vfs_object_create() was being called too early on devvp in the FFS mount
[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 $
7d618503 40 * $DragonFly: src/sys/vfs/ufs/ffs_softdep.c,v 1.19 2004/11/09 04:22:52 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 */
f719c866
DR
145static void softdep_error(char *, int);
146static void drain_output(struct vnode *, int);
147static int getdirtybuf(struct buf **, int);
148static void clear_remove(struct thread *);
149static void clear_inodedeps(struct thread *);
150static int flush_pagedep_deps(struct vnode *, struct mount *,
a6ee311a 151 struct diraddhd *);
f719c866
DR
152static int flush_inodedep_deps(struct fs *, ino_t);
153static int handle_written_filepage(struct pagedep *, struct buf *);
154static void diradd_inode_written(struct diradd *, struct inodedep *);
155static int handle_written_inodeblock(struct inodedep *, struct buf *);
156static void handle_allocdirect_partdone(struct allocdirect *);
157static void handle_allocindir_partdone(struct allocindir *);
158static void initiate_write_filepage(struct pagedep *, struct buf *);
159static void handle_written_mkdir(struct mkdir *, int);
160static void initiate_write_inodeblock(struct inodedep *, struct buf *);
161static void handle_workitem_freefile(struct freefile *);
162static void handle_workitem_remove(struct dirrem *);
163static struct dirrem *newdirrem(struct buf *, struct inode *,
a6ee311a 164 struct inode *, int, struct dirrem **);
f719c866
DR
165static void free_diradd(struct diradd *);
166static void free_allocindir(struct allocindir *, struct inodedep *);
a6ee311a
RG
167static int indir_trunc (struct inode *, ufs_daddr_t, int, ufs_lbn_t,
168 long *);
f719c866
DR
169static void deallocate_dependencies(struct buf *, struct inodedep *);
170static void free_allocdirect(struct allocdirectlst *,
a6ee311a 171 struct allocdirect *, int);
f719c866
DR
172static int check_inode_unwritten(struct inodedep *);
173static int free_inodedep(struct inodedep *);
174static void handle_workitem_freeblocks(struct freeblks *);
175static void merge_inode_lists(struct inodedep *);
176static void setup_allocindir_phase2(struct buf *, struct inode *,
a6ee311a 177 struct allocindir *);
f719c866 178static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t,
a6ee311a 179 ufs_daddr_t);
f719c866
DR
180static void handle_workitem_freefrag(struct freefrag *);
181static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long);
182static void allocdirect_merge(struct allocdirectlst *,
a6ee311a 183 struct allocdirect *, struct allocdirect *);
f719c866
DR
184static struct bmsafemap *bmsafemap_lookup(struct buf *);
185static int newblk_lookup(struct fs *, ufs_daddr_t, int,
a6ee311a 186 struct newblk **);
f719c866
DR
187static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
188static int pagedep_lookup(struct inode *, ufs_lbn_t, int,
a6ee311a 189 struct pagedep **);
f719c866
DR
190static void pause_timer(void *);
191static int request_cleanup(int, int);
192static int process_worklist_item(struct mount *, int);
193static void add_to_worklist(struct worklist *);
984263bc
MD
194
195/*
196 * Exported softdep operations.
197 */
f719c866
DR
198static void softdep_disk_io_initiation(struct buf *);
199static void softdep_disk_write_complete(struct buf *);
200static void softdep_deallocate_dependencies(struct buf *);
201static int softdep_fsync(struct vnode *);
202static int softdep_process_worklist(struct mount *);
203static void softdep_move_dependencies(struct buf *, struct buf *);
204static int softdep_count_dependencies(struct buf *bp, int);
984263bc 205
bc50d880 206static struct bio_ops softdep_bioops = {
984263bc
MD
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
f719c866
DR
247static void acquire_lock(struct lockit *);
248static void free_lock(struct lockit *);
249void softdep_panic(char *);
984263bc
MD
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
f719c866 298static int interlocked_sleep(struct lockit *, int, void *, int,
a6ee311a 299 const char *, int);
984263bc
MD
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};
f719c866
DR
366static void sema_init(struct sema *, char *, int, int);
367static int sema_get(struct sema *, struct lockit *);
368static void sema_release(struct sema *);
984263bc
MD
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 */
f719c866
DR
440static void worklist_insert(struct workhead *, struct worklist *);
441static void worklist_remove(struct worklist *);
442static void workitem_free(struct worklist *, int);
984263bc
MD
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 */
47979091 510static struct callout 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 }
5fd012e0 823 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY, td);
3b568787 824 error = VOP_FSYNC(devvp, MNT_WAIT, td);
5fd012e0 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
f719c866 1074 * mounting any filesystems.
984263bc
MD
1075 */
1076void
1077softdep_initialize()
1078{
47979091 1079 callout_init(&handle);
bc50d880 1080 bioops = softdep_bioops; /* XXX hack */
984263bc
MD
1081
1082 LIST_INIT(&mkdirlisthd);
1083 LIST_INIT(&softdep_workitem_pending);
1084 max_softdeps = min(desiredvnodes * 8,
1085 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep)));
1086 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1087 &pagedep_hash);
377d4740 1088 sema_init(&pagedep_in_progress, "pagedep", 0, 0);
984263bc 1089 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
377d4740 1090 sema_init(&inodedep_in_progress, "inodedep", 0, 0);
984263bc 1091 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
377d4740 1092 sema_init(&newblk_in_progress, "newblk", 0, 0);
984263bc
MD
1093}
1094
1095/*
1096 * Called at mount time to notify the dependency code that a
1097 * filesystem wishes to use it.
1098 */
1099int
3b568787 1100softdep_mount(devvp, mp, fs)
984263bc
MD
1101 struct vnode *devvp;
1102 struct mount *mp;
1103 struct fs *fs;
984263bc
MD
1104{
1105 struct csum cstotal;
1106 struct cg *cgp;
1107 struct buf *bp;
1108 int error, cyl;
1109
1110 mp->mnt_flag &= ~MNT_ASYNC;
1111 mp->mnt_flag |= MNT_SOFTDEP;
1112 /*
1113 * When doing soft updates, the counters in the
1114 * superblock may have gotten out of sync, so we have
1115 * to scan the cylinder groups and recalculate them.
1116 */
1117 if (fs->fs_clean != 0)
1118 return (0);
1119 bzero(&cstotal, sizeof cstotal);
1120 for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
1121 if ((error = bread(devvp, fsbtodb(fs, cgtod(fs, cyl)),
3b568787 1122 fs->fs_cgsize, &bp)) != 0) {
984263bc
MD
1123 brelse(bp);
1124 return (error);
1125 }
1126 cgp = (struct cg *)bp->b_data;
1127 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1128 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1129 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1130 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1131 fs->fs_cs(fs, cyl) = cgp->cg_cs;
1132 brelse(bp);
1133 }
1134#ifdef DEBUG
1135 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
1136 printf("ffs_mountfs: superblock updated for soft updates\n");
1137#endif
1138 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1139 return (0);
1140}
1141
1142/*
1143 * Protecting the freemaps (or bitmaps).
1144 *
f719c866 1145 * To eliminate the need to execute fsck before mounting a filesystem
984263bc
MD
1146 * after a power failure, one must (conservatively) guarantee that the
1147 * on-disk copy of the bitmaps never indicate that a live inode or block is
1148 * free. So, when a block or inode is allocated, the bitmap should be
1149 * updated (on disk) before any new pointers. When a block or inode is
1150 * freed, the bitmap should not be updated until all pointers have been
1151 * reset. The latter dependency is handled by the delayed de-allocation
1152 * approach described below for block and inode de-allocation. The former
1153 * dependency is handled by calling the following procedure when a block or
1154 * inode is allocated. When an inode is allocated an "inodedep" is created
1155 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1156 * Each "inodedep" is also inserted into the hash indexing structure so
1157 * that any additional link additions can be made dependent on the inode
1158 * allocation.
1159 *
f719c866 1160 * The ufs filesystem maintains a number of free block counts (e.g., per
984263bc
MD
1161 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1162 * in addition to the bitmaps. These counts are used to improve efficiency
1163 * during allocation and therefore must be consistent with the bitmaps.
1164 * There is no convenient way to guarantee post-crash consistency of these
1165 * counts with simple update ordering, for two main reasons: (1) The counts
1166 * and bitmaps for a single cylinder group block are not in the same disk
1167 * sector. If a disk write is interrupted (e.g., by power failure), one may
1168 * be written and the other not. (2) Some of the counts are located in the
1169 * superblock rather than the cylinder group block. So, we focus our soft
1170 * updates implementation on protecting the bitmaps. When mounting a
1171 * filesystem, we recompute the auxiliary counts from the bitmaps.
1172 */
1173
1174/*
1175 * Called just after updating the cylinder group block to allocate an inode.
1176 */
1177void
1178softdep_setup_inomapdep(bp, ip, newinum)
1179 struct buf *bp; /* buffer for cylgroup block with inode map */
1180 struct inode *ip; /* inode related to allocation */
1181 ino_t newinum; /* new inode number being allocated */
1182{
1183 struct inodedep *inodedep;
1184 struct bmsafemap *bmsafemap;
1185
1186 /*
1187 * Create a dependency for the newly allocated inode.
1188 * Panic if it already exists as something is seriously wrong.
1189 * Otherwise add it to the dependency list for the buffer holding
1190 * the cylinder group map from which it was allocated.
1191 */
1192 ACQUIRE_LOCK(&lk);
1193 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1194 FREE_LOCK(&lk);
1195 panic("softdep_setup_inomapdep: found inode");
1196 }
1197 inodedep->id_buf = bp;
1198 inodedep->id_state &= ~DEPCOMPLETE;
1199 bmsafemap = bmsafemap_lookup(bp);
1200 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1201 FREE_LOCK(&lk);
1202}
1203
1204/*
1205 * Called just after updating the cylinder group block to
1206 * allocate block or fragment.
1207 */
1208void
1209softdep_setup_blkmapdep(bp, fs, newblkno)
1210 struct buf *bp; /* buffer for cylgroup block with block map */
1211 struct fs *fs; /* filesystem doing allocation */
1212 ufs_daddr_t newblkno; /* number of newly allocated block */
1213{
1214 struct newblk *newblk;
1215 struct bmsafemap *bmsafemap;
1216
1217 /*
1218 * Create a dependency for the newly allocated block.
1219 * Add it to the dependency list for the buffer holding
1220 * the cylinder group map from which it was allocated.
1221 */
1222 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1223 panic("softdep_setup_blkmapdep: found block");
1224 ACQUIRE_LOCK(&lk);
1225 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1226 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1227 FREE_LOCK(&lk);
1228}
1229
1230/*
1231 * Find the bmsafemap associated with a cylinder group buffer.
1232 * If none exists, create one. The buffer must be locked when
1233 * this routine is called and this routine must be called with
1234 * splbio interrupts blocked.
1235 */
1236static struct bmsafemap *
1237bmsafemap_lookup(bp)
1238 struct buf *bp;
1239{
1240 struct bmsafemap *bmsafemap;
1241 struct worklist *wk;
1242
1243#ifdef DEBUG
dadab5e9 1244 if (lk.lkt_held == NOHOLDER)
984263bc
MD
1245 panic("bmsafemap_lookup: lock not held");
1246#endif
1247 LIST_FOREACH(wk, &bp->b_dep, wk_list)
1248 if (wk->wk_type == D_BMSAFEMAP)
1249 return (WK_BMSAFEMAP(wk));
1250 FREE_LOCK(&lk);
1251 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1252 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1253 bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1254 bmsafemap->sm_list.wk_state = 0;
1255 bmsafemap->sm_buf = bp;
1256 LIST_INIT(&bmsafemap->sm_allocdirecthd);
1257 LIST_INIT(&bmsafemap->sm_allocindirhd);
1258 LIST_INIT(&bmsafemap->sm_inodedephd);
1259 LIST_INIT(&bmsafemap->sm_newblkhd);
1260 ACQUIRE_LOCK(&lk);
1261 WORKLIST_INSERT(&bp->b_dep, &bmsafemap->sm_list);
1262 return (bmsafemap);
1263}
1264
1265/*
1266 * Direct block allocation dependencies.
1267 *
1268 * When a new block is allocated, the corresponding disk locations must be
1269 * initialized (with zeros or new data) before the on-disk inode points to
1270 * them. Also, the freemap from which the block was allocated must be
1271 * updated (on disk) before the inode's pointer. These two dependencies are
1272 * independent of each other and are needed for all file blocks and indirect
1273 * blocks that are pointed to directly by the inode. Just before the
1274 * "in-core" version of the inode is updated with a newly allocated block
1275 * number, a procedure (below) is called to setup allocation dependency
1276 * structures. These structures are removed when the corresponding
1277 * dependencies are satisfied or when the block allocation becomes obsolete
1278 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1279 * fragment that gets upgraded). All of these cases are handled in
1280 * procedures described later.
1281 *
1282 * When a file extension causes a fragment to be upgraded, either to a larger
1283 * fragment or to a full block, the on-disk location may change (if the
1284 * previous fragment could not simply be extended). In this case, the old
1285 * fragment must be de-allocated, but not until after the inode's pointer has
1286 * been updated. In most cases, this is handled by later procedures, which
1287 * will construct a "freefrag" structure to be added to the workitem queue
1288 * when the inode update is complete (or obsolete). The main exception to
1289 * this is when an allocation occurs while a pending allocation dependency
1290 * (for the same block pointer) remains. This case is handled in the main
1291 * allocation dependency setup procedure by immediately freeing the
1292 * unreferenced fragments.
1293 */
1294void
1295softdep_setup_allocdirect(ip, lbn, newblkno, oldblkno, newsize, oldsize, bp)
1296 struct inode *ip; /* inode to which block is being added */
1297 ufs_lbn_t lbn; /* block pointer within inode */
1298 ufs_daddr_t newblkno; /* disk block number being added */
1299 ufs_daddr_t oldblkno; /* previous block number, 0 unless frag */
1300 long newsize; /* size of new block */
1301 long oldsize; /* size of new block */
1302 struct buf *bp; /* bp for allocated block */
1303{
1304 struct allocdirect *adp, *oldadp;
1305 struct allocdirectlst *adphead;
1306 struct bmsafemap *bmsafemap;
1307 struct inodedep *inodedep;
1308 struct pagedep *pagedep;
1309 struct newblk *newblk;
1310
1311 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
1312 M_ALLOCDIRECT, M_SOFTDEP_FLAGS);
1313 bzero(adp, sizeof(struct allocdirect));
1314 adp->ad_list.wk_type = D_ALLOCDIRECT;
1315 adp->ad_lbn = lbn;
1316 adp->ad_newblkno = newblkno;
1317 adp->ad_oldblkno = oldblkno;
1318 adp->ad_newsize = newsize;
1319 adp->ad_oldsize = oldsize;
1320 adp->ad_state = ATTACHED;
1321 if (newblkno == oldblkno)
1322 adp->ad_freefrag = NULL;
1323 else
1324 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1325
1326 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1327 panic("softdep_setup_allocdirect: lost block");
1328
1329 ACQUIRE_LOCK(&lk);
1330 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1331 adp->ad_inodedep = inodedep;
1332
1333 if (newblk->nb_state == DEPCOMPLETE) {
1334 adp->ad_state |= DEPCOMPLETE;
1335 adp->ad_buf = NULL;
1336 } else {
1337 bmsafemap = newblk->nb_bmsafemap;
1338 adp->ad_buf = bmsafemap->sm_buf;
1339 LIST_REMOVE(newblk, nb_deps);
1340 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1341 }
1342 LIST_REMOVE(newblk, nb_hash);
1343 FREE(newblk, M_NEWBLK);
1344
1345 WORKLIST_INSERT(&bp->b_dep, &adp->ad_list);
1346 if (lbn >= NDADDR) {
1347 /* allocating an indirect block */
1348 if (oldblkno != 0) {
1349 FREE_LOCK(&lk);
1350 panic("softdep_setup_allocdirect: non-zero indir");
1351 }
1352 } else {
1353 /*
1354 * Allocating a direct block.
1355 *
1356 * If we are allocating a directory block, then we must
1357 * allocate an associated pagedep to track additions and
1358 * deletions.
1359 */
1360 if ((ip->i_mode & IFMT) == IFDIR &&
1361 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1362 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
1363 }
1364 /*
1365 * The list of allocdirects must be kept in sorted and ascending
1366 * order so that the rollback routines can quickly determine the
1367 * first uncommitted block (the size of the file stored on disk
1368 * ends at the end of the lowest committed fragment, or if there
1369 * are no fragments, at the end of the highest committed block).
1370 * Since files generally grow, the typical case is that the new
1371 * block is to be added at the end of the list. We speed this
1372 * special case by checking against the last allocdirect in the
1373 * list before laboriously traversing the list looking for the
1374 * insertion point.
1375 */
1376 adphead = &inodedep->id_newinoupdt;
1377 oldadp = TAILQ_LAST(adphead, allocdirectlst);
1378 if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1379 /* insert at end of list */
1380 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1381 if (oldadp != NULL && oldadp->ad_lbn == lbn)
1382 allocdirect_merge(adphead, adp, oldadp);
1383 FREE_LOCK(&lk);
1384 return;
1385 }
1386 TAILQ_FOREACH(oldadp, adphead, ad_next) {
1387 if (oldadp->ad_lbn >= lbn)
1388 break;
1389 }
1390 if (oldadp == NULL) {
1391 FREE_LOCK(&lk);
1392 panic("softdep_setup_allocdirect: lost entry");
1393 }
1394 /* insert in middle of list */
1395 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1396 if (oldadp->ad_lbn == lbn)
1397 allocdirect_merge(adphead, adp, oldadp);
1398 FREE_LOCK(&lk);
1399}
1400
1401/*
1402 * Replace an old allocdirect dependency with a newer one.
1403 * This routine must be called with splbio interrupts blocked.
1404 */
1405static void
1406allocdirect_merge(adphead, newadp, oldadp)
1407 struct allocdirectlst *adphead; /* head of list holding allocdirects */
1408 struct allocdirect *newadp; /* allocdirect being added */
1409 struct allocdirect *oldadp; /* existing allocdirect being checked */
1410{
1411 struct freefrag *freefrag;
1412
1413#ifdef DEBUG
dadab5e9 1414 if (lk.lkt_held == NOHOLDER)
984263bc
MD
1415 panic("allocdirect_merge: lock not held");
1416#endif
1417 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1418 newadp->ad_oldsize != oldadp->ad_newsize ||
1419 newadp->ad_lbn >= NDADDR) {
1420 FREE_LOCK(&lk);
1421 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1422 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1423 NDADDR);
1424 }
1425 newadp->ad_oldblkno = oldadp->ad_oldblkno;
1426 newadp->ad_oldsize = oldadp->ad_oldsize;
1427 /*
1428 * If the old dependency had a fragment to free or had never
1429 * previously had a block allocated, then the new dependency
1430 * can immediately post its freefrag and adopt the old freefrag.
1431 * This action is done by swapping the freefrag dependencies.
1432 * The new dependency gains the old one's freefrag, and the
1433 * old one gets the new one and then immediately puts it on
1434 * the worklist when it is freed by free_allocdirect. It is
1435 * not possible to do this swap when the old dependency had a
1436 * non-zero size but no previous fragment to free. This condition
1437 * arises when the new block is an extension of the old block.
1438 * Here, the first part of the fragment allocated to the new
1439 * dependency is part of the block currently claimed on disk by
1440 * the old dependency, so cannot legitimately be freed until the
1441 * conditions for the new dependency are fulfilled.
1442 */
1443 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1444 freefrag = newadp->ad_freefrag;
1445 newadp->ad_freefrag = oldadp->ad_freefrag;
1446 oldadp->ad_freefrag = freefrag;
1447 }
1448 free_allocdirect(adphead, oldadp, 0);
1449}
1450
1451/*
1452 * Allocate a new freefrag structure if needed.
1453 */
1454static struct freefrag *
1455newfreefrag(ip, blkno, size)
1456 struct inode *ip;
1457 ufs_daddr_t blkno;
1458 long size;
1459{
1460 struct freefrag *freefrag;
1461 struct fs *fs;
1462
1463 if (blkno == 0)
1464 return (NULL);
1465 fs = ip->i_fs;
1466 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1467 panic("newfreefrag: frag size");
1468 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1469 M_FREEFRAG, M_SOFTDEP_FLAGS);
1470 freefrag->ff_list.wk_type = D_FREEFRAG;
1471 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */
1472 freefrag->ff_inum = ip->i_number;
1473 freefrag->ff_fs = fs;
1474 freefrag->ff_devvp = ip->i_devvp;
1475 freefrag->ff_blkno = blkno;
1476 freefrag->ff_fragsize = size;
1477 return (freefrag);
1478}
1479
1480/*
1481 * This workitem de-allocates fragments that were replaced during
1482 * file block allocation.
1483 */
1484static void
1485handle_workitem_freefrag(freefrag)
1486 struct freefrag *freefrag;
1487{
1488 struct inode tip;
1489
1490 tip.i_fs = freefrag->ff_fs;
1491 tip.i_devvp = freefrag->ff_devvp;
1492 tip.i_dev = freefrag->ff_devvp->v_rdev;
1493 tip.i_number = freefrag->ff_inum;
1494 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */
1495 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1496 FREE(freefrag, M_FREEFRAG);
1497}
1498
1499/*
1500 * Indirect block allocation dependencies.
1501 *
1502 * The same dependencies that exist for a direct block also exist when
1503 * a new block is allocated and pointed to by an entry in a block of
1504 * indirect pointers. The undo/redo states described above are also
1505 * used here. Because an indirect block contains many pointers that
1506 * may have dependencies, a second copy of the entire in-memory indirect
1507 * block is kept. The buffer cache copy is always completely up-to-date.
1508 * The second copy, which is used only as a source for disk writes,
1509 * contains only the safe pointers (i.e., those that have no remaining
1510 * update dependencies). The second copy is freed when all pointers
1511 * are safe. The cache is not allowed to replace indirect blocks with
1512 * pending update dependencies. If a buffer containing an indirect
1513 * block with dependencies is written, these routines will mark it
1514 * dirty again. It can only be successfully written once all the
1515 * dependencies are removed. The ffs_fsync routine in conjunction with
1516 * softdep_sync_metadata work together to get all the dependencies
1517 * removed so that a file can be successfully written to disk. Three
1518 * procedures are used when setting up indirect block pointer
1519 * dependencies. The division is necessary because of the organization
1520 * of the "balloc" routine and because of the distinction between file
1521 * pages and file metadata blocks.
1522 */
1523
1524/*
1525 * Allocate a new allocindir structure.
1526 */
1527static struct allocindir *
1528newallocindir(ip, ptrno, newblkno, oldblkno)
1529 struct inode *ip; /* inode for file being extended */
1530 int ptrno; /* offset of pointer in indirect block */
1531 ufs_daddr_t newblkno; /* disk block number being added */
1532 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1533{
1534 struct allocindir *aip;
1535
1536 MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
1537 M_ALLOCINDIR, M_SOFTDEP_FLAGS);
1538 bzero(aip, sizeof(struct allocindir));
1539 aip->ai_list.wk_type = D_ALLOCINDIR;
1540 aip->ai_state = ATTACHED;
1541 aip->ai_offset = ptrno;
1542 aip->ai_newblkno = newblkno;
1543 aip->ai_oldblkno = oldblkno;
1544 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1545 return (aip);
1546}
1547
1548/*
1549 * Called just before setting an indirect block pointer
1550 * to a newly allocated file page.
1551 */
1552void
1553softdep_setup_allocindir_page(ip, lbn, bp, ptrno, newblkno, oldblkno, nbp)
1554 struct inode *ip; /* inode for file being extended */
1555 ufs_lbn_t lbn; /* allocated block number within file */
1556 struct buf *bp; /* buffer with indirect blk referencing page */
1557 int ptrno; /* offset of pointer in indirect block */
1558 ufs_daddr_t newblkno; /* disk block number being added */
1559 ufs_daddr_t oldblkno; /* previous block number, 0 if none */
1560 struct buf *nbp; /* buffer holding allocated page */
1561{
1562 struct allocindir *aip;
1563 struct pagedep *pagedep;
1564
1565 aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1566 ACQUIRE_LOCK(&lk);
1567 /*
1568 * If we are allocating a directory page, then we must
1569 * allocate an associated pagedep to track additions and
1570 * deletions.
1571 */
1572 if ((ip->i_mode & IFMT) == IFDIR &&
1573 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
1574 WORKLIST_INSERT(&nbp->b_dep, &pagedep->pd_list);
1575 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1576 FREE_LOCK(&lk);
1577 setup_allocindir_phase2(bp, ip, aip);
1578}
1579
1580/*
1581 * Called just before setting an indirect block pointer to a
1582 * newly allocated indirect block.
1583 */
1584void
1585softdep_setup_allocindir_meta(nbp, ip, bp, ptrno, newblkno)
1586 struct buf *nbp; /* newly allocated indirect block */
1587 struct inode *ip; /* inode for file being extended */
1588 struct buf *bp; /* indirect block referencing allocated block */
1589 int ptrno; /* offset of pointer in indirect block */
1590 ufs_daddr_t newblkno; /* disk block number being added */
1591{
1592 struct allocindir *aip;
1593
1594 aip = newallocindir(ip, ptrno, newblkno, 0);
1595 ACQUIRE_LOCK(&lk);
1596 WORKLIST_INSERT(&nbp->b_dep, &aip->ai_list);
1597 FREE_LOCK(&lk);
1598 setup_allocindir_phase2(bp, ip, aip);
1599}
1600
1601/*
1602 * Called to finish the allocation of the "aip" allocated
1603 * by one of the two routines above.
1604 */
1605static void
1606setup_allocindir_phase2(bp, ip, aip)
1607 struct buf *bp; /* in-memory copy of the indirect block */
1608 struct inode *ip; /* inode for file being extended */
1609 struct allocindir *aip; /* allocindir allocated by the above routines */
1610{
1611 struct worklist *wk;
1612 struct indirdep *indirdep, *newindirdep;
1613 struct bmsafemap *bmsafemap;
1614 struct allocindir *oldaip;
1615 struct freefrag *freefrag;
1616 struct newblk *newblk;
1617
1618 if (bp->b_lblkno >= 0)
1619 panic("setup_allocindir_phase2: not indir blk");
1620 for (indirdep = NULL, newindirdep = NULL; ; ) {
1621 ACQUIRE_LOCK(&lk);
1622 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1623 if (wk->wk_type != D_INDIRDEP)
1624 continue;
1625 indirdep = WK_INDIRDEP(wk);
1626 break;
1627 }
1628 if (indirdep == NULL && newindirdep) {
1629 indirdep = newindirdep;
1630 WORKLIST_INSERT(&bp->b_dep, &indirdep->ir_list);
1631 newindirdep = NULL;
1632 }
1633 FREE_LOCK(&lk);
1634 if (indirdep) {
1635 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1636 &newblk) == 0)
1637 panic("setup_allocindir: lost block");
1638 ACQUIRE_LOCK(&lk);
1639 if (newblk->nb_state == DEPCOMPLETE) {
1640 aip->ai_state |= DEPCOMPLETE;
1641 aip->ai_buf = NULL;
1642 } else {
1643 bmsafemap = newblk->nb_bmsafemap;
1644 aip->ai_buf = bmsafemap->sm_buf;
1645 LIST_REMOVE(newblk, nb_deps);
1646 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1647 aip, ai_deps);
1648 }
1649 LIST_REMOVE(newblk, nb_hash);
1650 FREE(newblk, M_NEWBLK);
1651 aip->ai_indirdep = indirdep;
1652 /*
1653 * Check to see if there is an existing dependency
1654 * for this block. If there is, merge the old
1655 * dependency into the new one.
1656 */
1657 if (aip->ai_oldblkno == 0)
1658 oldaip = NULL;
1659 else
1660
1661 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1662 if (oldaip->ai_offset == aip->ai_offset)
1663 break;
1664 if (oldaip != NULL) {
1665 if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1666 FREE_LOCK(&lk);
1667 panic("setup_allocindir_phase2: blkno");
1668 }
1669 aip->ai_oldblkno = oldaip->ai_oldblkno;
1670 freefrag = oldaip->ai_freefrag;
1671 oldaip->ai_freefrag = aip->ai_freefrag;
1672 aip->ai_freefrag = freefrag;
1673 free_allocindir(oldaip, NULL);
1674 }
1675 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1676 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1677 [aip->ai_offset] = aip->ai_oldblkno;
1678 FREE_LOCK(&lk);
1679 }
1680 if (newindirdep) {
7d618503
MD
1681 /*
1682 * We must invalidate ir_savebp to retire the original
1683 * getblk() that brought it in. VMIO backed buffers
1684 * will tend to do the right thing anyway, but
1685 * non-VMIO backed buffers must be properly retired
1686 * according to getblk()'s semantics because B_CACHE
1687 * is set in getblk() if the non-VMIO-backed bp
1688 * already exists.
1689 */
1690 if (indirdep->ir_savebp != NULL) {
1691 indirdep->ir_savebp->b_flags |=
1692 B_INVAL | B_NOCACHE;
984263bc 1693 brelse(newindirdep->ir_savebp);
7d618503 1694 }
984263bc
MD
1695 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1696 }
1697 if (indirdep)
1698 break;
1699 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1700 M_INDIRDEP, M_SOFTDEP_FLAGS);
1701 newindirdep->ir_list.wk_type = D_INDIRDEP;
1702 newindirdep->ir_state = ATTACHED;
1703 LIST_INIT(&newindirdep->ir_deplisthd);
1704 LIST_INIT(&newindirdep->ir_donehd);
1705 if (bp->b_blkno == bp->b_lblkno) {
1706 VOP_BMAP(bp->b_vp, bp->b_lblkno, NULL, &bp->b_blkno,
1707 NULL, NULL);
1708 }
1709 newindirdep->ir_savebp =
1710 getblk(ip->i_devvp, bp->b_blkno, bp->b_bcount, 0, 0);
1711 BUF_KERNPROC(newindirdep->ir_savebp);
1712 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1713 }
1714}
1715
1716/*
1717 * Block de-allocation dependencies.
1718 *
1719 * When blocks are de-allocated, the on-disk pointers must be nullified before
1720 * the blocks are made available for use by other files. (The true
1721 * requirement is that old pointers must be nullified before new on-disk
1722 * pointers are set. We chose this slightly more stringent requirement to
1723 * reduce complexity.) Our implementation handles this dependency by updating
1724 * the inode (or indirect block) appropriately but delaying the actual block
1725 * de-allocation (i.e., freemap and free space count manipulation) until
1726 * after the updated versions reach stable storage. After the disk is
1727 * updated, the blocks can be safely de-allocated whenever it is convenient.
1728 * This implementation handles only the common case of reducing a file's
1729 * length to zero. Other cases are handled by the conventional synchronous
1730 * write approach.
1731 *
1732 * The ffs implementation with which we worked double-checks
1733 * the state of the block pointers and file size as it reduces
1734 * a file's length. Some of this code is replicated here in our
1735 * soft updates implementation. The freeblks->fb_chkcnt field is
1736 * used to transfer a part of this information to the procedure
1737 * that eventually de-allocates the blocks.
1738 *
1739 * This routine should be called from the routine that shortens
1740 * a file's length, before the inode's size or block pointers
1741 * are modified. It will save the block pointer information for
1742 * later release and zero the inode so that the calling routine
1743 * can release it.
1744 */
1745void
1746softdep_setup_freeblocks(ip, length)
1747 struct inode *ip; /* The inode whose length is to be reduced */
1748 off_t length; /* The new length for the file */
1749{
1750 struct freeblks *freeblks;
1751 struct inodedep *inodedep;
1752 struct allocdirect *adp;
1753 struct vnode *vp;
1754 struct buf *bp;
1755 struct fs *fs;
1756 int i, error, delay;
1757
1758 fs = ip->i_fs;
1759 if (length != 0)
1760 panic("softde_setup_freeblocks: non-zero length");
1761 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
1762 M_FREEBLKS, M_SOFTDEP_FLAGS);
1763 bzero(freeblks, sizeof(struct freeblks));
1764 freeblks->fb_list.wk_type = D_FREEBLKS;
1765 freeblks->fb_uid = ip->i_uid;
1766 freeblks->fb_previousinum = ip->i_number;
1767 freeblks->fb_devvp = ip->i_devvp;
1768 freeblks->fb_fs = fs;
1769 freeblks->fb_oldsize = ip->i_size;
1770 freeblks->fb_newsize = length;
1771 freeblks->fb_chkcnt = ip->i_blocks;
1772 for (i = 0; i < NDADDR; i++) {
1773 freeblks->fb_dblks[i] = ip->i_db[i];
1774 ip->i_db[i] = 0;
1775 }
1776 for (i = 0; i < NIADDR; i++) {
1777 freeblks->fb_iblks[i] = ip->i_ib[i];
1778 ip->i_ib[i] = 0;
1779 }
1780 ip->i_blocks = 0;
1781 ip->i_size = 0;
1782 /*
1783 * Push the zero'ed inode to to its disk buffer so that we are free
1784 * to delete its dependencies below. Once the dependencies are gone
1785 * the buffer can be safely released.
1786 */
1787 if ((error = bread(ip->i_devvp,
1788 fsbtodb(fs, ino_to_fsba(fs, ip->i_number)),
3b568787 1789 (int)fs->fs_bsize, &bp)) != 0)
984263bc
MD
1790 softdep_error("softdep_setup_freeblocks", error);
1791 *((struct dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
1792 ip->i_din;
1793 /*
1794 * Find and eliminate any inode dependencies.
1795 */
1796 ACQUIRE_LOCK(&lk);
1797 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1798 if ((inodedep->id_state & IOSTARTED) != 0) {
1799 FREE_LOCK(&lk);
1800 panic("softdep_setup_freeblocks: inode busy");
1801 }
1802 /*
1803 * Add the freeblks structure to the list of operations that
1804 * must await the zero'ed inode being written to disk. If we
1805 * still have a bitmap dependency (delay == 0), then the inode
1806 * has never been written to disk, so we can process the
1807 * freeblks below once we have deleted the dependencies.
1808 */
1809 delay = (inodedep->id_state & DEPCOMPLETE);
1810 if (delay)
1811 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1812 /*
1813 * Because the file length has been truncated to zero, any
1814 * pending block allocation dependency structures associated
1815 * with this inode are obsolete and can simply be de-allocated.
1816 * We must first merge the two dependency lists to get rid of
1817 * any duplicate freefrag structures, then purge the merged list.
1818 */
1819 merge_inode_lists(inodedep);
1820 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1821 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
1822 FREE_LOCK(&lk);
1823 bdwrite(bp);
1824 /*
1825 * We must wait for any I/O in progress to finish so that
1826 * all potential buffers on the dirty list will be visible.
1827 * Once they are all there, walk the list and get rid of
1828 * any dependencies.
1829 */
1830 vp = ITOV(ip);
1831 ACQUIRE_LOCK(&lk);
1832 drain_output(vp, 1);
1833 while (getdirtybuf(&TAILQ_FIRST(&vp->v_dirtyblkhd), MNT_WAIT)) {
1834 bp = TAILQ_FIRST(&vp->v_dirtyblkhd);
1835 (void) inodedep_lookup(fs, ip->i_number, 0, &inodedep);
1836 deallocate_dependencies(bp, inodedep);
1837 bp->b_flags |= B_INVAL | B_NOCACHE;
1838 FREE_LOCK(&lk);
1839 brelse(bp);
1840 ACQUIRE_LOCK(&lk);
1841 }
1842 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
1843 (void)free_inodedep(inodedep);
1844 FREE_LOCK(&lk);
1845 /*
1846 * If the inode has never been written to disk (delay == 0),
1847 * then we can process the freeblks now that we have deleted
1848 * the dependencies.
1849 */
1850 if (!delay)
1851 handle_workitem_freeblocks(freeblks);
1852}
1853
1854/*
1855 * Reclaim any dependency structures from a buffer that is about to
1856 * be reallocated to a new vnode. The buffer must be locked, thus,
1857 * no I/O completion operations can occur while we are manipulating
1858 * its associated dependencies. The mutex is held so that other I/O's
1859 * associated with related dependencies do not occur.
1860 */
1861static void
1862deallocate_dependencies(bp, inodedep)
1863 struct buf *bp;
1864 struct inodedep *inodedep;
1865{
1866 struct worklist *wk;
1867 struct indirdep *indirdep;
1868 struct allocindir *aip;
1869 struct pagedep *pagedep;
1870 struct dirrem *dirrem;
1871 struct diradd *dap;
1872 int i;
1873
1874 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1875 switch (wk->wk_type) {
1876
1877 case D_INDIRDEP:
1878 indirdep = WK_INDIRDEP(wk);
1879 /*
1880 * None of the indirect pointers will ever be visible,
1881 * so they can simply be tossed. GOINGAWAY ensures
1882 * that allocated pointers will be saved in the buffer
1883 * cache until they are freed. Note that they will
1884 * only be able to be found by their physical address
1885 * since the inode mapping the logical address will
1886 * be gone. The save buffer used for the safe copy
1887 * was allocated in setup_allocindir_phase2 using
1888 * the physical address so it could be used for this
1889 * purpose. Hence we swap the safe copy with the real
1890 * copy, allowing the safe copy to be freed and holding
1891 * on to the real copy for later use in indir_trunc.
1892 */
1893 if (indirdep->ir_state & GOINGAWAY) {
1894 FREE_LOCK(&lk);
1895 panic("deallocate_dependencies: already gone");
1896 }
1897 indirdep->ir_state |= GOINGAWAY;
1898 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
1899 free_allocindir(aip, inodedep);
1900 if (bp->b_lblkno >= 0 ||
1901 bp->b_blkno != indirdep->ir_savebp->b_lblkno) {
1902 FREE_LOCK(&lk);
1903 panic("deallocate_dependencies: not indir");
1904 }
1905 bcopy(bp->b_data, indirdep->ir_savebp->b_data,
1906 bp->b_bcount);
1907 WORKLIST_REMOVE(wk);
1908 WORKLIST_INSERT(&indirdep->ir_savebp->b_dep, wk);
1909 continue;
1910
1911 case D_PAGEDEP:
1912 pagedep = WK_PAGEDEP(wk);
1913 /*
1914 * None of the directory additions will ever be
1915 * visible, so they can simply be tossed.
1916 */
1917 for (i = 0; i < DAHASHSZ; i++)
1918 while ((dap =
1919 LIST_FIRST(&pagedep->pd_diraddhd[i])))
1920 free_diradd(dap);
1921 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
1922 free_diradd(dap);
1923 /*
1924 * Copy any directory remove dependencies to the list
1925 * to be processed after the zero'ed inode is written.
1926 * If the inode has already been written, then they
1927 * can be dumped directly onto the work list.
1928 */
1929 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
1930 LIST_REMOVE(dirrem, dm_next);
1931 dirrem->dm_dirinum = pagedep->pd_ino;
1932 if (inodedep == NULL ||
1933 (inodedep->id_state & ALLCOMPLETE) ==
1934 ALLCOMPLETE)
1935 add_to_worklist(&dirrem->dm_list);
1936 else
1937 WORKLIST_INSERT(&inodedep->id_bufwait,
1938 &dirrem->dm_list);
1939 }
1940 WORKLIST_REMOVE(&pagedep->pd_list);
1941 LIST_REMOVE(pagedep, pd_hash);
1942 WORKITEM_FREE(pagedep, D_PAGEDEP);
1943 continue;
1944
1945 case D_ALLOCINDIR:
1946 free_allocindir(WK_ALLOCINDIR(wk), inodedep);
1947 continue;
1948
1949 case D_ALLOCDIRECT:
1950 case D_INODEDEP:
1951 FREE_LOCK(&lk);
1952 panic("deallocate_dependencies: Unexpected type %s",
1953 TYPENAME(wk->wk_type));
1954 /* NOTREACHED */
1955
1956 default:
1957 FREE_LOCK(&lk);
1958 panic("deallocate_dependencies: Unknown type %s",
1959 TYPENAME(wk->wk_type));
1960 /* NOTREACHED */
1961 }
1962 }
1963}
1964
1965/*
1966 * Free an allocdirect. Generate a new freefrag work request if appropriate.
1967 * This routine must be called with splbio interrupts blocked.
1968 */
1969static void
1970free_allocdirect(adphead, adp, delay)
1971 struct allocdirectlst *adphead;
1972 struct allocdirect *adp;
1973 int delay;
1974{
1975
1976#ifdef DEBUG
dadab5e9 1977 if (lk.lkt_held == NOHOLDER)
984263bc
MD
1978 panic("free_allocdirect: lock not held");
1979#endif
1980 if ((adp->ad_state & DEPCOMPLETE) == 0)
1981 LIST_REMOVE(adp, ad_deps);
1982 TAILQ_REMOVE(adphead, adp, ad_next);
1983 if ((adp->ad_state & COMPLETE) == 0)
1984 WORKLIST_REMOVE(&adp->ad_list);
1985 if (adp->ad_freefrag != NULL) {
1986 if (delay)
1987 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
1988 &adp->ad_freefrag->ff_list);
1989 else
1990 add_to_worklist(&adp->ad_freefrag->ff_list);
1991 }
1992 WORKITEM_FREE(adp, D_ALLOCDIRECT);
1993}
1994
1995/*
1996 * Prepare an inode to be freed. The actual free operation is not
1997 * done until the zero'ed inode has been written to disk.
1998 */
1999void
2000softdep_freefile(pvp, ino, mode)
2001 struct vnode *pvp;
2002 ino_t ino;
2003 int mode;
2004{
2005 struct inode *ip = VTOI(pvp);
2006 struct inodedep *inodedep;
2007 struct freefile *freefile;
2008
2009 /*
2010 * This sets up the inode de-allocation dependency.
2011 */
2012 MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2013 M_FREEFILE, M_SOFTDEP_FLAGS);
2014 freefile->fx_list.wk_type = D_FREEFILE;
2015 freefile->fx_list.wk_state = 0;
2016 freefile->fx_mode = mode;
2017 freefile->fx_oldinum = ino;
2018 freefile->fx_devvp = ip->i_devvp;
2019 freefile->fx_fs = ip->i_fs;
2020
2021 /*
2022 * If the inodedep does not exist, then the zero'ed inode has
2023 * been written to disk. If the allocated inode has never been
2024 * written to disk, then the on-disk inode is zero'ed. In either
2025 * case we can free the file immediately.
2026 */
2027 ACQUIRE_LOCK(&lk);
2028 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2029 check_inode_unwritten(inodedep)) {
2030 FREE_LOCK(&lk);
2031 handle_workitem_freefile(freefile);
2032 return;
2033 }
2034 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2035 FREE_LOCK(&lk);
2036}
2037
2038/*
2039 * Check to see if an inode has never been written to disk. If
2040 * so free the inodedep and return success, otherwise return failure.
2041 * This routine must be called with splbio interrupts blocked.
2042 *
2043 * If we still have a bitmap dependency, then the inode has never
2044 * been written to disk. Drop the dependency as it is no longer
2045 * necessary since the inode is being deallocated. We set the
2046 * ALLCOMPLETE flags since the bitmap now properly shows that the
2047 * inode is not allocated. Even if the inode is actively being
2048 * written, it has been rolled back to its zero'ed state, so we
2049 * are ensured that a zero inode is what is on the disk. For short
2050 * lived files, this change will usually result in removing all the
2051 * dependencies from the inode so that it can be freed immediately.
2052 */
2053static int
2054check_inode_unwritten(inodedep)
2055 struct inodedep *inodedep;
2056{
2057
2058 if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2059 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2060 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2061 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2062 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2063 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2064 inodedep->id_nlinkdelta != 0)
2065 return (0);
2066 inodedep->id_state |= ALLCOMPLETE;
2067 LIST_REMOVE(inodedep, id_deps);
2068 inodedep->id_buf = NULL;
2069 if (inodedep->id_state & ONWORKLIST)
2070 WORKLIST_REMOVE(&inodedep->id_list);
2071 if (inodedep->id_savedino != NULL) {
2072 FREE(inodedep->id_savedino, M_INODEDEP);
2073 inodedep->id_savedino = NULL;
2074 }
2075 if (free_inodedep(inodedep) == 0) {
2076 FREE_LOCK(&lk);
2077 panic("check_inode_unwritten: busy inode");
2078 }
2079 return (1);
2080}
2081
2082/*
2083 * Try to free an inodedep structure. Return 1 if it could be freed.
2084 */
2085static int
2086free_inodedep(inodedep)
2087 struct inodedep *inodedep;
2088{
2089
2090 if ((inodedep->id_state & ONWORKLIST) != 0 ||
2091 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2092 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2093 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2094 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2095 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2096 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2097 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
2098 return (0);
2099 LIST_REMOVE(inodedep, id_hash);
2100 WORKITEM_FREE(inodedep, D_INODEDEP);
2101 num_inodedep -= 1;
2102 return (1);
2103}
2104
2105/*
2106 * This workitem routine performs the block de-allocation.
2107 * The workitem is added to the pending list after the updated
2108 * inode block has been written to disk. As mentioned above,
2109 * checks regarding the number of blocks de-allocated (compared
2110 * to the number of blocks allocated for the file) are also
2111 * performed in this function.
2112 */
2113static void
2114handle_workitem_freeblocks(freeblks)
2115 struct freeblks *freeblks;
2116{
2117 struct inode tip;
2118 ufs_daddr_t bn;
2119 struct fs *fs;
2120 int i, level, bsize;
2121 long nblocks, blocksreleased = 0;
2122 int error, allerror = 0;
2123 ufs_lbn_t baselbns[NIADDR], tmpval;
2124
2125 tip.i_number = freeblks->fb_previousinum;
2126 tip.i_devvp = freeblks->fb_devvp;
2127 tip.i_dev = freeblks->fb_devvp->v_rdev;
2128 tip.i_fs = freeblks->fb_fs;
2129 tip.i_size = freeblks->fb_oldsize;
2130 tip.i_uid = freeblks->fb_uid;
2131 fs = freeblks->fb_fs;
2132 tmpval = 1;
2133 baselbns[0] = NDADDR;
2134 for (i = 1; i < NIADDR; i++) {
2135 tmpval *= NINDIR(fs);
2136 baselbns[i] = baselbns[i - 1] + tmpval;
2137 }
2138 nblocks = btodb(fs->fs_bsize);
2139 blocksreleased = 0;
2140 /*
2141 * Indirect blocks first.
2142 */
2143 for (level = (NIADDR - 1); level >= 0; level--) {
2144 if ((bn = freeblks->fb_iblks[level]) == 0)
2145 continue;
2146 if ((error = indir_trunc(&tip, fsbtodb(fs, bn), level,
2147 baselbns[level], &blocksreleased)) == 0)
2148 allerror = error;
2149 ffs_blkfree(&tip, bn, fs->fs_bsize);
2150 blocksreleased += nblocks;
2151 }
2152 /*
2153 * All direct blocks or frags.
2154 */
2155 for (i = (NDADDR - 1); i >= 0; i--) {
2156 if ((bn = freeblks->fb_dblks[i]) == 0)
2157 continue;
2158 bsize = blksize(fs, &tip, i);
2159 ffs_blkfree(&tip, bn, bsize);
2160 blocksreleased += btodb(bsize);
2161 }
2162
2163#ifdef DIAGNOSTIC
2164 if (freeblks->fb_chkcnt != blocksreleased)
2165 printf("handle_workitem_freeblocks: block count\n");
2166 if (allerror)
2167 softdep_error("handle_workitem_freeblks", allerror);
2168#endif /* DIAGNOSTIC */
2169 WORKITEM_FREE(freeblks, D_FREEBLKS);
2170}
2171
2172/*
2173 * Release blocks associated with the inode ip and stored in the indirect
2174 * block dbn. If level is greater than SINGLE, the block is an indirect block
2175 * and recursive calls to indirtrunc must be used to cleanse other indirect
2176 * blocks.
2177 */
2178static int
2179indir_trunc(ip, dbn, level, lbn, countp)
2180 struct inode *ip;
2181 ufs_daddr_t dbn;
2182 int level;
2183 ufs_lbn_t lbn;
2184 long *countp;
2185{
2186 struct buf *bp;
2187 ufs_daddr_t *bap;
2188 ufs_daddr_t nb;
2189 struct fs *fs;
2190 struct worklist *wk;
2191 struct indirdep *indirdep;
2192 int i, lbnadd, nblocks;
2193 int error, allerror = 0;
2194
2195 fs = ip->i_fs;
2196 lbnadd = 1;
2197 for (i = level; i > 0; i--)
2198 lbnadd *= NINDIR(fs);
2199 /*
2200 * Get buffer of block pointers to be freed. This routine is not
2201 * called until the zero'ed inode has been written, so it is safe
2202 * to free blocks as they are encountered. Because the inode has
2203 * been zero'ed, calls to bmap on these blocks will fail. So, we
2204 * have to use the on-disk address and the block device for the
2205 * filesystem to look them up. If the file was deleted before its
2206 * indirect blocks were all written to disk, the routine that set
2207 * us up (deallocate_dependencies) will have arranged to leave
2208 * a complete copy of the indirect block in memory for our use.
2209 * Otherwise we have to read the blocks in from the disk.
2210 */
2211 ACQUIRE_LOCK(&lk);
2212 if ((bp = incore(ip->i_devvp, dbn)) != NULL &&
2213 (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2214 if (wk->wk_type != D_INDIRDEP ||
2215 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2216 (indirdep->ir_state & GOINGAWAY) == 0) {
2217 FREE_LOCK(&lk);
2218 panic("indir_trunc: lost indirdep");
2219 }
2220 WORKLIST_REMOVE(wk);
2221 WORKITEM_FREE(indirdep, D_INDIRDEP);
2222 if (LIST_FIRST(&bp->b_dep) != NULL) {
2223 FREE_LOCK(&lk);
2224 panic("indir_trunc: dangling dep");
2225 }
2226 FREE_LOCK(&lk);
2227 } else {
2228 FREE_LOCK(&lk);
3b568787 2229 error = bread(ip->i_devvp, dbn, (int)fs->fs_bsize, &bp);
984263bc
MD
2230 if (error)
2231 return (error);
2232 }
2233 /*
2234 * Recursively free indirect blocks.
2235 */
2236 bap = (ufs_daddr_t *)bp->b_data;
2237 nblocks = btodb(fs->fs_bsize);
2238 for (i = NINDIR(fs) - 1; i >= 0; i--) {
2239 if ((nb = bap[i]) == 0)
2240 continue;
2241 if (level != 0) {
2242 if ((error = indir_trunc(ip, fsbtodb(fs, nb),
2243 level - 1, lbn + (i * lbnadd), countp)) != 0)
2244 allerror = error;
2245 }
2246 ffs_blkfree(ip, nb, fs->fs_bsize);
2247 *countp += nblocks;
2248 }
2249 bp->b_flags |= B_INVAL | B_NOCACHE;
2250 brelse(bp);
2251 return (allerror);
2252}
2253
2254/*
2255 * Free an allocindir.
2256 * This routine must be called with splbio interrupts blocked.
2257 */
2258static void
2259free_allocindir(aip, inodedep)
2260 struct allocindir *aip;
2261 struct inodedep *inodedep;
2262{
2263 struct freefrag *freefrag;
2264
2265#ifdef DEBUG
dadab5e9 2266 if (lk.lkt_held == NOHOLDER)
984263bc
MD
2267 panic("free_allocindir: lock not held");
2268#endif
2269 if ((aip->ai_state & DEPCOMPLETE) == 0)
2270 LIST_REMOVE(aip, ai_deps);
2271 if (aip->ai_state & ONWORKLIST)
2272 WORKLIST_REMOVE(&aip->ai_list);
2273 LIST_REMOVE(aip, ai_next);
2274 if ((freefrag = aip->ai_freefrag) != NULL) {
2275 if (inodedep == NULL)
2276 add_to_worklist(&freefrag->ff_list);
2277 else
2278 WORKLIST_INSERT(&inodedep->id_bufwait,
2279 &freefrag->ff_list);
2280 }
2281 WORKITEM_FREE(aip, D_ALLOCINDIR);
2282}
2283
2284/*
2285 * Directory entry addition dependencies.
2286 *
2287 * When adding a new directory entry, the inode (with its incremented link
2288 * count) must be written to disk before the directory entry's pointer to it.
2289 * Also, if the inode is newly allocated, the corresponding freemap must be
2290 * updated (on disk) before the directory entry's pointer. These requirements
2291 * are met via undo/redo on the directory entry's pointer, which consists
2292 * simply of the inode number.
2293 *
2294 * As directory entries are added and deleted, the free space within a
f719c866 2295 * directory block can become fragmented. The ufs filesystem will compact
984263bc
MD
2296 * a fragmented directory block to make space for a new entry. When this
2297 * occurs, the offsets of previously added entries change. Any "diradd"
2298 * dependency structures corresponding to these entries must be updated with
2299 * the new offsets.
2300 */
2301
2302/*
2303 * This routine is called after the in-memory inode's link
2304 * count has been incremented, but before the directory entry's
2305 * pointer to the inode has been set.
2306 */
2307void
2308softdep_setup_directory_add(bp, dp, diroffset, newinum, newdirbp)
2309 struct buf *bp; /* buffer containing directory block */
2310 struct inode *dp; /* inode for directory */
2311 off_t diroffset; /* offset of new entry in directory */
f719c866 2312 ino_t newinum; /* inode referenced by new directory entry */
984263bc
MD
2313 struct buf *newdirbp; /* non-NULL => contents of new mkdir */
2314{
2315 int offset; /* offset of new entry within directory block */
2316 ufs_lbn_t lbn; /* block in directory containing new entry */
2317 struct fs *fs;
2318 struct diradd *dap;
2319 struct pagedep *pagedep;
2320 struct inodedep *inodedep;
2321 struct mkdir *mkdir1, *mkdir2;
2322
2323 /*
2324 * Whiteouts have no dependencies.
2325 */
2326 if (newinum == WINO) {
2327 if (newdirbp != NULL)
2328 bdwrite(newdirbp);
2329 return;
2330 }
2331
2332 fs = dp->i_fs;
2333 lbn = lblkno(fs, diroffset);
2334 offset = blkoff(fs, diroffset);
2335 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
2336 M_SOFTDEP_FLAGS);
2337 bzero(dap, sizeof(struct diradd));
2338 dap->da_list.wk_type = D_DIRADD;
2339 dap->da_offset = offset;
2340 dap->da_newinum = newinum;
2341 dap->da_state = ATTACHED;
2342 if (newdirbp == NULL) {
2343 dap->da_state |= DEPCOMPLETE;
2344 ACQUIRE_LOCK(&lk);
2345 } else {
2346 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2347 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2348 M_SOFTDEP_FLAGS);
2349 mkdir1->md_list.wk_type = D_MKDIR;
2350 mkdir1->md_state = MKDIR_BODY;
2351 mkdir1->md_diradd = dap;
2352 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2353 M_SOFTDEP_FLAGS);
2354 mkdir2->md_list.wk_type = D_MKDIR;
2355 mkdir2->md_state = MKDIR_PARENT;
2356 mkdir2->md_diradd = dap;
2357 /*
2358 * Dependency on "." and ".." being written to disk.
2359 */
2360 mkdir1->md_buf = newdirbp;
2361 ACQUIRE_LOCK(&lk);
2362 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
2363 WORKLIST_INSERT(&newdirbp->b_dep, &mkdir1->md_list);
2364 FREE_LOCK(&lk);
2365 bdwrite(newdirbp);
2366 /*
2367 * Dependency on link count increase for parent directory
2368 */
2369 ACQUIRE_LOCK(&lk);
2370 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2371 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2372 dap->da_state &= ~MKDIR_PARENT;
2373 WORKITEM_FREE(mkdir2, D_MKDIR);
2374 } else {
2375 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2376 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2377 }
2378 }
2379 /*
2380 * Link into parent directory pagedep to await its being written.
2381 */
2382 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2383 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2384 dap->da_pagedep = pagedep;
2385 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2386 da_pdlist);
2387 /*
2388 * Link into its inodedep. Put it on the id_bufwait list if the inode
2389 * is not yet written. If it is written, do the post-inode write
2390 * processing to put it on the id_pendinghd list.
2391 */
2392 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2393 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2394 diradd_inode_written(dap, inodedep);
2395 else
2396 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2397 FREE_LOCK(&lk);
2398}
2399
2400/*
2401 * This procedure is called to change the offset of a directory
2402 * entry when compacting a directory block which must be owned
2403 * exclusively by the caller. Note that the actual entry movement
2404 * must be done in this procedure to ensure that no I/O completions
2405 * occur while the move is in progress.
2406 */
2407void
2408softdep_change_directoryentry_offset(dp, base, oldloc, newloc, entrysize)
2409 struct inode *dp; /* inode for directory */
2410 caddr_t base; /* address of dp->i_offset */
2411 caddr_t oldloc; /* address of old directory location */
2412 caddr_t newloc; /* address of new directory location */
2413 int entrysize; /* size of directory entry */
2414{
2415 int offset, oldoffset, newoffset;
2416 struct pagedep *pagedep;
2417 struct diradd *dap;
2418 ufs_lbn_t lbn;
2419
2420 ACQUIRE_LOCK(&lk);
2421 lbn = lblkno(dp->i_fs, dp->i_offset);
2422 offset = blkoff(dp->i_fs, dp->i_offset);
2423 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2424 goto done;
2425 oldoffset = offset + (oldloc - base);
2426 newoffset = offset + (newloc - base);
2427
2428 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2429 if (dap->da_offset != oldoffset)
2430 continue;
2431 dap->da_offset = newoffset;
2432 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2433 break;
2434 LIST_REMOVE(dap, da_pdlist);
2435 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2436 dap, da_pdlist);
2437 break;
2438 }
2439 if (dap == NULL) {
2440
2441 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2442 if (dap->da_offset == oldoffset) {
2443 dap->da_offset = newoffset;
2444 break;
2445 }
2446 }
2447 }
2448done:
2449 bcopy(oldloc, newloc, entrysize);
2450 FREE_LOCK(&lk);
2451}
2452
2453/*
2454 * Free a diradd dependency structure. This routine must be called
2455 * with splbio interrupts blocked.
2456 */
2457static void
2458free_diradd(dap)
2459 struct diradd *dap;
2460{
2461 struct dirrem *dirrem;
2462 struct pagedep *pagedep;
2463 struct inodedep *inodedep;
2464 struct mkdir *mkdir, *nextmd;
2465
2466#ifdef DEBUG
dadab5e9 2467 if (lk.lkt_held == NOHOLDER)
984263bc
MD
2468 panic("free_diradd: lock not held");
2469#endif
2470 WORKLIST_REMOVE(&dap->da_list);
2471 LIST_REMOVE(dap, da_pdlist);
2472 if ((dap->da_state & DIRCHG) == 0) {
2473 pagedep = dap->da_pagedep;
2474 } else {
2475 dirrem = dap->da_previous;
2476 pagedep = dirrem->dm_pagedep;
2477 dirrem->dm_dirinum = pagedep->pd_ino;
2478 add_to_worklist(&dirrem->dm_list);
2479 }
2480 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2481 0, &inodedep) != 0)
2482 (void) free_inodedep(inodedep);
2483 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2484 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2485 nextmd = LIST_NEXT(mkdir, md_mkdirs);
2486 if (mkdir->md_diradd != dap)
2487 continue;
2488 dap->da_state &= ~mkdir->md_state;
2489 WORKLIST_REMOVE(&mkdir->md_list);
2490 LIST_REMOVE(mkdir, md_mkdirs);
2491 WORKITEM_FREE(mkdir, D_MKDIR);
2492 }
2493 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2494 FREE_LOCK(&lk);
2495 panic("free_diradd: unfound ref");
2496 }
2497 }
2498 WORKITEM_FREE(dap, D_DIRADD);
2499}
2500
2501/*
2502 * Directory entry removal dependencies.
2503 *
2504 * When removing a directory entry, the entry's inode pointer must be
2505 * zero'ed on disk before the corresponding inode's link count is decremented
2506 * (possibly freeing the inode for re-use). This dependency is handled by
2507 * updating the directory entry but delaying the inode count reduction until
2508 * after the directory block has been written to disk. After this point, the
2509 * inode count can be decremented whenever it is convenient.
2510 */
2511
2512/*
2513 * This routine should be called immediately after removing
2514 * a directory entry. The inode's link count should not be
2515 * decremented by the calling procedure -- the soft updates
2516 * code will do this task when it is safe.
2517 */
2518void
2519softdep_setup_remove(bp, dp, ip, isrmdir)
2520 struct buf *bp; /* buffer containing directory block */
2521 struct inode *dp; /* inode for the directory being modified */
2522 struct inode *ip; /* inode for directory entry being removed */
2523 int isrmdir; /* indicates if doing RMDIR */
2524{
2525 struct dirrem *dirrem, *prevdirrem;
2526
2527 /*
2528 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2529 */
2530 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2531
2532 /*
2533 * If the COMPLETE flag is clear, then there were no active
2534 * entries and we want to roll back to a zeroed entry until
2535 * the new inode is committed to disk. If the COMPLETE flag is
2536 * set then we have deleted an entry that never made it to
2537 * disk. If the entry we deleted resulted from a name change,
2538 * then the old name still resides on disk. We cannot delete
2539 * its inode (returned to us in prevdirrem) until the zeroed
2540 * directory entry gets to disk. The new inode has never been
2541 * referenced on the disk, so can be deleted immediately.
2542 */
2543 if ((dirrem->dm_state & COMPLETE) == 0) {
2544 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2545 dm_next);
2546 FREE_LOCK(&lk);
2547 } else {
2548 if (prevdirrem != NULL)
2549 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2550 prevdirrem, dm_next);
2551 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2552 FREE_LOCK(&lk);
2553 handle_workitem_remove(dirrem);
2554 }
2555}
2556
2557/*
2558 * Allocate a new dirrem if appropriate and return it along with
2559 * its associated pagedep. Called without a lock, returns with lock.
2560 */
2561static long num_dirrem; /* number of dirrem allocated */
2562static struct dirrem *
2563newdirrem(bp, dp, ip, isrmdir, prevdirremp)
2564 struct buf *bp; /* buffer containing directory block */
2565 struct inode *dp; /* inode for the directory being modified */
2566 struct inode *ip; /* inode for directory entry being removed */
2567 int isrmdir; /* indicates if doing RMDIR */
2568 struct dirrem **prevdirremp; /* previously referenced inode, if any */
2569{
2570 int offset;
2571 ufs_lbn_t lbn;
2572 struct diradd *dap;
2573 struct dirrem *dirrem;
2574 struct pagedep *pagedep;
2575
2576 /*
2577 * Whiteouts have no deletion dependencies.
2578 */
2579 if (ip == NULL)
2580 panic("newdirrem: whiteout");
2581 /*
2582 * If we are over our limit, try to improve the situation.
2583 * Limiting the number of dirrem structures will also limit
2584 * the number of freefile and freeblks structures.
2585 */
2586 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2587 (void) request_cleanup(FLUSH_REMOVE, 0);
2588 num_dirrem += 1;
2589 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
2590 M_DIRREM, M_SOFTDEP_FLAGS);
2591 bzero(dirrem, sizeof(struct dirrem));
2592 dirrem->dm_list.wk_type = D_DIRREM;
2593 dirrem->dm_state = isrmdir ? RMDIR : 0;
2594 dirrem->dm_mnt = ITOV(ip)->v_mount;
2595 dirrem->dm_oldinum = ip->i_number;
2596 *prevdirremp = NULL;
2597
2598 ACQUIRE_LOCK(&lk);
2599 lbn = lblkno(dp->i_fs, dp->i_offset);
2600 offset = blkoff(dp->i_fs, dp->i_offset);
2601 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
2602 WORKLIST_INSERT(&bp->b_dep, &pagedep->pd_list);
2603 dirrem->dm_pagedep = pagedep;
2604 /*
2605 * Check for a diradd dependency for the same directory entry.
2606 * If present, then both dependencies become obsolete and can
2607 * be de-allocated. Check for an entry on both the pd_dirraddhd
2608 * list and the pd_pendinghd list.
2609 */
2610
2611 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
2612 if (dap->da_offset == offset)
2613 break;
2614 if (dap == NULL) {
2615
2616 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
2617 if (dap->da_offset == offset)
2618 break;
2619 if (dap == NULL)
2620 return (dirrem);
2621 }
2622 /*
2623 * Must be ATTACHED at this point.
2624 */
2625 if ((dap->da_state & ATTACHED) == 0) {
2626 FREE_LOCK(&lk);
2627 panic("newdirrem: not ATTACHED");
2628 }
2629 if (dap->da_newinum != ip->i_number) {
2630 FREE_LOCK(&lk);
2631 panic("newdirrem: inum %d should be %d",
2632 ip->i_number, dap->da_newinum);
2633 }
2634 /*
2635 * If we are deleting a changed name that never made it to disk,
2636 * then return the dirrem describing the previous inode (which
2637 * represents the inode currently referenced from this entry on disk).
2638 */
2639 if ((dap->da_state & DIRCHG) != 0) {
2640 *prevdirremp = dap->da_previous;
2641 dap->da_state &= ~DIRCHG;
2642 dap->da_pagedep = pagedep;
2643 }
2644 /*
2645 * We are deleting an entry that never made it to disk.
2646 * Mark it COMPLETE so we can delete its inode immediately.
2647 */
2648 dirrem->dm_state |= COMPLETE;
2649 free_diradd(dap);
2650 return (dirrem);
2651}
2652
2653/*
2654 * Directory entry change dependencies.
2655 *
2656 * Changing an existing directory entry requires that an add operation
2657 * be completed first followed by a deletion. The semantics for the addition
2658 * are identical to the description of adding a new entry above except
2659 * that the rollback is to the old inode number rather than zero. Once
2660 * the addition dependency is completed, the removal is done as described
2661 * in the removal routine above.
2662 */
2663
2664/*
2665 * This routine should be called immediately after changing
2666 * a directory entry. The inode's link count should not be
2667 * decremented by the calling procedure -- the soft updates
2668 * code will perform this task when it is safe.
2669 */
2670void
2671softdep_setup_directory_change(bp, dp, ip, newinum, isrmdir)
2672 struct buf *bp; /* buffer containing directory block */
2673 struct inode *dp; /* inode for the directory being modified */
2674 struct inode *ip; /* inode for directory entry being removed */
f719c866 2675 ino_t newinum; /* new inode number for changed entry */
984263bc
MD
2676 int isrmdir; /* indicates if doing RMDIR */
2677{
2678 int offset;
2679 struct diradd *dap = NULL;
2680 struct dirrem *dirrem, *prevdirrem;
2681 struct pagedep *pagedep;
2682 struct inodedep *inodedep;
2683
2684 offset = blkoff(dp->i_fs, dp->i_offset);
2685
2686 /*
2687 * Whiteouts do not need diradd dependencies.
2688 */
2689 if (newinum != WINO) {
2690 MALLOC(dap, struct diradd *, sizeof(struct diradd),
2691 M_DIRADD, M_SOFTDEP_FLAGS);
2692 bzero(dap, sizeof(struct diradd));
2693 dap->da_list.wk_type = D_DIRADD;
2694 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2695 dap->da_offset = offset;
2696 dap->da_newinum = newinum;
2697 }
2698
2699 /*
2700 * Allocate a new dirrem and ACQUIRE_LOCK.
2701 */
2702 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2703 pagedep = dirrem->dm_pagedep;
2704 /*
2705 * The possible values for isrmdir:
2706 * 0 - non-directory file rename
2707 * 1 - directory rename within same directory
2708 * inum - directory rename to new directory of given inode number
2709 * When renaming to a new directory, we are both deleting and
2710 * creating a new directory entry, so the link count on the new
2711 * directory should not change. Thus we do not need the followup
2712 * dirrem which is usually done in handle_workitem_remove. We set
2713 * the DIRCHG flag to tell handle_workitem_remove to skip the
2714 * followup dirrem.
2715 */
2716 if (isrmdir > 1)
2717 dirrem->dm_state |= DIRCHG;
2718
2719 /*
2720 * Whiteouts have no additional dependencies,
2721 * so just put the dirrem on the correct list.
2722 */
2723 if (newinum == WINO) {
2724 if ((dirrem->dm_state & COMPLETE) == 0) {
2725 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2726 dm_next);
2727 } else {
2728 dirrem->dm_dirinum = pagedep->pd_ino;
2729 add_to_worklist(&dirrem->dm_list);
2730 }
2731 FREE_LOCK(&lk);
2732 return;
2733 }
2734
2735 /*
2736 * If the COMPLETE flag is clear, then there were no active
2737 * entries and we want to roll back to the previous inode until
2738 * the new inode is committed to disk. If the COMPLETE flag is
2739 * set, then we have deleted an entry that never made it to disk.
2740 * If the entry we deleted resulted from a name change, then the old
2741 * inode reference still resides on disk. Any rollback that we do
2742 * needs to be to that old inode (returned to us in prevdirrem). If
2743 * the entry we deleted resulted from a create, then there is
2744 * no entry on the disk, so we want to roll back to zero rather
2745 * than the uncommitted inode. In either of the COMPLETE cases we
2746 * want to immediately free the unwritten and unreferenced inode.
2747 */
2748 if ((dirrem->dm_state & COMPLETE) == 0) {
2749 dap->da_previous = dirrem;
2750 } else {
2751 if (prevdirrem != NULL) {
2752 dap->da_previous = prevdirrem;
2753 } else {
2754 dap->da_state &= ~DIRCHG;
2755 dap->da_pagedep = pagedep;
2756 }
2757 dirrem->dm_dirinum = pagedep->pd_ino;
2758 add_to_worklist(&dirrem->dm_list);
2759 }
2760 /*
2761 * Link into its inodedep. Put it on the id_bufwait list if the inode
2762 * is not yet written. If it is written, do the post-inode write
2763 * processing to put it on the id_pendinghd list.
2764 */
2765 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2766 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2767 dap->da_state |= COMPLETE;
2768 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2769 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2770 } else {
2771 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2772 dap, da_pdlist);
2773 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2774 }
2775 FREE_LOCK(&lk);
2776}
2777
2778/*
2779 * Called whenever the link count on an inode is changed.
2780 * It creates an inode dependency so that the new reference(s)
2781 * to the inode cannot be committed to disk until the updated
2782 * inode has been written.
2783 */
2784void
2785softdep_change_linkcnt(ip)
2786 struct inode *ip; /* the inode with the increased link count */
2787{
2788 struct inodedep *inodedep;
2789
2790 ACQUIRE_LOCK(&lk);
2791 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2792 if (ip->i_nlink < ip->i_effnlink) {
2793 FREE_LOCK(&lk);
2794 panic("softdep_change_linkcnt: bad delta");
2795 }
2796 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2797 FREE_LOCK(&lk);
2798}
2799
2800/*
2801 * This workitem decrements the inode's link count.
2802 * If the link count reaches zero, the file is removed.
2803 */
2804static void
2805handle_workitem_remove(dirrem)
2806 struct dirrem *dirrem;
2807{
dadab5e9 2808 struct thread *td = curthread; /* XXX */
984263bc
MD
2809 struct inodedep *inodedep;
2810 struct vnode *vp;
2811 struct inode *ip;
2812 ino_t oldinum;
2813 int error;
2814
2815 if ((error = VFS_VGET(dirrem->dm_mnt, dirrem->dm_oldinum, &vp)) != 0) {
2816 softdep_error("handle_workitem_remove: vget", error);
2817 return;
2818 }
2819 ip = VTOI(vp);
2820 ACQUIRE_LOCK(&lk);
2821 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
2822 FREE_LOCK(&lk);
2823 panic("handle_workitem_remove: lost inodedep");
2824 }
2825 /*
2826 * Normal file deletion.
2827 */
2828 if ((dirrem->dm_state & RMDIR) == 0) {
2829 ip->i_nlink--;
2830 ip->i_flag |= IN_CHANGE;
2831 if (ip->i_nlink < ip->i_effnlink) {
2832 FREE_LOCK(&lk);
2833 panic("handle_workitem_remove: bad file delta");
2834 }
2835 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2836 FREE_LOCK(&lk);
2837 vput(vp);
2838 num_dirrem -= 1;
2839 WORKITEM_FREE(dirrem, D_DIRREM);
2840 return;
2841 }
2842 /*
2843 * Directory deletion. Decrement reference count for both the
2844 * just deleted parent directory entry and the reference for ".".
2845 * Next truncate the directory to length zero. When the
2846 * truncation completes, arrange to have the reference count on
2847 * the parent decremented to account for the loss of "..".
2848 */
2849 ip->i_nlink -= 2;
2850 ip->i_flag |= IN_CHANGE;
2851 if (ip->i_nlink < ip->i_effnlink) {
2852 FREE_LOCK(&lk);
2853 panic("handle_workitem_remove: bad dir delta");
2854 }
2855 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2856 FREE_LOCK(&lk);
0cfcada1 2857 if ((error = UFS_TRUNCATE(vp, (off_t)0, 0, proc0.p_ucred, td)) != 0)
984263bc
MD
2858 softdep_error("handle_workitem_remove: truncate", error);
2859 /*
2860 * Rename a directory to a new parent. Since, we are both deleting
2861 * and creating a new directory entry, the link count on the new
2862 * directory should not change. Thus we skip the followup dirrem.
2863 */
2864 if (dirrem->dm_state & DIRCHG) {
2865 vput(vp);
2866 num_dirrem -= 1;
2867 WORKITEM_FREE(dirrem, D_DIRREM);
2868 return;
2869 }
2870 /*
2871 * If the inodedep does not exist, then the zero'ed inode has
2872 * been written to disk. If the allocated inode has never been
2873 * written to disk, then the on-disk inode is zero'ed. In either
2874 * case we can remove the file immediately.
2875 */
2876 ACQUIRE_LOCK(&lk);
2877 dirrem->dm_state = 0;
2878 oldinum = dirrem->dm_oldinum;
2879 dirrem->dm_oldinum = dirrem->dm_dirinum;
2880 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
2881 check_inode_unwritten(inodedep)) {
2882 FREE_LOCK(&lk);
2883 vput(vp);
2884 handle_workitem_remove(dirrem);
2885 return;
2886 }
2887 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
2888 FREE_LOCK(&lk);
2889 vput(vp);
2890}
2891
2892/*
2893 * Inode de-allocation dependencies.
2894 *
2895 * When an inode's link count is reduced to zero, it can be de-allocated. We
2896 * found it convenient to postpone de-allocation until after the inode is
2897 * written to disk with its new link count (zero). At this point, all of the
2898 * on-disk inode's block pointers are nullified and, with careful dependency
2899 * list ordering, all dependencies related to the inode will be satisfied and
2900 * the corresponding dependency structures de-allocated. So, if/when the
2901 * inode is reused, there will be no mixing of old dependencies with new
2902 * ones. This artificial dependency is set up by the block de-allocation
2903 * procedure above (softdep_setup_freeblocks) and completed by the
2904 * following procedure.
2905 */
2906static void
2907handle_workitem_freefile(freefile)
2908 struct freefile *freefile;
2909{
2910 struct vnode vp;
2911 struct inode tip;
2912 struct inodedep *idp;
2913 int error;
2914
2915#ifdef DEBUG
2916 ACQUIRE_LOCK(&lk);
2917 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp);
2918 FREE_LOCK(&lk);
2919 if (error)
2920 panic("handle_workitem_freefile: inodedep survived");
2921#endif
2922 tip.i_devvp = freefile->fx_devvp;
2923 tip.i_dev = freefile->fx_devvp->v_rdev;
2924 tip.i_fs = freefile->fx_fs;
2925 vp.v_data = &tip;
2926 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
2927 softdep_error("handle_workitem_freefile", error);
2928 WORKITEM_FREE(freefile, D_FREEFILE);
2929}
2930
2931/*
2932 * Disk writes.
2933 *
2934 * The dependency structures constructed above are most actively used when file
2935 * system blocks are written to disk. No constraints are placed on when a
2936 * block can be written, but unsatisfied update dependencies are made safe by
2937 * modifying (or replacing) the source memory for the duration of the disk
2938 * write. When the disk write completes, the memory block is again brought
2939 * up-to-date.
2940 *
2941 * In-core inode structure reclamation.
2942 *
2943 * Because there are a finite number of "in-core" inode structures, they are
2944 * reused regularly. By transferring all inode-related dependencies to the
2945 * in-memory inode block and indexing them separately (via "inodedep"s), we
2946 * can allow "in-core" inode structures to be reused at any time and avoid
2947 * any increase in contention.
2948 *
2949 * Called just before entering the device driver to initiate a new disk I/O.
2950 * The buffer must be locked, thus, no I/O completion operations can occur
2951 * while we are manipulating its associated dependencies.
2952 */
2953static void
2954softdep_disk_io_initiation(bp)
2955 struct buf *bp; /* structure describing disk write to occur */
2956{
2957 struct worklist *wk, *nextwk;
2958 struct indirdep *indirdep;
2959
2960 /*
2961 * We only care about write operations. There should never
2962 * be dependencies for reads.
2963 */
2964 if (bp->b_flags & B_READ)
2965 panic("softdep_disk_io_initiation: read");
2966 /*
2967 * Do any necessary pre-I/O processing.
2968 */
2969 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = nextwk) {
2970 nextwk = LIST_NEXT(wk, wk_list);
2971 switch (wk->wk_type) {
2972
2973 case D_PAGEDEP:
2974 initiate_write_filepage(WK_PAGEDEP(wk), bp);
2975 continue;
2976
2977 case D_INODEDEP:
2978 initiate_write_inodeblock(WK_INODEDEP(wk), bp);
2979 continue;
2980
2981 case D_INDIRDEP:
2982 indirdep = WK_INDIRDEP(wk);
2983 if (indirdep->ir_state & GOINGAWAY)
2984 panic("disk_io_initiation: indirdep gone");
2985 /*
2986 * If there are no remaining dependencies, this
2987 * will be writing the real pointers, so the
2988 * dependency can be freed.
2989 */
2990 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
2991 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
2992 brelse(indirdep->ir_savebp);
2993 /* inline expand WORKLIST_REMOVE(wk); */
2994 wk->wk_state &= ~ONWORKLIST;
2995 LIST_REMOVE(wk, wk_list);
2996 WORKITEM_FREE(indirdep, D_INDIRDEP);
2997 continue;
2998 }
2999 /*
3000 * Replace up-to-date version with safe version.
3001 */
3002 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3003 M_INDIRDEP, M_SOFTDEP_FLAGS);
3004 ACQUIRE_LOCK(&lk);
3005 indirdep->ir_state &= ~ATTACHED;
3006 indirdep->ir_state |= UNDONE;
3007 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3008 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3009 bp->b_bcount);
3010 FREE_LOCK(&lk);
3011 continue;
3012
3013 case D_MKDIR:
3014 case D_BMSAFEMAP:
3015 case D_ALLOCDIRECT:
3016 case D_ALLOCINDIR:
3017 continue;
3018
3019 default:
3020 panic("handle_disk_io_initiation: Unexpected type %s",
3021 TYPENAME(wk->wk_type));
3022 /* NOTREACHED */
3023 }
3024 }
3025}
3026
3027/*
3028 * Called from within the procedure above to deal with unsatisfied
3029 * allocation dependencies in a directory. The buffer must be locked,
3030 * thus, no I/O completion operations can occur while we are
3031 * manipulating its associated dependencies.
3032 */
3033static void
3034initiate_write_filepage(pagedep, bp)
3035 struct pagedep *pagedep;
3036 struct buf *bp;
3037{
3038 struct diradd *dap;
3039 struct direct *ep;
3040 int i;
3041
3042 if (pagedep->pd_state & IOSTARTED) {
3043 /*
3044 * This can only happen if there is a driver that does not
3045 * understand chaining. Here biodone will reissue the call
3046 * to strategy for the incomplete buffers.
3047 */
3048 printf("initiate_write_filepage: already started\n");
3049 return;
3050 }
3051 pagedep->pd_state |= IOSTARTED;
3052 ACQUIRE_LOCK(&lk);
3053 for (i = 0; i < DAHASHSZ; i++) {
3054 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3055 ep = (struct direct *)
3056 ((char *)bp->b_data + dap->da_offset);
3057 if (ep->d_ino != dap->da_newinum) {
3058 FREE_LOCK(&lk);
3059 panic("%s: dir inum %d != new %d",
3060 "initiate_write_filepage",
3061 ep->d_ino, dap->da_newinum);
3062 }
3063 if (dap->da_state & DIRCHG)
3064 ep->d_ino = dap->da_previous->dm_oldinum;
3065 else
3066 ep->d_ino = 0;
3067 dap->da_state &= ~ATTACHED;
3068 dap->da_state |= UNDONE;
3069 }
3070 }
3071 FREE_LOCK(&lk);
3072}
3073
3074/*
3075 * Called from within the procedure above to deal with unsatisfied
3076 * allocation dependencies in an inodeblock. The buffer must be
3077 * locked, thus, no I/O completion operations can occur while we
3078 * are manipulating its associated dependencies.
3079 */
3080static void
3081initiate_write_inodeblock(inodedep, bp)
3082 struct inodedep *inodedep;
3083 struct buf *bp; /* The inode block */
3084{
3085 struct allocdirect *adp, *lastadp;
3086 struct dinode *dp;
3087 struct fs *fs;
3088 ufs_lbn_t prevlbn = 0;
3089 int i, deplist;
3090
3091 if (inodedep->id_state & IOSTARTED)
3092 panic("initiate_write_inodeblock: already started");
3093 inodedep->id_state |= IOSTARTED;
3094 fs = inodedep->id_fs;
3095 dp = (struct dinode *)bp->b_data +
3096 ino_to_fsbo(fs, inodedep->id_ino);
3097 /*
3098 * If the bitmap is not yet written, then the allocated
3099 * inode cannot be written to disk.
3100 */
3101 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3102 if (inodedep->id_savedino != NULL)
3103 panic("initiate_write_inodeblock: already doing I/O");
3104 MALLOC(inodedep->id_savedino, struct dinode *,
3105 sizeof(struct dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
3106 *inodedep->id_savedino = *dp;
3107 bzero((caddr_t)dp, sizeof(struct dinode));
3108 return;
3109 }
3110 /*
3111 * If no dependencies, then there is nothing to roll back.
3112 */
3113 inodedep->id_savedsize = dp->di_size;
3114 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3115 return;
3116 /*
3117 * Set the dependencies to busy.
3118 */
3119 ACQUIRE_LOCK(&lk);
3120 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3121 adp = TAILQ_NEXT(adp, ad_next)) {
3122#ifdef DIAGNOSTIC
3123 if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3124 FREE_LOCK(&lk);
3125 panic("softdep_write_inodeblock: lbn order");
3126 }
3127 prevlbn = adp->ad_lbn;
3128 if (adp->ad_lbn < NDADDR &&
3129 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3130 FREE_LOCK(&lk);
3131 panic("%s: direct pointer #%ld mismatch %d != %d",
3132 "softdep_write_inodeblock", adp->ad_lbn,
3133 dp->di_db[adp->ad_lbn], adp->ad_newblkno);
3134 }
3135 if (adp->ad_lbn >= NDADDR &&
3136 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3137 FREE_LOCK(&lk);
3138 panic("%s: indirect pointer #%ld mismatch %d != %d",
3139 "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
3140 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
3141 }
3142 deplist |= 1 << adp->ad_lbn;
3143 if ((adp->ad_state & ATTACHED) == 0) {
3144 FREE_LOCK(&lk);
3145 panic("softdep_write_inodeblock: Unknown state 0x%x",
3146 adp->ad_state);
3147 }
3148#endif /* DIAGNOSTIC */
3149 adp->ad_state &= ~ATTACHED;
3150 adp->ad_state |= UNDONE;
3151 }
3152 /*
3153 * The on-disk inode cannot claim to be any larger than the last
3154 * fragment that has been written. Otherwise, the on-disk inode
3155 * might have fragments that were not the last block in the file
3156 * which would corrupt the filesystem.
3157 */
3158 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3159 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3160 if (adp->ad_lbn >= NDADDR)
3161 break;
3162 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3163 /* keep going until hitting a rollback to a frag */
3164 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3165 continue;
3166 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3167 for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3168#ifdef DIAGNOSTIC
3169 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3170 FREE_LOCK(&lk);
3171 panic("softdep_write_inodeblock: lost dep1");
3172 }
3173#endif /* DIAGNOSTIC */
3174 dp->di_db[i] = 0;
3175 }
3176 for (i = 0; i < NIADDR; i++) {
3177#ifdef DIAGNOSTIC
3178 if (dp->di_ib[i] != 0 &&
3179 (deplist & ((1 << NDADDR) << i)) == 0) {
3180 FREE_LOCK(&lk);
3181 panic("softdep_write_inodeblock: lost dep2");
3182 }
3183#endif /* DIAGNOSTIC */
3184 dp->di_ib[i] = 0;
3185 }
3186 FREE_LOCK(&lk);
3187 return;
3188 }
3189 /*
3190 * If we have zero'ed out the last allocated block of the file,
3191 * roll back the size to the last currently allocated block.
3192 * We know that this last allocated block is a full-sized as
3193 * we already checked for fragments in the loop above.
3194 */
3195 if (lastadp != NULL &&
3196 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3197 for (i = lastadp->ad_lbn; i >= 0; i--)
3198 if (dp->di_db[i] != 0)
3199 break;
3200 dp->di_size = (i + 1) * fs->fs_bsize;
3201 }
3202 /*
3203 * The only dependencies are for indirect blocks.
3204 *
3205 * The file size for indirect block additions is not guaranteed.
3206 * Such a guarantee would be non-trivial to achieve. The conventional
3207 * synchronous write implementation also does not make this guarantee.
3208 * Fsck should catch and fix discrepancies. Arguably, the file size
3209 * can be over-estimated without destroying integrity when the file
3210 * moves into the indirect blocks (i.e., is large). If we want to
3211 * postpone fsck, we are stuck with this argument.
3212 */
3213 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3214 dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3215 FREE_LOCK(&lk);
3216}
3217
3218/*
3219 * This routine is called during the completion interrupt
3220 * service routine for a disk write (from the procedure called
f719c866 3221 * by the device driver to inform the filesystem caches of
984263bc
MD
3222 * a request completion). It should be called early in this
3223 * procedure, before the block is made available to other
3224 * processes or other routines are called.
3225 */
3226static void
3227softdep_disk_write_complete(bp)
3228 struct buf *bp; /* describes the completed disk write */
3229{
3230 struct worklist *wk;
3231 struct workhead reattach;
3232 struct newblk *newblk;
3233 struct allocindir *aip;
3234 struct allocdirect *adp;
3235 struct indirdep *indirdep;
3236 struct inodedep *inodedep;
3237 struct bmsafemap *bmsafemap;
3238
3239#ifdef DEBUG
dadab5e9 3240 if (lk.lkt_held != NOHOLDER)
984263bc 3241 panic("softdep_disk_write_complete: lock is held");
dadab5e9 3242 lk.lkt_held = SPECIAL_FLAG;
984263bc
MD
3243#endif
3244 LIST_INIT(&reattach);
3245 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3246 WORKLIST_REMOVE(wk);
3247 switch (wk->wk_type) {
3248
3249 case D_PAGEDEP:
3250 if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3251 WORKLIST_INSERT(&reattach, wk);
3252 continue;
3253
3254 case D_INODEDEP:
3255 if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3256 WORKLIST_INSERT(&reattach, wk);
3257 continue;
3258
3259 case D_BMSAFEMAP:
3260 bmsafemap = WK_BMSAFEMAP(wk);
3261 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3262 newblk->nb_state |= DEPCOMPLETE;
3263 newblk->nb_bmsafemap = NULL;
3264 LIST_REMOVE(newblk, nb_deps);
3265 }
3266 while ((adp =
3267 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3268 adp->ad_state |= DEPCOMPLETE;
3269 adp->ad_buf = NULL;
3270 LIST_REMOVE(adp, ad_deps);
3271 handle_allocdirect_partdone(adp);
3272 }
3273 while ((aip =
3274 LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3275 aip->ai_state |= DEPCOMPLETE;
3276 aip->ai_buf = NULL;
3277 LIST_REMOVE(aip, ai_deps);
3278 handle_allocindir_partdone(aip);
3279 }
3280 while ((inodedep =
3281 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3282 inodedep->id_state |= DEPCOMPLETE;
3283 LIST_REMOVE(inodedep, id_deps);
3284 inodedep->id_buf = NULL;
3285 }
3286 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3287 continue;
3288
3289 case D_MKDIR:
3290 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3291 continue;
3292
3293 case D_ALLOCDIRECT:
3294 adp = WK_ALLOCDIRECT(wk);
3295 adp->ad_state |= COMPLETE;
3296 handle_allocdirect_partdone(adp);
3297 continue;
3298
3299 case D_ALLOCINDIR:
3300 aip = WK_ALLOCINDIR(wk);
3301 aip->ai_state |= COMPLETE;
3302 handle_allocindir_partdone(aip);
3303 continue;
3304
3305 case D_INDIRDEP:
3306 indirdep = WK_INDIRDEP(wk);
3307 if (indirdep->ir_state & GOINGAWAY) {
dadab5e9 3308 lk.lkt_held = NOHOLDER;
984263bc
MD
3309 panic("disk_write_complete: indirdep gone");
3310 }
3311 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3312 FREE(indirdep->ir_saveddata, M_INDIRDEP);
3313 indirdep->ir_saveddata = 0;
3314 indirdep->ir_state &= ~UNDONE;
3315 indirdep->ir_state |= ATTACHED;
3316 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3317 handle_allocindir_partdone(aip);
3318 if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
dadab5e9 3319 lk.lkt_held = NOHOLDER;
984263bc
MD
3320 panic("disk_write_complete: not gone");
3321 }
3322 }
3323 WORKLIST_INSERT(&reattach, wk);
3324 if ((bp->b_flags & B_DELWRI) == 0)
3325 stat_indir_blk_ptrs++;
3326 bdirty(bp);
3327 continue;
3328
3329 default:
dadab5e9 3330 lk.lkt_held = NOHOLDER;
984263bc
MD
3331 panic("handle_disk_write_complete: Unknown type %s",
3332 TYPENAME(wk->wk_type));
3333 /* NOTREACHED */
3334 }
3335 }
3336 /*
3337 * Reattach any requests that must be redone.
3338 */
3339 while ((wk = LIST_FIRST(&reattach)) != NULL) {
3340 WORKLIST_REMOVE(wk);
3341 WORKLIST_INSERT(&bp->b_dep, wk);
3342 }
3343#ifdef DEBUG
dadab5e9 3344 if (lk.lkt_held != SPECIAL_FLAG)
984263bc 3345 panic("softdep_disk_write_complete: lock lost");
dadab5e9 3346 lk.lkt_held = NOHOLDER;
984263bc
MD
3347#endif
3348}
3349
3350/*
3351 * Called from within softdep_disk_write_complete above. Note that
3352 * this routine is always called from interrupt level with further
3353 * splbio interrupts blocked.
3354 */
3355static void
3356handle_allocdirect_partdone(adp)
3357 struct allocdirect *adp; /* the completed allocdirect */
3358{
3359 struct allocdirect *listadp;
3360 struct inodedep *inodedep;
3361 long bsize;
3362
3363 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3364 return;
3365 if (adp->ad_buf != NULL) {
dadab5e9 3366 lk.lkt_held = NOHOLDER;
984263bc
MD
3367 panic("handle_allocdirect_partdone: dangling dep");
3368 }
3369 /*
3370 * The on-disk inode cannot claim to be any larger than the last
3371 * fragment that has been written. Otherwise, the on-disk inode
3372 * might have fragments that were not the last block in the file
3373 * which would corrupt the filesystem. Thus, we cannot free any
3374 * allocdirects after one whose ad_oldblkno claims a fragment as
3375 * these blocks must be rolled back to zero before writing the inode.
3376 * We check the currently active set of allocdirects in id_inoupdt.
3377 */
3378 inodedep = adp->ad_inodedep;
3379 bsize = inodedep->id_fs->fs_bsize;
3380 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) {
3381 /* found our block */
3382 if (listadp == adp)
3383 break;
3384 /* continue if ad_oldlbn is not a fragment */
3385 if (listadp->ad_oldsize == 0 ||
3386 listadp->ad_oldsize == bsize)
3387 continue;
3388 /* hit a fragment */
3389 return;
3390 }
3391 /*
3392 * If we have reached the end of the current list without
3393 * finding the just finished dependency, then it must be
3394 * on the future dependency list. Future dependencies cannot
3395 * be freed until they are moved to the current list.
3396 */
3397 if (listadp == NULL) {
3398#ifdef DEBUG
3399 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next)
3400 /* found our block */
3401 if (listadp == adp)
3402 break;
3403 if (listadp == NULL) {
dadab5e9 3404 lk.lkt_held = NOHOLDER;
984263bc
MD
3405 panic("handle_allocdirect_partdone: lost dep");
3406 }
3407#endif /* DEBUG */
3408 return;
3409 }
3410 /*
3411 * If we have found the just finished dependency, then free
3412 * it along with anything that follows it that is complete.
3413 */
3414 for (; adp; adp = listadp) {
3415 listadp = TAILQ_NEXT(adp, ad_next);
3416 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3417 return;
3418 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
3419 }
3420}
3421
3422/*
3423 * Called from within softdep_disk_write_complete above. Note that
3424 * this routine is always called from interrupt level with further
3425 * splbio interrupts blocked.
3426 */
3427static void
3428handle_allocindir_partdone(aip)
3429 struct allocindir *aip; /* the completed allocindir */
3430{
3431 struct indirdep *indirdep;
3432
3433 if ((aip->ai_state & ALLCOMPLETE) != ALLCOMPLETE)
3434 return;
3435 if (aip->ai_buf != NULL) {
dadab5e9 3436 lk.lkt_held = NOHOLDER;
984263bc
MD
3437 panic("handle_allocindir_partdone: dangling dependency");
3438 }
3439 indirdep = aip->ai_indirdep;
3440 if (indirdep->ir_state & UNDONE) {
3441 LIST_REMOVE(aip, ai_next);
3442 LIST_INSERT_HEAD(&indirdep->ir_donehd, aip, ai_next);
3443 return;
3444 }
3445 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)[aip->ai_offset] =
3446 aip->ai_newblkno;
3447 LIST_REMOVE(aip, ai_next);
3448 if (aip->ai_freefrag != NULL)
3449 add_to_worklist(&aip->ai_freefrag->ff_list);
3450 WORKITEM_FREE(aip, D_ALLOCINDIR);
3451}
3452
3453/*
3454 * Called from within softdep_disk_write_complete above to restore
3455 * in-memory inode block contents to their most up-to-date state. Note
3456 * that this routine is always called from interrupt level with further
3457 * splbio interrupts blocked.
3458 */
3459static int
3460handle_written_inodeblock(inodedep, bp)
3461 struct inodedep *inodedep;
3462 struct buf *bp; /* buffer containing the inode block */
3463{
3464 struct worklist *wk, *filefree;
3465 struct allocdirect *adp, *nextadp;
3466 struct dinode *dp;
3467 int hadchanges;
3468
3469 if ((inodedep->id_state & IOSTARTED) == 0) {
dadab5e9 3470 lk.lkt_held = NOHOLDER;
984263bc
MD
3471 panic("handle_written_inodeblock: not started");
3472 }
3473 inodedep->id_state &= ~IOSTARTED;
3474 inodedep->id_state |= COMPLETE;
3475 dp = (struct dinode *)bp->b_data +
3476 ino_to_fsbo(inodedep->id_fs, inodedep->id_ino);
3477 /*
3478 * If we had to rollback the inode allocation because of
3479 * bitmaps being incomplete, then simply restore it.
3480 * Keep the block dirty so that it will not be reclaimed until
3481 * all associated dependencies have been cleared and the
3482 * corresponding updates written to disk.
3483 */
3484 if (inodedep->id_savedino != NULL) {
3485 *dp = *inodedep->id_savedino;
3486 FREE(inodedep->id_savedino, M_INODEDEP);
3487 inodedep->id_savedino = NULL;
3488 if ((bp->b_flags & B_DELWRI) == 0)
3489 stat_inode_bitmap++;
3490 bdirty(bp);
3491 return (1);
3492 }
3493 /*
3494 * Roll forward anything that had to be rolled back before
3495 * the inode could be updated.
3496 */
3497 hadchanges = 0;
3498 for (adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp; adp = nextadp) {
3499 nextadp = TAILQ_NEXT(adp, ad_next);
3500 if (adp->ad_state & ATTACHED) {
dadab5e9 3501 lk.lkt_held = NOHOLDER;
984263bc
MD
3502 panic("handle_written_inodeblock: new entry");
3503 }
3504 if (adp->ad_lbn < NDADDR) {
3505 if (dp->di_db[adp->ad_lbn] != adp->ad_oldblkno) {
dadab5e9 3506 lk.lkt_held = NOHOLDER;
984263bc
MD
3507 panic("%s: %s #%ld mismatch %d != %d",
3508 "handle_written_inodeblock",
3509 "direct pointer", adp->ad_lbn,
3510 dp->di_db[adp->ad_lbn], adp->ad_oldblkno);
3511 }
3512 dp->di_db[adp->ad_lbn] = adp->ad_newblkno;
3513 } else {
3514 if (dp->di_ib[adp->ad_lbn - NDADDR] != 0) {
dadab5e9 3515 lk.lkt_held = NOHOLDER;
984263bc
MD
3516 panic("%s: %s #%ld allocated as %d",
3517 "handle_written_inodeblock",
3518 "indirect pointer", adp->ad_lbn - NDADDR,
3519 dp->di_ib[adp->ad_lbn - NDADDR]);
3520 }
3521 dp->di_ib[adp->ad_lbn - NDADDR] = adp->ad_newblkno;
3522 }
3523 adp->ad_state &= ~UNDONE;
3524 adp->ad_state |= ATTACHED;
3525 hadchanges = 1;
3526 }
3527 if (hadchanges && (bp->b_flags & B_DELWRI) == 0)
3528 stat_direct_blk_ptrs++;
3529 /*
3530 * Reset the file size to its most up-to-date value.
3531 */
3532 if (inodedep->id_savedsize == -1) {
dadab5e9 3533 lk.lkt_held = NOHOLDER;
984263bc
MD
3534 panic("handle_written_inodeblock: bad size");
3535 }
3536 if (dp->di_size != inodedep->id_savedsize) {
3537 dp->di_size = inodedep->id_savedsize;
3538 hadchanges = 1;
3539 }
3540 inodedep->id_savedsize = -1;
3541 /*
3542 * If there were any rollbacks in the inode block, then it must be
3543 * marked dirty so that its will eventually get written back in
3544 * its correct form.
3545 */
3546 if (hadchanges)
3547 bdirty(bp);
3548 /*
3549 * Process any allocdirects that completed during the update.
3550 */
3551 if ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != NULL)
3552 handle_allocdirect_partdone(adp);
3553 /*
3554 * Process deallocations that were held pending until the
3555 * inode had been written to disk. Freeing of the inode
3556 * is delayed until after all blocks have been freed to
3557 * avoid creation of new <vfsid, inum, lbn> triples
3558 * before the old ones have been deleted.
3559 */
3560 filefree = NULL;
3561 while ((wk = LIST_FIRST(&inodedep->id_bufwait)) != NULL) {
3562 WORKLIST_REMOVE(wk);
3563 switch (wk->wk_type) {
3564
3565 case D_FREEFILE:
3566 /*
3567 * We defer adding filefree to the worklist until
3568 * all other additions have been made to ensure
3569 * that it will be done after all the old blocks
3570 * have been freed.
3571 */
3572 if (filefree != NULL) {
dadab5e9 3573 lk.lkt_held = NOHOLDER;
984263bc
MD
3574 panic("handle_written_inodeblock: filefree");
3575 }
3576 filefree = wk;
3577 continue;
3578
3579 case D_MKDIR:
3580 handle_written_mkdir(WK_MKDIR(wk), MKDIR_PARENT);
3581 continue;
3582
3583 case D_DIRADD:
3584 diradd_inode_written(WK_DIRADD(wk), inodedep);
3585 continue;
3586
3587 case D_FREEBLKS:
3588 case D_FREEFRAG:
3589 case D_DIRREM:
3590 add_to_worklist(wk);
3591 continue;
3592
3593 default:
dadab5e9 3594 lk.lkt_held = NOHOLDER;
984263bc
MD
3595 panic("handle_written_inodeblock: Unknown type %s",
3596 TYPENAME(wk->wk_type));
3597 /* NOTREACHED */
3598 }
3599 }
3600 if (filefree != NULL) {
3601 if (free_inodedep(inodedep) == 0) {
dadab5e9 3602 lk.lkt_held = NOHOLDER;
984263bc
MD
3603 panic("handle_written_inodeblock: live inodedep");
3604 }
3605 add_to_worklist(filefree);
3606 return (0);
3607 }
3608
3609 /*
3610 * If no outstanding dependencies, free it.
3611 */
3612 if (free_inodedep(inodedep) || TAILQ_FIRST(&inodedep->id_inoupdt) == 0)
3613 return (0);
3614 return (hadchanges);
3615}
3616
3617/*
3618 * Process a diradd entry after its dependent inode has been written.
3619 * This routine must be called with splbio interrupts blocked.
3620 */
3621static void
3622diradd_inode_written(dap, inodedep)
3623 struct diradd *dap;
3624 struct inodedep *inodedep;
3625{
3626 struct pagedep *pagedep;
3627
3628 dap->da_state |= COMPLETE;
3629 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3630 if (dap->da_state & DIRCHG)
3631 pagedep = dap->da_previous->dm_pagedep;
3632 else
3633 pagedep = dap->da_pagedep;
3634 LIST_REMOVE(dap, da_pdlist);
3635 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3636 }
3637 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
3638}
3639
3640/*
3641 * Handle the completion of a mkdir dependency.
3642 */
3643static void
3644handle_written_mkdir(mkdir, type)
3645 struct mkdir *mkdir;
3646 int type;
3647{
3648 struct diradd *dap;
3649 struct pagedep *pagedep;
3650
3651 if (mkdir->md_state != type) {
dadab5e9 3652 lk.lkt_held = NOHOLDER;
984263bc
MD
3653 panic("handle_written_mkdir: bad type");
3654 }
3655 dap = mkdir->md_diradd;
3656 dap->da_state &= ~type;
3657 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) == 0)
3658 dap->da_state |= DEPCOMPLETE;
3659 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3660 if (dap->da_state & DIRCHG)
3661 pagedep = dap->da_previous->dm_pagedep;
3662 else
3663 pagedep = dap->da_pagedep;
3664 LIST_REMOVE(dap, da_pdlist);
3665 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
3666 }
3667 LIST_REMOVE(mkdir, md_mkdirs);
3668 WORKITEM_FREE(mkdir, D_MKDIR);
3669}
3670
3671/*
3672 * Called from within softdep_disk_write_complete above.
3673 * A write operation was just completed. Removed inodes can
3674 * now be freed and associated block pointers may be committed.
3675 * Note that this routine is always called from interrupt level
3676 * with further splbio interrupts blocked.
3677 */
3678static int
3679handle_written_filepage(pagedep, bp)
3680 struct pagedep *pagedep;
3681 struct buf *bp; /* buffer containing the written page */
3682{
3683 struct dirrem *dirrem;
3684 struct diradd *dap, *nextdap;
3685 struct direct *ep;
3686 int i, chgs;
3687
3688 if ((pagedep->pd_state & IOSTARTED) == 0) {
dadab5e9 3689 lk.lkt_held = NOHOLDER;
984263bc
MD
3690 panic("handle_written_filepage: not started");
3691 }
3692 pagedep->pd_state &= ~IOSTARTED;
3693 /*
3694 * Process any directory removals that have been committed.
3695 */
3696 while ((dirrem = LIST_FIRST(&pagedep->pd_dirremhd)) != NULL) {
3697 LIST_REMOVE(dirrem, dm_next);
3698 dirrem->dm_dirinum = pagedep->pd_ino;
3699 add_to_worklist(&dirrem->dm_list);
3700 }
3701 /*
3702 * Free any directory additions that have been committed.
3703 */
3704 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != NULL)
3705 free_diradd(dap);
3706 /*
3707 * Uncommitted directory entries must be restored.
3708 */
3709 for (chgs = 0, i = 0; i < DAHASHSZ; i++) {
3710 for (dap = LIST_FIRST(&pagedep->pd_diraddhd[i]); dap;
3711 dap = nextdap) {
3712 nextdap = LIST_NEXT(dap, da_pdlist);
3713 if (dap->da_state & ATTACHED) {
dadab5e9 3714 lk.lkt_held = NOHOLDER;
984263bc
MD
3715 panic("handle_written_filepage: attached");
3716 }
3717 ep = (struct direct *)
3718 ((char *)bp->b_data + dap->da_offset);
3719 ep->d_ino = dap->da_newinum;
3720 dap->da_state &= ~UNDONE;
3721 dap->da_state |= ATTACHED;
3722 chgs = 1;
3723 /*
3724 * If the inode referenced by the directory has
3725 * been written out, then the dependency can be
3726 * moved to the pending list.
3727 */
3728 if ((dap->da_state & ALLCOMPLETE) == ALLCOMPLETE) {
3729 LIST_REMOVE(dap, da_pdlist);
3730 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap,
3731 da_pdlist);
3732 }
3733 }
3734 }
3735 /*
3736 * If there were any rollbacks in the directory, then it must be
3737 * marked dirty so that its will eventually get written back in
3738 * its correct form.
3739 */
3740 if (chgs) {
3741 if ((bp->b_flags & B_DELWRI) == 0)
3742 stat_dir_entry++;
3743 bdirty(bp);
3744 }
3745 /*
3746 * If no dependencies remain, the pagedep will be freed.
3747 * Otherwise it will remain to update the page before it
3748 * is written back to disk.
3749 */
3750 if (LIST_FIRST(&pagedep->pd_pendinghd) == 0) {
3751 for (i = 0; i < DAHASHSZ; i++)
3752 if (LIST_FIRST(&pagedep->pd_diraddhd[i]) != NULL)
3753 break;
3754 if (i == DAHASHSZ) {
3755 LIST_REMOVE(pagedep, pd_hash);
3756 WORKITEM_FREE(pagedep, D_PAGEDEP);
3757 return (0);
3758 }
3759 }
3760 return (1);
3761}
3762
3763/*
3764 * Writing back in-core inode structures.
3765 *
f719c866 3766 * The filesystem only accesses an inode's contents when it occupies an
984263bc
MD
3767 * "in-core" inode structure. These "in-core" structures are separate from
3768 * the page frames used to cache inode blocks. Only the latter are
3769 * transferred to/from the disk. So, when the updated contents of the
3770 * "in-core" inode structure are copied to the corresponding in-memory inode
3771 * block, the dependencies are also transferred. The following procedure is
3772 * called when copying a dirty "in-core" inode to a cached inode block.
3773 */
3774
3775/*
3776 * Called when an inode is loaded from disk. If the effective link count
3777 * differed from the actual link count when it was last flushed, then we
3778 * need to ensure that the correct effective link count is put back.
3779 */
3780void
3781softdep_load_inodeblock(ip)
3782 struct inode *ip; /* the "in_core" copy of the inode */
3783{
3784 struct inodedep *inodedep;
3785
3786 /*
3787 * Check for alternate nlink count.
3788 */
3789 ip->i_effnlink = ip->i_nlink;
3790 ACQUIRE_LOCK(&lk);
3791 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3792 FREE_LOCK(&lk);
3793 return;
3794 }
3795 ip->i_effnlink -= inodedep->id_nlinkdelta;
3796 FREE_LOCK(&lk);
3797}
3798
3799/*
3800 * This routine is called just before the "in-core" inode
3801 * information is to be copied to the in-memory inode block.
3802 * Recall that an inode block contains several inodes. If
3803 * the force flag is set, then the dependencies will be
3804 * cleared so that the update can always be made. Note that
3805 * the buffer is locked when this routine is called, so we
3806 * will never be in the middle of writing the inode block
3807 * to disk.
3808 */
3809void
3810softdep_update_inodeblock(ip, bp, waitfor)
3811 struct inode *ip; /* the "in_core" copy of the inode */
3812 struct buf *bp; /* the buffer containing the inode block */
3813 int waitfor; /* nonzero => update must be allowed */
3814{
3815 struct inodedep *inodedep;
3816 struct worklist *wk;
3817 int error, gotit;
3818
3819 /*
3820 * If the effective link count is not equal to the actual link
3821 * count, then we must track the difference in an inodedep while
3822 * the inode is (potentially) tossed out of the cache. Otherwise,
3823 * if there is no existing inodedep, then there are no dependencies
3824 * to track.
3825 */
3826 ACQUIRE_LOCK(&lk);
3827 if (inodedep_lookup(ip->i_fs, ip->i_number, 0, &inodedep) == 0) {
3828 FREE_LOCK(&lk);
3829 if (ip->i_effnlink != ip->i_nlink)
3830 panic("softdep_update_inodeblock: bad link count");
3831 return;
3832 }
3833 if (inodedep->id_nlinkdelta != ip->i_nlink - ip->i_effnlink) {
3834 FREE_LOCK(&lk);
3835 panic("softdep_update_inodeblock: bad delta");
3836 }
3837 /*
3838 * Changes have been initiated. Anything depending on these
3839 * changes cannot occur until this inode has been written.
3840 */
3841 inodedep->id_state &= ~COMPLETE;
3842 if ((inodedep->id_state & ONWORKLIST) == 0)
3843 WORKLIST_INSERT(&bp->b_dep, &inodedep->id_list);
3844 /*
3845 * Any new dependencies associated with the incore inode must
3846 * now be moved to the list associated with the buffer holding
3847 * the in-memory copy of the inode. Once merged process any
3848 * allocdirects that are completed by the merger.