kernel - Provide descriptions for lwkt.* and debug.* sysctl's
[dragonfly.git] / sys / vfs / ufs / ffs_softdep.c
CommitLineData
984263bc
MD
1/*
2 * Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
3 *
4 * The soft updates code is derived from the appendix of a University
5 * of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
6 * "Soft Updates: A Solution to the Metadata Update Problem in File
7 * Systems", CSE-TR-254-95, August 1995).
8 *
9 * Further information about soft updates can be obtained from:
10 *
11 * Marshall Kirk McKusick http://www.mckusick.com/softdep/
12 * 1614 Oxford Street mckusick@mckusick.com
13 * Berkeley, CA 94709-1608 +1-510-843-9542
14 * USA
15 *
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
18 * are met:
19 *
20 * 1. Redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer.
22 * 2. Redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution.
25 *
26 * THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
27 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
28 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
29 * DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
30 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * SUCH DAMAGE.
37 *
38 * from: @(#)ffs_softdep.c 9.59 (McKusick) 6/21/00
39 * $FreeBSD: src/sys/ufs/ffs/ffs_softdep.c,v 1.57.2.11 2002/02/05 18:46:53 dillon Exp $
c4df9635 40 * $DragonFly: src/sys/vfs/ufs/ffs_softdep.c,v 1.57 2008/06/28 17:59:51 dillon Exp $
984263bc
MD
41 */
42
43/*
44 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
45 */
46#ifndef DIAGNOSTIC
47#define DIAGNOSTIC
48#endif
49#ifndef DEBUG
50#define DEBUG
51#endif
52
53#include <sys/param.h>
54#include <sys/kernel.h>
55#include <sys/systm.h>
56#include <sys/buf.h>
57#include <sys/malloc.h>
58#include <sys/mount.h>
59#include <sys/proc.h>
60#include <sys/syslog.h>
61#include <sys/vnode.h>
62#include <sys/conf.h>
f91a71dd 63#include <machine/inttypes.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 72
59a647b1
MD
73#include <sys/buf2.h>
74#include <sys/mplock2.h>
165dba55
DR
75#include <sys/thread2.h>
76
984263bc
MD
77/*
78 * These definitions need to be adapted to the system to which
79 * this file is being ported.
80 */
81/*
82 * malloc types defined for the softdep system.
83 */
84MALLOC_DEFINE(M_PAGEDEP, "pagedep","File page dependencies");
85MALLOC_DEFINE(M_INODEDEP, "inodedep","Inode dependencies");
86MALLOC_DEFINE(M_NEWBLK, "newblk","New block allocation");
87MALLOC_DEFINE(M_BMSAFEMAP, "bmsafemap","Block or frag allocated from cyl group map");
88MALLOC_DEFINE(M_ALLOCDIRECT, "allocdirect","Block or frag dependency for an inode");
89MALLOC_DEFINE(M_INDIRDEP, "indirdep","Indirect block dependencies");
90MALLOC_DEFINE(M_ALLOCINDIR, "allocindir","Block dependency for an indirect block");
91MALLOC_DEFINE(M_FREEFRAG, "freefrag","Previously used frag for an inode");
92MALLOC_DEFINE(M_FREEBLKS, "freeblks","Blocks freed from an inode");
93MALLOC_DEFINE(M_FREEFILE, "freefile","Inode deallocated");
94MALLOC_DEFINE(M_DIRADD, "diradd","New directory entry");
95MALLOC_DEFINE(M_MKDIR, "mkdir","New directory");
96MALLOC_DEFINE(M_DIRREM, "dirrem","Directory entry deleted");
97
98#define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE)
99
100#define D_PAGEDEP 0
101#define D_INODEDEP 1
102#define D_NEWBLK 2
103#define D_BMSAFEMAP 3
104#define D_ALLOCDIRECT 4
105#define D_INDIRDEP 5
106#define D_ALLOCINDIR 6
107#define D_FREEFRAG 7
108#define D_FREEBLKS 8
109#define D_FREEFILE 9
110#define D_DIRADD 10
111#define D_MKDIR 11
112#define D_DIRREM 12
113#define D_LAST D_DIRREM
114
115/*
116 * translate from workitem type to memory type
117 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
118 */
119static struct malloc_type *memtype[] = {
120 M_PAGEDEP,
121 M_INODEDEP,
122 M_NEWBLK,
123 M_BMSAFEMAP,
124 M_ALLOCDIRECT,
125 M_INDIRDEP,
126 M_ALLOCINDIR,
127 M_FREEFRAG,
128 M_FREEBLKS,
129 M_FREEFILE,
130 M_DIRADD,
131 M_MKDIR,
132 M_DIRREM
133};
134
135#define DtoM(type) (memtype[type])
136
137/*
138 * Names of malloc types.
139 */
140#define TYPENAME(type) \
141 ((unsigned)(type) < D_LAST ? memtype[type]->ks_shortdesc : "???")
984263bc
MD
142/*
143 * End system adaptaion definitions.
144 */
145
146/*
147 * Internal function prototypes.
148 */
f719c866
DR
149static void softdep_error(char *, int);
150static void drain_output(struct vnode *, int);
151static int getdirtybuf(struct buf **, int);
152static void clear_remove(struct thread *);
153static void clear_inodedeps(struct thread *);
154static int flush_pagedep_deps(struct vnode *, struct mount *,
a6ee311a 155 struct diraddhd *);
f719c866
DR
156static int flush_inodedep_deps(struct fs *, ino_t);
157static int handle_written_filepage(struct pagedep *, struct buf *);
158static void diradd_inode_written(struct diradd *, struct inodedep *);
159static int handle_written_inodeblock(struct inodedep *, struct buf *);
160static void handle_allocdirect_partdone(struct allocdirect *);
161static void handle_allocindir_partdone(struct allocindir *);
162static void initiate_write_filepage(struct pagedep *, struct buf *);
163static void handle_written_mkdir(struct mkdir *, int);
164static void initiate_write_inodeblock(struct inodedep *, struct buf *);
165static void handle_workitem_freefile(struct freefile *);
166static void handle_workitem_remove(struct dirrem *);
167static struct dirrem *newdirrem(struct buf *, struct inode *,
a6ee311a 168 struct inode *, int, struct dirrem **);
f719c866
DR
169static void free_diradd(struct diradd *);
170static void free_allocindir(struct allocindir *, struct inodedep *);
54078292 171static int indir_trunc (struct inode *, off_t, int, ufs_lbn_t, long *);
f719c866
DR
172static void deallocate_dependencies(struct buf *, struct inodedep *);
173static void free_allocdirect(struct allocdirectlst *,
a6ee311a 174 struct allocdirect *, int);
f719c866
DR
175static int check_inode_unwritten(struct inodedep *);
176static int free_inodedep(struct inodedep *);
177static void handle_workitem_freeblocks(struct freeblks *);
178static void merge_inode_lists(struct inodedep *);
179static void setup_allocindir_phase2(struct buf *, struct inode *,
a6ee311a 180 struct allocindir *);
f719c866 181static struct allocindir *newallocindir(struct inode *, int, ufs_daddr_t,
a6ee311a 182 ufs_daddr_t);
f719c866
DR
183static void handle_workitem_freefrag(struct freefrag *);
184static struct freefrag *newfreefrag(struct inode *, ufs_daddr_t, long);
185static void allocdirect_merge(struct allocdirectlst *,
a6ee311a 186 struct allocdirect *, struct allocdirect *);
f719c866
DR
187static struct bmsafemap *bmsafemap_lookup(struct buf *);
188static int newblk_lookup(struct fs *, ufs_daddr_t, int,
a6ee311a 189 struct newblk **);
f719c866
DR
190static int inodedep_lookup(struct fs *, ino_t, int, struct inodedep **);
191static int pagedep_lookup(struct inode *, ufs_lbn_t, int,
a6ee311a 192 struct pagedep **);
f719c866
DR
193static void pause_timer(void *);
194static int request_cleanup(int, int);
195static int process_worklist_item(struct mount *, int);
196static void add_to_worklist(struct worklist *);
984263bc
MD
197
198/*
199 * Exported softdep operations.
200 */
f719c866
DR
201static void softdep_disk_io_initiation(struct buf *);
202static void softdep_disk_write_complete(struct buf *);
203static void softdep_deallocate_dependencies(struct buf *);
204static int softdep_fsync(struct vnode *);
205static int softdep_process_worklist(struct mount *);
206static void softdep_move_dependencies(struct buf *, struct buf *);
207static int softdep_count_dependencies(struct buf *bp, int);
27bc0cb1
MD
208static int softdep_checkread(struct buf *bp);
209static int softdep_checkwrite(struct buf *bp);
984263bc 210
bc50d880 211static struct bio_ops softdep_bioops = {
408357d8
MD
212 .io_start = softdep_disk_io_initiation,
213 .io_complete = softdep_disk_write_complete,
214 .io_deallocate = softdep_deallocate_dependencies,
215 .io_fsync = softdep_fsync,
216 .io_sync = softdep_process_worklist,
217 .io_movedeps = softdep_move_dependencies,
218 .io_countdeps = softdep_count_dependencies,
27bc0cb1
MD
219 .io_checkread = softdep_checkread,
220 .io_checkwrite = softdep_checkwrite
984263bc
MD
221};
222
223/*
224 * Locking primitives.
225 *
226 * For a uniprocessor, all we need to do is protect against disk
227 * interrupts. For a multiprocessor, this lock would have to be
228 * a mutex. A single mutex is used throughout this file, though
229 * finer grain locking could be used if contention warranted it.
230 *
231 * For a multiprocessor, the sleep call would accept a lock and
232 * release it after the sleep processing was complete. In a uniprocessor
233 * implementation there is no such interlock, so we simple mark
234 * the places where it needs to be done with the `interlocked' form
235 * of the lock calls. Since the uniprocessor sleep already interlocks
236 * the spl, there is nothing that really needs to be done.
237 */
238#ifndef /* NOT */ DEBUG
239static struct lockit {
984263bc 240} lk = { 0 };
02d8a449
MD
241#define ACQUIRE_LOCK(lk) crit_enter_id("softupdates");
242#define FREE_LOCK(lk) crit_exit_id("softupdates");
984263bc
MD
243
244#else /* DEBUG */
dadab5e9
MD
245#define NOHOLDER ((struct thread *)-1)
246#define SPECIAL_FLAG ((struct thread *)-2)
984263bc
MD
247static struct lockit {
248 int lkt_spl;
dadab5e9
MD
249 struct thread *lkt_held;
250} lk = { 0, NOHOLDER };
984263bc
MD
251static int lockcnt;
252
f719c866
DR
253static void acquire_lock(struct lockit *);
254static void free_lock(struct lockit *);
255void softdep_panic(char *);
984263bc
MD
256
257#define ACQUIRE_LOCK(lk) acquire_lock(lk)
258#define FREE_LOCK(lk) free_lock(lk)
259
260static void
3fcb1ab8 261acquire_lock(struct lockit *lk)
984263bc 262{
dadab5e9 263 thread_t holder;
984263bc 264
dadab5e9 265 if (lk->lkt_held != NOHOLDER) {
984263bc
MD
266 holder = lk->lkt_held;
267 FREE_LOCK(lk);
dadab5e9 268 if (holder == curthread)
984263bc
MD
269 panic("softdep_lock: locking against myself");
270 else
dadab5e9 271 panic("softdep_lock: lock held by %p", holder);
984263bc 272 }
02d8a449 273 crit_enter_id("softupdates");
dadab5e9 274 lk->lkt_held = curthread;
984263bc
MD
275 lockcnt++;
276}
277
278static void
3fcb1ab8 279free_lock(struct lockit *lk)
984263bc
MD
280{
281
dadab5e9 282 if (lk->lkt_held == NOHOLDER)
984263bc 283 panic("softdep_unlock: lock not held");
dadab5e9 284 lk->lkt_held = NOHOLDER;
02d8a449 285 crit_exit_id("softupdates");
984263bc
MD
286}
287
288/*
289 * Function to release soft updates lock and panic.
290 */
291void
3fcb1ab8 292softdep_panic(char *msg)
984263bc
MD
293{
294
dadab5e9 295 if (lk.lkt_held != NOHOLDER)
984263bc
MD
296 FREE_LOCK(&lk);
297 panic(msg);
298}
299#endif /* DEBUG */
300
f719c866 301static int interlocked_sleep(struct lockit *, int, void *, int,
a6ee311a 302 const char *, int);
984263bc
MD
303
304/*
305 * When going to sleep, we must save our SPL so that it does
306 * not get lost if some other process uses the lock while we
307 * are sleeping. We restore it after we have slept. This routine
308 * wraps the interlocking with functions that sleep. The list
309 * below enumerates the available set of operations.
310 */
311#define UNKNOWN 0
312#define SLEEP 1
313#define LOCKBUF 2
314
315static int
3fcb1ab8
SW
316interlocked_sleep(struct lockit *lk, int op, void *ident, int flags,
317 const char *wmesg, int timo)
984263bc 318{
dadab5e9 319 thread_t holder;
984263bc
MD
320 int s, retval;
321
322 s = lk->lkt_spl;
323# ifdef DEBUG
dadab5e9 324 if (lk->lkt_held == NOHOLDER)
984263bc 325 panic("interlocked_sleep: lock not held");
dadab5e9 326 lk->lkt_held = NOHOLDER;
984263bc
MD
327# endif /* DEBUG */
328 switch (op) {
329 case SLEEP:
330 retval = tsleep(ident, flags, wmesg, timo);
331 break;
332 case LOCKBUF:
333 retval = BUF_LOCK((struct buf *)ident, flags);
334 break;
335 default:
336 panic("interlocked_sleep: unknown operation");
337 }
338# ifdef DEBUG
dadab5e9 339 if (lk->lkt_held != NOHOLDER) {
984263bc
MD
340 holder = lk->lkt_held;
341 FREE_LOCK(lk);
dadab5e9 342 if (holder == curthread)
984263bc
MD
343 panic("interlocked_sleep: locking against self");
344 else
dadab5e9 345 panic("interlocked_sleep: lock held by %p", holder);
984263bc 346 }
dadab5e9 347 lk->lkt_held = curthread;
984263bc
MD
348 lockcnt++;
349# endif /* DEBUG */
350 lk->lkt_spl = s;
351 return (retval);
352}
353
354/*
355 * Place holder for real semaphores.
356 */
357struct sema {
358 int value;
dadab5e9 359 thread_t holder;
984263bc
MD
360 char *name;
361 int prio;
362 int timo;
363};
f719c866
DR
364static void sema_init(struct sema *, char *, int, int);
365static int sema_get(struct sema *, struct lockit *);
366static void sema_release(struct sema *);
984263bc
MD
367
368static void
3fcb1ab8 369sema_init(struct sema *semap, char *name, int prio, int timo)
984263bc
MD
370{
371
dadab5e9 372 semap->holder = NOHOLDER;
984263bc
MD
373 semap->value = 0;
374 semap->name = name;
375 semap->prio = prio;
376 semap->timo = timo;
377}
378
379static int
3fcb1ab8 380sema_get(struct sema *semap, struct lockit *interlock)
984263bc
MD
381{
382
383 if (semap->value++ > 0) {
384 if (interlock != NULL) {
385 interlocked_sleep(interlock, SLEEP, (caddr_t)semap,
386 semap->prio, semap->name, semap->timo);
387 FREE_LOCK(interlock);
388 } else {
389 tsleep((caddr_t)semap, semap->prio, semap->name,
390 semap->timo);
391 }
392 return (0);
393 }
dadab5e9 394 semap->holder = curthread;
984263bc
MD
395 if (interlock != NULL)
396 FREE_LOCK(interlock);
397 return (1);
398}
399
400static void
3fcb1ab8 401sema_release(struct sema *semap)
984263bc
MD
402{
403
dadab5e9
MD
404 if (semap->value <= 0 || semap->holder != curthread) {
405 if (lk.lkt_held != NOHOLDER)
984263bc
MD
406 FREE_LOCK(&lk);
407 panic("sema_release: not held");
408 }
409 if (--semap->value > 0) {
410 semap->value = 0;
411 wakeup(semap);
412 }
dadab5e9 413 semap->holder = NOHOLDER;
984263bc
MD
414}
415
416/*
417 * Worklist queue management.
418 * These routines require that the lock be held.
419 */
420#ifndef /* NOT */ DEBUG
421#define WORKLIST_INSERT(head, item) do { \
422 (item)->wk_state |= ONWORKLIST; \
423 LIST_INSERT_HEAD(head, item, wk_list); \
424} while (0)
408357d8
MD
425
426#define WORKLIST_INSERT_BP(bp, item) do { \
427 (item)->wk_state |= ONWORKLIST; \
428 (bp)->b_ops = &softdep_bioops; \
429 LIST_INSERT_HEAD(&(bp)->b_dep, item, wk_list); \
430} while (0)
431
984263bc
MD
432#define WORKLIST_REMOVE(item) do { \
433 (item)->wk_state &= ~ONWORKLIST; \
434 LIST_REMOVE(item, wk_list); \
435} while (0)
408357d8 436
984263bc
MD
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 443
408357d8
MD
444#define WORKLIST_INSERT_BP(bp, item) do { \
445 (bp)->b_ops = &softdep_bioops; \
446 worklist_insert(&(bp)->b_dep, item); \
447} while (0)
448
984263bc
MD
449#define WORKLIST_INSERT(head, item) worklist_insert(head, item)
450#define WORKLIST_REMOVE(item) worklist_remove(item)
451#define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
452
453static void
3fcb1ab8 454worklist_insert(struct workhead *head, struct worklist *item)
984263bc
MD
455{
456
dadab5e9 457 if (lk.lkt_held == NOHOLDER)
984263bc
MD
458 panic("worklist_insert: lock not held");
459 if (item->wk_state & ONWORKLIST) {
460 FREE_LOCK(&lk);
461 panic("worklist_insert: already on list");
462 }
463 item->wk_state |= ONWORKLIST;
464 LIST_INSERT_HEAD(head, item, wk_list);
465}
466
467static void
3fcb1ab8 468worklist_remove(struct worklist *item)
984263bc
MD
469{
470
dadab5e9 471 if (lk.lkt_held == NOHOLDER)
984263bc
MD
472 panic("worklist_remove: lock not held");
473 if ((item->wk_state & ONWORKLIST) == 0) {
474 FREE_LOCK(&lk);
475 panic("worklist_remove: not on list");
476 }
477 item->wk_state &= ~ONWORKLIST;
478 LIST_REMOVE(item, wk_list);
479}
480
481static void
3fcb1ab8 482workitem_free(struct worklist *item, int type)
984263bc
MD
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>
0c52fa62
SG
532SYSCTL_INT(_debug, OID_AUTO, max_softdeps, CTLFLAG_RW, &max_softdeps, 0,
533 "Maximum soft dependencies before slowdown occurs");
534SYSCTL_INT(_debug, OID_AUTO, tickdelay, CTLFLAG_RW, &tickdelay, 0,
535 "Ticks to delay before allocating during slowdown");
536SYSCTL_INT(_debug, OID_AUTO, worklist_push, CTLFLAG_RW, &stat_worklist_push, 0,
537 "Number of worklist cleanups");
538SYSCTL_INT(_debug, OID_AUTO, blk_limit_push, CTLFLAG_RW, &stat_blk_limit_push, 0,
539 "Number of times block limit neared");
540SYSCTL_INT(_debug, OID_AUTO, ino_limit_push, CTLFLAG_RW, &stat_ino_limit_push, 0,
541 "Number of times inode limit neared");
542SYSCTL_INT(_debug, OID_AUTO, blk_limit_hit, CTLFLAG_RW, &stat_blk_limit_hit, 0,
543 "Number of times block slowdown imposed");
544SYSCTL_INT(_debug, OID_AUTO, ino_limit_hit, CTLFLAG_RW, &stat_ino_limit_hit, 0,
545 "Number of times inode slowdown imposed ");
546SYSCTL_INT(_debug, OID_AUTO, sync_limit_hit, CTLFLAG_RW, &stat_sync_limit_hit, 0,
547 "Number of synchronous slowdowns imposed");
548SYSCTL_INT(_debug, OID_AUTO, indir_blk_ptrs, CTLFLAG_RW, &stat_indir_blk_ptrs, 0,
549 "Bufs redirtied as indir ptrs not written");
550SYSCTL_INT(_debug, OID_AUTO, inode_bitmap, CTLFLAG_RW, &stat_inode_bitmap, 0,
551 "Bufs redirtied as inode bitmap not written");
552SYSCTL_INT(_debug, OID_AUTO, direct_blk_ptrs, CTLFLAG_RW, &stat_direct_blk_ptrs, 0,
553 "Bufs redirtied as direct ptrs not written");
554SYSCTL_INT(_debug, OID_AUTO, dir_entry, CTLFLAG_RW, &stat_dir_entry, 0,
555 "Bufs redirtied as dir entry cannot write");
984263bc
MD
556#endif /* DEBUG */
557
558/*
559 * Add an item to the end of the work queue.
560 * This routine requires that the lock be held.
561 * This is the only routine that adds items to the list.
562 * The following routine is the only one that removes items
563 * and does so in order from first to last.
564 */
565static void
3fcb1ab8 566add_to_worklist(struct worklist *wk)
984263bc
MD
567{
568 static struct worklist *worklist_tail;
569
570 if (wk->wk_state & ONWORKLIST) {
dadab5e9 571 if (lk.lkt_held != NOHOLDER)
984263bc
MD
572 FREE_LOCK(&lk);
573 panic("add_to_worklist: already on list");
574 }
575 wk->wk_state |= ONWORKLIST;
576 if (LIST_FIRST(&softdep_workitem_pending) == NULL)
577 LIST_INSERT_HEAD(&softdep_workitem_pending, wk, wk_list);
578 else
579 LIST_INSERT_AFTER(worklist_tail, wk, wk_list);
580 worklist_tail = wk;
581 num_on_worklist += 1;
582}
583
584/*
585 * Process that runs once per second to handle items in the background queue.
586 *
587 * Note that we ensure that everything is done in the order in which they
588 * appear in the queue. The code below depends on this property to ensure
589 * that blocks of a file are freed before the inode itself is freed. This
590 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
591 * until all the old ones have been purged from the dependency lists.
59a647b1
MD
592 *
593 * bioops callback - hold io_token
984263bc
MD
594 */
595static int
3fcb1ab8 596softdep_process_worklist(struct mount *matchmnt)
984263bc 597{
dadab5e9 598 thread_t td = curthread;
984263bc
MD
599 int matchcnt, loopcount;
600 long starttime;
601
59a647b1
MD
602 get_mplock();
603
984263bc
MD
604 /*
605 * Record the process identifier of our caller so that we can give
606 * this process preferential treatment in request_cleanup below.
607 */
dadab5e9 608 filesys_syncer = td;
984263bc
MD
609 matchcnt = 0;
610
611 /*
612 * There is no danger of having multiple processes run this
613 * code, but we have to single-thread it when softdep_flushfiles()
614 * is in operation to get an accurate count of the number of items
615 * related to its mount point that are in the list.
616 */
617 if (matchmnt == NULL) {
59a647b1
MD
618 if (softdep_worklist_busy < 0) {
619 matchcnt = -1;
620 goto done;
621 }
984263bc
MD
622 softdep_worklist_busy += 1;
623 }
624
625 /*
626 * If requested, try removing inode or removal dependencies.
627 */
628 if (req_clear_inodedeps) {
dadab5e9 629 clear_inodedeps(td);
984263bc
MD
630 req_clear_inodedeps -= 1;
631 wakeup_one(&proc_waiting);
632 }
633 if (req_clear_remove) {
dadab5e9 634 clear_remove(td);
984263bc
MD
635 req_clear_remove -= 1;
636 wakeup_one(&proc_waiting);
637 }
638 loopcount = 1;
639 starttime = time_second;
640 while (num_on_worklist > 0) {
641 matchcnt += process_worklist_item(matchmnt, 0);
642
643 /*
644 * If a umount operation wants to run the worklist
645 * accurately, abort.
646 */
647 if (softdep_worklist_req && matchmnt == NULL) {
648 matchcnt = -1;
649 break;
650 }
651
652 /*
653 * If requested, try removing inode or removal dependencies.
654 */
655 if (req_clear_inodedeps) {
dadab5e9 656 clear_inodedeps(td);
984263bc
MD
657 req_clear_inodedeps -= 1;
658 wakeup_one(&proc_waiting);
659 }
660 if (req_clear_remove) {
dadab5e9 661 clear_remove(td);
984263bc
MD
662 req_clear_remove -= 1;
663 wakeup_one(&proc_waiting);
664 }
665 /*
666 * We do not generally want to stop for buffer space, but if
667 * we are really being a buffer hog, we will stop and wait.
668 */
669 if (loopcount++ % 128 == 0)
c4df9635 670 bwillinode(1);
984263bc
MD
671 /*
672 * Never allow processing to run for more than one
673 * second. Otherwise the other syncer tasks may get
674 * excessively backlogged.
675 */
676 if (starttime != time_second && matchmnt == NULL) {
677 matchcnt = -1;
678 break;
679 }
680 }
681 if (matchmnt == NULL) {
682 --softdep_worklist_busy;
683 if (softdep_worklist_req && softdep_worklist_busy == 0)
684 wakeup(&softdep_worklist_req);
685 }
59a647b1
MD
686done:
687 rel_mplock();
984263bc
MD
688 return (matchcnt);
689}
690
691/*
692 * Process one item on the worklist.
693 */
694static int
3fcb1ab8 695process_worklist_item(struct mount *matchmnt, int flags)
984263bc
MD
696{
697 struct worklist *wk;
698 struct dirrem *dirrem;
699 struct fs *matchfs;
700 struct vnode *vp;
701 int matchcnt = 0;
702
703 matchfs = NULL;
704 if (matchmnt != NULL)
705 matchfs = VFSTOUFS(matchmnt)->um_fs;
706 ACQUIRE_LOCK(&lk);
707 /*
708 * Normally we just process each item on the worklist in order.
709 * However, if we are in a situation where we cannot lock any
710 * inodes, we have to skip over any dirrem requests whose
711 * vnodes are resident and locked.
712 */
713 LIST_FOREACH(wk, &softdep_workitem_pending, wk_list) {
714 if ((flags & LK_NOWAIT) == 0 || wk->wk_type != D_DIRREM)
715 break;
716 dirrem = WK_DIRREM(wk);
717 vp = ufs_ihashlookup(VFSTOUFS(dirrem->dm_mnt)->um_dev,
718 dirrem->dm_oldinum);
a11aaa81 719 if (vp == NULL || !vn_islocked(vp))
984263bc
MD
720 break;
721 }
722 if (wk == 0) {
723 FREE_LOCK(&lk);
724 return (0);
725 }
726 WORKLIST_REMOVE(wk);
727 num_on_worklist -= 1;
728 FREE_LOCK(&lk);
729 switch (wk->wk_type) {
730
731 case D_DIRREM:
732 /* removal of a directory entry */
733 if (WK_DIRREM(wk)->dm_mnt == matchmnt)
734 matchcnt += 1;
735 handle_workitem_remove(WK_DIRREM(wk));
736 break;
737
738 case D_FREEBLKS:
739 /* releasing blocks and/or fragments from a file */
740 if (WK_FREEBLKS(wk)->fb_fs == matchfs)
741 matchcnt += 1;
742 handle_workitem_freeblocks(WK_FREEBLKS(wk));
743 break;
744
745 case D_FREEFRAG:
746 /* releasing a fragment when replaced as a file grows */
747 if (WK_FREEFRAG(wk)->ff_fs == matchfs)
748 matchcnt += 1;
749 handle_workitem_freefrag(WK_FREEFRAG(wk));
750 break;
751
752 case D_FREEFILE:
753 /* releasing an inode when its link count drops to 0 */
754 if (WK_FREEFILE(wk)->fx_fs == matchfs)
755 matchcnt += 1;
756 handle_workitem_freefile(WK_FREEFILE(wk));
757 break;
758
759 default:
760 panic("%s_process_worklist: Unknown type %s",
761 "softdep", TYPENAME(wk->wk_type));
762 /* NOTREACHED */
763 }
764 return (matchcnt);
765}
766
767/*
768 * Move dependencies from one buffer to another.
59a647b1
MD
769 *
770 * bioops callback - hold io_token
984263bc
MD
771 */
772static void
3fcb1ab8 773softdep_move_dependencies(struct buf *oldbp, struct buf *newbp)
984263bc
MD
774{
775 struct worklist *wk, *wktail;
776
59a647b1 777 get_mplock();
984263bc
MD
778 if (LIST_FIRST(&newbp->b_dep) != NULL)
779 panic("softdep_move_dependencies: need merge code");
408357d8 780 wktail = NULL;
984263bc
MD
781 ACQUIRE_LOCK(&lk);
782 while ((wk = LIST_FIRST(&oldbp->b_dep)) != NULL) {
783 LIST_REMOVE(wk, wk_list);
408357d8 784 if (wktail == NULL)
984263bc
MD
785 LIST_INSERT_HEAD(&newbp->b_dep, wk, wk_list);
786 else
787 LIST_INSERT_AFTER(wktail, wk, wk_list);
788 wktail = wk;
408357d8 789 newbp->b_ops = &softdep_bioops;
984263bc
MD
790 }
791 FREE_LOCK(&lk);
59a647b1 792 rel_mplock();
984263bc
MD
793}
794
795/*
796 * Purge the work list of all items associated with a particular mount point.
797 */
798int
2aa32050 799softdep_flushfiles(struct mount *oldmnt, int flags)
984263bc
MD
800{
801 struct vnode *devvp;
802 int error, loopcnt;
803
804 /*
805 * Await our turn to clear out the queue, then serialize access.
806 */
807 while (softdep_worklist_busy != 0) {
808 softdep_worklist_req += 1;
377d4740 809 tsleep(&softdep_worklist_req, 0, "softflush", 0);
984263bc
MD
810 softdep_worklist_req -= 1;
811 }
812 softdep_worklist_busy = -1;
813
2aa32050 814 if ((error = ffs_flushfiles(oldmnt, flags)) != 0) {
984263bc
MD
815 softdep_worklist_busy = 0;
816 if (softdep_worklist_req)
817 wakeup(&softdep_worklist_req);
818 return (error);
819 }
820 /*
821 * Alternately flush the block device associated with the mount
822 * point and process any dependencies that the flushing
823 * creates. In theory, this loop can happen at most twice,
824 * but we give it a few extra just to be sure.
825 */
826 devvp = VFSTOUFS(oldmnt)->um_devvp;
827 for (loopcnt = 10; loopcnt > 0; ) {
828 if (softdep_process_worklist(oldmnt) == 0) {
829 loopcnt--;
830 /*
831 * Do another flush in case any vnodes were brought in
832 * as part of the cleanup operations.
833 */
2aa32050 834 if ((error = ffs_flushfiles(oldmnt, flags)) != 0)
984263bc
MD
835 break;
836 /*
837 * If we still found nothing to do, we are really done.
838 */
839 if (softdep_process_worklist(oldmnt) == 0)
840 break;
841 }
ca466bae 842 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
52174f71 843 error = VOP_FSYNC(devvp, MNT_WAIT, 0);
a11aaa81 844 vn_unlock(devvp);
984263bc
MD
845 if (error)
846 break;
847 }
848 softdep_worklist_busy = 0;
849 if (softdep_worklist_req)
850 wakeup(&softdep_worklist_req);
851
852 /*
853 * If we are unmounting then it is an error to fail. If we
854 * are simply trying to downgrade to read-only, then filesystem
855 * activity can keep us busy forever, so we just fail with EBUSY.
856 */
857 if (loopcnt == 0) {
858 if (oldmnt->mnt_kern_flag & MNTK_UNMOUNT)
859 panic("softdep_flushfiles: looping");
860 error = EBUSY;
861 }
862 return (error);
863}
864
865/*
866 * Structure hashing.
867 *
868 * There are three types of structures that can be looked up:
869 * 1) pagedep structures identified by mount point, inode number,
870 * and logical block.
871 * 2) inodedep structures identified by mount point and inode number.
872 * 3) newblk structures identified by mount point and
873 * physical block number.
874 *
875 * The "pagedep" and "inodedep" dependency structures are hashed
876 * separately from the file blocks and inodes to which they correspond.
877 * This separation helps when the in-memory copy of an inode or
878 * file block must be replaced. It also obviates the need to access
879 * an inode or file page when simply updating (or de-allocating)
880 * dependency structures. Lookup of newblk structures is needed to
881 * find newly allocated blocks when trying to associate them with
882 * their allocdirect or allocindir structure.
883 *
884 * The lookup routines optionally create and hash a new instance when
885 * an existing entry is not found.
886 */
887#define DEPALLOC 0x0001 /* allocate structure if lookup fails */
888#define NODELAY 0x0002 /* cannot do background work */
889
890/*
891 * Structures and routines associated with pagedep caching.
892 */
893LIST_HEAD(pagedep_hashhead, pagedep) *pagedep_hashtbl;
894u_long pagedep_hash; /* size of hash table - 1 */
895#define PAGEDEP_HASH(mp, inum, lbn) \
896 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
897 pagedep_hash])
898static struct sema pagedep_in_progress;
899
900/*
98a74972
MD
901 * Helper routine for pagedep_lookup()
902 */
903static __inline
904struct pagedep *
905pagedep_find(struct pagedep_hashhead *pagedephd, ino_t ino, ufs_lbn_t lbn,
906 struct mount *mp)
907{
908 struct pagedep *pagedep;
909
910 LIST_FOREACH(pagedep, pagedephd, pd_hash) {
911 if (ino == pagedep->pd_ino &&
912 lbn == pagedep->pd_lbn &&
913 mp == pagedep->pd_mnt) {
914 return (pagedep);
915 }
916 }
917 return(NULL);
918}
919
920/*
984263bc
MD
921 * Look up a pagedep. Return 1 if found, 0 if not found.
922 * If not found, allocate if DEPALLOC flag is passed.
923 * Found or allocated entry is returned in pagedeppp.
924 * This routine must be called with splbio interrupts blocked.
925 */
926static int
3fcb1ab8
SW
927pagedep_lookup(struct inode *ip, ufs_lbn_t lbn, int flags,
928 struct pagedep **pagedeppp)
984263bc
MD
929{
930 struct pagedep *pagedep;
931 struct pagedep_hashhead *pagedephd;
932 struct mount *mp;
933 int i;
934
935#ifdef DEBUG
dadab5e9 936 if (lk.lkt_held == NOHOLDER)
984263bc
MD
937 panic("pagedep_lookup: lock not held");
938#endif
939 mp = ITOV(ip)->v_mount;
940 pagedephd = PAGEDEP_HASH(mp, ip->i_number, lbn);
941top:
98a74972
MD
942 *pagedeppp = pagedep_find(pagedephd, ip->i_number, lbn, mp);
943 if (*pagedeppp)
944 return(1);
945 if ((flags & DEPALLOC) == 0)
984263bc 946 return (0);
984263bc
MD
947 if (sema_get(&pagedep_in_progress, &lk) == 0) {
948 ACQUIRE_LOCK(&lk);
949 goto top;
950 }
951 MALLOC(pagedep, struct pagedep *, sizeof(struct pagedep), M_PAGEDEP,
98a74972
MD
952 M_SOFTDEP_FLAGS | M_ZERO);
953
954 if (pagedep_find(pagedephd, ip->i_number, lbn, mp)) {
086c1d7e 955 kprintf("pagedep_lookup: blocking race avoided\n");
98a74972
MD
956 ACQUIRE_LOCK(&lk);
957 sema_release(&pagedep_in_progress);
efda3bd0 958 kfree(pagedep, M_PAGEDEP);
98a74972
MD
959 goto top;
960 }
961
984263bc
MD
962 pagedep->pd_list.wk_type = D_PAGEDEP;
963 pagedep->pd_mnt = mp;
964 pagedep->pd_ino = ip->i_number;
965 pagedep->pd_lbn = lbn;
966 LIST_INIT(&pagedep->pd_dirremhd);
967 LIST_INIT(&pagedep->pd_pendinghd);
968 for (i = 0; i < DAHASHSZ; i++)
969 LIST_INIT(&pagedep->pd_diraddhd[i]);
970 ACQUIRE_LOCK(&lk);
971 LIST_INSERT_HEAD(pagedephd, pagedep, pd_hash);
972 sema_release(&pagedep_in_progress);
973 *pagedeppp = pagedep;
974 return (0);
975}
976
977/*
978 * Structures and routines associated with inodedep caching.
979 */
980LIST_HEAD(inodedep_hashhead, inodedep) *inodedep_hashtbl;
981static u_long inodedep_hash; /* size of hash table - 1 */
982static long num_inodedep; /* number of inodedep allocated */
983#define INODEDEP_HASH(fs, inum) \
984 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
985static struct sema inodedep_in_progress;
986
987/*
98a74972
MD
988 * Helper routine for inodedep_lookup()
989 */
990static __inline
991struct inodedep *
992inodedep_find(struct inodedep_hashhead *inodedephd, struct fs *fs, ino_t inum)
993{
994 struct inodedep *inodedep;
995
996 LIST_FOREACH(inodedep, inodedephd, id_hash) {
997 if (inum == inodedep->id_ino && fs == inodedep->id_fs)
998 return(inodedep);
999 }
1000 return (NULL);
1001}
1002
1003/*
984263bc
MD
1004 * Look up a inodedep. Return 1 if found, 0 if not found.
1005 * If not found, allocate if DEPALLOC flag is passed.
1006 * Found or allocated entry is returned in inodedeppp.
1007 * This routine must be called with splbio interrupts blocked.
1008 */
1009static int
3fcb1ab8
SW
1010inodedep_lookup(struct fs *fs, ino_t inum, int flags,
1011 struct inodedep **inodedeppp)
984263bc
MD
1012{
1013 struct inodedep *inodedep;
1014 struct inodedep_hashhead *inodedephd;
1015 int firsttry;
1016
1017#ifdef DEBUG
dadab5e9 1018 if (lk.lkt_held == NOHOLDER)
984263bc
MD
1019 panic("inodedep_lookup: lock not held");
1020#endif
1021 firsttry = 1;
1022 inodedephd = INODEDEP_HASH(fs, inum);
1023top:
98a74972
MD
1024 *inodedeppp = inodedep_find(inodedephd, fs, inum);
1025 if (*inodedeppp)
984263bc 1026 return (1);
98a74972 1027 if ((flags & DEPALLOC) == 0)
984263bc 1028 return (0);
984263bc
MD
1029 /*
1030 * If we are over our limit, try to improve the situation.
1031 */
1032 if (num_inodedep > max_softdeps && firsttry &&
1033 speedup_syncer() == 0 && (flags & NODELAY) == 0 &&
1034 request_cleanup(FLUSH_INODES, 1)) {
1035 firsttry = 0;
1036 goto top;
1037 }
1038 if (sema_get(&inodedep_in_progress, &lk) == 0) {
1039 ACQUIRE_LOCK(&lk);
1040 goto top;
1041 }
984263bc 1042 MALLOC(inodedep, struct inodedep *, sizeof(struct inodedep),
98a74972
MD
1043 M_INODEDEP, M_SOFTDEP_FLAGS | M_ZERO);
1044 if (inodedep_find(inodedephd, fs, inum)) {
086c1d7e 1045 kprintf("inodedep_lookup: blocking race avoided\n");
98a74972
MD
1046 ACQUIRE_LOCK(&lk);
1047 sema_release(&inodedep_in_progress);
efda3bd0 1048 kfree(inodedep, M_INODEDEP);
98a74972
MD
1049 goto top;
1050 }
984263bc
MD
1051 inodedep->id_list.wk_type = D_INODEDEP;
1052 inodedep->id_fs = fs;
1053 inodedep->id_ino = inum;
1054 inodedep->id_state = ALLCOMPLETE;
1055 inodedep->id_nlinkdelta = 0;
1056 inodedep->id_savedino = NULL;
1057 inodedep->id_savedsize = -1;
1058 inodedep->id_buf = NULL;
1059 LIST_INIT(&inodedep->id_pendinghd);
1060 LIST_INIT(&inodedep->id_inowait);
1061 LIST_INIT(&inodedep->id_bufwait);
1062 TAILQ_INIT(&inodedep->id_inoupdt);
1063 TAILQ_INIT(&inodedep->id_newinoupdt);
1064 ACQUIRE_LOCK(&lk);
98a74972 1065 num_inodedep += 1;
984263bc
MD
1066 LIST_INSERT_HEAD(inodedephd, inodedep, id_hash);
1067 sema_release(&inodedep_in_progress);
1068 *inodedeppp = inodedep;
1069 return (0);
1070}
1071
1072/*
1073 * Structures and routines associated with newblk caching.
1074 */
1075LIST_HEAD(newblk_hashhead, newblk) *newblk_hashtbl;
1076u_long newblk_hash; /* size of hash table - 1 */
1077#define NEWBLK_HASH(fs, inum) \
1078 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
1079static struct sema newblk_in_progress;
1080
1081/*
98a74972
MD
1082 * Helper routine for newblk_lookup()
1083 */
1084static __inline
1085struct newblk *
1086newblk_find(struct newblk_hashhead *newblkhd, struct fs *fs,
1087 ufs_daddr_t newblkno)
1088{
1089 struct newblk *newblk;
1090
1091 LIST_FOREACH(newblk, newblkhd, nb_hash) {
1092 if (newblkno == newblk->nb_newblkno && fs == newblk->nb_fs)
1093 return (newblk);
1094 }
1095 return(NULL);
1096}
1097
1098/*
984263bc
MD
1099 * Look up a newblk. Return 1 if found, 0 if not found.
1100 * If not found, allocate if DEPALLOC flag is passed.
1101 * Found or allocated entry is returned in newblkpp.
1102 */
1103static int
3fcb1ab8
SW
1104newblk_lookup(struct fs *fs, ufs_daddr_t newblkno, int flags,
1105 struct newblk **newblkpp)
984263bc
MD
1106{
1107 struct newblk *newblk;
1108 struct newblk_hashhead *newblkhd;
1109
1110 newblkhd = NEWBLK_HASH(fs, newblkno);
1111top:
98a74972
MD
1112 *newblkpp = newblk_find(newblkhd, fs, newblkno);
1113 if (*newblkpp)
1114 return(1);
1115 if ((flags & DEPALLOC) == 0)
984263bc 1116 return (0);
984263bc
MD
1117 if (sema_get(&newblk_in_progress, 0) == 0)
1118 goto top;
1119 MALLOC(newblk, struct newblk *, sizeof(struct newblk),
98a74972
MD
1120 M_NEWBLK, M_SOFTDEP_FLAGS | M_ZERO);
1121
1122 if (newblk_find(newblkhd, fs, newblkno)) {
086c1d7e 1123 kprintf("newblk_lookup: blocking race avoided\n");
98a74972 1124 sema_release(&pagedep_in_progress);
efda3bd0 1125 kfree(newblk, M_NEWBLK);
98a74972
MD
1126 goto top;
1127 }
984263bc
MD
1128 newblk->nb_state = 0;
1129 newblk->nb_fs = fs;
1130 newblk->nb_newblkno = newblkno;
1131 LIST_INSERT_HEAD(newblkhd, newblk, nb_hash);
1132 sema_release(&newblk_in_progress);
1133 *newblkpp = newblk;
1134 return (0);
1135}
1136
1137/*
1138 * Executed during filesystem system initialization before
f719c866 1139 * mounting any filesystems.
984263bc
MD
1140 */
1141void
3fcb1ab8 1142softdep_initialize(void)
984263bc 1143{
47979091 1144 callout_init(&handle);
984263bc
MD
1145
1146 LIST_INIT(&mkdirlisthd);
1147 LIST_INIT(&softdep_workitem_pending);
1148 max_softdeps = min(desiredvnodes * 8,
1149 M_INODEDEP->ks_limit / (2 * sizeof(struct inodedep)));
1150 pagedep_hashtbl = hashinit(desiredvnodes / 5, M_PAGEDEP,
1151 &pagedep_hash);
377d4740 1152 sema_init(&pagedep_in_progress, "pagedep", 0, 0);
984263bc 1153 inodedep_hashtbl = hashinit(desiredvnodes, M_INODEDEP, &inodedep_hash);
377d4740 1154 sema_init(&inodedep_in_progress, "inodedep", 0, 0);
984263bc 1155 newblk_hashtbl = hashinit(64, M_NEWBLK, &newblk_hash);
377d4740 1156 sema_init(&newblk_in_progress, "newblk", 0, 0);
408357d8 1157 add_bio_ops(&softdep_bioops);
984263bc
MD
1158}
1159
1160/*
1161 * Called at mount time to notify the dependency code that a
1162 * filesystem wishes to use it.
1163 */
1164int
3fcb1ab8 1165softdep_mount(struct vnode *devvp, struct mount *mp, struct fs *fs)
984263bc
MD
1166{
1167 struct csum cstotal;
1168 struct cg *cgp;
1169 struct buf *bp;
1170 int error, cyl;
1171
1172 mp->mnt_flag &= ~MNT_ASYNC;
1173 mp->mnt_flag |= MNT_SOFTDEP;
408357d8 1174 mp->mnt_bioops = &softdep_bioops;
984263bc
MD
1175 /*
1176 * When doing soft updates, the counters in the
1177 * superblock may have gotten out of sync, so we have
1178 * to scan the cylinder groups and recalculate them.
1179 */
1180 if (fs->fs_clean != 0)
1181 return (0);
1182 bzero(&cstotal, sizeof cstotal);
1183 for (cyl = 0; cyl < fs->fs_ncg; cyl++) {
54078292
MD
1184 if ((error = bread(devvp, fsbtodoff(fs, cgtod(fs, cyl)),
1185 fs->fs_cgsize, &bp)) != 0) {
984263bc
MD
1186 brelse(bp);
1187 return (error);
1188 }
1189 cgp = (struct cg *)bp->b_data;
1190 cstotal.cs_nffree += cgp->cg_cs.cs_nffree;
1191 cstotal.cs_nbfree += cgp->cg_cs.cs_nbfree;
1192 cstotal.cs_nifree += cgp->cg_cs.cs_nifree;
1193 cstotal.cs_ndir += cgp->cg_cs.cs_ndir;
1194 fs->fs_cs(fs, cyl) = cgp->cg_cs;
1195 brelse(bp);
1196 }
1197#ifdef DEBUG
1198 if (bcmp(&cstotal, &fs->fs_cstotal, sizeof cstotal))
086c1d7e 1199 kprintf("ffs_mountfs: superblock updated for soft updates\n");
984263bc
MD
1200#endif
1201 bcopy(&cstotal, &fs->fs_cstotal, sizeof cstotal);
1202 return (0);
1203}
1204
1205/*
1206 * Protecting the freemaps (or bitmaps).
1207 *
f719c866 1208 * To eliminate the need to execute fsck before mounting a filesystem
984263bc
MD
1209 * after a power failure, one must (conservatively) guarantee that the
1210 * on-disk copy of the bitmaps never indicate that a live inode or block is
1211 * free. So, when a block or inode is allocated, the bitmap should be
1212 * updated (on disk) before any new pointers. When a block or inode is
1213 * freed, the bitmap should not be updated until all pointers have been
1214 * reset. The latter dependency is handled by the delayed de-allocation
1215 * approach described below for block and inode de-allocation. The former
1216 * dependency is handled by calling the following procedure when a block or
1217 * inode is allocated. When an inode is allocated an "inodedep" is created
1218 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1219 * Each "inodedep" is also inserted into the hash indexing structure so
1220 * that any additional link additions can be made dependent on the inode
1221 * allocation.
1222 *
f719c866 1223 * The ufs filesystem maintains a number of free block counts (e.g., per
984263bc
MD
1224 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1225 * in addition to the bitmaps. These counts are used to improve efficiency
1226 * during allocation and therefore must be consistent with the bitmaps.
1227 * There is no convenient way to guarantee post-crash consistency of these
1228 * counts with simple update ordering, for two main reasons: (1) The counts
1229 * and bitmaps for a single cylinder group block are not in the same disk
1230 * sector. If a disk write is interrupted (e.g., by power failure), one may
1231 * be written and the other not. (2) Some of the counts are located in the
1232 * superblock rather than the cylinder group block. So, we focus our soft
1233 * updates implementation on protecting the bitmaps. When mounting a
1234 * filesystem, we recompute the auxiliary counts from the bitmaps.
1235 */
1236
1237/*
1238 * Called just after updating the cylinder group block to allocate an inode.
3fcb1ab8
SW
1239 *
1240 * Parameters:
1241 * bp: buffer for cylgroup block with inode map
1242 * ip: inode related to allocation
1243 * newinum: new inode number being allocated
984263bc
MD
1244 */
1245void
3fcb1ab8 1246softdep_setup_inomapdep(struct buf *bp, struct inode *ip, ino_t newinum)
984263bc
MD
1247{
1248 struct inodedep *inodedep;
1249 struct bmsafemap *bmsafemap;
1250
1251 /*
1252 * Create a dependency for the newly allocated inode.
1253 * Panic if it already exists as something is seriously wrong.
1254 * Otherwise add it to the dependency list for the buffer holding
1255 * the cylinder group map from which it was allocated.
1256 */
1257 ACQUIRE_LOCK(&lk);
1258 if ((inodedep_lookup(ip->i_fs, newinum, DEPALLOC|NODELAY, &inodedep))) {
1259 FREE_LOCK(&lk);
1260 panic("softdep_setup_inomapdep: found inode");
1261 }
1262 inodedep->id_buf = bp;
1263 inodedep->id_state &= ~DEPCOMPLETE;
1264 bmsafemap = bmsafemap_lookup(bp);
1265 LIST_INSERT_HEAD(&bmsafemap->sm_inodedephd, inodedep, id_deps);
1266 FREE_LOCK(&lk);
1267}
1268
1269/*
1270 * Called just after updating the cylinder group block to
1271 * allocate block or fragment.
3fcb1ab8
SW
1272 *
1273 * Parameters:
1274 * bp: buffer for cylgroup block with block map
1275 * fs: filesystem doing allocation
1276 * newblkno: number of newly allocated block
984263bc
MD
1277 */
1278void
3fcb1ab8
SW
1279softdep_setup_blkmapdep(struct buf *bp, struct fs *fs,
1280 ufs_daddr_t newblkno)
984263bc
MD
1281{
1282 struct newblk *newblk;
1283 struct bmsafemap *bmsafemap;
1284
1285 /*
1286 * Create a dependency for the newly allocated block.
1287 * Add it to the dependency list for the buffer holding
1288 * the cylinder group map from which it was allocated.
1289 */
1290 if (newblk_lookup(fs, newblkno, DEPALLOC, &newblk) != 0)
1291 panic("softdep_setup_blkmapdep: found block");
1292 ACQUIRE_LOCK(&lk);
1293 newblk->nb_bmsafemap = bmsafemap = bmsafemap_lookup(bp);
1294 LIST_INSERT_HEAD(&bmsafemap->sm_newblkhd, newblk, nb_deps);
1295 FREE_LOCK(&lk);
1296}
1297
1298/*
1299 * Find the bmsafemap associated with a cylinder group buffer.
1300 * If none exists, create one. The buffer must be locked when
1301 * this routine is called and this routine must be called with
1302 * splbio interrupts blocked.
1303 */
1304static struct bmsafemap *
3fcb1ab8 1305bmsafemap_lookup(struct buf *bp)
984263bc
MD
1306{
1307 struct bmsafemap *bmsafemap;
1308 struct worklist *wk;
1309
1310#ifdef DEBUG
dadab5e9 1311 if (lk.lkt_held == NOHOLDER)
984263bc
MD
1312 panic("bmsafemap_lookup: lock not held");
1313#endif
408357d8 1314 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
984263bc
MD
1315 if (wk->wk_type == D_BMSAFEMAP)
1316 return (WK_BMSAFEMAP(wk));
408357d8 1317 }
984263bc
MD
1318 FREE_LOCK(&lk);
1319 MALLOC(bmsafemap, struct bmsafemap *, sizeof(struct bmsafemap),
1320 M_BMSAFEMAP, M_SOFTDEP_FLAGS);
1321 bmsafemap->sm_list.wk_type = D_BMSAFEMAP;
1322 bmsafemap->sm_list.wk_state = 0;
1323 bmsafemap->sm_buf = bp;
1324 LIST_INIT(&bmsafemap->sm_allocdirecthd);
1325 LIST_INIT(&bmsafemap->sm_allocindirhd);
1326 LIST_INIT(&bmsafemap->sm_inodedephd);
1327 LIST_INIT(&bmsafemap->sm_newblkhd);
1328 ACQUIRE_LOCK(&lk);
408357d8 1329 WORKLIST_INSERT_BP(bp, &bmsafemap->sm_list);
984263bc
MD
1330 return (bmsafemap);
1331}
1332
1333/*
1334 * Direct block allocation dependencies.
1335 *
1336 * When a new block is allocated, the corresponding disk locations must be
1337 * initialized (with zeros or new data) before the on-disk inode points to
1338 * them. Also, the freemap from which the block was allocated must be
1339 * updated (on disk) before the inode's pointer. These two dependencies are
1340 * independent of each other and are needed for all file blocks and indirect
1341 * blocks that are pointed to directly by the inode. Just before the
1342 * "in-core" version of the inode is updated with a newly allocated block
1343 * number, a procedure (below) is called to setup allocation dependency
1344 * structures. These structures are removed when the corresponding
1345 * dependencies are satisfied or when the block allocation becomes obsolete
1346 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1347 * fragment that gets upgraded). All of these cases are handled in
1348 * procedures described later.
1349 *
1350 * When a file extension causes a fragment to be upgraded, either to a larger
1351 * fragment or to a full block, the on-disk location may change (if the
1352 * previous fragment could not simply be extended). In this case, the old
1353 * fragment must be de-allocated, but not until after the inode's pointer has
1354 * been updated. In most cases, this is handled by later procedures, which
1355 * will construct a "freefrag" structure to be added to the workitem queue
1356 * when the inode update is complete (or obsolete). The main exception to
1357 * this is when an allocation occurs while a pending allocation dependency
1358 * (for the same block pointer) remains. This case is handled in the main
1359 * allocation dependency setup procedure by immediately freeing the
1360 * unreferenced fragments.
3fcb1ab8
SW
1361 *
1362 * Parameters:
1363 * ip: inode to which block is being added
1364 * lbn: block pointer within inode
1365 * newblkno: disk block number being added
1366 * oldblkno: previous block number, 0 unless frag
1367 * newsize: size of new block
1368 * oldsize: size of new block
1369 * bp: bp for allocated block
984263bc
MD
1370 */
1371void
3fcb1ab8
SW
1372softdep_setup_allocdirect(struct inode *ip, ufs_lbn_t lbn, ufs_daddr_t newblkno,
1373 ufs_daddr_t oldblkno, long newsize, long oldsize,
1374 struct buf *bp)
984263bc
MD
1375{
1376 struct allocdirect *adp, *oldadp;
1377 struct allocdirectlst *adphead;
1378 struct bmsafemap *bmsafemap;
1379 struct inodedep *inodedep;
1380 struct pagedep *pagedep;
1381 struct newblk *newblk;
1382
1383 MALLOC(adp, struct allocdirect *, sizeof(struct allocdirect),
e7b4468c 1384 M_ALLOCDIRECT, M_SOFTDEP_FLAGS | M_ZERO);
984263bc
MD
1385 adp->ad_list.wk_type = D_ALLOCDIRECT;
1386 adp->ad_lbn = lbn;
1387 adp->ad_newblkno = newblkno;
1388 adp->ad_oldblkno = oldblkno;
1389 adp->ad_newsize = newsize;
1390 adp->ad_oldsize = oldsize;
1391 adp->ad_state = ATTACHED;
1392 if (newblkno == oldblkno)
1393 adp->ad_freefrag = NULL;
1394 else
1395 adp->ad_freefrag = newfreefrag(ip, oldblkno, oldsize);
1396
1397 if (newblk_lookup(ip->i_fs, newblkno, 0, &newblk) == 0)
1398 panic("softdep_setup_allocdirect: lost block");
1399
1400 ACQUIRE_LOCK(&lk);
1401 inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC | NODELAY, &inodedep);
1402 adp->ad_inodedep = inodedep;
1403
1404 if (newblk->nb_state == DEPCOMPLETE) {
1405 adp->ad_state |= DEPCOMPLETE;
1406 adp->ad_buf = NULL;
1407 } else {
1408 bmsafemap = newblk->nb_bmsafemap;
1409 adp->ad_buf = bmsafemap->sm_buf;
1410 LIST_REMOVE(newblk, nb_deps);
1411 LIST_INSERT_HEAD(&bmsafemap->sm_allocdirecthd, adp, ad_deps);
1412 }
1413 LIST_REMOVE(newblk, nb_hash);
1414 FREE(newblk, M_NEWBLK);
1415
408357d8 1416 WORKLIST_INSERT_BP(bp, &adp->ad_list);
984263bc
MD
1417 if (lbn >= NDADDR) {
1418 /* allocating an indirect block */
1419 if (oldblkno != 0) {
1420 FREE_LOCK(&lk);
1421 panic("softdep_setup_allocdirect: non-zero indir");
1422 }
1423 } else {
1424 /*
1425 * Allocating a direct block.
1426 *
1427 * If we are allocating a directory block, then we must
1428 * allocate an associated pagedep to track additions and
1429 * deletions.
1430 */
1431 if ((ip->i_mode & IFMT) == IFDIR &&
408357d8
MD
1432 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0) {
1433 WORKLIST_INSERT_BP(bp, &pagedep->pd_list);
1434 }
984263bc
MD
1435 }
1436 /*
1437 * The list of allocdirects must be kept in sorted and ascending
1438 * order so that the rollback routines can quickly determine the
1439 * first uncommitted block (the size of the file stored on disk
1440 * ends at the end of the lowest committed fragment, or if there
1441 * are no fragments, at the end of the highest committed block).
1442 * Since files generally grow, the typical case is that the new
1443 * block is to be added at the end of the list. We speed this
1444 * special case by checking against the last allocdirect in the
1445 * list before laboriously traversing the list looking for the
1446 * insertion point.
1447 */
1448 adphead = &inodedep->id_newinoupdt;
1449 oldadp = TAILQ_LAST(adphead, allocdirectlst);
1450 if (oldadp == NULL || oldadp->ad_lbn <= lbn) {
1451 /* insert at end of list */
1452 TAILQ_INSERT_TAIL(adphead, adp, ad_next);
1453 if (oldadp != NULL && oldadp->ad_lbn == lbn)
1454 allocdirect_merge(adphead, adp, oldadp);
1455 FREE_LOCK(&lk);
1456 return;
1457 }
1458 TAILQ_FOREACH(oldadp, adphead, ad_next) {
1459 if (oldadp->ad_lbn >= lbn)
1460 break;
1461 }
1462 if (oldadp == NULL) {
1463 FREE_LOCK(&lk);
1464 panic("softdep_setup_allocdirect: lost entry");
1465 }
1466 /* insert in middle of list */
1467 TAILQ_INSERT_BEFORE(oldadp, adp, ad_next);
1468 if (oldadp->ad_lbn == lbn)
1469 allocdirect_merge(adphead, adp, oldadp);
1470 FREE_LOCK(&lk);
1471}
1472
1473/*
1474 * Replace an old allocdirect dependency with a newer one.
1475 * This routine must be called with splbio interrupts blocked.
3fcb1ab8
SW
1476 *
1477 * Parameters:
1478 * adphead: head of list holding allocdirects
1479 * newadp: allocdirect being added
1480 * oldadp: existing allocdirect being checked
984263bc
MD
1481 */
1482static void
3fcb1ab8
SW
1483allocdirect_merge(struct allocdirectlst *adphead,
1484 struct allocdirect *newadp,
1485 struct allocdirect *oldadp)
984263bc
MD
1486{
1487 struct freefrag *freefrag;
1488
1489#ifdef DEBUG
dadab5e9 1490 if (lk.lkt_held == NOHOLDER)
984263bc
MD
1491 panic("allocdirect_merge: lock not held");
1492#endif
1493 if (newadp->ad_oldblkno != oldadp->ad_newblkno ||
1494 newadp->ad_oldsize != oldadp->ad_newsize ||
1495 newadp->ad_lbn >= NDADDR) {
1496 FREE_LOCK(&lk);
1497 panic("allocdirect_check: old %d != new %d || lbn %ld >= %d",
1498 newadp->ad_oldblkno, oldadp->ad_newblkno, newadp->ad_lbn,
1499 NDADDR);
1500 }
1501 newadp->ad_oldblkno = oldadp->ad_oldblkno;
1502 newadp->ad_oldsize = oldadp->ad_oldsize;
1503 /*
1504 * If the old dependency had a fragment to free or had never
1505 * previously had a block allocated, then the new dependency
1506 * can immediately post its freefrag and adopt the old freefrag.
1507 * This action is done by swapping the freefrag dependencies.
1508 * The new dependency gains the old one's freefrag, and the
1509 * old one gets the new one and then immediately puts it on
1510 * the worklist when it is freed by free_allocdirect. It is
1511 * not possible to do this swap when the old dependency had a
1512 * non-zero size but no previous fragment to free. This condition
1513 * arises when the new block is an extension of the old block.
1514 * Here, the first part of the fragment allocated to the new
1515 * dependency is part of the block currently claimed on disk by
1516 * the old dependency, so cannot legitimately be freed until the
1517 * conditions for the new dependency are fulfilled.
1518 */
1519 if (oldadp->ad_freefrag != NULL || oldadp->ad_oldblkno == 0) {
1520 freefrag = newadp->ad_freefrag;
1521 newadp->ad_freefrag = oldadp->ad_freefrag;
1522 oldadp->ad_freefrag = freefrag;
1523 }
1524 free_allocdirect(adphead, oldadp, 0);
1525}
1526
1527/*
1528 * Allocate a new freefrag structure if needed.
1529 */
1530static struct freefrag *
3fcb1ab8 1531newfreefrag(struct inode *ip, ufs_daddr_t blkno, long size)
984263bc
MD
1532{
1533 struct freefrag *freefrag;
1534 struct fs *fs;
1535
1536 if (blkno == 0)
1537 return (NULL);
1538 fs = ip->i_fs;
1539 if (fragnum(fs, blkno) + numfrags(fs, size) > fs->fs_frag)
1540 panic("newfreefrag: frag size");
1541 MALLOC(freefrag, struct freefrag *, sizeof(struct freefrag),
1542 M_FREEFRAG, M_SOFTDEP_FLAGS);
1543 freefrag->ff_list.wk_type = D_FREEFRAG;
1544 freefrag->ff_state = ip->i_uid & ~ONWORKLIST; /* XXX - used below */
1545 freefrag->ff_inum = ip->i_number;
1546 freefrag->ff_fs = fs;
1547 freefrag->ff_devvp = ip->i_devvp;
1548 freefrag->ff_blkno = blkno;
1549 freefrag->ff_fragsize = size;
1550 return (freefrag);
1551}
1552
1553/*
1554 * This workitem de-allocates fragments that were replaced during
1555 * file block allocation.
1556 */
1557static void
3fcb1ab8 1558handle_workitem_freefrag(struct freefrag *freefrag)
984263bc
MD
1559{
1560 struct inode tip;
1561
1562 tip.i_fs = freefrag->ff_fs;
1563 tip.i_devvp = freefrag->ff_devvp;
1564 tip.i_dev = freefrag->ff_devvp->v_rdev;
1565 tip.i_number = freefrag->ff_inum;
1566 tip.i_uid = freefrag->ff_state & ~ONWORKLIST; /* XXX - set above */
1567 ffs_blkfree(&tip, freefrag->ff_blkno, freefrag->ff_fragsize);
1568 FREE(freefrag, M_FREEFRAG);
1569}
1570
1571/*
1572 * Indirect block allocation dependencies.
1573 *
1574 * The same dependencies that exist for a direct block also exist when
1575 * a new block is allocated and pointed to by an entry in a block of
1576 * indirect pointers. The undo/redo states described above are also
1577 * used here. Because an indirect block contains many pointers that
1578 * may have dependencies, a second copy of the entire in-memory indirect
1579 * block is kept. The buffer cache copy is always completely up-to-date.
1580 * The second copy, which is used only as a source for disk writes,
1581 * contains only the safe pointers (i.e., those that have no remaining
1582 * update dependencies). The second copy is freed when all pointers
1583 * are safe. The cache is not allowed to replace indirect blocks with
1584 * pending update dependencies. If a buffer containing an indirect
1585 * block with dependencies is written, these routines will mark it
1586 * dirty again. It can only be successfully written once all the
1587 * dependencies are removed. The ffs_fsync routine in conjunction with
1588 * softdep_sync_metadata work together to get all the dependencies
1589 * removed so that a file can be successfully written to disk. Three
1590 * procedures are used when setting up indirect block pointer
1591 * dependencies. The division is necessary because of the organization
1592 * of the "balloc" routine and because of the distinction between file
1593 * pages and file metadata blocks.
1594 */
1595
1596/*
1597 * Allocate a new allocindir structure.
3fcb1ab8
SW
1598 *
1599 * Parameters:
1600 * ip: inode for file being extended
1601 * ptrno: offset of pointer in indirect block
1602 * newblkno: disk block number being added
1603 * oldblkno: previous block number, 0 if none
984263bc
MD
1604 */
1605static struct allocindir *
3fcb1ab8
SW
1606newallocindir(struct inode *ip, int ptrno, ufs_daddr_t newblkno,
1607 ufs_daddr_t oldblkno)
984263bc
MD
1608{
1609 struct allocindir *aip;
1610
1611 MALLOC(aip, struct allocindir *, sizeof(struct allocindir),
e7b4468c 1612 M_ALLOCINDIR, M_SOFTDEP_FLAGS | M_ZERO);
984263bc
MD
1613 aip->ai_list.wk_type = D_ALLOCINDIR;
1614 aip->ai_state = ATTACHED;
1615 aip->ai_offset = ptrno;
1616 aip->ai_newblkno = newblkno;
1617 aip->ai_oldblkno = oldblkno;
1618 aip->ai_freefrag = newfreefrag(ip, oldblkno, ip->i_fs->fs_bsize);
1619 return (aip);
1620}
1621
1622/*
1623 * Called just before setting an indirect block pointer
1624 * to a newly allocated file page.
3fcb1ab8
SW
1625 *
1626 * Parameters:
1627 * ip: inode for file being extended
1628 * lbn: allocated block number within file
1629 * bp: buffer with indirect blk referencing page
1630 * ptrno: offset of pointer in indirect block
1631 * newblkno: disk block number being added
1632 * oldblkno: previous block number, 0 if none
1633 * nbp: buffer holding allocated page
984263bc
MD
1634 */
1635void
3fcb1ab8
SW
1636softdep_setup_allocindir_page(struct inode *ip, ufs_lbn_t lbn,
1637 struct buf *bp, int ptrno,
1638 ufs_daddr_t newblkno, ufs_daddr_t oldblkno,
1639 struct buf *nbp)
984263bc
MD
1640{
1641 struct allocindir *aip;
1642 struct pagedep *pagedep;
1643
1644 aip = newallocindir(ip, ptrno, newblkno, oldblkno);
1645 ACQUIRE_LOCK(&lk);
1646 /*
1647 * If we are allocating a directory page, then we must
1648 * allocate an associated pagedep to track additions and
1649 * deletions.
1650 */
1651 if ((ip->i_mode & IFMT) == IFDIR &&
1652 pagedep_lookup(ip, lbn, DEPALLOC, &pagedep) == 0)
408357d8
MD
1653 WORKLIST_INSERT_BP(nbp, &pagedep->pd_list);
1654 WORKLIST_INSERT_BP(nbp, &aip->ai_list);
984263bc
MD
1655 FREE_LOCK(&lk);
1656 setup_allocindir_phase2(bp, ip, aip);
1657}
1658
1659/*
1660 * Called just before setting an indirect block pointer to a
1661 * newly allocated indirect block.
3fcb1ab8
SW
1662 * Parameters:
1663 * nbp: newly allocated indirect block
1664 * ip: inode for file being extended
1665 * bp: indirect block referencing allocated block
1666 * ptrno: offset of pointer in indirect block
1667 * newblkno: disk block number being added
984263bc
MD
1668 */
1669void
3fcb1ab8
SW
1670softdep_setup_allocindir_meta(struct buf *nbp, struct inode *ip,
1671 struct buf *bp, int ptrno,
1672 ufs_daddr_t newblkno)
984263bc
MD
1673{
1674 struct allocindir *aip;
1675
1676 aip = newallocindir(ip, ptrno, newblkno, 0);
1677 ACQUIRE_LOCK(&lk);
408357d8 1678 WORKLIST_INSERT_BP(nbp, &aip->ai_list);
984263bc
MD
1679 FREE_LOCK(&lk);
1680 setup_allocindir_phase2(bp, ip, aip);
1681}
1682
1683/*
1684 * Called to finish the allocation of the "aip" allocated
1685 * by one of the two routines above.
3fcb1ab8
SW
1686 *
1687 * Parameters:
1688 * bp: in-memory copy of the indirect block
1689 * ip: inode for file being extended
1690 * aip: allocindir allocated by the above routines
984263bc
MD
1691 */
1692static void
3fcb1ab8
SW
1693setup_allocindir_phase2(struct buf *bp, struct inode *ip,
1694 struct allocindir *aip)
984263bc
MD
1695{
1696 struct worklist *wk;
1697 struct indirdep *indirdep, *newindirdep;
1698 struct bmsafemap *bmsafemap;
1699 struct allocindir *oldaip;
1700 struct freefrag *freefrag;
1701 struct newblk *newblk;
1702
54078292 1703 if (bp->b_loffset >= 0)
984263bc
MD
1704 panic("setup_allocindir_phase2: not indir blk");
1705 for (indirdep = NULL, newindirdep = NULL; ; ) {
1706 ACQUIRE_LOCK(&lk);
1707 LIST_FOREACH(wk, &bp->b_dep, wk_list) {
1708 if (wk->wk_type != D_INDIRDEP)
1709 continue;
1710 indirdep = WK_INDIRDEP(wk);
1711 break;
1712 }
1713 if (indirdep == NULL && newindirdep) {
1714 indirdep = newindirdep;
408357d8 1715 WORKLIST_INSERT_BP(bp, &indirdep->ir_list);
984263bc
MD
1716 newindirdep = NULL;
1717 }
1718 FREE_LOCK(&lk);
1719 if (indirdep) {
1720 if (newblk_lookup(ip->i_fs, aip->ai_newblkno, 0,
1721 &newblk) == 0)
1722 panic("setup_allocindir: lost block");
1723 ACQUIRE_LOCK(&lk);
1724 if (newblk->nb_state == DEPCOMPLETE) {
1725 aip->ai_state |= DEPCOMPLETE;
1726 aip->ai_buf = NULL;
1727 } else {
1728 bmsafemap = newblk->nb_bmsafemap;
1729 aip->ai_buf = bmsafemap->sm_buf;
1730 LIST_REMOVE(newblk, nb_deps);
1731 LIST_INSERT_HEAD(&bmsafemap->sm_allocindirhd,
1732 aip, ai_deps);
1733 }
1734 LIST_REMOVE(newblk, nb_hash);
1735 FREE(newblk, M_NEWBLK);
1736 aip->ai_indirdep = indirdep;
1737 /*
1738 * Check to see if there is an existing dependency
1739 * for this block. If there is, merge the old
1740 * dependency into the new one.
1741 */
1742 if (aip->ai_oldblkno == 0)
1743 oldaip = NULL;
1744 else
1745
1746 LIST_FOREACH(oldaip, &indirdep->ir_deplisthd, ai_next)
1747 if (oldaip->ai_offset == aip->ai_offset)
1748 break;
1749 if (oldaip != NULL) {
1750 if (oldaip->ai_newblkno != aip->ai_oldblkno) {
1751 FREE_LOCK(&lk);
1752 panic("setup_allocindir_phase2: blkno");
1753 }
1754 aip->ai_oldblkno = oldaip->ai_oldblkno;
1755 freefrag = oldaip->ai_freefrag;
1756 oldaip->ai_freefrag = aip->ai_freefrag;
1757 aip->ai_freefrag = freefrag;
1758 free_allocindir(oldaip, NULL);
1759 }
1760 LIST_INSERT_HEAD(&indirdep->ir_deplisthd, aip, ai_next);
1761 ((ufs_daddr_t *)indirdep->ir_savebp->b_data)
1762 [aip->ai_offset] = aip->ai_oldblkno;
1763 FREE_LOCK(&lk);
1764 }
1765 if (newindirdep) {
7d618503 1766 /*
2ae68842
MD
1767 * Avoid any possibility of data corruption by
1768 * ensuring that our old version is thrown away.
7d618503 1769 */
2ae68842
MD
1770 newindirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
1771 brelse(newindirdep->ir_savebp);
984263bc
MD
1772 WORKITEM_FREE((caddr_t)newindirdep, D_INDIRDEP);
1773 }
1774 if (indirdep)
1775 break;
1776 MALLOC(newindirdep, struct indirdep *, sizeof(struct indirdep),
1777 M_INDIRDEP, M_SOFTDEP_FLAGS);
1778 newindirdep->ir_list.wk_type = D_INDIRDEP;
1779 newindirdep->ir_state = ATTACHED;
1780 LIST_INIT(&newindirdep->ir_deplisthd);
1781 LIST_INIT(&newindirdep->ir_donehd);
54078292
MD
1782 if (bp->b_bio2.bio_offset == NOOFFSET) {
1783 VOP_BMAP(bp->b_vp, bp->b_bio1.bio_offset,
e92ca23a
MD
1784 &bp->b_bio2.bio_offset, NULL, NULL,
1785 BUF_CMD_WRITE);
984263bc 1786 }
54078292
MD
1787 KKASSERT(bp->b_bio2.bio_offset != NOOFFSET);
1788 newindirdep->ir_savebp = getblk(ip->i_devvp,
1789 bp->b_bio2.bio_offset,
1790 bp->b_bcount, 0, 0);
984263bc
MD
1791 BUF_KERNPROC(newindirdep->ir_savebp);
1792 bcopy(bp->b_data, newindirdep->ir_savebp->b_data, bp->b_bcount);
1793 }
1794}
1795
1796/*
1797 * Block de-allocation dependencies.
1798 *
1799 * When blocks are de-allocated, the on-disk pointers must be nullified before
1800 * the blocks are made available for use by other files. (The true
1801 * requirement is that old pointers must be nullified before new on-disk
1802 * pointers are set. We chose this slightly more stringent requirement to
1803 * reduce complexity.) Our implementation handles this dependency by updating
1804 * the inode (or indirect block) appropriately but delaying the actual block
1805 * de-allocation (i.e., freemap and free space count manipulation) until
1806 * after the updated versions reach stable storage. After the disk is
1807 * updated, the blocks can be safely de-allocated whenever it is convenient.
1808 * This implementation handles only the common case of reducing a file's
1809 * length to zero. Other cases are handled by the conventional synchronous
1810 * write approach.
1811 *
1812 * The ffs implementation with which we worked double-checks
1813 * the state of the block pointers and file size as it reduces
1814 * a file's length. Some of this code is replicated here in our
1815 * soft updates implementation. The freeblks->fb_chkcnt field is
1816 * used to transfer a part of this information to the procedure
1817 * that eventually de-allocates the blocks.
1818 *
1819 * This routine should be called from the routine that shortens
1820 * a file's length, before the inode's size or block pointers
1821 * are modified. It will save the block pointer information for
1822 * later release and zero the inode so that the calling routine
1823 * can release it.
1824 */
6bae6177
MD
1825struct softdep_setup_freeblocks_info {
1826 struct fs *fs;
1827 struct inode *ip;
1828};
1829
1830static int softdep_setup_freeblocks_bp(struct buf *bp, void *data);
1831
3fcb1ab8
SW
1832/*
1833 * Parameters:
1834 * ip: The inode whose length is to be reduced
1835 * length: The new length for the file
1836 */
984263bc 1837void
3fcb1ab8 1838softdep_setup_freeblocks(struct inode *ip, off_t length)
984263bc 1839{
6bae6177 1840 struct softdep_setup_freeblocks_info info;
984263bc
MD
1841 struct freeblks *freeblks;
1842 struct inodedep *inodedep;
1843 struct allocdirect *adp;
1844 struct vnode *vp;
1845 struct buf *bp;
1846 struct fs *fs;
1847 int i, error, delay;
6bae6177 1848 int count;
984263bc
MD
1849
1850 fs = ip->i_fs;
1851 if (length != 0)
1852 panic("softde_setup_freeblocks: non-zero length");
1853 MALLOC(freeblks, struct freeblks *, sizeof(struct freeblks),
e7b4468c 1854 M_FREEBLKS, M_SOFTDEP_FLAGS | M_ZERO);
984263bc 1855 freeblks->fb_list.wk_type = D_FREEBLKS;
89a5de29 1856 freeblks->fb_state = ATTACHED;
984263bc
MD
1857 freeblks->fb_uid = ip->i_uid;
1858 freeblks->fb_previousinum = ip->i_number;
1859 freeblks->fb_devvp = ip->i_devvp;
1860 freeblks->fb_fs = fs;
1861 freeblks->fb_oldsize = ip->i_size;
1862 freeblks->fb_newsize = length;
1863 freeblks->fb_chkcnt = ip->i_blocks;
1864 for (i = 0; i < NDADDR; i++) {
1865 freeblks->fb_dblks[i] = ip->i_db[i];
1866 ip->i_db[i] = 0;
1867 }
1868 for (i = 0; i < NIADDR; i++) {
1869 freeblks->fb_iblks[i] = ip->i_ib[i];
1870 ip->i_ib[i] = 0;
1871 }
1872 ip->i_blocks = 0;
1873 ip->i_size = 0;
1874 /*
1875 * Push the zero'ed inode to to its disk buffer so that we are free
1876 * to delete its dependencies below. Once the dependencies are gone
1877 * the buffer can be safely released.
1878 */
1879 if ((error = bread(ip->i_devvp,
54078292 1880 fsbtodoff(fs, ino_to_fsba(fs, ip->i_number)),
3b568787 1881 (int)fs->fs_bsize, &bp)) != 0)
984263bc 1882 softdep_error("softdep_setup_freeblocks", error);
50e58362 1883 *((struct ufs1_dinode *)bp->b_data + ino_to_fsbo(fs, ip->i_number)) =
984263bc
MD
1884 ip->i_din;
1885 /*
1886 * Find and eliminate any inode dependencies.
1887 */
1888 ACQUIRE_LOCK(&lk);
1889 (void) inodedep_lookup(fs, ip->i_number, DEPALLOC, &inodedep);
1890 if ((inodedep->id_state & IOSTARTED) != 0) {
1891 FREE_LOCK(&lk);
1892 panic("softdep_setup_freeblocks: inode busy");
1893 }
1894 /*
1895 * Add the freeblks structure to the list of operations that
1896 * must await the zero'ed inode being written to disk. If we
1897 * still have a bitmap dependency (delay == 0), then the inode
1898 * has never been written to disk, so we can process the
1899 * freeblks below once we have deleted the dependencies.
1900 */
1901 delay = (inodedep->id_state & DEPCOMPLETE);
1902 if (delay)
1903 WORKLIST_INSERT(&inodedep->id_bufwait, &freeblks->fb_list);
1904 /*
1905 * Because the file length has been truncated to zero, any
1906 * pending block allocation dependency structures associated
1907 * with this inode are obsolete and can simply be de-allocated.
1908 * We must first merge the two dependency lists to get rid of
1909 * any duplicate freefrag structures, then purge the merged list.
1910 */
1911 merge_inode_lists(inodedep);
1912 while ((adp = TAILQ_FIRST(&inodedep->id_inoupdt)) != 0)
1913 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
1914 FREE_LOCK(&lk);
1915 bdwrite(bp);
1916 /*
1917 * We must wait for any I/O in progress to finish so that
1918 * all potential buffers on the dirty list will be visible.
1919 * Once they are all there, walk the list and get rid of
1920 * any dependencies.
1921 */
1922 vp = ITOV(ip);
1923 ACQUIRE_LOCK(&lk);
1924 drain_output(vp, 1);
6bae6177
MD
1925
1926 info.fs = fs;
1927 info.ip = ip;
3b998fa9 1928 lwkt_gettoken(&vp->v_token);
6bae6177
MD
1929 do {
1930 count = RB_SCAN(buf_rb_tree, &vp->v_rbdirty_tree, NULL,
1931 softdep_setup_freeblocks_bp, &info);
2ef16e04 1932 } while (count != 0);
3b998fa9 1933 lwkt_reltoken(&vp->v_token);
0202303b 1934
984263bc
MD
1935 if (inodedep_lookup(fs, ip->i_number, 0, &inodedep) != 0)
1936 (void)free_inodedep(inodedep);
89a5de29
MD
1937
1938 if (delay) {
1939 freeblks->fb_state |= DEPCOMPLETE;
1940 /*
1941 * If the inode with zeroed block pointers is now on disk
1942 * we can start freeing blocks. Add freeblks to the worklist
1943 * instead of calling handle_workitem_freeblocks directly as
1944 * it is more likely that additional IO is needed to complete
1945 * the request here than in the !delay case.
1946 */
1947 if ((freeblks->fb_state & ALLCOMPLETE) == ALLCOMPLETE)
1948 add_to_worklist(&freeblks->fb_list);
1949 }
1950
984263bc
MD
1951 FREE_LOCK(&lk);
1952 /*
1953 * If the inode has never been written to disk (delay == 0),
1954 * then we can process the freeblks now that we have deleted
1955 * the dependencies.
1956 */
1957 if (!delay)
1958 handle_workitem_freeblocks(freeblks);
1959}
1960
6bae6177
MD
1961static int
1962softdep_setup_freeblocks_bp(struct buf *bp, void *data)
1963{
1964 struct softdep_setup_freeblocks_info *info = data;
1965 struct inodedep *inodedep;
1966
2ef16e04 1967 if (getdirtybuf(&bp, MNT_WAIT) == 0) {
086c1d7e 1968 kprintf("softdep_setup_freeblocks_bp(1): caught bp %p going away\n", bp);
6bae6177 1969 return(-1);
2ef16e04
MD
1970 }
1971 if (bp->b_vp != ITOV(info->ip) || (bp->b_flags & B_DELWRI) == 0) {
086c1d7e 1972 kprintf("softdep_setup_freeblocks_bp(2): caught bp %p going away\n", bp);
2ef16e04
MD
1973 BUF_UNLOCK(bp);
1974 return(-1);
1975 }
6bae6177
MD
1976 (void) inodedep_lookup(info->fs, info->ip->i_number, 0, &inodedep);
1977 deallocate_dependencies(bp, inodedep);
1978 bp->b_flags |= B_INVAL | B_NOCACHE;
1979 FREE_LOCK(&lk);
1980 brelse(bp);
1981 ACQUIRE_LOCK(&lk);
1982 return(1);
1983}
1984
984263bc
MD
1985/*
1986 * Reclaim any dependency structures from a buffer that is about to
1987 * be reallocated to a new vnode. The buffer must be locked, thus,
1988 * no I/O completion operations can occur while we are manipulating
1989 * its associated dependencies. The mutex is held so that other I/O's
1990 * associated with related dependencies do not occur.
1991 */
1992static void
3fcb1ab8 1993deallocate_dependencies(struct buf *bp, struct inodedep *inodedep)
984263bc
MD
1994{
1995 struct worklist *wk;
1996 struct indirdep *indirdep;
1997 struct allocindir *aip;
1998 struct pagedep *pagedep;
1999 struct dirrem *dirrem;
2000 struct diradd *dap;
2001 int i;
2002
2003 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
2004 switch (wk->wk_type) {
2005
2006 case D_INDIRDEP:
2007 indirdep = WK_INDIRDEP(wk);
2008 /*
2009 * None of the indirect pointers will ever be visible,
2010 * so they can simply be tossed. GOINGAWAY ensures
2011 * that allocated pointers will be saved in the buffer
2012 * cache until they are freed. Note that they will
2013 * only be able to be found by their physical address
2014 * since the inode mapping the logical address will
2015 * be gone. The save buffer used for the safe copy
2016 * was allocated in setup_allocindir_phase2 using
2017 * the physical address so it could be used for this
2018 * purpose. Hence we swap the safe copy with the real
2019 * copy, allowing the safe copy to be freed and holding
2020 * on to the real copy for later use in indir_trunc.
81b5c339
MD
2021 *
2022 * NOTE: ir_savebp is relative to the block device
2023 * so b_bio1 contains the device block number.
984263bc
MD
2024 */
2025 if (indirdep->ir_state & GOINGAWAY) {
2026 FREE_LOCK(&lk);
2027 panic("deallocate_dependencies: already gone");
2028 }
2029 indirdep->ir_state |= GOINGAWAY;
2030 while ((aip = LIST_FIRST(&indirdep->ir_deplisthd)) != 0)
2031 free_allocindir(aip, inodedep);
54078292
MD
2032 if (bp->b_bio1.bio_offset >= 0 ||
2033 bp->b_bio2.bio_offset != indirdep->ir_savebp->b_bio1.bio_offset) {
984263bc
MD
2034 FREE_LOCK(&lk);
2035 panic("deallocate_dependencies: not indir");
2036 }
2037 bcopy(bp->b_data, indirdep->ir_savebp->b_data,
2038 bp->b_bcount);
2039 WORKLIST_REMOVE(wk);
408357d8 2040 WORKLIST_INSERT_BP(indirdep->ir_savebp, wk);
984263bc
MD
2041 continue;
2042
2043 case D_PAGEDEP:
2044 pagedep = WK_PAGEDEP(wk);
2045 /*
2046 * None of the directory additions will ever be
2047 * visible, so they can simply be tossed.
2048 */
2049 for (i = 0; i < DAHASHSZ; i++)
2050 while ((dap =
2051 LIST_FIRST(&pagedep->pd_diraddhd[i])))
2052 free_diradd(dap);
2053 while ((dap = LIST_FIRST(&pagedep->pd_pendinghd)) != 0)
2054 free_diradd(dap);
2055 /*
2056 * Copy any directory remove dependencies to the list
2057 * to be processed after the zero'ed inode is written.
2058 * If the inode has already been written, then they
2059 * can be dumped directly onto the work list.
2060 */
2061 LIST_FOREACH(dirrem, &pagedep->pd_dirremhd, dm_next) {
2062 LIST_REMOVE(dirrem, dm_next);
2063 dirrem->dm_dirinum = pagedep->pd_ino;
2064 if (inodedep == NULL ||
2065 (inodedep->id_state & ALLCOMPLETE) ==
2066 ALLCOMPLETE)
2067 add_to_worklist(&dirrem->dm_list);
2068 else
2069 WORKLIST_INSERT(&inodedep->id_bufwait,
2070 &dirrem->dm_list);
2071 }
2072 WORKLIST_REMOVE(&pagedep->pd_list);
2073 LIST_REMOVE(pagedep, pd_hash);
2074 WORKITEM_FREE(pagedep, D_PAGEDEP);
2075 continue;
2076
2077 case D_ALLOCINDIR:
2078 free_allocindir(WK_ALLOCINDIR(wk), inodedep);
2079 continue;
2080
2081 case D_ALLOCDIRECT:
2082 case D_INODEDEP:
2083 FREE_LOCK(&lk);
2084 panic("deallocate_dependencies: Unexpected type %s",
2085 TYPENAME(wk->wk_type));
2086 /* NOTREACHED */
2087
2088 default:
2089 FREE_LOCK(&lk);
2090 panic("deallocate_dependencies: Unknown type %s",
2091 TYPENAME(wk->wk_type));
2092 /* NOTREACHED */
2093 }
2094 }
2095}
2096
2097/*
2098 * Free an allocdirect. Generate a new freefrag work request if appropriate.
2099 * This routine must be called with splbio interrupts blocked.
2100 */
2101static void
3fcb1ab8
SW
2102free_allocdirect(struct allocdirectlst *adphead,
2103 struct allocdirect *adp, int delay)
984263bc
MD
2104{
2105
2106#ifdef DEBUG
dadab5e9 2107 if (lk.lkt_held == NOHOLDER)
984263bc
MD
2108 panic("free_allocdirect: lock not held");
2109#endif
2110 if ((adp->ad_state & DEPCOMPLETE) == 0)
2111 LIST_REMOVE(adp, ad_deps);
2112 TAILQ_REMOVE(adphead, adp, ad_next);
2113 if ((adp->ad_state & COMPLETE) == 0)
2114 WORKLIST_REMOVE(&adp->ad_list);
2115 if (adp->ad_freefrag != NULL) {
2116 if (delay)
2117 WORKLIST_INSERT(&adp->ad_inodedep->id_bufwait,
2118 &adp->ad_freefrag->ff_list);
2119 else
2120 add_to_worklist(&adp->ad_freefrag->ff_list);
2121 }
2122 WORKITEM_FREE(adp, D_ALLOCDIRECT);
2123}
2124
2125/*
2126 * Prepare an inode to be freed. The actual free operation is not
2127 * done until the zero'ed inode has been written to disk.
2128 */
2129void
3fcb1ab8 2130softdep_freefile(struct vnode *pvp, ino_t ino, int mode)
984263bc
MD
2131{
2132 struct inode *ip = VTOI(pvp);
2133 struct inodedep *inodedep;
2134 struct freefile *freefile;
2135
2136 /*
2137 * This sets up the inode de-allocation dependency.
2138 */
2139 MALLOC(freefile, struct freefile *, sizeof(struct freefile),
2140 M_FREEFILE, M_SOFTDEP_FLAGS);
2141 freefile->fx_list.wk_type = D_FREEFILE;
2142 freefile->fx_list.wk_state = 0;
2143 freefile->fx_mode = mode;
2144 freefile->fx_oldinum = ino;
2145 freefile->fx_devvp = ip->i_devvp;
2146 freefile->fx_fs = ip->i_fs;
2147
2148 /*
2149 * If the inodedep does not exist, then the zero'ed inode has
2150 * been written to disk. If the allocated inode has never been
2151 * written to disk, then the on-disk inode is zero'ed. In either
2152 * case we can free the file immediately.
2153 */
2154 ACQUIRE_LOCK(&lk);
2155 if (inodedep_lookup(ip->i_fs, ino, 0, &inodedep) == 0 ||
2156 check_inode_unwritten(inodedep)) {
2157 FREE_LOCK(&lk);
2158 handle_workitem_freefile(freefile);
2159 return;
2160 }
2161 WORKLIST_INSERT(&inodedep->id_inowait, &freefile->fx_list);
2162 FREE_LOCK(&lk);
2163}
2164
2165/*
2166 * Check to see if an inode has never been written to disk. If
2167 * so free the inodedep and return success, otherwise return failure.
2168 * This routine must be called with splbio interrupts blocked.
2169 *
2170 * If we still have a bitmap dependency, then the inode has never
2171 * been written to disk. Drop the dependency as it is no longer
2172 * necessary since the inode is being deallocated. We set the
2173 * ALLCOMPLETE flags since the bitmap now properly shows that the
2174 * inode is not allocated. Even if the inode is actively being
2175 * written, it has been rolled back to its zero'ed state, so we
2176 * are ensured that a zero inode is what is on the disk. For short
2177 * lived files, this change will usually result in removing all the
2178 * dependencies from the inode so that it can be freed immediately.
2179 */
2180static int
3fcb1ab8 2181check_inode_unwritten(struct inodedep *inodedep)
984263bc
MD
2182{
2183
2184 if ((inodedep->id_state & DEPCOMPLETE) != 0 ||
2185 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2186 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2187 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2188 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2189 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2190 inodedep->id_nlinkdelta != 0)
2191 return (0);
83b29fff
MD
2192
2193 /*
2194 * Another process might be in initiate_write_inodeblock
2195 * trying to allocate memory without holding "Softdep Lock".
2196 */
2197 if ((inodedep->id_state & IOSTARTED) != 0 &&
2198 inodedep->id_savedino == NULL)
2199 return(0);
2200
984263bc
MD
2201 inodedep->id_state |= ALLCOMPLETE;
2202 LIST_REMOVE(inodedep, id_deps);
2203 inodedep->id_buf = NULL;
2204 if (inodedep->id_state & ONWORKLIST)
2205 WORKLIST_REMOVE(&inodedep->id_list);
2206 if (inodedep->id_savedino != NULL) {
2207 FREE(inodedep->id_savedino, M_INODEDEP);
2208 inodedep->id_savedino = NULL;
2209 }
2210 if (free_inodedep(inodedep) == 0) {
2211 FREE_LOCK(&lk);
2212 panic("check_inode_unwritten: busy inode");
2213 }
2214 return (1);
2215}
2216
2217/*
2218 * Try to free an inodedep structure. Return 1 if it could be freed.
2219 */
2220static int
3fcb1ab8 2221free_inodedep(struct inodedep *inodedep)
984263bc
MD
2222{
2223
2224 if ((inodedep->id_state & ONWORKLIST) != 0 ||
2225 (inodedep->id_state & ALLCOMPLETE) != ALLCOMPLETE ||
2226 LIST_FIRST(&inodedep->id_pendinghd) != NULL ||
2227 LIST_FIRST(&inodedep->id_bufwait) != NULL ||
2228 LIST_FIRST(&inodedep->id_inowait) != NULL ||
2229 TAILQ_FIRST(&inodedep->id_inoupdt) != NULL ||
2230 TAILQ_FIRST(&inodedep->id_newinoupdt) != NULL ||
2231 inodedep->id_nlinkdelta != 0 || inodedep->id_savedino != NULL)
2232 return (0);
2233 LIST_REMOVE(inodedep, id_hash);
2234 WORKITEM_FREE(inodedep, D_INODEDEP);
2235 num_inodedep -= 1;
2236 return (1);
2237}
2238
2239/*
2240 * This workitem routine performs the block de-allocation.
2241 * The workitem is added to the pending list after the updated
2242 * inode block has been written to disk. As mentioned above,
2243 * checks regarding the number of blocks de-allocated (compared
2244 * to the number of blocks allocated for the file) are also
2245 * performed in this function.
2246 */
2247static void
3fcb1ab8 2248handle_workitem_freeblocks(struct freeblks *freeblks)
984263bc
MD
2249{
2250 struct inode tip;
2251 ufs_daddr_t bn;
2252 struct fs *fs;
2253 int i, level, bsize;
2254 long nblocks, blocksreleased = 0;
2255 int error, allerror = 0;
2256 ufs_lbn_t baselbns[NIADDR], tmpval;
2257
2258 tip.i_number = freeblks->fb_previousinum;
2259 tip.i_devvp = freeblks->fb_devvp;
2260 tip.i_dev = freeblks->fb_devvp->v_rdev;
2261 tip.i_fs = freeblks->fb_fs;
2262 tip.i_size = freeblks->fb_oldsize;
2263 tip.i_uid = freeblks->fb_uid;
2264 fs = freeblks->fb_fs;
2265 tmpval = 1;
2266 baselbns[0] = NDADDR;
2267 for (i = 1; i < NIADDR; i++) {
2268 tmpval *= NINDIR(fs);
2269 baselbns[i] = baselbns[i - 1] + tmpval;
2270 }
2271 nblocks = btodb(fs->fs_bsize);
2272 blocksreleased = 0;
2273 /*
2274 * Indirect blocks first.
2275 */
2276 for (level = (NIADDR - 1); level >= 0; level--) {
2277 if ((bn = freeblks->fb_iblks[level]) == 0)
2278 continue;
54078292 2279 if ((error = indir_trunc(&tip, fsbtodoff(fs, bn), level,
984263bc
MD
2280 baselbns[level], &blocksreleased)) == 0)
2281 allerror = error;
2282 ffs_blkfree(&tip, bn, fs->fs_bsize);
2283 blocksreleased += nblocks;
2284 }
2285 /*
2286 * All direct blocks or frags.
2287 */
2288 for (i = (NDADDR - 1); i >= 0; i--) {
2289 if ((bn = freeblks->fb_dblks[i]) == 0)
2290 continue;
2291 bsize = blksize(fs, &tip, i);
2292 ffs_blkfree(&tip, bn, bsize);
2293 blocksreleased += btodb(bsize);
2294 }
2295
2296#ifdef DIAGNOSTIC
2297 if (freeblks->fb_chkcnt != blocksreleased)
086c1d7e 2298 kprintf("handle_workitem_freeblocks: block count\n");
984263bc
MD
2299 if (allerror)
2300 softdep_error("handle_workitem_freeblks", allerror);
2301#endif /* DIAGNOSTIC */
2302 WORKITEM_FREE(freeblks, D_FREEBLKS);
2303}
2304
2305/*
2306 * Release blocks associated with the inode ip and stored in the indirect
54078292
MD
2307 * block at doffset. If level is greater than SINGLE, the block is an
2308 * indirect block and recursive calls to indirtrunc must be used to
2309 * cleanse other indirect blocks.
984263bc
MD
2310 */
2311static int
3fcb1ab8
SW
2312indir_trunc(struct inode *ip, off_t doffset, int level, ufs_lbn_t lbn,
2313 long *countp)
984263bc
MD
2314{
2315 struct buf *bp;
2316 ufs_daddr_t *bap;
2317 ufs_daddr_t nb;
2318 struct fs *fs;
2319 struct worklist *wk;
2320 struct indirdep *indirdep;
2321 int i, lbnadd, nblocks;
2322 int error, allerror = 0;
2323
2324 fs = ip->i_fs;
2325 lbnadd = 1;
2326 for (i = level; i > 0; i--)
2327 lbnadd *= NINDIR(fs);
2328 /*
2329 * Get buffer of block pointers to be freed. This routine is not
2330 * called until the zero'ed inode has been written, so it is safe
2331 * to free blocks as they are encountered. Because the inode has
2332 * been zero'ed, calls to bmap on these blocks will fail. So, we
2333 * have to use the on-disk address and the block device for the
2334 * filesystem to look them up. If the file was deleted before its
2335 * indirect blocks were all written to disk, the routine that set
2336 * us up (deallocate_dependencies) will have arranged to leave
2337 * a complete copy of the indirect block in memory for our use.
2338 * Otherwise we have to read the blocks in from the disk.
2339 */
2340 ACQUIRE_LOCK(&lk);
b1c20cfa 2341 if ((bp = findblk(ip->i_devvp, doffset, FINDBLK_TEST)) != NULL &&
984263bc 2342 (wk = LIST_FIRST(&bp->b_dep)) != NULL) {
1f1ea522
MD
2343 /*
2344 * bp must be ir_savebp, which is held locked for our use.
2345 */
984263bc
MD
2346 if (wk->wk_type != D_INDIRDEP ||
2347 (indirdep = WK_INDIRDEP(wk))->ir_savebp != bp ||
2348 (indirdep->ir_state & GOINGAWAY) == 0) {
2349 FREE_LOCK(&lk);
2350 panic("indir_trunc: lost indirdep");
2351 }
2352 WORKLIST_REMOVE(wk);
2353 WORKITEM_FREE(indirdep, D_INDIRDEP);
2354 if (LIST_FIRST(&bp->b_dep) != NULL) {
2355 FREE_LOCK(&lk);
2356 panic("indir_trunc: dangling dep");
2357 }
2358 FREE_LOCK(&lk);
2359 } else {
2360 FREE_LOCK(&lk);
54078292 2361 error = bread(ip->i_devvp, doffset, (int)fs->fs_bsize, &bp);
984263bc
MD
2362 if (error)
2363 return (error);
2364 }
2365 /*
2366 * Recursively free indirect blocks.
2367 */
2368 bap = (ufs_daddr_t *)bp->b_data;
2369 nblocks = btodb(fs->fs_bsize);
2370 for (i = NINDIR(fs) - 1; i >= 0; i--) {
2371 if ((nb = bap[i]) == 0)
2372 continue;
2373 if (level != 0) {
54078292 2374 if ((error = indir_trunc(ip, fsbtodoff(fs, nb),
984263bc
MD
2375 level - 1, lbn + (i * lbnadd), countp)) != 0)
2376 allerror = error;
2377 }
2378 ffs_blkfree(ip, nb, fs->fs_bsize);
2379 *countp += nblocks;
2380 }
2381 bp->b_flags |= B_INVAL | B_NOCACHE;
2382 brelse(bp);
2383 return (allerror);
2384}
2385
2386/*
2387 * Free an allocindir.
2388 * This routine must be called with splbio interrupts blocked.
2389 */
2390static void
3fcb1ab8 2391free_allocindir(struct allocindir *aip, struct inodedep *inodedep)
984263bc
MD
2392{
2393 struct freefrag *freefrag;
2394
2395#ifdef DEBUG
dadab5e9 2396 if (lk.lkt_held == NOHOLDER)
984263bc
MD
2397 panic("free_allocindir: lock not held");
2398#endif
2399 if ((aip->ai_state & DEPCOMPLETE) == 0)
2400 LIST_REMOVE(aip, ai_deps);
2401 if (aip->ai_state & ONWORKLIST)
2402 WORKLIST_REMOVE(&aip->ai_list);
2403 LIST_REMOVE(aip, ai_next);
2404 if ((freefrag = aip->ai_freefrag) != NULL) {
2405 if (inodedep == NULL)
2406 add_to_worklist(&freefrag->ff_list);
2407 else
2408 WORKLIST_INSERT(&inodedep->id_bufwait,
2409 &freefrag->ff_list);
2410 }
2411 WORKITEM_FREE(aip, D_ALLOCINDIR);
2412}
2413
2414/*
2415 * Directory entry addition dependencies.
2416 *
2417 * When adding a new directory entry, the inode (with its incremented link
2418 * count) must be written to disk before the directory entry's pointer to it.
2419 * Also, if the inode is newly allocated, the corresponding freemap must be
2420 * updated (on disk) before the directory entry's pointer. These requirements
2421 * are met via undo/redo on the directory entry's pointer, which consists
2422 * simply of the inode number.
2423 *
2424 * As directory entries are added and deleted, the free space within a
f719c866 2425 * directory block can become fragmented. The ufs filesystem will compact
984263bc
MD
2426 * a fragmented directory block to make space for a new entry. When this
2427 * occurs, the offsets of previously added entries change. Any "diradd"
2428 * dependency structures corresponding to these entries must be updated with
2429 * the new offsets.
2430 */
2431
2432/*
2433 * This routine is called after the in-memory inode's link
2434 * count has been incremented, but before the directory entry's
2435 * pointer to the inode has been set.
3fcb1ab8
SW
2436 *
2437 * Parameters:
2438 * bp: buffer containing directory block
2439 * dp: inode for directory
2440 * diroffset: offset of new entry in directory
2441 * newinum: inode referenced by new directory entry
2442 * newdirbp: non-NULL => contents of new mkdir
984263bc
MD
2443 */
2444void
3fcb1ab8
SW
2445softdep_setup_directory_add(struct buf *bp, struct inode *dp, off_t diroffset,
2446 ino_t newinum, struct buf *newdirbp)
984263bc
MD
2447{
2448 int offset; /* offset of new entry within directory block */
2449 ufs_lbn_t lbn; /* block in directory containing new entry */
2450 struct fs *fs;
2451 struct diradd *dap;
2452 struct pagedep *pagedep;
2453 struct inodedep *inodedep;
2454 struct mkdir *mkdir1, *mkdir2;
2455
2456 /*
2457 * Whiteouts have no dependencies.
2458 */
2459 if (newinum == WINO) {
2460 if (newdirbp != NULL)
2461 bdwrite(newdirbp);
2462 return;
2463 }
2464
2465 fs = dp->i_fs;
2466 lbn = lblkno(fs, diroffset);
2467 offset = blkoff(fs, diroffset);
2468 MALLOC(dap, struct diradd *, sizeof(struct diradd), M_DIRADD,
e7b4468c 2469 M_SOFTDEP_FLAGS | M_ZERO);
984263bc
MD
2470 dap->da_list.wk_type = D_DIRADD;
2471 dap->da_offset = offset;
2472 dap->da_newinum = newinum;
2473 dap->da_state = ATTACHED;
2474 if (newdirbp == NULL) {
2475 dap->da_state |= DEPCOMPLETE;
2476 ACQUIRE_LOCK(&lk);
2477 } else {
2478 dap->da_state |= MKDIR_BODY | MKDIR_PARENT;
2479 MALLOC(mkdir1, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2480 M_SOFTDEP_FLAGS);
2481 mkdir1->md_list.wk_type = D_MKDIR;
2482 mkdir1->md_state = MKDIR_BODY;
2483 mkdir1->md_diradd = dap;
2484 MALLOC(mkdir2, struct mkdir *, sizeof(struct mkdir), M_MKDIR,
2485 M_SOFTDEP_FLAGS);
2486 mkdir2->md_list.wk_type = D_MKDIR;
2487 mkdir2->md_state = MKDIR_PARENT;
2488 mkdir2->md_diradd = dap;
2489 /*
2490 * Dependency on "." and ".." being written to disk.
2491 */
2492 mkdir1->md_buf = newdirbp;
2493 ACQUIRE_LOCK(&lk);
2494 LIST_INSERT_HEAD(&mkdirlisthd, mkdir1, md_mkdirs);
408357d8 2495 WORKLIST_INSERT_BP(newdirbp, &mkdir1->md_list);
984263bc
MD
2496 FREE_LOCK(&lk);
2497 bdwrite(newdirbp);
2498 /*
2499 * Dependency on link count increase for parent directory
2500 */
2501 ACQUIRE_LOCK(&lk);
2502 if (inodedep_lookup(dp->i_fs, dp->i_number, 0, &inodedep) == 0
2503 || (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2504 dap->da_state &= ~MKDIR_PARENT;
2505 WORKITEM_FREE(mkdir2, D_MKDIR);
2506 } else {
2507 LIST_INSERT_HEAD(&mkdirlisthd, mkdir2, md_mkdirs);
2508 WORKLIST_INSERT(&inodedep->id_bufwait,&mkdir2->md_list);
2509 }
2510 }
2511 /*
2512 * Link into parent directory pagedep to await its being written.
2513 */
2514 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
408357d8 2515 WORKLIST_INSERT_BP(bp, &pagedep->pd_list);
984263bc
MD
2516 dap->da_pagedep = pagedep;
2517 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)], dap,
2518 da_pdlist);
2519 /*
2520 * Link into its inodedep. Put it on the id_bufwait list if the inode
2521 * is not yet written. If it is written, do the post-inode write
2522 * processing to put it on the id_pendinghd list.
2523 */
2524 (void) inodedep_lookup(fs, newinum, DEPALLOC, &inodedep);
2525 if ((inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE)
2526 diradd_inode_written(dap, inodedep);
2527 else
2528 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2529 FREE_LOCK(&lk);
2530}
2531
2532/*
2533 * This procedure is called to change the offset of a directory
2534 * entry when compacting a directory block which must be owned
2535 * exclusively by the caller. Note that the actual entry movement
2536 * must be done in this procedure to ensure that no I/O completions
2537 * occur while the move is in progress.
3fcb1ab8
SW
2538 *
2539 * Parameters:
2540 * dp: inode for directory
2541 * base: address of dp->i_offset
2542 * oldloc: address of old directory location
2543 * newloc: address of new directory location
2544 * entrysize: size of directory entry
984263bc
MD
2545 */
2546void
3fcb1ab8
SW
2547softdep_change_directoryentry_offset(struct inode *dp, caddr_t base,
2548 caddr_t oldloc, caddr_t newloc,
2549 int entrysize)
984263bc
MD
2550{
2551 int offset, oldoffset, newoffset;
2552 struct pagedep *pagedep;
2553 struct diradd *dap;
2554 ufs_lbn_t lbn;
2555
2556 ACQUIRE_LOCK(&lk);
2557 lbn = lblkno(dp->i_fs, dp->i_offset);
2558 offset = blkoff(dp->i_fs, dp->i_offset);
2559 if (pagedep_lookup(dp, lbn, 0, &pagedep) == 0)
2560 goto done;
2561 oldoffset = offset + (oldloc - base);
2562 newoffset = offset + (newloc - base);
2563
2564 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(oldoffset)], da_pdlist) {
2565 if (dap->da_offset != oldoffset)
2566 continue;
2567 dap->da_offset = newoffset;
2568 if (DIRADDHASH(newoffset) == DIRADDHASH(oldoffset))
2569 break;
2570 LIST_REMOVE(dap, da_pdlist);
2571 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(newoffset)],
2572 dap, da_pdlist);
2573 break;
2574 }
2575 if (dap == NULL) {
2576
2577 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist) {
2578 if (dap->da_offset == oldoffset) {
2579 dap->da_offset = newoffset;
2580 break;
2581 }
2582 }
2583 }
2584done:
2585 bcopy(oldloc, newloc, entrysize);
2586 FREE_LOCK(&lk);
2587}
2588
2589/*
2590 * Free a diradd dependency structure. This routine must be called
2591 * with splbio interrupts blocked.
2592 */
2593static void
3fcb1ab8 2594free_diradd(struct diradd *dap)
984263bc
MD
2595{
2596 struct dirrem *dirrem;
2597 struct pagedep *pagedep;
2598 struct inodedep *inodedep;
2599 struct mkdir *mkdir, *nextmd;
2600
2601#ifdef DEBUG
dadab5e9 2602 if (lk.lkt_held == NOHOLDER)
984263bc
MD
2603 panic("free_diradd: lock not held");
2604#endif
2605 WORKLIST_REMOVE(&dap->da_list);
2606 LIST_REMOVE(dap, da_pdlist);
2607 if ((dap->da_state & DIRCHG) == 0) {
2608 pagedep = dap->da_pagedep;
2609 } else {
2610 dirrem = dap->da_previous;
2611 pagedep = dirrem->dm_pagedep;
2612 dirrem->dm_dirinum = pagedep->pd_ino;
2613 add_to_worklist(&dirrem->dm_list);
2614 }
2615 if (inodedep_lookup(VFSTOUFS(pagedep->pd_mnt)->um_fs, dap->da_newinum,
2616 0, &inodedep) != 0)
2617 (void) free_inodedep(inodedep);
2618 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2619 for (mkdir = LIST_FIRST(&mkdirlisthd); mkdir; mkdir = nextmd) {
2620 nextmd = LIST_NEXT(mkdir, md_mkdirs);
2621 if (mkdir->md_diradd != dap)
2622 continue;
2623 dap->da_state &= ~mkdir->md_state;
2624 WORKLIST_REMOVE(&mkdir->md_list);
2625 LIST_REMOVE(mkdir, md_mkdirs);
2626 WORKITEM_FREE(mkdir, D_MKDIR);
2627 }
2628 if ((dap->da_state & (MKDIR_PARENT | MKDIR_BODY)) != 0) {
2629 FREE_LOCK(&lk);
2630 panic("free_diradd: unfound ref");
2631 }
2632 }
2633 WORKITEM_FREE(dap, D_DIRADD);
2634}
2635
2636/*
2637 * Directory entry removal dependencies.
2638 *
2639 * When removing a directory entry, the entry's inode pointer must be
2640 * zero'ed on disk before the corresponding inode's link count is decremented
2641 * (possibly freeing the inode for re-use). This dependency is handled by
2642 * updating the directory entry but delaying the inode count reduction until
2643 * after the directory block has been written to disk. After this point, the
2644 * inode count can be decremented whenever it is convenient.
2645 */
2646
2647/*
2648 * This routine should be called immediately after removing
2649 * a directory entry. The inode's link count should not be
2650 * decremented by the calling procedure -- the soft updates
2651 * code will do this task when it is safe.
3fcb1ab8
SW
2652 *
2653 * Parameters:
2654 * bp: buffer containing directory block
2655 * dp: inode for the directory being modified
2656 * ip: inode for directory entry being removed
2657 * isrmdir: indicates if doing RMDIR
984263bc
MD
2658 */
2659void
3fcb1ab8
SW
2660softdep_setup_remove(struct buf *bp, struct inode *dp, struct inode *ip,
2661 int isrmdir)
984263bc
MD
2662{
2663 struct dirrem *dirrem, *prevdirrem;
2664
2665 /*
2666 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2667 */
2668 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2669
2670 /*
2671 * If the COMPLETE flag is clear, then there were no active
2672 * entries and we want to roll back to a zeroed entry until
2673 * the new inode is committed to disk. If the COMPLETE flag is
2674 * set then we have deleted an entry that never made it to
2675 * disk. If the entry we deleted resulted from a name change,
2676 * then the old name still resides on disk. We cannot delete
2677 * its inode (returned to us in prevdirrem) until the zeroed
2678 * directory entry gets to disk. The new inode has never been
2679 * referenced on the disk, so can be deleted immediately.
2680 */
2681 if ((dirrem->dm_state & COMPLETE) == 0) {
2682 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd, dirrem,
2683 dm_next);
2684 FREE_LOCK(&lk);
2685 } else {
2686 if (prevdirrem != NULL)
2687 LIST_INSERT_HEAD(&dirrem->dm_pagedep->pd_dirremhd,
2688 prevdirrem, dm_next);
2689 dirrem->dm_dirinum = dirrem->dm_pagedep->pd_ino;
2690 FREE_LOCK(&lk);
2691 handle_workitem_remove(dirrem);
2692 }
2693}
2694
2695/*
2696 * Allocate a new dirrem if appropriate and return it along with
2697 * its associated pagedep. Called without a lock, returns with lock.
2698 */
2699static long num_dirrem; /* number of dirrem allocated */
3fcb1ab8
SW
2700
2701/*
2702 * Parameters:
2703 * bp: buffer containing directory block
2704 * dp: inode for the directory being modified
2705 * ip: inode for directory entry being removed
2706 * isrmdir: indicates if doing RMDIR
2707 * prevdirremp: previously referenced inode, if any
2708 */
984263bc 2709static struct dirrem *
3fcb1ab8
SW
2710newdirrem(struct buf *bp, struct inode *dp, struct inode *ip,
2711 int isrmdir, struct dirrem **prevdirremp)
984263bc
MD
2712{
2713 int offset;
2714 ufs_lbn_t lbn;
2715 struct diradd *dap;
2716 struct dirrem *dirrem;
2717 struct pagedep *pagedep;
2718
2719 /*
2720 * Whiteouts have no deletion dependencies.
2721 */
2722 if (ip == NULL)
2723 panic("newdirrem: whiteout");
2724 /*
2725 * If we are over our limit, try to improve the situation.
2726 * Limiting the number of dirrem structures will also limit
2727 * the number of freefile and freeblks structures.
2728 */
2729 if (num_dirrem > max_softdeps / 2 && speedup_syncer() == 0)
2730 (void) request_cleanup(FLUSH_REMOVE, 0);
2731 num_dirrem += 1;
2732 MALLOC(dirrem, struct dirrem *, sizeof(struct dirrem),
e7b4468c 2733 M_DIRREM, M_SOFTDEP_FLAGS | M_ZERO);
984263bc
MD
2734 dirrem->dm_list.wk_type = D_DIRREM;
2735 dirrem->dm_state = isrmdir ? RMDIR : 0;
2736 dirrem->dm_mnt = ITOV(ip)->v_mount;
2737 dirrem->dm_oldinum = ip->i_number;
2738 *prevdirremp = NULL;
2739
2740 ACQUIRE_LOCK(&lk);
2741 lbn = lblkno(dp->i_fs, dp->i_offset);
2742 offset = blkoff(dp->i_fs, dp->i_offset);
2743 if (pagedep_lookup(dp, lbn, DEPALLOC, &pagedep) == 0)
408357d8 2744 WORKLIST_INSERT_BP(bp, &pagedep->pd_list);
984263bc
MD
2745 dirrem->dm_pagedep = pagedep;
2746 /*
2747 * Check for a diradd dependency for the same directory entry.
2748 * If present, then both dependencies become obsolete and can
2749 * be de-allocated. Check for an entry on both the pd_dirraddhd
2750 * list and the pd_pendinghd list.
2751 */
2752
2753 LIST_FOREACH(dap, &pagedep->pd_diraddhd[DIRADDHASH(offset)], da_pdlist)
2754 if (dap->da_offset == offset)
2755 break;
2756 if (dap == NULL) {
2757
2758 LIST_FOREACH(dap, &pagedep->pd_pendinghd, da_pdlist)
2759 if (dap->da_offset == offset)
2760 break;
2761 if (dap == NULL)
2762 return (dirrem);
2763 }
2764 /*
2765 * Must be ATTACHED at this point.
2766 */
2767 if ((dap->da_state & ATTACHED) == 0) {
2768 FREE_LOCK(&lk);
2769 panic("newdirrem: not ATTACHED");
2770 }
2771 if (dap->da_newinum != ip->i_number) {
2772 FREE_LOCK(&lk);
f91a71dd 2773 panic("newdirrem: inum %"PRId64" should be %"PRId64,
984263bc
MD
2774 ip->i_number, dap->da_newinum);
2775 }
2776 /*
2777 * If we are deleting a changed name that never made it to disk,
2778 * then return the dirrem describing the previous inode (which
2779 * represents the inode currently referenced from this entry on disk).
2780 */
2781 if ((dap->da_state & DIRCHG) != 0) {
2782 *prevdirremp = dap->da_previous;
2783 dap->da_state &= ~DIRCHG;
2784 dap->da_pagedep = pagedep;
2785 }
2786 /*
2787 * We are deleting an entry that never made it to disk.
2788 * Mark it COMPLETE so we can delete its inode immediately.
2789 */
2790 dirrem->dm_state |= COMPLETE;
2791 free_diradd(dap);
2792 return (dirrem);
2793}
2794
2795/*
2796 * Directory entry change dependencies.
2797 *
2798 * Changing an existing directory entry requires that an add operation
2799 * be completed first followed by a deletion. The semantics for the addition
2800 * are identical to the description of adding a new entry above except
2801 * that the rollback is to the old inode number rather than zero. Once
2802 * the addition dependency is completed, the removal is done as described
2803 * in the removal routine above.
2804 */
2805
2806/*
2807 * This routine should be called immediately after changing
2808 * a directory entry. The inode's link count should not be
2809 * decremented by the calling procedure -- the soft updates
2810 * code will perform this task when it is safe.
3fcb1ab8
SW
2811 *
2812 * Parameters:
2813 * bp: buffer containing directory block
2814 * dp: inode for the directory being modified
2815 * ip: inode for directory entry being removed
2816 * newinum: new inode number for changed entry
2817 * isrmdir: indicates if doing RMDIR
984263bc
MD
2818 */
2819void
3fcb1ab8
SW
2820softdep_setup_directory_change(struct buf *bp, struct inode *dp,
2821 struct inode *ip, ino_t newinum,
2822 int isrmdir)
984263bc
MD
2823{
2824 int offset;
2825 struct diradd *dap = NULL;
2826 struct dirrem *dirrem, *prevdirrem;
2827 struct pagedep *pagedep;
2828 struct inodedep *inodedep;
2829
2830 offset = blkoff(dp->i_fs, dp->i_offset);
2831
2832 /*
2833 * Whiteouts do not need diradd dependencies.
2834 */
2835 if (newinum != WINO) {
2836 MALLOC(dap, struct diradd *, sizeof(struct diradd),
e7b4468c 2837 M_DIRADD, M_SOFTDEP_FLAGS | M_ZERO);
984263bc
MD
2838 dap->da_list.wk_type = D_DIRADD;
2839 dap->da_state = DIRCHG | ATTACHED | DEPCOMPLETE;
2840 dap->da_offset = offset;
2841 dap->da_newinum = newinum;
2842 }
2843
2844 /*
2845 * Allocate a new dirrem and ACQUIRE_LOCK.
2846 */
2847 dirrem = newdirrem(bp, dp, ip, isrmdir, &prevdirrem);
2848 pagedep = dirrem->dm_pagedep;
2849 /*
2850 * The possible values for isrmdir:
2851 * 0 - non-directory file rename
2852 * 1 - directory rename within same directory
2853 * inum - directory rename to new directory of given inode number
2854 * When renaming to a new directory, we are both deleting and
2855 * creating a new directory entry, so the link count on the new
2856 * directory should not change. Thus we do not need the followup
2857 * dirrem which is usually done in handle_workitem_remove. We set
2858 * the DIRCHG flag to tell handle_workitem_remove to skip the
2859 * followup dirrem.
2860 */
2861 if (isrmdir > 1)
2862 dirrem->dm_state |= DIRCHG;
2863
2864 /*
2865 * Whiteouts have no additional dependencies,
2866 * so just put the dirrem on the correct list.
2867 */
2868 if (newinum == WINO) {
2869 if ((dirrem->dm_state & COMPLETE) == 0) {
2870 LIST_INSERT_HEAD(&pagedep->pd_dirremhd, dirrem,
2871 dm_next);
2872 } else {
2873 dirrem->dm_dirinum = pagedep->pd_ino;
2874 add_to_worklist(&dirrem->dm_list);
2875 }
2876 FREE_LOCK(&lk);
2877 return;
2878 }
2879
2880 /*
2881 * If the COMPLETE flag is clear, then there were no active
2882 * entries and we want to roll back to the previous inode until
2883 * the new inode is committed to disk. If the COMPLETE flag is
2884 * set, then we have deleted an entry that never made it to disk.
2885 * If the entry we deleted resulted from a name change, then the old
2886 * inode reference still resides on disk. Any rollback that we do
2887 * needs to be to that old inode (returned to us in prevdirrem). If
2888 * the entry we deleted resulted from a create, then there is
2889 * no entry on the disk, so we want to roll back to zero rather
2890 * than the uncommitted inode. In either of the COMPLETE cases we
2891 * want to immediately free the unwritten and unreferenced inode.
2892 */
2893 if ((dirrem->dm_state & COMPLETE) == 0) {
2894 dap->da_previous = dirrem;
2895 } else {
2896 if (prevdirrem != NULL) {
2897 dap->da_previous = prevdirrem;
2898 } else {
2899 dap->da_state &= ~DIRCHG;
2900 dap->da_pagedep = pagedep;
2901 }
2902 dirrem->dm_dirinum = pagedep->pd_ino;
2903 add_to_worklist(&dirrem->dm_list);
2904 }
2905 /*
2906 * Link into its inodedep. Put it on the id_bufwait list if the inode
2907 * is not yet written. If it is written, do the post-inode write
2908 * processing to put it on the id_pendinghd list.
2909 */
2910 if (inodedep_lookup(dp->i_fs, newinum, DEPALLOC, &inodedep) == 0 ||
2911 (inodedep->id_state & ALLCOMPLETE) == ALLCOMPLETE) {
2912 dap->da_state |= COMPLETE;
2913 LIST_INSERT_HEAD(&pagedep->pd_pendinghd, dap, da_pdlist);
2914 WORKLIST_INSERT(&inodedep->id_pendinghd, &dap->da_list);
2915 } else {
2916 LIST_INSERT_HEAD(&pagedep->pd_diraddhd[DIRADDHASH(offset)],
2917 dap, da_pdlist);
2918 WORKLIST_INSERT(&inodedep->id_bufwait, &dap->da_list);
2919 }
2920 FREE_LOCK(&lk);
2921}
2922
2923/*
2924 * Called whenever the link count on an inode is changed.
2925 * It creates an inode dependency so that the new reference(s)
2926 * to the inode cannot be committed to disk until the updated
2927 * inode has been written.
3fcb1ab8
SW
2928 *
2929 * Parameters:
2930 * ip: the inode with the increased link count
984263bc
MD
2931 */
2932void
3fcb1ab8 2933softdep_change_linkcnt(struct inode *ip)
984263bc
MD
2934{
2935 struct inodedep *inodedep;
2936
2937 ACQUIRE_LOCK(&lk);
2938 (void) inodedep_lookup(ip->i_fs, ip->i_number, DEPALLOC, &inodedep);
2939 if (ip->i_nlink < ip->i_effnlink) {
2940 FREE_LOCK(&lk);
2941 panic("softdep_change_linkcnt: bad delta");
2942 }
2943 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2944 FREE_LOCK(&lk);
2945}
2946
2947/*
2948 * This workitem decrements the inode's link count.
2949 * If the link count reaches zero, the file is removed.
2950 */
2951static void
3fcb1ab8 2952handle_workitem_remove(struct dirrem *dirrem)
984263bc 2953{
984263bc
MD
2954 struct inodedep *inodedep;
2955 struct vnode *vp;
2956 struct inode *ip;
2957 ino_t oldinum;
2958 int error;
2959
b9b0a6d0
MD
2960 error = VFS_VGET(dirrem->dm_mnt, NULL, dirrem->dm_oldinum, &vp);
2961 if (error) {
984263bc
MD
2962 softdep_error("handle_workitem_remove: vget", error);
2963 return;
2964 }
2965 ip = VTOI(vp);
2966 ACQUIRE_LOCK(&lk);
2967 if ((inodedep_lookup(ip->i_fs, dirrem->dm_oldinum, 0, &inodedep)) == 0){
2968 FREE_LOCK(&lk);
2969 panic("handle_workitem_remove: lost inodedep");
2970 }
2971 /*
2972 * Normal file deletion.
2973 */
2974 if ((dirrem->dm_state & RMDIR) == 0) {
2975 ip->i_nlink--;
2976 ip->i_flag |= IN_CHANGE;
2977 if (ip->i_nlink < ip->i_effnlink) {
2978 FREE_LOCK(&lk);
2979 panic("handle_workitem_remove: bad file delta");
2980 }
2981 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
2982 FREE_LOCK(&lk);
2983 vput(vp);
2984 num_dirrem -= 1;
2985 WORKITEM_FREE(dirrem, D_DIRREM);
2986 return;
2987 }
2988 /*
2989 * Directory deletion. Decrement reference count for both the
2990 * just deleted parent directory entry and the reference for ".".
2991 * Next truncate the directory to length zero. When the
2992 * truncation completes, arrange to have the reference count on
2993 * the parent decremented to account for the loss of "..".
2994 */
2995 ip->i_nlink -= 2;
2996 ip->i_flag |= IN_CHANGE;
2997 if (ip->i_nlink < ip->i_effnlink) {
2998 FREE_LOCK(&lk);
2999 panic("handle_workitem_remove: bad dir delta");
3000 }
3001 inodedep->id_nlinkdelta = ip->i_nlink - ip->i_effnlink;
3002 FREE_LOCK(&lk);
ac690a1d 3003 if ((error = ffs_truncate(vp, (off_t)0, 0, proc0.p_ucred)) != 0)
984263bc
MD
3004 softdep_error("handle_workitem_remove: truncate", error);
3005 /*
3006 * Rename a directory to a new parent. Since, we are both deleting
3007 * and creating a new directory entry, the link count on the new
3008 * directory should not change. Thus we skip the followup dirrem.
3009 */
3010 if (dirrem->dm_state & DIRCHG) {
3011 vput(vp);
3012 num_dirrem -= 1;
3013 WORKITEM_FREE(dirrem, D_DIRREM);
3014 return;
3015 }
3016 /*
3017 * If the inodedep does not exist, then the zero'ed inode has
3018 * been written to disk. If the allocated inode has never been
3019 * written to disk, then the on-disk inode is zero'ed. In either
3020 * case we can remove the file immediately.
3021 */
3022 ACQUIRE_LOCK(&lk);
3023 dirrem->dm_state = 0;
3024 oldinum = dirrem->dm_oldinum;
3025 dirrem->dm_oldinum = dirrem->dm_dirinum;
3026 if (inodedep_lookup(ip->i_fs, oldinum, 0, &inodedep) == 0 ||
3027 check_inode_unwritten(inodedep)) {
3028 FREE_LOCK(&lk);
3029 vput(vp);
3030 handle_workitem_remove(dirrem);
3031 return;
3032 }
3033 WORKLIST_INSERT(&inodedep->id_inowait, &dirrem->dm_list);
3034 FREE_LOCK(&lk);
a0ff2e99
MD
3035 ip->i_flag |= IN_CHANGE;
3036 ffs_update(vp, 0);
984263bc
MD
3037 vput(vp);
3038}
3039
3040/*
3041 * Inode de-allocation dependencies.
3042 *
3043 * When an inode's link count is reduced to zero, it can be de-allocated. We
3044 * found it convenient to postpone de-allocation until after the inode is
3045 * written to disk with its new link count (zero). At this point, all of the
3046 * on-disk inode's block pointers are nullified and, with careful dependency
3047 * list ordering, all dependencies related to the inode will be satisfied and
3048 * the corresponding dependency structures de-allocated. So, if/when the
3049 * inode is reused, there will be no mixing of old dependencies with new
3050 * ones. This artificial dependency is set up by the block de-allocation
3051 * procedure above (softdep_setup_freeblocks) and completed by the
3052 * following procedure.
3053 */
3054static void
3fcb1ab8 3055handle_workitem_freefile(struct freefile *freefile)
984263bc
MD
3056{
3057 struct vnode vp;
3058 struct inode tip;
3059 struct inodedep *idp;
3060 int error;
3061
3062#ifdef DEBUG
3063 ACQUIRE_LOCK(&lk);
3064 error = inodedep_lookup(freefile->fx_fs, freefile->fx_oldinum, 0, &idp);
3065 FREE_LOCK(&lk);
3066 if (error)
3067 panic("handle_workitem_freefile: inodedep survived");
3068#endif
3069 tip.i_devvp = freefile->fx_devvp;
3070 tip.i_dev = freefile->fx_devvp->v_rdev;
3071 tip.i_fs = freefile->fx_fs;
3072 vp.v_data = &tip;
3073 if ((error = ffs_freefile(&vp, freefile->fx_oldinum, freefile->fx_mode)) != 0)
3074 softdep_error("handle_workitem_freefile", error);
3075 WORKITEM_FREE(freefile, D_FREEFILE);
3076}
3077
3078/*
83b29fff
MD
3079 * Helper function which unlinks marker element from work list and returns
3080 * the next element on the list.
3081 */
3082static __inline struct worklist *
3083markernext(struct worklist *marker)
3084{
3085 struct worklist *next;
3086
3087 next = LIST_NEXT(marker, wk_list);
3088 LIST_REMOVE(marker, wk_list);
3089 return next;
3090}
3091
3092/*
27bc0cb1
MD
3093 * checkread, checkwrite
3094 *
59a647b1 3095 * bioops callback - hold io_token
27bc0cb1
MD
3096 */
3097static int
3098softdep_checkread(struct buf *bp)
3099{
59a647b1 3100 /* nothing to do, mp lock not needed */
27bc0cb1
MD
3101 return(0);
3102}
3103
59a647b1
MD
3104/*
3105 * bioops callback - hold io_token
3106 */
27bc0cb1
MD
3107static int
3108softdep_checkwrite(struct buf *bp)
3109{
59a647b1 3110 /* nothing to do, mp lock not needed */
27bc0cb1
MD
3111 return(0);
3112}
3113
3114/*
984263bc
MD
3115 * Disk writes.
3116 *
3117 * The dependency structures constructed above are most actively used when file
3118 * system blocks are written to disk. No constraints are placed on when a
3119 * block can be written, but unsatisfied update dependencies are made safe by
3120 * modifying (or replacing) the source memory for the duration of the disk
3121 * write. When the disk write completes, the memory block is again brought
3122 * up-to-date.
3123 *
3124 * In-core inode structure reclamation.
3125 *
3126 * Because there are a finite number of "in-core" inode structures, they are
3127 * reused regularly. By transferring all inode-related dependencies to the
3128 * in-memory inode block and indexing them separately (via "inodedep"s), we
3129 * can allow "in-core" inode structures to be reused at any time and avoid
3130 * any increase in contention.
3131 *
3132 * Called just before entering the device driver to initiate a new disk I/O.
3133 * The buffer must be locked, thus, no I/O completion operations can occur
3134 * while we are manipulating its associated dependencies.
3fcb1ab8 3135 *
59a647b1
MD
3136 * bioops callback - hold io_token
3137 *
3fcb1ab8
SW
3138 * Parameters:
3139 * bp: structure describing disk write to occur
984263bc
MD
3140 */
3141static void
3fcb1ab8 3142softdep_disk_io_initiation(struct buf *bp)
984263bc 3143{
83b29fff
MD
3144 struct worklist *wk;
3145 struct worklist marker;
984263bc
MD
3146 struct indirdep *indirdep;
3147
3148 /*
3149 * We only care about write operations. There should never
3150 * be dependencies for reads.
3151 */
10f3fee5 3152 if (bp->b_cmd == BUF_CMD_READ)
984263bc 3153 panic("softdep_disk_io_initiation: read");
83b29fff 3154
59a647b1 3155 get_mplock();
83b29fff
MD
3156 marker.wk_type = D_LAST + 1; /* Not a normal workitem */
3157
984263bc
MD
3158 /*
3159 * Do any necessary pre-I/O processing.
3160 */
83b29fff
MD
3161 for (wk = LIST_FIRST(&bp->b_dep); wk; wk = markernext(&marker)) {
3162 LIST_INSERT_AFTER(wk, &marker, wk_list);
3163
984263bc 3164 switch (wk->wk_type) {
984263bc
MD
3165 case D_PAGEDEP:
3166 initiate_write_filepage(WK_PAGEDEP(wk), bp);
3167 continue;
3168
3169 case D_INODEDEP:
3170 initiate_write_inodeblock(WK_INODEDEP(wk), bp);
3171 continue;
3172
3173 case D_INDIRDEP:
3174 indirdep = WK_INDIRDEP(wk);
3175 if (indirdep->ir_state & GOINGAWAY)
3176 panic("disk_io_initiation: indirdep gone");
3177 /*
3178 * If there are no remaining dependencies, this
3179 * will be writing the real pointers, so the
3180 * dependency can be freed.
3181 */
3182 if (LIST_FIRST(&indirdep->ir_deplisthd) == NULL) {
3183 indirdep->ir_savebp->b_flags |= B_INVAL | B_NOCACHE;
3184 brelse(indirdep->ir_savebp);
3185 /* inline expand WORKLIST_REMOVE(wk); */
3186 wk->wk_state &= ~ONWORKLIST;
3187 LIST_REMOVE(wk, wk_list);
3188 WORKITEM_FREE(indirdep, D_INDIRDEP);
3189 continue;
3190 }
3191 /*
3192 * Replace up-to-date version with safe version.
3193 */
3194 MALLOC(indirdep->ir_saveddata, caddr_t, bp->b_bcount,
3195 M_INDIRDEP, M_SOFTDEP_FLAGS);
3196 ACQUIRE_LOCK(&lk);
3197 indirdep->ir_state &= ~ATTACHED;
3198 indirdep->ir_state |= UNDONE;
3199 bcopy(bp->b_data, indirdep->ir_saveddata, bp->b_bcount);
3200 bcopy(indirdep->ir_savebp->b_data, bp->b_data,
3201 bp->b_bcount);
3202 FREE_LOCK(&lk);
3203 continue;
3204
3205 case D_MKDIR:
3206 case D_BMSAFEMAP:
3207 case D_ALLOCDIRECT:
3208 case D_ALLOCINDIR:
3209 continue;
3210
3211 default:
3212 panic("handle_disk_io_initiation: Unexpected type %s",
3213 TYPENAME(wk->wk_type));
3214 /* NOTREACHED */
3215 }
3216 }
59a647b1 3217 rel_mplock();
984263bc
MD
3218}
3219
3220/*
3221 * Called from within the procedure above to deal with unsatisfied
3222 * allocation dependencies in a directory. The buffer must be locked,
3223 * thus, no I/O completion operations can occur while we are
3224 * manipulating its associated dependencies.
3225 */
3226static void
3fcb1ab8 3227initiate_write_filepage(struct pagedep *pagedep, struct buf *bp)
984263bc
MD
3228{
3229 struct diradd *dap;
3230 struct direct *ep;
3231 int i;
3232
3233 if (pagedep->pd_state & IOSTARTED) {
3234 /*
3235 * This can only happen if there is a driver that does not
3236 * understand chaining. Here biodone will reissue the call
3237 * to strategy for the incomplete buffers.
3238 */
086c1d7e 3239 kprintf("initiate_write_filepage: already started\n");
984263bc
MD
3240 return;
3241 }
3242 pagedep->pd_state |= IOSTARTED;
3243 ACQUIRE_LOCK(&lk);
3244 for (i = 0; i < DAHASHSZ; i++) {
3245 LIST_FOREACH(dap, &pagedep->pd_diraddhd[i], da_pdlist) {
3246 ep = (struct direct *)
3247 ((char *)bp->b_data + dap->da_offset);
3248 if (ep->d_ino != dap->da_newinum) {
3249 FREE_LOCK(&lk);
f91a71dd 3250 panic("%s: dir inum %d != new %"PRId64,
984263bc
MD
3251 "initiate_write_filepage",
3252 ep->d_ino, dap->da_newinum);
3253 }
3254 if (dap->da_state & DIRCHG)
3255 ep->d_ino = dap->da_previous->dm_oldinum;
3256 else
3257 ep->d_ino = 0;
3258 dap->da_state &= ~ATTACHED;
3259 dap->da_state |= UNDONE;
3260 }
3261 }
3262 FREE_LOCK(&lk);
3263}
3264
3265/*
3266 * Called from within the procedure above to deal with unsatisfied
3267 * allocation dependencies in an inodeblock. The buffer must be
3268 * locked, thus, no I/O completion operations can occur while we
3269 * are manipulating its associated dependencies.
3fcb1ab8
SW
3270 *
3271 * Parameters:
3272 * bp: The inode block
984263bc
MD
3273 */
3274static void
3fcb1ab8 3275initiate_write_inodeblock(struct inodedep *inodedep, struct buf *bp)
984263bc
MD
3276{
3277 struct allocdirect *adp, *lastadp;
50e58362
MD
3278 struct ufs1_dinode *dp;
3279 struct ufs1_dinode *sip;
984263bc
MD
3280 struct fs *fs;
3281 ufs_lbn_t prevlbn = 0;
3282 int i, deplist;
3283
3284 if (inodedep->id_state & IOSTARTED)
3285 panic("initiate_write_inodeblock: already started");
3286 inodedep->id_state |= IOSTARTED;
3287 fs = inodedep->id_fs;
50e58362 3288 dp = (struct ufs1_dinode *)bp->b_data +
984263bc
MD
3289 ino_to_fsbo(fs, inodedep->id_ino);
3290 /*
3291 * If the bitmap is not yet written, then the allocated
3292 * inode cannot be written to disk.
3293 */
3294 if ((inodedep->id_state & DEPCOMPLETE) == 0) {
3295 if (inodedep->id_savedino != NULL)
3296 panic("initiate_write_inodeblock: already doing I/O");
50e58362
MD
3297 MALLOC(sip, struct ufs1_dinode *,
3298 sizeof(struct ufs1_dinode), M_INODEDEP, M_SOFTDEP_FLAGS);
83b29fff 3299 inodedep->id_savedino = sip;
984263bc 3300 *inodedep->id_savedino = *dp;
50e58362 3301 bzero((caddr_t)dp, sizeof(struct ufs1_dinode));
83b29fff 3302 dp->di_gen = inodedep->id_savedino->di_gen;
984263bc
MD
3303 return;
3304 }
3305 /*
3306 * If no dependencies, then there is nothing to roll back.
3307 */
3308 inodedep->id_savedsize = dp->di_size;
3309 if (TAILQ_FIRST(&inodedep->id_inoupdt) == NULL)
3310 return;
3311 /*
3312 * Set the dependencies to busy.
3313 */
3314 ACQUIRE_LOCK(&lk);
3315 for (deplist = 0, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3316 adp = TAILQ_NEXT(adp, ad_next)) {
3317#ifdef DIAGNOSTIC
3318 if (deplist != 0 && prevlbn >= adp->ad_lbn) {
3319 FREE_LOCK(&lk);
3320 panic("softdep_write_inodeblock: lbn order");
3321 }
3322 prevlbn = adp->ad_lbn;
3323 if (adp->ad_lbn < NDADDR &&
3324 dp->di_db[adp->ad_lbn] != adp->ad_newblkno) {
3325 FREE_LOCK(&lk);
3326 panic("%s: direct pointer #%ld mismatch %d != %d",
3327 "softdep_write_inodeblock", adp->ad_lbn,
3328 dp->di_db[adp->ad_lbn], adp->ad_newblkno);
3329 }
3330 if (adp->ad_lbn >= NDADDR &&
3331 dp->di_ib[adp->ad_lbn - NDADDR] != adp->ad_newblkno) {
3332 FREE_LOCK(&lk);
3333 panic("%s: indirect pointer #%ld mismatch %d != %d",
3334 "softdep_write_inodeblock", adp->ad_lbn - NDADDR,
3335 dp->di_ib[adp->ad_lbn - NDADDR], adp->ad_newblkno);
3336 }
3337 deplist |= 1 << adp->ad_lbn;
3338 if ((adp->ad_state & ATTACHED) == 0) {
3339 FREE_LOCK(&lk);
3340 panic("softdep_write_inodeblock: Unknown state 0x%x",
3341 adp->ad_state);
3342 }
3343#endif /* DIAGNOSTIC */
3344 adp->ad_state &= ~ATTACHED;
3345 adp->ad_state |= UNDONE;
3346 }
3347 /*
3348 * The on-disk inode cannot claim to be any larger than the last
3349 * fragment that has been written. Otherwise, the on-disk inode
3350 * might have fragments that were not the last block in the file
3351 * which would corrupt the filesystem.
3352 */
3353 for (lastadp = NULL, adp = TAILQ_FIRST(&inodedep->id_inoupdt); adp;
3354 lastadp = adp, adp = TAILQ_NEXT(adp, ad_next)) {
3355 if (adp->ad_lbn >= NDADDR)
3356 break;
3357 dp->di_db[adp->ad_lbn] = adp->ad_oldblkno;
3358 /* keep going until hitting a rollback to a frag */
3359 if (adp->ad_oldsize == 0 || adp->ad_oldsize == fs->fs_bsize)
3360 continue;
3361 dp->di_size = fs->fs_bsize * adp->ad_lbn + adp->ad_oldsize;
3362 for (i = adp->ad_lbn + 1; i < NDADDR; i++) {
3363#ifdef DIAGNOSTIC
3364 if (dp->di_db[i] != 0 && (deplist & (1 << i)) == 0) {
3365 FREE_LOCK(&lk);
3366 panic("softdep_write_inodeblock: lost dep1");
3367 }
3368#endif /* DIAGNOSTIC */
3369 dp->di_db[i] = 0;
3370 }
3371 for (i = 0; i < NIADDR; i++) {
3372#ifdef DIAGNOSTIC
3373 if (dp->di_ib[i] != 0 &&
3374 (deplist & ((1 << NDADDR) << i)) == 0) {
3375 FREE_LOCK(&lk);
3376 panic("softdep_write_inodeblock: lost dep2");
3377 }
3378#endif /* DIAGNOSTIC */
3379 dp->di_ib[i] = 0;
3380 }
3381 FREE_LOCK(&lk);
3382 return;
3383 }
3384 /*
3385 * If we have zero'ed out the last allocated block of the file,
3386 * roll back the size to the last currently allocated block.
3387 * We know that this last allocated block is a full-sized as
3388 * we already checked for fragments in the loop above.
3389 */
3390 if (lastadp != NULL &&
3391 dp->di_size <= (lastadp->ad_lbn + 1) * fs->fs_bsize) {
3392 for (i = lastadp->ad_lbn; i >= 0; i--)
3393 if (dp->di_db[i] != 0)
3394 break;
3395 dp->di_size = (i + 1) * fs->fs_bsize;
3396 }
3397 /*
3398 * The only dependencies are for indirect blocks.
3399 *
3400 * The file size for indirect block additions is not guaranteed.
3401 * Such a guarantee would be non-trivial to achieve. The conventional
3402 * synchronous write implementation also does not make this guarantee.
3403 * Fsck should catch and fix discrepancies. Arguably, the file size
3404 * can be over-estimated without destroying integrity when the file
3405 * moves into the indirect blocks (i.e., is large). If we want to
3406 * postpone fsck, we are stuck with this argument.
3407 */
3408 for (; adp; adp = TAILQ_NEXT(adp, ad_next))
3409 dp->di_ib[adp->ad_lbn - NDADDR] = 0;
3410 FREE_LOCK(&lk);
3411}
3412
3413/*
3414 * This routine is called during the completion interrupt
3415 * service routine for a disk write (from the procedure called
f719c866 3416 * by the device driver to inform the filesystem caches of
984263bc
MD
3417 * a request completion). It should be called early in this
3418 * procedure, before the block is made available to other
3419 * processes or other routines are called.
3fcb1ab8 3420 *
59a647b1
MD
3421 * bioops callback - hold io_token
3422 *
3fcb1ab8
SW
3423 * Parameters:
3424 * bp: describes the completed disk write
984263bc
MD
3425 */
3426static void
3fcb1ab8 3427softdep_disk_write_complete(struct buf *bp)
984263bc
MD
3428{
3429 struct worklist *wk;
3430 struct workhead reattach;
3431 struct newblk *newblk;
3432 struct allocindir *aip;
3433 struct allocdirect *adp;
3434 struct indirdep *indirdep;
3435 struct inodedep *inodedep;
3436 struct bmsafemap *bmsafemap;
3437
59a647b1 3438 get_mplock();
984263bc 3439#ifdef DEBUG
dadab5e9 3440 if (lk.lkt_held != NOHOLDER)
984263bc 3441 panic("softdep_disk_write_complete: lock is held");
dadab5e9 3442 lk.lkt_held = SPECIAL_FLAG;
984263bc
MD
3443#endif
3444 LIST_INIT(&reattach);
3445 while ((wk = LIST_FIRST(&bp->b_dep)) != NULL) {
3446 WORKLIST_REMOVE(wk);
3447 switch (wk->wk_type) {
3448
3449 case D_PAGEDEP:
3450 if (handle_written_filepage(WK_PAGEDEP(wk), bp))
3451 WORKLIST_INSERT(&reattach, wk);
3452 continue;
3453
3454 case D_INODEDEP:
3455 if (handle_written_inodeblock(WK_INODEDEP(wk), bp))
3456 WORKLIST_INSERT(&reattach, wk);
3457 continue;
3458
3459 case D_BMSAFEMAP:
3460 bmsafemap = WK_BMSAFEMAP(wk);
3461 while ((newblk = LIST_FIRST(&bmsafemap->sm_newblkhd))) {
3462 newblk->nb_state |= DEPCOMPLETE;
3463 newblk->nb_bmsafemap = NULL;
3464 LIST_REMOVE(newblk, nb_deps);
3465 }
3466 while ((adp =
3467 LIST_FIRST(&bmsafemap->sm_allocdirecthd))) {
3468 adp->ad_state |= DEPCOMPLETE;
3469 adp->ad_buf = NULL;
3470 LIST_REMOVE(adp, ad_deps);
3471 handle_allocdirect_partdone(adp);
3472 }
3473 while ((aip =
3474 LIST_FIRST(&bmsafemap->sm_allocindirhd))) {
3475 aip->ai_state |= DEPCOMPLETE;
3476 aip->ai_buf = NULL;
3477 LIST_REMOVE(aip, ai_deps);
3478 handle_allocindir_partdone(aip);
3479 }
3480 while ((inodedep =
3481 LIST_FIRST(&bmsafemap->sm_inodedephd)) != NULL) {
3482 inodedep->id_state |= DEPCOMPLETE;
3483 LIST_REMOVE(inodedep, id_deps);
3484 inodedep->id_buf = NULL;
3485 }
3486 WORKITEM_FREE(bmsafemap, D_BMSAFEMAP);
3487 continue;
3488
3489 case D_MKDIR:
3490 handle_written_mkdir(WK_MKDIR(wk), MKDIR_BODY);
3491 continue;
3492
3493 case D_ALLOCDIRECT:
3494 adp = WK_ALLOCDIRECT(wk);
3495 adp->ad_state |= COMPLETE;
3496 handle_allocdirect_partdone(adp);
3497 continue;
3498
3499 case D_ALLOCINDIR:
3500 aip = WK_ALLOCINDIR(wk);
3501 aip->ai_state |= COMPLETE;
3502 handle_allocindir_partdone(aip);
3503 continue;
3504
3505 case D_INDIRDEP:
3506 indirdep = WK_INDIRDEP(wk);
3507 if (indirdep->ir_state & GOINGAWAY) {
dadab5e9 3508 lk.lkt_held = NOHOLDER;
984263bc
MD
3509 panic("disk_write_complete: indirdep gone");
3510 }
3511 bcopy(indirdep->ir_saveddata, bp->b_data, bp->b_bcount);
3512 FREE(indirdep->ir_saveddata, M_INDIRDEP);
3513 indirdep->ir_saveddata = 0;
3514 indirdep->ir_state &= ~UNDONE;
3515 indirdep->ir_state |= ATTACHED;
3516 while ((aip = LIST_FIRST(&indirdep->ir_donehd)) != 0) {
3517 handle_allocindir_partdone(aip);
3518 if (aip == LIST_FIRST(&indirdep->ir_donehd)) {
dadab5e9 3519 lk.lkt_held = NOHOLDER;
984263bc
MD
3520 panic("disk_write_complete: not gone");
3521 }
3522 }
3523 WORKLIST_INSERT(&reattach, wk);
3524 if ((bp->b_flags & B_DELWRI) == 0)
3525 stat_indir_blk_ptrs++;
3526 bdirty(bp);
3527 continue;
3528
3529 default:
dadab5e9 3530 lk.lkt_held = NOHOLDER;
984263bc
MD
3531 panic("handle_disk_write_complete: Unknown type %s",
3532 TYPENAME(wk->wk_type));
3533 /* NOTREACHED */
3534 }
3535 }
3536 /*
3537 * Reattach any requests that must be redone.
3538 */
3539 while ((wk = LIST_FIRST(&reattach)) != NULL) {
3540 WORKLIST_REMOVE(wk);
408357d8 3541 WORKLIST_INSERT_BP(bp, wk);
984263bc
MD
3542 }
3543#ifdef DEBUG
dadab5e9 3544 if (lk.lkt_held != SPECIAL_FLAG)
984263bc 3545 panic("softdep_disk_write_complete: lock lost");
dadab5e9 3546 lk.lkt_held = NOHOLDER;
984263bc 3547#endif
59a647b1 3548 rel_mplock();
984263bc
MD
3549}
3550
3551/*
3552 * Called from within softdep_disk_write_complete above. Note that
3553 * this routine is always called from interrupt level with further
3554 * splbio interrupts blocked.
3fcb1ab8
SW
3555 *
3556 * Parameters:
3557 * adp: the completed allocdirect
984263bc
MD
3558 */
3559static void
3fcb1ab8 3560handle_allocdirect_partdone(struct allocdirect *adp)
984263bc
MD
3561{
3562 struct allocdirect *listadp;
3563 struct inodedep *inodedep;
3564 long bsize;
3565
3566 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3567 return;
3568 if (adp->ad_buf != NULL) {
dadab5e9 3569 lk.lkt_held = NOHOLDER;
984263bc
MD
3570 panic("handle_allocdirect_partdone: dangling dep");
3571 }
3572 /*
3573 * The on-disk inode cannot claim to be any larger than the last
3574 * fragment that has been written. Otherwise, the on-disk inode
3575 * might have fragments that were not the last block in the file
3576 * which would corrupt the filesystem. Thus, we cannot free any
3577 * allocdirects after one whose ad_oldblkno claims a fragment as
3578 * these blocks must be rolled back to zero before writing the inode.
3579 * We check the currently active set of allocdirects in id_inoupdt.
3580 */
3581 inodedep = adp->ad_inodedep;
3582 bsize = inodedep->id_fs->fs_bsize;
3583 TAILQ_FOREACH(listadp, &inodedep->id_inoupdt, ad_next) {
3584 /* found our block */
3585 if (listadp == adp)
3586 break;
3587 /* continue if ad_oldlbn is not a fragment */
3588 if (listadp->ad_oldsize == 0 ||
3589 listadp->ad_oldsize == bsize)
3590 continue;
3591 /* hit a fragment */
3592 return;
3593 }
3594 /*
3595 * If we have reached the end of the current list without
3596 * finding the just finished dependency, then it must be
3597 * on the future dependency list. Future dependencies cannot
3598 * be freed until they are moved to the current list.
3599 */
3600 if (listadp == NULL) {
3601#ifdef DEBUG
3602 TAILQ_FOREACH(listadp, &inodedep->id_newinoupdt, ad_next)
3603 /* found our block */
3604 if (listadp == adp)
3605 break;
3606 if (listadp == NULL) {
dadab5e9 3607 lk.lkt_held = NOHOLDER;
984263bc
MD
3608 panic("handle_allocdirect_partdone: lost dep");
3609 }
3610#endif /* DEBUG */
3611 return;
3612 }
3613 /*
3614 * If we have found the just finished dependency, then free
3615 * it along with anything that follows it that is complete.
3616 */
3617 for (; adp; adp = listadp) {
3618 listadp = TAILQ_NEXT(adp, ad_next);
3619 if ((adp->ad_state & ALLCOMPLETE) != ALLCOMPLETE)
3620 return;
3621 free_allocdirect(&inodedep->id_inoupdt, adp, 1);
3622 }
3623}
3624
3625/*
3626 * Called from within softdep_disk_write_complete above. Note that
3627 * this routine is always called from interrupt level with further
3628 * splbio interrupts blocked.
3fcb1ab8
SW
3629 *
3630 * Parameters:
3631 * aip: the completed allocindir
984263bc
MD
3632 */
3633static void