| blob | 2967b73934151f2291b2d0123dacb2f125239020 |
1 /*
2 * Copyright (C) International Business Machines Corp., 2000-2004
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See
12 * the GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
17 */
19 #include <linux/fs.h>
28 /*
29 * SERIALIZATION of the Block Allocation Map.
30 *
31 * the working state of the block allocation map is accessed in
32 * two directions:
33 *
34 * 1) allocation and free requests that start at the dmap
35 * level and move up through the dmap control pages (i.e.
36 * the vast majority of requests).
37 *
38 * 2) allocation requests that start at dmap control page
39 * level and work down towards the dmaps.
40 *
41 * the serialization scheme used here is as follows.
42 *
43 * requests which start at the bottom are serialized against each
44 * other through buffers and each requests holds onto its buffers
45 * as it works it way up from a single dmap to the required level
46 * of dmap control page.
47 * requests that start at the top are serialized against each other
48 * and request that start from the bottom by the multiple read/single
49 * write inode lock of the bmap inode. requests starting at the top
50 * take this lock in write mode while request starting at the bottom
51 * take the lock in read mode. a single top-down request may proceed
52 * exclusively while multiple bottoms-up requests may proceed
53 * simultaneously (under the protection of busy buffers).
54 *
55 * in addition to information found in dmaps and dmap control pages,
56 * the working state of the block allocation map also includes read/
57 * write information maintained in the bmap descriptor (i.e. total
58 * free block count, allocation group level free block counts).
59 * a single exclusive lock (BMAP_LOCK) is used to guard this information
60 * in the face of multiple-bottoms up requests.
61 * (lock ordering: IREAD_LOCK, BMAP_LOCK);
62 *
63 * accesses to the persistent state of the block allocation map (limited
64 * to the persistent bitmaps in dmaps) is guarded by (busy) buffers.
65 */
67 #define BMAP_LOCK_INIT(bmp) init_MUTEX(&bmp->db_bmaplock)
68 #define BMAP_LOCK(bmp) down(&bmp->db_bmaplock)
69 #define BMAP_UNLOCK(bmp) up(&bmp->db_bmaplock)
71 /*
72 * forward references
73 */
120 /*
121 * buddy table
122 *
123 * table used for determining buddy sizes within characters of
124 * dmap bitmap words. the characters themselves serve as indexes
125 * into the table, with the table elements yielding the maximum
126 * binary buddy of free bits within the character.
127 */
145 };
148 /*
149 * NAME: dbMount()
150 *
151 * FUNCTION: initializate the block allocation map.
152 *
153 * memory is allocated for the in-core bmap descriptor and
154 * the in-core descriptor is initialized from disk.
155 *
156 * PARAMETERS:
157 * ipbmap - pointer to in-core inode for the block map.
158 *
159 * RETURN VALUES:
160 * 0 - success
161 * -ENOMEM - insufficient memory
162 * -EIO - i/o error
163 */
165 {
171 /*
172 * allocate/initialize the in-memory bmap descriptor
173 */
174 /* allocate memory for the in-memory bmap descriptor */
179 /* read the on-disk bmap descriptor. */
186 }
188 /* copy the on-disk bmap descriptor to its in-memory version. */
207 /* release the buffer. */
210 /* bind the bmap inode and the bmap descriptor to each other. */
216 /*
217 * allocate/initialize the bmap lock
218 */
222 }
225 /*
226 * NAME: dbUnmount()
227 *
228 * FUNCTION: terminate the block allocation map in preparation for
229 * file system unmount.
230 *
231 * the in-core bmap descriptor is written to disk and
232 * the memory for this descriptor is freed.
233 *
234 * PARAMETERS:
235 * ipbmap - pointer to in-core inode for the block map.
236 *
237 * RETURN VALUES:
238 * 0 - success
239 * -EIO - i/o error
240 */
242 {
248 /*
249 * Invalidate the page cache buffers
250 */
253 /* free the memory for the in-memory bmap. */
257 }
259 /*
260 * dbSync()
261 */
263 {
269 /*
270 * write bmap global control page
271 */
272 /* get the buffer for the on-disk bmap descriptor. */
279 }
280 /* copy the in-memory version of the bmap to the on-disk version */
299 /* write the buffer */
302 /*
303 * write out dirty pages of bmap
304 */
310 }
313 /*
314 * NAME: dbFree()
315 *
316 * FUNCTION: free the specified block range from the working block
317 * allocation map.
318 *
319 * the blocks will be free from the working map one dmap
320 * at a time.
321 *
322 * PARAMETERS:
323 * ip - pointer to in-core inode;
324 * blkno - starting block number to be freed.
325 * nblocks - number of blocks to be freed.
326 *
327 * RETURN VALUES:
328 * 0 - success
329 * -EIO - i/o error
330 */
332 {
342 /* block to be freed better be within the mapsize. */
351 }
353 /*
354 * free the blocks a dmap at a time.
355 */
358 /* release previous dmap if any */
361 }
363 /* get the buffer for the current dmap. */
369 }
372 /* determine the number of blocks to be freed from
373 * this dmap.
374 */
377 /* free the blocks. */
383 }
384 }
386 /* write the last buffer. */
392 }
395 /*
396 * NAME: dbUpdatePMap()
397 *
398 * FUNCTION: update the allocation state (free or allocate) of the
399 * specified block range in the persistent block allocation map.
400 *
401 * the blocks will be updated in the persistent map one
402 * dmap at a time.
403 *
404 * PARAMETERS:
405 * ipbmap - pointer to in-core inode for the block map.
406 * free - TRUE if block range is to be freed from the persistent
407 * map; FALSE if it is to be allocated.
408 * blkno - starting block number of the range.
409 * nblocks - number of contiguous blocks in the range.
410 * tblk - transaction block;
411 *
412 * RETURN VALUES:
413 * 0 - success
414 * -EIO - i/o error
415 */
416 int
419 {
424 u32 mask;
431 /* the blocks better be within the mapsize. */
439 }
441 /* compute delta of transaction lsn from log syncpt */
446 /*
447 * update the block state a dmap at a time.
448 */
452 /* get the buffer for the current dmap. */
457 }
464 }
467 /* determine the bit number and word within the dmap of
468 * the starting block. also determine how many blocks
469 * are to be updated within this dmap.
470 */
475 /* update the bits of the dmap words. the first and last
476 * words may only have a subset of their bits updated. if
477 * this is the case, we'll work against that word (i.e.
478 * partial first and/or last) only in a single pass. a
479 * single pass will also be used to update all words that
480 * are to have all their bits updated.
481 */
484 /* determine the bit number within the word and
485 * the number of bits within the word.
486 */
490 /* check if only part of the word is to be updated. */
492 /* update (free or allocate) the bits
493 * in this word.
494 */
495 mask =
500 else
506 /* one or more words are to have all
507 * their bits updated. determine how
508 * many words and how many bits.
509 */
513 /* update (free or allocate) the bits
514 * in these words.
515 */
519 else
524 }
525 }
527 /*
528 * update dmap lsn
529 */
536 /* inherit older/smaller lsn */
542 /* move bp after tblock in logsync list */
544 }
546 /* inherit younger/larger clsn */
556 /* insert bp after tblock in logsync list */
564 }
565 }
567 /* write the last buffer. */
570 }
573 }
576 /*
577 * NAME: dbNextAG()
578 *
579 * FUNCTION: find the preferred allocation group for new allocations.
580 *
581 * Within the allocation groups, we maintain a preferred
582 * allocation group which consists of a group with at least
583 * average free space. It is the preferred group that we target
584 * new inode allocation towards. The tie-in between inode
585 * allocation and block allocation occurs as we allocate the
586 * first (data) block of an inode and specify the inode (block)
587 * as the allocation hint for this block.
588 *
589 * We try to avoid having more than one open file growing in
590 * an allocation group, as this will lead to fragmentation.
591 * This differs from the old OS/2 method of trying to keep
592 * empty ags around for large allocations.
593 *
594 * PARAMETERS:
595 * ipbmap - pointer to in-core inode for the block map.
596 *
597 * RETURN VALUES:
598 * the preferred allocation group number.
599 */
601 {
602 s64 avgfree;
611 /* determine the average number of free blocks within the ags. */
614 /*
615 * if the current preferred ag does not have an active allocator
616 * and has at least average freespace, return it
617 */
623 /* From the last preferred ag, find the next one with at least
624 * average free space.
625 */
631 /* open file is currently growing in this ag */
634 /* Return this one */
638 /* Less than avg. freespace, but best so far */
641 }
642 }
644 /*
645 * If no inactive ag was found with average freespace, use the
646 * next best
647 */
650 /* else leave db_agpref unchanged */
651 unlock:
654 /* return the preferred group.
655 */
657 }
659 /*
660 * NAME: dbAlloc()
661 *
662 * FUNCTION: attempt to allocate a specified number of contiguous free
663 * blocks from the working allocation block map.
664 *
665 * the block allocation policy uses hints and a multi-step
666 * approach.
667 *
668 * for allocation requests smaller than the number of blocks
669 * per dmap, we first try to allocate the new blocks
670 * immediately following the hint. if these blocks are not
671 * available, we try to allocate blocks near the hint. if
672 * no blocks near the hint are available, we next try to
673 * allocate within the same dmap as contains the hint.
674 *
675 * if no blocks are available in the dmap or the allocation
676 * request is larger than the dmap size, we try to allocate
677 * within the same allocation group as contains the hint. if
678 * this does not succeed, we finally try to allocate anywhere
679 * within the aggregate.
680 *
681 * we also try to allocate anywhere within the aggregate for
682 * for allocation requests larger than the allocation group
683 * size or requests that specify no hint value.
684 *
685 * PARAMETERS:
686 * ip - pointer to in-core inode;
687 * hint - allocation hint.
688 * nblocks - number of contiguous blocks in the range.
689 * results - on successful return, set to the starting block number
690 * of the newly allocated contiguous range.
691 *
692 * RETURN VALUES:
693 * 0 - success
694 * -ENOSPC - insufficient disk resources
695 * -EIO - i/o error
696 */
698 {
706 s64 mapSize;
709 /* assert that nblocks is valid */
712 #ifdef _STILL_TO_PORT
713 /* DASD limit check F226941 */
718 /* get the log2 number of blocks to be allocated.
719 * if the number of blocks is not a log2 multiple,
720 * it will be rounded up to the next log2 multiple.
721 */
726 //retry: /* serialize w.r.t.extendfs() */
729 /* the hint should be within the map */
733 }
735 /* if the number of blocks to be allocated is greater than the
736 * allocation group size, try to allocate anywhere.
737 */
744 }
746 /*
747 * If no hint, let dbNextAG recommend an allocation group
748 */
752 /* we would like to allocate close to the hint. adjust the
753 * hint to the block following the hint since the allocators
754 * will start looking for free space starting at this point.
755 */
763 /* check if blkno crosses over into a new allocation group.
764 * if so, check if we should allow allocations within this
765 * allocation group.
766 */
768 /* check if the AG is currenly being written to.
769 * if so, call dbNextAG() to find a non-busy
770 * AG with sufficient free space.
771 */
775 /* check if the allocation request size can be satisfied from a
776 * single dmap. if so, try to allocate from the dmap containing
777 * the hint using a tiered strategy.
778 */
782 /* get the buffer for the dmap containing the hint.
783 */
792 /* first, try to satisfy the allocation request with the
793 * blocks beginning at the hint.
794 */
800 }
804 }
809 /*
810 * Someone else is writing in this allocation
811 * group. To avoid fragmenting, try another ag
812 */
816 }
818 /* next, try to satisfy the allocation request with blocks
819 * near the hint.
820 */
829 }
831 /* try to satisfy the allocation request with blocks within
832 * the same dmap as the hint.
833 */
841 }
845 }
847 /* try to satisfy the allocation request with blocks within
848 * the same allocation group as the hint.
849 */
857 pref_ag:
858 /*
859 * Let dbNextAG recommend a preferred allocation group
860 */
864 /* Try to allocate within this allocation group. if that fails, try to
865 * allocate anywhere in the map.
866 */
870 write_unlock:
875 read_unlock:
879 }
881 #ifdef _NOTYET
882 /*
883 * NAME: dbAllocExact()
884 *
885 * FUNCTION: try to allocate the requested extent;
886 *
887 * PARAMETERS:
888 * ip - pointer to in-core inode;
889 * blkno - extent address;
890 * nblocks - extent length;
891 *
892 * RETURN VALUES:
893 * 0 - success
894 * -ENOSPC - insufficient disk resources
895 * -EIO - i/o error
896 */
898 {
903 s64 lblkno;
908 /*
909 * validate extent request:
910 *
911 * note: defragfs policy:
912 * max 64 blocks will be moved.
913 * allocation request size must be satisfied from a single dmap.
914 */
918 }
921 /* the free space is no longer available */
924 }
926 /* read in the dmap covering the extent */
932 }
935 /* try to allocate the requested extent */
946 }
949 /*
950 * NAME: dbReAlloc()
951 *
952 * FUNCTION: attempt to extend a current allocation by a specified
953 * number of blocks.
954 *
955 * this routine attempts to satisfy the allocation request
956 * by first trying to extend the existing allocation in
957 * place by allocating the additional blocks as the blocks
958 * immediately following the current allocation. if these
959 * blocks are not available, this routine will attempt to
960 * allocate a new set of contiguous blocks large enough
961 * to cover the existing allocation plus the additional
962 * number of blocks required.
963 *
964 * PARAMETERS:
965 * ip - pointer to in-core inode requiring allocation.
966 * blkno - starting block of the current allocation.
967 * nblocks - number of contiguous blocks within the current
968 * allocation.
969 * addnblocks - number of blocks to add to the allocation.
970 * results - on successful return, set to the starting block number
971 * of the existing allocation if the existing allocation
972 * was extended in place or to a newly allocated contiguous
973 * range if the existing allocation could not be extended
974 * in place.
975 *
976 * RETURN VALUES:
977 * 0 - success
978 * -ENOSPC - insufficient disk resources
979 * -EIO - i/o error
980 */
981 int
984 {
987 /* try to extend the allocation in place.
988 */
995 }
997 /* could not extend the allocation in place, so allocate a
998 * new set of blocks for the entire request (i.e. try to get
999 * a range of contiguous blocks large enough to cover the
1000 * existing allocation plus the additional blocks.)
1001 */
1004 }
1007 /*
1008 * NAME: dbExtend()
1009 *
1010 * FUNCTION: attempt to extend a current allocation by a specified
1011 * number of blocks.
1012 *
1013 * this routine attempts to satisfy the allocation request
1014 * by first trying to extend the existing allocation in
1015 * place by allocating the additional blocks as the blocks
1016 * immediately following the current allocation.
1017 *
1018 * PARAMETERS:
1019 * ip - pointer to in-core inode requiring allocation.
1020 * blkno - starting block of the current allocation.
1021 * nblocks - number of contiguous blocks within the current
1022 * allocation.
1023 * addnblocks - number of blocks to add to the allocation.
1024 *
1025 * RETURN VALUES:
1026 * 0 - success
1027 * -ENOSPC - insufficient disk resources
1028 * -EIO - i/o error
1029 */
1031 {
1034 uint rel_block;
1041 /*
1042 * We don't want a non-aligned extent to cross a page boundary
1043 */
1048 /* get the last block of the current allocation */
1051 /* determine the block number of the block following
1052 * the existing allocation.
1053 */
1058 /* better be within the file system */
1065 }
1067 /* we'll attempt to extend the current allocation in place by
1068 * allocating the additional blocks as the blocks immediately
1069 * following the current allocation. we only try to extend the
1070 * current allocation in place if the number of additional blocks
1071 * can fit into a dmap, the last block of the current allocation
1072 * is not the last block of the file system, and the start of the
1073 * inplace extension is not on an allocation group boundary.
1074 */
1079 }
1081 /* get the buffer for the dmap containing the first block
1082 * of the extension.
1083 */
1089 }
1093 /* try to allocate the blocks immediately following the
1094 * current allocation.
1095 */
1100 /* were we successful ? */
1103 else
1104 /* we were not successful */
1109 }
1112 /*
1113 * NAME: dbAllocNext()
1114 *
1115 * FUNCTION: attempt to allocate the blocks of the specified block
1116 * range within a dmap.
1117 *
1118 * PARAMETERS:
1119 * bmp - pointer to bmap descriptor
1120 * dp - pointer to dmap.
1121 * blkno - starting block number of the range.
1122 * nblocks - number of contiguous free blocks of the range.
1123 *
1124 * RETURN VALUES:
1125 * 0 - success
1126 * -ENOSPC - insufficient disk resources
1127 * -EIO - i/o error
1128 *
1129 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1130 */
1133 {
1137 u32 mask;
1143 }
1145 /* pick up a pointer to the leaves of the dmap tree.
1146 */
1149 /* determine the bit number and word within the dmap of the
1150 * starting block.
1151 */
1155 /* check if the specified block range is contained within
1156 * this dmap.
1157 */
1161 /* check if the starting leaf indicates that anything
1162 * is free.
1163 */
1167 /* check the dmaps words corresponding to block range to see
1168 * if the block range is free. not all bits of the first and
1169 * last words may be contained within the block range. if this
1170 * is the case, we'll work against those words (i.e. partial first
1171 * and/or last) on an individual basis (a single pass) and examine
1172 * the actual bits to determine if they are free. a single pass
1173 * will be used for all dmap words fully contained within the
1174 * specified range. within this pass, the leaves of the dmap
1175 * tree will be examined to determine if the blocks are free. a
1176 * single leaf may describe the free space of multiple dmap
1177 * words, so we may visit only a subset of the actual leaves
1178 * corresponding to the dmap words of the block range.
1179 */
1181 /* determine the bit number within the word and
1182 * the number of bits within the word.
1183 */
1187 /* check if only part of the word is to be examined.
1188 */
1190 /* check if the bits are free.
1191 */
1198 /* one or more dmap words are fully contained
1199 * within the block range. determine how many
1200 * words and how many bits.
1201 */
1205 /* now examine the appropriate leaves to determine
1206 * if the blocks are free.
1207 */
1209 /* does the leaf describe any free space ?
1210 */
1214 /* determine the l2 number of bits provided
1215 * by this leaf.
1216 */
1217 l2size =
1220 /* determine how many words were handled.
1221 */
1226 }
1227 }
1228 }
1230 /* allocate the blocks.
1231 */
1233 }
1236 /*
1237 * NAME: dbAllocNear()
1238 *
1239 * FUNCTION: attempt to allocate a number of contiguous free blocks near
1240 * a specified block (hint) within a dmap.
1241 *
1242 * starting with the dmap leaf that covers the hint, we'll
1243 * check the next four contiguous leaves for sufficient free
1244 * space. if sufficient free space is found, we'll allocate
1245 * the desired free space.
1246 *
1247 * PARAMETERS:
1248 * bmp - pointer to bmap descriptor
1249 * dp - pointer to dmap.
1250 * blkno - block number to allocate near.
1251 * nblocks - actual number of contiguous free blocks desired.
1252 * l2nb - log2 number of contiguous free blocks desired.
1253 * results - on successful return, set to the starting block number
1254 * of the newly allocated range.
1255 *
1256 * RETURN VALUES:
1257 * 0 - success
1258 * -ENOSPC - insufficient disk resources
1259 * -EIO - i/o error
1260 *
1261 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1262 */
1263 static int
1266 {
1274 }
1278 /* determine the word within the dmap that holds the hint
1279 * (i.e. blkno). also, determine the last word in the dmap
1280 * that we'll include in our examination.
1281 */
1285 /* examine the leaves for sufficient free space.
1286 */
1288 /* does the leaf describe sufficient free space ?
1289 */
1293 /* determine the block number within the file system
1294 * of the first block described by this dmap word.
1295 */
1298 /* if not all bits of the dmap word are free, get the
1299 * starting bit number within the dmap word of the required
1300 * string of free bits and adjust the block number with the
1301 * value.
1302 */
1304 blkno +=
1307 /* allocate the blocks.
1308 */
1313 }
1316 }
1319 /*
1320 * NAME: dbAllocAG()
1321 *
1322 * FUNCTION: attempt to allocate the specified number of contiguous
1323 * free blocks within the specified allocation group.
1324 *
1325 * unless the allocation group size is equal to the number
1326 * of blocks per dmap, the dmap control pages will be used to
1327 * find the required free space, if available. we start the
1328 * search at the highest dmap control page level which
1329 * distinctly describes the allocation group's free space
1330 * (i.e. the highest level at which the allocation group's
1331 * free space is not mixed in with that of any other group).
1332 * in addition, we start the search within this level at a
1333 * height of the dmapctl dmtree at which the nodes distinctly
1334 * describe the allocation group's free space. at this height,
1335 * the allocation group's free space may be represented by 1
1336 * or two sub-trees, depending on the allocation group size.
1337 * we search the top nodes of these subtrees left to right for
1338 * sufficient free space. if sufficient free space is found,
1339 * the subtree is searched to find the leftmost leaf that
1340 * has free space. once we have made it to the leaf, we
1341 * move the search to the next lower level dmap control page
1342 * corresponding to this leaf. we continue down the dmap control
1343 * pages until we find the dmap that contains or starts the
1344 * sufficient free space and we allocate at this dmap.
1345 *
1346 * if the allocation group size is equal to the dmap size,
1347 * we'll start at the dmap corresponding to the allocation
1348 * group and attempt the allocation at this level.
1349 *
1350 * the dmap control page search is also not performed if the
1351 * allocation group is completely free and we go to the first
1352 * dmap of the allocation group to do the allocation. this is
1353 * done because the allocation group may be part (not the first
1354 * part) of a larger binary buddy system, causing the dmap
1355 * control pages to indicate no free space (NOFREE) within
1356 * the allocation group.
1357 *
1358 * PARAMETERS:
1359 * bmp - pointer to bmap descriptor
1360 * agno - allocation group number.
1361 * nblocks - actual number of contiguous free blocks desired.
1362 * l2nb - log2 number of contiguous free blocks desired.
1363 * results - on successful return, set to the starting block number
1364 * of the newly allocated range.
1365 *
1366 * RETURN VALUES:
1367 * 0 - success
1368 * -ENOSPC - insufficient disk resources
1369 * -EIO - i/o error
1370 *
1371 * note: IWRITE_LOCK(ipmap) held on entry/exit;
1372 */
1373 static int
1375 {
1382 /* allocation request should not be for more than the
1383 * allocation group size.
1384 */
1387 "dbAllocAG: allocation request is larger than the "
1390 }
1392 /* determine the starting block number of the allocation
1393 * group.
1394 */
1397 /* check if the allocation group size is the minimum allocation
1398 * group size or if the allocation group is completely free. if
1399 * the allocation group size is the minimum size of BPERDMAP (i.e.
1400 * 1 dmap), there is no need to search the dmap control page (below)
1401 * that fully describes the allocation group since the allocation
1402 * group is already fully described by a dmap. in this case, we
1403 * just call dbAllocCtl() to search the dmap tree and allocate the
1404 * required space if available.
1405 *
1406 * if the allocation group is completely free, dbAllocCtl() is
1407 * also called to allocate the required space. this is done for
1408 * two reasons. first, it makes no sense searching the dmap control
1409 * pages for free space when we know that free space exists. second,
1410 * the dmap control pages may indicate that the allocation group
1411 * has no free space if the allocation group is part (not the first
1412 * part) of a larger binary buddy system.
1413 */
1424 }
1426 }
1428 /* the buffer for the dmap control page that fully describes the
1429 * allocation group.
1430 */
1443 }
1445 /* search the subtree(s) of the dmap control page that describes
1446 * the allocation group, looking for sufficient free space. to begin,
1447 * determine how many allocation groups are represented in a dmap
1448 * control page at the control page level (i.e. L0, L1, L2) that
1449 * fully describes an allocation group. next, determine the starting
1450 * tree index of this allocation group within the control page.
1451 */
1452 agperlev =
1456 /* dmap control page trees fan-out by 4 and a single allocation
1457 * group may be described by 1 or 2 subtrees within the ag level
1458 * dmap control page, depending upon the ag size. examine the ag's
1459 * subtrees for sufficient free space, starting with the leftmost
1460 * subtree.
1461 */
1463 /* is there sufficient free space ?
1464 */
1468 /* sufficient free space found in a subtree. now search down
1469 * the subtree to find the leftmost leaf that describes this
1470 * free space.
1471 */
1477 }
1478 }
1484 }
1485 }
1487 /* determine the block number within the file system
1488 * that corresponds to this leaf.
1489 */
1497 blkno +=
1500 /* release the buffer in preparation for going down
1501 * the next level of dmap control pages.
1502 */
1505 /* check if we need to continue to search down the lower
1506 * level dmap control pages. we need to if the number of
1507 * blocks required is less than maximum number of blocks
1508 * described at the next lower level.
1509 */
1512 /* search the lower level dmap control pages to get
1513 * the starting block number of the the dmap that
1514 * contains or starts off the free space.
1515 */
1521 "dbAllocAG: control page "
1524 }
1526 }
1527 }
1529 /* allocate the blocks.
1530 */
1536 }
1538 }
1540 /* no space in the allocation group. release the buffer and
1541 * return -ENOSPC.
1542 */
1546 }
1549 /*
1550 * NAME: dbAllocAny()
1551 *
1552 * FUNCTION: attempt to allocate the specified number of contiguous
1553 * free blocks anywhere in the file system.
1554 *
1555 * dbAllocAny() attempts to find the sufficient free space by
1556 * searching down the dmap control pages, starting with the
1557 * highest level (i.e. L0, L1, L2) control page. if free space
1558 * large enough to satisfy the desired free space is found, the
1559 * desired free space is allocated.
1560 *
1561 * PARAMETERS:
1562 * bmp - pointer to bmap descriptor
1563 * nblocks - actual number of contiguous free blocks desired.
1564 * l2nb - log2 number of contiguous free blocks desired.
1565 * results - on successful return, set to the starting block number
1566 * of the newly allocated range.
1567 *
1568 * RETURN VALUES:
1569 * 0 - success
1570 * -ENOSPC - insufficient disk resources
1571 * -EIO - i/o error
1572 *
1573 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1574 */
1576 {
1580 /* starting with the top level dmap control page, search
1581 * down the dmap control levels for sufficient free space.
1582 * if free space is found, dbFindCtl() returns the starting
1583 * block number of the dmap that contains or starts off the
1584 * range of free space.
1585 */
1589 /* allocate the blocks.
1590 */
1596 }
1598 }
1601 /*
1602 * NAME: dbFindCtl()
1603 *
1604 * FUNCTION: starting at a specified dmap control page level and block
1605 * number, search down the dmap control levels for a range of
1606 * contiguous free blocks large enough to satisfy an allocation
1607 * request for the specified number of free blocks.
1608 *
1609 * if sufficient contiguous free blocks are found, this routine
1610 * returns the starting block number within a dmap page that
1611 * contains or starts a range of contiqious free blocks that
1612 * is sufficient in size.
1613 *
1614 * PARAMETERS:
1615 * bmp - pointer to bmap descriptor
1616 * level - starting dmap control page level.
1617 * l2nb - log2 number of contiguous free blocks desired.
1618 * *blkno - on entry, starting block number for conducting the search.
1619 * on successful return, the first block within a dmap page
1620 * that contains or starts a range of contiguous free blocks.
1621 *
1622 * RETURN VALUES:
1623 * 0 - success
1624 * -ENOSPC - insufficient disk resources
1625 * -EIO - i/o error
1626 *
1627 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1628 */
1630 {
1637 /* starting at the specified dmap control page level and block
1638 * number, search down the dmap control levels for the starting
1639 * block number of a dmap page that contains or starts off
1640 * sufficient free blocks.
1641 */
1643 /* get the buffer of the dmap control page for the block
1644 * number and level (i.e. L0, L1, L2).
1645 */
1658 }
1660 /* search the tree within the dmap control page for
1661 * sufficent free space. if sufficient free space is found,
1662 * dbFindLeaf() returns the index of the leaf at which
1663 * free space was found.
1664 */
1667 /* release the buffer.
1668 */
1671 /* space found ?
1672 */
1678 }
1680 }
1682 /* adjust the block number to reflect the location within
1683 * the dmap control page (i.e. the leaf) at which free
1684 * space was found.
1685 */
1688 /* we stop the search at this dmap control page level if
1689 * the number of blocks required is greater than or equal
1690 * to the maximum number of blocks described at the next
1691 * (lower) level.
1692 */
1695 }
1699 }
1702 /*
1703 * NAME: dbAllocCtl()
1704 *
1705 * FUNCTION: attempt to allocate a specified number of contiguous
1706 * blocks starting within a specific dmap.
1707 *
1708 * this routine is called by higher level routines that search
1709 * the dmap control pages above the actual dmaps for contiguous
1710 * free space. the result of successful searches by these
1711 * routines are the starting block numbers within dmaps, with
1712 * the dmaps themselves containing the desired contiguous free
1713 * space or starting a contiguous free space of desired size
1714 * that is made up of the blocks of one or more dmaps. these
1715 * calls should not fail due to insufficent resources.
1716 *
1717 * this routine is called in some cases where it is not known
1718 * whether it will fail due to insufficient resources. more
1719 * specifically, this occurs when allocating from an allocation
1720 * group whose size is equal to the number of blocks per dmap.
1721 * in this case, the dmap control pages are not examined prior
1722 * to calling this routine (to save pathlength) and the call
1723 * might fail.
1724 *
1725 * for a request size that fits within a dmap, this routine relies
1726 * upon the dmap's dmtree to find the requested contiguous free
1727 * space. for request sizes that are larger than a dmap, the
1728 * requested free space will start at the first block of the
1729 * first dmap (i.e. blkno).
1730 *
1731 * PARAMETERS:
1732 * bmp - pointer to bmap descriptor
1733 * nblocks - actual number of contiguous free blocks to allocate.
1734 * l2nb - log2 number of contiguous free blocks to allocate.
1735 * blkno - starting block number of the dmap to start the allocation
1736 * from.
1737 * results - on successful return, set to the starting block number
1738 * of the newly allocated range.
1739 *
1740 * RETURN VALUES:
1741 * 0 - success
1742 * -ENOSPC - insufficient disk resources
1743 * -EIO - i/o error
1744 *
1745 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1746 */
1747 static int
1749 {
1755 /* check if the allocation request is confined to a single dmap.
1756 */
1758 /* get the buffer for the dmap.
1759 */
1766 /* try to allocate the blocks.
1767 */
1775 }
1777 /* allocation request involving multiple dmaps. it must start on
1778 * a dmap boundary.
1779 */
1782 /* allocate the blocks dmap by dmap.
1783 */
1785 /* get the buffer for the dmap.
1786 */
1792 }
1795 /* the dmap better be all free.
1796 */
1803 }
1805 /* determine how many blocks to allocate from this dmap.
1806 */
1809 /* allocate the blocks from the dmap.
1810 */
1814 }
1816 /* write the buffer.
1817 */
1819 }
1821 /* set the results (starting block number) and return.
1822 */
1826 /* something failed in handling an allocation request involving
1827 * multiple dmaps. we'll try to clean up by backing out any
1828 * allocation that has already happened for this request. if
1829 * we fail in backing out the allocation, we'll mark the file
1830 * system to indicate that blocks have been leaked.
1831 */
1832 backout:
1834 /* try to backout the allocations dmap by dmap.
1835 */
1838 /* get the buffer for this dmap.
1839 */
1843 /* could not back out. mark the file system
1844 * to indicate that we have leaked blocks.
1845 */
1849 }
1852 /* free the blocks is this dmap.
1853 */
1855 /* could not back out. mark the file system
1856 * to indicate that we have leaked blocks.
1857 */
1862 }
1864 /* write the buffer.
1865 */
1867 }
1870 }
1873 /*
1874 * NAME: dbAllocDmapLev()
1875 *
1876 * FUNCTION: attempt to allocate a specified number of contiguous blocks
1877 * from a specified dmap.
1878 *
1879 * this routine checks if the contiguous blocks are available.
1880 * if so, nblocks of blocks are allocated; otherwise, ENOSPC is
1881 * returned.
1882 *
1883 * PARAMETERS:
1884 * mp - pointer to bmap descriptor
1885 * dp - pointer to dmap to attempt to allocate blocks from.
1886 * l2nb - log2 number of contiguous block desired.
1887 * nblocks - actual number of contiguous block desired.
1888 * results - on successful return, set to the starting block number
1889 * of the newly allocated range.
1890 *
1891 * RETURN VALUES:
1892 * 0 - success
1893 * -ENOSPC - insufficient disk resources
1894 * -EIO - i/o error
1895 *
1896 * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
1897 * IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
1898 */
1899 static int
1902 {
1903 s64 blkno;
1906 /* can't be more than a dmaps worth of blocks */
1909 /* search the tree within the dmap page for sufficient
1910 * free space. if sufficient free space is found, dbFindLeaf()
1911 * returns the index of the leaf at which free space was found.
1912 */
1916 /* determine the block number within the file system corresponding
1917 * to the leaf at which free space was found.
1918 */
1921 /* if not all bits of the dmap word are free, get the starting
1922 * bit number within the dmap word of the required string of free
1923 * bits and adjust the block number with this value.
1924 */
1928 /* allocate the blocks */
1933 }
1936 /*
1937 * NAME: dbAllocDmap()
1938 *
1939 * FUNCTION: adjust the disk allocation map to reflect the allocation
1940 * of a specified block range within a dmap.
1941 *
1942 * this routine allocates the specified blocks from the dmap
1943 * through a call to dbAllocBits(). if the allocation of the
1944 * block range causes the maximum string of free blocks within
1945 * the dmap to change (i.e. the value of the root of the dmap's
1946 * dmtree), this routine will cause this change to be reflected
1947 * up through the appropriate levels of the dmap control pages
1948 * by a call to dbAdjCtl() for the L0 dmap control page that
1949 * covers this dmap.
1950 *
1951 * PARAMETERS:
1952 * bmp - pointer to bmap descriptor
1953 * dp - pointer to dmap to allocate the block range from.
1954 * blkno - starting block number of the block to be allocated.
1955 * nblocks - number of blocks to be allocated.
1956 *
1957 * RETURN VALUES:
1958 * 0 - success
1959 * -EIO - i/o error
1960 *
1961 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
1962 */
1965 {
1966 s8 oldroot;
1969 /* save the current value of the root (i.e. maximum free string)
1970 * of the dmap tree.
1971 */
1974 /* allocate the specified (blocks) bits */
1977 /* if the root has not changed, done. */
1981 /* root changed. bubble the change up to the dmap control pages.
1982 * if the adjustment of the upper level control pages fails,
1983 * backout the bit allocation (thus making everything consistent).
1984 */
1989 }
1992 /*
1993 * NAME: dbFreeDmap()
1994 *
1995 * FUNCTION: adjust the disk allocation map to reflect the allocation
1996 * of a specified block range within a dmap.
1997 *
1998 * this routine frees the specified blocks from the dmap through
1999 * a call to dbFreeBits(). if the deallocation of the block range
2000 * causes the maximum string of free blocks within the dmap to
2001 * change (i.e. the value of the root of the dmap's dmtree), this
2002 * routine will cause this change to be reflected up through the
2003 * appropriate levels of the dmap control pages by a call to
2004 * dbAdjCtl() for the L0 dmap control page that covers this dmap.
2005 *
2006 * PARAMETERS:
2007 * bmp - pointer to bmap descriptor
2008 * dp - pointer to dmap to free the block range from.
2009 * blkno - starting block number of the block to be freed.
2010 * nblocks - number of blocks to be freed.
2011 *
2012 * RETURN VALUES:
2013 * 0 - success
2014 * -EIO - i/o error
2015 *
2016 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2017 */
2020 {
2021 s8 oldroot;
2024 /* save the current value of the root (i.e. maximum free string)
2025 * of the dmap tree.
2026 */
2029 /* free the specified (blocks) bits */
2032 /* if error or the root has not changed, done. */
2036 /* root changed. bubble the change up to the dmap control pages.
2037 * if the adjustment of the upper level control pages fails,
2038 * backout the deallocation.
2039 */
2043 /* as part of backing out the deallocation, we will have
2044 * to back split the dmap tree if the deallocation caused
2045 * the freed blocks to become part of a larger binary buddy
2046 * system.
2047 */
2052 }
2055 }
2058 /*
2059 * NAME: dbAllocBits()
2060 *
2061 * FUNCTION: allocate a specified block range from a dmap.
2062 *
2063 * this routine updates the dmap to reflect the working
2064 * state allocation of the specified block range. it directly
2065 * updates the bits of the working map and causes the adjustment
2066 * of the binary buddy system described by the dmap's dmtree
2067 * leaves to reflect the bits allocated. it also causes the
2068 * dmap's dmtree, as a whole, to reflect the allocated range.
2069 *
2070 * PARAMETERS:
2071 * bmp - pointer to bmap descriptor
2072 * dp - pointer to dmap to allocate bits from.
2073 * blkno - starting block number of the bits to be allocated.
2074 * nblocks - number of bits to be allocated.
2075 *
2076 * RETURN VALUES: none
2077 *
2078 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2079 */
2082 {
2088 /* pick up a pointer to the leaves of the dmap tree */
2091 /* determine the bit number and word within the dmap of the
2092 * starting block.
2093 */
2097 /* block range better be within the dmap */
2100 /* allocate the bits of the dmap's words corresponding to the block
2101 * range. not all bits of the first and last words may be contained
2102 * within the block range. if this is the case, we'll work against
2103 * those words (i.e. partial first and/or last) on an individual basis
2104 * (a single pass), allocating the bits of interest by hand and
2105 * updating the leaf corresponding to the dmap word. a single pass
2106 * will be used for all dmap words fully contained within the
2107 * specified range. within this pass, the bits of all fully contained
2108 * dmap words will be marked as free in a single shot and the leaves
2109 * will be updated. a single leaf may describe the free space of
2110 * multiple dmap words, so we may update only a subset of the actual
2111 * leaves corresponding to the dmap words of the block range.
2112 */
2114 /* determine the bit number within the word and
2115 * the number of bits within the word.
2116 */
2120 /* check if only part of a word is to be allocated.
2121 */
2123 /* allocate (set to 1) the appropriate bits within
2124 * this dmap word.
2125 */
2129 /* update the leaf for this dmap word. in addition
2130 * to setting the leaf value to the binary buddy max
2131 * of the updated dmap word, dbSplit() will split
2132 * the binary system of the leaves if need be.
2133 */
2139 /* one or more dmap words are fully contained
2140 * within the block range. determine how many
2141 * words and allocate (set to 1) the bits of these
2142 * words.
2143 */
2147 /* determine how many bits.
2148 */
2151 /* now update the appropriate leaves to reflect
2152 * the allocated words.
2153 */
2157 "dbAllocBits: leaf page "
2160 }
2162 /* determine what the leaf value should be
2163 * updated to as the minimum of the l2 number
2164 * of bits being allocated and the l2 number
2165 * of bits currently described by this leaf.
2166 */
2169 /* update the leaf to reflect the allocation.
2170 * in addition to setting the leaf value to
2171 * NOFREE, dbSplit() will split the binary
2172 * system of the leaves to reflect the current
2173 * allocation (size).
2174 */
2177 /* get the number of dmap words handled */
2180 }
2181 }
2182 }
2184 /* update the free count for this dmap */
2189 /* if this allocation group is completely free,
2190 * update the maximum allocation group number if this allocation
2191 * group is the new max.
2192 */
2197 /* update the free count for the allocation group and map */
2202 }
2205 /*
2206 * NAME: dbFreeBits()
2207 *
2208 * FUNCTION: free a specified block range from a dmap.
2209 *
2210 * this routine updates the dmap to reflect the working
2211 * state allocation of the specified block range. it directly
2212 * updates the bits of the working map and causes the adjustment
2213 * of the binary buddy system described by the dmap's dmtree
2214 * leaves to reflect the bits freed. it also causes the dmap's
2215 * dmtree, as a whole, to reflect the deallocated range.
2216 *
2217 * PARAMETERS:
2218 * bmp - pointer to bmap descriptor
2219 * dp - pointer to dmap to free bits from.
2220 * blkno - starting block number of the bits to be freed.
2221 * nblocks - number of bits to be freed.
2222 *
2223 * RETURN VALUES: 0 for success
2224 *
2225 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2226 */
2229 {
2235 /* determine the bit number and word within the dmap of the
2236 * starting block.
2237 */
2241 /* block range better be within the dmap.
2242 */
2245 /* free the bits of the dmaps words corresponding to the block range.
2246 * not all bits of the first and last words may be contained within
2247 * the block range. if this is the case, we'll work against those
2248 * words (i.e. partial first and/or last) on an individual basis
2249 * (a single pass), freeing the bits of interest by hand and updating
2250 * the leaf corresponding to the dmap word. a single pass will be used
2251 * for all dmap words fully contained within the specified range.
2252 * within this pass, the bits of all fully contained dmap words will
2253 * be marked as free in a single shot and the leaves will be updated. a
2254 * single leaf may describe the free space of multiple dmap words,
2255 * so we may update only a subset of the actual leaves corresponding
2256 * to the dmap words of the block range.
2257 *
2258 * dbJoin() is used to update leaf values and will join the binary
2259 * buddy system of the leaves if the new leaf values indicate this
2260 * should be done.
2261 */
2263 /* determine the bit number within the word and
2264 * the number of bits within the word.
2265 */
2269 /* check if only part of a word is to be freed.
2270 */
2272 /* free (zero) the appropriate bits within this
2273 * dmap word.
2274 */
2279 /* update the leaf for this dmap word.
2280 */
2288 /* one or more dmap words are fully contained
2289 * within the block range. determine how many
2290 * words and free (zero) the bits of these words.
2291 */
2295 /* determine how many bits.
2296 */
2299 /* now update the appropriate leaves to reflect
2300 * the freed words.
2301 */
2303 /* determine what the leaf value should be
2304 * updated to as the minimum of the l2 number
2305 * of bits being freed and the l2 (max) number
2306 * of bits that can be described by this leaf.
2307 */
2308 size =
2313 /* update the leaf.
2314 */
2319 /* get the number of dmap words handled.
2320 */
2323 }
2324 }
2325 }
2327 /* update the free count for this dmap.
2328 */
2333 /* update the free count for the allocation group and
2334 * map.
2335 */
2340 /* check if this allocation group is not completely free and
2341 * if it is currently the maximum (rightmost) allocation group.
2342 * if so, establish the new maximum allocation group number by
2343 * searching left for the first allocation group with allocation.
2344 */
2353 }
2355 /* re-establish the allocation group preference if the
2356 * current preference is right of the maximum allocation
2357 * group.
2358 */
2361 }
2366 }
2369 /*
2370 * NAME: dbAdjCtl()
2371 *
2372 * FUNCTION: adjust a dmap control page at a specified level to reflect
2373 * the change in a lower level dmap or dmap control page's
2374 * maximum string of free blocks (i.e. a change in the root
2375 * of the lower level object's dmtree) due to the allocation
2376 * or deallocation of a range of blocks with a single dmap.
2377 *
2378 * on entry, this routine is provided with the new value of
2379 * the lower level dmap or dmap control page root and the
2380 * starting block number of the block range whose allocation
2381 * or deallocation resulted in the root change. this range
2382 * is respresented by a single leaf of the current dmapctl
2383 * and the leaf will be updated with this value, possibly
2384 * causing a binary buddy system within the leaves to be
2385 * split or joined. the update may also cause the dmapctl's
2386 * dmtree to be updated.
2387 *
2388 * if the adjustment of the dmap control page, itself, causes its
2389 * root to change, this change will be bubbled up to the next dmap
2390 * control level by a recursive call to this routine, specifying
2391 * the new root value and the next dmap control page level to
2392 * be adjusted.
2393 * PARAMETERS:
2394 * bmp - pointer to bmap descriptor
2395 * blkno - the first block of a block range within a dmap. it is
2396 * the allocation or deallocation of this block range that
2397 * requires the dmap control page to be adjusted.
2398 * newval - the new value of the lower level dmap or dmap control
2399 * page root.
2400 * alloc - TRUE if adjustment is due to an allocation.
2401 * level - current level of dmap control page (i.e. L0, L1, L2) to
2402 * be adjusted.
2403 *
2404 * RETURN VALUES:
2405 * 0 - success
2406 * -EIO - i/o error
2407 *
2408 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2409 */
2410 static int
2412 {
2414 s8 oldroot;
2416 s64 lblkno;
2420 /* get the buffer for the dmap control page for the specified
2421 * block number and control page level.
2422 */
2434 }
2436 /* determine the leaf number corresponding to the block and
2437 * the index within the dmap control tree.
2438 */
2442 /* save the current leaf value and the current root level (i.e.
2443 * maximum l2 free string described by this dmapctl).
2444 */
2448 /* check if this is a control page update for an allocation.
2449 * if so, update the leaf to reflect the new leaf value using
2450 * dbSplit(); otherwise (deallocation), use dbJoin() to udpate
2451 * the leaf with the new value. in addition to updating the
2452 * leaf, dbSplit() will also split the binary buddy system of
2453 * the leaves, if required, and bubble new values within the
2454 * dmapctl tree, if required. similarly, dbJoin() will join
2455 * the binary buddy system of leaves and bubble new values up
2456 * the dmapctl tree as required by the new leaf value.
2457 */
2459 /* check if we are in the middle of a binary buddy
2460 * system. this happens when we are performing the
2461 * first allocation out of an allocation group that
2462 * is part (not the first part) of a larger binary
2463 * buddy system. if we are in the middle, back split
2464 * the system prior to calling dbSplit() which assumes
2465 * that it is at the front of a binary buddy system.
2466 */
2472 }
2478 }
2480 /* check if the root of the current dmap control page changed due
2481 * to the update and if the current dmap control page is not at
2482 * the current top level (i.e. L0, L1, L2) of the map. if so (i.e.
2483 * root changed and this is not the top level), call this routine
2484 * again (recursion) for the next higher level of the mapping to
2485 * reflect the change in root for the current dmap control page.
2486 */
2488 /* are we below the top level of the map. if so,
2489 * bubble the root up to the next higher level.
2490 */
2492 /* bubble up the new root of this dmap control page to
2493 * the next level.
2494 */
2498 /* something went wrong in bubbling up the new
2499 * root value, so backout the changes to the
2500 * current dmap control page.
2501 */
2504 oldval);
2506 /* the dbJoin() above might have
2507 * caused a larger binary buddy system
2508 * to form and we may now be in the
2509 * middle of it. if this is the case,
2510 * back split the buddies.
2511 */
2517 }
2519 /* release the buffer and return the error.
2520 */
2523 }
2525 /* we're at the top level of the map. update
2526 * the bmap control page to reflect the size
2527 * of the maximum free buddy system.
2528 */
2532 "dbAdjCtl: the maximum free buddy is "
2534 }
2536 }
2537 }
2539 /* write the buffer.
2540 */
2544 }
2547 /*
2548 * NAME: dbSplit()
2549 *
2550 * FUNCTION: update the leaf of a dmtree with a new value, splitting
2551 * the leaf from the binary buddy system of the dmtree's
2552 * leaves, as required.
2553 *
2554 * PARAMETERS:
2555 * tp - pointer to the tree containing the leaf.
2556 * leafno - the number of the leaf to be updated.
2557 * splitsz - the size the binary buddy system starting at the leaf
2558 * must be split to, specified as the log2 number of blocks.
2559 * newval - the new value for the leaf.
2560 *
2561 * RETURN VALUES: none
2562 *
2563 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2564 */
2566 {
2571 /* check if the leaf needs to be split.
2572 */
2574 /* the split occurs by cutting the buddy system in half
2575 * at the specified leaf until we reach the specified
2576 * size. pick up the starting split size (current size
2577 * - 1 in l2) and the corresponding buddy size.
2578 */
2582 /* split until we reach the specified size.
2583 */
2585 /* update the buddy's leaf with its new value.
2586 */
2589 /* on to the next size and buddy.
2590 */
2593 }
2594 }
2596 /* adjust the dmap tree to reflect the specified leaf's new
2597 * value.
2598 */
2600 }
2603 /*
2604 * NAME: dbBackSplit()
2605 *
2606 * FUNCTION: back split the binary buddy system of dmtree leaves
2607 * that hold a specified leaf until the specified leaf
2608 * starts its own binary buddy system.
2609 *
2610 * the allocators typically perform allocations at the start
2611 * of binary buddy systems and dbSplit() is used to accomplish
2612 * any required splits. in some cases, however, allocation
2613 * may occur in the middle of a binary system and requires a
2614 * back split, with the split proceeding out from the middle of
2615 * the system (less efficient) rather than the start of the
2616 * system (more efficient). the cases in which a back split
2617 * is required are rare and are limited to the first allocation
2618 * within an allocation group which is a part (not first part)
2619 * of a larger binary buddy system and a few exception cases
2620 * in which a previous join operation must be backed out.
2621 *
2622 * PARAMETERS:
2623 * tp - pointer to the tree containing the leaf.
2624 * leafno - the number of the leaf to be updated.
2625 *
2626 * RETURN VALUES: none
2627 *
2628 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2629 */
2631 {
2636 /* leaf should be part (not first part) of a binary
2637 * buddy system.
2638 */
2641 /* the back split is accomplished by iteratively finding the leaf
2642 * that starts the buddy system that contains the specified leaf and
2643 * splitting that system in two. this iteration continues until
2644 * the specified leaf becomes the start of a buddy system.
2645 *
2646 * determine maximum possible l2 size for the specified leaf.
2647 */
2648 size =
2652 /* determine the number of leaves covered by this size. this
2653 * is the buddy size that we will start with as we search for
2654 * the buddy system that contains the specified leaf.
2655 */
2658 /* back split.
2659 */
2661 /* find the leftmost buddy leaf.
2662 */
2668 }
2670 /* determine the buddy.
2671 */
2674 /* check if this buddy is the start of the system.
2675 */
2677 /* split the leaf at the start of the
2678 * system in two.
2679 */
2683 }
2684 }
2685 }
2690 }
2692 }
2695 /*
2696 * NAME: dbJoin()
2697 *
2698 * FUNCTION: update the leaf of a dmtree with a new value, joining
2699 * the leaf with other leaves of the dmtree into a multi-leaf
2700 * binary buddy system, as required.
2701 *
2702 * PARAMETERS:
2703 * tp - pointer to the tree containing the leaf.
2704 * leafno - the number of the leaf to be updated.
2705 * newval - the new value for the leaf.
2706 *
2707 * RETURN VALUES: none
2708 */
2710 {
2714 /* can the new leaf value require a join with other leaves ?
2715 */
2717 /* pickup a pointer to the leaves of the tree.
2718 */
2721 /* try to join the specified leaf into a large binary
2722 * buddy system. the join proceeds by attempting to join
2723 * the specified leafno with its buddy (leaf) at new value.
2724 * if the join occurs, we attempt to join the left leaf
2725 * of the joined buddies with its buddy at new value + 1.
2726 * we continue to join until we find a buddy that cannot be
2727 * joined (does not have a value equal to the size of the
2728 * last join) or until all leaves have been joined into a
2729 * single system.
2730 *
2731 * get the buddy size (number of words covered) of
2732 * the new value.
2733 */
2736 /* try to join.
2737 */
2739 /* get the buddy leaf.
2740 */
2743 /* if the leaf's new value is greater than its
2744 * buddy's value, we join no more.
2745 */
2749 /* It shouldn't be less */
2753 /* check which (leafno or buddy) is the left buddy.
2754 * the left buddy gets to claim the blocks resulting
2755 * from the join while the right gets to claim none.
2756 * the left buddy is also eligable to participate in
2757 * a join at the next higher level while the right
2758 * is not.
2759 *
2760 */
2762 /* leafno is the left buddy.
2763 */
2766 /* buddy is the left buddy and becomes
2767 * leafno.
2768 */
2771 }
2773 /* on to try the next join.
2774 */
2777 }
2778 }
2780 /* update the leaf value.
2781 */
2785 }
2788 /*
2789 * NAME: dbAdjTree()
2790 *
2791 * FUNCTION: update a leaf of a dmtree with a new value, adjusting
2792 * the dmtree, as required, to reflect the new leaf value.
2793 * the combination of any buddies must already be done before
2794 * this is called.
2795 *
2796 * PARAMETERS:
2797 * tp - pointer to the tree to be adjusted.
2798 * leafno - the number of the leaf to be updated.
2799 * newval - the new value for the leaf.
2800 *
2801 * RETURN VALUES: none
2802 */
2804 {
2808 /* pick up the index of the leaf for this leafno.
2809 */
2812 /* is the current value the same as the old value ? if so,
2813 * there is nothing to do.
2814 */
2818 /* set the new value.
2819 */
2822 /* bubble the new value up the tree as required.
2823 */
2825 /* get the index of the first leaf of the 4 leaf
2826 * group containing the specified leaf (leafno).
2827 */
2830 /* get the index of the parent of this 4 leaf group.
2831 */
2834 /* determine the maximum of the 4 leaves.
2835 */
2838 /* if the maximum of the 4 is the same as the
2839 * parent's value, we're done.
2840 */
2844 /* parent gets new value.
2845 */
2848 /* parent becomes leaf for next go-round.
2849 */
2851 }
2852 }
2855 /*
2856 * NAME: dbFindLeaf()
2857 *
2858 * FUNCTION: search a dmtree_t for sufficient free blocks, returning
2859 * the index of a leaf describing the free blocks if
2860 * sufficient free blocks are found.
2861 *
2862 * the search starts at the top of the dmtree_t tree and
2863 * proceeds down the tree to the leftmost leaf with sufficient
2864 * free space.
2865 *
2866 * PARAMETERS:
2867 * tp - pointer to the tree to be searched.
2868 * l2nb - log2 number of free blocks to search for.
2869 * leafidx - return pointer to be set to the index of the leaf
2870 * describing at least l2nb free blocks if sufficient
2871 * free blocks are found.
2872 *
2873 * RETURN VALUES:
2874 * 0 - success
2875 * -ENOSPC - insufficient free blocks.
2876 */
2878 {
2881 /* first check the root of the tree to see if there is
2882 * sufficient free space.
2883 */
2887 /* sufficient free space available. now search down the tree
2888 * starting at the next level for the leftmost leaf that
2889 * describes sufficient free space.
2890 */
2893 /* search the four nodes at this level, starting from
2894 * the left.
2895 */
2897 /* sufficient free space found. move to the next
2898 * level (or quit if this is the last level).
2899 */
2902 }
2904 /* better have found something since the higher
2905 * levels of the tree said it was here.
2906 */
2908 }
2910 /* set the return to the leftmost leaf describing sufficient
2911 * free space.
2912 */
2916 }
2919 /*
2920 * NAME: dbFindBits()
2921 *
2922 * FUNCTION: find a specified number of binary buddy free bits within a
2923 * dmap bitmap word value.
2924 *
2925 * this routine searches the bitmap value for (1 << l2nb) free
2926 * bits at (1 << l2nb) alignments within the value.
2927 *
2928 * PARAMETERS:
2929 * word - dmap bitmap word value.
2930 * l2nb - number of free bits specified as a log2 number.
2931 *
2932 * RETURN VALUES:
2933 * starting bit number of free bits.
2934 */
2936 {
2938 u32 mask;
2940 /* get the number of bits.
2941 */
2945 /* complement the word so we can use a mask (i.e. 0s represent
2946 * free bits) and compute the mask.
2947 */
2951 /* scan the word for nb free bits at nb alignments.
2952 */
2956 }
2960 /* return the bit number.
2961 */
2963 }
2966 /*
2967 * NAME: dbMaxBud(u8 *cp)
2968 *
2969 * FUNCTION: determine the largest binary buddy string of free
2970 * bits within 32-bits of the map.
2971 *
2972 * PARAMETERS:
2973 * cp - pointer to the 32-bit value.
2974 *
2975 * RETURN VALUES:
2976 * largest binary buddy of free bits within a dmap word.
2977 */
2979 {
2982 /* check if the wmap word is all free. if so, the
2983 * free buddy size is BUDMIN.
2984 */
2988 /* check if the wmap word is half free. if so, the
2989 * free buddy size is BUDMIN-1.
2990 */
2994 /* not all free or half free. determine the free buddy
2995 * size thru table lookup using quarters of the wmap word.
2996 */
3000 }
3003 /*
3004 * NAME: cnttz(uint word)
3005 *
3006 * FUNCTION: determine the number of trailing zeros within a 32-bit
3007 * value.
3008 *
3009 * PARAMETERS:
3010 * value - 32-bit value to be examined.
3011 *
3012 * RETURN VALUES:
3013 * count of trailing zeros
3014 */
3016 {
3022 }
3025 }
3028 /*
3029 * NAME: cntlz(u32 value)
3030 *
3031 * FUNCTION: determine the number of leading zeros within a 32-bit
3032 * value.
3033 *
3034 * PARAMETERS:
3035 * value - 32-bit value to be examined.
3036 *
3037 * RETURN VALUES:
3038 * count of leading zeros
3039 */
3041 {
3047 }
3049 }
3052 /*
3053 * NAME: blkstol2(s64 nb)
3054 *
3055 * FUNCTION: convert a block count to its log2 value. if the block
3056 * count is not a l2 multiple, it is rounded up to the next
3057 * larger l2 multiple.
3058 *
3059 * PARAMETERS:
3060 * nb - number of blocks
3061 *
3062 * RETURN VALUES:
3063 * log2 number of blocks
3064 */
3066 {
3072 /* count the leading bits.
3073 */
3075 /* leading bit found.
3076 */
3078 /* determine the l2 value.
3079 */
3082 /* check if we need to round up.
3083 */
3085 l2nb++;
3088 }
3089 }
3092 }
3095 /*
3096 * NAME: dbAllocBottomUp()
3097 *
3098 * FUNCTION: alloc the specified block range from the working block
3099 * allocation map.
3100 *
3101 * the blocks will be alloc from the working map one dmap
3102 * at a time.
3103 *
3104 * PARAMETERS:
3105 * ip - pointer to in-core inode;
3106 * blkno - starting block number to be freed.
3107 * nblocks - number of blocks to be freed.
3108 *
3109 * RETURN VALUES:
3110 * 0 - success
3111 * -EIO - i/o error
3112 */
3114 {
3124 /* block to be allocated better be within the mapsize. */
3127 /*
3128 * allocate the blocks a dmap at a time.
3129 */
3132 /* release previous dmap if any */
3135 }
3137 /* get the buffer for the current dmap. */
3143 }
3146 /* determine the number of blocks to be allocated from
3147 * this dmap.
3148 */
3151 /* allocate the blocks. */
3156 }
3157 }
3159 /* write the last buffer. */
3165 }
3170 {
3176 /* save the current value of the root (i.e. maximum free string)
3177 * of the dmap tree.
3178 */
3181 /* pick up a pointer to the leaves of the dmap tree */
3184 /* determine the bit number and word within the dmap of the
3185 * starting block.
3186 */
3190 /* block range better be within the dmap */
3193 /* allocate the bits of the dmap's words corresponding to the block
3194 * range. not all bits of the first and last words may be contained
3195 * within the block range. if this is the case, we'll work against
3196 * those words (i.e. partial first and/or last) on an individual basis
3197 * (a single pass), allocating the bits of interest by hand and
3198 * updating the leaf corresponding to the dmap word. a single pass
3199 * will be used for all dmap words fully contained within the
3200 * specified range. within this pass, the bits of all fully contained
3201 * dmap words will be marked as free in a single shot and the leaves
3202 * will be updated. a single leaf may describe the free space of
3203 * multiple dmap words, so we may update only a subset of the actual
3204 * leaves corresponding to the dmap words of the block range.
3205 */
3207 /* determine the bit number within the word and
3208 * the number of bits within the word.
3209 */
3213 /* check if only part of a word is to be allocated.
3214 */
3216 /* allocate (set to 1) the appropriate bits within
3217 * this dmap word.
3218 */
3222 word++;
3224 /* one or more dmap words are fully contained
3225 * within the block range. determine how many
3226 * words and allocate (set to 1) the bits of these
3227 * words.
3228 */
3232 /* determine how many bits */
3235 }
3236 }
3238 /* update the free count for this dmap */
3241 /* reconstruct summary tree */
3246 /* if this allocation group is completely free,
3247 * update the highest active allocation group number
3248 * if this allocation group is the new max.
3249 */
3254 /* update the free count for the allocation group and map */
3260 /* if the root has not changed, done. */
3264 /* root changed. bubble the change up to the dmap control pages.
3265 * if the adjustment of the upper level control pages fails,
3266 * backout the bit allocation (thus making everything consistent).
3267 */
3272 }
3275 /*
3276 * NAME: dbExtendFS()
3277 *
3278 * FUNCTION: extend bmap from blkno for nblocks;
3279 * dbExtendFS() updates bmap ready for dbAllocBottomUp();
3280 *
3281 * L2
3282 * |
3283 * L1---------------------------------L1
3284 * | |
3285 * L0---------L0---------L0 L0---------L0---------L0
3286 * | | | | | |
3287 * d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,.,dm;
3288 * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3289 *
3290 * <---old---><----------------------------extend----------------------->
3291 */
3293 {
3297 s64 newsize;
3298 s64 p;
3305 s64 ag_rem;
3312 /*
3313 * initialize bmap control page.
3314 *
3315 * all the data in bmap control page should exclude
3316 * the mkfs hidden dmap page.
3317 */
3319 /* update mapsize */
3323 /* compute new AG size */
3330 /* compute new number of AG */
3335 /*
3336 * reconfigure db_agfree[]
3337 * from old AG configuration to new AG configuration;
3338 *
3339 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3340 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3341 * note: new AG size = old AG size * (2**x).
3342 */
3350 /* coalesce cotiguous k AGs; */
3352 /* merge AGi to AGn */
3354 }
3355 }
3361 /*
3362 * update highest active ag number
3363 */
3367 /*
3368 * extend bmap
3369 *
3370 * update bit maps and corresponding level control pages;
3371 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3372 */
3373 extend:
3374 /* get L2 page */
3380 }
3383 /* compute start L1 */
3388 /*
3389 * extend each L1 in L2
3390 */
3392 /* get L1 page */
3394 /* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3400 /* compute start L0 */
3406 /* assign/init L1 page */
3413 /* compute start L0 */
3417 }
3419 /*
3420 * extend each L0 in L1
3421 */
3423 /* get L0 page */
3425 /* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3432 /* compute start dmap */
3434 L2BPERDMAP;
3440 /* assign/init L0 page */
3447 /* compute start dmap */
3451 }
3453 /*
3454 * extend each dmap in L0
3455 */
3457 /*
3458 * reconstruct the dmap page, and
3459 * initialize corresponding parent L0 leaf
3460 */
3462 /* read in dmap page: */
3469 /* assign/init dmap page */
3476 }
3487 l0leaf++;
3496 /*
3497 * build current L0 page from its leaves, and
3498 * initialize corresponding parent L1 leaf
3499 */
3507 /* more than 1 L0 ? */
3511 /* summarize in global bmap page */
3516 }
3517 }
3520 /*
3521 * build current L1 page from its leaves, and
3522 * initialize corresponding parent L2 leaf
3523 */
3531 /* more than 1 L1 ? */
3535 /* summarize in global bmap page */
3539 }
3540 }
3545 errout:
3553 /*
3554 * finalize bmap control page
3555 */
3556 finalize:
3559 }
3562 /*
3563 * dbFinalizeBmap()
3564 */
3566 {
3572 /*
3573 * finalize bmap control page
3574 */
3575 //finalize:
3576 /*
3577 * compute db_agpref: preferred ag to allocate from
3578 * (the leftmost ag with average free space in it);
3579 */
3580 //agpref:
3581 /* get the number of active ags and inacitve ags */
3586 /* determine how many blocks are in the inactive allocation
3587 * groups. in doing this, we must account for the fact that
3588 * the rightmost group might be a partial group (i.e. file
3589 * system size is not a multiple of the group size).
3590 */
3595 /* determine how many free blocks are in the active
3596 * allocation groups plus the average number of free blocks
3597 * within the active ags.
3598 */
3602 /* if the preferred allocation group has not average free space.
3603 * re-establish the preferred group as the leftmost
3604 * group with average free space.
3605 */
3611 }
3615 }
3616 }
3618 /*
3619 * compute db_aglevel, db_agheigth, db_width, db_agstart:
3620 * an ag is covered in aglevel dmapctl summary tree,
3621 * at agheight level height (from leaf) with agwidth number of nodes
3622 * each, which starts at agstart index node of the smmary tree node
3623 * array;
3624 */
3626 l2nl =
3631 i--) {
3634 }
3636 }
3639 /*
3640 * NAME: dbInitDmap()/ujfs_idmap_page()
3641 *
3642 * FUNCTION: initialize working/persistent bitmap of the dmap page
3643 * for the specified number of blocks:
3644 *
3645 * at entry, the bitmaps had been initialized as free (ZEROS);
3646 * The number of blocks will only account for the actually
3647 * existing blocks. Blocks which don't actually exist in
3648 * the aggregate will be marked as allocated (ONES);
3649 *
3650 * PARAMETERS:
3651 * dp - pointer to page of map
3652 * nblocks - number of blocks this page
3653 *
3654 * RETURNS: NONE
3655 */
3657 {
3660 /* starting block number within the dmap */
3671 }
3676 }
3678 /* word number containing start block number */
3681 /*
3682 * free the bits corresponding to the block range (ZEROS):
3683 * note: not all bits of the first and last words may be contained
3684 * within the block range.
3685 */
3687 /* number of bits preceding range to be freed in the word */
3689 /* number of bits to free in the word */
3692 /* is partial word to be freed ? */
3694 /* free (set to 0) from the bitmap word */
3700 /* skip the word freed */
3701 w++;
3703 /* free (set to 0) contiguous bitmap words */
3708 /* skip the words freed */
3711 }
3712 }
3714 /*
3715 * mark bits following the range to be freed (non-existing
3716 * blocks) as allocated (ONES)
3717 */
3722 /* the first word beyond the end of existing blocks */
3725 /* does nblocks fall on a 32-bit boundary ? */
3728 /* mark a partial word allocated */
3730 w++;
3731 }
3733 /* set the rest of the words in the page to allocated (ONES) */
3737 /*
3738 * init tree
3739 */
3740 initTree:
3742 }
3745 /*
3746 * NAME: dbInitDmapTree()/ujfs_complete_dmap()
3747 *
3748 * FUNCTION: initialize summary tree of the specified dmap:
3749 *
3750 * at entry, bitmap of the dmap has been initialized;
3751 *
3752 * PARAMETERS:
3753 * dp - dmap to complete
3754 * blkno - starting block number for this dmap
3755 * treemax - will be filled in with max free for this dmap
3756 *
3757 * RETURNS: max free string at the root of the tree
3758 */
3760 {
3765 /* init fixed info of tree */
3773 /* init each leaf from corresponding wmap word:
3774 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3775 * bitmap word are allocated.
3776 */
3781 /* build the dmap's binary buddy summary tree */
3783 }
3786 /*
3787 * NAME: dbInitTree()/ujfs_adjtree()
3788 *
3789 * FUNCTION: initialize binary buddy summary tree of a dmap or dmapctl.
3790 *
3791 * at entry, the leaves of the tree has been initialized
3792 * from corresponding bitmap word or root of summary tree
3793 * of the child control page;
3794 * configure binary buddy system at the leaf level, then
3795 * bubble up the values of the leaf nodes up the tree.
3796 *
3797 * PARAMETERS:
3798 * cp - Pointer to the root of the tree
3799 * l2leaves- Number of leaf nodes as a power of 2
3800 * l2min - Number of blocks that can be covered by a leaf
3801 * as a power of 2
3802 *
3803 * RETURNS: max free string at the root of the tree
3804 */
3806 {
3813 /* Determine the maximum free string possible for the leaves */
3816 /*
3817 * configure the leaf levevl into binary buddy system
3818 *
3819 * Try to combine buddies starting with a buddy size of 1
3820 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3821 * can be combined if both buddies have a maximum free of l2min;
3822 * the combination will result in the left-most buddy leaf having
3823 * a maximum free of l2min+1.
3824 * After processing all buddies for a given size, process buddies
3825 * at the next higher buddy size (i.e. current size * 2) and
3826 * the next maximum free (current free + 1).
3827 * This continues until the maximum possible buddy combination
3828 * yields maximum free.
3829 */
3832 /* get next buddy size == current buddy pair size */
3835 /* scan each adjacent buddy pair at current buddy size */
3839 /* coalesce if both adjacent buddies are max free */
3843 }
3844 }
3845 }
3847 /*
3848 * bubble summary information of leaves up the tree.
3849 *
3850 * Starting at the leaf node level, the four nodes described by
3851 * the higher level parent node are compared for a maximum free and
3852 * this maximum becomes the value of the parent node.
3853 * when all lower level nodes are processed in this fashion then
3854 * move up to the next level (parent becomes a lower level node) and
3855 * continue the process for that level.
3856 */
3860 /* get index of 1st node of parent level */
3863 /* set the value of the parent node as the maximum
3864 * of the four nodes of the current level.
3865 */
3869 }
3872 }
3875 /*
3876 * dbInitDmapCtl()
3877 *
3878 * function: initialize dmapctl page
3879 */
3890 /*
3891 * initialize the leaves of current level that were not covered
3892 * by the specified input block range (i.e. the leaves have no
3893 * low level dmapctl or dmap).
3894 */
3899 /* build the dmap's binary buddy summary tree */
3901 }
3904 /*
3905 * NAME: dbGetL2AGSize()/ujfs_getagl2size()
3906 *
3907 * FUNCTION: Determine log2(allocation group size) from aggregate size
3908 *
3909 * PARAMETERS:
3910 * nblocks - Number of blocks in aggregate
3911 *
3912 * RETURNS: log2(allocation group size) in aggregate blocks
3913 */
3915 {
3916 s64 sz;
3917 s64 m;
3923 /* round up aggregate size to power of 2 */
3928 }
3934 /* agsize = roundupSize/max_number_of_ag */
3936 }
3939 /*
3940 * NAME: dbMapFileSizeToMapSize()
3941 *
3942 * FUNCTION: compute number of blocks the block allocation map file
3943 * can cover from the map file size;
3944 *
3945 * RETURNS: Number of blocks which can be covered by this block map file;
3946 */
3948 /*
3949 * maximum number of map pages at each level including control pages
3950 */
3951 #define MAXL0PAGES (1 + LPERCTL)
3952 #define MAXL1PAGES (1 + LPERCTL * MAXL0PAGES)
3953 #define MAXL2PAGES (1 + LPERCTL * MAXL1PAGES)
3955 /*
3956 * convert number of map pages to the zero origin top dmapctl level
3957 */
3958 #define BMAPPGTOLEV(npages) \
3959 (((npages) <= 3 + MAXL0PAGES) ? 0 \
3960 : ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
3963 {
3965 s64 nblocks;
3974 /* At each level, accumulate the number of dmap pages covered by
3975 * the number of full child levels below it;
3976 * repeat for the last incomplete child level.
3977 */
3980 /* skip higher level control pages above top level covered by map */
3984 factor =
3990 /* pages in last/incomplete child */
3992 /* skip incomplete child's level control page */
3993 npages--;
3994 }
3996 /* convert the number of dmaps into the number of blocks
3997 * which can be covered by the dmaps;
3998 */
4002 }