1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright (C) International Business Machines Corp., 2000-2004
4 * Portions Copyright (C) Tino Reichardt, 2012
8 #include <linux/slab.h>
9 #include "jfs_incore.h"
10 #include "jfs_superblock.h"
14 #include "jfs_metapage.h"
15 #include "jfs_debug.h"
16 #include "jfs_discard.h"
19 * SERIALIZATION of the Block Allocation Map.
21 * the working state of the block allocation map is accessed in
24 * 1) allocation and free requests that start at the dmap
25 * level and move up through the dmap control pages (i.e.
26 * the vast majority of requests).
28 * 2) allocation requests that start at dmap control page
29 * level and work down towards the dmaps.
31 * the serialization scheme used here is as follows.
33 * requests which start at the bottom are serialized against each
34 * other through buffers and each requests holds onto its buffers
35 * as it works it way up from a single dmap to the required level
36 * of dmap control page.
37 * requests that start at the top are serialized against each other
38 * and request that start from the bottom by the multiple read/single
39 * write inode lock of the bmap inode. requests starting at the top
40 * take this lock in write mode while request starting at the bottom
41 * take the lock in read mode. a single top-down request may proceed
42 * exclusively while multiple bottoms-up requests may proceed
43 * simultaneously (under the protection of busy buffers).
45 * in addition to information found in dmaps and dmap control pages,
46 * the working state of the block allocation map also includes read/
47 * write information maintained in the bmap descriptor (i.e. total
48 * free block count, allocation group level free block counts).
49 * a single exclusive lock (BMAP_LOCK) is used to guard this information
50 * in the face of multiple-bottoms up requests.
51 * (lock ordering: IREAD_LOCK, BMAP_LOCK);
53 * accesses to the persistent state of the block allocation map (limited
54 * to the persistent bitmaps in dmaps) is guarded by (busy) buffers.
57 #define BMAP_LOCK_INIT(bmp) mutex_init(&bmp->db_bmaplock)
58 #define BMAP_LOCK(bmp) mutex_lock(&bmp->db_bmaplock)
59 #define BMAP_UNLOCK(bmp) mutex_unlock(&bmp->db_bmaplock)
64 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
66 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval);
67 static int dbBackSplit(dmtree_t * tp, int leafno);
68 static int dbJoin(dmtree_t * tp, int leafno, int newval);
69 static void dbAdjTree(dmtree_t * tp, int leafno, int newval);
70 static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc,
72 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results);
73 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
75 static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno,
77 int l2nb, s64 * results);
78 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
80 static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks,
83 static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb,
85 static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno,
87 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks);
88 static int dbFindBits(u32 word, int l2nb);
89 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno);
90 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx);
91 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
93 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
95 static int dbMaxBud(u8 * cp);
96 static int blkstol2(s64 nb);
98 static int cntlz(u32 value);
99 static int cnttz(u32 word);
101 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
103 static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks);
104 static int dbInitDmapTree(struct dmap * dp);
105 static int dbInitTree(struct dmaptree * dtp);
106 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i);
107 static int dbGetL2AGSize(s64 nblocks);
112 * table used for determining buddy sizes within characters of
113 * dmap bitmap words. the characters themselves serve as indexes
114 * into the table, with the table elements yielding the maximum
115 * binary buddy of free bits within the character.
117 static const s8 budtab[256] = {
118 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
119 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
120 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
121 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
122 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
123 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
124 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
125 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
126 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
127 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
128 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
129 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
130 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
131 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
132 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0,
133 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1
139 * FUNCTION: initializate the block allocation map.
141 * memory is allocated for the in-core bmap descriptor and
142 * the in-core descriptor is initialized from disk.
145 * ipbmap - pointer to in-core inode for the block map.
149 * -ENOMEM - insufficient memory
151 * -EINVAL - wrong bmap data
153 int dbMount(struct inode *ipbmap)
156 struct dbmap_disk *dbmp_le;
161 * allocate/initialize the in-memory bmap descriptor
163 /* allocate memory for the in-memory bmap descriptor */
164 bmp = kmalloc(sizeof(struct bmap), GFP_KERNEL);
168 /* read the on-disk bmap descriptor. */
169 mp = read_metapage(ipbmap,
170 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
177 /* copy the on-disk bmap descriptor to its in-memory version. */
178 dbmp_le = (struct dbmap_disk *) mp->data;
179 bmp->db_mapsize = le64_to_cpu(dbmp_le->dn_mapsize);
180 bmp->db_nfree = le64_to_cpu(dbmp_le->dn_nfree);
182 bmp->db_l2nbperpage = le32_to_cpu(dbmp_le->dn_l2nbperpage);
183 if (bmp->db_l2nbperpage > L2PSIZE - L2MINBLOCKSIZE) {
185 goto err_release_metapage;
188 bmp->db_numag = le32_to_cpu(dbmp_le->dn_numag);
189 if (!bmp->db_numag) {
191 goto err_release_metapage;
194 bmp->db_maxlevel = le32_to_cpu(dbmp_le->dn_maxlevel);
195 bmp->db_maxag = le32_to_cpu(dbmp_le->dn_maxag);
196 bmp->db_agpref = le32_to_cpu(dbmp_le->dn_agpref);
197 bmp->db_aglevel = le32_to_cpu(dbmp_le->dn_aglevel);
198 bmp->db_agheight = le32_to_cpu(dbmp_le->dn_agheight);
199 bmp->db_agwidth = le32_to_cpu(dbmp_le->dn_agwidth);
200 bmp->db_agstart = le32_to_cpu(dbmp_le->dn_agstart);
201 bmp->db_agl2size = le32_to_cpu(dbmp_le->dn_agl2size);
202 if (bmp->db_agl2size > L2MAXL2SIZE - L2MAXAG ||
203 bmp->db_agl2size < 0) {
205 goto err_release_metapage;
208 if (((bmp->db_mapsize - 1) >> bmp->db_agl2size) > MAXAG) {
210 goto err_release_metapage;
213 for (i = 0; i < MAXAG; i++)
214 bmp->db_agfree[i] = le64_to_cpu(dbmp_le->dn_agfree[i]);
215 bmp->db_agsize = le64_to_cpu(dbmp_le->dn_agsize);
216 bmp->db_maxfreebud = dbmp_le->dn_maxfreebud;
218 /* release the buffer. */
219 release_metapage(mp);
221 /* bind the bmap inode and the bmap descriptor to each other. */
222 bmp->db_ipbmap = ipbmap;
223 JFS_SBI(ipbmap->i_sb)->bmap = bmp;
225 memset(bmp->db_active, 0, sizeof(bmp->db_active));
228 * allocate/initialize the bmap lock
234 err_release_metapage:
235 release_metapage(mp);
245 * FUNCTION: terminate the block allocation map in preparation for
246 * file system unmount.
248 * the in-core bmap descriptor is written to disk and
249 * the memory for this descriptor is freed.
252 * ipbmap - pointer to in-core inode for the block map.
258 int dbUnmount(struct inode *ipbmap, int mounterror)
260 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
262 if (!(mounterror || isReadOnly(ipbmap)))
266 * Invalidate the page cache buffers
268 truncate_inode_pages(ipbmap->i_mapping, 0);
270 /* free the memory for the in-memory bmap. */
272 JFS_SBI(ipbmap->i_sb)->bmap = NULL;
280 int dbSync(struct inode *ipbmap)
282 struct dbmap_disk *dbmp_le;
283 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
288 * write bmap global control page
290 /* get the buffer for the on-disk bmap descriptor. */
291 mp = read_metapage(ipbmap,
292 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage,
295 jfs_err("dbSync: read_metapage failed!");
298 /* copy the in-memory version of the bmap to the on-disk version */
299 dbmp_le = (struct dbmap_disk *) mp->data;
300 dbmp_le->dn_mapsize = cpu_to_le64(bmp->db_mapsize);
301 dbmp_le->dn_nfree = cpu_to_le64(bmp->db_nfree);
302 dbmp_le->dn_l2nbperpage = cpu_to_le32(bmp->db_l2nbperpage);
303 dbmp_le->dn_numag = cpu_to_le32(bmp->db_numag);
304 dbmp_le->dn_maxlevel = cpu_to_le32(bmp->db_maxlevel);
305 dbmp_le->dn_maxag = cpu_to_le32(bmp->db_maxag);
306 dbmp_le->dn_agpref = cpu_to_le32(bmp->db_agpref);
307 dbmp_le->dn_aglevel = cpu_to_le32(bmp->db_aglevel);
308 dbmp_le->dn_agheight = cpu_to_le32(bmp->db_agheight);
309 dbmp_le->dn_agwidth = cpu_to_le32(bmp->db_agwidth);
310 dbmp_le->dn_agstart = cpu_to_le32(bmp->db_agstart);
311 dbmp_le->dn_agl2size = cpu_to_le32(bmp->db_agl2size);
312 for (i = 0; i < MAXAG; i++)
313 dbmp_le->dn_agfree[i] = cpu_to_le64(bmp->db_agfree[i]);
314 dbmp_le->dn_agsize = cpu_to_le64(bmp->db_agsize);
315 dbmp_le->dn_maxfreebud = bmp->db_maxfreebud;
317 /* write the buffer */
321 * write out dirty pages of bmap
323 filemap_write_and_wait(ipbmap->i_mapping);
325 diWriteSpecial(ipbmap, 0);
333 * FUNCTION: free the specified block range from the working block
336 * the blocks will be free from the working map one dmap
340 * ip - pointer to in-core inode;
341 * blkno - starting block number to be freed.
342 * nblocks - number of blocks to be freed.
348 int dbFree(struct inode *ip, s64 blkno, s64 nblocks)
354 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
355 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
356 struct super_block *sb = ipbmap->i_sb;
358 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
360 /* block to be freed better be within the mapsize. */
361 if (unlikely((blkno == 0) || (blkno + nblocks > bmp->db_mapsize))) {
362 IREAD_UNLOCK(ipbmap);
363 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
364 (unsigned long long) blkno,
365 (unsigned long long) nblocks);
366 jfs_error(ip->i_sb, "block to be freed is outside the map\n");
371 * TRIM the blocks, when mounted with discard option
373 if (JFS_SBI(sb)->flag & JFS_DISCARD)
374 if (JFS_SBI(sb)->minblks_trim <= nblocks)
375 jfs_issue_discard(ipbmap, blkno, nblocks);
378 * free the blocks a dmap at a time.
381 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
382 /* release previous dmap if any */
387 /* get the buffer for the current dmap. */
388 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
389 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
391 IREAD_UNLOCK(ipbmap);
394 dp = (struct dmap *) mp->data;
396 /* determine the number of blocks to be freed from
399 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
401 /* free the blocks. */
402 if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) {
403 jfs_error(ip->i_sb, "error in block map\n");
404 release_metapage(mp);
405 IREAD_UNLOCK(ipbmap);
410 /* write the last buffer. */
414 IREAD_UNLOCK(ipbmap);
421 * NAME: dbUpdatePMap()
423 * FUNCTION: update the allocation state (free or allocate) of the
424 * specified block range in the persistent block allocation map.
426 * the blocks will be updated in the persistent map one
430 * ipbmap - pointer to in-core inode for the block map.
431 * free - 'true' if block range is to be freed from the persistent
432 * map; 'false' if it is to be allocated.
433 * blkno - starting block number of the range.
434 * nblocks - number of contiguous blocks in the range.
435 * tblk - transaction block;
442 dbUpdatePMap(struct inode *ipbmap,
443 int free, s64 blkno, s64 nblocks, struct tblock * tblk)
445 int nblks, dbitno, wbitno, rbits;
446 int word, nbits, nwords;
447 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
448 s64 lblkno, rem, lastlblkno;
453 int lsn, difft, diffp;
456 /* the blocks better be within the mapsize. */
457 if (blkno + nblocks > bmp->db_mapsize) {
458 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n",
459 (unsigned long long) blkno,
460 (unsigned long long) nblocks);
461 jfs_error(ipbmap->i_sb, "blocks are outside the map\n");
465 /* compute delta of transaction lsn from log syncpt */
467 log = (struct jfs_log *) JFS_SBI(tblk->sb)->log;
468 logdiff(difft, lsn, log);
471 * update the block state a dmap at a time.
475 for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) {
476 /* get the buffer for the current dmap. */
477 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
478 if (lblkno != lastlblkno) {
483 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE,
487 metapage_wait_for_io(mp);
489 dp = (struct dmap *) mp->data;
491 /* determine the bit number and word within the dmap of
492 * the starting block. also determine how many blocks
493 * are to be updated within this dmap.
495 dbitno = blkno & (BPERDMAP - 1);
496 word = dbitno >> L2DBWORD;
497 nblks = min(rem, (s64)BPERDMAP - dbitno);
499 /* update the bits of the dmap words. the first and last
500 * words may only have a subset of their bits updated. if
501 * this is the case, we'll work against that word (i.e.
502 * partial first and/or last) only in a single pass. a
503 * single pass will also be used to update all words that
504 * are to have all their bits updated.
506 for (rbits = nblks; rbits > 0;
507 rbits -= nbits, dbitno += nbits) {
508 /* determine the bit number within the word and
509 * the number of bits within the word.
511 wbitno = dbitno & (DBWORD - 1);
512 nbits = min(rbits, DBWORD - wbitno);
514 /* check if only part of the word is to be updated. */
515 if (nbits < DBWORD) {
516 /* update (free or allocate) the bits
520 (ONES << (DBWORD - nbits) >> wbitno);
530 /* one or more words are to have all
531 * their bits updated. determine how
532 * many words and how many bits.
534 nwords = rbits >> L2DBWORD;
535 nbits = nwords << L2DBWORD;
537 /* update (free or allocate) the bits
541 memset(&dp->pmap[word], 0,
544 memset(&dp->pmap[word], (int) ONES,
554 if (lblkno == lastlblkno)
559 LOGSYNC_LOCK(log, flags);
561 /* inherit older/smaller lsn */
562 logdiff(diffp, mp->lsn, log);
566 /* move bp after tblock in logsync list */
567 list_move(&mp->synclist, &tblk->synclist);
570 /* inherit younger/larger clsn */
571 logdiff(difft, tblk->clsn, log);
572 logdiff(diffp, mp->clsn, log);
574 mp->clsn = tblk->clsn;
579 /* insert bp after tblock in logsync list */
581 list_add(&mp->synclist, &tblk->synclist);
583 mp->clsn = tblk->clsn;
585 LOGSYNC_UNLOCK(log, flags);
588 /* write the last buffer. */
600 * FUNCTION: find the preferred allocation group for new allocations.
602 * Within the allocation groups, we maintain a preferred
603 * allocation group which consists of a group with at least
604 * average free space. It is the preferred group that we target
605 * new inode allocation towards. The tie-in between inode
606 * allocation and block allocation occurs as we allocate the
607 * first (data) block of an inode and specify the inode (block)
608 * as the allocation hint for this block.
610 * We try to avoid having more than one open file growing in
611 * an allocation group, as this will lead to fragmentation.
612 * This differs from the old OS/2 method of trying to keep
613 * empty ags around for large allocations.
616 * ipbmap - pointer to in-core inode for the block map.
619 * the preferred allocation group number.
621 int dbNextAG(struct inode *ipbmap)
628 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
632 /* determine the average number of free blocks within the ags. */
633 avgfree = (u32)bmp->db_nfree / bmp->db_numag;
636 * if the current preferred ag does not have an active allocator
637 * and has at least average freespace, return it
639 agpref = bmp->db_agpref;
640 if ((atomic_read(&bmp->db_active[agpref]) == 0) &&
641 (bmp->db_agfree[agpref] >= avgfree))
644 /* From the last preferred ag, find the next one with at least
645 * average free space.
647 for (i = 0 ; i < bmp->db_numag; i++, agpref++) {
648 if (agpref == bmp->db_numag)
651 if (atomic_read(&bmp->db_active[agpref]))
652 /* open file is currently growing in this ag */
654 if (bmp->db_agfree[agpref] >= avgfree) {
655 /* Return this one */
656 bmp->db_agpref = agpref;
658 } else if (bmp->db_agfree[agpref] > hwm) {
659 /* Less than avg. freespace, but best so far */
660 hwm = bmp->db_agfree[agpref];
666 * If no inactive ag was found with average freespace, use the
670 bmp->db_agpref = next_best;
671 /* else leave db_agpref unchanged */
675 /* return the preferred group.
677 return (bmp->db_agpref);
683 * FUNCTION: attempt to allocate a specified number of contiguous free
684 * blocks from the working allocation block map.
686 * the block allocation policy uses hints and a multi-step
689 * for allocation requests smaller than the number of blocks
690 * per dmap, we first try to allocate the new blocks
691 * immediately following the hint. if these blocks are not
692 * available, we try to allocate blocks near the hint. if
693 * no blocks near the hint are available, we next try to
694 * allocate within the same dmap as contains the hint.
696 * if no blocks are available in the dmap or the allocation
697 * request is larger than the dmap size, we try to allocate
698 * within the same allocation group as contains the hint. if
699 * this does not succeed, we finally try to allocate anywhere
700 * within the aggregate.
702 * we also try to allocate anywhere within the aggregate
703 * for allocation requests larger than the allocation group
704 * size or requests that specify no hint value.
707 * ip - pointer to in-core inode;
708 * hint - allocation hint.
709 * nblocks - number of contiguous blocks in the range.
710 * results - on successful return, set to the starting block number
711 * of the newly allocated contiguous range.
715 * -ENOSPC - insufficient disk resources
718 int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results)
721 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
730 /* assert that nblocks is valid */
733 /* get the log2 number of blocks to be allocated.
734 * if the number of blocks is not a log2 multiple,
735 * it will be rounded up to the next log2 multiple.
737 l2nb = BLKSTOL2(nblocks);
739 bmp = JFS_SBI(ip->i_sb)->bmap;
741 mapSize = bmp->db_mapsize;
743 /* the hint should be within the map */
744 if (hint >= mapSize) {
745 jfs_error(ip->i_sb, "the hint is outside the map\n");
749 /* if the number of blocks to be allocated is greater than the
750 * allocation group size, try to allocate anywhere.
752 if (l2nb > bmp->db_agl2size) {
753 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
755 rc = dbAllocAny(bmp, nblocks, l2nb, results);
761 * If no hint, let dbNextAG recommend an allocation group
766 /* we would like to allocate close to the hint. adjust the
767 * hint to the block following the hint since the allocators
768 * will start looking for free space starting at this point.
772 if (blkno >= bmp->db_mapsize)
775 agno = blkno >> bmp->db_agl2size;
777 /* check if blkno crosses over into a new allocation group.
778 * if so, check if we should allow allocations within this
781 if ((blkno & (bmp->db_agsize - 1)) == 0)
782 /* check if the AG is currently being written to.
783 * if so, call dbNextAG() to find a non-busy
784 * AG with sufficient free space.
786 if (atomic_read(&bmp->db_active[agno]))
789 /* check if the allocation request size can be satisfied from a
790 * single dmap. if so, try to allocate from the dmap containing
791 * the hint using a tiered strategy.
793 if (nblocks <= BPERDMAP) {
794 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
796 /* get the buffer for the dmap containing the hint.
799 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
800 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
804 dp = (struct dmap *) mp->data;
806 /* first, try to satisfy the allocation request with the
807 * blocks beginning at the hint.
809 if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks))
813 mark_metapage_dirty(mp);
816 release_metapage(mp);
820 writers = atomic_read(&bmp->db_active[agno]);
822 ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) {
824 * Someone else is writing in this allocation
825 * group. To avoid fragmenting, try another ag
827 release_metapage(mp);
828 IREAD_UNLOCK(ipbmap);
832 /* next, try to satisfy the allocation request with blocks
836 dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results))
839 mark_metapage_dirty(mp);
841 release_metapage(mp);
845 /* try to satisfy the allocation request with blocks within
846 * the same dmap as the hint.
848 if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results))
851 mark_metapage_dirty(mp);
853 release_metapage(mp);
857 release_metapage(mp);
858 IREAD_UNLOCK(ipbmap);
861 /* try to satisfy the allocation request with blocks within
862 * the same allocation group as the hint.
864 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
865 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC)
868 IWRITE_UNLOCK(ipbmap);
873 * Let dbNextAG recommend a preferred allocation group
875 agno = dbNextAG(ipbmap);
876 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
878 /* Try to allocate within this allocation group. if that fails, try to
879 * allocate anywhere in the map.
881 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC)
882 rc = dbAllocAny(bmp, nblocks, l2nb, results);
885 IWRITE_UNLOCK(ipbmap);
890 IREAD_UNLOCK(ipbmap);
898 * FUNCTION: attempt to extend a current allocation by a specified
901 * this routine attempts to satisfy the allocation request
902 * by first trying to extend the existing allocation in
903 * place by allocating the additional blocks as the blocks
904 * immediately following the current allocation. if these
905 * blocks are not available, this routine will attempt to
906 * allocate a new set of contiguous blocks large enough
907 * to cover the existing allocation plus the additional
908 * number of blocks required.
911 * ip - pointer to in-core inode requiring allocation.
912 * blkno - starting block of the current allocation.
913 * nblocks - number of contiguous blocks within the current
915 * addnblocks - number of blocks to add to the allocation.
916 * results - on successful return, set to the starting block number
917 * of the existing allocation if the existing allocation
918 * was extended in place or to a newly allocated contiguous
919 * range if the existing allocation could not be extended
924 * -ENOSPC - insufficient disk resources
928 dbReAlloc(struct inode *ip,
929 s64 blkno, s64 nblocks, s64 addnblocks, s64 * results)
933 /* try to extend the allocation in place.
935 if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) {
943 /* could not extend the allocation in place, so allocate a
944 * new set of blocks for the entire request (i.e. try to get
945 * a range of contiguous blocks large enough to cover the
946 * existing allocation plus the additional blocks.)
949 (ip, blkno + nblocks - 1, addnblocks + nblocks, results));
956 * FUNCTION: attempt to extend a current allocation by a specified
959 * this routine attempts to satisfy the allocation request
960 * by first trying to extend the existing allocation in
961 * place by allocating the additional blocks as the blocks
962 * immediately following the current allocation.
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
969 * addnblocks - number of blocks to add to the allocation.
973 * -ENOSPC - insufficient disk resources
976 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks)
978 struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
979 s64 lblkno, lastblkno, extblkno;
984 struct inode *ipbmap = sbi->ipbmap;
988 * We don't want a non-aligned extent to cross a page boundary
990 if (((rel_block = blkno & (sbi->nbperpage - 1))) &&
991 (rel_block + nblocks + addnblocks > sbi->nbperpage))
994 /* get the last block of the current allocation */
995 lastblkno = blkno + nblocks - 1;
997 /* determine the block number of the block following
998 * the existing allocation.
1000 extblkno = lastblkno + 1;
1002 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
1004 /* better be within the file system */
1006 if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) {
1007 IREAD_UNLOCK(ipbmap);
1008 jfs_error(ip->i_sb, "the block is outside the filesystem\n");
1012 /* we'll attempt to extend the current allocation in place by
1013 * allocating the additional blocks as the blocks immediately
1014 * following the current allocation. we only try to extend the
1015 * current allocation in place if the number of additional blocks
1016 * can fit into a dmap, the last block of the current allocation
1017 * is not the last block of the file system, and the start of the
1018 * inplace extension is not on an allocation group boundary.
1020 if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize ||
1021 (extblkno & (bmp->db_agsize - 1)) == 0) {
1022 IREAD_UNLOCK(ipbmap);
1026 /* get the buffer for the dmap containing the first block
1029 lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage);
1030 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
1032 IREAD_UNLOCK(ipbmap);
1036 dp = (struct dmap *) mp->data;
1038 /* try to allocate the blocks immediately following the
1039 * current allocation.
1041 rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks);
1043 IREAD_UNLOCK(ipbmap);
1045 /* were we successful ? */
1049 /* we were not successful */
1050 release_metapage(mp);
1057 * NAME: dbAllocNext()
1059 * FUNCTION: attempt to allocate the blocks of the specified block
1060 * range within a dmap.
1063 * bmp - pointer to bmap descriptor
1064 * dp - pointer to dmap.
1065 * blkno - starting block number of the range.
1066 * nblocks - number of contiguous free blocks of the range.
1070 * -ENOSPC - insufficient disk resources
1073 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1075 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno,
1078 int dbitno, word, rembits, nb, nwords, wbitno, nw;
1083 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1084 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1088 /* pick up a pointer to the leaves of the dmap tree.
1090 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1092 /* determine the bit number and word within the dmap of the
1095 dbitno = blkno & (BPERDMAP - 1);
1096 word = dbitno >> L2DBWORD;
1098 /* check if the specified block range is contained within
1101 if (dbitno + nblocks > BPERDMAP)
1104 /* check if the starting leaf indicates that anything
1107 if (leaf[word] == NOFREE)
1110 /* check the dmaps words corresponding to block range to see
1111 * if the block range is free. not all bits of the first and
1112 * last words may be contained within the block range. if this
1113 * is the case, we'll work against those words (i.e. partial first
1114 * and/or last) on an individual basis (a single pass) and examine
1115 * the actual bits to determine if they are free. a single pass
1116 * will be used for all dmap words fully contained within the
1117 * specified range. within this pass, the leaves of the dmap
1118 * tree will be examined to determine if the blocks are free. a
1119 * single leaf may describe the free space of multiple dmap
1120 * words, so we may visit only a subset of the actual leaves
1121 * corresponding to the dmap words of the block range.
1123 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
1124 /* determine the bit number within the word and
1125 * the number of bits within the word.
1127 wbitno = dbitno & (DBWORD - 1);
1128 nb = min(rembits, DBWORD - wbitno);
1130 /* check if only part of the word is to be examined.
1133 /* check if the bits are free.
1135 mask = (ONES << (DBWORD - nb) >> wbitno);
1136 if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask)
1141 /* one or more dmap words are fully contained
1142 * within the block range. determine how many
1143 * words and how many bits.
1145 nwords = rembits >> L2DBWORD;
1146 nb = nwords << L2DBWORD;
1148 /* now examine the appropriate leaves to determine
1149 * if the blocks are free.
1151 while (nwords > 0) {
1152 /* does the leaf describe any free space ?
1154 if (leaf[word] < BUDMIN)
1157 /* determine the l2 number of bits provided
1161 min_t(int, leaf[word], NLSTOL2BSZ(nwords));
1163 /* determine how many words were handled.
1165 nw = BUDSIZE(l2size, BUDMIN);
1173 /* allocate the blocks.
1175 return (dbAllocDmap(bmp, dp, blkno, nblocks));
1180 * NAME: dbAllocNear()
1182 * FUNCTION: attempt to allocate a number of contiguous free blocks near
1183 * a specified block (hint) within a dmap.
1185 * starting with the dmap leaf that covers the hint, we'll
1186 * check the next four contiguous leaves for sufficient free
1187 * space. if sufficient free space is found, we'll allocate
1188 * the desired free space.
1191 * bmp - pointer to bmap descriptor
1192 * dp - pointer to dmap.
1193 * blkno - block number to allocate near.
1194 * nblocks - actual number of contiguous free blocks desired.
1195 * l2nb - log2 number of contiguous free blocks desired.
1196 * results - on successful return, set to the starting block number
1197 * of the newly allocated range.
1201 * -ENOSPC - insufficient disk resources
1204 * serialization: IREAD_LOCK(ipbmap) held on entry/exit;
1207 dbAllocNear(struct bmap * bmp,
1208 struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results)
1210 int word, lword, rc;
1213 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) {
1214 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n");
1218 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx);
1220 /* determine the word within the dmap that holds the hint
1221 * (i.e. blkno). also, determine the last word in the dmap
1222 * that we'll include in our examination.
1224 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
1225 lword = min(word + 4, LPERDMAP);
1227 /* examine the leaves for sufficient free space.
1229 for (; word < lword; word++) {
1230 /* does the leaf describe sufficient free space ?
1232 if (leaf[word] < l2nb)
1235 /* determine the block number within the file system
1236 * of the first block described by this dmap word.
1238 blkno = le64_to_cpu(dp->start) + (word << L2DBWORD);
1240 /* if not all bits of the dmap word are free, get the
1241 * starting bit number within the dmap word of the required
1242 * string of free bits and adjust the block number with the
1245 if (leaf[word] < BUDMIN)
1247 dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb);
1249 /* allocate the blocks.
1251 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1264 * FUNCTION: attempt to allocate the specified number of contiguous
1265 * free blocks within the specified allocation group.
1267 * unless the allocation group size is equal to the number
1268 * of blocks per dmap, the dmap control pages will be used to
1269 * find the required free space, if available. we start the
1270 * search at the highest dmap control page level which
1271 * distinctly describes the allocation group's free space
1272 * (i.e. the highest level at which the allocation group's
1273 * free space is not mixed in with that of any other group).
1274 * in addition, we start the search within this level at a
1275 * height of the dmapctl dmtree at which the nodes distinctly
1276 * describe the allocation group's free space. at this height,
1277 * the allocation group's free space may be represented by 1
1278 * or two sub-trees, depending on the allocation group size.
1279 * we search the top nodes of these subtrees left to right for
1280 * sufficient free space. if sufficient free space is found,
1281 * the subtree is searched to find the leftmost leaf that
1282 * has free space. once we have made it to the leaf, we
1283 * move the search to the next lower level dmap control page
1284 * corresponding to this leaf. we continue down the dmap control
1285 * pages until we find the dmap that contains or starts the
1286 * sufficient free space and we allocate at this dmap.
1288 * if the allocation group size is equal to the dmap size,
1289 * we'll start at the dmap corresponding to the allocation
1290 * group and attempt the allocation at this level.
1292 * the dmap control page search is also not performed if the
1293 * allocation group is completely free and we go to the first
1294 * dmap of the allocation group to do the allocation. this is
1295 * done because the allocation group may be part (not the first
1296 * part) of a larger binary buddy system, causing the dmap
1297 * control pages to indicate no free space (NOFREE) within
1298 * the allocation group.
1301 * bmp - pointer to bmap descriptor
1302 * agno - allocation group number.
1303 * nblocks - actual number of contiguous free blocks desired.
1304 * l2nb - log2 number of contiguous free blocks desired.
1305 * results - on successful return, set to the starting block number
1306 * of the newly allocated range.
1310 * -ENOSPC - insufficient disk resources
1313 * note: IWRITE_LOCK(ipmap) held on entry/exit;
1316 dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results)
1318 struct metapage *mp;
1319 struct dmapctl *dcp;
1320 int rc, ti, i, k, m, n, agperlev;
1324 /* allocation request should not be for more than the
1325 * allocation group size.
1327 if (l2nb > bmp->db_agl2size) {
1328 jfs_error(bmp->db_ipbmap->i_sb,
1329 "allocation request is larger than the allocation group size\n");
1333 /* determine the starting block number of the allocation
1336 blkno = (s64) agno << bmp->db_agl2size;
1338 /* check if the allocation group size is the minimum allocation
1339 * group size or if the allocation group is completely free. if
1340 * the allocation group size is the minimum size of BPERDMAP (i.e.
1341 * 1 dmap), there is no need to search the dmap control page (below)
1342 * that fully describes the allocation group since the allocation
1343 * group is already fully described by a dmap. in this case, we
1344 * just call dbAllocCtl() to search the dmap tree and allocate the
1345 * required space if available.
1347 * if the allocation group is completely free, dbAllocCtl() is
1348 * also called to allocate the required space. this is done for
1349 * two reasons. first, it makes no sense searching the dmap control
1350 * pages for free space when we know that free space exists. second,
1351 * the dmap control pages may indicate that the allocation group
1352 * has no free space if the allocation group is part (not the first
1353 * part) of a larger binary buddy system.
1355 if (bmp->db_agsize == BPERDMAP
1356 || bmp->db_agfree[agno] == bmp->db_agsize) {
1357 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1358 if ((rc == -ENOSPC) &&
1359 (bmp->db_agfree[agno] == bmp->db_agsize)) {
1360 printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n",
1361 (unsigned long long) blkno,
1362 (unsigned long long) nblocks);
1363 jfs_error(bmp->db_ipbmap->i_sb,
1364 "dbAllocCtl failed in free AG\n");
1369 /* the buffer for the dmap control page that fully describes the
1372 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel);
1373 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1376 dcp = (struct dmapctl *) mp->data;
1377 budmin = dcp->budmin;
1379 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1380 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
1381 release_metapage(mp);
1385 /* search the subtree(s) of the dmap control page that describes
1386 * the allocation group, looking for sufficient free space. to begin,
1387 * determine how many allocation groups are represented in a dmap
1388 * control page at the control page level (i.e. L0, L1, L2) that
1389 * fully describes an allocation group. next, determine the starting
1390 * tree index of this allocation group within the control page.
1393 (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth;
1394 ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1));
1396 /* dmap control page trees fan-out by 4 and a single allocation
1397 * group may be described by 1 or 2 subtrees within the ag level
1398 * dmap control page, depending upon the ag size. examine the ag's
1399 * subtrees for sufficient free space, starting with the leftmost
1402 for (i = 0; i < bmp->db_agwidth; i++, ti++) {
1403 /* is there sufficient free space ?
1405 if (l2nb > dcp->stree[ti])
1408 /* sufficient free space found in a subtree. now search down
1409 * the subtree to find the leftmost leaf that describes this
1412 for (k = bmp->db_agheight; k > 0; k--) {
1413 for (n = 0, m = (ti << 2) + 1; n < 4; n++) {
1414 if (l2nb <= dcp->stree[m + n]) {
1420 jfs_error(bmp->db_ipbmap->i_sb,
1421 "failed descending stree\n");
1422 release_metapage(mp);
1427 /* determine the block number within the file system
1428 * that corresponds to this leaf.
1430 if (bmp->db_aglevel == 2)
1432 else if (bmp->db_aglevel == 1)
1433 blkno &= ~(MAXL1SIZE - 1);
1434 else /* bmp->db_aglevel == 0 */
1435 blkno &= ~(MAXL0SIZE - 1);
1438 ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin;
1440 /* release the buffer in preparation for going down
1441 * the next level of dmap control pages.
1443 release_metapage(mp);
1445 /* check if we need to continue to search down the lower
1446 * level dmap control pages. we need to if the number of
1447 * blocks required is less than maximum number of blocks
1448 * described at the next lower level.
1450 if (l2nb < budmin) {
1452 /* search the lower level dmap control pages to get
1453 * the starting block number of the dmap that
1454 * contains or starts off the free space.
1457 dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1,
1459 if (rc == -ENOSPC) {
1460 jfs_error(bmp->db_ipbmap->i_sb,
1461 "control page inconsistent\n");
1468 /* allocate the blocks.
1470 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1471 if (rc == -ENOSPC) {
1472 jfs_error(bmp->db_ipbmap->i_sb,
1473 "unable to allocate blocks\n");
1479 /* no space in the allocation group. release the buffer and
1482 release_metapage(mp);
1489 * NAME: dbAllocAny()
1491 * FUNCTION: attempt to allocate the specified number of contiguous
1492 * free blocks anywhere in the file system.
1494 * dbAllocAny() attempts to find the sufficient free space by
1495 * searching down the dmap control pages, starting with the
1496 * highest level (i.e. L0, L1, L2) control page. if free space
1497 * large enough to satisfy the desired free space is found, the
1498 * desired free space is allocated.
1501 * bmp - pointer to bmap descriptor
1502 * nblocks - actual number of contiguous free blocks desired.
1503 * l2nb - log2 number of contiguous free blocks desired.
1504 * results - on successful return, set to the starting block number
1505 * of the newly allocated range.
1509 * -ENOSPC - insufficient disk resources
1512 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1514 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results)
1519 /* starting with the top level dmap control page, search
1520 * down the dmap control levels for sufficient free space.
1521 * if free space is found, dbFindCtl() returns the starting
1522 * block number of the dmap that contains or starts off the
1523 * range of free space.
1525 if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno)))
1528 /* allocate the blocks.
1530 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results);
1531 if (rc == -ENOSPC) {
1532 jfs_error(bmp->db_ipbmap->i_sb, "unable to allocate blocks\n");
1540 * NAME: dbDiscardAG()
1542 * FUNCTION: attempt to discard (TRIM) all free blocks of specific AG
1545 * 1) allocate blocks, as large as possible and save them
1546 * while holding IWRITE_LOCK on ipbmap
1547 * 2) trim all these saved block/length values
1548 * 3) mark the blocks free again
1551 * - we work only on one ag at some time, minimizing how long we
1552 * need to lock ipbmap
1553 * - reading / writing the fs is possible most time, even on
1557 * - we write two times to the dmapctl and dmap pages
1558 * - but for me, this seems the best way, better ideas?
1562 * ip - pointer to in-core inode
1564 * minlen - minimum value of contiguous blocks
1567 * s64 - actual number of blocks trimmed
1569 s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen)
1571 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
1572 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
1576 struct super_block *sb = ipbmap->i_sb;
1583 /* max blkno / nblocks pairs to trim */
1584 int count = 0, range_cnt;
1587 /* prevent others from writing new stuff here, while trimming */
1588 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP);
1590 nblocks = bmp->db_agfree[agno];
1591 max_ranges = nblocks;
1592 do_div(max_ranges, minlen);
1593 range_cnt = min_t(u64, max_ranges + 1, 32 * 1024);
1594 totrim = kmalloc_array(range_cnt, sizeof(struct range2trim), GFP_NOFS);
1595 if (totrim == NULL) {
1596 jfs_error(bmp->db_ipbmap->i_sb, "no memory for trim array\n");
1597 IWRITE_UNLOCK(ipbmap);
1602 while (nblocks >= minlen) {
1603 l2nb = BLKSTOL2(nblocks);
1605 /* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */
1606 rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno);
1609 tt->nblocks = nblocks;
1612 /* the whole ag is free, trim now */
1613 if (bmp->db_agfree[agno] == 0)
1616 /* give a hint for the next while */
1617 nblocks = bmp->db_agfree[agno];
1619 } else if (rc == -ENOSPC) {
1620 /* search for next smaller log2 block */
1621 l2nb = BLKSTOL2(nblocks) - 1;
1622 nblocks = 1LL << l2nb;
1624 /* Trim any already allocated blocks */
1625 jfs_error(bmp->db_ipbmap->i_sb, "-EIO\n");
1629 /* check, if our trim array is full */
1630 if (unlikely(count >= range_cnt - 1))
1633 IWRITE_UNLOCK(ipbmap);
1635 tt->nblocks = 0; /* mark the current end */
1636 for (tt = totrim; tt->nblocks != 0; tt++) {
1637 /* when mounted with online discard, dbFree() will
1638 * call jfs_issue_discard() itself */
1639 if (!(JFS_SBI(sb)->flag & JFS_DISCARD))
1640 jfs_issue_discard(ip, tt->blkno, tt->nblocks);
1641 dbFree(ip, tt->blkno, tt->nblocks);
1642 trimmed += tt->nblocks;
1652 * FUNCTION: starting at a specified dmap control page level and block
1653 * number, search down the dmap control levels for a range of
1654 * contiguous free blocks large enough to satisfy an allocation
1655 * request for the specified number of free blocks.
1657 * if sufficient contiguous free blocks are found, this routine
1658 * returns the starting block number within a dmap page that
1659 * contains or starts a range of contiqious free blocks that
1660 * is sufficient in size.
1663 * bmp - pointer to bmap descriptor
1664 * level - starting dmap control page level.
1665 * l2nb - log2 number of contiguous free blocks desired.
1666 * *blkno - on entry, starting block number for conducting the search.
1667 * on successful return, the first block within a dmap page
1668 * that contains or starts a range of contiguous free blocks.
1672 * -ENOSPC - insufficient disk resources
1675 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1677 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno)
1679 int rc, leafidx, lev;
1681 struct dmapctl *dcp;
1683 struct metapage *mp;
1685 /* starting at the specified dmap control page level and block
1686 * number, search down the dmap control levels for the starting
1687 * block number of a dmap page that contains or starts off
1688 * sufficient free blocks.
1690 for (lev = level, b = *blkno; lev >= 0; lev--) {
1691 /* get the buffer of the dmap control page for the block
1692 * number and level (i.e. L0, L1, L2).
1694 lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev);
1695 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1698 dcp = (struct dmapctl *) mp->data;
1699 budmin = dcp->budmin;
1701 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
1702 jfs_error(bmp->db_ipbmap->i_sb,
1703 "Corrupt dmapctl page\n");
1704 release_metapage(mp);
1708 /* search the tree within the dmap control page for
1709 * sufficient free space. if sufficient free space is found,
1710 * dbFindLeaf() returns the index of the leaf at which
1711 * free space was found.
1713 rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx);
1715 /* release the buffer.
1717 release_metapage(mp);
1723 jfs_error(bmp->db_ipbmap->i_sb,
1724 "dmap inconsistent\n");
1730 /* adjust the block number to reflect the location within
1731 * the dmap control page (i.e. the leaf) at which free
1734 b += (((s64) leafidx) << budmin);
1736 /* we stop the search at this dmap control page level if
1737 * the number of blocks required is greater than or equal
1738 * to the maximum number of blocks described at the next
1751 * NAME: dbAllocCtl()
1753 * FUNCTION: attempt to allocate a specified number of contiguous
1754 * blocks starting within a specific dmap.
1756 * this routine is called by higher level routines that search
1757 * the dmap control pages above the actual dmaps for contiguous
1758 * free space. the result of successful searches by these
1759 * routines are the starting block numbers within dmaps, with
1760 * the dmaps themselves containing the desired contiguous free
1761 * space or starting a contiguous free space of desired size
1762 * that is made up of the blocks of one or more dmaps. these
1763 * calls should not fail due to insufficent resources.
1765 * this routine is called in some cases where it is not known
1766 * whether it will fail due to insufficient resources. more
1767 * specifically, this occurs when allocating from an allocation
1768 * group whose size is equal to the number of blocks per dmap.
1769 * in this case, the dmap control pages are not examined prior
1770 * to calling this routine (to save pathlength) and the call
1773 * for a request size that fits within a dmap, this routine relies
1774 * upon the dmap's dmtree to find the requested contiguous free
1775 * space. for request sizes that are larger than a dmap, the
1776 * requested free space will start at the first block of the
1777 * first dmap (i.e. blkno).
1780 * bmp - pointer to bmap descriptor
1781 * nblocks - actual number of contiguous free blocks to allocate.
1782 * l2nb - log2 number of contiguous free blocks to allocate.
1783 * blkno - starting block number of the dmap to start the allocation
1785 * results - on successful return, set to the starting block number
1786 * of the newly allocated range.
1790 * -ENOSPC - insufficient disk resources
1793 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit;
1796 dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results)
1800 struct metapage *mp;
1803 /* check if the allocation request is confined to a single dmap.
1805 if (l2nb <= L2BPERDMAP) {
1806 /* get the buffer for the dmap.
1808 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
1809 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1812 dp = (struct dmap *) mp->data;
1814 /* try to allocate the blocks.
1816 rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results);
1818 mark_metapage_dirty(mp);
1820 release_metapage(mp);
1825 /* allocation request involving multiple dmaps. it must start on
1828 assert((blkno & (BPERDMAP - 1)) == 0);
1830 /* allocate the blocks dmap by dmap.
1832 for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) {
1833 /* get the buffer for the dmap.
1835 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1836 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1841 dp = (struct dmap *) mp->data;
1843 /* the dmap better be all free.
1845 if (dp->tree.stree[ROOT] != L2BPERDMAP) {
1846 release_metapage(mp);
1847 jfs_error(bmp->db_ipbmap->i_sb,
1848 "the dmap is not all free\n");
1853 /* determine how many blocks to allocate from this dmap.
1855 nb = min_t(s64, n, BPERDMAP);
1857 /* allocate the blocks from the dmap.
1859 if ((rc = dbAllocDmap(bmp, dp, b, nb))) {
1860 release_metapage(mp);
1864 /* write the buffer.
1869 /* set the results (starting block number) and return.
1874 /* something failed in handling an allocation request involving
1875 * multiple dmaps. we'll try to clean up by backing out any
1876 * allocation that has already happened for this request. if
1877 * we fail in backing out the allocation, we'll mark the file
1878 * system to indicate that blocks have been leaked.
1882 /* try to backout the allocations dmap by dmap.
1884 for (n = nblocks - n, b = blkno; n > 0;
1885 n -= BPERDMAP, b += BPERDMAP) {
1886 /* get the buffer for this dmap.
1888 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage);
1889 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
1891 /* could not back out. mark the file system
1892 * to indicate that we have leaked blocks.
1894 jfs_error(bmp->db_ipbmap->i_sb,
1895 "I/O Error: Block Leakage\n");
1898 dp = (struct dmap *) mp->data;
1900 /* free the blocks is this dmap.
1902 if (dbFreeDmap(bmp, dp, b, BPERDMAP)) {
1903 /* could not back out. mark the file system
1904 * to indicate that we have leaked blocks.
1906 release_metapage(mp);
1907 jfs_error(bmp->db_ipbmap->i_sb, "Block Leakage\n");
1911 /* write the buffer.
1921 * NAME: dbAllocDmapLev()
1923 * FUNCTION: attempt to allocate a specified number of contiguous blocks
1924 * from a specified dmap.
1926 * this routine checks if the contiguous blocks are available.
1927 * if so, nblocks of blocks are allocated; otherwise, ENOSPC is
1931 * mp - pointer to bmap descriptor
1932 * dp - pointer to dmap to attempt to allocate blocks from.
1933 * l2nb - log2 number of contiguous block desired.
1934 * nblocks - actual number of contiguous block desired.
1935 * results - on successful return, set to the starting block number
1936 * of the newly allocated range.
1940 * -ENOSPC - insufficient disk resources
1943 * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or
1944 * IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit;
1947 dbAllocDmapLev(struct bmap * bmp,
1948 struct dmap * dp, int nblocks, int l2nb, s64 * results)
1953 /* can't be more than a dmaps worth of blocks */
1954 assert(l2nb <= L2BPERDMAP);
1956 /* search the tree within the dmap page for sufficient
1957 * free space. if sufficient free space is found, dbFindLeaf()
1958 * returns the index of the leaf at which free space was found.
1960 if (dbFindLeaf((dmtree_t *) & dp->tree, l2nb, &leafidx))
1966 /* determine the block number within the file system corresponding
1967 * to the leaf at which free space was found.
1969 blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD);
1971 /* if not all bits of the dmap word are free, get the starting
1972 * bit number within the dmap word of the required string of free
1973 * bits and adjust the block number with this value.
1975 if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN)
1976 blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb);
1978 /* allocate the blocks */
1979 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0)
1987 * NAME: dbAllocDmap()
1989 * FUNCTION: adjust the disk allocation map to reflect the allocation
1990 * of a specified block range within a dmap.
1992 * this routine allocates the specified blocks from the dmap
1993 * through a call to dbAllocBits(). if the allocation of the
1994 * block range causes the maximum string of free blocks within
1995 * the dmap to change (i.e. the value of the root of the dmap's
1996 * dmtree), this routine will cause this change to be reflected
1997 * up through the appropriate levels of the dmap control pages
1998 * by a call to dbAdjCtl() for the L0 dmap control page that
2002 * bmp - pointer to bmap descriptor
2003 * dp - pointer to dmap to allocate the block range from.
2004 * blkno - starting block number of the block to be allocated.
2005 * nblocks - number of blocks to be allocated.
2011 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2013 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2019 /* save the current value of the root (i.e. maximum free string)
2022 oldroot = dp->tree.stree[ROOT];
2024 /* allocate the specified (blocks) bits */
2025 dbAllocBits(bmp, dp, blkno, nblocks);
2027 /* if the root has not changed, done. */
2028 if (dp->tree.stree[ROOT] == oldroot)
2031 /* root changed. bubble the change up to the dmap control pages.
2032 * if the adjustment of the upper level control pages fails,
2033 * backout the bit allocation (thus making everything consistent).
2035 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0)))
2036 dbFreeBits(bmp, dp, blkno, nblocks);
2043 * NAME: dbFreeDmap()
2045 * FUNCTION: adjust the disk allocation map to reflect the allocation
2046 * of a specified block range within a dmap.
2048 * this routine frees the specified blocks from the dmap through
2049 * a call to dbFreeBits(). if the deallocation of the block range
2050 * causes the maximum string of free blocks within the dmap to
2051 * change (i.e. the value of the root of the dmap's dmtree), this
2052 * routine will cause this change to be reflected up through the
2053 * appropriate levels of the dmap control pages by a call to
2054 * dbAdjCtl() for the L0 dmap control page that covers this dmap.
2057 * bmp - pointer to bmap descriptor
2058 * dp - pointer to dmap to free the block range from.
2059 * blkno - starting block number of the block to be freed.
2060 * nblocks - number of blocks to be freed.
2066 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2068 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno,
2074 /* save the current value of the root (i.e. maximum free string)
2077 oldroot = dp->tree.stree[ROOT];
2079 /* free the specified (blocks) bits */
2080 rc = dbFreeBits(bmp, dp, blkno, nblocks);
2082 /* if error or the root has not changed, done. */
2083 if (rc || (dp->tree.stree[ROOT] == oldroot))
2086 /* root changed. bubble the change up to the dmap control pages.
2087 * if the adjustment of the upper level control pages fails,
2088 * backout the deallocation.
2090 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) {
2091 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD;
2093 /* as part of backing out the deallocation, we will have
2094 * to back split the dmap tree if the deallocation caused
2095 * the freed blocks to become part of a larger binary buddy
2098 if (dp->tree.stree[word] == NOFREE)
2099 dbBackSplit((dmtree_t *) & dp->tree, word);
2101 dbAllocBits(bmp, dp, blkno, nblocks);
2109 * NAME: dbAllocBits()
2111 * FUNCTION: allocate a specified block range from a dmap.
2113 * this routine updates the dmap to reflect the working
2114 * state allocation of the specified block range. it directly
2115 * updates the bits of the working map and causes the adjustment
2116 * of the binary buddy system described by the dmap's dmtree
2117 * leaves to reflect the bits allocated. it also causes the
2118 * dmap's dmtree, as a whole, to reflect the allocated range.
2121 * bmp - pointer to bmap descriptor
2122 * dp - pointer to dmap to allocate bits from.
2123 * blkno - starting block number of the bits to be allocated.
2124 * nblocks - number of bits to be allocated.
2126 * RETURN VALUES: none
2128 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2130 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2133 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2134 dmtree_t *tp = (dmtree_t *) & dp->tree;
2138 /* pick up a pointer to the leaves of the dmap tree */
2139 leaf = dp->tree.stree + LEAFIND;
2141 /* determine the bit number and word within the dmap of the
2144 dbitno = blkno & (BPERDMAP - 1);
2145 word = dbitno >> L2DBWORD;
2147 /* block range better be within the dmap */
2148 assert(dbitno + nblocks <= BPERDMAP);
2150 /* allocate the bits of the dmap's words corresponding to the block
2151 * range. not all bits of the first and last words may be contained
2152 * within the block range. if this is the case, we'll work against
2153 * those words (i.e. partial first and/or last) on an individual basis
2154 * (a single pass), allocating the bits of interest by hand and
2155 * updating the leaf corresponding to the dmap word. a single pass
2156 * will be used for all dmap words fully contained within the
2157 * specified range. within this pass, the bits of all fully contained
2158 * dmap words will be marked as free in a single shot and the leaves
2159 * will be updated. a single leaf may describe the free space of
2160 * multiple dmap words, so we may update only a subset of the actual
2161 * leaves corresponding to the dmap words of the block range.
2163 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2164 /* determine the bit number within the word and
2165 * the number of bits within the word.
2167 wbitno = dbitno & (DBWORD - 1);
2168 nb = min(rembits, DBWORD - wbitno);
2170 /* check if only part of a word is to be allocated.
2173 /* allocate (set to 1) the appropriate bits within
2176 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
2179 /* update the leaf for this dmap word. in addition
2180 * to setting the leaf value to the binary buddy max
2181 * of the updated dmap word, dbSplit() will split
2182 * the binary system of the leaves if need be.
2184 dbSplit(tp, word, BUDMIN,
2185 dbMaxBud((u8 *) & dp->wmap[word]));
2189 /* one or more dmap words are fully contained
2190 * within the block range. determine how many
2191 * words and allocate (set to 1) the bits of these
2194 nwords = rembits >> L2DBWORD;
2195 memset(&dp->wmap[word], (int) ONES, nwords * 4);
2197 /* determine how many bits.
2199 nb = nwords << L2DBWORD;
2201 /* now update the appropriate leaves to reflect
2202 * the allocated words.
2204 for (; nwords > 0; nwords -= nw) {
2205 if (leaf[word] < BUDMIN) {
2206 jfs_error(bmp->db_ipbmap->i_sb,
2207 "leaf page corrupt\n");
2211 /* determine what the leaf value should be
2212 * updated to as the minimum of the l2 number
2213 * of bits being allocated and the l2 number
2214 * of bits currently described by this leaf.
2216 size = min_t(int, leaf[word],
2217 NLSTOL2BSZ(nwords));
2219 /* update the leaf to reflect the allocation.
2220 * in addition to setting the leaf value to
2221 * NOFREE, dbSplit() will split the binary
2222 * system of the leaves to reflect the current
2223 * allocation (size).
2225 dbSplit(tp, word, size, NOFREE);
2227 /* get the number of dmap words handled */
2228 nw = BUDSIZE(size, BUDMIN);
2234 /* update the free count for this dmap */
2235 le32_add_cpu(&dp->nfree, -nblocks);
2239 /* if this allocation group is completely free,
2240 * update the maximum allocation group number if this allocation
2241 * group is the new max.
2243 agno = blkno >> bmp->db_agl2size;
2244 if (agno > bmp->db_maxag)
2245 bmp->db_maxag = agno;
2247 /* update the free count for the allocation group and map */
2248 bmp->db_agfree[agno] -= nblocks;
2249 bmp->db_nfree -= nblocks;
2256 * NAME: dbFreeBits()
2258 * FUNCTION: free a specified block range from a dmap.
2260 * this routine updates the dmap to reflect the working
2261 * state allocation of the specified block range. it directly
2262 * updates the bits of the working map and causes the adjustment
2263 * of the binary buddy system described by the dmap's dmtree
2264 * leaves to reflect the bits freed. it also causes the dmap's
2265 * dmtree, as a whole, to reflect the deallocated range.
2268 * bmp - pointer to bmap descriptor
2269 * dp - pointer to dmap to free bits from.
2270 * blkno - starting block number of the bits to be freed.
2271 * nblocks - number of bits to be freed.
2273 * RETURN VALUES: 0 for success
2275 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2277 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno,
2280 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno;
2281 dmtree_t *tp = (dmtree_t *) & dp->tree;
2285 /* determine the bit number and word within the dmap of the
2288 dbitno = blkno & (BPERDMAP - 1);
2289 word = dbitno >> L2DBWORD;
2291 /* block range better be within the dmap.
2293 assert(dbitno + nblocks <= BPERDMAP);
2295 /* free the bits of the dmaps words corresponding to the block range.
2296 * not all bits of the first and last words may be contained within
2297 * the block range. if this is the case, we'll work against those
2298 * words (i.e. partial first and/or last) on an individual basis
2299 * (a single pass), freeing the bits of interest by hand and updating
2300 * the leaf corresponding to the dmap word. a single pass will be used
2301 * for all dmap words fully contained within the specified range.
2302 * within this pass, the bits of all fully contained dmap words will
2303 * be marked as free in a single shot and the leaves will be updated. a
2304 * single leaf may describe the free space of multiple dmap words,
2305 * so we may update only a subset of the actual leaves corresponding
2306 * to the dmap words of the block range.
2308 * dbJoin() is used to update leaf values and will join the binary
2309 * buddy system of the leaves if the new leaf values indicate this
2312 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
2313 /* determine the bit number within the word and
2314 * the number of bits within the word.
2316 wbitno = dbitno & (DBWORD - 1);
2317 nb = min(rembits, DBWORD - wbitno);
2319 /* check if only part of a word is to be freed.
2322 /* free (zero) the appropriate bits within this
2326 cpu_to_le32(~(ONES << (DBWORD - nb)
2329 /* update the leaf for this dmap word.
2331 rc = dbJoin(tp, word,
2332 dbMaxBud((u8 *) & dp->wmap[word]));
2338 /* one or more dmap words are fully contained
2339 * within the block range. determine how many
2340 * words and free (zero) the bits of these words.
2342 nwords = rembits >> L2DBWORD;
2343 memset(&dp->wmap[word], 0, nwords * 4);
2345 /* determine how many bits.
2347 nb = nwords << L2DBWORD;
2349 /* now update the appropriate leaves to reflect
2352 for (; nwords > 0; nwords -= nw) {
2353 /* determine what the leaf value should be
2354 * updated to as the minimum of the l2 number
2355 * of bits being freed and the l2 (max) number
2356 * of bits that can be described by this leaf.
2360 (word, L2LPERDMAP, BUDMIN),
2361 NLSTOL2BSZ(nwords));
2365 rc = dbJoin(tp, word, size);
2369 /* get the number of dmap words handled.
2371 nw = BUDSIZE(size, BUDMIN);
2377 /* update the free count for this dmap.
2379 le32_add_cpu(&dp->nfree, nblocks);
2383 /* update the free count for the allocation group and
2386 agno = blkno >> bmp->db_agl2size;
2387 bmp->db_nfree += nblocks;
2388 bmp->db_agfree[agno] += nblocks;
2390 /* check if this allocation group is not completely free and
2391 * if it is currently the maximum (rightmost) allocation group.
2392 * if so, establish the new maximum allocation group number by
2393 * searching left for the first allocation group with allocation.
2395 if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) ||
2396 (agno == bmp->db_numag - 1 &&
2397 bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) {
2398 while (bmp->db_maxag > 0) {
2400 if (bmp->db_agfree[bmp->db_maxag] !=
2405 /* re-establish the allocation group preference if the
2406 * current preference is right of the maximum allocation
2409 if (bmp->db_agpref > bmp->db_maxag)
2410 bmp->db_agpref = bmp->db_maxag;
2422 * FUNCTION: adjust a dmap control page at a specified level to reflect
2423 * the change in a lower level dmap or dmap control page's
2424 * maximum string of free blocks (i.e. a change in the root
2425 * of the lower level object's dmtree) due to the allocation
2426 * or deallocation of a range of blocks with a single dmap.
2428 * on entry, this routine is provided with the new value of
2429 * the lower level dmap or dmap control page root and the
2430 * starting block number of the block range whose allocation
2431 * or deallocation resulted in the root change. this range
2432 * is respresented by a single leaf of the current dmapctl
2433 * and the leaf will be updated with this value, possibly
2434 * causing a binary buddy system within the leaves to be
2435 * split or joined. the update may also cause the dmapctl's
2436 * dmtree to be updated.
2438 * if the adjustment of the dmap control page, itself, causes its
2439 * root to change, this change will be bubbled up to the next dmap
2440 * control level by a recursive call to this routine, specifying
2441 * the new root value and the next dmap control page level to
2444 * bmp - pointer to bmap descriptor
2445 * blkno - the first block of a block range within a dmap. it is
2446 * the allocation or deallocation of this block range that
2447 * requires the dmap control page to be adjusted.
2448 * newval - the new value of the lower level dmap or dmap control
2450 * alloc - 'true' if adjustment is due to an allocation.
2451 * level - current level of dmap control page (i.e. L0, L1, L2) to
2458 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2461 dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level)
2463 struct metapage *mp;
2467 struct dmapctl *dcp;
2470 /* get the buffer for the dmap control page for the specified
2471 * block number and control page level.
2473 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level);
2474 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0);
2477 dcp = (struct dmapctl *) mp->data;
2479 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) {
2480 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n");
2481 release_metapage(mp);
2485 /* determine the leaf number corresponding to the block and
2486 * the index within the dmap control tree.
2488 leafno = BLKTOCTLLEAF(blkno, dcp->budmin);
2489 ti = leafno + le32_to_cpu(dcp->leafidx);
2491 /* save the current leaf value and the current root level (i.e.
2492 * maximum l2 free string described by this dmapctl).
2494 oldval = dcp->stree[ti];
2495 oldroot = dcp->stree[ROOT];
2497 /* check if this is a control page update for an allocation.
2498 * if so, update the leaf to reflect the new leaf value using
2499 * dbSplit(); otherwise (deallocation), use dbJoin() to update
2500 * the leaf with the new value. in addition to updating the
2501 * leaf, dbSplit() will also split the binary buddy system of
2502 * the leaves, if required, and bubble new values within the
2503 * dmapctl tree, if required. similarly, dbJoin() will join
2504 * the binary buddy system of leaves and bubble new values up
2505 * the dmapctl tree as required by the new leaf value.
2508 /* check if we are in the middle of a binary buddy
2509 * system. this happens when we are performing the
2510 * first allocation out of an allocation group that
2511 * is part (not the first part) of a larger binary
2512 * buddy system. if we are in the middle, back split
2513 * the system prior to calling dbSplit() which assumes
2514 * that it is at the front of a binary buddy system.
2516 if (oldval == NOFREE) {
2517 rc = dbBackSplit((dmtree_t *) dcp, leafno);
2519 release_metapage(mp);
2522 oldval = dcp->stree[ti];
2524 dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval);
2526 rc = dbJoin((dmtree_t *) dcp, leafno, newval);
2528 release_metapage(mp);
2533 /* check if the root of the current dmap control page changed due
2534 * to the update and if the current dmap control page is not at
2535 * the current top level (i.e. L0, L1, L2) of the map. if so (i.e.
2536 * root changed and this is not the top level), call this routine
2537 * again (recursion) for the next higher level of the mapping to
2538 * reflect the change in root for the current dmap control page.
2540 if (dcp->stree[ROOT] != oldroot) {
2541 /* are we below the top level of the map. if so,
2542 * bubble the root up to the next higher level.
2544 if (level < bmp->db_maxlevel) {
2545 /* bubble up the new root of this dmap control page to
2549 dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc,
2551 /* something went wrong in bubbling up the new
2552 * root value, so backout the changes to the
2553 * current dmap control page.
2556 dbJoin((dmtree_t *) dcp, leafno,
2559 /* the dbJoin() above might have
2560 * caused a larger binary buddy system
2561 * to form and we may now be in the
2562 * middle of it. if this is the case,
2563 * back split the buddies.
2565 if (dcp->stree[ti] == NOFREE)
2566 dbBackSplit((dmtree_t *)
2568 dbSplit((dmtree_t *) dcp, leafno,
2569 dcp->budmin, oldval);
2572 /* release the buffer and return the error.
2574 release_metapage(mp);
2578 /* we're at the top level of the map. update
2579 * the bmap control page to reflect the size
2580 * of the maximum free buddy system.
2582 assert(level == bmp->db_maxlevel);
2583 if (bmp->db_maxfreebud != oldroot) {
2584 jfs_error(bmp->db_ipbmap->i_sb,
2585 "the maximum free buddy is not the old root\n");
2587 bmp->db_maxfreebud = dcp->stree[ROOT];
2591 /* write the buffer.
2602 * FUNCTION: update the leaf of a dmtree with a new value, splitting
2603 * the leaf from the binary buddy system of the dmtree's
2604 * leaves, as required.
2607 * tp - pointer to the tree containing the leaf.
2608 * leafno - the number of the leaf to be updated.
2609 * splitsz - the size the binary buddy system starting at the leaf
2610 * must be split to, specified as the log2 number of blocks.
2611 * newval - the new value for the leaf.
2613 * RETURN VALUES: none
2615 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2617 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval)
2621 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2623 /* check if the leaf needs to be split.
2625 if (leaf[leafno] > tp->dmt_budmin) {
2626 /* the split occurs by cutting the buddy system in half
2627 * at the specified leaf until we reach the specified
2628 * size. pick up the starting split size (current size
2629 * - 1 in l2) and the corresponding buddy size.
2631 cursz = leaf[leafno] - 1;
2632 budsz = BUDSIZE(cursz, tp->dmt_budmin);
2634 /* split until we reach the specified size.
2636 while (cursz >= splitsz) {
2637 /* update the buddy's leaf with its new value.
2639 dbAdjTree(tp, leafno ^ budsz, cursz);
2641 /* on to the next size and buddy.
2648 /* adjust the dmap tree to reflect the specified leaf's new
2651 dbAdjTree(tp, leafno, newval);
2656 * NAME: dbBackSplit()
2658 * FUNCTION: back split the binary buddy system of dmtree leaves
2659 * that hold a specified leaf until the specified leaf
2660 * starts its own binary buddy system.
2662 * the allocators typically perform allocations at the start
2663 * of binary buddy systems and dbSplit() is used to accomplish
2664 * any required splits. in some cases, however, allocation
2665 * may occur in the middle of a binary system and requires a
2666 * back split, with the split proceeding out from the middle of
2667 * the system (less efficient) rather than the start of the
2668 * system (more efficient). the cases in which a back split
2669 * is required are rare and are limited to the first allocation
2670 * within an allocation group which is a part (not first part)
2671 * of a larger binary buddy system and a few exception cases
2672 * in which a previous join operation must be backed out.
2675 * tp - pointer to the tree containing the leaf.
2676 * leafno - the number of the leaf to be updated.
2678 * RETURN VALUES: none
2680 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit;
2682 static int dbBackSplit(dmtree_t * tp, int leafno)
2684 int budsz, bud, w, bsz, size;
2686 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2688 /* leaf should be part (not first part) of a binary
2691 assert(leaf[leafno] == NOFREE);
2693 /* the back split is accomplished by iteratively finding the leaf
2694 * that starts the buddy system that contains the specified leaf and
2695 * splitting that system in two. this iteration continues until
2696 * the specified leaf becomes the start of a buddy system.
2698 * determine maximum possible l2 size for the specified leaf.
2701 LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs),
2704 /* determine the number of leaves covered by this size. this
2705 * is the buddy size that we will start with as we search for
2706 * the buddy system that contains the specified leaf.
2708 budsz = BUDSIZE(size, tp->dmt_budmin);
2712 while (leaf[leafno] == NOFREE) {
2713 /* find the leftmost buddy leaf.
2715 for (w = leafno, bsz = budsz;; bsz <<= 1,
2716 w = (w < bud) ? w : bud) {
2717 if (bsz >= le32_to_cpu(tp->dmt_nleafs)) {
2718 jfs_err("JFS: block map error in dbBackSplit");
2722 /* determine the buddy.
2726 /* check if this buddy is the start of the system.
2728 if (leaf[bud] != NOFREE) {
2729 /* split the leaf at the start of the
2732 cursz = leaf[bud] - 1;
2733 dbSplit(tp, bud, cursz, cursz);
2739 if (leaf[leafno] != size) {
2740 jfs_err("JFS: wrong leaf value in dbBackSplit");
2750 * FUNCTION: update the leaf of a dmtree with a new value, joining
2751 * the leaf with other leaves of the dmtree into a multi-leaf
2752 * binary buddy system, as required.
2755 * tp - pointer to the tree containing the leaf.
2756 * leafno - the number of the leaf to be updated.
2757 * newval - the new value for the leaf.
2759 * RETURN VALUES: none
2761 static int dbJoin(dmtree_t * tp, int leafno, int newval)
2766 /* can the new leaf value require a join with other leaves ?
2768 if (newval >= tp->dmt_budmin) {
2769 /* pickup a pointer to the leaves of the tree.
2771 leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx);
2773 /* try to join the specified leaf into a large binary
2774 * buddy system. the join proceeds by attempting to join
2775 * the specified leafno with its buddy (leaf) at new value.
2776 * if the join occurs, we attempt to join the left leaf
2777 * of the joined buddies with its buddy at new value + 1.
2778 * we continue to join until we find a buddy that cannot be
2779 * joined (does not have a value equal to the size of the
2780 * last join) or until all leaves have been joined into a
2783 * get the buddy size (number of words covered) of
2786 budsz = BUDSIZE(newval, tp->dmt_budmin);
2790 while (budsz < le32_to_cpu(tp->dmt_nleafs)) {
2791 /* get the buddy leaf.
2793 buddy = leafno ^ budsz;
2795 /* if the leaf's new value is greater than its
2796 * buddy's value, we join no more.
2798 if (newval > leaf[buddy])
2801 /* It shouldn't be less */
2802 if (newval < leaf[buddy])
2805 /* check which (leafno or buddy) is the left buddy.
2806 * the left buddy gets to claim the blocks resulting
2807 * from the join while the right gets to claim none.
2808 * the left buddy is also eligible to participate in
2809 * a join at the next higher level while the right
2813 if (leafno < buddy) {
2814 /* leafno is the left buddy.
2816 dbAdjTree(tp, buddy, NOFREE);
2818 /* buddy is the left buddy and becomes
2821 dbAdjTree(tp, leafno, NOFREE);
2825 /* on to try the next join.
2832 /* update the leaf value.
2834 dbAdjTree(tp, leafno, newval);
2843 * FUNCTION: update a leaf of a dmtree with a new value, adjusting
2844 * the dmtree, as required, to reflect the new leaf value.
2845 * the combination of any buddies must already be done before
2849 * tp - pointer to the tree to be adjusted.
2850 * leafno - the number of the leaf to be updated.
2851 * newval - the new value for the leaf.
2853 * RETURN VALUES: none
2855 static void dbAdjTree(dmtree_t * tp, int leafno, int newval)
2860 /* pick up the index of the leaf for this leafno.
2862 lp = leafno + le32_to_cpu(tp->dmt_leafidx);
2864 /* is the current value the same as the old value ? if so,
2865 * there is nothing to do.
2867 if (tp->dmt_stree[lp] == newval)
2870 /* set the new value.
2872 tp->dmt_stree[lp] = newval;
2874 /* bubble the new value up the tree as required.
2876 for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) {
2877 /* get the index of the first leaf of the 4 leaf
2878 * group containing the specified leaf (leafno).
2880 lp = ((lp - 1) & ~0x03) + 1;
2882 /* get the index of the parent of this 4 leaf group.
2886 /* determine the maximum of the 4 leaves.
2888 max = TREEMAX(&tp->dmt_stree[lp]);
2890 /* if the maximum of the 4 is the same as the
2891 * parent's value, we're done.
2893 if (tp->dmt_stree[pp] == max)
2896 /* parent gets new value.
2898 tp->dmt_stree[pp] = max;
2900 /* parent becomes leaf for next go-round.
2908 * NAME: dbFindLeaf()
2910 * FUNCTION: search a dmtree_t for sufficient free blocks, returning
2911 * the index of a leaf describing the free blocks if
2912 * sufficient free blocks are found.
2914 * the search starts at the top of the dmtree_t tree and
2915 * proceeds down the tree to the leftmost leaf with sufficient
2919 * tp - pointer to the tree to be searched.
2920 * l2nb - log2 number of free blocks to search for.
2921 * leafidx - return pointer to be set to the index of the leaf
2922 * describing at least l2nb free blocks if sufficient
2923 * free blocks are found.
2927 * -ENOSPC - insufficient free blocks.
2929 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx)
2931 int ti, n = 0, k, x = 0;
2933 /* first check the root of the tree to see if there is
2934 * sufficient free space.
2936 if (l2nb > tp->dmt_stree[ROOT])
2939 /* sufficient free space available. now search down the tree
2940 * starting at the next level for the leftmost leaf that
2941 * describes sufficient free space.
2943 for (k = le32_to_cpu(tp->dmt_height), ti = 1;
2944 k > 0; k--, ti = ((ti + n) << 2) + 1) {
2945 /* search the four nodes at this level, starting from
2948 for (x = ti, n = 0; n < 4; n++) {
2949 /* sufficient free space found. move to the next
2950 * level (or quit if this is the last level).
2952 if (l2nb <= tp->dmt_stree[x + n])
2956 /* better have found something since the higher
2957 * levels of the tree said it was here.
2962 /* set the return to the leftmost leaf describing sufficient
2965 *leafidx = x + n - le32_to_cpu(tp->dmt_leafidx);
2972 * NAME: dbFindBits()
2974 * FUNCTION: find a specified number of binary buddy free bits within a
2975 * dmap bitmap word value.
2977 * this routine searches the bitmap value for (1 << l2nb) free
2978 * bits at (1 << l2nb) alignments within the value.
2981 * word - dmap bitmap word value.
2982 * l2nb - number of free bits specified as a log2 number.
2985 * starting bit number of free bits.
2987 static int dbFindBits(u32 word, int l2nb)
2992 /* get the number of bits.
2995 assert(nb <= DBWORD);
2997 /* complement the word so we can use a mask (i.e. 0s represent
2998 * free bits) and compute the mask.
3001 mask = ONES << (DBWORD - nb);
3003 /* scan the word for nb free bits at nb alignments.
3005 for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) {
3006 if ((mask & word) == mask)
3012 /* return the bit number.
3019 * NAME: dbMaxBud(u8 *cp)
3021 * FUNCTION: determine the largest binary buddy string of free
3022 * bits within 32-bits of the map.
3025 * cp - pointer to the 32-bit value.
3028 * largest binary buddy of free bits within a dmap word.
3030 static int dbMaxBud(u8 * cp)
3032 signed char tmp1, tmp2;
3034 /* check if the wmap word is all free. if so, the
3035 * free buddy size is BUDMIN.
3037 if (*((uint *) cp) == 0)
3040 /* check if the wmap word is half free. if so, the
3041 * free buddy size is BUDMIN-1.
3043 if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0)
3044 return (BUDMIN - 1);
3046 /* not all free or half free. determine the free buddy
3047 * size thru table lookup using quarters of the wmap word.
3049 tmp1 = max(budtab[cp[2]], budtab[cp[3]]);
3050 tmp2 = max(budtab[cp[0]], budtab[cp[1]]);
3051 return (max(tmp1, tmp2));
3056 * NAME: cnttz(uint word)
3058 * FUNCTION: determine the number of trailing zeros within a 32-bit
3062 * value - 32-bit value to be examined.
3065 * count of trailing zeros
3067 static int cnttz(u32 word)
3071 for (n = 0; n < 32; n++, word >>= 1) {
3081 * NAME: cntlz(u32 value)
3083 * FUNCTION: determine the number of leading zeros within a 32-bit
3087 * value - 32-bit value to be examined.
3090 * count of leading zeros
3092 static int cntlz(u32 value)
3096 for (n = 0; n < 32; n++, value <<= 1) {
3097 if (value & HIGHORDER)
3105 * NAME: blkstol2(s64 nb)
3107 * FUNCTION: convert a block count to its log2 value. if the block
3108 * count is not a l2 multiple, it is rounded up to the next
3109 * larger l2 multiple.
3112 * nb - number of blocks
3115 * log2 number of blocks
3117 static int blkstol2(s64 nb)
3120 s64 mask; /* meant to be signed */
3122 mask = (s64) 1 << (64 - 1);
3124 /* count the leading bits.
3126 for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) {
3127 /* leading bit found.
3130 /* determine the l2 value.
3132 l2nb = (64 - 1) - l2nb;
3134 /* check if we need to round up.
3143 return 0; /* fix compiler warning */
3148 * NAME: dbAllocBottomUp()
3150 * FUNCTION: alloc the specified block range from the working block
3153 * the blocks will be alloc from the working map one dmap
3157 * ip - pointer to in-core inode;
3158 * blkno - starting block number to be freed.
3159 * nblocks - number of blocks to be freed.
3165 int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks)
3167 struct metapage *mp;
3171 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap;
3172 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap;
3174 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP);
3176 /* block to be allocated better be within the mapsize. */
3177 ASSERT(nblocks <= bmp->db_mapsize - blkno);
3180 * allocate the blocks a dmap at a time.
3183 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) {
3184 /* release previous dmap if any */
3189 /* get the buffer for the current dmap. */
3190 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage);
3191 mp = read_metapage(ipbmap, lblkno, PSIZE, 0);
3193 IREAD_UNLOCK(ipbmap);
3196 dp = (struct dmap *) mp->data;
3198 /* determine the number of blocks to be allocated from
3201 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1)));
3203 /* allocate the blocks. */
3204 if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) {
3205 release_metapage(mp);
3206 IREAD_UNLOCK(ipbmap);
3211 /* write the last buffer. */
3214 IREAD_UNLOCK(ipbmap);
3220 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno,
3224 int dbitno, word, rembits, nb, nwords, wbitno, agno;
3226 struct dmaptree *tp = (struct dmaptree *) & dp->tree;
3228 /* save the current value of the root (i.e. maximum free string)
3231 oldroot = tp->stree[ROOT];
3233 /* determine the bit number and word within the dmap of the
3236 dbitno = blkno & (BPERDMAP - 1);
3237 word = dbitno >> L2DBWORD;
3239 /* block range better be within the dmap */
3240 assert(dbitno + nblocks <= BPERDMAP);
3242 /* allocate the bits of the dmap's words corresponding to the block
3243 * range. not all bits of the first and last words may be contained
3244 * within the block range. if this is the case, we'll work against
3245 * those words (i.e. partial first and/or last) on an individual basis
3246 * (a single pass), allocating the bits of interest by hand and
3247 * updating the leaf corresponding to the dmap word. a single pass
3248 * will be used for all dmap words fully contained within the
3249 * specified range. within this pass, the bits of all fully contained
3250 * dmap words will be marked as free in a single shot and the leaves
3251 * will be updated. a single leaf may describe the free space of
3252 * multiple dmap words, so we may update only a subset of the actual
3253 * leaves corresponding to the dmap words of the block range.
3255 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) {
3256 /* determine the bit number within the word and
3257 * the number of bits within the word.
3259 wbitno = dbitno & (DBWORD - 1);
3260 nb = min(rembits, DBWORD - wbitno);
3262 /* check if only part of a word is to be allocated.
3265 /* allocate (set to 1) the appropriate bits within
3268 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb)
3273 /* one or more dmap words are fully contained
3274 * within the block range. determine how many
3275 * words and allocate (set to 1) the bits of these
3278 nwords = rembits >> L2DBWORD;
3279 memset(&dp->wmap[word], (int) ONES, nwords * 4);
3281 /* determine how many bits */
3282 nb = nwords << L2DBWORD;
3287 /* update the free count for this dmap */
3288 le32_add_cpu(&dp->nfree, -nblocks);
3290 /* reconstruct summary tree */
3295 /* if this allocation group is completely free,
3296 * update the highest active allocation group number
3297 * if this allocation group is the new max.
3299 agno = blkno >> bmp->db_agl2size;
3300 if (agno > bmp->db_maxag)
3301 bmp->db_maxag = agno;
3303 /* update the free count for the allocation group and map */
3304 bmp->db_agfree[agno] -= nblocks;
3305 bmp->db_nfree -= nblocks;
3309 /* if the root has not changed, done. */
3310 if (tp->stree[ROOT] == oldroot)
3313 /* root changed. bubble the change up to the dmap control pages.
3314 * if the adjustment of the upper level control pages fails,
3315 * backout the bit allocation (thus making everything consistent).
3317 if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0)))
3318 dbFreeBits(bmp, dp, blkno, nblocks);
3325 * NAME: dbExtendFS()
3327 * FUNCTION: extend bmap from blkno for nblocks;
3328 * dbExtendFS() updates bmap ready for dbAllocBottomUp();
3332 * L1---------------------------------L1
3334 * L0---------L0---------L0 L0---------L0---------L0
3336 * d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,.,dm;
3337 * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm
3339 * <---old---><----------------------------extend----------------------->
3341 int dbExtendFS(struct inode *ipbmap, s64 blkno, s64 nblocks)
3343 struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb);
3344 int nbperpage = sbi->nbperpage;
3345 int i, i0 = true, j, j0 = true, k, n;
3348 struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL;
3349 struct dmapctl *l2dcp, *l1dcp, *l0dcp;
3351 s8 *l0leaf, *l1leaf, *l2leaf;
3352 struct bmap *bmp = sbi->bmap;
3353 int agno, l2agsize, oldl2agsize;
3356 newsize = blkno + nblocks;
3358 jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld",
3359 (long long) blkno, (long long) nblocks, (long long) newsize);
3362 * initialize bmap control page.
3364 * all the data in bmap control page should exclude
3365 * the mkfs hidden dmap page.
3368 /* update mapsize */
3369 bmp->db_mapsize = newsize;
3370 bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize);
3372 /* compute new AG size */
3373 l2agsize = dbGetL2AGSize(newsize);
3374 oldl2agsize = bmp->db_agl2size;
3376 bmp->db_agl2size = l2agsize;
3377 bmp->db_agsize = 1 << l2agsize;
3379 /* compute new number of AG */
3380 agno = bmp->db_numag;
3381 bmp->db_numag = newsize >> l2agsize;
3382 bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0;
3385 * reconfigure db_agfree[]
3386 * from old AG configuration to new AG configuration;
3388 * coalesce contiguous k (newAGSize/oldAGSize) AGs;
3389 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn;
3390 * note: new AG size = old AG size * (2**x).
3392 if (l2agsize == oldl2agsize)
3394 k = 1 << (l2agsize - oldl2agsize);
3395 ag_rem = bmp->db_agfree[0]; /* save agfree[0] */
3396 for (i = 0, n = 0; i < agno; n++) {
3397 bmp->db_agfree[n] = 0; /* init collection point */
3399 /* coalesce contiguous k AGs; */
3400 for (j = 0; j < k && i < agno; j++, i++) {
3401 /* merge AGi to AGn */
3402 bmp->db_agfree[n] += bmp->db_agfree[i];
3405 bmp->db_agfree[0] += ag_rem; /* restore agfree[0] */
3407 for (; n < MAXAG; n++)
3408 bmp->db_agfree[n] = 0;
3411 * update highest active ag number
3414 bmp->db_maxag = bmp->db_maxag / k;
3419 * update bit maps and corresponding level control pages;
3420 * global control page db_nfree, db_agfree[agno], db_maxfreebud;
3424 p = BMAPBLKNO + nbperpage; /* L2 page */
3425 l2mp = read_metapage(ipbmap, p, PSIZE, 0);
3427 jfs_error(ipbmap->i_sb, "L2 page could not be read\n");
3430 l2dcp = (struct dmapctl *) l2mp->data;
3432 /* compute start L1 */
3433 k = blkno >> L2MAXL1SIZE;
3434 l2leaf = l2dcp->stree + CTLLEAFIND + k;
3435 p = BLKTOL1(blkno, sbi->l2nbperpage); /* L1 page */
3438 * extend each L1 in L2
3440 for (; k < LPERCTL; k++, p += nbperpage) {
3443 /* read in L1 page: (blkno & (MAXL1SIZE - 1)) */
3444 l1mp = read_metapage(ipbmap, p, PSIZE, 0);
3447 l1dcp = (struct dmapctl *) l1mp->data;
3449 /* compute start L0 */
3450 j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE;
3451 l1leaf = l1dcp->stree + CTLLEAFIND + j;
3452 p = BLKTOL0(blkno, sbi->l2nbperpage);
3455 /* assign/init L1 page */
3456 l1mp = get_metapage(ipbmap, p, PSIZE, 0);
3460 l1dcp = (struct dmapctl *) l1mp->data;
3462 /* compute start L0 */
3464 l1leaf = l1dcp->stree + CTLLEAFIND;
3465 p += nbperpage; /* 1st L0 of L1.k */
3469 * extend each L0 in L1
3471 for (; j < LPERCTL; j++) {
3474 /* read in L0 page: (blkno & (MAXL0SIZE - 1)) */
3476 l0mp = read_metapage(ipbmap, p, PSIZE, 0);
3479 l0dcp = (struct dmapctl *) l0mp->data;
3481 /* compute start dmap */
3482 i = (blkno & (MAXL0SIZE - 1)) >>
3484 l0leaf = l0dcp->stree + CTLLEAFIND + i;
3485 p = BLKTODMAP(blkno,
3489 /* assign/init L0 page */
3490 l0mp = get_metapage(ipbmap, p, PSIZE, 0);
3494 l0dcp = (struct dmapctl *) l0mp->data;
3496 /* compute start dmap */
3498 l0leaf = l0dcp->stree + CTLLEAFIND;
3499 p += nbperpage; /* 1st dmap of L0.j */
3503 * extend each dmap in L0
3505 for (; i < LPERCTL; i++) {
3507 * reconstruct the dmap page, and
3508 * initialize corresponding parent L0 leaf
3510 if ((n = blkno & (BPERDMAP - 1))) {
3511 /* read in dmap page: */
3512 mp = read_metapage(ipbmap, p,
3516 n = min(nblocks, (s64)BPERDMAP - n);
3518 /* assign/init dmap page */
3519 mp = read_metapage(ipbmap, p,
3524 n = min_t(s64, nblocks, BPERDMAP);
3527 dp = (struct dmap *) mp->data;
3528 *l0leaf = dbInitDmap(dp, blkno, n);
3531 agno = le64_to_cpu(dp->start) >> l2agsize;
3532 bmp->db_agfree[agno] += n;
3543 } /* for each dmap in a L0 */
3546 * build current L0 page from its leaves, and
3547 * initialize corresponding parent L1 leaf
3549 *l1leaf = dbInitDmapCtl(l0dcp, 0, ++i);
3550 write_metapage(l0mp);
3554 l1leaf++; /* continue for next L0 */
3556 /* more than 1 L0 ? */
3558 break; /* build L1 page */
3560 /* summarize in global bmap page */
3561 bmp->db_maxfreebud = *l1leaf;
3562 release_metapage(l1mp);
3563 release_metapage(l2mp);
3567 } /* for each L0 in a L1 */
3570 * build current L1 page from its leaves, and
3571 * initialize corresponding parent L2 leaf
3573 *l2leaf = dbInitDmapCtl(l1dcp, 1, ++j);
3574 write_metapage(l1mp);
3578 l2leaf++; /* continue for next L1 */
3580 /* more than 1 L1 ? */
3582 break; /* build L2 page */
3584 /* summarize in global bmap page */
3585 bmp->db_maxfreebud = *l2leaf;
3586 release_metapage(l2mp);
3590 } /* for each L1 in a L2 */
3592 jfs_error(ipbmap->i_sb, "function has not returned as expected\n");
3595 release_metapage(l0mp);
3597 release_metapage(l1mp);
3598 release_metapage(l2mp);
3602 * finalize bmap control page
3613 void dbFinalizeBmap(struct inode *ipbmap)
3615 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap;
3616 int actags, inactags, l2nl;
3617 s64 ag_rem, actfree, inactfree, avgfree;
3621 * finalize bmap control page
3625 * compute db_agpref: preferred ag to allocate from
3626 * (the leftmost ag with average free space in it);
3629 /* get the number of active ags and inactive ags */
3630 actags = bmp->db_maxag + 1;
3631 inactags = bmp->db_numag - actags;
3632 ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1); /* ??? */
3634 /* determine how many blocks are in the inactive allocation
3635 * groups. in doing this, we must account for the fact that
3636 * the rightmost group might be a partial group (i.e. file
3637 * system size is not a multiple of the group size).
3639 inactfree = (inactags && ag_rem) ?
3640 ((inactags - 1) << bmp->db_agl2size) + ag_rem
3641 : inactags << bmp->db_agl2size;
3643 /* determine how many free blocks are in the active
3644 * allocation groups plus the average number of free blocks
3645 * within the active ags.
3647 actfree = bmp->db_nfree - inactfree;
3648 avgfree = (u32) actfree / (u32) actags;
3650 /* if the preferred allocation group has not average free space.
3651 * re-establish the preferred group as the leftmost
3652 * group with average free space.
3654 if (bmp->db_agfree[bmp->db_agpref] < avgfree) {
3655 for (bmp->db_agpref = 0; bmp->db_agpref < actags;
3657 if (bmp->db_agfree[bmp->db_agpref] >= avgfree)
3660 if (bmp->db_agpref >= bmp->db_numag) {
3661 jfs_error(ipbmap->i_sb,
3662 "cannot find ag with average freespace\n");
3667 * compute db_aglevel, db_agheight, db_width, db_agstart:
3668 * an ag is covered in aglevel dmapctl summary tree,
3669 * at agheight level height (from leaf) with agwidth number of nodes
3670 * each, which starts at agstart index node of the smmary tree node
3673 bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize);
3675 bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL);
3676 bmp->db_agheight = l2nl >> 1;
3677 bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1));
3678 for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0;
3680 bmp->db_agstart += n;
3688 * NAME: dbInitDmap()/ujfs_idmap_page()
3690 * FUNCTION: initialize working/persistent bitmap of the dmap page
3691 * for the specified number of blocks:
3693 * at entry, the bitmaps had been initialized as free (ZEROS);
3694 * The number of blocks will only account for the actually
3695 * existing blocks. Blocks which don't actually exist in
3696 * the aggregate will be marked as allocated (ONES);
3699 * dp - pointer to page of map
3700 * nblocks - number of blocks this page
3704 static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks)
3706 int blkno, w, b, r, nw, nb, i;
3708 /* starting block number within the dmap */
3709 blkno = Blkno & (BPERDMAP - 1);
3712 dp->nblocks = dp->nfree = cpu_to_le32(nblocks);
3713 dp->start = cpu_to_le64(Blkno);
3715 if (nblocks == BPERDMAP) {
3716 memset(&dp->wmap[0], 0, LPERDMAP * 4);
3717 memset(&dp->pmap[0], 0, LPERDMAP * 4);
3721 le32_add_cpu(&dp->nblocks, nblocks);
3722 le32_add_cpu(&dp->nfree, nblocks);
3725 /* word number containing start block number */
3726 w = blkno >> L2DBWORD;
3729 * free the bits corresponding to the block range (ZEROS):
3730 * note: not all bits of the first and last words may be contained
3731 * within the block range.
3733 for (r = nblocks; r > 0; r -= nb, blkno += nb) {
3734 /* number of bits preceding range to be freed in the word */
3735 b = blkno & (DBWORD - 1);
3736 /* number of bits to free in the word */
3737 nb = min(r, DBWORD - b);
3739 /* is partial word to be freed ? */
3741 /* free (set to 0) from the bitmap word */
3742 dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3744 dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb)
3747 /* skip the word freed */
3750 /* free (set to 0) contiguous bitmap words */
3752 memset(&dp->wmap[w], 0, nw * 4);
3753 memset(&dp->pmap[w], 0, nw * 4);
3755 /* skip the words freed */
3756 nb = nw << L2DBWORD;
3762 * mark bits following the range to be freed (non-existing
3763 * blocks) as allocated (ONES)
3766 if (blkno == BPERDMAP)
3769 /* the first word beyond the end of existing blocks */
3770 w = blkno >> L2DBWORD;
3772 /* does nblocks fall on a 32-bit boundary ? */
3773 b = blkno & (DBWORD - 1);
3775 /* mark a partial word allocated */
3776 dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b);
3780 /* set the rest of the words in the page to allocated (ONES) */
3781 for (i = w; i < LPERDMAP; i++)
3782 dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES);
3788 return (dbInitDmapTree(dp));
3793 * NAME: dbInitDmapTree()/ujfs_complete_dmap()
3795 * FUNCTION: initialize summary tree of the specified dmap:
3797 * at entry, bitmap of the dmap has been initialized;
3800 * dp - dmap to complete
3801 * blkno - starting block number for this dmap
3802 * treemax - will be filled in with max free for this dmap
3804 * RETURNS: max free string at the root of the tree
3806 static int dbInitDmapTree(struct dmap * dp)
3808 struct dmaptree *tp;
3812 /* init fixed info of tree */
3814 tp->nleafs = cpu_to_le32(LPERDMAP);
3815 tp->l2nleafs = cpu_to_le32(L2LPERDMAP);
3816 tp->leafidx = cpu_to_le32(LEAFIND);
3817 tp->height = cpu_to_le32(4);
3818 tp->budmin = BUDMIN;
3820 /* init each leaf from corresponding wmap word:
3821 * note: leaf is set to NOFREE(-1) if all blocks of corresponding
3822 * bitmap word are allocated.
3824 cp = tp->stree + le32_to_cpu(tp->leafidx);
3825 for (i = 0; i < LPERDMAP; i++)
3826 *cp++ = dbMaxBud((u8 *) & dp->wmap[i]);
3828 /* build the dmap's binary buddy summary tree */
3829 return (dbInitTree(tp));
3834 * NAME: dbInitTree()/ujfs_adjtree()
3836 * FUNCTION: initialize binary buddy summary tree of a dmap or dmapctl.
3838 * at entry, the leaves of the tree has been initialized
3839 * from corresponding bitmap word or root of summary tree
3840 * of the child control page;
3841 * configure binary buddy system at the leaf level, then
3842 * bubble up the values of the leaf nodes up the tree.
3845 * cp - Pointer to the root of the tree
3846 * l2leaves- Number of leaf nodes as a power of 2
3847 * l2min - Number of blocks that can be covered by a leaf
3850 * RETURNS: max free string at the root of the tree
3852 static int dbInitTree(struct dmaptree * dtp)
3854 int l2max, l2free, bsize, nextb, i;
3855 int child, parent, nparent;
3860 /* Determine the maximum free string possible for the leaves */
3861 l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin;
3864 * configure the leaf level into binary buddy system
3866 * Try to combine buddies starting with a buddy size of 1
3867 * (i.e. two leaves). At a buddy size of 1 two buddy leaves
3868 * can be combined if both buddies have a maximum free of l2min;
3869 * the combination will result in the left-most buddy leaf having
3870 * a maximum free of l2min+1.
3871 * After processing all buddies for a given size, process buddies
3872 * at the next higher buddy size (i.e. current size * 2) and
3873 * the next maximum free (current free + 1).
3874 * This continues until the maximum possible buddy combination
3875 * yields maximum free.
3877 for (l2free = dtp->budmin, bsize = 1; l2free < l2max;
3878 l2free++, bsize = nextb) {
3879 /* get next buddy size == current buddy pair size */
3882 /* scan each adjacent buddy pair at current buddy size */
3883 for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx);
3884 i < le32_to_cpu(dtp->nleafs);
3885 i += nextb, cp += nextb) {
3886 /* coalesce if both adjacent buddies are max free */
3887 if (*cp == l2free && *(cp + bsize) == l2free) {
3888 *cp = l2free + 1; /* left take right */
3889 *(cp + bsize) = -1; /* right give left */
3895 * bubble summary information of leaves up the tree.
3897 * Starting at the leaf node level, the four nodes described by
3898 * the higher level parent node are compared for a maximum free and
3899 * this maximum becomes the value of the parent node.
3900 * when all lower level nodes are processed in this fashion then
3901 * move up to the next level (parent becomes a lower level node) and
3902 * continue the process for that level.
3904 for (child = le32_to_cpu(dtp->leafidx),
3905 nparent = le32_to_cpu(dtp->nleafs) >> 2;
3906 nparent > 0; nparent >>= 2, child = parent) {
3907 /* get index of 1st node of parent level */
3908 parent = (child - 1) >> 2;
3910 /* set the value of the parent node as the maximum
3911 * of the four nodes of the current level.
3913 for (i = 0, cp = tp + child, cp1 = tp + parent;
3914 i < nparent; i++, cp += 4, cp1++)
3925 * function: initialize dmapctl page
3927 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i)
3928 { /* start leaf index not covered by range */
3931 dcp->nleafs = cpu_to_le32(LPERCTL);
3932 dcp->l2nleafs = cpu_to_le32(L2LPERCTL);
3933 dcp->leafidx = cpu_to_le32(CTLLEAFIND);
3934 dcp->height = cpu_to_le32(5);
3935 dcp->budmin = L2BPERDMAP + L2LPERCTL * level;
3938 * initialize the leaves of current level that were not covered
3939 * by the specified input block range (i.e. the leaves have no
3940 * low level dmapctl or dmap).
3942 cp = &dcp->stree[CTLLEAFIND + i];
3943 for (; i < LPERCTL; i++)
3946 /* build the dmap's binary buddy summary tree */
3947 return (dbInitTree((struct dmaptree *) dcp));
3952 * NAME: dbGetL2AGSize()/ujfs_getagl2size()
3954 * FUNCTION: Determine log2(allocation group size) from aggregate size
3957 * nblocks - Number of blocks in aggregate
3959 * RETURNS: log2(allocation group size) in aggregate blocks
3961 static int dbGetL2AGSize(s64 nblocks)
3967 if (nblocks < BPERDMAP * MAXAG)
3968 return (L2BPERDMAP);
3970 /* round up aggregate size to power of 2 */
3971 m = ((u64) 1 << (64 - 1));
3972 for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) {
3977 sz = (s64) 1 << l2sz;
3981 /* agsize = roundupSize/max_number_of_ag */
3982 return (l2sz - L2MAXAG);
3987 * NAME: dbMapFileSizeToMapSize()
3989 * FUNCTION: compute number of blocks the block allocation map file
3990 * can cover from the map file size;
3992 * RETURNS: Number of blocks which can be covered by this block map file;
3996 * maximum number of map pages at each level including control pages
3998 #define MAXL0PAGES (1 + LPERCTL)
3999 #define MAXL1PAGES (1 + LPERCTL * MAXL0PAGES)
4002 * convert number of map pages to the zero origin top dmapctl level
4004 #define BMAPPGTOLEV(npages) \
4005 (((npages) <= 3 + MAXL0PAGES) ? 0 : \
4006 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2)
4008 s64 dbMapFileSizeToMapSize(struct inode * ipbmap)
4010 struct super_block *sb = ipbmap->i_sb;
4014 int complete, factor;
4016 nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize;
4017 npages = nblocks >> JFS_SBI(sb)->l2nbperpage;
4018 level = BMAPPGTOLEV(npages);
4020 /* At each level, accumulate the number of dmap pages covered by
4021 * the number of full child levels below it;
4022 * repeat for the last incomplete child level.
4025 npages--; /* skip the first global control page */
4026 /* skip higher level control pages above top level covered by map */
4027 npages -= (2 - level);
4028 npages--; /* skip top level's control page */
4029 for (i = level; i >= 0; i--) {
4031 (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1);
4032 complete = (u32) npages / factor;
4033 ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL :
4034 ((i == 1) ? LPERCTL : 1));
4036 /* pages in last/incomplete child */
4037 npages = (u32) npages % factor;
4038 /* skip incomplete child's level control page */
4042 /* convert the number of dmaps into the number of blocks
4043 * which can be covered by the dmaps;
4045 nblocks = ndmaps << L2BPERDMAP;