_mkdirs
);
2372 WORKLIST_INSERT_BP(newdirbp
, &mkdir1
->md_list
);
2376 * Dependency on link count increase for parent directory
2379 if (inodedep_lookup(dp
->i_fs
, dp
->i_number
, 0, &inodedep
) == 0
2380 || (inodedep
->id_state
& ALLCOMPLETE
) == ALLCOMPLETE
) {
2381 dap
->da_state
&= ~MKDIR_PARENT
;
2382 WORKITEM_FREE(mkdir2
, D_MKDIR
);
2384 LIST_INSERT_HEAD(&mkdirlisthd
, mkdir2
, md_mkdirs
);
2385 WORKLIST_INSERT(&inodedep
->id_bufwait
,&mkdir2
->md_list
);
2389 * Link into parent directory pagedep to await its being written.
2391 if (pagedep_lookup(dp
, lbn
, DEPALLOC
, &pagedep
) == 0)
2392 WORKLIST_INSERT_BP(bp
, &pagedep
->pd_list
);
2393 dap
->da_pagedep
= pagedep
;
2394 LIST_INSERT_HEAD(&pagedep
->pd_diraddhd
[DIRADDHASH(offset
)], dap
,
2397 * Link into its inodedep. Put it on the id_bufwait list if the inode
2398 * is not yet written. If it is written, do the post-inode write
2399 * processing to put it on the id_pendinghd list.
2401 (void) inodedep_lookup(fs
, newinum
, DEPALLOC
, &inodedep
);
2402 if ((inodedep
->id_state
& ALLCOMPLETE
) == ALLCOMPLETE
)
2403 diradd_inode_written(dap
, inodedep
);
2405 WORKLIST_INSERT(&inodedep
->id_bufwait
, &dap
->da_list
);
2410 * This procedure is called to change the offset of a directory
2411 * entry when compacting a directory block which must be owned
2412 * exclusively by the caller. Note that the actual entry movement
2413 * must be done in this procedure to ensure that no I/O completions
2414 * occur while the move is in progress.
2417 * dp: inode for directory
2418 * base: address of dp->i_offset
2419 * oldloc: address of old directory location
2420 * newloc: address of new directory location
2421 * entrysize: size of directory entry
2424 softdep_change_directoryentry_offset(struct inode
*dp
, caddr_t base
,
2425 caddr_t oldloc
, caddr_t newloc
,
2428 int offset
, oldoffset
, newoffset
;
2429 struct pagedep
*pagedep
;
2434 lbn
= lblkno(dp
->i_fs
, dp
->i_offset
);
2435 offset
= blkoff(dp
->i_fs
, dp
->i_offset
);
2436 if (pagedep_lookup(dp
, lbn
, 0, &pagedep
) == 0)
2438 oldoffset
= offset
+ (oldloc
- base
);
2439 newoffset
= offset
+ (newloc
- base
);
2441 LIST_FOREACH(dap
, &pagedep
->pd_diraddhd
[DIRADDHASH(oldoffset
)], da_pdlist
) {
2442 if (dap
->da_offset
!= oldoffset
)
2444 dap
->da_offset
= newoffset
;
2445 if (DIRADDHASH(newoffset
) == DIRADDHASH(oldoffset
))
2447 LIST_REMOVE(dap
, da_pdlist
);
2448 LIST_INSERT_HEAD(&pagedep
->pd_diraddhd
[DIRADDHASH(newoffset
)],
2454 LIST_FOREACH(dap
, &pagedep
->pd_pendinghd
, da_pdlist
) {
2455 if (dap
->da_offset
== oldoffset
) {
2456 dap
->da_offset
= newoffset
;
2462 bcopy(oldloc
, newloc
, entrysize
);
2467 * Free a diradd dependency structure. This routine must be called
2468 * with splbio interrupts blocked.
2471 free_diradd(struct diradd
*dap
)
2473 struct dirrem
*dirrem
;
2474 struct pagedep
*pagedep
;
2475 struct inodedep
*inodedep
;
2476 struct mkdir
*mkdir
, *nextmd
;
2478 KKASSERT(lock_held(&lk
) > 0);
2480 WORKLIST_REMOVE(&dap
->da_list
);
2481 LIST_REMOVE(dap
, da_pdlist
);
2482 if ((dap
->da_state
& DIRCHG
) == 0) {
2483 pagedep
= dap
->da_pagedep
;
2485 dirrem
= dap
->da_previous
;
2486 pagedep
= dirrem
->dm_pagedep
;
2487 dirrem
->dm_dirinum
= pagedep
->pd_ino
;
2488 add_to_worklist(&dirrem
->dm_list
);
2490 if (inodedep_lookup(VFSTOUFS(pagedep
->pd_mnt
)->um_fs
, dap
->da_newinum
,
2492 (void) free_inodedep(inodedep
);
2493 if ((dap
->da_state
& (MKDIR_PARENT
| MKDIR_BODY
)) != 0) {
2494 for (mkdir
= LIST_FIRST(&mkdirlisthd
); mkdir
; mkdir
= nextmd
) {
2495 nextmd
= LIST_NEXT(mkdir
, md_mkdirs
);
2496 if (mkdir
->md_diradd
!= dap
)
2498 dap
->da_state
&= ~mkdir
->md_state
;
2499 WORKLIST_REMOVE(&mkdir
->md_list
);
2500 LIST_REMOVE(mkdir
, md_mkdirs
);
2501 WORKITEM_FREE(mkdir
, D_MKDIR
);
2503 if ((dap
->da_state
& (MKDIR_PARENT
| MKDIR_BODY
)) != 0) {
2504 panic("free_diradd: unfound ref");
2507 WORKITEM_FREE(dap
, D_DIRADD
);
2511 * Directory entry removal dependencies.
2513 * When removing a directory entry, the entry's inode pointer must be
2514 * zero'ed on disk before the corresponding inode's link count is decremented
2515 * (possibly freeing the inode for re-use). This dependency is handled by
2516 * updating the directory entry but delaying the inode count reduction until
2517 * after the directory block has been written to disk. After this point, the
2518 * inode count can be decremented whenever it is convenient.
2522 * This routine should be called immediately after removing
2523 * a directory entry. The inode's link count should not be
2524 * decremented by the calling procedure -- the soft updates
2525 * code will do this task when it is safe.
2528 * bp: buffer containing directory block
2529 * dp: inode for the directory being modified
2530 * ip: inode for directory entry being removed
2531 * isrmdir: indicates if doing RMDIR
2534 softdep_setup_remove(struct buf
*bp
, struct inode
*dp
, struct inode
*ip
,
2537 struct dirrem
*dirrem
, *prevdirrem
;
2540 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2542 dirrem
= newdirrem(bp
, dp
, ip
, isrmdir
, &prevdirrem
);
2545 * If the COMPLETE flag is clear, then there were no active
2546 * entries and we want to roll back to a zeroed entry until
2547 * the new inode is committed to disk. If the COMPLETE flag is
2548 * set then we have deleted an entry that never made it to
2549 * disk. If the entry we deleted resulted from a name change,
2550 * then the old name still resides on disk. We cannot delete
2551 * its inode (returned to us in prevdirrem) until the zeroed
2552 * directory entry gets to disk. The new inode has never been
2553 * referenced on the disk, so can be deleted immediately.
2555 if ((dirrem
->dm_state
& COMPLETE
) == 0) {
2556 LIST_INSERT_HEAD(&dirrem
->dm_pagedep
->pd_dirremhd
, dirrem
,
2560 if (prevdirrem
!= NULL
)
2561 LIST_INSERT_HEAD(&dirrem
->dm_pagedep
->pd_dirremhd
,
2562 prevdirrem
, dm_next
);
2563 dirrem
->dm_dirinum
= dirrem
->dm_pagedep
->pd_ino
;
2565 handle_workitem_remove(dirrem
);
2570 * Allocate a new dirrem if appropriate and return it along with
2571 * its associated pagedep. Called without a lock, returns with lock.
2573 static long num_dirrem
; /* number of dirrem allocated */
2577 * bp: buffer containing directory block
2578 * dp: inode for the directory being modified
2579 * ip: inode for directory entry being removed
2580 * isrmdir: indicates if doing RMDIR
2581 * prevdirremp: previously referenced inode, if any
2583 static struct dirrem
*
2584 newdirrem(struct buf
*bp
, struct inode
*dp
, struct inode
*ip
,
2585 int isrmdir
, struct dirrem
**prevdirremp
)
2590 struct dirrem
*dirrem
;
2591 struct pagedep
*pagedep
;
2594 * Whiteouts have no deletion dependencies.
2597 panic("newdirrem: whiteout");
2599 * If we are over our limit, try to improve the situation.
2600 * Limiting the number of dirrem structures will also limit
2601 * the number of freefile and freeblks structures.
2603 if (num_dirrem
> max_softdeps
/ 4)
2604 speedup_syncer(NULL
);
2605 if (num_dirrem
> max_softdeps
/ 2) {
2607 request_cleanup(FLUSH_REMOVE
);
2612 dirrem
= kmalloc(sizeof(struct dirrem
), M_DIRREM
,
2613 M_SOFTDEP_FLAGS
| M_ZERO
);
2614 dirrem
->dm_list
.wk_type
= D_DIRREM
;
2615 dirrem
->dm_state
= isrmdir
? RMDIR
: 0;
2616 dirrem
->dm_mnt
= ITOV(ip
)->v_mount
;
2617 dirrem
->dm_oldinum
= ip
->i_number
;
2618 *prevdirremp
= NULL
;
2621 lbn
= lblkno(dp
->i_fs
, dp
->i_offset
);
2622 offset
= blkoff(dp
->i_fs
, dp
->i_offset
);
2623 if (pagedep_lookup(dp
, lbn
, DEPALLOC
, &pagedep
) == 0)
2624 WORKLIST_INSERT_BP(bp
, &pagedep
->pd_list
);
2625 dirrem
->dm_pagedep
= pagedep
;
2627 * Check for a diradd dependency for the same directory entry.
2628 * If present, then both dependencies become obsolete and can
2629 * be de-allocated. Check for an entry on both the pd_dirraddhd
2630 * list and the pd_pendinghd list.
2633 LIST_FOREACH(dap
, &pagedep
->pd_diraddhd
[DIRADDHASH(offset
)], da_pdlist
)
2634 if (dap
->da_offset
== offset
)
2638 LIST_FOREACH(dap
, &pagedep
->pd_pendinghd
, da_pdlist
)
2639 if (dap
->da_offset
== offset
)
2645 * Must be ATTACHED at this point.
2647 if ((dap
->da_state
& ATTACHED
) == 0) {
2648 panic("newdirrem: not ATTACHED");
2650 if (dap
->da_newinum
!= ip
->i_number
) {
2651 panic("newdirrem: inum %"PRId64
" should be %"PRId64
,
2652 ip
->i_number
, dap
->da_newinum
);
2655 * If we are deleting a changed name that never made it to disk,
2656 * then return the dirrem describing the previous inode (which
2657 * represents the inode currently referenced from this entry on disk).
2659 if ((dap
->da_state
& DIRCHG
) != 0) {
2660 *prevdirremp
= dap
->da_previous
;
2661 dap
->da_state
&= ~DIRCHG
;
2662 dap
->da_pagedep
= pagedep
;
2665 * We are deleting an entry that never made it to disk.
2666 * Mark it COMPLETE so we can delete its inode immediately.
2668 dirrem
->dm_state
|= COMPLETE
;
2674 * Directory entry change dependencies.
2676 * Changing an existing directory entry requires that an add operation
2677 * be completed first followed by a deletion. The semantics for the addition
2678 * are identical to the description of adding a new entry above except
2679 * that the rollback is to the old inode number rather than zero. Once
2680 * the addition dependency is completed, the removal is done as described
2681 * in the removal routine above.
2685 * This routine should be called immediately after changing
2686 * a directory entry. The inode's link count should not be
2687 * decremented by the calling procedure -- the soft updates
2688 * code will perform this task when it is safe.
2691 * bp: buffer containing directory block
2692 * dp: inode for the directory being modified
2693 * ip: inode for directory entry being removed
2694 * newinum: new inode number for changed entry
2695 * isrmdir: indicates if doing RMDIR
2698 softdep_setup_directory_change(struct buf
*bp
, struct inode
*dp
,
2699 struct inode
*ip
, ino_t newinum
,
2703 struct diradd
*dap
= NULL
;
2704 struct dirrem
*dirrem
, *prevdirrem
;
2705 struct pagedep
*pagedep
;
2706 struct inodedep
*inodedep
;
2708 offset
= blkoff(dp
->i_fs
, dp
->i_offset
);
2711 * Whiteouts do not need diradd dependencies.
2713 if (newinum
!= WINO
) {
2714 dap
= kmalloc(sizeof(struct diradd
), M_DIRADD
,
2715 M_SOFTDEP_FLAGS
| M_ZERO
);
2716 dap
->da_list
.wk_type
= D_DIRADD
;
2717 dap
->da_state
= DIRCHG
| ATTACHED
| DEPCOMPLETE
;
2718 dap
->da_offset
= offset
;
2719 dap
->da_newinum
= newinum
;
2723 * Allocate a new dirrem and ACQUIRE_LOCK.
2725 dirrem
= newdirrem(bp
, dp
, ip
, isrmdir
, &prevdirrem
);
2726 pagedep
= dirrem
->dm_pagedep
;
2728 * The possible values for isrmdir:
2729 * 0 - non-directory file rename
2730 * 1 - directory rename within same directory
2731 * inum - directory rename to new directory of given inode number
2732 * When renaming to a new directory, we are both deleting and
2733 * creating a new directory entry, so the link count on the new
2734 * directory should not change. Thus we do not need the followup
2735 * dirrem which is usually done in handle_workitem_remove. We set
2736 * the DIRCHG flag to tell handle_workitem_remove to skip the
2740 dirrem
->dm_state
|= DIRCHG
;
2743 * Whiteouts have no additional dependencies,
2744 * so just put the dirrem on the correct list.
2746 if (newinum
== WINO
) {
2747 if ((dirrem
->dm_state
& COMPLETE
) == 0) {
2748 LIST_INSERT_HEAD(&pagedep
->pd_dirremhd
, dirrem
,
2751 dirrem
->dm_dirinum
= pagedep
->pd_ino
;
2752 add_to_worklist(&dirrem
->dm_list
);
2759 * If the COMPLETE flag is clear, then there were no active
2760 * entries and we want to roll back to the previous inode until
2761 * the new inode is committed to disk. If the COMPLETE flag is
2762 * set, then we have deleted an entry that never made it to disk.
2763 * If the entry we deleted resulted from a name change, then the old
2764 * inode reference still resides on disk. Any rollback that we do
2765 * needs to be to that old inode (returned to us in prevdirrem). If
2766 * the entry we deleted resulted from a create, then there is
2767 * no entry on the disk, so we want to roll back to zero rather
2768 * than the uncommitted inode. In either of the COMPLETE cases we
2769 * want to immediately free the unwritten and unreferenced inode.
2771 if ((dirrem
->dm_state
& COMPLETE
) == 0) {
2772 dap
->da_previous
= dirrem
;
2774 if (prevdirrem
!= NULL
) {
2775 dap
->da_previous
= prevdirrem
;
2777 dap
->da_state
&= ~DIRCHG
;
2778 dap
->da_pagedep
= pagedep
;
2780 dirrem
->dm_dirinum
= pagedep
->pd_ino
;
2781 add_to_worklist(&dirrem
->dm_list
);
2784 * Link into its inodedep. Put it on the id_bufwait list if the inode
2785 * is not yet written. If it is written, do the post-inode write
2786 * processing to put it on the id_pendinghd list.
2788 if (inodedep_lookup(dp
->i_fs
, newinum
, DEPALLOC
, &inodedep
) == 0 ||
2789 (inodedep
->id_state
& ALLCOMPLETE
) == ALLCOMPLETE
) {
2790 dap
->da_state
|= COMPLETE
;
2791 LIST_INSERT_HEAD(&pagedep
->pd_pendinghd
, dap
, da_pdlist
);
2792 WORKLIST_INSERT(&inodedep
->id_pendinghd
, &dap
->da_list
);
2794 LIST_INSERT_HEAD(&pagedep
->pd_diraddhd
[DIRADDHASH(offset
)],
2796 WORKLIST_INSERT(&inodedep
->id_bufwait
, &dap
->da_list
);
2802 * Called whenever the link count on an inode is changed.
2803 * It creates an inode dependency so that the new reference(s)
2804 * to the inode cannot be committed to disk until the updated
2805 * inode has been written.
2808 * ip: the inode with the increased link count
2811 softdep_change_linkcnt(struct inode
*ip
)
2813 struct inodedep
*inodedep
;
2816 (void) inodedep_lookup(ip
->i_fs
, ip
->i_number
, DEPALLOC
, &inodedep
);
2817 if (ip
->i_nlink
< ip
->i_effnlink
) {
2818 panic("softdep_change_linkcnt: bad delta");
2820 inodedep
->id_nlinkdelta
= ip
->i_nlink
- ip
->i_effnlink
;
2825 * This workitem decrements the inode's link count.
2826 * If the link count reaches zero, the file is removed.
2829 handle_workitem_remove(struct dirrem
*dirrem
)
2831 struct inodedep
*inodedep
;
2837 error
= VFS_VGET(dirrem
->dm_mnt
, NULL
, dirrem
->dm_oldinum
, &vp
);
2839 softdep_error("handle_workitem_remove: vget", error
);
2844 if ((inodedep_lookup(ip
->i_fs
, dirrem
->dm_oldinum
, 0, &inodedep
)) == 0){
2845 panic("handle_workitem_remove: lost inodedep");
2848 * Normal file deletion.
2850 if ((dirrem
->dm_state
& RMDIR
) == 0) {
2852 ip
->i_flag
|= IN_CHANGE
;
2853 if (ip
->i_nlink
< ip
->i_effnlink
) {
2854 panic("handle_workitem_remove: bad file delta");
2856 inodedep
->id_nlinkdelta
= ip
->i_nlink
- ip
->i_effnlink
;
2860 WORKITEM_FREE(dirrem
, D_DIRREM
);
2864 * Directory deletion. Decrement reference count for both the
2865 * just deleted parent directory entry and the reference for ".".
2866 * Next truncate the directory to length zero. When the
2867 * truncation completes, arrange to have the reference count on
2868 * the parent decremented to account for the loss of "..".
2871 ip
->i_flag
|= IN_CHANGE
;
2872 if (ip
->i_nlink
< ip
->i_effnlink
) {
2873 panic("handle_workitem_remove: bad dir delta");
2875 inodedep
->id_nlinkdelta
= ip
->i_nlink
- ip
->i_effnlink
;
2877 if ((error
= ffs_truncate(vp
, (off_t
)0, 0, proc0
.p_ucred
)) != 0)
2878 softdep_error("handle_workitem_remove: truncate", error
);
2880 * Rename a directory to a new parent. Since, we are both deleting
2881 * and creating a new directory entry, the link count on the new
2882 * directory should not change. Thus we skip the followup dirrem.
2884 if (dirrem
->dm_state
& DIRCHG
) {
2887 WORKITEM_FREE(dirrem
, D_DIRREM
);
2891 * If the inodedep does not exist, then the zero'ed inode has
2892 * been written to disk. If the allocated inode has never been
2893 * written to disk, then the on-disk inode is zero'ed. In either
2894 * case we can remove the file immediately.
2897 dirrem
->dm_state
= 0;
2898 oldinum
= dirrem
->dm_oldinum
;
2899 dirrem
->dm_oldinum
= dirrem
->dm_dirinum
;
2900 if (inodedep_lookup(ip
->i_fs
, oldinum
, 0, &inodedep
) == 0 ||
2901 check_inode_unwritten(inodedep
)) {
2904 handle_workitem_remove(dirrem
);
2907 WORKLIST_INSERT(&inodedep
->id_inowait
, &dirrem
->dm_list
);
2909 ip
->i_flag
|= IN_CHANGE
;
2915 * Inode de-allocation dependencies.
2917 * When an inode's link count is reduced to zero, it can be de-allocated. We
2918 * found it convenient to postpone de-allocation until after the inode is
2919 * written to disk with its new link count (zero). At this point, all of the
2920 * on-disk inode's block pointers are nullified and, with careful dependency
2921 * list ordering, all dependencies related to the inode will be satisfied and
2922 * the corresponding dependency structures de-allocated. So, if/when the
2923 * inode is reused, there will be no mixing of old dependencies with new
2924 * ones. This artificial dependency is set up by the block de-allocation
2925 * procedure above (softdep_setup_freeblocks) and completed by the
2926 * following procedure.
2929 handle_workitem_freefile(struct freefile
*freefile
)
2933 struct inodedep
*idp
;
2938 error
= inodedep_lookup(freefile
->fx_fs
, freefile
->fx_oldinum
, 0, &idp
);
2941 panic("handle_workitem_freefile: inodedep survived");
2943 tip
.i_devvp
= freefile
->fx_devvp
;
2944 tip
.i_dev
= freefile
->fx_devvp
->v_rdev
;
2945 tip
.i_fs
= freefile
->fx_fs
;
2947 if ((error
= ffs_freefile(&vp
, freefile
->fx_oldinum
, freefile
->fx_mode
)) != 0)
2948 softdep_error("handle_workitem_freefile", error
);
2949 WORKITEM_FREE(freefile
, D_FREEFILE
);
2953 * Helper function which unlinks marker element from work list and returns
2954 * the next element on the list.
2956 static __inline
struct worklist
*
2957 markernext(struct worklist
*marker
)
2959 struct worklist
*next
;
2961 next
= LIST_NEXT(marker
, wk_list
);
2962 LIST_REMOVE(marker
, wk_list
);
2967 * checkread, checkwrite
2969 * bioops callback - hold io_token
2972 softdep_checkread(struct buf
*bp
)
2974 /* nothing to do, mp lock not needed */
2979 * bioops callback - hold io_token
2982 softdep_checkwrite(struct buf
*bp
)
2984 /* nothing to do, mp lock not needed */
2991 * The dependency structures constructed above are most actively used when file
2992 * system blocks are written to disk. No constraints are placed on when a
2993 * block can be written, but unsatisfied update dependencies are made safe by
2994 * modifying (or replacing) the source memory for the duration of the disk
2995 * write. When the disk write completes, the memory block is again brought
2998 * In-core inode structure reclamation.
3000 * Because there are a finite number of "in-core" inode structures, they are
3001 * reused regularly. By transferring all inode-related dependencies to the
3002 * in-memory inode block and indexing them separately (via "inodedep"s), we
3003 * can allow "in-core" inode structures to be reused at any time and avoid
3004 * any increase in contention.
3006 * Called just before entering the device driver to initiate a new disk I/O.
3007 * The buffer must be locked, thus, no I/O completion operations can occur
3008 * while we are manipulating its associated dependencies.
3010 * bioops callback - hold io_token
3013 * bp: structure describing disk write to occur
3016 softdep_disk_io_initiation(struct buf
*bp
)
3018 struct worklist
*wk
;
3019 struct worklist marker
;
3020 struct indirdep
*indirdep
;
3023 * We only care about write operations. There should never
3024 * be dependencies for reads.
3026 if (bp
->b_cmd
== BUF_CMD_READ
)
3027 panic("softdep_disk_io_initiation: read");
3030 marker
.wk_type
= D_LAST
+ 1; /* Not a normal workitem */
3033 * Do any necessary pre-I/O processing.
3035 for (wk
= LIST_FIRST(&bp
->b_dep
); wk
; wk
= markernext(&marker
)) {
3036 LIST_INSERT_AFTER(wk
, &marker
, wk_list
);
3038 switch (wk
->wk_type
) {
3040 initiate_write_filepage(WK_PAGEDEP(wk
), bp
);
3044 initiate_write_inodeblock(WK_INODEDEP(wk
), bp
);
3048 indirdep
= WK_INDIRDEP(wk
);
3049 if (indirdep
->ir_state
& GOINGAWAY
)
3050 panic("disk_io_initiation: indirdep gone");
3052 * If there are no remaining dependencies, this
3053 * will be writing the real pointers, so the
3054 * dependency can be freed.
3056 if (LIST_FIRST(&indirdep
->ir_deplisthd
) == NULL
) {
3057 indirdep
->ir_savebp
->b_flags
|= B_INVAL
| B_NOCACHE
;
3058 brelse(indirdep
->ir_savebp
);
3059 /* inline expand WORKLIST_REMOVE(wk); */
3060 wk
->wk_state
&= ~ONWORKLIST
;
3061 LIST_REMOVE(wk
, wk_list
);
3062 WORKITEM_FREE(indirdep
, D_INDIRDEP
);
3066 * Replace up-to-date version with safe version.
3068 indirdep
->ir_saveddata
= kmalloc(bp
->b_bcount
,
3072 indirdep
->ir_state
&= ~ATTACHED
;
3073 indirdep
->ir_state
|= UNDONE
;
3074 bcopy(bp
->b_data
, indirdep
->ir_saveddata
, bp
->b_bcount
);
3075 bcopy(indirdep
->ir_savebp
->b_data
, bp
->b_data
,
3087 panic("handle_disk_io_initiation: Unexpected type %s",
3088 TYPENAME(wk
->wk_type
));
3096 * Called from within the procedure above to deal with unsatisfied
3097 * allocation dependencies in a directory. The buffer must be locked,
3098 * thus, no I/O completion operations can occur while we are
3099 * manipulating its associated dependencies.
3102 initiate_write_filepage(struct pagedep
*pagedep
, struct buf
*bp
)
3108 if (pagedep
->pd_state
& IOSTARTED
) {
3110 * This can only happen if there is a driver that does not
3111 * understand chaining. Here biodone will reissue the call
3112 * to strategy for the incomplete buffers.
3114 kprintf("initiate_write_filepage: already started\n");
3117 pagedep
->pd_state
|= IOSTARTED
;
3119 for (i
= 0; i
< DAHASHSZ
; i
++) {
3120 LIST_FOREACH(dap
, &pagedep
->pd_diraddhd
[i
], da_pdlist
) {
3121 ep
= (struct direct
*)
3122 ((char *)bp
->b_data
+ dap
->da_offset
);
3123 if (ep
->d_ino
!= dap
->da_newinum
) {
3124 panic("%s: dir inum %d != new %"PRId64
,
3125 "initiate_write_filepage",
3126 ep
->d_ino
, dap
->da_newinum
);
3128 if (dap
->da_state
& DIRCHG
)
3129 ep
->d_ino
= dap
->da_previous
->dm_oldinum
;
3132 dap
->da_state
&= ~ATTACHED
;
3133 dap
->da_state
|= UNDONE
;
3140 * Called from within the procedure above to deal with unsatisfied
3141 * allocation dependencies in an inodeblock. The buffer must be
3142 * locked, thus, no I/O completion operations can occur while we
3143 * are manipulating its associated dependencies.
3146 * bp: The inode block
3149 initiate_write_inodeblock(struct inodedep
*inodedep
, struct buf
*bp
)
3151 struct allocdirect
*adp
, *lastadp
;
3152 struct ufs1_dinode
*dp
;
3153 struct ufs1_dinode
*sip
;
3155 ufs_lbn_t prevlbn
= 0;
3158 if (inodedep
->id_state
& IOSTARTED
)
3159 panic("initiate_write_inodeblock: already started");
3160 inodedep
->id_state
|= IOSTARTED
;
3161 fs
= inodedep
->id_fs
;
3162 dp
= (struct ufs1_dinode
*)bp
->b_data
+
3163 ino_to_fsbo(fs
, inodedep
->id_ino
);
3165 * If the bitmap is not yet written, then the allocated
3166 * inode cannot be written to disk.
3168 if ((inodedep
->id_state
& DEPCOMPLETE
) == 0) {
3169 if (inodedep
->id_savedino
!= NULL
)
3170 panic("initiate_write_inodeblock: already doing I/O");
3171 sip
= kmalloc(sizeof(struct ufs1_dinode
), M_INODEDEP
,
3173 inodedep
->id_savedino
= sip
;
3174 *inodedep
->id_savedino
= *dp
;
3175 bzero((caddr_t
)dp
, sizeof(struct ufs1_dinode
));
3176 dp
->di_gen
= inodedep
->id_savedino
->di_gen
;
3180 * If no dependencies, then there is nothing to roll back.
3182 inodedep
->id_savedsize
= dp
->di_size
;
3183 if (TAILQ_FIRST(&inodedep
->id_inoupdt
) == NULL
)
3186 * Set the dependencies to busy.
3189 for (deplist
= 0, adp
= TAILQ_FIRST(&inodedep
->id_inoupdt
); adp
;
3190 adp
= TAILQ_NEXT(adp
, ad_next
)) {
3192 if (deplist
!= 0 && prevlbn
>= adp
->ad_lbn
) {
3193 panic("softdep_write_inodeblock: lbn order");
3195 prevlbn
= adp
->ad_lbn
;
3196 if (adp
->ad_lbn
< NDADDR
&&
3197 dp
->di_db
[adp
->ad_lbn
] != adp
->ad_newblkno
) {
3198 panic("%s: direct pointer #%ld mismatch %d != %d",
3199 "softdep_write_inodeblock", adp
->ad_lbn
,
3200 dp
->di_db
[adp
->ad_lbn
], adp
->ad_newblkno
);
3202 if (adp
->ad_lbn
>= NDADDR
&&
3203 dp
->di_ib
[adp
->ad_lbn
- NDADDR
] != adp
->ad_newblkno
) {
3204 panic("%s: indirect pointer #%ld mismatch %d != %d",
3205 "softdep_write_inodeblock", adp
->ad_lbn
- NDADDR
,
3206 dp
->di_ib
[adp
->ad_lbn
- NDADDR
], adp
->ad_newblkno
);
3208 deplist
|= 1 << adp
->ad_lbn
;
3209 if ((adp
->ad_state
& ATTACHED
) == 0) {
3210 panic("softdep_write_inodeblock: Unknown state 0x%x",
3213 #endif /* DIAGNOSTIC */
3214 adp
->ad_state
&= ~ATTACHED
;
3215 adp
->ad_state
|= UNDONE
;
3218 * The on-disk inode cannot claim to be any larger than the last
3219 * fragment that has been written. Otherwise, the on-disk inode
3220 * might have fragments that were not the last block in the file
3221 * which would corrupt the filesystem.
3223 for (lastadp
= NULL
, adp
= TAILQ_FIRST(&inodedep
->id_inoupdt
); adp
;
3224 lastadp
= adp
, adp
= TAILQ_NEXT(adp
, ad_next
)) {
3225 if (adp
->ad_lbn
>= NDADDR
)
3227 dp
->di_db
[adp
->ad_lbn
] = adp
->ad_oldblkno
;
3228 /* keep going until hitting a rollback to a frag */
3229 if (adp
->ad_oldsize
== 0 || adp
->ad_oldsize
== fs
->fs_bsize
)
3231 dp
->di_size
= fs
->fs_bsize
* adp
->ad_lbn
+ adp
->ad_oldsize
;
3232 for (i
= adp
->ad_lbn
+ 1; i
< NDADDR
; i
++) {
3234 if (dp
->di_db
[i
] != 0 && (deplist
& (1 << i
)) == 0) {
3235 panic("softdep_write_inodeblock: lost dep1");
3237 #endif /* DIAGNOSTIC */
3240 for (i
= 0; i
< NIADDR
; i
++) {
3242 if (dp
->di_ib
[i
] != 0 &&
3243 (deplist
& ((1 << NDADDR
) << i
)) == 0) {
3244 panic("softdep_write_inodeblock: lost dep2");
3246 #endif /* DIAGNOSTIC */
3253 * If we have zero'ed out the last allocated block of the file,
3254 * roll back the size to the last currently allocated block.
3255 * We know that this last allocated block is a full-sized as
3256 * we already checked for fragments in the loop above.
3258 if (lastadp
!= NULL
&&
3259 dp
->di_size
<= (lastadp
->ad_lbn
+ 1) * fs
->fs_bsize
) {
3260 for (i
= lastadp
->ad_lbn
; i
>= 0; i
--)
3261 if (dp
->di_db
[i
] != 0)
3263 dp
->di_size
= (i
+ 1) * fs
->fs_bsize
;
3266 * The only dependencies are for indirect blocks.
3268 * The file size for indirect block additions is not guaranteed.
3269 * Such a guarantee would be non-trivial to achieve. The conventional
3270 * synchronous write implementation also does not make this guarantee.
3271 * Fsck should catch and fix discrepancies. Arguably, the file size
3272 * can be over-estimated without destroying integrity when the file
3273 * moves into the indirect blocks (i.e., is large). If we want to
3274 * postpone fsck, we are stuck with this argument.
3276 for (; adp
; adp
= TAILQ_NEXT(adp
, ad_next
))
3277 dp
->di_ib
[adp
->ad_lbn
- NDADDR
] = 0;
3282 * This routine is called during the completion interrupt
3283 * service routine for a disk write (from the procedure called
3284 * by the device driver to inform the filesystem caches of
3285 * a request completion). It should be called early in this
3286 * procedure, before the block is made available to other
3287 * processes or other routines are called.
3289 * bioops callback - hold io_token
3292 * bp: describes the completed disk write
3295 softdep_disk_write_complete(struct buf
*bp
)
3297 struct worklist
*wk
;
3298 struct workhead reattach
;
3299 struct newblk
*newblk
;
3300 struct allocindir
*aip
;
3301 struct allocdirect
*adp
;
3302 struct indirdep
*indirdep
;
3303 struct inodedep
*inodedep
;
3304 struct bmsafemap
*bmsafemap
;
3308 LIST_INIT(&reattach
);
3309 while ((wk
= LIST_FIRST(&bp
->b_dep
)) != NULL
) {
3310 WORKLIST_REMOVE(wk
);
3311 switch (wk
->wk_type
) {
3314 if (handle_written_filepage(WK_PAGEDEP(wk
), bp
))
3315 WORKLIST_INSERT(&reattach
, wk
);
3319 if (handle_written_inodeblock(WK_INODEDEP(wk
), bp
))
3320 WORKLIST_INSERT(&reattach
, wk
);
3324 bmsafemap
= WK_BMSAFEMAP(wk
);
3325 while ((newblk
= LIST_FIRST(&bmsafemap
->sm_newblkhd
))) {
3326 newblk
->nb_state
|= DEPCOMPLETE
;
3327 newblk
->nb_bmsafemap
= NULL
;
3328 LIST_REMOVE(newblk
, nb_deps
);
3331 LIST_FIRST(&bmsafemap
->sm_allocdirecthd
))) {
3332 adp
->ad_state
|= DEPCOMPLETE
;
3334 LIST_REMOVE(adp
, ad_deps
);
3335 handle_allocdirect_partdone(adp
);
3338 LIST_FIRST(&bmsafemap
->sm_allocindirhd
))) {
3339 aip
->ai_state
|= DEPCOMPLETE
;
3341 LIST_REMOVE(aip
, ai_deps
);
3342 handle_allocindir_partdone(aip
);
3345 LIST_FIRST(&bmsafemap
->sm_inodedephd
)) != NULL
) {
3346 inodedep
->id_state
|= DEPCOMPLETE
;
3347 LIST_REMOVE(inodedep
, id_deps
);
3348 inodedep
->id_buf
= NULL
;
3350 WORKITEM_FREE(bmsafemap
, D_BMSAFEMAP
);
3354 handle_written_mkdir(WK_MKDIR(wk
), MKDIR_BODY
);
3358 adp
= WK_ALLOCDIRECT(wk
);
3359 adp
->ad_state
|= COMPLETE
;
3360 handle_allocdirect_partdone(adp
);
3364 aip
= WK_ALLOCINDIR(wk
);
3365 aip
->ai_state
|= COMPLETE
;
3366 handle_allocindir_partdone(aip
);
3370 indirdep
= WK_INDIRDEP(wk
);
3371 if (indirdep
->ir_state
& GOINGAWAY
) {
3372 panic("disk_write_complete: indirdep gone");
3374 bcopy(indirdep
->ir_saveddata
, bp
->b_data
, bp
->b_bcount
);
3375 kfree(indirdep
->ir_saveddata
, M_INDIRDEP
);
3376 indirdep
->ir_saveddata
= NULL
;
3377 indirdep
->ir_state
&= ~UNDONE
;
3378 indirdep
->ir_state
|= ATTACHED
;
3379 while ((aip
= LIST_FIRST(&indirdep
->ir_donehd
)) != NULL
) {
3380 handle_allocindir_partdone(aip
);
3381 if (aip
== LIST_FIRST(&indirdep
->ir_donehd
)) {
3382 panic("disk_write_complete: not gone");
3385 WORKLIST_INSERT(&reattach
, wk
);
3386 if ((bp
->b_flags
& B_DELWRI
) == 0)
3387 stat_indir_blk_ptrs
++;
3392 panic("handle_disk_write_complete: Unknown type %s",
3393 TYPENAME(wk
->wk_type
));
3398 * Reattach any requests that must be redone.
3400 while ((wk
= LIST_FIRST(&reattach
)) != NULL
) {
3401 WORKLIST_REMOVE(wk
);
3402 WORKLIST_INSERT_BP(bp
, wk
);
3409 * Called from within softdep_disk_write_complete above. Note that
3410 * this routine is always called from interrupt level with further
3411 * splbio interrupts blocked.
3414 * adp: the completed allocdirect
3417 handle_allocdirect_partdone(struct allocdirect
*adp
)
3419 struct allocdirect
*listadp
;
3420 struct inodedep
*inodedep
;
3423 if ((adp
->ad_state
& ALLCOMPLETE
) != ALLCOMPLETE
)
3425 if (adp
->ad_buf
!= NULL
)
3426 panic("handle_allocdirect_partdone: dangling dep");
3429 * The on-disk inode cannot claim to be any larger than the last
3430 * fragment that has been written. Otherwise, the on-disk inode
3431 * might have fragments that were not the last block in the file
3432 * which would corrupt the filesystem. Thus, we cannot free any
3433 * allocdirects after one whose ad_oldblkno claims a fragment as
3434 * these blocks must be rolled back to zero before writing the inode.
3435 * We check the currently active set of allocdirects in id_inoupdt.
3437 inodedep
= adp
->ad_inodedep
;
3438 bsize
= inodedep
->id_fs
->fs_bsize
;
3439 TAILQ_FOREACH(listadp
, &inodedep
->id_inoupdt
, ad_next
) {
3440 /* found our block */
3443 /* continue if ad_oldlbn is not a fragment */
3444 if (listadp
->ad_oldsize
== 0 ||
3445 listadp
->ad_oldsize
== bsize
)
3447 /* hit a fragment */
3451 * If we have reached the end of the current list without
3452 * finding the just finished dependency, then it must be
3453 * on the future dependency list. Future dependencies cannot
3454 * be freed until they are moved to the current list.
3456 if (listadp
== NULL
) {
3458 TAILQ_FOREACH(listadp
, &inodedep
->id_newinoupdt
, ad_next
)
3459 /* found our block */
3462 if (listadp
== NULL
)
3463 panic("handle_allocdirect_partdone: lost dep");
3468 * If we have found the just finished dependency, then free
3469 * it along with anything that follows it that is complete.
3471 for (; adp
; adp
= listadp
) {
3472 listadp
= TAILQ_NEXT(adp
, ad_next
);
3473 if ((adp
->ad_state
& ALLCOMPLETE
) != ALLCOMPLETE
)
3475 free_allocdirect(&inodedep
->id_inoupdt
, adp
, 1);
3480 * Called from within softdep_disk_write_complete above. Note that
3481 * this routine is always called from interrupt level with further
3482 * splbio interrupts blocked.
3485 * aip: the completed allocindir
3488 handle_allocindir_partdone(struct allocindir
*aip
)
3490 struct indirdep
*indirdep
;
3492 if ((aip
->ai_state
& ALLCOMPLETE
) != ALLCOMPLETE
)
3494 if (aip
->ai_buf
!= NULL
)
3495 panic("handle_allocindir_partdone: dangling dependency");
3497 indirdep
= aip
->ai_indirdep
;
3498 if (indirdep
->ir_state
& UNDONE
) {
3499 LIST_REMOVE(aip
, ai_next
);
3500 LIST_INSERT_HEAD(&indirdep
->ir_donehd
, aip
, ai_next
);
3503 ((ufs_daddr_t
*)indirdep
->ir_savebp
->b_data
)[aip
->ai_offset
] =
3505 LIST_REMOVE(aip
, ai_next
);
3506 if (aip
->ai_freefrag
!= NULL
)
3507 add_to_worklist(&aip
->ai_freefrag
->ff_list
);
3508 WORKITEM_FREE(aip
, D_ALLOCINDIR
);
3512 * Called from within softdep_disk_write_complete above to restore
3513 * in-memory inode block contents to their most up-to-date state. Note
3514 * that this routine is always called from interrupt level with further
3515 * splbio interrupts blocked.
3518 * bp: buffer containing the inode block
3521 handle_written_inodeblock(struct inodedep
*inodedep
, struct buf
*bp
)
3523 struct worklist
*wk
, *filefree
;
3524 struct allocdirect
*adp
, *nextadp
;
3525 struct ufs1_dinode
*dp
;
3528 if ((inodedep
->id_state
& IOSTARTED
) == 0)
3529 panic("handle_written_inodeblock: not started");
3531 inodedep
->id_state
&= ~IOSTARTED
;
3532 dp
= (struct ufs1_dinode
*)bp
->b_data
+
3533 ino_to_fsbo(inodedep
->id_fs
, inodedep
->id_ino
);
3535 * If we had to rollback the inode allocation because of
3536 * bitmaps being incomplete, then simply restore it.
3537 * Keep the block dirty so that it will not be reclaimed until
3538 * all associated dependencies have been cleared and the
3539 * corresponding updates written to disk.
3541 if (inodedep
->id_savedino
!= NULL
) {
3542 *dp
= *inodedep
->id_savedino
;
3543 kfree(inodedep
->id_savedino
, M_INODEDEP
);
3544 inodedep
->id_savedino
= NULL
;
3545 if ((bp
->b_flags
& B_DELWRI
) == 0)
3546 stat_inode_bitmap
++;
3550 inodedep
->id_state
|= COMPLETE
;
3552 * Roll forward anything that had to be rolled back before
3553 * the inode could be updated.
3556 for (adp
= TAILQ_FIRST(&inodedep
->id_inoupdt
); adp
; adp
= nextadp
) {
3557 nextadp
= TAILQ_NEXT(adp
, ad_next
);
3558 if (adp
->ad_state
& ATTACHED
)
3559 panic("handle_written_inodeblock: new entry");
3561 if (adp
->ad_lbn
< NDADDR
) {
3562 if (dp
->di_db
[adp
->ad_lbn
] != adp
->ad_oldblkno
) {
3563 panic("%s: %s #%ld mismatch %d != %d",
3564 "handle_written_inodeblock",
3565 "direct pointer", adp
->ad_lbn
,
3566 dp
->di_db
[adp
->ad_lbn
], adp
->ad_oldblkno
);
3568 dp
->di_db
[adp
->ad_lbn
] = adp
->ad_newblkno
;
3570 if (dp
->di_ib
[adp
->ad_lbn
- NDADDR
] != 0) {
3571 panic("%s: %s #%ld allocated as %d",
3572 "handle_written_inodeblock",
3573 "indirect pointer", adp
->ad_lbn
- NDADDR
,
3574 dp
->di_ib
[adp
->ad_lbn
- NDADDR
]);
3576 dp
->di_ib
[adp
->ad_lbn
- NDADDR
] = adp
->ad_newblkno
;
3578 adp
->ad_state
&= ~UNDONE
;
3579 adp
->ad_state
|= ATTACHED
;
3582 if (hadchanges
&& (bp
->b_flags
& B_DELWRI
) == 0)
3583 stat_direct_blk_ptrs
++;
3585 * Reset the file size to its most up-to-date value.
3587 if (inodedep
->id_savedsize
== -1) {
3588 panic("handle_written_inodeblock: bad size");
3590 if (dp
->di_size
!= inodedep
->id_savedsize
) {
3591 dp
->di_size
= inodedep
->id_savedsize
;
3594 inodedep
->id_savedsize
= -1;
3596 * If there were any rollbacks in the inode block, then it must be
3597 * marked dirty so that its will eventually get written back in
3603 * Process any allocdirects that completed during the update.
3605 if ((adp
= TAILQ_FIRST(&inodedep
->id_inoupdt
)) != NULL
)
3606 handle_allocdirect_partdone(adp
);
3608 * Process deallocations that were held pending until the
3609 * inode had been written to disk. Freeing of the inode
3610 * is delayed until after all blocks have been freed to
3611 * avoid creation of new <vfsid, inum, lbn> triples
3612 * before the old ones have been deleted.
3615 while ((wk
= LIST_FIRST(&inodedep
->id_bufwait
)) != NULL
) {
3616 WORKLIST_REMOVE(wk
);
3617 switch (wk
->wk_type
) {
3621 * We defer adding filefree to the worklist until
3622 * all other additions have been made to ensure
3623 * that it will be done after all the old blocks
3626 if (filefree
!= NULL
) {
3627 panic("handle_written_inodeblock: filefree");
3633 handle_written_mkdir(WK_MKDIR(wk
), MKDIR_PARENT
);
3637 diradd_inode_written(WK_DIRADD(wk
), inodedep
);
3641 wk
->wk_state
|= COMPLETE
;
3642 if ((wk
->wk_state
& ALLCOMPLETE
) != ALLCOMPLETE
)
3644 /* -- fall through -- */
3647 add_to_worklist(wk
);
3651 panic("handle_written_inodeblock: Unknown type %s",
3652 TYPENAME(wk
->wk_type
));
3656 if (filefree
!= NULL
) {
3657 if (free_inodedep(inodedep
) == 0) {
3658 panic("handle_written_inodeblock: live inodedep");
3660 add_to_worklist(filefree
);
3665 * If no outstanding dependencies, free it.
3667 if (free_inodedep(inodedep
) || TAILQ_FIRST(&inodedep
->id_inoupdt
) == NULL
)
3669 return (hadchanges
);
3673 * Process a diradd entry after its dependent inode has been written.
3674 * This routine must be called with splbio interrupts blocked.
3677 diradd_inode_written(struct diradd
*dap
, struct inodedep
*inodedep
)
3679 struct pagedep
*pagedep
;
3681 dap
->da_state
|= COMPLETE
;
3682 if ((dap
->da_state
& ALLCOMPLETE
) == ALLCOMPLETE
) {
3683 if (dap
->da_state
& DIRCHG
)
3684 pagedep
= dap
->da_previous
->dm_pagedep
;
3686 pagedep
= dap
->da_pagedep
;
3687 LIST_REMOVE(dap
, da_pdlist
);
3688 LIST_INSERT_HEAD(&pagedep
->pd_pendinghd
, dap
, da_pdlist
);
3690 WORKLIST_INSERT(&inodedep
->id_pendinghd
, &dap
->da_list
);
3694 * Handle the completion of a mkdir dependency.
3697 handle_written_mkdir(struct mkdir
*mkdir
, int type
)
3700 struct pagedep
*pagedep
;
3702 if (mkdir
->md_state
!= type
) {
3703 panic("handle_written_mkdir: bad type");
3705 dap
= mkdir
->md_diradd
;
3706 dap
->da_state
&= ~type
;
3707 if ((dap
->da_state
& (MKDIR_PARENT
| MKDIR_BODY
)) == 0)
3708 dap
->da_state
|= DEPCOMPLETE
;
3709 if ((dap
->da_state
& ALLCOMPLETE
) == ALLCOMPLETE
) {
3710 if (dap
->da_state
& DIRCHG
)
3711 pagedep
= dap
->da_previous
->dm_pagedep
;
3713 pagedep
= dap
->da_pagedep
;
3714 LIST_REMOVE(dap
, da_pdlist
);
3715 LIST_INSERT_HEAD(&pagedep
->pd_pendinghd
, dap
, da_pdlist
);
3717 LIST_REMOVE(mkdir
, md_mkdirs
);
3718 WORKITEM_FREE(mkdir
, D_MKDIR
);
3722 * Called from within softdep_disk_write_complete above.
3723 * A write operation was just completed. Removed inodes can
3724 * now be freed and associated block pointers may be committed.
3725 * Note that this routine is always called from interrupt level
3726 * with further splbio interrupts blocked.
3729 * bp: buffer containing the written page
3732 handle_written_filepage(struct pagedep
*pagedep
, struct buf
*bp
)
3734 struct dirrem
*dirrem
;
3735 struct diradd
*dap
, *nextdap
;
3739 if ((pagedep
->pd_state
& IOSTARTED
) == 0) {
3740 panic("handle_written_filepage: not started");
3742 pagedep
->pd_state
&= ~IOSTARTED
;
3744 * Process any directory removals that have been committed.
3746 while ((dirrem
= LIST_FIRST(&pagedep
->pd_dirremhd
)) != NULL
) {
3747 LIST_REMOVE(dirrem
, dm_next
);
3748 dirrem
->dm_dirinum
= pagedep
->pd_ino
;
3749 add_to_worklist(&dirrem
->dm_list
);
3752 * Free any directory additions that have been committed.
3754 while ((dap
= LIST_FIRST(&pagedep
->pd_pendinghd
)) != NULL
)
3757 * Uncommitted directory entries must be restored.
3759 for (chgs
= 0, i
= 0; i
< DAHASHSZ
; i
++) {
3760 for (dap
= LIST_FIRST(&pagedep
->pd_diraddhd
[i
]); dap
;
3762 nextdap
= LIST_NEXT(dap
, da_pdlist
);
3763 if (dap
->da_state
& ATTACHED
) {
3764 panic("handle_written_filepage: attached");
3766 ep
= (struct direct
*)
3767 ((char *)bp
->b_data
+ dap
->da_offset
);
3768 ep
->d_ino
= dap
->da_newinum
;
3769 dap
->da_state
&= ~UNDONE
;
3770 dap
->da_state
|= ATTACHED
;
3773 * If the inode referenced by the directory has
3774 * been written out, then the dependency can be
3775 * moved to the pending list.
3777 if ((dap
->da_state
& ALLCOMPLETE
) == ALLCOMPLETE
) {
3778 LIST_REMOVE(dap
, da_pdlist
);
3779 LIST_INSERT_HEAD(&pagedep
->pd_pendinghd
, dap
,
3785 * If there were any rollbacks in the directory, then it must be
3786 * marked dirty so that its will eventually get written back in
3790 if ((bp
->b_flags
& B_DELWRI
) == 0)
3795 * If no dependencies remain, the pagedep will be freed.
3796 * Otherwise it will remain to update the page before it
3797 * is written back to disk.
3799 if (LIST_FIRST(&pagedep
->pd_pendinghd
) == NULL
) {
3800 for (i
= 0; i
< DAHASHSZ
; i
++)
3801 if (LIST_FIRST(&pagedep
->pd_diraddhd
[i
]) != NULL
)
3803 if (i
== DAHASHSZ
) {
3804 LIST_REMOVE(pagedep
, pd_hash
);
3805 WORKITEM_FREE(pagedep
, D_PAGEDEP
);
3813 * Writing back in-core inode structures.
3815 * The filesystem only accesses an inode's contents when it occupies an
3816 * "in-core" inode structure. These "in-core" structures are separate from
3817 * the page frames used to cache inode blocks. Only the latter are
3818 * transferred to/from the disk. So, when the updated contents of the
3819 * "in-core" inode structure are copied to the corresponding in-memory inode
3820 * block, the dependencies are also transferred. The following procedure is
3821 * called when copying a dirty "in-core" inode to a cached inode block.
3825 * Called when an inode is loaded from disk. If the effective link count
3826 * differed from the actual link count when it was last flushed, then we
3827 * need to ensure that the correct effective link count is put back.
3830 * ip: the "in_core" copy of the inode
3833 softdep_load_inodeblock(struct inode
*ip
)
3835 struct inodedep
*inodedep
;
3838 * Check for alternate nlink count.
3840 ip
->i_effnlink
= ip
->i_nlink
;
3842 if (inodedep_lookup(ip
->i_fs
, ip
->i_number
, 0, &inodedep
) == 0) {
3846 ip
->i_effnlink
-= inodedep
->id_nlinkdelta
;
3851 * This routine is called just before the "in-core" inode
3852 * information is to be copied to the in-memory inode block.
3853 * Recall that an inode block contains several inodes. If
3854 * the force flag is set, then the dependencies will be
3855 * cleared so that the update can always be made. Note that
3856 * the buffer is locked when this routine is called, so we
3857 * will never be in the middle of writing the inode block
3861 * ip: the "in_core" copy of the inode
3862 * bp: the buffer containing the inode block
3863 * waitfor: nonzero => update must be allowed
3866 softdep_update_inodeblock(struct inode
*ip
, struct buf
*bp
,
3869 struct inodedep
*inodedep
;
3870 struct worklist
*wk
;
3875 * If the effective link count is not equal to the actual link
3876 * count, then we must track the difference in an inodedep while
3877 * the inode is (potentially) tossed out of the cache. Otherwise,
3878 * if there is no existing inodedep, then there are no dependencies
3882 if (inodedep_lookup(ip
->i_fs
, ip
->i_number
, 0, &inodedep
) == 0) {
3884 if (ip
->i_effnlink
!= ip
->i_nlink
)
3885 panic("softdep_update_inodeblock: bad link count");
3888 if (inodedep
->id_nlinkdelta
!= ip
->i_nlink
- ip
->i_effnlink
) {
3889 panic("softdep_update_inodeblock: bad delta");
3892 * Changes have been initiated. Anything depending on these
3893 * changes cannot occur until this inode has been written.
3895 inodedep
->id_state
&= ~COMPLETE
;
3896 if ((inodedep
->id_state
& ONWORKLIST
) == 0)
3897 WORKLIST_INSERT_BP(bp
, &inodedep
->id_list
);
3899 * Any new dependencies associated with the incore inode must
3900 * now be moved to the list associated with the buffer holding
3901 * the in-memory copy of the inode. Once merged process any
3902 * allocdirects that are completed by the merger.
3904 merge_inode_lists(inodedep
);
3905 if (TAILQ_FIRST(&inodedep
->id_inoupdt
) != NULL
)
3906 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep
->id_inoupdt
));
3908 * Now that the inode has been pushed into the buffer, the
3909 * operations dependent on the inode being written to disk
3910 * can be moved to the id_bufwait so that they will be
3911 * processed when the buffer I/O completes.
3913 while ((wk
= LIST_FIRST(&inodedep
->id_inowait
)) != NULL
) {
3914 WORKLIST_REMOVE(wk
);
3915 WORKLIST_INSERT(&inodedep
->id_bufwait
, wk
);
3918 * Newly allocated inodes cannot be written until the bitmap
3919 * that allocates them have been written (indicated by
3920 * DEPCOMPLETE being set in id_state). If we are doing a
3921 * forced sync (e.g., an fsync on a file), we force the bitmap
3922 * to be written so that the update can be done.
3929 if ((inodedep
->id_state
& DEPCOMPLETE
) != 0) {
3933 gotit
= getdirtybuf(&inodedep
->id_buf
, MNT_WAIT
);
3935 if (inodedep_lookup(ip
->i_fs
, ip
->i_number
, 0, &inodedep
) != 0)
3940 ibp
= inodedep
->id_buf
;
3942 if ((error
= bwrite(ibp
)) != 0)
3943 softdep_error("softdep_update_inodeblock: bwrite", error
);
3947 * Merge the new inode dependency list (id_newinoupdt) into the old
3948 * inode dependency list (id_inoupdt). This routine must be called
3949 * with splbio interrupts blocked.
3952 merge_inode_lists(struct inodedep
*inodedep
)
3954 struct allocdirect
*listadp
, *newadp
;
3956 newadp
= TAILQ_FIRST(&inodedep
->id_newinoupdt
);
3957 for (listadp
= TAILQ_FIRST(&inodedep
->id_inoupdt
); listadp
&& newadp
;) {
3958 if (listadp
->ad_lbn
< newadp
->ad_lbn
) {
3959 listadp
= TAILQ_NEXT(listadp
, ad_next
);
3962 TAILQ_REMOVE(&inodedep
->id_newinoupdt
, newadp
, ad_next
);
3963 TAILQ_INSERT_BEFORE(listadp
, newadp
, ad_next
);
3964 if (listadp
->ad_lbn
== newadp
->ad_lbn
) {
3965 allocdirect_merge(&inodedep
->id_inoupdt
, newadp
,
3969 newadp
= TAILQ_FIRST(&inodedep
->id_newinoupdt
);
3971 while ((newadp
= TAILQ_FIRST(&inodedep
->id_newinoupdt
)) != NULL
) {
3972 TAILQ_REMOVE(&inodedep
->id_newinoupdt
, newadp
, ad_next
);
3973 TAILQ_INSERT_TAIL(&inodedep
->id_inoupdt
, newadp
, ad_next
);
3978 * If we are doing an fsync, then we must ensure that any directory
3979 * entries for the inode have been written after the inode gets to disk.
3981 * bioops callback - hold io_token
3984 * vp: the "in_core" copy of the inode
3987 softdep_fsync(struct vnode
*vp
)
3989 struct inodedep
*inodedep
;
3990 struct pagedep
*pagedep
;
3991 struct worklist
*wk
;
3998 int error
, flushparent
;
4003 * Move check from original kernel code, possibly not needed any
4004 * more with the per-mount bioops.
4006 if ((vp
->v_mount
->mnt_flag
& MNT_SOFTDEP
) == 0)
4012 if (inodedep_lookup(fs
, ip
->i_number
, 0, &inodedep
) == 0) {
4016 if (LIST_FIRST(&inodedep
->id_inowait
) != NULL
||
4017 LIST_FIRST(&inodedep
->id_bufwait
) != NULL
||
4018 TAILQ_FIRST(&inodedep
->id_inoupdt
) != NULL
||
4019 TAILQ_FIRST(&inodedep
->id_newinoupdt
) != NULL
) {
4020 panic("softdep_fsync: pending ops");
4022 for (error
= 0, flushparent
= 0; ; ) {
4023 if ((wk
= LIST_FIRST(&inodedep
->id_pendinghd
)) == NULL
)
4025 if (wk
->wk_type
!= D_DIRADD
) {
4026 panic("softdep_fsync: Unexpected type %s",
4027 TYPENAME(wk
->wk_type
));
4029 dap
= WK_DIRADD(wk
);
4031 * Flush our parent if this directory entry
4032 * has a MKDIR_PARENT dependency.
4034 if (dap
->da_state
& DIRCHG
)
4035 pagedep
= dap
->da_previous
->dm_pagedep
;
4037 pagedep
= dap
->da_pagedep
;
4038 mnt
= pagedep
->pd_mnt
;
4039 parentino
= pagedep
->pd_ino
;
4040 lbn
= pagedep
->pd_lbn
;
4041 if ((dap
->da_state
& (MKDIR_BODY
| COMPLETE
)) != COMPLETE
) {
4042 panic("softdep_fsync: dirty");
4044 flushparent
= dap
->da_state
& MKDIR_PARENT
;
4046 * If we are being fsync'ed as part of vgone'ing this vnode,
4047 * then we will not be able to release and recover the
4048 * vnode below, so we just have to give up on writing its
4049 * directory entry out. It will eventually be written, just
4050 * not now, but then the user was not asking to have it
4051 * written, so we are not breaking any promises.
4053 if (vp
->v_flag
& VRECLAIMED
)
4056 * We prevent deadlock by always fetching inodes from the
4057 * root, moving down the directory tree. Thus, when fetching
4058 * our parent directory, we must unlock ourselves before
4059 * requesting the lock on our parent. See the comment in
4060 * ufs_lookup for details on possible races.
4064 error
= VFS_VGET(mnt
, NULL
, parentino
, &pvp
);
4065 vn_lock(vp
, LK_EXCLUSIVE
| LK_RETRY
);
4070 if ((error
= ffs_update(pvp
, 1)) != 0) {
4076 * Flush directory page containing the inode's name.
4078 error
= bread(pvp
, lblktodoff(fs
, lbn
), blksize(fs
, VTOI(pvp
), lbn
), &bp
);
4086 if (inodedep_lookup(fs
, ip
->i_number
, 0, &inodedep
) == 0)
4094 * Flush all the dirty bitmaps associated with the block device
4095 * before flushing the rest of the dirty blocks so as to reduce
4096 * the number of dependencies that will have to be rolled back.
4098 static int softdep_fsync_mountdev_bp(struct buf
*bp
, void *data
);
4101 softdep_fsync_mountdev(struct vnode
*vp
)
4103 if (!vn_isdisk(vp
, NULL
))
4104 panic("softdep_fsync_mountdev: vnode not a disk");
4106 lwkt_gettoken(&vp
->v_token
);
4107 RB_SCAN(buf_rb_tree
, &vp
->v_rbdirty_tree
, NULL
,
4108 softdep_fsync_mountdev_bp
, vp
);
4109 lwkt_reltoken(&vp
->v_token
);
4110 drain_output(vp
, 1);
4115 softdep_fsync_mountdev_bp(struct buf
*bp
, void *data
)
4117 struct worklist
*wk
;
4118 struct vnode
*vp
= data
;
4121 * If it is already scheduled, skip to the next buffer.
4123 if (BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
))
4125 if (bp
->b_vp
!= vp
|| (bp
->b_flags
& B_DELWRI
) == 0) {
4127 kprintf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp
, vp
);
4131 * We are only interested in bitmaps with outstanding
4134 if ((wk
= LIST_FIRST(&bp
->b_dep
)) == NULL
||
4135 wk
->wk_type
!= D_BMSAFEMAP
) {
4147 * This routine is called when we are trying to synchronously flush a
4148 * file. This routine must eliminate any filesystem metadata dependencies
4149 * so that the syncing routine can succeed by pushing the dirty blocks
4150 * associated with the file. If any I/O errors occur, they are returned.
4152 struct softdep_sync_metadata_info
{
4157 static int softdep_sync_metadata_bp(struct buf
*bp
, void *data
);
4160 softdep_sync_metadata(struct vnode
*vp
, struct thread
*td
)
4162 struct softdep_sync_metadata_info info
;
4166 * Check whether this vnode is involved in a filesystem
4167 * that is doing soft dependency processing.
4169 if (!vn_isdisk(vp
, NULL
)) {
4170 if (!DOINGSOFTDEP(vp
))
4173 if (vp
->v_rdev
->si_mountpoint
== NULL
||
4174 (vp
->v_rdev
->si_mountpoint
->mnt_flag
& MNT_SOFTDEP
) == 0)
4177 * Ensure that any direct block dependencies have been cleared.
4180 if ((error
= flush_inodedep_deps(VTOI(vp
)->i_fs
, VTOI(vp
)->i_number
))) {
4185 * For most files, the only metadata dependencies are the
4186 * cylinder group maps that allocate their inode or blocks.
4187 * The block allocation dependencies can be found by traversing
4188 * the dependency lists for any buffers that remain on their
4189 * dirty buffer list. The inode allocation dependency will
4190 * be resolved when the inode is updated with MNT_WAIT.
4191 * This work is done in two passes. The first pass grabs most
4192 * of the buffers and begins asynchronously writing them. The
4193 * only way to wait for these asynchronous writes is to sleep
4194 * on the filesystem vnode which may stay busy for a long time
4195 * if the filesystem is active. So, instead, we make a second
4196 * pass over the dependencies blocking on each write. In the
4197 * usual case we will be blocking against a write that we
4198 * initiated, so when it is done the dependency will have been
4199 * resolved. Thus the second pass is expected to end quickly.
4201 waitfor
= MNT_NOWAIT
;
4204 * We must wait for any I/O in progress to finish so that
4205 * all potential buffers on the dirty list will be visible.
4207 drain_output(vp
, 1);
4210 info
.waitfor
= waitfor
;
4211 lwkt_gettoken(&vp
->v_token
);
4212 error
= RB_SCAN(buf_rb_tree
, &vp
->v_rbdirty_tree
, NULL
,
4213 softdep_sync_metadata_bp
, &info
);
4214 lwkt_reltoken(&vp
->v_token
);
4217 return(-error
); /* error code */
4221 * The brief unlock is to allow any pent up dependency
4222 * processing to be done. Then proceed with the second pass.
4224 if (waitfor
& MNT_NOWAIT
) {
4232 * If we have managed to get rid of all the dirty buffers,
4233 * then we are done. For certain directories and block
4234 * devices, we may need to do further work.
4236 * We must wait for any I/O in progress to finish so that
4237 * all potential buffers on the dirty list will be visible.
4239 drain_output(vp
, 1);
4240 if (RB_EMPTY(&vp
->v_rbdirty_tree
)) {
4247 * If we are trying to sync a block device, some of its buffers may
4248 * contain metadata that cannot be written until the contents of some
4249 * partially written files have been written to disk. The only easy
4250 * way to accomplish this is to sync the entire filesystem (luckily
4251 * this happens rarely).
4253 if (vn_isdisk(vp
, NULL
) &&
4255 vp
->v_rdev
->si_mountpoint
&& !vn_islocked(vp
) &&
4256 (error
= VFS_SYNC(vp
->v_rdev
->si_mountpoint
, MNT_WAIT
)) != 0)
4262 softdep_sync_metadata_bp(struct buf
*bp
, void *data
)
4264 struct softdep_sync_metadata_info
*info
= data
;
4265 struct pagedep
*pagedep
;
4266 struct allocdirect
*adp
;
4267 struct allocindir
*aip
;
4268 struct worklist
*wk
;
4273 if (getdirtybuf(&bp
, MNT_WAIT
) == 0) {
4274 kprintf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp
);
4277 if (bp
->b_vp
!= info
->vp
|| (bp
->b_flags
& B_DELWRI
) == 0) {
4278 kprintf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp
, info
->vp
);
4284 * As we hold the buffer locked, none of its dependencies
4287 LIST_FOREACH(wk
, &bp
->b_dep
, wk_list
) {
4288 switch (wk
->wk_type
) {
4291 adp
= WK_ALLOCDIRECT(wk
);
4292 if (adp
->ad_state
& DEPCOMPLETE
)
4295 if (getdirtybuf(&nbp
, info
->waitfor
) == 0)
4298 if (info
->waitfor
& MNT_NOWAIT
) {
4300 } else if ((error
= bwrite(nbp
)) != 0) {
4309 aip
= WK_ALLOCINDIR(wk
);
4310 if (aip
->ai_state
& DEPCOMPLETE
)
4313 if (getdirtybuf(&nbp
, info
->waitfor
) == 0)
4316 if (info
->waitfor
& MNT_NOWAIT
) {
4318 } else if ((error
= bwrite(nbp
)) != 0) {
4329 LIST_FOREACH(aip
, &WK_INDIRDEP(wk
)->ir_deplisthd
, ai_next
) {
4330 if (aip
->ai_state
& DEPCOMPLETE
)
4333 if (getdirtybuf(&nbp
, MNT_WAIT
) == 0)
4336 if ((error
= bwrite(nbp
)) != 0) {
4347 if ((error
= flush_inodedep_deps(WK_INODEDEP(wk
)->id_fs
,
4348 WK_INODEDEP(wk
)->id_ino
)) != 0) {
4358 * We are trying to sync a directory that may
4359 * have dependencies on both its own metadata
4360 * and/or dependencies on the inodes of any
4361 * recently allocated files. We walk its diradd
4362 * lists pushing out the associated inode.
4364 pagedep
= WK_PAGEDEP(wk
);
4365 for (i
= 0; i
< DAHASHSZ
; i
++) {
4366 if (LIST_FIRST(&pagedep
->pd_diraddhd
[i
]) == NULL
)
4369 flush_pagedep_deps(info
->vp
,
4371 &pagedep
->pd_diraddhd
[i
]))) {
4382 * This case should never happen if the vnode has
4383 * been properly sync'ed. However, if this function
4384 * is used at a place where the vnode has not yet
4385 * been sync'ed, this dependency can show up. So,
4386 * rather than panic, just flush it.
4388 nbp
= WK_MKDIR(wk
)->md_buf
;
4389 if (getdirtybuf(&nbp
, info
->waitfor
) == 0)
4392 if (info
->waitfor
& MNT_NOWAIT
) {
4394 } else if ((error
= bwrite(nbp
)) != 0) {
4404 * This case should never happen if the vnode has
4405 * been properly sync'ed. However, if this function
4406 * is used at a place where the vnode has not yet
4407 * been sync'ed, this dependency can show up. So,
4408 * rather than panic, just flush it.
4410 * nbp can wind up == bp if a device node for the
4411 * same filesystem is being fsynced at the same time,
4412 * leading to a panic if we don't catch the case.
4414 nbp
= WK_BMSAFEMAP(wk
)->sm_buf
;
4417 if (getdirtybuf(&nbp
, info
->waitfor
) == 0)
4420 if (info
->waitfor
& MNT_NOWAIT
) {
4422 } else if ((error
= bwrite(nbp
)) != 0) {
4431 panic("softdep_sync_metadata: Unknown type %s",
4432 TYPENAME(wk
->wk_type
));
4443 * Flush the dependencies associated with an inodedep.
4444 * Called with splbio blocked.
4447 flush_inodedep_deps(struct fs
*fs
, ino_t ino
)
4449 struct inodedep
*inodedep
;
4450 struct allocdirect
*adp
;
4455 * This work is done in two passes. The first pass grabs most
4456 * of the buffers and begins asynchronously writing them. The
4457 * only way to wait for these asynchronous writes is to sleep
4458 * on the filesystem vnode which may stay busy for a long time
4459 * if the filesystem is active. So, instead, we make a second
4460 * pass over the dependencies blocking on each write. In the
4461 * usual case we will be blocking against a write that we
4462 * initiated, so when it is done the dependency will have been
4463 * resolved. Thus the second pass is expected to end quickly.
4464 * We give a brief window at the top of the loop to allow
4465 * any pending I/O to complete.
4467 for (waitfor
= MNT_NOWAIT
; ; ) {
4470 if (inodedep_lookup(fs
, ino
, 0, &inodedep
) == 0)
4472 TAILQ_FOREACH(adp
, &inodedep
->id_inoupdt
, ad_next
) {
4473 if (adp
->ad_state
& DEPCOMPLETE
)
4476 if (getdirtybuf(&bp
, waitfor
) == 0) {
4477 if (waitfor
& MNT_NOWAIT
)
4482 if (waitfor
& MNT_NOWAIT
) {
4484 } else if ((error
= bwrite(bp
)) != 0) {
4493 TAILQ_FOREACH(adp
, &inodedep
->id_newinoupdt
, ad_next
) {
4494 if (adp
->ad_state
& DEPCOMPLETE
)
4497 if (getdirtybuf(&bp
, waitfor
) == 0) {
4498 if (waitfor
& MNT_NOWAIT
)
4503 if (waitfor
& MNT_NOWAIT
) {
4505 } else if ((error
= bwrite(bp
)) != 0) {
4515 * If pass2, we are done, otherwise do pass 2.
4517 if (waitfor
== MNT_WAIT
)
4522 * Try freeing inodedep in case all dependencies have been removed.
4524 if (inodedep_lookup(fs
, ino
, 0, &inodedep
) != 0)
4525 (void) free_inodedep(inodedep
);
4530 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
4531 * Called with splbio blocked.
4534 flush_pagedep_deps(struct vnode
*pvp
, struct mount
*mp
,
4535 struct diraddhd
*diraddhdp
)
4537 struct inodedep
*inodedep
;
4538 struct ufsmount
*ump
;
4540 struct worklist
*wk
;
4542 int gotit
, error
= 0;
4547 while ((dap
= LIST_FIRST(diraddhdp
)) != NULL
) {
4549 * Flush ourselves if this directory entry
4550 * has a MKDIR_PARENT dependency.
4552 if (dap
->da_state
& MKDIR_PARENT
) {
4554 if ((error
= ffs_update(pvp
, 1)) != 0)
4558 * If that cleared dependencies, go on to next.
4560 if (dap
!= LIST_FIRST(diraddhdp
))
4562 if (dap
->da_state
& MKDIR_PARENT
) {
4563 panic("flush_pagedep_deps: MKDIR_PARENT");
4567 * A newly allocated directory must have its "." and
4568 * ".." entries written out before its name can be
4569 * committed in its parent. We do not want or need
4570 * the full semantics of a synchronous VOP_FSYNC as
4571 * that may end up here again, once for each directory
4572 * level in the filesystem. Instead, we push the blocks
4573 * and wait for them to clear. We have to fsync twice
4574 * because the first call may choose to defer blocks
4575 * that still have dependencies, but deferral will
4576 * happen at most once.
4578 inum
= dap
->da_newinum
;
4579 if (dap
->da_state
& MKDIR_BODY
) {
4581 if ((error
= VFS_VGET(mp
, NULL
, inum
, &vp
)) != 0)
4583 if ((error
=VOP_FSYNC(vp
, MNT_NOWAIT
, 0)) ||
4584 (error
=VOP_FSYNC(vp
, MNT_NOWAIT
, 0))) {
4588 drain_output(vp
, 0);
4590 * If first block is still dirty with a D_MKDIR
4591 * dependency then it needs to be written now.
4595 bp
= findblk(vp
, 0, FINDBLK_TEST
);
4598 goto mkdir_body_continue
;
4600 LIST_FOREACH(wk
, &bp
->b_dep
, wk_list
)
4601 if (wk
->wk_type
== D_MKDIR
) {
4602 gotit
= getdirtybuf(&bp
, MNT_WAIT
);
4604 if (gotit
&& (error
= bwrite(bp
)) != 0)
4605 goto mkdir_body_continue
;
4610 mkdir_body_continue
:
4612 /* Flushing of first block failed. */
4617 * If that cleared dependencies, go on to next.
4619 if (dap
!= LIST_FIRST(diraddhdp
))
4621 if (dap
->da_state
& MKDIR_BODY
) {
4622 panic("flush_pagedep_deps: %p MKDIR_BODY", dap
);
4626 * Flush the inode on which the directory entry depends.
4627 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
4628 * the only remaining dependency is that the updated inode
4629 * count must get pushed to disk. The inode has already
4630 * been pushed into its inode buffer (via VOP_UPDATE) at
4631 * the time of the reference count change. So we need only
4632 * locate that buffer, ensure that there will be no rollback
4633 * caused by a bitmap dependency, then write the inode buffer.
4636 if (inodedep_lookup(ump
->um_fs
, inum
, 0, &inodedep
) == 0) {
4637 panic("flush_pagedep_deps: lost inode");
4640 * If the inode still has bitmap dependencies,
4641 * push them to disk.
4643 if ((inodedep
->id_state
& DEPCOMPLETE
) == 0) {
4644 gotit
= getdirtybuf(&inodedep
->id_buf
, MNT_WAIT
);
4648 if (gotit
&& (error
= bwrite(inodedep
->id_buf
)) != 0)
4651 if (dap
!= LIST_FIRST(diraddhdp
))
4655 * If the inode is still sitting in a buffer waiting
4656 * to be written, push it to disk.
4659 if ((error
= bread(ump
->um_devvp
,
4660 fsbtodoff(ump
->um_fs
, ino_to_fsba(ump
->um_fs
, inum
)),
4661 (int)ump
->um_fs
->fs_bsize
, &bp
)) != 0)
4663 if ((error
= bwrite(bp
)) != 0)
4667 * If we have failed to get rid of all the dependencies
4668 * then something is seriously wrong.
4670 if (dap
== LIST_FIRST(diraddhdp
)) {
4671 panic("flush_pagedep_deps: flush failed");
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.
4687 softdep_slowdown(struct vnode
*vp
)
4689 int max_softdeps_hard
;
4691 max_softdeps_hard
= max_softdeps
* 11 / 10;
4692 if (num_dirrem
< max_softdeps_hard
/ 2 &&
4693 num_inodedep
< max_softdeps_hard
)
4695 stat_sync_limit_hit
+= 1;
4700 * If memory utilization has gotten too high, deliberately slow things
4701 * down and speed up the I/O processing.
4704 request_cleanup(int resource
)
4706 struct thread
*td
= curthread
; /* XXX */
4708 KKASSERT(lock_held(&lk
) > 0);
4711 * We never hold up the filesystem syncer process.
4713 if (td
== filesys_syncer
)
4716 * First check to see if the work list has gotten backlogged.
4717 * If it has, co-opt this process to help clean up two entries.
4718 * Because this process may hold inodes locked, we cannot
4719 * handle any remove requests that might block on a locked
4720 * inode as that could lead to deadlock.
4722 if (num_on_worklist
> max_softdeps
/ 10) {
4723 process_worklist_item(NULL
, LK_NOWAIT
);
4724 process_worklist_item(NULL
, LK_NOWAIT
);
4725 stat_worklist_push
+= 2;
4730 * If we are resource constrained on inode dependencies, try
4731 * flushing some dirty inodes. Otherwise, we are constrained
4732 * by file deletions, so try accelerating flushes of directories
4733 * with removal dependencies. We would like to do the cleanup
4734 * here, but we probably hold an inode locked at this point and
4735 * that might deadlock against one that we try to clean. So,
4736 * the best that we can do is request the syncer daemon to do
4737 * the cleanup for us.
4742 stat_ino_limit_push
+= 1;
4743 req_clear_inodedeps
+= 1;
4744 stat_countp
= &stat_ino_limit_hit
;
4748 stat_blk_limit_push
+= 1;
4749 req_clear_remove
+= 1;
4750 stat_countp
= &stat_blk_limit_hit
;
4754 panic("request_cleanup: unknown type");
4757 * Hopefully the syncer daemon will catch up and awaken us.
4758 * We wait at most tickdelay before proceeding in any case.
4760 lksleep(&proc_waiting
, &lk
, 0, "softupdate",
4761 tickdelay
> 2 ? tickdelay
: 2);
4766 * Flush out a directory with at least one removal dependency in an effort to
4767 * reduce the number of dirrem, freefile, and freeblks dependency structures.
4770 clear_remove(struct thread
*td
)
4772 struct pagedep_hashhead
*pagedephd
;
4773 struct pagedep
*pagedep
;
4774 static int next
= 0;
4781 for (cnt
= 0; cnt
< pagedep_hash
; cnt
++) {
4782 pagedephd
= &pagedep_hashtbl
[next
++];
4783 if (next
>= pagedep_hash
)
4785 LIST_FOREACH(pagedep
, pagedephd
, pd_hash
) {
4786 if (LIST_FIRST(&pagedep
->pd_dirremhd
) == NULL
)
4788 mp
= pagedep
->pd_mnt
;
4789 ino
= pagedep
->pd_ino
;
4791 if ((error
= VFS_VGET(mp
, NULL
, ino
, &vp
)) != 0) {
4792 softdep_error("clear_remove: vget", error
);
4795 if ((error
= VOP_FSYNC(vp
, MNT_NOWAIT
, 0)))
4796 softdep_error("clear_remove: fsync", error
);
4797 drain_output(vp
, 0);
4806 * Clear out a block of dirty inodes in an effort to reduce
4807 * the number of inodedep dependency structures.
4809 struct clear_inodedeps_info
{
4815 clear_inodedeps_mountlist_callback(struct mount
*mp
, void *data
)
4817 struct clear_inodedeps_info
*info
= data
;
4819 if ((mp
->mnt_flag
& MNT_SOFTDEP
) && info
->fs
== VFSTOUFS(mp
)->um_fs
) {
4827 clear_inodedeps(struct thread
*td
)
4829 struct clear_inodedeps_info info
;
4830 struct inodedep_hashhead
*inodedephd
;
4831 struct inodedep
*inodedep
;
4832 static int next
= 0;
4836 ino_t firstino
, lastino
, ino
;
4840 * Pick a random inode dependency to be cleared.
4841 * We will then gather up all the inodes in its block
4842 * that have dependencies and flush them out.
4844 inodedep
= NULL
; /* avoid gcc warnings */
4845 for (cnt
= 0; cnt
< inodedep_hash
; cnt
++) {
4846 inodedephd
= &inodedep_hashtbl
[next
++];
4847 if (next
>= inodedep_hash
)
4849 if ((inodedep
= LIST_FIRST(inodedephd
)) != NULL
)
4852 if (inodedep
== NULL
) {
4857 * Ugly code to find mount point given pointer to superblock.
4859 fs
= inodedep
->id_fs
;
4862 mountlist_scan(clear_inodedeps_mountlist_callback
,
4863 &info
, MNTSCAN_FORWARD
|MNTSCAN_NOBUSY
);
4865 * Find the last inode in the block with dependencies.
4867 firstino
= inodedep
->id_ino
& ~(INOPB(fs
) - 1);
4868 for (lastino
= firstino
+ INOPB(fs
) - 1; lastino
> firstino
; lastino
--)
4869 if (inodedep_lookup(fs
, lastino
, 0, &inodedep
) != 0)
4872 * Asynchronously push all but the last inode with dependencies.
4873 * Synchronously push the last inode with dependencies to ensure
4874 * that the inode block gets written to free up the inodedeps.
4876 for (ino
= firstino
; ino
<= lastino
; ino
++) {
4877 if (inodedep_lookup(fs
, ino
, 0, &inodedep
) == 0)
4880 if ((error
= VFS_VGET(info
.mp
, NULL
, ino
, &vp
)) != 0) {
4881 softdep_error("clear_inodedeps: vget", error
);
4884 if (ino
== lastino
) {
4885 if ((error
= VOP_FSYNC(vp
, MNT_WAIT
, 0)))
4886 softdep_error("clear_inodedeps: fsync1", error
);
4888 if ((error
= VOP_FSYNC(vp
, MNT_NOWAIT
, 0)))
4889 softdep_error("clear_inodedeps: fsync2", error
);
4890 drain_output(vp
, 0);
4899 * Function to determine if the buffer has outstanding dependencies
4900 * that will cause a roll-back if the buffer is written. If wantcount
4901 * is set, return number of dependencies, otherwise just yes or no.
4903 * bioops callback - hold io_token
4906 softdep_count_dependencies(struct buf
*bp
, int wantcount
)
4908 struct worklist
*wk
;
4909 struct inodedep
*inodedep
;
4910 struct indirdep
*indirdep
;
4911 struct allocindir
*aip
;
4912 struct pagedep
*pagedep
;
4919 LIST_FOREACH(wk
, &bp
->b_dep
, wk_list
) {
4920 switch (wk
->wk_type
) {
4923 inodedep
= WK_INODEDEP(wk
);
4924 if ((inodedep
->id_state
& DEPCOMPLETE
) == 0) {
4925 /* bitmap allocation dependency */
4930 if (TAILQ_FIRST(&inodedep
->id_inoupdt
)) {
4931 /* direct block pointer dependency */
4939 indirdep
= WK_INDIRDEP(wk
);
4941 LIST_FOREACH(aip
, &indirdep
->ir_deplisthd
, ai_next
) {
4942 /* indirect block pointer dependency */
4950 pagedep
= WK_PAGEDEP(wk
);
4951 for (i
= 0; i
< DAHASHSZ
; i
++) {
4953 LIST_FOREACH(dap
, &pagedep
->pd_diraddhd
[i
], da_pdlist
) {
4954 /* directory entry dependency */
4966 /* never a dependency on these blocks */
4970 panic("softdep_check_for_rollback: Unexpected type %s",
4971 TYPENAME(wk
->wk_type
));
4982 * Acquire exclusive access to a buffer. Requires softdep lock
4983 * to be held on entry. If waitfor is MNT_WAIT, may release/reacquire
4986 * Returns 1 if the buffer was locked, 0 if it was not locked or
4987 * if we had to block.
4989 * NOTE! In order to return 1 we must acquire the buffer lock prior
4990 * to any release of &lk. Once we release &lk it's all over.
4991 * We may still have to block on the (type-stable) bp in that
4992 * case, but we must then unlock it and return 0.
4995 getdirtybuf(struct buf
**bpp
, int waitfor
)
5001 * If the contents of *bpp is NULL the caller presumably lost a race.
5008 * Try to obtain the buffer lock without deadlocking on &lk.
5010 KKASSERT(lock_held(&lk
) > 0);
5011 error
= BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_NOWAIT
);
5014 * If the buffer is no longer dirty the OS already wrote it
5015 * out, return failure.
5017 if ((bp
->b_flags
& B_DELWRI
) == 0) {
5023 * Finish nominal buffer locking sequence return success.
5032 * If we are not being asked to wait, return 0 immediately.
5034 if (waitfor
!= MNT_WAIT
)
5038 * Once we release the softdep lock we can never return success,
5039 * but we still have to block on the type-stable buf for the caller
5040 * to be able to retry without livelocking the system.
5042 * The caller will normally retry in this case.
5045 error
= BUF_LOCK(bp
, LK_EXCLUSIVE
| LK_SLEEPFAIL
);
5053 * Wait for pending output on a vnode to complete.
5054 * Must be called with vnode locked.
5057 drain_output(struct vnode
*vp
, int islocked
)
5062 while (bio_track_active(&vp
->v_track_write
)) {
5064 bio_track_wait(&vp
->v_track_write
, 0, 0);
5072 * Called whenever a buffer that is being invalidated or reallocated
5073 * contains dependencies. This should only happen if an I/O error has
5074 * occurred. The routine is called with the buffer locked.
5076 * bioops callback - hold io_token
5079 softdep_deallocate_dependencies(struct buf
*bp
)
5081 /* nothing to do, mp lock not needed */
5082 if ((bp
->b_flags
& B_ERROR
) == 0)
5083 panic("softdep_deallocate_dependencies: dangling deps");
5084 softdep_error(bp
->b_vp
->v_mount
->mnt_stat
.f_mntfromname
, bp
->b_error
);
5085 panic("softdep_deallocate_dependencies: unrecovered I/O error");
5089 * Function to handle asynchronous write errors in the filesystem.
5092 softdep_error(char *func
, int error
)
5094 /* XXX should do something better! */
5095 kprintf("%s: got error %d while accessing filesystem\n", func
, error
);