| blob | e459c8a69a9fc3e64c42685dc987affd49533f73 |
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 $
40 */
42 /*
43 * For now we want the safety net that the DIAGNOSTIC and DEBUG flags provide.
44 */
45 #ifndef DIAGNOSTIC
46 #define DIAGNOSTIC
47 #endif
48 #ifndef DEBUG
49 #define DEBUG
50 #endif
52 #include <sys/param.h>
53 #include <sys/kernel.h>
54 #include <sys/systm.h>
55 #include <sys/buf.h>
56 #include <sys/malloc.h>
57 #include <sys/mount.h>
58 #include <sys/proc.h>
59 #include <sys/syslog.h>
60 #include <sys/vnode.h>
61 #include <sys/conf.h>
62 #include <machine/inttypes.h>
72 #include <sys/buf2.h>
73 #include <sys/thread2.h>
74 #include <sys/lock.h>
76 /*
77 * These definitions need to be adapted to the system to which
78 * this file is being ported.
79 */
80 /*
81 * malloc types defined for the softdep system.
82 */
97 #define M_SOFTDEP_FLAGS (M_WAITOK | M_USE_RESERVE)
99 #define D_PAGEDEP 0
100 #define D_INODEDEP 1
101 #define D_NEWBLK 2
102 #define D_BMSAFEMAP 3
103 #define D_ALLOCDIRECT 4
104 #define D_INDIRDEP 5
105 #define D_ALLOCINDIR 6
106 #define D_FREEFRAG 7
107 #define D_FREEBLKS 8
108 #define D_FREEFILE 9
109 #define D_DIRADD 10
110 #define D_MKDIR 11
111 #define D_DIRREM 12
112 #define D_LAST D_DIRREM
114 /*
115 * translate from workitem type to memory type
116 * MUST match the defines above, such that memtype[D_XXX] == M_XXX
117 */
119 M_PAGEDEP,
120 M_INODEDEP,
121 M_NEWBLK,
122 M_BMSAFEMAP,
123 M_ALLOCDIRECT,
124 M_INDIRDEP,
125 M_ALLOCINDIR,
126 M_FREEFRAG,
127 M_FREEBLKS,
128 M_FREEFILE,
129 M_DIRADD,
130 M_MKDIR,
131 M_DIRREM
132 };
134 #define DtoM(type) (memtype[type])
136 /*
137 * Names of malloc types.
138 */
139 #define TYPENAME(type) \
141 /*
142 * End system adaptaion definitions.
143 */
145 /*
146 * Internal function prototypes.
147 */
181 ufs_daddr_t);
196 /*
197 * Exported softdep operations.
198 */
219 };
221 /*
222 * Locking primitives.
223 */
226 #ifdef INVARIANTS
228 #endif
232 #define ACQUIRE_LOCK(lkp) acquire_lock(lkp)
233 #define FREE_LOCK(lkp) free_lock(lkp)
235 static void
237 {
239 }
241 static void
243 {
245 }
247 #ifdef INVARIANTS
248 static int
250 {
252 }
253 #endif
255 /*
256 * Place holder for real semaphores.
257 */
260 thread_t holder;
264 };
269 #define NOHOLDER ((struct thread *) -1)
271 static void
273 {
279 }
281 /*
282 * Obtain exclusive access, semaphore is protected by the interlock.
283 * If interlock is NULL we must protect the semaphore ourselves.
284 */
285 static int
287 {
300 }
314 }
315 }
317 }
319 static void
321 {
329 }
339 }
340 }
341 }
343 /*
344 * Worklist queue management.
345 * These routines require that the lock be held.
346 */
351 #define WORKLIST_INSERT_BP(bp, item) do { \
352 (bp)->b_ops = &softdep_bioops; \
353 worklist_insert(&(bp)->b_dep, item); \
354 } while (0)
356 #define WORKLIST_INSERT(head, item) worklist_insert(head, item)
357 #define WORKLIST_REMOVE(item) worklist_remove(item)
358 #define WORKITEM_FREE(item, type) workitem_free((struct worklist *)item, type)
360 static void
362 {
367 }
370 }
372 static void
374 {
382 }
384 static void
386 {
394 }
396 /*
397 * Workitem queue management
398 */
409 #define FLUSH_INODES 1
411 #define FLUSH_REMOVE 2
412 /*
413 * runtime statistics
414 */
425 #ifdef DEBUG
426 #include <vm/vm.h>
427 #include <sys/sysctl.h>
454 /*
455 * Add an item to the end of the work queue.
456 * This routine requires that the lock be held.
457 * This is the only routine that adds items to the list.
458 * The following routine is the only one that removes items
459 * and does so in order from first to last.
460 */
461 static void
463 {
468 }
472 else
476 }
478 /*
479 * Process that runs once per second to handle items in the background queue.
480 *
481 * Note that we ensure that everything is done in the order in which they
482 * appear in the queue. The code below depends on this property to ensure
483 * that blocks of a file are freed before the inode itself is freed. This
484 * ordering ensures that no new <vfsid, inum, lbn> triples will be generated
485 * until all the old ones have been purged from the dependency lists.
486 *
487 * bioops callback - hold io_token
488 */
489 static int
491 {
498 /*
499 * Record the process identifier of our caller so that we can give
500 * this process preferential treatment in request_cleanup below.
501 */
505 /*
506 * There is no danger of having multiple processes run this
507 * code, but we have to single-thread it when softdep_flushfiles()
508 * is in operation to get an accurate count of the number of items
509 * related to its mount point that are in the list.
510 */
515 }
517 }
519 /*
520 * If requested, try removing inode or removal dependencies.
521 */
526 }
531 }
537 /*
538 * If a umount operation wants to run the worklist
539 * accurately, abort.
540 */
544 }
546 /*
547 * If requested, try removing inode or removal dependencies.
548 */
553 }
558 }
559 /*
560 * We do not generally want to stop for buffer space, but if
561 * we are really being a buffer hog, we will stop and wait.
562 */
567 }
569 /*
570 * Never allow processing to run for more than one
571 * second. Otherwise the other syncer tasks may get
572 * excessively backlogged.
573 *
574 * Use ticks to avoid boundary condition w/time_second or
575 * time_uptime.
576 */
580 }
581 }
586 }
587 done:
590 }
592 /*
593 * Process one item on the worklist.
594 */
595 static int
597 {
611 /*
612 * Normally we just process each item on the worklist in order.
613 * However, if we are in a situation where we cannot lock any
614 * inodes, we have to skip over any dirrem requests whose
615 * vnodes are resident and locked.
616 */
627 }
630 }
636 /* removal of a directory entry */
643 /* releasing blocks and/or fragments from a file */
650 /* releasing a fragment when replaced as a file grows */
657 /* releasing an inode when its link count drops to 0 */
666 /* NOTREACHED */
667 }
670 }
672 /*
673 * Move dependencies from one buffer to another.
674 *
675 * bioops callback - hold io_token
676 */
677 static void
679 {
690 else
694 }
696 }
698 /*
699 * Purge the work list of all items associated with a particular mount point.
700 */
701 int
703 {
707 /*
708 * Await our turn to clear out the queue, then serialize access.
709 */
715 }
724 }
725 /*
726 * Alternately flush the block device associated with the mount
727 * point and process any dependencies that the flushing
728 * creates. In theory, this loop can happen at most twice,
729 * but we give it a few extra just to be sure.
730 */
734 loopcnt--;
735 /*
736 * Do another flush in case any vnodes were brought in
737 * as part of the cleanup operations.
738 */
741 /*
742 * If we still found nothing to do, we are really done.
743 */
746 }
752 }
759 /*
760 * If we are unmounting then it is an error to fail. If we
761 * are simply trying to downgrade to read-only, then filesystem
762 * activity can keep us busy forever, so we just fail with EBUSY.
763 */
768 }
770 }
772 /*
773 * Structure hashing.
774 *
775 * There are three types of structures that can be looked up:
776 * 1) pagedep structures identified by mount point, inode number,
777 * and logical block.
778 * 2) inodedep structures identified by mount point and inode number.
779 * 3) newblk structures identified by mount point and
780 * physical block number.
781 *
782 * The "pagedep" and "inodedep" dependency structures are hashed
783 * separately from the file blocks and inodes to which they correspond.
784 * This separation helps when the in-memory copy of an inode or
785 * file block must be replaced. It also obviates the need to access
786 * an inode or file page when simply updating (or de-allocating)
787 * dependency structures. Lookup of newblk structures is needed to
788 * find newly allocated blocks when trying to associate them with
789 * their allocdirect or allocindir structure.
790 *
791 * The lookup routines optionally create and hash a new instance when
792 * an existing entry is not found.
793 */
797 /*
798 * Structures and routines associated with pagedep caching.
799 */
802 #define PAGEDEP_HASH(mp, inum, lbn) \
803 (&pagedep_hashtbl[((((register_t)(mp)) >> 13) + (inum) + (lbn)) & \
804 pagedep_hash])
807 /*
808 * Helper routine for pagedep_lookup()
809 */
810 static __inline
814 {
822 }
823 }
825 }
827 /*
828 * Look up a pagedep. Return 1 if found, 0 if not found.
829 * If not found, allocate if DEPALLOC flag is passed.
830 * Found or allocated entry is returned in pagedeppp.
831 * This routine must be called with splbio interrupts blocked.
832 */
833 static int
836 {
846 top:
864 }
878 }
880 /*
881 * Structures and routines associated with inodedep caching.
882 */
886 #define INODEDEP_HASH(fs, inum) \
887 (&inodedep_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & inodedep_hash])
890 /*
891 * Helper routine for inodedep_lookup()
892 */
893 static __inline
896 {
902 }
904 }
906 /*
907 * Look up a inodedep. Return 1 if found, 0 if not found.
908 * If not found, allocate if DEPALLOC flag is passed.
909 * Found or allocated entry is returned in inodedeppp.
910 * This routine must be called with splbio interrupts blocked.
911 */
912 static int
915 {
922 top:
929 /*
930 * If we are over our limit, try to improve the situation.
931 */
938 }
951 }
970 }
972 /*
973 * Structures and routines associated with newblk caching.
974 */
977 #define NEWBLK_HASH(fs, inum) \
978 (&newblk_hashtbl[((((register_t)(fs)) >> 13) + (inum)) & newblk_hash])
981 /*
982 * Helper routine for newblk_lookup()
983 */
984 static __inline
987 ufs_daddr_t newblkno)
988 {
994 }
996 }
998 /*
999 * Look up a newblk. Return 1 if found, 0 if not found.
1000 * If not found, allocate if DEPALLOC flag is passed.
1001 * Found or allocated entry is returned in newblkpp.
1002 */
1003 static int
1006 {
1011 top:
1028 }
1036 }
1038 /*
1039 * Executed during filesystem system initialization before
1040 * mounting any filesystems.
1041 */
1042 void
1044 {
1057 }
1059 /*
1060 * Called at mount time to notify the dependency code that a
1061 * filesystem wishes to use it.
1062 */
1063 int
1065 {
1074 /*
1075 * When doing soft updates, the counters in the
1076 * superblock may have gotten out of sync, so we have
1077 * to scan the cylinder groups and recalculate them.
1078 */
1087 }
1095 }
1096 #ifdef DEBUG
1099 #endif
1102 }
1104 /*
1105 * Protecting the freemaps (or bitmaps).
1106 *
1107 * To eliminate the need to execute fsck before mounting a filesystem
1108 * after a power failure, one must (conservatively) guarantee that the
1109 * on-disk copy of the bitmaps never indicate that a live inode or block is
1110 * free. So, when a block or inode is allocated, the bitmap should be
1111 * updated (on disk) before any new pointers. When a block or inode is
1112 * freed, the bitmap should not be updated until all pointers have been
1113 * reset. The latter dependency is handled by the delayed de-allocation
1114 * approach described below for block and inode de-allocation. The former
1115 * dependency is handled by calling the following procedure when a block or
1116 * inode is allocated. When an inode is allocated an "inodedep" is created
1117 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1118 * Each "inodedep" is also inserted into the hash indexing structure so
1119 * that any additional link additions can be made dependent on the inode
1120 * allocation.
1121 *
1122 * The ufs filesystem maintains a number of free block counts (e.g., per
1123 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1124 * in addition to the bitmaps. These counts are used to improve efficiency
1125 * during allocation and therefore must be consistent with the bitmaps.
1126 * There is no convenient way to guarantee post-crash consistency of these
1127 * counts with simple update ordering, for two main reasons: (1) The counts
1128 * and bitmaps for a single cylinder group block are not in the same disk
1129 * sector. If a disk write is interrupted (e.g., by power failure), one may
1130 * be written and the other not. (2) Some of the counts are located in the
1131 * superblock rather than the cylinder group block. So, we focus our soft
1132 * updates implementation on protecting the bitmaps. When mounting a
1133 * filesystem, we recompute the auxiliary counts from the bitmaps.
1134 */
1136 /*
1137 * Called just after updating the cylinder group block to allocate an inode.
1138 *
1139 * Parameters:
1140 * bp: buffer for cylgroup block with inode map
1141 * ip: inode related to allocation
1142 * newinum: new inode number being allocated
1143 */
1144 void
1146 {
1150 /*
1151 * Create a dependency for the newly allocated inode.
1152 * Panic if it already exists as something is seriously wrong.
1153 * Otherwise add it to the dependency list for the buffer holding
1154 * the cylinder group map from which it was allocated.
1155 */
1159 }
1165 }
1167 /*
1168 * Called just after updating the cylinder group block to
1169 * allocate block or fragment.
1170 *
1171 * Parameters:
1172 * bp: buffer for cylgroup block with block map
1173 * fs: filesystem doing allocation
1174 * newblkno: number of newly allocated block
1175 */
1176 void
1178 ufs_daddr_t newblkno)
1179 {
1183 /*
1184 * Create a dependency for the newly allocated block.
1185 * Add it to the dependency list for the buffer holding
1186 * the cylinder group map from which it was allocated.
1187 */
1194 }
1196 /*
1197 * Find the bmsafemap associated with a cylinder group buffer.
1198 * If none exists, create one. The buffer must be locked when
1199 * this routine is called and this routine must be called with
1200 * splbio interrupts blocked.
1201 */
1204 {
1213 }
1216 M_SOFTDEP_FLAGS);
1227 }
1229 /*
1230 * Direct block allocation dependencies.
1231 *
1232 * When a new block is allocated, the corresponding disk locations must be
1233 * initialized (with zeros or new data) before the on-disk inode points to
1234 * them. Also, the freemap from which the block was allocated must be
1235 * updated (on disk) before the inode's pointer. These two dependencies are
1236 * independent of each other and are needed for all file blocks and indirect
1237 * blocks that are pointed to directly by the inode. Just before the
1238 * "in-core" version of the inode is updated with a newly allocated block
1239 * number, a procedure (below) is called to setup allocation dependency
1240 * structures. These structures are removed when the corresponding
1241 * dependencies are satisfied or when the block allocation becomes obsolete
1242 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1243 * fragment that gets upgraded). All of these cases are handled in
1244 * procedures described later.
1245 *
1246 * When a file extension causes a fragment to be upgraded, either to a larger
1247 * fragment or to a full block, the on-disk location may change (if the
1248 * previous fragment could not simply be extended). In this case, the old
1249 * fragment must be de-allocated, but not until after the inode's pointer has
1250 * been updated. In most cases, this is handled by later procedures, which
1251 * will construct a "freefrag" structure to be added to the workitem queue
1252 * when the inode update is complete (or obsolete). The main exception to
1253 * this is when an allocation occurs while a pending allocation dependency
1254 * (for the same block pointer) remains. This case is handled in the main
1255 * allocation dependency setup procedure by immediately freeing the
1256 * unreferenced fragments.
1257 *
1258 * Parameters:
1259 * ip: inode to which block is being added
1260 * lbn: block pointer within inode
1261 * newblkno: disk block number being added
1262 * oldblkno: previous block number, 0 unless frag
1263 * newsize: size of new block
1264 * oldsize: size of new block
1265 * bp: bp for allocated block
1266 */
1267 void
1271 {
1290 else
1308 }
1314 /* allocating an indirect block */
1317 }
1319 /*
1320 * Allocating a direct block.
1321 *
1322 * If we are allocating a directory block, then we must
1323 * allocate an associated pagedep to track additions and
1324 * deletions.
1325 */
1329 }
1330 }
1331 /*
1332 * The list of allocdirects must be kept in sorted and ascending
1333 * order so that the rollback routines can quickly determine the
1334 * first uncommitted block (the size of the file stored on disk
1335 * ends at the end of the lowest committed fragment, or if there
1336 * are no fragments, at the end of the highest committed block).
1337 * Since files generally grow, the typical case is that the new
1338 * block is to be added at the end of the list. We speed this
1339 * special case by checking against the last allocdirect in the
1340 * list before laboriously traversing the list looking for the
1341 * insertion point.
1342 */
1346 /* insert at end of list */
1352 }
1356 }
1359 }
1360 /* insert in middle of list */
1365 }
1367 /*
1368 * Replace an old allocdirect dependency with a newer one.
1369 * This routine must be called with splbio interrupts blocked.
1370 *
1371 * Parameters:
1372 * adphead: head of list holding allocdirects
1373 * newadp: allocdirect being added
1374 * oldadp: existing allocdirect being checked
1375 */
1376 static void
1380 {
1390 NDADDR);
1391 }
1394 /*
1395 * If the old dependency had a fragment to free or had never
1396 * previously had a block allocated, then the new dependency
1397 * can immediately post its freefrag and adopt the old freefrag.
1398 * This action is done by swapping the freefrag dependencies.
1399 * The new dependency gains the old one's freefrag, and the
1400 * old one gets the new one and then immediately puts it on
1401 * the worklist when it is freed by free_allocdirect. It is
1402 * not possible to do this swap when the old dependency had a
1403 * non-zero size but no previous fragment to free. This condition
1404 * arises when the new block is an extension of the old block.
1405 * Here, the first part of the fragment allocated to the new
1406 * dependency is part of the block currently claimed on disk by
1407 * the old dependency, so cannot legitimately be freed until the
1408 * conditions for the new dependency are fulfilled.
1409 */
1414 }
1416 }
1418 /*
1419 * Allocate a new freefrag structure if needed.
1420 */
1423 {
1433 M_SOFTDEP_FLAGS);
1442 }
1444 /*
1445 * This workitem de-allocates fragments that were replaced during
1446 * file block allocation.
1447 */
1448 static void
1450 {
1460 }
1462 /*
1463 * Indirect block allocation dependencies.
1464 *
1465 * The same dependencies that exist for a direct block also exist when
1466 * a new block is allocated and pointed to by an entry in a block of
1467 * indirect pointers. The undo/redo states described above are also
1468 * used here. Because an indirect block contains many pointers that
1469 * may have dependencies, a second copy of the entire in-memory indirect
1470 * block is kept. The buffer cache copy is always completely up-to-date.
1471 * The second copy, which is used only as a source for disk writes,
1472 * contains only the safe pointers (i.e., those that have no remaining
1473 * update dependencies). The second copy is freed when all pointers
1474 * are safe. The cache is not allowed to replace indirect blocks with
1475 * pending update dependencies. If a buffer containing an indirect
1476 * block with dependencies is written, these routines will mark it
1477 * dirty again. It can only be successfully written once all the
1478 * dependencies are removed. The ffs_fsync routine in conjunction with
1479 * softdep_sync_metadata work together to get all the dependencies
1480 * removed so that a file can be successfully written to disk. Three
1481 * procedures are used when setting up indirect block pointer
1482 * dependencies. The division is necessary because of the organization
1483 * of the "balloc" routine and because of the distinction between file
1484 * pages and file metadata blocks.
1485 */
1487 /*
1488 * Allocate a new allocindir structure.
1489 *
1490 * Parameters:
1491 * ip: inode for file being extended
1492 * ptrno: offset of pointer in indirect block
1493 * newblkno: disk block number being added
1494 * oldblkno: previous block number, 0 if none
1495 */
1498 ufs_daddr_t oldblkno)
1499 {
1511 }
1513 /*
1514 * Called just before setting an indirect block pointer
1515 * to a newly allocated file page.
1516 *
1517 * Parameters:
1518 * ip: inode for file being extended
1519 * lbn: allocated block number within file
1520 * bp: buffer with indirect blk referencing page
1521 * ptrno: offset of pointer in indirect block
1522 * newblkno: disk block number being added
1523 * oldblkno: previous block number, 0 if none
1524 * nbp: buffer holding allocated page
1525 */
1526 void
1531 {
1537 /*
1538 * If we are allocating a directory page, then we must
1539 * allocate an associated pagedep to track additions and
1540 * deletions.
1541 */
1548 }
1550 /*
1551 * Called just before setting an indirect block pointer to a
1552 * newly allocated indirect block.
1553 * Parameters:
1554 * nbp: newly allocated indirect block
1555 * ip: inode for file being extended
1556 * bp: indirect block referencing allocated block
1557 * ptrno: offset of pointer in indirect block
1558 * newblkno: disk block number being added
1559 */
1560 void
1563 ufs_daddr_t newblkno)
1564 {
1572 }
1574 /*
1575 * Called to finish the allocation of the "aip" allocated
1576 * by one of the two routines above.
1577 *
1578 * Parameters:
1579 * bp: in-memory copy of the indirect block
1580 * ip: inode for file being extended
1581 * aip: allocindir allocated by the above routines
1582 */
1583 static void
1586 {
1603 }
1608 }
1624 }
1628 /*
1629 * Check to see if there is an existing dependency
1630 * for this block. If there is, merge the old
1631 * dependency into the new one.
1632 */
1635 else
1643 }
1649 }
1654 }
1656 /*
1657 * Avoid any possibility of data corruption by
1658 * ensuring that our old version is thrown away.
1659 */
1663 }
1667 M_SOFTDEP_FLAGS);
1675 BUF_CMD_WRITE);
1676 }
1683 }
1684 }
1686 /*
1687 * Block de-allocation dependencies.
1688 *
1689 * When blocks are de-allocated, the on-disk pointers must be nullified before
1690 * the blocks are made available for use by other files. (The true
1691 * requirement is that old pointers must be nullified before new on-disk
1692 * pointers are set. We chose this slightly more stringent requirement to
1693 * reduce complexity.) Our implementation handles this dependency by updating
1694 * the inode (or indirect block) appropriately but delaying the actual block
1695 * de-allocation (i.e., freemap and free space count manipulation) until
1696 * after the updated versions reach stable storage. After the disk is
1697 * updated, the blocks can be safely de-allocated whenever it is convenient.
1698 * This implementation handles only the common case of reducing a file's
1699 * length to zero. Other cases are handled by the conventional synchronous
1700 * write approach.
1701 *
1702 * The ffs implementation with which we worked double-checks
1703 * the state of the block pointers and file size as it reduces
1704 * a file's length. Some of this code is replicated here in our
1705 * soft updates implementation. The freeblks->fb_chkcnt field is
1706 * used to transfer a part of this information to the procedure
1707 * that eventually de-allocates the blocks.
1708 *
1709 * This routine should be called from the routine that shortens
1710 * a file's length, before the inode's size or block pointers
1711 * are modified. It will save the block pointer information for
1712 * later release and zero the inode so that the calling routine
1713 * can release it.
1714 */
1718 };
1722 /*
1723 * Parameters:
1724 * ip: The inode whose length is to be reduced
1725 * length: The new length for the file
1726 */
1727 void
1729 {
1757 }
1761 }
1764 /*
1765 * Push the zero'ed inode to to its disk buffer so that we are free
1766 * to delete its dependencies below. Once the dependencies are gone
1767 * the buffer can be safely released.
1768 */
1775 /*
1776 * Find and eliminate any inode dependencies.
1777 */
1782 }
1783 /*
1784 * Add the freeblks structure to the list of operations that
1785 * must await the zero'ed inode being written to disk. If we
1786 * still have a bitmap dependency (delay == 0), then the inode
1787 * has never been written to disk, so we can process the
1788 * freeblks below once we have deleted the dependencies.
1789 */
1793 /*
1794 * Because the file length has been truncated to zero, any
1795 * pending block allocation dependency structures associated
1796 * with this inode are obsolete and can simply be de-allocated.
1797 * We must first merge the two dependency lists to get rid of
1798 * any duplicate freefrag structures, then purge the merged list.
1799 */
1805 /*
1806 * We must wait for any I/O in progress to finish so that
1807 * all potential buffers on the dirty list will be visible.
1808 * Once they are all there, walk the list and get rid of
1809 * any dependencies.
1810 */
1829 /*
1830 * If the inode with zeroed block pointers is now on disk
1831 * we can start freeing blocks. Add freeblks to the worklist
1832 * instead of calling handle_workitem_freeblocks directly as
1833 * it is more likely that additional IO is needed to complete
1834 * the request here than in the !delay case.
1835 */
1838 }
1841 /*
1842 * If the inode has never been written to disk (delay == 0),
1843 * then we can process the freeblks now that we have deleted
1844 * the dependencies.
1845 */
1848 }
1850 static int
1852 {
1859 }
1864 }
1872 }
1874 /*
1875 * Reclaim any dependency structures from a buffer that is about to
1876 * be reallocated to a new vnode. The buffer must be locked, thus,
1877 * no I/O completion operations can occur while we are manipulating
1878 * its associated dependencies. The mutex is held so that other I/O's
1879 * associated with related dependencies do not occur.
1880 */
1881 static void
1883 {
1897 /*
1898 * None of the indirect pointers will ever be visible,
1899 * so they can simply be tossed. GOINGAWAY ensures
1900 * that allocated pointers will be saved in the buffer
1901 * cache until they are freed. Note that they will
1902 * only be able to be found by their physical address
1903 * since the inode mapping the logical address will
1904 * be gone. The save buffer used for the safe copy
1905 * was allocated in setup_allocindir_phase2 using
1906 * the physical address so it could be used for this
1907 * purpose. Hence we swap the safe copy with the real
1908 * copy, allowing the safe copy to be freed and holding
1909 * on to the real copy for later use in indir_trunc.
1910 *
1911 * NOTE: ir_savebp is relative to the block device
1912 * so b_bio1 contains the device block number.
1913 */
1916 }
1923 }
1932 /*
1933 * None of the directory additions will ever be
1934 * visible, so they can simply be tossed.
1935 */
1942 /*
1943 * Copy any directory remove dependencies to the list
1944 * to be processed after the zero'ed inode is written.
1945 * If the inode has already been written, then they
1946 * can be dumped directly onto the work list.
1947 */
1953 ALLCOMPLETE)
1955 else
1958 }
1972 /* NOTREACHED */
1977 /* NOTREACHED */
1978 }
1979 }
1980 }
1982 /*
1983 * Free an allocdirect. Generate a new freefrag work request if appropriate.
1984 * This routine must be called with splbio interrupts blocked.
1985 */
1986 static void
1989 {
2001 else
2003 }
2005 }
2007 /*
2008 * Prepare an inode to be freed. The actual free operation is not
2009 * done until the zero'ed inode has been written to disk.
2010 */
2011 void
2013 {
2018 /*
2019 * This sets up the inode de-allocation dependency.
2020 */
2022 M_SOFTDEP_FLAGS);
2030 /*
2031 * If the inodedep does not exist, then the zero'ed inode has
2032 * been written to disk. If the allocated inode has never been
2033 * written to disk, then the on-disk inode is zero'ed. In either
2034 * case we can free the file immediately.
2035 */
2042 }
2045 }
2047 /*
2048 * Check to see if an inode has never been written to disk. If
2049 * so free the inodedep and return success, otherwise return failure.
2050 * This routine must be called with splbio interrupts blocked.
2051 *
2052 * If we still have a bitmap dependency, then the inode has never
2053 * been written to disk. Drop the dependency as it is no longer
2054 * necessary since the inode is being deallocated. We set the
2055 * ALLCOMPLETE flags since the bitmap now properly shows that the
2056 * inode is not allocated. Even if the inode is actively being
2057 * written, it has been rolled back to its zero'ed state, so we
2058 * are ensured that a zero inode is what is on the disk. For short
2059 * lived files, this change will usually result in removing all the
2060 * dependencies from the inode so that it can be freed immediately.
2061 */
2062 static int
2064 {
2075 /*
2076 * Another process might be in initiate_write_inodeblock
2077 * trying to allocate memory without holding "Softdep Lock".
2078 */
2091 }
2094 }
2096 }
2098 /*
2099 * Try to free an inodedep structure. Return 1 if it could be freed.
2100 */
2101 static int
2103 {
2118 }
2120 /*
2121 * This workitem routine performs the block de-allocation.
2122 * The workitem is added to the pending list after the updated
2123 * inode block has been written to disk. As mentioned above,
2124 * checks regarding the number of blocks de-allocated (compared
2125 * to the number of blocks allocated for the file) are also
2126 * performed in this function.
2127 */
2128 static void
2130 {
2132 ufs_daddr_t bn;
2151 }
2154 /*
2155 * Indirect blocks first.
2156 */
2165 }
2166 /*
2167 * All direct blocks or frags.
2168 */
2175 }
2177 #ifdef DIAGNOSTIC
2184 }
2186 /*
2187 * Release blocks associated with the inode ip and stored in the indirect
2188 * block at doffset. If level is greater than SINGLE, the block is an
2189 * indirect block and recursive calls to indirtrunc must be used to
2190 * cleanse other indirect blocks.
2191 */
2192 static int
2195 {
2198 ufs_daddr_t nb;
2209 /*
2210 * Get buffer of block pointers to be freed. This routine is not
2211 * called until the zero'ed inode has been written, so it is safe
2212 * to free blocks as they are encountered. Because the inode has
2213 * been zero'ed, calls to bmap on these blocks will fail. So, we
2214 * have to use the on-disk address and the block device for the
2215 * filesystem to look them up. If the file was deleted before its
2216 * indirect blocks were all written to disk, the routine that set
2217 * us up (deallocate_dependencies) will have arranged to leave
2218 * a complete copy of the indirect block in memory for our use.
2219 * Otherwise we have to read the blocks in from the disk.
2220 */
2224 /*
2225 * bp must be ir_savebp, which is held locked for our use.
2226 */
2231 }
2236 }
2243 }
2244 /*
2245 * Recursively free indirect blocks.
2246 */
2256 }
2259 }
2263 }
2265 /*
2266 * Free an allocindir.
2267 * This routine must be called with splbio interrupts blocked.
2268 */
2269 static void
2271 {
2284 else
2287 }
2289 }
2291 /*
2292 * Directory entry addition dependencies.
2293 *
2294 * When adding a new directory entry, the inode (with its incremented link
2295 * count) must be written to disk before the directory entry's pointer to it.
2296 * Also, if the inode is newly allocated, the corresponding freemap must be
2297 * updated (on disk) before the directory entry's pointer. These requirements
2298 * are met via undo/redo on the directory entry's pointer, which consists
2299 * simply of the inode number.
2300 *
2301 * As directory entries are added and deleted, the free space within a
2302 * directory block can become fragmented. The ufs filesystem will compact
2303 * a fragmented directory block to make space for a new entry. When this
2304 * occurs, the offsets of previously added entries change. Any "diradd"
2305 * dependency structures corresponding to these entries must be updated with
2306 * the new offsets.
2307 */
2309 /*
2310 * This routine is called after the in-memory inode's link
2311 * count has been incremented, but before the directory entry's
2312 * pointer to the inode has been set.
2313 *
2314 * Parameters:
2315 * bp: buffer containing directory block
2316 * dp: inode for directory
2317 * diroffset: offset of new entry in directory
2318 * newinum: inode referenced by new directory entry
2319 * newdirbp: non-NULL => contents of new mkdir
2320 */
2321 void
2324 {
2333 /*
2334 * Whiteouts have no dependencies.
2335 */
2340 }
2357 M_SOFTDEP_FLAGS);
2362 M_SOFTDEP_FLAGS);
2366 /*
2367 * Dependency on "." and ".." being written to disk.
2368 */
2375 /*
2376 * Dependency on link count increase for parent directory
2377 */
2386 }
2387 }
2388 /*
2389 * Link into parent directory pagedep to await its being written.
2390 */
2395 da_pdlist);
2396 /*
2397 * Link into its inodedep. Put it on the id_bufwait list if the inode
2398 * is not yet written. If it is written, do the post-inode write
2399 * processing to put it on the id_pendinghd list.
2400 */
2404 else
2407 }
2409 /*
2410 * This procedure is called to change the offset of a directory
2411 * entry when compacting a directory block which must be owned
2412 * exclusively by the caller. Note that the actual entry movement
2413 * must be done in this procedure to ensure that no I/O completions
2414 * occur while the move is in progress.
2415 *
2416 * Parameters:
2417 * dp: inode for directory
2418 * base: address of dp->i_offset
2419 * oldloc: address of old directory location
2420 * newloc: address of new directory location
2421 * entrysize: size of directory entry
2422 */
2423 void
2427 {
2431 ufs_lbn_t lbn;
2451 }
2458 }
2459 }
2460 }
2461 done:
2464 }
2466 /*
2467 * Free a diradd dependency structure. This routine must be called
2468 * with splbio interrupts blocked.
2469 */
2470 static void
2472 {
2489 }
2502 }
2505 }
2506 }
2508 }
2510 /*
2511 * Directory entry removal dependencies.
2512 *
2513 * When removing a directory entry, the entry's inode pointer must be
2514 * zero'ed on disk before the corresponding inode's link count is decremented
2515 * (possibly freeing the inode for re-use). This dependency is handled by
2516 * updating the directory entry but delaying the inode count reduction until
2517 * after the directory block has been written to disk. After this point, the
2518 * inode count can be decremented whenever it is convenient.
2519 */
2521 /*
2522 * This routine should be called immediately after removing
2523 * a directory entry. The inode's link count should not be
2524 * decremented by the calling procedure -- the soft updates
2525 * code will do this task when it is safe.
2526 *
2527 * Parameters:
2528 * bp: buffer containing directory block
2529 * dp: inode for the directory being modified
2530 * ip: inode for directory entry being removed
2531 * isrmdir: indicates if doing RMDIR
2532 */
2533 void
2536 {
2539 /*
2540 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2541 */
2544 /*
2545 * If the COMPLETE flag is clear, then there were no active
2546 * entries and we want to roll back to a zeroed entry until
2547 * the new inode is committed to disk. If the COMPLETE flag is
2548 * set then we have deleted an entry that never made it to
2549 * disk. If the entry we deleted resulted from a name change,
2550 * then the old name still resides on disk. We cannot delete
2551 * its inode (returned to us in prevdirrem) until the zeroed
2552 * directory entry gets to disk. The new inode has never been
2553 * referenced on the disk, so can be deleted immediately.
2554 */
2557 dm_next);
2566 }
2567 }
2569 /*
2570 * Allocate a new dirrem if appropriate and return it along with
2571 * its associated pagedep. Called without a lock, returns with lock.
2572 */
2575 /*
2576 * Parameters:
2577 * bp: buffer containing directory block
2578 * dp: inode for the directory being modified
2579 * ip: inode for directory entry being removed
2580 * isrmdir: indicates if doing RMDIR
2581 * prevdirremp: previously referenced inode, if any
2582 */
2586 {
2588 ufs_lbn_t lbn;
2593 /*
2594 * Whiteouts have no deletion dependencies.
2595 */
2598 /*
2599 * If we are over our limit, try to improve the situation.
2600 * Limiting the number of dirrem structures will also limit
2601 * the number of freefile and freeblks structures.
2602 */
2609 }
2626 /*
2627 * Check for a diradd dependency for the same directory entry.
2628 * If present, then both dependencies become obsolete and can
2629 * be de-allocated. Check for an entry on both the pd_dirraddhd
2630 * list and the pd_pendinghd list.
2631 */
2643 }
2644 /*
2645 * Must be ATTACHED at this point.
2646 */
2649 }
2653 }
2654 /*
2655 * If we are deleting a changed name that never made it to disk,
2656 * then return the dirrem describing the previous inode (which
2657 * represents the inode currently referenced from this entry on disk).
2658 */
2663 }
2664 /*
2665 * We are deleting an entry that never made it to disk.
2666 * Mark it COMPLETE so we can delete its inode immediately.
2667 */
2671 }
2673 /*
2674 * Directory entry change dependencies.
2675 *
2676 * Changing an existing directory entry requires that an add operation
2677 * be completed first followed by a deletion. The semantics for the addition
2678 * are identical to the description of adding a new entry above except
2679 * that the rollback is to the old inode number rather than zero. Once
2680 * the addition dependency is completed, the removal is done as described
2681 * in the removal routine above.
2682 */
2684 /*
2685 * This routine should be called immediately after changing
2686 * a directory entry. The inode's link count should not be
2687 * decremented by the calling procedure -- the soft updates
2688 * code will perform this task when it is safe.
2689 *
2690 * Parameters:
2691 * bp: buffer containing directory block
2692 * dp: inode for the directory being modified
2693 * ip: inode for directory entry being removed
2694 * newinum: new inode number for changed entry
2695 * isrmdir: indicates if doing RMDIR
2696 */
2697 void
2701 {
2710 /*
2711 * Whiteouts do not need diradd dependencies.
2712 */
2720 }
2722 /*
2723 * Allocate a new dirrem and ACQUIRE_LOCK.
2724 */
2727 /*
2728 * The possible values for isrmdir:
2729 * 0 - non-directory file rename
2730 * 1 - directory rename within same directory
2731 * inum - directory rename to new directory of given inode number
2732 * When renaming to a new directory, we are both deleting and
2733 * creating a new directory entry, so the link count on the new
2734 * directory should not change. Thus we do not need the followup
2735 * dirrem which is usually done in handle_workitem_remove. We set
2736 * the DIRCHG flag to tell handle_workitem_remove to skip the
2737 * followup dirrem.
2738 */
2742 /*
2743 * Whiteouts have no additional dependencies,
2744 * so just put the dirrem on the correct list.
2745 */
2749 dm_next);
2753 }
2756 }
2758 /*
2759 * If the COMPLETE flag is clear, then there were no active
2760 * entries and we want to roll back to the previous inode until
2761 * the new inode is committed to disk. If the COMPLETE flag is
2762 * set, then we have deleted an entry that never made it to disk.
2763 * If the entry we deleted resulted from a name change, then the old
2764 * inode reference still resides on disk. Any rollback that we do
2765 * needs to be to that old inode (returned to us in prevdirrem). If
2766 * the entry we deleted resulted from a create, then there is
2767 * no entry on the disk, so we want to roll back to zero rather
2768 * than the uncommitted inode. In either of the COMPLETE cases we
2769 * want to immediately free the unwritten and unreferenced inode.
2770 */
2779 }
2782 }
2783 /*
2784 * Link into its inodedep. Put it on the id_bufwait list if the inode
2785 * is not yet written. If it is written, do the post-inode write
2786 * processing to put it on the id_pendinghd list.
2787 */
2797 }
2799 }
2801 /*
2802 * Called whenever the link count on an inode is changed.
2803 * It creates an inode dependency so that the new reference(s)
2804 * to the inode cannot be committed to disk until the updated
2805 * inode has been written.
2806 *
2807 * Parameters:
2808 * ip: the inode with the increased link count
2809 */
2810 void
2812 {
2819 }
2822 }
2824 /*
2825 * This workitem decrements the inode's link count.
2826 * If the link count reaches zero, the file is removed.
2827 */
2828 static void
2830 {
2834 ino_t oldinum;
2841 }
2846 }
2847 /*
2848 * Normal file deletion.
2849 */
2855 }
2862 }
2863 /*
2864 * Directory deletion. Decrement reference count for both the
2865 * just deleted parent directory entry and the reference for ".".
2866 * Next truncate the directory to length zero. When the
2867 * truncation completes, arrange to have the reference count on
2868 * the parent decremented to account for the loss of "..".
2869 */
2874 }
2879 /*
2880 * Rename a directory to a new parent. Since, we are both deleting
2881 * and creating a new directory entry, the link count on the new
2882 * directory should not change. Thus we skip the followup dirrem.
2883 */
2889 }
2890 /*
2891 * If the inodedep does not exist, then the zero'ed inode has
2892 * been written to disk. If the allocated inode has never been
2893 * written to disk, then the on-disk inode is zero'ed. In either
2894 * case we can remove the file immediately.
2895 */
2906 }
2912 }
2914 /*
2915 * Inode de-allocation dependencies.
2916 *
2917 * When an inode's link count is reduced to zero, it can be de-allocated. We
2918 * found it convenient to postpone de-allocation until after the inode is
2919 * written to disk with its new link count (zero). At this point, all of the
2920 * on-disk inode's block pointers are nullified and, with careful dependency
2921 * list ordering, all dependencies related to the inode will be satisfied and
2922 * the corresponding dependency structures de-allocated. So, if/when the
2923 * inode is reused, there will be no mixing of old dependencies with new
2924 * ones. This artificial dependency is set up by the block de-allocation
2925 * procedure above (softdep_setup_freeblocks) and completed by the
2926 * following procedure.
2927 */
2928 static void
2930 {
2936 #ifdef DEBUG
2942 #endif
2950 }
2952 /*
2953 * Helper function which unlinks marker element from work list and returns
2954 * the next element on the list.
2955 */
2958 {
2964 }
2966 /*
2967 * checkread, checkwrite
2968 *
2969 * bioops callback - hold io_token
2970 */
2971 static int
2973 {
2974 /* nothing to do, mp lock not needed */
2976 }
2978 /*
2979 * bioops callback - hold io_token
2980 */
2981 static int
2983 {
2984 /* nothing to do, mp lock not needed */
2986 }
2988 /*
2989 * Disk writes.
2990 *
2991 * The dependency structures constructed above are most actively used when file
2992 * system blocks are written to disk. No constraints are placed on when a
2993 * block can be written, but unsatisfied update dependencies are made safe by
2994 * modifying (or replacing) the source memory for the duration of the disk
2995 * write. When the disk write completes, the memory block is again brought
2996 * up-to-date.
2997 *
2998 * In-core inode structure reclamation.
2999 *
3000 * Because there are a finite number of "in-core" inode structures, they are
3001 * reused regularly. By transferring all inode-related dependencies to the
3002 * in-memory inode block and indexing them separately (via "inodedep"s), we
3003 * can allow "in-core" inode structures to be reused at any time and avoid
3004 * any increase in contention.
3005 *
3006 * Called just before entering the device driver to initiate a new disk I/O.
3007 * The buffer must be locked, thus, no I/O completion operations can occur
3008 * while we are manipulating its associated dependencies.
3009 *
3010 * bioops callback - hold io_token
3011 *
3012 * Parameters:
3013 * bp: structure describing disk write to occur
3014 */
3015 static void
3017 {
3022 /*
3023 * We only care about write operations. There should never
3024 * be dependencies for reads.
3025 */
3032 /*
3033 * Do any necessary pre-I/O processing.
3034 */
3051 /*
3052 * If there are no remaining dependencies, this
3053 * will be writing the real pointers, so the
3054 * dependency can be freed.
3055 */
3059 /* inline expand WORKLIST_REMOVE(wk); */
3064 }
3065 /*
3066 * Replace up-to-date version with safe version.
3067 */
3069 M_INDIRDEP,
3070 M_SOFTDEP_FLAGS);
3089 /* NOTREACHED */
3090 }
3091 }
3093 }
3095 /*
3096 * Called from within the procedure above to deal with unsatisfied
3097 * allocation dependencies in a directory. The buffer must be locked,
3098 * thus, no I/O completion operations can occur while we are
3099 * manipulating its associated dependencies.
3100 */
3101 static void
3103 {
3109 /*
3110 * This can only happen if there is a driver that does not
3111 * understand chaining. Here biodone will reissue the call
3112 * to strategy for the incomplete buffers.
3113 */
3116 }
3127 }
3130 else
3134 }
3135 }
3137 }
3139 /*
3140 * Called from within the procedure above to deal with unsatisfied
3141 * allocation dependencies in an inodeblock. The buffer must be
3142 * locked, thus, no I/O completion operations can occur while we
3143 * are manipulating its associated dependencies.
3144 *
3145 * Parameters:
3146 * bp: The inode block
3147 */
3148 static void
3150 {
3164 /*
3165 * If the bitmap is not yet written, then the allocated
3166 * inode cannot be written to disk.
3167 */
3172 M_SOFTDEP_FLAGS);
3178 }
3179 /*
3180 * If no dependencies, then there is nothing to roll back.
3181 */
3185 /*
3186 * Set the dependencies to busy.
3187 */
3191 #ifdef DIAGNOSTIC
3194 }
3201 }
3207 }
3212 }
3216 }
3217 /*
3218 * The on-disk inode cannot claim to be any larger than the last
3219 * fragment that has been written. Otherwise, the on-disk inode
3220 * might have fragments that were not the last block in the file
3221 * which would corrupt the filesystem.
3222 */
3228 /* keep going until hitting a rollback to a frag */
3233 #ifdef DIAGNOSTIC
3236 }
3239 }
3241 #ifdef DIAGNOSTIC
3245 }
3248 }
3251 }
3252 /*
3253 * If we have zero'ed out the last allocated block of the file,
3254 * roll back the size to the last currently allocated block.
3255 * We know that this last allocated block is a full-sized as
3256 * we already checked for fragments in the loop above.
3257 */
3264 }
3265 /*
3266 * The only dependencies are for indirect blocks.
3267 *
3268 * The file size for indirect block additions is not guaranteed.
3269 * Such a guarantee would be non-trivial to achieve. The conventional
3270 * synchronous write implementation also does not make this guarantee.
3271 * Fsck should catch and fix discrepancies. Arguably, the file size
3272 * can be over-estimated without destroying integrity when the file
3273 * moves into the indirect blocks (i.e., is large). If we want to
3274 * postpone fsck, we are stuck with this argument.
3275 */
3279 }
3281 /*
3282 * This routine is called during the completion interrupt
3283 * service routine for a disk write (from the procedure called
3284 * by the device driver to inform the filesystem caches of
3285 * a request completion). It should be called early in this
3286 * procedure, before the block is made available to other
3287 * processes or other routines are called.
3288 *
3289 * bioops callback - hold io_token
3290 *
3291 * Parameters:
3292 * bp: describes the completed disk write
3293 */
3294 static void
3296 {
3329 }
3336 }
3343 }
3349 }
3373 }
3383 }
3384 }
3387 stat_indir_blk_ptrs++;
3394 /* NOTREACHED */
3395 }
3396 }
3397 /*
3398 * Reattach any requests that must be redone.
3399 */
3403 }
3406 }
3408 /*
3409 * Called from within softdep_disk_write_complete above. Note that
3410 * this routine is always called from interrupt level with further
3411 * splbio interrupts blocked.
3412 *
3413 * Parameters:
3414 * adp: the completed allocdirect
3415 */
3416 static void
3418 {
3428 /*
3429 * The on-disk inode cannot claim to be any larger than the last
3430 * fragment that has been written. Otherwise, the on-disk inode
3431 * might have fragments that were not the last block in the file
3432 * which would corrupt the filesystem. Thus, we cannot free any
3433 * allocdirects after one whose ad_oldblkno claims a fragment as
3434 * these blocks must be rolled back to zero before writing the inode.
3435 * We check the currently active set of allocdirects in id_inoupdt.
3436 */
3440 /* found our block */
3443 /* continue if ad_oldlbn is not a fragment */
3447 /* hit a fragment */
3449 }
3450 /*
3451 * If we have reached the end of the current list without
3452 * finding the just finished dependency, then it must be
3453 * on the future dependency list. Future dependencies cannot
3454 * be freed until they are moved to the current list.
3455 */
3457 #ifdef DEBUG
3459 /* found our block */
3466 }
3467 /*
3468 * If we have found the just finished dependency, then free
3469 * it along with anything that follows it that is complete.
3470 */
3476 }
3477 }
3479 /*
3480 * Called from within softdep_disk_write_complete above. Note that
3481 * this routine is always called from interrupt level with further
3482 * splbio interrupts blocked.
3483 *
3484 * Parameters:
3485 * aip: the completed allocindir
3486 */
3487 static void
3489 {
3502 }
3509 }
3511 /*
3512 * Called from within softdep_disk_write_complete above to restore
3513 * in-memory inode block contents to their most up-to-date state. Note
3514 * that this routine is always called from interrupt level with further
3515 * splbio interrupts blocked.
3516 *
3517 * Parameters:
3518 * bp: buffer containing the inode block
3519 */
3520 static int
3522 {
3534 /*
3535 * If we had to rollback the inode allocation because of
3536 * bitmaps being incomplete, then simply restore it.
3537 * Keep the block dirty so that it will not be reclaimed until
3538 * all associated dependencies have been cleared and the
3539 * corresponding updates written to disk.
3540 */
3546 stat_inode_bitmap++;
3549 }
3551 /*
3552 * Roll forward anything that had to be rolled back before
3553 * the inode could be updated.
3554 */
3567 }
3575 }
3577 }
3581 }
3583 stat_direct_blk_ptrs++;
3584 /*
3585 * Reset the file size to its most up-to-date value.
3586 */
3589 }
3593 }
3595 /*
3596 * If there were any rollbacks in the inode block, then it must be
3597 * marked dirty so that its will eventually get written back in
3598 * its correct form.
3599 */
3602 /*
3603 * Process any allocdirects that completed during the update.
3604 */
3607 /*
3608 * Process deallocations that were held pending until the
3609 * inode had been written to disk. Freeing of the inode
3610 * is delayed until after all blocks have been freed to
3611 * avoid creation of new <vfsid, inum, lbn> triples
3612 * before the old ones have been deleted.
3613 */
3620 /*
3621 * We defer adding filefree to the worklist until
3622 * all other additions have been made to ensure
3623 * that it will be done after all the old blocks
3624 * have been freed.
3625 */
3628 }
3644 /* -- fall through -- */
3653 /* NOTREACHED */
3654 }
3655 }
3659 }
3662 }
3664 /*
3665 * If no outstanding dependencies, free it.
3666 */
3670 }
3672 /*
3673 * Process a diradd entry after its dependent inode has been written.
3674 * This routine must be called with splbio interrupts blocked.
3675 */
3676 static void
3678 {
3685 else
3689 }
3691 }
3693 /*
3694 * Handle the completion of a mkdir dependency.
3695 */
3696 static void
3698 {
3704 }
3712 else
3716 }
3719 }
3721 /*
3722 * Called from within softdep_disk_write_complete above.
3723 * A write operation was just completed. Removed inodes can
3724 * now be freed and associated block pointers may be committed.
3725 * Note that this routine is always called from interrupt level
3726 * with further splbio interrupts blocked.
3727 *
3728 * Parameters:
3729 * bp: buffer containing the written page
3730 */
3731 static int
3733 {
3741 }
3743 /*
3744 * Process any directory removals that have been committed.
3745 */
3750 }
3751 /*
3752 * Free any directory additions that have been committed.
3753 */
3756 /*
3757 * Uncommitted directory entries must be restored.
3758 */
3765 }
3772 /*
3773 * If the inode referenced by the directory has
3774 * been written out, then the dependency can be
3775 * moved to the pending list.
3776 */
3780 da_pdlist);
3781 }
3782 }
3783 }
3784 /*
3785 * If there were any rollbacks in the directory, then it must be
3786 * marked dirty so that its will eventually get written back in
3787 * its correct form.
3788 */
3791 stat_dir_entry++;
3793 }
3794 /*
3795 * If no dependencies remain, the pagedep will be freed.
3796 * Otherwise it will remain to update the page before it
3797 * is written back to disk.
3798 */
3807 }
3808 }
3810 }
3812 /*
3813 * Writing back in-core inode structures.
3814 *
3815 * The filesystem only accesses an inode's contents when it occupies an
3816 * "in-core" inode structure. These "in-core" structures are separate from
3817 * the page frames used to cache inode blocks. Only the latter are
3818 * transferred to/from the disk. So, when the updated contents of the
3819 * "in-core" inode structure are copied to the corresponding in-memory inode
3820 * block, the dependencies are also transferred. The following procedure is
3821 * called when copying a dirty "in-core" inode to a cached inode block.
3822 */
3824 /*
3825 * Called when an inode is loaded from disk. If the effective link count
3826 * differed from the actual link count when it was last flushed, then we
3827 * need to ensure that the correct effective link count is put back.
3828 *
3829 * Parameters:
3830 * ip: the "in_core" copy of the inode
3831 */
3832 void
3834 {
3837 /*
3838 * Check for alternate nlink count.
3839 */
3845 }
3848 }
3850 /*
3851 * This routine is called just before the "in-core" inode
3852 * information is to be copied to the in-memory inode block.
3853 * Recall that an inode block contains several inodes. If
3854 * the force flag is set, then the dependencies will be
3855 * cleared so that the update can always be made. Note that
3856 * the buffer is locked when this routine is called, so we
3857 * will never be in the middle of writing the inode block
3858 * to disk.
3859 *
3860 * Parameters:
3861 * ip: the "in_core" copy of the inode
3862 * bp: the buffer containing the inode block
3863 * waitfor: nonzero => update must be allowed
3864 */
3865 void
3868 {
3874 /*
3875 * If the effective link count is not equal to the actual link
3876 * count, then we must track the difference in an inodedep while
3877 * the inode is (potentially) tossed out of the cache. Otherwise,
3878 * if there is no existing inodedep, then there are no dependencies
3879 * to track.
3880 */
3887 }
3890 }
3891 /*
3892 * Changes have been initiated. Anything depending on these
3893 * changes cannot occur until this inode has been written.
3894 */
3898 /*
3899 * Any new dependencies associated with the incore inode must
3900 * now be moved to the list associated with the buffer holding
3901 * the in-memory copy of the inode. Once merged process any
3902 * allocdirects that are completed by the merger.
3903 */
3907 /*
3908 * Now that the inode has been pushed into the buffer, the
3909 * operations dependent on the inode being written to disk
3910 * can be moved to the id_bufwait so that they will be
3911 * processed when the buffer I/O completes.
3912 */
3916 }
3917 /*
3918 * Newly allocated inodes cannot be written until the bitmap
3919 * that allocates them have been written (indicated by
3920 * DEPCOMPLETE being set in id_state). If we are doing a
3921 * forced sync (e.g., an fsync on a file), we force the bitmap
3922 * to be written so that the update can be done.
3923 */
3927 }
3928 retry:
3932 }
3939 }
3944 }
3946 /*
3947 * Merge the new inode dependency list (id_newinoupdt) into the old
3948 * inode dependency list (id_inoupdt). This routine must be called
3949 * with splbio interrupts blocked.
3950 */
3951 static void
3953 {
3961 }
3966 listadp);
3968 }
3970 }
3974 }
3975 }
3977 /*
3978 * If we are doing an fsync, then we must ensure that any directory
3979 * entries for the inode have been written after the inode gets to disk.
3980 *
3981 * bioops callback - hold io_token
3982 *
3983 * Parameters:
3984 * vp: the "in_core" copy of the inode
3985 */
3986 static int
3988 {
3999 ino_t parentino;
4000 ufs_lbn_t lbn;
4002 /*
4003 * Move check from original kernel code, possibly not needed any
4004 * more with the per-mount bioops.
4005 */
4015 }
4021 }
4028 }
4030 /*
4031 * Flush our parent if this directory entry
4032 * has a MKDIR_PARENT dependency.
4033 */
4036 else
4043 }
4045 /*
4046 * If we are being fsync'ed as part of vgone'ing this vnode,
4047 * then we will not be able to release and recover the
4048 * vnode below, so we just have to give up on writing its
4049 * directory entry out. It will eventually be written, just
4050 * not now, but then the user was not asking to have it
4051 * written, so we are not breaking any promises.
4052 */
4055 /*
4056 * We prevent deadlock by always fetching inodes from the
4057 * root, moving down the directory tree. Thus, when fetching
4058 * our parent directory, we must unlock ourselves before
4059 * requesting the lock on our parent. See the comment in
4060 * ufs_lookup for details on possible races.
4061 */
4068 }
4073 }
4074 }
4075 /*
4076 * Flush directory page containing the inode's name.
4077 */
4084 }
4088 }
4091 }
4093 /*
4094 * Flush all the dirty bitmaps associated with the block device
4095 * before flushing the rest of the dirty blocks so as to reduce
4096 * the number of dependencies that will have to be rolled back.
4097 */
4100 void
4102 {
4112 }
4114 static int
4116 {
4120 /*
4121 * If it is already scheduled, skip to the next buffer.
4122 */
4127 kprintf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp);
4129 }
4130 /*
4131 * We are only interested in bitmaps with outstanding
4132 * dependencies.
4133 */
4138 }
4144 }
4146 /*
4147 * This routine is called when we are trying to synchronously flush a
4148 * file. This routine must eliminate any filesystem metadata dependencies
4149 * so that the syncing routine can succeed by pushing the dirty blocks
4150 * associated with the file. If any I/O errors occur, they are returned.
4151 */
4155 };
4159 int
4161 {
4165 /*
4166 * Check whether this vnode is involved in a filesystem
4167 * that is doing soft dependency processing.
4168 */
4176 /*
4177 * Ensure that any direct block dependencies have been cleared.
4178 */
4183 }
4184 /*
4185 * For most files, the only metadata dependencies are the
4186 * cylinder group maps that allocate their inode or blocks.
4187 * The block allocation dependencies can be found by traversing
4188 * the dependency lists for any buffers that remain on their
4189 * dirty buffer list. The inode allocation dependency will
4190 * be resolved when the inode is updated with MNT_WAIT.
4191 * This work is done in two passes. The first pass grabs most
4192 * of the buffers and begins asynchronously writing them. The
4193 * only way to wait for these asynchronous writes is to sleep
4194 * on the filesystem vnode which may stay busy for a long time
4195 * if the filesystem is active. So, instead, we make a second
4196 * pass over the dependencies blocking on each write. In the
4197 * usual case we will be blocking against a write that we
4198 * initiated, so when it is done the dependency will have been
4199 * resolved. Thus the second pass is expected to end quickly.
4200 */
4202 top:
4203 /*
4204 * We must wait for any I/O in progress to finish so that
4205 * all potential buffers on the dirty list will be visible.
4206 */
4218 }
4220 /*
4221 * The brief unlock is to allow any pent up dependency
4222 * processing to be done. Then proceed with the second pass.
4223 */
4229 }
4231 /*
4232 * If we have managed to get rid of all the dirty buffers,
4233 * then we are done. For certain directories and block
4234 * devices, we may need to do further work.
4235 *
4236 * We must wait for any I/O in progress to finish so that
4237 * all potential buffers on the dirty list will be visible.
4238 */
4243 }
4246 /*
4247 * If we are trying to sync a block device, some of its buffers may
4248 * contain metadata that cannot be written until the contents of some
4249 * partially written files have been written to disk. The only easy
4250 * way to accomplish this is to sync the entire filesystem (luckily
4251 * this happens rarely).
4252 */
4259 }
4261 static int
4263 {
4276 }
4281 }
4283 /*
4284 * As we hold the buffer locked, none of its dependencies
4285 * will disappear.
4286 */
4304 }
4322 }
4327 restart:
4340 }
4343 }
4353 }
4357 /*
4358 * We are trying to sync a directory that may
4359 * have dependencies on both its own metadata
4360 * and/or dependencies on the inodes of any
4361 * recently allocated files. We walk its diradd
4362 * lists pushing out the associated inode.
4363 */
4376 }
4377 }
4381 /*
4382 * This case should never happen if the vnode has
4383 * been properly sync'ed. However, if this function
4384 * is used at a place where the vnode has not yet
4385 * been sync'ed, this dependency can show up. So,
4386 * rather than panic, just flush it.
4387 */
4398 }
4403 /*
4404 * This case should never happen if the vnode has
4405 * been properly sync'ed. However, if this function
4406 * is used at a place where the vnode has not yet
4407 * been sync'ed, this dependency can show up. So,
4408 * rather than panic, just flush it.
4409 *
4410 * nbp can wind up == bp if a device node for the
4411 * same filesystem is being fsynced at the same time,
4412 * leading to a panic if we don't catch the case.
4413 */
4426 }
4433 /* NOTREACHED */
4434 }
4435 }
4440 }
4442 /*
4443 * Flush the dependencies associated with an inodedep.
4444 * Called with splbio blocked.
4445 */
4446 static int
4448 {
4454 /*
4455 * This work is done in two passes. The first pass grabs most
4456 * of the buffers and begins asynchronously writing them. The
4457 * only way to wait for these asynchronous writes is to sleep
4458 * on the filesystem vnode which may stay busy for a long time
4459 * if the filesystem is active. So, instead, we make a second
4460 * pass over the dependencies blocking on each write. In the
4461 * usual case we will be blocking against a write that we
4462 * initiated, so when it is done the dependency will have been
4463 * resolved. Thus the second pass is expected to end quickly.
4464 * We give a brief window at the top of the loop to allow
4465 * any pending I/O to complete.
4466 */
4480 }
4487 }
4490 }
4501 }
4508 }
4511 }
4514 /*
4515 * If pass2, we are done, otherwise do pass 2.
4516 */
4520 }
4521 /*
4522 * Try freeing inodedep in case all dependencies have been removed.
4523 */
4527 }
4529 /*
4530 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
4531 * Called with splbio blocked.
4532 */
4533 static int
4536 {
4544 ino_t inum;
4548 /*
4549 * Flush ourselves if this directory entry
4550 * has a MKDIR_PARENT dependency.
4551 */
4557 /*
4558 * If that cleared dependencies, go on to next.
4559 */
4564 }
4565 }
4566 /*
4567 * A newly allocated directory must have its "." and
4568 * ".." entries written out before its name can be
4569 * committed in its parent. We do not want or need
4570 * the full semantics of a synchronous VOP_FSYNC as
4571 * that may end up here again, once for each directory
4572 * level in the filesystem. Instead, we push the blocks
4573 * and wait for them to clear. We have to fsync twice
4574 * because the first call may choose to defer blocks
4575 * that still have dependencies, but deferral will
4576 * happen at most once.
4577 */
4587 }
4589 /*
4590 * If first block is still dirty with a D_MKDIR
4591 * dependency then it needs to be written now.
4592 */
4599 }
4607 }
4610 mkdir_body_continue:
4612 /* Flushing of first block failed. */
4616 /*
4617 * If that cleared dependencies, go on to next.
4618 */
4623 }
4624 }
4625 /*
4626 * Flush the inode on which the directory entry depends.
4627 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
4628 * the only remaining dependency is that the updated inode
4629 * count must get pushed to disk. The inode has already
4630 * been pushed into its inode buffer (via VOP_UPDATE) at
4631 * the time of the reference count change. So we need only
4632 * locate that buffer, ensure that there will be no rollback
4633 * caused by a bitmap dependency, then write the inode buffer.
4634 */
4635 retry_lookup:
4638 }
4639 /*
4640 * If the inode still has bitmap dependencies,
4641 * push them to disk.
4642 */
4653 }
4654 /*
4655 * If the inode is still sitting in a buffer waiting
4656 * to be written, push it to disk.
4657 */
4666 /*
4667 * If we have failed to get rid of all the dependencies
4668 * then something is seriously wrong.
4669 */
4672 }
4673 }
4677 }
4679 /*
4680 * A large burst of file addition or deletion activity can drive the
4681 * memory load excessively high. First attempt to slow things down
4682 * using the techniques below. If that fails, this routine requests
4683 * the offending operations to fall back to running synchronously
4684 * until the memory load returns to a reasonable level.
4685 */
4686 int
4688 {
4697 }
4699 /*
4700 * If memory utilization has gotten too high, deliberately slow things
4701 * down and speed up the I/O processing.
4702 */
4703 static int
4705 {
4710 /*
4711 * We never hold up the filesystem syncer process.
4712 */
4715 /*
4716 * First check to see if the work list has gotten backlogged.
4717 * If it has, co-opt this process to help clean up two entries.
4718 * Because this process may hold inodes locked, we cannot
4719 * handle any remove requests that might block on a locked
4720 * inode as that could lead to deadlock.
4721 */
4727 }
4729 /*
4730 * If we are resource constrained on inode dependencies, try
4731 * flushing some dirty inodes. Otherwise, we are constrained
4732 * by file deletions, so try accelerating flushes of directories
4733 * with removal dependencies. We would like to do the cleanup
4734 * here, but we probably hold an inode locked at this point and
4735 * that might deadlock against one that we try to clean. So,
4736 * the best that we can do is request the syncer daemon to do
4737 * the cleanup for us.
4738 */
4755 }
4756 /*
4757 * Hopefully the syncer daemon will catch up and awaken us.
4758 * We wait at most tickdelay before proceeding in any case.
4759 */
4763 }
4765 /*
4766 * Flush out a directory with at least one removal dependency in an effort to
4767 * reduce the number of dirrem, freefile, and freeblks dependency structures.
4768 */
4769 static void
4771 {
4778 ino_t ino;
4794 }
4800 }
4801 }
4803 }
4805 /*
4806 * Clear out a block of dirty inodes in an effort to reduce
4807 * the number of inodedep dependency structures.
4808 */
4812 };
4814 static int
4816 {
4822 }
4824 }
4826 static void
4828 {
4839 /*
4840 * Pick a random inode dependency to be cleared.
4841 * We will then gather up all the inodes in its block
4842 * that have dependencies and flush them out.
4843 */
4851 }
4855 }
4856 /*
4857 * Ugly code to find mount point given pointer to superblock.
4858 */
4864 /*
4865 * Find the last inode in the block with dependencies.
4866 */
4871 /*
4872 * Asynchronously push all but the last inode with dependencies.
4873 * Synchronously push the last inode with dependencies to ensure
4874 * that the inode block gets written to free up the inodedeps.
4875 */
4883 }
4891 }
4894 }
4896 }
4898 /*
4899 * Function to determine if the buffer has outstanding dependencies
4900 * that will cause a roll-back if the buffer is written. If wantcount
4901 * is set, return number of dependencies, otherwise just yes or no.
4902 *
4903 * bioops callback - hold io_token
4904 */
4905 static int
4907 {
4925 /* bitmap allocation dependency */
4929 }
4931 /* direct block pointer dependency */
4935 }
4942 /* indirect block pointer dependency */
4946 }
4954 /* directory entry dependency */
4958 }
4959 }
4966 /* never a dependency on these blocks */
4972 /* NOTREACHED */
4973 }
4974 }
4975 out:
4979 }
4981 /*
4982 * Acquire exclusive access to a buffer. Requires softdep lock
4983 * to be held on entry. If waitfor is MNT_WAIT, may release/reacquire
4984 * softdep lock.
4985 *
4986 * Returns 1 if the buffer was locked, 0 if it was not locked or
4987 * if we had to block.
4988 *
4989 * NOTE! In order to return 1 we must acquire the buffer lock prior
4990 * to any release of &lk. Once we release &lk it's all over.
4991 * We may still have to block on the (type-stable) bp in that
4992 * case, but we must then unlock it and return 0.
4993 */
4994 static int
4996 {
5000 /*
5001 * If the contents of *bpp is NULL the caller presumably lost a race.
5002 */
5007 /*
5008 * Try to obtain the buffer lock without deadlocking on &lk.
5009 */
5013 /*
5014 * If the buffer is no longer dirty the OS already wrote it
5015 * out, return failure.
5016 */
5020 }
5022 /*
5023 * Finish nominal buffer locking sequence return success.
5024 */
5027 }
5029 /*
5030 * Failure case.
5031 *
5032 * If we are not being asked to wait, return 0 immediately.
5033 */
5037 /*
5038 * Once we release the softdep lock we can never return success,
5039 * but we still have to block on the type-stable buf for the caller
5040 * to be able to retry without livelocking the system.
5041 *
5042 * The caller will normally retry in this case.
5043 */
5050 }
5052 /*
5053 * Wait for pending output on a vnode to complete.
5054 * Must be called with vnode locked.
5055 */
5056 static void
5058 {
5066 }
5069 }
5071 /*
5072 * Called whenever a buffer that is being invalidated or reallocated
5073 * contains dependencies. This should only happen if an I/O error has
5074 * occurred. The routine is called with the buffer locked.
5075 *
5076 * bioops callback - hold io_token
5077 */
5078 static void
5080 {
5081 /* nothing to do, mp lock not needed */
5086 }
5088 /*
5089 * Function to handle asynchronous write errors in the filesystem.
5090 */
5091 void
5093 {
5094 /* XXX should do something better! */
5096 }