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 1.2 2003/06/17 04:28:59 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>
63 #include <ufs/ufs/dir.h>
64 #include <ufs/ufs/quota.h>
65 #include <ufs/ufs/inode.h>
66 #include <ufs/ufs/ufsmount.h>
67 #include <ufs/ffs/fs.h>
68 #include <ufs/ffs/softdep.h>
69 #include <ufs/ffs/ffs_extern.h>
70 #include <ufs/ufs/ufs_extern.h>
73 * These definitions need to be adapted to the system to which
74 * this file is being ported.
77 * malloc types defined for the softdep system.
79 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
80 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
81 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
82 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
83 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
84 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
85 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
86 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
87 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
88 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
89 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
90 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
91 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
93 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE)
99 #define D_ALLOCDIRECT 4
101 #define D_ALLOCINDIR 6
108 #define D_LAST D_DIRREM
111 * translate from workitem type to memory type
112 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
114 static struct malloc_type *memtype[] = {
130 #define DtoM(type) (memtype[type])
133 * Names of malloc types.
135 #define TYPENAME(type) \
136 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
137 #define CURPROC curproc
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 proc *));
149 static void clear_inodedeps __P((struct proc *));
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 static struct lockit {
245 static void acquire_lock __P((struct lockit *));
246 static void free_lock __P((struct lockit *));
247 void softdep_panic __P((char *));
249 #define ACQUIRE_LOCK(lk) acquire_lock(lk)
250 #define FREE_LOCK(lk) free_lock(lk)
258 if (lk->lkt_held != -1) {
259 holder = lk->lkt_held;
261 if (holder == CURPROC->p_pid)
262 panic("softdep_lock: locking against myself");
264 panic("softdep_lock: lock held by %d", holder);
266 lk->lkt_spl = splbio();
267 lk->lkt_held = CURPROC->p_pid;
276 if (lk->lkt_held == -1)
277 panic("softdep_unlock: lock not held");
283 * Function to release soft updates lock and panic.
290 if (lk.lkt_held != -1)
296 static int interlocked_sleep __P((struct lockit *, int, void *, int,
300 * When going to sleep, we must save our SPL so that it does
301 * not get lost if some other process uses the lock while we
302 * are sleeping. We restore it after we have slept. This routine
303 * wraps the interlocking with functions that sleep. The list
304 * below enumerates the available set of operations.
311 interlocked_sleep(lk, op, ident, flags, wmesg, timo)
324 if (lk->lkt_held == -1)
325 panic("interlocked_sleep: lock not held");
330 retval = tsleep(ident, flags, wmesg, timo);
333 retval = BUF_LOCK((struct buf *)ident, flags);
336 panic("interlocked_sleep: unknown operation");
339 if (lk->lkt_held != -1) {
340 holder = lk->lkt_held;
342 if (holder == CURPROC->p_pid)
343 panic("interlocked_sleep: locking against self");
345 panic("interlocked_sleep: lock held by %d", holder);
347 lk->lkt_held = CURPROC->p_pid;
355 * Place holder for real semaphores.
364 static void sema_init __P((struct sema *, char *, int, int));
365 static int sema_get __P((struct sema *, struct lockit *));
366 static void sema_release __P((struct sema *));
369 sema_init(semap, name, prio, timo)
383 sema_get(semap, interlock)
385 struct lockit *interlock;
388 if (semap->value++ > 0) {
389 if (interlock != NULL) {
390 interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
391 semap->prio, semap->name, semap->timo);
392 FREE_LOCK(interlock);
394 tsleep((caddr_t)semap, semap->prio, semap->name,
399 semap->holder = CURPROC->p_pid;
400 if (interlock != NULL)
401 FREE_LOCK(interlock);
410 if (semap->value <= 0 || semap->holder != CURPROC->p_pid) {
411 if (lk.lkt_held != -1)
413 panic("sema_release: not held");
415 if (--semap->value > 0) {
423 * Worklist queue management.
424 * These routines require that the lock be held.
426 #ifndef /* NOT */ DEBUG
427 #define WORKLIST_INSERT(head, item) do { \
428 (item)->wk_state |= ONWORKLIST; \
429 LIST_INSERT_HEAD(head, item, wk_list); \
431 #define WORKLIST_REMOVE(item) do { \
432 (item)->wk_state &= ~ONWORKLIST; \
433 LIST_REMOVE(item, wk_list); \
435 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
438 static void worklist_insert __P((struct workhead *, struct worklist *));
439 static void worklist_remove __P((struct worklist *));
440 static void workitem_free __P((struct worklist *, int));
442 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
443 #define WORKLIST_REMOVE(item) worklist_remove(item)
444 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
447 worklist_insert(head, item)
448 struct workhead *head;
449 struct worklist *item;
452 if (lk.lkt_held == -1)
453 panic("worklist_insert: lock not held");
454 if (item->wk_state & ONWORKLIST) {
456 panic("worklist_insert: already on list");
458 item->wk_state |= ONWORKLIST;
459 LIST_INSERT_HEAD(head, item, wk_list);
463 worklist_remove(item)
464 struct worklist *item;
467 if (lk.lkt_held == -1)
468 panic("worklist_remove: lock not held");
469 if ((item->wk_state & ONWORKLIST) == 0) {
471 panic("worklist_remove: not on list");
473 item->wk_state &= ~ONWORKLIST;
474 LIST_REMOVE(item, wk_list);
478 workitem_free(item, type)
479 struct worklist *item;
483 if (item->wk_state & ONWORKLIST) {
484 if (lk.lkt_held != -1)
486 panic("workitem_free: still on list");
488 if (item->wk_type != type) {
489 if (lk.lkt_held != -1)
491 panic("workitem_free: type mismatch");
493 FREE(item, DtoM(type));
498 * Workitem queue management
500 static struct workhead softdep_workitem_pending;
501 static int num_on_worklist; /* number of worklist items to be processed */
502 static int softdep_worklist_busy; /* 1 => trying to do unmount */
503 static int softdep_worklist_req; /* serialized waiters */
504 static int max_softdeps; /* maximum number of structs before slowdown */
505 static int tickdelay = 2; /* number of ticks to pause during slowdown */
506 static int *stat_countp; /* statistic to count in proc_waiting timeout */
507 static int proc_waiting; /* tracks whether we have a timeout posted */
508 static struct callout_handle handle; /* handle on posted proc_waiting timeout */
509 static struct proc *filesys_syncer; /* proc of filesystem syncer process */
510 static int req_clear_inodedeps; /* syncer process flush some inodedeps */
511 #define FLUSH_INODES 1
512 static int req_clear_remove; /* syncer process flush some freeblks */
513 #define FLUSH_REMOVE 2
517 static int stat_worklist_push; /* number of worklist cleanups */
518 static int stat_blk_limit_push; /* number of times block limit neared */
519 static int stat_ino_limit_push; /* number of times inode limit neared */
520 static int stat_blk_limit_hit; /* number of times block slowdown imposed */
521 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */
522 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */
523 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */
524 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */
525 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
526 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */
529 #include <sys/sysctl.h>
530 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
531 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
532 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
533 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
534 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
535 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
536 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
537 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
538 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
539 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
540 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
541 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
545 * Add an item to the end of the work queue.
546 * This routine requires that the lock be held.
547 * This is the only routine that adds items to the list.
548 * The following routine is the only one that removes items
549 * and does so in order from first to last.
555 static struct worklist *worklist_tail;
557 if (wk->wk_state & ONWORKLIST) {
558 if (lk.lkt_held != -1)
560 panic("add_to_worklist: already on list");
562 wk->wk_state |= ONWORKLIST;
563 if (LIST_FIRST(&softdep_workitem_pending) == NULL)
564 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
566 LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
568 num_on_worklist += 1;
572 * Process that runs once per second to handle items in the background queue.
574 * Note that we ensure that everything is done in the order in which they
575 * appear in the queue. The code below depends on this property to ensure
576 * that blocks of a file are freed before the inode itself is freed. This
577 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
578 * until all the old ones have been purged from the dependency lists.
581 softdep_process_worklist(matchmnt)
582 struct mount *matchmnt;
584 struct proc *p = CURPROC;
585 int matchcnt, loopcount;
589 * Record the process identifier of our caller so that we can give
590 * this process preferential treatment in request_cleanup below.
596 * There is no danger of having multiple processes run this
597 * code, but we have to single-thread it when softdep_flushfiles()
598 * is in operation to get an accurate count of the number of items
599 * related to its mount point that are in the list.
601 if (matchmnt == NULL) {
602 if (softdep_worklist_busy < 0)
604 softdep_worklist_busy += 1;
608 * If requested, try removing inode or removal dependencies.
610 if (req_clear_inodedeps) {
612 req_clear_inodedeps -= 1;
613 wakeup_one(&proc_waiting);
615 if (req_clear_remove) {
617 req_clear_remove -= 1;
618 wakeup_one(&proc_waiting);
621 starttime = time_second;
622 while (num_on_worklist > 0) {
623 matchcnt += process_worklist_item(matchmnt, 0);
626 * If a umount operation wants to run the worklist
629 if (softdep_worklist_req && matchmnt == NULL) {
635 * If requested, try removing inode or removal dependencies.
637 if (req_clear_inodedeps) {
639 req_clear_inodedeps -= 1;
640 wakeup_one(&proc_waiting);
642 if (req_clear_remove) {
644 req_clear_remove -= 1;
645 wakeup_one(&proc_waiting);
648 * We do not generally want to stop for buffer space, but if
649 * we are really being a buffer hog, we will stop and wait.
651 if (loopcount++ % 128 == 0)
654 * Never allow processing to run for more than one
655 * second. Otherwise the other syncer tasks may get
656 * excessively backlogged.
658 if (starttime != time_second && matchmnt == NULL) {
663 if (matchmnt == NULL) {
664 --softdep_worklist_busy;
665 if (softdep_worklist_req && softdep_worklist_busy == 0)
666 wakeup(&softdep_worklist_req);
672 * Process one item on the worklist.
675 process_worklist_item(matchmnt, flags)
676 struct mount *matchmnt;
680 struct dirrem *dirrem;
686 if (matchmnt != NULL)
687 matchfs = VFSTOUFS(matchmnt)->um_fs;
690 * Normally we just process each item on the worklist in order.
691 * However, if we are in a situation where we cannot lock any
692 * inodes, we have to skip over any dirrem requests whose
693 * vnodes are resident and locked.
695 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
696 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
698 dirrem = WK_DIRREM(wk);
699 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev,
701 if (vp == NULL || !VOP_ISLOCKED(vp, CURPROC))
709 num_on_worklist -= 1;
711 switch (wk->wk_type) {
714 /* removal of a directory entry */
715 if (WK_DIRREM(wk)->dm_mnt == matchmnt)
717 handle_workitem_remove(WK_DIRREM(wk));
721 /* releasing blocks and/or fragments from a file */
722 if (WK_FREEBLKS(wk)->fb_fs == matchfs)
724 handle_workitem_freeblocks(WK_FREEBLKS(wk));
728 /* releasing a fragment when replaced as a file grows */
729 if (WK_FREEFRAG(wk)->ff_fs == matchfs)
731 handle_workitem_freefrag(WK_FREEFRAG(wk));
735 /* releasing an inode when its link count drops to 0 */
736 if (WK_FREEFILE(wk)->fx_fs == matchfs)
738 handle_workitem_freefile(WK_FREEFILE(wk));
742 panic("%s_process_worklist: Unknown type %s",
743 "softdep", TYPENAME(wk->wk_type));
750 * Move dependencies from one buffer to another.
753 softdep_move_dependencies(oldbp, newbp)
757 struct worklist *wk, *wktail;
759 if (LIST_FIRST(&newbp->b_dep) != NULL)
760 panic("softdep_move_dependencies: need merge code");
763 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
764 LIST_REMOVE(wk, wk_list);
766 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
768 LIST_INSERT_AFTER(wktail, wk, wk_list);
775 * Purge the work list of all items associated with a particular mount point.
778 softdep_flushfiles(oldmnt, flags, p)
779 struct mount *oldmnt;
787 * Await our turn to clear out the queue, then serialize access.
789 while (softdep_worklist_busy != 0) {
790 softdep_worklist_req += 1;
791 tsleep(&softdep_worklist_req, PRIBIO, "softflush", 0);
792 softdep_worklist_req -= 1;
794 softdep_worklist_busy = -1;
796 if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0) {
797 softdep_worklist_busy = 0;
798 if (softdep_worklist_req)
799 wakeup(&softdep_worklist_req);
803 * Alternately flush the block device associated with the mount
804 * point and process any dependencies that the flushing
805 * creates. In theory, this loop can happen at most twice,
806 * but we give it a few extra just to be sure.
808 devvp = VFSTOUFS(oldmnt)->um_devvp;
809 for (loopcnt = 10; loopcnt > 0; ) {
810 if (softdep_process_worklist(oldmnt) == 0) {
813 * Do another flush in case any vnodes were brought in
814 * as part of the cleanup operations.
816 if ((error = ffs_flushfiles(oldmnt, flags, p)) != 0)
819 * If we still found nothing to do, we are really done.
821 if (softdep_process_worklist(oldmnt) == 0)
824 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, p);
825 error = VOP_FSYNC(devvp, p->p_ucred, MNT_WAIT, p);
826 VOP_UNLOCK(devvp, 0, p);
830 softdep_worklist_busy = 0;
831 if (softdep_worklist_req)
832 wakeup(&softdep_worklist_req);
835 * If we are unmounting then it is an error to fail. If we
836 * are simply trying to downgrade to read-only, then filesystem
837 * activity can keep us busy forever, so we just fail with EBUSY.
840 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
841 panic("softdep_flushfiles: looping");
850 * There are three types of structures that can be looked up:
851 * 1) pagedep structures identified by mount point, inode number,
853 * 2) inodedep structures identified by mount point and inode number.
854 * 3) newblk structures identified by mount point and
855 * physical block number.
857 * The "pagedep" and "inodedep" dependency structures are hashed
858 * separately from the file blocks and inodes to which they correspond.
859 * This separation helps when the in-memory copy of an inode or
860 * file block must be replaced. It also obviates the need to access
861 * an inode or file page when simply updating (or de-allocating)
862 * dependency structures. Lookup of newblk structures is needed to
863 * find newly allocated blocks when trying to associate them with
864 * their allocdirect or allocindir structure.
866 * The lookup routines optionally create and hash a new instance when
867 * an existing entry is not found.
869 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */
870 #define NODELAY 0x0002 /* cannot do background work */
873 * Structures and routines associated with pagedep caching.
875 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
876 u_long pagedep_hash; /* size of hash table - 1 */
877 #define PAGEDEP_HASH(mp, inum, lbn) \
878 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
880 static struct sema pagedep_in_progress;
883 * Look up a pagedep. Return 1 if found, 0 if not found.
884 * If not found, allocate if DEPALLOC flag is passed.
885 * Found or allocated entry is returned in pagedeppp.
886 * This routine must be called with splbio interrupts blocked.
889 pagedep_lookup(ip, lbn, flags, pagedeppp)
893 struct pagedep **pagedeppp;
895 struct pagedep *pagedep;
896 struct pagedep_hashhead *pagedephd;
901 if (lk.lkt_held == -1)
902 panic("pagedep_lookup: lock not held");
904 mp = ITOV(ip)->v_mount;
905 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
907 LIST_FOREACH(pagedep, pagedephd, pd_hash)
908 if (ip->i_number == pagedep->pd_ino &&
909 lbn == pagedep->pd_lbn &&
910 mp == pagedep->pd_mnt)
913 *pagedeppp = pagedep;
916 if ((flags & DEPALLOC) == 0) {
920 if (sema_get(&pagedep_in_progress, &lk) == 0) {
924 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
926 bzero(pagedep, sizeof(struct pagedep));
927 pagedep->pd_list.wk_type = D_PAGEDEP;
928 pagedep->pd_mnt = mp;
929 pagedep->pd_ino = ip->i_number;
930 pagedep->pd_lbn = lbn;
931 LIST_INIT(&pagedep->pd_dirremhd);
932 LIST_INIT(&pagedep->pd_pendinghd);
933 for (i = 0; i < DAHASHSZ; i++)
934 LIST_INIT(&pagedep->pd_diraddhd[i]);
936 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
937 sema_release(&pagedep_in_progress);
938 *pagedeppp = pagedep;
943 * Structures and routines associated with inodedep caching.
945 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
946 static u_long inodedep_hash; /* size of hash table - 1 */
947 static long num_inodedep; /* number of inodedep allocated */
948 #define INODEDEP_HASH(fs, inum) \
949 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
950 static struct sema inodedep_in_progress;
953 * Look up a inodedep. Return 1 if found, 0 if not found.
954 * If not found, allocate if DEPALLOC flag is passed.
955 * Found or allocated entry is returned in inodedeppp.
956 * This routine must be called with splbio interrupts blocked.
959 inodedep_lookup(fs, inum, flags, inodedeppp)
963 struct inodedep **inodedeppp;
965 struct inodedep *inodedep;
966 struct inodedep_hashhead *inodedephd;
970 if (lk.lkt_held == -1)
971 panic("inodedep_lookup: lock not held");
974 inodedephd = INODEDEP_HASH(fs, inum);
976 LIST_FOREACH(inodedep, inodedephd, id_hash)
977 if (inum == inodedep->id_ino && fs == inodedep->id_fs)
980 *inodedeppp = inodedep;
983 if ((flags & DEPALLOC) == 0) {
988 * If we are over our limit, try to improve the situation.
990 if (num_inodedep > max_softdeps && firsttry &&
991 speedup_syncer() == 0 && (flags & NODELAY) == 0 &&
992 request_cleanup(FLUSH_INODES, 1)) {
996 if (sema_get(&inodedep_in_progress, &lk) == 0) {
1001 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1002 M_INODEDEP, M_SOFTDEP_FLAGS);
1003 inodedep->id_list.wk_type = D_INODEDEP;
1004 inodedep->id_fs = fs;
1005 inodedep->id_ino = inum;
1006 inodedep->id_state = ALLCOMPLETE;
1007 inodedep->id_nlinkdelta = 0;
1008 inodedep->id_savedino = NULL;
1009 inodedep->id_savedsize = -1;
1010 inodedep->id_buf = NULL;
1011 LIST_INIT(&inodedep->id_pendinghd);
1012 LIST_INIT(&inodedep->id_inowait);
1013 LIST_INIT(&inodedep->id_bufwait);
1014 TAILQ_INIT(&inodedep->id_inoupdt);
1015 TAILQ_INIT(&inodedep->id_newinoupdt);
1017 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1018 sema_release(&inodedep_in_progress);
1019 *inodedeppp = inodedep;
1024 * Structures and routines associated with newblk caching.
1026 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1027 u_long newblk_hash; /* size of hash table - 1 */
1028 #define NEWBLK_HASH(fs, inum) \
1029 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1030 static struct sema newblk_in_progress;
1033 * Look up a newblk. Return 1 if found, 0 if not found.
1034 * If not found, allocate if DEPALLOC flag is passed.
1035 * Found or allocated entry is returned in newblkpp.
1038 newblk_lookup(fs, newblkno, flags, newblkpp)
1040 ufs_daddr_t newblkno;
1042 struct newblk **newblkpp;
1044 struct newblk *newblk;
1045 struct newblk_hashhead *newblkhd;
1047 newblkhd = NEWBLK_HASH(fs, newblkno);
1049 LIST_FOREACH(newblk, newblkhd, nb_hash)
1050 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1056 if ((flags & DEPALLOC) == 0) {
1060 if (sema_get(&newblk_in_progress, 0) == 0)
1062 MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1063 M_NEWBLK, M_SOFTDEP_FLAGS);
1064 newblk->nb_state = 0;
1066 newblk->nb_newblkno = newblkno;
1067 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1068 sema_release(&newblk_in_progress);
1074 * Executed during filesystem system initialization before
1075 * mounting any file systems.
1078 softdep_initialize()
1081 LIST_INIT(&mkdirlisthd);
1082 LIST_INIT(&softdep_workitem_pending);
1083 max_softdeps = min(desiredvnodes * 8,
1084 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep)));
1085 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1087 sema_init(&pagedep_in_progress, "pagedep", PRIBIO, 0);
1088 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1089 sema_init(&inodedep_in_progress, "inodedep", PRIBIO, 0);
1090 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1091 sema_init(&newblk_in_progress, "newblk", PRIBIO, 0);
1095 * Called at mount time to notify the dependency code that a
1096 * filesystem wishes to use it.
1099 softdep_mount(devvp, mp, fs, cred)
1100 struct vnode *devvp;
1105 struct csum cstotal;
1110 mp->mnt_flag &= ~MNT_ASYNC;
1111 mp->mnt_flag |= MNT_SOFTDEP;
1113 * When doing soft updates, the counters in the
1114 * superblock may have gotten out of sync, so we have
1115 * to scan the cylinder groups and recalculate them.
1117 if (fs->fs_clean != 0)
1119 bzero(&cstotal, sizeof cstotal);
1120 for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1121 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1122 fs->fs_cgsize, cred, &bp)) != 0) {
1126 cgp = (struct cg *)bp->b_data;
1127 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1128 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1129 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1130 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1131 fs->fs_cs(fs, cyl) = cgp->cg_cs;
1135 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1136 printf("ffs_mountfs: superblock updated for soft updates\n");
1138 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1143 * Protecting the freemaps (or bitmaps).
1145 * To eliminate the need to execute fsck before mounting a file system
1146 * after a power failure, one must (conservatively) guarantee that the
1147 * on-disk copy of the bitmaps never indicate that a live inode or block is
1148 * free. So, when a block or inode is allocated, the bitmap should be
1149 * updated (on disk) before any new pointers. When a block or inode is
1150 * freed, the bitmap should not be updated until all pointers have been
1151 * reset. The latter dependency is handled by the delayed de-allocation
1152 * approach described below for block and inode de-allocation. The former
1153 * dependency is handled by calling the following procedure when a block or
1154 * inode is allocated. When an inode is allocated an "inodedep" is created
1155 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1156 * Each "inodedep" is also inserted into the hash indexing structure so
1157 * that any additional link additions can be made dependent on the inode
1160 * The ufs file system maintains a number of free block counts (e.g., per
1161 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1162 * in addition to the bitmaps. These counts are used to improve efficiency
1163 * during allocation and therefore must be consistent with the bitmaps.
1164 * There is no convenient way to guarantee post-crash consistency of these
1165 * counts with simple update ordering, for two main reasons: (1) The counts
1166 * and bitmaps for a single cylinder group block are not in the same disk
1167 * sector. If a disk write is interrupted (e.g., by power failure), one may
1168 * be written and the other not. (2) Some of the counts are located in the
1169 * superblock rather than the cylinder group block. So, we focus our soft
1170 * updates implementation on protecting the bitmaps. When mounting a
1171 * filesystem, we recompute the auxiliary counts from the bitmaps.
1175 * Called just after updating the cylinder group block to allocate an inode.
1178 softdep_setup_inomapdep(bp, ip, newinum)
1179 struct buf *bp; /* buffer for cylgroup block with inode map */
1180 struct inode *ip; /* inode related to allocation */
1181 ino_t newinum; /* new inode number being allocated */
1183 struct inodedep *inodedep;
1184 struct bmsafemap *bmsafemap;
1187 * Create a dependency for the newly allocated inode.
1188 * Panic if it already exists as something is seriously wrong.
1189 * Otherwise add it to the dependency list for the buffer holding
1190 * the cylinder group map from which it was allocated.
1193 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1195 panic("softdep_setup_inomapdep: found inode");
1197 inodedep->id_buf = bp;
1198 inodedep->id_state &= ~DEPCOMPLETE;
1199 bmsafemap = bmsafemap_lookup(bp);
1200 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1205 * Called just after updating the cylinder group block to
1206 * allocate block or fragment.
1209 softdep_setup_blkmapdep(bp, fs, newblkno)
1210 struct buf *bp; /* buffer for cylgroup block with block map */
1211 struct fs *fs; /* filesystem doing allocation */
1212 ufs_daddr_t newblkno; /* number of newly allocated block */
1214 struct newblk *newblk;
1215 struct bmsafemap *bmsafemap;
1218 * Create a dependency for the newly allocated block.
1219 * Add it to the dependency list for the buffer holding
1220 * the cylinder group map from which it was allocated.
1222 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1223 panic("softdep_setup_blkmapdep: found block");
1225 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1226 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1231 * Find the bmsafemap associated with a cylinder group buffer.
1232 * If none exists, create one. The buffer must be locked when
1233 * this routine is called and this routine must be called with
1234 * splbio interrupts blocked.
1236 static struct bmsafemap *
1237 bmsafemap_lookup(bp)
1240 struct bmsafemap *bmsafemap;
1241 struct worklist *wk;
1244 if (lk.lkt_held == -1)
1245 panic("bmsafemap_lookup: lock not held");
1247 LIST_FOREACH(wk, &bp->b_dep, wk_list)
1248 if (wk->wk_type == D_BMSAFEMAP)
1249 return (WK_BMSAFEMAP(wk));
1251 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1252 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1253 bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1254 bmsafemap->sm_list.wk_state = 0;
1255 bmsafemap->sm_buf = bp;
1256 LIST_INIT(&bmsafemap->sm_allocdirecthd);
1257 LIST_INIT(&bmsafemap->sm_allocindirhd);
1258 LIST_INIT(&bmsafemap->sm_inodedephd);
1259 LIST_INIT(&bmsafemap->sm_newblkhd);
1261 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1266 * Direct block allocation dependencies.
1268 * When a new block is allocated, the corresponding disk locations must be
1269 * initialized (with zeros or new data) before the on-disk inode points to
1270 * them. Also, the freemap from which the block was allocated must be
1271 * updated (on disk) before the inode's pointer. These two dependencies are
1272 * independent of each other and are needed for all file blocks and indirect
1273 * blocks that are pointed to directly by the inode. Just before the
1274 * "in-core" version of the inode is updated with a newly allocated block
1275 * number, a procedure (below) is called to setup allocation dependency
1276 * structures. These structures are removed when the corresponding
1277 * dependencies are satisfied or when the block allocation becomes obsolete
1278 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1279 * fragment that gets upgraded). All of these cases are handled in
1280 * procedures described later.
1282 * When a file extension causes a fragment to be upgraded, either to a larger
1283 * fragment or to a full block, the on-disk location may change (if the
1284 * previous fragment could not simply be extended). In this case, the old
1285 * fragment must be de-allocated, but not until after the inode's pointer has
1286 * been updated. In most cases, this is handled by later procedures, which
1287 * will construct a "freefrag" structure to be added to the workitem queue
1288 * when the inode update is complete (or obsolete). The main exception to
1289 * this is when an allocation occurs while a pending allocation dependency
1290 * (for the same block pointer) remains. This case is handled in the main
1291 * allocation dependency setup procedure by immediately freeing the
1292 * unreferenced fragments.
1295 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1296 struct inode *ip; /* inode to which block is being added */
1297 ufs_lbn_t lbn; /* block pointer within inode */
1298 ufs_daddr_t newblkno; /* disk block number being added */
1299 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */
1300 long newsize; /* size of new block */
1301 long oldsize; /* size of new block */
1302 struct buf *bp; /* bp for allocated block */
1304 struct allocdirect *adp, *oldadp;
1305 struct allocdirectlst *adphead;
1306 struct bmsafemap *bmsafemap;
1307 struct inodedep *inodedep;
1308 struct pagedep *pagedep;
1309 struct newblk *newblk;
1311 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1312 M_ALLOCDIRECT, M_SOFTDEP_FLAGS);
1313 bzero(adp, sizeof(struct allocdirect));
1314 adp->ad_list.wk_type = D_ALLOCDIRECT;
1316 adp->ad_newblkno = newblkno;
1317 adp->ad_oldblkno = oldblkno;
1318 adp->ad_newsize = newsize;
1319 adp->ad_oldsize = oldsize;
1320 adp->ad_state = ATTACHED;
1321 if (newblkno == oldblkno)
1322 adp->ad_freefrag = NULL;
1324 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1326 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1327 panic("softdep_setup_allocdirect: lost block");
1330 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1331 adp->ad_inodedep = inodedep;
1333 if (newblk->nb_state == DEPCOMPLETE) {
1334 adp->ad_state |= DEPCOMPLETE;
1337 bmsafemap = newblk->nb_bmsafemap;
1338 adp->ad_buf = bmsafemap->sm_buf;
1339 LIST_REMOVE(newblk, nb_deps);
1340 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1342 LIST_REMOVE(newblk, nb_hash);
1343 FREE(newblk, M_NEWBLK);
1345 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1346 if (lbn >= NDADDR) {
1347 /* allocating an indirect block */
1348 if (oldblkno != 0) {
1350 panic("softdep_setup_allocdirect: non-zero indir");
1354 * Allocating a direct block.
1356 * If we are allocating a directory block, then we must
1357 * allocate an associated pagedep to track additions and
1360 if ((ip->i_mode & IFMT) == IFDIR &&
1361 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1362 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1365 * The list of allocdirects must be kept in sorted and ascending
1366 * order so that the rollback routines can quickly determine the
1367 * first uncommitted block (the size of the file stored on disk
1368 * ends at the end of the lowest committed fragment, or if there
1369 * are no fragments, at the end of the highest committed block).
1370 * Since files generally grow, the typical case is that the new
1371 * block is to be added at the end of the list. We speed this
1372 * special case by checking against the last allocdirect in the
1373 * list before laboriously traversing the list looking for the
1376 adphead = &inodedep->id_newinoupdt;
1377 oldadp = TAILQ_LAST(adphead, allocdirectlst);
1378 if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1379 /* insert at end of list */
1380 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1381 if (oldadp != NULL && oldadp->ad_lbn == lbn)
1382 allocdirect_merge(adphead, adp, oldadp);
1386 TAILQ_FOREACH(oldadp, adphead, ad_next) {
1387 if (oldadp->ad_lbn >= lbn)
1390 if (oldadp == NULL) {
1392 panic("softdep_setup_allocdirect: lost entry");
1394 /* insert in middle of list */
1395 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1396 if (oldadp->ad_lbn == lbn)
1397 allocdirect_merge(adphead, adp, oldadp);
1402 * Replace an old allocdirect dependency with a newer one.
1403 * This routine must be called with splbio interrupts blocked.
1406 allocdirect_merge(adphead, newadp, oldadp)
1407 struct allocdirectlst *adphead; /* head of list holding allocdirects */
1408 struct allocdirect *newadp; /* allocdirect being added */
1409 struct allocdirect *oldadp; /* existing allocdirect being checked */
1411 struct freefrag *freefrag;
1414 if (lk.lkt_held == -1)
1415 panic("allocdirect_merge: lock not held");
1417 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1418 newadp->ad_oldsize != oldadp->ad_newsize ||
1419 newadp->ad_lbn >= NDADDR) {
1421 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1422 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1425 newadp->ad_oldblkno = oldadp->ad_oldblkno;
1426 newadp->ad_oldsize = oldadp->ad_oldsize;
1428 * If the old dependency had a fragment to free or had never
1429 * previously had a block allocated, then the new dependency
1430 * can immediately post its freefrag and adopt the old freefrag.
1431 * This action is done by swapping the freefrag dependencies.
1432 * The new dependency gains the old one's freefrag, and the
1433 * old one gets the new one and then immediately puts it on
1434 * the worklist when it is freed by free_allocdirect. It is
1435 * not possible to do this swap when the old dependency had a
1436 * non-zero size but no previous fragment to free. This condition
1437 * arises when the new block is an extension of the old block.
1438 * Here, the first part of the fragment allocated to the new
1439 * dependency is part of the block currently claimed on disk by
1440 * the old dependency, so cannot legitimately be freed until the
1441 * conditions for the new dependency are fulfilled.
1443 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1444 freefrag = newadp->ad_freefrag;
1445 newadp->ad_freefrag = oldadp->ad_freefrag;
1446 oldadp->ad_freefrag = freefrag;
1448 free_allocdirect(adphead, oldadp, 0);
1452 * Allocate a new freefrag structure if needed.
1454 static struct freefrag *
1455 newfreefrag(ip, blkno, size)
1460 struct freefrag *freefrag;
1466 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1467 panic("newfreefrag: frag size");
1468 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1469 M_FREEFRAG, M_SOFTDEP_FLAGS);
1470 freefrag->ff_list.wk_type = D_FREEFRAG;
1471 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */
1472 freefrag->ff_inum = ip->i_number;
1473 freefrag->ff_fs = fs;
1474 freefrag->ff_devvp = ip->i_devvp;
1475 freefrag->ff_blkno = blkno;
1476 freefrag->ff_fragsize = size;
1481 * This workitem de-allocates fragments that were replaced during
1482 * file block allocation.
1485 handle_workitem_freefrag(freefrag)
1486 struct freefrag *freefrag;
1490 tip.i_fs = freefrag->ff_fs;
1491 tip.i_devvp = freefrag->ff_devvp;
1492 tip.i_dev = freefrag->ff_devvp->v_rdev;
1493 tip.i_number = freefrag->ff_inum;
1494 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */
1495 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1496 FREE(freefrag, M_FREEFRAG);
1500 * Indirect block allocation dependencies.
1502 * The same dependencies that exist for a direct block also exist when
1503 * a new block is allocated and pointed to by an entry in a block of
1504 * indirect pointers. The undo/redo states described above are also
1505 * used here. Because an indirect block contains many pointers that
1506 * may have dependencies, a second copy of the entire in-memory indirect
1507 * block is kept. The buffer cache copy is always completely up-to-date.
1508 * The second copy, which is used only as a source for disk writes,
1509 * contains only the safe pointers (i.e., those that have no remaining
1510 * update dependencies). The second copy is freed when all pointers
1511 * are safe. The cache is not allowed to replace indirect blocks with
1512 * pending update dependencies. If a buffer containing an indirect
1513 * block with dependencies is written, these routines will mark it
1514 * dirty again. It can only be successfully written once all the
1515 * dependencies are removed. The ffs_fsync routine in conjunction with
1516 * softdep_sync_metadata work together to get all the dependencies
1517 * removed so that a file can be successfully written to disk. Three
1518 * procedures are used when setting up indirect block pointer
1519 * dependencies. The division is necessary because of the organization
1520 * of the "balloc" routine and because of the distinction between file
1521 * pages and file metadata blocks.
1525 * Allocate a new allocindir structure.
1527 static struct allocindir *
1528 newallocindir(ip, ptrno, newblkno, oldblkno)
1529 struct inode *ip; /* inode for file being extended */
1530 int ptrno; /* offset of pointer in indirect block */
1531 ufs_daddr_t newblkno; /* disk block number being added */
1532 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1534 struct allocindir *aip;
1536 MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1537 M_ALLOCINDIR, M_SOFTDEP_FLAGS);
1538 bzero(aip, sizeof(struct allocindir));
1539 aip->ai_list.wk_type = D_ALLOCINDIR;
1540 aip->ai_state = ATTACHED;
1541 aip->ai_offset = ptrno;
1542 aip->ai_newblkno = newblkno;
1543 aip->ai_oldblkno = oldblkno;
1544 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1549 * Called just before setting an indirect block pointer
1550 * to a newly allocated file page.
1553 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1554 struct inode *ip; /* inode for file being extended */
1555 ufs_lbn_t lbn; /* allocated block number within file */
1556 struct buf *bp; /* buffer with indirect blk referencing page */
1557 int ptrno; /* offset of pointer in indirect block */
1558 ufs_daddr_t newblkno; /* disk block number being added */
1559 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1560 struct buf *nbp; /* buffer holding allocated page */
1562 struct allocindir *aip;
1563 struct pagedep *pagedep;
1565 aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1568 * If we are allocating a directory page, then we must
1569 * allocate an associated pagedep to track additions and
1572 if ((ip->i_mode & IFMT) == IFDIR &&
1573 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1574 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1575 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1577 setup_allocindir_phase2(bp, ip, aip);
1581 * Called just before setting an indirect block pointer to a
1582 * newly allocated indirect block.
1585 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1586 struct buf *nbp; /* newly allocated indirect block */
1587 struct inode *ip; /* inode for file being extended */
1588 struct buf *bp; /* indirect block referencing allocated block */
1589 int ptrno; /* offset of pointer in indirect block */
1590 ufs_daddr_t newblkno; /* disk block number being added */
1592 struct allocindir *aip;
1594 aip = newallocindir(ip, ptrno, newblkno, 0);
1596 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1598 setup_allocindir_phase2(bp, ip, aip);
1602 * Called to finish the allocation of the "aip" allocated
1603 * by one of the two routines above.
1606 setup_allocindir_phase2(bp, ip, aip)
1607 struct buf *bp; /* in-memory copy of the indirect block */
1608 struct inode *ip; /* inode for file being extended */
1609 struct allocindir *aip; /* allocindir allocated by the above routines */
1611 struct worklist *wk;
1612 struct indirdep *indirdep, *newindirdep;
1613 struct bmsafemap *bmsafemap;
1614 struct allocindir *oldaip;
1615 struct freefrag *freefrag;
1616 struct newblk *newblk;
1618 if (bp->b_lblkno >= 0)
1619 panic("setup_allocindir_phase2: not indir blk");
1620 for (indirdep = NULL, newindirdep = NULL; ; ) {
1622 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1623 if (wk->wk_type != D_INDIRDEP)
1625 indirdep = WK_INDIRDEP(wk);
1628 if (indirdep == NULL && newindirdep) {
1629 indirdep = newindirdep;
1630 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1635 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1637 panic("setup_allocindir: lost block");
1639 if (newblk->nb_state == DEPCOMPLETE) {
1640 aip->ai_state |= DEPCOMPLETE;
1643 bmsafemap = newblk->nb_bmsafemap;
1644 aip->ai_buf = bmsafemap->sm_buf;
1645 LIST_REMOVE(newblk, nb_deps);
1646 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1649 LIST_REMOVE(newblk, nb_hash);
1650 FREE(newblk, M_NEWBLK);
1651 aip->ai_indirdep = indirdep;
1653 * Check to see if there is an existing dependency
1654 * for this block. If there is, merge the old
1655 * dependency into the new one.
1657 if (aip->ai_oldblkno == 0)
1661 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1662 if (oldaip->ai_offset == aip->ai_offset)
1664 if (oldaip != NULL) {
1665 if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1667 panic("setup_allocindir_phase2: blkno");
1669 aip->ai_oldblkno = oldaip->ai_oldblkno;
1670 freefrag = oldaip->ai_freefrag;
1671 oldaip->ai_freefrag = aip->ai_freefrag;
1672 aip->ai_freefrag = freefrag;
1673 free_allocindir(oldaip, NULL);
1675 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1676 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1677 [aip->ai_offset] = aip->ai_oldblkno;
1681 if (indirdep->ir_savebp != NULL)
1682 brelse(newindirdep->ir_savebp);
1683 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1687 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1688 M_INDIRDEP, M_SOFTDEP_FLAGS);
1689 newindirdep->ir_list.wk_type = D_INDIRDEP;
1690 newindirdep->ir_state = ATTACHED;
1691 LIST_INIT(&newindirdep->ir_deplisthd);
1692 LIST_INIT(&newindirdep->ir_donehd);
1693 if (bp->b_blkno == bp->b_lblkno) {
1694 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
1697 newindirdep->ir_savebp =
1698 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0);
1699 BUF_KERNPROC(newindirdep->ir_savebp);
1700 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1705 * Block de-allocation dependencies.
1707 * When blocks are de-allocated, the on-disk pointers must be nullified before
1708 * the blocks are made available for use by other files. (The true
1709 * requirement is that old pointers must be nullified before new on-disk
1710 * pointers are set. We chose this slightly more stringent requirement to
1711 * reduce complexity.) Our implementation handles this dependency by updating
1712 * the inode (or indirect block) appropriately but delaying the actual block
1713 * de-allocation (i.e., freemap and free space count manipulation) until
1714 * after the updated versions reach stable storage. After the disk is
1715 * updated, the blocks can be safely de-allocated whenever it is convenient.
1716 * This implementation handles only the common case of reducing a file's
1717 * length to zero. Other cases are handled by the conventional synchronous
1720 * The ffs implementation with which we worked double-checks
1721 * the state of the block pointers and file size as it reduces
1722 * a file's length. Some of this code is replicated here in our
1723 * soft updates implementation. The freeblks->fb_chkcnt field is
1724 * used to transfer a part of this information to the procedure
1725 * that eventually de-allocates the blocks.
1727 * This routine should be called from the routine that shortens
1728 * a file's length, before the inode's size or block pointers
1729 * are modified. It will save the block pointer information for
1730 * later release and zero the inode so that the calling routine
1734 softdep_setup_freeblocks(ip, length)
1735 struct inode *ip; /* The inode whose length is to be reduced */
1736 off_t length; /* The new length for the file */
1738 struct freeblks *freeblks;
1739 struct inodedep *inodedep;
1740 struct allocdirect *adp;
1744 int i, error, delay;
1748 panic("softde_setup_freeblocks: non-zero length");
1749 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1750 M_FREEBLKS, M_SOFTDEP_FLAGS);
1751 bzero(freeblks, sizeof(struct freeblks));
1752 freeblks->fb_list.wk_type = D_FREEBLKS;
1753 freeblks->fb_uid = ip->i_uid;
1754 freeblks->fb_previousinum = ip->i_number;
1755 freeblks->fb_devvp = ip->i_devvp;
1756 freeblks->fb_fs = fs;
1757 freeblks->fb_oldsize = ip->i_size;
1758 freeblks->fb_newsize = length;
1759 freeblks->fb_chkcnt = ip->i_blocks;
1760 for (i = 0; i < NDADDR; i++) {
1761 freeblks->fb_dblks[i] = ip->i_db[i];
1764 for (i = 0; i < NIADDR; i++) {
1765 freeblks->fb_iblks[i] = ip->i_ib[i];
1771 * Push the zero'ed inode to to its disk buffer so that we are free
1772 * to delete its dependencies below. Once the dependencies are gone
1773 * the buffer can be safely released.
1775 if ((error = bread(ip->i_devvp,
1776 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
1777 (int)fs->fs_bsize, NOCRED, &bp)) != 0)
1778 softdep_error("softdep_setup_freeblocks", error);
1779 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
1782 * Find and eliminate any inode dependencies.
1785 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1786 if ((inodedep->id_state & IOSTARTED) != 0) {
1788 panic("softdep_setup_freeblocks: inode busy");
1791 * Add the freeblks structure to the list of operations that
1792 * must await the zero'ed inode being written to disk. If we
1793 * still have a bitmap dependency (delay == 0), then the inode
1794 * has never been written to disk, so we can process the
1795 * freeblks below once we have deleted the dependencies.
1797 delay = (inodedep->id_state & DEPCOMPLETE);
1799 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1801 * Because the file length has been truncated to zero, any
1802 * pending block allocation dependency structures associated
1803 * with this inode are obsolete and can simply be de-allocated.
1804 * We must first merge the two dependency lists to get rid of
1805 * any duplicate freefrag structures, then purge the merged list.
1807 merge_inode_lists(inodedep);
1808 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1809 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
1813 * We must wait for any I/O in progress to finish so that
1814 * all potential buffers on the dirty list will be visible.
1815 * Once they are all there, walk the list and get rid of
1820 drain_output(vp, 1);
1821 while (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) {
1822 bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
1823 (void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
1824 deallocate_dependencies(bp, inodedep);
1825 bp->b_flags |= B_INVAL | B_NOCACHE;
1830 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
1831 (void)free_inodedep(inodedep);
1834 * If the inode has never been written to disk (delay == 0),
1835 * then we can process the freeblks now that we have deleted
1839 handle_workitem_freeblocks(freeblks);
1843 * Reclaim any dependency structures from a buffer that is about to
1844 * be reallocated to a new vnode. The buffer must be locked, thus,
1845 * no I/O completion operations can occur while we are manipulating
1846 * its associated dependencies. The mutex is held so that other I/O's
1847 * associated with related dependencies do not occur.
1850 deallocate_dependencies(bp, inodedep)
1852 struct inodedep *inodedep;
1854 struct worklist *wk;
1855 struct indirdep *indirdep;
1856 struct allocindir *aip;
1857 struct pagedep *pagedep;
1858 struct dirrem *dirrem;
1862 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1863 switch (wk->wk_type) {
1866 indirdep = WK_INDIRDEP(wk);
1868 * None of the indirect pointers will ever be visible,
1869 * so they can simply be tossed. GOINGAWAY ensures
1870 * that allocated pointers will be saved in the buffer
1871 * cache until they are freed. Note that they will
1872 * only be able to be found by their physical address
1873 * since the inode mapping the logical address will
1874 * be gone. The save buffer used for the safe copy
1875 * was allocated in setup_allocindir_phase2 using
1876 * the physical address so it could be used for this
1877 * purpose. Hence we swap the safe copy with the real
1878 * copy, allowing the safe copy to be freed and holding
1879 * on to the real copy for later use in indir_trunc.
1881 if (indirdep->ir_state & GOINGAWAY) {
1883 panic("deallocate_dependencies: already gone");
1885 indirdep->ir_state |= GOINGAWAY;
1886 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
1887 free_allocindir(aip, inodedep);
1888 if (bp->b_lblkno >= 0 ||
1889 bp->b_blkno != indirdep->ir_savebp->b_lblkno) {
1891 panic("deallocate_dependencies: not indir");
1893 bcopy(bp->b_data, indirdep->ir_savebp->b_data,
1895 WORKLIST_REMOVE(wk);
1896 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
1900 pagedep = WK_PAGEDEP(wk);
1902 * None of the directory additions will ever be
1903 * visible, so they can simply be tossed.
1905 for (i = 0; i < DAHASHSZ; i++)
1907 LIST_FIRST(&pagedep->pd_diraddhd[i])))
1909 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
1912 * Copy any directory remove dependencies to the list
1913 * to be processed after the zero'ed inode is written.
1914 * If the inode has already been written, then they
1915 * can be dumped directly onto the work list.
1917 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
1918 LIST_REMOVE(dirrem, dm_next);
1919 dirrem->dm_dirinum = pagedep->pd_ino;
1920 if (inodedep == NULL ||
1921 (inodedep->id_state & ALLCOMPLETE) ==
1923 add_to_worklist(&dirrem->dm_list);
1925 WORKLIST_INSERT(&inodedep->id_bufwait,
1928 WORKLIST_REMOVE(&pagedep->pd_list);
1929 LIST_REMOVE(pagedep, pd_hash);
1930 WORKITEM_FREE(pagedep, D_PAGEDEP);
1934 free_allocindir(WK_ALLOCINDIR(wk), inodedep);
1940 panic("deallocate_dependencies: Unexpected type %s",
1941 TYPENAME(wk->wk_type));
1946 panic("deallocate_dependencies: Unknown type %s",
1947 TYPENAME(wk->wk_type));
1954 * Free an allocdirect. Generate a new freefrag work request if appropriate.
1955 * This routine must be called with splbio interrupts blocked.
1958 free_allocdirect(adphead, adp, delay)
1959 struct allocdirectlst *adphead;
1960 struct allocdirect *adp;
1965 if (lk.lkt_held == -1)
1966 panic("free_allocdirect: lock not held");
1968 if ((adp->ad_state & DEPCOMPLETE) == 0)
1969 LIST_REMOVE(adp, ad_deps);
1970 TAILQ_REMOVE(adphead, adp, ad_next);
1971 if ((adp->ad_state & COMPLETE) == 0)
1972 WORKLIST_REMOVE(&adp->ad_list);
1973 if (adp->ad_freefrag != NULL) {
1975 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
1976 &adp->ad_freefrag->ff_list);
1978 add_to_worklist(&adp->ad_freefrag->ff_list);
1980 WORKITEM_FREE(adp, D_ALLOCDIRECT);
1984 * Prepare an inode to be freed. The actual free operation is not
1985 * done until the zero'ed inode has been written to disk.
1988 softdep_freefile(pvp, ino, mode)
1993 struct inode *ip = VTOI(pvp);
1994 struct inodedep *inodedep;
1995 struct freefile *freefile;
1998 * This sets up the inode de-allocation dependency.
2000 MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2001 M_FREEFILE, M_SOFTDEP_FLAGS);
2002 freefile->fx_list.wk_type = D_FREEFILE;
2003 freefile->fx_list.wk_state = 0;
2004 freefile->fx_mode = mode;
2005 freefile->fx_oldinum = ino;
2006 freefile->fx_devvp = ip->i_devvp;
2007 freefile->fx_fs = ip->i_fs;
2010 * If the inodedep does not exist, then the zero'ed inode has
2011 * been written to disk. If the allocated inode has never been
2012 * written to disk, then the on-disk inode is zero'ed. In either
2013 * case we can free the file immediately.
2016 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2017 check_inode_unwritten(inodedep)) {
2019 handle_workitem_freefile(freefile);
2022 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2027 * Check to see if an inode has never been written to disk. If
2028 * so free the inodedep and return success, otherwise return failure.
2029 * This routine must be called with splbio interrupts blocked.
2031 * If we still have a bitmap dependency, then the inode has never
2032 * been written to disk. Drop the dependency as it is no longer
2033 * necessary since the inode is being deallocated. We set the
2034 * ALLCOMPLETE flags since the bitmap now properly shows that the
2035 * inode is not allocated. Even if the inode is actively being
2036 * written, it has been rolled back to its zero'ed state, so we
2037 * are ensured that a zero inode is what is on the disk. For short
2038 * lived files, this change will usually result in removing all the
2039 * dependencies from the inode so that it can be freed immediately.
2042 check_inode_unwritten(inodedep)
2043 struct inodedep *inodedep;
2046 if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2047 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2048 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2049 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2050 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2051 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2052 inodedep->id_nlinkdelta != 0)
2054 inodedep->id_state |= ALLCOMPLETE;
2055 LIST_REMOVE(inodedep, id_deps);
2056 inodedep->id_buf = NULL;
2057 if (inodedep->id_state & ONWORKLIST)
2058 WORKLIST_REMOVE(&inodedep->id_list);
2059 if (inodedep->id_savedino != NULL) {
2060 FREE(inodedep->id_savedino, M_INODEDEP);
2061 inodedep->id_savedino = NULL;
2063 if (free_inodedep(inodedep) == 0) {
2065 panic("check_inode_unwritten: busy inode");
2071 * Try to free an inodedep structure. Return 1 if it could be freed.
2074 free_inodedep(inodedep)
2075 struct inodedep *inodedep;
2078 if ((inodedep->id_state & ONWORKLIST) != 0 ||
2079 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2080 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2081 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2082 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2083 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2084 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2085 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
2087 LIST_REMOVE(inodedep, id_hash);
2088 WORKITEM_FREE(inodedep, D_INODEDEP);
2094 * This workitem routine performs the block de-allocation.
2095 * The workitem is added to the pending list after the updated
2096 * inode block has been written to disk. As mentioned above,
2097 * checks regarding the number of blocks de-allocated (compared
2098 * to the number of blocks allocated for the file) are also
2099 * performed in this function.
2102 handle_workitem_freeblocks(freeblks)
2103 struct freeblks *freeblks;
2108 int i, level, bsize;
2109 long nblocks, blocksreleased = 0;
2110 int error, allerror = 0;
2111 ufs_lbn_t baselbns[NIADDR], tmpval;
2113 tip.i_number = freeblks->fb_previousinum;
2114 tip.i_devvp = freeblks->fb_devvp;
2115 tip.i_dev = freeblks->fb_devvp->v_rdev;
2116 tip.i_fs = freeblks->fb_fs;
2117 tip.i_size = freeblks->fb_oldsize;
2118 tip.i_uid = freeblks->fb_uid;
2119 fs = freeblks->fb_fs;
2121 baselbns[0] = NDADDR;
2122 for (i = 1; i < NIADDR; i++) {
2123 tmpval *= NINDIR(fs);
2124 baselbns[i] = baselbns[i - 1] + tmpval;
2126 nblocks = btodb(fs->fs_bsize);
2129 * Indirect blocks first.
2131 for (level = (NIADDR - 1); level >= 0; level--) {
2132 if ((bn = freeblks->fb_iblks[level]) == 0)
2134 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level,
2135 baselbns[level], &blocksreleased)) == 0)
2137 ffs_blkfree(&tip, bn, fs->fs_bsize);
2138 blocksreleased += nblocks;
2141 * All direct blocks or frags.
2143 for (i = (NDADDR - 1); i >= 0; i--) {
2144 if ((bn = freeblks->fb_dblks[i]) == 0)
2146 bsize = blksize(fs, &tip, i);
2147 ffs_blkfree(&tip, bn, bsize);
2148 blocksreleased += btodb(bsize);
2152 if (freeblks->fb_chkcnt != blocksreleased)
2153 printf("handle_workitem_freeblocks: block count\n");
2155 softdep_error("handle_workitem_freeblks", allerror);
2156 #endif /* DIAGNOSTIC */
2157 WORKITEM_FREE(freeblks, D_FREEBLKS);
2161 * Release blocks associated with the inode ip and stored in the indirect
2162 * block dbn. If level is greater than SINGLE, the block is an indirect block
2163 * and recursive calls to indirtrunc must be used to cleanse other indirect
2167 indir_trunc(ip, dbn, level, lbn, countp)
2178 struct worklist *wk;
2179 struct indirdep *indirdep;
2180 int i, lbnadd, nblocks;
2181 int error, allerror = 0;
2185 for (i = level; i > 0; i--)
2186 lbnadd *= NINDIR(fs);
2188 * Get buffer of block pointers to be freed. This routine is not
2189 * called until the zero'ed inode has been written, so it is safe
2190 * to free blocks as they are encountered. Because the inode has
2191 * been zero'ed, calls to bmap on these blocks will fail. So, we
2192 * have to use the on-disk address and the block device for the
2193 * filesystem to look them up. If the file was deleted before its
2194 * indirect blocks were all written to disk, the routine that set
2195 * us up (deallocate_dependencies) will have arranged to leave
2196 * a complete copy of the indirect block in memory for our use.
2197 * Otherwise we have to read the blocks in from the disk.
2200 if ((bp = incore(ip->i_devvp, dbn)) != NULL &&
2201 (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2202 if (wk->wk_type != D_INDIRDEP ||
2203 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2204 (indirdep->ir_state & GOINGAWAY) == 0) {
2206 panic("indir_trunc: lost indirdep");
2208 WORKLIST_REMOVE(wk);
2209 WORKITEM_FREE(indirdep, D_INDIRDEP);
2210 if (LIST_FIRST(&bp->b_dep) != NULL) {
2212 panic("indir_trunc: dangling dep");
2217 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, NOCRED, &bp);
2222 * Recursively free indirect blocks.
2224 bap = (ufs_daddr_t *)bp->b_data;
2225 nblocks = btodb(fs->fs_bsize);
2226 for (i = NINDIR(fs) - 1; i >= 0; i--) {
2227 if ((nb = bap[i]) == 0)
2230 if ((error = indir_trunc(ip, fsbtodb(fs, nb),
2231 level - 1, lbn + (i * lbnadd), countp)) != 0)
2234 ffs_blkfree(ip, nb, fs->fs_bsize);
2237 bp->b_flags |= B_INVAL | B_NOCACHE;
2243 * Free an allocindir.
2244 * This routine must be called with splbio interrupts blocked.
2247 free_allocindir(aip, inodedep)
2248 struct allocindir *aip;
2249 struct inodedep *inodedep;
2251 struct freefrag *freefrag;
2254 if (lk.lkt_held == -1)
2255 panic("free_allocindir: lock not held");
2257 if ((aip->ai_state & DEPCOMPLETE) == 0)
2258 LIST_REMOVE(aip, ai_deps);
2259 if (aip->ai_state & ONWORKLIST)
2260 WORKLIST_REMOVE(&aip->ai_list);
2261 LIST_REMOVE(aip, ai_next);
2262 if ((freefrag = aip->ai_freefrag) != NULL) {
2263 if (inodedep == NULL)
2264 add_to_worklist(&freefrag->ff_list);
2266 WORKLIST_INSERT(&inodedep->id_bufwait,
2267 &freefrag->ff_list);
2269 WORKITEM_FREE(aip, D_ALLOCINDIR);
2273 * Directory entry addition dependencies.
2275 * When adding a new directory entry, the inode (with its incremented link
2276 * count) must be written to disk before the directory entry's pointer to it.
2277 * Also, if the inode is newly allocated, the corresponding freemap must be
2278 * updated (on disk) before the directory entry's pointer. These requirements
2279 * are met via undo/redo on the directory entry's pointer, which consists
2280 * simply of the inode number.
2282 * As directory entries are added and deleted, the free space within a
2283 * directory block can become fragmented. The ufs file system will compact
2284 * a fragmented directory block to make space for a new entry. When this
2285 * occurs, the offsets of previously added entries change. Any "diradd"
2286 * dependency structures corresponding to these entries must be updated with
2291 * This routine is called after the in-memory inode's link
2292 * count has been incremented, but before the directory entry's
2293 * pointer to the inode has been set.
2296 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp)
2297 struct buf *bp; /* buffer containing directory block */
2298 struct inode *dp; /* inode for directory */
2299 off_t diroffset; /* offset of new entry in directory */
2300 long newinum; /* inode referenced by new directory entry */
2301 struct buf *newdirbp; /* non-NULL => contents of new mkdir */
2303 int offset; /* offset of new entry within directory block */
2304 ufs_lbn_t lbn; /* block in directory containing new entry */
2307 struct pagedep *pagedep;
2308 struct inodedep *inodedep;
2309 struct mkdir *mkdir1, *mkdir2;
2312 * Whiteouts have no dependencies.
2314 if (newinum == WINO) {
2315 if (newdirbp != NULL)
2321 lbn = lblkno(fs, diroffset);
2322 offset = blkoff(fs, diroffset);
2323 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2325 bzero(dap, sizeof(struct diradd));
2326 dap->da_list.wk_type = D_DIRADD;
2327 dap->da_offset = offset;
2328 dap->da_newinum = newinum;
2329 dap->da_state = ATTACHED;
2330 if (newdirbp == NULL) {
2331 dap->da_state |= DEPCOMPLETE;
2334 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2335 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2337 mkdir1->md_list.wk_type = D_MKDIR;
2338 mkdir1->md_state = MKDIR_BODY;
2339 mkdir1->md_diradd = dap;
2340 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2342 mkdir2->md_list.wk_type = D_MKDIR;
2343 mkdir2->md_state = MKDIR_PARENT;
2344 mkdir2->md_diradd = dap;
2346 * Dependency on "." and ".." being written to disk.
2348 mkdir1->md_buf = newdirbp;
2350 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2351 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2355 * Dependency on link count increase for parent directory
2358 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2359 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2360 dap->da_state &= ~MKDIR_PARENT;
2361 WORKITEM_FREE(mkdir2, D_MKDIR);
2363 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2364 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2368 * Link into parent directory pagedep to await its being written.
2370 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2371 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2372 dap->da_pagedep = pagedep;
2373 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2376 * Link into its inodedep. Put it on the id_bufwait list if the inode
2377 * is not yet written. If it is written, do the post-inode write
2378 * processing to put it on the id_pendinghd list.
2380 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2381 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2382 diradd_inode_written(dap, inodedep);
2384 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2389 * This procedure is called to change the offset of a directory
2390 * entry when compacting a directory block which must be owned
2391 * exclusively by the caller. Note that the actual entry movement
2392 * must be done in this procedure to ensure that no I/O completions
2393 * occur while the move is in progress.
2396 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2397 struct inode *dp; /* inode for directory */
2398 caddr_t base; /* address of dp->i_offset */
2399 caddr_t oldloc; /* address of old directory location */
2400 caddr_t newloc; /* address of new directory location */
2401 int entrysize; /* size of directory entry */
2403 int offset, oldoffset, newoffset;
2404 struct pagedep *pagedep;
2409 lbn = lblkno(dp->i_fs, dp->i_offset);
2410 offset = blkoff(dp->i_fs, dp->i_offset);
2411 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2413 oldoffset = offset + (oldloc - base);
2414 newoffset = offset + (newloc - base);
2416 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2417 if (dap->da_offset != oldoffset)
2419 dap->da_offset = newoffset;
2420 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2422 LIST_REMOVE(dap, da_pdlist);
2423 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2429 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2430 if (dap->da_offset == oldoffset) {
2431 dap->da_offset = newoffset;
2437 bcopy(oldloc, newloc, entrysize);
2442 * Free a diradd dependency structure. This routine must be called
2443 * with splbio interrupts blocked.
2449 struct dirrem *dirrem;
2450 struct pagedep *pagedep;
2451 struct inodedep *inodedep;
2452 struct mkdir *mkdir, *nextmd;
2455 if (lk.lkt_held == -1)
2456 panic("free_diradd: lock not held");
2458 WORKLIST_REMOVE(&dap->da_list);
2459 LIST_REMOVE(dap, da_pdlist);
2460 if ((dap->da_state & DIRCHG) == 0) {
2461 pagedep = dap->da_pagedep;
2463 dirrem = dap->da_previous;
2464 pagedep = dirrem->dm_pagedep;
2465 dirrem->dm_dirinum = pagedep->pd_ino;
2466 add_to_worklist(&dirrem->dm_list);
2468 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2470 (void) free_inodedep(inodedep);
2471 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2472 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2473 nextmd = LIST_NEXT(mkdir, md_mkdirs);
2474 if (mkdir->md_diradd != dap)
2476 dap->da_state &= ~mkdir->md_state;
2477 WORKLIST_REMOVE(&mkdir->md_list);
2478 LIST_REMOVE(mkdir, md_mkdirs);
2479 WORKITEM_FREE(mkdir, D_MKDIR);
2481 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2483 panic("free_diradd: unfound ref");
2486 WORKITEM_FREE(dap, D_DIRADD);
2490 * Directory entry removal dependencies.
2492 * When removing a directory entry, the entry's inode pointer must be
2493 * zero'ed on disk before the corresponding inode's link count is decremented
2494 * (possibly freeing the inode for re-use). This dependency is handled by
2495 * updating the directory entry but delaying the inode count reduction until
2496 * after the directory block has been written to disk. After this point, the
2497 * inode count can be decremented whenever it is convenient.
2501 * This routine should be called immediately after removing
2502 * a directory entry. The inode's link count should not be
2503 * decremented by the calling procedure -- the soft updates
2504 * code will do this task when it is safe.
2507 softdep_setup_remove(bp, dp, ip, isrmdir)
2508 struct buf *bp; /* buffer containing directory block */
2509 struct inode *dp; /* inode for the directory being modified */
2510 struct inode *ip; /* inode for directory entry being removed */
2511 int isrmdir; /* indicates if doing RMDIR */
2513 struct dirrem *dirrem, *prevdirrem;
2516 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2518 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2521 * If the COMPLETE flag is clear, then there were no active
2522 * entries and we want to roll back to a zeroed entry until
2523 * the new inode is committed to disk. If the COMPLETE flag is
2524 * set then we have deleted an entry that never made it to
2525 * disk. If the entry we deleted resulted from a name change,
2526 * then the old name still resides on disk. We cannot delete
2527 * its inode (returned to us in prevdirrem) until the zeroed
2528 * directory entry gets to disk. The new inode has never been
2529 * referenced on the disk, so can be deleted immediately.
2531 if ((dirrem->dm_state & COMPLETE) == 0) {
2532 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2536 if (prevdirrem != NULL)
2537 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2538 prevdirrem, dm_next);
2539 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2541 handle_workitem_remove(dirrem);
2546 * Allocate a new dirrem if appropriate and return it along with
2547 * its associated pagedep. Called without a lock, returns with lock.
2549 static long num_dirrem; /* number of dirrem allocated */
2550 static struct dirrem *
2551 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2552 struct buf *bp; /* buffer containing directory block */
2553 struct inode *dp; /* inode for the directory being modified */
2554 struct inode *ip; /* inode for directory entry being removed */
2555 int isrmdir; /* indicates if doing RMDIR */
2556 struct dirrem **prevdirremp; /* previously referenced inode, if any */
2561 struct dirrem *dirrem;
2562 struct pagedep *pagedep;
2565 * Whiteouts have no deletion dependencies.
2568 panic("newdirrem: whiteout");
2570 * If we are over our limit, try to improve the situation.
2571 * Limiting the number of dirrem structures will also limit
2572 * the number of freefile and freeblks structures.
2574 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2575 (void) request_cleanup(FLUSH_REMOVE, 0);
2577 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
2578 M_DIRREM, M_SOFTDEP_FLAGS);
2579 bzero(dirrem, sizeof(struct dirrem));
2580 dirrem->dm_list.wk_type = D_DIRREM;
2581 dirrem->dm_state = isrmdir ? RMDIR : 0;
2582 dirrem->dm_mnt = ITOV(ip)->v_mount;
2583 dirrem->dm_oldinum = ip->i_number;
2584 *prevdirremp = NULL;
2587 lbn = lblkno(dp->i_fs, dp->i_offset);
2588 offset = blkoff(dp->i_fs, dp->i_offset);
2589 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2590 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2591 dirrem->dm_pagedep = pagedep;
2593 * Check for a diradd dependency for the same directory entry.
2594 * If present, then both dependencies become obsolete and can
2595 * be de-allocated. Check for an entry on both the pd_dirraddhd
2596 * list and the pd_pendinghd list.
2599 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
2600 if (dap->da_offset == offset)
2604 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
2605 if (dap->da_offset == offset)
2611 * Must be ATTACHED at this point.
2613 if ((dap->da_state & ATTACHED) == 0) {
2615 panic("newdirrem: not ATTACHED");
2617 if (dap->da_newinum != ip->i_number) {
2619 panic("newdirrem: inum %d should be %d",
2620 ip->i_number, dap->da_newinum);
2623 * If we are deleting a changed name that never made it to disk,
2624 * then return the dirrem describing the previous inode (which
2625 * represents the inode currently referenced from this entry on disk).
2627 if ((dap->da_state & DIRCHG) != 0) {
2628 *prevdirremp = dap->da_previous;
2629 dap->da_state &= ~DIRCHG;
2630 dap->da_pagedep = pagedep;
2633 * We are deleting an entry that never made it to disk.
2634 * Mark it COMPLETE so we can delete its inode immediately.
2636 dirrem->dm_state |= COMPLETE;
2642 * Directory entry change dependencies.
2644 * Changing an existing directory entry requires that an add operation
2645 * be completed first followed by a deletion. The semantics for the addition
2646 * are identical to the description of adding a new entry above except
2647 * that the rollback is to the old inode number rather than zero. Once
2648 * the addition dependency is completed, the removal is done as described
2649 * in the removal routine above.
2653 * This routine should be called immediately after changing
2654 * a directory entry. The inode's link count should not be
2655 * decremented by the calling procedure -- the soft updates
2656 * code will perform this task when it is safe.
2659 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
2660 struct buf *bp; /* buffer containing directory block */
2661 struct inode *dp; /* inode for the directory being modified */
2662 struct inode *ip; /* inode for directory entry being removed */
2663 long newinum; /* new inode number for changed entry */
2664 int isrmdir; /* indicates if doing RMDIR */
2667 struct diradd *dap = NULL;
2668 struct dirrem *dirrem, *prevdirrem;
2669 struct pagedep *pagedep;
2670 struct inodedep *inodedep;
2672 offset = blkoff(dp->i_fs, dp->i_offset);
2675 * Whiteouts do not need diradd dependencies.
2677 if (newinum != WINO) {
2678 MALLOC(dap, struct diradd *, sizeof(struct diradd),
2679 M_DIRADD, M_SOFTDEP_FLAGS);
2680 bzero(dap, sizeof(struct diradd));
2681 dap->da_list.wk_type = D_DIRADD;
2682 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2683 dap->da_offset = offset;
2684 dap->da_newinum = newinum;
2688 * Allocate a new dirrem and ACQUIRE_LOCK.
2690 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2691 pagedep = dirrem->dm_pagedep;
2693 * The possible values for isrmdir:
2694 * 0 - non-directory file rename
2695 * 1 - directory rename within same directory
2696 * inum - directory rename to new directory of given inode number
2697 * When renaming to a new directory, we are both deleting and
2698 * creating a new directory entry, so the link count on the new
2699 * directory should not change. Thus we do not need the followup
2700 * dirrem which is usually done in handle_workitem_remove. We set
2701 * the DIRCHG flag to tell handle_workitem_remove to skip the
2705 dirrem->dm_state |= DIRCHG;
2708 * Whiteouts have no additional dependencies,
2709 * so just put the dirrem on the correct list.
2711 if (newinum == WINO) {
2712 if ((dirrem->dm_state & COMPLETE) == 0) {
2713 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2716 dirrem->dm_dirinum = pagedep->pd_ino;
2717 add_to_worklist(&dirrem->dm_list);
2724 * If the COMPLETE flag is clear, then there were no active
2725 * entries and we want to roll back to the previous inode until
2726 * the new inode is committed to disk. If the COMPLETE flag is
2727 * set, then we have deleted an entry that never made it to disk.
2728 * If the entry we deleted resulted from a name change, then the old
2729 * inode reference still resides on disk. Any rollback that we do
2730 * needs to be to that old inode (returned to us in prevdirrem). If
2731 * the entry we deleted resulted from a create, then there is
2732 * no entry on the disk, so we want to roll back to zero rather
2733 * than the uncommitted inode. In either of the COMPLETE cases we
2734 * want to immediately free the unwritten and unreferenced inode.
2736 if ((dirrem->dm_state & COMPLETE) == 0) {
2737 dap->da_previous = dirrem;
2739 if (prevdirrem != NULL) {
2740 dap->da_previous = prevdirrem;
2742 dap->da_state &= ~DIRCHG;
2743 dap->da_pagedep = pagedep;
2745 dirrem->dm_dirinum = pagedep->pd_ino;
2746 add_to_worklist(&dirrem->dm_list);
2749 * Link into its inodedep. Put it on the id_bufwait list if the inode
2750 * is not yet written. If it is written, do the post-inode write
2751 * processing to put it on the id_pendinghd list.
2753 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2754 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2755 dap->da_state |= COMPLETE;
2756 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2757 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2759 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2761 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2767 * Called whenever the link count on an inode is changed.
2768 * It creates an inode dependency so that the new reference(s)
2769 * to the inode cannot be committed to disk until the updated
2770 * inode has been written.
2773 softdep_change_linkcnt(ip)
2774 struct inode *ip; /* the inode with the increased link count */
2776 struct inodedep *inodedep;
2779 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2780 if (ip->i_nlink < ip->i_effnlink) {
2782 panic("softdep_change_linkcnt: bad delta");
2784 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2789 * This workitem decrements the inode's link count.
2790 * If the link count reaches zero, the file is removed.
2793 handle_workitem_remove(dirrem)
2794 struct dirrem *dirrem;
2796 struct proc *p = CURPROC; /* XXX */
2797 struct inodedep *inodedep;
2803 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) {
2804 softdep_error("handle_workitem_remove: vget", error);
2809 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
2811 panic("handle_workitem_remove: lost inodedep");
2814 * Normal file deletion.
2816 if ((dirrem->dm_state & RMDIR) == 0) {
2818 ip->i_flag |= IN_CHANGE;
2819 if (ip->i_nlink < ip->i_effnlink) {
2821 panic("handle_workitem_remove: bad file delta");
2823 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2827 WORKITEM_FREE(dirrem, D_DIRREM);
2831 * Directory deletion. Decrement reference count for both the
2832 * just deleted parent directory entry and the reference for ".".
2833 * Next truncate the directory to length zero. When the
2834 * truncation completes, arrange to have the reference count on
2835 * the parent decremented to account for the loss of "..".
2838 ip->i_flag |= IN_CHANGE;
2839 if (ip->i_nlink < ip->i_effnlink) {
2841 panic("handle_workitem_remove: bad dir delta");
2843 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2845 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, p->p_ucred, p)) != 0)
2846 softdep_error("handle_workitem_remove: truncate", error);
2848 * Rename a directory to a new parent. Since, we are both deleting
2849 * and creating a new directory entry, the link count on the new
2850 * directory should not change. Thus we skip the followup dirrem.
2852 if (dirrem->dm_state & DIRCHG) {
2855 WORKITEM_FREE(dirrem, D_DIRREM);
2859 * If the inodedep does not exist, then the zero'ed inode has
2860 * been written to disk. If the allocated inode has never been
2861 * written to disk, then the on-disk inode is zero'ed. In either
2862 * case we can remove the file immediately.
2865 dirrem->dm_state = 0;
2866 oldinum = dirrem->dm_oldinum;
2867 dirrem->dm_oldinum = dirrem->dm_dirinum;
2868 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
2869 check_inode_unwritten(inodedep)) {
2872 handle_workitem_remove(dirrem);
2875 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
2881 * Inode de-allocation dependencies.
2883 * When an inode's link count is reduced to zero, it can be de-allocated. We
2884 * found it convenient to postpone de-allocation until after the inode is
2885 * written to disk with its new link count (zero). At this point, all of the
2886 * on-disk inode's block pointers are nullified and, with careful dependency
2887 * list ordering, all dependencies related to the inode will be satisfied and
2888 * the corresponding dependency structures de-allocated. So, if/when the
2889 * inode is reused, there will be no mixing of old dependencies with new
2890 * ones. This artificial dependency is set up by the block de-allocation
2891 * procedure above (softdep_setup_freeblocks) and completed by the
2892 * following procedure.
2895 handle_workitem_freefile(freefile)
2896 struct freefile *freefile;
2900 struct inodedep *idp;
2905 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp);
2908 panic("handle_workitem_freefile: inodedep survived");
2910 tip.i_devvp = freefile->fx_devvp;
2911 tip.i_dev = freefile->fx_devvp->v_rdev;
2912 tip.i_fs = freefile->fx_fs;
2914 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
2915 softdep_error("handle_workitem_freefile", error);
2916 WORKITEM_FREE(freefile, D_FREEFILE);
2922 * The dependency structures constructed above are most actively used when file
2923 * system blocks are written to disk. No constraints are placed on when a
2924 * block can be written, but unsatisfied update dependencies are made safe by
2925 * modifying (or replacing) the source memory for the duration of the disk
2926 * write. When the disk write completes, the memory block is again brought
2929 * In-core inode structure reclamation.
2931 * Because there are a finite number of "in-core" inode structures, they are
2932 * reused regularly. By transferring all inode-related dependencies to the
2933 * in-memory inode block and indexing them separately (via "inodedep"s), we
2934 * can allow "in-core" inode structures to be reused at any time and avoid
2935 * any increase in contention.
2937 * Called just before entering the device driver to initiate a new disk I/O.
2938 * The buffer must be locked, thus, no I/O completion operations can occur
2939 * while we are manipulating its associated dependencies.
2942 softdep_disk_io_initiation(bp)
2943 struct buf *bp; /* structure describing disk write to occur */
2945 struct worklist *wk, *nextwk;
2946 struct indirdep *indirdep;
2949 * We only care about write operations. There should never
2950 * be dependencies for reads.
2952 if (bp->b_flags & B_READ)
2953 panic("softdep_disk_io_initiation: read");
2955 * Do any necessary pre-I/O processing.
2957 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
2958 nextwk = LIST_NEXT(wk, wk_list);
2959 switch (wk->wk_type) {
2962 initiate_write_filepage(WK_PAGEDEP(wk), bp);
2966 initiate_write_inodeblock(WK_INODEDEP(wk), bp);
2970 indirdep = WK_INDIRDEP(wk);
2971 if (indirdep->ir_state & GOINGAWAY)
2972 panic("disk_io_initiation: indirdep gone");
2974 * If there are no remaining dependencies, this
2975 * will be writing the real pointers, so the
2976 * dependency can be freed.
2978 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
2979 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
2980 brelse(indirdep->ir_savebp);
2981 /* inline expand WORKLIST_REMOVE(wk); */
2982 wk->wk_state &= ~ONWORKLIST;
2983 LIST_REMOVE(wk, wk_list);
2984 WORKITEM_FREE(indirdep, D_INDIRDEP);
2988 * Replace up-to-date version with safe version.
2990 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
2991 M_INDIRDEP, M_SOFTDEP_FLAGS);
2993 indirdep->ir_state &= ~ATTACHED;
2994 indirdep->ir_state |= UNDONE;
2995 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
2996 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3008 panic("handle_disk_io_initiation: Unexpected type %s",
3009 TYPENAME(wk->wk_type));
3016 * Called from within the procedure above to deal with unsatisfied
3017 * allocation dependencies in a directory. The buffer must be locked,
3018 * thus, no I/O completion operations can occur while we are
3019 * manipulating its associated dependencies.
3022 initiate_write_filepage(pagedep, bp)
3023 struct pagedep *pagedep;
3030 if (pagedep->pd_state & IOSTARTED) {
3032 * This can only happen if there is a driver that does not
3033 * understand chaining. Here biodone will reissue the call
3034 * to strategy for the incomplete buffers.
3036 printf("initiate_write_filepage: already started\n");
3039 pagedep->pd_state |= IOSTARTED;
3041 for (i = 0; i < DAHASHSZ; i++) {
3042 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3043 ep = (struct direct *)
3044 ((char *)bp->b_data + dap->da_offset);
3045 if (ep->d_ino != dap->da_newinum) {
3047 panic("%s: dir inum %d != new %d",
3048 "initiate_write_filepage",
3049 ep->d_ino, dap->da_newinum);
3051 if (dap->da_state & DIRCHG)
3052 ep->d_ino = dap->da_previous->dm_oldinum;
3055 dap->da_state &= ~ATTACHED;
3056 dap->da_state |= UNDONE;
3063 * Called from within the procedure above to deal with unsatisfied
3064 * allocation dependencies in an inodeblock. The buffer must be
3065 * locked, thus, no I/O completion operations can occur while we
3066 * are manipulating its associated dependencies.
3069 initiate_write_inodeblock(inodedep, bp)
3070 struct inodedep *inodedep;
3071 struct buf *bp; /* The inode block */
3073 struct allocdirect *adp, *lastadp;
3076 ufs_lbn_t prevlbn = 0;
3079 if (inodedep->id_state & IOSTARTED)
3080 panic("initiate_write_inodeblock: already started");
3081 inodedep->id_state |= IOSTARTED;
3082 fs = inodedep->id_fs;
3083 dp = (struct dinode *)bp->b_data +
3084 ino_to_fsbo(fs, inodedep->id_ino);
3086 * If the bitmap is not yet written, then the allocated
3087 * inode cannot be written to disk.
3089 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3090 if (inodedep->id_savedino != NULL)
3091 panic("initiate_write_inodeblock: already doing I/O");
3092 MALLOC(inodedep->id_savedino, struct dinode *,
3093 sizeof(struct dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3094 *inodedep->id_savedino = *dp;
3095 bzero((caddr_t)dp, sizeof(struct dinode));
3099 * If no dependencies, then there is nothing to roll back.
3101 inodedep->id_savedsize = dp->di_size;
3102 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3105 * Set the dependencies to busy.
3108 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3109 adp = TAILQ_NEXT(adp, ad_next)) {
3111 if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3113 panic("softdep_write_inodeblock: lbn order");
3115 prevlbn = adp->ad_lbn;
3116 if (adp->ad_lbn < NDADDR &&
3117 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3119 panic("%s: direct pointer #%ld mismatch %d != %d",
3120 "softdep_write_inodeblock", adp->ad_lbn,
3121 dp->di_db[adp->ad_lbn], adp->ad_newblkno);
3123 if (adp->ad_lbn >= NDADDR &&
3124 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3126 panic("%s: indirect pointer #%ld mismatch %d != %d",
3127 "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
3128 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
3130 deplist |= 1 << adp->ad_lbn;
3131 if ((adp->ad_state & ATTACHED) == 0) {
3133 panic("softdep_write_inodeblock: Unknown state 0x%x",
3136 #endif /* DIAGNOSTIC */
3137 adp->ad_state &= ~ATTACHED;
3138 adp->ad_state |= UNDONE;
3141 * The on-disk inode cannot claim to be any larger than the last
3142 * fragment that has been written. Otherwise, the on-disk inode
3143 * might have fragments that were not the last block in the file
3144 * which would corrupt the filesystem.
3146 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3147 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3148 if (adp->ad_lbn >= NDADDR)
3150 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3151 /* keep going until hitting a rollback to a frag */
3152 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3154 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3155 for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3157 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3159 panic("softdep_write_inodeblock: lost dep1");
3161 #endif /* DIAGNOSTIC */
3164 for (i = 0; i < NIADDR; i++) {
3166 if (dp->di_ib[i] != 0 &&
3167 (deplist & ((1 << NDADDR) << i)) == 0) {
3169 panic("softdep_write_inodeblock: lost dep2");
3171 #endif /* DIAGNOSTIC */
3178 * If we have zero'ed out the last allocated block of the file,
3179 * roll back the size to the last currently allocated block.
3180 * We know that this last allocated block is a full-sized as
3181 * we already checked for fragments in the loop above.
3183 if (lastadp != NULL &&
3184 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3185 for (i = lastadp->ad_lbn; i >= 0; i--)
3186 if (dp->di_db[i] != 0)
3188 dp->di_size = (i + 1) * fs->fs_bsize;
3191 * The only dependencies are for indirect blocks.
3193 * The file size for indirect block additions is not guaranteed.
3194 * Such a guarantee would be non-trivial to achieve. The conventional
3195 * synchronous write implementation also does not make this guarantee.
3196 * Fsck should catch and fix discrepancies. Arguably, the file size
3197 * can be over-estimated without destroying integrity when the file
3198 * moves into the indirect blocks (i.e., is large). If we want to
3199 * postpone fsck, we are stuck with this argument.
3201 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3202 dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3207 * This routine is called during the completion interrupt
3208 * service routine for a disk write (from the procedure called
3209 * by the device driver to inform the file system caches of
3210 * a request completion). It should be called early in this
3211 * procedure, before the block is made available to other
3212 * processes or other routines are called.
3215 softdep_disk_write_complete(bp)
3216 struct buf *bp; /* describes the completed disk write */
3218 struct worklist *wk;
3219 struct workhead reattach;
3220 struct newblk *newblk;
3221 struct allocindir *aip;
3222 struct allocdirect *adp;
3223 struct indirdep *indirdep;
3224 struct inodedep *inodedep;
3225 struct bmsafemap *bmsafemap;
3228 if (lk.lkt_held != -1)
3229 panic("softdep_disk_write_complete: lock is held");
3232 LIST_INIT(&reattach);
3233 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3234 WORKLIST_REMOVE(wk);
3235 switch (wk->wk_type) {
3238 if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3239 WORKLIST_INSERT(&reattach, wk);
3243 if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3244 WORKLIST_INSERT(&reattach, wk);
3248 bmsafemap = WK_BMSAFEMAP(wk);
3249 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3250 newblk->nb_state |= DEPCOMPLETE;
3251 newblk->nb_bmsafemap = NULL;
3252 LIST_REMOVE(newblk, nb_deps);
3255 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3256 adp->ad_state |= DEPCOMPLETE;
3258 LIST_REMOVE(adp, ad_deps);
3259 handle_allocdirect_partdone(adp);
3262 LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3263 aip->ai_state |= DEPCOMPLETE;
3265 LIST_REMOVE(aip, ai_deps);
3266 handle_allocindir_partdone(aip);
3269 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3270 inodedep->id_state |= DEPCOMPLETE;
3271 LIST_REMOVE(inodedep, id_deps);
3272 inodedep->id_buf = NULL;
3274 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3278 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3282 adp = WK_ALLOCDIRECT(wk);
3283 adp->ad_state |= COMPLETE;
3284 handle_allocdirect_partdone(adp);
3288 aip = WK_ALLOCINDIR(wk);
3289 aip->ai_state |= COMPLETE;
3290 handle_allocindir_partdone(aip);
3294 indirdep = WK_INDIRDEP(wk);
3295 if (indirdep->ir_state & GOINGAWAY) {
3297 panic("disk_write_complete: indirdep gone");
3299 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3300 FREE(indirdep->ir_saveddata, M_INDIRDEP);
3301 indirdep->ir_saveddata = 0;
3302 indirdep->ir_state &= ~UNDONE;
3303 indirdep->ir_state |= ATTACHED;
3304 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3305 handle_allocindir_partdone(aip);
3306 if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
3308 panic("disk_write_complete: not gone");
3311 WORKLIST_INSERT(&reattach, wk);
3312 if ((bp->b_flags & B_DELWRI) == 0)
3313 stat_indir_blk_ptrs++;
3319 panic("handle_disk_write_complete: Unknown type %s",
3320 TYPENAME(wk->wk_type));
3325 * Reattach any requests that must be redone.
3327 while ((wk = LIST_FIRST(&reattach)) != NULL) {
3328 WORKLIST_REMOVE(wk);
3329 WORKLIST_INSERT(&bp->b_dep, wk);
3332 if (lk.lkt_held != -2)
3333 panic("softdep_disk_write_complete: lock lost");
3339 * Called from within softdep_disk_write_complete above. Note that
3340 * this routine is always called from interrupt level with further
3341 * splbio interrupts blocked.
3344 handle_allocdirect_partdone(adp)
3345 struct allocdirect *adp; /* the completed allocdirect */
3347 struct allocdirect *listadp;
3348 struct inodedep *inodedep;
3351 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3353 if (adp->ad_buf != NULL) {
3355 panic("handle_allocdirect_partdone: dangling dep");
3358 * The on-disk inode cannot claim to be any larger than the last
3359 * fragment that has been written. Otherwise, the on-disk inode
3360 * might have fragments that were not the last block in the file
3361 * which would corrupt the filesystem. Thus, we cannot free any
3362 * allocdirects after one whose ad_oldblkno claims a fragment as
3363 * these blocks must be rolled back to zero before writing the inode.
3364 * We check the currently active set of allocdirects in id_inoupdt.
3366 inodedep = adp->ad_inodedep;
3367 bsize = inodedep->id_fs->fs_bsize;
3368 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) {
3369 /* found our block */
3372 /* continue if ad_oldlbn is not a fragment */
3373 if (listadp->ad_oldsize == 0 ||
3374 listadp->ad_oldsize == bsize)
3376 /* hit a fragment */
3380 * If we have reached the end of the current list without
3381 * finding the just finished dependency, then it must be
3382 * on the future dependency list. Future dependencies cannot
3383 * be freed until they are moved to the current list.
3385 if (listadp == NULL) {
3387 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next)
3388 /* found our block */
3391 if (listadp == NULL) {
3393 panic("handle_allocdirect_partdone: lost dep");
3399 * If we have found the just finished dependency, then free
3400 * it along with anything that follows it that is complete.
3402 for (; adp; adp = listadp) {
3403 listadp = TAILQ_NEXT(adp, ad_next);
3404 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3406 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
3411 * Called from within softdep_disk_write_complete above. Note that
3412 * this routine is always called from interrupt level with further
3413 * splbio interrupts blocked.
3416 handle_allocindir_partdone(aip)
3417 struct allocindir *aip; /* the completed allocindir */
3419 struct indirdep *indirdep;
3421 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
3423 if (aip->ai_buf != NULL) {
3425 panic("handle_allocindir_partdone: dangling dependency");
3427 indirdep = aip->ai_indirdep;
3428 if (indirdep->ir_state & UNDONE) {
3429 LIST_REMOVE(aip, ai_next);
3430 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
3433 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
3435 LIST_REMOVE(aip, ai_next);
3436 if (aip->ai_freefrag != NULL)
3437 add_to_worklist(&aip->ai_freefrag->ff_list);
3438 WORKITEM_FREE(aip, D_ALLOCINDIR);
3442 * Called from within softdep_disk_write_complete above to restore
3443 * in-memory inode block contents to their most up-to-date state. Note
3444 * that this routine is always called from interrupt level with further
3445 * splbio interrupts blocked.
3448 handle_written_inodeblock(inodedep, bp)
3449 struct inodedep *inodedep;
3450 struct buf *bp; /* buffer containing the inode block */
3452 struct worklist *wk, *filefree;
3453 struct allocdirect *adp, *nextadp;
3457 if ((inodedep->id_state & IOSTARTED) == 0) {
3459 panic("handle_written_inodeblock: not started");
3461 inodedep->id_state &= ~IOSTARTED;
3462 inodedep->id_state |= COMPLETE;
3463 dp = (struct dinode *)bp->b_data +
3464 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
3466 * If we had to rollback the inode allocation because of
3467 * bitmaps being incomplete, then simply restore it.
3468 * Keep the block dirty so that it will not be reclaimed until
3469 * all associated dependencies have been cleared and the
3470 * corresponding updates written to disk.
3472 if (inodedep->id_savedino != NULL) {
3473 *dp = *inodedep->id_savedino;
3474 FREE(inodedep->id_savedino, M_INODEDEP);
3475 inodedep->id_savedino = NULL;
3476 if ((bp->b_flags & B_DELWRI) == 0)
3477 stat_inode_bitmap++;
3482 * Roll forward anything that had to be rolled back before
3483 * the inode could be updated.
3486 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
3487 nextadp = TAILQ_NEXT(adp, ad_next);
3488 if (adp->ad_state & ATTACHED) {
3490 panic("handle_written_inodeblock: new entry");
3492 if (adp->ad_lbn < NDADDR) {
3493 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) {
3495 panic("%s: %s #%ld mismatch %d != %d",
3496 "handle_written_inodeblock",
3497 "direct pointer", adp->ad_lbn,
3498 dp->di_db[adp->ad_lbn], adp->ad_oldblkno);
3500 dp->di_db[adp->ad_lbn] = adp->ad_newblkno;
3502 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) {
3504 panic("%s: %s #%ld allocated as %d",
3505 "handle_written_inodeblock",
3506 "indirect pointer", adp->ad_lbn - NDADDR,
3507 dp->di_ib[adp->ad_lbn - NDADDR]);
3509 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno;
3511 adp->ad_state &= ~UNDONE;
3512 adp->ad_state |= ATTACHED;
3515 if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
3516 stat_direct_blk_ptrs++;
3518 * Reset the file size to its most up-to-date value.
3520 if (inodedep->id_savedsize == -1) {
3522 panic("handle_written_inodeblock: bad size");
3524 if (dp->di_size != inodedep->id_savedsize) {
3525 dp->di_size = inodedep->id_savedsize;
3528 inodedep->id_savedsize = -1;
3530 * If there were any rollbacks in the inode block, then it must be
3531 * marked dirty so that its will eventually get written back in
3537 * Process any allocdirects that completed during the update.
3539 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
3540 handle_allocdirect_partdone(adp);
3542 * Process deallocations that were held pending until the
3543 * inode had been written to disk. Freeing of the inode
3544 * is delayed until after all blocks have been freed to
3545 * avoid creation of new <vfsid, inum, lbn> triples
3546 * before the old ones have been deleted.
3549 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
3550 WORKLIST_REMOVE(wk);
3551 switch (wk->wk_type) {
3555 * We defer adding filefree to the worklist until
3556 * all other additions have been made to ensure
3557 * that it will be done after all the old blocks
3560 if (filefree != NULL) {
3562 panic("handle_written_inodeblock: filefree");
3568 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
3572 diradd_inode_written(WK_DIRADD(wk), inodedep);
3578 add_to_worklist(wk);
3583 panic("handle_written_inodeblock: Unknown type %s",
3584 TYPENAME(wk->wk_type));
3588 if (filefree != NULL) {
3589 if (free_inodedep(inodedep) == 0) {
3591 panic("handle_written_inodeblock: live inodedep");
3593 add_to_worklist(filefree);
3598 * If no outstanding dependencies, free it.
3600 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0)
3602 return (hadchanges);
3606 * Process a diradd entry after its dependent inode has been written.
3607 * This routine must be called with splbio interrupts blocked.
3610 diradd_inode_written(dap, inodedep)
3612 struct inodedep *inodedep;
3614 struct pagedep *pagedep;
3616 dap->da_state |= COMPLETE;
3617 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3618 if (dap->da_state & DIRCHG)
3619 pagedep = dap->da_previous->dm_pagedep;
3621 pagedep = dap->da_pagedep;
3622 LIST_REMOVE(dap, da_pdlist);
3623 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3625 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3629 * Handle the completion of a mkdir dependency.
3632 handle_written_mkdir(mkdir, type)
3633 struct mkdir *mkdir;
3637 struct pagedep *pagedep;
3639 if (mkdir->md_state != type) {
3641 panic("handle_written_mkdir: bad type");
3643 dap = mkdir->md_diradd;
3644 dap->da_state &= ~type;
3645 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
3646 dap->da_state |= DEPCOMPLETE;
3647 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3648 if (dap->da_state & DIRCHG)
3649 pagedep = dap->da_previous->dm_pagedep;
3651 pagedep = dap->da_pagedep;
3652 LIST_REMOVE(dap, da_pdlist);
3653 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3655 LIST_REMOVE(mkdir, md_mkdirs);
3656 WORKITEM_FREE(mkdir, D_MKDIR);
3660 * Called from within softdep_disk_write_complete above.
3661 * A write operation was just completed. Removed inodes can
3662 * now be freed and associated block pointers may be committed.
3663 * Note that this routine is always called from interrupt level
3664 * with further splbio interrupts blocked.
3667 handle_written_filepage(pagedep, bp)
3668 struct pagedep *pagedep;
3669 struct buf *bp; /* buffer containing the written page */
3671 struct dirrem *dirrem;
3672 struct diradd *dap, *nextdap;
3676 if ((pagedep->pd_state & IOSTARTED) == 0) {
3678 panic("handle_written_filepage: not started");
3680 pagedep->pd_state &= ~IOSTARTED;
3682 * Process any directory removals that have been committed.
3684 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
3685 LIST_REMOVE(dirrem, dm_next);
3686 dirrem->dm_dirinum = pagedep->pd_ino;
3687 add_to_worklist(&dirrem->dm_list);
3690 * Free any directory additions that have been committed.
3692 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
3695 * Uncommitted directory entries must be restored.
3697 for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
3698 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
3700 nextdap = LIST_NEXT(dap, da_pdlist);
3701 if (dap->da_state & ATTACHED) {
3703 panic("handle_written_filepage: attached");
3705 ep = (struct direct *)
3706 ((char *)bp->b_data + dap->da_offset);
3707 ep->d_ino = dap->da_newinum;
3708 dap->da_state &= ~UNDONE;
3709 dap->da_state |= ATTACHED;
3712 * If the inode referenced by the directory has
3713 * been written out, then the dependency can be
3714 * moved to the pending list.
3716 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3717 LIST_REMOVE(dap, da_pdlist);
3718 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
3724 * If there were any rollbacks in the directory, then it must be
3725 * marked dirty so that its will eventually get written back in
3729 if ((bp->b_flags & B_DELWRI) == 0)
3734 * If no dependencies remain, the pagedep will be freed.
3735 * Otherwise it will remain to update the page before it
3736 * is written back to disk.
3738 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) {
3739 for (i = 0; i < DAHASHSZ; i++)
3740 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
3742 if (i == DAHASHSZ) {
3743 LIST_REMOVE(pagedep, pd_hash);
3744 WORKITEM_FREE(pagedep, D_PAGEDEP);
3752 * Writing back in-core inode structures.
3754 * The file system only accesses an inode's contents when it occupies an
3755 * "in-core" inode structure. These "in-core" structures are separate from
3756 * the page frames used to cache inode blocks. Only the latter are
3757 * transferred to/from the disk. So, when the updated contents of the
3758 * "in-core" inode structure are copied to the corresponding in-memory inode
3759 * block, the dependencies are also transferred. The following procedure is
3760 * called when copying a dirty "in-core" inode to a cached inode block.
3764 * Called when an inode is loaded from disk. If the effective link count
3765 * differed from the actual link count when it was last flushed, then we
3766 * need to ensure that the correct effective link count is put back.
3769 softdep_load_inodeblock(ip)
3770 struct inode *ip; /* the "in_core" copy of the inode */
3772 struct inodedep *inodedep;
3775 * Check for alternate nlink count.
3777 ip->i_effnlink = ip->i_nlink;
3779 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3783 ip->i_effnlink -= inodedep->id_nlinkdelta;
3788 * This routine is called just before the "in-core" inode
3789 * information is to be copied to the in-memory inode block.
3790 * Recall that an inode block contains several inodes. If
3791 * the force flag is set, then the dependencies will be
3792 * cleared so that the update can always be made. Note that
3793 * the buffer is locked when this routine is called, so we
3794 * will never be in the middle of writing the inode block
3798 softdep_update_inodeblock(ip, bp, waitfor)
3799 struct inode *ip; /* the "in_core" copy of the inode */
3800 struct buf *bp; /* the buffer containing the inode block */
3801 int waitfor; /* nonzero => update must be allowed */
3803 struct inodedep *inodedep;
3804 struct worklist *wk;
3808 * If the effective link count is not equal to the actual link
3809 * count, then we must track the difference in an inodedep while
3810 * the inode is (potentially) tossed out of the cache. Otherwise,
3811 * if there is no existing inodedep, then there are no dependencies
3815 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3817 if (ip->i_effnlink != ip->i_nlink)
3818 panic("softdep_update_inodeblock: bad link count");
3821 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
3823 panic("softdep_update_inodeblock: bad delta");
3826 * Changes have been initiated. Anything depending on these
3827 * changes cannot occur until this inode has been written.
3829 inodedep->id_state &= ~COMPLETE;
3830 if ((inodedep->id_state & ONWORKLIST) == 0)
3831 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
3833 * Any new dependencies associated with the incore inode must
3834 * now be moved to the list associated with the buffer holding
3835 * the in-memory copy of the inode. Once merged process any
3836 * allocdirects that are completed by the merger.
3838 merge_inode_lists(inodedep);
3839 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
3840 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
3842 * Now that the inode has been pushed into the buffer, the
3843 * operations dependent on the inode being written to disk
3844 * can be moved to the id_bufwait so that they will be
3845 * processed when the buffer I/O completes.
3847 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
3848 WORKLIST_REMOVE(wk);
3849 WORKLIST_INSERT(&inodedep->id_bufwait, wk);
3852 * Newly allocated inodes cannot be written until the bitmap
3853 * that allocates them have been written (indicated by
3854 * DEPCOMPLETE being set in id_state). If we are doing a
3855 * forced sync (e.g., an fsync on a file), we force the bitmap
3856 * to be written so that the update can be done.
3858 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
3862 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
3865 (error = VOP_BWRITE(inodedep->id_buf->b_vp, inodedep->id_buf)) != 0)
3866 softdep_error("softdep_update_inodeblock: bwrite", error);
3867 if ((inodedep->id_state & DEPCOMPLETE) == 0)
3868 panic("softdep_update_inodeblock: update failed");
3872 * Merge the new inode dependency list (id_newinoupdt) into the old
3873 * inode dependency list (id_inoupdt). This routine must be called
3874 * with splbio interrupts blocked.
3877 merge_inode_lists(inodedep)
3878 struct inodedep *inodedep;
3880 struct allocdirect *listadp, *newadp;
3882 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3883 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) {
3884 if (listadp->ad_lbn < newadp->ad_lbn) {
3885 listadp = TAILQ_NEXT(listadp, ad_next);
3888 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3889 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
3890 if (listadp->ad_lbn == newadp->ad_lbn) {
3891 allocdirect_merge(&inodedep->id_inoupdt, newadp,
3895 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3897 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) {
3898 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3899 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next);
3904 * If we are doing an fsync, then we must ensure that any directory
3905 * entries for the inode have been written after the inode gets to disk.
3909 struct vnode *vp; /* the "in_core" copy of the inode */
3911 struct inodedep *inodedep;
3912 struct pagedep *pagedep;
3913 struct worklist *wk;
3920 struct proc *p = CURPROC; /* XXX */
3921 int error, flushparent;
3928 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
3932 if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
3933 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
3934 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
3935 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) {
3937 panic("softdep_fsync: pending ops");
3939 for (error = 0, flushparent = 0; ; ) {
3940 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
3942 if (wk->wk_type != D_DIRADD) {
3944 panic("softdep_fsync: Unexpected type %s",
3945 TYPENAME(wk->wk_type));
3947 dap = WK_DIRADD(wk);
3949 * Flush our parent if this directory entry
3950 * has a MKDIR_PARENT dependency.
3952 if (dap->da_state & DIRCHG)
3953 pagedep = dap->da_previous->dm_pagedep;
3955 pagedep = dap->da_pagedep;
3956 mnt = pagedep->pd_mnt;
3957 parentino = pagedep->pd_ino;
3958 lbn = pagedep->pd_lbn;
3959 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) {
3961 panic("softdep_fsync: dirty");
3963 flushparent = dap->da_state & MKDIR_PARENT;
3965 * If we are being fsync'ed as part of vgone'ing this vnode,
3966 * then we will not be able to release and recover the
3967 * vnode below, so we just have to give up on writing its
3968 * directory entry out. It will eventually be written, just
3969 * not now, but then the user was not asking to have it
3970 * written, so we are not breaking any promises.
3972 if (vp->v_flag & VXLOCK)
3975 * We prevent deadlock by always fetching inodes from the
3976 * root, moving down the directory tree. Thus, when fetching
3977 * our parent directory, we must unlock ourselves before
3978 * requesting the lock on our parent. See the comment in
3979 * ufs_lookup for details on possible races.
3982 VOP_UNLOCK(vp, 0, p);
3983 error = VFS_VGET(mnt, parentino, &pvp);
3984 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, p);
3988 if ((error = UFS_UPDATE(pvp, 1)) != 0) {
3994 * Flush directory page containing the inode's name.
3996 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), p->p_ucred,
3999 error = VOP_BWRITE(bp->b_vp, bp);
4004 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
4012 * Flush all the dirty bitmaps associated with the block device
4013 * before flushing the rest of the dirty blocks so as to reduce
4014 * the number of dependencies that will have to be rolled back.
4017 softdep_fsync_mountdev(vp)
4020 struct buf *bp, *nbp;
4021 struct worklist *wk;
4023 if (!vn_isdisk(vp, NULL))
4024 panic("softdep_fsync_mountdev: vnode not a disk");
4026 for (bp = TAILQ_FIRST(&vp->v_dirtyblkhd); bp; bp = nbp) {
4027 nbp = TAILQ_NEXT(bp, b_vnbufs);
4029 * If it is already scheduled, skip to the next buffer.
4031 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
4033 if ((bp->b_flags & B_DELWRI) == 0) {
4035 panic("softdep_fsync_mountdev: not dirty");
4038 * We are only interested in bitmaps with outstanding
4041 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
4042 wk->wk_type != D_BMSAFEMAP ||
4043 (bp->b_xflags & BX_BKGRDINPROG)) {
4052 * Since we may have slept during the I/O, we need
4053 * to start from a known point.
4055 nbp = TAILQ_FIRST(&vp->v_dirtyblkhd);
4057 drain_output(vp, 1);
4062 * This routine is called when we are trying to synchronously flush a
4063 * file. This routine must eliminate any filesystem metadata dependencies
4064 * so that the syncing routine can succeed by pushing the dirty blocks
4065 * associated with the file. If any I/O errors occur, they are returned.
4068 softdep_sync_metadata(ap)
4069 struct vop_fsync_args /* {
4071 struct ucred *a_cred;
4076 struct vnode *vp = ap->a_vp;
4077 struct pagedep *pagedep;
4078 struct allocdirect *adp;
4079 struct allocindir *aip;
4080 struct buf *bp, *nbp;
4081 struct worklist *wk;
4082 int i, error, waitfor;
4085 * Check whether this vnode is involved in a filesystem
4086 * that is doing soft dependency processing.
4088 if (!vn_isdisk(vp, NULL)) {
4089 if (!DOINGSOFTDEP(vp))
4092 if (vp->v_specmountpoint == NULL ||
4093 (vp->v_specmountpoint->mnt_flag & MNT_SOFTDEP) == 0)
4096 * Ensure that any direct block dependencies have been cleared.
4099 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
4104 * For most files, the only metadata dependencies are the
4105 * cylinder group maps that allocate their inode or blocks.
4106 * The block allocation dependencies can be found by traversing
4107 * the dependency lists for any buffers that remain on their
4108 * dirty buffer list. The inode allocation dependency will
4109 * be resolved when the inode is updated with MNT_WAIT.
4110 * This work is done in two passes. The first pass grabs most
4111 * of the buffers and begins asynchronously writing them. The
4112 * only way to wait for these asynchronous writes is to sleep
4113 * on the filesystem vnode which may stay busy for a long time
4114 * if the filesystem is active. So, instead, we make a second
4115 * pass over the dependencies blocking on each write. In the
4116 * usual case we will be blocking against a write that we
4117 * initiated, so when it is done the dependency will have been
4118 * resolved. Thus the second pass is expected to end quickly.
4120 waitfor = MNT_NOWAIT;
4123 * We must wait for any I/O in progress to finish so that
4124 * all potential buffers on the dirty list will be visible.
4126 drain_output(vp, 1);
4127 if (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT) == 0) {
4131 bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
4134 * As we hold the buffer locked, none of its dependencies
4137 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4138 switch (wk->wk_type) {
4141 adp = WK_ALLOCDIRECT(wk);
4142 if (adp->ad_state & DEPCOMPLETE)
4145 if (getdirtybuf(&nbp, waitfor) == 0)
4148 if (waitfor == MNT_NOWAIT) {
4150 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4158 aip = WK_ALLOCINDIR(wk);
4159 if (aip->ai_state & DEPCOMPLETE)
4162 if (getdirtybuf(&nbp, waitfor) == 0)
4165 if (waitfor == MNT_NOWAIT) {
4167 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4177 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
4178 if (aip->ai_state & DEPCOMPLETE)
4181 if (getdirtybuf(&nbp, MNT_WAIT) == 0)
4184 if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4194 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
4195 WK_INODEDEP(wk)->id_ino)) != 0) {
4204 * We are trying to sync a directory that may
4205 * have dependencies on both its own metadata
4206 * and/or dependencies on the inodes of any
4207 * recently allocated files. We walk its diradd
4208 * lists pushing out the associated inode.
4210 pagedep = WK_PAGEDEP(wk);
4211 for (i = 0; i < DAHASHSZ; i++) {
4212 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
4215 flush_pagedep_deps(vp, pagedep->pd_mnt,
4216 &pagedep->pd_diraddhd[i]))) {
4226 * This case should never happen if the vnode has
4227 * been properly sync'ed. However, if this function
4228 * is used at a place where the vnode has not yet
4229 * been sync'ed, this dependency can show up. So,
4230 * rather than panic, just flush it.
4232 nbp = WK_MKDIR(wk)->md_buf;
4233 if (getdirtybuf(&nbp, waitfor) == 0)
4236 if (waitfor == MNT_NOWAIT) {
4238 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4247 * This case should never happen if the vnode has
4248 * been properly sync'ed. However, if this function
4249 * is used at a place where the vnode has not yet
4250 * been sync'ed, this dependency can show up. So,
4251 * rather than panic, just flush it.
4253 nbp = WK_BMSAFEMAP(wk)->sm_buf;
4254 if (getdirtybuf(&nbp, waitfor) == 0)
4257 if (waitfor == MNT_NOWAIT) {
4259 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4268 panic("softdep_sync_metadata: Unknown type %s",
4269 TYPENAME(wk->wk_type));
4273 (void) getdirtybuf(&TAILQ_NEXT(bp, b_vnbufs), MNT_WAIT);
4274 nbp = TAILQ_NEXT(bp, b_vnbufs);
4283 * The brief unlock is to allow any pent up dependency
4284 * processing to be done. Then proceed with the second pass.
4286 if (waitfor == MNT_NOWAIT) {
4294 * If we have managed to get rid of all the dirty buffers,
4295 * then we are done. For certain directories and block
4296 * devices, we may need to do further work.
4298 if (TAILQ_FIRST(&vp->v_dirtyblkhd) == NULL) {
4305 * If we are trying to sync a block device, some of its buffers may
4306 * contain metadata that cannot be written until the contents of some
4307 * partially written files have been written to disk. The only easy
4308 * way to accomplish this is to sync the entire filesystem (luckily
4309 * this happens rarely).
4311 * We must wait for any I/O in progress to finish so that
4312 * all potential buffers on the dirty list will be visible.
4314 drain_output(vp, 1);
4315 if (vn_isdisk(vp, NULL) &&
4316 vp->v_specmountpoint && !VOP_ISLOCKED(vp, NULL) &&
4317 (error = VFS_SYNC(vp->v_specmountpoint, MNT_WAIT, ap->a_cred,
4324 * Flush the dependencies associated with an inodedep.
4325 * Called with splbio blocked.
4328 flush_inodedep_deps(fs, ino)
4332 struct inodedep *inodedep;
4333 struct allocdirect *adp;
4338 * This work is done in two passes. The first pass grabs most
4339 * of the buffers and begins asynchronously writing them. The
4340 * only way to wait for these asynchronous writes is to sleep
4341 * on the filesystem vnode which may stay busy for a long time
4342 * if the filesystem is active. So, instead, we make a second
4343 * pass over the dependencies blocking on each write. In the
4344 * usual case we will be blocking against a write that we
4345 * initiated, so when it is done the dependency will have been
4346 * resolved. Thus the second pass is expected to end quickly.
4347 * We give a brief window at the top of the loop to allow
4348 * any pending I/O to complete.
4350 for (waitfor = MNT_NOWAIT; ; ) {
4353 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4355 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) {
4356 if (adp->ad_state & DEPCOMPLETE)
4359 if (getdirtybuf(&bp, waitfor) == 0) {
4360 if (waitfor == MNT_NOWAIT)
4365 if (waitfor == MNT_NOWAIT) {
4367 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) {
4376 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) {
4377 if (adp->ad_state & DEPCOMPLETE)
4380 if (getdirtybuf(&bp, waitfor) == 0) {
4381 if (waitfor == MNT_NOWAIT)
4386 if (waitfor == MNT_NOWAIT) {
4388 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) {
4398 * If pass2, we are done, otherwise do pass 2.
4400 if (waitfor == MNT_WAIT)
4405 * Try freeing inodedep in case all dependencies have been removed.
4407 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
4408 (void) free_inodedep(inodedep);
4413 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
4414 * Called with splbio blocked.
4417 flush_pagedep_deps(pvp, mp, diraddhdp)
4420 struct diraddhd *diraddhdp;
4422 struct proc *p = CURPROC; /* XXX */
4423 struct inodedep *inodedep;
4424 struct ufsmount *ump;
4427 int gotit, error = 0;
4432 while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
4434 * Flush ourselves if this directory entry
4435 * has a MKDIR_PARENT dependency.
4437 if (dap->da_state & MKDIR_PARENT) {
4439 if ((error = UFS_UPDATE(pvp, 1)) != 0)
4443 * If that cleared dependencies, go on to next.
4445 if (dap != LIST_FIRST(diraddhdp))
4447 if (dap->da_state & MKDIR_PARENT) {
4449 panic("flush_pagedep_deps: MKDIR_PARENT");
4453 * A newly allocated directory must have its "." and
4454 * ".." entries written out before its name can be
4455 * committed in its parent. We do not want or need
4456 * the full semantics of a synchronous VOP_FSYNC as
4457 * that may end up here again, once for each directory
4458 * level in the filesystem. Instead, we push the blocks
4459 * and wait for them to clear. We have to fsync twice
4460 * because the first call may choose to defer blocks
4461 * that still have dependencies, but deferral will
4462 * happen at most once.
4464 inum = dap->da_newinum;
4465 if (dap->da_state & MKDIR_BODY) {
4467 if ((error = VFS_VGET(mp, inum, &vp)) != 0)
4469 if ((error=VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)) ||
4470 (error=VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p))) {
4474 drain_output(vp, 0);
4478 * If that cleared dependencies, go on to next.
4480 if (dap != LIST_FIRST(diraddhdp))
4482 if (dap->da_state & MKDIR_BODY) {
4484 panic("flush_pagedep_deps: MKDIR_BODY");
4488 * Flush the inode on which the directory entry depends.
4489 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
4490 * the only remaining dependency is that the updated inode
4491 * count must get pushed to disk. The inode has already
4492 * been pushed into its inode buffer (via VOP_UPDATE) at
4493 * the time of the reference count change. So we need only
4494 * locate that buffer, ensure that there will be no rollback
4495 * caused by a bitmap dependency, then write the inode buffer.
4497 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) {
4499 panic("flush_pagedep_deps: lost inode");
4502 * If the inode still has bitmap dependencies,
4503 * push them to disk.
4505 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4506 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4509 (error = VOP_BWRITE(inodedep->id_buf->b_vp,
4510 inodedep->id_buf)) != 0)
4513 if (dap != LIST_FIRST(diraddhdp))
4517 * If the inode is still sitting in a buffer waiting
4518 * to be written, push it to disk.
4521 if ((error = bread(ump->um_devvp,
4522 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
4523 (int)ump->um_fs->fs_bsize, NOCRED, &bp)) != 0)
4525 if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0)
4529 * If we have failed to get rid of all the dependencies
4530 * then something is seriously wrong.
4532 if (dap == LIST_FIRST(diraddhdp)) {
4534 panic("flush_pagedep_deps: flush failed");
4543 * A large burst of file addition or deletion activity can drive the
4544 * memory load excessively high. First attempt to slow things down
4545 * using the techniques below. If that fails, this routine requests
4546 * the offending operations to fall back to running synchronously
4547 * until the memory load returns to a reasonable level.
4550 softdep_slowdown(vp)
4553 int max_softdeps_hard;
4555 max_softdeps_hard = max_softdeps * 11 / 10;
4556 if (num_dirrem < max_softdeps_hard / 2 &&
4557 num_inodedep < max_softdeps_hard)
4559 stat_sync_limit_hit += 1;
4564 * If memory utilization has gotten too high, deliberately slow things
4565 * down and speed up the I/O processing.
4568 request_cleanup(resource, islocked)
4572 struct proc *p = CURPROC;
4575 * We never hold up the filesystem syncer process.
4577 if (p == filesys_syncer)
4580 * First check to see if the work list has gotten backlogged.
4581 * If it has, co-opt this process to help clean up two entries.
4582 * Because this process may hold inodes locked, we cannot
4583 * handle any remove requests that might block on a locked
4584 * inode as that could lead to deadlock.
4586 if (num_on_worklist > max_softdeps / 10) {
4589 process_worklist_item(NULL, LK_NOWAIT);
4590 process_worklist_item(NULL, LK_NOWAIT);
4591 stat_worklist_push += 2;
4598 * If we are resource constrained on inode dependencies, try
4599 * flushing some dirty inodes. Otherwise, we are constrained
4600 * by file deletions, so try accelerating flushes of directories
4601 * with removal dependencies. We would like to do the cleanup
4602 * here, but we probably hold an inode locked at this point and
4603 * that might deadlock against one that we try to clean. So,
4604 * the best that we can do is request the syncer daemon to do
4605 * the cleanup for us.
4610 stat_ino_limit_push += 1;
4611 req_clear_inodedeps += 1;
4612 stat_countp = &stat_ino_limit_hit;
4616 stat_blk_limit_push += 1;
4617 req_clear_remove += 1;
4618 stat_countp = &stat_blk_limit_hit;
4624 panic("request_cleanup: unknown type");
4627 * Hopefully the syncer daemon will catch up and awaken us.
4628 * We wait at most tickdelay before proceeding in any case.
4633 if (handle.callout == NULL)
4634 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
4635 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, PPAUSE,
4644 * Awaken processes pausing in request_cleanup and clear proc_waiting
4645 * to indicate that there is no longer a timer running.
4653 wakeup_one(&proc_waiting);
4654 if (proc_waiting > 0)
4655 handle = timeout(pause_timer, 0, tickdelay > 2 ? tickdelay : 2);
4657 handle.callout = NULL;
4661 * Flush out a directory with at least one removal dependency in an effort to
4662 * reduce the number of dirrem, freefile, and freeblks dependency structures.
4668 struct pagedep_hashhead *pagedephd;
4669 struct pagedep *pagedep;
4670 static int next = 0;
4677 for (cnt = 0; cnt < pagedep_hash; cnt++) {
4678 pagedephd = &pagedep_hashtbl[next++];
4679 if (next >= pagedep_hash)
4681 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
4682 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
4684 mp = pagedep->pd_mnt;
4685 ino = pagedep->pd_ino;
4687 if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4688 softdep_error("clear_remove: vget", error);
4691 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)))
4692 softdep_error("clear_remove: fsync", error);
4693 drain_output(vp, 0);
4702 * Clear out a block of dirty inodes in an effort to reduce
4703 * the number of inodedep dependency structures.
4709 struct inodedep_hashhead *inodedephd;
4710 struct inodedep *inodedep;
4711 static int next = 0;
4716 ino_t firstino, lastino, ino;
4720 * Pick a random inode dependency to be cleared.
4721 * We will then gather up all the inodes in its block
4722 * that have dependencies and flush them out.
4724 for (cnt = 0; cnt < inodedep_hash; cnt++) {
4725 inodedephd = &inodedep_hashtbl[next++];
4726 if (next >= inodedep_hash)
4728 if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
4731 if (inodedep == NULL)
4734 * Ugly code to find mount point given pointer to superblock.
4736 fs = inodedep->id_fs;
4737 TAILQ_FOREACH(mp, &mountlist, mnt_list)
4738 if ((mp->mnt_flag & MNT_SOFTDEP) && fs == VFSTOUFS(mp)->um_fs)
4741 * Find the last inode in the block with dependencies.
4743 firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
4744 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
4745 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
4748 * Asynchronously push all but the last inode with dependencies.
4749 * Synchronously push the last inode with dependencies to ensure
4750 * that the inode block gets written to free up the inodedeps.
4752 for (ino = firstino; ino <= lastino; ino++) {
4753 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4756 if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4757 softdep_error("clear_inodedeps: vget", error);
4760 if (ino == lastino) {
4761 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_WAIT, p)))
4762 softdep_error("clear_inodedeps: fsync1", error);
4764 if ((error = VOP_FSYNC(vp, p->p_ucred, MNT_NOWAIT, p)))
4765 softdep_error("clear_inodedeps: fsync2", error);
4766 drain_output(vp, 0);
4775 * Function to determine if the buffer has outstanding dependencies
4776 * that will cause a roll-back if the buffer is written. If wantcount
4777 * is set, return number of dependencies, otherwise just yes or no.
4780 softdep_count_dependencies(bp, wantcount)
4784 struct worklist *wk;
4785 struct inodedep *inodedep;
4786 struct indirdep *indirdep;
4787 struct allocindir *aip;
4788 struct pagedep *pagedep;
4794 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4795 switch (wk->wk_type) {
4798 inodedep = WK_INODEDEP(wk);
4799 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4800 /* bitmap allocation dependency */
4805 if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
4806 /* direct block pointer dependency */
4814 indirdep = WK_INDIRDEP(wk);
4816 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
4817 /* indirect block pointer dependency */
4825 pagedep = WK_PAGEDEP(wk);
4826 for (i = 0; i < DAHASHSZ; i++) {
4828 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
4829 /* directory entry dependency */
4841 /* never a dependency on these blocks */
4846 panic("softdep_check_for_rollback: Unexpected type %s",
4847 TYPENAME(wk->wk_type));
4857 * Acquire exclusive access to a buffer.
4858 * Must be called with splbio blocked.
4859 * Return 1 if buffer was acquired.
4862 getdirtybuf(bpp, waitfor)
4870 if ((bp = *bpp) == NULL)
4872 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
4873 if ((bp->b_xflags & BX_BKGRDINPROG) == 0)
4876 if (waitfor != MNT_WAIT)
4878 bp->b_xflags |= BX_BKGRDWAIT;
4879 interlocked_sleep(&lk, SLEEP, &bp->b_xflags, PRIBIO,
4883 if (waitfor != MNT_WAIT)
4885 error = interlocked_sleep(&lk, LOCKBUF, bp,
4886 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0);
4887 if (error != ENOLCK) {
4889 panic("getdirtybuf: inconsistent lock");
4892 if ((bp->b_flags & B_DELWRI) == 0) {
4901 * Wait for pending output on a vnode to complete.
4902 * Must be called with vnode locked.
4905 drain_output(vp, islocked)
4912 while (vp->v_numoutput) {
4913 vp->v_flag |= VBWAIT;
4914 interlocked_sleep(&lk, SLEEP, (caddr_t)&vp->v_numoutput,
4915 PRIBIO + 1, "drainvp", 0);
4922 * Called whenever a buffer that is being invalidated or reallocated
4923 * contains dependencies. This should only happen if an I/O error has
4924 * occurred. The routine is called with the buffer locked.
4927 softdep_deallocate_dependencies(bp)
4931 if ((bp->b_flags & B_ERROR) == 0)
4932 panic("softdep_deallocate_dependencies: dangling deps");
4933 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntonname, bp->b_error);
4934 panic("softdep_deallocate_dependencies: unrecovered I/O error");
4938 * Function to handle asynchronous write errors in the filesystem.
4941 softdep_error(func, error)
4946 /* XXX should do something better! */
4947 printf("%s: got error %d while accessing filesystem\n", func, error);