| blob | e1985066b1c63273cf5bf568ea6f9a417abb0a7a |
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) mutex_init(&bmp->db_bmaplock)
68 #define BMAP_LOCK(bmp) mutex_lock(&bmp->db_bmaplock)
69 #define BMAP_UNLOCK(bmp) mutex_unlock(&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 */
537 /* inherit older/smaller lsn */
542 /* move bp after tblock in logsync list */
544 }
546 /* inherit younger/larger clsn */
555 /* insert bp after tblock in logsync list */
560 }
562 }
564 /* write the last buffer. */
567 }
570 }
573 /*
574 * NAME: dbNextAG()
575 *
576 * FUNCTION: find the preferred allocation group for new allocations.
577 *
578 * Within the allocation groups, we maintain a preferred
579 * allocation group which consists of a group with at least
580 * average free space. It is the preferred group that we target
581 * new inode allocation towards. The tie-in between inode
582 * allocation and block allocation occurs as we allocate the
583 * first (data) block of an inode and specify the inode (block)
584 * as the allocation hint for this block.
585 *
586 * We try to avoid having more than one open file growing in
587 * an allocation group, as this will lead to fragmentation.
588 * This differs from the old OS/2 method of trying to keep
589 * empty ags around for large allocations.
590 *
591 * PARAMETERS:
592 * ipbmap - pointer to in-core inode for the block map.
593 *
594 * RETURN VALUES:
595 * the preferred allocation group number.
596 */
598 {
599 s64 avgfree;
608 /* determine the average number of free blocks within the ags. */
611 /*
612 * if the current preferred ag does not have an active allocator
613 * and has at least average freespace, return it
614 */
620 /* From the last preferred ag, find the next one with at least
621 * average free space.
622 */
628 /* open file is currently growing in this ag */
631 /* Return this one */
635 /* Less than avg. freespace, but best so far */
638 }
639 }
641 /*
642 * If no inactive ag was found with average freespace, use the
643 * next best
644 */
647 /* else leave db_agpref unchanged */
648 unlock:
651 /* return the preferred group.
652 */
654 }
656 /*
657 * NAME: dbAlloc()
658 *
659 * FUNCTION: attempt to allocate a specified number of contiguous free
660 * blocks from the working allocation block map.
661 *
662 * the block allocation policy uses hints and a multi-step
663 * approach.
664 *
665 * for allocation requests smaller than the number of blocks
666 * per dmap, we first try to allocate the new blocks
667 * immediately following the hint. if these blocks are not
668 * available, we try to allocate blocks near the hint. if
669 * no blocks near the hint are available, we next try to
670 * allocate within the same dmap as contains the hint.
671 *
672 * if no blocks are available in the dmap or the allocation
673 * request is larger than the dmap size, we try to allocate
674 * within the same allocation group as contains the hint. if
675 * this does not succeed, we finally try to allocate anywhere
676 * within the aggregate.
677 *
678 * we also try to allocate anywhere within the aggregate for
679 * for allocation requests larger than the allocation group
680 * size or requests that specify no hint value.
681 *
682 * PARAMETERS:
683 * ip - pointer to in-core inode;
684 * hint - allocation hint.
685 * nblocks - number of contiguous blocks in the range.
686 * results - on successful return, set to the starting block number
687 * of the newly allocated contiguous range.
688 *
689 * RETURN VALUES:
690 * 0 - success
691 * -ENOSPC - insufficient disk resources
692 * -EIO - i/o error
693 */
695 {
703 s64 mapSize;
706 /* assert that nblocks is valid */
709 /* get the log2 number of blocks to be allocated.
710 * if the number of blocks is not a log2 multiple,
711 * it will be rounded up to the next log2 multiple.
712 */
719 /* the hint should be within the map */
723 }
725 /* if the number of blocks to be allocated is greater than the
726 * allocation group size, try to allocate anywhere.
727 */
734 }
736 /*
737 * If no hint, let dbNextAG recommend an allocation group
738 */
742 /* we would like to allocate close to the hint. adjust the
743 * hint to the block following the hint since the allocators
744 * will start looking for free space starting at this point.
745 */
753 /* check if blkno crosses over into a new allocation group.
754 * if so, check if we should allow allocations within this
755 * allocation group.
756 */
758 /* check if the AG is currenly being written to.
759 * if so, call dbNextAG() to find a non-busy
760 * AG with sufficient free space.
761 */
765 /* check if the allocation request size can be satisfied from a
766 * single dmap. if so, try to allocate from the dmap containing
767 * the hint using a tiered strategy.
768 */
772 /* get the buffer for the dmap containing the hint.
773 */
782 /* first, try to satisfy the allocation request with the
783 * blocks beginning at the hint.
784 */
790 }
794 }
799 /*
800 * Someone else is writing in this allocation
801 * group. To avoid fragmenting, try another ag
802 */
806 }
808 /* next, try to satisfy the allocation request with blocks
809 * near the hint.
810 */
819 }
821 /* try to satisfy the allocation request with blocks within
822 * the same dmap as the hint.
823 */
831 }
835 }
837 /* try to satisfy the allocation request with blocks within
838 * the same allocation group as the hint.
839 */
847 pref_ag:
848 /*
849 * Let dbNextAG recommend a preferred allocation group
850 */
854 /* Try to allocate within this allocation group. if that fails, try to
855 * allocate anywhere in the map.
856 */
860 write_unlock:
865 read_unlock:
869 }
871 #ifdef _NOTYET
872 /*
873 * NAME: dbAllocExact()
874 *
875 * FUNCTION: try to allocate the requested extent;
876 *
877 * PARAMETERS:
878 * ip - pointer to in-core inode;
879 * blkno - extent address;
880 * nblocks - extent length;
881 *
882 * RETURN VALUES:
883 * 0 - success
884 * -ENOSPC - insufficient disk resources
885 * -EIO - i/o error
886 */
888 {
893 s64 lblkno;
898 /*
899 * validate extent request:
900 *
901 * note: defragfs policy:
902 * max 64 blocks will be moved.
903 * allocation request size must be satisfied from a single dmap.
904 */
908 }
911 /* the free space is no longer available */
914 }
916 /* read in the dmap covering the extent */
922 }
925 /* try to allocate the requested extent */
936 }
939 /*
940 * NAME: dbReAlloc()
941 *
942 * FUNCTION: attempt to extend a current allocation by a specified
943 * number of blocks.
944 *
945 * this routine attempts to satisfy the allocation request
946 * by first trying to extend the existing allocation in
947 * place by allocating the additional blocks as the blocks
948 * immediately following the current allocation. if these
949 * blocks are not available, this routine will attempt to
950 * allocate a new set of contiguous blocks large enough
951 * to cover the existing allocation plus the additional
952 * number of blocks required.
953 *
954 * PARAMETERS:
955 * ip - pointer to in-core inode requiring allocation.
956 * blkno - starting block of the current allocation.
957 * nblocks - number of contiguous blocks within the current
958 * allocation.
959 * addnblocks - number of blocks to add to the allocation.
960 * results - on successful return, set to the starting block number
961 * of the existing allocation if the existing allocation
962 * was extended in place or to a newly allocated contiguous
963 * range if the existing allocation could not be extended
964 * in place.
965 *
966 * RETURN VALUES:
967 * 0 - success
968 * -ENOSPC - insufficient disk resources
969 * -EIO - i/o error
970 */
971 int
974 {
977 /* try to extend the allocation in place.
978 */
985 }
987 /* could not extend the allocation in place, so allocate a
988 * new set of blocks for the entire request (i.e. try to get
989 * a range of contiguous blocks large enough to cover the
990 * existing allocation plus the additional blocks.)
991 */
994 }
997 /*
998 * NAME: dbExtend()
999 *
1000 * FUNCTION: attempt to extend a current allocation by a specified
1001 * number of blocks.
1002 *
1003 * this routine attempts to satisfy the allocation request
1004 * by first trying to extend the existing allocation in
1005 * place by allocating the additional blocks as the blocks
1006 * immediately following the current allocation.
1007 *
1008 * PARAMETERS:
1009 * ip - pointer to in-core inode requiring allocation.
1010 * blkno - starting block of the current allocation.
1011 * nblocks - number of contiguous blocks within the current
1012 * allocation.
1013 * addnblocks - number of blocks to add to the allocation.
1014 *
1015 * RETURN VALUES:
1016 * 0 - success
1017 * -ENOSPC - insufficient disk resources
1018 * -EIO - i/o error
1019 */
1021 {
1024 uint rel_block;
1031 /*
1032 * We don't want a non-aligned extent to cross a page boundary
1033 */
1038 /* get the last block of the current allocation */
1041 /* determine the block number of the block following
1042 * the existing allocation.
1043 */
1048 /* better be within the file system */
1055 }
1057 /* we'll attempt to extend the current allocation in place by
1058 * allocating the additional blocks as the blocks immediately
1059 * following the current allocation. we only try to extend the
1060 * current allocation in place if the number of additional blocks
1061 * can fit into a dmap, the last block of the current allocation
1062 * is not the last block of the file system, and the start of the
1063 * inplace extension is not on an allocation group boundary.
1064 */
1069 }
1071 /* get the buffer for the dmap containing the first block
1072 * of the extension.
1073 */
1079 }
1083 /* try to allocate the blocks immediately following the
1084 * current allocation.
1085 */
1090 /* were we successful ? */
1093 else
1094 /* we were not successful */
1099 }
1102 /*
1103 * NAME: dbAllocNext()
1104 *
1105 * FUNCTION: attempt to allocate the blocks of the specified block
1106 * range within a dmap.
1107 *
1108 * PARAMETERS:
1109 * bmp - pointer to bmap descriptor
1110 * dp - pointer to dmap.
1111 * blkno - starting block number of the range.
1112 * nblocks - number of contiguous free blocks of the range.
1113 *
1114 * RETURN VALUES:
1115 * 0 - success
1116 * -ENOSPC - insufficient disk resources
1117 * -EIO - i/o error
1118 *
1119 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1120 */
1123 {
1127 u32 mask;
1133 }
1135 /* pick up a pointer to the leaves of the dmap tree.
1136 */
1139 /* determine the bit number and word within the dmap of the
1140 * starting block.
1141 */
1145 /* check if the specified block range is contained within
1146 * this dmap.
1147 */
1151 /* check if the starting leaf indicates that anything
1152 * is free.
1153 */
1157 /* check the dmaps words corresponding to block range to see
1158 * if the block range is free. not all bits of the first and
1159 * last words may be contained within the block range. if this
1160 * is the case, we'll work against those words (i.e. partial first
1161 * and/or last) on an individual basis (a single pass) and examine
1162 * the actual bits to determine if they are free. a single pass
1163 * will be used for all dmap words fully contained within the
1164 * specified range. within this pass, the leaves of the dmap
1165 * tree will be examined to determine if the blocks are free. a
1166 * single leaf may describe the free space of multiple dmap
1167 * words, so we may visit only a subset of the actual leaves
1168 * corresponding to the dmap words of the block range.
1169 */
1171 /* determine the bit number within the word and
1172 * the number of bits within the word.
1173 */
1177 /* check if only part of the word is to be examined.
1178 */
1180 /* check if the bits are free.
1181 */
1188 /* one or more dmap words are fully contained
1189 * within the block range. determine how many
1190 * words and how many bits.
1191 */
1195 /* now examine the appropriate leaves to determine
1196 * if the blocks are free.
1197 */
1199 /* does the leaf describe any free space ?
1200 */
1204 /* determine the l2 number of bits provided
1205 * by this leaf.
1206 */
1207 l2size =
1210 /* determine how many words were handled.
1211 */
1216 }
1217 }
1218 }
1220 /* allocate the blocks.
1221 */
1223 }
1226 /*
1227 * NAME: dbAllocNear()
1228 *
1229 * FUNCTION: attempt to allocate a number of contiguous free blocks near
1230 * a specified block (hint) within a dmap.
1231 *
1232 * starting with the dmap leaf that covers the hint, we'll
1233 * check the next four contiguous leaves for sufficient free
1234 * space. if sufficient free space is found, we'll allocate
1235 * the desired free space.
1236 *
1237 * PARAMETERS:
1238 * bmp - pointer to bmap descriptor
1239 * dp - pointer to dmap.
1240 * blkno - block number to allocate near.
1241 * nblocks - actual number of contiguous free blocks desired.
1242 * l2nb - log2 number of contiguous free blocks desired.
1243 * results - on successful return, set to the starting block number
1244 * of the newly allocated range.
1245 *
1246 * RETURN VALUES:
1247 * 0 - success
1248 * -ENOSPC - insufficient disk resources
1249 * -EIO - i/o error
1250 *
1251 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1252 */
1253 static int
1256 {
1264 }
1268 /* determine the word within the dmap that holds the hint
1269 * (i.e. blkno). also, determine the last word in the dmap
1270 * that we'll include in our examination.
1271 */
1275 /* examine the leaves for sufficient free space.
1276 */
1278 /* does the leaf describe sufficient free space ?
1279 */
1283 /* determine the block number within the file system
1284 * of the first block described by this dmap word.
1285 */
1288 /* if not all bits of the dmap word are free, get the
1289 * starting bit number within the dmap word of the required
1290 * string of free bits and adjust the block number with the
1291 * value.
1292 */
1294 blkno +=
1297 /* allocate the blocks.
1298 */
1303 }
1306 }
1309 /*
1310 * NAME: dbAllocAG()
1311 *
1312 * FUNCTION: attempt to allocate the specified number of contiguous
1313 * free blocks within the specified allocation group.
1314 *
1315 * unless the allocation group size is equal to the number
1316 * of blocks per dmap, the dmap control pages will be used to
1317 * find the required free space, if available. we start the
1318 * search at the highest dmap control page level which
1319 * distinctly describes the allocation group's free space
1320 * (i.e. the highest level at which the allocation group's
1321 * free space is not mixed in with that of any other group).
1322 * in addition, we start the search within this level at a
1323 * height of the dmapctl dmtree at which the nodes distinctly
1324 * describe the allocation group's free space. at this height,
1325 * the allocation group's free space may be represented by 1
1326 * or two sub-trees, depending on the allocation group size.
1327 * we search the top nodes of these subtrees left to right for
1328 * sufficient free space. if sufficient free space is found,
1329 * the subtree is searched to find the leftmost leaf that
1330 * has free space. once we have made it to the leaf, we
1331 * move the search to the next lower level dmap control page
1332 * corresponding to this leaf. we continue down the dmap control
1333 * pages until we find the dmap that contains or starts the
1334 * sufficient free space and we allocate at this dmap.
1335 *
1336 * if the allocation group size is equal to the dmap size,
1337 * we'll start at the dmap corresponding to the allocation
1338 * group and attempt the allocation at this level.
1339 *
1340 * the dmap control page search is also not performed if the
1341 * allocation group is completely free and we go to the first
1342 * dmap of the allocation group to do the allocation. this is
1343 * done because the allocation group may be part (not the first
1344 * part) of a larger binary buddy system, causing the dmap
1345 * control pages to indicate no free space (NOFREE) within
1346 * the allocation group.
1347 *
1348 * PARAMETERS:
1349 * bmp - pointer to bmap descriptor
1350 * agno - allocation group number.
1351 * nblocks - actual number of contiguous free blocks desired.
1352 * l2nb - log2 number of contiguous free blocks desired.
1353 * results - on successful return, set to the starting block number
1354 * of the newly allocated range.
1355 *
1356 * RETURN VALUES:
1357 * 0 - success
1358 * -ENOSPC - insufficient disk resources
1359 * -EIO - i/o error
1360 *
1361 * note: IWRITE_LOCK(ipmap) held on entry/exit;
1362 */
1363 static int
1365 {
1372 /* allocation request should not be for more than the
1373 * allocation group size.
1374 */
1377 "dbAllocAG: allocation request is larger than the "
1380 }
1382 /* determine the starting block number of the allocation
1383 * group.
1384 */
1387 /* check if the allocation group size is the minimum allocation
1388 * group size or if the allocation group is completely free. if
1389 * the allocation group size is the minimum size of BPERDMAP (i.e.
1390 * 1 dmap), there is no need to search the dmap control page (below)
1391 * that fully describes the allocation group since the allocation
1392 * group is already fully described by a dmap. in this case, we
1393 * just call dbAllocCtl() to search the dmap tree and allocate the
1394 * required space if available.
1395 *
1396 * if the allocation group is completely free, dbAllocCtl() is
1397 * also called to allocate the required space. this is done for
1398 * two reasons. first, it makes no sense searching the dmap control
1399 * pages for free space when we know that free space exists. second,
1400 * the dmap control pages may indicate that the allocation group
1401 * has no free space if the allocation group is part (not the first
1402 * part) of a larger binary buddy system.
1403 */
1414 }
1416 }
1418 /* the buffer for the dmap control page that fully describes the
1419 * allocation group.
1420 */
1433 }
1435 /* search the subtree(s) of the dmap control page that describes
1436 * the allocation group, looking for sufficient free space. to begin,
1437 * determine how many allocation groups are represented in a dmap
1438 * control page at the control page level (i.e. L0, L1, L2) that
1439 * fully describes an allocation group. next, determine the starting
1440 * tree index of this allocation group within the control page.
1441 */
1442 agperlev =
1446 /* dmap control page trees fan-out by 4 and a single allocation
1447 * group may be described by 1 or 2 subtrees within the ag level
1448 * dmap control page, depending upon the ag size. examine the ag's
1449 * subtrees for sufficient free space, starting with the leftmost
1450 * subtree.
1451 */
1453 /* is there sufficient free space ?
1454 */
1458 /* sufficient free space found in a subtree. now search down
1459 * the subtree to find the leftmost leaf that describes this
1460 * free space.
1461 */
1467 }
1468 }
1474 }
1475 }
1477 /* determine the block number within the file system
1478 * that corresponds to this leaf.
1479 */
1487 blkno +=
1490 /* release the buffer in preparation for going down
1491 * the next level of dmap control pages.
1492 */
1495 /* check if we need to continue to search down the lower
1496 * level dmap control pages. we need to if the number of
1497 * blocks required is less than maximum number of blocks
1498 * described at the next lower level.
1499 */
1502 /* search the lower level dmap control pages to get
1503 * the starting block number of the dmap that
1504 * contains or starts off the free space.
1505 */
1511 "dbAllocAG: control page "
1514 }
1516 }
1517 }
1519 /* allocate the blocks.
1520 */
1526 }
1528 }
1530 /* no space in the allocation group. release the buffer and
1531 * return -ENOSPC.
1532 */
1536 }
1539 /*
1540 * NAME: dbAllocAny()
1541 *
1542 * FUNCTION: attempt to allocate the specified number of contiguous
1543 * free blocks anywhere in the file system.
1544 *
1545 * dbAllocAny() attempts to find the sufficient free space by
1546 * searching down the dmap control pages, starting with the
1547 * highest level (i.e. L0, L1, L2) control page. if free space
1548 * large enough to satisfy the desired free space is found, the
1549 * desired free space is allocated.
1550 *
1551 * PARAMETERS:
1552 * bmp - pointer to bmap descriptor
1553 * nblocks - actual number of contiguous free blocks desired.
1554 * l2nb - log2 number of contiguous free blocks desired.
1555 * results - on successful return, set to the starting block number
1556 * of the newly allocated range.
1557 *
1558 * RETURN VALUES:
1559 * 0 - success
1560 * -ENOSPC - insufficient disk resources
1561 * -EIO - i/o error
1562 *
1563 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1564 */
1566 {
1570 /* starting with the top level dmap control page, search
1571 * down the dmap control levels for sufficient free space.
1572 * if free space is found, dbFindCtl() returns the starting
1573 * block number of the dmap that contains or starts off the
1574 * range of free space.
1575 */
1579 /* allocate the blocks.
1580 */
1586 }
1588 }
1591 /*
1592 * NAME: dbFindCtl()
1593 *
1594 * FUNCTION: starting at a specified dmap control page level and block
1595 * number, search down the dmap control levels for a range of
1596 * contiguous free blocks large enough to satisfy an allocation
1597 * request for the specified number of free blocks.
1598 *
1599 * if sufficient contiguous free blocks are found, this routine
1600 * returns the starting block number within a dmap page that
1601 * contains or starts a range of contiqious free blocks that
1602 * is sufficient in size.
1603 *
1604 * PARAMETERS:
1605 * bmp - pointer to bmap descriptor
1606 * level - starting dmap control page level.
1607 * l2nb - log2 number of contiguous free blocks desired.
1608 * *blkno - on entry, starting block number for conducting the search.
1609 * on successful return, the first block within a dmap page
1610 * that contains or starts a range of contiguous free blocks.
1611 *
1612 * RETURN VALUES:
1613 * 0 - success
1614 * -ENOSPC - insufficient disk resources
1615 * -EIO - i/o error
1616 *
1617 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1618 */
1620 {
1627 /* starting at the specified dmap control page level and block
1628 * number, search down the dmap control levels for the starting
1629 * block number of a dmap page that contains or starts off
1630 * sufficient free blocks.
1631 */
1633 /* get the buffer of the dmap control page for the block
1634 * number and level (i.e. L0, L1, L2).
1635 */
1648 }
1650 /* search the tree within the dmap control page for
1651 * sufficent free space. if sufficient free space is found,
1652 * dbFindLeaf() returns the index of the leaf at which
1653 * free space was found.
1654 */
1657 /* release the buffer.
1658 */
1661 /* space found ?
1662 */
1668 }
1670 }
1672 /* adjust the block number to reflect the location within
1673 * the dmap control page (i.e. the leaf) at which free
1674 * space was found.
1675 */
1678 /* we stop the search at this dmap control page level if
1679 * the number of blocks required is greater than or equal
1680 * to the maximum number of blocks described at the next
1681 * (lower) level.
1682 */
1685 }
1689 }
1692 /*
1693 * NAME: dbAllocCtl()
1694 *
1695 * FUNCTION: attempt to allocate a specified number of contiguous
1696 * blocks starting within a specific dmap.
1697 *
1698 * this routine is called by higher level routines that search
1699 * the dmap control pages above the actual dmaps for contiguous
1700 * free space. the result of successful searches by these
1701 * routines are the starting block numbers within dmaps, with
1702 * the dmaps themselves containing the desired contiguous free
1703 * space or starting a contiguous free space of desired size
1704 * that is made up of the blocks of one or more dmaps. these
1705 * calls should not fail due to insufficent resources.
1706 *
1707 * this routine is called in some cases where it is not known
1708 * whether it will fail due to insufficient resources. more
1709 * specifically, this occurs when allocating from an allocation
1710 * group whose size is equal to the number of blocks per dmap.
1711 * in this case, the dmap control pages are not examined prior
1712 * to calling this routine (to save pathlength) and the call
1713 * might fail.
1714 *
1715 * for a request size that fits within a dmap, this routine relies
1716 * upon the dmap's dmtree to find the requested contiguous free
1717 * space. for request sizes that are larger than a dmap, the
1718 * requested free space will start at the first block of the
1719 * first dmap (i.e. blkno).
1720 *
1721 * PARAMETERS:
1722 * bmp - pointer to bmap descriptor
1723 * nblocks - actual number of contiguous free blocks to allocate.
1724 * l2nb - log2 number of contiguous free blocks to allocate.
1725 * blkno - starting block number of the dmap to start the allocation
1726 * from.
1727 * results - on successful return, set to the starting block number
1728 * of the newly allocated range.
1729 *
1730 * RETURN VALUES:
1731 * 0 - success
1732 * -ENOSPC - insufficient disk resources
1733 * -EIO - i/o error
1734 *
1735 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1736 */
1737 static int
1739 {
1745 /* check if the allocation request is confined to a single dmap.
1746 */
1748 /* get the buffer for the dmap.
1749 */
1756 /* try to allocate the blocks.
1757 */
1765 }
1767 /* allocation request involving multiple dmaps. it must start on
1768 * a dmap boundary.
1769 */
1772 /* allocate the blocks dmap by dmap.
1773 */
1775 /* get the buffer for the dmap.
1776 */
1782 }
1785 /* the dmap better be all free.
1786 */
1793 }
1795 /* determine how many blocks to allocate from this dmap.
1796 */
1799 /* allocate the blocks from the dmap.
1800 */
1804 }
1806 /* write the buffer.
1807 */
1809 }
1811 /* set the results (starting block number) and return.
1812 */
1816 /* something failed in handling an allocation request involving
1817 * multiple dmaps. we'll try to clean up by backing out any
1818 * allocation that has already happened for this request. if
1819 * we fail in backing out the allocation, we'll mark the file
1820 * system to indicate that blocks have been leaked.
1821 */
1822 backout:
1824 /* try to backout the allocations dmap by dmap.
1825 */
1828 /* get the buffer for this dmap.
1829 */
1833 /* could not back out. mark the file system
1834 * to indicate that we have leaked blocks.
1835 */
1839 }
1842 /* free the blocks is this dmap.
1843 */
1845 /* could not back out. mark the file system
1846 * to indicate that we have leaked blocks.
1847 */
1852 }
1854 /* write the buffer.
1855 */
1857 }
1860 }
1863 /*
1864 * NAME: dbAllocDmapLev()
1865 *
1866 * FUNCTION: attempt to allocate a specified number of contiguous blocks
1867 * from a specified dmap.
1868 *
1869 * this routine checks if the contiguous blocks are available.
1870 * if so, nblocks of blocks are allocated; otherwise, ENOSPC is
1871 * returned.
1872 *
1873 * PARAMETERS:
1874 * mp - pointer to bmap descriptor
1875 * dp - pointer to dmap to attempt to allocate blocks from.
1876 * l2nb - log2 number of contiguous block desired.
1877 * nblocks - actual number of contiguous block desired.
1878 * results - on successful return, set to the starting block number
1879 * of the newly allocated range.
1880 *
1881 * RETURN VALUES:
1882 * 0 - success
1883 * -ENOSPC - insufficient disk resources
1884 * -EIO - i/o error
1885 *
1886 * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
1887 * IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
1888 */
1889 static int
1892 {
1893 s64 blkno;
1896 /* can't be more than a dmaps worth of blocks */
1899 /* search the tree within the dmap page for sufficient
1900 * free space. if sufficient free space is found, dbFindLeaf()
1901 * returns the index of the leaf at which free space was found.
1902 */
1906 /* determine the block number within the file system corresponding
1907 * to the leaf at which free space was found.
1908 */
1911 /* if not all bits of the dmap word are free, get the starting
1912 * bit number within the dmap word of the required string of free
1913 * bits and adjust the block number with this value.
1914 */
1918 /* allocate the blocks */
1923 }
1926 /*
1927 * NAME: dbAllocDmap()
1928 *
1929 * FUNCTION: adjust the disk allocation map to reflect the allocation
1930 * of a specified block range within a dmap.
1931 *
1932 * this routine allocates the specified blocks from the dmap
1933 * through a call to dbAllocBits(). if the allocation of the
1934 * block range causes the maximum string of free blocks within
1935 * the dmap to change (i.e. the value of the root of the dmap's
1936 * dmtree), this routine will cause this change to be reflected
1937 * up through the appropriate levels of the dmap control pages
1938 * by a call to dbAdjCtl() for the L0 dmap control page that
1939 * covers this dmap.
1940 *
1941 * PARAMETERS:
1942 * bmp - pointer to bmap descriptor
1943 * dp - pointer to dmap to allocate the block range from.
1944 * blkno - starting block number of the block to be allocated.
1945 * nblocks - number of blocks to be allocated.
1946 *
1947 * RETURN VALUES:
1948 * 0 - success
1949 * -EIO - i/o error
1950 *
1951 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
1952 */
1955 {
1956 s8 oldroot;
1959 /* save the current value of the root (i.e. maximum free string)
1960 * of the dmap tree.
1961 */
1964 /* allocate the specified (blocks) bits */
1967 /* if the root has not changed, done. */
1971 /* root changed. bubble the change up to the dmap control pages.
1972 * if the adjustment of the upper level control pages fails,
1973 * backout the bit allocation (thus making everything consistent).
1974 */
1979 }
1982 /*
1983 * NAME: dbFreeDmap()
1984 *
1985 * FUNCTION: adjust the disk allocation map to reflect the allocation
1986 * of a specified block range within a dmap.
1987 *
1988 * this routine frees the specified blocks from the dmap through
1989 * a call to dbFreeBits(). if the deallocation of the block range
1990 * causes the maximum string of free blocks within the dmap to
1991 * change (i.e. the value of the root of the dmap's dmtree), this
1992 * routine will cause this change to be reflected up through the
1993 * appropriate levels of the dmap control pages by a call to
1994 * dbAdjCtl() for the L0 dmap control page that covers this dmap.
1995 *
1996 * PARAMETERS:
1997 * bmp - pointer to bmap descriptor
1998 * dp - pointer to dmap to free the block range from.
1999 * blkno - starting block number of the block to be freed.
2000 * nblocks - number of blocks to be freed.
2001 *
2002 * RETURN VALUES:
2003 * 0 - success
2004 * -EIO - i/o error
2005 *
2006 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2007 */
2010 {
2011 s8 oldroot;
2014 /* save the current value of the root (i.e. maximum free string)
2015 * of the dmap tree.
2016 */
2019 /* free the specified (blocks) bits */
2022 /* if error or the root has not changed, done. */
2026 /* root changed. bubble the change up to the dmap control pages.
2027 * if the adjustment of the upper level control pages fails,
2028 * backout the deallocation.
2029 */
2033 /* as part of backing out the deallocation, we will have
2034 * to back split the dmap tree if the deallocation caused
2035 * the freed blocks to become part of a larger binary buddy
2036 * system.
2037 */
2042 }
2045 }
2048 /*
2049 * NAME: dbAllocBits()
2050 *
2051 * FUNCTION: allocate a specified block range from a dmap.
2052 *
2053 * this routine updates the dmap to reflect the working
2054 * state allocation of the specified block range. it directly
2055 * updates the bits of the working map and causes the adjustment
2056 * of the binary buddy system described by the dmap's dmtree
2057 * leaves to reflect the bits allocated. it also causes the
2058 * dmap's dmtree, as a whole, to reflect the allocated range.
2059 *
2060 * PARAMETERS:
2061 * bmp - pointer to bmap descriptor
2062 * dp - pointer to dmap to allocate bits from.
2063 * blkno - starting block number of the bits to be allocated.
2064 * nblocks - number of bits to be allocated.
2065 *
2066 * RETURN VALUES: none
2067 *
2068 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2069 */
2072 {
2078 /* pick up a pointer to the leaves of the dmap tree */
2081 /* determine the bit number and word within the dmap of the
2082 * starting block.
2083 */
2087 /* block range better be within the dmap */
2090 /* allocate the bits of the dmap's words corresponding to the block
2091 * range. not all bits of the first and last words may be contained
2092 * within the block range. if this is the case, we'll work against
2093 * those words (i.e. partial first and/or last) on an individual basis
2094 * (a single pass), allocating the bits of interest by hand and
2095 * updating the leaf corresponding to the dmap word. a single pass
2096 * will be used for all dmap words fully contained within the
2097 * specified range. within this pass, the bits of all fully contained
2098 * dmap words will be marked as free in a single shot and the leaves
2099 * will be updated. a single leaf may describe the free space of
2100 * multiple dmap words, so we may update only a subset of the actual
2101 * leaves corresponding to the dmap words of the block range.
2102 */
2104 /* determine the bit number within the word and
2105 * the number of bits within the word.
2106 */
2110 /* check if only part of a word is to be allocated.
2111 */
2113 /* allocate (set to 1) the appropriate bits within
2114 * this dmap word.
2115 */
2119 /* update the leaf for this dmap word. in addition
2120 * to setting the leaf value to the binary buddy max
2121 * of the updated dmap word, dbSplit() will split
2122 * the binary system of the leaves if need be.
2123 */
2129 /* one or more dmap words are fully contained
2130 * within the block range. determine how many
2131 * words and allocate (set to 1) the bits of these
2132 * words.
2133 */
2137 /* determine how many bits.
2138 */
2141 /* now update the appropriate leaves to reflect
2142 * the allocated words.
2143 */
2147 "dbAllocBits: leaf page "
2150 }
2152 /* determine what the leaf value should be
2153 * updated to as the minimum of the l2 number
2154 * of bits being allocated and the l2 number
2155 * of bits currently described by this leaf.
2156 */
2159 /* update the leaf to reflect the allocation.
2160 * in addition to setting the leaf value to
2161 * NOFREE, dbSplit() will split the binary
2162 * system of the leaves to reflect the current
2163 * allocation (size).
2164 */
2167 /* get the number of dmap words handled */
2170 }
2171 }
2172 }
2174 /* update the free count for this dmap */
2179 /* if this allocation group is completely free,
2180 * update the maximum allocation group number if this allocation
2181 * group is the new max.
2182 */
2187 /* update the free count for the allocation group and map */
2192 }
2195 /*
2196 * NAME: dbFreeBits()
2197 *
2198 * FUNCTION: free a specified block range from a dmap.
2199 *
2200 * this routine updates the dmap to reflect the working
2201 * state allocation of the specified block range. it directly
2202 * updates the bits of the working map and causes the adjustment
2203 * of the binary buddy system described by the dmap's dmtree
2204 * leaves to reflect the bits freed. it also causes the dmap's
2205 * dmtree, as a whole, to reflect the deallocated range.
2206 *
2207 * PARAMETERS:
2208 * bmp - pointer to bmap descriptor
2209 * dp - pointer to dmap to free bits from.
2210 * blkno - starting block number of the bits to be freed.
2211 * nblocks - number of bits to be freed.
2212 *
2213 * RETURN VALUES: 0 for success
2214 *
2215 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2216 */
2219 {
2225 /* determine the bit number and word within the dmap of the
2226 * starting block.
2227 */
2231 /* block range better be within the dmap.
2232 */
2235 /* free the bits of the dmaps words corresponding to the block range.
2236 * not all bits of the first and last words may be contained within
2237 * the block range. if this is the case, we'll work against those
2238 * words (i.e. partial first and/or last) on an individual basis
2239 * (a single pass), freeing the bits of interest by hand and updating
2240 * the leaf corresponding to the dmap word. a single pass will be used
2241 * for all dmap words fully contained within the specified range.
2242 * within this pass, the bits of all fully contained dmap words will
2243 * be marked as free in a single shot and the leaves will be updated. a
2244 * single leaf may describe the free space of multiple dmap words,
2245 * so we may update only a subset of the actual leaves corresponding
2246 * to the dmap words of the block range.
2247 *
2248 * dbJoin() is used to update leaf values and will join the binary
2249 * buddy system of the leaves if the new leaf values indicate this
2250 * should be done.
2251 */
2253 /* determine the bit number within the word and
2254 * the number of bits within the word.
2255 */
2259 /* check if only part of a word is to be freed.
2260 */
2262 /* free (zero) the appropriate bits within this
2263 * dmap word.
2264 */
2269 /* update the leaf for this dmap word.
2270 */
2278 /* one or more dmap words are fully contained
2279 * within the block range. determine how many
2280 * words and free (zero) the bits of these words.
2281 */
2285 /* determine how many bits.
2286 */
2289 /* now update the appropriate leaves to reflect
2290 * the freed words.
2291 */
2293 /* determine what the leaf value should be
2294 * updated to as the minimum of the l2 number
2295 * of bits being freed and the l2 (max) number
2296 * of bits that can be described by this leaf.
2297 */
2298 size =
2303 /* update the leaf.
2304 */
2309 /* get the number of dmap words handled.
2310 */
2313 }
2314 }
2315 }
2317 /* update the free count for this dmap.
2318 */
2323 /* update the free count for the allocation group and
2324 * map.
2325 */
2330 /* check if this allocation group is not completely free and
2331 * if it is currently the maximum (rightmost) allocation group.
2332 * if so, establish the new maximum allocation group number by
2333 * searching left for the first allocation group with allocation.
2334 */
2343 }
2345 /* re-establish the allocation group preference if the
2346 * current preference is right of the maximum allocation
2347 * group.
2348 */
2351 }
2356 }
2359 /*
2360 * NAME: dbAdjCtl()
2361 *
2362 * FUNCTION: adjust a dmap control page at a specified level to reflect
2363 * the change in a lower level dmap or dmap control page's
2364 * maximum string of free blocks (i.e. a change in the root
2365 * of the lower level object's dmtree) due to the allocation
2366 * or deallocation of a range of blocks with a single dmap.
2367 *
2368 * on entry, this routine is provided with the new value of
2369 * the lower level dmap or dmap control page root and the
2370 * starting block number of the block range whose allocation
2371 * or deallocation resulted in the root change. this range
2372 * is respresented by a single leaf of the current dmapctl
2373 * and the leaf will be updated with this value, possibly
2374 * causing a binary buddy system within the leaves to be
2375 * split or joined. the update may also cause the dmapctl's
2376 * dmtree to be updated.
2377 *
2378 * if the adjustment of the dmap control page, itself, causes its
2379 * root to change, this change will be bubbled up to the next dmap
2380 * control level by a recursive call to this routine, specifying
2381 * the new root value and the next dmap control page level to
2382 * be adjusted.
2383 * PARAMETERS:
2384 * bmp - pointer to bmap descriptor
2385 * blkno - the first block of a block range within a dmap. it is
2386 * the allocation or deallocation of this block range that
2387 * requires the dmap control page to be adjusted.
2388 * newval - the new value of the lower level dmap or dmap control
2389 * page root.
2390 * alloc - 'true' if adjustment is due to an allocation.
2391 * level - current level of dmap control page (i.e. L0, L1, L2) to
2392 * be adjusted.
2393 *
2394 * RETURN VALUES:
2395 * 0 - success
2396 * -EIO - i/o error
2397 *
2398 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2399 */
2400 static int
2402 {
2404 s8 oldroot;
2406 s64 lblkno;
2410 /* get the buffer for the dmap control page for the specified
2411 * block number and control page level.
2412 */
2424 }
2426 /* determine the leaf number corresponding to the block and
2427 * the index within the dmap control tree.
2428 */
2432 /* save the current leaf value and the current root level (i.e.
2433 * maximum l2 free string described by this dmapctl).
2434 */
2438 /* check if this is a control page update for an allocation.
2439 * if so, update the leaf to reflect the new leaf value using
2440 * dbSplit(); otherwise (deallocation), use dbJoin() to udpate
2441 * the leaf with the new value. in addition to updating the
2442 * leaf, dbSplit() will also split the binary buddy system of
2443 * the leaves, if required, and bubble new values within the
2444 * dmapctl tree, if required. similarly, dbJoin() will join
2445 * the binary buddy system of leaves and bubble new values up
2446 * the dmapctl tree as required by the new leaf value.
2447 */
2449 /* check if we are in the middle of a binary buddy
2450 * system. this happens when we are performing the
2451 * first allocation out of an allocation group that
2452 * is part (not the first part) of a larger binary
2453 * buddy system. if we are in the middle, back split
2454 * the system prior to calling dbSplit() which assumes
2455 * that it is at the front of a binary buddy system.
2456 */
2462 }
2468 }
2470 /* check if the root of the current dmap control page changed due
2471 * to the update and if the current dmap control page is not at
2472 * the current top level (i.e. L0, L1, L2) of the map. if so (i.e.
2473 * root changed and this is not the top level), call this routine
2474 * again (recursion) for the next higher level of the mapping to
2475 * reflect the change in root for the current dmap control page.
2476 */
2478 /* are we below the top level of the map. if so,
2479 * bubble the root up to the next higher level.
2480 */
2482 /* bubble up the new root of this dmap control page to
2483 * the next level.
2484 */
2488 /* something went wrong in bubbling up the new
2489 * root value, so backout the changes to the
2490 * current dmap control page.
2491 */
2494 oldval);
2496 /* the dbJoin() above might have
2497 * caused a larger binary buddy system
2498 * to form and we may now be in the
2499 * middle of it. if this is the case,
2500 * back split the buddies.
2501 */
2507 }
2509 /* release the buffer and return the error.
2510 */
2513 }
2515 /* we're at the top level of the map. update
2516 * the bmap control page to reflect the size
2517 * of the maximum free buddy system.
2518 */
2522 "dbAdjCtl: the maximum free buddy is "
2524 }
2526 }
2527 }
2529 /* write the buffer.
2530 */
2534 }
2537 /*
2538 * NAME: dbSplit()
2539 *
2540 * FUNCTION: update the leaf of a dmtree with a new value, splitting
2541 * the leaf from the binary buddy system of the dmtree's
2542 * leaves, as required.
2543 *
2544 * PARAMETERS:
2545 * tp - pointer to the tree containing the leaf.
2546 * leafno - the number of the leaf to be updated.
2547 * splitsz - the size the binary buddy system starting at the leaf
2548 * must be split to, specified as the log2 number of blocks.
2549 * newval - the new value for the leaf.
2550 *
2551 * RETURN VALUES: none
2552 *
2553 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2554 */
2556 {
2561 /* check if the leaf needs to be split.
2562 */
2564 /* the split occurs by cutting the buddy system in half
2565 * at the specified leaf until we reach the specified
2566 * size. pick up the starting split size (current size
2567 * - 1 in l2) and the corresponding buddy size.
2568 */
2572 /* split until we reach the specified size.
2573 */
2575 /* update the buddy's leaf with its new value.
2576 */
2579 /* on to the next size and buddy.
2580 */
2583 }
2584 }
2586 /* adjust the dmap tree to reflect the specified leaf's new
2587 * value.
2588 */
2590 }
2593 /*
2594 * NAME: dbBackSplit()
2595 *
2596 * FUNCTION: back split the binary buddy system of dmtree leaves
2597 * that hold a specified leaf until the specified leaf
2598 * starts its own binary buddy system.
2599 *
2600 * the allocators typically perform allocations at the start
2601 * of binary buddy systems and dbSplit() is used to accomplish
2602 * any required splits. in some cases, however, allocation
2603 * may occur in the middle of a binary system and requires a
2604 * back split, with the split proceeding out from the middle of
2605 * the system (less efficient) rather than the start of the
2606 * system (more efficient). the cases in which a back split
2607 * is required are rare and are limited to the first allocation
2608 * within an allocation group which is a part (not first part)
2609 * of a larger binary buddy system and a few exception cases
2610 * in which a previous join operation must be backed out.
2611 *
2612 * PARAMETERS:
2613 * tp - pointer to the tree containing the leaf.
2614 * leafno - the number of the leaf to be updated.
2615 *
2616 * RETURN VALUES: none
2617 *
2618 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2619 */
2621 {
2626 /* leaf should be part (not first part) of a binary
2627 * buddy system.
2628 */
2631 /* the back split is accomplished by iteratively finding the leaf
2632 * that starts the buddy system that contains the specified leaf and
2633 * splitting that system in two. this iteration continues until
2634 * the specified leaf becomes the start of a buddy system.
2635 *
2636 * determine maximum possible l2 size for the specified leaf.
2637 */
2638 size =
2642 /* determine the number of leaves covered by this size. this
2643 * is the buddy size that we will start with as we search for
2644 * the buddy system that contains the specified leaf.
2645 */
2648 /* back split.
2649 */
2651 /* find the leftmost buddy leaf.
2652 */
2658 }
2660 /* determine the buddy.
2661 */
2664 /* check if this buddy is the start of the system.
2665 */
2667 /* split the leaf at the start of the
2668 * system in two.
2669 */
2673 }
2674 }
2675 }
2680 }
2682 }
2685 /*
2686 * NAME: dbJoin()
2687 *
2688 * FUNCTION: update the leaf of a dmtree with a new value, joining
2689 * the leaf with other leaves of the dmtree into a multi-leaf
2690 * binary buddy system, as required.
2691 *
2692 * PARAMETERS:
2693 * tp - pointer to the tree containing the leaf.
2694 * leafno - the number of the leaf to be updated.
2695 * newval - the new value for the leaf.
2696 *
2697 * RETURN VALUES: none
2698 */
2700 {
2704 /* can the new leaf value require a join with other leaves ?
2705 */
2707 /* pickup a pointer to the leaves of the tree.
2708 */
2711 /* try to join the specified leaf into a large binary
2712 * buddy system. the join proceeds by attempting to join
2713 * the specified leafno with its buddy (leaf) at new value.
2714 * if the join occurs, we attempt to join the left leaf
2715 * of the joined buddies with its buddy at new value + 1.
2716 * we continue to join until we find a buddy that cannot be
2717 * joined (does not have a value equal to the size of the
2718 * last join) or until all leaves have been joined into a
2719 * single system.
2720 *
2721 * get the buddy size (number of words covered) of
2722 * the new value.
2723 */
2726 /* try to join.
2727 */
2729 /* get the buddy leaf.
2730 */
2733 /* if the leaf's new value is greater than its
2734 * buddy's value, we join no more.
2735 */
2739 /* It shouldn't be less */
2743 /* check which (leafno or buddy) is the left buddy.
2744 * the left buddy gets to claim the blocks resulting
2745 * from the join while the right gets to claim none.
2746 * the left buddy is also eligable to participate in
2747 * a join at the next higher level while the right
2748 * is not.
2749 *
2750 */
2752 /* leafno is the left buddy.
2753 */
2756 /* buddy is the left buddy and becomes
2757 * leafno.
2758 */
2761 }
2763 /* on to try the next join.
2764 */
2767 }
2768 }
2770 /* update the leaf value.
2771 */
2775 }
2778 /*
2779 * NAME: dbAdjTree()
2780 *
2781 * FUNCTION: update a leaf of a dmtree with a new value, adjusting
2782 * the dmtree, as required, to reflect the new leaf value.
2783 * the combination of any buddies must already be done before
2784 * this is called.
2785 *
2786 * PARAMETERS:
2787 * tp - pointer to the tree to be adjusted.
2788 * leafno - the number of the leaf to be updated.
2789 * newval - the new value for the leaf.
2790 *
2791 * RETURN VALUES: none
2792 */
2794 {
2798 /* pick up the index of the leaf for this leafno.
2799 */
2802 /* is the current value the same as the old value ? if so,
2803 * there is nothing to do.
2804 */
2808 /* set the new value.
2809 */
2812 /* bubble the new value up the tree as required.
2813 */
2815 /* get the index of the first leaf of the 4 leaf
2816 * group containing the specified leaf (leafno).
2817 */
2820 /* get the index of the parent of this 4 leaf group.
2821 */
2824 /* determine the maximum of the 4 leaves.
2825 */
2828 /* if the maximum of the 4 is the same as the
2829 * parent's value, we're done.
2830 */
2834 /* parent gets new value.
2835 */
2838 /* parent becomes leaf for next go-round.
2839 */
2841 }
2842 }
2845 /*
2846 * NAME: dbFindLeaf()
2847 *
2848 * FUNCTION: search a dmtree_t for sufficient free blocks, returning
2849 * the index of a leaf describing the free blocks if
2850 * sufficient free blocks are found.
2851 *
2852 * the search starts at the top of the dmtree_t tree and
2853 * proceeds down the tree to the leftmost leaf with sufficient
2854 * free space.
2855 *
2856 * PARAMETERS:
2857 * tp - pointer to the tree to be searched.
2858 * l2nb - log2 number of free blocks to search for.
2859 * leafidx - return pointer to be set to the index of the leaf
2860 * describing at least l2nb free blocks if sufficient
2861 * free blocks are found.
2862 *
2863 * RETURN VALUES:
2864 * 0 - success
2865 * -ENOSPC - insufficient free blocks.
2866 */
2868 {
2871 /* first check the root of the tree to see if there is
2872 * sufficient free space.
2873 */
2877 /* sufficient free space available. now search down the tree
2878 * starting at the next level for the leftmost leaf that
2879 * describes sufficient free space.
2880 */
2883 /* search the four nodes at this level, starting from
2884 * the left.
2885 */
2887 /* sufficient free space found. move to the next
2888 * level (or quit if this is the last level).
2889 */
2892 }
2894 /* better have found something since the higher
2895 * levels of the tree said it was here.
2896 */
2898 }
2900 /* set the return to the leftmost leaf describing sufficient
2901 * free space.
2902 */
2906 }
2909 /*
2910 * NAME: dbFindBits()
2911 *
2912 * FUNCTION: find a specified number of binary buddy free bits within a
2913 * dmap bitmap word value.
2914 *
2915 * this routine searches the bitmap value for (1 << l2nb) free
2916 * bits at (1 << l2nb) alignments within the value.
2917 *
2918 * PARAMETERS:
2919 * word - dmap bitmap word value.
2920 * l2nb - number of free bits specified as a log2 number.
2921 *
2922 * RETURN VALUES:
2923 * starting bit number of free bits.
2924 */
2926 {
2928 u32 mask;
2930 /* get the number of bits.
2931 */
2935 /* complement the word so we can use a mask (i.e. 0s represent
2936 * free bits) and compute the mask.
2937 */
2941 /* scan the word for nb free bits at nb alignments.
2942 */
2946 }
2950 /* return the bit number.
2951 */
2953 }
2956 /*
2957 * NAME: dbMaxBud(u8 *cp)
2958 *
2959 * FUNCTION: determine the largest binary buddy string of free
2960 * bits within 32-bits of the map.
2961 *
2962 * PARAMETERS:
2963 * cp - pointer to the 32-bit value.
2964 *
2965 * RETURN VALUES:
2966 * largest binary buddy of free bits within a dmap word.
2967 */
2969 {
2972 /* check if the wmap word is all free. if so, the
2973 * free buddy size is BUDMIN.
2974 */
2978 /* check if the wmap word is half free. if so, the
2979 * free buddy size is BUDMIN-1.
2980 */
2984 /* not all free or half free. determine the free buddy
2985 * size thru table lookup using quarters of the wmap word.
2986 */
2990 }
2993 /*
2994 * NAME: cnttz(uint word)
2995 *
2996 * FUNCTION: determine the number of trailing zeros within a 32-bit
2997 * value.
2998 *
2999 * PARAMETERS:
3000 * value - 32-bit value to be examined.
3001 *
3002 * RETURN VALUES:
3003 * count of trailing zeros
3004 */
3006 {
3012 }
3015 }
3018 /*
3019 * NAME: cntlz(u32 value)
3020 *
3021 * FUNCTION: determine the number of leading zeros within a 32-bit
3022 * value.
3023 *
3024 * PARAMETERS:
3025 * value - 32-bit value to be examined.
3026 *
3027 * RETURN VALUES:
3028 * count of leading zeros
3029 */
3031 {
3037 }
3039 }
3042 /*
3043 * NAME: blkstol2(s64 nb)
3044 *
3045 * FUNCTION: convert a block count to its log2 value. if the block
3046 * count is not a l2 multiple, it is rounded up to the next
3047 * larger l2 multiple.
3048 *
3049 * PARAMETERS:
3050 * nb - number of blocks
3051 *
3052 * RETURN VALUES:
3053 * log2 number of blocks
3054 */
3056 {
3062 /* count the leading bits.
3063 */
3065 /* leading bit found.
3066 */
3068 /* determine the l2 value.
3069 */
3072 /* check if we need to round up.
3073 */
3075 l2nb++;
3078 }
3079 }
3082 }
3085 /*
3086 * NAME: dbAllocBottomUp()
3087 *
3088 * FUNCTION: alloc the specified block range from the working block
3089 * allocation map.
3090 *
3091 * the blocks will be alloc from the working map one dmap
3092 * at a time.
3093 *
3094 * PARAMETERS:
3095 * ip - pointer to in-core inode;
3096 * blkno - starting block number to be freed.
3097 * nblocks - number of blocks to be freed.
3098 *
3099 * RETURN VALUES:
3100 * 0 - success
3101 * -EIO - i/o error
3102 */
3104 {
3114 /* block to be allocated better be within the mapsize. */
3117 /*
3118 * allocate the blocks a dmap at a time.
3119 */
3122 /* release previous dmap if any */
3125 }
3127 /* get the buffer for the current dmap. */
3133 }
3136 /* determine the number of blocks to be allocated from
3137 * this dmap.
3138 */
3141 /* allocate the blocks. */
3146 }
3147 }
3149 /* write the last buffer. */
3155 }
3160 {
3166 /* save the current value of the root (i.e. maximum free string)
3167 * of the dmap tree.
3168 */
3171 /* pick up a pointer to the leaves of the dmap tree */
3174 /* determine the bit number and word within the dmap of the
3175 * starting block.
3176 */
3180 /* block range better be within the dmap */
3183 /* allocate the bits of the dmap's words corresponding to the block
3184 * range. not all bits of the first and last words may be contained
3185 * within the block range. if this is the case, we'll work against
3186 * those words (i.e. partial first and/or last) on an individual basis
3187 * (a single pass), allocating the bits of interest by hand and
3188 * updating the leaf corresponding to the dmap word. a single pass
3189 * will be used for all dmap words fully contained within the
3190 * specified range. within this pass, the bits of all fully contained
3191 * dmap words will be marked as free in a single shot and the leaves
3192 * will be updated. a single leaf may describe the free space of
3193 * multiple dmap words, so we may update only a subset of the actual
3194 * leaves corresponding to the dmap words of the block range.
3195 */
3197 /* determine the bit number within the word and
3198 * the number of bits within the word.
3199 */
3203 /* check if only part of a word is to be allocated.
3204 */
3206 /* allocate (set to 1) the appropriate bits within
3207 * this dmap word.
3208 */
3212 word++;
3214 /* one or more dmap words are fully contained
3215 * within the block range. determine how many
3216 * words and allocate (set to 1) the bits of these
3217 * words.
3218 */
3222 /* determine how many bits */
3225 }
3226 }
3228 /* update the free count for this dmap */
3231 /* reconstruct summary tree */
3236 /* if this allocation group is completely free,
3237 * update the highest active allocation group number
3238 * if this allocation group is the new max.
3239 */
3244 /* update the free count for the allocation group and map */
3250 /* if the root has not changed, done. */
3254 /* root changed. bubble the change up to the dmap control pages.
3255 * if the adjustment of the upper level control pages fails,
3256 * backout the bit allocation (thus making everything consistent).
3257 */
3262 }
3265 /*
3266 * NAME: dbExtendFS()
3267 *
3268 * FUNCTION: extend bmap from blkno for nblocks;
3269 * dbExtendFS() updates bmap ready for dbAllocBottomUp();
3270 *
3271 * L2
3272 * |
3273 * L1---------------------------------L1
3274 * | |
3275 * L0---------L0---------L0 L0---------L0---------L0
3276 * | | | | | |
3277 * d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,.,dm;
3278 * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3279 *
3280 * <---old---><----------------------------extend----------------------->
3281 */
3283 {
3287 s64 newsize;
3288 s64 p;
3295 s64 ag_rem;
3302 /*
3303 * initialize bmap control page.
3304 *
3305 * all the data in bmap control page should exclude
3306 * the mkfs hidden dmap page.
3307 */
3309 /* update mapsize */
3313 /* compute new AG size */
3320 /* compute new number of AG */
3325 /*
3326 * reconfigure db_agfree[]
3327 * from old AG configuration to new AG configuration;
3328 *
3329 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3330 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3331 * note: new AG size = old AG size * (2**x).
3332 */
3340 /* coalesce cotiguous k AGs; */
3342 /* merge AGi to AGn */
3344 }
3345 }
3351 /*
3352 * update highest active ag number
3353 */
3357 /*
3358 * extend bmap
3359 *
3360 * update bit maps and corresponding level control pages;
3361 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3362 */
3363 extend:
3364 /* get L2 page */
3370 }
3373 /* compute start L1 */
3378 /*
3379 * extend each L1 in L2
3380 */
3382 /* get L1 page */
3384 /* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3390 /* compute start L0 */
3396 /* assign/init L1 page */
3403 /* compute start L0 */
3407 }
3409 /*
3410 * extend each L0 in L1
3411 */
3413 /* get L0 page */
3415 /* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3422 /* compute start dmap */
3424 L2BPERDMAP;
3430 /* assign/init L0 page */
3437 /* compute start dmap */
3441 }
3443 /*
3444 * extend each dmap in L0
3445 */
3447 /*
3448 * reconstruct the dmap page, and
3449 * initialize corresponding parent L0 leaf
3450 */
3452 /* read in dmap page: */
3459 /* assign/init dmap page */
3466 }
3477 l0leaf++;
3486 /*
3487 * build current L0 page from its leaves, and
3488 * initialize corresponding parent L1 leaf
3489 */
3497 /* more than 1 L0 ? */
3501 /* summarize in global bmap page */
3506 }
3507 }
3510 /*
3511 * build current L1 page from its leaves, and
3512 * initialize corresponding parent L2 leaf
3513 */
3521 /* more than 1 L1 ? */
3525 /* summarize in global bmap page */
3529 }
3530 }
3535 errout:
3543 /*
3544 * finalize bmap control page
3545 */
3546 finalize:
3549 }
3552 /*
3553 * dbFinalizeBmap()
3554 */
3556 {
3562 /*
3563 * finalize bmap control page
3564 */
3565 //finalize:
3566 /*
3567 * compute db_agpref: preferred ag to allocate from
3568 * (the leftmost ag with average free space in it);
3569 */
3570 //agpref:
3571 /* get the number of active ags and inacitve ags */
3576 /* determine how many blocks are in the inactive allocation
3577 * groups. in doing this, we must account for the fact that
3578 * the rightmost group might be a partial group (i.e. file
3579 * system size is not a multiple of the group size).
3580 */
3585 /* determine how many free blocks are in the active
3586 * allocation groups plus the average number of free blocks
3587 * within the active ags.
3588 */
3592 /* if the preferred allocation group has not average free space.
3593 * re-establish the preferred group as the leftmost
3594 * group with average free space.
3595 */
3601 }
3605 }
3606 }
3608 /*
3609 * compute db_aglevel, db_agheigth, db_width, db_agstart:
3610 * an ag is covered in aglevel dmapctl summary tree,
3611 * at agheight level height (from leaf) with agwidth number of nodes
3612 * each, which starts at agstart index node of the smmary tree node
3613 * array;
3614 */
3616 l2nl =
3621 i--) {
3624 }
3626 }
3629 /*
3630 * NAME: dbInitDmap()/ujfs_idmap_page()
3631 *
3632 * FUNCTION: initialize working/persistent bitmap of the dmap page
3633 * for the specified number of blocks:
3634 *
3635 * at entry, the bitmaps had been initialized as free (ZEROS);
3636 * The number of blocks will only account for the actually
3637 * existing blocks. Blocks which don't actually exist in
3638 * the aggregate will be marked as allocated (ONES);
3639 *
3640 * PARAMETERS:
3641 * dp - pointer to page of map
3642 * nblocks - number of blocks this page
3643 *
3644 * RETURNS: NONE
3645 */
3647 {
3650 /* starting block number within the dmap */
3661 }
3666 }
3668 /* word number containing start block number */
3671 /*
3672 * free the bits corresponding to the block range (ZEROS):
3673 * note: not all bits of the first and last words may be contained
3674 * within the block range.
3675 */
3677 /* number of bits preceding range to be freed in the word */
3679 /* number of bits to free in the word */
3682 /* is partial word to be freed ? */
3684 /* free (set to 0) from the bitmap word */
3690 /* skip the word freed */
3691 w++;
3693 /* free (set to 0) contiguous bitmap words */
3698 /* skip the words freed */
3701 }
3702 }
3704 /*
3705 * mark bits following the range to be freed (non-existing
3706 * blocks) as allocated (ONES)
3707 */
3712 /* the first word beyond the end of existing blocks */
3715 /* does nblocks fall on a 32-bit boundary ? */
3718 /* mark a partial word allocated */
3720 w++;
3721 }
3723 /* set the rest of the words in the page to allocated (ONES) */
3727 /*
3728 * init tree
3729 */
3730 initTree:
3732 }
3735 /*
3736 * NAME: dbInitDmapTree()/ujfs_complete_dmap()
3737 *
3738 * FUNCTION: initialize summary tree of the specified dmap:
3739 *
3740 * at entry, bitmap of the dmap has been initialized;
3741 *
3742 * PARAMETERS:
3743 * dp - dmap to complete
3744 * blkno - starting block number for this dmap
3745 * treemax - will be filled in with max free for this dmap
3746 *
3747 * RETURNS: max free string at the root of the tree
3748 */
3750 {
3755 /* init fixed info of tree */
3763 /* init each leaf from corresponding wmap word:
3764 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3765 * bitmap word are allocated.
3766 */
3771 /* build the dmap's binary buddy summary tree */
3773 }
3776 /*
3777 * NAME: dbInitTree()/ujfs_adjtree()
3778 *
3779 * FUNCTION: initialize binary buddy summary tree of a dmap or dmapctl.
3780 *
3781 * at entry, the leaves of the tree has been initialized
3782 * from corresponding bitmap word or root of summary tree
3783 * of the child control page;
3784 * configure binary buddy system at the leaf level, then
3785 * bubble up the values of the leaf nodes up the tree.
3786 *
3787 * PARAMETERS:
3788 * cp - Pointer to the root of the tree
3789 * l2leaves- Number of leaf nodes as a power of 2
3790 * l2min - Number of blocks that can be covered by a leaf
3791 * as a power of 2
3792 *
3793 * RETURNS: max free string at the root of the tree
3794 */
3796 {
3803 /* Determine the maximum free string possible for the leaves */
3806 /*
3807 * configure the leaf levevl into binary buddy system
3808 *
3809 * Try to combine buddies starting with a buddy size of 1
3810 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3811 * can be combined if both buddies have a maximum free of l2min;
3812 * the combination will result in the left-most buddy leaf having
3813 * a maximum free of l2min+1.
3814 * After processing all buddies for a given size, process buddies
3815 * at the next higher buddy size (i.e. current size * 2) and
3816 * the next maximum free (current free + 1).
3817 * This continues until the maximum possible buddy combination
3818 * yields maximum free.
3819 */
3822 /* get next buddy size == current buddy pair size */
3825 /* scan each adjacent buddy pair at current buddy size */
3829 /* coalesce if both adjacent buddies are max free */
3833 }
3834 }
3835 }
3837 /*
3838 * bubble summary information of leaves up the tree.
3839 *
3840 * Starting at the leaf node level, the four nodes described by
3841 * the higher level parent node are compared for a maximum free and
3842 * this maximum becomes the value of the parent node.
3843 * when all lower level nodes are processed in this fashion then
3844 * move up to the next level (parent becomes a lower level node) and
3845 * continue the process for that level.
3846 */
3850 /* get index of 1st node of parent level */
3853 /* set the value of the parent node as the maximum
3854 * of the four nodes of the current level.
3855 */
3859 }
3862 }
3865 /*
3866 * dbInitDmapCtl()
3867 *
3868 * function: initialize dmapctl page
3869 */
3880 /*
3881 * initialize the leaves of current level that were not covered
3882 * by the specified input block range (i.e. the leaves have no
3883 * low level dmapctl or dmap).
3884 */
3889 /* build the dmap's binary buddy summary tree */
3891 }
3894 /*
3895 * NAME: dbGetL2AGSize()/ujfs_getagl2size()
3896 *
3897 * FUNCTION: Determine log2(allocation group size) from aggregate size
3898 *
3899 * PARAMETERS:
3900 * nblocks - Number of blocks in aggregate
3901 *
3902 * RETURNS: log2(allocation group size) in aggregate blocks
3903 */
3905 {
3906 s64 sz;
3907 s64 m;
3913 /* round up aggregate size to power of 2 */
3918 }
3924 /* agsize = roundupSize/max_number_of_ag */
3926 }
3929 /*
3930 * NAME: dbMapFileSizeToMapSize()
3931 *
3932 * FUNCTION: compute number of blocks the block allocation map file
3933 * can cover from the map file size;
3934 *
3935 * RETURNS: Number of blocks which can be covered by this block map file;
3936 */
3938 /*
3939 * maximum number of map pages at each level including control pages
3940 */
3941 #define MAXL0PAGES (1 + LPERCTL)
3942 #define MAXL1PAGES (1 + LPERCTL * MAXL0PAGES)
3943 #define MAXL2PAGES (1 + LPERCTL * MAXL1PAGES)
3945 /*
3946 * convert number of map pages to the zero origin top dmapctl level
3947 */
3948 #define BMAPPGTOLEV(npages) \
3949 (((npages) <= 3 + MAXL0PAGES) ? 0 : \
3950 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
3953 {
3955 s64 nblocks;
3964 /* At each level, accumulate the number of dmap pages covered by
3965 * the number of full child levels below it;
3966 * repeat for the last incomplete child level.
3967 */
3970 /* skip higher level control pages above top level covered by map */
3974 factor =
3980 /* pages in last/incomplete child */
3982 /* skip incomplete child's level control page */
3983 npages--;
3984 }
3986 /* convert the number of dmaps into the number of blocks
3987 * which can be covered by the dmaps;
3988 */
3992 }