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 1.53 2007/11/06 03:50:02 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 .io_start = softdep_disk_io_initiation,
210 .io_complete = softdep_disk_write_complete,
211 .io_deallocate = softdep_deallocate_dependencies,
212 .io_fsync = softdep_fsync,
213 .io_sync = softdep_process_worklist,
214 .io_movedeps = softdep_move_dependencies,
215 .io_countdeps = softdep_count_dependencies,
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)
256 acquire_lock(struct lockit *lk)
260 if (lk->lkt_held != NOHOLDER) {
261 holder = lk->lkt_held;
263 if (holder == curthread)
264 panic("softdep_lock: locking against myself");
266 panic("softdep_lock: lock held by %p", holder);
268 crit_enter_id("softupdates");
269 lk->lkt_held = curthread;
274 free_lock(struct lockit *lk)
277 if (lk->lkt_held == NOHOLDER)
278 panic("softdep_unlock: lock not held");
279 lk->lkt_held = NOHOLDER;
280 crit_exit_id("softupdates");
284 * Function to release soft updates lock and panic.
287 softdep_panic(char *msg)
290 if (lk.lkt_held != NOHOLDER)
296 static int interlocked_sleep(struct lockit *, int, void *, int,
300 * When going to sleep, we must save our SPL so that it does
301 * not get lost if some other process uses the lock while we
302 * are sleeping. We restore it after we have slept. This routine
303 * wraps the interlocking with functions that sleep. The list
304 * below enumerates the available set of operations.
311 interlocked_sleep(struct lockit *lk, int op, void *ident, int flags,
312 const char *wmesg, int timo)
319 if (lk->lkt_held == NOHOLDER)
320 panic("interlocked_sleep: lock not held");
321 lk->lkt_held = NOHOLDER;
325 retval = tsleep(ident, flags, wmesg, timo);
328 retval = BUF_LOCK((struct buf *)ident, flags);
331 panic("interlocked_sleep: unknown operation");
334 if (lk->lkt_held != NOHOLDER) {
335 holder = lk->lkt_held;
337 if (holder == curthread)
338 panic("interlocked_sleep: locking against self");
340 panic("interlocked_sleep: lock held by %p", holder);
342 lk->lkt_held = curthread;
350 * Place holder for real semaphores.
359 static void sema_init(struct sema *, char *, int, int);
360 static int sema_get(struct sema *, struct lockit *);
361 static void sema_release(struct sema *);
364 sema_init(struct sema *semap, char *name, int prio, int timo)
367 semap->holder = NOHOLDER;
375 sema_get(struct sema *semap, struct lockit *interlock)
378 if (semap->value++ > 0) {
379 if (interlock != NULL) {
380 interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
381 semap->prio, semap->name, semap->timo);
382 FREE_LOCK(interlock);
384 tsleep((caddr_t)semap, semap->prio, semap->name,
389 semap->holder = curthread;
390 if (interlock != NULL)
391 FREE_LOCK(interlock);
396 sema_release(struct sema *semap)
399 if (semap->value <= 0 || semap->holder != curthread) {
400 if (lk.lkt_held != NOHOLDER)
402 panic("sema_release: not held");
404 if (--semap->value > 0) {
408 semap->holder = NOHOLDER;
412 * Worklist queue management.
413 * These routines require that the lock be held.
415 #ifndef /* NOT */ DEBUG
416 #define WORKLIST_INSERT(head, item) do { \
417 (item)->wk_state |= ONWORKLIST; \
418 LIST_INSERT_HEAD(head, item, wk_list); \
421 #define WORKLIST_INSERT_BP(bp, item) do { \
422 (item)->wk_state |= ONWORKLIST; \
423 (bp)->b_ops = &softdep_bioops; \
424 LIST_INSERT_HEAD(&(bp)->b_dep, item, wk_list); \
427 #define WORKLIST_REMOVE(item) do { \
428 (item)->wk_state &= ~ONWORKLIST; \
429 LIST_REMOVE(item, wk_list); \
432 #define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
435 static void worklist_insert(struct workhead *, struct worklist *);
436 static void worklist_remove(struct worklist *);
437 static void workitem_free(struct worklist *, int);
439 #define WORKLIST_INSERT_BP(bp, item) do { \
440 (bp)->b_ops = &softdep_bioops; \
441 worklist_insert(&(bp)->b_dep, item); \
444 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
445 #define WORKLIST_REMOVE(item) worklist_remove(item)
446 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
449 worklist_insert(struct workhead *head, struct worklist *item)
452 if (lk.lkt_held == NOHOLDER)
453 panic("worklist_insert: lock not held");
454 if (item->wk_state & ONWORKLIST) {
456 panic("worklist_insert: already on list");
458 item->wk_state |= ONWORKLIST;
459 LIST_INSERT_HEAD(head, item, wk_list);
463 worklist_remove(struct worklist *item)
466 if (lk.lkt_held == NOHOLDER)
467 panic("worklist_remove: lock not held");
468 if ((item->wk_state & ONWORKLIST) == 0) {
470 panic("worklist_remove: not on list");
472 item->wk_state &= ~ONWORKLIST;
473 LIST_REMOVE(item, wk_list);
477 workitem_free(struct worklist *item, int type)
480 if (item->wk_state & ONWORKLIST) {
481 if (lk.lkt_held != NOHOLDER)
483 panic("workitem_free: still on list");
485 if (item->wk_type != type) {
486 if (lk.lkt_held != NOHOLDER)
488 panic("workitem_free: type mismatch");
490 FREE(item, DtoM(type));
495 * Workitem queue management
497 static struct workhead softdep_workitem_pending;
498 static int num_on_worklist; /* number of worklist items to be processed */
499 static int softdep_worklist_busy; /* 1 => trying to do unmount */
500 static int softdep_worklist_req; /* serialized waiters */
501 static int max_softdeps; /* maximum number of structs before slowdown */
502 static int tickdelay = 2; /* number of ticks to pause during slowdown */
503 static int *stat_countp; /* statistic to count in proc_waiting timeout */
504 static int proc_waiting; /* tracks whether we have a timeout posted */
505 static struct callout handle; /* handle on posted proc_waiting timeout */
506 static struct thread *filesys_syncer; /* proc of filesystem syncer process */
507 static int req_clear_inodedeps; /* syncer process flush some inodedeps */
508 #define FLUSH_INODES 1
509 static int req_clear_remove; /* syncer process flush some freeblks */
510 #define FLUSH_REMOVE 2
514 static int stat_worklist_push; /* number of worklist cleanups */
515 static int stat_blk_limit_push; /* number of times block limit neared */
516 static int stat_ino_limit_push; /* number of times inode limit neared */
517 static int stat_blk_limit_hit; /* number of times block slowdown imposed */
518 static int stat_ino_limit_hit; /* number of times inode slowdown imposed */
519 static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */
520 static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */
521 static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */
522 static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
523 static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */
526 #include <sys/sysctl.h>
527 SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
528 SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
529 SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
530 SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
531 SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
532 SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
533 SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
534 SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
535 SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
536 SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
537 SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
538 SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
542 * Add an item to the end of the work queue.
543 * This routine requires that the lock be held.
544 * This is the only routine that adds items to the list.
545 * The following routine is the only one that removes items
546 * and does so in order from first to last.
549 add_to_worklist(struct worklist *wk)
551 static struct worklist *worklist_tail;
553 if (wk->wk_state & ONWORKLIST) {
554 if (lk.lkt_held != NOHOLDER)
556 panic("add_to_worklist: already on list");
558 wk->wk_state |= ONWORKLIST;
559 if (LIST_FIRST(&softdep_workitem_pending) == NULL)
560 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
562 LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
564 num_on_worklist += 1;
568 * Process that runs once per second to handle items in the background queue.
570 * Note that we ensure that everything is done in the order in which they
571 * appear in the queue. The code below depends on this property to ensure
572 * that blocks of a file are freed before the inode itself is freed. This
573 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
574 * until all the old ones have been purged from the dependency lists.
577 softdep_process_worklist(struct mount *matchmnt)
579 thread_t td = curthread;
580 int matchcnt, loopcount;
584 * Record the process identifier of our caller so that we can give
585 * this process preferential treatment in request_cleanup below.
591 * There is no danger of having multiple processes run this
592 * code, but we have to single-thread it when softdep_flushfiles()
593 * is in operation to get an accurate count of the number of items
594 * related to its mount point that are in the list.
596 if (matchmnt == NULL) {
597 if (softdep_worklist_busy < 0)
599 softdep_worklist_busy += 1;
603 * If requested, try removing inode or removal dependencies.
605 if (req_clear_inodedeps) {
607 req_clear_inodedeps -= 1;
608 wakeup_one(&proc_waiting);
610 if (req_clear_remove) {
612 req_clear_remove -= 1;
613 wakeup_one(&proc_waiting);
616 starttime = time_second;
617 while (num_on_worklist > 0) {
618 matchcnt += process_worklist_item(matchmnt, 0);
621 * If a umount operation wants to run the worklist
624 if (softdep_worklist_req && matchmnt == NULL) {
630 * If requested, try removing inode or removal dependencies.
632 if (req_clear_inodedeps) {
634 req_clear_inodedeps -= 1;
635 wakeup_one(&proc_waiting);
637 if (req_clear_remove) {
639 req_clear_remove -= 1;
640 wakeup_one(&proc_waiting);
643 * We do not generally want to stop for buffer space, but if
644 * we are really being a buffer hog, we will stop and wait.
646 if (loopcount++ % 128 == 0)
649 * Never allow processing to run for more than one
650 * second. Otherwise the other syncer tasks may get
651 * excessively backlogged.
653 if (starttime != time_second && matchmnt == NULL) {
658 if (matchmnt == NULL) {
659 --softdep_worklist_busy;
660 if (softdep_worklist_req && softdep_worklist_busy == 0)
661 wakeup(&softdep_worklist_req);
667 * Process one item on the worklist.
670 process_worklist_item(struct mount *matchmnt, int flags)
673 struct dirrem *dirrem;
679 if (matchmnt != NULL)
680 matchfs = VFSTOUFS(matchmnt)->um_fs;
683 * Normally we just process each item on the worklist in order.
684 * However, if we are in a situation where we cannot lock any
685 * inodes, we have to skip over any dirrem requests whose
686 * vnodes are resident and locked.
688 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
689 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
691 dirrem = WK_DIRREM(wk);
692 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev,
694 if (vp == NULL || !vn_islocked(vp))
702 num_on_worklist -= 1;
704 switch (wk->wk_type) {
707 /* removal of a directory entry */
708 if (WK_DIRREM(wk)->dm_mnt == matchmnt)
710 handle_workitem_remove(WK_DIRREM(wk));
714 /* releasing blocks and/or fragments from a file */
715 if (WK_FREEBLKS(wk)->fb_fs == matchfs)
717 handle_workitem_freeblocks(WK_FREEBLKS(wk));
721 /* releasing a fragment when replaced as a file grows */
722 if (WK_FREEFRAG(wk)->ff_fs == matchfs)
724 handle_workitem_freefrag(WK_FREEFRAG(wk));
728 /* releasing an inode when its link count drops to 0 */
729 if (WK_FREEFILE(wk)->fx_fs == matchfs)
731 handle_workitem_freefile(WK_FREEFILE(wk));
735 panic("%s_process_worklist: Unknown type %s",
736 "softdep", TYPENAME(wk->wk_type));
743 * Move dependencies from one buffer to another.
746 softdep_move_dependencies(struct buf *oldbp, struct buf *newbp)
748 struct worklist *wk, *wktail;
750 if (LIST_FIRST(&newbp->b_dep) != NULL)
751 panic("softdep_move_dependencies: need merge code");
754 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
755 LIST_REMOVE(wk, wk_list);
757 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
759 LIST_INSERT_AFTER(wktail, wk, wk_list);
761 newbp->b_ops = &softdep_bioops;
767 * Purge the work list of all items associated with a particular mount point.
770 softdep_flushfiles(struct mount *oldmnt, int flags)
776 * Await our turn to clear out the queue, then serialize access.
778 while (softdep_worklist_busy != 0) {
779 softdep_worklist_req += 1;
780 tsleep(&softdep_worklist_req, 0, "softflush", 0);
781 softdep_worklist_req -= 1;
783 softdep_worklist_busy = -1;
785 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) {
786 softdep_worklist_busy = 0;
787 if (softdep_worklist_req)
788 wakeup(&softdep_worklist_req);
792 * Alternately flush the block device associated with the mount
793 * point and process any dependencies that the flushing
794 * creates. In theory, this loop can happen at most twice,
795 * but we give it a few extra just to be sure.
797 devvp = VFSTOUFS(oldmnt)->um_devvp;
798 for (loopcnt = 10; loopcnt > 0; ) {
799 if (softdep_process_worklist(oldmnt) == 0) {
802 * Do another flush in case any vnodes were brought in
803 * as part of the cleanup operations.
805 if ((error = ffs_flushfiles(oldmnt, flags)) != 0)
808 * If we still found nothing to do, we are really done.
810 if (softdep_process_worklist(oldmnt) == 0)
813 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
814 error = VOP_FSYNC(devvp, MNT_WAIT);
819 softdep_worklist_busy = 0;
820 if (softdep_worklist_req)
821 wakeup(&softdep_worklist_req);
824 * If we are unmounting then it is an error to fail. If we
825 * are simply trying to downgrade to read-only, then filesystem
826 * activity can keep us busy forever, so we just fail with EBUSY.
829 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
830 panic("softdep_flushfiles: looping");
839 * There are three types of structures that can be looked up:
840 * 1) pagedep structures identified by mount point, inode number,
842 * 2) inodedep structures identified by mount point and inode number.
843 * 3) newblk structures identified by mount point and
844 * physical block number.
846 * The "pagedep" and "inodedep" dependency structures are hashed
847 * separately from the file blocks and inodes to which they correspond.
848 * This separation helps when the in-memory copy of an inode or
849 * file block must be replaced. It also obviates the need to access
850 * an inode or file page when simply updating (or de-allocating)
851 * dependency structures. Lookup of newblk structures is needed to
852 * find newly allocated blocks when trying to associate them with
853 * their allocdirect or allocindir structure.
855 * The lookup routines optionally create and hash a new instance when
856 * an existing entry is not found.
858 #define DEPALLOC 0x0001 /* allocate structure if lookup fails */
859 #define NODELAY 0x0002 /* cannot do background work */
862 * Structures and routines associated with pagedep caching.
864 LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
865 u_long pagedep_hash; /* size of hash table - 1 */
866 #define PAGEDEP_HASH(mp, inum, lbn) \
867 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
869 static struct sema pagedep_in_progress;
872 * Helper routine for pagedep_lookup()
876 pagedep_find(struct pagedep_hashhead *pagedephd, ino_t ino, ufs_lbn_t lbn,
879 struct pagedep *pagedep;
881 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
882 if (ino == pagedep->pd_ino &&
883 lbn == pagedep->pd_lbn &&
884 mp == pagedep->pd_mnt) {
892 * Look up a pagedep. Return 1 if found, 0 if not found.
893 * If not found, allocate if DEPALLOC flag is passed.
894 * Found or allocated entry is returned in pagedeppp.
895 * This routine must be called with splbio interrupts blocked.
898 pagedep_lookup(struct inode *ip, ufs_lbn_t lbn, int flags,
899 struct pagedep **pagedeppp)
901 struct pagedep *pagedep;
902 struct pagedep_hashhead *pagedephd;
907 if (lk.lkt_held == NOHOLDER)
908 panic("pagedep_lookup: lock not held");
910 mp = ITOV(ip)->v_mount;
911 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
913 *pagedeppp = pagedep_find(pagedephd, ip->i_number, lbn, mp);
916 if ((flags & DEPALLOC) == 0)
918 if (sema_get(&pagedep_in_progress, &lk) == 0) {
922 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
923 M_SOFTDEP_FLAGS | M_ZERO);
925 if (pagedep_find(pagedephd, ip->i_number, lbn, mp)) {
926 kprintf("pagedep_lookup: blocking race avoided\n");
928 sema_release(&pagedep_in_progress);
929 kfree(pagedep, M_PAGEDEP);
933 pagedep->pd_list.wk_type = D_PAGEDEP;
934 pagedep->pd_mnt = mp;
935 pagedep->pd_ino = ip->i_number;
936 pagedep->pd_lbn = lbn;
937 LIST_INIT(&pagedep->pd_dirremhd);
938 LIST_INIT(&pagedep->pd_pendinghd);
939 for (i = 0; i < DAHASHSZ; i++)
940 LIST_INIT(&pagedep->pd_diraddhd[i]);
942 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
943 sema_release(&pagedep_in_progress);
944 *pagedeppp = pagedep;
949 * Structures and routines associated with inodedep caching.
951 LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
952 static u_long inodedep_hash; /* size of hash table - 1 */
953 static long num_inodedep; /* number of inodedep allocated */
954 #define INODEDEP_HASH(fs, inum) \
955 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
956 static struct sema inodedep_in_progress;
959 * Helper routine for inodedep_lookup()
963 inodedep_find(struct inodedep_hashhead *inodedephd, struct fs *fs, ino_t inum)
965 struct inodedep *inodedep;
967 LIST_FOREACH(inodedep, inodedephd, id_hash) {
968 if (inum == inodedep->id_ino && fs == inodedep->id_fs)
975 * Look up a inodedep. Return 1 if found, 0 if not found.
976 * If not found, allocate if DEPALLOC flag is passed.
977 * Found or allocated entry is returned in inodedeppp.
978 * This routine must be called with splbio interrupts blocked.
981 inodedep_lookup(struct fs *fs, ino_t inum, int flags,
982 struct inodedep **inodedeppp)
984 struct inodedep *inodedep;
985 struct inodedep_hashhead *inodedephd;
989 if (lk.lkt_held == NOHOLDER)
990 panic("inodedep_lookup: lock not held");
993 inodedephd = INODEDEP_HASH(fs, inum);
995 *inodedeppp = inodedep_find(inodedephd, fs, inum);
998 if ((flags & DEPALLOC) == 0)
1001 * If we are over our limit, try to improve the situation.
1003 if (num_inodedep > max_softdeps && firsttry &&
1004 speedup_syncer() == 0 && (flags & NODELAY) == 0 &&
1005 request_cleanup(FLUSH_INODES, 1)) {
1009 if (sema_get(&inodedep_in_progress, &lk) == 0) {
1013 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1014 M_INODEDEP, M_SOFTDEP_FLAGS | M_ZERO);
1015 if (inodedep_find(inodedephd, fs, inum)) {
1016 kprintf("inodedep_lookup: blocking race avoided\n");
1018 sema_release(&inodedep_in_progress);
1019 kfree(inodedep, M_INODEDEP);
1022 inodedep->id_list.wk_type = D_INODEDEP;
1023 inodedep->id_fs = fs;
1024 inodedep->id_ino = inum;
1025 inodedep->id_state = ALLCOMPLETE;
1026 inodedep->id_nlinkdelta = 0;
1027 inodedep->id_savedino = NULL;
1028 inodedep->id_savedsize = -1;
1029 inodedep->id_buf = NULL;
1030 LIST_INIT(&inodedep->id_pendinghd);
1031 LIST_INIT(&inodedep->id_inowait);
1032 LIST_INIT(&inodedep->id_bufwait);
1033 TAILQ_INIT(&inodedep->id_inoupdt);
1034 TAILQ_INIT(&inodedep->id_newinoupdt);
1037 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1038 sema_release(&inodedep_in_progress);
1039 *inodedeppp = inodedep;
1044 * Structures and routines associated with newblk caching.
1046 LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1047 u_long newblk_hash; /* size of hash table - 1 */
1048 #define NEWBLK_HASH(fs, inum) \
1049 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1050 static struct sema newblk_in_progress;
1053 * Helper routine for newblk_lookup()
1057 newblk_find(struct newblk_hashhead *newblkhd, struct fs *fs,
1058 ufs_daddr_t newblkno)
1060 struct newblk *newblk;
1062 LIST_FOREACH(newblk, newblkhd, nb_hash) {
1063 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1070 * Look up a newblk. Return 1 if found, 0 if not found.
1071 * If not found, allocate if DEPALLOC flag is passed.
1072 * Found or allocated entry is returned in newblkpp.
1075 newblk_lookup(struct fs *fs, ufs_daddr_t newblkno, int flags,
1076 struct newblk **newblkpp)
1078 struct newblk *newblk;
1079 struct newblk_hashhead *newblkhd;
1081 newblkhd = NEWBLK_HASH(fs, newblkno);
1083 *newblkpp = newblk_find(newblkhd, fs, newblkno);
1086 if ((flags & DEPALLOC) == 0)
1088 if (sema_get(&newblk_in_progress, 0) == 0)
1090 MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1091 M_NEWBLK, M_SOFTDEP_FLAGS | M_ZERO);
1093 if (newblk_find(newblkhd, fs, newblkno)) {
1094 kprintf("newblk_lookup: blocking race avoided\n");
1095 sema_release(&pagedep_in_progress);
1096 kfree(newblk, M_NEWBLK);
1099 newblk->nb_state = 0;
1101 newblk->nb_newblkno = newblkno;
1102 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1103 sema_release(&newblk_in_progress);
1109 * Executed during filesystem system initialization before
1110 * mounting any filesystems.
1113 softdep_initialize(void)
1115 callout_init(&handle);
1117 LIST_INIT(&mkdirlisthd);
1118 LIST_INIT(&softdep_workitem_pending);
1119 max_softdeps = min(desiredvnodes * 8,
1120 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep)));
1121 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1123 sema_init(&pagedep_in_progress, "pagedep", 0, 0);
1124 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
1125 sema_init(&inodedep_in_progress, "inodedep", 0, 0);
1126 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
1127 sema_init(&newblk_in_progress, "newblk", 0, 0);
1128 add_bio_ops(&softdep_bioops);
1132 * Called at mount time to notify the dependency code that a
1133 * filesystem wishes to use it.
1136 softdep_mount(struct vnode *devvp, struct mount *mp, struct fs *fs)
1138 struct csum cstotal;
1143 mp->mnt_flag &= ~MNT_ASYNC;
1144 mp->mnt_flag |= MNT_SOFTDEP;
1145 mp->mnt_bioops = &softdep_bioops;
1147 * When doing soft updates, the counters in the
1148 * superblock may have gotten out of sync, so we have
1149 * to scan the cylinder groups and recalculate them.
1151 if (fs->fs_clean != 0)
1153 bzero(&cstotal, sizeof cstotal);
1154 for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1155 if ((error = bread(devvp, fsbtodoff(fs, cgtod(fs, cyl)),
1156 fs->fs_cgsize, &bp)) != 0) {
1160 cgp = (struct cg *)bp->b_data;
1161 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1162 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1163 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1164 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1165 fs->fs_cs(fs, cyl) = cgp->cg_cs;
1169 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1170 kprintf("ffs_mountfs: superblock updated for soft updates\n");
1172 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1177 * Protecting the freemaps (or bitmaps).
1179 * To eliminate the need to execute fsck before mounting a filesystem
1180 * after a power failure, one must (conservatively) guarantee that the
1181 * on-disk copy of the bitmaps never indicate that a live inode or block is
1182 * free. So, when a block or inode is allocated, the bitmap should be
1183 * updated (on disk) before any new pointers. When a block or inode is
1184 * freed, the bitmap should not be updated until all pointers have been
1185 * reset. The latter dependency is handled by the delayed de-allocation
1186 * approach described below for block and inode de-allocation. The former
1187 * dependency is handled by calling the following procedure when a block or
1188 * inode is allocated. When an inode is allocated an "inodedep" is created
1189 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1190 * Each "inodedep" is also inserted into the hash indexing structure so
1191 * that any additional link additions can be made dependent on the inode
1194 * The ufs filesystem maintains a number of free block counts (e.g., per
1195 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1196 * in addition to the bitmaps. These counts are used to improve efficiency
1197 * during allocation and therefore must be consistent with the bitmaps.
1198 * There is no convenient way to guarantee post-crash consistency of these
1199 * counts with simple update ordering, for two main reasons: (1) The counts
1200 * and bitmaps for a single cylinder group block are not in the same disk
1201 * sector. If a disk write is interrupted (e.g., by power failure), one may
1202 * be written and the other not. (2) Some of the counts are located in the
1203 * superblock rather than the cylinder group block. So, we focus our soft
1204 * updates implementation on protecting the bitmaps. When mounting a
1205 * filesystem, we recompute the auxiliary counts from the bitmaps.
1209 * Called just after updating the cylinder group block to allocate an inode.
1212 * bp: buffer for cylgroup block with inode map
1213 * ip: inode related to allocation
1214 * newinum: new inode number being allocated
1217 softdep_setup_inomapdep(struct buf *bp, struct inode *ip, ino_t newinum)
1219 struct inodedep *inodedep;
1220 struct bmsafemap *bmsafemap;
1223 * Create a dependency for the newly allocated inode.
1224 * Panic if it already exists as something is seriously wrong.
1225 * Otherwise add it to the dependency list for the buffer holding
1226 * the cylinder group map from which it was allocated.
1229 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1231 panic("softdep_setup_inomapdep: found inode");
1233 inodedep->id_buf = bp;
1234 inodedep->id_state &= ~DEPCOMPLETE;
1235 bmsafemap = bmsafemap_lookup(bp);
1236 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1241 * Called just after updating the cylinder group block to
1242 * allocate block or fragment.
1245 * bp: buffer for cylgroup block with block map
1246 * fs: filesystem doing allocation
1247 * newblkno: number of newly allocated block
1250 softdep_setup_blkmapdep(struct buf *bp, struct fs *fs,
1251 ufs_daddr_t newblkno)
1253 struct newblk *newblk;
1254 struct bmsafemap *bmsafemap;
1257 * Create a dependency for the newly allocated block.
1258 * Add it to the dependency list for the buffer holding
1259 * the cylinder group map from which it was allocated.
1261 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1262 panic("softdep_setup_blkmapdep: found block");
1264 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1265 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1270 * Find the bmsafemap associated with a cylinder group buffer.
1271 * If none exists, create one. The buffer must be locked when
1272 * this routine is called and this routine must be called with
1273 * splbio interrupts blocked.
1275 static struct bmsafemap *
1276 bmsafemap_lookup(struct buf *bp)
1278 struct bmsafemap *bmsafemap;
1279 struct worklist *wk;
1282 if (lk.lkt_held == NOHOLDER)
1283 panic("bmsafemap_lookup: lock not held");
1285 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1286 if (wk->wk_type == D_BMSAFEMAP)
1287 return (WK_BMSAFEMAP(wk));
1290 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1291 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1292 bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1293 bmsafemap->sm_list.wk_state = 0;
1294 bmsafemap->sm_buf = bp;
1295 LIST_INIT(&bmsafemap->sm_allocdirecthd);
1296 LIST_INIT(&bmsafemap->sm_allocindirhd);
1297 LIST_INIT(&bmsafemap->sm_inodedephd);
1298 LIST_INIT(&bmsafemap->sm_newblkhd);
1300 WORKLIST_INSERT_BP(bp, &bmsafemap->sm_list);
1305 * Direct block allocation dependencies.
1307 * When a new block is allocated, the corresponding disk locations must be
1308 * initialized (with zeros or new data) before the on-disk inode points to
1309 * them. Also, the freemap from which the block was allocated must be
1310 * updated (on disk) before the inode's pointer. These two dependencies are
1311 * independent of each other and are needed for all file blocks and indirect
1312 * blocks that are pointed to directly by the inode. Just before the
1313 * "in-core" version of the inode is updated with a newly allocated block
1314 * number, a procedure (below) is called to setup allocation dependency
1315 * structures. These structures are removed when the corresponding
1316 * dependencies are satisfied or when the block allocation becomes obsolete
1317 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1318 * fragment that gets upgraded). All of these cases are handled in
1319 * procedures described later.
1321 * When a file extension causes a fragment to be upgraded, either to a larger
1322 * fragment or to a full block, the on-disk location may change (if the
1323 * previous fragment could not simply be extended). In this case, the old
1324 * fragment must be de-allocated, but not until after the inode's pointer has
1325 * been updated. In most cases, this is handled by later procedures, which
1326 * will construct a "freefrag" structure to be added to the workitem queue
1327 * when the inode update is complete (or obsolete). The main exception to
1328 * this is when an allocation occurs while a pending allocation dependency
1329 * (for the same block pointer) remains. This case is handled in the main
1330 * allocation dependency setup procedure by immediately freeing the
1331 * unreferenced fragments.
1334 * ip: inode to which block is being added
1335 * lbn: block pointer within inode
1336 * newblkno: disk block number being added
1337 * oldblkno: previous block number, 0 unless frag
1338 * newsize: size of new block
1339 * oldsize: size of new block
1340 * bp: bp for allocated block
1343 softdep_setup_allocdirect(struct inode *ip, ufs_lbn_t lbn, ufs_daddr_t newblkno,
1344 ufs_daddr_t oldblkno, long newsize, long oldsize,
1347 struct allocdirect *adp, *oldadp;
1348 struct allocdirectlst *adphead;
1349 struct bmsafemap *bmsafemap;
1350 struct inodedep *inodedep;
1351 struct pagedep *pagedep;
1352 struct newblk *newblk;
1354 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1355 M_ALLOCDIRECT, M_SOFTDEP_FLAGS);
1356 bzero(adp, sizeof(struct allocdirect));
1357 adp->ad_list.wk_type = D_ALLOCDIRECT;
1359 adp->ad_newblkno = newblkno;
1360 adp->ad_oldblkno = oldblkno;
1361 adp->ad_newsize = newsize;
1362 adp->ad_oldsize = oldsize;
1363 adp->ad_state = ATTACHED;
1364 if (newblkno == oldblkno)
1365 adp->ad_freefrag = NULL;
1367 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1369 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1370 panic("softdep_setup_allocdirect: lost block");
1373 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1374 adp->ad_inodedep = inodedep;
1376 if (newblk->nb_state == DEPCOMPLETE) {
1377 adp->ad_state |= DEPCOMPLETE;
1380 bmsafemap = newblk->nb_bmsafemap;
1381 adp->ad_buf = bmsafemap->sm_buf;
1382 LIST_REMOVE(newblk, nb_deps);
1383 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1385 LIST_REMOVE(newblk, nb_hash);
1386 FREE(newblk, M_NEWBLK);
1388 WORKLIST_INSERT_BP(bp, &adp->ad_list);
1389 if (lbn >= NDADDR) {
1390 /* allocating an indirect block */
1391 if (oldblkno != 0) {
1393 panic("softdep_setup_allocdirect: non-zero indir");
1397 * Allocating a direct block.
1399 * If we are allocating a directory block, then we must
1400 * allocate an associated pagedep to track additions and
1403 if ((ip->i_mode & IFMT) == IFDIR &&
1404 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) {
1405 WORKLIST_INSERT_BP(bp, &pagedep->pd_list);
1409 * The list of allocdirects must be kept in sorted and ascending
1410 * order so that the rollback routines can quickly determine the
1411 * first uncommitted block (the size of the file stored on disk
1412 * ends at the end of the lowest committed fragment, or if there
1413 * are no fragments, at the end of the highest committed block).
1414 * Since files generally grow, the typical case is that the new
1415 * block is to be added at the end of the list. We speed this
1416 * special case by checking against the last allocdirect in the
1417 * list before laboriously traversing the list looking for the
1420 adphead = &inodedep->id_newinoupdt;
1421 oldadp = TAILQ_LAST(adphead, allocdirectlst);
1422 if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1423 /* insert at end of list */
1424 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1425 if (oldadp != NULL && oldadp->ad_lbn == lbn)
1426 allocdirect_merge(adphead, adp, oldadp);
1430 TAILQ_FOREACH(oldadp, adphead, ad_next) {
1431 if (oldadp->ad_lbn >= lbn)
1434 if (oldadp == NULL) {
1436 panic("softdep_setup_allocdirect: lost entry");
1438 /* insert in middle of list */
1439 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1440 if (oldadp->ad_lbn == lbn)
1441 allocdirect_merge(adphead, adp, oldadp);
1446 * Replace an old allocdirect dependency with a newer one.
1447 * This routine must be called with splbio interrupts blocked.
1450 * adphead: head of list holding allocdirects
1451 * newadp: allocdirect being added
1452 * oldadp: existing allocdirect being checked
1455 allocdirect_merge(struct allocdirectlst *adphead,
1456 struct allocdirect *newadp,
1457 struct allocdirect *oldadp)
1459 struct freefrag *freefrag;
1462 if (lk.lkt_held == NOHOLDER)
1463 panic("allocdirect_merge: lock not held");
1465 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1466 newadp->ad_oldsize != oldadp->ad_newsize ||
1467 newadp->ad_lbn >= NDADDR) {
1469 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1470 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1473 newadp->ad_oldblkno = oldadp->ad_oldblkno;
1474 newadp->ad_oldsize = oldadp->ad_oldsize;
1476 * If the old dependency had a fragment to free or had never
1477 * previously had a block allocated, then the new dependency
1478 * can immediately post its freefrag and adopt the old freefrag.
1479 * This action is done by swapping the freefrag dependencies.
1480 * The new dependency gains the old one's freefrag, and the
1481 * old one gets the new one and then immediately puts it on
1482 * the worklist when it is freed by free_allocdirect. It is
1483 * not possible to do this swap when the old dependency had a
1484 * non-zero size but no previous fragment to free. This condition
1485 * arises when the new block is an extension of the old block.
1486 * Here, the first part of the fragment allocated to the new
1487 * dependency is part of the block currently claimed on disk by
1488 * the old dependency, so cannot legitimately be freed until the
1489 * conditions for the new dependency are fulfilled.
1491 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1492 freefrag = newadp->ad_freefrag;
1493 newadp->ad_freefrag = oldadp->ad_freefrag;
1494 oldadp->ad_freefrag = freefrag;
1496 free_allocdirect(adphead, oldadp, 0);
1500 * Allocate a new freefrag structure if needed.
1502 static struct freefrag *
1503 newfreefrag(struct inode *ip, ufs_daddr_t blkno, long size)
1505 struct freefrag *freefrag;
1511 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1512 panic("newfreefrag: frag size");
1513 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1514 M_FREEFRAG, M_SOFTDEP_FLAGS);
1515 freefrag->ff_list.wk_type = D_FREEFRAG;
1516 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */
1517 freefrag->ff_inum = ip->i_number;
1518 freefrag->ff_fs = fs;
1519 freefrag->ff_devvp = ip->i_devvp;
1520 freefrag->ff_blkno = blkno;
1521 freefrag->ff_fragsize = size;
1526 * This workitem de-allocates fragments that were replaced during
1527 * file block allocation.
1530 handle_workitem_freefrag(struct freefrag *freefrag)
1534 tip.i_fs = freefrag->ff_fs;
1535 tip.i_devvp = freefrag->ff_devvp;
1536 tip.i_dev = freefrag->ff_devvp->v_rdev;
1537 tip.i_number = freefrag->ff_inum;
1538 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */
1539 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1540 FREE(freefrag, M_FREEFRAG);
1544 * Indirect block allocation dependencies.
1546 * The same dependencies that exist for a direct block also exist when
1547 * a new block is allocated and pointed to by an entry in a block of
1548 * indirect pointers. The undo/redo states described above are also
1549 * used here. Because an indirect block contains many pointers that
1550 * may have dependencies, a second copy of the entire in-memory indirect
1551 * block is kept. The buffer cache copy is always completely up-to-date.
1552 * The second copy, which is used only as a source for disk writes,
1553 * contains only the safe pointers (i.e., those that have no remaining
1554 * update dependencies). The second copy is freed when all pointers
1555 * are safe. The cache is not allowed to replace indirect blocks with
1556 * pending update dependencies. If a buffer containing an indirect
1557 * block with dependencies is written, these routines will mark it
1558 * dirty again. It can only be successfully written once all the
1559 * dependencies are removed. The ffs_fsync routine in conjunction with
1560 * softdep_sync_metadata work together to get all the dependencies
1561 * removed so that a file can be successfully written to disk. Three
1562 * procedures are used when setting up indirect block pointer
1563 * dependencies. The division is necessary because of the organization
1564 * of the "balloc" routine and because of the distinction between file
1565 * pages and file metadata blocks.
1569 * Allocate a new allocindir structure.
1572 * ip: inode for file being extended
1573 * ptrno: offset of pointer in indirect block
1574 * newblkno: disk block number being added
1575 * oldblkno: previous block number, 0 if none
1577 static struct allocindir *
1578 newallocindir(struct inode *ip, int ptrno, ufs_daddr_t newblkno,
1579 ufs_daddr_t oldblkno)
1581 struct allocindir *aip;
1583 MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1584 M_ALLOCINDIR, M_SOFTDEP_FLAGS);
1585 bzero(aip, sizeof(struct allocindir));
1586 aip->ai_list.wk_type = D_ALLOCINDIR;
1587 aip->ai_state = ATTACHED;
1588 aip->ai_offset = ptrno;
1589 aip->ai_newblkno = newblkno;
1590 aip->ai_oldblkno = oldblkno;
1591 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1596 * Called just before setting an indirect block pointer
1597 * to a newly allocated file page.
1600 * ip: inode for file being extended
1601 * lbn: allocated block number within file
1602 * bp: buffer with indirect blk referencing page
1603 * ptrno: offset of pointer in indirect block
1604 * newblkno: disk block number being added
1605 * oldblkno: previous block number, 0 if none
1606 * nbp: buffer holding allocated page
1609 softdep_setup_allocindir_page(struct inode *ip, ufs_lbn_t lbn,
1610 struct buf *bp, int ptrno,
1611 ufs_daddr_t newblkno, ufs_daddr_t oldblkno,
1614 struct allocindir *aip;
1615 struct pagedep *pagedep;
1617 aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1620 * If we are allocating a directory page, then we must
1621 * allocate an associated pagedep to track additions and
1624 if ((ip->i_mode & IFMT) == IFDIR &&
1625 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1626 WORKLIST_INSERT_BP(nbp, &pagedep->pd_list);
1627 WORKLIST_INSERT_BP(nbp, &aip->ai_list);
1629 setup_allocindir_phase2(bp, ip, aip);
1633 * Called just before setting an indirect block pointer to a
1634 * newly allocated indirect block.
1636 * nbp: newly allocated indirect block
1637 * ip: inode for file being extended
1638 * bp: indirect block referencing allocated block
1639 * ptrno: offset of pointer in indirect block
1640 * newblkno: disk block number being added
1643 softdep_setup_allocindir_meta(struct buf *nbp, struct inode *ip,
1644 struct buf *bp, int ptrno,
1645 ufs_daddr_t newblkno)
1647 struct allocindir *aip;
1649 aip = newallocindir(ip, ptrno, newblkno, 0);
1651 WORKLIST_INSERT_BP(nbp, &aip->ai_list);
1653 setup_allocindir_phase2(bp, ip, aip);
1657 * Called to finish the allocation of the "aip" allocated
1658 * by one of the two routines above.
1661 * bp: in-memory copy of the indirect block
1662 * ip: inode for file being extended
1663 * aip: allocindir allocated by the above routines
1666 setup_allocindir_phase2(struct buf *bp, struct inode *ip,
1667 struct allocindir *aip)
1669 struct worklist *wk;
1670 struct indirdep *indirdep, *newindirdep;
1671 struct bmsafemap *bmsafemap;
1672 struct allocindir *oldaip;
1673 struct freefrag *freefrag;
1674 struct newblk *newblk;
1676 if (bp->b_loffset >= 0)
1677 panic("setup_allocindir_phase2: not indir blk");
1678 for (indirdep = NULL, newindirdep = NULL; ; ) {
1680 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1681 if (wk->wk_type != D_INDIRDEP)
1683 indirdep = WK_INDIRDEP(wk);
1686 if (indirdep == NULL && newindirdep) {
1687 indirdep = newindirdep;
1688 WORKLIST_INSERT_BP(bp, &indirdep->ir_list);
1693 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1695 panic("setup_allocindir: lost block");
1697 if (newblk->nb_state == DEPCOMPLETE) {
1698 aip->ai_state |= DEPCOMPLETE;
1701 bmsafemap = newblk->nb_bmsafemap;
1702 aip->ai_buf = bmsafemap->sm_buf;
1703 LIST_REMOVE(newblk, nb_deps);
1704 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1707 LIST_REMOVE(newblk, nb_hash);
1708 FREE(newblk, M_NEWBLK);
1709 aip->ai_indirdep = indirdep;
1711 * Check to see if there is an existing dependency
1712 * for this block. If there is, merge the old
1713 * dependency into the new one.
1715 if (aip->ai_oldblkno == 0)
1719 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1720 if (oldaip->ai_offset == aip->ai_offset)
1722 if (oldaip != NULL) {
1723 if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1725 panic("setup_allocindir_phase2: blkno");
1727 aip->ai_oldblkno = oldaip->ai_oldblkno;
1728 freefrag = oldaip->ai_freefrag;
1729 oldaip->ai_freefrag = aip->ai_freefrag;
1730 aip->ai_freefrag = freefrag;
1731 free_allocindir(oldaip, NULL);
1733 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1734 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1735 [aip->ai_offset] = aip->ai_oldblkno;
1740 * Avoid any possibility of data corruption by
1741 * ensuring that our old version is thrown away.
1743 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
1744 brelse(newindirdep->ir_savebp);
1745 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1749 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1750 M_INDIRDEP, M_SOFTDEP_FLAGS);
1751 newindirdep->ir_list.wk_type = D_INDIRDEP;
1752 newindirdep->ir_state = ATTACHED;
1753 LIST_INIT(&newindirdep->ir_deplisthd);
1754 LIST_INIT(&newindirdep->ir_donehd);
1755 if (bp->b_bio2.bio_offset == NOOFFSET) {
1756 VOP_BMAP(bp->b_vp, bp->b_bio1.bio_offset,
1757 &bp->b_bio2.bio_offset, NULL, NULL);
1759 KKASSERT(bp->b_bio2.bio_offset != NOOFFSET);
1760 newindirdep->ir_savebp = getblk(ip->i_devvp,
1761 bp->b_bio2.bio_offset,
1762 bp->b_bcount, 0, 0);
1763 BUF_KERNPROC(newindirdep->ir_savebp);
1764 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1769 * Block de-allocation dependencies.
1771 * When blocks are de-allocated, the on-disk pointers must be nullified before
1772 * the blocks are made available for use by other files. (The true
1773 * requirement is that old pointers must be nullified before new on-disk
1774 * pointers are set. We chose this slightly more stringent requirement to
1775 * reduce complexity.) Our implementation handles this dependency by updating
1776 * the inode (or indirect block) appropriately but delaying the actual block
1777 * de-allocation (i.e., freemap and free space count manipulation) until
1778 * after the updated versions reach stable storage. After the disk is
1779 * updated, the blocks can be safely de-allocated whenever it is convenient.
1780 * This implementation handles only the common case of reducing a file's
1781 * length to zero. Other cases are handled by the conventional synchronous
1784 * The ffs implementation with which we worked double-checks
1785 * the state of the block pointers and file size as it reduces
1786 * a file's length. Some of this code is replicated here in our
1787 * soft updates implementation. The freeblks->fb_chkcnt field is
1788 * used to transfer a part of this information to the procedure
1789 * that eventually de-allocates the blocks.
1791 * This routine should be called from the routine that shortens
1792 * a file's length, before the inode's size or block pointers
1793 * are modified. It will save the block pointer information for
1794 * later release and zero the inode so that the calling routine
1797 struct softdep_setup_freeblocks_info {
1802 static int softdep_setup_freeblocks_bp(struct buf *bp, void *data);
1806 * ip: The inode whose length is to be reduced
1807 * length: The new length for the file
1810 softdep_setup_freeblocks(struct inode *ip, off_t length)
1812 struct softdep_setup_freeblocks_info info;
1813 struct freeblks *freeblks;
1814 struct inodedep *inodedep;
1815 struct allocdirect *adp;
1819 int i, error, delay;
1824 panic("softde_setup_freeblocks: non-zero length");
1825 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1826 M_FREEBLKS, M_SOFTDEP_FLAGS);
1827 bzero(freeblks, sizeof(struct freeblks));
1828 freeblks->fb_list.wk_type = D_FREEBLKS;
1829 freeblks->fb_state = ATTACHED;
1830 freeblks->fb_uid = ip->i_uid;
1831 freeblks->fb_previousinum = ip->i_number;
1832 freeblks->fb_devvp = ip->i_devvp;
1833 freeblks->fb_fs = fs;
1834 freeblks->fb_oldsize = ip->i_size;
1835 freeblks->fb_newsize = length;
1836 freeblks->fb_chkcnt = ip->i_blocks;
1837 for (i = 0; i < NDADDR; i++) {
1838 freeblks->fb_dblks[i] = ip->i_db[i];
1841 for (i = 0; i < NIADDR; i++) {
1842 freeblks->fb_iblks[i] = ip->i_ib[i];
1848 * Push the zero'ed inode to to its disk buffer so that we are free
1849 * to delete its dependencies below. Once the dependencies are gone
1850 * the buffer can be safely released.
1852 if ((error = bread(ip->i_devvp,
1853 fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)),
1854 (int)fs->fs_bsize, &bp)) != 0)
1855 softdep_error("softdep_setup_freeblocks", error);
1856 *((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
1859 * Find and eliminate any inode dependencies.
1862 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1863 if ((inodedep->id_state & IOSTARTED) != 0) {
1865 panic("softdep_setup_freeblocks: inode busy");
1868 * Add the freeblks structure to the list of operations that
1869 * must await the zero'ed inode being written to disk. If we
1870 * still have a bitmap dependency (delay == 0), then the inode
1871 * has never been written to disk, so we can process the
1872 * freeblks below once we have deleted the dependencies.
1874 delay = (inodedep->id_state & DEPCOMPLETE);
1876 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1878 * Because the file length has been truncated to zero, any
1879 * pending block allocation dependency structures associated
1880 * with this inode are obsolete and can simply be de-allocated.
1881 * We must first merge the two dependency lists to get rid of
1882 * any duplicate freefrag structures, then purge the merged list.
1884 merge_inode_lists(inodedep);
1885 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1886 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
1890 * We must wait for any I/O in progress to finish so that
1891 * all potential buffers on the dirty list will be visible.
1892 * Once they are all there, walk the list and get rid of
1897 drain_output(vp, 1);
1902 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
1903 softdep_setup_freeblocks_bp, &info);
1904 } while (count != 0);
1905 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
1906 (void)free_inodedep(inodedep);
1909 freeblks->fb_state |= DEPCOMPLETE;
1911 * If the inode with zeroed block pointers is now on disk
1912 * we can start freeing blocks. Add freeblks to the worklist
1913 * instead of calling handle_workitem_freeblocks directly as
1914 * it is more likely that additional IO is needed to complete
1915 * the request here than in the !delay case.
1917 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
1918 add_to_worklist(&freeblks->fb_list);
1923 * If the inode has never been written to disk (delay == 0),
1924 * then we can process the freeblks now that we have deleted
1928 handle_workitem_freeblocks(freeblks);
1932 softdep_setup_freeblocks_bp(struct buf *bp, void *data)
1934 struct softdep_setup_freeblocks_info *info = data;
1935 struct inodedep *inodedep;
1937 if (getdirtybuf(&bp, MNT_WAIT) == 0) {
1938 kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp);
1941 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) {
1942 kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp);
1946 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep);
1947 deallocate_dependencies(bp, inodedep);
1948 bp->b_flags |= B_INVAL | B_NOCACHE;
1956 * Reclaim any dependency structures from a buffer that is about to
1957 * be reallocated to a new vnode. The buffer must be locked, thus,
1958 * no I/O completion operations can occur while we are manipulating
1959 * its associated dependencies. The mutex is held so that other I/O's
1960 * associated with related dependencies do not occur.
1963 deallocate_dependencies(struct buf *bp, struct inodedep *inodedep)
1965 struct worklist *wk;
1966 struct indirdep *indirdep;
1967 struct allocindir *aip;
1968 struct pagedep *pagedep;
1969 struct dirrem *dirrem;
1973 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1974 switch (wk->wk_type) {
1977 indirdep = WK_INDIRDEP(wk);
1979 * None of the indirect pointers will ever be visible,
1980 * so they can simply be tossed. GOINGAWAY ensures
1981 * that allocated pointers will be saved in the buffer
1982 * cache until they are freed. Note that they will
1983 * only be able to be found by their physical address
1984 * since the inode mapping the logical address will
1985 * be gone. The save buffer used for the safe copy
1986 * was allocated in setup_allocindir_phase2 using
1987 * the physical address so it could be used for this
1988 * purpose. Hence we swap the safe copy with the real
1989 * copy, allowing the safe copy to be freed and holding
1990 * on to the real copy for later use in indir_trunc.
1992 * NOTE: ir_savebp is relative to the block device
1993 * so b_bio1 contains the device block number.
1995 if (indirdep->ir_state & GOINGAWAY) {
1997 panic("deallocate_dependencies: already gone");
1999 indirdep->ir_state |= GOINGAWAY;
2000 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
2001 free_allocindir(aip, inodedep);
2002 if (bp->b_bio1.bio_offset >= 0 ||
2003 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) {
2005 panic("deallocate_dependencies: not indir");
2007 bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2009 WORKLIST_REMOVE(wk);
2010 WORKLIST_INSERT_BP(indirdep->ir_savebp, wk);
2014 pagedep = WK_PAGEDEP(wk);
2016 * None of the directory additions will ever be
2017 * visible, so they can simply be tossed.
2019 for (i = 0; i < DAHASHSZ; i++)
2021 LIST_FIRST(&pagedep->pd_diraddhd[i])))
2023 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2026 * Copy any directory remove dependencies to the list
2027 * to be processed after the zero'ed inode is written.
2028 * If the inode has already been written, then they
2029 * can be dumped directly onto the work list.
2031 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2032 LIST_REMOVE(dirrem, dm_next);
2033 dirrem->dm_dirinum = pagedep->pd_ino;
2034 if (inodedep == NULL ||
2035 (inodedep->id_state & ALLCOMPLETE) ==
2037 add_to_worklist(&dirrem->dm_list);
2039 WORKLIST_INSERT(&inodedep->id_bufwait,
2042 WORKLIST_REMOVE(&pagedep->pd_list);
2043 LIST_REMOVE(pagedep, pd_hash);
2044 WORKITEM_FREE(pagedep, D_PAGEDEP);
2048 free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2054 panic("deallocate_dependencies: Unexpected type %s",
2055 TYPENAME(wk->wk_type));
2060 panic("deallocate_dependencies: Unknown type %s",
2061 TYPENAME(wk->wk_type));
2068 * Free an allocdirect. Generate a new freefrag work request if appropriate.
2069 * This routine must be called with splbio interrupts blocked.
2072 free_allocdirect(struct allocdirectlst *adphead,
2073 struct allocdirect *adp, int delay)
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(struct vnode *pvp, ino_t ino, int mode)
2102 struct inode *ip = VTOI(pvp);
2103 struct inodedep *inodedep;
2104 struct freefile *freefile;
2107 * This sets up the inode de-allocation dependency.
2109 MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2110 M_FREEFILE, M_SOFTDEP_FLAGS);
2111 freefile->fx_list.wk_type = D_FREEFILE;
2112 freefile->fx_list.wk_state = 0;
2113 freefile->fx_mode = mode;
2114 freefile->fx_oldinum = ino;
2115 freefile->fx_devvp = ip->i_devvp;
2116 freefile->fx_fs = ip->i_fs;
2119 * If the inodedep does not exist, then the zero'ed inode has
2120 * been written to disk. If the allocated inode has never been
2121 * written to disk, then the on-disk inode is zero'ed. In either
2122 * case we can free the file immediately.
2125 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2126 check_inode_unwritten(inodedep)) {
2128 handle_workitem_freefile(freefile);
2131 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2136 * Check to see if an inode has never been written to disk. If
2137 * so free the inodedep and return success, otherwise return failure.
2138 * This routine must be called with splbio interrupts blocked.
2140 * If we still have a bitmap dependency, then the inode has never
2141 * been written to disk. Drop the dependency as it is no longer
2142 * necessary since the inode is being deallocated. We set the
2143 * ALLCOMPLETE flags since the bitmap now properly shows that the
2144 * inode is not allocated. Even if the inode is actively being
2145 * written, it has been rolled back to its zero'ed state, so we
2146 * are ensured that a zero inode is what is on the disk. For short
2147 * lived files, this change will usually result in removing all the
2148 * dependencies from the inode so that it can be freed immediately.
2151 check_inode_unwritten(struct inodedep *inodedep)
2154 if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2155 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2156 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2157 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2158 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2159 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2160 inodedep->id_nlinkdelta != 0)
2164 * Another process might be in initiate_write_inodeblock
2165 * trying to allocate memory without holding "Softdep Lock".
2167 if ((inodedep->id_state & IOSTARTED) != 0 &&
2168 inodedep->id_savedino == NULL)
2171 inodedep->id_state |= ALLCOMPLETE;
2172 LIST_REMOVE(inodedep, id_deps);
2173 inodedep->id_buf = NULL;
2174 if (inodedep->id_state & ONWORKLIST)
2175 WORKLIST_REMOVE(&inodedep->id_list);
2176 if (inodedep->id_savedino != NULL) {
2177 FREE(inodedep->id_savedino, M_INODEDEP);
2178 inodedep->id_savedino = NULL;
2180 if (free_inodedep(inodedep) == 0) {
2182 panic("check_inode_unwritten: busy inode");
2188 * Try to free an inodedep structure. Return 1 if it could be freed.
2191 free_inodedep(struct inodedep *inodedep)
2194 if ((inodedep->id_state & ONWORKLIST) != 0 ||
2195 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2196 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2197 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2198 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2199 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2200 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2201 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
2203 LIST_REMOVE(inodedep, id_hash);
2204 WORKITEM_FREE(inodedep, D_INODEDEP);
2210 * This workitem routine performs the block de-allocation.
2211 * The workitem is added to the pending list after the updated
2212 * inode block has been written to disk. As mentioned above,
2213 * checks regarding the number of blocks de-allocated (compared
2214 * to the number of blocks allocated for the file) are also
2215 * performed in this function.
2218 handle_workitem_freeblocks(struct freeblks *freeblks)
2223 int i, level, bsize;
2224 long nblocks, blocksreleased = 0;
2225 int error, allerror = 0;
2226 ufs_lbn_t baselbns[NIADDR], tmpval;
2228 tip.i_number = freeblks->fb_previousinum;
2229 tip.i_devvp = freeblks->fb_devvp;
2230 tip.i_dev = freeblks->fb_devvp->v_rdev;
2231 tip.i_fs = freeblks->fb_fs;
2232 tip.i_size = freeblks->fb_oldsize;
2233 tip.i_uid = freeblks->fb_uid;
2234 fs = freeblks->fb_fs;
2236 baselbns[0] = NDADDR;
2237 for (i = 1; i < NIADDR; i++) {
2238 tmpval *= NINDIR(fs);
2239 baselbns[i] = baselbns[i - 1] + tmpval;
2241 nblocks = btodb(fs->fs_bsize);
2244 * Indirect blocks first.
2246 for (level = (NIADDR - 1); level >= 0; level--) {
2247 if ((bn = freeblks->fb_iblks[level]) == 0)
2249 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level,
2250 baselbns[level], &blocksreleased)) == 0)
2252 ffs_blkfree(&tip, bn, fs->fs_bsize);
2253 blocksreleased += nblocks;
2256 * All direct blocks or frags.
2258 for (i = (NDADDR - 1); i >= 0; i--) {
2259 if ((bn = freeblks->fb_dblks[i]) == 0)
2261 bsize = blksize(fs, &tip, i);
2262 ffs_blkfree(&tip, bn, bsize);
2263 blocksreleased += btodb(bsize);
2267 if (freeblks->fb_chkcnt != blocksreleased)
2268 kprintf("handle_workitem_freeblocks: block count\n");
2270 softdep_error("handle_workitem_freeblks", allerror);
2271 #endif /* DIAGNOSTIC */
2272 WORKITEM_FREE(freeblks, D_FREEBLKS);
2276 * Release blocks associated with the inode ip and stored in the indirect
2277 * block at doffset. If level is greater than SINGLE, the block is an
2278 * indirect block and recursive calls to indirtrunc must be used to
2279 * cleanse other indirect blocks.
2282 indir_trunc(struct inode *ip, off_t doffset, int level, ufs_lbn_t lbn,
2289 struct worklist *wk;
2290 struct indirdep *indirdep;
2291 int i, lbnadd, nblocks;
2292 int error, allerror = 0;
2296 for (i = level; i > 0; i--)
2297 lbnadd *= NINDIR(fs);
2299 * Get buffer of block pointers to be freed. This routine is not
2300 * called until the zero'ed inode has been written, so it is safe
2301 * to free blocks as they are encountered. Because the inode has
2302 * been zero'ed, calls to bmap on these blocks will fail. So, we
2303 * have to use the on-disk address and the block device for the
2304 * filesystem to look them up. If the file was deleted before its
2305 * indirect blocks were all written to disk, the routine that set
2306 * us up (deallocate_dependencies) will have arranged to leave
2307 * a complete copy of the indirect block in memory for our use.
2308 * Otherwise we have to read the blocks in from the disk.
2311 if ((bp = findblk(ip->i_devvp, doffset)) != NULL &&
2312 (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2314 * bp must be ir_savebp, which is held locked for our use.
2316 if (wk->wk_type != D_INDIRDEP ||
2317 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2318 (indirdep->ir_state & GOINGAWAY) == 0) {
2320 panic("indir_trunc: lost indirdep");
2322 WORKLIST_REMOVE(wk);
2323 WORKITEM_FREE(indirdep, D_INDIRDEP);
2324 if (LIST_FIRST(&bp->b_dep) != NULL) {
2326 panic("indir_trunc: dangling dep");
2331 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp);
2336 * Recursively free indirect blocks.
2338 bap = (ufs_daddr_t *)bp->b_data;
2339 nblocks = btodb(fs->fs_bsize);
2340 for (i = NINDIR(fs) - 1; i >= 0; i--) {
2341 if ((nb = bap[i]) == 0)
2344 if ((error = indir_trunc(ip, fsbtodoff(fs, nb),
2345 level - 1, lbn + (i * lbnadd), countp)) != 0)
2348 ffs_blkfree(ip, nb, fs->fs_bsize);
2351 bp->b_flags |= B_INVAL | B_NOCACHE;
2357 * Free an allocindir.
2358 * This routine must be called with splbio interrupts blocked.
2361 free_allocindir(struct allocindir *aip, struct inodedep *inodedep)
2363 struct freefrag *freefrag;
2366 if (lk.lkt_held == NOHOLDER)
2367 panic("free_allocindir: lock not held");
2369 if ((aip->ai_state & DEPCOMPLETE) == 0)
2370 LIST_REMOVE(aip, ai_deps);
2371 if (aip->ai_state & ONWORKLIST)
2372 WORKLIST_REMOVE(&aip->ai_list);
2373 LIST_REMOVE(aip, ai_next);
2374 if ((freefrag = aip->ai_freefrag) != NULL) {
2375 if (inodedep == NULL)
2376 add_to_worklist(&freefrag->ff_list);
2378 WORKLIST_INSERT(&inodedep->id_bufwait,
2379 &freefrag->ff_list);
2381 WORKITEM_FREE(aip, D_ALLOCINDIR);
2385 * Directory entry addition dependencies.
2387 * When adding a new directory entry, the inode (with its incremented link
2388 * count) must be written to disk before the directory entry's pointer to it.
2389 * Also, if the inode is newly allocated, the corresponding freemap must be
2390 * updated (on disk) before the directory entry's pointer. These requirements
2391 * are met via undo/redo on the directory entry's pointer, which consists
2392 * simply of the inode number.
2394 * As directory entries are added and deleted, the free space within a
2395 * directory block can become fragmented. The ufs filesystem will compact
2396 * a fragmented directory block to make space for a new entry. When this
2397 * occurs, the offsets of previously added entries change. Any "diradd"
2398 * dependency structures corresponding to these entries must be updated with
2403 * This routine is called after the in-memory inode's link
2404 * count has been incremented, but before the directory entry's
2405 * pointer to the inode has been set.
2408 * bp: buffer containing directory block
2409 * dp: inode for directory
2410 * diroffset: offset of new entry in directory
2411 * newinum: inode referenced by new directory entry
2412 * newdirbp: non-NULL => contents of new mkdir
2415 softdep_setup_directory_add(struct buf *bp, struct inode *dp, off_t diroffset,
2416 ino_t newinum, struct buf *newdirbp)
2418 int offset; /* offset of new entry within directory block */
2419 ufs_lbn_t lbn; /* block in directory containing new entry */
2422 struct pagedep *pagedep;
2423 struct inodedep *inodedep;
2424 struct mkdir *mkdir1, *mkdir2;
2427 * Whiteouts have no dependencies.
2429 if (newinum == WINO) {
2430 if (newdirbp != NULL)
2436 lbn = lblkno(fs, diroffset);
2437 offset = blkoff(fs, diroffset);
2438 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2440 bzero(dap, sizeof(struct diradd));
2441 dap->da_list.wk_type = D_DIRADD;
2442 dap->da_offset = offset;
2443 dap->da_newinum = newinum;
2444 dap->da_state = ATTACHED;
2445 if (newdirbp == NULL) {
2446 dap->da_state |= DEPCOMPLETE;
2449 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2450 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2452 mkdir1->md_list.wk_type = D_MKDIR;
2453 mkdir1->md_state = MKDIR_BODY;
2454 mkdir1->md_diradd = dap;
2455 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2457 mkdir2->md_list.wk_type = D_MKDIR;
2458 mkdir2->md_state = MKDIR_PARENT;
2459 mkdir2->md_diradd = dap;
2461 * Dependency on "." and ".." being written to disk.
2463 mkdir1->md_buf = newdirbp;
2465 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2466 WORKLIST_INSERT_BP(newdirbp, &mkdir1->md_list);
2470 * Dependency on link count increase for parent directory
2473 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2474 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2475 dap->da_state &= ~MKDIR_PARENT;
2476 WORKITEM_FREE(mkdir2, D_MKDIR);
2478 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2479 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2483 * Link into parent directory pagedep to await its being written.
2485 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2486 WORKLIST_INSERT_BP(bp, &pagedep->pd_list);
2487 dap->da_pagedep = pagedep;
2488 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2491 * Link into its inodedep. Put it on the id_bufwait list if the inode
2492 * is not yet written. If it is written, do the post-inode write
2493 * processing to put it on the id_pendinghd list.
2495 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2496 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2497 diradd_inode_written(dap, inodedep);
2499 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2504 * This procedure is called to change the offset of a directory
2505 * entry when compacting a directory block which must be owned
2506 * exclusively by the caller. Note that the actual entry movement
2507 * must be done in this procedure to ensure that no I/O completions
2508 * occur while the move is in progress.
2511 * dp: inode for directory
2512 * base: address of dp->i_offset
2513 * oldloc: address of old directory location
2514 * newloc: address of new directory location
2515 * entrysize: size of directory entry
2518 softdep_change_directoryentry_offset(struct inode *dp, caddr_t base,
2519 caddr_t oldloc, caddr_t newloc,
2522 int offset, oldoffset, newoffset;
2523 struct pagedep *pagedep;
2528 lbn = lblkno(dp->i_fs, dp->i_offset);
2529 offset = blkoff(dp->i_fs, dp->i_offset);
2530 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2532 oldoffset = offset + (oldloc - base);
2533 newoffset = offset + (newloc - base);
2535 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2536 if (dap->da_offset != oldoffset)
2538 dap->da_offset = newoffset;
2539 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2541 LIST_REMOVE(dap, da_pdlist);
2542 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2548 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2549 if (dap->da_offset == oldoffset) {
2550 dap->da_offset = newoffset;
2556 bcopy(oldloc, newloc, entrysize);
2561 * Free a diradd dependency structure. This routine must be called
2562 * with splbio interrupts blocked.
2565 free_diradd(struct diradd *dap)
2567 struct dirrem *dirrem;
2568 struct pagedep *pagedep;
2569 struct inodedep *inodedep;
2570 struct mkdir *mkdir, *nextmd;
2573 if (lk.lkt_held == NOHOLDER)
2574 panic("free_diradd: lock not held");
2576 WORKLIST_REMOVE(&dap->da_list);
2577 LIST_REMOVE(dap, da_pdlist);
2578 if ((dap->da_state & DIRCHG) == 0) {
2579 pagedep = dap->da_pagedep;
2581 dirrem = dap->da_previous;
2582 pagedep = dirrem->dm_pagedep;
2583 dirrem->dm_dirinum = pagedep->pd_ino;
2584 add_to_worklist(&dirrem->dm_list);
2586 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2588 (void) free_inodedep(inodedep);
2589 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2590 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2591 nextmd = LIST_NEXT(mkdir, md_mkdirs);
2592 if (mkdir->md_diradd != dap)
2594 dap->da_state &= ~mkdir->md_state;
2595 WORKLIST_REMOVE(&mkdir->md_list);
2596 LIST_REMOVE(mkdir, md_mkdirs);
2597 WORKITEM_FREE(mkdir, D_MKDIR);
2599 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2601 panic("free_diradd: unfound ref");
2604 WORKITEM_FREE(dap, D_DIRADD);
2608 * Directory entry removal dependencies.
2610 * When removing a directory entry, the entry's inode pointer must be
2611 * zero'ed on disk before the corresponding inode's link count is decremented
2612 * (possibly freeing the inode for re-use). This dependency is handled by
2613 * updating the directory entry but delaying the inode count reduction until
2614 * after the directory block has been written to disk. After this point, the
2615 * inode count can be decremented whenever it is convenient.
2619 * This routine should be called immediately after removing
2620 * a directory entry. The inode's link count should not be
2621 * decremented by the calling procedure -- the soft updates
2622 * code will do this task when it is safe.
2625 * bp: buffer containing directory block
2626 * dp: inode for the directory being modified
2627 * ip: inode for directory entry being removed
2628 * isrmdir: indicates if doing RMDIR
2631 softdep_setup_remove(struct buf *bp, struct inode *dp, struct inode *ip,
2634 struct dirrem *dirrem, *prevdirrem;
2637 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2639 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2642 * If the COMPLETE flag is clear, then there were no active
2643 * entries and we want to roll back to a zeroed entry until
2644 * the new inode is committed to disk. If the COMPLETE flag is
2645 * set then we have deleted an entry that never made it to
2646 * disk. If the entry we deleted resulted from a name change,
2647 * then the old name still resides on disk. We cannot delete
2648 * its inode (returned to us in prevdirrem) until the zeroed
2649 * directory entry gets to disk. The new inode has never been
2650 * referenced on the disk, so can be deleted immediately.
2652 if ((dirrem->dm_state & COMPLETE) == 0) {
2653 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2657 if (prevdirrem != NULL)
2658 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2659 prevdirrem, dm_next);
2660 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2662 handle_workitem_remove(dirrem);
2667 * Allocate a new dirrem if appropriate and return it along with
2668 * its associated pagedep. Called without a lock, returns with lock.
2670 static long num_dirrem; /* number of dirrem allocated */
2674 * bp: buffer containing directory block
2675 * dp: inode for the directory being modified
2676 * ip: inode for directory entry being removed
2677 * isrmdir: indicates if doing RMDIR
2678 * prevdirremp: previously referenced inode, if any
2680 static struct dirrem *
2681 newdirrem(struct buf *bp, struct inode *dp, struct inode *ip,
2682 int isrmdir, struct dirrem **prevdirremp)
2687 struct dirrem *dirrem;
2688 struct pagedep *pagedep;
2691 * Whiteouts have no deletion dependencies.
2694 panic("newdirrem: whiteout");
2696 * If we are over our limit, try to improve the situation.
2697 * Limiting the number of dirrem structures will also limit
2698 * the number of freefile and freeblks structures.
2700 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2701 (void) request_cleanup(FLUSH_REMOVE, 0);
2703 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
2704 M_DIRREM, M_SOFTDEP_FLAGS);
2705 bzero(dirrem, sizeof(struct dirrem));
2706 dirrem->dm_list.wk_type = D_DIRREM;
2707 dirrem->dm_state = isrmdir ? RMDIR : 0;
2708 dirrem->dm_mnt = ITOV(ip)->v_mount;
2709 dirrem->dm_oldinum = ip->i_number;
2710 *prevdirremp = NULL;
2713 lbn = lblkno(dp->i_fs, dp->i_offset);
2714 offset = blkoff(dp->i_fs, dp->i_offset);
2715 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2716 WORKLIST_INSERT_BP(bp, &pagedep->pd_list);
2717 dirrem->dm_pagedep = pagedep;
2719 * Check for a diradd dependency for the same directory entry.
2720 * If present, then both dependencies become obsolete and can
2721 * be de-allocated. Check for an entry on both the pd_dirraddhd
2722 * list and the pd_pendinghd list.
2725 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
2726 if (dap->da_offset == offset)
2730 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
2731 if (dap->da_offset == offset)
2737 * Must be ATTACHED at this point.
2739 if ((dap->da_state & ATTACHED) == 0) {
2741 panic("newdirrem: not ATTACHED");
2743 if (dap->da_newinum != ip->i_number) {
2745 panic("newdirrem: inum %"PRId64" should be %"PRId64,
2746 ip->i_number, dap->da_newinum);
2749 * If we are deleting a changed name that never made it to disk,
2750 * then return the dirrem describing the previous inode (which
2751 * represents the inode currently referenced from this entry on disk).
2753 if ((dap->da_state & DIRCHG) != 0) {
2754 *prevdirremp = dap->da_previous;
2755 dap->da_state &= ~DIRCHG;
2756 dap->da_pagedep = pagedep;
2759 * We are deleting an entry that never made it to disk.
2760 * Mark it COMPLETE so we can delete its inode immediately.
2762 dirrem->dm_state |= COMPLETE;
2768 * Directory entry change dependencies.
2770 * Changing an existing directory entry requires that an add operation
2771 * be completed first followed by a deletion. The semantics for the addition
2772 * are identical to the description of adding a new entry above except
2773 * that the rollback is to the old inode number rather than zero. Once
2774 * the addition dependency is completed, the removal is done as described
2775 * in the removal routine above.
2779 * This routine should be called immediately after changing
2780 * a directory entry. The inode's link count should not be
2781 * decremented by the calling procedure -- the soft updates
2782 * code will perform this task when it is safe.
2785 * bp: buffer containing directory block
2786 * dp: inode for the directory being modified
2787 * ip: inode for directory entry being removed
2788 * newinum: new inode number for changed entry
2789 * isrmdir: indicates if doing RMDIR
2792 softdep_setup_directory_change(struct buf *bp, struct inode *dp,
2793 struct inode *ip, ino_t newinum,
2797 struct diradd *dap = NULL;
2798 struct dirrem *dirrem, *prevdirrem;
2799 struct pagedep *pagedep;
2800 struct inodedep *inodedep;
2802 offset = blkoff(dp->i_fs, dp->i_offset);
2805 * Whiteouts do not need diradd dependencies.
2807 if (newinum != WINO) {
2808 MALLOC(dap, struct diradd *, sizeof(struct diradd),
2809 M_DIRADD, M_SOFTDEP_FLAGS);
2810 bzero(dap, sizeof(struct diradd));
2811 dap->da_list.wk_type = D_DIRADD;
2812 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2813 dap->da_offset = offset;
2814 dap->da_newinum = newinum;
2818 * Allocate a new dirrem and ACQUIRE_LOCK.
2820 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2821 pagedep = dirrem->dm_pagedep;
2823 * The possible values for isrmdir:
2824 * 0 - non-directory file rename
2825 * 1 - directory rename within same directory
2826 * inum - directory rename to new directory of given inode number
2827 * When renaming to a new directory, we are both deleting and
2828 * creating a new directory entry, so the link count on the new
2829 * directory should not change. Thus we do not need the followup
2830 * dirrem which is usually done in handle_workitem_remove. We set
2831 * the DIRCHG flag to tell handle_workitem_remove to skip the
2835 dirrem->dm_state |= DIRCHG;
2838 * Whiteouts have no additional dependencies,
2839 * so just put the dirrem on the correct list.
2841 if (newinum == WINO) {
2842 if ((dirrem->dm_state & COMPLETE) == 0) {
2843 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2846 dirrem->dm_dirinum = pagedep->pd_ino;
2847 add_to_worklist(&dirrem->dm_list);
2854 * If the COMPLETE flag is clear, then there were no active
2855 * entries and we want to roll back to the previous inode until
2856 * the new inode is committed to disk. If the COMPLETE flag is
2857 * set, then we have deleted an entry that never made it to disk.
2858 * If the entry we deleted resulted from a name change, then the old
2859 * inode reference still resides on disk. Any rollback that we do
2860 * needs to be to that old inode (returned to us in prevdirrem). If
2861 * the entry we deleted resulted from a create, then there is
2862 * no entry on the disk, so we want to roll back to zero rather
2863 * than the uncommitted inode. In either of the COMPLETE cases we
2864 * want to immediately free the unwritten and unreferenced inode.
2866 if ((dirrem->dm_state & COMPLETE) == 0) {
2867 dap->da_previous = dirrem;
2869 if (prevdirrem != NULL) {
2870 dap->da_previous = prevdirrem;
2872 dap->da_state &= ~DIRCHG;
2873 dap->da_pagedep = pagedep;
2875 dirrem->dm_dirinum = pagedep->pd_ino;
2876 add_to_worklist(&dirrem->dm_list);
2879 * Link into its inodedep. Put it on the id_bufwait list if the inode
2880 * is not yet written. If it is written, do the post-inode write
2881 * processing to put it on the id_pendinghd list.
2883 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2884 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2885 dap->da_state |= COMPLETE;
2886 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2887 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2889 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2891 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2897 * Called whenever the link count on an inode is changed.
2898 * It creates an inode dependency so that the new reference(s)
2899 * to the inode cannot be committed to disk until the updated
2900 * inode has been written.
2903 * ip: the inode with the increased link count
2906 softdep_change_linkcnt(struct inode *ip)
2908 struct inodedep *inodedep;
2911 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2912 if (ip->i_nlink < ip->i_effnlink) {
2914 panic("softdep_change_linkcnt: bad delta");
2916 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2921 * This workitem decrements the inode's link count.
2922 * If the link count reaches zero, the file is removed.
2925 handle_workitem_remove(struct dirrem *dirrem)
2927 struct inodedep *inodedep;
2933 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) {
2934 softdep_error("handle_workitem_remove: vget", error);
2939 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
2941 panic("handle_workitem_remove: lost inodedep");
2944 * Normal file deletion.
2946 if ((dirrem->dm_state & RMDIR) == 0) {
2948 ip->i_flag |= IN_CHANGE;
2949 if (ip->i_nlink < ip->i_effnlink) {
2951 panic("handle_workitem_remove: bad file delta");
2953 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2957 WORKITEM_FREE(dirrem, D_DIRREM);
2961 * Directory deletion. Decrement reference count for both the
2962 * just deleted parent directory entry and the reference for ".".
2963 * Next truncate the directory to length zero. When the
2964 * truncation completes, arrange to have the reference count on
2965 * the parent decremented to account for the loss of "..".
2968 ip->i_flag |= IN_CHANGE;
2969 if (ip->i_nlink < ip->i_effnlink) {
2971 panic("handle_workitem_remove: bad dir delta");
2973 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2975 if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0)
2976 softdep_error("handle_workitem_remove: truncate", error);
2978 * Rename a directory to a new parent. Since, we are both deleting
2979 * and creating a new directory entry, the link count on the new
2980 * directory should not change. Thus we skip the followup dirrem.
2982 if (dirrem->dm_state & DIRCHG) {
2985 WORKITEM_FREE(dirrem, D_DIRREM);
2989 * If the inodedep does not exist, then the zero'ed inode has
2990 * been written to disk. If the allocated inode has never been
2991 * written to disk, then the on-disk inode is zero'ed. In either
2992 * case we can remove the file immediately.
2995 dirrem->dm_state = 0;
2996 oldinum = dirrem->dm_oldinum;
2997 dirrem->dm_oldinum = dirrem->dm_dirinum;
2998 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
2999 check_inode_unwritten(inodedep)) {
3002 handle_workitem_remove(dirrem);
3005 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3007 ip->i_flag |= IN_CHANGE;
3013 * Inode de-allocation dependencies.
3015 * When an inode's link count is reduced to zero, it can be de-allocated. We
3016 * found it convenient to postpone de-allocation until after the inode is
3017 * written to disk with its new link count (zero). At this point, all of the
3018 * on-disk inode's block pointers are nullified and, with careful dependency
3019 * list ordering, all dependencies related to the inode will be satisfied and
3020 * the corresponding dependency structures de-allocated. So, if/when the
3021 * inode is reused, there will be no mixing of old dependencies with new
3022 * ones. This artificial dependency is set up by the block de-allocation
3023 * procedure above (softdep_setup_freeblocks) and completed by the
3024 * following procedure.
3027 handle_workitem_freefile(struct freefile *freefile)
3031 struct inodedep *idp;
3036 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp);
3039 panic("handle_workitem_freefile: inodedep survived");
3041 tip.i_devvp = freefile->fx_devvp;
3042 tip.i_dev = freefile->fx_devvp->v_rdev;
3043 tip.i_fs = freefile->fx_fs;
3045 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
3046 softdep_error("handle_workitem_freefile", error);
3047 WORKITEM_FREE(freefile, D_FREEFILE);
3051 * Helper function which unlinks marker element from work list and returns
3052 * the next element on the list.
3054 static __inline struct worklist *
3055 markernext(struct worklist *marker)
3057 struct worklist *next;
3059 next = LIST_NEXT(marker, wk_list);
3060 LIST_REMOVE(marker, wk_list);
3067 * The dependency structures constructed above are most actively used when file
3068 * system blocks are written to disk. No constraints are placed on when a
3069 * block can be written, but unsatisfied update dependencies are made safe by
3070 * modifying (or replacing) the source memory for the duration of the disk
3071 * write. When the disk write completes, the memory block is again brought
3074 * In-core inode structure reclamation.
3076 * Because there are a finite number of "in-core" inode structures, they are
3077 * reused regularly. By transferring all inode-related dependencies to the
3078 * in-memory inode block and indexing them separately (via "inodedep"s), we
3079 * can allow "in-core" inode structures to be reused at any time and avoid
3080 * any increase in contention.
3082 * Called just before entering the device driver to initiate a new disk I/O.
3083 * The buffer must be locked, thus, no I/O completion operations can occur
3084 * while we are manipulating its associated dependencies.
3087 * bp: structure describing disk write to occur
3090 softdep_disk_io_initiation(struct buf *bp)
3092 struct worklist *wk;
3093 struct worklist marker;
3094 struct indirdep *indirdep;
3097 * We only care about write operations. There should never
3098 * be dependencies for reads.
3100 if (bp->b_cmd == BUF_CMD_READ)
3101 panic("softdep_disk_io_initiation: read");
3103 marker.wk_type = D_LAST + 1; /* Not a normal workitem */
3106 * Do any necessary pre-I/O processing.
3108 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) {
3109 LIST_INSERT_AFTER(wk, &marker, wk_list);
3111 switch (wk->wk_type) {
3114 initiate_write_filepage(WK_PAGEDEP(wk), bp);
3118 initiate_write_inodeblock(WK_INODEDEP(wk), bp);
3122 indirdep = WK_INDIRDEP(wk);
3123 if (indirdep->ir_state & GOINGAWAY)
3124 panic("disk_io_initiation: indirdep gone");
3126 * If there are no remaining dependencies, this
3127 * will be writing the real pointers, so the
3128 * dependency can be freed.
3130 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
3131 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
3132 brelse(indirdep->ir_savebp);
3133 /* inline expand WORKLIST_REMOVE(wk); */
3134 wk->wk_state &= ~ONWORKLIST;
3135 LIST_REMOVE(wk, wk_list);
3136 WORKITEM_FREE(indirdep, D_INDIRDEP);
3140 * Replace up-to-date version with safe version.
3142 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3143 M_INDIRDEP, M_SOFTDEP_FLAGS);
3145 indirdep->ir_state &= ~ATTACHED;
3146 indirdep->ir_state |= UNDONE;
3147 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3148 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3160 panic("handle_disk_io_initiation: Unexpected type %s",
3161 TYPENAME(wk->wk_type));
3168 * Called from within the procedure above to deal with unsatisfied
3169 * allocation dependencies in a directory. The buffer must be locked,
3170 * thus, no I/O completion operations can occur while we are
3171 * manipulating its associated dependencies.
3174 initiate_write_filepage(struct pagedep *pagedep, struct buf *bp)
3180 if (pagedep->pd_state & IOSTARTED) {
3182 * This can only happen if there is a driver that does not
3183 * understand chaining. Here biodone will reissue the call
3184 * to strategy for the incomplete buffers.
3186 kprintf("initiate_write_filepage: already started\n");
3189 pagedep->pd_state |= IOSTARTED;
3191 for (i = 0; i < DAHASHSZ; i++) {
3192 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3193 ep = (struct direct *)
3194 ((char *)bp->b_data + dap->da_offset);
3195 if (ep->d_ino != dap->da_newinum) {
3197 panic("%s: dir inum %d != new %"PRId64,
3198 "initiate_write_filepage",
3199 ep->d_ino, dap->da_newinum);
3201 if (dap->da_state & DIRCHG)
3202 ep->d_ino = dap->da_previous->dm_oldinum;
3205 dap->da_state &= ~ATTACHED;
3206 dap->da_state |= UNDONE;
3213 * Called from within the procedure above to deal with unsatisfied
3214 * allocation dependencies in an inodeblock. The buffer must be
3215 * locked, thus, no I/O completion operations can occur while we
3216 * are manipulating its associated dependencies.
3219 * bp: The inode block
3222 initiate_write_inodeblock(struct inodedep *inodedep, struct buf *bp)
3224 struct allocdirect *adp, *lastadp;
3225 struct ufs1_dinode *dp;
3226 struct ufs1_dinode *sip;
3228 ufs_lbn_t prevlbn = 0;
3231 if (inodedep->id_state & IOSTARTED)
3232 panic("initiate_write_inodeblock: already started");
3233 inodedep->id_state |= IOSTARTED;
3234 fs = inodedep->id_fs;
3235 dp = (struct ufs1_dinode *)bp->b_data +
3236 ino_to_fsbo(fs, inodedep->id_ino);
3238 * If the bitmap is not yet written, then the allocated
3239 * inode cannot be written to disk.
3241 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3242 if (inodedep->id_savedino != NULL)
3243 panic("initiate_write_inodeblock: already doing I/O");
3244 MALLOC(sip, struct ufs1_dinode *,
3245 sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3246 inodedep->id_savedino = sip;
3247 *inodedep->id_savedino = *dp;
3248 bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
3249 dp->di_gen = inodedep->id_savedino->di_gen;
3253 * If no dependencies, then there is nothing to roll back.
3255 inodedep->id_savedsize = dp->di_size;
3256 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3259 * Set the dependencies to busy.
3262 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3263 adp = TAILQ_NEXT(adp, ad_next)) {
3265 if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3267 panic("softdep_write_inodeblock: lbn order");
3269 prevlbn = adp->ad_lbn;
3270 if (adp->ad_lbn < NDADDR &&
3271 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3273 panic("%s: direct pointer #%ld mismatch %d != %d",
3274 "softdep_write_inodeblock", adp->ad_lbn,
3275 dp->di_db[adp->ad_lbn], adp->ad_newblkno);
3277 if (adp->ad_lbn >= NDADDR &&
3278 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3280 panic("%s: indirect pointer #%ld mismatch %d != %d",
3281 "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
3282 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
3284 deplist |= 1 << adp->ad_lbn;
3285 if ((adp->ad_state & ATTACHED) == 0) {
3287 panic("softdep_write_inodeblock: Unknown state 0x%x",
3290 #endif /* DIAGNOSTIC */
3291 adp->ad_state &= ~ATTACHED;
3292 adp->ad_state |= UNDONE;
3295 * The on-disk inode cannot claim to be any larger than the last
3296 * fragment that has been written. Otherwise, the on-disk inode
3297 * might have fragments that were not the last block in the file
3298 * which would corrupt the filesystem.
3300 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3301 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3302 if (adp->ad_lbn >= NDADDR)
3304 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3305 /* keep going until hitting a rollback to a frag */
3306 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3308 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3309 for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3311 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3313 panic("softdep_write_inodeblock: lost dep1");
3315 #endif /* DIAGNOSTIC */
3318 for (i = 0; i < NIADDR; i++) {
3320 if (dp->di_ib[i] != 0 &&
3321 (deplist & ((1 << NDADDR) << i)) == 0) {
3323 panic("softdep_write_inodeblock: lost dep2");
3325 #endif /* DIAGNOSTIC */
3332 * If we have zero'ed out the last allocated block of the file,
3333 * roll back the size to the last currently allocated block.
3334 * We know that this last allocated block is a full-sized as
3335 * we already checked for fragments in the loop above.
3337 if (lastadp != NULL &&
3338 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3339 for (i = lastadp->ad_lbn; i >= 0; i--)
3340 if (dp->di_db[i] != 0)
3342 dp->di_size = (i + 1) * fs->fs_bsize;
3345 * The only dependencies are for indirect blocks.
3347 * The file size for indirect block additions is not guaranteed.
3348 * Such a guarantee would be non-trivial to achieve. The conventional
3349 * synchronous write implementation also does not make this guarantee.
3350 * Fsck should catch and fix discrepancies. Arguably, the file size
3351 * can be over-estimated without destroying integrity when the file
3352 * moves into the indirect blocks (i.e., is large). If we want to
3353 * postpone fsck, we are stuck with this argument.
3355 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3356 dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3361 * This routine is called during the completion interrupt
3362 * service routine for a disk write (from the procedure called
3363 * by the device driver to inform the filesystem caches of
3364 * a request completion). It should be called early in this
3365 * procedure, before the block is made available to other
3366 * processes or other routines are called.
3369 * bp: describes the completed disk write
3372 softdep_disk_write_complete(struct buf *bp)
3374 struct worklist *wk;
3375 struct workhead reattach;
3376 struct newblk *newblk;
3377 struct allocindir *aip;
3378 struct allocdirect *adp;
3379 struct indirdep *indirdep;
3380 struct inodedep *inodedep;
3381 struct bmsafemap *bmsafemap;
3384 if (lk.lkt_held != NOHOLDER)
3385 panic("softdep_disk_write_complete: lock is held");
3386 lk.lkt_held = SPECIAL_FLAG;
3388 LIST_INIT(&reattach);
3389 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3390 WORKLIST_REMOVE(wk);
3391 switch (wk->wk_type) {
3394 if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3395 WORKLIST_INSERT(&reattach, wk);
3399 if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3400 WORKLIST_INSERT(&reattach, wk);
3404 bmsafemap = WK_BMSAFEMAP(wk);
3405 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3406 newblk->nb_state |= DEPCOMPLETE;
3407 newblk->nb_bmsafemap = NULL;
3408 LIST_REMOVE(newblk, nb_deps);
3411 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3412 adp->ad_state |= DEPCOMPLETE;
3414 LIST_REMOVE(adp, ad_deps);
3415 handle_allocdirect_partdone(adp);
3418 LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3419 aip->ai_state |= DEPCOMPLETE;
3421 LIST_REMOVE(aip, ai_deps);
3422 handle_allocindir_partdone(aip);
3425 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3426 inodedep->id_state |= DEPCOMPLETE;
3427 LIST_REMOVE(inodedep, id_deps);
3428 inodedep->id_buf = NULL;
3430 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3434 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3438 adp = WK_ALLOCDIRECT(wk);
3439 adp->ad_state |= COMPLETE;
3440 handle_allocdirect_partdone(adp);
3444 aip = WK_ALLOCINDIR(wk);
3445 aip->ai_state |= COMPLETE;
3446 handle_allocindir_partdone(aip);
3450 indirdep = WK_INDIRDEP(wk);
3451 if (indirdep->ir_state & GOINGAWAY) {
3452 lk.lkt_held = NOHOLDER;
3453 panic("disk_write_complete: indirdep gone");
3455 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3456 FREE(indirdep->ir_saveddata, M_INDIRDEP);
3457 indirdep->ir_saveddata = 0;
3458 indirdep->ir_state &= ~UNDONE;
3459 indirdep->ir_state |= ATTACHED;
3460 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3461 handle_allocindir_partdone(aip);
3462 if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
3463 lk.lkt_held = NOHOLDER;
3464 panic("disk_write_complete: not gone");
3467 WORKLIST_INSERT(&reattach, wk);
3468 if ((bp->b_flags & B_DELWRI) == 0)
3469 stat_indir_blk_ptrs++;
3474 lk.lkt_held = NOHOLDER;
3475 panic("handle_disk_write_complete: Unknown type %s",
3476 TYPENAME(wk->wk_type));
3481 * Reattach any requests that must be redone.
3483 while ((wk = LIST_FIRST(&reattach)) != NULL) {
3484 WORKLIST_REMOVE(wk);
3485 WORKLIST_INSERT_BP(bp, wk);
3488 if (lk.lkt_held != SPECIAL_FLAG)
3489 panic("softdep_disk_write_complete: lock lost");
3490 lk.lkt_held = NOHOLDER;
3495 * Called from within softdep_disk_write_complete above. Note that
3496 * this routine is always called from interrupt level with further
3497 * splbio interrupts blocked.
3500 * adp: the completed allocdirect
3503 handle_allocdirect_partdone(struct allocdirect *adp)
3505 struct allocdirect *listadp;
3506 struct inodedep *inodedep;
3509 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3511 if (adp->ad_buf != NULL) {
3512 lk.lkt_held = NOHOLDER;
3513 panic("handle_allocdirect_partdone: dangling dep");
3516 * The on-disk inode cannot claim to be any larger than the last
3517 * fragment that has been written. Otherwise, the on-disk inode
3518 * might have fragments that were not the last block in the file
3519 * which would corrupt the filesystem. Thus, we cannot free any
3520 * allocdirects after one whose ad_oldblkno claims a fragment as
3521 * these blocks must be rolled back to zero before writing the inode.
3522 * We check the currently active set of allocdirects in id_inoupdt.
3524 inodedep = adp->ad_inodedep;
3525 bsize = inodedep->id_fs->fs_bsize;
3526 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) {
3527 /* found our block */
3530 /* continue if ad_oldlbn is not a fragment */
3531 if (listadp->ad_oldsize == 0 ||
3532 listadp->ad_oldsize == bsize)
3534 /* hit a fragment */
3538 * If we have reached the end of the current list without
3539 * finding the just finished dependency, then it must be
3540 * on the future dependency list. Future dependencies cannot
3541 * be freed until they are moved to the current list.
3543 if (listadp == NULL) {
3545 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next)
3546 /* found our block */
3549 if (listadp == NULL) {
3550 lk.lkt_held = NOHOLDER;
3551 panic("handle_allocdirect_partdone: lost dep");
3557 * If we have found the just finished dependency, then free
3558 * it along with anything that follows it that is complete.
3560 for (; adp; adp = listadp) {
3561 listadp = TAILQ_NEXT(adp, ad_next);
3562 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3564 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
3569 * Called from within softdep_disk_write_complete above. Note that
3570 * this routine is always called from interrupt level with further
3571 * splbio interrupts blocked.
3574 * aip: the completed allocindir
3577 handle_allocindir_partdone(struct allocindir *aip)
3579 struct indirdep *indirdep;
3581 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
3583 if (aip->ai_buf != NULL) {
3584 lk.lkt_held = NOHOLDER;
3585 panic("handle_allocindir_partdone: dangling dependency");
3587 indirdep = aip->ai_indirdep;
3588 if (indirdep->ir_state & UNDONE) {
3589 LIST_REMOVE(aip, ai_next);
3590 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
3593 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
3595 LIST_REMOVE(aip, ai_next);
3596 if (aip->ai_freefrag != NULL)
3597 add_to_worklist(&aip->ai_freefrag->ff_list);
3598 WORKITEM_FREE(aip, D_ALLOCINDIR);
3602 * Called from within softdep_disk_write_complete above to restore
3603 * in-memory inode block contents to their most up-to-date state. Note
3604 * that this routine is always called from interrupt level with further
3605 * splbio interrupts blocked.
3608 * bp: buffer containing the inode block
3611 handle_written_inodeblock(struct inodedep *inodedep, struct buf *bp)
3613 struct worklist *wk, *filefree;
3614 struct allocdirect *adp, *nextadp;
3615 struct ufs1_dinode *dp;
3618 if ((inodedep->id_state & IOSTARTED) == 0) {
3619 lk.lkt_held = NOHOLDER;
3620 panic("handle_written_inodeblock: not started");
3622 inodedep->id_state &= ~IOSTARTED;
3623 dp = (struct ufs1_dinode *)bp->b_data +
3624 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
3626 * If we had to rollback the inode allocation because of
3627 * bitmaps being incomplete, then simply restore it.
3628 * Keep the block dirty so that it will not be reclaimed until
3629 * all associated dependencies have been cleared and the
3630 * corresponding updates written to disk.
3632 if (inodedep->id_savedino != NULL) {
3633 *dp = *inodedep->id_savedino;
3634 FREE(inodedep->id_savedino, M_INODEDEP);
3635 inodedep->id_savedino = NULL;
3636 if ((bp->b_flags & B_DELWRI) == 0)
3637 stat_inode_bitmap++;
3641 inodedep->id_state |= COMPLETE;
3643 * Roll forward anything that had to be rolled back before
3644 * the inode could be updated.
3647 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
3648 nextadp = TAILQ_NEXT(adp, ad_next);
3649 if (adp->ad_state & ATTACHED) {
3650 lk.lkt_held = NOHOLDER;
3651 panic("handle_written_inodeblock: new entry");
3653 if (adp->ad_lbn < NDADDR) {
3654 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) {
3655 lk.lkt_held = NOHOLDER;
3656 panic("%s: %s #%ld mismatch %d != %d",
3657 "handle_written_inodeblock",
3658 "direct pointer", adp->ad_lbn,
3659 dp->di_db[adp->ad_lbn], adp->ad_oldblkno);
3661 dp->di_db[adp->ad_lbn] = adp->ad_newblkno;
3663 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) {
3664 lk.lkt_held = NOHOLDER;
3665 panic("%s: %s #%ld allocated as %d",
3666 "handle_written_inodeblock",
3667 "indirect pointer", adp->ad_lbn - NDADDR,
3668 dp->di_ib[adp->ad_lbn - NDADDR]);
3670 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno;
3672 adp->ad_state &= ~UNDONE;
3673 adp->ad_state |= ATTACHED;
3676 if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
3677 stat_direct_blk_ptrs++;
3679 * Reset the file size to its most up-to-date value.
3681 if (inodedep->id_savedsize == -1) {
3682 lk.lkt_held = NOHOLDER;
3683 panic("handle_written_inodeblock: bad size");
3685 if (dp->di_size != inodedep->id_savedsize) {
3686 dp->di_size = inodedep->id_savedsize;
3689 inodedep->id_savedsize = -1;
3691 * If there were any rollbacks in the inode block, then it must be
3692 * marked dirty so that its will eventually get written back in
3698 * Process any allocdirects that completed during the update.
3700 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
3701 handle_allocdirect_partdone(adp);
3703 * Process deallocations that were held pending until the
3704 * inode had been written to disk. Freeing of the inode
3705 * is delayed until after all blocks have been freed to
3706 * avoid creation of new <vfsid, inum, lbn> triples
3707 * before the old ones have been deleted.
3710 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
3711 WORKLIST_REMOVE(wk);
3712 switch (wk->wk_type) {
3716 * We defer adding filefree to the worklist until
3717 * all other additions have been made to ensure
3718 * that it will be done after all the old blocks
3721 if (filefree != NULL) {
3722 lk.lkt_held = NOHOLDER;
3723 panic("handle_written_inodeblock: filefree");
3729 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
3733 diradd_inode_written(WK_DIRADD(wk), inodedep);
3737 wk->wk_state |= COMPLETE;
3738 if ((wk->wk_state & ALLCOMPLETE) != ALLCOMPLETE)
3740 /* -- fall through -- */
3743 add_to_worklist(wk);
3747 lk.lkt_held = NOHOLDER;
3748 panic("handle_written_inodeblock: Unknown type %s",
3749 TYPENAME(wk->wk_type));
3753 if (filefree != NULL) {
3754 if (free_inodedep(inodedep) == 0) {
3755 lk.lkt_held = NOHOLDER;
3756 panic("handle_written_inodeblock: live inodedep");
3758 add_to_worklist(filefree);
3763 * If no outstanding dependencies, free it.
3765 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0)
3767 return (hadchanges);
3771 * Process a diradd entry after its dependent inode has been written.
3772 * This routine must be called with splbio interrupts blocked.
3775 diradd_inode_written(struct diradd *dap, struct inodedep *inodedep)
3777 struct pagedep *pagedep;
3779 dap->da_state |= COMPLETE;
3780 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3781 if (dap->da_state & DIRCHG)
3782 pagedep = dap->da_previous->dm_pagedep;
3784 pagedep = dap->da_pagedep;
3785 LIST_REMOVE(dap, da_pdlist);
3786 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3788 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3792 * Handle the completion of a mkdir dependency.
3795 handle_written_mkdir(struct mkdir *mkdir, int type)
3798 struct pagedep *pagedep;
3800 if (mkdir->md_state != type) {
3801 lk.lkt_held = NOHOLDER;
3802 panic("handle_written_mkdir: bad type");
3804 dap = mkdir->md_diradd;
3805 dap->da_state &= ~type;
3806 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
3807 dap->da_state |= DEPCOMPLETE;
3808 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3809 if (dap->da_state & DIRCHG)
3810 pagedep = dap->da_previous->dm_pagedep;
3812 pagedep = dap->da_pagedep;
3813 LIST_REMOVE(dap, da_pdlist);
3814 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3816 LIST_REMOVE(mkdir, md_mkdirs);
3817 WORKITEM_FREE(mkdir, D_MKDIR);
3821 * Called from within softdep_disk_write_complete above.
3822 * A write operation was just completed. Removed inodes can
3823 * now be freed and associated block pointers may be committed.
3824 * Note that this routine is always called from interrupt level
3825 * with further splbio interrupts blocked.
3828 * bp: buffer containing the written page
3831 handle_written_filepage(struct pagedep *pagedep, struct buf *bp)
3833 struct dirrem *dirrem;
3834 struct diradd *dap, *nextdap;
3838 if ((pagedep->pd_state & IOSTARTED) == 0) {
3839 lk.lkt_held = NOHOLDER;
3840 panic("handle_written_filepage: not started");
3842 pagedep->pd_state &= ~IOSTARTED;
3844 * Process any directory removals that have been committed.
3846 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
3847 LIST_REMOVE(dirrem, dm_next);
3848 dirrem->dm_dirinum = pagedep->pd_ino;
3849 add_to_worklist(&dirrem->dm_list);
3852 * Free any directory additions that have been committed.
3854 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
3857 * Uncommitted directory entries must be restored.
3859 for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
3860 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
3862 nextdap = LIST_NEXT(dap, da_pdlist);
3863 if (dap->da_state & ATTACHED) {
3864 lk.lkt_held = NOHOLDER;
3865 panic("handle_written_filepage: attached");
3867 ep = (struct direct *)
3868 ((char *)bp->b_data + dap->da_offset);
3869 ep->d_ino = dap->da_newinum;
3870 dap->da_state &= ~UNDONE;
3871 dap->da_state |= ATTACHED;
3874 * If the inode referenced by the directory has
3875 * been written out, then the dependency can be
3876 * moved to the pending list.
3878 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3879 LIST_REMOVE(dap, da_pdlist);
3880 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
3886 * If there were any rollbacks in the directory, then it must be
3887 * marked dirty so that its will eventually get written back in
3891 if ((bp->b_flags & B_DELWRI) == 0)
3896 * If no dependencies remain, the pagedep will be freed.
3897 * Otherwise it will remain to update the page before it
3898 * is written back to disk.
3900 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) {
3901 for (i = 0; i < DAHASHSZ; i++)
3902 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
3904 if (i == DAHASHSZ) {
3905 LIST_REMOVE(pagedep, pd_hash);
3906 WORKITEM_FREE(pagedep, D_PAGEDEP);
3914 * Writing back in-core inode structures.
3916 * The filesystem only accesses an inode's contents when it occupies an
3917 * "in-core" inode structure. These "in-core" structures are separate from
3918 * the page frames used to cache inode blocks. Only the latter are
3919 * transferred to/from the disk. So, when the updated contents of the
3920 * "in-core" inode structure are copied to the corresponding in-memory inode
3921 * block, the dependencies are also transferred. The following procedure is
3922 * called when copying a dirty "in-core" inode to a cached inode block.
3926 * Called when an inode is loaded from disk. If the effective link count
3927 * differed from the actual link count when it was last flushed, then we
3928 * need to ensure that the correct effective link count is put back.
3931 * ip: the "in_core" copy of the inode
3934 softdep_load_inodeblock(struct inode *ip)
3936 struct inodedep *inodedep;
3939 * Check for alternate nlink count.
3941 ip->i_effnlink = ip->i_nlink;
3943 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3947 ip->i_effnlink -= inodedep->id_nlinkdelta;
3952 * This routine is called just before the "in-core" inode
3953 * information is to be copied to the in-memory inode block.
3954 * Recall that an inode block contains several inodes. If
3955 * the force flag is set, then the dependencies will be
3956 * cleared so that the update can always be made. Note that
3957 * the buffer is locked when this routine is called, so we
3958 * will never be in the middle of writing the inode block
3962 * ip: the "in_core" copy of the inode
3963 * bp: the buffer containing the inode block
3964 * waitfor: nonzero => update must be allowed
3967 softdep_update_inodeblock(struct inode *ip, struct buf *bp,
3970 struct inodedep *inodedep;
3971 struct worklist *wk;
3975 * If the effective link count is not equal to the actual link
3976 * count, then we must track the difference in an inodedep while
3977 * the inode is (potentially) tossed out of the cache. Otherwise,
3978 * if there is no existing inodedep, then there are no dependencies
3982 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3984 if (ip->i_effnlink != ip->i_nlink)
3985 panic("softdep_update_inodeblock: bad link count");
3988 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
3990 panic("softdep_update_inodeblock: bad delta");
3993 * Changes have been initiated. Anything depending on these
3994 * changes cannot occur until this inode has been written.
3996 inodedep->id_state &= ~COMPLETE;
3997 if ((inodedep->id_state & ONWORKLIST) == 0)
3998 WORKLIST_INSERT_BP(bp, &inodedep->id_list);
4000 * Any new dependencies associated with the incore inode must
4001 * now be moved to the list associated with the buffer holding
4002 * the in-memory copy of the inode. Once merged process any
4003 * allocdirects that are completed by the merger.
4005 merge_inode_lists(inodedep);
4006 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
4007 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
4009 * Now that the inode has been pushed into the buffer, the
4010 * operations dependent on the inode being written to disk
4011 * can be moved to the id_bufwait so that they will be
4012 * processed when the buffer I/O completes.
4014 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
4015 WORKLIST_REMOVE(wk);
4016 WORKLIST_INSERT(&inodedep->id_bufwait, wk);
4019 * Newly allocated inodes cannot be written until the bitmap
4020 * that allocates them have been written (indicated by
4021 * DEPCOMPLETE being set in id_state). If we are doing a
4022 * forced sync (e.g., an fsync on a file), we force the bitmap
4023 * to be written so that the update can be done.
4025 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
4029 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4032 (error = bwrite(inodedep->id_buf)) != 0)
4033 softdep_error("softdep_update_inodeblock: bwrite", error);
4037 * Merge the new inode dependency list (id_newinoupdt) into the old
4038 * inode dependency list (id_inoupdt). This routine must be called
4039 * with splbio interrupts blocked.
4042 merge_inode_lists(struct inodedep *inodedep)
4044 struct allocdirect *listadp, *newadp;
4046 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
4047 for (listadp = TAILQ_FIRST(&inodedep->id_inoupdt); listadp && newadp;) {
4048 if (listadp->ad_lbn < newadp->ad_lbn) {
4049 listadp = TAILQ_NEXT(listadp, ad_next);
4052 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
4053 TAILQ_INSERT_BEFORE(listadp, newadp, ad_next);
4054 if (listadp->ad_lbn == newadp->ad_lbn) {
4055 allocdirect_merge(&inodedep->id_inoupdt, newadp,
4059 newadp = TAILQ_FIRST(&inodedep->id_newinoupdt);
4061 while ((newadp = TAILQ_FIRST(&inodedep->id_newinoupdt)) != NULL) {
4062 TAILQ_REMOVE(&inodedep->id_newinoupdt, newadp, ad_next);
4063 TAILQ_INSERT_TAIL(&inodedep->id_inoupdt, newadp, ad_next);
4068 * If we are doing an fsync, then we must ensure that any directory
4069 * entries for the inode have been written after the inode gets to disk.
4072 * vp: the "in_core" copy of the inode
4075 softdep_fsync(struct vnode *vp)
4077 struct inodedep *inodedep;
4078 struct pagedep *pagedep;
4079 struct worklist *wk;
4086 int error, flushparent;
4091 * Move check from original kernel code, possibly not needed any
4092 * more with the per-mount bioops.
4094 if ((vp->v_mount->mnt_flag & MNT_SOFTDEP) == 0)
4100 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0) {
4104 if (LIST_FIRST(&inodedep->id_inowait) != NULL ||
4105 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
4106 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
4107 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL) {
4109 panic("softdep_fsync: pending ops");
4111 for (error = 0, flushparent = 0; ; ) {
4112 if ((wk = LIST_FIRST(&inodedep->id_pendinghd)) == NULL)
4114 if (wk->wk_type != D_DIRADD) {
4116 panic("softdep_fsync: Unexpected type %s",
4117 TYPENAME(wk->wk_type));
4119 dap = WK_DIRADD(wk);
4121 * Flush our parent if this directory entry
4122 * has a MKDIR_PARENT dependency.
4124 if (dap->da_state & DIRCHG)
4125 pagedep = dap->da_previous->dm_pagedep;
4127 pagedep = dap->da_pagedep;
4128 mnt = pagedep->pd_mnt;
4129 parentino = pagedep->pd_ino;
4130 lbn = pagedep->pd_lbn;
4131 if ((dap->da_state & (MKDIR_BODY | COMPLETE)) != COMPLETE) {
4133 panic("softdep_fsync: dirty");
4135 flushparent = dap->da_state & MKDIR_PARENT;
4137 * If we are being fsync'ed as part of vgone'ing this vnode,
4138 * then we will not be able to release and recover the
4139 * vnode below, so we just have to give up on writing its
4140 * directory entry out. It will eventually be written, just
4141 * not now, but then the user was not asking to have it
4142 * written, so we are not breaking any promises.
4144 if (vp->v_flag & VRECLAIMED)
4147 * We prevent deadlock by always fetching inodes from the
4148 * root, moving down the directory tree. Thus, when fetching
4149 * our parent directory, we must unlock ourselves before
4150 * requesting the lock on our parent. See the comment in
4151 * ufs_lookup for details on possible races.
4155 error = VFS_VGET(mnt, parentino, &pvp);
4156 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4160 if ((error = ffs_update(pvp, 1)) != 0) {
4166 * Flush directory page containing the inode's name.
4168 error = bread(pvp, lblktodoff(fs, lbn), blksize(fs, VTOI(pvp), lbn), &bp);
4175 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) == 0)
4183 * Flush all the dirty bitmaps associated with the block device
4184 * before flushing the rest of the dirty blocks so as to reduce
4185 * the number of dependencies that will have to be rolled back.
4187 static int softdep_fsync_mountdev_bp(struct buf *bp, void *data);
4190 softdep_fsync_mountdev(struct vnode *vp)
4192 if (!vn_isdisk(vp, NULL))
4193 panic("softdep_fsync_mountdev: vnode not a disk");
4195 RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
4196 softdep_fsync_mountdev_bp, vp);
4197 drain_output(vp, 1);
4202 softdep_fsync_mountdev_bp(struct buf *bp, void *data)
4204 struct worklist *wk;
4205 struct vnode *vp = data;
4208 * If it is already scheduled, skip to the next buffer.
4210 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT))
4212 if (bp->b_vp != vp || (bp->b_flags & B_DELWRI) == 0) {
4214 kprintf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp);
4218 * We are only interested in bitmaps with outstanding
4221 if ((wk = LIST_FIRST(&bp->b_dep)) == NULL ||
4222 wk->wk_type != D_BMSAFEMAP) {
4234 * This routine is called when we are trying to synchronously flush a
4235 * file. This routine must eliminate any filesystem metadata dependencies
4236 * so that the syncing routine can succeed by pushing the dirty blocks
4237 * associated with the file. If any I/O errors occur, they are returned.
4239 struct softdep_sync_metadata_info {
4244 static int softdep_sync_metadata_bp(struct buf *bp, void *data);
4247 softdep_sync_metadata(struct vnode *vp, struct thread *td)
4249 struct softdep_sync_metadata_info info;
4253 * Check whether this vnode is involved in a filesystem
4254 * that is doing soft dependency processing.
4256 if (!vn_isdisk(vp, NULL)) {
4257 if (!DOINGSOFTDEP(vp))
4260 if (vp->v_rdev->si_mountpoint == NULL ||
4261 (vp->v_rdev->si_mountpoint->mnt_flag & MNT_SOFTDEP) == 0)
4264 * Ensure that any direct block dependencies have been cleared.
4267 if ((error = flush_inodedep_deps(VTOI(vp)->i_fs, VTOI(vp)->i_number))) {
4272 * For most files, the only metadata dependencies are the
4273 * cylinder group maps that allocate their inode or blocks.
4274 * The block allocation dependencies can be found by traversing
4275 * the dependency lists for any buffers that remain on their
4276 * dirty buffer list. The inode allocation dependency will
4277 * be resolved when the inode is updated with MNT_WAIT.
4278 * This work is done in two passes. The first pass grabs most
4279 * of the buffers and begins asynchronously writing them. The
4280 * only way to wait for these asynchronous writes is to sleep
4281 * on the filesystem vnode which may stay busy for a long time
4282 * if the filesystem is active. So, instead, we make a second
4283 * pass over the dependencies blocking on each write. In the
4284 * usual case we will be blocking against a write that we
4285 * initiated, so when it is done the dependency will have been
4286 * resolved. Thus the second pass is expected to end quickly.
4288 waitfor = MNT_NOWAIT;
4291 * We must wait for any I/O in progress to finish so that
4292 * all potential buffers on the dirty list will be visible.
4294 drain_output(vp, 1);
4296 info.waitfor = waitfor;
4297 error = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
4298 softdep_sync_metadata_bp, &info);
4301 return(-error); /* error code */
4305 * The brief unlock is to allow any pent up dependency
4306 * processing to be done. Then proceed with the second pass.
4308 if (waitfor == MNT_NOWAIT) {
4316 * If we have managed to get rid of all the dirty buffers,
4317 * then we are done. For certain directories and block
4318 * devices, we may need to do further work.
4320 * We must wait for any I/O in progress to finish so that
4321 * all potential buffers on the dirty list will be visible.
4323 drain_output(vp, 1);
4324 if (RB_EMPTY(&vp->v_rbdirty_tree)) {
4331 * If we are trying to sync a block device, some of its buffers may
4332 * contain metadata that cannot be written until the contents of some
4333 * partially written files have been written to disk. The only easy
4334 * way to accomplish this is to sync the entire filesystem (luckily
4335 * this happens rarely).
4337 if (vn_isdisk(vp, NULL) &&
4339 vp->v_rdev->si_mountpoint && !vn_islocked(vp) &&
4340 (error = VFS_SYNC(vp->v_rdev->si_mountpoint, MNT_WAIT)) != 0)
4346 softdep_sync_metadata_bp(struct buf *bp, void *data)
4348 struct softdep_sync_metadata_info *info = data;
4349 struct pagedep *pagedep;
4350 struct allocdirect *adp;
4351 struct allocindir *aip;
4352 struct worklist *wk;
4357 if (getdirtybuf(&bp, MNT_WAIT) == 0) {
4358 kprintf("softdep_sync_metadata_bp(1): caught buf %p going away\n", bp);
4361 if (bp->b_vp != info->vp || (bp->b_flags & B_DELWRI) == 0) {
4362 kprintf("softdep_sync_metadata_bp(2): caught buf %p going away vp %p\n", bp, info->vp);
4368 * As we hold the buffer locked, none of its dependencies
4371 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
4372 switch (wk->wk_type) {
4375 adp = WK_ALLOCDIRECT(wk);
4376 if (adp->ad_state & DEPCOMPLETE)
4379 if (getdirtybuf(&nbp, info->waitfor) == 0)
4382 if (info->waitfor == MNT_NOWAIT) {
4384 } else if ((error = bwrite(nbp)) != 0) {
4393 aip = WK_ALLOCINDIR(wk);
4394 if (aip->ai_state & DEPCOMPLETE)
4397 if (getdirtybuf(&nbp, info->waitfor) == 0)
4400 if (info->waitfor == MNT_NOWAIT) {
4402 } else if ((error = bwrite(nbp)) != 0) {
4413 LIST_FOREACH(aip, &WK_INDIRDEP(wk)->ir_deplisthd, ai_next) {
4414 if (aip->ai_state & DEPCOMPLETE)
4417 if (getdirtybuf(&nbp, MNT_WAIT) == 0)
4420 if ((error = bwrite(nbp)) != 0) {
4431 if ((error = flush_inodedep_deps(WK_INODEDEP(wk)->id_fs,
4432 WK_INODEDEP(wk)->id_ino)) != 0) {
4442 * We are trying to sync a directory that may
4443 * have dependencies on both its own metadata
4444 * and/or dependencies on the inodes of any
4445 * recently allocated files. We walk its diradd
4446 * lists pushing out the associated inode.
4448 pagedep = WK_PAGEDEP(wk);
4449 for (i = 0; i < DAHASHSZ; i++) {
4450 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) == 0)
4453 flush_pagedep_deps(info->vp,
4455 &pagedep->pd_diraddhd[i]))) {
4466 * This case should never happen if the vnode has
4467 * been properly sync'ed. However, if this function
4468 * is used at a place where the vnode has not yet
4469 * been sync'ed, this dependency can show up. So,
4470 * rather than panic, just flush it.
4472 nbp = WK_MKDIR(wk)->md_buf;
4473 if (getdirtybuf(&nbp, info->waitfor) == 0)
4476 if (info->waitfor == MNT_NOWAIT) {
4478 } else if ((error = bwrite(nbp)) != 0) {
4488 * This case should never happen if the vnode has
4489 * been properly sync'ed. However, if this function
4490 * is used at a place where the vnode has not yet
4491 * been sync'ed, this dependency can show up. So,
4492 * rather than panic, just flush it.
4494 * nbp can wind up == bp if a device node for the
4495 * same filesystem is being fsynced at the same time,
4496 * leading to a panic if we don't catch the case.
4498 nbp = WK_BMSAFEMAP(wk)->sm_buf;
4501 if (getdirtybuf(&nbp, info->waitfor) == 0)
4504 if (info->waitfor == MNT_NOWAIT) {
4506 } else if ((error = bwrite(nbp)) != 0) {
4516 panic("softdep_sync_metadata: Unknown type %s",
4517 TYPENAME(wk->wk_type));
4528 * Flush the dependencies associated with an inodedep.
4529 * Called with splbio blocked.
4532 flush_inodedep_deps(struct fs *fs, ino_t ino)
4534 struct inodedep *inodedep;
4535 struct allocdirect *adp;
4540 * This work is done in two passes. The first pass grabs most
4541 * of the buffers and begins asynchronously writing them. The
4542 * only way to wait for these asynchronous writes is to sleep
4543 * on the filesystem vnode which may stay busy for a long time
4544 * if the filesystem is active. So, instead, we make a second
4545 * pass over the dependencies blocking on each write. In the
4546 * usual case we will be blocking against a write that we
4547 * initiated, so when it is done the dependency will have been
4548 * resolved. Thus the second pass is expected to end quickly.
4549 * We give a brief window at the top of the loop to allow
4550 * any pending I/O to complete.
4552 for (waitfor = MNT_NOWAIT; ; ) {
4555 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4557 TAILQ_FOREACH(adp, &inodedep->id_inoupdt, 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) {
4578 TAILQ_FOREACH(adp, &inodedep->id_newinoupdt, ad_next) {
4579 if (adp->ad_state & DEPCOMPLETE)
4582 if (getdirtybuf(&bp, waitfor) == 0) {
4583 if (waitfor == MNT_NOWAIT)
4588 if (waitfor == MNT_NOWAIT) {
4590 } else if ((error = bwrite(bp)) != 0) {
4600 * If pass2, we are done, otherwise do pass 2.
4602 if (waitfor == MNT_WAIT)
4607 * Try freeing inodedep in case all dependencies have been removed.
4609 if (inodedep_lookup(fs, ino, 0, &inodedep) != 0)
4610 (void) free_inodedep(inodedep);
4615 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
4616 * Called with splbio blocked.
4619 flush_pagedep_deps(struct vnode *pvp, struct mount *mp,
4620 struct diraddhd *diraddhdp)
4622 struct inodedep *inodedep;
4623 struct ufsmount *ump;
4626 int gotit, error = 0;
4631 while ((dap = LIST_FIRST(diraddhdp)) != NULL) {
4633 * Flush ourselves if this directory entry
4634 * has a MKDIR_PARENT dependency.
4636 if (dap->da_state & MKDIR_PARENT) {
4638 if ((error = ffs_update(pvp, 1)) != 0)
4642 * If that cleared dependencies, go on to next.
4644 if (dap != LIST_FIRST(diraddhdp))
4646 if (dap->da_state & MKDIR_PARENT) {
4648 panic("flush_pagedep_deps: MKDIR_PARENT");
4652 * A newly allocated directory must have its "." and
4653 * ".." entries written out before its name can be
4654 * committed in its parent. We do not want or need
4655 * the full semantics of a synchronous VOP_FSYNC as
4656 * that may end up here again, once for each directory
4657 * level in the filesystem. Instead, we push the blocks
4658 * and wait for them to clear. We have to fsync twice
4659 * because the first call may choose to defer blocks
4660 * that still have dependencies, but deferral will
4661 * happen at most once.
4663 inum = dap->da_newinum;
4664 if (dap->da_state & MKDIR_BODY) {
4666 if ((error = VFS_VGET(mp, inum, &vp)) != 0)
4668 if ((error=VOP_FSYNC(vp, MNT_NOWAIT)) ||
4669 (error=VOP_FSYNC(vp, MNT_NOWAIT))) {
4673 drain_output(vp, 0);
4677 * If that cleared dependencies, go on to next.
4679 if (dap != LIST_FIRST(diraddhdp))
4681 if (dap->da_state & MKDIR_BODY) {
4683 panic("flush_pagedep_deps: MKDIR_BODY");
4687 * Flush the inode on which the directory entry depends.
4688 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
4689 * the only remaining dependency is that the updated inode
4690 * count must get pushed to disk. The inode has already
4691 * been pushed into its inode buffer (via VOP_UPDATE) at
4692 * the time of the reference count change. So we need only
4693 * locate that buffer, ensure that there will be no rollback
4694 * caused by a bitmap dependency, then write the inode buffer.
4696 if (inodedep_lookup(ump->um_fs, inum, 0, &inodedep) == 0) {
4698 panic("flush_pagedep_deps: lost inode");
4701 * If the inode still has bitmap dependencies,
4702 * push them to disk.
4704 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
4705 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
4707 if (gotit && (error = bwrite(inodedep->id_buf)) != 0)
4710 if (dap != LIST_FIRST(diraddhdp))
4714 * If the inode is still sitting in a buffer waiting
4715 * to be written, push it to disk.
4718 if ((error = bread(ump->um_devvp,
4719 fsbtodoff(ump->um_fs, ino_to_fsba(ump->um_fs, inum)),
4720 (int)ump->um_fs->fs_bsize, &bp)) != 0)
4722 if ((error = bwrite(bp)) != 0)
4726 * If we have failed to get rid of all the dependencies
4727 * then something is seriously wrong.
4729 if (dap == LIST_FIRST(diraddhdp)) {
4731 panic("flush_pagedep_deps: flush failed");
4740 * A large burst of file addition or deletion activity can drive the
4741 * memory load excessively high. First attempt to slow things down
4742 * using the techniques below. If that fails, this routine requests
4743 * the offending operations to fall back to running synchronously
4744 * until the memory load returns to a reasonable level.
4747 softdep_slowdown(struct vnode *vp)
4749 int max_softdeps_hard;
4751 max_softdeps_hard = max_softdeps * 11 / 10;
4752 if (num_dirrem < max_softdeps_hard / 2 &&
4753 num_inodedep < max_softdeps_hard)
4755 stat_sync_limit_hit += 1;
4760 * If memory utilization has gotten too high, deliberately slow things
4761 * down and speed up the I/O processing.
4764 request_cleanup(int resource, int islocked)
4766 struct thread *td = curthread; /* XXX */
4769 * We never hold up the filesystem syncer process.
4771 if (td == filesys_syncer)
4774 * First check to see if the work list has gotten backlogged.
4775 * If it has, co-opt this process to help clean up two entries.
4776 * Because this process may hold inodes locked, we cannot
4777 * handle any remove requests that might block on a locked
4778 * inode as that could lead to deadlock.
4780 if (num_on_worklist > max_softdeps / 10) {
4783 process_worklist_item(NULL, LK_NOWAIT);
4784 process_worklist_item(NULL, LK_NOWAIT);
4785 stat_worklist_push += 2;
4792 * If we are resource constrained on inode dependencies, try
4793 * flushing some dirty inodes. Otherwise, we are constrained
4794 * by file deletions, so try accelerating flushes of directories
4795 * with removal dependencies. We would like to do the cleanup
4796 * here, but we probably hold an inode locked at this point and
4797 * that might deadlock against one that we try to clean. So,
4798 * the best that we can do is request the syncer daemon to do
4799 * the cleanup for us.
4804 stat_ino_limit_push += 1;
4805 req_clear_inodedeps += 1;
4806 stat_countp = &stat_ino_limit_hit;
4810 stat_blk_limit_push += 1;
4811 req_clear_remove += 1;
4812 stat_countp = &stat_blk_limit_hit;
4818 panic("request_cleanup: unknown type");
4821 * Hopefully the syncer daemon will catch up and awaken us.
4822 * We wait at most tickdelay before proceeding in any case.
4827 if (!callout_active(&handle))
4828 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2,
4830 interlocked_sleep(&lk, SLEEP, (caddr_t)&proc_waiting, 0,
4839 * Awaken processes pausing in request_cleanup and clear proc_waiting
4840 * to indicate that there is no longer a timer running.
4843 pause_timer(void *arg)
4846 wakeup_one(&proc_waiting);
4847 if (proc_waiting > 0)
4848 callout_reset(&handle, tickdelay > 2 ? tickdelay : 2,
4851 callout_deactivate(&handle);
4855 * Flush out a directory with at least one removal dependency in an effort to
4856 * reduce the number of dirrem, freefile, and freeblks dependency structures.
4859 clear_remove(struct thread *td)
4861 struct pagedep_hashhead *pagedephd;
4862 struct pagedep *pagedep;
4863 static int next = 0;
4870 for (cnt = 0; cnt < pagedep_hash; cnt++) {
4871 pagedephd = &pagedep_hashtbl[next++];
4872 if (next >= pagedep_hash)
4874 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
4875 if (LIST_FIRST(&pagedep->pd_dirremhd) == NULL)
4877 mp = pagedep->pd_mnt;
4878 ino = pagedep->pd_ino;
4880 if ((error = VFS_VGET(mp, ino, &vp)) != 0) {
4881 softdep_error("clear_remove: vget", error);
4884 if ((error = VOP_FSYNC(vp, MNT_NOWAIT)))
4885 softdep_error("clear_remove: fsync", error);
4886 drain_output(vp, 0);
4895 * Clear out a block of dirty inodes in an effort to reduce
4896 * the number of inodedep dependency structures.
4898 struct clear_inodedeps_info {
4904 clear_inodedeps_mountlist_callback(struct mount *mp, void *data)
4906 struct clear_inodedeps_info *info = data;
4908 if ((mp->mnt_flag & MNT_SOFTDEP) && info->fs == VFSTOUFS(mp)->um_fs) {
4916 clear_inodedeps(struct thread *td)
4918 struct clear_inodedeps_info info;
4919 struct inodedep_hashhead *inodedephd;
4920 struct inodedep *inodedep;
4921 static int next = 0;
4925 ino_t firstino, lastino, ino;
4929 * Pick a random inode dependency to be cleared.
4930 * We will then gather up all the inodes in its block
4931 * that have dependencies and flush them out.
4933 for (cnt = 0; cnt < inodedep_hash; cnt++) {
4934 inodedephd = &inodedep_hashtbl[next++];
4935 if (next >= inodedep_hash)
4937 if ((inodedep = LIST_FIRST(inodedephd)) != NULL)
4940 if (inodedep == NULL) {
4945 * Ugly code to find mount point given pointer to superblock.
4947 fs = inodedep->id_fs;
4950 mountlist_scan(clear_inodedeps_mountlist_callback,
4951 &info, MNTSCAN_FORWARD|MNTSCAN_NOBUSY);
4953 * Find the last inode in the block with dependencies.
4955 firstino = inodedep->id_ino & ~(INOPB(fs) - 1);
4956 for (lastino = firstino + INOPB(fs) - 1; lastino > firstino; lastino--)
4957 if (inodedep_lookup(fs, lastino, 0, &inodedep) != 0)
4960 * Asynchronously push all but the last inode with dependencies.
4961 * Synchronously push the last inode with dependencies to ensure
4962 * that the inode block gets written to free up the inodedeps.
4964 for (ino = firstino; ino <= lastino; ino++) {
4965 if (inodedep_lookup(fs, ino, 0, &inodedep) == 0)
4968 if ((error = VFS_VGET(info.mp, ino, &vp)) != 0) {
4969 softdep_error("clear_inodedeps: vget", error);
4972 if (ino == lastino) {
4973 if ((error = VOP_FSYNC(vp, MNT_WAIT)))
4974 softdep_error("clear_inodedeps: fsync1", error);
4976 if ((error = VOP_FSYNC(vp, MNT_NOWAIT)))
4977 softdep_error("clear_inodedeps: fsync2", error);
4978 drain_output(vp, 0);
4987 * Function to determine if the buffer has outstanding dependencies
4988 * that will cause a roll-back if the buffer is written. If wantcount
4989 * is set, return number of dependencies, otherwise just yes or no.
4992 softdep_count_dependencies(struct buf *bp, int wantcount)
4994 struct worklist *wk;
4995 struct inodedep *inodedep;
4996 struct indirdep *indirdep;
4997 struct allocindir *aip;
4998 struct pagedep *pagedep;
5004 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
5005 switch (wk->wk_type) {
5008 inodedep = WK_INODEDEP(wk);
5009 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
5010 /* bitmap allocation dependency */
5015 if (TAILQ_FIRST(&inodedep->id_inoupdt)) {
5016 /* direct block pointer dependency */
5024 indirdep = WK_INDIRDEP(wk);
5026 LIST_FOREACH(aip, &indirdep->ir_deplisthd, ai_next) {
5027 /* indirect block pointer dependency */
5035 pagedep = WK_PAGEDEP(wk);
5036 for (i = 0; i < DAHASHSZ; i++) {
5038 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
5039 /* directory entry dependency */
5051 /* never a dependency on these blocks */
5056 panic("softdep_check_for_rollback: Unexpected type %s",
5057 TYPENAME(wk->wk_type));
5067 * Acquire exclusive access to a buffer.
5068 * Must be called with splbio blocked.
5069 * Return 1 if buffer was acquired.
5072 getdirtybuf(struct buf **bpp, int waitfor)
5078 if ((bp = *bpp) == NULL)
5080 if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) == 0)
5082 if (waitfor != MNT_WAIT)
5084 error = interlocked_sleep(&lk, LOCKBUF, bp,
5085 LK_EXCLUSIVE | LK_SLEEPFAIL, 0, 0);
5086 if (error != ENOLCK) {
5088 panic("getdirtybuf: inconsistent lock");
5091 if ((bp->b_flags & B_DELWRI) == 0) {
5100 * Wait for pending output on a vnode to complete.
5101 * Must be called with vnode locked.
5104 drain_output(struct vnode *vp, int islocked)
5109 while (vp->v_track_write.bk_active) {
5110 vp->v_track_write.bk_waitflag = 1;
5111 interlocked_sleep(&lk, SLEEP, &vp->v_track_write,
5119 * Called whenever a buffer that is being invalidated or reallocated
5120 * contains dependencies. This should only happen if an I/O error has
5121 * occurred. The routine is called with the buffer locked.
5124 softdep_deallocate_dependencies(struct buf *bp)
5126 if ((bp->b_flags & B_ERROR) == 0)
5127 panic("softdep_deallocate_dependencies: dangling deps");
5128 softdep_error(bp->b_vp->v_mount->mnt_stat.f_mntfromname, bp->b_error);
5129 panic("softdep_deallocate_dependencies: unrecovered I/O error");
5133 * Function to handle asynchronous write errors in the filesystem.
5136 softdep_error(char *func, int error)
5139 /* XXX should do something better! */
5140 kprintf("%s: got error %d while accessing filesystem\n", func, error);