kernel - lwkt_token revamp
[dragonfly.git] / sys / vfs / ufs / ffs_softdep.c
CommitLineData
984263bc
MD
1/*
2 * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
3 *
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).
8 *
9 * Further information about soft updates can be obtained from:
10 *
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
14 * USA
15 *
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
18 * are met:
19 *
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.
25 *
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
36 * SUCH DAMAGE.
37 *
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 $
c4df9635 40 * $DragonFly: src/sys/vfs/ufs/ffs_softdep.c,v 1.57 2008/06/28 17:59:51 dillon Exp $
984263bc
MD
41 */
42
43/*
44 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
45 */
46#ifndef DIAGNOSTIC
47#define DIAGNOSTIC
48#endif
49#ifndef DEBUG
50#define DEBUG
51#endif
52
53#include <sys/param.h>
54#include <sys/kernel.h>
55#include <sys/systm.h>
56#include <sys/buf.h>
57#include <sys/malloc.h>
58#include <sys/mount.h>
59#include <sys/proc.h>
60#include <sys/syslog.h>
61#include <sys/vnode.h>
62#include <sys/conf.h>
3020e3be 63#include <sys/buf2.h>
f91a71dd 64#include <machine/inttypes.h>
1f2de5d4
MD
65#include "dir.h"
66#include "quota.h"
67#include "inode.h"
68#include "ufsmount.h"
69#include "fs.h"
70#include "softdep.h"
71#include "ffs_extern.h"
72#include "ufs_extern.h"
984263bc 73
165dba55
DR
74#include <sys/thread2.h>
75
984263bc
MD
76/*
77 * These definitions need to be adapted to the system to which
78 * this file is being ported.
79 */
80/*
81 * malloc types defined for the softdep system.
82 */
83MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
84MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
85MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
86MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
87MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
88MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
89MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
90MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
91MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
92MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
93MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
94MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
95MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
96
97#define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE)
98
99#define D_PAGEDEP 0
100#define D_INODEDEP 1
101#define D_NEWBLK 2
102#define D_BMSAFEMAP 3
103#define D_ALLOCDIRECT 4
104#define D_INDIRDEP 5
105#define D_ALLOCINDIR 6
106#define D_FREEFRAG 7
107#define D_FREEBLKS 8
108#define D_FREEFILE 9
109#define D_DIRADD 10
110#define D_MKDIR 11
111#define D_DIRREM 12
112#define D_LAST D_DIRREM
113
114/*
115 * translate from workitem type to memory type
116 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
117 */
118static struct malloc_type *memtype[] = {
119 M_PAGEDEP,
120 M_INODEDEP,
121 M_NEWBLK,
122 M_BMSAFEMAP,
123 M_ALLOCDIRECT,
124 M_INDIRDEP,
125 M_ALLOCINDIR,
126 M_FREEFRAG,
127 M_FREEBLKS,
128 M_FREEFILE,
129 M_DIRADD,
130 M_MKDIR,
131 M_DIRREM
132};
133
134#define DtoM(type) (memtype[type])
135
136/*
137 * Names of malloc types.
138 */
139#define TYPENAME(type) \
140 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
984263bc
MD
141/*
142 * End system adaptaion definitions.
143 */
144
145/*
146 * Internal function prototypes.
147 */
f719c866
DR
148static void softdep_error(char *, int);
149static void drain_output(struct vnode *, int);
150static int getdirtybuf(struct buf **, int);
151static void clear_remove(struct thread *);
152static void clear_inodedeps(struct thread *);
153static int flush_pagedep_deps(struct vnode *, struct mount *,
a6ee311a 154 struct diraddhd *);
f719c866
DR
155static int flush_inodedep_deps(struct fs *, ino_t);
156static int handle_written_filepage(struct pagedep *, struct buf *);
157static void diradd_inode_written(struct diradd *, struct inodedep *);
158static int handle_written_inodeblock(struct inodedep *, struct buf *);
159static void handle_allocdirect_partdone(struct allocdirect *);
160static void handle_allocindir_partdone(struct allocindir *);
161static void initiate_write_filepage(struct pagedep *, struct buf *);
162static void handle_written_mkdir(struct mkdir *, int);
163static void initiate_write_inodeblock(struct inodedep *, struct buf *);
164static void handle_workitem_freefile(struct freefile *);
165static void handle_workitem_remove(struct dirrem *);
166static struct dirrem *newdirrem(struct buf *, struct inode *,
a6ee311a 167 struct inode *, int, struct dirrem **);
f719c866
DR
168static void free_diradd(struct diradd *);
169static void free_allocindir(struct allocindir *, struct inodedep *);
54078292 170static int indir_trunc (struct inode *, off_t, int, ufs_lbn_t, long *);
f719c866
DR
171static void deallocate_dependencies(struct buf *, struct inodedep *);
172static void free_allocdirect(struct allocdirectlst *,
a6ee311a 173 struct allocdirect *, int);
f719c866
DR
174static int check_inode_unwritten(struct inodedep *);
175static int free_inodedep(struct inodedep *);
176static void handle_workitem_freeblocks(struct freeblks *);
177static void merge_inode_lists(struct inodedep *);
178static void setup_allocindir_phase2(struct buf *, struct inode *,
a6ee311a 179 struct allocindir *);
f719c866 180static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t,
a6ee311a 181 ufs_daddr_t);
f719c866
DR
182static void handle_workitem_freefrag(struct freefrag *);
183static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long);
184static void allocdirect_merge(struct allocdirectlst *,
a6ee311a 185 struct allocdirect *, struct allocdirect *);
f719c866
DR
186static struct bmsafemap *bmsafemap_lookup(struct buf *);
187static int newblk_lookup(struct fs *, ufs_daddr_t, int,
a6ee311a 188 struct newblk **);
f719c866
DR
189static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
190static int pagedep_lookup(struct inode *, ufs_lbn_t, int,
a6ee311a 191 struct pagedep **);
f719c866
DR
192static void pause_timer(void *);
193static int request_cleanup(int, int);
194static int process_worklist_item(struct mount *, int);
195static void add_to_worklist(struct worklist *);
984263bc
MD
196
197/*
198 * Exported softdep operations.
199 */
f719c866
DR
200static void softdep_disk_io_initiation(struct buf *);
201static void softdep_disk_write_complete(struct buf *);
202static void softdep_deallocate_dependencies(struct buf *);
203static int softdep_fsync(struct vnode *);
204static int softdep_process_worklist(struct mount *);
205static void softdep_move_dependencies(struct buf *, struct buf *);
206static int softdep_count_dependencies(struct buf *bp, int);
27bc0cb1
MD
207static int softdep_checkread(struct buf *bp);
208static int softdep_checkwrite(struct buf *bp);
984263bc 209
bc50d880 210static struct bio_ops softdep_bioops = {
408357d8
MD
211 .io_start = softdep_disk_io_initiation,
212 .io_complete = softdep_disk_write_complete,
213 .io_deallocate = softdep_deallocate_dependencies,
214 .io_fsync = softdep_fsync,
215 .io_sync = softdep_process_worklist,
216 .io_movedeps = softdep_move_dependencies,
217 .io_countdeps = softdep_count_dependencies,
27bc0cb1
MD
218 .io_checkread = softdep_checkread,
219 .io_checkwrite = softdep_checkwrite
984263bc
MD
220};
221
222/*
223 * Locking primitives.
224 *
225 * For a uniprocessor, all we need to do is protect against disk
226 * interrupts. For a multiprocessor, this lock would have to be
227 * a mutex. A single mutex is used throughout this file, though
228 * finer grain locking could be used if contention warranted it.
229 *
230 * For a multiprocessor, the sleep call would accept a lock and
231 * release it after the sleep processing was complete. In a uniprocessor
232 * implementation there is no such interlock, so we simple mark
233 * the places where it needs to be done with the `interlocked' form
234 * of the lock calls. Since the uniprocessor sleep already interlocks
235 * the spl, there is nothing that really needs to be done.
236 */
237#ifndef /* NOT */ DEBUG
238static struct lockit {
984263bc 239} lk = { 0 };
02d8a449
MD
240#define ACQUIRE_LOCK(lk) crit_enter_id("softupdates");
241#define FREE_LOCK(lk) crit_exit_id("softupdates");
984263bc
MD
242
243#else /* DEBUG */
dadab5e9
MD
244#define NOHOLDER ((struct thread *)-1)
245#define SPECIAL_FLAG ((struct thread *)-2)
984263bc
MD
246static struct lockit {
247 int lkt_spl;
dadab5e9
MD
248 struct thread *lkt_held;
249} lk = { 0, NOHOLDER };
984263bc
MD
250static int lockcnt;
251
f719c866
DR
252static void acquire_lock(struct lockit *);
253static void free_lock(struct lockit *);
254void softdep_panic(char *);
984263bc
MD
255
256#define ACQUIRE_LOCK(lk) acquire_lock(lk)
257#define FREE_LOCK(lk) free_lock(lk)
258
259static void
3fcb1ab8 260acquire_lock(struct lockit *lk)
984263bc 261{
dadab5e9 262 thread_t holder;
984263bc 263
dadab5e9 264 if (lk->lkt_held != NOHOLDER) {
984263bc
MD
265 holder = lk->lkt_held;
266 FREE_LOCK(lk);
dadab5e9 267 if (holder == curthread)
984263bc
MD
268 panic("softdep_lock: locking against myself");
269 else
dadab5e9 270 panic("softdep_lock: lock held by %p", holder);
984263bc 271 }
02d8a449 272 crit_enter_id("softupdates");
dadab5e9 273 lk->lkt_held = curthread;
984263bc
MD
274 lockcnt++;
275}
276
277static void
3fcb1ab8 278free_lock(struct lockit *lk)
984263bc
MD
279{
280
dadab5e9 281 if (lk->lkt_held == NOHOLDER)
984263bc 282 panic("softdep_unlock: lock not held");
dadab5e9 283 lk->lkt_held = NOHOLDER;
02d8a449 284 crit_exit_id("softupdates");
984263bc
MD
285}
286
287/*
288 * Function to release soft updates lock and panic.
289 */
290void
3fcb1ab8 291softdep_panic(char *msg)
984263bc
MD
292{
293
dadab5e9 294 if (lk.lkt_held != NOHOLDER)
984263bc
MD
295 FREE_LOCK(&lk);
296 panic(msg);
297}
298#endif /* DEBUG */
299
f719c866 300static int interlocked_sleep(struct lockit *, int, void *, int,
a6ee311a 301 const char *, int);
984263bc
MD
302
303/*
304 * When going to sleep, we must save our SPL so that it does
305 * not get lost if some other process uses the lock while we
306 * are sleeping. We restore it after we have slept. This routine
307 * wraps the interlocking with functions that sleep. The list
308 * below enumerates the available set of operations.
309 */
310#define UNKNOWN 0
311#define SLEEP 1
312#define LOCKBUF 2
313
314static int
3fcb1ab8
SW
315interlocked_sleep(struct lockit *lk, int op, void *ident, int flags,
316 const char *wmesg, int timo)
984263bc 317{
dadab5e9 318 thread_t holder;
984263bc
MD
319 int s, retval;
320
321 s = lk->lkt_spl;
322# ifdef DEBUG
dadab5e9 323 if (lk->lkt_held == NOHOLDER)
984263bc 324 panic("interlocked_sleep: lock not held");
dadab5e9 325 lk->lkt_held = NOHOLDER;
984263bc
MD
326# endif /* DEBUG */
327 switch (op) {
328 case SLEEP:
329 retval = tsleep(ident, flags, wmesg, timo);
330 break;
331 case LOCKBUF:
332 retval = BUF_LOCK((struct buf *)ident, flags);
333 break;
334 default:
335 panic("interlocked_sleep: unknown operation");
336 }
337# ifdef DEBUG
dadab5e9 338 if (lk->lkt_held != NOHOLDER) {
984263bc
MD
339 holder = lk->lkt_held;
340 FREE_LOCK(lk);
dadab5e9 341 if (holder == curthread)
984263bc
MD
342 panic("interlocked_sleep: locking against self");
343 else
dadab5e9 344 panic("interlocked_sleep: lock held by %p", holder);
984263bc 345 }
dadab5e9 346 lk->lkt_held = curthread;
984263bc
MD
347 lockcnt++;
348# endif /* DEBUG */
349 lk->lkt_spl = s;
350 return (retval);
351}
352
353/*
354 * Place holder for real semaphores.
355 */
356struct sema {
357 int value;
dadab5e9 358 thread_t holder;
984263bc
MD
359 char *name;
360 int prio;
361 int timo;
362};
f719c866
DR
363static void sema_init(struct sema *, char *, int, int);
364static int sema_get(struct sema *, struct lockit *);
365static void sema_release(struct sema *);
984263bc
MD
366
367static void
3fcb1ab8 368sema_init(struct sema *semap, char *name, int prio, int timo)
984263bc
MD
369{
370
dadab5e9 371 semap->holder = NOHOLDER;
984263bc
MD
372 semap->value = 0;
373 semap->name = name;
374 semap->prio = prio;
375 semap->timo = timo;
376}
377
378static int
3fcb1ab8 379sema_get(struct sema *semap, struct lockit *interlock)
984263bc
MD
380{
381
382 if (semap->value++ > 0) {
383 if (interlock != NULL) {
384 interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
385 semap->prio, semap->name, semap->timo);
386 FREE_LOCK(interlock);
387 } else {
388 tsleep((caddr_t)semap, semap->prio, semap->name,
389 semap->timo);
390 }
391 return (0);
392 }
dadab5e9 393 semap->holder = curthread;
984263bc
MD
394 if (interlock != NULL)
395 FREE_LOCK(interlock);
396 return (1);
397}
398
399static void
3fcb1ab8 400sema_release(struct sema *semap)
984263bc
MD
401{
402
dadab5e9
MD
403 if (semap->value <= 0 || semap->holder != curthread) {
404 if (lk.lkt_held != NOHOLDER)
984263bc
MD
405 FREE_LOCK(&lk);
406 panic("sema_release: not held");
407 }
408 if (--semap->value > 0) {
409 semap->value = 0;
410 wakeup(semap);
411 }
dadab5e9 412 semap->holder = NOHOLDER;
984263bc
MD
413}
414
415/*
416 * Worklist queue management.
417 * These routines require that the lock be held.
418 */
419#ifndef /* NOT */ DEBUG
420#define WORKLIST_INSERT(head, item) do { \
421 (item)->wk_state |= ONWORKLIST; \
422 LIST_INSERT_HEAD(head, item, wk_list); \
423} while (0)
408357d8
MD
424
425#define WORKLIST_INSERT_BP(bp, item) do { \
426 (item)->wk_state |= ONWORKLIST; \
427 (bp)->b_ops = &softdep_bioops; \
428 LIST_INSERT_HEAD(&(bp)->b_dep, item, wk_list); \
429} while (0)
430
984263bc
MD
431#define WORKLIST_REMOVE(item) do { \
432 (item)->wk_state &= ~ONWORKLIST; \
433 LIST_REMOVE(item, wk_list); \
434} while (0)
408357d8 435
984263bc
MD
436#define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
437
438#else /* DEBUG */
f719c866
DR
439static void worklist_insert(struct workhead *, struct worklist *);
440static void worklist_remove(struct worklist *);
441static void workitem_free(struct worklist *, int);
984263bc 442
408357d8
MD
443#define WORKLIST_INSERT_BP(bp, item) do { \
444 (bp)->b_ops = &softdep_bioops; \
445 worklist_insert(&(bp)->b_dep, item); \
446} while (0)
447
984263bc
MD
448#define WORKLIST_INSERT(head, item) worklist_insert(head, item)
449#define WORKLIST_REMOVE(item) worklist_remove(item)
450#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
451
452static void
3fcb1ab8 453worklist_insert(struct workhead *head, struct worklist *item)
984263bc
MD
454{
455
dadab5e9 456 if (lk.lkt_held == NOHOLDER)
984263bc
MD
457 panic("worklist_insert: lock not held");
458 if (item->wk_state & ONWORKLIST) {
459 FREE_LOCK(&lk);
460 panic("worklist_insert: already on list");
461 }
462 item->wk_state |= ONWORKLIST;
463 LIST_INSERT_HEAD(head, item, wk_list);
464}
465
466static void
3fcb1ab8 467worklist_remove(struct worklist *item)
984263bc
MD
468{
469
dadab5e9 470 if (lk.lkt_held == NOHOLDER)
984263bc
MD
471 panic("worklist_remove: lock not held");
472 if ((item->wk_state & ONWORKLIST) == 0) {
473 FREE_LOCK(&lk);
474 panic("worklist_remove: not on list");
475 }
476 item->wk_state &= ~ONWORKLIST;
477 LIST_REMOVE(item, wk_list);
478}
479
480static void
3fcb1ab8 481workitem_free(struct worklist *item, int type)
984263bc
MD
482{
483
484 if (item->wk_state & ONWORKLIST) {
dadab5e9 485 if (lk.lkt_held != NOHOLDER)
984263bc
MD
486 FREE_LOCK(&lk);
487 panic("workitem_free: still on list");
488 }
489 if (item->wk_type != type) {
dadab5e9 490 if (lk.lkt_held != NOHOLDER)
984263bc
MD
491 FREE_LOCK(&lk);
492 panic("workitem_free: type mismatch");
493 }
494 FREE(item, DtoM(type));
495}
496#endif /* DEBUG */
497
498/*
499 * Workitem queue management
500 */
501static struct workhead softdep_workitem_pending;
502static int num_on_worklist; /* number of worklist items to be processed */
503static int softdep_worklist_busy; /* 1 => trying to do unmount */
504static int softdep_worklist_req; /* serialized waiters */
505static int max_softdeps; /* maximum number of structs before slowdown */
506static int tickdelay = 2; /* number of ticks to pause during slowdown */
507static int *stat_countp; /* statistic to count in proc_waiting timeout */
508static int proc_waiting; /* tracks whether we have a timeout posted */
47979091 509static struct callout handle; /* handle on posted proc_waiting timeout */
dadab5e9 510static struct thread *filesys_syncer; /* proc of filesystem syncer process */
984263bc
MD
511static int req_clear_inodedeps; /* syncer process flush some inodedeps */
512#define FLUSH_INODES 1
513static int req_clear_remove; /* syncer process flush some freeblks */
514#define FLUSH_REMOVE 2
515/*
516 * runtime statistics
517 */
518static int stat_worklist_push; /* number of worklist cleanups */
519static int stat_blk_limit_push; /* number of times block limit neared */
520static int stat_ino_limit_push; /* number of times inode limit neared */
521static int stat_blk_limit_hit; /* number of times block slowdown imposed */
522static int stat_ino_limit_hit; /* number of times inode slowdown imposed */
523static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */
524static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */
525static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */
526static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
527static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */
528#ifdef DEBUG
529#include <vm/vm.h>
530#include <sys/sysctl.h>
531SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
532SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
533SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
534SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
535SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
536SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
537SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
538SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
539SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
540SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
541SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
542SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
543#endif /* DEBUG */
544
545/*
546 * Add an item to the end of the work queue.
547 * This routine requires that the lock be held.
548 * This is the only routine that adds items to the list.
549 * The following routine is the only one that removes items
550 * and does so in order from first to last.
551 */
552static void
3fcb1ab8 553add_to_worklist(struct worklist *wk)
984263bc
MD
554{
555 static struct worklist *worklist_tail;
556
557 if (wk->wk_state & ONWORKLIST) {
dadab5e9 558 if (lk.lkt_held != NOHOLDER)
984263bc
MD
559 FREE_LOCK(&lk);
560 panic("add_to_worklist: already on list");
561 }
562 wk->wk_state |= ONWORKLIST;
563 if (LIST_FIRST(&softdep_workitem_pending) == NULL)
564 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
565 else
566 LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
567 worklist_tail = wk;
568 num_on_worklist += 1;
569}
570
571/*
572 * Process that runs once per second to handle items in the background queue.
573 *
574 * Note that we ensure that everything is done in the order in which they
575 * appear in the queue. The code below depends on this property to ensure
576 * that blocks of a file are freed before the inode itself is freed. This
577 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
578 * until all the old ones have been purged from the dependency lists.
579 */
580static int
3fcb1ab8 581softdep_process_worklist(struct mount *matchmnt)
984263bc 582{
dadab5e9 583 thread_t td = curthread;
984263bc
MD
584 int matchcnt, loopcount;
585 long starttime;
586
587 /*
588 * Record the process identifier of our caller so that we can give
589 * this process preferential treatment in request_cleanup below.
590 */
dadab5e9 591 filesys_syncer = td;
984263bc
MD
592 matchcnt = 0;
593
594 /*
595 * There is no danger of having multiple processes run this
596 * code, but we have to single-thread it when softdep_flushfiles()
597 * is in operation to get an accurate count of the number of items
598 * related to its mount point that are in the list.
599 */
600 if (matchmnt == NULL) {
601 if (softdep_worklist_busy < 0)
602 return(-1);
603 softdep_worklist_busy += 1;
604 }
605
606 /*
607 * If requested, try removing inode or removal dependencies.
608 */
609 if (req_clear_inodedeps) {
dadab5e9 610 clear_inodedeps(td);
984263bc
MD
611 req_clear_inodedeps -= 1;
612 wakeup_one(&proc_waiting);
613 }
614 if (req_clear_remove) {
dadab5e9 615 clear_remove(td);
984263bc
MD
616 req_clear_remove -= 1;
617 wakeup_one(&proc_waiting);
618 }
619 loopcount = 1;
620 starttime = time_second;
621 while (num_on_worklist > 0) {
622 matchcnt += process_worklist_item(matchmnt, 0);
623
624 /*
625 * If a umount operation wants to run the worklist
626 * accurately, abort.
627 */
628 if (softdep_worklist_req && matchmnt == NULL) {
629 matchcnt = -1;
630 break;
631 }
632
633 /*
634 * If requested, try removing inode or removal dependencies.
635 */
636 if (req_clear_inodedeps) {
dadab5e9 637 clear_inodedeps(td);
984263bc
MD
638 req_clear_inodedeps -= 1;
639 wakeup_one(&proc_waiting);
640 }
641 if (req_clear_remove) {
dadab5e9 642 clear_remove(td);
984263bc
MD
643 req_clear_remove -= 1;
644 wakeup_one(&proc_waiting);
645 }
646 /*
647 * We do not generally want to stop for buffer space, but if
648 * we are really being a buffer hog, we will stop and wait.
649 */
650 if (loopcount++ % 128 == 0)
c4df9635 651 bwillinode(1);
984263bc
MD
652 /*
653 * Never allow processing to run for more than one
654 * second. Otherwise the other syncer tasks may get
655 * excessively backlogged.
656 */
657 if (starttime != time_second && matchmnt == NULL) {
658 matchcnt = -1;
659 break;
660 }
661 }
662 if (matchmnt == NULL) {
663 --softdep_worklist_busy;
664 if (softdep_worklist_req && softdep_worklist_busy == 0)
665 wakeup(&softdep_worklist_req);
666 }
667 return (matchcnt);
668}
669
670/*
671 * Process one item on the worklist.
672 */
673static int
3fcb1ab8 674process_worklist_item(struct mount *matchmnt, int flags)
984263bc
MD
675{
676 struct worklist *wk;
677 struct dirrem *dirrem;
678 struct fs *matchfs;
679 struct vnode *vp;
680 int matchcnt = 0;
681
682 matchfs = NULL;
683 if (matchmnt != NULL)
684 matchfs = VFSTOUFS(matchmnt)->um_fs;
685 ACQUIRE_LOCK(&lk);
686 /*
687 * Normally we just process each item on the worklist in order.
688 * However, if we are in a situation where we cannot lock any
689 * inodes, we have to skip over any dirrem requests whose
690 * vnodes are resident and locked.
691 */
692 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
693 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
694 break;
695 dirrem = WK_DIRREM(wk);
696 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev,
697 dirrem->dm_oldinum);
a11aaa81 698 if (vp == NULL || !vn_islocked(vp))
984263bc
MD
699 break;
700 }
701 if (wk == 0) {
702 FREE_LOCK(&lk);
703 return (0);
704 }
705 WORKLIST_REMOVE(wk);
706 num_on_worklist -= 1;
707 FREE_LOCK(&lk);
708 switch (wk->wk_type) {
709
710 case D_DIRREM:
711 /* removal of a directory entry */
712 if (WK_DIRREM(wk)->dm_mnt == matchmnt)
713 matchcnt += 1;
714 handle_workitem_remove(WK_DIRREM(wk));
715 break;
716
717 case D_FREEBLKS:
718 /* releasing blocks and/or fragments from a file */
719 if (WK_FREEBLKS(wk)->fb_fs == matchfs)
720 matchcnt += 1;
721 handle_workitem_freeblocks(WK_FREEBLKS(wk));
722 break;
723
724 case D_FREEFRAG:
725 /* releasing a fragment when replaced as a file grows */
726 if (WK_FREEFRAG(wk)->ff_fs == matchfs)
727 matchcnt += 1;
728 handle_workitem_freefrag(WK_FREEFRAG(wk));
729 break;
730
731 case D_FREEFILE:
732 /* releasing an inode when its link count drops to 0 */
733 if (WK_FREEFILE(wk)->fx_fs == matchfs)
734 matchcnt += 1;
735 handle_workitem_freefile(WK_FREEFILE(wk));
736 break;
737
738 default:
739 panic("%s_process_worklist: Unknown type %s",
740 "softdep", TYPENAME(wk->wk_type));
741 /* NOTREACHED */
742 }
743 return (matchcnt);
744}
745
746/*
747 * Move dependencies from one buffer to another.
748 */
749static void
3fcb1ab8 750softdep_move_dependencies(struct buf *oldbp, struct buf *newbp)
984263bc
MD
751{
752 struct worklist *wk, *wktail;
753
754 if (LIST_FIRST(&newbp->b_dep) != NULL)
755 panic("softdep_move_dependencies: need merge code");
408357d8 756 wktail = NULL;
984263bc
MD
757 ACQUIRE_LOCK(&lk);
758 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
759 LIST_REMOVE(wk, wk_list);
408357d8 760 if (wktail == NULL)
984263bc
MD
761 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
762 else
763 LIST_INSERT_AFTER(wktail, wk, wk_list);
764 wktail = wk;
408357d8 765 newbp->b_ops = &softdep_bioops;
984263bc
MD
766 }
767 FREE_LOCK(&lk);
768}
769
770/*
771 * Purge the work list of all items associated with a particular mount point.
772 */
773int
2aa32050 774softdep_flushfiles(struct mount *oldmnt, int flags)
984263bc
MD
775{
776 struct vnode *devvp;
777 int error, loopcnt;
778
779 /*
780 * Await our turn to clear out the queue, then serialize access.
781 */
782 while (softdep_worklist_busy != 0) {
783 softdep_worklist_req += 1;
377d4740 784 tsleep(&softdep_worklist_req, 0, "softflush", 0);
984263bc
MD
785 softdep_worklist_req -= 1;
786 }
787 softdep_worklist_busy = -1;
788
2aa32050 789 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) {
984263bc
MD
790 softdep_worklist_busy = 0;
791 if (softdep_worklist_req)
792 wakeup(&softdep_worklist_req);
793 return (error);
794 }
795 /*
796 * Alternately flush the block device associated with the mount
797 * point and process any dependencies that the flushing
798 * creates. In theory, this loop can happen at most twice,
799 * but we give it a few extra just to be sure.
800 */
801 devvp = VFSTOUFS(oldmnt)->um_devvp;
802 for (loopcnt = 10; loopcnt > 0; ) {
803 if (softdep_process_worklist(oldmnt) == 0) {
804 loopcnt--;
805 /*
806 * Do another flush in case any vnodes were brought in
807 * as part of the cleanup operations.
808 */
2aa32050 809 if ((error = ffs_flushfiles(oldmnt, flags)) != 0)
984263bc
MD
810 break;
811 /*
812 * If we still found nothing to do, we are really done.
813 */
814 if (softdep_process_worklist(oldmnt) == 0)
815 break;
816 }
ca466bae 817 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
52174f71 818 error = VOP_FSYNC(devvp, MNT_WAIT, 0);
a11aaa81 819 vn_unlock(devvp);
984263bc
MD
820 if (error)
821 break;
822 }
823 softdep_worklist_busy = 0;
824 if (softdep_worklist_req)
825 wakeup(&softdep_worklist_req);
826
827 /*
828 * If we are unmounting then it is an error to fail. If we
829 * are simply trying to downgrade to read-only, then filesystem
830 * activity can keep us busy forever, so we just fail with EBUSY.
831 */
832 if (loopcnt == 0) {
833 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
834 panic("softdep_flushfiles: looping");
835 error = EBUSY;
836 }
837 return (error);
838}
839
840/*
841 * Structure hashing.
842 *
843 * There are three types of structures that can be looked up:
844 * 1) pagedep structures identified by mount point, inode number,
845 * and logical block.
846 * 2) inodedep structures identified by mount point and inode number.
847 * 3) newblk structures identified by mount point and
848 * physical block number.
849 *
850 * The "pagedep" and "inodedep" dependency structures are hashed
851 * separately from the file blocks and inodes to which they correspond.
852 * This separation helps when the in-memory copy of an inode or
853 * file block must be replaced. It also obviates the need to access
854 * an inode or file page when simply updating (or de-allocating)
855 * dependency structures. Lookup of newblk structures is needed to
856 * find newly allocated blocks when trying to associate them with
857 * their allocdirect or allocindir structure.
858 *
859 * The lookup routines optionally create and hash a new instance when
860 * an existing entry is not found.
861 */
862#define DEPALLOC 0x0001 /* allocate structure if lookup fails */
863#define NODELAY 0x0002 /* cannot do background work */
864
865/*
866 * Structures and routines associated with pagedep caching.
867 */
868LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
869u_long pagedep_hash; /* size of hash table - 1 */
870#define PAGEDEP_HASH(mp, inum, lbn) \
871 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
872 pagedep_hash])
873static struct sema pagedep_in_progress;
874
875/*
98a74972
MD
876 * Helper routine for pagedep_lookup()
877 */
878static __inline
879struct pagedep *
880pagedep_find(struct pagedep_hashhead *pagedephd, ino_t ino, ufs_lbn_t lbn,
881 struct mount *mp)
882{
883 struct pagedep *pagedep;
884
885 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
886 if (ino == pagedep->pd_ino &&
887 lbn == pagedep->pd_lbn &&
888 mp == pagedep->pd_mnt) {
889 return (pagedep);
890 }
891 }
892 return(NULL);
893}
894
895/*
984263bc
MD
896 * Look up a pagedep. Return 1 if found, 0 if not found.
897 * If not found, allocate if DEPALLOC flag is passed.
898 * Found or allocated entry is returned in pagedeppp.
899 * This routine must be called with splbio interrupts blocked.
900 */
901static int
3fcb1ab8
SW
902pagedep_lookup(struct inode *ip, ufs_lbn_t lbn, int flags,
903 struct pagedep **pagedeppp)
984263bc
MD
904{
905 struct pagedep *pagedep;
906 struct pagedep_hashhead *pagedephd;
907 struct mount *mp;
908 int i;
909
910#ifdef DEBUG
dadab5e9 911 if (lk.lkt_held == NOHOLDER)
984263bc
MD
912 panic("pagedep_lookup: lock not held");
913#endif
914 mp = ITOV(ip)->v_mount;
915 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
916top:
98a74972
MD
917 *pagedeppp = pagedep_find(pagedephd, ip->i_number, lbn, mp);
918 if (*pagedeppp)
919 return(1);
920 if ((flags & DEPALLOC) == 0)
984263bc 921 return (0);
984263bc
MD
922 if (sema_get(&pagedep_in_progress, &lk) == 0) {
923 ACQUIRE_LOCK(&lk);
924 goto top;
925 }
926 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
98a74972
MD
927 M_SOFTDEP_FLAGS | M_ZERO);
928
929 if (pagedep_find(pagedephd, ip->i_number, lbn, mp)) {
086c1d7e 930 kprintf("pagedep_lookup: blocking race avoided\n");
98a74972
MD
931 ACQUIRE_LOCK(&lk);
932 sema_release(&pagedep_in_progress);
efda3bd0 933 kfree(pagedep, M_PAGEDEP);
98a74972
MD
934 goto top;
935 }
936
984263bc
MD
937 pagedep->pd_list.wk_type = D_PAGEDEP;
938 pagedep->pd_mnt = mp;
939 pagedep->pd_ino = ip->i_number;
940 pagedep->pd_lbn = lbn;
941 LIST_INIT(&pagedep->pd_dirremhd);
942 LIST_INIT(&pagedep->pd_pendinghd);
943 for (i = 0; i < DAHASHSZ; i++)
944 LIST_INIT(&pagedep->pd_diraddhd[i]);
945 ACQUIRE_LOCK(&lk);
946 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
947 sema_release(&pagedep_in_progress);
948 *pagedeppp = pagedep;
949 return (0);
950}
951
952/*
953 * Structures and routines associated with inodedep caching.
954 */
955LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
956static u_long inodedep_hash; /* size of hash table - 1 */
957static long num_inodedep; /* number of inodedep allocated */
958#define INODEDEP_HASH(fs, inum) \
959 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
960static struct sema inodedep_in_progress;
961
962/*
98a74972
MD
963 * Helper routine for inodedep_lookup()
964 */
965static __inline
966struct inodedep *
967inodedep_find(struct inodedep_hashhead *inodedephd, struct fs *fs, ino_t inum)
968{
969 struct inodedep *inodedep;
970
971 LIST_FOREACH(inodedep, inodedephd, id_hash) {
972 if (inum == inodedep->id_ino && fs == inodedep->id_fs)
973 return(inodedep);
974 }
975 return (NULL);
976}
977
978/*
984263bc
MD
979 * Look up a inodedep. Return 1 if found, 0 if not found.
980 * If not found, allocate if DEPALLOC flag is passed.
981 * Found or allocated entry is returned in inodedeppp.
982 * This routine must be called with splbio interrupts blocked.
983 */
984static int
3fcb1ab8
SW
985inodedep_lookup(struct fs *fs, ino_t inum, int flags,
986 struct inodedep **inodedeppp)
984263bc
MD
987{
988 struct inodedep *inodedep;
989 struct inodedep_hashhead *inodedephd;
990 int firsttry;
991
992#ifdef DEBUG
dadab5e9 993 if (lk.lkt_held == NOHOLDER)
984263bc
MD
994 panic("inodedep_lookup: lock not held");
995#endif
996 firsttry = 1;
997 inodedephd = INODEDEP_HASH(fs, inum);
998top:
98a74972
MD
999 *inodedeppp = inodedep_find(inodedephd, fs, inum);
1000 if (*inodedeppp)
984263bc 1001 return (1);
98a74972 1002 if ((flags & DEPALLOC) == 0)
984263bc 1003 return (0);
984263bc
MD
1004 /*
1005 * If we are over our limit, try to improve the situation.
1006 */
1007 if (num_inodedep > max_softdeps && firsttry &&
1008 speedup_syncer() == 0 && (flags & NODELAY) == 0 &&
1009 request_cleanup(FLUSH_INODES, 1)) {
1010 firsttry = 0;
1011 goto top;
1012 }
1013 if (sema_get(&inodedep_in_progress, &lk) == 0) {
1014 ACQUIRE_LOCK(&lk);
1015 goto top;
1016 }
984263bc 1017 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
98a74972
MD
1018 M_INODEDEP, M_SOFTDEP_FLAGS | M_ZERO);
1019 if (inodedep_find(inodedephd, fs, inum)) {
086c1d7e 1020 kprintf("inodedep_lookup: blocking race avoided\n");
98a74972
MD
1021 ACQUIRE_LOCK(&lk);
1022 sema_release(&inodedep_in_progress);
efda3bd0 1023 kfree(inodedep, M_INODEDEP);
98a74972
MD
1024 goto top;
1025 }
984263bc
MD
1026 inodedep->id_list.wk_type = D_INODEDEP;
1027 inodedep->id_fs = fs;
1028 inodedep->id_ino = inum;
1029 inodedep->id_state = ALLCOMPLETE;
1030 inodedep->id_nlinkdelta = 0;
1031 inodedep->id_savedino = NULL;
1032 inodedep->id_savedsize = -1;
1033 inodedep->id_buf = NULL;
1034 LIST_INIT(&inodedep->id_pendinghd);
1035 LIST_INIT(&inodedep->id_inowait);
1036 LIST_INIT(&inodedep->id_bufwait);
1037 TAILQ_INIT(&inodedep->id_inoupdt);
1038 TAILQ_INIT(&inodedep->id_newinoupdt);
1039 ACQUIRE_LOCK(&lk);
98a74972 1040 num_inodedep += 1;
984263bc
MD
1041 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1042 sema_release(&inodedep_in_progress);
1043 *inodedeppp = inodedep;
1044 return (0);
1045}
1046
1047/*
1048 * Structures and routines associated with newblk caching.
1049 */
1050LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1051u_long newblk_hash; /* size of hash table - 1 */
1052#define NEWBLK_HASH(fs, inum) \
1053 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1054static struct sema newblk_in_progress;
1055
1056/*
98a74972
MD
1057 * Helper routine for newblk_lookup()
1058 */
1059static __inline
1060struct newblk *
1061newblk_find(struct newblk_hashhead *newblkhd, struct fs *fs,
1062 ufs_daddr_t newblkno)
1063{
1064 struct newblk *newblk;
1065
1066 LIST_FOREACH(newblk, newblkhd, nb_hash) {
1067 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1068 return (newblk);
1069 }
1070 return(NULL);
1071}
1072
1073/*
984263bc
MD
1074 * Look up a newblk. Return 1 if found, 0 if not found.
1075 * If not found, allocate if DEPALLOC flag is passed.
1076 * Found or allocated entry is returned in newblkpp.
1077 */
1078static int
3fcb1ab8
SW
1079newblk_lookup(struct fs *fs, ufs_daddr_t newblkno, int flags,
1080 struct newblk **newblkpp)
984263bc
MD
1081{
1082 struct newblk *newblk;
1083 struct newblk_hashhead *newblkhd;
1084
1085 newblkhd = NEWBLK_HASH(fs, newblkno);
1086top:
98a74972
MD
1087 *newblkpp = newblk_find(newblkhd, fs, newblkno);
1088 if (*newblkpp)
1089 return(1);
1090 if ((flags & DEPALLOC) == 0)
984263bc 1091 return (0);
984263bc
MD
1092 if (sema_get(&newblk_in_progress, 0) == 0)
1093 goto top;
1094 MALLOC(newblk, struct newblk *, sizeof(struct newblk),
98a74972
MD
1095 M_NEWBLK, M_SOFTDEP_FLAGS | M_ZERO);
1096
1097 if (newblk_find(newblkhd, fs, newblkno)) {
086c1d7e 1098 kprintf("newblk_lookup: blocking race avoided\n");
98a74972 1099 sema_release(&pagedep_in_progress);
efda3bd0 1100 kfree(newblk, M_NEWBLK);
98a74972
MD
1101 goto top;
1102 }
984263bc
MD
1103 newblk->nb_state = 0;
1104 newblk->nb_fs = fs;
1105 newblk->nb_newblkno = newblkno;
1106 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1107 sema_release(&newblk_in_progress);
1108 *newblkpp = newblk;
1109 return (0);
1110}
1111
1112/*
1113 * Executed during filesystem system initialization before
f719c866 1114 * mounting any filesystems.
984263bc
MD
1115 */
1116void
3fcb1ab8 1117softdep_initialize(void)
984263bc 1118{
47979091 1119 callout_init(&handle);
984263bc
MD
1120
1121 LIST_INIT(&mkdirlisthd);
1122 LIST_INIT(&softdep_workitem_pending);
1123 max_softdeps = min(desiredvnodes * 8,
1124 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep)));
1125 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1126 &pagedep_hash);
377d4740 1127 sema_init(&pagedep_in_progress, "pagedep", 0, 0);
984263bc 1128 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
377d4740 1129 sema_init(&inodedep_in_progress, "inodedep", 0, 0);
984263bc 1130 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
377d4740 1131 sema_init(&newblk_in_progress, "newblk", 0, 0);
408357d8 1132 add_bio_ops(&softdep_bioops);
984263bc
MD
1133}
1134
1135/*
1136 * Called at mount time to notify the dependency code that a
1137 * filesystem wishes to use it.
1138 */
1139int
3fcb1ab8 1140softdep_mount(struct vnode *devvp, struct mount *mp, struct fs *fs)
984263bc
MD
1141{
1142 struct csum cstotal;
1143 struct cg *cgp;
1144 struct buf *bp;
1145 int error, cyl;
1146
1147 mp->mnt_flag &= ~MNT_ASYNC;
1148 mp->mnt_flag |= MNT_SOFTDEP;
408357d8 1149 mp->mnt_bioops = &softdep_bioops;
984263bc
MD
1150 /*
1151 * When doing soft updates, the counters in the
1152 * superblock may have gotten out of sync, so we have
1153 * to scan the cylinder groups and recalculate them.
1154 */
1155 if (fs->fs_clean != 0)
1156 return (0);
1157 bzero(&cstotal, sizeof cstotal);
1158 for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
54078292
MD
1159 if ((error = bread(devvp, fsbtodoff(fs, cgtod(fs, cyl)),
1160 fs->fs_cgsize, &bp)) != 0) {
984263bc
MD
1161 brelse(bp);
1162 return (error);
1163 }
1164 cgp = (struct cg *)bp->b_data;
1165 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1166 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1167 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1168 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1169 fs->fs_cs(fs, cyl) = cgp->cg_cs;
1170 brelse(bp);
1171 }
1172#ifdef DEBUG
1173 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
086c1d7e 1174 kprintf("ffs_mountfs: superblock updated for soft updates\n");
984263bc
MD
1175#endif
1176 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1177 return (0);
1178}
1179
1180/*
1181 * Protecting the freemaps (or bitmaps).
1182 *
f719c866 1183 * To eliminate the need to execute fsck before mounting a filesystem
984263bc
MD
1184 * after a power failure, one must (conservatively) guarantee that the
1185 * on-disk copy of the bitmaps never indicate that a live inode or block is
1186 * free. So, when a block or inode is allocated, the bitmap should be
1187 * updated (on disk) before any new pointers. When a block or inode is
1188 * freed, the bitmap should not be updated until all pointers have been
1189 * reset. The latter dependency is handled by the delayed de-allocation
1190 * approach described below for block and inode de-allocation. The former
1191 * dependency is handled by calling the following procedure when a block or
1192 * inode is allocated. When an inode is allocated an "inodedep" is created
1193 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1194 * Each "inodedep" is also inserted into the hash indexing structure so
1195 * that any additional link additions can be made dependent on the inode
1196 * allocation.
1197 *
f719c866 1198 * The ufs filesystem maintains a number of free block counts (e.g., per
984263bc
MD
1199 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1200 * in addition to the bitmaps. These counts are used to improve efficiency
1201 * during allocation and therefore must be consistent with the bitmaps.
1202 * There is no convenient way to guarantee post-crash consistency of these
1203 * counts with simple update ordering, for two main reasons: (1) The counts
1204 * and bitmaps for a single cylinder group block are not in the same disk
1205 * sector. If a disk write is interrupted (e.g., by power failure), one may
1206 * be written and the other not. (2) Some of the counts are located in the
1207 * superblock rather than the cylinder group block. So, we focus our soft
1208 * updates implementation on protecting the bitmaps. When mounting a
1209 * filesystem, we recompute the auxiliary counts from the bitmaps.
1210 */
1211
1212/*
1213 * Called just after updating the cylinder group block to allocate an inode.
3fcb1ab8
SW
1214 *
1215 * Parameters:
1216 * bp: buffer for cylgroup block with inode map
1217 * ip: inode related to allocation
1218 * newinum: new inode number being allocated
984263bc
MD
1219 */
1220void
3fcb1ab8 1221softdep_setup_inomapdep(struct buf *bp, struct inode *ip, ino_t newinum)
984263bc
MD
1222{
1223 struct inodedep *inodedep;
1224 struct bmsafemap *bmsafemap;
1225
1226 /*
1227 * Create a dependency for the newly allocated inode.
1228 * Panic if it already exists as something is seriously wrong.
1229 * Otherwise add it to the dependency list for the buffer holding
1230 * the cylinder group map from which it was allocated.
1231 */
1232 ACQUIRE_LOCK(&lk);
1233 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1234 FREE_LOCK(&lk);
1235 panic("softdep_setup_inomapdep: found inode");
1236 }
1237 inodedep->id_buf = bp;
1238 inodedep->id_state &= ~DEPCOMPLETE;
1239 bmsafemap = bmsafemap_lookup(bp);
1240 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1241 FREE_LOCK(&lk);
1242}
1243
1244/*
1245 * Called just after updating the cylinder group block to
1246 * allocate block or fragment.
3fcb1ab8
SW
1247 *
1248 * Parameters:
1249 * bp: buffer for cylgroup block with block map
1250 * fs: filesystem doing allocation
1251 * newblkno: number of newly allocated block
984263bc
MD
1252 */
1253void
3fcb1ab8
SW
1254softdep_setup_blkmapdep(struct buf *bp, struct fs *fs,
1255 ufs_daddr_t newblkno)
984263bc
MD
1256{
1257 struct newblk *newblk;
1258 struct bmsafemap *bmsafemap;
1259
1260 /*
1261 * Create a dependency for the newly allocated block.
1262 * Add it to the dependency list for the buffer holding
1263 * the cylinder group map from which it was allocated.
1264 */
1265 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1266 panic("softdep_setup_blkmapdep: found block");
1267 ACQUIRE_LOCK(&lk);
1268 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1269 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1270 FREE_LOCK(&lk);
1271}
1272
1273/*
1274 * Find the bmsafemap associated with a cylinder group buffer.
1275 * If none exists, create one. The buffer must be locked when
1276 * this routine is called and this routine must be called with
1277 * splbio interrupts blocked.
1278 */
1279static struct bmsafemap *
3fcb1ab8 1280bmsafemap_lookup(struct buf *bp)
984263bc
MD
1281{
1282 struct bmsafemap *bmsafemap;
1283 struct worklist *wk;
1284
1285#ifdef DEBUG
dadab5e9 1286 if (lk.lkt_held == NOHOLDER)
984263bc
MD
1287 panic("bmsafemap_lookup: lock not held");
1288#endif
408357d8 1289 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
984263bc
MD
1290 if (wk->wk_type == D_BMSAFEMAP)
1291 return (WK_BMSAFEMAP(wk));
408357d8 1292 }
984263bc
MD
1293 FREE_LOCK(&lk);
1294 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1295 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1296 bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1297 bmsafemap->sm_list.wk_state = 0;
1298 bmsafemap->sm_buf = bp;
1299 LIST_INIT(&bmsafemap->sm_allocdirecthd);
1300 LIST_INIT(&bmsafemap->sm_allocindirhd);
1301 LIST_INIT(&bmsafemap->sm_inodedephd);
1302 LIST_INIT(&bmsafemap->sm_newblkhd);
1303 ACQUIRE_LOCK(&lk);
408357d8 1304 WORKLIST_INSERT_BP(bp, &bmsafemap->sm_list);
984263bc
MD
1305 return (bmsafemap);
1306}
1307
1308/*
1309 * Direct block allocation dependencies.
1310 *
1311 * When a new block is allocated, the corresponding disk locations must be
1312 * initialized (with zeros or new data) before the on-disk inode points to
1313 * them. Also, the freemap from which the block was allocated must be
1314 * updated (on disk) before the inode's pointer. These two dependencies are
1315 * independent of each other and are needed for all file blocks and indirect
1316 * blocks that are pointed to directly by the inode. Just before the
1317 * "in-core" version of the inode is updated with a newly allocated block
1318 * number, a procedure (below) is called to setup allocation dependency
1319 * structures. These structures are removed when the corresponding
1320 * dependencies are satisfied or when the block allocation becomes obsolete
1321 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1322 * fragment that gets upgraded). All of these cases are handled in
1323 * procedures described later.
1324 *
1325 * When a file extension causes a fragment to be upgraded, either to a larger
1326 * fragment or to a full block, the on-disk location may change (if the
1327 * previous fragment could not simply be extended). In this case, the old
1328 * fragment must be de-allocated, but not until after the inode's pointer has
1329 * been updated. In most cases, this is handled by later procedures, which
1330 * will construct a "freefrag" structure to be added to the workitem queue
1331 * when the inode update is complete (or obsolete). The main exception to
1332 * this is when an allocation occurs while a pending allocation dependency
1333 * (for the same block pointer) remains. This case is handled in the main
1334 * allocation dependency setup procedure by immediately freeing the
1335 * unreferenced fragments.
3fcb1ab8
SW
1336 *
1337 * Parameters:
1338 * ip: inode to which block is being added
1339 * lbn: block pointer within inode
1340 * newblkno: disk block number being added
1341 * oldblkno: previous block number, 0 unless frag
1342 * newsize: size of new block
1343 * oldsize: size of new block
1344 * bp: bp for allocated block
984263bc
MD
1345 */
1346void
3fcb1ab8
SW
1347softdep_setup_allocdirect(struct inode *ip, ufs_lbn_t lbn, ufs_daddr_t newblkno,
1348 ufs_daddr_t oldblkno, long newsize, long oldsize,
1349 struct buf *bp)
984263bc
MD
1350{
1351 struct allocdirect *adp, *oldadp;
1352 struct allocdirectlst *adphead;
1353 struct bmsafemap *bmsafemap;
1354 struct inodedep *inodedep;
1355 struct pagedep *pagedep;
1356 struct newblk *newblk;
1357
1358 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
e7b4468c 1359 M_ALLOCDIRECT, M_SOFTDEP_FLAGS | M_ZERO);
984263bc
MD
1360 adp->ad_list.wk_type = D_ALLOCDIRECT;
1361 adp->ad_lbn = lbn;
1362 adp->ad_newblkno = newblkno;
1363 adp->ad_oldblkno = oldblkno;
1364 adp->ad_newsize = newsize;
1365 adp->ad_oldsize = oldsize;
1366 adp->ad_state = ATTACHED;
1367 if (newblkno == oldblkno)
1368 adp->ad_freefrag = NULL;
1369 else
1370 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1371
1372 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1373 panic("softdep_setup_allocdirect: lost block");
1374
1375 ACQUIRE_LOCK(&lk);
1376 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1377 adp->ad_inodedep = inodedep;
1378
1379 if (newblk->nb_state == DEPCOMPLETE) {
1380 adp->ad_state |= DEPCOMPLETE;
1381 adp->ad_buf = NULL;
1382 } else {
1383 bmsafemap = newblk->nb_bmsafemap;
1384 adp->ad_buf = bmsafemap->sm_buf;
1385 LIST_REMOVE(newblk, nb_deps);
1386 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1387 }
1388 LIST_REMOVE(newblk, nb_hash);
1389 FREE(newblk, M_NEWBLK);
1390
408357d8 1391 WORKLIST_INSERT_BP(bp, &adp->ad_list);
984263bc
MD
1392 if (lbn >= NDADDR) {
1393 /* allocating an indirect block */
1394 if (oldblkno != 0) {
1395 FREE_LOCK(&lk);
1396 panic("softdep_setup_allocdirect: non-zero indir");
1397 }
1398 } else {
1399 /*
1400 * Allocating a direct block.
1401 *
1402 * If we are allocating a directory block, then we must
1403 * allocate an associated pagedep to track additions and
1404 * deletions.
1405 */
1406 if ((ip->i_mode & IFMT) == IFDIR &&
408357d8
MD
1407 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) {
1408 WORKLIST_INSERT_BP(bp, &pagedep->pd_list);
1409 }
984263bc
MD
1410 }
1411 /*
1412 * The list of allocdirects must be kept in sorted and ascending
1413 * order so that the rollback routines can quickly determine the
1414 * first uncommitted block (the size of the file stored on disk
1415 * ends at the end of the lowest committed fragment, or if there
1416 * are no fragments, at the end of the highest committed block).
1417 * Since files generally grow, the typical case is that the new
1418 * block is to be added at the end of the list. We speed this
1419 * special case by checking against the last allocdirect in the
1420 * list before laboriously traversing the list looking for the
1421 * insertion point.
1422 */
1423 adphead = &inodedep->id_newinoupdt;
1424 oldadp = TAILQ_LAST(adphead, allocdirectlst);
1425 if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1426 /* insert at end of list */
1427 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1428 if (oldadp != NULL && oldadp->ad_lbn == lbn)
1429 allocdirect_merge(adphead, adp, oldadp);
1430 FREE_LOCK(&lk);
1431 return;
1432 }
1433 TAILQ_FOREACH(oldadp, adphead, ad_next) {
1434 if (oldadp->ad_lbn >= lbn)
1435 break;
1436 }
1437 if (oldadp == NULL) {
1438 FREE_LOCK(&lk);
1439 panic("softdep_setup_allocdirect: lost entry");
1440 }
1441 /* insert in middle of list */
1442 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1443 if (oldadp->ad_lbn == lbn)
1444 allocdirect_merge(adphead, adp, oldadp);
1445 FREE_LOCK(&lk);
1446}
1447
1448/*
1449 * Replace an old allocdirect dependency with a newer one.
1450 * This routine must be called with splbio interrupts blocked.
3fcb1ab8
SW
1451 *
1452 * Parameters:
1453 * adphead: head of list holding allocdirects
1454 * newadp: allocdirect being added
1455 * oldadp: existing allocdirect being checked
984263bc
MD
1456 */
1457static void
3fcb1ab8
SW
1458allocdirect_merge(struct allocdirectlst *adphead,
1459 struct allocdirect *newadp,
1460 struct allocdirect *oldadp)
984263bc
MD
1461{
1462 struct freefrag *freefrag;
1463
1464#ifdef DEBUG
dadab5e9 1465 if (lk.lkt_held == NOHOLDER)
984263bc
MD
1466 panic("allocdirect_merge: lock not held");
1467#endif
1468 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1469 newadp->ad_oldsize != oldadp->ad_newsize ||
1470 newadp->ad_lbn >= NDADDR) {
1471 FREE_LOCK(&lk);
1472 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1473 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1474 NDADDR);
1475 }
1476 newadp->ad_oldblkno = oldadp->ad_oldblkno;
1477 newadp->ad_oldsize = oldadp->ad_oldsize;
1478 /*
1479 * If the old dependency had a fragment to free or had never
1480 * previously had a block allocated, then the new dependency
1481 * can immediately post its freefrag and adopt the old freefrag.
1482 * This action is done by swapping the freefrag dependencies.
1483 * The new dependency gains the old one's freefrag, and the
1484 * old one gets the new one and then immediately puts it on
1485 * the worklist when it is freed by free_allocdirect. It is
1486 * not possible to do this swap when the old dependency had a
1487 * non-zero size but no previous fragment to free. This condition
1488 * arises when the new block is an extension of the old block.
1489 * Here, the first part of the fragment allocated to the new
1490 * dependency is part of the block currently claimed on disk by
1491 * the old dependency, so cannot legitimately be freed until the
1492 * conditions for the new dependency are fulfilled.
1493 */
1494 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1495 freefrag = newadp->ad_freefrag;
1496 newadp->ad_freefrag = oldadp->ad_freefrag;
1497 oldadp->ad_freefrag = freefrag;
1498 }
1499 free_allocdirect(adphead, oldadp, 0);
1500}
1501
1502/*
1503 * Allocate a new freefrag structure if needed.
1504 */
1505static struct freefrag *
3fcb1ab8 1506newfreefrag(struct inode *ip, ufs_daddr_t blkno, long size)
984263bc
MD
1507{
1508 struct freefrag *freefrag;
1509 struct fs *fs;
1510
1511 if (blkno == 0)
1512 return (NULL);
1513 fs = ip->i_fs;
1514 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1515 panic("newfreefrag: frag size");
1516 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1517 M_FREEFRAG, M_SOFTDEP_FLAGS);
1518 freefrag->ff_list.wk_type = D_FREEFRAG;
1519 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */
1520 freefrag->ff_inum = ip->i_number;
1521 freefrag->ff_fs = fs;
1522 freefrag->ff_devvp = ip->i_devvp;
1523 freefrag->ff_blkno = blkno;
1524 freefrag->ff_fragsize = size;
1525 return (freefrag);
1526}
1527
1528/*
1529 * This workitem de-allocates fragments that were replaced during
1530 * file block allocation.
1531 */
1532static void
3fcb1ab8 1533handle_workitem_freefrag(struct freefrag *freefrag)
984263bc
MD
1534{
1535 struct inode tip;
1536
1537 tip.i_fs = freefrag->ff_fs;
1538 tip.i_devvp = freefrag->ff_devvp;
1539 tip.i_dev = freefrag->ff_devvp->v_rdev;
1540 tip.i_number = freefrag->ff_inum;
1541 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */
1542 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1543 FREE(freefrag, M_FREEFRAG);
1544}
1545
1546/*
1547 * Indirect block allocation dependencies.
1548 *
1549 * The same dependencies that exist for a direct block also exist when
1550 * a new block is allocated and pointed to by an entry in a block of
1551 * indirect pointers. The undo/redo states described above are also
1552 * used here. Because an indirect block contains many pointers that
1553 * may have dependencies, a second copy of the entire in-memory indirect
1554 * block is kept. The buffer cache copy is always completely up-to-date.
1555 * The second copy, which is used only as a source for disk writes,
1556 * contains only the safe pointers (i.e., those that have no remaining
1557 * update dependencies). The second copy is freed when all pointers
1558 * are safe. The cache is not allowed to replace indirect blocks with
1559 * pending update dependencies. If a buffer containing an indirect
1560 * block with dependencies is written, these routines will mark it
1561 * dirty again. It can only be successfully written once all the
1562 * dependencies are removed. The ffs_fsync routine in conjunction with
1563 * softdep_sync_metadata work together to get all the dependencies
1564 * removed so that a file can be successfully written to disk. Three
1565 * procedures are used when setting up indirect block pointer
1566 * dependencies. The division is necessary because of the organization
1567 * of the "balloc" routine and because of the distinction between file
1568 * pages and file metadata blocks.
1569 */
1570
1571/*
1572 * Allocate a new allocindir structure.
3fcb1ab8
SW
1573 *
1574 * Parameters:
1575 * ip: inode for file being extended
1576 * ptrno: offset of pointer in indirect block
1577 * newblkno: disk block number being added
1578 * oldblkno: previous block number, 0 if none
984263bc
MD
1579 */
1580static struct allocindir *
3fcb1ab8
SW
1581newallocindir(struct inode *ip, int ptrno, ufs_daddr_t newblkno,
1582 ufs_daddr_t oldblkno)
984263bc
MD
1583{
1584 struct allocindir *aip;
1585
1586 MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
e7b4468c 1587 M_ALLOCINDIR, M_SOFTDEP_FLAGS | M_ZERO);
984263bc
MD
1588 aip->ai_list.wk_type = D_ALLOCINDIR;
1589 aip->ai_state = ATTACHED;
1590 aip->ai_offset = ptrno;
1591 aip->ai_newblkno = newblkno;
1592 aip->ai_oldblkno = oldblkno;
1593 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1594 return (aip);
1595}
1596
1597/*
1598 * Called just before setting an indirect block pointer
1599 * to a newly allocated file page.
3fcb1ab8
SW
1600 *
1601 * Parameters:
1602 * ip: inode for file being extended
1603 * lbn: allocated block number within file
1604 * bp: buffer with indirect blk referencing page
1605 * ptrno: offset of pointer in indirect block
1606 * newblkno: disk block number being added
1607 * oldblkno: previous block number, 0 if none
1608 * nbp: buffer holding allocated page
984263bc
MD
1609 */
1610void
3fcb1ab8
SW
1611softdep_setup_allocindir_page(struct inode *ip, ufs_lbn_t lbn,
1612 struct buf *bp, int ptrno,
1613 ufs_daddr_t newblkno, ufs_daddr_t oldblkno,
1614 struct buf *nbp)
984263bc
MD
1615{
1616 struct allocindir *aip;
1617 struct pagedep *pagedep;
1618
1619 aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1620 ACQUIRE_LOCK(&lk);
1621 /*
1622 * If we are allocating a directory page, then we must
1623 * allocate an associated pagedep to track additions and
1624 * deletions.
1625 */
1626 if ((ip->i_mode & IFMT) == IFDIR &&
1627 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
408357d8
MD
1628 WORKLIST_INSERT_BP(nbp, &pagedep->pd_list);
1629 WORKLIST_INSERT_BP(nbp, &aip->ai_list);
984263bc
MD
1630 FREE_LOCK(&lk);
1631 setup_allocindir_phase2(bp, ip, aip);
1632}
1633
1634/*
1635 * Called just before setting an indirect block pointer to a
1636 * newly allocated indirect block.
3fcb1ab8
SW
1637 * Parameters:
1638 * nbp: newly allocated indirect block
1639 * ip: inode for file being extended
1640 * bp: indirect block referencing allocated block
1641 * ptrno: offset of pointer in indirect block
1642 * newblkno: disk block number being added
984263bc
MD
1643 */
1644void
3fcb1ab8
SW
1645softdep_setup_allocindir_meta(struct buf *nbp, struct inode *ip,
1646 struct buf *bp, int ptrno,
1647 ufs_daddr_t newblkno)
984263bc
MD
1648{
1649 struct allocindir *aip;
1650
1651 aip = newallocindir(ip, ptrno, newblkno, 0);
1652 ACQUIRE_LOCK(&lk);
408357d8 1653 WORKLIST_INSERT_BP(nbp, &aip->ai_list);
984263bc
MD
1654 FREE_LOCK(&lk);
1655 setup_allocindir_phase2(bp, ip, aip);
1656}
1657
1658/*
1659 * Called to finish the allocation of the "aip" allocated
1660 * by one of the two routines above.
3fcb1ab8
SW
1661 *
1662 * Parameters:
1663 * bp: in-memory copy of the indirect block
1664 * ip: inode for file being extended
1665 * aip: allocindir allocated by the above routines
984263bc
MD
1666 */
1667static void
3fcb1ab8
SW
1668setup_allocindir_phase2(struct buf *bp, struct inode *ip,
1669 struct allocindir *aip)
984263bc
MD
1670{
1671 struct worklist *wk;
1672 struct indirdep *indirdep, *newindirdep;
1673 struct bmsafemap *bmsafemap;
1674 struct allocindir *oldaip;
1675 struct freefrag *freefrag;
1676 struct newblk *newblk;
1677
54078292 1678 if (bp->b_loffset >= 0)
984263bc
MD
1679 panic("setup_allocindir_phase2: not indir blk");
1680 for (indirdep = NULL, newindirdep = NULL; ; ) {
1681 ACQUIRE_LOCK(&lk);
1682 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1683 if (wk->wk_type != D_INDIRDEP)
1684 continue;
1685 indirdep = WK_INDIRDEP(wk);
1686 break;
1687 }
1688 if (indirdep == NULL && newindirdep) {
1689 indirdep = newindirdep;
408357d8 1690 WORKLIST_INSERT_BP(bp, &indirdep->ir_list);
984263bc
MD
1691 newindirdep = NULL;
1692 }
1693 FREE_LOCK(&lk);
1694 if (indirdep) {
1695 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1696 &newblk) == 0)
1697 panic("setup_allocindir: lost block");
1698 ACQUIRE_LOCK(&lk);
1699 if (newblk->nb_state == DEPCOMPLETE) {
1700 aip->ai_state |= DEPCOMPLETE;
1701 aip->ai_buf = NULL;
1702 } else {
1703 bmsafemap = newblk->nb_bmsafemap;
1704 aip->ai_buf = bmsafemap->sm_buf;
1705 LIST_REMOVE(newblk, nb_deps);
1706 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1707 aip, ai_deps);
1708 }
1709 LIST_REMOVE(newblk, nb_hash);
1710 FREE(newblk, M_NEWBLK);
1711 aip->ai_indirdep = indirdep;
1712 /*
1713 * Check to see if there is an existing dependency
1714 * for this block. If there is, merge the old
1715 * dependency into the new one.
1716 */
1717 if (aip->ai_oldblkno == 0)
1718 oldaip = NULL;
1719 else
1720
1721 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1722 if (oldaip->ai_offset == aip->ai_offset)
1723 break;
1724 if (oldaip != NULL) {
1725 if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1726 FREE_LOCK(&lk);
1727 panic("setup_allocindir_phase2: blkno");
1728 }
1729 aip->ai_oldblkno = oldaip->ai_oldblkno;
1730 freefrag = oldaip->ai_freefrag;
1731 oldaip->ai_freefrag = aip->ai_freefrag;
1732 aip->ai_freefrag = freefrag;
1733 free_allocindir(oldaip, NULL);
1734 }
1735 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1736 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1737 [aip->ai_offset] = aip->ai_oldblkno;
1738 FREE_LOCK(&lk);
1739 }
1740 if (newindirdep) {
7d618503 1741 /*
2ae68842
MD
1742 * Avoid any possibility of data corruption by
1743 * ensuring that our old version is thrown away.
7d618503 1744 */
2ae68842
MD
1745 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
1746 brelse(newindirdep->ir_savebp);
984263bc
MD
1747 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1748 }
1749 if (indirdep)
1750 break;
1751 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1752 M_INDIRDEP, M_SOFTDEP_FLAGS);
1753 newindirdep->ir_list.wk_type = D_INDIRDEP;
1754 newindirdep->ir_state = ATTACHED;
1755 LIST_INIT(&newindirdep->ir_deplisthd);
1756 LIST_INIT(&newindirdep->ir_donehd);
54078292
MD
1757 if (bp->b_bio2.bio_offset == NOOFFSET) {
1758 VOP_BMAP(bp->b_vp, bp->b_bio1.bio_offset,
e92ca23a
MD
1759 &bp->b_bio2.bio_offset, NULL, NULL,
1760 BUF_CMD_WRITE);
984263bc 1761 }
54078292
MD
1762 KKASSERT(bp->b_bio2.bio_offset != NOOFFSET);
1763 newindirdep->ir_savebp = getblk(ip->i_devvp,
1764 bp->b_bio2.bio_offset,
1765 bp->b_bcount, 0, 0);
984263bc
MD
1766 BUF_KERNPROC(newindirdep->ir_savebp);
1767 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1768 }
1769}
1770
1771/*
1772 * Block de-allocation dependencies.
1773 *
1774 * When blocks are de-allocated, the on-disk pointers must be nullified before
1775 * the blocks are made available for use by other files. (The true
1776 * requirement is that old pointers must be nullified before new on-disk
1777 * pointers are set. We chose this slightly more stringent requirement to
1778 * reduce complexity.) Our implementation handles this dependency by updating
1779 * the inode (or indirect block) appropriately but delaying the actual block
1780 * de-allocation (i.e., freemap and free space count manipulation) until
1781 * after the updated versions reach stable storage. After the disk is
1782 * updated, the blocks can be safely de-allocated whenever it is convenient.
1783 * This implementation handles only the common case of reducing a file's
1784 * length to zero. Other cases are handled by the conventional synchronous
1785 * write approach.
1786 *
1787 * The ffs implementation with which we worked double-checks
1788 * the state of the block pointers and file size as it reduces
1789 * a file's length. Some of this code is replicated here in our
1790 * soft updates implementation. The freeblks->fb_chkcnt field is
1791 * used to transfer a part of this information to the procedure
1792 * that eventually de-allocates the blocks.
1793 *
1794 * This routine should be called from the routine that shortens
1795 * a file's length, before the inode's size or block pointers
1796 * are modified. It will save the block pointer information for
1797 * later release and zero the inode so that the calling routine
1798 * can release it.
1799 */
6bae6177
MD
1800struct softdep_setup_freeblocks_info {
1801 struct fs *fs;
1802 struct inode *ip;
1803};
1804
1805static int softdep_setup_freeblocks_bp(struct buf *bp, void *data);
1806
3fcb1ab8
SW
1807/*
1808 * Parameters:
1809 * ip: The inode whose length is to be reduced
1810 * length: The new length for the file
1811 */
984263bc 1812void
3fcb1ab8 1813softdep_setup_freeblocks(struct inode *ip, off_t length)
984263bc 1814{
6bae6177 1815 struct softdep_setup_freeblocks_info info;
984263bc
MD
1816 struct freeblks *freeblks;
1817 struct inodedep *inodedep;
1818 struct allocdirect *adp;
1819 struct vnode *vp;
1820 struct buf *bp;
1821 struct fs *fs;
1822 int i, error, delay;
6bae6177 1823 int count;
984263bc
MD
1824
1825 fs = ip->i_fs;
1826 if (length != 0)
1827 panic("softde_setup_freeblocks: non-zero length");
1828 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
e7b4468c 1829 M_FREEBLKS, M_SOFTDEP_FLAGS | M_ZERO);
984263bc 1830 freeblks->fb_list.wk_type = D_FREEBLKS;
89a5de29 1831 freeblks->fb_state = ATTACHED;
984263bc
MD
1832 freeblks->fb_uid = ip->i_uid;
1833 freeblks->fb_previousinum = ip->i_number;
1834 freeblks->fb_devvp = ip->i_devvp;
1835 freeblks->fb_fs = fs;
1836 freeblks->fb_oldsize = ip->i_size;
1837 freeblks->fb_newsize = length;
1838 freeblks->fb_chkcnt = ip->i_blocks;
1839 for (i = 0; i < NDADDR; i++) {
1840 freeblks->fb_dblks[i] = ip->i_db[i];
1841 ip->i_db[i] = 0;
1842 }
1843 for (i = 0; i < NIADDR; i++) {
1844 freeblks->fb_iblks[i] = ip->i_ib[i];
1845 ip->i_ib[i] = 0;
1846 }
1847 ip->i_blocks = 0;
1848 ip->i_size = 0;
1849 /*
1850 * Push the zero'ed inode to to its disk buffer so that we are free
1851 * to delete its dependencies below. Once the dependencies are gone
1852 * the buffer can be safely released.
1853 */
1854 if ((error = bread(ip->i_devvp,
54078292 1855 fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)),
3b568787 1856 (int)fs->fs_bsize, &bp)) != 0)
984263bc 1857 softdep_error("softdep_setup_freeblocks", error);
50e58362 1858 *((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
984263bc
MD
1859 ip->i_din;
1860 /*
1861 * Find and eliminate any inode dependencies.
1862 */
1863 ACQUIRE_LOCK(&lk);
1864 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1865 if ((inodedep->id_state & IOSTARTED) != 0) {
1866 FREE_LOCK(&lk);
1867 panic("softdep_setup_freeblocks: inode busy");
1868 }
1869 /*
1870 * Add the freeblks structure to the list of operations that
1871 * must await the zero'ed inode being written to disk. If we
1872 * still have a bitmap dependency (delay == 0), then the inode
1873 * has never been written to disk, so we can process the
1874 * freeblks below once we have deleted the dependencies.
1875 */
1876 delay = (inodedep->id_state & DEPCOMPLETE);
1877 if (delay)
1878 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1879 /*
1880 * Because the file length has been truncated to zero, any
1881 * pending block allocation dependency structures associated
1882 * with this inode are obsolete and can simply be de-allocated.
1883 * We must first merge the two dependency lists to get rid of
1884 * any duplicate freefrag structures, then purge the merged list.
1885 */
1886 merge_inode_lists(inodedep);
1887 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1888 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
1889 FREE_LOCK(&lk);
1890 bdwrite(bp);
1891 /*
1892 * We must wait for any I/O in progress to finish so that
1893 * all potential buffers on the dirty list will be visible.
1894 * Once they are all there, walk the list and get rid of
1895 * any dependencies.
1896 */
1897 vp = ITOV(ip);
1898 ACQUIRE_LOCK(&lk);
1899 drain_output(vp, 1);
6bae6177
MD
1900
1901 info.fs = fs;
1902 info.ip = ip;
3b998fa9 1903 lwkt_gettoken(&vp->v_token);
6bae6177
MD
1904 do {
1905 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
1906 softdep_setup_freeblocks_bp, &info);
2ef16e04 1907 } while (count != 0);
3b998fa9 1908 lwkt_reltoken(&vp->v_token);
0202303b 1909
984263bc
MD
1910 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
1911 (void)free_inodedep(inodedep);
89a5de29
MD
1912
1913 if (delay) {
1914 freeblks->fb_state |= DEPCOMPLETE;
1915 /*
1916 * If the inode with zeroed block pointers is now on disk
1917 * we can start freeing blocks. Add freeblks to the worklist
1918 * instead of calling handle_workitem_freeblocks directly as
1919 * it is more likely that additional IO is needed to complete
1920 * the request here than in the !delay case.
1921 */
1922 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
1923 add_to_worklist(&freeblks->fb_list);
1924 }
1925
984263bc
MD
1926 FREE_LOCK(&lk);
1927 /*
1928 * If the inode has never been written to disk (delay == 0),
1929 * then we can process the freeblks now that we have deleted
1930 * the dependencies.
1931 */
1932 if (!delay)
1933 handle_workitem_freeblocks(freeblks);
1934}
1935
6bae6177
MD
1936static int
1937softdep_setup_freeblocks_bp(struct buf *bp, void *data)
1938{
1939 struct softdep_setup_freeblocks_info *info = data;
1940 struct inodedep *inodedep;
1941
2ef16e04 1942 if (getdirtybuf(&bp, MNT_WAIT) == 0) {
086c1d7e 1943 kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp);
6bae6177 1944 return(-1);
2ef16e04
MD
1945 }
1946 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) {
086c1d7e 1947 kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp);
2ef16e04
MD
1948 BUF_UNLOCK(bp);
1949 return(-1);
1950 }
6bae6177
MD
1951 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep);
1952 deallocate_dependencies(bp, inodedep);
1953 bp->b_flags |= B_INVAL | B_NOCACHE;
1954 FREE_LOCK(&lk);
1955 brelse(bp);
1956 ACQUIRE_LOCK(&lk);
1957 return(1);
1958}
1959
984263bc
MD
1960/*
1961 * Reclaim any dependency structures from a buffer that is about to
1962 * be reallocated to a new vnode. The buffer must be locked, thus,
1963 * no I/O completion operations can occur while we are manipulating
1964 * its associated dependencies. The mutex is held so that other I/O's
1965 * associated with related dependencies do not occur.
1966 */
1967static void
3fcb1ab8 1968deallocate_dependencies(struct buf *bp, struct inodedep *inodedep)
984263bc
MD
1969{
1970 struct worklist *wk;
1971 struct indirdep *indirdep;
1972 struct allocindir *aip;
1973 struct pagedep *pagedep;
1974 struct dirrem *dirrem;
1975 struct diradd *dap;
1976 int i;
1977
1978 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1979 switch (wk->wk_type) {
1980
1981 case D_INDIRDEP:
1982 indirdep = WK_INDIRDEP(wk);
1983 /*
1984 * None of the indirect pointers will ever be visible,
1985 * so they can simply be tossed. GOINGAWAY ensures
1986 * that allocated pointers will be saved in the buffer
1987 * cache until they are freed. Note that they will
1988 * only be able to be found by their physical address
1989 * since the inode mapping the logical address will
1990 * be gone. The save buffer used for the safe copy
1991 * was allocated in setup_allocindir_phase2 using
1992 * the physical address so it could be used for this
1993 * purpose. Hence we swap the safe copy with the real
1994 * copy, allowing the safe copy to be freed and holding
1995 * on to the real copy for later use in indir_trunc.
81b5c339
MD
1996 *
1997 * NOTE: ir_savebp is relative to the block device
1998 * so b_bio1 contains the device block number.
984263bc
MD
1999 */
2000 if (indirdep->ir_state & GOINGAWAY) {
2001 FREE_LOCK(&lk);
2002 panic("deallocate_dependencies: already gone");
2003 }
2004 indirdep->ir_state |= GOINGAWAY;
2005 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
2006 free_allocindir(aip, inodedep);
54078292
MD
2007 if (bp->b_bio1.bio_offset >= 0 ||
2008 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) {
984263bc
MD
2009 FREE_LOCK(&lk);
2010 panic("deallocate_dependencies: not indir");
2011 }
2012 bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2013 bp->b_bcount);
2014 WORKLIST_REMOVE(wk);
408357d8 2015 WORKLIST_INSERT_BP(indirdep->ir_savebp, wk);
984263bc
MD
2016 continue;
2017
2018 case D_PAGEDEP:
2019 pagedep = WK_PAGEDEP(wk);
2020 /*
2021 * None of the directory additions will ever be
2022 * visible, so they can simply be tossed.
2023 */
2024 for (i = 0; i < DAHASHSZ; i++)
2025 while ((dap =
2026 LIST_FIRST(&pagedep->pd_diraddhd[i])))
2027 free_diradd(dap);
2028 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2029 free_diradd(dap);
2030 /*
2031 * Copy any directory remove dependencies to the list
2032 * to be processed after the zero'ed inode is written.
2033 * If the inode has already been written, then they
2034 * can be dumped directly onto the work list.
2035 */
2036 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2037 LIST_REMOVE(dirrem, dm_next);
2038 dirrem->dm_dirinum = pagedep->pd_ino;
2039 if (inodedep == NULL ||
2040 (inodedep->id_state & ALLCOMPLETE) ==
2041 ALLCOMPLETE)
2042 add_to_worklist(&dirrem->dm_list);
2043 else
2044 WORKLIST_INSERT(&inodedep->id_bufwait,
2045 &dirrem->dm_list);
2046 }
2047 WORKLIST_REMOVE(&pagedep->pd_list);
2048 LIST_REMOVE(pagedep, pd_hash);
2049 WORKITEM_FREE(pagedep, D_PAGEDEP);
2050 continue;
2051
2052 case D_ALLOCINDIR:
2053 free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2054 continue;
2055
2056 case D_ALLOCDIRECT:
2057 case D_INODEDEP:
2058 FREE_LOCK(&lk);
2059 panic("deallocate_dependencies: Unexpected type %s",
2060 TYPENAME(wk->wk_type));
2061 /* NOTREACHED */
2062
2063 default:
2064 FREE_LOCK(&lk);
2065 panic("deallocate_dependencies: Unknown type %s",
2066 TYPENAME(wk->wk_type));
2067 /* NOTREACHED */
2068 }
2069 }
2070}
2071
2072/*
2073 * Free an allocdirect. Generate a new freefrag work request if appropriate.
2074 * This routine must be called with splbio interrupts blocked.
2075 */
2076static void
3fcb1ab8
SW
2077free_allocdirect(struct allocdirectlst *adphead,
2078 struct allocdirect *adp, int delay)
984263bc
MD
2079{
2080
2081#ifdef DEBUG
dadab5e9 2082 if (lk.lkt_held == NOHOLDER)
984263bc
MD
2083 panic("free_allocdirect: lock not held");
2084#endif
2085 if ((adp->ad_state & DEPCOMPLETE) == 0)
2086 LIST_REMOVE(adp, ad_deps);
2087 TAILQ_REMOVE(adphead, adp, ad_next);
2088 if ((adp->ad_state & COMPLETE) == 0)
2089 WORKLIST_REMOVE(&adp->ad_list);
2090 if (adp->ad_freefrag != NULL) {
2091 if (delay)
2092 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2093 &adp->ad_freefrag->ff_list);
2094 else
2095 add_to_worklist(&adp->ad_freefrag->ff_list);
2096 }
2097 WORKITEM_FREE(adp, D_ALLOCDIRECT);
2098}
2099
2100/*
2101 * Prepare an inode to be freed. The actual free operation is not
2102 * done until the zero'ed inode has been written to disk.
2103 */
2104void
3fcb1ab8 2105softdep_freefile(struct vnode *pvp, ino_t ino, int mode)
984263bc
MD
2106{
2107 struct inode *ip = VTOI(pvp);
2108 struct inodedep *inodedep;
2109 struct freefile *freefile;
2110
2111 /*
2112 * This sets up the inode de-allocation dependency.
2113 */
2114 MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2115 M_FREEFILE, M_SOFTDEP_FLAGS);
2116 freefile->fx_list.wk_type = D_FREEFILE;
2117 freefile->fx_list.wk_state = 0;
2118 freefile->fx_mode = mode;
2119 freefile->fx_oldinum = ino;
2120 freefile->fx_devvp = ip->i_devvp;
2121 freefile->fx_fs = ip->i_fs;
2122
2123 /*
2124 * If the inodedep does not exist, then the zero'ed inode has
2125 * been written to disk. If the allocated inode has never been
2126 * written to disk, then the on-disk inode is zero'ed. In either
2127 * case we can free the file immediately.
2128 */
2129 ACQUIRE_LOCK(&lk);
2130 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2131 check_inode_unwritten(inodedep)) {
2132 FREE_LOCK(&lk);
2133 handle_workitem_freefile(freefile);
2134 return;
2135 }
2136 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2137 FREE_LOCK(&lk);
2138}
2139
2140/*
2141 * Check to see if an inode has never been written to disk. If
2142 * so free the inodedep and return success, otherwise return failure.
2143 * This routine must be called with splbio interrupts blocked.
2144 *
2145 * If we still have a bitmap dependency, then the inode has never
2146 * been written to disk. Drop the dependency as it is no longer
2147 * necessary since the inode is being deallocated. We set the
2148 * ALLCOMPLETE flags since the bitmap now properly shows that the
2149 * inode is not allocated. Even if the inode is actively being
2150 * written, it has been rolled back to its zero'ed state, so we
2151 * are ensured that a zero inode is what is on the disk. For short
2152 * lived files, this change will usually result in removing all the
2153 * dependencies from the inode so that it can be freed immediately.
2154 */
2155static int
3fcb1ab8 2156check_inode_unwritten(struct inodedep *inodedep)
984263bc
MD
2157{
2158
2159 if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2160 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2161 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2162 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2163 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2164 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2165 inodedep->id_nlinkdelta != 0)
2166 return (0);
83b29fff
MD
2167
2168 /*
2169 * Another process might be in initiate_write_inodeblock
2170 * trying to allocate memory without holding "Softdep Lock".
2171 */
2172 if ((inodedep->id_state & IOSTARTED) != 0 &&
2173 inodedep->id_savedino == NULL)
2174 return(0);
2175
984263bc
MD
2176 inodedep->id_state |= ALLCOMPLETE;
2177 LIST_REMOVE(inodedep, id_deps);
2178 inodedep->id_buf = NULL;
2179 if (inodedep->id_state & ONWORKLIST)
2180 WORKLIST_REMOVE(&inodedep->id_list);
2181 if (inodedep->id_savedino != NULL) {
2182 FREE(inodedep->id_savedino, M_INODEDEP);
2183 inodedep->id_savedino = NULL;
2184 }
2185 if (free_inodedep(inodedep) == 0) {
2186 FREE_LOCK(&lk);
2187 panic("check_inode_unwritten: busy inode");
2188 }
2189 return (1);
2190}
2191
2192/*
2193 * Try to free an inodedep structure. Return 1 if it could be freed.
2194 */
2195static int
3fcb1ab8 2196free_inodedep(struct inodedep *inodedep)
984263bc
MD
2197{
2198
2199 if ((inodedep->id_state & ONWORKLIST) != 0 ||
2200 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2201 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2202 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2203 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2204 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2205 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2206 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
2207 return (0);
2208 LIST_REMOVE(inodedep, id_hash);
2209 WORKITEM_FREE(inodedep, D_INODEDEP);
2210 num_inodedep -= 1;
2211 return (1);
2212}
2213
2214/*
2215 * This workitem routine performs the block de-allocation.
2216 * The workitem is added to the pending list after the updated
2217 * inode block has been written to disk. As mentioned above,
2218 * checks regarding the number of blocks de-allocated (compared
2219 * to the number of blocks allocated for the file) are also
2220 * performed in this function.
2221 */
2222static void
3fcb1ab8 2223handle_workitem_freeblocks(struct freeblks *freeblks)
984263bc
MD
2224{
2225 struct inode tip;
2226 ufs_daddr_t bn;
2227 struct fs *fs;
2228 int i, level, bsize;
2229 long nblocks, blocksreleased = 0;
2230 int error, allerror = 0;
2231 ufs_lbn_t baselbns[NIADDR], tmpval;
2232
2233 tip.i_number = freeblks->fb_previousinum;
2234 tip.i_devvp = freeblks->fb_devvp;
2235 tip.i_dev = freeblks->fb_devvp->v_rdev;
2236 tip.i_fs = freeblks->fb_fs;
2237 tip.i_size = freeblks->fb_oldsize;
2238 tip.i_uid = freeblks->fb_uid;
2239 fs = freeblks->fb_fs;
2240 tmpval = 1;
2241 baselbns[0] = NDADDR;
2242 for (i = 1; i < NIADDR; i++) {
2243 tmpval *= NINDIR(fs);
2244 baselbns[i] = baselbns[i - 1] + tmpval;
2245 }
2246 nblocks = btodb(fs->fs_bsize);
2247 blocksreleased = 0;
2248 /*
2249 * Indirect blocks first.
2250 */
2251 for (level = (NIADDR - 1); level >= 0; level--) {
2252 if ((bn = freeblks->fb_iblks[level]) == 0)
2253 continue;
54078292 2254 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level,
984263bc
MD
2255 baselbns[level], &blocksreleased)) == 0)
2256 allerror = error;
2257 ffs_blkfree(&tip, bn, fs->fs_bsize);
2258 blocksreleased += nblocks;
2259 }
2260 /*
2261 * All direct blocks or frags.
2262 */
2263 for (i = (NDADDR - 1); i >= 0; i--) {
2264 if ((bn = freeblks->fb_dblks[i]) == 0)
2265 continue;
2266 bsize = blksize(fs, &tip, i);
2267 ffs_blkfree(&tip, bn, bsize);
2268 blocksreleased += btodb(bsize);
2269 }
2270
2271#ifdef DIAGNOSTIC
2272 if (freeblks->fb_chkcnt != blocksreleased)
086c1d7e 2273 kprintf("handle_workitem_freeblocks: block count\n");
984263bc
MD
2274 if (allerror)
2275 softdep_error("handle_workitem_freeblks", allerror);
2276#endif /* DIAGNOSTIC */
2277 WORKITEM_FREE(freeblks, D_FREEBLKS);
2278}
2279
2280/*
2281 * Release blocks associated with the inode ip and stored in the indirect
54078292
MD
2282 * block at doffset. If level is greater than SINGLE, the block is an
2283 * indirect block and recursive calls to indirtrunc must be used to
2284 * cleanse other indirect blocks.
984263bc
MD
2285 */
2286static int
3fcb1ab8
SW
2287indir_trunc(struct inode *ip, off_t doffset, int level, ufs_lbn_t lbn,
2288 long *countp)
984263bc
MD
2289{
2290 struct buf *bp;
2291 ufs_daddr_t *bap;
2292 ufs_daddr_t nb;
2293 struct fs *fs;
2294 struct worklist *wk;
2295 struct indirdep *indirdep;
2296 int i, lbnadd, nblocks;
2297 int error, allerror = 0;
2298
2299 fs = ip->i_fs;
2300 lbnadd = 1;
2301 for (i = level; i > 0; i--)
2302 lbnadd *= NINDIR(fs);
2303 /*
2304 * Get buffer of block pointers to be freed. This routine is not
2305 * called until the zero'ed inode has been written, so it is safe
2306 * to free blocks as they are encountered. Because the inode has
2307 * been zero'ed, calls to bmap on these blocks will fail. So, we
2308 * have to use the on-disk address and the block device for the
2309 * filesystem to look them up. If the file was deleted before its
2310 * indirect blocks were all written to disk, the routine that set
2311 * us up (deallocate_dependencies) will have arranged to leave
2312 * a complete copy of the indirect block in memory for our use.
2313 * Otherwise we have to read the blocks in from the disk.
2314 */
2315 ACQUIRE_LOCK(&lk);
b1c20cfa 2316 if ((bp = findblk(ip->i_devvp, doffset, FINDBLK_TEST)) != NULL &&
984263bc 2317 (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1f1ea522
MD
2318 /*
2319 * bp must be ir_savebp, which is held locked for our use.
2320 */
984263bc
MD
2321 if (wk->wk_type != D_INDIRDEP ||
2322 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2323 (indirdep->ir_state & GOINGAWAY) == 0) {
2324 FREE_LOCK(&lk);
2325 panic("indir_trunc: lost indirdep");
2326 }
2327 WORKLIST_REMOVE(wk);
2328 WORKITEM_FREE(indirdep, D_INDIRDEP);
2329 if (LIST_FIRST(&bp->b_dep) != NULL) {
2330 FREE_LOCK(&lk);
2331 panic("indir_trunc: dangling dep");
2332 }
2333 FREE_LOCK(&lk);
2334 } else {
2335 FREE_LOCK(&lk);
54078292 2336 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp);
984263bc
MD
2337 if (error)
2338 return (error);
2339 }
2340 /*
2341 * Recursively free indirect blocks.
2342 */
2343 bap = (ufs_daddr_t *)bp->b_data;
2344 nblocks = btodb(fs->fs_bsize);
2345 for (i = NINDIR(fs) - 1; i >= 0; i--) {
2346 if ((nb = bap[i]) == 0)
2347 continue;
2348 if (level != 0) {
54078292 2349 if ((error = indir_trunc(ip, fsbtodoff(fs, nb),
984263bc
MD
2350 level - 1, lbn + (i * lbnadd), countp)) != 0)
2351 allerror = error;
2352 }
2353 ffs_blkfree(ip, nb, fs->fs_bsize);
2354 *countp += nblocks;
2355 }
2356 bp->b_flags |= B_INVAL | B_NOCACHE;
2357 brelse(bp);
2358 return (allerror);
2359}
2360
2361/*
2362 * Free an allocindir.
2363 * This routine must be called with splbio interrupts blocked.
2364 */
2365static void
3fcb1ab8 2366free_allocindir(struct allocindir *aip, struct inodedep *inodedep)
984263bc
MD
2367{
2368 struct freefrag *freefrag;
2369
2370#ifdef DEBUG
dadab5e9 2371 if (lk.lkt_held == NOHOLDER)
984263bc
MD
2372 panic("free_allocindir: lock not held");
2373#endif
2374 if ((aip->ai_state & DEPCOMPLETE) == 0)
2375 LIST_REMOVE(aip, ai_deps);
2376 if (aip->ai_state & ONWORKLIST)
2377 WORKLIST_REMOVE(&aip->ai_list);
2378 LIST_REMOVE(aip, ai_next);
2379 if ((freefrag = aip->ai_freefrag) != NULL) {
2380 if (inodedep == NULL)
2381 add_to_worklist(&freefrag->ff_list);
2382 else
2383 WORKLIST_INSERT(&inodedep->id_bufwait,
2384 &freefrag->ff_list);
2385 }
2386 WORKITEM_FREE(aip, D_ALLOCINDIR);
2387}
2388
2389/*
2390 * Directory entry addition dependencies.
2391 *
2392 * When adding a new directory entry, the inode (with its incremented link
2393 * count) must be written to disk before the directory entry's pointer to it.
2394 * Also, if the inode is newly allocated, the corresponding freemap must be
2395 * updated (on disk) before the directory entry's pointer. These requirements
2396 * are met via undo/redo on the directory entry's pointer, which consists
2397 * simply of the inode number.
2398 *
2399 * As directory entries are added and deleted, the free space within a
f719c866 2400 * directory block can become fragmented. The ufs filesystem will compact
984263bc
MD
2401 * a fragmented directory block to make space for a new entry. When this
2402 * occurs, the offsets of previously added entries change. Any "diradd"
2403 * dependency structures corresponding to these entries must be updated with
2404 * the new offsets.
2405 */
2406
2407/*
2408 * This routine is called after the in-memory inode's link
2409 * count has been incremented, but before the directory entry's
2410 * pointer to the inode has been set.
3fcb1ab8
SW
2411 *
2412 * Parameters:
2413 * bp: buffer containing directory block
2414 * dp: inode for directory
2415 * diroffset: offset of new entry in directory
2416 * newinum: inode referenced by new directory entry
2417 * newdirbp: non-NULL => contents of new mkdir
984263bc
MD
2418 */
2419void
3fcb1ab8
SW
2420softdep_setup_directory_add(struct buf *bp, struct inode *dp, off_t diroffset,
2421 ino_t newinum, struct buf *newdirbp)
984263bc
MD
2422{
2423 int offset; /* offset of new entry within directory block */
2424 ufs_lbn_t lbn; /* block in directory containing new entry */
2425 struct fs *fs;
2426 struct diradd *dap;
2427 struct pagedep *pagedep;
2428 struct inodedep *inodedep;
2429 struct mkdir *mkdir1, *mkdir2;
2430
2431 /*
2432 * Whiteouts have no dependencies.
2433 */
2434 if (newinum == WINO) {
2435 if (newdirbp != NULL)
2436 bdwrite(newdirbp);
2437 return;
2438 }
2439
2440 fs = dp->i_fs;
2441 lbn = lblkno(fs, diroffset);
2442 offset = blkoff(fs, diroffset);
2443 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
e7b4468c 2444 M_SOFTDEP_FLAGS | M_ZERO);
984263bc
MD
2445 dap->da_list.wk_type = D_DIRADD;
2446 dap->da_offset = offset;
2447 dap->da_newinum = newinum;
2448 dap->da_state = ATTACHED;
2449 if (newdirbp == NULL) {
2450 dap->da_state |= DEPCOMPLETE;
2451 ACQUIRE_LOCK(&lk);
2452 } else {
2453 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2454 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2455 M_SOFTDEP_FLAGS);
2456 mkdir1->md_list.wk_type = D_MKDIR;
2457 mkdir1->md_state = MKDIR_BODY;
2458 mkdir1->md_diradd = dap;
2459 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2460 M_SOFTDEP_FLAGS);
2461 mkdir2->md_list.wk_type = D_MKDIR;
2462 mkdir2->md_state = MKDIR_PARENT;
2463 mkdir2->md_diradd = dap;
2464 /*
2465 * Dependency on "." and ".." being written to disk.
2466 */
2467 mkdir1->md_buf = newdirbp;
2468 ACQUIRE_LOCK(&lk);
2469 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
408357d8 2470 WORKLIST_INSERT_BP(newdirbp, &mkdir1->md_list);
984263bc
MD
2471 FREE_LOCK(&lk);
2472 bdwrite(newdirbp);
2473 /*
2474 * Dependency on link count increase for parent directory
2475 */
2476 ACQUIRE_LOCK(&lk);
2477 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2478 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2479 dap->da_state &= ~MKDIR_PARENT;
2480 WORKITEM_FREE(mkdir2, D_MKDIR);
2481 } else {
2482 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2483 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2484 }
2485 }
2486 /*
2487 * Link into parent directory pagedep to await its being written.
2488 */
2489 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
408357d8 2490 WORKLIST_INSERT_BP(bp, &pagedep->pd_list);
984263bc
MD
2491 dap->da_pagedep = pagedep;
2492 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2493 da_pdlist);
2494 /*
2495 * Link into its inodedep. Put it on the id_bufwait list if the inode
2496 * is not yet written. If it is written, do the post-inode write
2497 * processing to put it on the id_pendinghd list.
2498 */
2499 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2500 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2501 diradd_inode_written(dap, inodedep);
2502 else
2503 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2504 FREE_LOCK(&lk);
2505}
2506
2507/*
2508 * This procedure is called to change the offset of a directory
2509 * entry when compacting a directory block which must be owned
2510 * exclusively by the caller. Note that the actual entry movement
2511 * must be done in this procedure to ensure that no I/O completions
2512 * occur while the move is in progress.
3fcb1ab8
SW
2513 *
2514 * Parameters:
2515 * dp: inode for directory
2516 * base: address of dp->i_offset
2517 * oldloc: address of old directory location
2518 * newloc: address of new directory location
2519 * entrysize: size of directory entry
984263bc
MD
2520 */
2521void
3fcb1ab8
SW
2522softdep_change_directoryentry_offset(struct inode *dp, caddr_t base,
2523 caddr_t oldloc, caddr_t newloc,
2524 int entrysize)
984263bc
MD
2525{
2526 int offset, oldoffset, newoffset;
2527 struct pagedep *pagedep;
2528 struct diradd *dap;
2529 ufs_lbn_t lbn;
2530
2531 ACQUIRE_LOCK(&lk);
2532 lbn = lblkno(dp->i_fs, dp->i_offset);
2533 offset = blkoff(dp->i_fs, dp->i_offset);
2534 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2535 goto done;
2536 oldoffset = offset + (oldloc - base);
2537 newoffset = offset + (newloc - base);
2538
2539 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2540 if (dap->da_offset != oldoffset)
2541 continue;
2542 dap->da_offset = newoffset;
2543 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2544 break;
2545 LIST_REMOVE(dap, da_pdlist);
2546 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2547 dap, da_pdlist);
2548 break;
2549 }
2550 if (dap == NULL) {
2551
2552 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2553 if (dap->da_offset == oldoffset) {
2554 dap->da_offset = newoffset;
2555 break;
2556 }
2557 }
2558 }
2559done:
2560 bcopy(oldloc, newloc, entrysize);
2561 FREE_LOCK(&lk);
2562}
2563
2564/*
2565 * Free a diradd dependency structure. This routine must be called
2566 * with splbio interrupts blocked.
2567 */
2568static void
3fcb1ab8 2569free_diradd(struct diradd *dap)
984263bc
MD
2570{
2571 struct dirrem *dirrem;
2572 struct pagedep *pagedep;
2573 struct inodedep *inodedep;
2574 struct mkdir *mkdir, *nextmd;
2575
2576#ifdef DEBUG
dadab5e9 2577 if (lk.lkt_held == NOHOLDER)
984263bc
MD
2578 panic("free_diradd: lock not held");
2579#endif
2580 WORKLIST_REMOVE(&dap->da_list);
2581 LIST_REMOVE(dap, da_pdlist);
2582 if ((dap->da_state & DIRCHG) == 0) {
2583 pagedep = dap->da_pagedep;
2584 } else {
2585 dirrem = dap->da_previous;
2586 pagedep = dirrem->dm_pagedep;
2587 dirrem->dm_dirinum = pagedep->pd_ino;
2588 add_to_worklist(&dirrem->dm_list);
2589 }
2590 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2591 0, &inodedep) != 0)
2592 (void) free_inodedep(inodedep);
2593 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2594 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2595 nextmd = LIST_NEXT(mkdir, md_mkdirs);
2596 if (mkdir->md_diradd != dap)
2597 continue;
2598 dap->da_state &= ~mkdir->md_state;
2599 WORKLIST_REMOVE(&mkdir->md_list);
2600 LIST_REMOVE(mkdir, md_mkdirs);
2601 WORKITEM_FREE(mkdir, D_MKDIR);
2602 }
2603 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2604 FREE_LOCK(&lk);
2605 panic("free_diradd: unfound ref");
2606 }
2607 }
2608 WORKITEM_FREE(dap, D_DIRADD);
2609}
2610
2611/*
2612 * Directory entry removal dependencies.
2613 *
2614 * When removing a directory entry, the entry's inode pointer must be
2615 * zero'ed on disk before the corresponding inode's link count is decremented
2616 * (possibly freeing the inode for re-use). This dependency is handled by
2617 * updating the directory entry but delaying the inode count reduction until
2618 * after the directory block has been written to disk. After this point, the
2619 * inode count can be decremented whenever it is convenient.
2620 */
2621
2622/*
2623 * This routine should be called immediately after removing
2624 * a directory entry. The inode's link count should not be
2625 * decremented by the calling procedure -- the soft updates
2626 * code will do this task when it is safe.
3fcb1ab8
SW
2627 *
2628 * Parameters:
2629 * bp: buffer containing directory block
2630 * dp: inode for the directory being modified
2631 * ip: inode for directory entry being removed
2632 * isrmdir: indicates if doing RMDIR
984263bc
MD
2633 */
2634void
3fcb1ab8
SW
2635softdep_setup_remove(struct buf *bp, struct inode *dp, struct inode *ip,
2636 int isrmdir)
984263bc
MD
2637{
2638 struct dirrem *dirrem, *prevdirrem;
2639
2640 /*
2641 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2642 */
2643 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2644
2645 /*
2646 * If the COMPLETE flag is clear, then there were no active
2647 * entries and we want to roll back to a zeroed entry until
2648 * the new inode is committed to disk. If the COMPLETE flag is
2649 * set then we have deleted an entry that never made it to
2650 * disk. If the entry we deleted resulted from a name change,
2651 * then the old name still resides on disk. We cannot delete
2652 * its inode (returned to us in prevdirrem) until the zeroed
2653 * directory entry gets to disk. The new inode has never been
2654 * referenced on the disk, so can be deleted immediately.
2655 */
2656 if ((dirrem->dm_state & COMPLETE) == 0) {
2657 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2658 dm_next);
2659 FREE_LOCK(&lk);
2660 } else {
2661 if (prevdirrem != NULL)
2662 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2663 prevdirrem, dm_next);
2664 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2665 FREE_LOCK(&lk);
2666 handle_workitem_remove(dirrem);
2667 }
2668}
2669
2670/*
2671 * Allocate a new dirrem if appropriate and return it along with
2672 * its associated pagedep. Called without a lock, returns with lock.
2673 */
2674static long num_dirrem; /* number of dirrem allocated */
3fcb1ab8
SW
2675
2676/*
2677 * Parameters:
2678 * bp: buffer containing directory block
2679 * dp: inode for the directory being modified
2680 * ip: inode for directory entry being removed
2681 * isrmdir: indicates if doing RMDIR
2682 * prevdirremp: previously referenced inode, if any
2683 */
984263bc 2684static struct dirrem *
3fcb1ab8
SW
2685newdirrem(struct buf *bp, struct inode *dp, struct inode *ip,
2686 int isrmdir, struct dirrem **prevdirremp)
984263bc
MD
2687{
2688 int offset;
2689 ufs_lbn_t lbn;
2690 struct diradd *dap;
2691 struct dirrem *dirrem;
2692 struct pagedep *pagedep;
2693
2694 /*
2695 * Whiteouts have no deletion dependencies.
2696 */
2697 if (ip == NULL)
2698 panic("newdirrem: whiteout");
2699 /*
2700 * If we are over our limit, try to improve the situation.
2701 * Limiting the number of dirrem structures will also limit
2702 * the number of freefile and freeblks structures.
2703 */
2704 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2705 (void) request_cleanup(FLUSH_REMOVE, 0);
2706 num_dirrem += 1;
2707 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
e7b4468c 2708 M_DIRREM, M_SOFTDEP_FLAGS | M_ZERO);
984263bc
MD
2709 dirrem->dm_list.wk_type = D_DIRREM;
2710 dirrem->dm_state = isrmdir ? RMDIR : 0;
2711 dirrem->dm_mnt = ITOV(ip)->v_mount;
2712 dirrem->dm_oldinum = ip->i_number;
2713 *prevdirremp = NULL;
2714
2715 ACQUIRE_LOCK(&lk);
2716 lbn = lblkno(dp->i_fs, dp->i_offset);
2717 offset = blkoff(dp->i_fs, dp->i_offset);
2718 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
408357d8 2719 WORKLIST_INSERT_BP(bp, &pagedep->pd_list);
984263bc
MD
2720 dirrem->dm_pagedep = pagedep;
2721 /*
2722 * Check for a diradd dependency for the same directory entry.
2723 * If present, then both dependencies become obsolete and can
2724 * be de-allocated. Check for an entry on both the pd_dirraddhd
2725 * list and the pd_pendinghd list.
2726 */
2727
2728 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
2729 if (dap->da_offset == offset)
2730 break;
2731 if (dap == NULL) {
2732
2733 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
2734 if (dap->da_offset == offset)
2735 break;
2736 if (dap == NULL)
2737 return (dirrem);
2738 }
2739 /*
2740 * Must be ATTACHED at this point.
2741 */
2742 if ((dap->da_state & ATTACHED) == 0) {
2743 FREE_LOCK(&lk);
2744 panic("newdirrem: not ATTACHED");
2745 }
2746 if (dap->da_newinum != ip->i_number) {
2747 FREE_LOCK(&lk);
f91a71dd 2748 panic("newdirrem: inum %"PRId64" should be %"PRId64,
984263bc
MD
2749 ip->i_number, dap->da_newinum);
2750 }
2751 /*
2752 * If we are deleting a changed name that never made it to disk,
2753 * then return the dirrem describing the previous inode (which
2754 * represents the inode currently referenced from this entry on disk).
2755 */
2756 if ((dap->da_state & DIRCHG) != 0) {
2757 *prevdirremp = dap->da_previous;
2758 dap->da_state &= ~DIRCHG;
2759 dap->da_pagedep = pagedep;
2760 }
2761 /*
2762 * We are deleting an entry that never made it to disk.
2763 * Mark it COMPLETE so we can delete its inode immediately.
2764 */
2765 dirrem->dm_state |= COMPLETE;
2766 free_diradd(dap);
2767 return (dirrem);
2768}
2769
2770/*
2771 * Directory entry change dependencies.
2772 *
2773 * Changing an existing directory entry requires that an add operation
2774 * be completed first followed by a deletion. The semantics for the addition
2775 * are identical to the description of adding a new entry above except
2776 * that the rollback is to the old inode number rather than zero. Once
2777 * the addition dependency is completed, the removal is done as described
2778 * in the removal routine above.
2779 */
2780
2781/*
2782 * This routine should be called immediately after changing
2783 * a directory entry. The inode's link count should not be
2784 * decremented by the calling procedure -- the soft updates
2785 * code will perform this task when it is safe.
3fcb1ab8
SW
2786 *
2787 * Parameters:
2788 * bp: buffer containing directory block
2789 * dp: inode for the directory being modified
2790 * ip: inode for directory entry being removed
2791 * newinum: new inode number for changed entry
2792 * isrmdir: indicates if doing RMDIR
984263bc
MD
2793 */
2794void
3fcb1ab8
SW
2795softdep_setup_directory_change(struct buf *bp, struct inode *dp,
2796 struct inode *ip, ino_t newinum,
2797 int isrmdir)
984263bc
MD
2798{
2799 int offset;
2800 struct diradd *dap = NULL;
2801 struct dirrem *dirrem, *prevdirrem;
2802 struct pagedep *pagedep;
2803 struct inodedep *inodedep;
2804
2805 offset = blkoff(dp->i_fs, dp->i_offset);
2806
2807 /*
2808 * Whiteouts do not need diradd dependencies.
2809 */
2810 if (newinum != WINO) {
2811 MALLOC(dap, struct diradd *, sizeof(struct diradd),
e7b4468c 2812 M_DIRADD, M_SOFTDEP_FLAGS | M_ZERO);
984263bc
MD
2813 dap->da_list.wk_type = D_DIRADD;
2814 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2815 dap->da_offset = offset;
2816 dap->da_newinum = newinum;
2817 }
2818
2819 /*
2820 * Allocate a new dirrem and ACQUIRE_LOCK.
2821 */
2822 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2823 pagedep = dirrem->dm_pagedep;
2824 /*
2825 * The possible values for isrmdir:
2826 * 0 - non-directory file rename
2827 * 1 - directory rename within same directory
2828 * inum - directory rename to new directory of given inode number
2829 * When renaming to a new directory, we are both deleting and
2830 * creating a new directory entry, so the link count on the new
2831 * directory should not change. Thus we do not need the followup
2832 * dirrem which is usually done in handle_workitem_remove. We set
2833 * the DIRCHG flag to tell handle_workitem_remove to skip the
2834 * followup dirrem.
2835 */
2836 if (isrmdir > 1)
2837 dirrem->dm_state |= DIRCHG;
2838
2839 /*
2840 * Whiteouts have no additional dependencies,
2841 * so just put the dirrem on the correct list.
2842 */
2843 if (newinum == WINO) {
2844 if ((dirrem->dm_state & COMPLETE) == 0) {
2845 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2846 dm_next);
2847 } else {
2848 dirrem->dm_dirinum = pagedep->pd_ino;
2849 add_to_worklist(&dirrem->dm_list);
2850 }
2851 FREE_LOCK(&lk);
2852 return;
2853 }
2854
2855 /*
2856 * If the COMPLETE flag is clear, then there were no active
2857 * entries and we want to roll back to the previous inode until
2858 * the new inode is committed to disk. If the COMPLETE flag is
2859 * set, then we have deleted an entry that never made it to disk.
2860 * If the entry we deleted resulted from a name change, then the old
2861 * inode reference still resides on disk. Any rollback that we do
2862 * needs to be to that old inode (returned to us in prevdirrem). If
2863 * the entry we deleted resulted from a create, then there is
2864 * no entry on the disk, so we want to roll back to zero rather
2865 * than the uncommitted inode. In either of the COMPLETE cases we
2866 * want to immediately free the unwritten and unreferenced inode.
2867 */
2868 if ((dirrem->dm_state & COMPLETE) == 0) {
2869 dap->da_previous = dirrem;
2870 } else {
2871 if (prevdirrem != NULL) {
2872 dap->da_previous = prevdirrem;
2873 } else {
2874 dap->da_state &= ~DIRCHG;
2875 dap->da_pagedep = pagedep;
2876 }
2877 dirrem->dm_dirinum = pagedep->pd_ino;
2878 add_to_worklist(&dirrem->dm_list);
2879 }
2880 /*
2881 * Link into its inodedep. Put it on the id_bufwait list if the inode
2882 * is not yet written. If it is written, do the post-inode write
2883 * processing to put it on the id_pendinghd list.
2884 */
2885 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2886 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2887 dap->da_state |= COMPLETE;
2888 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2889 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2890 } else {
2891 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2892 dap, da_pdlist);
2893 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2894 }
2895 FREE_LOCK(&lk);
2896}
2897
2898/*
2899 * Called whenever the link count on an inode is changed.
2900 * It creates an inode dependency so that the new reference(s)
2901 * to the inode cannot be committed to disk until the updated
2902 * inode has been written.
3fcb1ab8
SW
2903 *
2904 * Parameters:
2905 * ip: the inode with the increased link count
984263bc
MD
2906 */
2907void
3fcb1ab8 2908softdep_change_linkcnt(struct inode *ip)
984263bc
MD
2909{
2910 struct inodedep *inodedep;
2911
2912 ACQUIRE_LOCK(&lk);
2913 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2914 if (ip->i_nlink < ip->i_effnlink) {
2915 FREE_LOCK(&lk);
2916 panic("softdep_change_linkcnt: bad delta");
2917 }
2918 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2919 FREE_LOCK(&lk);
2920}
2921
2922/*
2923 * This workitem decrements the inode's link count.
2924 * If the link count reaches zero, the file is removed.
2925 */
2926static void
3fcb1ab8 2927handle_workitem_remove(struct dirrem *dirrem)
984263bc 2928{
984263bc
MD
2929 struct inodedep *inodedep;
2930 struct vnode *vp;
2931 struct inode *ip;
2932 ino_t oldinum;
2933 int error;
2934
b9b0a6d0
MD
2935 error = VFS_VGET(dirrem->dm_mnt, NULL, dirrem->dm_oldinum, &vp);
2936 if (error) {
984263bc
MD
2937 softdep_error("handle_workitem_remove: vget", error);
2938 return;
2939 }
2940 ip = VTOI(vp);
2941 ACQUIRE_LOCK(&lk);
2942 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
2943 FREE_LOCK(&lk);
2944 panic("handle_workitem_remove: lost inodedep");
2945 }
2946 /*
2947 * Normal file deletion.
2948 */
2949 if ((dirrem->dm_state & RMDIR) == 0) {
2950 ip->i_nlink--;
2951 ip->i_flag |= IN_CHANGE;
2952 if (ip->i_nlink < ip->i_effnlink) {
2953 FREE_LOCK(&lk);
2954 panic("handle_workitem_remove: bad file delta");
2955 }
2956 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2957 FREE_LOCK(&lk);
2958 vput(vp);
2959 num_dirrem -= 1;
2960 WORKITEM_FREE(dirrem, D_DIRREM);
2961 return;
2962 }
2963 /*
2964 * Directory deletion. Decrement reference count for both the
2965 * just deleted parent directory entry and the reference for ".".
2966 * Next truncate the directory to length zero. When the
2967 * truncation completes, arrange to have the reference count on
2968 * the parent decremented to account for the loss of "..".
2969 */
2970 ip->i_nlink -= 2;
2971 ip->i_flag |= IN_CHANGE;
2972 if (ip->i_nlink < ip->i_effnlink) {
2973 FREE_LOCK(&lk);
2974 panic("handle_workitem_remove: bad dir delta");
2975 }
2976 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2977 FREE_LOCK(&lk);
ac690a1d 2978 if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0)
984263bc
MD
2979 softdep_error("handle_workitem_remove: truncate", error);
2980 /*
2981 * Rename a directory to a new parent. Since, we are both deleting
2982 * and creating a new directory entry, the link count on the new
2983 * directory should not change. Thus we skip the followup dirrem.
2984 */
2985 if (dirrem->dm_state & DIRCHG) {
2986 vput(vp);
2987 num_dirrem -= 1;
2988 WORKITEM_FREE(dirrem, D_DIRREM);
2989 return;
2990 }
2991 /*
2992 * If the inodedep does not exist, then the zero'ed inode has
2993 * been written to disk. If the allocated inode has never been
2994 * written to disk, then the on-disk inode is zero'ed. In either
2995 * case we can remove the file immediately.
2996 */
2997 ACQUIRE_LOCK(&lk);
2998 dirrem->dm_state = 0;
2999 oldinum = dirrem->dm_oldinum;
3000 dirrem->dm_oldinum = dirrem->dm_dirinum;
3001 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
3002 check_inode_unwritten(inodedep)) {
3003 FREE_LOCK(&lk);
3004 vput(vp);
3005 handle_workitem_remove(dirrem);
3006 return;
3007 }
3008 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3009 FREE_LOCK(&lk);
a0ff2e99
MD
3010 ip->i_flag |= IN_CHANGE;
3011 ffs_update(vp, 0);
984263bc
MD
3012 vput(vp);
3013}
3014
3015/*
3016 * Inode de-allocation dependencies.
3017 *
3018 * When an inode's link count is reduced to zero, it can be de-allocated. We
3019 * found it convenient to postpone de-allocation until after the inode is
3020 * written to disk with its new link count (zero). At this point, all of the
3021 * on-disk inode's block pointers are nullified and, with careful dependency
3022 * list ordering, all dependencies related to the inode will be satisfied and
3023 * the corresponding dependency structures de-allocated. So, if/when the
3024 * inode is reused, there will be no mixing of old dependencies with new
3025 * ones. This artificial dependency is set up by the block de-allocation
3026 * procedure above (softdep_setup_freeblocks) and completed by the
3027 * following procedure.
3028 */
3029static void
3fcb1ab8 3030handle_workitem_freefile(struct freefile *freefile)
984263bc
MD
3031{
3032 struct vnode vp;
3033 struct inode tip;
3034 struct inodedep *idp;
3035 int error;
3036
3037#ifdef DEBUG
3038 ACQUIRE_LOCK(&lk);
3039 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp);
3040 FREE_LOCK(&lk);
3041 if (error)
3042 panic("handle_workitem_freefile: inodedep survived");
3043#endif
3044 tip.i_devvp = freefile->fx_devvp;
3045 tip.i_dev = freefile->fx_devvp->v_rdev;
3046 tip.i_fs = freefile->fx_fs;
3047 vp.v_data = &tip;
3048 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
3049 softdep_error("handle_workitem_freefile", error);
3050 WORKITEM_FREE(freefile, D_FREEFILE);
3051}
3052
3053/*
83b29fff
MD
3054 * Helper function which unlinks marker element from work list and returns
3055 * the next element on the list.
3056 */
3057static __inline struct worklist *
3058markernext(struct worklist *marker)
3059{
3060 struct worklist *next;
3061
3062 next = LIST_NEXT(marker, wk_list);
3063 LIST_REMOVE(marker, wk_list);
3064 return next;
3065}
3066
3067/*
27bc0cb1
MD
3068 * checkread, checkwrite
3069 *
3070 */
3071static int
3072softdep_checkread(struct buf *bp)
3073{
3074 return(0);
3075}
3076
3077static int
3078softdep_checkwrite(struct buf *bp)
3079{
3080 return(0);
3081}
3082
3083/*
984263bc
MD
3084 * Disk writes.
3085 *
3086 * The dependency structures constructed above are most actively used when file
3087 * system blocks are written to disk. No constraints are placed on when a
3088 * block can be written, but unsatisfied update dependencies are made safe by
3089 * modifying (or replacing) the source memory for the duration of the disk
3090 * write. When the disk write completes, the memory block is again brought
3091 * up-to-date.
3092 *
3093 * In-core inode structure reclamation.
3094 *
3095 * Because there are a finite number of "in-core" inode structures, they are
3096 * reused regularly. By transferring all inode-related dependencies to the
3097 * in-memory inode block and indexing them separately (via "inodedep"s), we
3098 * can allow "in-core" inode structures to be reused at any time and avoid
3099 * any increase in contention.
3100 *
3101 * Called just before entering the device driver to initiate a new disk I/O.
3102 * The buffer must be locked, thus, no I/O completion operations can occur
3103 * while we are manipulating its associated dependencies.
3fcb1ab8
SW
3104 *
3105 * Parameters:
3106 * bp: structure describing disk write to occur
984263bc
MD
3107 */
3108static void
3fcb1ab8 3109softdep_disk_io_initiation(struct buf *bp)
984263bc 3110{
83b29fff
MD
3111 struct worklist *wk;
3112 struct worklist marker;
984263bc
MD
3113 struct indirdep *indirdep;
3114
3115 /*
3116 * We only care about write operations. There should never
3117 * be dependencies for reads.
3118 */
10f3fee5 3119 if (bp->b_cmd == BUF_CMD_READ)
984263bc 3120 panic("softdep_disk_io_initiation: read");
83b29fff
MD
3121
3122 marker.wk_type = D_LAST + 1; /* Not a normal workitem */
3123
984263bc
MD
3124 /*
3125 * Do any necessary pre-I/O processing.
3126 */
83b29fff
MD
3127 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) {
3128 LIST_INSERT_AFTER(wk, &marker, wk_list);
3129
984263bc
MD
3130 switch (wk->wk_type) {
3131
3132 case D_PAGEDEP:
3133 initiate_write_filepage(WK_PAGEDEP(wk), bp);
3134 continue;
3135
3136 case D_INODEDEP:
3137 initiate_write_inodeblock(WK_INODEDEP(wk), bp);
3138 continue;
3139
3140 case D_INDIRDEP:
3141 indirdep = WK_INDIRDEP(wk);
3142 if (indirdep->ir_state & GOINGAWAY)
3143 panic("disk_io_initiation: indirdep gone");
3144 /*
3145 * If there are no remaining dependencies, this
3146 * will be writing the real pointers, so the
3147 * dependency can be freed.
3148 */
3149 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
3150 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
3151 brelse(indirdep->ir_savebp);
3152 /* inline expand WORKLIST_REMOVE(wk); */
3153 wk->wk_state &= ~ONWORKLIST;
3154 LIST_REMOVE(wk, wk_list);
3155 WORKITEM_FREE(indirdep, D_INDIRDEP);
3156 continue;
3157 }
3158 /*
3159 * Replace up-to-date version with safe version.
3160 */
3161 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3162 M_INDIRDEP, M_SOFTDEP_FLAGS);
3163 ACQUIRE_LOCK(&lk);
3164 indirdep->ir_state &= ~ATTACHED;
3165 indirdep->ir_state |= UNDONE;
3166 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3167 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3168 bp->b_bcount);
3169 FREE_LOCK(&lk);
3170 continue;
3171
3172 case D_MKDIR:
3173 case D_BMSAFEMAP:
3174 case D_ALLOCDIRECT:
3175 case D_ALLOCINDIR:
3176 continue;
3177
3178 default:
3179 panic("handle_disk_io_initiation: Unexpected type %s",
3180 TYPENAME(wk->wk_type));
3181 /* NOTREACHED */
3182 }
3183 }
3184}
3185
3186/*
3187 * Called from within the procedure above to deal with unsatisfied
3188 * allocation dependencies in a directory. The buffer must be locked,
3189 * thus, no I/O completion operations can occur while we are
3190 * manipulating its associated dependencies.
3191 */
3192static void
3fcb1ab8 3193initiate_write_filepage(struct pagedep *pagedep, struct buf *bp)
984263bc
MD
3194{
3195 struct diradd *dap;
3196 struct direct *ep;
3197 int i;
3198
3199 if (pagedep->pd_state & IOSTARTED) {
3200 /*
3201 * This can only happen if there is a driver that does not
3202 * understand chaining. Here biodone will reissue the call
3203 * to strategy for the incomplete buffers.
3204 */
086c1d7e 3205 kprintf("initiate_write_filepage: already started\n");
984263bc
MD
3206 return;
3207 }
3208 pagedep->pd_state |= IOSTARTED;
3209 ACQUIRE_LOCK(&lk);
3210 for (i = 0; i < DAHASHSZ; i++) {
3211 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3212 ep = (struct direct *)
3213 ((char *)bp->b_data + dap->da_offset);
3214 if (ep->d_ino != dap->da_newinum) {
3215 FREE_LOCK(&lk);
f91a71dd 3216 panic("%s: dir inum %d != new %"PRId64,
984263bc
MD
3217 "initiate_write_filepage",
3218 ep->d_ino, dap->da_newinum);
3219 }
3220 if (dap->da_state & DIRCHG)
3221 ep->d_ino = dap->da_previous->dm_oldinum;
3222 else
3223 ep->d_ino = 0;
3224 dap->da_state &= ~ATTACHED;
3225 dap->da_state |= UNDONE;
3226 }
3227 }
3228 FREE_LOCK(&lk);
3229}
3230
3231/*
3232 * Called from within the procedure above to deal with unsatisfied
3233 * allocation dependencies in an inodeblock. The buffer must be
3234 * locked, thus, no I/O completion operations can occur while we
3235 * are manipulating its associated dependencies.
3fcb1ab8
SW
3236 *
3237 * Parameters:
3238 * bp: The inode block
984263bc
MD
3239 */
3240static void
3fcb1ab8 3241initiate_write_inodeblock(struct inodedep *inodedep, struct buf *bp)
984263bc
MD
3242{
3243 struct allocdirect *adp, *lastadp;
50e58362
MD
3244 struct ufs1_dinode *dp;
3245 struct ufs1_dinode *sip;
984263bc
MD
3246 struct fs *fs;
3247 ufs_lbn_t prevlbn = 0;
3248 int i, deplist;
3249
3250 if (inodedep->id_state & IOSTARTED)
3251 panic("initiate_write_inodeblock: already started");
3252 inodedep->id_state |= IOSTARTED;
3253 fs = inodedep->id_fs;
50e58362 3254 dp = (struct ufs1_dinode *)bp->b_data +
984263bc
MD
3255 ino_to_fsbo(fs, inodedep->id_ino);
3256 /*
3257 * If the bitmap is not yet written, then the allocated
3258 * inode cannot be written to disk.
3259 */
3260 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3261 if (inodedep->id_savedino != NULL)
3262 panic("initiate_write_inodeblock: already doing I/O");
50e58362
MD
3263 MALLOC(sip, struct ufs1_dinode *,
3264 sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
83b29fff 3265 inodedep->id_savedino = sip;
984263bc 3266 *inodedep->id_savedino = *dp;
50e58362 3267 bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
83b29fff 3268 dp->di_gen = inodedep->id_savedino->di_gen;
984263bc
MD
3269 return;
3270 }
3271 /*
3272 * If no dependencies, then there is nothing to roll back.
3273 */
3274 inodedep->id_savedsize = dp->di_size;
3275 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3276 return;
3277 /*
3278 * Set the dependencies to busy.
3279 */
3280 ACQUIRE_LOCK(&lk);
3281 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3282 adp = TAILQ_NEXT(adp, ad_next)) {
3283#ifdef DIAGNOSTIC
3284 if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3285 FREE_LOCK(&lk);
3286 panic("softdep_write_inodeblock: lbn order");
3287 }
3288 prevlbn = adp->ad_lbn;
3289 if (adp->ad_lbn < NDADDR &&
3290 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3291 FREE_LOCK(&lk);
3292 panic("%s: direct pointer #%ld mismatch %d != %d",
3293 "softdep_write_inodeblock", adp->ad_lbn,
3294 dp->di_db[adp->ad_lbn], adp->ad_newblkno);
3295 }
3296 if (adp->ad_lbn >= NDADDR &&
3297 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3298 FREE_LOCK(&lk);
3299 panic("%s: indirect pointer #%ld mismatch %d != %d",
3300 "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
3301 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
3302 }
3303 deplist |= 1 << adp->ad_lbn;
3304 if ((adp->ad_state & ATTACHED) == 0) {
3305 FREE_LOCK(&lk);
3306 panic("softdep_write_inodeblock: Unknown state 0x%x",
3307 adp->ad_state);
3308 }
3309#endif /* DIAGNOSTIC */
3310 adp->ad_state &= ~ATTACHED;
3311 adp->ad_state |= UNDONE;
3312 }
3313 /*
3314 * The on-disk inode cannot claim to be any larger than the last
3315 * fragment that has been written. Otherwise, the on-disk inode
3316 * might have fragments that were not the last block in the file
3317 * which would corrupt the filesystem.
3318 */
3319 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3320 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3321 if (adp->ad_lbn >= NDADDR)
3322 break;
3323 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3324 /* keep going until hitting a rollback to a frag */
3325 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3326 continue;
3327 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3328 for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3329#ifdef DIAGNOSTIC
3330 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3331 FREE_LOCK(&lk);
3332 panic("softdep_write_inodeblock: lost dep1");
3333 }
3334#endif /* DIAGNOSTIC */
3335 dp->di_db[i] = 0;
3336 }
3337 for (i = 0; i < NIADDR; i++) {
3338#ifdef DIAGNOSTIC
3339 if (dp->di_ib[i] != 0 &&
3340 (deplist & ((1 << NDADDR) << i)) == 0) {
3341 FREE_LOCK(&lk);
3342 panic("softdep_write_inodeblock: lost dep2");
3343 }
3344#endif /* DIAGNOSTIC */
3345 dp->di_ib[i] = 0;
3346 }
3347 FREE_LOCK(&lk);
3348 return;
3349 }
3350 /*
3351 * If we have zero'ed out the last allocated block of the file,
3352 * roll back the size to the last currently allocated block.
3353 * We know that this last allocated block is a full-sized as
3354 * we already checked for fragments in the loop above.
3355 */
3356 if (lastadp != NULL &&
3357 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3358 for (i = lastadp->ad_lbn; i >= 0; i--)
3359 if (dp->di_db[i] != 0)
3360 break;
3361 dp->di_size = (i + 1) * fs->fs_bsize;
3362 }
3363 /*
3364 * The only dependencies are for indirect blocks.
3365 *
3366 * The file size for indirect block additions is not guaranteed.
3367 * Such a guarantee would be non-trivial to achieve. The conventional
3368 * synchronous write implementation also does not make this guarantee.
3369 * Fsck should catch and fix discrepancies. Arguably, the file size
3370 * can be over-estimated without destroying integrity when the file
3371 * moves into the indirect blocks (i.e., is large). If we want to
3372 * postpone fsck, we are stuck with this argument.
3373 */
3374 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3375 dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3376 FREE_LOCK(&lk);
3377}
3378
3379/*
3380 * This routine is called during the completion interrupt
3381 * service routine for a disk write (from the procedure called
f719c866 3382 * by the device driver to inform the filesystem caches of
984263bc
MD
3383 * a request completion). It should be called early in this
3384 * procedure, before the block is made available to other
3385 * processes or other routines are called.
3fcb1ab8
SW
3386 *
3387 * Parameters:
3388 * bp: describes the completed disk write
984263bc
MD
3389 */
3390static void
3fcb1ab8 3391softdep_disk_write_complete(struct buf *bp)
984263bc
MD
3392{
3393 struct worklist *wk;
3394 struct workhead reattach;
3395 struct newblk *newblk;
3396 struct allocindir *aip;
3397 struct allocdirect *adp;
3398 struct indirdep *indirdep;
3399 struct inodedep *inodedep;
3400 struct bmsafemap *bmsafemap;
3401
3402#ifdef DEBUG
dadab5e9 3403 if (lk.lkt_held != NOHOLDER)
984263bc 3404 panic("softdep_disk_write_complete: lock is held");
dadab5e9 3405 lk.lkt_held = SPECIAL_FLAG;
984263bc
MD
3406#endif
3407 LIST_INIT(&reattach);
3408 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3409 WORKLIST_REMOVE(wk);
3410 switch (wk->wk_type) {
3411
3412 case D_PAGEDEP:
3413 if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3414 WORKLIST_INSERT(&reattach, wk);
3415 continue;
3416
3417 case D_INODEDEP:
3418 if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3419 WORKLIST_INSERT(&reattach, wk);
3420 continue;
3421
3422 case D_BMSAFEMAP:
3423 bmsafemap = WK_BMSAFEMAP(wk);
3424 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3425 newblk->nb_state |= DEPCOMPLETE;
3426 newblk->nb_bmsafemap = NULL;
3427 LIST_REMOVE(newblk, nb_deps);
3428 }
3429 while ((adp =
3430 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3431 adp->ad_state |= DEPCOMPLETE;
3432 adp->ad_buf = NULL;
3433 LIST_REMOVE(adp, ad_deps);
3434 handle_allocdirect_partdone(adp);
3435 }
3436 while ((aip =
3437 LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3438 aip->ai_state |= DEPCOMPLETE;
3439 aip->ai_buf = NULL;
3440 LIST_REMOVE(aip, ai_deps);
3441 handle_allocindir_partdone(aip);
3442 }
3443 while ((inodedep =
3444 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3445 inodedep->id_state |= DEPCOMPLETE;
3446 LIST_REMOVE(inodedep, id_deps);
3447 inodedep->id_buf = NULL;
3448 }
3449 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3450 continue;
3451
3452 case D_MKDIR:
3453 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3454 continue;
3455
3456 case D_ALLOCDIRECT:
3457 adp = WK_ALLOCDIRECT(wk);
3458 adp->ad_state |= COMPLETE;
3459 handle_allocdirect_partdone(adp);
3460 continue;
3461
3462 case D_ALLOCINDIR:
3463 aip = WK_ALLOCINDIR(wk);
3464 aip->ai_state |= COMPLETE;
3465 handle_allocindir_partdone(aip);
3466 continue;
3467
3468 case D_INDIRDEP:
3469 indirdep = WK_INDIRDEP(wk);
3470 if (indirdep->ir_state & GOINGAWAY) {
dadab5e9 3471 lk.lkt_held = NOHOLDER;
984263bc
MD
3472 panic("disk_write_complete: indirdep gone");
3473 }
3474 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3475 FREE(indirdep->ir_saveddata, M_INDIRDEP);
3476 indirdep->ir_saveddata = 0;
3477 indirdep->ir_state &= ~UNDONE;
3478 indirdep->ir_state |= ATTACHED;
3479 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3480 handle_allocindir_partdone(aip);
3481 if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
dadab5e9 3482 lk.lkt_held = NOHOLDER;
984263bc
MD
3483 panic("disk_write_complete: not gone");
3484 }
3485 }
3486 WORKLIST_INSERT(&reattach, wk);
3487 if ((bp->b_flags & B_DELWRI) == 0)
3488 stat_indir_blk_ptrs++;
3489 bdirty(bp);
3490 continue;
3491
3492 default:
dadab5e9 3493 lk.lkt_held = NOHOLDER;
984263bc
MD
3494 panic("handle_disk_write_complete: Unknown type %s",
3495 TYPENAME(wk->wk_type));
3496 /* NOTREACHED */
3497 }
3498 }
3499 /*
3500 * Reattach any requests that must be redone.
3501 */
3502 while ((wk = LIST_FIRST(&reattach)) != NULL) {
3503 WORKLIST_REMOVE(wk);
408357d8 3504 WORKLIST_INSERT_BP(bp, wk);
984263bc
MD
3505 }
3506#ifdef DEBUG
dadab5e9 3507 if (lk.lkt_held != SPECIAL_FLAG)
984263bc 3508 panic("softdep_disk_write_complete: lock lost");
dadab5e9 3509 lk.lkt_held = NOHOLDER;
984263bc
MD
3510#endif
3511}
3512
3513/*
3514 * Called from within softdep_disk_write_complete above. Note that
3515 * this routine is always called from interrupt level with further
3516 * splbio interrupts blocked.
3fcb1ab8
SW
3517 *
3518 * Parameters:
3519 * adp: the completed allocdirect
984263bc
MD
3520 */
3521static void
3fcb1ab8 3522handle_allocdirect_partdone(struct allocdirect *adp)
984263bc
MD
3523{
3524 struct allocdirect *listadp;
3525 struct inodedep *inodedep;
3526 long bsize;
3527
3528 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3529 return;
3530 if (adp->ad_buf != NULL) {
dadab5e9 3531 lk.lkt_held = NOHOLDER;
984263bc
MD
3532 panic("handle_allocdirect_partdone: dangling dep");
3533 }
3534 /*
3535 * The on-disk inode cannot claim to be any larger than the last
3536 * fragment that has been written. Otherwise, the on-disk inode
3537 * might have fragments that were not the last block in the file
3538 * which would corrupt the filesystem. Thus, we cannot free any
3539 * allocdirects after one whose ad_oldblkno claims a fragment as
3540 * these blocks must be rolled back to zero before writing the inode.
3541 * We check the currently active set of allocdirects in id_inoupdt.
3542 */
3543 inodedep = adp->ad_inodedep;
3544 bsize = inodedep->id_fs->fs_bsize;
3545 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) {
3546 /* found our block */
3547 if (listadp == adp)
3548 break;
3549 /* continue if ad_oldlbn is not a fragment */
3550 if (listadp->ad_oldsize == 0 ||
3551 listadp->ad_oldsize == bsize)
3552 continue;
3553 /* hit a fragment */
3554 return;
3555 }
3556 /*
3557 * If we have reached the end of the current list without
3558 * finding the just finished dependency, then it must be
3559 * on the future dependency list. Future dependencies cannot
3560 * be freed until they are moved to the current list.
3561 */
3562 if (listadp == NULL) {
3563#ifdef DEBUG
3564 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next)
3565 /* found our block */
3566 if (listadp == adp)
3567 break;
3568 if (listadp == NULL) {
dadab5e9 3569 lk.lkt_held = NOHOLDER;
984263bc
MD
3570 panic("handle_allocdirect_partdone: lost dep");
3571 }
3572#endif /* DEBUG */
3573 return;
3574 }
3575 /*
3576 * If we have found the just finished dependency, then free
3577 * it along with anything that follows it that is complete.
3578 */
3579 for (; adp; adp = listadp) {
3580 listadp = TAILQ_NEXT(adp, ad_next);
3581 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3582 return;
3583 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
3584 }
3585}
3586
3587/*
3588 * Called from within softdep_disk_write_complete above. Note that
3589 * this routine is always called from interrupt level with further
3590 * splbio interrupts blocked.
3fcb1ab8
SW
3591 *
3592 * Parameters:
3593 * aip: the completed allocindir
984263bc
MD
3594 */
3595static void
3fcb1ab8 3596handle_allocindir_partdone(struct allocindir *aip)
984263bc
MD
3597{
3598 struct indirdep *indirdep;
3599
3600 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
3601 return;
3602 if (aip->ai_buf != NULL) {
dadab5e9 3603 lk.lkt_held = NOHOLDER;
984263bc
MD
3604 panic("handle_allocindir_partdone: dangling dependency");
3605 }
3606 indirdep = aip->ai_indirdep;
3607 if (indirdep->ir_state & UNDONE) {
3608 LIST_REMOVE(aip, ai_next);
3609 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
3610 return;
3611 }
3612 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
3613 aip->ai_newblkno;
3614 LIST_REMOVE(aip, ai_next);
3615 if (aip->ai_freefrag != NULL)
3616 add_to_worklist(&aip->ai_freefrag->ff_list);
3617 WORKITEM_FREE(aip, D_ALLOCINDIR);
3618}
3619
3620/*
3621 * Called from within softdep_disk_write_complete above to restore
3622 * in-memory inode block contents to their most up-to-date state. Note
3623 * that this routine is always called from interrupt level with further
3624 * splbio interrupts blocked.
3fcb1ab8
SW
3625 *
3626 * Parameters:
3627 * bp: buffer containing the inode block
984263bc
MD
3628 */
3629static int
3fcb1ab8 3630handle_written_inodeblock(struct inodedep *inodedep, struct buf *bp)
984263bc
MD
3631{
3632 struct worklist *wk, *filefree;
3633 struct allocdirect *adp, *nextadp;
50e58362 3634 struct ufs1_dinode *dp;
984263bc
MD
3635 int hadchanges;
3636
3637 if ((inodedep->id_state & IOSTARTED) == 0) {
dadab5e9 3638 lk.lkt_held = NOHOLDER;
984263bc
MD
3639 panic("handle_written_inodeblock: not started");
3640 }
3641 inodedep->id_state &= ~IOSTARTED;
50e58362 3642 dp = (struct ufs1_dinode *)bp->b_data +
984263bc
MD
3643 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
3644 /*
3645 * If we had to rollback the inode allocation because of
3646 * bitmaps being incomplete, then simply restore it.
3647 * Keep the block dirty so that it will not be reclaimed until
3648 * all associated dependencies have been cleared and the