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 1.49 2006/09/05 00:55:51 dillon Exp $
44 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
53 #include <sys/param.h>
54 #include <sys/kernel.h>
55 #include <sys/systm.h>
57 #include <sys/malloc.h>
58 #include <sys/mount.h>
60 #include <sys/syslog.h>
61 #include <sys/vnode.h>
64 #include <machine/inttypes.h>
71 #include "ffs_extern.h"
72 #include "ufs_extern.h"
74 #include <sys/thread2.h>
77 * These definitions need to be adapted to the system to which
78 * this file is being ported.
81 * malloc types defined for the softdep system.
83 MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
84 MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
85 MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
86 MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
87 MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
88 MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
89 MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
90 MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
91 MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
92 MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
93 MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
94 MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
95 MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
97 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE)
102 #define D_BMSAFEMAP 3
103 #define D_ALLOCDIRECT 4
105 #define D_ALLOCINDIR 6
112 #define D_LAST D_DIRREM
115 * translate from workitem type to memory type
116 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
118 static struct malloc_type *memtype[] = {
134 #define DtoM(type) (memtype[type])
137 * Names of malloc types.
139 #define TYPENAME(type) \
140 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
142 * End system adaptaion definitions.
146 * Internal function prototypes.
148 static void softdep_error(char *, int);
149 static void drain_output(struct vnode *, int);
150 static int getdirtybuf(struct buf **, int);
151 static void clear_remove(struct thread *);
152 static void clear_inodedeps(struct thread *);
153 static int flush_pagedep_deps(struct vnode *, struct mount *,
155 static int flush_inodedep_deps(struct fs *, ino_t);
156 static int handle_written_filepage(struct pagedep *, struct buf *);
157 static void diradd_inode_written(struct diradd *, struct inodedep *);
158 static int handle_written_inodeblock(struct inodedep *, struct buf *);
159 static void handle_allocdirect_partdone(struct allocdirect *);
160 static void handle_allocindir_partdone(struct allocindir *);
161 static void initiate_write_filepage(struct pagedep *, struct buf *);
162 static void handle_written_mkdir(struct mkdir *, int);
163 static void initiate_write_inodeblock(struct inodedep *, struct buf *);
164 static void handle_workitem_freefile(struct freefile *);
165 static void handle_workitem_remove(struct dirrem *);
166 static struct dirrem *newdirrem(struct buf *, struct inode *,
167 struct inode *, int, struct dirrem **);
168 static void free_diradd(struct diradd *);
169 static void free_allocindir(struct allocindir *, struct inodedep *);
170 static int indir_trunc (struct inode *, off_t, int, ufs_lbn_t, long *);
171 static void deallocate_dependencies(struct buf *, struct inodedep *);
172 static void free_allocdirect(struct allocdirectlst *,
173 struct allocdirect *, int);
174 static int check_inode_unwritten(struct inodedep *);
175 static int free_inodedep(struct inodedep *);
176 static void handle_workitem_freeblocks(struct freeblks *);
177 static void merge_inode_lists(struct inodedep *);
178 static void setup_allocindir_phase2(struct buf *, struct inode *,
179 struct allocindir *);
180 static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t,
182 static void handle_workitem_freefrag(struct freefrag *);
183 static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long);
184 static void allocdirect_merge(struct allocdirectlst *,
185 struct allocdirect *, struct allocdirect *);
186 static struct bmsafemap *bmsafemap_lookup(struct buf *);
187 static int newblk_lookup(struct fs *, ufs_daddr_t, int,
189 static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
190 static int pagedep_lookup(struct inode *, ufs_lbn_t, int,
192 static 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 || !vn_islocked(vp))
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)
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)) != 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)) != 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);
825 error = VOP_FSYNC(devvp, MNT_WAIT);
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 * Helper routine for pagedep_lookup()
887 pagedep_find(struct pagedep_hashhead *pagedephd, ino_t ino, ufs_lbn_t lbn,
890 struct pagedep *pagedep;
892 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
893 if (ino == pagedep->pd_ino &&
894 lbn == pagedep->pd_lbn &&
895 mp == pagedep->pd_mnt) {
903 * Look up a pagedep. Return 1 if found, 0 if not found.
904 * If not found, allocate if DEPALLOC flag is passed.
905 * Found or allocated entry is returned in pagedeppp.
906 * This routine must be called with splbio interrupts blocked.
909 pagedep_lookup(ip, lbn, flags, pagedeppp)
913 struct pagedep **pagedeppp;
915 struct pagedep *pagedep;
916 struct pagedep_hashhead *pagedephd;
921 if (lk.lkt_held == NOHOLDER)
922 panic("pagedep_lookup: lock not held");
924 mp = ITOV(ip)->v_mount;
925 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
927 *pagedeppp = pagedep_find(pagedephd, ip->i_number, lbn, mp);
930 if ((flags & DEPALLOC) == 0)
932 if (sema_get(&pagedep_in_progress, &lk) == 0) {
936 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
937 M_SOFTDEP_FLAGS | M_ZERO);
939 if (pagedep_find(pagedephd, ip->i_number, lbn, mp)) {
940 printf("pagedep_lookup: blocking race avoided\n");
942 sema_release(&pagedep_in_progress);
943 kfree(pagedep, M_PAGEDEP);
947 pagedep->pd_list.wk_type = D_PAGEDEP;
948 pagedep->pd_mnt = mp;
949 pagedep->pd_ino = ip->i_number;
950 pagedep->pd_lbn = lbn;
951 LIST_INIT(&pagedep->pd_dirremhd);
952 LIST_INIT(&pagedep->pd_pendinghd);
953 for (i = 0; i < DAHASHSZ; i++)
954 LIST_INIT(&pagedep->pd_diraddhd[i]);
956 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
957 sema_release(&pagedep_in_progress);
958 *pagedeppp = pagedep;
963 * Structures and routines associated with inodedep caching.
965 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
966 static u_long inodedep_hash; /* size of hash table - 1 */
967 static long num_inodedep; /* number of inodedep allocated */
968 #define INODEDEP_HASH(fs, inum) \
969 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
970 static struct sema inodedep_in_progress;
973 * Helper routine for inodedep_lookup()
977 inodedep_find(struct inodedep_hashhead *inodedephd, struct fs *fs, ino_t inum)
979 struct inodedep *inodedep;
981 LIST_FOREACH(inodedep, inodedephd, id_hash) {
982 if (inum == inodedep->id_ino && fs == inodedep->id_fs)
989 * Look up a inodedep. Return 1 if found, 0 if not found.
990 * If not found, allocate if DEPALLOC flag is passed.
991 * Found or allocated entry is returned in inodedeppp.
992 * This routine must be called with splbio interrupts blocked.
995 inodedep_lookup(fs, inum, flags, inodedeppp)
999 struct inodedep **inodedeppp;
1001 struct inodedep *inodedep;
1002 struct inodedep_hashhead *inodedephd;
1006 if (lk.lkt_held == NOHOLDER)
1007 panic("inodedep_lookup: lock not held");
1010 inodedephd = INODEDEP_HASH(fs, inum);
1012 *inodedeppp = inodedep_find(inodedephd, fs, inum);
1015 if ((flags & DEPALLOC) == 0)
1018 * If we are over our limit, try to improve the situation.
1020 if (num_inodedep > max_softdeps && firsttry &&
1021 speedup_syncer() == 0 && (flags & NODELAY) == 0 &&
1022 request_cleanup(FLUSH_INODES, 1)) {
1026 if (sema_get(&inodedep_in_progress, &lk) == 0) {
1030 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1031 M_INODEDEP, M_SOFTDEP_FLAGS | M_ZERO);
1032 if (inodedep_find(inodedephd, fs, inum)) {
1033 printf("inodedep_lookup: blocking race avoided\n");
1035 sema_release(&inodedep_in_progress);
1036 kfree(inodedep, M_INODEDEP);
1039 inodedep->id_list.wk_type = D_INODEDEP;
1040 inodedep->id_fs = fs;
1041 inodedep->id_ino = inum;
1042 inodedep->id_state = ALLCOMPLETE;
1043 inodedep->id_nlinkdelta = 0;
1044 inodedep->id_savedino = NULL;
1045 inodedep->id_savedsize = -1;
1046 inodedep->id_buf = NULL;
1047 LIST_INIT(&inodedep->id_pendinghd);
1048 LIST_INIT(&inodedep->id_inowait);
1049 LIST_INIT(&inodedep->id_bufwait);
1050 TAILQ_INIT(&inodedep->id_inoupdt);
1051 TAILQ_INIT(&inodedep->id_newinoupdt);
1054 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1055 sema_release(&inodedep_in_progress);
1056 *inodedeppp = inodedep;
1061 * Structures and routines associated with newblk caching.
1063 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1064 u_long newblk_hash; /* size of hash table - 1 */
1065 #define NEWBLK_HASH(fs, inum) \
1066 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1067 static struct sema newblk_in_progress;
1070 * Helper routine for newblk_lookup()
1074 newblk_find(struct newblk_hashhead *newblkhd, struct fs *fs,
1075 ufs_daddr_t newblkno)
1077 struct newblk *newblk;
1079 LIST_FOREACH(newblk, newblkhd, nb_hash) {
1080 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1087 * Look up a newblk. Return 1 if found, 0 if not found.
1088 * If not found, allocate if DEPALLOC flag is passed.
1089 * Found or allocated entry is returned in newblkpp.
1092 newblk_lookup(fs, newblkno, flags, newblkpp)
1094 ufs_daddr_t newblkno;
1096 struct newblk **newblkpp;
1098 struct newblk *newblk;
1099 struct newblk_hashhead *newblkhd;
1101 newblkhd = NEWBLK_HASH(fs, newblkno);
1103 *newblkpp = newblk_find(newblkhd, fs, newblkno);
1106 if ((flags & DEPALLOC) == 0)
1108 if (sema_get(&newblk_in_progress, 0) == 0)
1110 MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1111 M_NEWBLK, M_SOFTDEP_FLAGS | M_ZERO);
1113 if (newblk_find(newblkhd, fs, newblkno)) {
1114 printf("newblk_lookup: blocking race avoided\n");
1115 sema_release(&pagedep_in_progress);
1116 kfree(newblk, M_NEWBLK);
1119 newblk->nb_state = 0;
1121 newblk->nb_newblkno = newblkno;
1122 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1123 sema_release(&newblk_in_progress);
1129 * Executed during filesystem system initialization before
1130 * mounting any filesystems.
1133 softdep_initialize()
1135 callout_init(&handle);
1136 bioops = softdep_bioops; /* XXX hack */
1138 LIST_INIT(&mkdirlisthd);
1139 LIST_INIT(&softdep_workitem_pending);
1140 max_softdeps = min(desiredvnodes * 8,
1141 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep)));
1142 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1144 sema_init(&pagedep_in_progress, "pagedep", 0, 0);
1145 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1146 sema_init(&inodedep_in_progress, "inodedep", 0, 0);
1147 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1148 sema_init(&newblk_in_progress, "newblk", 0, 0);
1152 * Called at mount time to notify the dependency code that a
1153 * filesystem wishes to use it.
1156 softdep_mount(devvp, mp, fs)
1157 struct vnode *devvp;
1161 struct csum cstotal;
1166 mp->mnt_flag &= ~MNT_ASYNC;
1167 mp->mnt_flag |= MNT_SOFTDEP;
1169 * When doing soft updates, the counters in the
1170 * superblock may have gotten out of sync, so we have
1171 * to scan the cylinder groups and recalculate them.
1173 if (fs->fs_clean != 0)
1175 bzero(&cstotal, sizeof cstotal);
1176 for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1177 if ((error = bread(devvp, fsbtodoff(fs, cgtod(fs, cyl)),
1178 fs->fs_cgsize, &bp)) != 0) {
1182 cgp = (struct cg *)bp->b_data;
1183 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1184 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1185 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1186 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1187 fs->fs_cs(fs, cyl) = cgp->cg_cs;
1191 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1192 printf("ffs_mountfs: superblock updated for soft updates\n");
1194 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1199 * Protecting the freemaps (or bitmaps).
1201 * To eliminate the need to execute fsck before mounting a filesystem
1202 * after a power failure, one must (conservatively) guarantee that the
1203 * on-disk copy of the bitmaps never indicate that a live inode or block is
1204 * free. So, when a block or inode is allocated, the bitmap should be
1205 * updated (on disk) before any new pointers. When a block or inode is
1206 * freed, the bitmap should not be updated until all pointers have been
1207 * reset. The latter dependency is handled by the delayed de-allocation
1208 * approach described below for block and inode de-allocation. The former
1209 * dependency is handled by calling the following procedure when a block or
1210 * inode is allocated. When an inode is allocated an "inodedep" is created
1211 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1212 * Each "inodedep" is also inserted into the hash indexing structure so
1213 * that any additional link additions can be made dependent on the inode
1216 * The ufs filesystem maintains a number of free block counts (e.g., per
1217 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1218 * in addition to the bitmaps. These counts are used to improve efficiency
1219 * during allocation and therefore must be consistent with the bitmaps.
1220 * There is no convenient way to guarantee post-crash consistency of these
1221 * counts with simple update ordering, for two main reasons: (1) The counts
1222 * and bitmaps for a single cylinder group block are not in the same disk
1223 * sector. If a disk write is interrupted (e.g., by power failure), one may
1224 * be written and the other not. (2) Some of the counts are located in the
1225 * superblock rather than the cylinder group block. So, we focus our soft
1226 * updates implementation on protecting the bitmaps. When mounting a
1227 * filesystem, we recompute the auxiliary counts from the bitmaps.
1231 * Called just after updating the cylinder group block to allocate an inode.
1234 softdep_setup_inomapdep(bp, ip, newinum)
1235 struct buf *bp; /* buffer for cylgroup block with inode map */
1236 struct inode *ip; /* inode related to allocation */
1237 ino_t newinum; /* new inode number being allocated */
1239 struct inodedep *inodedep;
1240 struct bmsafemap *bmsafemap;
1243 * Create a dependency for the newly allocated inode.
1244 * Panic if it already exists as something is seriously wrong.
1245 * Otherwise add it to the dependency list for the buffer holding
1246 * the cylinder group map from which it was allocated.
1249 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1251 panic("softdep_setup_inomapdep: found inode");
1253 inodedep->id_buf = bp;
1254 inodedep->id_state &= ~DEPCOMPLETE;
1255 bmsafemap = bmsafemap_lookup(bp);
1256 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1261 * Called just after updating the cylinder group block to
1262 * allocate block or fragment.
1265 softdep_setup_blkmapdep(bp, fs, newblkno)
1266 struct buf *bp; /* buffer for cylgroup block with block map */
1267 struct fs *fs; /* filesystem doing allocation */
1268 ufs_daddr_t newblkno; /* number of newly allocated block */
1270 struct newblk *newblk;
1271 struct bmsafemap *bmsafemap;
1274 * Create a dependency for the newly allocated block.
1275 * Add it to the dependency list for the buffer holding
1276 * the cylinder group map from which it was allocated.
1278 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1279 panic("softdep_setup_blkmapdep: found block");
1281 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1282 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1287 * Find the bmsafemap associated with a cylinder group buffer.
1288 * If none exists, create one. The buffer must be locked when
1289 * this routine is called and this routine must be called with
1290 * splbio interrupts blocked.
1292 static struct bmsafemap *
1293 bmsafemap_lookup(bp)
1296 struct bmsafemap *bmsafemap;
1297 struct worklist *wk;
1300 if (lk.lkt_held == NOHOLDER)
1301 panic("bmsafemap_lookup: lock not held");
1303 LIST_FOREACH(wk, &bp->b_dep, wk_list)
1304 if (wk->wk_type == D_BMSAFEMAP)
1305 return (WK_BMSAFEMAP(wk));
1307 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1308 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1309 bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1310 bmsafemap->sm_list.wk_state = 0;
1311 bmsafemap->sm_buf = bp;
1312 LIST_INIT(&bmsafemap->sm_allocdirecthd);
1313 LIST_INIT(&bmsafemap->sm_allocindirhd);
1314 LIST_INIT(&bmsafemap->sm_inodedephd);
1315 LIST_INIT(&bmsafemap->sm_newblkhd);
1317 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1322 * Direct block allocation dependencies.
1324 * When a new block is allocated, the corresponding disk locations must be
1325 * initialized (with zeros or new data) before the on-disk inode points to
1326 * them. Also, the freemap from which the block was allocated must be
1327 * updated (on disk) before the inode's pointer. These two dependencies are
1328 * independent of each other and are needed for all file blocks and indirect
1329 * blocks that are pointed to directly by the inode. Just before the
1330 * "in-core" version of the inode is updated with a newly allocated block
1331 * number, a procedure (below) is called to setup allocation dependency
1332 * structures. These structures are removed when the corresponding
1333 * dependencies are satisfied or when the block allocation becomes obsolete
1334 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1335 * fragment that gets upgraded). All of these cases are handled in
1336 * procedures described later.
1338 * When a file extension causes a fragment to be upgraded, either to a larger
1339 * fragment or to a full block, the on-disk location may change (if the
1340 * previous fragment could not simply be extended). In this case, the old
1341 * fragment must be de-allocated, but not until after the inode's pointer has
1342 * been updated. In most cases, this is handled by later procedures, which
1343 * will construct a "freefrag" structure to be added to the workitem queue
1344 * when the inode update is complete (or obsolete). The main exception to
1345 * this is when an allocation occurs while a pending allocation dependency
1346 * (for the same block pointer) remains. This case is handled in the main
1347 * allocation dependency setup procedure by immediately freeing the
1348 * unreferenced fragments.
1351 softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1352 struct inode *ip; /* inode to which block is being added */
1353 ufs_lbn_t lbn; /* block pointer within inode */
1354 ufs_daddr_t newblkno; /* disk block number being added */
1355 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */
1356 long newsize; /* size of new block */
1357 long oldsize; /* size of new block */
1358 struct buf *bp; /* bp for allocated block */
1360 struct allocdirect *adp, *oldadp;
1361 struct allocdirectlst *adphead;
1362 struct bmsafemap *bmsafemap;
1363 struct inodedep *inodedep;
1364 struct pagedep *pagedep;
1365 struct newblk *newblk;
1367 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1368 M_ALLOCDIRECT, M_SOFTDEP_FLAGS);
1369 bzero(adp, sizeof(struct allocdirect));
1370 adp->ad_list.wk_type = D_ALLOCDIRECT;
1372 adp->ad_newblkno = newblkno;
1373 adp->ad_oldblkno = oldblkno;
1374 adp->ad_newsize = newsize;
1375 adp->ad_oldsize = oldsize;
1376 adp->ad_state = ATTACHED;
1377 if (newblkno == oldblkno)
1378 adp->ad_freefrag = NULL;
1380 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1382 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1383 panic("softdep_setup_allocdirect: lost block");
1386 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1387 adp->ad_inodedep = inodedep;
1389 if (newblk->nb_state == DEPCOMPLETE) {
1390 adp->ad_state |= DEPCOMPLETE;
1393 bmsafemap = newblk->nb_bmsafemap;
1394 adp->ad_buf = bmsafemap->sm_buf;
1395 LIST_REMOVE(newblk, nb_deps);
1396 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1398 LIST_REMOVE(newblk, nb_hash);
1399 FREE(newblk, M_NEWBLK);
1401 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1402 if (lbn >= NDADDR) {
1403 /* allocating an indirect block */
1404 if (oldblkno != 0) {
1406 panic("softdep_setup_allocdirect: non-zero indir");
1410 * Allocating a direct block.
1412 * If we are allocating a directory block, then we must
1413 * allocate an associated pagedep to track additions and
1416 if ((ip->i_mode & IFMT) == IFDIR &&
1417 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1418 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1421 * The list of allocdirects must be kept in sorted and ascending
1422 * order so that the rollback routines can quickly determine the
1423 * first uncommitted block (the size of the file stored on disk
1424 * ends at the end of the lowest committed fragment, or if there
1425 * are no fragments, at the end of the highest committed block).
1426 * Since files generally grow, the typical case is that the new
1427 * block is to be added at the end of the list. We speed this
1428 * special case by checking against the last allocdirect in the
1429 * list before laboriously traversing the list looking for the
1432 adphead = &inodedep->id_newinoupdt;
1433 oldadp = TAILQ_LAST(adphead, allocdirectlst);
1434 if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1435 /* insert at end of list */
1436 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1437 if (oldadp != NULL && oldadp->ad_lbn == lbn)
1438 allocdirect_merge(adphead, adp, oldadp);
1442 TAILQ_FOREACH(oldadp, adphead, ad_next) {
1443 if (oldadp->ad_lbn >= lbn)
1446 if (oldadp == NULL) {
1448 panic("softdep_setup_allocdirect: lost entry");
1450 /* insert in middle of list */
1451 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1452 if (oldadp->ad_lbn == lbn)
1453 allocdirect_merge(adphead, adp, oldadp);
1458 * Replace an old allocdirect dependency with a newer one.
1459 * This routine must be called with splbio interrupts blocked.
1462 allocdirect_merge(adphead, newadp, oldadp)
1463 struct allocdirectlst *adphead; /* head of list holding allocdirects */
1464 struct allocdirect *newadp; /* allocdirect being added */
1465 struct allocdirect *oldadp; /* existing allocdirect being checked */
1467 struct freefrag *freefrag;
1470 if (lk.lkt_held == NOHOLDER)
1471 panic("allocdirect_merge: lock not held");
1473 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1474 newadp->ad_oldsize != oldadp->ad_newsize ||
1475 newadp->ad_lbn >= NDADDR) {
1477 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1478 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1481 newadp->ad_oldblkno = oldadp->ad_oldblkno;
1482 newadp->ad_oldsize = oldadp->ad_oldsize;
1484 * If the old dependency had a fragment to free or had never
1485 * previously had a block allocated, then the new dependency
1486 * can immediately post its freefrag and adopt the old freefrag.
1487 * This action is done by swapping the freefrag dependencies.
1488 * The new dependency gains the old one's freefrag, and the
1489 * old one gets the new one and then immediately puts it on
1490 * the worklist when it is freed by free_allocdirect. It is
1491 * not possible to do this swap when the old dependency had a
1492 * non-zero size but no previous fragment to free. This condition
1493 * arises when the new block is an extension of the old block.
1494 * Here, the first part of the fragment allocated to the new
1495 * dependency is part of the block currently claimed on disk by
1496 * the old dependency, so cannot legitimately be freed until the
1497 * conditions for the new dependency are fulfilled.
1499 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1500 freefrag = newadp->ad_freefrag;
1501 newadp->ad_freefrag = oldadp->ad_freefrag;
1502 oldadp->ad_freefrag = freefrag;
1504 free_allocdirect(adphead, oldadp, 0);
1508 * Allocate a new freefrag structure if needed.
1510 static struct freefrag *
1511 newfreefrag(ip, blkno, size)
1516 struct freefrag *freefrag;
1522 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1523 panic("newfreefrag: frag size");
1524 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1525 M_FREEFRAG, M_SOFTDEP_FLAGS);
1526 freefrag->ff_list.wk_type = D_FREEFRAG;
1527 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */
1528 freefrag->ff_inum = ip->i_number;
1529 freefrag->ff_fs = fs;
1530 freefrag->ff_devvp = ip->i_devvp;
1531 freefrag->ff_blkno = blkno;
1532 freefrag->ff_fragsize = size;
1537 * This workitem de-allocates fragments that were replaced during
1538 * file block allocation.
1541 handle_workitem_freefrag(freefrag)
1542 struct freefrag *freefrag;
1546 tip.i_fs = freefrag->ff_fs;
1547 tip.i_devvp = freefrag->ff_devvp;
1548 tip.i_dev = freefrag->ff_devvp->v_rdev;
1549 tip.i_number = freefrag->ff_inum;
1550 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */
1551 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1552 FREE(freefrag, M_FREEFRAG);
1556 * Indirect block allocation dependencies.
1558 * The same dependencies that exist for a direct block also exist when
1559 * a new block is allocated and pointed to by an entry in a block of
1560 * indirect pointers. The undo/redo states described above are also
1561 * used here. Because an indirect block contains many pointers that
1562 * may have dependencies, a second copy of the entire in-memory indirect
1563 * block is kept. The buffer cache copy is always completely up-to-date.
1564 * The second copy, which is used only as a source for disk writes,
1565 * contains only the safe pointers (i.e., those that have no remaining
1566 * update dependencies). The second copy is freed when all pointers
1567 * are safe. The cache is not allowed to replace indirect blocks with
1568 * pending update dependencies. If a buffer containing an indirect
1569 * block with dependencies is written, these routines will mark it
1570 * dirty again. It can only be successfully written once all the
1571 * dependencies are removed. The ffs_fsync routine in conjunction with
1572 * softdep_sync_metadata work together to get all the dependencies
1573 * removed so that a file can be successfully written to disk. Three
1574 * procedures are used when setting up indirect block pointer
1575 * dependencies. The division is necessary because of the organization
1576 * of the "balloc" routine and because of the distinction between file
1577 * pages and file metadata blocks.
1581 * Allocate a new allocindir structure.
1583 static struct allocindir *
1584 newallocindir(ip, ptrno, newblkno, oldblkno)
1585 struct inode *ip; /* inode for file being extended */
1586 int ptrno; /* offset of pointer in indirect block */
1587 ufs_daddr_t newblkno; /* disk block number being added */
1588 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1590 struct allocindir *aip;
1592 MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1593 M_ALLOCINDIR, M_SOFTDEP_FLAGS);
1594 bzero(aip, sizeof(struct allocindir));
1595 aip->ai_list.wk_type = D_ALLOCINDIR;
1596 aip->ai_state = ATTACHED;
1597 aip->ai_offset = ptrno;
1598 aip->ai_newblkno = newblkno;
1599 aip->ai_oldblkno = oldblkno;
1600 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1605 * Called just before setting an indirect block pointer
1606 * to a newly allocated file page.
1609 softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1610 struct inode *ip; /* inode for file being extended */
1611 ufs_lbn_t lbn; /* allocated block number within file */
1612 struct buf *bp; /* buffer with indirect blk referencing page */
1613 int ptrno; /* offset of pointer in indirect block */
1614 ufs_daddr_t newblkno; /* disk block number being added */
1615 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1616 struct buf *nbp; /* buffer holding allocated page */
1618 struct allocindir *aip;
1619 struct pagedep *pagedep;
1621 aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1624 * If we are allocating a directory page, then we must
1625 * allocate an associated pagedep to track additions and
1628 if ((ip->i_mode & IFMT) == IFDIR &&
1629 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1630 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1631 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1633 setup_allocindir_phase2(bp, ip, aip);
1637 * Called just before setting an indirect block pointer to a
1638 * newly allocated indirect block.
1641 softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1642 struct buf *nbp; /* newly allocated indirect block */
1643 struct inode *ip; /* inode for file being extended */
1644 struct buf *bp; /* indirect block referencing allocated block */
1645 int ptrno; /* offset of pointer in indirect block */
1646 ufs_daddr_t newblkno; /* disk block number being added */
1648 struct allocindir *aip;
1650 aip = newallocindir(ip, ptrno, newblkno, 0);
1652 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1654 setup_allocindir_phase2(bp, ip, aip);
1658 * Called to finish the allocation of the "aip" allocated
1659 * by one of the two routines above.
1662 setup_allocindir_phase2(bp, ip, aip)
1663 struct buf *bp; /* in-memory copy of the indirect block */
1664 struct inode *ip; /* inode for file being extended */
1665 struct allocindir *aip; /* allocindir allocated by the above routines */
1667 struct worklist *wk;
1668 struct indirdep *indirdep, *newindirdep;
1669 struct bmsafemap *bmsafemap;
1670 struct allocindir *oldaip;
1671 struct freefrag *freefrag;
1672 struct newblk *newblk;
1674 if (bp->b_loffset >= 0)
1675 panic("setup_allocindir_phase2: not indir blk");
1676 for (indirdep = NULL, newindirdep = NULL; ; ) {
1678 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1679 if (wk->wk_type != D_INDIRDEP)
1681 indirdep = WK_INDIRDEP(wk);
1684 if (indirdep == NULL && newindirdep) {
1685 indirdep = newindirdep;
1686 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1691 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1693 panic("setup_allocindir: lost block");
1695 if (newblk->nb_state == DEPCOMPLETE) {
1696 aip->ai_state |= DEPCOMPLETE;
1699 bmsafemap = newblk->nb_bmsafemap;
1700 aip->ai_buf = bmsafemap->sm_buf;
1701 LIST_REMOVE(newblk, nb_deps);
1702 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1705 LIST_REMOVE(newblk, nb_hash);
1706 FREE(newblk, M_NEWBLK);
1707 aip->ai_indirdep = indirdep;
1709 * Check to see if there is an existing dependency
1710 * for this block. If there is, merge the old
1711 * dependency into the new one.
1713 if (aip->ai_oldblkno == 0)
1717 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1718 if (oldaip->ai_offset == aip->ai_offset)
1720 if (oldaip != NULL) {
1721 if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1723 panic("setup_allocindir_phase2: blkno");
1725 aip->ai_oldblkno = oldaip->ai_oldblkno;
1726 freefrag = oldaip->ai_freefrag;
1727 oldaip->ai_freefrag = aip->ai_freefrag;
1728 aip->ai_freefrag = freefrag;
1729 free_allocindir(oldaip, NULL);
1731 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1732 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1733 [aip->ai_offset] = aip->ai_oldblkno;
1738 * Avoid any possibility of data corruption by
1739 * ensuring that our old version is thrown away.
1741 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
1742 brelse(newindirdep->ir_savebp);
1743 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1747 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1748 M_INDIRDEP, M_SOFTDEP_FLAGS);
1749 newindirdep->ir_list.wk_type = D_INDIRDEP;
1750 newindirdep->ir_state = ATTACHED;
1751 LIST_INIT(&newindirdep->ir_deplisthd);
1752 LIST_INIT(&newindirdep->ir_donehd);
1753 if (bp->b_bio2.bio_offset == NOOFFSET) {
1754 VOP_BMAP(bp->b_vp, bp->b_bio1.bio_offset,
1755 NULL, &bp->b_bio2.bio_offset,
1758 KKASSERT(bp->b_bio2.bio_offset != NOOFFSET);
1759 newindirdep->ir_savebp = getblk(ip->i_devvp,
1760 bp->b_bio2.bio_offset,
1761 bp->b_bcount, 0, 0);
1762 BUF_KERNPROC(newindirdep->ir_savebp);
1763 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1768 * Block de-allocation dependencies.
1770 * When blocks are de-allocated, the on-disk pointers must be nullified before
1771 * the blocks are made available for use by other files. (The true
1772 * requirement is that old pointers must be nullified before new on-disk
1773 * pointers are set. We chose this slightly more stringent requirement to
1774 * reduce complexity.) Our implementation handles this dependency by updating
1775 * the inode (or indirect block) appropriately but delaying the actual block
1776 * de-allocation (i.e., freemap and free space count manipulation) until
1777 * after the updated versions reach stable storage. After the disk is
1778 * updated, the blocks can be safely de-allocated whenever it is convenient.
1779 * This implementation handles only the common case of reducing a file's
1780 * length to zero. Other cases are handled by the conventional synchronous
1783 * The ffs implementation with which we worked double-checks
1784 * the state of the block pointers and file size as it reduces
1785 * a file's length. Some of this code is replicated here in our
1786 * soft updates implementation. The freeblks->fb_chkcnt field is
1787 * used to transfer a part of this information to the procedure
1788 * that eventually de-allocates the blocks.
1790 * This routine should be called from the routine that shortens
1791 * a file's length, before the inode's size or block pointers
1792 * are modified. It will save the block pointer information for
1793 * later release and zero the inode so that the calling routine
1796 struct softdep_setup_freeblocks_info {
1801 static int softdep_setup_freeblocks_bp(struct buf *bp, void *data);
1804 softdep_setup_freeblocks(ip, length)
1805 struct inode *ip; /* The inode whose length is to be reduced */
1806 off_t length; /* The new length for the file */
1808 struct softdep_setup_freeblocks_info info;
1809 struct freeblks *freeblks;
1810 struct inodedep *inodedep;
1811 struct allocdirect *adp;
1815 int i, error, delay;
1820 panic("softde_setup_freeblocks: non-zero length");
1821 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1822 M_FREEBLKS, M_SOFTDEP_FLAGS);
1823 bzero(freeblks, sizeof(struct freeblks));
1824 freeblks->fb_list.wk_type = D_FREEBLKS;
1825 freeblks->fb_state = ATTACHED;
1826 freeblks->fb_uid = ip->i_uid;
1827 freeblks->fb_previousinum = ip->i_number;
1828 freeblks->fb_devvp = ip->i_devvp;
1829 freeblks->fb_fs = fs;
1830 freeblks->fb_oldsize = ip->i_size;
1831 freeblks->fb_newsize = length;
1832 freeblks->fb_chkcnt = ip->i_blocks;
1833 for (i = 0; i < NDADDR; i++) {
1834 freeblks->fb_dblks[i] = ip->i_db[i];
1837 for (i = 0; i < NIADDR; i++) {
1838 freeblks->fb_iblks[i] = ip->i_ib[i];
1844 * Push the zero'ed inode to to its disk buffer so that we are free
1845 * to delete its dependencies below. Once the dependencies are gone
1846 * the buffer can be safely released.
1848 if ((error = bread(ip->i_devvp,
1849 fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)),
1850 (int)fs->fs_bsize, &bp)) != 0)
1851 softdep_error("softdep_setup_freeblocks", error);
1852 *((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
1855 * Find and eliminate any inode dependencies.
1858 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1859 if ((inodedep->id_state & IOSTARTED) != 0) {
1861 panic("softdep_setup_freeblocks: inode busy");
1864 * Add the freeblks structure to the list of operations that
1865 * must await the zero'ed inode being written to disk. If we
1866 * still have a bitmap dependency (delay == 0), then the inode
1867 * has never been written to disk, so we can process the
1868 * freeblks below once we have deleted the dependencies.
1870 delay = (inodedep->id_state & DEPCOMPLETE);
1872 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1874 * Because the file length has been truncated to zero, any
1875 * pending block allocation dependency structures associated
1876 * with this inode are obsolete and can simply be de-allocated.
1877 * We must first merge the two dependency lists to get rid of
1878 * any duplicate freefrag structures, then purge the merged list.
1880 merge_inode_lists(inodedep);
1881 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1882 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
1886 * We must wait for any I/O in progress to finish so that
1887 * all potential buffers on the dirty list will be visible.
1888 * Once they are all there, walk the list and get rid of
1893 drain_output(vp, 1);
1898 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
1899 softdep_setup_freeblocks_bp, &info);
1900 } while (count != 0);
1901 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
1902 (void)free_inodedep(inodedep);
1905 freeblks->fb_state |= DEPCOMPLETE;
1907 * If the inode with zeroed block pointers is now on disk
1908 * we can start freeing blocks. Add freeblks to the worklist
1909 * instead of calling handle_workitem_freeblocks directly as
1910 * it is more likely that additional IO is needed to complete
1911 * the request here than in the !delay case.
1913 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
1914 add_to_worklist(&freeblks->fb_list);
1919 * If the inode has never been written to disk (delay == 0),
1920 * then we can process the freeblks now that we have deleted
1924 handle_workitem_freeblocks(freeblks);
1928 softdep_setup_freeblocks_bp(struct buf *bp, void *data)
1930 struct softdep_setup_freeblocks_info *info = data;
1931 struct inodedep *inodedep;
1933 if (getdirtybuf(&bp, MNT_WAIT) == 0) {
1934 printf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp);
1937 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) {
1938 printf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp);
1942 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep);
1943 deallocate_dependencies(bp, inodedep);
1944 bp->b_flags |= B_INVAL | B_NOCACHE;
1952 * Reclaim any dependency structures from a buffer that is about to
1953 * be reallocated to a new vnode. The buffer must be locked, thus,
1954 * no I/O completion operations can occur while we are manipulating
1955 * its associated dependencies. The mutex is held so that other I/O's
1956 * associated with related dependencies do not occur.
1959 deallocate_dependencies(bp, inodedep)
1961 struct inodedep *inodedep;
1963 struct worklist *wk;
1964 struct indirdep *indirdep;
1965 struct allocindir *aip;
1966 struct pagedep *pagedep;
1967 struct dirrem *dirrem;
1971 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1972 switch (wk->wk_type) {
1975 indirdep = WK_INDIRDEP(wk);
1977 * None of the indirect pointers will ever be visible,
1978 * so they can simply be tossed. GOINGAWAY ensures
1979 * that allocated pointers will be saved in the buffer
1980 * cache until they are freed. Note that they will
1981 * only be able to be found by their physical address
1982 * since the inode mapping the logical address will
1983 * be gone. The save buffer used for the safe copy
1984 * was allocated in setup_allocindir_phase2 using
1985 * the physical address so it could be used for this
1986 * purpose. Hence we swap the safe copy with the real
1987 * copy, allowing the safe copy to be freed and holding
1988 * on to the real copy for later use in indir_trunc.
1990 * NOTE: ir_savebp is relative to the block device
1991 * so b_bio1 contains the device block number.
1993 if (indirdep->ir_state & GOINGAWAY) {
1995 panic("deallocate_dependencies: already gone");
1997 indirdep->ir_state |= GOINGAWAY;
1998 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
1999 free_allocindir(aip, inodedep);
2000 if (bp->b_bio1.bio_offset >= 0 ||
2001 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) {
2003 panic("deallocate_dependencies: not indir");
2005 bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2007 WORKLIST_REMOVE(wk);
2008 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
2012 pagedep = WK_PAGEDEP(wk);
2014 * None of the directory additions will ever be
2015 * visible, so they can simply be tossed.
2017 for (i = 0; i < DAHASHSZ; i++)
2019 LIST_FIRST(&pagedep->pd_diraddhd[i])))
2021 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2024 * Copy any directory remove dependencies to the list
2025 * to be processed after the zero'ed inode is written.
2026 * If the inode has already been written, then they
2027 * can be dumped directly onto the work list.
2029 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2030 LIST_REMOVE(dirrem, dm_next);
2031 dirrem->dm_dirinum = pagedep->pd_ino;
2032 if (inodedep == NULL ||
2033 (inodedep->id_state & ALLCOMPLETE) ==
2035 add_to_worklist(&dirrem->dm_list);
2037 WORKLIST_INSERT(&inodedep->id_bufwait,
2040 WORKLIST_REMOVE(&pagedep->pd_list);
2041 LIST_REMOVE(pagedep, pd_hash);
2042 WORKITEM_FREE(pagedep, D_PAGEDEP);
2046 free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2052 panic("deallocate_dependencies: Unexpected type %s",
2053 TYPENAME(wk->wk_type));
2058 panic("deallocate_dependencies: Unknown type %s",
2059 TYPENAME(wk->wk_type));
2066 * Free an allocdirect. Generate a new freefrag work request if appropriate.
2067 * This routine must be called with splbio interrupts blocked.
2070 free_allocdirect(adphead, adp, delay)
2071 struct allocdirectlst *adphead;
2072 struct allocdirect *adp;
2077 if (lk.lkt_held == NOHOLDER)
2078 panic("free_allocdirect: lock not held");
2080 if ((adp->ad_state & DEPCOMPLETE) == 0)
2081 LIST_REMOVE(adp, ad_deps);
2082 TAILQ_REMOVE(adphead, adp, ad_next);
2083 if ((adp->ad_state & COMPLETE) == 0)
2084 WORKLIST_REMOVE(&adp->ad_list);
2085 if (adp->ad_freefrag != NULL) {
2087 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2088 &adp->ad_freefrag->ff_list);
2090 add_to_worklist(&adp->ad_freefrag->ff_list);
2092 WORKITEM_FREE(adp, D_ALLOCDIRECT);
2096 * Prepare an inode to be freed. The actual free operation is not
2097 * done until the zero'ed inode has been written to disk.
2100 softdep_freefile(pvp, ino, mode)
2105 struct inode *ip = VTOI(pvp);
2106 struct inodedep *inodedep;
2107 struct freefile *freefile;
2110 * This sets up the inode de-allocation dependency.
2112 MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2113 M_FREEFILE, M_SOFTDEP_FLAGS);
2114 freefile->fx_list.wk_type = D_FREEFILE;
2115 freefile->fx_list.wk_state = 0;
2116 freefile->fx_mode = mode;
2117 freefile->fx_oldinum = ino;
2118 freefile->fx_devvp = ip->i_devvp;
2119 freefile->fx_fs = ip->i_fs;
2122 * If the inodedep does not exist, then the zero'ed inode has
2123 * been written to disk. If the allocated inode has never been
2124 * written to disk, then the on-disk inode is zero'ed. In either
2125 * case we can free the file immediately.
2128 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2129 check_inode_unwritten(inodedep)) {
2131 handle_workitem_freefile(freefile);
2134 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2139 * Check to see if an inode has never been written to disk. If
2140 * so free the inodedep and return success, otherwise return failure.
2141 * This routine must be called with splbio interrupts blocked.
2143 * If we still have a bitmap dependency, then the inode has never
2144 * been written to disk. Drop the dependency as it is no longer
2145 * necessary since the inode is being deallocated. We set the
2146 * ALLCOMPLETE flags since the bitmap now properly shows that the
2147 * inode is not allocated. Even if the inode is actively being
2148 * written, it has been rolled back to its zero'ed state, so we
2149 * are ensured that a zero inode is what is on the disk. For short
2150 * lived files, this change will usually result in removing all the
2151 * dependencies from the inode so that it can be freed immediately.
2154 check_inode_unwritten(inodedep)
2155 struct inodedep *inodedep;
2158 if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2159 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2160 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2161 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2162 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2163 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2164 inodedep->id_nlinkdelta != 0)
2168 * Another process might be in initiate_write_inodeblock
2169 * trying to allocate memory without holding "Softdep Lock".
2171 if ((inodedep->id_state & IOSTARTED) != 0 &&
2172 inodedep->id_savedino == NULL)
2175 inodedep->id_state |= ALLCOMPLETE;
2176 LIST_REMOVE(inodedep, id_deps);
2177 inodedep->id_buf = NULL;
2178 if (inodedep->id_state & ONWORKLIST)
2179 WORKLIST_REMOVE(&inodedep->id_list);
2180 if (inodedep->id_savedino != NULL) {
2181 FREE(inodedep->id_savedino, M_INODEDEP);
2182 inodedep->id_savedino = NULL;
2184 if (free_inodedep(inodedep) == 0) {
2186 panic("check_inode_unwritten: busy inode");
2192 * Try to free an inodedep structure. Return 1 if it could be freed.
2195 free_inodedep(inodedep)
2196 struct inodedep *inodedep;
2199 if ((inodedep->id_state & ONWORKLIST) != 0 ||
2200 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2201 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2202 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2203 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2204 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2205 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2206 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
2208 LIST_REMOVE(inodedep, id_hash);
2209 WORKITEM_FREE(inodedep, D_INODEDEP);
2215 * This workitem routine performs the block de-allocation.
2216 * The workitem is added to the pending list after the updated
2217 * inode block has been written to disk. As mentioned above,
2218 * checks regarding the number of blocks de-allocated (compared
2219 * to the number of blocks allocated for the file) are also
2220 * performed in this function.
2223 handle_workitem_freeblocks(freeblks)
2224 struct freeblks *freeblks;
2229 int i, level, bsize;
2230 long nblocks, blocksreleased = 0;
2231 int error, allerror = 0;
2232 ufs_lbn_t baselbns[NIADDR], tmpval;
2234 tip.i_number = freeblks->fb_previousinum;
2235 tip.i_devvp = freeblks->fb_devvp;
2236 tip.i_dev = freeblks->fb_devvp->v_rdev;
2237 tip.i_fs = freeblks->fb_fs;
2238 tip.i_size = freeblks->fb_oldsize;
2239 tip.i_uid = freeblks->fb_uid;
2240 fs = freeblks->fb_fs;
2242 baselbns[0] = NDADDR;
2243 for (i = 1; i < NIADDR; i++) {
2244 tmpval *= NINDIR(fs);
2245 baselbns[i] = baselbns[i - 1] + tmpval;
2247 nblocks = btodb(fs->fs_bsize);
2250 * Indirect blocks first.
2252 for (level = (NIADDR - 1); level >= 0; level--) {
2253 if ((bn = freeblks->fb_iblks[level]) == 0)
2255 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level,
2256 baselbns[level], &blocksreleased)) == 0)
2258 ffs_blkfree(&tip, bn, fs->fs_bsize);
2259 blocksreleased += nblocks;
2262 * All direct blocks or frags.
2264 for (i = (NDADDR - 1); i >= 0; i--) {
2265 if ((bn = freeblks->fb_dblks[i]) == 0)
2267 bsize = blksize(fs, &tip, i);
2268 ffs_blkfree(&tip, bn, bsize);
2269 blocksreleased += btodb(bsize);
2273 if (freeblks->fb_chkcnt != blocksreleased)
2274 printf("handle_workitem_freeblocks: block count\n");
2276 softdep_error("handle_workitem_freeblks", allerror);
2277 #endif /* DIAGNOSTIC */
2278 WORKITEM_FREE(freeblks, D_FREEBLKS);
2282 * Release blocks associated with the inode ip and stored in the indirect
2283 * block at doffset. If level is greater than SINGLE, the block is an
2284 * indirect block and recursive calls to indirtrunc must be used to
2285 * cleanse other indirect blocks.
2288 indir_trunc(ip, doffset, level, lbn, countp)
2299 struct worklist *wk;
2300 struct indirdep *indirdep;
2301 int i, lbnadd, nblocks;
2302 int error, allerror = 0;
2306 for (i = level; i > 0; i--)
2307 lbnadd *= NINDIR(fs);
2309 * Get buffer of block pointers to be freed. This routine is not
2310 * called until the zero'ed inode has been written, so it is safe
2311 * to free blocks as they are encountered. Because the inode has
2312 * been zero'ed, calls to bmap on these blocks will fail. So, we
2313 * have to use the on-disk address and the block device for the
2314 * filesystem to look them up. If the file was deleted before its
2315 * indirect blocks were all written to disk, the routine that set
2316 * us up (deallocate_dependencies) will have arranged to leave
2317 * a complete copy of the indirect block in memory for our use.
2318 * Otherwise we have to read the blocks in from the disk.
2321 if ((bp = findblk(ip->i_devvp, doffset)) != NULL &&
2322 (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2324 * bp must be ir_savebp, which is held locked for our use.
2326 if (wk->wk_type != D_INDIRDEP ||
2327 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2328 (indirdep->ir_state & GOINGAWAY) == 0) {
2330 panic("indir_trunc: lost indirdep");
2332 WORKLIST_REMOVE(wk);
2333 WORKITEM_FREE(indirdep, D_INDIRDEP);
2334 if (LIST_FIRST(&bp->b_dep) != NULL) {
2336 panic("indir_trunc: dangling dep");
2341 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp);
2346 * Recursively free indirect blocks.
2348 bap = (ufs_daddr_t *)bp->b_data;
2349 nblocks = btodb(fs->fs_bsize);
2350 for (i = NINDIR(fs) - 1; i >= 0; i--) {
2351 if ((nb = bap[i]) == 0)
2354 if ((error = indir_trunc(ip, fsbtodoff(fs, nb),
2355 level - 1, lbn + (i * lbnadd), countp)) != 0)
2358 ffs_blkfree(ip, nb, fs->fs_bsize);
2361 bp->b_flags |= B_INVAL | B_NOCACHE;
2367 * Free an allocindir.
2368 * This routine must be called with splbio interrupts blocked.
2371 free_allocindir(aip, inodedep)
2372 struct allocindir *aip;
2373 struct inodedep *inodedep;
2375 struct freefrag *freefrag;
2378 if (lk.lkt_held == NOHOLDER)
2379 panic("free_allocindir: lock not held");
2381 if ((aip->ai_state & DEPCOMPLETE) == 0)
2382 LIST_REMOVE(aip, ai_deps);
2383 if (aip->ai_state & ONWORKLIST)
2384 WORKLIST_REMOVE(&aip->ai_list);
2385 LIST_REMOVE(aip, ai_next);
2386 if ((freefrag = aip->ai_freefrag) != NULL) {
2387 if (inodedep == NULL)
2388 add_to_worklist(&freefrag->ff_list);
2390 WORKLIST_INSERT(&inodedep->id_bufwait,
2391 &freefrag->ff_list);
2393 WORKITEM_FREE(aip, D_ALLOCINDIR);
2397 * Directory entry addition dependencies.
2399 * When adding a new directory entry, the inode (with its incremented link
2400 * count) must be written to disk before the directory entry's pointer to it.
2401 * Also, if the inode is newly allocated, the corresponding freemap must be
2402 * updated (on disk) before the directory entry's pointer. These requirements
2403 * are met via undo/redo on the directory entry's pointer, which consists
2404 * simply of the inode number.
2406 * As directory entries are added and deleted, the free space within a
2407 * directory block can become fragmented. The ufs filesystem will compact
2408 * a fragmented directory block to make space for a new entry. When this
2409 * occurs, the offsets of previously added entries change. Any "diradd"
2410 * dependency structures corresponding to these entries must be updated with
2415 * This routine is called after the in-memory inode's link
2416 * count has been incremented, but before the directory entry's
2417 * pointer to the inode has been set.
2420 softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp)
2421 struct buf *bp; /* buffer containing directory block */
2422 struct inode *dp; /* inode for directory */
2423 off_t diroffset; /* offset of new entry in directory */
2424 ino_t newinum; /* inode referenced by new directory entry */
2425 struct buf *newdirbp; /* non-NULL => contents of new mkdir */
2427 int offset; /* offset of new entry within directory block */
2428 ufs_lbn_t lbn; /* block in directory containing new entry */
2431 struct pagedep *pagedep;
2432 struct inodedep *inodedep;
2433 struct mkdir *mkdir1, *mkdir2;
2436 * Whiteouts have no dependencies.
2438 if (newinum == WINO) {
2439 if (newdirbp != NULL)
2445 lbn = lblkno(fs, diroffset);
2446 offset = blkoff(fs, diroffset);
2447 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2449 bzero(dap, sizeof(struct diradd));
2450 dap->da_list.wk_type = D_DIRADD;
2451 dap->da_offset = offset;
2452 dap->da_newinum = newinum;
2453 dap->da_state = ATTACHED;
2454 if (newdirbp == NULL) {
2455 dap->da_state |= DEPCOMPLETE;
2458 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2459 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2461 mkdir1->md_list.wk_type = D_MKDIR;
2462 mkdir1->md_state = MKDIR_BODY;
2463 mkdir1->md_diradd = dap;
2464 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2466 mkdir2->md_list.wk_type = D_MKDIR;
2467 mkdir2->md_state = MKDIR_PARENT;
2468 mkdir2->md_diradd = dap;
2470 * Dependency on "." and ".." being written to disk.
2472 mkdir1->md_buf = newdirbp;
2474 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2475 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2479 * Dependency on link count increase for parent directory
2482 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2483 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2484 dap->da_state &= ~MKDIR_PARENT;
2485 WORKITEM_FREE(mkdir2, D_MKDIR);
2487 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2488 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2492 * Link into parent directory pagedep to await its being written.
2494 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2495 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2496 dap->da_pagedep = pagedep;
2497 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2500 * Link into its inodedep. Put it on the id_bufwait list if the inode
2501 * is not yet written. If it is written, do the post-inode write
2502 * processing to put it on the id_pendinghd list.
2504 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2505 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2506 diradd_inode_written(dap, inodedep);
2508 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2513 * This procedure is called to change the offset of a directory
2514 * entry when compacting a directory block which must be owned
2515 * exclusively by the caller. Note that the actual entry movement
2516 * must be done in this procedure to ensure that no I/O completions
2517 * occur while the move is in progress.
2520 softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2521 struct inode *dp; /* inode for directory */
2522 caddr_t base; /* address of dp->i_offset */
2523 caddr_t oldloc; /* address of old directory location */
2524 caddr_t newloc; /* address of new directory location */
2525 int entrysize; /* size of directory entry */
2527 int offset, oldoffset, newoffset;
2528 struct pagedep *pagedep;
2533 lbn = lblkno(dp->i_fs, dp->i_offset);
2534 offset = blkoff(dp->i_fs, dp->i_offset);
2535 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2537 oldoffset = offset + (oldloc - base);
2538 newoffset = offset + (newloc - base);
2540 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2541 if (dap->da_offset != oldoffset)
2543 dap->da_offset = newoffset;
2544 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2546 LIST_REMOVE(dap, da_pdlist);
2547 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2553 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2554 if (dap->da_offset == oldoffset) {
2555 dap->da_offset = newoffset;
2561 bcopy(oldloc, newloc, entrysize);
2566 * Free a diradd dependency structure. This routine must be called
2567 * with splbio interrupts blocked.
2573 struct dirrem *dirrem;
2574 struct pagedep *pagedep;
2575 struct inodedep *inodedep;
2576 struct mkdir *mkdir, *nextmd;
2579 if (lk.lkt_held == NOHOLDER)
2580 panic("free_diradd: lock not held");
2582 WORKLIST_REMOVE(&dap->da_list);
2583 LIST_REMOVE(dap, da_pdlist);
2584 if ((dap->da_state & DIRCHG) == 0) {
2585 pagedep = dap->da_pagedep;
2587 dirrem = dap->da_previous;
2588 pagedep = dirrem->dm_pagedep;
2589 dirrem->dm_dirinum = pagedep->pd_ino;
2590 add_to_worklist(&dirrem->dm_list);
2592 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2594 (void) free_inodedep(inodedep);
2595 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2596 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2597 nextmd = LIST_NEXT(mkdir, md_mkdirs);
2598 if (mkdir->md_diradd != dap)
2600 dap->da_state &= ~mkdir->md_state;
2601 WORKLIST_REMOVE(&mkdir->md_list);
2602 LIST_REMOVE(mkdir, md_mkdirs);
2603 WORKITEM_FREE(mkdir, D_MKDIR);
2605 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2607 panic("free_diradd: unfound ref");
2610 WORKITEM_FREE(dap, D_DIRADD);
2614 * Directory entry removal dependencies.
2616 * When removing a directory entry, the entry's inode pointer must be
2617 * zero'ed on disk before the corresponding inode's link count is decremented
2618 * (possibly freeing the inode for re-use). This dependency is handled by
2619 * updating the directory entry but delaying the inode count reduction until
2620 * after the directory block has been written to disk. After this point, the
2621 * inode count can be decremented whenever it is convenient.
2625 * This routine should be called immediately after removing
2626 * a directory entry. The inode's link count should not be
2627 * decremented by the calling procedure -- the soft updates
2628 * code will do this task when it is safe.
2631 softdep_setup_remove(bp, dp, ip, isrmdir)
2632 struct buf *bp; /* buffer containing directory block */
2633 struct inode *dp; /* inode for the directory being modified */
2634 struct inode *ip; /* inode for directory entry being removed */
2635 int isrmdir; /* indicates if doing RMDIR */
2637 struct dirrem *dirrem, *prevdirrem;
2640 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2642 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2645 * If the COMPLETE flag is clear, then there were no active
2646 * entries and we want to roll back to a zeroed entry until
2647 * the new inode is committed to disk. If the COMPLETE flag is
2648 * set then we have deleted an entry that never made it to
2649 * disk. If the entry we deleted resulted from a name change,
2650 * then the old name still resides on disk. We cannot delete
2651 * its inode (returned to us in prevdirrem) until the zeroed
2652 * directory entry gets to disk. The new inode has never been
2653 * referenced on the disk, so can be deleted immediately.
2655 if ((dirrem->dm_state & COMPLETE) == 0) {
2656 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2660 if (prevdirrem != NULL)
2661 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2662 prevdirrem, dm_next);
2663 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2665 handle_workitem_remove(dirrem);
2670 * Allocate a new dirrem if appropriate and return it along with
2671 * its associated pagedep. Called without a lock, returns with lock.
2673 static long num_dirrem; /* number of dirrem allocated */
2674 static struct dirrem *
2675 newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2676 struct buf *bp; /* buffer containing directory block */
2677 struct inode *dp; /* inode for the directory being modified */
2678 struct inode *ip; /* inode for directory entry being removed */
2679 int isrmdir; /* indicates if doing RMDIR */
2680 struct dirrem **prevdirremp; /* previously referenced inode, if any */
2685 struct dirrem *dirrem;
2686 struct pagedep *pagedep;
2689 * Whiteouts have no deletion dependencies.
2692 panic("newdirrem: whiteout");
2694 * If we are over our limit, try to improve the situation.
2695 * Limiting the number of dirrem structures will also limit
2696 * the number of freefile and freeblks structures.
2698 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2699 (void) request_cleanup(FLUSH_REMOVE, 0);
2701 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
2702 M_DIRREM, M_SOFTDEP_FLAGS);
2703 bzero(dirrem, sizeof(struct dirrem));
2704 dirrem->dm_list.wk_type = D_DIRREM;
2705 dirrem->dm_state = isrmdir ? RMDIR : 0;
2706 dirrem->dm_mnt = ITOV(ip)->v_mount;
2707 dirrem->dm_oldinum = ip->i_number;
2708 *prevdirremp = NULL;
2711 lbn = lblkno(dp->i_fs, dp->i_offset);
2712 offset = blkoff(dp->i_fs, dp->i_offset);
2713 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2714 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2715 dirrem->dm_pagedep = pagedep;
2717 * Check for a diradd dependency for the same directory entry.
2718 * If present, then both dependencies become obsolete and can
2719 * be de-allocated. Check for an entry on both the pd_dirraddhd
2720 * list and the pd_pendinghd list.
2723 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
2724 if (dap->da_offset == offset)
2728 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
2729 if (dap->da_offset == offset)
2735 * Must be ATTACHED at this point.
2737 if ((dap->da_state & ATTACHED) == 0) {
2739 panic("newdirrem: not ATTACHED");
2741 if (dap->da_newinum != ip->i_number) {
2743 panic("newdirrem: inum %"PRId64" should be %"PRId64,
2744 ip->i_number, dap->da_newinum);
2747 * If we are deleting a changed name that never made it to disk,
2748 * then return the dirrem describing the previous inode (which
2749 * represents the inode currently referenced from this entry on disk).
2751 if ((dap->da_state & DIRCHG) != 0) {
2752 *prevdirremp = dap->da_previous;
2753 dap->da_state &= ~DIRCHG;
2754 dap->da_pagedep = pagedep;
2757 * We are deleting an entry that never made it to disk.
2758 * Mark it COMPLETE so we can delete its inode immediately.
2760 dirrem->dm_state |= COMPLETE;
2766 * Directory entry change dependencies.
2768 * Changing an existing directory entry requires that an add operation
2769 * be completed first followed by a deletion. The semantics for the addition
2770 * are identical to the description of adding a new entry above except
2771 * that the rollback is to the old inode number rather than zero. Once
2772 * the addition dependency is completed, the removal is done as described
2773 * in the removal routine above.
2777 * This routine should be called immediately after changing
2778 * a directory entry. The inode's link count should not be
2779 * decremented by the calling procedure -- the soft updates
2780 * code will perform this task when it is safe.
2783 softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
2784 struct buf *bp; /* buffer containing directory block */
2785 struct inode *dp; /* inode for the directory being modified */
2786 struct inode *ip; /* inode for directory entry being removed */
2787 ino_t newinum; /* new inode number for changed entry */
2788 int isrmdir; /* indicates if doing RMDIR */
2791 struct diradd *dap = NULL;
2792 struct dirrem *dirrem, *prevdirrem;
2793 struct pagedep *pagedep;
2794 struct inodedep *inodedep;
2796 offset = blkoff(dp->i_fs, dp->i_offset);
2799 * Whiteouts do not need diradd dependencies.
2801 if (newinum != WINO) {
2802 MALLOC(dap, struct diradd *, sizeof(struct diradd),
2803 M_DIRADD, M_SOFTDEP_FLAGS);
2804 bzero(dap, sizeof(struct diradd));
2805 dap->da_list.wk_type = D_DIRADD;
2806 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2807 dap->da_offset = offset;
2808 dap->da_newinum = newinum;
2812 * Allocate a new dirrem and ACQUIRE_LOCK.
2814 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2815 pagedep = dirrem->dm_pagedep;
2817 * The possible values for isrmdir:
2818 * 0 - non-directory file rename
2819 * 1 - directory rename within same directory
2820 * inum - directory rename to new directory of given inode number
2821 * When renaming to a new directory, we are both deleting and
2822 * creating a new directory entry, so the link count on the new
2823 * directory should not change. Thus we do not need the followup
2824 * dirrem which is usually done in handle_workitem_remove. We set
2825 * the DIRCHG flag to tell handle_workitem_remove to skip the
2829 dirrem->dm_state |= DIRCHG;
2832 * Whiteouts have no additional dependencies,
2833 * so just put the dirrem on the correct list.
2835 if (newinum == WINO) {
2836 if ((dirrem->dm_state & COMPLETE) == 0) {
2837 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2840 dirrem->dm_dirinum = pagedep->pd_ino;
2841 add_to_worklist(&dirrem->dm_list);
2848 * If the COMPLETE flag is clear, then there were no active
2849 * entries and we want to roll back to the previous inode until
2850 * the new inode is committed to disk. If the COMPLETE flag is
2851 * set, then we have deleted an entry that never made it to disk.
2852 * If the entry we deleted resulted from a name change, then the old
2853 * inode reference still resides on disk. Any rollback that we do
2854 * needs to be to that old inode (returned to us in prevdirrem). If
2855 * the entry we deleted resulted from a create, then there is
2856 * no entry on the disk, so we want to roll back to zero rather
2857 * than the uncommitted inode. In either of the COMPLETE cases we
2858 * want to immediately free the unwritten and unreferenced inode.
2860 if ((dirrem->dm_state & COMPLETE) == 0) {
2861 dap->da_previous = dirrem;
2863 if (prevdirrem != NULL) {
2864 dap->da_previous = prevdirrem;
2866 dap->da_state &= ~DIRCHG;
2867 dap->da_pagedep = pagedep;
2869 dirrem->dm_dirinum = pagedep->pd_ino;
2870 add_to_worklist(&dirrem->dm_list);
2873 * Link into its inodedep. Put it on the id_bufwait list if the inode
2874 * is not yet written. If it is written, do the post-inode write
2875 * processing to put it on the id_pendinghd list.
2877 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2878 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2879 dap->da_state |= COMPLETE;
2880 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2881 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2883 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2885 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2891 * Called whenever the link count on an inode is changed.
2892 * It creates an inode dependency so that the new reference(s)
2893 * to the inode cannot be committed to disk until the updated
2894 * inode has been written.
2897 softdep_change_linkcnt(ip)
2898 struct inode *ip; /* the inode with the increased link count */
2900 struct inodedep *inodedep;
2903 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2904 if (ip->i_nlink < ip->i_effnlink) {
2906 panic("softdep_change_linkcnt: bad delta");
2908 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2913 * This workitem decrements the inode's link count.
2914 * If the link count reaches zero, the file is removed.
2917 handle_workitem_remove(dirrem)
2918 struct dirrem *dirrem;
2920 struct inodedep *inodedep;
2926 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) {
2927 softdep_error("handle_workitem_remove: vget", error);
2932 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
2934 panic("handle_workitem_remove: lost inodedep");
2937 * Normal file deletion.
2939 if ((dirrem->dm_state & RMDIR) == 0) {
2941 ip->i_flag |= IN_CHANGE;
2942 if (ip->i_nlink < ip->i_effnlink) {
2944 panic("handle_workitem_remove: bad file delta");
2946 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2950 WORKITEM_FREE(dirrem, D_DIRREM);
2954 * Directory deletion. Decrement reference count for both the
2955 * just deleted parent directory entry and the reference for ".".
2956 * Next truncate the directory to length zero. When the
2957 * truncation completes, arrange to have the reference count on
2958 * the parent decremented to account for the loss of "..".
2961 ip->i_flag |= IN_CHANGE;
2962 if (ip->i_nlink < ip->i_effnlink) {
2964 panic("handle_workitem_remove: bad dir delta");
2966 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2968 if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0)
2969 softdep_error("handle_workitem_remove: truncate", error);
2971 * Rename a directory to a new parent. Since, we are both deleting
2972 * and creating a new directory entry, the link count on the new
2973 * directory should not change. Thus we skip the followup dirrem.
2975 if (dirrem->dm_state & DIRCHG) {
2978 WORKITEM_FREE(dirrem, D_DIRREM);
2982 * If the inodedep does not exist, then the zero'ed inode has
2983 * been written to disk. If the allocated inode has never been
2984 * written to disk, then the on-disk inode is zero'ed. In either
2985 * case we can remove the file immediately.
2988 dirrem->dm_state = 0;
2989 oldinum = dirrem->dm_oldinum;
2990 dirrem->dm_oldinum = dirrem->dm_dirinum;
2991 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
2992 check_inode_unwritten(inodedep)) {
2995 handle_workitem_remove(dirrem);
2998 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3000 ip->i_flag |= IN_CHANGE;
3006 * Inode de-allocation dependencies.
3008 * When an inode's link count is reduced to zero, it can be de-allocated. We
3009 * found it convenient to postpone de-allocation until after the inode is
3010 * written to disk with its new link count (zero). At this point, all of the
3011 * on-disk inode's block pointers are nullified and, with careful dependency
3012 * list ordering, all dependencies related to the inode will be satisfied and
3013 * the corresponding dependency structures de-allocated. So, if/when the
3014 * inode is reused, there will be no mixing of old dependencies with new
3015 * ones. This artificial dependency is set up by the block de-allocation
3016 * procedure above (softdep_setup_freeblocks) and completed by the
3017 * following procedure.
3020 handle_workitem_freefile(freefile)
3021 struct freefile *freefile;
3025 struct inodedep *idp;
3030 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp);
3033 panic("handle_workitem_freefile: inodedep survived");
3035 tip.i_devvp = freefile->fx_devvp;
3036 tip.i_dev = freefile->fx_devvp->v_rdev;
3037 tip.i_fs = freefile->fx_fs;
3039 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
3040 softdep_error("handle_workitem_freefile", error);
3041 WORKITEM_FREE(freefile, D_FREEFILE);
3045 * Helper function which unlinks marker element from work list and returns
3046 * the next element on the list.
3048 static __inline struct worklist *
3049 markernext(struct worklist *marker)
3051 struct worklist *next;
3053 next = LIST_NEXT(marker, wk_list);
3054 LIST_REMOVE(marker, wk_list);
3061 * The dependency structures constructed above are most actively used when file
3062 * system blocks are written to disk. No constraints are placed on when a
3063 * block can be written, but unsatisfied update dependencies are made safe by
3064 * modifying (or replacing) the source memory for the duration of the disk
3065 * write. When the disk write completes, the memory block is again brought
3068 * In-core inode structure reclamation.
3070 * Because there are a finite number of "in-core" inode structures, they are
3071 * reused regularly. By transferring all inode-related dependencies to the
3072 * in-memory inode block and indexing them separately (via "inodedep"s), we
3073 * can allow "in-core" inode structures to be reused at any time and avoid
3074 * any increase in contention.
3076 * Called just before entering the device driver to initiate a new disk I/O.
3077 * The buffer must be locked, thus, no I/O completion operations can occur
3078 * while we are manipulating its associated dependencies.
3081 softdep_disk_io_initiation(bp)
3082 struct buf *bp; /* structure describing disk write to occur */
3084 struct worklist *wk;
3085 struct worklist marker;
3086 struct indirdep *indirdep;
3089 * We only care about write operations. There should never
3090 * be dependencies for reads.
3092 if (bp->b_cmd == BUF_CMD_READ)
3093 panic("softdep_disk_io_initiation: read");
3095 marker.wk_type = D_LAST + 1; /* Not a normal workitem */
3098 * Do any necessary pre-I/O processing.
3100 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) {
3101 LIST_INSERT_AFTER(wk, &marker, wk_list);
3103 switch (wk->wk_type) {
3106 initiate_write_filepage(WK_PAGEDEP(wk), bp);
3110 initiate_write_inodeblock(WK_INODEDEP(wk), bp);
3114 indirdep = WK_INDIRDEP(wk);
3115 if (indirdep->ir_state & GOINGAWAY)
3116 panic("disk_io_initiation: indirdep gone");
3118 * If there are no remaining dependencies, this
3119 * will be writing the real pointers, so the
3120 * dependency can be freed.
3122 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
3123 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
3124 brelse(indirdep->ir_savebp);
3125 /* inline expand WORKLIST_REMOVE(wk); */
3126 wk->wk_state &= ~ONWORKLIST;
3127 LIST_REMOVE(wk, wk_list);
3128 WORKITEM_FREE(indirdep, D_INDIRDEP);
3132 * Replace up-to-date version with safe version.
3134 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3135 M_INDIRDEP, M_SOFTDEP_FLAGS);
3137 indirdep->ir_state &= ~ATTACHED;
3138 indirdep->ir_state |= UNDONE;
3139 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3140 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3152 panic("handle_disk_io_initiation: Unexpected type %s",
3153 TYPENAME(wk->wk_type));
3160 * Called from within the procedure above to deal with unsatisfied
3161 * allocation dependencies in a directory. The buffer must be locked,
3162 * thus, no I/O completion operations can occur while we are
3163 * manipulating its associated dependencies.
3166 initiate_write_filepage(pagedep, bp)
3167 struct pagedep *pagedep;
3174 if (pagedep->pd_state & IOSTARTED) {
3176 * This can only happen if there is a driver that does not
3177 * understand chaining. Here biodone will reissue the call
3178 * to strategy for the incomplete buffers.
3180 printf("initiate_write_filepage: already started\n");
3183 pagedep->pd_state |= IOSTARTED;
3185 for (i = 0; i < DAHASHSZ; i++) {
3186 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3187 ep = (struct direct *)
3188 ((char *)bp->b_data + dap->da_offset);
3189 if (ep->d_ino != dap->da_newinum) {
3191 panic("%s: dir inum %d != new %"PRId64,
3192 "initiate_write_filepage",
3193 ep->d_ino, dap->da_newinum);
3195 if (dap->da_state & DIRCHG)
3196 ep->d_ino = dap->da_previous->dm_oldinum;
3199 dap->da_state &= ~ATTACHED;
3200 dap->da_state |= UNDONE;
3207 * Called from within the procedure above to deal with unsatisfied
3208 * allocation dependencies in an inodeblock. The buffer must be
3209 * locked, thus, no I/O completion operations can occur while we
3210 * are manipulating its associated dependencies.
3213 initiate_write_inodeblock(inodedep, bp)
3214 struct inodedep *inodedep;
3215 struct buf *bp; /* The inode block */
3217 struct allocdirect *adp, *lastadp;
3218 struct ufs1_dinode *dp;
3219 struct ufs1_dinode *sip;
3221 ufs_lbn_t prevlbn = 0;
3224 if (inodedep->id_state & IOSTARTED)
3225 panic("initiate_write_inodeblock: already started");
3226 inodedep->id_state |= IOSTARTED;
3227 fs = inodedep->id_fs;
3228 dp = (struct ufs1_dinode *)bp->b_data +
3229 ino_to_fsbo(fs, inodedep->id_ino);
3231 * If the bitmap is not yet written, then the allocated
3232 * inode cannot be written to disk.
3234 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3235 if (inodedep->id_savedino != NULL)
3236 panic("initiate_write_inodeblock: already doing I/O");
3237 MALLOC(sip, struct ufs1_dinode *,
3238 sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3239 inodedep->id_savedino = sip;
3240 *inodedep->id_savedino = *dp;
3241 bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
3242 dp->di_gen = inodedep->id_savedino->di_gen;
3246 * If no dependencies, then there is nothing to roll back.
3248 inodedep->id_savedsize = dp->di_size;
3249 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3252 * Set the dependencies to busy.
3255 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3256 adp = TAILQ_NEXT(adp, ad_next)) {
3258 if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3260 panic("softdep_write_inodeblock: lbn order");
3262 prevlbn = adp->ad_lbn;
3263 if (adp->ad_lbn < NDADDR &&
3264 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3266 panic("%s: direct pointer #%ld mismatch %d != %d",
3267 "softdep_write_inodeblock", adp->ad_lbn,
3268 dp->di_db[adp->ad_lbn], adp->ad_newblkno);
3270 if (adp->ad_lbn >= NDADDR &&
3271 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3273 panic("%s: indirect pointer #%ld mismatch %d != %d",
3274 "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
3275 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
3277 deplist |= 1 << adp->ad_lbn;
3278 if ((adp->ad_state & ATTACHED) == 0) {
3280 panic("softdep_write_inodeblock: Unknown state 0x%x",
3283 #endif /* DIAGNOSTIC */
3284 adp->ad_state &= ~ATTACHED;
3285 adp->ad_state |= UNDONE;
3288 * The on-disk inode cannot claim to be any larger than the last
3289 * fragment that has been written. Otherwise, the on-disk inode
3290 * might have fragments that were not the last block in the file
3291 * which would corrupt the filesystem.
3293 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3294 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3295 if (adp->ad_lbn >= NDADDR)
3297 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3298 /* keep going until hitting a rollback to a frag */
3299 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3301 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3302 for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3304 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3306 panic("softdep_write_inodeblock: lost dep1");
3308 #endif /* DIAGNOSTIC */
3311 for (i = 0; i < NIADDR; i++) {
3313 if (dp->di_ib[i] != 0 &&
3314 (deplist & ((1 << NDADDR) << i)) == 0) {
3316 panic("softdep_write_inodeblock: lost dep2");
3318 #endif /* DIAGNOSTIC */
3325 * If we have zero'ed out the last allocated block of the file,
3326 * roll back the size to the last currently allocated block.
3327 * We know that this last allocated block is a full-sized as
3328 * we already checked for fragments in the loop above.
3330 if (lastadp != NULL &&
3331 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3332 for (i = lastadp->ad_lbn; i >= 0; i--)
3333 if (dp->di_db[i] != 0)
3335 dp->di_size = (i + 1) * fs->fs_bsize;
3338 * The only dependencies are for indirect blocks.
3340 * The file size for indirect block additions is not guaranteed.
3341 * Such a guarantee would be non-trivial to achieve. The conventional
3342 * synchronous write implementation also does not make this guarantee.
3343 * Fsck should catch and fix discrepancies. Arguably, the file size
3344 * can be over-estimated without destroying integrity when the file
3345 * moves into the indirect blocks (i.e., is large). If we want to
3346 * postpone fsck, we are stuck with this argument.
3348 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3349 dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3354 * This routine is called during the completion interrupt
3355 * service routine for a disk write (from the procedure called
3356 * by the device driver to inform the filesystem caches of
3357 * a request completion). It should be called early in this
3358 * procedure, before the block is made available to other
3359 * processes or other routines are called.
3362 softdep_disk_write_complete(bp)
3363 struct buf *bp; /* describes the completed disk write */
3365 struct worklist *wk;
3366 struct workhead reattach;
3367 struct newblk *newblk;
3368 struct allocindir *aip;
3369 struct allocdirect *adp;
3370 struct indirdep *indirdep;
3371 struct inodedep *inodedep;
3372 struct bmsafemap *bmsafemap;
3375 if (lk.lkt_held != NOHOLDER)
3376 panic("softdep_disk_write_complete: lock is held");
3377 lk.lkt_held = SPECIAL_FLAG;
3379 LIST_INIT(&reattach);
3380 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3381 WORKLIST_REMOVE(wk);
3382 switch (wk->wk_type) {
3385 if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3386 WORKLIST_INSERT(&reattach, wk);
3390 if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3391 WORKLIST_INSERT(&reattach, wk);
3395 bmsafemap = WK_BMSAFEMAP(wk);
3396 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3397 newblk->nb_state |= DEPCOMPLETE;
3398 newblk->nb_bmsafemap = NULL;
3399 LIST_REMOVE(newblk, nb_deps);
3402 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3403 adp->ad_state |= DEPCOMPLETE;
3405 LIST_REMOVE(adp, ad_deps);
3406 handle_allocdirect_partdone(adp);
3409 LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3410 aip->ai_state |= DEPCOMPLETE;
3412 LIST_REMOVE(aip, ai_deps);
3413 handle_allocindir_partdone(aip);
3416 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3417 inodedep->id_state |= DEPCOMPLETE;
3418 LIST_REMOVE(inodedep, id_deps);
3419 inodedep->id_buf = NULL;
3421 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3425 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3429 adp = WK_ALLOCDIRECT(wk);
3430 adp->ad_state |= COMPLETE;
3431 handle_allocdirect_partdone(adp);
3435 aip = WK_ALLOCINDIR(wk);
3436 aip->ai_state |= COMPLETE;
3437 handle_allocindir_partdone(aip);
3441 indirdep = WK_INDIRDEP(wk);
3442 if (indirdep->ir_state & GOINGAWAY) {
3443 lk.lkt_held = NOHOLDER;
3444 panic("disk_write_complete: indirdep gone");
3446 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3447 FREE(indirdep->ir_saveddata, M_INDIRDEP);
3448 indirdep->ir_saveddata = 0;
3449 indirdep->ir_state &= ~UNDONE;
3450 indirdep->ir_state |= ATTACHED;
3451 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3452 handle_allocindir_partdone(aip);
3453 if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
3454 lk.lkt_held = NOHOLDER;
3455 panic("disk_write_complete: not gone");
3458 WORKLIST_INSERT(&reattach, wk);
3459 if ((bp->b_flags & B_DELWRI) == 0)
3460 stat_indir_blk_ptrs++;
3465 lk.lkt_held = NOHOLDER;
3466 panic("handle_disk_write_complete: Unknown type %s",
3467 TYPENAME(wk->wk_type));
3472 * Reattach any requests that must be redone.
3474 while ((wk = LIST_FIRST(&reattach)) != NULL) {
3475 WORKLIST_REMOVE(wk);
3476 WORKLIST_INSERT(&bp->b_dep, wk);
3479 if (lk.lkt_held != SPECIAL_FLAG)
3480 panic("softdep_disk_write_complete: lock lost");
3481 lk.lkt_held = NOHOLDER;
3486 * Called from within softdep_disk_write_complete above. Note that
3487 * this routine is always called from interrupt level with further
3488 * splbio interrupts blocked.
3491 handle_allocdirect_partdone(adp)
3492 struct allocdirect *adp; /* the completed allocdirect */
3494 struct allocdirect *listadp;
3495 struct inodedep *inodedep;
3498 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3500 if (adp->ad_buf != NULL) {
3501 lk.lkt_held = NOHOLDER;
3502 panic("handle_allocdirect_partdone: dangling dep");
3505 * The on-disk inode cannot claim to be any larger than the last
3506 * fragment that has been written. Otherwise, the on-disk inode
3507 * might have fragments that were not the last block in the file
3508 * which would corrupt the filesystem. Thus, we cannot free any
3509 * allocdirects after one whose ad_oldblkno claims a fragment as
3510 * these blocks must be rolled back to zero before writing the inode.
3511 * We check the currently active set of allocdirects in id_inoupdt.
3513 inodedep = adp->ad_inodedep;
3514 bsize = inodedep->id_fs->fs_bsize;
3515 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) {
3516 /* found our block */
3519 /* continue if ad_oldlbn is not a fragment */
3520 if (listadp->ad_oldsize == 0 ||
3521 listadp->ad_oldsize == bsize)
3523 /* hit a fragment */
3527 * If we have reached the end of the current list without
3528 * finding the just finished dependency, then it must be
3529 * on the future dependency list. Future dependencies cannot
3530 * be freed until they are moved to the current list.
3532 if (listadp == NULL) {
3534 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next)
3535 /* found our block */
3538 if (listadp == NULL) {
3539 lk.lkt_held = NOHOLDER;
3540 panic("handle_allocdirect_partdone: lost dep");
3546 * If we have found the just finished dependency, then free
3547 * it along with anything that follows it that is complete.
3549 for (; adp; adp = listadp) {
3550 listadp = TAILQ_NEXT(adp, ad_next);
3551 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3553 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
3558 * Called from within softdep_disk_write_complete above. Note that
3559 * this routine is always called from interrupt level with further
3560 * splbio interrupts blocked.
3563 handle_allocindir_partdone(aip)
3564 struct allocindir *aip; /* the completed allocindir */
3566 struct indirdep *indirdep;
3568 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
3570 if (aip->ai_buf != NULL) {
3571 lk.lkt_held = NOHOLDER;
3572 panic("handle_allocindir_partdone: dangling dependency");
3574 indirdep = aip->ai_indirdep;
3575 if (indirdep->ir_state & UNDONE) {
3576 LIST_REMOVE(aip, ai_next);
3577 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
3580 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
3582 LIST_REMOVE(aip, ai_next);
3583 if (aip->ai_freefrag != NULL)
3584 add_to_worklist(&aip->ai_freefrag->ff_list);
3585 WORKITEM_FREE(aip, D_ALLOCINDIR);
3589 * Called from within softdep_disk_write_complete above to restore
3590 * in-memory inode block contents to their most up-to-date state. Note
3591 * that this routine is always called from interrupt level with further
3592 * splbio interrupts blocked.
3595 handle_written_inodeblock(inodedep, bp)
3596 struct inodedep *inodedep;
3597 struct buf *bp; /* buffer containing the inode block */
3599 struct worklist *wk, *filefree;
3600 struct allocdirect *adp, *nextadp;
3601 struct ufs1_dinode *dp;
3604 if ((inodedep->id_state & IOSTARTED) == 0) {
3605 lk.lkt_held = NOHOLDER;
3606 panic("handle_written_inodeblock: not started");
3608 inodedep->id_state &= ~IOSTARTED;
3609 dp = (struct ufs1_dinode *)bp->b_data +
3610 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
3612 * If we had to rollback the inode allocation because of
3613 * bitmaps being incomplete, then simply restore it.
3614 * Keep the block dirty so that it will not be reclaimed until
3615 * all associated dependencies have been cleared and the
3616 * corresponding updates written to disk.
3618 if (inodedep->id_savedino != NULL) {
3619 *dp = *inodedep->id_savedino;
3620 FREE(inodedep->id_savedino, M_INODEDEP);
3621 inodedep->id_savedino = NULL;
3622 if ((bp->b_flags & B_DELWRI) == 0)
3623 stat_inode_bitmap++;
3627 inodedep->id_state |= COMPLETE;
3629 * Roll forward anything that had to be rolled back before
3630 * the inode could be updated.
3633 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
3634 nextadp = TAILQ_NEXT(adp, ad_next);
3635 if (adp->ad_state & ATTACHED) {
3636 lk.lkt_held = NOHOLDER;
3637 panic("handle_written_inodeblock: new entry");
3639 if (adp->ad_lbn < NDADDR) {
3640 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) {
3641 lk.lkt_held = NOHOLDER;
3642 panic("%s: %s #%ld mismatch %d != %d",
3643 "handle_written_inodeblock",
3644 "direct pointer", adp->ad_lbn,
3645 dp->di_db[adp->ad_lbn], adp->ad_oldblkno);
3647 dp->di_db[adp->ad_lbn] = adp->ad_newblkno;
3649 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) {
3650 lk.lkt_held = NOHOLDER;
3651 panic("%s: %s #%ld allocated as %d",
3652 "handle_written_inodeblock",
3653 "indirect pointer", adp->ad_lbn - NDADDR,
3654 dp->di_ib[adp->ad_lbn - NDADDR]);
3656 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno;
3658 adp->ad_state &= ~UNDONE;
3659 adp->ad_state |= ATTACHED;
3662 if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
3663 stat_direct_blk_ptrs++;
3665 * Reset the file size to its most up-to-date value.
3667 if (inodedep->id_savedsize == -1) {
3668 lk.lkt_held = NOHOLDER;
3669 panic("handle_written_inodeblock: bad size");
3671 if (dp->di_size != inodedep->id_savedsize) {
3672 dp->di_size = inodedep->id_savedsize;
3675 inodedep->id_savedsize = -1;
3677 * If there were any rollbacks in the inode block, then it must be
3678 * marked dirty so that its will eventually get written back in
3684 * Process any allocdirects that completed during the update.
3686 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
3687 handle_allocdirect_partdone(adp);
3689 * Process deallocations that were held pending until the
3690 * inode had been written to disk. Freeing of the inode
3691 * is delayed until after all blocks have been freed to
3692 * avoid creation of new <vfsid, inum, lbn> triples
3693 * before the old ones have been deleted.
3696 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
3697 WORKLIST_REMOVE(wk);
3698 switch (wk->wk_type) {
3702 * We defer adding filefree to the worklist until
3703 * all other additions have been made to ensure
3704 * that it will be done after all the old blocks
3707 if (filefree != NULL) {
3708 lk.lkt_held = NOHOLDER;
3709 panic("handle_written_inodeblock: filefree");
3715 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
3719 diradd_inode_written(WK_DIRADD(wk), inodedep);
3723 wk->wk_state |= COMPLETE;
3724 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE)
3726 /* -- fall through -- */
3729 add_to_worklist(wk);
3733 lk.lkt_held = NOHOLDER;
3734 panic("handle_written_inodeblock: Unknown type %s",
3735 TYPENAME(wk->wk_type));
3739 if (filefree != NULL) {
3740 if (free_inodedep(inodedep) == 0) {
3741 lk.lkt_held = NOHOLDER;
3742 panic("handle_written_inodeblock: live inodedep");
3744 add_to_worklist(filefree);
3749 * If no outstanding dependencies, free it.
3751 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0)
3753 return (hadchanges);
3757 * Process a diradd entry after its dependent inode has been written.
3758 * This routine must be called with splbio interrupts blocked.
3761 diradd_inode_written(dap, inodedep)
3763 struct inodedep *inodedep;
3765 struct pagedep *pagedep;
3767 dap->da_state |= COMPLETE;
3768 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3769 if (dap->da_state & DIRCHG)
3770 pagedep = dap->da_previous->dm_pagedep;
3772 pagedep = dap->da_pagedep;
3773 LIST_REMOVE(dap, da_pdlist);
3774 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3776 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3780 * Handle the completion of a mkdir dependency.
3783 handle_written_mkdir(mkdir, type)
3784 struct mkdir *mkdir;
3788 struct pagedep *pagedep;
3790 if (mkdir->md_state != type) {
3791 lk.lkt_held = NOHOLDER;
3792 panic("handle_written_mkdir: bad type");
3794 dap = mkdir->md_diradd;
3795 dap->da_state &= ~type;
3796 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
3797 dap->da_state |= DEPCOMPLETE;
3798 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3799 if (dap->da_state & DIRCHG)
3800 pagedep = dap->da_previous->dm_pagedep;
3802 pagedep = dap->da_pagedep;
3803 LIST_REMOVE(dap, da_pdlist);
3804 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3806 LIST_REMOVE(mkdir, md_mkdirs);
3807 WORKITEM_FREE(mkdir, D_MKDIR);
3811 * Called from within softdep_disk_write_complete above.
3812 * A write operation was just completed. Removed inodes can
3813 * now be freed and associated block pointers may be committed.
3814 * Note that this routine is always called from interrupt level
3815 * with further splbio interrupts blocked.
3818 handle_written_filepage(pagedep, bp)
3819 struct pagedep *pagedep;
3820 struct buf *bp; /* buffer containing the written page */
3822 struct dirrem *dirrem;
3823 struct diradd *dap, *nextdap;
3827 if ((pagedep->pd_state & IOSTARTED) == 0) {
3828 lk.lkt_held = NOHOLDER;
3829 panic("handle_written_filepage: not started");
3831 pagedep->pd_state &= ~IOSTARTED;
3833 * Process any directory removals that have been committed.
3835 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
3836 LIST_REMOVE(dirrem, dm_next);
3837 dirrem->dm_dirinum = pagedep->pd_ino;
3838 add_to_worklist(&dirrem->dm_list);
3841 * Free any directory additions that have been committed.
3843 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
3846 * Uncommitted directory entries must be restored.
3848 for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
3849 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
3851 nextdap = LIST_NEXT(dap, da_pdlist);
3852 if (dap->da_state & ATTACHED) {
3853 lk.lkt_held = NOHOLDER;
3854 panic("handle_written_filepage: attached");
3856 ep = (struct direct *)
3857 ((char *)bp->b_data + dap->da_offset);
3858 ep->d_ino = dap->da_newinum;
3859 dap->da_state &= ~UNDONE;
3860 dap->da_state |= ATTACHED;
3863 * If the inode referenced by the directory has
3864 * been written out, then the dependency can be
3865 * moved to the pending list.
3867 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3868 LIST_REMOVE(dap, da_pdlist);
3869 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
3875 * If there were any rollbacks in the directory, then it must be
3876 * marked dirty so that its will eventually get written back in
3880 if ((bp->b_flags & B_DELWRI) == 0)
3885 * If no dependencies remain, the pagedep will be freed.
3886 * Otherwise it will remain to update the page before it
3887 * is written back to disk.
3889 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) {
3890 for (i = 0; i < DAHASHSZ; i++)
3891 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
3893 if (i == DAHASHSZ) {
3894 LIST_REMOVE(pagedep, pd_hash);
3895 WORKITEM_FREE(pagedep, D_PAGEDEP);
3903 * Writing back in-core inode structures.
3905 * The filesystem only accesses an inode's contents when it occupies an
3906 * "in-core" inode structure. These "in-core" structures are separate from
3907 * the page frames used to cache inode blocks. Only the latter are
3908 * transferred to/from the disk. So, when the updated contents of the
3909 * "in-core" inode structure are copied to the corresponding in-memory inode
3910 * block, the dependencies are also transferred. The following procedure is
3911 * called when copying a dirty "in-core" inode to a cached inode block.
3915 * Called when an inode is loaded from disk. If the effective link count
3916 * differed from the actual link count when it was last flushed, then we
3917 * need to ensure that the correct effective link count is put back.
3920 softdep_load_inodeblock(ip)
3921 struct inode *ip; /* the "in_core" copy of the inode */
3923 struct inodedep *inodedep;
3926 * Check for alternate nlink count.
3928 ip->i_effnlink = ip->i_nlink;
3930 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3934 ip->i_effnlink -= inodedep->id_nlinkdelta;
3939 * This routine is called just before the "in-core" inode
3940 * information is to be copied to the in-memory inode block.
3941 * Recall that an inode block contains several inodes. If
3942 * the force flag is set, then the dependencies will be
3943 * cleared so that the update can always be made. Note that
3944 * the buffer is locked when this routine is called, so we
3945 * will never be in the middle of writing the inode block
3949 softdep_update_inodeblock(ip, bp, waitfor)
3950 struct inode *ip; /* the "in_core" copy of the inode */
3951 struct buf *bp; /* the buffer containing the inode block */
3952 int waitfor; /* nonzero => update must be allowed */
3954 struct inodedep *inodedep;
3955 struct worklist *wk;
3959 * If the effective link count is not equal to the actual link
3960 * count, then we must track the difference in an inodedep while
3961 * the inode is (potentially) tossed out of the cache. Otherwise,
3962 * if there is no existing inodedep, then there are no dependencies
3966 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3968 if (ip->i_effnlink != ip->i_nlink)
3969 panic("softdep_update_inodeblock: bad link count");
3972 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
3974 panic("softdep_update_inodeblock: bad delta");
3977 * Changes have been initiated. Anything depending on these
3978 * changes cannot occur until this inode has been written.
3980 inodedep->id_state &= ~COMPLETE;
3981 if ((inodedep->id_state & ONWORKLIST) == 0)
3982 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
3984 * Any new dependencies associated with the incore inode must
3985 * now be moved to the list associated with the buffer holding
3986 * the in-memory copy of the inode. Once merged process any
3987 * allocdirects that are completed by the merger.
3989 merge_inode_lists(inodedep);
3990 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
3991 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
3993 * Now that the inode has been pushed into the buffer, the
3994 * operations dependent on the inode being written to disk
3995 * can be moved to the id_bufwait so that they will be
3996 * processed when the buffer I/O completes.
3998 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
3999 WORKLIST_REMOVE(wk);
4000 WORKLIST_INSERT(&inodedep->id_bufwait, wk);
4003 * Newly allocated inodes cannot be written until the bitmap
4004 * that allocates them have been written (indicated by
4005 * DEPCOMPLETE being set in id_state). If we are doing a
4006 * forced sync (e.g., an fsync on a file), we force the bitmap
4007 * to be written so that the update can be done.
4009 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
4013 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4016 (error = bwrite(inodedep->id_buf)) != 0)
4017 softdep_error("softdep_update_inodeblock: bwrite", error);
4021 * Merge the new inode dependency list (id_newinoupdt) into the old
4022 * inode dependency list (id_inoupdt). This routine must be called
4023 * with splbio interrupts blocked.
4026 merge_inode_lists(inodedep)
4027 struct inodedep *inodedep;
4029 struct allocdirect *listadp, *newadp;
4031 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
4032 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) {
4033 if (listadp->ad_lbn < newadp->ad_lbn) {
4034 listadp = TAILQ_NEXT(listadp, ad_next);
4037 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
4038 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
4039 if (listadp->ad_lbn == newadp->ad_lbn) {
4040 allocdirect_merge(&inodedep->id_inoupdt, newadp,
4044 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
4046 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) {
4047 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
4048 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next);
4053 * If we are doing an fsync, then we must ensure that any directory
4054 * entries for the inode have been written after the inode gets to disk.
4058 struct vnode *vp; /* the "in_core" copy of the inode */
4060 struct inodedep *inodedep;
4061 struct pagedep *pagedep;
4062 struct worklist *wk;
4069 int error, flushparent;
4076 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
4080 if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
4081 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
4082 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
4083 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) {
4085 panic("softdep_fsync: pending ops");
4087 for (error = 0, flushparent = 0; ; ) {
4088 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
4090 if (wk->wk_type != D_DIRADD) {
4092 panic("softdep_fsync: Unexpected type %s",
4093 TYPENAME(wk->wk_type));
4095 dap = WK_DIRADD(wk);
4097 * Flush our parent if this directory entry
4098 * has a MKDIR_PARENT dependency.
4100 if (dap->da_state & DIRCHG)
4101 pagedep = dap->da_previous->dm_pagedep;
4103 pagedep = dap->da_pagedep;
4104 mnt = pagedep->pd_mnt;
4105 parentino = pagedep->pd_ino;
4106 lbn = pagedep->pd_lbn;
4107 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) {
4109 panic("softdep_fsync: dirty");
4111 flushparent = dap->da_state & MKDIR_PARENT;
4113 * If we are being fsync'ed as part of vgone'ing this vnode,
4114 * then we will not be able to release and recover the
4115 * vnode below, so we just have to give up on writing its
4116 * directory entry out. It will eventually be written, just
4117 * not now, but then the user was not asking to have it
4118 * written, so we are not breaking any promises.
4120 if (vp->v_flag & VRECLAIMED)
4123 * We prevent deadlock by always fetching inodes from the
4124 * root, moving down the directory tree. Thus, when fetching
4125 * our parent directory, we must unlock ourselves before
4126 * requesting the lock on our parent. See the comment in
4127 * ufs_lookup for details on possible races.
4131 error = VFS_VGET(mnt, parentino, &pvp);
4132 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4136 if ((error = ffs_update(pvp, 1)) != 0) {
4142 * Flush directory page containing the inode's name.
4144 error = bread(pvp, lblktodoff(fs, lbn), blksize(fs, VTOI(pvp), lbn), &bp);
4151 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
4159 * Flush all the dirty bitmaps associated with the block device
4160 * before flushing the rest of the dirty blocks so as to reduce
4161 * the number of dependencies that will have to be rolled back.
4163 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data);
4166 softdep_fsync_mountdev(vp)
4169 if (!vn_isdisk(vp, NULL))
4170 panic("softdep_fsync_mountdev: vnode not a disk");
4172 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
4173 softdep_fsync_mountdev_bp, vp);
4174 drain_output(vp, 1);
4179 softdep_fsync_mountdev_bp(struct buf *bp, void *data)
4181 struct worklist *wk;
4182 struct vnode *vp = data;
4185 * If it is already scheduled, skip to the next buffer.
4187 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
4189 if (bp->b_vp != vp || (bp->b_flags & B_DELWRI) == 0) {
4191 printf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp);
4195 * We are only interested in bitmaps with outstanding
4198 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
4199 wk->wk_type != D_BMSAFEMAP) {
4211 * This routine is called when we are trying to synchronously flush a
4212 * file. This routine must eliminate any filesystem metadata dependencies
4213 * so that the syncing routine can succeed by pushing the dirty blocks
4214 * associated with the file. If any I/O errors occur, they are returned.
4216 struct softdep_sync_metadata_info {
4221 static int softdep_sync_metadata_bp(struct buf *bp, void *data);
4224 softdep_sync_metadata(struct vnode *vp, struct thread *td)
4226 struct softdep_sync_metadata_info info;
4230 * Check whether this vnode is involved in a filesystem
4231 * that is doing soft dependency processing.
4233 if (!vn_isdisk(vp, NULL)) {
4234 if (!DOINGSOFTDEP(vp))
4237 if (vp->v_rdev->si_mountpoint == NULL ||
4238 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0)
4241 * Ensure that any direct block dependencies have been cleared.
4244 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
4249 * For most files, the only metadata dependencies are the
4250 * cylinder group maps that allocate their inode or blocks.
4251 * The block allocation dependencies can be found by traversing
4252 * the dependency lists for any buffers that remain on their
4253 * dirty buffer list. The inode allocation dependency will
4254 * be resolved when the inode is updated with MNT_WAIT.
4255 * This work is done in two passes. The first pass grabs most
4256 * of the buffers and begins asynchronously writing them. The
4257 * only way to wait for these asynchronous writes is to sleep
4258 * on the filesystem vnode which may stay busy for a long time
4259 * if the filesystem is active. So, instead, we make a second
4260 * pass over the dependencies blocking on each write. In the
4261 * usual case we will be blocking against a write that we
4262 * initiated, so when it is done the dependency will have been
4263 * resolved. Thus the second pass is expected to end quickly.
4265 waitfor = MNT_NOWAIT;
4268 * We must wait for any I/O in progress to finish so that
4269 * all potential buffers on the dirty list will be visible.
4271 drain_output(vp, 1);
4273 info.waitfor = waitfor;
4274 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
4275 softdep_sync_metadata_bp, &info);
4278 return(-error); /* error code */
4282 * The brief unlock is to allow any pent up dependency
4283 * processing to be done. Then proceed with the second pass.
4285 if (waitfor == MNT_NOWAIT) {
4293 * If we have managed to get rid of all the dirty buffers,
4294 * then we are done. For certain directories and block
4295 * devices, we may need to do further work.
4297 * We must wait for any I/O in progress to finish so that
4298 * all potential buffers on the dirty list will be visible.
4300 drain_output(vp, 1);
4301 if (RB_EMPTY(&vp->v_rbdirty_tree)) {
4308 * If we are trying to sync a block device, some of its buffers may
4309 * contain metadata that cannot be written until the contents of some
4310 * partially written files have been written to disk. The only easy
4311 * way to accomplish this is to sync the entire filesystem (luckily
4312 * this happens rarely).
4314 if (vn_isdisk(vp, NULL) &&
4316 vp->v_rdev->si_mountpoint && !vn_islocked(vp) &&
4317 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT)) != 0)
4323 softdep_sync_metadata_bp(struct buf *bp, void *data)
4325 struct softdep_sync_metadata_info *info = data;
4326 struct pagedep *pagedep;
4327 struct allocdirect *adp;
4328 struct allocindir *aip;
4329 struct worklist *wk;
4334 if (getdirtybuf(&bp, MNT_WAIT) == 0) {
4335 printf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp);
4338 if (bp->b_vp != info->vp || (bp->b_flags & B_DELWRI) == 0) {
4339 printf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp, info->vp);
4345 * As we hold the buffer locked, none of its dependencies
4348 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4349 switch (wk->wk_type) {
4352 adp = WK_ALLOCDIRECT(wk);
4353 if (adp->ad_state & DEPCOMPLETE)
4356 if (getdirtybuf(&nbp, info->waitfor) == 0)
4359 if (info->waitfor == MNT_NOWAIT) {
4361 } else if ((error = bwrite(nbp)) != 0) {
4370 aip = WK_ALLOCINDIR(wk);
4371 if (aip->ai_state & DEPCOMPLETE)
4374 if (getdirtybuf(&nbp, info->waitfor) == 0)
4377 if (info->waitfor == MNT_NOWAIT) {
4379 } else if ((error = bwrite(nbp)) != 0) {
4390 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
4391 if (aip->ai_state & DEPCOMPLETE)
4394 if (getdirtybuf(&nbp, MNT_WAIT) == 0)
4397 if ((error = bwrite(nbp)) != 0) {
4408 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
4409 WK_INODEDEP(wk)->id_ino)) != 0) {
4419 * We are trying to sync a directory that may
4420 * have dependencies on both its own metadata
4421 * and/or dependencies on the inodes of any
4422 * recently allocated files. We walk its diradd
4423 * lists pushing out the associated inode.
4425 pagedep = WK_PAGEDEP(wk);
4426 for (i = 0; i < DAHASHSZ; i++) {
4427 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
4430 flush_pagedep_deps(info->vp,
4432 &pagedep->pd_diraddhd[i]))) {
4443 * This case should never happen if the vnode has
4444 * been properly sync'ed. However, if this function
4445 * is used at a place where the vnode has not yet
4446 * been sync'ed, this dependency can show up. So,
4447 * rather than panic, just flush it.
4449 nbp = WK_MKDIR(wk)->md_buf;
4450 if (getdirtybuf(&nbp, info->waitfor) == 0)
4453 if (info->waitfor == MNT_NOWAIT) {
4455 } else if ((error = bwrite(nbp)) != 0) {
4465 * This case should never happen if the vnode has
4466 * been properly sync'ed. However, if this function
4467 * is used at a place where the vnode has not yet
4468 * been sync'ed, this dependency can show up. So,
4469 * rather than panic, just flush it.
4471 * nbp can wind up == bp if a device node for the
4472 * same filesystem is being fsynced at the same time,
4473 * leading to a panic if we don't catch the case.
4475 nbp = WK_BMSAFEMAP(wk)->sm_buf;
4478 if (getdirtybuf(&nbp, info->waitfor) == 0)
4481 if (info->waitfor == MNT_NOWAIT) {
4483 } else if ((error = bwrite(nbp)) != 0) {
4493 panic("softdep_sync_metadata: Unknown type %s",
4494 TYPENAME(wk->wk_type));
4505 * Flush the dependencies associated with an inodedep.
4506 * Called with splbio blocked.
4509 flush_inodedep_deps(fs, ino)
4513 struct inodedep *inodedep;
4514 struct allocdirect *adp;
4519 * This work is done in two passes. The first pass grabs most
4520 * of the buffers and begins asynchronously writing them. The
4521 * only way to wait for these asynchronous writes is to sleep
4522 * on the filesystem vnode which may stay busy for a long time
4523 * if the filesystem is active. So, instead, we make a second
4524 * pass over the dependencies blocking on each write. In the
4525 * usual case we will be blocking against a write that we
4526 * initiated, so when it is done the dependency will have been
4527 * resolved. Thus the second pass is expected to end quickly.
4528 * We give a brief window at the top of the loop to allow
4529 * any pending I/O to complete.
4531 for (waitfor = MNT_NOWAIT; ; ) {
4534 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4536 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, ad_next) {
4537 if (adp->ad_state & DEPCOMPLETE)
4540 if (getdirtybuf(&bp, waitfor) == 0) {
4541 if (waitfor == MNT_NOWAIT)
4546 if (waitfor == MNT_NOWAIT) {
4548 } else if ((error = bwrite(bp)) != 0) {
4557 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) {
4558 if (adp->ad_state & DEPCOMPLETE)
4561 if (getdirtybuf(&bp, waitfor) == 0) {
4562 if (waitfor == MNT_NOWAIT)
4567 if (waitfor == MNT_NOWAIT) {
4569 } else if ((error = bwrite(bp)) != 0) {
4579 * If pass2, we are done, otherwise do pass 2.
4581 if (waitfor == MNT_WAIT)
4586 * Try freeing inodedep in case all dependencies have been removed.
4588 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
4589 (void) free_inodedep(inodedep);
4594 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
4595 * Called with splbio blocked.
4598 flush_pagedep_deps(pvp, mp, diraddhdp)
4601 struct diraddhd *diraddhdp;
4603 struct inodedep *inodedep;
4604 struct ufsmount *ump;
4607 int gotit, error = 0;
4612 while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
4614 * Flush ourselves if this directory entry
4615 * has a MKDIR_PARENT dependency.
4617 if (dap->da_state & MKDIR_PARENT) {
4619 if ((error = ffs_update(pvp, 1)) != 0)
4623 * If that cleared dependencies, go on to next.
4625 if (dap != LIST_FIRST(diraddhdp))
4627 if (dap->da_state & MKDIR_PARENT) {
4629 panic("flush_pagedep_deps: MKDIR_PARENT");
4633 * A newly allocated directory must have its "." and
4634 * ".." entries written out before its name can be
4635 * committed in its parent. We do not want or need
4636 * the full semantics of a synchronous VOP_FSYNC as
4637 * that may end up here again, once for each directory
4638 * level in the filesystem. Instead, we push the blocks
4639 * and wait for them to clear. We have to fsync twice
4640 * because the first call may choose to defer blocks
4641 * that still have dependencies, but deferral will
4642 * happen at most once.
4644 inum = dap->da_newinum;
4645 if (dap->da_state & MKDIR_BODY) {
4647 if ((error = VFS_VGET(mp, inum, &vp)) != 0)
4649 if ((error=VOP_FSYNC(vp, MNT_NOWAIT)) ||
4650 (error=VOP_FSYNC(vp, MNT_NOWAIT))) {
4654 drain_output(vp, 0);
4658 * If that cleared dependencies, go on to next.
4660 if (dap != LIST_FIRST(diraddhdp))
4662 if (dap->da_state & MKDIR_BODY) {
4664 panic("flush_pagedep_deps: MKDIR_BODY");
4668 * Flush the inode on which the directory entry depends.
4669 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
4670 * the only remaining dependency is that the updated inode
4671 * count must get pushed to disk. The inode has already
4672 * been pushed into its inode buffer (via VOP_UPDATE) at
4673 * the time of the reference count change. So we need only
4674 * locate that buffer, ensure that there will be no rollback
4675 * caused by a bitmap dependency, then write the inode buffer.
4677 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) {
4679 panic("flush_pagedep_deps: lost inode");
4682 * If the inode still has bitmap dependencies,
4683 * push them to disk.
4685 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4686 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4688 if (gotit && (error = bwrite(inodedep->id_buf)) != 0)
4691 if (dap != LIST_FIRST(diraddhdp))
4695 * If the inode is still sitting in a buffer waiting
4696 * to be written, push it to disk.
4699 if ((error = bread(ump->um_devvp,
4700 fsbtodoff(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
4701 (int)ump->um_fs->fs_bsize, &bp)) != 0)
4703 if ((error = bwrite(bp)) != 0)
4707 * If we have failed to get rid of all the dependencies
4708 * then something is seriously wrong.
4710 if (dap == LIST_FIRST(diraddhdp)) {
4712 panic("flush_pagedep_deps: flush failed");
4721 * A large burst of file addition or deletion activity can drive the
4722 * memory load excessively high. First attempt to slow things down
4723 * using the techniques below. If that fails, this routine requests
4724 * the offending operations to fall back to running synchronously
4725 * until the memory load returns to a reasonable level.
4728 softdep_slowdown(vp)
4731 int max_softdeps_hard;
4733 max_softdeps_hard = max_softdeps * 11 / 10;
4734 if (num_dirrem < max_softdeps_hard / 2 &&
4735 num_inodedep < max_softdeps_hard)
4737 stat_sync_limit_hit += 1;
4742 * If memory utilization has gotten too high, deliberately slow things
4743 * down and speed up the I/O processing.
4746 request_cleanup(resource, islocked)
4750 struct thread *td = curthread; /* XXX */
4753 * We never hold up the filesystem syncer process.
4755 if (td == filesys_syncer)
4758 * First check to see if the work list has gotten backlogged.
4759 * If it has, co-opt this process to help clean up two entries.
4760 * Because this process may hold inodes locked, we cannot
4761 * handle any remove requests that might block on a locked
4762 * inode as that could lead to deadlock.
4764 if (num_on_worklist > max_softdeps / 10) {
4767 process_worklist_item(NULL, LK_NOWAIT);
4768 process_worklist_item(NULL, LK_NOWAIT);
4769 stat_worklist_push += 2;
4776 * If we are resource constrained on inode dependencies, try
4777 * flushing some dirty inodes. Otherwise, we are constrained
4778 * by file deletions, so try accelerating flushes of directories
4779 * with removal dependencies. We would like to do the cleanup
4780 * here, but we probably hold an inode locked at this point and
4781 * that might deadlock against one that we try to clean. So,
4782 * the best that we can do is request the syncer daemon to do
4783 * the cleanup for us.
4788 stat_ino_limit_push += 1;
4789 req_clear_inodedeps += 1;
4790 stat_countp = &stat_ino_limit_hit;
4794 stat_blk_limit_push += 1;
4795 req_clear_remove += 1;
4796 stat_countp = &stat_blk_limit_hit;
4802 panic("request_cleanup: unknown type");
4805 * Hopefully the syncer daemon will catch up and awaken us.
4806 * We wait at most tickdelay before proceeding in any case.
4811 if (!callout_active(&handle))
4812 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2,
4814 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0,
4823 * Awaken processes pausing in request_cleanup and clear proc_waiting
4824 * to indicate that there is no longer a timer running.
4831 wakeup_one(&proc_waiting);
4832 if (proc_waiting > 0)
4833 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2,
4836 callout_deactivate(&handle);
4840 * Flush out a directory with at least one removal dependency in an effort to
4841 * reduce the number of dirrem, freefile, and freeblks dependency structures.
4844 clear_remove(struct thread *td)
4846 struct pagedep_hashhead *pagedephd;
4847 struct pagedep *pagedep;
4848 static int next = 0;
4855 for (cnt = 0; cnt < pagedep_hash; cnt++) {
4856 pagedephd = &pagedep_hashtbl[next++];
4857 if (next >= pagedep_hash)
4859 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
4860 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
4862 mp = pagedep->pd_mnt;
4863 ino = pagedep->pd_ino;
4865 if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4866 softdep_error("clear_remove: vget", error);
4869 if ((error = VOP_FSYNC(vp, MNT_NOWAIT)))
4870 softdep_error("clear_remove: fsync", error);
4871 drain_output(vp, 0);
4880 * Clear out a block of dirty inodes in an effort to reduce
4881 * the number of inodedep dependency structures.
4883 struct clear_inodedeps_info {
4889 clear_inodedeps_mountlist_callback(struct mount *mp, void *data)
4891 struct clear_inodedeps_info *info = data;
4893 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) {
4901 clear_inodedeps(struct thread *td)
4903 struct clear_inodedeps_info info;
4904 struct inodedep_hashhead *inodedephd;
4905 struct inodedep *inodedep;
4906 static int next = 0;
4910 ino_t firstino, lastino, ino;
4914 * Pick a random inode dependency to be cleared.
4915 * We will then gather up all the inodes in its block
4916 * that have dependencies and flush them out.
4918 for (cnt = 0; cnt < inodedep_hash; cnt++) {
4919 inodedephd = &inodedep_hashtbl[next++];
4920 if (next >= inodedep_hash)
4922 if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
4925 if (inodedep == NULL) {
4930 * Ugly code to find mount point given pointer to superblock.
4932 fs = inodedep->id_fs;
4935 mountlist_scan(clear_inodedeps_mountlist_callback,
4936 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
4938 * Find the last inode in the block with dependencies.
4940 firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
4941 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
4942 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
4945 * Asynchronously push all but the last inode with dependencies.
4946 * Synchronously push the last inode with dependencies to ensure
4947 * that the inode block gets written to free up the inodedeps.
4949 for (ino = firstino; ino <= lastino; ino++) {
4950 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4953 if ((error = VFS_VGET(info.mp, ino, &vp)) != 0) {
4954 softdep_error("clear_inodedeps: vget", error);
4957 if (ino == lastino) {
4958 if ((error = VOP_FSYNC(vp, MNT_WAIT)))
4959 softdep_error("clear_inodedeps: fsync1", error);
4961 if ((error = VOP_FSYNC(vp, MNT_NOWAIT)))
4962 softdep_error("clear_inodedeps: fsync2", error);
4963 drain_output(vp, 0);
4972 * Function to determine if the buffer has outstanding dependencies
4973 * that will cause a roll-back if the buffer is written. If wantcount
4974 * is set, return number of dependencies, otherwise just yes or no.
4977 softdep_count_dependencies(bp, wantcount)
4981 struct worklist *wk;
4982 struct inodedep *inodedep;
4983 struct indirdep *indirdep;
4984 struct allocindir *aip;
4985 struct pagedep *pagedep;
4991 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4992 switch (wk->wk_type) {
4995 inodedep = WK_INODEDEP(wk);
4996 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4997 /* bitmap allocation dependency */
5002 if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
5003 /* direct block pointer dependency */
5011 indirdep = WK_INDIRDEP(wk);
5013 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
5014 /* indirect block pointer dependency */
5022 pagedep = WK_PAGEDEP(wk);
5023 for (i = 0; i < DAHASHSZ; i++) {
5025 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
5026 /* directory entry dependency */
5038 /* never a dependency on these blocks */
5043 panic("softdep_check_for_rollback: Unexpected type %s",
5044 TYPENAME(wk->wk_type));
5054 * Acquire exclusive access to a buffer.
5055 * Must be called with splbio blocked.
5056 * Return 1 if buffer was acquired.
5059 getdirtybuf(bpp, waitfor)
5067 if ((bp = *bpp) == NULL)
5069 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0)
5071 if (waitfor != MNT_WAIT)
5073 error = interlocked_sleep(&lk, LOCKBUF, bp,
5074 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0);
5075 if (error != ENOLCK) {
5077 panic("getdirtybuf: inconsistent lock");
5080 if ((bp->b_flags & B_DELWRI) == 0) {
5089 * Wait for pending output on a vnode to complete.
5090 * Must be called with vnode locked.
5093 drain_output(vp, islocked)
5100 while (vp->v_track_write.bk_active) {
5101 vp->v_track_write.bk_waitflag = 1;
5102 interlocked_sleep(&lk, SLEEP, &vp->v_track_write,
5110 * Called whenever a buffer that is being invalidated or reallocated
5111 * contains dependencies. This should only happen if an I/O error has
5112 * occurred. The routine is called with the buffer locked.
5115 softdep_deallocate_dependencies(bp)
5119 if ((bp->b_flags & B_ERROR) == 0)
5120 panic("softdep_deallocate_dependencies: dangling deps");
5121 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error);
5122 panic("softdep_deallocate_dependencies: unrecovered I/O error");
5126 * Function to handle asynchronous write errors in the filesystem.
5129 softdep_error(func, error)
5134 /* XXX should do something better! */
5135 printf("%s: got error %d while accessing filesystem\n", func, error);