2
* Copyright 1998, 2000 Marshall Kirk McKusick. All Rights Reserved.
4
* The soft updates code is derived from the appendix of a University
5
* of Michigan technical report (Gregory R. Ganger and Yale N. Patt,
6
* "Soft Updates: A Solution to the Metadata Update Problem in File
7
* Systems", CSE-TR-254-95, August 1995).
9
* Further information about soft updates can be obtained from:
11
* Marshall Kirk McKusick http://www.mckusick.com/softdep/
12
* 1614 Oxford Street mckusick@mckusick.com
13
* Berkeley, CA 94709-1608 +1-510-843-9542
16
* Redistribution and use in source and binary forms, with or without
17
* modification, are permitted provided that the following conditions
20
* 1. Redistributions of source code must retain the above copyright
21
* notice, this list of conditions and the following disclaimer.
22
* 2. Redistributions in binary form must reproduce the above copyright
23
* notice, this list of conditions and the following disclaimer in the
24
* documentation and/or other materials provided with the distribution.
26
* THIS SOFTWARE IS PROVIDED BY MARSHALL KIRK MCKUSICK ``AS IS'' AND ANY
27
* EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
28
* WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
29
* DISCLAIMED. IN NO EVENT SHALL MARSHALL KIRK MCKUSICK BE LIABLE FOR
30
* ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38
* @(#)softdep.h 9.7 (McKusick) 6/21/00
39
* $FreeBSD: src/sys/ufs/ffs/softdep.h,v 1.19.12.1 2010/02/10 00:26:20 kensmith Exp $
42
#include <sys/queue.h>
45
* Allocation dependencies are handled with undo/redo on the in-memory
46
* copy of the data. A particular data dependency is eliminated when
47
* it is ALLCOMPLETE: that is ATTACHED, DEPCOMPLETE, and COMPLETE.
49
* ATTACHED means that the data is not currently being written to
50
* disk. UNDONE means that the data has been rolled back to a safe
51
* state for writing to the disk. When the I/O completes, the data is
52
* restored to its current form and the state reverts to ATTACHED.
53
* The data must be locked throughout the rollback, I/O, and roll
54
* forward so that the rolled back information is never visible to
55
* user processes. The COMPLETE flag indicates that the item has been
56
* written. For example, a dependency that requires that an inode be
57
* written will be marked COMPLETE after the inode has been written
58
* to disk. The DEPCOMPLETE flag indicates the completion of any other
59
* dependencies such as the writing of a cylinder group map has been
60
* completed. A dependency structure may be freed only when both it
61
* and its dependencies have completed and any rollbacks that are in
62
* progress have finished as indicated by the set of ALLCOMPLETE flags
63
* all being set. The two MKDIR flags indicate additional dependencies
64
* that must be done when creating a new directory. MKDIR_BODY is
65
* cleared when the directory data block containing the "." and ".."
66
* entries has been written. MKDIR_PARENT is cleared when the parent
67
* inode with the increased link count for ".." has been written. When
68
* both MKDIR flags have been cleared, the DEPCOMPLETE flag is set to
69
* indicate that the directory dependencies have been completed. The
70
* writing of the directory inode itself sets the COMPLETE flag which
71
* then allows the directory entry for the new directory to be written
72
* to disk. The RMDIR flag marks a dirrem structure as representing
73
* the removal of a directory rather than a file. When the removal
74
* dependencies are completed, additional work needs to be done
75
* (truncation of the "." and ".." entries, an additional decrement
76
* of the associated inode, and a decrement of the parent inode). The
77
* DIRCHG flag marks a diradd structure as representing the changing
78
* of an existing entry rather than the addition of a new one. When
79
* the update is complete the dirrem associated with the inode for
80
* the old name must be added to the worklist to do the necessary
81
* reference count decrement. The GOINGAWAY flag indicates that the
82
* data structure is frozen from further change until its dependencies
83
* have been completed and its resources freed after which it will be
84
* discarded. The IOSTARTED flag prevents multiple calls to the I/O
85
* start routine from doing multiple rollbacks. The SPACECOUNTED flag
86
* says that the files space has been accounted to the pending free
87
* space count. The NEWBLOCK flag marks pagedep structures that have
88
* just been allocated, so must be claimed by the inode before all
89
* dependencies are complete. The INPROGRESS flag marks worklist
90
* structures that are still on the worklist, but are being considered
91
* for action by some process. The UFS1FMT flag indicates that the
92
* inode being processed is a ufs1 format. The EXTDATA flag indicates
93
* that the allocdirect describes an extended-attributes dependency.
94
* The ONWORKLIST flag shows whether the structure is currently linked
97
#define ATTACHED 0x0001
99
#define COMPLETE 0x0004
100
#define DEPCOMPLETE 0x0008
101
#define MKDIR_PARENT 0x0010 /* diradd & mkdir only */
102
#define MKDIR_BODY 0x0020 /* diradd & mkdir only */
103
#define RMDIR 0x0040 /* dirrem only */
104
#define DIRCHG 0x0080 /* diradd & dirrem only */
105
#define GOINGAWAY 0x0100 /* indirdep only */
106
#define IOSTARTED 0x0200 /* inodedep & pagedep only */
107
#define SPACECOUNTED 0x0400 /* inodedep only */
108
#define NEWBLOCK 0x0800 /* pagedep only */
109
#define INPROGRESS 0x1000 /* dirrem, freeblks, freefrag, freefile only */
110
#define UFS1FMT 0x2000 /* indirdep only */
111
#define EXTDATA 0x4000 /* allocdirect only */
112
#define ONWORKLIST 0x8000
114
#define ALLCOMPLETE (ATTACHED | COMPLETE | DEPCOMPLETE)
117
* The workitem queue.
119
* It is sometimes useful and/or necessary to clean up certain dependencies
120
* in the background rather than during execution of an application process
121
* or interrupt service routine. To realize this, we append dependency
122
* structures corresponding to such tasks to a "workitem" queue. In a soft
123
* updates implementation, most pending workitems should not wait for more
124
* than a couple of seconds, so the filesystem syncer process awakens once
125
* per second to process the items on the queue.
128
/* LIST_HEAD(workhead, worklist); -- declared in buf.h */
131
* Each request can be linked onto a work queue through its worklist structure.
132
* To avoid the need for a pointer to the structure itself, this structure
133
* MUST be declared FIRST in each type in which it appears! If more than one
134
* worklist is needed in the structure, then a wk_data field must be added
135
* and the macros below changed to use it.
138
struct mount *wk_mp; /* Mount we live in */
139
LIST_ENTRY(worklist) wk_list; /* list of work requests */
140
unsigned short wk_type; /* type of request */
141
unsigned short wk_state; /* state flags */
143
#define WK_DATA(wk) ((void *)(wk))
144
#define WK_PAGEDEP(wk) ((struct pagedep *)(wk))
145
#define WK_INODEDEP(wk) ((struct inodedep *)(wk))
146
#define WK_BMSAFEMAP(wk) ((struct bmsafemap *)(wk))
147
#define WK_ALLOCDIRECT(wk) ((struct allocdirect *)(wk))
148
#define WK_INDIRDEP(wk) ((struct indirdep *)(wk))
149
#define WK_ALLOCINDIR(wk) ((struct allocindir *)(wk))
150
#define WK_FREEFRAG(wk) ((struct freefrag *)(wk))
151
#define WK_FREEBLKS(wk) ((struct freeblks *)(wk))
152
#define WK_FREEFILE(wk) ((struct freefile *)(wk))
153
#define WK_DIRADD(wk) ((struct diradd *)(wk))
154
#define WK_MKDIR(wk) ((struct mkdir *)(wk))
155
#define WK_DIRREM(wk) ((struct dirrem *)(wk))
156
#define WK_NEWDIRBLK(wk) ((struct newdirblk *)(wk))
159
* Various types of lists
161
LIST_HEAD(dirremhd, dirrem);
162
LIST_HEAD(diraddhd, diradd);
163
LIST_HEAD(newblkhd, newblk);
164
LIST_HEAD(inodedephd, inodedep);
165
LIST_HEAD(allocindirhd, allocindir);
166
LIST_HEAD(allocdirecthd, allocdirect);
167
TAILQ_HEAD(allocdirectlst, allocdirect);
170
* The "pagedep" structure tracks the various dependencies related to
171
* a particular directory page. If a directory page has any dependencies,
172
* it will have a pagedep linked to its associated buffer. The
173
* pd_dirremhd list holds the list of dirrem requests which decrement
174
* inode reference counts. These requests are processed after the
175
* directory page with the corresponding zero'ed entries has been
176
* written. The pd_diraddhd list maintains the list of diradd requests
177
* which cannot be committed until their corresponding inode has been
178
* written to disk. Because a directory may have many new entries
179
* being created, several lists are maintained hashed on bits of the
180
* offset of the entry into the directory page to keep the lists from
181
* getting too long. Once a new directory entry has been cleared to
182
* be written, it is moved to the pd_pendinghd list. After the new
183
* entry has been written to disk it is removed from the pd_pendinghd
184
* list, any removed operations are done, and the dependency structure
188
#define DIRADDHASH(offset) (((offset) >> 2) % DAHASHSZ)
190
struct worklist pd_list; /* page buffer */
191
# define pd_state pd_list.wk_state /* check for multiple I/O starts */
192
LIST_ENTRY(pagedep) pd_hash; /* hashed lookup */
193
ino_t pd_ino; /* associated file */
194
ufs_lbn_t pd_lbn; /* block within file */
195
struct dirremhd pd_dirremhd; /* dirrem's waiting for page */
196
struct diraddhd pd_diraddhd[DAHASHSZ]; /* diradd dir entry updates */
197
struct diraddhd pd_pendinghd; /* directory entries awaiting write */
201
* The "inodedep" structure tracks the set of dependencies associated
202
* with an inode. One task that it must manage is delayed operations
203
* (i.e., work requests that must be held until the inodedep's associated
204
* inode has been written to disk). Getting an inode from its incore
205
* state to the disk requires two steps to be taken by the filesystem
206
* in this order: first the inode must be copied to its disk buffer by
207
* the VOP_UPDATE operation; second the inode's buffer must be written
208
* to disk. To ensure that both operations have happened in the required
209
* order, the inodedep maintains two lists. Delayed operations are
210
* placed on the id_inowait list. When the VOP_UPDATE is done, all
211
* operations on the id_inowait list are moved to the id_bufwait list.
212
* When the buffer is written, the items on the id_bufwait list can be
213
* safely moved to the work queue to be processed. A second task of the
214
* inodedep structure is to track the status of block allocation within
215
* the inode. Each block that is allocated is represented by an
216
* "allocdirect" structure (see below). It is linked onto the id_newinoupdt
217
* list until both its contents and its allocation in the cylinder
218
* group map have been written to disk. Once these dependencies have been
219
* satisfied, it is removed from the id_newinoupdt list and any followup
220
* actions such as releasing the previous block or fragment are placed
221
* on the id_inowait list. When an inode is updated (a VOP_UPDATE is
222
* done), the "inodedep" structure is linked onto the buffer through
223
* its worklist. Thus, it will be notified when the buffer is about
224
* to be written and when it is done. At the update time, all the
225
* elements on the id_newinoupdt list are moved to the id_inoupdt list
226
* since those changes are now relevant to the copy of the inode in the
227
* buffer. Also at update time, the tasks on the id_inowait list are
228
* moved to the id_bufwait list so that they will be executed when
229
* the updated inode has been written to disk. When the buffer containing
230
* the inode is written to disk, any updates listed on the id_inoupdt
231
* list are rolled back as they are not yet safe. Following the write,
232
* the changes are once again rolled forward and any actions on the
233
* id_bufwait list are processed (since those actions are now safe).
234
* The entries on the id_inoupdt and id_newinoupdt lists must be kept
235
* sorted by logical block number to speed the calculation of the size
236
* of the rolled back inode (see explanation in initiate_write_inodeblock).
237
* When a directory entry is created, it is represented by a diradd.
238
* The diradd is added to the id_inowait list as it cannot be safely
239
* written to disk until the inode that it represents is on disk. After
240
* the inode is written, the id_bufwait list is processed and the diradd
241
* entries are moved to the id_pendinghd list where they remain until
242
* the directory block containing the name has been written to disk.
243
* The purpose of keeping the entries on the id_pendinghd list is so that
244
* the softdep_fsync function can find and push the inode's directory
245
* name(s) as part of the fsync operation for that file.
248
struct worklist id_list; /* buffer holding inode block */
249
# define id_state id_list.wk_state /* inode dependency state */
250
LIST_ENTRY(inodedep) id_hash; /* hashed lookup */
251
struct fs *id_fs; /* associated filesystem */
252
ino_t id_ino; /* dependent inode */
253
nlink_t id_nlinkdelta; /* saved effective link count */
254
LIST_ENTRY(inodedep) id_deps; /* bmsafemap's list of inodedep's */
255
struct buf *id_buf; /* related bmsafemap (if pending) */
256
long id_savedextsize; /* ext size saved during rollback */
257
off_t id_savedsize; /* file size saved during rollback */
258
struct workhead id_pendinghd; /* entries awaiting directory write */
259
struct workhead id_bufwait; /* operations after inode written */
260
struct workhead id_inowait; /* operations waiting inode update */
261
struct allocdirectlst id_inoupdt; /* updates before inode written */
262
struct allocdirectlst id_newinoupdt; /* updates when inode written */
263
struct allocdirectlst id_extupdt; /* extdata updates pre-inode write */
264
struct allocdirectlst id_newextupdt; /* extdata updates at ino write */
266
struct ufs1_dinode *idu_savedino1; /* saved ufs1_dinode contents */
267
struct ufs2_dinode *idu_savedino2; /* saved ufs2_dinode contents */
270
#define id_savedino1 id_un.idu_savedino1
271
#define id_savedino2 id_un.idu_savedino2
274
* A "newblk" structure is attached to a bmsafemap structure when a block
275
* or fragment is allocated from a cylinder group. Its state is set to
276
* DEPCOMPLETE when its cylinder group map is written. It is consumed by
277
* an associated allocdirect or allocindir allocation which will attach
278
* themselves to the bmsafemap structure if the newblk's DEPCOMPLETE flag
279
* is not set (i.e., its cylinder group map has not been written).
282
LIST_ENTRY(newblk) nb_hash; /* hashed lookup */
283
struct fs *nb_fs; /* associated filesystem */
284
int nb_state; /* state of bitmap dependency */
285
ufs2_daddr_t nb_newblkno; /* allocated block number */
286
LIST_ENTRY(newblk) nb_deps; /* bmsafemap's list of newblk's */
287
struct bmsafemap *nb_bmsafemap; /* associated bmsafemap */
291
* A "bmsafemap" structure maintains a list of dependency structures
292
* that depend on the update of a particular cylinder group map.
293
* It has lists for newblks, allocdirects, allocindirs, and inodedeps.
294
* It is attached to the buffer of a cylinder group block when any of
295
* these things are allocated from the cylinder group. It is freed
296
* after the cylinder group map is written and the state of its
297
* dependencies are updated with DEPCOMPLETE to indicate that it has
301
struct worklist sm_list; /* cylgrp buffer */
302
struct buf *sm_buf; /* associated buffer */
303
struct allocdirecthd sm_allocdirecthd; /* allocdirect deps */
304
struct allocindirhd sm_allocindirhd; /* allocindir deps */
305
struct inodedephd sm_inodedephd; /* inodedep deps */
306
struct newblkhd sm_newblkhd; /* newblk deps */
310
* An "allocdirect" structure is attached to an "inodedep" when a new block
311
* or fragment is allocated and pointed to by the inode described by
312
* "inodedep". The worklist is linked to the buffer that holds the block.
313
* When the block is first allocated, it is linked to the bmsafemap
314
* structure associated with the buffer holding the cylinder group map
315
* from which it was allocated. When the cylinder group map is written
316
* to disk, ad_state has the DEPCOMPLETE flag set. When the block itself
317
* is written, the COMPLETE flag is set. Once both the cylinder group map
318
* and the data itself have been written, it is safe to write the inode
319
* that claims the block. If there was a previous fragment that had been
320
* allocated before the file was increased in size, the old fragment may
321
* be freed once the inode claiming the new block is written to disk.
322
* This ad_fragfree request is attached to the id_inowait list of the
323
* associated inodedep (pointed to by ad_inodedep) for processing after
324
* the inode is written. When a block is allocated to a directory, an
325
* fsync of a file whose name is within that block must ensure not only
326
* that the block containing the file name has been written, but also
327
* that the on-disk inode references that block. When a new directory
328
* block is created, we allocate a newdirblk structure which is linked
329
* to the associated allocdirect (on its ad_newdirblk list). When the
330
* allocdirect has been satisfied, the newdirblk structure is moved to
331
* the inodedep id_bufwait list of its directory to await the inode
332
* being written. When the inode is written, the directory entries are
333
* fully committed and can be deleted from their pagedep->id_pendinghd
334
* and inodedep->id_pendinghd lists.
337
struct worklist ad_list; /* buffer holding block */
338
# define ad_state ad_list.wk_state /* block pointer state */
339
TAILQ_ENTRY(allocdirect) ad_next; /* inodedep's list of allocdirect's */
340
ufs_lbn_t ad_lbn; /* block within file */
341
ufs2_daddr_t ad_newblkno; /* new value of block pointer */
342
ufs2_daddr_t ad_oldblkno; /* old value of block pointer */
343
long ad_newsize; /* size of new block */
344
long ad_oldsize; /* size of old block */
345
LIST_ENTRY(allocdirect) ad_deps; /* bmsafemap's list of allocdirect's */
346
struct buf *ad_buf; /* cylgrp buffer (if pending) */
347
struct inodedep *ad_inodedep; /* associated inodedep */
348
struct freefrag *ad_freefrag; /* fragment to be freed (if any) */
349
struct workhead ad_newdirblk; /* dir block to notify when written */
353
* A single "indirdep" structure manages all allocation dependencies for
354
* pointers in an indirect block. The up-to-date state of the indirect
355
* block is stored in ir_savedata. The set of pointers that may be safely
356
* written to the disk is stored in ir_safecopy. The state field is used
357
* only to track whether the buffer is currently being written (in which
358
* case it is not safe to update ir_safecopy). Ir_deplisthd contains the
359
* list of allocindir structures, one for each block that needs to be
360
* written to disk. Once the block and its bitmap allocation have been
361
* written the safecopy can be updated to reflect the allocation and the
362
* allocindir structure freed. If ir_state indicates that an I/O on the
363
* indirect block is in progress when ir_safecopy is to be updated, the
364
* update is deferred by placing the allocindir on the ir_donehd list.
365
* When the I/O on the indirect block completes, the entries on the
366
* ir_donehd list are processed by updating their corresponding ir_safecopy
367
* pointers and then freeing the allocindir structure.
370
struct worklist ir_list; /* buffer holding indirect block */
371
# define ir_state ir_list.wk_state /* indirect block pointer state */
372
caddr_t ir_saveddata; /* buffer cache contents */
373
struct buf *ir_savebp; /* buffer holding safe copy */
374
struct allocindirhd ir_donehd; /* done waiting to update safecopy */
375
struct allocindirhd ir_deplisthd; /* allocindir deps for this block */
379
* An "allocindir" structure is attached to an "indirdep" when a new block
380
* is allocated and pointed to by the indirect block described by the
381
* "indirdep". The worklist is linked to the buffer that holds the new block.
382
* When the block is first allocated, it is linked to the bmsafemap
383
* structure associated with the buffer holding the cylinder group map
384
* from which it was allocated. When the cylinder group map is written
385
* to disk, ai_state has the DEPCOMPLETE flag set. When the block itself
386
* is written, the COMPLETE flag is set. Once both the cylinder group map
387
* and the data itself have been written, it is safe to write the entry in
388
* the indirect block that claims the block; the "allocindir" dependency
389
* can then be freed as it is no longer applicable.
392
struct worklist ai_list; /* buffer holding indirect block */
393
# define ai_state ai_list.wk_state /* indirect block pointer state */
394
LIST_ENTRY(allocindir) ai_next; /* indirdep's list of allocindir's */
395
int ai_offset; /* pointer offset in indirect block */
396
ufs2_daddr_t ai_newblkno; /* new block pointer value */
397
ufs2_daddr_t ai_oldblkno; /* old block pointer value */
398
struct freefrag *ai_freefrag; /* block to be freed when complete */
399
struct indirdep *ai_indirdep; /* address of associated indirdep */
400
LIST_ENTRY(allocindir) ai_deps; /* bmsafemap's list of allocindir's */
401
struct buf *ai_buf; /* cylgrp buffer (if pending) */
405
* A "freefrag" structure is attached to an "inodedep" when a previously
406
* allocated fragment is replaced with a larger fragment, rather than extended.
407
* The "freefrag" structure is constructed and attached when the replacement
408
* block is first allocated. It is processed after the inode claiming the
409
* bigger block that replaces it has been written to disk. Note that the
410
* ff_state field is is used to store the uid, so may lose data. However,
411
* the uid is used only in printing an error message, so is not critical.
412
* Keeping it in a short keeps the data structure down to 32 bytes.
415
struct worklist ff_list; /* id_inowait or delayed worklist */
416
# define ff_state ff_list.wk_state /* owning user; should be uid_t */
417
ufs2_daddr_t ff_blkno; /* fragment physical block number */
418
long ff_fragsize; /* size of fragment being deleted */
419
ino_t ff_inum; /* owning inode number */
423
* A "freeblks" structure is attached to an "inodedep" when the
424
* corresponding file's length is reduced to zero. It records all
425
* the information needed to free the blocks of a file after its
426
* zero'ed inode has been written to disk.
429
struct worklist fb_list; /* id_inowait or delayed worklist */
430
# define fb_state fb_list.wk_state /* inode and dirty block state */
431
ino_t fb_previousinum; /* inode of previous owner of blocks */
432
uid_t fb_uid; /* uid of previous owner of blocks */
433
struct vnode *fb_devvp; /* filesystem device vnode */
434
long fb_oldextsize; /* previous ext data size */
435
off_t fb_oldsize; /* previous file size */
436
ufs2_daddr_t fb_chkcnt; /* used to check cnt of blks released */
437
ufs2_daddr_t fb_dblks[NDADDR]; /* direct blk ptrs to deallocate */
438
ufs2_daddr_t fb_iblks[NIADDR]; /* indirect blk ptrs to deallocate */
439
ufs2_daddr_t fb_eblks[NXADDR]; /* indirect blk ptrs to deallocate */
443
* A "freefile" structure is attached to an inode when its
444
* link count is reduced to zero. It marks the inode as free in
445
* the cylinder group map after the zero'ed inode has been written
446
* to disk and any associated blocks and fragments have been freed.
449
struct worklist fx_list; /* id_inowait or delayed worklist */
450
mode_t fx_mode; /* mode of inode */
451
ino_t fx_oldinum; /* inum of the unlinked file */
452
struct vnode *fx_devvp; /* filesystem device vnode */
456
* A "diradd" structure is linked to an "inodedep" id_inowait list when a
457
* new directory entry is allocated that references the inode described
458
* by "inodedep". When the inode itself is written (either the initial
459
* allocation for new inodes or with the increased link count for
460
* existing inodes), the COMPLETE flag is set in da_state. If the entry
461
* is for a newly allocated inode, the "inodedep" structure is associated
462
* with a bmsafemap which prevents the inode from being written to disk
463
* until the cylinder group has been updated. Thus the da_state COMPLETE
464
* flag cannot be set until the inode bitmap dependency has been removed.
465
* When creating a new file, it is safe to write the directory entry that
466
* claims the inode once the referenced inode has been written. Since
467
* writing the inode clears the bitmap dependencies, the DEPCOMPLETE flag
468
* in the diradd can be set unconditionally when creating a file. When
469
* creating a directory, there are two additional dependencies described by
470
* mkdir structures (see their description below). When these dependencies
471
* are resolved the DEPCOMPLETE flag is set in the diradd structure.
472
* If there are multiple links created to the same inode, there will be
473
* a separate diradd structure created for each link. The diradd is
474
* linked onto the pg_diraddhd list of the pagedep for the directory
475
* page that contains the entry. When a directory page is written,
476
* the pg_diraddhd list is traversed to rollback any entries that are
477
* not yet ready to be written to disk. If a directory entry is being
478
* changed (by rename) rather than added, the DIRCHG flag is set and
479
* the da_previous entry points to the entry that will be "removed"
480
* once the new entry has been committed. During rollback, entries
481
* with da_previous are replaced with the previous inode number rather
484
* The overlaying of da_pagedep and da_previous is done to keep the
485
* structure down to 32 bytes in size on a 32-bit machine. If a
486
* da_previous entry is present, the pointer to its pagedep is available
487
* in the associated dirrem entry. If the DIRCHG flag is set, the
488
* da_previous entry is valid; if not set the da_pagedep entry is valid.
489
* The DIRCHG flag never changes; it is set when the structure is created
490
* if appropriate and is never cleared.
493
struct worklist da_list; /* id_inowait or id_pendinghd list */
494
# define da_state da_list.wk_state /* state of the new directory entry */
495
LIST_ENTRY(diradd) da_pdlist; /* pagedep holding directory block */
496
doff_t da_offset; /* offset of new dir entry in dir blk */
497
ino_t da_newinum; /* inode number for the new dir entry */
499
struct dirrem *dau_previous; /* entry being replaced in dir change */
500
struct pagedep *dau_pagedep; /* pagedep dependency for addition */
503
#define da_previous da_un.dau_previous
504
#define da_pagedep da_un.dau_pagedep
507
* Two "mkdir" structures are needed to track the additional dependencies
508
* associated with creating a new directory entry. Normally a directory
509
* addition can be committed as soon as the newly referenced inode has been
510
* written to disk with its increased link count. When a directory is
511
* created there are two additional dependencies: writing the directory
512
* data block containing the "." and ".." entries (MKDIR_BODY) and writing
513
* the parent inode with the increased link count for ".." (MKDIR_PARENT).
514
* These additional dependencies are tracked by two mkdir structures that
515
* reference the associated "diradd" structure. When they have completed,
516
* they set the DEPCOMPLETE flag on the diradd so that it knows that its
517
* extra dependencies have been completed. The md_state field is used only
518
* to identify which type of dependency the mkdir structure is tracking.
519
* It is not used in the mainline code for any purpose other than consistency
520
* checking. All the mkdir structures in the system are linked together on
521
* a list. This list is needed so that a diradd can find its associated
522
* mkdir structures and deallocate them if it is prematurely freed (as for
523
* example if a mkdir is immediately followed by a rmdir of the same directory).
524
* Here, the free of the diradd must traverse the list to find the associated
525
* mkdir structures that reference it. The deletion would be faster if the
526
* diradd structure were simply augmented to have two pointers that referenced
527
* the associated mkdir's. However, this would increase the size of the diradd
528
* structure from 32 to 64-bits to speed a very infrequent operation.
531
struct worklist md_list; /* id_inowait or buffer holding dir */
532
# define md_state md_list.wk_state /* type: MKDIR_PARENT or MKDIR_BODY */
533
struct diradd *md_diradd; /* associated diradd */
534
struct buf *md_buf; /* MKDIR_BODY: buffer holding dir */
535
LIST_ENTRY(mkdir) md_mkdirs; /* list of all mkdirs */
537
LIST_HEAD(mkdirlist, mkdir) mkdirlisthd;
540
* A "dirrem" structure describes an operation to decrement the link
541
* count on an inode. The dirrem structure is attached to the pg_dirremhd
542
* list of the pagedep for the directory page that contains the entry.
543
* It is processed after the directory page with the deleted entry has
544
* been written to disk.
546
* The overlaying of dm_pagedep and dm_dirinum is done to keep the
547
* structure down to 32 bytes in size on a 32-bit machine. It works
548
* because they are never used concurrently.
551
struct worklist dm_list; /* delayed worklist */
552
# define dm_state dm_list.wk_state /* state of the old directory entry */
553
LIST_ENTRY(dirrem) dm_next; /* pagedep's list of dirrem's */
554
ino_t dm_oldinum; /* inum of the removed dir entry */
556
struct pagedep *dmu_pagedep; /* pagedep dependency for remove */
557
ino_t dmu_dirinum; /* parent inode number (for rmdir) */
560
#define dm_pagedep dm_un.dmu_pagedep
561
#define dm_dirinum dm_un.dmu_dirinum
564
* A "newdirblk" structure tracks the progress of a newly allocated
565
* directory block from its creation until it is claimed by its on-disk
566
* inode. When a block is allocated to a directory, an fsync of a file
567
* whose name is within that block must ensure not only that the block
568
* containing the file name has been written, but also that the on-disk
569
* inode references that block. When a new directory block is created,
570
* we allocate a newdirblk structure which is linked to the associated
571
* allocdirect (on its ad_newdirblk list). When the allocdirect has been
572
* satisfied, the newdirblk structure is moved to the inodedep id_bufwait
573
* list of its directory to await the inode being written. When the inode
574
* is written, the directory entries are fully committed and can be
575
* deleted from their pagedep->id_pendinghd and inodedep->id_pendinghd
576
* lists. Note that we could track directory blocks allocated to indirect
577
* blocks using a similar scheme with the allocindir structures. Rather
578
* than adding this level of complexity, we simply write those newly
579
* allocated indirect blocks synchronously as such allocations are rare.
582
struct worklist db_list; /* id_inowait or pg_newdirblk */
583
# define db_state db_list.wk_state /* unused */
584
struct pagedep *db_pagedep; /* associated pagedep */