| blob | fd4651af84e44d2dfc6dd4578a7e5297a78f8fe3 |
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/spinlock2.h>
60 #include <sys/syslog.h>
61 #include <sys/vnode.h>
62 #include <sys/conf.h>
63 #include <machine/inttypes.h>
73 #include <sys/buf2.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 {
1052 /*
1053 * Cap it at 100,000, having more just gets kinda silly.
1054 */
1065 }
1067 /*
1068 * Called at mount time to notify the dependency code that a
1069 * filesystem wishes to use it.
1070 */
1071 int
1073 {
1082 /*
1083 * When doing soft updates, the counters in the
1084 * superblock may have gotten out of sync, so we have
1085 * to scan the cylinder groups and recalculate them.
1086 */
1095 }
1103 }
1104 #ifdef DEBUG
1107 #endif
1110 }
1112 /*
1113 * Protecting the freemaps (or bitmaps).
1114 *
1115 * To eliminate the need to execute fsck before mounting a filesystem
1116 * after a power failure, one must (conservatively) guarantee that the
1117 * on-disk copy of the bitmaps never indicate that a live inode or block is
1118 * free. So, when a block or inode is allocated, the bitmap should be
1119 * updated (on disk) before any new pointers. When a block or inode is
1120 * freed, the bitmap should not be updated until all pointers have been
1121 * reset. The latter dependency is handled by the delayed de-allocation
1122 * approach described below for block and inode de-allocation. The former
1123 * dependency is handled by calling the following procedure when a block or
1124 * inode is allocated. When an inode is allocated an "inodedep" is created
1125 * with its DEPCOMPLETE flag cleared until its bitmap is written to disk.
1126 * Each "inodedep" is also inserted into the hash indexing structure so
1127 * that any additional link additions can be made dependent on the inode
1128 * allocation.
1129 *
1130 * The ufs filesystem maintains a number of free block counts (e.g., per
1131 * cylinder group, per cylinder and per <cylinder, rotational position> pair)
1132 * in addition to the bitmaps. These counts are used to improve efficiency
1133 * during allocation and therefore must be consistent with the bitmaps.
1134 * There is no convenient way to guarantee post-crash consistency of these
1135 * counts with simple update ordering, for two main reasons: (1) The counts
1136 * and bitmaps for a single cylinder group block are not in the same disk
1137 * sector. If a disk write is interrupted (e.g., by power failure), one may
1138 * be written and the other not. (2) Some of the counts are located in the
1139 * superblock rather than the cylinder group block. So, we focus our soft
1140 * updates implementation on protecting the bitmaps. When mounting a
1141 * filesystem, we recompute the auxiliary counts from the bitmaps.
1142 */
1144 /*
1145 * Called just after updating the cylinder group block to allocate an inode.
1146 *
1147 * Parameters:
1148 * bp: buffer for cylgroup block with inode map
1149 * ip: inode related to allocation
1150 * newinum: new inode number being allocated
1151 */
1152 void
1154 {
1158 /*
1159 * Create a dependency for the newly allocated inode.
1160 * Panic if it already exists as something is seriously wrong.
1161 * Otherwise add it to the dependency list for the buffer holding
1162 * the cylinder group map from which it was allocated.
1163 */
1167 }
1173 }
1175 /*
1176 * Called just after updating the cylinder group block to
1177 * allocate block or fragment.
1178 *
1179 * Parameters:
1180 * bp: buffer for cylgroup block with block map
1181 * fs: filesystem doing allocation
1182 * newblkno: number of newly allocated block
1183 */
1184 void
1186 ufs_daddr_t newblkno)
1187 {
1191 /*
1192 * Create a dependency for the newly allocated block.
1193 * Add it to the dependency list for the buffer holding
1194 * the cylinder group map from which it was allocated.
1195 */
1202 }
1204 /*
1205 * Find the bmsafemap associated with a cylinder group buffer.
1206 * If none exists, create one. The buffer must be locked when
1207 * this routine is called and this routine must be called with
1208 * splbio interrupts blocked.
1209 */
1212 {
1221 }
1224 M_SOFTDEP_FLAGS);
1235 }
1237 /*
1238 * Direct block allocation dependencies.
1239 *
1240 * When a new block is allocated, the corresponding disk locations must be
1241 * initialized (with zeros or new data) before the on-disk inode points to
1242 * them. Also, the freemap from which the block was allocated must be
1243 * updated (on disk) before the inode's pointer. These two dependencies are
1244 * independent of each other and are needed for all file blocks and indirect
1245 * blocks that are pointed to directly by the inode. Just before the
1246 * "in-core" version of the inode is updated with a newly allocated block
1247 * number, a procedure (below) is called to setup allocation dependency
1248 * structures. These structures are removed when the corresponding
1249 * dependencies are satisfied or when the block allocation becomes obsolete
1250 * (i.e., the file is deleted, the block is de-allocated, or the block is a
1251 * fragment that gets upgraded). All of these cases are handled in
1252 * procedures described later.
1253 *
1254 * When a file extension causes a fragment to be upgraded, either to a larger
1255 * fragment or to a full block, the on-disk location may change (if the
1256 * previous fragment could not simply be extended). In this case, the old
1257 * fragment must be de-allocated, but not until after the inode's pointer has
1258 * been updated. In most cases, this is handled by later procedures, which
1259 * will construct a "freefrag" structure to be added to the workitem queue
1260 * when the inode update is complete (or obsolete). The main exception to
1261 * this is when an allocation occurs while a pending allocation dependency
1262 * (for the same block pointer) remains. This case is handled in the main
1263 * allocation dependency setup procedure by immediately freeing the
1264 * unreferenced fragments.
1265 *
1266 * Parameters:
1267 * ip: inode to which block is being added
1268 * lbn: block pointer within inode
1269 * newblkno: disk block number being added
1270 * oldblkno: previous block number, 0 unless frag
1271 * newsize: size of new block
1272 * oldsize: size of new block
1273 * bp: bp for allocated block
1274 */
1275 void
1279 {
1298 else
1316 }
1322 /* allocating an indirect block */
1325 }
1327 /*
1328 * Allocating a direct block.
1329 *
1330 * If we are allocating a directory block, then we must
1331 * allocate an associated pagedep to track additions and
1332 * deletions.
1333 */
1337 }
1338 }
1339 /*
1340 * The list of allocdirects must be kept in sorted and ascending
1341 * order so that the rollback routines can quickly determine the
1342 * first uncommitted block (the size of the file stored on disk
1343 * ends at the end of the lowest committed fragment, or if there
1344 * are no fragments, at the end of the highest committed block).
1345 * Since files generally grow, the typical case is that the new
1346 * block is to be added at the end of the list. We speed this
1347 * special case by checking against the last allocdirect in the
1348 * list before laboriously traversing the list looking for the
1349 * insertion point.
1350 */
1354 /* insert at end of list */
1360 }
1364 }
1367 }
1368 /* insert in middle of list */
1373 }
1375 /*
1376 * Replace an old allocdirect dependency with a newer one.
1377 * This routine must be called with splbio interrupts blocked.
1378 *
1379 * Parameters:
1380 * adphead: head of list holding allocdirects
1381 * newadp: allocdirect being added
1382 * oldadp: existing allocdirect being checked
1383 */
1384 static void
1388 {
1398 UFS_NDADDR);
1399 }
1402 /*
1403 * If the old dependency had a fragment to free or had never
1404 * previously had a block allocated, then the new dependency
1405 * can immediately post its freefrag and adopt the old freefrag.
1406 * This action is done by swapping the freefrag dependencies.
1407 * The new dependency gains the old one's freefrag, and the
1408 * old one gets the new one and then immediately puts it on
1409 * the worklist when it is freed by free_allocdirect. It is
1410 * not possible to do this swap when the old dependency had a
1411 * non-zero size but no previous fragment to free. This condition
1412 * arises when the new block is an extension of the old block.
1413 * Here, the first part of the fragment allocated to the new
1414 * dependency is part of the block currently claimed on disk by
1415 * the old dependency, so cannot legitimately be freed until the
1416 * conditions for the new dependency are fulfilled.
1417 */
1422 }
1424 }
1426 /*
1427 * Allocate a new freefrag structure if needed.
1428 */
1431 {
1441 M_SOFTDEP_FLAGS);
1450 }
1452 /*
1453 * This workitem de-allocates fragments that were replaced during
1454 * file block allocation.
1455 */
1456 static void
1458 {
1468 }
1470 /*
1471 * Indirect block allocation dependencies.
1472 *
1473 * The same dependencies that exist for a direct block also exist when
1474 * a new block is allocated and pointed to by an entry in a block of
1475 * indirect pointers. The undo/redo states described above are also
1476 * used here. Because an indirect block contains many pointers that
1477 * may have dependencies, a second copy of the entire in-memory indirect
1478 * block is kept. The buffer cache copy is always completely up-to-date.
1479 * The second copy, which is used only as a source for disk writes,
1480 * contains only the safe pointers (i.e., those that have no remaining
1481 * update dependencies). The second copy is freed when all pointers
1482 * are safe. The cache is not allowed to replace indirect blocks with
1483 * pending update dependencies. If a buffer containing an indirect
1484 * block with dependencies is written, these routines will mark it
1485 * dirty again. It can only be successfully written once all the
1486 * dependencies are removed. The ffs_fsync routine in conjunction with
1487 * softdep_sync_metadata work together to get all the dependencies
1488 * removed so that a file can be successfully written to disk. Three
1489 * procedures are used when setting up indirect block pointer
1490 * dependencies. The division is necessary because of the organization
1491 * of the "balloc" routine and because of the distinction between file
1492 * pages and file metadata blocks.
1493 */
1495 /*
1496 * Allocate a new allocindir structure.
1497 *
1498 * Parameters:
1499 * ip: inode for file being extended
1500 * ptrno: offset of pointer in indirect block
1501 * newblkno: disk block number being added
1502 * oldblkno: previous block number, 0 if none
1503 */
1506 ufs_daddr_t oldblkno)
1507 {
1519 }
1521 /*
1522 * Called just before setting an indirect block pointer
1523 * to a newly allocated file page.
1524 *
1525 * Parameters:
1526 * ip: inode for file being extended
1527 * lbn: allocated block number within file
1528 * bp: buffer with indirect blk referencing page
1529 * ptrno: offset of pointer in indirect block
1530 * newblkno: disk block number being added
1531 * oldblkno: previous block number, 0 if none
1532 * nbp: buffer holding allocated page
1533 */
1534 void
1539 {
1545 /*
1546 * If we are allocating a directory page, then we must
1547 * allocate an associated pagedep to track additions and
1548 * deletions.
1549 */
1556 }
1558 /*
1559 * Called just before setting an indirect block pointer to a
1560 * newly allocated indirect block.
1561 * Parameters:
1562 * nbp: newly allocated indirect block
1563 * ip: inode for file being extended
1564 * bp: indirect block referencing allocated block
1565 * ptrno: offset of pointer in indirect block
1566 * newblkno: disk block number being added
1567 */
1568 void
1571 ufs_daddr_t newblkno)
1572 {
1580 }
1582 /*
1583 * Called to finish the allocation of the "aip" allocated
1584 * by one of the two routines above.
1585 *
1586 * Parameters:
1587 * bp: in-memory copy of the indirect block
1588 * ip: inode for file being extended
1589 * aip: allocindir allocated by the above routines
1590 */
1591 static void
1594 {
1611 }
1616 }
1632 }
1636 /*
1637 * Check to see if there is an existing dependency
1638 * for this block. If there is, merge the old
1639 * dependency into the new one.
1640 */
1643 else
1651 }
1657 }
1662 }
1664 /*
1665 * Avoid any possibility of data corruption by
1666 * ensuring that our old version is thrown away.
1667 */
1671 }
1675 M_SOFTDEP_FLAGS);
1683 BUF_CMD_WRITE);
1684 }
1691 }
1692 }
1694 /*
1695 * Block de-allocation dependencies.
1696 *
1697 * When blocks are de-allocated, the on-disk pointers must be nullified before
1698 * the blocks are made available for use by other files. (The true
1699 * requirement is that old pointers must be nullified before new on-disk
1700 * pointers are set. We chose this slightly more stringent requirement to
1701 * reduce complexity.) Our implementation handles this dependency by updating
1702 * the inode (or indirect block) appropriately but delaying the actual block
1703 * de-allocation (i.e., freemap and free space count manipulation) until
1704 * after the updated versions reach stable storage. After the disk is
1705 * updated, the blocks can be safely de-allocated whenever it is convenient.
1706 * This implementation handles only the common case of reducing a file's
1707 * length to zero. Other cases are handled by the conventional synchronous
1708 * write approach.
1709 *
1710 * The ffs implementation with which we worked double-checks
1711 * the state of the block pointers and file size as it reduces
1712 * a file's length. Some of this code is replicated here in our
1713 * soft updates implementation. The freeblks->fb_chkcnt field is
1714 * used to transfer a part of this information to the procedure
1715 * that eventually de-allocates the blocks.
1716 *
1717 * This routine should be called from the routine that shortens
1718 * a file's length, before the inode's size or block pointers
1719 * are modified. It will save the block pointer information for
1720 * later release and zero the inode so that the calling routine
1721 * can release it.
1722 */
1726 };
1730 /*
1731 * Parameters:
1732 * ip: The inode whose length is to be reduced
1733 * length: The new length for the file
1734 */
1735 void
1737 {
1765 }
1769 }
1772 /*
1773 * Push the zero'ed inode to to its disk buffer so that we are free
1774 * to delete its dependencies below. Once the dependencies are gone
1775 * the buffer can be safely released.
1776 */
1783 /*
1784 * Find and eliminate any inode dependencies.
1785 */
1790 }
1791 /*
1792 * Add the freeblks structure to the list of operations that
1793 * must await the zero'ed inode being written to disk. If we
1794 * still have a bitmap dependency (delay == 0), then the inode
1795 * has never been written to disk, so we can process the
1796 * freeblks below once we have deleted the dependencies.
1797 */
1801 /*
1802 * Because the file length has been truncated to zero, any
1803 * pending block allocation dependency structures associated
1804 * with this inode are obsolete and can simply be de-allocated.
1805 * We must first merge the two dependency lists to get rid of
1806 * any duplicate freefrag structures, then purge the merged list.
1807 */
1813 /*
1814 * We must wait for any I/O in progress to finish so that
1815 * all potential buffers on the dirty list will be visible.
1816 * Once they are all there, walk the list and get rid of
1817 * any dependencies.
1818 */
1837 /*
1838 * If the inode with zeroed block pointers is now on disk
1839 * we can start freeing blocks. Add freeblks to the worklist
1840 * instead of calling handle_workitem_freeblocks directly as
1841 * it is more likely that additional IO is needed to complete
1842 * the request here than in the !delay case.
1843 */
1846 }
1849 /*
1850 * If the inode has never been written to disk (delay == 0),
1851 * then we can process the freeblks now that we have deleted
1852 * the dependencies.
1853 */
1856 }
1858 static int
1860 {
1867 }
1872 }
1880 }
1882 /*
1883 * Reclaim any dependency structures from a buffer that is about to
1884 * be reallocated to a new vnode. The buffer must be locked, thus,
1885 * no I/O completion operations can occur while we are manipulating
1886 * its associated dependencies. The mutex is held so that other I/O's
1887 * associated with related dependencies do not occur.
1888 */
1889 static void
1891 {
1905 /*
1906 * None of the indirect pointers will ever be visible,
1907 * so they can simply be tossed. GOINGAWAY ensures
1908 * that allocated pointers will be saved in the buffer
1909 * cache until they are freed. Note that they will
1910 * only be able to be found by their physical address
1911 * since the inode mapping the logical address will
1912 * be gone. The save buffer used for the safe copy
1913 * was allocated in setup_allocindir_phase2 using
1914 * the physical address so it could be used for this
1915 * purpose. Hence we swap the safe copy with the real
1916 * copy, allowing the safe copy to be freed and holding
1917 * on to the real copy for later use in indir_trunc.
1918 *
1919 * NOTE: ir_savebp is relative to the block device
1920 * so b_bio1 contains the device block number.
1921 */
1924 }
1931 }
1940 /*
1941 * None of the directory additions will ever be
1942 * visible, so they can simply be tossed.
1943 */
1950 /*
1951 * Copy any directory remove dependencies to the list
1952 * to be processed after the zero'ed inode is written.
1953 * If the inode has already been written, then they
1954 * can be dumped directly onto the work list.
1955 */
1961 ALLCOMPLETE)
1963 else
1966 }
1980 /* NOTREACHED */
1985 /* NOTREACHED */
1986 }
1987 }
1988 }
1990 /*
1991 * Free an allocdirect. Generate a new freefrag work request if appropriate.
1992 * This routine must be called with splbio interrupts blocked.
1993 */
1994 static void
1997 {
2009 else
2011 }
2013 }
2015 /*
2016 * Prepare an inode to be freed. The actual free operation is not
2017 * done until the zero'ed inode has been written to disk.
2018 */
2019 void
2021 {
2026 /*
2027 * This sets up the inode de-allocation dependency.
2028 */
2030 M_SOFTDEP_FLAGS);
2038 /*
2039 * If the inodedep does not exist, then the zero'ed inode has
2040 * been written to disk. If the allocated inode has never been
2041 * written to disk, then the on-disk inode is zero'ed. In either
2042 * case we can free the file immediately.
2043 */
2050 }
2053 }
2055 /*
2056 * Check to see if an inode has never been written to disk. If
2057 * so free the inodedep and return success, otherwise return failure.
2058 * This routine must be called with splbio interrupts blocked.
2059 *
2060 * If we still have a bitmap dependency, then the inode has never
2061 * been written to disk. Drop the dependency as it is no longer
2062 * necessary since the inode is being deallocated. We set the
2063 * ALLCOMPLETE flags since the bitmap now properly shows that the
2064 * inode is not allocated. Even if the inode is actively being
2065 * written, it has been rolled back to its zero'ed state, so we
2066 * are ensured that a zero inode is what is on the disk. For short
2067 * lived files, this change will usually result in removing all the
2068 * dependencies from the inode so that it can be freed immediately.
2069 */
2070 static int
2072 {
2083 /*
2084 * Another process might be in initiate_write_inodeblock
2085 * trying to allocate memory without holding "Softdep Lock".
2086 */
2099 }
2102 }
2104 }
2106 /*
2107 * Try to free an inodedep structure. Return 1 if it could be freed.
2108 */
2109 static int
2111 {
2126 }
2128 /*
2129 * This workitem routine performs the block de-allocation.
2130 * The workitem is added to the pending list after the updated
2131 * inode block has been written to disk. As mentioned above,
2132 * checks regarding the number of blocks de-allocated (compared
2133 * to the number of blocks allocated for the file) are also
2134 * performed in this function.
2135 */
2136 static void
2138 {
2140 ufs_daddr_t bn;
2159 }
2162 /*
2163 * Indirect blocks first.
2164 */
2173 }
2174 /*
2175 * All direct blocks or frags.
2176 */
2183 }
2185 #ifdef DIAGNOSTIC
2192 }
2194 /*
2195 * Release blocks associated with the inode ip and stored in the indirect
2196 * block at doffset. If level is greater than SINGLE, the block is an
2197 * indirect block and recursive calls to indirtrunc must be used to
2198 * cleanse other indirect blocks.
2199 */
2200 static int
2203 {
2206 ufs_daddr_t nb;
2217 /*
2218 * Get buffer of block pointers to be freed. This routine is not
2219 * called until the zero'ed inode has been written, so it is safe
2220 * to free blocks as they are encountered. Because the inode has
2221 * been zero'ed, calls to bmap on these blocks will fail. So, we
2222 * have to use the on-disk address and the block device for the
2223 * filesystem to look them up. If the file was deleted before its
2224 * indirect blocks were all written to disk, the routine that set
2225 * us up (deallocate_dependencies) will have arranged to leave
2226 * a complete copy of the indirect block in memory for our use.
2227 * Otherwise we have to read the blocks in from the disk.
2228 */
2232 /*
2233 * bp must be ir_savebp, which is held locked for our use.
2234 */
2239 }
2244 }
2251 }
2252 /*
2253 * Recursively free indirect blocks.
2254 */
2264 }
2267 }
2271 }
2273 /*
2274 * Free an allocindir.
2275 * This routine must be called with splbio interrupts blocked.
2276 */
2277 static void
2279 {
2292 else
2295 }
2297 }
2299 /*
2300 * Directory entry addition dependencies.
2301 *
2302 * When adding a new directory entry, the inode (with its incremented link
2303 * count) must be written to disk before the directory entry's pointer to it.
2304 * Also, if the inode is newly allocated, the corresponding freemap must be
2305 * updated (on disk) before the directory entry's pointer. These requirements
2306 * are met via undo/redo on the directory entry's pointer, which consists
2307 * simply of the inode number.
2308 *
2309 * As directory entries are added and deleted, the free space within a
2310 * directory block can become fragmented. The ufs filesystem will compact
2311 * a fragmented directory block to make space for a new entry. When this
2312 * occurs, the offsets of previously added entries change. Any "diradd"
2313 * dependency structures corresponding to these entries must be updated with
2314 * the new offsets.
2315 */
2317 /*
2318 * This routine is called after the in-memory inode's link
2319 * count has been incremented, but before the directory entry's
2320 * pointer to the inode has been set.
2321 *
2322 * Parameters:
2323 * bp: buffer containing directory block
2324 * dp: inode for directory
2325 * diroffset: offset of new entry in directory
2326 * newinum: inode referenced by new directory entry
2327 * newdirbp: non-NULL => contents of new mkdir
2328 */
2329 void
2332 {
2341 /*
2342 * Whiteouts have no dependencies.
2343 */
2348 }
2365 M_SOFTDEP_FLAGS);
2370 M_SOFTDEP_FLAGS);
2374 /*
2375 * Dependency on "." and ".." being written to disk.
2376 */
2383 /*
2384 * Dependency on link count increase for parent directory
2385 */
2394 }
2395 }
2396 /*
2397 * Link into parent directory pagedep to await its being written.
2398 */
2403 da_pdlist);
2404 /*
2405 * Link into its inodedep. Put it on the id_bufwait list if the inode
2406 * is not yet written. If it is written, do the post-inode write
2407 * processing to put it on the id_pendinghd list.
2408 */
2412 else
2415 }
2417 /*
2418 * This procedure is called to change the offset of a directory
2419 * entry when compacting a directory block which must be owned
2420 * exclusively by the caller. Note that the actual entry movement
2421 * must be done in this procedure to ensure that no I/O completions
2422 * occur while the move is in progress.
2423 *
2424 * Parameters:
2425 * dp: inode for directory
2426 * base: address of dp->i_offset
2427 * oldloc: address of old directory location
2428 * newloc: address of new directory location
2429 * entrysize: size of directory entry
2430 */
2431 void
2435 {
2439 ufs_lbn_t lbn;
2459 }
2466 }
2467 }
2468 }
2469 done:
2472 }
2474 /*
2475 * Free a diradd dependency structure. This routine must be called
2476 * with splbio interrupts blocked.
2477 */
2478 static void
2480 {
2497 }
2510 }
2513 }
2514 }
2516 }
2518 /*
2519 * Directory entry removal dependencies.
2520 *
2521 * When removing a directory entry, the entry's inode pointer must be
2522 * zero'ed on disk before the corresponding inode's link count is decremented
2523 * (possibly freeing the inode for re-use). This dependency is handled by
2524 * updating the directory entry but delaying the inode count reduction until
2525 * after the directory block has been written to disk. After this point, the
2526 * inode count can be decremented whenever it is convenient.
2527 */
2529 /*
2530 * This routine should be called immediately after removing
2531 * a directory entry. The inode's link count should not be
2532 * decremented by the calling procedure -- the soft updates
2533 * code will do this task when it is safe.
2534 *
2535 * Parameters:
2536 * bp: buffer containing directory block
2537 * dp: inode for the directory being modified
2538 * ip: inode for directory entry being removed
2539 * isrmdir: indicates if doing RMDIR
2540 */
2541 void
2544 {
2547 /*
2548 * Allocate a new dirrem if appropriate and ACQUIRE_LOCK.
2549 */
2552 /*
2553 * If the COMPLETE flag is clear, then there were no active
2554 * entries and we want to roll back to a zeroed entry until
2555 * the new inode is committed to disk. If the COMPLETE flag is
2556 * set then we have deleted an entry that never made it to
2557 * disk. If the entry we deleted resulted from a name change,
2558 * then the old name still resides on disk. We cannot delete
2559 * its inode (returned to us in prevdirrem) until the zeroed
2560 * directory entry gets to disk. The new inode has never been
2561 * referenced on the disk, so can be deleted immediately.
2562 */
2565 dm_next);
2574 }
2575 }
2577 /*
2578 * Allocate a new dirrem if appropriate and return it along with
2579 * its associated pagedep. Called without a lock, returns with lock.
2580 */
2583 /*
2584 * Parameters:
2585 * bp: buffer containing directory block
2586 * dp: inode for the directory being modified
2587 * ip: inode for directory entry being removed
2588 * isrmdir: indicates if doing RMDIR
2589 * prevdirremp: previously referenced inode, if any
2590 */
2594 {
2596 ufs_lbn_t lbn;
2601 /*
2602 * Whiteouts have no deletion dependencies.
2603 */
2606 /*
2607 * If we are over our limit, try to improve the situation.
2608 * Limiting the number of dirrem structures will also limit
2609 * the number of freefile and freeblks structures.
2610 */
2617 }
2634 /*
2635 * Check for a diradd dependency for the same directory entry.
2636 * If present, then both dependencies become obsolete and can
2637 * be de-allocated. Check for an entry on both the pd_dirraddhd
2638 * list and the pd_pendinghd list.
2639 */
2651 }
2652 /*
2653 * Must be ATTACHED at this point.
2654 */
2657 }
2661 }
2662 /*
2663 * If we are deleting a changed name that never made it to disk,
2664 * then return the dirrem describing the previous inode (which
2665 * represents the inode currently referenced from this entry on disk).
2666 */
2671 }
2672 /*
2673 * We are deleting an entry that never made it to disk.
2674 * Mark it COMPLETE so we can delete its inode immediately.
2675 */
2679 }
2681 /*
2682 * Directory entry change dependencies.
2683 *
2684 * Changing an existing directory entry requires that an add operation
2685 * be completed first followed by a deletion. The semantics for the addition
2686 * are identical to the description of adding a new entry above except
2687 * that the rollback is to the old inode number rather than zero. Once
2688 * the addition dependency is completed, the removal is done as described
2689 * in the removal routine above.
2690 */
2692 /*
2693 * This routine should be called immediately after changing
2694 * a directory entry. The inode's link count should not be
2695 * decremented by the calling procedure -- the soft updates
2696 * code will perform this task when it is safe.
2697 *
2698 * Parameters:
2699 * bp: buffer containing directory block
2700 * dp: inode for the directory being modified
2701 * ip: inode for directory entry being removed
2702 * newinum: new inode number for changed entry
2703 * isrmdir: indicates if doing RMDIR
2704 */
2705 void
2709 {
2718 /*
2719 * Whiteouts do not need diradd dependencies.
2720 */
2728 }
2730 /*
2731 * Allocate a new dirrem and ACQUIRE_LOCK.
2732 */
2735 /*
2736 * The possible values for isrmdir:
2737 * 0 - non-directory file rename
2738 * 1 - directory rename within same directory
2739 * inum - directory rename to new directory of given inode number
2740 * When renaming to a new directory, we are both deleting and
2741 * creating a new directory entry, so the link count on the new
2742 * directory should not change. Thus we do not need the followup
2743 * dirrem which is usually done in handle_workitem_remove. We set
2744 * the DIRCHG flag to tell handle_workitem_remove to skip the
2745 * followup dirrem.
2746 */
2750 /*
2751 * Whiteouts have no additional dependencies,
2752 * so just put the dirrem on the correct list.
2753 */
2757 dm_next);
2761 }
2764 }
2766 /*
2767 * If the COMPLETE flag is clear, then there were no active
2768 * entries and we want to roll back to the previous inode until
2769 * the new inode is committed to disk. If the COMPLETE flag is
2770 * set, then we have deleted an entry that never made it to disk.
2771 * If the entry we deleted resulted from a name change, then the old
2772 * inode reference still resides on disk. Any rollback that we do
2773 * needs to be to that old inode (returned to us in prevdirrem). If
2774 * the entry we deleted resulted from a create, then there is
2775 * no entry on the disk, so we want to roll back to zero rather
2776 * than the uncommitted inode. In either of the COMPLETE cases we
2777 * want to immediately free the unwritten and unreferenced inode.
2778 */
2787 }
2790 }
2791 /*
2792 * Link into its inodedep. Put it on the id_bufwait list if the inode
2793 * is not yet written. If it is written, do the post-inode write
2794 * processing to put it on the id_pendinghd list.
2795 */
2805 }
2807 }
2809 /*
2810 * Called whenever the link count on an inode is changed.
2811 * It creates an inode dependency so that the new reference(s)
2812 * to the inode cannot be committed to disk until the updated
2813 * inode has been written.
2814 *
2815 * Parameters:
2816 * ip: the inode with the increased link count
2817 */
2818 void
2820 {
2827 }
2830 }
2832 /*
2833 * This workitem decrements the inode's link count.
2834 * If the link count reaches zero, the file is removed.
2835 */
2836 static void
2838 {
2842 ino_t oldinum;
2849 }
2854 }
2855 /*
2856 * Normal file deletion.
2857 */
2863 }
2870 }
2871 /*
2872 * Directory deletion. Decrement reference count for both the
2873 * just deleted parent directory entry and the reference for ".".
2874 * Next truncate the directory to length zero. When the
2875 * truncation completes, arrange to have the reference count on
2876 * the parent decremented to account for the loss of "..".
2877 */
2882 }
2887 /*
2888 * Rename a directory to a new parent. Since, we are both deleting
2889 * and creating a new directory entry, the link count on the new
2890 * directory should not change. Thus we skip the followup dirrem.
2891 */
2897 }
2898 /*
2899 * If the inodedep does not exist, then the zero'ed inode has
2900 * been written to disk. If the allocated inode has never been
2901 * written to disk, then the on-disk inode is zero'ed. In either
2902 * case we can remove the file immediately.
2903 */
2914 }
2920 }
2922 /*
2923 * Inode de-allocation dependencies.
2924 *
2925 * When an inode's link count is reduced to zero, it can be de-allocated. We
2926 * found it convenient to postpone de-allocation until after the inode is
2927 * written to disk with its new link count (zero). At this point, all of the
2928 * on-disk inode's block pointers are nullified and, with careful dependency
2929 * list ordering, all dependencies related to the inode will be satisfied and
2930 * the corresponding dependency structures de-allocated. So, if/when the
2931 * inode is reused, there will be no mixing of old dependencies with new
2932 * ones. This artificial dependency is set up by the block de-allocation
2933 * procedure above (softdep_setup_freeblocks) and completed by the
2934 * following procedure.
2935 */
2936 static void
2938 {
2944 #ifdef DEBUG
2950 #endif
2958 }
2960 /*
2961 * Helper function which unlinks marker element from work list and returns
2962 * the next element on the list.
2963 */
2966 {
2972 }
2974 /*
2975 * checkread, checkwrite
2976 *
2977 * bioops callback - hold io_token
2978 */
2979 static int
2981 {
2982 /* nothing to do, mp lock not needed */
2984 }
2986 /*
2987 * bioops callback - hold io_token
2988 */
2989 static int
2991 {
2992 /* nothing to do, mp lock not needed */
2994 }
2996 /*
2997 * Disk writes.
2998 *
2999 * The dependency structures constructed above are most actively used when file
3000 * system blocks are written to disk. No constraints are placed on when a
3001 * block can be written, but unsatisfied update dependencies are made safe by
3002 * modifying (or replacing) the source memory for the duration of the disk
3003 * write. When the disk write completes, the memory block is again brought
3004 * up-to-date.
3005 *
3006 * In-core inode structure reclamation.
3007 *
3008 * Because there are a finite number of "in-core" inode structures, they are
3009 * reused regularly. By transferring all inode-related dependencies to the
3010 * in-memory inode block and indexing them separately (via "inodedep"s), we
3011 * can allow "in-core" inode structures to be reused at any time and avoid
3012 * any increase in contention.
3013 *
3014 * Called just before entering the device driver to initiate a new disk I/O.
3015 * The buffer must be locked, thus, no I/O completion operations can occur
3016 * while we are manipulating its associated dependencies.
3017 *
3018 * bioops callback - hold io_token
3019 *
3020 * Parameters:
3021 * bp: structure describing disk write to occur
3022 */
3023 static void
3025 {
3030 /*
3031 * We only care about write operations. There should never
3032 * be dependencies for reads.
3033 */
3040 /*
3041 * Do any necessary pre-I/O processing.
3042 */
3059 /*
3060 * If there are no remaining dependencies, this
3061 * will be writing the real pointers, so the
3062 * dependency can be freed.
3063 */
3067 /* inline expand WORKLIST_REMOVE(wk); */
3072 }
3073 /*
3074 * Replace up-to-date version with safe version.
3075 */
3077 M_INDIRDEP,
3078 M_SOFTDEP_FLAGS);
3097 /* NOTREACHED */
3098 }
3099 }
3101 }
3103 /*
3104 * Called from within the procedure above to deal with unsatisfied
3105 * allocation dependencies in a directory. The buffer must be locked,
3106 * thus, no I/O completion operations can occur while we are
3107 * manipulating its associated dependencies.
3108 */
3109 static void
3111 {
3117 /*
3118 * This can only happen if there is a driver that does not
3119 * understand chaining. Here biodone will reissue the call
3120 * to strategy for the incomplete buffers.
3121 */
3124 }
3135 }
3138 else
3142 }
3143 }
3145 }
3147 /*
3148 * Called from within the procedure above to deal with unsatisfied
3149 * allocation dependencies in an inodeblock. The buffer must be
3150 * locked, thus, no I/O completion operations can occur while we
3151 * are manipulating its associated dependencies.
3152 *
3153 * Parameters:
3154 * bp: The inode block
3155 */
3156 static void
3158 {
3172 /*
3173 * If the bitmap is not yet written, then the allocated
3174 * inode cannot be written to disk.
3175 */
3180 M_SOFTDEP_FLAGS);
3186 }
3187 /*
3188 * If no dependencies, then there is nothing to roll back.
3189 */
3193 /*
3194 * Set the dependencies to busy.
3195 */
3199 #ifdef DIAGNOSTIC
3202 }
3209 }
3217 }
3222 }
3226 }
3227 /*
3228 * The on-disk inode cannot claim to be any larger than the last
3229 * fragment that has been written. Otherwise, the on-disk inode
3230 * might have fragments that were not the last block in the file
3231 * which would corrupt the filesystem.
3232 */
3238 /* keep going until hitting a rollback to a frag */
3243 #ifdef DIAGNOSTIC
3246 }
3249 }
3251 #ifdef DIAGNOSTIC
3255 }
3258 }
3261 }
3262 /*
3263 * If we have zero'ed out the last allocated block of the file,
3264 * roll back the size to the last currently allocated block.
3265 * We know that this last allocated block is a full-sized as
3266 * we already checked for fragments in the loop above.
3267 */
3274 }
3275 /*
3276 * The only dependencies are for indirect blocks.
3277 *
3278 * The file size for indirect block additions is not guaranteed.
3279 * Such a guarantee would be non-trivial to achieve. The conventional
3280 * synchronous write implementation also does not make this guarantee.
3281 * Fsck should catch and fix discrepancies. Arguably, the file size
3282 * can be over-estimated without destroying integrity when the file
3283 * moves into the indirect blocks (i.e., is large). If we want to
3284 * postpone fsck, we are stuck with this argument.
3285 */
3289 }
3291 /*
3292 * This routine is called during the completion interrupt
3293 * service routine for a disk write (from the procedure called
3294 * by the device driver to inform the filesystem caches of
3295 * a request completion). It should be called early in this
3296 * procedure, before the block is made available to other
3297 * processes or other routines are called.
3298 *
3299 * bioops callback - hold io_token
3300 *
3301 * Parameters:
3302 * bp: describes the completed disk write
3303 */
3304 static void
3306 {
3339 }
3346 }
3353 }
3359 }
3383 }
3393 }
3394 }
3397 stat_indir_blk_ptrs++;
3404 /* NOTREACHED */
3405 }
3406 }
3407 /*
3408 * Reattach any requests that must be redone.
3409 */
3413 }
3416 }
3418 /*
3419 * Called from within softdep_disk_write_complete above. Note that
3420 * this routine is always called from interrupt level with further
3421 * splbio interrupts blocked.
3422 *
3423 * Parameters:
3424 * adp: the completed allocdirect
3425 */
3426 static void
3428 {
3438 /*
3439 * The on-disk inode cannot claim to be any larger than the last
3440 * fragment that has been written. Otherwise, the on-disk inode
3441 * might have fragments that were not the last block in the file
3442 * which would corrupt the filesystem. Thus, we cannot free any
3443 * allocdirects after one whose ad_oldblkno claims a fragment as
3444 * these blocks must be rolled back to zero before writing the inode.
3445 * We check the currently active set of allocdirects in id_inoupdt.
3446 */
3450 /* found our block */
3453 /* continue if ad_oldlbn is not a fragment */
3457 /* hit a fragment */
3459 }
3460 /*
3461 * If we have reached the end of the current list without
3462 * finding the just finished dependency, then it must be
3463 * on the future dependency list. Future dependencies cannot
3464 * be freed until they are moved to the current list.
3465 */
3467 #ifdef DEBUG
3469 /* found our block */
3476 }
3477 /*
3478 * If we have found the just finished dependency, then free
3479 * it along with anything that follows it that is complete.
3480 */
3486 }
3487 }
3489 /*
3490 * Called from within softdep_disk_write_complete above. Note that
3491 * this routine is always called from interrupt level with further
3492 * splbio interrupts blocked.
3493 *
3494 * Parameters:
3495 * aip: the completed allocindir
3496 */
3497 static void
3499 {
3512 }
3519 }
3521 /*
3522 * Called from within softdep_disk_write_complete above to restore
3523 * in-memory inode block contents to their most up-to-date state. Note
3524 * that this routine is always called from interrupt level with further
3525 * splbio interrupts blocked.
3526 *
3527 * Parameters:
3528 * bp: buffer containing the inode block
3529 */
3530 static int
3532 {
3544 /*
3545 * If we had to rollback the inode allocation because of
3546 * bitmaps being incomplete, then simply restore it.
3547 * Keep the block dirty so that it will not be reclaimed until
3548 * all associated dependencies have been cleared and the
3549 * corresponding updates written to disk.
3550 */
3556 stat_inode_bitmap++;
3559 }
3561 /*
3562 * Roll forward anything that had to be rolled back before
3563 * the inode could be updated.
3564 */
3577 }
3586 }
3588 }
3592 }
3594 stat_direct_blk_ptrs++;
3595 /*
3596 * Reset the file size to its most up-to-date value.
3597 */
3600 }
3604 }
3606 /*
3607 * If there were any rollbacks in the inode block, then it must be
3608 * marked dirty so that its will eventually get written back in
3609 * its correct form.
3610 */
3613 /*
3614 * Process any allocdirects that completed during the update.
3615 */
3618 /*
3619 * Process deallocations that were held pending until the
3620 * inode had been written to disk. Freeing of the inode
3621 * is delayed until after all blocks have been freed to
3622 * avoid creation of new <vfsid, inum, lbn> triples
3623 * before the old ones have been deleted.
3624 */
3631 /*
3632 * We defer adding filefree to the worklist until
3633 * all other additions have been made to ensure
3634 * that it will be done after all the old blocks
3635 * have been freed.
3636 */
3639 }
3655 /* -- fall through -- */
3664 /* NOTREACHED */
3665 }
3666 }
3670 }
3673 }
3675 /*
3676 * If no outstanding dependencies, free it.
3677 */
3681 }
3683 /*
3684 * Process a diradd entry after its dependent inode has been written.
3685 * This routine must be called with splbio interrupts blocked.
3686 */
3687 static void
3689 {
3696 else
3700 }
3702 }
3704 /*
3705 * Handle the completion of a mkdir dependency.
3706 */
3707 static void
3709 {
3715 }
3723 else
3727 }
3730 }
3732 /*
3733 * Called from within softdep_disk_write_complete above.
3734 * A write operation was just completed. Removed inodes can
3735 * now be freed and associated block pointers may be committed.
3736 * Note that this routine is always called from interrupt level
3737 * with further splbio interrupts blocked.
3738 *
3739 * Parameters:
3740 * bp: buffer containing the written page
3741 */
3742 static int
3744 {
3752 }
3754 /*
3755 * Process any directory removals that have been committed.
3756 */
3761 }
3762 /*
3763 * Free any directory additions that have been committed.
3764 */
3767 /*
3768 * Uncommitted directory entries must be restored.
3769 */
3776 }
3783 /*
3784 * If the inode referenced by the directory has
3785 * been written out, then the dependency can be
3786 * moved to the pending list.
3787 */
3791 da_pdlist);
3792 }
3793 }
3794 }
3795 /*
3796 * If there were any rollbacks in the directory, then it must be
3797 * marked dirty so that its will eventually get written back in
3798 * its correct form.
3799 */
3802 stat_dir_entry++;
3804 }
3805 /*
3806 * If no dependencies remain, the pagedep will be freed.
3807 * Otherwise it will remain to update the page before it
3808 * is written back to disk.
3809 */
3818 }
3819 }
3821 }
3823 /*
3824 * Writing back in-core inode structures.
3825 *
3826 * The filesystem only accesses an inode's contents when it occupies an
3827 * "in-core" inode structure. These "in-core" structures are separate from
3828 * the page frames used to cache inode blocks. Only the latter are
3829 * transferred to/from the disk. So, when the updated contents of the
3830 * "in-core" inode structure are copied to the corresponding in-memory inode
3831 * block, the dependencies are also transferred. The following procedure is
3832 * called when copying a dirty "in-core" inode to a cached inode block.
3833 */
3835 /*
3836 * Called when an inode is loaded from disk. If the effective link count
3837 * differed from the actual link count when it was last flushed, then we
3838 * need to ensure that the correct effective link count is put back.
3839 *
3840 * Parameters:
3841 * ip: the "in_core" copy of the inode
3842 */
3843 void
3845 {
3848 /*
3849 * Check for alternate nlink count.
3850 */
3856 }
3859 }
3861 /*
3862 * This routine is called just before the "in-core" inode
3863 * information is to be copied to the in-memory inode block.
3864 * Recall that an inode block contains several inodes. If
3865 * the force flag is set, then the dependencies will be
3866 * cleared so that the update can always be made. Note that
3867 * the buffer is locked when this routine is called, so we
3868 * will never be in the middle of writing the inode block
3869 * to disk.
3870 *
3871 * Parameters:
3872 * ip: the "in_core" copy of the inode
3873 * bp: the buffer containing the inode block
3874 * waitfor: nonzero => update must be allowed
3875 */
3876 void
3879 {
3885 /*
3886 * If the effective link count is not equal to the actual link
3887 * count, then we must track the difference in an inodedep while
3888 * the inode is (potentially) tossed out of the cache. Otherwise,
3889 * if there is no existing inodedep, then there are no dependencies
3890 * to track.
3891 */
3898 }
3901 }
3902 /*
3903 * Changes have been initiated. Anything depending on these
3904 * changes cannot occur until this inode has been written.
3905 */
3909 /*
3910 * Any new dependencies associated with the incore inode must
3911 * now be moved to the list associated with the buffer holding
3912 * the in-memory copy of the inode. Once merged process any
3913 * allocdirects that are completed by the merger.
3914 */
3918 /*
3919 * Now that the inode has been pushed into the buffer, the
3920 * operations dependent on the inode being written to disk
3921 * can be moved to the id_bufwait so that they will be
3922 * processed when the buffer I/O completes.
3923 */
3927 }
3928 /*
3929 * Newly allocated inodes cannot be written until the bitmap
3930 * that allocates them have been written (indicated by
3931 * DEPCOMPLETE being set in id_state). If we are doing a
3932 * forced sync (e.g., an fsync on a file), we force the bitmap
3933 * to be written so that the update can be done.
3934 */
3938 }
3939 retry:
3943 }
3950 }
3955 }
3957 /*
3958 * Merge the new inode dependency list (id_newinoupdt) into the old
3959 * inode dependency list (id_inoupdt). This routine must be called
3960 * with splbio interrupts blocked.
3961 */
3962 static void
3964 {
3972 }
3977 listadp);
3979 }
3981 }
3985 }
3986 }
3988 /*
3989 * If we are doing an fsync, then we must ensure that any directory
3990 * entries for the inode have been written after the inode gets to disk.
3991 *
3992 * bioops callback - hold io_token
3993 *
3994 * Parameters:
3995 * vp: the "in_core" copy of the inode
3996 */
3997 static int
3999 {
4010 ino_t parentino;
4011 ufs_lbn_t lbn;
4013 /*
4014 * Move check from original kernel code, possibly not needed any
4015 * more with the per-mount bioops.
4016 */
4026 }
4032 }
4039 }
4041 /*
4042 * Flush our parent if this directory entry
4043 * has a MKDIR_PARENT dependency.
4044 */
4047 else
4054 }
4056 /*
4057 * If we are being fsync'ed as part of vgone'ing this vnode,
4058 * then we will not be able to release and recover the
4059 * vnode below, so we just have to give up on writing its
4060 * directory entry out. It will eventually be written, just
4061 * not now, but then the user was not asking to have it
4062 * written, so we are not breaking any promises.
4063 */
4066 /*
4067 * We prevent deadlock by always fetching inodes from the
4068 * root, moving down the directory tree. Thus, when fetching
4069 * our parent directory, we must unlock ourselves before
4070 * requesting the lock on our parent. See the comment in
4071 * ufs_lookup for details on possible races.
4072 */
4079 }
4084 }
4085 }
4086 /*
4087 * Flush directory page containing the inode's name.
4088 */
4095 }
4099 }
4102 }
4104 /*
4105 * Flush all the dirty bitmaps associated with the block device
4106 * before flushing the rest of the dirty blocks so as to reduce
4107 * the number of dependencies that will have to be rolled back.
4108 */
4111 void
4113 {
4123 }
4125 static int
4127 {
4131 /*
4132 * If it is already scheduled, skip to the next buffer.
4133 */
4138 kprintf("softdep_fsync_mountdev_bp: warning, buffer %p ripped out from under vnode %p\n", bp, vp);
4140 }
4141 /*
4142 * We are only interested in bitmaps with outstanding
4143 * dependencies.
4144 */
4149 }
4155 }
4157 /*
4158 * This routine is called when we are trying to synchronously flush a
4159 * file. This routine must eliminate any filesystem metadata dependencies
4160 * so that the syncing routine can succeed by pushing the dirty blocks
4161 * associated with the file. If any I/O errors occur, they are returned.
4162 */
4166 };
4170 int
4172 {
4176 /*
4177 * Check whether this vnode is involved in a filesystem
4178 * that is doing soft dependency processing.
4179 */
4187 /*
4188 * Ensure that any direct block dependencies have been cleared.
4189 */
4194 }
4195 /*
4196 * For most files, the only metadata dependencies are the
4197 * cylinder group maps that allocate their inode or blocks.
4198 * The block allocation dependencies can be found by traversing
4199 * the dependency lists for any buffers that remain on their
4200 * dirty buffer list. The inode allocation dependency will
4201 * be resolved when the inode is updated with MNT_WAIT.
4202 * This work is done in two passes. The first pass grabs most
4203 * of the buffers and begins asynchronously writing them. The
4204 * only way to wait for these asynchronous writes is to sleep
4205 * on the filesystem vnode which may stay busy for a long time
4206 * if the filesystem is active. So, instead, we make a second
4207 * pass over the dependencies blocking on each write. In the
4208 * usual case we will be blocking against a write that we
4209 * initiated, so when it is done the dependency will have been
4210 * resolved. Thus the second pass is expected to end quickly.
4211 */
4213 top:
4214 /*
4215 * We must wait for any I/O in progress to finish so that
4216 * all potential buffers on the dirty list will be visible.
4217 */
4229 }
4231 /*
4232 * The brief unlock is to allow any pent up dependency
4233 * processing to be done. Then proceed with the second pass.
4234 */
4240 }
4242 /*
4243 * If we have managed to get rid of all the dirty buffers,
4244 * then we are done. For certain directories and block
4245 * devices, we may need to do further work.
4246 *
4247 * We must wait for any I/O in progress to finish so that
4248 * all potential buffers on the dirty list will be visible.
4249 */
4254 }
4257 /*
4258 * If we are trying to sync a block device, some of its buffers may
4259 * contain metadata that cannot be written until the contents of some
4260 * partially written files have been written to disk. The only easy
4261 * way to accomplish this is to sync the entire filesystem (luckily
4262 * this happens rarely).
4263 */
4270 }
4272 static int
4274 {
4287 }
4292 }
4294 /*
4295 * As we hold the buffer locked, none of its dependencies
4296 * will disappear.
4297 */
4315 }
4333 }
4338 restart:
4351 }
4354 }
4364 }
4368 /*
4369 * We are trying to sync a directory that may
4370 * have dependencies on both its own metadata
4371 * and/or dependencies on the inodes of any
4372 * recently allocated files. We walk its diradd
4373 * lists pushing out the associated inode.
4374 */
4387 }
4388 }
4392 /*
4393 * This case should never happen if the vnode has
4394 * been properly sync'ed. However, if this function
4395 * is used at a place where the vnode has not yet
4396 * been sync'ed, this dependency can show up. So,
4397 * rather than panic, just flush it.
4398 */
4409 }
4414 /*
4415 * This case should never happen if the vnode has
4416 * been properly sync'ed. However, if this function
4417 * is used at a place where the vnode has not yet
4418 * been sync'ed, this dependency can show up. So,
4419 * rather than panic, just flush it.
4420 *
4421 * nbp can wind up == bp if a device node for the
4422 * same filesystem is being fsynced at the same time,
4423 * leading to a panic if we don't catch the case.
4424 */
4437 }
4444 /* NOTREACHED */
4445 }
4446 }
4451 }
4453 /*
4454 * Flush the dependencies associated with an inodedep.
4455 * Called with splbio blocked.
4456 */
4457 static int
4459 {
4465 /*
4466 * This work is done in two passes. The first pass grabs most
4467 * of the buffers and begins asynchronously writing them. The
4468 * only way to wait for these asynchronous writes is to sleep
4469 * on the filesystem vnode which may stay busy for a long time
4470 * if the filesystem is active. So, instead, we make a second
4471 * pass over the dependencies blocking on each write. In the
4472 * usual case we will be blocking against a write that we
4473 * initiated, so when it is done the dependency will have been
4474 * resolved. Thus the second pass is expected to end quickly.
4475 * We give a brief window at the top of the loop to allow
4476 * any pending I/O to complete.
4477 */
4491 }
4498 }
4501 }
4512 }
4519 }
4522 }
4525 /*
4526 * If pass2, we are done, otherwise do pass 2.
4527 */
4531 }
4532 /*
4533 * Try freeing inodedep in case all dependencies have been removed.
4534 */
4538 }
4540 /*
4541 * Eliminate a pagedep dependency by flushing out all its diradd dependencies.
4542 * Called with splbio blocked.
4543 */
4544 static int
4547 {
4555 ino_t inum;
4559 /*
4560 * Flush ourselves if this directory entry
4561 * has a MKDIR_PARENT dependency.
4562 */
4568 /*
4569 * If that cleared dependencies, go on to next.
4570 */
4575 }
4576 }
4577 /*
4578 * A newly allocated directory must have its "." and
4579 * ".." entries written out before its name can be
4580 * committed in its parent. We do not want or need
4581 * the full semantics of a synchronous VOP_FSYNC as
4582 * that may end up here again, once for each directory
4583 * level in the filesystem. Instead, we push the blocks
4584 * and wait for them to clear. We have to fsync twice
4585 * because the first call may choose to defer blocks
4586 * that still have dependencies, but deferral will
4587 * happen at most once.
4588 */
4598 }
4600 /*
4601 * If first block is still dirty with a D_MKDIR
4602 * dependency then it needs to be written now.
4603 */
4610 }
4618 }
4621 mkdir_body_continue:
4623 /* Flushing of first block failed. */
4627 /*
4628 * If that cleared dependencies, go on to next.
4629 */
4634 }
4635 }
4636 /*
4637 * Flush the inode on which the directory entry depends.
4638 * Having accounted for MKDIR_PARENT and MKDIR_BODY above,
4639 * the only remaining dependency is that the updated inode
4640 * count must get pushed to disk. The inode has already
4641 * been pushed into its inode buffer (via VOP_UPDATE) at
4642 * the time of the reference count change. So we need only
4643 * locate that buffer, ensure that there will be no rollback
4644 * caused by a bitmap dependency, then write the inode buffer.
4645 */
4646 retry_lookup:
4649 }
4650 /*
4651 * If the inode still has bitmap dependencies,
4652 * push them to disk.
4653 */
4664 }
4665 /*
4666 * If the inode is still sitting in a buffer waiting
4667 * to be written, push it to disk.
4668 */
4677 /*
4678 * If we have failed to get rid of all the dependencies
4679 * then something is seriously wrong.
4680 */
4683 }
4684 }
4688 }
4690 /*
4691 * A large burst of file addition or deletion activity can drive the
4692 * memory load excessively high. First attempt to slow things down
4693 * using the techniques below. If that fails, this routine requests
4694 * the offending operations to fall back to running synchronously
4695 * until the memory load returns to a reasonable level.
4696 */
4697 int
4699 {
4708 }
4710 /*
4711 * If memory utilization has gotten too high, deliberately slow things
4712 * down and speed up the I/O processing.
4713 */
4714 static int
4716 {
4721 /*
4722 * We never hold up the filesystem syncer process.
4723 */
4726 /*
4727 * First check to see if the work list has gotten backlogged.
4728 * If it has, co-opt this process to help clean up two entries.
4729 * Because this process may hold inodes locked, we cannot
4730 * handle any remove requests that might block on a locked
4731 * inode as that could lead to deadlock.
4732 */
4738 }
4740 /*
4741 * If we are resource constrained on inode dependencies, try
4742 * flushing some dirty inodes. Otherwise, we are constrained
4743 * by file deletions, so try accelerating flushes of directories
4744 * with removal dependencies. We would like to do the cleanup
4745 * here, but we probably hold an inode locked at this point and
4746 * that might deadlock against one that we try to clean. So,
4747 * the best that we can do is request the syncer daemon to do
4748 * the cleanup for us.
4749 */
4766 }
4767 /*
4768 * Hopefully the syncer daemon will catch up and awaken us.
4769 * We wait at most tickdelay before proceeding in any case.
4770 */
4774 }
4776 /*
4777 * Flush out a directory with at least one removal dependency in an effort to
4778 * reduce the number of dirrem, freefile, and freeblks dependency structures.
4779 */
4780 static void
4782 {
4789 ino_t ino;
4805 }
4811 }
4812 }
4814 }
4816 /*
4817 * Clear out a block of dirty inodes in an effort to reduce
4818 * the number of inodedep dependency structures.
4819 */
4823 };
4825 static int
4827 {
4833 }
4835 }
4837 static void
4839 {
4850 /*
4851 * Pick a random inode dependency to be cleared.
4852 * We will then gather up all the inodes in its block
4853 * that have dependencies and flush them out.
4854 */
4862 }
4866 }
4867 /*
4868 * Ugly code to find mount point given pointer to superblock.
4869 */
4875 /*
4876 * Find the last inode in the block with dependencies.
4877 */
4882 /*
4883 * Asynchronously push all but the last inode with dependencies.
4884 * Synchronously push the last inode with dependencies to ensure
4885 * that the inode block gets written to free up the inodedeps.
4886 */
4894 }
4902 }
4905 }
4907 }
4909 /*
4910 * Function to determine if the buffer has outstanding dependencies
4911 * that will cause a roll-back if the buffer is written. If wantcount
4912 * is set, return number of dependencies, otherwise just yes or no.
4913 *
4914 * bioops callback - hold io_token
4915 */
4916 static int
4918 {
4936 /* bitmap allocation dependency */
4940 }
4942 /* direct block pointer dependency */
4946 }
4953 /* indirect block pointer dependency */
4957 }
4965 /* directory entry dependency */
4969 }
4970 }
4977 /* never a dependency on these blocks */
4983 /* NOTREACHED */
4984 }
4985 }
4986 out:
4990 }
4992 /*
4993 * Acquire exclusive access to a buffer. Requires softdep lock
4994 * to be held on entry. If waitfor is MNT_WAIT, may release/reacquire
4995 * softdep lock.
4996 *
4997 * Returns 1 if the buffer was locked, 0 if it was not locked or
4998 * if we had to block.
4999 *
5000 * NOTE! In order to return 1 we must acquire the buffer lock prior
5001 * to any release of &lk. Once we release &lk it's all over.
5002 * We may still have to block on the (type-stable) bp in that
5003 * case, but we must then unlock it and return 0.
5004 */
5005 static int
5007 {
5011 /*
5012 * If the contents of *bpp is NULL the caller presumably lost a race.
5013 */
5018 /*
5019 * Try to obtain the buffer lock without deadlocking on &lk.
5020 */
5024 /*
5025 * If the buffer is no longer dirty the OS already wrote it
5026 * out, return failure.
5027 */
5031 }
5033 /*
5034 * Finish nominal buffer locking sequence return success.
5035 *
5036 * Since we are not using a normal getblk(), and UFS
5037 * isn't KVABIO aware, we must make sure that the bp
5038 * is synchronized before returning it.
5039 */
5043 }
5045 /*
5046 * Failure case.
5047 *
5048 * If we are not being asked to wait, return 0 immediately.
5049 */
5053 /*
5054 * Once we release the softdep lock we can never return success,
5055 * but we still have to block on the type-stable buf for the caller
5056 * to be able to retry without livelocking the system.
5057 *
5058 * The caller will normally retry in this case.
5059 */
5066 }
5068 /*
5069 * Wait for pending output on a vnode to complete.
5070 * Must be called with vnode locked.
5071 */
5072 static void
5074 {
5082 }
5085 }
5087 /*
5088 * Called whenever a buffer that is being invalidated or reallocated
5089 * contains dependencies. This should only happen if an I/O error has
5090 * occurred. The routine is called with the buffer locked.
5091 *
5092 * bioops callback - hold io_token
5093 */
5094 static void
5096 {
5097 /* nothing to do, mp lock not needed */
5102 }
5104 /*
5105 * Function to handle asynchronous write errors in the filesystem.
5106 */
5107 void
5109 {
5110 /* XXX should do something better! */
5112 }