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 1.27 2005/07/31 22:25:46 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>
70 #include "ffs_extern.h"
71 #include "ufs_extern.h"
73 #include <sys/thread2.h>
76 * These definitions need to be adapted to the system to which
77 * this file is being ported.
80 * malloc types defined for the softdep system.
82 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
83 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
84 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
85 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
86 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
87 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
88 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
89 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
90 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
91 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
92 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
93 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
94 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
96 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE)
101 #define D_BMSAFEMAP 3
102 #define D_ALLOCDIRECT 4
104 #define D_ALLOCINDIR 6
111 #define D_LAST D_DIRREM
114 * translate from workitem type to memory type
115 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
117 static struct malloc_type *memtype[] = {
133 #define DtoM(type) (memtype[type])
136 * Names of malloc types.
138 #define TYPENAME(type) \
139 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
141 * End system adaptaion definitions.
145 * Internal function prototypes.
147 static void softdep_error(char *, int);
148 static void drain_output(struct vnode *, int);
149 static int getdirtybuf(struct buf **, int);
150 static void clear_remove(struct thread *);
151 static void clear_inodedeps(struct thread *);
152 static int flush_pagedep_deps(struct vnode *, struct mount *,
154 static int flush_inodedep_deps(struct fs *, ino_t);
155 static int handle_written_filepage(struct pagedep *, struct buf *);
156 static void diradd_inode_written(struct diradd *, struct inodedep *);
157 static int handle_written_inodeblock(struct inodedep *, struct buf *);
158 static void handle_allocdirect_partdone(struct allocdirect *);
159 static void handle_allocindir_partdone(struct allocindir *);
160 static void initiate_write_filepage(struct pagedep *, struct buf *);
161 static void handle_written_mkdir(struct mkdir *, int);
162 static void initiate_write_inodeblock(struct inodedep *, struct buf *);
163 static void handle_workitem_freefile(struct freefile *);
164 static void handle_workitem_remove(struct dirrem *);
165 static struct dirrem *newdirrem(struct buf *, struct inode *,
166 struct inode *, int, struct dirrem **);
167 static void free_diradd(struct diradd *);
168 static void free_allocindir(struct allocindir *, struct inodedep *);
169 static int indir_trunc (struct inode *, ufs_daddr_t, int, ufs_lbn_t,
171 static void deallocate_dependencies(struct buf *, struct inodedep *);
172 static void free_allocdirect(struct allocdirectlst *,
173 struct allocdirect *, int);
174 static int check_inode_unwritten(struct inodedep *);
175 static int free_inodedep(struct inodedep *);
176 static void handle_workitem_freeblocks(struct freeblks *);
177 static void merge_inode_lists(struct inodedep *);
178 static void setup_allocindir_phase2(struct buf *, struct inode *,
179 struct allocindir *);
180 static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t,
182 static void handle_workitem_freefrag(struct freefrag *);
183 static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long);
184 static void allocdirect_merge(struct allocdirectlst *,
185 struct allocdirect *, struct allocdirect *);
186 static struct bmsafemap *bmsafemap_lookup(struct buf *);
187 static int newblk_lookup(struct fs *, ufs_daddr_t, int,
189 static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
190 static int pagedep_lookup(struct inode *, ufs_lbn_t, int,
192 static void pause_timer(void *);
193 static int request_cleanup(int, int);
194 static int process_worklist_item(struct mount *, int);
195 static void add_to_worklist(struct worklist *);
198 * Exported softdep operations.
200 static void softdep_disk_io_initiation(struct buf *);
201 static void softdep_disk_write_complete(struct buf *);
202 static void softdep_deallocate_dependencies(struct buf *);
203 static int softdep_fsync(struct vnode *);
204 static int softdep_process_worklist(struct mount *);
205 static void softdep_move_dependencies(struct buf *, struct buf *);
206 static int softdep_count_dependencies(struct buf *bp, int);
208 static struct bio_ops softdep_bioops = {
209 softdep_disk_io_initiation, /* io_start */
210 softdep_disk_write_complete, /* io_complete */
211 softdep_deallocate_dependencies, /* io_deallocate */
212 softdep_fsync, /* io_fsync */
213 softdep_process_worklist, /* io_sync */
214 softdep_move_dependencies, /* io_movedeps */
215 softdep_count_dependencies, /* io_countdeps */
219 * Locking primitives.
221 * For a uniprocessor, all we need to do is protect against disk
222 * interrupts. For a multiprocessor, this lock would have to be
223 * a mutex. A single mutex is used throughout this file, though
224 * finer grain locking could be used if contention warranted it.
226 * For a multiprocessor, the sleep call would accept a lock and
227 * release it after the sleep processing was complete. In a uniprocessor
228 * implementation there is no such interlock, so we simple mark
229 * the places where it needs to be done with the `interlocked' form
230 * of the lock calls. Since the uniprocessor sleep already interlocks
231 * the spl, there is nothing that really needs to be done.
233 #ifndef /* NOT */ DEBUG
234 static struct lockit {
236 #define ACQUIRE_LOCK(lk) crit_enter_id("softupdates");
237 #define FREE_LOCK(lk) crit_exit_id("softupdates");
240 #define NOHOLDER ((struct thread *)-1)
241 #define SPECIAL_FLAG ((struct thread *)-2)
242 static struct lockit {
244 struct thread *lkt_held;
245 } lk = { 0, NOHOLDER };
248 static void acquire_lock(struct lockit *);
249 static void free_lock(struct lockit *);
250 void softdep_panic(char *);
252 #define ACQUIRE_LOCK(lk) acquire_lock(lk)
253 #define FREE_LOCK(lk) free_lock(lk)
261 if (lk->lkt_held != NOHOLDER) {
262 holder = lk->lkt_held;
264 if (holder == curthread)
265 panic("softdep_lock: locking against myself");
267 panic("softdep_lock: lock held by %p", holder);
269 crit_enter_id("softupdates");
270 lk->lkt_held = curthread;
279 if (lk->lkt_held == NOHOLDER)
280 panic("softdep_unlock: lock not held");
281 lk->lkt_held = NOHOLDER;
282 crit_exit_id("softupdates");
286 * Function to release soft updates lock and panic.
293 if (lk.lkt_held != NOHOLDER)
299 static int interlocked_sleep(struct lockit *, int, void *, int,
303 * When going to sleep, we must save our SPL so that it does
304 * not get lost if some other process uses the lock while we
305 * are sleeping. We restore it after we have slept. This routine
306 * wraps the interlocking with functions that sleep. The list
307 * below enumerates the available set of operations.
314 interlocked_sleep(lk, op, ident, flags, wmesg, timo)
327 if (lk->lkt_held == NOHOLDER)
328 panic("interlocked_sleep: lock not held");
329 lk->lkt_held = NOHOLDER;
333 retval = tsleep(ident, flags, wmesg, timo);
336 retval = BUF_LOCK((struct buf *)ident, flags);
339 panic("interlocked_sleep: unknown operation");
342 if (lk->lkt_held != NOHOLDER) {
343 holder = lk->lkt_held;
345 if (holder == curthread)
346 panic("interlocked_sleep: locking against self");
348 panic("interlocked_sleep: lock held by %p", holder);
350 lk->lkt_held = curthread;
358 * Place holder for real semaphores.
367 static void sema_init(struct sema *, char *, int, int);
368 static int sema_get(struct sema *, struct lockit *);
369 static void sema_release(struct sema *);
372 sema_init(semap, name, prio, timo)
378 semap->holder = NOHOLDER;
386 sema_get(semap, interlock)
388 struct lockit *interlock;
391 if (semap->value++ > 0) {
392 if (interlock != NULL) {
393 interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
394 semap->prio, semap->name, semap->timo);
395 FREE_LOCK(interlock);
397 tsleep((caddr_t)semap, semap->prio, semap->name,
402 semap->holder = curthread;
403 if (interlock != NULL)
404 FREE_LOCK(interlock);
413 if (semap->value <= 0 || semap->holder != curthread) {
414 if (lk.lkt_held != NOHOLDER)
416 panic("sema_release: not held");
418 if (--semap->value > 0) {
422 semap->holder = NOHOLDER;
426 * Worklist queue management.
427 * These routines require that the lock be held.
429 #ifndef /* NOT */ DEBUG
430 #define WORKLIST_INSERT(head, item) do { \
431 (item)->wk_state |= ONWORKLIST; \
432 LIST_INSERT_HEAD(head, item, wk_list); \
434 #define WORKLIST_REMOVE(item) do { \
435 (item)->wk_state &= ~ONWORKLIST; \
436 LIST_REMOVE(item, wk_list); \
438 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
441 static void worklist_insert(struct workhead *, struct worklist *);
442 static void worklist_remove(struct worklist *);
443 static void workitem_free(struct worklist *, int);
445 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
446 #define WORKLIST_REMOVE(item) worklist_remove(item)
447 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
450 worklist_insert(head, item)
451 struct workhead *head;
452 struct worklist *item;
455 if (lk.lkt_held == NOHOLDER)
456 panic("worklist_insert: lock not held");
457 if (item->wk_state & ONWORKLIST) {
459 panic("worklist_insert: already on list");
461 item->wk_state |= ONWORKLIST;
462 LIST_INSERT_HEAD(head, item, wk_list);
466 worklist_remove(item)
467 struct worklist *item;
470 if (lk.lkt_held == NOHOLDER)
471 panic("worklist_remove: lock not held");
472 if ((item->wk_state & ONWORKLIST) == 0) {
474 panic("worklist_remove: not on list");
476 item->wk_state &= ~ONWORKLIST;
477 LIST_REMOVE(item, wk_list);
481 workitem_free(item, type)
482 struct worklist *item;
486 if (item->wk_state & ONWORKLIST) {
487 if (lk.lkt_held != NOHOLDER)
489 panic("workitem_free: still on list");
491 if (item->wk_type != type) {
492 if (lk.lkt_held != NOHOLDER)
494 panic("workitem_free: type mismatch");
496 FREE(item, DtoM(type));
501 * Workitem queue management
503 static struct workhead softdep_workitem_pending;
504 static int num_on_worklist; /* number of worklist items to be processed */
505 static int softdep_worklist_busy; /* 1 => trying to do unmount */
506 static int softdep_worklist_req; /* serialized waiters */
507 static int max_softdeps; /* maximum number of structs before slowdown */
508 static int tickdelay = 2; /* number of ticks to pause during slowdown */
509 static int *stat_countp; /* statistic to count in proc_waiting timeout */
510 static int proc_waiting; /* tracks whether we have a timeout posted */
511 static struct callout handle; /* handle on posted proc_waiting timeout */
512 static struct thread *filesys_syncer; /* proc of filesystem syncer process */
513 static int req_clear_inodedeps; /* syncer process flush some inodedeps */
514 #define FLUSH_INODES 1
515 static int req_clear_remove; /* syncer process flush some freeblks */
516 #define FLUSH_REMOVE 2
520 static int stat_worklist_push; /* number of worklist cleanups */
521 static int stat_blk_limit_push; /* number of times block limit neared */
522 static int stat_ino_limit_push; /* number of times inode limit neared */
523 static int stat_blk_limit_hit; /* number of times block slowdown imposed */
524 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */
525 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */
526 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */
527 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */
528 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
529 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */
532 #include <sys/sysctl.h>
533 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
534 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
535 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
536 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
537 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
538 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
539 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
540 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
541 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
542 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
543 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
544 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
548 * Add an item to the end of the work queue.
549 * This routine requires that the lock be held.
550 * This is the only routine that adds items to the list.
551 * The following routine is the only one that removes items
552 * and does so in order from first to last.
558 static struct worklist *worklist_tail;
560 if (wk->wk_state & ONWORKLIST) {
561 if (lk.lkt_held != NOHOLDER)
563 panic("add_to_worklist: already on list");
565 wk->wk_state |= ONWORKLIST;
566 if (LIST_FIRST(&softdep_workitem_pending) == NULL)
567 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
569 LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
571 num_on_worklist += 1;
575 * Process that runs once per second to handle items in the background queue.
577 * Note that we ensure that everything is done in the order in which they
578 * appear in the queue. The code below depends on this property to ensure
579 * that blocks of a file are freed before the inode itself is freed. This
580 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
581 * until all the old ones have been purged from the dependency lists.
584 softdep_process_worklist(matchmnt)
585 struct mount *matchmnt;
587 thread_t td = curthread;
588 int matchcnt, loopcount;
592 * Record the process identifier of our caller so that we can give
593 * this process preferential treatment in request_cleanup below.
599 * There is no danger of having multiple processes run this
600 * code, but we have to single-thread it when softdep_flushfiles()
601 * is in operation to get an accurate count of the number of items
602 * related to its mount point that are in the list.
604 if (matchmnt == NULL) {
605 if (softdep_worklist_busy < 0)
607 softdep_worklist_busy += 1;
611 * If requested, try removing inode or removal dependencies.
613 if (req_clear_inodedeps) {
615 req_clear_inodedeps -= 1;
616 wakeup_one(&proc_waiting);
618 if (req_clear_remove) {
620 req_clear_remove -= 1;
621 wakeup_one(&proc_waiting);
624 starttime = time_second;
625 while (num_on_worklist > 0) {
626 matchcnt += process_worklist_item(matchmnt, 0);
629 * If a umount operation wants to run the worklist
632 if (softdep_worklist_req && matchmnt == NULL) {
638 * If requested, try removing inode or removal dependencies.
640 if (req_clear_inodedeps) {
642 req_clear_inodedeps -= 1;
643 wakeup_one(&proc_waiting);
645 if (req_clear_remove) {
647 req_clear_remove -= 1;
648 wakeup_one(&proc_waiting);
651 * We do not generally want to stop for buffer space, but if
652 * we are really being a buffer hog, we will stop and wait.
654 if (loopcount++ % 128 == 0)
657 * Never allow processing to run for more than one
658 * second. Otherwise the other syncer tasks may get
659 * excessively backlogged.
661 if (starttime != time_second && matchmnt == NULL) {
666 if (matchmnt == NULL) {
667 --softdep_worklist_busy;
668 if (softdep_worklist_req && softdep_worklist_busy == 0)
669 wakeup(&softdep_worklist_req);
675 * Process one item on the worklist.
678 process_worklist_item(matchmnt, flags)
679 struct mount *matchmnt;
683 struct dirrem *dirrem;
689 if (matchmnt != NULL)
690 matchfs = VFSTOUFS(matchmnt)->um_fs;
693 * Normally we just process each item on the worklist in order.
694 * However, if we are in a situation where we cannot lock any
695 * inodes, we have to skip over any dirrem requests whose
696 * vnodes are resident and locked.
698 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
699 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
701 dirrem = WK_DIRREM(wk);
702 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev,
704 if (vp == NULL || !VOP_ISLOCKED(vp, curthread))
712 num_on_worklist -= 1;
714 switch (wk->wk_type) {
717 /* removal of a directory entry */
718 if (WK_DIRREM(wk)->dm_mnt == matchmnt)
720 handle_workitem_remove(WK_DIRREM(wk));
724 /* releasing blocks and/or fragments from a file */
725 if (WK_FREEBLKS(wk)->fb_fs == matchfs)
727 handle_workitem_freeblocks(WK_FREEBLKS(wk));
731 /* releasing a fragment when replaced as a file grows */
732 if (WK_FREEFRAG(wk)->ff_fs == matchfs)
734 handle_workitem_freefrag(WK_FREEFRAG(wk));
738 /* releasing an inode when its link count drops to 0 */
739 if (WK_FREEFILE(wk)->fx_fs == matchfs)
741 handle_workitem_freefile(WK_FREEFILE(wk));
745 panic("%s_process_worklist: Unknown type %s",
746 "softdep", TYPENAME(wk->wk_type));
753 * Move dependencies from one buffer to another.
756 softdep_move_dependencies(oldbp, newbp)
760 struct worklist *wk, *wktail;
762 if (LIST_FIRST(&newbp->b_dep) != NULL)
763 panic("softdep_move_dependencies: need merge code");
766 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
767 LIST_REMOVE(wk, wk_list);
769 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
771 LIST_INSERT_AFTER(wktail, wk, wk_list);
778 * Purge the work list of all items associated with a particular mount point.
781 softdep_flushfiles(struct mount *oldmnt, int flags, struct thread *td)
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, 0, "softflush", 0);
792 softdep_worklist_req -= 1;
794 softdep_worklist_busy = -1;
796 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 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, td)) != 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, td);
825 error = VOP_FSYNC(devvp, MNT_WAIT, td);
826 VOP_UNLOCK(devvp, 0, td);
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 == NOHOLDER)
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 == NOHOLDER)
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 filesystems.
1078 softdep_initialize()
1080 callout_init(&handle);
1081 bioops = softdep_bioops; /* XXX hack */
1083 LIST_INIT(&mkdirlisthd);
1084 LIST_INIT(&softdep_workitem_pending);
1085 max_softdeps = min(desiredvnodes * 8,
1086 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep)));
1087 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1089 sema_init(&pagedep_in_progress, "pagedep", 0, 0);
1090 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1091 sema_init(&inodedep_in_progress, "inodedep", 0, 0);
1092 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1093 sema_init(&newblk_in_progress, "newblk", 0, 0);
1097 * Called at mount time to notify the dependency code that a
1098 * filesystem wishes to use it.
1101 softdep_mount(devvp, mp, fs)
1102 struct vnode *devvp;
1106 struct csum cstotal;
1111 mp->mnt_flag &= ~MNT_ASYNC;
1112 mp->mnt_flag |= MNT_SOFTDEP;
1114 * When doing soft updates, the counters in the
1115 * superblock may have gotten out of sync, so we have
1116 * to scan the cylinder groups and recalculate them.
1118 if (fs->fs_clean != 0)
1120 bzero(&cstotal, sizeof cstotal);
1121 for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1122 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
1123 fs->fs_cgsize, &bp)) != 0) {
1127 cgp = (struct cg *)bp->b_data;
1128 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1129 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1130 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1131 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1132 fs->fs_cs(fs, cyl) = cgp->cg_cs;
1136 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1137 printf("ffs_mountfs: superblock updated for soft updates\n");
1139 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1144 * Protecting the freemaps (or bitmaps).
1146 * To eliminate the need to execute fsck before mounting a filesystem
1147 * after a power failure, one must (conservatively) guarantee that the
1148 * on-disk copy of the bitmaps never indicate that a live inode or block is
1149 * free. So, when a block or inode is allocated, the bitmap should be
1150 * updated (on disk) before any new pointers. When a block or inode is
1151 * freed, the bitmap should not be updated until all pointers have been
1152 * reset. The latter dependency is handled by the delayed de-allocation
1153 * approach described below for block and inode de-allocation. The former
1154 * dependency is handled by calling the following procedure when a block or
1155 * inode is allocated. When an inode is allocated an "inodedep" is created
1156 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1157 * Each "inodedep" is also inserted into the hash indexing structure so
1158 * that any additional link additions can be made dependent on the inode
1161 * The ufs filesystem maintains a number of free block counts (e.g., per
1162 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1163 * in addition to the bitmaps. These counts are used to improve efficiency
1164 * during allocation and therefore must be consistent with the bitmaps.
1165 * There is no convenient way to guarantee post-crash consistency of these
1166 * counts with simple update ordering, for two main reasons: (1) The counts
1167 * and bitmaps for a single cylinder group block are not in the same disk
1168 * sector. If a disk write is interrupted (e.g., by power failure), one may
1169 * be written and the other not. (2) Some of the counts are located in the
1170 * superblock rather than the cylinder group block. So, we focus our soft
1171 * updates implementation on protecting the bitmaps. When mounting a
1172 * filesystem, we recompute the auxiliary counts from the bitmaps.
1176 * Called just after updating the cylinder group block to allocate an inode.
1179 softdep_setup_inomapdep(bp, ip, newinum)
1180 struct buf *bp; /* buffer for cylgroup block with inode map */
1181 struct inode *ip; /* inode related to allocation */
1182 ino_t newinum; /* new inode number being allocated */
1184 struct inodedep *inodedep;
1185 struct bmsafemap *bmsafemap;
1188 * Create a dependency for the newly allocated inode.
1189 * Panic if it already exists as something is seriously wrong.
1190 * Otherwise add it to the dependency list for the buffer holding
1191 * the cylinder group map from which it was allocated.
1194 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1196 panic("softdep_setup_inomapdep: found inode");
1198 inodedep->id_buf = bp;
1199 inodedep->id_state &= ~DEPCOMPLETE;
1200 bmsafemap = bmsafemap_lookup(bp);
1201 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1206 * Called just after updating the cylinder group block to
1207 * allocate block or fragment.
1210 softdep_setup_blkmapdep(bp, fs, newblkno)
1211 struct buf *bp; /* buffer for cylgroup block with block map */
1212 struct fs *fs; /* filesystem doing allocation */
1213 ufs_daddr_t newblkno; /* number of newly allocated block */
1215 struct newblk *newblk;
1216 struct bmsafemap *bmsafemap;
1219 * Create a dependency for the newly allocated block.
1220 * Add it to the dependency list for the buffer holding
1221 * the cylinder group map from which it was allocated.
1223 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1224 panic("softdep_setup_blkmapdep: found block");
1226 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1227 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1232 * Find the bmsafemap associated with a cylinder group buffer.
1233 * If none exists, create one. The buffer must be locked when
1234 * this routine is called and this routine must be called with
1235 * splbio interrupts blocked.
1237 static struct bmsafemap *
1238 bmsafemap_lookup(bp)
1241 struct bmsafemap *bmsafemap;
1242 struct worklist *wk;
1245 if (lk.lkt_held == NOHOLDER)
1246 panic("bmsafemap_lookup: lock not held");
1248 LIST_FOREACH(wk, &bp->b_dep, wk_list)
1249 if (wk->wk_type == D_BMSAFEMAP)
1250 return (WK_BMSAFEMAP(wk));
1252 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1253 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1254 bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1255 bmsafemap->sm_list.wk_state = 0;
1256 bmsafemap->sm_buf = bp;
1257 LIST_INIT(&bmsafemap->sm_allocdirecthd);
1258 LIST_INIT(&bmsafemap->sm_allocindirhd);
1259 LIST_INIT(&bmsafemap->sm_inodedephd);
1260 LIST_INIT(&bmsafemap->sm_newblkhd);
1262 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1267 * Direct block allocation dependencies.
1269 * When a new block is allocated, the corresponding disk locations must be
1270 * initialized (with zeros or new data) before the on-disk inode points to
1271 * them. Also, the freemap from which the block was allocated must be
1272 * updated (on disk) before the inode's pointer. These two dependencies are
1273 * independent of each other and are needed for all file blocks and indirect
1274 * blocks that are pointed to directly by the inode. Just before the
1275 * "in-core" version of the inode is updated with a newly allocated block
1276 * number, a procedure (below) is called to setup allocation dependency
1277 * structures. These structures are removed when the corresponding
1278 * dependencies are satisfied or when the block allocation becomes obsolete
1279 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1280 * fragment that gets upgraded). All of these cases are handled in
1281 * procedures described later.
1283 * When a file extension causes a fragment to be upgraded, either to a larger
1284 * fragment or to a full block, the on-disk location may change (if the
1285 * previous fragment could not simply be extended). In this case, the old
1286 * fragment must be de-allocated, but not until after the inode's pointer has
1287 * been updated. In most cases, this is handled by later procedures, which
1288 * will construct a "freefrag" structure to be added to the workitem queue
1289 * when the inode update is complete (or obsolete). The main exception to
1290 * this is when an allocation occurs while a pending allocation dependency
1291 * (for the same block pointer) remains. This case is handled in the main
1292 * allocation dependency setup procedure by immediately freeing the
1293 * unreferenced fragments.
1296 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1297 struct inode *ip; /* inode to which block is being added */
1298 ufs_lbn_t lbn; /* block pointer within inode */
1299 ufs_daddr_t newblkno; /* disk block number being added */
1300 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */
1301 long newsize; /* size of new block */
1302 long oldsize; /* size of new block */
1303 struct buf *bp; /* bp for allocated block */
1305 struct allocdirect *adp, *oldadp;
1306 struct allocdirectlst *adphead;
1307 struct bmsafemap *bmsafemap;
1308 struct inodedep *inodedep;
1309 struct pagedep *pagedep;
1310 struct newblk *newblk;
1312 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1313 M_ALLOCDIRECT, M_SOFTDEP_FLAGS);
1314 bzero(adp, sizeof(struct allocdirect));
1315 adp->ad_list.wk_type = D_ALLOCDIRECT;
1317 adp->ad_newblkno = newblkno;
1318 adp->ad_oldblkno = oldblkno;
1319 adp->ad_newsize = newsize;
1320 adp->ad_oldsize = oldsize;
1321 adp->ad_state = ATTACHED;
1322 if (newblkno == oldblkno)
1323 adp->ad_freefrag = NULL;
1325 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1327 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1328 panic("softdep_setup_allocdirect: lost block");
1331 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1332 adp->ad_inodedep = inodedep;
1334 if (newblk->nb_state == DEPCOMPLETE) {
1335 adp->ad_state |= DEPCOMPLETE;
1338 bmsafemap = newblk->nb_bmsafemap;
1339 adp->ad_buf = bmsafemap->sm_buf;
1340 LIST_REMOVE(newblk, nb_deps);
1341 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1343 LIST_REMOVE(newblk, nb_hash);
1344 FREE(newblk, M_NEWBLK);
1346 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1347 if (lbn >= NDADDR) {
1348 /* allocating an indirect block */
1349 if (oldblkno != 0) {
1351 panic("softdep_setup_allocdirect: non-zero indir");
1355 * Allocating a direct block.
1357 * If we are allocating a directory block, then we must
1358 * allocate an associated pagedep to track additions and
1361 if ((ip->i_mode & IFMT) == IFDIR &&
1362 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1363 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1366 * The list of allocdirects must be kept in sorted and ascending
1367 * order so that the rollback routines can quickly determine the
1368 * first uncommitted block (the size of the file stored on disk
1369 * ends at the end of the lowest committed fragment, or if there
1370 * are no fragments, at the end of the highest committed block).
1371 * Since files generally grow, the typical case is that the new
1372 * block is to be added at the end of the list. We speed this
1373 * special case by checking against the last allocdirect in the
1374 * list before laboriously traversing the list looking for the
1377 adphead = &inodedep->id_newinoupdt;
1378 oldadp = TAILQ_LAST(adphead, allocdirectlst);
1379 if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1380 /* insert at end of list */
1381 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1382 if (oldadp != NULL && oldadp->ad_lbn == lbn)
1383 allocdirect_merge(adphead, adp, oldadp);
1387 TAILQ_FOREACH(oldadp, adphead, ad_next) {
1388 if (oldadp->ad_lbn >= lbn)
1391 if (oldadp == NULL) {
1393 panic("softdep_setup_allocdirect: lost entry");
1395 /* insert in middle of list */
1396 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1397 if (oldadp->ad_lbn == lbn)
1398 allocdirect_merge(adphead, adp, oldadp);
1403 * Replace an old allocdirect dependency with a newer one.
1404 * This routine must be called with splbio interrupts blocked.
1407 allocdirect_merge(adphead, newadp, oldadp)
1408 struct allocdirectlst *adphead; /* head of list holding allocdirects */
1409 struct allocdirect *newadp; /* allocdirect being added */
1410 struct allocdirect *oldadp; /* existing allocdirect being checked */
1412 struct freefrag *freefrag;
1415 if (lk.lkt_held == NOHOLDER)
1416 panic("allocdirect_merge: lock not held");
1418 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1419 newadp->ad_oldsize != oldadp->ad_newsize ||
1420 newadp->ad_lbn >= NDADDR) {
1422 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1423 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1426 newadp->ad_oldblkno = oldadp->ad_oldblkno;
1427 newadp->ad_oldsize = oldadp->ad_oldsize;
1429 * If the old dependency had a fragment to free or had never
1430 * previously had a block allocated, then the new dependency
1431 * can immediately post its freefrag and adopt the old freefrag.
1432 * This action is done by swapping the freefrag dependencies.
1433 * The new dependency gains the old one's freefrag, and the
1434 * old one gets the new one and then immediately puts it on
1435 * the worklist when it is freed by free_allocdirect. It is
1436 * not possible to do this swap when the old dependency had a
1437 * non-zero size but no previous fragment to free. This condition
1438 * arises when the new block is an extension of the old block.
1439 * Here, the first part of the fragment allocated to the new
1440 * dependency is part of the block currently claimed on disk by
1441 * the old dependency, so cannot legitimately be freed until the
1442 * conditions for the new dependency are fulfilled.
1444 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1445 freefrag = newadp->ad_freefrag;
1446 newadp->ad_freefrag = oldadp->ad_freefrag;
1447 oldadp->ad_freefrag = freefrag;
1449 free_allocdirect(adphead, oldadp, 0);
1453 * Allocate a new freefrag structure if needed.
1455 static struct freefrag *
1456 newfreefrag(ip, blkno, size)
1461 struct freefrag *freefrag;
1467 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1468 panic("newfreefrag: frag size");
1469 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1470 M_FREEFRAG, M_SOFTDEP_FLAGS);
1471 freefrag->ff_list.wk_type = D_FREEFRAG;
1472 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */
1473 freefrag->ff_inum = ip->i_number;
1474 freefrag->ff_fs = fs;
1475 freefrag->ff_devvp = ip->i_devvp;
1476 freefrag->ff_blkno = blkno;
1477 freefrag->ff_fragsize = size;
1482 * This workitem de-allocates fragments that were replaced during
1483 * file block allocation.
1486 handle_workitem_freefrag(freefrag)
1487 struct freefrag *freefrag;
1491 tip.i_fs = freefrag->ff_fs;
1492 tip.i_devvp = freefrag->ff_devvp;
1493 tip.i_dev = freefrag->ff_devvp->v_rdev;
1494 tip.i_number = freefrag->ff_inum;
1495 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */
1496 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1497 FREE(freefrag, M_FREEFRAG);
1501 * Indirect block allocation dependencies.
1503 * The same dependencies that exist for a direct block also exist when
1504 * a new block is allocated and pointed to by an entry in a block of
1505 * indirect pointers. The undo/redo states described above are also
1506 * used here. Because an indirect block contains many pointers that
1507 * may have dependencies, a second copy of the entire in-memory indirect
1508 * block is kept. The buffer cache copy is always completely up-to-date.
1509 * The second copy, which is used only as a source for disk writes,
1510 * contains only the safe pointers (i.e., those that have no remaining
1511 * update dependencies). The second copy is freed when all pointers
1512 * are safe. The cache is not allowed to replace indirect blocks with
1513 * pending update dependencies. If a buffer containing an indirect
1514 * block with dependencies is written, these routines will mark it
1515 * dirty again. It can only be successfully written once all the
1516 * dependencies are removed. The ffs_fsync routine in conjunction with
1517 * softdep_sync_metadata work together to get all the dependencies
1518 * removed so that a file can be successfully written to disk. Three
1519 * procedures are used when setting up indirect block pointer
1520 * dependencies. The division is necessary because of the organization
1521 * of the "balloc" routine and because of the distinction between file
1522 * pages and file metadata blocks.
1526 * Allocate a new allocindir structure.
1528 static struct allocindir *
1529 newallocindir(ip, ptrno, newblkno, oldblkno)
1530 struct inode *ip; /* inode for file being extended */
1531 int ptrno; /* offset of pointer in indirect block */
1532 ufs_daddr_t newblkno; /* disk block number being added */
1533 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1535 struct allocindir *aip;
1537 MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1538 M_ALLOCINDIR, M_SOFTDEP_FLAGS);
1539 bzero(aip, sizeof(struct allocindir));
1540 aip->ai_list.wk_type = D_ALLOCINDIR;
1541 aip->ai_state = ATTACHED;
1542 aip->ai_offset = ptrno;
1543 aip->ai_newblkno = newblkno;
1544 aip->ai_oldblkno = oldblkno;
1545 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1550 * Called just before setting an indirect block pointer
1551 * to a newly allocated file page.
1554 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1555 struct inode *ip; /* inode for file being extended */
1556 ufs_lbn_t lbn; /* allocated block number within file */
1557 struct buf *bp; /* buffer with indirect blk referencing page */
1558 int ptrno; /* offset of pointer in indirect block */
1559 ufs_daddr_t newblkno; /* disk block number being added */
1560 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1561 struct buf *nbp; /* buffer holding allocated page */
1563 struct allocindir *aip;
1564 struct pagedep *pagedep;
1566 aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1569 * If we are allocating a directory page, then we must
1570 * allocate an associated pagedep to track additions and
1573 if ((ip->i_mode & IFMT) == IFDIR &&
1574 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1575 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1576 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1578 setup_allocindir_phase2(bp, ip, aip);
1582 * Called just before setting an indirect block pointer to a
1583 * newly allocated indirect block.
1586 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1587 struct buf *nbp; /* newly allocated indirect block */
1588 struct inode *ip; /* inode for file being extended */
1589 struct buf *bp; /* indirect block referencing allocated block */
1590 int ptrno; /* offset of pointer in indirect block */
1591 ufs_daddr_t newblkno; /* disk block number being added */
1593 struct allocindir *aip;
1595 aip = newallocindir(ip, ptrno, newblkno, 0);
1597 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1599 setup_allocindir_phase2(bp, ip, aip);
1603 * Called to finish the allocation of the "aip" allocated
1604 * by one of the two routines above.
1607 setup_allocindir_phase2(bp, ip, aip)
1608 struct buf *bp; /* in-memory copy of the indirect block */
1609 struct inode *ip; /* inode for file being extended */
1610 struct allocindir *aip; /* allocindir allocated by the above routines */
1612 struct worklist *wk;
1613 struct indirdep *indirdep, *newindirdep;
1614 struct bmsafemap *bmsafemap;
1615 struct allocindir *oldaip;
1616 struct freefrag *freefrag;
1617 struct newblk *newblk;
1619 if (bp->b_lblkno >= 0)
1620 panic("setup_allocindir_phase2: not indir blk");
1621 for (indirdep = NULL, newindirdep = NULL; ; ) {
1623 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1624 if (wk->wk_type != D_INDIRDEP)
1626 indirdep = WK_INDIRDEP(wk);
1629 if (indirdep == NULL && newindirdep) {
1630 indirdep = newindirdep;
1631 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1636 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1638 panic("setup_allocindir: lost block");
1640 if (newblk->nb_state == DEPCOMPLETE) {
1641 aip->ai_state |= DEPCOMPLETE;
1644 bmsafemap = newblk->nb_bmsafemap;
1645 aip->ai_buf = bmsafemap->sm_buf;
1646 LIST_REMOVE(newblk, nb_deps);
1647 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1650 LIST_REMOVE(newblk, nb_hash);
1651 FREE(newblk, M_NEWBLK);
1652 aip->ai_indirdep = indirdep;
1654 * Check to see if there is an existing dependency
1655 * for this block. If there is, merge the old
1656 * dependency into the new one.
1658 if (aip->ai_oldblkno == 0)
1662 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1663 if (oldaip->ai_offset == aip->ai_offset)
1665 if (oldaip != NULL) {
1666 if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1668 panic("setup_allocindir_phase2: blkno");
1670 aip->ai_oldblkno = oldaip->ai_oldblkno;
1671 freefrag = oldaip->ai_freefrag;
1672 oldaip->ai_freefrag = aip->ai_freefrag;
1673 aip->ai_freefrag = freefrag;
1674 free_allocindir(oldaip, NULL);
1676 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1677 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1678 [aip->ai_offset] = aip->ai_oldblkno;
1683 * Avoid any possibility of data corruption by
1684 * ensuring that our old version is thrown away.
1686 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
1687 brelse(newindirdep->ir_savebp);
1688 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1692 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1693 M_INDIRDEP, M_SOFTDEP_FLAGS);
1694 newindirdep->ir_list.wk_type = D_INDIRDEP;
1695 newindirdep->ir_state = ATTACHED;
1696 LIST_INIT(&newindirdep->ir_deplisthd);
1697 LIST_INIT(&newindirdep->ir_donehd);
1698 if (bp->b_blkno == bp->b_lblkno) {
1699 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
1702 newindirdep->ir_savebp =
1703 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0);
1704 BUF_KERNPROC(newindirdep->ir_savebp);
1705 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1710 * Block de-allocation dependencies.
1712 * When blocks are de-allocated, the on-disk pointers must be nullified before
1713 * the blocks are made available for use by other files. (The true
1714 * requirement is that old pointers must be nullified before new on-disk
1715 * pointers are set. We chose this slightly more stringent requirement to
1716 * reduce complexity.) Our implementation handles this dependency by updating
1717 * the inode (or indirect block) appropriately but delaying the actual block
1718 * de-allocation (i.e., freemap and free space count manipulation) until
1719 * after the updated versions reach stable storage. After the disk is
1720 * updated, the blocks can be safely de-allocated whenever it is convenient.
1721 * This implementation handles only the common case of reducing a file's
1722 * length to zero. Other cases are handled by the conventional synchronous
1725 * The ffs implementation with which we worked double-checks
1726 * the state of the block pointers and file size as it reduces
1727 * a file's length. Some of this code is replicated here in our
1728 * soft updates implementation. The freeblks->fb_chkcnt field is
1729 * used to transfer a part of this information to the procedure
1730 * that eventually de-allocates the blocks.
1732 * This routine should be called from the routine that shortens
1733 * a file's length, before the inode's size or block pointers
1734 * are modified. It will save the block pointer information for
1735 * later release and zero the inode so that the calling routine
1738 struct softdep_setup_freeblocks_info {
1743 static int softdep_setup_freeblocks_bp(struct buf *bp, void *data);
1746 softdep_setup_freeblocks(ip, length)
1747 struct inode *ip; /* The inode whose length is to be reduced */
1748 off_t length; /* The new length for the file */
1750 struct softdep_setup_freeblocks_info info;
1751 struct freeblks *freeblks;
1752 struct inodedep *inodedep;
1753 struct allocdirect *adp;
1757 int i, error, delay;
1762 panic("softde_setup_freeblocks: non-zero length");
1763 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1764 M_FREEBLKS, M_SOFTDEP_FLAGS);
1765 bzero(freeblks, sizeof(struct freeblks));
1766 freeblks->fb_list.wk_type = D_FREEBLKS;
1767 freeblks->fb_state = ATTACHED;
1768 freeblks->fb_uid = ip->i_uid;
1769 freeblks->fb_previousinum = ip->i_number;
1770 freeblks->fb_devvp = ip->i_devvp;
1771 freeblks->fb_fs = fs;
1772 freeblks->fb_oldsize = ip->i_size;
1773 freeblks->fb_newsize = length;
1774 freeblks->fb_chkcnt = ip->i_blocks;
1775 for (i = 0; i < NDADDR; i++) {
1776 freeblks->fb_dblks[i] = ip->i_db[i];
1779 for (i = 0; i < NIADDR; i++) {
1780 freeblks->fb_iblks[i] = ip->i_ib[i];
1786 * Push the zero'ed inode to to its disk buffer so that we are free
1787 * to delete its dependencies below. Once the dependencies are gone
1788 * the buffer can be safely released.
1790 if ((error = bread(ip->i_devvp,
1791 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
1792 (int)fs->fs_bsize, &bp)) != 0)
1793 softdep_error("softdep_setup_freeblocks", error);
1794 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
1797 * Find and eliminate any inode dependencies.
1800 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1801 if ((inodedep->id_state & IOSTARTED) != 0) {
1803 panic("softdep_setup_freeblocks: inode busy");
1806 * Add the freeblks structure to the list of operations that
1807 * must await the zero'ed inode being written to disk. If we
1808 * still have a bitmap dependency (delay == 0), then the inode
1809 * has never been written to disk, so we can process the
1810 * freeblks below once we have deleted the dependencies.
1812 delay = (inodedep->id_state & DEPCOMPLETE);
1814 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1816 * Because the file length has been truncated to zero, any
1817 * pending block allocation dependency structures associated
1818 * with this inode are obsolete and can simply be de-allocated.
1819 * We must first merge the two dependency lists to get rid of
1820 * any duplicate freefrag structures, then purge the merged list.
1822 merge_inode_lists(inodedep);
1823 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1824 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
1828 * We must wait for any I/O in progress to finish so that
1829 * all potential buffers on the dirty list will be visible.
1830 * Once they are all there, walk the list and get rid of
1835 drain_output(vp, 1);
1840 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
1841 softdep_setup_freeblocks_bp, &info);
1842 } while (count > 0);
1843 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
1844 (void)free_inodedep(inodedep);
1847 freeblks->fb_state |= DEPCOMPLETE;
1849 * If the inode with zeroed block pointers is now on disk
1850 * we can start freeing blocks. Add freeblks to the worklist
1851 * instead of calling handle_workitem_freeblocks directly as
1852 * it is more likely that additional IO is needed to complete
1853 * the request here than in the !delay case.
1855 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
1856 add_to_worklist(&freeblks->fb_list);
1861 * If the inode has never been written to disk (delay == 0),
1862 * then we can process the freeblks now that we have deleted
1866 handle_workitem_freeblocks(freeblks);
1870 softdep_setup_freeblocks_bp(struct buf *bp, void *data)
1872 struct softdep_setup_freeblocks_info *info = data;
1873 struct inodedep *inodedep;
1875 if (getdirtybuf(&bp, MNT_WAIT) == 0)
1877 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep);
1878 deallocate_dependencies(bp, inodedep);
1879 bp->b_flags |= B_INVAL | B_NOCACHE;
1887 * Reclaim any dependency structures from a buffer that is about to
1888 * be reallocated to a new vnode. The buffer must be locked, thus,
1889 * no I/O completion operations can occur while we are manipulating
1890 * its associated dependencies. The mutex is held so that other I/O's
1891 * associated with related dependencies do not occur.
1894 deallocate_dependencies(bp, inodedep)
1896 struct inodedep *inodedep;
1898 struct worklist *wk;
1899 struct indirdep *indirdep;
1900 struct allocindir *aip;
1901 struct pagedep *pagedep;
1902 struct dirrem *dirrem;
1906 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1907 switch (wk->wk_type) {
1910 indirdep = WK_INDIRDEP(wk);
1912 * None of the indirect pointers will ever be visible,
1913 * so they can simply be tossed. GOINGAWAY ensures
1914 * that allocated pointers will be saved in the buffer
1915 * cache until they are freed. Note that they will
1916 * only be able to be found by their physical address
1917 * since the inode mapping the logical address will
1918 * be gone. The save buffer used for the safe copy
1919 * was allocated in setup_allocindir_phase2 using
1920 * the physical address so it could be used for this
1921 * purpose. Hence we swap the safe copy with the real
1922 * copy, allowing the safe copy to be freed and holding
1923 * on to the real copy for later use in indir_trunc.
1925 if (indirdep->ir_state & GOINGAWAY) {
1927 panic("deallocate_dependencies: already gone");
1929 indirdep->ir_state |= GOINGAWAY;
1930 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
1931 free_allocindir(aip, inodedep);
1932 if (bp->b_lblkno >= 0 ||
1933 bp->b_blkno != indirdep->ir_savebp->b_lblkno) {
1935 panic("deallocate_dependencies: not indir");
1937 bcopy(bp->b_data, indirdep->ir_savebp->b_data,
1939 WORKLIST_REMOVE(wk);
1940 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
1944 pagedep = WK_PAGEDEP(wk);
1946 * None of the directory additions will ever be
1947 * visible, so they can simply be tossed.
1949 for (i = 0; i < DAHASHSZ; i++)
1951 LIST_FIRST(&pagedep->pd_diraddhd[i])))
1953 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
1956 * Copy any directory remove dependencies to the list
1957 * to be processed after the zero'ed inode is written.
1958 * If the inode has already been written, then they
1959 * can be dumped directly onto the work list.
1961 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
1962 LIST_REMOVE(dirrem, dm_next);
1963 dirrem->dm_dirinum = pagedep->pd_ino;
1964 if (inodedep == NULL ||
1965 (inodedep->id_state & ALLCOMPLETE) ==
1967 add_to_worklist(&dirrem->dm_list);
1969 WORKLIST_INSERT(&inodedep->id_bufwait,
1972 WORKLIST_REMOVE(&pagedep->pd_list);
1973 LIST_REMOVE(pagedep, pd_hash);
1974 WORKITEM_FREE(pagedep, D_PAGEDEP);
1978 free_allocindir(WK_ALLOCINDIR(wk), inodedep);
1984 panic("deallocate_dependencies: Unexpected type %s",
1985 TYPENAME(wk->wk_type));
1990 panic("deallocate_dependencies: Unknown type %s",
1991 TYPENAME(wk->wk_type));
1998 * Free an allocdirect. Generate a new freefrag work request if appropriate.
1999 * This routine must be called with splbio interrupts blocked.
2002 free_allocdirect(adphead, adp, delay)
2003 struct allocdirectlst *adphead;
2004 struct allocdirect *adp;
2009 if (lk.lkt_held == NOHOLDER)
2010 panic("free_allocdirect: lock not held");
2012 if ((adp->ad_state & DEPCOMPLETE) == 0)
2013 LIST_REMOVE(adp, ad_deps);
2014 TAILQ_REMOVE(adphead, adp, ad_next);
2015 if ((adp->ad_state & COMPLETE) == 0)
2016 WORKLIST_REMOVE(&adp->ad_list);
2017 if (adp->ad_freefrag != NULL) {
2019 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2020 &adp->ad_freefrag->ff_list);
2022 add_to_worklist(&adp->ad_freefrag->ff_list);
2024 WORKITEM_FREE(adp, D_ALLOCDIRECT);
2028 * Prepare an inode to be freed. The actual free operation is not
2029 * done until the zero'ed inode has been written to disk.
2032 softdep_freefile(pvp, ino, mode)
2037 struct inode *ip = VTOI(pvp);
2038 struct inodedep *inodedep;
2039 struct freefile *freefile;
2042 * This sets up the inode de-allocation dependency.
2044 MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2045 M_FREEFILE, M_SOFTDEP_FLAGS);
2046 freefile->fx_list.wk_type = D_FREEFILE;
2047 freefile->fx_list.wk_state = 0;
2048 freefile->fx_mode = mode;
2049 freefile->fx_oldinum = ino;
2050 freefile->fx_devvp = ip->i_devvp;
2051 freefile->fx_fs = ip->i_fs;
2054 * If the inodedep does not exist, then the zero'ed inode has
2055 * been written to disk. If the allocated inode has never been
2056 * written to disk, then the on-disk inode is zero'ed. In either
2057 * case we can free the file immediately.
2060 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2061 check_inode_unwritten(inodedep)) {
2063 handle_workitem_freefile(freefile);
2066 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2071 * Check to see if an inode has never been written to disk. If
2072 * so free the inodedep and return success, otherwise return failure.
2073 * This routine must be called with splbio interrupts blocked.
2075 * If we still have a bitmap dependency, then the inode has never
2076 * been written to disk. Drop the dependency as it is no longer
2077 * necessary since the inode is being deallocated. We set the
2078 * ALLCOMPLETE flags since the bitmap now properly shows that the
2079 * inode is not allocated. Even if the inode is actively being
2080 * written, it has been rolled back to its zero'ed state, so we
2081 * are ensured that a zero inode is what is on the disk. For short
2082 * lived files, this change will usually result in removing all the
2083 * dependencies from the inode so that it can be freed immediately.
2086 check_inode_unwritten(inodedep)
2087 struct inodedep *inodedep;
2090 if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2091 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2092 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2093 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2094 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2095 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2096 inodedep->id_nlinkdelta != 0)
2098 inodedep->id_state |= ALLCOMPLETE;
2099 LIST_REMOVE(inodedep, id_deps);
2100 inodedep->id_buf = NULL;
2101 if (inodedep->id_state & ONWORKLIST)
2102 WORKLIST_REMOVE(&inodedep->id_list);
2103 if (inodedep->id_savedino != NULL) {
2104 FREE(inodedep->id_savedino, M_INODEDEP);
2105 inodedep->id_savedino = NULL;
2107 if (free_inodedep(inodedep) == 0) {
2109 panic("check_inode_unwritten: busy inode");
2115 * Try to free an inodedep structure. Return 1 if it could be freed.
2118 free_inodedep(inodedep)
2119 struct inodedep *inodedep;
2122 if ((inodedep->id_state & ONWORKLIST) != 0 ||
2123 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2124 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2125 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2126 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2127 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2128 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2129 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
2131 LIST_REMOVE(inodedep, id_hash);
2132 WORKITEM_FREE(inodedep, D_INODEDEP);
2138 * This workitem routine performs the block de-allocation.
2139 * The workitem is added to the pending list after the updated
2140 * inode block has been written to disk. As mentioned above,
2141 * checks regarding the number of blocks de-allocated (compared
2142 * to the number of blocks allocated for the file) are also
2143 * performed in this function.
2146 handle_workitem_freeblocks(freeblks)
2147 struct freeblks *freeblks;
2152 int i, level, bsize;
2153 long nblocks, blocksreleased = 0;
2154 int error, allerror = 0;
2155 ufs_lbn_t baselbns[NIADDR], tmpval;
2157 tip.i_number = freeblks->fb_previousinum;
2158 tip.i_devvp = freeblks->fb_devvp;
2159 tip.i_dev = freeblks->fb_devvp->v_rdev;
2160 tip.i_fs = freeblks->fb_fs;
2161 tip.i_size = freeblks->fb_oldsize;
2162 tip.i_uid = freeblks->fb_uid;
2163 fs = freeblks->fb_fs;
2165 baselbns[0] = NDADDR;
2166 for (i = 1; i < NIADDR; i++) {
2167 tmpval *= NINDIR(fs);
2168 baselbns[i] = baselbns[i - 1] + tmpval;
2170 nblocks = btodb(fs->fs_bsize);
2173 * Indirect blocks first.
2175 for (level = (NIADDR - 1); level >= 0; level--) {
2176 if ((bn = freeblks->fb_iblks[level]) == 0)
2178 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level,
2179 baselbns[level], &blocksreleased)) == 0)
2181 ffs_blkfree(&tip, bn, fs->fs_bsize);
2182 blocksreleased += nblocks;
2185 * All direct blocks or frags.
2187 for (i = (NDADDR - 1); i >= 0; i--) {
2188 if ((bn = freeblks->fb_dblks[i]) == 0)
2190 bsize = blksize(fs, &tip, i);
2191 ffs_blkfree(&tip, bn, bsize);
2192 blocksreleased += btodb(bsize);
2196 if (freeblks->fb_chkcnt != blocksreleased)
2197 printf("handle_workitem_freeblocks: block count\n");
2199 softdep_error("handle_workitem_freeblks", allerror);
2200 #endif /* DIAGNOSTIC */
2201 WORKITEM_FREE(freeblks, D_FREEBLKS);
2205 * Release blocks associated with the inode ip and stored in the indirect
2206 * block dbn. If level is greater than SINGLE, the block is an indirect block
2207 * and recursive calls to indirtrunc must be used to cleanse other indirect
2211 indir_trunc(ip, dbn, level, lbn, countp)
2222 struct worklist *wk;
2223 struct indirdep *indirdep;
2224 int i, lbnadd, nblocks;
2225 int error, allerror = 0;
2229 for (i = level; i > 0; i--)
2230 lbnadd *= NINDIR(fs);
2232 * Get buffer of block pointers to be freed. This routine is not
2233 * called until the zero'ed inode has been written, so it is safe
2234 * to free blocks as they are encountered. Because the inode has
2235 * been zero'ed, calls to bmap on these blocks will fail. So, we
2236 * have to use the on-disk address and the block device for the
2237 * filesystem to look them up. If the file was deleted before its
2238 * indirect blocks were all written to disk, the routine that set
2239 * us up (deallocate_dependencies) will have arranged to leave
2240 * a complete copy of the indirect block in memory for our use.
2241 * Otherwise we have to read the blocks in from the disk.
2244 if ((bp = incore(ip->i_devvp, dbn)) != NULL &&
2245 (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2246 if (wk->wk_type != D_INDIRDEP ||
2247 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2248 (indirdep->ir_state & GOINGAWAY) == 0) {
2250 panic("indir_trunc: lost indirdep");
2252 WORKLIST_REMOVE(wk);
2253 WORKITEM_FREE(indirdep, D_INDIRDEP);
2254 if (LIST_FIRST(&bp->b_dep) != NULL) {
2256 panic("indir_trunc: dangling dep");
2261 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, &bp);
2266 * Recursively free indirect blocks.
2268 bap = (ufs_daddr_t *)bp->b_data;
2269 nblocks = btodb(fs->fs_bsize);
2270 for (i = NINDIR(fs) - 1; i >= 0; i--) {
2271 if ((nb = bap[i]) == 0)
2274 if ((error = indir_trunc(ip, fsbtodb(fs, nb),
2275 level - 1, lbn + (i * lbnadd), countp)) != 0)
2278 ffs_blkfree(ip, nb, fs->fs_bsize);
2281 bp->b_flags |= B_INVAL | B_NOCACHE;
2287 * Free an allocindir.
2288 * This routine must be called with splbio interrupts blocked.
2291 free_allocindir(aip, inodedep)
2292 struct allocindir *aip;
2293 struct inodedep *inodedep;
2295 struct freefrag *freefrag;
2298 if (lk.lkt_held == NOHOLDER)
2299 panic("free_allocindir: lock not held");
2301 if ((aip->ai_state & DEPCOMPLETE) == 0)
2302 LIST_REMOVE(aip, ai_deps);
2303 if (aip->ai_state & ONWORKLIST)
2304 WORKLIST_REMOVE(&aip->ai_list);
2305 LIST_REMOVE(aip, ai_next);
2306 if ((freefrag = aip->ai_freefrag) != NULL) {
2307 if (inodedep == NULL)
2308 add_to_worklist(&freefrag->ff_list);
2310 WORKLIST_INSERT(&inodedep->id_bufwait,
2311 &freefrag->ff_list);
2313 WORKITEM_FREE(aip, D_ALLOCINDIR);
2317 * Directory entry addition dependencies.
2319 * When adding a new directory entry, the inode (with its incremented link
2320 * count) must be written to disk before the directory entry's pointer to it.
2321 * Also, if the inode is newly allocated, the corresponding freemap must be
2322 * updated (on disk) before the directory entry's pointer. These requirements
2323 * are met via undo/redo on the directory entry's pointer, which consists
2324 * simply of the inode number.
2326 * As directory entries are added and deleted, the free space within a
2327 * directory block can become fragmented. The ufs filesystem will compact
2328 * a fragmented directory block to make space for a new entry. When this
2329 * occurs, the offsets of previously added entries change. Any "diradd"
2330 * dependency structures corresponding to these entries must be updated with
2335 * This routine is called after the in-memory inode's link
2336 * count has been incremented, but before the directory entry's
2337 * pointer to the inode has been set.
2340 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp)
2341 struct buf *bp; /* buffer containing directory block */
2342 struct inode *dp; /* inode for directory */
2343 off_t diroffset; /* offset of new entry in directory */
2344 ino_t newinum; /* inode referenced by new directory entry */
2345 struct buf *newdirbp; /* non-NULL => contents of new mkdir */
2347 int offset; /* offset of new entry within directory block */
2348 ufs_lbn_t lbn; /* block in directory containing new entry */
2351 struct pagedep *pagedep;
2352 struct inodedep *inodedep;
2353 struct mkdir *mkdir1, *mkdir2;
2356 * Whiteouts have no dependencies.
2358 if (newinum == WINO) {
2359 if (newdirbp != NULL)
2365 lbn = lblkno(fs, diroffset);
2366 offset = blkoff(fs, diroffset);
2367 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2369 bzero(dap, sizeof(struct diradd));
2370 dap->da_list.wk_type = D_DIRADD;
2371 dap->da_offset = offset;
2372 dap->da_newinum = newinum;
2373 dap->da_state = ATTACHED;
2374 if (newdirbp == NULL) {
2375 dap->da_state |= DEPCOMPLETE;
2378 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2379 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2381 mkdir1->md_list.wk_type = D_MKDIR;
2382 mkdir1->md_state = MKDIR_BODY;
2383 mkdir1->md_diradd = dap;
2384 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2386 mkdir2->md_list.wk_type = D_MKDIR;
2387 mkdir2->md_state = MKDIR_PARENT;
2388 mkdir2->md_diradd = dap;
2390 * Dependency on "." and ".." being written to disk.
2392 mkdir1->md_buf = newdirbp;
2394 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2395 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2399 * Dependency on link count increase for parent directory
2402 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2403 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2404 dap->da_state &= ~MKDIR_PARENT;
2405 WORKITEM_FREE(mkdir2, D_MKDIR);
2407 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2408 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2412 * Link into parent directory pagedep to await its being written.
2414 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2415 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2416 dap->da_pagedep = pagedep;
2417 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2420 * Link into its inodedep. Put it on the id_bufwait list if the inode
2421 * is not yet written. If it is written, do the post-inode write
2422 * processing to put it on the id_pendinghd list.
2424 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2425 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2426 diradd_inode_written(dap, inodedep);
2428 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2433 * This procedure is called to change the offset of a directory
2434 * entry when compacting a directory block which must be owned
2435 * exclusively by the caller. Note that the actual entry movement
2436 * must be done in this procedure to ensure that no I/O completions
2437 * occur while the move is in progress.
2440 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2441 struct inode *dp; /* inode for directory */
2442 caddr_t base; /* address of dp->i_offset */
2443 caddr_t oldloc; /* address of old directory location */
2444 caddr_t newloc; /* address of new directory location */
2445 int entrysize; /* size of directory entry */
2447 int offset, oldoffset, newoffset;
2448 struct pagedep *pagedep;
2453 lbn = lblkno(dp->i_fs, dp->i_offset);
2454 offset = blkoff(dp->i_fs, dp->i_offset);
2455 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2457 oldoffset = offset + (oldloc - base);
2458 newoffset = offset + (newloc - base);
2460 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2461 if (dap->da_offset != oldoffset)
2463 dap->da_offset = newoffset;
2464 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2466 LIST_REMOVE(dap, da_pdlist);
2467 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2473 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2474 if (dap->da_offset == oldoffset) {
2475 dap->da_offset = newoffset;
2481 bcopy(oldloc, newloc, entrysize);
2486 * Free a diradd dependency structure. This routine must be called
2487 * with splbio interrupts blocked.
2493 struct dirrem *dirrem;
2494 struct pagedep *pagedep;
2495 struct inodedep *inodedep;
2496 struct mkdir *mkdir, *nextmd;
2499 if (lk.lkt_held == NOHOLDER)
2500 panic("free_diradd: lock not held");
2502 WORKLIST_REMOVE(&dap->da_list);
2503 LIST_REMOVE(dap, da_pdlist);
2504 if ((dap->da_state & DIRCHG) == 0) {
2505 pagedep = dap->da_pagedep;
2507 dirrem = dap->da_previous;
2508 pagedep = dirrem->dm_pagedep;
2509 dirrem->dm_dirinum = pagedep->pd_ino;
2510 add_to_worklist(&dirrem->dm_list);
2512 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2514 (void) free_inodedep(inodedep);
2515 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2516 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2517 nextmd = LIST_NEXT(mkdir, md_mkdirs);
2518 if (mkdir->md_diradd != dap)
2520 dap->da_state &= ~mkdir->md_state;
2521 WORKLIST_REMOVE(&mkdir->md_list);
2522 LIST_REMOVE(mkdir, md_mkdirs);
2523 WORKITEM_FREE(mkdir, D_MKDIR);
2525 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2527 panic("free_diradd: unfound ref");
2530 WORKITEM_FREE(dap, D_DIRADD);
2534 * Directory entry removal dependencies.
2536 * When removing a directory entry, the entry's inode pointer must be
2537 * zero'ed on disk before the corresponding inode's link count is decremented
2538 * (possibly freeing the inode for re-use). This dependency is handled by
2539 * updating the directory entry but delaying the inode count reduction until
2540 * after the directory block has been written to disk. After this point, the
2541 * inode count can be decremented whenever it is convenient.
2545 * This routine should be called immediately after removing
2546 * a directory entry. The inode's link count should not be
2547 * decremented by the calling procedure -- the soft updates
2548 * code will do this task when it is safe.
2551 softdep_setup_remove(bp, dp, ip, isrmdir)
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 */
2557 struct dirrem *dirrem, *prevdirrem;
2560 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2562 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2565 * If the COMPLETE flag is clear, then there were no active
2566 * entries and we want to roll back to a zeroed entry until
2567 * the new inode is committed to disk. If the COMPLETE flag is
2568 * set then we have deleted an entry that never made it to
2569 * disk. If the entry we deleted resulted from a name change,
2570 * then the old name still resides on disk. We cannot delete
2571 * its inode (returned to us in prevdirrem) until the zeroed
2572 * directory entry gets to disk. The new inode has never been
2573 * referenced on the disk, so can be deleted immediately.
2575 if ((dirrem->dm_state & COMPLETE) == 0) {
2576 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2580 if (prevdirrem != NULL)
2581 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2582 prevdirrem, dm_next);
2583 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2585 handle_workitem_remove(dirrem);
2590 * Allocate a new dirrem if appropriate and return it along with
2591 * its associated pagedep. Called without a lock, returns with lock.
2593 static long num_dirrem; /* number of dirrem allocated */
2594 static struct dirrem *
2595 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2596 struct buf *bp; /* buffer containing directory block */
2597 struct inode *dp; /* inode for the directory being modified */
2598 struct inode *ip; /* inode for directory entry being removed */
2599 int isrmdir; /* indicates if doing RMDIR */
2600 struct dirrem **prevdirremp; /* previously referenced inode, if any */
2605 struct dirrem *dirrem;
2606 struct pagedep *pagedep;
2609 * Whiteouts have no deletion dependencies.
2612 panic("newdirrem: whiteout");
2614 * If we are over our limit, try to improve the situation.
2615 * Limiting the number of dirrem structures will also limit
2616 * the number of freefile and freeblks structures.
2618 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2619 (void) request_cleanup(FLUSH_REMOVE, 0);
2621 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
2622 M_DIRREM, M_SOFTDEP_FLAGS);
2623 bzero(dirrem, sizeof(struct dirrem));
2624 dirrem->dm_list.wk_type = D_DIRREM;
2625 dirrem->dm_state = isrmdir ? RMDIR : 0;
2626 dirrem->dm_mnt = ITOV(ip)->v_mount;
2627 dirrem->dm_oldinum = ip->i_number;
2628 *prevdirremp = NULL;
2631 lbn = lblkno(dp->i_fs, dp->i_offset);
2632 offset = blkoff(dp->i_fs, dp->i_offset);
2633 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2634 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2635 dirrem->dm_pagedep = pagedep;
2637 * Check for a diradd dependency for the same directory entry.
2638 * If present, then both dependencies become obsolete and can
2639 * be de-allocated. Check for an entry on both the pd_dirraddhd
2640 * list and the pd_pendinghd list.
2643 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
2644 if (dap->da_offset == offset)
2648 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
2649 if (dap->da_offset == offset)
2655 * Must be ATTACHED at this point.
2657 if ((dap->da_state & ATTACHED) == 0) {
2659 panic("newdirrem: not ATTACHED");
2661 if (dap->da_newinum != ip->i_number) {
2663 panic("newdirrem: inum %d should be %d",
2664 ip->i_number, dap->da_newinum);
2667 * If we are deleting a changed name that never made it to disk,
2668 * then return the dirrem describing the previous inode (which
2669 * represents the inode currently referenced from this entry on disk).
2671 if ((dap->da_state & DIRCHG) != 0) {
2672 *prevdirremp = dap->da_previous;
2673 dap->da_state &= ~DIRCHG;
2674 dap->da_pagedep = pagedep;
2677 * We are deleting an entry that never made it to disk.
2678 * Mark it COMPLETE so we can delete its inode immediately.
2680 dirrem->dm_state |= COMPLETE;
2686 * Directory entry change dependencies.
2688 * Changing an existing directory entry requires that an add operation
2689 * be completed first followed by a deletion. The semantics for the addition
2690 * are identical to the description of adding a new entry above except
2691 * that the rollback is to the old inode number rather than zero. Once
2692 * the addition dependency is completed, the removal is done as described
2693 * in the removal routine above.
2697 * This routine should be called immediately after changing
2698 * a directory entry. The inode's link count should not be
2699 * decremented by the calling procedure -- the soft updates
2700 * code will perform this task when it is safe.
2703 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
2704 struct buf *bp; /* buffer containing directory block */
2705 struct inode *dp; /* inode for the directory being modified */
2706 struct inode *ip; /* inode for directory entry being removed */
2707 ino_t newinum; /* new inode number for changed entry */
2708 int isrmdir; /* indicates if doing RMDIR */
2711 struct diradd *dap = NULL;
2712 struct dirrem *dirrem, *prevdirrem;
2713 struct pagedep *pagedep;
2714 struct inodedep *inodedep;
2716 offset = blkoff(dp->i_fs, dp->i_offset);
2719 * Whiteouts do not need diradd dependencies.
2721 if (newinum != WINO) {
2722 MALLOC(dap, struct diradd *, sizeof(struct diradd),
2723 M_DIRADD, M_SOFTDEP_FLAGS);
2724 bzero(dap, sizeof(struct diradd));
2725 dap->da_list.wk_type = D_DIRADD;
2726 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2727 dap->da_offset = offset;
2728 dap->da_newinum = newinum;
2732 * Allocate a new dirrem and ACQUIRE_LOCK.
2734 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2735 pagedep = dirrem->dm_pagedep;
2737 * The possible values for isrmdir:
2738 * 0 - non-directory file rename
2739 * 1 - directory rename within same directory
2740 * inum - directory rename to new directory of given inode number
2741 * When renaming to a new directory, we are both deleting and
2742 * creating a new directory entry, so the link count on the new
2743 * directory should not change. Thus we do not need the followup
2744 * dirrem which is usually done in handle_workitem_remove. We set
2745 * the DIRCHG flag to tell handle_workitem_remove to skip the
2749 dirrem->dm_state |= DIRCHG;
2752 * Whiteouts have no additional dependencies,
2753 * so just put the dirrem on the correct list.
2755 if (newinum == WINO) {
2756 if ((dirrem->dm_state & COMPLETE) == 0) {
2757 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2760 dirrem->dm_dirinum = pagedep->pd_ino;
2761 add_to_worklist(&dirrem->dm_list);
2768 * If the COMPLETE flag is clear, then there were no active
2769 * entries and we want to roll back to the previous inode until
2770 * the new inode is committed to disk. If the COMPLETE flag is
2771 * set, then we have deleted an entry that never made it to disk.
2772 * If the entry we deleted resulted from a name change, then the old
2773 * inode reference still resides on disk. Any rollback that we do
2774 * needs to be to that old inode (returned to us in prevdirrem). If
2775 * the entry we deleted resulted from a create, then there is
2776 * no entry on the disk, so we want to roll back to zero rather
2777 * than the uncommitted inode. In either of the COMPLETE cases we
2778 * want to immediately free the unwritten and unreferenced inode.
2780 if ((dirrem->dm_state & COMPLETE) == 0) {
2781 dap->da_previous = dirrem;
2783 if (prevdirrem != NULL) {
2784 dap->da_previous = prevdirrem;
2786 dap->da_state &= ~DIRCHG;
2787 dap->da_pagedep = pagedep;
2789 dirrem->dm_dirinum = pagedep->pd_ino;
2790 add_to_worklist(&dirrem->dm_list);
2793 * Link into its inodedep. Put it on the id_bufwait list if the inode
2794 * is not yet written. If it is written, do the post-inode write
2795 * processing to put it on the id_pendinghd list.
2797 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2798 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2799 dap->da_state |= COMPLETE;
2800 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2801 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2803 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2805 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2811 * Called whenever the link count on an inode is changed.
2812 * It creates an inode dependency so that the new reference(s)
2813 * to the inode cannot be committed to disk until the updated
2814 * inode has been written.
2817 softdep_change_linkcnt(ip)
2818 struct inode *ip; /* the inode with the increased link count */
2820 struct inodedep *inodedep;
2823 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2824 if (ip->i_nlink < ip->i_effnlink) {
2826 panic("softdep_change_linkcnt: bad delta");
2828 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2833 * This workitem decrements the inode's link count.
2834 * If the link count reaches zero, the file is removed.
2837 handle_workitem_remove(dirrem)
2838 struct dirrem *dirrem;
2840 struct thread *td = curthread; /* XXX */
2841 struct inodedep *inodedep;
2847 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) {
2848 softdep_error("handle_workitem_remove: vget", error);
2853 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
2855 panic("handle_workitem_remove: lost inodedep");
2858 * Normal file deletion.
2860 if ((dirrem->dm_state & RMDIR) == 0) {
2862 ip->i_flag |= IN_CHANGE;
2863 if (ip->i_nlink < ip->i_effnlink) {
2865 panic("handle_workitem_remove: bad file delta");
2867 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2871 WORKITEM_FREE(dirrem, D_DIRREM);
2875 * Directory deletion. Decrement reference count for both the
2876 * just deleted parent directory entry and the reference for ".".
2877 * Next truncate the directory to length zero. When the
2878 * truncation completes, arrange to have the reference count on
2879 * the parent decremented to account for the loss of "..".
2882 ip->i_flag |= IN_CHANGE;
2883 if (ip->i_nlink < ip->i_effnlink) {
2885 panic("handle_workitem_remove: bad dir delta");
2887 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2889 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, proc0.p_ucred, td)) != 0)
2890 softdep_error("handle_workitem_remove: truncate", error);
2892 * Rename a directory to a new parent. Since, we are both deleting
2893 * and creating a new directory entry, the link count on the new
2894 * directory should not change. Thus we skip the followup dirrem.
2896 if (dirrem->dm_state & DIRCHG) {
2899 WORKITEM_FREE(dirrem, D_DIRREM);
2903 * If the inodedep does not exist, then the zero'ed inode has
2904 * been written to disk. If the allocated inode has never been
2905 * written to disk, then the on-disk inode is zero'ed. In either
2906 * case we can remove the file immediately.
2909 dirrem->dm_state = 0;
2910 oldinum = dirrem->dm_oldinum;
2911 dirrem->dm_oldinum = dirrem->dm_dirinum;
2912 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
2913 check_inode_unwritten(inodedep)) {
2916 handle_workitem_remove(dirrem);
2919 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
2925 * Inode de-allocation dependencies.
2927 * When an inode's link count is reduced to zero, it can be de-allocated. We
2928 * found it convenient to postpone de-allocation until after the inode is
2929 * written to disk with its new link count (zero). At this point, all of the
2930 * on-disk inode's block pointers are nullified and, with careful dependency
2931 * list ordering, all dependencies related to the inode will be satisfied and
2932 * the corresponding dependency structures de-allocated. So, if/when the
2933 * inode is reused, there will be no mixing of old dependencies with new
2934 * ones. This artificial dependency is set up by the block de-allocation
2935 * procedure above (softdep_setup_freeblocks) and completed by the
2936 * following procedure.
2939 handle_workitem_freefile(freefile)
2940 struct freefile *freefile;
2944 struct inodedep *idp;
2949 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp);
2952 panic("handle_workitem_freefile: inodedep survived");
2954 tip.i_devvp = freefile->fx_devvp;
2955 tip.i_dev = freefile->fx_devvp->v_rdev;
2956 tip.i_fs = freefile->fx_fs;
2958 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
2959 softdep_error("handle_workitem_freefile", error);
2960 WORKITEM_FREE(freefile, D_FREEFILE);
2966 * The dependency structures constructed above are most actively used when file
2967 * system blocks are written to disk. No constraints are placed on when a
2968 * block can be written, but unsatisfied update dependencies are made safe by
2969 * modifying (or replacing) the source memory for the duration of the disk
2970 * write. When the disk write completes, the memory block is again brought
2973 * In-core inode structure reclamation.
2975 * Because there are a finite number of "in-core" inode structures, they are
2976 * reused regularly. By transferring all inode-related dependencies to the
2977 * in-memory inode block and indexing them separately (via "inodedep"s), we
2978 * can allow "in-core" inode structures to be reused at any time and avoid
2979 * any increase in contention.
2981 * Called just before entering the device driver to initiate a new disk I/O.
2982 * The buffer must be locked, thus, no I/O completion operations can occur
2983 * while we are manipulating its associated dependencies.
2986 softdep_disk_io_initiation(bp)
2987 struct buf *bp; /* structure describing disk write to occur */
2989 struct worklist *wk, *nextwk;
2990 struct indirdep *indirdep;
2993 * We only care about write operations. There should never
2994 * be dependencies for reads.
2996 if (bp->b_flags & B_READ)
2997 panic("softdep_disk_io_initiation: read");
2999 * Do any necessary pre-I/O processing.
3001 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
3002 nextwk = LIST_NEXT(wk, wk_list);
3003 switch (wk->wk_type) {
3006 initiate_write_filepage(WK_PAGEDEP(wk), bp);
3010 initiate_write_inodeblock(WK_INODEDEP(wk), bp);
3014 indirdep = WK_INDIRDEP(wk);
3015 if (indirdep->ir_state & GOINGAWAY)
3016 panic("disk_io_initiation: indirdep gone");
3018 * If there are no remaining dependencies, this
3019 * will be writing the real pointers, so the
3020 * dependency can be freed.
3022 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
3023 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
3024 brelse(indirdep->ir_savebp);
3025 /* inline expand WORKLIST_REMOVE(wk); */
3026 wk->wk_state &= ~ONWORKLIST;
3027 LIST_REMOVE(wk, wk_list);
3028 WORKITEM_FREE(indirdep, D_INDIRDEP);
3032 * Replace up-to-date version with safe version.
3034 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3035 M_INDIRDEP, M_SOFTDEP_FLAGS);
3037 indirdep->ir_state &= ~ATTACHED;
3038 indirdep->ir_state |= UNDONE;
3039 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3040 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3052 panic("handle_disk_io_initiation: Unexpected type %s",
3053 TYPENAME(wk->wk_type));
3060 * Called from within the procedure above to deal with unsatisfied
3061 * allocation dependencies in a directory. The buffer must be locked,
3062 * thus, no I/O completion operations can occur while we are
3063 * manipulating its associated dependencies.
3066 initiate_write_filepage(pagedep, bp)
3067 struct pagedep *pagedep;
3074 if (pagedep->pd_state & IOSTARTED) {
3076 * This can only happen if there is a driver that does not
3077 * understand chaining. Here biodone will reissue the call
3078 * to strategy for the incomplete buffers.
3080 printf("initiate_write_filepage: already started\n");
3083 pagedep->pd_state |= IOSTARTED;
3085 for (i = 0; i < DAHASHSZ; i++) {
3086 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3087 ep = (struct direct *)
3088 ((char *)bp->b_data + dap->da_offset);
3089 if (ep->d_ino != dap->da_newinum) {
3091 panic("%s: dir inum %d != new %d",
3092 "initiate_write_filepage",
3093 ep->d_ino, dap->da_newinum);
3095 if (dap->da_state & DIRCHG)
3096 ep->d_ino = dap->da_previous->dm_oldinum;
3099 dap->da_state &= ~ATTACHED;
3100 dap->da_state |= UNDONE;
3107 * Called from within the procedure above to deal with unsatisfied
3108 * allocation dependencies in an inodeblock. The buffer must be
3109 * locked, thus, no I/O completion operations can occur while we
3110 * are manipulating its associated dependencies.
3113 initiate_write_inodeblock(inodedep, bp)
3114 struct inodedep *inodedep;
3115 struct buf *bp; /* The inode block */
3117 struct allocdirect *adp, *lastadp;
3120 ufs_lbn_t prevlbn = 0;
3123 if (inodedep->id_state & IOSTARTED)
3124 panic("initiate_write_inodeblock: already started");
3125 inodedep->id_state |= IOSTARTED;
3126 fs = inodedep->id_fs;
3127 dp = (struct dinode *)bp->b_data +
3128 ino_to_fsbo(fs, inodedep->id_ino);
3130 * If the bitmap is not yet written, then the allocated
3131 * inode cannot be written to disk.
3133 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3134 if (inodedep->id_savedino != NULL)
3135 panic("initiate_write_inodeblock: already doing I/O");
3136 MALLOC(inodedep->id_savedino, struct dinode *,
3137 sizeof(struct dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3138 *inodedep->id_savedino = *dp;
3139 bzero((caddr_t)dp, sizeof(struct dinode));
3143 * If no dependencies, then there is nothing to roll back.
3145 inodedep->id_savedsize = dp->di_size;
3146 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3149 * Set the dependencies to busy.
3152 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3153 adp = TAILQ_NEXT(adp, ad_next)) {
3155 if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3157 panic("softdep_write_inodeblock: lbn order");
3159 prevlbn = adp->ad_lbn;
3160 if (adp->ad_lbn < NDADDR &&
3161 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3163 panic("%s: direct pointer #%ld mismatch %d != %d",
3164 "softdep_write_inodeblock", adp->ad_lbn,
3165 dp->di_db[adp->ad_lbn], adp->ad_newblkno);
3167 if (adp->ad_lbn >= NDADDR &&
3168 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3170 panic("%s: indirect pointer #%ld mismatch %d != %d",
3171 "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
3172 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
3174 deplist |= 1 << adp->ad_lbn;
3175 if ((adp->ad_state & ATTACHED) == 0) {
3177 panic("softdep_write_inodeblock: Unknown state 0x%x",
3180 #endif /* DIAGNOSTIC */
3181 adp->ad_state &= ~ATTACHED;
3182 adp->ad_state |= UNDONE;
3185 * The on-disk inode cannot claim to be any larger than the last
3186 * fragment that has been written. Otherwise, the on-disk inode
3187 * might have fragments that were not the last block in the file
3188 * which would corrupt the filesystem.
3190 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3191 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3192 if (adp->ad_lbn >= NDADDR)
3194 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3195 /* keep going until hitting a rollback to a frag */
3196 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3198 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3199 for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3201 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3203 panic("softdep_write_inodeblock: lost dep1");
3205 #endif /* DIAGNOSTIC */
3208 for (i = 0; i < NIADDR; i++) {
3210 if (dp->di_ib[i] != 0 &&
3211 (deplist & ((1 << NDADDR) << i)) == 0) {
3213 panic("softdep_write_inodeblock: lost dep2");
3215 #endif /* DIAGNOSTIC */
3222 * If we have zero'ed out the last allocated block of the file,
3223 * roll back the size to the last currently allocated block.
3224 * We know that this last allocated block is a full-sized as
3225 * we already checked for fragments in the loop above.
3227 if (lastadp != NULL &&
3228 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3229 for (i = lastadp->ad_lbn; i >= 0; i--)
3230 if (dp->di_db[i] != 0)
3232 dp->di_size = (i + 1) * fs->fs_bsize;
3235 * The only dependencies are for indirect blocks.
3237 * The file size for indirect block additions is not guaranteed.
3238 * Such a guarantee would be non-trivial to achieve. The conventional
3239 * synchronous write implementation also does not make this guarantee.
3240 * Fsck should catch and fix discrepancies. Arguably, the file size
3241 * can be over-estimated without destroying integrity when the file
3242 * moves into the indirect blocks (i.e., is large). If we want to
3243 * postpone fsck, we are stuck with this argument.
3245 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3246 dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3251 * This routine is called during the completion interrupt
3252 * service routine for a disk write (from the procedure called
3253 * by the device driver to inform the filesystem caches of
3254 * a request completion). It should be called early in this
3255 * procedure, before the block is made available to other
3256 * processes or other routines are called.
3259 softdep_disk_write_complete(bp)
3260 struct buf *bp; /* describes the completed disk write */
3262 struct worklist *wk;
3263 struct workhead reattach;
3264 struct newblk *newblk;
3265 struct allocindir *aip;
3266 struct allocdirect *adp;
3267 struct indirdep *indirdep;
3268 struct inodedep *inodedep;
3269 struct bmsafemap *bmsafemap;
3272 if (lk.lkt_held != NOHOLDER)
3273 panic("softdep_disk_write_complete: lock is held");
3274 lk.lkt_held = SPECIAL_FLAG;
3276 LIST_INIT(&reattach);
3277 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3278 WORKLIST_REMOVE(wk);
3279 switch (wk->wk_type) {
3282 if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3283 WORKLIST_INSERT(&reattach, wk);
3287 if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3288 WORKLIST_INSERT(&reattach, wk);
3292 bmsafemap = WK_BMSAFEMAP(wk);
3293 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3294 newblk->nb_state |= DEPCOMPLETE;
3295 newblk->nb_bmsafemap = NULL;
3296 LIST_REMOVE(newblk, nb_deps);
3299 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3300 adp->ad_state |= DEPCOMPLETE;
3302 LIST_REMOVE(adp, ad_deps);
3303 handle_allocdirect_partdone(adp);
3306 LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3307 aip->ai_state |= DEPCOMPLETE;
3309 LIST_REMOVE(aip, ai_deps);
3310 handle_allocindir_partdone(aip);
3313 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3314 inodedep->id_state |= DEPCOMPLETE;
3315 LIST_REMOVE(inodedep, id_deps);
3316 inodedep->id_buf = NULL;
3318 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3322 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3326 adp = WK_ALLOCDIRECT(wk);
3327 adp->ad_state |= COMPLETE;
3328 handle_allocdirect_partdone(adp);
3332 aip = WK_ALLOCINDIR(wk);
3333 aip->ai_state |= COMPLETE;
3334 handle_allocindir_partdone(aip);
3338 indirdep = WK_INDIRDEP(wk);
3339 if (indirdep->ir_state & GOINGAWAY) {
3340 lk.lkt_held = NOHOLDER;
3341 panic("disk_write_complete: indirdep gone");
3343 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3344 FREE(indirdep->ir_saveddata, M_INDIRDEP);
3345 indirdep->ir_saveddata = 0;
3346 indirdep->ir_state &= ~UNDONE;
3347 indirdep->ir_state |= ATTACHED;
3348 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3349 handle_allocindir_partdone(aip);
3350 if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
3351 lk.lkt_held = NOHOLDER;
3352 panic("disk_write_complete: not gone");
3355 WORKLIST_INSERT(&reattach, wk);
3356 if ((bp->b_flags & B_DELWRI) == 0)
3357 stat_indir_blk_ptrs++;
3362 lk.lkt_held = NOHOLDER;
3363 panic("handle_disk_write_complete: Unknown type %s",
3364 TYPENAME(wk->wk_type));
3369 * Reattach any requests that must be redone.
3371 while ((wk = LIST_FIRST(&reattach)) != NULL) {
3372 WORKLIST_REMOVE(wk);
3373 WORKLIST_INSERT(&bp->b_dep, wk);
3376 if (lk.lkt_held != SPECIAL_FLAG)
3377 panic("softdep_disk_write_complete: lock lost");
3378 lk.lkt_held = NOHOLDER;
3383 * Called from within softdep_disk_write_complete above. Note that
3384 * this routine is always called from interrupt level with further
3385 * splbio interrupts blocked.
3388 handle_allocdirect_partdone(adp)
3389 struct allocdirect *adp; /* the completed allocdirect */
3391 struct allocdirect *listadp;
3392 struct inodedep *inodedep;
3395 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3397 if (adp->ad_buf != NULL) {
3398 lk.lkt_held = NOHOLDER;
3399 panic("handle_allocdirect_partdone: dangling dep");
3402 * The on-disk inode cannot claim to be any larger than the last
3403 * fragment that has been written. Otherwise, the on-disk inode
3404 * might have fragments that were not the last block in the file
3405 * which would corrupt the filesystem. Thus, we cannot free any
3406 * allocdirects after one whose ad_oldblkno claims a fragment as
3407 * these blocks must be rolled back to zero before writing the inode.
3408 * We check the currently active set of allocdirects in id_inoupdt.
3410 inodedep = adp->ad_inodedep;
3411 bsize = inodedep->id_fs->fs_bsize;
3412 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) {
3413 /* found our block */
3416 /* continue if ad_oldlbn is not a fragment */
3417 if (listadp->ad_oldsize == 0 ||
3418 listadp->ad_oldsize == bsize)
3420 /* hit a fragment */
3424 * If we have reached the end of the current list without
3425 * finding the just finished dependency, then it must be
3426 * on the future dependency list. Future dependencies cannot
3427 * be freed until they are moved to the current list.
3429 if (listadp == NULL) {
3431 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next)
3432 /* found our block */
3435 if (listadp == NULL) {
3436 lk.lkt_held = NOHOLDER;
3437 panic("handle_allocdirect_partdone: lost dep");
3443 * If we have found the just finished dependency, then free
3444 * it along with anything that follows it that is complete.
3446 for (; adp; adp = listadp) {
3447 listadp = TAILQ_NEXT(adp, ad_next);
3448 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3450 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
3455 * Called from within softdep_disk_write_complete above. Note that
3456 * this routine is always called from interrupt level with further
3457 * splbio interrupts blocked.
3460 handle_allocindir_partdone(aip)
3461 struct allocindir *aip; /* the completed allocindir */
3463 struct indirdep *indirdep;
3465 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
3467 if (aip->ai_buf != NULL) {
3468 lk.lkt_held = NOHOLDER;
3469 panic("handle_allocindir_partdone: dangling dependency");
3471 indirdep = aip->ai_indirdep;
3472 if (indirdep->ir_state & UNDONE) {
3473 LIST_REMOVE(aip, ai_next);
3474 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
3477 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
3479 LIST_REMOVE(aip, ai_next);
3480 if (aip->ai_freefrag != NULL)
3481 add_to_worklist(&aip->ai_freefrag->ff_list);
3482 WORKITEM_FREE(aip, D_ALLOCINDIR);
3486 * Called from within softdep_disk_write_complete above to restore
3487 * in-memory inode block contents to their most up-to-date state. Note
3488 * that this routine is always called from interrupt level with further
3489 * splbio interrupts blocked.
3492 handle_written_inodeblock(inodedep, bp)
3493 struct inodedep *inodedep;
3494 struct buf *bp; /* buffer containing the inode block */
3496 struct worklist *wk, *filefree;
3497 struct allocdirect *adp, *nextadp;
3501 if ((inodedep->id_state & IOSTARTED) == 0) {
3502 lk.lkt_held = NOHOLDER;
3503 panic("handle_written_inodeblock: not started");
3505 inodedep->id_state &= ~IOSTARTED;
3506 inodedep->id_state |= COMPLETE;
3507 dp = (struct dinode *)bp->b_data +
3508 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
3510 * If we had to rollback the inode allocation because of
3511 * bitmaps being incomplete, then simply restore it.
3512 * Keep the block dirty so that it will not be reclaimed until
3513 * all associated dependencies have been cleared and the
3514 * corresponding updates written to disk.
3516 if (inodedep->id_savedino != NULL) {
3517 *dp = *inodedep->id_savedino;
3518 FREE(inodedep->id_savedino, M_INODEDEP);
3519 inodedep->id_savedino = NULL;
3520 if ((bp->b_flags & B_DELWRI) == 0)
3521 stat_inode_bitmap++;
3526 * Roll forward anything that had to be rolled back before
3527 * the inode could be updated.
3530 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
3531 nextadp = TAILQ_NEXT(adp, ad_next);
3532 if (adp->ad_state & ATTACHED) {
3533 lk.lkt_held = NOHOLDER;
3534 panic("handle_written_inodeblock: new entry");
3536 if (adp->ad_lbn < NDADDR) {
3537 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) {
3538 lk.lkt_held = NOHOLDER;
3539 panic("%s: %s #%ld mismatch %d != %d",
3540 "handle_written_inodeblock",
3541 "direct pointer", adp->ad_lbn,
3542 dp->di_db[adp->ad_lbn], adp->ad_oldblkno);
3544 dp->di_db[adp->ad_lbn] = adp->ad_newblkno;
3546 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) {
3547 lk.lkt_held = NOHOLDER;
3548 panic("%s: %s #%ld allocated as %d",
3549 "handle_written_inodeblock",
3550 "indirect pointer", adp->ad_lbn - NDADDR,
3551 dp->di_ib[adp->ad_lbn - NDADDR]);
3553 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno;
3555 adp->ad_state &= ~UNDONE;
3556 adp->ad_state |= ATTACHED;
3559 if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
3560 stat_direct_blk_ptrs++;
3562 * Reset the file size to its most up-to-date value.
3564 if (inodedep->id_savedsize == -1) {
3565 lk.lkt_held = NOHOLDER;
3566 panic("handle_written_inodeblock: bad size");
3568 if (dp->di_size != inodedep->id_savedsize) {
3569 dp->di_size = inodedep->id_savedsize;
3572 inodedep->id_savedsize = -1;
3574 * If there were any rollbacks in the inode block, then it must be
3575 * marked dirty so that its will eventually get written back in
3581 * Process any allocdirects that completed during the update.
3583 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
3584 handle_allocdirect_partdone(adp);
3586 * Process deallocations that were held pending until the
3587 * inode had been written to disk. Freeing of the inode
3588 * is delayed until after all blocks have been freed to
3589 * avoid creation of new <vfsid, inum, lbn> triples
3590 * before the old ones have been deleted.
3593 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
3594 WORKLIST_REMOVE(wk);
3595 switch (wk->wk_type) {
3599 * We defer adding filefree to the worklist until
3600 * all other additions have been made to ensure
3601 * that it will be done after all the old blocks
3604 if (filefree != NULL) {
3605 lk.lkt_held = NOHOLDER;
3606 panic("handle_written_inodeblock: filefree");
3612 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
3616 diradd_inode_written(WK_DIRADD(wk), inodedep);
3620 wk->wk_state |= COMPLETE;
3621 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE)
3623 /* -- fall through -- */
3626 add_to_worklist(wk);
3630 lk.lkt_held = NOHOLDER;
3631 panic("handle_written_inodeblock: Unknown type %s",
3632 TYPENAME(wk->wk_type));
3636 if (filefree != NULL) {
3637 if (free_inodedep(inodedep) == 0) {
3638 lk.lkt_held = NOHOLDER;
3639 panic("handle_written_inodeblock: live inodedep");
3641 add_to_worklist(filefree);
3646 * If no outstanding dependencies, free it.
3648 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0)
3650 return (hadchanges);
3654 * Process a diradd entry after its dependent inode has been written.
3655 * This routine must be called with splbio interrupts blocked.
3658 diradd_inode_written(dap, inodedep)
3660 struct inodedep *inodedep;
3662 struct pagedep *pagedep;
3664 dap->da_state |= COMPLETE;
3665 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3666 if (dap->da_state & DIRCHG)
3667 pagedep = dap->da_previous->dm_pagedep;
3669 pagedep = dap->da_pagedep;
3670 LIST_REMOVE(dap, da_pdlist);
3671 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3673 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3677 * Handle the completion of a mkdir dependency.
3680 handle_written_mkdir(mkdir, type)
3681 struct mkdir *mkdir;
3685 struct pagedep *pagedep;
3687 if (mkdir->md_state != type) {
3688 lk.lkt_held = NOHOLDER;
3689 panic("handle_written_mkdir: bad type");
3691 dap = mkdir->md_diradd;
3692 dap->da_state &= ~type;
3693 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
3694 dap->da_state |= DEPCOMPLETE;
3695 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3696 if (dap->da_state & DIRCHG)
3697 pagedep = dap->da_previous->dm_pagedep;
3699 pagedep = dap->da_pagedep;
3700 LIST_REMOVE(dap, da_pdlist);
3701 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3703 LIST_REMOVE(mkdir, md_mkdirs);
3704 WORKITEM_FREE(mkdir, D_MKDIR);
3708 * Called from within softdep_disk_write_complete above.
3709 * A write operation was just completed. Removed inodes can
3710 * now be freed and associated block pointers may be committed.
3711 * Note that this routine is always called from interrupt level
3712 * with further splbio interrupts blocked.
3715 handle_written_filepage(pagedep, bp)
3716 struct pagedep *pagedep;
3717 struct buf *bp; /* buffer containing the written page */
3719 struct dirrem *dirrem;
3720 struct diradd *dap, *nextdap;
3724 if ((pagedep->pd_state & IOSTARTED) == 0) {
3725 lk.lkt_held = NOHOLDER;
3726 panic("handle_written_filepage: not started");
3728 pagedep->pd_state &= ~IOSTARTED;
3730 * Process any directory removals that have been committed.
3732 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
3733 LIST_REMOVE(dirrem, dm_next);
3734 dirrem->dm_dirinum = pagedep->pd_ino;
3735 add_to_worklist(&dirrem->dm_list);
3738 * Free any directory additions that have been committed.
3740 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
3743 * Uncommitted directory entries must be restored.
3745 for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
3746 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
3748 nextdap = LIST_NEXT(dap, da_pdlist);
3749 if (dap->da_state & ATTACHED) {
3750 lk.lkt_held = NOHOLDER;
3751 panic("handle_written_filepage: attached");
3753 ep = (struct direct *)
3754 ((char *)bp->b_data + dap->da_offset);
3755 ep->d_ino = dap->da_newinum;
3756 dap->da_state &= ~UNDONE;
3757 dap->da_state |= ATTACHED;
3760 * If the inode referenced by the directory has
3761 * been written out, then the dependency can be
3762 * moved to the pending list.
3764 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3765 LIST_REMOVE(dap, da_pdlist);
3766 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
3772 * If there were any rollbacks in the directory, then it must be
3773 * marked dirty so that its will eventually get written back in
3777 if ((bp->b_flags & B_DELWRI) == 0)
3782 * If no dependencies remain, the pagedep will be freed.
3783 * Otherwise it will remain to update the page before it
3784 * is written back to disk.
3786 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) {
3787 for (i = 0; i < DAHASHSZ; i++)
3788 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
3790 if (i == DAHASHSZ) {
3791 LIST_REMOVE(pagedep, pd_hash);
3792 WORKITEM_FREE(pagedep, D_PAGEDEP);
3800 * Writing back in-core inode structures.
3802 * The filesystem only accesses an inode's contents when it occupies an
3803 * "in-core" inode structure. These "in-core" structures are separate from
3804 * the page frames used to cache inode blocks. Only the latter are
3805 * transferred to/from the disk. So, when the updated contents of the
3806 * "in-core" inode structure are copied to the corresponding in-memory inode
3807 * block, the dependencies are also transferred. The following procedure is
3808 * called when copying a dirty "in-core" inode to a cached inode block.
3812 * Called when an inode is loaded from disk. If the effective link count
3813 * differed from the actual link count when it was last flushed, then we
3814 * need to ensure that the correct effective link count is put back.
3817 softdep_load_inodeblock(ip)
3818 struct inode *ip; /* the "in_core" copy of the inode */
3820 struct inodedep *inodedep;
3823 * Check for alternate nlink count.
3825 ip->i_effnlink = ip->i_nlink;
3827 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3831 ip->i_effnlink -= inodedep->id_nlinkdelta;
3836 * This routine is called just before the "in-core" inode
3837 * information is to be copied to the in-memory inode block.
3838 * Recall that an inode block contains several inodes. If
3839 * the force flag is set, then the dependencies will be
3840 * cleared so that the update can always be made. Note that
3841 * the buffer is locked when this routine is called, so we
3842 * will never be in the middle of writing the inode block
3846 softdep_update_inodeblock(ip, bp, waitfor)
3847 struct inode *ip; /* the "in_core" copy of the inode */
3848 struct buf *bp; /* the buffer containing the inode block */
3849 int waitfor; /* nonzero => update must be allowed */
3851 struct inodedep *inodedep;
3852 struct worklist *wk;
3856 * If the effective link count is not equal to the actual link
3857 * count, then we must track the difference in an inodedep while
3858 * the inode is (potentially) tossed out of the cache. Otherwise,
3859 * if there is no existing inodedep, then there are no dependencies
3863 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3865 if (ip->i_effnlink != ip->i_nlink)
3866 panic("softdep_update_inodeblock: bad link count");
3869 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
3871 panic("softdep_update_inodeblock: bad delta");
3874 * Changes have been initiated. Anything depending on these
3875 * changes cannot occur until this inode has been written.
3877 inodedep->id_state &= ~COMPLETE;
3878 if ((inodedep->id_state & ONWORKLIST) == 0)
3879 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
3881 * Any new dependencies associated with the incore inode must
3882 * now be moved to the list associated with the buffer holding
3883 * the in-memory copy of the inode. Once merged process any
3884 * allocdirects that are completed by the merger.
3886 merge_inode_lists(inodedep);
3887 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
3888 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
3890 * Now that the inode has been pushed into the buffer, the
3891 * operations dependent on the inode being written to disk
3892 * can be moved to the id_bufwait so that they will be
3893 * processed when the buffer I/O completes.
3895 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
3896 WORKLIST_REMOVE(wk);
3897 WORKLIST_INSERT(&inodedep->id_bufwait, wk);
3900 * Newly allocated inodes cannot be written until the bitmap
3901 * that allocates them have been written (indicated by
3902 * DEPCOMPLETE being set in id_state). If we are doing a
3903 * forced sync (e.g., an fsync on a file), we force the bitmap
3904 * to be written so that the update can be done.
3906 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
3910 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
3913 (error = VOP_BWRITE(inodedep->id_buf->b_vp, inodedep->id_buf)) != 0)
3914 softdep_error("softdep_update_inodeblock: bwrite", error);
3915 if ((inodedep->id_state & DEPCOMPLETE) == 0)
3916 panic("softdep_update_inodeblock: update failed");
3920 * Merge the new inode dependency list (id_newinoupdt) into the old
3921 * inode dependency list (id_inoupdt). This routine must be called
3922 * with splbio interrupts blocked.
3925 merge_inode_lists(inodedep)
3926 struct inodedep *inodedep;
3928 struct allocdirect *listadp, *newadp;
3930 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3931 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) {
3932 if (listadp->ad_lbn < newadp->ad_lbn) {
3933 listadp = TAILQ_NEXT(listadp, ad_next);
3936 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3937 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
3938 if (listadp->ad_lbn == newadp->ad_lbn) {
3939 allocdirect_merge(&inodedep->id_inoupdt, newadp,
3943 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
3945 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) {
3946 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
3947 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next);
3952 * If we are doing an fsync, then we must ensure that any directory
3953 * entries for the inode have been written after the inode gets to disk.
3957 struct vnode *vp; /* the "in_core" copy of the inode */
3959 struct inodedep *inodedep;
3960 struct pagedep *pagedep;
3961 struct worklist *wk;
3968 struct thread *td = curthread; /* XXX */
3969 int error, flushparent;
3976 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
3980 if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
3981 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
3982 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
3983 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) {
3985 panic("softdep_fsync: pending ops");
3987 for (error = 0, flushparent = 0; ; ) {
3988 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
3990 if (wk->wk_type != D_DIRADD) {
3992 panic("softdep_fsync: Unexpected type %s",
3993 TYPENAME(wk->wk_type));
3995 dap = WK_DIRADD(wk);
3997 * Flush our parent if this directory entry
3998 * has a MKDIR_PARENT dependency.
4000 if (dap->da_state & DIRCHG)
4001 pagedep = dap->da_previous->dm_pagedep;
4003 pagedep = dap->da_pagedep;
4004 mnt = pagedep->pd_mnt;
4005 parentino = pagedep->pd_ino;
4006 lbn = pagedep->pd_lbn;
4007 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) {
4009 panic("softdep_fsync: dirty");
4011 flushparent = dap->da_state & MKDIR_PARENT;
4013 * If we are being fsync'ed as part of vgone'ing this vnode,
4014 * then we will not be able to release and recover the
4015 * vnode below, so we just have to give up on writing its
4016 * directory entry out. It will eventually be written, just
4017 * not now, but then the user was not asking to have it
4018 * written, so we are not breaking any promises.
4020 if (vp->v_flag & VRECLAIMED)
4023 * We prevent deadlock by always fetching inodes from the
4024 * root, moving down the directory tree. Thus, when fetching
4025 * our parent directory, we must unlock ourselves before
4026 * requesting the lock on our parent. See the comment in
4027 * ufs_lookup for details on possible races.
4030 VOP_UNLOCK(vp, 0, td);
4031 error = VFS_VGET(mnt, parentino, &pvp);
4032 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
4036 if ((error = UFS_UPDATE(pvp, 1)) != 0) {
4042 * Flush directory page containing the inode's name.
4044 error = bread(pvp, lbn, blksize(fs, VTOI(pvp), lbn), &bp);
4046 error = VOP_BWRITE(bp->b_vp, bp);
4051 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
4059 * Flush all the dirty bitmaps associated with the block device
4060 * before flushing the rest of the dirty blocks so as to reduce
4061 * the number of dependencies that will have to be rolled back.
4063 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data);
4066 softdep_fsync_mountdev(vp)
4069 if (!vn_isdisk(vp, NULL))
4070 panic("softdep_fsync_mountdev: vnode not a disk");
4072 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
4073 softdep_fsync_mountdev_bp, NULL);
4074 drain_output(vp, 1);
4079 softdep_fsync_mountdev_bp(struct buf *bp, void *data)
4081 struct worklist *wk;
4084 * If it is already scheduled, skip to the next buffer.
4086 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
4088 if ((bp->b_flags & B_DELWRI) == 0) {
4090 panic("softdep_fsync_mountdev: not dirty");
4093 * We are only interested in bitmaps with outstanding
4096 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
4097 wk->wk_type != D_BMSAFEMAP ||
4098 (bp->b_xflags & BX_BKGRDINPROG)) {
4110 * This routine is called when we are trying to synchronously flush a
4111 * file. This routine must eliminate any filesystem metadata dependencies
4112 * so that the syncing routine can succeed by pushing the dirty blocks
4113 * associated with the file. If any I/O errors occur, they are returned.
4115 struct softdep_sync_metadata_info {
4120 static int softdep_sync_metadata_bp(struct buf *bp, void *data);
4123 softdep_sync_metadata(struct vnode *vp, struct thread *td)
4125 struct softdep_sync_metadata_info info;
4129 * Check whether this vnode is involved in a filesystem
4130 * that is doing soft dependency processing.
4132 if (!vn_isdisk(vp, NULL)) {
4133 if (!DOINGSOFTDEP(vp))
4136 if (vp->v_rdev->si_mountpoint == NULL ||
4137 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0)
4140 * Ensure that any direct block dependencies have been cleared.
4143 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
4148 * For most files, the only metadata dependencies are the
4149 * cylinder group maps that allocate their inode or blocks.
4150 * The block allocation dependencies can be found by traversing
4151 * the dependency lists for any buffers that remain on their
4152 * dirty buffer list. The inode allocation dependency will
4153 * be resolved when the inode is updated with MNT_WAIT.
4154 * This work is done in two passes. The first pass grabs most
4155 * of the buffers and begins asynchronously writing them. The
4156 * only way to wait for these asynchronous writes is to sleep
4157 * on the filesystem vnode which may stay busy for a long time
4158 * if the filesystem is active. So, instead, we make a second
4159 * pass over the dependencies blocking on each write. In the
4160 * usual case we will be blocking against a write that we
4161 * initiated, so when it is done the dependency will have been
4162 * resolved. Thus the second pass is expected to end quickly.
4164 waitfor = MNT_NOWAIT;
4167 * We must wait for any I/O in progress to finish so that
4168 * all potential buffers on the dirty list will be visible.
4170 drain_output(vp, 1);
4172 info.waitfor = waitfor;
4173 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
4174 softdep_sync_metadata_bp, &info);
4177 return(-error); /* error code */
4181 * The brief unlock is to allow any pent up dependency
4182 * processing to be done. Then proceed with the second pass.
4184 if (waitfor == MNT_NOWAIT) {
4192 * If we have managed to get rid of all the dirty buffers,
4193 * then we are done. For certain directories and block
4194 * devices, we may need to do further work.
4196 * We must wait for any I/O in progress to finish so that
4197 * all potential buffers on the dirty list will be visible.
4199 drain_output(vp, 1);
4200 if (RB_EMPTY(&vp->v_rbdirty_tree)) {
4207 * If we are trying to sync a block device, some of its buffers may
4208 * contain metadata that cannot be written until the contents of some
4209 * partially written files have been written to disk. The only easy
4210 * way to accomplish this is to sync the entire filesystem (luckily
4211 * this happens rarely).
4213 if (vn_isdisk(vp, NULL) &&
4215 vp->v_rdev->si_mountpoint && !VOP_ISLOCKED(vp, NULL) &&
4216 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT, td)) != 0)
4222 softdep_sync_metadata_bp(struct buf *bp, void *data)
4224 struct softdep_sync_metadata_info *info = data;
4225 struct pagedep *pagedep;
4226 struct allocdirect *adp;
4227 struct allocindir *aip;
4228 struct worklist *wk;
4233 if (getdirtybuf(&bp, MNT_WAIT) == 0)
4237 * As we hold the buffer locked, none of its dependencies
4240 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4241 switch (wk->wk_type) {
4244 adp = WK_ALLOCDIRECT(wk);
4245 if (adp->ad_state & DEPCOMPLETE)
4248 if (getdirtybuf(&nbp, info->waitfor) == 0)
4251 if (info->waitfor == MNT_NOWAIT) {
4253 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4262 aip = WK_ALLOCINDIR(wk);
4263 if (aip->ai_state & DEPCOMPLETE)
4266 if (getdirtybuf(&nbp, info->waitfor) == 0)
4269 if (info->waitfor == MNT_NOWAIT) {
4271 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4282 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
4283 if (aip->ai_state & DEPCOMPLETE)
4286 if (getdirtybuf(&nbp, MNT_WAIT) == 0)
4289 if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4300 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
4301 WK_INODEDEP(wk)->id_ino)) != 0) {
4311 * We are trying to sync a directory that may
4312 * have dependencies on both its own metadata
4313 * and/or dependencies on the inodes of any
4314 * recently allocated files. We walk its diradd
4315 * lists pushing out the associated inode.
4317 pagedep = WK_PAGEDEP(wk);
4318 for (i = 0; i < DAHASHSZ; i++) {
4319 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
4322 flush_pagedep_deps(info->vp,
4324 &pagedep->pd_diraddhd[i]))) {
4335 * This case should never happen if the vnode has
4336 * been properly sync'ed. However, if this function
4337 * is used at a place where the vnode has not yet
4338 * been sync'ed, this dependency can show up. So,
4339 * rather than panic, just flush it.
4341 nbp = WK_MKDIR(wk)->md_buf;
4342 if (getdirtybuf(&nbp, info->waitfor) == 0)
4345 if (info->waitfor == MNT_NOWAIT) {
4347 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4357 * This case should never happen if the vnode has
4358 * been properly sync'ed. However, if this function
4359 * is used at a place where the vnode has not yet
4360 * been sync'ed, this dependency can show up. So,
4361 * rather than panic, just flush it.
4363 nbp = WK_BMSAFEMAP(wk)->sm_buf;
4364 if (getdirtybuf(&nbp, info->waitfor) == 0)
4367 if (info->waitfor == MNT_NOWAIT) {
4369 } else if ((error = VOP_BWRITE(nbp->b_vp, nbp)) != 0) {
4379 panic("softdep_sync_metadata: Unknown type %s",
4380 TYPENAME(wk->wk_type));
4391 * Flush the dependencies associated with an inodedep.
4392 * Called with splbio blocked.
4395 flush_inodedep_deps(fs, ino)
4399 struct inodedep *inodedep;
4400 struct allocdirect *adp;
4405 * This work is done in two passes. The first pass grabs most
4406 * of the buffers and begins asynchronously writing them. The
4407 * only way to wait for these asynchronous writes is to sleep
4408 * on the filesystem vnode which may stay busy for a long time
4409 * if the filesystem is active. So, instead, we make a second
4410 * pass over the dependencies blocking on each write. In the
4411 * usual case we will be blocking against a write that we
4412 * initiated, so when it is done the dependency will have been
4413 * resolved. Thus the second pass is expected to end quickly.
4414 * We give a brief window at the top of the loop to allow
4415 * any pending I/O to complete.
4417 for (waitfor = MNT_NOWAIT; ; ) {
4420 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4422 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) {
4423 if (adp->ad_state & DEPCOMPLETE)
4426 if (getdirtybuf(&bp, waitfor) == 0) {
4427 if (waitfor == MNT_NOWAIT)
4432 if (waitfor == MNT_NOWAIT) {
4434 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) {
4443 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) {
4444 if (adp->ad_state & DEPCOMPLETE)
4447 if (getdirtybuf(&bp, waitfor) == 0) {
4448 if (waitfor == MNT_NOWAIT)
4453 if (waitfor == MNT_NOWAIT) {
4455 } else if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0) {
4465 * If pass2, we are done, otherwise do pass 2.
4467 if (waitfor == MNT_WAIT)
4472 * Try freeing inodedep in case all dependencies have been removed.
4474 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
4475 (void) free_inodedep(inodedep);
4480 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
4481 * Called with splbio blocked.
4484 flush_pagedep_deps(pvp, mp, diraddhdp)
4487 struct diraddhd *diraddhdp;
4489 struct thread *td = curthread; /* XXX */
4490 struct inodedep *inodedep;
4491 struct ufsmount *ump;
4494 int gotit, error = 0;
4499 while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
4501 * Flush ourselves if this directory entry
4502 * has a MKDIR_PARENT dependency.
4504 if (dap->da_state & MKDIR_PARENT) {
4506 if ((error = UFS_UPDATE(pvp, 1)) != 0)
4510 * If that cleared dependencies, go on to next.
4512 if (dap != LIST_FIRST(diraddhdp))
4514 if (dap->da_state & MKDIR_PARENT) {
4516 panic("flush_pagedep_deps: MKDIR_PARENT");
4520 * A newly allocated directory must have its "." and
4521 * ".." entries written out before its name can be
4522 * committed in its parent. We do not want or need
4523 * the full semantics of a synchronous VOP_FSYNC as
4524 * that may end up here again, once for each directory
4525 * level in the filesystem. Instead, we push the blocks
4526 * and wait for them to clear. We have to fsync twice
4527 * because the first call may choose to defer blocks
4528 * that still have dependencies, but deferral will
4529 * happen at most once.
4531 inum = dap->da_newinum;
4532 if (dap->da_state & MKDIR_BODY) {
4534 if ((error = VFS_VGET(mp, inum, &vp)) != 0)
4536 if ((error=VOP_FSYNC(vp, MNT_NOWAIT, td)) ||
4537 (error=VOP_FSYNC(vp, MNT_NOWAIT, td))) {
4541 drain_output(vp, 0);
4545 * If that cleared dependencies, go on to next.
4547 if (dap != LIST_FIRST(diraddhdp))
4549 if (dap->da_state & MKDIR_BODY) {
4551 panic("flush_pagedep_deps: MKDIR_BODY");
4555 * Flush the inode on which the directory entry depends.
4556 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
4557 * the only remaining dependency is that the updated inode
4558 * count must get pushed to disk. The inode has already
4559 * been pushed into its inode buffer (via VOP_UPDATE) at
4560 * the time of the reference count change. So we need only
4561 * locate that buffer, ensure that there will be no rollback
4562 * caused by a bitmap dependency, then write the inode buffer.
4564 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) {
4566 panic("flush_pagedep_deps: lost inode");
4569 * If the inode still has bitmap dependencies,
4570 * push them to disk.
4572 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4573 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4576 (error = VOP_BWRITE(inodedep->id_buf->b_vp,
4577 inodedep->id_buf)) != 0)
4580 if (dap != LIST_FIRST(diraddhdp))
4584 * If the inode is still sitting in a buffer waiting
4585 * to be written, push it to disk.
4588 if ((error = bread(ump->um_devvp,
4589 fsbtodb(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
4590 (int)ump->um_fs->fs_bsize, &bp)) != 0)
4592 if ((error = VOP_BWRITE(bp->b_vp, bp)) != 0)
4596 * If we have failed to get rid of all the dependencies
4597 * then something is seriously wrong.
4599 if (dap == LIST_FIRST(diraddhdp)) {
4601 panic("flush_pagedep_deps: flush failed");
4610 * A large burst of file addition or deletion activity can drive the
4611 * memory load excessively high. First attempt to slow things down
4612 * using the techniques below. If that fails, this routine requests
4613 * the offending operations to fall back to running synchronously
4614 * until the memory load returns to a reasonable level.
4617 softdep_slowdown(vp)
4620 int max_softdeps_hard;
4622 max_softdeps_hard = max_softdeps * 11 / 10;
4623 if (num_dirrem < max_softdeps_hard / 2 &&
4624 num_inodedep < max_softdeps_hard)
4626 stat_sync_limit_hit += 1;
4631 * If memory utilization has gotten too high, deliberately slow things
4632 * down and speed up the I/O processing.
4635 request_cleanup(resource, islocked)
4639 struct thread *td = curthread; /* XXX */
4642 * We never hold up the filesystem syncer process.
4644 if (td == filesys_syncer)
4647 * First check to see if the work list has gotten backlogged.
4648 * If it has, co-opt this process to help clean up two entries.
4649 * Because this process may hold inodes locked, we cannot
4650 * handle any remove requests that might block on a locked
4651 * inode as that could lead to deadlock.
4653 if (num_on_worklist > max_softdeps / 10) {
4656 process_worklist_item(NULL, LK_NOWAIT);
4657 process_worklist_item(NULL, LK_NOWAIT);
4658 stat_worklist_push += 2;
4665 * If we are resource constrained on inode dependencies, try
4666 * flushing some dirty inodes. Otherwise, we are constrained
4667 * by file deletions, so try accelerating flushes of directories
4668 * with removal dependencies. We would like to do the cleanup
4669 * here, but we probably hold an inode locked at this point and
4670 * that might deadlock against one that we try to clean. So,
4671 * the best that we can do is request the syncer daemon to do
4672 * the cleanup for us.
4677 stat_ino_limit_push += 1;
4678 req_clear_inodedeps += 1;
4679 stat_countp = &stat_ino_limit_hit;
4683 stat_blk_limit_push += 1;
4684 req_clear_remove += 1;
4685 stat_countp = &stat_blk_limit_hit;
4691 panic("request_cleanup: unknown type");
4694 * Hopefully the syncer daemon will catch up and awaken us.
4695 * We wait at most tickdelay before proceeding in any case.
4701 if (!callout_active(&handle))
4702 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2,
4704 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0,
4714 * Awaken processes pausing in request_cleanup and clear proc_waiting
4715 * to indicate that there is no longer a timer running.
4722 wakeup_one(&proc_waiting);
4723 if (proc_waiting > 0)
4724 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2,
4727 callout_deactivate(&handle);
4731 * Flush out a directory with at least one removal dependency in an effort to
4732 * reduce the number of dirrem, freefile, and freeblks dependency structures.
4735 clear_remove(struct thread *td)
4737 struct pagedep_hashhead *pagedephd;
4738 struct pagedep *pagedep;
4739 static int next = 0;
4746 for (cnt = 0; cnt < pagedep_hash; cnt++) {
4747 pagedephd = &pagedep_hashtbl[next++];
4748 if (next >= pagedep_hash)
4750 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
4751 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
4753 mp = pagedep->pd_mnt;
4754 ino = pagedep->pd_ino;
4756 if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4757 softdep_error("clear_remove: vget", error);
4760 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, td)))
4761 softdep_error("clear_remove: fsync", error);
4762 drain_output(vp, 0);
4771 * Clear out a block of dirty inodes in an effort to reduce
4772 * the number of inodedep dependency structures.
4774 struct clear_inodedeps_info {
4780 clear_inodedeps_mountlist_callback(struct mount *mp, void *data)
4782 struct clear_inodedeps_info *info = data;
4784 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) {
4792 clear_inodedeps(struct thread *td)
4794 struct clear_inodedeps_info info;
4795 struct inodedep_hashhead *inodedephd;
4796 struct inodedep *inodedep;
4797 static int next = 0;
4801 ino_t firstino, lastino, ino;
4805 * Pick a random inode dependency to be cleared.
4806 * We will then gather up all the inodes in its block
4807 * that have dependencies and flush them out.
4809 for (cnt = 0; cnt < inodedep_hash; cnt++) {
4810 inodedephd = &inodedep_hashtbl[next++];
4811 if (next >= inodedep_hash)
4813 if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
4816 if (inodedep == NULL) {
4821 * Ugly code to find mount point given pointer to superblock.
4823 fs = inodedep->id_fs;
4826 mountlist_scan(clear_inodedeps_mountlist_callback,
4827 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
4829 * Find the last inode in the block with dependencies.
4831 firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
4832 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
4833 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
4836 * Asynchronously push all but the last inode with dependencies.
4837 * Synchronously push the last inode with dependencies to ensure
4838 * that the inode block gets written to free up the inodedeps.
4840 for (ino = firstino; ino <= lastino; ino++) {
4841 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4844 if ((error = VFS_VGET(info.mp, ino, &vp)) != 0) {
4845 softdep_error("clear_inodedeps: vget", error);
4848 if (ino == lastino) {
4849 if ((error = VOP_FSYNC(vp, MNT_WAIT, td)))
4850 softdep_error("clear_inodedeps: fsync1", error);
4852 if ((error = VOP_FSYNC(vp, MNT_NOWAIT, td)))
4853 softdep_error("clear_inodedeps: fsync2", error);
4854 drain_output(vp, 0);
4863 * Function to determine if the buffer has outstanding dependencies
4864 * that will cause a roll-back if the buffer is written. If wantcount
4865 * is set, return number of dependencies, otherwise just yes or no.
4868 softdep_count_dependencies(bp, wantcount)
4872 struct worklist *wk;
4873 struct inodedep *inodedep;
4874 struct indirdep *indirdep;
4875 struct allocindir *aip;
4876 struct pagedep *pagedep;
4882 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4883 switch (wk->wk_type) {
4886 inodedep = WK_INODEDEP(wk);
4887 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4888 /* bitmap allocation dependency */
4893 if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
4894 /* direct block pointer dependency */
4902 indirdep = WK_INDIRDEP(wk);
4904 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
4905 /* indirect block pointer dependency */
4913 pagedep = WK_PAGEDEP(wk);
4914 for (i = 0; i < DAHASHSZ; i++) {
4916 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
4917 /* directory entry dependency */
4929 /* never a dependency on these blocks */
4934 panic("softdep_check_for_rollback: Unexpected type %s",
4935 TYPENAME(wk->wk_type));
4945 * Acquire exclusive access to a buffer.
4946 * Must be called with splbio blocked.
4947 * Return 1 if buffer was acquired.
4950 getdirtybuf(bpp, waitfor)
4958 if ((bp = *bpp) == NULL)
4960 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0) {
4961 if ((bp->b_xflags & BX_BKGRDINPROG) == 0)
4964 if (waitfor != MNT_WAIT)
4966 bp->b_xflags |= BX_BKGRDWAIT;
4967 interlocked_sleep(&lk, SLEEP, &bp->b_xflags, 0,
4971 if (waitfor != MNT_WAIT)
4973 error = interlocked_sleep(&lk, LOCKBUF, bp,
4974 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0);
4975 if (error != ENOLCK) {
4977 panic("getdirtybuf: inconsistent lock");
4980 if ((bp->b_flags & B_DELWRI) == 0) {
4989 * Wait for pending output on a vnode to complete.
4990 * Must be called with vnode locked.
4993 drain_output(vp, islocked)
5000 while (vp->v_numoutput) {
5001 vp->v_flag |= VBWAIT;
5002 interlocked_sleep(&lk, SLEEP, (caddr_t)&vp->v_numoutput,
5010 * Called whenever a buffer that is being invalidated or reallocated
5011 * contains dependencies. This should only happen if an I/O error has
5012 * occurred. The routine is called with the buffer locked.
5015 softdep_deallocate_dependencies(bp)
5019 if ((bp->b_flags & B_ERROR) == 0)
5020 panic("softdep_deallocate_dependencies: dangling deps");
5021 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error);
5022 panic("softdep_deallocate_dependencies: unrecovered I/O error");
5026 * Function to handle asynchronous write errors in the filesystem.
5029 softdep_error(func, error)
5034 /* XXX should do something better! */
5035 printf("%s: got error %d while accessing filesystem\n", func, error);