__P()!=wanted, clean up the vm subsystem
[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 $
1f2de5d4 40 * $DragonFly: src/sys/vfs/ufs/ffs_softdep.c,v 1.10 2003/08/07 21:17:44 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>
1f2de5d4
MD
64#include "dir.h"
65#include "quota.h"
66#include "inode.h"
67#include "ufsmount.h"
68#include "fs.h"
69#include "softdep.h"
70#include "ffs_extern.h"
71#include "ufs_extern.h"
984263bc
MD
72
73/*
74 * These definitions need to be adapted to the system to which
75 * this file is being ported.
76 */
77/*
78 * malloc types defined for the softdep system.
79 */
80MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
81MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
82MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
83MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
84MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
85MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
86MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
87MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
88MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
89MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
90MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
91MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
92MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
93
94#define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE)
95
96#define D_PAGEDEP 0
97#define D_INODEDEP 1
98#define D_NEWBLK 2
99#define D_BMSAFEMAP 3
100#define D_ALLOCDIRECT 4
101#define D_INDIRDEP 5
102#define D_ALLOCINDIR 6
103#define D_FREEFRAG 7
104#define D_FREEBLKS 8
105#define D_FREEFILE 9
106#define D_DIRADD 10
107#define D_MKDIR 11
108#define D_DIRREM 12
109#define D_LAST D_DIRREM
110
111/*
112 * translate from workitem type to memory type
113 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
114 */
115static struct malloc_type *memtype[] = {
116 M_PAGEDEP,
117 M_INODEDEP,
118 M_NEWBLK,
119 M_BMSAFEMAP,
120 M_ALLOCDIRECT,
121 M_INDIRDEP,
122 M_ALLOCINDIR,
123 M_FREEFRAG,
124 M_FREEBLKS,
125 M_FREEFILE,
126 M_DIRADD,
127 M_MKDIR,
128 M_DIRREM
129};
130
131#define DtoM(type) (memtype[type])
132
133/*
134 * Names of malloc types.
135 */
136#define TYPENAME(type) \
137 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
984263bc
MD
138/*
139 * End system adaptaion definitions.
140 */
141
142/*
143 * Internal function prototypes.
144 */
145static void softdep_error __P((char *, int));
146static void drain_output __P((struct vnode *, int));
147static int getdirtybuf __P((struct buf **, int));
dadab5e9
MD
148static void clear_remove __P((struct thread *));
149static void clear_inodedeps __P((struct thread *));
984263bc
MD
150static int flush_pagedep_deps __P((struct vnode *, struct mount *,
151 struct diraddhd *));
152static int flush_inodedep_deps __P((struct fs *, ino_t));
153static int handle_written_filepage __P((struct pagedep *, struct buf *));
154static void diradd_inode_written __P((struct diradd *, struct inodedep *));
155static int handle_written_inodeblock __P((struct inodedep *, struct buf *));
156static void handle_allocdirect_partdone __P((struct allocdirect *));
157static void handle_allocindir_partdone __P((struct allocindir *));
158static void initiate_write_filepage __P((struct pagedep *, struct buf *));
159static void handle_written_mkdir __P((struct mkdir *, int));
160static void initiate_write_inodeblock __P((struct inodedep *, struct buf *));
161static void handle_workitem_freefile __P((struct freefile *));
162static void handle_workitem_remove __P((struct dirrem *));
163static struct dirrem *newdirrem __P((struct buf *, struct inode *,
164 struct inode *, int, struct dirrem **));
165static void free_diradd __P((struct diradd *));
166static void free_allocindir __P((struct allocindir *, struct inodedep *));
167static int indir_trunc __P((struct inode *, ufs_daddr_t, int, ufs_lbn_t,
168 long *));
169static void deallocate_dependencies __P((struct buf *, struct inodedep *));
170static void free_allocdirect __P((struct allocdirectlst *,
171 struct allocdirect *, int));
172static int check_inode_unwritten __P((struct inodedep *));
173static int free_inodedep __P((struct inodedep *));
174static void handle_workitem_freeblocks __P((struct freeblks *));
175static void merge_inode_lists __P((struct inodedep *));
176static void setup_allocindir_phase2 __P((struct buf *, struct inode *,
177 struct allocindir *));
178static struct allocindir *newallocindir __P((struct inode *, int, ufs_daddr_t,
179 ufs_daddr_t));
180static void handle_workitem_freefrag __P((struct freefrag *));
181static struct freefrag *newfreefrag __P((struct inode *, ufs_daddr_t, long));
182static void allocdirect_merge __P((struct allocdirectlst *,
183 struct allocdirect *, struct allocdirect *));
184static struct bmsafemap *bmsafemap_lookup __P((struct buf *));
185static int newblk_lookup __P((struct fs *, ufs_daddr_t, int,
186 struct newblk **));
187static int inodedep_lookup __P((struct fs *, ino_t, int, struct inodedep **));
188static int pagedep_lookup __P((struct inode *, ufs_lbn_t, int,
189 struct pagedep **));
190static void pause_timer __P((void *));
191static int request_cleanup __P((int, int));
192static int process_worklist_item __P((struct mount *, int));
193static void add_to_worklist __P((struct worklist *));
194
195/*
196 * Exported softdep operations.
197 */
198static void softdep_disk_io_initiation __P((struct buf *));
199static void softdep_disk_write_complete __P((struct buf *));
200static void softdep_deallocate_dependencies __P((struct buf *));
201static int softdep_fsync __P((struct vnode *));
202static int softdep_process_worklist __P((struct mount *));
203static void softdep_move_dependencies __P((struct buf *, struct buf *));
204static int softdep_count_dependencies __P((struct buf *bp, int));
205
206struct bio_ops bioops = {
207 softdep_disk_io_initiation, /* io_start */
208 softdep_disk_write_complete, /* io_complete */
209 softdep_deallocate_dependencies, /* io_deallocate */
210 softdep_fsync, /* io_fsync */
211 softdep_process_worklist, /* io_sync */
212 softdep_move_dependencies, /* io_movedeps */
213 softdep_count_dependencies, /* io_countdeps */
214};
215
216/*
217 * Locking primitives.
218 *
219 * For a uniprocessor, all we need to do is protect against disk
220 * interrupts. For a multiprocessor, this lock would have to be
221 * a mutex. A single mutex is used throughout this file, though
222 * finer grain locking could be used if contention warranted it.
223 *
224 * For a multiprocessor, the sleep call would accept a lock and
225 * release it after the sleep processing was complete. In a uniprocessor
226 * implementation there is no such interlock, so we simple mark
227 * the places where it needs to be done with the `interlocked' form
228 * of the lock calls. Since the uniprocessor sleep already interlocks
229 * the spl, there is nothing that really needs to be done.
230 */
231#ifndef /* NOT */ DEBUG
232static struct lockit {
233 int lkt_spl;
234} lk = { 0 };
235#define ACQUIRE_LOCK(lk) (lk)->lkt_spl = splbio()
236#define FREE_LOCK(lk) splx((lk)->lkt_spl)
237
238#else /* DEBUG */
dadab5e9
MD
239#define NOHOLDER ((struct thread *)-1)
240#define SPECIAL_FLAG ((struct thread *)-2)
984263bc
MD
241static struct lockit {
242 int lkt_spl;
dadab5e9
MD
243 struct thread *lkt_held;
244} lk = { 0, NOHOLDER };
984263bc
MD
245static int lockcnt;
246
247static void acquire_lock __P((struct lockit *));
248static void free_lock __P((struct lockit *));
249void softdep_panic __P((char *));
250
251#define ACQUIRE_LOCK(lk) acquire_lock(lk)
252#define FREE_LOCK(lk) free_lock(lk)
253
254static void
255acquire_lock(lk)
256 struct lockit *lk;
257{
dadab5e9 258 thread_t holder;
984263bc 259
dadab5e9 260 if (lk->lkt_held != NOHOLDER) {
984263bc
MD
261 holder = lk->lkt_held;
262 FREE_LOCK(lk);
dadab5e9 263 if (holder == curthread)
984263bc
MD
264 panic("softdep_lock: locking against myself");
265 else
dadab5e9 266 panic("softdep_lock: lock held by %p", holder);
984263bc
MD
267 }
268 lk->lkt_spl = splbio();
dadab5e9 269 lk->lkt_held = curthread;
984263bc
MD
270 lockcnt++;
271}
272
273static void
274free_lock(lk)
275 struct lockit *lk;
276{
277
dadab5e9 278 if (lk->lkt_held == NOHOLDER)
984263bc 279 panic("softdep_unlock: lock not held");
dadab5e9 280 lk->lkt_held = NOHOLDER;
984263bc
MD
281 splx(lk->lkt_spl);
282}
283
284/*
285 * Function to release soft updates lock and panic.
286 */
287void
288softdep_panic(msg)
289 char *msg;
290{
291
dadab5e9 292 if (lk.lkt_held != NOHOLDER)
984263bc
MD
293 FREE_LOCK(&lk);
294 panic(msg);
295}
296#endif /* DEBUG */
297
298static int interlocked_sleep __P((struct lockit *, int, void *, int,
299 const char *, int));
300
301/*
302 * When going to sleep, we must save our SPL so that it does
303 * not get lost if some other process uses the lock while we
304 * are sleeping. We restore it after we have slept. This routine
305 * wraps the interlocking with functions that sleep. The list
306 * below enumerates the available set of operations.
307 */
308#define UNKNOWN 0
309#define SLEEP 1
310#define LOCKBUF 2
311
312static int
313interlocked_sleep(lk, op, ident, flags, wmesg, timo)
314 struct lockit *lk;
315 int op;
316 void *ident;
317 int flags;
318 const char *wmesg;
319 int timo;
320{
dadab5e9 321 thread_t holder;
984263bc
MD
322 int s, retval;
323
324 s = lk->lkt_spl;
325# ifdef DEBUG
dadab5e9 326 if (lk->lkt_held == NOHOLDER)
984263bc 327 panic("interlocked_sleep: lock not held");
dadab5e9 328 lk->lkt_held = NOHOLDER;
984263bc
MD
329# endif /* DEBUG */
330 switch (op) {
331 case SLEEP:
332 retval = tsleep(ident, flags, wmesg, timo);
333 break;
334 case LOCKBUF:
335 retval = BUF_LOCK((struct buf *)ident, flags);
336 break;
337 default:
338 panic("interlocked_sleep: unknown operation");
339 }
340# ifdef DEBUG
dadab5e9 341 if (lk->lkt_held != NOHOLDER) {
984263bc
MD
342 holder = lk->lkt_held;
343 FREE_LOCK(lk);
dadab5e9 344 if (holder == curthread)
984263bc
MD
345 panic("interlocked_sleep: locking against self");
346 else
dadab5e9 347 panic("interlocked_sleep: lock held by %p", holder);
984263bc 348 }
dadab5e9 349 lk->lkt_held = curthread;
984263bc
MD
350 lockcnt++;
351# endif /* DEBUG */
352 lk->lkt_spl = s;
353 return (retval);
354}
355
356/*
357 * Place holder for real semaphores.
358 */
359struct sema {
360 int value;
dadab5e9 361 thread_t holder;
984263bc
MD
362 char *name;
363 int prio;
364 int timo;
365};
366static void sema_init __P((struct sema *, char *, int, int));
367static int sema_get __P((struct sema *, struct lockit *));
368static void sema_release __P((struct sema *));
369
370static void
371sema_init(semap, name, prio, timo)
372 struct sema *semap;
373 char *name;
374 int prio, timo;
375{
376
dadab5e9 377 semap->holder = NOHOLDER;
984263bc
MD
378 semap->value = 0;
379 semap->name = name;
380 semap->prio = prio;
381 semap->timo = timo;
382}
383
384static int
385sema_get(semap, interlock)
386 struct sema *semap;
387 struct lockit *interlock;
388{
389
390 if (semap->value++ > 0) {
391 if (interlock != NULL) {
392 interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
393 semap->prio, semap->name, semap->timo);
394 FREE_LOCK(interlock);
395 } else {
396 tsleep((caddr_t)semap, semap->prio, semap->name,
397 semap->timo);
398 }
399 return (0);
400 }
dadab5e9 401 semap->holder = curthread;
984263bc
MD
402 if (interlock != NULL)
403 FREE_LOCK(interlock);
404 return (1);
405}
406
407static void
408sema_release(semap)
409 struct sema *semap;
410{
411
dadab5e9
MD
412 if (semap->value <= 0 || semap->holder != curthread) {
413 if (lk.lkt_held != NOHOLDER)
984263bc
MD
414 FREE_LOCK(&lk);
415 panic("sema_release: not held");
416 }
417 if (--semap->value > 0) {
418 semap->value = 0;
419 wakeup(semap);
420 }
dadab5e9 421 semap->holder = NOHOLDER;
984263bc
MD
422}
423
424/*
425 * Worklist queue management.
426 * These routines require that the lock be held.
427 */
428#ifndef /* NOT */ DEBUG
429#define WORKLIST_INSERT(head, item) do { \
430 (item)->wk_state |= ONWORKLIST; \
431 LIST_INSERT_HEAD(head, item, wk_list); \
432} while (0)
433#define WORKLIST_REMOVE(item) do { \
434 (item)->wk_state &= ~ONWORKLIST; \
435 LIST_REMOVE(item, wk_list); \
436} while (0)
437#define WORKITEM_FREE(item, type) FREE(item, DtoM(type))
438
439#else /* DEBUG */
440static void worklist_insert __P((struct workhead *, struct worklist *));
441static void worklist_remove __P((struct worklist *));
442static void workitem_free __P((struct worklist *, int));
443
444#define WORKLIST_INSERT(head, item) worklist_insert(head, item)
445#define WORKLIST_REMOVE(item) worklist_remove(item)
446#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
447
448static void
449worklist_insert(head, item)
450 struct workhead *head;
451 struct worklist *item;
452{
453
dadab5e9 454 if (lk.lkt_held == NOHOLDER)
984263bc
MD
455 panic("worklist_insert: lock not held");
456 if (item->wk_state & ONWORKLIST) {
457 FREE_LOCK(&lk);
458 panic("worklist_insert: already on list");
459 }
460 item->wk_state |= ONWORKLIST;
461 LIST_INSERT_HEAD(head, item, wk_list);
462}
463
464static void
465worklist_remove(item)
466 struct worklist *item;
467{
468
dadab5e9 469 if (lk.lkt_held == NOHOLDER)
984263bc
MD
470 panic("worklist_remove: lock not held");
471 if ((item->wk_state & ONWORKLIST) == 0) {
472 FREE_LOCK(&lk);
473 panic("worklist_remove: not on list");
474 }
475 item->wk_state &= ~ONWORKLIST;
476 LIST_REMOVE(item, wk_list);
477}
478
479static void
480workitem_free(item, type)
481 struct worklist *item;
482 int type;
483{
484
485 if (item->wk_state & ONWORKLIST) {
dadab5e9 486 if (lk.lkt_held != NOHOLDER)
984263bc
MD
487 FREE_LOCK(&lk);
488 panic("workitem_free: still on list");
489 }
490 if (item->wk_type != type) {
dadab5e9 491 if (lk.lkt_held != NOHOLDER)
984263bc
MD
492 FREE_LOCK(&lk);
493 panic("workitem_free: type mismatch");
494 }
495 FREE(item, DtoM(type));
496}
497#endif /* DEBUG */
498
499/*
500 * Workitem queue management
501 */
502static struct workhead softdep_workitem_pending;
503static int num_on_worklist; /* number of worklist items to be processed */
504static int softdep_worklist_busy; /* 1 => trying to do unmount */
505static int softdep_worklist_req; /* serialized waiters */
506static int max_softdeps; /* maximum number of structs before slowdown */
507static int tickdelay = 2; /* number of ticks to pause during slowdown */
508static int *stat_countp; /* statistic to count in proc_waiting timeout */
509static int proc_waiting; /* tracks whether we have a timeout posted */
510static struct callout_handle handle; /* handle on posted proc_waiting timeout */
dadab5e9 511static struct thread *filesys_syncer; /* proc of filesystem syncer process */
984263bc
MD
512static int req_clear_inodedeps; /* syncer process flush some inodedeps */
513#define FLUSH_INODES 1
514static int req_clear_remove; /* syncer process flush some freeblks */
515#define FLUSH_REMOVE 2
516/*
517 * runtime statistics
518 */
519static int stat_worklist_push; /* number of worklist cleanups */
520static int stat_blk_limit_push; /* number of times block limit neared */
521static int stat_ino_limit_push; /* number of times inode limit neared */
522static int stat_blk_limit_hit; /* number of times block slowdown imposed */
523static int stat_ino_limit_hit; /* number of times inode slowdown imposed */
524static int stat_sync_limit_hit; /* number of synchronous slowdowns imposed */
525static int stat_indir_blk_ptrs; /* bufs redirtied as indir ptrs not written */
526static int stat_inode_bitmap; /* bufs redirtied as inode bitmap not written */
527static int stat_direct_blk_ptrs;/* bufs redirtied as direct ptrs not written */
528static int stat_dir_entry; /* bufs redirtied as dir entry cannot write */
529#ifdef DEBUG
530#include <vm/vm.h>
531#include <sys/sysctl.h>
532SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0, "");
533SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0, "");
534SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,"");
535SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,"");
536SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,"");
537SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0, "");
538SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0, "");
539SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0, "");
540SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0, "");
541SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0, "");
542SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0, "");
543SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0, "");
544#endif /* DEBUG */
545
546/*
547 * Add an item to the end of the work queue.
548 * This routine requires that the lock be held.
549 * This is the only routine that adds items to the list.
550 * The following routine is the only one that removes items
551 * and does so in order from first to last.
552 */
553static void
554add_to_worklist(wk)
555 struct worklist *wk;
556{
557 static struct worklist *worklist_tail;
558
559 if (wk->wk_state & ONWORKLIST) {
dadab5e9 560 if (lk.lkt_held != NOHOLDER)
984263bc
MD
561 FREE_LOCK(&lk);
562 panic("add_to_worklist: already on list");
563 }
564 wk->wk_state |= ONWORKLIST;
565 if (LIST_FIRST(&softdep_workitem_pending) == NULL)
566 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
567 else
568 LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
569 worklist_tail = wk;
570 num_on_worklist += 1;
571}
572
573/*
574 * Process that runs once per second to handle items in the background queue.
575 *
576 * Note that we ensure that everything is done in the order in which they
577 * appear in the queue. The code below depends on this property to ensure
578 * that blocks of a file are freed before the inode itself is freed. This
579 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
580 * until all the old ones have been purged from the dependency lists.
581 */
582static int
583softdep_process_worklist(matchmnt)
584 struct mount *matchmnt;
585{
dadab5e9 586 thread_t td = curthread;
984263bc
MD
587 int matchcnt, loopcount;
588 long starttime;
589
590 /*
591 * Record the process identifier of our caller so that we can give
592 * this process preferential treatment in request_cleanup below.
593 */
dadab5e9 594 filesys_syncer = td;
984263bc
MD
595 matchcnt = 0;
596
597 /*
598 * There is no danger of having multiple processes run this
599 * code, but we have to single-thread it when softdep_flushfiles()
600 * is in operation to get an accurate count of the number of items
601 * related to its mount point that are in the list.
602 */
603 if (matchmnt == NULL) {
604 if (softdep_worklist_busy < 0)
605 return(-1);
606 softdep_worklist_busy += 1;
607 }
608
609 /*
610 * If requested, try removing inode or removal dependencies.
611 */
612 if (req_clear_inodedeps) {
dadab5e9 613 clear_inodedeps(td);
984263bc
MD
614 req_clear_inodedeps -= 1;
615 wakeup_one(&proc_waiting);
616 }
617 if (req_clear_remove) {
dadab5e9 618 clear_remove(td);
984263bc
MD
619 req_clear_remove -= 1;
620 wakeup_one(&proc_waiting);
621 }
622 loopcount = 1;
623 starttime = time_second;
624 while (num_on_worklist > 0) {
625 matchcnt += process_worklist_item(matchmnt, 0);
626
627 /*
628 * If a umount operation wants to run the worklist
629 * accurately, abort.
630 */
631 if (softdep_worklist_req && matchmnt == NULL) {
632 matchcnt = -1;
633 break;
634 }
635
636 /*
637 * If requested, try removing inode or removal dependencies.
638 */
639 if (req_clear_inodedeps) {
dadab5e9 640 clear_inodedeps(td);
984263bc
MD
641 req_clear_inodedeps -= 1;
642 wakeup_one(&proc_waiting);
643 }
644 if (req_clear_remove) {
dadab5e9 645 clear_remove(td);
984263bc
MD
646 req_clear_remove -= 1;
647 wakeup_one(&proc_waiting);
648 }
649 /*
650 * We do not generally want to stop for buffer space, but if
651 * we are really being a buffer hog, we will stop and wait.
652 */
653 if (loopcount++ % 128 == 0)
654 bwillwrite();
655 /*
656 * Never allow processing to run for more than one
657 * second. Otherwise the other syncer tasks may get
658 * excessively backlogged.
659 */
660 if (starttime != time_second && matchmnt == NULL) {
661 matchcnt = -1;
662 break;
663 }
664 }
665 if (matchmnt == NULL) {
666 --softdep_worklist_busy;
667 if (softdep_worklist_req && softdep_worklist_busy == 0)
668 wakeup(&softdep_worklist_req);
669 }
670 return (matchcnt);
671}
672
673/*
674 * Process one item on the worklist.
675 */
676static int
677process_worklist_item(matchmnt, flags)
678 struct mount *matchmnt;
679 int flags;
680{
681 struct worklist *wk;
682 struct dirrem *dirrem;
683 struct fs *matchfs;
684 struct vnode *vp;
685 int matchcnt = 0;
686
687 matchfs = NULL;
688 if (matchmnt != NULL)
689 matchfs = VFSTOUFS(matchmnt)->um_fs;
690 ACQUIRE_LOCK(&lk);
691 /*
692 * Normally we just process each item on the worklist in order.
693 * However, if we are in a situation where we cannot lock any
694 * inodes, we have to skip over any dirrem requests whose
695 * vnodes are resident and locked.
696 */
697 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
698 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
699 break;
700 dirrem = WK_DIRREM(wk);
701 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev,
702 dirrem->dm_oldinum);
dadab5e9 703 if (vp == NULL || !VOP_ISLOCKED(vp, curthread))
984263bc
MD
704 break;
705 }
706 if (wk == 0) {
707 FREE_LOCK(&lk);
708 return (0);
709 }
710 WORKLIST_REMOVE(wk);
711 num_on_worklist -= 1;
712 FREE_LOCK(&lk);
713 switch (wk->wk_type) {
714
715 case D_DIRREM:
716 /* removal of a directory entry */
717 if (WK_DIRREM(wk)->dm_mnt == matchmnt)
718 matchcnt += 1;
719 handle_workitem_remove(WK_DIRREM(wk));
720 break;
721
722 case D_FREEBLKS:
723 /* releasing blocks and/or fragments from a file */
724 if (WK_FREEBLKS(wk)->fb_fs == matchfs)
725 matchcnt += 1;
726 handle_workitem_freeblocks(WK_FREEBLKS(wk));
727 break;
728
729 case D_FREEFRAG:
730 /* releasing a fragment when replaced as a file grows */
731 if (WK_FREEFRAG(wk)->ff_fs == matchfs)
732 matchcnt += 1;
733 handle_workitem_freefrag(WK_FREEFRAG(wk));
734 break;
735
736 case D_FREEFILE:
737 /* releasing an inode when its link count drops to 0 */
738 if (WK_FREEFILE(wk)->fx_fs == matchfs)
739 matchcnt += 1;
740 handle_workitem_freefile(WK_FREEFILE(wk));
741 break;
742
743 default:
744 panic("%s_process_worklist: Unknown type %s",
745 "softdep", TYPENAME(wk->wk_type));
746 /* NOTREACHED */
747 }
748 return (matchcnt);
749}
750
751/*
752 * Move dependencies from one buffer to another.
753 */
754static void
755softdep_move_dependencies(oldbp, newbp)
756 struct buf *oldbp;
757 struct buf *newbp;
758{
759 struct worklist *wk, *wktail;
760
761 if (LIST_FIRST(&newbp->b_dep) != NULL)
762 panic("softdep_move_dependencies: need merge code");
763 wktail = 0;
764 ACQUIRE_LOCK(&lk);
765 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
766 LIST_REMOVE(wk, wk_list);
767 if (wktail == 0)
768 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
769 else
770 LIST_INSERT_AFTER(wktail, wk, wk_list);
771 wktail = wk;
772 }
773 FREE_LOCK(&lk);
774}
775
776/*
777 * Purge the work list of all items associated with a particular mount point.
778 */
779int
dadab5e9 780softdep_flushfiles(struct mount *oldmnt, int flags, struct thread *td)
984263bc
MD
781{
782 struct vnode *devvp;
783 int error, loopcnt;
784
785 /*
786 * Await our turn to clear out the queue, then serialize access.
787 */
788 while (softdep_worklist_busy != 0) {
789 softdep_worklist_req += 1;
377d4740 790 tsleep(&softdep_worklist_req, 0, "softflush", 0);
984263bc
MD
791 softdep_worklist_req -= 1;
792 }
793 softdep_worklist_busy = -1;
794
dadab5e9 795 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0) {
984263bc
MD
796 softdep_worklist_busy = 0;
797 if (softdep_worklist_req)
798 wakeup(&softdep_worklist_req);
799 return (error);
800 }
801 /*
802 * Alternately flush the block device associated with the mount
803 * point and process any dependencies that the flushing
804 * creates. In theory, this loop can happen at most twice,
805 * but we give it a few extra just to be sure.
806 */
807 devvp = VFSTOUFS(oldmnt)->um_devvp;
808 for (loopcnt = 10; loopcnt > 0; ) {
809 if (softdep_process_worklist(oldmnt) == 0) {
810 loopcnt--;
811 /*
812 * Do another flush in case any vnodes were brought in
813 * as part of the cleanup operations.
814 */
dadab5e9 815 if ((error = ffs_flushfiles(oldmnt, flags, td)) != 0)
984263bc
MD
816 break;
817 /*
818 * If we still found nothing to do, we are really done.
819 */
820 if (softdep_process_worklist(oldmnt) == 0)
821 break;
822 }
dadab5e9 823 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td);
3b568787 824 error = VOP_FSYNC(devvp, MNT_WAIT, td);
dadab5e9 825 VOP_UNLOCK(devvp, 0, td);
984263bc
MD
826 if (error)
827 break;
828 }
829 softdep_worklist_busy = 0;
830 if (softdep_worklist_req)
831 wakeup(&softdep_worklist_req);
832
833 /*
834 * If we are unmounting then it is an error to fail. If we
835 * are simply trying to downgrade to read-only, then filesystem
836 * activity can keep us busy forever, so we just fail with EBUSY.
837 */
838 if (loopcnt == 0) {
839 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
840 panic("softdep_flushfiles: looping");
841 error = EBUSY;
842 }
843 return (error);
844}
845
846/*
847 * Structure hashing.
848 *
849 * There are three types of structures that can be looked up:
850 * 1) pagedep structures identified by mount point, inode number,
851 * and logical block.
852 * 2) inodedep structures identified by mount point and inode number.
853 * 3) newblk structures identified by mount point and
854 * physical block number.
855 *
856 * The "pagedep" and "inodedep" dependency structures are hashed
857 * separately from the file blocks and inodes to which they correspond.
858 * This separation helps when the in-memory copy of an inode or
859 * file block must be replaced. It also obviates the need to access
860 * an inode or file page when simply updating (or de-allocating)
861 * dependency structures. Lookup of newblk structures is needed to
862 * find newly allocated blocks when trying to associate them with
863 * their allocdirect or allocindir structure.
864 *
865 * The lookup routines optionally create and hash a new instance when
866 * an existing entry is not found.
867 */
868#define DEPALLOC 0x0001 /* allocate structure if lookup fails */
869#define NODELAY 0x0002 /* cannot do background work */
870
871/*
872 * Structures and routines associated with pagedep caching.
873 */
874LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
875u_long pagedep_hash; /* size of hash table - 1 */
876#define PAGEDEP_HASH(mp, inum, lbn) \
877 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
878 pagedep_hash])
879static struct sema pagedep_in_progress;
880
881/*
882 * Look up a pagedep. Return 1 if found, 0 if not found.
883 * If not found, allocate if DEPALLOC flag is passed.
884 * Found or allocated entry is returned in pagedeppp.
885 * This routine must be called with splbio interrupts blocked.
886 */
887static int
888pagedep_lookup(ip, lbn, flags, pagedeppp)
889 struct inode *ip;
890 ufs_lbn_t lbn;
891 int flags;
892 struct pagedep **pagedeppp;
893{
894 struct pagedep *pagedep;
895 struct pagedep_hashhead *pagedephd;
896 struct mount *mp;
897 int i;
898
899#ifdef DEBUG
dadab5e9 900 if (lk.lkt_held == NOHOLDER)
984263bc
MD
901 panic("pagedep_lookup: lock not held");
902#endif
903 mp = ITOV(ip)->v_mount;
904 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
905top:
906 LIST_FOREACH(pagedep, pagedephd, pd_hash)
907 if (ip->i_number == pagedep->pd_ino &&
908 lbn == pagedep->pd_lbn &&
909 mp == pagedep->pd_mnt)
910 break;
911 if (pagedep) {
912 *pagedeppp = pagedep;
913 return (1);
914 }
915 if ((flags & DEPALLOC) == 0) {
916 *pagedeppp = NULL;
917 return (0);
918 }
919 if (sema_get(&pagedep_in_progress, &lk) == 0) {
920 ACQUIRE_LOCK(&lk);
921 goto top;
922 }
923 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
924 M_SOFTDEP_FLAGS);
925 bzero(pagedep, sizeof(struct pagedep));
926 pagedep->pd_list.wk_type = D_PAGEDEP;
927 pagedep->pd_mnt = mp;
928 pagedep->pd_ino = ip->i_number;
929 pagedep->pd_lbn = lbn;
930 LIST_INIT(&pagedep->pd_dirremhd);
931 LIST_INIT(&pagedep->pd_pendinghd);
932 for (i = 0; i < DAHASHSZ; i++)
933 LIST_INIT(&pagedep->pd_diraddhd[i]);
934 ACQUIRE_LOCK(&lk);
935 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
936 sema_release(&pagedep_in_progress);
937 *pagedeppp = pagedep;
938 return (0);
939}
940
941/*
942 * Structures and routines associated with inodedep caching.
943 */
944LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
945static u_long inodedep_hash; /* size of hash table - 1 */
946static long num_inodedep; /* number of inodedep allocated */
947#define INODEDEP_HASH(fs, inum) \
948 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
949static struct sema inodedep_in_progress;
950
951/*
952 * Look up a inodedep. Return 1 if found, 0 if not found.
953 * If not found, allocate if DEPALLOC flag is passed.
954 * Found or allocated entry is returned in inodedeppp.
955 * This routine must be called with splbio interrupts blocked.
956 */
957static int
958inodedep_lookup(fs, inum, flags, inodedeppp)
959 struct fs *fs;
960 ino_t inum;
961 int flags;
962 struct inodedep **inodedeppp;
963{
964 struct inodedep *inodedep;
965 struct inodedep_hashhead *inodedephd;
966 int firsttry;
967
968#ifdef DEBUG
dadab5e9 969 if (lk.lkt_held == NOHOLDER)
984263bc
MD
970 panic("inodedep_lookup: lock not held");
971#endif
972 firsttry = 1;
973 inodedephd = INODEDEP_HASH(fs, inum);
974top:
975 LIST_FOREACH(inodedep, inodedephd, id_hash)
976 if (inum == inodedep->id_ino && fs == inodedep->id_fs)
977 break;
978 if (inodedep) {
979 *inodedeppp = inodedep;
980 return (1);
981 }
982 if ((flags & DEPALLOC) == 0) {
983 *inodedeppp = NULL;
984 return (0);
985 }
986 /*
987 * If we are over our limit, try to improve the situation.
988 */
989 if (num_inodedep > max_softdeps && firsttry &&
990 speedup_syncer() == 0 && (flags & NODELAY) == 0 &&
991 request_cleanup(FLUSH_INODES, 1)) {
992 firsttry = 0;
993 goto top;
994 }
995 if (sema_get(&inodedep_in_progress, &lk) == 0) {
996 ACQUIRE_LOCK(&lk);
997 goto top;
998 }
999 num_inodedep += 1;
1000 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
1001 M_INODEDEP, M_SOFTDEP_FLAGS);
1002 inodedep->id_list.wk_type = D_INODEDEP;
1003 inodedep->id_fs = fs;
1004 inodedep->id_ino = inum;
1005 inodedep->id_state = ALLCOMPLETE;
1006 inodedep->id_nlinkdelta = 0;
1007 inodedep->id_savedino = NULL;
1008 inodedep->id_savedsize = -1;
1009 inodedep->id_buf = NULL;
1010 LIST_INIT(&inodedep->id_pendinghd);
1011 LIST_INIT(&inodedep->id_inowait);
1012 LIST_INIT(&inodedep->id_bufwait);
1013 TAILQ_INIT(&inodedep->id_inoupdt);
1014 TAILQ_INIT(&inodedep->id_newinoupdt);
1015 ACQUIRE_LOCK(&lk);
1016 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1017 sema_release(&inodedep_in_progress);
1018 *inodedeppp = inodedep;
1019 return (0);
1020}
1021
1022/*
1023 * Structures and routines associated with newblk caching.
1024 */
1025LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1026u_long newblk_hash; /* size of hash table - 1 */
1027#define NEWBLK_HASH(fs, inum) \
1028 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1029static struct sema newblk_in_progress;
1030
1031/*
1032 * Look up a newblk. Return 1 if found, 0 if not found.
1033 * If not found, allocate if DEPALLOC flag is passed.
1034 * Found or allocated entry is returned in newblkpp.
1035 */
1036static int
1037newblk_lookup(fs, newblkno, flags, newblkpp)
1038 struct fs *fs;
1039 ufs_daddr_t newblkno;
1040 int flags;
1041 struct newblk **newblkpp;
1042{
1043 struct newblk *newblk;
1044 struct newblk_hashhead *newblkhd;
1045
1046 newblkhd = NEWBLK_HASH(fs, newblkno);
1047top:
1048 LIST_FOREACH(newblk, newblkhd, nb_hash)
1049 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1050 break;
1051 if (newblk) {
1052 *newblkpp = newblk;
1053 return (1);
1054 }
1055 if ((flags & DEPALLOC) == 0) {
1056 *newblkpp = NULL;
1057 return (0);
1058 }
1059 if (sema_get(&newblk_in_progress, 0) == 0)
1060 goto top;
1061 MALLOC(newblk, struct newblk *, sizeof(struct newblk),
1062 M_NEWBLK, M_SOFTDEP_FLAGS);
1063 newblk->nb_state = 0;
1064 newblk->nb_fs = fs;
1065 newblk->nb_newblkno = newblkno;
1066 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1067 sema_release(&newblk_in_progress);
1068 *newblkpp = newblk;
1069 return (0);
1070}
1071
1072/*
1073 * Executed during filesystem system initialization before
1074 * mounting any file systems.
1075 */
1076void
1077softdep_initialize()
1078{
1079
1080 LIST_INIT(&mkdirlisthd);
1081 LIST_INIT(&softdep_workitem_pending);
1082 max_softdeps = min(desiredvnodes * 8,
1083 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep)));
1084 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1085 &pagedep_hash);
377d4740 1086 sema_init(&pagedep_in_progress, "pagedep", 0, 0);
984263bc 1087 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
377d4740 1088 sema_init(&inodedep_in_progress, "inodedep", 0, 0);
984263bc 1089 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
377d4740 1090 sema_init(&newblk_in_progress, "newblk", 0, 0);
984263bc
MD
1091}
1092
1093/*
1094 * Called at mount time to notify the dependency code that a
1095 * filesystem wishes to use it.
1096 */
1097int
3b568787 1098softdep_mount(devvp, mp, fs)
984263bc
MD
1099 struct vnode *devvp;
1100 struct mount *mp;
1101 struct fs *fs;
984263bc
MD
1102{
1103 struct csum cstotal;
1104 struct cg *cgp;
1105 struct buf *bp;
1106 int error, cyl;
1107
1108 mp->mnt_flag &= ~MNT_ASYNC;
1109 mp->mnt_flag |= MNT_SOFTDEP;
1110 /*
1111 * When doing soft updates, the counters in the
1112 * superblock may have gotten out of sync, so we have
1113 * to scan the cylinder groups and recalculate them.
1114 */
1115 if (fs->fs_clean != 0)
1116 return (0);
1117 bzero(&cstotal, sizeof cstotal);
1118 for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1119 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
3b568787 1120 fs->fs_cgsize, &bp)) != 0) {
984263bc
MD
1121 brelse(bp);
1122 return (error);
1123 }
1124 cgp = (struct cg *)bp->b_data;
1125 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1126 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1127 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1128 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1129 fs->fs_cs(fs, cyl) = cgp->cg_cs;
1130 brelse(bp);
1131 }
1132#ifdef DEBUG
1133 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1134 printf("ffs_mountfs: superblock updated for soft updates\n");
1135#endif
1136 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1137 return (0);
1138}
1139
1140/*
1141 * Protecting the freemaps (or bitmaps).
1142 *
1143 * To eliminate the need to execute fsck before mounting a file system
1144 * after a power failure, one must (conservatively) guarantee that the
1145 * on-disk copy of the bitmaps never indicate that a live inode or block is
1146 * free. So, when a block or inode is allocated, the bitmap should be
1147 * updated (on disk) before any new pointers. When a block or inode is
1148 * freed, the bitmap should not be updated until all pointers have been
1149 * reset. The latter dependency is handled by the delayed de-allocation
1150 * approach described below for block and inode de-allocation. The former
1151 * dependency is handled by calling the following procedure when a block or
1152 * inode is allocated. When an inode is allocated an "inodedep" is created
1153 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1154 * Each "inodedep" is also inserted into the hash indexing structure so
1155 * that any additional link additions can be made dependent on the inode
1156 * allocation.
1157 *
1158 * The ufs file system maintains a number of free block counts (e.g., per
1159 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1160 * in addition to the bitmaps. These counts are used to improve efficiency
1161 * during allocation and therefore must be consistent with the bitmaps.
1162 * There is no convenient way to guarantee post-crash consistency of these
1163 * counts with simple update ordering, for two main reasons: (1) The counts
1164 * and bitmaps for a single cylinder group block are not in the same disk
1165 * sector. If a disk write is interrupted (e.g., by power failure), one may
1166 * be written and the other not. (2) Some of the counts are located in the
1167 * superblock rather than the cylinder group block. So, we focus our soft
1168 * updates implementation on protecting the bitmaps. When mounting a
1169 * filesystem, we recompute the auxiliary counts from the bitmaps.
1170 */
1171
1172/*
1173 * Called just after updating the cylinder group block to allocate an inode.
1174 */
1175void
1176softdep_setup_inomapdep(bp, ip, newinum)
1177 struct buf *bp; /* buffer for cylgroup block with inode map */
1178 struct inode *ip; /* inode related to allocation */
1179 ino_t newinum; /* new inode number being allocated */
1180{
1181 struct inodedep *inodedep;
1182 struct bmsafemap *bmsafemap;
1183
1184 /*
1185 * Create a dependency for the newly allocated inode.
1186 * Panic if it already exists as something is seriously wrong.
1187 * Otherwise add it to the dependency list for the buffer holding
1188 * the cylinder group map from which it was allocated.
1189 */
1190 ACQUIRE_LOCK(&lk);
1191 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1192 FREE_LOCK(&lk);
1193 panic("softdep_setup_inomapdep: found inode");
1194 }
1195 inodedep->id_buf = bp;
1196 inodedep->id_state &= ~DEPCOMPLETE;
1197 bmsafemap = bmsafemap_lookup(bp);
1198 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1199 FREE_LOCK(&lk);
1200}
1201
1202/*
1203 * Called just after updating the cylinder group block to
1204 * allocate block or fragment.
1205 */
1206void
1207softdep_setup_blkmapdep(bp, fs, newblkno)
1208 struct buf *bp; /* buffer for cylgroup block with block map */
1209 struct fs *fs; /* filesystem doing allocation */
1210 ufs_daddr_t newblkno; /* number of newly allocated block */
1211{
1212 struct newblk *newblk;
1213 struct bmsafemap *bmsafemap;
1214
1215 /*
1216 * Create a dependency for the newly allocated block.
1217 * Add it to the dependency list for the buffer holding
1218 * the cylinder group map from which it was allocated.
1219 */
1220 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1221 panic("softdep_setup_blkmapdep: found block");
1222 ACQUIRE_LOCK(&lk);
1223 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1224 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1225 FREE_LOCK(&lk);
1226}
1227
1228/*
1229 * Find the bmsafemap associated with a cylinder group buffer.
1230 * If none exists, create one. The buffer must be locked when
1231 * this routine is called and this routine must be called with
1232 * splbio interrupts blocked.
1233 */
1234static struct bmsafemap *
1235bmsafemap_lookup(bp)
1236 struct buf *bp;
1237{
1238 struct bmsafemap *bmsafemap;
1239 struct worklist *wk;
1240
1241#ifdef DEBUG
dadab5e9 1242 if (lk.lkt_held == NOHOLDER)
984263bc
MD
1243 panic("bmsafemap_lookup: lock not held");
1244#endif
1245 LIST_FOREACH(wk, &bp->b_dep, wk_list)
1246 if (wk->wk_type == D_BMSAFEMAP)
1247 return (WK_BMSAFEMAP(wk));
1248 FREE_LOCK(&lk);
1249 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1250 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1251 bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1252 bmsafemap->sm_list.wk_state = 0;
1253 bmsafemap->sm_buf = bp;
1254 LIST_INIT(&bmsafemap->sm_allocdirecthd);
1255 LIST_INIT(&bmsafemap->sm_allocindirhd);
1256 LIST_INIT(&bmsafemap->sm_inodedephd);
1257 LIST_INIT(&bmsafemap->sm_newblkhd);
1258 ACQUIRE_LOCK(&lk);
1259 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1260 return (bmsafemap);
1261}
1262
1263/*
1264 * Direct block allocation dependencies.
1265 *
1266 * When a new block is allocated, the corresponding disk locations must be
1267 * initialized (with zeros or new data) before the on-disk inode points to
1268 * them. Also, the freemap from which the block was allocated must be
1269 * updated (on disk) before the inode's pointer. These two dependencies are
1270 * independent of each other and are needed for all file blocks and indirect
1271 * blocks that are pointed to directly by the inode. Just before the
1272 * "in-core" version of the inode is updated with a newly allocated block
1273 * number, a procedure (below) is called to setup allocation dependency
1274 * structures. These structures are removed when the corresponding
1275 * dependencies are satisfied or when the block allocation becomes obsolete
1276 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1277 * fragment that gets upgraded). All of these cases are handled in
1278 * procedures described later.
1279 *
1280 * When a file extension causes a fragment to be upgraded, either to a larger
1281 * fragment or to a full block, the on-disk location may change (if the
1282 * previous fragment could not simply be extended). In this case, the old
1283 * fragment must be de-allocated, but not until after the inode's pointer has
1284 * been updated. In most cases, this is handled by later procedures, which
1285 * will construct a "freefrag" structure to be added to the workitem queue
1286 * when the inode update is complete (or obsolete). The main exception to
1287 * this is when an allocation occurs while a pending allocation dependency
1288 * (for the same block pointer) remains. This case is handled in the main
1289 * allocation dependency setup procedure by immediately freeing the
1290 * unreferenced fragments.
1291 */
1292void
1293softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1294 struct inode *ip; /* inode to which block is being added */
1295 ufs_lbn_t lbn; /* block pointer within inode */
1296 ufs_daddr_t newblkno; /* disk block number being added */
1297 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */
1298 long newsize; /* size of new block */
1299 long oldsize; /* size of new block */
1300 struct buf *bp; /* bp for allocated block */
1301{
1302 struct allocdirect *adp, *oldadp;
1303 struct allocdirectlst *adphead;
1304 struct bmsafemap *bmsafemap;
1305 struct inodedep *inodedep;
1306 struct pagedep *pagedep;
1307 struct newblk *newblk;
1308
1309 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1310 M_ALLOCDIRECT, M_SOFTDEP_FLAGS);
1311 bzero(adp, sizeof(struct allocdirect));
1312 adp->ad_list.wk_type = D_ALLOCDIRECT;
1313 adp->ad_lbn = lbn;
1314 adp->ad_newblkno = newblkno;
1315 adp->ad_oldblkno = oldblkno;
1316 adp->ad_newsize = newsize;
1317 adp->ad_oldsize = oldsize;
1318 adp->ad_state = ATTACHED;
1319 if (newblkno == oldblkno)
1320 adp->ad_freefrag = NULL;
1321 else
1322 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1323
1324 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1325 panic("softdep_setup_allocdirect: lost block");
1326
1327 ACQUIRE_LOCK(&lk);
1328 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1329 adp->ad_inodedep = inodedep;
1330
1331 if (newblk->nb_state == DEPCOMPLETE) {
1332 adp->ad_state |= DEPCOMPLETE;
1333 adp->ad_buf = NULL;
1334 } else {
1335 bmsafemap = newblk->nb_bmsafemap;
1336 adp->ad_buf = bmsafemap->sm_buf;
1337 LIST_REMOVE(newblk, nb_deps);
1338 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1339 }
1340 LIST_REMOVE(newblk, nb_hash);
1341 FREE(newblk, M_NEWBLK);
1342
1343 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1344 if (lbn >= NDADDR) {
1345 /* allocating an indirect block */
1346 if (oldblkno != 0) {
1347 FREE_LOCK(&lk);
1348 panic("softdep_setup_allocdirect: non-zero indir");
1349 }
1350 } else {
1351 /*
1352 * Allocating a direct block.
1353 *
1354 * If we are allocating a directory block, then we must
1355 * allocate an associated pagedep to track additions and
1356 * deletions.
1357 */
1358 if ((ip->i_mode & IFMT) == IFDIR &&
1359 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1360 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1361 }
1362 /*
1363 * The list of allocdirects must be kept in sorted and ascending
1364 * order so that the rollback routines can quickly determine the
1365 * first uncommitted block (the size of the file stored on disk
1366 * ends at the end of the lowest committed fragment, or if there
1367 * are no fragments, at the end of the highest committed block).
1368 * Since files generally grow, the typical case is that the new
1369 * block is to be added at the end of the list. We speed this
1370 * special case by checking against the last allocdirect in the
1371 * list before laboriously traversing the list looking for the
1372 * insertion point.
1373 */
1374 adphead = &inodedep->id_newinoupdt;
1375 oldadp = TAILQ_LAST(adphead, allocdirectlst);
1376 if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1377 /* insert at end of list */
1378 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1379 if (oldadp != NULL && oldadp->ad_lbn == lbn)
1380 allocdirect_merge(adphead, adp, oldadp);
1381 FREE_LOCK(&lk);
1382 return;
1383 }
1384 TAILQ_FOREACH(oldadp, adphead, ad_next) {
1385 if (oldadp->ad_lbn >= lbn)
1386 break;
1387 }
1388 if (oldadp == NULL) {
1389 FREE_LOCK(&lk);
1390 panic("softdep_setup_allocdirect: lost entry");
1391 }
1392 /* insert in middle of list */
1393 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1394 if (oldadp->ad_lbn == lbn)
1395 allocdirect_merge(adphead, adp, oldadp);
1396 FREE_LOCK(&lk);
1397}
1398
1399/*
1400 * Replace an old allocdirect dependency with a newer one.
1401 * This routine must be called with splbio interrupts blocked.
1402 */
1403static void
1404allocdirect_merge(adphead, newadp, oldadp)
1405 struct allocdirectlst *adphead; /* head of list holding allocdirects */
1406 struct allocdirect *newadp; /* allocdirect being added */
1407 struct allocdirect *oldadp; /* existing allocdirect being checked */
1408{
1409 struct freefrag *freefrag;
1410
1411#ifdef DEBUG
dadab5e9 1412 if (lk.lkt_held == NOHOLDER)
984263bc
MD
1413 panic("allocdirect_merge: lock not held");
1414#endif
1415 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1416 newadp->ad_oldsize != oldadp->ad_newsize ||
1417 newadp->ad_lbn >= NDADDR) {
1418 FREE_LOCK(&lk);
1419 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1420 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1421 NDADDR);
1422 }
1423 newadp->ad_oldblkno = oldadp->ad_oldblkno;
1424 newadp->ad_oldsize = oldadp->ad_oldsize;
1425 /*
1426 * If the old dependency had a fragment to free or had never
1427 * previously had a block allocated, then the new dependency
1428 * can immediately post its freefrag and adopt the old freefrag.
1429 * This action is done by swapping the freefrag dependencies.
1430 * The new dependency gains the old one's freefrag, and the
1431 * old one gets the new one and then immediately puts it on
1432 * the worklist when it is freed by free_allocdirect. It is
1433 * not possible to do this swap when the old dependency had a
1434 * non-zero size but no previous fragment to free. This condition
1435 * arises when the new block is an extension of the old block.
1436 * Here, the first part of the fragment allocated to the new
1437 * dependency is part of the block currently claimed on disk by
1438 * the old dependency, so cannot legitimately be freed until the
1439 * conditions for the new dependency are fulfilled.
1440 */
1441 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1442 freefrag = newadp->ad_freefrag;
1443 newadp->ad_freefrag = oldadp->ad_freefrag;
1444 oldadp->ad_freefrag = freefrag;
1445 }
1446 free_allocdirect(adphead, oldadp, 0);
1447}
1448
1449/*
1450 * Allocate a new freefrag structure if needed.
1451 */
1452static struct freefrag *
1453newfreefrag(ip, blkno, size)
1454 struct inode *ip;
1455 ufs_daddr_t blkno;
1456 long size;
1457{
1458 struct freefrag *freefrag;
1459 struct fs *fs;
1460
1461 if (blkno == 0)
1462 return (NULL);
1463 fs = ip->i_fs;
1464 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1465 panic("newfreefrag: frag size");
1466 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1467 M_FREEFRAG, M_SOFTDEP_FLAGS);
1468 freefrag->ff_list.wk_type = D_FREEFRAG;
1469 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */
1470 freefrag->ff_inum = ip->i_number;
1471 freefrag->ff_fs = fs;
1472 freefrag->ff_devvp = ip->i_devvp;
1473 freefrag->ff_blkno = blkno;
1474 freefrag->ff_fragsize = size;
1475 return (freefrag);
1476}
1477
1478/*
1479 * This workitem de-allocates fragments that were replaced during
1480 * file block allocation.
1481 */
1482static void
1483handle_workitem_freefrag(freefrag)
1484 struct freefrag *freefrag;
1485{
1486 struct inode tip;
1487
1488 tip.i_fs = freefrag->ff_fs;
1489 tip.i_devvp = freefrag->ff_devvp;
1490 tip.i_dev = freefrag->ff_devvp->v_rdev;
1491 tip.i_number = freefrag->ff_inum;
1492 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */
1493 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1494 FREE(freefrag, M_FREEFRAG);
1495}
1496
1497/*
1498 * Indirect block allocation dependencies.
1499 *
1500 * The same dependencies that exist for a direct block also exist when
1501 * a new block is allocated and pointed to by an entry in a block of
1502 * indirect pointers. The undo/redo states described above are also
1503 * used here. Because an indirect block contains many pointers that
1504 * may have dependencies, a second copy of the entire in-memory indirect
1505 * block is kept. The buffer cache copy is always completely up-to-date.
1506 * The second copy, which is used only as a source for disk writes,
1507 * contains only the safe pointers (i.e., those that have no remaining
1508 * update dependencies). The second copy is freed when all pointers
1509 * are safe. The cache is not allowed to replace indirect blocks with
1510 * pending update dependencies. If a buffer containing an indirect
1511 * block with dependencies is written, these routines will mark it
1512 * dirty again. It can only be successfully written once all the
1513 * dependencies are removed. The ffs_fsync routine in conjunction with
1514 * softdep_sync_metadata work together to get all the dependencies
1515 * removed so that a file can be successfully written to disk. Three
1516 * procedures are used when setting up indirect block pointer
1517 * dependencies. The division is necessary because of the organization
1518 * of the "balloc" routine and because of the distinction between file
1519 * pages and file metadata blocks.
1520 */
1521
1522/*
1523 * Allocate a new allocindir structure.
1524 */
1525static struct allocindir *
1526newallocindir(ip, ptrno, newblkno, oldblkno)
1527 struct inode *ip; /* inode for file being extended */
1528 int ptrno; /* offset of pointer in indirect block */
1529 ufs_daddr_t newblkno; /* disk block number being added */
1530 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1531{
1532 struct allocindir *aip;
1533
1534 MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1535 M_ALLOCINDIR, M_SOFTDEP_FLAGS);
1536 bzero(aip, sizeof(struct allocindir));
1537 aip->ai_list.wk_type = D_ALLOCINDIR;
1538 aip->ai_state = ATTACHED;
1539 aip->ai_offset = ptrno;
1540 aip->ai_newblkno = newblkno;
1541 aip->ai_oldblkno = oldblkno;
1542 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1543 return (aip);
1544}
1545
1546/*
1547 * Called just before setting an indirect block pointer
1548 * to a newly allocated file page.
1549 */
1550void
1551softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1552 struct inode *ip; /* inode for file being extended */
1553 ufs_lbn_t lbn; /* allocated block number within file */
1554 struct buf *bp; /* buffer with indirect blk referencing page */
1555 int ptrno; /* offset of pointer in indirect block */
1556 ufs_daddr_t newblkno; /* disk block number being added */
1557 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1558 struct buf *nbp; /* buffer holding allocated page */
1559{
1560 struct allocindir *aip;
1561 struct pagedep *pagedep;
1562
1563 aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1564 ACQUIRE_LOCK(&lk);
1565 /*
1566 * If we are allocating a directory page, then we must
1567 * allocate an associated pagedep to track additions and
1568 * deletions.
1569 */
1570 if ((ip->i_mode & IFMT) == IFDIR &&
1571 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1572 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1573 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1574 FREE_LOCK(&lk);
1575 setup_allocindir_phase2(bp, ip, aip);
1576}
1577
1578/*
1579 * Called just before setting an indirect block pointer to a
1580 * newly allocated indirect block.
1581 */
1582void
1583softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1584 struct buf *nbp; /* newly allocated indirect block */
1585 struct inode *ip; /* inode for file being extended */
1586 struct buf *bp; /* indirect block referencing allocated block */
1587 int ptrno; /* offset of pointer in indirect block */
1588 ufs_daddr_t newblkno; /* disk block number being added */
1589{
1590 struct allocindir *aip;
1591
1592 aip = newallocindir(ip, ptrno, newblkno, 0);
1593 ACQUIRE_LOCK(&lk);
1594 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1595 FREE_LOCK(&lk);
1596 setup_allocindir_phase2(bp, ip, aip);
1597}
1598
1599/*
1600 * Called to finish the allocation of the "aip" allocated
1601 * by one of the two routines above.
1602 */
1603static void
1604setup_allocindir_phase2(bp, ip, aip)
1605 struct buf *bp; /* in-memory copy of the indirect block */
1606 struct inode *ip; /* inode for file being extended */
1607 struct allocindir *aip; /* allocindir allocated by the above routines */
1608{
1609 struct worklist *wk;
1610 struct indirdep *indirdep, *newindirdep;
1611 struct bmsafemap *bmsafemap;
1612 struct allocindir *oldaip;
1613 struct freefrag *freefrag;
1614 struct newblk *newblk;
1615
1616 if (bp->b_lblkno >= 0)
1617 panic("setup_allocindir_phase2: not indir blk");
1618 for (indirdep = NULL, newindirdep = NULL; ; ) {
1619 ACQUIRE_LOCK(&lk);
1620 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1621 if (wk->wk_type != D_INDIRDEP)
1622 continue;
1623 indirdep = WK_INDIRDEP(wk);
1624 break;
1625 }
1626 if (indirdep == NULL && newindirdep) {
1627 indirdep = newindirdep;
1628 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1629 newindirdep = NULL;
1630 }
1631 FREE_LOCK(&lk);
1632 if (indirdep) {
1633 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1634 &newblk) == 0)
1635 panic("setup_allocindir: lost block");
1636 ACQUIRE_LOCK(&lk);
1637 if (newblk->nb_state == DEPCOMPLETE) {
1638 aip->ai_state |= DEPCOMPLETE;
1639 aip->ai_buf = NULL;
1640 } else {
1641 bmsafemap = newblk->nb_bmsafemap;
1642 aip->ai_buf = bmsafemap->sm_buf;
1643 LIST_REMOVE(newblk, nb_deps);
1644 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1645 aip, ai_deps);
1646 }
1647 LIST_REMOVE(newblk, nb_hash);
1648 FREE(newblk, M_NEWBLK);
1649 aip->ai_indirdep = indirdep;
1650 /*
1651 * Check to see if there is an existing dependency
1652 * for this block. If there is, merge the old
1653 * dependency into the new one.
1654 */
1655 if (aip->ai_oldblkno == 0)
1656 oldaip = NULL;
1657 else
1658
1659 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1660 if (oldaip->ai_offset == aip->ai_offset)
1661 break;
1662 if (oldaip != NULL) {
1663 if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1664 FREE_LOCK(&lk);
1665 panic("setup_allocindir_phase2: blkno");
1666 }
1667 aip->ai_oldblkno = oldaip->ai_oldblkno;
1668 freefrag = oldaip->ai_freefrag;
1669 oldaip->ai_freefrag = aip->ai_freefrag;
1670 aip->ai_freefrag = freefrag;
1671 free_allocindir(oldaip, NULL);
1672 }
1673 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1674 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1675 [aip->ai_offset] = aip->ai_oldblkno;
1676 FREE_LOCK(&lk);
1677 }
1678 if (newindirdep) {
1679 if (indirdep->ir_savebp != NULL)
1680 brelse(newindirdep->ir_savebp);
1681 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1682 }
1683 if (indirdep)
1684 break;
1685 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1686 M_INDIRDEP, M_SOFTDEP_FLAGS);
1687 newindirdep->ir_list.wk_type = D_INDIRDEP;
1688 newindirdep->ir_state = ATTACHED;
1689 LIST_INIT(&newindirdep->ir_deplisthd);
1690 LIST_INIT(&newindirdep->ir_donehd);
1691 if (bp->b_blkno == bp->b_lblkno) {
1692 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
1693 NULL, NULL);
1694 }
1695 newindirdep->ir_savebp =
1696 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0);
1697 BUF_KERNPROC(newindirdep->ir_savebp);
1698 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1699 }
1700}
1701
1702/*
1703 * Block de-allocation dependencies.
1704 *
1705 * When blocks are de-allocated, the on-disk pointers must be nullified before
1706 * the blocks are made available for use by other files. (The true
1707 * requirement is that old pointers must be nullified before new on-disk
1708 * pointers are set. We chose this slightly more stringent requirement to
1709 * reduce complexity.) Our implementation handles this dependency by updating
1710 * the inode (or indirect block) appropriately but delaying the actual block
1711 * de-allocation (i.e., freemap and free space count manipulation) until
1712 * after the updated versions reach stable storage. After the disk is
1713 * updated, the blocks can be safely de-allocated whenever it is convenient.
1714 * This implementation handles only the common case of reducing a file's
1715 * length to zero. Other cases are handled by the conventional synchronous
1716 * write approach.
1717 *
1718 * The ffs implementation with which we worked double-checks
1719 * the state of the block pointers and file size as it reduces
1720 * a file's length. Some of this code is replicated here in our
1721 * soft updates implementation. The freeblks->fb_chkcnt field is
1722 * used to transfer a part of this information to the procedure
1723 * that eventually de-allocates the blocks.
1724 *
1725 * This routine should be called from the routine that shortens
1726 * a file's length, before the inode's size or block pointers
1727 * are modified. It will save the block pointer information for
1728 * later release and zero the inode so that the calling routine
1729 * can release it.
1730 */
1731void
1732softdep_setup_freeblocks(ip, length)
1733 struct inode *ip; /* The inode whose length is to be reduced */
1734 off_t length; /* The new length for the file */
1735{
1736 struct freeblks *freeblks;
1737 struct inodedep *inodedep;
1738 struct allocdirect *adp;
1739 struct vnode *vp;
1740 struct buf *bp;
1741 struct fs *fs;
1742 int i, error, delay;
1743
1744 fs = ip->i_fs;
1745 if (length != 0)
1746 panic("softde_setup_freeblocks: non-zero length");
1747 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1748 M_FREEBLKS, M_SOFTDEP_FLAGS);
1749 bzero(freeblks, sizeof(struct freeblks));
1750 freeblks->fb_list.wk_type = D_FREEBLKS;
1751 freeblks->fb_uid = ip->i_uid;
1752 freeblks->fb_previousinum = ip->i_number;
1753 freeblks->fb_devvp = ip->i_devvp;
1754 freeblks->fb_fs = fs;
1755 freeblks->fb_oldsize = ip->i_size;
1756 freeblks->fb_newsize = length;
1757 freeblks->fb_chkcnt = ip->i_blocks;
1758 for (i = 0; i < NDADDR; i++) {
1759 freeblks->fb_dblks[i] = ip->i_db[i];
1760 ip->i_db[i] = 0;
1761 }
1762 for (i = 0; i < NIADDR; i++) {
1763 freeblks->fb_iblks[i] = ip->i_ib[i];
1764 ip->i_ib[i] = 0;
1765 }
1766 ip->i_blocks = 0;
1767 ip->i_size = 0;
1768 /*
1769 * Push the zero'ed inode to to its disk buffer so that we are free
1770 * to delete its dependencies below. Once the dependencies are gone
1771 * the buffer can be safely released.
1772 */
1773 if ((error = bread(ip->i_devvp,
1774 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
3b568787 1775 (int)fs->fs_bsize, &bp)) != 0)
984263bc
MD
1776 softdep_error("softdep_setup_freeblocks", error);
1777 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
1778 ip->i_din;
1779 /*
1780 * Find and eliminate any inode dependencies.
1781 */
1782 ACQUIRE_LOCK(&lk);
1783 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1784 if ((inodedep->id_state & IOSTARTED) != 0) {
1785 FREE_LOCK(&lk);
1786 panic("softdep_setup_freeblocks: inode busy");
1787 }
1788 /*
1789 * Add the freeblks structure to the list of operations that
1790 * must await the zero'ed inode being written to disk. If we
1791 * still have a bitmap dependency (delay == 0), then the inode
1792 * has never been written to disk, so we can process the
1793 * freeblks below once we have deleted the dependencies.
1794 */
1795 delay = (inodedep->id_state & DEPCOMPLETE);
1796 if (delay)
1797 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1798 /*
1799 * Because the file length has been truncated to zero, any
1800 * pending block allocation dependency structures associated
1801 * with this inode are obsolete and can simply be de-allocated.
1802 * We must first merge the two dependency lists to get rid of
1803 * any duplicate freefrag structures, then purge the merged list.
1804 */
1805 merge_inode_lists(inodedep);
1806 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1807 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
1808 FREE_LOCK(&lk);
1809 bdwrite(bp);
1810 /*
1811 * We must wait for any I/O in progress to finish so that
1812 * all potential buffers on the dirty list will be visible.
1813 * Once they are all there, walk the list and get rid of
1814 * any dependencies.
1815 */
1816 vp = ITOV(ip);
1817 ACQUIRE_LOCK(&lk);
1818 drain_output(vp, 1);
1819 while (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) {
1820 bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
1821 (void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
1822 deallocate_dependencies(bp, inodedep);
1823 bp->b_flags |= B_INVAL | B_NOCACHE;
1824 FREE_LOCK(&lk);
1825 brelse(bp);
1826 ACQUIRE_LOCK(&lk);
1827 }
1828 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
1829 (void)free_inodedep(inodedep);
1830 FREE_LOCK(&lk);
1831 /*
1832 * If the inode has never been written to disk (delay == 0),
1833 * then we can process the freeblks now that we have deleted
1834 * the dependencies.
1835 */
1836 if (!delay)
1837 handle_workitem_freeblocks(freeblks);
1838}
1839
1840/*
1841 * Reclaim any dependency structures from a buffer that is about to
1842 * be reallocated to a new vnode. The buffer must be locked, thus,
1843 * no I/O completion operations can occur while we are manipulating
1844 * its associated dependencies. The mutex is held so that other I/O's
1845 * associated with related dependencies do not occur.
1846 */
1847static void
1848deallocate_dependencies(bp, inodedep)
1849 struct buf *bp;
1850 struct inodedep *inodedep;
1851{
1852 struct worklist *wk;
1853 struct indirdep *indirdep;
1854 struct allocindir *aip;
1855 struct pagedep *pagedep;
1856 struct dirrem *dirrem;
1857 struct diradd *dap;
1858 int i;
1859
1860 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1861 switch (wk->wk_type) {
1862
1863 case D_INDIRDEP:
1864 indirdep = WK_INDIRDEP(wk);
1865 /*
1866 * None of the indirect pointers will ever be visible,
1867 * so they can simply be tossed. GOINGAWAY ensures
1868 * that allocated pointers will be saved in the buffer
1869 * cache until they are freed. Note that they will
1870 * only be able to be found by their physical address
1871 * since the inode mapping the logical address will
1872 * be gone. The save buffer used for the safe copy
1873 * was allocated in setup_allocindir_phase2 using
1874 * the physical address so it could be used for this
1875 * purpose. Hence we swap the safe copy with the real
1876 * copy, allowing the safe copy to be freed and holding
1877 * on to the real copy for later use in indir_trunc.
1878 */
1879 if (indirdep->ir_state & GOINGAWAY) {
1880 FREE_LOCK(&lk);
1881 panic("deallocate_dependencies: already gone");
1882 }
1883 indirdep->ir_state |= GOINGAWAY;
1884 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
1885 free_allocindir(aip, inodedep);
1886 if (bp->b_lblkno >= 0 ||
1887 bp->b_blkno != indirdep->ir_savebp->b_lblkno) {
1888 FREE_LOCK(&lk);
1889 panic("deallocate_dependencies: not indir");
1890 }
1891 bcopy(bp->b_data, indirdep->ir_savebp->b_data,
1892 bp->b_bcount);
1893 WORKLIST_REMOVE(wk);
1894 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
1895 continue;
1896
1897 case D_PAGEDEP:
1898 pagedep = WK_PAGEDEP(wk);
1899 /*
1900 * None of the directory additions will ever be
1901 * visible, so they can simply be tossed.
1902 */
1903 for (i = 0; i < DAHASHSZ; i++)
1904 while ((dap =
1905 LIST_FIRST(&pagedep->pd_diraddhd[i])))
1906 free_diradd(dap);
1907 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
1908 free_diradd(dap);
1909 /*
1910 * Copy any directory remove dependencies to the list
1911 * to be processed after the zero'ed inode is written.
1912 * If the inode has already been written, then they
1913 * can be dumped directly onto the work list.
1914 */
1915 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
1916 LIST_REMOVE(dirrem, dm_next);
1917 dirrem->dm_dirinum = pagedep->pd_ino;
1918 if (inodedep == NULL ||
1919 (inodedep->id_state & ALLCOMPLETE) ==
1920 ALLCOMPLETE)
1921 add_to_worklist(&dirrem->dm_list);
1922 else
1923 WORKLIST_INSERT(&inodedep->id_bufwait,
1924 &dirrem->dm_list);
1925 }
1926 WORKLIST_REMOVE(&pagedep->pd_list);
1927 LIST_REMOVE(pagedep, pd_hash);
1928 WORKITEM_FREE(pagedep, D_PAGEDEP);
1929 continue;
1930
1931 case D_ALLOCINDIR:
1932 free_allocindir(WK_ALLOCINDIR(wk), inodedep);
1933 continue;
1934
1935 case D_ALLOCDIRECT:
1936 case D_INODEDEP:
1937 FREE_LOCK(&lk);
1938 panic("deallocate_dependencies: Unexpected type %s",
1939 TYPENAME(wk->wk_type));
1940 /* NOTREACHED */
1941
1942 default:
1943 FREE_LOCK(&lk);
1944 panic("deallocate_dependencies: Unknown type %s",
1945 TYPENAME(wk->wk_type));
1946 /* NOTREACHED */
1947 }
1948 }
1949}
1950
1951/*
1952 * Free an allocdirect. Generate a new freefrag work request if appropriate.
1953 * This routine must be called with splbio interrupts blocked.
1954 */
1955static void
1956free_allocdirect(adphead, adp, delay)
1957 struct allocdirectlst *adphead;
1958 struct allocdirect *adp;
1959 int delay;
1960{
1961
1962#ifdef DEBUG
dadab5e9 1963 if (lk.lkt_held == NOHOLDER)
984263bc
MD
1964 panic("free_allocdirect: lock not held");
1965#endif
1966 if ((adp->ad_state & DEPCOMPLETE) == 0)
1967 LIST_REMOVE(adp, ad_deps);
1968 TAILQ_REMOVE(adphead, adp, ad_next);
1969 if ((adp->ad_state & COMPLETE) == 0)
1970 WORKLIST_REMOVE(&adp->ad_list);
1971 if (adp->ad_freefrag != NULL) {
1972 if (delay)
1973 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
1974 &adp->ad_freefrag->ff_list);
1975 else
1976 add_to_worklist(&adp->ad_freefrag->ff_list);
1977 }
1978 WORKITEM_FREE(adp, D_ALLOCDIRECT);
1979}
1980
1981/*
1982 * Prepare an inode to be freed. The actual free operation is not
1983 * done until the zero'ed inode has been written to disk.
1984 */
1985void
1986softdep_freefile(pvp, ino, mode)
1987 struct vnode *pvp;
1988 ino_t ino;
1989 int mode;
1990{
1991 struct inode *ip = VTOI(pvp);
1992 struct inodedep *inodedep;
1993 struct freefile *freefile;
1994
1995 /*
1996 * This sets up the inode de-allocation dependency.
1997 */
1998 MALLOC(freefile, struct freefile *, sizeof(struct freefile),
1999 M_FREEFILE, M_SOFTDEP_FLAGS);
2000 freefile->fx_list.wk_type = D_FREEFILE;
2001 freefile->fx_list.wk_state = 0;
2002 freefile->fx_mode = mode;
2003 freefile->fx_oldinum = ino;
2004 freefile->fx_devvp = ip->i_devvp;
2005 freefile->fx_fs = ip->i_fs;
2006
2007 /*
2008 * If the inodedep does not exist, then the zero'ed inode has
2009 * been written to disk. If the allocated inode has never been
2010 * written to disk, then the on-disk inode is zero'ed. In either
2011 * case we can free the file immediately.
2012 */
2013 ACQUIRE_LOCK(&lk);
2014 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2015 check_inode_unwritten(inodedep)) {
2016 FREE_LOCK(&lk);
2017 handle_workitem_freefile(freefile);
2018 return;
2019 }
2020 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2021 FREE_LOCK(&lk);
2022}
2023
2024/*
2025 * Check to see if an inode has never been written to disk. If
2026 * so free the inodedep and return success, otherwise return failure.
2027 * This routine must be called with splbio interrupts blocked.
2028 *
2029 * If we still have a bitmap dependency, then the inode has never
2030 * been written to disk. Drop the dependency as it is no longer
2031 * necessary since the inode is being deallocated. We set the
2032 * ALLCOMPLETE flags since the bitmap now properly shows that the
2033 * inode is not allocated. Even if the inode is actively being
2034 * written, it has been rolled back to its zero'ed state, so we
2035 * are ensured that a zero inode is what is on the disk. For short
2036 * lived files, this change will usually result in removing all the
2037 * dependencies from the inode so that it can be freed immediately.
2038 */
2039static int
2040check_inode_unwritten(inodedep)
2041 struct inodedep *inodedep;
2042{
2043
2044 if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2045 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2046 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2047 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2048 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2049 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2050 inodedep->id_nlinkdelta != 0)
2051 return (0);
2052 inodedep->id_state |= ALLCOMPLETE;
2053 LIST_REMOVE(inodedep, id_deps);
2054 inodedep->id_buf = NULL;
2055 if (inodedep->id_state & ONWORKLIST)
2056 WORKLIST_REMOVE(&inodedep->id_list);
2057 if (inodedep->id_savedino != NULL) {
2058 FREE(inodedep->id_savedino, M_INODEDEP);
2059 inodedep->id_savedino = NULL;
2060 }
2061 if (free_inodedep(inodedep) == 0) {
2062 FREE_LOCK(&lk);
2063 panic("check_inode_unwritten: busy inode");
2064 }
2065 return (1);
2066}
2067
2068/*
2069 * Try to free an inodedep structure. Return 1 if it could be freed.
2070 */
2071static int
2072free_inodedep(inodedep)
2073 struct inodedep *inodedep;
2074{
2075
2076 if ((inodedep->id_state & ONWORKLIST) != 0 ||
2077 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2078 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2079 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2080 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2081 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2082 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2083 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
2084 return (0);
2085 LIST_REMOVE(inodedep, id_hash);
2086 WORKITEM_FREE(inodedep, D_INODEDEP);
2087 num_inodedep -= 1;
2088 return (1);
2089}
2090
2091/*
2092 * This workitem routine performs the block de-allocation.
2093 * The workitem is added to the pending list after the updated
2094 * inode block has been written to disk. As mentioned above,
2095 * checks regarding the number of blocks de-allocated (compared
2096 * to the number of blocks allocated for the file) are also
2097 * performed in this function.
2098 */
2099static void
2100handle_workitem_freeblocks(freeblks)
2101 struct freeblks *freeblks;
2102{
2103 struct inode tip;
2104 ufs_daddr_t bn;
2105 struct fs *fs;
2106 int i, level, bsize;
2107 long nblocks, blocksreleased = 0;
2108 int error, allerror = 0;
2109 ufs_lbn_t baselbns[NIADDR], tmpval;
2110
2111 tip.i_number = freeblks->fb_previousinum;
2112 tip.i_devvp = freeblks->fb_devvp;
2113 tip.i_dev = freeblks->fb_devvp->v_rdev;
2114 tip.i_fs = freeblks->fb_fs;
2115 tip.i_size = freeblks->fb_oldsize;
2116 tip.i_uid = freeblks->fb_uid;
2117 fs = freeblks->fb_fs;
2118 tmpval = 1;
2119 baselbns[0] = NDADDR;
2120 for (i = 1; i < NIADDR; i++) {
2121 tmpval *= NINDIR(fs);
2122 baselbns[i] = baselbns[i - 1] + tmpval;
2123 }
2124 nblocks = btodb(fs->fs_bsize);
2125 blocksreleased = 0;
2126 /*
2127 * Indirect blocks first.
2128 */
2129 for (level = (NIADDR - 1); level >= 0; level--) {
2130 if ((bn = freeblks->fb_iblks[level]) == 0)
2131 continue;
2132 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level,
2133 baselbns[level], &blocksreleased)) == 0)
2134 allerror = error;
2135 ffs_blkfree(&tip, bn, fs->fs_bsize);
2136 blocksreleased += nblocks;
2137 }
2138 /*
2139 * All direct blocks or frags.
2140 */
2141 for (i = (NDADDR - 1); i >= 0; i--) {
2142 if ((bn = freeblks->fb_dblks[i]) == 0)
2143 continue;
2144 bsize = blksize(fs, &tip, i);
2145 ffs_blkfree(&tip, bn, bsize);
2146 blocksreleased += btodb(bsize);
2147 }
2148
2149#ifdef DIAGNOSTIC
2150 if (freeblks->fb_chkcnt != blocksreleased)
2151 printf("handle_workitem_freeblocks: block count\n");
2152 if (allerror)
2153 softdep_error("handle_workitem_freeblks", allerror);
2154#endif /* DIAGNOSTIC */
2155 WORKITEM_FREE(freeblks, D_FREEBLKS);
2156}
2157
2158/*
2159 * Release blocks associated with the inode ip and stored in the indirect
2160 * block dbn. If level is greater than SINGLE, the block is an indirect block
2161 * and recursive calls to indirtrunc must be used to cleanse other indirect
2162 * blocks.
2163 */
2164static int
2165indir_trunc(ip, dbn, level, lbn, countp)
2166 struct inode *ip;
2167 ufs_daddr_t dbn;
2168 int level;
2169 ufs_lbn_t lbn;
2170 long *countp;
2171{
2172 struct buf *bp;
2173 ufs_daddr_t *bap;
2174 ufs_daddr_t nb;
2175 struct fs *fs;
2176 struct worklist *wk;
2177 struct indirdep *indirdep;
2178 int i, lbnadd, nblocks;
2179 int error, allerror = 0;
2180
2181 fs = ip->i_fs;
2182 lbnadd = 1;
2183 for (i = level; i > 0; i--)
2184 lbnadd *= NINDIR(fs);
2185 /*
2186 * Get buffer of block pointers to be freed. This routine is not
2187 * called until the zero'ed inode has been written, so it is safe
2188 * to free blocks as they are encountered. Because the inode has
2189 * been zero'ed, calls to bmap on these blocks will fail. So, we
2190 * have to use the on-disk address and the block device for the
2191 * filesystem to look them up. If the file was deleted before its
2192 * indirect blocks were all written to disk, the routine that set
2193 * us up (deallocate_dependencies) will have arranged to leave
2194 * a complete copy of the indirect block in memory for our use.
2195 * Otherwise we have to read the blocks in from the disk.
2196 */
2197 ACQUIRE_LOCK(&lk);
2198 if ((bp = incore(ip->i_devvp, dbn)) != NULL &&
2199 (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2200 if (wk->wk_type != D_INDIRDEP ||
2201 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2202 (indirdep->ir_state & GOINGAWAY) == 0) {
2203 FREE_LOCK(&lk);
2204 panic("indir_trunc: lost indirdep");
2205 }
2206 WORKLIST_REMOVE(wk);
2207 WORKITEM_FREE(indirdep, D_INDIRDEP);
2208 if (LIST_FIRST(&bp->b_dep) != NULL) {
2209 FREE_LOCK(&lk);
2210 panic("indir_trunc: dangling dep");
2211 }
2212 FREE_LOCK(&lk);
2213 } else {
2214 FREE_LOCK(&lk);
3b568787 2215 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, &bp);
984263bc
MD
2216 if (error)
2217 return (error);
2218 }
2219 /*
2220 * Recursively free indirect blocks.
2221 */
2222 bap = (ufs_daddr_t *)bp->b_data;
2223 nblocks = btodb(fs->fs_bsize);
2224 for (i = NINDIR(fs) - 1; i >= 0; i--) {
2225 if ((nb = bap[i]) == 0)
2226 continue;
2227 if (level != 0) {
2228 if ((error = indir_trunc(ip, fsbtodb(fs, nb),
2229 level - 1, lbn + (i * lbnadd), countp)) != 0)
2230 allerror = error;
2231 }
2232 ffs_blkfree(ip, nb, fs->fs_bsize);
2233 *countp += nblocks;
2234 }
2235 bp->b_flags |= B_INVAL | B_NOCACHE;
2236 brelse(bp);
2237 return (allerror);
2238}
2239
2240/*
2241 * Free an allocindir.
2242 * This routine must be called with splbio interrupts blocked.
2243 */
2244static void
2245free_allocindir(aip, inodedep)
2246 struct allocindir *aip;
2247 struct inodedep *inodedep;
2248{
2249 struct freefrag *freefrag;
2250
2251#ifdef DEBUG
dadab5e9 2252 if (lk.lkt_held == NOHOLDER)
984263bc
MD
2253 panic("free_allocindir: lock not held");
2254#endif
2255 if ((aip->ai_state & DEPCOMPLETE) == 0)
2256 LIST_REMOVE(aip, ai_deps);
2257 if (aip->ai_state & ONWORKLIST)
2258 WORKLIST_REMOVE(&aip->ai_list);
2259 LIST_REMOVE(aip, ai_next);
2260 if ((freefrag = aip->ai_freefrag) != NULL) {
2261 if (inodedep == NULL)
2262 add_to_worklist(&freefrag->ff_list);
2263 else
2264 WORKLIST_INSERT(&inodedep->id_bufwait,
2265 &freefrag->ff_list);
2266 }
2267 WORKITEM_FREE(aip, D_ALLOCINDIR);
2268}
2269
2270/*
2271 * Directory entry addition dependencies.
2272 *
2273 * When adding a new directory entry, the inode (with its incremented link
2274 * count) must be written to disk before the directory entry's pointer to it.
2275 * Also, if the inode is newly allocated, the corresponding freemap must be
2276 * updated (on disk) before the directory entry's pointer. These requirements
2277 * are met via undo/redo on the directory entry's pointer, which consists
2278 * simply of the inode number.
2279 *
2280 * As directory entries are added and deleted, the free space within a
2281 * directory block can become fragmented. The ufs file system will compact
2282 * a fragmented directory block to make space for a new entry. When this
2283 * occurs, the offsets of previously added entries change. Any "diradd"
2284 * dependency structures corresponding to these entries must be updated with
2285 * the new offsets.
2286 */
2287
2288/*
2289 * This routine is called after the in-memory inode's link
2290 * count has been incremented, but before the directory entry's
2291 * pointer to the inode has been set.
2292 */
2293void
2294softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp)
2295 struct buf *bp; /* buffer containing directory block */
2296 struct inode *dp; /* inode for directory */
2297 off_t diroffset; /* offset of new entry in directory */
2298 long newinum; /* inode referenced by new directory entry */
2299 struct buf *newdirbp; /* non-NULL => contents of new mkdir */
2300{
2301 int offset; /* offset of new entry within directory block */
2302 ufs_lbn_t lbn; /* block in directory containing new entry */
2303 struct fs *fs;
2304 struct diradd *dap;
2305 struct pagedep *pagedep;
2306 struct inodedep *inodedep;
2307 struct mkdir *mkdir1, *mkdir2;
2308
2309 /*
2310 * Whiteouts have no dependencies.
2311 */
2312 if (newinum == WINO) {
2313 if (newdirbp != NULL)
2314 bdwrite(newdirbp);
2315 return;
2316 }
2317
2318 fs = dp->i_fs;
2319 lbn = lblkno(fs, diroffset);
2320 offset = blkoff(fs, diroffset);
2321 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2322 M_SOFTDEP_FLAGS);
2323 bzero(dap, sizeof(struct diradd));
2324 dap->da_list.wk_type = D_DIRADD;
2325 dap->da_offset = offset;
2326 dap->da_newinum = newinum;
2327 dap->da_state = ATTACHED;
2328 if (newdirbp == NULL) {
2329 dap->da_state |= DEPCOMPLETE;
2330 ACQUIRE_LOCK(&lk);
2331 } else {
2332 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2333 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2334 M_SOFTDEP_FLAGS);
2335 mkdir1->md_list.wk_type = D_MKDIR;
2336 mkdir1->md_state = MKDIR_BODY;
2337 mkdir1->md_diradd = dap;
2338 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2339 M_SOFTDEP_FLAGS);
2340 mkdir2->md_list.wk_type = D_MKDIR;
2341 mkdir2->md_state = MKDIR_PARENT;
2342 mkdir2->md_diradd = dap;
2343 /*
2344 * Dependency on "." and ".." being written to disk.
2345 */
2346 mkdir1->md_buf = newdirbp;
2347 ACQUIRE_LOCK(&lk);
2348 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2349 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2350 FREE_LOCK(&lk);
2351 bdwrite(newdirbp);
2352 /*
2353 * Dependency on link count increase for parent directory
2354 */
2355 ACQUIRE_LOCK(&lk);
2356 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2357 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2358 dap->da_state &= ~MKDIR_PARENT;
2359 WORKITEM_FREE(mkdir2, D_MKDIR);
2360 } else {
2361 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2362 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2363 }
2364 }
2365 /*
2366 * Link into parent directory pagedep to await its being written.
2367 */
2368 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2369 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2370 dap->da_pagedep = pagedep;
2371 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2372 da_pdlist);
2373 /*
2374 * Link into its inodedep. Put it on the id_bufwait list if the inode
2375 * is not yet written. If it is written, do the post-inode write
2376 * processing to put it on the id_pendinghd list.
2377 */
2378 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2379 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2380 diradd_inode_written(dap, inodedep);
2381 else
2382 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2383 FREE_LOCK(&lk);
2384}
2385
2386/*
2387 * This procedure is called to change the offset of a directory
2388 * entry when compacting a directory block which must be owned
2389 * exclusively by the caller. Note that the actual entry movement
2390 * must be done in this procedure to ensure that no I/O completions
2391 * occur while the move is in progress.
2392 */
2393void
2394softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2395 struct inode *dp; /* inode for directory */
2396 caddr_t base; /* address of dp->i_offset */
2397 caddr_t oldloc; /* address of old directory location */
2398 caddr_t newloc; /* address of new directory location */
2399 int entrysize; /* size of directory entry */
2400{
2401 int offset, oldoffset, newoffset;
2402 struct pagedep *pagedep;
2403 struct diradd *dap;
2404 ufs_lbn_t lbn;
2405
2406 ACQUIRE_LOCK(&lk);
2407 lbn = lblkno(dp->i_fs, dp->i_offset);
2408 offset = blkoff(dp->i_fs, dp->i_offset);
2409 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2410 goto done;
2411 oldoffset = offset + (oldloc - base);
2412 newoffset = offset + (newloc - base);
2413
2414 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2415 if (dap->da_offset != oldoffset)
2416 continue;
2417 dap->da_offset = newoffset;
2418 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2419 break;
2420 LIST_REMOVE(dap, da_pdlist);
2421 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2422 dap, da_pdlist);
2423 break;
2424 }
2425 if (dap == NULL) {
2426
2427 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2428 if (dap->da_offset == oldoffset) {
2429 dap->da_offset = newoffset;
2430 break;
2431 }
2432 }
2433 }
2434done:
2435 bcopy(oldloc, newloc, entrysize);
2436 FREE_LOCK(&lk);
2437}
2438
2439/*
2440 * Free a diradd dependency structure. This routine must be called
2441 * with splbio interrupts blocked.
2442 */
2443static void
2444free_diradd(dap)
2445 struct diradd *dap;
2446{
2447 struct dirrem *dirrem;
2448 struct pagedep *pagedep;
2449 struct inodedep *inodedep;
2450 struct mkdir *mkdir, *nextmd;
2451
2452#ifdef DEBUG
dadab5e9 2453 if (lk.lkt_held == NOHOLDER)
984263bc
MD
2454 panic("free_diradd: lock not held");
2455#endif
2456 WORKLIST_REMOVE(&dap->da_list);
2457 LIST_REMOVE(dap, da_pdlist);
2458 if ((dap->da_state & DIRCHG) == 0) {
2459 pagedep = dap->da_pagedep;
2460 } else {
2461 dirrem = dap->da_previous;
2462 pagedep = dirrem->dm_pagedep;
2463 dirrem->dm_dirinum = pagedep->pd_ino;
2464 add_to_worklist(&dirrem->dm_list);
2465 }
2466 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2467 0, &inodedep) != 0)
2468 (void) free_inodedep(inodedep);
2469 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2470 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2471 nextmd = LIST_NEXT(mkdir, md_mkdirs);
2472 if (mkdir->md_diradd != dap)
2473 continue;
2474 dap->da_state &= ~mkdir->md_state;
2475 WORKLIST_REMOVE(&mkdir->md_list);
2476 LIST_REMOVE(mkdir, md_mkdirs);
2477 WORKITEM_FREE(mkdir, D_MKDIR);
2478 }
2479 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2480 FREE_LOCK(&lk);
2481 panic("free_diradd: unfound ref");
2482 }
2483 }
2484 WORKITEM_FREE(dap, D_DIRADD);
2485}
2486
2487/*
2488 * Directory entry removal dependencies.
2489 *
2490 * When removing a directory entry, the entry's inode pointer must be
2491 * zero'ed on disk before the corresponding inode's link count is decremented
2492 * (possibly freeing the inode for re-use). This dependency is handled by
2493 * updating the directory entry but delaying the inode count reduction until
2494 * after the directory block has been written to disk. After this point, the
2495 * inode count can be decremented whenever it is convenient.
2496 */
2497
2498/*
2499 * This routine should be called immediately after removing
2500 * a directory entry. The inode's link count should not be
2501 * decremented by the calling procedure -- the soft updates
2502 * code will do this task when it is safe.
2503 */
2504void
2505softdep_setup_remove(bp, dp, ip, isrmdir)
2506 struct buf *bp; /* buffer containing directory block */
2507 struct inode *dp; /* inode for the directory being modified */
2508 struct inode *ip; /* inode for directory entry being removed */
2509 int isrmdir; /* indicates if doing RMDIR */
2510{
2511 struct dirrem *dirrem, *prevdirrem;
2512
2513 /*
2514 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2515 */
2516 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2517
2518 /*
2519 * If the COMPLETE flag is clear, then there were no active
2520 * entries and we want to roll back to a zeroed entry until
2521 * the new inode is committed to disk. If the COMPLETE flag is
2522 * set then we have deleted an entry that never made it to
2523 * disk. If the entry we deleted resulted from a name change,
2524 * then the old name still resides on disk. We cannot delete
2525 * its inode (returned to us in prevdirrem) until the zeroed
2526 * directory entry gets to disk. The new inode has never been
2527 * referenced on the disk, so can be deleted immediately.
2528 */
2529 if ((dirrem->dm_state & COMPLETE) == 0) {
2530 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2531 dm_next);
2532 FREE_LOCK(&lk);
2533 } else {
2534 if (prevdirrem != NULL)
2535 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2536 prevdirrem, dm_next);
2537 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2538 FREE_LOCK(&lk);
2539 handle_workitem_remove(dirrem);
2540 }
2541}
2542
2543/*
2544 * Allocate a new dirrem if appropriate and return it along with
2545 * its associated pagedep. Called without a lock, returns with lock.
2546 */
2547static long num_dirrem; /* number of dirrem allocated */
2548static struct dirrem *
2549newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2550 struct buf *bp; /* buffer containing directory block */
2551 struct inode *dp; /* inode for the directory being modified */
2552 struct inode *ip; /* inode for directory entry being removed */
2553 int isrmdir; /* indicates if doing RMDIR */
2554 struct dirrem **prevdirremp; /* previously referenced inode, if any */
2555{
2556 int offset;
2557 ufs_lbn_t lbn;
2558 struct diradd *dap;
2559 struct dirrem *dirrem;
2560 struct pagedep *pagedep;
2561
2562 /*
2563 * Whiteouts have no deletion dependencies.
2564 */
2565 if (ip == NULL)
2566 panic("newdirrem: whiteout");
2567 /*
2568 * If we are over our limit, try to improve the situation.
2569 * Limiting the number of dirrem structures will also limit
2570 * the number of freefile and freeblks structures.
2571 */
2572 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2573 (void) request_cleanup(FLUSH_REMOVE, 0);
2574 num_dirrem += 1;
2575 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
2576 M_DIRREM, M_SOFTDEP_FLAGS);
2577 bzero(dirrem, sizeof(struct dirrem));
2578 dirrem->dm_list.wk_type = D_DIRREM;
2579 dirrem->dm_state = isrmdir ? RMDIR : 0;
2580 dirrem->dm_mnt = ITOV(ip)->v_mount;
2581 dirrem->dm_oldinum = ip->i_number;
2582 *prevdirremp = NULL;
2583
2584 ACQUIRE_LOCK(&lk);
2585 lbn = lblkno(dp->i_fs, dp->i_offset);
2586 offset = blkoff(dp->i_fs, dp->i_offset);
2587 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2588 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2589 dirrem->dm_pagedep = pagedep;
2590 /*
2591 * Check for a diradd dependency for the same directory entry.
2592 * If present, then both dependencies become obsolete and can
2593 * be de-allocated. Check for an entry on both the pd_dirraddhd
2594 * list and the pd_pendinghd list.
2595 */
2596
2597 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
2598 if (dap->da_offset == offset)
2599 break;
2600 if (dap == NULL) {
2601
2602 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
2603 if (dap->da_offset == offset)
2604 break;
2605 if (dap == NULL)
2606 return (dirrem);
2607 }
2608 /*
2609 * Must be ATTACHED at this point.
2610 */
2611 if ((dap->da_state & ATTACHED) == 0) {
2612 FREE_LOCK(&lk);
2613 panic("newdirrem: not ATTACHED");
2614 }
2615 if (dap->da_newinum != ip->i_number) {
2616 FREE_LOCK(&lk);
2617 panic("newdirrem: inum %d should be %d",
2618 ip->i_number, dap->da_newinum);
2619 }
2620 /*
2621 * If we are deleting a changed name that never made it to disk,
2622 * then return the dirrem describing the previous inode (which
2623 * represents the inode currently referenced from this entry on disk).
2624 */
2625 if ((dap->da_state & DIRCHG) != 0) {
2626 *prevdirremp = dap->da_previous;
2627 dap->da_state &= ~DIRCHG;
2628 dap->da_pagedep = pagedep;
2629 }
2630 /*
2631 * We are deleting an entry that never made it to disk.
2632 * Mark it COMPLETE so we can delete its inode immediately.
2633 */
2634 dirrem->dm_state |= COMPLETE;
2635 free_diradd(dap);
2636 return (dirrem);
2637}
2638
2639/*
2640 * Directory entry change dependencies.
2641 *
2642 * Changing an existing directory entry requires that an add operation
2643 * be completed first followed by a deletion. The semantics for the addition
2644 * are identical to the description of adding a new entry above except
2645 * that the rollback is to the old inode number rather than zero. Once
2646 * the addition dependency is completed, the removal is done as described
2647 * in the removal routine above.
2648 */
2649
2650/*
2651 * This routine should be called immediately after changing
2652 * a directory entry. The inode's link count should not be
2653 * decremented by the calling procedure -- the soft updates
2654 * code will perform this task when it is safe.
2655 */
2656void
2657softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
2658 struct buf *bp; /* buffer containing directory block */
2659 struct inode *dp; /* inode for the directory being modified */
2660 struct inode *ip; /* inode for directory entry being removed */
2661 long newinum; /* new inode number for changed entry */
2662 int isrmdir; /* indicates if doing RMDIR */
2663{
2664 int offset;
2665 struct diradd *dap = NULL;
2666 struct dirrem *dirrem, *prevdirrem;
2667 struct pagedep *pagedep;
2668 struct inodedep *inodedep;
2669
2670 offset = blkoff(dp->i_fs, dp->i_offset);
2671
2672 /*
2673 * Whiteouts do not need diradd dependencies.
2674 */
2675 if (newinum != WINO) {
2676 MALLOC(dap, struct diradd *, sizeof(struct diradd),
2677 M_DIRADD, M_SOFTDEP_FLAGS);
2678 bzero(dap, sizeof(struct diradd));
2679 dap->da_list.wk_type = D_DIRADD;
2680 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2681 dap->da_offset = offset;
2682 dap->da_newinum = newinum;
2683 }
2684
2685 /*
2686 * Allocate a new dirrem and ACQUIRE_LOCK.
2687 */
2688 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2689 pagedep = dirrem->dm_pagedep;
2690 /*
2691 * The possible values for isrmdir:
2692 * 0 - non-directory file rename
2693 * 1 - directory rename within same directory
2694 * inum - directory rename to new directory of given inode number
2695 * When renaming to a new directory, we are both deleting and
2696 * creating a new directory entry, so the link count on the new
2697 * directory should not change. Thus we do not need the followup
2698 * dirrem which is usually done in handle_workitem_remove. We set
2699 * the DIRCHG flag to tell handle_workitem_remove to skip the
2700 * followup dirrem.
2701 */
2702 if (isrmdir > 1)
2703 dirrem->dm_state |= DIRCHG;
2704
2705 /*
2706 * Whiteouts have no additional dependencies,
2707 * so just put the dirrem on the correct list.
2708 */
2709 if (newinum == WINO) {
2710 if ((dirrem->dm_state & COMPLETE) == 0) {
2711 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2712 dm_next);
2713 } else {
2714 dirrem->dm_dirinum = pagedep->pd_ino;
2715 add_to_worklist(&dirrem->dm_list);
2716 }
2717 FREE_LOCK(&lk);
2718 return;
2719 }
2720
2721 /*
2722 * If the COMPLETE flag is clear, then there were no active
2723 * entries and we want to roll back to the previous inode until
2724 * the new inode is committed to disk. If the COMPLETE flag is
2725 * set, then we have deleted an entry that never made it to disk.
2726 * If the entry we deleted resulted from a name change, then the old
2727 * inode reference still resides on disk. Any rollback that we do
2728 * needs to be to that old inode (returned to us in prevdirrem). If
2729 * the entry we deleted resulted from a create, then there is
2730 * no entry on the disk, so we want to roll back to zero rather
2731 * than the uncommitted inode. In either of the COMPLETE cases we
2732 * want to immediately free the unwritten and unreferenced inode.
2733 */
2734 if ((dirrem->dm_state & COMPLETE) == 0) {
2735 dap->da_previous = dirrem;
2736 } else {
2737 if (prevdirrem != NULL) {
2738 dap->da_previous = prevdirrem;
2739 } else {
2740 dap->da_state &= ~DIRCHG;
2741 dap->da_pagedep = pagedep;
2742 }
2743 dirrem->dm_dirinum = pagedep->pd_ino;
2744 add_to_worklist(&dirrem->dm_list);
2745 }
2746 /*
2747 * Link into its inodedep. Put it on the id_bufwait list if the inode
2748 * is not yet written. If it is written, do the post-inode write
2749 * processing to put it on the id_pendinghd list.
2750 */
2751 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2752 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2753 dap->da_state |= COMPLETE;
2754 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2755 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2756 } else {
2757 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2758 dap, da_pdlist);
2759 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2760 }
2761 FREE_LOCK(&lk);
2762}
2763
2764/*
2765 * Called whenever the link count on an inode is changed.
2766 * It creates an inode dependency so that the new reference(s)
2767 * to the inode cannot be committed to disk until the updated
2768 * inode has been written.
2769 */
2770void
2771softdep_change_linkcnt(ip)
2772 struct inode *ip; /* the inode with the increased link count */
2773{
2774 struct inodedep *inodedep;
2775
2776 ACQUIRE_LOCK(&lk);
2777 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2778 if (ip->i_nlink < ip->i_effnlink) {
2779 FREE_LOCK(&lk);
2780 panic("softdep_change_linkcnt: bad delta");
2781 }
2782 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2783 FREE_LOCK(&lk);
2784}
2785
2786/*
2787 * This workitem decrements the inode's link count.
2788 * If the link count reaches zero, the file is removed.
2789 */
2790static void
2791handle_workitem_remove(dirrem)
2792 struct dirrem *dirrem;
2793{
dadab5e9 2794 struct thread *td = curthread; /* XXX */
984263bc
MD
2795 struct inodedep *inodedep;
2796 struct vnode *vp;
2797 struct inode *ip;
2798 ino_t oldinum;
2799 int error;
2800
2801 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) {
2802 softdep_error("handle_workitem_remove: vget", error);
2803 return;
2804 }
2805 ip = VTOI(vp);
2806 ACQUIRE_LOCK(&lk);
2807 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
2808 FREE_LOCK(&lk);
2809 panic("handle_workitem_remove: lost inodedep");
2810 }
2811 /*
2812 * Normal file deletion.
2813 */
2814 if ((dirrem->dm_state & RMDIR) == 0) {
2815 ip->i_nlink--;
2816 ip->i_flag |= IN_CHANGE;
2817 if (ip->i_nlink < ip->i_effnlink) {
2818 FREE_LOCK(&lk);
2819 panic("handle_workitem_remove: bad file delta");
2820 }
2821 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2822 FREE_LOCK(&lk);
2823 vput(vp);
2824 num_dirrem -= 1;
2825 WORKITEM_FREE(dirrem, D_DIRREM);
2826 return;
2827 }
2828 /*
2829 * Directory deletion. Decrement reference count for both the
2830 * just deleted parent directory entry and the reference for ".".
2831 * Next truncate the directory to length zero. When the
2832 * truncation completes, arrange to have the reference count on
2833 * the parent decremented to account for the loss of "..".
2834 */
2835 ip->i_nlink -= 2;
2836 ip->i_flag |= IN_CHANGE;
2837 if (ip->i_nlink < ip->i_effnlink) {
2838 FREE_LOCK(&lk);
2839 panic("handle_workitem_remove: bad dir delta");
2840 }
2841 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2842 FREE_LOCK(&lk);
0cfcada1 2843 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, proc0.p_ucred, td)) != 0)
984263bc
MD
2844 softdep_error("handle_workitem_remove: truncate", error);
2845 /*
2846 * Rename a directory to a new parent. Since, we are both deleting
2847 * and creating a new directory entry, the link count on the new
2848 * directory should not change. Thus we skip the followup dirrem.
2849 */
2850 if (dirrem->dm_state & DIRCHG) {
2851 vput(vp);
2852 num_dirrem -= 1;
2853 WORKITEM_FREE(dirrem, D_DIRREM);
2854 return;
2855 }
2856 /*
2857 * If the inodedep does not exist, then the zero'ed inode has
2858 * been written to disk. If the allocated inode has never been
2859 * written to disk, then the on-disk inode is zero'ed. In either
2860 * case we can remove the file immediately.
2861 */
2862 ACQUIRE_LOCK(&lk);
2863 dirrem->dm_state = 0;
2864 oldinum = dirrem->dm_oldinum;
2865 dirrem->dm_oldinum = dirrem->dm_dirinum;
2866 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
2867 check_inode_unwritten(inodedep)) {
2868 FREE_LOCK(&lk);
2869 vput(vp);
2870 handle_workitem_remove(dirrem);
2871 return;
2872 }
2873 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
2874 FREE_LOCK(&lk);
2875 vput(vp);
2876}
2877
2878/*
2879 * Inode de-allocation dependencies.
2880 *
2881 * When an inode's link count is reduced to zero, it can be de-allocated. We
2882 * found it convenient to postpone de-allocation until after the inode is
2883 * written to disk with its new link count (zero). At this point, all of the
2884 * on-disk inode's block pointers are nullified and, with careful dependency
2885 * list ordering, all dependencies related to the inode will be satisfied and
2886 * the corresponding dependency structures de-allocated. So, if/when the
2887 * inode is reused, there will be no mixing of old dependencies with new
2888 * ones. This artificial dependency is set up by the block de-allocation
2889 * procedure above (softdep_setup_freeblocks) and completed by the
2890 * following procedure.
2891 */
2892static void
2893handle_workitem_freefile(freefile)
2894 struct freefile *freefile;
2895{
2896 struct vnode vp;
2897 struct inode tip;
2898 struct inodedep *idp;
2899 int error;
2900
2901#ifdef DEBUG
2902 ACQUIRE_LOCK(&lk);
2903 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp);
2904 FREE_LOCK(&lk);
2905 if (error)
2906 panic("handle_workitem_freefile: inodedep survived");
2907#endif
2908 tip.i_devvp = freefile->fx_devvp;
2909 tip.i_dev = freefile->fx_devvp->v_rdev;
2910 tip.i_fs = freefile->fx_fs;
2911 vp.v_data = &tip;
2912 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
2913 softdep_error("handle_workitem_freefile", error);
2914 WORKITEM_FREE(freefile, D_FREEFILE);
2915}
2916
2917/*
2918 * Disk writes.
2919 *
2920 * The dependency structures constructed above are most actively used when file
2921 * system blocks are written to disk. No constraints are placed on when a
2922 * block can be written, but unsatisfied update dependencies are made safe by
2923 * modifying (or replacing) the source memory for the duration of the disk
2924 * write. When the disk write completes, the memory block is again brought
2925 * up-to-date.
2926 *
2927 * In-core inode structure reclamation.
2928 *
2929 * Because there are a finite number of "in-core" inode structures, they are
2930 * reused regularly. By transferring all inode-related dependencies to the
2931 * in-memory inode block and indexing them separately (via "inodedep"s), we
2932 * can allow "in-core" inode structures to be reused at any time and avoid
2933 * any increase in contention.
2934 *
2935 * Called just before entering the device driver to initiate a new disk I/O.
2936 * The buffer must be locked, thus, no I/O completion operations can occur
2937 * while we are manipulating its associated dependencies.
2938 */
2939static void
2940softdep_disk_io_initiation(bp)
2941 struct buf *bp; /* structure describing disk write to occur */
2942{
2943 struct worklist *wk, *nextwk;
2944 struct indirdep *indirdep;
2945
2946 /*
2947 * We only care about write operations. There should never
2948 * be dependencies for reads.
2949 */
2950 if (bp->b_flags & B_READ)
2951 panic("softdep_disk_io_initiation: read");
2952 /*
2953 * Do any necessary pre-I/O processing.
2954 */
2955 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
2956 nextwk = LIST_NEXT(wk, wk_list);
2957 switch (wk->wk_type) {
2958
2959 case D_PAGEDEP:
2960 initiate_write_filepage(WK_PAGEDEP(wk), bp);
2961 continue;
2962
2963 case D_INODEDEP:
2964 initiate_write_inodeblock(WK_INODEDEP(wk), bp);
2965 continue;
2966
2967 case D_INDIRDEP:
2968 indirdep = WK_INDIRDEP(wk);
2969 if (indirdep->ir_state & GOINGAWAY)
2970 panic("disk_io_initiation: indirdep gone");
2971 /*
2972 * If there are no remaining dependencies, this
2973 * will be writing the real pointers, so the
2974 * dependency can be freed.
2975 */
2976 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
2977 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
2978 brelse(indirdep->ir_savebp);
2979 /* inline expand WORKLIST_REMOVE(wk); */
2980 wk->wk_state &= ~ONWORKLIST;
2981 LIST_REMOVE(wk, wk_list);
2982 WORKITEM_FREE(indirdep, D_INDIRDEP);
2983 continue;
2984 }
2985 /*
2986 * Replace up-to-date version with safe version.
2987 */
2988 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
2989 M_INDIRDEP, M_SOFTDEP_FLAGS);
2990 ACQUIRE_LOCK(&lk);
2991 indirdep->ir_state &= ~ATTACHED;
2992 indirdep->ir_state |= UNDONE;
2993 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
2994 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
2995 bp->b_bcount);
2996 FREE_LOCK(&lk);
2997 continue;
2998
2999 case D_MKDIR:
3000 case D_BMSAFEMAP:
3001 case D_ALLOCDIRECT:
3002 case D_ALLOCINDIR:
3003 continue;
3004
3005 default:
3006 panic("handle_disk_io_initiation: Unexpected type %s",
3007 TYPENAME(wk->wk_type));
3008 /* NOTREACHED */
3009 }
3010 }
3011}
3012
3013/*
3014 * Called from within the procedure above to deal with unsatisfied
3015 * allocation dependencies in a directory. The buffer must be locked,
3016 * thus, no I/O completion operations can occur while we are
3017 * manipulating its associated dependencies.
3018 */
3019static void
3020initiate_write_filepage(pagedep, bp)
3021 struct pagedep *pagedep;
3022 struct buf *bp;
3023{
3024 struct diradd *dap;
3025 struct direct *ep;
3026 int i;
3027
3028 if (pagedep->pd_state & IOSTARTED) {
3029 /*
3030 * This can only happen if there is a driver that does not
3031 * understand chaining. Here biodone will reissue the call
3032 * to strategy for the incomplete buffers.
3033 */
3034 printf("initiate_write_filepage: already started\n");
3035 return;
3036 }
3037 pagedep->pd_state |= IOSTARTED;
3038 ACQUIRE_LOCK(&lk);
3039 for (i = 0; i < DAHASHSZ; i++) {
3040 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3041 ep = (struct direct *)
3042 ((char *)bp->b_data + dap->da_offset);
3043 if (ep->d_ino != dap->da_newinum) {
3044 FREE_LOCK(&lk);
3045 panic("%s: dir inum %d != new %d",
3046 "initiate_write_filepage",
3047 ep->d_ino, dap->da_newinum);
3048 }
3049 if (dap->da_state & DIRCHG)
3050 ep->d_ino = dap->da_previous->dm_oldinum;
3051 else
3052 ep->d_ino = 0;
3053 dap->da_state &= ~ATTACHED;
3054 dap->da_state |= UNDONE;
3055 }
3056 }
3057 FREE_LOCK(&lk);
3058}
3059
3060/*
3061 * Called from within the procedure above to deal with unsatisfied
3062 * allocation dependencies in an inodeblock. The buffer must be
3063 * locked, thus, no I/O completion operations can occur while we
3064 * are manipulating its associated dependencies.
3065 */
3066static void
3067initiate_write_inodeblock(inodedep, bp)
3068 struct inodedep *inodedep;
3069 struct buf *bp; /* The inode block */
3070{
3071 struct allocdirect *adp, *lastadp;
3072 struct dinode *dp;
3073 struct fs *fs;
3074 ufs_lbn_t prevlbn = 0;
3075 int i, deplist;
3076
3077 if (inodedep->id_state & IOSTARTED)
3078 panic("initiate_write_inodeblock: already started");
3079 inodedep->id_state |= IOSTARTED;
3080 fs = inodedep->id_fs;
3081 dp = (struct dinode *)bp->b_data +
3082 ino_to_fsbo(fs, inodedep->id_ino);
3083 /*
3084 * If the bitmap is not yet written, then the allocated
3085 * inode cannot be written to disk.
3086 */
3087 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3088 if (inodedep->id_savedino != NULL)
3089 panic("initiate_write_inodeblock: already doing I/O");
3090 MALLOC(inodedep->id_savedino, struct dinode *,
3091 sizeof(struct dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3092 *inodedep->id_savedino = *dp;
3093 bzero((caddr_t)dp, sizeof(struct dinode));
3094 return;
3095 }
3096 /*
3097 * If no dependencies, then there is nothing to roll back.
3098 */
3099 inodedep->id_savedsize = dp->di_size;
3100 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3101 return;
3102 /*
3103 * Set the dependencies to busy.
3104 */
3105 ACQUIRE_LOCK(&lk);
3106 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3107 adp = TAILQ_NEXT(adp, ad_next)) {
3108#ifdef DIAGNOSTIC
3109 if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3110 FREE_LOCK(&lk);
3111 panic("softdep_write_inodeblock: lbn order");
3112 }
3113 prevlbn = adp->ad_lbn;
3114 if (adp->ad_lbn < NDADDR &&
3115 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3116 FREE_LOCK(&lk);
3117 panic("%s: direct pointer #%ld mismatch %d != %d",
3118 "softdep_write_inodeblock", adp->ad_lbn,
3119 dp->di_db[adp->ad_lbn], adp->ad_newblkno);
3120 }
3121 if (adp->ad_lbn >= NDADDR &&
3122 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3123 FREE_LOCK(&lk);
3124 panic("%s: indirect pointer #%ld mismatch %d != %d",
3125 "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
3126 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
3127 }
3128 deplist |= 1 << adp->ad_lbn;
3129 if ((adp->ad_state & ATTACHED) == 0) {
3130 FREE_LOCK(&lk);
3131 panic("softdep_write_inodeblock: Unknown state 0x%x",
3132 adp->ad_state);
3133 }
3134#endif /* DIAGNOSTIC */
3135 adp->ad_state &= ~ATTACHED;
3136 adp->ad_state |= UNDONE;
3137 }
3138 /*
3139 * The on-disk inode cannot claim to be any larger than the last
3140 * fragment that has been written. Otherwise, the on-disk inode
3141 * might have fragments that were not the last block in the file
3142 * which would corrupt the filesystem.
3143 */
3144 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3145 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3146 if (adp->ad_lbn >= NDADDR)
3147 break;
3148 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3149 /* keep going until hitting a rollback to a frag */
3150 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3151 continue;
3152 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3153 for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3154#ifdef DIAGNOSTIC
3155 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3156 FREE_LOCK(&lk);
3157 panic("softdep_write_inodeblock: lost dep1");
3158 }
3159#endif /* DIAGNOSTIC */
3160 dp->di_db[i] = 0;
3161 }
3162 for (i = 0; i < NIADDR; i++) {
3163#ifdef DIAGNOSTIC
3164 if (dp->di_ib[i] != 0 &&
3165 (deplist & ((1 << NDADDR) << i)) == 0) {
3166 FREE_LOCK(&lk);
3167 panic("softdep_write_inodeblock: lost dep2");
3168 }
3169#endif /* DIAGNOSTIC */
3170 dp->di_ib[i] = 0;
3171 }
3172 FREE_LOCK(&lk);
3173 return;
3174 }
3175 /*
3176 * If we have zero'ed out the last allocated block of the file,
3177 * roll back the size to the last currently allocated block.
3178 * We know that this last allocated block is a full-sized as
3179 * we already checked for fragments in the loop above.
3180 */
3181 if (lastadp != NULL &&
3182 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3183 for (i = lastadp->ad_lbn; i >= 0; i--)
3184 if (dp->di_db[i] != 0)
3185 break;
3186 dp->di_size = (i + 1) * fs->fs_bsize;
3187 }
3188 /*
3189 * The only dependencies are for indirect blocks.
3190 *
3191 * The file size for indirect block additions is not guaranteed.
3192 * Such a guarantee would be non-trivial to achieve. The conventional
3193 * synchronous write implementation also does not make this guarantee.
3194 * Fsck should catch and fix discrepancies. Arguably, the file size
3195 * can be over-estimated without destroying integrity when the file
3196 * moves into the indirect blocks (i.e., is large). If we want to
3197 * postpone fsck, we are stuck with this argument.
3198 */
3199 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3200 dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3201 FREE_LOCK(&lk);
3202}
3203
3204/*
3205 * This routine is called during the completion interrupt
3206 * service routine for a disk write (from the procedure called
3207 * by the device driver to inform the file system caches of
3208 * a request completion). It should be called early in this
3209 * procedure, before the block is made available to other
3210 * processes or other routines are called.
3211 */
3212static void
3213softdep_disk_write_complete(bp)
3214 struct buf *bp; /* describes the completed disk write */
3215{
3216 struct worklist *wk;
3217 struct workhead reattach;
3218 struct newblk *newblk;
3219 struct allocindir *aip;
3220 struct allocdirect *adp;
3221 struct indirdep *indirdep;
3222 struct inodedep *inodedep;
3223 struct bmsafemap *bmsafemap;
3224
3225#ifdef DEBUG
dadab5e9 3226 if (lk.lkt_held != NOHOLDER)
984263bc 3227 panic("softdep_disk_write_complete: lock is held");
dadab5e9 3228 lk.lkt_held = SPECIAL_FLAG;
984263bc
MD
3229#endif
3230 LIST_INIT(&reattach);
3231 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3232 WORKLIST_REMOVE(wk);
3233 switch (wk->wk_type) {
3234
3235 case D_PAGEDEP:
3236 if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3237 WORKLIST_INSERT(&reattach, wk);
3238 continue;
3239
3240 case D_INODEDEP:
3241 if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3242 WORKLIST_INSERT(&reattach, wk);
3243 continue;
3244
3245 case D_BMSAFEMAP:
3246 bmsafemap = WK_BMSAFEMAP(wk);
3247 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3248 newblk->nb_state |= DEPCOMPLETE;
3249 newblk->nb_bmsafemap = NULL;
3250 LIST_REMOVE(newblk, nb_deps);
3251 }
3252 while ((adp =
3253 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3254 adp->ad_state |= DEPCOMPLETE;
3255 adp->ad_buf = NULL;
3256 LIST_REMOVE(adp, ad_deps);
3257 handle_allocdirect_partdone(adp);
3258 }
3259 while ((aip =
3260 LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3261 aip->ai_state |= DEPCOMPLETE;
3262 aip->ai_buf = NULL;
3263 LIST_REMOVE(aip, ai_deps);
3264 handle_allocindir_partdone(aip);
3265 }
3266 while ((inodedep =
3267 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3268 inodedep->id_state |= DEPCOMPLETE;
3269 LIST_REMOVE(inodedep, id_deps);
3270 inodedep->id_buf = NULL;
3271 }
3272 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3273 continue;
3274
3275 case D_MKDIR:
3276 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3277 continue;
3278
3279 case D_ALLOCDIRECT:
3280 adp = WK_ALLOCDIRECT(wk);
3281 adp->ad_state |= COMPLETE;
3282 handle_allocdirect_partdone(adp);
3283 continue;
3284
3285 case D_ALLOCINDIR:
3286 aip = WK_ALLOCINDIR(wk);
3287 aip->ai_state |= COMPLETE;
3288 handle_allocindir_partdone(aip);
3289 continue;
3290
3291 case D_INDIRDEP:
3292 indirdep = WK_INDIRDEP(wk);
3293 if (indirdep->ir_state & GOINGAWAY) {
dadab5e9 3294 lk.lkt_held = NOHOLDER;
984263bc
MD
3295 panic("disk_write_complete: indirdep gone");
3296 }
3297 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3298 FREE(indirdep->ir_saveddata, M_INDIRDEP);
3299 indirdep->ir_saveddata = 0;
3300 indirdep->ir_state &= ~UNDONE;
3301 indirdep->ir_state |= ATTACHED;
3302 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3303 handle_allocindir_partdone(aip);
3304 if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
dadab5e9 3305 lk.lkt_held = NOHOLDER;
984263bc
MD
3306 panic("disk_write_complete: not gone");
3307 }
3308 }
3309 WORKLIST_INSERT(&reattach, wk);
3310 if ((bp->b_flags & B_DELWRI) == 0)
3311 stat_indir_blk_ptrs++;
3312 bdirty(bp);
3313 continue;
3314
3315 default:
dadab5e9 3316 lk.lkt_held = NOHOLDER;
984263bc
MD
3317 panic("handle_disk_write_complete: Unknown type %s",
3318 TYPENAME(wk->wk_type));
3319 /* NOTREACHED */
3320 }
3321 }
3322 /*
3323 * Reattach any requests that must be redone.
3324 */
3325 while ((wk = LIST_FIRST(&reattach)) != NULL) {
3326 WORKLIST_REMOVE(wk);
3327 WORKLIST_INSERT(&bp->b_dep, wk);
3328 }
3329#ifdef DEBUG
dadab5e9 3330 if (lk.lkt_held != SPECIAL_FLAG)
984263bc 3331 panic("softdep_disk_write_complete: lock lost");
dadab5e9 3332 lk.lkt_held = NOHOLDER;
984263bc
MD
3333#endif
3334}
3335
3336/*
3337 * Called from within softdep_disk_write_complete above. Note that
3338 * this routine is always called from interrupt level with further
3339 * splbio interrupts blocked.
3340 */
3341static void
3342handle_allocdirect_partdone(adp)
3343 struct allocdirect *adp; /* the completed allocdirect */
3344{
3345 struct allocdirect *listadp;
3346 struct inodedep *inodedep;
3347 long bsize;
3348
3349 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3350 return;
3351 if (adp->ad_buf != NULL) {
dadab5e9 3352 lk.lkt_held = NOHOLDER;
984263bc
MD
3353 panic("handle_allocdirect_partdone: dangling dep");
3354 }
3355 /*
3356 * The on-disk inode cannot claim to be any larger than the last
3357 * fragment that has been written. Otherwise, the on-disk inode
3358 * might have fragments that were not the last block in the file
3359 * which would corrupt the filesystem. Thus, we cannot free any
3360 * allocdirects after one whose ad_oldblkno claims a fragment as
3361 * these blocks must be rolled back to zero before writing the inode.
3362 * We check the currently active set of allocdirects in id_inoupdt.
3363 */
3364 inodedep = adp->ad_inodedep;
3365 bsize = inodedep->id_fs->fs_bsize;
3366 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) {
3367 /* found our block */
3368 if (listadp == adp)
3369 break;
3370 /* continue if ad_oldlbn is not a fragment */
3371 if (listadp->ad_oldsize == 0 ||
3372 listadp->ad_oldsize == bsize)
3373 continue;
3374 /* hit a fragment */
3375 return;
3376 }
3377 /*
3378 * If we have reached the end of the current list without
3379 * finding the just finished dependency, then it must be
3380 * on the future dependency list. Future dependencies cannot
3381 * be freed until they are moved to the current list.
3382 */
3383 if (listadp == NULL) {
3384#ifdef DEBUG
3385 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next)
3386 /* found our block */
3387 if (listadp == adp)
3388 break;
3389 if (listadp == NULL) {
dadab5e9 3390 lk.lkt_held = NOHOLDER;
984263bc
MD
3391 panic("handle_allocdirect_partdone: lost dep");
3392 }
3393#endif /* DEBUG */
3394 return;
3395 }
3396 /*
3397 * If we have found the just finished dependency, then free
3398 * it along with anything that follows it that is complete.
3399 */
3400 for (; adp; adp = listadp) {
3401 listadp = TAILQ_NEXT(adp, ad_next);
3402 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3403 return;
3404 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
3405 }
3406}
3407
3408/*
3409 * Called from within softdep_disk_write_complete above. Note that
3410 * this routine is always called from interrupt level with further
3411 * splbio interrupts blocked.
3412 */
3413static void
3414handle_allocindir_partdone(aip)
3415 struct allocindir *aip; /* the completed allocindir */
3416{
3417 struct indirdep *indirdep;
3418
3419 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
3420 return;
3421 if (aip->ai_buf != NULL) {
dadab5e9 3422 lk.lkt_held = NOHOLDER;
984263bc
MD
3423 panic("handle_allocindir_partdone: dangling dependency");
3424 }
3425 indirdep = aip->ai_indirdep;
3426 if (indirdep->ir_state & UNDONE) {
3427 LIST_REMOVE(aip, ai_next);
3428 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
3429 return;
3430 }
3431 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
3432 aip->ai_newblkno;
3433 LIST_REMOVE(aip, ai_next);
3434 if (aip->ai_freefrag != NULL)
3435 add_to_worklist(&aip->ai_freefrag->ff_list);
3436 WORKITEM_FREE(aip, D_ALLOCINDIR);
3437}
3438
3439/*
3440 * Called from within softdep_disk_write_complete above to restore
3441 * in-memory inode block contents to their most up-to-date state. Note
3442 * that this routine is always called from interrupt level with further
3443 * splbio interrupts blocked.
3444 */
3445static int
3446handle_written_inodeblock(inodedep, bp)
3447 struct inodedep *inodedep;
3448 struct buf *bp; /* buffer containing the inode block */
3449{
3450 struct worklist *wk, *filefree;
3451 struct allocdirect *adp, *nextadp;
3452 struct dinode *dp;
3453 int hadchanges;
3454
3455 if ((inodedep->id_state & IOSTARTED) == 0) {
dadab5e9 3456 lk.lkt_held = NOHOLDER;
984263bc
MD
3457 panic("handle_written_inodeblock: not started");
3458 }
3459 inodedep->id_state &= ~IOSTARTED;
3460 inodedep->id_state |= COMPLETE;
3461 dp = (struct dinode *)bp->b_data +
3462 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
3463 /*
3464 * If we had to rollback the inode allocation because of
3465 * bitmaps being incomplete, then simply restore it.
3466 * Keep the block dirty so that it will not be reclaimed until
3467 * all associated dependencies have been cleared and the
3468 * corresponding updates written to disk.
3469 */
3470 if (inodedep->id_savedino != NULL) {
3471 *dp = *inodedep->id_savedino;
3472 FREE(inodedep->id_savedino, M_INODEDEP);
3473 inodedep->id_savedino = NULL;
3474 if ((bp->b_flags & B_DELWRI) == 0)
3475 stat_inode_bitmap++;
3476 bdirty(bp);
3477 return (1);
3478 }
3479 /*
3480 * Roll forward anything that had to be rolled back before
3481 * the inode could be updated.
3482 */
3483 hadchanges = 0;
3484 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
3485 nextadp = TAILQ_NEXT(adp, ad_next);
3486 if (adp->ad_state & ATTACHED) {
dadab5e9 3487 lk.lkt_held = NOHOLDER;
984263bc
MD
3488 panic("handle_written_inodeblock: new entry");
3489 }
3490 if (adp->ad_lbn < NDADDR) {
3491 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) {
dadab5e9 3492 lk.lkt_held = NOHOLDER;
984263bc
MD
3493 panic("%s: %s #%ld mismatch %d != %d",
3494 "handle_written_inodeblock",
3495 "direct pointer", adp->ad_lbn,
3496 dp->di_db[adp->ad_lbn], adp->ad_oldblkno);
3497 }
3498 dp->di_db[adp->ad_lbn] = adp->ad_newblkno;
3499 } else {
3500 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) {
dadab5e9 3501 lk.lkt_held = NOHOLDER;
984263bc
MD
3502 panic("%s: %s #%ld allocated as %d",
3503 "handle_written_inodeblock",
3504 "indirect pointer", adp->ad_lbn - NDADDR,
3505 dp->di_ib[adp->ad_lbn - NDADDR]);
3506 }
3507 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno;
3508 }
3509 adp->ad_state &= ~UNDONE;
3510 adp->ad_state |= ATTACHED;
3511 hadchanges = 1;
3512 }
3513 if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
3514 stat_direct_blk_ptrs++;
3515 /*
3516 * Reset the file size to its most up-to-date value.
3517 */
3518 if (inodedep->id_savedsize == -1) {
dadab5e9 3519 lk.lkt_held = NOHOLDER;
984263bc
MD
3520 panic("handle_written_inodeblock: bad size");
3521 }
3522 if (dp->di_size != inodedep->id_savedsize) {
3523 dp->di_size = inodedep->id_savedsize;
3524 hadchanges = 1;
3525 }
3526 inodedep->id_savedsize = -1;
3527 /*
3528 * If there were any rollbacks in the inode block, then it must be
3529 * marked dirty so that its will eventually get written back in
3530 * its correct form.
3531 */
3532 if (hadchanges)
3533 bdirty(bp);
3534 /*
3535 * Process any allocdirects that completed during the update.
3536 */
3537 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
3538 handle_allocdirect_partdone(adp);
3539 /*
3540 * Process deallocations that were held pending until the
3541 * inode had been written to disk. Freeing of the inode
3542 * is delayed until after all blocks have been freed to
3543 * avoid creation of new <vfsid, inum, lbn> triples
3544 * before the old ones have been deleted.
3545 */
3546 filefree = NULL;
3547 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
3548 WORKLIST_REMOVE(wk);
3549 switch (wk->wk_type) {
3550
3551 case D_FREEFILE:
3552 /*
3553 * We defer adding filefree to the worklist until
3554 * all other additions have been made to ensure
3555 * that it will be done after all the old blocks
3556 * have been freed.
3557 */
3558 if (filefree != NULL) {
dadab5e9 3559 lk.lkt_held = NOHOLDER;
984263bc
MD
3560 panic("handle_written_inodeblock: filefree");
3561 }
3562 filefree = wk;
3563 continue;
3564
3565 case D_MKDIR:
3566 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
3567 continue;
3568
3569 case D_DIRADD:
3570 diradd_inode_written(WK_DIRADD(wk), inodedep);
3571 continue;
3572
3573 case D_FREEBLKS:
3574 case D_FREEFRAG:
3575 case D_DIRREM:
3576 add_to_worklist(wk);
3577 continue;
3578
3579 default:
dadab5e9 3580 lk.lkt_held = NOHOLDER;
984263bc
MD
3581 panic("handle_written_inodeblock: Unknown type %s",
3582 TYPENAME(wk->wk_type));
3583 /* NOTREACHED */
3584 }
3585 }
3586 if (filefree != NULL) {
3587 if (free_inodedep(inodedep) == 0) {
dadab5e9 3588 lk.lkt_held = NOHOLDER;
984263bc
MD
3589 panic("handle_written_inodeblock: live inodedep");
3590 }
3591 add_to_worklist(filefree);
3592 return (0);
3593 }
3594
3595 /*
3596 * If no outstanding dependencies, free it.
3597 */
3598 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0)
3599 return (0);
3600 return (hadchanges);
3601}
3602
3603/*
3604 * Process a diradd entry after its dependent inode has been written.
3605 * This routine must be called with splbio interrupts blocked.
3606 */
3607static void
3608diradd_inode_written(dap, inodedep)
3609 struct diradd *dap;
3610 struct inodedep *inodedep;
3611{
3612 struct pagedep *pagedep;
3613
3614 dap->da_state |= COMPLETE;
3615 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3616 if (dap->da_state & DIRCHG)
3617 pagedep = dap->da_previous->dm_pagedep;
3618 else
3619 pagedep = dap->da_pagedep;
3620 LIST_REMOVE(dap, da_pdlist);
3621 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3622 }
3623 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3624}
3625
3626/*
3627 * Handle the completion of a mkdir dependency.
3628 */
3629static void
3630handle_written_mkdir(mkdir, type)
3631 struct mkdir *mkdir;
3632 int type;
3633{
3634 struct diradd *dap;
3635 struct pagedep *pagedep;
3636
3637 if (mkdir->md_state != type) {
dadab5e9 3638 lk.lkt_held = NOHOLDER;
984263bc
MD
3639 panic("handle_written_mkdir: bad type");
3640 }
3641 dap = mkdir->md_diradd;
3642 dap->da_state &= ~type;
3643 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
3644 dap->da_state |= DEPCOMPLETE;
3645 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3646 if (dap->da_state & DIRCHG)
3647 pagedep = dap->da_previous->dm_pagedep;
3648 else
3649 pagedep = dap->da_pagedep;
3650 LIST_REMOVE(dap, da_pdlist);
3651 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3652 }
3653 LIST_REMOVE(mkdir, md_mkdirs);
3654 WORKITEM_FREE(mkdir, D_MKDIR);
3655}
3656
3657/*
3658 * Called from within softdep_disk_write_complete above.
3659 * A write operation was just completed. Removed inodes can
3660 * now be freed and associated block pointers may be committed.
3661 * Note that this routine is always called from interrupt level
3662 * with further splbio interrupts blocked.
3663 */
3664static int
3665handle_written_filepage(pagedep, bp)
3666 struct pagedep *pagedep;
3667 struct buf *bp; /* buffer containing the written page */
3668{
3669 struct dirrem *dirrem;
3670 struct diradd *dap, *nextdap;
3671 struct direct *ep;
3672 int i, chgs;
3673
3674 if ((pagedep->pd_state & IOSTARTED) == 0) {
dadab5e9 3675 lk.lkt_held = NOHOLDER;
984263bc
MD
3676 panic("handle_written_filepage: not started");
3677 }
3678 pagedep->pd_state &= ~IOSTARTED;
3679 /*
3680 * Process any directory removals that have been committed.
3681 */
3682 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
3683 LIST_REMOVE(dirrem, dm_next);
3684 dirrem->dm_dirinum = pagedep->pd_ino;
3685 add_to_worklist(&dirrem->dm_list);
3686 }
3687 /*
3688 * Free any directory additions that have been committed.
3689 */
3690 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
3691 free_diradd(dap);
3692 /*
3693 * Uncommitted directory entries must be restored.
3694 */
3695 for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
3696 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
3697 dap = nextdap) {
3698 nextdap = LIST_NEXT(dap, da_pdlist);
3699 if (dap->da_state & ATTACHED) {
dadab5e9 3700 lk.lkt_held = NOHOLDER;
984263bc
MD
3701 panic("handle_written_filepage: attached");
3702 }
3703 ep = (struct direct *)
3704 ((char *)bp->b_data + dap->da_offset);
3705 ep->d_ino = dap->da_newinum;
3706 dap->da_state &= ~UNDONE;
3707 dap->da_state |= ATTACHED;
3708 chgs = 1;
3709 /*
3710 * If the inode referenced by the directory has
3711 * been written out, then the dependency can be
3712 * moved to the pending list.
3713 */
3714 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3715 LIST_REMOVE(dap, da_pdlist);
3716 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
3717 da_pdlist);
3718 }
3719 }
3720 }
3721 /*
3722 * If there were any rollbacks in the directory, then it must be
3723 * marked dirty so that its will eventually get written back in
3724 * its correct form.
3725 */
3726 if (chgs) {
3727 if ((bp->b_flags & B_DELWRI) == 0)
3728 stat_dir_entry++;
3729 bdirty(bp);
3730 }
3731 /*
3732 * If no dependencies remain, the pagedep will be freed.
3733 * Otherwise it will remain to update the page before it
3734 * is written back to disk.
3735 */
3736 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) {
3737 for (i = 0; i < DAHASHSZ; i++)
3738 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
3739 break;
3740 if (i == DAHASHSZ) {
3741 LIST_REMOVE(pagedep, pd_hash);
3742 WORKITEM_FREE(pagedep, D_PAGEDEP);
3743 return (0);
3744 }
3745 }
3746 return (1);
3747}
3748
3749/*
3750 * Writing back in-core inode structures.
3751 *
3752 * The file system only accesses an inode's contents when it occupies an
3753 * "in-core" inode structure. These "in-core" structures are separate from
3754 * the page frames used to cache inode blocks. Only the latter are
3755 * transferred to/from the disk. So, when the updated contents of the
3756 * "in-core" inode structure are copied to the corresponding in-memory inode
3757 * block, the dependencies are also transferred. The following procedure is
3758 * called when copying a dirty "in-core" inode to a cached inode block.
3759 */
3760
3761/*
3762 * Called when an inode is loaded from disk. If the effective link count
3763 * differed from the actual link count when it was last flushed, then we
3764 * need to ensure that the correct effective link count is put back.
3765 */
3766void
3767softdep_load_inodeblock(ip)
3768 struct inode *ip; /* the "in_core" copy of the inode */
3769{
3770 struct inodedep *inodedep;
3771
3772 /*
3773 * Check for alternate nlink count.
3774 */
3775 ip->i_effnlink = ip->i_nlink;
3776 ACQUIRE_LOCK(&lk);
3777 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3778 FREE_LOCK(&lk);
3779 return;
3780 }
3781 ip->i_effnlink -= inodedep->id_nlinkdelta;
3782 FREE_LOCK(&lk);
3783}
3784
3785/*
3786 * This routine is called just before the "in-core" inode
3787 * information is to be copied to the in-memory inode block.
3788 * Recall that an inode block contains several inodes. If
3789 * the force flag is set, then the dependencies will be
3790 * cleared so that the update can always be made. Note that
3791 * the buffer is locked when this routine is called, so we
3792 * will never be in the middle of writing the inode block
3793 * to disk.
3794 */
3795void
3796softdep_update_inodeblock(ip, bp, waitfor)
3797 struct inode *ip; /* the "in_core" copy of the inode */
3798 struct buf *bp; /* the buffer containing the inode block */
3799 int waitfor; /* nonzero => update must be allowed */
3800{
3801 struct inodedep *inodedep;
3802 struct worklist *wk;
3803 int error, gotit;
3804
3805 /*
3806 * If the effective link count is not equal to the actual link
3807 * count, then we must track the difference in an inodedep while
3808 * the inode is (potentially) tossed out of the cache. Otherwise,
3809 * if there is no existing inodedep, then there are no dependencies
3810 * to track.
3811 */
3812 ACQUIRE_LOCK(&lk);
3813 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3814 FREE_LOCK(&lk);
3815 if (ip->i_effnlink != ip->i_nlink)
3816 panic("softdep_update_inodeblock: bad link count");
3817 return;
3818 }
3819 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
3820 FREE_LOCK(&lk);
3821 panic("softdep_update_inodeblock: bad delta");
3822 }
3823 /*
3824 * Changes have been initiated. Anything depending on these
3825 * changes cannot occur until this inode has been written.
3826 */
3827 inodedep->id_state &= ~COMPLETE;
3828 if ((inodedep->id_state & ONWORKLIST) == 0)
3829 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
3830 /*
3831 * Any new dependencies associated with the incore inode must
3832 * now be moved to the list associated with the buffer holding
3833 * the in-memory copy of the inode. Once merged process any
3834 * allocdirects that are completed by the merger.
3835 */
3836 merge_inode_lists(inodedep);
3837 if (TAILQ_FIRST(&inodedep->id_inoupdt) != NULL)
3838 handle_allocdirect_partdone(TAILQ_FIRST(&inodedep->id_inoupdt));
3839 /*
3840 * Now that the inode has been pushed into the buffer, the
3841 * operations dependent on the inode being written to disk
3842 * can be moved to the id_bufwait so that they will be
3843 * processed when the buffer I/O completes.
3844 */
3845 while ((wk = LIST_FIRST(&inodedep->id_inowait)) != NULL) {
3846 WORKLIST_REMOVE(wk);
3847 WORKLIST_INSERT(&inodedep->id_bufwait, wk);
3848 }
3849 /*
3850 * Newly allocated inodes cannot be written until the bitmap
3851 * that allocates them have been written (indicated by
3852 * DEPCOMPLETE being set in id_state). If we are doing a
3853 * forced sync (e.g., an fsync on a file), we force the bitmap
3854 * to be written so that the update can be done.
3855 */
3856 if ((inodedep->id_state & DEPCOMPLETE) != 0 || waitfor == 0) {
3857 FREE_LOCK(&lk);
3858 return;
3859 }
3860 gotit = getdirtybuf(&inodedep->id_buf, MNT_WAIT);
3861 FREE_LOCK(&lk);
3862 if (gotit &&
3863 (error = VOP_BWRITE(inodedep->id_buf->b_vp, inodedep->id_buf)) != 0)
3864 softdep_error("softdep_update_inodeblock: bwrite", error);
3865 if ((inodedep->id_state & DEPCOMPLETE) == 0)
3866 panic("softdep_update_inodeblock: update failed");
3867}
3868
3869/*
3870 * Merge the new inode dependency list (id_newinoupdt) into the old
3871 * inode dependency list (id_inoupdt). This routine must be called