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 $
40 * $DragonFly: src/sys/vfs/ufs/ffs_softdep.c,v 1.4 2003/06/25 03:56:11 dillon Exp $
44 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
53 #include <sys/param.h>
54 #include <sys/kernel.h>
55 #include <sys/systm.h>
57 #include <sys/malloc.h>
58 #include <sys/mount.h>
60 #include <sys/syslog.h>
61 #include <sys/vnode.h>
64 #include <ufs/ufs/dir.h>
65 #include <ufs/ufs/quota.h>
66 #include <ufs/ufs/inode.h>
67 #include <ufs/ufs/ufsmount.h>
68 #include <ufs/ffs/fs.h>
69 #include <ufs/ffs/softdep.h>
70 #include <ufs/ffs/ffs_extern.h>
71 #include <ufs/ufs/ufs_extern.h>
74 * These definitions need to be adapted to the system to which
75 * this file is being ported.
78 * malloc types defined for the softdep system.
80 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
81 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
82 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
83 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
84 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
85 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
86 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
87 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
88 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
89 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
90 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
91 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
92 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
94 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE)
100 #define D_ALLOCDIRECT 4
102 #define D_ALLOCINDIR 6
109 #define D_LAST D_DIRREM
112 * translate from workitem type to memory type
113 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
115 static struct malloc_type *memtype[] = {
131 #define DtoM(type) (memtype[type])
134 * Names of malloc types.
136 #define TYPENAME(type) \
137 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
139 * End system adaptaion definitions.
143 * Internal function prototypes.
145 static void softdep_error __P((char *, int));
146 static void drain_output __P((struct vnode *, int));
147 static int getdirtybuf __P((struct buf **, int));
148 static void clear_remove __P((struct thread *));
149 static void clear_inodedeps __P((struct thread *));
150 static int flush_pagedep_deps __P((struct vnode *, struct mount *,
152 static int flush_inodedep_deps __P((struct fs *, ino_t));
153 static int handle_written_filepage __P((struct pagedep *, struct buf *));
154 static void diradd_inode_written __P((struct diradd *, struct inodedep *));
155 static int handle_written_inodeblock __P((struct inodedep *, struct buf *));
156 static void handle_allocdirect_partdone __P((struct allocdirect *));
157 static void handle_allocindir_partdone __P((struct allocindir *));
158 static void initiate_write_filepage __P((struct pagedep *, struct buf *));
159 static void handle_written_mkdir __P((struct mkdir *, int));
160 static void initiate_write_inodeblock __P((struct inodedep *, struct buf *));
161 static void handle_workitem_freefile __P((struct freefile *));
162 static void handle_workitem_remove __P((struct dirrem *));
163 static struct dirrem *newdirrem __P((struct buf *, struct inode *,
164 struct inode *, int, struct dirrem **));
165 static void free_diradd __P((struct diradd *));
166 static void free_allocindir __P((struct allocindir *, struct inodedep *));
167 static int indir_trunc __P((struct inode *, ufs_daddr_t, int, ufs_lbn_t,
169 static void deallocate_dependencies __P((struct buf *, struct inodedep *));
170 static void free_allocdirect __P((struct allocdirectlst *,
171 struct allocdirect *, int));
172 static int check_inode_unwritten __P((struct inodedep *));
173 static int free_inodedep __P((struct inodedep *));
174 static void handle_workitem_freeblocks __P((struct freeblks *));
175 static void merge_inode_lists __P((struct inodedep *));
176 static void setup_allocindir_phase2 __P((struct buf *, struct inode *,
177 struct allocindir *));
178 static struct allocindir *newallocindir __P((struct inode *, int, ufs_daddr_t,
180 static void handle_workitem_freefrag __P((struct freefrag *));
181 static struct freefrag *newfreefrag __P((struct inode *, ufs_daddr_t, long));
182 static void allocdirect_merge __P((struct allocdirectlst *,
183 struct allocdirect *, struct allocdirect *));
184 static struct bmsafemap *bmsafemap_lookup __P((struct buf *));
185 static int newblk_lookup __P((struct fs *, ufs_daddr_t, int,
187 static int inodedep_lookup __P((struct fs *, ino_t, int, struct inodedep **));
188 static int pagedep_lookup __P((struct inode *, ufs_lbn_t, int,
190 static void pause_timer __P((void *));
191 static int request_cleanup __P((int, int));
192 static int process_worklist_item __P((struct mount *, int));
193 static void add_to_worklist __P((struct worklist *));
196 * Exported softdep operations.
198 static void softdep_disk_io_initiation __P((struct buf *));
199 static void softdep_disk_write_complete __P((struct buf *));
200 static void softdep_deallocate_dependencies __P((struct buf *));
201 static int softdep_fsync __P((struct vnode *));
202 static int softdep_process_worklist __P((struct mount *));
203 static void softdep_move_dependencies __P((struct buf *, struct buf *));
204 static int softdep_count_dependencies __P((struct buf *bp, int));
206 struct bio_ops bioops = {
207 softdep_disk_io_initiation, /* io_start */
208 softdep_disk_write_complete, /* io_complete */
209 softdep_deallocate_dependencies, /* io_deallocate */
210 softdep_fsync, /* io_fsync */
211 softdep_process_worklist, /* io_sync */
212 softdep_move_dependencies, /* io_movedeps */
213 softdep_count_dependencies, /* io_countdeps */
217 * Locking primitives.
219 * For a uniprocessor, all we need to do is protect against disk
220 * interrupts. For a multiprocessor, this lock would have to be
221 * a mutex. A single mutex is used throughout this file, though
222 * finer grain locking could be used if contention warranted it.
224 * For a multiprocessor, the sleep call would accept a lock and
225 * release it after the sleep processing was complete. In a uniprocessor
226 * implementation there is no such interlock, so we simple mark
227 * the places where it needs to be done with the `interlocked' form
228 * of the lock calls. Since the uniprocessor sleep already interlocks
229 * the spl, there is nothing that really needs to be done.
231 #ifndef /* NOT */ DEBUG
232 static struct lockit {
235 #define ACQUIRE_LOCK(lk) (lk)->lkt_spl = splbio()
236 #define FREE_LOCK(lk) splx((lk)->lkt_spl)
239 #define NOHOLDER ((struct thread *)-1)
240 #define SPECIAL_FLAG ((struct thread *)-2)
241 static struct lockit {
243 struct thread *lkt_held;
244 } lk = { 0, NOHOLDER };
247 static void acquire_lock __P((struct lockit *));
248 static void free_lock __P((struct lockit *));
249 void softdep_panic __P((char *));
251 #define ACQUIRE_LOCK(lk) acquire_lock(lk)
252 #define FREE_LOCK(lk) free_lock(lk)
260 if (lk->lkt_held != NOHOLDER) {
261 holder = lk->lkt_held;
263 if (holder == curthread)
264 panic("softdep_lock: locking against myself");
266 panic("softdep_lock: lock held by %p", holder);
268 lk->lkt_spl = splbio();
269 lk->lkt_held = curthread;
278 if (lk->lkt_held == NOHOLDER)
279 panic("softdep_unlock: lock not held");
280 lk->lkt_held = NOHOLDER;
285 * Function to release soft updates lock and panic.
292 if (lk.lkt_held != NOHOLDER)
298 static int interlocked_sleep __P((struct lockit *, int, void *, int,
302 * When going to sleep, we must save our SPL so that it does
303 * not get lost if some other process uses the lock while we
304 * are sleeping. We restore it after we have slept. This routine
305 * wraps the interlocking with functions that sleep. The list
306 * below enumerates the available set of operations.
313 interlocked_sleep(lk, op, ident, flags, wmesg, timo)
326 if (lk->lkt_held == NOHOLDER)
327 panic("interlocked_sleep: lock not held");
328 lk->lkt_held = NOHOLDER;
332 retval = tsleep(ident, flags, wmesg, timo);
335 retval = BUF_LOCK((struct buf *)ident, flags);
338 panic("interlocked_sleep: unknown operation");
341 if (lk->lkt_held != NOHOLDER) {
342 holder = lk->lkt_held;
344 if (holder == curthread)
345 panic("interlocked_sleep: locking against self");
347 panic("interlocked_sleep: lock held by %p", holder);
349 lk->lkt_held = curthread;
357 * Place holder for real semaphores.
366 static void sema_init __P((struct sema *, char *, int, int));
367 static int sema_get __P((struct sema *, struct lockit *));
368 static void sema_release __P((struct sema *));
371 sema_init(semap, name, prio, timo)
377 semap->holder = NOHOLDER;
385 sema_get(semap, interlock)
387 struct lockit *interlock;
390 if (semap->value++ > 0) {
391 if (interlock != NULL) {
392 interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
393 semap->prio, semap->name, semap->timo);
394 FREE_LOCK(interlock);
396 tsleep((caddr_t)semap, semap->prio, semap->name,
401 semap->holder = curthread;
402 if (interlock != NULL)
403 FREE_LOCK(interlock);
412 if (semap->value <= 0 || semap->holder != curthread) {
413 if (lk.lkt_held != NOHOLDER)
415 panic("sema_release: not held");
417 if (--semap->value > 0) {
421 semap->holder = NOHOLDER;
425 * Worklist queue management.
426 * These routines require that the lock be held.
428 #ifndef /* NOT */ DEBUG
429 #define WORKLIST_INSERT(head, item) do { \
430 (item)->wk_state |= ONWORKLIST; \
431 LIST_INSERT_HEAD(head, item, wk_list); \
433 #define WORKLIST_REMOVE(item) do { \
434 (item)->wk_state &= ~ONWORKLIST; \
435 LIST_REMOVE(item, wk_list); \
437 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
440 static void worklist_insert __P((struct workhead *, struct worklist *));
441 static void worklist_remove __P((struct worklist *));
442 static void workitem_free __P((struct worklist *, int));
444 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
445 #define WORKLIST_REMOVE(item) worklist_remove(item)
446 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
449 worklist_insert(head, item)
450 struct workhead *head;
451 struct worklist *item;
454 if (lk.lkt_held == NOHOLDER)
455 panic("worklist_insert: lock not held");
456 if (item->wk_state & ONWORKLIST) {
458 panic("worklist_insert: already on list");
460 item->wk_state |= ONWORKLIST;
461 LIST_INSERT_HEAD(head, item, wk_list);
465 worklist_remove(item)
466 struct worklist *item;
469 if (lk.lkt_held == NOHOLDER)
470 panic("worklist_remove: lock not held");
471 if ((item->wk_state & ONWORKLIST) == 0) {
473 panic("worklist_remove: not on list");
475 item->wk_state &= ~ONWORKLIST;
476 LIST_REMOVE(item, wk_list);
480 workitem_free(item, type)
481 struct worklist *item;
485 if (item->wk_state & ONWORKLIST) {
486 if (lk.lkt_held != NOHOLDER)
488 panic("workitem_free: still on list");
490 if (item->wk_type != type) {
491 if (lk.lkt_held != NOHOLDER)
493 panic("workitem_free: type mismatch");
495 FREE(item, DtoM(type));
500 * Workitem queue management
502 static struct workhead softdep_workitem_pending;
503 static int num_on_worklist; /* number of worklist items to be processed */
504 static int softdep_worklist_busy; /* 1 => trying to do unmount */
505 static int softdep_worklist_req; /* serialized waiters */
506 static int max_softdeps; /* maximum number of structs before slowdown */
507 static int tickdelay = 2; /* number of ticks to pause during slowdown */
508 static int *stat_countp; /* statistic to count in proc_waiting timeout */
509 static int proc_waiting; /* tracks whether we have a timeout posted */
510 static struct callout_handle handle; /* handle on posted proc_waiting timeout */
511 static struct thread *filesys_syncer; /* proc of filesystem syncer process */
512 static int req_clear_inodedeps; /* syncer process flush some inodedeps */
513 #define FLUSH_INODES 1
514 static int req_clear_remove; /* syncer process flush some freeblks */
515 #define FLUSH_REMOVE 2
519 static int stat_worklist_push; /* number of worklist cleanups */
520 static int stat_blk_limit_push; /* number of times block limit neared */
521 static int stat_ino_limit_push; /* number of times inode limit neared */
522 static int stat_blk_limit_hit; /* number of times block slowdown imposed */
523 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */
524 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */
525 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */
526 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */
527 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
528 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */
531 #include <sys/sysctl.h>
532 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
533 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
534 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
535 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
536 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
537 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
538 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
539 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
540 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
541 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
542 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
543 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
547 * Add an item to the end of the work queue.
548 * This routine requires that the lock be held.
549 * This is the only routine that adds items to the list.
550 * The following routine is the only one that removes items
551 * and does so in order from first to last.
557 static struct worklist *worklist_tail;
559 if (wk->wk_state & ONWORKLIST) {
560 if (lk.lkt_held != NOHOLDER)
562 panic("add_to_worklist: already on list");
564 wk->wk_state |= ONWORKLIST;
565 if (LIST_FIRST(&softdep_workitem_pending) == NULL)
566 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
568 LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
570 num_on_worklist += 1;
574 * Process that runs once per second to handle items in the background queue.
576 * Note that we ensure that everything is done in the order in which they
577 * appear in the queue. The code below depends on this property to ensure
578 * that blocks of a file are freed before the inode itself is freed. This
579 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
580 * until all the old ones have been purged from the dependency lists.
583 softdep_process_worklist(matchmnt)
584 struct mount *matchmnt;
586 thread_t td = curthread;
587 int matchcnt, loopcount;
591 * Record the process identifier of our caller so that we can give
592 * this process preferential treatment in request_cleanup below.
598 * There is no danger of having multiple processes run this
599 * code, but we have to single-thread it when softdep_flushfiles()
600 * is in operation to get an accurate count of the number of items
601 * related to its mount point that are in the list.
603 if (matchmnt == NULL) {
604 if (softdep_worklist_busy < 0)
606 softdep_worklist_busy += 1;
610 * If requested, try removing inode or removal dependencies.
612 if (req_clear_inodedeps) {
614 req_clear_inodedeps -= 1;
615 wakeup_one(&proc_waiting);
617 if (req_clear_remove) {
619 req_clear_remove -= 1;
620 wakeup_one(&proc_waiting);
623 starttime = time_second;
624 while (num_on_worklist > 0) {
625 matchcnt += process_worklist_item(matchmnt, 0);
628 * If a umount operation wants to run the worklist
631 if (softdep_worklist_req && matchmnt == NULL) {
637 * If requested, try removing inode or removal dependencies.
639 if (req_clear_inodedeps) {
641 req_clear_inodedeps -= 1;
642 wakeup_one(&proc_waiting);
644 if (req_clear_remove) {
646 req_clear_remove -= 1;
647 wakeup_one(&proc_waiting);
650 * We do not generally want to stop for buffer space, but if
651 * we are really being a buffer hog, we will stop and wait.
653 if (loopcount++ % 128 == 0)
656 * Never allow processing to run for more than one
657 * second. Otherwise the other syncer tasks may get
658 * excessively backlogged.
660 if (starttime != time_second && matchmnt == NULL) {
665 if (matchmnt == NULL) {
666 --softdep_worklist_busy;
667 if (softdep_worklist_req && softdep_worklist_busy == 0)
668 wakeup(&softdep_worklist_req);
674 * Process one item on the worklist.
677 process_worklist_item(matchmnt, flags)
678 struct mount *matchmnt;
682 struct dirrem *dirrem;
688 if (matchmnt != NULL)
689 matchfs = VFSTOUFS(matchmnt)->um_fs;
692 * Normally we just process each item on the worklist in order.
693 * However, if we are in a situation where we cannot lock any
694 * inodes, we have to skip over any dirrem requests whose
695 * vnodes are resident and locked.
697 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
698 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
700 dirrem = WK_DIRREM(wk);
701 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev,
703 if (vp == NULL || !VOP_ISLOCKED(vp, curthread))
711 num_on_worklist -= 1;
713 switch (wk->wk_type) {
716 /* removal of a directory entry */
717 if (WK_DIRREM(wk)->dm_mnt == matchmnt)
719 handle_workitem_remove(WK_DIRREM(wk));
723 /* releasing blocks and/or fragments from a file */
724 if (WK_FREEBLKS(wk)->fb_fs == matchfs)
726 handle_workitem_freeblocks(WK_FREEBLKS(wk));
730 /* releasing a fragment when replaced as a file grows */
731 if (WK_FREEFRAG(wk)->ff_fs == matchfs)
733 handle_workitem_freefrag(WK_FREEFRAG(wk));
737 /* releasing an inode when its link count drops to 0 */
738 if (WK_FREEFILE(wk)->fx_fs == matchfs)
740 handle_workitem_freefile(WK_FREEFILE(wk));
744 panic("%s_process_worklist: Unknown type %s",
745 "softdep", TYPENAME(wk->wk_type));
752 * Move dependencies from one buffer to another.
755 softdep_move_dependencies(oldbp, newbp)
759 struct worklist *wk, *wktail;
761 if (LIST_FIRST(&newbp->b_dep) != NULL)
762 panic("softdep_move_dependencies: need merge code");
765 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
766 LIST_REMOVE(wk, wk_list);
768 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
770 LIST_INSERT_AFTER(wktail, wk, wk_list);
777 * Purge the work list of all items associated with a particular mount point.
780 softdep_flushfiles(struct mount *oldmnt, int flags, struct thread *td)
786 KKASSERT(td->td_proc);
787 cred = td->td_proc->p_ucred;
790 * Await our turn to clear out the queue, then serialize access.
792 while (softdep_worklist_busy != 0) {
793 softdep_worklist_req += 1;
794 tsleep(&softdep_worklist_req, PRIBIO, "softflush", 0);
795 softdep_worklist_req -= 1;
797 softdep_worklist_busy = -1;
799 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) {
800 softdep_worklist_busy = 0;
801 if (softdep_worklist_req)
802 wakeup(&softdep_worklist_req);
806 * Alternately flush the block device associated with the mount
807 * point and process any dependencies that the flushing
808 * creates. In theory, this loop can happen at most twice,
809 * but we give it a few extra just to be sure.
811 devvp = VFSTOUFS(oldmnt)->um_devvp;
812 for (loopcnt = 10; loopcnt > 0; ) {
813 if (softdep_process_worklist(oldmnt) == 0) {
816 * Do another flush in case any vnodes were brought in
817 * as part of the cleanup operations.
819 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
822 * If we still found nothing to do, we are really done.
824 if (softdep_process_worklist(oldmnt) == 0)
827 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td);
828 error = VOP_FSYNC(devvp, cred, MNT_WAIT, td);
829 VOP_UNLOCK(devvp, 0, td);
833 softdep_worklist_busy = 0;
834 if (softdep_worklist_req)
835 wakeup(&softdep_worklist_req);
838 * If we are unmounting then it is an error to fail. If we
839 * are simply trying to downgrade to read-only, then filesystem
840 * activity can keep us busy forever, so we just fail with EBUSY.
843 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
844 panic("softdep_flushfiles: looping");
853 * There are three types of structures that can be looked up:
854 * 1) pagedep structures identified by mount point, inode number,
856 * 2) inodedep structures identified by mount point and inode number.
857 * 3) newblk structures identified by mount point and
858 * physical block number.
860 * The "pagedep" and "inodedep" dependency structures are hashed
861 * separately from the file blocks and inodes to which they correspond.
862 * This separation helps when the in-memory copy of an inode or
863 * file block must be replaced. It also obviates the need to access
864 * an inode or file page when simply updating (or de-allocating)
865 * dependency structures. Lookup of newblk structures is needed to
866 * find newly allocated blocks when trying to associate them with
867 * their allocdirect or allocindir structure.
869 * The lookup routines optionally create and hash a new instance when
870 * an existing entry is not found.
872 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */
873 #define NODELAY 0x0002 /* cannot do background work */
876 * Structures and routines associated with pagedep caching.
878 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
879 u_long pagedep_hash; /* size of hash table - 1 */
880 #define PAGEDEP_HASH(mp, inum, lbn) \
881 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
883 static struct sema pagedep_in_progress;
886 * Look up a pagedep. Return 1 if found, 0 if not found.
887 * If not found, allocate if DEPALLOC flag is passed.
888 * Found or allocated entry is returned in pagedeppp.
889 * This routine must be called with splbio interrupts blocked.
892 pagedep_lookup(ip, lbn, flags, pagedeppp)
896 struct pagedep **pagedeppp;
898 struct pagedep *pagedep;
899 struct pagedep_hashhead *pagedephd;
904 if (lk.lkt_held == NOHOLDER)
905 panic("pagedep_lookup: lock not held");
907 mp = ITOV(ip)->v_mount;
908 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
910 LIST_FOREACH(pagedep, pagedephd, pd_hash)
911 if (ip->i_number == pagedep->pd_ino &&
912 lbn == pagedep->pd_lbn &&
913 mp == pagedep->pd_mnt)
916 *pagedeppp = pagedep;
919 if ((flags & DEPALLOC) == 0) {
923 if (sema_get(&pagedep_in_progress, &lk) == 0) {
927 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
929 bzero(pagedep, sizeof(struct pagedep));
930 pagedep->pd_list.wk_type = D_PAGEDEP;
931 pagedep->pd_mnt = mp;
932 pagedep->pd_ino = ip->i_number;
933 pagedep->pd_lbn = lbn;
934 LIST_INIT(&pagedep->pd_dirremhd);
935 LIST_INIT(&pagedep->pd_pendinghd);
936 for (i = 0; i < DAHASHSZ; i++)
937 LIST_INIT(&pagedep->pd_diraddhd[i]);
939 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
940 sema_release(&pagedep_in_progress);
941 *pagedeppp = pagedep;
946 * Structures and routines associated with inodedep caching.
948 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
949 static u_long inodedep_hash; /* size of hash table - 1 */
950 static long num_inodedep; /* number of inodedep allocated */
951 #define INODEDEP_HASH(fs, inum) \
952 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
953 static struct sema inodedep_in_progress;
956 * Look up a inodedep. Return 1 if found, 0 if not found.
957 * If not found, allocate if DEPALLOC flag is passed.
958 * Found or allocated entry is returned in inodedeppp.
959 * This routine must be called with splbio interrupts blocked.
962 inodedep_lookup(fs, inum, flags, inodedeppp)
966 struct inodedep **inodedeppp;
968 struct inodedep *inodedep;
969 struct inodedep_hashhead *inodedephd;
973 if (lk.lkt_held == NOHOLDER)
974 panic("inodedep_lookup: lock not held");
977 inodedephd = INODEDEP_HASH(fs, inum);
979 LIST_FOREACH(inodedep, inodedephd, id_hash)
980 if (inum == inodedep->id_ino && fs == inodedep->id_fs)
983 *inodedeppp = inodedep;
986 if ((flags & DEPALLOC) == 0) {
991 * If we are over our limit, try to improve the situation.
993 if (num_inodedep > max_softdeps && firsttry &&
994 speedup_syncer() == 0 && (flags & NODELAY) == 0 &&
995 request_cleanup(FLUSH_INODES, 1)) {
999 if (sema_get(&inodedep_in_progress, &lk) == 0) {
1004 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1005 M_INODEDEP, M_SOFTDEP_FLAGS);
1006 inodedep->id_list.wk_type = D_INODEDEP;
1007 inodedep->id_fs = fs;
1008 inodedep->id_ino = inum;
1009 inodedep->id_state = ALLCOMPLETE;
1010 inodedep->id_nlinkdelta = 0;
1011 inodedep->id_savedino = NULL;
1012 inodedep->id_savedsize = -1;
1013 inodedep->id_buf = NULL;
1014 LIST_INIT(&inodedep->id_pendinghd);
1015 LIST_INIT(&inodedep->id_inowait);
1016 LIST_INIT(&inodedep->id_bufwait);
1017 TAILQ_INIT(&inodedep->id_inoupdt);
1018 TAILQ_INIT(&inodedep->id_newinoupdt);
1020 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1021 sema_release(&inodedep_in_progress);
1022 *inodedeppp = inodedep;
1027 * Structures and routines associated with newblk caching.
1029 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1030 u_long newblk_hash; /* size of hash table - 1 */
1031 #define NEWBLK_HASH(fs, inum) \
1032 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1033 static struct sema newblk_in_progress;
1036 * Look up a newblk. Return 1 if found, 0 if not found.
1037 * If not found, allocate if DEPALLOC flag is passed.
1038 * Found or allocated entry is returned in newblkpp.
1041 newblk_lookup(fs, newblkno, flags, newblkpp)
1043 ufs_daddr_t newblkno;
1045 struct newblk **newblkpp;
1047 struct newblk *newblk;
1048 struct newblk_hashhead *newblkhd;
1050 newblkhd = NEWBLK_HASH(fs, newblkno);
1052 LIST_FOREACH(newblk, newblkhd, nb_hash)
1053 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1059 if ((flags & DEPALLOC) == 0) {
1063 if (sema_get(&newblk_in_progress, 0) == 0)
1065 MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1066 M_NEWBLK, M_SOFTDEP_FLAGS);
1067 newblk->nb_state = 0;
1069 newblk->nb_newblkno = newblkno;
1070 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1071 sema_release(&newblk_in_progress);
1077 * Executed during filesystem system initialization before
1078 * mounting any file systems.
1081 softdep_initialize()
1084 LIST_INIT(&mkdirlisthd);
1085 LIST_INIT(&softdep_workitem_pending);
1086 max_softdeps = min(desiredvnodes * 8,
1087 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep)));
1088 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1090 sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0);
1091 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1092 sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0);
1093 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1094 sema_init(&newblk_in_progress, "newblk", PRIBIO, 0);
1098 * Called at mount time to notify the dependency code that a
1099 * filesystem wishes to use it.
1102 softdep_mount(devvp, mp, fs, cred)
1103 struct vnode *devvp;
1108 struct csum cstotal;
1113 mp->mnt_flag &= ~MNT_ASYNC;
1114 mp->mnt_flag |= MNT_SOFTDEP;
1116 * When doing soft updates, the counters in the
1117 * superblock may have gotten out of sync, so we have
1118 * to scan the cylinder groups and recalculate them.
1120 if (fs->fs_clean != 0)
1122 bzero(&cstotal, sizeof cstotal);
1123 for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1124 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1125 fs->fs_cgsize, cred, &bp)) != 0) {
1129 cgp = (struct cg *)bp->b_data;
1130 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1131 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1132 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1133 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1134 fs->fs_cs(fs, cyl) = cgp->cg_cs;
1138 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1139 printf("ffs_mountfs: superblock updated for soft updates\n");
1141 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1146 * Protecting the freemaps (or bitmaps).
1148 * To eliminate the need to execute fsck before mounting a file system
1149 * after a power failure, one must (conservatively) guarantee that the
1150 * on-disk copy of the bitmaps never indicate that a live inode or block is
1151 * free. So, when a block or inode is allocated, the bitmap should be
1152 * updated (on disk) before any new pointers. When a block or inode is
1153 * freed, the bitmap should not be updated until all pointers have been
1154 * reset. The latter dependency is handled by the delayed de-allocation
1155 * approach described below for block and inode de-allocation. The former
1156 * dependency is handled by calling the following procedure when a block or
1157 * inode is allocated. When an inode is allocated an "inodedep" is created
1158 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1159 * Each "inodedep" is also inserted into the hash indexing structure so
1160 * that any additional link additions can be made dependent on the inode
1163 * The ufs file system maintains a number of free block counts (e.g., per
1164 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1165 * in addition to the bitmaps. These counts are used to improve efficiency
1166 * during allocation and therefore must be consistent with the bitmaps.
1167 * There is no convenient way to guarantee post-crash consistency of these
1168 * counts with simple update ordering, for two main reasons: (1) The counts
1169 * and bitmaps for a single cylinder group block are not in the same disk
1170 * sector. If a disk write is interrupted (e.g., by power failure), one may
1171 * be written and the other not. (2) Some of the counts are located in the
1172 * superblock rather than the cylinder group block. So, we focus our soft
1173 * updates implementation on protecting the bitmaps. When mounting a
1174 * filesystem, we recompute the auxiliary counts from the bitmaps.
1178 * Called just after updating the cylinder group block to allocate an inode.
1181 softdep_setup_inomapdep(bp, ip, newinum)
1182 struct buf *bp; /* buffer for cylgroup block with inode map */
1183 struct inode *ip; /* inode related to allocation */
1184 ino_t newinum; /* new inode number being allocated */
1186 struct inodedep *inodedep;
1187 struct bmsafemap *bmsafemap;
1190 * Create a dependency for the newly allocated inode.
1191 * Panic if it already exists as something is seriously wrong.
1192 * Otherwise add it to the dependency list for the buffer holding
1193 * the cylinder group map from which it was allocated.
1196 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1198 panic("softdep_setup_inomapdep: found inode");
1200 inodedep->id_buf = bp;
1201 inodedep->id_state &= ~DEPCOMPLETE;
1202 bmsafemap = bmsafemap_lookup(bp);
1203 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1208 * Called just after updating the cylinder group block to
1209 * allocate block or fragment.
1212 softdep_setup_blkmapdep(bp, fs, newblkno)
1213 struct buf *bp; /* buffer for cylgroup block with block map */
1214 struct fs *fs; /* filesystem doing allocation */
1215 ufs_daddr_t newblkno; /* number of newly allocated block */
1217 struct newblk *newblk;
1218 struct bmsafemap *bmsafemap;
1221 * Create a dependency for the newly allocated block.
1222 * Add it to the dependency list for the buffer holding
1223 * the cylinder group map from which it was allocated.
1225 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1226 panic("softdep_setup_blkmapdep: found block");
1228 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1229 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1234 * Find the bmsafemap associated with a cylinder group buffer.
1235 * If none exists, create one. The buffer must be locked when
1236 * this routine is called and this routine must be called with
1237 * splbio interrupts blocked.
1239 static struct bmsafemap *
1240 bmsafemap_lookup(bp)
1243 struct bmsafemap *bmsafemap;
1244 struct worklist *wk;
1247 if (lk.lkt_held == NOHOLDER)
1248 panic("bmsafemap_lookup: lock not held");
1250 LIST_FOREACH(wk, &bp->b_dep, wk_list)
1251 if (wk->wk_type == D_BMSAFEMAP)
1252 return (WK_BMSAFEMAP(wk));
1254 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1255 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1256 bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1257 bmsafemap->sm_list.wk_state = 0;
1258 bmsafemap->sm_buf = bp;
1259 LIST_INIT(&bmsafemap->sm_allocdirecthd);
1260 LIST_INIT(&bmsafemap->sm_allocindirhd);
1261 LIST_INIT(&bmsafemap->sm_inodedephd);
1262 LIST_INIT(&bmsafemap->sm_newblkhd);
1264 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1269 * Direct block allocation dependencies.
1271 * When a new block is allocated, the corresponding disk locations must be
1272 * initialized (with zeros or new data) before the on-disk inode points to
1273 * them. Also, the freemap from which the block was allocated must be
1274 * updated (on disk) before the inode's pointer. These two dependencies are
1275 * independent of each other and are needed for all file blocks and indirect
1276 * blocks that are pointed to directly by the inode. Just before the
1277 * "in-core" version of the inode is updated with a newly allocated block
1278 * number, a procedure (below) is called to setup allocation dependency
1279 * structures. These structures are removed when the corresponding
1280 * dependencies are satisfied or when the block allocation becomes obsolete
1281 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1282 * fragment that gets upgraded). All of these cases are handled in
1283 * procedures described later.
1285 * When a file extension causes a fragment to be upgraded, either to a larger
1286 * fragment or to a full block, the on-disk location may change (if the
1287 * previous fragment could not simply be extended). In this case, the old
1288 * fragment must be de-allocated, but not until after the inode's pointer has
1289 * been updated. In most cases, this is handled by later procedures, which
1290 * will construct a "freefrag" structure to be added to the workitem queue
1291 * when the inode update is complete (or obsolete). The main exception to
1292 * this is when an allocation occurs while a pending allocation dependency
1293 * (for the same block pointer) remains. This case is handled in the main
1294 * allocation dependency setup procedure by immediately freeing the
1295 * unreferenced fragments.
1298 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1299 struct inode *ip; /* inode to which block is being added */
1300 ufs_lbn_t lbn; /* block pointer within inode */
1301 ufs_daddr_t newblkno; /* disk block number being added */
1302 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */
1303 long newsize; /* size of new block */
1304 long oldsize; /* size of new block */
1305 struct buf *bp; /* bp for allocated block */
1307 struct allocdirect *adp, *oldadp;
1308 struct allocdirectlst *adphead;
1309 struct bmsafemap *bmsafemap;
1310 struct inodedep *inodedep;
1311 struct pagedep *pagedep;
1312 struct newblk *newblk;
1314 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1315 M_ALLOCDIRECT, M_SOFTDEP_FLAGS);
1316 bzero(adp, sizeof(struct allocdirect));
1317 adp->ad_list.wk_type = D_ALLOCDIRECT;
1319 adp->ad_newblkno = newblkno;
1320 adp->ad_oldblkno = oldblkno;
1321 adp->ad_newsize = newsize;
1322 adp->ad_oldsize = oldsize;
1323 adp->ad_state = ATTACHED;
1324 if (newblkno == oldblkno)
1325 adp->ad_freefrag = NULL;
1327 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1329 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1330 panic("softdep_setup_allocdirect: lost block");
1333 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1334 adp->ad_inodedep = inodedep;
1336 if (newblk->nb_state == DEPCOMPLETE) {
1337 adp->ad_state |= DEPCOMPLETE;
1340 bmsafemap = newblk->nb_bmsafemap;
1341 adp->ad_buf = bmsafemap->sm_buf;
1342 LIST_REMOVE(newblk, nb_deps);
1343 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1345 LIST_REMOVE(newblk, nb_hash);
1346 FREE(newblk, M_NEWBLK);
1348 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1349 if (lbn >= NDADDR) {
1350 /* allocating an indirect block */
1351 if (oldblkno != 0) {
1353 panic("softdep_setup_allocdirect: non-zero indir");
1357 * Allocating a direct block.
1359 * If we are allocating a directory block, then we must
1360 * allocate an associated pagedep to track additions and
1363 if ((ip->i_mode & IFMT) == IFDIR &&
1364 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1365 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1368 * The list of allocdirects must be kept in sorted and ascending
1369 * order so that the rollback routines can quickly determine the
1370 * first uncommitted block (the size of the file stored on disk
1371 * ends at the end of the lowest committed fragment, or if there
1372 * are no fragments, at the end of the highest committed block).
1373 * Since files generally grow, the typical case is that the new
1374 * block is to be added at the end of the list. We speed this
1375 * special case by checking against the last allocdirect in the
1376 * list before laboriously traversing the list looking for the
1379 adphead = &inodedep->id_newinoupdt;
1380 oldadp = TAILQ_LAST(adphead, allocdirectlst);
1381 if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1382 /* insert at end of list */
1383 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1384 if (oldadp != NULL && oldadp->ad_lbn == lbn)
1385 allocdirect_merge(adphead, adp, oldadp);
1389 TAILQ_FOREACH(oldadp, adphead, ad_next) {
1390 if (oldadp->ad_lbn >= lbn)
1393 if (oldadp == NULL) {
1395 panic("softdep_setup_allocdirect: lost entry");
1397 /* insert in middle of list */
1398 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1399 if (oldadp->ad_lbn == lbn)
1400 allocdirect_merge(adphead, adp, oldadp);
1405 * Replace an old allocdirect dependency with a newer one.
1406 * This routine must be called with splbio interrupts blocked.
1409 allocdirect_merge(adphead, newadp, oldadp)
1410 struct allocdirectlst *adphead; /* head of list holding allocdirects */
1411 struct allocdirect *newadp; /* allocdirect being added */
1412 struct allocdirect *oldadp; /* existing allocdirect being checked */
1414 struct freefrag *freefrag;
1417 if (lk.lkt_held == NOHOLDER)
1418 panic("allocdirect_merge: lock not held");
1420 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1421 newadp->ad_oldsize != oldadp->ad_newsize ||
1422 newadp->ad_lbn >= NDADDR) {
1424 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1425 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1428 newadp->ad_oldblkno = oldadp->ad_oldblkno;
1429 newadp->ad_oldsize = oldadp->ad_oldsize;
1431 * If the old dependency had a fragment to free or had never
1432 * previously had a block allocated, then the new dependency
1433 * can immediately post its freefrag and adopt the old freefrag.
1434 * This action is done by swapping the freefrag dependencies.
1435 * The new dependency gains the old one's freefrag, and the
1436 * old one gets the new one and then immediately puts it on
1437 * the worklist when it is freed by free_allocdirect. It is
1438 * not possible to do this swap when the old dependency had a
1439 * non-zero size but no previous fragment to free. This condition
1440 * arises when the new block is an extension of the old block.
1441 * Here, the first part of the fragment allocated to the new
1442 * dependency is part of the block currently claimed on disk by
1443 * the old dependency, so cannot legitimately be freed until the
1444 * conditions for the new dependency are fulfilled.
1446 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1447 freefrag = newadp->ad_freefrag;
1448 newadp->ad_freefrag = oldadp->ad_freefrag;
1449 oldadp->ad_freefrag = freefrag;
1451 free_allocdirect(adphead, oldadp, 0);
1455 * Allocate a new freefrag structure if needed.
1457 static struct freefrag *
1458 newfreefrag(ip, blkno, size)
1463 struct freefrag *freefrag;
1469 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1470 panic("newfreefrag: frag size");
1471 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1472 M_FREEFRAG, M_SOFTDEP_FLAGS);
1473 freefrag->ff_list.wk_type = D_FREEFRAG;
1474 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */
1475 freefrag->ff_inum = ip->i_number;
1476 freefrag->ff_fs = fs;
1477 freefrag->ff_devvp = ip->i_devvp;
1478 freefrag->ff_blkno = blkno;
1479 freefrag->ff_fragsize = size;
1484 * This workitem de-allocates fragments that were replaced during
1485 * file block allocation.
1488 handle_workitem_freefrag(freefrag)
1489 struct freefrag *freefrag;
1493 tip.i_fs = freefrag->ff_fs;
1494 tip.i_devvp = freefrag->ff_devvp;
1495 tip.i_dev = freefrag->ff_devvp->v_rdev;
1496 tip.i_number = freefrag->ff_inum;
1497 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */
1498 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1499 FREE(freefrag, M_FREEFRAG);
1503 * Indirect block allocation dependencies.
1505 * The same dependencies that exist for a direct block also exist when
1506 * a new block is allocated and pointed to by an entry in a block of
1507 * indirect pointers. The undo/redo states described above are also
1508 * used here. Because an indirect block contains many pointers that
1509 * may have dependencies, a second copy of the entire in-memory indirect
1510 * block is kept. The buffer cache copy is always completely up-to-date.
1511 * The second copy, which is used only as a source for disk writes,
1512 * contains only the safe pointers (i.e., those that have no remaining
1513 * update dependencies). The second copy is freed when all pointers
1514 * are safe. The cache is not allowed to replace indirect blocks with
1515 * pending update dependencies. If a buffer containing an indirect
1516 * block with dependencies is written, these routines will mark it
1517 * dirty again. It can only be successfully written once all the
1518 * dependencies are removed. The ffs_fsync routine in conjunction with
1519 * softdep_sync_metadata work together to get all the dependencies
1520 * removed so that a file can be successfully written to disk. Three
1521 * procedures are used when setting up indirect block pointer
1522 * dependencies. The division is necessary because of the organization
1523 * of the "balloc" routine and because of the distinction between file
1524 * pages and file metadata blocks.
1528 * Allocate a new allocindir structure.
1530 static struct allocindir *
1531 newallocindir(ip, ptrno, newblkno, oldblkno)
1532 struct inode *ip; /* inode for file being extended */
1533 int ptrno; /* offset of pointer in indirect block */
1534 ufs_daddr_t newblkno; /* disk block number being added */
1535 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1537 struct allocindir *aip;
1539 MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1540 M_ALLOCINDIR, M_SOFTDEP_FLAGS);
1541 bzero(aip, sizeof(struct allocindir));
1542 aip->ai_list.wk_type = D_ALLOCINDIR;
1543 aip->ai_state = ATTACHED;
1544 aip->ai_offset = ptrno;
1545 aip->ai_newblkno = newblkno;
1546 aip->ai_oldblkno = oldblkno;
1547 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1552 * Called just before setting an indirect block pointer
1553 * to a newly allocated file page.
1556 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1557 struct inode *ip; /* inode for file being extended */
1558 ufs_lbn_t lbn; /* allocated block number within file */
1559 struct buf *bp; /* buffer with indirect blk referencing page */
1560 int ptrno; /* offset of pointer in indirect block */
1561 ufs_daddr_t newblkno; /* disk block number being added */
1562 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1563 struct buf *nbp; /* buffer holding allocated page */
1565 struct allocindir *aip;
1566 struct pagedep *pagedep;
1568 aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1571 * If we are allocating a directory page, then we must
1572 * allocate an associated pagedep to track additions and
1575 if ((ip->i_mode & IFMT) == IFDIR &&
1576 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1577 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1578 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1580 setup_allocindir_phase2(bp, ip, aip);
1584 * Called just before setting an indirect block pointer to a
1585 * newly allocated indirect block.
1588 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1589 struct buf *nbp; /* newly allocated indirect block */
1590 struct inode *ip; /* inode for file being extended */
1591 struct buf *bp; /* indirect block referencing allocated block */
1592 int ptrno; /* offset of pointer in indirect block */
1593 ufs_daddr_t newblkno; /* disk block number being added */
1595 struct allocindir *aip;
1597 aip = newallocindir(ip, ptrno, newblkno, 0);
1599 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1601 setup_allocindir_phase2(bp, ip, aip);
1605 * Called to finish the allocation of the "aip" allocated
1606 * by one of the two routines above.
1609 setup_allocindir_phase2(bp, ip, aip)
1610 struct buf *bp; /* in-memory copy of the indirect block */
1611 struct inode *ip; /* inode for file being extended */
1612 struct allocindir *aip; /* allocindir allocated by the above routines */
1614 struct worklist *wk;
1615 struct indirdep *indirdep, *newindirdep;
1616 struct bmsafemap *bmsafemap;
1617 struct allocindir *oldaip;
1618 struct freefrag *freefrag;
1619 struct newblk *newblk;
1621 if (bp->b_lblkno >= 0)
1622 panic("setup_allocindir_phase2: not indir blk");
1623 for (indirdep = NULL, newindirdep = NULL; ; ) {
1625 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1626 if (wk->wk_type != D_INDIRDEP)
1628 indirdep = WK_INDIRDEP(wk);
1631 if (indirdep == NULL && newindirdep) {
1632 indirdep = newindirdep;
1633 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1638 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1640 panic("setup_allocindir: lost block");
1642 if (newblk->nb_state == DEPCOMPLETE) {
1643 aip->ai_state |= DEPCOMPLETE;
1646 bmsafemap = newblk->nb_bmsafemap;
1647 aip->ai_buf = bmsafemap->sm_buf;
1648 LIST_REMOVE(newblk, nb_deps);
1649 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1652 LIST_REMOVE(newblk, nb_hash);
1653 FREE(newblk, M_NEWBLK);
1654 aip->ai_indirdep = indirdep;
1656 * Check to see if there is an existing dependency
1657 * for this block. If there is, merge the old
1658 * dependency into the new one.
1660 if (aip->ai_oldblkno == 0)
1664 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1665 if (oldaip->ai_offset == aip->ai_offset)
1667 if (oldaip != NULL) {
1668 if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1670 panic("setup_allocindir_phase2: blkno");
1672 aip->ai_oldblkno = oldaip->ai_oldblkno;
1673 freefrag = oldaip->ai_freefrag;
1674 oldaip->ai_freefrag = aip->ai_freefrag;
1675 aip->ai_freefrag = freefrag;
1676 free_allocindir(oldaip, NULL);
1678 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1679 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1680 [aip->ai_offset] = aip->ai_oldblkno;
1684 if (indirdep->ir_savebp != NULL)
1685 brelse(newindirdep->ir_savebp);
1686 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1690 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1691 M_INDIRDEP, M_SOFTDEP_FLAGS);
1692 newindirdep->ir_list.wk_type = D_INDIRDEP;
1693 newindirdep->ir_state = ATTACHED;
1694 LIST_INIT(&newindirdep->ir_deplisthd);
1695 LIST_INIT(&newindirdep->ir_donehd);
1696 if (bp->b_blkno == bp->b_lblkno) {
1697 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
1700 newindirdep->ir_savebp =
1701 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0);
1702 BUF_KERNPROC(newindirdep->ir_savebp);
1703 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1708 * Block de-allocation dependencies.
1710 * When blocks are de-allocated, the on-disk pointers must be nullified before
1711 * the blocks are made available for use by other files. (The true
1712 * requirement is that old pointers must be nullified before new on-disk
1713 * pointers are set. We chose this slightly more stringent requirement to
1714 * reduce complexity.) Our implementation handles this dependency by updating
1715 * the inode (or indirect block) appropriately but delaying the actual block
1716 * de-allocation (i.e., freemap and free space count manipulation) until
1717 * after the updated versions reach stable storage. After the disk is
1718 * updated, the blocks can be safely de-allocated whenever it is convenient.
1719 * This implementation handles only the common case of reducing a file's
1720 * length to zero. Other cases are handled by the conventional synchronous
1723 * The ffs implementation with which we worked double-checks
1724 * the state of the block pointers and file size as it reduces
1725 * a file's length. Some of this code is replicated here in our
1726 * soft updates implementation. The freeblks->fb_chkcnt field is
1727 * used to transfer a part of this information to the procedure
1728 * that eventually de-allocates the blocks.
1730 * This routine should be called from the routine that shortens
1731 * a file's length, before the inode's size or block pointers
1732 * are modified. It will save the block pointer information for
1733 * later release and zero the inode so that the calling routine
1737 softdep_setup_freeblocks(ip, length)
1738 struct inode *ip; /* The inode whose length is to be reduced */
1739 off_t length; /* The new length for the file */
1741 struct freeblks *freeblks;
1742 struct inodedep *inodedep;
1743 struct allocdirect *adp;
1747 int i, error, delay;
1751 panic("softde_setup_freeblocks: non-zero length");
1752 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1753 M_FREEBLKS, M_SOFTDEP_FLAGS);
1754 bzero(freeblks, sizeof(struct freeblks));
1755 freeblks->fb_list.wk_type = D_FREEBLKS;
1756 freeblks->fb_uid = ip->i_uid;
1757 freeblks->fb_previousinum = ip->i_number;
1758 freeblks->fb_devvp = ip->i_devvp;
1759 freeblks->fb_fs = fs;
1760 freeblks->fb_oldsize = ip->i_size;
1761 freeblks->fb_newsize = length;
1762 freeblks->fb_chkcnt = ip->i_blocks;
1763 for (i = 0; i < NDADDR; i++) {
1764 freeblks->fb_dblks[i] = ip->i_db[i];
1767 for (i = 0; i < NIADDR; i++) {
1768 freeblks->fb_iblks[i] = ip->i_ib[i];
1774 * Push the zero'ed inode to to its disk buffer so that we are free
1775 * to delete its dependencies below. Once the dependencies are gone
1776 * the buffer can be safely released.
1778 if ((error = bread(ip->i_devvp,
1779 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
1780 (int)fs->fs_bsize, NOCRED, &bp)) != 0)
1781 softdep_error("softdep_setup_freeblocks", error);
1782 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
1785 * Find and eliminate any inode dependencies.
1788 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1789 if ((inodedep->id_state & IOSTARTED) != 0) {
1791 panic("softdep_setup_freeblocks: inode busy");
1794 * Add the freeblks structure to the list of operations that
1795 * must await the zero'ed inode being written to disk. If we
1796 * still have a bitmap dependency (delay == 0), then the inode
1797 * has never been written to disk, so we can process the
1798 * freeblks below once we have deleted the dependencies.
1800 delay = (inodedep->id_state & DEPCOMPLETE);
1802 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1804 * Because the file length has been truncated to zero, any
1805 * pending block allocation dependency structures associated
1806 * with this inode are obsolete and can simply be de-allocated.
1807 * We must first merge the two dependency lists to get rid of
1808 * any duplicate freefrag structures, then purge the merged list.
1810 merge_inode_lists(inodedep);
1811 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1812 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
1816 * We must wait for any I/O in progress to finish so that
1817 * all potential buffers on the dirty list will be visible.
1818 * Once they are all there, walk the list and get rid of
1823 drain_output(vp, 1);
1824 while (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) {
1825 bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
1826 (void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
1827 deallocate_dependencies(bp, inodedep);
1828 bp->b_flags |= B_INVAL | B_NOCACHE;
1833 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
1834 (void)free_inodedep(inodedep);
1837 * If the inode has never been written to disk (delay == 0),
1838 * then we can process the freeblks now that we have deleted
1842 handle_workitem_freeblocks(freeblks);
1846 * Reclaim any dependency structures from a buffer that is about to
1847 * be reallocated to a new vnode. The buffer must be locked, thus,
1848 * no I/O completion operations can occur while we are manipulating
1849 * its associated dependencies. The mutex is held so that other I/O's
1850 * associated with related dependencies do not occur.
1853 deallocate_dependencies(bp, inodedep)
1855 struct inodedep *inodedep;
1857 struct worklist *wk;
1858 struct indirdep *indirdep;
1859 struct allocindir *aip;
1860 struct pagedep *pagedep;
1861 struct dirrem *dirrem;
1865 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1866 switch (wk->wk_type) {
1869 indirdep = WK_INDIRDEP(wk);
1871 * None of the indirect pointers will ever be visible,
1872 * so they can simply be tossed. GOINGAWAY ensures
1873 * that allocated pointers will be saved in the buffer
1874 * cache until they are freed. Note that they will
1875 * only be able to be found by their physical address
1876 * since the inode mapping the logical address will
1877 * be gone. The save buffer used for the safe copy
1878 * was allocated in setup_allocindir_phase2 using
1879 * the physical address so it could be used for this
1880 * purpose. Hence we swap the safe copy with the real
1881 * copy, allowing the safe copy to be freed and holding
1882 * on to the real copy for later use in indir_trunc.
1884 if (indirdep->ir_state & GOINGAWAY) {
1886 panic("deallocate_dependencies: already gone");
1888 indirdep->ir_state |= GOINGAWAY;
1889 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
1890 free_allocindir(aip, inodedep);
1891 if (bp->b_lblkno >= 0 ||
1892 bp->b_blkno != indirdep->ir_savebp->b_lblkno) {
1894 panic("deallocate_dependencies: not indir");
1896 bcopy(bp->b_data, indirdep->ir_savebp->b_data,
1898 WORKLIST_REMOVE(wk);
1899 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
1903 pagedep = WK_PAGEDEP(wk);
1905 * None of the directory additions will ever be
1906 * visible, so they can simply be tossed.
1908 for (i = 0; i < DAHASHSZ; i++)
1910 LIST_FIRST(&pagedep->pd_diraddhd[i])))
1912 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
1915 * Copy any directory remove dependencies to the list
1916 * to be processed after the zero'ed inode is written.
1917 * If the inode has already been written, then they
1918 * can be dumped directly onto the work list.
1920 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
1921 LIST_REMOVE(dirrem, dm_next);
1922 dirrem->dm_dirinum = pagedep->pd_ino;
1923 if (inodedep == NULL ||
1924 (inodedep->id_state & ALLCOMPLETE) ==
1926 add_to_worklist(&dirrem->dm_list);
1928 WORKLIST_INSERT(&inodedep->id_bufwait,
1931 WORKLIST_REMOVE(&pagedep->pd_list);
1932 LIST_REMOVE(pagedep, pd_hash);
1933 WORKITEM_FREE(pagedep, D_PAGEDEP);
1937 free_allocindir(WK_ALLOCINDIR(wk), inodedep);
1943 panic("deallocate_dependencies: Unexpected type %s",
1944 TYPENAME(wk->wk_type));
1949 panic("deallocate_dependencies: Unknown type %s",
1950 TYPENAME(wk->wk_type));
1957 * Free an allocdirect. Generate a new freefrag work request if appropriate.
1958 * This routine must be called with splbio interrupts blocked.
1961 free_allocdirect(adphead, adp, delay)
1962 struct allocdirectlst *adphead;
1963 struct allocdirect *adp;
1968 if (lk.lkt_held == NOHOLDER)
1969 panic("free_allocdirect: lock not held");
1971 if ((adp->ad_state & DEPCOMPLETE) == 0)
1972 LIST_REMOVE(adp, ad_deps);
1973 TAILQ_REMOVE(adphead, adp, ad_next);
1974 if ((adp->ad_state & COMPLETE) == 0)
1975 WORKLIST_REMOVE(&adp->ad_list);
1976 if (adp->ad_freefrag != NULL) {
1978 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
1979 &adp->ad_freefrag->ff_list);
1981 add_to_worklist(&adp->ad_freefrag->ff_list);
1983 WORKITEM_FREE(adp, D_ALLOCDIRECT);
1987 * Prepare an inode to be freed. The actual free operation is not
1988 * done until the zero'ed inode has been written to disk.
1991 softdep_freefile(pvp, ino, mode)
1996 struct inode *ip = VTOI(pvp);
1997 struct inodedep *inodedep;
1998 struct freefile *freefile;
2001 * This sets up the inode de-allocation dependency.
2003 MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2004 M_FREEFILE, M_SOFTDEP_FLAGS);
2005 freefile->fx_list.wk_type = D_FREEFILE;
2006 freefile->fx_list.wk_state = 0;
2007 freefile->fx_mode = mode;
2008 freefile->fx_oldinum = ino;
2009 freefile->fx_devvp = ip->i_devvp;
2010 freefile->fx_fs = ip->i_fs;
2013 * If the inodedep does not exist, then the zero'ed inode has
2014 * been written to disk. If the allocated inode has never been
2015 * written to disk, then the on-disk inode is zero'ed. In either
2016 * case we can free the file immediately.
2019 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2020 check_inode_unwritten(inodedep)) {
2022 handle_workitem_freefile(freefile);
2025 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2030 * Check to see if an inode has never been written to disk. If
2031 * so free the inodedep and return success, otherwise return failure.
2032 * This routine must be called with splbio interrupts blocked.
2034 * If we still have a bitmap dependency, then the inode has never
2035 * been written to disk. Drop the dependency as it is no longer
2036 * necessary since the inode is being deallocated. We set the
2037 * ALLCOMPLETE flags since the bitmap now properly shows that the
2038 * inode is not allocated. Even if the inode is actively being
2039 * written, it has been rolled back to its zero'ed state, so we
2040 * are ensured that a zero inode is what is on the disk. For short
2041 * lived files, this change will usually result in removing all the
2042 * dependencies from the inode so that it can be freed immediately.
2045 check_inode_unwritten(inodedep)
2046 struct inodedep *inodedep;
2049 if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2050 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2051 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2052 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2053 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2054 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2055 inodedep->id_nlinkdelta != 0)
2057 inodedep->id_state |= ALLCOMPLETE;
2058 LIST_REMOVE(inodedep, id_deps);
2059 inodedep->id_buf = NULL;
2060 if (inodedep->id_state & ONWORKLIST)
2061 WORKLIST_REMOVE(&inodedep->id_list);
2062 if (inodedep->id_savedino != NULL) {
2063 FREE(inodedep->id_savedino, M_INODEDEP);
2064 inodedep->id_savedino = NULL;
2066 if (free_inodedep(inodedep) == 0) {
2068 panic("check_inode_unwritten: busy inode");
2074 * Try to free an inodedep structure. Return 1 if it could be freed.
2077 free_inodedep(inodedep)
2078 struct inodedep *inodedep;
2081 if ((inodedep->id_state & ONWORKLIST) != 0 ||
2082 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2083 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2084 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2085 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2086 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2087 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2088 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
2090 LIST_REMOVE(inodedep, id_hash);
2091 WORKITEM_FREE(inodedep, D_INODEDEP);
2097 * This workitem routine performs the block de-allocation.
2098 * The workitem is added to the pending list after the updated
2099 * inode block has been written to disk. As mentioned above,
2100 * checks regarding the number of blocks de-allocated (compared
2101 * to the number of blocks allocated for the file) are also
2102 * performed in this function.
2105 handle_workitem_freeblocks(freeblks)
2106 struct freeblks *freeblks;
2111 int i, level, bsize;
2112 long nblocks, blocksreleased = 0;
2113 int error, allerror = 0;
2114 ufs_lbn_t baselbns[NIADDR], tmpval;
2116 tip.i_number = freeblks->fb_previousinum;
2117 tip.i_devvp = freeblks->fb_devvp;
2118 tip.i_dev = freeblks->fb_devvp->v_rdev;
2119 tip.i_fs = freeblks->fb_fs;
2120 tip.i_size = freeblks->fb_oldsize;
2121 tip.i_uid = freeblks->fb_uid;
2122 fs = freeblks->fb_fs;
2124 baselbns[0] = NDADDR;
2125 for (i = 1; i < NIADDR; i++) {
2126 tmpval *= NINDIR(fs);
2127 baselbns[i] = baselbns[i - 1] + tmpval;
2129 nblocks = btodb(fs->fs_bsize);
2132 * Indirect blocks first.
2134 for (level = (NIADDR - 1); level >= 0; level--) {
2135 if ((bn = freeblks->fb_iblks[level]) == 0)
2137 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level,
2138 baselbns[level], &blocksreleased)) == 0)
2140 ffs_blkfree(&tip, bn, fs->fs_bsize);
2141 blocksreleased += nblocks;
2144 * All direct blocks or frags.
2146 for (i = (NDADDR - 1); i >= 0; i--) {
2147 if ((bn = freeblks->fb_dblks[i]) == 0)
2149 bsize = blksize(fs, &tip, i);
2150 ffs_blkfree(&tip, bn, bsize);
2151 blocksreleased += btodb(bsize);
2155 if (freeblks->fb_chkcnt != blocksreleased)
2156 printf("handle_workitem_freeblocks: block count\n");
2158 softdep_error("handle_workitem_freeblks", allerror);
2159 #endif /* DIAGNOSTIC */
2160 WORKITEM_FREE(freeblks, D_FREEBLKS);
2164 * Release blocks associated with the inode ip and stored in the indirect
2165 * block dbn. If level is greater than SINGLE, the block is an indirect block
2166 * and recursive calls to indirtrunc must be used to cleanse other indirect
2170 indir_trunc(ip, dbn, level, lbn, countp)
2181 struct worklist *wk;
2182 struct indirdep *indirdep;
2183 int i, lbnadd, nblocks;
2184 int error, allerror = 0;
2188 for (i = level; i > 0; i--)
2189 lbnadd *= NINDIR(fs);
2191 * Get buffer of block pointers to be freed. This routine is not
2192 * called until the zero'ed inode has been written, so it is safe
2193 * to free blocks as they are encountered. Because the inode has
2194 * been zero'ed, calls to bmap on these blocks will fail. So, we
2195 * have to use the on-disk address and the block device for the
2196 * filesystem to look them up. If the file was deleted before its
2197 * indirect blocks were all written to disk, the routine that set
2198 * us up (deallocate_dependencies) will have arranged to leave
2199 * a complete copy of the indirect block in memory for our use.
2200 * Otherwise we have to read the blocks in from the disk.
2203 if ((bp = incore(ip->i_devvp, dbn)) != NULL &&
2204 (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2205 if (wk->wk_type != D_INDIRDEP ||
2206 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2207 (indirdep->ir_state & GOINGAWAY) == 0) {
2209 panic("indir_trunc: lost indirdep");
2211 WORKLIST_REMOVE(wk);
2212 WORKITEM_FREE(indirdep, D_INDIRDEP);
2213 if (LIST_FIRST(&bp->b_dep) != NULL) {
2215 panic("indir_trunc: dangling dep");
2220 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, NOCRED, &bp);
2225 * Recursively free indirect blocks.
2227 bap = (ufs_daddr_t *)bp->b_data;
2228 nblocks = btodb(fs->fs_bsize);
2229 for (i = NINDIR(fs) - 1; i >= 0; i--) {
2230 if ((nb = bap[i]) == 0)
2233 if ((error = indir_trunc(ip, fsbtodb(fs, nb),
2234 level - 1, lbn + (i * lbnadd), countp)) != 0)
2237 ffs_blkfree(ip, nb, fs->fs_bsize);
2240 bp->b_flags |= B_INVAL | B_NOCACHE;
2246 * Free an allocindir.
2247 * This routine must be called with splbio interrupts blocked.
2250 free_allocindir(aip, inodedep)
2251 struct allocindir *aip;
2252 struct inodedep *inodedep;
2254 struct freefrag *freefrag;
2257 if (lk.lkt_held == NOHOLDER)
2258 panic("free_allocindir: lock not held");
2260 if ((aip->ai_state & DEPCOMPLETE) == 0)
2261 LIST_REMOVE(aip, ai_deps);
2262 if (aip->ai_state & ONWORKLIST)
2263 WORKLIST_REMOVE(&aip->ai_list);
2264 LIST_REMOVE(aip, ai_next);
2265 if ((freefrag = aip->ai_freefrag) != NULL) {
2266 if (inodedep == NULL)
2267 add_to_worklist(&freefrag->ff_list);
2269 WORKLIST_INSERT(&inodedep->id_bufwait,
2270 &freefrag->ff_list);
2272 WORKITEM_FREE(aip, D_ALLOCINDIR);
2276 * Directory entry addition dependencies.
2278 * When adding a new directory entry, the inode (with its incremented link
2279 * count) must be written to disk before the directory entry's pointer to it.
2280 * Also, if the inode is newly allocated, the corresponding freemap must be
2281 * updated (on disk) before the directory entry's pointer. These requirements
2282 * are met via undo/redo on the directory entry's pointer, which consists
2283 * simply of the inode number.
2285 * As directory entries are added and deleted, the free space within a
2286 * directory block can become fragmented. The ufs file system will compact
2287 * a fragmented directory block to make space for a new entry. When this
2288 * occurs, the offsets of previously added entries change. Any "diradd"
2289 * dependency structures corresponding to these entries must be updated with
2294 * This routine is called after the in-memory inode's link
2295 * count has been incremented, but before the directory entry's
2296 * pointer to the inode has been set.
2299 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp)
2300 struct buf *bp; /* buffer containing directory block */
2301 struct inode *dp; /* inode for directory */
2302 off_t diroffset; /* offset of new entry in directory */
2303 long newinum; /* inode referenced by new directory entry */
2304 struct buf *newdirbp; /* non-NULL => contents of new mkdir */
2306 int offset; /* offset of new entry within directory block */
2307 ufs_lbn_t lbn; /* block in directory containing new entry */
2310 struct pagedep *pagedep;
2311 struct inodedep *inodedep;
2312 struct mkdir *mkdir1, *mkdir2;
2315 * Whiteouts have no dependencies.
2317 if (newinum == WINO) {
2318 if (newdirbp != NULL)
2324 lbn = lblkno(fs, diroffset);
2325 offset = blkoff(fs, diroffset);
2326 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2328 bzero(dap, sizeof(struct diradd));
2329 dap->da_list.wk_type = D_DIRADD;
2330 dap->da_offset = offset;
2331 dap->da_newinum = newinum;
2332 dap->da_state = ATTACHED;
2333 if (newdirbp == NULL) {
2334 dap->da_state |= DEPCOMPLETE;
2337 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2338 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2340 mkdir1->md_list.wk_type = D_MKDIR;
2341 mkdir1->md_state = MKDIR_BODY;
2342 mkdir1->md_diradd = dap;
2343 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2345 mkdir2->md_list.wk_type = D_MKDIR;
2346 mkdir2->md_state = MKDIR_PARENT;
2347 mkdir2->md_diradd = dap;
2349 * Dependency on "." and ".." being written to disk.
2351 mkdir1->md_buf = newdirbp;
2353 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2354 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2358 * Dependency on link count increase for parent directory
2361 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2362 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2363 dap->da_state &= ~MKDIR_PARENT;
2364 WORKITEM_FREE(mkdir2, D_MKDIR);
2366 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2367 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2371 * Link into parent directory pagedep to await its being written.
2373 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2374 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2375 dap->da_pagedep = pagedep;
2376 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2379 * Link into its inodedep. Put it on the id_bufwait list if the inode
2380 * is not yet written. If it is written, do the post-inode write
2381 * processing to put it on the id_pendinghd list.
2383 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2384 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2385 diradd_inode_written(dap, inodedep);
2387 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2392 * This procedure is called to change the offset of a directory
2393 * entry when compacting a directory block which must be owned
2394 * exclusively by the caller. Note that the actual entry movement
2395 * must be done in this procedure to ensure that no I/O completions
2396 * occur while the move is in progress.
2399 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2400 struct inode *dp; /* inode for directory */
2401 caddr_t base; /* address of dp->i_offset */
2402 caddr_t oldloc; /* address of old directory location */
2403 caddr_t newloc; /* address of new directory location */
2404 int entrysize; /* size of directory entry */
2406 int offset, oldoffset, newoffset;
2407 struct pagedep *pagedep;
2412 lbn = lblkno(dp->i_fs, dp->i_offset);
2413 offset = blkoff(dp->i_fs, dp->i_offset);
2414 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2416 oldoffset = offset + (oldloc - base);
2417 newoffset = offset + (newloc - base);
2419 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2420 if (dap->da_offset != oldoffset)
2422 dap->da_offset = newoffset;
2423 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2425 LIST_REMOVE(dap, da_pdlist);
2426 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2432 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2433 if (dap->da_offset == oldoffset) {
2434 dap->da_offset = newoffset;
2440 bcopy(oldloc, newloc, entrysize);
2445 * Free a diradd dependency structure. This routine must be called
2446 * with splbio interrupts blocked.
2452 struct dirrem *dirrem;
2453 struct pagedep *pagedep;
2454 struct inodedep *inodedep;
2455 struct mkdir *mkdir, *nextmd;
2458 if (lk.lkt_held == NOHOLDER)
2459 panic("free_diradd: lock not held");
2461 WORKLIST_REMOVE(&dap->da_list);
2462 LIST_REMOVE(dap, da_pdlist);
2463 if ((dap->da_state & DIRCHG) == 0) {
2464 pagedep = dap->da_pagedep;
2466 dirrem = dap->da_previous;
2467 pagedep = dirrem->dm_pagedep;
2468 dirrem->dm_dirinum = pagedep->pd_ino;
2469 add_to_worklist(&dirrem->dm_list);
2471 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2473 (void) free_inodedep(inodedep);
2474 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2475 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2476 nextmd = LIST_NEXT(mkdir, md_mkdirs);
2477 if (mkdir->md_diradd != dap)
2479 dap->da_state &= ~mkdir->md_state;
2480 WORKLIST_REMOVE(&mkdir->md_list);
2481 LIST_REMOVE(mkdir, md_mkdirs);
2482 WORKITEM_FREE(mkdir, D_MKDIR);
2484 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2486 panic("free_diradd: unfound ref");
2489 WORKITEM_FREE(dap, D_DIRADD);
2493 * Directory entry removal dependencies.
2495 * When removing a directory entry, the entry's inode pointer must be
2496 * zero'ed on disk before the corresponding inode's link count is decremented
2497 * (possibly freeing the inode for re-use). This dependency is handled by
2498 * updating the directory entry but delaying the inode count reduction until
2499 * after the directory block has been written to disk. After this point, the
2500 * inode count can be decremented whenever it is convenient.
2504 * This routine should be called immediately after removing
2505 * a directory entry. The inode's link count should not be
2506 * decremented by the calling procedure -- the soft updates
2507 * code will do this task when it is safe.
2510 softdep_setup_remove(bp, dp, ip, isrmdir)
2511 struct buf *bp; /* buffer containing directory block */
2512 struct inode *dp; /* inode for the directory being modified */
2513 struct inode *ip; /* inode for directory entry being removed */
2514 int isrmdir; /* indicates if doing RMDIR */
2516 struct dirrem *dirrem, *prevdirrem;
2519 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2521 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2524 * If the COMPLETE flag is clear, then there were no active
2525 * entries and we want to roll back to a zeroed entry until
2526 * the new inode is committed to disk. If the COMPLETE flag is
2527 * set then we have deleted an entry that never made it to
2528 * disk. If the entry we deleted resulted from a name change,
2529 * then the old name still resides on disk. We cannot delete
2530 * its inode (returned to us in prevdirrem) until the zeroed
2531 * directory entry gets to disk. The new inode has never been
2532 * referenced on the disk, so can be deleted immediately.
2534 if ((dirrem->dm_state & COMPLETE) == 0) {
2535 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2539 if (prevdirrem != NULL)
2540 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2541 prevdirrem, dm_next);
2542 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2544 handle_workitem_remove(dirrem);
2549 * Allocate a new dirrem if appropriate and return it along with
2550 * its associated pagedep. Called without a lock, returns with lock.
2552 static long num_dirrem; /* number of dirrem allocated */
2553 static struct dirrem *
2554 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2555 struct buf *bp; /* buffer containing directory block */
2556 struct inode *dp; /* inode for the directory being modified */
2557 struct inode *ip; /* inode for directory entry being removed */
2558 int isrmdir; /* indicates if doing RMDIR */
2559 struct dirrem **prevdirremp; /* previously referenced inode, if any */
2564 struct dirrem *dirrem;
2565 struct pagedep *pagedep;
2568 * Whiteouts have no deletion dependencies.
2571 panic("newdirrem: whiteout");
2573 * If we are over our limit, try to improve the situation.
2574 * Limiting the number of dirrem structures will also limit
2575 * the number of freefile and freeblks structures.
2577 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2578 (void) request_cleanup(FLUSH_REMOVE, 0);
2580 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
2581 M_DIRREM, M_SOFTDEP_FLAGS);
2582 bzero(dirrem, sizeof(struct dirrem));
2583 dirrem->dm_list.wk_type = D_DIRREM;
2584 dirrem->dm_state = isrmdir ? RMDIR : 0;
2585 dirrem->dm_mnt = ITOV(ip)->v_mount;
2586 dirrem->dm_oldinum = ip->i_number;
2587 *prevdirremp = NULL;
2590 lbn = lblkno(dp->i_fs, dp->i_offset);
2591 offset = blkoff(dp->i_fs, dp->i_offset);
2592 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2593 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2594 dirrem->dm_pagedep = pagedep;
2596 * Check for a diradd dependency for the same directory entry.
2597 * If present, then both dependencies become obsolete and can
2598 * be de-allocated. Check for an entry on both the pd_dirraddhd
2599 * list and the pd_pendinghd list.
2602 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
2603 if (dap->da_offset == offset)
2607 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
2608 if (dap->da_offset == offset)
2614 * Must be ATTACHED at this point.
2616 if ((dap->da_state & ATTACHED) == 0) {
2618 panic("newdirrem: not ATTACHED");
2620 if (dap->da_newinum != ip->i_number) {
2622 panic("newdirrem: inum %d should be %d",
2623 ip->i_number, dap->da_newinum);
2626 * If we are deleting a changed name that never made it to disk,
2627 * then return the dirrem describing the previous inode (which
2628 * represents the inode currently referenced from this entry on disk).
2630 if ((dap->da_state & DIRCHG) != 0) {
2631 *prevdirremp = dap->da_previous;
2632 dap->da_state &= ~DIRCHG;
2633 dap->da_pagedep = pagedep;
2636 * We are deleting an entry that never made it to disk.
2637 * Mark it COMPLETE so we can delete its inode immediately.
2639 dirrem->dm_state |= COMPLETE;
2645 * Directory entry change dependencies.
2647 * Changing an existing directory entry requires that an add operation
2648 * be completed first followed by a deletion. The semantics for the addition
2649 * are identical to the description of adding a new entry above except
2650 * that the rollback is to the old inode number rather than zero. Once
2651 * the addition dependency is completed, the removal is done as described
2652 * in the removal routine above.
2656 * This routine should be called immediately after changing
2657 * a directory entry. The inode's link count should not be
2658 * decremented by the calling procedure -- the soft updates
2659 * code will perform this task when it is safe.
2662 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
2663 struct buf *bp; /* buffer containing directory block */
2664 struct inode *dp; /* inode for the directory being modified */
2665 struct inode *ip; /* inode for directory entry being removed */
2666 long newinum; /* new inode number for changed entry */
2667 int isrmdir; /* indicates if doing RMDIR */
2670 struct diradd *dap = NULL;
2671 struct dirrem *dirrem, *prevdirrem;
2672 struct pagedep *pagedep;
2673 struct inodedep *inodedep;
2675 offset = blkoff(dp->i_fs, dp->i_offset);
2678 * Whiteouts do not need diradd dependencies.
2680 if (newinum != WINO) {
2681 MALLOC(dap, struct diradd *, sizeof(struct diradd),
2682 M_DIRADD, M_SOFTDEP_FLAGS);
2683 bzero(dap, sizeof(struct diradd));
2684 dap->da_list.wk_type = D_DIRADD;
2685 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2686 dap->da_offset = offset;
2687 dap->da_newinum = newinum;
2691 * Allocate a new dirrem and ACQUIRE_LOCK.
2693 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2694 pagedep = dirrem->dm_pagedep;
2696 * The possible values for isrmdir:
2697 * 0 - non-directory file rename
2698 * 1 - directory rename within same directory
2699 * inum - directory rename to new directory of given inode number
2700 * When renaming to a new directory, we are both deleting and
2701 * creating a new directory entry, so the link count on the new
2702 * directory should not change. Thus we do not need the followup
2703 * dirrem which is usually done in handle_workitem_remove. We set
2704 * the DIRCHG flag to tell handle_workitem_remove to skip the
2708 dirrem->dm_state |= DIRCHG;
2711 * Whiteouts have no additional dependencies,
2712 * so just put the dirrem on the correct list.
2714 if (newinum == WINO) {
2715 if ((dirrem->dm_state & COMPLETE) == 0) {
2716 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2719 dirrem->dm_dirinum = pagedep->pd_ino;
2720 add_to_worklist(&dirrem->dm_list);
2727 * If the COMPLETE flag is clear, then there were no active
2728 * entries and we want to roll back to the previous inode until
2729 * the new inode is committed to disk. If the COMPLETE flag is
2730 * set, then we have deleted an entry that never made it to disk.
2731 * If the entry we deleted resulted from a name change, then the old
2732 * inode reference still resides on disk. Any rollback that we do
2733 * needs to be to that old inode (returned to us in prevdirrem). If
2734 * the entry we deleted resulted from a create, then there is
2735 * no entry on the disk, so we want to roll back to zero rather
2736 * than the uncommitted inode. In either of the COMPLETE cases we
2737 * want to immediately free the unwritten and unreferenced inode.
2739 if ((dirrem->dm_state & COMPLETE) == 0) {
2740 dap->da_previous = dirrem;
2742 if (prevdirrem != NULL) {
2743 dap->da_previous = prevdirrem;
2745 dap->da_state &= ~DIRCHG;
2746 dap->da_pagedep = pagedep;
2748 dirrem->dm_dirinum = pagedep->pd_ino;
2749 add_to_worklist(&dirrem->dm_list);
2752 * Link into its inodedep. Put it on the id_bufwait list if the inode
2753 * is not yet written. If it is written, do the post-inode write
2754 * processing to put it on the id_pendinghd list.
2756 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2757 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2758 dap->da_state |= COMPLETE;
2759 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2760 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2762 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2764 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2770 * Called whenever the link count on an inode is changed.
2771 * It creates an inode dependency so that the new reference(s)
2772 * to the inode cannot be committed to disk until the updated
2773 * inode has been written.
2776 softdep_change_linkcnt(ip)
2777 struct inode *ip; /* the inode with the increased link count */
2779 struct inodedep *inodedep;
2782 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2783 if (ip->i_nlink < ip->i_effnlink) {
2785 panic("softdep_change_linkcnt: bad delta");
2787 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2792 * This workitem decrements the inode's link count.
2793 * If the link count reaches zero, the file is removed.
2796 handle_workitem_remove(dirrem)
2797 struct dirrem *dirrem;
2799 struct thread *td = curthread; /* XXX */
2801 struct inodedep *inodedep;
2807 KKASSERT(td->td_proc);
2808 cred = td->td_proc->p_ucred;
2809 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) {
2810 softdep_error("handle_workitem_remove: vget", error);
2815 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
2817 panic("handle_workitem_remove: lost inodedep");
2820 * Normal file deletion.
2822 if ((dirrem->dm_state & RMDIR) == 0) {
2824 ip->i_flag |= IN_CHANGE;
2825 if (ip->i_nlink < ip->i_effnlink) {
2827 panic("handle_workitem_remove: bad file delta");
2829 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2833 WORKITEM_FREE(dirrem, D_DIRREM);
2837 * Directory deletion. Decrement reference count for both the
2838 * just deleted parent directory entry and the reference for ".".
2839 * Next truncate the directory to length zero. When the
2840 * truncation completes, arrange to have the reference count on
2841 * the parent decremented to account for the loss of "..".
2844 ip->i_flag |= IN_CHANGE;
2845 if (ip->i_nlink < ip->i_effnlink) {
2847 panic("handle_workitem_remove: bad dir delta");
2849 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2851 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0,cred, td)) != 0)
2852 softdep_error("handle_workitem_remove: truncate", error);
2854 * Rename a directory to a new parent. Since, we are both deleting
2855 * and creating a new directory entry, the link count on the new
2856 * directory should not change. Thus we skip the followup dirrem.
2858 if (dirrem->dm_state & DIRCHG) {
2861 WORKITEM_FREE(dirrem, D_DIRREM);
2865 * If the inodedep does not exist, then the zero'ed inode has
2866 * been written to disk. If the allocated inode has never been
2867 * written to disk, then the on-disk inode is zero'ed. In either
2868 * case we can remove the file immediately.
2871 dirrem->dm_state = 0;
2872 oldinum = dirrem->dm_oldinum;
2873 dirrem->dm_oldinum = dirrem->dm_dirinum;
2874 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
2875 check_inode_unwritten(inodedep)) {
2878 handle_workitem_remove(dirrem);
2881 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
2887 * Inode de-allocation dependencies.
2889 * When an inode's link count is reduced to zero, it can be de-allocated. We
2890 * found it convenient to postpone de-allocation until after the inode is
2891 * written to disk with its new link count (zero). At this point, all of the
2892 * on-disk inode's block pointers are nullified and, with careful dependency
2893 * list ordering, all dependencies related to the inode will be satisfied and
2894 * the corresponding dependency structures de-allocated. So, if/when the
2895 * inode is reused, there will be no mixing of old dependencies with new
2896 * ones. This artificial dependency is set up by the block de-allocation
2897 * procedure above (softdep_setup_freeblocks) and completed by the
2898 * following procedure.
2901 handle_workitem_freefile(freefile)
2902 struct freefile *freefile;
2906 struct inodedep *idp;
2911 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp);
2914 panic("handle_workitem_freefile: inodedep survived");
2916 tip.i_devvp = freefile->fx_devvp;
2917 tip.i_dev = freefile->fx_devvp->v_rdev;
2918 tip.i_fs = freefile->fx_fs;
2920 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
2921 softdep_error("handle_workitem_freefile", error);
2922 WORKITEM_FREE(freefile, D_FREEFILE);
2928 * The dependency structures constructed above are most actively used when file
2929 * system blocks are written to disk. No constraints are placed on when a
2930 * block can be written, but unsatisfied update dependencies are made safe by
2931 * modifying (or replacing) the source memory for the duration of the disk
2932 * write. When the disk write completes, the memory block is again brought
2935 * In-core inode structure reclamation.
2937 * Because there are a finite number of "in-core" inode structures, they are
2938 * reused regularly. By transferring all inode-related dependencies to the
2939 * in-memory inode block and indexing them separately (via "inodedep"s), we
2940 * can allow "in-core" inode structures to be reused at any time and avoid
2941 * any increase in contention.
2943 * Called just before entering the device driver to initiate a new disk I/O.
2944 * The buffer must be locked, thus, no I/O completion operations can occur
2945 * while we are manipulating its associated dependencies.
2948 softdep_disk_io_initiation(bp)
2949 struct buf *bp; /* structure describing disk write to occur */
2951 struct worklist *wk, *nextwk;
2952 struct indirdep *indirdep;
2955 * We only care about write operations. There should never
2956 * be dependencies for reads.
2958 if (bp->b_flags & B_READ)
2959 panic("softdep_disk_io_initiation: read");
2961 * Do any necessary pre-I/O processing.
2963 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
2964 nextwk = LIST_NEXT(wk, wk_list);
2965 switch (wk->wk_type) {
2968 initiate_write_filepage(WK_PAGEDEP(wk), bp);
2972 initiate_write_inodeblock(WK_INODEDEP(wk), bp);
2976 indirdep = WK_INDIRDEP(wk);
2977 if (indirdep->ir_state & GOINGAWAY)
2978 panic("disk_io_initiation: indirdep gone");
2980 * If there are no remaining dependencies, this
2981 * will be writing the real pointers, so the
2982 * dependency can be freed.
2984 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
2985 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
2986 brelse(indirdep->ir_savebp);
2987 /* inline expand WORKLIST_REMOVE(wk); */
2988 wk->wk_state &= ~ONWORKLIST;
2989 LIST_REMOVE(wk, wk_list);
2990 WORKITEM_FREE(indirdep, D_INDIRDEP);
2994 * Replace up-to-date version with safe version.
2996 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
2997 M_INDIRDEP, M_SOFTDEP_FLAGS);
2999 indirdep->ir_state &= ~ATTACHED;
3000 indirdep->ir_state |= UNDONE;
3001 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3002 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3014 panic("handle_disk_io_initiation: Unexpected type %s",
3015 TYPENAME(wk->wk_type));
3022 * Called from within the procedure above to deal with unsatisfied
3023 * allocation dependencies in a directory. The buffer must be locked,
3024 * thus, no I/O completion operations can occur while we are
3025 * manipulating its associated dependencies.
3028 initiate_write_filepage(pagedep, bp)
3029 struct pagedep *pagedep;
3036 if (pagedep->pd_state & IOSTARTED) {
3038 * This can only happen if there is a driver that does not
3039 * understand chaining. Here biodone will reissue the call
3040 * to strategy for the incomplete buffers.
3042 printf("initiate_write_filepage: already started\n");
3045 pagedep->pd_state |= IOSTARTED;
3047 for (i = 0; i < DAHASHSZ; i++) {
3048 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3049 ep = (struct direct *)
3050 ((char *)bp->b_data + dap->da_offset);
3051 if (ep->d_ino != dap->da_newinum) {
3053 panic("%s: dir inum %d != new %d",
3054 "initiate_write_filepage",
3055 ep->d_ino, dap->da_newinum);
3057 if (dap->da_state & DIRCHG)
3058 ep->d_ino = dap->da_previous->dm_oldinum;
3061 dap->da_state &= ~ATTACHED;
3062 dap->da_state |= UNDONE;
3069 * Called from within the procedure above to deal with unsatisfied
3070 * allocation dependencies in an inodeblock. The buffer must be
3071 * locked, thus, no I/O completion operations can occur while we
3072 * are manipulating its associated dependencies.
3075 initiate_write_inodeblock(inodedep, bp)
3076 struct inodedep *inodedep;
3077 struct buf *bp; /* The inode block */
3079 struct allocdirect *adp, *lastadp;
3082 ufs_lbn_t prevlbn = 0;
3085 if (inodedep->id_state & IOSTARTED)
3086 panic("initiate_write_inodeblock: already started");
3087 inodedep->id_state |= IOSTARTED;
3088 fs = inodedep->id_fs;
3089 dp = (struct dinode *)bp->b_data +
3090 ino_to_fsbo(fs, inodedep->id_ino);
3092 * If the bitmap is not yet written, then the allocated
3093 * inode cannot be written to disk.
3095 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3096 if (inodedep->id_savedino != NULL)
3097 panic("initiate_write_inodeblock: already doing I/O");
3098 MALLOC(inodedep->id_savedino, struct dinode *,
3099 sizeof(struct dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3100 *inodedep->id_savedino = *dp;
3101 bzero((caddr_t)dp, sizeof(struct dinode));
3105 * If no dependencies, then there is nothing to roll back.
3107 inodedep->id_savedsize = dp->di_size;
3108 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3111 * Set the dependencies to busy.
3114 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3115 adp = TAILQ_NEXT(adp, ad_next)) {
3117 if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3119 panic("softdep_write_inodeblock: lbn order");
3121 prevlbn = adp->ad_lbn;
3122 if (adp->ad_lbn < NDADDR &&
3123 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3125 panic("%s: direct pointer #%ld mismatch %d != %d",
3126 "softdep_write_inodeblock", adp->ad_lbn,
3127 dp->di_db[adp->ad_lbn], adp->ad_newblkno);
3129 if (adp->ad_lbn >= NDADDR &&
3130 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3132 panic("%s: indirect pointer #%ld mismatch %d != %d",
3133 "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
3134 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
3136 deplist |= 1 << adp->ad_lbn;
3137 if ((adp->ad_state & ATTACHED) == 0) {
3139 panic("softdep_write_inodeblock: Unknown state 0x%x",
3142 #endif /* DIAGNOSTIC */
3143 adp->ad_state &= ~ATTACHED;
3144 adp->ad_state |= UNDONE;
3147 * The on-disk inode cannot claim to be any larger than the last
3148 * fragment that has been written. Otherwise, the on-disk inode
3149 * might have fragments that were not the last block in the file
3150 * which would corrupt the filesystem.
3152 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3153 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3154 if (adp->ad_lbn >= NDADDR)
3156 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3157 /* keep going until hitting a rollback to a frag */
3158 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3160 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3161 for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3163 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3165 panic("softdep_write_inodeblock: lost dep1");
3167 #endif /* DIAGNOSTIC */
3170 for (i = 0; i < NIADDR; i++) {
3172 if (dp->di_ib[i] != 0 &&
3173 (deplist & ((1 << NDADDR) << i)) == 0) {
3175 panic("softdep_write_inodeblock: lost dep2");
3177 #endif /* DIAGNOSTIC */
3184 * If we have zero'ed out the last allocated block of the file,
3185 * roll back the size to the last currently allocated block.
3186 * We know that this last allocated block is a full-sized as
3187 * we already checked for fragments in the loop above.
3189 if (lastadp != NULL &&
3190 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3191 for (i = lastadp->ad_lbn; i >= 0; i--)
3192 if (dp->di_db[i] != 0)
3194 dp->di_size = (i + 1) * fs->fs_bsize;
3197 * The only dependencies are for indirect blocks.
3199 * The file size for indirect block additions is not guaranteed.
3200 * Such a guarantee would be non-trivial to achieve. The conventional
3201 * synchronous write implementation also does not make this guarantee.
3202 * Fsck should catch and fix discrepancies. Arguably, the file size
3203 * can be over-estimated without destroying integrity when the file
3204 * moves into the indirect blocks (i.e., is large). If we want to
3205 * postpone fsck, we are stuck with this argument.
3207 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3208 dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3213 * This routine is called during the completion interrupt
3214 * service routine for a disk write (from the procedure called
3215 * by the device driver to inform the file system caches of
3216 * a request completion). It should be called early in this
3217 * procedure, before the block is made available to other
3218 * processes or other routines are called.
3221 softdep_disk_write_complete(bp)
3222 struct buf *bp; /* describes the completed disk write */
3224 struct worklist *wk;
3225 struct workhead reattach;
3226 struct newblk *newblk;
3227 struct allocindir *aip;
3228 struct allocdirect *adp;
3229 struct indirdep *indirdep;
3230 struct inodedep *inodedep;
3231 struct bmsafemap *bmsafemap;
3234 if (lk.lkt_held != NOHOLDER)
3235 panic("softdep_disk_write_complete: lock is held");
3236 lk.lkt_held = SPECIAL_FLAG;
3238 LIST_INIT(&reattach);
3239 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3240 WORKLIST_REMOVE(wk);
3241 switch (wk->wk_type) {
3244 if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3245 WORKLIST_INSERT(&reattach, wk);
3249 if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3250 WORKLIST_INSERT(&reattach, wk);
3254 bmsafemap = WK_BMSAFEMAP(wk);
3255 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3256 newblk->nb_state |= DEPCOMPLETE;
3257 newblk->nb_bmsafemap = NULL;
3258 LIST_REMOVE(newblk, nb_deps);
3261 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3262 adp->ad_state |= DEPCOMPLETE;
3264 LIST_REMOVE(adp, ad_deps);
3265 handle_allocdirect_partdone(adp);
3268 LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3269 aip->ai_state |= DEPCOMPLETE;
3271 LIST_REMOVE(aip, ai_deps);
3272 handle_allocindir_partdone(aip);
3275 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3276 inodedep->id_state |= DEPCOMPLETE;
3277 LIST_REMOVE(inodedep, id_deps);
3278 inodedep->id_buf = NULL;
3280 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3284 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3288 adp = WK_ALLOCDIRECT(wk);
3289 adp->ad_state |= COMPLETE;
3290 handle_allocdirect_partdone(adp);
3294 aip = WK_ALLOCINDIR(wk);
3295 aip->ai_state |= COMPLETE;
3296 handle_allocindir_partdone(aip);
3300 indirdep = WK_INDIRDEP(wk);
3301 if (indirdep->ir_state & GOINGAWAY) {
3302 lk.lkt_held = NOHOLDER;
3303 panic("disk_write_complete: indirdep gone");
3305 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3306 FREE(indirdep->ir_saveddata, M_INDIRDEP);
3307 indirdep->ir_saveddata = 0;
3308 indirdep->ir_state &= ~UNDONE;
3309 indirdep->ir_state |= ATTACHED;
3310 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3311 handle_allocindir_partdone(aip);
3312 if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
3313 lk.lkt_held = NOHOLDER;
3314 panic("disk_write_complete: not gone");
3317 WORKLIST_INSERT(&reattach, wk);
3318 if ((bp->b_flags & B_DELWRI) == 0)
3319 stat_indir_blk_ptrs++;
3324 lk.lkt_held = NOHOLDER;
3325 panic("handle_disk_write_complete: Unknown type %s",
3326 TYPENAME(wk->wk_type));
3331 * Reattach any requests that must be redone.
3333 while ((wk = LIST_FIRST(&reattach)) != NULL) {
3334 WORKLIST_REMOVE(wk);
3335 WORKLIST_INSERT(&bp->b_dep, wk);
3338 if (lk.lkt_held != SPECIAL_FLAG)
3339 panic("softdep_disk_write_complete: lock lost");
3340 lk.lkt_held = NOHOLDER;
3345 * Called from within softdep_disk_write_complete above. Note that
3346 * this routine is always called from interrupt level with further
3347 * splbio interrupts blocked.
3350 handle_allocdirect_partdone(adp)
3351 struct allocdirect *adp; /* the completed allocdirect */
3353 struct allocdirect *listadp;
3354 struct inodedep *inodedep;
3357 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3359 if (adp->ad_buf != NULL) {
3360 lk.lkt_held = NOHOLDER;
3361 panic("handle_allocdirect_partdone: dangling dep");
3364 * The on-disk inode cannot claim to be any larger than the last
3365 * fragment that has been written. Otherwise, the on-disk inode
3366 * might have fragments that were not the last block in the file
3367 * which would corrupt the filesystem. Thus, we cannot free any
3368 * allocdirects after one whose ad_oldblkno claims a fragment as
3369 * these blocks must be rolled back to zero before writing the inode.
3370 * We check the currently active set of allocdirects in id_inoupdt.
3372 inodedep = adp->ad_inodedep;
3373 bsize = inodedep->id_fs->fs_bsize;
3374 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) {
3375 /* found our block */
3378 /* continue if ad_oldlbn is not a fragment */
3379 if (listadp->ad_oldsize == 0 ||
3380 listadp->ad_oldsize == bsize)
3382 /* hit a fragment */
3386 * If we have reached the end of the current list without
3387 * finding the just finished dependency, then it must be
3388 * on the future dependency list. Future dependencies cannot
3389 * be freed until they are moved to the current list.
3391 if (listadp == NULL) {
3393 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next)
3394 /* found our block */
3397 if (listadp == NULL) {
3398 lk.lkt_held = NOHOLDER;
3399 panic("handle_allocdirect_partdone: lost dep");
3405 * If we have found the just finished dependency, then free
3406 * it along with anything that follows it that is complete.
3408 for (; adp; adp = listadp) {
3409 listadp = TAILQ_NEXT(adp, ad_next);
3410 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3412 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
3417 * Called from within softdep_disk_write_complete above. Note that
3418 * this routine is always called from interrupt level with further
3419 * splbio interrupts blocked.
3422 handle_allocindir_partdone(aip)
3423 struct allocindir *aip; /* the completed allocindir */
3425 struct indirdep *indirdep;
3427 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
3429 if (aip->ai_buf != NULL) {
3430 lk.lkt_held = NOHOLDER;
3431 panic("handle_allocindir_partdone: dangling dependency");
3433 indirdep = aip->ai_indirdep;
3434 if (indirdep->ir_state & UNDONE) {
3435 LIST_REMOVE(aip, ai_next);
3436 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
3439 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
3441 LIST_REMOVE(aip, ai_next);
3442 if (aip->ai_freefrag != NULL)
3443 add_to_worklist(&aip->ai_freefrag->ff_list);
3444 WORKITEM_FREE(aip, D_ALLOCINDIR);
3448 * Called from within softdep_disk_write_complete above to restore
3449 * in-memory inode block contents to their most up-to-date state. Note
3450 * that this routine is always called from interrupt level with further
3451 * splbio interrupts blocked.
3454 handle_written_inodeblock(inodedep, bp)
3455 struct inodedep *inodedep;
3456 struct buf *bp; /* buffer containing the inode block */
3458 struct worklist *wk, *filefree;
3459 struct allocdirect *adp, *nextadp;
3463 if ((inodedep->id_state & IOSTARTED) == 0) {
3464 lk.lkt_held = NOHOLDER;
3465 panic("handle_written_inodeblock: not started");
3467 inodedep->id_state &= ~IOSTARTED;
3468 inodedep->id_state |= COMPLETE;
3469 dp = (struct dinode *)bp->b_data +
3470 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
3472 * If we had to rollback the inode allocation because of
3473 * bitmaps being incomplete, then simply restore it.
3474 * Keep the block dirty so that it will not be reclaimed until
3475 * all associated dependencies have been cleared and the
3476 * corresponding updates written to disk.
3478 if (inodedep->id_savedino != NULL) {
3479 *dp = *inodedep->id_savedino;
3480 FREE(inodedep->id_savedino, M_INODEDEP);
3481 inodedep->id_savedino = NULL;
3482 if ((bp->b_flags & B_DELWRI) == 0)
3483 stat_inode_bitmap++;
3488 * Roll forward anything that had to be rolled back before
3489 * the inode could be updated.
3492 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
3493 nextadp = TAILQ_NEXT(adp, ad_next);
3494 if (adp->ad_state & ATTACHED) {
3495 lk.lkt_held = NOHOLDER;
3496 panic("handle_written_inodeblock: new entry");
3498 if (adp->ad_lbn < NDADDR) {
3499 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) {
3500 lk.lkt_held = NOHOLDER;
3501 panic("%s: %s #%ld mismatch %d != %d",
3502 "handle_written_inodeblock",
3503 "direct pointer", adp->ad_lbn,
3504 dp->di_db[adp->ad_lbn], adp->ad_oldblkno);
3506 dp->di_db[adp->ad_lbn] = adp->ad_newblkno;
3508 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) {
3509 lk.lkt_held = NOHOLDER;
3510 panic("%s: %s #%ld allocated as %d",
3511 "handle_written_inodeblock",
3512 "indirect pointer", adp->ad_lbn - NDADDR,
3513 dp->di_ib[adp->ad_lbn - NDADDR]);
3515 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno;
3517 adp->ad_state &= ~UNDONE;
3518 adp->ad_state |= ATTACHED;
3521 if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
3522 stat_direct_blk_ptrs++;
3524 * Reset the file size to its most up-to-date value.
3526 if (inodedep->id_savedsize == -1) {
3527 lk.lkt_held = NOHOLDER;
3528 panic("handle_written_inodeblock: bad size");
3530 if (dp->di_size != inodedep->id_savedsize) {
3531 dp->di_size = inodedep->id_savedsize;
3534 inodedep->id_savedsize = -1;
3536 * If there were any rollbacks in the inode block, then it must be
3537 * marked dirty so that its will eventually get written back in
3543 * Process any allocdirects that completed during the update.
3545 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
3546 handle_allocdirect_partdone(adp);
3548 * Process deallocations that were held pending until the
3549 * inode had been written to disk. Freeing of the inode
3550 * is delayed until after all blocks have been freed to
3551 * avoid creation of new <vfsid, inum, lbn> triples
3552 * before the old ones have been deleted.
3555 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
3556 WORKLIST_REMOVE(wk);
3557 switch (wk->wk_type) {
3561 * We defer adding filefree to the worklist until
3562 * all other additions have been made to ensure
3563 * that it will be done after all the old blocks
3566 if (filefree != NULL) {
3567 lk.lkt_held = NOHOLDER;
3568 panic("handle_written_inodeblock: filefree");
3574 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
3578 diradd_inode_written(WK_DIRADD(wk), inodedep);
3584 add_to_worklist(wk);
3588 lk.lkt_held = NOHOLDER;
3589 panic("handle_written_inodeblock: Unknown type %s",
3590 TYPENAME(wk->wk_type));
3594 if (filefree != NULL) {
3595 if (free_inodedep(inodedep) == 0) {
3596 lk.lkt_held = NOHOLDER;
3597 panic("handle_written_inodeblock: live inodedep");
3599 add_to_worklist(filefree);
3604 * If no outstanding dependencies, free it.
3606 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0)
3608 return (hadchanges);
3612 * Process a diradd entry after its dependent inode has been written.
3613 * This routine must be called with splbio interrupts blocked.
3616 diradd_inode_written(dap, inodedep)
3618 struct inodedep *inodedep;
3620 struct pagedep *pagedep;
3622 dap->da_state |= COMPLETE;
3623 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3624 if (dap->da_state & DIRCHG)
3625 pagedep = dap->da_previous->dm_pagedep;
3627 pagedep = dap->da_pagedep;
3628 LIST_REMOVE(dap, da_pdlist);
3629 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3631 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3635 * Handle the completion of a mkdir dependency.
3638 handle_written_mkdir(mkdir, type)
3639 struct mkdir *mkdir;
3643 struct pagedep *pagedep;
3645 if (mkdir->md_state != type) {
3646 lk.lkt_held = NOHOLDER;
3647 panic("handle_written_mkdir: bad type");
3649 dap = mkdir->md_diradd;
3650 dap->da_state &= ~type;
3651 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
3652 dap->da_state |= DEPCOMPLETE;
3653 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3654 if (dap->da_state & DIRCHG)
3655 pagedep = dap->da_previous->dm_pagedep;
3657 pagedep = dap->da_pagedep;
3658 LIST_REMOVE(dap, da_pdlist);
3659 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3661 LIST_REMOVE(mkdir, md_mkdirs);
3662 WORKITEM_FREE(mkdir, D_MKDIR);
3666 * Called from within softdep_disk_write_complete above.
3667 * A write operation was just completed. Removed inodes can
3668 * now be freed and associated block pointers may be committed.
3669 * Note that this routine is always called from interrupt level
3670 * with further splbio interrupts blocked.
3673 handle_written_filepage(pagedep, bp)
3674 struct pagedep *pagedep;
3675 struct buf *bp; /* buffer containing the written page */
3677 struct dirrem *dirrem;
3678 struct diradd *dap, *nextdap;
3682 if ((pagedep->pd_state & IOSTARTED) == 0) {
3683 lk.lkt_held = NOHOLDER;
3684 panic("handle_written_filepage: not started");
3686 pagedep->pd_state &= ~IOSTARTED;
3688 * Process any directory removals that have been committed.
3690 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
3691 LIST_REMOVE(dirrem, dm_next);
3692 dirrem->dm_dirinum = pagedep->pd_ino;
3693 add_to_worklist(&dirrem->dm_list);
3696 * Free any directory additions that have been committed.
3698 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
3701 * Uncommitted directory entries must be restored.
3703 for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
3704 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
3706 nextdap = LIST_NEXT(dap, da_pdlist);
3707 if (dap->da_state & ATTACHED) {
3708 lk.lkt_held = NOHOLDER;
3709 panic("handle_written_filepage: attached");
3711 ep = (struct direct *)
3712 ((char *)bp->b_data + dap->da_offset);
3713 ep->d_ino = dap->da_newinum;
3714 dap->da_state &= ~UNDONE;
3715 dap->da_state |= ATTACHED;
3718 * If the inode referenced by the directory has
3719 * been written out, then the dependency can be
3720 * moved to the pending list.
3722 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3723 LIST_REMOVE(dap, da_pdlist);
3724 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
3730 * If there were any rollbacks in the directory, then it must be
3731 * marked dirty so that its will eventually get written back in
3735 if ((bp->b_flags & B_DELWRI) == 0)
3740 * If no dependencies remain, the pagedep will be freed.
3741 * Otherwise it will remain to update the page before it
3742 * is written back to disk.
3744 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) {
3745 for (i = 0; i < DAHASHSZ; i++)
3746 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
3748 if (i == DAHASHSZ) {
3749 LIST_REMOVE(pagedep, pd_hash);
3750 WORKITEM_FREE(pagedep, D_PAGEDEP);
3758 * Writing back in-core inode structures.
3760 * The file system only accesses an inode's contents when it occupies an
3761 * "in-core" inode structure. These "in-core" structures are separate from
3762 * the page frames used to cache inode blocks. Only the latter are
3763 * transferred to/from the disk. So, when the updated contents of the
3764 * "in-core" inode structure are copied to the corresponding in-memory inode
3765 * block, the dependencies are also transferred. The following procedure is
3766 * called when copying a dirty "in-core" inode to a cached inode block.
3770 * Called when an inode is loaded from disk. If the effective link count
3771 * differed from the actual link count when it was last flushed, then we
3772 * need to ensure that the correct effective link count is put back.
3775 softdep_load_inodeblock(ip)
3776 struct inode *ip; /* the "in_core" copy of the inode */
3778 struct inodedep *inodedep;
3781 * Check for alternate nlink count.
3783 ip->i_effnlink = ip->i_nlink;
3785 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3789 ip->i_effnlink -= inodedep->id_nlinkdelta;
3794 * This routine is called just before the "in-core" inode
3795 * information is to be copied to the in-memory inode block.
3796 * Recall that an inode block contains several inodes. If
3797 * the force flag is set, then the dependencies will be
3798 * cleared so that the update can always be made. Note that
3799 * the buffer is locked when this routine is called, so we
3800 * will never be in the middle of writing the inode block
3804 softdep_update_inodeblock(ip, bp, waitfor)
3805 struct inode *ip; /* the "in_core" copy of the inode */
3806 struct buf *bp; /* the buffer containing the inode block */
3807 int waitfor; /* nonzero => update must be allowed */
3809 struct inodedep *inodedep;
3810 struct worklist *wk;
3814 * If the effective link count is not equal to the actual link
3815 * count, then we must track the difference in an inodedep while
3816 * the inode is (potentially) tossed out of the cache. Otherwise,
3817 * if there is no existing inodedep, then there are no dependencies
3821 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3823 if (ip->i_effnlink != ip->i_nlink)
3824 panic("softdep_update_inodeblock: bad link count");
3827 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
3829 panic("softdep_update_inodeblock: bad delta");
3832 * Changes have been initiated. Anything depending on these
3833 * changes cannot occur until this inode has been written.
3835 inodedep->id_state &= ~COMPLETE;
3836 if ((inodedep->id_state & ONWORKLIST) == 0)
3837 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
3839 * Any new dependencies associated with the incore inode must
3840 * now be moved to the list associated with the buffer holding
3841 * the in-memory copy of the inode. Once merged process any
3842 * allocdirects that are completed by the merger.
3844 merge_inode_lists(inodedep);
3845 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
3846 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
3848 * Now that the inode has been pushed into the buffer, the
3849 * operations dependent on the inode being written to disk
3850 * can be moved to the id_bufwait so that they will be
3851 * processed when the buffer I/O completes.
3853 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
3854 WORKLIST_REMOVE(wk);
3855 WORKLIST_INSERT(&inodedep->id_bufwait, wk);
3858 * Newly allocated inodes cannot be written until the bitmap
3859 * that allocates them have been written (indicated by
3860 * DEPCOMPLETE being set in id_state). If we are doing a
3861 * forced sync (e.g., an fsync on a file), we force the bitmap
3862 * to be written so that the update can be done.
3864 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
3868 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
3871 (error = VOP_BWRITE(inodedep->id_buf->b_vp, inodedep->id_buf)) != 0)
3872 softdep_error("softdep_update_inodeblock: bwrite", error);
3873 if ((inodedep->id_state & DEPCOMPLETE) == 0)
3874 panic("softdep_update_inodeblock: update failed");
3878 * Merge the new inode dependency list (id_newinoupdt) into the old
3879 * inode dependency list (id_inoupdt). This routine must be called
3880 * with splbio interrupts blocked.
3883 merge_inode_lists(inodedep)
3884 struct inodedep *inodedep;
3886 struct allocdirect *listadp, *newadp;
3888 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3889 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) {
3890 if (listadp->ad_lbn < newadp->ad_lbn) {
3891 listadp = TAILQ_NEXT(listadp, ad_next);
3894 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3895 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
3896 if (listadp->ad_lbn == newadp->ad_lbn) {
3897 allocdirect_merge(&inodedep->id_inoupdt, newadp,
3901 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3903 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) {
3904 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3905 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next);
3910 * If we are doing an fsync, then we must ensure that any directory
3911 * entries for the inode have been written after the inode gets to disk.
3915 struct vnode *vp; /* the "in_core" copy of the inode */
3917 struct inodedep *inodedep;
3918 struct pagedep *pagedep;
3919 struct worklist *wk;
3926 struct thread *td = curthread; /* XXX */
3927 struct proc *p = td->td_proc;
3928 int error, flushparent;
3937 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
3941 if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
3942 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
3943 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
3944 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) {
3946 panic("softdep_fsync: pending ops");
3948 for (error = 0, flushparent = 0; ; ) {
3949 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
3951 if (wk->wk_type != D_DIRADD) {
3953 panic("softdep_fsync: Unexpected type %s",
3954 TYPENAME(wk->wk_type));
3956 dap = WK_DIRADD(wk);
3958 * Flush our parent if this directory entry
3959 * has a MKDIR_PARENT dependency.
3961 if (dap->da_state & DIRCHG)
3962 pagedep = dap->da_previous->dm_pagedep;
3964 pagedep = dap->da_pagedep;
3965 mnt = pagedep->pd_mnt;
3966 parentino = pagedep->pd_ino;
3967 lbn = pagedep->pd_lbn;
3968 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) {
3970 panic("softdep_fsync: dirty");
3972 flushparent = dap->da_state & MKDIR_PARENT;
3974 * If we are being fsync'ed as part of vgone'ing this vnode,
3975 * then we will not be able to release and recover the
3976 * vnode below, so we just have to give up on writing its
3977 * directory entry out. It will eventually be written, just
3978 * not now, but then the user was not asking to have it
3979 * written, so we are not breaking any promises.
3981 if (vp->v_flag & VXLOCK)
3984 * We prevent deadlock by always fetching inodes from the
3985 * root, moving down the directory tree. Thus, when fetching
3986 * our parent directory, we must unlock ourselves before
3987 * requesting the lock on our parent. See the comment in
3988 * ufs_lookup for details on possible races.
3991 VOP_UNLOCK(vp, 0, td);
3992 error = VFS_VGET(mnt, parentino, &pvp);
3993 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
3997 if ((error = UFS_UPDATE(pvp, 1)) != 0) {
4003 * Flush directory page containing the inode's name.
4005 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), p->p_ucred,
4008 error = VOP_BWRITE(bp->b_vp, bp);
4013 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
4021 * Flush all the dirty bitmaps associated with the block device
4022 * before flushing the rest of the dirty blocks so as to reduce
4023 * the number of dependencies that will have to be rolled back.
4026 softdep_fsync_mountdev(vp)
4029 struct buf *bp, *nbp;
4030 struct worklist *wk;
4032 if (!vn_isdisk(vp, NULL))
4033 panic("softdep_fsync_mountdev: vnode not a disk");
4035 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
4036 nbp = TAILQ_NEXT(bp, b_vnbufs);
4038 * If it is already scheduled, skip to the next buffer.
4040 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
4042 if ((bp->b_flags & B_DELWRI) == 0) {
4044 panic("softdep_fsync_mountdev: not dirty");
4047 * We are only interested in bitmaps with outstanding
4050 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
4051 wk->wk_type != D_BMSAFEMAP ||
4052 (bp->b_xflags & BX_BKGRDINPROG)) {
4061 * Since we may have slept during the I/O, we need
4062 * to start from a known point.
4064 nbp = TAILQ_FIRST(&vp->v_dirtyblkhd);
4066 drain_output(vp, 1);
4071 * This routine is called when we are trying to synchronously flush a
4072 * file. This routine must eliminate any filesystem metadata dependencies
4073 * so that the syncing routine can succeed by pushing the dirty blocks
4074 * associated with the file. If any I/O errors occur, they are returned.
4077 softdep_sync_metadata(ap)
4078 struct vop_fsync_args /* {
4080 struct ucred *a_cred;
4085 struct vnode *vp = ap->a_vp;
4086 struct pagedep *pagedep;
4087 struct allocdirect *adp;
4088 struct allocindir *aip;
4089 struct buf *bp, *nbp;
4090 struct worklist *wk;
4091 int i, error, waitfor;
4094 * Check whether this vnode is involved in a filesystem
4095 * that is doing soft dependency processing.
4097 if (!vn_isdisk(vp, NULL)) {
4098 if (!DOINGSOFTDEP(vp))
4101 if (vp->v_specmountpoint == NULL ||
4102 (vp->v_specmountpoint->mnt_flag & MNT_SOFTDEP) == 0)
4105 * Ensure that any direct block dependencies have been cleared.
4108 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
4113 * For most files, the only metadata dependencies are the
4114 * cylinder group maps that allocate their inode or blocks.
4115 * The block allocation dependencies can be found by traversing
4116 * the dependency lists for any buffers that remain on their
4117 * dirty buffer list. The inode allocation dependency will
4118 * be resolved when the inode is updated with MNT_WAIT.
4119 * This work is done in two passes. The first pass grabs most
4120 * of the buffers and begins asynchronously writing them. The
4121 * only way to wait for these asynchronous writes is to sleep
4122 * on the filesystem vnode which may stay busy for a long time
4123 * if the filesystem is active. So, instead, we make a second
4124 * pass over the dependencies blocking on each write. In the
4125 * usual case we will be blocking against a write that we
4126 * initiated, so when it is done the dependency will have been
4127 * resolved. Thus the second pass is expected to end quickly.
4129 waitfor = MNT_NOWAIT;
4132 * We must wait for any I/O in progress to finish so that
4133 * all potential buffers on the dirty list will be visible.
4135 drain_output(vp, 1);
4136 if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) {
4140 bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
4143 * As we hold the buffer locked, none of its dependencies
4146 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4147 switch (wk->wk_type) {
4150 adp = WK_ALLOCDIRECT(wk);
4151 if (adp->ad_state & DEPCOMPLETE)
4154 if (getdirtybuf(&nbp, waitfor) == 0)
4157 if (waitfor == MNT_NOWAIT) {
4159 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4167 aip = WK_ALLOCINDIR(wk);
4168 if (aip->ai_state & DEPCOMPLETE)
4171 if (getdirtybuf(&nbp, waitfor) == 0)
4174 if (waitfor == MNT_NOWAIT) {
4176 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4186 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
4187 if (aip->ai_state & DEPCOMPLETE)
4190 if (getdirtybuf(&nbp, MNT_WAIT) == 0)
4193 if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4203 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
4204 WK_INODEDEP(wk)->id_ino)) != 0) {
4213 * We are trying to sync a directory that may
4214 * have dependencies on both its own metadata
4215 * and/or dependencies on the inodes of any
4216 * recently allocated files. We walk its diradd
4217 * lists pushing out the associated inode.
4219 pagedep = WK_PAGEDEP(wk);
4220 for (i = 0; i < DAHASHSZ; i++) {
4221 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
4224 flush_pagedep_deps(vp, pagedep->pd_mnt,
4225 &pagedep->pd_diraddhd[i]))) {
4235 * This case should never happen if the vnode has
4236 * been properly sync'ed. However, if this function
4237 * is used at a place where the vnode has not yet
4238 * been sync'ed, this dependency can show up. So,
4239 * rather than panic, just flush it.
4241 nbp = WK_MKDIR(wk)->md_buf;
4242 if (getdirtybuf(&nbp, waitfor) == 0)
4245 if (waitfor == MNT_NOWAIT) {
4247 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4256 * This case should never happen if the vnode has
4257 * been properly sync'ed. However, if this function
4258 * is used at a place where the vnode has not yet
4259 * been sync'ed, this dependency can show up. So,
4260 * rather than panic, just flush it.
4262 nbp = WK_BMSAFEMAP(wk)->sm_buf;
4263 if (getdirtybuf(&nbp, waitfor) == 0)
4266 if (waitfor == MNT_NOWAIT) {
4268 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4277 panic("softdep_sync_metadata: Unknown type %s",
4278 TYPENAME(wk->wk_type));
4282 (void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT);
4283 nbp = TAILQ_NEXT(bp, b_vnbufs);
4292 * The brief unlock is to allow any pent up dependency
4293 * processing to be done. Then proceed with the second pass.
4295 if (waitfor == MNT_NOWAIT) {
4303 * If we have managed to get rid of all the dirty buffers,
4304 * then we are done. For certain directories and block
4305 * devices, we may need to do further work.
4307 if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) {
4314 * If we are trying to sync a block device, some of its buffers may
4315 * contain metadata that cannot be written until the contents of some
4316 * partially written files have been written to disk. The only easy
4317 * way to accomplish this is to sync the entire filesystem (luckily
4318 * this happens rarely).
4320 * We must wait for any I/O in progress to finish so that
4321 * all potential buffers on the dirty list will be visible.
4323 drain_output(vp, 1);
4324 if (vn_isdisk(vp, NULL) &&
4325 vp->v_specmountpoint && !VOP_ISLOCKED(vp, NULL) &&
4326 (error = VFS_SYNC(vp->v_specmountpoint, MNT_WAIT, ap->a_cred,
4333 * Flush the dependencies associated with an inodedep.
4334 * Called with splbio blocked.
4337 flush_inodedep_deps(fs, ino)
4341 struct inodedep *inodedep;
4342 struct allocdirect *adp;
4347 * This work is done in two passes. The first pass grabs most
4348 * of the buffers and begins asynchronously writing them. The
4349 * only way to wait for these asynchronous writes is to sleep
4350 * on the filesystem vnode which may stay busy for a long time
4351 * if the filesystem is active. So, instead, we make a second
4352 * pass over the dependencies blocking on each write. In the
4353 * usual case we will be blocking against a write that we
4354 * initiated, so when it is done the dependency will have been
4355 * resolved. Thus the second pass is expected to end quickly.
4356 * We give a brief window at the top of the loop to allow
4357 * any pending I/O to complete.
4359 for (waitfor = MNT_NOWAIT; ; ) {
4362 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4364 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) {
4365 if (adp->ad_state & DEPCOMPLETE)
4368 if (getdirtybuf(&bp, waitfor) == 0) {
4369 if (waitfor == MNT_NOWAIT)
4374 if (waitfor == MNT_NOWAIT) {
4376 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) {
4385 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) {
4386 if (adp->ad_state & DEPCOMPLETE)
4389 if (getdirtybuf(&bp, waitfor) == 0) {
4390 if (waitfor == MNT_NOWAIT)
4395 if (waitfor == MNT_NOWAIT) {
4397 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) {
4407 * If pass2, we are done, otherwise do pass 2.
4409 if (waitfor == MNT_WAIT)
4414 * Try freeing inodedep in case all dependencies have been removed.
4416 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
4417 (void) free_inodedep(inodedep);
4422 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
4423 * Called with splbio blocked.
4426 flush_pagedep_deps(pvp, mp, diraddhdp)
4429 struct diraddhd *diraddhdp;
4431 struct thread *td = curthread; /* XXX */
4433 struct inodedep *inodedep;
4434 struct ufsmount *ump;
4437 int gotit, error = 0;
4441 KKASSERT(td->td_proc);
4442 cr = td->td_proc->p_ucred;
4445 while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
4447 * Flush ourselves if this directory entry
4448 * has a MKDIR_PARENT dependency.
4450 if (dap->da_state & MKDIR_PARENT) {
4452 if ((error = UFS_UPDATE(pvp, 1)) != 0)
4456 * If that cleared dependencies, go on to next.
4458 if (dap != LIST_FIRST(diraddhdp))
4460 if (dap->da_state & MKDIR_PARENT) {
4462 panic("flush_pagedep_deps: MKDIR_PARENT");
4466 * A newly allocated directory must have its "." and
4467 * ".." entries written out before its name can be
4468 * committed in its parent. We do not want or need
4469 * the full semantics of a synchronous VOP_FSYNC as
4470 * that may end up here again, once for each directory
4471 * level in the filesystem. Instead, we push the blocks
4472 * and wait for them to clear. We have to fsync twice
4473 * because the first call may choose to defer blocks
4474 * that still have dependencies, but deferral will
4475 * happen at most once.
4477 inum = dap->da_newinum;
4478 if (dap->da_state & MKDIR_BODY) {
4480 if ((error = VFS_VGET(mp, inum, &vp)) != 0)
4482 if ((error=VOP_FSYNC(vp, cr, MNT_NOWAIT, td)) ||
4483 (error=VOP_FSYNC(vp, cr, MNT_NOWAIT, td))) {
4487 drain_output(vp, 0);
4491 * If that cleared dependencies, go on to next.
4493 if (dap != LIST_FIRST(diraddhdp))
4495 if (dap->da_state & MKDIR_BODY) {
4497 panic("flush_pagedep_deps: MKDIR_BODY");
4501 * Flush the inode on which the directory entry depends.
4502 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
4503 * the only remaining dependency is that the updated inode
4504 * count must get pushed to disk. The inode has already
4505 * been pushed into its inode buffer (via VOP_UPDATE) at
4506 * the time of the reference count change. So we need only
4507 * locate that buffer, ensure that there will be no rollback
4508 * caused by a bitmap dependency, then write the inode buffer.
4510 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) {
4512 panic("flush_pagedep_deps: lost inode");
4515 * If the inode still has bitmap dependencies,
4516 * push them to disk.
4518 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4519 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4522 (error = VOP_BWRITE(inodedep->id_buf->b_vp,
4523 inodedep->id_buf)) != 0)
4526 if (dap != LIST_FIRST(diraddhdp))
4530 * If the inode is still sitting in a buffer waiting
4531 * to be written, push it to disk.
4534 if ((error = bread(ump->um_devvp,
4535 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
4536 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0)
4538 if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0)
4542 * If we have failed to get rid of all the dependencies
4543 * then something is seriously wrong.
4545 if (dap == LIST_FIRST(diraddhdp)) {
4547 panic("flush_pagedep_deps: flush failed");
4556 * A large burst of file addition or deletion activity can drive the
4557 * memory load excessively high. First attempt to slow things down
4558 * using the techniques below. If that fails, this routine requests
4559 * the offending operations to fall back to running synchronously
4560 * until the memory load returns to a reasonable level.
4563 softdep_slowdown(vp)
4566 int max_softdeps_hard;
4568 max_softdeps_hard = max_softdeps * 11 / 10;
4569 if (num_dirrem < max_softdeps_hard / 2 &&
4570 num_inodedep < max_softdeps_hard)
4572 stat_sync_limit_hit += 1;
4577 * If memory utilization has gotten too high, deliberately slow things
4578 * down and speed up the I/O processing.
4581 request_cleanup(resource, islocked)
4585 struct thread *td = curthread; /* XXX */
4588 * We never hold up the filesystem syncer process.
4590 if (td == filesys_syncer)
4593 * First check to see if the work list has gotten backlogged.
4594 * If it has, co-opt this process to help clean up two entries.
4595 * Because this process may hold inodes locked, we cannot
4596 * handle any remove requests that might block on a locked
4597 * inode as that could lead to deadlock.
4599 if (num_on_worklist > max_softdeps / 10) {
4602 process_worklist_item(NULL, LK_NOWAIT);
4603 process_worklist_item(NULL, LK_NOWAIT);
4604 stat_worklist_push += 2;
4611 * If we are resource constrained on inode dependencies, try
4612 * flushing some dirty inodes. Otherwise, we are constrained
4613 * by file deletions, so try accelerating flushes of directories
4614 * with removal dependencies. We would like to do the cleanup
4615 * here, but we probably hold an inode locked at this point and
4616 * that might deadlock against one that we try to clean. So,
4617 * the best that we can do is request the syncer daemon to do
4618 * the cleanup for us.
4623 stat_ino_limit_push += 1;
4624 req_clear_inodedeps += 1;
4625 stat_countp = &stat_ino_limit_hit;
4629 stat_blk_limit_push += 1;
4630 req_clear_remove += 1;
4631 stat_countp = &stat_blk_limit_hit;
4637 panic("request_cleanup: unknown type");
4640 * Hopefully the syncer daemon will catch up and awaken us.
4641 * We wait at most tickdelay before proceeding in any case.
4646 if (handle.callout == NULL)
4647 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
4648 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, PPAUSE,
4657 * Awaken processes pausing in request_cleanup and clear proc_waiting
4658 * to indicate that there is no longer a timer running.
4666 wakeup_one(&proc_waiting);
4667 if (proc_waiting > 0)
4668 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
4670 handle.callout = NULL;
4674 * Flush out a directory with at least one removal dependency in an effort to
4675 * reduce the number of dirrem, freefile, and freeblks dependency structures.
4678 clear_remove(struct thread *td)
4680 struct pagedep_hashhead *pagedephd;
4681 struct pagedep *pagedep;
4682 static int next = 0;
4689 KKASSERT(td->td_proc);
4690 cred = td->td_proc->p_ucred;
4693 for (cnt = 0; cnt < pagedep_hash; cnt++) {
4694 pagedephd = &pagedep_hashtbl[next++];
4695 if (next >= pagedep_hash)
4697 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
4698 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
4700 mp = pagedep->pd_mnt;
4701 ino = pagedep->pd_ino;
4703 if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4704 softdep_error("clear_remove: vget", error);
4707 if ((error = VOP_FSYNC(vp, cred, MNT_NOWAIT, td)))
4708 softdep_error("clear_remove: fsync", error);
4709 drain_output(vp, 0);
4718 * Clear out a block of dirty inodes in an effort to reduce
4719 * the number of inodedep dependency structures.
4722 clear_inodedeps(struct thread *td)
4725 struct inodedep_hashhead *inodedephd;
4726 struct inodedep *inodedep;
4727 static int next = 0;
4732 ino_t firstino, lastino, ino;
4734 KKASSERT(td->td_proc);
4735 cred = td->td_proc->p_ucred;
4739 * Pick a random inode dependency to be cleared.
4740 * We will then gather up all the inodes in its block
4741 * that have dependencies and flush them out.
4743 for (cnt = 0; cnt < inodedep_hash; cnt++) {
4744 inodedephd = &inodedep_hashtbl[next++];
4745 if (next >= inodedep_hash)
4747 if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
4750 if (inodedep == NULL)
4753 * Ugly code to find mount point given pointer to superblock.
4755 fs = inodedep->id_fs;
4756 TAILQ_FOREACH(mp, &mountlist, mnt_list)
4757 if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs)
4760 * Find the last inode in the block with dependencies.
4762 firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
4763 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
4764 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
4767 * Asynchronously push all but the last inode with dependencies.
4768 * Synchronously push the last inode with dependencies to ensure
4769 * that the inode block gets written to free up the inodedeps.
4771 for (ino = firstino; ino <= lastino; ino++) {
4772 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4775 if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4776 softdep_error("clear_inodedeps: vget", error);
4779 if (ino == lastino) {
4780 if ((error = VOP_FSYNC(vp, cred, MNT_WAIT, td)))
4781 softdep_error("clear_inodedeps: fsync1", error);
4783 if ((error = VOP_FSYNC(vp, cred, MNT_NOWAIT, td)))
4784 softdep_error("clear_inodedeps: fsync2", error);
4785 drain_output(vp, 0);
4794 * Function to determine if the buffer has outstanding dependencies
4795 * that will cause a roll-back if the buffer is written. If wantcount
4796 * is set, return number of dependencies, otherwise just yes or no.
4799 softdep_count_dependencies(bp, wantcount)
4803 struct worklist *wk;
4804 struct inodedep *inodedep;
4805 struct indirdep *indirdep;
4806 struct allocindir *aip;
4807 struct pagedep *pagedep;
4813 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4814 switch (wk->wk_type) {
4817 inodedep = WK_INODEDEP(wk);
4818 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4819 /* bitmap allocation dependency */
4824 if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
4825 /* direct block pointer dependency */
4833 indirdep = WK_INDIRDEP(wk);
4835 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
4836 /* indirect block pointer dependency */
4844 pagedep = WK_PAGEDEP(wk);
4845 for (i = 0; i < DAHASHSZ; i++) {
4847 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
4848 /* directory entry dependency */
4860 /* never a dependency on these blocks */
4865 panic("softdep_check_for_rollback: Unexpected type %s",
4866 TYPENAME(wk->wk_type));
4876 * Acquire exclusive access to a buffer.
4877 * Must be called with splbio blocked.
4878 * Return 1 if buffer was acquired.
4881 getdirtybuf(bpp, waitfor)
4889 if ((bp = *bpp) == NULL)
4891 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
4892 if ((bp->b_xflags & BX_BKGRDINPROG) == 0)
4895 if (waitfor != MNT_WAIT)
4897 bp->b_xflags |= BX_BKGRDWAIT;
4898 interlocked_sleep(&lk, SLEEP, &bp->b_xflags, PRIBIO,
4902 if (waitfor != MNT_WAIT)
4904 error = interlocked_sleep(&lk, LOCKBUF, bp,
4905 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0);
4906 if (error != ENOLCK) {
4908 panic("getdirtybuf: inconsistent lock");
4911 if ((bp->b_flags & B_DELWRI) == 0) {
4920 * Wait for pending output on a vnode to complete.
4921 * Must be called with vnode locked.
4924 drain_output(vp, islocked)
4931 while (vp->v_numoutput) {
4932 vp->v_flag |= VBWAIT;
4933 interlocked_sleep(&lk, SLEEP, (caddr_t)&vp->v_numoutput,
4934 PRIBIO + 1, "drainvp", 0);
4941 * Called whenever a buffer that is being invalidated or reallocated
4942 * contains dependencies. This should only happen if an I/O error has
4943 * occurred. The routine is called with the buffer locked.
4946 softdep_deallocate_dependencies(bp)
4950 if ((bp->b_flags & B_ERROR) == 0)
4951 panic("softdep_deallocate_dependencies: dangling deps");
4952 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
4953 panic("softdep_deallocate_dependencies: unrecovered I/O error");
4957 * Function to handle asynchronous write errors in the filesystem.
4960 softdep_error(func, error)
4965 /* XXX should do something better! */
4966 printf("%s: got error %d while accessing filesystem\n", func, error);